runtime: only sleep before stealing work from a running P

The sleep in question does not make sense if the stolen-from P cannot
run the stolen G. The usleep(3) has been observed delaying execution of
woken G's by ~60us; skipping it reduces the wakeup-to-execution latency
to ~7us in these cases, improving CPU utilization.

Benchmarks added by this change:

name                             old time/op  new time/op  delta
WakeupParallelSpinning/0s-12     14.4µs ± 1%  14.3µs ± 1%     ~     (p=0.227 n=19+20)
WakeupParallelSpinning/1µs-12    18.3µs ± 0%  18.3µs ± 1%     ~     (p=0.950 n=20+19)
WakeupParallelSpinning/2µs-12    22.3µs ± 1%  22.3µs ± 1%     ~     (p=0.670 n=20+18)
WakeupParallelSpinning/5µs-12    31.7µs ± 0%  31.7µs ± 0%     ~     (p=0.460 n=20+17)
WakeupParallelSpinning/10µs-12   51.8µs ± 0%  51.8µs ± 0%     ~     (p=0.883 n=20+20)
WakeupParallelSpinning/20µs-12   91.9µs ± 0%  91.9µs ± 0%     ~     (p=0.245 n=20+20)
WakeupParallelSpinning/50µs-12    214µs ± 0%   214µs ± 0%     ~     (p=0.509 n=19+20)
WakeupParallelSpinning/100µs-12   335µs ± 0%   335µs ± 0%   -0.05%  (p=0.006 n=17+15)
WakeupParallelSyscall/0s-12       228µs ± 2%   129µs ± 1%  -43.32%  (p=0.000 n=20+19)
WakeupParallelSyscall/1µs-12      232µs ± 1%   131µs ± 1%  -43.60%  (p=0.000 n=19+20)
WakeupParallelSyscall/2µs-12      236µs ± 1%   133µs ± 1%  -43.44%  (p=0.000 n=18+19)
WakeupParallelSyscall/5µs-12      248µs ± 2%   139µs ± 1%  -43.68%  (p=0.000 n=18+19)
WakeupParallelSyscall/10µs-12     263µs ± 3%   150µs ± 2%  -42.97%  (p=0.000 n=18+20)
WakeupParallelSyscall/20µs-12     281µs ± 2%   170µs ± 1%  -39.43%  (p=0.000 n=19+19)
WakeupParallelSyscall/50µs-12     345µs ± 4%   246µs ± 7%  -28.85%  (p=0.000 n=20+20)
WakeupParallelSyscall/100µs-12    460µs ± 5%   350µs ± 4%  -23.85%  (p=0.000 n=20+20)

Benchmarks associated with the change that originally added this sleep
(see https://golang.org/s/go15gomaxprocs):

name        old time/op  new time/op  delta
Chain       19.4µs ± 2%  19.3µs ± 1%    ~     (p=0.101 n=19+20)
ChainBuf    19.5µs ± 2%  19.4µs ± 2%    ~     (p=0.840 n=19+19)
Chain-2     19.9µs ± 1%  19.9µs ± 2%    ~     (p=0.734 n=19+19)
ChainBuf-2  20.0µs ± 2%  20.0µs ± 2%    ~     (p=0.175 n=19+17)
Chain-4     20.3µs ± 1%  20.1µs ± 1%  -0.62%  (p=0.010 n=19+18)
ChainBuf-4  20.3µs ± 1%  20.2µs ± 1%  -0.52%  (p=0.023 n=19+19)
Powser       2.09s ± 1%   2.10s ± 3%    ~     (p=0.908 n=19+19)
Powser-2     2.21s ± 1%   2.20s ± 1%  -0.35%  (p=0.010 n=19+18)
Powser-4     2.31s ± 2%   2.31s ± 2%    ~     (p=0.578 n=18+19)
Sieve        13.6s ± 1%   13.6s ± 1%    ~     (p=0.909 n=17+18)
Sieve-2      8.02s ±52%   7.28s ±15%    ~     (p=0.336 n=20+16)
Sieve-4      4.00s ±35%   3.98s ±26%    ~     (p=0.654 n=20+18)

Change-Id: I58edd8ce01075859d871e2348fc0833e9c01f70f
Reviewed-on: https://go-review.googlesource.com/78538
Reviewed-by: Austin Clements <austin@google.com>

src/runtime/proc.go

@@ -4773,22 +4773,25 @@ func runqgrab(_p_ *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool
 			if stealRunNextG {
 				// Try to steal from _p_.runnext.
 				if next := _p_.runnext; next != 0 {
-					// Sleep to ensure that _p_ isn't about to run the g we
-					// are about to steal.
-					// The important use case here is when the g running on _p_
-					// ready()s another g and then almost immediately blocks.
-					// Instead of stealing runnext in this window, back off
-					// to give _p_ a chance to schedule runnext. This will avoid
-					// thrashing gs between different Ps.
-					// A sync chan send/recv takes ~50ns as of time of writing,
-					// so 3us gives ~50x overshoot.
-					if GOOS != "windows" {
-						usleep(3)
-					} else {
-						// On windows system timer granularity is 1-15ms,
-						// which is way too much for this optimization.
-						// So just yield.
-						osyield()
+					if _p_.status == _Prunning {
+						// Sleep to ensure that _p_ isn't about to run the g
+						// we are about to steal.
+						// The important use case here is when the g running
+						// on _p_ ready()s another g and then almost
+						// immediately blocks. Instead of stealing runnext
+						// in this window, back off to give _p_ a chance to
+						// schedule runnext. This will avoid thrashing gs
+						// between different Ps.
+						// A sync chan send/recv takes ~50ns as of time of
+						// writing, so 3us gives ~50x overshoot.
+						if GOOS != "windows" {
+							usleep(3)
+						} else {
+							// On windows system timer granularity is
+							// 1-15ms, which is way too much for this
+							// optimization. So just yield.
+							osyield()
+						}
 					}
 					if !_p_.runnext.cas(next, 0) {
 						continue