Re: [PATCH 1/4] gpio: Remove VLA from gpiolib

[Rasmus Villemoes]

On 2018-03-18 15:23, Lukas Wunner wrote:
>>>
>>> Other random thoughts: maybe two allocations for each loop iteration is
>>> a bit much. Maybe do a first pass over the array and collect the maximal
>>> chip->ngpio, do the memory allocation and freeing outside the loop (then
>>> you'd of course need to preserve the memset() with appropriate length
>>> computed). And maybe even just do one allocation, making bits point at
>>> the second half.
>>
>> I think those are great ideas because the function is kind of a hotpath
>> and usage of VLAs was motivated by the desire to make it fast.
>>
>> I'd go one step further and store the maximum ngpio of all registered
>> chips in a global variable (and update it in gpiochip_add_data_with_key()),
>> then allocate 2 * max_ngpio once before entering the loop (as you've
>> suggested).  That would avoid the first pass to determine the maximum
>> chip->ngpio.  In most systems max_ngpio will be < 64, so one or two
>> unsigned longs depending on the arch's bitness.
> 
> Actually, scratch that.  If ngpio is usually smallish, we can just
> allocate reasonably sized space for mask and bits on the stack,

Yes.

> and fall back to the kcalloc slowpath only if chip->ngpio exceeds
> that limit.

Well, I'd suggest not adding that fallback code now, but simply add a
check in gpiochip_add_data_with_key to ensure ngpio is sane (and refuse
to register the chip otherwise), at least if we know that every
currently supported/known chip is covered by the 256 (?). That keeps the
code simple and fast, and then if somebody has a chip with 40000 gpio
lines, we can add a fallback path. Or we could consider alternative
solutions, to avoid a 10000 byte GFP_ATOMIC allocation (maybe hang a
pre-allocation off the gpio_chip; that's only two more bits per
descriptor, and there's already a whole gpio_desc for each - but not
sure about the locking in that case).

Rasmus