Sync up i386's pmap_enter() with amd64's
ClosedPublic
Actions

Authored by alc on Jul 8 2018, 11:10 PM.

Details

Reviewers

kib
markj

Commits

rS336175: Eliminate unnecessary differences between i386's pmap_enter() and amd64's.

Summary

Eliminate inconsequential differences between i386's pmap_enter() and amd64's. For example, fully construct the new PTE before entering the critical section. This change is a stepping stone to psind == 1 support.

Test Plan

Ask Peter for help. :-)

Diff Detail

Repository

rS FreeBSD src repository - subversion

Lint

Lint Not Applicable

Unit

Tests Not Applicable

Event Timeline

alc created this revision.Jul 8 2018, 11:10 PM

You can easily see the differences with diff -u amd64/amd64/pmap.c i386/i386/pmap.c.

Peter, could you please test this patch?

If nothing else. the new version is 96 bytes smaller.

Before:

# size /local/obj/freebsd/base/head/i386.i386/sys/GENERIC-NODEBUG/pmap.o
   text   data   bss     dec      hex   filename
  40523   1020   408   41951   0xa3df   /local/obj/freebsd/base/head/i386.i386/sys/GENERIC-NODEBUG/pmap.o

After:

# size /local/obj/freebsd/base/head/i386.i386/sys/GENERIC-NODEBUG/pmap.o 
   text   data   bss     dec      hex   filename
  40427   1020   408   41855   0xa37f   /local/obj/freebsd/base/head/i386.i386/sys/GENERIC-NODEBUG/pmap.o

In D16188#343237, @alc wrote:

Peter, could you please test this patch?

Sure.

kib accepted this revision.Jul 9 2018, 3:10 PM

This revision is now accepted and ready to land.Jul 9 2018, 3:10 PM

FWIW, I started working on pmap_enter(psind = 1) for arm64 in the background. Had you already started working on that?

In D16188#343550, @markj wrote:

FWIW, I started working on pmap_enter(psind = 1) for arm64 in the background. Had you already started working on that?

No, I have not. Go for it.

I ran all of the stress2 tests on i386. No problems seen.

In D16188#343550, @markj wrote:

FWIW, I started working on pmap_enter(psind = 1) for arm64 in the background. Had you already started working on that?

FYI, on amd64, over the course of a buildworld, a few 2MB regions within the code segment of clang will be promoted to superpages. Essentially, I/O clustering on page faults, fills out theses superpages even if we don't execute instructions from each of the pages. And, then, pmap_enter_object() premaps these 2MB regions as superpages on subsequent executions of clang. However, on arm64, I believe that the default layout used by the new linker is going to sabotage this, because the code segment no longer starts at a 2MB-aligned address.

Closed by commit rS336175: Eliminate unnecessary differences between i386's pmap_enter() and amd64's. (authored by alc). · Explain WhyJul 10 2018, 6:01 PM

This revision was automatically updated to reflect the committed changes.

Herald added a subscriber: imp. · View Herald TranscriptJul 10 2018, 6:01 PM