linux-kernel - Re: [RFC PATCH v5] sched: Fix performance regression introduced by mm

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

Message-ID: <20230414025926.GA250825@ziqianlu-desk2>
Date:   Fri, 14 Apr 2023 10:59:26 +0800
From:   Aaron Lu <aaron.lu@...el.com>
To:     Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
CC:     Peter Zijlstra <peterz@...radead.org>,
        <linux-kernel@...r.kernel.org>, Olivier Dion <odion@...icios.com>,
        <michael.christie@...cle.com>
Subject: Re: [RFC PATCH v5] sched: Fix performance regression introduced by
 mm_cid

On Thu, Apr 13, 2023 at 11:11:14AM -0400, Mathieu Desnoyers wrote:
> On 2023-04-13 10:37, Aaron Lu wrote:
> > On Thu, Apr 13, 2023 at 10:02:48AM -0400, Mathieu Desnoyers wrote:
> > > Introduce per-mm/cpu current concurrency id (mm_cid) to fix a PostgreSQL
> > > sysbench regression reported by Aaron Lu.
> > Initial test using postgres_sysbench showed the lock contention is gone :-)
> > Will check how it affects hackbench tomorrow.
> 
> That's good news! Not unexpected though.
> 
> One thing to look closely at is whether it's just that the qspinlocks are
> gone from the profiles but replaced by time spent in __mm_cid_get() due to
> cache-line bouncing of find-first-zero/test-and-set on the per-mm cid
> bitmap.

Found the below two appeared in profile when using cid as keyword, full
profile attached for your reference.

# Children      Self  Shared Object    Symbol
     0.97%     0.62%  [kernel.vmlinux] [k] sched_mm_cid_migrate_from
     0.48%     0.47%  [kernel.vmlinux] [k] mm_cid_get

Thanks,
Aaron

Download attachment "node0_0.profile.gz" of type "application/gzip" (101523 bytes)