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/linux/linux-misc.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_ERROR_MSG(0x16b,
  95                                    "jobid: '%s' value too large (%d)\n",
  96                                    obd_jobid_var, *jobid_len);
  97                 printed = true;
  98                 rc = 0;
  99         }
 100         if (rc) {
 101                 CDEBUG((rc == -ENOENT || rc == -EINVAL ||
 102                         rc == -EDEADLK) ? D_INFO : D_ERROR,
 103                        "jobid: get '%s' failed: rc = %d\n",
 104                        obd_jobid_var, rc);
 105         }
 106
 107 out:
 108         return rc;
 109 }
 110
 111 /*
 112  * jobid_should_free_item
 113  *
 114  * Each item is checked to see if it should be released
 115  * Removed from hash table by caller
 116  * Actually freed in jobid_put_locked
 117  *
 118  * Returns 1 if item is to be freed, 0 if it is to be kept
 119  */
 120
 121 static int jobid_should_free_item(void *obj, void *data)
 122 {
 123         char *jobid = data;
 124         struct jobid_pid_map *pidmap = obj;
 125         int rc = 0;
 126
 127         if (obj == NULL)
 128                 return 0;
 129
 130         spin_lock(&pidmap->jp_lock);
 131         if (jobid == NULL)
 132                 rc = 1;
 133         else if (jobid[0] == '\0')
 134                 rc = 1;
 135         else if (ktime_get_real_seconds() - pidmap->jp_time > DELETE_INTERVAL)
 136                 rc = 1;
 137         else if (strcmp(pidmap->jp_jobid, jobid) == 0)
 138                 rc = 1;
 139         spin_unlock(&pidmap->jp_lock);
 140
 141         return rc;
 142 }
 143
 144 /*
 145  * jobid_name_is_valid
 146  *
 147  * Checks if the jobid is a Lustre process
 148  *
 149  * Returns true if jobid is valid
 150  * Returns false if jobid looks like it's a Lustre process
 151  */
 152 static bool jobid_name_is_valid(char *jobid)
 153 {
 154         const char *const lustre_reserved[] = { "ll_ping", "ptlrpc",
 155                                                 "ldlm", "ll_sa", NULL };
 156         int i;
 157
 158         if (jobid[0] == '\0')
 159                 return false;
 160
 161         for (i = 0; lustre_reserved[i] != NULL; i++) {
 162                 if (strncmp(jobid, lustre_reserved[i],
 163                             strlen(lustre_reserved[i])) == 0)
 164                         return false;
 165         }
 166         return true;
 167 }
 168
 169 /*
 170  * jobid_get_from_cache()
 171  *
 172  * Returns contents of jobid_var from process environment for current PID.
 173  * This will be cached for some time to avoid overhead scanning environment.
 174  *
 175  * Return: -ENOMEM if allocating a new pidmap fails
 176  *         -ENOENT if no entry could be found
 177  *         +ve string length for success (something was returned in jobid)
 178  */
 179 static int jobid_get_from_cache(char *jobid, size_t joblen)
 180 {
 181         static time64_t last_expire;
 182         bool expire_cache = false;
 183         pid_t pid = current_pid();
 184         struct jobid_pid_map *pidmap = NULL;
 185         time64_t now = ktime_get_real_seconds();
 186         int rc = 0;
 187         ENTRY;
 188
 189         LASSERT(jobid_hash != NULL);
 190
 191         /* scan hash periodically to remove old PID entries from cache */
 192         spin_lock(&jobid_hash_lock);
 193         if (unlikely(last_expire + DELETE_INTERVAL <= now)) {
 194                 expire_cache = true;
 195                 last_expire = now;
 196         }
 197         spin_unlock(&jobid_hash_lock);
 198
 199         if (expire_cache)
 200                 cfs_hash_cond_del(jobid_hash, jobid_should_free_item,
 201                                   "intentionally_bad_jobid");
 202
 203         /* first try to find PID in the hash and use that value */
 204         pidmap = cfs_hash_lookup(jobid_hash, &pid);
 205         if (pidmap == NULL) {
 206                 struct jobid_pid_map *pidmap2;
 207
 208                 OBD_ALLOC_PTR(pidmap);
 209                 if (pidmap == NULL)
 210                         GOTO(out, rc = -ENOMEM);
 211
 212                 pidmap->jp_pid = pid;
 213                 pidmap->jp_time = 0;
 214                 pidmap->jp_jobid[0] = '\0';
 215                 spin_lock_init(&pidmap->jp_lock);
 216                 INIT_HLIST_NODE(&pidmap->jp_hash);
 217
 218                 /*
 219                  * Add the newly created map to the hash, on key collision we
 220                  * lost a racing addition and must destroy our newly allocated
 221                  * map.  The object which exists in the hash will be returned.
 222                  */
 223                 pidmap2 = cfs_hash_findadd_unique(jobid_hash, &pid,
 224                                                   &pidmap->jp_hash);
 225                 if (unlikely(pidmap != pidmap2)) {
 226                         CDEBUG(D_INFO, "jobid: duplicate found for PID=%u\n",
 227                                pid);
 228                         OBD_FREE_PTR(pidmap);
 229                         pidmap = pidmap2;
 230                 } else {
 231                         cfs_hash_get(jobid_hash, &pidmap->jp_hash);
 232                 }
 233         }
 234
 235         /*
 236          * If pidmap is old (this is always true for new entries) refresh it.
 237          * If obd_jobid_var is not found, cache empty entry and try again
 238          * later, to avoid repeat lookups for PID if obd_jobid_var missing.
 239          */
 240         spin_lock(&pidmap->jp_lock);
 241         if (pidmap->jp_time + RESCAN_INTERVAL <= now) {
 242                 char env_jobid[LUSTRE_JOBID_SIZE] = "";
 243                 int env_len = sizeof(env_jobid);
 244
 245                 pidmap->jp_time = now;
 246
 247                 spin_unlock(&pidmap->jp_lock);
 248                 rc = jobid_get_from_environ(obd_jobid_var, env_jobid, &env_len);
 249
 250                 CDEBUG(D_INFO, "jobid: PID mapping established: %d->%s\n",
 251                        pidmap->jp_pid, env_jobid);
 252                 spin_lock(&pidmap->jp_lock);
 253                 if (!rc) {
 254                         pidmap->jp_joblen = env_len;
 255                         strlcpy(pidmap->jp_jobid, env_jobid,
 256                                 sizeof(pidmap->jp_jobid));
 257                         rc = 0;
 258                 } else if (rc == -ENOENT) {
 259                         /* It might have been deleted, clear out old entry */
 260                         pidmap->jp_joblen = 0;
 261                         pidmap->jp_jobid[0] = '\0';
 262                 }
 263         }
 264
 265         /*
 266          * Regardless of how pidmap was found, if it contains a valid entry
 267          * use that for now.  If there was a technical error (e.g. -ENOMEM)
 268          * use the old cached value until it can be looked up again properly.
 269          * If a cached missing entry was found, return -ENOENT.
 270          */
 271         if (pidmap->jp_joblen) {
 272                 strlcpy(jobid, pidmap->jp_jobid, joblen);
 273                 joblen = pidmap->jp_joblen;
 274                 rc = 0;
 275         } else if (!rc) {
 276                 rc = -ENOENT;
 277         }
 278         spin_unlock(&pidmap->jp_lock);
 279
 280         cfs_hash_put(jobid_hash, &pidmap->jp_hash);
 281
 282         EXIT;
 283 out:
 284         return rc < 0 ? rc : joblen;
 285 }
 286
 287 /*
 288  * jobid_interpret_string()
 289  *
 290  * Interpret the jobfmt string to expand specified fields, like coredumps do:
 291  *   %e = executable
 292  *   %g = gid
 293  *   %h = hostname
 294  *   %j = jobid from environment
 295  *   %p = pid
 296  *   %u = uid
 297  *
 298  * Unknown escape strings are dropped.  Other characters are copied through,
 299  * excluding whitespace (to avoid making jobid parsing difficult).
 300  *
 301  * Return: -EOVERFLOW if the expanded string does not fit within @joblen
 302  *         0 for success
 303  */
 304 static int jobid_interpret_string(const char *jobfmt, char *jobid,
 305                                   ssize_t joblen)
 306 {
 307         char c;
 308
 309         while ((c = *jobfmt++) && joblen > 1) {
 310                 char f;
 311                 int l;
 312
 313                 if (isspace(c)) /* Don't allow embedded spaces */
 314                         continue;
 315
 316                 if (c != '%') {
 317                         *jobid = c;
 318                         joblen--;
 319                         jobid++;
 320                         continue;
 321                 }
 322
 323                 switch ((f = *jobfmt++)) {
 324                 case 'e': /* executable name */
 325                         l = snprintf(jobid, joblen, "%s", current_comm());
 326                         break;
 327                 case 'g': /* group ID */
 328                         l = snprintf(jobid, joblen, "%u",
 329                                      from_kgid(&init_user_ns, current_fsgid()));
 330                         break;
 331                 case 'h': /* hostname */
 332                         l = snprintf(jobid, joblen, "%s",
 333                                      init_utsname()->nodename);
 334                         break;
 335                 case 'j': /* jobid stored in process environment */
 336                         l = jobid_get_from_cache(jobid, joblen);
 337                         if (l < 0)
 338                                 l = 0;
 339                         break;
 340                 case 'p': /* process ID */
 341                         l = snprintf(jobid, joblen, "%u", current_pid());
 342                         break;
 343                 case 'u': /* user ID */
 344                         l = snprintf(jobid, joblen, "%u",
 345                                      from_kuid(&init_user_ns, current_fsuid()));
 346                         break;
 347                 case '\0': /* '%' at end of format string */
 348                         l = 0;
 349                         goto out;
 350                 default: /* drop unknown %x format strings */
 351                         l = 0;
 352                         break;
 353                 }
 354                 jobid += l;
 355                 joblen -= l;
 356         }
 357         /*
 358          * This points at the end of the buffer, so long as jobid is always
 359          * incremented the same amount as joblen is decremented.
 360          */
 361 out:
 362         jobid[joblen - 1] = '\0';
 363
 364         return joblen < 0 ? -EOVERFLOW : 0;
 365 }
 366
 367 /*
 368  * Hash initialization, copied from server-side job stats bucket sizes
 369  */
 370 #define HASH_JOBID_BKT_BITS 5
 371 #define HASH_JOBID_CUR_BITS 7
 372 #define HASH_JOBID_MAX_BITS 12
 373
 374 int jobid_cache_init(void)
 375 {
 376         int rc = 0;
 377         ENTRY;
 378
 379         if (jobid_hash)
 380                 return 0;
 381
 382         spin_lock_init(&jobid_hash_lock);
 383         jobid_hash = cfs_hash_create("JOBID_HASH", HASH_JOBID_CUR_BITS,
 384                                      HASH_JOBID_MAX_BITS, HASH_JOBID_BKT_BITS,
 385                                      0, CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
 386                                      &jobid_hash_ops, CFS_HASH_DEFAULT);
 387         if (!jobid_hash)
 388                 rc = -ENOMEM;
 389
 390         RETURN(rc);
 391 }
 392 EXPORT_SYMBOL(jobid_cache_init);
 393
 394 void jobid_cache_fini(void)
 395 {
 396         struct cfs_hash *tmp_hash;
 397         ENTRY;
 398
 399         spin_lock(&jobid_hash_lock);
 400         tmp_hash = jobid_hash;
 401         jobid_hash = NULL;
 402         spin_unlock(&jobid_hash_lock);
 403
 404         if (tmp_hash != NULL) {
 405                 cfs_hash_cond_del(tmp_hash, jobid_should_free_item, NULL);
 406                 cfs_hash_putref(tmp_hash);
 407         }
 408
 409         EXIT;
 410 }
 411 EXPORT_SYMBOL(jobid_cache_fini);
 412
 413 /*
 414  * Hash operations for pid<->jobid
 415  */
 416 static unsigned jobid_hashfn(struct cfs_hash *hs, const void *key,
 417                              unsigned mask)
 418 {
 419         return cfs_hash_djb2_hash(key, sizeof(pid_t), mask);
 420 }
 421
 422 static void *jobid_key(struct hlist_node *hnode)
 423 {
 424         struct jobid_pid_map *pidmap;
 425
 426         pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 427         return &pidmap->jp_pid;
 428 }
 429
 430 static int jobid_keycmp(const void *key, struct hlist_node *hnode)
 431 {
 432         const pid_t *pid_key1;
 433         const pid_t *pid_key2;
 434
 435         LASSERT(key != NULL);
 436         pid_key1 = (pid_t *)key;
 437         pid_key2 = (pid_t *)jobid_key(hnode);
 438
 439         return *pid_key1 == *pid_key2;
 440 }
 441
 442 static void *jobid_object(struct hlist_node *hnode)
 443 {
 444         return hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 445 }
 446
 447 static void jobid_get(struct cfs_hash *hs, struct hlist_node *hnode)
 448 {
 449         struct jobid_pid_map *pidmap;
 450
 451         pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 452
 453         atomic_inc(&pidmap->jp_refcount);
 454 }
 455
 456 static void jobid_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 457 {
 458         struct jobid_pid_map *pidmap;
 459
 460         if (hnode == NULL)
 461                 return;
 462
 463         pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 464         LASSERT(atomic_read(&pidmap->jp_refcount) > 0);
 465         if (atomic_dec_and_test(&pidmap->jp_refcount)) {
 466                 CDEBUG(D_INFO, "Freeing: %d->%s\n",
 467                        pidmap->jp_pid, pidmap->jp_jobid);
 468
 469                 OBD_FREE_PTR(pidmap);
 470         }
 471 }
 472
 473 static struct cfs_hash_ops jobid_hash_ops = {
 474         .hs_hash        = jobid_hashfn,
 475         .hs_keycmp      = jobid_keycmp,
 476         .hs_key         = jobid_key,
 477         .hs_object      = jobid_object,
 478         .hs_get         = jobid_get,
 479         .hs_put         = jobid_put_locked,
 480         .hs_put_locked  = jobid_put_locked,
 481 };
 482
 483 /**
 484  * Generate the job identifier string for this process for tracking purposes.
 485  *
 486  * Fill in @jobid string based on the value of obd_jobid_var:
 487  * JOBSTATS_DISABLE:      none
 488  * JOBSTATS_NODELOCAL:    content of obd_jobid_node (jobid_interpret_string())
 489  * JOBSTATS_PROCNAME_UID: process name/UID
 490  * anything else:         look up obd_jobid_var in the processes environment
 491  *
 492  * Return -ve error number, 0 on success.
 493  */
 494 int lustre_get_jobid(char *jobid, size_t joblen)
 495 {
 496         int rc = 0;
 497         ENTRY;
 498
 499         if (unlikely(joblen < 2)) {
 500                 if (joblen == 1)
 501                         jobid[0] = '\0';
 502                 RETURN(-EINVAL);
 503         }
 504
 505         if (strcmp(obd_jobid_var, JOBSTATS_DISABLE) == 0) {
 506                 /* Jobstats isn't enabled */
 507                 memset(jobid, 0, joblen);
 508         } else if (strcmp(obd_jobid_var, JOBSTATS_NODELOCAL) == 0) {
 509                 /* Whole node dedicated to single job */
 510                 rc = jobid_interpret_string(obd_jobid_name, jobid, joblen);
 511         } else if (strcmp(obd_jobid_var, JOBSTATS_PROCNAME_UID) == 0) {
 512                 rc = jobid_interpret_string("%e.%u", jobid, joblen);
 513         } else if (jobid_name_is_valid(current_comm())) {
 514                 /*
 515                  * obd_jobid_var holds the jobid environment variable name.
 516                  * Skip initial check if obd_jobid_name already uses "%j",
 517                  * otherwise try just "%j" first, then fall back to whatever
 518                  * is in obd_jobid_name if obd_jobid_var is not found.
 519                  */
 520                 rc = -EAGAIN;
 521                 if (!strnstr(obd_jobid_name, "%j", joblen))
 522                         rc = jobid_get_from_cache(jobid, joblen);
 523
 524                 /* fall back to jobid_node if jobid_var not in environment */
 525                 if (rc < 0) {
 526                         int rc2 = jobid_interpret_string(obd_jobid_name,
 527                                                          jobid, joblen);
 528                         if (!rc2)
 529                                 rc = 0;
 530                 }
 531         }
 532
 533         RETURN(rc);
 534 }
 535 EXPORT_SYMBOL(lustre_get_jobid);
 536
 537 /*
 538  * lustre_jobid_clear
 539  *
 540  * Search cache for JobID given by @find_jobid.
 541  * If any entries in the hash table match the value, they are removed
 542  */
 543 void lustre_jobid_clear(const char *find_jobid)
 544 {
 545         char jobid[LUSTRE_JOBID_SIZE];
 546         char *end;
 547
 548         if (jobid_hash == NULL)
 549                 return;
 550
 551         strlcpy(jobid, find_jobid, sizeof(jobid));
 552         /* trim \n off the end of the incoming jobid */
 553         end = strchr(jobid, '\n');
 554         if (end && *end == '\n')
 555                 *end = '\0';
 556
 557         CDEBUG(D_INFO, "Clearing Jobid: %s\n", jobid);
 558         cfs_hash_cond_del(jobid_hash, jobid_should_free_item, jobid);
 559
 560         CDEBUG(D_INFO, "%d items remain in jobID table\n",
 561                atomic_read(&jobid_hash->hs_count));
 562 }