On 5/17/23 17:58, Yu Kuai wrote:
> Hi,
>
> 在 2023/05/17 16:44, Pradeep P V K 写道:
>> There is a potential race between ioc_clear_fn() and
>> exit_io_context() as shown below, due to which below
>> crash is observed. It can also result into use-after-free
>> issue.
>>
>> context#1: context#2:
>> ioc_release_fn() do_exit();
>> ->spin_lock(&ioc->lock); ->exit_io_context();
>> ->ioc_destroy_icq(icq); ->ioc_exit_icqs();
>> ->list_del_init(&icq->q_node); ->spin_lock_irq(&ioc->lock);
>> ->call_rcu(&icq->__rcu_head,
>> icq_free_icq_rcu);
>> ->spin_unlock(&ioc->lock);
>> ->ioc_exit_icq(); gets the same icq
> I don't understand how is this possible, the list is protected by
> 'ioc->lock', both hlist_del_init and hlist_for_each_entry are called
> inside the lock.
Given that ioc_destroy_icq() calls ioc_exit_icq(), ioc_exit_icqs() should ignore
all icqs that have been destroyed already, otherwise, ioc_exit_icq() gets called
twice for the same icq. The missing rcu lock in ioc_exit_icqs() already was in
itself a bug, and the missing flag check is another.
>
> Thanks,
> Kuai
>> ->bfq_exit_icq();
>> This results into below crash as bic
>> is NULL as it is derived from icq.
>> There is a chance that icq could be
>> free'd as well.
>>
>> [33.245722][ T8666] Unable to handle kernel NULL pointer dereference
>> at virtual address 0000000000000018.
>> ...
>> Call trace:
>> [33.325782][ T8666] bfq_exit_icq+0x28/0xa8
>> [33.325785][ T8666] exit_io_context+0xcc/0x100
>> [33.325786][ T8666] do_exit+0x764/0xa58
>> [33.325791][ T8666] do_group_exit+0x0/0xa0
>> [33.325793][ T8666] invoke_syscall+0x48/0x114
>> [33.325802][ T8666] el0_svc_common+0xcc/0x118
>> [33.325805][ T8666] do_el0_svc+0x34/0xd0
>> [33.325807][ T8666] el0_svc+0x38/0xd0
>> [33.325812][ T8666] el0t_64_sync_handler+0x8c/0xfc
>> [33.325813][ T8666] el0t_64_sync+0x1a0/0x1a4
>>
>> Fix this by checking with ICQ_DESTROYED flags in ioc_exit_icqs().
>> Also, ensure ioc_exit_icq() is accessing icq within rcu_read_lock/unlock
>> so that icq doesn't get free'd up while it is still using it.
>>
>> Signed-off-by: Pradeep P V K <quic_pragalla@xxxxxxxxxxx>
Pradeep, this needs a Fixes tag and cc-stable I think.
>> ---
>> block/blk-ioc.c | 8 ++++++--
>> 1 file changed, 6 insertions(+), 2 deletions(-)
>>
>> diff --git a/block/blk-ioc.c b/block/blk-ioc.c
>> index 63fc02042408..1aa34fd46ac8 100644
>> --- a/block/blk-ioc.c
>> +++ b/block/blk-ioc.c
>> @@ -60,10 +60,14 @@ static void ioc_exit_icqs(struct io_context *ioc)
>> {
>> struct io_cq *icq;
>>
>> + rcu_read_lock();
>> spin_lock_irq(&ioc->lock);
>> - hlist_for_each_entry(icq, &ioc->icq_list, ioc_node)
>> - ioc_exit_icq(icq);
>> + hlist_for_each_entry(icq, &ioc->icq_list, ioc_node) {
>> + if (!(icq->flags & ICQ_DESTROYED))
>> + ioc_exit_icq(icq);
>> + }
>> spin_unlock_irq(&ioc->lock);
>> + rcu_read_unlock();
>> }
>>
>> /*
>>
>
--
Damien Le Moal
Western Digital Research
