Re: [RFCv2][PATCH] flexible array implementation

[Li Zefan <lizf@cn.fujitsu.com>]

12:34, Dave Hansen wrote:
> On Wed, 2009-07-22 at 11:25 +0800, Li Zefan wrote:
>>> +/**
>>> + * flex_array_put - copy data into the array at @element_nr
>>> + * @src:	address of data to copy into the array
>>> + * @element_nr:	index of the position in which to insert
>>> + * 		the new element.
>> @fa and @flags are not documented.
> 
> True...  But one of my pet peeves are kerneldocs like this:
> 
> 	@fa: the flex array
> 	@flags: GFP flags
> 
> It's so trivially obvious from looking at the types and the variable
> names that I'm not sure it's worth the cost of the lines.  
> 

I'm not kernel-doc expert, but ./scripts/kernel-doc will warn
on this. And from time to time, we receive patches to fix
kernel-doc.

>>> + *
>>> + * Note that this *copies* the contents of @src into
>>> + * the array.  If you are trying to store an array of
>>> + * pointers, make sure to pass in &ptr instead of ptr.
>>> + *
>>> + * Locking must be provided by the caller.
>>> + */
>>> +int flex_array_put(struct flex_array *fa, int element_nr, void *src, gfp_t flags)
>>> +{
>>> +	int part_nr = fa_element_to_part_nr(fa, element_nr);
>>> +	struct flex_array_part *part;
>>> +	void *dst;
>>> +
>>> +	part = __fa_get_part(fa, part_nr, flags);
>>> +	if (!part)
>>> +		return -ENOMEM;
>> So this may allocate memory, and has disavantages:
>>
>> - If flex_array_put() is called in atomic context, flags has to be GFP_ATOMIC.
>> - and thus it may fail.
>>
>> Since we pass the total_elem to flex_array_alloc(), how about add a flag,
>> and if the flag is set, the alloc() will also allocate all fa_parts?
>>
>> And add __flex_array_put(), which assumes fa_parts has been allocated.
> 
> How about flex_array_prealloc()?  It seems to work for all the radix
> tree users.
> 

I have no strong opinion. I just want a non-fail version of
flex_array_put() (I mean "void __flex_array_put()").