Re: [syzbot] WARNING in call_timer_fn

From: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
To: Luiz Augusto von Dentz <luiz.dentz@gmail.com>,
	Thomas Gleixner <tglx@linutronix.de>
Cc: Hillf Danton <hdanton@sina.com>,
	syzbot <syzbot+6fb78d577e89e69602f9@syzkaller.appspotmail.com>,
	linux-kernel@vger.kernel.org, pbonzini@redhat.com,
	syzkaller-bugs@googlegroups.com,
	Steven Rostedt <rosted@goodmis.org>,
	Marcel Holtmann <marcel@holtmann.org>
Subject: Re: [syzbot] WARNING in call_timer_fn
Date: Fri, 18 Nov 2022 09:53:41 +0900	[thread overview]
Message-ID: <d38fe3c3-1f0b-b5f6-2895-aee9476b20bf@I-love.SAKURA.ne.jp> (raw)
In-Reply-To: <CABBYNZJYbzX9xoqU8_KEArTJ2=5eMTiDTjMbNjEYNVmi1rkp+A@mail.gmail.com>

On 2022/11/18 6:16, Luiz Augusto von Dentz wrote:
> Wasn't the following patch suppose to address such problem:
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/bluetooth/bluetooth-next.git/commit/?id=deee93d13d385103205879a8a0915036ecd83261
> 
> It was merged in the last pull request to net-next:
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/bluetooth/bluetooth-next.git/commit/?id=a507ea32b9c2c407012bf89109ac0cf89fae313c
> 

No. Commit deee93d13d38 ("Bluetooth: use hdev->workqueue when queuing hdev->{cmd,ncmd}_timer works")
is for handling queue_work() from hci_cmd_timeout(struct work_struct *work) from process_one_work()
(that is, called from a kernel workqueue thread from process context).

But this report says that queue_work() is called from timer interrupt handler from interrupt context
while drain_workqueue(hdev->workqueue) is in progress from process context.

But... is is_chained_work() check appropriate?
Why can't we exclude "timer interrupt handler" from "somebody else" ?

The comment for drain_workqueue() says

 * Wait until the workqueue becomes empty.  While draining is in progress,
 * only chain queueing is allowed.  IOW, only currently pending or running
 * work items on @wq can queue further work items on it.  @wq is flushed
 * repeatedly until it becomes empty.  The number of flushing is determined
 * by the depth of chaining and should be relatively short.  Whine if it
 * takes too long.

but why limited to "only currently pending or running work items on @wq" (that is,
only process context) ?

Although drain_workqueue() is also called from destroy_workqueue() (which would cause
use-after-free bug if interrupt handler calls queue_work() some time after drain_workqueue()),
I think that we can wait for drain_workqueue() to call __flush_workqueue() again if a further
work item is queued from interrupt handler...

Anyway, stopping all works and delayed works before calling drain_workqueue() would address
this problem.