Use invlpg instead of total invalidation, when setting PG_G.
ClosedPublic
Actions

Authored by kib on Dec 4 2015, 8:35 AM.

Details

Reviewers

alc

Commits

Summary

Continuation of the D4346.

Repository

rS FreeBSD src repository - subversion

Lint

Lint Skipped

Unit

Tests Skipped

kib retitled this revision from to Use invlpg instead of total invalidation, when setting PG_G..

kib updated this object.

kib edited the test plan for this revision. (Show Details)

kib added a reviewer: alc.

kib set the repository for this revision to rS FreeBSD src repository - subversion.

kib added subscribers: jhb, cem.

cem added inline comments.Dec 4 2015, 4:36 PM

sys/i386/i386/pmap.c
660–661	Here we'll invalidate 512-1024 pages; is that really cheaper than an `invltlb_glob()`?

jhb added inline comments.Dec 4 2015, 5:38 PM

sys/i386/i386/pmap.c
641	Note that this isn't a critical path, it's only called once during boot, so I'm not sure it's worth optimizing.

cem added inline comments.Dec 4 2015, 6:02 PM

sys/i386/i386/pmap.c
641	Might as well boot as fast as possible.

alc added inline comments.Dec 4 2015, 6:03 PM

sys/i386/i386/pmap.c
641	My issue with this isn't performance, but the logical inconsistency of issuing a PG_G flush when there are no PG_G mappings to flush. Except the ones created on the previous iteration.
660–661	No, not really, and there is little state in the TLB worth preserving. I'll suggest this line: invltlb(); /* Flush non-PG_G entries. */