root/lib/pacemaker/pcmk_sched_utilization.c

/* */

DEFINITIONS

1. utilization_value
2. compare_utilization_value
3. pcmk__compare_node_capacities
4. update_utilization_value
5. pcmk__consume_node_capacity
6. pcmk__release_node_capacity
7. check_capacity
8. have_enough_capacity
9. sum_resource_utilization
10. pcmk__ban_insufficient_capacity
11. new_load_stopped_op
12. pcmk__create_utilization_constraints
13. pcmk__show_node_capacities

   1 /*
   2  * Copyright 2014-2024 the Pacemaker project contributors
   3  *
   4  * The version control history for this file may have further details.
   5  *
   6  * This source code is licensed under the GNU General Public License version 2
   7  * or later (GPLv2+) WITHOUT ANY WARRANTY.
   8  */
   9 
  10 #include <crm_internal.h>
  11 
  12 #include <limits.h>                 // INT_MIN, INT_MAX
  13 
  14 #include <crm/common/xml.h>
  15 #include <pacemaker-internal.h>
  16 
  17 #include "libpacemaker_private.h"
  18 
  19 /*!
  20  * \internal
  21  * \brief Get integer utilization from a string
  22  *
  23  * \param[in] s  String representation of a node utilization value
  24  *
  25  * \return Integer equivalent of \p s
  26  * \todo It would make sense to restrict utilization values to nonnegative
  27  *       integers, but the documentation just says "integers" and we didn't
  28  *       restrict them initially, so for backward compatibility, allow any
  29  *       integer.
  30  */
  31 static int
  32 utilization_value(const char *s)
     /*  */
  33 {
  34     int value = 0;
  35 
  36     if ((s != NULL) && (pcmk__scan_min_int(s, &value, INT_MIN) == EINVAL)) {
  37         pcmk__config_warn("Using 0 for utilization instead of "
  38                           "invalid value '%s'", s);
  39         value = 0;
  40     }
  41     return value;
  42 }
  43 
  44 
  45 /*
  46  * Functions for comparing node capacities
  47  */
  48 
  49 struct compare_data {
  50     const pcmk_node_t *node1;
  51     const pcmk_node_t *node2;
  52     bool node2_only;
  53     int result;
  54 };
  55 
  56 /*!
  57  * \internal
  58  * \brief Compare a single utilization attribute for two nodes
  59  *
  60  * Compare one utilization attribute for two nodes, decrementing the result if
  61  * the first node has greater capacity, and incrementing it if the second node
  62  * has greater capacity.
  63  *
  64  * \param[in]     key        Utilization attribute name to compare
  65  * \param[in]     value      Utilization attribute value to compare
  66  * \param[in,out] user_data  Comparison data (as struct compare_data*)
  67  */
  68 static void
  69 compare_utilization_value(gpointer key, gpointer value, gpointer user_data)
     /*  */
  70 {
  71     int node1_capacity = 0;
  72     int node2_capacity = 0;
  73     struct compare_data *data = user_data;
  74     const char *node2_value = NULL;
  75 
  76     if (data->node2_only) {
  77         if (g_hash_table_lookup(data->node1->details->utilization, key)) {
  78             return; // We've already compared this attribute
  79         }
  80     } else {
  81         node1_capacity = utilization_value((const char *) value);
  82     }
  83 
  84     node2_value = g_hash_table_lookup(data->node2->details->utilization, key);
  85     node2_capacity = utilization_value(node2_value);
  86 
  87     if (node1_capacity > node2_capacity) {
  88         data->result--;
  89     } else if (node1_capacity < node2_capacity) {
  90         data->result++;
  91     }
  92 }
  93 
  94 /*!
  95  * \internal
  96  * \brief Compare utilization capacities of two nodes
  97  *
  98  * \param[in] node1  First node to compare
  99  * \param[in] node2  Second node to compare
 100  *
 101  * \return Negative integer if node1 has more free capacity,
 102  *         0 if the capacities are equal, or a positive integer
 103  *         if node2 has more free capacity
 104  */
 105 int
 106 pcmk__compare_node_capacities(const pcmk_node_t *node1,
     /*  */
 107                               const pcmk_node_t *node2)
 108 {
 109     struct compare_data data = {
 110         .node1      = node1,
 111         .node2      = node2,
 112         .node2_only = false,
 113         .result     = 0,
 114     };
 115 
 116     // Compare utilization values that node1 and maybe node2 have
 117     g_hash_table_foreach(node1->details->utilization, compare_utilization_value,
 118                          &data);
 119 
 120     // Compare utilization values that only node2 has
 121     data.node2_only = true;
 122     g_hash_table_foreach(node2->details->utilization, compare_utilization_value,
 123                          &data);
 124 
 125     return data.result;
 126 }
 127 
 128 
 129 /*
 130  * Functions for updating node capacities
 131  */
 132 
 133 struct calculate_data {
 134     GHashTable *current_utilization;
 135     bool plus;
 136 };
 137 
 138 /*!
 139  * \internal
 140  * \brief Update a single utilization attribute with a new value
 141  *
 142  * \param[in]     key        Name of utilization attribute to update
 143  * \param[in]     value      Value to add or substract
 144  * \param[in,out] user_data  Calculation data (as struct calculate_data *)
 145  */
 146 static void
 147 update_utilization_value(gpointer key, gpointer value, gpointer user_data)
     /*  */
 148 {
 149     struct calculate_data *data = user_data;
 150     const char *current = g_hash_table_lookup(data->current_utilization, key);
 151     long long result = utilization_value(current)
 152                        + (data->plus? 1LL : -1LL) * utilization_value(value);
 153 
 154     if (result < INT_MIN) {
 155         result = INT_MIN;
 156     } else if (result > INT_MAX) {
 157         result = INT_MAX;
 158     }
 159     g_hash_table_replace(data->current_utilization,
 160                          strdup(key), pcmk__itoa((int) result));
 161 }
 162 
 163 /*!
 164  * \internal
 165  * \brief Subtract a resource's utilization from node capacity
 166  *
 167  * \param[in,out] current_utilization  Current node utilization attributes
 168  * \param[in]     rsc                  Resource with utilization to subtract
 169  */
 170 void
 171 pcmk__consume_node_capacity(GHashTable *current_utilization,
     /*  */
 172                             const pcmk_resource_t *rsc)
 173 {
 174     struct calculate_data data = {
 175         .current_utilization = current_utilization,
 176         .plus = false,
 177     };
 178 
 179     g_hash_table_foreach(rsc->utilization, update_utilization_value, &data);
 180 }
 181 
 182 /*!
 183  * \internal
 184  * \brief Add a resource's utilization to node capacity
 185  *
 186  * \param[in,out] current_utilization  Current node utilization attributes
 187  * \param[in]     rsc                  Resource with utilization to add
 188  */
 189 void
 190 pcmk__release_node_capacity(GHashTable *current_utilization,
     /*  */
 191                             const pcmk_resource_t *rsc)
 192 {
 193     struct calculate_data data = {
 194         .current_utilization = current_utilization,
 195         .plus = true,
 196     };
 197 
 198     g_hash_table_foreach(rsc->utilization, update_utilization_value, &data);
 199 }
 200 
 201 
 202 /*
 203  * Functions for checking for sufficient node capacity
 204  */
 205 
 206 struct capacity_data {
 207     const pcmk_node_t *node;
 208     const char *rsc_id;
 209     bool is_enough;
 210 };
 211 
 212 /*!
 213  * \internal
 214  * \brief Check whether a single utilization attribute has sufficient capacity
 215  *
 216  * \param[in]     key        Name of utilization attribute to check
 217  * \param[in]     value      Amount of utilization required
 218  * \param[in,out] user_data  Capacity data (as struct capacity_data *)
 219  */
 220 static void
 221 check_capacity(gpointer key, gpointer value, gpointer user_data)
     /*  */
 222 {
 223     int required = 0;
 224     int remaining = 0;
 225     const char *node_value_s = NULL;
 226     struct capacity_data *data = user_data;
 227 
 228     node_value_s = g_hash_table_lookup(data->node->details->utilization, key);
 229 
 230     required = utilization_value(value);
 231     remaining = utilization_value(node_value_s);
 232 
 233     if (required > remaining) {
 234         crm_debug("Remaining capacity for %s on %s (%d) is insufficient "
 235                   "for resource %s usage (%d)",
 236                   (const char *) key, pcmk__node_name(data->node), remaining,
 237                   data->rsc_id, required);
 238         data->is_enough = false;
 239     }
 240 }
 241 
 242 /*!
 243  * \internal
 244  * \brief Check whether a node has sufficient capacity for a resource
 245  *
 246  * \param[in] node         Node to check
 247  * \param[in] rsc_id       ID of resource to check (for debug logs only)
 248  * \param[in] utilization  Required utilization amounts
 249  *
 250  * \return true if node has sufficient capacity for resource, otherwise false
 251  */
 252 static bool
 253 have_enough_capacity(const pcmk_node_t *node, const char *rsc_id,
     /*  */
 254                      GHashTable *utilization)
 255 {
 256     struct capacity_data data = {
 257         .node = node,
 258         .rsc_id = rsc_id,
 259         .is_enough = true,
 260     };
 261 
 262     g_hash_table_foreach(utilization, check_capacity, &data);
 263     return data.is_enough;
 264 }
 265 
 266 /*!
 267  * \internal
 268  * \brief Sum the utilization requirements of a list of resources
 269  *
 270  * \param[in] orig_rsc  Resource being assigned (for logging purposes)
 271  * \param[in] rscs      Resources whose utilization should be summed
 272  *
 273  * \return Newly allocated hash table with sum of all utilization values
 274  * \note It is the caller's responsibility to free the return value using
 275  *       g_hash_table_destroy().
 276  */
 277 static GHashTable *
 278 sum_resource_utilization(const pcmk_resource_t *orig_rsc, GList *rscs)
     /*  */
 279 {
 280     GHashTable *utilization = pcmk__strkey_table(free, free);
 281 
 282     for (GList *iter = rscs; iter != NULL; iter = iter->next) {
 283         pcmk_resource_t *rsc = (pcmk_resource_t *) iter->data;
 284 
 285         rsc->cmds->add_utilization(rsc, orig_rsc, rscs, utilization);
 286     }
 287     return utilization;
 288 }
 289 
 290 /*!
 291  * \internal
 292  * \brief Ban resource from nodes with insufficient utilization capacity
 293  *
 294  * \param[in,out] rsc  Resource to check
 295  *
 296  * \return Allowed node for \p rsc with most spare capacity, if there are no
 297  *         nodes with enough capacity for \p rsc and all its colocated resources
 298  */
 299 const pcmk_node_t *
 300 pcmk__ban_insufficient_capacity(pcmk_resource_t *rsc)
     /*  */
 301 {
 302     bool any_capable = false;
 303     char *rscs_id = NULL;
 304     pcmk_node_t *node = NULL;
 305     const pcmk_node_t *most_capable_node = NULL;
 306     GList *colocated_rscs = NULL;
 307     GHashTable *unassigned_utilization = NULL;
 308     GHashTableIter iter;
 309 
 310     CRM_CHECK(rsc != NULL, return NULL);
 311 
 312     // The default placement strategy ignores utilization
 313     if (pcmk__str_eq(rsc->cluster->placement_strategy, PCMK_VALUE_DEFAULT,
 314                      pcmk__str_casei)) {
 315         return NULL;
 316     }
 317 
 318     // Check whether any resources are colocated with this one
 319     colocated_rscs = rsc->cmds->colocated_resources(rsc, NULL, NULL);
 320     if (colocated_rscs == NULL) {
 321         return NULL;
 322     }
 323 
 324     rscs_id = crm_strdup_printf("%s and its colocated resources", rsc->id);
 325 
 326     // If rsc isn't in the list, add it so we include its utilization
 327     if (g_list_find(colocated_rscs, rsc) == NULL) {
 328         colocated_rscs = g_list_append(colocated_rscs, rsc);
 329     }
 330 
 331     // Sum utilization of colocated resources that haven't been assigned yet
 332     unassigned_utilization = sum_resource_utilization(rsc, colocated_rscs);
 333 
 334     // Check whether any node has enough capacity for all the resources
 335     g_hash_table_iter_init(&iter, rsc->allowed_nodes);
 336     while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) {
 337         if (!pcmk__node_available(node, true, false)) {
 338             continue;
 339         }
 340 
 341         if (have_enough_capacity(node, rscs_id, unassigned_utilization)) {
 342             any_capable = true;
 343         }
 344 
 345         // Keep track of node with most free capacity
 346         if ((most_capable_node == NULL)
 347             || (pcmk__compare_node_capacities(node, most_capable_node) < 0)) {
 348             most_capable_node = node;
 349         }
 350     }
 351 
 352     if (any_capable) {
 353         // If so, ban resource from any node with insufficient capacity
 354         g_hash_table_iter_init(&iter, rsc->allowed_nodes);
 355         while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) {
 356             if (pcmk__node_available(node, true, false)
 357                 && !have_enough_capacity(node, rscs_id,
 358                                          unassigned_utilization)) {
 359                 pcmk__rsc_debug(rsc, "%s does not have enough capacity for %s",
 360                                 pcmk__node_name(node), rscs_id);
 361                 resource_location(rsc, node, -PCMK_SCORE_INFINITY,
 362                                   "__limit_utilization__", rsc->cluster);
 363             }
 364         }
 365         most_capable_node = NULL;
 366 
 367     } else {
 368         // Otherwise, ban from nodes with insufficient capacity for rsc alone
 369         g_hash_table_iter_init(&iter, rsc->allowed_nodes);
 370         while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) {
 371             if (pcmk__node_available(node, true, false)
 372                 && !have_enough_capacity(node, rsc->id, rsc->utilization)) {
 373                 pcmk__rsc_debug(rsc, "%s does not have enough capacity for %s",
 374                                 pcmk__node_name(node), rsc->id);
 375                 resource_location(rsc, node, -PCMK_SCORE_INFINITY,
 376                                   "__limit_utilization__", rsc->cluster);
 377             }
 378         }
 379     }
 380 
 381     g_hash_table_destroy(unassigned_utilization);
 382     g_list_free(colocated_rscs);
 383     free(rscs_id);
 384 
 385     pe__show_node_scores(true, rsc, "Post-utilization", rsc->allowed_nodes,
 386                          rsc->cluster);
 387     return most_capable_node;
 388 }
 389 
 390 /*!
 391  * \internal
 392  * \brief Create a new load_stopped pseudo-op for a node
 393  *
 394  * \param[in,out] node  Node to create op for
 395  *
 396  * \return Newly created load_stopped op
 397  */
 398 static pcmk_action_t *
 399 new_load_stopped_op(pcmk_node_t *node)
     /*  */
 400 {
 401     char *load_stopped_task = crm_strdup_printf(PCMK_ACTION_LOAD_STOPPED "_%s",
 402                                                 node->details->uname);
 403     pcmk_action_t *load_stopped = get_pseudo_op(load_stopped_task,
 404                                               node->details->data_set);
 405 
 406     if (load_stopped->node == NULL) {
 407         load_stopped->node = pe__copy_node(node);
 408         pcmk__clear_action_flags(load_stopped, pcmk_action_optional);
 409     }
 410     free(load_stopped_task);
 411     return load_stopped;
 412 }
 413 
 414 /*!
 415  * \internal
 416  * \brief Create utilization-related internal constraints for a resource
 417  *
 418  * \param[in,out] rsc            Resource to create constraints for
 419  * \param[in]     allowed_nodes  List of allowed next nodes for \p rsc
 420  */
 421 void
 422 pcmk__create_utilization_constraints(pcmk_resource_t *rsc,
     /*  */
 423                                      const GList *allowed_nodes)
 424 {
 425     const GList *iter = NULL;
 426     pcmk_action_t *load_stopped = NULL;
 427 
 428     pcmk__rsc_trace(rsc,
 429                     "Creating utilization constraints for %s - strategy: %s",
 430                     rsc->id, rsc->cluster->placement_strategy);
 431 
 432     // "stop rsc then load_stopped" constraints for current nodes
 433     for (iter = rsc->running_on; iter != NULL; iter = iter->next) {
 434         load_stopped = new_load_stopped_op(iter->data);
 435         pcmk__new_ordering(rsc, stop_key(rsc), NULL, NULL, NULL, load_stopped,
 436                            pcmk__ar_if_on_same_node_or_target, rsc->cluster);
 437     }
 438 
 439     // "load_stopped then start/migrate_to rsc" constraints for allowed nodes
 440     for (iter = allowed_nodes; iter; iter = iter->next) {
 441         load_stopped = new_load_stopped_op(iter->data);
 442         pcmk__new_ordering(NULL, NULL, load_stopped, rsc, start_key(rsc), NULL,
 443                            pcmk__ar_if_on_same_node_or_target, rsc->cluster);
 444         pcmk__new_ordering(NULL, NULL, load_stopped,
 445                            rsc,
 446                            pcmk__op_key(rsc->id, PCMK_ACTION_MIGRATE_TO, 0),
 447                            NULL,
 448                            pcmk__ar_if_on_same_node_or_target, rsc->cluster);
 449     }
 450 }
 451 
 452 /*!
 453  * \internal
 454  * \brief Output node capacities if enabled
 455  *
 456  * \param[in]     desc       Prefix for output
 457  * \param[in,out] scheduler  Scheduler data
 458  */
 459 void
 460 pcmk__show_node_capacities(const char *desc, pcmk_scheduler_t *scheduler)
     /*  */
 461 {
 462     if (!pcmk_is_set(scheduler->flags, pcmk_sched_show_utilization)) {
 463         return;
 464     }
 465     for (const GList *iter = scheduler->nodes;
 466          iter != NULL; iter = iter->next) {
 467         const pcmk_node_t *node = (const pcmk_node_t *) iter->data;
 468         pcmk__output_t *out = scheduler->priv;
 469 
 470         out->message(out, "node-capacity", node, desc);
 471     }
 472 }

/* */