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.
Thanks,
Ming
