Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Fri, 28 Jan 2022 09:34:56 +0100
From:   Sebastian Andrzej Siewior <bigeasy@linutronix.de>
To:     "Jason A. Donenfeld" <Jason@zx2c4.com>
Cc:     Jonathan =?utf-8?Q?Neusch=C3=A4fer?= <j.neuschaefer@gmx.net>,
        Andy Lutomirski <luto@amacapital.net>,
        LKML <linux-kernel@vger.kernel.org>,
        Theodore Ts'o <tytso@mit.edu>,
        Peter Zijlstra <peterz@infradead.org>,
        Ingo Molnar <mingo@redhat.com>, Will Deacon <will@kernel.org>,
        Waiman Long <longman@redhat.com>,
        Boqun Feng <boqun.feng@gmail.com>
Subject: Re: "BUG: Invalid wait context" in invalidate_batched_entropy
Message-ID: <YfOqsOiNfURyvFRX@linutronix.de>
References: <YfMa0QgsjCVdRAvJ@latitude>
 <CAHmME9pb9A4SN6TTjNvvxKqw1L3gXVOX7KKihfEH4AgKGNGZ2A@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
In-Reply-To: <CAHmME9pb9A4SN6TTjNvvxKqw1L3gXVOX7KKihfEH4AgKGNGZ2A@mail.gmail.com>
Precedence: bulk

On 2022-01-27 23:26:32 [+0100], Jason A. Donenfeld wrote:
> Hi Jonathan,
Hi Jason,

> Thanks for the report. I'll try to reproduce this and see what's going on.
> 
> I'm emailing back right away, though, so that I can CC in Andy
> Lutomirski, who I know has been sitting on a stack of patches that fix
> up (actually, remove) the locking, so this might be one path to fixing
> this.

This report is due to CONFIG_PROVE_LOCKING=y _and_
CONFIG_PROVE_RAW_LOCK_NESTING=y. It reports a nesting problem
(raw_spinlock_t -> spinlock_t lock ordering) which becomes a real
problem on PREEMPT_RT.

I've been testing my old series on top of 5.17-rc1. With blake2 the
numbers lowered a little. I'm gettin 3-6us on average and 16-26us
worstcase and with NUMA it still goes up to 40-50us.
If you still object the previous approach and neither tglx/peterz
disagree we could try making the lock raw_spinlock_t and add a mutex_t
around the userspace interface to lower the lock contention. But even
then we need to find a way to move the crng init part (crng_fast_load())
out of the hard-IRQ.

> Thanks,
> Jason

Sebastian