runtime: fatal error: unaligned sysUnused on Linux/PPC64LE

[cherrymui]

What version of Go are you using (go version)?

tip (33dfd35)

Does this issue reproduce with the latest release?

No

What operating system and processor architecture are you using (go env)?

Linux/PPC64LE

What did you do?

https://build.golang.org/log/fb55a05d088fb47d6fd65e7a10a685dea7aea1fb

fatal error: unaligned sysUnused

runtime stack:
runtime.throw(0x344d85, 0x13)
	/workdir/go/src/runtime/panic.go:1106 +0x5c
runtime.sysUnused(0xc00100a000, 0x66000)
	/workdir/go/src/runtime/mem_linux.go:102 +0x220
runtime.(*pageAlloc).scavengeRangeLocked(0x5a7a68, 0x30004, 0x5, 0x33)
	/workdir/go/src/runtime/mgcscavenge.go:488 +0xac
runtime.(*pageAlloc).scavengeOne(0x5a7a68, 0x66000, 0x30001, 0xf0000)
	/workdir/go/src/runtime/mgcscavenge.go:412 +0x39c
runtime.(*pageAlloc).scavenge(0x5a7a68, 0x3e6000, 0x400001, 0x5a7a68)
	/workdir/go/src/runtime/mgcscavenge.go:330 +0x5c
runtime.(*mheap).grow(0x5a7a60, 0x179, 0x0)
	/workdir/go/src/runtime/mheap.go:1188 +0xf0
runtime.(*mheap).allocSpanLocked(0x5a7a60, 0x179, 0x5c5648, 0x1c)
	/workdir/go/src/runtime/mheap.go:1093 +0x1ac
runtime.(*mheap).alloc_m(0x5a7a60, 0x179, 0x30100, 0x48a60)
	/workdir/go/src/runtime/mheap.go:881 +0x114
runtime.(*mheap).alloc.func1()
	/workdir/go/src/runtime/mheap.go:969 +0x4c
runtime.systemstack(0xc000071ee8)
	/workdir/go/src/runtime/asm_ppc64x.s:295 +0xd4
runtime.(*mheap).alloc(0x5a7a60, 0x179, 0x73b5d3010100, 0x73b5d3ac6650)
	/workdir/go/src/runtime/mheap.go:968 +0x7c
runtime.largeAlloc(0x2f0800, 0x688a500000001, 0x73b5d3ad8fc0)
	/workdir/go/src/runtime/malloc.go:1141 +0x8c
runtime.mallocgc.func1()
	/workdir/go/src/runtime/malloc.go:1036 +0x48
runtime.systemstack(0x45000)
	/workdir/go/src/runtime/asm_ppc64x.s:269 +0x94
runtime.mstart()
	/workdir/go/src/runtime/proc.go:1069

I also ran into this on Linux/PPC64LE gomote while testing async preemption.

cc @mknyszek

[bradfitz]

Another:
https://build.golang.org/log/fc44897bdb508f9fafcec73b390d425e8b57de01

[mknyszek]

Looking into it. I expect this to be fairly easy to diagnose and fix.

[mknyszek]

OK so the error is definitely correct, the problem is findScavengeCandidate is returning an unaligned base + length. I verified the physical page size on this builder is 64 KiB, so this indicates a broader problem for systems which have larger page sizes. Despite all the tests something is not quite right.

However, I can probably figure out what's wrong just with some diagnostics on crash, and this seems to be reproducing relatively easily. I'm going to dig deeper now.

[mknyszek]

Unfortunately I've been unable to reproduce so far running all.bash in a gomote...

But, I've found the problem. The max/min logic in findScavengeCandidate is slightly wrong.

Fixed, and added a test. Will upload shortly.

[gopherbot]

Change https://golang.org/cl/206277 mentions this issue: runtime: fix min/max logic in findScavengeCandidate

[mengzhuo]

Occasionally on Linux/mips64le too:
https://build.golang.org/log/a4f5091f520457351400a459789de45543522fa3

[ceseo]

@laboger FYI

[laboger]

I'm sure this is because of CLs 201765 and 195700 because we never saw this error before that. I believe CL 206277 fixes it on ppc64le.

[mknyszek]

@laboger @mengzhuo The problem occurs on machines with a physical page size > 8 KiB. It should now be fixed on all such platforms.