Re: [PATCH V2] ublk: fix deadlock when reading partition table

[Ming Lei <ming.lei@redhat.com>]

On Fri, Dec 12, 2025 at 12:49:49PM -0700, Jens Axboe wrote:
> On 12/12/25 7:34 AM, Ming Lei wrote:
> > diff --git a/drivers/block/ublk_drv.c b/drivers/block/ublk_drv.c
> > index df9831783a13..38f138f248e6 100644
> > --- a/drivers/block/ublk_drv.c
> > +++ b/drivers/block/ublk_drv.c
> > @@ -1080,12 +1080,20 @@ static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
> >  	return io_uring_cmd_to_pdu(ioucmd, struct ublk_uring_cmd_pdu);
> >  }
> >  
> > +static void ublk_end_request(struct request *req, blk_status_t error)
> > +{
> > +	local_bh_disable();
> > +	blk_mq_end_request(req, error);
> > +	local_bh_enable();
> > +}
> 
> This is really almost too ugly to live, as a work-around for what just
> happens to be in __fput_deferred()... Surely we can come up with
> something better here? Heck even a PF_ flag would be better than this,
> imho.

task flag will switch to release all files opened by current from wq context,
and there may be chance to cause regression, especially this fix needs to
backport to stable.

So I'd suggest to take it for safe stable purpose.


> 
> > @@ -1117,14 +1125,26 @@ static inline void __ublk_complete_rq(struct request *req, struct ublk_io *io,
> >  	if (unlikely(unmapped_bytes < io->res))
> >  		io->res = unmapped_bytes;
> >  
> > -	if (blk_update_request(req, BLK_STS_OK, io->res))
> > +	/*
> > +	 * Run bio->bi_end_io() from softirq context for preventing this
> > +	 * ublk's blkdev_release() from being called on current's task
> > +	 * work, see fput() implementation.
> 
> But that's not what it does, running it from softirq context. It simply
> disables local bottomhalf interrupts to trick __fput_deferred(), it's
> not scheduling ->bi_end_io() to run from softirq context itself.

Yes, I can twink the words if you don't object this approach, but it isn't
different from core code viewpoint, in_interrupt() returns true.

Thanks,
Ming