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/*
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 <linux/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;
28
29/*
30 * To implement hierarchical throttling, throtl_grps form a tree and bios
31 * are dispatched upwards level by level until they reach the top and get
32 * issued. When dispatching bios from the children and local group at each
33 * level, if the bios are dispatched into a single bio_list, there's a risk
34 * of a local or child group which can queue many bios at once filling up
35 * the list starving others.
36 *
37 * To avoid such starvation, dispatched bios are queued separately
38 * according to where they came from. When they are again dispatched to
39 * the parent, they're popped in round-robin order so that no single source
40 * hogs the dispatch window.
41 *
42 * throtl_qnode is used to keep the queued bios separated by their sources.
43 * Bios are queued to throtl_qnode which in turn is queued to
44 * throtl_service_queue and then dispatched in round-robin order.
45 *
46 * It's also used to track the reference counts on blkg's. A qnode always
47 * belongs to a throtl_grp and gets queued on itself or the parent, so
48 * incrementing the reference of the associated throtl_grp when a qnode is
49 * queued and decrementing when dequeued is enough to keep the whole blkg
50 * tree pinned while bios are in flight.
51 */
52struct throtl_qnode {
53 struct list_head node; /* service_queue->queued[] */
54 struct bio_list bios; /* queued bios */
55 struct throtl_grp *tg; /* tg this qnode belongs to */
56};
57
58struct throtl_service_queue {
59 struct throtl_service_queue *parent_sq; /* the parent service_queue */
60
61 /*
62 * Bios queued directly to this service_queue or dispatched from
63 * children throtl_grp's.
64 */
65 struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */
66 unsigned int nr_queued[2]; /* number of queued bios */
67
68 /*
69 * RB tree of active children throtl_grp's, which are sorted by
70 * their ->disptime.
71 */
72 struct rb_root pending_tree; /* RB tree of active tgs */
73 struct rb_node *first_pending; /* first node in the tree */
74 unsigned int nr_pending; /* # queued in the tree */
75 unsigned long first_pending_disptime; /* disptime of the first tg */
76 struct timer_list pending_timer; /* fires on first_pending_disptime */
77};
78
79enum tg_state_flags {
80 THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */
81 THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */
82};
83
84#define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
85
86struct throtl_grp {
87 /* must be the first member */
88 struct blkg_policy_data pd;
89
90 /* active throtl group service_queue member */
91 struct rb_node rb_node;
92
93 /* throtl_data this group belongs to */
94 struct throtl_data *td;
95
96 /* this group's service queue */
97 struct throtl_service_queue service_queue;
98
99 /*
100 * qnode_on_self is used when bios are directly queued to this
101 * throtl_grp so that local bios compete fairly with bios
102 * dispatched from children. qnode_on_parent is used when bios are
103 * dispatched from this throtl_grp into its parent and will compete
104 * with the sibling qnode_on_parents and the parent's
105 * qnode_on_self.
106 */
107 struct throtl_qnode qnode_on_self[2];
108 struct throtl_qnode qnode_on_parent[2];
109
110 /*
111 * Dispatch time in jiffies. This is the estimated time when group
112 * will unthrottle and is ready to dispatch more bio. It is used as
113 * key to sort active groups in service tree.
114 */
115 unsigned long disptime;
116
117 unsigned int flags;
118
119 /* are there any throtl rules between this group and td? */
120 bool has_rules[2];
121
122 /* bytes per second rate limits */
123 uint64_t bps[2];
124
125 /* IOPS limits */
126 unsigned int iops[2];
127
128 /* Number of bytes disptached in current slice */
129 uint64_t bytes_disp[2];
130 /* Number of bio's dispatched in current slice */
131 unsigned int io_disp[2];
132
133 /* When did we start a new slice */
134 unsigned long slice_start[2];
135 unsigned long slice_end[2];
136};
137
138struct throtl_data
139{
140 /* service tree for active throtl groups */
141 struct throtl_service_queue service_queue;
142
143 struct request_queue *queue;
144
145 /* Total Number of queued bios on READ and WRITE lists */
146 unsigned int nr_queued[2];
147
148 /* Work for dispatching throttled bios */
149 struct work_struct dispatch_work;
150};
151
152static void throtl_pending_timer_fn(unsigned long arg);
153
154static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
155{
156 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
157}
158
159static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
160{
161 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
162}
163
164static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
165{
166 return pd_to_blkg(&tg->pd);
167}
168
169/**
170 * sq_to_tg - return the throl_grp the specified service queue belongs to
171 * @sq: the throtl_service_queue of interest
172 *
173 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
174 * embedded in throtl_data, %NULL is returned.
175 */
176static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
177{
178 if (sq && sq->parent_sq)
179 return container_of(sq, struct throtl_grp, service_queue);
180 else
181 return NULL;
182}
183
184/**
185 * sq_to_td - return throtl_data the specified service queue belongs to
186 * @sq: the throtl_service_queue of interest
187 *
188 * A service_queue can be embeded in either a throtl_grp or throtl_data.
189 * Determine the associated throtl_data accordingly and return it.
190 */
191static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
192{
193 struct throtl_grp *tg = sq_to_tg(sq);
194
195 if (tg)
196 return tg->td;
197 else
198 return container_of(sq, struct throtl_data, service_queue);
199}
200
201/**
202 * throtl_log - log debug message via blktrace
203 * @sq: the service_queue being reported
204 * @fmt: printf format string
205 * @args: printf args
206 *
207 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
208 * throtl_grp; otherwise, just "throtl".
209 */
210#define throtl_log(sq, fmt, args...) do { \
211 struct throtl_grp *__tg = sq_to_tg((sq)); \
212 struct throtl_data *__td = sq_to_td((sq)); \
213 \
214 (void)__td; \
215 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
216 break; \
217 if ((__tg)) { \
218 char __pbuf[128]; \
219 \
220 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
221 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
222 } else { \
223 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
224 } \
225} while (0)
226
227static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
228{
229 INIT_LIST_HEAD(&qn->node);
230 bio_list_init(&qn->bios);
231 qn->tg = tg;
232}
233
234/**
235 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
236 * @bio: bio being added
237 * @qn: qnode to add bio to
238 * @queued: the service_queue->queued[] list @qn belongs to
239 *
240 * Add @bio to @qn and put @qn on @queued if it's not already on.
241 * @qn->tg's reference count is bumped when @qn is activated. See the
242 * comment on top of throtl_qnode definition for details.
243 */
244static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
245 struct list_head *queued)
246{
247 bio_list_add(&qn->bios, bio);
248 if (list_empty(&qn->node)) {
249 list_add_tail(&qn->node, queued);
250 blkg_get(tg_to_blkg(qn->tg));
251 }
252}
253
254/**
255 * throtl_peek_queued - peek the first bio on a qnode list
256 * @queued: the qnode list to peek
257 */
258static struct bio *throtl_peek_queued(struct list_head *queued)
259{
260 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
261 struct bio *bio;
262
263 if (list_empty(queued))
264 return NULL;
265
266 bio = bio_list_peek(&qn->bios);
267 WARN_ON_ONCE(!bio);
268 return bio;
269}
270
271/**
272 * throtl_pop_queued - pop the first bio form a qnode list
273 * @queued: the qnode list to pop a bio from
274 * @tg_to_put: optional out argument for throtl_grp to put
275 *
276 * Pop the first bio from the qnode list @queued. After popping, the first
277 * qnode is removed from @queued if empty or moved to the end of @queued so
278 * that the popping order is round-robin.
279 *
280 * When the first qnode is removed, its associated throtl_grp should be put
281 * too. If @tg_to_put is NULL, this function automatically puts it;
282 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
283 * responsible for putting it.
284 */
285static struct bio *throtl_pop_queued(struct list_head *queued,
286 struct throtl_grp **tg_to_put)
287{
288 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
289 struct bio *bio;
290
291 if (list_empty(queued))
292 return NULL;
293
294 bio = bio_list_pop(&qn->bios);
295 WARN_ON_ONCE(!bio);
296
297 if (bio_list_empty(&qn->bios)) {
298 list_del_init(&qn->node);
299 if (tg_to_put)
300 *tg_to_put = qn->tg;
301 else
302 blkg_put(tg_to_blkg(qn->tg));
303 } else {
304 list_move_tail(&qn->node, queued);
305 }
306
307 return bio;
308}
309
310/* init a service_queue, assumes the caller zeroed it */
311static void throtl_service_queue_init(struct throtl_service_queue *sq)
312{
313 INIT_LIST_HEAD(&sq->queued[0]);
314 INIT_LIST_HEAD(&sq->queued[1]);
315 sq->pending_tree = RB_ROOT;
316 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
317 (unsigned long)sq);
318}
319
320static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
321{
322 struct throtl_grp *tg;
323 int rw;
324
325 tg = kzalloc_node(sizeof(*tg), gfp, node);
326 if (!tg)
327 return NULL;
328
329 throtl_service_queue_init(&tg->service_queue);
330
331 for (rw = READ; rw <= WRITE; rw++) {
332 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
333 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
334 }
335
336 RB_CLEAR_NODE(&tg->rb_node);
337 tg->bps[READ] = -1;
338 tg->bps[WRITE] = -1;
339 tg->iops[READ] = -1;
340 tg->iops[WRITE] = -1;
341
342 return &tg->pd;
343}
344
345static void throtl_pd_init(struct blkg_policy_data *pd)
346{
347 struct throtl_grp *tg = pd_to_tg(pd);
348 struct blkcg_gq *blkg = tg_to_blkg(tg);
349 struct throtl_data *td = blkg->q->td;
350 struct throtl_service_queue *sq = &tg->service_queue;
351
352 /*
353 * If on the default hierarchy, we switch to properly hierarchical
354 * behavior where limits on a given throtl_grp are applied to the
355 * whole subtree rather than just the group itself. e.g. If 16M
356 * read_bps limit is set on the root group, the whole system can't
357 * exceed 16M for the device.
358 *
359 * If not on the default hierarchy, the broken flat hierarchy
360 * behavior is retained where all throtl_grps are treated as if
361 * they're all separate root groups right below throtl_data.
362 * Limits of a group don't interact with limits of other groups
363 * regardless of the position of the group in the hierarchy.
364 */
365 sq->parent_sq = &td->service_queue;
366 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
367 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
368 tg->td = td;
369}
370
371/*
372 * Set has_rules[] if @tg or any of its parents have limits configured.
373 * This doesn't require walking up to the top of the hierarchy as the
374 * parent's has_rules[] is guaranteed to be correct.
375 */
376static void tg_update_has_rules(struct throtl_grp *tg)
377{
378 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
379 int rw;
380
381 for (rw = READ; rw <= WRITE; rw++)
382 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
383 (tg->bps[rw] != -1 || tg->iops[rw] != -1);
384}
385
386static void throtl_pd_online(struct blkg_policy_data *pd)
387{
388 /*
389 * We don't want new groups to escape the limits of its ancestors.
390 * Update has_rules[] after a new group is brought online.
391 */
392 tg_update_has_rules(pd_to_tg(pd));
393}
394
395static void throtl_pd_free(struct blkg_policy_data *pd)
396{
397 struct throtl_grp *tg = pd_to_tg(pd);
398
399 del_timer_sync(&tg->service_queue.pending_timer);
400 kfree(tg);
401}
402
403static struct throtl_grp *
404throtl_rb_first(struct throtl_service_queue *parent_sq)
405{
406 /* Service tree is empty */
407 if (!parent_sq->nr_pending)
408 return NULL;
409
410 if (!parent_sq->first_pending)
411 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
412
413 if (parent_sq->first_pending)
414 return rb_entry_tg(parent_sq->first_pending);
415
416 return NULL;
417}
418
419static void rb_erase_init(struct rb_node *n, struct rb_root *root)
420{
421 rb_erase(n, root);
422 RB_CLEAR_NODE(n);
423}
424
425static void throtl_rb_erase(struct rb_node *n,
426 struct throtl_service_queue *parent_sq)
427{
428 if (parent_sq->first_pending == n)
429 parent_sq->first_pending = NULL;
430 rb_erase_init(n, &parent_sq->pending_tree);
431 --parent_sq->nr_pending;
432}
433
434static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
435{
436 struct throtl_grp *tg;
437
438 tg = throtl_rb_first(parent_sq);
439 if (!tg)
440 return;
441
442 parent_sq->first_pending_disptime = tg->disptime;
443}
444
445static void tg_service_queue_add(struct throtl_grp *tg)
446{
447 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
448 struct rb_node **node = &parent_sq->pending_tree.rb_node;
449 struct rb_node *parent = NULL;
450 struct throtl_grp *__tg;
451 unsigned long key = tg->disptime;
452 int left = 1;
453
454 while (*node != NULL) {
455 parent = *node;
456 __tg = rb_entry_tg(parent);
457
458 if (time_before(key, __tg->disptime))
459 node = &parent->rb_left;
460 else {
461 node = &parent->rb_right;
462 left = 0;
463 }
464 }
465
466 if (left)
467 parent_sq->first_pending = &tg->rb_node;
468
469 rb_link_node(&tg->rb_node, parent, node);
470 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
471}
472
473static void __throtl_enqueue_tg(struct throtl_grp *tg)
474{
475 tg_service_queue_add(tg);
476 tg->flags |= THROTL_TG_PENDING;
477 tg->service_queue.parent_sq->nr_pending++;
478}
479
480static void throtl_enqueue_tg(struct throtl_grp *tg)
481{
482 if (!(tg->flags & THROTL_TG_PENDING))
483 __throtl_enqueue_tg(tg);
484}
485
486static void __throtl_dequeue_tg(struct throtl_grp *tg)
487{
488 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
489 tg->flags &= ~THROTL_TG_PENDING;
490}
491
492static void throtl_dequeue_tg(struct throtl_grp *tg)
493{
494 if (tg->flags & THROTL_TG_PENDING)
495 __throtl_dequeue_tg(tg);
496}
497
498/* Call with queue lock held */
499static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
500 unsigned long expires)
501{
502 mod_timer(&sq->pending_timer, expires);
503 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
504 expires - jiffies, jiffies);
505}
506
507/**
508 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
509 * @sq: the service_queue to schedule dispatch for
510 * @force: force scheduling
511 *
512 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
513 * dispatch time of the first pending child. Returns %true if either timer
514 * is armed or there's no pending child left. %false if the current
515 * dispatch window is still open and the caller should continue
516 * dispatching.
517 *
518 * If @force is %true, the dispatch timer is always scheduled and this
519 * function is guaranteed to return %true. This is to be used when the
520 * caller can't dispatch itself and needs to invoke pending_timer
521 * unconditionally. Note that forced scheduling is likely to induce short
522 * delay before dispatch starts even if @sq->first_pending_disptime is not
523 * in the future and thus shouldn't be used in hot paths.
524 */
525static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
526 bool force)
527{
528 /* any pending children left? */
529 if (!sq->nr_pending)
530 return true;
531
532 update_min_dispatch_time(sq);
533
534 /* is the next dispatch time in the future? */
535 if (force || time_after(sq->first_pending_disptime, jiffies)) {
536 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
537 return true;
538 }
539
540 /* tell the caller to continue dispatching */
541 return false;
542}
543
544static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
545 bool rw, unsigned long start)
546{
547 tg->bytes_disp[rw] = 0;
548 tg->io_disp[rw] = 0;
549
550 /*
551 * Previous slice has expired. We must have trimmed it after last
552 * bio dispatch. That means since start of last slice, we never used
553 * that bandwidth. Do try to make use of that bandwidth while giving
554 * credit.
555 */
556 if (time_after_eq(start, tg->slice_start[rw]))
557 tg->slice_start[rw] = start;
558
559 tg->slice_end[rw] = jiffies + throtl_slice;
560 throtl_log(&tg->service_queue,
561 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
562 rw == READ ? 'R' : 'W', tg->slice_start[rw],
563 tg->slice_end[rw], jiffies);
564}
565
566static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
567{
568 tg->bytes_disp[rw] = 0;
569 tg->io_disp[rw] = 0;
570 tg->slice_start[rw] = jiffies;
571 tg->slice_end[rw] = jiffies + throtl_slice;
572 throtl_log(&tg->service_queue,
573 "[%c] new slice start=%lu end=%lu jiffies=%lu",
574 rw == READ ? 'R' : 'W', tg->slice_start[rw],
575 tg->slice_end[rw], jiffies);
576}
577
578static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
579 unsigned long jiffy_end)
580{
581 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
582}
583
584static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
585 unsigned long jiffy_end)
586{
587 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
588 throtl_log(&tg->service_queue,
589 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
590 rw == READ ? 'R' : 'W', tg->slice_start[rw],
591 tg->slice_end[rw], jiffies);
592}
593
594/* Determine if previously allocated or extended slice is complete or not */
595static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
596{
597 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
598 return false;
599
600 return 1;
601}
602
603/* Trim the used slices and adjust slice start accordingly */
604static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
605{
606 unsigned long nr_slices, time_elapsed, io_trim;
607 u64 bytes_trim, tmp;
608
609 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
610
611 /*
612 * If bps are unlimited (-1), then time slice don't get
613 * renewed. Don't try to trim the slice if slice is used. A new
614 * slice will start when appropriate.
615 */
616 if (throtl_slice_used(tg, rw))
617 return;
618
619 /*
620 * A bio has been dispatched. Also adjust slice_end. It might happen
621 * that initially cgroup limit was very low resulting in high
622 * slice_end, but later limit was bumped up and bio was dispached
623 * sooner, then we need to reduce slice_end. A high bogus slice_end
624 * is bad because it does not allow new slice to start.
625 */
626
627 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
628
629 time_elapsed = jiffies - tg->slice_start[rw];
630
631 nr_slices = time_elapsed / throtl_slice;
632
633 if (!nr_slices)
634 return;
635 tmp = tg->bps[rw] * throtl_slice * nr_slices;
636 do_div(tmp, HZ);
637 bytes_trim = tmp;
638
639 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
640
641 if (!bytes_trim && !io_trim)
642 return;
643
644 if (tg->bytes_disp[rw] >= bytes_trim)
645 tg->bytes_disp[rw] -= bytes_trim;
646 else
647 tg->bytes_disp[rw] = 0;
648
649 if (tg->io_disp[rw] >= io_trim)
650 tg->io_disp[rw] -= io_trim;
651 else
652 tg->io_disp[rw] = 0;
653
654 tg->slice_start[rw] += nr_slices * throtl_slice;
655
656 throtl_log(&tg->service_queue,
657 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
658 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
659 tg->slice_start[rw], tg->slice_end[rw], jiffies);
660}
661
662static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
663 unsigned long *wait)
664{
665 bool rw = bio_data_dir(bio);
666 unsigned int io_allowed;
667 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
668 u64 tmp;
669
670 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
671
672 /* Slice has just started. Consider one slice interval */
673 if (!jiffy_elapsed)
674 jiffy_elapsed_rnd = throtl_slice;
675
676 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
677
678 /*
679 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
680 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
681 * will allow dispatch after 1 second and after that slice should
682 * have been trimmed.
683 */
684
685 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
686 do_div(tmp, HZ);
687
688 if (tmp > UINT_MAX)
689 io_allowed = UINT_MAX;
690 else
691 io_allowed = tmp;
692
693 if (tg->io_disp[rw] + 1 <= io_allowed) {
694 if (wait)
695 *wait = 0;
696 return true;
697 }
698
699 /* Calc approx time to dispatch */
700 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
701
702 if (jiffy_wait > jiffy_elapsed)
703 jiffy_wait = jiffy_wait - jiffy_elapsed;
704 else
705 jiffy_wait = 1;
706
707 if (wait)
708 *wait = jiffy_wait;
709 return 0;
710}
711
712static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
713 unsigned long *wait)
714{
715 bool rw = bio_data_dir(bio);
716 u64 bytes_allowed, extra_bytes, tmp;
717 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
718
719 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
720
721 /* Slice has just started. Consider one slice interval */
722 if (!jiffy_elapsed)
723 jiffy_elapsed_rnd = throtl_slice;
724
725 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
726
727 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
728 do_div(tmp, HZ);
729 bytes_allowed = tmp;
730
731 if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
732 if (wait)
733 *wait = 0;
734 return true;
735 }
736
737 /* Calc approx time to dispatch */
738 extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
739 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
740
741 if (!jiffy_wait)
742 jiffy_wait = 1;
743
744 /*
745 * This wait time is without taking into consideration the rounding
746 * up we did. Add that time also.
747 */
748 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
749 if (wait)
750 *wait = jiffy_wait;
751 return 0;
752}
753
754/*
755 * Returns whether one can dispatch a bio or not. Also returns approx number
756 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
757 */
758static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
759 unsigned long *wait)
760{
761 bool rw = bio_data_dir(bio);
762 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
763
764 /*
765 * Currently whole state machine of group depends on first bio
766 * queued in the group bio list. So one should not be calling
767 * this function with a different bio if there are other bios
768 * queued.
769 */
770 BUG_ON(tg->service_queue.nr_queued[rw] &&
771 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
772
773 /* If tg->bps = -1, then BW is unlimited */
774 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
775 if (wait)
776 *wait = 0;
777 return true;
778 }
779
780 /*
781 * If previous slice expired, start a new one otherwise renew/extend
782 * existing slice to make sure it is at least throtl_slice interval
783 * long since now. New slice is started only for empty throttle group.
784 * If there is queued bio, that means there should be an active
785 * slice and it should be extended instead.
786 */
787 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
788 throtl_start_new_slice(tg, rw);
789 else {
790 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
791 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
792 }
793
794 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
795 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
796 if (wait)
797 *wait = 0;
798 return 1;
799 }
800
801 max_wait = max(bps_wait, iops_wait);
802
803 if (wait)
804 *wait = max_wait;
805
806 if (time_before(tg->slice_end[rw], jiffies + max_wait))
807 throtl_extend_slice(tg, rw, jiffies + max_wait);
808
809 return 0;
810}
811
812static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
813{
814 bool rw = bio_data_dir(bio);
815
816 /* Charge the bio to the group */
817 tg->bytes_disp[rw] += bio->bi_iter.bi_size;
818 tg->io_disp[rw]++;
819
820 /*
821 * BIO_THROTTLED is used to prevent the same bio to be throttled
822 * more than once as a throttled bio will go through blk-throtl the
823 * second time when it eventually gets issued. Set it when a bio
824 * is being charged to a tg.
825 */
826 if (!bio_flagged(bio, BIO_THROTTLED))
827 bio_set_flag(bio, BIO_THROTTLED);
828}
829
830/**
831 * throtl_add_bio_tg - add a bio to the specified throtl_grp
832 * @bio: bio to add
833 * @qn: qnode to use
834 * @tg: the target throtl_grp
835 *
836 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
837 * tg->qnode_on_self[] is used.
838 */
839static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
840 struct throtl_grp *tg)
841{
842 struct throtl_service_queue *sq = &tg->service_queue;
843 bool rw = bio_data_dir(bio);
844
845 if (!qn)
846 qn = &tg->qnode_on_self[rw];
847
848 /*
849 * If @tg doesn't currently have any bios queued in the same
850 * direction, queueing @bio can change when @tg should be
851 * dispatched. Mark that @tg was empty. This is automatically
852 * cleaered on the next tg_update_disptime().
853 */
854 if (!sq->nr_queued[rw])
855 tg->flags |= THROTL_TG_WAS_EMPTY;
856
857 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
858
859 sq->nr_queued[rw]++;
860 throtl_enqueue_tg(tg);
861}
862
863static void tg_update_disptime(struct throtl_grp *tg)
864{
865 struct throtl_service_queue *sq = &tg->service_queue;
866 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
867 struct bio *bio;
868
869 if ((bio = throtl_peek_queued(&sq->queued[READ])))
870 tg_may_dispatch(tg, bio, &read_wait);
871
872 if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
873 tg_may_dispatch(tg, bio, &write_wait);
874
875 min_wait = min(read_wait, write_wait);
876 disptime = jiffies + min_wait;
877
878 /* Update dispatch time */
879 throtl_dequeue_tg(tg);
880 tg->disptime = disptime;
881 throtl_enqueue_tg(tg);
882
883 /* see throtl_add_bio_tg() */
884 tg->flags &= ~THROTL_TG_WAS_EMPTY;
885}
886
887static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
888 struct throtl_grp *parent_tg, bool rw)
889{
890 if (throtl_slice_used(parent_tg, rw)) {
891 throtl_start_new_slice_with_credit(parent_tg, rw,
892 child_tg->slice_start[rw]);
893 }
894
895}
896
897static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
898{
899 struct throtl_service_queue *sq = &tg->service_queue;
900 struct throtl_service_queue *parent_sq = sq->parent_sq;
901 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
902 struct throtl_grp *tg_to_put = NULL;
903 struct bio *bio;
904
905 /*
906 * @bio is being transferred from @tg to @parent_sq. Popping a bio
907 * from @tg may put its reference and @parent_sq might end up
908 * getting released prematurely. Remember the tg to put and put it
909 * after @bio is transferred to @parent_sq.
910 */
911 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
912 sq->nr_queued[rw]--;
913
914 throtl_charge_bio(tg, bio);
915
916 /*
917 * If our parent is another tg, we just need to transfer @bio to
918 * the parent using throtl_add_bio_tg(). If our parent is
919 * @td->service_queue, @bio is ready to be issued. Put it on its
920 * bio_lists[] and decrease total number queued. The caller is
921 * responsible for issuing these bios.
922 */
923 if (parent_tg) {
924 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
925 start_parent_slice_with_credit(tg, parent_tg, rw);
926 } else {
927 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
928 &parent_sq->queued[rw]);
929 BUG_ON(tg->td->nr_queued[rw] <= 0);
930 tg->td->nr_queued[rw]--;
931 }
932
933 throtl_trim_slice(tg, rw);
934
935 if (tg_to_put)
936 blkg_put(tg_to_blkg(tg_to_put));
937}
938
939static int throtl_dispatch_tg(struct throtl_grp *tg)
940{
941 struct throtl_service_queue *sq = &tg->service_queue;
942 unsigned int nr_reads = 0, nr_writes = 0;
943 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
944 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
945 struct bio *bio;
946
947 /* Try to dispatch 75% READS and 25% WRITES */
948
949 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
950 tg_may_dispatch(tg, bio, NULL)) {
951
952 tg_dispatch_one_bio(tg, bio_data_dir(bio));
953 nr_reads++;
954
955 if (nr_reads >= max_nr_reads)
956 break;
957 }
958
959 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
960 tg_may_dispatch(tg, bio, NULL)) {
961
962 tg_dispatch_one_bio(tg, bio_data_dir(bio));
963 nr_writes++;
964
965 if (nr_writes >= max_nr_writes)
966 break;
967 }
968
969 return nr_reads + nr_writes;
970}
971
972static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
973{
974 unsigned int nr_disp = 0;
975
976 while (1) {
977 struct throtl_grp *tg = throtl_rb_first(parent_sq);
978 struct throtl_service_queue *sq = &tg->service_queue;
979
980 if (!tg)
981 break;
982
983 if (time_before(jiffies, tg->disptime))
984 break;
985
986 throtl_dequeue_tg(tg);
987
988 nr_disp += throtl_dispatch_tg(tg);
989
990 if (sq->nr_queued[0] || sq->nr_queued[1])
991 tg_update_disptime(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 * @arg: 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(unsigned long arg)
1016{
1017 struct throtl_service_queue *sq = (void *)arg;
1018 struct throtl_grp *tg = sq_to_tg(sq);
1019 struct throtl_data *td = sq_to_td(sq);
1020 struct request_queue *q = td->queue;
1021 struct throtl_service_queue *parent_sq;
1022 bool dispatched;
1023 int ret;
1024
1025 spin_lock_irq(q->queue_lock);
1026again:
1027 parent_sq = sq->parent_sq;
1028 dispatched = false;
1029
1030 while (true) {
1031 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1032 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1033 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1034
1035 ret = throtl_select_dispatch(sq);
1036 if (ret) {
1037 throtl_log(sq, "bios disp=%u", ret);
1038 dispatched = true;
1039 }
1040
1041 if (throtl_schedule_next_dispatch(sq, false))
1042 break;
1043
1044 /* this dispatch windows is still open, relax and repeat */
1045 spin_unlock_irq(q->queue_lock);
1046 cpu_relax();
1047 spin_lock_irq(q->queue_lock);
1048 }
1049
1050 if (!dispatched)
1051 goto out_unlock;
1052
1053 if (parent_sq) {
1054 /* @parent_sq is another throl_grp, propagate dispatch */
1055 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1056 tg_update_disptime(tg);
1057 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1058 /* window is already open, repeat dispatching */
1059 sq = parent_sq;
1060 tg = sq_to_tg(sq);
1061 goto again;
1062 }
1063 }
1064 } else {
1065 /* reached the top-level, queue issueing */
1066 queue_work(kthrotld_workqueue, &td->dispatch_work);
1067 }
1068out_unlock:
1069 spin_unlock_irq(q->queue_lock);
1070}
1071
1072/**
1073 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1074 * @work: work item being executed
1075 *
1076 * This function is queued for execution when bio's reach the bio_lists[]
1077 * of throtl_data->service_queue. Those bio's are ready and issued by this
1078 * function.
1079 */
1080static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1081{
1082 struct throtl_data *td = container_of(work, struct throtl_data,
1083 dispatch_work);
1084 struct throtl_service_queue *td_sq = &td->service_queue;
1085 struct request_queue *q = td->queue;
1086 struct bio_list bio_list_on_stack;
1087 struct bio *bio;
1088 struct blk_plug plug;
1089 int rw;
1090
1091 bio_list_init(&bio_list_on_stack);
1092
1093 spin_lock_irq(q->queue_lock);
1094 for (rw = READ; rw <= WRITE; rw++)
1095 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1096 bio_list_add(&bio_list_on_stack, bio);
1097 spin_unlock_irq(q->queue_lock);
1098
1099 if (!bio_list_empty(&bio_list_on_stack)) {
1100 blk_start_plug(&plug);
1101 while((bio = bio_list_pop(&bio_list_on_stack)))
1102 generic_make_request(bio);
1103 blk_finish_plug(&plug);
1104 }
1105}
1106
1107static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1108 int off)
1109{
1110 struct throtl_grp *tg = pd_to_tg(pd);
1111 u64 v = *(u64 *)((void *)tg + off);
1112
1113 if (v == -1)
1114 return 0;
1115 return __blkg_prfill_u64(sf, pd, v);
1116}
1117
1118static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1119 int off)
1120{
1121 struct throtl_grp *tg = pd_to_tg(pd);
1122 unsigned int v = *(unsigned int *)((void *)tg + off);
1123
1124 if (v == -1)
1125 return 0;
1126 return __blkg_prfill_u64(sf, pd, v);
1127}
1128
1129static int tg_print_conf_u64(struct seq_file *sf, void *v)
1130{
1131 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1132 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1133 return 0;
1134}
1135
1136static int tg_print_conf_uint(struct seq_file *sf, void *v)
1137{
1138 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1139 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1140 return 0;
1141}
1142
1143static void tg_conf_updated(struct throtl_grp *tg)
1144{
1145 struct throtl_service_queue *sq = &tg->service_queue;
1146 struct cgroup_subsys_state *pos_css;
1147 struct blkcg_gq *blkg;
1148
1149 throtl_log(&tg->service_queue,
1150 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1151 tg->bps[READ], tg->bps[WRITE],
1152 tg->iops[READ], tg->iops[WRITE]);
1153
1154 /*
1155 * Update has_rules[] flags for the updated tg's subtree. A tg is
1156 * considered to have rules if either the tg itself or any of its
1157 * ancestors has rules. This identifies groups without any
1158 * restrictions in the whole hierarchy and allows them to bypass
1159 * blk-throttle.
1160 */
1161 blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg))
1162 tg_update_has_rules(blkg_to_tg(blkg));
1163
1164 /*
1165 * We're already holding queue_lock and know @tg is valid. Let's
1166 * apply the new config directly.
1167 *
1168 * Restart the slices for both READ and WRITES. It might happen
1169 * that a group's limit are dropped suddenly and we don't want to
1170 * account recently dispatched IO with new low rate.
1171 */
1172 throtl_start_new_slice(tg, 0);
1173 throtl_start_new_slice(tg, 1);
1174
1175 if (tg->flags & THROTL_TG_PENDING) {
1176 tg_update_disptime(tg);
1177 throtl_schedule_next_dispatch(sq->parent_sq, true);
1178 }
1179}
1180
1181static ssize_t tg_set_conf(struct kernfs_open_file *of,
1182 char *buf, size_t nbytes, loff_t off, bool is_u64)
1183{
1184 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1185 struct blkg_conf_ctx ctx;
1186 struct throtl_grp *tg;
1187 int ret;
1188 u64 v;
1189
1190 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1191 if (ret)
1192 return ret;
1193
1194 ret = -EINVAL;
1195 if (sscanf(ctx.body, "%llu", &v) != 1)
1196 goto out_finish;
1197 if (!v)
1198 v = -1;
1199
1200 tg = blkg_to_tg(ctx.blkg);
1201
1202 if (is_u64)
1203 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1204 else
1205 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1206
1207 tg_conf_updated(tg);
1208 ret = 0;
1209out_finish:
1210 blkg_conf_finish(&ctx);
1211 return ret ?: nbytes;
1212}
1213
1214static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1215 char *buf, size_t nbytes, loff_t off)
1216{
1217 return tg_set_conf(of, buf, nbytes, off, true);
1218}
1219
1220static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1221 char *buf, size_t nbytes, loff_t off)
1222{
1223 return tg_set_conf(of, buf, nbytes, off, false);
1224}
1225
1226static struct cftype throtl_legacy_files[] = {
1227 {
1228 .name = "throttle.read_bps_device",
1229 .private = offsetof(struct throtl_grp, bps[READ]),
1230 .seq_show = tg_print_conf_u64,
1231 .write = tg_set_conf_u64,
1232 },
1233 {
1234 .name = "throttle.write_bps_device",
1235 .private = offsetof(struct throtl_grp, bps[WRITE]),
1236 .seq_show = tg_print_conf_u64,
1237 .write = tg_set_conf_u64,
1238 },
1239 {
1240 .name = "throttle.read_iops_device",
1241 .private = offsetof(struct throtl_grp, iops[READ]),
1242 .seq_show = tg_print_conf_uint,
1243 .write = tg_set_conf_uint,
1244 },
1245 {
1246 .name = "throttle.write_iops_device",
1247 .private = offsetof(struct throtl_grp, iops[WRITE]),
1248 .seq_show = tg_print_conf_uint,
1249 .write = tg_set_conf_uint,
1250 },
1251 {
1252 .name = "throttle.io_service_bytes",
1253 .private = (unsigned long)&blkcg_policy_throtl,
1254 .seq_show = blkg_print_stat_bytes,
1255 },
1256 {
1257 .name = "throttle.io_serviced",
1258 .private = (unsigned long)&blkcg_policy_throtl,
1259 .seq_show = blkg_print_stat_ios,
1260 },
1261 { } /* terminate */
1262};
1263
1264static u64 tg_prfill_max(struct seq_file *sf, struct blkg_policy_data *pd,
1265 int off)
1266{
1267 struct throtl_grp *tg = pd_to_tg(pd);
1268 const char *dname = blkg_dev_name(pd->blkg);
1269 char bufs[4][21] = { "max", "max", "max", "max" };
1270
1271 if (!dname)
1272 return 0;
1273 if (tg->bps[READ] == -1 && tg->bps[WRITE] == -1 &&
1274 tg->iops[READ] == -1 && tg->iops[WRITE] == -1)
1275 return 0;
1276
1277 if (tg->bps[READ] != -1)
1278 snprintf(bufs[0], sizeof(bufs[0]), "%llu", tg->bps[READ]);
1279 if (tg->bps[WRITE] != -1)
1280 snprintf(bufs[1], sizeof(bufs[1]), "%llu", tg->bps[WRITE]);
1281 if (tg->iops[READ] != -1)
1282 snprintf(bufs[2], sizeof(bufs[2]), "%u", tg->iops[READ]);
1283 if (tg->iops[WRITE] != -1)
1284 snprintf(bufs[3], sizeof(bufs[3]), "%u", tg->iops[WRITE]);
1285
1286 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s\n",
1287 dname, bufs[0], bufs[1], bufs[2], bufs[3]);
1288 return 0;
1289}
1290
1291static int tg_print_max(struct seq_file *sf, void *v)
1292{
1293 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_max,
1294 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1295 return 0;
1296}
1297
1298static ssize_t tg_set_max(struct kernfs_open_file *of,
1299 char *buf, size_t nbytes, loff_t off)
1300{
1301 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1302 struct blkg_conf_ctx ctx;
1303 struct throtl_grp *tg;
1304 u64 v[4];
1305 int ret;
1306
1307 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1308 if (ret)
1309 return ret;
1310
1311 tg = blkg_to_tg(ctx.blkg);
1312
1313 v[0] = tg->bps[READ];
1314 v[1] = tg->bps[WRITE];
1315 v[2] = tg->iops[READ];
1316 v[3] = tg->iops[WRITE];
1317
1318 while (true) {
1319 char tok[27]; /* wiops=18446744073709551616 */
1320 char *p;
1321 u64 val = -1;
1322 int len;
1323
1324 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1325 break;
1326 if (tok[0] == '\0')
1327 break;
1328 ctx.body += len;
1329
1330 ret = -EINVAL;
1331 p = tok;
1332 strsep(&p, "=");
1333 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1334 goto out_finish;
1335
1336 ret = -ERANGE;
1337 if (!val)
1338 goto out_finish;
1339
1340 ret = -EINVAL;
1341 if (!strcmp(tok, "rbps"))
1342 v[0] = val;
1343 else if (!strcmp(tok, "wbps"))
1344 v[1] = val;
1345 else if (!strcmp(tok, "riops"))
1346 v[2] = min_t(u64, val, UINT_MAX);
1347 else if (!strcmp(tok, "wiops"))
1348 v[3] = min_t(u64, val, UINT_MAX);
1349 else
1350 goto out_finish;
1351 }
1352
1353 tg->bps[READ] = v[0];
1354 tg->bps[WRITE] = v[1];
1355 tg->iops[READ] = v[2];
1356 tg->iops[WRITE] = v[3];
1357
1358 tg_conf_updated(tg);
1359 ret = 0;
1360out_finish:
1361 blkg_conf_finish(&ctx);
1362 return ret ?: nbytes;
1363}
1364
1365static struct cftype throtl_files[] = {
1366 {
1367 .name = "max",
1368 .flags = CFTYPE_NOT_ON_ROOT,
1369 .seq_show = tg_print_max,
1370 .write = tg_set_max,
1371 },
1372 { } /* terminate */
1373};
1374
1375static void throtl_shutdown_wq(struct request_queue *q)
1376{
1377 struct throtl_data *td = q->td;
1378
1379 cancel_work_sync(&td->dispatch_work);
1380}
1381
1382static struct blkcg_policy blkcg_policy_throtl = {
1383 .dfl_cftypes = throtl_files,
1384 .legacy_cftypes = throtl_legacy_files,
1385
1386 .pd_alloc_fn = throtl_pd_alloc,
1387 .pd_init_fn = throtl_pd_init,
1388 .pd_online_fn = throtl_pd_online,
1389 .pd_free_fn = throtl_pd_free,
1390};
1391
1392bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
1393 struct bio *bio)
1394{
1395 struct throtl_qnode *qn = NULL;
1396 struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
1397 struct throtl_service_queue *sq;
1398 bool rw = bio_data_dir(bio);
1399 bool throttled = false;
1400
1401 WARN_ON_ONCE(!rcu_read_lock_held());
1402
1403 /* see throtl_charge_bio() */
1404 if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw])
1405 goto out;
1406
1407 spin_lock_irq(q->queue_lock);
1408
1409 if (unlikely(blk_queue_bypass(q)))
1410 goto out_unlock;
1411
1412 sq = &tg->service_queue;
1413
1414 while (true) {
1415 /* throtl is FIFO - if bios are already queued, should queue */
1416 if (sq->nr_queued[rw])
1417 break;
1418
1419 /* if above limits, break to queue */
1420 if (!tg_may_dispatch(tg, bio, NULL))
1421 break;
1422
1423 /* within limits, let's charge and dispatch directly */
1424 throtl_charge_bio(tg, bio);
1425
1426 /*
1427 * We need to trim slice even when bios are not being queued
1428 * otherwise it might happen that a bio is not queued for
1429 * a long time and slice keeps on extending and trim is not
1430 * called for a long time. Now if limits are reduced suddenly
1431 * we take into account all the IO dispatched so far at new
1432 * low rate and * newly queued IO gets a really long dispatch
1433 * time.
1434 *
1435 * So keep on trimming slice even if bio is not queued.
1436 */
1437 throtl_trim_slice(tg, rw);
1438
1439 /*
1440 * @bio passed through this layer without being throttled.
1441 * Climb up the ladder. If we''re already at the top, it
1442 * can be executed directly.
1443 */
1444 qn = &tg->qnode_on_parent[rw];
1445 sq = sq->parent_sq;
1446 tg = sq_to_tg(sq);
1447 if (!tg)
1448 goto out_unlock;
1449 }
1450
1451 /* out-of-limit, queue to @tg */
1452 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1453 rw == READ ? 'R' : 'W',
1454 tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw],
1455 tg->io_disp[rw], tg->iops[rw],
1456 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1457
1458 bio_associate_current(bio);
1459 tg->td->nr_queued[rw]++;
1460 throtl_add_bio_tg(bio, qn, tg);
1461 throttled = true;
1462
1463 /*
1464 * Update @tg's dispatch time and force schedule dispatch if @tg
1465 * was empty before @bio. The forced scheduling isn't likely to
1466 * cause undue delay as @bio is likely to be dispatched directly if
1467 * its @tg's disptime is not in the future.
1468 */
1469 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1470 tg_update_disptime(tg);
1471 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1472 }
1473
1474out_unlock:
1475 spin_unlock_irq(q->queue_lock);
1476out:
1477 /*
1478 * As multiple blk-throtls may stack in the same issue path, we
1479 * don't want bios to leave with the flag set. Clear the flag if
1480 * being issued.
1481 */
1482 if (!throttled)
1483 bio_clear_flag(bio, BIO_THROTTLED);
1484 return throttled;
1485}
1486
1487/*
1488 * Dispatch all bios from all children tg's queued on @parent_sq. On
1489 * return, @parent_sq is guaranteed to not have any active children tg's
1490 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1491 */
1492static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1493{
1494 struct throtl_grp *tg;
1495
1496 while ((tg = throtl_rb_first(parent_sq))) {
1497 struct throtl_service_queue *sq = &tg->service_queue;
1498 struct bio *bio;
1499
1500 throtl_dequeue_tg(tg);
1501
1502 while ((bio = throtl_peek_queued(&sq->queued[READ])))
1503 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1504 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1505 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1506 }
1507}
1508
1509/**
1510 * blk_throtl_drain - drain throttled bios
1511 * @q: request_queue to drain throttled bios for
1512 *
1513 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1514 */
1515void blk_throtl_drain(struct request_queue *q)
1516 __releases(q->queue_lock) __acquires(q->queue_lock)
1517{
1518 struct throtl_data *td = q->td;
1519 struct blkcg_gq *blkg;
1520 struct cgroup_subsys_state *pos_css;
1521 struct bio *bio;
1522 int rw;
1523
1524 queue_lockdep_assert_held(q);
1525 rcu_read_lock();
1526
1527 /*
1528 * Drain each tg while doing post-order walk on the blkg tree, so
1529 * that all bios are propagated to td->service_queue. It'd be
1530 * better to walk service_queue tree directly but blkg walk is
1531 * easier.
1532 */
1533 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
1534 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1535
1536 /* finally, transfer bios from top-level tg's into the td */
1537 tg_drain_bios(&td->service_queue);
1538
1539 rcu_read_unlock();
1540 spin_unlock_irq(q->queue_lock);
1541
1542 /* all bios now should be in td->service_queue, issue them */
1543 for (rw = READ; rw <= WRITE; rw++)
1544 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1545 NULL)))
1546 generic_make_request(bio);
1547
1548 spin_lock_irq(q->queue_lock);
1549}
1550
1551int blk_throtl_init(struct request_queue *q)
1552{
1553 struct throtl_data *td;
1554 int ret;
1555
1556 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1557 if (!td)
1558 return -ENOMEM;
1559
1560 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1561 throtl_service_queue_init(&td->service_queue);
1562
1563 q->td = td;
1564 td->queue = q;
1565
1566 /* activate policy */
1567 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1568 if (ret)
1569 kfree(td);
1570 return ret;
1571}
1572
1573void blk_throtl_exit(struct request_queue *q)
1574{
1575 BUG_ON(!q->td);
1576 throtl_shutdown_wq(q);
1577 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1578 kfree(q->td);
1579}
1580
1581static int __init throtl_init(void)
1582{
1583 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1584 if (!kthrotld_workqueue)
1585 panic("Failed to create kthrotld\n");
1586
1587 return blkcg_policy_register(&blkcg_policy_throtl);
1588}
1589
1590module_init(throtl_init);