Simimlar to AIO_WAIT_WHILE(), bdrv_drain_poll_top_level() needs to
release the AioContext lock of the node to be drained before calling
aio_poll(). Otherwise, callbacks called by aio_poll() would possibly
take the lock a second time and run into a deadlock with a nested
AIO_WAIT_WHILE() call.

Signed-off-by: Kevin Wolf <kwolf@redhat.com>
---
 block/io.c | 25 ++++++++++++++++++++++++-
 1 file changed, 24 insertions(+), 1 deletion(-)

diff --git a/block/io.c b/block/io.c
index 7100344c7b..832d2536bf 100644
--- a/block/io.c
+++ b/block/io.c
@@ -268,9 +268,32 @@ bool bdrv_drain_poll(BlockDriverState *bs, bool recursive,
 static bool bdrv_drain_poll_top_level(BlockDriverState *bs, bool recursive,
                                       BdrvChild *ignore_parent)
 {
+    AioContext *ctx = bdrv_get_aio_context(bs);
+
+    /*
+     * We cannot easily release the lock unconditionally here because many
+     * callers of drain function (like qemu initialisation, tools, etc.) don't
+     * even hold the main context lock.
+     *
+     * This means that we fix potential deadlocks for the case where we are in
+     * the main context and polling a BDS in a different AioContext, but
+     * draining a BDS in the main context from a different I/O thread would
+     * still have this problem. Fortunately, this isn't supposed to happen
+     * anyway.
+     */
+    if (ctx != qemu_get_aio_context()) {
+        aio_context_release(ctx);
+    } else {
+        assert(qemu_get_current_aio_context() == qemu_get_aio_context());
+    }
+
     /* Execute pending BHs first and check everything else only after the BHs
      * have executed. */
-    while (aio_poll(bs->aio_context, false));
+    while (aio_poll(ctx, false));
+
+    if (ctx != qemu_get_aio_context()) {
+        aio_context_acquire(ctx);
+    }
 
     return bdrv_drain_poll(bs, recursive, ignore_parent, false);
 }
-- 
2.13.6

On Fri, 08/17 19:02, Kevin Wolf wrote:
> Simimlar to AIO_WAIT_WHILE(), bdrv_drain_poll_top_level() needs to
> release the AioContext lock of the node to be drained before calling
> aio_poll(). Otherwise, callbacks called by aio_poll() would possibly
> take the lock a second time and run into a deadlock with a nested
> AIO_WAIT_WHILE() call.
> 
> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
> ---
>  block/io.c | 25 ++++++++++++++++++++++++-
>  1 file changed, 24 insertions(+), 1 deletion(-)
> 
> diff --git a/block/io.c b/block/io.c
> index 7100344c7b..832d2536bf 100644
> --- a/block/io.c
> +++ b/block/io.c
> @@ -268,9 +268,32 @@ bool bdrv_drain_poll(BlockDriverState *bs, bool recursive,
>  static bool bdrv_drain_poll_top_level(BlockDriverState *bs, bool recursive,
>                                        BdrvChild *ignore_parent)
>  {
> +    AioContext *ctx = bdrv_get_aio_context(bs);
> +
> +    /*
> +     * We cannot easily release the lock unconditionally here because many
> +     * callers of drain function (like qemu initialisation, tools, etc.) don't
> +     * even hold the main context lock.
> +     *
> +     * This means that we fix potential deadlocks for the case where we are in
> +     * the main context and polling a BDS in a different AioContext, but
> +     * draining a BDS in the main context from a different I/O thread would
> +     * still have this problem. Fortunately, this isn't supposed to happen
> +     * anyway.
> +     */
> +    if (ctx != qemu_get_aio_context()) {
> +        aio_context_release(ctx);
> +    } else {
> +        assert(qemu_get_current_aio_context() == qemu_get_aio_context());
> +    }
> +
>      /* Execute pending BHs first and check everything else only after the BHs
>       * have executed. */
> -    while (aio_poll(bs->aio_context, false));
> +    while (aio_poll(ctx, false));
> +
> +    if (ctx != qemu_get_aio_context()) {
> +        aio_context_acquire(ctx);
> +    }
>  
>      return bdrv_drain_poll(bs, recursive, ignore_parent, false);
>  }
> -- 
> 2.13.6
> 

The same question as patch 3: why not just use AIO_WAIT_WHILE() here? It takes
care to not release any lock if both running and polling in the main context
(taking the in_aio_context_home_thread() branch).

Fam

Am 24.08.2018 um 09:24 hat Fam Zheng geschrieben:
> On Fri, 08/17 19:02, Kevin Wolf wrote:
> > Simimlar to AIO_WAIT_WHILE(), bdrv_drain_poll_top_level() needs to
> > release the AioContext lock of the node to be drained before calling
> > aio_poll(). Otherwise, callbacks called by aio_poll() would possibly
> > take the lock a second time and run into a deadlock with a nested
> > AIO_WAIT_WHILE() call.
> > 
> > Signed-off-by: Kevin Wolf <kwolf@redhat.com>
> > ---
> >  block/io.c | 25 ++++++++++++++++++++++++-
> >  1 file changed, 24 insertions(+), 1 deletion(-)
> > 
> > diff --git a/block/io.c b/block/io.c
> > index 7100344c7b..832d2536bf 100644
> > --- a/block/io.c
> > +++ b/block/io.c
> > @@ -268,9 +268,32 @@ bool bdrv_drain_poll(BlockDriverState *bs, bool recursive,
> >  static bool bdrv_drain_poll_top_level(BlockDriverState *bs, bool recursive,
> >                                        BdrvChild *ignore_parent)
> >  {
> > +    AioContext *ctx = bdrv_get_aio_context(bs);
> > +
> > +    /*
> > +     * We cannot easily release the lock unconditionally here because many
> > +     * callers of drain function (like qemu initialisation, tools, etc.) don't
> > +     * even hold the main context lock.
> > +     *
> > +     * This means that we fix potential deadlocks for the case where we are in
> > +     * the main context and polling a BDS in a different AioContext, but
> > +     * draining a BDS in the main context from a different I/O thread would
> > +     * still have this problem. Fortunately, this isn't supposed to happen
> > +     * anyway.
> > +     */
> > +    if (ctx != qemu_get_aio_context()) {
> > +        aio_context_release(ctx);
> > +    } else {
> > +        assert(qemu_get_current_aio_context() == qemu_get_aio_context());
> > +    }
> > +
> >      /* Execute pending BHs first and check everything else only after the BHs
> >       * have executed. */
> > -    while (aio_poll(bs->aio_context, false));
> > +    while (aio_poll(ctx, false));
> > +
> > +    if (ctx != qemu_get_aio_context()) {
> > +        aio_context_acquire(ctx);
> > +    }
> >  
> >      return bdrv_drain_poll(bs, recursive, ignore_parent, false);
> >  }
> 
> The same question as patch 3: why not just use AIO_WAIT_WHILE() here? It takes
> care to not release any lock if both running and polling in the main context
> (taking the in_aio_context_home_thread() branch).

I don't think AIO_WAIT_WHILE() can be non-blocking, though?

There is also no real condition here to check. It's just polling as long
as there is activity to get the pending BH callbacks run. I'm not sure
how I could possibly write this as a AIO_WAIT_WHILE() condition.

After all, this one just doesn't feel like the right use case for
AIO_WAIT_WHILE().

Kevin