runtime: stoplockedm: inconsistent locking

[prattmic]

2021-04-16T21:50:38-9481789/freebsd-arm-paulzhol
2021-04-09T12:56:04-519f223/netbsd-amd64-9_0

fatal error: stoplockedm: inconsistent locking

runtime stack:
runtime.throw(0x5a78b0, 0x21)
        /tmp/workdir/go/src/runtime/panic.go:1193 +0x74 fp=0x7384fbbffeb0 sp=0x7384fbbffe80 pc=0x43ee94
runtime.stoplockedm()
        /tmp/workdir/go/src/runtime/proc.go:2479 +0x2b0 fp=0x7384fbbfff00 sp=0x7384fbbffeb0 pc=0x446970
runtime.exitsyscall0(0xc00019a300)
        /tmp/workdir/go/src/runtime/proc.go:3884 +0x179 fp=0x7384fbbfff30 sp=0x7384fbbfff00 pc=0x44a599
runtime.mcall(0x400000)
        /tmp/workdir/go/src/runtime/asm_amd64.s:341 +0x5a fp=0x7384fbbfff40 sp=0x7384fbbfff30 pc=0x471dfa

goroutine 1 [chan receive]:
runtime.gopark(0x5ae208, 0xc000198368, 0x170e, 0x2)
        /tmp/workdir/go/src/runtime/proc.go:334 +0x105 fp=0xc000098b48 sp=0xc000098b28 pc=0x441e65
runtime.chanrecv(0xc000198310, 0xc000098c4f, 0xc000000101, 0x4d85cb)
        /tmp/workdir/go/src/runtime/chan.go:576 +0x5f7 fp=0xc000098bd8 sp=0xc000098b48 pc=0x40d557
runtime.chanrecv1(0xc000198310, 0xc000098c4f)
        /tmp/workdir/go/src/runtime/chan.go:439 +0x2b fp=0xc000098c08 sp=0xc000098bd8 pc=0x40cf4b
testing.(*T).Run(0xc0000bec00, 0x5a1c03, 0xc, 0x5ab7b0, 0x488201)
        /tmp/workdir/go/src/testing/testing.go:1281 +0x3d2 fp=0xc000098ca0 sp=0xc000098c08 pc=0x4d85f2
testing.runTests.func1(0xc00019a000)
        /tmp/workdir/go/src/testing/testing.go:1553 +0x85 fp=0xc000098cf0 sp=0xc000098ca0 pc=0x4d9f65
testing.tRunner(0xc00019a000, 0xc00007dde0)
        /tmp/workdir/go/src/testing/testing.go:1235 +0x102 fp=0xc000098d40 sp=0xc000098cf0 pc=0x4d7862
testing.runTests(0xc000194000, 0x8ab7e0, 0x59, 0x59, 0xc0143355ed0a8eb6, 0x8bb3a3bba2, 0x8c2fa0, 0x0)
        /tmp/workdir/go/src/testing/testing.go:1551 +0x450 fp=0xc000098e10 sp=0xc000098d40 pc=0x4d9e10
testing.(*M).Run(0xc00018e000, 0x0)
        /tmp/workdir/go/src/testing/testing.go:1459 +0x25b fp=0xc000098f20 sp=0xc000098e10 pc=0x4d8abb
main.main()
        _testmain.go:225 +0x165 fp=0xc000098f88 sp=0xc000098f20 pc=0x55d3a5
runtime.main()
        /tmp/workdir/go/src/runtime/proc.go:223 +0x248 fp=0xc000098fe0 sp=0xc000098f88 pc=0x441a08
runtime.goexit()
        /tmp/workdir/go/src/runtime/asm_amd64.s:1500 +0x1 fp=0xc000098fe8 sp=0xc000098fe0 pc=0x473f81

An internal test triggered this same crash and a bisect identified http://golang.org/cl/291329 as the cause. Likely the assertion in the comment is not always true, but I'll need to look more closely.

cc @mknyszek @ianlancetaylor

[prattmic]

The problem is that earlier in exitsyscall0 we put gp on the global runq, allowing us to race reading gp.lockedm with another M running gp (which wasn't locked) and having it call lockOSThread, thus setting gp.lockedm.

CL forthcoming.

[gopherbot]

Change https://golang.org/cl/324350 mentions this issue: runtime: avoid gp.lockedm race in exitsyscall0