On 04/27/2017 10:04 PM, Ming Lei wrote:
> On Thu, Apr 27, 2017 at 04:51:32PM -0600, Jens Axboe wrote:
>> Instead of bypassing the scheduler for insertion of reserved requests,
>> we ensure that the request is marked as RQF_RESERVED so they driver
>> knows where it came from.
>>
>> Usually we just use the tag to know if it's reserved or not,
>> but that only works when the request has a driver tag assigned.
>> Using RQF_RESERVED can be done independently of whether or not
>> scheduling is used.
>>
>> Signed-off-by: Jens Axboe <axboe@xxxxxx>
>> ---
>> block/blk-mq-sched.c | 8 +++-----
>> block/blk-mq.c | 3 +++
>> include/linux/blkdev.h | 2 ++
>> 3 files changed, 8 insertions(+), 5 deletions(-)
>>
>> diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c
>> index 8b361e192e8a..27c67465f856 100644
>> --- a/block/blk-mq-sched.c
>> +++ b/block/blk-mq-sched.c
>> @@ -82,11 +82,7 @@ struct request *blk_mq_sched_get_request(struct request_queue *q,
>> if (likely(!data->hctx))
>> data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
>>
>> - /*
>> - * For a reserved tag, allocate a normal request since we might
>> - * have driver dependencies on the value of the internal tag.
>> - */
>> - if (e && !(data->flags & BLK_MQ_REQ_RESERVED)) {
>> + if (e) {
>> data->flags |= BLK_MQ_REQ_INTERNAL;
>>
>> /*
>> @@ -104,6 +100,8 @@ struct request *blk_mq_sched_get_request(struct request_queue *q,
>> }
>>
>> if (rq) {
>> + if (data->flags & BLK_MQ_REQ_RESERVED)
>> + rq->rq_flags |= RQF_RESERVED;
>
> I think this flag may not be needed, becasue driver can
> decide if one rq is from reversed pool just by the tag, for example
> of mtip32xx, it can be done easily by checking if rq->tag is zero.
>
> So I suggest to not introduce this flag until it is necessary.
But that only works after get_driver_tag() has been run, which is why
I added the flag. That may or may not be a big deal, depending on
what path is called before the request is sent off to be executed.
--
Jens Axboe
