2 * Interface for controlling IO bandwidth on a request queue
4 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
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"
15 /* Max dispatch from a group in 1 round */
16 static int throtl_grp_quantum = 8;
18 /* Total max dispatch from all groups in one round */
19 static int throtl_quantum = 32;
21 /* Throttling is performed over 100ms slice and after that slice is renewed */
22 static unsigned long throtl_slice = HZ/10; /* 100 ms */
24 static struct blkcg_policy blkcg_policy_throtl;
26 /* A workqueue to queue throttle related work */
27 static struct workqueue_struct *kthrotld_workqueue;
28 static void throtl_schedule_delayed_work(struct throtl_data *td,
31 struct throtl_rb_root {
35 unsigned long min_disptime;
38 #define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \
39 .count = 0, .min_disptime = 0}
41 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
43 /* Per-cpu group stats */
45 /* total bytes transferred */
46 struct blkg_rwstat service_bytes;
47 /* total IOs serviced, post merge */
48 struct blkg_rwstat serviced;
52 /* must be the first member */
53 struct blkg_policy_data pd;
55 /* active throtl group service_tree member */
56 struct rb_node rb_node;
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.
63 unsigned long disptime;
67 /* Two lists for READ and WRITE */
68 struct bio_list bio_lists[2];
70 /* Number of queued bios on READ and WRITE lists */
71 unsigned int nr_queued[2];
73 /* bytes per second rate limits */
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];
84 /* When did we start a new slice */
85 unsigned long slice_start[2];
86 unsigned long slice_end[2];
88 /* Some throttle limits got updated for the group */
91 /* Per cpu stats pointer */
92 struct tg_stats_cpu __percpu *stats_cpu;
94 /* List of tgs waiting for per cpu stats memory to be allocated */
95 struct list_head stats_alloc_node;
100 /* service tree for active throtl groups */
101 struct throtl_rb_root tg_service_tree;
103 struct request_queue *queue;
105 /* Total Number of queued bios on READ and WRITE lists */
106 unsigned int nr_queued[2];
109 * number of total undestroyed groups
111 unsigned int nr_undestroyed_grps;
113 /* Work for dispatching throttled bios */
114 struct delayed_work throtl_work;
119 /* list and work item to allocate percpu group stats */
120 static DEFINE_SPINLOCK(tg_stats_alloc_lock);
121 static LIST_HEAD(tg_stats_alloc_list);
123 static void tg_stats_alloc_fn(struct work_struct *);
124 static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn);
126 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
128 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
131 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
133 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
136 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
138 return pd_to_blkg(&tg->pd);
141 static inline struct throtl_grp *td_root_tg(struct throtl_data *td)
143 return blkg_to_tg(td->queue->root_blkg);
146 enum tg_state_flags {
147 THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */
150 #define THROTL_TG_FNS(name) \
151 static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \
153 (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \
155 static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \
157 (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \
159 static inline int throtl_tg_##name(const struct throtl_grp *tg) \
161 return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \
164 THROTL_TG_FNS(on_rr);
166 #define throtl_log_tg(td, tg, fmt, args...) do { \
169 blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \
170 blk_add_trace_msg((td)->queue, "throtl %s " fmt, __pbuf, ##args); \
173 #define throtl_log(td, fmt, args...) \
174 blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
176 static inline unsigned int total_nr_queued(struct throtl_data *td)
178 return td->nr_queued[0] + td->nr_queued[1];
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
186 static void tg_stats_alloc_fn(struct work_struct *work)
188 static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */
189 struct delayed_work *dwork = to_delayed_work(work);
194 stats_cpu = alloc_percpu(struct tg_stats_cpu);
196 /* allocation failed, try again after some time */
197 queue_delayed_work(system_nrt_wq, dwork,
198 msecs_to_jiffies(10));
203 spin_lock_irq(&tg_stats_alloc_lock);
205 if (!list_empty(&tg_stats_alloc_list)) {
206 struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list,
209 swap(tg->stats_cpu, stats_cpu);
210 list_del_init(&tg->stats_alloc_node);
213 empty = list_empty(&tg_stats_alloc_list);
214 spin_unlock_irq(&tg_stats_alloc_lock);
219 static void throtl_pd_init(struct blkcg_gq *blkg)
221 struct throtl_grp *tg = blkg_to_tg(blkg);
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;
232 tg->iops[WRITE] = -1;
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.
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);
245 static void throtl_pd_exit(struct blkcg_gq *blkg)
247 struct throtl_grp *tg = blkg_to_tg(blkg);
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);
254 free_percpu(tg->stats_cpu);
257 static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
259 struct throtl_grp *tg = blkg_to_tg(blkg);
262 if (tg->stats_cpu == NULL)
265 for_each_possible_cpu(cpu) {
266 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
268 blkg_rwstat_reset(&sc->service_bytes);
269 blkg_rwstat_reset(&sc->serviced);
273 static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
277 * This is the common case when there are no blkcgs. Avoid lookup
280 if (blkcg == &blkcg_root)
281 return td_root_tg(td);
283 return blkg_to_tg(blkg_lookup(blkcg, td->queue));
286 static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
289 struct request_queue *q = td->queue;
290 struct throtl_grp *tg = NULL;
293 * This is the common case when there are no blkcgs. Avoid lookup
296 if (blkcg == &blkcg_root) {
299 struct blkcg_gq *blkg;
301 blkg = blkg_lookup_create(blkcg, q);
303 /* if %NULL and @q is alive, fall back to root_tg */
305 tg = blkg_to_tg(blkg);
306 else if (!blk_queue_dead(q))
313 static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root)
315 /* Service tree is empty */
320 root->left = rb_first(&root->rb);
323 return rb_entry_tg(root->left);
328 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
334 static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root)
338 rb_erase_init(n, &root->rb);
342 static void update_min_dispatch_time(struct throtl_rb_root *st)
344 struct throtl_grp *tg;
346 tg = throtl_rb_first(st);
350 st->min_disptime = tg->disptime;
354 tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg)
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;
362 while (*node != NULL) {
364 __tg = rb_entry_tg(parent);
366 if (time_before(key, __tg->disptime))
367 node = &parent->rb_left;
369 node = &parent->rb_right;
375 st->left = &tg->rb_node;
377 rb_link_node(&tg->rb_node, parent, node);
378 rb_insert_color(&tg->rb_node, &st->rb);
381 static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
383 struct throtl_rb_root *st = &td->tg_service_tree;
385 tg_service_tree_add(st, tg);
386 throtl_mark_tg_on_rr(tg);
390 static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
392 if (!throtl_tg_on_rr(tg))
393 __throtl_enqueue_tg(td, tg);
396 static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
398 throtl_rb_erase(&tg->rb_node, &td->tg_service_tree);
399 throtl_clear_tg_on_rr(tg);
402 static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
404 if (throtl_tg_on_rr(tg))
405 __throtl_dequeue_tg(td, tg);
408 static void throtl_schedule_next_dispatch(struct throtl_data *td)
410 struct throtl_rb_root *st = &td->tg_service_tree;
413 * If there are more bios pending, schedule more work.
415 if (!total_nr_queued(td))
420 update_min_dispatch_time(st);
422 if (time_before_eq(st->min_disptime, jiffies))
423 throtl_schedule_delayed_work(td, 0);
425 throtl_schedule_delayed_work(td, (st->min_disptime - jiffies));
429 throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
431 tg->bytes_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);
440 static inline void throtl_set_slice_end(struct throtl_data *td,
441 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
443 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
446 static inline void throtl_extend_slice(struct throtl_data *td,
447 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
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);
455 /* Determine if previously allocated or extended slice is complete or not */
457 throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
459 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
465 /* Trim the used slices and adjust slice start accordingly */
467 throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
469 unsigned long nr_slices, time_elapsed, io_trim;
472 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
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.
479 if (throtl_slice_used(td, tg, rw))
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.
490 throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice);
492 time_elapsed = jiffies - tg->slice_start[rw];
494 nr_slices = time_elapsed / throtl_slice;
498 tmp = tg->bps[rw] * throtl_slice * nr_slices;
502 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
504 if (!bytes_trim && !io_trim)
507 if (tg->bytes_disp[rw] >= bytes_trim)
508 tg->bytes_disp[rw] -= bytes_trim;
510 tg->bytes_disp[rw] = 0;
512 if (tg->io_disp[rw] >= io_trim)
513 tg->io_disp[rw] -= io_trim;
517 tg->slice_start[rw] += nr_slices * throtl_slice;
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);
525 static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
526 struct bio *bio, unsigned long *wait)
528 bool rw = bio_data_dir(bio);
529 unsigned int io_allowed;
530 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
533 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
535 /* Slice has just started. Consider one slice interval */
537 jiffy_elapsed_rnd = throtl_slice;
539 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
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
548 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
552 io_allowed = UINT_MAX;
556 if (tg->io_disp[rw] + 1 <= io_allowed) {
562 /* Calc approx time to dispatch */
563 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
565 if (jiffy_wait > jiffy_elapsed)
566 jiffy_wait = jiffy_wait - jiffy_elapsed;
575 static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
576 struct bio *bio, unsigned long *wait)
578 bool rw = bio_data_dir(bio);
579 u64 bytes_allowed, extra_bytes, tmp;
580 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
582 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
584 /* Slice has just started. Consider one slice interval */
586 jiffy_elapsed_rnd = throtl_slice;
588 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
590 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
594 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
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]);
608 * This wait time is without taking into consideration the rounding
609 * up we did. Add that time also.
611 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
617 static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
618 if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
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
627 static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
628 struct bio *bio, unsigned long *wait)
630 bool rw = bio_data_dir(bio);
631 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
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
639 BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
641 /* If tg->bps = -1, then BW is unlimited */
642 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
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
653 if (throtl_slice_used(td, tg, rw))
654 throtl_start_new_slice(td, tg, rw);
656 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
657 throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
660 if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
661 && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
667 max_wait = max(bps_wait, iops_wait);
672 if (time_before(tg->slice_end[rw], jiffies + max_wait))
673 throtl_extend_slice(td, tg, rw, jiffies + max_wait);
678 static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
681 struct throtl_grp *tg = blkg_to_tg(blkg);
682 struct tg_stats_cpu *stats_cpu;
685 /* If per cpu stats are not allocated yet, don't do any accounting. */
686 if (tg->stats_cpu == NULL)
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.
694 local_irq_save(flags);
696 stats_cpu = this_cpu_ptr(tg->stats_cpu);
698 blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
699 blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
701 local_irq_restore(flags);
704 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
706 bool rw = bio_data_dir(bio);
708 /* Charge the bio to the group */
709 tg->bytes_disp[rw] += bio->bi_size;
712 throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw);
715 static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
718 bool rw = bio_data_dir(bio);
720 bio_list_add(&tg->bio_lists[rw], bio);
721 /* Take a bio reference on tg */
722 blkg_get(tg_to_blkg(tg));
725 throtl_enqueue_tg(td, tg);
728 static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
730 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
733 if ((bio = bio_list_peek(&tg->bio_lists[READ])))
734 tg_may_dispatch(td, tg, bio, &read_wait);
736 if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
737 tg_may_dispatch(td, tg, bio, &write_wait);
739 min_wait = min(read_wait, write_wait);
740 disptime = jiffies + min_wait;
742 /* Update dispatch time */
743 throtl_dequeue_tg(td, tg);
744 tg->disptime = disptime;
745 throtl_enqueue_tg(td, tg);
748 static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
749 bool rw, struct bio_list *bl)
753 bio = bio_list_pop(&tg->bio_lists[rw]);
755 /* Drop bio reference on blkg */
756 blkg_put(tg_to_blkg(tg));
758 BUG_ON(td->nr_queued[rw] <= 0);
761 throtl_charge_bio(tg, bio);
762 bio_list_add(bl, bio);
763 bio->bi_rw |= REQ_THROTTLED;
765 throtl_trim_slice(td, tg, rw);
768 static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
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;
776 /* Try to dispatch 75% READS and 25% WRITES */
778 while ((bio = bio_list_peek(&tg->bio_lists[READ]))
779 && tg_may_dispatch(td, tg, bio, NULL)) {
781 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
784 if (nr_reads >= max_nr_reads)
788 while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
789 && tg_may_dispatch(td, tg, bio, NULL)) {
791 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
794 if (nr_writes >= max_nr_writes)
798 return nr_reads + nr_writes;
801 static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
803 unsigned int nr_disp = 0;
804 struct throtl_grp *tg;
805 struct throtl_rb_root *st = &td->tg_service_tree;
808 tg = throtl_rb_first(st);
813 if (time_before(jiffies, tg->disptime))
816 throtl_dequeue_tg(td, tg);
818 nr_disp += throtl_dispatch_tg(td, tg, bl);
820 if (tg->nr_queued[0] || tg->nr_queued[1]) {
821 tg_update_disptime(td, tg);
822 throtl_enqueue_tg(td, tg);
825 if (nr_disp >= throtl_quantum)
832 static void throtl_process_limit_change(struct throtl_data *td)
834 struct request_queue *q = td->queue;
835 struct blkcg_gq *blkg, *n;
837 if (!td->limits_changed)
840 xchg(&td->limits_changed, false);
842 throtl_log(td, "limits changed");
844 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
845 struct throtl_grp *tg = blkg_to_tg(blkg);
847 if (!tg->limits_changed)
850 if (!xchg(&tg->limits_changed, false))
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]);
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
863 throtl_start_new_slice(td, tg, 0);
864 throtl_start_new_slice(td, tg, 1);
866 if (throtl_tg_on_rr(tg))
867 tg_update_disptime(td, tg);
871 /* Dispatch throttled bios. Should be called without queue lock held. */
872 static int throtl_dispatch(struct request_queue *q)
874 struct throtl_data *td = q->td;
875 unsigned int nr_disp = 0;
876 struct bio_list bio_list_on_stack;
878 struct blk_plug plug;
880 spin_lock_irq(q->queue_lock);
882 throtl_process_limit_change(td);
884 if (!total_nr_queued(td))
887 bio_list_init(&bio_list_on_stack);
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]);
893 nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
896 throtl_log(td, "bios disp=%u", nr_disp);
898 throtl_schedule_next_dispatch(td);
900 spin_unlock_irq(q->queue_lock);
903 * If we dispatched some requests, unplug the queue to make sure
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);
915 void blk_throtl_work(struct work_struct *work)
917 struct throtl_data *td = container_of(work, struct throtl_data,
919 struct request_queue *q = td->queue;
924 /* Call with queue lock held */
926 throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay)
929 struct delayed_work *dwork = &td->throtl_work;
931 /* schedule work if limits changed even if no bio is queued */
932 if (total_nr_queued(td) || td->limits_changed) {
934 * We might have a work scheduled to be executed in future.
935 * Cancel that and schedule a new one.
937 __cancel_delayed_work(dwork);
938 queue_delayed_work(kthrotld_workqueue, dwork, delay);
939 throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
944 static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
945 struct blkg_policy_data *pd, int off)
947 struct throtl_grp *tg = pd_to_tg(pd);
948 struct blkg_rwstat rwstat = { }, tmp;
951 for_each_possible_cpu(cpu) {
952 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
954 tmp = blkg_rwstat_read((void *)sc + off);
955 for (i = 0; i < BLKG_RWSTAT_NR; i++)
956 rwstat.cnt[i] += tmp.cnt[i];
959 return __blkg_prfill_rwstat(sf, pd, &rwstat);
962 static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft,
965 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
967 blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl,
972 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
975 struct throtl_grp *tg = pd_to_tg(pd);
976 u64 v = *(u64 *)((void *)tg + off);
980 return __blkg_prfill_u64(sf, pd, v);
983 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
986 struct throtl_grp *tg = pd_to_tg(pd);
987 unsigned int v = *(unsigned int *)((void *)tg + off);
991 return __blkg_prfill_u64(sf, pd, v);
994 static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft,
997 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64,
998 &blkcg_policy_throtl, cft->private, false);
1002 static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1003 struct seq_file *sf)
1005 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint,
1006 &blkcg_policy_throtl, cft->private, false);
1010 static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf,
1013 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1014 struct blkg_conf_ctx ctx;
1015 struct throtl_grp *tg;
1016 struct throtl_data *td;
1019 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1023 tg = blkg_to_tg(ctx.blkg);
1024 td = ctx.blkg->q->td;
1030 *(u64 *)((void *)tg + cft->private) = ctx.v;
1032 *(unsigned int *)((void *)tg + cft->private) = ctx.v;
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);
1039 blkg_conf_finish(&ctx);
1043 static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft,
1046 return tg_set_conf(cgrp, cft, buf, true);
1049 static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1052 return tg_set_conf(cgrp, cft, buf, false);
1055 static struct cftype throtl_files[] = {
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,
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,
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,
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,
1085 .name = "throttle.io_service_bytes",
1086 .private = offsetof(struct tg_stats_cpu, service_bytes),
1087 .read_seq_string = tg_print_cpu_rwstat,
1090 .name = "throttle.io_serviced",
1091 .private = offsetof(struct tg_stats_cpu, serviced),
1092 .read_seq_string = tg_print_cpu_rwstat,
1097 static void throtl_shutdown_wq(struct request_queue *q)
1099 struct throtl_data *td = q->td;
1101 cancel_delayed_work_sync(&td->throtl_work);
1104 static struct blkcg_policy blkcg_policy_throtl = {
1105 .pd_size = sizeof(struct throtl_grp),
1106 .cftypes = throtl_files,
1108 .pd_init_fn = throtl_pd_init,
1109 .pd_exit_fn = throtl_pd_exit,
1110 .pd_reset_stats_fn = throtl_pd_reset_stats,
1113 bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
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;
1121 if (bio->bi_rw & REQ_THROTTLED) {
1122 bio->bi_rw &= ~REQ_THROTTLED;
1127 * A throtl_grp pointer retrieved under rcu can be used to access
1128 * basic fields like stats and io rates. If a group has no rules,
1129 * just update the dispatch stats in lockless manner and return.
1132 blkcg = bio_blkcg(bio);
1133 tg = throtl_lookup_tg(td, blkcg);
1135 if (tg_no_rule_group(tg, rw)) {
1136 throtl_update_dispatch_stats(tg_to_blkg(tg),
1137 bio->bi_size, bio->bi_rw);
1138 goto out_unlock_rcu;
1143 * Either group has not been allocated yet or it is not an unlimited
1146 spin_lock_irq(q->queue_lock);
1147 tg = throtl_lookup_create_tg(td, blkcg);
1151 if (tg->nr_queued[rw]) {
1153 * There is already another bio queued in same dir. No
1154 * need to update dispatch time.
1156 update_disptime = false;
1161 /* Bio is with-in rate limit of group */
1162 if (tg_may_dispatch(td, tg, bio, NULL)) {
1163 throtl_charge_bio(tg, bio);
1166 * We need to trim slice even when bios are not being queued
1167 * otherwise it might happen that a bio is not queued for
1168 * a long time and slice keeps on extending and trim is not
1169 * called for a long time. Now if limits are reduced suddenly
1170 * we take into account all the IO dispatched so far at new
1171 * low rate and * newly queued IO gets a really long dispatch
1174 * So keep on trimming slice even if bio is not queued.
1176 throtl_trim_slice(td, tg, rw);
1181 throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
1182 " iodisp=%u iops=%u queued=%d/%d",
1183 rw == READ ? 'R' : 'W',
1184 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1185 tg->io_disp[rw], tg->iops[rw],
1186 tg->nr_queued[READ], tg->nr_queued[WRITE]);
1188 bio_associate_current(bio);
1189 throtl_add_bio_tg(q->td, tg, bio);
1192 if (update_disptime) {
1193 tg_update_disptime(td, tg);
1194 throtl_schedule_next_dispatch(td);
1198 spin_unlock_irq(q->queue_lock);
1206 * blk_throtl_drain - drain throttled bios
1207 * @q: request_queue to drain throttled bios for
1209 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1211 void blk_throtl_drain(struct request_queue *q)
1212 __releases(q->queue_lock) __acquires(q->queue_lock)
1214 struct throtl_data *td = q->td;
1215 struct throtl_rb_root *st = &td->tg_service_tree;
1216 struct throtl_grp *tg;
1220 queue_lockdep_assert_held(q);
1224 while ((tg = throtl_rb_first(st))) {
1225 throtl_dequeue_tg(td, tg);
1227 while ((bio = bio_list_peek(&tg->bio_lists[READ])))
1228 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl);
1229 while ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
1230 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl);
1232 spin_unlock_irq(q->queue_lock);
1234 while ((bio = bio_list_pop(&bl)))
1235 generic_make_request(bio);
1237 spin_lock_irq(q->queue_lock);
1240 int blk_throtl_init(struct request_queue *q)
1242 struct throtl_data *td;
1245 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1249 td->tg_service_tree = THROTL_RB_ROOT;
1250 td->limits_changed = false;
1251 INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);
1256 /* activate policy */
1257 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1263 void blk_throtl_exit(struct request_queue *q)
1266 throtl_shutdown_wq(q);
1267 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1271 static int __init throtl_init(void)
1273 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1274 if (!kthrotld_workqueue)
1275 panic("Failed to create kthrotld\n");
1277 return blkcg_policy_register(&blkcg_policy_throtl);
1280 module_init(throtl_init);