[v1] Fix deadlock with bdrv_open of self-served NBD

[PATCH 0/8] Fix deadlock with bdrv_open of self-served NBD

Eric Blake posted 8 patches 1 week, 5 days ago

Diff against v2
Download series mbox

Failed in applying to current master (apply log)

Maintainers: "Daniel P. Berrangé" <berrange@redhat.com>, Eric Blake <eblake@redhat.com>, Vladimir Sementsov-Ogievskiy <vsementsov@yandex-team.ru>, Kevin Wolf <kwolf@redhat.com>, Hanna Reitz <hreitz@redhat.com>

There is a newer version of this series

include/io/channel-socket.h                   |   1 +
include/io/net-listener.h                     |   1 +
io/channel-socket.c                           |   1 +
io/net-listener.c                             | 136 +++++++++++-------
io/trace-events                               |   5 +
tests/qemu-iotests/tests/nbd-in-qcow2-chain   |  84 +++++++++++
.../qemu-iotests/tests/nbd-in-qcow2-chain.out |  56 ++++++++
tests/qemu-iotests/tests/vvfat.out            |   0
8 files changed, 236 insertions(+), 48 deletions(-)
create mode 100755 tests/qemu-iotests/tests/nbd-in-qcow2-chain
create mode 100644 tests/qemu-iotests/tests/nbd-in-qcow2-chain.out
mode change 100755 => 100644 tests/qemu-iotests/tests/vvfat.out

Expand all Fold all

[PATCH 0/8] Fix deadlock with bdrv_open of self-served NBD

Posted by Eric Blake 1 week, 5 days ago

https://gitlab.com/qemu-project/qemu/-/issues/3169 was an excellent
bug report of a deadlock scenario when qemu opens both the server and
client end of an NBD intermediary in a qcow2 backing chain.  It took
me quite a bit of time to understand why the deadlock was even
happening, and then some mental gymnastics on how best to break the
deadlock.  Ideally - we'd get rid of all nested event loops, and have
the main loop do EVERYTHING event-driven.  But that's a much bigger
code change to the entire code base of QEMU, so I went with the next
best thing of keeping the AioContext nested loop but ensuring that
accepting NBD clients is now properly part of the AioContext rather
than gated by only the main loop progress.

I have not researched how long the deadlock has been present, to
determine if it has ever worked before earlier refactorings when we
started using AioContext more heavily, or if it has always been broken
until now to try and connect QEMU as a client to a self-served NBD
server.  But either way, I think this series should be part of the
10.2 release; I'm awfully close to soft freeze, but I think this
counts as a bug fix worth having even if it doesn't make it in before
-rc1.

Eric Blake (8):
  qio: Add trace points to net_listener
  qio: Minor optimization when callback function is unchanged
  qio: Remember context of qio_net_listener_set_client_func_full
  qio: Factor out helpers qio_net_listener_[un]watch
  qio: Let listening sockets remember their owning QIONetListener
  qio: Hoist ref of listener outside loop
  qio: Use AioContext for default-context QIONetListener
  iotests: Add coverage of recent NBD qio deadlock fix

 include/io/channel-socket.h                   |   1 +
 include/io/net-listener.h                     |   1 +
 io/channel-socket.c                           |   1 +
 io/net-listener.c                             | 136 +++++++++++-------
 io/trace-events                               |   5 +
 tests/qemu-iotests/tests/nbd-in-qcow2-chain   |  84 +++++++++++
 .../qemu-iotests/tests/nbd-in-qcow2-chain.out |  56 ++++++++
 tests/qemu-iotests/tests/vvfat.out            |   0
 8 files changed, 236 insertions(+), 48 deletions(-)
 create mode 100755 tests/qemu-iotests/tests/nbd-in-qcow2-chain
 create mode 100644 tests/qemu-iotests/tests/nbd-in-qcow2-chain.out
 mode change 100755 => 100644 tests/qemu-iotests/tests/vvfat.out

-- 
2.51.1