On Wed, Nov 06, 2024 at 11:25:07AM +0800, Yu Kuai wrote:
> Hi,
>
> 在 2024/11/06 11:11, Ming Lei 写道:
> > On Wed, Nov 06, 2024 at 10:58:40AM +0800, Yu Kuai wrote:
> > > Hi,Ming and Christoph
> > >
> > > 在 2024/11/05 19:19, Christoph Hellwig 写道:
> > > > On Mon, Nov 04, 2024 at 07:00:05PM +0800, Yu Kuai wrote:
> > > > > From: Yu Kuai <yukuai3@xxxxxxxxxx>
> > > > >
> > > > > blk_mq_clear_flush_rq_mapping() is not called during scsi probe, by
> > > > > checking blk_queue_init_done(). However, QUEUE_FLAG_INIT_DONE is cleared
> > > > > in del_gendisk by commit aec89dc5d421 ("block: keep q_usage_counter in
> > > > > atomic mode after del_gendisk"), hence for disk like scsi, following
> > > > > blk_mq_destroy_queue() will not clear flush rq from tags->rqs[] as well,
> > > > > cause following uaf that is found by our syzkaller for v6.6:
> > > >
> > > > Which means we leave the flush request lingering after del_gendisk,
> > > > which sounds like the real bug. I suspect we just need to move the
> > > > call to blk_mq_clear_flush_rq_mapping so that it is called from
> > > > del_gendisk and doesn't leave the flush tag lingering around.
> > > >
> > >
> > > This remind me that del_gendisk is still too late to do that. Noted that
> > > flush_rq can acquire different tags, so if the multiple flush_rq is done
> > > and those tags are not reused, the flush_rq can exist in multiple
> > > entries in tags->rqs[]. The consequence I can think of is that iterating
> > > tags can found the same flush_rq multiple times, and the flush_rq can be
> > > inflight.
> >
> > How can that be one problem?
> >
> > Please look at
> >
> > commit 364b61818f65 ("blk-mq: clearing flush request reference in tags->rqs[]")
> > commit bd63141d585b ("blk-mq: clear stale request in tags->rq[] before freeing one request pool")
> >
> > and understand the motivation.
> >
> > That also means it is just fine to delay blk_mq_clear_flush_rq_mapping()
> > after disk is deleted.
>
> I do understand what you mean. Let's see if you want this to be avoided,
> for example(no disk is deleted):
It is definitely another issue, and not supposed to be covered by
blk_mq_clear_flush_rq_mapping().
>
> 1) issue a flush, and tag 0 is used, after the flush is done,
> tags->rqs[0] = flush_rq
> 2) issue another flush, and tag 1 is used, after the flush is done,
> tags->rqs[1] = flush_rq
> 3) issue a flush again, and tag 2 is used, and the flush_rq is
> dispatched to disk;
Yes, this kind of thing exists since blk-mq begins, and you can't expect
blk_mq_in_flight_rw() to get accurate inflight requests.
> 4) Then in this case, blk_mq_in_flight_rw() will found the same flush_rq
> 3 times and think there are 3 inflight request, same for
> hctx_busy_show()...
But we have tried to avoid it, such as in blk_mq_find_and_get_req()
req->tag is checked against the current 'bitnr' of sbitmap when walking
over tags. Then the same flush_rq won't be counted 3 times.
Thanks,
Ming
