Fix incorrect header prediction flags documentation. Don't use output calculated tp->tcp_header_len for input decisions. While the output header is usually the same as the input (same options in both directions), that's a poor assumption. In particular, Sack will be different. Newer options are not guaranteed. Moreover, in the fast path, that only saved a shift or two. The other efficiencies in this patch more than make up the difference. Instead, use tp->rx_opt.tstamp_ok to accurately predict header length. Likewise, use tp->rx_opt.tstamp_ok for received MSS calculations. Don't use "sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED" to guess that the timestamp is present. This may have been OK in the days with fewer possible options, but various combinations of newer options may yield the same header length. Instead, use tp->rx_opt.saw_tstamp to determine a timestamp is present. (This bug is in 3 places.) There's no need to test buffer length against header length, already checked by tcp_v[4,6]_rcv(). Straighten code for minor efficiency gain. Don't overload tp->tcp_header_len by temporarily storing predicted header length. It's wiped later by the output code, and very confusing. Instead, the simpler __tcp_fast_path_header_length() can also be used for the same purpose, at the cost of a branch, but the substantial savings of removing the extra store and re-load. Finally, don't use TCPOLEN_MD5SIG_ALIGNED in the tcp_minisock.c last_seg_size calculations. It's not used in all other places, and may result in a too short estimate. Instead, use the same calculations as tp->advmss in tcp_input.c. Stand-alone patch, originally developed for TCPCT. Requires: net: tcp_header_len_th and tcp_option_len_th tcp: harmonize tcp_vx_rcv header length assumptions Signed-off-by: William.Allen.Simpson@gmail.com --- include/linux/tcp.h | 6 ++- include/net/tcp.h | 9 ++++- net/ipv4/tcp_input.c | 92 ++++++++++++++++------------------------------ net/ipv4/tcp_minisocks.c | 10 ++--- 4 files changed, 49 insertions(+), 68 deletions(-)