From ac77a3f33fee7b7da2f257e7d672a40e9157bc2f Mon Sep 17 00:00:00 2001
From: Moe Jette <jette1@llnl.gov>
Date: Wed, 16 Feb 2005 02:07:31 +0000
Subject: [PATCH] Fix race condition in pgid set for spawned tasks.
---
NEWS | 5 +++++
src/slurmd/smgr.c | 45 ++++++++++++++++++++++++++++++++++++++-------
2 files changed, 43 insertions(+), 7 deletions(-)
diff --git a/NEWS b/NEWS
index c0398d927eb..431b7f3492f 100644
--- a/NEWS
+++ b/NEWS
@@ -1,6 +1,11 @@
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)
diff --git a/src/slurmd/smgr.c b/src/slurmd/smgr.c
index f5061545497..fb70c70d7aa 100644
--- a/src/slurmd/smgr.c
+++ b/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,
--
GitLab