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_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 schedule_delayed_work(dwork, msecs_to_jiffies(10));
202 spin_lock_irq(&tg_stats_alloc_lock);
204 if (!list_empty(&tg_stats_alloc_list)) {
205 struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list,
208 swap(tg->stats_cpu, stats_cpu);
209 list_del_init(&tg->stats_alloc_node);
212 empty = list_empty(&tg_stats_alloc_list);
213 spin_unlock_irq(&tg_stats_alloc_lock);
218 static void throtl_pd_init(struct blkcg_gq *blkg)
220 struct throtl_grp *tg = blkg_to_tg(blkg);
223 RB_CLEAR_NODE(&tg->rb_node);
224 bio_list_init(&tg->bio_lists[0]);
225 bio_list_init(&tg->bio_lists[1]);
226 tg->limits_changed = false;
231 tg->iops[WRITE] = -1;
234 * Ugh... We need to perform per-cpu allocation for tg->stats_cpu
235 * but percpu allocator can't be called from IO path. Queue tg on
236 * tg_stats_alloc_list and allocate from work item.
238 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
239 list_add(&tg->stats_alloc_node, &tg_stats_alloc_list);
240 schedule_delayed_work(&tg_stats_alloc_work, 0);
241 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
244 static void throtl_pd_exit(struct blkcg_gq *blkg)
246 struct throtl_grp *tg = blkg_to_tg(blkg);
249 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
250 list_del_init(&tg->stats_alloc_node);
251 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
253 free_percpu(tg->stats_cpu);
256 static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
258 struct throtl_grp *tg = blkg_to_tg(blkg);
261 if (tg->stats_cpu == NULL)
264 for_each_possible_cpu(cpu) {
265 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
267 blkg_rwstat_reset(&sc->service_bytes);
268 blkg_rwstat_reset(&sc->serviced);
272 static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
276 * This is the common case when there are no blkcgs. Avoid lookup
279 if (blkcg == &blkcg_root)
280 return td_root_tg(td);
282 return blkg_to_tg(blkg_lookup(blkcg, td->queue));
285 static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
288 struct request_queue *q = td->queue;
289 struct throtl_grp *tg = NULL;
292 * This is the common case when there are no blkcgs. Avoid lookup
295 if (blkcg == &blkcg_root) {
298 struct blkcg_gq *blkg;
300 blkg = blkg_lookup_create(blkcg, q);
302 /* if %NULL and @q is alive, fall back to root_tg */
304 tg = blkg_to_tg(blkg);
305 else if (!blk_queue_dying(q))
312 static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root)
314 /* Service tree is empty */
319 root->left = rb_first(&root->rb);
322 return rb_entry_tg(root->left);
327 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
333 static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root)
337 rb_erase_init(n, &root->rb);
341 static void update_min_dispatch_time(struct throtl_rb_root *st)
343 struct throtl_grp *tg;
345 tg = throtl_rb_first(st);
349 st->min_disptime = tg->disptime;
353 tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg)
355 struct rb_node **node = &st->rb.rb_node;
356 struct rb_node *parent = NULL;
357 struct throtl_grp *__tg;
358 unsigned long key = tg->disptime;
361 while (*node != NULL) {
363 __tg = rb_entry_tg(parent);
365 if (time_before(key, __tg->disptime))
366 node = &parent->rb_left;
368 node = &parent->rb_right;
374 st->left = &tg->rb_node;
376 rb_link_node(&tg->rb_node, parent, node);
377 rb_insert_color(&tg->rb_node, &st->rb);
380 static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
382 struct throtl_rb_root *st = &td->tg_service_tree;
384 tg_service_tree_add(st, tg);
385 throtl_mark_tg_on_rr(tg);
389 static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
391 if (!throtl_tg_on_rr(tg))
392 __throtl_enqueue_tg(td, tg);
395 static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
397 throtl_rb_erase(&tg->rb_node, &td->tg_service_tree);
398 throtl_clear_tg_on_rr(tg);
401 static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
403 if (throtl_tg_on_rr(tg))
404 __throtl_dequeue_tg(td, tg);
407 static void throtl_schedule_next_dispatch(struct throtl_data *td)
409 struct throtl_rb_root *st = &td->tg_service_tree;
412 * If there are more bios pending, schedule more work.
414 if (!total_nr_queued(td))
419 update_min_dispatch_time(st);
421 if (time_before_eq(st->min_disptime, jiffies))
422 throtl_schedule_delayed_work(td, 0);
424 throtl_schedule_delayed_work(td, (st->min_disptime - jiffies));
428 throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
430 tg->bytes_disp[rw] = 0;
432 tg->slice_start[rw] = jiffies;
433 tg->slice_end[rw] = jiffies + throtl_slice;
434 throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
435 rw == READ ? 'R' : 'W', tg->slice_start[rw],
436 tg->slice_end[rw], jiffies);
439 static inline void throtl_set_slice_end(struct throtl_data *td,
440 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
442 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
445 static inline void throtl_extend_slice(struct throtl_data *td,
446 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
448 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
449 throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
450 rw == READ ? 'R' : 'W', tg->slice_start[rw],
451 tg->slice_end[rw], jiffies);
454 /* Determine if previously allocated or extended slice is complete or not */
456 throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
458 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
464 /* Trim the used slices and adjust slice start accordingly */
466 throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
468 unsigned long nr_slices, time_elapsed, io_trim;
471 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
474 * If bps are unlimited (-1), then time slice don't get
475 * renewed. Don't try to trim the slice if slice is used. A new
476 * slice will start when appropriate.
478 if (throtl_slice_used(td, tg, rw))
482 * A bio has been dispatched. Also adjust slice_end. It might happen
483 * that initially cgroup limit was very low resulting in high
484 * slice_end, but later limit was bumped up and bio was dispached
485 * sooner, then we need to reduce slice_end. A high bogus slice_end
486 * is bad because it does not allow new slice to start.
489 throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice);
491 time_elapsed = jiffies - tg->slice_start[rw];
493 nr_slices = time_elapsed / throtl_slice;
497 tmp = tg->bps[rw] * throtl_slice * nr_slices;
501 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
503 if (!bytes_trim && !io_trim)
506 if (tg->bytes_disp[rw] >= bytes_trim)
507 tg->bytes_disp[rw] -= bytes_trim;
509 tg->bytes_disp[rw] = 0;
511 if (tg->io_disp[rw] >= io_trim)
512 tg->io_disp[rw] -= io_trim;
516 tg->slice_start[rw] += nr_slices * throtl_slice;
518 throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
519 " start=%lu end=%lu jiffies=%lu",
520 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
521 tg->slice_start[rw], tg->slice_end[rw], jiffies);
524 static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
525 struct bio *bio, unsigned long *wait)
527 bool rw = bio_data_dir(bio);
528 unsigned int io_allowed;
529 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
532 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
534 /* Slice has just started. Consider one slice interval */
536 jiffy_elapsed_rnd = throtl_slice;
538 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
541 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
542 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
543 * will allow dispatch after 1 second and after that slice should
547 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
551 io_allowed = UINT_MAX;
555 if (tg->io_disp[rw] + 1 <= io_allowed) {
561 /* Calc approx time to dispatch */
562 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
564 if (jiffy_wait > jiffy_elapsed)
565 jiffy_wait = jiffy_wait - jiffy_elapsed;
574 static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
575 struct bio *bio, unsigned long *wait)
577 bool rw = bio_data_dir(bio);
578 u64 bytes_allowed, extra_bytes, tmp;
579 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
581 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
583 /* Slice has just started. Consider one slice interval */
585 jiffy_elapsed_rnd = throtl_slice;
587 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
589 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
593 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
599 /* Calc approx time to dispatch */
600 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
601 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
607 * This wait time is without taking into consideration the rounding
608 * up we did. Add that time also.
610 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
616 static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
617 if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
623 * Returns whether one can dispatch a bio or not. Also returns approx number
624 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
626 static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
627 struct bio *bio, unsigned long *wait)
629 bool rw = bio_data_dir(bio);
630 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
633 * Currently whole state machine of group depends on first bio
634 * queued in the group bio list. So one should not be calling
635 * this function with a different bio if there are other bios
638 BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
640 /* If tg->bps = -1, then BW is unlimited */
641 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
648 * If previous slice expired, start a new one otherwise renew/extend
649 * existing slice to make sure it is at least throtl_slice interval
652 if (throtl_slice_used(td, tg, rw))
653 throtl_start_new_slice(td, tg, rw);
655 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
656 throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
659 if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
660 && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
666 max_wait = max(bps_wait, iops_wait);
671 if (time_before(tg->slice_end[rw], jiffies + max_wait))
672 throtl_extend_slice(td, tg, rw, jiffies + max_wait);
677 static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
680 struct throtl_grp *tg = blkg_to_tg(blkg);
681 struct tg_stats_cpu *stats_cpu;
684 /* If per cpu stats are not allocated yet, don't do any accounting. */
685 if (tg->stats_cpu == NULL)
689 * Disabling interrupts to provide mutual exclusion between two
690 * writes on same cpu. It probably is not needed for 64bit. Not
691 * optimizing that case yet.
693 local_irq_save(flags);
695 stats_cpu = this_cpu_ptr(tg->stats_cpu);
697 blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
698 blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
700 local_irq_restore(flags);
703 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
705 bool rw = bio_data_dir(bio);
707 /* Charge the bio to the group */
708 tg->bytes_disp[rw] += bio->bi_size;
711 throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw);
714 static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
717 bool rw = bio_data_dir(bio);
719 bio_list_add(&tg->bio_lists[rw], bio);
720 /* Take a bio reference on tg */
721 blkg_get(tg_to_blkg(tg));
724 throtl_enqueue_tg(td, tg);
727 static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
729 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
732 if ((bio = bio_list_peek(&tg->bio_lists[READ])))
733 tg_may_dispatch(td, tg, bio, &read_wait);
735 if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
736 tg_may_dispatch(td, tg, bio, &write_wait);
738 min_wait = min(read_wait, write_wait);
739 disptime = jiffies + min_wait;
741 /* Update dispatch time */
742 throtl_dequeue_tg(td, tg);
743 tg->disptime = disptime;
744 throtl_enqueue_tg(td, tg);
747 static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
748 bool rw, struct bio_list *bl)
752 bio = bio_list_pop(&tg->bio_lists[rw]);
754 /* Drop bio reference on blkg */
755 blkg_put(tg_to_blkg(tg));
757 BUG_ON(td->nr_queued[rw] <= 0);
760 throtl_charge_bio(tg, bio);
761 bio_list_add(bl, bio);
762 bio->bi_rw |= REQ_THROTTLED;
764 throtl_trim_slice(td, tg, rw);
767 static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
770 unsigned int nr_reads = 0, nr_writes = 0;
771 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
772 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
775 /* Try to dispatch 75% READS and 25% WRITES */
777 while ((bio = bio_list_peek(&tg->bio_lists[READ]))
778 && tg_may_dispatch(td, tg, bio, NULL)) {
780 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
783 if (nr_reads >= max_nr_reads)
787 while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
788 && tg_may_dispatch(td, tg, bio, NULL)) {
790 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
793 if (nr_writes >= max_nr_writes)
797 return nr_reads + nr_writes;
800 static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
802 unsigned int nr_disp = 0;
803 struct throtl_grp *tg;
804 struct throtl_rb_root *st = &td->tg_service_tree;
807 tg = throtl_rb_first(st);
812 if (time_before(jiffies, tg->disptime))
815 throtl_dequeue_tg(td, tg);
817 nr_disp += throtl_dispatch_tg(td, tg, bl);
819 if (tg->nr_queued[0] || tg->nr_queued[1]) {
820 tg_update_disptime(td, tg);
821 throtl_enqueue_tg(td, tg);
824 if (nr_disp >= throtl_quantum)
831 static void throtl_process_limit_change(struct throtl_data *td)
833 struct request_queue *q = td->queue;
834 struct blkcg_gq *blkg, *n;
836 if (!td->limits_changed)
839 xchg(&td->limits_changed, false);
841 throtl_log(td, "limits changed");
843 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
844 struct throtl_grp *tg = blkg_to_tg(blkg);
846 if (!tg->limits_changed)
849 if (!xchg(&tg->limits_changed, false))
852 throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
853 " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
854 tg->iops[READ], tg->iops[WRITE]);
857 * Restart the slices for both READ and WRITES. It
858 * might happen that a group's limit are dropped
859 * suddenly and we don't want to account recently
860 * dispatched IO with new low rate
862 throtl_start_new_slice(td, tg, 0);
863 throtl_start_new_slice(td, tg, 1);
865 if (throtl_tg_on_rr(tg))
866 tg_update_disptime(td, tg);
870 /* Dispatch throttled bios. Should be called without queue lock held. */
871 static int throtl_dispatch(struct request_queue *q)
873 struct throtl_data *td = q->td;
874 unsigned int nr_disp = 0;
875 struct bio_list bio_list_on_stack;
877 struct blk_plug plug;
879 spin_lock_irq(q->queue_lock);
881 throtl_process_limit_change(td);
883 if (!total_nr_queued(td))
886 bio_list_init(&bio_list_on_stack);
888 throtl_log(td, "dispatch nr_queued=%u read=%u write=%u",
889 total_nr_queued(td), td->nr_queued[READ],
890 td->nr_queued[WRITE]);
892 nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
895 throtl_log(td, "bios disp=%u", nr_disp);
897 throtl_schedule_next_dispatch(td);
899 spin_unlock_irq(q->queue_lock);
902 * If we dispatched some requests, unplug the queue to make sure
906 blk_start_plug(&plug);
907 while((bio = bio_list_pop(&bio_list_on_stack)))
908 generic_make_request(bio);
909 blk_finish_plug(&plug);
914 void blk_throtl_work(struct work_struct *work)
916 struct throtl_data *td = container_of(work, struct throtl_data,
918 struct request_queue *q = td->queue;
923 /* Call with queue lock held */
925 throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay)
928 struct delayed_work *dwork = &td->throtl_work;
930 /* schedule work if limits changed even if no bio is queued */
931 if (total_nr_queued(td) || td->limits_changed) {
932 mod_delayed_work(kthrotld_workqueue, dwork, delay);
933 throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
938 static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
939 struct blkg_policy_data *pd, int off)
941 struct throtl_grp *tg = pd_to_tg(pd);
942 struct blkg_rwstat rwstat = { }, tmp;
945 for_each_possible_cpu(cpu) {
946 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
948 tmp = blkg_rwstat_read((void *)sc + off);
949 for (i = 0; i < BLKG_RWSTAT_NR; i++)
950 rwstat.cnt[i] += tmp.cnt[i];
953 return __blkg_prfill_rwstat(sf, pd, &rwstat);
956 static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft,
959 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
961 blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl,
966 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
969 struct throtl_grp *tg = pd_to_tg(pd);
970 u64 v = *(u64 *)((void *)tg + off);
974 return __blkg_prfill_u64(sf, pd, v);
977 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
980 struct throtl_grp *tg = pd_to_tg(pd);
981 unsigned int v = *(unsigned int *)((void *)tg + off);
985 return __blkg_prfill_u64(sf, pd, v);
988 static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft,
991 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64,
992 &blkcg_policy_throtl, cft->private, false);
996 static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft,
999 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint,
1000 &blkcg_policy_throtl, cft->private, false);
1004 static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf,
1007 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1008 struct blkg_conf_ctx ctx;
1009 struct throtl_grp *tg;
1010 struct throtl_data *td;
1013 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1017 tg = blkg_to_tg(ctx.blkg);
1018 td = ctx.blkg->q->td;
1024 *(u64 *)((void *)tg + cft->private) = ctx.v;
1026 *(unsigned int *)((void *)tg + cft->private) = ctx.v;
1028 /* XXX: we don't need the following deferred processing */
1029 xchg(&tg->limits_changed, true);
1030 xchg(&td->limits_changed, true);
1031 throtl_schedule_delayed_work(td, 0);
1033 blkg_conf_finish(&ctx);
1037 static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft,
1040 return tg_set_conf(cgrp, cft, buf, true);
1043 static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1046 return tg_set_conf(cgrp, cft, buf, false);
1049 static struct cftype throtl_files[] = {
1051 .name = "throttle.read_bps_device",
1052 .private = offsetof(struct throtl_grp, bps[READ]),
1053 .read_seq_string = tg_print_conf_u64,
1054 .write_string = tg_set_conf_u64,
1055 .max_write_len = 256,
1058 .name = "throttle.write_bps_device",
1059 .private = offsetof(struct throtl_grp, bps[WRITE]),
1060 .read_seq_string = tg_print_conf_u64,
1061 .write_string = tg_set_conf_u64,
1062 .max_write_len = 256,
1065 .name = "throttle.read_iops_device",
1066 .private = offsetof(struct throtl_grp, iops[READ]),
1067 .read_seq_string = tg_print_conf_uint,
1068 .write_string = tg_set_conf_uint,
1069 .max_write_len = 256,
1072 .name = "throttle.write_iops_device",
1073 .private = offsetof(struct throtl_grp, iops[WRITE]),
1074 .read_seq_string = tg_print_conf_uint,
1075 .write_string = tg_set_conf_uint,
1076 .max_write_len = 256,
1079 .name = "throttle.io_service_bytes",
1080 .private = offsetof(struct tg_stats_cpu, service_bytes),
1081 .read_seq_string = tg_print_cpu_rwstat,
1084 .name = "throttle.io_serviced",
1085 .private = offsetof(struct tg_stats_cpu, serviced),
1086 .read_seq_string = tg_print_cpu_rwstat,
1091 static void throtl_shutdown_wq(struct request_queue *q)
1093 struct throtl_data *td = q->td;
1095 cancel_delayed_work_sync(&td->throtl_work);
1098 static struct blkcg_policy blkcg_policy_throtl = {
1099 .pd_size = sizeof(struct throtl_grp),
1100 .cftypes = throtl_files,
1102 .pd_init_fn = throtl_pd_init,
1103 .pd_exit_fn = throtl_pd_exit,
1104 .pd_reset_stats_fn = throtl_pd_reset_stats,
1107 bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
1109 struct throtl_data *td = q->td;
1110 struct throtl_grp *tg;
1111 bool rw = bio_data_dir(bio), update_disptime = true;
1112 struct blkcg *blkcg;
1113 bool throttled = false;
1115 if (bio->bi_rw & REQ_THROTTLED) {
1116 bio->bi_rw &= ~REQ_THROTTLED;
1121 * A throtl_grp pointer retrieved under rcu can be used to access
1122 * basic fields like stats and io rates. If a group has no rules,
1123 * just update the dispatch stats in lockless manner and return.
1126 blkcg = bio_blkcg(bio);
1127 tg = throtl_lookup_tg(td, blkcg);
1129 if (tg_no_rule_group(tg, rw)) {
1130 throtl_update_dispatch_stats(tg_to_blkg(tg),
1131 bio->bi_size, bio->bi_rw);
1132 goto out_unlock_rcu;
1137 * Either group has not been allocated yet or it is not an unlimited
1140 spin_lock_irq(q->queue_lock);
1141 tg = throtl_lookup_create_tg(td, blkcg);
1145 if (tg->nr_queued[rw]) {
1147 * There is already another bio queued in same dir. No
1148 * need to update dispatch time.
1150 update_disptime = false;
1155 /* Bio is with-in rate limit of group */
1156 if (tg_may_dispatch(td, tg, bio, NULL)) {
1157 throtl_charge_bio(tg, bio);
1160 * We need to trim slice even when bios are not being queued
1161 * otherwise it might happen that a bio is not queued for
1162 * a long time and slice keeps on extending and trim is not
1163 * called for a long time. Now if limits are reduced suddenly
1164 * we take into account all the IO dispatched so far at new
1165 * low rate and * newly queued IO gets a really long dispatch
1168 * So keep on trimming slice even if bio is not queued.
1170 throtl_trim_slice(td, tg, rw);
1175 throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
1176 " iodisp=%u iops=%u queued=%d/%d",
1177 rw == READ ? 'R' : 'W',
1178 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1179 tg->io_disp[rw], tg->iops[rw],
1180 tg->nr_queued[READ], tg->nr_queued[WRITE]);
1182 bio_associate_current(bio);
1183 throtl_add_bio_tg(q->td, tg, bio);
1186 if (update_disptime) {
1187 tg_update_disptime(td, tg);
1188 throtl_schedule_next_dispatch(td);
1192 spin_unlock_irq(q->queue_lock);
1200 * blk_throtl_drain - drain throttled bios
1201 * @q: request_queue to drain throttled bios for
1203 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1205 void blk_throtl_drain(struct request_queue *q)
1206 __releases(q->queue_lock) __acquires(q->queue_lock)
1208 struct throtl_data *td = q->td;
1209 struct throtl_rb_root *st = &td->tg_service_tree;
1210 struct throtl_grp *tg;
1214 queue_lockdep_assert_held(q);
1218 while ((tg = throtl_rb_first(st))) {
1219 throtl_dequeue_tg(td, tg);
1221 while ((bio = bio_list_peek(&tg->bio_lists[READ])))
1222 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl);
1223 while ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
1224 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl);
1226 spin_unlock_irq(q->queue_lock);
1228 while ((bio = bio_list_pop(&bl)))
1229 generic_make_request(bio);
1231 spin_lock_irq(q->queue_lock);
1234 int blk_throtl_init(struct request_queue *q)
1236 struct throtl_data *td;
1239 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1243 td->tg_service_tree = THROTL_RB_ROOT;
1244 td->limits_changed = false;
1245 INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);
1250 /* activate policy */
1251 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1257 void blk_throtl_exit(struct request_queue *q)
1260 throtl_shutdown_wq(q);
1261 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1265 static int __init throtl_init(void)
1267 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1268 if (!kthrotld_workqueue)
1269 panic("Failed to create kthrotld\n");
1271 return blkcg_policy_register(&blkcg_policy_throtl);
1274 module_init(throtl_init);