From 0344cdf364739af09e2f00b199aa163c60700a1c Mon Sep 17 00:00:00 2001
From: Morris Jette <jette@schedmd.com>
Date: Tue, 26 Apr 2011 21:36:56 -0700
Subject: [PATCH] salloc: support for Cray container IDs
This adds detection and use of SGI process aggregate job container IDs for
salloc interactive sessions.
The preferred and documented way to support this on a Cray system is by
enabling the provided pam_job.so via /etc/pam.d/common-session.
There is a header dependency on job.h. This file depends on the optional
cray-libjob-devel package, which installs into /opt/cray/job/<version>.
This package is however not always installed or may not be up-to-date.
Hence the patch "cheats" by duplicating the known prototype of job_getjid().
---
src/salloc/salloc.c | 31 +++++++++++++++++++++++++++----
src/slurmctld/proc_req.c | 6 +++---
2 files changed, 30 insertions(+), 7 deletions(-)
diff --git a/src/salloc/salloc.c b/src/salloc/salloc.c
index 8b5645b7298..68510b40024 100644
--- a/src/salloc/salloc.c
+++ b/src/salloc/salloc.c
@@ -70,10 +70,14 @@
#ifdef HAVE_BG
#include "src/common/node_select.h"
#include "src/plugins/select/bluegene/bg_enums.h"
-#endif
-
-#ifndef __USE_XOPEN_EXTENDED
-extern pid_t getsid(pid_t pid); /* missing from <unistd.h> */
+#elif defined(HAVE_CRAY)
+#include "src/common/node_select.h"
+/*
+ * On Cray installations, the libjob headers are not automatically installed
+ * by default, while libjob.so always is, and kernels are > 2.6. Hence it is
+ * simpler to just duplicate the single declaration here.
+ */
+extern uint64_t job_getjid(pid_t pid);
#endif
#define MAX_RETRIES 10
@@ -543,6 +547,25 @@ static void _set_submit_dir_env(void)
/* Returns 0 on success, -1 on failure */
static int _fill_job_desc_from_opts(job_desc_msg_t *desc)
{
+#ifdef HAVE_CRAY
+ uint64_t pagg_id = job_getjid(getpid());
+ /*
+ * Interactive sessions require pam_job.so in /etc/pam.d/common-session
+ * since creating sgi_job containers requires root permissions. This is
+ * the only exception where we allow the fallback of using the SID to
+ * confirm the reservation (caught later, in do_basil_confirm).
+ */
+ if (pagg_id == (uint64_t)-1) {
+ error("No SGI job container ID detected - please enable the "
+ "Cray job service via /etc/init.d/job");
+ } else {
+ if (!desc->select_jobinfo)
+ desc->select_jobinfo = select_g_select_jobinfo_alloc();
+
+ select_g_select_jobinfo_set(desc->select_jobinfo,
+ SELECT_JOBDATA_PAGG_ID, &pagg_id);
+ }
+#endif
desc->contiguous = opt.contiguous ? 1 : 0;
desc->features = opt.constraints;
desc->gres = opt.gres;
diff --git a/src/slurmctld/proc_req.c b/src/slurmctld/proc_req.c
index 80722779576..44d0c65138c 100644
--- a/src/slurmctld/proc_req.c
+++ b/src/slurmctld/proc_req.c
@@ -777,9 +777,9 @@ static void _slurm_rpc_allocate_resources(slurm_msg_t * msg)
}
#if HAVE_CRAY
/*
- * We are using the alloc_sid as unique identifier to confirm the ALPS
- * reservation. ALPS will refuse any attempt to create a second session
- * with the same identifier, hence sessions may not be nested.
+ * Catch attempts to nest salloc sessions. It is not possible to use an
+ * ALPS session which has the same alloc_sid, it fails even if PAGG
+ * container IDs are used.
*/
if (allocated_session_in_use(job_desc_msg)) {
error_code = ESLURM_RESERVATION_BUSY;
--
GitLab