runtime: replace assist sleep loop with park/ready
GC assists must block until the assist can be satisfied (either
through stealing credit or doing work) or the GC cycle ends.
Currently, this is implemented as a retry loop with a 100 µs delay.
This obviously isn't ideal, as it wastes CPU and delays mutator
execution. It also has the somewhat peculiar downside that sleeping a
G requires allocation, and this requires working around recursive
allocation.
Replace this timed delay with a proper scheduling queue. When an
assist can't be satisfied immediately, it adds the allocating G to a
queue and parks it. Any time background scan credit is flushed, it
consults this queue, directly satisfies the debt of queued assists,
and wakes up satisfied assists before flushing any remaining credit to
the background credit pool.
No effect on the go1 benchmarks. Slightly speeds up the garbage
benchmark.
name old time/op new time/op delta
XBenchGarbage-12 5.81ms ± 1% 5.72ms ± 4% -1.65% (p=0.011 n=20+20)
Updates #12041.
Change-Id: I8ee3b6274dd097b12b10a8030796a958a4b0e7b7
Reviewed-on: https://go-review.googlesource.com/15890
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
diff --git a/src/runtime/mgcmark.go b/src/runtime/mgcmark.go
index 7603085..ab1af21 100644
--- a/src/runtime/mgcmark.go
+++ b/src/runtime/mgcmark.go
@@ -364,7 +364,7 @@
// stack to determine if we should preform an assist.
// GC is done, so ignore any remaining debt.
- scanWork = 0
+ gp.gcAssistBytes = 0
return
}
// Track time spent in this assist. Since we're on the
@@ -389,7 +389,7 @@
}
// Record that we did this much scan work.
- scanWork -= workDone
+ //
// Back out the number of bytes of assist credit that
// this scan work counts for. The "1+" is a poor man's
// round-up, to ensure this adds credit even if
@@ -432,33 +432,135 @@
// We called complete() above, so we should yield to
// the now-runnable GC coordinator.
Gosched()
-
- // It's likely that this assist wasn't able to pay off
- // its debt, but it's also likely that the Gosched let
- // the GC finish this cycle and there's no point in
- // waiting. If the GC finished, skip the delay below.
- if atomicload(&gcBlackenEnabled) == 0 {
- scanWork = 0
- }
}
- if scanWork > 0 {
+ if gp.gcAssistBytes < 0 {
// We were unable steal enough credit or perform
// enough work to pay off the assist debt. We need to
// do one of these before letting the mutator allocate
- // more, so go around again after performing an
- // interruptible sleep for 100 us (the same as the
- // getfull barrier) to let other mutators run.
+ // more to prevent over-allocation.
+ //
+ // Add this G to an assist queue and park. When the GC
+ // has more background credit, it will satisfy queued
+ // assists before flushing to the global credit pool.
+ //
+ // Note that this does *not* get woken up when more
+ // work is added to the work list. The theory is that
+ // there wasn't enough work to do anyway, so we might
+ // as well let background marking take care of the
+ // work that is available.
+ lock(&work.assistQueue.lock)
- // timeSleep may allocate, so avoid recursive assist.
- gcAssistBytes := gp.gcAssistBytes
- gp.gcAssistBytes = int64(^uint64(0) >> 1)
- timeSleep(100 * 1000)
- gp.gcAssistBytes = gcAssistBytes
- goto retry
+ // If the GC cycle is over, just return. This is the
+ // likely path if we called Gosched above. We do this
+ // under the lock to prevent a GC cycle from ending
+ // between this check and queuing the assist.
+ if atomicload(&gcBlackenEnabled) == 0 {
+ unlock(&work.assistQueue.lock)
+ return
+ }
+
+ oldHead, oldTail := work.assistQueue.head, work.assistQueue.tail
+ if oldHead == 0 {
+ work.assistQueue.head.set(gp)
+ } else {
+ oldTail.ptr().schedlink.set(gp)
+ }
+ work.assistQueue.tail.set(gp)
+ gp.schedlink.set(nil)
+ // Recheck for background credit now that this G is in
+ // the queue, but can still back out. This avoids a
+ // race in case background marking has flushed more
+ // credit since we checked above.
+ if atomicloadint64(&gcController.bgScanCredit) > 0 {
+ work.assistQueue.head = oldHead
+ work.assistQueue.tail = oldTail
+ if oldTail != 0 {
+ oldTail.ptr().schedlink.set(nil)
+ }
+ unlock(&work.assistQueue.lock)
+ goto retry
+ }
+ // Park for real.
+ goparkunlock(&work.assistQueue.lock, "GC assist", traceEvGoBlock, 2)
+
+ // At this point either background GC has satisfied
+ // this G's assist debt, or the GC cycle is over.
}
}
+// gcWakeAllAssists wakes all currently blocked assists. This is used
+// at the end of a GC cycle.
+func gcWakeAllAssists() {
+ lock(&work.assistQueue.lock)
+ injectglist(work.assistQueue.head.ptr())
+ work.assistQueue.head.set(nil)
+ work.assistQueue.tail.set(nil)
+ unlock(&work.assistQueue.lock)
+}
+
+// gcFlushBgCredit flushes scanWork units of background scan work
+// credit. This first satisfies blocked assists on the
+// work.assistQueue and then flushes any remaining credit to
+// gcController.bgScanCredit.
+func gcFlushBgCredit(scanWork int64) {
+ if work.assistQueue.head == 0 {
+ // Fast path; there are no blocked assists. There's a
+ // small window here where an assist may add itself to
+ // the blocked queue and park. If that happens, we'll
+ // just get it on the next flush.
+ xaddint64(&gcController.bgScanCredit, scanWork)
+ return
+ }
+
+ scanBytes := int64(float64(scanWork) * gcController.assistBytesPerWork)
+
+ lock(&work.assistQueue.lock)
+ gp := work.assistQueue.head.ptr()
+ for gp != nil && scanBytes > 0 {
+ // Note that gp.gcAssistBytes is negative because gp
+ // is in debt. Think carefully about the signs below.
+ if scanBytes+gp.gcAssistBytes >= 0 {
+ // Satisfy this entire assist debt.
+ scanBytes += gp.gcAssistBytes
+ gp.gcAssistBytes = 0
+ xgp := gp
+ gp = gp.schedlink.ptr()
+ ready(xgp, 0)
+ } else {
+ // Partially satisfy this assist.
+ gp.gcAssistBytes += scanBytes
+ scanBytes = 0
+ // As a heuristic, we move this assist to the
+ // back of the queue so that large assists
+ // can't clog up the assist queue and
+ // substantially delay small assists.
+ xgp := gp
+ gp = gp.schedlink.ptr()
+ if gp == nil {
+ // gp is the only assist in the queue.
+ gp = xgp
+ } else {
+ xgp.schedlink = 0
+ work.assistQueue.tail.ptr().schedlink.set(xgp)
+ work.assistQueue.tail.set(xgp)
+ }
+ break
+ }
+ }
+ work.assistQueue.head.set(gp)
+ if gp == nil {
+ work.assistQueue.tail.set(nil)
+ }
+
+ if scanBytes > 0 {
+ // Convert from scan bytes back to work.
+ scanWork = int64(float64(scanBytes) * gcController.assistWorkPerByte)
+ xaddint64(&gcController.bgScanCredit, scanWork)
+ }
+ unlock(&work.assistQueue.lock)
+}
+
//go:nowritebarrier
func scanstack(gp *g) {
if gp.gcscanvalid {
@@ -725,7 +827,7 @@
if gcw.scanWork >= gcCreditSlack {
xaddint64(&gcController.scanWork, gcw.scanWork)
if flushBgCredit {
- xaddint64(&gcController.bgScanCredit, gcw.scanWork-initScanWork)
+ gcFlushBgCredit(gcw.scanWork - initScanWork)
initScanWork = 0
}
gcw.scanWork = 0
@@ -736,7 +838,7 @@
if gcw.scanWork > 0 {
xaddint64(&gcController.scanWork, gcw.scanWork)
if flushBgCredit {
- xaddint64(&gcController.bgScanCredit, gcw.scanWork-initScanWork)
+ gcFlushBgCredit(gcw.scanWork - initScanWork)
}
gcw.scanWork = 0
}
