NOTE: Not be picked directly. After getting necessary acks, I'll be working
      out merging with Jakub and Jens.

The patchset implements io_uring zerocopy send. It works with both registered
and normal buffers, mixing is allowed but not recommended. Apart from usual
request completions, just as with MSG_ZEROCOPY, io_uring separately notifies
the userspace when buffers are freed and can be reused (see API design below),
which is delivered into io_uring's Completion Queue. Those "buffer-free"
notifications are not necessarily per request, but the userspace has control
over it and should explicitly attaching a number of requests to a single
notification. The series also adds some internal optimisations when used with
registered buffers like removing page referencing.

From the kernel networking perspective there are two main changes. The first
one is passing ubuf_info into the network layer from io_uring (inside of an
in kernel struct msghdr). This allows extra optimisations, e.g. ubuf_info
caching on the io_uring side, but also helps to avoid cross-referencing
and synchronisation problems. The second part is an optional optimisation
removing page referencing for requests with registered buffers.

Benchmarking with an optimised version of the selftest (see [1]), which sends
a bunch of requests, waits for completions and repeats. "+ flush" column posts
one additional "buffer-free" notification per request, and just "zc" doesn't
post buffer notifications at all.

NIC (requests / second):
IO size | non-zc    | zc             | zc + flush
4000    | 495134    | 606420 (+22%)  | 558971 (+12%)
1500    | 551808    | 577116 (+4.5%) | 565803 (+2.5%)
1000    | 584677    | 592088 (+1.2%) | 560885 (-4%)
600     | 596292    | 598550 (+0.4%) | 555366 (-6.7%)

dummy (requests / second):
IO size | non-zc    | zc             | zc + flush
8000    | 1299916   | 2396600 (+84%) | 2224219 (+71%)
4000    | 1869230   | 2344146 (+25%) | 2170069 (+16%)
1200    | 2071617   | 2361960 (+14%) | 2203052 (+6%)
600     | 2106794   | 2381527 (+13%) | 2195295 (+4%)

Previously it also brought a massive performance speedup compared to the
msg_zerocopy tool (see [3]), which is probably not super interesting.

There is an additional bunch of refcounting optimisations that was omitted from
the series for simplicity and as they don't change the picture drastically,
they will be sent as follow up, as well as flushing optimisations closing the
performance gap b/w two last columns.

Note: the series is based on net-next + for-5.20/io_uring, but as vanilla
net-next fails for me the repo (see [2]) is on top of for-5.20/io_uring.

Links:

  liburing (benchmark + tests):
  [1] https://github.com/isilence/liburing/tree/zc_v4

  kernel repo:
  [2] https://github.com/isilence/linux/tree/zc_v4

  RFC v1:
  [3] https://lore.kernel.org/io-uring/cover.1638282789.git.asml.silence@gmail.com/

  RFC v2:
  https://lore.kernel.org/io-uring/cover.1640029579.git.asml.silence@gmail.com/

  Net patches based
  git@github.com:isilence/linux.git zc_v4-net-base

API design overview:

  The series introduces an io_uring concept of notifactors. From the userspace
  perspective it's an entity to which it can bind one or more requests and then
  requesting to flush it. Flushing a notifier makes it impossible to attach new
  requests to it, and instructs the notifier to post a completion once all
  requests attached to it are completed and the kernel doesn't need the buffers
  anymore.

  Notifications are stored in notification slots, which should be registered as
  an array in io_uring. Each slot stores only one notifier at any particular
  moment. Flushing removes it from the slot and the slot automatically replaces
  it with a new notifier. All operations with notifiers are done by specifying
  an index of a slot it's currently in.

  When registering a notification the userspace specifies a u64 tag for each
  slot, which will be copied in notification completion entries as
  cqe::user_data. cqe::res is 0 and cqe::flags is equal to wrap around u32
  sequence number counting notifiers of a slot.

Changelog:

  v3 -> v4
    custom iov_iter handling

  RFC v2 -> v3:
    mem accounting for non-registered buffers
    allow mixing registered and normal requests per notifier
    notification flushing via IORING_OP_RSRC_UPDATE
    TCP support
    fix buffer indexing
    fix io-wq ->uring_lock locking
    fix bugs when mixing with MSG_ZEROCOPY
    fix managed refs bugs in skbuff.c

  RFC -> RFC v2:
    remove additional overhead for non-zc from skb_release_data()
    avoid msg propagation, hide extra bits of non-zc overhead
    task_work based "buffer free" notifications
    improve io_uring's notification refcounting
    added 5/19, (no pfmemalloc tracking)
    added 8/19 and 9/19 preventing small copies with zc
    misc small changes

David Ahern (1):
  net: Allow custom iter handler in msghdr

Pavel Begunkov (26):
  ipv4: avoid partial copy for zc
  ipv6: avoid partial copy for zc
  skbuff: don't mix ubuf_info from different sources
  skbuff: add SKBFL_DONT_ORPHAN flag
  skbuff: carry external ubuf_info in msghdr
  net: introduce managed frags infrastructure
  net: introduce __skb_fill_page_desc_noacc
  ipv4/udp: support externally provided ubufs
  ipv6/udp: support externally provided ubufs
  tcp: support externally provided ubufs
  io_uring: initialise msghdr::msg_ubuf
  io_uring: export io_put_task()
  io_uring: add zc notification infrastructure
  io_uring: cache struct io_notif
  io_uring: complete notifiers in tw
  io_uring: add rsrc referencing for notifiers
  io_uring: add notification slot registration
  io_uring: wire send zc request type
  io_uring: account locked pages for non-fixed zc
  io_uring: allow to pass addr into sendzc
  io_uring: sendzc with fixed buffers
  io_uring: flush notifiers after sendzc
  io_uring: rename IORING_OP_FILES_UPDATE
  io_uring: add zc notification flush requests
  io_uring: enable managed frags with register buffers
  selftests/io_uring: test zerocopy send

 include/linux/io_uring_types.h                |  37 ++
 include/linux/skbuff.h                        |  66 +-
 include/linux/socket.h                        |   5 +
 include/uapi/linux/io_uring.h                 |  45 +-
 io_uring/Makefile                             |   2 +-
 io_uring/io_uring.c                           |  42 +-
 io_uring/io_uring.h                           |  22 +
 io_uring/net.c                                | 187 ++++++
 io_uring/net.h                                |   4 +
 io_uring/notif.c                              | 215 +++++++
 io_uring/notif.h                              |  87 +++
 io_uring/opdef.c                              |  24 +-
 io_uring/rsrc.c                               |  55 +-
 io_uring/rsrc.h                               |  16 +-
 io_uring/tctx.h                               |  26 -
 net/compat.c                                  |   1 +
 net/core/datagram.c                           |  14 +-
 net/core/skbuff.c                             |  37 +-
 net/ipv4/ip_output.c                          |  50 +-
 net/ipv4/tcp.c                                |  32 +-
 net/ipv6/ip6_output.c                         |  49 +-
 net/socket.c                                  |   3 +
 tools/testing/selftests/net/Makefile          |   1 +
 .../selftests/net/io_uring_zerocopy_tx.c      | 605 ++++++++++++++++++
 .../selftests/net/io_uring_zerocopy_tx.sh     | 131 ++++
 25 files changed, 1628 insertions(+), 128 deletions(-)
 create mode 100644 io_uring/notif.c
 create mode 100644 io_uring/notif.h
 create mode 100644 tools/testing/selftests/net/io_uring_zerocopy_tx.c
 create mode 100755 tools/testing/selftests/net/io_uring_zerocopy_tx.sh

-- 
2.36.1

On 7/7/22 5:49 AM, Pavel Begunkov wrote:
> NOTE: Not be picked directly. After getting necessary acks, I'll be working
>       out merging with Jakub and Jens.
> 
> The patchset implements io_uring zerocopy send. It works with both registered
> and normal buffers, mixing is allowed but not recommended. Apart from usual
> request completions, just as with MSG_ZEROCOPY, io_uring separately notifies
> the userspace when buffers are freed and can be reused (see API design below),
> which is delivered into io_uring's Completion Queue. Those "buffer-free"
> notifications are not necessarily per request, but the userspace has control
> over it and should explicitly attaching a number of requests to a single
> notification. The series also adds some internal optimisations when used with
> registered buffers like removing page referencing.
> 
> From the kernel networking perspective there are two main changes. The first
> one is passing ubuf_info into the network layer from io_uring (inside of an
> in kernel struct msghdr). This allows extra optimisations, e.g. ubuf_info
> caching on the io_uring side, but also helps to avoid cross-referencing
> and synchronisation problems. The second part is an optional optimisation
> removing page referencing for requests with registered buffers.
> 
> Benchmarking with an optimised version of the selftest (see [1]), which sends
> a bunch of requests, waits for completions and repeats. "+ flush" column posts
> one additional "buffer-free" notification per request, and just "zc" doesn't
> post buffer notifications at all.
> 
> NIC (requests / second):
> IO size | non-zc    | zc             | zc + flush
> 4000    | 495134    | 606420 (+22%)  | 558971 (+12%)
> 1500    | 551808    | 577116 (+4.5%) | 565803 (+2.5%)
> 1000    | 584677    | 592088 (+1.2%) | 560885 (-4%)
> 600     | 596292    | 598550 (+0.4%) | 555366 (-6.7%)
> 
> dummy (requests / second):
> IO size | non-zc    | zc             | zc + flush
> 8000    | 1299916   | 2396600 (+84%) | 2224219 (+71%)
> 4000    | 1869230   | 2344146 (+25%) | 2170069 (+16%)
> 1200    | 2071617   | 2361960 (+14%) | 2203052 (+6%)
> 600     | 2106794   | 2381527 (+13%) | 2195295 (+4%)
> 
> Previously it also brought a massive performance speedup compared to the
> msg_zerocopy tool (see [3]), which is probably not super interesting.
> 

can you add a comment that the above results are for UDP.

You dropped comments about TCP testing; any progress there? If not, can
you relay any issues you are hitting?

On 7/8/22 05:10, David Ahern wrote:
> On 7/7/22 5:49 AM, Pavel Begunkov wrote:
>> NOTE: Not be picked directly. After getting necessary acks, I'll be working
>>        out merging with Jakub and Jens.
>>
>> The patchset implements io_uring zerocopy send. It works with both registered
>> and normal buffers, mixing is allowed but not recommended. Apart from usual
>> request completions, just as with MSG_ZEROCOPY, io_uring separately notifies
>> the userspace when buffers are freed and can be reused (see API design below),
>> which is delivered into io_uring's Completion Queue. Those "buffer-free"
>> notifications are not necessarily per request, but the userspace has control
>> over it and should explicitly attaching a number of requests to a single
>> notification. The series also adds some internal optimisations when used with
>> registered buffers like removing page referencing.
>>
>>  From the kernel networking perspective there are two main changes. The first
>> one is passing ubuf_info into the network layer from io_uring (inside of an
>> in kernel struct msghdr). This allows extra optimisations, e.g. ubuf_info
>> caching on the io_uring side, but also helps to avoid cross-referencing
>> and synchronisation problems. The second part is an optional optimisation
>> removing page referencing for requests with registered buffers.
>>
>> Benchmarking with an optimised version of the selftest (see [1]), which sends
>> a bunch of requests, waits for completions and repeats. "+ flush" column posts
>> one additional "buffer-free" notification per request, and just "zc" doesn't
>> post buffer notifications at all.
>>
>> NIC (requests / second):
>> IO size | non-zc    | zc             | zc + flush
>> 4000    | 495134    | 606420 (+22%)  | 558971 (+12%)
>> 1500    | 551808    | 577116 (+4.5%) | 565803 (+2.5%)
>> 1000    | 584677    | 592088 (+1.2%) | 560885 (-4%)
>> 600     | 596292    | 598550 (+0.4%) | 555366 (-6.7%)
>>
>> dummy (requests / second):
>> IO size | non-zc    | zc             | zc + flush
>> 8000    | 1299916   | 2396600 (+84%) | 2224219 (+71%)
>> 4000    | 1869230   | 2344146 (+25%) | 2170069 (+16%)
>> 1200    | 2071617   | 2361960 (+14%) | 2203052 (+6%)
>> 600     | 2106794   | 2381527 (+13%) | 2195295 (+4%)
>>
>> Previously it also brought a massive performance speedup compared to the
>> msg_zerocopy tool (see [3]), which is probably not super interesting.
>>
> 
> can you add a comment that the above results are for UDP.

Oh, right, forgot to add it


> You dropped comments about TCP testing; any progress there? If not, can
> you relay any issues you are hitting?

Not really a problem, but for me it's bottle necked at NIC bandwidth
(~3GB/s) for both zc and non-zc and doesn't even nearly saturate a CPU.
Was actually benchmarked by my colleague quite a while ago, but can't
find numbers. Probably need to at least add localhost numbers or grab
a better server.

-- 
Pavel Begunkov

On 7/8/22 15:26, Pavel Begunkov wrote:
> On 7/8/22 05:10, David Ahern wrote:
>> On 7/7/22 5:49 AM, Pavel Begunkov wrote:
>>> NOTE: Not be picked directly. After getting necessary acks, I'll be working
>>>        out merging with Jakub and Jens.
>>>
>>> The patchset implements io_uring zerocopy send. It works with both registered
>>> and normal buffers, mixing is allowed but not recommended. Apart from usual
>>> request completions, just as with MSG_ZEROCOPY, io_uring separately notifies
>>> the userspace when buffers are freed and can be reused (see API design below),
>>> which is delivered into io_uring's Completion Queue. Those "buffer-free"
>>> notifications are not necessarily per request, but the userspace has control
>>> over it and should explicitly attaching a number of requests to a single
>>> notification. The series also adds some internal optimisations when used with
>>> registered buffers like removing page referencing.
>>>
>>>  From the kernel networking perspective there are two main changes. The first
>>> one is passing ubuf_info into the network layer from io_uring (inside of an
>>> in kernel struct msghdr). This allows extra optimisations, e.g. ubuf_info
>>> caching on the io_uring side, but also helps to avoid cross-referencing
>>> and synchronisation problems. The second part is an optional optimisation
>>> removing page referencing for requests with registered buffers.
>>>
>>> Benchmarking with an optimised version of the selftest (see [1]), which sends
>>> a bunch of requests, waits for completions and repeats. "+ flush" column posts
>>> one additional "buffer-free" notification per request, and just "zc" doesn't
>>> post buffer notifications at all.
>>>
>>> NIC (requests / second):
>>> IO size | non-zc    | zc             | zc + flush
>>> 4000    | 495134    | 606420 (+22%)  | 558971 (+12%)
>>> 1500    | 551808    | 577116 (+4.5%) | 565803 (+2.5%)
>>> 1000    | 584677    | 592088 (+1.2%) | 560885 (-4%)
>>> 600     | 596292    | 598550 (+0.4%) | 555366 (-6.7%)
>>>
>>> dummy (requests / second):
>>> IO size | non-zc    | zc             | zc + flush
>>> 8000    | 1299916   | 2396600 (+84%) | 2224219 (+71%)
>>> 4000    | 1869230   | 2344146 (+25%) | 2170069 (+16%)
>>> 1200    | 2071617   | 2361960 (+14%) | 2203052 (+6%)
>>> 600     | 2106794   | 2381527 (+13%) | 2195295 (+4%)
>>>
>>> Previously it also brought a massive performance speedup compared to the
>>> msg_zerocopy tool (see [3]), which is probably not super interesting.
>>>
>>
>> can you add a comment that the above results are for UDP.
> 
> Oh, right, forgot to add it
> 
> 
>> You dropped comments about TCP testing; any progress there? If not, can
>> you relay any issues you are hitting?
> 
> Not really a problem, but for me it's bottle necked at NIC bandwidth
> (~3GB/s) for both zc and non-zc and doesn't even nearly saturate a CPU.
> Was actually benchmarked by my colleague quite a while ago, but can't
> find numbers. Probably need to at least add localhost numbers or grab
> a better server.

Testing localhost TCP with a hack (see below), it doesn't include
refcounting optimisations I was testing UDP with and that will be
sent afterwards. Numbers are in MB/s

IO size | non-zc    | zc
1200    | 4174      | 4148
4096    | 7597      | 11228

Because it's localhost, we also spend cycles here for the recv side.
Using a real NIC 1200 bytes, zc is worse than non-zc ~5-10%, maybe the
omitted optimisations will somewhat help. I don't consider it to be a
blocker. but would be interesting to poke into later. One thing helping
non-zc is that it squeezes a number of requests into a single page
whenever zerocopy adds a new frag for every request.

Can't say anything new for larger payloads, I'm still NIC-bound but
looking at CPU utilisation zc doesn't drain as much cycles as non-zc.
Also, I don't remember if mentioned before, but another catch is that
with TCP it expects users to not be flushing notifications too much,
because it forces it to allocate a new skb and lose a good chunk of
benefits from using TCP.


diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
index 1111adefd906..c4b781b2c3b1 100644
--- a/include/linux/skbuff.h
+++ b/include/linux/skbuff.h
@@ -3218,9 +3218,7 @@ static inline int skb_orphan_frags(struct sk_buff *skb, gfp_t gfp_mask)
  /* Frags must be orphaned, even if refcounted, if skb might loop to rx path */
  static inline int skb_orphan_frags_rx(struct sk_buff *skb, gfp_t gfp_mask)
  {
-	if (likely(!skb_zcopy(skb)))
-		return 0;
-	return skb_copy_ubufs(skb, gfp_mask);
+	return skb_orphan_frags(skb, gfp_mask);
  }

-- 
Pavel Begunkov

On 7/11/22 5:56 AM, Pavel Begunkov wrote:
> On 7/8/22 15:26, Pavel Begunkov wrote:
>> On 7/8/22 05:10, David Ahern wrote:
>>> On 7/7/22 5:49 AM, Pavel Begunkov wrote:
>>>> NOTE: Not be picked directly. After getting necessary acks, I'll be
>>>> working
>>>>        out merging with Jakub and Jens.
>>>>
>>>> The patchset implements io_uring zerocopy send. It works with both
>>>> registered
>>>> and normal buffers, mixing is allowed but not recommended. Apart
>>>> from usual
>>>> request completions, just as with MSG_ZEROCOPY, io_uring separately
>>>> notifies
>>>> the userspace when buffers are freed and can be reused (see API
>>>> design below),
>>>> which is delivered into io_uring's Completion Queue. Those
>>>> "buffer-free"
>>>> notifications are not necessarily per request, but the userspace has
>>>> control
>>>> over it and should explicitly attaching a number of requests to a
>>>> single
>>>> notification. The series also adds some internal optimisations when
>>>> used with
>>>> registered buffers like removing page referencing.
>>>>
>>>>  From the kernel networking perspective there are two main changes.
>>>> The first
>>>> one is passing ubuf_info into the network layer from io_uring
>>>> (inside of an
>>>> in kernel struct msghdr). This allows extra optimisations, e.g.
>>>> ubuf_info
>>>> caching on the io_uring side, but also helps to avoid cross-referencing
>>>> and synchronisation problems. The second part is an optional
>>>> optimisation
>>>> removing page referencing for requests with registered buffers.
>>>>
>>>> Benchmarking with an optimised version of the selftest (see [1]),
>>>> which sends
>>>> a bunch of requests, waits for completions and repeats. "+ flush"
>>>> column posts
>>>> one additional "buffer-free" notification per request, and just "zc"
>>>> doesn't
>>>> post buffer notifications at all.
>>>>
>>>> NIC (requests / second):
>>>> IO size | non-zc    | zc             | zc + flush
>>>> 4000    | 495134    | 606420 (+22%)  | 558971 (+12%)
>>>> 1500    | 551808    | 577116 (+4.5%) | 565803 (+2.5%)
>>>> 1000    | 584677    | 592088 (+1.2%) | 560885 (-4%)
>>>> 600     | 596292    | 598550 (+0.4%) | 555366 (-6.7%)
>>>>
>>>> dummy (requests / second):
>>>> IO size | non-zc    | zc             | zc + flush
>>>> 8000    | 1299916   | 2396600 (+84%) | 2224219 (+71%)
>>>> 4000    | 1869230   | 2344146 (+25%) | 2170069 (+16%)
>>>> 1200    | 2071617   | 2361960 (+14%) | 2203052 (+6%)
>>>> 600     | 2106794   | 2381527 (+13%) | 2195295 (+4%)
>>>>
>>>> Previously it also brought a massive performance speedup compared to
>>>> the
>>>> msg_zerocopy tool (see [3]), which is probably not super interesting.
>>>>
>>>
>>> can you add a comment that the above results are for UDP.
>>
>> Oh, right, forgot to add it
>>
>>
>>> You dropped comments about TCP testing; any progress there? If not, can
>>> you relay any issues you are hitting?
>>
>> Not really a problem, but for me it's bottle necked at NIC bandwidth
>> (~3GB/s) for both zc and non-zc and doesn't even nearly saturate a CPU.
>> Was actually benchmarked by my colleague quite a while ago, but can't
>> find numbers. Probably need to at least add localhost numbers or grab
>> a better server.
> 
> Testing localhost TCP with a hack (see below), it doesn't include
> refcounting optimisations I was testing UDP with and that will be
> sent afterwards. Numbers are in MB/s
> 
> IO size | non-zc    | zc
> 1200    | 4174      | 4148
> 4096    | 7597      | 11228

I am surprised by the low numbers; you should be able to saturate a 100G
link with TCP and ZC TX API.

> 
> Because it's localhost, we also spend cycles here for the recv side.
> Using a real NIC 1200 bytes, zc is worse than non-zc ~5-10%, maybe the
> omitted optimisations will somewhat help. I don't consider it to be a
> blocker. but would be interesting to poke into later. One thing helping
> non-zc is that it squeezes a number of requests into a single page
> whenever zerocopy adds a new frag for every request.
> 
> Can't say anything new for larger payloads, I'm still NIC-bound but
> looking at CPU utilisation zc doesn't drain as much cycles as non-zc.
> Also, I don't remember if mentioned before, but another catch is that
> with TCP it expects users to not be flushing notifications too much,
> because it forces it to allocate a new skb and lose a good chunk of
> benefits from using TCP.

I had issues with TCP sockets and io_uring at the end of 2020:
https://www.spinics.net/lists/io-uring/msg05125.html

have not tried anything recent (from 2022).

On 7/14/22 00:45, David Ahern wrote:
> On 7/11/22 5:56 AM, Pavel Begunkov wrote:
>> On 7/8/22 15:26, Pavel Begunkov wrote:
>>> On 7/8/22 05:10, David Ahern wrote:
>>>> On 7/7/22 5:49 AM, Pavel Begunkov wrote:
>>>>> NOTE: Not be picked directly. After getting necessary acks, I'll be
>>>>> working
>>>>>         out merging with Jakub and Jens.
>>>>>
>>>>> The patchset implements io_uring zerocopy send. It works with both
>>>>> registered
>>>>> and normal buffers, mixing is allowed but not recommended. Apart
>>>>> from usual
>>>>> request completions, just as with MSG_ZEROCOPY, io_uring separately
>>>>> notifies
>>>>> the userspace when buffers are freed and can be reused (see API
>>>>> design below),
>>>>> which is delivered into io_uring's Completion Queue. Those
>>>>> "buffer-free"
>>>>> notifications are not necessarily per request, but the userspace has
>>>>> control
>>>>> over it and should explicitly attaching a number of requests to a
>>>>> single
>>>>> notification. The series also adds some internal optimisations when
>>>>> used with
>>>>> registered buffers like removing page referencing.
>>>>>
>>>>>   From the kernel networking perspective there are two main changes.
>>>>> The first
>>>>> one is passing ubuf_info into the network layer from io_uring
>>>>> (inside of an
>>>>> in kernel struct msghdr). This allows extra optimisations, e.g.
>>>>> ubuf_info
>>>>> caching on the io_uring side, but also helps to avoid cross-referencing
>>>>> and synchronisation problems. The second part is an optional
>>>>> optimisation
>>>>> removing page referencing for requests with registered buffers.
>>>>>
>>>>> Benchmarking with an optimised version of the selftest (see [1]),
>>>>> which sends
>>>>> a bunch of requests, waits for completions and repeats. "+ flush"
>>>>> column posts
>>>>> one additional "buffer-free" notification per request, and just "zc"
>>>>> doesn't
>>>>> post buffer notifications at all.
>>>>>
>>>>> NIC (requests / second):
>>>>> IO size | non-zc    | zc             | zc + flush
>>>>> 4000    | 495134    | 606420 (+22%)  | 558971 (+12%)
>>>>> 1500    | 551808    | 577116 (+4.5%) | 565803 (+2.5%)
>>>>> 1000    | 584677    | 592088 (+1.2%) | 560885 (-4%)
>>>>> 600     | 596292    | 598550 (+0.4%) | 555366 (-6.7%)
>>>>>
>>>>> dummy (requests / second):
>>>>> IO size | non-zc    | zc             | zc + flush
>>>>> 8000    | 1299916   | 2396600 (+84%) | 2224219 (+71%)
>>>>> 4000    | 1869230   | 2344146 (+25%) | 2170069 (+16%)
>>>>> 1200    | 2071617   | 2361960 (+14%) | 2203052 (+6%)
>>>>> 600     | 2106794   | 2381527 (+13%) | 2195295 (+4%)
>>>>>
>>>>> Previously it also brought a massive performance speedup compared to
>>>>> the
>>>>> msg_zerocopy tool (see [3]), which is probably not super interesting.
>>>>>
>>>>
>>>> can you add a comment that the above results are for UDP.
>>>
>>> Oh, right, forgot to add it
>>>
>>>
>>>> You dropped comments about TCP testing; any progress there? If not, can
>>>> you relay any issues you are hitting?
>>>
>>> Not really a problem, but for me it's bottle necked at NIC bandwidth
>>> (~3GB/s) for both zc and non-zc and doesn't even nearly saturate a CPU.
>>> Was actually benchmarked by my colleague quite a while ago, but can't
>>> find numbers. Probably need to at least add localhost numbers or grab
>>> a better server.
>>
>> Testing localhost TCP with a hack (see below), it doesn't include
>> refcounting optimisations I was testing UDP with and that will be
>> sent afterwards. Numbers are in MB/s
>>
>> IO size | non-zc    | zc
>> 1200    | 4174      | 4148
>> 4096    | 7597      | 11228
> 
> I am surprised by the low numbers; you should be able to saturate a 100G
> link with TCP and ZC TX API.

It was a quick test with my laptop, not a super fast CPU, preemptible
kernel, etc., and considering that the fact that it processes receives
from in the same send syscall roughly doubles the overhead, 87Gb/s
looks ok. It's not like MSG_ZEROCOPY would look much different, even
more to that all sends here will be executed sequentially in io_uring,
so no extra parallelism or so. As for 1200, I think 4GB/s is reasonable,
it's just the kernel overhead per byte is too high, should be same with
just send(2).

>> Because it's localhost, we also spend cycles here for the recv side.
>> Using a real NIC 1200 bytes, zc is worse than non-zc ~5-10%, maybe the
>> omitted optimisations will somewhat help. I don't consider it to be a
>> blocker. but would be interesting to poke into later. One thing helping
>> non-zc is that it squeezes a number of requests into a single page
>> whenever zerocopy adds a new frag for every request.
>>
>> Can't say anything new for larger payloads, I'm still NIC-bound but
>> looking at CPU utilisation zc doesn't drain as much cycles as non-zc.
>> Also, I don't remember if mentioned before, but another catch is that
>> with TCP it expects users to not be flushing notifications too much,
>> because it forces it to allocate a new skb and lose a good chunk of
>> benefits from using TCP.
> 
> I had issues with TCP sockets and io_uring at the end of 2020:
> https://www.spinics.net/lists/io-uring/msg05125.html
> 
> have not tried anything recent (from 2022).

Haven't seen it back then. In general io_uring doesn't stop submitting
requests if one request fails, at least because we're trying to execute
requests asynchronously. And in general, requests can get executed
out of order, so most probably submitting a bunch of requests to a single
TCP sock without any ordering on io_uring side is likely a bug.

You can link io_uring requests, i.e. IOSQE_IO_LINK, guaranteeing
execution ordering. And if you meant links in the message, I agree
that it was not the best decision to consider len < sqe->len not
an error and not breaking links, but it was later added that
MSG_WAITALL would also change the success condition to
len==sqe->len. But all that is relevant if you was using linking.

-- 
Pavel Begunkov

On 7/14/22 12:55 PM, Pavel Begunkov wrote:
>>>>> You dropped comments about TCP testing; any progress there? If not,
>>>>> can
>>>>> you relay any issues you are hitting?
>>>>
>>>> Not really a problem, but for me it's bottle necked at NIC bandwidth
>>>> (~3GB/s) for both zc and non-zc and doesn't even nearly saturate a CPU.
>>>> Was actually benchmarked by my colleague quite a while ago, but can't
>>>> find numbers. Probably need to at least add localhost numbers or grab
>>>> a better server.
>>>
>>> Testing localhost TCP with a hack (see below), it doesn't include
>>> refcounting optimisations I was testing UDP with and that will be
>>> sent afterwards. Numbers are in MB/s
>>>
>>> IO size | non-zc    | zc
>>> 1200    | 4174      | 4148
>>> 4096    | 7597      | 11228
>>
>> I am surprised by the low numbers; you should be able to saturate a 100G
>> link with TCP and ZC TX API.
> 
> It was a quick test with my laptop, not a super fast CPU, preemptible
> kernel, etc., and considering that the fact that it processes receives
> from in the same send syscall roughly doubles the overhead, 87Gb/s
> looks ok. It's not like MSG_ZEROCOPY would look much different, even
> more to that all sends here will be executed sequentially in io_uring,
> so no extra parallelism or so. As for 1200, I think 4GB/s is reasonable,
> it's just the kernel overhead per byte is too high, should be same with
> just send(2).

?
It's a stream socket so those sends are coalesced into MTU sized packets.

> 
>>> Because it's localhost, we also spend cycles here for the recv side.
>>> Using a real NIC 1200 bytes, zc is worse than non-zc ~5-10%, maybe the
>>> omitted optimisations will somewhat help. I don't consider it to be a
>>> blocker. but would be interesting to poke into later. One thing helping
>>> non-zc is that it squeezes a number of requests into a single page
>>> whenever zerocopy adds a new frag for every request.
>>>
>>> Can't say anything new for larger payloads, I'm still NIC-bound but
>>> looking at CPU utilisation zc doesn't drain as much cycles as non-zc.
>>> Also, I don't remember if mentioned before, but another catch is that
>>> with TCP it expects users to not be flushing notifications too much,
>>> because it forces it to allocate a new skb and lose a good chunk of
>>> benefits from using TCP.
>>
>> I had issues with TCP sockets and io_uring at the end of 2020:
>> https://www.spinics.net/lists/io-uring/msg05125.html
>>
>> have not tried anything recent (from 2022).
> 
> Haven't seen it back then. In general io_uring doesn't stop submitting
> requests if one request fails, at least because we're trying to execute
> requests asynchronously. And in general, requests can get executed
> out of order, so most probably submitting a bunch of requests to a single
> TCP sock without any ordering on io_uring side is likely a bug.

TCP socket buffer fills resulting in a partial send (i.e, for a given
sqe submission only part of the write/send succeeded). io_uring was not
handling that case.

I'll try to find some time to resurrect the iperf3 patch and try top of
tree kernel.

> 
> You can link io_uring requests, i.e. IOSQE_IO_LINK, guaranteeing
> execution ordering. And if you meant links in the message, I agree
> that it was not the best decision to consider len < sqe->len not
> an error and not breaking links, but it was later added that
> MSG_WAITALL would also change the success condition to
> len==sqe->len. But all that is relevant if you was using linking.
> 

On 7/17/22 8:19 PM, David Ahern wrote:
>>
>> Haven't seen it back then. In general io_uring doesn't stop submitting
>> requests if one request fails, at least because we're trying to execute
>> requests asynchronously. And in general, requests can get executed
>> out of order, so most probably submitting a bunch of requests to a single
>> TCP sock without any ordering on io_uring side is likely a bug.
> 
> TCP socket buffer fills resulting in a partial send (i.e, for a given
> sqe submission only part of the write/send succeeded). io_uring was not
> handling that case.
> 
> I'll try to find some time to resurrect the iperf3 patch and try top of
> tree kernel.

With your zc_v5 branch (plus the init fix on using msg->sg_from_iter),
iperf3 with io_uring support (non-ZC case) no longer shows completions
with incomplete sends. So that is good improvement over the last time I
tried it.

However, adding in the ZC support and that problem resurfaces - a lot of
completions are for an incomplete size.

liburing comes from your tree, zc_v4 branch. Upstream does not have
support for notifications yet, so I can not move to it.

Changes to iperf3 are here:
   https://github.com/dsahern/iperf mods-3.10-io_uring

On 7/24/22 19:28, David Ahern wrote:
> On 7/17/22 8:19 PM, David Ahern wrote:
>>>
>>> Haven't seen it back then. In general io_uring doesn't stop submitting
>>> requests if one request fails, at least because we're trying to execute
>>> requests asynchronously. And in general, requests can get executed
>>> out of order, so most probably submitting a bunch of requests to a single
>>> TCP sock without any ordering on io_uring side is likely a bug.
>>
>> TCP socket buffer fills resulting in a partial send (i.e, for a given
>> sqe submission only part of the write/send succeeded). io_uring was not
>> handling that case.
>>
>> I'll try to find some time to resurrect the iperf3 patch and try top of
>> tree kernel.
> 
> With your zc_v5 branch (plus the init fix on using msg->sg_from_iter),
> iperf3 with io_uring support (non-ZC case) no longer shows completions
> with incomplete sends. So that is good improvement over the last time I
> tried it.
> 
> However, adding in the ZC support and that problem resurfaces - a lot of
> completions are for an incomplete size.
> 
> liburing comes from your tree, zc_v4 branch. Upstream does not have
> support for notifications yet, so I can not move to it.
> 
> Changes to iperf3 are here:
>     https://github.com/dsahern/iperf mods-3.10-io_uring

Tried it out, the branch below fixes a small problem, adds a couple
of extra optimisations and now it actually uses registered buffers.

     https://github.com/isilence/iperf iou-sendzc

Still, the submission loop looked a bit weird, i.e. it submits I/O
to io_uring only when it exhausts sqes instead of sending right
away with some notion of QD and/or sending in batches. The approach
is good for batching (SQ size =16 here), but not so for latency.

I also see some CPU cycles being burnt in select(2). io_uring wait
would be more natural and perhaps more performant, but I didn't
spend enough time with iperf to say for sure.

-- 
Pavel Begunkov

On 9/26/22 1:08 PM, Pavel Begunkov wrote:
> Tried it out, the branch below fixes a small problem, adds a couple
> of extra optimisations and now it actually uses registered buffers.
> 
>     https://github.com/isilence/iperf iou-sendzc

thanks for the patch; will it pull it in.

> 
> Still, the submission loop looked a bit weird, i.e. it submits I/O
> to io_uring only when it exhausts sqes instead of sending right
> away with some notion of QD and/or sending in batches. The approach
> is good for batching (SQ size =16 here), but not so for latency.
> 
> I also see some CPU cycles being burnt in select(2). io_uring wait
> would be more natural and perhaps more performant, but I didn't
> spend enough time with iperf to say for sure.

ok. It will be a while before I have time to come back to it. In the
meantime it seems like some io_uring changes happened between your dev
branch and what was merged into liburing (compile worked on your branch
but fails with upstream). Is the ZC support in liburing now?

On 9/28/22 20:31, David Ahern wrote:
> On 9/26/22 1:08 PM, Pavel Begunkov wrote:
>> Tried it out, the branch below fixes a small problem, adds a couple
>> of extra optimisations and now it actually uses registered buffers.
>>
>>      https://github.com/isilence/iperf iou-sendzc
> 
> thanks for the patch; will it pull it in.
> 
>> Still, the submission loop looked a bit weird, i.e. it submits I/O
>> to io_uring only when it exhausts sqes instead of sending right
>> away with some notion of QD and/or sending in batches. The approach
>> is good for batching (SQ size =16 here), but not so for latency.
>>
>> I also see some CPU cycles being burnt in select(2). io_uring wait
>> would be more natural and perhaps more performant, but I didn't
>> spend enough time with iperf to say for sure.
> 
> ok. It will be a while before I have time to come back to it. In the
> meantime it seems like some io_uring changes happened between your dev
> branch and what was merged into liburing (compile worked on your branch
> but fails with upstream). Is the ZC support in liburing now?

It is. I forgot to put a note that I also adapted your patches
to uapi changes.No more notification slots but a zc send request
now can post a second CQE if IORING_CQE_F_MORE is set in the
first one. Better described in io_uring_enter(2) man, e.g.

https://git.kernel.dk/cgit/liburing/tree/man/io_uring_enter.2#n1063

-- 
Pavel Begunkov

On 7/24/22 19:28, David Ahern wrote:
> On 7/17/22 8:19 PM, David Ahern wrote:
>>>
>>> Haven't seen it back then. In general io_uring doesn't stop submitting
>>> requests if one request fails, at least because we're trying to execute
>>> requests asynchronously. And in general, requests can get executed
>>> out of order, so most probably submitting a bunch of requests to a single
>>> TCP sock without any ordering on io_uring side is likely a bug.
>>
>> TCP socket buffer fills resulting in a partial send (i.e, for a given
>> sqe submission only part of the write/send succeeded). io_uring was not
>> handling that case.
>>
>> I'll try to find some time to resurrect the iperf3 patch and try top of
>> tree kernel.
> 
> With your zc_v5 branch (plus the init fix on using msg->sg_from_iter),
> iperf3 with io_uring support (non-ZC case) no longer shows completions
> with incomplete sends. So that is good improvement over the last time I
> tried it.
> 
> However, adding in the ZC support and that problem resurfaces - a lot of
> completions are for an incomplete size.

Makes sense, it explicitly retries with normal sends but I didn't
implement it for zc. Might be a good thing to add.

> liburing comes from your tree, zc_v4 branch. Upstream does not have
> support for notifications yet, so I can not move to it.

Upstreamed it

> Changes to iperf3 are here:
>     https://github.com/dsahern/iperf mods-3.10-io_uring

-- 
Pavel Begunkov

On 7/27/22 4:51 AM, Pavel Begunkov wrote:
>> With your zc_v5 branch (plus the init fix on using msg->sg_from_iter),
>> iperf3 with io_uring support (non-ZC case) no longer shows completions
>> with incomplete sends. So that is good improvement over the last time I
>> tried it.
>>
>> However, adding in the ZC support and that problem resurfaces - a lot of
>> completions are for an incomplete size.
> 
> Makes sense, it explicitly retries with normal sends but I didn't
> implement it for zc. Might be a good thing to add.
> 

Yes, before this goes it. It will be confusing to users to get
incomplete completions when using the ZC option.

On 7/18/22 03:19, David Ahern wrote:
> On 7/14/22 12:55 PM, Pavel Begunkov wrote:
>>>>>> You dropped comments about TCP testing; any progress there? If not,
>>>>>> can
>>>>>> you relay any issues you are hitting?
>>>>>
>>>>> Not really a problem, but for me it's bottle necked at NIC bandwidth
>>>>> (~3GB/s) for both zc and non-zc and doesn't even nearly saturate a CPU.
>>>>> Was actually benchmarked by my colleague quite a while ago, but can't
>>>>> find numbers. Probably need to at least add localhost numbers or grab
>>>>> a better server.
>>>>
>>>> Testing localhost TCP with a hack (see below), it doesn't include
>>>> refcounting optimisations I was testing UDP with and that will be
>>>> sent afterwards. Numbers are in MB/s
>>>>
>>>> IO size | non-zc    | zc
>>>> 1200    | 4174      | 4148
>>>> 4096    | 7597      | 11228
>>>
>>> I am surprised by the low numbers; you should be able to saturate a 100G
>>> link with TCP and ZC TX API.
>>
>> It was a quick test with my laptop, not a super fast CPU, preemptible
>> kernel, etc., and considering that the fact that it processes receives
>> from in the same send syscall roughly doubles the overhead, 87Gb/s
>> looks ok. It's not like MSG_ZEROCOPY would look much different, even
>> more to that all sends here will be executed sequentially in io_uring,
>> so no extra parallelism or so. As for 1200, I think 4GB/s is reasonable,
>> it's just the kernel overhead per byte is too high, should be same with
>> just send(2).
> 
> ?
> It's a stream socket so those sends are coalesced into MTU sized packets.

That leaves syscall and io_uring overhead, locking the socket, etc.,
which still requires more cycles than just copying 1200 bytes. And
the used CPU is not blazingly fast, could be that a better CPU/setup
will saturate 100G

>>>> Because it's localhost, we also spend cycles here for the recv side.
>>>> Using a real NIC 1200 bytes, zc is worse than non-zc ~5-10%, maybe the
>>>> omitted optimisations will somewhat help. I don't consider it to be a
>>>> blocker. but would be interesting to poke into later. One thing helping
>>>> non-zc is that it squeezes a number of requests into a single page
>>>> whenever zerocopy adds a new frag for every request.
>>>>
>>>> Can't say anything new for larger payloads, I'm still NIC-bound but
>>>> looking at CPU utilisation zc doesn't drain as much cycles as non-zc.
>>>> Also, I don't remember if mentioned before, but another catch is that
>>>> with TCP it expects users to not be flushing notifications too much,
>>>> because it forces it to allocate a new skb and lose a good chunk of
>>>> benefits from using TCP.
>>>
>>> I had issues with TCP sockets and io_uring at the end of 2020:
>>> https://www.spinics.net/lists/io-uring/msg05125.html
>>>
>>> have not tried anything recent (from 2022).
>>
>> Haven't seen it back then. In general io_uring doesn't stop submitting
>> requests if one request fails, at least because we're trying to execute
>> requests asynchronously. And in general, requests can get executed
>> out of order, so most probably submitting a bunch of requests to a single
>> TCP sock without any ordering on io_uring side is likely a bug.
> 
> TCP socket buffer fills resulting in a partial send (i.e, for a given
> sqe submission only part of the write/send succeeded). io_uring was not
> handling that case.

Shouldn't have been different from send(2) with MSG_NOWAIT, can be short
and the user should handle it. Also I believe Jens pushed just recently
in-kernel retries on the io_uring side for TCP in such cases.

> I'll try to find some time to resurrect the iperf3 patch and try top of
> tree kernel.

Awesome


>> You can link io_uring requests, i.e. IOSQE_IO_LINK, guaranteeing
>> execution ordering. And if you meant links in the message, I agree
>> that it was not the best decision to consider len < sqe->len not
>> an error and not breaking links, but it was later added that
>> MSG_WAITALL would also change the success condition to
>> len==sqe->len. But all that is relevant if you was using linking.

-- 
Pavel Begunkov