Page MenuHomeFreeBSD

hyperv/hn: Avoid mbuf cluster allocation on reception path, if the packet is small
ClosedPublic

Authored by sepherosa_gmail.com on Jan 11 2016, 5:38 AM.
Tags
None
Referenced Files
F101886409: D4853.id.diff
Tue, Nov 5, 5:03 AM
Unknown Object (File)
Wed, Oct 16, 5:50 AM
Unknown Object (File)
Oct 1 2024, 1:30 PM
Unknown Object (File)
Sep 21 2024, 10:11 AM
Unknown Object (File)
Sep 21 2024, 4:57 AM
Unknown Object (File)
Sep 19 2024, 7:57 AM
Unknown Object (File)
Sep 19 2024, 7:52 AM
Unknown Object (File)
Sep 17 2024, 8:39 PM

Details

Summary

This gives me ~200Mbps improvement, when running iperf3 sending test w/ 16 connections.

While I'm here, nuke the unnecessary pkthdr.csum_flags zeroing out.

Sponsored by: Microsoft OSTC

Diff Detail

Repository
rS FreeBSD src repository - subversion
Lint
Lint Not Applicable
Unit
Tests Not Applicable

Event Timeline

sepherosa_gmail.com retitled this revision from to hyperv/hn: Avoid mbuf cluster allocation on reception path, if the packet is small.
sepherosa_gmail.com updated this object.
sepherosa_gmail.com edited the test plan for this revision. (Show Details)

If no objection comes, this will be committed this week.

adrian edited edge metadata.

You should explain more why the change improves performance more dramatically when you commit. other than that, it's fine!

This revision is now accepted and ready to land.Jan 11 2016, 6:48 AM

You should explain more why the change improves performance more dramatically when you commit. other than that, it's fine!

Sure, thanks.

jtl added inline comments.
sys/dev/hyperv/netvsc/hv_netvsc_drv_freebsd.c
475 โ†—(On Diff #12102)

Is there a reason this is read/write?

1036 โ†—(On Diff #12102)

Do you need an atomic operation here to guarantee accurate statistics? (And, even if you do, is it that important?)

This revision was automatically updated to reflect the committed changes.