[slurm-dev] A problem with devices cgroup subsystem.

[Александр Берсенев]

Hello,

I use task/cgroup plugin. When I enabling ConstrainDevices=yes in
cgroups.conf tasks refuses to run.
The error is: sinfo: bitstring.c:238: bit_nclear: Assertion `(start) <
((b)[1])' failed.

The error occurs while calling
gres_conf_lines = gres_plugin_node_config_devices_path(dev_path,
gres_name,  PATH_MAX);
from src/plugins/task/cgroup/task_cgroup_devices.c.

This function assumes that another function
gres_plugin_node_config_load was called before.

I've added gres_plugin_node_config_load call into
task_cgroup_devices_init function and problem disappears. The patch is
in attachment.

Alexander Bersenev, Institute of Mathematics and Mechanics, Russia

--- slurm-2.3.1/src/plugins/task/cgroup/task_cgroup_devices.c   2011-10-24 23:15:42.000000000 +0600
+++ slurm-2.3.1_patched/src/plugins/task/cgroup/task_cgroup_devices.c      2011-12-02 17:21:25.055846044 +0600
@@ -83,6 +83,8 @@
 {
        char release_agent_path[PATH_MAX];
 
+       uint16_t cpunum;
+
        /* initialize cpuinfo internal data */
        if ( xcpuinfo_init() != XCPUINFO_SUCCESS )
                return SLURM_ERROR;
@@ -95,6 +97,15 @@
        release_agent_path[0] = '\0';
        /* initialize allowed_devices_filename */
        cgroup_allowed_devices_file[0] = '\0';
+
+       if(get_procs(&cpunum) != 0) {
+               error("task/cgroup: unable to get a number of CPU");
+               goto error;
+       }
+
+       (void) gres_plugin_node_config_load(cpunum);
 
        strcpy(cgroup_allowed_devices_file, slurm_cgroup_conf->allowed_devices_file);
 

slurm.conf
Description: Binary data

gres.conf
Description: Binary data

cgroup.conf
Description: Binary data