net: more accurate skb truesize
authorEric Dumazet <eric.dumazet@gmail.com>
Thu, 13 Oct 2011 07:28:54 +0000 (07:28 +0000)
committerDavid S. Miller <davem@davemloft.net>
Thu, 13 Oct 2011 20:05:07 +0000 (16:05 -0400)
commit87fb4b7b533073eeeaed0b6bf7c2328995f6c075
treebe4b37f08d7fe2d018ae68bae4577b1b2bafd0fc
parent97ba0eb64ca690165f945a83e609102fcefa99cb
net: more accurate skb truesize

skb truesize currently accounts for sk_buff struct and part of skb head.
kmalloc() roundings are also ignored.

Considering that skb_shared_info is larger than sk_buff, its time to
take it into account for better memory accounting.

This patch introduces SKB_TRUESIZE(X) macro to centralize various
assumptions into a single place.

At skb alloc phase, we put skb_shared_info struct at the exact end of
skb head, to allow a better use of memory (lowering number of
reallocations), since kmalloc() gives us power-of-two memory blocks.

Unless SLUB/SLUB debug is active, both skb->head and skb_shared_info are
aligned to cache lines, as before.

Note: This patch might trigger performance regressions because of
misconfigured protocol stacks, hitting per socket or global memory
limits that were previously not reached. But its a necessary step for a
more accurate memory accounting.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Andi Kleen <ak@linux.intel.com>
CC: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/skbuff.h
net/core/skbuff.c
net/core/sock.c
net/ipv4/icmp.c
net/ipv4/tcp_input.c
net/ipv6/icmp.c
net/iucv/af_iucv.c
net/sctp/protocol.c