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