Fix race condition in pgid set for spawned tasks.

NEWS

 This file describes changes in recent versions of SLURM. It primarily
 documents those changes that are of interest to users and admins. 
 
+* Changes in SLURM 0.4.4
+========================
+ -- Fix race condition in slurmd seting pgid of spawned tasks for 
+    process tracking.
+
 * Changes in SLURM 0.4.3
 ========================
  -- Turn off some RPM build checks (bug in RPM, see slurm.spec.in)

src/slurmd/smgr.c

@@ -75,7 +75,7 @@ static int smgr_sigarray[] = {
  */
 static void  _session_mgr(slurmd_job_t *job);
 static int   _exec_all_tasks(slurmd_job_t *job);
-static void  _exec_task(slurmd_job_t *job, int i);
+static void  _exec_task(slurmd_job_t *job, int i, int fd);
 static int   _become_user(slurmd_job_t *job);
 static void  _make_tmpdir(slurmd_job_t *job);
 static int   _child_exited(void);
@@ -273,6 +273,8 @@ _exec_all_tasks(slurmd_job_t *job)
 {
 	int i;
 	int fd = job->fdpair[1];
+	int cpipe[2];
+	uint8_t tid;
 
 	xassert(job != NULL);
 	xassert(fd >= 0);
@@ -285,6 +287,9 @@ _exec_all_tasks(slurmd_job_t *job)
 	if (xsignal_block(smgr_sigarray) < 0)
 		return error ("Unable to block signals");
 
+	if (pipe (cpipe) < 0)
+		return error ("Unable to open child pipe: %m");
+
 	for (i = 0; i < job->ntasks; i++) {
 		pid_t pid = fork();
 
@@ -292,7 +297,7 @@ _exec_all_tasks(slurmd_job_t *job)
 			error("fork: %m");
 			return SLURM_ERROR;
 		} else if (pid == 0)  /* child */
-			_exec_task(job, i);
+			_exec_task(job, i, cpipe[1]);
 
 		/* Parent continues: 
 		 */
@@ -310,6 +315,12 @@ _exec_all_tasks(slurmd_job_t *job)
 
 		job->task[i]->pid = pid;
 
+		/*
+		 * For task 0, wait until it has created a new pgrp.
+		 */
+		if (i == 0)
+			read (cpipe[0], &tid, sizeof (tid));
+
 		/*
 		 * Prepare process for attach by parallel debugger 
 		 * (if specified and able)
@@ -317,18 +328,30 @@ _exec_all_tasks(slurmd_job_t *job)
 		_pdebug_trace_process(job, pid);
 	}
 
-   again:
-	for (i = 1; i < job->ntasks; i++)
-		if (getpgid (job->task[i]->pid) != job->task[0]->pid) 
-			goto again;
+	/*
+	 * Wait for all tasks to finish joining new process group
+	 */
+	for (i = 1; i < job->ntasks; i++)  
+		read (cpipe[0], &tid, sizeof (tid));
+
+	close (cpipe[0]);
+	close (cpipe[1]);
 
 	return SLURM_SUCCESS;
 }
 
 
 static void
-_exec_task(slurmd_job_t *job, int i)
+_exec_task(slurmd_job_t *job, int i, int fd)
 {
+	/*
+	 * Fit taskid into single byte for writing back to
+	 *  initiating slurmd. Doesn't matter if the values
+	 *  are truncated, we just need the number of bytes to
+	 *  be right.
+	 */
+	uint8_t tid = i;
+
 	if (xsignal_unblock(smgr_sigarray) < 0) {
 		error("unable to unblock signals");
 		exit(1);
@@ -341,6 +364,14 @@ _exec_task(slurmd_job_t *job, int i)
 		error ("Unable to put task %d into pgrp %ld: %m",
 			i, job->task[0]->pid);
 
+	/*
+	 * Notify slurmd that pgid has been set for this task
+	 */
+	if (write (fd, &tid, sizeof (tid)) < 0)
+		error ("Unable to notify slurmd that task %d started\n", i);
+
+	close (fd);
+
 	if (!job->batch) {
 		if (interconnect_attach(job->switch_job, &job->env,
 				job->nodeid, (uint32_t) i, job->nnodes,