Re: [RFC PATCH] block: Make request_queue lockdep splats show up earlier

["Thomas Hellström" <thomas.hellstrom@linux.intel.com>]

On Mon, 2025-03-17 at 11:28 -0700, Bart Van Assche wrote:
> On 3/17/25 11:13 AM, Thomas Hellström wrote:
> > On Mon, 2025-03-17 at 10:37 -0700, Bart Van Assche wrote:
> > > On 3/17/25 10:11 AM, Thomas Hellström wrote:
> > > > diff --git a/block/blk-core.c b/block/blk-core.c
> > > > index d6c4fa3943b5..4aa439309406 100644
> > > > --- a/block/blk-core.c
> > > > +++ b/block/blk-core.c
> > > > @@ -455,6 +455,12 @@ struct request_queue
> > > > *blk_alloc_queue(struct
> > > > queue_limits *lim, int node_id)
> > > >    	lockdep_init_map(&q->q_lockdep_map, "&q-
> > > > > q_usage_counter(queue)",
> > > >    			 &q->q_lock_cls_key, 0);
> > > >    
> > > > +	/* Prime io_lockdep_map as reclaim tainted */
> > > > +	fs_reclaim_acquire(GFP_KERNEL);
> > > > +	rwsem_acquire_read(&q->io_lockdep_map, 0, 0,
> > > > _RET_IP_);
> > > > +	rwsem_release(&q->io_lockdep_map, _RET_IP_);
> > > > +	fs_reclaim_release(GFP_KERNEL);
> > > > +
> > > >    	q->nr_requests = BLKDEV_DEFAULT_RQ;
> > > >    
> > > >    	return q;
> > > 
> > > Hmm ... my understanding is that it is fine if FS code calls
> > > block
> > > layer
> > > code but also that block layer code never should call FS code.
> > 
> > That added code only mimics the locking sequence that happens
> > during
> > reclaim with the existing code to register the locking order
> > expected
> > by the reclaim code. If anything violates that, lockdep splat [2]
> > will
> > appear.
> > 
> > So I'm not quite following your comment?
> Shouldn't the above code be added in the VFS code rather than in the
> block layer?

It registers a known locking order WRT reclaim(GFP_KERNEL) for the q-
>io_lockdep_map, which is itself initialized in this function. I
believe any known locking orders should be registered at the place the
lockdep map is initialized.

Thanks,
Thomas


> 
> Thanks,
> 
> Bart.