This little number makes all of the flow dissection functions take raw input data pointer as const (1-5) and shuffles the branches in __skb_header_pointer() according to their hit probability. The result is +20 Mbps per flow/core with one Flow Dissector pass per packet. This affects RPS (with software hashing), drivers that use eth_get_headlen() on their Rx path and so on. >From v2 [1]: - reword some commit messages as a potential fix for NIPA; - no functional changes. >From v1 [0]: - rebase on top of the latest net-next. This was super-weird, but I double-checked that the series applies with no conflicts, and then on Patchwork it didn't; - no other changes. [0] https://lore.kernel.org/netdev/20210312194538.337504-1-alobakin@xxxxx [1] https://lore.kernel.org/netdev/20210313113645.5949-1-alobakin@xxxxx Alexander Lobakin (6): flow_dissector: constify bpf_flow_dissector's data pointers skbuff: make __skb_header_pointer()'s data argument const flow_dissector: constify raw input data argument linux/etherdevice.h: misc trailing whitespace cleanup ethernet: constify eth_get_headlen()'s data argument skbuff: micro-optimize {,__}skb_header_pointer() include/linux/etherdevice.h | 4 ++-- include/linux/skbuff.h | 26 +++++++++++------------ include/net/flow_dissector.h | 6 +++--- net/core/flow_dissector.c | 41 +++++++++++++++++++----------------- net/ethernet/eth.c | 2 +- 5 files changed, 40 insertions(+), 39 deletions(-) -- 2.30.2
