From a failing net/http test:

runtime stack:
runtime.gothrow(0x9d3ed0, 0x24)
    c:/go/src/runtime/panic.go:409 +0x8e fp=0x44bfd68 sp=0x44bfd50
runtime.badmcall(0xa469b8)
    c:/go/src/runtime/proc.go:184 +0x3b fp=0x44bfd80 sp=0x44bfd68
runtime.gopark(0x4365f0, 0x155d380, 0x97e2f0, 0xa)
    c:/go/src/runtime/proc.go:129 +0x116 fp=0x44bfdc0 sp=0x44bfd80
runtime.goparkunlock(0x155d380, 0x97e2f0, 0xa)
    c:/go/src/runtime/proc.go:135 +0x4f fp=0x44bfde8 sp=0x44bfdc0
runtime.semacquire(0x1540310, 0x27b5e100)
    c:/go/src/runtime/sema.go:84 +0x203 fp=0x44bfe20 sp=0x44bfde8
readmemstats_m()
    c:/go/src/runtime/mgc0.c:1438 +0x7b fp=0x44bfe48 sp=0x44bfe20
runtime.onM(0xc082015200)
    c:/go/src/runtime/asm_amd64.s:257 +0x6d fp=0x44bfe50 sp=0x44bfe48
runtime.mstart()
    c:/go/src/runtime/proc.c:828 fp=0x44bfe58 sp=0x44bfe50

goroutine 269 [running]:
runtime.switchtoM()
    c:/go/src/runtime/asm_amd64.s:198 fp=0xc082083ec8 sp=0xc082083ec0
runtime.ReadMemStats(0x15633e0)
    c:/go/src/runtime/mgc0.c:1423 +0x41 fp=0xc082083ee0 sp=0xc082083ec8
testing.(*B).StopTimer(0xc082064400)
    c:/go/src/testing/benchmark.go:74 +0x136 fp=0xc082083f50 sp=0xc082083ee0
testing.(*B).runN(0xc082064400, 0x2)
    c:/go/src/testing/benchmark.go:125 +0x100 fp=0xc082083f60 sp=0xc082083f50
testing.(*B).launch(0xc082064400)
    c:/go/src/testing/benchmark.go:216 +0x1d1 fp=0xc082083fd8 sp=0xc082083f60
runtime.goexit()
    c:/go/src/runtime/proc.c:1705 fp=0xc082083fe0 sp=0xc082083fd8
created by testing.(*B).run
    c:/go/src/testing/benchmark.go:179 +0x53

There is a call to ReadMemStats.  It switches to the M stack and then tries to stop the
world.  The first part of stopping the world involves acquiring the worldsema.  That
calls semacquire, and that calls goparkunlock.  We can't goparkunlock from the M stack -
we need a clean G stack to suspend the current execution.

gopark correctly asserts that it can't run from the M stack.  We need to convert all
calls to semacquire to come from the G stack.

This currently happens in the following 3 routines:
gomaxprocs_m
readmemstats_m
writeheapdump_m

So I think we have to move the semacquire in these routines earlier, to the other side
of the G/M switch.

Then we need to assert at the start of semacquire that we are still on the G stack.