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;
29 struct throtl_service_queue {
31 * Bios queued directly to this service_queue or dispatched from
32 * children throtl_grp's.
34 struct bio_list bio_lists[2]; /* queued bios [READ/WRITE] */
35 unsigned int nr_queued[2]; /* number of queued bios */
38 * RB tree of active children throtl_grp's, which are sorted by
41 struct rb_root pending_tree; /* RB tree of active tgs */
42 struct rb_node *first_pending; /* first node in the tree */
43 unsigned int nr_pending; /* # queued in the tree */
44 unsigned long first_pending_disptime; /* disptime of the first tg */
48 THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */
51 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
53 /* Per-cpu group stats */
55 /* total bytes transferred */
56 struct blkg_rwstat service_bytes;
57 /* total IOs serviced, post merge */
58 struct blkg_rwstat serviced;
62 /* must be the first member */
63 struct blkg_policy_data pd;
65 /* active throtl group service_queue member */
66 struct rb_node rb_node;
68 /* throtl_data this group belongs to */
69 struct throtl_data *td;
71 /* this group's service queue */
72 struct throtl_service_queue service_queue;
75 * Dispatch time in jiffies. This is the estimated time when group
76 * will unthrottle and is ready to dispatch more bio. It is used as
77 * key to sort active groups in service tree.
79 unsigned long disptime;
83 /* bytes per second rate limits */
89 /* Number of bytes disptached in current slice */
90 uint64_t bytes_disp[2];
91 /* Number of bio's dispatched in current slice */
92 unsigned int io_disp[2];
94 /* When did we start a new slice */
95 unsigned long slice_start[2];
96 unsigned long slice_end[2];
98 /* Per cpu stats pointer */
99 struct tg_stats_cpu __percpu *stats_cpu;
101 /* List of tgs waiting for per cpu stats memory to be allocated */
102 struct list_head stats_alloc_node;
107 /* service tree for active throtl groups */
108 struct throtl_service_queue service_queue;
110 struct request_queue *queue;
112 /* Total Number of queued bios on READ and WRITE lists */
113 unsigned int nr_queued[2];
116 * number of total undestroyed groups
118 unsigned int nr_undestroyed_grps;
120 /* Work for dispatching throttled bios */
121 struct delayed_work dispatch_work;
124 /* list and work item to allocate percpu group stats */
125 static DEFINE_SPINLOCK(tg_stats_alloc_lock);
126 static LIST_HEAD(tg_stats_alloc_list);
128 static void tg_stats_alloc_fn(struct work_struct *);
129 static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn);
131 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
133 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
136 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
138 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
141 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
143 return pd_to_blkg(&tg->pd);
146 static inline struct throtl_grp *td_root_tg(struct throtl_data *td)
148 return blkg_to_tg(td->queue->root_blkg);
151 #define throtl_log_tg(tg, fmt, args...) do { \
154 blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \
155 blk_add_trace_msg((tg)->td->queue, "throtl %s " fmt, __pbuf, ##args); \
158 #define throtl_log(td, fmt, args...) \
159 blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
162 * Worker for allocating per cpu stat for tgs. This is scheduled on the
163 * system_wq once there are some groups on the alloc_list waiting for
166 static void tg_stats_alloc_fn(struct work_struct *work)
168 static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */
169 struct delayed_work *dwork = to_delayed_work(work);
174 stats_cpu = alloc_percpu(struct tg_stats_cpu);
176 /* allocation failed, try again after some time */
177 schedule_delayed_work(dwork, msecs_to_jiffies(10));
182 spin_lock_irq(&tg_stats_alloc_lock);
184 if (!list_empty(&tg_stats_alloc_list)) {
185 struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list,
188 swap(tg->stats_cpu, stats_cpu);
189 list_del_init(&tg->stats_alloc_node);
192 empty = list_empty(&tg_stats_alloc_list);
193 spin_unlock_irq(&tg_stats_alloc_lock);
198 /* init a service_queue, assumes the caller zeroed it */
199 static void throtl_service_queue_init(struct throtl_service_queue *sq)
201 bio_list_init(&sq->bio_lists[0]);
202 bio_list_init(&sq->bio_lists[1]);
203 sq->pending_tree = RB_ROOT;
206 static void throtl_pd_init(struct blkcg_gq *blkg)
208 struct throtl_grp *tg = blkg_to_tg(blkg);
211 throtl_service_queue_init(&tg->service_queue);
212 RB_CLEAR_NODE(&tg->rb_node);
213 tg->td = blkg->q->td;
218 tg->iops[WRITE] = -1;
221 * Ugh... We need to perform per-cpu allocation for tg->stats_cpu
222 * but percpu allocator can't be called from IO path. Queue tg on
223 * tg_stats_alloc_list and allocate from work item.
225 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
226 list_add(&tg->stats_alloc_node, &tg_stats_alloc_list);
227 schedule_delayed_work(&tg_stats_alloc_work, 0);
228 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
231 static void throtl_pd_exit(struct blkcg_gq *blkg)
233 struct throtl_grp *tg = blkg_to_tg(blkg);
236 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
237 list_del_init(&tg->stats_alloc_node);
238 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
240 free_percpu(tg->stats_cpu);
243 static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
245 struct throtl_grp *tg = blkg_to_tg(blkg);
248 if (tg->stats_cpu == NULL)
251 for_each_possible_cpu(cpu) {
252 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
254 blkg_rwstat_reset(&sc->service_bytes);
255 blkg_rwstat_reset(&sc->serviced);
259 static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
263 * This is the common case when there are no blkcgs. Avoid lookup
266 if (blkcg == &blkcg_root)
267 return td_root_tg(td);
269 return blkg_to_tg(blkg_lookup(blkcg, td->queue));
272 static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
275 struct request_queue *q = td->queue;
276 struct throtl_grp *tg = NULL;
279 * This is the common case when there are no blkcgs. Avoid lookup
282 if (blkcg == &blkcg_root) {
285 struct blkcg_gq *blkg;
287 blkg = blkg_lookup_create(blkcg, q);
289 /* if %NULL and @q is alive, fall back to root_tg */
291 tg = blkg_to_tg(blkg);
292 else if (!blk_queue_dying(q))
299 static struct throtl_grp *
300 throtl_rb_first(struct throtl_service_queue *parent_sq)
302 /* Service tree is empty */
303 if (!parent_sq->nr_pending)
306 if (!parent_sq->first_pending)
307 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
309 if (parent_sq->first_pending)
310 return rb_entry_tg(parent_sq->first_pending);
315 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
321 static void throtl_rb_erase(struct rb_node *n,
322 struct throtl_service_queue *parent_sq)
324 if (parent_sq->first_pending == n)
325 parent_sq->first_pending = NULL;
326 rb_erase_init(n, &parent_sq->pending_tree);
327 --parent_sq->nr_pending;
330 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
332 struct throtl_grp *tg;
334 tg = throtl_rb_first(parent_sq);
338 parent_sq->first_pending_disptime = tg->disptime;
341 static void tg_service_queue_add(struct throtl_grp *tg,
342 struct throtl_service_queue *parent_sq)
344 struct rb_node **node = &parent_sq->pending_tree.rb_node;
345 struct rb_node *parent = NULL;
346 struct throtl_grp *__tg;
347 unsigned long key = tg->disptime;
350 while (*node != NULL) {
352 __tg = rb_entry_tg(parent);
354 if (time_before(key, __tg->disptime))
355 node = &parent->rb_left;
357 node = &parent->rb_right;
363 parent_sq->first_pending = &tg->rb_node;
365 rb_link_node(&tg->rb_node, parent, node);
366 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
369 static void __throtl_enqueue_tg(struct throtl_grp *tg,
370 struct throtl_service_queue *parent_sq)
372 tg_service_queue_add(tg, parent_sq);
373 tg->flags |= THROTL_TG_PENDING;
374 parent_sq->nr_pending++;
377 static void throtl_enqueue_tg(struct throtl_grp *tg,
378 struct throtl_service_queue *parent_sq)
380 if (!(tg->flags & THROTL_TG_PENDING))
381 __throtl_enqueue_tg(tg, parent_sq);
384 static void __throtl_dequeue_tg(struct throtl_grp *tg,
385 struct throtl_service_queue *parent_sq)
387 throtl_rb_erase(&tg->rb_node, parent_sq);
388 tg->flags &= ~THROTL_TG_PENDING;
391 static void throtl_dequeue_tg(struct throtl_grp *tg,
392 struct throtl_service_queue *parent_sq)
394 if (tg->flags & THROTL_TG_PENDING)
395 __throtl_dequeue_tg(tg, parent_sq);
398 /* Call with queue lock held */
399 static void throtl_schedule_delayed_work(struct throtl_data *td,
402 struct delayed_work *dwork = &td->dispatch_work;
404 mod_delayed_work(kthrotld_workqueue, dwork, delay);
405 throtl_log(td, "schedule work. delay=%lu jiffies=%lu", delay, jiffies);
408 static void throtl_schedule_next_dispatch(struct throtl_data *td)
410 struct throtl_service_queue *sq = &td->service_queue;
412 /* any pending children left? */
416 update_min_dispatch_time(sq);
418 if (time_before_eq(sq->first_pending_disptime, jiffies))
419 throtl_schedule_delayed_work(td, 0);
421 throtl_schedule_delayed_work(td, sq->first_pending_disptime - jiffies);
424 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
426 tg->bytes_disp[rw] = 0;
428 tg->slice_start[rw] = jiffies;
429 tg->slice_end[rw] = jiffies + throtl_slice;
430 throtl_log_tg(tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
431 rw == READ ? 'R' : 'W', tg->slice_start[rw],
432 tg->slice_end[rw], jiffies);
435 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
436 unsigned long jiffy_end)
438 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
441 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
442 unsigned long jiffy_end)
444 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
445 throtl_log_tg(tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
446 rw == READ ? 'R' : 'W', tg->slice_start[rw],
447 tg->slice_end[rw], jiffies);
450 /* Determine if previously allocated or extended slice is complete or not */
451 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
453 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
459 /* Trim the used slices and adjust slice start accordingly */
460 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
462 unsigned long nr_slices, time_elapsed, io_trim;
465 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
468 * If bps are unlimited (-1), then time slice don't get
469 * renewed. Don't try to trim the slice if slice is used. A new
470 * slice will start when appropriate.
472 if (throtl_slice_used(tg, rw))
476 * A bio has been dispatched. Also adjust slice_end. It might happen
477 * that initially cgroup limit was very low resulting in high
478 * slice_end, but later limit was bumped up and bio was dispached
479 * sooner, then we need to reduce slice_end. A high bogus slice_end
480 * is bad because it does not allow new slice to start.
483 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
485 time_elapsed = jiffies - tg->slice_start[rw];
487 nr_slices = time_elapsed / throtl_slice;
491 tmp = tg->bps[rw] * throtl_slice * nr_slices;
495 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
497 if (!bytes_trim && !io_trim)
500 if (tg->bytes_disp[rw] >= bytes_trim)
501 tg->bytes_disp[rw] -= bytes_trim;
503 tg->bytes_disp[rw] = 0;
505 if (tg->io_disp[rw] >= io_trim)
506 tg->io_disp[rw] -= io_trim;
510 tg->slice_start[rw] += nr_slices * throtl_slice;
512 throtl_log_tg(tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
513 " start=%lu end=%lu jiffies=%lu",
514 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
515 tg->slice_start[rw], tg->slice_end[rw], jiffies);
518 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
521 bool rw = bio_data_dir(bio);
522 unsigned int io_allowed;
523 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
526 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
528 /* Slice has just started. Consider one slice interval */
530 jiffy_elapsed_rnd = throtl_slice;
532 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
535 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
536 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
537 * will allow dispatch after 1 second and after that slice should
541 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
545 io_allowed = UINT_MAX;
549 if (tg->io_disp[rw] + 1 <= io_allowed) {
555 /* Calc approx time to dispatch */
556 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
558 if (jiffy_wait > jiffy_elapsed)
559 jiffy_wait = jiffy_wait - jiffy_elapsed;
568 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
571 bool rw = bio_data_dir(bio);
572 u64 bytes_allowed, extra_bytes, tmp;
573 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
575 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
577 /* Slice has just started. Consider one slice interval */
579 jiffy_elapsed_rnd = throtl_slice;
581 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
583 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
587 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
593 /* Calc approx time to dispatch */
594 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
595 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
601 * This wait time is without taking into consideration the rounding
602 * up we did. Add that time also.
604 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
610 static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
611 if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
617 * Returns whether one can dispatch a bio or not. Also returns approx number
618 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
620 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
623 bool rw = bio_data_dir(bio);
624 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
627 * Currently whole state machine of group depends on first bio
628 * queued in the group bio list. So one should not be calling
629 * this function with a different bio if there are other bios
632 BUG_ON(tg->service_queue.nr_queued[rw] &&
633 bio != bio_list_peek(&tg->service_queue.bio_lists[rw]));
635 /* If tg->bps = -1, then BW is unlimited */
636 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
643 * If previous slice expired, start a new one otherwise renew/extend
644 * existing slice to make sure it is at least throtl_slice interval
647 if (throtl_slice_used(tg, rw))
648 throtl_start_new_slice(tg, rw);
650 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
651 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
654 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
655 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
661 max_wait = max(bps_wait, iops_wait);
666 if (time_before(tg->slice_end[rw], jiffies + max_wait))
667 throtl_extend_slice(tg, rw, jiffies + max_wait);
672 static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
675 struct throtl_grp *tg = blkg_to_tg(blkg);
676 struct tg_stats_cpu *stats_cpu;
679 /* If per cpu stats are not allocated yet, don't do any accounting. */
680 if (tg->stats_cpu == NULL)
684 * Disabling interrupts to provide mutual exclusion between two
685 * writes on same cpu. It probably is not needed for 64bit. Not
686 * optimizing that case yet.
688 local_irq_save(flags);
690 stats_cpu = this_cpu_ptr(tg->stats_cpu);
692 blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
693 blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
695 local_irq_restore(flags);
698 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
700 bool rw = bio_data_dir(bio);
702 /* Charge the bio to the group */
703 tg->bytes_disp[rw] += bio->bi_size;
706 throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw);
709 static void throtl_add_bio_tg(struct bio *bio, struct throtl_grp *tg,
710 struct throtl_service_queue *parent_sq)
712 struct throtl_service_queue *sq = &tg->service_queue;
713 bool rw = bio_data_dir(bio);
715 bio_list_add(&sq->bio_lists[rw], bio);
716 /* Take a bio reference on tg */
717 blkg_get(tg_to_blkg(tg));
719 tg->td->nr_queued[rw]++;
720 throtl_enqueue_tg(tg, parent_sq);
723 static void tg_update_disptime(struct throtl_grp *tg,
724 struct throtl_service_queue *parent_sq)
726 struct throtl_service_queue *sq = &tg->service_queue;
727 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
730 if ((bio = bio_list_peek(&sq->bio_lists[READ])))
731 tg_may_dispatch(tg, bio, &read_wait);
733 if ((bio = bio_list_peek(&sq->bio_lists[WRITE])))
734 tg_may_dispatch(tg, bio, &write_wait);
736 min_wait = min(read_wait, write_wait);
737 disptime = jiffies + min_wait;
739 /* Update dispatch time */
740 throtl_dequeue_tg(tg, parent_sq);
741 tg->disptime = disptime;
742 throtl_enqueue_tg(tg, parent_sq);
745 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw,
746 struct throtl_service_queue *parent_sq)
748 struct throtl_service_queue *sq = &tg->service_queue;
751 bio = bio_list_pop(&sq->bio_lists[rw]);
753 /* Drop bio reference on blkg */
754 blkg_put(tg_to_blkg(tg));
756 BUG_ON(tg->td->nr_queued[rw] <= 0);
757 tg->td->nr_queued[rw]--;
759 throtl_charge_bio(tg, bio);
760 bio_list_add(&parent_sq->bio_lists[rw], bio);
761 bio->bi_rw |= REQ_THROTTLED;
763 throtl_trim_slice(tg, rw);
766 static int throtl_dispatch_tg(struct throtl_grp *tg,
767 struct throtl_service_queue *parent_sq)
769 struct throtl_service_queue *sq = &tg->service_queue;
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(&sq->bio_lists[READ])) &&
778 tg_may_dispatch(tg, bio, NULL)) {
780 tg_dispatch_one_bio(tg, bio_data_dir(bio), parent_sq);
783 if (nr_reads >= max_nr_reads)
787 while ((bio = bio_list_peek(&sq->bio_lists[WRITE])) &&
788 tg_may_dispatch(tg, bio, NULL)) {
790 tg_dispatch_one_bio(tg, bio_data_dir(bio), parent_sq);
793 if (nr_writes >= max_nr_writes)
797 return nr_reads + nr_writes;
800 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
802 unsigned int nr_disp = 0;
805 struct throtl_grp *tg = throtl_rb_first(parent_sq);
806 struct throtl_service_queue *sq = &tg->service_queue;
811 if (time_before(jiffies, tg->disptime))
814 throtl_dequeue_tg(tg, parent_sq);
816 nr_disp += throtl_dispatch_tg(tg, parent_sq);
818 if (sq->nr_queued[0] || sq->nr_queued[1])
819 tg_update_disptime(tg, parent_sq);
821 if (nr_disp >= throtl_quantum)
828 /* work function to dispatch throttled bios */
829 void blk_throtl_dispatch_work_fn(struct work_struct *work)
831 struct throtl_data *td = container_of(to_delayed_work(work),
832 struct throtl_data, dispatch_work);
833 struct throtl_service_queue *sq = &td->service_queue;
834 struct request_queue *q = td->queue;
835 unsigned int nr_disp = 0;
836 struct bio_list bio_list_on_stack;
838 struct blk_plug plug;
841 spin_lock_irq(q->queue_lock);
843 bio_list_init(&bio_list_on_stack);
845 throtl_log(td, "dispatch nr_queued=%u read=%u write=%u",
846 td->nr_queued[READ] + td->nr_queued[WRITE],
847 td->nr_queued[READ], td->nr_queued[WRITE]);
849 nr_disp = throtl_select_dispatch(sq);
852 for (rw = READ; rw <= WRITE; rw++) {
853 bio_list_merge(&bio_list_on_stack, &sq->bio_lists[rw]);
854 bio_list_init(&sq->bio_lists[rw]);
856 throtl_log(td, "bios disp=%u", nr_disp);
859 throtl_schedule_next_dispatch(td);
861 spin_unlock_irq(q->queue_lock);
864 * If we dispatched some requests, unplug the queue to make sure
868 blk_start_plug(&plug);
869 while((bio = bio_list_pop(&bio_list_on_stack)))
870 generic_make_request(bio);
871 blk_finish_plug(&plug);
875 static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
876 struct blkg_policy_data *pd, int off)
878 struct throtl_grp *tg = pd_to_tg(pd);
879 struct blkg_rwstat rwstat = { }, tmp;
882 for_each_possible_cpu(cpu) {
883 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
885 tmp = blkg_rwstat_read((void *)sc + off);
886 for (i = 0; i < BLKG_RWSTAT_NR; i++)
887 rwstat.cnt[i] += tmp.cnt[i];
890 return __blkg_prfill_rwstat(sf, pd, &rwstat);
893 static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft,
896 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
898 blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl,
903 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
906 struct throtl_grp *tg = pd_to_tg(pd);
907 u64 v = *(u64 *)((void *)tg + off);
911 return __blkg_prfill_u64(sf, pd, v);
914 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
917 struct throtl_grp *tg = pd_to_tg(pd);
918 unsigned int v = *(unsigned int *)((void *)tg + off);
922 return __blkg_prfill_u64(sf, pd, v);
925 static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft,
928 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64,
929 &blkcg_policy_throtl, cft->private, false);
933 static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft,
936 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint,
937 &blkcg_policy_throtl, cft->private, false);
941 static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf,
944 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
945 struct blkg_conf_ctx ctx;
946 struct throtl_grp *tg;
947 struct throtl_data *td;
950 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
954 tg = blkg_to_tg(ctx.blkg);
955 td = ctx.blkg->q->td;
961 *(u64 *)((void *)tg + cft->private) = ctx.v;
963 *(unsigned int *)((void *)tg + cft->private) = ctx.v;
965 throtl_log_tg(tg, "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
966 tg->bps[READ], tg->bps[WRITE],
967 tg->iops[READ], tg->iops[WRITE]);
970 * We're already holding queue_lock and know @tg is valid. Let's
971 * apply the new config directly.
973 * Restart the slices for both READ and WRITES. It might happen
974 * that a group's limit are dropped suddenly and we don't want to
975 * account recently dispatched IO with new low rate.
977 throtl_start_new_slice(tg, 0);
978 throtl_start_new_slice(tg, 1);
980 if (tg->flags & THROTL_TG_PENDING) {
981 tg_update_disptime(tg, &td->service_queue);
982 throtl_schedule_next_dispatch(td);
985 blkg_conf_finish(&ctx);
989 static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft,
992 return tg_set_conf(cgrp, cft, buf, true);
995 static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft,
998 return tg_set_conf(cgrp, cft, buf, false);
1001 static struct cftype throtl_files[] = {
1003 .name = "throttle.read_bps_device",
1004 .private = offsetof(struct throtl_grp, bps[READ]),
1005 .read_seq_string = tg_print_conf_u64,
1006 .write_string = tg_set_conf_u64,
1007 .max_write_len = 256,
1010 .name = "throttle.write_bps_device",
1011 .private = offsetof(struct throtl_grp, bps[WRITE]),
1012 .read_seq_string = tg_print_conf_u64,
1013 .write_string = tg_set_conf_u64,
1014 .max_write_len = 256,
1017 .name = "throttle.read_iops_device",
1018 .private = offsetof(struct throtl_grp, iops[READ]),
1019 .read_seq_string = tg_print_conf_uint,
1020 .write_string = tg_set_conf_uint,
1021 .max_write_len = 256,
1024 .name = "throttle.write_iops_device",
1025 .private = offsetof(struct throtl_grp, iops[WRITE]),
1026 .read_seq_string = tg_print_conf_uint,
1027 .write_string = tg_set_conf_uint,
1028 .max_write_len = 256,
1031 .name = "throttle.io_service_bytes",
1032 .private = offsetof(struct tg_stats_cpu, service_bytes),
1033 .read_seq_string = tg_print_cpu_rwstat,
1036 .name = "throttle.io_serviced",
1037 .private = offsetof(struct tg_stats_cpu, serviced),
1038 .read_seq_string = tg_print_cpu_rwstat,
1043 static void throtl_shutdown_wq(struct request_queue *q)
1045 struct throtl_data *td = q->td;
1047 cancel_delayed_work_sync(&td->dispatch_work);
1050 static struct blkcg_policy blkcg_policy_throtl = {
1051 .pd_size = sizeof(struct throtl_grp),
1052 .cftypes = throtl_files,
1054 .pd_init_fn = throtl_pd_init,
1055 .pd_exit_fn = throtl_pd_exit,
1056 .pd_reset_stats_fn = throtl_pd_reset_stats,
1059 bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
1061 struct throtl_data *td = q->td;
1062 struct throtl_grp *tg;
1063 struct throtl_service_queue *sq;
1064 bool rw = bio_data_dir(bio), update_disptime = true;
1065 struct blkcg *blkcg;
1066 bool throttled = false;
1068 if (bio->bi_rw & REQ_THROTTLED) {
1069 bio->bi_rw &= ~REQ_THROTTLED;
1074 * A throtl_grp pointer retrieved under rcu can be used to access
1075 * basic fields like stats and io rates. If a group has no rules,
1076 * just update the dispatch stats in lockless manner and return.
1079 blkcg = bio_blkcg(bio);
1080 tg = throtl_lookup_tg(td, blkcg);
1082 if (tg_no_rule_group(tg, rw)) {
1083 throtl_update_dispatch_stats(tg_to_blkg(tg),
1084 bio->bi_size, bio->bi_rw);
1085 goto out_unlock_rcu;
1090 * Either group has not been allocated yet or it is not an unlimited
1093 spin_lock_irq(q->queue_lock);
1094 tg = throtl_lookup_create_tg(td, blkcg);
1098 sq = &tg->service_queue;
1100 if (sq->nr_queued[rw]) {
1102 * There is already another bio queued in same dir. No
1103 * need to update dispatch time.
1105 update_disptime = false;
1110 /* Bio is with-in rate limit of group */
1111 if (tg_may_dispatch(tg, bio, NULL)) {
1112 throtl_charge_bio(tg, bio);
1115 * We need to trim slice even when bios are not being queued
1116 * otherwise it might happen that a bio is not queued for
1117 * a long time and slice keeps on extending and trim is not
1118 * called for a long time. Now if limits are reduced suddenly
1119 * we take into account all the IO dispatched so far at new
1120 * low rate and * newly queued IO gets a really long dispatch
1123 * So keep on trimming slice even if bio is not queued.
1125 throtl_trim_slice(tg, rw);
1130 throtl_log_tg(tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
1131 " iodisp=%u iops=%u queued=%d/%d",
1132 rw == READ ? 'R' : 'W',
1133 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1134 tg->io_disp[rw], tg->iops[rw],
1135 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1137 bio_associate_current(bio);
1138 throtl_add_bio_tg(bio, tg, &q->td->service_queue);
1141 if (update_disptime) {
1142 tg_update_disptime(tg, &td->service_queue);
1143 throtl_schedule_next_dispatch(td);
1147 spin_unlock_irq(q->queue_lock);
1155 * blk_throtl_drain - drain throttled bios
1156 * @q: request_queue to drain throttled bios for
1158 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1160 void blk_throtl_drain(struct request_queue *q)
1161 __releases(q->queue_lock) __acquires(q->queue_lock)
1163 struct throtl_data *td = q->td;
1164 struct throtl_service_queue *parent_sq = &td->service_queue;
1165 struct throtl_grp *tg;
1169 queue_lockdep_assert_held(q);
1171 while ((tg = throtl_rb_first(parent_sq))) {
1172 struct throtl_service_queue *sq = &tg->service_queue;
1174 throtl_dequeue_tg(tg, parent_sq);
1176 while ((bio = bio_list_peek(&sq->bio_lists[READ])))
1177 tg_dispatch_one_bio(tg, bio_data_dir(bio), parent_sq);
1178 while ((bio = bio_list_peek(&sq->bio_lists[WRITE])))
1179 tg_dispatch_one_bio(tg, bio_data_dir(bio), parent_sq);
1181 spin_unlock_irq(q->queue_lock);
1183 for (rw = READ; rw <= WRITE; rw++)
1184 while ((bio = bio_list_pop(&parent_sq->bio_lists[rw])))
1185 generic_make_request(bio);
1187 spin_lock_irq(q->queue_lock);
1190 int blk_throtl_init(struct request_queue *q)
1192 struct throtl_data *td;
1195 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1199 INIT_DELAYED_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1200 throtl_service_queue_init(&td->service_queue);
1205 /* activate policy */
1206 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1212 void blk_throtl_exit(struct request_queue *q)
1215 throtl_shutdown_wq(q);
1216 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1220 static int __init throtl_init(void)
1222 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1223 if (!kthrotld_workqueue)
1224 panic("Failed to create kthrotld\n");
1226 return blkcg_policy_register(&blkcg_policy_throtl);
1229 module_init(throtl_init);