3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 only,
7 * as published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful, but
10 * WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License version 2 for more details (a copy is included
13 * in the LICENSE file that accompanied this code).
15 * You should have received a copy of the GNU General Public License
16 * version 2 along with this program; If not, see
17 * http://www.gnu.org/licenses/gpl-2.0.html
22 * Copyright (c) 2012, 2016, Intel Corporation.
23 * Use is subject to license terms.
25 * Author: Niu Yawei <niu@whamcloud.com>
28 * lustre/obdclass/lprocfs_jobstats.c
31 #define DEBUG_SUBSYSTEM S_CLASS
33 #include <obd_class.h>
34 #include <lprocfs_status.h>
39 * JobID formats & JobID environment variable names for supported
43 * JobID format: 32 bit integer.
44 * JobID env var: SLURM_JOB_ID.
46 * JobID format: Decimal integer range to 99999.
47 * JobID env var: JOB_ID.
49 * JobID format: 6 digit integer by default (up to 999999), can be
50 * increased to 10 digit (up to 2147483646).
51 * JobID env var: LSB_JOBID.
53 * JobID format: String of machine_name.cluster_id.process_id, for
54 * example: fr2n02.32.0
55 * JobID env var: LOADL_STEP_ID.
57 * JobID format: String of sequence_number[.server_name][@server].
58 * JobID env var: PBS_JOBID.
60 * JobID format: Same as PBS.
61 * JobID env var: Same as PBS.
65 struct hlist_node js_hash; /* hash struct for this jobid */
66 struct list_head js_list; /* on ojs_list, with ojs_lock */
67 atomic_t js_refcount; /* num users of this struct */
68 char js_jobid[LUSTRE_JOBID_SIZE]; /* job name + NUL*/
69 ktime_t js_time_init; /* time of initial stat*/
70 ktime_t js_time_latest; /* time of most recent stat*/
71 struct lprocfs_stats *js_stats; /* per-job statistics */
72 struct obd_job_stats *js_jobstats; /* for accessing ojs_lock */
76 job_stat_hash(struct cfs_hash *hs, const void *key, unsigned mask)
78 return cfs_hash_djb2_hash(key, strlen(key), mask);
81 static void *job_stat_key(struct hlist_node *hnode)
84 job = hlist_entry(hnode, struct job_stat, js_hash);
88 static int job_stat_keycmp(const void *key, struct hlist_node *hnode)
91 job = hlist_entry(hnode, struct job_stat, js_hash);
92 return (strlen(job->js_jobid) == strlen(key)) &&
93 !strncmp(job->js_jobid, key, strlen(key));
96 static void *job_stat_object(struct hlist_node *hnode)
98 return hlist_entry(hnode, struct job_stat, js_hash);
101 static void job_stat_get(struct cfs_hash *hs, struct hlist_node *hnode)
103 struct job_stat *job;
104 job = hlist_entry(hnode, struct job_stat, js_hash);
105 atomic_inc(&job->js_refcount);
108 static void job_free(struct job_stat *job)
110 LASSERT(atomic_read(&job->js_refcount) == 0);
111 LASSERT(job->js_jobstats != NULL);
113 write_lock(&job->js_jobstats->ojs_lock);
114 list_del_init(&job->js_list);
115 write_unlock(&job->js_jobstats->ojs_lock);
117 lprocfs_stats_free(&job->js_stats);
121 static void job_putref(struct job_stat *job)
123 LASSERT(atomic_read(&job->js_refcount) > 0);
124 if (atomic_dec_and_test(&job->js_refcount))
128 static void job_stat_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
130 struct job_stat *job;
132 job = hlist_entry(hnode, struct job_stat, js_hash);
136 static void job_stat_exit(struct cfs_hash *hs, struct hlist_node *hnode)
138 CERROR("should not have any items\n");
141 static struct cfs_hash_ops job_stats_hash_ops = {
142 .hs_hash = job_stat_hash,
143 .hs_key = job_stat_key,
144 .hs_keycmp = job_stat_keycmp,
145 .hs_object = job_stat_object,
146 .hs_get = job_stat_get,
147 .hs_put_locked = job_stat_put_locked,
148 .hs_exit = job_stat_exit,
152 * Jobstats expiry iterator to clean up old jobids
154 * Called for each job_stat structure on this device, it should delete stats
155 * older than the specified \a oldest_time in seconds. If \a oldest_time is
156 * in the future then this will delete all statistics (e.g. during shutdown).
158 * \param[in] hs hash of all jobids on this device
159 * \param[in] bd hash bucket containing this jobid
160 * \param[in] hnode hash structure for this jobid
161 * \param[in] data pointer to stats expiry time in seconds
163 static int job_cleanup_iter_callback(struct cfs_hash *hs,
164 struct cfs_hash_bd *bd,
165 struct hlist_node *hnode, void *data)
167 ktime_t oldest_time = *((ktime_t *)data);
168 struct job_stat *job;
170 job = hlist_entry(hnode, struct job_stat, js_hash);
171 if (ktime_before(job->js_time_latest, oldest_time))
172 cfs_hash_bd_del_locked(hs, bd, hnode);
178 * Clean up jobstats that were updated more than \a before seconds ago.
180 * Since this function may be called frequently, do not scan all of the
181 * jobstats on each call, only twice per cleanup interval. That means stats
182 * may be on average around cleanup_interval / 4 older than the cleanup
183 * interval, but that is not considered harmful.
185 * The value stored in ojs_cleanup_interval is how often to perform a cleanup
186 * scan, and 1/2 of the maximum age of the individual statistics. This is
187 * done rather than dividing the interval by two each time, because it is
188 * much easier to do the division when the value is initially set (in seconds)
189 * rather than after it has been converted to ktime_t, and maybe a bit faster.
191 * If \a clear is true then this will force clean up all jobstats
192 * (e.g. at shutdown).
194 * If there is already another thread doing jobstats cleanup, don't try to
195 * do this again in the current thread unless this is a force cleanup.
197 * \param[in] stats stucture tracking all job stats for this device
198 * \param[in] clear clear all job stats if true
200 static void lprocfs_job_cleanup(struct obd_job_stats *stats, bool clear)
202 ktime_t cleanup_interval = stats->ojs_cleanup_interval;
203 ktime_t now = ktime_get_real();
206 if (likely(!clear)) {
207 /* ojs_cleanup_interval of zero means never clean up stats */
208 if (ktime_to_ns(cleanup_interval) == 0)
211 if (ktime_before(now, ktime_add(stats->ojs_cleanup_last,
215 if (stats->ojs_cleaning)
219 write_lock(&stats->ojs_lock);
220 if (!clear && stats->ojs_cleaning) {
221 write_unlock(&stats->ojs_lock);
225 stats->ojs_cleaning = true;
226 write_unlock(&stats->ojs_lock);
228 /* Can't hold ojs_lock over hash iteration, since it is grabbed by
229 * job_cleanup_iter_callback()
230 * ->cfs_hash_bd_del_locked()
234 * Holding ojs_lock isn't necessary for safety of the hash iteration,
235 * since locking of the hash is handled internally, but there isn't
236 * any benefit to having multiple threads doing cleanup at one time.
238 * Subtract or add twice the cleanup_interval, since it is 1/2 the
239 * maximum age. When clearing all stats, push oldest into the future.
241 cleanup_interval = ktime_add(cleanup_interval, cleanup_interval);
243 oldest = ktime_sub(now, cleanup_interval);
245 oldest = ktime_add(now, cleanup_interval);
246 cfs_hash_for_each_safe(stats->ojs_hash, job_cleanup_iter_callback,
249 write_lock(&stats->ojs_lock);
250 stats->ojs_cleaning = false;
251 stats->ojs_cleanup_last = ktime_get_real();
252 write_unlock(&stats->ojs_lock);
255 static struct job_stat *job_alloc(char *jobid, struct obd_job_stats *jobs)
257 struct job_stat *job;
263 job->js_stats = lprocfs_stats_alloc(jobs->ojs_cntr_num, 0);
264 if (job->js_stats == NULL) {
269 jobs->ojs_cntr_init_fn(job->js_stats, 0, 0);
271 memcpy(job->js_jobid, jobid, sizeof(job->js_jobid));
272 job->js_time_latest = job->js_stats->ls_init;
273 job->js_jobstats = jobs;
274 INIT_HLIST_NODE(&job->js_hash);
275 INIT_LIST_HEAD(&job->js_list);
276 atomic_set(&job->js_refcount, 1);
281 int lprocfs_job_stats_log(struct obd_device *obd, char *jobid,
282 int event, long amount)
284 struct obd_job_stats *stats = &obd2obt(obd)->obt_jobstats;
285 struct job_stat *job, *job2;
288 LASSERT(stats != NULL);
289 LASSERT(stats->ojs_hash != NULL);
291 if (event >= stats->ojs_cntr_num)
294 if (jobid == NULL || strlen(jobid) == 0)
297 /* unterminated jobid should be handled in lustre_msg_get_jobid() */
298 if (strlen(jobid) >= LUSTRE_JOBID_SIZE) {
299 CERROR("%s: invalid jobid size %lu, expect %d\n", obd->obd_name,
300 (unsigned long)strlen(jobid) + 1, LUSTRE_JOBID_SIZE);
304 job = cfs_hash_lookup(stats->ojs_hash, jobid);
308 lprocfs_job_cleanup(stats, false);
310 job = job_alloc(jobid, stats);
314 job2 = cfs_hash_findadd_unique(stats->ojs_hash, job->js_jobid,
319 /* We cannot LASSERT(!list_empty(&job->js_list)) here,
320 * since we just lost the race for inserting "job" into the
321 * ojs_list, and some other thread is doing it _right_now_.
322 * Instead, be content the other thread is doing this, since
323 * "job2" was initialized in job_alloc() already. LU-2163 */
325 LASSERT(list_empty(&job->js_list));
326 write_lock(&stats->ojs_lock);
327 list_add_tail(&job->js_list, &stats->ojs_list);
328 write_unlock(&stats->ojs_lock);
332 LASSERT(stats == job->js_jobstats);
333 job->js_time_latest = ktime_get_real();
334 lprocfs_counter_add(job->js_stats, event, amount);
340 EXPORT_SYMBOL(lprocfs_job_stats_log);
342 void lprocfs_job_stats_fini(struct obd_device *obd)
344 struct obd_job_stats *stats = &obd2obt(obd)->obt_jobstats;
346 if (stats->ojs_hash == NULL)
349 lprocfs_job_cleanup(stats, true);
350 cfs_hash_putref(stats->ojs_hash);
351 stats->ojs_hash = NULL;
352 LASSERT(list_empty(&stats->ojs_list));
354 EXPORT_SYMBOL(lprocfs_job_stats_fini);
356 static void *lprocfs_jobstats_seq_start(struct seq_file *p, loff_t *pos)
358 struct obd_job_stats *stats = p->private;
360 struct job_stat *job;
362 read_lock(&stats->ojs_lock);
364 return SEQ_START_TOKEN;
366 list_for_each_entry(job, &stats->ojs_list, js_list) {
373 static void lprocfs_jobstats_seq_stop(struct seq_file *p, void *v)
375 struct obd_job_stats *stats = p->private;
377 read_unlock(&stats->ojs_lock);
380 static void *lprocfs_jobstats_seq_next(struct seq_file *p, void *v, loff_t *pos)
382 struct obd_job_stats *stats = p->private;
383 struct job_stat *job;
384 struct list_head *next;
387 if (v == SEQ_START_TOKEN) {
388 next = stats->ojs_list.next;
390 job = (struct job_stat *)v;
391 next = job->js_list.next;
394 return next == &stats->ojs_list ? NULL :
395 list_entry(next, struct job_stat, js_list);
399 * Example of output on MDT:
403 * snapshot_time: 1322494486.123456789
404 * start_time: 1322494476.012345678
405 * elapsed_time: 10.111111111
406 * open: { samples: 1, unit: reqs }
407 * close: { samples: 1, unit: reqs }
408 * mknod: { samples: 0, unit: reqs }
409 * link: { samples: 0, unit: reqs }
410 * unlink: { samples: 0, unit: reqs }
411 * mkdir: { samples: 0, unit: reqs }
412 * rmdir: { samples: 0, unit: reqs }
413 * rename: { samples: 0, unit: reqs }
414 * getattr: { samples: 1, unit: reqs }
415 * setattr: { samples: 0, unit: reqs }
416 * getxattr: { samples: 0, unit: reqs }
417 * setxattr: { samples: 0, unit: reqs }
418 * statfs: { samples: 0, unit: reqs }
419 * sync: { samples: 0, unit: reqs }
421 * Example of output on OST:
425 * snapshot_time: 1322494602.123456789
426 * start_time: 1322494592.987654321
427 * elapsed_time: 9.135802468
428 * read: { samples: 0, unit: bytes, min: 0, max: 0, sum: 0 }
429 * write: { samples: 1, unit: bytes, min: 4096, max: 4096, sum: 4096 }
430 * setattr: { samples: 0, unit: reqs }
431 * punch: { samples: 0, unit: reqs }
432 * sync: { samples: 0, unit: reqs }
435 static const char spaces[] = " ";
437 static int inline width(const char *str, int len)
439 return len - min((int)strlen(str), 15);
442 static int lprocfs_jobstats_seq_show(struct seq_file *p, void *v)
444 struct job_stat *job = v;
445 struct lprocfs_stats *s;
446 struct lprocfs_counter ret;
447 struct lprocfs_counter_header *cntr_header;
448 char escaped[LUSTRE_JOBID_SIZE * 4] = "";
449 char *quote = "", *c, *end;
452 if (v == SEQ_START_TOKEN) {
453 seq_puts(p, "job_stats:\n");
457 /* Quote and escape jobid characters to escape hex codes "\xHH" if
458 * it contains any non-standard characters (space, newline, etc),
459 * so it will be confined to single line and not break parsing.
461 for (c = job->js_jobid, end = job->js_jobid + sizeof(job->js_jobid);
462 c < end && *c != '\0';
464 if (!isalnum(*c) && strchr(".@-_:/", *c) == NULL) {
466 snprintf(escaped + joblen, sizeof(escaped), "\\x%02X",
470 escaped[joblen] = *c;
471 /* if jobid has ':', it should be quoted too */
477 seq_printf(p, "- %-16s %s%*s%s\n",
478 "job_id:", quote, joblen, escaped, quote);
479 lprocfs_stats_header(p, job->js_time_latest, job->js_stats->ls_init,
483 for (i = 0; i < s->ls_num; i++) {
484 struct obd_histogram *hist;
486 cntr_header = &s->ls_cnt_header[i];
487 lprocfs_stats_collect(s, i, &ret);
489 seq_printf(p, " %s:%.*s { samples: %11llu",
490 cntr_header->lc_name,
491 width(cntr_header->lc_name, 15), spaces,
493 if (cntr_header->lc_units[0] != '\0')
494 seq_printf(p, ", unit: %5s", cntr_header->lc_units);
496 if (cntr_header->lc_config & LPROCFS_CNTR_AVGMINMAX) {
497 seq_printf(p, ", min: %8llu, max: %8llu, sum: %16llu",
498 ret.lc_count ? ret.lc_min : 0,
499 ret.lc_count ? ret.lc_max : 0,
500 ret.lc_count ? ret.lc_sum : 0);
502 if (cntr_header->lc_config & LPROCFS_CNTR_STDDEV) {
503 seq_printf(p, ", sumsq: %18llu",
504 ret.lc_count ? ret.lc_sumsquare : 0);
507 /* show obd_histogram */
508 hist = s->ls_cnt_header[i].lc_hist;
513 seq_puts(p, ", hist: { ");
514 for (j = 0; j < ARRAY_SIZE(hist->oh_buckets); j++) {
515 unsigned long val = hist->oh_buckets[j];
525 seq_printf(p, "%lu: %lu", BIT(j), val);
527 seq_printf(p, "%luK: %lu", BIT(j - 10),
530 seq_printf(p, "%luM: %lu", BIT(j - 20),
533 seq_printf(p, "%luG: %lu", BIT(j - 30),
544 static const struct seq_operations lprocfs_jobstats_seq_sops = {
545 .start = lprocfs_jobstats_seq_start,
546 .stop = lprocfs_jobstats_seq_stop,
547 .next = lprocfs_jobstats_seq_next,
548 .show = lprocfs_jobstats_seq_show,
551 static int lprocfs_jobstats_seq_open(struct inode *inode, struct file *file)
553 struct seq_file *seq;
556 rc = seq_open(file, &lprocfs_jobstats_seq_sops);
559 seq = file->private_data;
560 seq->private = pde_data(inode);
564 static ssize_t lprocfs_jobstats_seq_write(struct file *file,
565 const char __user *buf,
566 size_t len, loff_t *off)
568 struct seq_file *seq = file->private_data;
569 struct obd_job_stats *stats = seq->private;
570 char jobid[4 * LUSTRE_JOBID_SIZE]; /* all escaped chars, plus ""\n\0 */
571 char *p1, *p2, *last;
573 struct job_stat *job;
575 if (len == 0 || len >= 4 * LUSTRE_JOBID_SIZE)
578 if (stats->ojs_hash == NULL)
581 if (copy_from_user(jobid, buf, len))
584 last = jobid + len - 1;
586 /* Trim '\n' if any */
590 /* decode escaped chars if jobid is a quoted string */
591 if (jobid[0] == '"' && *last == '"') {
594 for (p1 = jobid, p2 = jobid + 1; p2 <= last; p1++, p2++) {
597 } else if (p2 + 3 <= last && *(p2 + 1) == 'x' &&
598 sscanf(p2 + 2, "%02X", &c) == 1) {
608 jobid[LUSTRE_JOBID_SIZE - 1] = 0;
610 if (strcmp(jobid, "clear") == 0) {
611 lprocfs_job_cleanup(stats, true);
616 if (strlen(jobid) == 0)
619 job = cfs_hash_lookup(stats->ojs_hash, jobid);
623 cfs_hash_del_key(stats->ojs_hash, jobid);
630 * Clean up the seq file state when the /proc file is closed.
632 * This also expires old job stats from the cache after they have been
633 * printed in case the system is idle and not generating new jobstats.
635 * \param[in] inode struct inode for seq file being closed
636 * \param[in] file struct file for seq file being closed
638 * \retval 0 on success
639 * \retval negative errno on failure
641 static int lprocfs_jobstats_seq_release(struct inode *inode, struct file *file)
643 struct seq_file *seq = file->private_data;
644 struct obd_job_stats *stats = seq->private;
646 lprocfs_job_cleanup(stats, false);
648 return lprocfs_seq_release(inode, file);
651 static const struct proc_ops lprocfs_jobstats_seq_fops = {
652 PROC_OWNER(THIS_MODULE)
653 .proc_open = lprocfs_jobstats_seq_open,
654 .proc_read = seq_read,
655 .proc_write = lprocfs_jobstats_seq_write,
656 .proc_lseek = seq_lseek,
657 .proc_release = lprocfs_jobstats_seq_release,
660 int lprocfs_job_stats_init(struct obd_device *obd, int cntr_num,
661 cntr_init_callback init_fn)
663 struct proc_dir_entry *entry;
664 struct obd_job_stats *stats;
667 LASSERT(obd->obd_proc_entry != NULL);
668 LASSERT(obd->obd_type->typ_name);
676 /* Currently needs to be a target due to the use of obt_jobstats. */
677 if (strcmp(obd->obd_type->typ_name, LUSTRE_MDT_NAME) != 0 &&
678 strcmp(obd->obd_type->typ_name, LUSTRE_OST_NAME) != 0) {
679 CERROR("%s: invalid device type %s for job stats: rc = %d\n",
680 obd->obd_name, obd->obd_type->typ_name, -EINVAL);
683 stats = &obd2obt(obd)->obt_jobstats;
685 LASSERT(stats->ojs_hash == NULL);
686 stats->ojs_hash = cfs_hash_create("JOB_STATS",
687 HASH_JOB_STATS_CUR_BITS,
688 HASH_JOB_STATS_MAX_BITS,
689 HASH_JOB_STATS_BKT_BITS, 0,
694 if (stats->ojs_hash == NULL)
697 INIT_LIST_HEAD(&stats->ojs_list);
698 rwlock_init(&stats->ojs_lock);
699 stats->ojs_cntr_num = cntr_num;
700 stats->ojs_cntr_init_fn = init_fn;
701 /* Store 1/2 the actual interval, since we use that the most, and
702 * it is easier to work with.
704 stats->ojs_cleanup_interval = ktime_set(600 / 2, 0); /* default 10 min*/
705 stats->ojs_cleanup_last = ktime_get_real();
707 entry = lprocfs_add_simple(obd->obd_proc_entry, "job_stats", stats,
708 &lprocfs_jobstats_seq_fops);
710 lprocfs_job_stats_fini(obd);
715 EXPORT_SYMBOL(lprocfs_job_stats_init);
716 #endif /* CONFIG_PROC_FS*/
718 ssize_t job_cleanup_interval_show(struct kobject *kobj, struct attribute *attr,
721 struct obd_device *obd = container_of(kobj, struct obd_device,
723 struct obd_job_stats *stats;
724 struct timespec64 ts;
726 stats = &obd2obt(obd)->obt_jobstats;
727 ts = ktime_to_timespec64(stats->ojs_cleanup_interval);
729 return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)ts.tv_sec * 2);
731 EXPORT_SYMBOL(job_cleanup_interval_show);
733 ssize_t job_cleanup_interval_store(struct kobject *kobj,
734 struct attribute *attr,
735 const char *buffer, size_t count)
737 struct obd_device *obd = container_of(kobj, struct obd_device,
739 struct obd_job_stats *stats;
743 stats = &obd2obt(obd)->obt_jobstats;
745 rc = kstrtouint(buffer, 0, &val);
749 stats->ojs_cleanup_interval = ktime_set(val / 2, 0);
750 lprocfs_job_cleanup(stats, false);
754 EXPORT_SYMBOL(job_cleanup_interval_store);