arm64: Elide coherent busdma maps
Needs ReviewPublic
Actions

Authored by gallatin on Mon, Sep 28, 5:44 PM.

Details

Reviewers

andrew
jhb
markj
manu

Summary

Avoid allocating per-transfer maps for coherent tags that cannot
bounce. Retain maps for cache synchronization, CCA realms, and KMSAN.
These un-used maps carry with them memory and cache miss overheads.

This saves close to 1% CPU on my tiny N1 setup serving ~80Gb/s of Netflix
traffic.

Diff Detail

Repository

rG FreeBSD src repository

Lint

Lint Skipped

Unit

Tests Skipped

Event Timeline

gallatin created this revision.Mon, Sep 28, 5:44 PM

Herald added a reviewer: manu. · View Herald TranscriptMon, Sep 28, 5:44 PM

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

gallatin requested review of this revision.Mon, Sep 28, 5:44 PM

gallatin added a child revision: D60099: arm64: Use only required ordering for coherent dma.Mon, Sep 28, 5:48 PM

jhb added inline comments.Mon, Sep 28, 6:16 PM

FreeBSD/sys/arm64/arm64/busdma_bounce.c
980	Can this be called with a NULL map? Hmm, I guess it can. FWIW, x86 handles this in the header file so the check gets inlined sooner in sys/kern/subr_busdma_bounce.c. That is, x86 checks it in sys/x86/include/bus_dma.h:_bus_dmamap_waitok(). Presumably could do the same here?

handled null maps for _bus_dmamap_waitok() and bus_dmamap_unload() in the linline wrappers as suggested by @jhb

gallatin marked an inline comment as done.Mon, Sep 28, 9:25 PM

gallatin added inline comments.

FreeBSD/sys/arm64/arm64/busdma_bounce.c
980	Great idea.. Just implemented it. Thank you!

Revision Contents
Changeset List

Path

Size

FreeBSD/

sys/

arm64/

busdma_bounce.c

82 lines

include/

bus_dma.h

12 lines

Diff 187929

View Options

FreeBSD/sys/arm64/arm64/busdma_bounce.c