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         spin_lock(&pidmap->jp_lock);
 130         if (jobid == NULL)
 131                 rc = 1;
 132         else if (jobid[0] == '\0')
 133                 rc = 1;
 134         else if (ktime_get_real_seconds() - pidmap->jp_time > DELETE_INTERVAL)
 135                 rc = 1;
 136         else if (strcmp(pidmap->jp_jobid, jobid) == 0)
 137                 rc = 1;
 138         spin_unlock(&pidmap->jp_lock);
 139
 140         return rc;
 141 }
 142
 143 /*
 144  * jobid_name_is_valid
 145  *
 146  * Checks if the jobid is a Lustre process
 147  *
 148  * Returns true if jobid is valid
 149  * Returns false if jobid looks like it's a Lustre process
 150  */
 151 static bool jobid_name_is_valid(char *jobid)
 152 {
 153         const char *const lustre_reserved[] = { "ll_ping", "ptlrpc",
 154                                                 "ldlm", "ll_sa", NULL };
 155         int i;
 156
 157         if (jobid[0] == '\0')
 158                 return false;
 159
 160         for (i = 0; lustre_reserved[i] != NULL; i++) {
 161                 if (strncmp(jobid, lustre_reserved[i],
 162                             strlen(lustre_reserved[i])) == 0)
 163                         return false;
 164         }
 165         return true;
 166 }
 167
 168 /*
 169  * jobid_get_from_cache()
 170  *
 171  * Returns contents of jobid_var from process environment for current PID.
 172  * This will be cached for some time to avoid overhead scanning environment.
 173  *
 174  * Return: -ENOMEM if allocating a new pidmap fails
 175  *         -ENOENT if no entry could be found
 176  *         +ve string length for success (something was returned in jobid)
 177  */
 178 static int jobid_get_from_cache(char *jobid, size_t joblen)
 179 {
 180         static time64_t last_expire;
 181         bool expire_cache = false;
 182         pid_t pid = current_pid();
 183         struct jobid_pid_map *pidmap = NULL;
 184         time64_t now = ktime_get_real_seconds();
 185         int rc = 0;
 186         ENTRY;
 187
 188         LASSERT(jobid_hash != NULL);
 189
 190         /* scan hash periodically to remove old PID entries from cache */
 191         spin_lock(&jobid_hash_lock);
 192         if (unlikely(last_expire + DELETE_INTERVAL <= now)) {
 193                 expire_cache = true;
 194                 last_expire = now;
 195         }
 196         spin_unlock(&jobid_hash_lock);
 197
 198         if (expire_cache)
 199                 cfs_hash_cond_del(jobid_hash, jobid_should_free_item,
 200                                   "intentionally_bad_jobid");
 201
 202         /* first try to find PID in the hash and use that value */
 203         pidmap = cfs_hash_lookup(jobid_hash, &pid);
 204         if (pidmap == NULL) {
 205                 struct jobid_pid_map *pidmap2;
 206
 207                 OBD_ALLOC_PTR(pidmap);
 208                 if (pidmap == NULL)
 209                         GOTO(out, rc = -ENOMEM);
 210
 211                 pidmap->jp_pid = pid;
 212                 pidmap->jp_time = 0;
 213                 pidmap->jp_jobid[0] = '\0';
 214                 spin_lock_init(&pidmap->jp_lock);
 215                 INIT_HLIST_NODE(&pidmap->jp_hash);
 216
 217                 /*
 218                  * Add the newly created map to the hash, on key collision we
 219                  * lost a racing addition and must destroy our newly allocated
 220                  * map.  The object which exists in the hash will be returned.
 221                  */
 222                 pidmap2 = cfs_hash_findadd_unique(jobid_hash, &pid,
 223                                                   &pidmap->jp_hash);
 224                 if (unlikely(pidmap != pidmap2)) {
 225                         CDEBUG(D_INFO, "jobid: duplicate found for PID=%u\n",
 226                                pid);
 227                         OBD_FREE_PTR(pidmap);
 228                         pidmap = pidmap2;
 229                 } else {
 230                         cfs_hash_get(jobid_hash, &pidmap->jp_hash);
 231                 }
 232         }
 233
 234         /*
 235          * If pidmap is old (this is always true for new entries) refresh it.
 236          * If obd_jobid_var is not found, cache empty entry and try again
 237          * later, to avoid repeat lookups for PID if obd_jobid_var missing.
 238          */
 239         spin_lock(&pidmap->jp_lock);
 240         if (pidmap->jp_time + RESCAN_INTERVAL <= now) {
 241                 char env_jobid[LUSTRE_JOBID_SIZE] = "";
 242                 int env_len = sizeof(env_jobid);
 243
 244                 pidmap->jp_time = now;
 245
 246                 spin_unlock(&pidmap->jp_lock);
 247                 rc = jobid_get_from_environ(obd_jobid_var, env_jobid, &env_len);
 248
 249                 CDEBUG(D_INFO, "jobid: PID mapping established: %d->%s\n",
 250                        pidmap->jp_pid, env_jobid);
 251                 spin_lock(&pidmap->jp_lock);
 252                 if (!rc) {
 253                         pidmap->jp_joblen = env_len;
 254                         strlcpy(pidmap->jp_jobid, env_jobid,
 255                                 sizeof(pidmap->jp_jobid));
 256                         rc = 0;
 257                 } else if (rc == -ENOENT) {
 258                         /* It might have been deleted, clear out old entry */
 259                         pidmap->jp_joblen = 0;
 260                         pidmap->jp_jobid[0] = '\0';
 261                 }
 262         }
 263
 264         /*
 265          * Regardless of how pidmap was found, if it contains a valid entry
 266          * use that for now.  If there was a technical error (e.g. -ENOMEM)
 267          * use the old cached value until it can be looked up again properly.
 268          * If a cached missing entry was found, return -ENOENT.
 269          */
 270         if (pidmap->jp_joblen) {
 271                 strlcpy(jobid, pidmap->jp_jobid, joblen);
 272                 joblen = pidmap->jp_joblen;
 273                 rc = 0;
 274         } else if (!rc) {
 275                 rc = -ENOENT;
 276         }
 277         spin_unlock(&pidmap->jp_lock);
 278
 279         cfs_hash_put(jobid_hash, &pidmap->jp_hash);
 280
 281         EXIT;
 282 out:
 283         return rc < 0 ? rc : joblen;
 284 }
 285
 286 /*
 287  * jobid_interpret_string()
 288  *
 289  * Interpret the jobfmt string to expand specified fields, like coredumps do:
 290  *   %e = executable
 291  *   %g = gid
 292  *   %h = hostname
 293  *   %j = jobid from environment
 294  *   %p = pid
 295  *   %u = uid
 296  *
 297  * Unknown escape strings are dropped.  Other characters are copied through,
 298  * excluding whitespace (to avoid making jobid parsing difficult).
 299  *
 300  * Return: -EOVERFLOW if the expanded string does not fit within @joblen
 301  *         0 for success
 302  */
 303 static int jobid_interpret_string(const char *jobfmt, char *jobid,
 304                                   ssize_t joblen)
 305 {
 306         char c;
 307
 308         while ((c = *jobfmt++) && joblen > 1) {
 309                 char f;
 310                 int l;
 311
 312                 if (isspace(c)) /* Don't allow embedded spaces */
 313                         continue;
 314
 315                 if (c != '%') {
 316                         *jobid = c;
 317                         joblen--;
 318                         jobid++;
 319                         continue;
 320                 }
 321
 322                 switch ((f = *jobfmt++)) {
 323                 case 'e': /* executable name */
 324                         l = snprintf(jobid, joblen, "%s", current_comm());
 325                         break;
 326                 case 'g': /* group ID */
 327                         l = snprintf(jobid, joblen, "%u",
 328                                      from_kgid(&init_user_ns, current_fsgid()));
 329                         break;
 330                 case 'h': /* hostname */
 331                         l = snprintf(jobid, joblen, "%s",
 332                                      init_utsname()->nodename);
 333                         break;
 334                 case 'j': /* jobid stored in process environment */
 335                         l = jobid_get_from_cache(jobid, joblen);
 336                         if (l < 0)
 337                                 l = 0;
 338                         break;
 339                 case 'p': /* process ID */
 340                         l = snprintf(jobid, joblen, "%u", current_pid());
 341                         break;
 342                 case 'u': /* user ID */
 343                         l = snprintf(jobid, joblen, "%u",
 344                                      from_kuid(&init_user_ns, current_fsuid()));
 345                         break;
 346                 case '\0': /* '%' at end of format string */
 347                         l = 0;
 348                         goto out;
 349                 default: /* drop unknown %x format strings */
 350                         l = 0;
 351                         break;
 352                 }
 353                 jobid += l;
 354                 joblen -= l;
 355         }
 356         /*
 357          * This points at the end of the buffer, so long as jobid is always
 358          * incremented the same amount as joblen is decremented.
 359          */
 360 out:
 361         jobid[joblen - 1] = '\0';
 362
 363         return joblen < 0 ? -EOVERFLOW : 0;
 364 }
 365
 366 /*
 367  * Hash initialization, copied from server-side job stats bucket sizes
 368  */
 369 #define HASH_JOBID_BKT_BITS 5
 370 #define HASH_JOBID_CUR_BITS 7
 371 #define HASH_JOBID_MAX_BITS 12
 372
 373 int jobid_cache_init(void)
 374 {
 375         int rc = 0;
 376         ENTRY;
 377
 378         if (jobid_hash)
 379                 return 0;
 380
 381         spin_lock_init(&jobid_hash_lock);
 382         jobid_hash = cfs_hash_create("JOBID_HASH", HASH_JOBID_CUR_BITS,
 383                                      HASH_JOBID_MAX_BITS, HASH_JOBID_BKT_BITS,
 384                                      0, CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
 385                                      &jobid_hash_ops, CFS_HASH_DEFAULT);
 386         if (!jobid_hash)
 387                 rc = -ENOMEM;
 388
 389         RETURN(rc);
 390 }
 391 EXPORT_SYMBOL(jobid_cache_init);
 392
 393 void jobid_cache_fini(void)
 394 {
 395         struct cfs_hash *tmp_hash;
 396         ENTRY;
 397
 398         spin_lock(&jobid_hash_lock);
 399         tmp_hash = jobid_hash;
 400         jobid_hash = NULL;
 401         spin_unlock(&jobid_hash_lock);
 402
 403         if (tmp_hash != NULL) {
 404                 cfs_hash_cond_del(tmp_hash, jobid_should_free_item, NULL);
 405                 cfs_hash_putref(tmp_hash);
 406         }
 407
 408         EXIT;
 409 }
 410 EXPORT_SYMBOL(jobid_cache_fini);
 411
 412 /*
 413  * Hash operations for pid<->jobid
 414  */
 415 static unsigned jobid_hashfn(struct cfs_hash *hs, const void *key,
 416                              unsigned mask)
 417 {
 418         return cfs_hash_djb2_hash(key, sizeof(pid_t), mask);
 419 }
 420
 421 static void *jobid_key(struct hlist_node *hnode)
 422 {
 423         struct jobid_pid_map *pidmap;
 424
 425         pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 426         return &pidmap->jp_pid;
 427 }
 428
 429 static int jobid_keycmp(const void *key, struct hlist_node *hnode)
 430 {
 431         const pid_t *pid_key1;
 432         const pid_t *pid_key2;
 433
 434         LASSERT(key != NULL);
 435         pid_key1 = (pid_t *)key;
 436         pid_key2 = (pid_t *)jobid_key(hnode);
 437
 438         return *pid_key1 == *pid_key2;
 439 }
 440
 441 static void *jobid_object(struct hlist_node *hnode)
 442 {
 443         return hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 444 }
 445
 446 static void jobid_get(struct cfs_hash *hs, struct hlist_node *hnode)
 447 {
 448         struct jobid_pid_map *pidmap;
 449
 450         pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 451
 452         atomic_inc(&pidmap->jp_refcount);
 453 }
 454
 455 static void jobid_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 456 {
 457         struct jobid_pid_map *pidmap;
 458
 459         if (hnode == NULL)
 460                 return;
 461
 462         pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 463         LASSERT(atomic_read(&pidmap->jp_refcount) > 0);
 464         if (atomic_dec_and_test(&pidmap->jp_refcount)) {
 465                 CDEBUG(D_INFO, "Freeing: %d->%s\n",
 466                        pidmap->jp_pid, pidmap->jp_jobid);
 467
 468                 OBD_FREE_PTR(pidmap);
 469         }
 470 }
 471
 472 static struct cfs_hash_ops jobid_hash_ops = {
 473         .hs_hash        = jobid_hashfn,
 474         .hs_keycmp      = jobid_keycmp,
 475         .hs_key         = jobid_key,
 476         .hs_object      = jobid_object,
 477         .hs_get         = jobid_get,
 478         .hs_put         = jobid_put_locked,
 479         .hs_put_locked  = jobid_put_locked,
 480 };
 481
 482 /**
 483  * Generate the job identifier string for this process for tracking purposes.
 484  *
 485  * Fill in @jobid string based on the value of obd_jobid_var:
 486  * JOBSTATS_DISABLE:      none
 487  * JOBSTATS_NODELOCAL:    content of obd_jobid_node (jobid_interpret_string())
 488  * JOBSTATS_PROCNAME_UID: process name/UID
 489  * anything else:         look up obd_jobid_var in the processes environment
 490  *
 491  * Return -ve error number, 0 on success.
 492  */
 493 int lustre_get_jobid(char *jobid, size_t joblen)
 494 {
 495         int rc = 0;
 496         ENTRY;
 497
 498         if (unlikely(joblen < 2)) {
 499                 if (joblen == 1)
 500                         jobid[0] = '\0';
 501                 RETURN(-EINVAL);
 502         }
 503
 504         if (strcmp(obd_jobid_var, JOBSTATS_DISABLE) == 0) {
 505                 /* Jobstats isn't enabled */
 506                 memset(jobid, 0, joblen);
 507         } else if (strcmp(obd_jobid_var, JOBSTATS_NODELOCAL) == 0) {
 508                 /* Whole node dedicated to single job */
 509                 rc = jobid_interpret_string(obd_jobid_name, jobid, joblen);
 510         } else if (strcmp(obd_jobid_var, JOBSTATS_PROCNAME_UID) == 0) {
 511                 rc = jobid_interpret_string("%e.%u", jobid, joblen);
 512         } else if (jobid_name_is_valid(current_comm())) {
 513                 /*
 514                  * obd_jobid_var holds the jobid environment variable name.
 515                  * Skip initial check if obd_jobid_name already uses "%j",
 516                  * otherwise try just "%j" first, then fall back to whatever
 517                  * is in obd_jobid_name if obd_jobid_var is not found.
 518                  */
 519                 rc = -EAGAIN;
 520                 if (!strnstr(obd_jobid_name, "%j", joblen))
 521                         rc = jobid_get_from_cache(jobid, joblen);
 522
 523                 /* fall back to jobid_node if jobid_var not in environment */
 524                 if (rc < 0) {
 525                         int rc2 = jobid_interpret_string(obd_jobid_name,
 526                                                          jobid, joblen);
 527                         if (!rc2)
 528                                 rc = 0;
 529                 }
 530         }
 531
 532         RETURN(rc);
 533 }
 534 EXPORT_SYMBOL(lustre_get_jobid);
 535
 536 /*
 537  * lustre_jobid_clear
 538  *
 539  * Search cache for JobID given by @find_jobid.
 540  * If any entries in the hash table match the value, they are removed
 541  */
 542 void lustre_jobid_clear(const char *find_jobid)
 543 {
 544         char jobid[LUSTRE_JOBID_SIZE];
 545         char *end;
 546
 547         if (jobid_hash == NULL)
 548                 return;
 549
 550         strlcpy(jobid, find_jobid, sizeof(jobid));
 551         /* trim \n off the end of the incoming jobid */
 552         end = strchr(jobid, '\n');
 553         if (end && *end == '\n')
 554                 *end = '\0';
 555
 556         CDEBUG(D_INFO, "Clearing Jobid: %s\n", jobid);
 557         cfs_hash_cond_del(jobid_hash, jobid_should_free_item, jobid);
 558
 559         CDEBUG(D_INFO, "%d items remain in jobID table\n",
 560                atomic_read(&jobid_hash->hs_count));
 561 }