[Emacs-diffs] /srv/bzr/emacs/trunk r110784: Fix a race condition that causes Emacs to mess up glib.

[Paul Eggert]

------------------------------------------------------------
revno: 110784
fixes bug: http://debbugs.gnu.org/8855
committer: Paul Eggert <address@hidden>
branch nick: trunk
timestamp: Sat 2012-11-03 11:32:41 -0700
message:
  Fix a race condition that causes Emacs to mess up glib.
  
  The symptom is a diagnostic "GLib-WARNING **: In call to
  g_spawn_sync(), exit status of a child process was requested but
  SIGCHLD action was set to SIG_IGN and ECHILD was received by
  waitpid(), so exit status can't be returned."  The diagnostic
  is partly wrong, as the SIGCHLD action is not set to SIG_IGN.
  The real bug is a race condition between Emacs and glib: Emacs
  does a waitpid (-1, ...) and reaps glib's subprocess by mistake,
  so that glib can't find it.  Work around the bug by invoking
  waitpid only on subprocesses that Emacs itself creates.
  * process.c (create_process, record_child_status_change):
  Don't use special value -1 in pid field, as the caller now must
  know the pid rather than having the callee infer it.  The
  inference was sometimes incorrect anyway, due to another race.
  (create_process): Set new 'alive' member if child is created.
  (process_status_retrieved): New function.
  (record_child_status_change): Use it.
  Accept negative 1st argument, which means to wait for the
  processes that Emacs already knows about.  Move special-case code
  for DOS_NT (which lacks WNOHANG) here, from caller.  Keep track of
  processes that have already been waited for, by testing and
  clearing new 'alive' member.
  (CAN_HANDLE_MULTIPLE_CHILDREN): Remove, as record_child_status_change
  now does this internally.
  (handle_child_signal): Let record_child_status_change do all
  the work, since we do not want to reap all exited child processes,
  only the child processes that Emacs itself created.
  * process.h (Lisp_Process): New boolean member 'alive'.
modified:
  src/ChangeLog
  src/process.c
  src/process.h

=== modified file 'src/ChangeLog'
--- a/src/ChangeLog     2012-11-03 18:19:50 +0000
+++ b/src/ChangeLog     2012-11-03 18:32:41 +0000
@@ -1,5 +1,34 @@
 2012-11-03  Paul Eggert  <address@hidden>
 
+       Fix a race condition that causes Emacs to mess up glib (Bug#8855).
+       The symptom is a diagnostic "GLib-WARNING **: In call to
+       g_spawn_sync(), exit status of a child process was requested but
+       SIGCHLD action was set to SIG_IGN and ECHILD was received by
+       waitpid(), so exit status can't be returned."  The diagnostic
+       is partly wrong, as the SIGCHLD action is not set to SIG_IGN.
+       The real bug is a race condition between Emacs and glib: Emacs
+       does a waitpid (-1, ...) and reaps glib's subprocess by mistake,
+       so that glib can't find it.  Work around the bug by invoking
+       waitpid only on subprocesses that Emacs itself creates.
+       * process.c (create_process, record_child_status_change):
+       Don't use special value -1 in pid field, as the caller now must
+       know the pid rather than having the callee infer it.  The
+       inference was sometimes incorrect anyway, due to another race.
+       (create_process): Set new 'alive' member if child is created.
+       (process_status_retrieved): New function.
+       (record_child_status_change): Use it.
+       Accept negative 1st argument, which means to wait for the
+       processes that Emacs already knows about.  Move special-case code
+       for DOS_NT (which lacks WNOHANG) here, from caller.  Keep track of
+       processes that have already been waited for, by testing and
+       clearing new 'alive' member.
+       (CAN_HANDLE_MULTIPLE_CHILDREN): Remove, as record_child_status_change
+       now does this internally.
+       (handle_child_signal): Let record_child_status_change do all
+       the work, since we do not want to reap all exited child processes,
+       only the child processes that Emacs itself created.
+       * process.h (Lisp_Process): New boolean member 'alive'.
+
        Omit duplicate definitions no longer needed with gcc -g3.
        * lisp.h (GCTYPEBITS, GCALIGNMENT, ARRAY_MARK_FLAG, PSEUDOVECTOR_FLAG)
        (VALMASK, MOST_POSITIVE_FIXNUM, MOST_NEGATIVE_FIXNUM):

=== modified file 'src/process.c'
--- a/src/process.c     2012-11-03 16:36:23 +0000
+++ b/src/process.c     2012-11-03 18:32:41 +0000
@@ -130,6 +130,10 @@
                       EMACS_TIME *, void *);
 #endif
 
+/* This is for DOS_NT ports.  FIXME: Remove this old portability cruft
+   by having DOS_NT ports implement waitpid instead of wait.  Nowadays
+   POSIXish hosts all define waitpid, WNOHANG, and WUNTRACED, as these
+   have been standard since POSIX.1-1988.  */
 #ifndef WNOHANG
 # undef waitpid
 # define waitpid(pid, status, options) wait (status)
@@ -795,9 +799,8 @@
 #ifdef SIGCHLD
 /* Fdelete_process promises to immediately forget about the process, but in
    reality, Emacs needs to remember those processes until they have been
-   treated by the SIGCHLD handler; otherwise this handler would consider the
-   process as being synchronous and say that the synchronous process is
-   dead.  */
+   treated by the SIGCHLD handler and waitpid has been invoked on them;
+   otherwise they might fill up the kernel's process table.  */
 static Lisp_Object deleted_pid_list;
 #endif
 
@@ -1704,16 +1707,7 @@
   if (inchannel > max_process_desc)
     max_process_desc = inchannel;
 
-  /* Until we store the proper pid, enable the SIGCHLD handler
-     to recognize an unknown pid as standing for this process.
-     It is very important not to let this `marker' value stay
-     in the table after this function has returned; if it does
-     it might cause call-process to hang and subsequent asynchronous
-     processes to get their return values scrambled.  */
-  XPROCESS (process)->pid = -1;
-
-  /* This must be called after the above line because it may signal an
-     error. */
+  /* This may signal an error. */
   setup_process_coding_systems (process);
 
   encoded_current_dir = ENCODE_FILE (current_dir);
@@ -1880,6 +1874,8 @@
 #endif
 
   XPROCESS (process)->pid = pid;
+  if (0 <= pid)
+    XPROCESS (process)->alive = 1;
 
   /* Stop blocking signals in the parent.  */
 #ifdef SIGCHLD
@@ -6271,9 +6267,35 @@
   return process;
 }
 
-/* On receipt of a signal that a child status has changed, loop asking
-   about children with changed statuses until the system says there
-   are no more.
+/* If the status of the process DESIRED has changed, return true and
+   set *STATUS to its exit status; otherwise, return false.
+   If HAVE is nonnegative, assume that HAVE = waitpid (HAVE, STATUS, ...)
+   has already been invoked, and do not invoke waitpid again.  */
+
+static bool
+process_status_retrieved (pid_t desired, pid_t have, int *status)
+{
+  if (have < 0)
+    {
+      /* Invoke waitpid only with a known process ID; do not invoke
+        waitpid with a nonpositive argument.  Otherwise, Emacs might
+        reap an unwanted process by mistake.  For example, invoking
+        waitpid (-1, ...) can mess up glib by reaping glib's subprocesses,
+        so that another thread running glib won't find them.  */
+      do
+       have = waitpid (desired, status, WNOHANG | WUNTRACED);
+      while (have < 0 && errno == EINTR);
+    }
+
+  return have == desired;
+}
+
+/* If PID is nonnegative, the child process PID with wait status W has
+   changed its status; record this and return true.
+
+   If PID is negative, ignore W, and look for known child processes
+   of Emacs whose status have changed.  For each one found, record its new
+   status.
 
    All we do is change the status; we do not run sentinels or print
    notifications.  That is saved for the next time keyboard input is
@@ -6296,13 +6318,23 @@
    ** Malloc WARNING: This should never call malloc either directly or
    indirectly; if it does, that is a bug  */
 
-/* Record the changed status of the child process PID with wait status W.  */
 void
 record_child_status_change (pid_t pid, int w)
 {
 #ifdef SIGCHLD
-  Lisp_Object proc;
-  struct Lisp_Process *p;
+
+# ifdef WNOHANG
+  /* On POSIXish hosts, record at most one child only if we already
+     know one child that has exited.  */
+  bool record_at_most_one_child = 0 <= pid;
+# else
+  /* On DOS_NT (the only porting target that lacks WNOHANG),
+     record the status of at most one child process, since the SIGCHLD
+     handler must return right away.  If any more processes want to
+     signal us, we will get another signal.  */
+  bool record_at_most_one_child = 1;
+# endif
+
   Lisp_Object tail;
 
   /* Find the process that signaled us, and record its status.  */
@@ -6310,68 +6342,69 @@
   /* The process can have been deleted by Fdelete_process.  */
   for (tail = deleted_pid_list; CONSP (tail); tail = XCDR (tail))
     {
+      bool all_pids_are_fixnums
+       = (MOST_NEGATIVE_FIXNUM <= TYPE_MINIMUM (pid_t)
+          && TYPE_MAXIMUM (pid_t) <= MOST_POSITIVE_FIXNUM);
       Lisp_Object xpid = XCAR (tail);
-      if ((INTEGERP (xpid) && pid == XINT (xpid))
-         || (FLOATP (xpid) && pid == XFLOAT_DATA (xpid)))
+      if (all_pids_are_fixnums ? INTEGERP (xpid) : NUMBERP (xpid))
        {
-         XSETCAR (tail, Qnil);
-         return;
+         pid_t deleted_pid;
+         if (INTEGERP (xpid))
+           deleted_pid = XINT (xpid);
+         else
+           deleted_pid = XFLOAT_DATA (xpid);
+         if (process_status_retrieved (deleted_pid, pid, &w))
+           {
+             XSETCAR (tail, Qnil);
+             if (record_at_most_one_child)
+               return;
+           }
        }
     }
 
   /* Otherwise, if it is asynchronous, it is in Vprocess_alist.  */
-  p = 0;
   for (tail = Vprocess_alist; CONSP (tail); tail = XCDR (tail))
     {
-      proc = XCDR (XCAR (tail));
-      p = XPROCESS (proc);
-      if (EQ (p->type, Qreal) && p->pid == pid)
-       break;
-      p = 0;
-    }
-
-  /* Look for an asynchronous process whose pid hasn't been filled
-     in yet.  */
-  if (! p)
-    for (tail = Vprocess_alist; CONSP (tail); tail = XCDR (tail))
-      {
-       proc = XCDR (XCAR (tail));
-       p = XPROCESS (proc);
-       if (p->pid == -1)
-         break;
-       p = 0;
-      }
-
-  /* Change the status of the process that was found.  */
-  if (p)
-    {
-      int clear_desc_flag = 0;
-
-      p->tick = ++process_tick;
-      p->raw_status = w;
-      p->raw_status_new = 1;
-
-      /* If process has terminated, stop waiting for its output.  */
-      if ((WIFSIGNALED (w) || WIFEXITED (w))
-         && p->infd >= 0)
-       clear_desc_flag = 1;
-
-      /* We use clear_desc_flag to avoid a compiler bug in Microsoft C.  */
-      if (clear_desc_flag)
+      Lisp_Object proc = XCDR (XCAR (tail));
+      struct Lisp_Process *p = XPROCESS (proc);
+      if (p->alive && process_status_retrieved (p->pid, pid, &w))
        {
-         FD_CLR (p->infd, &input_wait_mask);
-         FD_CLR (p->infd, &non_keyboard_wait_mask);
+         /* Change the status of the process that was found.  */
+         p->tick = ++process_tick;
+         p->raw_status = w;
+         p->raw_status_new = 1;
+
+         /* If process has terminated, stop waiting for its output.  */
+         if (WIFSIGNALED (w) || WIFEXITED (w))
+           {
+             int clear_desc_flag = 0;
+             p->alive = 0;
+             if (p->infd >= 0)
+               clear_desc_flag = 1;
+
+             /* clear_desc_flag avoids a compiler bug in Microsoft C.  */
+             if (clear_desc_flag)
+               {
+                 FD_CLR (p->infd, &input_wait_mask);
+                 FD_CLR (p->infd, &non_keyboard_wait_mask);
+               }
+           }
+
+         /* Tell wait_reading_process_output that it needs to wake up and
+            look around.  */
+         if (input_available_clear_time)
+           *input_available_clear_time = make_emacs_time (0, 0);
+
+         if (record_at_most_one_child)
+           return;
        }
-
-      /* Tell wait_reading_process_output that it needs to wake up and
-        look around.  */
-      if (input_available_clear_time)
-       *input_available_clear_time = make_emacs_time (0, 0);
     }
-  /* There was no asynchronous process found for that pid: we have
-     a synchronous process.  */
-  else
+
+  if (0 <= pid)
     {
+      /* The caller successfully waited for a pid but no asynchronous
+        process was found for it, so this is a synchronous process.  */
+
       synch_process_alive = 0;
 
       /* Report the status of the synchronous process.  */
@@ -6390,38 +6423,10 @@
 
 #ifdef SIGCHLD
 
-/* On some systems, the SIGCHLD handler must return right away.  If
-   any more processes want to signal us, we will get another signal.
-   Otherwise, loop around to use up all the processes that have
-   something to tell us.  */
-#if (defined WINDOWSNT \
-     || (defined USG && !defined GNU_LINUX \
-        && !(defined HPUX && defined WNOHANG)))
-enum { CAN_HANDLE_MULTIPLE_CHILDREN = 0 };
-#else
-enum { CAN_HANDLE_MULTIPLE_CHILDREN = 1 };
-#endif
-
 static void
 handle_child_signal (int sig)
 {
-  do
-    {
-      pid_t pid;
-      int status;
-
-      do
-       pid = waitpid (-1, &status, WNOHANG | WUNTRACED);
-      while (pid < 0 && errno == EINTR);
-
-      /* PID == 0 means no processes found, PID == -1 means a real failure.
-        Either way, we have done all our job.  */
-      if (pid <= 0)
-       break;
-
-      record_child_status_change (pid, status);
-    }
-  while (CAN_HANDLE_MULTIPLE_CHILDREN);
+  record_child_status_change (-1, 0);
 }
 
 static void

=== modified file 'src/process.h'
--- a/src/process.h     2012-08-27 17:23:48 +0000
+++ b/src/process.h     2012-11-03 18:32:41 +0000
@@ -142,6 +142,9 @@
     /* Flag to set coding-system of the process buffer from the
        coding_system used to decode process output.  */
     unsigned int inherit_coding_system_flag : 1;
+    /* Whether the process is alive, i.e., can be waited for.  Running
+       processes can be waited for, but exited and fake processes cannot.  */
+    unsigned int alive : 1;
     /* Record the process status in the raw form in which it comes from `wait'.
        This is to avoid consing in a signal handler.  The `raw_status_new'
        flag indicates that `raw_status' contains a new status that still