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 {
30 struct rb_root pending_tree; /* RB tree of active tgs */
31 struct rb_node *first_pending; /* first node in the tree */
32 unsigned int nr_pending; /* # queued in the tree */
33 unsigned long first_pending_disptime; /* disptime of the first tg */
36 #define THROTL_SERVICE_QUEUE_INITIALIZER \
37 (struct throtl_service_queue){ .pending_tree = RB_ROOT }
40 THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */
43 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
45 /* Per-cpu group stats */
47 /* total bytes transferred */
48 struct blkg_rwstat service_bytes;
49 /* total IOs serviced, post merge */
50 struct blkg_rwstat serviced;
54 /* must be the first member */
55 struct blkg_policy_data pd;
57 /* active throtl group service_queue member */
58 struct rb_node rb_node;
60 /* throtl_data this group belongs to */
61 struct throtl_data *td;
64 * Dispatch time in jiffies. This is the estimated time when group
65 * will unthrottle and is ready to dispatch more bio. It is used as
66 * key to sort active groups in service tree.
68 unsigned long disptime;
72 /* Two lists for READ and WRITE */
73 struct bio_list bio_lists[2];
75 /* Number of queued bios on READ and WRITE lists */
76 unsigned int nr_queued[2];
78 /* bytes per second rate limits */
84 /* Number of bytes disptached in current slice */
85 uint64_t bytes_disp[2];
86 /* Number of bio's dispatched in current slice */
87 unsigned int io_disp[2];
89 /* When did we start a new slice */
90 unsigned long slice_start[2];
91 unsigned long slice_end[2];
93 /* Per cpu stats pointer */
94 struct tg_stats_cpu __percpu *stats_cpu;
96 /* List of tgs waiting for per cpu stats memory to be allocated */
97 struct list_head stats_alloc_node;
102 /* service tree for active throtl groups */
103 struct throtl_service_queue service_queue;
105 struct request_queue *queue;
107 /* Total Number of queued bios on READ and WRITE lists */
108 unsigned int nr_queued[2];
111 * number of total undestroyed groups
113 unsigned int nr_undestroyed_grps;
115 /* Work for dispatching throttled bios */
116 struct delayed_work dispatch_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 #define throtl_log_tg(tg, fmt, args...) do { \
149 blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \
150 blk_add_trace_msg((tg)->td->queue, "throtl %s " fmt, __pbuf, ##args); \
153 #define throtl_log(td, fmt, args...) \
154 blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
157 * Worker for allocating per cpu stat for tgs. This is scheduled on the
158 * system_wq once there are some groups on the alloc_list waiting for
161 static void tg_stats_alloc_fn(struct work_struct *work)
163 static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */
164 struct delayed_work *dwork = to_delayed_work(work);
169 stats_cpu = alloc_percpu(struct tg_stats_cpu);
171 /* allocation failed, try again after some time */
172 schedule_delayed_work(dwork, msecs_to_jiffies(10));
177 spin_lock_irq(&tg_stats_alloc_lock);
179 if (!list_empty(&tg_stats_alloc_list)) {
180 struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list,
183 swap(tg->stats_cpu, stats_cpu);
184 list_del_init(&tg->stats_alloc_node);
187 empty = list_empty(&tg_stats_alloc_list);
188 spin_unlock_irq(&tg_stats_alloc_lock);
193 static void throtl_pd_init(struct blkcg_gq *blkg)
195 struct throtl_grp *tg = blkg_to_tg(blkg);
198 RB_CLEAR_NODE(&tg->rb_node);
199 tg->td = blkg->q->td;
200 bio_list_init(&tg->bio_lists[0]);
201 bio_list_init(&tg->bio_lists[1]);
206 tg->iops[WRITE] = -1;
209 * Ugh... We need to perform per-cpu allocation for tg->stats_cpu
210 * but percpu allocator can't be called from IO path. Queue tg on
211 * tg_stats_alloc_list and allocate from work item.
213 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
214 list_add(&tg->stats_alloc_node, &tg_stats_alloc_list);
215 schedule_delayed_work(&tg_stats_alloc_work, 0);
216 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
219 static void throtl_pd_exit(struct blkcg_gq *blkg)
221 struct throtl_grp *tg = blkg_to_tg(blkg);
224 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
225 list_del_init(&tg->stats_alloc_node);
226 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
228 free_percpu(tg->stats_cpu);
231 static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
233 struct throtl_grp *tg = blkg_to_tg(blkg);
236 if (tg->stats_cpu == NULL)
239 for_each_possible_cpu(cpu) {
240 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
242 blkg_rwstat_reset(&sc->service_bytes);
243 blkg_rwstat_reset(&sc->serviced);
247 static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
251 * This is the common case when there are no blkcgs. Avoid lookup
254 if (blkcg == &blkcg_root)
255 return td_root_tg(td);
257 return blkg_to_tg(blkg_lookup(blkcg, td->queue));
260 static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
263 struct request_queue *q = td->queue;
264 struct throtl_grp *tg = NULL;
267 * This is the common case when there are no blkcgs. Avoid lookup
270 if (blkcg == &blkcg_root) {
273 struct blkcg_gq *blkg;
275 blkg = blkg_lookup_create(blkcg, q);
277 /* if %NULL and @q is alive, fall back to root_tg */
279 tg = blkg_to_tg(blkg);
280 else if (!blk_queue_dying(q))
287 static struct throtl_grp *
288 throtl_rb_first(struct throtl_service_queue *parent_sq)
290 /* Service tree is empty */
291 if (!parent_sq->nr_pending)
294 if (!parent_sq->first_pending)
295 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
297 if (parent_sq->first_pending)
298 return rb_entry_tg(parent_sq->first_pending);
303 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
309 static void throtl_rb_erase(struct rb_node *n,
310 struct throtl_service_queue *parent_sq)
312 if (parent_sq->first_pending == n)
313 parent_sq->first_pending = NULL;
314 rb_erase_init(n, &parent_sq->pending_tree);
315 --parent_sq->nr_pending;
318 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
320 struct throtl_grp *tg;
322 tg = throtl_rb_first(parent_sq);
326 parent_sq->first_pending_disptime = tg->disptime;
329 static void tg_service_queue_add(struct throtl_grp *tg,
330 struct throtl_service_queue *parent_sq)
332 struct rb_node **node = &parent_sq->pending_tree.rb_node;
333 struct rb_node *parent = NULL;
334 struct throtl_grp *__tg;
335 unsigned long key = tg->disptime;
338 while (*node != NULL) {
340 __tg = rb_entry_tg(parent);
342 if (time_before(key, __tg->disptime))
343 node = &parent->rb_left;
345 node = &parent->rb_right;
351 parent_sq->first_pending = &tg->rb_node;
353 rb_link_node(&tg->rb_node, parent, node);
354 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
357 static void __throtl_enqueue_tg(struct throtl_grp *tg,
358 struct throtl_service_queue *parent_sq)
360 tg_service_queue_add(tg, parent_sq);
361 tg->flags |= THROTL_TG_PENDING;
362 parent_sq->nr_pending++;
365 static void throtl_enqueue_tg(struct throtl_grp *tg,
366 struct throtl_service_queue *parent_sq)
368 if (!(tg->flags & THROTL_TG_PENDING))
369 __throtl_enqueue_tg(tg, parent_sq);
372 static void __throtl_dequeue_tg(struct throtl_grp *tg,
373 struct throtl_service_queue *parent_sq)
375 throtl_rb_erase(&tg->rb_node, parent_sq);
376 tg->flags &= ~THROTL_TG_PENDING;
379 static void throtl_dequeue_tg(struct throtl_grp *tg,
380 struct throtl_service_queue *parent_sq)
382 if (tg->flags & THROTL_TG_PENDING)
383 __throtl_dequeue_tg(tg, parent_sq);
386 /* Call with queue lock held */
387 static void throtl_schedule_delayed_work(struct throtl_data *td,
390 struct delayed_work *dwork = &td->dispatch_work;
392 mod_delayed_work(kthrotld_workqueue, dwork, delay);
393 throtl_log(td, "schedule work. delay=%lu jiffies=%lu", delay, jiffies);
396 static void throtl_schedule_next_dispatch(struct throtl_data *td)
398 struct throtl_service_queue *sq = &td->service_queue;
400 /* any pending children left? */
404 update_min_dispatch_time(sq);
406 if (time_before_eq(sq->first_pending_disptime, jiffies))
407 throtl_schedule_delayed_work(td, 0);
409 throtl_schedule_delayed_work(td, sq->first_pending_disptime - jiffies);
412 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
414 tg->bytes_disp[rw] = 0;
416 tg->slice_start[rw] = jiffies;
417 tg->slice_end[rw] = jiffies + throtl_slice;
418 throtl_log_tg(tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
419 rw == READ ? 'R' : 'W', tg->slice_start[rw],
420 tg->slice_end[rw], jiffies);
423 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
424 unsigned long jiffy_end)
426 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
429 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
430 unsigned long jiffy_end)
432 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
433 throtl_log_tg(tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
434 rw == READ ? 'R' : 'W', tg->slice_start[rw],
435 tg->slice_end[rw], jiffies);
438 /* Determine if previously allocated or extended slice is complete or not */
439 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
441 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
447 /* Trim the used slices and adjust slice start accordingly */
448 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
450 unsigned long nr_slices, time_elapsed, io_trim;
453 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
456 * If bps are unlimited (-1), then time slice don't get
457 * renewed. Don't try to trim the slice if slice is used. A new
458 * slice will start when appropriate.
460 if (throtl_slice_used(tg, rw))
464 * A bio has been dispatched. Also adjust slice_end. It might happen
465 * that initially cgroup limit was very low resulting in high
466 * slice_end, but later limit was bumped up and bio was dispached
467 * sooner, then we need to reduce slice_end. A high bogus slice_end
468 * is bad because it does not allow new slice to start.
471 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
473 time_elapsed = jiffies - tg->slice_start[rw];
475 nr_slices = time_elapsed / throtl_slice;
479 tmp = tg->bps[rw] * throtl_slice * nr_slices;
483 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
485 if (!bytes_trim && !io_trim)
488 if (tg->bytes_disp[rw] >= bytes_trim)
489 tg->bytes_disp[rw] -= bytes_trim;
491 tg->bytes_disp[rw] = 0;
493 if (tg->io_disp[rw] >= io_trim)
494 tg->io_disp[rw] -= io_trim;
498 tg->slice_start[rw] += nr_slices * throtl_slice;
500 throtl_log_tg(tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
501 " start=%lu end=%lu jiffies=%lu",
502 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
503 tg->slice_start[rw], tg->slice_end[rw], jiffies);
506 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
509 bool rw = bio_data_dir(bio);
510 unsigned int io_allowed;
511 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
514 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
516 /* Slice has just started. Consider one slice interval */
518 jiffy_elapsed_rnd = throtl_slice;
520 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
523 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
524 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
525 * will allow dispatch after 1 second and after that slice should
529 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
533 io_allowed = UINT_MAX;
537 if (tg->io_disp[rw] + 1 <= io_allowed) {
543 /* Calc approx time to dispatch */
544 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
546 if (jiffy_wait > jiffy_elapsed)
547 jiffy_wait = jiffy_wait - jiffy_elapsed;
556 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
559 bool rw = bio_data_dir(bio);
560 u64 bytes_allowed, extra_bytes, tmp;
561 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
563 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
565 /* Slice has just started. Consider one slice interval */
567 jiffy_elapsed_rnd = throtl_slice;
569 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
571 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
575 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
581 /* Calc approx time to dispatch */
582 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
583 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
589 * This wait time is without taking into consideration the rounding
590 * up we did. Add that time also.
592 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
598 static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
599 if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
605 * Returns whether one can dispatch a bio or not. Also returns approx number
606 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
608 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
611 bool rw = bio_data_dir(bio);
612 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
615 * Currently whole state machine of group depends on first bio
616 * queued in the group bio list. So one should not be calling
617 * this function with a different bio if there are other bios
620 BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
622 /* If tg->bps = -1, then BW is unlimited */
623 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
630 * If previous slice expired, start a new one otherwise renew/extend
631 * existing slice to make sure it is at least throtl_slice interval
634 if (throtl_slice_used(tg, rw))
635 throtl_start_new_slice(tg, rw);
637 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
638 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
641 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
642 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
648 max_wait = max(bps_wait, iops_wait);
653 if (time_before(tg->slice_end[rw], jiffies + max_wait))
654 throtl_extend_slice(tg, rw, jiffies + max_wait);
659 static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
662 struct throtl_grp *tg = blkg_to_tg(blkg);
663 struct tg_stats_cpu *stats_cpu;
666 /* If per cpu stats are not allocated yet, don't do any accounting. */
667 if (tg->stats_cpu == NULL)
671 * Disabling interrupts to provide mutual exclusion between two
672 * writes on same cpu. It probably is not needed for 64bit. Not
673 * optimizing that case yet.
675 local_irq_save(flags);
677 stats_cpu = this_cpu_ptr(tg->stats_cpu);
679 blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
680 blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
682 local_irq_restore(flags);
685 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
687 bool rw = bio_data_dir(bio);
689 /* Charge the bio to the group */
690 tg->bytes_disp[rw] += bio->bi_size;
693 throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw);
696 static void throtl_add_bio_tg(struct bio *bio, struct throtl_grp *tg,
697 struct throtl_service_queue *parent_sq)
699 bool rw = bio_data_dir(bio);
701 bio_list_add(&tg->bio_lists[rw], bio);
702 /* Take a bio reference on tg */
703 blkg_get(tg_to_blkg(tg));
705 tg->td->nr_queued[rw]++;
706 throtl_enqueue_tg(tg, parent_sq);
709 static void tg_update_disptime(struct throtl_grp *tg,
710 struct throtl_service_queue *parent_sq)
712 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
715 if ((bio = bio_list_peek(&tg->bio_lists[READ])))
716 tg_may_dispatch(tg, bio, &read_wait);
718 if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
719 tg_may_dispatch(tg, bio, &write_wait);
721 min_wait = min(read_wait, write_wait);
722 disptime = jiffies + min_wait;
724 /* Update dispatch time */
725 throtl_dequeue_tg(tg, parent_sq);
726 tg->disptime = disptime;
727 throtl_enqueue_tg(tg, parent_sq);
730 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw,
735 bio = bio_list_pop(&tg->bio_lists[rw]);
737 /* Drop bio reference on blkg */
738 blkg_put(tg_to_blkg(tg));
740 BUG_ON(tg->td->nr_queued[rw] <= 0);
741 tg->td->nr_queued[rw]--;
743 throtl_charge_bio(tg, bio);
744 bio_list_add(bl, bio);
745 bio->bi_rw |= REQ_THROTTLED;
747 throtl_trim_slice(tg, rw);
750 static int throtl_dispatch_tg(struct throtl_grp *tg, struct bio_list *bl)
752 unsigned int nr_reads = 0, nr_writes = 0;
753 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
754 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
757 /* Try to dispatch 75% READS and 25% WRITES */
759 while ((bio = bio_list_peek(&tg->bio_lists[READ])) &&
760 tg_may_dispatch(tg, bio, NULL)) {
762 tg_dispatch_one_bio(tg, bio_data_dir(bio), bl);
765 if (nr_reads >= max_nr_reads)
769 while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) &&
770 tg_may_dispatch(tg, bio, NULL)) {
772 tg_dispatch_one_bio(tg, bio_data_dir(bio), bl);
775 if (nr_writes >= max_nr_writes)
779 return nr_reads + nr_writes;
782 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq,
785 unsigned int nr_disp = 0;
786 struct throtl_grp *tg;
789 tg = throtl_rb_first(parent_sq);
794 if (time_before(jiffies, tg->disptime))
797 throtl_dequeue_tg(tg, parent_sq);
799 nr_disp += throtl_dispatch_tg(tg, bl);
801 if (tg->nr_queued[0] || tg->nr_queued[1])
802 tg_update_disptime(tg, parent_sq);
804 if (nr_disp >= throtl_quantum)
811 /* work function to dispatch throttled bios */
812 void blk_throtl_dispatch_work_fn(struct work_struct *work)
814 struct throtl_data *td = container_of(to_delayed_work(work),
815 struct throtl_data, dispatch_work);
816 struct request_queue *q = td->queue;
817 unsigned int nr_disp = 0;
818 struct bio_list bio_list_on_stack;
820 struct blk_plug plug;
822 spin_lock_irq(q->queue_lock);
824 bio_list_init(&bio_list_on_stack);
826 throtl_log(td, "dispatch nr_queued=%u read=%u write=%u",
827 td->nr_queued[READ] + td->nr_queued[WRITE],
828 td->nr_queued[READ], td->nr_queued[WRITE]);
830 nr_disp = throtl_select_dispatch(&td->service_queue, &bio_list_on_stack);
833 throtl_log(td, "bios disp=%u", nr_disp);
835 throtl_schedule_next_dispatch(td);
837 spin_unlock_irq(q->queue_lock);
840 * If we dispatched some requests, unplug the queue to make sure
844 blk_start_plug(&plug);
845 while((bio = bio_list_pop(&bio_list_on_stack)))
846 generic_make_request(bio);
847 blk_finish_plug(&plug);
851 static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
852 struct blkg_policy_data *pd, int off)
854 struct throtl_grp *tg = pd_to_tg(pd);
855 struct blkg_rwstat rwstat = { }, tmp;
858 for_each_possible_cpu(cpu) {
859 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
861 tmp = blkg_rwstat_read((void *)sc + off);
862 for (i = 0; i < BLKG_RWSTAT_NR; i++)
863 rwstat.cnt[i] += tmp.cnt[i];
866 return __blkg_prfill_rwstat(sf, pd, &rwstat);
869 static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft,
872 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
874 blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl,
879 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
882 struct throtl_grp *tg = pd_to_tg(pd);
883 u64 v = *(u64 *)((void *)tg + off);
887 return __blkg_prfill_u64(sf, pd, v);
890 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
893 struct throtl_grp *tg = pd_to_tg(pd);
894 unsigned int v = *(unsigned int *)((void *)tg + off);
898 return __blkg_prfill_u64(sf, pd, v);
901 static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft,
904 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64,
905 &blkcg_policy_throtl, cft->private, false);
909 static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft,
912 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint,
913 &blkcg_policy_throtl, cft->private, false);
917 static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf,
920 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
921 struct blkg_conf_ctx ctx;
922 struct throtl_grp *tg;
923 struct throtl_data *td;
926 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
930 tg = blkg_to_tg(ctx.blkg);
931 td = ctx.blkg->q->td;
937 *(u64 *)((void *)tg + cft->private) = ctx.v;
939 *(unsigned int *)((void *)tg + cft->private) = ctx.v;
941 throtl_log_tg(tg, "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
942 tg->bps[READ], tg->bps[WRITE],
943 tg->iops[READ], tg->iops[WRITE]);
946 * We're already holding queue_lock and know @tg is valid. Let's
947 * apply the new config directly.
949 * Restart the slices for both READ and WRITES. It might happen
950 * that a group's limit are dropped suddenly and we don't want to
951 * account recently dispatched IO with new low rate.
953 throtl_start_new_slice(tg, 0);
954 throtl_start_new_slice(tg, 1);
956 if (tg->flags & THROTL_TG_PENDING) {
957 tg_update_disptime(tg, &td->service_queue);
958 throtl_schedule_next_dispatch(td);
961 blkg_conf_finish(&ctx);
965 static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft,
968 return tg_set_conf(cgrp, cft, buf, true);
971 static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft,
974 return tg_set_conf(cgrp, cft, buf, false);
977 static struct cftype throtl_files[] = {
979 .name = "throttle.read_bps_device",
980 .private = offsetof(struct throtl_grp, bps[READ]),
981 .read_seq_string = tg_print_conf_u64,
982 .write_string = tg_set_conf_u64,
983 .max_write_len = 256,
986 .name = "throttle.write_bps_device",
987 .private = offsetof(struct throtl_grp, bps[WRITE]),
988 .read_seq_string = tg_print_conf_u64,
989 .write_string = tg_set_conf_u64,
990 .max_write_len = 256,
993 .name = "throttle.read_iops_device",
994 .private = offsetof(struct throtl_grp, iops[READ]),
995 .read_seq_string = tg_print_conf_uint,
996 .write_string = tg_set_conf_uint,
997 .max_write_len = 256,
1000 .name = "throttle.write_iops_device",
1001 .private = offsetof(struct throtl_grp, iops[WRITE]),
1002 .read_seq_string = tg_print_conf_uint,
1003 .write_string = tg_set_conf_uint,
1004 .max_write_len = 256,
1007 .name = "throttle.io_service_bytes",
1008 .private = offsetof(struct tg_stats_cpu, service_bytes),
1009 .read_seq_string = tg_print_cpu_rwstat,
1012 .name = "throttle.io_serviced",
1013 .private = offsetof(struct tg_stats_cpu, serviced),
1014 .read_seq_string = tg_print_cpu_rwstat,
1019 static void throtl_shutdown_wq(struct request_queue *q)
1021 struct throtl_data *td = q->td;
1023 cancel_delayed_work_sync(&td->dispatch_work);
1026 static struct blkcg_policy blkcg_policy_throtl = {
1027 .pd_size = sizeof(struct throtl_grp),
1028 .cftypes = throtl_files,
1030 .pd_init_fn = throtl_pd_init,
1031 .pd_exit_fn = throtl_pd_exit,
1032 .pd_reset_stats_fn = throtl_pd_reset_stats,
1035 bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
1037 struct throtl_data *td = q->td;
1038 struct throtl_grp *tg;
1039 bool rw = bio_data_dir(bio), update_disptime = true;
1040 struct blkcg *blkcg;
1041 bool throttled = false;
1043 if (bio->bi_rw & REQ_THROTTLED) {
1044 bio->bi_rw &= ~REQ_THROTTLED;
1049 * A throtl_grp pointer retrieved under rcu can be used to access
1050 * basic fields like stats and io rates. If a group has no rules,
1051 * just update the dispatch stats in lockless manner and return.
1054 blkcg = bio_blkcg(bio);
1055 tg = throtl_lookup_tg(td, blkcg);
1057 if (tg_no_rule_group(tg, rw)) {
1058 throtl_update_dispatch_stats(tg_to_blkg(tg),
1059 bio->bi_size, bio->bi_rw);
1060 goto out_unlock_rcu;
1065 * Either group has not been allocated yet or it is not an unlimited
1068 spin_lock_irq(q->queue_lock);
1069 tg = throtl_lookup_create_tg(td, blkcg);
1073 if (tg->nr_queued[rw]) {
1075 * There is already another bio queued in same dir. No
1076 * need to update dispatch time.
1078 update_disptime = false;
1083 /* Bio is with-in rate limit of group */
1084 if (tg_may_dispatch(tg, bio, NULL)) {
1085 throtl_charge_bio(tg, bio);
1088 * We need to trim slice even when bios are not being queued
1089 * otherwise it might happen that a bio is not queued for
1090 * a long time and slice keeps on extending and trim is not
1091 * called for a long time. Now if limits are reduced suddenly
1092 * we take into account all the IO dispatched so far at new
1093 * low rate and * newly queued IO gets a really long dispatch
1096 * So keep on trimming slice even if bio is not queued.
1098 throtl_trim_slice(tg, rw);
1103 throtl_log_tg(tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
1104 " iodisp=%u iops=%u queued=%d/%d",
1105 rw == READ ? 'R' : 'W',
1106 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1107 tg->io_disp[rw], tg->iops[rw],
1108 tg->nr_queued[READ], tg->nr_queued[WRITE]);
1110 bio_associate_current(bio);
1111 throtl_add_bio_tg(bio, tg, &q->td->service_queue);
1114 if (update_disptime) {
1115 tg_update_disptime(tg, &td->service_queue);
1116 throtl_schedule_next_dispatch(td);
1120 spin_unlock_irq(q->queue_lock);
1128 * blk_throtl_drain - drain throttled bios
1129 * @q: request_queue to drain throttled bios for
1131 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1133 void blk_throtl_drain(struct request_queue *q)
1134 __releases(q->queue_lock) __acquires(q->queue_lock)
1136 struct throtl_data *td = q->td;
1137 struct throtl_service_queue *parent_sq = &td->service_queue;
1138 struct throtl_grp *tg;
1142 queue_lockdep_assert_held(q);
1146 while ((tg = throtl_rb_first(parent_sq))) {
1147 throtl_dequeue_tg(tg, parent_sq);
1149 while ((bio = bio_list_peek(&tg->bio_lists[READ])))
1150 tg_dispatch_one_bio(tg, bio_data_dir(bio), &bl);
1151 while ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
1152 tg_dispatch_one_bio(tg, bio_data_dir(bio), &bl);
1154 spin_unlock_irq(q->queue_lock);
1156 while ((bio = bio_list_pop(&bl)))
1157 generic_make_request(bio);
1159 spin_lock_irq(q->queue_lock);
1162 int blk_throtl_init(struct request_queue *q)
1164 struct throtl_data *td;
1167 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1171 td->service_queue = THROTL_SERVICE_QUEUE_INITIALIZER;
1172 INIT_DELAYED_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1177 /* activate policy */
1178 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1184 void blk_throtl_exit(struct request_queue *q)
1187 throtl_shutdown_wq(q);
1188 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1192 static int __init throtl_init(void)
1194 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1195 if (!kthrotld_workqueue)
1196 panic("Failed to create kthrotld\n");
1198 return blkcg_policy_register(&blkcg_policy_throtl);
1201 module_init(throtl_init);