Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Thu, 17 Feb 2022 18:33:07 +0100
From:   Sebastian Andrzej Siewior <bigeasy@linutronix.de>
To:     "Jason A. Donenfeld" <Jason@zx2c4.com>
Cc:     Jann Horn <jannh@google.com>,
        Andy Lutomirski <luto@amacapital.net>,
        Boqun Feng <boqun.feng@gmail.com>,
        Will Deacon <will@kernel.org>,
        Peter Zijlstra <peterz@infradead.org>,
        Ingo Molnar <mingo@redhat.com>,
        Waiman Long <longman@redhat.com>,
        Sultan Alsawaf <sultan@kerneltoast.com>,
        Theodore Ts'o <tytso@mit.edu>,
        Andy Lutomirski <luto@kernel.org>,
        Jonathan =?utf-8?Q?Neusch=C3=A4fer?= <j.neuschaefer@gmx.net>,
        LKML <linux-kernel@vger.kernel.org>,
        Thomas Gleixner <tglx@linutronix.de>
Subject: Re: [PATCH v3] random: remove batched entropy locking
Message-ID: <Yg6G08tFN5kjamG2@linutronix.de>
References: <CAHmME9pO41uwYExSROc5X2+RE=a5tZfE=c=bAxVbhCHfa7=zSA@mail.gmail.com>
 <20220204155142.56419-1-Jason@zx2c4.com>
 <Yf1M3YGVq71oC9BM@linutronix.de>
 <CAG48ez1ucJikx_6GzK2XUfCGoTeL+R418riTn+ECj_ud5BSFow@mail.gmail.com>
 <CAHmME9q6bm_pXer0aVq907-RevtH8nLPH=Uo2UU6gEKU6GaFwA@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
In-Reply-To: <CAHmME9q6bm_pXer0aVq907-RevtH8nLPH=Uo2UU6gEKU6GaFwA@mail.gmail.com>
Precedence: bulk

On 2022-02-16 21:58:14 [+0100], Jason A. Donenfeld wrote:
> > Why raw_cpu_ptr? include/linux/percpu-defs.h says about raw_*() operations:
> >
> >  * Operations for contexts where we do not want to do any checks for
> >  * preemptions.  Unless strictly necessary, always use [__]this_cpu_*()
> >  * instead.
> >
> > So when I see a raw_*() percpu thing, I read it as "it is expected
> > that we can migrate after this point (or we're in some really weird
> > context where the normal context check doesn't work)". Is that
> > incorrect?
> 
> If it says "contexts where we do not want to do any checks for
> preemptions", then that would apply here I would think? We're taking a
> local lock, which means afterwards there are no preemptions. For
> context, the code that got committed after Sebastian's final review
> is:
> 
>         local_lock_irqsave(&batched_entropy_u32.lock, flags);
>         batch = raw_cpu_ptr(&batched_entropy_u32);
> 
> However, I think most other code uses this_cpu_ptr() instead? So not sure.

It depends what you are looking for.
raw_cpu_ptr(&batched_entropy_u32) give you the pointer to
&batched_entropy_u32 of _this_ CPU - the CPU you are currently running.
this_cpu_ptr() does the same except that it has smp_processor_id() in
it. smp_processor_id() will yell at you (given you enabled
CONFIG_DEBUG_PREEMPT) if the code can migrate to another CPU.
So it will yell at you unless:
- you disable preemption / migration so code can't migrate
- you run on a per-CPU thread i.e. a thread which is bound to a single
  CPU and therefore can't migrate to another.

The local_lock_irqsave() acquires a lock / disables interrupt and
therefore can't migrate. So I suggested to use raw_cpu_ptr() as an
optimisation in the debug-case.

If you disable preemption before accessing a per-CPU variable with
this_cpu_ptr() then you _need_ to ensure that nobody is accessing the
variable from in-IRQ context. Nobody will yell at you. But if you use a
local-lock then you get lockdep annotation and lockdep will complain if
you use that variable always in process context and sometimes in-IRQ.

> Jason

Sebastian