Differential D60567

hwpmc: let MD code see every context switch and exec
AcceptedPublic
Actions

Authored by afscoelho_gmail.com on Oct 10 2026, 1:22 AM.

Project Tags

None

Referenced Files

None

Subscribers

imp

olce

Details

Reviewers

mhorne
gnn
ali_mashtizadeh.com
kib
markj
olce

Group Reviewers

pmc

Summary

Branch records do not track address spaces, so the MD code must clear
them on every context switch and exec. Add optional pmd_lbr_csw()
and pmd_lbr_exec() hooks.

While a system-mode PMC that needs them runs, call the switch hooks for
all threads. Otherwise the scheduler fast path is unchanged.

Depends on D60566

Test Plan

Built amd64 GENERIC + hwpmc.ko, GENERIC without HWPMC_HOOKS, and arm64.
Ran tests/sys/pmc on AMD Zen 6 (LBR v2).

Diff Detail

Repository

rG FreeBSD src repository

Lint

Lint Skipped

Unit

Tests Skipped

Build Status

Buildable 78142
Build 75025: arc lint + arc unit

Event Timeline

afscoelho_gmail.com created this revision.Oct 10 2026, 1:22 AM

Herald added a reviewer: olce. · View Herald TranscriptOct 10 2026, 1:22 AM

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

afscoelho_gmail.com requested review of this revision.Oct 10 2026, 1:22 AM

Harbormaster completed remote builds in B77940: Diff 189322.Oct 10 2026, 1:22 AM

afscoelho_gmail.com added a child revision: D60568: hwpmc: add the LBR allocation policy.Oct 10 2026, 1:22 AM

kib added inline comments.Oct 10 2026, 9:37 AM

sys/dev/hwpmc/hwpmc_mod.c
1781	Why do you need the critical section there? PMC_FN_CSW_OUT is always called under some spinlock: either the thread lock when done during normal context switch, or the process spinlock if called during thread exit. As is, the context switch is not possible already there. You might add CRITICAL_ASSERT() there instead, if needed, I believe.

afscoelho_gmail.com added inline comments.Oct 13 2026, 7:34 PM

sys/dev/hwpmc/hwpmc_mod.c
1781	Good catch, thanks! You're right: we're already under a spinlock there, so the critical section isn't needed. I'll replace it with CRITICAL_ASSERT() and do the same in the CSW_IN early path, which has the same problem. Thanks for taking the time to review this!