On Wed, 2015-04-22 at 15:35 -0400, David Miller wrote:
> From: Eric Dumazet <eric.duma...@gmail.com>
> Date: Wed, 22 Apr 2015 07:33:36 -0700
>
> > Also, we must refrain using high order pages from the pfmemalloc
> > reserve, so __page_frag_refill() must also use __GFP_NOMEMALLOC for
> > them. Under memory pressure, using order-0 pages is probably the best
> > strategy.
> ...
> > @@ -348,7 +352,8 @@ static struct page *__page_frag_refill(struct
> > netdev_alloc_cache *nc,
> > gfp_t gfp = gfp_mask;
> >
> > if (order) {
> > - gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY;
> > + gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
> > + __GFP_NOMEMALLOC;
> > page = alloc_pages_node(NUMA_NO_NODE, gfp_mask, order);
>
> Hmmm, skbuff.h says:
>
> * @gfp_mask: allocation priority. Set __GFP_NOMEMALLOC if not for network Rx
>
> But that's exactly what this __page_frag_refill() code is primarily
> being used for, network RX, right?
>
> If we believe the comment, you should not be adding the
> __GFP_NOMEMALLOC flag here.
I do not see a problem here.
We attempt high order pages allocations only if they bring a performance
gain. Under memory pressure, there is the high risk a whole order-3 page
is consumed by a tiny buffer using 200 bytes but not consumed.
Only high order pages allocation attempts should request this
__GFP_NOMEMALLOC flag set.
order-0 pages will have __GFP_MEMALLOC if requested by at least one
socket in the host.
About __dev_alloc_pages() comment / users, they have to know if they
need high order pages, or could fallback to order-0 ones.
We can update the comments in skbuff.h, but so far the only user has no
need to assert __GFP_NOMEMALLOC.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
