On Thu, Apr 28, 2011 at 04:19:33PM +1000, NeilBrown wrote:
> On Wed, 27 Apr 2011 17:08:06 +0100 Mel Gorman <mgorman@xxxxxxx> wrote:
>
>
> > @@ -1578,7 +1589,7 @@ static inline struct sk_buff *netdev_alloc_skb_ip_align(struct net_device *dev,
> > */
> > static inline struct page *__netdev_alloc_page(struct net_device *dev, gfp_t gfp_mask)
> > {
> > - return alloc_pages_node(NUMA_NO_NODE, gfp_mask, 0);
> > + return alloc_pages_node(NUMA_NO_NODE, gfp_mask | __GFP_MEMALLOC, 0);
> > }
> >
>
> I'm puzzling a bit over this change.
> __netdev_alloc_page appears to be used to get pages to put in ring buffer
> for a network card to DMA received packets into. So it is OK to use
> __GFP_MEMALLOC for these allocations providing we mark the resulting skb as
> 'pfmemalloc' if a reserved page was used.
>
> However I don't see where that marking is done.
> I think it should be in skb_fill_page_desc, something like:
>
> if (page->pfmemalloc)
> skb->pfmemalloc = true;
>
> Is this covered somewhere else that I am missing?
>
You're not missing anything.
>From the context of __netdev_alloc_page, we do not know if the skb
is suitable for marking pfmemalloc or not (we don't have SKB_ALLOC_RX
flag for example that __alloc_skb has). The reserves are potentially
being dipped into for an unsuitable packet but it gets dropped in
__netif_receive_skb() and the memory is returned. If we mark the skb
pfmemalloc as a result of __netdev_alloc_page using a reserve page, the
packets would not get dropped as expected.
--
Mel Gorman
SUSE Labs
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>