Re: [PATCH] blktrace: Avoid sparse warnings when assigning q->blk_trace

[Luis Chamberlain <mcgrof@kernel.org>]

On Fri, May 29, 2020 at 11:04:48AM +0200, Jan Kara wrote:
> On Fri 29-05-20 08:00:56, Luis Chamberlain wrote:
> > On Thu, May 28, 2020 at 08:55:39PM +0200, Jan Kara wrote:
> > > On Thu 28-05-20 18:43:33, Luis Chamberlain wrote:
> > > > On Thu, May 28, 2020 at 08:31:52PM +0200, Jan Kara wrote:
> > > > > On Thu 28-05-20 07:44:38, Bart Van Assche wrote:
> > > > > > (+Luis)
> > > > > > 
> > > > > > On 2020-05-28 02:29, Jan Kara wrote:
> > > > > > > Mostly for historical reasons, q->blk_trace is assigned through xchg()
> > > > > > > and cmpxchg() atomic operations. Although this is correct, sparse
> > > > > > > complains about this because it violates rcu annotations. Furthermore
> > > > > > > there's no real need for atomic operations anymore since all changes to
> > > > > > > q->blk_trace happen under q->blk_trace_mutex. So let's just replace
> > > > > > > xchg() with rcu_replace_pointer() and cmpxchg() with explicit check and
> > > > > > > rcu_assign_pointer(). This makes the code more efficient and sparse
> > > > > > > happy.
> > > > > > > 
> > > > > > > Reported-by: kbuild test robot <lkp@intel.com>
> > > > > > > Signed-off-by: Jan Kara <jack@suse.cz>
> > > > > > 
> > > > > > How about adding a reference to commit c780e86dd48e ("blktrace: Protect
> > > > > > q->blk_trace with RCU") in the description of this patch?
> > > > > 
> > > > > Yes, that's probably a good idea.
> > > > > 
> > > > > > > @@ -1669,10 +1672,7 @@ static int blk_trace_setup_queue(struct request_queue *q,
> > > > > > >  
> > > > > > >  	blk_trace_setup_lba(bt, bdev);
> > > > > > >  
> > > > > > > -	ret = -EBUSY;
> > > > > > > -	if (cmpxchg(&q->blk_trace, NULL, bt))
> > > > > > > -		goto free_bt;
> > > > > > > -
> > > > > > > +	rcu_assign_pointer(q->blk_trace, bt);
> > > > > > >  	get_probe_ref();
> > > > > > >  	return 0;
> > > > > > 
> > > > > > This changes a conditional assignment of q->blk_trace into an
> > > > > > unconditional assignment. Shouldn't q->blk_trace only be assigned if
> > > > > > q->blk_trace == NULL?
> > > > > 
> > > > > Yes but both callers of blk_trace_setup_queue() actually check that
> > > > > q->blk_trace is NULL before calling blk_trace_setup_queue() and since we
> > > > > hold blk_trace_mutex all the time, the value of q->blk_trace cannot change.
> > > > > So the conditional assignment was just bogus.
> > > > 
> > > > If you run a blktrace against a different partition the check does have
> > > > an effect today. This is because the request_queue is shared between
> > > > partitions implicitly, even though they end up using a different struct
> > > > dentry. So the check is actually still needed, however my change adds
> > > > this check early as well so we don't do a memory allocation just to
> > > > throw it away.
> > > 
> > > I'm not sure we are speaking about the same check but I might be missing
> > > something. blk_trace_setup_queue() is only called from
> > > sysfs_blk_trace_attr_store(). That does:
> > > 
> > >         mutex_lock(&q->blk_trace_mutex);
> > > 
> > >         bt = rcu_dereference_protected(q->blk_trace,
> > >                                        lockdep_is_held(&q->blk_trace_mutex));
> > >         if (attr == &dev_attr_enable) {
> > >                 if (!!value == !!bt) {
> > >                         ret = 0;
> > >                         goto out_unlock_bdev;
> > >                 }
> > > 
> > > 		^^^ So if 'bt' is non-NULL, and we are enabling, we bail
> > > instead of calling blk_trace_setup_queue().
> > > 
> > > Similarly later:
> > > 
> > >         if (bt == NULL) {
> > >                 ret = blk_trace_setup_queue(q, bdev);
> > > 	...
> > > so we again call blk_trace_setup_queue() only if bt is NULL. So IMO the
> > > cmpxchg() in blk_trace_setup_queue() could never fail to set the value.
> > > Am I missing something?
> > 
> > I believe we are talking about the same check indeed. Consider the
> > situation not as a race, but instead consider the state machine of
> > the ioctl. The BLKTRACESETUP goes first, and when that is over we
> > have not ran BLKTRACESTART. So, prior to BLKTRACESTART we can have
> > another BLKTRACESETUP run but against another partition.
> 
> So first note that BLKTRACESETUP goes through do_blk_trace_setup() while
> 'echo 1 >/sys/block/../trace/enable' goes through blk_trace_setup_queue().
> Although these operations achieve a very similar things, they are completely
> separate code paths. I was speaking about the second case while you are now
> speaking about the first one.
> 
> WRT to your BLKTRACESETUP example, the first BLKTRACESETUP will end up
> setting q->blk_trace to 'bt' so the second BLKTRACESETUP will see
> q->blk_trace is not NULL (my patch adds this check to do_blk_trace_setup()
> so we bail out earlier than during cmpxchg()) and fails. Again I don't see
> any problem here...

Ah, the patch I was CC'd on didn't contain this hunk! It only had the
change from cmpxchg() to the rcu_assign_pointer(), so I misunderstood
your intention, sorry!

In that case, I already proposed a patch to do that, and it also adds
a tiny bit of verbiage given we currently don't inform the user about
why this fails [0].

Let me know how you folks would like to proceed.

[0] https://lkml.kernel.org/r/20200516031956.2605-7-mcgrof@kernel.org

  Luis