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) 2011 Whamcloud, Inc.
23 * Use is subject to license terms.
25 * Author: Niu Yawei <niu@whamcloud.com>
28 * lustre/obdclass/lprocfs_jobstats.c
32 # define EXPORT_SYMTAB
34 #define DEBUG_SUBSYSTEM S_CLASS
37 # include <liblustre.h>
40 #include <obd_class.h>
41 #include <lprocfs_status.h>
42 #include <lustre/lustre_idl.h>
47 * JobID formats & JobID environment variable names for supported
51 * JobID format: 32 bit integer.
52 * JobID env var: SLURM_JOB_ID.
54 * JobID format: Decimal integer range to 99999.
55 * JobID env var: JOB_ID.
57 * JobID format: 6 digit integer by default (up to 999999), can be
58 * increased to 10 digit (up to 2147483646).
59 * JobID env var: LSB_JOBID.
61 * JobID format: String of machine_name.cluster_id.process_id, for
62 * example: fr2n02.32.0
63 * JobID env var: LOADL_STEP_ID.
65 * JobID format: String of sequence_number[.server_name][@server].
66 * JobID env var: PBS_JOBID.
68 * JobID format: Same as PBS.
69 * JobID env var: Same as PBS.
73 cfs_hlist_node_t js_hash;
75 cfs_atomic_t js_refcount;
76 char js_jobid[JOBSTATS_JOBID_SIZE];
77 time_t js_timestamp; /* seconds */
78 struct lprocfs_stats *js_stats;
79 struct obd_job_stats *js_jobstats;
82 static unsigned job_stat_hash(cfs_hash_t *hs, const void *key, unsigned mask)
84 return cfs_hash_djb2_hash(key, strlen(key), mask);
87 static void *job_stat_key(cfs_hlist_node_t *hnode)
90 job = cfs_hlist_entry(hnode, struct job_stat, js_hash);
94 static int job_stat_keycmp(const void *key, cfs_hlist_node_t *hnode)
97 job = cfs_hlist_entry(hnode, struct job_stat, js_hash);
98 return (strlen(job->js_jobid) == strlen(key)) &&
99 !strncmp(job->js_jobid, key, strlen(key));
102 static void *job_stat_object(cfs_hlist_node_t *hnode)
104 return cfs_hlist_entry(hnode, struct job_stat, js_hash);
107 static void job_stat_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
109 struct job_stat *job;
110 job = cfs_hlist_entry(hnode, struct job_stat, js_hash);
111 cfs_atomic_inc(&job->js_refcount);
114 static void job_free(struct job_stat *job)
116 LASSERT(atomic_read(&job->js_refcount) == 0);
117 LASSERT(job->js_jobstats);
119 write_lock(&job->js_jobstats->ojs_lock);
120 cfs_list_del_init(&job->js_list);
121 write_unlock(&job->js_jobstats->ojs_lock);
123 lprocfs_free_stats(&job->js_stats);
127 static void job_putref(struct job_stat *job)
129 LASSERT(atomic_read(&job->js_refcount) > 0);
130 if (atomic_dec_and_test(&job->js_refcount))
134 static void job_stat_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
136 struct job_stat *job;
137 job = cfs_hlist_entry(hnode, struct job_stat, js_hash);
141 static void job_stat_exit(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
143 CERROR("Should not have any items!");
146 static cfs_hash_ops_t job_stats_hash_ops = {
147 .hs_hash = job_stat_hash,
148 .hs_key = job_stat_key,
149 .hs_keycmp = job_stat_keycmp,
150 .hs_object = job_stat_object,
151 .hs_get = job_stat_get,
152 .hs_put_locked = job_stat_put_locked,
153 .hs_exit = job_stat_exit,
156 static int job_iter_callback(cfs_hash_t *hs, cfs_hash_bd_t *bd,
157 cfs_hlist_node_t *hnode, void *data)
159 time_t oldest = *((time_t *)data);
160 struct job_stat *job;
162 job = cfs_hlist_entry(hnode, struct job_stat, js_hash);
163 if (!oldest || job->js_timestamp < oldest)
164 cfs_hash_bd_del_locked(hs, bd, hnode);
169 static void lprocfs_job_cleanup(struct obd_job_stats *stats, bool force)
173 if (stats->ojs_cleanup_interval == 0)
176 now = cfs_time_current_sec();
177 if (!force && now < stats->ojs_last_cleanup +
178 stats->ojs_cleanup_interval)
181 oldest = now - stats->ojs_cleanup_interval;
182 cfs_hash_for_each_safe(stats->ojs_hash, job_iter_callback,
184 stats->ojs_last_cleanup = cfs_time_current_sec();
187 static struct job_stat *job_alloc(char *jobid, struct obd_job_stats *jobs)
189 struct job_stat *job;
191 LASSERT(jobs->ojs_cntr_num && jobs->ojs_cntr_init_fn);
197 job->js_stats = lprocfs_alloc_stats(jobs->ojs_cntr_num, 0);
198 if (job->js_stats == NULL) {
203 jobs->ojs_cntr_init_fn(job->js_stats);
205 memcpy(job->js_jobid, jobid, JOBSTATS_JOBID_SIZE);
206 job->js_timestamp = cfs_time_current_sec();
207 job->js_jobstats = jobs;
208 CFS_INIT_HLIST_NODE(&job->js_hash);
209 CFS_INIT_LIST_HEAD(&job->js_list);
210 cfs_atomic_set(&job->js_refcount, 1);
215 int lprocfs_job_stats_log(struct obd_device *obd, char *jobid,
216 int event, long amount)
218 struct obd_job_stats *stats = &obd->u.obt.obt_jobstats;
219 struct job_stat *job, *job2;
222 LASSERT(stats && stats->ojs_hash);
224 lprocfs_job_cleanup(stats, false);
226 if (!jobid || !strlen(jobid))
229 if (strlen(jobid) >= JOBSTATS_JOBID_SIZE) {
230 CERROR("Invalid jobid size (%lu), expect(%d)\n",
231 (unsigned long)strlen(jobid) + 1, JOBSTATS_JOBID_SIZE);
235 job = cfs_hash_lookup(stats->ojs_hash, jobid);
239 job = job_alloc(jobid, stats);
243 job2 = cfs_hash_findadd_unique(stats->ojs_hash, job->js_jobid,
248 /* We cannot LASSERT(!cfs_list_empty(&job->js_list)) here,
249 * since we just lost the race for inserting "job" into the
250 * ojs_list, and some other thread is doing it _right_now_.
251 * Instead, be content the other thread is doing this, since
252 * "job2" was initialized in job_alloc() already. LU-2163 */
254 LASSERT(cfs_list_empty(&job->js_list));
255 write_lock(&stats->ojs_lock);
256 cfs_list_add_tail(&job->js_list, &stats->ojs_list);
257 write_unlock(&stats->ojs_lock);
261 LASSERT(stats == job->js_jobstats);
262 LASSERT(stats->ojs_cntr_num > event);
263 job->js_timestamp = cfs_time_current_sec();
264 lprocfs_counter_add(job->js_stats, event, amount);
269 EXPORT_SYMBOL(lprocfs_job_stats_log);
271 void lprocfs_job_stats_fini(struct obd_device *obd)
273 struct obd_job_stats *stats = &obd->u.obt.obt_jobstats;
276 if (stats->ojs_hash == NULL)
278 cfs_hash_for_each_safe(stats->ojs_hash, job_iter_callback, &oldest);
279 cfs_hash_putref(stats->ojs_hash);
280 stats->ojs_hash = NULL;
281 LASSERT(cfs_list_empty(&stats->ojs_list));
283 EXPORT_SYMBOL(lprocfs_job_stats_fini);
285 static void *lprocfs_jobstats_seq_start(struct seq_file *p, loff_t *pos)
287 struct obd_job_stats *stats = p->private;
289 struct job_stat *job;
291 read_lock(&stats->ojs_lock);
293 return SEQ_START_TOKEN;
295 cfs_list_for_each_entry(job, &stats->ojs_list, js_list) {
302 static void lprocfs_jobstats_seq_stop(struct seq_file *p, void *v)
304 struct obd_job_stats *stats = p->private;
306 read_unlock(&stats->ojs_lock);
309 static void *lprocfs_jobstats_seq_next(struct seq_file *p, void *v, loff_t *pos)
311 struct obd_job_stats *stats = p->private;
312 struct job_stat *job;
316 if (v == SEQ_START_TOKEN) {
317 next = stats->ojs_list.next;
319 job = (struct job_stat *)v;
320 next = job->js_list.next;
323 return next == &stats->ojs_list ? NULL :
324 cfs_list_entry(next, struct job_stat, js_list);
328 * Example of output on MDT:
331 * - job_id: test_id.222.25844
332 * snapshot_time: 1322494486
333 * open: { samples: 3, unit: reqs }
334 * close: { samples: 3, unit: reqs }
335 * mknod: { samples: 0, unit: reqs }
336 * link: { samples: 0, unit: reqs }
337 * unlink: { samples: 0, unit: reqs }
338 * mkdir: { samples: 0, unit: reqs }
339 * rmdir: { samples: 0, unit: reqs }
340 * rename: { samples: 1, unit: reqs }
341 * getattr: { samples: 7, unit: reqs }
342 * setattr: { samples: 0, unit: reqs }
343 * getxattr: { samples: 0, unit: reqs }
344 * setxattr: { samples: 0, unit: reqs }
345 * statfs: { samples: 0, unit: reqs }
346 * sync: { samples: 0, unit: reqs }
348 * Example of output on OST:
352 * snapshot_time: 1322494602
353 * read: { samples: 0, unit: bytes, min: 0, max: 0, sum: 0 }
354 * write: { samples: 1, unit: bytes, min: 10, max: 10, sum: 10 }
355 * setattr: { samples: 0, unit: reqs }
356 * punch: { samples: 0, unit: reqs }
357 * sync: { samples: 0, unit: reqs }
360 static const char spaces[] = " ";
362 static int inline width(const char *str, int len)
364 return len - min((int)strlen(str), 15);
367 static int lprocfs_jobstats_seq_show(struct seq_file *p, void *v)
369 struct job_stat *job = v;
370 struct lprocfs_stats *s;
371 struct lprocfs_counter ret, *cntr;
374 if (v == SEQ_START_TOKEN) {
375 seq_printf(p, "job_stats:\n");
379 seq_printf(p, "- %-16s %s\n", "job_id:", job->js_jobid);
380 seq_printf(p, " %-16s %ld\n", "snapshot_time:", job->js_timestamp);
383 for (i = 0; i < s->ls_num; i++) {
384 cntr = &(s->ls_percpu[0]->lp_cntr[i]);
385 lprocfs_stats_collect(s, i, &ret);
387 seq_printf(p, " %s:%.*s { samples: %11"LPF64"u",
388 cntr->lc_name, width(cntr->lc_name, 15), spaces,
390 if (cntr->lc_units[0] != '\0')
391 seq_printf(p, ", unit: %5s", cntr->lc_units);
393 if (cntr->lc_config & LPROCFS_CNTR_AVGMINMAX) {
394 seq_printf(p, ", min:%8"LPF64"u, max:%8"LPF64"u,"
396 ret.lc_count ? ret.lc_min : 0,
397 ret.lc_count ? ret.lc_max : 0,
398 ret.lc_count ? ret.lc_sum : 0);
400 if (cntr->lc_config & LPROCFS_CNTR_STDDEV) {
401 seq_printf(p, ", sumsq: %18"LPF64"u",
402 ret.lc_count ? ret.lc_sumsquare : 0);
405 seq_printf(p, " }\n");
411 struct seq_operations lprocfs_jobstats_seq_sops = {
412 start: lprocfs_jobstats_seq_start,
413 stop: lprocfs_jobstats_seq_stop,
414 next: lprocfs_jobstats_seq_next,
415 show: lprocfs_jobstats_seq_show,
418 static int lprocfs_jobstats_seq_open(struct inode *inode, struct file *file)
420 struct proc_dir_entry *dp = PDE(inode);
421 struct seq_file *seq;
424 if (LPROCFS_ENTRY_AND_CHECK(dp))
427 rc = seq_open(file, &lprocfs_jobstats_seq_sops);
432 seq = file->private_data;
433 seq->private = dp->data;
437 static ssize_t lprocfs_jobstats_seq_write(struct file *file, const char *buf,
438 size_t len, loff_t *off)
440 struct seq_file *seq = file->private_data;
441 struct obd_job_stats *stats = seq->private;
442 char jobid[JOBSTATS_JOBID_SIZE];
444 struct job_stat *job;
446 if (!memcmp(buf, "clear", strlen("clear"))) {
448 } else if (len < JOBSTATS_JOBID_SIZE) {
449 memset(jobid, 0, JOBSTATS_JOBID_SIZE);
450 /* Trim '\n' if any */
451 if (buf[len - 1] == '\n')
452 memcpy(jobid, buf, len - 1);
454 memcpy(jobid, buf, len);
459 LASSERT(stats->ojs_hash);
462 cfs_hash_for_each_safe(stats->ojs_hash, job_iter_callback,
470 job = cfs_hash_lookup(stats->ojs_hash, jobid);
474 cfs_hash_del_key(stats->ojs_hash, jobid);
480 struct file_operations lprocfs_jobstats_seq_fops = {
481 .owner = THIS_MODULE,
482 .open = lprocfs_jobstats_seq_open,
484 .write = lprocfs_jobstats_seq_write,
486 .release = lprocfs_seq_release,
489 int lprocfs_job_stats_init(struct obd_device *obd, int cntr_num,
490 cntr_init_callback init_fn)
492 struct proc_dir_entry *entry;
493 struct obd_job_stats *stats;
496 LASSERT(obd->obd_proc_entry != NULL);
497 LASSERT(obd->obd_type->typ_name);
499 if (strcmp(obd->obd_type->typ_name, LUSTRE_MDT_NAME) &&
500 strcmp(obd->obd_type->typ_name, LUSTRE_OST_NAME)) {
501 CERROR("Invalid obd device type.\n");
504 stats = &obd->u.obt.obt_jobstats;
506 LASSERT(stats->ojs_hash == NULL);
507 stats->ojs_hash = cfs_hash_create("JOB_STATS",
508 HASH_JOB_STATS_CUR_BITS,
509 HASH_JOB_STATS_MAX_BITS,
510 HASH_JOB_STATS_BKT_BITS, 0,
515 if (stats->ojs_hash == NULL)
518 CFS_INIT_LIST_HEAD(&stats->ojs_list);
519 rwlock_init(&stats->ojs_lock);
520 stats->ojs_cntr_num = cntr_num;
521 stats->ojs_cntr_init_fn = init_fn;
522 stats->ojs_cleanup_interval = 600; /* 10 mins by default */
523 stats->ojs_last_cleanup = cfs_time_current_sec();
525 LPROCFS_WRITE_ENTRY();
526 entry = create_proc_entry("job_stats", 0644, obd->obd_proc_entry);
527 LPROCFS_WRITE_EXIT();
529 entry->proc_fops = &lprocfs_jobstats_seq_fops;
533 lprocfs_job_stats_fini(obd);
537 EXPORT_SYMBOL(lprocfs_job_stats_init);
539 int lprocfs_rd_job_interval(char *page, char **start, off_t off,
540 int count, int *eof, void *data)
542 struct obd_device *obd = (struct obd_device *)data;
543 struct obd_job_stats *stats;
545 LASSERT(obd != NULL);
546 stats = &obd->u.obt.obt_jobstats;
548 return snprintf(page, count, "%d\n", stats->ojs_cleanup_interval);
550 EXPORT_SYMBOL(lprocfs_rd_job_interval);
552 int lprocfs_wr_job_interval(struct file *file, const char *buffer,
553 unsigned long count, void *data)
555 struct obd_device *obd = (struct obd_device *)data;
556 struct obd_job_stats *stats;
559 LASSERT(obd != NULL);
560 stats = &obd->u.obt.obt_jobstats;
562 rc = lprocfs_write_helper(buffer, count, &val);
566 stats->ojs_cleanup_interval = val;
567 lprocfs_job_cleanup(stats, true);
572 EXPORT_SYMBOL(lprocfs_wr_job_interval);