commit 2f91c16e68b8c82e4ca2e422bf9a520ec46f817c
author Michael Pratt <mpratt@google.com> Wed Dec 13 15:07:28 2023 -0500
committer Cherry Mui <cherryyz@google.com> Thu Jan 25 17:20:02 2024 +0000
tree 7a1205483b7d1684c50d9edecfff0fa3d477f91e
parent 2540b1436ff452529b1668a8310411ddea826c52

[release-branch.go1.21] runtime: properly model rwmutex in lock ranking

(This cherry-pick combines CL 549536 and the follow-up fix CL 555055.)

Currently, lock ranking doesn't really try to model rwmutex. It records
the internal locks rLock and wLock, but in a subpar fashion:

1. wLock is held from lock to unlock, so it works OK, but it conflates
   write locks of all rwmutexes as rwmutexW, rather than allowing
   different rwmutexes to have different rankings.
2. rLock is an internal implementation detail that is only taken when
   there is contention in rlock. As as result, the reader lock path is
   almost never checked.

Add proper modeling. rwmutexR and rwmutexW remain as the ranks of the
internal locks, which have their own ordering. The new init method is
passed the ranks of the higher level lock that this represents, just
like lockInit for mutex.

execW ordered before MALLOC captures the case from #64722. i.e., there
can be allocation between BeforeFork and AfterFork.

For #64722.
Fixes #64761.

------

runtime: replace rwmutexR/W with per-rwmutex lock rank

CL 549536 intended to decouple the internal implementation of rwmutex
from the semantic meaning of an rwmutex read/write lock in the static
lock ranking.

Unfortunately, it was not thought through well enough. The internals
were represented with the rwmutexR and rwmutexW lock ranks. The idea was
that the internal lock ranks need not model the higher-level ordering,
since those have separate rankings. That is incorrect; rwmutexW is held
for the duration of a write lock, so it must be ranked before any lock
taken while any write lock is held, which is precisely what we were
trying to avoid.

This is visible in violations like:

        0 : execW 11 0x0
        1 : rwmutexW 51 0x111d9c8
        2 : fin 30 0x111d3a0
        fatal error: lock ordering problem

execW < fin is modeled, but rwmutexW < fin is missing.

Fix this by eliminating the rwmutexR/W lock ranks shared across
different types of rwmutex. Instead require users to define an
additional "internal" lock rank to represent the implementation details
of rwmutex.rLock. We can avoid an additional "internal" lock rank for
rwmutex.wLock because the existing writeRank has the same semantics for
semantic and internal locking. i.e., writeRank is held for the duration
of a write lock, which is exactly how rwmutex.wLock is used, so we can
use writeRank directly on wLock.

For #64722.

Cq-Include-Trybots: luci.golang.try:go1.21-linux-amd64-staticlockranking
Change-Id: I23335b28faa42fb04f1bc9da02fdf54d1616cd28
Reviewed-on: https://go-review.googlesource.com/c/go/+/549536
Reviewed-by: Michael Knyszek <mknyszek@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
(cherry picked from commit 9b4b3e5acca2dabe107fa2c3ed963097d78a4562)
(cherry picked from commit dcbe77246922fe7ef41f07df228f47a37803f360)
Reviewed-on: https://go-review.googlesource.com/c/go/+/554976