A Slurm plugin is a dynamically linked code object which is loaded explicitly at run time by the Slurm libraries. A plugin provides a customized implementation of a well-defined API connected to tasks such as authentication, interconnect fabric, and task scheduling.
A Slurm plugin identifies itself by a short character string formatted similarly to a MIME type: <major>/<minor>. The major type identifies which API the plugin implements. The minor type uniquely distinguishes a plugin from other plugins that implement that same API, by such means as the intended platform or the internal algorithm. For example, a plugin to interface to the Maui scheduler would give its type as "sched/maui." It would implement the Slurm Scheduler API.
Slurm plugin version numbers comprise a major, minor and micro revision number. If the major and/or minor revision number changes, this indicates major changes to the Slurm functionality including changes to APIs, command options, and plugins. These plugin changes may include new functions and/or function arguments. If only the micro revision number changes, this is indicative of bug fixes and possibly minor enhancements which should not adversely impact users. In all cases, rebuilding and installing all Slurm plugins is recommended at upgrade time. Not all compute nodes in a cluster need be updated at the same time, but all Slurm APIs, commands, plugins, etc. on a compute node should represent the same version of Slurm.
A plugin must define and export the following symbols:
- char plugin_type
a unique, short, formatted string to identify the plugin's purpose as described above. A "null" plugin (i.e., one that implements the desired API as stubs) should have a minor type of "none."
- char plugin_name
a free-form string that identifies the plugin in human-readable terms, such as "Kerberos authentication." Slurm will use this string to identify the plugin to end users.
A plugin may optionally define and export the following symbols:
- const uint32_t plugin_version
If specified, identifies the version of Slurm used to build this plugin and any attempt to load the plugin from a different version of Slurm will result in an error. If not specified, then the plugin may be loadeed by Slurm commands and daemons from any version, however this may result in difficult to diagnose failures due to changes in the arguments to plugin functions or changes in other Slurm functions used by the plugin.
API Functions in All Plugins
int init (void);
Description: If present, this function is called just after the plugin is loaded. This allows the plugin to perform any global initialization prior to any actual API calls.
Returns: SLURM_SUCCESS if the plugin's initialization was successful. Any other return value indicates to Slurm that the plugin should be unloaded and not used.
void fini (void);
Description: If present, this function is called just before the plugin is unloaded. This allows the plugin to do any finalization after the last plugin-specific API call is made.
Note: These init and fini functions are not the same as those described in the dlopen (3) system library. The C run-time system co-opts those symbols for its own initialization. The system _init() is called before the Slurm init(), and the Slurm fini() is called before the system's _fini().
The functions need not appear. The plugin may provide either init() or fini() or both.
Slurm is a multithreaded application. The Slurm plugin library may exercise the plugin functions in a re-entrant fashion. It is the responsibility of the plugin author to provide the necessarily mutual exclusion and synchronization in order to avoid the pitfalls of re-entrant code.
The standard system libraries are available to the plugin. The Slurm libraries are also available and plugin authors are encouraged to make use of them rather than develop their own substitutes. Plugins should use the Slurm log to print error messages.
The plugin author is responsible for specifying any specific non-standard libraries needed for correct operation. Plugins will not load if their dependent libraries are not available, so it is the installer's job to make sure the specified libraries are available.
All plugin functions are expected to execute very quickly. If any function entails delays (e.g. transactions with other systems), it should be written to utilize a thread for that functionality. This thread may be created by the init() function and deleted by the fini() functions. See plugins/sched/backfill for an example of how to do this.
Data Structure Consistency
In certain situations Slurm iterates over different data structures elements using counters. For example, with environment variable arrays. In order to avoid buffer overflows and other undesired situations, when a plugin modifies certain elements it must also update these counters accordingly. Other situations may require other types of changes.
The following advice indicates which structures have arrays with associated counters that must be mantained when modifying data, plus other possible important information to take in consideration when manipulating these structures. This list is not fully exhaustive due to constant modifications in code, but it is a first start point and basic guideline for most common situations. Complete structure inforation can be seen in the slurm/slurm.h.in file.
slurm_job_info_t (job_info_t) Data Structure
uint32_t env_size; char **environment; uint32_t spank_job_env_size; char **spank_job_env; uint32_t gres_detail_cnt; char **gres_detail_str;
These pairs of array pointers and element counters must kept updated in order to avoid subsequent buffer overflows, so if you update the array you must also update the related counter.
char *nodes; int32_t *node_inx; int32_t *req_node_inx; char *req_nodes;
node_inx and req_node_inx represents a list of index pairs for ranges of nodes defined in the nodes and req_nodes fields respectively. In each case, both array variables must match the count.
uint32_t pack_job_id; char *pack_job_id_set;
The pack_job_id field should be the first element of the pack_job_id_set array.
job_step_info_t Data Structure
char *nodes; int32_t *node_inx;
node_inx represents a list of index pairs for range of nodes defined in nodes. Both variables must match the node count.
priority_factors_object_t Data Structure
uint32_t tres_cnt; char **tres_names; double *tres_weights;
This value must match the configured TRES on the system, otherwise iteration over the tres_names or tres_weights arrays can cause buffer overflows.
job_step_pids_t Data Structure
uint32_t pid_cnt; uint32_t *pid;
Array pid represents the list of Process IDs for the job step, and pid_cnt is the counter that must match the size of the array.
slurm_step_layout_t Data Structure
uint32_t node_cnt; char *node_list;
The node_list array size must match node_cnt.
uint16_t *tasks; uint32_t node_cnt; uint32_t task_cnt;
In the tasks array, each element is the number of tasks assigned to the corresponding node, to its size must match node_cnt. Moreover task_cnt represents the sum of tasks registered in tasks.
tids is an array of length node_cnt of task ID arrays. Each subarray is designated by the corresponding value in the tasks array, so tasks, tids and task_cnt must be set to match this layout.
slurm_step_launch_params_t Data Structure
uint32_t envc; char **env;
When modifying the environment variables in the env array, you must also modify the envc counter accordingly to prevent buffer overflows in subsequent loops over that array.
uint32_t pack_nnodes; uint32_t pack_ntasks; uint16_t *pack_task_cnts; uint32_t **pack_tids; uint32_t *pack_node_list;
This pack_* related variables must match the current heterogeneous job
For example, if for whatever reason you are reducing the number of tasks for a node in a heterogeneous job, you should at least remove that task ID from pack_tids, decrement pack_ntasks and pack_task_cnts, and possibly decrement the number of nodes of the heterogeneous job in pack_nnodes and pack_node_list.
char **spank_job_env; uint32_t spank_job_env_size;
When modifying the spank_job_env structure, the spank_job_env_size field must be updated to prevent buffer overflows in subsequent loops over that array.
node_info_t Data Structure
char *features; char *features_act;
In a system containing Intel KNL processors the features_act field is set by the plugin to match the currently running modes on the node. On other systems the features_act is not usually used. If you program such a plugin you must ensure that features_act contains a subset of features.
char *reason; time_t reason_time; uint32_t reason_uid;
If reason is modified then reason_time and reason_uid should be updated.
reserve_info_t Data Structure
int32_t *node_inx; uint32_t node_cnt;
node_inx represents a list of index pairs for range of nodes associated with the reservation and its count must equal node_cnt.
partition_info_t Data Structure
No special advice.
slurm_step_layout_req_t Data Structure
No special advice.
No special advice.
Last modified 20 December 2017