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.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Copyright (c) 2012, 2013, 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
34 #include <obd_class.h>
35 #include <lprocfs_status.h>
36 #include <lustre/lustre_idl.h>
41 * JobID formats & JobID environment variable names for supported
45 * JobID format: 32 bit integer.
46 * JobID env var: SLURM_JOB_ID.
48 * JobID format: Decimal integer range to 99999.
49 * JobID env var: JOB_ID.
51 * JobID format: 6 digit integer by default (up to 999999), can be
52 * increased to 10 digit (up to 2147483646).
53 * JobID env var: LSB_JOBID.
55 * JobID format: String of machine_name.cluster_id.process_id, for
56 * example: fr2n02.32.0
57 * JobID env var: LOADL_STEP_ID.
59 * JobID format: String of sequence_number[.server_name][@server].
60 * JobID env var: PBS_JOBID.
62 * JobID format: Same as PBS.
63 * JobID env var: Same as PBS.
67 struct hlist_node js_hash;
68 struct list_head js_list;
70 char js_jobid[JOBSTATS_JOBID_SIZE];
71 time_t js_timestamp; /* seconds */
72 struct lprocfs_stats *js_stats;
73 struct obd_job_stats *js_jobstats;
76 static unsigned job_stat_hash(cfs_hash_t *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(cfs_hash_t *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);
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_free_stats(&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(cfs_hash_t *hs, struct hlist_node *hnode)
130 struct job_stat *job;
131 job = hlist_entry(hnode, struct job_stat, js_hash);
135 static void job_stat_exit(cfs_hash_t *hs, struct hlist_node *hnode)
137 CERROR("should not have any items\n");
140 static cfs_hash_ops_t job_stats_hash_ops = {
141 .hs_hash = job_stat_hash,
142 .hs_key = job_stat_key,
143 .hs_keycmp = job_stat_keycmp,
144 .hs_object = job_stat_object,
145 .hs_get = job_stat_get,
146 .hs_put_locked = job_stat_put_locked,
147 .hs_exit = job_stat_exit,
150 static int job_iter_callback(cfs_hash_t *hs, cfs_hash_bd_t *bd,
151 struct hlist_node *hnode, void *data)
153 time_t oldest = *((time_t *)data);
154 struct job_stat *job;
156 job = hlist_entry(hnode, struct job_stat, js_hash);
157 if (!oldest || job->js_timestamp < oldest)
158 cfs_hash_bd_del_locked(hs, bd, hnode);
163 static void lprocfs_job_cleanup(struct obd_job_stats *stats, bool force)
167 if (stats->ojs_cleanup_interval == 0)
170 now = cfs_time_current_sec();
171 if (!force && now < stats->ojs_last_cleanup +
172 stats->ojs_cleanup_interval)
175 oldest = now - stats->ojs_cleanup_interval;
176 cfs_hash_for_each_safe(stats->ojs_hash, job_iter_callback,
178 stats->ojs_last_cleanup = cfs_time_current_sec();
181 static struct job_stat *job_alloc(char *jobid, struct obd_job_stats *jobs)
183 struct job_stat *job;
185 LASSERT(jobs->ojs_cntr_num && jobs->ojs_cntr_init_fn);
191 job->js_stats = lprocfs_alloc_stats(jobs->ojs_cntr_num, 0);
192 if (job->js_stats == NULL) {
197 jobs->ojs_cntr_init_fn(job->js_stats);
199 memcpy(job->js_jobid, jobid, JOBSTATS_JOBID_SIZE);
200 job->js_timestamp = cfs_time_current_sec();
201 job->js_jobstats = jobs;
202 INIT_HLIST_NODE(&job->js_hash);
203 INIT_LIST_HEAD(&job->js_list);
204 atomic_set(&job->js_refcount, 1);
209 int lprocfs_job_stats_log(struct obd_device *obd, char *jobid,
210 int event, long amount)
212 struct obd_job_stats *stats = &obd->u.obt.obt_jobstats;
213 struct job_stat *job, *job2;
216 LASSERT(stats && stats->ojs_hash);
218 lprocfs_job_cleanup(stats, false);
220 if (!jobid || !strlen(jobid))
223 if (strlen(jobid) >= JOBSTATS_JOBID_SIZE) {
224 CERROR("Invalid jobid size (%lu), expect(%d)\n",
225 (unsigned long)strlen(jobid) + 1, JOBSTATS_JOBID_SIZE);
229 job = cfs_hash_lookup(stats->ojs_hash, jobid);
233 job = job_alloc(jobid, stats);
237 job2 = cfs_hash_findadd_unique(stats->ojs_hash, job->js_jobid,
242 /* We cannot LASSERT(!list_empty(&job->js_list)) here,
243 * since we just lost the race for inserting "job" into the
244 * ojs_list, and some other thread is doing it _right_now_.
245 * Instead, be content the other thread is doing this, since
246 * "job2" was initialized in job_alloc() already. LU-2163 */
248 LASSERT(list_empty(&job->js_list));
249 write_lock(&stats->ojs_lock);
250 list_add_tail(&job->js_list, &stats->ojs_list);
251 write_unlock(&stats->ojs_lock);
255 LASSERT(stats == job->js_jobstats);
256 LASSERT(stats->ojs_cntr_num > event);
257 job->js_timestamp = cfs_time_current_sec();
258 lprocfs_counter_add(job->js_stats, event, amount);
263 EXPORT_SYMBOL(lprocfs_job_stats_log);
265 void lprocfs_job_stats_fini(struct obd_device *obd)
267 struct obd_job_stats *stats = &obd->u.obt.obt_jobstats;
270 if (stats->ojs_hash == NULL)
272 cfs_hash_for_each_safe(stats->ojs_hash, job_iter_callback, &oldest);
273 cfs_hash_putref(stats->ojs_hash);
274 stats->ojs_hash = NULL;
275 LASSERT(list_empty(&stats->ojs_list));
277 EXPORT_SYMBOL(lprocfs_job_stats_fini);
279 static void *lprocfs_jobstats_seq_start(struct seq_file *p, loff_t *pos)
281 struct obd_job_stats *stats = p->private;
283 struct job_stat *job;
285 read_lock(&stats->ojs_lock);
287 return SEQ_START_TOKEN;
289 list_for_each_entry(job, &stats->ojs_list, js_list) {
296 static void lprocfs_jobstats_seq_stop(struct seq_file *p, void *v)
298 struct obd_job_stats *stats = p->private;
300 read_unlock(&stats->ojs_lock);
303 static void *lprocfs_jobstats_seq_next(struct seq_file *p, void *v, loff_t *pos)
305 struct obd_job_stats *stats = p->private;
306 struct job_stat *job;
307 struct list_head *next;
310 if (v == SEQ_START_TOKEN) {
311 next = stats->ojs_list.next;
313 job = (struct job_stat *)v;
314 next = job->js_list.next;
317 return next == &stats->ojs_list ? NULL :
318 list_entry(next, struct job_stat, js_list);
322 * Example of output on MDT:
325 * - job_id: test_id.222.25844
326 * snapshot_time: 1322494486
327 * open: { samples: 3, unit: reqs }
328 * close: { samples: 3, unit: reqs }
329 * mknod: { samples: 0, unit: reqs }
330 * link: { samples: 0, unit: reqs }
331 * unlink: { samples: 0, unit: reqs }
332 * mkdir: { samples: 0, unit: reqs }
333 * rmdir: { samples: 0, unit: reqs }
334 * rename: { samples: 1, unit: reqs }
335 * getattr: { samples: 7, unit: reqs }
336 * setattr: { samples: 0, unit: reqs }
337 * getxattr: { samples: 0, unit: reqs }
338 * setxattr: { samples: 0, unit: reqs }
339 * statfs: { samples: 0, unit: reqs }
340 * sync: { samples: 0, unit: reqs }
342 * Example of output on OST:
346 * snapshot_time: 1322494602
347 * read: { samples: 0, unit: bytes, min: 0, max: 0, sum: 0 }
348 * write: { samples: 1, unit: bytes, min: 10, max: 10, sum: 10 }
349 * setattr: { samples: 0, unit: reqs }
350 * punch: { samples: 0, unit: reqs }
351 * sync: { samples: 0, unit: reqs }
354 static const char spaces[] = " ";
356 static int inline width(const char *str, int len)
358 return len - min((int)strlen(str), 15);
361 static int lprocfs_jobstats_seq_show(struct seq_file *p, void *v)
363 struct job_stat *job = v;
364 struct lprocfs_stats *s;
365 struct lprocfs_counter ret;
366 struct lprocfs_counter_header *cntr_header;
369 if (v == SEQ_START_TOKEN) {
370 seq_printf(p, "job_stats:\n");
374 seq_printf(p, "- %-16s %s\n", "job_id:", job->js_jobid);
375 seq_printf(p, " %-16s %ld\n", "snapshot_time:", job->js_timestamp);
378 for (i = 0; i < s->ls_num; i++) {
379 cntr_header = &s->ls_cnt_header[i];
380 lprocfs_stats_collect(s, i, &ret);
382 seq_printf(p, " %s:%.*s { samples: %11"LPF64"u",
383 cntr_header->lc_name,
384 width(cntr_header->lc_name, 15), spaces,
386 if (cntr_header->lc_units[0] != '\0')
387 seq_printf(p, ", unit: %5s", cntr_header->lc_units);
389 if (cntr_header->lc_config & LPROCFS_CNTR_AVGMINMAX) {
390 seq_printf(p, ", min:%8"LPF64"u, max:%8"LPF64"u,"
392 ret.lc_count ? ret.lc_min : 0,
393 ret.lc_count ? ret.lc_max : 0,
394 ret.lc_count ? ret.lc_sum : 0);
396 if (cntr_header->lc_config & LPROCFS_CNTR_STDDEV) {
397 seq_printf(p, ", sumsq: %18"LPF64"u",
398 ret.lc_count ? ret.lc_sumsquare : 0);
401 seq_printf(p, " }\n");
407 struct seq_operations lprocfs_jobstats_seq_sops = {
408 start: lprocfs_jobstats_seq_start,
409 stop: lprocfs_jobstats_seq_stop,
410 next: lprocfs_jobstats_seq_next,
411 show: lprocfs_jobstats_seq_show,
414 static int lprocfs_jobstats_seq_open(struct inode *inode, struct file *file)
416 struct seq_file *seq;
419 if (LPROCFS_ENTRY_CHECK(PDE(inode)))
422 rc = seq_open(file, &lprocfs_jobstats_seq_sops);
425 seq = file->private_data;
426 seq->private = PDE_DATA(inode);
430 static ssize_t lprocfs_jobstats_seq_write(struct file *file,
431 const char __user *buf,
432 size_t len, loff_t *off)
434 struct seq_file *seq = file->private_data;
435 struct obd_job_stats *stats = seq->private;
436 char jobid[JOBSTATS_JOBID_SIZE];
438 struct job_stat *job;
440 if (len == 0 || len >= JOBSTATS_JOBID_SIZE)
443 if (copy_from_user(jobid, buf, len))
447 /* Trim '\n' if any */
448 if (jobid[len - 1] == '\n')
451 if (strcmp(jobid, "clear") == 0)
454 LASSERT(stats->ojs_hash);
457 cfs_hash_for_each_safe(stats->ojs_hash, job_iter_callback,
465 job = cfs_hash_lookup(stats->ojs_hash, jobid);
469 cfs_hash_del_key(stats->ojs_hash, jobid);
475 struct file_operations lprocfs_jobstats_seq_fops = {
476 .owner = THIS_MODULE,
477 .open = lprocfs_jobstats_seq_open,
479 .write = lprocfs_jobstats_seq_write,
481 .release = lprocfs_seq_release,
484 int lprocfs_job_stats_init(struct obd_device *obd, int cntr_num,
485 cntr_init_callback init_fn)
487 struct proc_dir_entry *entry;
488 struct obd_job_stats *stats;
491 LASSERT(obd->obd_proc_entry != NULL);
492 LASSERT(obd->obd_type->typ_name);
494 if (strcmp(obd->obd_type->typ_name, LUSTRE_MDT_NAME) &&
495 strcmp(obd->obd_type->typ_name, LUSTRE_OST_NAME)) {
496 CERROR("Invalid obd device type.\n");
499 stats = &obd->u.obt.obt_jobstats;
501 LASSERT(stats->ojs_hash == NULL);
502 stats->ojs_hash = cfs_hash_create("JOB_STATS",
503 HASH_JOB_STATS_CUR_BITS,
504 HASH_JOB_STATS_MAX_BITS,
505 HASH_JOB_STATS_BKT_BITS, 0,
510 if (stats->ojs_hash == NULL)
513 INIT_LIST_HEAD(&stats->ojs_list);
514 rwlock_init(&stats->ojs_lock);
515 stats->ojs_cntr_num = cntr_num;
516 stats->ojs_cntr_init_fn = init_fn;
517 stats->ojs_cleanup_interval = 600; /* 10 mins by default */
518 stats->ojs_last_cleanup = cfs_time_current_sec();
520 LPROCFS_WRITE_ENTRY();
521 entry = proc_create_data("job_stats", 0644, obd->obd_proc_entry,
522 &lprocfs_jobstats_seq_fops, stats);
523 LPROCFS_WRITE_EXIT();
525 lprocfs_job_stats_fini(obd);
530 EXPORT_SYMBOL(lprocfs_job_stats_init);
532 #ifndef HAVE_ONLY_PROCFS_SEQ
533 int lprocfs_rd_job_interval(char *page, char **start, off_t off,
534 int count, int *eof, void *data)
536 struct obd_device *obd = (struct obd_device *)data;
537 struct obd_job_stats *stats;
539 LASSERT(obd != NULL);
540 stats = &obd->u.obt.obt_jobstats;
542 return snprintf(page, count, "%d\n", stats->ojs_cleanup_interval);
544 EXPORT_SYMBOL(lprocfs_rd_job_interval);
546 int lprocfs_wr_job_interval(struct file *file, const char __user *buffer,
547 unsigned long count, void *data)
549 struct obd_device *obd = (struct obd_device *)data;
550 struct obd_job_stats *stats;
553 LASSERT(obd != NULL);
554 stats = &obd->u.obt.obt_jobstats;
556 rc = lprocfs_write_helper(buffer, count, &val);
560 stats->ojs_cleanup_interval = val;
561 lprocfs_job_cleanup(stats, true);
566 EXPORT_SYMBOL(lprocfs_wr_job_interval);
568 int lprocfs_job_interval_seq_show(struct seq_file *m, void *data)
570 struct obd_device *obd = m->private;
571 struct obd_job_stats *stats;
573 LASSERT(obd != NULL);
574 stats = &obd->u.obt.obt_jobstats;
575 return seq_printf(m, "%d\n", stats->ojs_cleanup_interval);
577 EXPORT_SYMBOL(lprocfs_job_interval_seq_show);
580 lprocfs_job_interval_seq_write(struct file *file, const char *buffer,
581 size_t count, loff_t *off)
583 struct obd_device *obd = ((struct seq_file *)file->private_data)->private;
584 struct obd_job_stats *stats;
587 LASSERT(obd != NULL);
588 stats = &obd->u.obt.obt_jobstats;
590 rc = lprocfs_write_helper(buffer, count, &val);
594 stats->ojs_cleanup_interval = val;
595 lprocfs_job_cleanup(stats, true);
598 EXPORT_SYMBOL(lprocfs_job_interval_seq_write);