Recently, a few bugs were reported that resulted from an inconsistent
state regarding AioContexts. Block nodes can end up in different
contexts than their users expect - the AioContext of a node can even
change under the feet of a device with no way for the device to forbid
this. We recently added a few basic checks to scsi-disk and virtio-blk,
but they are by far not enough.

This is the first part of my work to actually properly manage
AioContexts in the block layer rather than just doing some ad-hoc calls
to bdrv_set_aio_context() and hoping that everything will work out.

The goal of this first part is that bdrv_set_aio_context() propagates
the AioContext change not only to the children of the node like we
already do, but also to any other affected nodes, such as additional
parents or nodes connected to the requested one only through a block job
that operates on both nodes.

Keep in mind that a second part will follow and that this is visible in
some functions that may not seem that useful in this series. In
particular, bdrv_try_set_aio_context() isn't used much outside of test
cases in this series. This will change in the second part.

Kevin Wolf (10):
  block: Add bdrv_try_set_aio_context()
  block: Make bdrv_attach/detach_aio_context() static
  block: Move recursion to bdrv_set_aio_context()
  block: Propagate AioContext change to parents
  test-block-iothread: Test AioContext propagation through the tree
  block: Implement .(can_)set_aio_ctx for BlockBackend
  block: Add blk_set_allow_aio_context_change()
  blockjob: Propagate AioContext change to all job nodes
  blockjob: Remove AioContext notifiers
  test-block-iothread: Test AioContext propagation for block jobs

 include/block/block.h          |  10 ++
 include/block/block_int.h      |  25 +---
 include/sysemu/block-backend.h |   1 +
 block.c                        | 157 ++++++++++++++++++++++---
 block/backup.c                 |   8 --
 block/block-backend.c          |  55 ++++++++-
 block/mirror.c                 |  10 +-
 blockjob.c                     |  77 ++++++-------
 tests/test-block-iothread.c    | 202 +++++++++++++++++++++++++++++++++
 9 files changed, 449 insertions(+), 96 deletions(-)

-- 
2.20.1

Am 06.05.2019 um 19:17 hat Kevin Wolf geschrieben:
> Recently, a few bugs were reported that resulted from an inconsistent
> state regarding AioContexts. Block nodes can end up in different
> contexts than their users expect - the AioContext of a node can even
> change under the feet of a device with no way for the device to forbid
> this. We recently added a few basic checks to scsi-disk and virtio-blk,
> but they are by far not enough.
> 
> This is the first part of my work to actually properly manage
> AioContexts in the block layer rather than just doing some ad-hoc calls
> to bdrv_set_aio_context() and hoping that everything will work out.
> 
> The goal of this first part is that bdrv_set_aio_context() propagates
> the AioContext change not only to the children of the node like we
> already do, but also to any other affected nodes, such as additional
> parents or nodes connected to the requested one only through a block job
> that operates on both nodes.
> 
> Keep in mind that a second part will follow and that this is visible in
> some functions that may not seem that useful in this series. In
> particular, bdrv_try_set_aio_context() isn't used much outside of test
> cases in this series. This will change in the second part.

Applied to the block branch.

Kevin