runtime: hold sched.lock across atomic pidleget/pidleput

As a cleanup, golang.org/cl/307914 unintentionally caused the idle GC
work recheck to drop sched.lock between acquiring a P and committing to
keep it (once a worker G was found).

This is unsafe, as releasing a P requires extra checks once sched.lock
is taken (such as for runSafePointFn). Since checkIdleGCNoP does not
perform these extra checks, we can now race with other users.

In the case of #45975, we may hang with this sequence:

1. M1: checkIdleGCNoP takes sched.lock, gets P1, releases sched.lock.
2. M2: forEachP takes sched.lock, iterates over sched.pidle without
   finding P1, releases sched.lock.
3. M1: checkIdleGCNoP puts P1 back in sched.pidle.
4. M2: forEachP waits forever for P1 to run the safePointFn.

Change back to the old behavior of releasing sched.lock only after we
are certain we will keep the P. Thus if we put it back its removal from
sched.pidle was never visible.

Fixes #45975
For #45916
For #45885
For #45884

Change-Id: I191a1800923b206ccaf96bdcdd0bfdad17b532e9
Reviewed-on: https://go-review.googlesource.com/c/go/+/318569
Trust: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>

1 files changed, 21 insertions, 11 deletions

diff --git a/src/runtime/proc.go b/src/runtime/proc.go
index ba02b14995..378d5e32f5 100644
--- a/src/runtime/proc.go
+++ b/src/runtime/proc.go
@@ -3132,24 +3132,33 @@ func checkIdleGCNoP() (*p, *g) {
 		return nil, nil
 	}
 
-	// Work is available; we can start an idle GC worker only if
-	// there is an available P and available worker G.
+	// Work is available; we can start an idle GC worker only if there is
+	// an available P and available worker G.
 	//
-	// We can attempt to acquire these in either order. Workers are
-	// almost always available (see comment in findRunnableGCWorker
-	// for the one case there may be none). Since we're slightly
-	// less likely to find a P, check for that first.
+	// We can attempt to acquire these in either order, though both have
+	// synchonization concerns (see below). Workers are almost always
+	// available (see comment in findRunnableGCWorker for the one case
+	// there may be none). Since we're slightly less likely to find a P,
+	// check for that first.
+	//
+	// Synchronization: note that we must hold sched.lock until we are
+	// committed to keeping it. Otherwise we cannot put the unnecessary P
+	// back in sched.pidle without performing the full set of idle
+	// transition checks.
+	//
+	// If we were to check gcBgMarkWorkerPool first, we must somehow handle
+	// the assumption in gcControllerState.findRunnableGCWorker that an
+	// empty gcBgMarkWorkerPool is only possible if gcMarkDone is running.
 	lock(&sched.lock)
 	pp := pidleget()
-	unlock(&sched.lock)
 	if pp == nil {
+		unlock(&sched.lock)
 		return nil, nil
 	}
 
-	// Now that we own a P, gcBlackenEnabled can't change
-	// (as it requires STW).
+	// Now that we own a P, gcBlackenEnabled can't change (as it requires
+	// STW).
 	if gcBlackenEnabled == 0 {
-		lock(&sched.lock)
 		pidleput(pp)
 		unlock(&sched.lock)
 		return nil, nil
@@ -3157,12 +3166,13 @@ func checkIdleGCNoP() (*p, *g) {
 
 	node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
 	if node == nil {
-		lock(&sched.lock)
 		pidleput(pp)
 		unlock(&sched.lock)
 		return nil, nil
 	}
 
+	unlock(&sched.lock)
+
 	return pp, node.gp.ptr()
 }