Re: [syzbot] KASAN: use-after-free Read in srcu_invoke_callbacks

[Ming Lei]

On Fri, Jan 14, 2022 at 08:11:38AM -0800, Paul E. McKenney wrote:
> On Fri, Jan 14, 2022 at 11:58:24PM +0800, Ming Lei wrote:
> > On Fri, Jan 14, 2022 at 07:27:52AM -0800, Paul E. McKenney wrote:
> > > On Fri, Jan 14, 2022 at 10:38:42PM +0800, Boqun Feng wrote:
> > > > Hi,
> > > > 
> > > > On Tue, Jan 11, 2022 at 11:05:00AM -0800, Paul E. McKenney wrote:
> > > > [...]
> > > > > > > The buggy address belongs to the object at ffff8880189b5c70
> > > > > > >  which belongs to the cache request_queue_srcu of size 3816
> > > > 
> > > > This cache name drew my attention when I was trying to look into this,
> > > > because I couldn't find it in v5.16, later on I realized the UAF was
> > > > found in linux-next and the commit introducing the cache was merged into
> > > > mainline if 5.17 merge window:
> > > > 
> > > > 	704b914f15fb blk-mq: move srcu from blk_mq_hw_ctx to request_queue
> > > > 
> > > > I think the UAF is actually a bug introduced by the commit, because in
> > > > that commit srcu structure was moved from blk_mq_hw_ctx to
> > > > request_queue, and therefore the cleanup_srcu_struct() should be moved
> > > > from blk_mq_hw_sysfs_release() to blk_release_queue(), however the above
> > > > commit only deleted the one in blk_mq_hw_sysfs_release() but didn't add
> > > > a new one in blk_release_queue(). As a result when a request queue is
> > > > freed, the srcu structure is not fully clean up, therefore the UAF.
> > > > 
> > > > IOW, something like below (untested) should fix this. Copy the auther
> > > > and block maintainers.
> > > 
> > > One question for the author and block maintainers...  Why not simply have
> > > a single srcu_struct for all of the queues?  Or is there some situation
> > > where you need one queue's reader to avoid blocking other queues' SRCU
> > > grace periods?
> > 
> > Because srcu_struct is too fat, and only few drivers need it, and
> > most block drivers needn't it.
> 
> Fair points.
> 
> But would it make sense to dynamically allocate a single srcu_struct
> when the first need arose, and only remove it when the last need passed?

q->srcu is referred in fast IO path, we want to avoid to add indirect
reference so not allocating it dynamically.


Thanks,
Ming