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 #include <linux/uidgid.h>
  37 #include <linux/utsname.h>
  38
  39 #include <libcfs/libcfs.h>
  40 #include <obd_support.h>
  41 #include <obd_class.h>
  42 #include <lustre_net.h>
  43
  44 static struct cfs_hash *jobid_hash;
  45 static struct cfs_hash_ops jobid_hash_ops;
  46 spinlock_t jobid_hash_lock;
  47
  48 #define RESCAN_INTERVAL 30
  49 #define DELETE_INTERVAL 300
  50
  51 char obd_jobid_var[JOBSTATS_JOBID_VAR_MAX_LEN + 1] = JOBSTATS_DISABLE;
  52 char obd_jobid_name[LUSTRE_JOBID_SIZE] = "%e.%u";
  53
  54 /**
  55  * Structure to store a single PID->JobID mapping
  56  */
  57 struct jobid_pid_map {
  58         struct hlist_node       jp_hash;
  59         time64_t                jp_time;
  60         spinlock_t              jp_lock; /* protects jp_jobid */
  61         char                    jp_jobid[LUSTRE_JOBID_SIZE];
  62         unsigned int            jp_joblen;
  63         struct kref             jp_refcount;
  64         pid_t                   jp_pid;
  65 };
  66
  67 /*
  68  * Jobid can be set for a session (see setsid(2)) by writing to
  69  * a sysfs file from any process in that session.
  70  * The jobids are stored in a hash table indexed by the relevant
  71  * struct pid.  We periodically look for entries where the pid has
  72  * no PIDTYPE_SID tasks any more, and prune them.  This happens within
  73  * 5 seconds of a jobid being added, and every 5 minutes when jobids exist,
  74  * but none are added.
  75  */
  76 #define JOBID_EXPEDITED_CLEAN (5)
  77 #define JOBID_BACKGROUND_CLEAN (5 * 60)
  78
  79 struct session_jobid {
  80         struct pid              *sj_session;
  81         struct rhash_head       sj_linkage;
  82         struct rcu_head         sj_rcu;
  83         char                    sj_jobid[1];
  84 };
  85
  86 static const struct rhashtable_params jobid_params = {
  87         .key_len        = sizeof(struct pid *),
  88         .key_offset     = offsetof(struct session_jobid, sj_session),
  89         .head_offset    = offsetof(struct session_jobid, sj_linkage),
  90 };
  91
  92 static struct rhashtable session_jobids;
  93
  94 /*
  95  * jobid_current must be called with rcu_read_lock held.
  96  * if it returns non-NULL, the string can only be used
  97  * until rcu_read_unlock is called.
  98  */
  99 char *jobid_current(void)
 100 {
 101         struct pid *sid = task_session(current);
 102         struct session_jobid *sj;
 103
 104         sj = rhashtable_lookup_fast(&session_jobids, &sid, jobid_params);
 105         if (sj)
 106                 return sj->sj_jobid;
 107         return NULL;
 108 }
 109
 110 static void jobid_prune_expedite(void);
 111 /*
 112  * jobid_set_current will try to add a new entry
 113  * to the table.  If one exists with the same key, the
 114  * jobid will be replaced
 115  */
 116 int jobid_set_current(char *jobid)
 117 {
 118         struct pid *sid;
 119         struct session_jobid *sj, *origsj;
 120         int ret;
 121         int len = strlen(jobid);
 122
 123         sj = kmalloc(sizeof(*sj) + len, GFP_KERNEL);
 124         if (!sj)
 125                 return -ENOMEM;
 126         rcu_read_lock();
 127         sid = task_session(current);
 128         sj->sj_session = get_pid(sid);
 129         strncpy(sj->sj_jobid, jobid, len+1);
 130         origsj = rhashtable_lookup_get_insert_fast(&session_jobids,
 131                                                    &sj->sj_linkage,
 132                                                    jobid_params);
 133         if (origsj == NULL) {
 134                 /* successful insert */
 135                 rcu_read_unlock();
 136                 jobid_prune_expedite();
 137                 return 0;
 138         }
 139
 140         if (IS_ERR(origsj)) {
 141                 put_pid(sj->sj_session);
 142                 kfree(sj);
 143                 rcu_read_unlock();
 144                 return PTR_ERR(origsj);
 145         }
 146         ret = rhashtable_replace_fast(&session_jobids,
 147                                       &origsj->sj_linkage,
 148                                       &sj->sj_linkage,
 149                                       jobid_params);
 150         if (ret) {
 151                 put_pid(sj->sj_session);
 152                 kfree(sj);
 153                 rcu_read_unlock();
 154                 return ret;
 155         }
 156         put_pid(origsj->sj_session);
 157         rcu_read_unlock();
 158         kfree_rcu(origsj, sj_rcu);
 159         jobid_prune_expedite();
 160
 161         return 0;
 162 }
 163
 164 static void jobid_free(void *vsj, void *arg)
 165 {
 166         struct session_jobid *sj = vsj;
 167
 168         put_pid(sj->sj_session);
 169         kfree(sj);
 170 }
 171
 172 static void jobid_prune(struct work_struct *work);
 173 static DECLARE_DELAYED_WORK(jobid_prune_work, jobid_prune);
 174 static int jobid_prune_expedited;
 175 static void jobid_prune(struct work_struct *work)
 176 {
 177         int remaining = 0;
 178         struct rhashtable_iter iter;
 179         struct session_jobid *sj;
 180
 181         jobid_prune_expedited = 0;
 182         rhashtable_walk_enter(&session_jobids, &iter);
 183         rhashtable_walk_start(&iter);
 184         while ((sj = rhashtable_walk_next(&iter)) != NULL) {
 185                 if (IS_ERR(sj)) {
 186                         if (PTR_ERR(sj) == -EAGAIN)
 187                                 continue;
 188                         break;
 189                 }
 190                 if (!hlist_empty(&sj->sj_session->tasks[PIDTYPE_SID])) {
 191                         remaining++;
 192                         continue;
 193                 }
 194                 if (rhashtable_remove_fast(&session_jobids,
 195                                            &sj->sj_linkage,
 196                                            jobid_params) == 0) {
 197                         put_pid(sj->sj_session);
 198                         kfree_rcu(sj, sj_rcu);
 199                 }
 200         }
 201         rhashtable_walk_stop(&iter);
 202         rhashtable_walk_exit(&iter);
 203         if (remaining)
 204                 schedule_delayed_work(&jobid_prune_work,
 205                                       cfs_time_seconds(JOBID_BACKGROUND_CLEAN));
 206 }
 207
 208 static void jobid_prune_expedite(void)
 209 {
 210         if (!jobid_prune_expedited) {
 211                 jobid_prune_expedited = 1;
 212                 mod_delayed_work(system_wq, &jobid_prune_work,
 213                                  cfs_time_seconds(JOBID_EXPEDITED_CLEAN));
 214         }
 215 }
 216
 217 static int cfs_access_process_vm(struct task_struct *tsk,
 218                                  struct mm_struct *mm,
 219                                  unsigned long addr,
 220                                  void *buf, int len, int write)
 221 {
 222         /* Just copied from kernel for the kernels which doesn't
 223          * have access_process_vm() exported
 224          */
 225         struct vm_area_struct *vma = NULL;
 226         struct page *page;
 227         void *old_buf = buf;
 228
 229         /* Avoid deadlocks on mmap_sem if called from sys_mmap_pgoff(),
 230          * which is already holding mmap_sem for writes.  If some other
 231          * thread gets the write lock in the meantime, this thread will
 232          * block, but at least it won't deadlock on itself.  LU-1735
 233          */
 234         if (!mmap_read_trylock(mm))
 235                 return -EDEADLK;
 236
 237         /* ignore errors, just check how much was successfully transferred */
 238         while (len) {
 239                 int bytes, rc, offset;
 240                 void *maddr;
 241
 242 #if defined(HAVE_GET_USER_PAGES_WITHOUT_VMA)
 243                 rc = get_user_pages(addr, 1, write ? FOLL_WRITE : 0, &page);
 244                 if (rc > 0)
 245                         vma = vma_lookup(mm, addr);
 246 #elif defined(HAVE_GET_USER_PAGES_GUP_FLAGS)
 247                 rc = get_user_pages(addr, 1, write ? FOLL_WRITE : 0, &page,
 248                                     &vma);
 249 #elif defined(HAVE_GET_USER_PAGES_6ARG)
 250                 rc = get_user_pages(addr, 1, write, 1, &page, &vma);
 251 #else
 252                 rc = get_user_pages(tsk, mm, addr, 1, write, 1, &page, &vma);
 253 #endif
 254                 if (rc <= 0 || !vma)
 255                         break;
 256
 257                 bytes = len;
 258                 offset = addr & (PAGE_SIZE-1);
 259                 if (bytes > PAGE_SIZE-offset)
 260                         bytes = PAGE_SIZE-offset;
 261
 262                 maddr = kmap(page);
 263                 if (write) {
 264                         copy_to_user_page(vma, page, addr,
 265                                           maddr + offset, buf, bytes);
 266                         set_page_dirty_lock(page);
 267                 } else {
 268                         copy_from_user_page(vma, page, addr,
 269                                             buf, maddr + offset, bytes);
 270                 }
 271                 kunmap(page);
 272                 put_page(page);
 273                 len -= bytes;
 274                 buf += bytes;
 275                 addr += bytes;
 276         }
 277         mmap_read_unlock(mm);
 278
 279         return buf - old_buf;
 280 }
 281
 282 /* Read the environment variable of current process specified by @key. */
 283 static int cfs_get_environ(const char *key, char *value, int *val_len)
 284 {
 285         struct mm_struct *mm;
 286         char *buffer;
 287         int buf_len = PAGE_SIZE;
 288         int key_len = strlen(key);
 289         unsigned long addr;
 290         int rc;
 291         bool skip = false;
 292
 293         ENTRY;
 294         buffer = kmalloc(buf_len, GFP_USER);
 295         if (!buffer)
 296                 RETURN(-ENOMEM);
 297
 298         mm = get_task_mm(current);
 299         if (!mm) {
 300                 kfree(buffer);
 301                 RETURN(-EINVAL);
 302         }
 303
 304         addr = mm->env_start;
 305         while (addr < mm->env_end) {
 306                 int this_len, retval, scan_len;
 307                 char *env_start, *env_end;
 308
 309                 memset(buffer, 0, buf_len);
 310
 311                 this_len = min_t(int, mm->env_end - addr, buf_len);
 312                 retval = cfs_access_process_vm(current, mm, addr, buffer,
 313                                                this_len, 0);
 314                 if (retval < 0)
 315                         GOTO(out, rc = retval);
 316                 else if (retval != this_len)
 317                         break;
 318
 319                 addr += retval;
 320
 321                 /* Parse the buffer to find out the specified key/value pair.
 322                  * The "key=value" entries are separated by '\0'.
 323                  */
 324                 env_start = buffer;
 325                 scan_len = this_len;
 326                 while (scan_len) {
 327                         char *entry;
 328                         int entry_len;
 329
 330                         env_end = memscan(env_start, '\0', scan_len);
 331                         LASSERT(env_end >= env_start &&
 332                                 env_end <= env_start + scan_len);
 333
 334                         /* The last entry of this buffer cross the buffer
 335                          * boundary, reread it in next cycle.
 336                          */
 337                         if (unlikely(env_end - env_start == scan_len)) {
 338                                 /* Just skip the entry larger than page size,
 339                                  * it can't be jobID env variable.
 340                                  */
 341                                 if (unlikely(scan_len == this_len))
 342                                         skip = true;
 343                                 else
 344                                         addr -= scan_len;
 345                                 break;
 346                         } else if (unlikely(skip)) {
 347                                 skip = false;
 348                                 goto skip;
 349                         }
 350                         entry = env_start;
 351                         entry_len = env_end - env_start;
 352                         CDEBUG(D_INFO, "key: %s, entry: %s\n", key, entry);
 353
 354                         /* Key length + length of '=' */
 355                         if (entry_len > key_len + 1 &&
 356                             entry[key_len] == '='  &&
 357                             !memcmp(entry, key, key_len)) {
 358                                 entry += key_len + 1;
 359                                 entry_len -= key_len + 1;
 360
 361                                 /* The 'value' buffer passed in is too small.
 362                                  * Copy what fits, but return -EOVERFLOW.
 363                                  */
 364                                 if (entry_len >= *val_len) {
 365                                         memcpy(value, entry, *val_len);
 366                                         value[*val_len - 1] = 0;
 367                                         GOTO(out, rc = -EOVERFLOW);
 368                                 }
 369
 370                                 memcpy(value, entry, entry_len);
 371                                 *val_len = entry_len;
 372                                 GOTO(out, rc = 0);
 373                         }
 374 skip:
 375                         scan_len -= (env_end - env_start + 1);
 376                         env_start = env_end + 1;
 377                 }
 378         }
 379         GOTO(out, rc = -ENOENT);
 380
 381 out:
 382         mmput(mm);
 383         kfree((void *)buffer);
 384         return rc;
 385 }
 386
 387 /*
 388  * Get jobid of current process by reading the environment variable
 389  * stored in between the "env_start" & "env_end" of task struct.
 390  *
 391  * If some job scheduler doesn't store jobid in the "env_start/end",
 392  * then an upcall could be issued here to get the jobid by utilizing
 393  * the userspace tools/API. Then, the jobid must be cached.
 394  */
 395 static int jobid_get_from_environ(char *jobid_var, char *jobid, int *jobid_len)
 396 {
 397         int rc;
 398
 399         rc = cfs_get_environ(jobid_var, jobid, jobid_len);
 400         if (!rc)
 401                 goto out;
 402
 403         if (rc == -EOVERFLOW) {
 404                 /* For the PBS_JOBID and LOADL_STEP_ID keys (which are
 405                  * variable length strings instead of just numbers), it
 406                  * might make sense to keep the unique parts for JobID,
 407                  * instead of just returning an error.  That means a
 408                  * larger temp buffer for cfs_get_environ(), then
 409                  * truncating the string at some separator to fit into
 410                  * the specified jobid_len.  Fix later if needed. */
 411                 static ktime_t printed;
 412
 413                 if (unlikely(ktime_to_ns(printed) == 0 ||
 414                              ktime_after(ktime_get(),
 415                                          ktime_add_ns(printed,
 416                                              3600ULL * 24 * NSEC_PER_SEC)))) {
 417                         LCONSOLE_WARN("jobid: '%s' value too large (%d)\n",
 418                                       obd_jobid_var, *jobid_len);
 419                         printed = ktime_get();
 420                 }
 421
 422                 rc = 0;
 423         } else {
 424                 CDEBUG_LIMIT((rc == -ENOENT || rc == -EINVAL ||
 425                               rc == -EDEADLK) ? D_INFO : D_ERROR,
 426                              "jobid: get '%s' failed: rc = %d\n",
 427                              obd_jobid_var, rc);
 428         }
 429
 430 out:
 431         return rc;
 432 }
 433
 434 /*
 435  * jobid_should_free_item
 436  *
 437  * Each item is checked to see if it should be released
 438  * Removed from hash table by caller
 439  * Actually freed in jobid_put_locked
 440  *
 441  * Returns 1 if item is to be freed, 0 if it is to be kept
 442  */
 443
 444 static int jobid_should_free_item(void *obj, void *data)
 445 {
 446         char *jobid = data;
 447         struct jobid_pid_map *pidmap = obj;
 448         int rc = 0;
 449
 450         if (obj == NULL)
 451                 return 0;
 452
 453         if (jobid == NULL) {
 454                 WARN_ON_ONCE(kref_read(&pidmap->jp_refcount) != 1);
 455                 return 1;
 456         }
 457
 458         spin_lock(&pidmap->jp_lock);
 459         /* prevent newly inserted items from deleting */
 460         if (jobid[0] == '\0' && kref_read(&pidmap->jp_refcount) == 1)
 461                 rc = 1;
 462         else if (ktime_get_real_seconds() - pidmap->jp_time > DELETE_INTERVAL)
 463                 rc = 1;
 464         else if (strcmp(pidmap->jp_jobid, jobid) == 0)
 465                 rc = 1;
 466         spin_unlock(&pidmap->jp_lock);
 467
 468         return rc;
 469 }
 470
 471 /*
 472  * jobid_name_is_valid
 473  *
 474  * Checks if the jobid is a Lustre process
 475  *
 476  * Returns true if jobid is valid
 477  * Returns false if jobid looks like it's a Lustre process
 478  */
 479 static bool jobid_name_is_valid(char *jobid)
 480 {
 481         const char *const lustre_reserved[] = { "ll_ping", "ptlrpc",
 482                                                 "ldlm", "ll_sa", "kworker",
 483                                                 "kswapd", "writeback", "irq",
 484                                                 "ksoftirq", NULL };
 485         int i;
 486
 487         if (jobid[0] == '\0')
 488                 return false;
 489
 490         for (i = 0; lustre_reserved[i] != NULL; i++) {
 491                 if (strncmp(jobid, lustre_reserved[i],
 492                             strlen(lustre_reserved[i])) == 0)
 493                         return false;
 494         }
 495         return true;
 496 }
 497
 498 /*
 499  * jobid_get_from_cache()
 500  *
 501  * Returns contents of jobid_var from process environment for current PID,
 502  * or from the per-session jobid table.
 503  * Values fetch from process environment will be cached for some time to avoid
 504  * the overhead of scanning the environment.
 505  *
 506  * Return: -ENOMEM if allocating a new pidmap fails
 507  *         -ENOENT if no entry could be found
 508  *         +ve string length for success (something was returned in jobid)
 509  */
 510 static int jobid_get_from_cache(char *jobid, size_t joblen)
 511 {
 512         static time64_t last_expire;
 513         bool expire_cache = false;
 514         pid_t pid = current->pid;
 515         struct jobid_pid_map *pidmap = NULL;
 516         time64_t now = ktime_get_real_seconds();
 517         int rc = 0;
 518         ENTRY;
 519
 520         if (strcmp(obd_jobid_var, JOBSTATS_SESSION) == 0) {
 521                 char *jid;
 522
 523                 rcu_read_lock();
 524                 jid = jobid_current();
 525                 if (jid) {
 526                         strscpy(jobid, jid, joblen);
 527                         joblen = strlen(jobid);
 528                 } else {
 529                         rc = -ENOENT;
 530                 }
 531                 rcu_read_unlock();
 532                 GOTO(out, rc);
 533         }
 534
 535         LASSERT(jobid_hash != NULL);
 536
 537         /* scan hash periodically to remove old PID entries from cache */
 538         spin_lock(&jobid_hash_lock);
 539         if (unlikely(last_expire + DELETE_INTERVAL <= now)) {
 540                 expire_cache = true;
 541                 last_expire = now;
 542         }
 543         spin_unlock(&jobid_hash_lock);
 544
 545         if (expire_cache)
 546                 cfs_hash_cond_del(jobid_hash, jobid_should_free_item,
 547                                   "intentionally_bad_jobid");
 548
 549         /* first try to find PID in the hash and use that value */
 550         pidmap = cfs_hash_lookup(jobid_hash, &pid);
 551         if (pidmap == NULL) {
 552                 struct jobid_pid_map *pidmap2;
 553
 554                 OBD_ALLOC_PTR(pidmap);
 555                 if (pidmap == NULL)
 556                         GOTO(out, rc = -ENOMEM);
 557
 558                 pidmap->jp_pid = pid;
 559                 pidmap->jp_time = 0;
 560                 pidmap->jp_jobid[0] = '\0';
 561                 spin_lock_init(&pidmap->jp_lock);
 562                 INIT_HLIST_NODE(&pidmap->jp_hash);
 563                 /*
 564                  * @pidmap might be reclaimed just after it is added into
 565                  * hash list, init @jp_refcount as 1 to make sure memory
 566                  * could be not freed during access.
 567                  */
 568                 kref_init(&pidmap->jp_refcount);
 569
 570                 /*
 571                  * Add the newly created map to the hash, on key collision we
 572                  * lost a racing addition and must destroy our newly allocated
 573                  * map.  The object which exists in the hash will be returned.
 574                  */
 575                 pidmap2 = cfs_hash_findadd_unique(jobid_hash, &pid,
 576                                                   &pidmap->jp_hash);
 577                 if (unlikely(pidmap != pidmap2)) {
 578                         CDEBUG(D_INFO, "jobid: duplicate found for PID=%u\n",
 579                                pid);
 580                         OBD_FREE_PTR(pidmap);
 581                         pidmap = pidmap2;
 582                 }
 583         }
 584
 585         /*
 586          * If pidmap is old (this is always true for new entries) refresh it.
 587          * If obd_jobid_var is not found, cache empty entry and try again
 588          * later, to avoid repeat lookups for PID if obd_jobid_var missing.
 589          */
 590         spin_lock(&pidmap->jp_lock);
 591         if (pidmap->jp_time + RESCAN_INTERVAL <= now) {
 592                 char env_jobid[LUSTRE_JOBID_SIZE] = "";
 593                 int env_len = sizeof(env_jobid);
 594
 595                 pidmap->jp_time = now;
 596
 597                 spin_unlock(&pidmap->jp_lock);
 598                 rc = jobid_get_from_environ(obd_jobid_var, env_jobid, &env_len);
 599
 600                 CDEBUG(D_INFO, "jobid: PID mapping established: %d->%s\n",
 601                        pidmap->jp_pid, env_jobid);
 602                 spin_lock(&pidmap->jp_lock);
 603                 if (!rc) {
 604                         pidmap->jp_joblen = env_len;
 605                         strscpy(pidmap->jp_jobid, env_jobid,
 606                                 sizeof(pidmap->jp_jobid));
 607                         rc = 0;
 608                 } else if (rc == -ENOENT) {
 609                         /* It might have been deleted, clear out old entry */
 610                         pidmap->jp_joblen = 0;
 611                         pidmap->jp_jobid[0] = '\0';
 612                 }
 613         }
 614
 615         /*
 616          * Regardless of how pidmap was found, if it contains a valid entry
 617          * use that for now.  If there was a technical error (e.g. -ENOMEM)
 618          * use the old cached value until it can be looked up again properly.
 619          * If a cached missing entry was found, return -ENOENT.
 620          */
 621         if (pidmap->jp_joblen) {
 622                 strscpy(jobid, pidmap->jp_jobid, joblen);
 623                 joblen = pidmap->jp_joblen;
 624                 rc = 0;
 625         } else if (!rc) {
 626                 rc = -ENOENT;
 627         }
 628         spin_unlock(&pidmap->jp_lock);
 629
 630         cfs_hash_put(jobid_hash, &pidmap->jp_hash);
 631
 632         EXIT;
 633 out:
 634         return rc < 0 ? rc : joblen;
 635 }
 636
 637 /*
 638  * jobid_print_current_comm()
 639  *
 640  * Print current comm name into the provided jobid buffer, and trim names of
 641  * kernel threads like "kworker/0:0" to "kworker" or "ll_sa_12345" to "ll_sa"
 642  *
 643  * Return: number of chars printed to jobid
 644  */
 645 static int jobid_print_current_comm(char *jobid, ssize_t joblen)
 646 {
 647         const char *const names[] = {"kworker", "kswapd", "ll_sa", "ll_agl",
 648                                      "ldlm_bl", NULL};
 649         int i;
 650
 651         if (current->flags & PF_KTHREAD) {
 652                 for (i = 0; names[i] != NULL; i++) {
 653                         if (strncmp(current->comm, names[i],
 654                                     strlen(names[i])) == 0)
 655                                 return snprintf(jobid, joblen, "%s", names[i]);
 656                 }
 657         }
 658
 659         return snprintf(jobid, joblen, "%s", current->comm);
 660 }
 661
 662 /*
 663  * jobid_interpret_string()
 664  *
 665  * Interpret the jobfmt string to expand specified fields, like coredumps do:
 666  *   %e = executable
 667  *   %g = gid
 668  *   %h = hostname
 669  *   %H = short hostname
 670  *   %j = jobid from environment
 671  *   %p = pid
 672  *   %u = uid
 673  *
 674  * Unknown escape strings are dropped.  Other characters are copied through,
 675  * excluding whitespace (to avoid making jobid parsing difficult).
 676  *
 677  * Return: -EOVERFLOW if the expanded string does not fit within @joblen
 678  *         0 for success
 679  */
 680 static int jobid_interpret_string(const char *jobfmt, char *jobid,
 681                                   ssize_t joblen)
 682 {
 683         char c;
 684
 685         while ((c = *jobfmt++) && joblen > 1) {
 686                 char f, *p;
 687                 int l;
 688
 689                 if (isspace(c)) /* Don't allow embedded spaces */
 690                         continue;
 691
 692                 if (c != '%') {
 693                         *jobid = c;
 694                         joblen--;
 695                         jobid++;
 696                         *jobid = '\0';
 697                         continue;
 698                 }
 699
 700                 switch ((f = *jobfmt++)) {
 701                 case 'e': /* executable name */
 702                         l = jobid_print_current_comm(jobid, joblen);
 703                         break;
 704                 case 'g': /* group ID */
 705                         l = snprintf(jobid, joblen, "%u",
 706                                      from_kgid(&init_user_ns, current_fsgid()));
 707                         break;
 708                 case 'h': /* hostname */
 709                         l = snprintf(jobid, joblen, "%s",
 710                                      init_utsname()->nodename);
 711                         break;
 712                 case 'H': /* short hostname. Cut at first dot */
 713                         l = snprintf(jobid, joblen, "%s",
 714                                      init_utsname()->nodename);
 715                         p = strnchr(jobid, joblen, '.');
 716                         if (p) {
 717                                 *p = '\0';
 718                                 l = p - jobid;
 719                         }
 720                         break;
 721                 case 'j': /* jobid stored in process environment */
 722                         l = jobid_get_from_cache(jobid, joblen);
 723                         if (l < 0)
 724                                 l = 0;
 725                         break;
 726                 case 'p': /* process ID */
 727                         l = snprintf(jobid, joblen, "%u", current->pid);
 728                         break;
 729                 case 'u': /* user ID */
 730                         l = snprintf(jobid, joblen, "%u",
 731                                      from_kuid(&init_user_ns, current_fsuid()));
 732                         break;
 733                 case '\0': /* '%' at end of format string */
 734                         l = 0;
 735                         goto out;
 736                 default: /* drop unknown %x format strings */
 737                         l = 0;
 738                         break;
 739                 }
 740                 jobid += l;
 741                 joblen -= l;
 742         }
 743         /*
 744          * This points at the end of the buffer, so long as jobid is always
 745          * incremented the same amount as joblen is decremented.
 746          */
 747 out:
 748         jobid[joblen - 1] = '\0';
 749
 750         return joblen < 0 ? -EOVERFLOW : 0;
 751 }
 752
 753 /*
 754  * Hash initialization, copied from server-side job stats bucket sizes
 755  */
 756 #define HASH_JOBID_BKT_BITS 5
 757 #define HASH_JOBID_CUR_BITS 7
 758 #define HASH_JOBID_MAX_BITS 12
 759
 760 int jobid_cache_init(void)
 761 {
 762         int rc = 0;
 763         ENTRY;
 764
 765         if (jobid_hash)
 766                 return 0;
 767
 768         spin_lock_init(&jobid_hash_lock);
 769         jobid_hash = cfs_hash_create("JOBID_HASH", HASH_JOBID_CUR_BITS,
 770                                      HASH_JOBID_MAX_BITS, HASH_JOBID_BKT_BITS,
 771                                      0, CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
 772                                      &jobid_hash_ops, CFS_HASH_DEFAULT);
 773         if (!jobid_hash) {
 774                 rc = -ENOMEM;
 775         } else {
 776                 rc = rhashtable_init(&session_jobids, &jobid_params);
 777                 if (rc) {
 778                         cfs_hash_putref(jobid_hash);
 779                         jobid_hash = NULL;
 780                 }
 781         }
 782
 783         RETURN(rc);
 784 }
 785 EXPORT_SYMBOL(jobid_cache_init);
 786
 787 void jobid_cache_fini(void)
 788 {
 789         struct cfs_hash *tmp_hash;
 790         ENTRY;
 791
 792         spin_lock(&jobid_hash_lock);
 793         tmp_hash = jobid_hash;
 794         jobid_hash = NULL;
 795         spin_unlock(&jobid_hash_lock);
 796
 797         cancel_delayed_work_sync(&jobid_prune_work);
 798
 799         if (tmp_hash != NULL) {
 800                 cfs_hash_cond_del(tmp_hash, jobid_should_free_item, NULL);
 801                 cfs_hash_putref(tmp_hash);
 802
 803                 rhashtable_free_and_destroy(&session_jobids, jobid_free, NULL);
 804         }
 805
 806
 807         EXIT;
 808 }
 809 EXPORT_SYMBOL(jobid_cache_fini);
 810
 811 /*
 812  * Hash operations for pid<->jobid
 813  */
 814 static unsigned int
 815 jobid_hashfn(struct cfs_hash *hs, const void *key, const unsigned int bits)
 816 {
 817         return cfs_hash_djb2_hash(key, sizeof(pid_t), bits);
 818 }
 819
 820 static void *jobid_key(struct hlist_node *hnode)
 821 {
 822         struct jobid_pid_map *pidmap;
 823
 824         pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 825         return &pidmap->jp_pid;
 826 }
 827
 828 static int jobid_keycmp(const void *key, struct hlist_node *hnode)
 829 {
 830         const pid_t *pid_key1;
 831         const pid_t *pid_key2;
 832
 833         LASSERT(key != NULL);
 834         pid_key1 = (pid_t *)key;
 835         pid_key2 = (pid_t *)jobid_key(hnode);
 836
 837         return *pid_key1 == *pid_key2;
 838 }
 839
 840 static void *jobid_object(struct hlist_node *hnode)
 841 {
 842         return hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 843 }
 844
 845 static void jobid_get(struct cfs_hash *hs, struct hlist_node *hnode)
 846 {
 847         struct jobid_pid_map *pidmap;
 848
 849         pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 850
 851         kref_get(&pidmap->jp_refcount);
 852 }
 853
 854 static void jobid_put_locked_free(struct kref *kref)
 855 {
 856         struct jobid_pid_map *pidmap = container_of(kref, struct jobid_pid_map,
 857                                                     jp_refcount);
 858
 859         CDEBUG(D_INFO, "Freeing: %d->%s\n", pidmap->jp_pid, pidmap->jp_jobid);
 860         OBD_FREE_PTR(pidmap);
 861 }
 862
 863 static void jobid_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 864 {
 865         struct jobid_pid_map *pidmap;
 866
 867         if (hnode == NULL)
 868                 return;
 869
 870         pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
 871         LASSERT(kref_read(&pidmap->jp_refcount) > 0);
 872         kref_put(&pidmap->jp_refcount, jobid_put_locked_free);
 873 }
 874
 875 static struct cfs_hash_ops jobid_hash_ops = {
 876         .hs_hash        = jobid_hashfn,
 877         .hs_keycmp      = jobid_keycmp,
 878         .hs_key         = jobid_key,
 879         .hs_object      = jobid_object,
 880         .hs_get         = jobid_get,
 881         .hs_put         = jobid_put_locked,
 882         .hs_put_locked  = jobid_put_locked,
 883 };
 884
 885 /**
 886  * Generate the job identifier string for this process for tracking purposes.
 887  *
 888  * Fill in @jobid string based on the value of obd_jobid_var:
 889  * JOBSTATS_DISABLE:      none
 890  * JOBSTATS_NODELOCAL:    content of obd_jobid_name (jobid_interpret_string())
 891  * JOBSTATS_PROCNAME_UID: process name/UID
 892  * JOBSTATS_SESSION       per-session value set by
 893  *                            /sys/fs/lustre/jobid_this_session
 894  * anything else:         look up obd_jobid_var in the processes environment
 895  *
 896  * Return -ve error number, 0 on success.
 897  */
 898 int lustre_get_jobid(char *jobid, size_t joblen)
 899 {
 900         char id[LUSTRE_JOBID_SIZE] = "";
 901         int len = min_t(int, joblen, LUSTRE_JOBID_SIZE);
 902         int rc = 0;
 903         ENTRY;
 904
 905         if (unlikely(joblen < 2)) {
 906                 if (joblen == 1)
 907                         jobid[0] = '\0';
 908                 RETURN(-EINVAL);
 909         }
 910
 911         if (strcmp(obd_jobid_var, JOBSTATS_DISABLE) == 0) {
 912                 /* Jobstats isn't enabled */
 913                 memset(jobid, 0, joblen);
 914                 RETURN(0);
 915         }
 916
 917         if (strcmp(obd_jobid_var, JOBSTATS_NODELOCAL) == 0) {
 918                 /* Whole node dedicated to single job */
 919                 rc = jobid_interpret_string(obd_jobid_name, id, len);
 920         } else if (strcmp(obd_jobid_var, JOBSTATS_PROCNAME_UID) == 0) {
 921                 rc = jobid_interpret_string("%e.%u", id, len);
 922         } else if (strcmp(obd_jobid_var, JOBSTATS_SESSION) == 0 ||
 923                    jobid_name_is_valid(current->comm)) {
 924                 /*
 925                  * per-process jobid wanted, either from environment or from
 926                  * per-session setting.
 927                  * If obd_jobid_name contains "%j" or if getting the per-process
 928                  * jobid directly fails, fall back to using obd_jobid_name.
 929                  */
 930                 rc = -EAGAIN;
 931                 if (!strnstr(obd_jobid_name, "%j", joblen))
 932                         rc = jobid_get_from_cache(id, len);
 933
 934                 /* fall back to jobid_name if jobid_var not available */
 935                 if (rc < 0) {
 936                         int rc2 = jobid_interpret_string(obd_jobid_name,
 937                                                          id, len);
 938                         if (!rc2)
 939                                 rc = 0;
 940                 }
 941         }
 942
 943         memcpy(jobid, id, len);
 944         RETURN(rc);
 945 }
 946 EXPORT_SYMBOL(lustre_get_jobid);
 947
 948 /*
 949  * lustre_jobid_clear
 950  *
 951  * Search cache for JobID given by @find_jobid.
 952  * If any entries in the hash table match the value, they are removed
 953  */
 954 void lustre_jobid_clear(const char *find_jobid)
 955 {
 956         char jobid[LUSTRE_JOBID_SIZE];
 957         char *end;
 958
 959         if (jobid_hash == NULL)
 960                 return;
 961
 962         strscpy(jobid, find_jobid, sizeof(jobid));
 963         /* trim \n off the end of the incoming jobid */
 964         end = strchr(jobid, '\n');
 965         if (end && *end == '\n')
 966                 *end = '\0';
 967
 968         CDEBUG(D_INFO, "Clearing Jobid: %s\n", jobid);
 969         cfs_hash_cond_del(jobid_hash, jobid_should_free_item, jobid);
 970
 971         CDEBUG(D_INFO, "%d items remain in jobID table\n",
 972                atomic_read(&jobid_hash->hs_count));
 973 }