1
2
3
4
5
6
7#include <linux/module.h>
8#include <linux/slab.h>
9#include <linux/blkdev.h>
10#include <linux/bio.h>
11#include <linux/blktrace_api.h>
12#include <linux/blk-cgroup.h>
13#include "blk.h"
14
15
16static int throtl_grp_quantum = 8;
17
18
19static int throtl_quantum = 32;
20
21
22static unsigned long throtl_slice = HZ/10;
23
24static struct blkcg_policy blkcg_policy_throtl;
25
26
27static struct workqueue_struct *kthrotld_workqueue;
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52struct throtl_qnode {
53 struct list_head node;
54 struct bio_list bios;
55 struct throtl_grp *tg;
56};
57
58struct throtl_service_queue {
59 struct throtl_service_queue *parent_sq;
60
61
62
63
64
65 struct list_head queued[2];
66 unsigned int nr_queued[2];
67
68
69
70
71
72 struct rb_root pending_tree;
73 struct rb_node *first_pending;
74 unsigned int nr_pending;
75 unsigned long first_pending_disptime;
76 struct timer_list pending_timer;
77};
78
79enum tg_state_flags {
80 THROTL_TG_PENDING = 1 << 0,
81 THROTL_TG_WAS_EMPTY = 1 << 1,
82};
83
84#define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
85
86
87struct tg_stats_cpu {
88
89 struct blkg_rwstat service_bytes;
90
91 struct blkg_rwstat serviced;
92};
93
94struct throtl_grp {
95
96 struct blkg_policy_data pd;
97
98
99 struct rb_node rb_node;
100
101
102 struct throtl_data *td;
103
104
105 struct throtl_service_queue service_queue;
106
107
108
109
110
111
112
113
114
115 struct throtl_qnode qnode_on_self[2];
116 struct throtl_qnode qnode_on_parent[2];
117
118
119
120
121
122
123 unsigned long disptime;
124
125 unsigned int flags;
126
127
128 bool has_rules[2];
129
130
131 uint64_t bps[2];
132
133
134 unsigned int iops[2];
135
136
137 uint64_t bytes_disp[2];
138
139 unsigned int io_disp[2];
140
141
142 unsigned long slice_start[2];
143 unsigned long slice_end[2];
144
145
146 struct tg_stats_cpu __percpu *stats_cpu;
147
148
149 struct list_head stats_alloc_node;
150};
151
152struct throtl_data
153{
154
155 struct throtl_service_queue service_queue;
156
157 struct request_queue *queue;
158
159
160 unsigned int nr_queued[2];
161
162
163
164
165 unsigned int nr_undestroyed_grps;
166
167
168 struct work_struct dispatch_work;
169};
170
171
172static DEFINE_SPINLOCK(tg_stats_alloc_lock);
173static LIST_HEAD(tg_stats_alloc_list);
174
175static void tg_stats_alloc_fn(struct work_struct *);
176static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn);
177
178static void throtl_pending_timer_fn(unsigned long arg);
179
180static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
181{
182 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
183}
184
185static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
186{
187 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
188}
189
190static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
191{
192 return pd_to_blkg(&tg->pd);
193}
194
195static inline struct throtl_grp *td_root_tg(struct throtl_data *td)
196{
197 return blkg_to_tg(td->queue->root_blkg);
198}
199
200
201
202
203
204
205
206
207static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
208{
209 if (sq && sq->parent_sq)
210 return container_of(sq, struct throtl_grp, service_queue);
211 else
212 return NULL;
213}
214
215
216
217
218
219
220
221
222static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
223{
224 struct throtl_grp *tg = sq_to_tg(sq);
225
226 if (tg)
227 return tg->td;
228 else
229 return container_of(sq, struct throtl_data, service_queue);
230}
231
232
233
234
235
236
237
238
239
240
241
242
243
244#define throtl_log(sq, fmt, args...) do { \
245 struct throtl_grp *__tg = sq_to_tg((sq)); \
246 struct throtl_data *__td = sq_to_td((sq)); \
247 \
248 (void)__td; \
249 if ((__tg)) { \
250 char __pbuf[128]; \
251 \
252 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
253 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
254 } else { \
255 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
256 } \
257} while (0)
258
259static void tg_stats_init(struct tg_stats_cpu *tg_stats)
260{
261 blkg_rwstat_init(&tg_stats->service_bytes);
262 blkg_rwstat_init(&tg_stats->serviced);
263}
264
265
266
267
268
269
270static void tg_stats_alloc_fn(struct work_struct *work)
271{
272 static struct tg_stats_cpu *stats_cpu;
273 struct delayed_work *dwork = to_delayed_work(work);
274 bool empty = false;
275
276alloc_stats:
277 if (!stats_cpu) {
278 int cpu;
279
280 stats_cpu = alloc_percpu(struct tg_stats_cpu);
281 if (!stats_cpu) {
282
283 schedule_delayed_work(dwork, msecs_to_jiffies(10));
284 return;
285 }
286 for_each_possible_cpu(cpu)
287 tg_stats_init(per_cpu_ptr(stats_cpu, cpu));
288 }
289
290 spin_lock_irq(&tg_stats_alloc_lock);
291
292 if (!list_empty(&tg_stats_alloc_list)) {
293 struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list,
294 struct throtl_grp,
295 stats_alloc_node);
296 swap(tg->stats_cpu, stats_cpu);
297 list_del_init(&tg->stats_alloc_node);
298 }
299
300 empty = list_empty(&tg_stats_alloc_list);
301 spin_unlock_irq(&tg_stats_alloc_lock);
302 if (!empty)
303 goto alloc_stats;
304}
305
306static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
307{
308 INIT_LIST_HEAD(&qn->node);
309 bio_list_init(&qn->bios);
310 qn->tg = tg;
311}
312
313
314
315
316
317
318
319
320
321
322
323static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
324 struct list_head *queued)
325{
326 bio_list_add(&qn->bios, bio);
327 if (list_empty(&qn->node)) {
328 list_add_tail(&qn->node, queued);
329 blkg_get(tg_to_blkg(qn->tg));
330 }
331}
332
333
334
335
336
337static struct bio *throtl_peek_queued(struct list_head *queued)
338{
339 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
340 struct bio *bio;
341
342 if (list_empty(queued))
343 return NULL;
344
345 bio = bio_list_peek(&qn->bios);
346 WARN_ON_ONCE(!bio);
347 return bio;
348}
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364static struct bio *throtl_pop_queued(struct list_head *queued,
365 struct throtl_grp **tg_to_put)
366{
367 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
368 struct bio *bio;
369
370 if (list_empty(queued))
371 return NULL;
372
373 bio = bio_list_pop(&qn->bios);
374 WARN_ON_ONCE(!bio);
375
376 if (bio_list_empty(&qn->bios)) {
377 list_del_init(&qn->node);
378 if (tg_to_put)
379 *tg_to_put = qn->tg;
380 else
381 blkg_put(tg_to_blkg(qn->tg));
382 } else {
383 list_move_tail(&qn->node, queued);
384 }
385
386 return bio;
387}
388
389
390static void throtl_service_queue_init(struct throtl_service_queue *sq,
391 struct throtl_service_queue *parent_sq)
392{
393 INIT_LIST_HEAD(&sq->queued[0]);
394 INIT_LIST_HEAD(&sq->queued[1]);
395 sq->pending_tree = RB_ROOT;
396 sq->parent_sq = parent_sq;
397 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
398 (unsigned long)sq);
399}
400
401static void throtl_service_queue_exit(struct throtl_service_queue *sq)
402{
403 del_timer_sync(&sq->pending_timer);
404}
405
406static void throtl_pd_init(struct blkcg_gq *blkg)
407{
408 struct throtl_grp *tg = blkg_to_tg(blkg);
409 struct throtl_data *td = blkg->q->td;
410 struct throtl_service_queue *parent_sq;
411 unsigned long flags;
412 int rw;
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427 parent_sq = &td->service_queue;
428
429 if (cgroup_on_dfl(blkg->blkcg->css.cgroup) && blkg->parent)
430 parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
431
432 throtl_service_queue_init(&tg->service_queue, parent_sq);
433
434 for (rw = READ; rw <= WRITE; rw++) {
435 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
436 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
437 }
438
439 RB_CLEAR_NODE(&tg->rb_node);
440 tg->td = td;
441
442 tg->bps[READ] = -1;
443 tg->bps[WRITE] = -1;
444 tg->iops[READ] = -1;
445 tg->iops[WRITE] = -1;
446
447
448
449
450
451
452 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
453 list_add(&tg->stats_alloc_node, &tg_stats_alloc_list);
454 schedule_delayed_work(&tg_stats_alloc_work, 0);
455 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
456}
457
458
459
460
461
462
463static void tg_update_has_rules(struct throtl_grp *tg)
464{
465 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
466 int rw;
467
468 for (rw = READ; rw <= WRITE; rw++)
469 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
470 (tg->bps[rw] != -1 || tg->iops[rw] != -1);
471}
472
473static void throtl_pd_online(struct blkcg_gq *blkg)
474{
475
476
477
478
479 tg_update_has_rules(blkg_to_tg(blkg));
480}
481
482static void throtl_pd_exit(struct blkcg_gq *blkg)
483{
484 struct throtl_grp *tg = blkg_to_tg(blkg);
485 unsigned long flags;
486
487 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
488 list_del_init(&tg->stats_alloc_node);
489 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
490
491 free_percpu(tg->stats_cpu);
492
493 throtl_service_queue_exit(&tg->service_queue);
494}
495
496static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
497{
498 struct throtl_grp *tg = blkg_to_tg(blkg);
499 int cpu;
500
501 if (tg->stats_cpu == NULL)
502 return;
503
504 for_each_possible_cpu(cpu) {
505 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
506
507 blkg_rwstat_reset(&sc->service_bytes);
508 blkg_rwstat_reset(&sc->serviced);
509 }
510}
511
512static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
513 struct blkcg *blkcg)
514{
515
516
517
518
519 if (blkcg == &blkcg_root)
520 return td_root_tg(td);
521
522 return blkg_to_tg(blkg_lookup(blkcg, td->queue));
523}
524
525static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
526 struct blkcg *blkcg)
527{
528 struct request_queue *q = td->queue;
529 struct throtl_grp *tg = NULL;
530
531
532
533
534
535 if (blkcg == &blkcg_root) {
536 tg = td_root_tg(td);
537 } else {
538 struct blkcg_gq *blkg;
539
540 blkg = blkg_lookup_create(blkcg, q);
541
542
543 if (!IS_ERR(blkg))
544 tg = blkg_to_tg(blkg);
545 else if (!blk_queue_dying(q))
546 tg = td_root_tg(td);
547 }
548
549 return tg;
550}
551
552static struct throtl_grp *
553throtl_rb_first(struct throtl_service_queue *parent_sq)
554{
555
556 if (!parent_sq->nr_pending)
557 return NULL;
558
559 if (!parent_sq->first_pending)
560 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
561
562 if (parent_sq->first_pending)
563 return rb_entry_tg(parent_sq->first_pending);
564
565 return NULL;
566}
567
568static void rb_erase_init(struct rb_node *n, struct rb_root *root)
569{
570 rb_erase(n, root);
571 RB_CLEAR_NODE(n);
572}
573
574static void throtl_rb_erase(struct rb_node *n,
575 struct throtl_service_queue *parent_sq)
576{
577 if (parent_sq->first_pending == n)
578 parent_sq->first_pending = NULL;
579 rb_erase_init(n, &parent_sq->pending_tree);
580 --parent_sq->nr_pending;
581}
582
583static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
584{
585 struct throtl_grp *tg;
586
587 tg = throtl_rb_first(parent_sq);
588 if (!tg)
589 return;
590
591 parent_sq->first_pending_disptime = tg->disptime;
592}
593
594static void tg_service_queue_add(struct throtl_grp *tg)
595{
596 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
597 struct rb_node **node = &parent_sq->pending_tree.rb_node;
598 struct rb_node *parent = NULL;
599 struct throtl_grp *__tg;
600 unsigned long key = tg->disptime;
601 int left = 1;
602
603 while (*node != NULL) {
604 parent = *node;
605 __tg = rb_entry_tg(parent);
606
607 if (time_before(key, __tg->disptime))
608 node = &parent->rb_left;
609 else {
610 node = &parent->rb_right;
611 left = 0;
612 }
613 }
614
615 if (left)
616 parent_sq->first_pending = &tg->rb_node;
617
618 rb_link_node(&tg->rb_node, parent, node);
619 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
620}
621
622static void __throtl_enqueue_tg(struct throtl_grp *tg)
623{
624 tg_service_queue_add(tg);
625 tg->flags |= THROTL_TG_PENDING;
626 tg->service_queue.parent_sq->nr_pending++;
627}
628
629static void throtl_enqueue_tg(struct throtl_grp *tg)
630{
631 if (!(tg->flags & THROTL_TG_PENDING))
632 __throtl_enqueue_tg(tg);
633}
634
635static void __throtl_dequeue_tg(struct throtl_grp *tg)
636{
637 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
638 tg->flags &= ~THROTL_TG_PENDING;
639}
640
641static void throtl_dequeue_tg(struct throtl_grp *tg)
642{
643 if (tg->flags & THROTL_TG_PENDING)
644 __throtl_dequeue_tg(tg);
645}
646
647
648static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
649 unsigned long expires)
650{
651 mod_timer(&sq->pending_timer, expires);
652 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
653 expires - jiffies, jiffies);
654}
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
675 bool force)
676{
677
678 if (!sq->nr_pending)
679 return true;
680
681 update_min_dispatch_time(sq);
682
683
684 if (force || time_after(sq->first_pending_disptime, jiffies)) {
685 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
686 return true;
687 }
688
689
690 return false;
691}
692
693static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
694 bool rw, unsigned long start)
695{
696 tg->bytes_disp[rw] = 0;
697 tg->io_disp[rw] = 0;
698
699
700
701
702
703
704
705 if (time_after_eq(start, tg->slice_start[rw]))
706 tg->slice_start[rw] = start;
707
708 tg->slice_end[rw] = jiffies + throtl_slice;
709 throtl_log(&tg->service_queue,
710 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
711 rw == READ ? 'R' : 'W', tg->slice_start[rw],
712 tg->slice_end[rw], jiffies);
713}
714
715static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
716{
717 tg->bytes_disp[rw] = 0;
718 tg->io_disp[rw] = 0;
719 tg->slice_start[rw] = jiffies;
720 tg->slice_end[rw] = jiffies + throtl_slice;
721 throtl_log(&tg->service_queue,
722 "[%c] new slice start=%lu end=%lu jiffies=%lu",
723 rw == READ ? 'R' : 'W', tg->slice_start[rw],
724 tg->slice_end[rw], jiffies);
725}
726
727static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
728 unsigned long jiffy_end)
729{
730 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
731}
732
733static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
734 unsigned long jiffy_end)
735{
736 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
737 throtl_log(&tg->service_queue,
738 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
739 rw == READ ? 'R' : 'W', tg->slice_start[rw],
740 tg->slice_end[rw], jiffies);
741}
742
743
744static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
745{
746 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
747 return false;
748
749 return 1;
750}
751
752
753static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
754{
755 unsigned long nr_slices, time_elapsed, io_trim;
756 u64 bytes_trim, tmp;
757
758 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
759
760
761
762
763
764
765 if (throtl_slice_used(tg, rw))
766 return;
767
768
769
770
771
772
773
774
775
776 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
777
778 time_elapsed = jiffies - tg->slice_start[rw];
779
780 nr_slices = time_elapsed / throtl_slice;
781
782 if (!nr_slices)
783 return;
784 tmp = tg->bps[rw] * throtl_slice * nr_slices;
785 do_div(tmp, HZ);
786 bytes_trim = tmp;
787
788 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
789
790 if (!bytes_trim && !io_trim)
791 return;
792
793 if (tg->bytes_disp[rw] >= bytes_trim)
794 tg->bytes_disp[rw] -= bytes_trim;
795 else
796 tg->bytes_disp[rw] = 0;
797
798 if (tg->io_disp[rw] >= io_trim)
799 tg->io_disp[rw] -= io_trim;
800 else
801 tg->io_disp[rw] = 0;
802
803 tg->slice_start[rw] += nr_slices * throtl_slice;
804
805 throtl_log(&tg->service_queue,
806 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
807 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
808 tg->slice_start[rw], tg->slice_end[rw], jiffies);
809}
810
811static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
812 unsigned long *wait)
813{
814 bool rw = bio_data_dir(bio);
815 unsigned int io_allowed;
816 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
817 u64 tmp;
818
819 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
820
821
822 if (!jiffy_elapsed)
823 jiffy_elapsed_rnd = throtl_slice;
824
825 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
826
827
828
829
830
831
832
833
834 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
835 do_div(tmp, HZ);
836
837 if (tmp > UINT_MAX)
838 io_allowed = UINT_MAX;
839 else
840 io_allowed = tmp;
841
842 if (tg->io_disp[rw] + 1 <= io_allowed) {
843 if (wait)
844 *wait = 0;
845 return true;
846 }
847
848
849 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
850
851 if (jiffy_wait > jiffy_elapsed)
852 jiffy_wait = jiffy_wait - jiffy_elapsed;
853 else
854 jiffy_wait = 1;
855
856 if (wait)
857 *wait = jiffy_wait;
858 return 0;
859}
860
861static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
862 unsigned long *wait)
863{
864 bool rw = bio_data_dir(bio);
865 u64 bytes_allowed, extra_bytes, tmp;
866 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
867
868 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
869
870
871 if (!jiffy_elapsed)
872 jiffy_elapsed_rnd = throtl_slice;
873
874 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
875
876 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
877 do_div(tmp, HZ);
878 bytes_allowed = tmp;
879
880 if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
881 if (wait)
882 *wait = 0;
883 return true;
884 }
885
886
887 extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
888 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
889
890 if (!jiffy_wait)
891 jiffy_wait = 1;
892
893
894
895
896
897 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
898 if (wait)
899 *wait = jiffy_wait;
900 return 0;
901}
902
903
904
905
906
907static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
908 unsigned long *wait)
909{
910 bool rw = bio_data_dir(bio);
911 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
912
913
914
915
916
917
918
919 BUG_ON(tg->service_queue.nr_queued[rw] &&
920 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
921
922
923 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
924 if (wait)
925 *wait = 0;
926 return true;
927 }
928
929
930
931
932
933
934 if (throtl_slice_used(tg, rw))
935 throtl_start_new_slice(tg, rw);
936 else {
937 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
938 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
939 }
940
941 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
942 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
943 if (wait)
944 *wait = 0;
945 return 1;
946 }
947
948 max_wait = max(bps_wait, iops_wait);
949
950 if (wait)
951 *wait = max_wait;
952
953 if (time_before(tg->slice_end[rw], jiffies + max_wait))
954 throtl_extend_slice(tg, rw, jiffies + max_wait);
955
956 return 0;
957}
958
959static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
960 int rw)
961{
962 struct throtl_grp *tg = blkg_to_tg(blkg);
963 struct tg_stats_cpu *stats_cpu;
964 unsigned long flags;
965
966
967 if (tg->stats_cpu == NULL)
968 return;
969
970
971
972
973
974
975 local_irq_save(flags);
976
977 stats_cpu = this_cpu_ptr(tg->stats_cpu);
978
979 blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
980 blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
981
982 local_irq_restore(flags);
983}
984
985static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
986{
987 bool rw = bio_data_dir(bio);
988
989
990 tg->bytes_disp[rw] += bio->bi_iter.bi_size;
991 tg->io_disp[rw]++;
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004 if (!(bio->bi_rw & REQ_THROTTLED)) {
1005 bio->bi_rw |= REQ_THROTTLED;
1006 throtl_update_dispatch_stats(tg_to_blkg(tg),
1007 bio->bi_iter.bi_size, bio->bi_rw);
1008 }
1009}
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1021 struct throtl_grp *tg)
1022{
1023 struct throtl_service_queue *sq = &tg->service_queue;
1024 bool rw = bio_data_dir(bio);
1025
1026 if (!qn)
1027 qn = &tg->qnode_on_self[rw];
1028
1029
1030
1031
1032
1033
1034
1035 if (!sq->nr_queued[rw])
1036 tg->flags |= THROTL_TG_WAS_EMPTY;
1037
1038 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1039
1040 sq->nr_queued[rw]++;
1041 throtl_enqueue_tg(tg);
1042}
1043
1044static void tg_update_disptime(struct throtl_grp *tg)
1045{
1046 struct throtl_service_queue *sq = &tg->service_queue;
1047 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1048 struct bio *bio;
1049
1050 if ((bio = throtl_peek_queued(&sq->queued[READ])))
1051 tg_may_dispatch(tg, bio, &read_wait);
1052
1053 if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1054 tg_may_dispatch(tg, bio, &write_wait);
1055
1056 min_wait = min(read_wait, write_wait);
1057 disptime = jiffies + min_wait;
1058
1059
1060 throtl_dequeue_tg(tg);
1061 tg->disptime = disptime;
1062 throtl_enqueue_tg(tg);
1063
1064
1065 tg->flags &= ~THROTL_TG_WAS_EMPTY;
1066}
1067
1068static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1069 struct throtl_grp *parent_tg, bool rw)
1070{
1071 if (throtl_slice_used(parent_tg, rw)) {
1072 throtl_start_new_slice_with_credit(parent_tg, rw,
1073 child_tg->slice_start[rw]);
1074 }
1075
1076}
1077
1078static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1079{
1080 struct throtl_service_queue *sq = &tg->service_queue;
1081 struct throtl_service_queue *parent_sq = sq->parent_sq;
1082 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1083 struct throtl_grp *tg_to_put = NULL;
1084 struct bio *bio;
1085
1086
1087
1088
1089
1090
1091
1092 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1093 sq->nr_queued[rw]--;
1094
1095 throtl_charge_bio(tg, bio);
1096
1097
1098
1099
1100
1101
1102
1103
1104 if (parent_tg) {
1105 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1106 start_parent_slice_with_credit(tg, parent_tg, rw);
1107 } else {
1108 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1109 &parent_sq->queued[rw]);
1110 BUG_ON(tg->td->nr_queued[rw] <= 0);
1111 tg->td->nr_queued[rw]--;
1112 }
1113
1114 throtl_trim_slice(tg, rw);
1115
1116 if (tg_to_put)
1117 blkg_put(tg_to_blkg(tg_to_put));
1118}
1119
1120static int throtl_dispatch_tg(struct throtl_grp *tg)
1121{
1122 struct throtl_service_queue *sq = &tg->service_queue;
1123 unsigned int nr_reads = 0, nr_writes = 0;
1124 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1125 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1126 struct bio *bio;
1127
1128
1129
1130 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1131 tg_may_dispatch(tg, bio, NULL)) {
1132
1133 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1134 nr_reads++;
1135
1136 if (nr_reads >= max_nr_reads)
1137 break;
1138 }
1139
1140 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1141 tg_may_dispatch(tg, bio, NULL)) {
1142
1143 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1144 nr_writes++;
1145
1146 if (nr_writes >= max_nr_writes)
1147 break;
1148 }
1149
1150 return nr_reads + nr_writes;
1151}
1152
1153static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1154{
1155 unsigned int nr_disp = 0;
1156
1157 while (1) {
1158 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1159 struct throtl_service_queue *sq = &tg->service_queue;
1160
1161 if (!tg)
1162 break;
1163
1164 if (time_before(jiffies, tg->disptime))
1165 break;
1166
1167 throtl_dequeue_tg(tg);
1168
1169 nr_disp += throtl_dispatch_tg(tg);
1170
1171 if (sq->nr_queued[0] || sq->nr_queued[1])
1172 tg_update_disptime(tg);
1173
1174 if (nr_disp >= throtl_quantum)
1175 break;
1176 }
1177
1178 return nr_disp;
1179}
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196static void throtl_pending_timer_fn(unsigned long arg)
1197{
1198 struct throtl_service_queue *sq = (void *)arg;
1199 struct throtl_grp *tg = sq_to_tg(sq);
1200 struct throtl_data *td = sq_to_td(sq);
1201 struct request_queue *q = td->queue;
1202 struct throtl_service_queue *parent_sq;
1203 bool dispatched;
1204 int ret;
1205
1206 spin_lock_irq(q->queue_lock);
1207again:
1208 parent_sq = sq->parent_sq;
1209 dispatched = false;
1210
1211 while (true) {
1212 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1213 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1214 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1215
1216 ret = throtl_select_dispatch(sq);
1217 if (ret) {
1218 throtl_log(sq, "bios disp=%u", ret);
1219 dispatched = true;
1220 }
1221
1222 if (throtl_schedule_next_dispatch(sq, false))
1223 break;
1224
1225
1226 spin_unlock_irq(q->queue_lock);
1227 cpu_relax();
1228 spin_lock_irq(q->queue_lock);
1229 }
1230
1231 if (!dispatched)
1232 goto out_unlock;
1233
1234 if (parent_sq) {
1235
1236 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1237 tg_update_disptime(tg);
1238 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1239
1240 sq = parent_sq;
1241 tg = sq_to_tg(sq);
1242 goto again;
1243 }
1244 }
1245 } else {
1246
1247 queue_work(kthrotld_workqueue, &td->dispatch_work);
1248 }
1249out_unlock:
1250 spin_unlock_irq(q->queue_lock);
1251}
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1262{
1263 struct throtl_data *td = container_of(work, struct throtl_data,
1264 dispatch_work);
1265 struct throtl_service_queue *td_sq = &td->service_queue;
1266 struct request_queue *q = td->queue;
1267 struct bio_list bio_list_on_stack;
1268 struct bio *bio;
1269 struct blk_plug plug;
1270 int rw;
1271
1272 bio_list_init(&bio_list_on_stack);
1273
1274 spin_lock_irq(q->queue_lock);
1275 for (rw = READ; rw <= WRITE; rw++)
1276 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1277 bio_list_add(&bio_list_on_stack, bio);
1278 spin_unlock_irq(q->queue_lock);
1279
1280 if (!bio_list_empty(&bio_list_on_stack)) {
1281 blk_start_plug(&plug);
1282 while((bio = bio_list_pop(&bio_list_on_stack)))
1283 generic_make_request(bio);
1284 blk_finish_plug(&plug);
1285 }
1286}
1287
1288static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
1289 struct blkg_policy_data *pd, int off)
1290{
1291 struct throtl_grp *tg = pd_to_tg(pd);
1292 struct blkg_rwstat rwstat = { }, tmp;
1293 int i, cpu;
1294
1295 if (tg->stats_cpu == NULL)
1296 return 0;
1297
1298 for_each_possible_cpu(cpu) {
1299 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
1300
1301 tmp = blkg_rwstat_read((void *)sc + off);
1302 for (i = 0; i < BLKG_RWSTAT_NR; i++)
1303 rwstat.cnt[i] += tmp.cnt[i];
1304 }
1305
1306 return __blkg_prfill_rwstat(sf, pd, &rwstat);
1307}
1308
1309static int tg_print_cpu_rwstat(struct seq_file *sf, void *v)
1310{
1311 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_cpu_rwstat,
1312 &blkcg_policy_throtl, seq_cft(sf)->private, true);
1313 return 0;
1314}
1315
1316static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1317 int off)
1318{
1319 struct throtl_grp *tg = pd_to_tg(pd);
1320 u64 v = *(u64 *)((void *)tg + off);
1321
1322 if (v == -1)
1323 return 0;
1324 return __blkg_prfill_u64(sf, pd, v);
1325}
1326
1327static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1328 int off)
1329{
1330 struct throtl_grp *tg = pd_to_tg(pd);
1331 unsigned int v = *(unsigned int *)((void *)tg + off);
1332
1333 if (v == -1)
1334 return 0;
1335 return __blkg_prfill_u64(sf, pd, v);
1336}
1337
1338static int tg_print_conf_u64(struct seq_file *sf, void *v)
1339{
1340 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1341 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1342 return 0;
1343}
1344
1345static int tg_print_conf_uint(struct seq_file *sf, void *v)
1346{
1347 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1348 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1349 return 0;
1350}
1351
1352static ssize_t tg_set_conf(struct kernfs_open_file *of,
1353 char *buf, size_t nbytes, loff_t off, bool is_u64)
1354{
1355 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1356 struct blkg_conf_ctx ctx;
1357 struct throtl_grp *tg;
1358 struct throtl_service_queue *sq;
1359 struct blkcg_gq *blkg;
1360 struct cgroup_subsys_state *pos_css;
1361 int ret;
1362
1363 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1364 if (ret)
1365 return ret;
1366
1367 tg = blkg_to_tg(ctx.blkg);
1368 sq = &tg->service_queue;
1369
1370 if (!ctx.v)
1371 ctx.v = -1;
1372
1373 if (is_u64)
1374 *(u64 *)((void *)tg + of_cft(of)->private) = ctx.v;
1375 else
1376 *(unsigned int *)((void *)tg + of_cft(of)->private) = ctx.v;
1377
1378 throtl_log(&tg->service_queue,
1379 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1380 tg->bps[READ], tg->bps[WRITE],
1381 tg->iops[READ], tg->iops[WRITE]);
1382
1383
1384
1385
1386
1387
1388
1389
1390 blkg_for_each_descendant_pre(blkg, pos_css, ctx.blkg)
1391 tg_update_has_rules(blkg_to_tg(blkg));
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401 throtl_start_new_slice(tg, 0);
1402 throtl_start_new_slice(tg, 1);
1403
1404 if (tg->flags & THROTL_TG_PENDING) {
1405 tg_update_disptime(tg);
1406 throtl_schedule_next_dispatch(sq->parent_sq, true);
1407 }
1408
1409 blkg_conf_finish(&ctx);
1410 return nbytes;
1411}
1412
1413static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1414 char *buf, size_t nbytes, loff_t off)
1415{
1416 return tg_set_conf(of, buf, nbytes, off, true);
1417}
1418
1419static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1420 char *buf, size_t nbytes, loff_t off)
1421{
1422 return tg_set_conf(of, buf, nbytes, off, false);
1423}
1424
1425static struct cftype throtl_files[] = {
1426 {
1427 .name = "throttle.read_bps_device",
1428 .private = offsetof(struct throtl_grp, bps[READ]),
1429 .seq_show = tg_print_conf_u64,
1430 .write = tg_set_conf_u64,
1431 },
1432 {
1433 .name = "throttle.write_bps_device",
1434 .private = offsetof(struct throtl_grp, bps[WRITE]),
1435 .seq_show = tg_print_conf_u64,
1436 .write = tg_set_conf_u64,
1437 },
1438 {
1439 .name = "throttle.read_iops_device",
1440 .private = offsetof(struct throtl_grp, iops[READ]),
1441 .seq_show = tg_print_conf_uint,
1442 .write = tg_set_conf_uint,
1443 },
1444 {
1445 .name = "throttle.write_iops_device",
1446 .private = offsetof(struct throtl_grp, iops[WRITE]),
1447 .seq_show = tg_print_conf_uint,
1448 .write = tg_set_conf_uint,
1449 },
1450 {
1451 .name = "throttle.io_service_bytes",
1452 .private = offsetof(struct tg_stats_cpu, service_bytes),
1453 .seq_show = tg_print_cpu_rwstat,
1454 },
1455 {
1456 .name = "throttle.io_serviced",
1457 .private = offsetof(struct tg_stats_cpu, serviced),
1458 .seq_show = tg_print_cpu_rwstat,
1459 },
1460 { }
1461};
1462
1463static void throtl_shutdown_wq(struct request_queue *q)
1464{
1465 struct throtl_data *td = q->td;
1466
1467 cancel_work_sync(&td->dispatch_work);
1468}
1469
1470static struct blkcg_policy blkcg_policy_throtl = {
1471 .pd_size = sizeof(struct throtl_grp),
1472 .cftypes = throtl_files,
1473
1474 .pd_init_fn = throtl_pd_init,
1475 .pd_online_fn = throtl_pd_online,
1476 .pd_exit_fn = throtl_pd_exit,
1477 .pd_reset_stats_fn = throtl_pd_reset_stats,
1478};
1479
1480bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
1481{
1482 struct throtl_data *td = q->td;
1483 struct throtl_qnode *qn = NULL;
1484 struct throtl_grp *tg;
1485 struct throtl_service_queue *sq;
1486 bool rw = bio_data_dir(bio);
1487 struct blkcg *blkcg;
1488 bool throttled = false;
1489
1490
1491 if (bio->bi_rw & REQ_THROTTLED)
1492 goto out;
1493
1494
1495
1496
1497
1498
1499 rcu_read_lock();
1500 blkcg = bio_blkcg(bio);
1501 tg = throtl_lookup_tg(td, blkcg);
1502 if (tg) {
1503 if (!tg->has_rules[rw]) {
1504 throtl_update_dispatch_stats(tg_to_blkg(tg),
1505 bio->bi_iter.bi_size, bio->bi_rw);
1506 goto out_unlock_rcu;
1507 }
1508 }
1509
1510
1511
1512
1513
1514 spin_lock_irq(q->queue_lock);
1515 tg = throtl_lookup_create_tg(td, blkcg);
1516 if (unlikely(!tg))
1517 goto out_unlock;
1518
1519 sq = &tg->service_queue;
1520
1521 while (true) {
1522
1523 if (sq->nr_queued[rw])
1524 break;
1525
1526
1527 if (!tg_may_dispatch(tg, bio, NULL))
1528 break;
1529
1530
1531 throtl_charge_bio(tg, bio);
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544 throtl_trim_slice(tg, rw);
1545
1546
1547
1548
1549
1550
1551 qn = &tg->qnode_on_parent[rw];
1552 sq = sq->parent_sq;
1553 tg = sq_to_tg(sq);
1554 if (!tg)
1555 goto out_unlock;
1556 }
1557
1558
1559 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1560 rw == READ ? 'R' : 'W',
1561 tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw],
1562 tg->io_disp[rw], tg->iops[rw],
1563 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1564
1565 bio_associate_current(bio);
1566 tg->td->nr_queued[rw]++;
1567 throtl_add_bio_tg(bio, qn, tg);
1568 throttled = true;
1569
1570
1571
1572
1573
1574
1575
1576 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1577 tg_update_disptime(tg);
1578 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1579 }
1580
1581out_unlock:
1582 spin_unlock_irq(q->queue_lock);
1583out_unlock_rcu:
1584 rcu_read_unlock();
1585out:
1586
1587
1588
1589
1590
1591 if (!throttled)
1592 bio->bi_rw &= ~REQ_THROTTLED;
1593 return throttled;
1594}
1595
1596
1597
1598
1599
1600
1601static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1602{
1603 struct throtl_grp *tg;
1604
1605 while ((tg = throtl_rb_first(parent_sq))) {
1606 struct throtl_service_queue *sq = &tg->service_queue;
1607 struct bio *bio;
1608
1609 throtl_dequeue_tg(tg);
1610
1611 while ((bio = throtl_peek_queued(&sq->queued[READ])))
1612 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1613 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1614 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1615 }
1616}
1617
1618
1619
1620
1621
1622
1623
1624void blk_throtl_drain(struct request_queue *q)
1625 __releases(q->queue_lock) __acquires(q->queue_lock)
1626{
1627 struct throtl_data *td = q->td;
1628 struct blkcg_gq *blkg;
1629 struct cgroup_subsys_state *pos_css;
1630 struct bio *bio;
1631 int rw;
1632
1633 queue_lockdep_assert_held(q);
1634 rcu_read_lock();
1635
1636
1637
1638
1639
1640
1641
1642 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
1643 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1644
1645
1646 tg_drain_bios(&td->service_queue);
1647
1648 rcu_read_unlock();
1649 spin_unlock_irq(q->queue_lock);
1650
1651
1652 for (rw = READ; rw <= WRITE; rw++)
1653 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1654 NULL)))
1655 generic_make_request(bio);
1656
1657 spin_lock_irq(q->queue_lock);
1658}
1659
1660int blk_throtl_init(struct request_queue *q)
1661{
1662 struct throtl_data *td;
1663 int ret;
1664
1665 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1666 if (!td)
1667 return -ENOMEM;
1668
1669 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1670 throtl_service_queue_init(&td->service_queue, NULL);
1671
1672 q->td = td;
1673 td->queue = q;
1674
1675
1676 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1677 if (ret)
1678 kfree(td);
1679 return ret;
1680}
1681
1682void blk_throtl_exit(struct request_queue *q)
1683{
1684 BUG_ON(!q->td);
1685 throtl_shutdown_wq(q);
1686 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1687 kfree(q->td);
1688}
1689
1690static int __init throtl_init(void)
1691{
1692 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1693 if (!kthrotld_workqueue)
1694 panic("Failed to create kthrotld\n");
1695
1696 return blkcg_policy_register(&blkcg_policy_throtl);
1697}
1698
1699module_init(throtl_init);
1700