Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Pavel Tatashin <pasha.tatashin@soleen.com>
To:     pasha.tatashin@soleen.com, tyhicks@linux.microsoft.com,
        axboe@kernel.dk, linux-block@vger.kernel.org,
        linux-kernel@vger.kernel.org, sashal@kernel.org, jmorris@namei.org,
        lukas.bulwahn@gmail.com, hch@lst.de, pvorel@suse.cz,
        ming.lei@redhat.com, mzxreary@0pointer.de, mcgrof@kernel.org,
        zhengbin13@huawei.com, maco@android.com, colin.king@canonical.com,
        evgreen@chromium.org
Subject: [PATCH v3 0/1] scale loop device lock
Date:   Mon, 25 Jan 2021 15:11:55 -0500
Message-Id: <20210125201156.1330164-1-pasha.tatashin@soleen.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

Changelog
v3
  - Added review-by Tyler
  - Sync with mainline
v2
  - Addressed Tyler Hicks comments
  - added mutex_destroy()
  - comment in lo_open()
  - added lock around lo_disk
===

In our environment we are using systemd portable containers in
squashfs formats, convert them into loop device, and mount.

NAME                      MAJ:MIN RM   SIZE RO TYPE  MOUNTPOINT
loop5                       7:5    0  76.4M  0 loop
`-BaseImageM1908          252:3    0  76.4M  1 crypt /BaseImageM1908
loop6                       7:6    0    20K  0 loop
`-test_launchperf20       252:17   0   1.3M  1 crypt /app/test_launchperf20
loop7                       7:7    0    20K  0 loop
`-test_launchperf18       252:4    0   1.5M  1 crypt /app/test_launchperf18
loop8                       7:8    0     8K  0 loop
`-test_launchperf8        252:25   0    28K  1 crypt app/test_launchperf8
loop9                       7:9    0   376K  0 loop
`-test_launchperf14       252:29   0  45.7M  1 crypt /app/test_launchperf14
loop10                      7:10   0    16K  0 loop
`-test_launchperf4        252:11   0   968K  1 crypt app/test_launchperf4
loop11                      7:11   0   1.2M  0 loop
`-test_launchperf17       252:26   0 150.4M  1 crypt /app/test_launchperf17
loop12                      7:12   0    36K  0 loop
`-test_launchperf19       252:13   0   3.3M  1 crypt /app/test_launchperf19
loop13                      7:13   0     8K  0 loop
...

We have over 50 loop devices which are mounted  during boot.

We observed contentions around loop_ctl_mutex.

The sample contentions stacks:

Contention 1:
__blkdev_get()
   bdev->bd_disk->fops->open()
      lo_open()
         mutex_lock_killable(&loop_ctl_mutex); <- contention

Contention 2:
__blkdev_put()
   disk->fops->release()
      lo_release()
         mutex_lock(&loop_ctl_mutex); <- contention

With total time waiting for loop_ctl_mutex ~18.8s during boot (across 8
CPUs) on our machine (69 loop devices): 2.35s per CPU.

Scaling this lock eliminates this contention entirely, and improves the boot
performance by 2s on our machine.

v2 https://lore.kernel.org/lkml/20200723211748.13139-1-pasha.tatashin@soleen.com
v1 https://lore.kernel.org/lkml/20200717205322.127694-1-pasha.tatashin@soleen.com

Pavel Tatashin (1):
  loop: scale loop device by introducing per device lock

 drivers/block/loop.c | 92 +++++++++++++++++++++++++-------------------
 drivers/block/loop.h |  1 +
 2 files changed, 54 insertions(+), 39 deletions(-)

-- 
2.25.1