[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20200915155150.GD2674@hirez.programming.kicks-ass.net>
Date: Tue, 15 Sep 2020 17:51:50 +0200
From: peterz@...radead.org
To: Oleg Nesterov <oleg@...hat.com>
Cc: Hou Tao <houtao1@...wei.com>, Ingo Molnar <mingo@...hat.com>,
Will Deacon <will@...nel.org>, Dennis Zhou <dennis@...nel.org>,
Tejun Heo <tj@...nel.org>, Christoph Lameter <cl@...ux.com>,
linux-kernel@...r.kernel.org, linux-fsdevel@...r.kernel.org
Subject: Re: [RFC PATCH] locking/percpu-rwsem: use this_cpu_{inc|dec}() for
read_count
On Tue, Sep 15, 2020 at 05:31:14PM +0200, Oleg Nesterov wrote:
> > So yeah, fs/super totally abuses percpu_rwsem, and yes, using it from
> > IRQ context is totally out of spec. That said, we've (grudgingly)
> > accomodated them before.
>
> Yes, I didn't expect percpu_up_ can be called from IRQ :/
Yeah, me neither. That's well out of spec for a blocking primitive in
general.
> > This seems to be a fairly long standing issue, and certainly not unique
> > to ARM64 either (Power, and anyone else using asm-gemeric/percpu.h,
> > should be similarly affected I think). The issue seems to stem from
> > Oleg's original rewrite:
> >
> > a1fd3e24d8a4 ("percpu_rw_semaphore: reimplement to not block the readers unnecessarily")
>
> Not really... I think it was 70fe2f48152e ("aio: fix freeze protection of aio writes").
Ah, that came later? Fair enough, I'll change the Fixes line.
> And iiuc io_uring does the same.
Indeed, I just went through a bunch of the file_end_write() callers.
> > and is certainly an understandable mistake.
> >
> > I'm torn on what to do, using this_cpu over __this_cpu is going to
> > adversely affect code-gen (and possibly performance) for all the
> > percpu-rwsem users that are not quite so 'creative'.
>
> Yes, but what else can we do?
Well, I just talked about it with Will, there's a bunch of things we
could do, but they're all quite ugly.
My leading alternative was adding: percpu_down_read_irqsafe() /
percpu_up_read_irqsafe(), which use local_irq_save() instead of
preempt_disable().
But blergh.. Will also argued that by going with this patch, we'll get
an affected workload when someone reports a performance regression,
which I suppose is a bonus.
Anyway, I'll rewrite the Changelog and stuff it in locking/urgent.
Powered by blists - more mailing lists