tcp: For hostcache performance, use atomics instead of counters
ClosedPublic
Actions

Authored by rscheff on Mar 31 2021, 10:29 PM.

Details

Reviewers

jtl
tuexen
jhb

Group Reviewers

transport

Commits

rG254b56ac2a9e: tcp: For hostcache performance, use atomics instead of counters
rGe9ca38abc79b: tcp: For hostcache performance, use atomics instead of counters
rG4052275c2fc5: tcp: For hostcache performance, use atomics instead of counters
rG529a2a0f2765: tcp: For hostcache performance, use atomics instead of counters

Summary

As accessing the tcp hostcache happens frequently on some
classes of servers, it was recommended to use atomic_add/subtract
rather than (per-CPU distributed) counters, which have to be
summed up at high cost to cache efficiency.

This eliminates a counter_u64_fetch() from a hot path, introduced
by D29510.

PR: 254333
MFC after: 2 weeks
Sponsored by: NetApp, Inc.

Diff Detail

Repository

rG FreeBSD src repository

Lint

Lint Not Applicable

Unit

Tests Not Applicable

Event Timeline

rscheff created this revision.Mar 31 2021, 10:29 PM

Herald added 1 blocking reviewer(s): transport. · View Herald TranscriptMar 31 2021, 10:29 PM

Herald added subscribers: melifaro, imp. · View Herald Transcript

Harbormaster completed remote builds in B38228: Diff 86647.Mar 31 2021, 10:29 PM

rscheff requested review of this revision.Mar 31 2021, 10:29 PM

tuexen accepted this revision.Mar 31 2021, 10:50 PM

This revision is now accepted and ready to land.Mar 31 2021, 10:50 PM

Thanks for doing this so quickly!

sys/netinet/tcp_hostcache.c
377	I think you need an atomic_load_int() here.
643–644	I think you need an atomic_load_int() here, although I don't think it functionally probably matters too much. In fact, this code already allows races between when the sysctl function is called to get the length, and when the code is called with another buffer. But, it seems like it is probably good to be consistent in using the atomic* functions to access or set atomic variables.