Received-SPF: pass (google.com: domain of linux-kernel+bounces-144971-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Message-ID: <a0a220da-ebd4-4ce3-ae26-e1f26d374146@suse.com>
Date: Mon, 15 Apr 2024 13:08:57 +0200
Precedence: bulk
MIME-Version: 1.0
User-Agent: Mozilla Thunderbird
Subject: Re: [PATCH] usb: cdc-wdm: close race between read and workqueue
To: =?UTF-8?Q?Bj=C3=B8rn_Mork?= <bjorn@mork.no>,
 Oliver Neukum <oneukum@suse.com>
Cc: Aleksander Morgado <aleksandermj@chromium.org>,
 linux-usb@vger.kernel.org, gregkh@linuxfoundation.org, linux@roeck-us.net,
 linux-kernel@vger.kernel.org, ejcaruso@chromium.org
References: <385a3519-b45d-48c5-a6fd-a3fdb6bec92f@chromium.org>
 <87mspvi0lk.fsf@miraculix.mork.no>
 <a6ad34d3-9cce-4178-8271-0e09ced2b6f4@suse.com>
 <878r1fht93.fsf@miraculix.mork.no>
 <c35f98be-23a3-41c3-bee5-f394ce504545@suse.com>
 <8734rmj3ro.fsf@miraculix.mork.no>
Content-Language: en-US
From: Oliver Neukum <oneukum@suse.com>
In-Reply-To: <8734rmj3ro.fsf@miraculix.mork.no>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 8bit


On 15.04.24 12:53, Bjørn Mork wrote:
> Oliver Neukum <oneukum@suse.com> writes:
>> On 15.04.24 11:26, Bjørn Mork wrote:
>>> Oliver Neukum <oneukum@suse.com> writes:
>>>> On 15.04.24 08:47, Bjørn Mork wrote:
>>>>
>>>>> urb from service_outstanding_interrupt(). That's why it was added. See
>>>>> the explanation Robert wrote when introducing it:
>>>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/drivers/usb/class/cdc-wdm.c?id=c1da59dad0ebd3f9bd238f3fff82b1f7ffda7829
>>>>
>>>> Well, the explanation is correct in that we must read
>>>> data available. However, if the RESPONDING flag is set
>>>> and the URB submitted, we are already doing so.
>>> Sounds reasonable.  Except that the bug proves we didn't.
>>
>> Why? I am afraid I do not get that part.
> 
> I don't get how it happens either.  But that's the only thing changed by
> the patch.

Now you have lost me. I agree that this is the only thing that the patch
changes, but how do you derive the consequences from that?

>>   > If you are right that service_outstanding_interrupt can race
>>    againts
>>> itself (and I don't doubt that), then I guess this could also happen
>>> between failure to submit the URB and clearing the flag?
>>
>> Yes, it can. In fact in this case the behavior should not change.
>> I am afraid we have a misunderstanding. It seems to me that in the
>> unchanged driver the result of service_outstanding_interrupt()
>> is undefined.
>> Please explain.
> 
> Sorry, I am so lost here that I am probably only confusing things.  I doresp_count
> not understand why we unlock &desc->iuspin around the usb_submit_urb
> call.  And git tells me I wrote that.

Dropping iuspin there allowed you to call usb_submit_urb() with GFP_KERNEL.
clear_wdm_read_flag(), as it then existed, could not race with itself because
its only caller wdm_read() is holding a mutex.

That, however, is not very material to the question at hand. iuspin at that
time protected only resp_count. Even today the URB itself is protected by
WDM_RESPONDING. (Which is why I think that test_and_set_bit is required)

Now, if we say that service_outstanding_interrupt() is racing with itself,
we have to ask why this is helpful. Do we at least agree that the regression
Aleksander is seeing is due to the removal of a race or are we looking at a side effect?

	Regards
		Oliver