| Message ID | 1457431876-8475-1-git-send-email-stefanha@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 08/03/2016 11:11, Stefan Hajnoczi wrote: > This reverts commit df9a681dc9ad41c9cdeb9ecc5d060ba9abd27e01. > > Note that commit df9a681dc9ad41c9cdeb9ecc5d060ba9abd27e01 included some > unrelated hunks, possibly due to a merge failure or an overlooked > squash. This only reverts the qed .bdrv_drain() implementation. > > The qed .bdrv_drain() implementation is unsafe and can lead to a double > request completion. > > Paolo Bonzini reports: > "The problem is that bdrv_qed_drain calls qed_plug_allocating_write_reqs > unconditionally, but this is not correct if an allocating write is > queued. In this case, qed_unplug_allocating_write_reqs will restart the > allocating write and possibly cause it to complete. The aiocb however > is still in use for the L2/L1 table writes, and will then be completed > again as soon as the table writes are stable." > > For QEMU 2.6 we can simply revert this commit. A full solution for the > qed need check timer may be added if the bdrv_drain() implementation is > extended. > > Reported-by: Paolo Bonzini <pbonzini@redhat.com> > Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> Acked-by: Paolo Bonzini <pbonzini@redhat.com> Paolo > --- > block/qed.c | 13 ------------- > 1 file changed, 13 deletions(-) > > Passes ./check -qed. > > diff --git a/block/qed.c b/block/qed.c > index 404be1e..3da8021 100644 > --- a/block/qed.c > +++ b/block/qed.c > @@ -376,18 +376,6 @@ static void bdrv_qed_attach_aio_context(BlockDriverState *bs, > } > } > > -static void bdrv_qed_drain(BlockDriverState *bs) > -{ > - BDRVQEDState *s = bs->opaque; > - > - /* Cancel timer and start doing I/O that were meant to happen as if it > - * fired, that way we get bdrv_drain() taking care of the ongoing requests > - * correctly. */ > - qed_cancel_need_check_timer(s); > - qed_plug_allocating_write_reqs(s); > - bdrv_aio_flush(s->bs, qed_clear_need_check, s); > -} > - > static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags, > Error **errp) > { > @@ -1692,7 +1680,6 @@ static BlockDriver bdrv_qed = { > .bdrv_check = bdrv_qed_check, > .bdrv_detach_aio_context = bdrv_qed_detach_aio_context, > .bdrv_attach_aio_context = bdrv_qed_attach_aio_context, > - .bdrv_drain = bdrv_qed_drain, > }; > > static void bdrv_qed_init(void) >
On Tue, Mar 08, 2016 at 10:11:16AM +0000, Stefan Hajnoczi wrote: > This reverts commit df9a681dc9ad41c9cdeb9ecc5d060ba9abd27e01. > > Note that commit df9a681dc9ad41c9cdeb9ecc5d060ba9abd27e01 included some > unrelated hunks, possibly due to a merge failure or an overlooked > squash. This only reverts the qed .bdrv_drain() implementation. > > The qed .bdrv_drain() implementation is unsafe and can lead to a double > request completion. > > Paolo Bonzini reports: > "The problem is that bdrv_qed_drain calls qed_plug_allocating_write_reqs > unconditionally, but this is not correct if an allocating write is > queued. In this case, qed_unplug_allocating_write_reqs will restart the > allocating write and possibly cause it to complete. The aiocb however > is still in use for the L2/L1 table writes, and will then be completed > again as soon as the table writes are stable." > > For QEMU 2.6 we can simply revert this commit. A full solution for the > qed need check timer may be added if the bdrv_drain() implementation is > extended. > > Reported-by: Paolo Bonzini <pbonzini@redhat.com> > Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> > --- > block/qed.c | 13 ------------- > 1 file changed, 13 deletions(-) > > Passes ./check -qed. Thanks, applied to my block tree: https://github.com/stefanha/qemu/commits/block Stefan
diff --git a/block/qed.c b/block/qed.c index 404be1e..3da8021 100644 --- a/block/qed.c +++ b/block/qed.c @@ -376,18 +376,6 @@ static void bdrv_qed_attach_aio_context(BlockDriverState *bs, } } -static void bdrv_qed_drain(BlockDriverState *bs) -{ - BDRVQEDState *s = bs->opaque; - - /* Cancel timer and start doing I/O that were meant to happen as if it - * fired, that way we get bdrv_drain() taking care of the ongoing requests - * correctly. */ - qed_cancel_need_check_timer(s); - qed_plug_allocating_write_reqs(s); - bdrv_aio_flush(s->bs, qed_clear_need_check, s); -} - static int bdrv_qed_open(BlockDriverState *bs, QDict *options, int flags, Error **errp) { @@ -1692,7 +1680,6 @@ static BlockDriver bdrv_qed = { .bdrv_check = bdrv_qed_check, .bdrv_detach_aio_context = bdrv_qed_detach_aio_context, .bdrv_attach_aio_context = bdrv_qed_attach_aio_context, - .bdrv_drain = bdrv_qed_drain, }; static void bdrv_qed_init(void)
This reverts commit df9a681dc9ad41c9cdeb9ecc5d060ba9abd27e01. Note that commit df9a681dc9ad41c9cdeb9ecc5d060ba9abd27e01 included some unrelated hunks, possibly due to a merge failure or an overlooked squash. This only reverts the qed .bdrv_drain() implementation. The qed .bdrv_drain() implementation is unsafe and can lead to a double request completion. Paolo Bonzini reports: "The problem is that bdrv_qed_drain calls qed_plug_allocating_write_reqs unconditionally, but this is not correct if an allocating write is queued. In this case, qed_unplug_allocating_write_reqs will restart the allocating write and possibly cause it to complete. The aiocb however is still in use for the L2/L1 table writes, and will then be completed again as soon as the table writes are stable." For QEMU 2.6 we can simply revert this commit. A full solution for the qed need check timer may be added if the bdrv_drain() implementation is extended. Reported-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> --- block/qed.c | 13 ------------- 1 file changed, 13 deletions(-) Passes ./check -qed.