Received-SPF: pass (google.com: domain of linux-kernel+bounces-125968-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.48.161 as permitted sender) client-ip=147.75.48.161;
From: Hillf Danton <hdanton@sina.com>
To: Uladzislau Rezki <urezki@gmail.com>
Cc: syzbot <syzbot+1fa663a2100308ab6eab@syzkaller.appspotmail.com>,
	Alexei Starovoitov <alexei.starovoitov@gmail.com>,
	"Paul E. McKenney" <paulmck@kernel.org>,
	linux-kernel@vger.kernel.org,
	syzkaller-bugs@googlegroups.com
Subject: Re: [syzbot] [bpf?] possible deadlock in kvfree_call_rcu
Date: Sun, 31 Mar 2024 14:23:14 +0800
Message-Id: <20240331062314.3394-1-hdanton@sina.com>
In-Reply-To: <ZghSHWexmBHXpyA-@pc638.lan>
References: <000000000000f2f0c0061494e610@google.com>
Precedence: bulk
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit

On Sat, 30 Mar 2024 18:55:41 +0100 Uladzislau Rezki <urezki@gmail.com>
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index d9642dd06c25..8867aac3668c 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -3467,19 +3467,19 @@ void kvfree_call_rcu(struct rcu_head *head, void *ptr)
>  	 */
>  	kmemleak_ignore(ptr);
>  
> -	// Set timer to drain after KFREE_DRAIN_JIFFIES.
> -	if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING)
> -		schedule_delayed_monitor_work(krcp);
> -

This is not enough at least WRT run_page_cache_worker() [1]

[1] https://lore.kernel.org/lkml/0000000000007a44120614e27cb7@google.com/

while the reason why syzbot failed to catch the zone->per_cpu_pageset in
setup_zone_pageset() in mm/page_alloc.c is trylock [2]

[2] https://lore.kernel.org/lkml/000000000000a5ee4e0614ee586e@google.com/

>  unlock_return:
>  	krc_this_cpu_unlock(krcp, flags);
>  
> -	/*
> -	 * Inline kvfree() after synchronize_rcu(). We can do
> -	 * it from might_sleep() context only, so the current
> -	 * CPU can pass the QS state.
> -	 */
> -	if (!success) {
> +	if (success) {
> +		// Set timer to drain after KFREE_DRAIN_JIFFIES.
> +		if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING)
> +			schedule_delayed_monitor_work(krcp);
> +	} else {
> +		/*
> +		 * Inline kvfree() after synchronize_rcu(). We can do
> +		 * it from might_sleep() context only, so the current
> +		 * CPU can pass the QS state.
> +		 */
>  		debug_rcu_head_unqueue((struct rcu_head *) ptr);
>  		synchronize_rcu();
>  		kvfree(ptr);
> 
> 
> --