On 9/24/22 10:01, Jens Axboe wrote:
> On 9/23/22 6:59 PM, Damien Le Moal wrote:
>> On 9/24/22 05:54, Jens Axboe wrote:
>>> On 9/23/22 9:13 AM, Pankaj Raghav wrote:
>>>> On 2022-09-23 16:52, Pankaj Raghav wrote:
>>>>> On Thu, Sep 22, 2022 at 12:28:01PM -0600, Jens Axboe wrote:
>>>>>> The filesystem IO path can take advantage of allocating batches of
>>>>>> requests, if the underlying submitter tells the block layer about it
>>>>>> through the blk_plug. For passthrough IO, the exported API is the
>>>>>> blk_mq_alloc_request() helper, and that one does not allow for
>>>>>> request caching.
>>>>>>
>>>>>> Wire up request caching for blk_mq_alloc_request(), which is generally
>>>>>> done without having a bio available upfront.
>>>>>>
>>>>>> Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>
>>>>>> ---
>>>>>> block/blk-mq.c | 80 ++++++++++++++++++++++++++++++++++++++++++++------
>>>>>> 1 file changed, 71 insertions(+), 9 deletions(-)
>>>>>>
>>>>> I think we need this patch to ensure correct behaviour for passthrough:
>>>>>
>>>>> diff --git a/block/blk-mq.c b/block/blk-mq.c
>>>>> index c11949d66163..840541c1ab40 100644
>>>>> --- a/block/blk-mq.c
>>>>> +++ b/block/blk-mq.c
>>>>> @@ -1213,7 +1213,7 @@ void blk_execute_rq_nowait(struct request *rq, bool at_head)
>>>>> WARN_ON(!blk_rq_is_passthrough(rq));
>>>>>
>>>>> blk_account_io_start(rq);
>>>>> - if (current->plug)
>>>>> + if (blk_mq_plug(rq->bio))
>>>>> blk_add_rq_to_plug(current->plug, rq);
>>>>> else
>>>>> blk_mq_sched_insert_request(rq, at_head, true, false);
>>>>>
>>>>> As the passthrough path can now support request caching via blk_mq_alloc_request(),
>>>>> and it uses blk_execute_rq_nowait(), bad things can happen at least for zoned
>>>>> devices:
>>>>>
>>>>> static inline struct blk_plug *blk_mq_plug( struct bio *bio)
>>>>> {
>>>>> /* Zoned block device write operation case: do not plug the BIO */
>>>>> if (bdev_is_zoned(bio->bi_bdev) && op_is_write(bio_op(bio)))
>>>>> return NULL;
>>>>> ..
>>>>
>>>> Thinking more about it, even this will not fix it because op is
>>>> REQ_OP_DRV_OUT if it is a NVMe write for passthrough requests.
>>>>
>>>> @Damien Should the condition in blk_mq_plug() be changed to:
>>>>
>>>> static inline struct blk_plug *blk_mq_plug( struct bio *bio)
>>>> {
>>>> /* Zoned block device write operation case: do not plug the BIO */
>>>> if (bdev_is_zoned(bio->bi_bdev) && !op_is_read(bio_op(bio)))
>>>> return NULL;
>>>
>>> That looks reasonable to me. It'll prevent plug optimizations even
>>> for passthrough on zoned devices, but that's probably fine.
>>
>> Could do:
>>
>> if (blk_op_is_passthrough(bio_op(bio)) ||
>> (bdev_is_zoned(bio->bi_bdev) && op_is_write(bio_op(bio))))
>> return NULL;
>>
>> Which I think is way cleaner. No ?
>> Unless you want to preserve plugging with passthrough commands on regular
>> (not zoned) drives ?
>
> We most certainly do, without plugging this whole patchset is not
> functional. Nor is batched dispatch, for example.
OK. Then the change to !op_is_read() is fine then.
>
--
Damien Le Moal
Western Digital Research
