4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2014, Intel Corporation.
28 * Copyright 2017 Cray Inc, all rights reserved.
31 * Store PID->JobID mappings
34 #define DEBUG_SUBSYSTEM S_RPC
35 #include <linux/user_namespace.h>
36 #include <linux/uidgid.h>
37 #include <linux/utsname.h>
39 #include <libcfs/libcfs.h>
40 #include <obd_support.h>
41 #include <obd_class.h>
42 #include <lustre_net.h>
44 static struct cfs_hash *jobid_hash;
45 static struct cfs_hash_ops jobid_hash_ops;
46 spinlock_t jobid_hash_lock;
48 #define RESCAN_INTERVAL 30
49 #define DELETE_INTERVAL 300
51 char obd_jobid_var[JOBSTATS_JOBID_VAR_MAX_LEN + 1] = JOBSTATS_DISABLE;
52 char obd_jobid_name[LUSTRE_JOBID_SIZE] = "%e.%u";
55 * Structure to store a single PID->JobID mapping
57 struct jobid_pid_map {
58 struct hlist_node jp_hash;
60 spinlock_t jp_lock; /* protects jp_jobid */
61 char jp_jobid[LUSTRE_JOBID_SIZE];
62 unsigned int jp_joblen;
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,
76 #define JOBID_EXPEDITED_CLEAN (5)
77 #define JOBID_BACKGROUND_CLEAN (5 * 60)
79 struct session_jobid {
80 struct pid *sj_session;
81 struct rhash_head sj_linkage;
82 struct rcu_head sj_rcu;
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),
92 static struct rhashtable session_jobids;
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.
99 char *jobid_current(void)
101 struct pid *sid = task_session(current);
102 struct session_jobid *sj;
104 sj = rhashtable_lookup_fast(&session_jobids, &sid, jobid_params);
110 static void jobid_prune_expedite(void);
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
116 int jobid_set_current(char *jobid)
119 struct session_jobid *sj, *origsj;
121 int len = strlen(jobid);
123 sj = kmalloc(sizeof(*sj) + len, GFP_KERNEL);
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,
133 if (origsj == NULL) {
134 /* successful insert */
136 jobid_prune_expedite();
140 if (IS_ERR(origsj)) {
141 put_pid(sj->sj_session);
144 return PTR_ERR(origsj);
146 ret = rhashtable_replace_fast(&session_jobids,
151 put_pid(sj->sj_session);
156 put_pid(origsj->sj_session);
158 kfree_rcu(origsj, sj_rcu);
159 jobid_prune_expedite();
164 static void jobid_free(void *vsj, void *arg)
166 struct session_jobid *sj = vsj;
168 put_pid(sj->sj_session);
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)
178 struct rhashtable_iter iter;
179 struct session_jobid *sj;
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 (!hlist_empty(&sj->sj_session->tasks[PIDTYPE_SID])) {
189 if (rhashtable_remove_fast(&session_jobids,
191 jobid_params) == 0) {
192 put_pid(sj->sj_session);
193 kfree_rcu(sj, sj_rcu);
196 rhashtable_walk_stop(&iter);
197 rhashtable_walk_exit(&iter);
199 schedule_delayed_work(&jobid_prune_work,
200 cfs_time_seconds(JOBID_BACKGROUND_CLEAN));
203 static void jobid_prune_expedite(void)
205 if (!jobid_prune_expedited) {
206 jobid_prune_expedited = 1;
207 mod_delayed_work(system_wq, &jobid_prune_work,
208 cfs_time_seconds(JOBID_EXPEDITED_CLEAN));
213 * Get jobid of current process by reading the environment variable
214 * stored in between the "env_start" & "env_end" of task struct.
216 * If some job scheduler doesn't store jobid in the "env_start/end",
217 * then an upcall could be issued here to get the jobid by utilizing
218 * the userspace tools/API. Then, the jobid must be cached.
220 int jobid_get_from_environ(char *jobid_var, char *jobid, int *jobid_len)
224 rc = cfs_get_environ(jobid_var, jobid, jobid_len);
228 if (rc == -EOVERFLOW) {
229 /* For the PBS_JOBID and LOADL_STEP_ID keys (which are
230 * variable length strings instead of just numbers), it
231 * might make sense to keep the unique parts for JobID,
232 * instead of just returning an error. That means a
233 * larger temp buffer for cfs_get_environ(), then
234 * truncating the string at some separator to fit into
235 * the specified jobid_len. Fix later if needed. */
236 static ktime_t printed;
238 if (unlikely(ktime_to_ns(printed) == 0 ||
239 ktime_after(ktime_get(),
240 ktime_add_ns(printed,
241 3600*24*NSEC_PER_SEC)))) {
242 LCONSOLE_WARN("jobid: '%s' value too large (%d)\n",
243 obd_jobid_var, *jobid_len);
244 printed = ktime_get();
249 CDEBUG_LIMIT((rc == -ENOENT || rc == -EINVAL ||
250 rc == -EDEADLK) ? D_INFO : D_ERROR,
251 "jobid: get '%s' failed: rc = %d\n",
260 * jobid_should_free_item
262 * Each item is checked to see if it should be released
263 * Removed from hash table by caller
264 * Actually freed in jobid_put_locked
266 * Returns 1 if item is to be freed, 0 if it is to be kept
269 static int jobid_should_free_item(void *obj, void *data)
272 struct jobid_pid_map *pidmap = obj;
279 WARN_ON_ONCE(atomic_read(&pidmap->jp_refcount) != 1);
283 spin_lock(&pidmap->jp_lock);
284 /* prevent newly inserted items from deleting */
285 if (jobid[0] == '\0' && atomic_read(&pidmap->jp_refcount) == 1)
287 else if (ktime_get_real_seconds() - pidmap->jp_time > DELETE_INTERVAL)
289 else if (strcmp(pidmap->jp_jobid, jobid) == 0)
291 spin_unlock(&pidmap->jp_lock);
297 * jobid_name_is_valid
299 * Checks if the jobid is a Lustre process
301 * Returns true if jobid is valid
302 * Returns false if jobid looks like it's a Lustre process
304 static bool jobid_name_is_valid(char *jobid)
306 const char *const lustre_reserved[] = { "ll_ping", "ptlrpc",
307 "ldlm", "ll_sa", NULL };
310 if (jobid[0] == '\0')
313 for (i = 0; lustre_reserved[i] != NULL; i++) {
314 if (strncmp(jobid, lustre_reserved[i],
315 strlen(lustre_reserved[i])) == 0)
322 * jobid_get_from_cache()
324 * Returns contents of jobid_var from process environment for current PID,
325 * or from the per-session jobid table.
326 * Values fetch from process environment will be cached for some time to avoid
327 * the overhead of scanning the environment.
329 * Return: -ENOMEM if allocating a new pidmap fails
330 * -ENOENT if no entry could be found
331 * +ve string length for success (something was returned in jobid)
333 static int jobid_get_from_cache(char *jobid, size_t joblen)
335 static time64_t last_expire;
336 bool expire_cache = false;
337 pid_t pid = current_pid();
338 struct jobid_pid_map *pidmap = NULL;
339 time64_t now = ktime_get_real_seconds();
343 if (strcmp(obd_jobid_var, JOBSTATS_SESSION) == 0) {
347 jid = jobid_current();
349 strlcpy(jobid, jid, joblen);
350 joblen = strlen(jobid);
358 LASSERT(jobid_hash != NULL);
360 /* scan hash periodically to remove old PID entries from cache */
361 spin_lock(&jobid_hash_lock);
362 if (unlikely(last_expire + DELETE_INTERVAL <= now)) {
366 spin_unlock(&jobid_hash_lock);
369 cfs_hash_cond_del(jobid_hash, jobid_should_free_item,
370 "intentionally_bad_jobid");
372 /* first try to find PID in the hash and use that value */
373 pidmap = cfs_hash_lookup(jobid_hash, &pid);
374 if (pidmap == NULL) {
375 struct jobid_pid_map *pidmap2;
377 OBD_ALLOC_PTR(pidmap);
379 GOTO(out, rc = -ENOMEM);
381 pidmap->jp_pid = pid;
383 pidmap->jp_jobid[0] = '\0';
384 spin_lock_init(&pidmap->jp_lock);
385 INIT_HLIST_NODE(&pidmap->jp_hash);
387 * @pidmap might be reclaimed just after it is added into
388 * hash list, init @jp_refcount as 1 to make sure memory
389 * could be not freed during access.
391 atomic_set(&pidmap->jp_refcount, 1);
394 * Add the newly created map to the hash, on key collision we
395 * lost a racing addition and must destroy our newly allocated
396 * map. The object which exists in the hash will be returned.
398 pidmap2 = cfs_hash_findadd_unique(jobid_hash, &pid,
400 if (unlikely(pidmap != pidmap2)) {
401 CDEBUG(D_INFO, "jobid: duplicate found for PID=%u\n",
403 OBD_FREE_PTR(pidmap);
409 * If pidmap is old (this is always true for new entries) refresh it.
410 * If obd_jobid_var is not found, cache empty entry and try again
411 * later, to avoid repeat lookups for PID if obd_jobid_var missing.
413 spin_lock(&pidmap->jp_lock);
414 if (pidmap->jp_time + RESCAN_INTERVAL <= now) {
415 char env_jobid[LUSTRE_JOBID_SIZE] = "";
416 int env_len = sizeof(env_jobid);
418 pidmap->jp_time = now;
420 spin_unlock(&pidmap->jp_lock);
421 rc = jobid_get_from_environ(obd_jobid_var, env_jobid, &env_len);
423 CDEBUG(D_INFO, "jobid: PID mapping established: %d->%s\n",
424 pidmap->jp_pid, env_jobid);
425 spin_lock(&pidmap->jp_lock);
427 pidmap->jp_joblen = env_len;
428 strlcpy(pidmap->jp_jobid, env_jobid,
429 sizeof(pidmap->jp_jobid));
431 } else if (rc == -ENOENT) {
432 /* It might have been deleted, clear out old entry */
433 pidmap->jp_joblen = 0;
434 pidmap->jp_jobid[0] = '\0';
439 * Regardless of how pidmap was found, if it contains a valid entry
440 * use that for now. If there was a technical error (e.g. -ENOMEM)
441 * use the old cached value until it can be looked up again properly.
442 * If a cached missing entry was found, return -ENOENT.
444 if (pidmap->jp_joblen) {
445 strlcpy(jobid, pidmap->jp_jobid, joblen);
446 joblen = pidmap->jp_joblen;
451 spin_unlock(&pidmap->jp_lock);
453 cfs_hash_put(jobid_hash, &pidmap->jp_hash);
457 return rc < 0 ? rc : joblen;
461 * jobid_interpret_string()
463 * Interpret the jobfmt string to expand specified fields, like coredumps do:
467 * %j = jobid from environment
471 * Unknown escape strings are dropped. Other characters are copied through,
472 * excluding whitespace (to avoid making jobid parsing difficult).
474 * Return: -EOVERFLOW if the expanded string does not fit within @joblen
477 static int jobid_interpret_string(const char *jobfmt, char *jobid,
482 while ((c = *jobfmt++) && joblen > 1) {
486 if (isspace(c)) /* Don't allow embedded spaces */
496 switch ((f = *jobfmt++)) {
497 case 'e': /* executable name */
498 l = snprintf(jobid, joblen, "%s", current_comm());
500 case 'g': /* group ID */
501 l = snprintf(jobid, joblen, "%u",
502 from_kgid(&init_user_ns, current_fsgid()));
504 case 'h': /* hostname */
505 l = snprintf(jobid, joblen, "%s",
506 init_utsname()->nodename);
508 case 'j': /* jobid stored in process environment */
509 l = jobid_get_from_cache(jobid, joblen);
513 case 'p': /* process ID */
514 l = snprintf(jobid, joblen, "%u", current_pid());
516 case 'u': /* user ID */
517 l = snprintf(jobid, joblen, "%u",
518 from_kuid(&init_user_ns, current_fsuid()));
520 case '\0': /* '%' at end of format string */
523 default: /* drop unknown %x format strings */
531 * This points at the end of the buffer, so long as jobid is always
532 * incremented the same amount as joblen is decremented.
535 jobid[joblen - 1] = '\0';
537 return joblen < 0 ? -EOVERFLOW : 0;
541 * Hash initialization, copied from server-side job stats bucket sizes
543 #define HASH_JOBID_BKT_BITS 5
544 #define HASH_JOBID_CUR_BITS 7
545 #define HASH_JOBID_MAX_BITS 12
547 int jobid_cache_init(void)
555 spin_lock_init(&jobid_hash_lock);
556 jobid_hash = cfs_hash_create("JOBID_HASH", HASH_JOBID_CUR_BITS,
557 HASH_JOBID_MAX_BITS, HASH_JOBID_BKT_BITS,
558 0, CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
559 &jobid_hash_ops, CFS_HASH_DEFAULT);
563 rc = rhashtable_init(&session_jobids, &jobid_params);
565 cfs_hash_putref(jobid_hash);
572 EXPORT_SYMBOL(jobid_cache_init);
574 void jobid_cache_fini(void)
576 struct cfs_hash *tmp_hash;
579 spin_lock(&jobid_hash_lock);
580 tmp_hash = jobid_hash;
582 spin_unlock(&jobid_hash_lock);
584 cancel_delayed_work_sync(&jobid_prune_work);
586 if (tmp_hash != NULL) {
587 cfs_hash_cond_del(tmp_hash, jobid_should_free_item, NULL);
588 cfs_hash_putref(tmp_hash);
590 rhashtable_free_and_destroy(&session_jobids, jobid_free, NULL);
596 EXPORT_SYMBOL(jobid_cache_fini);
599 * Hash operations for pid<->jobid
601 static unsigned jobid_hashfn(struct cfs_hash *hs, const void *key,
604 return cfs_hash_djb2_hash(key, sizeof(pid_t), mask);
607 static void *jobid_key(struct hlist_node *hnode)
609 struct jobid_pid_map *pidmap;
611 pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
612 return &pidmap->jp_pid;
615 static int jobid_keycmp(const void *key, struct hlist_node *hnode)
617 const pid_t *pid_key1;
618 const pid_t *pid_key2;
620 LASSERT(key != NULL);
621 pid_key1 = (pid_t *)key;
622 pid_key2 = (pid_t *)jobid_key(hnode);
624 return *pid_key1 == *pid_key2;
627 static void *jobid_object(struct hlist_node *hnode)
629 return hlist_entry(hnode, struct jobid_pid_map, jp_hash);
632 static void jobid_get(struct cfs_hash *hs, struct hlist_node *hnode)
634 struct jobid_pid_map *pidmap;
636 pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
638 atomic_inc(&pidmap->jp_refcount);
641 static void jobid_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
643 struct jobid_pid_map *pidmap;
648 pidmap = hlist_entry(hnode, struct jobid_pid_map, jp_hash);
649 LASSERT(atomic_read(&pidmap->jp_refcount) > 0);
650 if (atomic_dec_and_test(&pidmap->jp_refcount)) {
651 CDEBUG(D_INFO, "Freeing: %d->%s\n",
652 pidmap->jp_pid, pidmap->jp_jobid);
654 OBD_FREE_PTR(pidmap);
658 static struct cfs_hash_ops jobid_hash_ops = {
659 .hs_hash = jobid_hashfn,
660 .hs_keycmp = jobid_keycmp,
662 .hs_object = jobid_object,
664 .hs_put = jobid_put_locked,
665 .hs_put_locked = jobid_put_locked,
669 * Generate the job identifier string for this process for tracking purposes.
671 * Fill in @jobid string based on the value of obd_jobid_var:
672 * JOBSTATS_DISABLE: none
673 * JOBSTATS_NODELOCAL: content of obd_jobid_name (jobid_interpret_string())
674 * JOBSTATS_PROCNAME_UID: process name/UID
675 * JOBSTATS_SESSION per-session value set by
676 * /sys/fs/lustre/jobid_this_session
677 * anything else: look up obd_jobid_var in the processes environment
679 * Return -ve error number, 0 on success.
681 int lustre_get_jobid(char *jobid, size_t joblen)
686 if (unlikely(joblen < 2)) {
692 if (strcmp(obd_jobid_var, JOBSTATS_DISABLE) == 0) {
693 /* Jobstats isn't enabled */
694 memset(jobid, 0, joblen);
695 } else if (strcmp(obd_jobid_var, JOBSTATS_NODELOCAL) == 0) {
696 /* Whole node dedicated to single job */
697 rc = jobid_interpret_string(obd_jobid_name, jobid, joblen);
698 } else if (strcmp(obd_jobid_var, JOBSTATS_PROCNAME_UID) == 0) {
699 rc = jobid_interpret_string("%e.%u", jobid, joblen);
700 } else if (strcmp(obd_jobid_var, JOBSTATS_SESSION) == 0) {
704 jid = jobid_current();
706 strlcpy(jobid, jid, sizeof(jobid));
708 } else if (jobid_name_is_valid(current_comm())) {
710 * obd_jobid_var holds the jobid environment variable name.
711 * Skip initial check if obd_jobid_name already uses "%j",
712 * otherwise try just "%j" first, then fall back to whatever
713 * is in obd_jobid_name if obd_jobid_var is not found.
716 if (!strnstr(obd_jobid_name, "%j", joblen))
717 rc = jobid_get_from_cache(jobid, joblen);
719 /* fall back to jobid_node if jobid_var not in environment */
721 int rc2 = jobid_interpret_string(obd_jobid_name,
730 EXPORT_SYMBOL(lustre_get_jobid);
735 * Search cache for JobID given by @find_jobid.
736 * If any entries in the hash table match the value, they are removed
738 void lustre_jobid_clear(const char *find_jobid)
740 char jobid[LUSTRE_JOBID_SIZE];
743 if (jobid_hash == NULL)
746 strlcpy(jobid, find_jobid, sizeof(jobid));
747 /* trim \n off the end of the incoming jobid */
748 end = strchr(jobid, '\n');
749 if (end && *end == '\n')
752 CDEBUG(D_INFO, "Clearing Jobid: %s\n", jobid);
753 cfs_hash_cond_del(jobid_hash, jobid_should_free_item, jobid);
755 CDEBUG(D_INFO, "%d items remain in jobID table\n",
756 atomic_read(&jobid_hash->hs_count));