linux-kernel - Re: [mm] 2d146aa3aa: vm-scalability.throughput -36.4% regression

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]

Date:   Tue, 17 Aug 2021 18:47:37 +0200
From:   Michal Koutný <mkoutny@...e.com>
To:     Feng Tang <feng.tang@...el.com>
Cc:     Johannes Weiner <hannes@...xchg.org>,
        Linus Torvalds <torvalds@...ux-foundation.org>,
        kernel test robot <oliver.sang@...el.com>,
        Roman Gushchin <guro@...com>, Michal Hocko <mhocko@...e.com>,
        Shakeel Butt <shakeelb@...gle.com>,
        Balbir Singh <bsingharora@...il.com>,
        Tejun Heo <tj@...nel.org>,
        Andrew Morton <akpm@...ux-foundation.org>,
        LKML <linux-kernel@...r.kernel.org>, lkp@...ts.01.org,
        kernel test robot <lkp@...el.com>,
        "Huang, Ying" <ying.huang@...el.com>,
        Zhengjun Xing <zhengjun.xing@...ux.intel.com>,
        andi.kleen@...el.com
Subject: Re: [mm] 2d146aa3aa: vm-scalability.throughput -36.4% regression

On Tue, Aug 17, 2021 at 10:45:00AM +0800, Feng Tang <feng.tang@...el.com> wrote:
> Initially from the perf-c2c data, the in-cacheline hotspots are only
> 0x0, and 0x10, and if we extends to 2 cachelines, there is one more
> offset 0x54 (css.flags), but still I can't figure out which member
> inside the 128 bytes range is written frequenty.

Is it certain that perf-c2c reported offsets are the cacheline of the
first bytes of struct cgroup_subsys_state? (Yeah, it looks to me so,
given what code accesses those and your padding fixing it. I'm just
raising it in case there was anything non-obvious.)

> 
> /* pah info for cgroup_subsys_state */
> struct cgroup_subsys_state {
> 	struct cgroup *            cgroup;               /*     0     8 */
> 	struct cgroup_subsys *     ss;                   /*     8     8 */
> 	struct percpu_ref          refcnt;               /*    16    16 */
> 	struct list_head           sibling;              /*    32    16 */
> 	struct list_head           children;             /*    48    16 */
> 	/* --- cacheline 1 boundary (64 bytes) --- */
> 	struct list_head           rstat_css_node;       /*    64    16 */
> 	int                        id;                   /*    80     4 */
> 	unsigned int               flags;                /*    84     4 */
> 	u64                        serial_nr;            /*    88     8 */
> 	atomic_t                   online_cnt;           /*    96     4 */
> 
> 	/* XXX 4 bytes hole, try to pack */
> 
> 	struct work_struct         destroy_work;         /*   104    32 */
> 	/* --- cacheline 2 boundary (128 bytes) was 8 bytes ago --- */
> 
> Since the test run implies this is cacheline related, and I'm not very
> familiar with the mem_cgroup code, the original perf-c2c log is attached
> which may give more hints.

As noted by Johannes, even in atomic mode, the refcnt would have the
atomic part elsewhere. The other members shouldn't be written frequently
unless there are some intense modifications of the cgroup tree in
parallel.
Does the benchmark create lots of memory cgroups in such a fashion?

Regards,
Michal