netdev - Re: [PATCH net-next v2 4/4] net: dqs: make struct dql more cache efficient

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: <ZhaZuOjq+b5nzqcs@gmail.com>
Date: Wed, 10 Apr 2024 06:52:56 -0700
From: Breno Leitao <leitao@...ian.org>
To: Jakub Kicinski <kuba@...nel.org>
Cc: davem@...emloft.net, pabeni@...hat.com, edumazet@...gle.com,
	linux-kernel@...r.kernel.org, netdev@...r.kernel.org
Subject: Re: [PATCH net-next v2 4/4] net: dqs: make struct dql more cache
 efficient

On Tue, Apr 09, 2024 at 05:21:49PM -0700, Jakub Kicinski wrote:
> On Mon,  8 Apr 2024 10:25:56 -0700 Breno Leitao wrote:
> > With the previous change, struct dqs->stall_thrs will be in the hot path
> > (at queue side), even if DQS is disabled.
> > 
> > The other fields accessed in this function (last_obj_cnt and num_queued)
> > are in the first cache line, let's move this field  (stall_thrs) to the
> > very first cache line, since there is a hole there.
> > 
> > This does not change the structure size, since it moves an short (2
> > bytes) to 4-bytes whole in the first cache line.
> 
> Doesn't this move the cache line bouncing problem to the other side?

I think so. Looking at dql_check_stall(), it only uses fields in the
second cache line, except now 'dql->stall_thrs' that is in the first
cache line.

> Eric said "copy" I read that as "have two fields with the same value".

Sorry, I misunderstood it. I can create two fields, and update them
together at the only place where they will be updated
(bql_set_stall_thrs).

> I think it's single digit number of alu instructions we'd be saving
> here, not super convinced patch 3 is the right trade off...

Right. I was more concerned about the write barriers (smp_wmb()) inside
the loop which happen quite frequently.

But, if you think that the this is not the right approach, I can drop
this whole patchset. Do you think a profiler will us here?

Thanks!