[v1] block: Protect AIO context change with perm API

[Qemu-devel] [PATCH RFC 00/16] block: Protect AIO context change with perm API

Posted by Fam Zheng 7 years, 1 month ago

Eject / change of scsi-cd on a virtio-scsi dataplane bus causes abort() because
the new BDS doesn't get proper bdrv_set_aio_context().

Store the AioContext in BB and do it in blk_insert_bs. That is done by
Vladimir's patch.

Other patches are to make sure such a bdrv_set_aio_context() doesn't interfere
with other BBs using other nodes from this graph.

RFC note:

Unfortunately, a use-after-free crash in iotests 030 appears since patch 7,
which I believe is a latent bug that bdrv_reopen is "reentered" in existing
code, rather than from this series:

> #4  0x0000561ab90425a7 in bdrv_reopen
> #5  0x0000561ab8e1d28e in stream_complete
> #6  0x0000561ab9048543 in block_job_defer_to_main_loop_bh
> #7  0x0000561ab91305bc in aio_bh_call
> #8  0x0000561ab9130659 in aio_bh_poll
> #9  0x0000561ab9135656 in aio_poll
> #10 0x0000561ab90a6cf5 in bdrv_flush
> #11 0x0000561ab904285a in bdrv_reopen_prepare
> #12 0x0000561ab90423f0 in bdrv_reopen_multiple
> #13 0x0000561ab90425ef in bdrv_reopen
> #14 0x0000561ab909fa49 in commit_active_start
> #15 0x0000561ab8dd6ffb in qmp_block_commit
> #16 0x0000561ab8ded485 in qmp_marshal_block_commit
> #17 0x0000561ab9123e6c in do_qmp_dispatch
> #18 0x0000561ab9123fa4 in qmp_dispatch
> #19 0x0000561ab8ca26b7 in handle_qmp_command

I have a fix that I'll post separately.

The last patches are an alternative to patch 7, to "workaround" this in a
really non-obvious way.

Fam

Fam Zheng (15):
  block: Define BLK_PERM_AIO_CONTEXT_CHANGE
  block-backend: Add blk_request_perm
  blockjob: Add BLK_PERM_AIO_CONTEXT_CHANGE shared perm on bs
  block: Propagate BLK_PERM_AIO_CONTEXT_CHANGE down the graph
  backup: Request BLK_PERM_AIO_CONTEXT_CHANGE on target
  backup: Do initial aio context move of target via BB interface
  mirror: Request aio context change permission on target
  commit: Allow aio context change on s->base
  mirror: Do initial aio context move of target via BB interface
  virtio-scsi: Request BLK_PERM_AIO_CONTEXT_CHANGE for dataplane
  virtio-blk: Request BLK_PERM_AIO_CONTEXT_CHANGE for dataplane
  nbd: Allow BLK_PERM_AIO_CONTEXT_CHANGE on BB
  block: Add perm assertion on blk_set_aio_context
  mirror: Lazily request aio context change permission on target
  Revert "mirror: Request aio context change permission on target"

Vladimir Sementsov-Ogievskiy (1):
  blk: fix aio context loss on media change

 block.c                         | 20 ++++++++++++--------
 block/backup.c                  | 11 ++++++++++-
 block/block-backend.c           | 23 +++++++++++++++++++++++
 block/commit.c                  |  3 ++-
 block/mirror.c                  | 10 ++++++++++
 block/vvfat.c                   |  2 +-
 blockdev.c                      | 18 ------------------
 blockjob.c                      |  3 +++
 hw/block/dataplane/virtio-blk.c | 15 +++++++++++----
 hw/scsi/virtio-scsi.c           |  4 ++++
 include/block/block.h           |  7 ++++++-
 include/sysemu/block-backend.h  |  1 +
 nbd/server.c                    |  6 ++++--
 13 files changed, 87 insertions(+), 36 deletions(-)

-- 
2.9.3

Re: [Qemu-devel] [Qemu-block] [PATCH RFC 00/16] block: Protect AIO context change with perm API

Posted by Stefan Hajnoczi 7 years ago

On Tue, Mar 21, 2017 at 11:16:19AM +0800, Fam Zheng wrote:
> Eject / change of scsi-cd on a virtio-scsi dataplane bus causes abort() because
> the new BDS doesn't get proper bdrv_set_aio_context().
> 
> Store the AioContext in BB and do it in blk_insert_bs. That is done by
> Vladimir's patch.
> 
> Other patches are to make sure such a bdrv_set_aio_context() doesn't interfere
> with other BBs using other nodes from this graph.
> 
> RFC note:
> 
> Unfortunately, a use-after-free crash in iotests 030 appears since patch 7,
> which I believe is a latent bug that bdrv_reopen is "reentered" in existing
> code, rather than from this series:
> 
> > #4  0x0000561ab90425a7 in bdrv_reopen
> > #5  0x0000561ab8e1d28e in stream_complete
> > #6  0x0000561ab9048543 in block_job_defer_to_main_loop_bh
> > #7  0x0000561ab91305bc in aio_bh_call
> > #8  0x0000561ab9130659 in aio_bh_poll
> > #9  0x0000561ab9135656 in aio_poll
> > #10 0x0000561ab90a6cf5 in bdrv_flush
> > #11 0x0000561ab904285a in bdrv_reopen_prepare
> > #12 0x0000561ab90423f0 in bdrv_reopen_multiple
> > #13 0x0000561ab90425ef in bdrv_reopen
> > #14 0x0000561ab909fa49 in commit_active_start
> > #15 0x0000561ab8dd6ffb in qmp_block_commit
> > #16 0x0000561ab8ded485 in qmp_marshal_block_commit
> > #17 0x0000561ab9123e6c in do_qmp_dispatch
> > #18 0x0000561ab9123fa4 in qmp_dispatch
> > #19 0x0000561ab8ca26b7 in handle_qmp_command
> 
> I have a fix that I'll post separately.
> 
> The last patches are an alternative to patch 7, to "workaround" this in a
> really non-obvious way.

Are there qemu-iotests that cover both success and failure scenarios for
the new permission?

Stefan

Re: [Qemu-devel] [Qemu-block] [PATCH RFC 00/16] block: Protect AIO context change with perm API

Posted by Fam Zheng 7 years ago

On Mon, 04/10 10:04, Stefan Hajnoczi wrote:
> On Tue, Mar 21, 2017 at 11:16:19AM +0800, Fam Zheng wrote:
> > Eject / change of scsi-cd on a virtio-scsi dataplane bus causes abort() because
> > the new BDS doesn't get proper bdrv_set_aio_context().
> > 
> > Store the AioContext in BB and do it in blk_insert_bs. That is done by
> > Vladimir's patch.
> > 
> > Other patches are to make sure such a bdrv_set_aio_context() doesn't interfere
> > with other BBs using other nodes from this graph.
> > 
> > RFC note:
> > 
> > Unfortunately, a use-after-free crash in iotests 030 appears since patch 7,
> > which I believe is a latent bug that bdrv_reopen is "reentered" in existing
> > code, rather than from this series:
> > 
> > > #4  0x0000561ab90425a7 in bdrv_reopen
> > > #5  0x0000561ab8e1d28e in stream_complete
> > > #6  0x0000561ab9048543 in block_job_defer_to_main_loop_bh
> > > #7  0x0000561ab91305bc in aio_bh_call
> > > #8  0x0000561ab9130659 in aio_bh_poll
> > > #9  0x0000561ab9135656 in aio_poll
> > > #10 0x0000561ab90a6cf5 in bdrv_flush
> > > #11 0x0000561ab904285a in bdrv_reopen_prepare
> > > #12 0x0000561ab90423f0 in bdrv_reopen_multiple
> > > #13 0x0000561ab90425ef in bdrv_reopen
> > > #14 0x0000561ab909fa49 in commit_active_start
> > > #15 0x0000561ab8dd6ffb in qmp_block_commit
> > > #16 0x0000561ab8ded485 in qmp_marshal_block_commit
> > > #17 0x0000561ab9123e6c in do_qmp_dispatch
> > > #18 0x0000561ab9123fa4 in qmp_dispatch
> > > #19 0x0000561ab8ca26b7 in handle_qmp_command
> > 
> > I have a fix that I'll post separately.
> > 
> > The last patches are an alternative to patch 7, to "workaround" this in a
> > really non-obvious way.
> 
> Are there qemu-iotests that cover both success and failure scenarios for
> the new permission?

No, I'll add them.

Fam