Message-ID: <a236ab6b-d38c-3974-d4cb-5e92d0877abc@I-love.SAKURA.ne.jp>
Date:   Sat, 4 Feb 2023 23:21:27 +0900
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:102.0) Gecko/20100101
 Thunderbird/102.6.1
Subject: Re: Converting dev->mutex into dev->spinlock ?
Content-Language: en-US
To:     Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc:     "Rafael J. Wysocki" <rafael@kernel.org>,
        LKML <linux-kernel@vger.kernel.org>,
        USB list <linux-usb@vger.kernel.org>
References: <28a82f50-39d5-a45f-7c7a-57a66cec0741@I-love.SAKURA.ne.jp>
 <Y95h7Vop9t5Li0HD@kroah.com>
From:   Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
In-Reply-To: <Y95h7Vop9t5Li0HD@kroah.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 2023/02/04 22:47, Greg Kroah-Hartman wrote:
> On Sat, Feb 04, 2023 at 10:32:11PM +0900, Tetsuo Handa wrote:
>> Hello.
>>
>> There is a long-standing deadlock problem in driver core code caused by
>> "struct device"->mutex being marked as "do not apply lockdep checks".
> 
> The marking of a lock does not cause a deadlock problem, so what do you
> mean exactly by this?  Where is the actual deadlock?

A few of examples:

  https://syzkaller.appspot.com/bug?extid=2d6ac90723742279e101
  https://syzkaller.appspot.com/bug?extid=2e39bc6569d281acbcfb
  https://syzkaller.appspot.com/bug?extid=9ef743bba3a17c756174

> 
>> We can make this deadlock visible by applying [1], and we can confirm that
>> there is a deadlock problem that I think needs to be addressed in core code [2].
> 
> Any reason why you didn't cc: us on these patches?

We can't apply this "drivers/core: Remove lockdep_set_novalidate_class() usage" patch
until we fix all lockdep warnings that happen during the boot stage; otherwise syzbot
testing can't work which is more painful than applying this patch now.

Therefore, I locally tested this patch (in order not to be applied now). And I got
a lockdep warning on the perf_event code. I got next lockdep warning on the driver core
code when I tried a fix for the perf_event code suggested by Peter Zijlstra. Since Peter
confirmed that this is a problem that led to commit 1704f47b50b5 ("lockdep: Add novalidate
class for dev->mutex conversion"), this time I'm reporting this problem to you (so that
you can propose a fix for the driver core code).