On 10/25/18 7:38 PM, jianchao.wang wrote:
> Hi Jens
>
> On 10/26/18 12:25 AM, Jens Axboe wrote:
>> On 10/24/18 9:20 AM, Jianchao Wang wrote:
>>> When issue request directly and the task is migrated out of the
>>> original cpu where it allocates request, hctx could be ran on
>>> the cpu where it is not mapped. To fix this, insert the request
>>> if BLK_MQ_F_BLOCKING is set, check whether the current is mapped
>>> to the hctx and invoke __blk_mq_issue_directly under preemption
>>> disabled.
>>>
>>> Signed-off-by: Jianchao Wang <jianchao.w.wang@xxxxxxxxxx>
>>> ---
>>> block/blk-mq.c | 17 ++++++++++++++++-
>>> 1 file changed, 16 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/block/blk-mq.c b/block/blk-mq.c
>>> index e3c39ea..0cdc306 100644
>>> --- a/block/blk-mq.c
>>> +++ b/block/blk-mq.c
>>> @@ -1717,6 +1717,12 @@ static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
>>> {
>>> struct request_queue *q = rq->q;
>>> bool run_queue = true;
>>> + blk_status_t ret;
>>> +
>>> + if (hctx->flags & BLK_MQ_F_BLOCKING) {
>>> + bypass_insert = false;
>>> + goto insert;
>>> + }
>>
>> I'd do a prep patch that moves the insert logic out of this function,
>> and just have the caller do it by return BLK_STS_RESOURCE, for instance.
>> It's silly that we have that in both the caller and inside this function.
>
> Yes.
>
>>
>>> @@ -1734,6 +1740,11 @@ static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
>>> if (q->elevator && !bypass_insert)
>>> goto insert;
>>>
>>> + if (!cpumask_test_cpu(get_cpu(), hctx->cpumask)) {
>>> + bypass_insert = false;
>>> + goto insert;
>>> + }
>>
>> Should be fine to just do smp_processor_id() here, as we're inside
>> hctx_lock() here.
>>
>
> If the rcu is preemptible, smp_processor_id will not enough here.
True, for some reason I keep forgetting that rcu_*_lock() doesn't
imply preempt_disable() anymore.
--
Jens Axboe
