fix(debug): handle mapping edges, stray traps and dying debuggees

Assisted-by: GLM 5.3

debug/ptrace_linux.go

@@ -91,6 +91,16 @@ func LaunchWithBuffers(gasmBin, asmPath, funcName string, args []byte, bufSpec s
 		if _, err := os.Stat(readyFile); err == nil {
 			break
 		}
+		// A debuggee that died before signalling readiness (unknown
+		// function, unparseable source) writes its failure notice to the
+		// handshake directory; read it and fail fast.  The poll never
+		// waits on the child: a wait here could consume the SIGSTOP park
+		// that waitStopped below must receive, hanging the launch.
+		if err := s.deadReason(); err != nil {
+			cmd.Wait()
+			os.RemoveAll(tmpDir)
+			return nil, nil, err
+		}
 		time.Sleep(5 * time.Millisecond)
 	}
 
@@ -131,6 +141,18 @@ func LaunchWithBuffers(gasmBin, asmPath, funcName string, args []byte, bufSpec s
 	return s, bufAddrs, nil
 }
 
+// deadReason reports the debuggee's own failure notice, the file its
+// failure paths write before exiting.  A debuggee killed without a notice
+// (a crash, SIGKILL) surfaces through waitStopped after the poll instead,
+// which is why the poll's budget stays finite.
+func (s *Session) deadReason() error {
+	data, err := os.ReadFile(filepath.Join(s.tmpDir, "dead"))
+	if err != nil {
+		return nil
+	}
+	return fmt.Errorf("debug: debuggee failed before signalling readiness: %s", strings.TrimSpace(string(data)))
+}
+
 // waitStopped consumes ptrace-stop events until one the debugger cares
 // about arrives: SIGTRAP (a breakpoint or a completed single-step), the
 // debuggee's own SIGSTOP, or a genuine signal-delivery-stop.  A Go tracee's
@@ -219,40 +241,41 @@ func (s *Session) Poke(addr, val uint64) error {
 	return nil
 }
 
-// ReadMemory reads len bytes from the debuggee's memory at addr.
+// ReadMemory reads len bytes from the debuggee's memory at addr.  The read
+// covers exactly the requested range: the old word-at-a-time loop read a
+// whole 8-byte word for the final partial word, so a request that ended
+// inside the last mapped page failed whenever the following page was
+// unmapped, even though every requested byte was readable.
 func (s *Session) ReadMemory(addr uint64, length int) ([]byte, error) {
 	out := make([]byte, length)
-	for i := 0; i < length; i += 8 {
-		word, err := s.Peek(addr + uint64(i))
-		if err != nil {
-			return out[:i], err
-		}
-		for j := 0; j < 8 && i+j < length; j++ {
-			out[i+j] = byte(word >> (8 * j))
-		}
+	mem, err := os.OpenFile(fmt.Sprintf("/proc/%d/mem", s.pid), os.O_RDONLY, 0)
+	if err != nil {
+		return out, fmt.Errorf("debug: open /proc/%d/mem: %w", s.pid, err)
+	}
+	defer mem.Close()
+	n, err := mem.ReadAt(out, int64(addr))
+	if err != nil {
+		return out[:n], fmt.Errorf("debug: read mem %#x: %w", addr, err)
 	}
 	return out, nil
 }
 
-// WriteMemory writes bytes to the debuggee's memory at addr.
+// WriteMemory writes bytes to the debuggee's memory at addr.  The write
+// covers exactly the given bytes: /proc/pid/mem accepts writes of any
+// length at any offset, so the word loop's read-modify-write of the final
+// partial word (which read past the requested range and failed on an
+// unmapped following page) is unnecessary.
 func (s *Session) WriteMemory(addr uint64, data []byte) error {
-	for i := 0; i < len(data); i += 8 {
-		end := min(i+8, len(data))
-		var word uint64
-		for j := 0; j < end-i; j++ {
-			word |= uint64(data[i+j]) << (8 * j)
-		}
-		if end-i < 8 {
-			existing, err := s.Peek(addr + uint64(i))
-			if err != nil {
-				return err
-			}
-			mask := ^((uint64(1) << (8 * (end - i))) - 1)
-			word = (existing & mask) | word
-		}
-		if err := s.Poke(addr+uint64(i), word); err != nil {
-			return err
-		}
+	if len(data) == 0 {
+		return nil
+	}
+	mem, err := os.OpenFile(fmt.Sprintf("/proc/%d/mem", s.pid), os.O_WRONLY, 0)
+	if err != nil {
+		return fmt.Errorf("debug: open /proc/%d/mem: %w", s.pid, err)
+	}
+	defer mem.Close()
+	if _, err := mem.WriteAt(data, int64(addr)); err != nil {
+		return fmt.Errorf("debug: write mem %#x: %w", addr, err)
 	}
 	return nil
 }