Re: [Qemu-devel] [Qemu-block] [PATCH for-2.9-rc5 v2] block: Drain BH in bdrv_drained_begin

[Stefan Hajnoczi <stefanha@gmail.com>]

On Fri, Apr 14, 2017 at 9:02 AM, Fam Zheng <famz@redhat.com> wrote:
> During block job completion, nothing is preventing
> block_job_defer_to_main_loop_bh from being called in a nested
> aio_poll(), which is a trouble, such as in this code path:
>
>     qmp_block_commit
>       commit_active_start
>         bdrv_reopen
>           bdrv_reopen_multiple
>             bdrv_reopen_prepare
>               bdrv_flush
>                 aio_poll
>                   aio_bh_poll
>                     aio_bh_call
>                       block_job_defer_to_main_loop_bh
>                         stream_complete
>                           bdrv_reopen
>
> block_job_defer_to_main_loop_bh is the last step of the stream job,
> which should have been "paused" by the bdrv_drained_begin/end in
> bdrv_reopen_multiple, but it is not done because it's in the form of a
> main loop BH.
>
> Similar to why block jobs should be paused between drained_begin and
> drained_end, BHs they schedule must be excluded as well.  To achieve
> this, this patch forces draining the BH in BDRV_POLL_WHILE.
>
> Also because the BH in question can do bdrv_unref and child replacing,
> protect @bs carefully to avoid use-after-free.
>
> As a side effect this fixes a hang in block_job_detach_aio_context
> during system_reset when a block job is ready:
>
>     #0  0x0000555555aa79f3 in bdrv_drain_recurse
>     #1  0x0000555555aa825d in bdrv_drained_begin
>     #2  0x0000555555aa8449 in bdrv_drain
>     #3  0x0000555555a9c356 in blk_drain
>     #4  0x0000555555aa3cfd in mirror_drain
>     #5  0x0000555555a66e11 in block_job_detach_aio_context
>     #6  0x0000555555a62f4d in bdrv_detach_aio_context
>     #7  0x0000555555a63116 in bdrv_set_aio_context
>     #8  0x0000555555a9d326 in blk_set_aio_context
>     #9  0x00005555557e38da in virtio_blk_data_plane_stop
>     #10 0x00005555559f9d5f in virtio_bus_stop_ioeventfd
>     #11 0x00005555559fa49b in virtio_bus_stop_ioeventfd
>     #12 0x00005555559f6a18 in virtio_pci_stop_ioeventfd
>     #13 0x00005555559f6a18 in virtio_pci_reset
>     #14 0x00005555559139a9 in qdev_reset_one
>     #15 0x0000555555916738 in qbus_walk_children
>     #16 0x0000555555913318 in qdev_walk_children
>     #17 0x0000555555916738 in qbus_walk_children
>     #18 0x00005555559168ca in qemu_devices_reset
>     #19 0x000055555581fcbb in pc_machine_reset
>     #20 0x00005555558a4d96 in qemu_system_reset
>     #21 0x000055555577157a in main_loop_should_exit
>     #22 0x000055555577157a in main_loop
>     #23 0x000055555577157a in main
>
> The rationale is that the loop in block_job_detach_aio_context cannot
> make any progress in pausing/completing the job, because bs->in_flight
> is 0, so bdrv_drain doesn't process the block_job_defer_to_main_loop
> BH. With this patch, it does.
>
> Reported-by: Jeff Cody <jcody@redhat.com>
> Signed-off-by: Fam Zheng <famz@redhat.com>
>
> ---
>
> v2: Do the BH poll in BDRV_POLL_WHILE to cover bdrv_drain_all_begin as
> well. [Kevin]
> ---
>  block/io.c            | 10 +++++++---
>  include/block/block.h | 21 +++++++++++++--------
>  2 files changed, 20 insertions(+), 11 deletions(-)

Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>

>
> diff --git a/block/io.c b/block/io.c
> index 8706bfa..a472157 100644
> --- a/block/io.c
> +++ b/block/io.c
> @@ -158,7 +158,7 @@ bool bdrv_requests_pending(BlockDriverState *bs)
>
>  static bool bdrv_drain_recurse(BlockDriverState *bs)
>  {
> -    BdrvChild *child;
> +    BdrvChild *child, *tmp;
>      bool waited;
>
>      waited = BDRV_POLL_WHILE(bs, atomic_read(&bs->in_flight) > 0);
> @@ -167,8 +167,12 @@ static bool bdrv_drain_recurse(BlockDriverState *bs)
>          bs->drv->bdrv_drain(bs);
>      }
>
> -    QLIST_FOREACH(child, &bs->children, next) {
> -        waited |= bdrv_drain_recurse(child->bs);
> +    QLIST_FOREACH_SAFE(child, &bs->children, next, tmp) {
> +        BlockDriverState *bs = child->bs;
> +        assert(bs->refcnt > 0);
> +        bdrv_ref(bs);
> +        waited |= bdrv_drain_recurse(bs);
> +        bdrv_unref(bs);
>      }

There are other loops over bs->children in the block layer code.  If a
they indirectly call aio_poll() then the child bs could disappear.
They should probably also take areference.

That said, I wonder if this really solves the problem of graph changes
during bs->children traversal.  We now have a reference but are
iterating over stale data.

This is a separate issue and doesn't need to hold up this patch.

>
>      return waited;
> diff --git a/include/block/block.h b/include/block/block.h
> index 97d4330..3f05e71 100644
> --- a/include/block/block.h
> +++ b/include/block/block.h
> @@ -381,12 +381,13 @@ void bdrv_drain_all(void);
>
>  #define BDRV_POLL_WHILE(bs, cond) ({                       \
>      bool waited_ = false;                                  \
> +    bool busy_ = true;                                     \
>      BlockDriverState *bs_ = (bs);                          \
>      AioContext *ctx_ = bdrv_get_aio_context(bs_);          \
>      if (aio_context_in_iothread(ctx_)) {                   \
> -        while ((cond)) {                                   \
> -            aio_poll(ctx_, true);                          \
> -            waited_ = true;                                \
> +        while ((cond) || busy_) {                          \
> +            busy_ = aio_poll(ctx_, (cond));                \
> +            waited_ |= !!(cond);                           \
>          }                                                  \
>      } else {                                               \
>          assert(qemu_get_current_aio_context() ==           \
> @@ -398,11 +399,15 @@ void bdrv_drain_all(void);
>           */                                                \
>          assert(!bs_->wakeup);                              \
>          bs_->wakeup = true;                                \
> -        while ((cond)) {                                   \
> -            aio_context_release(ctx_);                     \
> -            aio_poll(qemu_get_aio_context(), true);        \
> -            aio_context_acquire(ctx_);                     \
> -            waited_ = true;                                \
> +        while (busy_) {                                    \
> +            if ((cond)) {                                  \
> +                waited_ = busy_ = true;                    \
> +                aio_context_release(ctx_);                 \
> +                aio_poll(qemu_get_aio_context(), true);    \
> +                aio_context_acquire(ctx_);                 \
> +            } else {                                       \
> +                busy_ = aio_poll(ctx_, false);             \
> +            }                                              \
>          }                                                  \
>          bs_->wakeup = false;                               \
>      }                                                      \
> --
> 2.9.3
>
>