Re: [PATCH] usbhid: Fix lockdep unannotated irqs-off warning

[Sedat Dilek]

On 3/1/16, Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> wrote:
> On Tue, 1 Mar 2016, Sedat Dilek wrote:
>
>> On Tue, Oct 13, 2015 at 2:57 AM, Steven Rostedt <rostedt@xxxxxxxxxxx>
>> wrote:
>> > On Sat, 3 Oct 2015 12:05:42 +0200
>> > Sedat Dilek <sedat.dilek@xxxxxxxxx> wrote:
>> >
>> >> So, at the beginning... dunno WTF is causing the problems - no
>> >> workaround for CLANG.
>> >
>> > Probably need to compile with gcc and with clang and look at the binary
>> > differences. Or at least what objdump shows.
>> >
>>
>> [ Hope to address this issue to the correct people - CCed some people
>> I taped on their nerves ]
>>
>> Not sure if I should open a new thread?
>> Please, some clear statements on this.
>> Thanks.
>>
>> The issue is still visible and alive.
>
> ...
>
>> [ FACT #3: TEST-CASE #2 ]
>>
>> The most reliable test-case is to simply unplug my external Logitech
>> USB mouse - saw this by accident.
>> YES, it was so simple.
>>
>> --- dmesg_4.5.0-rc6-2-llvmlinux-amd64.txt       2016-02-29
>> 21:23:56.399691975 +0100
>> +++ dmesg_4.5.0-rc6-2-llvmlinux-amd64_usbmouse-unplugged.txt
>> 2016-02-29 21:28:14.401832240 +0100
>> @@ -832,3 +832,75 @@
>>  [   66.529779] PPP BSD Compression module registered
>>  [   66.563013] PPP Deflate Compression module registered
>>  [   66.978977] usb 2-1.5: USB disconnect, device number 4
>> +[  321.937369] usb 2-1.4: USB disconnect, device number 3
>> +[  321.950810] BUG: sleeping function called from invalid context at
>> kernel/workqueue.c:2785
>> +[  321.950816] in_atomic(): 0, irqs_disabled(): 1, pid: 44, name:
>> kworker/2:1
>
>> +[  321.950885] hardirqs last  enabled at (47769):
>> [<ffffffff819426a2>] _raw_spin_unlock_irq+0x32/0x60
>> +[  321.950889] hardirqs last disabled at (47770):
>> [<ffffffff81115bdc>] del_timer_sync+0x3c/0x110
>> +[  321.950894] softirqs last  enabled at (47112):
>> [<ffffffff810871a2>] __do_softirq+0x5a2/0x610
>> +[  321.950898] softirqs last disabled at (47097):
>> [<ffffffff8108747c>] irq_exit+0x9c/0x120
>> +[  321.950903] CPU: 2 PID: 44 Comm: kworker/2:1 Not tainted
>> 4.5.0-rc6-2-llvmlinux-amd64 #1
>> +[  321.950905] Hardware name: SAMSUNG ELECTRONICS CO., LTD.
>> 530U3BI/530U4BI/530U4BH/530U3BI/530U4BI/530U4BH, BIOS 13XK 03/28/2013
>> +[  321.950908] Workqueue: usb_hub_wq hub_event
>> +[  321.950911]  ffff8800d3326948 0000000000000092 0000000000000000
>> ffff8800d4347568
>> +[  321.950915]  ffffffff814ba7d4 ffff8800d43474f8 ffff8800d4340cc0
>> ffff8800d4347568
>> +[  321.950919]  ffffffff810e28fd 0000000000000092 0000000000000096
>> ffff8800d43475a8
>> +[  321.950923] Call Trace:
>> +[  321.950927]  [<ffffffff814ba7d4>] dump_stack+0xa4/0xf0
>> +[  321.950931]  [<ffffffff810e28fd>] ? print_irqtrace_events+0xcd/0xe0
>> +[  321.950936]  [<ffffffff810bf495>] ___might_sleep+0x295/0x2b0
>> +[  321.950939]  [<ffffffff810bf18f>] __might_sleep+0x4f/0xc0
>> +[  321.950943]  [<ffffffff810a12ef>] start_flush_work+0x2f/0x2a0
>> +[  321.950946]  [<ffffffff810a129c>] flush_work+0x5c/0x80
>> +[  321.950949]  [<ffffffff810a125a>] ? flush_work+0x1a/0x80
>> +[  321.950953]  [<ffffffff810e247d>] ? trace_hardirqs_off+0xd/0x10
>> +[  321.950956]  [<ffffffff810a032a>] ? try_to_grab_pending+0x4a/0x260
>> +[  321.950960]  [<ffffffff810a1717>] __cancel_work_timer+0x197/0x290
>> +[  321.950963]  [<ffffffff81115b8d>] ? try_to_del_timer_sync+0xad/0xc0
>> +[  321.950966]  [<ffffffff810a1578>] cancel_work_sync+0x18/0x20
>
> It's possible that this could be a compiler-related error connected
> with local_irq_save().  __cancel_work_timer() calls
>
> 		ret = try_to_grab_pending(work, is_dwork, &flags);
>
> and try_to_grab_pending() starts this way:
>
> static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
> 			       unsigned long *flags)
> {
> 	struct worker_pool *pool;
> 	struct pool_workqueue *pwq;
>
> 	local_irq_save(*flags);
>
> Then later on, __cancel_work_timer() does
>
> 	local_irq_restore(flags);
>
> Maybe CLANG doesn't like local_irq_save() applied to a pointer as
> opposed to a stack variable.
>
> As a quick test, try applying the patch below.  (Note that there is a
> similar construction in kernel/signal.c.)
>
> Alan Stern
>
>
>
> Index: usb-4.4/kernel/workqueue.c
> ===================================================================
> --- usb-4.4.orig/kernel/workqueue.c
> +++ usb-4.4/kernel/workqueue.c
> @@ -1175,12 +1175,14 @@ out_put:
>   * This function is safe to call from any context including IRQ handler.
>   */
>  static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
> -			       unsigned long *flags)
> +			       unsigned long *pflags)
>  {
>  	struct worker_pool *pool;
>  	struct pool_workqueue *pwq;
> +	unsigned long flags;
>
> -	local_irq_save(*flags);
> +	local_irq_save(flags);
> +	*pflags = flags;
>
>  	/* try to steal the timer if it exists */
>  	if (is_dwork) {
> @@ -1241,7 +1243,7 @@ static int try_to_grab_pending(struct wo
>  	}
>  	spin_unlock(&pool->lock);
>  fail:
> -	local_irq_restore(*flags);
> +	local_irq_restore(flags);
>  	if (work_is_canceling(work))
>  		return -ENOENT;
>  	cpu_relax();
>
>

might_sleep() is invoked in start_flush_work(), but there are no
"flags" passed to local_irq_{disable,enable}.

What about something like this?

@@ -2781,13 +2783,14 @@ static bool start_flush_work(struct
work_struct *work, struct wq_barrier *barr)
        struct worker *worker = NULL;
        struct worker_pool *pool;
        struct pool_workqueue *pwq;
+       unsigned long flags;

        might_sleep();

-       local_irq_disable();
+       local_irq_disable(flags);
        pool = get_work_pool(work);
        if (!pool) {
-               local_irq_enable();
+               local_irq_enable(flags);
                return false;
        }

- Sedat -