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 2014-2022 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 #include <crm/msg_xml.h>
  12 #include <pacemaker-internal.h>
  13 
  14 #include "libpacemaker_private.h"
  15 
  16 // Name for a pseudo-op to use in ordering constraints for utilization
  17 #define LOAD_STOPPED "load_stopped"
  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         pe_warn("Using 0 for utilization instead of invalid value '%s'", value);
  38         value = 0;
  39     }
  40     return value;
  41 }
  42 
  43 
  44 /*
  45  * Functions for comparing node capacities
  46  */
  47 
  48 struct compare_data {
  49     const pe_node_t *node1;
  50     const pe_node_t *node2;
  51     bool node2_only;
  52     int result;
  53 };
  54 
  55 /*!
  56  * \internal
  57  * \brief Compare a single utilization attribute for two nodes
  58  *
  59  * Compare one utilization attribute for two nodes, incrementing the result if
  60  * the first node has greater capacity, and decrementing it if the second node
  61  * has greater capacity.
  62  *
  63  * \param[in] key        Utilization attribute name to compare
  64  * \param[in] value      Utilization attribute value to compare
  65  * \param[in] user_data  Comparison data (as struct compare_data*)
  66  */
  67 static void
  68 compare_utilization_value(gpointer key, gpointer value, gpointer user_data)
     /*  */
  69 {
  70     int node1_capacity = 0;
  71     int node2_capacity = 0;
  72     struct compare_data *data = user_data;
  73     const char *node2_value = NULL;
  74 
  75     if (data->node2_only) {
  76         if (g_hash_table_lookup(data->node1->details->utilization, key)) {
  77             return; // We've already compared this attribute
  78         }
  79     } else {
  80         node1_capacity = utilization_value((const char *) value);
  81     }
  82 
  83     node2_value = g_hash_table_lookup(data->node2->details->utilization, key);
  84     node2_capacity = utilization_value(node2_value);
  85 
  86     if (node1_capacity > node2_capacity) {
  87         data->result--;
  88     } else if (node1_capacity < node2_capacity) {
  89         data->result++;
  90     }
  91 }
  92 
  93 /*!
  94  * \internal
  95  * \brief Compare utilization capacities of two nodes
  96  *
  97  * \param[in] node1  First node to compare
  98  * \param[in] node2  Second node to compare
  99  *
 100  * \return Negative integer if node1 has more free capacity,
 101  *         0 if the capacities are equal, or a positive integer
 102  *         if node2 has more free capacity
 103  */
 104 int
 105 pcmk__compare_node_capacities(const pe_node_t *node1, const pe_node_t *node2)
     /*  */
 106 {
 107     struct compare_data data = {
 108         .node1      = node1,
 109         .node2      = node2,
 110         .node2_only = false,
 111         .result     = 0,
 112     };
 113 
 114     // Compare utilization values that node1 and maybe node2 have
 115     g_hash_table_foreach(node1->details->utilization, compare_utilization_value,
 116                          &data);
 117 
 118     // Compare utilization values that only node2 has
 119     data.node2_only = true;
 120     g_hash_table_foreach(node2->details->utilization, compare_utilization_value,
 121                          &data);
 122 
 123     return data.result;
 124 }
 125 
 126 
 127 /*
 128  * Functions for updating node capacities
 129  */
 130 
 131 struct calculate_data {
 132     GHashTable *current_utilization;
 133     bool plus;
 134 };
 135 
 136 /*!
 137  * \internal
 138  * \brief Update a single utilization attribute with a new value
 139  *
 140  * \param[in] key        Name of utilization attribute to update
 141  * \param[in] value      Value to add or substract
 142  * \param[in] user_data  Calculation data (as struct calculate_data *)
 143  */
 144 static void
 145 update_utilization_value(gpointer key, gpointer value, gpointer user_data)
     /*  */
 146 {
 147     int result = 0;
 148     const char *current = NULL;
 149     struct calculate_data *data = user_data;
 150 
 151     current = g_hash_table_lookup(data->current_utilization, key);
 152     if (data->plus) {
 153         result = utilization_value(current) + utilization_value(value);
 154     } else if (current) {
 155         result = utilization_value(current) - utilization_value(value);
 156     }
 157     g_hash_table_replace(data->current_utilization,
 158                          strdup(key), pcmk__itoa(result));
 159 }
 160 
 161 /*!
 162  * \internal
 163  * \brief Subtract a resource's utilization from node capacity
 164  *
 165  * \param[in] current_utilization  Current node utilization attributes
 166  * \param[in] rsc                  Resource with utilization to subtract
 167  */
 168 void
 169 pcmk__consume_node_capacity(GHashTable *current_utilization, pe_resource_t *rsc)
     /*  */
 170 {
 171     struct calculate_data data = {
 172         .current_utilization = current_utilization,
 173         .plus = false,
 174     };
 175 
 176     g_hash_table_foreach(rsc->utilization, update_utilization_value, &data);
 177 }
 178 
 179 /*!
 180  * \internal
 181  * \brief Add a resource's utilization to node capacity
 182  *
 183  * \param[in] current_utilization  Current node utilization attributes
 184  * \param[in] rsc                  Resource with utilization to add
 185  */
 186 void
 187 pcmk__release_node_capacity(GHashTable *current_utilization,
     /*  */
 188                             const pe_resource_t *rsc)
 189 {
 190     struct calculate_data data = {
 191         .current_utilization = current_utilization,
 192         .plus = true,
 193     };
 194 
 195     g_hash_table_foreach(rsc->utilization, update_utilization_value, &data);
 196 }
 197 
 198 
 199 /*
 200  * Functions for checking for sufficient node capacity
 201  */
 202 
 203 struct capacity_data {
 204     pe_node_t *node;
 205     const char *rsc_id;
 206     bool is_enough;
 207 };
 208 
 209 /*!
 210  * \internal
 211  * \brief Check whether a single utilization attribute has sufficient capacity
 212  *
 213  * \param[in] key        Name of utilization attribute to check
 214  * \param[in] value      Amount of utilization required
 215  * \param[in] user_data  Capacity data (as struct capacity_data *)
 216  */
 217 static void
 218 check_capacity(gpointer key, gpointer value, gpointer user_data)
     /*  */
 219 {
 220     int required = 0;
 221     int remaining = 0;
 222     const char *node_value_s = NULL;
 223     struct capacity_data *data = user_data;
 224 
 225     node_value_s = g_hash_table_lookup(data->node->details->utilization, key);
 226 
 227     required = utilization_value(value);
 228     remaining = utilization_value(node_value_s);
 229 
 230     if (required > remaining) {
 231         crm_debug("Remaining capacity for %s on %s (%d) is insufficient "
 232                   "for resource %s usage (%d)",
 233                   (const char *) key, pe__node_name(data->node), remaining,
 234                   data->rsc_id, required);
 235         data->is_enough = false;
 236     }
 237 }
 238 
 239 /*!
 240  * \internal
 241  * \brief Check whether a node has sufficient capacity for a resource
 242  *
 243  * \param[in] node         Node to check
 244  * \param[in] rsc_id       ID of resource to check (for debug logs only)
 245  * \param[in] utilization  Required utilization amounts
 246  *
 247  * \return true if node has sufficient capacity for resource, otherwise false
 248  */
 249 static bool
 250 have_enough_capacity(pe_node_t *node, const char *rsc_id,
     /*  */
 251                      GHashTable *utilization)
 252 {
 253     struct capacity_data data = {
 254         .node = node,
 255         .rsc_id = rsc_id,
 256         .is_enough = true,
 257     };
 258 
 259     g_hash_table_foreach(utilization, check_capacity, &data);
 260     return data.is_enough;
 261 }
 262 
 263 /*!
 264  * \internal
 265  * \brief Sum the utilization requirements of a list of resources
 266  *
 267  * \param[in] orig_rsc  Resource being allocated (for logging purposes)
 268  * \param[in] rscs      Resources whose utilization should be summed
 269  *
 270  * \return Newly allocated hash table with sum of all utilization values
 271  * \note It is the caller's responsibility to free the return value using
 272  *       g_hash_table_destroy().
 273  */
 274 static GHashTable *
 275 sum_resource_utilization(pe_resource_t *orig_rsc, GList *rscs)
     /*  */
 276 {
 277     GHashTable *utilization = pcmk__strkey_table(free, free);
 278 
 279     for (GList *iter = rscs; iter != NULL; iter = iter->next) {
 280         pe_resource_t *rsc = (pe_resource_t *) iter->data;
 281 
 282         rsc->cmds->add_utilization(rsc, orig_rsc, rscs, utilization);
 283     }
 284     return utilization;
 285 }
 286 
 287 /*!
 288  * \internal
 289  * \brief Ban resource from nodes with insufficient utilization capacity
 290  *
 291  * \param[in] rsc  Resource to check
 292  *
 293  * \return Allowed node for \p rsc with most spare capacity, if there are no
 294  *         nodes with enough capacity for \p rsc and all its colocated resources
 295  */
 296 const pe_node_t *
 297 pcmk__ban_insufficient_capacity(pe_resource_t *rsc)
     /*  */
 298 {
 299     bool any_capable = false;
 300     char *rscs_id = NULL;
 301     pe_node_t *node = NULL;
 302     const pe_node_t *most_capable_node = NULL;
 303     GList *colocated_rscs = NULL;
 304     GHashTable *unallocated_utilization = NULL;
 305     GHashTableIter iter;
 306 
 307     CRM_CHECK(rsc != NULL, return NULL);
 308 
 309     // The default placement strategy ignores utilization
 310     if (pcmk__str_eq(rsc->cluster->placement_strategy, "default",
 311                      pcmk__str_casei)) {
 312         return NULL;
 313     }
 314 
 315     // Check whether any resources are colocated with this one
 316     colocated_rscs = rsc->cmds->colocated_resources(rsc, NULL, NULL);
 317     if (colocated_rscs == NULL) {
 318         return NULL;
 319     }
 320 
 321     rscs_id = crm_strdup_printf("%s and its colocated resources", rsc->id);
 322 
 323     // If rsc isn't in the list, add it so we include its utilization
 324     if (g_list_find(colocated_rscs, rsc) == NULL) {
 325         colocated_rscs = g_list_append(colocated_rscs, rsc);
 326     }
 327 
 328     // Sum utilization of colocated resources that haven't been allocated yet
 329     unallocated_utilization = sum_resource_utilization(rsc, colocated_rscs);
 330 
 331     // Check whether any node has enough capacity for all the resources
 332     g_hash_table_iter_init(&iter, rsc->allowed_nodes);
 333     while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) {
 334         if (!pcmk__node_available(node, true, false)) {
 335             continue;
 336         }
 337 
 338         if (have_enough_capacity(node, rscs_id, unallocated_utilization)) {
 339             any_capable = true;
 340         }
 341 
 342         // Keep track of node with most free capacity
 343         if ((most_capable_node == NULL)
 344             || (pcmk__compare_node_capacities(node, most_capable_node) < 0)) {
 345             most_capable_node = node;
 346         }
 347     }
 348 
 349     if (any_capable) {
 350         // If so, ban resource from any node with insufficient capacity
 351         g_hash_table_iter_init(&iter, rsc->allowed_nodes);
 352         while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) {
 353             if (pcmk__node_available(node, true, false)
 354                 && !have_enough_capacity(node, rscs_id,
 355                                          unallocated_utilization)) {
 356                 pe_rsc_debug(rsc, "%s does not have enough capacity for %s",
 357                              pe__node_name(node), rscs_id);
 358                 resource_location(rsc, node, -INFINITY, "__limit_utilization__",
 359                                   rsc->cluster);
 360             }
 361         }
 362         most_capable_node = NULL;
 363 
 364     } else {
 365         // Otherwise, ban from nodes with insufficient capacity for rsc alone
 366         g_hash_table_iter_init(&iter, rsc->allowed_nodes);
 367         while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) {
 368             if (pcmk__node_available(node, true, false)
 369                 && !have_enough_capacity(node, rsc->id, rsc->utilization)) {
 370                 pe_rsc_debug(rsc, "%s does not have enough capacity for %s",
 371                              pe__node_name(node), rsc->id);
 372                 resource_location(rsc, node, -INFINITY, "__limit_utilization__",
 373                                   rsc->cluster);
 374             }
 375         }
 376     }
 377 
 378     g_hash_table_destroy(unallocated_utilization);
 379     g_list_free(colocated_rscs);
 380     free(rscs_id);
 381 
 382     pe__show_node_weights(true, rsc, "Post-utilization",
 383                           rsc->allowed_nodes, rsc->cluster);
 384     return most_capable_node;
 385 }
 386 
 387 /*!
 388  * \internal
 389  * \brief Create a new load_stopped pseudo-op for a node
 390  *
 391  * \param[in] node      Node to create op for
 392  * \param[in] data_set  Cluster working set
 393  *
 394  * \return Newly created load_stopped op
 395  */
 396 static pe_action_t *
 397 new_load_stopped_op(const pe_node_t *node, pe_working_set_t *data_set)
     /*  */
 398 {
 399     char *load_stopped_task = crm_strdup_printf(LOAD_STOPPED "_%s",
 400                                                 node->details->uname);
 401     pe_action_t *load_stopped = get_pseudo_op(load_stopped_task, data_set);
 402 
 403     if (load_stopped->node == NULL) {
 404         load_stopped->node = pe__copy_node(node);
 405         pe__clear_action_flags(load_stopped, pe_action_optional);
 406     }
 407     free(load_stopped_task);
 408     return load_stopped;
 409 }
 410 
 411 /*!
 412  * \internal
 413  * \brief Create utilization-related internal constraints for a resource
 414  *
 415  * \param[in] rsc            Resource to create constraints for
 416  * \param[in] allowed_nodes  List of allowed next nodes for \p rsc
 417  */
 418 void
 419 pcmk__create_utilization_constraints(pe_resource_t *rsc, GList *allowed_nodes)
     /*  */
 420 {
 421     GList *iter = NULL;
 422     pe_node_t *node = NULL;
 423     pe_action_t *load_stopped = NULL;
 424 
 425     pe_rsc_trace(rsc, "Creating utilization constraints for %s - strategy: %s",
 426                  rsc->id, rsc->cluster->placement_strategy);
 427 
 428     // "stop rsc then load_stopped" constraints for current nodes
 429     for (iter = rsc->running_on; iter != NULL; iter = iter->next) {
 430         node = (pe_node_t *) iter->data;
 431         load_stopped = new_load_stopped_op(node, rsc->cluster);
 432         pcmk__new_ordering(rsc, stop_key(rsc), NULL, NULL, NULL, load_stopped,
 433                            pe_order_load, rsc->cluster);
 434     }
 435 
 436     // "load_stopped then start/migrate_to rsc" constraints for allowed nodes
 437     for (GList *iter = allowed_nodes; iter; iter = iter->next) {
 438         node = (pe_node_t *) iter->data;
 439         load_stopped = new_load_stopped_op(node, rsc->cluster);
 440         pcmk__new_ordering(NULL, NULL, load_stopped, rsc, start_key(rsc), NULL,
 441                            pe_order_load, rsc->cluster);
 442         pcmk__new_ordering(NULL, NULL, load_stopped,
 443                            rsc, pcmk__op_key(rsc->id, RSC_MIGRATE, 0), NULL,
 444                            pe_order_load, rsc->cluster);
 445     }
 446 }
 447 
 448 /*!
 449  * \internal
 450  * \brief Output node capacities if enabled
 451  *
 452  * \param[in] desc      Prefix for output
 453  * \param[in] data_set  Cluster working set
 454  */
 455 void
 456 pcmk__show_node_capacities(const char *desc, pe_working_set_t *data_set)
     /*  */
 457 {
 458     if (!pcmk_is_set(data_set->flags, pe_flag_show_utilization)) {
 459         return;
 460     }
 461     for (GList *iter = data_set->nodes; iter != NULL; iter = iter->next) {
 462         pe_node_t *node = (pe_node_t *) iter->data;
 463         pcmk__output_t *out = data_set->priv;
 464 
 465         out->message(out, "node-capacity", node, desc);
 466     }
 467 }

/* */