mbox series

[RFC,v5,00/17] vsock: MSG_ZEROCOPY flag support

Message ID 20230701063947.3422088-1-AVKrasnov@sberdevices.ru (mailing list archive)
Headers show
Series vsock: MSG_ZEROCOPY flag support | expand

Message

Arseniy Krasnov July 1, 2023, 6:39 a.m. UTC
Hello,

                           DESCRIPTION

this is MSG_ZEROCOPY feature support for virtio/vsock. I tried to follow
current implementation for TCP as much as possible:

1) Sender must enable SO_ZEROCOPY flag to use this feature. Without this
   flag, data will be sent in "classic" copy manner and MSG_ZEROCOPY
   flag will be ignored (e.g. without completion).

2) Kernel uses completions from socket's error queue. Single completion
   for single tx syscall (or it can merge several completions to single
   one). I used already implemented logic for MSG_ZEROCOPY support:
   'msg_zerocopy_realloc()' etc.

Difference with copy way is not significant. During packet allocation,
non-linear skb is created and filled with pinned user pages.
There are also some updates for vhost and guest parts of transport - in
both cases i've added handling of non-linear skb for virtio part. vhost
copies data from such skb to the guest's rx virtio buffers. In the guest,
virtio transport fills tx virtio queue with pages from skb.

Head of this patchset is:
https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next.git/commit/?id=d20dd0ea14072e8a90ff864b2c1603bd68920b4b


This version has several limits/problems (all resolved at v5):

1) As this feature totally depends on transport, there is no way (or it
   is difficult) to check whether transport is able to handle it or not
   during SO_ZEROCOPY setting. Seems I need to call AF_VSOCK specific
   setsockopt callback from setsockopt callback for SOL_SOCKET, but this
   leads to lock problem, because both AF_VSOCK and SOL_SOCKET callback
   are not considered to be called from each other. So in current version
   SO_ZEROCOPY is set successfully to any type (e.g. transport) of
   AF_VSOCK socket, but if transport does not support MSG_ZEROCOPY,
   tx routine will fail with EOPNOTSUPP.

   ^^^ fixed in v5. Thanks to Bobby Eshleman.

2) When MSG_ZEROCOPY is used, for each tx system call we need to enqueue
   one completion. In each completion there is flag which shows how tx
   was performed: zerocopy or copy. This leads that whole message must
   be send in zerocopy or copy way - we can't send part of message with
   copying and rest of message with zerocopy mode (or vice versa). Now,
   we need to account vsock credit logic, e.g. we can't send whole data
   once - only allowed number of bytes could sent at any moment. In case
   of copying way there is no problem as in worst case we can send single
   bytes, but zerocopy is more complex because smallest transmission
   unit is single page. So if there is not enough space at peer's side
   to send integer number of pages (at least one) - we will wait, thus
   stalling tx side. To overcome this problem i've added simple rule -
   zerocopy is possible only when there is enough space at another side
   for whole message (to check, that current 'msghdr' was already used
   in previous tx iterations i use 'iov_offset' field of it's iov iter).

   ^^^
   Discussed as ok during v2. Link:
   https://lore.kernel.org/netdev/23guh3txkghxpgcrcjx7h62qsoj3xgjhfzgtbmqp2slrz3rxr4@zya2z7kwt75l/

3) loopback transport is not supported, because it requires to implement
   non-linear skb handling in dequeue logic (as we "send" fragged skb
   and "receive" it from the same queue). I'm going to implement it in
   next versions.

   ^^^ fixed in v2

4) Current implementation sets max length of packet to 64KB. IIUC this
   is due to 'kmalloc()' allocated data buffers. I think, in case of
   MSG_ZEROCOPY this value could be increased, because 'kmalloc()' is
   not touched for data - user space pages are used as buffers. Also
   this limit trims every message which is > 64KB, thus such messages
   will be send in copy mode due to 'iov_offset' check in 2).

   ^^^ fixed in v2

                         PATCHSET STRUCTURE

Patchset has the following structure:
1) Handle non-linear skbuff on receive in virtio/vhost.
2) Handle non-linear skbuff on send in virtio/vhost.
3) Updates for AF_VSOCK.
4) Enable MSG_ZEROCOPY support on transports.
5) Tests/tools/docs updates.

                            PERFORMANCE

Performance: it is a little bit tricky to compare performance between
copy and zerocopy transmissions. In zerocopy way we need to wait when
user buffers will be released by kernel, so it is like synchronous
path (wait until device driver will process it), while in copy way we
can feed data to kernel as many as we want, don't care about device
driver. So I compared only time which we spend in the 'send()' syscall.
Then if this value will be combined with total number of transmitted
bytes, we can get Gbit/s parameter. Also to avoid tx stalls due to not
enough credit, receiver allocates same amount of space as sender needs.

Sender:
./vsock_perf --sender <CID> --buf-size <buf size> --bytes 256M [--zc]

Receiver:
./vsock_perf --vsk-size 256M

I run tests on two setups: desktop with Core i7 - I use this PC for
development and in this case guest is nested guest, and host is normal
guest. Another hardware is some embedded board with Atom - here I don't
have nested virtualization - host runs on hw, and guest is normal guest.

G2H transmission (values are Gbit/s):

   Core i7 with nested guest.            Atom with normal guest.

*-------------------------------*   *-------------------------------*
|          |         |          |   |          |         |          |
| buf size |   copy  | zerocopy |   | buf size |   copy  | zerocopy |
|          |         |          |   |          |         |          |
*-------------------------------*   *-------------------------------*
|   4KB    |    3    |    10    |   |   4KB    |   0.8   |   1.9    |
*-------------------------------*   *-------------------------------*
|   32KB   |   20    |    61    |   |   32KB   |   6.8   |   20.2   |
*-------------------------------*   *-------------------------------*
|   256KB  |   33    |   244    |   |   256KB  |   7.8   |   55     |
*-------------------------------*   *-------------------------------*
|    1M    |   30    |   373    |   |    1M    |   7     |   95     |
*-------------------------------*   *-------------------------------*
|    8M    |   22    |   475    |   |    8M    |   7     |   114    |
*-------------------------------*   *-------------------------------*

H2G:

   Core i7 with nested guest.            Atom with normal guest.

*-------------------------------*   *-------------------------------*
|          |         |          |   |          |         |          |
| buf size |   copy  | zerocopy |   | buf size |   copy  | zerocopy |
|          |         |          |   |          |         |          |
*-------------------------------*   *-------------------------------*
|   4KB    |   20    |    10    |   |   4KB    |   4.37  |    3     |
*-------------------------------*   *-------------------------------*
|   32KB   |   37    |    75    |   |   32KB   |   11    |   18     |
*-------------------------------*   *-------------------------------*
|   256KB  |   44    |   299    |   |   256KB  |   11    |   62     |
*-------------------------------*   *-------------------------------*
|    1M    |   28    |   335    |   |    1M    |   9     |   77     |
*-------------------------------*   *-------------------------------*
|    8M    |   27    |   417    |   |    8M    |  9.35   |  115     |
*-------------------------------*   *-------------------------------*

 * Let's look to the first line of both tables - where copy is better
   than zerocopy. I analyzed this case more deeply and found that
   bottleneck is function 'vhost_work_queue()'. With 4K buffer size,
   caller spends too much time in it with zerocopy mode (comparing to
   copy mode). This happens only with 4K buffer size. This function just
   calls 'wake_up_process()' and its internal logic does not depends on
   skb, so i think potential reason (may be) is interval between two
   calls of this function (e.g. how often it is called). Note, that
   'vhost_work_queue()' differs from the same function at guest's side of
   transport: 'virtio_transport_send_pkt()' uses 'queue_work()' which
   i think is more optimized for worker purposes, than direct call to
   'wake_up_process()'. But again - this is just my assumption.

Loopback:

   Core i7 with nested guest.            Atom with normal guest.

*-------------------------------*   *-------------------------------*
|          |         |          |   |          |         |          |
| buf size |   copy  | zerocopy |   | buf size |   copy  | zerocopy |
|          |         |          |   |          |         |          |
*-------------------------------*   *-------------------------------*
|   4KB    |    8    |     7    |   |   4KB    |   1.8   |   1.3    |
*-------------------------------*   *-------------------------------*
|   32KB   |   38    |    44    |   |   32KB   |   10    |   10     |
*-------------------------------*   *-------------------------------*
|   256KB  |   55    |   168    |   |   256KB  |   15    |   36     |
*-------------------------------*   *-------------------------------*
|    1M    |   53    |   250    |   |    1M    |   12    |   45     |
*-------------------------------*   *-------------------------------*
|    8M    |   40    |   344    |   |    8M    |   11    |   74     |
*-------------------------------*   *-------------------------------*

I analyzed performace difference more deeply for the following setup:
server: ./vsock_perf --vsk-size 16M
client: ./vsock_perf --sender 2 --bytes 16M --buf-size 16K/4K [--zc]

In other words I send 16M of data from guest to host in copy/zerocopy
modes and with two different sizes of buffer - 4K and 64K. Let's see
to tx path for both modes - it consists of two steps:

copy:
1) Allocate skb of buffer's length.
2) Copy data to skb from buffer.

zerocopy:
1) Allocate skb with header space only.
2) Pin pages of the buffer and insert them to skb.

I measured average number of ns (returned by 'ktime_get()') for each
step above:
1) Skb allocation (for both copy and zerocopy modes).
2) For copy mode in 'memcpy_to_msg()' - copying.
3) For zerocopy mode in '__zerocopy_sg_from_iter()' - pinning.

Here are results for copy mode:
*-------------------------------------*
| buf | skb alloc | 'memcpy_to_msg()' |
*-------------------------------------*
|     |           |                   |
| 64K |  5000ns   |      25000ns      |
|     |           |                   |
*-------------------------------------*
|     |           |                   |
| 4K  |  800ns    |      2200ns       |
|     |           |                   |
*-------------------------------------*

Here are results for zerocopy mode:
*-----------------------------------------------*
| buf | skb alloc | '__zerocopy_sg_from_iter()' |
*-----------------------------------------------*
|     |           |                             |
| 64K |  250ns    |          3500ns             |
|     |           |                             |
*-----------------------------------------------*
|     |           |                             |
| 4K  |  250ns    |          3000ns             |
|     |           |                             |
*-----------------------------------------------*

I guess that reason of zerocopy performance is low overhead for page
pinning: there is big difference between 4K and 64K in case of copying
(25000 vs 2200), but in pinning case - just 3000 vs 3500.

So, zerocopy is faster than classic copy mode, but of course it requires
specific architecture of application due to user pages pinning, buffer
size and alignment.

                             NOTES

If host fails to send data with "Cannot allocate memory", check value
/proc/sys/net/core/optmem_max - it is accounted during completion skb
allocation. Try to update it to for example 1M and try send again:
"echo 1048576 > /proc/sys/net/core/optmem_max" (as root).

                            TESTING

This patchset includes set of tests for MSG_ZEROCOPY feature. I tried to
cover new code as much as possible so there are different cases for
MSG_ZEROCOPY transmissions: with disabled SO_ZEROCOPY and several io
vector types (different sizes, alignments, with unmapped pages). I also
run tests with loopback transport and run vsockmon. In v3 i've added
io_uring test as separated application.

           LET'S SPLIT PATCHSET TO MAKE REVIEW EASIER

In v3 Stefano Garzarella <sgarzare@redhat.com> asked to split this patchset
for several parts, because it looks too big for review. I think in this
version (v4) we can do it in the following way:

[0001 - 0005] - this is preparation for virtio/vhost part.
[0006 - 0009] - this is preparation for AF_VSOCK part.
[0010 - 0014] - these patches allows to trigger logic from the previous
                two parts. In addition 0014 is patch for Documentation.
[0015 - rest] - updates for tests, utils. This part doesn't touch kernel
                code and looks not critical.

Thanks, Arseniy

Link to v1:
https://lore.kernel.org/netdev/0e7c6fc4-b4a6-a27b-36e9-359597bba2b5@sberdevices.ru/
Link to v2:
https://lore.kernel.org/netdev/20230423192643.1537470-1-AVKrasnov@sberdevices.ru/
Link to v3:
https://lore.kernel.org/netdev/20230522073950.3574171-1-AVKrasnov@sberdevices.ru/
Link to v4:
https://lore.kernel.org/netdev/20230603204939.1598818-1-AVKrasnov@sberdevices.ru/

Changelog:
v1 -> v2:
 - Replace 'get_user_pages()' with 'pin_user_pages()'.
 - Loopback transport support.

v2 -> v3
 - Use 'get_user_pages()' instead of 'pin_user_pages()'. I think this
   is right approach, because i'm using '__zerocopy_sg_from_iter()'
   function. It is already implemented and used by io_uring zerocopy
   tx logic to 'pin' pages of user's buffer.

 - Use 'skb_copy_datagram_iter()' to copy data from both linear and
   non-linear skb to user's iov iter. It already has support for copying
   data from paged part of skb (by calling 'kmap()'). In v2 i used my
   own "from scratch" implemented function. With this and previous thing
   I significantly reduced LOC number in kernel part.

 - Add io_uring test for AF_VSOCK. It is implemented as separated util,
   because it depends on liburing (i think there is no need to link
   'vsock_test' with liburing, because io_uring functionality depends
   on environment - both in kernel and userspace).

 - Values from PERFORMANCE section are updated for all transports, but
   I didn't found any significant difference with v2.

 - More details in commit messages.

v3 -> v4:
 - Requirement for buffers to have page aligned base and size is removed,
   because virtio can handle such buffers.

 - Crash with SOCK_SEQPACKET is fixed. This is done by setting owner of
   new 'skb' before passing it to '__zerocopy_sg_from_iter()'. Last one
   dereferences owner of the passed skb without any checks (it was NULL).

 - Type of "owning" of the newly created skb is also changed: in v3 and
   before it was 'skb_set_owner_sk_safe()'. I replace it with this one:
   'skb_set_owner_w()'. This is because '__zerocopy_sg_from_iter()'
   increments 'sk_wmem_alloc' of socket which owns skb, thus we need a
   proper destructor which decrements it back - it is 'sock_wfree()'.
   This destructor is set by 'skb_set_owner_w()'. Otherwise we get leak
   of resource - such socket will be never deallocated.

 - Use ITER_KVEC instead of ITER_IOVEC when skb is copied to another one
   for passing to TAP device. Reason of this update is that ITER_IOVEC
   considered as userspace memory, while we have only kernel memory here.

v4 -> v5:
 - Problem 1) with dependency of SO_ZEROCOPY from the current transport
   is fixed.

 - See per patch changelog (after ---).

Arseniy Krasnov (17):
  vsock/virtio: read data from non-linear skb
  vhost/vsock: read data from non-linear skb
  vsock/virtio: support to send non-linear skb
  vsock/virtio: non-linear skb handling for tap
  vsock/virtio: MSG_ZEROCOPY flag support
  vsock: fix EPOLLERR set on non-empty error queue
  vsock: read from socket's error queue
  vsock: check for MSG_ZEROCOPY support on send
  vsock: enable SOCK_SUPPORT_ZC bit
  vhost/vsock: support MSG_ZEROCOPY for transport
  vsock/virtio: support MSG_ZEROCOPY for transport
  vsock/loopback: support MSG_ZEROCOPY for transport
  vsock: enable setting SO_ZEROCOPY
  docs: net: description of MSG_ZEROCOPY for AF_VSOCK
  test/vsock: MSG_ZEROCOPY flag tests
  test/vsock: MSG_ZEROCOPY support for vsock_perf
  test/vsock: io_uring rx/tx tests

 Documentation/networking/msg_zerocopy.rst |  12 +-
 drivers/vhost/vsock.c                     |  21 +-
 include/linux/socket.h                    |   1 +
 include/linux/virtio_vsock.h              |   1 +
 include/net/af_vsock.h                    |   7 +
 net/vmw_vsock/af_vsock.c                  |  61 +++-
 net/vmw_vsock/virtio_transport.c          |  47 +++-
 net/vmw_vsock/virtio_transport_common.c   | 313 ++++++++++++++++-----
 net/vmw_vsock/vsock_loopback.c            |   6 +
 tools/testing/vsock/Makefile              |   9 +-
 tools/testing/vsock/util.c                | 218 +++++++++++++++
 tools/testing/vsock/util.h                |  18 ++
 tools/testing/vsock/vsock_perf.c          | 139 +++++++++-
 tools/testing/vsock/vsock_test.c          |  16 ++
 tools/testing/vsock/vsock_test_zerocopy.c | 312 +++++++++++++++++++++
 tools/testing/vsock/vsock_test_zerocopy.h |  15 +
 tools/testing/vsock/vsock_uring_test.c    | 321 ++++++++++++++++++++++
 17 files changed, 1423 insertions(+), 94 deletions(-)
 create mode 100644 tools/testing/vsock/vsock_test_zerocopy.c
 create mode 100644 tools/testing/vsock/vsock_test_zerocopy.h
 create mode 100644 tools/testing/vsock/vsock_uring_test.c

Comments

Stefano Garzarella July 6, 2023, 5:07 p.m. UTC | #1
On Sat, Jul 01, 2023 at 09:39:30AM +0300, Arseniy Krasnov wrote:
>Hello,
>
>                           DESCRIPTION
>
>this is MSG_ZEROCOPY feature support for virtio/vsock. I tried to follow
>current implementation for TCP as much as possible:
>
>1) Sender must enable SO_ZEROCOPY flag to use this feature. Without this
>   flag, data will be sent in "classic" copy manner and MSG_ZEROCOPY
>   flag will be ignored (e.g. without completion).
>
>2) Kernel uses completions from socket's error queue. Single completion
>   for single tx syscall (or it can merge several completions to single
>   one). I used already implemented logic for MSG_ZEROCOPY support:
>   'msg_zerocopy_realloc()' etc.
>
>Difference with copy way is not significant. During packet allocation,
>non-linear skb is created and filled with pinned user pages.
>There are also some updates for vhost and guest parts of transport - in
>both cases i've added handling of non-linear skb for virtio part. vhost
>copies data from such skb to the guest's rx virtio buffers. In the guest,
>virtio transport fills tx virtio queue with pages from skb.
>
>Head of this patchset is:
>https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next.git/commit/?id=d20dd0ea14072e8a90ff864b2c1603bd68920b4b
>
>
>This version has several limits/problems (all resolved at v5):
>
>1) As this feature totally depends on transport, there is no way (or it
>   is difficult) to check whether transport is able to handle it or not
>   during SO_ZEROCOPY setting. Seems I need to call AF_VSOCK specific
>   setsockopt callback from setsockopt callback for SOL_SOCKET, but this
>   leads to lock problem, because both AF_VSOCK and SOL_SOCKET callback
>   are not considered to be called from each other. So in current version
>   SO_ZEROCOPY is set successfully to any type (e.g. transport) of
>   AF_VSOCK socket, but if transport does not support MSG_ZEROCOPY,
>   tx routine will fail with EOPNOTSUPP.
>
>   ^^^ fixed in v5. Thanks to Bobby Eshleman.
>
>2) When MSG_ZEROCOPY is used, for each tx system call we need to enqueue
>   one completion. In each completion there is flag which shows how tx
>   was performed: zerocopy or copy. This leads that whole message must
>   be send in zerocopy or copy way - we can't send part of message with
>   copying and rest of message with zerocopy mode (or vice versa). Now,
>   we need to account vsock credit logic, e.g. we can't send whole data
>   once - only allowed number of bytes could sent at any moment. In case
>   of copying way there is no problem as in worst case we can send single
>   bytes, but zerocopy is more complex because smallest transmission
>   unit is single page. So if there is not enough space at peer's side
>   to send integer number of pages (at least one) - we will wait, thus
>   stalling tx side. To overcome this problem i've added simple rule -
>   zerocopy is possible only when there is enough space at another side
>   for whole message (to check, that current 'msghdr' was already used
>   in previous tx iterations i use 'iov_offset' field of it's iov iter).
>
>   ^^^
>   Discussed as ok during v2. Link:
>   https://lore.kernel.org/netdev/23guh3txkghxpgcrcjx7h62qsoj3xgjhfzgtbmqp2slrz3rxr4@zya2z7kwt75l/
>
>3) loopback transport is not supported, because it requires to implement
>   non-linear skb handling in dequeue logic (as we "send" fragged skb
>   and "receive" it from the same queue). I'm going to implement it in
>   next versions.
>
>   ^^^ fixed in v2
>
>4) Current implementation sets max length of packet to 64KB. IIUC this
>   is due to 'kmalloc()' allocated data buffers. I think, in case of
>   MSG_ZEROCOPY this value could be increased, because 'kmalloc()' is
>   not touched for data - user space pages are used as buffers. Also
>   this limit trims every message which is > 64KB, thus such messages
>   will be send in copy mode due to 'iov_offset' check in 2).
>
>   ^^^ fixed in v2
>
>                         PATCHSET STRUCTURE
>
>Patchset has the following structure:
>1) Handle non-linear skbuff on receive in virtio/vhost.
>2) Handle non-linear skbuff on send in virtio/vhost.
>3) Updates for AF_VSOCK.
>4) Enable MSG_ZEROCOPY support on transports.
>5) Tests/tools/docs updates.
>
>                            PERFORMANCE
>
>Performance: it is a little bit tricky to compare performance between
>copy and zerocopy transmissions. In zerocopy way we need to wait when
>user buffers will be released by kernel, so it is like synchronous
>path (wait until device driver will process it), while in copy way we
>can feed data to kernel as many as we want, don't care about device
>driver. So I compared only time which we spend in the 'send()' syscall.
>Then if this value will be combined with total number of transmitted
>bytes, we can get Gbit/s parameter. Also to avoid tx stalls due to not
>enough credit, receiver allocates same amount of space as sender needs.
>
>Sender:
>./vsock_perf --sender <CID> --buf-size <buf size> --bytes 256M [--zc]
>
>Receiver:
>./vsock_perf --vsk-size 256M
>
>I run tests on two setups: desktop with Core i7 - I use this PC for
>development and in this case guest is nested guest, and host is normal
>guest. Another hardware is some embedded board with Atom - here I don't
>have nested virtualization - host runs on hw, and guest is normal guest.
>
>G2H transmission (values are Gbit/s):
>
>   Core i7 with nested guest.            Atom with normal guest.
>
>*-------------------------------*   *-------------------------------*
>|          |         |          |   |          |         |          |
>| buf size |   copy  | zerocopy |   | buf size |   copy  | zerocopy |
>|          |         |          |   |          |         |          |
>*-------------------------------*   *-------------------------------*
>|   4KB    |    3    |    10    |   |   4KB    |   0.8   |   1.9    |
>*-------------------------------*   *-------------------------------*
>|   32KB   |   20    |    61    |   |   32KB   |   6.8   |   20.2   |
>*-------------------------------*   *-------------------------------*
>|   256KB  |   33    |   244    |   |   256KB  |   7.8   |   55     |
>*-------------------------------*   *-------------------------------*
>|    1M    |   30    |   373    |   |    1M    |   7     |   95     |
>*-------------------------------*   *-------------------------------*
>|    8M    |   22    |   475    |   |    8M    |   7     |   114    |
>*-------------------------------*   *-------------------------------*
>
>H2G:
>
>   Core i7 with nested guest.            Atom with normal guest.
>
>*-------------------------------*   *-------------------------------*
>|          |         |          |   |          |         |          |
>| buf size |   copy  | zerocopy |   | buf size |   copy  | zerocopy |
>|          |         |          |   |          |         |          |
>*-------------------------------*   *-------------------------------*
>|   4KB    |   20    |    10    |   |   4KB    |   4.37  |    3     |
>*-------------------------------*   *-------------------------------*
>|   32KB   |   37    |    75    |   |   32KB   |   11    |   18     |
>*-------------------------------*   *-------------------------------*
>|   256KB  |   44    |   299    |   |   256KB  |   11    |   62     |
>*-------------------------------*   *-------------------------------*
>|    1M    |   28    |   335    |   |    1M    |   9     |   77     |
>*-------------------------------*   *-------------------------------*
>|    8M    |   27    |   417    |   |    8M    |  9.35   |  115     |
>*-------------------------------*   *-------------------------------*
>
> * Let's look to the first line of both tables - where copy is better
>   than zerocopy. I analyzed this case more deeply and found that
>   bottleneck is function 'vhost_work_queue()'. With 4K buffer size,
>   caller spends too much time in it with zerocopy mode (comparing to
>   copy mode). This happens only with 4K buffer size. This function just
>   calls 'wake_up_process()' and its internal logic does not depends on
>   skb, so i think potential reason (may be) is interval between two
>   calls of this function (e.g. how often it is called). Note, that
>   'vhost_work_queue()' differs from the same function at guest's side of
>   transport: 'virtio_transport_send_pkt()' uses 'queue_work()' which
>   i think is more optimized for worker purposes, than direct call to
>   'wake_up_process()'. But again - this is just my assumption.
>
>Loopback:
>
>   Core i7 with nested guest.            Atom with normal guest.
>
>*-------------------------------*   *-------------------------------*
>|          |         |          |   |          |         |          |
>| buf size |   copy  | zerocopy |   | buf size |   copy  | zerocopy |
>|          |         |          |   |          |         |          |
>*-------------------------------*   *-------------------------------*
>|   4KB    |    8    |     7    |   |   4KB    |   1.8   |   1.3    |
>*-------------------------------*   *-------------------------------*
>|   32KB   |   38    |    44    |   |   32KB   |   10    |   10     |
>*-------------------------------*   *-------------------------------*
>|   256KB  |   55    |   168    |   |   256KB  |   15    |   36     |
>*-------------------------------*   *-------------------------------*
>|    1M    |   53    |   250    |   |    1M    |   12    |   45     |
>*-------------------------------*   *-------------------------------*
>|    8M    |   40    |   344    |   |    8M    |   11    |   74     |
>*-------------------------------*   *-------------------------------*
>
>I analyzed performace difference more deeply for the following setup:
>server: ./vsock_perf --vsk-size 16M
>client: ./vsock_perf --sender 2 --bytes 16M --buf-size 16K/4K [--zc]
>
>In other words I send 16M of data from guest to host in copy/zerocopy
>modes and with two different sizes of buffer - 4K and 64K. Let's see
>to tx path for both modes - it consists of two steps:
>
>copy:
>1) Allocate skb of buffer's length.
>2) Copy data to skb from buffer.
>
>zerocopy:
>1) Allocate skb with header space only.
>2) Pin pages of the buffer and insert them to skb.
>
>I measured average number of ns (returned by 'ktime_get()') for each
>step above:
>1) Skb allocation (for both copy and zerocopy modes).
>2) For copy mode in 'memcpy_to_msg()' - copying.
>3) For zerocopy mode in '__zerocopy_sg_from_iter()' - pinning.
>
>Here are results for copy mode:
>*-------------------------------------*
>| buf | skb alloc | 'memcpy_to_msg()' |
>*-------------------------------------*
>|     |           |                   |
>| 64K |  5000ns   |      25000ns      |
>|     |           |                   |
>*-------------------------------------*
>|     |           |                   |
>| 4K  |  800ns    |      2200ns       |
>|     |           |                   |
>*-------------------------------------*
>
>Here are results for zerocopy mode:
>*-----------------------------------------------*
>| buf | skb alloc | '__zerocopy_sg_from_iter()' |
>*-----------------------------------------------*
>|     |           |                             |
>| 64K |  250ns    |          3500ns             |
>|     |           |                             |
>*-----------------------------------------------*
>|     |           |                             |
>| 4K  |  250ns    |          3000ns             |
>|     |           |                             |
>*-----------------------------------------------*
>
>I guess that reason of zerocopy performance is low overhead for page
>pinning: there is big difference between 4K and 64K in case of copying
>(25000 vs 2200), but in pinning case - just 3000 vs 3500.
>
>So, zerocopy is faster than classic copy mode, but of course it requires
>specific architecture of application due to user pages pinning, buffer
>size and alignment.
>
>                             NOTES
>
>If host fails to send data with "Cannot allocate memory", check value
>/proc/sys/net/core/optmem_max - it is accounted during completion skb
>allocation. Try to update it to for example 1M and try send again:
>"echo 1048576 > /proc/sys/net/core/optmem_max" (as root).
>
>                            TESTING
>
>This patchset includes set of tests for MSG_ZEROCOPY feature. I tried to
>cover new code as much as possible so there are different cases for
>MSG_ZEROCOPY transmissions: with disabled SO_ZEROCOPY and several io
>vector types (different sizes, alignments, with unmapped pages). I also
>run tests with loopback transport and run vsockmon. In v3 i've added
>io_uring test as separated application.
>
>           LET'S SPLIT PATCHSET TO MAKE REVIEW EASIER
>
>In v3 Stefano Garzarella <sgarzare@redhat.com> asked to split this patchset
>for several parts, because it looks too big for review. I think in this
>version (v4) we can do it in the following way:
>
>[0001 - 0005] - this is preparation for virtio/vhost part.
>[0006 - 0009] - this is preparation for AF_VSOCK part.
>[0010 - 0014] - these patches allows to trigger logic from the previous
>                two parts. In addition 0014 is patch for Documentation.
>[0015 - rest] - updates for tests, utils. This part doesn't touch kernel
>                code and looks not critical.

Great!

So IIUC all the issues are fixed. I left some comments, but I think
you can start sending the virtio/vhost preparation patches to net-next
(when it will re-open).

I just pointend out something to fix, and that maybe we can merge
the first 2 patches.

I think you can restart with v0, describing in the cover letter that
the patches was part of this RFC.

Thanks,
Stefano
Arseniy Krasnov July 7, 2023, 4:41 a.m. UTC | #2
On 06.07.2023 20:07, Stefano Garzarella wrote:
> On Sat, Jul 01, 2023 at 09:39:30AM +0300, Arseniy Krasnov wrote:
>> Hello,
>>
>>                           DESCRIPTION
>>
>> this is MSG_ZEROCOPY feature support for virtio/vsock. I tried to follow
>> current implementation for TCP as much as possible:
>>
>> 1) Sender must enable SO_ZEROCOPY flag to use this feature. Without this
>>   flag, data will be sent in "classic" copy manner and MSG_ZEROCOPY
>>   flag will be ignored (e.g. without completion).
>>
>> 2) Kernel uses completions from socket's error queue. Single completion
>>   for single tx syscall (or it can merge several completions to single
>>   one). I used already implemented logic for MSG_ZEROCOPY support:
>>   'msg_zerocopy_realloc()' etc.
>>
>> Difference with copy way is not significant. During packet allocation,
>> non-linear skb is created and filled with pinned user pages.
>> There are also some updates for vhost and guest parts of transport - in
>> both cases i've added handling of non-linear skb for virtio part. vhost
>> copies data from such skb to the guest's rx virtio buffers. In the guest,
>> virtio transport fills tx virtio queue with pages from skb.
>>
>> Head of this patchset is:
>> https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next.git/commit/?id=d20dd0ea14072e8a90ff864b2c1603bd68920b4b
>>
>>
>> This version has several limits/problems (all resolved at v5):
>>
>> 1) As this feature totally depends on transport, there is no way (or it
>>   is difficult) to check whether transport is able to handle it or not
>>   during SO_ZEROCOPY setting. Seems I need to call AF_VSOCK specific
>>   setsockopt callback from setsockopt callback for SOL_SOCKET, but this
>>   leads to lock problem, because both AF_VSOCK and SOL_SOCKET callback
>>   are not considered to be called from each other. So in current version
>>   SO_ZEROCOPY is set successfully to any type (e.g. transport) of
>>   AF_VSOCK socket, but if transport does not support MSG_ZEROCOPY,
>>   tx routine will fail with EOPNOTSUPP.
>>
>>   ^^^ fixed in v5. Thanks to Bobby Eshleman.
>>
>> 2) When MSG_ZEROCOPY is used, for each tx system call we need to enqueue
>>   one completion. In each completion there is flag which shows how tx
>>   was performed: zerocopy or copy. This leads that whole message must
>>   be send in zerocopy or copy way - we can't send part of message with
>>   copying and rest of message with zerocopy mode (or vice versa). Now,
>>   we need to account vsock credit logic, e.g. we can't send whole data
>>   once - only allowed number of bytes could sent at any moment. In case
>>   of copying way there is no problem as in worst case we can send single
>>   bytes, but zerocopy is more complex because smallest transmission
>>   unit is single page. So if there is not enough space at peer's side
>>   to send integer number of pages (at least one) - we will wait, thus
>>   stalling tx side. To overcome this problem i've added simple rule -
>>   zerocopy is possible only when there is enough space at another side
>>   for whole message (to check, that current 'msghdr' was already used
>>   in previous tx iterations i use 'iov_offset' field of it's iov iter).
>>
>>   ^^^
>>   Discussed as ok during v2. Link:
>>   https://lore.kernel.org/netdev/23guh3txkghxpgcrcjx7h62qsoj3xgjhfzgtbmqp2slrz3rxr4@zya2z7kwt75l/
>>
>> 3) loopback transport is not supported, because it requires to implement
>>   non-linear skb handling in dequeue logic (as we "send" fragged skb
>>   and "receive" it from the same queue). I'm going to implement it in
>>   next versions.
>>
>>   ^^^ fixed in v2
>>
>> 4) Current implementation sets max length of packet to 64KB. IIUC this
>>   is due to 'kmalloc()' allocated data buffers. I think, in case of
>>   MSG_ZEROCOPY this value could be increased, because 'kmalloc()' is
>>   not touched for data - user space pages are used as buffers. Also
>>   this limit trims every message which is > 64KB, thus such messages
>>   will be send in copy mode due to 'iov_offset' check in 2).
>>
>>   ^^^ fixed in v2
>>
>>                         PATCHSET STRUCTURE
>>
>> Patchset has the following structure:
>> 1) Handle non-linear skbuff on receive in virtio/vhost.
>> 2) Handle non-linear skbuff on send in virtio/vhost.
>> 3) Updates for AF_VSOCK.
>> 4) Enable MSG_ZEROCOPY support on transports.
>> 5) Tests/tools/docs updates.
>>
>>                            PERFORMANCE
>>
>> Performance: it is a little bit tricky to compare performance between
>> copy and zerocopy transmissions. In zerocopy way we need to wait when
>> user buffers will be released by kernel, so it is like synchronous
>> path (wait until device driver will process it), while in copy way we
>> can feed data to kernel as many as we want, don't care about device
>> driver. So I compared only time which we spend in the 'send()' syscall.
>> Then if this value will be combined with total number of transmitted
>> bytes, we can get Gbit/s parameter. Also to avoid tx stalls due to not
>> enough credit, receiver allocates same amount of space as sender needs.
>>
>> Sender:
>> ./vsock_perf --sender <CID> --buf-size <buf size> --bytes 256M [--zc]
>>
>> Receiver:
>> ./vsock_perf --vsk-size 256M
>>
>> I run tests on two setups: desktop with Core i7 - I use this PC for
>> development and in this case guest is nested guest, and host is normal
>> guest. Another hardware is some embedded board with Atom - here I don't
>> have nested virtualization - host runs on hw, and guest is normal guest.
>>
>> G2H transmission (values are Gbit/s):
>>
>>   Core i7 with nested guest.            Atom with normal guest.
>>
>> *-------------------------------*   *-------------------------------*
>> |          |         |          |   |          |         |          |
>> | buf size |   copy  | zerocopy |   | buf size |   copy  | zerocopy |
>> |          |         |          |   |          |         |          |
>> *-------------------------------*   *-------------------------------*
>> |   4KB    |    3    |    10    |   |   4KB    |   0.8   |   1.9    |
>> *-------------------------------*   *-------------------------------*
>> |   32KB   |   20    |    61    |   |   32KB   |   6.8   |   20.2   |
>> *-------------------------------*   *-------------------------------*
>> |   256KB  |   33    |   244    |   |   256KB  |   7.8   |   55     |
>> *-------------------------------*   *-------------------------------*
>> |    1M    |   30    |   373    |   |    1M    |   7     |   95     |
>> *-------------------------------*   *-------------------------------*
>> |    8M    |   22    |   475    |   |    8M    |   7     |   114    |
>> *-------------------------------*   *-------------------------------*
>>
>> H2G:
>>
>>   Core i7 with nested guest.            Atom with normal guest.
>>
>> *-------------------------------*   *-------------------------------*
>> |          |         |          |   |          |         |          |
>> | buf size |   copy  | zerocopy |   | buf size |   copy  | zerocopy |
>> |          |         |          |   |          |         |          |
>> *-------------------------------*   *-------------------------------*
>> |   4KB    |   20    |    10    |   |   4KB    |   4.37  |    3     |
>> *-------------------------------*   *-------------------------------*
>> |   32KB   |   37    |    75    |   |   32KB   |   11    |   18     |
>> *-------------------------------*   *-------------------------------*
>> |   256KB  |   44    |   299    |   |   256KB  |   11    |   62     |
>> *-------------------------------*   *-------------------------------*
>> |    1M    |   28    |   335    |   |    1M    |   9     |   77     |
>> *-------------------------------*   *-------------------------------*
>> |    8M    |   27    |   417    |   |    8M    |  9.35   |  115     |
>> *-------------------------------*   *-------------------------------*
>>
>> * Let's look to the first line of both tables - where copy is better
>>   than zerocopy. I analyzed this case more deeply and found that
>>   bottleneck is function 'vhost_work_queue()'. With 4K buffer size,
>>   caller spends too much time in it with zerocopy mode (comparing to
>>   copy mode). This happens only with 4K buffer size. This function just
>>   calls 'wake_up_process()' and its internal logic does not depends on
>>   skb, so i think potential reason (may be) is interval between two
>>   calls of this function (e.g. how often it is called). Note, that
>>   'vhost_work_queue()' differs from the same function at guest's side of
>>   transport: 'virtio_transport_send_pkt()' uses 'queue_work()' which
>>   i think is more optimized for worker purposes, than direct call to
>>   'wake_up_process()'. But again - this is just my assumption.
>>
>> Loopback:
>>
>>   Core i7 with nested guest.            Atom with normal guest.
>>
>> *-------------------------------*   *-------------------------------*
>> |          |         |          |   |          |         |          |
>> | buf size |   copy  | zerocopy |   | buf size |   copy  | zerocopy |
>> |          |         |          |   |          |         |          |
>> *-------------------------------*   *-------------------------------*
>> |   4KB    |    8    |     7    |   |   4KB    |   1.8   |   1.3    |
>> *-------------------------------*   *-------------------------------*
>> |   32KB   |   38    |    44    |   |   32KB   |   10    |   10     |
>> *-------------------------------*   *-------------------------------*
>> |   256KB  |   55    |   168    |   |   256KB  |   15    |   36     |
>> *-------------------------------*   *-------------------------------*
>> |    1M    |   53    |   250    |   |    1M    |   12    |   45     |
>> *-------------------------------*   *-------------------------------*
>> |    8M    |   40    |   344    |   |    8M    |   11    |   74     |
>> *-------------------------------*   *-------------------------------*
>>
>> I analyzed performace difference more deeply for the following setup:
>> server: ./vsock_perf --vsk-size 16M
>> client: ./vsock_perf --sender 2 --bytes 16M --buf-size 16K/4K [--zc]
>>
>> In other words I send 16M of data from guest to host in copy/zerocopy
>> modes and with two different sizes of buffer - 4K and 64K. Let's see
>> to tx path for both modes - it consists of two steps:
>>
>> copy:
>> 1) Allocate skb of buffer's length.
>> 2) Copy data to skb from buffer.
>>
>> zerocopy:
>> 1) Allocate skb with header space only.
>> 2) Pin pages of the buffer and insert them to skb.
>>
>> I measured average number of ns (returned by 'ktime_get()') for each
>> step above:
>> 1) Skb allocation (for both copy and zerocopy modes).
>> 2) For copy mode in 'memcpy_to_msg()' - copying.
>> 3) For zerocopy mode in '__zerocopy_sg_from_iter()' - pinning.
>>
>> Here are results for copy mode:
>> *-------------------------------------*
>> | buf | skb alloc | 'memcpy_to_msg()' |
>> *-------------------------------------*
>> |     |           |                   |
>> | 64K |  5000ns   |      25000ns      |
>> |     |           |                   |
>> *-------------------------------------*
>> |     |           |                   |
>> | 4K  |  800ns    |      2200ns       |
>> |     |           |                   |
>> *-------------------------------------*
>>
>> Here are results for zerocopy mode:
>> *-----------------------------------------------*
>> | buf | skb alloc | '__zerocopy_sg_from_iter()' |
>> *-----------------------------------------------*
>> |     |           |                             |
>> | 64K |  250ns    |          3500ns             |
>> |     |           |                             |
>> *-----------------------------------------------*
>> |     |           |                             |
>> | 4K  |  250ns    |          3000ns             |
>> |     |           |                             |
>> *-----------------------------------------------*
>>
>> I guess that reason of zerocopy performance is low overhead for page
>> pinning: there is big difference between 4K and 64K in case of copying
>> (25000 vs 2200), but in pinning case - just 3000 vs 3500.
>>
>> So, zerocopy is faster than classic copy mode, but of course it requires
>> specific architecture of application due to user pages pinning, buffer
>> size and alignment.
>>
>>                             NOTES
>>
>> If host fails to send data with "Cannot allocate memory", check value
>> /proc/sys/net/core/optmem_max - it is accounted during completion skb
>> allocation. Try to update it to for example 1M and try send again:
>> "echo 1048576 > /proc/sys/net/core/optmem_max" (as root).
>>
>>                            TESTING
>>
>> This patchset includes set of tests for MSG_ZEROCOPY feature. I tried to
>> cover new code as much as possible so there are different cases for
>> MSG_ZEROCOPY transmissions: with disabled SO_ZEROCOPY and several io
>> vector types (different sizes, alignments, with unmapped pages). I also
>> run tests with loopback transport and run vsockmon. In v3 i've added
>> io_uring test as separated application.
>>
>>           LET'S SPLIT PATCHSET TO MAKE REVIEW EASIER
>>
>> In v3 Stefano Garzarella <sgarzare@redhat.com> asked to split this patchset
>> for several parts, because it looks too big for review. I think in this
>> version (v4) we can do it in the following way:
>>
>> [0001 - 0005] - this is preparation for virtio/vhost part.
>> [0006 - 0009] - this is preparation for AF_VSOCK part.
>> [0010 - 0014] - these patches allows to trigger logic from the previous
>>                two parts. In addition 0014 is patch for Documentation.
>> [0015 - rest] - updates for tests, utils. This part doesn't touch kernel
>>                code and looks not critical.
> 
> Great!

Thanks for review! All comments are clear for me.

> 
> So IIUC all the issues are fixed. I left some comments, but I think
> you can start sending the virtio/vhost preparation patches to net-next
> (when it will re-open).
> 
> I just pointend out something to fix, and that maybe we can merge
> the first 2 patches.
> 
> I think you can restart with v0, describing in the cover letter that
> the patches was part of this RFC.

Ok, I'll fix comments and send 0001-0005 (with first two merged) in a single
net-next patchset!

Thanks, Arseniy

> 
> Thanks,
> Stefano
>