lustre/obdclass/jobid.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2014, Intel Corporation.
  27  *
  28  * Copyright 2017 Cray Inc, all rights reserved.
  29  * Author: Ben Evans.
  30  *
  31  * Store PID->JobID mappings
  32  */
  33
  34 #define DEBUG_SUBSYSTEM S_RPC
  35 #include <linux/user_namespace.h>
  36 #ifdef HAVE_UIDGID_HEADER
  37 #include <linux/uidgid.h>
  38 #endif
  39 #include <linux/utsname.h>
  40
  41 #include <libcfs/libcfs.h>
  42 #include <obd_support.h>
  43 #include <obd_class.h>
  44 #include <lustre_net.h>
  45
  46 static struct cfs_hash *jobid_hash;
  47 static struct cfs_hash_ops jobid_hash_ops;
  48 spinlock_t jobid_hash_lock;
  49
  50 #define RESCAN_INTERVAL 30
  51 #define DELETE_INTERVAL 300
  52
  53 char obd_jobid_var[JOBSTATS_JOBID_VAR_MAX_LEN + 1] = JOBSTATS_DISABLE;
  54 char obd_jobid_name[LUSTRE_JOBID_SIZE] = "%e.%u";
  55
  56 /**
  57  * Structure to store a single PID->JobID mapping
  58  */
  59 struct jobid_pid_map {
  60         struct hlist_node       jp_hash;
  61         time64_t                jp_time;
  62         spinlock_t              jp_lock; /* protects jp_jobid */
  63         char                    jp_jobid[LUSTRE_JOBID_SIZE];
  64         unsigned int            jp_joblen;
  65         atomic_t                jp_refcount;
  66         pid_t                   jp_pid;
  67 };
  68
  69 /*
  70  * Get jobid of current process by reading the environment variable
  71  * stored in between the "env_start" & "env_end" of task struct.
  72  *
  73  * If some job scheduler doesn't store jobid in the "env_start/end",
  74  * then an upcall could be issued here to get the jobid by utilizing
  75  * the userspace tools/API. Then, the jobid must be cached.
  76  */
  77 int jobid_get_from_environ(char *jobid_var, char *jobid, int *jobid_len)
  78 {
  79         static bool printed;
  80         int rc;
  81
  82         rc = cfs_get_environ(jobid_var, jobid, jobid_len);
  83         if (!rc)
  84                 goto out;
  85
  86         if (unlikely(rc == -EOVERFLOW && !printed)) {
  87                 /* For the PBS_JOBID and LOADL_STEP_ID keys (which are
  88                  * variable length strings instead of just numbers), it
  89                  * might make sense to keep the unique parts for JobID,
  90                  * instead of just returning an error.  That means a
  91                  * larger temp buffer for cfs_get_environ(), then
  92                  * truncating the string at some separator to fit into
  93                  * the specified jobid_len.  Fix later if needed. */
  94                 LCONSOLE_WARN("jobid: '%s' value too large (%d)\n",
  95                               obd_jobid_var, *jobid_len);
  96                 printed = true;
  97                 rc = 0;
  98         }
  99         if (rc) {
 100                 CDEBUG((rc == -ENOENT || rc == -EINVAL ||
 101                         rc == -EDEADLK) ? D_INFO : D_ERROR,
 102                        "jobid: get '%s' failed: rc = %d\n",
 103                        obd_jobid_var, rc);
 104         }
 105
 106 out:
 107         return rc;
 108 }
 109
 110 /*
 111  * jobid_should_free_item
 112  *
 113  * Each item is checked to see if it should be released
 114  * Removed from hash table by caller
 115  * Actually freed in jobid_put_locked
 116  *
 117  * Returns 1 if item is to be freed, 0 if it is to be kept
 118  */
 119
 120 static int jobid_should_free_item(void *obj, void *data)
 121 {
 122         char *jobid = data;
 123         struct jobid_pid_map *pidmap = obj;
 124         int rc = 0;
 125
 126         if (obj == NULL)
 127                 return 0;
 128
 129         if (jobid == NULL) {
 130                 WARN_ON_ONCE(atomic_read(&pidmap->jp_refcount) != 1);
 131                 return 1;
 132         }
 133
 134         spin_lock(&pidmap->jp_lock);
 135         /* prevent newly inserted items from deleting */
 136         if (jobid[0] == '\0' && atomic_read(&pidmap->jp_refcount) == 1)
 137                 rc = 1;
 138         else if (ktime_get_real_seconds() - pidmap->jp_time > DELETE_INTERVAL)
 139                 rc = 1;
 140         else if (strcmp(pidmap->jp_jobid, jobid) == 0)
 141                 rc = 1;
 142         spin_unlock(&pidmap->jp_lock);
 143
 144         return rc;
 145 }
 146
 147 /*
 148  * jobid_name_is_valid
 149  *
 150  * Checks if the jobid is a Lustre process
 151  *
 152  * Returns true if jobid is valid
 153  * Returns false if jobid looks like it's a Lustre process
 154  */
 155 static bool jobid_name_is_valid(char *jobid)
 156 {
 157         const char *const lustre_reserved[] = { "ll_ping", "ptlrpc",
 158                                                 "ldlm", "ll_sa", NULL };
 159         int i;
 160
 161         if (jobid[0] == '\0')
 162                 return false;
 163
 164         for (i = 0; lustre_reserved[i] != NULL; i++) {
 165                 if (strncmp(jobid, lustre_reserved[i],
 166                             strlen(lustre_reserved[i])) == 0)
 167                         return false;
 168         }
 169         return true;
 170 }
 171
 172 /*
 173  * jobid_get_from_cache()
 174  *
 175  * Returns contents of jobid_var from process environment for current PID.
 176  * This will be cached for some time to avoid overhead scanning environment.
 177  *
 178  * Return: -ENOMEM if allocating a new pidmap fails
 179  *         -ENOENT if no entry could be found
 180  *         +ve string length for success (something was returned in jobid)
 181  */
 182 static int jobid_get_from_cache(char *jobid, size_t joblen)
 183 {
 184         static time64_t last_expire;
 185         bool expire_cache = false;
 186         pid_t pid = current_pid();
 187         struct jobid_pid_map *pidmap = NULL;
 188         time64_t now = ktime_get_real_seconds();
 189         int rc = 0;
 190         ENTRY;
 191
 192         LASSERT(jobid_hash != NULL);
 193
 194         /* scan hash periodically to remove old PID entries from cache */
 195         spin_lock(&jobid_hash_lock);
 196         if (unlikely(last_expire + DELETE_INTERVAL <= now)) {
 197                 expire_cache = true;
 198                 last_expire = now;
 199         }
 200         spin_unlock(&jobid_hash_lock);
 201
 202         if (expire_cache)
 203                 cfs_hash_cond_del(jobid_hash, jobid_should_free_item,
 204                                   "intentionally_bad_jobid");
 205
 206         /* first try to find PID in the hash and use that value */
 207         pidmap = cfs_hash_lookup(jobid_hash, &pid);
 208         if (pidmap == NULL) {
 209                 struct jobid_pid_map *pidmap2;
 210
 211                 OBD_ALLOC_PTR(pidmap);
 212                 if (pidmap == NULL)
 213                         GOTO(out, rc = -ENOMEM);
 214
 215                 pidmap->jp_pid = pid;
 216                 pidmap->jp_time = 0;
 217                 pidmap->jp_jobid[0] = '\0';
 218                 spin_lock_init(&pidmap->jp_lock);
 219                 INIT_HLIST_NODE(&pidmap->jp_hash);
 220                 /*
 221                  * @pidmap might be reclaimed just after it is added into
 222                  * hash list, init @jp_refcount as 1 to make sure memory
 223                  * could be not freed during access.
 224                  */
 225                 atomic_set(&pidmap->jp_refcount, 1);
 226
 227                 /*
 228                  * Add the newly created map to the hash, on key collision we
 229                  * lost a racing addition and must destroy our newly allocated
 230                  * map.  The object which exists in the hash will be returned.
 231                  */
 232                 pidmap2 = cfs_hash_findadd_unique(jobid_hash, &pid,
 233                                                   &pidmap->jp_hash);
 234                 if (unlikely(pidmap != pidmap2)) {
 235                         CDEBUG(D_INFO, "jobid: duplicate found for PID=%u\n",
 236                                pid);
 237                         OBD_FREE_PTR(pidmap);
 238                         pidmap = pidmap2;
 239                 }
 240         }
 241
 242         /*
 243          * If pidmap is old (this is always true for new entries) refresh it.
 244          * If obd_jobid_var is not found, cache empty entry and try again
 245          * later, to avoid repeat lookups for PID if obd_jobid_var missing.
 246          */
 247         spin_lock(&pidmap->jp_lock);
 248         if (pidmap->jp_time + RESCAN_INTERVAL <= now) {
 249                 char env_jobid[LUSTRE_JOBID_SIZE] = "";
 250                 int env_len = sizeof(env_jobid);
 251
 252                 pidmap->jp_time = now;
 253
 254                 spin_unlock(&pidmap->jp_lock);
 255                 rc = jobid_get_from_environ(obd_jobid_var, env_jobid, &env_len);
 256
 257                 CDEBUG(D_INFO, "jobid: PID mapping established: %d->%s\n",
 258                        pidmap->jp_pid, env_jobid);
 259                 spin_lock(&pidmap->jp_lock);
 260                 if (!rc) {
 261                         pidmap->jp_joblen = env_len;
 262                         strlcpy(pidmap->jp_jobid, env_jobid,
 263                                 sizeof(pidmap->jp_jobid));
 264                         rc = 0;
 265                 } else if (rc == -ENOENT) {
 266                         /* It might have been deleted, clear out old entry */
 267                         pidmap->jp_joblen = 0;
 268                         pidmap->jp_jobid[0] = '\0';
 269                 }
 270         }
 271
 272         /*
 273          * Regardless of how pidmap was found, if it contains a valid entry
 274          * use that for now.  If there was a technical error (e.g. -ENOMEM)
 275          * use the old cached value until it can be looked up again properly.
 276          * If a cached missing entry was found, return -ENOENT.
 277          */
 278         if (pidmap->jp_joblen) {
 279                 strlcpy(jobid, pidmap->jp_jobid, joblen);
 280                 joblen = pidmap->jp_joblen;
 281                 rc = 0;
 282         } else if (!rc) {
 283                 rc = -ENOENT;
 284         }
 285         spin_unlock(&pidmap->jp_lock);
 286
 287         cfs_hash_put(jobid_hash, &pidmap->jp_hash);
 288
 289         EXIT;
 290 out:
 291         return rc < 0 ? rc : joblen;
 292 }
 293
 294 /*
 295  * jobid_interpret_string()
 296  *
 297  * Interpret the jobfmt string to expand specified fields, like coredumps do:
 298  *   %e = executable
 299  *   %g = gid
 300  *   %h = hostname
 301  *   %j = jobid from environment
 302  *   %p = pid
 303  *   %u = uid
 304  *
 305  * Unknown escape strings are dropped.  Other characters are copied through,
 306  * excluding whitespace (to avoid making jobid parsing difficult).
 307  *
 308  * Return: -EOVERFLOW if the expanded string does not fit within @joblen
 309  *         0 for success
 310  */
 311 static int jobid_interpret_string(const char *jobfmt, char *jobid,
 312                                   ssize_t joblen)
 313 {
 314         char c;
 315
 316         while ((c = *jobfmt++) && joblen > 1) {
 317                 char f;
 318                 int l;
 319
 320                 if (isspace(c)) /* Don't allow embedded spaces */
 321                         continue;
 322
 323                 if (c != '%') {
 324                         *jobid = c;
 325                         joblen--;
 326                         jobid++;
 327                         continue;
 328                 }
 329
 330                 switch ((f = *jobfmt++)) {
 331                 case 'e': /* executable name */
 332                         l = snprintf(jobid, joblen, "%s", current_comm());
 333                         break;
 334                 case 'g': /* group ID */
 335                         l = snprintf(jobid, joblen, "%u",
 336                                      from_kgid(&init_user_ns, current_fsgid()));
 337                         break;
 338                 case 'h': /* hostname */
 339                         l = snprintf(jobid, joblen, "%s",
 340                                      init_utsname()->nodename);
 341                         break;
 342                 case 'j': /* jobid stored in process environment */
 343                         l = jobid_get_from_cache(jobid, joblen);
 344                         if (l < 0)
 345                                 l = 0;
 346                         break;
 347                 case 'p': /* process ID */
 348                         l = snprintf(jobid, joblen, "%u", current_pid());
 349                         break;
 350                 case 'u': /* user ID */
 351                         l = snprintf(jobid, joblen, "%u",
 352                                      from_kuid(&init_user_ns, current_fsuid()));
 353                         break;
 354                 case '\0': /* '%' at end of format string */
 355                         l = 0;
 356                         goto out;
 357                 default: /* drop unknown %x format strings */
 358                         l = 0;
 359                         break;
 360                 }
 361                 jobid += l;
 362                 joblen -= l;
 363         }
 364         /*
 365          * This points at the end of the buffer, so long as jobid is always
 366          * incremented the same amount as joblen is decremented.
 367          */
 368 out:
 369         jobid[joblen - 1] = '\0';
 370
 371         return joblen < 0 ? -EOVERFLOW : 0;
 372 }
 373
 374 /*
 375  * Hash initialization, copied from server-side job stats bucket sizes
 376  */
 377 #define HASH_JOBID_BKT_BITS 5
 378 #define HASH_JOBID_CUR_BITS 7
 379 #define HASH_JOBID_MAX_BITS 12
 380
 381 int jobid_cache_init(void)
 382 {
 383         int rc = 0;
 384         ENTRY;
 385
 386         if (jobid_hash)
 387                 return 0;
 388
 389         spin_lock_init(&jobid_hash_lock);
 390         jobid_hash = cfs_hash_create("JOBID_HASH", HASH_JOBID_CUR_BITS,
 391                                      HASH_JOBID_MAX_BITS, HASH_JOBID_BKT_BITS,
 392                                      0, CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
 393                                      &jobid_hash_ops, CFS_HASH_DEFAULT);
 394         if (!jobid_hash)
 395                 rc = -ENOMEM;
 396
 397         RETURN(rc);
 398 }
 399 EXPORT_SYMBOL(jobid_cache_init);
 400
 401 void jobid_cache_fini(void)
 402 {
 403         struct cfs_hash *tmp_hash;
 404         ENTRY;
 405
 406         spin_lock(&jobid_hash_lock);
 407         tmp_hash = jobid_hash;
 408         jobid_hash = NULL;
 409         spin_unlock(&jobid_hash_lock);
 410
 411         if (tmp_hash != NULL) {
 412                 cfs_hash_cond_del(tmp_hash, jobid_should_free_item, NULL);
 413                 cfs_hash_putref(tmp_hash);
 414         }
 415
 416         EXIT;
 417 }
 418 EXPORT_SYMBOL(jobid_cache_fini);
 419
 420 /*
 421  * Hash operations for pid<->jobid
 422  */
 423 static unsigned jobid_hashfn(struct cfs_hash *hs, const void *key,
 424                              unsigned mask)
 425 {
 426         return cfs_hash_djb2_hash(key, sizeof(pid_t), mask);
 427 }
 428
 429 static void *jobid_key(struct hlist_node *hnode)
 430 {
 431         struct jobid_pid_map *pidmap;
 432
 433         pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 434         return &pidmap->jp_pid;
 435 }
 436
 437 static int jobid_keycmp(const void *key, struct hlist_node *hnode)
 438 {
 439         const pid_t *pid_key1;
 440         const pid_t *pid_key2;
 441
 442         LASSERT(key != NULL);
 443         pid_key1 = (pid_t *)key;
 444         pid_key2 = (pid_t *)jobid_key(hnode);
 445
 446         return *pid_key1 == *pid_key2;
 447 }
 448
 449 static void *jobid_object(struct hlist_node *hnode)
 450 {
 451         return hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 452 }
 453
 454 static void jobid_get(struct cfs_hash *hs, struct hlist_node *hnode)
 455 {
 456         struct jobid_pid_map *pidmap;
 457
 458         pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 459
 460         atomic_inc(&pidmap->jp_refcount);
 461 }
 462
 463 static void jobid_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 464 {
 465         struct jobid_pid_map *pidmap;
 466
 467         if (hnode == NULL)
 468                 return;
 469
 470         pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 471         LASSERT(atomic_read(&pidmap->jp_refcount) > 0);
 472         if (atomic_dec_and_test(&pidmap->jp_refcount)) {
 473                 CDEBUG(D_INFO, "Freeing: %d->%s\n",
 474                        pidmap->jp_pid, pidmap->jp_jobid);
 475
 476                 OBD_FREE_PTR(pidmap);
 477         }
 478 }
 479
 480 static struct cfs_hash_ops jobid_hash_ops = {
 481         .hs_hash        = jobid_hashfn,
 482         .hs_keycmp      = jobid_keycmp,
 483         .hs_key         = jobid_key,
 484         .hs_object      = jobid_object,
 485         .hs_get         = jobid_get,
 486         .hs_put         = jobid_put_locked,
 487         .hs_put_locked  = jobid_put_locked,
 488 };
 489
 490 /**
 491  * Generate the job identifier string for this process for tracking purposes.
 492  *
 493  * Fill in @jobid string based on the value of obd_jobid_var:
 494  * JOBSTATS_DISABLE:      none
 495  * JOBSTATS_NODELOCAL:    content of obd_jobid_node (jobid_interpret_string())
 496  * JOBSTATS_PROCNAME_UID: process name/UID
 497  * anything else:         look up obd_jobid_var in the processes environment
 498  *
 499  * Return -ve error number, 0 on success.
 500  */
 501 int lustre_get_jobid(char *jobid, size_t joblen)
 502 {
 503         int rc = 0;
 504         ENTRY;
 505
 506         if (unlikely(joblen < 2)) {
 507                 if (joblen == 1)
 508                         jobid[0] = '\0';
 509                 RETURN(-EINVAL);
 510         }
 511
 512         if (strcmp(obd_jobid_var, JOBSTATS_DISABLE) == 0) {
 513                 /* Jobstats isn't enabled */
 514                 memset(jobid, 0, joblen);
 515         } else if (strcmp(obd_jobid_var, JOBSTATS_NODELOCAL) == 0) {
 516                 /* Whole node dedicated to single job */
 517                 rc = jobid_interpret_string(obd_jobid_name, jobid, joblen);
 518         } else if (strcmp(obd_jobid_var, JOBSTATS_PROCNAME_UID) == 0) {
 519                 rc = jobid_interpret_string("%e.%u", jobid, joblen);
 520         } else if (jobid_name_is_valid(current_comm())) {
 521                 /*
 522                  * obd_jobid_var holds the jobid environment variable name.
 523                  * Skip initial check if obd_jobid_name already uses "%j",
 524                  * otherwise try just "%j" first, then fall back to whatever
 525                  * is in obd_jobid_name if obd_jobid_var is not found.
 526                  */
 527                 rc = -EAGAIN;
 528                 if (!strnstr(obd_jobid_name, "%j", joblen))
 529                         rc = jobid_get_from_cache(jobid, joblen);
 530
 531                 /* fall back to jobid_node if jobid_var not in environment */
 532                 if (rc < 0) {
 533                         int rc2 = jobid_interpret_string(obd_jobid_name,
 534                                                          jobid, joblen);
 535                         if (!rc2)
 536                                 rc = 0;
 537                 }
 538         }
 539
 540         RETURN(rc);
 541 }
 542 EXPORT_SYMBOL(lustre_get_jobid);
 543
 544 /*
 545  * lustre_jobid_clear
 546  *
 547  * Search cache for JobID given by @find_jobid.
 548  * If any entries in the hash table match the value, they are removed
 549  */
 550 void lustre_jobid_clear(const char *find_jobid)
 551 {
 552         char jobid[LUSTRE_JOBID_SIZE];
 553         char *end;
 554
 555         if (jobid_hash == NULL)
 556                 return;
 557
 558         strlcpy(jobid, find_jobid, sizeof(jobid));
 559         /* trim \n off the end of the incoming jobid */
 560         end = strchr(jobid, '\n');
 561         if (end && *end == '\n')
 562                 *end = '\0';
 563
 564         CDEBUG(D_INFO, "Clearing Jobid: %s\n", jobid);
 565         cfs_hash_cond_del(jobid_hash, jobid_should_free_item, jobid);
 566
 567         CDEBUG(D_INFO, "%d items remain in jobID table\n",
 568                atomic_read(&jobid_hash->hs_count));
 569 }