On 6/02/19 2:00 PM, Ulf Hansson wrote:
> + Adrian, Linus
>
> On Tue, 5 Feb 2019 at 17:53, Zachary Hays <zhays@xxxxxxxxxxx> wrote:
>>
>> The kblockd workqueue is created with the WQ_MEM_RECLAIM flag set.
>> This generates a rescuer thread for that queue that will trigger when
>> the CPU is under heavy load and collect the uncompleted work.
>>
>> In the case of mmc, this creates the possibility of a deadlock as
>> other blk-mq work is also run on the same queue. For example:
>>
>> - worker 0 claims the mmc host
>> - worker 1 attempts to claim the host
>> - worker 0 schedules complete_work to release the host
>> - rescuer thread is triggered after time-out and collects the dangling
>> work
>> - rescuer thread attempts to complete the work in order starting with
>> claim host
>> - the task to release host is now blocked by a task to claim it and
>> will never be called
>>
>
> Adrian, I need your help to understand more of this. The above
> description doesn't add up to me.
>
> In principle, already when "worker 1 attempts to claim the host" as
> described above, it should succeed and should *not* need to wait for
> the host to be released. Right?
If it is the same queue, then yes. Although in that case there is only 1
work for the hw queue so there cannot be another worker. There could be
another attempt to send a request directly, but that will not block - if the
host controller is busy, BLK_STS_RESOURCE will be returned from ->queue_rq().
>
> The hole point with the commit 6c0cedd1ef95 ("mmc: core: Introduce
> host claiming by context"), was to allow the mmc host to be
> re-claimable for blk I/O requests, no matter from what worker/thread
> the claim/release is done from.
>
> Is it not working as expected you think? What am I missing here?
I assumed we were talking about a situation where there are multiple
internal eMMC partitions each with their own disk and queue. In that case,
a queue waits if there is another queue that is using the eMMC.
We should clarify whether that is the scenario we are looking at. Zachary?
>
>> The above results in multiple hung tasks that lead to failures to boot.
>
> Of course, during boot there is also a card detect work running in
> parallel with blk I/O requests. This work is being punted to the
> "system_freezable_wq" and it claims/releases the host as well.
>
> Perhaps, that could have something to do with the problem....
>
>>
>> Handling complete_work on a separate workqueue avoids this by keeping
>> the work completion tasks separate from the other blk-mq work. This
>> allows the host to be released without getting blocked by other tasks
>> attempting to claim the host.
>>
>> Signed-off-by: Zachary Hays <zhays@xxxxxxxxxxx>
>> ---
>> drivers/mmc/core/block.c | 10 +++++++++-
>> include/linux/mmc/card.h | 1 +
>> 2 files changed, 10 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/mmc/core/block.c b/drivers/mmc/core/block.c
>> index aef1185f383d..14f3fdb8c6bb 100644
>> --- a/drivers/mmc/core/block.c
>> +++ b/drivers/mmc/core/block.c
>> @@ -2112,7 +2112,7 @@ static void mmc_blk_mq_req_done(struct mmc_request *mrq)
>> if (waiting)
>> wake_up(&mq->wait);
>> else
>> - kblockd_schedule_work(&mq->complete_work);
>> + queue_work(mq->card->complete_wq, &mq->complete_work);
>>
>> return;
>> }
>> @@ -2924,6 +2924,13 @@ static int mmc_blk_probe(struct mmc_card *card)
>>
>> mmc_fixup_device(card, mmc_blk_fixups);
>>
>> + card->complete_wq = alloc_workqueue("mmc_complete",
>> + WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
>> + if (unlikely(!card->complete_wq)) {
>> + pr_err("Failed to create mmc completion workqueue");
>> + return -ENOMEM;
>> + }
>> +
>> md = mmc_blk_alloc(card);
>> if (IS_ERR(md))
>> return PTR_ERR(md);
>> @@ -2987,6 +2994,7 @@ static void mmc_blk_remove(struct mmc_card *card)
>> pm_runtime_put_noidle(&card->dev);
>> mmc_blk_remove_req(md);
>> dev_set_drvdata(&card->dev, NULL);
>> + destroy_workqueue(card->complete_wq);
>> }
>>
>> static int _mmc_blk_suspend(struct mmc_card *card)
>> diff --git a/include/linux/mmc/card.h b/include/linux/mmc/card.h
>> index de7377815b6b..8ef330027b13 100644
>> --- a/include/linux/mmc/card.h
>> +++ b/include/linux/mmc/card.h
>> @@ -308,6 +308,7 @@ struct mmc_card {
>> unsigned int nr_parts;
>>
>> unsigned int bouncesz; /* Bounce buffer size */
>> + struct workqueue_struct *complete_wq; /* Private workqueue */
>> };
>>
>> static inline bool mmc_large_sector(struct mmc_card *card)
>> --
>> 2.7.4
>>
>
> So this change seems to solve the problem, which is really great.
> However, I think we really need to understand what goes wrong, before
> we apply this as fix.
>
> Unfortunate, I am also in busy period, so I need a couple of more days
> before I can help out running tests.
>
> Kind regards
> Uffe
>
