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
14/* Max dispatch from a group in 1 round */
15static int throtl_grp_quantum = 8;
16
17/* Total max dispatch from all groups in one round */
18static int throtl_quantum = 32;
19
20/* Throttling is performed over 100ms slice and after that slice is renewed */
21static unsigned long throtl_slice = HZ/10; /* 100 ms */
22
23/* A workqueue to queue throttle related work */
24static struct workqueue_struct *kthrotld_workqueue;
25static void throtl_schedule_delayed_work(struct throtl_data *td,
26 unsigned long delay);
27
28struct throtl_rb_root {
29 struct rb_root rb;
30 struct rb_node *left;
31 unsigned int count;
32 unsigned long min_disptime;
33};
34
35#define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \
36 .count = 0, .min_disptime = 0}
37
38#define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
39
40struct throtl_grp {
41 /* List of throtl groups on the request queue*/
42 struct hlist_node tg_node;
43
44 /* active throtl group service_tree member */
45 struct rb_node rb_node;
46
47 /*
48 * Dispatch time in jiffies. This is the estimated time when group
49 * will unthrottle and is ready to dispatch more bio. It is used as
50 * key to sort active groups in service tree.
51 */
52 unsigned long disptime;
53
54 struct blkio_group blkg;
55 atomic_t ref;
56 unsigned int flags;
57
58 /* Two lists for READ and WRITE */
59 struct bio_list bio_lists[2];
60
61 /* Number of queued bios on READ and WRITE lists */
62 unsigned int nr_queued[2];
63
64 /* bytes per second rate limits */
65 uint64_t bps[2];
66
67 /* IOPS limits */
68 unsigned int iops[2];
69
70 /* Number of bytes disptached in current slice */
71 uint64_t bytes_disp[2];
72 /* Number of bio's dispatched in current slice */
73 unsigned int io_disp[2];
74
75 /* When did we start a new slice */
76 unsigned long slice_start[2];
77 unsigned long slice_end[2];
78
79 /* Some throttle limits got updated for the group */
80 int limits_changed;
81
82 struct rcu_head rcu_head;
83};
84
85struct throtl_data
86{
87 /* List of throtl groups */
88 struct hlist_head tg_list;
89
90 /* service tree for active throtl groups */
91 struct throtl_rb_root tg_service_tree;
92
93 struct throtl_grp *root_tg;
94 struct request_queue *queue;
95
96 /* Total Number of queued bios on READ and WRITE lists */
97 unsigned int nr_queued[2];
98
99 /*
100 * number of total undestroyed groups
101 */
102 unsigned int nr_undestroyed_grps;
103
104 /* Work for dispatching throttled bios */
105 struct delayed_work throtl_work;
106
107 int limits_changed;
108};
109
110enum tg_state_flags {
111 THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */
112};
113
114#define THROTL_TG_FNS(name) \
115static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \
116{ \
117 (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \
118} \
119static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \
120{ \
121 (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \
122} \
123static inline int throtl_tg_##name(const struct throtl_grp *tg) \
124{ \
125 return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \
126}
127
128THROTL_TG_FNS(on_rr);
129
130#define throtl_log_tg(td, tg, fmt, args...) \
131 blk_add_trace_msg((td)->queue, "throtl %s " fmt, \
132 blkg_path(&(tg)->blkg), ##args); \
133
134#define throtl_log(td, fmt, args...) \
135 blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
136
137static inline struct throtl_grp *tg_of_blkg(struct blkio_group *blkg)
138{
139 if (blkg)
140 return container_of(blkg, struct throtl_grp, blkg);
141
142 return NULL;
143}
144
145static inline unsigned int total_nr_queued(struct throtl_data *td)
146{
147 return td->nr_queued[0] + td->nr_queued[1];
148}
149
150static inline struct throtl_grp *throtl_ref_get_tg(struct throtl_grp *tg)
151{
152 atomic_inc(&tg->ref);
153 return tg;
154}
155
156static void throtl_free_tg(struct rcu_head *head)
157{
158 struct throtl_grp *tg;
159
160 tg = container_of(head, struct throtl_grp, rcu_head);
161 free_percpu(tg->blkg.stats_cpu);
162 kfree(tg);
163}
164
165static void throtl_put_tg(struct throtl_grp *tg)
166{
167 BUG_ON(atomic_read(&tg->ref) <= 0);
168 if (!atomic_dec_and_test(&tg->ref))
169 return;
170
171 /*
172 * A group is freed in rcu manner. But having an rcu lock does not
173 * mean that one can access all the fields of blkg and assume these
174 * are valid. For example, don't try to follow throtl_data and
175 * request queue links.
176 *
177 * Having a reference to blkg under an rcu allows acess to only
178 * values local to groups like group stats and group rate limits
179 */
180 call_rcu(&tg->rcu_head, throtl_free_tg);
181}
182
183static void throtl_init_group(struct throtl_grp *tg)
184{
185 INIT_HLIST_NODE(&tg->tg_node);
186 RB_CLEAR_NODE(&tg->rb_node);
187 bio_list_init(&tg->bio_lists[0]);
188 bio_list_init(&tg->bio_lists[1]);
189 tg->limits_changed = false;
190
191 /* Practically unlimited BW */
192 tg->bps[0] = tg->bps[1] = -1;
193 tg->iops[0] = tg->iops[1] = -1;
194
195 /*
196 * Take the initial reference that will be released on destroy
197 * This can be thought of a joint reference by cgroup and
198 * request queue which will be dropped by either request queue
199 * exit or cgroup deletion path depending on who is exiting first.
200 */
201 atomic_set(&tg->ref, 1);
202}
203
204/* Should be called with rcu read lock held (needed for blkcg) */
205static void
206throtl_add_group_to_td_list(struct throtl_data *td, struct throtl_grp *tg)
207{
208 hlist_add_head(&tg->tg_node, &td->tg_list);
209 td->nr_undestroyed_grps++;
210}
211
212static void
213__throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg)
214{
215 struct backing_dev_info *bdi = &td->queue->backing_dev_info;
216 unsigned int major, minor;
217
218 if (!tg || tg->blkg.dev)
219 return;
220
221 /*
222 * Fill in device details for a group which might not have been
223 * filled at group creation time as queue was being instantiated
224 * and driver had not attached a device yet
225 */
226 if (bdi->dev && dev_name(bdi->dev)) {
227 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
228 tg->blkg.dev = MKDEV(major, minor);
229 }
230}
231
232/*
233 * Should be called with without queue lock held. Here queue lock will be
234 * taken rarely. It will be taken only once during life time of a group
235 * if need be
236 */
237static void
238throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg)
239{
240 if (!tg || tg->blkg.dev)
241 return;
242
243 spin_lock_irq(td->queue->queue_lock);
244 __throtl_tg_fill_dev_details(td, tg);
245 spin_unlock_irq(td->queue->queue_lock);
246}
247
248static void throtl_init_add_tg_lists(struct throtl_data *td,
249 struct throtl_grp *tg, struct blkio_cgroup *blkcg)
250{
251 __throtl_tg_fill_dev_details(td, tg);
252
253 /* Add group onto cgroup list */
254 blkiocg_add_blkio_group(blkcg, &tg->blkg, (void *)td,
255 tg->blkg.dev, BLKIO_POLICY_THROTL);
256
257 tg->bps[READ] = blkcg_get_read_bps(blkcg, tg->blkg.dev);
258 tg->bps[WRITE] = blkcg_get_write_bps(blkcg, tg->blkg.dev);
259 tg->iops[READ] = blkcg_get_read_iops(blkcg, tg->blkg.dev);
260 tg->iops[WRITE] = blkcg_get_write_iops(blkcg, tg->blkg.dev);
261
262 throtl_add_group_to_td_list(td, tg);
263}
264
265/* Should be called without queue lock and outside of rcu period */
266static struct throtl_grp *throtl_alloc_tg(struct throtl_data *td)
267{
268 struct throtl_grp *tg = NULL;
269 int ret;
270
271 tg = kzalloc_node(sizeof(*tg), GFP_ATOMIC, td->queue->node);
272 if (!tg)
273 return NULL;
274
275 ret = blkio_alloc_blkg_stats(&tg->blkg);
276
277 if (ret) {
278 kfree(tg);
279 return NULL;
280 }
281
282 throtl_init_group(tg);
283 return tg;
284}
285
286static struct
287throtl_grp *throtl_find_tg(struct throtl_data *td, struct blkio_cgroup *blkcg)
288{
289 struct throtl_grp *tg = NULL;
290 void *key = td;
291
292 /*
293 * This is the common case when there are no blkio cgroups.
294 * Avoid lookup in this case
295 */
296 if (blkcg == &blkio_root_cgroup)
297 tg = td->root_tg;
298 else
299 tg = tg_of_blkg(blkiocg_lookup_group(blkcg, key));
300
301 __throtl_tg_fill_dev_details(td, tg);
302 return tg;
303}
304
305/*
306 * This function returns with queue lock unlocked in case of error, like
307 * request queue is no more
308 */
309static struct throtl_grp * throtl_get_tg(struct throtl_data *td)
310{
311 struct throtl_grp *tg = NULL, *__tg = NULL;
312 struct blkio_cgroup *blkcg;
313 struct request_queue *q = td->queue;
314
315 rcu_read_lock();
316 blkcg = task_blkio_cgroup(current);
317 tg = throtl_find_tg(td, blkcg);
318 if (tg) {
319 rcu_read_unlock();
320 return tg;
321 }
322
323 /*
324 * Need to allocate a group. Allocation of group also needs allocation
325 * of per cpu stats which in-turn takes a mutex() and can block. Hence
326 * we need to drop rcu lock and queue_lock before we call alloc
327 *
328 * Take the request queue reference to make sure queue does not
329 * go away once we return from allocation.
330 */
331 blk_get_queue(q);
332 rcu_read_unlock();
333 spin_unlock_irq(q->queue_lock);
334
335 tg = throtl_alloc_tg(td);
336 /*
337 * We might have slept in group allocation. Make sure queue is not
338 * dead
339 */
340 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
341 blk_put_queue(q);
342 if (tg)
343 kfree(tg);
344
345 return ERR_PTR(-ENODEV);
346 }
347 blk_put_queue(q);
348
349 /* Group allocated and queue is still alive. take the lock */
350 spin_lock_irq(q->queue_lock);
351
352 /*
353 * Initialize the new group. After sleeping, read the blkcg again.
354 */
355 rcu_read_lock();
356 blkcg = task_blkio_cgroup(current);
357
358 /*
359 * If some other thread already allocated the group while we were
360 * not holding queue lock, free up the group
361 */
362 __tg = throtl_find_tg(td, blkcg);
363
364 if (__tg) {
365 kfree(tg);
366 rcu_read_unlock();
367 return __tg;
368 }
369
370 /* Group allocation failed. Account the IO to root group */
371 if (!tg) {
372 tg = td->root_tg;
373 return tg;
374 }
375
376 throtl_init_add_tg_lists(td, tg, blkcg);
377 rcu_read_unlock();
378 return tg;
379}
380
381static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root)
382{
383 /* Service tree is empty */
384 if (!root->count)
385 return NULL;
386
387 if (!root->left)
388 root->left = rb_first(&root->rb);
389
390 if (root->left)
391 return rb_entry_tg(root->left);
392
393 return NULL;
394}
395
396static void rb_erase_init(struct rb_node *n, struct rb_root *root)
397{
398 rb_erase(n, root);
399 RB_CLEAR_NODE(n);
400}
401
402static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root)
403{
404 if (root->left == n)
405 root->left = NULL;
406 rb_erase_init(n, &root->rb);
407 --root->count;
408}
409
410static void update_min_dispatch_time(struct throtl_rb_root *st)
411{
412 struct throtl_grp *tg;
413
414 tg = throtl_rb_first(st);
415 if (!tg)
416 return;
417
418 st->min_disptime = tg->disptime;
419}
420
421static void
422tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg)
423{
424 struct rb_node **node = &st->rb.rb_node;
425 struct rb_node *parent = NULL;
426 struct throtl_grp *__tg;
427 unsigned long key = tg->disptime;
428 int left = 1;
429
430 while (*node != NULL) {
431 parent = *node;
432 __tg = rb_entry_tg(parent);
433
434 if (time_before(key, __tg->disptime))
435 node = &parent->rb_left;
436 else {
437 node = &parent->rb_right;
438 left = 0;
439 }
440 }
441
442 if (left)
443 st->left = &tg->rb_node;
444
445 rb_link_node(&tg->rb_node, parent, node);
446 rb_insert_color(&tg->rb_node, &st->rb);
447}
448
449static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
450{
451 struct throtl_rb_root *st = &td->tg_service_tree;
452
453 tg_service_tree_add(st, tg);
454 throtl_mark_tg_on_rr(tg);
455 st->count++;
456}
457
458static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
459{
460 if (!throtl_tg_on_rr(tg))
461 __throtl_enqueue_tg(td, tg);
462}
463
464static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
465{
466 throtl_rb_erase(&tg->rb_node, &td->tg_service_tree);
467 throtl_clear_tg_on_rr(tg);
468}
469
470static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
471{
472 if (throtl_tg_on_rr(tg))
473 __throtl_dequeue_tg(td, tg);
474}
475
476static void throtl_schedule_next_dispatch(struct throtl_data *td)
477{
478 struct throtl_rb_root *st = &td->tg_service_tree;
479
480 /*
481 * If there are more bios pending, schedule more work.
482 */
483 if (!total_nr_queued(td))
484 return;
485
486 BUG_ON(!st->count);
487
488 update_min_dispatch_time(st);
489
490 if (time_before_eq(st->min_disptime, jiffies))
491 throtl_schedule_delayed_work(td, 0);
492 else
493 throtl_schedule_delayed_work(td, (st->min_disptime - jiffies));
494}
495
496static inline void
497throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
498{
499 tg->bytes_disp[rw] = 0;
500 tg->io_disp[rw] = 0;
501 tg->slice_start[rw] = jiffies;
502 tg->slice_end[rw] = jiffies + throtl_slice;
503 throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
504 rw == READ ? 'R' : 'W', tg->slice_start[rw],
505 tg->slice_end[rw], jiffies);
506}
507
508static inline void throtl_set_slice_end(struct throtl_data *td,
509 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
510{
511 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
512}
513
514static inline void throtl_extend_slice(struct throtl_data *td,
515 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
516{
517 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
518 throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
519 rw == READ ? 'R' : 'W', tg->slice_start[rw],
520 tg->slice_end[rw], jiffies);
521}
522
523/* Determine if previously allocated or extended slice is complete or not */
524static bool
525throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
526{
527 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
528 return 0;
529
530 return 1;
531}
532
533/* Trim the used slices and adjust slice start accordingly */
534static inline void
535throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
536{
537 unsigned long nr_slices, time_elapsed, io_trim;
538 u64 bytes_trim, tmp;
539
540 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
541
542 /*
543 * If bps are unlimited (-1), then time slice don't get
544 * renewed. Don't try to trim the slice if slice is used. A new
545 * slice will start when appropriate.
546 */
547 if (throtl_slice_used(td, tg, rw))
548 return;
549
550 /*
551 * A bio has been dispatched. Also adjust slice_end. It might happen
552 * that initially cgroup limit was very low resulting in high
553 * slice_end, but later limit was bumped up and bio was dispached
554 * sooner, then we need to reduce slice_end. A high bogus slice_end
555 * is bad because it does not allow new slice to start.
556 */
557
558 throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice);
559
560 time_elapsed = jiffies - tg->slice_start[rw];
561
562 nr_slices = time_elapsed / throtl_slice;
563
564 if (!nr_slices)
565 return;
566 tmp = tg->bps[rw] * throtl_slice * nr_slices;
567 do_div(tmp, HZ);
568 bytes_trim = tmp;
569
570 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
571
572 if (!bytes_trim && !io_trim)
573 return;
574
575 if (tg->bytes_disp[rw] >= bytes_trim)
576 tg->bytes_disp[rw] -= bytes_trim;
577 else
578 tg->bytes_disp[rw] = 0;
579
580 if (tg->io_disp[rw] >= io_trim)
581 tg->io_disp[rw] -= io_trim;
582 else
583 tg->io_disp[rw] = 0;
584
585 tg->slice_start[rw] += nr_slices * throtl_slice;
586
587 throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
588 " start=%lu end=%lu jiffies=%lu",
589 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
590 tg->slice_start[rw], tg->slice_end[rw], jiffies);
591}
592
593static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
594 struct bio *bio, unsigned long *wait)
595{
596 bool rw = bio_data_dir(bio);
597 unsigned int io_allowed;
598 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
599 u64 tmp;
600
601 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
602
603 /* Slice has just started. Consider one slice interval */
604 if (!jiffy_elapsed)
605 jiffy_elapsed_rnd = throtl_slice;
606
607 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
608
609 /*
610 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
611 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
612 * will allow dispatch after 1 second and after that slice should
613 * have been trimmed.
614 */
615
616 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
617 do_div(tmp, HZ);
618
619 if (tmp > UINT_MAX)
620 io_allowed = UINT_MAX;
621 else
622 io_allowed = tmp;
623
624 if (tg->io_disp[rw] + 1 <= io_allowed) {
625 if (wait)
626 *wait = 0;
627 return 1;
628 }
629
630 /* Calc approx time to dispatch */
631 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
632
633 if (jiffy_wait > jiffy_elapsed)
634 jiffy_wait = jiffy_wait - jiffy_elapsed;
635 else
636 jiffy_wait = 1;
637
638 if (wait)
639 *wait = jiffy_wait;
640 return 0;
641}
642
643static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
644 struct bio *bio, unsigned long *wait)
645{
646 bool rw = bio_data_dir(bio);
647 u64 bytes_allowed, extra_bytes, tmp;
648 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
649
650 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
651
652 /* Slice has just started. Consider one slice interval */
653 if (!jiffy_elapsed)
654 jiffy_elapsed_rnd = throtl_slice;
655
656 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
657
658 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
659 do_div(tmp, HZ);
660 bytes_allowed = tmp;
661
662 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
663 if (wait)
664 *wait = 0;
665 return 1;
666 }
667
668 /* Calc approx time to dispatch */
669 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
670 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
671
672 if (!jiffy_wait)
673 jiffy_wait = 1;
674
675 /*
676 * This wait time is without taking into consideration the rounding
677 * up we did. Add that time also.
678 */
679 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
680 if (wait)
681 *wait = jiffy_wait;
682 return 0;
683}
684
685static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
686 if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
687 return 1;
688 return 0;
689}
690
691/*
692 * Returns whether one can dispatch a bio or not. Also returns approx number
693 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
694 */
695static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
696 struct bio *bio, unsigned long *wait)
697{
698 bool rw = bio_data_dir(bio);
699 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
700
701 /*
702 * Currently whole state machine of group depends on first bio
703 * queued in the group bio list. So one should not be calling
704 * this function with a different bio if there are other bios
705 * queued.
706 */
707 BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
708
709 /* If tg->bps = -1, then BW is unlimited */
710 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
711 if (wait)
712 *wait = 0;
713 return 1;
714 }
715
716 /*
717 * If previous slice expired, start a new one otherwise renew/extend
718 * existing slice to make sure it is at least throtl_slice interval
719 * long since now.
720 */
721 if (throtl_slice_used(td, tg, rw))
722 throtl_start_new_slice(td, tg, rw);
723 else {
724 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
725 throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
726 }
727
728 if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
729 && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
730 if (wait)
731 *wait = 0;
732 return 1;
733 }
734
735 max_wait = max(bps_wait, iops_wait);
736
737 if (wait)
738 *wait = max_wait;
739
740 if (time_before(tg->slice_end[rw], jiffies + max_wait))
741 throtl_extend_slice(td, tg, rw, jiffies + max_wait);
742
743 return 0;
744}
745
746static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
747{
748 bool rw = bio_data_dir(bio);
749 bool sync = rw_is_sync(bio->bi_rw);
750
751 /* Charge the bio to the group */
752 tg->bytes_disp[rw] += bio->bi_size;
753 tg->io_disp[rw]++;
754
755 blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync);
756}
757
758static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
759 struct bio *bio)
760{
761 bool rw = bio_data_dir(bio);
762
763 bio_list_add(&tg->bio_lists[rw], bio);
764 /* Take a bio reference on tg */
765 throtl_ref_get_tg(tg);
766 tg->nr_queued[rw]++;
767 td->nr_queued[rw]++;
768 throtl_enqueue_tg(td, tg);
769}
770
771static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
772{
773 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
774 struct bio *bio;
775
776 if ((bio = bio_list_peek(&tg->bio_lists[READ])))
777 tg_may_dispatch(td, tg, bio, &read_wait);
778
779 if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
780 tg_may_dispatch(td, tg, bio, &write_wait);
781
782 min_wait = min(read_wait, write_wait);
783 disptime = jiffies + min_wait;
784
785 /* Update dispatch time */
786 throtl_dequeue_tg(td, tg);
787 tg->disptime = disptime;
788 throtl_enqueue_tg(td, tg);
789}
790
791static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
792 bool rw, struct bio_list *bl)
793{
794 struct bio *bio;
795
796 bio = bio_list_pop(&tg->bio_lists[rw]);
797 tg->nr_queued[rw]--;
798 /* Drop bio reference on tg */
799 throtl_put_tg(tg);
800
801 BUG_ON(td->nr_queued[rw] <= 0);
802 td->nr_queued[rw]--;
803
804 throtl_charge_bio(tg, bio);
805 bio_list_add(bl, bio);
806 bio->bi_rw |= REQ_THROTTLED;
807
808 throtl_trim_slice(td, tg, rw);
809}
810
811static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
812 struct bio_list *bl)
813{
814 unsigned int nr_reads = 0, nr_writes = 0;
815 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
816 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
817 struct bio *bio;
818
819 /* Try to dispatch 75% READS and 25% WRITES */
820
821 while ((bio = bio_list_peek(&tg->bio_lists[READ]))
822 && tg_may_dispatch(td, tg, bio, NULL)) {
823
824 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
825 nr_reads++;
826
827 if (nr_reads >= max_nr_reads)
828 break;
829 }
830
831 while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
832 && tg_may_dispatch(td, tg, bio, NULL)) {
833
834 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
835 nr_writes++;
836
837 if (nr_writes >= max_nr_writes)
838 break;
839 }
840
841 return nr_reads + nr_writes;
842}
843
844static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
845{
846 unsigned int nr_disp = 0;
847 struct throtl_grp *tg;
848 struct throtl_rb_root *st = &td->tg_service_tree;
849
850 while (1) {
851 tg = throtl_rb_first(st);
852
853 if (!tg)
854 break;
855
856 if (time_before(jiffies, tg->disptime))
857 break;
858
859 throtl_dequeue_tg(td, tg);
860
861 nr_disp += throtl_dispatch_tg(td, tg, bl);
862
863 if (tg->nr_queued[0] || tg->nr_queued[1]) {
864 tg_update_disptime(td, tg);
865 throtl_enqueue_tg(td, tg);
866 }
867
868 if (nr_disp >= throtl_quantum)
869 break;
870 }
871
872 return nr_disp;
873}
874
875static void throtl_process_limit_change(struct throtl_data *td)
876{
877 struct throtl_grp *tg;
878 struct hlist_node *pos, *n;
879
880 if (!td->limits_changed)
881 return;
882
883 xchg(&td->limits_changed, false);
884
885 throtl_log(td, "limits changed");
886
887 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
888 if (!tg->limits_changed)
889 continue;
890
891 if (!xchg(&tg->limits_changed, false))
892 continue;
893
894 throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
895 " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
896 tg->iops[READ], tg->iops[WRITE]);
897
898 /*
899 * Restart the slices for both READ and WRITES. It
900 * might happen that a group's limit are dropped
901 * suddenly and we don't want to account recently
902 * dispatched IO with new low rate
903 */
904 throtl_start_new_slice(td, tg, 0);
905 throtl_start_new_slice(td, tg, 1);
906
907 if (throtl_tg_on_rr(tg))
908 tg_update_disptime(td, tg);
909 }
910}
911
912/* Dispatch throttled bios. Should be called without queue lock held. */
913static int throtl_dispatch(struct request_queue *q)
914{
915 struct throtl_data *td = q->td;
916 unsigned int nr_disp = 0;
917 struct bio_list bio_list_on_stack;
918 struct bio *bio;
919 struct blk_plug plug;
920
921 spin_lock_irq(q->queue_lock);
922
923 throtl_process_limit_change(td);
924
925 if (!total_nr_queued(td))
926 goto out;
927
928 bio_list_init(&bio_list_on_stack);
929
930 throtl_log(td, "dispatch nr_queued=%u read=%u write=%u",
931 total_nr_queued(td), td->nr_queued[READ],
932 td->nr_queued[WRITE]);
933
934 nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
935
936 if (nr_disp)
937 throtl_log(td, "bios disp=%u", nr_disp);
938
939 throtl_schedule_next_dispatch(td);
940out:
941 spin_unlock_irq(q->queue_lock);
942
943 /*
944 * If we dispatched some requests, unplug the queue to make sure
945 * immediate dispatch
946 */
947 if (nr_disp) {
948 blk_start_plug(&plug);
949 while((bio = bio_list_pop(&bio_list_on_stack)))
950 generic_make_request(bio);
951 blk_finish_plug(&plug);
952 }
953 return nr_disp;
954}
955
956void blk_throtl_work(struct work_struct *work)
957{
958 struct throtl_data *td = container_of(work, struct throtl_data,
959 throtl_work.work);
960 struct request_queue *q = td->queue;
961
962 throtl_dispatch(q);
963}
964
965/* Call with queue lock held */
966static void
967throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay)
968{
969
970 struct delayed_work *dwork = &td->throtl_work;
971
972 /* schedule work if limits changed even if no bio is queued */
973 if (total_nr_queued(td) || td->limits_changed) {
974 /*
975 * We might have a work scheduled to be executed in future.
976 * Cancel that and schedule a new one.
977 */
978 __cancel_delayed_work(dwork);
979 queue_delayed_work(kthrotld_workqueue, dwork, delay);
980 throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
981 delay, jiffies);
982 }
983}
984
985static void
986throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg)
987{
988 /* Something wrong if we are trying to remove same group twice */
989 BUG_ON(hlist_unhashed(&tg->tg_node));
990
991 hlist_del_init(&tg->tg_node);
992
993 /*
994 * Put the reference taken at the time of creation so that when all
995 * queues are gone, group can be destroyed.
996 */
997 throtl_put_tg(tg);
998 td->nr_undestroyed_grps--;
999}
1000
1001static void throtl_release_tgs(struct throtl_data *td)
1002{
1003 struct hlist_node *pos, *n;
1004 struct throtl_grp *tg;
1005
1006 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
1007 /*
1008 * If cgroup removal path got to blk_group first and removed
1009 * it from cgroup list, then it will take care of destroying
1010 * cfqg also.
1011 */
1012 if (!blkiocg_del_blkio_group(&tg->blkg))
1013 throtl_destroy_tg(td, tg);
1014 }
1015}
1016
1017static void throtl_td_free(struct throtl_data *td)
1018{
1019 kfree(td);
1020}
1021
1022/*
1023 * Blk cgroup controller notification saying that blkio_group object is being
1024 * delinked as associated cgroup object is going away. That also means that
1025 * no new IO will come in this group. So get rid of this group as soon as
1026 * any pending IO in the group is finished.
1027 *
1028 * This function is called under rcu_read_lock(). key is the rcu protected
1029 * pointer. That means "key" is a valid throtl_data pointer as long as we are
1030 * rcu read lock.
1031 *
1032 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1033 * it should not be NULL as even if queue was going away, cgroup deltion
1034 * path got to it first.
1035 */
1036void throtl_unlink_blkio_group(void *key, struct blkio_group *blkg)
1037{
1038 unsigned long flags;
1039 struct throtl_data *td = key;
1040
1041 spin_lock_irqsave(td->queue->queue_lock, flags);
1042 throtl_destroy_tg(td, tg_of_blkg(blkg));
1043 spin_unlock_irqrestore(td->queue->queue_lock, flags);
1044}
1045
1046static void throtl_update_blkio_group_common(struct throtl_data *td,
1047 struct throtl_grp *tg)
1048{
1049 xchg(&tg->limits_changed, true);
1050 xchg(&td->limits_changed, true);
1051 /* Schedule a work now to process the limit change */
1052 throtl_schedule_delayed_work(td, 0);
1053}
1054
1055/*
1056 * For all update functions, key should be a valid pointer because these
1057 * update functions are called under blkcg_lock, that means, blkg is
1058 * valid and in turn key is valid. queue exit path can not race because
1059 * of blkcg_lock
1060 *
1061 * Can not take queue lock in update functions as queue lock under blkcg_lock
1062 * is not allowed. Under other paths we take blkcg_lock under queue_lock.
1063 */
1064static void throtl_update_blkio_group_read_bps(void *key,
1065 struct blkio_group *blkg, u64 read_bps)
1066{
1067 struct throtl_data *td = key;
1068 struct throtl_grp *tg = tg_of_blkg(blkg);
1069
1070 tg->bps[READ] = read_bps;
1071 throtl_update_blkio_group_common(td, tg);
1072}
1073
1074static void throtl_update_blkio_group_write_bps(void *key,
1075 struct blkio_group *blkg, u64 write_bps)
1076{
1077 struct throtl_data *td = key;
1078 struct throtl_grp *tg = tg_of_blkg(blkg);
1079
1080 tg->bps[WRITE] = write_bps;
1081 throtl_update_blkio_group_common(td, tg);
1082}
1083
1084static void throtl_update_blkio_group_read_iops(void *key,
1085 struct blkio_group *blkg, unsigned int read_iops)
1086{
1087 struct throtl_data *td = key;
1088 struct throtl_grp *tg = tg_of_blkg(blkg);
1089
1090 tg->iops[READ] = read_iops;
1091 throtl_update_blkio_group_common(td, tg);
1092}
1093
1094static void throtl_update_blkio_group_write_iops(void *key,
1095 struct blkio_group *blkg, unsigned int write_iops)
1096{
1097 struct throtl_data *td = key;
1098 struct throtl_grp *tg = tg_of_blkg(blkg);
1099
1100 tg->iops[WRITE] = write_iops;
1101 throtl_update_blkio_group_common(td, tg);
1102}
1103
1104static void throtl_shutdown_wq(struct request_queue *q)
1105{
1106 struct throtl_data *td = q->td;
1107
1108 cancel_delayed_work_sync(&td->throtl_work);
1109}
1110
1111static struct blkio_policy_type blkio_policy_throtl = {
1112 .ops = {
1113 .blkio_unlink_group_fn = throtl_unlink_blkio_group,
1114 .blkio_update_group_read_bps_fn =
1115 throtl_update_blkio_group_read_bps,
1116 .blkio_update_group_write_bps_fn =
1117 throtl_update_blkio_group_write_bps,
1118 .blkio_update_group_read_iops_fn =
1119 throtl_update_blkio_group_read_iops,
1120 .blkio_update_group_write_iops_fn =
1121 throtl_update_blkio_group_write_iops,
1122 },
1123 .plid = BLKIO_POLICY_THROTL,
1124};
1125
1126int blk_throtl_bio(struct request_queue *q, struct bio **biop)
1127{
1128 struct throtl_data *td = q->td;
1129 struct throtl_grp *tg;
1130 struct bio *bio = *biop;
1131 bool rw = bio_data_dir(bio), update_disptime = true;
1132 struct blkio_cgroup *blkcg;
1133
1134 if (bio->bi_rw & REQ_THROTTLED) {
1135 bio->bi_rw &= ~REQ_THROTTLED;
1136 return 0;
1137 }
1138
1139 /*
1140 * A throtl_grp pointer retrieved under rcu can be used to access
1141 * basic fields like stats and io rates. If a group has no rules,
1142 * just update the dispatch stats in lockless manner and return.
1143 */
1144
1145 rcu_read_lock();
1146 blkcg = task_blkio_cgroup(current);
1147 tg = throtl_find_tg(td, blkcg);
1148 if (tg) {
1149 throtl_tg_fill_dev_details(td, tg);
1150
1151 if (tg_no_rule_group(tg, rw)) {
1152 blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size,
1153 rw, rw_is_sync(bio->bi_rw));
1154 rcu_read_unlock();
1155 return 0;
1156 }
1157 }
1158 rcu_read_unlock();
1159
1160 /*
1161 * Either group has not been allocated yet or it is not an unlimited
1162 * IO group
1163 */
1164
1165 spin_lock_irq(q->queue_lock);
1166 tg = throtl_get_tg(td);
1167
1168 if (IS_ERR(tg)) {
1169 if (PTR_ERR(tg) == -ENODEV) {
1170 /*
1171 * Queue is gone. No queue lock held here.
1172 */
1173 return -ENODEV;
1174 }
1175 }
1176
1177 if (tg->nr_queued[rw]) {
1178 /*
1179 * There is already another bio queued in same dir. No
1180 * need to update dispatch time.
1181 */
1182 update_disptime = false;
1183 goto queue_bio;
1184
1185 }
1186
1187 /* Bio is with-in rate limit of group */
1188 if (tg_may_dispatch(td, tg, bio, NULL)) {
1189 throtl_charge_bio(tg, bio);
1190
1191 /*
1192 * We need to trim slice even when bios are not being queued
1193 * otherwise it might happen that a bio is not queued for
1194 * a long time and slice keeps on extending and trim is not
1195 * called for a long time. Now if limits are reduced suddenly
1196 * we take into account all the IO dispatched so far at new
1197 * low rate and * newly queued IO gets a really long dispatch
1198 * time.
1199 *
1200 * So keep on trimming slice even if bio is not queued.
1201 */
1202 throtl_trim_slice(td, tg, rw);
1203 goto out;
1204 }
1205
1206queue_bio:
1207 throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
1208 " iodisp=%u iops=%u queued=%d/%d",
1209 rw == READ ? 'R' : 'W',
1210 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1211 tg->io_disp[rw], tg->iops[rw],
1212 tg->nr_queued[READ], tg->nr_queued[WRITE]);
1213
1214 throtl_add_bio_tg(q->td, tg, bio);
1215 *biop = NULL;
1216
1217 if (update_disptime) {
1218 tg_update_disptime(td, tg);
1219 throtl_schedule_next_dispatch(td);
1220 }
1221
1222out:
1223 spin_unlock_irq(q->queue_lock);
1224 return 0;
1225}
1226
1227int blk_throtl_init(struct request_queue *q)
1228{
1229 struct throtl_data *td;
1230 struct throtl_grp *tg;
1231
1232 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1233 if (!td)
1234 return -ENOMEM;
1235
1236 INIT_HLIST_HEAD(&td->tg_list);
1237 td->tg_service_tree = THROTL_RB_ROOT;
1238 td->limits_changed = false;
1239 INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);
1240
1241 /* alloc and Init root group. */
1242 td->queue = q;
1243 tg = throtl_alloc_tg(td);
1244
1245 if (!tg) {
1246 kfree(td);
1247 return -ENOMEM;
1248 }
1249
1250 td->root_tg = tg;
1251
1252 rcu_read_lock();
1253 throtl_init_add_tg_lists(td, tg, &blkio_root_cgroup);
1254 rcu_read_unlock();
1255
1256 /* Attach throtl data to request queue */
1257 q->td = td;
1258 return 0;
1259}
1260
1261void blk_throtl_exit(struct request_queue *q)
1262{
1263 struct throtl_data *td = q->td;
1264 bool wait = false;
1265
1266 BUG_ON(!td);
1267
1268 throtl_shutdown_wq(q);
1269
1270 spin_lock_irq(q->queue_lock);
1271 throtl_release_tgs(td);
1272
1273 /* If there are other groups */
1274 if (td->nr_undestroyed_grps > 0)
1275 wait = true;
1276
1277 spin_unlock_irq(q->queue_lock);
1278
1279 /*
1280 * Wait for tg->blkg->key accessors to exit their grace periods.
1281 * Do this wait only if there are other undestroyed groups out
1282 * there (other than root group). This can happen if cgroup deletion
1283 * path claimed the responsibility of cleaning up a group before
1284 * queue cleanup code get to the group.
1285 *
1286 * Do not call synchronize_rcu() unconditionally as there are drivers
1287 * which create/delete request queue hundreds of times during scan/boot
1288 * and synchronize_rcu() can take significant time and slow down boot.
1289 */
1290 if (wait)
1291 synchronize_rcu();
1292
1293 /*
1294 * Just being safe to make sure after previous flush if some body did
1295 * update limits through cgroup and another work got queued, cancel
1296 * it.
1297 */
1298 throtl_shutdown_wq(q);
1299 throtl_td_free(td);
1300}
1301
1302static int __init throtl_init(void)
1303{
1304 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1305 if (!kthrotld_workqueue)
1306 panic("Failed to create kthrotld\n");
1307
1308 blkio_policy_register(&blkio_policy_throtl);
1309 return 0;
1310}
1311
1312module_init(throtl_init);