linux-kernel - [PATCHv2 0/2] blk-mq: add CPU latency limit control

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [thread-next>] [day] [month] [year] [list]

Message-ID: <20241018075416.436916-1-tero.kristo@linux.intel.com>
Date: Fri, 18 Oct 2024 10:30:36 +0300
From: Tero Kristo <tero.kristo@...ux.intel.com>
To: axboe@...nel.dk
Cc: hch@....de,
	linux-block@...r.kernel.org,
	linux-kernel@...r.kernel.org
Subject: [PATCHv2 0/2] blk-mq: add CPU latency limit control

Hello,

Try #2 of the patches here. I sent earlier an RFC version against block/bio
[1], and then an isolated patch for NVMe driver only [2].

After feedback from maintainers, I've reworked the patches again to block
layer, this time to multiqueue support only. Once a blk-mq request is about
to be dispatched to the driver layer, the PM QoS variables for active CPUs
are tweaked based on configuration, and a timeout is launched as a delayed
work that drops the PM QoS limits once the queue is idle. The mechanism is
disabled by default, and only enabled once user activates it via the
provided sysfs knobs.

Some measurement data provided below as a reference for the
results, measured with 'fio' on an Intel Icelake Xeon platform, with an
extra NVMe card on the system. Both latency and bandwidth values are
provided, to showcase that the latency is reduced and bandwidth is not
impacted negatively due to overhead. C6 residency measurement is not
very accurate in my test leading to somewhat glitchy result on its
value (c6%).

key:
  slat: start latency max, in us
  clat: completion latency max, in us
  lat: overall latency max, in us
  bw: min-avg-max bandwidth values
  c6%: c6 (deep idle) residency for the active CPU during the test

cpu_lat_limit_us=10 (enabled)
  slat: 63, clat: 107, lat: 115, bw: 1177-1367-1397, c6%: 11.9
  slat: 30, clat: 129, lat: 137, bw: 1196-1380-1409, c6%: 0.9
  slat: 60, clat: 101, lat: 109, bw: 1193-1372-1407, c6%: 0.9
  slat: 29, clat: 135, lat: 143, bw: 1184-1369-1398, c6%: 1.0
  slat: 29, clat: 112, lat: 120, bw: 1188-1368-1397, c6%: 1.0
cpu_lat_limit_us=-1 (disabled)
  slat: 106, clat: 281, lat: 353, bw: 1183-1363-1403, c6%: 79.9
  slat: 107, clat: 270, lat: 319, bw: 1192-1370-1406, c6%: 79.8
  slat: 156, clat: 269, lat: 323, bw: 1187-1363-1398, c6%: 80.4
  slat: 106, clat: 267, lat: 316, bw: 1183-1367-1402, c6%: 80.5
  slat: 108, clat: 247, lat: 313, bw: 1186-1368-1404, c6%: 80.0
  slat: 107, clat: 274, lat: 323, bw: 1188-1361-1399, c6%: 80.0

-Tero

[1] https://lore.kernel.org/lkml/ZtHWkn2FJhAa+Vvo@fedora/T/
[2] https://lore.kernel.org/lkml/20241004101014.3716006-1-tero.kristo@linux.intel.com/