Re: [PATCH 4/5] block: move queue enter logic into blk_mq_submit_bio()

[Jens Axboe <axboe@kernel.dk>]

On 11/11/21 8:23 AM, Ming Lei wrote:
> Hi Geert,
> 
> On Thu, Nov 11, 2021 at 03:51:28PM +0100, Geert Uytterhoeven wrote:
>> Hi Ming,
>>
>> On Thu, Nov 11, 2021 at 2:45 PM Ming Lei <ming.lei@redhat.com> wrote:
>>> On Thu, Nov 11, 2021 at 01:58:38PM +0100, Geert Uytterhoeven wrote:
>>>> On Thu, 4 Nov 2021, Jens Axboe wrote:
>>>>> On 11/4/21 1:04 PM, Christoph Hellwig wrote:
>>>>>> On Thu, Nov 04, 2021 at 01:02:54PM -0600, Jens Axboe wrote:
>>>>>>> On 11/4/21 12:52 PM, Christoph Hellwig wrote:
>>>>>>>> Looks good:
>>>>>>>>
>>>>>>>> Reviewed-by: Christoph Hellwig <hch@lst.de>
>>>>>>>
>>>>>>> So these two are now:
>>>>>>>
>>>>>>> https://git.kernel.dk/cgit/linux-block/commit/?h=for-5.16/block&id=c98cb5bbdab10d187aff9b4e386210eb2332af96
>>>>>>>
>>>>>>> which is the one I sent here, and then the next one gets cleaned up to
>>>>>>> remove that queue enter helper:
>>>>>>>
>>>>>>> https://git.kernel.dk/cgit/linux-block/commit/?h=for-5.16/block&id=7f930eb31eeb07f1b606b3316d8ad3ab6a92905b
>>>>>>>
>>>>>>> Can I add your reviewed-by to this last one as well? Only change is the
>>>>>>> removal of blk_mq_enter_queue() and the weird construct there, it's just
>>>>>>> bio_queue_enter() now.
>>>>>>
>>>>>> Sure.
>>>>>
>>>>> Thanks, prematurely already done, as you could tell :-)
>>>>
>>>> The updated version is now commit 900e080752025f00 ("block: move queue
>>>> enter logic into blk_mq_submit_bio()") in Linus' tree.
>>>>
>>>> I have bisected failures on m68k/atari (on ARAnyM, using nfhd as the
>>>> root device) to this commit, e.g.:
>>>>
>>>>     sd 0:0:0:0: [sda] tag#0 FAILED Result: hostbyte=DID_OK driverbyte=DRIVER_OK cmd_age=0s
>>>>     sd 0:0:0:0: [sda] tag#0 Sense Key : Illegal Request [current]
>>>>     sd 0:0:0:0: [sda] tag#0 Add. Sense: Invalid field in cdb
>>>>     sd 0:0:0:0: [sda] tag#0 CDB: Write(10) 2a 08 00 00 00 01 00 00 08 00
>>>>     critical target error, dev sda, sector 1 op 0x1:(WRITE) flags 0x20800 phys_seg 1 prio class 0
>>>>     Buffer I/O error on dev sda1, logical block 0, lost sync page write
>>>>
>>>>     EXT4-fs (sda1): I/O error while writing superblock
>>>>     sd 0:0:0:0: [sda] tag#0 FAILED Result: hostbyte=DID_OK driverbyte=DRIVER_OK cmd_age=0s
>>>>     sd 0:0:0:0: [sda] tag#0 Sense Key : Illegal Request [current]
>>>>     sd 0:0:0:0: [sda] tag#0 Add. Sense: Invalid field in cdb
>>>>     sd 0:0:0:0: [sda] tag#0 CDB: Write(10) 2a 08 00 00 00 01 00 00 08 00
>>>>     critical target error, dev sda, sector 1 op 0x1:(WRITE) flags 0x20800 phys_seg 1 prio class 0
>>>>     Buffer I/O error on dev sda1, logical block 0, lost sync page write
>>>>     EXT4-fs (sda1): I/O error while writing superblock
>>>>
>>>> This may happen either when mounting the root file system (leading to an
>>>> unable to mount root fs panic), or later (leading to a read-only
>>>> rootfs).
>>>
>>> BTW, today I just found that hang in blk_mq_freeze_queue_wait() is
>>> caused by commit 900e080752025f00, and the following patch can fix it:
>>>
>>> - blk-mq: don't grab ->q_usage_counter in blk_mq_sched_bio_merge
>>>
>>> https://lore.kernel.org/linux-block/20211111085650.GA476@lst.de/T/#m759b88fda094a65ebf29bc81b780967cdaf9cf28
>>>
>>> Maybe you can try the above patch.
>>
>> Thanks! I have applied both patches, but it doesn't make a difference.
> 
> Thanks for your test!
> 
> Can you try the following patch?
> 
> 
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index f511db395c7f..a5ab2f2e9f67 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -2517,7 +2517,6 @@ static struct request *blk_mq_get_new_requests(struct request_queue *q,
>  	struct blk_mq_alloc_data data = {
>  		.q		= q,
>  		.nr_tags	= 1,
> -		.cmd_flags	= bio->bi_opf,
>  	};
>  	struct request *rq;
>  
> @@ -2525,6 +2524,7 @@ static struct request *blk_mq_get_new_requests(struct request_queue *q,
>  		return NULL;
>  	if (unlikely(!submit_bio_checks(bio)))
>  		goto put_exit;
> +	data.cmd_flags	= bio->bi_opf;
>  	if (blk_mq_attempt_bio_merge(q, bio, nsegs, same_queue_rq))
>  		goto put_exit;

That's definitely a real fix, akin to the other pre-enter variants, this
one just post checks. Geert, can you give this a whirl?

Ming, would you mind sending this as a real patch?

-- 
Jens Axboe