Re: [RESEND PATCH] ION: Sys_heap: Add cached pool to spead up cached buffer alloc

[Chen Feng <puck.chen@hisilicon.com>]

Hi Laura,

On 2016/5/10 7:50, Laura Abbott wrote:
> On 05/09/2016 01:37 AM, Chen Feng wrote:
>> Add ion cached pool in system heap.
>>
>> Signed-off-by: Chen Feng <puck.chen@hisilicon.com>
>> Signed-off-by: Xia  Qing <saberlily.xia@hisilicon.com>
>> Reviewed-by: Fu Jun <oliver.fu@hisilicon.com>
>> ---
>>  drivers/staging/android/ion/ion_system_heap.c | 145 +++++++++++++++++---------
>>  1 file changed, 95 insertions(+), 50 deletions(-)
>>
>> diff --git a/drivers/staging/android/ion/ion_system_heap.c b/drivers/staging/android/ion/ion_system_heap.c
>> index b69dfc7..c633252 100644
>> --- a/drivers/staging/android/ion/ion_system_heap.c
>> +++ b/drivers/staging/android/ion/ion_system_heap.c
>> @@ -49,7 +49,8 @@ static inline unsigned int order_to_size(int order)
>>
>>  struct ion_system_heap {
>>      struct ion_heap heap;
>> -    struct ion_page_pool *pools[0];
>> +    struct ion_page_pool *uncached_pools[0];
>> +    struct ion_page_pool *cached_pools[0];
>>  };
>>
>>  static struct page *alloc_buffer_page(struct ion_system_heap *heap,
>> @@ -57,39 +58,36 @@ static struct page *alloc_buffer_page(struct ion_system_heap *heap,
>>                        unsigned long order)
>>  {
>>      bool cached = ion_buffer_cached(buffer);
>> -    struct ion_page_pool *pool = heap->pools[order_to_index(order)];
>> +    struct ion_page_pool *pool;
>>      struct page *page;
>>
>> -    if (!cached) {
>> -        page = ion_page_pool_alloc(pool);
>> -    } else {
>> -        gfp_t gfp_flags = low_order_gfp_flags;
>> -
>> -        if (order > 4)
>> -            gfp_flags = high_order_gfp_flags;
>> -        page = alloc_pages(gfp_flags | __GFP_COMP, order);
>> -        if (!page)
>> -            return NULL;
>> -        ion_pages_sync_for_device(NULL, page, PAGE_SIZE << order,
>> -                        DMA_BIDIRECTIONAL);
>> -    }
>> +    if (!cached)
>> +        pool = heap->uncached_pools[order_to_index(order)];
>> +    else
>> +        pool = heap->cached_pools[order_to_index(order)];
>>
>> +    page = ion_page_pool_alloc(pool);
>>      return page;
>>  }
>>
> 
> This is a change in behavior. The page is no longer guaranteed to be synced
> in the cache. If the page came from the pool (not just freshly allocated)
> the cache state is unknown. Do you have a good explanation why we no longer
> need to do the cache flush here on every allocation?
> 
Yes, no more explanation here.
Please see my v1[1] version.

Thanks for review!

[1] http://thread.gmane.org/gmane.linux.kernel/2217505

> Thanks,
> Laura
> 
> .
>