]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commit
net: add skb_data_unref() helper
authorEric Dumazet <edumazet@google.com>
Thu, 7 Mar 2024 12:34:46 +0000 (12:34 +0000)
committerJakub Kicinski <kuba@kernel.org>
Fri, 8 Mar 2024 19:38:45 +0000 (11:38 -0800)
commit1cface552a5b5f6e53a855de1a503ff958e2e253
treeb645765112af7b170abdac469f17ca0cf2955135
parent75c2946db360e625f1447a37f47dbbb38b1dd478
net: add skb_data_unref() helper

Similar to skb_unref(), add skb_data_unref() to save an expensive
atomic operation (and cache line dirtying) when last reference
on shinfo->dataref is released.

I saw this opportunity on hosts with RAW sockets accidentally
bound to UDP protocol, forcing an skb_clone() on all received packets.

These RAW sockets had their receive queue full, so all clone
packets were immediately dropped.

When UDP recvmsg() consumes later the original skb, skb_release_data()
is hitting atomic_sub_return() quite badly, because skb->clone
has been set permanently.

Note that this patch helps TCP TX performance, because
TCP stack also use (fast) clones.

This means that at least one of the two packets (the main skb or
its clone) will no longer have to perform this atomic operation
in skb_release_data().

Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240307123446.2302230-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
include/linux/skbuff.h
net/core/skbuff.c