Date:   Mon, 6 Feb 2023 06:51:55 +0000
From:   Arseniy Krasnov <AVKrasnov@...rdevices.ru>
To:     Stefan Hajnoczi <stefanha@...hat.com>,
        Stefano Garzarella <sgarzare@...hat.com>,
        "Michael S. Tsirkin" <mst@...hat.com>,
        Jason Wang <jasowang@...hat.com>,
        "David S. Miller" <davem@...emloft.net>,
        Eric Dumazet <edumazet@...gle.com>,
        Jakub Kicinski <kuba@...nel.org>,
        Paolo Abeni <pabeni@...hat.com>,
        Arseniy Krasnov <AVKrasnov@...rdevices.ru>,
        "Krasnov Arseniy" <oxffffaa@...il.com>
CC:     "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "kvm@...r.kernel.org" <kvm@...r.kernel.org>,
        "virtualization@...ts.linux-foundation.org" 
        <virtualization@...ts.linux-foundation.org>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        kernel <kernel@...rdevices.ru>
Subject: [RFC PATCH v1 00/12] vsock: MSG_ZEROCOPY flag support

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, then I call 'get_user_pages()' for each page
from user's iov iterator (I think i don't need 'pin_user_pages()' as
there is no backing storage for these pages) and add each returned page
to the skb as fragment. 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 virtio queue with pages
from skb.

I think doc in Documentation/networking/msg_zerocopy.rst could be also
updated in next versions.

This version has several limits/problems:

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.

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).

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.

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).

                            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 something 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 'sendmsg()' syscall.
Also there is limit from 4) above so max buffer size is 64KB. I've
tested this patchset in the nested VM, but i think for V1 it is not a
big deal.

Sender:
./vsock_perf --sender <CID> --buf-size <buf size> --bytes 60M [--zc]

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

Number in cell is seconds which senders spends inside tx syscall.

Guest to host transmission:

*-------------------------------*
|          |         |          |
| buf size |   copy  | zerocopy |
|          |         |          |
*-------------------------------*
|   4KB    |  0.26   |   0.042  |
*-------------------------------*
|   16KB   |  0.11   |   0.014  |
*-------------------------------*
|   32KB   |  0.05   |   0.009  |
*-------------------------------*
|   64KB   |  0.04   |   0.005  |
*-------------------------------*

Host to guest transmission:

*--------------------------------*
|          |          |          |
| buf size |   copy   | zerocopy |
|          |          |          |
*--------------------------------*
|   4KB    |   0.049  |   0.034  |
*--------------------------------*
|   16KB   |   0.03   |   0.024  |
*--------------------------------*
|   32KB   |   0.025  |   0.01   |
*--------------------------------*
|   64KB   |   0.028  |   0.01   |
*--------------------------------*

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.

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. In next versions i'm going to fix 64KB barrier to
perform tests with bigger buffer sizes.

                            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).

Thanks, Arseniy

Arseniy Krasnov(12):
 vsock: check error queue to set EPOLLERR
 vsock: read from socket's error queue
 vsock: check for MSG_ZEROCOPY support
 vhost/vsock: non-linear skb handling support
 vsock/virtio: non-linear skb support
 vsock/virtio: non-linear skb handling for TAP dev
 vsock/virtio: MGS_ZEROCOPY flag support
 vhost/vsock: support MSG_ZEROCOPY for transport
 vsock/virtio: support MSG_ZEROCOPY for transport
 net/sock: enable setting SO_ZEROCOPY for PF_VSOCK
 test/vsock: MSG_ZEROCOPY flag tests
 test/vsock: MSG_ZEROCOPY support for vsock_perf

 drivers/vhost/vsock.c                     |  62 +++-
 include/linux/socket.h                    |   1 +
 include/linux/virtio_vsock.h              |  12 +
 include/net/af_vsock.h                    |   2 +
 net/core/sock.c                           |   4 +-
 net/vmw_vsock/af_vsock.c                  |  35 ++-
 net/vmw_vsock/virtio_transport.c          |  38 ++-
 net/vmw_vsock/virtio_transport_common.c   | 255 ++++++++++++++--
 tools/testing/vsock/Makefile              |   2 +-
 tools/testing/vsock/util.h                |   1 +
 tools/testing/vsock/vsock_perf.c          | 127 +++++++-
 tools/testing/vsock/vsock_test.c          |  11 +
 tools/testing/vsock/vsock_test_zerocopy.c | 470 ++++++++++++++++++++++++++++++
 tools/testing/vsock/vsock_test_zerocopy.h |  12 +
 14 files changed, 991 insertions(+), 41 deletions(-)

-- 
2.25.1

Powered by blists - more mailing lists