Re: [PATCH] mm: cma: honor __GFP_ZERO flag in cma_alloc()

[Marek Szyprowski <m.szyprowski@samsung.com>]

Hi Michal,

On 2018-07-02 15:32, Michal Hocko wrote:
> On Mon 02-07-18 15:23:34, Marek Szyprowski wrote:
>> On 2018-06-13 15:39, Michal Hocko wrote:
>>> On Wed 13-06-18 05:55:46, Christoph Hellwig wrote:
>>>> On Wed, Jun 13, 2018 at 02:40:00PM +0200, Marek Szyprowski wrote:
>>>>> It is not only the matter of the spinlocks. GFP_ATOMIC is not supported
>>>>> by the
>>>>> memory compaction code, which is used in alloc_contig_range(). Right, this
>>>>> should be also noted in the documentation.
>>>> Documentation is good, asserts are better.  The code should reject any
>>>> flag not explicitly supported, or even better have its own flags type
>>>> with the few actually supported flags.
>>> Agreed. Is the cma allocator used for anything other than GFP_KERNEL
>>> btw.? If not then, shouldn't we simply drop the gfp argument altogether
>>> rather than give users a false hope for differen gfp modes that are not
>>> really supported and grow broken code?
>> Nope, all cma_alloc() callers are expected to use it with GFP_KERNEL gfp
>> mask.
>> The only flag which is now checked is __GFP_NOWARN. I can change the
>> function
>> signature of cma_alloc to:
>> struct page *cma_alloc(struct cma *cma, size_t count, unsigned int
>> align, bool no_warn);
> Are there any __GFP_NOWARN users? I have quickly hit the indirection
> trap and searching for alloc callback didn't tell me really much.

They might be via dma_alloc_from_contiguous() and dma_alloc_*() path.

>> What about clearing the allocated buffer? Should it be another bool
>> parameter, done unconditionally or moved to the callers?
> That really depends on callers. I have no idea what they actually ask
> for.

Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland