1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98#define DEBUG_SUBSYSTEM S_LDLM
99
100# include <lustre_dlm.h>
101
102#include <cl_object.h>
103
104#include <obd_class.h>
105#include <obd_support.h>
106#include "ldlm_internal.h"
107
108
109
110
111
112#define LDLM_POOL_HOST_L ((NUM_CACHEPAGES >> (20 - PAGE_CACHE_SHIFT)) * 50)
113
114
115
116
117#define LDLM_POOL_MAX_GSP (30)
118
119
120
121
122#define LDLM_POOL_MIN_GSP (1)
123
124
125
126
127
128#define LDLM_POOL_GSP_STEP_SHIFT (2)
129
130
131
132
133#define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100)
134
135
136
137
138#define LDLM_POOL_MAX_AGE (36000)
139
140
141
142
143#define LDLM_POOL_SLV_SHIFT (10)
144
145extern proc_dir_entry_t *ldlm_ns_proc_dir;
146
147static inline __u64 dru(__u64 val, __u32 shift, int round_up)
148{
149 return (val + (round_up ? (1 << shift) - 1 : 0)) >> shift;
150}
151
152static inline __u64 ldlm_pool_slv_max(__u32 L)
153{
154
155
156
157
158 __u64 lim = (__u64)L * LDLM_POOL_MAX_AGE / 1;
159 return lim;
160}
161
162static inline __u64 ldlm_pool_slv_min(__u32 L)
163{
164 return 1;
165}
166
167enum {
168 LDLM_POOL_FIRST_STAT = 0,
169 LDLM_POOL_GRANTED_STAT = LDLM_POOL_FIRST_STAT,
170 LDLM_POOL_GRANT_STAT,
171 LDLM_POOL_CANCEL_STAT,
172 LDLM_POOL_GRANT_RATE_STAT,
173 LDLM_POOL_CANCEL_RATE_STAT,
174 LDLM_POOL_GRANT_PLAN_STAT,
175 LDLM_POOL_SLV_STAT,
176 LDLM_POOL_SHRINK_REQTD_STAT,
177 LDLM_POOL_SHRINK_FREED_STAT,
178 LDLM_POOL_RECALC_STAT,
179 LDLM_POOL_TIMING_STAT,
180 LDLM_POOL_LAST_STAT
181};
182
183static inline struct ldlm_namespace *ldlm_pl2ns(struct ldlm_pool *pl)
184{
185 return container_of(pl, struct ldlm_namespace, ns_pool);
186}
187
188
189
190
191
192static inline int ldlm_pool_t2gsp(unsigned int t)
193{
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214 return LDLM_POOL_MAX_GSP -
215 ((LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) >>
216 (t >> LDLM_POOL_GSP_STEP_SHIFT));
217}
218
219
220
221
222
223
224static void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl)
225{
226 int granted, grant_step, limit;
227
228 limit = ldlm_pool_get_limit(pl);
229 granted = atomic_read(&pl->pl_granted);
230
231 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
232 grant_step = ((limit - granted) * grant_step) / 100;
233 pl->pl_grant_plan = granted + grant_step;
234 limit = (limit * 5) >> 2;
235 if (pl->pl_grant_plan > limit)
236 pl->pl_grant_plan = limit;
237}
238
239
240
241
242
243
244static void ldlm_pool_recalc_slv(struct ldlm_pool *pl)
245{
246 int granted;
247 int grant_plan;
248 int round_up;
249 __u64 slv;
250 __u64 slv_factor;
251 __u64 grant_usage;
252 __u32 limit;
253
254 slv = pl->pl_server_lock_volume;
255 grant_plan = pl->pl_grant_plan;
256 limit = ldlm_pool_get_limit(pl);
257 granted = atomic_read(&pl->pl_granted);
258 round_up = granted < limit;
259
260 grant_usage = max_t(int, limit - (granted - grant_plan), 1);
261
262
263
264
265
266
267
268
269
270 slv_factor = (grant_usage << LDLM_POOL_SLV_SHIFT);
271 do_div(slv_factor, limit);
272 slv = slv * slv_factor;
273 slv = dru(slv, LDLM_POOL_SLV_SHIFT, round_up);
274
275 if (slv > ldlm_pool_slv_max(limit)) {
276 slv = ldlm_pool_slv_max(limit);
277 } else if (slv < ldlm_pool_slv_min(limit)) {
278 slv = ldlm_pool_slv_min(limit);
279 }
280
281 pl->pl_server_lock_volume = slv;
282}
283
284
285
286
287
288
289static void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
290{
291 int grant_plan = pl->pl_grant_plan;
292 __u64 slv = pl->pl_server_lock_volume;
293 int granted = atomic_read(&pl->pl_granted);
294 int grant_rate = atomic_read(&pl->pl_grant_rate);
295 int cancel_rate = atomic_read(&pl->pl_cancel_rate);
296
297 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT,
298 slv);
299 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
300 granted);
301 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
302 grant_rate);
303 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
304 grant_plan);
305 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
306 cancel_rate);
307}
308
309
310
311
312static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl)
313{
314 struct obd_device *obd;
315
316
317
318
319
320
321
322
323 obd = ldlm_pl2ns(pl)->ns_obd;
324 LASSERT(obd != NULL);
325 write_lock(&obd->obd_pool_lock);
326 obd->obd_pool_slv = pl->pl_server_lock_volume;
327 write_unlock(&obd->obd_pool_lock);
328}
329
330
331
332
333
334
335static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
336{
337 time_t recalc_interval_sec;
338 ENTRY;
339
340 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
341 if (recalc_interval_sec < pl->pl_recalc_period)
342 RETURN(0);
343
344 spin_lock(&pl->pl_lock);
345 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
346 if (recalc_interval_sec < pl->pl_recalc_period) {
347 spin_unlock(&pl->pl_lock);
348 RETURN(0);
349 }
350
351
352
353
354 ldlm_pool_recalc_slv(pl);
355
356
357
358
359 ldlm_srv_pool_push_slv(pl);
360
361
362
363
364 ldlm_pool_recalc_grant_plan(pl);
365
366 pl->pl_recalc_time = cfs_time_current_sec();
367 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
368 recalc_interval_sec);
369 spin_unlock(&pl->pl_lock);
370 RETURN(0);
371}
372
373
374
375
376
377
378
379
380
381static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
382 int nr, unsigned int gfp_mask)
383{
384 __u32 limit;
385
386
387
388
389 if (nr == 0)
390 return atomic_read(&pl->pl_granted);
391
392
393
394
395
396 if (atomic_read(&pl->pl_granted) == 0)
397 RETURN(0);
398
399 spin_lock(&pl->pl_lock);
400
401
402
403
404
405
406
407
408
409
410
411
412
413 if (nr < pl->pl_server_lock_volume) {
414 pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr;
415 } else {
416 limit = ldlm_pool_get_limit(pl);
417 pl->pl_server_lock_volume = ldlm_pool_slv_min(limit);
418 }
419
420
421
422
423 ldlm_srv_pool_push_slv(pl);
424 spin_unlock(&pl->pl_lock);
425
426
427
428
429
430 return 0;
431}
432
433
434
435
436static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit)
437{
438 struct obd_device *obd;
439
440 obd = ldlm_pl2ns(pl)->ns_obd;
441 LASSERT(obd != NULL && obd != LP_POISON);
442 LASSERT(obd->obd_type != LP_POISON);
443 write_lock(&obd->obd_pool_lock);
444 obd->obd_pool_limit = limit;
445 write_unlock(&obd->obd_pool_lock);
446
447 ldlm_pool_set_limit(pl, limit);
448 return 0;
449}
450
451
452
453
454static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
455{
456 struct obd_device *obd;
457
458
459
460
461
462 obd = ldlm_pl2ns(pl)->ns_obd;
463 LASSERT(obd != NULL);
464 read_lock(&obd->obd_pool_lock);
465 pl->pl_server_lock_volume = obd->obd_pool_slv;
466 ldlm_pool_set_limit(pl, obd->obd_pool_limit);
467 read_unlock(&obd->obd_pool_lock);
468}
469
470
471
472
473static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
474{
475 time_t recalc_interval_sec;
476 ENTRY;
477
478 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
479 if (recalc_interval_sec < pl->pl_recalc_period)
480 RETURN(0);
481
482 spin_lock(&pl->pl_lock);
483
484
485
486 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
487 if (recalc_interval_sec < pl->pl_recalc_period) {
488 spin_unlock(&pl->pl_lock);
489 RETURN(0);
490 }
491
492
493
494
495 ldlm_cli_pool_pop_slv(pl);
496
497 pl->pl_recalc_time = cfs_time_current_sec();
498 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
499 recalc_interval_sec);
500 spin_unlock(&pl->pl_lock);
501
502
503
504
505 if (!ns_connect_lru_resize(ldlm_pl2ns(pl)))
506 RETURN(0);
507
508
509
510
511
512
513
514 RETURN(ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC,
515 LDLM_CANCEL_LRUR));
516}
517
518
519
520
521
522
523static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
524 int nr, unsigned int gfp_mask)
525{
526 struct ldlm_namespace *ns;
527 int canceled = 0, unused;
528
529 ns = ldlm_pl2ns(pl);
530
531
532
533
534 if (!ns_connect_lru_resize(ns))
535 RETURN(0);
536
537
538
539
540 ldlm_cli_pool_pop_slv(pl);
541
542 spin_lock(&ns->ns_lock);
543 unused = ns->ns_nr_unused;
544 spin_unlock(&ns->ns_lock);
545
546 if (nr) {
547 canceled = ldlm_cancel_lru(ns, nr, LCF_ASYNC,
548 LDLM_CANCEL_SHRINK);
549 }
550
551
552
553 return ((unused - canceled) / 100) * sysctl_vfs_cache_pressure;
554}
555
556struct ldlm_pool_ops ldlm_srv_pool_ops = {
557 .po_recalc = ldlm_srv_pool_recalc,
558 .po_shrink = ldlm_srv_pool_shrink,
559 .po_setup = ldlm_srv_pool_setup
560};
561
562struct ldlm_pool_ops ldlm_cli_pool_ops = {
563 .po_recalc = ldlm_cli_pool_recalc,
564 .po_shrink = ldlm_cli_pool_shrink
565};
566
567
568
569
570
571int ldlm_pool_recalc(struct ldlm_pool *pl)
572{
573 time_t recalc_interval_sec;
574 int count;
575
576 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
577 if (recalc_interval_sec <= 0)
578 goto recalc;
579
580 spin_lock(&pl->pl_lock);
581 recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
582 if (recalc_interval_sec > 0) {
583
584
585
586 ldlm_pool_recalc_stats(pl);
587
588
589
590
591 atomic_set(&pl->pl_grant_rate, 0);
592 atomic_set(&pl->pl_cancel_rate, 0);
593 }
594 spin_unlock(&pl->pl_lock);
595
596 recalc:
597 if (pl->pl_ops->po_recalc != NULL) {
598 count = pl->pl_ops->po_recalc(pl);
599 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
600 count);
601 return count;
602 }
603
604 return 0;
605}
606EXPORT_SYMBOL(ldlm_pool_recalc);
607
608
609
610
611
612int ldlm_pool_shrink(struct ldlm_pool *pl, int nr,
613 unsigned int gfp_mask)
614{
615 int cancel = 0;
616
617 if (pl->pl_ops->po_shrink != NULL) {
618 cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask);
619 if (nr > 0) {
620 lprocfs_counter_add(pl->pl_stats,
621 LDLM_POOL_SHRINK_REQTD_STAT,
622 nr);
623 lprocfs_counter_add(pl->pl_stats,
624 LDLM_POOL_SHRINK_FREED_STAT,
625 cancel);
626 CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, "
627 "shrunk %d\n", pl->pl_name, nr, cancel);
628 }
629 }
630 return cancel;
631}
632EXPORT_SYMBOL(ldlm_pool_shrink);
633
634
635
636
637
638
639
640int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
641{
642 if (pl->pl_ops->po_setup != NULL)
643 return(pl->pl_ops->po_setup(pl, limit));
644 return 0;
645}
646EXPORT_SYMBOL(ldlm_pool_setup);
647
648static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
649{
650 int granted, grant_rate, cancel_rate, grant_step;
651 int grant_speed, grant_plan, lvf;
652 struct ldlm_pool *pl = m->private;
653 __u64 slv, clv;
654 __u32 limit;
655
656 spin_lock(&pl->pl_lock);
657 slv = pl->pl_server_lock_volume;
658 clv = pl->pl_client_lock_volume;
659 limit = ldlm_pool_get_limit(pl);
660 grant_plan = pl->pl_grant_plan;
661 granted = atomic_read(&pl->pl_granted);
662 grant_rate = atomic_read(&pl->pl_grant_rate);
663 cancel_rate = atomic_read(&pl->pl_cancel_rate);
664 grant_speed = grant_rate - cancel_rate;
665 lvf = atomic_read(&pl->pl_lock_volume_factor);
666 grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
667 spin_unlock(&pl->pl_lock);
668
669 seq_printf(m, "LDLM pool state (%s):\n"
670 " SLV: "LPU64"\n"
671 " CLV: "LPU64"\n"
672 " LVF: %d\n",
673 pl->pl_name, slv, clv, lvf);
674
675 if (ns_is_server(ldlm_pl2ns(pl))) {
676 seq_printf(m, " GSP: %d%%\n"
677 " GP: %d\n",
678 grant_step, grant_plan);
679 }
680 seq_printf(m, " GR: %d\n" " CR: %d\n" " GS: %d\n"
681 " G: %d\n" " L: %d\n",
682 grant_rate, cancel_rate, grant_speed,
683 granted, limit);
684
685 return 0;
686}
687LPROC_SEQ_FOPS_RO(lprocfs_pool_state);
688
689static int lprocfs_grant_speed_seq_show(struct seq_file *m, void *unused)
690{
691 struct ldlm_pool *pl = m->private;
692 int grant_speed;
693
694 spin_lock(&pl->pl_lock);
695
696 grant_speed = atomic_read(&pl->pl_grant_rate) -
697 atomic_read(&pl->pl_cancel_rate);
698 spin_unlock(&pl->pl_lock);
699 return lprocfs_rd_uint(m, &grant_speed);
700}
701
702LDLM_POOL_PROC_READER_SEQ_SHOW(grant_plan, int);
703LPROC_SEQ_FOPS_RO(lprocfs_grant_plan);
704
705LDLM_POOL_PROC_READER_SEQ_SHOW(recalc_period, int);
706LDLM_POOL_PROC_WRITER(recalc_period, int);
707static ssize_t lprocfs_recalc_period_seq_write(struct file *file, const char *buf,
708 size_t len, loff_t *off)
709{
710 struct seq_file *seq = file->private_data;
711
712 return lprocfs_wr_recalc_period(file, buf, len, seq->private);
713}
714LPROC_SEQ_FOPS(lprocfs_recalc_period);
715
716LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, u64);
717LPROC_SEQ_FOPS_RO_TYPE(ldlm_pool, atomic);
718LPROC_SEQ_FOPS_RW_TYPE(ldlm_pool_rw, atomic);
719
720LPROC_SEQ_FOPS_RO(lprocfs_grant_speed);
721
722#define LDLM_POOL_ADD_VAR(name, var, ops) \
723 do { \
724 snprintf(var_name, MAX_STRING_SIZE, #name); \
725 pool_vars[0].data = var; \
726 pool_vars[0].fops = ops; \
727 lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);\
728 } while (0)
729
730static int ldlm_pool_proc_init(struct ldlm_pool *pl)
731{
732 struct ldlm_namespace *ns = ldlm_pl2ns(pl);
733 struct proc_dir_entry *parent_ns_proc;
734 struct lprocfs_vars pool_vars[2];
735 char *var_name = NULL;
736 int rc = 0;
737 ENTRY;
738
739 OBD_ALLOC(var_name, MAX_STRING_SIZE + 1);
740 if (!var_name)
741 RETURN(-ENOMEM);
742
743 parent_ns_proc = ns->ns_proc_dir_entry;
744 if (parent_ns_proc == NULL) {
745 CERROR("%s: proc entry is not initialized\n",
746 ldlm_ns_name(ns));
747 GOTO(out_free_name, rc = -EINVAL);
748 }
749 pl->pl_proc_dir = lprocfs_register("pool", parent_ns_proc,
750 NULL, NULL);
751 if (IS_ERR(pl->pl_proc_dir)) {
752 CERROR("LProcFS failed in ldlm-pool-init\n");
753 rc = PTR_ERR(pl->pl_proc_dir);
754 GOTO(out_free_name, rc);
755 }
756
757 var_name[MAX_STRING_SIZE] = '\0';
758 memset(pool_vars, 0, sizeof(pool_vars));
759 pool_vars[0].name = var_name;
760
761 LDLM_POOL_ADD_VAR("server_lock_volume", &pl->pl_server_lock_volume,
762 &ldlm_pool_u64_fops);
763 LDLM_POOL_ADD_VAR("limit", &pl->pl_limit, &ldlm_pool_rw_atomic_fops);
764 LDLM_POOL_ADD_VAR("granted", &pl->pl_granted, &ldlm_pool_atomic_fops);
765 LDLM_POOL_ADD_VAR("grant_speed", pl, &lprocfs_grant_speed_fops);
766 LDLM_POOL_ADD_VAR("cancel_rate", &pl->pl_cancel_rate,
767 &ldlm_pool_atomic_fops);
768 LDLM_POOL_ADD_VAR("grant_rate", &pl->pl_grant_rate,
769 &ldlm_pool_atomic_fops);
770 LDLM_POOL_ADD_VAR("grant_plan", pl, &lprocfs_grant_plan_fops);
771 LDLM_POOL_ADD_VAR("recalc_period", pl, &lprocfs_recalc_period_fops);
772 LDLM_POOL_ADD_VAR("lock_volume_factor", &pl->pl_lock_volume_factor,
773 &ldlm_pool_rw_atomic_fops);
774 LDLM_POOL_ADD_VAR("state", pl, &lprocfs_pool_state_fops);
775
776 pl->pl_stats = lprocfs_alloc_stats(LDLM_POOL_LAST_STAT -
777 LDLM_POOL_FIRST_STAT, 0);
778 if (!pl->pl_stats)
779 GOTO(out_free_name, rc = -ENOMEM);
780
781 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT,
782 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
783 "granted", "locks");
784 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT,
785 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
786 "grant", "locks");
787 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT,
788 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
789 "cancel", "locks");
790 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT,
791 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
792 "grant_rate", "locks/s");
793 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_RATE_STAT,
794 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
795 "cancel_rate", "locks/s");
796 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_PLAN_STAT,
797 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
798 "grant_plan", "locks/s");
799 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SLV_STAT,
800 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
801 "slv", "slv");
802 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT,
803 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
804 "shrink_request", "locks");
805 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT,
806 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
807 "shrink_freed", "locks");
808 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_RECALC_STAT,
809 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
810 "recalc_freed", "locks");
811 lprocfs_counter_init(pl->pl_stats, LDLM_POOL_TIMING_STAT,
812 LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV,
813 "recalc_timing", "sec");
814 rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
815
816 EXIT;
817out_free_name:
818 OBD_FREE(var_name, MAX_STRING_SIZE + 1);
819 return rc;
820}
821
822static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
823{
824 if (pl->pl_stats != NULL) {
825 lprocfs_free_stats(&pl->pl_stats);
826 pl->pl_stats = NULL;
827 }
828 if (pl->pl_proc_dir != NULL) {
829 lprocfs_remove(&pl->pl_proc_dir);
830 pl->pl_proc_dir = NULL;
831 }
832}
833
834int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
835 int idx, ldlm_side_t client)
836{
837 int rc;
838 ENTRY;
839
840 spin_lock_init(&pl->pl_lock);
841 atomic_set(&pl->pl_granted, 0);
842 pl->pl_recalc_time = cfs_time_current_sec();
843 atomic_set(&pl->pl_lock_volume_factor, 1);
844
845 atomic_set(&pl->pl_grant_rate, 0);
846 atomic_set(&pl->pl_cancel_rate, 0);
847 pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
848
849 snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
850 ldlm_ns_name(ns), idx);
851
852 if (client == LDLM_NAMESPACE_SERVER) {
853 pl->pl_ops = &ldlm_srv_pool_ops;
854 ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L);
855 pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD;
856 pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L);
857 } else {
858 ldlm_pool_set_limit(pl, 1);
859 pl->pl_server_lock_volume = 0;
860 pl->pl_ops = &ldlm_cli_pool_ops;
861 pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
862 }
863 pl->pl_client_lock_volume = 0;
864 rc = ldlm_pool_proc_init(pl);
865 if (rc)
866 RETURN(rc);
867
868 CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name);
869
870 RETURN(rc);
871}
872EXPORT_SYMBOL(ldlm_pool_init);
873
874void ldlm_pool_fini(struct ldlm_pool *pl)
875{
876 ENTRY;
877 ldlm_pool_proc_fini(pl);
878
879
880
881
882
883
884 POISON(pl, 0x5a, sizeof(*pl));
885 EXIT;
886}
887EXPORT_SYMBOL(ldlm_pool_fini);
888
889
890
891
892void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
893{
894
895
896
897
898
899
900 if (lock->l_resource->lr_type == LDLM_FLOCK)
901 return;
902
903 atomic_inc(&pl->pl_granted);
904 atomic_inc(&pl->pl_grant_rate);
905 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT);
906
907
908
909
910
911
912 if (ns_is_server(ldlm_pl2ns(pl)))
913 ldlm_pool_recalc(pl);
914}
915EXPORT_SYMBOL(ldlm_pool_add);
916
917
918
919
920void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
921{
922
923
924
925 if (lock->l_resource->lr_type == LDLM_FLOCK)
926 return;
927
928 LASSERT(atomic_read(&pl->pl_granted) > 0);
929 atomic_dec(&pl->pl_granted);
930 atomic_inc(&pl->pl_cancel_rate);
931
932 lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT);
933
934 if (ns_is_server(ldlm_pl2ns(pl)))
935 ldlm_pool_recalc(pl);
936}
937EXPORT_SYMBOL(ldlm_pool_del);
938
939
940
941
942
943
944__u64 ldlm_pool_get_slv(struct ldlm_pool *pl)
945{
946 __u64 slv;
947 spin_lock(&pl->pl_lock);
948 slv = pl->pl_server_lock_volume;
949 spin_unlock(&pl->pl_lock);
950 return slv;
951}
952EXPORT_SYMBOL(ldlm_pool_get_slv);
953
954
955
956
957
958
959void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv)
960{
961 spin_lock(&pl->pl_lock);
962 pl->pl_server_lock_volume = slv;
963 spin_unlock(&pl->pl_lock);
964}
965EXPORT_SYMBOL(ldlm_pool_set_slv);
966
967
968
969
970
971
972__u64 ldlm_pool_get_clv(struct ldlm_pool *pl)
973{
974 __u64 slv;
975 spin_lock(&pl->pl_lock);
976 slv = pl->pl_client_lock_volume;
977 spin_unlock(&pl->pl_lock);
978 return slv;
979}
980EXPORT_SYMBOL(ldlm_pool_get_clv);
981
982
983
984
985
986
987void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv)
988{
989 spin_lock(&pl->pl_lock);
990 pl->pl_client_lock_volume = clv;
991 spin_unlock(&pl->pl_lock);
992}
993EXPORT_SYMBOL(ldlm_pool_set_clv);
994
995
996
997
998__u32 ldlm_pool_get_limit(struct ldlm_pool *pl)
999{
1000 return atomic_read(&pl->pl_limit);
1001}
1002EXPORT_SYMBOL(ldlm_pool_get_limit);
1003
1004
1005
1006
1007void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit)
1008{
1009 atomic_set(&pl->pl_limit, limit);
1010}
1011EXPORT_SYMBOL(ldlm_pool_set_limit);
1012
1013
1014
1015
1016__u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
1017{
1018 return atomic_read(&pl->pl_lock_volume_factor);
1019}
1020EXPORT_SYMBOL(ldlm_pool_get_lvf);
1021
1022static int ldlm_pool_granted(struct ldlm_pool *pl)
1023{
1024 return atomic_read(&pl->pl_granted);
1025}
1026
1027static struct ptlrpc_thread *ldlm_pools_thread;
1028static struct shrinker *ldlm_pools_srv_shrinker;
1029static struct shrinker *ldlm_pools_cli_shrinker;
1030static struct completion ldlm_pools_comp;
1031
1032
1033
1034
1035
1036
1037static int ldlm_pools_shrink(ldlm_side_t client, int nr,
1038 unsigned int gfp_mask)
1039{
1040 int total = 0, cached = 0, nr_ns;
1041 struct ldlm_namespace *ns;
1042 void *cookie;
1043
1044 if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
1045 !(gfp_mask & __GFP_FS))
1046 return -1;
1047
1048 CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n",
1049 nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
1050
1051 cookie = cl_env_reenter();
1052
1053
1054
1055
1056 for (nr_ns = atomic_read(ldlm_namespace_nr(client));
1057 nr_ns > 0; nr_ns--)
1058 {
1059 mutex_lock(ldlm_namespace_lock(client));
1060 if (list_empty(ldlm_namespace_list(client))) {
1061 mutex_unlock(ldlm_namespace_lock(client));
1062 cl_env_reexit(cookie);
1063 return 0;
1064 }
1065 ns = ldlm_namespace_first_locked(client);
1066 ldlm_namespace_get(ns);
1067 ldlm_namespace_move_locked(ns, client);
1068 mutex_unlock(ldlm_namespace_lock(client));
1069 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
1070 ldlm_namespace_put(ns);
1071 }
1072
1073 if (nr == 0 || total == 0) {
1074 cl_env_reexit(cookie);
1075 return total;
1076 }
1077
1078
1079
1080
1081 for (nr_ns = atomic_read(ldlm_namespace_nr(client));
1082 nr_ns > 0; nr_ns--)
1083 {
1084 int cancel, nr_locks;
1085
1086
1087
1088
1089 mutex_lock(ldlm_namespace_lock(client));
1090 if (list_empty(ldlm_namespace_list(client))) {
1091 mutex_unlock(ldlm_namespace_lock(client));
1092
1093
1094
1095
1096
1097 cached = 0;
1098 break;
1099 }
1100 ns = ldlm_namespace_first_locked(client);
1101 ldlm_namespace_get(ns);
1102 ldlm_namespace_move_locked(ns, client);
1103 mutex_unlock(ldlm_namespace_lock(client));
1104
1105 nr_locks = ldlm_pool_granted(&ns->ns_pool);
1106 cancel = 1 + nr_locks * nr / total;
1107 ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
1108 cached += ldlm_pool_granted(&ns->ns_pool);
1109 ldlm_namespace_put(ns);
1110 }
1111 cl_env_reexit(cookie);
1112
1113
1114 return (client == LDLM_NAMESPACE_SERVER) ? -1 : cached;
1115}
1116
1117static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1118{
1119 return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER,
1120 shrink_param(sc, nr_to_scan),
1121 shrink_param(sc, gfp_mask));
1122}
1123
1124static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1125{
1126 return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT,
1127 shrink_param(sc, nr_to_scan),
1128 shrink_param(sc, gfp_mask));
1129}
1130
1131void ldlm_pools_recalc(ldlm_side_t client)
1132{
1133 __u32 nr_l = 0, nr_p = 0, l;
1134 struct ldlm_namespace *ns;
1135 int nr, equal = 0;
1136
1137
1138
1139
1140 if (client == LDLM_NAMESPACE_SERVER) {
1141
1142
1143
1144 mutex_lock(ldlm_namespace_lock(client));
1145 list_for_each_entry(ns, ldlm_namespace_list(client),
1146 ns_list_chain)
1147 {
1148 if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
1149 continue;
1150
1151 l = ldlm_pool_granted(&ns->ns_pool);
1152 if (l == 0)
1153 l = 1;
1154
1155
1156
1157
1158
1159 l += dru(l, LDLM_POOLS_MODEST_MARGIN_SHIFT, 0);
1160 ldlm_pool_setup(&ns->ns_pool, l);
1161 nr_l += l;
1162 nr_p++;
1163 }
1164
1165
1166
1167
1168
1169 if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) {
1170 CWARN("\"Modest\" pools eat out 2/3 of server locks "
1171 "limit (%d of %lu). This means that you have too "
1172 "many clients for this amount of server RAM. "
1173 "Upgrade server!\n", nr_l, LDLM_POOL_HOST_L);
1174 equal = 1;
1175 }
1176
1177
1178
1179
1180 list_for_each_entry(ns, ldlm_namespace_list(client),
1181 ns_list_chain)
1182 {
1183 if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
1184 continue;
1185
1186 if (equal) {
1187
1188
1189
1190
1191
1192 l = LDLM_POOL_HOST_L /
1193 atomic_read(
1194 ldlm_namespace_nr(client));
1195 } else {
1196
1197
1198
1199
1200 l = (LDLM_POOL_HOST_L - nr_l) /
1201 (atomic_read(
1202 ldlm_namespace_nr(client)) -
1203 nr_p);
1204 }
1205 ldlm_pool_setup(&ns->ns_pool, l);
1206 }
1207 mutex_unlock(ldlm_namespace_lock(client));
1208 }
1209
1210
1211
1212
1213 for (nr = atomic_read(ldlm_namespace_nr(client)); nr > 0; nr--) {
1214 int skip;
1215
1216
1217
1218
1219
1220
1221
1222 mutex_lock(ldlm_namespace_lock(client));
1223 if (list_empty(ldlm_namespace_list(client))) {
1224 mutex_unlock(ldlm_namespace_lock(client));
1225 break;
1226 }
1227 ns = ldlm_namespace_first_locked(client);
1228
1229 spin_lock(&ns->ns_lock);
1230
1231
1232
1233
1234 if (ns->ns_stopping) {
1235 skip = 1;
1236 } else {
1237 skip = 0;
1238 ldlm_namespace_get(ns);
1239 }
1240 spin_unlock(&ns->ns_lock);
1241
1242 ldlm_namespace_move_locked(ns, client);
1243 mutex_unlock(ldlm_namespace_lock(client));
1244
1245
1246
1247
1248 if (!skip) {
1249 ldlm_pool_recalc(&ns->ns_pool);
1250 ldlm_namespace_put(ns);
1251 }
1252 }
1253}
1254EXPORT_SYMBOL(ldlm_pools_recalc);
1255
1256static int ldlm_pools_thread_main(void *arg)
1257{
1258 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
1259 ENTRY;
1260
1261 thread_set_flags(thread, SVC_RUNNING);
1262 wake_up(&thread->t_ctl_waitq);
1263
1264 CDEBUG(D_DLMTRACE, "%s: pool thread starting, process %d\n",
1265 "ldlm_poold", current_pid());
1266
1267 while (1) {
1268 struct l_wait_info lwi;
1269
1270
1271
1272
1273 ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
1274 ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
1275
1276
1277
1278
1279
1280 lwi = LWI_TIMEOUT(cfs_time_seconds(LDLM_POOLS_THREAD_PERIOD),
1281 NULL, NULL);
1282 l_wait_event(thread->t_ctl_waitq,
1283 thread_is_stopping(thread) ||
1284 thread_is_event(thread),
1285 &lwi);
1286
1287 if (thread_test_and_clear_flags(thread, SVC_STOPPING))
1288 break;
1289 else
1290 thread_test_and_clear_flags(thread, SVC_EVENT);
1291 }
1292
1293 thread_set_flags(thread, SVC_STOPPED);
1294 wake_up(&thread->t_ctl_waitq);
1295
1296 CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n",
1297 "ldlm_poold", current_pid());
1298
1299 complete_and_exit(&ldlm_pools_comp, 0);
1300}
1301
1302static int ldlm_pools_thread_start(void)
1303{
1304 struct l_wait_info lwi = { 0 };
1305 task_t *task;
1306 ENTRY;
1307
1308 if (ldlm_pools_thread != NULL)
1309 RETURN(-EALREADY);
1310
1311 OBD_ALLOC_PTR(ldlm_pools_thread);
1312 if (ldlm_pools_thread == NULL)
1313 RETURN(-ENOMEM);
1314
1315 init_completion(&ldlm_pools_comp);
1316 init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq);
1317
1318 task = kthread_run(ldlm_pools_thread_main, ldlm_pools_thread,
1319 "ldlm_poold");
1320 if (IS_ERR(task)) {
1321 CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task));
1322 OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread));
1323 ldlm_pools_thread = NULL;
1324 RETURN(PTR_ERR(task));
1325 }
1326 l_wait_event(ldlm_pools_thread->t_ctl_waitq,
1327 thread_is_running(ldlm_pools_thread), &lwi);
1328 RETURN(0);
1329}
1330
1331static void ldlm_pools_thread_stop(void)
1332{
1333 ENTRY;
1334
1335 if (ldlm_pools_thread == NULL) {
1336 EXIT;
1337 return;
1338 }
1339
1340 thread_set_flags(ldlm_pools_thread, SVC_STOPPING);
1341 wake_up(&ldlm_pools_thread->t_ctl_waitq);
1342
1343
1344
1345
1346
1347
1348 wait_for_completion(&ldlm_pools_comp);
1349 OBD_FREE_PTR(ldlm_pools_thread);
1350 ldlm_pools_thread = NULL;
1351 EXIT;
1352}
1353
1354int ldlm_pools_init(void)
1355{
1356 int rc;
1357 ENTRY;
1358
1359 rc = ldlm_pools_thread_start();
1360 if (rc == 0) {
1361 ldlm_pools_srv_shrinker =
1362 set_shrinker(DEFAULT_SEEKS,
1363 ldlm_pools_srv_shrink);
1364 ldlm_pools_cli_shrinker =
1365 set_shrinker(DEFAULT_SEEKS,
1366 ldlm_pools_cli_shrink);
1367 }
1368 RETURN(rc);
1369}
1370EXPORT_SYMBOL(ldlm_pools_init);
1371
1372void ldlm_pools_fini(void)
1373{
1374 if (ldlm_pools_srv_shrinker != NULL) {
1375 remove_shrinker(ldlm_pools_srv_shrinker);
1376 ldlm_pools_srv_shrinker = NULL;
1377 }
1378 if (ldlm_pools_cli_shrinker != NULL) {
1379 remove_shrinker(ldlm_pools_cli_shrinker);
1380 ldlm_pools_cli_shrinker = NULL;
1381 }
1382 ldlm_pools_thread_stop();
1383}
1384EXPORT_SYMBOL(ldlm_pools_fini);
1385