go/test/recover4.go
Michael Anthony Knyszek 95d0657670 test/recover4.go: use mprotect to create a hole instead of munmap
Currently the recover4 test, which recovers from a panic created from a
fault, generates a fault by creating a hole in a mapping. It does this
via munmap. However, it's possible the runtime can create a new mapping
that ends up in that hole, for example if the GC executes, causing the
test to fail.

In fact, this is the case now with a smaller minimum heap size.

Modify the test to use mprotect, and clean up the code a little while
we're here: define everything in terms of the length of original
mapping, deduplicate some constants and expressions, and have the test
recover properly even if recover() returns nil (right now it panics
because it fails to type assert nil as error).

Fixes #49381.

Change-Id: If399eca564466e5e8aeb2dc6f86a246d0fce7b5d
Reviewed-on: https://go-review.googlesource.com/c/go/+/363534
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cherry Mui <cherryyz@google.com>
2021-11-12 16:58:34 +00:00

75 lines
2.2 KiB
Go

// +build linux darwin
// run
// Copyright 2015 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// Test that if a slice access causes a fault, a deferred func
// sees the most recent value of the variables it accesses.
// This is true today; the role of the test is to ensure it stays true.
//
// In the test, memcopy is the function that will fault, during dst[i] = src[i].
// The deferred func recovers from the error and returns, making memcopy
// return the current value of n. If n is not being flushed to memory
// after each modification, the result will be a stale value of n.
//
// The test is set up by mmapping a 64 kB block of memory and then
// unmapping a 16 kB hole in the middle of it. Running memcopy
// on the resulting slice will fault when it reaches the hole.
package main
import (
"log"
"runtime/debug"
"syscall"
)
func memcopy(dst, src []byte) (n int, err error) {
defer func() {
if r, ok := recover().(error); ok {
err = r
}
}()
for i := 0; i < len(dst) && i < len(src); i++ {
dst[i] = src[i]
n++
}
return
}
func main() {
// Turn the eventual fault into a panic, not a program crash,
// so that memcopy can recover.
debug.SetPanicOnFault(true)
size := syscall.Getpagesize()
// Map 16 pages of data with a 4-page hole in the middle.
data, err := syscall.Mmap(-1, 0, 16*size, syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_ANON|syscall.MAP_PRIVATE)
if err != nil {
log.Fatalf("mmap: %v", err)
}
// Create a hole in the mapping that's PROT_NONE.
// Note that we can't use munmap here because the Go runtime
// could create a mapping that ends up in this hole otherwise,
// invalidating the test.
hole := data[len(data)/2 : 3*(len(data)/4)]
if err := syscall.Mprotect(hole, syscall.PROT_NONE); err != nil {
log.Fatalf("mprotect: %v", err)
}
// Check that memcopy returns the actual amount copied
// before the fault.
const offset = 5
n, err := memcopy(data[offset:], make([]byte, len(data)))
if err == nil {
log.Fatal("no error from memcopy across memory hole")
}
if expect := len(data)/2 - offset; n != expect {
log.Fatalf("memcopy returned %d, want %d", n, expect)
}
}