v2:
- fixed "make check" error that patchew reported
- moved the thread_join upper in monitor_data_destroy(), before
  resources are released
- added one new patch (current patch 3) that fixes a nasty risk
  condition with IOWatchPoll.  Please see commit message for more
  information.
- added a g_main_context_wakeup() to make sure the separate loop
  thread can be kicked always when we want to destroy the per-monitor
  threads.
- added one new patch (current patch 8) to introduce migration mgmt
  lock for migrate_incoming.

This is an extended work for migration postcopy recovery. This series
is tested with the following series to make sure it solves the monitor
hang problem that we have encountered for postcopy recovery:

  [RFC 00/29] Migration: postcopy failure recovery
  [RFC 0/6] migration: re-use migrate_incoming for postcopy recovery

The root problem is that, monitor commands are all handled in main
loop thread now, no matter how many monitors we specify. And, if main
loop thread hangs due to some reason, all monitors will be stuck.
This can be done in reversed order as well: if any of the monitor
hangs, it will hang the main loop, and the rest of the monitors (if
there is any).

That affects postcopy recovery, since the recovery requires user input
on destination side.  If monitors hang, the destination VM dies and
lose hope for even a final recovery.

So, sometimes we need to make sure the monitor be alive, at least one
of them.

The whole idea of this series is that instead if handling monitor
commands all in main loop thread, we do it separately in per-monitor
threads.  Then, even if main loop thread hangs at any point by any
reason, per-monitor thread can still survive.  Further, we add hint in
QMP/HMP to show whether a command can be executed without QMP, if so,
we avoid taking BQL when running that command.  It greatly reduced
contention of BQL.  Now the only user of that new parameter (currently
I call it "without-bql") is "migrate-incoming" command, which is the
only command to rescue a paused postcopy migration.

However, even with the series, it does not mean that per-monitor
threads will never hang.  One example is that we can still run "info
vcpus" in per-monitor threads during a paused postcopy (in that state,
page faults are never handled, and "info cpus" will never return since
it tries to sync every vcpus).  So to make sure it does not hang, we
not only need the per-monitor thread, the user should be careful as
well on how to use it.

For postcopy recovery, we may need dedicated monitor channel for
recovery.  In other words, a destination VM that supports postcopy
recovery would possibly need:

  -qmp MAIN_CHANNEL -qmp RECOVERY_CHANNEL

Here, the MAIN_CHANNEL can be MUXed and shared by other chardev
frontends, while the RECOVERY_CHANNEL should *ONLY* be used to input
the "migrate-incoming" command (similar thing applies to HMP
channels).  As long as we are following this rule, the
RECOVERY_CHANNEL can never hang.

Some details on each patch:

Patch 1: a simple cleanup only

Patch 2: allow monitors to create per-monitor thread to handle monitor
         command requests. Since monitor is only one type of chardev
         frontend, we only do this if the backend is dedicated, say,
         if MUX is not turned on (if MUX is on, it's still using main
         loop thread).

Patch 3: based on patch 2, this patch introduced a new parameter for
         QMP commands called "without-bql", it is a hint that the
         command does not need BQL.

Patch 4: Let QMP command "migrate-incoming" avoid taking BQL.

Patch 5: Introduced sister parameter for HMP "without_bql", which
         works just like QMP "without-bql".

Patch 6: Let HMP command "migrate-incoming" avoid taking BQL.

Please review. Thanks,

Peter Xu (8):
  monitor: move skip_flush into monitor_data_init
  monitor: allow monitor to create thread to poll
  char-io: fix possible risk on IOWatchPoll
  QAPI: new QMP command option "without-bql"
  hmp: support "without_bql"
  migration: qmp: migrate_incoming don't need BQL
  migration: hmp: migrate_incoming don't need BQL
  migration: add incoming mgmt lock

 chardev/char-io.c              | 15 +++++++-
 docs/devel/qapi-code-gen.txt   | 10 ++++-
 hmp-commands.hx                |  1 +
 include/qapi/qmp/dispatch.h    |  1 +
 migration/migration.c          |  6 +++
 migration/migration.h          |  3 ++
 monitor.c                      | 87 +++++++++++++++++++++++++++++++++++++++---
 qapi-schema.json               |  3 +-
 qapi/qmp-dispatch.c            | 26 +++++++++++++
 scripts/qapi-commands.py       | 18 ++++++---
 scripts/qapi-introspect.py     |  2 +-
 scripts/qapi.py                | 15 +++++---
 scripts/qapi2texi.py           |  2 +-
 tests/qapi-schema/test-qapi.py |  2 +-
 14 files changed, 168 insertions(+), 23 deletions(-)

-- 
2.7.4

On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> v2:
> - fixed "make check" error that patchew reported
> - moved the thread_join upper in monitor_data_destroy(), before
>   resources are released
> - added one new patch (current patch 3) that fixes a nasty risk
>   condition with IOWatchPoll.  Please see commit message for more
>   information.
> - added a g_main_context_wakeup() to make sure the separate loop
>   thread can be kicked always when we want to destroy the per-monitor
>   threads.
> - added one new patch (current patch 8) to introduce migration mgmt
>   lock for migrate_incoming.
> 
> This is an extended work for migration postcopy recovery. This series
> is tested with the following series to make sure it solves the monitor
> hang problem that we have encountered for postcopy recovery:
> 
>   [RFC 00/29] Migration: postcopy failure recovery
>   [RFC 0/6] migration: re-use migrate_incoming for postcopy recovery
> 
> The root problem is that, monitor commands are all handled in main
> loop thread now, no matter how many monitors we specify. And, if main
> loop thread hangs due to some reason, all monitors will be stuck.
> This can be done in reversed order as well: if any of the monitor
> hangs, it will hang the main loop, and the rest of the monitors (if
> there is any).
> 
> That affects postcopy recovery, since the recovery requires user input
> on destination side.  If monitors hang, the destination VM dies and
> lose hope for even a final recovery.
> 
> So, sometimes we need to make sure the monitor be alive, at least one
> of them.
> 
> The whole idea of this series is that instead if handling monitor
> commands all in main loop thread, we do it separately in per-monitor
> threads.  Then, even if main loop thread hangs at any point by any
> reason, per-monitor thread can still survive.  Further, we add hint in
> QMP/HMP to show whether a command can be executed without QMP, if so,
> we avoid taking BQL when running that command.  It greatly reduced
> contention of BQL.  Now the only user of that new parameter (currently
> I call it "without-bql") is "migrate-incoming" command, which is the
> only command to rescue a paused postcopy migration.
> 
> However, even with the series, it does not mean that per-monitor
> threads will never hang.  One example is that we can still run "info
> vcpus" in per-monitor threads during a paused postcopy (in that state,
> page faults are never handled, and "info cpus" will never return since
> it tries to sync every vcpus).  So to make sure it does not hang, we
> not only need the per-monitor thread, the user should be careful as
> well on how to use it.
> 
> For postcopy recovery, we may need dedicated monitor channel for
> recovery.  In other words, a destination VM that supports postcopy
> recovery would possibly need:
> 
>   -qmp MAIN_CHANNEL -qmp RECOVERY_CHANNEL

I think this is a really horrible thing to expose to management applications.
They should not need to be aware of fact that QEMU is buggy and thus requires
that certain commands be run on different monitors to work around the bug.

I'd much prefer to see the problem described handled transparently inside
QEMU. One approach is have a dedicated thread in QEMU responsible for all
monitor I/O. This thread should never actually execute monitor commands
though, it would simply parse the command request and put data onto a queue
of pending commands, thus it could never hang. The command queue could be
processed by the main thread, or by another thread that is interested.
eg the migration thread could process any queued commands related to
migration directly.

Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

"Daniel P. Berrange" <berrange@redhat.com> writes:

> On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
>> v2:
>> - fixed "make check" error that patchew reported
>> - moved the thread_join upper in monitor_data_destroy(), before
>>   resources are released
>> - added one new patch (current patch 3) that fixes a nasty risk
>>   condition with IOWatchPoll.  Please see commit message for more
>>   information.
>> - added a g_main_context_wakeup() to make sure the separate loop
>>   thread can be kicked always when we want to destroy the per-monitor
>>   threads.
>> - added one new patch (current patch 8) to introduce migration mgmt
>>   lock for migrate_incoming.
>> 
>> This is an extended work for migration postcopy recovery. This series
>> is tested with the following series to make sure it solves the monitor
>> hang problem that we have encountered for postcopy recovery:
>> 
>>   [RFC 00/29] Migration: postcopy failure recovery
>>   [RFC 0/6] migration: re-use migrate_incoming for postcopy recovery
>> 
>> The root problem is that, monitor commands are all handled in main
>> loop thread now, no matter how many monitors we specify. And, if main
>> loop thread hangs due to some reason, all monitors will be stuck.
>> This can be done in reversed order as well: if any of the monitor
>> hangs, it will hang the main loop, and the rest of the monitors (if
>> there is any).

Yes.

>> That affects postcopy recovery, since the recovery requires user input
>> on destination side.  If monitors hang, the destination VM dies and
>> lose hope for even a final recovery.
>> 
>> So, sometimes we need to make sure the monitor be alive, at least one
>> of them.
>> 
>> The whole idea of this series is that instead if handling monitor
>> commands all in main loop thread, we do it separately in per-monitor
>> threads.  Then, even if main loop thread hangs at any point by any
>> reason, per-monitor thread can still survive.

This takes care of "monitor hangs because other parts of the main loop
(including other monitors) hang".  It doesn't take care of "monitor
hangs because the current monitor command hangs".

>>                                                Further, we add hint in
>> QMP/HMP to show whether a command can be executed without QMP, if so,
>> we avoid taking BQL when running that command.  It greatly reduced
>> contention of BQL.  Now the only user of that new parameter (currently
>> I call it "without-bql") is "migrate-incoming" command, which is the
>> only command to rescue a paused postcopy migration.

This takes care of one way commands can hang.  There are other ways;
NFS server going AWOL is a classic.  I don't know whether any other way
applies to migrate-incoming.

>> However, even with the series, it does not mean that per-monitor
>> threads will never hang.  One example is that we can still run "info
>> vcpus" in per-monitor threads during a paused postcopy (in that state,
>> page faults are never handled, and "info cpus" will never return since
>> it tries to sync every vcpus).  So to make sure it does not hang, we
>> not only need the per-monitor thread, the user should be careful as
>> well on how to use it.
>> 
>> For postcopy recovery, we may need dedicated monitor channel for
>> recovery.  In other words, a destination VM that supports postcopy
>> recovery would possibly need:
>> 
>>   -qmp MAIN_CHANNEL -qmp RECOVERY_CHANNEL

Where RECOVERY_CHANNEL isn't necessarily just for postcopy, but for any
"emergency" QMP access.  If you use it only for commands that cannot
hang (i.e. terminate in bounded time), then you'll always be able to get
commands accepted there in bounded time.

> I think this is a really horrible thing to expose to management applications.
> They should not need to be aware of fact that QEMU is buggy and thus requires
> that certain commands be run on different monitors to work around the bug.

These are (serious) design limitations, not bugs in the narrow sense of
the word.

However, I quite agree that the need for clients to know whether a
monitor command can hang is impractical for the general case.  What
might be practical is a QMP monitor mode that accepts only known
hang-free commands.  Hang-free could be introspectable.

In case you consider that ugly: it's best to explore the design space
first, and recoil from "ugly" second.

> I'd much prefer to see the problem described handled transparently inside
> QEMU. One approach is have a dedicated thread in QEMU responsible for all
> monitor I/O. This thread should never actually execute monitor commands
> though, it would simply parse the command request and put data onto a queue
> of pending commands, thus it could never hang. The command queue could be
> processed by the main thread, or by another thread that is interested.
> eg the migration thread could process any queued commands related to
> migration directly.

The monitor itself can't hang then, but the thread(s) dequeuing parsed
commands can.

To maintain commands' synchronous semantics, their replies need to be
sent in order, which of course reintroduces the hangs.

Let's take a step back from the implementation, and talk about
*behavior* instead.

You prefer to have "the problem described handled transparently inside
QEMU".  I read that as "QEMU must ensure the QMP monitor is available at
all times".  "Available" means it accepts commands in bounded time.
Some commands will always finish in bounded time once accepted, others
may not, and whether they do may depend on the commands currently in
flight.

Commands that can always start and always terminate in bounded time are
no problem.

All the other commands have to become "job-starting": the QMP command
kicks off a "job", which runs concurrently with the QMP monitor for some
(possibly unbounded) time, then finishes.  Jobs can be examined (say to
monitor progress, if the job supports that) and controlled (say to
cancel, if the job supports that).

A few commands are already job-starting: migrate, the block job family,
dump-guest-memory with detach=true.  Whether they're already hang-free I
can't say; they could do risky work in their synchronous part.

Many commands that can hang are not job-starting.

Changing a command from "do the job" to "merely start the job" is a
compatibility break.

We could make the change opt-in to preserve compatibility.  But is
preserving a compatible QMP monitor that is prone to hang wortwhile?

If no, we may choose to use the resulting compatibility break to also
switch the packaging of jobs from the current "synchronous command +
broadcast message when done" to some variation of asynchronous command.
But that should be discussed in a separate thread, and only after we
know how we plan to ensure monitor availability.

On Wed, Aug 30, 2017 at 09:06:20AM +0200, Markus Armbruster wrote:
> "Daniel P. Berrange" <berrange@redhat.com> writes:
> 
> > On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> 
> >> However, even with the series, it does not mean that per-monitor
> >> threads will never hang.  One example is that we can still run "info
> >> vcpus" in per-monitor threads during a paused postcopy (in that state,
> >> page faults are never handled, and "info cpus" will never return since
> >> it tries to sync every vcpus).  So to make sure it does not hang, we
> >> not only need the per-monitor thread, the user should be careful as
> >> well on how to use it.
> >> 
> >> For postcopy recovery, we may need dedicated monitor channel for
> >> recovery.  In other words, a destination VM that supports postcopy
> >> recovery would possibly need:
> >> 
> >>   -qmp MAIN_CHANNEL -qmp RECOVERY_CHANNEL
> 
> Where RECOVERY_CHANNEL isn't necessarily just for postcopy, but for any
> "emergency" QMP access.  If you use it only for commands that cannot
> hang (i.e. terminate in bounded time), then you'll always be able to get
> commands accepted there in bounded time.
> 
> > I think this is a really horrible thing to expose to management applications.
> > They should not need to be aware of fact that QEMU is buggy and thus requires
> > that certain commands be run on different monitors to work around the bug.
> 
> These are (serious) design limitations, not bugs in the narrow sense of
> the word.
> 
> However, I quite agree that the need for clients to know whether a
> monitor command can hang is impractical for the general case.  What
> might be practical is a QMP monitor mode that accepts only known
> hang-free commands.  Hang-free could be introspectable.
> 
> In case you consider that ugly: it's best to explore the design space
> first, and recoil from "ugly" second.

Actually you slightly mis-interpreted me there. I think it is ok for
applications to have knowledge about whether a particular command
may hang or not. Given that knowledge it should *not*, however, require
that the application issue such commands on separate monitor channels.
It is entirely possible to handle hang-free commands on the existing
channel.

> > I'd much prefer to see the problem described handled transparently inside
> > QEMU. One approach is have a dedicated thread in QEMU responsible for all
> > monitor I/O. This thread should never actually execute monitor commands
> > though, it would simply parse the command request and put data onto a queue
> > of pending commands, thus it could never hang. The command queue could be
> > processed by the main thread, or by another thread that is interested.
> > eg the migration thread could process any queued commands related to
> > migration directly.
> 
> The monitor itself can't hang then, but the thread(s) dequeuing parsed
> commands can.

If certain commands are hang-free then you can have a dedicated thread
that only de-queues & processes the hang-free commands. The approach I
outlined is exactly how libvirt deals with its own RPC dispatch. We have
certain commands that are guaranteed to not hang, which are processed by
a dedicated pool of threads. So even if all normal RPC commands have
hung, you can still run a subset of hang-free RPC commands.

> 
> To maintain commands' synchronous semantics, their replies need to be
> sent in order, which of course reintroduces the hangs.

The requirement for such ordering is just an arbitrary restriction that
QEMU currently imposes. It is reasonable to allow arbitrary ordering of
responses (which is what libvirt does in its RPC layer). Admittedly at
this stage though, we would likely require some "opt in" handshake when
initializing QMP for the app to tell QEMU it can cope with out of order
replies. It would require that each command request has a unique serial
number, which is included in the associated reply, so apps can match
them up. We used to have that but iirc it was then removed.

There's other ways to deal with this, such as the job starting idea you
mention below.

The key point though is that I don't think creating multiple monitor
servers is a desirable approach - it is just a hack to avoid dealing
with the root cause problems. 

> Let's take a step back from the implementation, and talk about
> *behavior* instead.
> 
> You prefer to have "the problem described handled transparently inside
> QEMU".  I read that as "QEMU must ensure the QMP monitor is available at
> all times".  "Available" means it accepts commands in bounded time.
> Some commands will always finish in bounded time once accepted, others
> may not, and whether they do may depend on the commands currently in
> flight.
> 
> Commands that can always start and always terminate in bounded time are
> no problem.
> 
> All the other commands have to become "job-starting": the QMP command
> kicks off a "job", which runs concurrently with the QMP monitor for some
> (possibly unbounded) time, then finishes.  Jobs can be examined (say to
> monitor progress, if the job supports that) and controlled (say to
> cancel, if the job supports that).
> 
> A few commands are already job-starting: migrate, the block job family,
> dump-guest-memory with detach=true.  Whether they're already hang-free I
> can't say; they could do risky work in their synchronous part.
> 
> Many commands that can hang are not job-starting.
> 
> Changing a command from "do the job" to "merely start the job" is a
> compatibility break.
> 
> We could make the change opt-in to preserve compatibility.  But is
> preserving a compatible QMP monitor that is prone to hang wortwhile?
> 
> If no, we may choose to use the resulting compatibility break to also
> switch the packaging of jobs from the current "synchronous command +
> broadcast message when done" to some variation of asynchronous command.
> But that should be discussed in a separate thread, and only after we
> know how we plan to ensure monitor availability.

Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

On Wed, Aug 30, 2017 at 11:13:11AM +0100, Daniel P. Berrange wrote:
> On Wed, Aug 30, 2017 at 09:06:20AM +0200, Markus Armbruster wrote:
> > "Daniel P. Berrange" <berrange@redhat.com> writes:
> > 
> > > On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> > 
> > >> However, even with the series, it does not mean that per-monitor
> > >> threads will never hang.  One example is that we can still run "info
> > >> vcpus" in per-monitor threads during a paused postcopy (in that state,
> > >> page faults are never handled, and "info cpus" will never return since
> > >> it tries to sync every vcpus).  So to make sure it does not hang, we
> > >> not only need the per-monitor thread, the user should be careful as
> > >> well on how to use it.
> > >> 
> > >> For postcopy recovery, we may need dedicated monitor channel for
> > >> recovery.  In other words, a destination VM that supports postcopy
> > >> recovery would possibly need:
> > >> 
> > >>   -qmp MAIN_CHANNEL -qmp RECOVERY_CHANNEL
> > 
> > Where RECOVERY_CHANNEL isn't necessarily just for postcopy, but for any
> > "emergency" QMP access.  If you use it only for commands that cannot
> > hang (i.e. terminate in bounded time), then you'll always be able to get
> > commands accepted there in bounded time.
> > 
> > > I think this is a really horrible thing to expose to management applications.
> > > They should not need to be aware of fact that QEMU is buggy and thus requires
> > > that certain commands be run on different monitors to work around the bug.
> > 
> > These are (serious) design limitations, not bugs in the narrow sense of
> > the word.
> > 
> > However, I quite agree that the need for clients to know whether a
> > monitor command can hang is impractical for the general case.  What
> > might be practical is a QMP monitor mode that accepts only known
> > hang-free commands.  Hang-free could be introspectable.
> > 
> > In case you consider that ugly: it's best to explore the design space
> > first, and recoil from "ugly" second.
> 
> Actually you slightly mis-interpreted me there. I think it is ok for
> applications to have knowledge about whether a particular command
> may hang or not. Given that knowledge it should *not*, however, require
> that the application issue such commands on separate monitor channels.
> It is entirely possible to handle hang-free commands on the existing
> channel.
> 
> > > I'd much prefer to see the problem described handled transparently inside
> > > QEMU. One approach is have a dedicated thread in QEMU responsible for all
> > > monitor I/O. This thread should never actually execute monitor commands
> > > though, it would simply parse the command request and put data onto a queue
> > > of pending commands, thus it could never hang. The command queue could be
> > > processed by the main thread, or by another thread that is interested.
> > > eg the migration thread could process any queued commands related to
> > > migration directly.
> > 
> > The monitor itself can't hang then, but the thread(s) dequeuing parsed
> > commands can.
> 
> If certain commands are hang-free then you can have a dedicated thread
> that only de-queues & processes the hang-free commands. The approach I
> outlined is exactly how libvirt deals with its own RPC dispatch. We have
> certain commands that are guaranteed to not hang, which are processed by
> a dedicated pool of threads. So even if all normal RPC commands have
> hung, you can still run a subset of hang-free RPC commands.
> 
> > 
> > To maintain commands' synchronous semantics, their replies need to be
> > sent in order, which of course reintroduces the hangs.
> 
> The requirement for such ordering is just an arbitrary restriction that
> QEMU currently imposes. It is reasonable to allow arbitrary ordering of
> responses (which is what libvirt does in its RPC layer). Admittedly at
> this stage though, we would likely require some "opt in" handshake when
> initializing QMP for the app to tell QEMU it can cope with out of order
> replies. It would require that each command request has a unique serial
> number, which is included in the associated reply, so apps can match
> them up. We used to have that but iirc it was then removed.
> 
> There's other ways to deal with this, such as the job starting idea you
> mention below.
> 
> The key point though is that I don't think creating multiple monitor
> servers is a desirable approach - it is just a hack to avoid dealing
> with the root cause problems. 

Yeah I kindly agree.  It's not the root problem, but AFAIU that's the
simplest way for now to solve the problem.  But I think I understand
the major problem here - an extra channel is an interface change, and
it affects users of monitors.  So I agree we'd better be patient on
choosing a good enough interface, looks like we have two:

- dedicated "hang-able" and "hang-free" channel, or,

- async command handling, then we will have one single dedicated
  command parser (possibly as well in separate thread rather than main
  thread), per-command ID, and possibly slightly more work.  For this
  one, I believe there are different implementations.

So looks like what we need to do is firstly choose an interface, and
if we choose the 2nd, we need to further choose the implementaion.

Before getting to an conclusion, just want to make sure we have got a
consensus on that at least we should start to move the monitor command
handling into a separate thread rather than main thread, am I correct?

Thanks,

> 
> > Let's take a step back from the implementation, and talk about
> > *behavior* instead.
> > 
> > You prefer to have "the problem described handled transparently inside
> > QEMU".  I read that as "QEMU must ensure the QMP monitor is available at
> > all times".  "Available" means it accepts commands in bounded time.
> > Some commands will always finish in bounded time once accepted, others
> > may not, and whether they do may depend on the commands currently in
> > flight.
> > 
> > Commands that can always start and always terminate in bounded time are
> > no problem.
> > 
> > All the other commands have to become "job-starting": the QMP command
> > kicks off a "job", which runs concurrently with the QMP monitor for some
> > (possibly unbounded) time, then finishes.  Jobs can be examined (say to
> > monitor progress, if the job supports that) and controlled (say to
> > cancel, if the job supports that).
> > 
> > A few commands are already job-starting: migrate, the block job family,
> > dump-guest-memory with detach=true.  Whether they're already hang-free I
> > can't say; they could do risky work in their synchronous part.
> > 
> > Many commands that can hang are not job-starting.
> > 
> > Changing a command from "do the job" to "merely start the job" is a
> > compatibility break.
> > 
> > We could make the change opt-in to preserve compatibility.  But is
> > preserving a compatible QMP monitor that is prone to hang wortwhile?
> > 
> > If no, we may choose to use the resulting compatibility break to also
> > switch the packaging of jobs from the current "synchronous command +
> > broadcast message when done" to some variation of asynchronous command.
> > But that should be discussed in a separate thread, and only after we
> > know how we plan to ensure monitor availability.
> 
> Regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

-- 
Peter Xu

On Thu, Aug 31, 2017 at 11:31:55AM +0800, Peter Xu wrote:
> Before getting to an conclusion, just want to make sure we have got a
> consensus on that at least we should start to move the monitor command
> handling into a separate thread rather than main thread, am I correct?

Certainly agree on that, moving dispatch of monitor commands out of
the main thread is critical IMHO. The main thread should only ever
be doing work that is gauranteed non-blockable and completable in a
short, finite amount of time. This means at most it should do I/O only
for the monitor.

Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

* Daniel P. Berrange (berrange@redhat.com) wrote:
> On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> > v2:
> > - fixed "make check" error that patchew reported
> > - moved the thread_join upper in monitor_data_destroy(), before
> >   resources are released
> > - added one new patch (current patch 3) that fixes a nasty risk
> >   condition with IOWatchPoll.  Please see commit message for more
> >   information.
> > - added a g_main_context_wakeup() to make sure the separate loop
> >   thread can be kicked always when we want to destroy the per-monitor
> >   threads.
> > - added one new patch (current patch 8) to introduce migration mgmt
> >   lock for migrate_incoming.
> > 
> > This is an extended work for migration postcopy recovery. This series
> > is tested with the following series to make sure it solves the monitor
> > hang problem that we have encountered for postcopy recovery:
> > 
> >   [RFC 00/29] Migration: postcopy failure recovery
> >   [RFC 0/6] migration: re-use migrate_incoming for postcopy recovery
> > 
> > The root problem is that, monitor commands are all handled in main
> > loop thread now, no matter how many monitors we specify. And, if main
> > loop thread hangs due to some reason, all monitors will be stuck.
> > This can be done in reversed order as well: if any of the monitor
> > hangs, it will hang the main loop, and the rest of the monitors (if
> > there is any).
> > 
> > That affects postcopy recovery, since the recovery requires user input
> > on destination side.  If monitors hang, the destination VM dies and
> > lose hope for even a final recovery.
> > 
> > So, sometimes we need to make sure the monitor be alive, at least one
> > of them.
> > 
> > The whole idea of this series is that instead if handling monitor
> > commands all in main loop thread, we do it separately in per-monitor
> > threads.  Then, even if main loop thread hangs at any point by any
> > reason, per-monitor thread can still survive.  Further, we add hint in
> > QMP/HMP to show whether a command can be executed without QMP, if so,
> > we avoid taking BQL when running that command.  It greatly reduced
> > contention of BQL.  Now the only user of that new parameter (currently
> > I call it "without-bql") is "migrate-incoming" command, which is the
> > only command to rescue a paused postcopy migration.
> > 
> > However, even with the series, it does not mean that per-monitor
> > threads will never hang.  One example is that we can still run "info
> > vcpus" in per-monitor threads during a paused postcopy (in that state,
> > page faults are never handled, and "info cpus" will never return since
> > it tries to sync every vcpus).  So to make sure it does not hang, we
> > not only need the per-monitor thread, the user should be careful as
> > well on how to use it.
> > 
> > For postcopy recovery, we may need dedicated monitor channel for
> > recovery.  In other words, a destination VM that supports postcopy
> > recovery would possibly need:
> > 
> >   -qmp MAIN_CHANNEL -qmp RECOVERY_CHANNEL
> 
> I think this is a really horrible thing to expose to management applications.
> They should not need to be aware of fact that QEMU is buggy and thus requires
> that certain commands be run on different monitors to work around the bug.

It's unfortunately baked in way too deep to fix in the near term; the
BQL is just too cantagious and we have a fundamental design of running
all the main IO emulation in one thread.

> I'd much prefer to see the problem described handled transparently inside
> QEMU. One approach is have a dedicated thread in QEMU responsible for all
> monitor I/O. This thread should never actually execute monitor commands
> though, it would simply parse the command request and put data onto a queue
> of pending commands, thus it could never hang. The command queue could be
> processed by the main thread, or by another thread that is interested.
> eg the migration thread could process any queued commands related to
> migration directly.

That requires a change in the current API to allow async command
completion (OK that is something Marc-Andre's world has) so that
from the one connection you can have multiple outstanding commands.
Hmm unless....

We've also got problems that some commands don't like being run outside
of the main thread (see Fam's reply on the 21st pointing out that a lot
of block commands would assert).

I think the way to move to what you describe would be:
  a) A separate thread for monitor IO
      This seems a separate problem
      How hard is that?  Will all the current IO mechanisms used
      for monitors just work if we run them in a separate thread?
      What about mux?

  b) Initially all commands get dispatched to the main thread
     so nothing changes about the API.

  c) We create a new thread for the lock-free commands, and route
      lock-free commands down it.

  d) We start with a rule that on any one monitor connection we
  don't allow you to start a command until the previous one has
  finished

(d) allows us to avoid any API changes, but allows us to do lock-free
stuff on a separate connection like Peter's world.
We can drop (d) once we have a way of doing async commands.
We can add dispatching to more threads once someone describes
what they want from those threads.

Does that work for you Dan?

(IMHO this is still more complex than Peter's world and I don't
really see the advantage).

Dave


> Regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

On Wed, Sep 06, 2017 at 10:48:46AM +0100, Dr. David Alan Gilbert wrote:
> * Daniel P. Berrange (berrange@redhat.com) wrote:
> > On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> > > v2:
> > > - fixed "make check" error that patchew reported
> > > - moved the thread_join upper in monitor_data_destroy(), before
> > >   resources are released
> > > - added one new patch (current patch 3) that fixes a nasty risk
> > >   condition with IOWatchPoll.  Please see commit message for more
> > >   information.
> > > - added a g_main_context_wakeup() to make sure the separate loop
> > >   thread can be kicked always when we want to destroy the per-monitor
> > >   threads.
> > > - added one new patch (current patch 8) to introduce migration mgmt
> > >   lock for migrate_incoming.
> > > 
> > > This is an extended work for migration postcopy recovery. This series
> > > is tested with the following series to make sure it solves the monitor
> > > hang problem that we have encountered for postcopy recovery:
> > > 
> > >   [RFC 00/29] Migration: postcopy failure recovery
> > >   [RFC 0/6] migration: re-use migrate_incoming for postcopy recovery
> > > 
> > > The root problem is that, monitor commands are all handled in main
> > > loop thread now, no matter how many monitors we specify. And, if main
> > > loop thread hangs due to some reason, all monitors will be stuck.
> > > This can be done in reversed order as well: if any of the monitor
> > > hangs, it will hang the main loop, and the rest of the monitors (if
> > > there is any).
> > > 
> > > That affects postcopy recovery, since the recovery requires user input
> > > on destination side.  If monitors hang, the destination VM dies and
> > > lose hope for even a final recovery.
> > > 
> > > So, sometimes we need to make sure the monitor be alive, at least one
> > > of them.
> > > 
> > > The whole idea of this series is that instead if handling monitor
> > > commands all in main loop thread, we do it separately in per-monitor
> > > threads.  Then, even if main loop thread hangs at any point by any
> > > reason, per-monitor thread can still survive.  Further, we add hint in
> > > QMP/HMP to show whether a command can be executed without QMP, if so,
> > > we avoid taking BQL when running that command.  It greatly reduced
> > > contention of BQL.  Now the only user of that new parameter (currently
> > > I call it "without-bql") is "migrate-incoming" command, which is the
> > > only command to rescue a paused postcopy migration.
> > > 
> > > However, even with the series, it does not mean that per-monitor
> > > threads will never hang.  One example is that we can still run "info
> > > vcpus" in per-monitor threads during a paused postcopy (in that state,
> > > page faults are never handled, and "info cpus" will never return since
> > > it tries to sync every vcpus).  So to make sure it does not hang, we
> > > not only need the per-monitor thread, the user should be careful as
> > > well on how to use it.
> > > 
> > > For postcopy recovery, we may need dedicated monitor channel for
> > > recovery.  In other words, a destination VM that supports postcopy
> > > recovery would possibly need:
> > > 
> > >   -qmp MAIN_CHANNEL -qmp RECOVERY_CHANNEL
> > 
> > I think this is a really horrible thing to expose to management applications.
> > They should not need to be aware of fact that QEMU is buggy and thus requires
> > that certain commands be run on different monitors to work around the bug.
> 
> It's unfortunately baked in way too deep to fix in the near term; the
> BQL is just too cantagious and we have a fundamental design of running
> all the main IO emulation in one thread.
> 
> > I'd much prefer to see the problem described handled transparently inside
> > QEMU. One approach is have a dedicated thread in QEMU responsible for all
> > monitor I/O. This thread should never actually execute monitor commands
> > though, it would simply parse the command request and put data onto a queue
> > of pending commands, thus it could never hang. The command queue could be
> > processed by the main thread, or by another thread that is interested.
> > eg the migration thread could process any queued commands related to
> > migration directly.
> 
> That requires a change in the current API to allow async command
> completion (OK that is something Marc-Andre's world has) so that
> from the one connection you can have multiple outstanding commands.
> Hmm unless....
> 
> We've also got problems that some commands don't like being run outside
> of the main thread (see Fam's reply on the 21st pointing out that a lot
> of block commands would assert).
> 
> I think the way to move to what you describe would be:
>   a) A separate thread for monitor IO
>       This seems a separate problem
>       How hard is that?  Will all the current IO mechanisms used
>       for monitors just work if we run them in a separate thread?
>       What about mux?
> 
>   b) Initially all commands get dispatched to the main thread
>      so nothing changes about the API.
> 
>   c) We create a new thread for the lock-free commands, and route
>       lock-free commands down it.
> 
>   d) We start with a rule that on any one monitor connection we
>   don't allow you to start a command until the previous one has
>   finished
> 
> (d) allows us to avoid any API changes, but allows us to do lock-free
> stuff on a separate connection like Peter's world.
> We can drop (d) once we have a way of doing async commands.
> We can add dispatching to more threads once someone describes
> what they want from those threads.
> 
> Does that work for you Dan?

It would *provided* that we do (c) for the commands Peter wants for
this migration series.  IOW, I don't want to have to have logic in
libvirt that either needs to add a 2nd monitor server, or open a 2nd
monitor connection, to deal with migration post-copy recovery in some
versions of QEMU.  So whatever is needed to make post-copy recovery
work has to be done for (c).

Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

* Daniel P. Berrange (berrange@redhat.com) wrote:
> On Wed, Sep 06, 2017 at 10:48:46AM +0100, Dr. David Alan Gilbert wrote:
> > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> > > > v2:
> > > > - fixed "make check" error that patchew reported
> > > > - moved the thread_join upper in monitor_data_destroy(), before
> > > >   resources are released
> > > > - added one new patch (current patch 3) that fixes a nasty risk
> > > >   condition with IOWatchPoll.  Please see commit message for more
> > > >   information.
> > > > - added a g_main_context_wakeup() to make sure the separate loop
> > > >   thread can be kicked always when we want to destroy the per-monitor
> > > >   threads.
> > > > - added one new patch (current patch 8) to introduce migration mgmt
> > > >   lock for migrate_incoming.
> > > > 
> > > > This is an extended work for migration postcopy recovery. This series
> > > > is tested with the following series to make sure it solves the monitor
> > > > hang problem that we have encountered for postcopy recovery:
> > > > 
> > > >   [RFC 00/29] Migration: postcopy failure recovery
> > > >   [RFC 0/6] migration: re-use migrate_incoming for postcopy recovery
> > > > 
> > > > The root problem is that, monitor commands are all handled in main
> > > > loop thread now, no matter how many monitors we specify. And, if main
> > > > loop thread hangs due to some reason, all monitors will be stuck.
> > > > This can be done in reversed order as well: if any of the monitor
> > > > hangs, it will hang the main loop, and the rest of the monitors (if
> > > > there is any).
> > > > 
> > > > That affects postcopy recovery, since the recovery requires user input
> > > > on destination side.  If monitors hang, the destination VM dies and
> > > > lose hope for even a final recovery.
> > > > 
> > > > So, sometimes we need to make sure the monitor be alive, at least one
> > > > of them.
> > > > 
> > > > The whole idea of this series is that instead if handling monitor
> > > > commands all in main loop thread, we do it separately in per-monitor
> > > > threads.  Then, even if main loop thread hangs at any point by any
> > > > reason, per-monitor thread can still survive.  Further, we add hint in
> > > > QMP/HMP to show whether a command can be executed without QMP, if so,
> > > > we avoid taking BQL when running that command.  It greatly reduced
> > > > contention of BQL.  Now the only user of that new parameter (currently
> > > > I call it "without-bql") is "migrate-incoming" command, which is the
> > > > only command to rescue a paused postcopy migration.
> > > > 
> > > > However, even with the series, it does not mean that per-monitor
> > > > threads will never hang.  One example is that we can still run "info
> > > > vcpus" in per-monitor threads during a paused postcopy (in that state,
> > > > page faults are never handled, and "info cpus" will never return since
> > > > it tries to sync every vcpus).  So to make sure it does not hang, we
> > > > not only need the per-monitor thread, the user should be careful as
> > > > well on how to use it.
> > > > 
> > > > For postcopy recovery, we may need dedicated monitor channel for
> > > > recovery.  In other words, a destination VM that supports postcopy
> > > > recovery would possibly need:
> > > > 
> > > >   -qmp MAIN_CHANNEL -qmp RECOVERY_CHANNEL
> > > 
> > > I think this is a really horrible thing to expose to management applications.
> > > They should not need to be aware of fact that QEMU is buggy and thus requires
> > > that certain commands be run on different monitors to work around the bug.
> > 
> > It's unfortunately baked in way too deep to fix in the near term; the
> > BQL is just too cantagious and we have a fundamental design of running
> > all the main IO emulation in one thread.
> > 
> > > I'd much prefer to see the problem described handled transparently inside
> > > QEMU. One approach is have a dedicated thread in QEMU responsible for all
> > > monitor I/O. This thread should never actually execute monitor commands
> > > though, it would simply parse the command request and put data onto a queue
> > > of pending commands, thus it could never hang. The command queue could be
> > > processed by the main thread, or by another thread that is interested.
> > > eg the migration thread could process any queued commands related to
> > > migration directly.
> > 
> > That requires a change in the current API to allow async command
> > completion (OK that is something Marc-Andre's world has) so that
> > from the one connection you can have multiple outstanding commands.
> > Hmm unless....
> > 
> > We've also got problems that some commands don't like being run outside
> > of the main thread (see Fam's reply on the 21st pointing out that a lot
> > of block commands would assert).
> > 
> > I think the way to move to what you describe would be:
> >   a) A separate thread for monitor IO
> >       This seems a separate problem
> >       How hard is that?  Will all the current IO mechanisms used
> >       for monitors just work if we run them in a separate thread?
> >       What about mux?
> > 
> >   b) Initially all commands get dispatched to the main thread
> >      so nothing changes about the API.
> > 
> >   c) We create a new thread for the lock-free commands, and route
> >       lock-free commands down it.
> > 
> >   d) We start with a rule that on any one monitor connection we
> >   don't allow you to start a command until the previous one has
> >   finished
> > 
> > (d) allows us to avoid any API changes, but allows us to do lock-free
> > stuff on a separate connection like Peter's world.
> > We can drop (d) once we have a way of doing async commands.
> > We can add dispatching to more threads once someone describes
> > what they want from those threads.
> > 
> > Does that work for you Dan?
> 
> It would *provided* that we do (c) for the commands Peter wants for
> this migration series.  IOW, I don't want to have to have logic in
> libvirt that either needs to add a 2nd monitor server, or open a 2nd
> monitor connection, to deal with migration post-copy recovery in some
> versions of QEMU.  So whatever is needed to make post-copy recovery
> work has to be done for (c).

But then doesn't that mean you're requiring us to break (d) and change
the QMP interface to libvirt so it can do async stuff?

Dave

> Regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

On Wed, Sep 06, 2017 at 11:48:51AM +0100, Dr. David Alan Gilbert wrote:
> * Daniel P. Berrange (berrange@redhat.com) wrote:
> > On Wed, Sep 06, 2017 at 10:48:46AM +0100, Dr. David Alan Gilbert wrote:
> > > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > > On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> > > > > v2:
> > > > > - fixed "make check" error that patchew reported
> > > > > - moved the thread_join upper in monitor_data_destroy(), before
> > > > >   resources are released
> > > > > - added one new patch (current patch 3) that fixes a nasty risk
> > > > >   condition with IOWatchPoll.  Please see commit message for more
> > > > >   information.
> > > > > - added a g_main_context_wakeup() to make sure the separate loop
> > > > >   thread can be kicked always when we want to destroy the per-monitor
> > > > >   threads.
> > > > > - added one new patch (current patch 8) to introduce migration mgmt
> > > > >   lock for migrate_incoming.
> > > > > 
> > > > > This is an extended work for migration postcopy recovery. This series
> > > > > is tested with the following series to make sure it solves the monitor
> > > > > hang problem that we have encountered for postcopy recovery:
> > > > > 
> > > > >   [RFC 00/29] Migration: postcopy failure recovery
> > > > >   [RFC 0/6] migration: re-use migrate_incoming for postcopy recovery
> > > > > 
> > > > > The root problem is that, monitor commands are all handled in main
> > > > > loop thread now, no matter how many monitors we specify. And, if main
> > > > > loop thread hangs due to some reason, all monitors will be stuck.
> > > > > This can be done in reversed order as well: if any of the monitor
> > > > > hangs, it will hang the main loop, and the rest of the monitors (if
> > > > > there is any).
> > > > > 
> > > > > That affects postcopy recovery, since the recovery requires user input
> > > > > on destination side.  If monitors hang, the destination VM dies and
> > > > > lose hope for even a final recovery.
> > > > > 
> > > > > So, sometimes we need to make sure the monitor be alive, at least one
> > > > > of them.
> > > > > 
> > > > > The whole idea of this series is that instead if handling monitor
> > > > > commands all in main loop thread, we do it separately in per-monitor
> > > > > threads.  Then, even if main loop thread hangs at any point by any
> > > > > reason, per-monitor thread can still survive.  Further, we add hint in
> > > > > QMP/HMP to show whether a command can be executed without QMP, if so,
> > > > > we avoid taking BQL when running that command.  It greatly reduced
> > > > > contention of BQL.  Now the only user of that new parameter (currently
> > > > > I call it "without-bql") is "migrate-incoming" command, which is the
> > > > > only command to rescue a paused postcopy migration.
> > > > > 
> > > > > However, even with the series, it does not mean that per-monitor
> > > > > threads will never hang.  One example is that we can still run "info
> > > > > vcpus" in per-monitor threads during a paused postcopy (in that state,
> > > > > page faults are never handled, and "info cpus" will never return since
> > > > > it tries to sync every vcpus).  So to make sure it does not hang, we
> > > > > not only need the per-monitor thread, the user should be careful as
> > > > > well on how to use it.
> > > > > 
> > > > > For postcopy recovery, we may need dedicated monitor channel for
> > > > > recovery.  In other words, a destination VM that supports postcopy
> > > > > recovery would possibly need:
> > > > > 
> > > > >   -qmp MAIN_CHANNEL -qmp RECOVERY_CHANNEL
> > > > 
> > > > I think this is a really horrible thing to expose to management applications.
> > > > They should not need to be aware of fact that QEMU is buggy and thus requires
> > > > that certain commands be run on different monitors to work around the bug.
> > > 
> > > It's unfortunately baked in way too deep to fix in the near term; the
> > > BQL is just too cantagious and we have a fundamental design of running
> > > all the main IO emulation in one thread.
> > > 
> > > > I'd much prefer to see the problem described handled transparently inside
> > > > QEMU. One approach is have a dedicated thread in QEMU responsible for all
> > > > monitor I/O. This thread should never actually execute monitor commands
> > > > though, it would simply parse the command request and put data onto a queue
> > > > of pending commands, thus it could never hang. The command queue could be
> > > > processed by the main thread, or by another thread that is interested.
> > > > eg the migration thread could process any queued commands related to
> > > > migration directly.
> > > 
> > > That requires a change in the current API to allow async command
> > > completion (OK that is something Marc-Andre's world has) so that
> > > from the one connection you can have multiple outstanding commands.
> > > Hmm unless....
> > > 
> > > We've also got problems that some commands don't like being run outside
> > > of the main thread (see Fam's reply on the 21st pointing out that a lot
> > > of block commands would assert).
> > > 
> > > I think the way to move to what you describe would be:
> > >   a) A separate thread for monitor IO
> > >       This seems a separate problem
> > >       How hard is that?  Will all the current IO mechanisms used
> > >       for monitors just work if we run them in a separate thread?
> > >       What about mux?
> > > 
> > >   b) Initially all commands get dispatched to the main thread
> > >      so nothing changes about the API.
> > > 
> > >   c) We create a new thread for the lock-free commands, and route
> > >       lock-free commands down it.
> > > 
> > >   d) We start with a rule that on any one monitor connection we
> > >   don't allow you to start a command until the previous one has
> > >   finished
> > > 
> > > (d) allows us to avoid any API changes, but allows us to do lock-free
> > > stuff on a separate connection like Peter's world.
> > > We can drop (d) once we have a way of doing async commands.
> > > We can add dispatching to more threads once someone describes
> > > what they want from those threads.
> > > 
> > > Does that work for you Dan?
> > 
> > It would *provided* that we do (c) for the commands Peter wants for
> > this migration series.  IOW, I don't want to have to have logic in
> > libvirt that either needs to add a 2nd monitor server, or open a 2nd
> > monitor connection, to deal with migration post-copy recovery in some
> > versions of QEMU.  So whatever is needed to make post-copy recovery
> > work has to be done for (c).
> 
> But then doesn't that mean you're requiring us to break (d) and change
> the QMP interface to libvirt so it can do async stuff?

Depends on your definition of break - I'm assuming there's either a way
to opt-in to use of a async mode for existing commands in (c), or that
async commands would be added in parallel with existing sync commands.
IOW, its not a API breakage - its an opt-in extension of existing
functionality.

Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

* Daniel P. Berrange (berrange@redhat.com) wrote:
> On Wed, Sep 06, 2017 at 11:48:51AM +0100, Dr. David Alan Gilbert wrote:
> > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > On Wed, Sep 06, 2017 at 10:48:46AM +0100, Dr. David Alan Gilbert wrote:
> > > > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > > > On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> > > > > > v2:
> > > > > > - fixed "make check" error that patchew reported
> > > > > > - moved the thread_join upper in monitor_data_destroy(), before
> > > > > >   resources are released
> > > > > > - added one new patch (current patch 3) that fixes a nasty risk
> > > > > >   condition with IOWatchPoll.  Please see commit message for more
> > > > > >   information.
> > > > > > - added a g_main_context_wakeup() to make sure the separate loop
> > > > > >   thread can be kicked always when we want to destroy the per-monitor
> > > > > >   threads.
> > > > > > - added one new patch (current patch 8) to introduce migration mgmt
> > > > > >   lock for migrate_incoming.
> > > > > > 
> > > > > > This is an extended work for migration postcopy recovery. This series
> > > > > > is tested with the following series to make sure it solves the monitor
> > > > > > hang problem that we have encountered for postcopy recovery:
> > > > > > 
> > > > > >   [RFC 00/29] Migration: postcopy failure recovery
> > > > > >   [RFC 0/6] migration: re-use migrate_incoming for postcopy recovery
> > > > > > 
> > > > > > The root problem is that, monitor commands are all handled in main
> > > > > > loop thread now, no matter how many monitors we specify. And, if main
> > > > > > loop thread hangs due to some reason, all monitors will be stuck.
> > > > > > This can be done in reversed order as well: if any of the monitor
> > > > > > hangs, it will hang the main loop, and the rest of the monitors (if
> > > > > > there is any).
> > > > > > 
> > > > > > That affects postcopy recovery, since the recovery requires user input
> > > > > > on destination side.  If monitors hang, the destination VM dies and
> > > > > > lose hope for even a final recovery.
> > > > > > 
> > > > > > So, sometimes we need to make sure the monitor be alive, at least one
> > > > > > of them.
> > > > > > 
> > > > > > The whole idea of this series is that instead if handling monitor
> > > > > > commands all in main loop thread, we do it separately in per-monitor
> > > > > > threads.  Then, even if main loop thread hangs at any point by any
> > > > > > reason, per-monitor thread can still survive.  Further, we add hint in
> > > > > > QMP/HMP to show whether a command can be executed without QMP, if so,
> > > > > > we avoid taking BQL when running that command.  It greatly reduced
> > > > > > contention of BQL.  Now the only user of that new parameter (currently
> > > > > > I call it "without-bql") is "migrate-incoming" command, which is the
> > > > > > only command to rescue a paused postcopy migration.
> > > > > > 
> > > > > > However, even with the series, it does not mean that per-monitor
> > > > > > threads will never hang.  One example is that we can still run "info
> > > > > > vcpus" in per-monitor threads during a paused postcopy (in that state,
> > > > > > page faults are never handled, and "info cpus" will never return since
> > > > > > it tries to sync every vcpus).  So to make sure it does not hang, we
> > > > > > not only need the per-monitor thread, the user should be careful as
> > > > > > well on how to use it.
> > > > > > 
> > > > > > For postcopy recovery, we may need dedicated monitor channel for
> > > > > > recovery.  In other words, a destination VM that supports postcopy
> > > > > > recovery would possibly need:
> > > > > > 
> > > > > >   -qmp MAIN_CHANNEL -qmp RECOVERY_CHANNEL
> > > > > 
> > > > > I think this is a really horrible thing to expose to management applications.
> > > > > They should not need to be aware of fact that QEMU is buggy and thus requires
> > > > > that certain commands be run on different monitors to work around the bug.
> > > > 
> > > > It's unfortunately baked in way too deep to fix in the near term; the
> > > > BQL is just too cantagious and we have a fundamental design of running
> > > > all the main IO emulation in one thread.
> > > > 
> > > > > I'd much prefer to see the problem described handled transparently inside
> > > > > QEMU. One approach is have a dedicated thread in QEMU responsible for all
> > > > > monitor I/O. This thread should never actually execute monitor commands
> > > > > though, it would simply parse the command request and put data onto a queue
> > > > > of pending commands, thus it could never hang. The command queue could be
> > > > > processed by the main thread, or by another thread that is interested.
> > > > > eg the migration thread could process any queued commands related to
> > > > > migration directly.
> > > > 
> > > > That requires a change in the current API to allow async command
> > > > completion (OK that is something Marc-Andre's world has) so that
> > > > from the one connection you can have multiple outstanding commands.
> > > > Hmm unless....
> > > > 
> > > > We've also got problems that some commands don't like being run outside
> > > > of the main thread (see Fam's reply on the 21st pointing out that a lot
> > > > of block commands would assert).
> > > > 
> > > > I think the way to move to what you describe would be:
> > > >   a) A separate thread for monitor IO
> > > >       This seems a separate problem
> > > >       How hard is that?  Will all the current IO mechanisms used
> > > >       for monitors just work if we run them in a separate thread?
> > > >       What about mux?
> > > > 
> > > >   b) Initially all commands get dispatched to the main thread
> > > >      so nothing changes about the API.
> > > > 
> > > >   c) We create a new thread for the lock-free commands, and route
> > > >       lock-free commands down it.
> > > > 
> > > >   d) We start with a rule that on any one monitor connection we
> > > >   don't allow you to start a command until the previous one has
> > > >   finished
> > > > 
> > > > (d) allows us to avoid any API changes, but allows us to do lock-free
> > > > stuff on a separate connection like Peter's world.
> > > > We can drop (d) once we have a way of doing async commands.
> > > > We can add dispatching to more threads once someone describes
> > > > what they want from those threads.
> > > > 
> > > > Does that work for you Dan?
> > > 
> > > It would *provided* that we do (c) for the commands Peter wants for
> > > this migration series.  IOW, I don't want to have to have logic in
> > > libvirt that either needs to add a 2nd monitor server, or open a 2nd
> > > monitor connection, to deal with migration post-copy recovery in some
> > > versions of QEMU.  So whatever is needed to make post-copy recovery
> > > work has to be done for (c).
> > 
> > But then doesn't that mean you're requiring us to break (d) and change
> > the QMP interface to libvirt so it can do async stuff?
> 
> Depends on your definition of break - I'm assuming there's either a way
> to opt-in to use of a async mode for existing commands in (c), or that
> async commands would be added in parallel with existing sync commands.
> IOW, its not a API breakage - its an opt-in extension of existing
> functionality.

But you'd need to do async commands for all commands you issued to avoid
blocking the io thread so that you could then issue the recovery
commands.

Dave

> Regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

On Wed, Sep 06, 2017 at 11:57:05AM +0100, Dr. David Alan Gilbert wrote:
> * Daniel P. Berrange (berrange@redhat.com) wrote:
> > On Wed, Sep 06, 2017 at 11:48:51AM +0100, Dr. David Alan Gilbert wrote:
> > > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > > On Wed, Sep 06, 2017 at 10:48:46AM +0100, Dr. David Alan Gilbert wrote:
> > > > > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > > > > On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> > > > > > > v2:
> > > > > > > - fixed "make check" error that patchew reported
> > > > > > > - moved the thread_join upper in monitor_data_destroy(), before
> > > > > > >   resources are released
> > > > > > > - added one new patch (current patch 3) that fixes a nasty risk
> > > > > > >   condition with IOWatchPoll.  Please see commit message for more
> > > > > > >   information.
> > > > > > > - added a g_main_context_wakeup() to make sure the separate loop
> > > > > > >   thread can be kicked always when we want to destroy the per-monitor
> > > > > > >   threads.
> > > > > > > - added one new patch (current patch 8) to introduce migration mgmt
> > > > > > >   lock for migrate_incoming.
> > > > > > > 
> > > > > > > This is an extended work for migration postcopy recovery. This series
> > > > > > > is tested with the following series to make sure it solves the monitor
> > > > > > > hang problem that we have encountered for postcopy recovery:
> > > > > > > 
> > > > > > >   [RFC 00/29] Migration: postcopy failure recovery
> > > > > > >   [RFC 0/6] migration: re-use migrate_incoming for postcopy recovery
> > > > > > > 
> > > > > > > The root problem is that, monitor commands are all handled in main
> > > > > > > loop thread now, no matter how many monitors we specify. And, if main
> > > > > > > loop thread hangs due to some reason, all monitors will be stuck.
> > > > > > > This can be done in reversed order as well: if any of the monitor
> > > > > > > hangs, it will hang the main loop, and the rest of the monitors (if
> > > > > > > there is any).
> > > > > > > 
> > > > > > > That affects postcopy recovery, since the recovery requires user input
> > > > > > > on destination side.  If monitors hang, the destination VM dies and
> > > > > > > lose hope for even a final recovery.
> > > > > > > 
> > > > > > > So, sometimes we need to make sure the monitor be alive, at least one
> > > > > > > of them.
> > > > > > > 
> > > > > > > The whole idea of this series is that instead if handling monitor
> > > > > > > commands all in main loop thread, we do it separately in per-monitor
> > > > > > > threads.  Then, even if main loop thread hangs at any point by any
> > > > > > > reason, per-monitor thread can still survive.  Further, we add hint in
> > > > > > > QMP/HMP to show whether a command can be executed without QMP, if so,
> > > > > > > we avoid taking BQL when running that command.  It greatly reduced
> > > > > > > contention of BQL.  Now the only user of that new parameter (currently
> > > > > > > I call it "without-bql") is "migrate-incoming" command, which is the
> > > > > > > only command to rescue a paused postcopy migration.
> > > > > > > 
> > > > > > > However, even with the series, it does not mean that per-monitor
> > > > > > > threads will never hang.  One example is that we can still run "info
> > > > > > > vcpus" in per-monitor threads during a paused postcopy (in that state,
> > > > > > > page faults are never handled, and "info cpus" will never return since
> > > > > > > it tries to sync every vcpus).  So to make sure it does not hang, we
> > > > > > > not only need the per-monitor thread, the user should be careful as
> > > > > > > well on how to use it.
> > > > > > > 
> > > > > > > For postcopy recovery, we may need dedicated monitor channel for
> > > > > > > recovery.  In other words, a destination VM that supports postcopy
> > > > > > > recovery would possibly need:
> > > > > > > 
> > > > > > >   -qmp MAIN_CHANNEL -qmp RECOVERY_CHANNEL
> > > > > > 
> > > > > > I think this is a really horrible thing to expose to management applications.
> > > > > > They should not need to be aware of fact that QEMU is buggy and thus requires
> > > > > > that certain commands be run on different monitors to work around the bug.
> > > > > 
> > > > > It's unfortunately baked in way too deep to fix in the near term; the
> > > > > BQL is just too cantagious and we have a fundamental design of running
> > > > > all the main IO emulation in one thread.
> > > > > 
> > > > > > I'd much prefer to see the problem described handled transparently inside
> > > > > > QEMU. One approach is have a dedicated thread in QEMU responsible for all
> > > > > > monitor I/O. This thread should never actually execute monitor commands
> > > > > > though, it would simply parse the command request and put data onto a queue
> > > > > > of pending commands, thus it could never hang. The command queue could be
> > > > > > processed by the main thread, or by another thread that is interested.
> > > > > > eg the migration thread could process any queued commands related to
> > > > > > migration directly.
> > > > > 
> > > > > That requires a change in the current API to allow async command
> > > > > completion (OK that is something Marc-Andre's world has) so that
> > > > > from the one connection you can have multiple outstanding commands.
> > > > > Hmm unless....
> > > > > 
> > > > > We've also got problems that some commands don't like being run outside
> > > > > of the main thread (see Fam's reply on the 21st pointing out that a lot
> > > > > of block commands would assert).
> > > > > 
> > > > > I think the way to move to what you describe would be:
> > > > >   a) A separate thread for monitor IO
> > > > >       This seems a separate problem
> > > > >       How hard is that?  Will all the current IO mechanisms used
> > > > >       for monitors just work if we run them in a separate thread?
> > > > >       What about mux?
> > > > > 
> > > > >   b) Initially all commands get dispatched to the main thread
> > > > >      so nothing changes about the API.
> > > > > 
> > > > >   c) We create a new thread for the lock-free commands, and route
> > > > >       lock-free commands down it.
> > > > > 
> > > > >   d) We start with a rule that on any one monitor connection we
> > > > >   don't allow you to start a command until the previous one has
> > > > >   finished
> > > > > 
> > > > > (d) allows us to avoid any API changes, but allows us to do lock-free
> > > > > stuff on a separate connection like Peter's world.
> > > > > We can drop (d) once we have a way of doing async commands.
> > > > > We can add dispatching to more threads once someone describes
> > > > > what they want from those threads.
> > > > > 
> > > > > Does that work for you Dan?
> > > > 
> > > > It would *provided* that we do (c) for the commands Peter wants for
> > > > this migration series.  IOW, I don't want to have to have logic in
> > > > libvirt that either needs to add a 2nd monitor server, or open a 2nd
> > > > monitor connection, to deal with migration post-copy recovery in some
> > > > versions of QEMU.  So whatever is needed to make post-copy recovery
> > > > work has to be done for (c).
> > > 
> > > But then doesn't that mean you're requiring us to break (d) and change
> > > the QMP interface to libvirt so it can do async stuff?
> > 
> > Depends on your definition of break - I'm assuming there's either a way
> > to opt-in to use of a async mode for existing commands in (c), or that
> > async commands would be added in parallel with existing sync commands.
> > IOW, its not a API breakage - its an opt-in extension of existing
> > functionality.
> 
> But you'd need to do async commands for all commands you issued to avoid
> blocking the io thread so that you could then issue the recovery
> commands.

I don't see why that has to be the case. In order to issue an async command
all that needs to be the case is that command replies should be allowed to
be sent out of order.

IOW if command A is blocking and command B is async, then we shoudl be
allowed to have the following

   req A
   req B
   res A
   res B

Or

   req A
   req B
   res B
   res A

Or

   req B
   req A
   res B
   res A

etc.

This does imply that you need a separate monitor I/O processing, from the
command execution thread, but I see no need for all commands to suddenly
become async. Just allowing interleaved replies is sufficient from the
POV of the protocol definition. This interleaving is easy to handle from
the client POV - just requires a unique 'serial' in the request by the
client, that is copied into the reply by QEMU.

Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

* Daniel P. Berrange (berrange@redhat.com) wrote:
> On Wed, Sep 06, 2017 at 11:57:05AM +0100, Dr. David Alan Gilbert wrote:
> > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > On Wed, Sep 06, 2017 at 11:48:51AM +0100, Dr. David Alan Gilbert wrote:
> > > > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > > > On Wed, Sep 06, 2017 at 10:48:46AM +0100, Dr. David Alan Gilbert wrote:
> > > > > > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > > > > > On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> > > > > > > > v2:
> > > > > > > > - fixed "make check" error that patchew reported
> > > > > > > > - moved the thread_join upper in monitor_data_destroy(), before
> > > > > > > >   resources are released
> > > > > > > > - added one new patch (current patch 3) that fixes a nasty risk
> > > > > > > >   condition with IOWatchPoll.  Please see commit message for more
> > > > > > > >   information.
> > > > > > > > - added a g_main_context_wakeup() to make sure the separate loop
> > > > > > > >   thread can be kicked always when we want to destroy the per-monitor
> > > > > > > >   threads.
> > > > > > > > - added one new patch (current patch 8) to introduce migration mgmt
> > > > > > > >   lock for migrate_incoming.
> > > > > > > > 
> > > > > > > > This is an extended work for migration postcopy recovery. This series
> > > > > > > > is tested with the following series to make sure it solves the monitor
> > > > > > > > hang problem that we have encountered for postcopy recovery:
> > > > > > > > 
> > > > > > > >   [RFC 00/29] Migration: postcopy failure recovery
> > > > > > > >   [RFC 0/6] migration: re-use migrate_incoming for postcopy recovery
> > > > > > > > 
> > > > > > > > The root problem is that, monitor commands are all handled in main
> > > > > > > > loop thread now, no matter how many monitors we specify. And, if main
> > > > > > > > loop thread hangs due to some reason, all monitors will be stuck.
> > > > > > > > This can be done in reversed order as well: if any of the monitor
> > > > > > > > hangs, it will hang the main loop, and the rest of the monitors (if
> > > > > > > > there is any).
> > > > > > > > 
> > > > > > > > That affects postcopy recovery, since the recovery requires user input
> > > > > > > > on destination side.  If monitors hang, the destination VM dies and
> > > > > > > > lose hope for even a final recovery.
> > > > > > > > 
> > > > > > > > So, sometimes we need to make sure the monitor be alive, at least one
> > > > > > > > of them.
> > > > > > > > 
> > > > > > > > The whole idea of this series is that instead if handling monitor
> > > > > > > > commands all in main loop thread, we do it separately in per-monitor
> > > > > > > > threads.  Then, even if main loop thread hangs at any point by any
> > > > > > > > reason, per-monitor thread can still survive.  Further, we add hint in
> > > > > > > > QMP/HMP to show whether a command can be executed without QMP, if so,
> > > > > > > > we avoid taking BQL when running that command.  It greatly reduced
> > > > > > > > contention of BQL.  Now the only user of that new parameter (currently
> > > > > > > > I call it "without-bql") is "migrate-incoming" command, which is the
> > > > > > > > only command to rescue a paused postcopy migration.
> > > > > > > > 
> > > > > > > > However, even with the series, it does not mean that per-monitor
> > > > > > > > threads will never hang.  One example is that we can still run "info
> > > > > > > > vcpus" in per-monitor threads during a paused postcopy (in that state,
> > > > > > > > page faults are never handled, and "info cpus" will never return since
> > > > > > > > it tries to sync every vcpus).  So to make sure it does not hang, we
> > > > > > > > not only need the per-monitor thread, the user should be careful as
> > > > > > > > well on how to use it.
> > > > > > > > 
> > > > > > > > For postcopy recovery, we may need dedicated monitor channel for
> > > > > > > > recovery.  In other words, a destination VM that supports postcopy
> > > > > > > > recovery would possibly need:
> > > > > > > > 
> > > > > > > >   -qmp MAIN_CHANNEL -qmp RECOVERY_CHANNEL
> > > > > > > 
> > > > > > > I think this is a really horrible thing to expose to management applications.
> > > > > > > They should not need to be aware of fact that QEMU is buggy and thus requires
> > > > > > > that certain commands be run on different monitors to work around the bug.
> > > > > > 
> > > > > > It's unfortunately baked in way too deep to fix in the near term; the
> > > > > > BQL is just too cantagious and we have a fundamental design of running
> > > > > > all the main IO emulation in one thread.
> > > > > > 
> > > > > > > I'd much prefer to see the problem described handled transparently inside
> > > > > > > QEMU. One approach is have a dedicated thread in QEMU responsible for all
> > > > > > > monitor I/O. This thread should never actually execute monitor commands
> > > > > > > though, it would simply parse the command request and put data onto a queue
> > > > > > > of pending commands, thus it could never hang. The command queue could be
> > > > > > > processed by the main thread, or by another thread that is interested.
> > > > > > > eg the migration thread could process any queued commands related to
> > > > > > > migration directly.
> > > > > > 
> > > > > > That requires a change in the current API to allow async command
> > > > > > completion (OK that is something Marc-Andre's world has) so that
> > > > > > from the one connection you can have multiple outstanding commands.
> > > > > > Hmm unless....
> > > > > > 
> > > > > > We've also got problems that some commands don't like being run outside
> > > > > > of the main thread (see Fam's reply on the 21st pointing out that a lot
> > > > > > of block commands would assert).
> > > > > > 
> > > > > > I think the way to move to what you describe would be:
> > > > > >   a) A separate thread for monitor IO
> > > > > >       This seems a separate problem
> > > > > >       How hard is that?  Will all the current IO mechanisms used
> > > > > >       for monitors just work if we run them in a separate thread?
> > > > > >       What about mux?
> > > > > > 
> > > > > >   b) Initially all commands get dispatched to the main thread
> > > > > >      so nothing changes about the API.
> > > > > > 
> > > > > >   c) We create a new thread for the lock-free commands, and route
> > > > > >       lock-free commands down it.
> > > > > > 
> > > > > >   d) We start with a rule that on any one monitor connection we
> > > > > >   don't allow you to start a command until the previous one has
> > > > > >   finished
> > > > > > 
> > > > > > (d) allows us to avoid any API changes, but allows us to do lock-free
> > > > > > stuff on a separate connection like Peter's world.
> > > > > > We can drop (d) once we have a way of doing async commands.
> > > > > > We can add dispatching to more threads once someone describes
> > > > > > what they want from those threads.
> > > > > > 
> > > > > > Does that work for you Dan?
> > > > > 
> > > > > It would *provided* that we do (c) for the commands Peter wants for
> > > > > this migration series.  IOW, I don't want to have to have logic in
> > > > > libvirt that either needs to add a 2nd monitor server, or open a 2nd
> > > > > monitor connection, to deal with migration post-copy recovery in some
> > > > > versions of QEMU.  So whatever is needed to make post-copy recovery
> > > > > work has to be done for (c).
> > > > 
> > > > But then doesn't that mean you're requiring us to break (d) and change
> > > > the QMP interface to libvirt so it can do async stuff?
> > > 
> > > Depends on your definition of break - I'm assuming there's either a way
> > > to opt-in to use of a async mode for existing commands in (c), or that
> > > async commands would be added in parallel with existing sync commands.
> > > IOW, its not a API breakage - its an opt-in extension of existing
> > > functionality.
> > 
> > But you'd need to do async commands for all commands you issued to avoid
> > blocking the io thread so that you could then issue the recovery
> > commands.
> 
> I don't see why that has to be the case. In order to issue an async command
> all that needs to be the case is that command replies should be allowed to
> be sent out of order.
> 
> IOW if command A is blocking and command B is async, then we shoudl be
> allowed to have the following
> 
>    req A
>    req B
>    res A
>    res B
> 
> Or
> 
>    req A
>    req B
>    res B
>    res A
> 
> Or
> 
>    req B
>    req A
>    res B
>    res A
> 
> etc.
> 
> This does imply that you need a separate monitor I/O processing, from the
> command execution thread, but I see no need for all commands to suddenly
> become async. Just allowing interleaved replies is sufficient from the
> POV of the protocol definition. This interleaving is easy to handle from
> the client POV - just requires a unique 'serial' in the request by the
> client, that is copied into the reply by QEMU.

OK, so for that we can just take Marc-André's syntax and call it 'id':
  https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html

then it's upto the caller to ensure those id's are unique.

I do worry about two things:
  a) With this the caller doesn't really know which commands could be
  in parallel - for example if we've got a recovery command that's
  executed by this non-locking thread that's OK, we expect that
  to be doable in parallel.  If in the future though we do
  what you initially suggested and have a bunch of commands get
  routed to the migration thread (say) then those would suddenly
  operate in parallel with other commands that we're previously
  synchronous.

  b) I still worry how the various IO channels will behave on another
  thread.  But that's more a general feeling rather than anything
  specific.

Dave

> Regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> * Daniel P. Berrange (berrange@redhat.com) wrote:
> > This does imply that you need a separate monitor I/O processing, from the
> > command execution thread, but I see no need for all commands to suddenly
> > become async. Just allowing interleaved replies is sufficient from the
> > POV of the protocol definition. This interleaving is easy to handle from
> > the client POV - just requires a unique 'serial' in the request by the
> > client, that is copied into the reply by QEMU.
> 
> OK, so for that we can just take Marc-André's syntax and call it 'id':
>   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> 
> then it's upto the caller to ensure those id's are unique.

Libvirt has in fact generated a unique 'id' for every monitor command
since day 1 of supporting QMP.

> I do worry about two things:
>   a) With this the caller doesn't really know which commands could be
>   in parallel - for example if we've got a recovery command that's
>   executed by this non-locking thread that's OK, we expect that
>   to be doable in parallel.  If in the future though we do
>   what you initially suggested and have a bunch of commands get
>   routed to the migration thread (say) then those would suddenly
>   operate in parallel with other commands that we're previously
>   synchronous.

We could still have an opt-in for async commands. eg default to executing
all commands in the main thread, unless the client issues an explicit
"make it async" command, to switch to allowing the migration thread to
process it async.

 { "execute": "qmp_allow_async",
   "data": { "commands": [
       "migrate_cancel",
   ] } }


 { "return": { "commands": [
       "migrate_cancel",
   ] } }

The server response contains the subset of commands from the request
for which async is supported.

That gives good negotiation ability going forward as we incrementally
support async on more commands.

>   b) I still worry how the various IO channels will behave on another
>   thread.  But that's more a general feeling rather than anything
>   specific.

The only complexity will be around making sure the Chardev code uses
the right GMainContext for any watches on the underlying QIOChannel,
so that we poll() from the custom thread instead of the main thread.
IOW, as long as all I/O is done from the single thread everything
should work fine.

Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
> On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > This does imply that you need a separate monitor I/O processing, from the
> > > command execution thread, but I see no need for all commands to suddenly
> > > become async. Just allowing interleaved replies is sufficient from the
> > > POV of the protocol definition. This interleaving is easy to handle from
> > > the client POV - just requires a unique 'serial' in the request by the
> > > client, that is copied into the reply by QEMU.
> > 
> > OK, so for that we can just take Marc-André's syntax and call it 'id':
> >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> > 
> > then it's upto the caller to ensure those id's are unique.
> 
> Libvirt has in fact generated a unique 'id' for every monitor command
> since day 1 of supporting QMP.
> 
> > I do worry about two things:
> >   a) With this the caller doesn't really know which commands could be
> >   in parallel - for example if we've got a recovery command that's
> >   executed by this non-locking thread that's OK, we expect that
> >   to be doable in parallel.  If in the future though we do
> >   what you initially suggested and have a bunch of commands get
> >   routed to the migration thread (say) then those would suddenly
> >   operate in parallel with other commands that we're previously
> >   synchronous.
> 
> We could still have an opt-in for async commands. eg default to executing
> all commands in the main thread, unless the client issues an explicit
> "make it async" command, to switch to allowing the migration thread to
> process it async.
> 
>  { "execute": "qmp_allow_async",
>    "data": { "commands": [
>        "migrate_cancel",
>    ] } }
> 
> 
>  { "return": { "commands": [
>        "migrate_cancel",
>    ] } }
> 
> The server response contains the subset of commands from the request
> for which async is supported.
> 
> That gives good negotiation ability going forward as we incrementally
> support async on more commands.

I think this goes back to the discussion on which design we'd like to
choose.  IMHO the whole async idea plus the per-command-id is indeed
cleaner and nicer, and I believe that can benefit not only libvirt,
but also other QMP users.  The problem is, I have no idea how long
it'll take to let us have such a feature - I believe that will include
QEMU and Libvirt to both support that.  And it'll be a pity if the
postcopy recovery cannot work only because we cannot guarantee a
stable monitor.

I'm curious whether there are other requirements (besides postcopy
recovery) that would want an always-alive monitor to run some
lock-free commands?  If there is, I'd be more inclined to first
provide a work-around solution like "-qmp-lockfree", and we can
provide a better solution afterwards until when the whole async QMP
work ready.

> 
> >   b) I still worry how the various IO channels will behave on another
> >   thread.  But that's more a general feeling rather than anything
> >   specific.
> 
> The only complexity will be around making sure the Chardev code uses
> the right GMainContext for any watches on the underlying QIOChannel,
> so that we poll() from the custom thread instead of the main thread.
> IOW, as long as all I/O is done from the single thread everything
> should work fine.
> 
> Regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

-- 
Peter Xu

On Thu, Sep 7, 2017 at 9:13 AM, Peter Xu <peterx@redhat.com> wrote:
> On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
>> On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
>> > * Daniel P. Berrange (berrange@redhat.com) wrote:
>> > > This does imply that you need a separate monitor I/O processing, from the
>> > > command execution thread, but I see no need for all commands to suddenly
>> > > become async. Just allowing interleaved replies is sufficient from the
>> > > POV of the protocol definition. This interleaving is easy to handle from
>> > > the client POV - just requires a unique 'serial' in the request by the
>> > > client, that is copied into the reply by QEMU.
>> >
>> > OK, so for that we can just take Marc-André's syntax and call it 'id':
>> >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
>> >
>> > then it's upto the caller to ensure those id's are unique.
>>
>> Libvirt has in fact generated a unique 'id' for every monitor command
>> since day 1 of supporting QMP.
>>
>> > I do worry about two things:
>> >   a) With this the caller doesn't really know which commands could be
>> >   in parallel - for example if we've got a recovery command that's
>> >   executed by this non-locking thread that's OK, we expect that
>> >   to be doable in parallel.  If in the future though we do
>> >   what you initially suggested and have a bunch of commands get
>> >   routed to the migration thread (say) then those would suddenly
>> >   operate in parallel with other commands that we're previously
>> >   synchronous.
>>
>> We could still have an opt-in for async commands. eg default to executing
>> all commands in the main thread, unless the client issues an explicit
>> "make it async" command, to switch to allowing the migration thread to
>> process it async.
>>
>>  { "execute": "qmp_allow_async",
>>    "data": { "commands": [
>>        "migrate_cancel",
>>    ] } }
>>
>>
>>  { "return": { "commands": [
>>        "migrate_cancel",
>>    ] } }
>>
>> The server response contains the subset of commands from the request
>> for which async is supported.
>>
>> That gives good negotiation ability going forward as we incrementally
>> support async on more commands.
>
> I think this goes back to the discussion on which design we'd like to
> choose.  IMHO the whole async idea plus the per-command-id is indeed
> cleaner and nicer, and I believe that can benefit not only libvirt,
> but also other QMP users.  The problem is, I have no idea how long
> it'll take to let us have such a feature - I believe that will include
> QEMU and Libvirt to both support that.  And it'll be a pity if the
> postcopy recovery cannot work only because we cannot guarantee a
> stable monitor.

Please don't rush in a hack, they often introduce new bugs that we
have to support long-term when they are part of the QMP API.

In your original email you mentioned "info cpus".  Have you considered
modifying this command so it does not sync the CPU?  I'm not sure
callers really need to sync the CPU, typically they just want to know
the vcpu numbers, thread IDs, and current state (halted, running,
etc).

The next step after that would be to audit other monitor commands for
unnecessary vcpu synchronization.

Stefan

* Stefan Hajnoczi (stefanha@gmail.com) wrote:
> On Thu, Sep 7, 2017 at 9:13 AM, Peter Xu <peterx@redhat.com> wrote:
> > On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
> >> On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> >> > * Daniel P. Berrange (berrange@redhat.com) wrote:
> >> > > This does imply that you need a separate monitor I/O processing, from the
> >> > > command execution thread, but I see no need for all commands to suddenly
> >> > > become async. Just allowing interleaved replies is sufficient from the
> >> > > POV of the protocol definition. This interleaving is easy to handle from
> >> > > the client POV - just requires a unique 'serial' in the request by the
> >> > > client, that is copied into the reply by QEMU.
> >> >
> >> > OK, so for that we can just take Marc-André's syntax and call it 'id':
> >> >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> >> >
> >> > then it's upto the caller to ensure those id's are unique.
> >>
> >> Libvirt has in fact generated a unique 'id' for every monitor command
> >> since day 1 of supporting QMP.
> >>
> >> > I do worry about two things:
> >> >   a) With this the caller doesn't really know which commands could be
> >> >   in parallel - for example if we've got a recovery command that's
> >> >   executed by this non-locking thread that's OK, we expect that
> >> >   to be doable in parallel.  If in the future though we do
> >> >   what you initially suggested and have a bunch of commands get
> >> >   routed to the migration thread (say) then those would suddenly
> >> >   operate in parallel with other commands that we're previously
> >> >   synchronous.
> >>
> >> We could still have an opt-in for async commands. eg default to executing
> >> all commands in the main thread, unless the client issues an explicit
> >> "make it async" command, to switch to allowing the migration thread to
> >> process it async.
> >>
> >>  { "execute": "qmp_allow_async",
> >>    "data": { "commands": [
> >>        "migrate_cancel",
> >>    ] } }
> >>
> >>
> >>  { "return": { "commands": [
> >>        "migrate_cancel",
> >>    ] } }
> >>
> >> The server response contains the subset of commands from the request
> >> for which async is supported.
> >>
> >> That gives good negotiation ability going forward as we incrementally
> >> support async on more commands.
> >
> > I think this goes back to the discussion on which design we'd like to
> > choose.  IMHO the whole async idea plus the per-command-id is indeed
> > cleaner and nicer, and I believe that can benefit not only libvirt,
> > but also other QMP users.  The problem is, I have no idea how long
> > it'll take to let us have such a feature - I believe that will include
> > QEMU and Libvirt to both support that.  And it'll be a pity if the
> > postcopy recovery cannot work only because we cannot guarantee a
> > stable monitor.
> 
> Please don't rush in a hack, they often introduce new bugs that we
> have to support long-term when they are part of the QMP API.
> 
> In your original email you mentioned "info cpus".  Have you considered
> modifying this command so it does not sync the CPU?  I'm not sure
> callers really need to sync the CPU, typically they just want to know
> the vcpu numbers, thread IDs, and current state (halted, running,
> etc).

But it has the pc as well, so that's actual state.

Dave

> The next step after that would be to audit other monitor commands for
> unnecessary vcpu synchronization.
> 
> Stefan
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

On Thu, Sep 7, 2017 at 10:18 AM, Dr. David Alan Gilbert
<dgilbert@redhat.com> wrote:
> * Stefan Hajnoczi (stefanha@gmail.com) wrote:
>> On Thu, Sep 7, 2017 at 9:13 AM, Peter Xu <peterx@redhat.com> wrote:
>> > On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
>> >> On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
>> >> > * Daniel P. Berrange (berrange@redhat.com) wrote:
>> >> > > This does imply that you need a separate monitor I/O processing, from the
>> >> > > command execution thread, but I see no need for all commands to suddenly
>> >> > > become async. Just allowing interleaved replies is sufficient from the
>> >> > > POV of the protocol definition. This interleaving is easy to handle from
>> >> > > the client POV - just requires a unique 'serial' in the request by the
>> >> > > client, that is copied into the reply by QEMU.
>> >> >
>> >> > OK, so for that we can just take Marc-André's syntax and call it 'id':
>> >> >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
>> >> >
>> >> > then it's upto the caller to ensure those id's are unique.
>> >>
>> >> Libvirt has in fact generated a unique 'id' for every monitor command
>> >> since day 1 of supporting QMP.
>> >>
>> >> > I do worry about two things:
>> >> >   a) With this the caller doesn't really know which commands could be
>> >> >   in parallel - for example if we've got a recovery command that's
>> >> >   executed by this non-locking thread that's OK, we expect that
>> >> >   to be doable in parallel.  If in the future though we do
>> >> >   what you initially suggested and have a bunch of commands get
>> >> >   routed to the migration thread (say) then those would suddenly
>> >> >   operate in parallel with other commands that we're previously
>> >> >   synchronous.
>> >>
>> >> We could still have an opt-in for async commands. eg default to executing
>> >> all commands in the main thread, unless the client issues an explicit
>> >> "make it async" command, to switch to allowing the migration thread to
>> >> process it async.
>> >>
>> >>  { "execute": "qmp_allow_async",
>> >>    "data": { "commands": [
>> >>        "migrate_cancel",
>> >>    ] } }
>> >>
>> >>
>> >>  { "return": { "commands": [
>> >>        "migrate_cancel",
>> >>    ] } }
>> >>
>> >> The server response contains the subset of commands from the request
>> >> for which async is supported.
>> >>
>> >> That gives good negotiation ability going forward as we incrementally
>> >> support async on more commands.
>> >
>> > I think this goes back to the discussion on which design we'd like to
>> > choose.  IMHO the whole async idea plus the per-command-id is indeed
>> > cleaner and nicer, and I believe that can benefit not only libvirt,
>> > but also other QMP users.  The problem is, I have no idea how long
>> > it'll take to let us have such a feature - I believe that will include
>> > QEMU and Libvirt to both support that.  And it'll be a pity if the
>> > postcopy recovery cannot work only because we cannot guarantee a
>> > stable monitor.
>>
>> Please don't rush in a hack, they often introduce new bugs that we
>> have to support long-term when they are part of the QMP API.
>>
>> In your original email you mentioned "info cpus".  Have you considered
>> modifying this command so it does not sync the CPU?  I'm not sure
>> callers really need to sync the CPU, typically they just want to know
>> the vcpu numbers, thread IDs, and current state (halted, running,
>> etc).
>
> But it has the pc as well, so that's actual state.

In what circumstances is the pc useful?

If the client just wants the vcpu -> thread ID mapping, it doesn't
matter at all.

If the CPU is halted, then the PC is already accurate and
synchronization isn't a problem.

If the CPU is running, then an accurate PC is meaningless since it
will have changed the moment the monitor command completes.  We might
as well just keep a copy of the last PC when entering QEMU in a vcpu
thread.

So I think we can offer a perfectly useful PC value without syncing.

Stefan

On Thu, Sep 07, 2017 at 10:18:16AM +0100, Dr. David Alan Gilbert wrote:
> * Stefan Hajnoczi (stefanha@gmail.com) wrote:
> > On Thu, Sep 7, 2017 at 9:13 AM, Peter Xu <peterx@redhat.com> wrote:
> > > On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
> > >> On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> > >> > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > >> > > This does imply that you need a separate monitor I/O processing, from the
> > >> > > command execution thread, but I see no need for all commands to suddenly
> > >> > > become async. Just allowing interleaved replies is sufficient from the
> > >> > > POV of the protocol definition. This interleaving is easy to handle from
> > >> > > the client POV - just requires a unique 'serial' in the request by the
> > >> > > client, that is copied into the reply by QEMU.
> > >> >
> > >> > OK, so for that we can just take Marc-André's syntax and call it 'id':
> > >> >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> > >> >
> > >> > then it's upto the caller to ensure those id's are unique.
> > >>
> > >> Libvirt has in fact generated a unique 'id' for every monitor command
> > >> since day 1 of supporting QMP.
> > >>
> > >> > I do worry about two things:
> > >> >   a) With this the caller doesn't really know which commands could be
> > >> >   in parallel - for example if we've got a recovery command that's
> > >> >   executed by this non-locking thread that's OK, we expect that
> > >> >   to be doable in parallel.  If in the future though we do
> > >> >   what you initially suggested and have a bunch of commands get
> > >> >   routed to the migration thread (say) then those would suddenly
> > >> >   operate in parallel with other commands that we're previously
> > >> >   synchronous.
> > >>
> > >> We could still have an opt-in for async commands. eg default to executing
> > >> all commands in the main thread, unless the client issues an explicit
> > >> "make it async" command, to switch to allowing the migration thread to
> > >> process it async.
> > >>
> > >>  { "execute": "qmp_allow_async",
> > >>    "data": { "commands": [
> > >>        "migrate_cancel",
> > >>    ] } }
> > >>
> > >>
> > >>  { "return": { "commands": [
> > >>        "migrate_cancel",
> > >>    ] } }
> > >>
> > >> The server response contains the subset of commands from the request
> > >> for which async is supported.
> > >>
> > >> That gives good negotiation ability going forward as we incrementally
> > >> support async on more commands.
> > >
> > > I think this goes back to the discussion on which design we'd like to
> > > choose.  IMHO the whole async idea plus the per-command-id is indeed
> > > cleaner and nicer, and I believe that can benefit not only libvirt,
> > > but also other QMP users.  The problem is, I have no idea how long
> > > it'll take to let us have such a feature - I believe that will include
> > > QEMU and Libvirt to both support that.  And it'll be a pity if the
> > > postcopy recovery cannot work only because we cannot guarantee a
> > > stable monitor.
> > 
> > Please don't rush in a hack, they often introduce new bugs that we
> > have to support long-term when they are part of the QMP API.

Sorry, I wasn't meant to push anything.  I was trying to see what
would be the best way to go.

> > 
> > In your original email you mentioned "info cpus".  Have you considered
> > modifying this command so it does not sync the CPU?  I'm not sure
> > callers really need to sync the CPU, typically they just want to know
> > the vcpu numbers, thread IDs, and current state (halted, running,
> > etc).
> 
> But it has the pc as well, so that's actual state.

Yes.  Even if we don't need to sync pc regs for this single "info
cpus" command, I do feel slightly awkward if we don't allow things
like syncing CPU to happen in any command.  IMHO we just need to make
sure these commands may block.

> 
> Dave
> 
> > The next step after that would be to audit other monitor commands for
> > unnecessary vcpu synchronization.

It's really hard to do this for every single command, at least to me.

Comparing to this, I think now I more prefer what Dan has suggested in
the other reply to have an extra way to request async commands while
keep the rest of commands compatible (though obviously I misunderstood
the email when I was writting up previous reply...).

Thanks,

-- 
Peter Xu

On Thu, Sep 07, 2017 at 04:13:41PM +0800, Peter Xu wrote:
> On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
> > On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> > > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > > This does imply that you need a separate monitor I/O processing, from the
> > > > command execution thread, but I see no need for all commands to suddenly
> > > > become async. Just allowing interleaved replies is sufficient from the
> > > > POV of the protocol definition. This interleaving is easy to handle from
> > > > the client POV - just requires a unique 'serial' in the request by the
> > > > client, that is copied into the reply by QEMU.
> > > 
> > > OK, so for that we can just take Marc-André's syntax and call it 'id':
> > >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> > > 
> > > then it's upto the caller to ensure those id's are unique.
> > 
> > Libvirt has in fact generated a unique 'id' for every monitor command
> > since day 1 of supporting QMP.
> > 
> > > I do worry about two things:
> > >   a) With this the caller doesn't really know which commands could be
> > >   in parallel - for example if we've got a recovery command that's
> > >   executed by this non-locking thread that's OK, we expect that
> > >   to be doable in parallel.  If in the future though we do
> > >   what you initially suggested and have a bunch of commands get
> > >   routed to the migration thread (say) then those would suddenly
> > >   operate in parallel with other commands that we're previously
> > >   synchronous.
> > 
> > We could still have an opt-in for async commands. eg default to executing
> > all commands in the main thread, unless the client issues an explicit
> > "make it async" command, to switch to allowing the migration thread to
> > process it async.
> > 
> >  { "execute": "qmp_allow_async",
> >    "data": { "commands": [
> >        "migrate_cancel",
> >    ] } }
> > 
> > 
> >  { "return": { "commands": [
> >        "migrate_cancel",
> >    ] } }
> > 
> > The server response contains the subset of commands from the request
> > for which async is supported.
> > 
> > That gives good negotiation ability going forward as we incrementally
> > support async on more commands.
> 
> I think this goes back to the discussion on which design we'd like to
> choose.  IMHO the whole async idea plus the per-command-id is indeed
> cleaner and nicer, and I believe that can benefit not only libvirt,
> but also other QMP users.  The problem is, I have no idea how long
> it'll take to let us have such a feature - I believe that will include
> QEMU and Libvirt to both support that.  And it'll be a pity if the
> postcopy recovery cannot work only because we cannot guarantee a
> stable monitor.

This is not a blocker for having postcopy recovery feature merged.
It merely means that in a situation where the mainloop is blocked,
then we can't recover, in other situations we'll be able to recover
fine. Sure it would be nice to fix that problem too, but I don't
see it as a block.

I don't think the hacks proposed are a good tradeoff, compared to
fixing the fundamental problem with the monitor impl in QEMU. We
have discussed this monitor problem for years pretty much since
day 1 of QMP being designed, but it never gets serious attention.
IMHO it is well overdue to change that and focus attention on the
root problem and not just punt it down the road yet again by adding
short term hacks.

Adding an extra monitor channel, even as a short term hack, is
*not* short term from libvirt's POV - we'll have to carry that
code for many years into the future, even after QEMU provides
a real fix. So even if QEMU provides such a short term hack, I
would none the less be strongly against libvirt using it.


Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

* Daniel P. Berrange (berrange@redhat.com) wrote:
> On Thu, Sep 07, 2017 at 04:13:41PM +0800, Peter Xu wrote:
> > On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
> > > On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> > > > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > > > This does imply that you need a separate monitor I/O processing, from the
> > > > > command execution thread, but I see no need for all commands to suddenly
> > > > > become async. Just allowing interleaved replies is sufficient from the
> > > > > POV of the protocol definition. This interleaving is easy to handle from
> > > > > the client POV - just requires a unique 'serial' in the request by the
> > > > > client, that is copied into the reply by QEMU.
> > > > 
> > > > OK, so for that we can just take Marc-André's syntax and call it 'id':
> > > >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> > > > 
> > > > then it's upto the caller to ensure those id's are unique.
> > > 
> > > Libvirt has in fact generated a unique 'id' for every monitor command
> > > since day 1 of supporting QMP.
> > > 
> > > > I do worry about two things:
> > > >   a) With this the caller doesn't really know which commands could be
> > > >   in parallel - for example if we've got a recovery command that's
> > > >   executed by this non-locking thread that's OK, we expect that
> > > >   to be doable in parallel.  If in the future though we do
> > > >   what you initially suggested and have a bunch of commands get
> > > >   routed to the migration thread (say) then those would suddenly
> > > >   operate in parallel with other commands that we're previously
> > > >   synchronous.
> > > 
> > > We could still have an opt-in for async commands. eg default to executing
> > > all commands in the main thread, unless the client issues an explicit
> > > "make it async" command, to switch to allowing the migration thread to
> > > process it async.
> > > 
> > >  { "execute": "qmp_allow_async",
> > >    "data": { "commands": [
> > >        "migrate_cancel",
> > >    ] } }
> > > 
> > > 
> > >  { "return": { "commands": [
> > >        "migrate_cancel",
> > >    ] } }
> > > 
> > > The server response contains the subset of commands from the request
> > > for which async is supported.
> > > 
> > > That gives good negotiation ability going forward as we incrementally
> > > support async on more commands.
> > 
> > I think this goes back to the discussion on which design we'd like to
> > choose.  IMHO the whole async idea plus the per-command-id is indeed
> > cleaner and nicer, and I believe that can benefit not only libvirt,
> > but also other QMP users.  The problem is, I have no idea how long
> > it'll take to let us have such a feature - I believe that will include
> > QEMU and Libvirt to both support that.  And it'll be a pity if the
> > postcopy recovery cannot work only because we cannot guarantee a
> > stable monitor.
> 
> This is not a blocker for having postcopy recovery feature merged.
> It merely means that in a situation where the mainloop is blocked,
> then we can't recover, in other situations we'll be able to recover
> fine. Sure it would be nice to fix that problem too, but I don't
> see it as a block.

It's probably OK to merge the recovery code before the monitor code;
but I don't think it's something you'd want to tell users about -
a 'postcopy recovery that only works rarely' isn't much use.

Dave

> I don't think the hacks proposed are a good tradeoff, compared to
> fixing the fundamental problem with the monitor impl in QEMU. We
> have discussed this monitor problem for years pretty much since
> day 1 of QMP being designed, but it never gets serious attention.
> IMHO it is well overdue to change that and focus attention on the
> root problem and not just punt it down the road yet again by adding
> short term hacks.
> 
> Adding an extra monitor channel, even as a short term hack, is
> *not* short term from libvirt's POV - we'll have to carry that
> code for many years into the future, even after QEMU provides
> a real fix. So even if QEMU provides such a short term hack, I
> would none the less be strongly against libvirt using it.
> 
> 
> Regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

On Thu, Sep 07, 2017 at 10:19:47AM +0100, Dr. David Alan Gilbert wrote:
> * Daniel P. Berrange (berrange@redhat.com) wrote:
> > On Thu, Sep 07, 2017 at 04:13:41PM +0800, Peter Xu wrote:
> > > On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
> > > > On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> > > > > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > > > > This does imply that you need a separate monitor I/O processing, from the
> > > > > > command execution thread, but I see no need for all commands to suddenly
> > > > > > become async. Just allowing interleaved replies is sufficient from the
> > > > > > POV of the protocol definition. This interleaving is easy to handle from
> > > > > > the client POV - just requires a unique 'serial' in the request by the
> > > > > > client, that is copied into the reply by QEMU.
> > > > > 
> > > > > OK, so for that we can just take Marc-André's syntax and call it 'id':
> > > > >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> > > > > 
> > > > > then it's upto the caller to ensure those id's are unique.
> > > > 
> > > > Libvirt has in fact generated a unique 'id' for every monitor command
> > > > since day 1 of supporting QMP.
> > > > 
> > > > > I do worry about two things:
> > > > >   a) With this the caller doesn't really know which commands could be
> > > > >   in parallel - for example if we've got a recovery command that's
> > > > >   executed by this non-locking thread that's OK, we expect that
> > > > >   to be doable in parallel.  If in the future though we do
> > > > >   what you initially suggested and have a bunch of commands get
> > > > >   routed to the migration thread (say) then those would suddenly
> > > > >   operate in parallel with other commands that we're previously
> > > > >   synchronous.
> > > > 
> > > > We could still have an opt-in for async commands. eg default to executing
> > > > all commands in the main thread, unless the client issues an explicit
> > > > "make it async" command, to switch to allowing the migration thread to
> > > > process it async.
> > > > 
> > > >  { "execute": "qmp_allow_async",
> > > >    "data": { "commands": [
> > > >        "migrate_cancel",
> > > >    ] } }
> > > > 
> > > > 
> > > >  { "return": { "commands": [
> > > >        "migrate_cancel",
> > > >    ] } }
> > > > 
> > > > The server response contains the subset of commands from the request
> > > > for which async is supported.
> > > > 
> > > > That gives good negotiation ability going forward as we incrementally
> > > > support async on more commands.
> > > 
> > > I think this goes back to the discussion on which design we'd like to
> > > choose.  IMHO the whole async idea plus the per-command-id is indeed
> > > cleaner and nicer, and I believe that can benefit not only libvirt,
> > > but also other QMP users.  The problem is, I have no idea how long
> > > it'll take to let us have such a feature - I believe that will include
> > > QEMU and Libvirt to both support that.  And it'll be a pity if the
> > > postcopy recovery cannot work only because we cannot guarantee a
> > > stable monitor.
> > 
> > This is not a blocker for having postcopy recovery feature merged.
> > It merely means that in a situation where the mainloop is blocked,
> > then we can't recover, in other situations we'll be able to recover
> > fine. Sure it would be nice to fix that problem too, but I don't
> > see it as a block.
> 
> It's probably OK to merge the recovery code before the monitor code;
> but I don't think it's something you'd want to tell users about -
> a 'postcopy recovery that only works rarely' isn't much use.

I dunno. Compared to today where there's zero post-copy recovery,
I think even an incremental improvement is useful. Its a choice
between "your VM is dead" and "you've a 50/50 chance of life".


Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

* Daniel P. Berrange (berrange@redhat.com) wrote:
> On Thu, Sep 07, 2017 at 10:19:47AM +0100, Dr. David Alan Gilbert wrote:
> > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > On Thu, Sep 07, 2017 at 04:13:41PM +0800, Peter Xu wrote:
> > > > On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
> > > > > On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> > > > > > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > > > > > This does imply that you need a separate monitor I/O processing, from the
> > > > > > > command execution thread, but I see no need for all commands to suddenly
> > > > > > > become async. Just allowing interleaved replies is sufficient from the
> > > > > > > POV of the protocol definition. This interleaving is easy to handle from
> > > > > > > the client POV - just requires a unique 'serial' in the request by the
> > > > > > > client, that is copied into the reply by QEMU.
> > > > > > 
> > > > > > OK, so for that we can just take Marc-André's syntax and call it 'id':
> > > > > >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> > > > > > 
> > > > > > then it's upto the caller to ensure those id's are unique.
> > > > > 
> > > > > Libvirt has in fact generated a unique 'id' for every monitor command
> > > > > since day 1 of supporting QMP.
> > > > > 
> > > > > > I do worry about two things:
> > > > > >   a) With this the caller doesn't really know which commands could be
> > > > > >   in parallel - for example if we've got a recovery command that's
> > > > > >   executed by this non-locking thread that's OK, we expect that
> > > > > >   to be doable in parallel.  If in the future though we do
> > > > > >   what you initially suggested and have a bunch of commands get
> > > > > >   routed to the migration thread (say) then those would suddenly
> > > > > >   operate in parallel with other commands that we're previously
> > > > > >   synchronous.
> > > > > 
> > > > > We could still have an opt-in for async commands. eg default to executing
> > > > > all commands in the main thread, unless the client issues an explicit
> > > > > "make it async" command, to switch to allowing the migration thread to
> > > > > process it async.
> > > > > 
> > > > >  { "execute": "qmp_allow_async",
> > > > >    "data": { "commands": [
> > > > >        "migrate_cancel",
> > > > >    ] } }
> > > > > 
> > > > > 
> > > > >  { "return": { "commands": [
> > > > >        "migrate_cancel",
> > > > >    ] } }
> > > > > 
> > > > > The server response contains the subset of commands from the request
> > > > > for which async is supported.
> > > > > 
> > > > > That gives good negotiation ability going forward as we incrementally
> > > > > support async on more commands.
> > > > 
> > > > I think this goes back to the discussion on which design we'd like to
> > > > choose.  IMHO the whole async idea plus the per-command-id is indeed
> > > > cleaner and nicer, and I believe that can benefit not only libvirt,
> > > > but also other QMP users.  The problem is, I have no idea how long
> > > > it'll take to let us have such a feature - I believe that will include
> > > > QEMU and Libvirt to both support that.  And it'll be a pity if the
> > > > postcopy recovery cannot work only because we cannot guarantee a
> > > > stable monitor.
> > > 
> > > This is not a blocker for having postcopy recovery feature merged.
> > > It merely means that in a situation where the mainloop is blocked,
> > > then we can't recover, in other situations we'll be able to recover
> > > fine. Sure it would be nice to fix that problem too, but I don't
> > > see it as a block.
> > 
> > It's probably OK to merge the recovery code before the monitor code;
> > but I don't think it's something you'd want to tell users about -
> > a 'postcopy recovery that only works rarely' isn't much use.
> 
> I dunno. Compared to today where there's zero post-copy recovery,
> I think even an incremental improvement is useful. Its a choice
> between "your VM is dead" and "you've a 50/50 chance of life".

There's a chunk of people who wont use postcopy because they regard
it as dangerous; they need something that works in most cases before
they'll use it.

Dave

> Regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

"Dr. David Alan Gilbert" <dgilbert@redhat.com> writes:

> * Daniel P. Berrange (berrange@redhat.com) wrote:
>> On Thu, Sep 07, 2017 at 04:13:41PM +0800, Peter Xu wrote:
>> > On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
>> > > On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
>> > > > * Daniel P. Berrange (berrange@redhat.com) wrote:
>> > > > > This does imply that you need a separate monitor I/O processing, from the
>> > > > > command execution thread, but I see no need for all commands to suddenly
>> > > > > become async. Just allowing interleaved replies is sufficient from the
>> > > > > POV of the protocol definition. This interleaving is easy to handle from
>> > > > > the client POV - just requires a unique 'serial' in the request by the
>> > > > > client, that is copied into the reply by QEMU.
>> > > > 
>> > > > OK, so for that we can just take Marc-André's syntax and call it 'id':
>> > > >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
>> > > > 
>> > > > then it's upto the caller to ensure those id's are unique.
>> > > 
>> > > Libvirt has in fact generated a unique 'id' for every monitor command
>> > > since day 1 of supporting QMP.
>> > > 
>> > > > I do worry about two things:
>> > > >   a) With this the caller doesn't really know which commands could be
>> > > >   in parallel - for example if we've got a recovery command that's
>> > > >   executed by this non-locking thread that's OK, we expect that
>> > > >   to be doable in parallel.  If in the future though we do
>> > > >   what you initially suggested and have a bunch of commands get
>> > > >   routed to the migration thread (say) then those would suddenly
>> > > >   operate in parallel with other commands that we're previously
>> > > >   synchronous.
>> > > 
>> > > We could still have an opt-in for async commands. eg default to executing
>> > > all commands in the main thread, unless the client issues an explicit
>> > > "make it async" command, to switch to allowing the migration thread to
>> > > process it async.
>> > > 
>> > >  { "execute": "qmp_allow_async",
>> > >    "data": { "commands": [
>> > >        "migrate_cancel",
>> > >    ] } }
>> > > 
>> > > 
>> > >  { "return": { "commands": [
>> > >        "migrate_cancel",
>> > >    ] } }
>> > > 
>> > > The server response contains the subset of commands from the request
>> > > for which async is supported.
>> > > 
>> > > That gives good negotiation ability going forward as we incrementally
>> > > support async on more commands.
>> > 
>> > I think this goes back to the discussion on which design we'd like to
>> > choose.  IMHO the whole async idea plus the per-command-id is indeed
>> > cleaner and nicer, and I believe that can benefit not only libvirt,
>> > but also other QMP users.  The problem is, I have no idea how long
>> > it'll take to let us have such a feature - I believe that will include
>> > QEMU and Libvirt to both support that.  And it'll be a pity if the
>> > postcopy recovery cannot work only because we cannot guarantee a
>> > stable monitor.
>> 
>> This is not a blocker for having postcopy recovery feature merged.
>> It merely means that in a situation where the mainloop is blocked,
>> then we can't recover, in other situations we'll be able to recover
>> fine. Sure it would be nice to fix that problem too, but I don't
>> see it as a block.
>
> It's probably OK to merge the recovery code before the monitor code;
> but I don't think it's something you'd want to tell users about -
> a 'postcopy recovery that only works rarely' isn't much use.

"Rarely"?  Are main loop hangs *that* common?

Can we quantify the problem to help gauge urgency?

* Markus Armbruster (armbru@redhat.com) wrote:
> "Dr. David Alan Gilbert" <dgilbert@redhat.com> writes:
> 
> > * Daniel P. Berrange (berrange@redhat.com) wrote:
> >> On Thu, Sep 07, 2017 at 04:13:41PM +0800, Peter Xu wrote:
> >> > On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
> >> > > On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> >> > > > * Daniel P. Berrange (berrange@redhat.com) wrote:
> >> > > > > This does imply that you need a separate monitor I/O processing, from the
> >> > > > > command execution thread, but I see no need for all commands to suddenly
> >> > > > > become async. Just allowing interleaved replies is sufficient from the
> >> > > > > POV of the protocol definition. This interleaving is easy to handle from
> >> > > > > the client POV - just requires a unique 'serial' in the request by the
> >> > > > > client, that is copied into the reply by QEMU.
> >> > > > 
> >> > > > OK, so for that we can just take Marc-André's syntax and call it 'id':
> >> > > >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> >> > > > 
> >> > > > then it's upto the caller to ensure those id's are unique.
> >> > > 
> >> > > Libvirt has in fact generated a unique 'id' for every monitor command
> >> > > since day 1 of supporting QMP.
> >> > > 
> >> > > > I do worry about two things:
> >> > > >   a) With this the caller doesn't really know which commands could be
> >> > > >   in parallel - for example if we've got a recovery command that's
> >> > > >   executed by this non-locking thread that's OK, we expect that
> >> > > >   to be doable in parallel.  If in the future though we do
> >> > > >   what you initially suggested and have a bunch of commands get
> >> > > >   routed to the migration thread (say) then those would suddenly
> >> > > >   operate in parallel with other commands that we're previously
> >> > > >   synchronous.
> >> > > 
> >> > > We could still have an opt-in for async commands. eg default to executing
> >> > > all commands in the main thread, unless the client issues an explicit
> >> > > "make it async" command, to switch to allowing the migration thread to
> >> > > process it async.
> >> > > 
> >> > >  { "execute": "qmp_allow_async",
> >> > >    "data": { "commands": [
> >> > >        "migrate_cancel",
> >> > >    ] } }
> >> > > 
> >> > > 
> >> > >  { "return": { "commands": [
> >> > >        "migrate_cancel",
> >> > >    ] } }
> >> > > 
> >> > > The server response contains the subset of commands from the request
> >> > > for which async is supported.
> >> > > 
> >> > > That gives good negotiation ability going forward as we incrementally
> >> > > support async on more commands.
> >> > 
> >> > I think this goes back to the discussion on which design we'd like to
> >> > choose.  IMHO the whole async idea plus the per-command-id is indeed
> >> > cleaner and nicer, and I believe that can benefit not only libvirt,
> >> > but also other QMP users.  The problem is, I have no idea how long
> >> > it'll take to let us have such a feature - I believe that will include
> >> > QEMU and Libvirt to both support that.  And it'll be a pity if the
> >> > postcopy recovery cannot work only because we cannot guarantee a
> >> > stable monitor.
> >> 
> >> This is not a blocker for having postcopy recovery feature merged.
> >> It merely means that in a situation where the mainloop is blocked,
> >> then we can't recover, in other situations we'll be able to recover
> >> fine. Sure it would be nice to fix that problem too, but I don't
> >> see it as a block.
> >
> > It's probably OK to merge the recovery code before the monitor code;
> > but I don't think it's something you'd want to tell users about -
> > a 'postcopy recovery that only works rarely' isn't much use.
> 
> "Rarely"?  Are main loop hangs *that* common?
> 
> Can we quantify the problem to help gauge urgency?

Not really; it depends on workload and behaviour.  The people who worry
about postcopy recovery actually care about it working in almost every
case.
So I'm OK to add the recovery code, it's just not something we should
be shouting about to users until we have the monitor fixed.

Dave

--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

* Peter Xu (peterx@redhat.com) wrote:
> On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
> > On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> > > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > > This does imply that you need a separate monitor I/O processing, from the
> > > > command execution thread, but I see no need for all commands to suddenly
> > > > become async. Just allowing interleaved replies is sufficient from the
> > > > POV of the protocol definition. This interleaving is easy to handle from
> > > > the client POV - just requires a unique 'serial' in the request by the
> > > > client, that is copied into the reply by QEMU.
> > > 
> > > OK, so for that we can just take Marc-André's syntax and call it 'id':
> > >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> > > 
> > > then it's upto the caller to ensure those id's are unique.
> > 
> > Libvirt has in fact generated a unique 'id' for every monitor command
> > since day 1 of supporting QMP.
> > 
> > > I do worry about two things:
> > >   a) With this the caller doesn't really know which commands could be
> > >   in parallel - for example if we've got a recovery command that's
> > >   executed by this non-locking thread that's OK, we expect that
> > >   to be doable in parallel.  If in the future though we do
> > >   what you initially suggested and have a bunch of commands get
> > >   routed to the migration thread (say) then those would suddenly
> > >   operate in parallel with other commands that we're previously
> > >   synchronous.
> > 
> > We could still have an opt-in for async commands. eg default to executing
> > all commands in the main thread, unless the client issues an explicit
> > "make it async" command, to switch to allowing the migration thread to
> > process it async.
> > 
> >  { "execute": "qmp_allow_async",
> >    "data": { "commands": [
> >        "migrate_cancel",
> >    ] } }
> > 
> > 
> >  { "return": { "commands": [
> >        "migrate_cancel",
> >    ] } }
> > 
> > The server response contains the subset of commands from the request
> > for which async is supported.
> > 
> > That gives good negotiation ability going forward as we incrementally
> > support async on more commands.
> 
> I think this goes back to the discussion on which design we'd like to
> choose.  IMHO the whole async idea plus the per-command-id is indeed
> cleaner and nicer, and I believe that can benefit not only libvirt,
> but also other QMP users.  The problem is, I have no idea how long
> it'll take to let us have such a feature - I believe that will include
> QEMU and Libvirt to both support that.  And it'll be a pity if the
> postcopy recovery cannot work only because we cannot guarantee a
> stable monitor.

libvirt will need changes for postcopy recovery however we do it;
so we need to do it the way they want.

I think Dan's suggestion isn't as hard as it initially sounded;  a first
thing to try would be taking all the monitor IO into another thread
and feeding all commands to the main thread for execution - that sounds
like the hard part.
(I'm not sure how multiple monitors interact for this).

Dave

> I'm curious whether there are other requirements (besides postcopy
> recovery) that would want an always-alive monitor to run some
> lock-free commands?  If there is, I'd be more inclined to first
> provide a work-around solution like "-qmp-lockfree", and we can
> provide a better solution afterwards until when the whole async QMP
> work ready.
> 
> > 
> > >   b) I still worry how the various IO channels will behave on another
> > >   thread.  But that's more a general feeling rather than anything
> > >   specific.
> > 
> > The only complexity will be around making sure the Chardev code uses
> > the right GMainContext for any watches on the underlying QIOChannel,
> > so that we poll() from the custom thread instead of the main thread.
> > IOW, as long as all I/O is done from the single thread everything
> > should work fine.
> > 
> > Regards,
> > Daniel
> > -- 
> > |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> > |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> > |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
> 
> -- 
> Peter Xu
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

On Thu, Sep 07, 2017 at 10:15:09AM +0100, Dr. David Alan Gilbert wrote:
> * Peter Xu (peterx@redhat.com) wrote:
> > On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
> > > On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> > > > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > > > This does imply that you need a separate monitor I/O processing, from the
> > > > > command execution thread, but I see no need for all commands to suddenly
> > > > > become async. Just allowing interleaved replies is sufficient from the
> > > > > POV of the protocol definition. This interleaving is easy to handle from
> > > > > the client POV - just requires a unique 'serial' in the request by the
> > > > > client, that is copied into the reply by QEMU.
> > > > 
> > > > OK, so for that we can just take Marc-André's syntax and call it 'id':
> > > >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> > > > 
> > > > then it's upto the caller to ensure those id's are unique.
> > > 
> > > Libvirt has in fact generated a unique 'id' for every monitor command
> > > since day 1 of supporting QMP.
> > > 
> > > > I do worry about two things:
> > > >   a) With this the caller doesn't really know which commands could be
> > > >   in parallel - for example if we've got a recovery command that's
> > > >   executed by this non-locking thread that's OK, we expect that
> > > >   to be doable in parallel.  If in the future though we do
> > > >   what you initially suggested and have a bunch of commands get
> > > >   routed to the migration thread (say) then those would suddenly
> > > >   operate in parallel with other commands that we're previously
> > > >   synchronous.
> > > 
> > > We could still have an opt-in for async commands. eg default to executing
> > > all commands in the main thread, unless the client issues an explicit
> > > "make it async" command, to switch to allowing the migration thread to
> > > process it async.
> > > 
> > >  { "execute": "qmp_allow_async",
> > >    "data": { "commands": [
> > >        "migrate_cancel",
> > >    ] } }
> > > 
> > > 
> > >  { "return": { "commands": [
> > >        "migrate_cancel",
> > >    ] } }
> > > 
> > > The server response contains the subset of commands from the request
> > > for which async is supported.
> > > 
> > > That gives good negotiation ability going forward as we incrementally
> > > support async on more commands.
> > 
> > I think this goes back to the discussion on which design we'd like to
> > choose.  IMHO the whole async idea plus the per-command-id is indeed
> > cleaner and nicer, and I believe that can benefit not only libvirt,
> > but also other QMP users.  The problem is, I have no idea how long
> > it'll take to let us have such a feature - I believe that will include
> > QEMU and Libvirt to both support that.  And it'll be a pity if the
> > postcopy recovery cannot work only because we cannot guarantee a
> > stable monitor.
> 
> libvirt will need changes for postcopy recovery however we do it;
> so we need to do it the way they want.
> 
> I think Dan's suggestion isn't as hard as it initially sounded;  a first
> thing to try would be taking all the monitor IO into another thread
> and feeding all commands to the main thread for execution - that sounds
> like the hard part.
> (I'm not sure how multiple monitors interact for this).

Multiple monitors is probably not as hard as it sounds. No matter how
many monitors you have configured today, they'll all serviced by the
main loop so commands from each monitor are strictly serialized.

So if you moved I/O processing to a separate thread, and had a single
queue of commands to be executed by the main thread, you would still
have the exact same serialized processing across multiple monitors.

Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

Peter Xu <peterx@redhat.com> writes:

> On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
>> On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
>> > * Daniel P. Berrange (berrange@redhat.com) wrote:
>> > > This does imply that you need a separate monitor I/O processing, from the
>> > > command execution thread, but I see no need for all commands to suddenly
>> > > become async. Just allowing interleaved replies is sufficient from the
>> > > POV of the protocol definition. This interleaving is easy to handle from
>> > > the client POV - just requires a unique 'serial' in the request by the
>> > > client, that is copied into the reply by QEMU.
>> > 
>> > OK, so for that we can just take Marc-André's syntax and call it 'id':
>> >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
>> > 
>> > then it's upto the caller to ensure those id's are unique.
>> 
>> Libvirt has in fact generated a unique 'id' for every monitor command
>> since day 1 of supporting QMP.
>> 
>> > I do worry about two things:
>> >   a) With this the caller doesn't really know which commands could be
>> >   in parallel - for example if we've got a recovery command that's
>> >   executed by this non-locking thread that's OK, we expect that
>> >   to be doable in parallel.  If in the future though we do
>> >   what you initially suggested and have a bunch of commands get
>> >   routed to the migration thread (say) then those would suddenly
>> >   operate in parallel with other commands that we're previously
>> >   synchronous.
>> 
>> We could still have an opt-in for async commands. eg default to executing
>> all commands in the main thread, unless the client issues an explicit
>> "make it async" command, to switch to allowing the migration thread to
>> process it async.
>> 
>>  { "execute": "qmp_allow_async",
>>    "data": { "commands": [
>>        "migrate_cancel",
>>    ] } }
>> 
>> 
>>  { "return": { "commands": [
>>        "migrate_cancel",
>>    ] } }
>> 
>> The server response contains the subset of commands from the request
>> for which async is supported.
>> 
>> That gives good negotiation ability going forward as we incrementally
>> support async on more commands.
>
> I think this goes back to the discussion on which design we'd like to
> choose.  IMHO the whole async idea plus the per-command-id is indeed
> cleaner and nicer, and I believe that can benefit not only libvirt,

The following may be a bit harsh in places.  I apologize in advance.  A
better writer than me wouldn't have to resort to that.  I've tried a few
times to make my point that "async QMP" is neither necessary nor
sufficient for monitor availability, but apparently without luck, since
there's still talk like it was.  I hope this attempt will work.

> but also other QMP users.  The problem is, I have no idea how long
> it'll take to let us have such a feature - I believe that will include
> QEMU and Libvirt to both support that.  And it'll be a pity if the
> postcopy recovery cannot work only because we cannot guarantee a
> stable monitor.
>
> I'm curious whether there are other requirements (besides postcopy
> recovery) that would want an always-alive monitor to run some
> lock-free commands?  If there is, I'd be more inclined to first
> provide a work-around solution like "-qmp-lockfree", and we can
> provide a better solution afterwards until when the whole async QMP
> work ready.

Yes, there are other requirements for "async QMP", and no, "async QMP"
isn't a solution, but at best a part of a solution.

Before I talk about QMP requirements, I need to ask a whole raft of
questions, because so far this thread feels like dreaming up grand
designs with only superficial understanding of the subject matter.
Quite possibly because *my* understanding is superficial.  If yours
isn't, great!  Go answer my questions :)

The root problem are main loop hangs.  QMP monitor hangs are merely a
special case.

The main loop should not hang.  We've always violated that design
assumption in places, e.g. in monitor commands that write to disk, and
thus can hang indefinitely with NFS.  Post-copy adds more violations, as
Stefan pointed out.

I can't say whether solving the special case "QMP monitor hangs" without
also solving "main loop hangs" is useful.  A perfectly available QMP
monitor buys you nothing if it feeds a command queue that isn't being
emptied because its consumers all hang.

So, what exactly is going to drain the command queue?  If there's more
than one consumer, how exactly are commands from the queue dispatched to
the consumers?

What are the "no hang" guarantees (if any) and conditions for each of
these consumers?

We can have any number of QMP monitors today.  Would each of them feed
its own queue?  Would they all feed a shared queue?

How exactly is opt-in asynchronous to work?  Per QMP monitor?  Per
command?

What does it mean when an asynchronous command follows a synchronous
command in the same QMP monitor?  I would expect the synchronous command
to complete before the asynchronous command, because that's what
synchronous means, isn't it?  To keep your QMP monitor available, you
then must not send synchronous commands that can hang.

How can we determine whether a certain synchronous command can hang?
Note that with opt-in async, *all* commands are also synchronous
commands.

In short, explain to me how exactly you plan to ensure that certain QMP
commands (such as post-copy recovery) can always "get through", in the
presence of multiple monitors, hanging main loop, hanging synchronous
commands, hanging whatever-else-can-now-hang-in-this-post-copy-world.


Now let's talk about QMP requirements.

Any addition to QMP must consider what exists already.

You may add more of the same.

You may generalize existing stuff.

You may change existing stuff if you have sufficient reason, subject to
backward compatibility constraints.

But attempts to add new ways to do the same old stuff without properly
integrating the existing ways are not going to fly.

In particular, any new way to start some job, monitor and control it
while it lives, get notified about its state changes and so forth must
integrate the existing ways.  These include block jobs (probably the
most sophisticated of the lot), migration, dump-guest-memory, and
possibly more.  They all work the same way: synchronous command to kick
off the job, more synchronous commands to monitor and control, events to
notify.  They do differ in detail.

Asynchronous commands are a new way to do this.  When you only need to
be notified on "done", and don't need to monitor / control, they fit the
bill quite neatly.

However, we can't just ignore the cases where we need more than that!
For those, we want a single generic solution instead of the several ad
hoc solutions we have now.

If we add asynchronous commands *now*, and for simple cases only, we add
yet another special case for a future generic solution to integrate.
I'm not going to let that happen.

I figure the closest to a generic solution we have is block jobs.
Perhaps a generic solution could be had by abstracting away the "block"
from "block jobs", leaving just "jobs".

Another approach is generalizing the asynchronous command proposal to
fully cover the not-so-simple cases.

If you'd rather want to make progress on monitor availability without
cracking the "jobs" problem, you're in luck!  Use your license to "add
more of the same": synchronous command to start a job, query to monitor,
event to notify.  

If you insist on tying your monitor availability solution to
asynchronous commands, then I'm in luck!  I just found volunteers to
solve the "jobs" problem for me.

On Thu, Sep 07, 2017 at 02:59:28PM +0200, Markus Armbruster wrote:
> So, what exactly is going to drain the command queue?  If there's more
> than one consumer, how exactly are commands from the queue dispatched to
> the consumers?

In terms of my proposal, for any single command there should only ever
be a single consumer. The default consumer would be the main event loop
thread, such that we have no semantic change to QMP operation from today.

Some commands that are capable of being made "async", would have a
different consumer. For example, if the client requested the 'migrate-cancel'
be made async, this would change things such that the migration thread is
now responsible for consuming the "migrate-cancel" command, instead of the
default main loop.

> What are the "no hang" guarantees (if any) and conditions for each of
> these consumers?

The non-main thread consumers would have to have some reasonable
guarantee that they won't block on a lock held by the main loop,
otherwise the whole feature is largely useless.

> We can have any number of QMP monitors today.  Would each of them feed
> its own queue?  Would they all feed a shared queue?

Currently with multiple QMP monitors, everything runs in the main
loop, so commands arriving across  multiple monitors are 100%
serialized and processed strictly in the order in which QEMU reads
them off the wire.  To maintain these semantics, we would need to
have a single shared queue for the default main loop consumer, so
that ordering does not change.

> How exactly is opt-in asynchronous to work?  Per QMP monitor?  Per
> command?

Per monitor+command. ie just because libvirt knows how to cope with
async execution on the monitor it has open, does not mean that a
different app on the 2nd monitor command can cope. So in my proposal
the switch to async must be scoped to the particular command only
for the monitor connection that requesteed it.

> What does it mean when an asynchronous command follows a synchronous
> command in the same QMP monitor?  I would expect the synchronous command
> to complete before the asynchronous command, because that's what
> synchronous means, isn't it?  To keep your QMP monitor available, you
> then must not send synchronous commands that can hang.

No, that is not what I described. All synchronous commands are
serialized wrt each other, just as today. An asychronous command
can run as soon as it is received, regardless of whether any
earlier sent sync commands are still executing or pending. This
is trivial to achieve when you separate monitor I/O from command
execution in separate threads, provided of course the async
command consumers are not in the main loop.

> How can we determine whether a certain synchronous command can hang?
> Note that with opt-in async, *all* commands are also synchronous
> commands.
> 
> In short, explain to me how exactly you plan to ensure that certain QMP
> commands (such as post-copy recovery) can always "get through", in the
> presence of multiple monitors, hanging main loop, hanging synchronous
> commands, hanging whatever-else-can-now-hang-in-this-post-copy-world.

Taking migrate-cancel as the example. The migration code already has
a background thread doing work independantly onthe main loop. Upon
marking the migrate-cancel command as async, the migration control
thread would become the consumer of migrate-cancel. This allows the
migration operation to be cancelled immediately, regardless of whether
there are earlier monitor commands blocked in the main loop.

Of course this assumes the migration control thread can't block
for locks held by the main thread. 

> Now let's talk about QMP requirements.
> 
> Any addition to QMP must consider what exists already.
> 
> You may add more of the same.
> 
> You may generalize existing stuff.
> 
> You may change existing stuff if you have sufficient reason, subject to
> backward compatibility constraints.
> 
> But attempts to add new ways to do the same old stuff without properly
> integrating the existing ways are not going to fly.
> 
> In particular, any new way to start some job, monitor and control it
> while it lives, get notified about its state changes and so forth must
> integrate the existing ways.  These include block jobs (probably the
> most sophisticated of the lot), migration, dump-guest-memory, and
> possibly more.  They all work the same way: synchronous command to kick
> off the job, more synchronous commands to monitor and control, events to
> notify.  They do differ in detail.
> 
> Asynchronous commands are a new way to do this.  When you only need to
> be notified on "done", and don't need to monitor / control, they fit the
> bill quite neatly.
> 
> However, we can't just ignore the cases where we need more than that!
> For those, we want a single generic solution instead of the several ad
> hoc solutions we have now.
> 
> If we add asynchronous commands *now*, and for simple cases only, we add
> yet another special case for a future generic solution to integrate.
> I'm not going to let that happen.

With the async commands suggestion, while it would initially not
provide a way to query incremental status, that could easily be
fitted in.  Because command replies from async commands may be
out-of-order wrt the original requests, clients would need to
provide a unique ID for each command run. This originally was
part of QMP spec but then dropped, but libvirt still actually
generates a uniqe ID for every QMP command.

Given this, one option is to actually use the QMP command ID as
a job ID, and let you query ongoing status via some new QMP
command that accepts the ID of the job to be queried. A complexity
with this is how to make the jobs visible across multiple QMP
monitors. The job ID might actually have to be a combination of
the serial ID from the QMP command, and the ID of the monitor
chardev combined.


Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

"Daniel P. Berrange" <berrange@redhat.com> writes:

> On Thu, Sep 07, 2017 at 02:59:28PM +0200, Markus Armbruster wrote:
>> So, what exactly is going to drain the command queue?  If there's more
>> than one consumer, how exactly are commands from the queue dispatched to
>> the consumers?
>
> In terms of my proposal, for any single command there should only ever
> be a single consumer. The default consumer would be the main event loop
> thread, such that we have no semantic change to QMP operation from today.
>
> Some commands that are capable of being made "async", would have a
> different consumer. For example, if the client requested the 'migrate-cancel'
> be made async, this would change things such that the migration thread is
> now responsible for consuming the "migrate-cancel" command, instead of the
> default main loop.
>
>> What are the "no hang" guarantees (if any) and conditions for each of
>> these consumers?
>
> The non-main thread consumers would have to have some reasonable
> guarantee that they won't block on a lock held by the main loop,
> otherwise the whole feature is largely useless.

Same if they block indefinitely on anything else, actually.  In other
words, we need to talk about liveness.

Threads by themselves don't buy us liveness.  Being careful with
operations that may block does.  That care may lead to farming out
certain operations to other threads, where they may block without harm.

You only talk about "the non-main thread consumers".  What about the
main thread?  Is it okay for the main thread to block?  If yes, why?

>> We can have any number of QMP monitors today.  Would each of them feed
>> its own queue?  Would they all feed a shared queue?
>
> Currently with multiple QMP monitors, everything runs in the main
> loop, so commands arriving across  multiple monitors are 100%
> serialized and processed strictly in the order in which QEMU reads
> them off the wire.  To maintain these semantics, we would need to
> have a single shared queue for the default main loop consumer, so
> that ordering does not change.
>
>> How exactly is opt-in asynchronous to work?  Per QMP monitor?  Per
>> command?
>
> Per monitor+command. ie just because libvirt knows how to cope with
> async execution on the monitor it has open, does not mean that a
> different app on the 2nd monitor command can cope. So in my proposal
> the switch to async must be scoped to the particular command only
> for the monitor connection that requesteed it.
>
>> What does it mean when an asynchronous command follows a synchronous
>> command in the same QMP monitor?  I would expect the synchronous command
>> to complete before the asynchronous command, because that's what
>> synchronous means, isn't it?  To keep your QMP monitor available, you
>> then must not send synchronous commands that can hang.
>
> No, that is not what I described. All synchronous commands are
> serialized wrt each other, just as today. An asychronous command
> can run as soon as it is received, regardless of whether any
> earlier sent sync commands are still executing or pending. This
> is trivial to achieve when you separate monitor I/O from command
> execution in separate threads, provided of course the async
> command consumers are not in the main loop.

So, a synchronous command is synchronous with respect to other commands,
except for certain non-blocking commands.  The distinctive feature of
the latter isn't so much an asynchronous reply, but out-of-band
dispatch.

Out-of-band dispatch of commands that cannot block in fact orthogonal to
asynchronous replies.  I can't see why out-of-band dispatch of
synchronous non-blocking commands wouldn't work, too.

>> How can we determine whether a certain synchronous command can hang?
>> Note that with opt-in async, *all* commands are also synchronous
>> commands.
>> 
>> In short, explain to me how exactly you plan to ensure that certain QMP
>> commands (such as post-copy recovery) can always "get through", in the
>> presence of multiple monitors, hanging main loop, hanging synchronous
>> commands, hanging whatever-else-can-now-hang-in-this-post-copy-world.
>
> Taking migrate-cancel as the example. The migration code already has
> a background thread doing work independantly onthe main loop. Upon
> marking the migrate-cancel command as async, the migration control
> thread would become the consumer of migrate-cancel.

From 30,000 feet, the QMP monitor sends a "cancel" message to the
migration thread, and later receives a "canceled" message from the
migration thread.

From 300 feet, we use the migrate-cancel QMP command as the cancel
message, and its success response as the "canceled" message.

In other words, we're pressing the external QM-Protocol into service as
internal message passing protocol.

>                                                     This allows the
> migration operation to be cancelled immediately, regardless of whether
> there are earlier monitor commands blocked in the main loop.

The necessary part is moving all operations that can block out of
whatever loop runs the monitor, be it the main loop, some other event
loop, or a dedicated monitor thread's monitor loop.

Moving out non-blocking operations isn't necessary.  migrate-cancel
could communicate with the migration thread by any suitable mechanism or
protocol.  It doesn't have to be QMP.  Why would we want it to be QMP?

> Of course this assumes the migration control thread can't block
> for locks held by the main thread.

Thanks for your answers, they help.

>> Now let's talk about QMP requirements.
>> 
>> Any addition to QMP must consider what exists already.
>> 
>> You may add more of the same.
>> 
>> You may generalize existing stuff.
>> 
>> You may change existing stuff if you have sufficient reason, subject to
>> backward compatibility constraints.
>> 
>> But attempts to add new ways to do the same old stuff without properly
>> integrating the existing ways are not going to fly.
>> 
>> In particular, any new way to start some job, monitor and control it
>> while it lives, get notified about its state changes and so forth must
>> integrate the existing ways.  These include block jobs (probably the
>> most sophisticated of the lot), migration, dump-guest-memory, and
>> possibly more.  They all work the same way: synchronous command to kick
>> off the job, more synchronous commands to monitor and control, events to
>> notify.  They do differ in detail.
>> 
>> Asynchronous commands are a new way to do this.  When you only need to
>> be notified on "done", and don't need to monitor / control, they fit the
>> bill quite neatly.
>> 
>> However, we can't just ignore the cases where we need more than that!
>> For those, we want a single generic solution instead of the several ad
>> hoc solutions we have now.
>> 
>> If we add asynchronous commands *now*, and for simple cases only, we add
>> yet another special case for a future generic solution to integrate.
>> I'm not going to let that happen.
>
> With the async commands suggestion, while it would initially not
> provide a way to query incremental status, that could easily be
> fitted in.

This is [*] below.

>             Because command replies from async commands may be
> out-of-order wrt the original requests, clients would need to
> provide a unique ID for each command run. This originally was
> part of QMP spec but then dropped, but libvirt still actually
> generates a uniqe ID for every QMP command.
>
> Given this, one option is to actually use the QMP command ID as
> a job ID, and let you query ongoing status via some new QMP
> command that accepts the ID of the job to be queried. A complexity
> with this is how to make the jobs visible across multiple QMP
> monitors. The job ID might actually have to be a combination of
> the serial ID from the QMP command, and the ID of the monitor
> chardev combined.

Yes.  The job ID must be unique across all QMP monitors to make
broadcast notifications work.

>> I figure the closest to a generic solution we have is block jobs.
>> Perhaps a generic solution could be had by abstracting away the "block"
>> from "block jobs", leaving just "jobs".

[*] starts here:

>> Another approach is generalizing the asynchronous command proposal to
>> fully cover the not-so-simple cases.

We know asynchronous commands "fully cover" when we can use them to
replace all the existing job-like commands.

Until then, they enlarge rather than solve our jobs problem.

I get the need for an available monitor.  But I need to balance it with
other needs.  Can we find a solution for our monitor availability
problem that doesn't enlarge our jobs problem?

>> If you'd rather want to make progress on monitor availability without
>> cracking the "jobs" problem, you're in luck!  Use your license to "add
>> more of the same": synchronous command to start a job, query to monitor,
>> event to notify.  
>> 
>> If you insist on tying your monitor availability solution to
>> asynchronous commands, then I'm in luck!  I just found volunteers to
>> solve the "jobs" problem for me.

* Markus Armbruster (armbru@redhat.com) wrote:
> "Daniel P. Berrange" <berrange@redhat.com> writes:
> 
> > On Thu, Sep 07, 2017 at 02:59:28PM +0200, Markus Armbruster wrote:
> >> So, what exactly is going to drain the command queue?  If there's more
> >> than one consumer, how exactly are commands from the queue dispatched to
> >> the consumers?
> >
> > In terms of my proposal, for any single command there should only ever
> > be a single consumer. The default consumer would be the main event loop
> > thread, such that we have no semantic change to QMP operation from today.
> >
> > Some commands that are capable of being made "async", would have a
> > different consumer. For example, if the client requested the 'migrate-cancel'
> > be made async, this would change things such that the migration thread is
> > now responsible for consuming the "migrate-cancel" command, instead of the
> > default main loop.
> >
> >> What are the "no hang" guarantees (if any) and conditions for each of
> >> these consumers?
> >
> > The non-main thread consumers would have to have some reasonable
> > guarantee that they won't block on a lock held by the main loop,
> > otherwise the whole feature is largely useless.
> 
> Same if they block indefinitely on anything else, actually.  In other
> words, we need to talk about liveness.
> 
> Threads by themselves don't buy us liveness.  Being careful with
> operations that may block does.  That care may lead to farming out
> certain operations to other threads, where they may block without harm.
> 
> You only talk about "the non-main thread consumers".  What about the
> main thread?  Is it okay for the main thread to block?  If yes, why?

It would be great if the main thread never blocked; but IMHO that's
a huge task that we'll never get done [challenge].

> >> We can have any number of QMP monitors today.  Would each of them feed
> >> its own queue?  Would they all feed a shared queue?
> >
> > Currently with multiple QMP monitors, everything runs in the main
> > loop, so commands arriving across  multiple monitors are 100%
> > serialized and processed strictly in the order in which QEMU reads
> > them off the wire.  To maintain these semantics, we would need to
> > have a single shared queue for the default main loop consumer, so
> > that ordering does not change.
> >
> >> How exactly is opt-in asynchronous to work?  Per QMP monitor?  Per
> >> command?
> >
> > Per monitor+command. ie just because libvirt knows how to cope with
> > async execution on the monitor it has open, does not mean that a
> > different app on the 2nd monitor command can cope. So in my proposal
> > the switch to async must be scoped to the particular command only
> > for the monitor connection that requesteed it.
> >
> >> What does it mean when an asynchronous command follows a synchronous
> >> command in the same QMP monitor?  I would expect the synchronous command
> >> to complete before the asynchronous command, because that's what
> >> synchronous means, isn't it?  To keep your QMP monitor available, you
> >> then must not send synchronous commands that can hang.
> >
> > No, that is not what I described. All synchronous commands are
> > serialized wrt each other, just as today. An asychronous command
> > can run as soon as it is received, regardless of whether any
> > earlier sent sync commands are still executing or pending. This
> > is trivial to achieve when you separate monitor I/O from command
> > execution in separate threads, provided of course the async
> > command consumers are not in the main loop.
> 
> So, a synchronous command is synchronous with respect to other commands,
> except for certain non-blocking commands.  The distinctive feature of
> the latter isn't so much an asynchronous reply, but out-of-band
> dispatch.
> 
> Out-of-band dispatch of commands that cannot block in fact orthogonal to
> asynchronous replies.  I can't see why out-of-band dispatch of
> synchronous non-blocking commands wouldn't work, too.
> 
> >> How can we determine whether a certain synchronous command can hang?
> >> Note that with opt-in async, *all* commands are also synchronous
> >> commands.
> >> 
> >> In short, explain to me how exactly you plan to ensure that certain QMP
> >> commands (such as post-copy recovery) can always "get through", in the
> >> presence of multiple monitors, hanging main loop, hanging synchronous
> >> commands, hanging whatever-else-can-now-hang-in-this-post-copy-world.
> >
> > Taking migrate-cancel as the example. The migration code already has
> > a background thread doing work independantly onthe main loop. Upon
> > marking the migrate-cancel command as async, the migration control
> > thread would become the consumer of migrate-cancel.
> 
> From 30,000 feet, the QMP monitor sends a "cancel" message to the
> migration thread, and later receives a "canceled" message from the
> migration thread.
> 
> From 300 feet, we use the migrate-cancel QMP command as the cancel
> message, and its success response as the "canceled" message.
> 
> In other words, we're pressing the external QM-Protocol into service as
> internal message passing protocol.

Be careful; it's not a cancel in the postcopy recovery case, it's a
restart.  The command is very much like the migration-incoming command.
The management layer has to provide data with the request, so it's not
an internal command.

> >                                                     This allows the
> > migration operation to be cancelled immediately, regardless of whether
> > there are earlier monitor commands blocked in the main loop.
> 
> The necessary part is moving all operations that can block out of
> whatever loop runs the monitor, be it the main loop, some other event
> loop, or a dedicated monitor thread's monitor loop.
> 
> Moving out non-blocking operations isn't necessary.  migrate-cancel
> could communicate with the migration thread by any suitable mechanism or
> protocol.  It doesn't have to be QMP.  Why would we want it to be QMP?

Because why invent another wheel?
This is a command that the management layer has to issue to qemu for
it to recover, including passing data, in a way similar to other
commands - so it looks like a QMP command, so why not use QMP.

Also, I think making other commands lock-free is advantageous - 
some of the 'info' commands just dont really need locks, making them
not use locks removes latency effects caused by the management layer
prodding qemu.

> > Of course this assumes the migration control thread can't block
> > for locks held by the main thread.
> 
> Thanks for your answers, they help.
> 
> >> Now let's talk about QMP requirements.
> >> 
> >> Any addition to QMP must consider what exists already.
> >> 
> >> You may add more of the same.
> >> 
> >> You may generalize existing stuff.
> >> 
> >> You may change existing stuff if you have sufficient reason, subject to
> >> backward compatibility constraints.
> >> 
> >> But attempts to add new ways to do the same old stuff without properly
> >> integrating the existing ways are not going to fly.
> >> 
> >> In particular, any new way to start some job, monitor and control it
> >> while it lives, get notified about its state changes and so forth must
> >> integrate the existing ways.  These include block jobs (probably the
> >> most sophisticated of the lot), migration, dump-guest-memory, and
> >> possibly more.  They all work the same way: synchronous command to kick
> >> off the job, more synchronous commands to monitor and control, events to
> >> notify.  They do differ in detail.
> >> 
> >> Asynchronous commands are a new way to do this.  When you only need to
> >> be notified on "done", and don't need to monitor / control, they fit the
> >> bill quite neatly.
> >> 
> >> However, we can't just ignore the cases where we need more than that!
> >> For those, we want a single generic solution instead of the several ad
> >> hoc solutions we have now.
> >> 
> >> If we add asynchronous commands *now*, and for simple cases only, we add
> >> yet another special case for a future generic solution to integrate.
> >> I'm not going to let that happen.
> >
> > With the async commands suggestion, while it would initially not
> > provide a way to query incremental status, that could easily be
> > fitted in.
> 
> This is [*] below.
> 
> >             Because command replies from async commands may be
> > out-of-order wrt the original requests, clients would need to
> > provide a unique ID for each command run. This originally was
> > part of QMP spec but then dropped, but libvirt still actually
> > generates a uniqe ID for every QMP command.
> >
> > Given this, one option is to actually use the QMP command ID as
> > a job ID, and let you query ongoing status via some new QMP
> > command that accepts the ID of the job to be queried. A complexity
> > with this is how to make the jobs visible across multiple QMP
> > monitors. The job ID might actually have to be a combination of
> > the serial ID from the QMP command, and the ID of the monitor
> > chardev combined.
> 
> Yes.  The job ID must be unique across all QMP monitors to make
> broadcast notifications work.
> 
> >> I figure the closest to a generic solution we have is block jobs.
> >> Perhaps a generic solution could be had by abstracting away the "block"
> >> from "block jobs", leaving just "jobs".
> 
> [*] starts here:
> 
> >> Another approach is generalizing the asynchronous command proposal to
> >> fully cover the not-so-simple cases.
> 
> We know asynchronous commands "fully cover" when we can use them to
> replace all the existing job-like commands.
> 
> Until then, they enlarge rather than solve our jobs problem.
> 
> I get the need for an available monitor.  But I need to balance it with
> other needs.  Can we find a solution for our monitor availability
> problem that doesn't enlarge our jobs problem?

Hopefully!

Dave

> >> If you'd rather want to make progress on monitor availability without
> >> cracking the "jobs" problem, you're in luck!  Use your license to "add
> >> more of the same": synchronous command to start a job, query to monitor,
> >> event to notify.  
> >> 
> >> If you insist on tying your monitor availability solution to
> >> asynchronous commands, then I'm in luck!  I just found volunteers to
> >> solve the "jobs" problem for me.
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

"Dr. David Alan Gilbert" <dgilbert@redhat.com> writes:

> * Markus Armbruster (armbru@redhat.com) wrote:
>> "Daniel P. Berrange" <berrange@redhat.com> writes:
>> 
>> > On Thu, Sep 07, 2017 at 02:59:28PM +0200, Markus Armbruster wrote:
>> >> So, what exactly is going to drain the command queue?  If there's more
>> >> than one consumer, how exactly are commands from the queue dispatched to
>> >> the consumers?
>> >
>> > In terms of my proposal, for any single command there should only ever
>> > be a single consumer. The default consumer would be the main event loop
>> > thread, such that we have no semantic change to QMP operation from today.
>> >
>> > Some commands that are capable of being made "async", would have a
>> > different consumer. For example, if the client requested the 'migrate-cancel'
>> > be made async, this would change things such that the migration thread is
>> > now responsible for consuming the "migrate-cancel" command, instead of the
>> > default main loop.
>> >
>> >> What are the "no hang" guarantees (if any) and conditions for each of
>> >> these consumers?
>> >
>> > The non-main thread consumers would have to have some reasonable
>> > guarantee that they won't block on a lock held by the main loop,
>> > otherwise the whole feature is largely useless.
>> 
>> Same if they block indefinitely on anything else, actually.  In other
>> words, we need to talk about liveness.
>> 
>> Threads by themselves don't buy us liveness.  Being careful with
>> operations that may block does.  That care may lead to farming out
>> certain operations to other threads, where they may block without harm.
>> 
>> You only talk about "the non-main thread consumers".  What about the
>> main thread?  Is it okay for the main thread to block?  If yes, why?
>
> It would be great if the main thread never blocked; but IMHO that's
> a huge task that we'll never get done [challenge].

This is perhaps starting to wander off the topic, but here goes anyway.

What unpleasant things can happen when the main loop hangs?

What are the known causes of main loop hangs?  Any ideas on fixing them?

Are the unknown main loop hangs relevant in practice?

If we can't eliminate main loop hangs, any ideas on reducing their
impact?

>> >> We can have any number of QMP monitors today.  Would each of them feed
>> >> its own queue?  Would they all feed a shared queue?
>> >
>> > Currently with multiple QMP monitors, everything runs in the main
>> > loop, so commands arriving across  multiple monitors are 100%
>> > serialized and processed strictly in the order in which QEMU reads
>> > them off the wire.  To maintain these semantics, we would need to
>> > have a single shared queue for the default main loop consumer, so
>> > that ordering does not change.
>> >
>> >> How exactly is opt-in asynchronous to work?  Per QMP monitor?  Per
>> >> command?
>> >
>> > Per monitor+command. ie just because libvirt knows how to cope with
>> > async execution on the monitor it has open, does not mean that a
>> > different app on the 2nd monitor command can cope. So in my proposal
>> > the switch to async must be scoped to the particular command only
>> > for the monitor connection that requesteed it.
>> >
>> >> What does it mean when an asynchronous command follows a synchronous
>> >> command in the same QMP monitor?  I would expect the synchronous command
>> >> to complete before the asynchronous command, because that's what
>> >> synchronous means, isn't it?  To keep your QMP monitor available, you
>> >> then must not send synchronous commands that can hang.
>> >
>> > No, that is not what I described. All synchronous commands are
>> > serialized wrt each other, just as today. An asychronous command
>> > can run as soon as it is received, regardless of whether any
>> > earlier sent sync commands are still executing or pending. This
>> > is trivial to achieve when you separate monitor I/O from command
>> > execution in separate threads, provided of course the async
>> > command consumers are not in the main loop.
>> 
>> So, a synchronous command is synchronous with respect to other commands,
>> except for certain non-blocking commands.  The distinctive feature of
>> the latter isn't so much an asynchronous reply, but out-of-band
>> dispatch.
>> 
>> Out-of-band dispatch of commands that cannot block in fact orthogonal to
>> asynchronous replies.  I can't see why out-of-band dispatch of
>> synchronous non-blocking commands wouldn't work, too.
>> 
>> >> How can we determine whether a certain synchronous command can hang?
>> >> Note that with opt-in async, *all* commands are also synchronous
>> >> commands.
>> >> 
>> >> In short, explain to me how exactly you plan to ensure that certain QMP
>> >> commands (such as post-copy recovery) can always "get through", in the
>> >> presence of multiple monitors, hanging main loop, hanging synchronous
>> >> commands, hanging whatever-else-can-now-hang-in-this-post-copy-world.
>> >
>> > Taking migrate-cancel as the example. The migration code already has
>> > a background thread doing work independantly onthe main loop. Upon
>> > marking the migrate-cancel command as async, the migration control
>> > thread would become the consumer of migrate-cancel.
>> 
>> From 30,000 feet, the QMP monitor sends a "cancel" message to the
>> migration thread, and later receives a "canceled" message from the
>> migration thread.
>> 
>> From 300 feet, we use the migrate-cancel QMP command as the cancel
>> message, and its success response as the "canceled" message.
>> 
>> In other words, we're pressing the external QM-Protocol into service as
>> internal message passing protocol.
>
> Be careful; it's not a cancel in the postcopy recovery case, it's a
> restart.  The command is very much like the migration-incoming command.
> The management layer has to provide data with the request, so it's not
> an internal command.

It's still a message.

>> >                                                     This allows the
>> > migration operation to be cancelled immediately, regardless of whether
>> > there are earlier monitor commands blocked in the main loop.
>> 
>> The necessary part is moving all operations that can block out of
>> whatever loop runs the monitor, be it the main loop, some other event
>> loop, or a dedicated monitor thread's monitor loop.
>> 
>> Moving out non-blocking operations isn't necessary.  migrate-cancel
>> could communicate with the migration thread by any suitable mechanism or
>> protocol.  It doesn't have to be QMP.  Why would we want it to be QMP?
>
> Because why invent another wheel?
> This is a command that the management layer has to issue to qemu for
> it to recover, including passing data, in a way similar to other
> commands - so it looks like a QMP command, so why not use QMP.

Point taken.

Minor terminology remark: I'd prefer to call this a reuse of QAPI rather
than QMP, because QMP makes me think of sockets and JSON, while QAPI
makes me think of generated data types and marshaling code.

> Also, I think making other commands lock-free is advantageous - 
> some of the 'info' commands just dont really need locks, making them
> not use locks removes latency effects caused by the management layer
> prodding qemu.

I get the desire to move commands that can block out of whatever loop
runs the monitor.  But moving out commands that always complete quickly
seems pointless: by the time you're done queuing them, you could be done
*executing* them.  More on that below.

>> > Of course this assumes the migration control thread can't block
>> > for locks held by the main thread.
>> 
>> Thanks for your answers, they help.
>> 
>> >> Now let's talk about QMP requirements.
>> >> 
>> >> Any addition to QMP must consider what exists already.
>> >> 
>> >> You may add more of the same.
>> >> 
>> >> You may generalize existing stuff.
>> >> 
>> >> You may change existing stuff if you have sufficient reason, subject to
>> >> backward compatibility constraints.
>> >> 
>> >> But attempts to add new ways to do the same old stuff without properly
>> >> integrating the existing ways are not going to fly.
>> >> 
>> >> In particular, any new way to start some job, monitor and control it
>> >> while it lives, get notified about its state changes and so forth must
>> >> integrate the existing ways.  These include block jobs (probably the
>> >> most sophisticated of the lot), migration, dump-guest-memory, and
>> >> possibly more.  They all work the same way: synchronous command to kick
>> >> off the job, more synchronous commands to monitor and control, events to
>> >> notify.  They do differ in detail.
>> >> 
>> >> Asynchronous commands are a new way to do this.  When you only need to
>> >> be notified on "done", and don't need to monitor / control, they fit the
>> >> bill quite neatly.
>> >> 
>> >> However, we can't just ignore the cases where we need more than that!
>> >> For those, we want a single generic solution instead of the several ad
>> >> hoc solutions we have now.
>> >> 
>> >> If we add asynchronous commands *now*, and for simple cases only, we add
>> >> yet another special case for a future generic solution to integrate.
>> >> I'm not going to let that happen.
>> >
>> > With the async commands suggestion, while it would initially not
>> > provide a way to query incremental status, that could easily be
>> > fitted in.
>> 
>> This is [*] below.
>> 
>> >             Because command replies from async commands may be
>> > out-of-order wrt the original requests, clients would need to
>> > provide a unique ID for each command run. This originally was
>> > part of QMP spec but then dropped, but libvirt still actually
>> > generates a uniqe ID for every QMP command.
>> >
>> > Given this, one option is to actually use the QMP command ID as
>> > a job ID, and let you query ongoing status via some new QMP
>> > command that accepts the ID of the job to be queried. A complexity
>> > with this is how to make the jobs visible across multiple QMP
>> > monitors. The job ID might actually have to be a combination of
>> > the serial ID from the QMP command, and the ID of the monitor
>> > chardev combined.
>> 
>> Yes.  The job ID must be unique across all QMP monitors to make
>> broadcast notifications work.
>> 
>> >> I figure the closest to a generic solution we have is block jobs.
>> >> Perhaps a generic solution could be had by abstracting away the "block"
>> >> from "block jobs", leaving just "jobs".
>> 
>> [*] starts here:
>> 
>> >> Another approach is generalizing the asynchronous command proposal to
>> >> fully cover the not-so-simple cases.
>> 
>> We know asynchronous commands "fully cover" when we can use them to
>> replace all the existing job-like commands.
>> 
>> Until then, they enlarge rather than solve our jobs problem.
>> 
>> I get the need for an available monitor.  But I need to balance it with
>> other needs.  Can we find a solution for our monitor availability
>> problem that doesn't enlarge our jobs problem?
>
> Hopefully!
>
> Dave
>
>> >> If you'd rather want to make progress on monitor availability without
>> >> cracking the "jobs" problem, you're in luck!  Use your license to "add
>> >> more of the same": synchronous command to start a job, query to monitor,
>> >> event to notify.  
>> >> 
>> >> If you insist on tying your monitor availability solution to
>> >> asynchronous commands, then I'm in luck!  I just found volunteers to
>> >> solve the "jobs" problem for me.

Let me try to distill the discussion so far into a design sketch.

1. A QMP monitor runs in a loop.  The loop may execute other stuff, but
   this must not unduly delay the monitor's work.  Thus, everything in
   this loop must complete "quickly".

   All QMP monitors currently run in the main loop, which really should
   satisfy "quickly", but doesn't.  Since fixing that to a tolerable
   degree is beyond our means (is it?), we move them out.

   Design alternative: either one loop and thread per monitor, or one
   loop and thread for all monitors, or something in between.

   I'm wary of "one thread per software artifact" designs.  "One
   (preemptable) thread per activity, all sharing state" is a lousy way
   to structure software.

2. A QMP monitor receives and dispatches commands, and sends command
   responses and events.

   What if sending a response or event would block?  See 6.

3. Arbitrary code can trigger QMP events.  Events are broadcast to all
   QMP monitors.  Each QMP monitor provides an event queue.  When an
   event is triggered, it gets put into all queues, subject to rate
   limiting.

   Rate limiting and queuing needs some shared data, which is protected
   by a mutex.  The critical sections guarded by this mutex must be
   "quick".

   Nothing new here, it's how events work today.

   We could easily add events that go to just one monitor, if there's a
   need.

4. Commands are normally dispatched to a worker thread, where they can
   take their own sweet time to complete.

   Currently, the monitor runs in the main loop and executes commands
   directly.  This is effectively dispatching commands to the main loop.
   Dispatch to main loop is wrong, because it can make the main loop
   hang.  If it was the only relevant cause for main loop hangs, we'd
   move the command work out and be done.  Since it isn't (see 1.) we
   *also* have to move the monitor out to prevent main loop hangs from
   hanging the monitor.

   Moving monitor and command work to separate threads changes the
   dispatch from function call to queuing.  Need a pair of queues, one
   for commands, one for responses.

   Design alternative: one worker per monitor, or one worker for all
   monitors, or main loop is the one worker for all monitors.  The
   latter leaves the main loop hangs unaddressed.  No worse than before,
   so it could be okay as a first step.

   The worker provides the pair of queues.  It executes commands in
   order.  If a command blocks, the command queue stalls.

   The command queue can therefore grow without bounds.  See 6.

5. Certain commands that always complete "quickly" can instead be
   executed directly, at the QMP client's request.  This direct
   execution is out-of-band; the response can "overtake" prior in-band
   commands' responses.

   The amount of work these out-of-band commands do themselves is up to
   them.  A qiuck query command would do all the work.  migrate-cancel
   could perhaps notify the migration thread and be done.  Postcopy
   recovery could perhaps send its argument struct to whatever thread
   handles recovery.

6. Flow control

   We currently leave flow control to the underlying character device.
   If the client sends more quickly than the monitor can execute, the
   client's send eventually blocks or fails with EAGAIN.  If the monitor
   sends more quickly than the client accepts, the monitor buffers
   without bounds (I think).

   Buffering monitor output without bounds is bad.  We could perhaps
   kill a monitor when it exceeds its limit.

   Buffering monitor input (in the command queue) without bounds is just
   as bad.  It also destroys the existing flow control mechanism: the
   client can no longer detect that it's sending too much.  Not an issue
   for fully synchronous clients, i.e. clients that wait for the
   previous command's response before they send the next command.  Such
   clients cannot use of out-of-band command execution.

   The obvious way to limit the command queue is to fail commands when
   the queue is "full".

   Note that we can't send an error response right away then, because
   the command is in-band (if it wasn't, we wouldn't queue it), so its
   response has to go after all all the respones to the (in-band)
   commands currently in the queue.

   To tell the client right away, we could send an event.

   Delaying the "queue full" response until the correct time to send it
   requires state: at least the command ID.  We can just as well enqueue
   and pray memory will suffice.

   Note that the only reason for the command queue is out-of-band
   commands.  Without them, reading the next command is pointless.  This
   leads me to a possible solution: separate out-of-band mode, default
   off, QMP client can switch it on.  When off, we read monitor input
   just like we do now (no queue, no problem).  When on, we read and
   queue.  If the queue is full, we send a "queue full" event with the
   IDs of the commands we dropped on the floor.  By switching on
   out-of-band-mode, the QMP client also opts into this event.

   Switching could be done with QMP capabilities negotiation.

7. How all this is related to "jobs"

   Out-of-band execution is a limited special case of asynchronous
   execution.  With general asynchronous execution, responses can be
   sent in any order.  With out-of-band execution, only the out-of-band
   responses can "jump" order, and only over in-band responses.

   "All commands are (to be treated as) asynchronous" is arguably more
   elegant than this out-of-band thing.  However, it runs into two
   roadblocks that don't apply to out-of-band.

   One, backward compatibility.  That's a roadblock only as much as we
   make it one.

   Two, consistency.  "All asynchronous, but we do most job-like things
   with commands + events anyway" is not acceptable to me.  I'd be
   willing to accept "all asynchronous" when it solves the jobs problem.

   You asked for a solution to the monitor availability problem that
   doesn't require you to solve the jobs problem first.  Well, here's my
   best try.  Go shoot some holes into it :)

* Markus Armbruster (armbru@redhat.com) wrote:
> "Dr. David Alan Gilbert" <dgilbert@redhat.com> writes:
> 
> > * Markus Armbruster (armbru@redhat.com) wrote:
> >> "Daniel P. Berrange" <berrange@redhat.com> writes:
> >> 
> >> > On Thu, Sep 07, 2017 at 02:59:28PM +0200, Markus Armbruster wrote:
> >> >> So, what exactly is going to drain the command queue?  If there's more
> >> >> than one consumer, how exactly are commands from the queue dispatched to
> >> >> the consumers?
> >> >
> >> > In terms of my proposal, for any single command there should only ever
> >> > be a single consumer. The default consumer would be the main event loop
> >> > thread, such that we have no semantic change to QMP operation from today.
> >> >
> >> > Some commands that are capable of being made "async", would have a
> >> > different consumer. For example, if the client requested the 'migrate-cancel'
> >> > be made async, this would change things such that the migration thread is
> >> > now responsible for consuming the "migrate-cancel" command, instead of the
> >> > default main loop.
> >> >
> >> >> What are the "no hang" guarantees (if any) and conditions for each of
> >> >> these consumers?
> >> >
> >> > The non-main thread consumers would have to have some reasonable
> >> > guarantee that they won't block on a lock held by the main loop,
> >> > otherwise the whole feature is largely useless.
> >> 
> >> Same if they block indefinitely on anything else, actually.  In other
> >> words, we need to talk about liveness.
> >> 
> >> Threads by themselves don't buy us liveness.  Being careful with
> >> operations that may block does.  That care may lead to farming out
> >> certain operations to other threads, where they may block without harm.
> >> 
> >> You only talk about "the non-main thread consumers".  What about the
> >> main thread?  Is it okay for the main thread to block?  If yes, why?
> >
> > It would be great if the main thread never blocked; but IMHO that's
> > a huge task that we'll never get done [challenge].
> 
> This is perhaps starting to wander off the topic, but here goes anyway.
> 
> What unpleasant things can happen when the main loop hangs?

  a) We can't interact with the monitor to fix the cause of the hang
     (Which is my main interest here)
  b) IO emulation might also be blocked because it's waiting on the bql

> What are the known causes of main loop hangs?  Any ideas on fixing them?

  c) hangs on networking while under BQL; there's at least one case near
  the end of migrate
  d) hangs on storage devices while under BQL; I think there are similar
  cases in migrate and possibly elsewhere
  e) postcopy pages not yet arrived - then a problem if the postcopy
  dies and needs recovery (because of a)

> Are the unknown main loop hangs relevant in practice?

Well, the unknown ones are unknown; the known ones however seem
relevant:
  f) I can't recover a failed postcopy
  g) A COLO synchronisation might hang at a bad point in migrate and
  you can't kill it off to cause one side to continue
  h) A failure of networking at just the wrong point in migrate can
  cause the source to be paused for a long time - but I don't think
  I've seen it in practice.

> If we can't eliminate main loop hangs, any ideas on reducing their
> impact?

Note there's two related things; main loop hangs and bql hangs; I'm not
sure that the two are always the same.

Stefan mentioned some ways of doing asynchronous memory lookups/accesses
though I'm not sure they'd work in the postcopy case; but they'd need
work in lots of devices.
Some of the IO under the BQL might be fixable; IMHO in a lot of places
we don't really need the full BQL, we just need a 'you aren't going to
change the config' lock.

> >> >> We can have any number of QMP monitors today.  Would each of them feed
> >> >> its own queue?  Would they all feed a shared queue?
> >> >
> >> > Currently with multiple QMP monitors, everything runs in the main
> >> > loop, so commands arriving across  multiple monitors are 100%
> >> > serialized and processed strictly in the order in which QEMU reads
> >> > them off the wire.  To maintain these semantics, we would need to
> >> > have a single shared queue for the default main loop consumer, so
> >> > that ordering does not change.
> >> >
> >> >> How exactly is opt-in asynchronous to work?  Per QMP monitor?  Per
> >> >> command?
> >> >
> >> > Per monitor+command. ie just because libvirt knows how to cope with
> >> > async execution on the monitor it has open, does not mean that a
> >> > different app on the 2nd monitor command can cope. So in my proposal
> >> > the switch to async must be scoped to the particular command only
> >> > for the monitor connection that requesteed it.
> >> >
> >> >> What does it mean when an asynchronous command follows a synchronous
> >> >> command in the same QMP monitor?  I would expect the synchronous command
> >> >> to complete before the asynchronous command, because that's what
> >> >> synchronous means, isn't it?  To keep your QMP monitor available, you
> >> >> then must not send synchronous commands that can hang.
> >> >
> >> > No, that is not what I described. All synchronous commands are
> >> > serialized wrt each other, just as today. An asychronous command
> >> > can run as soon as it is received, regardless of whether any
> >> > earlier sent sync commands are still executing or pending. This
> >> > is trivial to achieve when you separate monitor I/O from command
> >> > execution in separate threads, provided of course the async
> >> > command consumers are not in the main loop.
> >> 
> >> So, a synchronous command is synchronous with respect to other commands,
> >> except for certain non-blocking commands.  The distinctive feature of
> >> the latter isn't so much an asynchronous reply, but out-of-band
> >> dispatch.
> >> 
> >> Out-of-band dispatch of commands that cannot block in fact orthogonal to
> >> asynchronous replies.  I can't see why out-of-band dispatch of
> >> synchronous non-blocking commands wouldn't work, too.
> >> 
> >> >> How can we determine whether a certain synchronous command can hang?
> >> >> Note that with opt-in async, *all* commands are also synchronous
> >> >> commands.
> >> >> 
> >> >> In short, explain to me how exactly you plan to ensure that certain QMP
> >> >> commands (such as post-copy recovery) can always "get through", in the
> >> >> presence of multiple monitors, hanging main loop, hanging synchronous
> >> >> commands, hanging whatever-else-can-now-hang-in-this-post-copy-world.
> >> >
> >> > Taking migrate-cancel as the example. The migration code already has
> >> > a background thread doing work independantly onthe main loop. Upon
> >> > marking the migrate-cancel command as async, the migration control
> >> > thread would become the consumer of migrate-cancel.
> >> 
> >> From 30,000 feet, the QMP monitor sends a "cancel" message to the
> >> migration thread, and later receives a "canceled" message from the
> >> migration thread.
> >> 
> >> From 300 feet, we use the migrate-cancel QMP command as the cancel
> >> message, and its success response as the "canceled" message.
> >> 
> >> In other words, we're pressing the external QM-Protocol into service as
> >> internal message passing protocol.
> >
> > Be careful; it's not a cancel in the postcopy recovery case, it's a
> > restart.  The command is very much like the migration-incoming command.
> > The management layer has to provide data with the request, so it's not
> > an internal command.
> 
> It's still a message.
> 
> >> >                                                     This allows the
> >> > migration operation to be cancelled immediately, regardless of whether
> >> > there are earlier monitor commands blocked in the main loop.
> >> 
> >> The necessary part is moving all operations that can block out of
> >> whatever loop runs the monitor, be it the main loop, some other event
> >> loop, or a dedicated monitor thread's monitor loop.
> >> 
> >> Moving out non-blocking operations isn't necessary.  migrate-cancel
> >> could communicate with the migration thread by any suitable mechanism or
> >> protocol.  It doesn't have to be QMP.  Why would we want it to be QMP?
> >
> > Because why invent another wheel?
> > This is a command that the management layer has to issue to qemu for
> > it to recover, including passing data, in a way similar to other
> > commands - so it looks like a QMP command, so why not use QMP.
> 
> Point taken.
> 
> Minor terminology remark: I'd prefer to call this a reuse of QAPI rather
> than QMP, because QMP makes me think of sockets and JSON, while QAPI
> makes me think of generated data types and marshaling code.

Well it's a command that's got to come over the socket from management,
so I'm still thinking sockets and JSON. A lot of the problems you
describe below are more to do with the pain of managing the messages
squeezed through a socket.

> > Also, I think making other commands lock-free is advantageous - 
> > some of the 'info' commands just dont really need locks, making them
> > not use locks removes latency effects caused by the management layer
> > prodding qemu.
> 
> I get the desire to move commands that can block out of whatever loop
> runs the monitor.  But moving out commands that always complete quickly
> seems pointless: by the time you're done queuing them, you could be done
> *executing* them.  More on that below.

My thinking here wasn't about the speed of executing the command, my
interest was more on the performance of the guest/IO - avoiding taking
the BQL would have less impact on IO emulation, as would keeping the
main thread free.

> >> > Of course this assumes the migration control thread can't block
> >> > for locks held by the main thread.
> >> 
> >> Thanks for your answers, they help.
> >> 
> >> >> Now let's talk about QMP requirements.
> >> >> 
> >> >> Any addition to QMP must consider what exists already.
> >> >> 
> >> >> You may add more of the same.
> >> >> 
> >> >> You may generalize existing stuff.
> >> >> 
> >> >> You may change existing stuff if you have sufficient reason, subject to
> >> >> backward compatibility constraints.
> >> >> 
> >> >> But attempts to add new ways to do the same old stuff without properly
> >> >> integrating the existing ways are not going to fly.
> >> >> 
> >> >> In particular, any new way to start some job, monitor and control it
> >> >> while it lives, get notified about its state changes and so forth must
> >> >> integrate the existing ways.  These include block jobs (probably the
> >> >> most sophisticated of the lot), migration, dump-guest-memory, and
> >> >> possibly more.  They all work the same way: synchronous command to kick
> >> >> off the job, more synchronous commands to monitor and control, events to
> >> >> notify.  They do differ in detail.
> >> >> 
> >> >> Asynchronous commands are a new way to do this.  When you only need to
> >> >> be notified on "done", and don't need to monitor / control, they fit the
> >> >> bill quite neatly.
> >> >> 
> >> >> However, we can't just ignore the cases where we need more than that!
> >> >> For those, we want a single generic solution instead of the several ad
> >> >> hoc solutions we have now.
> >> >> 
> >> >> If we add asynchronous commands *now*, and for simple cases only, we add
> >> >> yet another special case for a future generic solution to integrate.
> >> >> I'm not going to let that happen.
> >> >
> >> > With the async commands suggestion, while it would initially not
> >> > provide a way to query incremental status, that could easily be
> >> > fitted in.
> >> 
> >> This is [*] below.
> >> 
> >> >             Because command replies from async commands may be
> >> > out-of-order wrt the original requests, clients would need to
> >> > provide a unique ID for each command run. This originally was
> >> > part of QMP spec but then dropped, but libvirt still actually
> >> > generates a uniqe ID for every QMP command.
> >> >
> >> > Given this, one option is to actually use the QMP command ID as
> >> > a job ID, and let you query ongoing status via some new QMP
> >> > command that accepts the ID of the job to be queried. A complexity
> >> > with this is how to make the jobs visible across multiple QMP
> >> > monitors. The job ID might actually have to be a combination of
> >> > the serial ID from the QMP command, and the ID of the monitor
> >> > chardev combined.
> >> 
> >> Yes.  The job ID must be unique across all QMP monitors to make
> >> broadcast notifications work.
> >> 
> >> >> I figure the closest to a generic solution we have is block jobs.
> >> >> Perhaps a generic solution could be had by abstracting away the "block"
> >> >> from "block jobs", leaving just "jobs".
> >> 
> >> [*] starts here:
> >> 
> >> >> Another approach is generalizing the asynchronous command proposal to
> >> >> fully cover the not-so-simple cases.
> >> 
> >> We know asynchronous commands "fully cover" when we can use them to
> >> replace all the existing job-like commands.
> >> 
> >> Until then, they enlarge rather than solve our jobs problem.
> >> 
> >> I get the need for an available monitor.  But I need to balance it with
> >> other needs.  Can we find a solution for our monitor availability
> >> problem that doesn't enlarge our jobs problem?
> >
> > Hopefully!
> >
> > Dave
> >
> >> >> If you'd rather want to make progress on monitor availability without
> >> >> cracking the "jobs" problem, you're in luck!  Use your license to "add
> >> >> more of the same": synchronous command to start a job, query to monitor,
> >> >> event to notify.  
> >> >> 
> >> >> If you insist on tying your monitor availability solution to
> >> >> asynchronous commands, then I'm in luck!  I just found volunteers to
> >> >> solve the "jobs" problem for me.
> 
> Let me try to distill the discussion so far into a design sketch.
> 
> 1. A QMP monitor runs in a loop.  The loop may execute other stuff, but
>    this must not unduly delay the monitor's work.  Thus, everything in
>    this loop must complete "quickly".
> 
>    All QMP monitors currently run in the main loop, which really should
>    satisfy "quickly", but doesn't.  Since fixing that to a tolerable
>    degree is beyond our means (is it?), we move them out.
> 
>    Design alternative: either one loop and thread per monitor, or one
>    loop and thread for all monitors, or something in between.
> 
>    I'm wary of "one thread per software artifact" designs.  "One
>    (preemptable) thread per activity, all sharing state" is a lousy way
>    to structure software.

Shrug; I've always thought of it as an easy solution unless you'd
get into hundreds of threads, which given the number of monitors, we
wont.

> 2. A QMP monitor receives and dispatches commands, and sends command
>    responses and events.
> 
>    What if sending a response or event would block?  See 6.
> 
> 3. Arbitrary code can trigger QMP events.  Events are broadcast to all
>    QMP monitors.  Each QMP monitor provides an event queue.  When an
>    event is triggered, it gets put into all queues, subject to rate
>    limiting.
> 
>    Rate limiting and queuing needs some shared data, which is protected
>    by a mutex.  The critical sections guarded by this mutex must be
>    "quick".
> 
>    Nothing new here, it's how events work today.
> 
>    We could easily add events that go to just one monitor, if there's a
>    need.

I don't think events could cause a problem here since they're always
outbound - so they could never block inbound commands?

> 4. Commands are normally dispatched to a worker thread, where they can
>    take their own sweet time to complete.
> 
>    Currently, the monitor runs in the main loop and executes commands
>    directly.  This is effectively dispatching commands to the main loop.
>    Dispatch to main loop is wrong, because it can make the main loop
>    hang.  If it was the only relevant cause for main loop hangs, we'd
>    move the command work out and be done.  Since it isn't (see 1.) we
>    *also* have to move the monitor out to prevent main loop hangs from
>    hanging the monitor.
> 
>    Moving monitor and command work to separate threads changes the
>    dispatch from function call to queuing.  Need a pair of queues, one
>    for commands, one for responses.
> 
>    Design alternative: one worker per monitor, or one worker for all
>    monitors, or main loop is the one worker for all monitors.  The
>    latter leaves the main loop hangs unaddressed.  No worse than before,
>    so it could be okay as a first step.
> 
>    The worker provides the pair of queues.  It executes commands in
>    order.  If a command blocks, the command queue stalls.
> 
>    The command queue can therefore grow without bounds.  See 6.
> 
> 5. Certain commands that always complete "quickly" can instead be
>    executed directly, at the QMP client's request.  This direct
>    execution is out-of-band; the response can "overtake" prior in-band
>    commands' responses.
> 
>    The amount of work these out-of-band commands do themselves is up to
>    them.  A qiuck query command would do all the work.  migrate-cancel
>    could perhaps notify the migration thread and be done.  Postcopy
>    recovery could perhaps send its argument struct to whatever thread
>    handles recovery.

Yes.

> 6. Flow control

I think this is potentially the tricky bit!

>    We currently leave flow control to the underlying character device.
>    If the client sends more quickly than the monitor can execute, the
>    client's send eventually blocks or fails with EAGAIN.  If the monitor
>    sends more quickly than the client accepts, the monitor buffers
>    without bounds (I think).
> 
>    Buffering monitor output without bounds is bad.  We could perhaps
>    kill a monitor when it exceeds its limit.

I'm not sure it's possible to define that limit; for example
'query-block' gives a list of information for all devices; there are
people running with 200+ block devices so the output for that would be
huge.

>    Buffering monitor input (in the command queue) without bounds is just
>    as bad.  It also destroys the existing flow control mechanism: the
>    client can no longer detect that it's sending too much.  Not an issue
>    for fully synchronous clients, i.e. clients that wait for the
>    previous command's response before they send the next command.  Such
>    clients cannot use of out-of-band command execution.
> 
>    The obvious way to limit the command queue is to fail commands when
>    the queue is "full".
> 
>    Note that we can't send an error response right away then, because
>    the command is in-band (if it wasn't, we wouldn't queue it), so its
>    response has to go after all all the respones to the (in-band)
>    commands currently in the queue.
> 
>    To tell the client right away, we could send an event.
> 
>    Delaying the "queue full" response until the correct time to send it
>    requires state: at least the command ID.  We can just as well enqueue
>    and pray memory will suffice.
> 
>    Note that the only reason for the command queue is out-of-band
>    commands.  Without them, reading the next command is pointless.  This
>    leads me to a possible solution: separate out-of-band mode, default
>    off, QMP client can switch it on.  When off, we read monitor input
>    just like we do now (no queue, no problem).  When on, we read and
>    queue.  If the queue is full, we send a "queue full" event with the
>    IDs of the commands we dropped on the floor.  By switching on
>    out-of-band-mode, the QMP client also opts into this event.
> 
>    Switching could be done with QMP capabilities negotiation.

I'm not sure how this queue interacts for multiple monitors using the
single IO thread.  It's currently legal for each monitor to send one
command and for that command to be outstanding; so 'queue full' mustn't
happen in that case, because we still want to allow any of the monitors
to issue one of the non-locking commands.
So I think we need 2x 1 entry input queues per monitor; one for normal
command and one for non-locking commands; I think that's different
from what we've previously suggested which is 2 central queues.

> 7. How all this is related to "jobs"
> 
>    Out-of-band execution is a limited special case of asynchronous
>    execution.  With general asynchronous execution, responses can be
>    sent in any order.  With out-of-band execution, only the out-of-band
>    responses can "jump" order, and only over in-band responses.
> 
>    "All commands are (to be treated as) asynchronous" is arguably more
>    elegant than this out-of-band thing.  However, it runs into two
>    roadblocks that don't apply to out-of-band.
> 
>    One, backward compatibility.  That's a roadblock only as much as we
>    make it one.
> 
>    Two, consistency.  "All asynchronous, but we do most job-like things
>    with commands + events anyway" is not acceptable to me.  I'd be
>    willing to accept "all asynchronous" when it solves the jobs problem.

I suspect there are other things that limit making everything
asynchronous; for example commands that currently only expect to be
executing in the main thread; if you wanted to make an existing command
async you'd have to audit it for all the possible places it could hang.

I also see the other problem as keeping the management level
understanding of which commands are asynchronous; Dan's suggestion is
that command where the management layer specifies which commands it
expects to be asynchronous, and qemu responds with which ones actually
are.

>    You asked for a solution to the monitor availability problem that
>    doesn't require you to solve the jobs problem first.  Well, here's my
>    best try.  Go shoot some holes into it :)

Hopefully we're running out of holes.

Dave

--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

"Dr. David Alan Gilbert" <dgilbert@redhat.com> writes:

> * Markus Armbruster (armbru@redhat.com) wrote:
>> "Dr. David Alan Gilbert" <dgilbert@redhat.com> writes:
>> 
>> > * Markus Armbruster (armbru@redhat.com) wrote:
>> >> "Daniel P. Berrange" <berrange@redhat.com> writes:
>> >> 
>> >> > On Thu, Sep 07, 2017 at 02:59:28PM +0200, Markus Armbruster wrote:
>> >> >> So, what exactly is going to drain the command queue?  If there's more
>> >> >> than one consumer, how exactly are commands from the queue dispatched to
>> >> >> the consumers?
>> >> >
>> >> > In terms of my proposal, for any single command there should only ever
>> >> > be a single consumer. The default consumer would be the main event loop
>> >> > thread, such that we have no semantic change to QMP operation from today.
>> >> >
>> >> > Some commands that are capable of being made "async", would have a
>> >> > different consumer. For example, if the client requested the 'migrate-cancel'
>> >> > be made async, this would change things such that the migration thread is
>> >> > now responsible for consuming the "migrate-cancel" command, instead of the
>> >> > default main loop.
>> >> >
>> >> >> What are the "no hang" guarantees (if any) and conditions for each of
>> >> >> these consumers?
>> >> >
>> >> > The non-main thread consumers would have to have some reasonable
>> >> > guarantee that they won't block on a lock held by the main loop,
>> >> > otherwise the whole feature is largely useless.
>> >> 
>> >> Same if they block indefinitely on anything else, actually.  In other
>> >> words, we need to talk about liveness.
>> >> 
>> >> Threads by themselves don't buy us liveness.  Being careful with
>> >> operations that may block does.  That care may lead to farming out
>> >> certain operations to other threads, where they may block without harm.
>> >> 
>> >> You only talk about "the non-main thread consumers".  What about the
>> >> main thread?  Is it okay for the main thread to block?  If yes, why?
>> >
>> > It would be great if the main thread never blocked; but IMHO that's
>> > a huge task that we'll never get done [challenge].
>> 
>> This is perhaps starting to wander off the topic, but here goes anyway.
>> 
>> What unpleasant things can happen when the main loop hangs?
>
>   a) We can't interact with the monitor to fix the cause of the hang
>      (Which is my main interest here)
>   b) IO emulation might also be blocked because it's waiting on the bql

To readers other than Dave: anything else?

>> What are the known causes of main loop hangs?  Any ideas on fixing them?
>
>   c) hangs on networking while under BQL; there's at least one case near
>   the end of migrate
>   d) hangs on storage devices while under BQL; I think there are similar
>   cases in migrate and possibly elsewhere
>   e) postcopy pages not yet arrived - then a problem if the postcopy
>   dies and needs recovery (because of a)

Any ideas on fixing them?

To readers other than Dave: anything else?

>> Are the unknown main loop hangs relevant in practice?
>
> Well, the unknown ones are unknown; the known ones however seem
> relevant:
>   f) I can't recover a failed postcopy
>   g) A COLO synchronisation might hang at a bad point in migrate and
>   you can't kill it off to cause one side to continue
>   h) A failure of networking at just the wrong point in migrate can
>   cause the source to be paused for a long time - but I don't think
>   I've seen it in practice.

I don't doubt your assertion that the known ones are relevant; I assume
you've run into them.

The purpose of my question is to find out how serious a problem the
unknown causes are.  I'm afraid the answer is "we don't know".

>> If we can't eliminate main loop hangs, any ideas on reducing their
>> impact?
>
> Note there's two related things; main loop hangs and bql hangs; I'm not
> sure that the two are always the same.
>
> Stefan mentioned some ways of doing asynchronous memory lookups/accesses
> though I'm not sure they'd work in the postcopy case; but they'd need
> work in lots of devices.
> Some of the IO under the BQL might be fixable; IMHO in a lot of places
> we don't really need the full BQL, we just need a 'you aren't going to
> change the config' lock.

This is all about reducing main loop hangs.  Another one is moving
"slow" code out of the main loop, e.g. monitor commands.

My question was aiming in a slightly different direction, however: given
that the main loop can hang, is there anything we can do to mitigate
known bad consequences of such hangs?

We're actually discussing one thing we can do to mitigate: moving the
monitor core out of the main loop, to keep the monitor available.  Any
other ideas?

>> >> >> We can have any number of QMP monitors today.  Would each of them feed
>> >> >> its own queue?  Would they all feed a shared queue?
>> >> >
>> >> > Currently with multiple QMP monitors, everything runs in the main
>> >> > loop, so commands arriving across  multiple monitors are 100%
>> >> > serialized and processed strictly in the order in which QEMU reads
>> >> > them off the wire.  To maintain these semantics, we would need to
>> >> > have a single shared queue for the default main loop consumer, so
>> >> > that ordering does not change.
>> >> >
>> >> >> How exactly is opt-in asynchronous to work?  Per QMP monitor?  Per
>> >> >> command?
>> >> >
>> >> > Per monitor+command. ie just because libvirt knows how to cope with
>> >> > async execution on the monitor it has open, does not mean that a
>> >> > different app on the 2nd monitor command can cope. So in my proposal
>> >> > the switch to async must be scoped to the particular command only
>> >> > for the monitor connection that requesteed it.
>> >> >
>> >> >> What does it mean when an asynchronous command follows a synchronous
>> >> >> command in the same QMP monitor?  I would expect the synchronous command
>> >> >> to complete before the asynchronous command, because that's what
>> >> >> synchronous means, isn't it?  To keep your QMP monitor available, you
>> >> >> then must not send synchronous commands that can hang.
>> >> >
>> >> > No, that is not what I described. All synchronous commands are
>> >> > serialized wrt each other, just as today. An asychronous command
>> >> > can run as soon as it is received, regardless of whether any
>> >> > earlier sent sync commands are still executing or pending. This
>> >> > is trivial to achieve when you separate monitor I/O from command
>> >> > execution in separate threads, provided of course the async
>> >> > command consumers are not in the main loop.
>> >> 
>> >> So, a synchronous command is synchronous with respect to other commands,
>> >> except for certain non-blocking commands.  The distinctive feature of
>> >> the latter isn't so much an asynchronous reply, but out-of-band
>> >> dispatch.
>> >> 
>> >> Out-of-band dispatch of commands that cannot block in fact orthogonal to
>> >> asynchronous replies.  I can't see why out-of-band dispatch of
>> >> synchronous non-blocking commands wouldn't work, too.
>> >> 
>> >> >> How can we determine whether a certain synchronous command can hang?
>> >> >> Note that with opt-in async, *all* commands are also synchronous
>> >> >> commands.
>> >> >> 
>> >> >> In short, explain to me how exactly you plan to ensure that certain QMP
>> >> >> commands (such as post-copy recovery) can always "get through", in the
>> >> >> presence of multiple monitors, hanging main loop, hanging synchronous
>> >> >> commands, hanging whatever-else-can-now-hang-in-this-post-copy-world.
>> >> >
>> >> > Taking migrate-cancel as the example. The migration code already has
>> >> > a background thread doing work independantly onthe main loop. Upon
>> >> > marking the migrate-cancel command as async, the migration control
>> >> > thread would become the consumer of migrate-cancel.
>> >> 
>> >> From 30,000 feet, the QMP monitor sends a "cancel" message to the
>> >> migration thread, and later receives a "canceled" message from the
>> >> migration thread.
>> >> 
>> >> From 300 feet, we use the migrate-cancel QMP command as the cancel
>> >> message, and its success response as the "canceled" message.
>> >> 
>> >> In other words, we're pressing the external QM-Protocol into service as
>> >> internal message passing protocol.
>> >
>> > Be careful; it's not a cancel in the postcopy recovery case, it's a
>> > restart.  The command is very much like the migration-incoming command.
>> > The management layer has to provide data with the request, so it's not
>> > an internal command.
>> 
>> It's still a message.
>> 
>> >> >                                                     This allows the
>> >> > migration operation to be cancelled immediately, regardless of whether
>> >> > there are earlier monitor commands blocked in the main loop.
>> >> 
>> >> The necessary part is moving all operations that can block out of
>> >> whatever loop runs the monitor, be it the main loop, some other event
>> >> loop, or a dedicated monitor thread's monitor loop.
>> >> 
>> >> Moving out non-blocking operations isn't necessary.  migrate-cancel
>> >> could communicate with the migration thread by any suitable mechanism or
>> >> protocol.  It doesn't have to be QMP.  Why would we want it to be QMP?
>> >
>> > Because why invent another wheel?
>> > This is a command that the management layer has to issue to qemu for
>> > it to recover, including passing data, in a way similar to other
>> > commands - so it looks like a QMP command, so why not use QMP.
>> 
>> Point taken.
>> 
>> Minor terminology remark: I'd prefer to call this a reuse of QAPI rather
>> than QMP, because QMP makes me think of sockets and JSON, while QAPI
>> makes me think of generated data types and marshaling code.
>
> Well it's a command that's got to come over the socket from management,
> so I'm still thinking sockets and JSON. A lot of the problems you
> describe below are more to do with the pain of managing the messages
> squeezed through a socket.
>
>> > Also, I think making other commands lock-free is advantageous - 
>> > some of the 'info' commands just dont really need locks, making them
>> > not use locks removes latency effects caused by the management layer
>> > prodding qemu.
>> 
>> I get the desire to move commands that can block out of whatever loop
>> runs the monitor.  But moving out commands that always complete quickly
>> seems pointless: by the time you're done queuing them, you could be done
>> *executing* them.  More on that below.
>
> My thinking here wasn't about the speed of executing the command, my
> interest was more on the performance of the guest/IO - avoiding taking
> the BQL would have less impact on IO emulation, as would keeping the
> main thread free.
>
>> >> > Of course this assumes the migration control thread can't block
>> >> > for locks held by the main thread.
>> >> 
>> >> Thanks for your answers, they help.
>> >> 
>> >> >> Now let's talk about QMP requirements.
>> >> >> 
>> >> >> Any addition to QMP must consider what exists already.
>> >> >> 
>> >> >> You may add more of the same.
>> >> >> 
>> >> >> You may generalize existing stuff.
>> >> >> 
>> >> >> You may change existing stuff if you have sufficient reason, subject to
>> >> >> backward compatibility constraints.
>> >> >> 
>> >> >> But attempts to add new ways to do the same old stuff without properly
>> >> >> integrating the existing ways are not going to fly.
>> >> >> 
>> >> >> In particular, any new way to start some job, monitor and control it
>> >> >> while it lives, get notified about its state changes and so forth must
>> >> >> integrate the existing ways.  These include block jobs (probably the
>> >> >> most sophisticated of the lot), migration, dump-guest-memory, and
>> >> >> possibly more.  They all work the same way: synchronous command to kick
>> >> >> off the job, more synchronous commands to monitor and control, events to
>> >> >> notify.  They do differ in detail.
>> >> >> 
>> >> >> Asynchronous commands are a new way to do this.  When you only need to
>> >> >> be notified on "done", and don't need to monitor / control, they fit the
>> >> >> bill quite neatly.
>> >> >> 
>> >> >> However, we can't just ignore the cases where we need more than that!
>> >> >> For those, we want a single generic solution instead of the several ad
>> >> >> hoc solutions we have now.
>> >> >> 
>> >> >> If we add asynchronous commands *now*, and for simple cases only, we add
>> >> >> yet another special case for a future generic solution to integrate.
>> >> >> I'm not going to let that happen.
>> >> >
>> >> > With the async commands suggestion, while it would initially not
>> >> > provide a way to query incremental status, that could easily be
>> >> > fitted in.
>> >> 
>> >> This is [*] below.
>> >> 
>> >> >             Because command replies from async commands may be
>> >> > out-of-order wrt the original requests, clients would need to
>> >> > provide a unique ID for each command run. This originally was
>> >> > part of QMP spec but then dropped, but libvirt still actually
>> >> > generates a uniqe ID for every QMP command.
>> >> >
>> >> > Given this, one option is to actually use the QMP command ID as
>> >> > a job ID, and let you query ongoing status via some new QMP
>> >> > command that accepts the ID of the job to be queried. A complexity
>> >> > with this is how to make the jobs visible across multiple QMP
>> >> > monitors. The job ID might actually have to be a combination of
>> >> > the serial ID from the QMP command, and the ID of the monitor
>> >> > chardev combined.
>> >> 
>> >> Yes.  The job ID must be unique across all QMP monitors to make
>> >> broadcast notifications work.
>> >> 
>> >> >> I figure the closest to a generic solution we have is block jobs.
>> >> >> Perhaps a generic solution could be had by abstracting away the "block"
>> >> >> from "block jobs", leaving just "jobs".
>> >> 
>> >> [*] starts here:
>> >> 
>> >> >> Another approach is generalizing the asynchronous command proposal to
>> >> >> fully cover the not-so-simple cases.
>> >> 
>> >> We know asynchronous commands "fully cover" when we can use them to
>> >> replace all the existing job-like commands.
>> >> 
>> >> Until then, they enlarge rather than solve our jobs problem.
>> >> 
>> >> I get the need for an available monitor.  But I need to balance it with
>> >> other needs.  Can we find a solution for our monitor availability
>> >> problem that doesn't enlarge our jobs problem?
>> >
>> > Hopefully!
>> >
>> > Dave
>> >
>> >> >> If you'd rather want to make progress on monitor availability without
>> >> >> cracking the "jobs" problem, you're in luck!  Use your license to "add
>> >> >> more of the same": synchronous command to start a job, query to monitor,
>> >> >> event to notify.  
>> >> >> 
>> >> >> If you insist on tying your monitor availability solution to
>> >> >> asynchronous commands, then I'm in luck!  I just found volunteers to
>> >> >> solve the "jobs" problem for me.
>> 
>> Let me try to distill the discussion so far into a design sketch.
>> 
>> 1. A QMP monitor runs in a loop.  The loop may execute other stuff, but
>>    this must not unduly delay the monitor's work.  Thus, everything in
>>    this loop must complete "quickly".
>> 
>>    All QMP monitors currently run in the main loop, which really should
>>    satisfy "quickly", but doesn't.  Since fixing that to a tolerable
>>    degree is beyond our means (is it?), we move them out.
>> 
>>    Design alternative: either one loop and thread per monitor, or one
>>    loop and thread for all monitors, or something in between.
>> 
>>    I'm wary of "one thread per software artifact" designs.  "One
>>    (preemptable) thread per activity, all sharing state" is a lousy way
>>    to structure software.
>
> Shrug; I've always thought of it as an easy solution unless you'd
> get into hundreds of threads, which given the number of monitors, we
> wont.
>
>> 2. A QMP monitor receives and dispatches commands, and sends command
>>    responses and events.
>> 
>>    What if sending a response or event would block?  See 6.
>> 
>> 3. Arbitrary code can trigger QMP events.  Events are broadcast to all
>>    QMP monitors.  Each QMP monitor provides an event queue.  When an
>>    event is triggered, it gets put into all queues, subject to rate
>>    limiting.

Correction: only events that are rate-limited go through the queue.  The
other bypass it.  This is an optimization.

>>    Rate limiting and queuing needs some shared data, which is protected
>>    by a mutex.  The critical sections guarded by this mutex must be
>>    "quick".

There's another mutex guarding the monitor's output buffer (see 6.),
among other things.

>>    Nothing new here, it's how events work today.
>> 
>>    We could easily add events that go to just one monitor, if there's a
>>    need.
>
> I don't think events could cause a problem here since they're always
> outbound - so they could never block inbound commands?

Events are indeed okay as they are.  I merely wanted to mention that
they don't *have* to broadcast.  The "queue full" event mentioned under
6. probably shouldn't be broadcast.

>> 4. Commands are normally dispatched to a worker thread, where they can
>>    take their own sweet time to complete.
>> 
>>    Currently, the monitor runs in the main loop and executes commands
>>    directly.  This is effectively dispatching commands to the main loop.
>>    Dispatch to main loop is wrong, because it can make the main loop
>>    hang.  If it was the only relevant cause for main loop hangs, we'd
>>    move the command work out and be done.  Since it isn't (see 1.) we
>>    *also* have to move the monitor out to prevent main loop hangs from
>>    hanging the monitor.
>> 
>>    Moving monitor and command work to separate threads changes the
>>    dispatch from function call to queuing.  Need a pair of queues, one
>>    for commands, one for responses.
>> 
>>    Design alternative: one worker per monitor, or one worker for all
>>    monitors, or main loop is the one worker for all monitors.  The
>>    latter leaves the main loop hangs unaddressed.  No worse than before,
>>    so it could be okay as a first step.
>> 
>>    The worker provides the pair of queues.  It executes commands in
>>    order.  If a command blocks, the command queue stalls.
>> 
>>    The command queue can therefore grow without bounds.  See 6.
>> 
>> 5. Certain commands that always complete "quickly" can instead be
>>    executed directly, at the QMP client's request.  This direct
>>    execution is out-of-band; the response can "overtake" prior in-band
>>    commands' responses.
>> 
>>    The amount of work these out-of-band commands do themselves is up to
>>    them.  A qiuck query command would do all the work.  migrate-cancel
>>    could perhaps notify the migration thread and be done.  Postcopy
>>    recovery could perhaps send its argument struct to whatever thread
>>    handles recovery.
>
> Yes.

Message sending needs to be non-blocking.  If the message can't be sent,
the command should fail.  Queuing instead is a problematic idea, because
then you get to deal with the same flow control problems we're
discussing below.

>> 6. Flow control
>
> I think this is potentially the tricky bit!
>
>>    We currently leave flow control to the underlying character device.
>>    If the client sends more quickly than the monitor can execute, the
>>    client's send eventually blocks or fails with EAGAIN.  If the monitor
>>    sends more quickly than the client accepts, the monitor buffers
>>    without bounds (I think).
>> 
>>    Buffering monitor output without bounds is bad.  We could perhaps
>>    kill a monitor when it exceeds its limit.
>
> I'm not sure it's possible to define that limit; for example
> 'query-block' gives a list of information for all devices; there are
> people running with 200+ block devices so the output for that would be
> huge.

For comparison, here are our current input limits:

* Sum of JSON token size: 64MiB
* JSON token count: 2Mi
* JSON nesting depth 1024

The first two limit heap usage, the third limits stack usage.  The first
and the last go back to Anthony (commit 29c75dd).  I added the second
because the first is insufficient (commit df64983).

If we hit these generous limits, surely something has gone haywire.

An output limit of 64MiB should be good for ~100k block devices with
MiBs to spare.  Generous enough for a "if you hit this limit, you're
abusing QMP way too much" argument?  If not, how far left would you like
me to shift the limit?

>>    Buffering monitor input (in the command queue) without bounds is just
>>    as bad.  It also destroys the existing flow control mechanism: the
>>    client can no longer detect that it's sending too much.  Not an issue
>>    for fully synchronous clients, i.e. clients that wait for the
>>    previous command's response before they send the next command.  Such
>>    clients cannot use of out-of-band command execution.
>> 
>>    The obvious way to limit the command queue is to fail commands when
>>    the queue is "full".
>> 
>>    Note that we can't send an error response right away then, because
>>    the command is in-band (if it wasn't, we wouldn't queue it), so its
>>    response has to go after all all the respones to the (in-band)
>>    commands currently in the queue.
>> 
>>    To tell the client right away, we could send an event.
>> 
>>    Delaying the "queue full" response until the correct time to send it
>>    requires state: at least the command ID.  We can just as well enqueue
>>    and pray memory will suffice.
>> 
>>    Note that the only reason for the command queue is out-of-band
>>    commands.  Without them, reading the next command is pointless.  This
>>    leads me to a possible solution: separate out-of-band mode, default
>>    off, QMP client can switch it on.  When off, we read monitor input
>>    just like we do now (no queue, no problem).  When on, we read and
>>    queue.  If the queue is full, we send a "queue full" event with the
>>    IDs of the commands we dropped on the floor.  By switching on
>>    out-of-band-mode, the QMP client also opts into this event.
>> 
>>    Switching could be done with QMP capabilities negotiation.
>
> I'm not sure how this queue interacts for multiple monitors using the
> single IO thread.  It's currently legal for each monitor to send one
> command and for that command to be outstanding; so 'queue full' mustn't
> happen in that case, because we still want to allow any of the monitors
> to issue one of the non-locking commands.

Right, the "queue full" condition must be per monitor, and it must not
apply to in-band commands (which aren't queued anyway).

> So I think we need 2x 1 entry input queues per monitor; one for normal
> command and one for non-locking commands; I think that's different
> from what we've previously suggested which is 2 central queues.

Perhaps I was less than clear under 4., but I meant to propose design
alternatives one shared worker fed by one pair of queues, and one worker
per monitor, each fed by its own pair of queues.  Another alternative
would be one shared worker fed by one pair of queues per monitor.

Pair of queues means one for in-band commands, one for their responses.
There is no queue for out-of-band commands, because out-of-band commands
are not queued.

>> 7. How all this is related to "jobs"
>> 
>>    Out-of-band execution is a limited special case of asynchronous
>>    execution.  With general asynchronous execution, responses can be
>>    sent in any order.  With out-of-band execution, only the out-of-band
>>    responses can "jump" order, and only over in-band responses.
>> 
>>    "All commands are (to be treated as) asynchronous" is arguably more
>>    elegant than this out-of-band thing.  However, it runs into two
>>    roadblocks that don't apply to out-of-band.
>> 
>>    One, backward compatibility.  That's a roadblock only as much as we
>>    make it one.
>> 
>>    Two, consistency.  "All asynchronous, but we do most job-like things
>>    with commands + events anyway" is not acceptable to me.  I'd be
>>    willing to accept "all asynchronous" when it solves the jobs problem.
>
> I suspect there are other things that limit making everything
> asynchronous; for example commands that currently only expect to be
> executing in the main thread; if you wanted to make an existing command
> async you'd have to audit it for all the possible places it could hang.

You're right.

> I also see the other problem as keeping the management level
> understanding of which commands are asynchronous; Dan's suggestion is
> that command where the management layer specifies which commands it
> expects to be asynchronous, and qemu responds with which ones actually
> are.

"Command supports out-of-band dispatch" would be visible in
query-qmp-schema.

Design alternative: either switching on out-of-band mode (see 6.)
switches all out-of-band commands to out-of-band dispatch, or it
doesn't, and the client has to request out-of-band dispatch explicitly.
The explicit request could either be per execute (say send {'exec-oob':
COMMAND-NAME ...} instead of {'execute': COMMAND-NAME...}), or per
session, i.e. with a new command to enable oob dispatch for a list of
oob-capable commands.

I figure explicit is safer, because it lets us make more commands
oob-capable without upsetting existing oob-aware QMP clients.

>>    You asked for a solution to the monitor availability problem that
>>    doesn't require you to solve the jobs problem first.  Well, here's my
>>    best try.  Go shoot some holes into it :)
>
> Hopefully we're running out of holes.

Thanks!

On Fri, Sep 8, 2017 at 12:49 PM, Markus Armbruster <armbru@redhat.com> wrote:
> "Dr. David Alan Gilbert" <dgilbert@redhat.com> writes:
>
>> * Markus Armbruster (armbru@redhat.com) wrote:
>>> "Dr. David Alan Gilbert" <dgilbert@redhat.com> writes:
>>>
>>> > * Markus Armbruster (armbru@redhat.com) wrote:
>>> >> "Daniel P. Berrange" <berrange@redhat.com> writes:
>>> >>
>>> >> > On Thu, Sep 07, 2017 at 02:59:28PM +0200, Markus Armbruster wrote:
>>> If we can't eliminate main loop hangs, any ideas on reducing their
>>> impact?
>>
>> Note there's two related things; main loop hangs and bql hangs; I'm not
>> sure that the two are always the same.
>>
>> Stefan mentioned some ways of doing asynchronous memory lookups/accesses
>> though I'm not sure they'd work in the postcopy case; but they'd need
>> work in lots of devices.
>> Some of the IO under the BQL might be fixable; IMHO in a lot of places
>> we don't really need the full BQL, we just need a 'you aren't going to
>> change the config' lock.
>
> This is all about reducing main loop hangs.  Another one is moving
> "slow" code out of the main loop, e.g. monitor commands.
>
> My question was aiming in a slightly different direction, however: given
> that the main loop can hang, is there anything we can do to mitigate
> known bad consequences of such hangs?

I don't think we can mitigate it completely but we can make it visible
and easier to study.

There were discussions about making the event loop observable in the
past.  In other words, logging which handler functions are firing.
That way you can debug scenarios where the loop is spinning
("main-loop: WARNING: I/O thread spun for 1000 iterations\n") and also
latency.  Collecting event handler latencies and looking at the
histogram would be interesting.  The outliers (e.g. 250+ microseconds)
are things that we should know about and consider refactoring.

On Fri, Sep 08, 2017 at 01:49:41PM +0200, Markus Armbruster wrote:
> "Dr. David Alan Gilbert" <dgilbert@redhat.com> writes:
> 
> > * Markus Armbruster (armbru@redhat.com) wrote:
> >> "Dr. David Alan Gilbert" <dgilbert@redhat.com> writes:
> >> 
> >> > * Markus Armbruster (armbru@redhat.com) wrote:
> >> >> "Daniel P. Berrange" <berrange@redhat.com> writes:
> >> >> 
> >> >> > On Thu, Sep 07, 2017 at 02:59:28PM +0200, Markus Armbruster wrote:
> >> >> >> So, what exactly is going to drain the command queue?  If there's more
> >> >> >> than one consumer, how exactly are commands from the queue dispatched to
> >> >> >> the consumers?
> >> >> >
> >> >> > In terms of my proposal, for any single command there should only ever
> >> >> > be a single consumer. The default consumer would be the main event loop
> >> >> > thread, such that we have no semantic change to QMP operation from today.
> >> >> >
> >> >> > Some commands that are capable of being made "async", would have a
> >> >> > different consumer. For example, if the client requested the 'migrate-cancel'
> >> >> > be made async, this would change things such that the migration thread is
> >> >> > now responsible for consuming the "migrate-cancel" command, instead of the
> >> >> > default main loop.
> >> >> >
> >> >> >> What are the "no hang" guarantees (if any) and conditions for each of
> >> >> >> these consumers?
> >> >> >
> >> >> > The non-main thread consumers would have to have some reasonable
> >> >> > guarantee that they won't block on a lock held by the main loop,
> >> >> > otherwise the whole feature is largely useless.
> >> >> 
> >> >> Same if they block indefinitely on anything else, actually.  In other
> >> >> words, we need to talk about liveness.
> >> >> 
> >> >> Threads by themselves don't buy us liveness.  Being careful with
> >> >> operations that may block does.  That care may lead to farming out
> >> >> certain operations to other threads, where they may block without harm.
> >> >> 
> >> >> You only talk about "the non-main thread consumers".  What about the
> >> >> main thread?  Is it okay for the main thread to block?  If yes, why?
> >> >
> >> > It would be great if the main thread never blocked; but IMHO that's
> >> > a huge task that we'll never get done [challenge].
> >> 
> >> This is perhaps starting to wander off the topic, but here goes anyway.
> >> 
> >> What unpleasant things can happen when the main loop hangs?
> >
> >   a) We can't interact with the monitor to fix the cause of the hang
> >      (Which is my main interest here)
> >   b) IO emulation might also be blocked because it's waiting on the bql
> 
> To readers other than Dave: anything else?
> 
> >> What are the known causes of main loop hangs?  Any ideas on fixing them?
> >
> >   c) hangs on networking while under BQL; there's at least one case near
> >   the end of migrate
> >   d) hangs on storage devices while under BQL; I think there are similar
> >   cases in migrate and possibly elsewhere
> >   e) postcopy pages not yet arrived - then a problem if the postcopy
> >   dies and needs recovery (because of a)
> 
> Any ideas on fixing them?
> 
> To readers other than Dave: anything else?
> 
> >> Are the unknown main loop hangs relevant in practice?
> >
> > Well, the unknown ones are unknown; the known ones however seem
> > relevant:
> >   f) I can't recover a failed postcopy
> >   g) A COLO synchronisation might hang at a bad point in migrate and
> >   you can't kill it off to cause one side to continue
> >   h) A failure of networking at just the wrong point in migrate can
> >   cause the source to be paused for a long time - but I don't think
> >   I've seen it in practice.
> 
> I don't doubt your assertion that the known ones are relevant; I assume
> you've run into them.
> 
> The purpose of my question is to find out how serious a problem the
> unknown causes are.  I'm afraid the answer is "we don't know".
> 
> >> If we can't eliminate main loop hangs, any ideas on reducing their
> >> impact?
> >
> > Note there's two related things; main loop hangs and bql hangs; I'm not
> > sure that the two are always the same.
> >
> > Stefan mentioned some ways of doing asynchronous memory lookups/accesses
> > though I'm not sure they'd work in the postcopy case; but they'd need
> > work in lots of devices.
> > Some of the IO under the BQL might be fixable; IMHO in a lot of places
> > we don't really need the full BQL, we just need a 'you aren't going to
> > change the config' lock.
> 
> This is all about reducing main loop hangs.  Another one is moving
> "slow" code out of the main loop, e.g. monitor commands.
> 
> My question was aiming in a slightly different direction, however: given
> that the main loop can hang, is there anything we can do to mitigate
> known bad consequences of such hangs?
> 
> We're actually discussing one thing we can do to mitigate: moving the
> monitor core out of the main loop, to keep the monitor available.  Any
> other ideas?
> 
> >> >> >> We can have any number of QMP monitors today.  Would each of them feed
> >> >> >> its own queue?  Would they all feed a shared queue?
> >> >> >
> >> >> > Currently with multiple QMP monitors, everything runs in the main
> >> >> > loop, so commands arriving across  multiple monitors are 100%
> >> >> > serialized and processed strictly in the order in which QEMU reads
> >> >> > them off the wire.  To maintain these semantics, we would need to
> >> >> > have a single shared queue for the default main loop consumer, so
> >> >> > that ordering does not change.
> >> >> >
> >> >> >> How exactly is opt-in asynchronous to work?  Per QMP monitor?  Per
> >> >> >> command?
> >> >> >
> >> >> > Per monitor+command. ie just because libvirt knows how to cope with
> >> >> > async execution on the monitor it has open, does not mean that a
> >> >> > different app on the 2nd monitor command can cope. So in my proposal
> >> >> > the switch to async must be scoped to the particular command only
> >> >> > for the monitor connection that requesteed it.
> >> >> >
> >> >> >> What does it mean when an asynchronous command follows a synchronous
> >> >> >> command in the same QMP monitor?  I would expect the synchronous command
> >> >> >> to complete before the asynchronous command, because that's what
> >> >> >> synchronous means, isn't it?  To keep your QMP monitor available, you
> >> >> >> then must not send synchronous commands that can hang.
> >> >> >
> >> >> > No, that is not what I described. All synchronous commands are
> >> >> > serialized wrt each other, just as today. An asychronous command
> >> >> > can run as soon as it is received, regardless of whether any
> >> >> > earlier sent sync commands are still executing or pending. This
> >> >> > is trivial to achieve when you separate monitor I/O from command
> >> >> > execution in separate threads, provided of course the async
> >> >> > command consumers are not in the main loop.
> >> >> 
> >> >> So, a synchronous command is synchronous with respect to other commands,
> >> >> except for certain non-blocking commands.  The distinctive feature of
> >> >> the latter isn't so much an asynchronous reply, but out-of-band
> >> >> dispatch.
> >> >> 
> >> >> Out-of-band dispatch of commands that cannot block in fact orthogonal to
> >> >> asynchronous replies.  I can't see why out-of-band dispatch of
> >> >> synchronous non-blocking commands wouldn't work, too.
> >> >> 
> >> >> >> How can we determine whether a certain synchronous command can hang?
> >> >> >> Note that with opt-in async, *all* commands are also synchronous
> >> >> >> commands.
> >> >> >> 
> >> >> >> In short, explain to me how exactly you plan to ensure that certain QMP
> >> >> >> commands (such as post-copy recovery) can always "get through", in the
> >> >> >> presence of multiple monitors, hanging main loop, hanging synchronous
> >> >> >> commands, hanging whatever-else-can-now-hang-in-this-post-copy-world.
> >> >> >
> >> >> > Taking migrate-cancel as the example. The migration code already has
> >> >> > a background thread doing work independantly onthe main loop. Upon
> >> >> > marking the migrate-cancel command as async, the migration control
> >> >> > thread would become the consumer of migrate-cancel.
> >> >> 
> >> >> From 30,000 feet, the QMP monitor sends a "cancel" message to the
> >> >> migration thread, and later receives a "canceled" message from the
> >> >> migration thread.
> >> >> 
> >> >> From 300 feet, we use the migrate-cancel QMP command as the cancel
> >> >> message, and its success response as the "canceled" message.
> >> >> 
> >> >> In other words, we're pressing the external QM-Protocol into service as
> >> >> internal message passing protocol.
> >> >
> >> > Be careful; it's not a cancel in the postcopy recovery case, it's a
> >> > restart.  The command is very much like the migration-incoming command.
> >> > The management layer has to provide data with the request, so it's not
> >> > an internal command.
> >> 
> >> It's still a message.
> >> 
> >> >> >                                                     This allows the
> >> >> > migration operation to be cancelled immediately, regardless of whether
> >> >> > there are earlier monitor commands blocked in the main loop.
> >> >> 
> >> >> The necessary part is moving all operations that can block out of
> >> >> whatever loop runs the monitor, be it the main loop, some other event
> >> >> loop, or a dedicated monitor thread's monitor loop.
> >> >> 
> >> >> Moving out non-blocking operations isn't necessary.  migrate-cancel
> >> >> could communicate with the migration thread by any suitable mechanism or
> >> >> protocol.  It doesn't have to be QMP.  Why would we want it to be QMP?
> >> >
> >> > Because why invent another wheel?
> >> > This is a command that the management layer has to issue to qemu for
> >> > it to recover, including passing data, in a way similar to other
> >> > commands - so it looks like a QMP command, so why not use QMP.
> >> 
> >> Point taken.
> >> 
> >> Minor terminology remark: I'd prefer to call this a reuse of QAPI rather
> >> than QMP, because QMP makes me think of sockets and JSON, while QAPI
> >> makes me think of generated data types and marshaling code.
> >
> > Well it's a command that's got to come over the socket from management,
> > so I'm still thinking sockets and JSON. A lot of the problems you
> > describe below are more to do with the pain of managing the messages
> > squeezed through a socket.
> >
> >> > Also, I think making other commands lock-free is advantageous - 
> >> > some of the 'info' commands just dont really need locks, making them
> >> > not use locks removes latency effects caused by the management layer
> >> > prodding qemu.
> >> 
> >> I get the desire to move commands that can block out of whatever loop
> >> runs the monitor.  But moving out commands that always complete quickly
> >> seems pointless: by the time you're done queuing them, you could be done
> >> *executing* them.  More on that below.
> >
> > My thinking here wasn't about the speed of executing the command, my
> > interest was more on the performance of the guest/IO - avoiding taking
> > the BQL would have less impact on IO emulation, as would keeping the
> > main thread free.
> >
> >> >> > Of course this assumes the migration control thread can't block
> >> >> > for locks held by the main thread.
> >> >> 
> >> >> Thanks for your answers, they help.
> >> >> 
> >> >> >> Now let's talk about QMP requirements.
> >> >> >> 
> >> >> >> Any addition to QMP must consider what exists already.
> >> >> >> 
> >> >> >> You may add more of the same.
> >> >> >> 
> >> >> >> You may generalize existing stuff.
> >> >> >> 
> >> >> >> You may change existing stuff if you have sufficient reason, subject to
> >> >> >> backward compatibility constraints.
> >> >> >> 
> >> >> >> But attempts to add new ways to do the same old stuff without properly
> >> >> >> integrating the existing ways are not going to fly.
> >> >> >> 
> >> >> >> In particular, any new way to start some job, monitor and control it
> >> >> >> while it lives, get notified about its state changes and so forth must
> >> >> >> integrate the existing ways.  These include block jobs (probably the
> >> >> >> most sophisticated of the lot), migration, dump-guest-memory, and
> >> >> >> possibly more.  They all work the same way: synchronous command to kick
> >> >> >> off the job, more synchronous commands to monitor and control, events to
> >> >> >> notify.  They do differ in detail.
> >> >> >> 
> >> >> >> Asynchronous commands are a new way to do this.  When you only need to
> >> >> >> be notified on "done", and don't need to monitor / control, they fit the
> >> >> >> bill quite neatly.
> >> >> >> 
> >> >> >> However, we can't just ignore the cases where we need more than that!
> >> >> >> For those, we want a single generic solution instead of the several ad
> >> >> >> hoc solutions we have now.
> >> >> >> 
> >> >> >> If we add asynchronous commands *now*, and for simple cases only, we add
> >> >> >> yet another special case for a future generic solution to integrate.
> >> >> >> I'm not going to let that happen.
> >> >> >
> >> >> > With the async commands suggestion, while it would initially not
> >> >> > provide a way to query incremental status, that could easily be
> >> >> > fitted in.
> >> >> 
> >> >> This is [*] below.
> >> >> 
> >> >> >             Because command replies from async commands may be
> >> >> > out-of-order wrt the original requests, clients would need to
> >> >> > provide a unique ID for each command run. This originally was
> >> >> > part of QMP spec but then dropped, but libvirt still actually
> >> >> > generates a uniqe ID for every QMP command.
> >> >> >
> >> >> > Given this, one option is to actually use the QMP command ID as
> >> >> > a job ID, and let you query ongoing status via some new QMP
> >> >> > command that accepts the ID of the job to be queried. A complexity
> >> >> > with this is how to make the jobs visible across multiple QMP
> >> >> > monitors. The job ID might actually have to be a combination of
> >> >> > the serial ID from the QMP command, and the ID of the monitor
> >> >> > chardev combined.
> >> >> 
> >> >> Yes.  The job ID must be unique across all QMP monitors to make
> >> >> broadcast notifications work.
> >> >> 
> >> >> >> I figure the closest to a generic solution we have is block jobs.
> >> >> >> Perhaps a generic solution could be had by abstracting away the "block"
> >> >> >> from "block jobs", leaving just "jobs".
> >> >> 
> >> >> [*] starts here:
> >> >> 
> >> >> >> Another approach is generalizing the asynchronous command proposal to
> >> >> >> fully cover the not-so-simple cases.
> >> >> 
> >> >> We know asynchronous commands "fully cover" when we can use them to
> >> >> replace all the existing job-like commands.
> >> >> 
> >> >> Until then, they enlarge rather than solve our jobs problem.
> >> >> 
> >> >> I get the need for an available monitor.  But I need to balance it with
> >> >> other needs.  Can we find a solution for our monitor availability
> >> >> problem that doesn't enlarge our jobs problem?
> >> >
> >> > Hopefully!
> >> >
> >> > Dave
> >> >
> >> >> >> If you'd rather want to make progress on monitor availability without
> >> >> >> cracking the "jobs" problem, you're in luck!  Use your license to "add
> >> >> >> more of the same": synchronous command to start a job, query to monitor,
> >> >> >> event to notify.  
> >> >> >> 
> >> >> >> If you insist on tying your monitor availability solution to
> >> >> >> asynchronous commands, then I'm in luck!  I just found volunteers to
> >> >> >> solve the "jobs" problem for me.
> >> 
> >> Let me try to distill the discussion so far into a design sketch.
> >> 
> >> 1. A QMP monitor runs in a loop.  The loop may execute other stuff, but
> >>    this must not unduly delay the monitor's work.  Thus, everything in
> >>    this loop must complete "quickly".
> >> 
> >>    All QMP monitors currently run in the main loop, which really should
> >>    satisfy "quickly", but doesn't.  Since fixing that to a tolerable
> >>    degree is beyond our means (is it?), we move them out.
> >> 
> >>    Design alternative: either one loop and thread per monitor, or one
> >>    loop and thread for all monitors, or something in between.
> >> 
> >>    I'm wary of "one thread per software artifact" designs.  "One
> >>    (preemptable) thread per activity, all sharing state" is a lousy way
> >>    to structure software.
> >
> > Shrug; I've always thought of it as an easy solution unless you'd
> > get into hundreds of threads, which given the number of monitors, we
> > wont.
> >
> >> 2. A QMP monitor receives and dispatches commands, and sends command
> >>    responses and events.
> >> 
> >>    What if sending a response or event would block?  See 6.
> >> 
> >> 3. Arbitrary code can trigger QMP events.  Events are broadcast to all
> >>    QMP monitors.  Each QMP monitor provides an event queue.  When an
> >>    event is triggered, it gets put into all queues, subject to rate
> >>    limiting.
> 
> Correction: only events that are rate-limited go through the queue.  The
> other bypass it.  This is an optimization.
> 
> >>    Rate limiting and queuing needs some shared data, which is protected
> >>    by a mutex.  The critical sections guarded by this mutex must be
> >>    "quick".
> 
> There's another mutex guarding the monitor's output buffer (see 6.),
> among other things.
> 
> >>    Nothing new here, it's how events work today.
> >> 
> >>    We could easily add events that go to just one monitor, if there's a
> >>    need.
> >
> > I don't think events could cause a problem here since they're always
> > outbound - so they could never block inbound commands?
> 
> Events are indeed okay as they are.  I merely wanted to mention that
> they don't *have* to broadcast.  The "queue full" event mentioned under
> 6. probably shouldn't be broadcast.
> 
> >> 4. Commands are normally dispatched to a worker thread, where they can
> >>    take their own sweet time to complete.
> >> 
> >>    Currently, the monitor runs in the main loop and executes commands
> >>    directly.  This is effectively dispatching commands to the main loop.
> >>    Dispatch to main loop is wrong, because it can make the main loop
> >>    hang.  If it was the only relevant cause for main loop hangs, we'd
> >>    move the command work out and be done.  Since it isn't (see 1.) we
> >>    *also* have to move the monitor out to prevent main loop hangs from
> >>    hanging the monitor.
> >> 
> >>    Moving monitor and command work to separate threads changes the
> >>    dispatch from function call to queuing.  Need a pair of queues, one
> >>    for commands, one for responses.
> >> 
> >>    Design alternative: one worker per monitor, or one worker for all
> >>    monitors, or main loop is the one worker for all monitors.  The
> >>    latter leaves the main loop hangs unaddressed.  No worse than before,
> >>    so it could be okay as a first step.
> >> 
> >>    The worker provides the pair of queues.  It executes commands in
> >>    order.  If a command blocks, the command queue stalls.
> >> 
> >>    The command queue can therefore grow without bounds.  See 6.
> >> 
> >> 5. Certain commands that always complete "quickly" can instead be
> >>    executed directly, at the QMP client's request.  This direct
> >>    execution is out-of-band; the response can "overtake" prior in-band
> >>    commands' responses.
> >> 
> >>    The amount of work these out-of-band commands do themselves is up to
> >>    them.  A qiuck query command would do all the work.  migrate-cancel
> >>    could perhaps notify the migration thread and be done.  Postcopy
> >>    recovery could perhaps send its argument struct to whatever thread
> >>    handles recovery.
> >
> > Yes.
> 
> Message sending needs to be non-blocking.  If the message can't be sent,
> the command should fail.  Queuing instead is a problematic idea, because
> then you get to deal with the same flow control problems we're
> discussing below.
> 
> >> 6. Flow control
> >
> > I think this is potentially the tricky bit!
> >
> >>    We currently leave flow control to the underlying character device.
> >>    If the client sends more quickly than the monitor can execute, the
> >>    client's send eventually blocks or fails with EAGAIN.  If the monitor
> >>    sends more quickly than the client accepts, the monitor buffers
> >>    without bounds (I think).
> >> 
> >>    Buffering monitor output without bounds is bad.  We could perhaps
> >>    kill a monitor when it exceeds its limit.
> >
> > I'm not sure it's possible to define that limit; for example
> > 'query-block' gives a list of information for all devices; there are
> > people running with 200+ block devices so the output for that would be
> > huge.
> 
> For comparison, here are our current input limits:
> 
> * Sum of JSON token size: 64MiB
> * JSON token count: 2Mi
> * JSON nesting depth 1024
> 
> The first two limit heap usage, the third limits stack usage.  The first
> and the last go back to Anthony (commit 29c75dd).  I added the second
> because the first is insufficient (commit df64983).
> 
> If we hit these generous limits, surely something has gone haywire.
> 
> An output limit of 64MiB should be good for ~100k block devices with
> MiBs to spare.  Generous enough for a "if you hit this limit, you're
> abusing QMP way too much" argument?  If not, how far left would you like
> me to shift the limit?
> 
> >>    Buffering monitor input (in the command queue) without bounds is just
> >>    as bad.  It also destroys the existing flow control mechanism: the
> >>    client can no longer detect that it's sending too much.  Not an issue
> >>    for fully synchronous clients, i.e. clients that wait for the
> >>    previous command's response before they send the next command.  Such
> >>    clients cannot use of out-of-band command execution.
> >> 
> >>    The obvious way to limit the command queue is to fail commands when
> >>    the queue is "full".
> >> 
> >>    Note that we can't send an error response right away then, because
> >>    the command is in-band (if it wasn't, we wouldn't queue it), so its
> >>    response has to go after all all the respones to the (in-band)
> >>    commands currently in the queue.
> >> 
> >>    To tell the client right away, we could send an event.
> >> 
> >>    Delaying the "queue full" response until the correct time to send it
> >>    requires state: at least the command ID.  We can just as well enqueue
> >>    and pray memory will suffice.
> >> 
> >>    Note that the only reason for the command queue is out-of-band
> >>    commands.  Without them, reading the next command is pointless.  This
> >>    leads me to a possible solution: separate out-of-band mode, default
> >>    off, QMP client can switch it on.  When off, we read monitor input
> >>    just like we do now (no queue, no problem).  When on, we read and
> >>    queue.  If the queue is full, we send a "queue full" event with the
> >>    IDs of the commands we dropped on the floor.  By switching on
> >>    out-of-band-mode, the QMP client also opts into this event.
> >> 
> >>    Switching could be done with QMP capabilities negotiation.
> >
> > I'm not sure how this queue interacts for multiple monitors using the
> > single IO thread.  It's currently legal for each monitor to send one
> > command and for that command to be outstanding; so 'queue full' mustn't
> > happen in that case, because we still want to allow any of the monitors
> > to issue one of the non-locking commands.
> 
> Right, the "queue full" condition must be per monitor, and it must not
> apply to in-band commands (which aren't queued anyway).
> 
> > So I think we need 2x 1 entry input queues per monitor; one for normal
> > command and one for non-locking commands; I think that's different
> > from what we've previously suggested which is 2 central queues.
> 
> Perhaps I was less than clear under 4., but I meant to propose design
> alternatives one shared worker fed by one pair of queues, and one worker
> per monitor, each fed by its own pair of queues.  Another alternative
> would be one shared worker fed by one pair of queues per monitor.
> 
> Pair of queues means one for in-band commands, one for their responses.
> There is no queue for out-of-band commands, because out-of-band commands
> are not queued.
> 
> >> 7. How all this is related to "jobs"
> >> 
> >>    Out-of-band execution is a limited special case of asynchronous
> >>    execution.  With general asynchronous execution, responses can be
> >>    sent in any order.  With out-of-band execution, only the out-of-band
> >>    responses can "jump" order, and only over in-band responses.
> >> 
> >>    "All commands are (to be treated as) asynchronous" is arguably more
> >>    elegant than this out-of-band thing.  However, it runs into two
> >>    roadblocks that don't apply to out-of-band.
> >> 
> >>    One, backward compatibility.  That's a roadblock only as much as we
> >>    make it one.
> >> 
> >>    Two, consistency.  "All asynchronous, but we do most job-like things
> >>    with commands + events anyway" is not acceptable to me.  I'd be
> >>    willing to accept "all asynchronous" when it solves the jobs problem.
> >
> > I suspect there are other things that limit making everything
> > asynchronous; for example commands that currently only expect to be
> > executing in the main thread; if you wanted to make an existing command
> > async you'd have to audit it for all the possible places it could hang.
> 
> You're right.
> 
> > I also see the other problem as keeping the management level
> > understanding of which commands are asynchronous; Dan's suggestion is
> > that command where the management layer specifies which commands it
> > expects to be asynchronous, and qemu responds with which ones actually
> > are.
> 
> "Command supports out-of-band dispatch" would be visible in
> query-qmp-schema.
> 
> Design alternative: either switching on out-of-band mode (see 6.)
> switches all out-of-band commands to out-of-band dispatch, or it
> doesn't, and the client has to request out-of-band dispatch explicitly.
> The explicit request could either be per execute (say send {'exec-oob':
> COMMAND-NAME ...} instead of {'execute': COMMAND-NAME...}), or per
> session, i.e. with a new command to enable oob dispatch for a list of
> oob-capable commands.
> 
> I figure explicit is safer, because it lets us make more commands
> oob-capable without upsetting existing oob-aware QMP clients.

I think this OOB solution should work for us, though I'm still trying
to digest this whole thing.  Thanks Markus for this design, much
appreciated.  Meanwhile, sorry to have troubled you on this. I really
didn't mean to!

Considering that we may still have some commands (like what Fam has
mentioned in block layer) that may need to be run only in main thread,
I think a first attempt may need to have one IO/parser thread (parses
monitor input stream, and also responsible to run out-of-band
commands), plus no worker thread, then I'll feed the dispatching work
back to main thread again to make sure that assumption still keeps.

Dan, do you think this will work from libvirt POV?  I won't try to
prototype anything if without your confirmation as well.

Thanks!

> 
> >>    You asked for a solution to the monitor availability problem that
> >>    doesn't require you to solve the jobs problem first.  Well, here's my
> >>    best try.  Go shoot some holes into it :)
> >
> > Hopefully we're running out of holes.
> 
> Thanks!

-- 
Peter Xu

On Mon, Sep 11, 2017 at 06:32:03PM +0800, Peter Xu wrote:

[...]

> I think this OOB solution should work for us, though I'm still trying
> to digest this whole thing.  Thanks Markus for this design, much
> appreciated.  Meanwhile, sorry to have troubled you on this. I really
> didn't mean to!
> 
> Considering that we may still have some commands (like what Fam has
> mentioned in block layer) that may need to be run only in main thread,
> I think a first attempt may need to have one IO/parser thread (parses
> monitor input stream, and also responsible to run out-of-band
> commands), plus no worker thread, then I'll feed the dispatching work
> back to main thread again to make sure that assumption still keeps.
> 
> Dan, do you think this will work from libvirt POV?  I won't try to
> prototype anything if without your confirmation as well.
> 
> Thanks!

CC Daniel.

-- 
Peter Xu

On Fri, Sep 08, 2017 at 01:49:41PM +0200, Markus Armbruster wrote:
> > I also see the other problem as keeping the management level
> > understanding of which commands are asynchronous; Dan's suggestion is
> > that command where the management layer specifies which commands it
> > expects to be asynchronous, and qemu responds with which ones actually
> > are.
> 
> "Command supports out-of-band dispatch" would be visible in
> query-qmp-schema.
> 
> Design alternative: either switching on out-of-band mode (see 6.)
> switches all out-of-band commands to out-of-band dispatch, or it
> doesn't, and the client has to request out-of-band dispatch explicitly.
> The explicit request could either be per execute (say send {'exec-oob':
> COMMAND-NAME ...} instead of {'execute': COMMAND-NAME...}), or per
> session, i.e. with a new command to enable oob dispatch for a list of
> oob-capable commands.
> 
> I figure explicit is safer, because it lets us make more commands
> oob-capable without upsetting existing oob-aware QMP clients.

Yep, this is fine too - it achieves the same end goals as the approach
I suggest. Namely

 - clients can detect which commands can do OOB (via the schema)
 - clients can choose which commands to run OOB (via exec vs exec-oob)

Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

On Thu, Sep 07, 2017 at 07:41:29PM +0200, Markus Armbruster wrote:
> "Daniel P. Berrange" <berrange@redhat.com> writes:
> 
> > On Thu, Sep 07, 2017 at 02:59:28PM +0200, Markus Armbruster wrote:
> >> So, what exactly is going to drain the command queue?  If there's more
> >> than one consumer, how exactly are commands from the queue dispatched to
> >> the consumers?
> >
> > In terms of my proposal, for any single command there should only ever
> > be a single consumer. The default consumer would be the main event loop
> > thread, such that we have no semantic change to QMP operation from today.
> >
> > Some commands that are capable of being made "async", would have a
> > different consumer. For example, if the client requested the 'migrate-cancel'
> > be made async, this would change things such that the migration thread is
> > now responsible for consuming the "migrate-cancel" command, instead of the
> > default main loop.
> >
> >> What are the "no hang" guarantees (if any) and conditions for each of
> >> these consumers?
> >
> > The non-main thread consumers would have to have some reasonable
> > guarantee that they won't block on a lock held by the main loop,
> > otherwise the whole feature is largely useless.
> 
> Same if they block indefinitely on anything else, actually.  In other
> words, we need to talk about liveness.
> 
> Threads by themselves don't buy us liveness.  Being careful with
> operations that may block does.  That care may lead to farming out
> certain operations to other threads, where they may block without harm.
> 
> You only talk about "the non-main thread consumers".  What about the
> main thread?  Is it okay for the main thread to block?  If yes, why?

It isn't ok, but I feel that challenge is intractable in the short to
medium term. Agree that having separate threads doesn't automatically
give liveness, but I think it makes the problem tractble to solve for
at least a subset of scenarios.

> > No, that is not what I described. All synchronous commands are
> > serialized wrt each other, just as today. An asychronous command
> > can run as soon as it is received, regardless of whether any
> > earlier sent sync commands are still executing or pending. This
> > is trivial to achieve when you separate monitor I/O from command
> > execution in separate threads, provided of course the async
> > command consumers are not in the main loop.
> 
> So, a synchronous command is synchronous with respect to other commands,
> except for certain non-blocking commands.  The distinctive feature of
> the latter isn't so much an asynchronous reply, but out-of-band
> dispatch.

The terminology synchronous vs asynchronous is not a great fit for
what I was describing. The distinction is really closer to being
serialized vs parallelizable commands.

> >                                                     This allows the
> > migration operation to be cancelled immediately, regardless of whether
> > there are earlier monitor commands blocked in the main loop.
> 
> The necessary part is moving all operations that can block out of
> whatever loop runs the monitor, be it the main loop, some other event
> loop, or a dedicated monitor thread's monitor loop.
> 
> Moving out non-blocking operations isn't necessary.  migrate-cancel
> could communicate with the migration thread by any suitable mechanism or
> protocol.  It doesn't have to be QMP.  Why would we want it to be QMP?

I don't think we really want to invent yet another way of controlling
QEMU, that isn't QMP do we, particularly not if it is special cased
to just one operationg ?

Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

* Markus Armbruster (armbru@redhat.com) wrote:
> Peter Xu <peterx@redhat.com> writes:
> 
> > On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
> >> On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> >> > * Daniel P. Berrange (berrange@redhat.com) wrote:
> >> > > This does imply that you need a separate monitor I/O processing, from the
> >> > > command execution thread, but I see no need for all commands to suddenly
> >> > > become async. Just allowing interleaved replies is sufficient from the
> >> > > POV of the protocol definition. This interleaving is easy to handle from
> >> > > the client POV - just requires a unique 'serial' in the request by the
> >> > > client, that is copied into the reply by QEMU.
> >> > 
> >> > OK, so for that we can just take Marc-André's syntax and call it 'id':
> >> >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> >> > 
> >> > then it's upto the caller to ensure those id's are unique.
> >> 
> >> Libvirt has in fact generated a unique 'id' for every monitor command
> >> since day 1 of supporting QMP.
> >> 
> >> > I do worry about two things:
> >> >   a) With this the caller doesn't really know which commands could be
> >> >   in parallel - for example if we've got a recovery command that's
> >> >   executed by this non-locking thread that's OK, we expect that
> >> >   to be doable in parallel.  If in the future though we do
> >> >   what you initially suggested and have a bunch of commands get
> >> >   routed to the migration thread (say) then those would suddenly
> >> >   operate in parallel with other commands that we're previously
> >> >   synchronous.
> >> 
> >> We could still have an opt-in for async commands. eg default to executing
> >> all commands in the main thread, unless the client issues an explicit
> >> "make it async" command, to switch to allowing the migration thread to
> >> process it async.
> >> 
> >>  { "execute": "qmp_allow_async",
> >>    "data": { "commands": [
> >>        "migrate_cancel",
> >>    ] } }
> >> 
> >> 
> >>  { "return": { "commands": [
> >>        "migrate_cancel",
> >>    ] } }
> >> 
> >> The server response contains the subset of commands from the request
> >> for which async is supported.
> >> 
> >> That gives good negotiation ability going forward as we incrementally
> >> support async on more commands.
> >
> > I think this goes back to the discussion on which design we'd like to
> > choose.  IMHO the whole async idea plus the per-command-id is indeed
> > cleaner and nicer, and I believe that can benefit not only libvirt,
> 
> The following may be a bit harsh in places.  I apologize in advance.  A
> better writer than me wouldn't have to resort to that.  I've tried a few
> times to make my point that "async QMP" is neither necessary nor
> sufficient for monitor availability, but apparently without luck, since
> there's still talk like it was.  I hope this attempt will work.
> 
> > but also other QMP users.  The problem is, I have no idea how long
> > it'll take to let us have such a feature - I believe that will include
> > QEMU and Libvirt to both support that.  And it'll be a pity if the
> > postcopy recovery cannot work only because we cannot guarantee a
> > stable monitor.
> >
> > I'm curious whether there are other requirements (besides postcopy
> > recovery) that would want an always-alive monitor to run some
> > lock-free commands?  If there is, I'd be more inclined to first
> > provide a work-around solution like "-qmp-lockfree", and we can
> > provide a better solution afterwards until when the whole async QMP
> > work ready.
> 
> Yes, there are other requirements for "async QMP", and no, "async QMP"
> isn't a solution, but at best a part of a solution.
> 
> Before I talk about QMP requirements, I need to ask a whole raft of
> questions, because so far this thread feels like dreaming up grand
> designs with only superficial understanding of the subject matter.

I think Dan's suggestions are pretty good; while I prefered Peter's
implementation, I think Dan's will work fine and if that's good for
libvirt I'm OK with that.  I think we have a reasonable understanding
of the problem.

> Quite possibly because *my* understanding is superficial.  If yours
> isn't, great!  Go answer my questions :)
> 
> The root problem are main loop hangs.  QMP monitor hangs are merely a
> special case.
> 
> The main loop should not hang.  We've always violated that design
> assumption in places, e.g. in monitor commands that write to disk, and
> thus can hang indefinitely with NFS.  Post-copy adds more violations, as
> Stefan pointed out.
> 
> I can't say whether solving the special case "QMP monitor hangs" without
> also solving "main loop hangs" is useful.  A perfectly available QMP
> monitor buys you nothing if it feeds a command queue that isn't being
> emptied because its consumers all hang.

Correct.

> So, what exactly is going to drain the command queue?  If there's more
> than one consumer, how exactly are commands from the queue dispatched to
> the consumers?

The idea is to have 2 extra threads:
   a) An IO thread
   b) A thread that deals with non-blocking commands
   the existing main thread.

   The IO thread dispatches most commands to the main thread
but doesn't wait for the response.  When responses arrive it forwards
the response back.
   A class of commands is forwarded to the non-blocking command thread.

   More threads may be added in the future with some set of the commands
being moved off the main thread to these other threads.  Eventually
maybe no commands would be handled on the main thread.

> What are the "no hang" guarantees (if any) and conditions for each of
> these consumers?

Commands sent to the main thread are as they are now.
The non-blocking-command thread *shall not block*, it will not access
guest memory, it wont take any lock that is taken by any other thread
that can block on the main thread or main memory.  Commands that run
on it can:
   a) Access state that can be read atomically - e.g. 
      'info status'
   b) Store parameters and then wake another thread
   c) Issue a non-blocking system call.


  In the case of postcopy recovery I see a command issued which starts
the new migration stream;  the command parses the path and makes sure
it's valid, and then stores it and kicks a recovery thread.
  In the case of a COLO failover I'd see something that does a
shutdown(2) on the migration stream.

> We can have any number of QMP monitors today.  Would each of them feed
> its own queue?  Would they all feed a shared queue?

I see two queues; one which is the set of commands being forwarded
to the main thread, the other is the set of commands being forwarded
to the non-blocking thread.

> How exactly is opt-in asynchronous to work?  Per QMP monitor?  Per
> command?

The command that Dan suggested is the opt-in; I think it's per monitor;
now we're starting to get a bit more fuzzy.

> What does it mean when an asynchronous command follows a synchronous
> command in the same QMP monitor?  I would expect the synchronous command
> to complete before the asynchronous command, because that's what
> synchronous means, isn't it?  To keep your QMP monitor available, you
> then must not send synchronous commands that can hang.

Once you opt-in, all commands operate in a semi-asynchronous fashion;
that is they don't block the IO thread, but at the same time there's
never any more than one command outstanding on any one thread.
You can issue any command you like; one command at a time waiting
for the response with the knowledge that you can then always
issue one of the non-blocking-commands after it.

> How can we determine whether a certain synchronous command can hang?
> Note that with opt-in async, *all* commands are also synchronous
> commands.

You regard all commands as blockable unless told otherwise.  The result
from Dan's command is a list of truly async commands.

> In short, explain to me how exactly you plan to ensure that certain QMP
> commands (such as post-copy recovery) can always "get through", in the
> presence of multiple monitors, hanging main loop, hanging synchronous
> commands, hanging whatever-else-can-now-hang-in-this-post-copy-world.

Have I missed anything?

> 
> Now let's talk about QMP requirements.
> 
> Any addition to QMP must consider what exists already.

Yes.

> You may add more of the same.

Yes

> You may generalize existing stuff.

Yes

> You may change existing stuff if you have sufficient reason, subject to
> backward compatibility constraints.

Yes

> But attempts to add new ways to do the same old stuff without properly
> integrating the existing ways are not going to fly.

Agreed; that's why I'm following Dan's recommendations.

> In particular, any new way to start some job, monitor and control it
> while it lives, get notified about its state changes and so forth must
> integrate the existing ways.  These include block jobs (probably the
> most sophisticated of the lot), migration, dump-guest-memory, and
> possibly more.  They all work the same way: synchronous command to kick
> off the job, more synchronous commands to monitor and control, events to
> notify.  They do differ in detail.

And that's why we have the rule that all existing commands go onto the
main thread and only one of those is outstanding at a time.  That way
the actual behaviour of the existing commands doesn't change at all -
however you do require the 'id' field in the command to put into the
response so that you can distinguish the response of a command from each
thread.  Even if you enable async, if you don't use any of the
non-blocking commands the stream is just the same - send a command, get
a response, send a command, get a response....

> Asynchronous commands are a new way to do this.  When you only need to
> be notified on "done", and don't need to monitor / control, they fit the
> bill quite neatly.
> 
> However, we can't just ignore the cases where we need more than that!
> For those, we want a single generic solution instead of the several ad
> hoc solutions we have now.
> 
> If we add asynchronous commands *now*, and for simple cases only, we add
> yet another special case for a future generic solution to integrate.
> I'm not going to let that happen.
> 
> I figure the closest to a generic solution we have is block jobs.
> Perhaps a generic solution could be had by abstracting away the "block"
> from "block jobs", leaving just "jobs".

I don't know block jobs well enough to answer that.
I would suggest you could add a thread for asynchronous commands
and you could shuffle commands onto that thread as and when you feel
like it.

> Another approach is generalizing the asynchronous command proposal to
> fully cover the not-so-simple cases.
> 
> If you'd rather want to make progress on monitor availability without
> cracking the "jobs" problem, you're in luck!  Use your license to "add
> more of the same": synchronous command to start a job, query to monitor,
> event to notify.  
> 
> If you insist on tying your monitor availability solution to
> asynchronous commands, then I'm in luck!  I just found volunteers to
> solve the "jobs" problem for me.

I'm looking for minimal change here while keeping the door open for
the future, if there's anything you think we should do to make that
easy then tell us - but I'd rather this didn't turn into a 'fix all
known monitor problems' because frankly we may as well give up now.
So i don't see this as solving the 'jobs' problem, but if we can
do something to make it easier to solve in the future then lets do it.

Dave

--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

"Dr. David Alan Gilbert" <dgilbert@redhat.com> writes:

> * Markus Armbruster (armbru@redhat.com) wrote:
>> Peter Xu <peterx@redhat.com> writes:
>> 
>> > On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
>> >> On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
>> >> > * Daniel P. Berrange (berrange@redhat.com) wrote:
>> >> > > This does imply that you need a separate monitor I/O processing, from the
>> >> > > command execution thread, but I see no need for all commands to suddenly
>> >> > > become async. Just allowing interleaved replies is sufficient from the
>> >> > > POV of the protocol definition. This interleaving is easy to handle from
>> >> > > the client POV - just requires a unique 'serial' in the request by the
>> >> > > client, that is copied into the reply by QEMU.
>> >> > 
>> >> > OK, so for that we can just take Marc-André's syntax and call it 'id':
>> >> >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
>> >> > 
>> >> > then it's upto the caller to ensure those id's are unique.
>> >> 
>> >> Libvirt has in fact generated a unique 'id' for every monitor command
>> >> since day 1 of supporting QMP.
>> >> 
>> >> > I do worry about two things:
>> >> >   a) With this the caller doesn't really know which commands could be
>> >> >   in parallel - for example if we've got a recovery command that's
>> >> >   executed by this non-locking thread that's OK, we expect that
>> >> >   to be doable in parallel.  If in the future though we do
>> >> >   what you initially suggested and have a bunch of commands get
>> >> >   routed to the migration thread (say) then those would suddenly
>> >> >   operate in parallel with other commands that we're previously
>> >> >   synchronous.
>> >> 
>> >> We could still have an opt-in for async commands. eg default to executing
>> >> all commands in the main thread, unless the client issues an explicit
>> >> "make it async" command, to switch to allowing the migration thread to
>> >> process it async.
>> >> 
>> >>  { "execute": "qmp_allow_async",
>> >>    "data": { "commands": [
>> >>        "migrate_cancel",
>> >>    ] } }
>> >> 
>> >> 
>> >>  { "return": { "commands": [
>> >>        "migrate_cancel",
>> >>    ] } }
>> >> 
>> >> The server response contains the subset of commands from the request
>> >> for which async is supported.
>> >> 
>> >> That gives good negotiation ability going forward as we incrementally
>> >> support async on more commands.
>> >
>> > I think this goes back to the discussion on which design we'd like to
>> > choose.  IMHO the whole async idea plus the per-command-id is indeed
>> > cleaner and nicer, and I believe that can benefit not only libvirt,
>> 
>> The following may be a bit harsh in places.  I apologize in advance.  A
>> better writer than me wouldn't have to resort to that.  I've tried a few
>> times to make my point that "async QMP" is neither necessary nor
>> sufficient for monitor availability, but apparently without luck, since
>> there's still talk like it was.  I hope this attempt will work.
>> 
>> > but also other QMP users.  The problem is, I have no idea how long
>> > it'll take to let us have such a feature - I believe that will include
>> > QEMU and Libvirt to both support that.  And it'll be a pity if the
>> > postcopy recovery cannot work only because we cannot guarantee a
>> > stable monitor.
>> >
>> > I'm curious whether there are other requirements (besides postcopy
>> > recovery) that would want an always-alive monitor to run some
>> > lock-free commands?  If there is, I'd be more inclined to first
>> > provide a work-around solution like "-qmp-lockfree", and we can
>> > provide a better solution afterwards until when the whole async QMP
>> > work ready.
>> 
>> Yes, there are other requirements for "async QMP", and no, "async QMP"
>> isn't a solution, but at best a part of a solution.
>> 
>> Before I talk about QMP requirements, I need to ask a whole raft of
>> questions, because so far this thread feels like dreaming up grand
>> designs with only superficial understanding of the subject matter.
>
> I think Dan's suggestions are pretty good; while I prefered Peter's
> implementation, I think Dan's will work fine and if that's good for
> libvirt I'm OK with that.  I think we have a reasonable understanding
> of the problem.
>
>> Quite possibly because *my* understanding is superficial.  If yours
>> isn't, great!  Go answer my questions :)
>> 
>> The root problem are main loop hangs.  QMP monitor hangs are merely a
>> special case.
>> 
>> The main loop should not hang.  We've always violated that design
>> assumption in places, e.g. in monitor commands that write to disk, and
>> thus can hang indefinitely with NFS.  Post-copy adds more violations, as
>> Stefan pointed out.
>> 
>> I can't say whether solving the special case "QMP monitor hangs" without
>> also solving "main loop hangs" is useful.  A perfectly available QMP
>> monitor buys you nothing if it feeds a command queue that isn't being
>> emptied because its consumers all hang.
>
> Correct.
>
>> So, what exactly is going to drain the command queue?  If there's more
>> than one consumer, how exactly are commands from the queue dispatched to
>> the consumers?
>
> The idea is to have 2 extra threads:
>    a) An IO thread
>    b) A thread that deals with non-blocking commands

These are the two extra threads, and ...

>    the existing main thread.

... they are "extras" to the existing main thread, right?

>    The IO thread dispatches most commands to the main thread
> but doesn't wait for the response.  When responses arrive it forwards
> the response back.

The QMP monitor runs in this I/O thread?

>    A class of commands is forwarded to the non-blocking command thread.

Since the non-blocking commands by definition don't block, why can't we
simply execute them in the I/O thread?

>    More threads may be added in the future with some set of the commands
> being moved off the main thread to these other threads.  Eventually
> maybe no commands would be handled on the main thread.
>
>> What are the "no hang" guarantees (if any) and conditions for each of
>> these consumers?
>
> Commands sent to the main thread are as they are now.
> The non-blocking-command thread *shall not block*, it will not access
> guest memory, it wont take any lock that is taken by any other thread
> that can block on the main thread or main memory.  Commands that run
> on it can:
>    a) Access state that can be read atomically - e.g. 
>       'info status'
>    b) Store parameters and then wake another thread
>    c) Issue a non-blocking system call.
>
>
>   In the case of postcopy recovery I see a command issued which starts
> the new migration stream;  the command parses the path and makes sure
> it's valid, and then stores it and kicks a recovery thread.
>   In the case of a COLO failover I'd see something that does a
> shutdown(2) on the migration stream.
>
>> We can have any number of QMP monitors today.  Would each of them feed
>> its own queue?  Would they all feed a shared queue?
>
> I see two queues; one which is the set of commands being forwarded
> to the main thread, the other is the set of commands being forwarded
> to the non-blocking thread.
>
>> How exactly is opt-in asynchronous to work?  Per QMP monitor?  Per
>> command?
>
> The command that Dan suggested is the opt-in; I think it's per monitor;
> now we're starting to get a bit more fuzzy.
>
>> What does it mean when an asynchronous command follows a synchronous
>> command in the same QMP monitor?  I would expect the synchronous command
>> to complete before the asynchronous command, because that's what
>> synchronous means, isn't it?  To keep your QMP monitor available, you
>> then must not send synchronous commands that can hang.
>
> Once you opt-in, all commands operate in a semi-asynchronous fashion;
> that is they don't block the IO thread, but at the same time there's
> never any more than one command outstanding on any one thread.
> You can issue any command you like; one command at a time waiting
> for the response with the knowledge that you can then always
> issue one of the non-blocking-commands after it.
>
>> How can we determine whether a certain synchronous command can hang?
>> Note that with opt-in async, *all* commands are also synchronous
>> commands.
>
> You regard all commands as blockable unless told otherwise.  The result
> from Dan's command is a list of truly async commands.
>
>> In short, explain to me how exactly you plan to ensure that certain QMP
>> commands (such as post-copy recovery) can always "get through", in the
>> presence of multiple monitors, hanging main loop, hanging synchronous
>> commands, hanging whatever-else-can-now-hang-in-this-post-copy-world.
>
> Have I missed anything?

I'm trying to square this with Dan's reply, but it's probably too late
in my day for me to succeed.

>> Now let's talk about QMP requirements.
>> 
>> Any addition to QMP must consider what exists already.
>
> Yes.
>
>> You may add more of the same.
>
> Yes
>
>> You may generalize existing stuff.
>
> Yes
>
>> You may change existing stuff if you have sufficient reason, subject to
>> backward compatibility constraints.
>
> Yes
>
>> But attempts to add new ways to do the same old stuff without properly
>> integrating the existing ways are not going to fly.
>
> Agreed; that's why I'm following Dan's recommendations.
>
>> In particular, any new way to start some job, monitor and control it
>> while it lives, get notified about its state changes and so forth must
>> integrate the existing ways.  These include block jobs (probably the
>> most sophisticated of the lot), migration, dump-guest-memory, and
>> possibly more.  They all work the same way: synchronous command to kick
>> off the job, more synchronous commands to monitor and control, events to
>> notify.  They do differ in detail.
>
> And that's why we have the rule that all existing commands go onto the
> main thread and only one of those is outstanding at a time.  That way
> the actual behaviour of the existing commands doesn't change at all -
> however you do require the 'id' field in the command to put into the
> response so that you can distinguish the response of a command from each
> thread.  Even if you enable async, if you don't use any of the
> non-blocking commands the stream is just the same - send a command, get
> a response, send a command, get a response....
>
>> Asynchronous commands are a new way to do this.  When you only need to
>> be notified on "done", and don't need to monitor / control, they fit the
>> bill quite neatly.
>> 
>> However, we can't just ignore the cases where we need more than that!
>> For those, we want a single generic solution instead of the several ad
>> hoc solutions we have now.
>> 
>> If we add asynchronous commands *now*, and for simple cases only, we add
>> yet another special case for a future generic solution to integrate.
>> I'm not going to let that happen.
>> 
>> I figure the closest to a generic solution we have is block jobs.
>> Perhaps a generic solution could be had by abstracting away the "block"
>> from "block jobs", leaving just "jobs".
>
> I don't know block jobs well enough to answer that.
> I would suggest you could add a thread for asynchronous commands
> and you could shuffle commands onto that thread as and when you feel
> like it.
>
>> Another approach is generalizing the asynchronous command proposal to
>> fully cover the not-so-simple cases.
>> 
>> If you'd rather want to make progress on monitor availability without
>> cracking the "jobs" problem, you're in luck!  Use your license to "add
>> more of the same": synchronous command to start a job, query to monitor,
>> event to notify.  
>> 
>> If you insist on tying your monitor availability solution to
>> asynchronous commands, then I'm in luck!  I just found volunteers to
>> solve the "jobs" problem for me.
>
> I'm looking for minimal change here while keeping the door open for
> the future, if there's anything you think we should do to make that
> easy then tell us - but I'd rather this didn't turn into a 'fix all
> known monitor problems' because frankly we may as well give up now.
> So i don't see this as solving the 'jobs' problem, but if we can
> do something to make it easier to solve in the future then lets do it.

Forget about asynchronous commands, jobs and the whole shebang of
distractions, and consider what you really need: I suspect it could be
out-of-band dispatch of non-blocking commands.  More on that in my reply
to Daniel.

* Markus Armbruster (armbru@redhat.com) wrote:
> "Dr. David Alan Gilbert" <dgilbert@redhat.com> writes:
> 
> > * Markus Armbruster (armbru@redhat.com) wrote:
> >> Peter Xu <peterx@redhat.com> writes:
> >> 
> >> > On Wed, Sep 06, 2017 at 12:54:28PM +0100, Daniel P. Berrange wrote:
> >> >> On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> >> >> > * Daniel P. Berrange (berrange@redhat.com) wrote:
> >> >> > > This does imply that you need a separate monitor I/O processing, from the
> >> >> > > command execution thread, but I see no need for all commands to suddenly
> >> >> > > become async. Just allowing interleaved replies is sufficient from the
> >> >> > > POV of the protocol definition. This interleaving is easy to handle from
> >> >> > > the client POV - just requires a unique 'serial' in the request by the
> >> >> > > client, that is copied into the reply by QEMU.
> >> >> > 
> >> >> > OK, so for that we can just take Marc-André's syntax and call it 'id':
> >> >> >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> >> >> > 
> >> >> > then it's upto the caller to ensure those id's are unique.
> >> >> 
> >> >> Libvirt has in fact generated a unique 'id' for every monitor command
> >> >> since day 1 of supporting QMP.
> >> >> 
> >> >> > I do worry about two things:
> >> >> >   a) With this the caller doesn't really know which commands could be
> >> >> >   in parallel - for example if we've got a recovery command that's
> >> >> >   executed by this non-locking thread that's OK, we expect that
> >> >> >   to be doable in parallel.  If in the future though we do
> >> >> >   what you initially suggested and have a bunch of commands get
> >> >> >   routed to the migration thread (say) then those would suddenly
> >> >> >   operate in parallel with other commands that we're previously
> >> >> >   synchronous.
> >> >> 
> >> >> We could still have an opt-in for async commands. eg default to executing
> >> >> all commands in the main thread, unless the client issues an explicit
> >> >> "make it async" command, to switch to allowing the migration thread to
> >> >> process it async.
> >> >> 
> >> >>  { "execute": "qmp_allow_async",
> >> >>    "data": { "commands": [
> >> >>        "migrate_cancel",
> >> >>    ] } }
> >> >> 
> >> >> 
> >> >>  { "return": { "commands": [
> >> >>        "migrate_cancel",
> >> >>    ] } }
> >> >> 
> >> >> The server response contains the subset of commands from the request
> >> >> for which async is supported.
> >> >> 
> >> >> That gives good negotiation ability going forward as we incrementally
> >> >> support async on more commands.
> >> >
> >> > I think this goes back to the discussion on which design we'd like to
> >> > choose.  IMHO the whole async idea plus the per-command-id is indeed
> >> > cleaner and nicer, and I believe that can benefit not only libvirt,
> >> 
> >> The following may be a bit harsh in places.  I apologize in advance.  A
> >> better writer than me wouldn't have to resort to that.  I've tried a few
> >> times to make my point that "async QMP" is neither necessary nor
> >> sufficient for monitor availability, but apparently without luck, since
> >> there's still talk like it was.  I hope this attempt will work.
> >> 
> >> > but also other QMP users.  The problem is, I have no idea how long
> >> > it'll take to let us have such a feature - I believe that will include
> >> > QEMU and Libvirt to both support that.  And it'll be a pity if the
> >> > postcopy recovery cannot work only because we cannot guarantee a
> >> > stable monitor.
> >> >
> >> > I'm curious whether there are other requirements (besides postcopy
> >> > recovery) that would want an always-alive monitor to run some
> >> > lock-free commands?  If there is, I'd be more inclined to first
> >> > provide a work-around solution like "-qmp-lockfree", and we can
> >> > provide a better solution afterwards until when the whole async QMP
> >> > work ready.
> >> 
> >> Yes, there are other requirements for "async QMP", and no, "async QMP"
> >> isn't a solution, but at best a part of a solution.
> >> 
> >> Before I talk about QMP requirements, I need to ask a whole raft of
> >> questions, because so far this thread feels like dreaming up grand
> >> designs with only superficial understanding of the subject matter.
> >
> > I think Dan's suggestions are pretty good; while I prefered Peter's
> > implementation, I think Dan's will work fine and if that's good for
> > libvirt I'm OK with that.  I think we have a reasonable understanding
> > of the problem.
> >
> >> Quite possibly because *my* understanding is superficial.  If yours
> >> isn't, great!  Go answer my questions :)
> >> 
> >> The root problem are main loop hangs.  QMP monitor hangs are merely a
> >> special case.
> >> 
> >> The main loop should not hang.  We've always violated that design
> >> assumption in places, e.g. in monitor commands that write to disk, and
> >> thus can hang indefinitely with NFS.  Post-copy adds more violations, as
> >> Stefan pointed out.
> >> 
> >> I can't say whether solving the special case "QMP monitor hangs" without
> >> also solving "main loop hangs" is useful.  A perfectly available QMP
> >> monitor buys you nothing if it feeds a command queue that isn't being
> >> emptied because its consumers all hang.
> >
> > Correct.
> >
> >> So, what exactly is going to drain the command queue?  If there's more
> >> than one consumer, how exactly are commands from the queue dispatched to
> >> the consumers?
> >
> > The idea is to have 2 extra threads:
> >    a) An IO thread
> >    b) A thread that deals with non-blocking commands
> 
> These are the two extra threads, and ...
> 
> >    the existing main thread.
> 
> ... they are "extras" to the existing main thread, right?

Yes; three threads total.

> >    The IO thread dispatches most commands to the main thread
> > but doesn't wait for the response.  When responses arrive it forwards
> > the response back.
> 
> The QMP monitor runs in this I/O thread?

Yes; all the output formatting, basic checking of the input stream
(assuming that is that's all lock free)

> >    A class of commands is forwarded to the non-blocking command thread.
> 
> Since the non-blocking commands by definition don't block, why can't we
> simply execute them in the I/O thread?

Yes, I think that's possible.
I think it came from Dan's shape of suggestion so that in future
maybe we would add more threads for other types of things (e.g.
a thread that handled block commands)
(I don't think that ends up looking too different from Peter's world -
it's got one thread per monitor connection which is the other way up
from this but I think the behaviour is very similar).

> >    More threads may be added in the future with some set of the commands
> > being moved off the main thread to these other threads.  Eventually
> > maybe no commands would be handled on the main thread.
> >
> >> What are the "no hang" guarantees (if any) and conditions for each of
> >> these consumers?
> >
> > Commands sent to the main thread are as they are now.
> > The non-blocking-command thread *shall not block*, it will not access
> > guest memory, it wont take any lock that is taken by any other thread
> > that can block on the main thread or main memory.  Commands that run
> > on it can:
> >    a) Access state that can be read atomically - e.g. 
> >       'info status'
> >    b) Store parameters and then wake another thread
> >    c) Issue a non-blocking system call.
> >
> >
> >   In the case of postcopy recovery I see a command issued which starts
> > the new migration stream;  the command parses the path and makes sure
> > it's valid, and then stores it and kicks a recovery thread.
> >   In the case of a COLO failover I'd see something that does a
> > shutdown(2) on the migration stream.
> >
> >> We can have any number of QMP monitors today.  Would each of them feed
> >> its own queue?  Would they all feed a shared queue?
> >
> > I see two queues; one which is the set of commands being forwarded
> > to the main thread, the other is the set of commands being forwarded
> > to the non-blocking thread.
> >
> >> How exactly is opt-in asynchronous to work?  Per QMP monitor?  Per
> >> command?
> >
> > The command that Dan suggested is the opt-in; I think it's per monitor;
> > now we're starting to get a bit more fuzzy.
> >
> >> What does it mean when an asynchronous command follows a synchronous
> >> command in the same QMP monitor?  I would expect the synchronous command
> >> to complete before the asynchronous command, because that's what
> >> synchronous means, isn't it?  To keep your QMP monitor available, you
> >> then must not send synchronous commands that can hang.
> >
> > Once you opt-in, all commands operate in a semi-asynchronous fashion;
> > that is they don't block the IO thread, but at the same time there's
> > never any more than one command outstanding on any one thread.
> > You can issue any command you like; one command at a time waiting
> > for the response with the knowledge that you can then always
> > issue one of the non-blocking-commands after it.
> >
> >> How can we determine whether a certain synchronous command can hang?
> >> Note that with opt-in async, *all* commands are also synchronous
> >> commands.
> >
> > You regard all commands as blockable unless told otherwise.  The result
> > from Dan's command is a list of truly async commands.
> >
> >> In short, explain to me how exactly you plan to ensure that certain QMP
> >> commands (such as post-copy recovery) can always "get through", in the
> >> presence of multiple monitors, hanging main loop, hanging synchronous
> >> commands, hanging whatever-else-can-now-hang-in-this-post-copy-world.
> >
> > Have I missed anything?
> 
> I'm trying to square this with Dan's reply, but it's probably too late
> in my day for me to succeed.

Nod.

> >> Now let's talk about QMP requirements.
> >> 
> >> Any addition to QMP must consider what exists already.
> >
> > Yes.
> >
> >> You may add more of the same.
> >
> > Yes
> >
> >> You may generalize existing stuff.
> >
> > Yes
> >
> >> You may change existing stuff if you have sufficient reason, subject to
> >> backward compatibility constraints.
> >
> > Yes
> >
> >> But attempts to add new ways to do the same old stuff without properly
> >> integrating the existing ways are not going to fly.
> >
> > Agreed; that's why I'm following Dan's recommendations.
> >
> >> In particular, any new way to start some job, monitor and control it
> >> while it lives, get notified about its state changes and so forth must
> >> integrate the existing ways.  These include block jobs (probably the
> >> most sophisticated of the lot), migration, dump-guest-memory, and
> >> possibly more.  They all work the same way: synchronous command to kick
> >> off the job, more synchronous commands to monitor and control, events to
> >> notify.  They do differ in detail.
> >
> > And that's why we have the rule that all existing commands go onto the
> > main thread and only one of those is outstanding at a time.  That way
> > the actual behaviour of the existing commands doesn't change at all -
> > however you do require the 'id' field in the command to put into the
> > response so that you can distinguish the response of a command from each
> > thread.  Even if you enable async, if you don't use any of the
> > non-blocking commands the stream is just the same - send a command, get
> > a response, send a command, get a response....
> >
> >> Asynchronous commands are a new way to do this.  When you only need to
> >> be notified on "done", and don't need to monitor / control, they fit the
> >> bill quite neatly.
> >> 
> >> However, we can't just ignore the cases where we need more than that!
> >> For those, we want a single generic solution instead of the several ad
> >> hoc solutions we have now.
> >> 
> >> If we add asynchronous commands *now*, and for simple cases only, we add
> >> yet another special case for a future generic solution to integrate.
> >> I'm not going to let that happen.
> >> 
> >> I figure the closest to a generic solution we have is block jobs.
> >> Perhaps a generic solution could be had by abstracting away the "block"
> >> from "block jobs", leaving just "jobs".
> >
> > I don't know block jobs well enough to answer that.
> > I would suggest you could add a thread for asynchronous commands
> > and you could shuffle commands onto that thread as and when you feel
> > like it.
> >
> >> Another approach is generalizing the asynchronous command proposal to
> >> fully cover the not-so-simple cases.
> >> 
> >> If you'd rather want to make progress on monitor availability without
> >> cracking the "jobs" problem, you're in luck!  Use your license to "add
> >> more of the same": synchronous command to start a job, query to monitor,
> >> event to notify.  
> >> 
> >> If you insist on tying your monitor availability solution to
> >> asynchronous commands, then I'm in luck!  I just found volunteers to
> >> solve the "jobs" problem for me.
> >
> > I'm looking for minimal change here while keeping the door open for
> > the future, if there's anything you think we should do to make that
> > easy then tell us - but I'd rather this didn't turn into a 'fix all
> > known monitor problems' because frankly we may as well give up now.
> > So i don't see this as solving the 'jobs' problem, but if we can
> > do something to make it easier to solve in the future then lets do it.
> 
> Forget about asynchronous commands, jobs and the whole shebang of
> distractions, and consider what you really need: I suspect it could be
> out-of-band dispatch of non-blocking commands.  More on that in my reply
> to Daniel.

Right, and then the only thing we need to do is make sure the caller
doesn't get the replies to those commands mixed up with the replies
to the blocking commands.

Dave

--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

* Daniel P. Berrange (berrange@redhat.com) wrote:
> On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > This does imply that you need a separate monitor I/O processing, from the
> > > command execution thread, but I see no need for all commands to suddenly
> > > become async. Just allowing interleaved replies is sufficient from the
> > > POV of the protocol definition. This interleaving is easy to handle from
> > > the client POV - just requires a unique 'serial' in the request by the
> > > client, that is copied into the reply by QEMU.
> > 
> > OK, so for that we can just take Marc-André's syntax and call it 'id':
> >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> > 
> > then it's upto the caller to ensure those id's are unique.
> 
> Libvirt has in fact generated a unique 'id' for every monitor command
> since day 1 of supporting QMP.
> 
> > I do worry about two things:
> >   a) With this the caller doesn't really know which commands could be
> >   in parallel - for example if we've got a recovery command that's
> >   executed by this non-locking thread that's OK, we expect that
> >   to be doable in parallel.  If in the future though we do
> >   what you initially suggested and have a bunch of commands get
> >   routed to the migration thread (say) then those would suddenly
> >   operate in parallel with other commands that we're previously
> >   synchronous.
> 
> We could still have an opt-in for async commands. eg default to executing
> all commands in the main thread, unless the client issues an explicit
> "make it async" command, to switch to allowing the migration thread to
> process it async.
> 
>  { "execute": "qmp_allow_async",
>    "data": { "commands": [
>        "migrate_cancel",
>    ] } }
> 
> 
>  { "return": { "commands": [
>        "migrate_cancel",
>    ] } }
> 
> The server response contains the subset of commands from the request
> for which async is supported.
> 
> That gives good negotiation ability going forward as we incrementally
> support async on more commands.

Is that 'qmp_allow_async' a command purely to query whether a command
is async or is it a wrapper to cause that command to be executed async?

> >   b) I still worry how the various IO channels will behave on another
> >   thread.  But that's more a general feeling rather than anything
> >   specific.
> 
> The only complexity will be around making sure the Chardev code uses
> the right GMainContext for any watches on the underlying QIOChannel,
> so that we poll() from the custom thread instead of the main thread.
> IOW, as long as all I/O is done from the single thread everything
> should work fine.

Dave

> Regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

On Thu, Sep 07, 2017 at 11:04:02AM +0100, Dr. David Alan Gilbert wrote:
> * Daniel P. Berrange (berrange@redhat.com) wrote:
> > On Wed, Sep 06, 2017 at 12:31:58PM +0100, Dr. David Alan Gilbert wrote:
> > > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > > This does imply that you need a separate monitor I/O processing, from the
> > > > command execution thread, but I see no need for all commands to suddenly
> > > > become async. Just allowing interleaved replies is sufficient from the
> > > > POV of the protocol definition. This interleaving is easy to handle from
> > > > the client POV - just requires a unique 'serial' in the request by the
> > > > client, that is copied into the reply by QEMU.
> > > 
> > > OK, so for that we can just take Marc-André's syntax and call it 'id':
> > >   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> > > 
> > > then it's upto the caller to ensure those id's are unique.
> > 
> > Libvirt has in fact generated a unique 'id' for every monitor command
> > since day 1 of supporting QMP.
> > 
> > > I do worry about two things:
> > >   a) With this the caller doesn't really know which commands could be
> > >   in parallel - for example if we've got a recovery command that's
> > >   executed by this non-locking thread that's OK, we expect that
> > >   to be doable in parallel.  If in the future though we do
> > >   what you initially suggested and have a bunch of commands get
> > >   routed to the migration thread (say) then those would suddenly
> > >   operate in parallel with other commands that we're previously
> > >   synchronous.
> > 
> > We could still have an opt-in for async commands. eg default to executing
> > all commands in the main thread, unless the client issues an explicit
> > "make it async" command, to switch to allowing the migration thread to
> > process it async.
> > 
> >  { "execute": "qmp_allow_async",
> >    "data": { "commands": [
> >        "migrate_cancel",
> >    ] } }
> > 
> > 
> >  { "return": { "commands": [
> >        "migrate_cancel",
> >    ] } }
> > 
> > The server response contains the subset of commands from the request
> > for which async is supported.
> > 
> > That gives good negotiation ability going forward as we incrementally
> > support async on more commands.
> 
> Is that 'qmp_allow_async' a command purely to query whether a command
> is async or is it a wrapper to cause that command to be executed async?

The former.

It merely used by the client to tell QEMU that it wants the command(s)
listed to have async processing enabled. QEMU reports back which commands
it has actually enabled async for.

IOW, before executing this, everything is still processed synchronously,
even if QEMU has support for async. This ensures back compat as we enable
support for async per command. After executing this command, then future
usage of 'migrate_cancel' would be run async.


Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

On 09/06/2017 06:31 AM, Dr. David Alan Gilbert wrote:

>> This does imply that you need a separate monitor I/O processing, from the
>> command execution thread, but I see no need for all commands to suddenly
>> become async. Just allowing interleaved replies is sufficient from the
>> POV of the protocol definition. This interleaving is easy to handle from
>> the client POV - just requires a unique 'serial' in the request by the
>> client, that is copied into the reply by QEMU.
> 
> OK, so for that we can just take Marc-André's syntax and call it 'id':
>   https://lists.gnu.org/archive/html/qemu-devel/2017-01/msg03634.html
> 
> then it's upto the caller to ensure those id's are unique.

We ALREADY support 'id', and it is already up to the caller to ensure
those id's are unique, even without Marc-André's additions.

> 
> I do worry about two things:
>   a) With this the caller doesn't really know which commands could be
>   in parallel - for example if we've got a recovery command that's
>   executed by this non-locking thread that's OK, we expect that
>   to be doable in parallel.  If in the future though we do
>   what you initially suggested and have a bunch of commands get
>   routed to the migration thread (say) then those would suddenly
>   operate in parallel with other commands that we're previously
>   synchronous.

Presumably, all existing commands are NOT async, and introspection via
query-qmp-schema will let you query which new commands ARE async.  Or
existing commands will gain an optional parameter to opt-in to async
behavior for that command, defaulting to sync by default.  Thus, an old
libvirt will never call an async command, and never notice the
difference, but a new libvirt that is aware of async commands will opt
in to the commands that it wants to use in an async manner.

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3266
Virtualization:  qemu.org | libvirt.org

On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> The root problem is that, monitor commands are all handled in main
> loop thread now, no matter how many monitors we specify. And, if main
> loop thread hangs due to some reason, all monitors will be stuck.

I see a larger issue with postcopy: existing QEMU code assumes that
guest memory access is instantaneous.

Postcopy breaks this assumption and introduces blocking points that can
now take unbounded time.

This problem isn't specific to the monitor.  It can also happen to other
components in QEMU like the gdbstub.

Do we need an asynchronous memory API?  Synchronous memory access should
only be allowed in vcpu threads.

Stefan

* Stefan Hajnoczi (stefanha@gmail.com) wrote:
> On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> > The root problem is that, monitor commands are all handled in main
> > loop thread now, no matter how many monitors we specify. And, if main
> > loop thread hangs due to some reason, all monitors will be stuck.
> 
> I see a larger issue with postcopy: existing QEMU code assumes that
> guest memory access is instantaneous.
> 
> Postcopy breaks this assumption and introduces blocking points that can
> now take unbounded time.
> 
> This problem isn't specific to the monitor.  It can also happen to other
> components in QEMU like the gdbstub.
> 
> Do we need an asynchronous memory API?  Synchronous memory access should
> only be allowed in vcpu threads.

It would probably be useful for gdbstub where the overhead of async
doesn't matter;  but doing that for all IO emulation is hard.

Dave

> Stefan
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

On Wed, Sep 06, 2017 at 04:14:37PM +0100, Dr. David Alan Gilbert wrote:
> * Stefan Hajnoczi (stefanha@gmail.com) wrote:
> > On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> > > The root problem is that, monitor commands are all handled in main
> > > loop thread now, no matter how many monitors we specify. And, if main
> > > loop thread hangs due to some reason, all monitors will be stuck.
> > 
> > I see a larger issue with postcopy: existing QEMU code assumes that
> > guest memory access is instantaneous.
> > 
> > Postcopy breaks this assumption and introduces blocking points that can
> > now take unbounded time.
> > 
> > This problem isn't specific to the monitor.  It can also happen to other
> > components in QEMU like the gdbstub.
> > 
> > Do we need an asynchronous memory API?  Synchronous memory access should
> > only be allowed in vcpu threads.
> 
> It would probably be useful for gdbstub where the overhead of async
> doesn't matter;  but doing that for all IO emulation is hard.

Agreed.

IIUC one problem is that we should have code that cached the HVA for
specific GPA, then when it wants to write to that GPA, it directly
writes to corresponding HVA.  No memory API is used.  I am not sure
whether it's possible to convert all these usages into memory APIs
(even if it supports async operations).

-- 
Peter Xu

On Wed, Sep 6, 2017 at 4:14 PM, Dr. David Alan Gilbert
<dgilbert@redhat.com> wrote:
> * Stefan Hajnoczi (stefanha@gmail.com) wrote:
>> On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
>> > The root problem is that, monitor commands are all handled in main
>> > loop thread now, no matter how many monitors we specify. And, if main
>> > loop thread hangs due to some reason, all monitors will be stuck.
>>
>> I see a larger issue with postcopy: existing QEMU code assumes that
>> guest memory access is instantaneous.
>>
>> Postcopy breaks this assumption and introduces blocking points that can
>> now take unbounded time.
>>
>> This problem isn't specific to the monitor.  It can also happen to other
>> components in QEMU like the gdbstub.
>>
>> Do we need an asynchronous memory API?  Synchronous memory access should
>> only be allowed in vcpu threads.
>
> It would probably be useful for gdbstub where the overhead of async
> doesn't matter;  but doing that for all IO emulation is hard.

Why is it hard?

Memory access can be synchronous in the vcpu thread.  That eliminates
a lot of code straight away.

Anything using dma-helpers.c is already async.  They just don't know
that the memory access part is being made async too :).

The remaining cases are virtio and some other devices.

If you are worried about performance, the first rule is that async
memory access is only needed on the destination side when post-copy is
active.  Maybe use setjmp to return from the signal handler and queue
a callback for when the page has been loaded.

Stefan

* Stefan Hajnoczi (stefanha@gmail.com) wrote:
> On Wed, Sep 6, 2017 at 4:14 PM, Dr. David Alan Gilbert
> <dgilbert@redhat.com> wrote:
> > * Stefan Hajnoczi (stefanha@gmail.com) wrote:
> >> On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> >> > The root problem is that, monitor commands are all handled in main
> >> > loop thread now, no matter how many monitors we specify. And, if main
> >> > loop thread hangs due to some reason, all monitors will be stuck.
> >>
> >> I see a larger issue with postcopy: existing QEMU code assumes that
> >> guest memory access is instantaneous.
> >>
> >> Postcopy breaks this assumption and introduces blocking points that can
> >> now take unbounded time.
> >>
> >> This problem isn't specific to the monitor.  It can also happen to other
> >> components in QEMU like the gdbstub.
> >>
> >> Do we need an asynchronous memory API?  Synchronous memory access should
> >> only be allowed in vcpu threads.
> >
> > It would probably be useful for gdbstub where the overhead of async
> > doesn't matter;  but doing that for all IO emulation is hard.
> 
> Why is it hard?
> 
> Memory access can be synchronous in the vcpu thread.  That eliminates
> a lot of code straight away.
> 
> Anything using dma-helpers.c is already async.  They just don't know
> that the memory access part is being made async too :).

Can you point me to some info on that ?

> The remaining cases are virtio and some other devices.
> 
> If you are worried about performance, the first rule is that async
> memory access is only needed on the destination side when post-copy is
> active.  Maybe use setjmp to return from the signal handler and queue
> a callback for when the page has been loaded.

I'm not sure it's worth trying to be too clever at avoiding this;
I see the fact that we're doing IO with the bql held as a more
fundamental problem.

Dave

> Stefan
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

On Thu, Sep 7, 2017 at 10:35 AM, Dr. David Alan Gilbert
<dgilbert@redhat.com> wrote:
> * Stefan Hajnoczi (stefanha@gmail.com) wrote:
>> On Wed, Sep 6, 2017 at 4:14 PM, Dr. David Alan Gilbert
>> <dgilbert@redhat.com> wrote:
>> > * Stefan Hajnoczi (stefanha@gmail.com) wrote:
>> >> On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
>> >> > The root problem is that, monitor commands are all handled in main
>> >> > loop thread now, no matter how many monitors we specify. And, if main
>> >> > loop thread hangs due to some reason, all monitors will be stuck.
>> >>
>> >> I see a larger issue with postcopy: existing QEMU code assumes that
>> >> guest memory access is instantaneous.
>> >>
>> >> Postcopy breaks this assumption and introduces blocking points that can
>> >> now take unbounded time.
>> >>
>> >> This problem isn't specific to the monitor.  It can also happen to other
>> >> components in QEMU like the gdbstub.
>> >>
>> >> Do we need an asynchronous memory API?  Synchronous memory access should
>> >> only be allowed in vcpu threads.
>> >
>> > It would probably be useful for gdbstub where the overhead of async
>> > doesn't matter;  but doing that for all IO emulation is hard.
>>
>> Why is it hard?
>>
>> Memory access can be synchronous in the vcpu thread.  That eliminates
>> a lot of code straight away.
>>
>> Anything using dma-helpers.c is already async.  They just don't know
>> that the memory access part is being made async too :).
>
> Can you point me to some info on that ?

IDE and SCSI use dma-helpers.c to perform I/O:
hw/ide/core.c:892:        s->bus->dma->aiocb =
dma_blk_io(blk_get_aio_context(s->blk),
hw/ide/macio.c:189:        s->bus->dma->aiocb =
dma_blk_io(blk_get_aio_context(s->blk), &s->sg,
hw/scsi/scsi-disk.c:348:        r->req.aiocb =
dma_blk_io(blk_get_aio_context(s->qdev.conf.blk),
hw/scsi/scsi-disk.c:551:        r->req.aiocb =
dma_blk_io(blk_get_aio_context(s->qdev.conf.blk),

They pass a scatter-gather list of guest RAM addresses to
dma-helpers.c.  They receive a callback when I/O has finished.

Try following the code path.  Request submission may be from a vcpu
thread or IOThread.  Completion occurs in the main loop or an
IOThread.

The main point is that this API is already asynchronous.  If any
changes are needed for async guest memory access (not sure, I haven't
checked), then at least the dma-helpers.c users do not need to be
modified.

>> The remaining cases are virtio and some other devices.
>>
>> If you are worried about performance, the first rule is that async
>> memory access is only needed on the destination side when post-copy is
>> active.  Maybe use setjmp to return from the signal handler and queue
>> a callback for when the page has been loaded.
>
> I'm not sure it's worth trying to be too clever at avoiding this;
> I see the fact that we're doing IO with the bql held as a more
> fundamental problem.

QEMU should be doing I/O syscalls in async fashion or threadpool
workers (no BQL) so the BQL is not an issue.  Anything else could
cause unbounded waits even without postcopy.

Can you explain what you are worried about?

Stefan

On Thu, Sep 07, 2017 at 11:09:29AM +0100, Stefan Hajnoczi wrote:
> On Thu, Sep 7, 2017 at 10:35 AM, Dr. David Alan Gilbert
> <dgilbert@redhat.com> wrote:
> > * Stefan Hajnoczi (stefanha@gmail.com) wrote:
> >> On Wed, Sep 6, 2017 at 4:14 PM, Dr. David Alan Gilbert
> >> <dgilbert@redhat.com> wrote:
> >> > * Stefan Hajnoczi (stefanha@gmail.com) wrote:
> >> >> On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> >> >> > The root problem is that, monitor commands are all handled in main
> >> >> > loop thread now, no matter how many monitors we specify. And, if main
> >> >> > loop thread hangs due to some reason, all monitors will be stuck.
> >> >>
> >> >> I see a larger issue with postcopy: existing QEMU code assumes that
> >> >> guest memory access is instantaneous.
> >> >>
> >> >> Postcopy breaks this assumption and introduces blocking points that can
> >> >> now take unbounded time.
> >> >>
> >> >> This problem isn't specific to the monitor.  It can also happen to other
> >> >> components in QEMU like the gdbstub.
> >> >>
> >> >> Do we need an asynchronous memory API?  Synchronous memory access should
> >> >> only be allowed in vcpu threads.
> >> >
> >> > It would probably be useful for gdbstub where the overhead of async
> >> > doesn't matter;  but doing that for all IO emulation is hard.
> >>
> >> Why is it hard?
> >>
> >> Memory access can be synchronous in the vcpu thread.  That eliminates
> >> a lot of code straight away.
> >>
> >> Anything using dma-helpers.c is already async.  They just don't know
> >> that the memory access part is being made async too :).
> >
> > Can you point me to some info on that ?
> 
> IDE and SCSI use dma-helpers.c to perform I/O:
> hw/ide/core.c:892:        s->bus->dma->aiocb =
> dma_blk_io(blk_get_aio_context(s->blk),
> hw/ide/macio.c:189:        s->bus->dma->aiocb =
> dma_blk_io(blk_get_aio_context(s->blk), &s->sg,
> hw/scsi/scsi-disk.c:348:        r->req.aiocb =
> dma_blk_io(blk_get_aio_context(s->qdev.conf.blk),
> hw/scsi/scsi-disk.c:551:        r->req.aiocb =
> dma_blk_io(blk_get_aio_context(s->qdev.conf.blk),
> 
> They pass a scatter-gather list of guest RAM addresses to
> dma-helpers.c.  They receive a callback when I/O has finished.
> 
> Try following the code path.  Request submission may be from a vcpu
> thread or IOThread.  Completion occurs in the main loop or an
> IOThread.
> 
> The main point is that this API is already asynchronous.  If any
> changes are needed for async guest memory access (not sure, I haven't
> checked), then at least the dma-helpers.c users do not need to be
> modified.
> 
> >> The remaining cases are virtio and some other devices.
> >>
> >> If you are worried about performance, the first rule is that async
> >> memory access is only needed on the destination side when post-copy is
> >> active.  Maybe use setjmp to return from the signal handler and queue
> >> a callback for when the page has been loaded.
> >
> > I'm not sure it's worth trying to be too clever at avoiding this;
> > I see the fact that we're doing IO with the bql held as a more
> > fundamental problem.
> 
> QEMU should be doing I/O syscalls in async fashion or threadpool
> workers (no BQL) so the BQL is not an issue.  Anything else could
> cause unbounded waits even without postcopy.

E.g. when vcpu got page faulted with BQL taken, while the main thread
needs the BQL to dispatch anything, including monitor commands.

So I think it's a multiplex problem - we need to solve both (1) main
thread accessing guest memories which is still missing, and (2) BQL
deadlocks between vcpu threads and main thread.

Thanks,

-- 
Peter Xu

On Thu, Sep 7, 2017 at 1:02 PM, Peter Xu <peterx@redhat.com> wrote:
> On Thu, Sep 07, 2017 at 11:09:29AM +0100, Stefan Hajnoczi wrote:
>> On Thu, Sep 7, 2017 at 10:35 AM, Dr. David Alan Gilbert
>> <dgilbert@redhat.com> wrote:
>> > * Stefan Hajnoczi (stefanha@gmail.com) wrote:
>> >> On Wed, Sep 6, 2017 at 4:14 PM, Dr. David Alan Gilbert
>> >> <dgilbert@redhat.com> wrote:
>> >> > * Stefan Hajnoczi (stefanha@gmail.com) wrote:
>> >> >> On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
>> >> >> > The root problem is that, monitor commands are all handled in main
>> >> >> > loop thread now, no matter how many monitors we specify. And, if main
>> >> >> > loop thread hangs due to some reason, all monitors will be stuck.
>> >> >>
>> >> >> I see a larger issue with postcopy: existing QEMU code assumes that
>> >> >> guest memory access is instantaneous.
>> >> >>
>> >> >> Postcopy breaks this assumption and introduces blocking points that can
>> >> >> now take unbounded time.
>> >> >>
>> >> >> This problem isn't specific to the monitor.  It can also happen to other
>> >> >> components in QEMU like the gdbstub.
>> >> >>
>> >> >> Do we need an asynchronous memory API?  Synchronous memory access should
>> >> >> only be allowed in vcpu threads.
>> >> >
>> >> > It would probably be useful for gdbstub where the overhead of async
>> >> > doesn't matter;  but doing that for all IO emulation is hard.
>> >>
>> >> Why is it hard?
>> >>
>> >> Memory access can be synchronous in the vcpu thread.  That eliminates
>> >> a lot of code straight away.
>> >>
>> >> Anything using dma-helpers.c is already async.  They just don't know
>> >> that the memory access part is being made async too :).
>> >
>> > Can you point me to some info on that ?
>>
>> IDE and SCSI use dma-helpers.c to perform I/O:
>> hw/ide/core.c:892:        s->bus->dma->aiocb =
>> dma_blk_io(blk_get_aio_context(s->blk),
>> hw/ide/macio.c:189:        s->bus->dma->aiocb =
>> dma_blk_io(blk_get_aio_context(s->blk), &s->sg,
>> hw/scsi/scsi-disk.c:348:        r->req.aiocb =
>> dma_blk_io(blk_get_aio_context(s->qdev.conf.blk),
>> hw/scsi/scsi-disk.c:551:        r->req.aiocb =
>> dma_blk_io(blk_get_aio_context(s->qdev.conf.blk),
>>
>> They pass a scatter-gather list of guest RAM addresses to
>> dma-helpers.c.  They receive a callback when I/O has finished.
>>
>> Try following the code path.  Request submission may be from a vcpu
>> thread or IOThread.  Completion occurs in the main loop or an
>> IOThread.
>>
>> The main point is that this API is already asynchronous.  If any
>> changes are needed for async guest memory access (not sure, I haven't
>> checked), then at least the dma-helpers.c users do not need to be
>> modified.
>>
>> >> The remaining cases are virtio and some other devices.
>> >>
>> >> If you are worried about performance, the first rule is that async
>> >> memory access is only needed on the destination side when post-copy is
>> >> active.  Maybe use setjmp to return from the signal handler and queue
>> >> a callback for when the page has been loaded.
>> >
>> > I'm not sure it's worth trying to be too clever at avoiding this;
>> > I see the fact that we're doing IO with the bql held as a more
>> > fundamental problem.
>>
>> QEMU should be doing I/O syscalls in async fashion or threadpool
>> workers (no BQL) so the BQL is not an issue.  Anything else could
>> cause unbounded waits even without postcopy.
>
> E.g. when vcpu got page faulted with BQL taken, while the main thread
> needs the BQL to dispatch anything, including monitor commands.
>
> So I think it's a multiplex problem - we need to solve both (1) main
> thread accessing guest memories which is still missing, and (2) BQL
> deadlocks between vcpu threads and main thread.

I think we need a single solution and cannot treat these as separate.
This is because the same virtio device emulation code may run in 3
contexts:
1. vcpu thread (ioeventfd=off)
2. main loop thread (ioeventfd=on)
3. IOThread (ioeventfd=on, iothread=<id>)

If you try to solve them separately then the code won't work in all 3
contexts anymore.

Stefan

* Stefan Hajnoczi (stefanha@gmail.com) wrote:
> On Thu, Sep 7, 2017 at 1:02 PM, Peter Xu <peterx@redhat.com> wrote:
> > On Thu, Sep 07, 2017 at 11:09:29AM +0100, Stefan Hajnoczi wrote:
> >> On Thu, Sep 7, 2017 at 10:35 AM, Dr. David Alan Gilbert
> >> <dgilbert@redhat.com> wrote:
> >> > * Stefan Hajnoczi (stefanha@gmail.com) wrote:
> >> >> On Wed, Sep 6, 2017 at 4:14 PM, Dr. David Alan Gilbert
> >> >> <dgilbert@redhat.com> wrote:
> >> >> > * Stefan Hajnoczi (stefanha@gmail.com) wrote:
> >> >> >> On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
> >> >> >> > The root problem is that, monitor commands are all handled in main
> >> >> >> > loop thread now, no matter how many monitors we specify. And, if main
> >> >> >> > loop thread hangs due to some reason, all monitors will be stuck.
> >> >> >>
> >> >> >> I see a larger issue with postcopy: existing QEMU code assumes that
> >> >> >> guest memory access is instantaneous.
> >> >> >>
> >> >> >> Postcopy breaks this assumption and introduces blocking points that can
> >> >> >> now take unbounded time.
> >> >> >>
> >> >> >> This problem isn't specific to the monitor.  It can also happen to other
> >> >> >> components in QEMU like the gdbstub.
> >> >> >>
> >> >> >> Do we need an asynchronous memory API?  Synchronous memory access should
> >> >> >> only be allowed in vcpu threads.
> >> >> >
> >> >> > It would probably be useful for gdbstub where the overhead of async
> >> >> > doesn't matter;  but doing that for all IO emulation is hard.
> >> >>
> >> >> Why is it hard?
> >> >>
> >> >> Memory access can be synchronous in the vcpu thread.  That eliminates
> >> >> a lot of code straight away.
> >> >>
> >> >> Anything using dma-helpers.c is already async.  They just don't know
> >> >> that the memory access part is being made async too :).
> >> >
> >> > Can you point me to some info on that ?
> >>
> >> IDE and SCSI use dma-helpers.c to perform I/O:
> >> hw/ide/core.c:892:        s->bus->dma->aiocb =
> >> dma_blk_io(blk_get_aio_context(s->blk),
> >> hw/ide/macio.c:189:        s->bus->dma->aiocb =
> >> dma_blk_io(blk_get_aio_context(s->blk), &s->sg,
> >> hw/scsi/scsi-disk.c:348:        r->req.aiocb =
> >> dma_blk_io(blk_get_aio_context(s->qdev.conf.blk),
> >> hw/scsi/scsi-disk.c:551:        r->req.aiocb =
> >> dma_blk_io(blk_get_aio_context(s->qdev.conf.blk),
> >>
> >> They pass a scatter-gather list of guest RAM addresses to
> >> dma-helpers.c.  They receive a callback when I/O has finished.
> >>
> >> Try following the code path.  Request submission may be from a vcpu
> >> thread or IOThread.  Completion occurs in the main loop or an
> >> IOThread.
> >>
> >> The main point is that this API is already asynchronous.  If any
> >> changes are needed for async guest memory access (not sure, I haven't
> >> checked), then at least the dma-helpers.c users do not need to be
> >> modified.
> >>
> >> >> The remaining cases are virtio and some other devices.
> >> >>
> >> >> If you are worried about performance, the first rule is that async
> >> >> memory access is only needed on the destination side when post-copy is
> >> >> active.  Maybe use setjmp to return from the signal handler and queue
> >> >> a callback for when the page has been loaded.
> >> >
> >> > I'm not sure it's worth trying to be too clever at avoiding this;
> >> > I see the fact that we're doing IO with the bql held as a more
> >> > fundamental problem.
> >>
> >> QEMU should be doing I/O syscalls in async fashion or threadpool
> >> workers (no BQL) so the BQL is not an issue.  Anything else could
> >> cause unbounded waits even without postcopy.
> >
> > E.g. when vcpu got page faulted with BQL taken, while the main thread
> > needs the BQL to dispatch anything, including monitor commands.
> >
> > So I think it's a multiplex problem - we need to solve both (1) main
> > thread accessing guest memories which is still missing, and (2) BQL
> > deadlocks between vcpu threads and main thread.
> 
> I think we need a single solution and cannot treat these as separate.
> This is because the same virtio device emulation code may run in 3
> contexts:
> 1. vcpu thread (ioeventfd=off)
> 2. main loop thread (ioeventfd=on)
> 3. IOThread (ioeventfd=on, iothread=<id>)
> 
> If you try to solve them separately then the code won't work in all 3
> contexts anymore.

I think you can also get main loop thread hangs on things like
network packet reception.

Dave

> 
> Stefan
> 
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

On Thu, Sep 7, 2017 at 6:14 PM, Dr. David Alan Gilbert
<dgilbert@redhat.com> wrote:
> * Stefan Hajnoczi (stefanha@gmail.com) wrote:
>> On Thu, Sep 7, 2017 at 1:02 PM, Peter Xu <peterx@redhat.com> wrote:
>> > On Thu, Sep 07, 2017 at 11:09:29AM +0100, Stefan Hajnoczi wrote:
>> >> On Thu, Sep 7, 2017 at 10:35 AM, Dr. David Alan Gilbert
>> >> <dgilbert@redhat.com> wrote:
>> >> > * Stefan Hajnoczi (stefanha@gmail.com) wrote:
>> >> >> On Wed, Sep 6, 2017 at 4:14 PM, Dr. David Alan Gilbert
>> >> >> <dgilbert@redhat.com> wrote:
>> >> >> > * Stefan Hajnoczi (stefanha@gmail.com) wrote:
>> >> >> >> On Wed, Aug 23, 2017 at 02:51:03PM +0800, Peter Xu wrote:
>> >> >> >> > The root problem is that, monitor commands are all handled in main
>> >> >> >> > loop thread now, no matter how many monitors we specify. And, if main
>> >> >> >> > loop thread hangs due to some reason, all monitors will be stuck.
>> >> >> >>
>> >> >> >> I see a larger issue with postcopy: existing QEMU code assumes that
>> >> >> >> guest memory access is instantaneous.
>> >> >> >>
>> >> >> >> Postcopy breaks this assumption and introduces blocking points that can
>> >> >> >> now take unbounded time.
>> >> >> >>
>> >> >> >> This problem isn't specific to the monitor.  It can also happen to other
>> >> >> >> components in QEMU like the gdbstub.
>> >> >> >>
>> >> >> >> Do we need an asynchronous memory API?  Synchronous memory access should
>> >> >> >> only be allowed in vcpu threads.
>> >> >> >
>> >> >> > It would probably be useful for gdbstub where the overhead of async
>> >> >> > doesn't matter;  but doing that for all IO emulation is hard.
>> >> >>
>> >> >> Why is it hard?
>> >> >>
>> >> >> Memory access can be synchronous in the vcpu thread.  That eliminates
>> >> >> a lot of code straight away.
>> >> >>
>> >> >> Anything using dma-helpers.c is already async.  They just don't know
>> >> >> that the memory access part is being made async too :).
>> >> >
>> >> > Can you point me to some info on that ?
>> >>
>> >> IDE and SCSI use dma-helpers.c to perform I/O:
>> >> hw/ide/core.c:892:        s->bus->dma->aiocb =
>> >> dma_blk_io(blk_get_aio_context(s->blk),
>> >> hw/ide/macio.c:189:        s->bus->dma->aiocb =
>> >> dma_blk_io(blk_get_aio_context(s->blk), &s->sg,
>> >> hw/scsi/scsi-disk.c:348:        r->req.aiocb =
>> >> dma_blk_io(blk_get_aio_context(s->qdev.conf.blk),
>> >> hw/scsi/scsi-disk.c:551:        r->req.aiocb =
>> >> dma_blk_io(blk_get_aio_context(s->qdev.conf.blk),
>> >>
>> >> They pass a scatter-gather list of guest RAM addresses to
>> >> dma-helpers.c.  They receive a callback when I/O has finished.
>> >>
>> >> Try following the code path.  Request submission may be from a vcpu
>> >> thread or IOThread.  Completion occurs in the main loop or an
>> >> IOThread.
>> >>
>> >> The main point is that this API is already asynchronous.  If any
>> >> changes are needed for async guest memory access (not sure, I haven't
>> >> checked), then at least the dma-helpers.c users do not need to be
>> >> modified.
>> >>
>> >> >> The remaining cases are virtio and some other devices.
>> >> >>
>> >> >> If you are worried about performance, the first rule is that async
>> >> >> memory access is only needed on the destination side when post-copy is
>> >> >> active.  Maybe use setjmp to return from the signal handler and queue
>> >> >> a callback for when the page has been loaded.
>> >> >
>> >> > I'm not sure it's worth trying to be too clever at avoiding this;
>> >> > I see the fact that we're doing IO with the bql held as a more
>> >> > fundamental problem.
>> >>
>> >> QEMU should be doing I/O syscalls in async fashion or threadpool
>> >> workers (no BQL) so the BQL is not an issue.  Anything else could
>> >> cause unbounded waits even without postcopy.
>> >
>> > E.g. when vcpu got page faulted with BQL taken, while the main thread
>> > needs the BQL to dispatch anything, including monitor commands.
>> >
>> > So I think it's a multiplex problem - we need to solve both (1) main
>> > thread accessing guest memories which is still missing, and (2) BQL
>> > deadlocks between vcpu threads and main thread.
>>
>> I think we need a single solution and cannot treat these as separate.
>> This is because the same virtio device emulation code may run in 3
>> contexts:
>> 1. vcpu thread (ioeventfd=off)
>> 2. main loop thread (ioeventfd=on)
>> 3. IOThread (ioeventfd=on, iothread=<id>)
>>
>> If you try to solve them separately then the code won't work in all 3
>> contexts anymore.
>
> I think you can also get main loop thread hangs on things like
> network packet reception.

That is case #2.  The QEMU net subsystem reads receive packets into a
temporary buffer (it's not zero-copy) and invokes the virtio-net
receive handler function from the main loop.

Stefan