[v9,0/7] MSG_ZEROCOPY + multifd

Message

Leonardo Bras April 25, 2022, 9:50 p.m. UTC

This patch series intends to enable MSG_ZEROCOPY in QIOChannel, and make
use of it for multifd migration performance improvement, by reducing cpu
usage.

Patch #1 creates new callbacks for QIOChannel, allowing the implementation
of zero copy writing.

Patch #2 implements io_writev flags and io_flush() on QIOChannelSocket,
making use of MSG_ZEROCOPY on Linux.

Patch #3 adds a "zero_copy_send" migration property, only available with
CONFIG_LINUX, and compiled-out in any other architectures.
This migration property has to be enabled before multifd migration starts.

Patch #4 adds a helper function that allows to see if TLS is going to be used.
This helper will be later used in patch #5.

Patch #5 changes multifd_send_sync_main() so it returns int instead of void.
The return value is used to understand if any error happened in the function,
allowing migration to possible fail earlier.

Patch #6 implements an workaround: The behavior introduced in d48c3a0445 is
hard to deal with in zerocopy, so a workaround is introduced to send the
header in a different syscall, without MSG_ZEROCOPY.

Patch #7 Makes use of QIOChannelSocket zero_copy implementation on
nocomp multifd migration.

Results:
In preliminary tests, the resource usage of __sys_sendmsg() reduced 15 times,
and the overall migration took 13-22% less time, based in synthetic cpu
workload.

In further tests, it was noted that, on multifd migration with 8 channels:
- On idle hosts, migration time reduced in 10% to 21%.
- On hosts busy with heavy cpu stress (1 stress thread per cpu, but
  not cpu-pinned) migration time reduced in ~25% by enabling zero-copy.
- On hosts with heavy cpu-pinned workloads (1 stress thread per cpu, 
  cpu-pinned), migration time reducted in ~66% by enabling zero-copy.

Above tests setup:
- Sending and Receiving hosts:
  - CPU : Intel(R) Xeon(R) Platinum 8276L CPU @ 2.20GHz (448 CPUS)
  - Network card: E810-C (100Gbps)
  - >1TB RAM
  - QEMU: Upstream master branch + This patchset
  - Linux: Upstream v5.15 
- VM configuration:
  - 28 VCPUs
  - 512GB RAM


---
Changes since v8:
- Inserted two new patches #5 & #6, previous patch #5 is now #7.
- Workaround an optimization introduced in d48c3a0445
- Removed unnecessary assert in qio_channel_writev_full_all

Changes since v7:
- Migration property renamed from zero-copy to zero-copy-send
- A few early tests added to help misconfigurations to fail earlier
- qio_channel_full*_flags() renamed back to qio_channel_full*()
- multifd_send_sync_main() reverted back to not receiving a flag,
  so it always sync zero-copy when enabled.
- Improve code quality on a few points

Changes since v6:
- Remove io_writev_zero_copy(), and makes use of io_writev() new flags
  to achieve the same results.
- Rename io_flush_zero_copy() to io_flush()
- Previous patch #2 became too small, so it was squashed in previous
  patch #3 (now patch #2)

Changes since v5:
- flush_zero_copy now returns -1 on fail, 0 on success, and 1 when all
  processed writes were not able to use zerocopy in kernel.
- qio_channel_socket_poll() removed, using qio_channel_wait() instead
- ENOBUFS is now processed inside qio_channel_socket_writev_flags()
- Most zerocopy parameter validation moved to migrate_params_check(),
  leaving only feature test to socket_outgoing_migration() callback
- Naming went from *zerocopy to *zero_copy or *zero-copy, due to QAPI/QMP
  preferences
- Improved docs

Changes since v4:
- 3 patches got splitted in 6
- Flush is used for syncing after each iteration, instead of only at the end
- If zerocopy is not available, fail in connect instead of failing on write
- 'multifd-zerocopy' property renamed to 'zerocopy'
- Fail migrations that don't support zerocopy, if it's enabled.
- Instead of checking for zerocopy at each write, save the flags in
  MultiFDSendParams->write_flags and use them on write
- Reorganized flag usage in QIOChannelSocket 
- A lot of typos fixed
- More doc on buffer restrictions

Changes since v3:
- QIOChannel interface names changed from io_async_{writev,flush} to
  io_{writev,flush}_zerocopy
- Instead of falling back in case zerocopy is not implemented, return
  error and abort operation.
- Flush now waits as long as needed, or return error in case anything
  goes wrong, aborting the operation.
- Zerocopy is now conditional in multifd, being set by parameter
  multifd-zerocopy
- Moves zerocopy_flush to multifd_send_sync_main() from multifd_save_cleanup
  so migration can abort if flush goes wrong.
- Several other small improvements

Changes since v2:
- Patch #1: One more fallback
- Patch #2: Fall back to sync if fails to lock buffer memory in MSG_ZEROCOPY send.

Changes since v1:
- Reimplemented the patchset using async_write + async_flush approach.
- Implemented a flush to be able to tell whenever all data was written.


Leonardo Bras (7):
  QIOChannel: Add flags on io_writev and introduce io_flush callback
  QIOChannelSocket: Implement io_writev zero copy flag & io_flush for
    CONFIG_LINUX
  migration: Add zero-copy-send parameter for QMP/HMP for Linux
  migration: Add migrate_use_tls() helper
  multifd: multifd_send_sync_main now returns negative on error
  multifd: Send header packet without flags if zero-copy-send is enabled
  multifd: Implement zero copy write in multifd migration
    (multifd-zero-copy)

 qapi/migration.json                 |  24 ++++++
 include/io/channel-socket.h         |   2 +
 include/io/channel.h                |  38 +++++++++-
 migration/migration.h               |   6 ++
 migration/multifd.h                 |   4 +-
 chardev/char-io.c                   |   2 +-
 hw/remote/mpqemu-link.c             |   2 +-
 io/channel-buffer.c                 |   1 +
 io/channel-command.c                |   1 +
 io/channel-file.c                   |   1 +
 io/channel-socket.c                 | 110 +++++++++++++++++++++++++++-
 io/channel-tls.c                    |   1 +
 io/channel-websock.c                |   1 +
 io/channel.c                        |  49 ++++++++++---
 migration/channel.c                 |   3 +-
 migration/migration.c               |  52 ++++++++++++-
 migration/multifd.c                 |  72 +++++++++++++++---
 migration/ram.c                     |  29 ++++++--
 migration/rdma.c                    |   1 +
 migration/socket.c                  |  12 ++-
 monitor/hmp-cmds.c                  |   6 ++
 scsi/pr-manager-helper.c            |   2 +-
 tests/unit/test-io-channel-socket.c |   1 +
 23 files changed, 377 insertions(+), 43 deletions(-)