Re: [PATCH v2] workqueue.c: Increase workqueue name length

[Rasmus Villemoes]

On 10/01/2024 22.52, Rafael Aquini wrote:
> On Wed, Jan 10, 2024 at 09:47:56PM +0100, Rasmus Villemoes wrote:
>> On 10/01/2024 21.29, Audra Mitchell wrote:
>>
>>> @@ -4663,9 +4663,10 @@ struct workqueue_struct *alloc_workqueue(const char *fmt,
>>>  					 unsigned int flags,
>>>  					 int max_active, ...)
>>>  {
>>> -	va_list args;
>>> +	va_list args, args_copy;
>>>  	struct workqueue_struct *wq;
>>>  	struct pool_workqueue *pwq;
>>> +	int len;
>>>  
>>>  	/*
>>>  	 * Unbound && max_active == 1 used to imply ordered, which is no longer
>>> @@ -4692,6 +4693,13 @@ struct workqueue_struct *alloc_workqueue(const char *fmt,
>>>  	}
>>>  
>>>  	va_start(args, max_active);
>>> +	va_copy(args_copy, args);
>>> +	len = vsnprintf(NULL, 0, fmt, args_copy);
>>> +	WARN(len > WQ_NAME_LEN,
>>> +		"workqueue: wq->name too long (%d). Truncated to WQ_NAME_LEN (%d)\n",
>>> +		len, WQ_NAME_LEN);
>>> +
>>> +	va_end(args_copy);
>>>  	vsnprintf(wq->name, sizeof(wq->name), fmt, args);
>>
>> Eh, why not just _not_ throw away the return value from the existing
>> vsnprintf() and do "len >= sizeof(wq->name)" to know if truncation
>> happened? There's really no need need to do vsnprintf() twice. (And yes,
>> you want >=, not >).
>>
> 
> The extra vsnprintf call is required because the return of the existing 
> vsnprintf() is going to be already capped by sizeof(wq->name).

No, it is not. vsnprintf() returns the length of the would-be-created
string if the buffer was big enough. That is independent of whether one
does a dummy NULL,0 call or just calls it with a real, but possibly too
small, buffer.

This is true for userspace (as required by posix) as well as the kernel
implementation of vsnprintf(). What makes you think otherwise?

The kernel _also_ happens to have a non-standardized function called
vscnprintf (note the c) which returns the possibly-truncated result. But
that's irrelevant here.

>> Oh, and definitely not WARN,  pr_warn() or pr_warn_once() please.
>>
> 
> Then you lose the ability to figure out what was trying to create the
> wq with the inflated name. Also, the _once variants don't seem to do
> good here, because alloc_workqueue() can be called by different 
> drivers.

I assume that whatever creates the wq will do so on every boot, and the
name is most likely some fixed thing. So you're essentially setting up
some configurations to do a WARN on every single boot, not to mention
that for some machines that implies a panic... It really is not
something that warrants a WARN.

As for figuring out what caused that too-long name, well, I'd hope that
the 31 meaningful bytes that did get produced would provide a
sufficiently good hint.

Rasmus