[v2] block: fix disordered IO in the case recursive split

[PATCH RFC v2 02/10] md/raid0: convert raid0_handle_discard() to use bio_submit_split_bioset()

Posted by Yu Kuai 5 months, 2 weeks ago

From: Yu Kuai <yukuai3@huawei.com>

On the one hand unify bio split code, prepare to fix disordered split
IO; On the other hand fix missing blkcg_bio_issue_init() and
trace_block_split() for split IO.

Noted commit 319ff40a5427 ("md/raid0: Fix performance regression for large
sequential writes") already fix disordered split IO by converting bio to
underlying disks before submit_bio_noacct(), with the respect
md_submit_bio() already split by sectors, and raid0_make_request() will
split at most once for unaligned IO. This is a bit hacky and we'll convert
this to solution in general later.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
---
 drivers/md/raid0.c | 19 +++++++------------
 1 file changed, 7 insertions(+), 12 deletions(-)

diff --git a/drivers/md/raid0.c b/drivers/md/raid0.c
index f1d8811a542a..4dcc5133d679 100644
--- a/drivers/md/raid0.c
+++ b/drivers/md/raid0.c
@@ -463,21 +463,16 @@ static void raid0_handle_discard(struct mddev *mddev, struct bio *bio)
 	zone = find_zone(conf, &start);
 
 	if (bio_end_sector(bio) > zone->zone_end) {
-		struct bio *split = bio_split(bio,
-			zone->zone_end - bio->bi_iter.bi_sector, GFP_NOIO,
-			&mddev->bio_set);
-
-		if (IS_ERR(split)) {
-			bio->bi_status = errno_to_blk_status(PTR_ERR(split));
-			bio_endio(bio);
+		bio = bio_submit_split_bioset(bio,
+				zone->zone_end - bio->bi_iter.bi_sector,
+				&mddev->bio_set);
+		if (!bio)
 			return;
-		}
-		bio_chain(split, bio);
-		submit_bio_noacct(bio);
-		bio = split;
+
 		end = zone->zone_end;
-	} else
+	} else {
 		end = bio_end_sector(bio);
+	}
 
 	orig_end = end;
 	if (zone != conf->strip_zone)
-- 
2.39.2

Re: [PATCH RFC v2 02/10] md/raid0: convert raid0_handle_discard() to use bio_submit_split_bioset()

Posted by Damien Le Moal 5 months, 2 weeks ago

On 8/28/25 15:57, Yu Kuai wrote:
> From: Yu Kuai <yukuai3@huawei.com>
> 
> On the one hand unify bio split code, prepare to fix disordered split
> IO; On the other hand fix missing blkcg_bio_issue_init() and
> trace_block_split() for split IO.

Hmmm... Shouldn't that be a prep patch with a fixes tag for backport ?
Because that "fix" here is not done directly but is the result of calling
bio_submit_split_bioset().

> 
> Noted commit 319ff40a5427 ("md/raid0: Fix performance regression for large
> sequential writes") already fix disordered split IO by converting bio to
> underlying disks before submit_bio_noacct(), with the respect
> md_submit_bio() already split by sectors, and raid0_make_request() will
> split at most once for unaligned IO. This is a bit hacky and we'll convert
> this to solution in general later.
> 
> Signed-off-by: Yu Kuai <yukuai3@huawei.com>
> ---
>  drivers/md/raid0.c | 19 +++++++------------
>  1 file changed, 7 insertions(+), 12 deletions(-)
> 
> diff --git a/drivers/md/raid0.c b/drivers/md/raid0.c
> index f1d8811a542a..4dcc5133d679 100644
> --- a/drivers/md/raid0.c
> +++ b/drivers/md/raid0.c
> @@ -463,21 +463,16 @@ static void raid0_handle_discard(struct mddev *mddev, struct bio *bio)
>  	zone = find_zone(conf, &start);
>  
>  	if (bio_end_sector(bio) > zone->zone_end) {
> -		struct bio *split = bio_split(bio,
> -			zone->zone_end - bio->bi_iter.bi_sector, GFP_NOIO,
> -			&mddev->bio_set);
> -
> -		if (IS_ERR(split)) {
> -			bio->bi_status = errno_to_blk_status(PTR_ERR(split));
> -			bio_endio(bio);
> +		bio = bio_submit_split_bioset(bio,
> +				zone->zone_end - bio->bi_iter.bi_sector,

Can this ever be negative (of course not I think)? But if
bio_submit_split_bioset() is changed to have an unsigned int sectors count,
maybe add a sanity check before calling bio_submit_split_bioset() ?

> +				&mddev->bio_set);
> +		if (!bio)
>  			return;
> -		}
> -		bio_chain(split, bio);
> -		submit_bio_noacct(bio);
> -		bio = split;
> +
>  		end = zone->zone_end;
> -	} else
> +	} else {
>  		end = bio_end_sector(bio);
> +	}
>  
>  	orig_end = end;
>  	if (zone != conf->strip_zone)


-- 
Damien Le Moal
Western Digital Research

Re: [PATCH RFC v2 02/10] md/raid0: convert raid0_handle_discard() to use bio_submit_split_bioset()

Posted by Yu Kuai 5 months, 2 weeks ago

Hi,

在 2025/8/30 8:41, Damien Le Moal 写道:
> On 8/28/25 15:57, Yu Kuai wrote:
>> From: Yu Kuai <yukuai3@huawei.com>
>>
>> On the one hand unify bio split code, prepare to fix disordered split
>> IO; On the other hand fix missing blkcg_bio_issue_init() and
>> trace_block_split() for split IO.
> Hmmm... Shouldn't that be a prep patch with a fixes tag for backport ?
> Because that "fix" here is not done directly but is the result of calling
> bio_submit_split_bioset().

I can add a fix tag as blkcg_bio_issue_init() and trace_block_split() is missed,
however, if we consider stable backport, should we fix this directly from caller
first? As this is better for backport. Later this patch can be just considered
cleanup.

>> Noted commit 319ff40a5427 ("md/raid0: Fix performance regression for large
>> sequential writes") already fix disordered split IO by converting bio to
>> underlying disks before submit_bio_noacct(), with the respect
>> md_submit_bio() already split by sectors, and raid0_make_request() will
>> split at most once for unaligned IO. This is a bit hacky and we'll convert
>> this to solution in general later.
>>
>> Signed-off-by: Yu Kuai <yukuai3@huawei.com>
>> ---
>>   drivers/md/raid0.c | 19 +++++++------------
>>   1 file changed, 7 insertions(+), 12 deletions(-)
>>
>> diff --git a/drivers/md/raid0.c b/drivers/md/raid0.c
>> index f1d8811a542a..4dcc5133d679 100644
>> --- a/drivers/md/raid0.c
>> +++ b/drivers/md/raid0.c
>> @@ -463,21 +463,16 @@ static void raid0_handle_discard(struct mddev *mddev, struct bio *bio)
>>   	zone = find_zone(conf, &start);
>>   
>>   	if (bio_end_sector(bio) > zone->zone_end) {
>> -		struct bio *split = bio_split(bio,
>> -			zone->zone_end - bio->bi_iter.bi_sector, GFP_NOIO,
>> -			&mddev->bio_set);
>> -
>> -		if (IS_ERR(split)) {
>> -			bio->bi_status = errno_to_blk_status(PTR_ERR(split));
>> -			bio_endio(bio);
>> +		bio = bio_submit_split_bioset(bio,
>> +				zone->zone_end - bio->bi_iter.bi_sector,
> Can this ever be negative (of course not I think)? But if
> bio_submit_split_bioset() is changed to have an unsigned int sectors count,
> maybe add a sanity check before calling bio_submit_split_bioset() ?

Yes, this can never be negative.

Thanks,
Kuai

>
>> +				&mddev->bio_set);
>> +		if (!bio)
>>   			return;
>> -		}
>> -		bio_chain(split, bio);
>> -		submit_bio_noacct(bio);
>> -		bio = split;
>> +
>>   		end = zone->zone_end;
>> -	} else
>> +	} else {
>>   		end = bio_end_sector(bio);
>> +	}
>>   
>>   	orig_end = end;
>>   	if (zone != conf->strip_zone)
>

Re: [PATCH RFC v2 02/10] md/raid0: convert raid0_handle_discard() to use bio_submit_split_bioset()

Posted by Damien Le Moal 5 months, 2 weeks ago

On 8/30/25 13:10, Yu Kuai wrote:
> Hi,
> 
> 在 2025/8/30 8:41, Damien Le Moal 写道:
>> On 8/28/25 15:57, Yu Kuai wrote:
>>> From: Yu Kuai <yukuai3@huawei.com>
>>>
>>> On the one hand unify bio split code, prepare to fix disordered split
>>> IO; On the other hand fix missing blkcg_bio_issue_init() and
>>> trace_block_split() for split IO.
>> Hmmm... Shouldn't that be a prep patch with a fixes tag for backport ?
>> Because that "fix" here is not done directly but is the result of calling
>> bio_submit_split_bioset().
> 
> I can add a fix tag as blkcg_bio_issue_init() and trace_block_split() is missed,
> however, if we consider stable backport, should we fix this directly from caller
> first? As this is better for backport. Later this patch can be just considered
> cleanup.

That is what I was suggesting: fix the blkcg issue first withe fixes tag and
then do the conversion to using bio_submit_split_bioset() in later patch that is
not to be backported.

-- 
Damien Le Moal
Western Digital Research