runtime/debug: SetCrashOutput sets the FD for fatal panics

This feature makes it possible to record unhandled panics
in any goroutine through a watchdog process (e.g. the same
application forked+exec'd as a child in a special mode)
that can process the panic report, for example by sending
it to a crash-reporting system such as Go telemetry
or Sentry.

Fixes #42888

Change-Id: I5aa7be8f726bbc70fc650540bd1a14ab60c62ecb
Reviewed-on: https://go-review.googlesource.com/c/go/+/547978
Reviewed-by: Michael Pratt <mpratt@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Auto-Submit: Alan Donovan <adonovan@google.com>
Reviewed-by: Russ Cox <rsc@golang.org>

1 files changed, 53 insertions, 0 deletions

diff --git a/src/runtime/debug/stack.go b/src/runtime/debug/stack.go
index 3999840d3c..508afe1f97 100644
--- a/src/runtime/debug/stack.go
+++ b/src/runtime/debug/stack.go
@@ -7,8 +7,10 @@
 package debug
 
 import (
+	"internal/poll"
 	"os"
 	"runtime"
+	_ "unsafe" // for linkname
 )
 
 // PrintStack prints to standard error the stack trace returned by runtime.Stack.
@@ -28,3 +30,54 @@ func Stack() []byte {
 		buf = make([]byte, 2*len(buf))
 	}
 }
+
+// SetCrashOutput configures a single additional file where unhandled
+// panics and other fatal errors are printed, in addition to standard error.
+// There is only one additional file: calling SetCrashOutput again
+// overrides any earlier call; it does not close the previous file.
+// SetCrashOutput(nil) disables the use of any additional file.
+func SetCrashOutput(f *os.File) error {
+	fd := ^uintptr(0)
+	if f != nil {
+		// The runtime will write to this file descriptor from
+		// low-level routines during a panic, possibly without
+		// a G, so we must call f.Fd() eagerly. This creates a
+		// danger that that the file descriptor is no longer
+		// valid at the time of the write, because the caller
+		// (incorrectly) called f.Close() and the kernel
+		// reissued the fd in a later call to open(2), leading
+		// to crashes being written to the wrong file.
+		//
+		// So, we duplicate the fd to obtain a private one
+		// that cannot be closed by the user.
+		// This also alleviates us from concerns about the
+		// lifetime and finalization of f.
+		// (DupCloseOnExec returns an fd, not a *File, so
+		// there is no finalizer, and we are responsible for
+		// closing it.)
+		//
+		// The new fd must be close-on-exec, otherwise if the
+		// crash monitor is a child process, it may inherit
+		// it, so it will never see EOF from the pipe even
+		// when this process crashes.
+		//
+		// A side effect of Fd() is that it calls SetBlocking,
+		// which is important so that writes of a crash report
+		// to a full pipe buffer don't get lost.
+		fd2, _, err := poll.DupCloseOnExec(int(f.Fd()))
+		if err != nil {
+			return err
+		}
+		runtime.KeepAlive(f) // prevent finalization before dup
+		fd = uintptr(fd2)
+	}
+	if prev := runtime_setCrashFD(fd); prev != ^uintptr(0) {
+		// We use NewFile+Close because it is portable
+		// unlike syscall.Close, whose parameter type varies.
+		os.NewFile(prev, "").Close() // ignore error
+	}
+	return nil
+}
+
+//go:linkname runtime_setCrashFD runtime.setCrashFD
+func runtime_setCrashFD(uintptr) uintptr