This change makes it so that worldsema isn't held across the mark phase.
This means that various operations like ReadMemStats may now stop the
world during the mark phase, reducing latency on such operations.
Only three such operations are still no longer allowed to occur during
marking: GOMAXPROCS, StartTrace, and StopTrace.
For the former it's because any change to GOMAXPROCS impacts GC mark
background worker scheduling and the details there are tricky.
For the latter two it's because tracing needs to observe consistent GC
start and GC end events, and if StartTrace or StopTrace may stop the
world during marking, then it's possible for it to see a GC end event
without a start or GC start event without an end, respectively.
To ensure that GOMAXPROCS and StartTrace/StopTrace cannot proceed until
marking is complete, the runtime now holds a new semaphore, gcsema,
across the mark phase just like it used to with worldsema.
This change is being landed once more after being reverted in the Go
1.14 release cycle, since CL 215157 allows it to have a positive
effect on system performance.
For the benchmark BenchmarkReadMemStatsLatency in the runtime, which
measures ReadMemStats latencies while the GC is exercised, the tail of
these latencies reduced dramatically on an 8-core machine:
name old 50%tile-ns new 50%tile-ns delta
ReadMemStatsLatency-8 4.40M ±74% 0.12M ± 2% -97.35% (p=0.008 n=5+5)
name old 90%tile-ns new 90%tile-ns delta
ReadMemStatsLatency-8 102M ± 6% 0M ±14% -99.79% (p=0.008 n=5+5)
name old 99%tile-ns new 99%tile-ns delta
ReadMemStatsLatency-8 147M ±18% 4M ±57% -97.43% (p=0.008 n=5+5)