Hello,

On Fri, Sep 10, 2021 at 11:46:42AM +0800, Li Jinlin wrote:

Freeing bfqg is protected by queue lock in blkcg_deactivate_policy(), 
but getting/using bfqg is protected by blkcg lock in
bfq_io_set_weight_legacy(). If bfq_io_set_weight_legacy() get bfqg
before freeing bfqg and use bfqg in the after, the use-after-free
will occur.

CPU0                             CPU1
blkcg_deactivate_policy
  spin_lock_irq(&q->queue_lock)
                                 bfq_io_set_weight_legacy  
                                   spin_lock_irq(&blkcg->lock)
                                   blkg_to_bfqg(blkg)
                                     pd_to_bfqg(blkg->pd[pol->plid])
                                     ^^^^^^blkg->pd[pol->plid] != NULL
                                           bfqg != NULL
  pol->pd_free_fn(blkg->pd[pol->plid])
    pd_to_bfqg(blkg->pd[pol->plid])
    bfqg_put(bfqg)
      kfree(bfqg)
  blkg->pd[pol->plid] = NULL
  spin_unlock_irq(q->queue_lock);
                                   bfq_group_set_weight(bfqg, val, 0)
                                     bfqg->entity.new_weight
                                     ^^^^^^trigger uaf here 
                                   spin_unlock_irq(&blkcg->lock);

To fix this use-after-free, instead of holding blkcg->lock while
walking ->blkg_list and getting/using bfqg, RCU walk ->blkg_list and
hold the blkg's queue lock while getting/using bfqg.

I think this is a bug in blkcg_deactivate_policy() than the other way
around. blkgs are protected by both q and blkcg locks and holding either
should stabilize them. The blkcg lock nests inside q lock, so I think
blkcg_deactivate_policy() just needs to grab the matching blkcg lock before
trying to destroy blkgs.

Thanks.

-- 
tejun