On Wed, 03/09 17:17, Paolo Bonzini wrote:
>
>
> On 09/03/2016 15:29, Christian Borntraeger wrote:
> > FWIW, it seems that this patch triggers this error, the
> > "tracked_request_begin"
> > that I reported yesterday and / or some early read issues from the
> > bootloader
> > in a random fashion.
> > Using 2906cddfecff21af20eedab43288b485a679f9ac^ seems to work all the time,
> > moving around vblk->dataplane_started = true also triggers all 3 types
> > of bugs
>
> In all likelihood, the bug is that virtio_blk_handle_output is being
> called in two threads.
>
> It's not clear to me how that's possible, though.
The aio_poll() inside "blk_set_aio_context(s->conf->conf.blk, s->ctx)" looks
suspicious:
main thread iothread
----------------------------------------------------------------------------
virtio_blk_handle_output()
virtio_blk_data_plane_start()
vblk->dataplane_started = true;
blk_set_aio_context()
bdrv_set_aio_context()
bdrv_drain()
aio_poll()
<snip...>
virtio_blk_handle_output()
/* s->dataplane_started is true */
!!! -> virtio_blk_handle_request()
event_notifier_set(ioeventfd)
aio_poll()
virtio_blk_handle_request()
Christian, could you try the followed patch? The aio_poll above is replaced
with a "limited aio_poll" that doesn't disptach ioeventfd.
(Note: perhaps moving "vblk->dataplane_started = true;" after
blk_set_aio_context() also *works around* this.)
---
diff --git a/block.c b/block.c
index ba24b8e..e37e8f7 100644
--- a/block.c
+++ b/block.c
@@ -4093,7 +4093,9 @@ void bdrv_attach_aio_context(BlockDriverState *bs,
void bdrv_set_aio_context(BlockDriverState *bs, AioContext *new_context)
{
- bdrv_drain(bs); /* ensure there are no in-flight requests */
+ /* ensure there are no in-flight requests */
+ bdrv_drained_begin(bs);
+ bdrv_drained_end(bs);
bdrv_detach_aio_context(bs);
