Re: [PATCH] locking/semaphore: Use wake_q to wake up processes outside lock critical section

From: Peter Zijlstra <peterz@infradead.org>
To: Waiman Long <longman@redhat.com>
Cc: Ingo Molnar <mingo@redhat.com>, Will Deacon <will.deacon@arm.com>,
	Boqun Feng <boqun.feng@gmail.com>,
	linux-kernel@vger.kernel.org
Subject: Re: [PATCH] locking/semaphore: Use wake_q to wake up processes outside lock critical section
Date: Thu, 10 Feb 2022 11:19:31 +0100	[thread overview]
Message-ID: <YgTms3tGYeQ4HYFZ@hirez.programming.kicks-ass.net> (raw)
In-Reply-To: <20220118153254.358748-1-longman@redhat.com>

On Tue, Jan 18, 2022 at 10:32:54AM -0500, Waiman Long wrote:
> The following lockdep splat was observed:
> 
> [ 9776.459819] ======================================================
> [ 9776.459820] WARNING: possible circular locking dependency detected
> [ 9776.459821] 5.14.0-0.rc4.35.el9.x86_64+debug #1 Not tainted
> [ 9776.459823] ------------------------------------------------------
> [ 9776.459824] stress-ng/117708 is trying to acquire lock:
> [ 9776.459825] ffffffff892d41d8 ((console_sem).lock){-...}-{2:2}, at: down_trylock+0x13/0x70
> 
> [ 9776.459831] but task is already holding lock:
> [ 9776.459832] ffff888e005f6d18 (&rq->__lock){-.-.}-{2:2}, at: raw_spin_rq_lock_nested+0x27/0x130
> 
> [ 9776.459837] which lock already depends on the new lock.
>       :
> [ 9776.459857] -> #1 (&p->pi_lock){-.-.}-{2:2}:
> [ 9776.459860]        __lock_acquire+0xb72/0x1870
> [ 9776.459861]        lock_acquire+0x1ca/0x570
> [ 9776.459862]        _raw_spin_lock_irqsave+0x40/0x90
> [ 9776.459863]        try_to_wake_up+0x9d/0x1210
> [ 9776.459864]        up+0x7a/0xb0
> [ 9776.459864]        __up_console_sem+0x33/0x70
> [ 9776.459865]        console_unlock+0x3a1/0x5f0
> [ 9776.459866]        vprintk_emit+0x23b/0x2b0
> [ 9776.459867]        devkmsg_emit.constprop.0+0xab/0xdc
> [ 9776.459868]        devkmsg_write.cold+0x4e/0x78
> [ 9776.459869]        do_iter_readv_writev+0x343/0x690
> [ 9776.459870]        do_iter_write+0x123/0x340
> [ 9776.459871]        vfs_writev+0x19d/0x520
> [ 9776.459871]        do_writev+0x110/0x290
> [ 9776.459872]        do_syscall_64+0x3b/0x90
> [ 9776.459873]        entry_SYSCALL_64_after_hwframe+0x44/0xae
>       :
> [ 9776.459905] Chain exists of:
> [ 9776.459906]   (console_sem).lock --> &p->pi_lock --> &rq->__lock
> 
> [ 9776.459911]  Possible unsafe locking scenario:
> 
> [ 9776.459913]        CPU0                    CPU1
> [ 9776.459914]        ----                    ----
> [ 9776.459914]   lock(&rq->__lock);
> [ 9776.459917]                                lock(&p->pi_lock);
> [ 9776.459919]                                lock(&rq->__lock);
> [ 9776.459921]   lock((console_sem).lock);
> 
> [ 9776.459923]  *** DEADLOCK ***

Is this new due to the ongoing printk rewrite? Also the above doesn't
actually describe the whole invesion, the one where console sem is taken
inside pi_lock is missing.

More concerning, that ordering is invalid to begin with, so the above
seems like a very poor justification for this patch.

> The problematic locking sequence ((console_sem).lock --> &p->pi_lock)
> was caused by the fact the semaphore up() function is calling
> wake_up_process() while holding the semaphore raw spinlock.
> 
> By moving the wake_up_processs() call out of the raw spinlock critical
> section using wake_q, it will break the problematic locking sequence as
> well as reducing raw spinlock hold time which will be good for
> PREEMPT_RT.
> 
> Signed-off-by: Waiman Long <longman@redhat.com>

I have no problem with the patch, just the justification / Changelog.