Re: [Qemu-block] [PATCH 3/3] migration: avoid recursive AioContext locking in save_vmstate()

[Kevin Wolf]

Am 17.05.2017 um 19:09 hat Stefan Hajnoczi geschrieben:
> AioContext was designed to allow nested acquire/release calls.  It uses
> a recursive mutex so callers don't need to worry about nesting...or so
> we thought.
> 
> BDRV_POLL_WHILE() is used to wait for block I/O requests.  It releases
> the AioContext temporarily around aio_poll().  This gives IOThreads a
> chance to acquire the AioContext to process I/O completions.
> 
> It turns out that recursive locking and BDRV_POLL_WHILE() don't mix.
> BDRV_POLL_WHILE() only releases the AioContext once, so the IOThread
> will not be able to acquire the AioContext if it was acquired
> multiple times.
> 
> Instead of trying to release AioContext n times in BDRV_POLL_WHILE(),
> this patch simply avoids nested locking in save_vmstate().  It's the
> simplest fix and we should step back to consider the big picture with
> all the recent changes to block layer threading.
> 
> This patch is the final fix to solve 'savevm' hanging with -object
> iothread.
> 
> Signed-off-by: Stefan Hajnoczi <address@hidden>
> ---
>  migration/savevm.c | 12 +++++++++++-
>  1 file changed, 11 insertions(+), 1 deletion(-)
> 
> diff --git a/migration/savevm.c b/migration/savevm.c
> index 7f66d58..a70ba20 100644
> --- a/migration/savevm.c
> +++ b/migration/savevm.c
> @@ -2153,6 +2153,14 @@ int save_vmstate(const char *name)
>          goto the_end;
>      }
>  
> +    /* The bdrv_all_create_snapshot() call that follows acquires the 
> AioContext
> +     * for itself.  BDRV_POLL_WHILE() does not support nested locking because
> +     * it only releases the lock once.  Therefore synchronous I/O will 
> deadlock
> +     * unless we release the AioContext before bdrv_all_create_snapshot().
> +     */
> +    aio_context_release(aio_context);
> +    aio_context = NULL;
> +
>      ret = bdrv_all_create_snapshot(sn, bs, vm_state_size, &bs);
>      if (ret < 0) {
>          error_report("Error while creating snapshot on '%s'",
> @@ -2163,7 +2171,9 @@ int save_vmstate(const char *name)
>      ret = 0;
>  
>   the_end:
> -    aio_context_release(aio_context);
> +    if (aio_context) {
> +        aio_context_release(aio_context);
> +    }
>      if (saved_vm_running) {
>          vm_start();
>      }

It might actually even be true before this patch because the lock is
already only taken for some parts of the function, but don't we need to
call bdrv_drain_all_begin/end() around the whole function now?

We're stopping the VM, so hopefully no device is continuing to process
requests, but can't we still have block jobs, NBD server requests etc.?

And the same is probably true for qemu_loadvm_state().

Kevin