Improve support for ptrace emulation during group stops. (#3926)

Firefox's minidump writer expects to be able to attach to tasks in group stops
and invoke ptrace commands on them (e.g. PTRACE_GETREGS)

Fixes #3741.

CMakeLists.txt

@@ -1261,6 +1261,7 @@ set(BASIC_TESTS
   x86/ptrace_debug_regs
   ptrace_exec
   x86/ptrace_exec32
+  ptrace_group_stop
   ptrace_kill_grandtracee
   x86/ptrace_tls
   ptrace_seize

src/RecordTask.cc

@@ -1185,34 +1185,28 @@ void RecordTask::emulate_SIGCONT() {
 }
 
 void RecordTask::signal_delivered(int sig) {
+  bool needs_SIGCHLD = true;
   Sighandler& h = sighandlers->get(sig);
   if (h.resethand) {
     reset_handler(&h, arch());
   }
 
-  if (!is_sig_ignored(sig)) {
-    switch (sig) {
-      case SIGTSTP:
-      case SIGTTIN:
-      case SIGTTOU:
-        if (h.disposition() == SIGNAL_HANDLER) {
-          break;
-        }
-        RR_FALLTHROUGH;
-      case SIGSTOP:
-        // All threads in the process are stopped.
-        for (Task* t : thread_group()->task_set()) {
-          auto rt = static_cast<RecordTask*>(t);
-          rt->apply_group_stop(sig);
-        }
-        break;
-      case SIGCONT:
-        emulate_SIGCONT();
-        break;
+  if (is_sig_stopping(sig)) {
+    // All threads in the process are stopped.
+    for (Task* t : thread_group()->task_set()) {
+      auto rt = static_cast<RecordTask*>(t);
+      rt->apply_group_stop(sig);
     }
+    // apply_group_stop calls send_synthetic_SIGCHLD_if_necessary(). Don't
+    // do it again.
+    needs_SIGCHLD = false;
+  } else if (sig == SIGCONT && !is_sig_ignored(sig)) {
+    emulate_SIGCONT();
   }
 
-  send_synthetic_SIGCHLD_if_necessary();
+  if (needs_SIGCHLD) {
+    send_synthetic_SIGCHLD_if_necessary();
+  }
 }
 
 bool RecordTask::signal_has_user_handler(int sig) const {
@@ -1259,6 +1253,21 @@ bool RecordTask::is_sig_ignored(int sig) const {
   }
 }
 
+bool RecordTask::is_sig_stopping(int sig) const {
+  switch (sig) {
+    case SIGTSTP:
+    case SIGTTIN:
+    case SIGTTOU:
+      if (sig_disposition(sig) != SIGNAL_DEFAULT) {
+        break;
+      }
+      RR_FALLTHROUGH;
+    case SIGSTOP:
+      return true;
+  }
+  return false;
+}
+
 SignalDisposition RecordTask::sig_disposition(int sig) const {
   return sighandlers->get(sig).disposition();
 }

src/RecordTask.h

@@ -246,6 +246,10 @@ class RecordTask final : public Task {
    * default disposition is "ignore".
    */
   bool is_sig_ignored(int sig) const;
+  /**
+   * Return true iff |sig| is a stopping signal.
+   */
+  bool is_sig_stopping(int sig) const;
   /**
    * Return the applications current disposition of |sig|.
    */

src/Scheduler.cc

@@ -345,7 +345,22 @@ bool Scheduler::is_task_runnable(RecordTask* t, WaitAggregator& wait_aggregator,
       WaitAggregator::try_wait_exit(t);
       // N.B.: If we supported ptrace exit notifications for killed tracee's
       // that would need handling here, but we don't at the moment.
-      return t->seen_ptrace_exit_event();
+      if (t->seen_ptrace_exit_event()) {
+        LOGM(debug) << "  ... but it died";
+        return true;
+      }
+      if (t->is_stopped()) {
+        return false;
+      }
+      // If we're not stopped, we need to get to the stop.
+      // AFAIK we can only get here with group stops, which are eagerly applied
+      // to every task in the group. If I'm wrong, die here.
+      ASSERT(t, t->emulated_stop_type == GROUP_STOP);
+      LOGM(debug) << "  interrupting and waiting";
+      t->do_ptrace_interrupt();
+      // Wait on the task to get the kernel to kick it into the group stop.
+      // If it died, we can deal with it later.
+      return t->wait();
     }
   }
 
@@ -787,8 +802,8 @@ Scheduler::Rescheduled Scheduler::reschedule(Switchable switchable) {
 #ifdef MONITOR_UNSWITCHABLE_WAITS
       double wait_duration = monotonic_now_sec() - now;
       if (wait_duration >= 0.010) {
-        log_warn("Waiting for unswitchable %s took %g ms",
-                 strevent(current_->event), 1000.0 * wait_duration);
+        LOGM(warn) << "Waiting for unswitchable " << current_->ev()
+                   << " took " << 1000.0 * wait_duration << "ms";
       }
 #endif
     }

src/test/ptrace_group_stop.c

@@ -0,0 +1,64 @@
+/* -*- Mode: C; tab-width: 8; c-basic-offset: 2; indent-tabs-mode: nil; -*- */
+
+#include "util.h"
+#include "ptrace_util.h"
+
+static void* do_thread(void* arg) {
+  int pipe_fd = *(int*)arg;
+  uint32_t tid = gettid();
+
+  write(pipe_fd, &tid, 4);
+  /* Sleep long enough that it will be noticed if it's not interrupted. */
+  sleep(1000);
+
+  return NULL;
+}
+
+int main(void) {
+  pid_t child, child2;
+  uint32_t msg;
+  int status;
+  int pipe_fds[2];
+  struct user_regs_struct regs;
+
+  test_assert(0 == pipe(pipe_fds));
+
+  if (0 == (child = fork())) {
+    pthread_t t;
+
+    pthread_create(&t, NULL, do_thread, &pipe_fds[1]);
+    pthread_join(t, NULL);
+
+    return 77;
+  }
+
+  test_assert(4 == read(pipe_fds[0], &msg, 4));
+  child2 = (pid_t)msg;
+  close(pipe_fds[0]);
+  sched_yield();
+
+  /* Hit the entire process group with a SIGSTOP. */
+  tgkill(child, child, SIGSTOP);
+
+  /* Force the rr scheduler to run. */
+  sched_yield();
+
+  /* Now seize the stopped task. */
+  test_assert(0 == ptrace(PTRACE_SEIZE, child2, 0, 0));
+  test_assert(child2 == waitpid(child2, &status, 0));
+  test_assert(WIFSTOPPED(status) && WSTOPSIG(status) == SIGSTOP);
+
+  /* Do something that requires the task to be stopped. */
+  ptrace_getregs(child2, &regs);
+
+  /* Verify that we can resume from group stops. */
+  test_assert(0 == ptrace(PTRACE_CONT, child2, 0, 0));
+  /* Force the rr scheduler to run. */
+  sched_yield();
+  test_assert(0 == ptrace(PTRACE_INTERRUPT, child2, 0, 0));
+  test_assert(child2 == waitpid(child2, &status, 0));
+  test_assert(WIFSTOPPED(status) && WSTOPSIG(status) == SIGSTOP);
+
+  atomic_puts("EXIT-SUCCESS");
+  return 0;
+}