lib/pengine/failcounts.c

/* */
This source file includes following definitions.
is_matched_failure
block_failure
rsc_fail_name
generate_fail_regex
generate_fail_regexes
update_failcount_for_attr
update_launched_failcount
pe_get_failcount
pe__clear_failcount
   1 /*
   2  * Copyright 2008-2024 the Pacemaker project contributors
   3  *
   4  * This source code is licensed under the GNU Lesser General Public License
   5  * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY.
   6  */
   7 
   8 #include <crm_internal.h>
   9 
  10 #include <sys/types.h>
  11 #include <regex.h>
  12 #include <glib.h>
  13 
  14 #include <crm/crm.h>
  15 #include <crm/common/xml.h>
  16 #include <crm/common/util.h>
  17 #include <crm/pengine/internal.h>
  18 
  19 static gboolean
  20 is_matched_failure(const char *rsc_id, const xmlNode *conf_op_xml,
     /*  */
  21                    const xmlNode *lrm_op_xml)
  22 {
  23     gboolean matched = FALSE;
  24     const char *conf_op_name = NULL;
  25     const char *lrm_op_task = NULL;
  26     const char *conf_op_interval_spec = NULL;
  27     guint conf_op_interval_ms = 0;
  28     guint lrm_op_interval_ms = 0;
  29     const char *lrm_op_id = NULL;
  30     char *last_failure_key = NULL;
  31 
  32     if (rsc_id == NULL || conf_op_xml == NULL || lrm_op_xml == NULL) {
  33         return FALSE;
  34     }
  35 
  36     // Get name and interval from configured op
  37     conf_op_name = crm_element_value(conf_op_xml, PCMK_XA_NAME);
  38     conf_op_interval_spec = crm_element_value(conf_op_xml, PCMK_META_INTERVAL);
  39     pcmk_parse_interval_spec(conf_op_interval_spec, &conf_op_interval_ms);
  40 
  41     // Get name and interval from op history entry
  42     lrm_op_task = crm_element_value(lrm_op_xml, PCMK_XA_OPERATION);
  43     crm_element_value_ms(lrm_op_xml, PCMK_META_INTERVAL, &lrm_op_interval_ms);
  44 
  45     if ((conf_op_interval_ms != lrm_op_interval_ms)
  46         || !pcmk__str_eq(conf_op_name, lrm_op_task, pcmk__str_casei)) {
  47         return FALSE;
  48     }
  49 
  50     lrm_op_id = pcmk__xe_id(lrm_op_xml);
  51     last_failure_key = pcmk__op_key(rsc_id, "last_failure", 0);
  52 
  53     if (pcmk__str_eq(last_failure_key, lrm_op_id, pcmk__str_casei)) {
  54         matched = TRUE;
  55 
  56     } else {
  57         char *expected_op_key = pcmk__op_key(rsc_id, conf_op_name,
  58                                                 conf_op_interval_ms);
  59 
  60         if (pcmk__str_eq(expected_op_key, lrm_op_id, pcmk__str_casei)) {
  61             int rc = 0;
  62             int target_rc = pe__target_rc_from_xml(lrm_op_xml);
  63 
  64             crm_element_value_int(lrm_op_xml, PCMK__XA_RC_CODE, &rc);
  65             if (rc != target_rc) {
  66                 matched = TRUE;
  67             }
  68         }
  69         free(expected_op_key);
  70     }
  71 
  72     free(last_failure_key);
  73     return matched;
  74 }
  75 
  76 static gboolean
  77 block_failure(const pcmk_node_t *node, pcmk_resource_t *rsc,
     /*  */
  78               const xmlNode *xml_op)
  79 {
  80     char *xml_name = clone_strip(rsc->id);
  81 
  82     /* @TODO This xpath search occurs after template expansion, but it is unable
  83      * to properly detect on-fail in id-ref, operation meta-attributes, or
  84      * op_defaults, or evaluate rules.
  85      *
  86      * Also, PCMK_META_ON_FAIL defaults to PCMK_VALUE_BLOCK (in
  87      * unpack_operation()) for stop actions when stonith is disabled.
  88      *
  89      * Ideally, we'd unpack the operation before this point, and pass in a
  90      * meta-attributes table that takes all that into consideration.
  91      */
  92     char *xpath = crm_strdup_printf("//" PCMK_XE_PRIMITIVE
  93                                     "[@" PCMK_XA_ID "='%s']"
  94                                     "//" PCMK_XE_OP
  95                                     "[@" PCMK_META_ON_FAIL
  96                                         "='" PCMK_VALUE_BLOCK "']",
  97                                     xml_name);
  98 
  99     xmlXPathObject *xpathObj = xpath_search(rsc->priv->xml, xpath);
 100     gboolean should_block = FALSE;
 101 
 102     free(xpath);
 103 
 104     if (xpathObj) {
 105         int max = numXpathResults(xpathObj);
 106         int lpc = 0;
 107 
 108         for (lpc = 0; lpc < max; lpc++) {
 109             xmlNode *pref = getXpathResult(xpathObj, lpc);
 110 
 111             if (xml_op) {
 112                 should_block = is_matched_failure(xml_name, pref, xml_op);
 113                 if (should_block) {
 114                     break;
 115                 }
 116 
 117             } else {
 118                 const char *conf_op_name = NULL;
 119                 const char *conf_op_interval_spec = NULL;
 120                 guint conf_op_interval_ms = 0;
 121                 char *lrm_op_xpath = NULL;
 122                 xmlXPathObject *lrm_op_xpathObj = NULL;
 123 
 124                 // Get name and interval from configured op
 125                 conf_op_name = crm_element_value(pref, PCMK_XA_NAME);
 126                 conf_op_interval_spec = crm_element_value(pref,
 127                                                           PCMK_META_INTERVAL);
 128                 pcmk_parse_interval_spec(conf_op_interval_spec,
 129                                          &conf_op_interval_ms);
 130 
 131 #define XPATH_FMT "//" PCMK__XE_NODE_STATE "[@" PCMK_XA_UNAME "='%s']"      \
 132                   "//" PCMK__XE_LRM_RESOURCE "[@" PCMK_XA_ID "='%s']"       \
 133                   "/" PCMK__XE_LRM_RSC_OP "[@" PCMK_XA_OPERATION "='%s']"   \
 134                   "[@" PCMK_META_INTERVAL "='%u']"
 135 
 136                 lrm_op_xpath = crm_strdup_printf(XPATH_FMT,
 137                                                  node->priv->name, xml_name,
 138                                                  conf_op_name,
 139                                                  conf_op_interval_ms);
 140                 lrm_op_xpathObj = xpath_search(rsc->priv->scheduler->input,
 141                                                lrm_op_xpath);
 142 
 143                 free(lrm_op_xpath);
 144 
 145                 if (lrm_op_xpathObj) {
 146                     int max2 = numXpathResults(lrm_op_xpathObj);
 147                     int lpc2 = 0;
 148 
 149                     for (lpc2 = 0; lpc2 < max2; lpc2++) {
 150                         xmlNode *lrm_op_xml = getXpathResult(lrm_op_xpathObj,
 151                                                              lpc2);
 152 
 153                         should_block = is_matched_failure(xml_name, pref,
 154                                                           lrm_op_xml);
 155                         if (should_block) {
 156                             break;
 157                         }
 158                     }
 159                 }
 160                 freeXpathObject(lrm_op_xpathObj);
 161 
 162                 if (should_block) {
 163                     break;
 164                 }
 165             }
 166         }
 167     }
 168 
 169     free(xml_name);
 170     freeXpathObject(xpathObj);
 171 
 172     return should_block;
 173 }
 174 
 175 /*!
 176  * \internal
 177  * \brief Get resource name as used in failure-related node attributes
 178  *
 179  * \param[in] rsc  Resource to check
 180  *
 181  * \return Newly allocated string containing resource's fail name
 182  * \note The caller is responsible for freeing the result.
 183  */
 184 static inline char *
 185 rsc_fail_name(const pcmk_resource_t *rsc)
     /*  */
 186 {
 187     const char *name = pcmk__s(rsc->priv->history_id, rsc->id);
 188 
 189     return pcmk_is_set(rsc->flags, pcmk__rsc_unique)? strdup(name) : clone_strip(name);
 190 }
 191 
 192 /*!
 193  * \internal
 194  * \brief Compile regular expression to match a failure-related node attribute
 195  *
 196  * \param[in]  prefix    Attribute prefix to match
 197  * \param[in]  rsc_name  Resource name to match as used in failure attributes
 198  * \param[in]  is_unique Whether the resource is a globally unique clone
 199  * \param[out] re        Where to store resulting regular expression
 200  *
 201  * \return Standard Pacemaker return code
 202  * \note Fail attributes are named like PREFIX-RESOURCE#OP_INTERVAL.
 203  *       The caller is responsible for freeing re with regfree().
 204  */
 205 static int
 206 generate_fail_regex(const char *prefix, const char *rsc_name, bool is_unique,
     /*  */
 207                     regex_t *re)
 208 {
 209     char *pattern = NULL;
 210     const char *op_pattern = "#.+_[0-9]+";
 211 
 212     /* Ignore instance numbers for anything other than globally unique clones.
 213      * Anonymous clone fail counts could contain an instance number if the
 214      * clone was initially unique, failed, then was converted to anonymous.
 215      */
 216     const char *instance_pattern = (is_unique? "" : "(:[0-9]+)?");
 217 
 218     pattern = crm_strdup_printf("^%s-%s%s%s$", prefix, rsc_name,
 219                                 instance_pattern, op_pattern);
 220     if (regcomp(re, pattern, REG_EXTENDED|REG_NOSUB) != 0) {
 221         free(pattern);
 222         return EINVAL;
 223     }
 224 
 225     free(pattern);
 226     return pcmk_rc_ok;
 227 }
 228 
 229 /*!
 230  * \internal
 231  * \brief Compile regular expressions to match failure-related node attributes
 232  *
 233  * \param[in]  rsc             Resource being checked for failures
 234  * \param[out] failcount_re    Storage for regular expression for fail count
 235  * \param[out] lastfailure_re  Storage for regular expression for last failure
 236  *
 237  * \return Standard Pacemaker return code
 238  * \note On success, the caller is responsible for freeing the expressions with
 239  *       regfree().
 240  */
 241 static int
 242 generate_fail_regexes(const pcmk_resource_t *rsc, regex_t *failcount_re,
     /*  */
 243                       regex_t *lastfailure_re)
 244 {
 245     int rc = pcmk_rc_ok;
 246     char *rsc_name = rsc_fail_name(rsc);
 247 
 248     if (generate_fail_regex(PCMK__FAIL_COUNT_PREFIX, rsc_name,
 249                             pcmk_is_set(rsc->flags, pcmk__rsc_unique),
 250                             failcount_re) != pcmk_rc_ok) {
 251         rc = EINVAL;
 252 
 253     } else if (generate_fail_regex(PCMK__LAST_FAILURE_PREFIX, rsc_name,
 254                                    pcmk_is_set(rsc->flags, pcmk__rsc_unique),
 255                                    lastfailure_re) != pcmk_rc_ok) {
 256         rc = EINVAL;
 257         regfree(failcount_re);
 258     }
 259 
 260     free(rsc_name);
 261     return rc;
 262 }
 263 
 264 // Data for fail-count-related iterators
 265 struct failcount_data {
 266     const pcmk_node_t *node;// Node to check for fail count
 267     pcmk_resource_t *rsc;     // Resource to check for fail count
 268     uint32_t flags;         // Fail count flags
 269     const xmlNode *xml_op;  // History entry for expiration purposes (or NULL)
 270     regex_t failcount_re;   // Fail count regular expression to match
 271     regex_t lastfailure_re; // Last failure regular expression to match
 272     int failcount;          // Fail count so far
 273     time_t last_failure;    // Time of most recent failure so far
 274 };
 275 
 276 /*!
 277  * \internal
 278  * \brief Update fail count and last failure appropriately for a node attribute
 279  *
 280  * \param[in] key        Node attribute name
 281  * \param[in] value      Node attribute value
 282  * \param[in] user_data  Fail count data to update
 283  */
 284 static void
 285 update_failcount_for_attr(gpointer key, gpointer value, gpointer user_data)
     /*  */
 286 {
 287     struct failcount_data *fc_data = user_data;
 288 
 289     // If this is a matching fail count attribute, update fail count
 290     if (regexec(&(fc_data->failcount_re), (const char *) key, 0, NULL, 0) == 0) {
 291         int score = 0;
 292         int rc = pcmk_parse_score(value, &score, 0);
 293 
 294         if (rc != pcmk_rc_ok) {
 295             crm_warn("Ignoring %s for %s "
 296                      "because '%s' is not a valid fail count: %s",
 297                      (const char *) key, pcmk__node_name(fc_data->node),
 298                      value, pcmk_rc_str(rc));
 299             return;
 300         }
 301         fc_data->failcount = pcmk__add_scores(fc_data->failcount, score);
 302         pcmk__rsc_trace(fc_data->rsc, "Added %s (%s) to %s fail count (now %s)",
 303                         (const char *) key, (const char *) value,
 304                         fc_data->rsc->id,
 305                         pcmk_readable_score(fc_data->failcount));
 306         return;
 307     }
 308 
 309     // If this is a matching last failure attribute, update last failure
 310     if (regexec(&(fc_data->lastfailure_re), (const char *) key, 0, NULL,
 311                 0) == 0) {
 312         long long last_ll;
 313         int rc = pcmk__scan_ll(value, &last_ll, 0LL);
 314 
 315         if (rc != pcmk_rc_ok) {
 316             crm_info("Ignoring invalid value '%s' for %s: %s",
 317                      (const char *) value, (const char *) key, pcmk_rc_str(rc));
 318             return;
 319         }
 320         fc_data->last_failure = (time_t) QB_MAX(fc_data->last_failure, last_ll);
 321     }
 322 }
 323 
 324 /*!
 325  * \internal
 326  * \brief Update fail count and last failure appropriately for launched resource
 327  *
 328  * \param[in] data       Launched resource
 329  * \param[in] user_data  Fail count data to update
 330  */
 331 static void
 332 update_launched_failcount(gpointer data, gpointer user_data)
     /*  */
 333 {
 334     pcmk_resource_t *launched = data;
 335     struct failcount_data *fc_data = user_data;
 336     time_t launched_last_failure = 0;
 337 
 338     fc_data->failcount += pe_get_failcount(fc_data->node, launched,
 339                                            &launched_last_failure,
 340                                            fc_data->flags, fc_data->xml_op);
 341     fc_data->last_failure = QB_MAX(fc_data->last_failure, launched_last_failure);
 342 }
 343 
 344 #define readable_expiration(rsc)    \
 345     pcmk__readable_interval((rsc)->priv->failure_expiration_ms)
 346 
 347 /*!
 348  * \internal
 349  * \brief Get a resource's fail count on a node
 350  *
 351  * \param[in]     node          Node to check
 352  * \param[in,out] rsc           Resource to check
 353  * \param[out]    last_failure  If not NULL, where to set time of most recent
 354  *                              failure of \p rsc on \p node
 355  * \param[in]     flags         Group of enum pcmk__fc_flags
 356  * \param[in]     xml_op        If not NULL, consider only the action in this
 357  *                              history entry when determining whether on-fail
 358  *                              is configured as "blocked", otherwise consider
 359  *                              all actions configured for \p rsc
 360  *
 361  * \return Fail count for \p rsc on \p node according to \p flags
 362  */
 363 int
 364 pe_get_failcount(const pcmk_node_t *node, pcmk_resource_t *rsc,
     /*  */
 365                  time_t *last_failure, uint32_t flags, const xmlNode *xml_op)
 366 {
 367     struct failcount_data fc_data = {
 368         .node = node,
 369         .rsc = rsc,
 370         .flags = flags,
 371         .xml_op = xml_op,
 372         .failcount = 0,
 373         .last_failure = (time_t) 0,
 374     };
 375 
 376     // Calculate resource failcount as sum of all matching operation failcounts
 377     CRM_CHECK(generate_fail_regexes(rsc, &fc_data.failcount_re,
 378                                     &fc_data.lastfailure_re) == pcmk_rc_ok,
 379               return 0);
 380     g_hash_table_foreach(node->priv->attrs, update_failcount_for_attr,
 381                          &fc_data);
 382     regfree(&(fc_data.failcount_re));
 383     regfree(&(fc_data.lastfailure_re));
 384 
 385     // If failure blocks the resource, disregard any failure timeout
 386     if ((fc_data.failcount > 0) && (rsc->priv->failure_expiration_ms > 0)
 387         && block_failure(node, rsc, xml_op)) {
 388 
 389         pcmk__config_warn("Ignoring failure timeout (%s) for %s "
 390                           "because it conflicts with "
 391                           PCMK_META_ON_FAIL "=" PCMK_VALUE_BLOCK,
 392                           readable_expiration(rsc), rsc->id);
 393         rsc->priv->failure_expiration_ms = 0;
 394     }
 395 
 396     // If all failures have expired, ignore fail count
 397     if (pcmk_is_set(flags, pcmk__fc_effective) && (fc_data.failcount > 0)
 398         && (fc_data.last_failure > 0)
 399         && (rsc->priv->failure_expiration_ms > 0)) {
 400 
 401         time_t now = get_effective_time(rsc->priv->scheduler);
 402         const guint expiration = pcmk__timeout_ms2s(rsc->priv->failure_expiration_ms);
 403 
 404         if (now > (fc_data.last_failure + expiration)) {
 405             pcmk__rsc_debug(rsc, "Failcount for %s on %s expired after %s",
 406                             rsc->id, pcmk__node_name(node),
 407                             readable_expiration(rsc));
 408             fc_data.failcount = 0;
 409         }
 410     }
 411 
 412     /* Add the fail count of any launched resources, except that we never want
 413      * the fail counts of a bundle container's launched resources to count
 414      * towards the container's fail count.
 415      *
 416      * Most importantly, a Pacemaker Remote connection to a bundle container
 417      * is launched by the container, but can reside on a different node than the
 418      * container itself. Counting its fail count on its node towards the
 419      * container's fail count on that node could lead to attempting to stop the
 420      * container on the wrong node.
 421      */
 422     if (pcmk_is_set(flags, pcmk__fc_launched)
 423         && (rsc->priv->launched != NULL) && !pcmk__is_bundled(rsc)) {
 424 
 425         g_list_foreach(rsc->priv->launched, update_launched_failcount,
 426                        &fc_data);
 427         if (fc_data.failcount > 0) {
 428             pcmk__rsc_info(rsc,
 429                            "Container %s and the resources within it "
 430                            "have failed %s time%s on %s",
 431                            rsc->id, pcmk_readable_score(fc_data.failcount),
 432                            pcmk__plural_s(fc_data.failcount),
 433                            pcmk__node_name(node));
 434         }
 435 
 436     } else if (fc_data.failcount > 0) {
 437         pcmk__rsc_info(rsc, "%s has failed %s time%s on %s",
 438                        rsc->id, pcmk_readable_score(fc_data.failcount),
 439                        pcmk__plural_s(fc_data.failcount),
 440                        pcmk__node_name(node));
 441     }
 442 
 443     if (last_failure != NULL) {
 444         if ((fc_data.failcount > 0) && (fc_data.last_failure > 0)) {
 445             *last_failure = fc_data.last_failure;
 446         } else  {
 447             *last_failure = 0;
 448         }
 449     }
 450     return fc_data.failcount;
 451 }
 452 
 453 /*!
 454  * \brief Schedule a controller operation to clear a fail count
 455  *
 456  * \param[in,out] rsc        Resource with failure
 457  * \param[in]     node       Node failure occurred on
 458  * \param[in]     reason     Readable description why needed (for logging)
 459  * \param[in,out] scheduler  Scheduler data cluster
 460  *
 461  * \return Scheduled action
 462  */
 463 pcmk_action_t *
 464 pe__clear_failcount(pcmk_resource_t *rsc, const pcmk_node_t *node,
     /*  */
 465                     const char *reason, pcmk_scheduler_t *scheduler)
 466 {
 467     char *key = NULL;
 468     pcmk_action_t *clear = NULL;
 469 
 470     CRM_CHECK(rsc && node && reason && scheduler, return NULL);
 471 
 472     key = pcmk__op_key(rsc->id, PCMK_ACTION_CLEAR_FAILCOUNT, 0);
 473     clear = custom_action(rsc, key, PCMK_ACTION_CLEAR_FAILCOUNT, node, FALSE,
 474                           scheduler);
 475     pcmk__insert_meta(clear, PCMK__META_OP_NO_WAIT, PCMK_VALUE_TRUE);
 476     crm_notice("Clearing failure of %s on %s because %s " QB_XS " %s",
 477                rsc->id, pcmk__node_name(node), reason, clear->uuid);
 478     return clear;
 479 }
/* */
root/lib/pengine/failcounts.c

DEFINITIONS
