diff --git a/src/common/gres.c b/src/common/gres.c
index b6cdbec92bbd0cb88223af33d204a185b56a05d3..eb648f58423115fbaf2afc3d1346c26b142c406f 100644
--- a/src/common/gres.c
+++ b/src/common/gres.c
@@ -1720,7 +1720,7 @@ static void _node_state_dealloc(gres_state_t *gres_ptr)
break;
}
}
- error("gres_plugin_node_state_dealloc: gres/%s topo_cnt!=0 "
+ error("gres_plugin_node_state_dealloc_all: gres/%s topo_cnt!=0 "
"and topo_gres_cnt_alloc is NULL", gres_name);
} else {
for (i=0; i<gres_node_ptr->topo_cnt; i++) {
@@ -1735,7 +1735,7 @@ static void _node_state_dealloc(gres_state_t *gres_ptr)
* is reconfigured.
* IN gres_list - node gres state information
*/
-extern void gres_plugin_node_state_dealloc(List gres_list)
+extern void gres_plugin_node_state_dealloc_all(List gres_list)
{
ListIterator gres_iter;
gres_state_t *gres_ptr;
@@ -1754,120 +1754,6 @@ extern void gres_plugin_node_state_dealloc(List gres_list)
slurm_mutex_unlock(&gres_context_lock);
}
-static int _node_state_realloc(void *job_gres_data, int node_offset,
- void *node_gres_data, char *gres_name)
-{
- gres_job_state_t *job_gres_ptr = (gres_job_state_t *) job_gres_data;
- gres_node_state_t *node_gres_ptr = (gres_node_state_t *) node_gres_data;
- int i, job_bit_size, node_bit_size;
-
- xassert(job_gres_ptr);
- xassert(node_gres_ptr);
-
- if (node_offset >= job_gres_ptr->node_cnt) {
- error("gres: %s job node offset is bad (%d >= %u)",
- gres_name, node_offset, job_gres_ptr->node_cnt);
- return EINVAL;
- }
-
- node_gres_ptr->gres_cnt_alloc += job_gres_ptr->gres_cnt_alloc;
-
- if (node_gres_ptr->gres_bit_alloc && job_gres_ptr->gres_bit_alloc &&
- job_gres_ptr->gres_bit_alloc[node_offset]) {
- job_bit_size = bit_size(job_gres_ptr->
- gres_bit_alloc[node_offset]);
- node_bit_size = bit_size(node_gres_ptr->gres_bit_alloc);
- if (job_bit_size > node_bit_size) {
- error("gres/%s: job/node bit size mismatch (%d != %d)",
- gres_name, job_bit_size, node_bit_size);
- /* Node needs to register with more resources, expand
- * node's bitmap now so we can merge the data */
- node_gres_ptr->gres_bit_alloc =
- bit_realloc(node_gres_ptr->gres_bit_alloc,
- job_bit_size);
- if (node_gres_ptr->gres_bit_alloc == NULL)
- fatal("bit_realloc: malloc failure");
- node_bit_size = job_bit_size;
- }
- if (job_bit_size < node_bit_size) {
- error("gres/%s: job/node bit size mismatch (%d != %d)",
- gres_name, job_bit_size, node_bit_size);
- /* Update what we can */
- node_bit_size = MIN(job_bit_size, node_bit_size);
- for (i=0; i<node_bit_size; i++) {
- if (!bit_test(job_gres_ptr->
- gres_bit_alloc[node_offset], i))
- continue;
- node_gres_ptr->gres_cnt_alloc++;
- bit_set(node_gres_ptr->gres_bit_alloc, i);
- }
- } else {
- bit_or(node_gres_ptr->gres_bit_alloc,
- job_gres_ptr->gres_bit_alloc[node_offset]);
- node_gres_ptr->gres_cnt_alloc =
- bit_set_count(node_gres_ptr->gres_bit_alloc);
- }
- }
-
- return SLURM_SUCCESS;
-}
-
-/*
- * Allocate in this nodes record the resources previously allocated to this
- * job. This function isused to synchronize state after slurmctld restarts
- * or is reconfigured.
- * IN job_gres_list - job gres state information
- * IN node_offset - zero-origin index of this node in the job's allocation
- * IN node_gres_list - node gres state information
- * RET SLURM_SUCCESS or error code
- */
-extern int gres_plugin_node_state_realloc(List job_gres_list, int node_offset,
- List node_gres_list)
-{
- ListIterator job_gres_iter, node_gres_iter;
- gres_state_t *job_gres_ptr, *node_gres_ptr;
- int i;
-
- if (job_gres_list == NULL)
- return SLURM_SUCCESS;
- if (node_gres_list == NULL)
- return SLURM_ERROR;
-
- (void) gres_plugin_init();
-
- slurm_mutex_lock(&gres_context_lock);
- job_gres_iter = list_iterator_create(job_gres_list);
- while ((job_gres_ptr = (gres_state_t *) list_next(job_gres_iter))) {
- node_gres_iter = list_iterator_create(node_gres_list);
- while ((node_gres_ptr = (gres_state_t *)
- list_next(node_gres_iter))) {
- if (job_gres_ptr->plugin_id == node_gres_ptr->plugin_id)
- break;
- }
- list_iterator_destroy(node_gres_iter);
- if (node_gres_ptr == NULL) {
- error("Could not find plugin id %u to realloc job",
- job_gres_ptr->plugin_id);
- continue;
- }
-
- for (i=0; i<gres_context_cnt; i++) {
- if (job_gres_ptr->plugin_id !=
- gres_context[i].plugin_id)
- continue;
- _node_state_realloc(job_gres_ptr->gres_data,
- node_offset,
- node_gres_ptr->gres_data,
- gres_context[i].gres_name);
- break;
- }
- }
- list_iterator_destroy(job_gres_iter);
- slurm_mutex_unlock(&gres_context_lock);
-
- return SLURM_SUCCESS;
-}
-
static void _node_state_log(void *gres_data, char *node_name, char *gres_name)
{
gres_node_state_t *gres_node_ptr;
@@ -2146,12 +2032,15 @@ List gres_plugin_job_state_dup(List gres_list)
* IN gres_list - generated by gres_plugin_job_config_validate()
* IN/OUT buffer - location to write state to
* IN job_id - job's ID
+ * IN details - if set then pack job step allocation details (only needed to
+ * save/restore job state, not needed in job credential for
+ * slurmd task binding)
*
* NOTE: A job's allocation to steps is not recorded here, but recovered with
* the job step state information upon slurmctld restart.
*/
extern int gres_plugin_job_state_pack(List gres_list, Buf buffer,
- uint32_t job_id)
+ uint32_t job_id, bool details)
{
int i, rc = SLURM_SUCCESS;
uint32_t top_offset, tail_offset;
@@ -2186,6 +2075,24 @@ extern int gres_plugin_job_state_pack(List gres_list, Buf buffer,
} else {
pack8((uint8_t) 0, buffer);
}
+ if (details && gres_job_ptr->gres_bit_step_alloc) {
+ pack8((uint8_t) 1, buffer);
+ for (i=0; i<gres_job_ptr->node_cnt; i++) {
+ pack_bit_str(gres_job_ptr->
+ gres_bit_step_alloc[i], buffer);
+ }
+ } else {
+ pack8((uint8_t) 0, buffer);
+ }
+ if (details && gres_job_ptr->gres_cnt_step_alloc) {
+ pack8((uint8_t) 1, buffer);
+ for (i=0; i<gres_job_ptr->node_cnt; i++) {
+ pack32(gres_job_ptr->gres_cnt_step_alloc[i],
+ buffer);
+ }
+ } else {
+ pack8((uint8_t) 0, buffer);
+ }
rec_cnt++;
}
list_iterator_destroy(gres_iter);
@@ -2211,7 +2118,7 @@ extern int gres_plugin_job_state_unpack(List *gres_list, Buf buffer,
int i, rc;
uint32_t magic, plugin_id;
uint16_t rec_cnt;
- uint8_t has_bitmap;
+ uint8_t has_more;
gres_state_t *gres_ptr;
gres_job_state_t *gres_job_ptr = NULL;
@@ -2239,8 +2146,8 @@ extern int gres_plugin_job_state_unpack(List *gres_list, Buf buffer,
gres_job_ptr = xmalloc(sizeof(gres_job_state_t));
safe_unpack32(&gres_job_ptr->gres_cnt_alloc, buffer);
safe_unpack32(&gres_job_ptr->node_cnt, buffer);
- safe_unpack8(&has_bitmap, buffer);
- if (has_bitmap) {
+ safe_unpack8(&has_more, buffer);
+ if (has_more) {
gres_job_ptr->gres_bit_alloc =
xmalloc(sizeof(bitstr_t *) *
gres_job_ptr->node_cnt);
@@ -2249,6 +2156,26 @@ extern int gres_plugin_job_state_unpack(List *gres_list, Buf buffer,
buffer);
}
}
+ safe_unpack8(&has_more, buffer);
+ if (has_more) {
+ gres_job_ptr->gres_bit_step_alloc =
+ xmalloc(sizeof(bitstr_t *) *
+ gres_job_ptr->node_cnt);
+ for (i=0; i<gres_job_ptr->node_cnt; i++) {
+ unpack_bit_str(&gres_job_ptr->
+ gres_bit_step_alloc[i], buffer);
+ }
+ }
+ safe_unpack8(&has_more, buffer);
+ if (has_more) {
+ gres_job_ptr->gres_cnt_step_alloc =
+ xmalloc(sizeof(uint32_t) *
+ gres_job_ptr->node_cnt);
+ for (i=0; i<gres_job_ptr->node_cnt; i++) {
+ safe_unpack32(&gres_job_ptr->
+ gres_cnt_step_alloc[i], buffer);
+ }
+ }
for (i=0; i<gres_context_cnt; i++) {
if (gres_context[i].plugin_id == plugin_id)
break;
diff --git a/src/common/gres.h b/src/common/gres.h
index a4a91ad0b27148e4a0f06fb67745991e259608b1..5460da019b909df32cd34ebb1cde59f69bacd99b 100644
--- a/src/common/gres.h
+++ b/src/common/gres.h
@@ -252,19 +252,7 @@ extern List gres_plugin_node_state_dup(List gres_list);
* is reconfigured.
* IN gres_list - node gres state information
*/
-extern void gres_plugin_node_state_dealloc(List gres_list);
-
-/*
- * Allocate in this nodes record the resources previously allocated to this
- * job. This function isused to synchronize state after slurmctld restarts
- * or is reconfigured.
- * IN job_gres_list - job gres state information
- * IN node_offset - zero-origin index of this node in the job's allocation
- * IN node_gres_list - node gres state information
- * RET SLURM_SUCCESS or error code
- */
-extern int gres_plugin_node_state_realloc(List job_gres_list, int node_offset,
- List node_gres_list);
+extern void gres_plugin_node_state_dealloc_all(List gres_list);
/*
* Log a node's current gres state
@@ -295,12 +283,15 @@ List gres_plugin_job_state_dup(List gres_list);
* IN gres_list - generated by gres_plugin_job_config_validate()
* IN/OUT buffer - location to write state to
* IN job_id - job's ID
+ * IN details - if set then pack job step allocation details (only needed to
+ * save/restore job state, not needed in job credential for
+ * slurmd task binding)
*
* NOTE: A job's allocation to steps is not recorded here, but recovered with
* the job step state information upon slurmctld restart.
*/
extern int gres_plugin_job_state_pack(List gres_list, Buf buffer,
- uint32_t job_id);
+ uint32_t job_id, bool details);
/*
* Unpack a job's current gres status, called from slurmctld for save/restore
diff --git a/src/common/slurm_cred.c b/src/common/slurm_cred.c
index a382e5a7a866c2140458eb38a8560e65ca2e756b..324956376ae20190a90e5ce4995c1baccb8ab1a6 100644
--- a/src/common/slurm_cred.c
+++ b/src/common/slurm_cred.c
@@ -1645,7 +1645,8 @@ _pack_cred(slurm_cred_t *cred, Buf buffer)
pack32(cred->stepid, buffer);
pack32(cred_uid, buffer);
- gres_plugin_job_state_pack(cred->job_gres_list, buffer, cred->jobid);
+ (void) gres_plugin_job_state_pack(cred->job_gres_list, buffer,
+ cred->jobid, false);
gres_plugin_step_state_pack(cred->step_gres_list, buffer,
cred->jobid, cred->stepid);
pack32(cred->job_mem_limit, buffer);
diff --git a/src/plugins/select/cons_res/select_cons_res.c b/src/plugins/select/cons_res/select_cons_res.c
index 55a305a71d17544018daa7cb9a8f57e05e469745..5fb2364d8e467c7a69bc55a1d5d440214e332234 100644
--- a/src/plugins/select/cons_res/select_cons_res.c
+++ b/src/plugins/select/cons_res/select_cons_res.c
@@ -1701,8 +1701,8 @@ extern int select_p_node_init(struct node_record *node_ptr, int node_cnt)
real_memory;
}
select_node_usage[i].node_state = NODE_CR_AVAILABLE;
- gres_plugin_node_state_dealloc(select_node_record[i].node_ptr->
- gres_list);
+ gres_plugin_node_state_dealloc_all(select_node_record[i].
+ node_ptr->gres_list);
}
_create_part_data();
diff --git a/src/plugins/select/linear/select_linear.c b/src/plugins/select/linear/select_linear.c
index 5a8782a6da8ff7f8328bc29fd041bd65c421bd7a..89aad0fc0d5d03614059c17b090926a08c68fe1f 100644
--- a/src/plugins/select/linear/select_linear.c
+++ b/src/plugins/select/linear/select_linear.c
@@ -1848,7 +1848,7 @@ static void _init_node_cr(void)
/* Clear existing node Gres allocations */
for (i = 0, node_ptr = node_record_table_ptr; i < node_record_count;
i++, node_ptr++) {
- gres_plugin_node_state_dealloc(node_ptr->gres_list);
+ gres_plugin_node_state_dealloc_all(node_ptr->gres_list);
}
/* record running and suspended jobs in node_cr_records */
@@ -1913,9 +1913,12 @@ static void _init_node_cr(void)
}
if (bit_test(job_ptr->node_bitmap, i)) {
- gres_plugin_node_state_realloc(
- job_ptr->gres_list, node_offset,
- node_ptr->gres_list);
+ gres_plugin_job_alloc(job_ptr->gres_list,
+ node_ptr->gres_list,
+ job_resrcs_ptr->nhosts,
+ node_offset,
+ job_resrcs_ptr->
+ cpus[node_offset]);
}
part_cr_ptr = cr_ptr->nodes[i].parts;
diff --git a/src/slurmctld/job_mgr.c b/src/slurmctld/job_mgr.c
index a9fb55abd7e3c41c51ed4bffbada529a31905c28..c47f449cd5a4d68109ec469ce4184495e8ccd5a2 100644
--- a/src/slurmctld/job_mgr.c
+++ b/src/slurmctld/job_mgr.c
@@ -787,7 +787,7 @@ static void _dump_job_state(struct job_record *dump_job_ptr, Buf buffer)
dump_job_ptr->spank_job_env_size, buffer);
(void) gres_plugin_job_state_pack(dump_job_ptr->gres_list, buffer,
- dump_job_ptr->job_id);
+ dump_job_ptr->job_id, true);
/* Dump job details, if available */
detail_ptr = dump_job_ptr->details;