linux-kernel - Re: [RFC][PATCH 0/3] sched: User Managed Concurrency Groups

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: <YbnyaCT4wh0II/Ew@casper.infradead.org>
Date:   Wed, 15 Dec 2021 13:49:28 +0000
From:   Matthew Wilcox <willy@...radead.org>
To:     Peter Zijlstra <peterz@...radead.org>
Cc:     Peter Oskolkov <posk@...k.io>, Ingo Molnar <mingo@...hat.com>,
        Thomas Gleixner <tglx@...utronix.de>, juri.lelli@...hat.com,
        Vincent Guittot <vincent.guittot@...aro.org>,
        dietmar.eggemann@....com, Steven Rostedt <rostedt@...dmis.org>,
        Ben Segall <bsegall@...gle.com>, mgorman@...e.de,
        bristot@...hat.com,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Linux Memory Management List <linux-mm@...ck.org>,
        linux-api@...r.kernel.org, x86@...nel.org,
        Paul Turner <pjt@...gle.com>, Peter Oskolkov <posk@...gle.com>,
        Andrei Vagin <avagin@...gle.com>, Jann Horn <jannh@...gle.com>,
        Thierry Delisle <tdelisle@...terloo.ca>
Subject: Re: [RFC][PATCH 0/3] sched: User Managed Concurrency Groups

On Wed, Dec 15, 2021 at 11:44:49AM +0100, Peter Zijlstra wrote:
> On Tue, Dec 14, 2021 at 07:46:25PM -0800, Peter Oskolkov wrote:
> 
> > Anyway, I'll test your patchset over the next week or so and let you
> > know if anything really needed is missing (other than waking an idle
> > server if there is one on a worker wakeup; this piece is definitely
> > needed).
> 
> Right, so the problem I'm having is that a single idle server ptr like
> before can trivially miss waking annother idle server.
> 
> Suppose:
> 
> 	umcg::idle_server_tid_ptr
> 
> Then the enqueue_and_wake() thing from the last patch would:
> 
> 	idle_server_tid = xchg((pid_t __user *)self->idle_server_tid_ptr, 0);
> 
> to consume the tid, and then use that to enqueue and wake. But what if a
> second wakeup happens right after that? There might be a second idle
> server, but we'll never find it, because userspace hasn't had time to
> update the field again.
> 
> Alternatively, we do a linked list of servers, but then every such
> wakeup needs to iterate the whole list, looking for one that has
> UMCG_TF_IDLE set, or something like that, but that lookup is bad for
> performance.
> 
> So I'm really not sure what way to go yet.

1. Linked lists are fugly and bad for the CPU.

2. I'm not sure how big the 'N' in 'M:N' is supposed to be.  Might be
one per hardware thread?  So it could be hundreds-to-thousands,
depending on the scale of system.

3. The interface between user-kernel could be an array of idle tids,
maybe 16 entries long (16 * 4 = 64 bytes, just one cacheline).  As a
server finishes work, it looks for a 0 tid in the batch and stores
its tid in the slot (cmpxchg, I guess, since the array will be shared
between processes).  If there are no free slots in the array, then we
definitely have 16 threads already waiting for work, so it can park itself
in whatever data structure userspace wants to use to manage idle servers.
It's up to userspace to decide when to repopulate the array of available
servers from its data structure of idle servers.