Re: grep-use-null-device

[Stefan Monnier <monnier@iro.umontreal.ca>]

>>> `compilation-start' needs to check if the process is running
>>> before calling `process-send-eof':
>> 
>> That's odd.  AFAICT no blobking operation takes place between the
>> start-process and the process-send-eof, so the process-status should still
>> be `run' no matter how quickly the process exits (because Emacs shouldn't
>> process the SIGCHLD it receives until later).
>> 
>> What am I missing?

> The process exits during execution of create_process.  The gdb log below
> with a breakpoint on sigchld_handler demonstrates what really happens:

> Breakpoint 4, sigchld_handler (signo=17) at process.c:6249
> 6249              XSETINT (p->raw_status_low, u.i & 0xffff);
> (gdb) n
> 6250              XSETINT (p->raw_status_high, u.i >> 16);
> (gdb) n
> 6253              if ((WIFSIGNALED (w) || WIFEXITED (w))
> (gdb) n
> 6260                  FD_CLR (XINT (p->infd), &input_wait_mask);
> (gdb) n
> 6261                  FD_CLR (XINT (p->infd), &non_keyboard_wait_mask);
> (gdb) p w
> $1 = 512     <-- WIFEXITED
> (gdb) bt
> #0  sigchld_handler (signo=17) at process.c:6261
> #1  <signal handler called>
> #2  0x4031f784 in sigprocmask () from /lib/libc.so.6
> #3  0x0817af28 in create_process (process=141365940, new_argv=0xbfffe954, current_dir=140249731) at process.c:2153
> #4  0x0817a97d in Fstart_process (nargs=5, args=0xbfffea94) at process.c:1695
> ...
> (gdb) fr 3
> #3  0x0817af28 in create_process (process=141365940, new_argv=0xbfffe954, current_dir=140249731) at process.c:2153
> 2153      sigprocmask (SIG_SETMASK, &procmask, 0);

Oh, I now see that process-send-eof does:

  if (! NILP (XPROCESS (proc)->raw_status_low))
    update_status (XPROCESS (proc));

which basically copies the asynchronously updated proc->raw_status_* to the
synchronously updated proc->status.

Since this process status can change asynchronously, adding your test
for (eq (process-status proc) 'run) before calling process-send-eof doesn't
fix the bug but just narrows the window of the race condition because the
status can still change between the call to process-status and the call to
process-send-eof.

So I suggest the patch below instead,


        Stefan


--- compile.el	01 sep 2005 10:04:39 -0400	1.379
+++ compile.el	01 sep 2005 10:58:22 -0400	
@@ -1038,11 +1038,14 @@
 						       outbuf command))))
 	      ;; Make the buffer's mode line show process state.
 	      (setq mode-line-process '(":%s"))
-	      (when compilation-disable-input
-		(process-send-eof proc))
 	      (set-process-sentinel proc 'compilation-sentinel)
 	      (set-process-filter proc 'compilation-filter)
 	      (set-marker (process-mark proc) (point) outbuf)
+	      (when compilation-disable-input
+                (condition-case nil
+                    (process-send-eof proc)
+                  ;; The process may have exited already.
+                  (error nil)))
 	      (setq compilation-in-progress
 		    (cons proc compilation-in-progress)))
 	  ;; No asynchronous processes available.