Loading...
1/*
2 * Interface for controlling IO bandwidth on a request queue
3 *
4 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
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 "blk-cgroup.h"
13#include "blk.h"
14
15/* Max dispatch from a group in 1 round */
16static int throtl_grp_quantum = 8;
17
18/* Total max dispatch from all groups in one round */
19static int throtl_quantum = 32;
20
21/* Throttling is performed over 100ms slice and after that slice is renewed */
22static unsigned long throtl_slice = HZ/10; /* 100 ms */
23
24static struct blkcg_policy blkcg_policy_throtl;
25
26/* A workqueue to queue throttle related work */
27static struct workqueue_struct *kthrotld_workqueue;
28static void throtl_schedule_delayed_work(struct throtl_data *td,
29 unsigned long delay);
30
31struct throtl_rb_root {
32 struct rb_root rb;
33 struct rb_node *left;
34 unsigned int count;
35 unsigned long min_disptime;
36};
37
38#define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \
39 .count = 0, .min_disptime = 0}
40
41#define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
42
43/* Per-cpu group stats */
44struct tg_stats_cpu {
45 /* total bytes transferred */
46 struct blkg_rwstat service_bytes;
47 /* total IOs serviced, post merge */
48 struct blkg_rwstat serviced;
49};
50
51struct throtl_grp {
52 /* must be the first member */
53 struct blkg_policy_data pd;
54
55 /* active throtl group service_tree member */
56 struct rb_node rb_node;
57
58 /*
59 * Dispatch time in jiffies. This is the estimated time when group
60 * will unthrottle and is ready to dispatch more bio. It is used as
61 * key to sort active groups in service tree.
62 */
63 unsigned long disptime;
64
65 unsigned int flags;
66
67 /* Two lists for READ and WRITE */
68 struct bio_list bio_lists[2];
69
70 /* Number of queued bios on READ and WRITE lists */
71 unsigned int nr_queued[2];
72
73 /* bytes per second rate limits */
74 uint64_t bps[2];
75
76 /* IOPS limits */
77 unsigned int iops[2];
78
79 /* Number of bytes disptached in current slice */
80 uint64_t bytes_disp[2];
81 /* Number of bio's dispatched in current slice */
82 unsigned int io_disp[2];
83
84 /* When did we start a new slice */
85 unsigned long slice_start[2];
86 unsigned long slice_end[2];
87
88 /* Some throttle limits got updated for the group */
89 int limits_changed;
90
91 /* Per cpu stats pointer */
92 struct tg_stats_cpu __percpu *stats_cpu;
93
94 /* List of tgs waiting for per cpu stats memory to be allocated */
95 struct list_head stats_alloc_node;
96};
97
98struct throtl_data
99{
100 /* service tree for active throtl groups */
101 struct throtl_rb_root tg_service_tree;
102
103 struct request_queue *queue;
104
105 /* Total Number of queued bios on READ and WRITE lists */
106 unsigned int nr_queued[2];
107
108 /*
109 * number of total undestroyed groups
110 */
111 unsigned int nr_undestroyed_grps;
112
113 /* Work for dispatching throttled bios */
114 struct delayed_work throtl_work;
115
116 int limits_changed;
117};
118
119/* list and work item to allocate percpu group stats */
120static DEFINE_SPINLOCK(tg_stats_alloc_lock);
121static LIST_HEAD(tg_stats_alloc_list);
122
123static void tg_stats_alloc_fn(struct work_struct *);
124static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn);
125
126static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
127{
128 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
129}
130
131static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
132{
133 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
134}
135
136static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
137{
138 return pd_to_blkg(&tg->pd);
139}
140
141static inline struct throtl_grp *td_root_tg(struct throtl_data *td)
142{
143 return blkg_to_tg(td->queue->root_blkg);
144}
145
146enum tg_state_flags {
147 THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */
148};
149
150#define THROTL_TG_FNS(name) \
151static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \
152{ \
153 (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \
154} \
155static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \
156{ \
157 (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \
158} \
159static inline int throtl_tg_##name(const struct throtl_grp *tg) \
160{ \
161 return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \
162}
163
164THROTL_TG_FNS(on_rr);
165
166#define throtl_log_tg(td, tg, fmt, args...) do { \
167 char __pbuf[128]; \
168 \
169 blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \
170 blk_add_trace_msg((td)->queue, "throtl %s " fmt, __pbuf, ##args); \
171} while (0)
172
173#define throtl_log(td, fmt, args...) \
174 blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
175
176static inline unsigned int total_nr_queued(struct throtl_data *td)
177{
178 return td->nr_queued[0] + td->nr_queued[1];
179}
180
181/*
182 * Worker for allocating per cpu stat for tgs. This is scheduled on the
183 * system_nrt_wq once there are some groups on the alloc_list waiting for
184 * allocation.
185 */
186static void tg_stats_alloc_fn(struct work_struct *work)
187{
188 static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */
189 struct delayed_work *dwork = to_delayed_work(work);
190 bool empty = false;
191
192alloc_stats:
193 if (!stats_cpu) {
194 stats_cpu = alloc_percpu(struct tg_stats_cpu);
195 if (!stats_cpu) {
196 /* allocation failed, try again after some time */
197 queue_delayed_work(system_nrt_wq, dwork,
198 msecs_to_jiffies(10));
199 return;
200 }
201 }
202
203 spin_lock_irq(&tg_stats_alloc_lock);
204
205 if (!list_empty(&tg_stats_alloc_list)) {
206 struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list,
207 struct throtl_grp,
208 stats_alloc_node);
209 swap(tg->stats_cpu, stats_cpu);
210 list_del_init(&tg->stats_alloc_node);
211 }
212
213 empty = list_empty(&tg_stats_alloc_list);
214 spin_unlock_irq(&tg_stats_alloc_lock);
215 if (!empty)
216 goto alloc_stats;
217}
218
219static void throtl_pd_init(struct blkcg_gq *blkg)
220{
221 struct throtl_grp *tg = blkg_to_tg(blkg);
222 unsigned long flags;
223
224 RB_CLEAR_NODE(&tg->rb_node);
225 bio_list_init(&tg->bio_lists[0]);
226 bio_list_init(&tg->bio_lists[1]);
227 tg->limits_changed = false;
228
229 tg->bps[READ] = -1;
230 tg->bps[WRITE] = -1;
231 tg->iops[READ] = -1;
232 tg->iops[WRITE] = -1;
233
234 /*
235 * Ugh... We need to perform per-cpu allocation for tg->stats_cpu
236 * but percpu allocator can't be called from IO path. Queue tg on
237 * tg_stats_alloc_list and allocate from work item.
238 */
239 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
240 list_add(&tg->stats_alloc_node, &tg_stats_alloc_list);
241 queue_delayed_work(system_nrt_wq, &tg_stats_alloc_work, 0);
242 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
243}
244
245static void throtl_pd_exit(struct blkcg_gq *blkg)
246{
247 struct throtl_grp *tg = blkg_to_tg(blkg);
248 unsigned long flags;
249
250 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
251 list_del_init(&tg->stats_alloc_node);
252 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
253
254 free_percpu(tg->stats_cpu);
255}
256
257static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
258{
259 struct throtl_grp *tg = blkg_to_tg(blkg);
260 int cpu;
261
262 if (tg->stats_cpu == NULL)
263 return;
264
265 for_each_possible_cpu(cpu) {
266 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
267
268 blkg_rwstat_reset(&sc->service_bytes);
269 blkg_rwstat_reset(&sc->serviced);
270 }
271}
272
273static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
274 struct blkcg *blkcg)
275{
276 /*
277 * This is the common case when there are no blkcgs. Avoid lookup
278 * in this case
279 */
280 if (blkcg == &blkcg_root)
281 return td_root_tg(td);
282
283 return blkg_to_tg(blkg_lookup(blkcg, td->queue));
284}
285
286static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
287 struct blkcg *blkcg)
288{
289 struct request_queue *q = td->queue;
290 struct throtl_grp *tg = NULL;
291
292 /*
293 * This is the common case when there are no blkcgs. Avoid lookup
294 * in this case
295 */
296 if (blkcg == &blkcg_root) {
297 tg = td_root_tg(td);
298 } else {
299 struct blkcg_gq *blkg;
300
301 blkg = blkg_lookup_create(blkcg, q);
302
303 /* if %NULL and @q is alive, fall back to root_tg */
304 if (!IS_ERR(blkg))
305 tg = blkg_to_tg(blkg);
306 else if (!blk_queue_dead(q))
307 tg = td_root_tg(td);
308 }
309
310 return tg;
311}
312
313static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root)
314{
315 /* Service tree is empty */
316 if (!root->count)
317 return NULL;
318
319 if (!root->left)
320 root->left = rb_first(&root->rb);
321
322 if (root->left)
323 return rb_entry_tg(root->left);
324
325 return NULL;
326}
327
328static void rb_erase_init(struct rb_node *n, struct rb_root *root)
329{
330 rb_erase(n, root);
331 RB_CLEAR_NODE(n);
332}
333
334static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root)
335{
336 if (root->left == n)
337 root->left = NULL;
338 rb_erase_init(n, &root->rb);
339 --root->count;
340}
341
342static void update_min_dispatch_time(struct throtl_rb_root *st)
343{
344 struct throtl_grp *tg;
345
346 tg = throtl_rb_first(st);
347 if (!tg)
348 return;
349
350 st->min_disptime = tg->disptime;
351}
352
353static void
354tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg)
355{
356 struct rb_node **node = &st->rb.rb_node;
357 struct rb_node *parent = NULL;
358 struct throtl_grp *__tg;
359 unsigned long key = tg->disptime;
360 int left = 1;
361
362 while (*node != NULL) {
363 parent = *node;
364 __tg = rb_entry_tg(parent);
365
366 if (time_before(key, __tg->disptime))
367 node = &parent->rb_left;
368 else {
369 node = &parent->rb_right;
370 left = 0;
371 }
372 }
373
374 if (left)
375 st->left = &tg->rb_node;
376
377 rb_link_node(&tg->rb_node, parent, node);
378 rb_insert_color(&tg->rb_node, &st->rb);
379}
380
381static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
382{
383 struct throtl_rb_root *st = &td->tg_service_tree;
384
385 tg_service_tree_add(st, tg);
386 throtl_mark_tg_on_rr(tg);
387 st->count++;
388}
389
390static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
391{
392 if (!throtl_tg_on_rr(tg))
393 __throtl_enqueue_tg(td, tg);
394}
395
396static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
397{
398 throtl_rb_erase(&tg->rb_node, &td->tg_service_tree);
399 throtl_clear_tg_on_rr(tg);
400}
401
402static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
403{
404 if (throtl_tg_on_rr(tg))
405 __throtl_dequeue_tg(td, tg);
406}
407
408static void throtl_schedule_next_dispatch(struct throtl_data *td)
409{
410 struct throtl_rb_root *st = &td->tg_service_tree;
411
412 /*
413 * If there are more bios pending, schedule more work.
414 */
415 if (!total_nr_queued(td))
416 return;
417
418 BUG_ON(!st->count);
419
420 update_min_dispatch_time(st);
421
422 if (time_before_eq(st->min_disptime, jiffies))
423 throtl_schedule_delayed_work(td, 0);
424 else
425 throtl_schedule_delayed_work(td, (st->min_disptime - jiffies));
426}
427
428static inline void
429throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
430{
431 tg->bytes_disp[rw] = 0;
432 tg->io_disp[rw] = 0;
433 tg->slice_start[rw] = jiffies;
434 tg->slice_end[rw] = jiffies + throtl_slice;
435 throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
436 rw == READ ? 'R' : 'W', tg->slice_start[rw],
437 tg->slice_end[rw], jiffies);
438}
439
440static inline void throtl_set_slice_end(struct throtl_data *td,
441 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
442{
443 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
444}
445
446static inline void throtl_extend_slice(struct throtl_data *td,
447 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
448{
449 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
450 throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
451 rw == READ ? 'R' : 'W', tg->slice_start[rw],
452 tg->slice_end[rw], jiffies);
453}
454
455/* Determine if previously allocated or extended slice is complete or not */
456static bool
457throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
458{
459 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
460 return 0;
461
462 return 1;
463}
464
465/* Trim the used slices and adjust slice start accordingly */
466static inline void
467throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
468{
469 unsigned long nr_slices, time_elapsed, io_trim;
470 u64 bytes_trim, tmp;
471
472 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
473
474 /*
475 * If bps are unlimited (-1), then time slice don't get
476 * renewed. Don't try to trim the slice if slice is used. A new
477 * slice will start when appropriate.
478 */
479 if (throtl_slice_used(td, tg, rw))
480 return;
481
482 /*
483 * A bio has been dispatched. Also adjust slice_end. It might happen
484 * that initially cgroup limit was very low resulting in high
485 * slice_end, but later limit was bumped up and bio was dispached
486 * sooner, then we need to reduce slice_end. A high bogus slice_end
487 * is bad because it does not allow new slice to start.
488 */
489
490 throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice);
491
492 time_elapsed = jiffies - tg->slice_start[rw];
493
494 nr_slices = time_elapsed / throtl_slice;
495
496 if (!nr_slices)
497 return;
498 tmp = tg->bps[rw] * throtl_slice * nr_slices;
499 do_div(tmp, HZ);
500 bytes_trim = tmp;
501
502 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
503
504 if (!bytes_trim && !io_trim)
505 return;
506
507 if (tg->bytes_disp[rw] >= bytes_trim)
508 tg->bytes_disp[rw] -= bytes_trim;
509 else
510 tg->bytes_disp[rw] = 0;
511
512 if (tg->io_disp[rw] >= io_trim)
513 tg->io_disp[rw] -= io_trim;
514 else
515 tg->io_disp[rw] = 0;
516
517 tg->slice_start[rw] += nr_slices * throtl_slice;
518
519 throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
520 " start=%lu end=%lu jiffies=%lu",
521 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
522 tg->slice_start[rw], tg->slice_end[rw], jiffies);
523}
524
525static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
526 struct bio *bio, unsigned long *wait)
527{
528 bool rw = bio_data_dir(bio);
529 unsigned int io_allowed;
530 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
531 u64 tmp;
532
533 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
534
535 /* Slice has just started. Consider one slice interval */
536 if (!jiffy_elapsed)
537 jiffy_elapsed_rnd = throtl_slice;
538
539 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
540
541 /*
542 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
543 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
544 * will allow dispatch after 1 second and after that slice should
545 * have been trimmed.
546 */
547
548 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
549 do_div(tmp, HZ);
550
551 if (tmp > UINT_MAX)
552 io_allowed = UINT_MAX;
553 else
554 io_allowed = tmp;
555
556 if (tg->io_disp[rw] + 1 <= io_allowed) {
557 if (wait)
558 *wait = 0;
559 return 1;
560 }
561
562 /* Calc approx time to dispatch */
563 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
564
565 if (jiffy_wait > jiffy_elapsed)
566 jiffy_wait = jiffy_wait - jiffy_elapsed;
567 else
568 jiffy_wait = 1;
569
570 if (wait)
571 *wait = jiffy_wait;
572 return 0;
573}
574
575static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
576 struct bio *bio, unsigned long *wait)
577{
578 bool rw = bio_data_dir(bio);
579 u64 bytes_allowed, extra_bytes, tmp;
580 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
581
582 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
583
584 /* Slice has just started. Consider one slice interval */
585 if (!jiffy_elapsed)
586 jiffy_elapsed_rnd = throtl_slice;
587
588 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
589
590 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
591 do_div(tmp, HZ);
592 bytes_allowed = tmp;
593
594 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
595 if (wait)
596 *wait = 0;
597 return 1;
598 }
599
600 /* Calc approx time to dispatch */
601 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
602 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
603
604 if (!jiffy_wait)
605 jiffy_wait = 1;
606
607 /*
608 * This wait time is without taking into consideration the rounding
609 * up we did. Add that time also.
610 */
611 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
612 if (wait)
613 *wait = jiffy_wait;
614 return 0;
615}
616
617static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
618 if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
619 return 1;
620 return 0;
621}
622
623/*
624 * Returns whether one can dispatch a bio or not. Also returns approx number
625 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
626 */
627static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
628 struct bio *bio, unsigned long *wait)
629{
630 bool rw = bio_data_dir(bio);
631 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
632
633 /*
634 * Currently whole state machine of group depends on first bio
635 * queued in the group bio list. So one should not be calling
636 * this function with a different bio if there are other bios
637 * queued.
638 */
639 BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
640
641 /* If tg->bps = -1, then BW is unlimited */
642 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
643 if (wait)
644 *wait = 0;
645 return 1;
646 }
647
648 /*
649 * If previous slice expired, start a new one otherwise renew/extend
650 * existing slice to make sure it is at least throtl_slice interval
651 * long since now.
652 */
653 if (throtl_slice_used(td, tg, rw))
654 throtl_start_new_slice(td, tg, rw);
655 else {
656 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
657 throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
658 }
659
660 if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
661 && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
662 if (wait)
663 *wait = 0;
664 return 1;
665 }
666
667 max_wait = max(bps_wait, iops_wait);
668
669 if (wait)
670 *wait = max_wait;
671
672 if (time_before(tg->slice_end[rw], jiffies + max_wait))
673 throtl_extend_slice(td, tg, rw, jiffies + max_wait);
674
675 return 0;
676}
677
678static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
679 int rw)
680{
681 struct throtl_grp *tg = blkg_to_tg(blkg);
682 struct tg_stats_cpu *stats_cpu;
683 unsigned long flags;
684
685 /* If per cpu stats are not allocated yet, don't do any accounting. */
686 if (tg->stats_cpu == NULL)
687 return;
688
689 /*
690 * Disabling interrupts to provide mutual exclusion between two
691 * writes on same cpu. It probably is not needed for 64bit. Not
692 * optimizing that case yet.
693 */
694 local_irq_save(flags);
695
696 stats_cpu = this_cpu_ptr(tg->stats_cpu);
697
698 blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
699 blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
700
701 local_irq_restore(flags);
702}
703
704static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
705{
706 bool rw = bio_data_dir(bio);
707
708 /* Charge the bio to the group */
709 tg->bytes_disp[rw] += bio->bi_size;
710 tg->io_disp[rw]++;
711
712 throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw);
713}
714
715static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
716 struct bio *bio)
717{
718 bool rw = bio_data_dir(bio);
719
720 bio_list_add(&tg->bio_lists[rw], bio);
721 /* Take a bio reference on tg */
722 blkg_get(tg_to_blkg(tg));
723 tg->nr_queued[rw]++;
724 td->nr_queued[rw]++;
725 throtl_enqueue_tg(td, tg);
726}
727
728static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
729{
730 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
731 struct bio *bio;
732
733 if ((bio = bio_list_peek(&tg->bio_lists[READ])))
734 tg_may_dispatch(td, tg, bio, &read_wait);
735
736 if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
737 tg_may_dispatch(td, tg, bio, &write_wait);
738
739 min_wait = min(read_wait, write_wait);
740 disptime = jiffies + min_wait;
741
742 /* Update dispatch time */
743 throtl_dequeue_tg(td, tg);
744 tg->disptime = disptime;
745 throtl_enqueue_tg(td, tg);
746}
747
748static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
749 bool rw, struct bio_list *bl)
750{
751 struct bio *bio;
752
753 bio = bio_list_pop(&tg->bio_lists[rw]);
754 tg->nr_queued[rw]--;
755 /* Drop bio reference on blkg */
756 blkg_put(tg_to_blkg(tg));
757
758 BUG_ON(td->nr_queued[rw] <= 0);
759 td->nr_queued[rw]--;
760
761 throtl_charge_bio(tg, bio);
762 bio_list_add(bl, bio);
763 bio->bi_rw |= REQ_THROTTLED;
764
765 throtl_trim_slice(td, tg, rw);
766}
767
768static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
769 struct bio_list *bl)
770{
771 unsigned int nr_reads = 0, nr_writes = 0;
772 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
773 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
774 struct bio *bio;
775
776 /* Try to dispatch 75% READS and 25% WRITES */
777
778 while ((bio = bio_list_peek(&tg->bio_lists[READ]))
779 && tg_may_dispatch(td, tg, bio, NULL)) {
780
781 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
782 nr_reads++;
783
784 if (nr_reads >= max_nr_reads)
785 break;
786 }
787
788 while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
789 && tg_may_dispatch(td, tg, bio, NULL)) {
790
791 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
792 nr_writes++;
793
794 if (nr_writes >= max_nr_writes)
795 break;
796 }
797
798 return nr_reads + nr_writes;
799}
800
801static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
802{
803 unsigned int nr_disp = 0;
804 struct throtl_grp *tg;
805 struct throtl_rb_root *st = &td->tg_service_tree;
806
807 while (1) {
808 tg = throtl_rb_first(st);
809
810 if (!tg)
811 break;
812
813 if (time_before(jiffies, tg->disptime))
814 break;
815
816 throtl_dequeue_tg(td, tg);
817
818 nr_disp += throtl_dispatch_tg(td, tg, bl);
819
820 if (tg->nr_queued[0] || tg->nr_queued[1]) {
821 tg_update_disptime(td, tg);
822 throtl_enqueue_tg(td, tg);
823 }
824
825 if (nr_disp >= throtl_quantum)
826 break;
827 }
828
829 return nr_disp;
830}
831
832static void throtl_process_limit_change(struct throtl_data *td)
833{
834 struct request_queue *q = td->queue;
835 struct blkcg_gq *blkg, *n;
836
837 if (!td->limits_changed)
838 return;
839
840 xchg(&td->limits_changed, false);
841
842 throtl_log(td, "limits changed");
843
844 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
845 struct throtl_grp *tg = blkg_to_tg(blkg);
846
847 if (!tg->limits_changed)
848 continue;
849
850 if (!xchg(&tg->limits_changed, false))
851 continue;
852
853 throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
854 " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
855 tg->iops[READ], tg->iops[WRITE]);
856
857 /*
858 * Restart the slices for both READ and WRITES. It
859 * might happen that a group's limit are dropped
860 * suddenly and we don't want to account recently
861 * dispatched IO with new low rate
862 */
863 throtl_start_new_slice(td, tg, 0);
864 throtl_start_new_slice(td, tg, 1);
865
866 if (throtl_tg_on_rr(tg))
867 tg_update_disptime(td, tg);
868 }
869}
870
871/* Dispatch throttled bios. Should be called without queue lock held. */
872static int throtl_dispatch(struct request_queue *q)
873{
874 struct throtl_data *td = q->td;
875 unsigned int nr_disp = 0;
876 struct bio_list bio_list_on_stack;
877 struct bio *bio;
878 struct blk_plug plug;
879
880 spin_lock_irq(q->queue_lock);
881
882 throtl_process_limit_change(td);
883
884 if (!total_nr_queued(td))
885 goto out;
886
887 bio_list_init(&bio_list_on_stack);
888
889 throtl_log(td, "dispatch nr_queued=%u read=%u write=%u",
890 total_nr_queued(td), td->nr_queued[READ],
891 td->nr_queued[WRITE]);
892
893 nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
894
895 if (nr_disp)
896 throtl_log(td, "bios disp=%u", nr_disp);
897
898 throtl_schedule_next_dispatch(td);
899out:
900 spin_unlock_irq(q->queue_lock);
901
902 /*
903 * If we dispatched some requests, unplug the queue to make sure
904 * immediate dispatch
905 */
906 if (nr_disp) {
907 blk_start_plug(&plug);
908 while((bio = bio_list_pop(&bio_list_on_stack)))
909 generic_make_request(bio);
910 blk_finish_plug(&plug);
911 }
912 return nr_disp;
913}
914
915void blk_throtl_work(struct work_struct *work)
916{
917 struct throtl_data *td = container_of(work, struct throtl_data,
918 throtl_work.work);
919 struct request_queue *q = td->queue;
920
921 throtl_dispatch(q);
922}
923
924/* Call with queue lock held */
925static void
926throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay)
927{
928
929 struct delayed_work *dwork = &td->throtl_work;
930
931 /* schedule work if limits changed even if no bio is queued */
932 if (total_nr_queued(td) || td->limits_changed) {
933 /*
934 * We might have a work scheduled to be executed in future.
935 * Cancel that and schedule a new one.
936 */
937 __cancel_delayed_work(dwork);
938 queue_delayed_work(kthrotld_workqueue, dwork, delay);
939 throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
940 delay, jiffies);
941 }
942}
943
944static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
945 struct blkg_policy_data *pd, int off)
946{
947 struct throtl_grp *tg = pd_to_tg(pd);
948 struct blkg_rwstat rwstat = { }, tmp;
949 int i, cpu;
950
951 for_each_possible_cpu(cpu) {
952 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
953
954 tmp = blkg_rwstat_read((void *)sc + off);
955 for (i = 0; i < BLKG_RWSTAT_NR; i++)
956 rwstat.cnt[i] += tmp.cnt[i];
957 }
958
959 return __blkg_prfill_rwstat(sf, pd, &rwstat);
960}
961
962static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft,
963 struct seq_file *sf)
964{
965 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
966
967 blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl,
968 cft->private, true);
969 return 0;
970}
971
972static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
973 int off)
974{
975 struct throtl_grp *tg = pd_to_tg(pd);
976 u64 v = *(u64 *)((void *)tg + off);
977
978 if (v == -1)
979 return 0;
980 return __blkg_prfill_u64(sf, pd, v);
981}
982
983static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
984 int off)
985{
986 struct throtl_grp *tg = pd_to_tg(pd);
987 unsigned int v = *(unsigned int *)((void *)tg + off);
988
989 if (v == -1)
990 return 0;
991 return __blkg_prfill_u64(sf, pd, v);
992}
993
994static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft,
995 struct seq_file *sf)
996{
997 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64,
998 &blkcg_policy_throtl, cft->private, false);
999 return 0;
1000}
1001
1002static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1003 struct seq_file *sf)
1004{
1005 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint,
1006 &blkcg_policy_throtl, cft->private, false);
1007 return 0;
1008}
1009
1010static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf,
1011 bool is_u64)
1012{
1013 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1014 struct blkg_conf_ctx ctx;
1015 struct throtl_grp *tg;
1016 struct throtl_data *td;
1017 int ret;
1018
1019 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1020 if (ret)
1021 return ret;
1022
1023 tg = blkg_to_tg(ctx.blkg);
1024 td = ctx.blkg->q->td;
1025
1026 if (!ctx.v)
1027 ctx.v = -1;
1028
1029 if (is_u64)
1030 *(u64 *)((void *)tg + cft->private) = ctx.v;
1031 else
1032 *(unsigned int *)((void *)tg + cft->private) = ctx.v;
1033
1034 /* XXX: we don't need the following deferred processing */
1035 xchg(&tg->limits_changed, true);
1036 xchg(&td->limits_changed, true);
1037 throtl_schedule_delayed_work(td, 0);
1038
1039 blkg_conf_finish(&ctx);
1040 return 0;
1041}
1042
1043static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft,
1044 const char *buf)
1045{
1046 return tg_set_conf(cgrp, cft, buf, true);
1047}
1048
1049static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1050 const char *buf)
1051{
1052 return tg_set_conf(cgrp, cft, buf, false);
1053}
1054
1055static struct cftype throtl_files[] = {
1056 {
1057 .name = "throttle.read_bps_device",
1058 .private = offsetof(struct throtl_grp, bps[READ]),
1059 .read_seq_string = tg_print_conf_u64,
1060 .write_string = tg_set_conf_u64,
1061 .max_write_len = 256,
1062 },
1063 {
1064 .name = "throttle.write_bps_device",
1065 .private = offsetof(struct throtl_grp, bps[WRITE]),
1066 .read_seq_string = tg_print_conf_u64,
1067 .write_string = tg_set_conf_u64,
1068 .max_write_len = 256,
1069 },
1070 {
1071 .name = "throttle.read_iops_device",
1072 .private = offsetof(struct throtl_grp, iops[READ]),
1073 .read_seq_string = tg_print_conf_uint,
1074 .write_string = tg_set_conf_uint,
1075 .max_write_len = 256,
1076 },
1077 {
1078 .name = "throttle.write_iops_device",
1079 .private = offsetof(struct throtl_grp, iops[WRITE]),
1080 .read_seq_string = tg_print_conf_uint,
1081 .write_string = tg_set_conf_uint,
1082 .max_write_len = 256,
1083 },
1084 {
1085 .name = "throttle.io_service_bytes",
1086 .private = offsetof(struct tg_stats_cpu, service_bytes),
1087 .read_seq_string = tg_print_cpu_rwstat,
1088 },
1089 {
1090 .name = "throttle.io_serviced",
1091 .private = offsetof(struct tg_stats_cpu, serviced),
1092 .read_seq_string = tg_print_cpu_rwstat,
1093 },
1094 { } /* terminate */
1095};
1096
1097static void throtl_shutdown_wq(struct request_queue *q)
1098{
1099 struct throtl_data *td = q->td;
1100
1101 cancel_delayed_work_sync(&td->throtl_work);
1102}
1103
1104static struct blkcg_policy blkcg_policy_throtl = {
1105 .pd_size = sizeof(struct throtl_grp),
1106 .cftypes = throtl_files,
1107
1108 .pd_init_fn = throtl_pd_init,
1109 .pd_exit_fn = throtl_pd_exit,
1110 .pd_reset_stats_fn = throtl_pd_reset_stats,
1111};
1112
1113bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
1114{
1115 struct throtl_data *td = q->td;
1116 struct throtl_grp *tg;
1117 bool rw = bio_data_dir(bio), update_disptime = true;
1118 struct blkcg *blkcg;
1119 bool throttled = false;
1120
1121 if (bio->bi_rw & REQ_THROTTLED) {
1122 bio->bi_rw &= ~REQ_THROTTLED;
1123 goto out;
1124 }
1125
1126 /* bio_associate_current() needs ioc, try creating */
1127 create_io_context(GFP_ATOMIC, q->node);
1128
1129 /*
1130 * A throtl_grp pointer retrieved under rcu can be used to access
1131 * basic fields like stats and io rates. If a group has no rules,
1132 * just update the dispatch stats in lockless manner and return.
1133 */
1134 rcu_read_lock();
1135 blkcg = bio_blkcg(bio);
1136 tg = throtl_lookup_tg(td, blkcg);
1137 if (tg) {
1138 if (tg_no_rule_group(tg, rw)) {
1139 throtl_update_dispatch_stats(tg_to_blkg(tg),
1140 bio->bi_size, bio->bi_rw);
1141 goto out_unlock_rcu;
1142 }
1143 }
1144
1145 /*
1146 * Either group has not been allocated yet or it is not an unlimited
1147 * IO group
1148 */
1149 spin_lock_irq(q->queue_lock);
1150 tg = throtl_lookup_create_tg(td, blkcg);
1151 if (unlikely(!tg))
1152 goto out_unlock;
1153
1154 if (tg->nr_queued[rw]) {
1155 /*
1156 * There is already another bio queued in same dir. No
1157 * need to update dispatch time.
1158 */
1159 update_disptime = false;
1160 goto queue_bio;
1161
1162 }
1163
1164 /* Bio is with-in rate limit of group */
1165 if (tg_may_dispatch(td, tg, bio, NULL)) {
1166 throtl_charge_bio(tg, bio);
1167
1168 /*
1169 * We need to trim slice even when bios are not being queued
1170 * otherwise it might happen that a bio is not queued for
1171 * a long time and slice keeps on extending and trim is not
1172 * called for a long time. Now if limits are reduced suddenly
1173 * we take into account all the IO dispatched so far at new
1174 * low rate and * newly queued IO gets a really long dispatch
1175 * time.
1176 *
1177 * So keep on trimming slice even if bio is not queued.
1178 */
1179 throtl_trim_slice(td, tg, rw);
1180 goto out_unlock;
1181 }
1182
1183queue_bio:
1184 throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
1185 " iodisp=%u iops=%u queued=%d/%d",
1186 rw == READ ? 'R' : 'W',
1187 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1188 tg->io_disp[rw], tg->iops[rw],
1189 tg->nr_queued[READ], tg->nr_queued[WRITE]);
1190
1191 bio_associate_current(bio);
1192 throtl_add_bio_tg(q->td, tg, bio);
1193 throttled = true;
1194
1195 if (update_disptime) {
1196 tg_update_disptime(td, tg);
1197 throtl_schedule_next_dispatch(td);
1198 }
1199
1200out_unlock:
1201 spin_unlock_irq(q->queue_lock);
1202out_unlock_rcu:
1203 rcu_read_unlock();
1204out:
1205 return throttled;
1206}
1207
1208/**
1209 * blk_throtl_drain - drain throttled bios
1210 * @q: request_queue to drain throttled bios for
1211 *
1212 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1213 */
1214void blk_throtl_drain(struct request_queue *q)
1215 __releases(q->queue_lock) __acquires(q->queue_lock)
1216{
1217 struct throtl_data *td = q->td;
1218 struct throtl_rb_root *st = &td->tg_service_tree;
1219 struct throtl_grp *tg;
1220 struct bio_list bl;
1221 struct bio *bio;
1222
1223 queue_lockdep_assert_held(q);
1224
1225 bio_list_init(&bl);
1226
1227 while ((tg = throtl_rb_first(st))) {
1228 throtl_dequeue_tg(td, tg);
1229
1230 while ((bio = bio_list_peek(&tg->bio_lists[READ])))
1231 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl);
1232 while ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
1233 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl);
1234 }
1235 spin_unlock_irq(q->queue_lock);
1236
1237 while ((bio = bio_list_pop(&bl)))
1238 generic_make_request(bio);
1239
1240 spin_lock_irq(q->queue_lock);
1241}
1242
1243int blk_throtl_init(struct request_queue *q)
1244{
1245 struct throtl_data *td;
1246 int ret;
1247
1248 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1249 if (!td)
1250 return -ENOMEM;
1251
1252 td->tg_service_tree = THROTL_RB_ROOT;
1253 td->limits_changed = false;
1254 INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);
1255
1256 q->td = td;
1257 td->queue = q;
1258
1259 /* activate policy */
1260 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1261 if (ret)
1262 kfree(td);
1263 return ret;
1264}
1265
1266void blk_throtl_exit(struct request_queue *q)
1267{
1268 BUG_ON(!q->td);
1269 throtl_shutdown_wq(q);
1270 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1271 kfree(q->td);
1272}
1273
1274static int __init throtl_init(void)
1275{
1276 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1277 if (!kthrotld_workqueue)
1278 panic("Failed to create kthrotld\n");
1279
1280 return blkcg_policy_register(&blkcg_policy_throtl);
1281}
1282
1283module_init(throtl_init);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Interface for controlling IO bandwidth on a request queue
4 *
5 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
6 */
7
8#include <linux/module.h>
9#include <linux/slab.h>
10#include <linux/blkdev.h>
11#include <linux/bio.h>
12#include <linux/blktrace_api.h>
13#include "blk.h"
14#include "blk-cgroup-rwstat.h"
15#include "blk-stat.h"
16#include "blk-throttle.h"
17
18/* Max dispatch from a group in 1 round */
19#define THROTL_GRP_QUANTUM 8
20
21/* Total max dispatch from all groups in one round */
22#define THROTL_QUANTUM 32
23
24/* Throttling is performed over a slice and after that slice is renewed */
25#define DFL_THROTL_SLICE_HD (HZ / 10)
26#define DFL_THROTL_SLICE_SSD (HZ / 50)
27#define MAX_THROTL_SLICE (HZ)
28
29/* A workqueue to queue throttle related work */
30static struct workqueue_struct *kthrotld_workqueue;
31
32#define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
33
34struct throtl_data
35{
36 /* service tree for active throtl groups */
37 struct throtl_service_queue service_queue;
38
39 struct request_queue *queue;
40
41 /* Total Number of queued bios on READ and WRITE lists */
42 unsigned int nr_queued[2];
43
44 unsigned int throtl_slice;
45
46 /* Work for dispatching throttled bios */
47 struct work_struct dispatch_work;
48
49 bool track_bio_latency;
50};
51
52static void throtl_pending_timer_fn(struct timer_list *t);
53
54static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
55{
56 return pd_to_blkg(&tg->pd);
57}
58
59/**
60 * sq_to_tg - return the throl_grp the specified service queue belongs to
61 * @sq: the throtl_service_queue of interest
62 *
63 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
64 * embedded in throtl_data, %NULL is returned.
65 */
66static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
67{
68 if (sq && sq->parent_sq)
69 return container_of(sq, struct throtl_grp, service_queue);
70 else
71 return NULL;
72}
73
74/**
75 * sq_to_td - return throtl_data the specified service queue belongs to
76 * @sq: the throtl_service_queue of interest
77 *
78 * A service_queue can be embedded in either a throtl_grp or throtl_data.
79 * Determine the associated throtl_data accordingly and return it.
80 */
81static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
82{
83 struct throtl_grp *tg = sq_to_tg(sq);
84
85 if (tg)
86 return tg->td;
87 else
88 return container_of(sq, struct throtl_data, service_queue);
89}
90
91static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
92{
93 struct blkcg_gq *blkg = tg_to_blkg(tg);
94
95 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
96 return U64_MAX;
97
98 return tg->bps[rw];
99}
100
101static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
102{
103 struct blkcg_gq *blkg = tg_to_blkg(tg);
104
105 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
106 return UINT_MAX;
107
108 return tg->iops[rw];
109}
110
111/**
112 * throtl_log - log debug message via blktrace
113 * @sq: the service_queue being reported
114 * @fmt: printf format string
115 * @args: printf args
116 *
117 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
118 * throtl_grp; otherwise, just "throtl".
119 */
120#define throtl_log(sq, fmt, args...) do { \
121 struct throtl_grp *__tg = sq_to_tg((sq)); \
122 struct throtl_data *__td = sq_to_td((sq)); \
123 \
124 (void)__td; \
125 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
126 break; \
127 if ((__tg)) { \
128 blk_add_cgroup_trace_msg(__td->queue, \
129 &tg_to_blkg(__tg)->blkcg->css, "throtl " fmt, ##args);\
130 } else { \
131 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
132 } \
133} while (0)
134
135static inline unsigned int throtl_bio_data_size(struct bio *bio)
136{
137 /* assume it's one sector */
138 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
139 return 512;
140 return bio->bi_iter.bi_size;
141}
142
143static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
144{
145 INIT_LIST_HEAD(&qn->node);
146 bio_list_init(&qn->bios);
147 qn->tg = tg;
148}
149
150/**
151 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
152 * @bio: bio being added
153 * @qn: qnode to add bio to
154 * @queued: the service_queue->queued[] list @qn belongs to
155 *
156 * Add @bio to @qn and put @qn on @queued if it's not already on.
157 * @qn->tg's reference count is bumped when @qn is activated. See the
158 * comment on top of throtl_qnode definition for details.
159 */
160static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
161 struct list_head *queued)
162{
163 bio_list_add(&qn->bios, bio);
164 if (list_empty(&qn->node)) {
165 list_add_tail(&qn->node, queued);
166 blkg_get(tg_to_blkg(qn->tg));
167 }
168}
169
170/**
171 * throtl_peek_queued - peek the first bio on a qnode list
172 * @queued: the qnode list to peek
173 */
174static struct bio *throtl_peek_queued(struct list_head *queued)
175{
176 struct throtl_qnode *qn;
177 struct bio *bio;
178
179 if (list_empty(queued))
180 return NULL;
181
182 qn = list_first_entry(queued, struct throtl_qnode, node);
183 bio = bio_list_peek(&qn->bios);
184 WARN_ON_ONCE(!bio);
185 return bio;
186}
187
188/**
189 * throtl_pop_queued - pop the first bio form a qnode list
190 * @queued: the qnode list to pop a bio from
191 * @tg_to_put: optional out argument for throtl_grp to put
192 *
193 * Pop the first bio from the qnode list @queued. After popping, the first
194 * qnode is removed from @queued if empty or moved to the end of @queued so
195 * that the popping order is round-robin.
196 *
197 * When the first qnode is removed, its associated throtl_grp should be put
198 * too. If @tg_to_put is NULL, this function automatically puts it;
199 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
200 * responsible for putting it.
201 */
202static struct bio *throtl_pop_queued(struct list_head *queued,
203 struct throtl_grp **tg_to_put)
204{
205 struct throtl_qnode *qn;
206 struct bio *bio;
207
208 if (list_empty(queued))
209 return NULL;
210
211 qn = list_first_entry(queued, struct throtl_qnode, node);
212 bio = bio_list_pop(&qn->bios);
213 WARN_ON_ONCE(!bio);
214
215 if (bio_list_empty(&qn->bios)) {
216 list_del_init(&qn->node);
217 if (tg_to_put)
218 *tg_to_put = qn->tg;
219 else
220 blkg_put(tg_to_blkg(qn->tg));
221 } else {
222 list_move_tail(&qn->node, queued);
223 }
224
225 return bio;
226}
227
228/* init a service_queue, assumes the caller zeroed it */
229static void throtl_service_queue_init(struct throtl_service_queue *sq)
230{
231 INIT_LIST_HEAD(&sq->queued[READ]);
232 INIT_LIST_HEAD(&sq->queued[WRITE]);
233 sq->pending_tree = RB_ROOT_CACHED;
234 timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
235}
236
237static struct blkg_policy_data *throtl_pd_alloc(struct gendisk *disk,
238 struct blkcg *blkcg, gfp_t gfp)
239{
240 struct throtl_grp *tg;
241 int rw;
242
243 tg = kzalloc_node(sizeof(*tg), gfp, disk->node_id);
244 if (!tg)
245 return NULL;
246
247 if (blkg_rwstat_init(&tg->stat_bytes, gfp))
248 goto err_free_tg;
249
250 if (blkg_rwstat_init(&tg->stat_ios, gfp))
251 goto err_exit_stat_bytes;
252
253 throtl_service_queue_init(&tg->service_queue);
254
255 for (rw = READ; rw <= WRITE; rw++) {
256 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
257 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
258 }
259
260 RB_CLEAR_NODE(&tg->rb_node);
261 tg->bps[READ] = U64_MAX;
262 tg->bps[WRITE] = U64_MAX;
263 tg->iops[READ] = UINT_MAX;
264 tg->iops[WRITE] = UINT_MAX;
265
266 return &tg->pd;
267
268err_exit_stat_bytes:
269 blkg_rwstat_exit(&tg->stat_bytes);
270err_free_tg:
271 kfree(tg);
272 return NULL;
273}
274
275static void throtl_pd_init(struct blkg_policy_data *pd)
276{
277 struct throtl_grp *tg = pd_to_tg(pd);
278 struct blkcg_gq *blkg = tg_to_blkg(tg);
279 struct throtl_data *td = blkg->q->td;
280 struct throtl_service_queue *sq = &tg->service_queue;
281
282 /*
283 * If on the default hierarchy, we switch to properly hierarchical
284 * behavior where limits on a given throtl_grp are applied to the
285 * whole subtree rather than just the group itself. e.g. If 16M
286 * read_bps limit is set on a parent group, summary bps of
287 * parent group and its subtree groups can't exceed 16M for the
288 * device.
289 *
290 * If not on the default hierarchy, the broken flat hierarchy
291 * behavior is retained where all throtl_grps are treated as if
292 * they're all separate root groups right below throtl_data.
293 * Limits of a group don't interact with limits of other groups
294 * regardless of the position of the group in the hierarchy.
295 */
296 sq->parent_sq = &td->service_queue;
297 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
298 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
299 tg->td = td;
300}
301
302/*
303 * Set has_rules[] if @tg or any of its parents have limits configured.
304 * This doesn't require walking up to the top of the hierarchy as the
305 * parent's has_rules[] is guaranteed to be correct.
306 */
307static void tg_update_has_rules(struct throtl_grp *tg)
308{
309 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
310 int rw;
311
312 for (rw = READ; rw <= WRITE; rw++) {
313 tg->has_rules_iops[rw] =
314 (parent_tg && parent_tg->has_rules_iops[rw]) ||
315 tg_iops_limit(tg, rw) != UINT_MAX;
316 tg->has_rules_bps[rw] =
317 (parent_tg && parent_tg->has_rules_bps[rw]) ||
318 tg_bps_limit(tg, rw) != U64_MAX;
319 }
320}
321
322static void throtl_pd_online(struct blkg_policy_data *pd)
323{
324 struct throtl_grp *tg = pd_to_tg(pd);
325 /*
326 * We don't want new groups to escape the limits of its ancestors.
327 * Update has_rules[] after a new group is brought online.
328 */
329 tg_update_has_rules(tg);
330}
331
332static void throtl_pd_free(struct blkg_policy_data *pd)
333{
334 struct throtl_grp *tg = pd_to_tg(pd);
335
336 del_timer_sync(&tg->service_queue.pending_timer);
337 blkg_rwstat_exit(&tg->stat_bytes);
338 blkg_rwstat_exit(&tg->stat_ios);
339 kfree(tg);
340}
341
342static struct throtl_grp *
343throtl_rb_first(struct throtl_service_queue *parent_sq)
344{
345 struct rb_node *n;
346
347 n = rb_first_cached(&parent_sq->pending_tree);
348 WARN_ON_ONCE(!n);
349 if (!n)
350 return NULL;
351 return rb_entry_tg(n);
352}
353
354static void throtl_rb_erase(struct rb_node *n,
355 struct throtl_service_queue *parent_sq)
356{
357 rb_erase_cached(n, &parent_sq->pending_tree);
358 RB_CLEAR_NODE(n);
359}
360
361static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
362{
363 struct throtl_grp *tg;
364
365 tg = throtl_rb_first(parent_sq);
366 if (!tg)
367 return;
368
369 parent_sq->first_pending_disptime = tg->disptime;
370}
371
372static void tg_service_queue_add(struct throtl_grp *tg)
373{
374 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
375 struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
376 struct rb_node *parent = NULL;
377 struct throtl_grp *__tg;
378 unsigned long key = tg->disptime;
379 bool leftmost = true;
380
381 while (*node != NULL) {
382 parent = *node;
383 __tg = rb_entry_tg(parent);
384
385 if (time_before(key, __tg->disptime))
386 node = &parent->rb_left;
387 else {
388 node = &parent->rb_right;
389 leftmost = false;
390 }
391 }
392
393 rb_link_node(&tg->rb_node, parent, node);
394 rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
395 leftmost);
396}
397
398static void throtl_enqueue_tg(struct throtl_grp *tg)
399{
400 if (!(tg->flags & THROTL_TG_PENDING)) {
401 tg_service_queue_add(tg);
402 tg->flags |= THROTL_TG_PENDING;
403 tg->service_queue.parent_sq->nr_pending++;
404 }
405}
406
407static void throtl_dequeue_tg(struct throtl_grp *tg)
408{
409 if (tg->flags & THROTL_TG_PENDING) {
410 struct throtl_service_queue *parent_sq =
411 tg->service_queue.parent_sq;
412
413 throtl_rb_erase(&tg->rb_node, parent_sq);
414 --parent_sq->nr_pending;
415 tg->flags &= ~THROTL_TG_PENDING;
416 }
417}
418
419/* Call with queue lock held */
420static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
421 unsigned long expires)
422{
423 unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
424
425 /*
426 * Since we are adjusting the throttle limit dynamically, the sleep
427 * time calculated according to previous limit might be invalid. It's
428 * possible the cgroup sleep time is very long and no other cgroups
429 * have IO running so notify the limit changes. Make sure the cgroup
430 * doesn't sleep too long to avoid the missed notification.
431 */
432 if (time_after(expires, max_expire))
433 expires = max_expire;
434 mod_timer(&sq->pending_timer, expires);
435 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
436 expires - jiffies, jiffies);
437}
438
439/**
440 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
441 * @sq: the service_queue to schedule dispatch for
442 * @force: force scheduling
443 *
444 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
445 * dispatch time of the first pending child. Returns %true if either timer
446 * is armed or there's no pending child left. %false if the current
447 * dispatch window is still open and the caller should continue
448 * dispatching.
449 *
450 * If @force is %true, the dispatch timer is always scheduled and this
451 * function is guaranteed to return %true. This is to be used when the
452 * caller can't dispatch itself and needs to invoke pending_timer
453 * unconditionally. Note that forced scheduling is likely to induce short
454 * delay before dispatch starts even if @sq->first_pending_disptime is not
455 * in the future and thus shouldn't be used in hot paths.
456 */
457static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
458 bool force)
459{
460 /* any pending children left? */
461 if (!sq->nr_pending)
462 return true;
463
464 update_min_dispatch_time(sq);
465
466 /* is the next dispatch time in the future? */
467 if (force || time_after(sq->first_pending_disptime, jiffies)) {
468 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
469 return true;
470 }
471
472 /* tell the caller to continue dispatching */
473 return false;
474}
475
476static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
477 bool rw, unsigned long start)
478{
479 tg->bytes_disp[rw] = 0;
480 tg->io_disp[rw] = 0;
481 tg->carryover_bytes[rw] = 0;
482 tg->carryover_ios[rw] = 0;
483
484 /*
485 * Previous slice has expired. We must have trimmed it after last
486 * bio dispatch. That means since start of last slice, we never used
487 * that bandwidth. Do try to make use of that bandwidth while giving
488 * credit.
489 */
490 if (time_after(start, tg->slice_start[rw]))
491 tg->slice_start[rw] = start;
492
493 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
494 throtl_log(&tg->service_queue,
495 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
496 rw == READ ? 'R' : 'W', tg->slice_start[rw],
497 tg->slice_end[rw], jiffies);
498}
499
500static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw,
501 bool clear_carryover)
502{
503 tg->bytes_disp[rw] = 0;
504 tg->io_disp[rw] = 0;
505 tg->slice_start[rw] = jiffies;
506 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
507 if (clear_carryover) {
508 tg->carryover_bytes[rw] = 0;
509 tg->carryover_ios[rw] = 0;
510 }
511
512 throtl_log(&tg->service_queue,
513 "[%c] new slice start=%lu end=%lu jiffies=%lu",
514 rw == READ ? 'R' : 'W', tg->slice_start[rw],
515 tg->slice_end[rw], jiffies);
516}
517
518static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
519 unsigned long jiffy_end)
520{
521 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
522}
523
524static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
525 unsigned long jiffy_end)
526{
527 throtl_set_slice_end(tg, rw, jiffy_end);
528 throtl_log(&tg->service_queue,
529 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
530 rw == READ ? 'R' : 'W', tg->slice_start[rw],
531 tg->slice_end[rw], jiffies);
532}
533
534/* Determine if previously allocated or extended slice is complete or not */
535static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
536{
537 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
538 return false;
539
540 return true;
541}
542
543static unsigned int calculate_io_allowed(u32 iops_limit,
544 unsigned long jiffy_elapsed)
545{
546 unsigned int io_allowed;
547 u64 tmp;
548
549 /*
550 * jiffy_elapsed should not be a big value as minimum iops can be
551 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
552 * will allow dispatch after 1 second and after that slice should
553 * have been trimmed.
554 */
555
556 tmp = (u64)iops_limit * jiffy_elapsed;
557 do_div(tmp, HZ);
558
559 if (tmp > UINT_MAX)
560 io_allowed = UINT_MAX;
561 else
562 io_allowed = tmp;
563
564 return io_allowed;
565}
566
567static u64 calculate_bytes_allowed(u64 bps_limit, unsigned long jiffy_elapsed)
568{
569 /*
570 * Can result be wider than 64 bits?
571 * We check against 62, not 64, due to ilog2 truncation.
572 */
573 if (ilog2(bps_limit) + ilog2(jiffy_elapsed) - ilog2(HZ) > 62)
574 return U64_MAX;
575 return mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed, (u64)HZ);
576}
577
578/* Trim the used slices and adjust slice start accordingly */
579static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
580{
581 unsigned long time_elapsed;
582 long long bytes_trim;
583 int io_trim;
584
585 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
586
587 /*
588 * If bps are unlimited (-1), then time slice don't get
589 * renewed. Don't try to trim the slice if slice is used. A new
590 * slice will start when appropriate.
591 */
592 if (throtl_slice_used(tg, rw))
593 return;
594
595 /*
596 * A bio has been dispatched. Also adjust slice_end. It might happen
597 * that initially cgroup limit was very low resulting in high
598 * slice_end, but later limit was bumped up and bio was dispatched
599 * sooner, then we need to reduce slice_end. A high bogus slice_end
600 * is bad because it does not allow new slice to start.
601 */
602
603 throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
604
605 time_elapsed = rounddown(jiffies - tg->slice_start[rw],
606 tg->td->throtl_slice);
607 if (!time_elapsed)
608 return;
609
610 bytes_trim = calculate_bytes_allowed(tg_bps_limit(tg, rw),
611 time_elapsed) +
612 tg->carryover_bytes[rw];
613 io_trim = calculate_io_allowed(tg_iops_limit(tg, rw), time_elapsed) +
614 tg->carryover_ios[rw];
615 if (bytes_trim <= 0 && io_trim <= 0)
616 return;
617
618 tg->carryover_bytes[rw] = 0;
619 if ((long long)tg->bytes_disp[rw] >= bytes_trim)
620 tg->bytes_disp[rw] -= bytes_trim;
621 else
622 tg->bytes_disp[rw] = 0;
623
624 tg->carryover_ios[rw] = 0;
625 if ((int)tg->io_disp[rw] >= io_trim)
626 tg->io_disp[rw] -= io_trim;
627 else
628 tg->io_disp[rw] = 0;
629
630 tg->slice_start[rw] += time_elapsed;
631
632 throtl_log(&tg->service_queue,
633 "[%c] trim slice nr=%lu bytes=%lld io=%d start=%lu end=%lu jiffies=%lu",
634 rw == READ ? 'R' : 'W', time_elapsed / tg->td->throtl_slice,
635 bytes_trim, io_trim, tg->slice_start[rw], tg->slice_end[rw],
636 jiffies);
637}
638
639static void __tg_update_carryover(struct throtl_grp *tg, bool rw)
640{
641 unsigned long jiffy_elapsed = jiffies - tg->slice_start[rw];
642 u64 bps_limit = tg_bps_limit(tg, rw);
643 u32 iops_limit = tg_iops_limit(tg, rw);
644
645 /*
646 * If config is updated while bios are still throttled, calculate and
647 * accumulate how many bytes/ios are waited across changes. And
648 * carryover_bytes/ios will be used to calculate new wait time under new
649 * configuration.
650 */
651 if (bps_limit != U64_MAX)
652 tg->carryover_bytes[rw] +=
653 calculate_bytes_allowed(bps_limit, jiffy_elapsed) -
654 tg->bytes_disp[rw];
655 if (iops_limit != UINT_MAX)
656 tg->carryover_ios[rw] +=
657 calculate_io_allowed(iops_limit, jiffy_elapsed) -
658 tg->io_disp[rw];
659}
660
661static void tg_update_carryover(struct throtl_grp *tg)
662{
663 if (tg->service_queue.nr_queued[READ])
664 __tg_update_carryover(tg, READ);
665 if (tg->service_queue.nr_queued[WRITE])
666 __tg_update_carryover(tg, WRITE);
667
668 /* see comments in struct throtl_grp for meaning of these fields. */
669 throtl_log(&tg->service_queue, "%s: %lld %lld %d %d\n", __func__,
670 tg->carryover_bytes[READ], tg->carryover_bytes[WRITE],
671 tg->carryover_ios[READ], tg->carryover_ios[WRITE]);
672}
673
674static unsigned long tg_within_iops_limit(struct throtl_grp *tg, struct bio *bio,
675 u32 iops_limit)
676{
677 bool rw = bio_data_dir(bio);
678 int io_allowed;
679 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
680
681 if (iops_limit == UINT_MAX) {
682 return 0;
683 }
684
685 jiffy_elapsed = jiffies - tg->slice_start[rw];
686
687 /* Round up to the next throttle slice, wait time must be nonzero */
688 jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
689 io_allowed = calculate_io_allowed(iops_limit, jiffy_elapsed_rnd) +
690 tg->carryover_ios[rw];
691 if (io_allowed > 0 && tg->io_disp[rw] + 1 <= io_allowed)
692 return 0;
693
694 /* Calc approx time to dispatch */
695 jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
696
697 /* make sure at least one io can be dispatched after waiting */
698 jiffy_wait = max(jiffy_wait, HZ / iops_limit + 1);
699 return jiffy_wait;
700}
701
702static unsigned long tg_within_bps_limit(struct throtl_grp *tg, struct bio *bio,
703 u64 bps_limit)
704{
705 bool rw = bio_data_dir(bio);
706 long long bytes_allowed;
707 u64 extra_bytes;
708 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
709 unsigned int bio_size = throtl_bio_data_size(bio);
710
711 /* no need to throttle if this bio's bytes have been accounted */
712 if (bps_limit == U64_MAX || bio_flagged(bio, BIO_BPS_THROTTLED)) {
713 return 0;
714 }
715
716 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
717
718 /* Slice has just started. Consider one slice interval */
719 if (!jiffy_elapsed)
720 jiffy_elapsed_rnd = tg->td->throtl_slice;
721
722 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
723 bytes_allowed = calculate_bytes_allowed(bps_limit, jiffy_elapsed_rnd) +
724 tg->carryover_bytes[rw];
725 if (bytes_allowed > 0 && tg->bytes_disp[rw] + bio_size <= bytes_allowed)
726 return 0;
727
728 /* Calc approx time to dispatch */
729 extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
730 jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
731
732 if (!jiffy_wait)
733 jiffy_wait = 1;
734
735 /*
736 * This wait time is without taking into consideration the rounding
737 * up we did. Add that time also.
738 */
739 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
740 return jiffy_wait;
741}
742
743/*
744 * Returns whether one can dispatch a bio or not. Also returns approx number
745 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
746 */
747static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
748 unsigned long *wait)
749{
750 bool rw = bio_data_dir(bio);
751 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
752 u64 bps_limit = tg_bps_limit(tg, rw);
753 u32 iops_limit = tg_iops_limit(tg, rw);
754
755 /*
756 * Currently whole state machine of group depends on first bio
757 * queued in the group bio list. So one should not be calling
758 * this function with a different bio if there are other bios
759 * queued.
760 */
761 BUG_ON(tg->service_queue.nr_queued[rw] &&
762 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
763
764 /* If tg->bps = -1, then BW is unlimited */
765 if ((bps_limit == U64_MAX && iops_limit == UINT_MAX) ||
766 tg->flags & THROTL_TG_CANCELING) {
767 if (wait)
768 *wait = 0;
769 return true;
770 }
771
772 /*
773 * If previous slice expired, start a new one otherwise renew/extend
774 * existing slice to make sure it is at least throtl_slice interval
775 * long since now. New slice is started only for empty throttle group.
776 * If there is queued bio, that means there should be an active
777 * slice and it should be extended instead.
778 */
779 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
780 throtl_start_new_slice(tg, rw, true);
781 else {
782 if (time_before(tg->slice_end[rw],
783 jiffies + tg->td->throtl_slice))
784 throtl_extend_slice(tg, rw,
785 jiffies + tg->td->throtl_slice);
786 }
787
788 bps_wait = tg_within_bps_limit(tg, bio, bps_limit);
789 iops_wait = tg_within_iops_limit(tg, bio, iops_limit);
790 if (bps_wait + iops_wait == 0) {
791 if (wait)
792 *wait = 0;
793 return true;
794 }
795
796 max_wait = max(bps_wait, iops_wait);
797
798 if (wait)
799 *wait = max_wait;
800
801 if (time_before(tg->slice_end[rw], jiffies + max_wait))
802 throtl_extend_slice(tg, rw, jiffies + max_wait);
803
804 return false;
805}
806
807static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
808{
809 bool rw = bio_data_dir(bio);
810 unsigned int bio_size = throtl_bio_data_size(bio);
811
812 /* Charge the bio to the group */
813 if (!bio_flagged(bio, BIO_BPS_THROTTLED)) {
814 tg->bytes_disp[rw] += bio_size;
815 tg->last_bytes_disp[rw] += bio_size;
816 }
817
818 tg->io_disp[rw]++;
819 tg->last_io_disp[rw]++;
820}
821
822/**
823 * throtl_add_bio_tg - add a bio to the specified throtl_grp
824 * @bio: bio to add
825 * @qn: qnode to use
826 * @tg: the target throtl_grp
827 *
828 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
829 * tg->qnode_on_self[] is used.
830 */
831static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
832 struct throtl_grp *tg)
833{
834 struct throtl_service_queue *sq = &tg->service_queue;
835 bool rw = bio_data_dir(bio);
836
837 if (!qn)
838 qn = &tg->qnode_on_self[rw];
839
840 /*
841 * If @tg doesn't currently have any bios queued in the same
842 * direction, queueing @bio can change when @tg should be
843 * dispatched. Mark that @tg was empty. This is automatically
844 * cleared on the next tg_update_disptime().
845 */
846 if (!sq->nr_queued[rw])
847 tg->flags |= THROTL_TG_WAS_EMPTY;
848
849 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
850
851 sq->nr_queued[rw]++;
852 throtl_enqueue_tg(tg);
853}
854
855static void tg_update_disptime(struct throtl_grp *tg)
856{
857 struct throtl_service_queue *sq = &tg->service_queue;
858 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
859 struct bio *bio;
860
861 bio = throtl_peek_queued(&sq->queued[READ]);
862 if (bio)
863 tg_may_dispatch(tg, bio, &read_wait);
864
865 bio = throtl_peek_queued(&sq->queued[WRITE]);
866 if (bio)
867 tg_may_dispatch(tg, bio, &write_wait);
868
869 min_wait = min(read_wait, write_wait);
870 disptime = jiffies + min_wait;
871
872 /* Update dispatch time */
873 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
874 tg->disptime = disptime;
875 tg_service_queue_add(tg);
876
877 /* see throtl_add_bio_tg() */
878 tg->flags &= ~THROTL_TG_WAS_EMPTY;
879}
880
881static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
882 struct throtl_grp *parent_tg, bool rw)
883{
884 if (throtl_slice_used(parent_tg, rw)) {
885 throtl_start_new_slice_with_credit(parent_tg, rw,
886 child_tg->slice_start[rw]);
887 }
888
889}
890
891static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
892{
893 struct throtl_service_queue *sq = &tg->service_queue;
894 struct throtl_service_queue *parent_sq = sq->parent_sq;
895 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
896 struct throtl_grp *tg_to_put = NULL;
897 struct bio *bio;
898
899 /*
900 * @bio is being transferred from @tg to @parent_sq. Popping a bio
901 * from @tg may put its reference and @parent_sq might end up
902 * getting released prematurely. Remember the tg to put and put it
903 * after @bio is transferred to @parent_sq.
904 */
905 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
906 sq->nr_queued[rw]--;
907
908 throtl_charge_bio(tg, bio);
909
910 /*
911 * If our parent is another tg, we just need to transfer @bio to
912 * the parent using throtl_add_bio_tg(). If our parent is
913 * @td->service_queue, @bio is ready to be issued. Put it on its
914 * bio_lists[] and decrease total number queued. The caller is
915 * responsible for issuing these bios.
916 */
917 if (parent_tg) {
918 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
919 start_parent_slice_with_credit(tg, parent_tg, rw);
920 } else {
921 bio_set_flag(bio, BIO_BPS_THROTTLED);
922 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
923 &parent_sq->queued[rw]);
924 BUG_ON(tg->td->nr_queued[rw] <= 0);
925 tg->td->nr_queued[rw]--;
926 }
927
928 throtl_trim_slice(tg, rw);
929
930 if (tg_to_put)
931 blkg_put(tg_to_blkg(tg_to_put));
932}
933
934static int throtl_dispatch_tg(struct throtl_grp *tg)
935{
936 struct throtl_service_queue *sq = &tg->service_queue;
937 unsigned int nr_reads = 0, nr_writes = 0;
938 unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
939 unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
940 struct bio *bio;
941
942 /* Try to dispatch 75% READS and 25% WRITES */
943
944 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
945 tg_may_dispatch(tg, bio, NULL)) {
946
947 tg_dispatch_one_bio(tg, READ);
948 nr_reads++;
949
950 if (nr_reads >= max_nr_reads)
951 break;
952 }
953
954 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
955 tg_may_dispatch(tg, bio, NULL)) {
956
957 tg_dispatch_one_bio(tg, WRITE);
958 nr_writes++;
959
960 if (nr_writes >= max_nr_writes)
961 break;
962 }
963
964 return nr_reads + nr_writes;
965}
966
967static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
968{
969 unsigned int nr_disp = 0;
970
971 while (1) {
972 struct throtl_grp *tg;
973 struct throtl_service_queue *sq;
974
975 if (!parent_sq->nr_pending)
976 break;
977
978 tg = throtl_rb_first(parent_sq);
979 if (!tg)
980 break;
981
982 if (time_before(jiffies, tg->disptime))
983 break;
984
985 nr_disp += throtl_dispatch_tg(tg);
986
987 sq = &tg->service_queue;
988 if (sq->nr_queued[READ] || sq->nr_queued[WRITE])
989 tg_update_disptime(tg);
990 else
991 throtl_dequeue_tg(tg);
992
993 if (nr_disp >= THROTL_QUANTUM)
994 break;
995 }
996
997 return nr_disp;
998}
999
1000/**
1001 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1002 * @t: the pending_timer member of the throtl_service_queue being serviced
1003 *
1004 * This timer is armed when a child throtl_grp with active bio's become
1005 * pending and queued on the service_queue's pending_tree and expires when
1006 * the first child throtl_grp should be dispatched. This function
1007 * dispatches bio's from the children throtl_grps to the parent
1008 * service_queue.
1009 *
1010 * If the parent's parent is another throtl_grp, dispatching is propagated
1011 * by either arming its pending_timer or repeating dispatch directly. If
1012 * the top-level service_tree is reached, throtl_data->dispatch_work is
1013 * kicked so that the ready bio's are issued.
1014 */
1015static void throtl_pending_timer_fn(struct timer_list *t)
1016{
1017 struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1018 struct throtl_grp *tg = sq_to_tg(sq);
1019 struct throtl_data *td = sq_to_td(sq);
1020 struct throtl_service_queue *parent_sq;
1021 struct request_queue *q;
1022 bool dispatched;
1023 int ret;
1024
1025 /* throtl_data may be gone, so figure out request queue by blkg */
1026 if (tg)
1027 q = tg->pd.blkg->q;
1028 else
1029 q = td->queue;
1030
1031 spin_lock_irq(&q->queue_lock);
1032
1033 if (!q->root_blkg)
1034 goto out_unlock;
1035
1036again:
1037 parent_sq = sq->parent_sq;
1038 dispatched = false;
1039
1040 while (true) {
1041 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1042 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1043 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1044
1045 ret = throtl_select_dispatch(sq);
1046 if (ret) {
1047 throtl_log(sq, "bios disp=%u", ret);
1048 dispatched = true;
1049 }
1050
1051 if (throtl_schedule_next_dispatch(sq, false))
1052 break;
1053
1054 /* this dispatch windows is still open, relax and repeat */
1055 spin_unlock_irq(&q->queue_lock);
1056 cpu_relax();
1057 spin_lock_irq(&q->queue_lock);
1058 }
1059
1060 if (!dispatched)
1061 goto out_unlock;
1062
1063 if (parent_sq) {
1064 /* @parent_sq is another throl_grp, propagate dispatch */
1065 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1066 tg_update_disptime(tg);
1067 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1068 /* window is already open, repeat dispatching */
1069 sq = parent_sq;
1070 tg = sq_to_tg(sq);
1071 goto again;
1072 }
1073 }
1074 } else {
1075 /* reached the top-level, queue issuing */
1076 queue_work(kthrotld_workqueue, &td->dispatch_work);
1077 }
1078out_unlock:
1079 spin_unlock_irq(&q->queue_lock);
1080}
1081
1082/**
1083 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1084 * @work: work item being executed
1085 *
1086 * This function is queued for execution when bios reach the bio_lists[]
1087 * of throtl_data->service_queue. Those bios are ready and issued by this
1088 * function.
1089 */
1090static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1091{
1092 struct throtl_data *td = container_of(work, struct throtl_data,
1093 dispatch_work);
1094 struct throtl_service_queue *td_sq = &td->service_queue;
1095 struct request_queue *q = td->queue;
1096 struct bio_list bio_list_on_stack;
1097 struct bio *bio;
1098 struct blk_plug plug;
1099 int rw;
1100
1101 bio_list_init(&bio_list_on_stack);
1102
1103 spin_lock_irq(&q->queue_lock);
1104 for (rw = READ; rw <= WRITE; rw++)
1105 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1106 bio_list_add(&bio_list_on_stack, bio);
1107 spin_unlock_irq(&q->queue_lock);
1108
1109 if (!bio_list_empty(&bio_list_on_stack)) {
1110 blk_start_plug(&plug);
1111 while ((bio = bio_list_pop(&bio_list_on_stack)))
1112 submit_bio_noacct_nocheck(bio);
1113 blk_finish_plug(&plug);
1114 }
1115}
1116
1117static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1118 int off)
1119{
1120 struct throtl_grp *tg = pd_to_tg(pd);
1121 u64 v = *(u64 *)((void *)tg + off);
1122
1123 if (v == U64_MAX)
1124 return 0;
1125 return __blkg_prfill_u64(sf, pd, v);
1126}
1127
1128static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1129 int off)
1130{
1131 struct throtl_grp *tg = pd_to_tg(pd);
1132 unsigned int v = *(unsigned int *)((void *)tg + off);
1133
1134 if (v == UINT_MAX)
1135 return 0;
1136 return __blkg_prfill_u64(sf, pd, v);
1137}
1138
1139static int tg_print_conf_u64(struct seq_file *sf, void *v)
1140{
1141 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1142 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1143 return 0;
1144}
1145
1146static int tg_print_conf_uint(struct seq_file *sf, void *v)
1147{
1148 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1149 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1150 return 0;
1151}
1152
1153static void tg_conf_updated(struct throtl_grp *tg, bool global)
1154{
1155 struct throtl_service_queue *sq = &tg->service_queue;
1156 struct cgroup_subsys_state *pos_css;
1157 struct blkcg_gq *blkg;
1158
1159 throtl_log(&tg->service_queue,
1160 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1161 tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1162 tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1163
1164 rcu_read_lock();
1165 /*
1166 * Update has_rules[] flags for the updated tg's subtree. A tg is
1167 * considered to have rules if either the tg itself or any of its
1168 * ancestors has rules. This identifies groups without any
1169 * restrictions in the whole hierarchy and allows them to bypass
1170 * blk-throttle.
1171 */
1172 blkg_for_each_descendant_pre(blkg, pos_css,
1173 global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1174 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1175
1176 tg_update_has_rules(this_tg);
1177 /* ignore root/second level */
1178 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1179 !blkg->parent->parent)
1180 continue;
1181 }
1182 rcu_read_unlock();
1183
1184 /*
1185 * We're already holding queue_lock and know @tg is valid. Let's
1186 * apply the new config directly.
1187 *
1188 * Restart the slices for both READ and WRITES. It might happen
1189 * that a group's limit are dropped suddenly and we don't want to
1190 * account recently dispatched IO with new low rate.
1191 */
1192 throtl_start_new_slice(tg, READ, false);
1193 throtl_start_new_slice(tg, WRITE, false);
1194
1195 if (tg->flags & THROTL_TG_PENDING) {
1196 tg_update_disptime(tg);
1197 throtl_schedule_next_dispatch(sq->parent_sq, true);
1198 }
1199}
1200
1201static int blk_throtl_init(struct gendisk *disk)
1202{
1203 struct request_queue *q = disk->queue;
1204 struct throtl_data *td;
1205 int ret;
1206
1207 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1208 if (!td)
1209 return -ENOMEM;
1210
1211 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1212 throtl_service_queue_init(&td->service_queue);
1213
1214 /*
1215 * Freeze queue before activating policy, to synchronize with IO path,
1216 * which is protected by 'q_usage_counter'.
1217 */
1218 blk_mq_freeze_queue(disk->queue);
1219 blk_mq_quiesce_queue(disk->queue);
1220
1221 q->td = td;
1222 td->queue = q;
1223
1224 /* activate policy */
1225 ret = blkcg_activate_policy(disk, &blkcg_policy_throtl);
1226 if (ret) {
1227 q->td = NULL;
1228 kfree(td);
1229 goto out;
1230 }
1231
1232 if (blk_queue_nonrot(q))
1233 td->throtl_slice = DFL_THROTL_SLICE_SSD;
1234 else
1235 td->throtl_slice = DFL_THROTL_SLICE_HD;
1236 td->track_bio_latency = !queue_is_mq(q);
1237 if (!td->track_bio_latency)
1238 blk_stat_enable_accounting(q);
1239
1240out:
1241 blk_mq_unquiesce_queue(disk->queue);
1242 blk_mq_unfreeze_queue(disk->queue);
1243
1244 return ret;
1245}
1246
1247
1248static ssize_t tg_set_conf(struct kernfs_open_file *of,
1249 char *buf, size_t nbytes, loff_t off, bool is_u64)
1250{
1251 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1252 struct blkg_conf_ctx ctx;
1253 struct throtl_grp *tg;
1254 int ret;
1255 u64 v;
1256
1257 blkg_conf_init(&ctx, buf);
1258
1259 ret = blkg_conf_open_bdev(&ctx);
1260 if (ret)
1261 goto out_finish;
1262
1263 if (!blk_throtl_activated(ctx.bdev->bd_queue)) {
1264 ret = blk_throtl_init(ctx.bdev->bd_disk);
1265 if (ret)
1266 goto out_finish;
1267 }
1268
1269 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
1270 if (ret)
1271 goto out_finish;
1272
1273 ret = -EINVAL;
1274 if (sscanf(ctx.body, "%llu", &v) != 1)
1275 goto out_finish;
1276 if (!v)
1277 v = U64_MAX;
1278
1279 tg = blkg_to_tg(ctx.blkg);
1280 tg_update_carryover(tg);
1281
1282 if (is_u64)
1283 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1284 else
1285 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1286
1287 tg_conf_updated(tg, false);
1288 ret = 0;
1289out_finish:
1290 blkg_conf_exit(&ctx);
1291 return ret ?: nbytes;
1292}
1293
1294static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1295 char *buf, size_t nbytes, loff_t off)
1296{
1297 return tg_set_conf(of, buf, nbytes, off, true);
1298}
1299
1300static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1301 char *buf, size_t nbytes, loff_t off)
1302{
1303 return tg_set_conf(of, buf, nbytes, off, false);
1304}
1305
1306static int tg_print_rwstat(struct seq_file *sf, void *v)
1307{
1308 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1309 blkg_prfill_rwstat, &blkcg_policy_throtl,
1310 seq_cft(sf)->private, true);
1311 return 0;
1312}
1313
1314static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1315 struct blkg_policy_data *pd, int off)
1316{
1317 struct blkg_rwstat_sample sum;
1318
1319 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1320 &sum);
1321 return __blkg_prfill_rwstat(sf, pd, &sum);
1322}
1323
1324static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1325{
1326 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1327 tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1328 seq_cft(sf)->private, true);
1329 return 0;
1330}
1331
1332static struct cftype throtl_legacy_files[] = {
1333 {
1334 .name = "throttle.read_bps_device",
1335 .private = offsetof(struct throtl_grp, bps[READ]),
1336 .seq_show = tg_print_conf_u64,
1337 .write = tg_set_conf_u64,
1338 },
1339 {
1340 .name = "throttle.write_bps_device",
1341 .private = offsetof(struct throtl_grp, bps[WRITE]),
1342 .seq_show = tg_print_conf_u64,
1343 .write = tg_set_conf_u64,
1344 },
1345 {
1346 .name = "throttle.read_iops_device",
1347 .private = offsetof(struct throtl_grp, iops[READ]),
1348 .seq_show = tg_print_conf_uint,
1349 .write = tg_set_conf_uint,
1350 },
1351 {
1352 .name = "throttle.write_iops_device",
1353 .private = offsetof(struct throtl_grp, iops[WRITE]),
1354 .seq_show = tg_print_conf_uint,
1355 .write = tg_set_conf_uint,
1356 },
1357 {
1358 .name = "throttle.io_service_bytes",
1359 .private = offsetof(struct throtl_grp, stat_bytes),
1360 .seq_show = tg_print_rwstat,
1361 },
1362 {
1363 .name = "throttle.io_service_bytes_recursive",
1364 .private = offsetof(struct throtl_grp, stat_bytes),
1365 .seq_show = tg_print_rwstat_recursive,
1366 },
1367 {
1368 .name = "throttle.io_serviced",
1369 .private = offsetof(struct throtl_grp, stat_ios),
1370 .seq_show = tg_print_rwstat,
1371 },
1372 {
1373 .name = "throttle.io_serviced_recursive",
1374 .private = offsetof(struct throtl_grp, stat_ios),
1375 .seq_show = tg_print_rwstat_recursive,
1376 },
1377 { } /* terminate */
1378};
1379
1380static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1381 int off)
1382{
1383 struct throtl_grp *tg = pd_to_tg(pd);
1384 const char *dname = blkg_dev_name(pd->blkg);
1385 u64 bps_dft;
1386 unsigned int iops_dft;
1387
1388 if (!dname)
1389 return 0;
1390
1391 bps_dft = U64_MAX;
1392 iops_dft = UINT_MAX;
1393
1394 if (tg->bps[READ] == bps_dft &&
1395 tg->bps[WRITE] == bps_dft &&
1396 tg->iops[READ] == iops_dft &&
1397 tg->iops[WRITE] == iops_dft)
1398 return 0;
1399
1400 seq_printf(sf, "%s", dname);
1401 if (tg->bps[READ] == U64_MAX)
1402 seq_printf(sf, " rbps=max");
1403 else
1404 seq_printf(sf, " rbps=%llu", tg->bps[READ]);
1405
1406 if (tg->bps[WRITE] == U64_MAX)
1407 seq_printf(sf, " wbps=max");
1408 else
1409 seq_printf(sf, " wbps=%llu", tg->bps[WRITE]);
1410
1411 if (tg->iops[READ] == UINT_MAX)
1412 seq_printf(sf, " riops=max");
1413 else
1414 seq_printf(sf, " riops=%u", tg->iops[READ]);
1415
1416 if (tg->iops[WRITE] == UINT_MAX)
1417 seq_printf(sf, " wiops=max");
1418 else
1419 seq_printf(sf, " wiops=%u", tg->iops[WRITE]);
1420
1421 seq_printf(sf, "\n");
1422 return 0;
1423}
1424
1425static int tg_print_limit(struct seq_file *sf, void *v)
1426{
1427 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1428 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1429 return 0;
1430}
1431
1432static ssize_t tg_set_limit(struct kernfs_open_file *of,
1433 char *buf, size_t nbytes, loff_t off)
1434{
1435 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1436 struct blkg_conf_ctx ctx;
1437 struct throtl_grp *tg;
1438 u64 v[4];
1439 int ret;
1440
1441 blkg_conf_init(&ctx, buf);
1442
1443 ret = blkg_conf_open_bdev(&ctx);
1444 if (ret)
1445 goto out_finish;
1446
1447 if (!blk_throtl_activated(ctx.bdev->bd_queue)) {
1448 ret = blk_throtl_init(ctx.bdev->bd_disk);
1449 if (ret)
1450 goto out_finish;
1451 }
1452
1453 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
1454 if (ret)
1455 goto out_finish;
1456
1457 tg = blkg_to_tg(ctx.blkg);
1458 tg_update_carryover(tg);
1459
1460 v[0] = tg->bps[READ];
1461 v[1] = tg->bps[WRITE];
1462 v[2] = tg->iops[READ];
1463 v[3] = tg->iops[WRITE];
1464
1465 while (true) {
1466 char tok[27]; /* wiops=18446744073709551616 */
1467 char *p;
1468 u64 val = U64_MAX;
1469 int len;
1470
1471 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1472 break;
1473 if (tok[0] == '\0')
1474 break;
1475 ctx.body += len;
1476
1477 ret = -EINVAL;
1478 p = tok;
1479 strsep(&p, "=");
1480 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1481 goto out_finish;
1482
1483 ret = -ERANGE;
1484 if (!val)
1485 goto out_finish;
1486
1487 ret = -EINVAL;
1488 if (!strcmp(tok, "rbps"))
1489 v[0] = val;
1490 else if (!strcmp(tok, "wbps"))
1491 v[1] = val;
1492 else if (!strcmp(tok, "riops"))
1493 v[2] = min_t(u64, val, UINT_MAX);
1494 else if (!strcmp(tok, "wiops"))
1495 v[3] = min_t(u64, val, UINT_MAX);
1496 else
1497 goto out_finish;
1498 }
1499
1500 tg->bps[READ] = v[0];
1501 tg->bps[WRITE] = v[1];
1502 tg->iops[READ] = v[2];
1503 tg->iops[WRITE] = v[3];
1504
1505 tg_conf_updated(tg, false);
1506 ret = 0;
1507out_finish:
1508 blkg_conf_exit(&ctx);
1509 return ret ?: nbytes;
1510}
1511
1512static struct cftype throtl_files[] = {
1513 {
1514 .name = "max",
1515 .flags = CFTYPE_NOT_ON_ROOT,
1516 .seq_show = tg_print_limit,
1517 .write = tg_set_limit,
1518 },
1519 { } /* terminate */
1520};
1521
1522static void throtl_shutdown_wq(struct request_queue *q)
1523{
1524 struct throtl_data *td = q->td;
1525
1526 cancel_work_sync(&td->dispatch_work);
1527}
1528
1529static void tg_flush_bios(struct throtl_grp *tg)
1530{
1531 struct throtl_service_queue *sq = &tg->service_queue;
1532
1533 if (tg->flags & THROTL_TG_CANCELING)
1534 return;
1535 /*
1536 * Set the flag to make sure throtl_pending_timer_fn() won't
1537 * stop until all throttled bios are dispatched.
1538 */
1539 tg->flags |= THROTL_TG_CANCELING;
1540
1541 /*
1542 * Do not dispatch cgroup without THROTL_TG_PENDING or cgroup
1543 * will be inserted to service queue without THROTL_TG_PENDING
1544 * set in tg_update_disptime below. Then IO dispatched from
1545 * child in tg_dispatch_one_bio will trigger double insertion
1546 * and corrupt the tree.
1547 */
1548 if (!(tg->flags & THROTL_TG_PENDING))
1549 return;
1550
1551 /*
1552 * Update disptime after setting the above flag to make sure
1553 * throtl_select_dispatch() won't exit without dispatching.
1554 */
1555 tg_update_disptime(tg);
1556
1557 throtl_schedule_pending_timer(sq, jiffies + 1);
1558}
1559
1560static void throtl_pd_offline(struct blkg_policy_data *pd)
1561{
1562 tg_flush_bios(pd_to_tg(pd));
1563}
1564
1565struct blkcg_policy blkcg_policy_throtl = {
1566 .dfl_cftypes = throtl_files,
1567 .legacy_cftypes = throtl_legacy_files,
1568
1569 .pd_alloc_fn = throtl_pd_alloc,
1570 .pd_init_fn = throtl_pd_init,
1571 .pd_online_fn = throtl_pd_online,
1572 .pd_offline_fn = throtl_pd_offline,
1573 .pd_free_fn = throtl_pd_free,
1574};
1575
1576void blk_throtl_cancel_bios(struct gendisk *disk)
1577{
1578 struct request_queue *q = disk->queue;
1579 struct cgroup_subsys_state *pos_css;
1580 struct blkcg_gq *blkg;
1581
1582 if (!blk_throtl_activated(q))
1583 return;
1584
1585 spin_lock_irq(&q->queue_lock);
1586 /*
1587 * queue_lock is held, rcu lock is not needed here technically.
1588 * However, rcu lock is still held to emphasize that following
1589 * path need RCU protection and to prevent warning from lockdep.
1590 */
1591 rcu_read_lock();
1592 blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) {
1593 /*
1594 * disk_release will call pd_offline_fn to cancel bios.
1595 * However, disk_release can't be called if someone get
1596 * the refcount of device and issued bios which are
1597 * inflight after del_gendisk.
1598 * Cancel bios here to ensure no bios are inflight after
1599 * del_gendisk.
1600 */
1601 tg_flush_bios(blkg_to_tg(blkg));
1602 }
1603 rcu_read_unlock();
1604 spin_unlock_irq(&q->queue_lock);
1605}
1606
1607static bool tg_within_limit(struct throtl_grp *tg, struct bio *bio, bool rw)
1608{
1609 /* throtl is FIFO - if bios are already queued, should queue */
1610 if (tg->service_queue.nr_queued[rw])
1611 return false;
1612
1613 return tg_may_dispatch(tg, bio, NULL);
1614}
1615
1616static void tg_dispatch_in_debt(struct throtl_grp *tg, struct bio *bio, bool rw)
1617{
1618 if (!bio_flagged(bio, BIO_BPS_THROTTLED))
1619 tg->carryover_bytes[rw] -= throtl_bio_data_size(bio);
1620 tg->carryover_ios[rw]--;
1621}
1622
1623bool __blk_throtl_bio(struct bio *bio)
1624{
1625 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
1626 struct blkcg_gq *blkg = bio->bi_blkg;
1627 struct throtl_qnode *qn = NULL;
1628 struct throtl_grp *tg = blkg_to_tg(blkg);
1629 struct throtl_service_queue *sq;
1630 bool rw = bio_data_dir(bio);
1631 bool throttled = false;
1632 struct throtl_data *td = tg->td;
1633
1634 rcu_read_lock();
1635 spin_lock_irq(&q->queue_lock);
1636 sq = &tg->service_queue;
1637
1638 while (true) {
1639 if (tg_within_limit(tg, bio, rw)) {
1640 /* within limits, let's charge and dispatch directly */
1641 throtl_charge_bio(tg, bio);
1642
1643 /*
1644 * We need to trim slice even when bios are not being
1645 * queued otherwise it might happen that a bio is not
1646 * queued for a long time and slice keeps on extending
1647 * and trim is not called for a long time. Now if limits
1648 * are reduced suddenly we take into account all the IO
1649 * dispatched so far at new low rate and * newly queued
1650 * IO gets a really long dispatch time.
1651 *
1652 * So keep on trimming slice even if bio is not queued.
1653 */
1654 throtl_trim_slice(tg, rw);
1655 } else if (bio_issue_as_root_blkg(bio)) {
1656 /*
1657 * IOs which may cause priority inversions are
1658 * dispatched directly, even if they're over limit.
1659 * Debts are handled by carryover_bytes/ios while
1660 * calculating wait time.
1661 */
1662 tg_dispatch_in_debt(tg, bio, rw);
1663 } else {
1664 /* if above limits, break to queue */
1665 break;
1666 }
1667
1668 /*
1669 * @bio passed through this layer without being throttled.
1670 * Climb up the ladder. If we're already at the top, it
1671 * can be executed directly.
1672 */
1673 qn = &tg->qnode_on_parent[rw];
1674 sq = sq->parent_sq;
1675 tg = sq_to_tg(sq);
1676 if (!tg) {
1677 bio_set_flag(bio, BIO_BPS_THROTTLED);
1678 goto out_unlock;
1679 }
1680 }
1681
1682 /* out-of-limit, queue to @tg */
1683 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1684 rw == READ ? 'R' : 'W',
1685 tg->bytes_disp[rw], bio->bi_iter.bi_size,
1686 tg_bps_limit(tg, rw),
1687 tg->io_disp[rw], tg_iops_limit(tg, rw),
1688 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1689
1690 td->nr_queued[rw]++;
1691 throtl_add_bio_tg(bio, qn, tg);
1692 throttled = true;
1693
1694 /*
1695 * Update @tg's dispatch time and force schedule dispatch if @tg
1696 * was empty before @bio. The forced scheduling isn't likely to
1697 * cause undue delay as @bio is likely to be dispatched directly if
1698 * its @tg's disptime is not in the future.
1699 */
1700 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1701 tg_update_disptime(tg);
1702 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1703 }
1704
1705out_unlock:
1706 spin_unlock_irq(&q->queue_lock);
1707
1708 rcu_read_unlock();
1709 return throttled;
1710}
1711
1712void blk_throtl_exit(struct gendisk *disk)
1713{
1714 struct request_queue *q = disk->queue;
1715
1716 if (!blk_throtl_activated(q))
1717 return;
1718
1719 del_timer_sync(&q->td->service_queue.pending_timer);
1720 throtl_shutdown_wq(q);
1721 blkcg_deactivate_policy(disk, &blkcg_policy_throtl);
1722 kfree(q->td);
1723}
1724
1725static int __init throtl_init(void)
1726{
1727 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1728 if (!kthrotld_workqueue)
1729 panic("Failed to create kthrotld\n");
1730
1731 return blkcg_policy_register(&blkcg_policy_throtl);
1732}
1733
1734module_init(throtl_init);