Re: [Qemu-devel] [PATCH] block: Cleanup BMDS in bdrv_close_all

[Juan Quintela <quintela@redhat.com>]

Kevin Wolf <kwolf@redhat.com> wrote:
> Am 05.09.2017 um 10:54 hat Fam Zheng geschrieben:
>> On Tue, 09/05 10:44, Kevin Wolf wrote:
>> > Am 30.08.2017 um 12:06 hat Fam Zheng geschrieben:
>> > > This fixes the assertion due to op blockers added by BMDS:
>> > > 
>> > > block.c:3248: bdrv_delete: Assertion `bdrv_op_blocker_is_empty(bs)' failed.
>> > > 
>> > > Reproducer: simply start block migration and quit QEMU before it ends.
>> > > 
>> > > Cc: qemu-stable@nongnu.org
>> > > Signed-off-by: Fam Zheng <famz@redhat.com>
>> > > ---
>> > >  block.c                 | 2 ++
>> > >  migration/block.c       | 2 +-
>> > >  migration/block.h       | 1 +
>> > >  stubs/Makefile.objs     | 1 +
>> > >  stubs/block-migration.c | 6 ++++++
>> > >  5 files changed, 11 insertions(+), 1 deletion(-)
>> > >  create mode 100644 stubs/block-migration.c
>> > > 
>> > > diff --git a/block.c b/block.c
>> > > index 3308814bba..508a57274d 100644
>> > > --- a/block.c
>> > > +++ b/block.c
>> > > @@ -43,6 +43,7 @@
>> > >  #include "qemu/cutils.h"
>> > >  #include "qemu/id.h"
>> > >  #include "qapi/util.h"
>> > > +#include "migration/block.h"
>> > >  
>> > >  #ifdef CONFIG_BSD
>> > >  #include <sys/ioctl.h>
>> > > @@ -3111,6 +3112,7 @@ static void bdrv_close(BlockDriverState *bs)
>> > >  
>> > >  void bdrv_close_all(void)
>> > >  {
>> > > +    block_migration_cleanup_bmds();
>> > >      block_job_cancel_sync_all();
>> > >      nbd_export_close_all();
>> > 
>> > This is before bdrv_drain_all(). Can't we still have a block migration
>> > request in flight, whose callback will then dereference a stale pointer?
>> 
>> You're right, bdrv_drain_all should be called first.
>
> Actually, looking a bit closer, what prevents the migration thread from
> starting new requests even after bdrv_drain_all()? Maybe what we really
> need to do is cancelling the migration before calling bdrv_close_all().

I was wondering *where* to put this call inside the migration cleanup
code, but I got to the conclusion that I was not sure that the migration
cancellation code got called when you just do a "quit".

Later, Juan.