Re: [PATCH RFC stable 4.14 1/1] mmc: core: fix hung task caused by race condition on context_info

["Luis Claudio R. Goncalves" <lgoncalv@redhat.com>]

On Thu, Sep 29, 2022 at 02:41:26PM +0200, Ulf Hansson wrote:
> On Mon, 5 Sept 2022 at 08:22, dinggao.pan <dinggao.pan@horizon.ai> wrote:
> >
> > Hi,
> > After applying rt patches to our 4.14 kernel and enabling preempt-rt, we met a hung task during boot caused by race condition on context_info stored in struct mmc_host.
> > From our investigation, context_info should not be changed by threads that have not claimed the host, hence the following fix.
> >
> > Any comments are much appreciated.
> > Dinggao Pan
> 
> Hi Dinggao,
> 
> Apologize for the delay.
> 
> The 4.14 kernel is too old for me to be able to comment. In
> particular, the mmc block layer moved to blk-mq in v4.16, which means
> the path you are investigating doesn't exist any more, sorry.

And the new code has the queue operations protected by a spinlock
(queue_lock), which I believe is necessary to fix the issue reported
here.

Luis
 
> Kind regards
> Uffe
> 
> >
> > From: "Dinggao Pan" <mailto:dinggao.pan@horizon.ai>
> >
> > 　　A race condition happens under following circumstances:
> >     (mmc_thread1)               |              (mmc_thread2)
> >     mmc_issue_rq(req1)          |
> >       > qcnt++ for req1         |
> >         host handling req1      |
> >     mmc_queue_thread(req=null)  |
> >       > enter queue thread      |
> >         again, fetches blk req  |
> >         (return null), sets     |
> >         is_waiting_last_req 1   |  mmc_request_fn(req1) -> set is_new_req 1
> >                                 |                   and wake_up wait_queue
> >     mmc_issue_rq(req2)          |   > mmc_thread2 tries to claim host
> >       > **qcnt++ for req2**     |
> >       mmc_finalize_req(req2)    |
> >         > should wait for req1  |
> >           done but req2 return  |
> >           MMC_BLK_NEW_REQ       |
> >           due to is_new_req     |
> >           already set to 1      |
> >                                 |
> >                                 |
> >     req1 done                   |
> >       > qcnt-- for req1         |
> >     mmc_issue_rq(req3)          |
> >       > qcnt++ for req3         |
> > req2 is not handled but qcnt is already added(noted by **),
> > thus mmc_thread1 will never release host, causing mmc_threads
> > except thread1 to hung. Fix race by moving wake_up to the front of
> > context_info update.
> >
> > Reviewed By: Yunqian Wang <mailto:yunqian.wang@horizon.ai>
> > Signed-off-by: Dinggao Pan <mailto:dinggao.pan@horizon.ai>
> > Signed-off-by: Ming Yu <mailto:ming.yu@horizon.ai>
> > ---
> > drivers/mmc/core/queue.c | 7 +++++--
> > 1 file changed, 5 insertions(+), 2 deletions(-)
> >
> > diff --git a/drivers/mmc/core/queue.c b/drivers/mmc/core/queue.c
> > index 0a4e77a5b..58318c102 100644
> > --- a/drivers/mmc/core/queue.c
> > +++ b/drivers/mmc/core/queue.c
> > @@ -107,6 +107,11 @@ static void mmc_request_fn(struct request_queue *q)
> >                return;
> >       }
> >
> > +      if (mq->asleep) {
> > +               wake_up_process(mq->thread);
> > +               return;
> > +      }
> > +
> >       cntx = &mq->card->host->context_info;
> >
> >       if (cntx->is_waiting_last_req) {
> > @@ -114,8 +119,6 @@ static void mmc_request_fn(struct request_queue *q)
> >                wake_up_interruptible(&cntx->wait);
> >       }
> >
> > -       if (mq->asleep)
> > -                wake_up_process(mq->thread);
> > }
> >
> > static struct scatterlist *mmc_alloc_sg(int sg_len, gfp_t gfp)
> > --
> > 2.36.1
> 
---end quoted text---