[PATCH v2 0/3] mm: Make unregistration of super_block shrinker more faster

Kirill Tkhai posted 3 patches 2 years, 8 months ago
fs/super.c               |    3 ++-
include/linux/shrinker.h |    4 ++++
mm/vmscan.c              |   39 +++++++++++++++++++++++++++++++--------
3 files changed, 37 insertions(+), 9 deletions(-)
[PATCH v2 0/3] mm: Make unregistration of super_block shrinker more faster
Posted by Kirill Tkhai 2 years, 8 months ago
This patch set introduces a new scheme of shrinker unregistration. It allows to split
the unregistration in two parts: fast and slow. This allows to hide slow part from
a user, so user-visible unregistration becomes fast.

This fixes the -88.8% regression of stress-ng.ramfs.ops_per_sec noticed
by kernel test robot:

https://lore.kernel.org/lkml/202305230837.db2c233f-yujie.liu@intel.com/

---

Kirill Tkhai (2):
      mm: Split unregister_shrinker() in fast and slow part
      fs: Use delayed shrinker unregistration

Qi Zheng (1):
      mm: vmscan: move shrinker_debugfs_remove() before synchronize_srcu()


 fs/super.c               |    3 ++-
 include/linux/shrinker.h |    4 ++++
 mm/vmscan.c              |   39 +++++++++++++++++++++++++++++++--------
 3 files changed, 37 insertions(+), 9 deletions(-)

--
Signed-off-by: Kirill Tkhai <tkhai@ya.ru>
Re: [PATCH v2 0/3] mm: Make unregistration of super_block shrinker more faster
Posted by Dave Chinner 2 years, 8 months ago
On Mon, Jun 05, 2023 at 10:02:46PM +0300, Kirill Tkhai wrote:
> This patch set introduces a new scheme of shrinker unregistration. It allows to split
> the unregistration in two parts: fast and slow. This allows to hide slow part from
> a user, so user-visible unregistration becomes fast.
> 
> This fixes the -88.8% regression of stress-ng.ramfs.ops_per_sec noticed
> by kernel test robot:
> 
> https://lore.kernel.org/lkml/202305230837.db2c233f-yujie.liu@intel.com/
> 
> ---
> 
> Kirill Tkhai (2):
>       mm: Split unregister_shrinker() in fast and slow part
>       fs: Use delayed shrinker unregistration

Did you test any filesystem other than ramfs?

Filesystems more complex than ramfs have internal shrinkers, and so
they will still be running the slow synchronize_srcu() - potentially
multiple times! - in every unmount. Both XFS and ext4 have 3
internal shrinker instances per mount, so they will still call
synchronize_srcu() at least 3 times per unmount after this change.

What about any other subsystem that runs a shrinker - do they have
context depedent shrinker instances that get frequently created and
destroyed? They'll need the same treatment.

Seriously, part of changing shrinker infrastructure is doing an
audit of all the shrinker instances to determine how the change will
impact those shrinkers, and if the same structural changes are
needed to those implementations.

I don't see any of this being done - this looks like a "slap a bandaid
over the visible symptom" patch set without any deeper investigation
of the scope of the issue having been gained.

Along with all shrinkers now running under a SRCU critical region
and requiring a machine wide synchronisation point for every
unregister_shrinker() call made, the ability to repeated abort
global shrinker passes via external SRCU expediting, and now an
intricate locking and state dance in do_shrink_slab() vs
unregister_shrinker, I can't say I'm particularly liking any of
this, regardles of the benefits it supposedly provides.

-Dave.
-- 
Dave Chinner
david@fromorbit.com
Re: [PATCH v2 0/3] mm: Make unregistration of super_block shrinker more faster
Posted by Kirill Tkhai 2 years, 8 months ago
On 06.06.2023 01:32, Dave Chinner wrote:
> On Mon, Jun 05, 2023 at 10:02:46PM +0300, Kirill Tkhai wrote:
>> This patch set introduces a new scheme of shrinker unregistration. It allows to split
>> the unregistration in two parts: fast and slow. This allows to hide slow part from
>> a user, so user-visible unregistration becomes fast.
>>
>> This fixes the -88.8% regression of stress-ng.ramfs.ops_per_sec noticed
>> by kernel test robot:
>>
>> https://lore.kernel.org/lkml/202305230837.db2c233f-yujie.liu@intel.com/
>>
>> ---
>>
>> Kirill Tkhai (2):
>>       mm: Split unregister_shrinker() in fast and slow part
>>       fs: Use delayed shrinker unregistration
> 
> Did you test any filesystem other than ramfs?
> 
> Filesystems more complex than ramfs have internal shrinkers, and so
> they will still be running the slow synchronize_srcu() - potentially
> multiple times! - in every unmount. Both XFS and ext4 have 3
> internal shrinker instances per mount, so they will still call
> synchronize_srcu() at least 3 times per unmount after this change.
> 
> What about any other subsystem that runs a shrinker - do they have
> context depedent shrinker instances that get frequently created and
> destroyed? They'll need the same treatment.

Of course, all of shrinkers should be fixed. This patch set just aims to describe
the idea more wider, because I'm not sure most people read replys to kernel robot reports.

This is my suggestion of way to go. Probably, Qi is right person to ask whether
we're going to extend this and to maintain f95bdb700bc6 in tree.

There is not much time. Unfortunately, kernel test robot reported this significantly late.

> Seriously, part of changing shrinker infrastructure is doing an
> audit of all the shrinker instances to determine how the change will
> impact those shrinkers, and if the same structural changes are
> needed to those implementations.
> 
> I don't see any of this being done - this looks like a "slap a bandaid
> over the visible symptom" patch set without any deeper investigation
> of the scope of the issue having been gained.
> 
> Along with all shrinkers now running under a SRCU critical region
> and requiring a machine wide synchronisation point for every
> unregister_shrinker() call made, the ability to repeated abort
> global shrinker passes via external SRCU expediting, and now an
> intricate locking and state dance in do_shrink_slab() vs
> unregister_shrinker, I can't say I'm particularly liking any of
> this, regardles of the benefits it supposedly provides.
> 
> -Dave.
Re: [PATCH v2 0/3] mm: Make unregistration of super_block shrinker more faster
Posted by Dave Chinner 2 years, 8 months ago
On Wed, Jun 07, 2023 at 12:06:03AM +0300, Kirill Tkhai wrote:
> On 06.06.2023 01:32, Dave Chinner wrote:
> > On Mon, Jun 05, 2023 at 10:02:46PM +0300, Kirill Tkhai wrote:
> >> This patch set introduces a new scheme of shrinker unregistration. It allows to split
> >> the unregistration in two parts: fast and slow. This allows to hide slow part from
> >> a user, so user-visible unregistration becomes fast.
> >>
> >> This fixes the -88.8% regression of stress-ng.ramfs.ops_per_sec noticed
> >> by kernel test robot:
> >>
> >> https://lore.kernel.org/lkml/202305230837.db2c233f-yujie.liu@intel.com/
> >>
> >> ---
> >>
> >> Kirill Tkhai (2):
> >>       mm: Split unregister_shrinker() in fast and slow part
> >>       fs: Use delayed shrinker unregistration
> > 
> > Did you test any filesystem other than ramfs?
> > 
> > Filesystems more complex than ramfs have internal shrinkers, and so
> > they will still be running the slow synchronize_srcu() - potentially
> > multiple times! - in every unmount. Both XFS and ext4 have 3
> > internal shrinker instances per mount, so they will still call
> > synchronize_srcu() at least 3 times per unmount after this change.
> > 
> > What about any other subsystem that runs a shrinker - do they have
> > context depedent shrinker instances that get frequently created and
> > destroyed? They'll need the same treatment.
> 
> Of course, all of shrinkers should be fixed. This patch set just aims to describe
> the idea more wider, because I'm not sure most people read replys to kernel robot reports.
> 
> This is my suggestion of way to go. Probably, Qi is right person to ask whether
> we're going to extend this and to maintain f95bdb700bc6 in tree.
> 
> There is not much time. Unfortunately, kernel test robot reported this significantly late.

And that's why it should be reverted rather than trying to rush to
try to fix it.

I'm kind of tired of finding out about mm reclaim regressions only
when I see patches making naive and/or broken changes to subsystem
shrinker implementations without any real clue about what they are
doing.  If people/subsystems who maintain shrinker implementations
were cc'd on the changes to the shrinker implementation, this would
have all been resolved before merging occurred....

Lockless shrinker lists need a heap of supporting changes to be done
first so that they aren't reliant on synchronise_srcu() *at all*. If
the code was properly designed in the first place (i.e. dynamic
shrinker structures freed via call_rcu()), we wouldn't be in rushing
to fix weird regressions right now. 

Can we please revert this and start again with a properly throught
out and reveiwed design?

-Dave.
-- 
Dave Chinner
david@fromorbit.com
Re: [PATCH v2 0/3] mm: Make unregistration of super_block shrinker more faster
Posted by Qi Zheng 2 years, 8 months ago

On 2023/6/7 06:02, Dave Chinner wrote:
> On Wed, Jun 07, 2023 at 12:06:03AM +0300, Kirill Tkhai wrote:
>> On 06.06.2023 01:32, Dave Chinner wrote:
>>> On Mon, Jun 05, 2023 at 10:02:46PM +0300, Kirill Tkhai wrote:
>>>> This patch set introduces a new scheme of shrinker unregistration. It allows to split
>>>> the unregistration in two parts: fast and slow. This allows to hide slow part from
>>>> a user, so user-visible unregistration becomes fast.
>>>>
>>>> This fixes the -88.8% regression of stress-ng.ramfs.ops_per_sec noticed
>>>> by kernel test robot:
>>>>
>>>> https://lore.kernel.org/lkml/202305230837.db2c233f-yujie.liu@intel.com/
>>>>
>>>> ---
>>>>
>>>> Kirill Tkhai (2):
>>>>        mm: Split unregister_shrinker() in fast and slow part
>>>>        fs: Use delayed shrinker unregistration
>>>
>>> Did you test any filesystem other than ramfs?
>>>
>>> Filesystems more complex than ramfs have internal shrinkers, and so
>>> they will still be running the slow synchronize_srcu() - potentially
>>> multiple times! - in every unmount. Both XFS and ext4 have 3
>>> internal shrinker instances per mount, so they will still call
>>> synchronize_srcu() at least 3 times per unmount after this change.
>>>
>>> What about any other subsystem that runs a shrinker - do they have
>>> context depedent shrinker instances that get frequently created and
>>> destroyed? They'll need the same treatment.
>>
>> Of course, all of shrinkers should be fixed. This patch set just aims to describe
>> the idea more wider, because I'm not sure most people read replys to kernel robot reports.

Thank you, Kirill.

>>
>> This is my suggestion of way to go. Probably, Qi is right person to ask whether
>> we're going to extend this and to maintain f95bdb700bc6 in tree.
>>
>> There is not much time. Unfortunately, kernel test robot reported this significantly late.
> 
> And that's why it should be reverted rather than trying to rush to
> try to fix it.
> 
> I'm kind of tired of finding out about mm reclaim regressions only
> when I see patches making naive and/or broken changes to subsystem
> shrinker implementations without any real clue about what they are
> doing.  If people/subsystems who maintain shrinker implementations
> were cc'd on the changes to the shrinker implementation, this would
> have all been resolved before merging occurred....
> 
> Lockless shrinker lists need a heap of supporting changes to be done
> first so that they aren't reliant on synchronise_srcu() *at all*. If
> the code was properly designed in the first place (i.e. dynamic
> shrinker structures freed via call_rcu()), we wouldn't be in rushing
> to fix weird regressions right now.
> 
> Can we please revert this and start again with a properly throught
> out and reveiwed design?

I have no idea on whether to revert this, I follow the final decision of
the community.

 From my personal point of view, I think it is worth sacrificing the
speed of unregistration alone compared to the benefits it brings
(lockless shrink, etc).

Of course, it would be better if there is a more perfect solution.
If you have a better idea, it might be better to post the code first for
discussion. Otherwise, I am afraid that if we just revert it, the
problem of shrinker_rwsem will continue for many years.

And hi Dave, I know you're mad that I didn't cc you in the original
patch. Sorry again. How about splitting shrinker-related codes into
the separate files? Then we can add a MAINTAINERS entry to it and add
linux-fsdevel@vger.kernel.org to this entry? So that future people
will not miss to cc fs folks.

Qi.

> 
> -Dave.
Re: [PATCH v2 0/3] mm: Make unregistration of super_block shrinker more faster
Posted by Dave Chinner 2 years, 8 months ago
On Wed, Jun 07, 2023 at 10:51:35AM +0800, Qi Zheng wrote:
> From my personal point of view, I think it is worth sacrificing the
> speed of unregistration alone compared to the benefits it brings
> (lockless shrink, etc).

Nobody is questioning whether this is a worthwhile improvement. The
lockless shrinker instance iteration is definitely a good direction
to move in. The problem is the -process- that has been followed has
lead to a very sub-optimal result.

> Of course, it would be better if there is a more perfect solution.
> If you have a better idea, it might be better to post the code first for
> discussion. Otherwise, I am afraid that if we just revert it, the
> problem of shrinker_rwsem will continue for many years.

No, a revert doesn't mean we don't want the change; a revert means
the way the change was attempted has caused unexpected problems.
We need to go back to the drawing board and work out a better way to
do this.

> And hi Dave, I know you're mad that I didn't cc you in the original
> patch.

No, I'm not mad at you.

If I'm annoyed at anyone, it's the senior mm developers and
maintainers that I'm annoyed at - not informing relevant parties
about modifications to shrinker infrastructure or implementations
has lead to regressions escaping out to user systems multiple times
in the past. 

Yet here we are again....

> Sorry again. How about splitting shrinker-related codes into
> the separate files? Then we can add a MAINTAINERS entry to it and add
> linux-fsdevel@vger.kernel.org to this entry? So that future people
> will not miss to cc fs folks.

I don't think that fixes the problem, because the scope if much
wider than fs-devel:  look at all the different subsystems
that have a shrinker.

The whole kernel development process has always worked by the rule
that we're changing common infrastructure, all the subsystems using
that infrastructure need to be cc'd on the changes to the
infrastructure they are using. Just cc'ing -fsdevel isn't enough,
we've also got shrinkers in graphics driver infrastructure, *RCU*,
virtio, DM, bcache and various other subsystems.

And I'm betting most of them don't know that significant changes
have been made to how the shrinkers work....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com