Re: [syzbot] WARNING: locking bug in umh_complete

[Schspa Shi]

Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:

> On Mon, Feb 06, 2023 at 07:51:16AM -0800, Luis Chamberlain wrote:
>
>> I think this seems to be the same issue that Schspa Shi reported / provided a
>> fix sugggestion for [0]. This lead me to ask if:
>> 
>>   a) incorrect usage of completion on stack could be generic and;
>>   b) if we should instead have an API helper for that?
>> 
>> Although he already implemented a suggestion for b) to answer a) we need
>> some SmPL constructs yet to be written by Schspa. The reason I asked for
>> b) is that if this is a regular pattern it begs for a) as this sort of
>> issue could be prevalent in other places. So the status of Schspa's work
>> was that he was going to work on the SmPL grammar to check how frequent
>> this incorrect patern could be found.
>
> Do I read correctly, from you above alphabet-soup, that someone is
> working on some static analysis for on-stack completions or something?
>

Yes, I was trying to do this.

> If so, perhaps the simplest rule would to be ensure there is an
> unconditional uninterruptible wait-for-completion() before going out of
> scope.
>
> This latter can be spelled like wait_for_completion() or
> wait_for_completion_state(TASK_UNINTERRUPTIBLE). More specifically,
> TASK_INTERRUPTIBLE and TASK_WAKEKILL must not be set in the state mask
> for the wait to be uninterruptible.
>
> If it cannot be proven, raise a warning and audit or somesuch.

This is a good suggestion. I have written a SmPL patch to complete this
check, and now I need to rule out the situation that the driver has
added an additional lock to protect it.

And I have found a lot of bad usage, should we consider adding a new
helper API to simplify the fix this?

Such as:

+
+void complete_on_stack(struct completion **x)
+{
+       struct completion *comp = xchg(*x, NULL);
+
+       if (comp)
+               complete(comp);
+}
+EXPORT_SYMBOL(complete_on_stack);
+
+int __sched wait_for_completion_state_on_stack(struct completion **x,
+                                       unsigned int state)
+{
+       struct completion *comp = *x;
+       int retval;
+
+       retval = wait_for_completion_state(comp, state);
+       if (retval) {
+               if (xchg(*x, NULL))
+                       return retval;
+
+               /*
+                * complete_on_stack will call complete shortly.
+                */
+               wait_for_completion(comp);
+       }
+
+       return retval;
+}
+EXPORT_SYMBOL(wait_for_completion_state_on_stack);

Link: https://lore.kernel.org/all/20221115140233.21981-1-schspa@xxxxxxxxx/T/#mf6a41a7009bb47af1b15adf2b7b355e495f609c4

-- 
BRs
Schspa Shi