mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
From: "yu kuai" <yukuai@fygo.io>
To: "Herman Semenoff" <germanaizek0@gmail.com>, <song@kernel.org>,
	 <magiclinan@didiglobal.com>, <xiao@kernel.org>,
	 <linux-raid@vger.kernel.org>, <linux-kernel@vger.kernel.org>,
	 "yu kuai" <yukuai@fygo.io>
Subject: Re: [PATCH] md/raid0: optimize stripe mapping with bitwise shifts for power-of-2 devices
Date: Sat, 21 Nov 2026 18:58:01 +0800	[thread overview]
Message-ID: <f1aa2a91-0859-44d7-9cc1-b64ed4def889@fygo.io> (raw)
In-Reply-To: <CAMnMt5Z+BwOaPWQGTTUeSr8uaT=dysMi42xRpFpUruRRcO7=wA@mail.gmail.com>

Hi,

在 2026/11/17 13:15, Herman Semenoff 写道:
>  From 086678202572260393864f72c06f4d67032a4b76 Mon Sep 17 00:00:00 2001
> From: Herman Semenoff <GermanAizek@yandex.ru>
> Date: Sun, 8 Nov 2026 22:43:29 +0300
> Subject: [PATCH] md/raid0: optimize stripe mapping with bitwise shifts
> for power-of-2 devices
>
> In drivers/md/raid0.c, map_sector() performs two 64-bit integer divisions
> (sector_div) on every submitted bio:
>    1. sector_div(chunk, zone->nb_dev << chunksect_bits) to compute the chunk
>       offset within the member device.
>    2. sector_div(sector, zone->nb_dev) to select the target member disk.
>
> For ultra-low latency solid-state storage devices, such as Intel Optane
> (3D XPoint NVMe SSDs) capable of sub-10us latencies and multi-million IOPS,
> these un-pipelined 64-bit division instructions create unnecessary CPU
> pipeline stalls and latency overhead in the I/O submission hot path.
>
> When the number of member devices in a strip zone is a power of two
> (e.g., 2, 4, 8, 16 disks, which is the predominant configuration for
> NVMe arrays), both divisions can be replaced with bitwise shifts and
> masks:
>    - chunk >>= (zone->nb_dev_shift + chunksect_bits)
>    - disk_idx = sector & (zone->nb_dev - 1)
>    - *sector_offset = (chunk << chunksect_bits) + sect_in_chunk
>
> Store nb_dev_shift (ilog2(zone->nb_dev)) in struct strip_zone during zone
> initialization when is_power_of_2(zone->nb_dev) is true, or -1 otherwise.
> If nb_dev_shift is non-negative and chunk_sects is a power of two, fast
> bitwise math is used with zero divisions. Arrays with non-power-of-2 disk
> counts fall back to the existing sector_div() path without any behavior
> or compatibility change.
>
> In addition, annotate the bio split check in raid0_make_request() with
> unlikely(), as standard MD devices already enforce chunk_sectors limits
> via bio_split_to_limits() in md_submit_bio().
>
> Tested on an 8-disk Intel Optane Memory M10 (MEMPEK1J016GAD) RAID0 array:
>    - Random 4K read QD1 latency:  avg lat 22.06 us -> 21.10 us (-4.4%)
>    - Random 4K read QD1 IOPS:     42.6k IOPS       -> 44.6k IOPS (+4.7%)
>    - Random 4K read QD32x8 IOPS:  1.11M IOPS       -> 1.72M IOPS (+54.5%)
>    - Non-power-of-2 disk arrays:  unaffected, identical fallback path
>
> Signed-off-by: Herman Semenoff <GermanAizek@yandex.ru>
> Signed-off-by: Herman Semenoff <GermanAizek@aizekgroup.ru>
> ---
>   drivers/md/raid0.c | 26 +++++++++++++++++++-------
>   drivers/md/raid0.h |  1 +
>   2 files changed, 20 insertions(+), 7 deletions(-)

This patch is unacceptable, you need to regenerate it with correct format.

>
> diff --git a/drivers/md/raid0.c b/drivers/md/raid0.c
> index 35e103f0c2c3..993df48eaa4c 100644
> --- a/drivers/md/raid0.c
> +++ b/drivers/md/raid0.c
> @@ -207,6 +207,8 @@ static int create_strip_zones(struct mddev *mddev,
> struct r0conf **private_conf)
>    }
>    zone->nb_dev = cnt;
>    zone->zone_end = smallest->sectors * cnt;
> + zone->nb_dev_shift = is_power_of_2(zone->nb_dev) ?
> +     ilog2(zone->nb_dev) : -1;
>
>    curr_zone_end = zone->zone_end;
>
> @@ -246,6 +248,8 @@ static int create_strip_zones(struct mddev *mddev,
> struct r0conf **private_conf)
>    }
>
>    zone->nb_dev = c;
> + zone->nb_dev_shift = is_power_of_2(zone->nb_dev) ?
> +     ilog2(zone->nb_dev) : -1;
>    sectors = (smallest->sectors - zone->dev_start) * c;
>    pr_debug("md/raid0:%s: zone->nb_dev: %d, sectors: %llu\n",
>    mdname(mddev),
> @@ -330,6 +334,7 @@ static struct md_rdev *map_sector(struct mddev
> *mddev, struct strip_zone *zone,
>    struct r0conf *conf = mddev->private;
>    int raid_disks = conf->strip_zone[0].nb_dev;
>    unsigned int chunk_sects = mddev->chunk_sectors;
> + unsigned int disk_idx;
>
>    if (is_power_of_2(chunk_sects)) {
>    int chunksect_bits = ffz(~chunk_sects);
> @@ -338,21 +343,28 @@ static struct md_rdev *map_sector(struct mddev
> *mddev, struct strip_zone *zone,
>    sector >>= chunksect_bits;
>    /* chunk in zone */
>    chunk = *sector_offset;
> - /* quotient is the chunk in real device*/
> - sector_div(chunk, zone->nb_dev << chunksect_bits);
> - } else{
> + if (zone->nb_dev_shift >= 0) {
> + chunk >>= (zone->nb_dev_shift + chunksect_bits);
> + disk_idx = sector & (zone->nb_dev - 1);
> + } else {
> + /* quotient is the chunk in real device*/
> + sector_div(chunk, zone->nb_dev << chunksect_bits);
> + disk_idx = sector_div(sector, zone->nb_dev);
> + }
> + *sector_offset = (chunk << chunksect_bits) + sect_in_chunk;
> + } else {
>    sect_in_chunk = sector_div(sector, chunk_sects);
>    chunk = *sector_offset;
>    sector_div(chunk, chunk_sects * zone->nb_dev);
> + disk_idx = sector_div(sector, zone->nb_dev);
> + *sector_offset = (chunk * chunk_sects) + sect_in_chunk;
>    }
>    /*
>    *  position the bio over the real device
>    *  real sector = chunk in device + starting of zone
>    * + the position in the chunk
>    */
> - *sector_offset = (chunk * chunk_sects) + sect_in_chunk;
> - return conf->devlist[(zone - conf->strip_zone)*raid_disks
> -     + sector_div(sector, zone->nb_dev)];
> + return conf->devlist[(zone - conf->strip_zone) * raid_disks + disk_idx];
>   }
>
>   static sector_t raid0_size(struct mddev *mddev, sector_t sectors, int
> raid_disks)
> @@ -616,7 +628,7 @@ static bool raid0_make_request(struct mddev
> *mddev, struct bio *bio)
>    ? (sector & (chunk_sects-1))
>    : sector_div(sector, chunk_sects));
>
> - if (sectors < bio_sectors(bio)) {
> + if (unlikely(sectors < bio_sectors(bio))) {
>    bio = bio_submit_split_bioset(bio, sectors,
>         &mddev->bio_set);
>    if (!bio)
> diff --git a/drivers/md/raid0.h b/drivers/md/raid0.h
> index 8cc761ca7423..70b73511fb14 100644
> --- a/drivers/md/raid0.h
> +++ b/drivers/md/raid0.h
> @@ -7,6 +7,7 @@ struct strip_zone {
>    sector_t dev_start; /* Zone offset in real dev (in sectors) */
>    int nb_dev; /* # of devices attached to the zone */
>    int disk_shift; /* start disk for the original layout */
> + int nb_dev_shift; /* log2(nb_dev) if power of 2, else -1 */
>   };
>
>   /* Linux 3.14 (20d0189b101) made an unintended change to

-- 
Thanks,
Kuai

      reply	other threads:[~2026-11-21 10:58 UTC|newest]

Thread overview: 2+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-17  5:15 Herman Semenoff
2026-11-21 10:58 ` yu kuai [this message]

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=f1aa2a91-0859-44d7-9cc1-b64ed4def889@fygo.io \
    --to=yukuai@fygo.io \
    --cc=germanaizek0@gmail.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-raid@vger.kernel.org \
    --cc=magiclinan@didiglobal.com \
    --cc=song@kernel.org \
    --cc=xiao@kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox

all inboxes | Powered by JetHome®