Re: [net-next PATCH 1/6] net: Add skb_free_frag to replace use of put_page in freeing skb->head

[Alexander Duyck <alexander.h.duyck@redhat.com>]

On 05/06/2015 01:41 PM, Andrew Morton wrote:
> On Wed, 06 May 2015 13:27:43 -0700 Alexander Duyck <alexander.h.duyck@redhat.com> wrote:
>
>>>> +void skb_free_frag(void *head)
>>>> +{
>>>> +	struct page *page = virt_to_head_page(head);
>>>> +
>>>> +	if (unlikely(put_page_testzero(page))) {
>>>> +		if (likely(PageHead(page)))
>>>> +			__free_pages_ok(page, compound_order(page));
>>>> +		else
>>>> +			free_hot_cold_page(page, false);
>>>> +	}
>>>> +}
>>> Why are we testing for PageHead in here?  If the code were to simply do
>>>
>>> 	if (unlikely(put_page_testzero(page)))
>>> 		__free_pages_ok(page, compound_order(page));
>>>
>>> that would still work?
>> My assumption was that there was a performance difference between
>> __free_pages_ok and free_hot_cold_page for order 0 pages.  From what I
>> can tell free_hot_cold_page will do bulk cleanup via free_pcppages_bulk
>> while __free_pages_ok just calls free_one_page.
> Could be.  Plus there's hopefully some performance advantage if the
> page is genuinely cache-hot.  I don't think that anyone has verified
> the benefits of the hot/cold optimisation in the last decade or two,
> and it was always pretty marginal..

Either way it doesn't make much difference.  If you would prefer I can 
probably just call __free_pages_ok for all cases.

> Is the PageHead thing really "likely"?  We're usually dealing with
> order>0 pages here?

On any system that only supports 4K pages the default is to allocate an 
order 3 page (32K) and then pull the fragments out of that.  So if 
__free_pages_ok works for an order 0 page I'll just call it since it 
shouldn't be a very common occurrence anyway unless we are under memory 
pressure.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>