runtime: don't crash on mlock failure

Instead, note that mlock has failed, start trying the mitigation of
touching the signal stack before sending a preemption signal, and,
if the program crashes, mention the possible problem and a wiki page
describing the issue (https://golang.org/wiki/LinuxKernelSignalVectorBug).

Tested on a kernel in the buggy version range, but with the patch,
by using `ulimit -l 0`.

Fixes #37436

Change-Id: I072aadb2101496dffd655e442fa5c367dad46ce8
Reviewed-on: https://go-review.googlesource.com/c/go/+/223121
Run-TryBot: Ian Lance Taylor <[email protected]>
Reviewed-by: Austin Clements <[email protected]>
Reviewed-by: Keith Randall <[email protected]>

Author: ianlancetaylor

src/runtime/os_linux.go

@@ -5,6 +5,7 @@
 package runtime
 
 import (
+	"runtime/internal/atomic"
 	"runtime/internal/sys"
 	"unsafe"
 )
@@ -479,7 +480,21 @@ func rt_sigaction(sig uintptr, new, old *sigactiont, size uintptr) int32
 func getpid() int
 func tgkill(tgid, tid, sig int)
 
+// touchStackBeforeSignal stores an errno value. If non-zero, it means
+// that we should touch the signal stack before sending a signal.
+// This is used on systems that have a bug when the signal stack must
+// be faulted in.  See #35777 and #37436.
+//
+// This is accessed atomically as it is set and read in different threads.
+//
+// TODO(austin): Remove this after Go 1.15 when we remove the
+// mlockGsignal workaround.
+var touchStackBeforeSignal uint32
+
 // signalM sends a signal to mp.
 func signalM(mp *m, sig int) {
+	if atomic.Load(&touchStackBeforeSignal) != 0 {
+		atomic.Cas((*uint32)(unsafe.Pointer(mp.gsignal.stack.hi-4)), 0, 0)
+	}
 	tgkill(getpid(), int(mp.procid), sig)
 }

src/runtime/os_linux_x86.go

@@ -7,6 +7,8 @@
 
 package runtime
 
+import "runtime/internal/atomic"
+
 //go:noescape
 func uname(utsname *new_utsname) int
 
@@ -58,17 +60,34 @@ func osArchInit() {
 		if m0.gsignal != nil {
 			throw("gsignal quirk too late")
 		}
+		throwReportQuirk = throwBadKernel
 	}
 }
 
 func mlockGsignal(gsignal *g) {
-	if err := mlock(gsignal.stack.hi-physPageSize, physPageSize); err < 0 {
-		printlock()
-		println("runtime: mlock of signal stack failed:", -err)
-		if err == -_ENOMEM {
-			println("runtime: increase the mlock limit (ulimit -l) or")
-		}
-		println("runtime: update your kernel to 5.3.15+, 5.4.2+, or 5.5+")
-		throw("mlock failed")
+	if atomic.Load(&touchStackBeforeSignal) != 0 {
+		// mlock has already failed, don't try again.
+		return
+	}
+
+	// This mlock call may fail, but we don't report the failure.
+	// Instead, if something goes badly wrong, we rely on prepareSignalM
+	// and throwBadKernel to do further mitigation and to report a problem
+	// to the user if mitigation fails. This is because many
+	// systems have a limit on the total mlock size, and many kernels
+	// that appear to have bad versions are actually patched to avoid the
+	// bug described above. We want Go 1.14 to run on those systems.
+	// See #37436.
+	if errno := mlock(gsignal.stack.hi-physPageSize, physPageSize); errno < 0 {
+		atomic.Store(&touchStackBeforeSignal, uint32(-errno))
+	}
+}
+
+// throwBadKernel is called, via throwReportQuirk, by throw.
+func throwBadKernel() {
+	if errno := atomic.Load(&touchStackBeforeSignal); errno != 0 {
+		println("runtime: note: your Linux kernel may be buggy")
+		println("runtime: note: see https://golang.org/wiki/LinuxKernelSignalVectorBug")
+		println("runtime: note: mlock workaround for kernel bug failed with errno", errno)
 	}
 }

src/runtime/panic.go

@@ -1281,6 +1281,12 @@ func startpanic_m() bool {
 	}
 }
 
+// throwReportQuirk, if non-nil, is called by throw after dumping the stacks.
+//
+// TODO(austin): Remove this after Go 1.15 when we remove the
+// mlockGsignal workaround.
+var throwReportQuirk func()
+
 var didothers bool
 var deadlock mutex
 
@@ -1327,6 +1333,10 @@ func dopanic_m(gp *g, pc, sp uintptr) bool {
 
 	printDebugLog()
 
+	if throwReportQuirk != nil {
+		throwReportQuirk()
+	}
+
 	return docrash
 }