Re: [PATCH] workqueue: unregister wq lockdep on error path in alloc_workqueue()

[Kefeng Wang]

On 2019/3/8 22:45, Bart Van Assche wrote:
> On 3/7/19 11:37 PM, Kefeng Wang wrote:
>> syzkaller report an issue "KASAN: use-after-free Read in alloc_workqueue",
>>
>> alloc_workqueue
>> Â - kzalloc wq
>> Â - wq_init_lockdep(wq);
>> ÂÂÂ - lockdep_register_key(&wq->key);Â // add to hlist
>> Â - kfree wq
>>
>> But forget to call wq_unregister_lockdep()->lockdep_unregister_key(), it
>> will delete the entry from hlist.
>>
>> Reported-by: syzbot+17335689e239ce135d8b@xxxxxxxxxxxxxxxxxxxxxxxxx
>> Fixes: 669de8bda87b ("kernel/workqueue: Use dynamic lockdep keys for workqueues")
>> Signed-off-by: Kefeng Wang <wangkefeng.wang@xxxxxxxxxx>
>> ---
>> Â kernel/workqueue.c | 1 +
>> Â 1 file changed, 1 insertion(+)
>>
>> diff --git a/kernel/workqueue.c b/kernel/workqueue.c
>> index 7abbeed13421..9209d25dfade 100644
>> --- a/kernel/workqueue.c
>> +++ b/kernel/workqueue.c
>> @@ -4291,6 +4291,7 @@ struct workqueue_struct *alloc_workqueue(const char *fmt,
>> ÂÂÂÂÂ return wq;
>> Â Â err_free_wq:
>> +ÂÂÂ wq_unregister_lockdep(wq);
>> ÂÂÂÂÂ free_workqueue_attrs(wq->unbound_attrs);
>> ÂÂÂÂÂ kfree(wq);
>> ÂÂÂÂÂ return NULL;
>
> Hi Kefeng,
>
> Five days ago I had posted a similar but slightly different patch. Can you have a look at it? See also https://lore.kernel.org/lkml/20190303220046.29448-1-bvanassche@xxxxxxx/Â; 
I don't notice your patch due to no subscribe lkml, that is better, wq->lock_name should be freed too; )
>
> Thanks,
>
> Bart.
>
>
>
> .
>