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 throtl_service_queue *parent_sq; /* the parent service_queue */
33 * Bios queued directly to this service_queue or dispatched from
34 * children throtl_grp's.
36 struct bio_list bio_lists[2]; /* queued bios [READ/WRITE] */
37 unsigned int nr_queued[2]; /* number of queued bios */
40 * RB tree of active children throtl_grp's, which are sorted by
43 struct rb_root pending_tree; /* RB tree of active tgs */
44 struct rb_node *first_pending; /* first node in the tree */
45 unsigned int nr_pending; /* # queued in the tree */
46 unsigned long first_pending_disptime; /* disptime of the first tg */
47 struct timer_list pending_timer; /* fires on first_pending_disptime */
51 THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */
52 THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */
55 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
57 /* Per-cpu group stats */
59 /* total bytes transferred */
60 struct blkg_rwstat service_bytes;
61 /* total IOs serviced, post merge */
62 struct blkg_rwstat serviced;
66 /* must be the first member */
67 struct blkg_policy_data pd;
69 /* active throtl group service_queue member */
70 struct rb_node rb_node;
72 /* throtl_data this group belongs to */
73 struct throtl_data *td;
75 /* this group's service queue */
76 struct throtl_service_queue service_queue;
79 * Dispatch time in jiffies. This is the estimated time when group
80 * will unthrottle and is ready to dispatch more bio. It is used as
81 * key to sort active groups in service tree.
83 unsigned long disptime;
87 /* bytes per second rate limits */
93 /* Number of bytes disptached in current slice */
94 uint64_t bytes_disp[2];
95 /* Number of bio's dispatched in current slice */
96 unsigned int io_disp[2];
98 /* When did we start a new slice */
99 unsigned long slice_start[2];
100 unsigned long slice_end[2];
102 /* Per cpu stats pointer */
103 struct tg_stats_cpu __percpu *stats_cpu;
105 /* List of tgs waiting for per cpu stats memory to be allocated */
106 struct list_head stats_alloc_node;
111 /* service tree for active throtl groups */
112 struct throtl_service_queue service_queue;
114 struct request_queue *queue;
116 /* Total Number of queued bios on READ and WRITE lists */
117 unsigned int nr_queued[2];
120 * number of total undestroyed groups
122 unsigned int nr_undestroyed_grps;
124 /* Work for dispatching throttled bios */
125 struct work_struct dispatch_work;
128 /* list and work item to allocate percpu group stats */
129 static DEFINE_SPINLOCK(tg_stats_alloc_lock);
130 static LIST_HEAD(tg_stats_alloc_list);
132 static void tg_stats_alloc_fn(struct work_struct *);
133 static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn);
135 static void throtl_pending_timer_fn(unsigned long arg);
137 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
139 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
142 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
144 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
147 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
149 return pd_to_blkg(&tg->pd);
152 static inline struct throtl_grp *td_root_tg(struct throtl_data *td)
154 return blkg_to_tg(td->queue->root_blkg);
158 * sq_to_tg - return the throl_grp the specified service queue belongs to
159 * @sq: the throtl_service_queue of interest
161 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
162 * embedded in throtl_data, %NULL is returned.
164 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
166 if (sq && sq->parent_sq)
167 return container_of(sq, struct throtl_grp, service_queue);
173 * sq_to_td - return throtl_data the specified service queue belongs to
174 * @sq: the throtl_service_queue of interest
176 * A service_queue can be embeded in either a throtl_grp or throtl_data.
177 * Determine the associated throtl_data accordingly and return it.
179 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
181 struct throtl_grp *tg = sq_to_tg(sq);
186 return container_of(sq, struct throtl_data, service_queue);
190 * throtl_log - log debug message via blktrace
191 * @sq: the service_queue being reported
192 * @fmt: printf format string
195 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
196 * throtl_grp; otherwise, just "throtl".
198 * TODO: this should be made a function and name formatting should happen
199 * after testing whether blktrace is enabled.
201 #define throtl_log(sq, fmt, args...) do { \
202 struct throtl_grp *__tg = sq_to_tg((sq)); \
203 struct throtl_data *__td = sq_to_td((sq)); \
209 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
210 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
212 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
217 * Worker for allocating per cpu stat for tgs. This is scheduled on the
218 * system_wq once there are some groups on the alloc_list waiting for
221 static void tg_stats_alloc_fn(struct work_struct *work)
223 static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */
224 struct delayed_work *dwork = to_delayed_work(work);
229 stats_cpu = alloc_percpu(struct tg_stats_cpu);
231 /* allocation failed, try again after some time */
232 schedule_delayed_work(dwork, msecs_to_jiffies(10));
237 spin_lock_irq(&tg_stats_alloc_lock);
239 if (!list_empty(&tg_stats_alloc_list)) {
240 struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list,
243 swap(tg->stats_cpu, stats_cpu);
244 list_del_init(&tg->stats_alloc_node);
247 empty = list_empty(&tg_stats_alloc_list);
248 spin_unlock_irq(&tg_stats_alloc_lock);
253 /* init a service_queue, assumes the caller zeroed it */
254 static void throtl_service_queue_init(struct throtl_service_queue *sq,
255 struct throtl_service_queue *parent_sq)
257 bio_list_init(&sq->bio_lists[0]);
258 bio_list_init(&sq->bio_lists[1]);
259 sq->pending_tree = RB_ROOT;
260 sq->parent_sq = parent_sq;
261 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
265 static void throtl_service_queue_exit(struct throtl_service_queue *sq)
267 del_timer_sync(&sq->pending_timer);
270 static void throtl_pd_init(struct blkcg_gq *blkg)
272 struct throtl_grp *tg = blkg_to_tg(blkg);
273 struct throtl_data *td = blkg->q->td;
276 throtl_service_queue_init(&tg->service_queue, &td->service_queue);
277 RB_CLEAR_NODE(&tg->rb_node);
283 tg->iops[WRITE] = -1;
286 * Ugh... We need to perform per-cpu allocation for tg->stats_cpu
287 * but percpu allocator can't be called from IO path. Queue tg on
288 * tg_stats_alloc_list and allocate from work item.
290 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
291 list_add(&tg->stats_alloc_node, &tg_stats_alloc_list);
292 schedule_delayed_work(&tg_stats_alloc_work, 0);
293 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
296 static void throtl_pd_exit(struct blkcg_gq *blkg)
298 struct throtl_grp *tg = blkg_to_tg(blkg);
301 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
302 list_del_init(&tg->stats_alloc_node);
303 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
305 free_percpu(tg->stats_cpu);
307 throtl_service_queue_exit(&tg->service_queue);
310 static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
312 struct throtl_grp *tg = blkg_to_tg(blkg);
315 if (tg->stats_cpu == NULL)
318 for_each_possible_cpu(cpu) {
319 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
321 blkg_rwstat_reset(&sc->service_bytes);
322 blkg_rwstat_reset(&sc->serviced);
326 static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
330 * This is the common case when there are no blkcgs. Avoid lookup
333 if (blkcg == &blkcg_root)
334 return td_root_tg(td);
336 return blkg_to_tg(blkg_lookup(blkcg, td->queue));
339 static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
342 struct request_queue *q = td->queue;
343 struct throtl_grp *tg = NULL;
346 * This is the common case when there are no blkcgs. Avoid lookup
349 if (blkcg == &blkcg_root) {
352 struct blkcg_gq *blkg;
354 blkg = blkg_lookup_create(blkcg, q);
356 /* if %NULL and @q is alive, fall back to root_tg */
358 tg = blkg_to_tg(blkg);
359 else if (!blk_queue_dying(q))
366 static struct throtl_grp *
367 throtl_rb_first(struct throtl_service_queue *parent_sq)
369 /* Service tree is empty */
370 if (!parent_sq->nr_pending)
373 if (!parent_sq->first_pending)
374 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
376 if (parent_sq->first_pending)
377 return rb_entry_tg(parent_sq->first_pending);
382 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
388 static void throtl_rb_erase(struct rb_node *n,
389 struct throtl_service_queue *parent_sq)
391 if (parent_sq->first_pending == n)
392 parent_sq->first_pending = NULL;
393 rb_erase_init(n, &parent_sq->pending_tree);
394 --parent_sq->nr_pending;
397 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
399 struct throtl_grp *tg;
401 tg = throtl_rb_first(parent_sq);
405 parent_sq->first_pending_disptime = tg->disptime;
408 static void tg_service_queue_add(struct throtl_grp *tg)
410 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
411 struct rb_node **node = &parent_sq->pending_tree.rb_node;
412 struct rb_node *parent = NULL;
413 struct throtl_grp *__tg;
414 unsigned long key = tg->disptime;
417 while (*node != NULL) {
419 __tg = rb_entry_tg(parent);
421 if (time_before(key, __tg->disptime))
422 node = &parent->rb_left;
424 node = &parent->rb_right;
430 parent_sq->first_pending = &tg->rb_node;
432 rb_link_node(&tg->rb_node, parent, node);
433 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
436 static void __throtl_enqueue_tg(struct throtl_grp *tg)
438 tg_service_queue_add(tg);
439 tg->flags |= THROTL_TG_PENDING;
440 tg->service_queue.parent_sq->nr_pending++;
443 static void throtl_enqueue_tg(struct throtl_grp *tg)
445 if (!(tg->flags & THROTL_TG_PENDING))
446 __throtl_enqueue_tg(tg);
449 static void __throtl_dequeue_tg(struct throtl_grp *tg)
451 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
452 tg->flags &= ~THROTL_TG_PENDING;
455 static void throtl_dequeue_tg(struct throtl_grp *tg)
457 if (tg->flags & THROTL_TG_PENDING)
458 __throtl_dequeue_tg(tg);
461 /* Call with queue lock held */
462 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
463 unsigned long expires)
465 mod_timer(&sq->pending_timer, expires);
466 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
467 expires - jiffies, jiffies);
470 static void throtl_schedule_next_dispatch(struct throtl_service_queue *sq)
472 struct throtl_data *td = sq_to_td(sq);
474 /* any pending children left? */
478 update_min_dispatch_time(sq);
480 /* is the next dispatch time in the future? */
481 if (time_after(sq->first_pending_disptime, jiffies)) {
482 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
486 /* kick immediate execution */
487 queue_work(kthrotld_workqueue, &td->dispatch_work);
490 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
492 tg->bytes_disp[rw] = 0;
494 tg->slice_start[rw] = jiffies;
495 tg->slice_end[rw] = jiffies + throtl_slice;
496 throtl_log(&tg->service_queue,
497 "[%c] new slice start=%lu end=%lu jiffies=%lu",
498 rw == READ ? 'R' : 'W', tg->slice_start[rw],
499 tg->slice_end[rw], jiffies);
502 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
503 unsigned long jiffy_end)
505 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
508 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
509 unsigned long jiffy_end)
511 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
512 throtl_log(&tg->service_queue,
513 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
514 rw == READ ? 'R' : 'W', tg->slice_start[rw],
515 tg->slice_end[rw], jiffies);
518 /* Determine if previously allocated or extended slice is complete or not */
519 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
521 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
527 /* Trim the used slices and adjust slice start accordingly */
528 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
530 unsigned long nr_slices, time_elapsed, io_trim;
533 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
536 * If bps are unlimited (-1), then time slice don't get
537 * renewed. Don't try to trim the slice if slice is used. A new
538 * slice will start when appropriate.
540 if (throtl_slice_used(tg, rw))
544 * A bio has been dispatched. Also adjust slice_end. It might happen
545 * that initially cgroup limit was very low resulting in high
546 * slice_end, but later limit was bumped up and bio was dispached
547 * sooner, then we need to reduce slice_end. A high bogus slice_end
548 * is bad because it does not allow new slice to start.
551 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
553 time_elapsed = jiffies - tg->slice_start[rw];
555 nr_slices = time_elapsed / throtl_slice;
559 tmp = tg->bps[rw] * throtl_slice * nr_slices;
563 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
565 if (!bytes_trim && !io_trim)
568 if (tg->bytes_disp[rw] >= bytes_trim)
569 tg->bytes_disp[rw] -= bytes_trim;
571 tg->bytes_disp[rw] = 0;
573 if (tg->io_disp[rw] >= io_trim)
574 tg->io_disp[rw] -= io_trim;
578 tg->slice_start[rw] += nr_slices * throtl_slice;
580 throtl_log(&tg->service_queue,
581 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
582 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
583 tg->slice_start[rw], tg->slice_end[rw], jiffies);
586 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
589 bool rw = bio_data_dir(bio);
590 unsigned int io_allowed;
591 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
594 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
596 /* Slice has just started. Consider one slice interval */
598 jiffy_elapsed_rnd = throtl_slice;
600 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
603 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
604 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
605 * will allow dispatch after 1 second and after that slice should
609 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
613 io_allowed = UINT_MAX;
617 if (tg->io_disp[rw] + 1 <= io_allowed) {
623 /* Calc approx time to dispatch */
624 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
626 if (jiffy_wait > jiffy_elapsed)
627 jiffy_wait = jiffy_wait - jiffy_elapsed;
636 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
639 bool rw = bio_data_dir(bio);
640 u64 bytes_allowed, extra_bytes, tmp;
641 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
643 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
645 /* Slice has just started. Consider one slice interval */
647 jiffy_elapsed_rnd = throtl_slice;
649 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
651 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
655 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
661 /* Calc approx time to dispatch */
662 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
663 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
669 * This wait time is without taking into consideration the rounding
670 * up we did. Add that time also.
672 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
678 static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
679 if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
685 * Returns whether one can dispatch a bio or not. Also returns approx number
686 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
688 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
691 bool rw = bio_data_dir(bio);
692 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
695 * Currently whole state machine of group depends on first bio
696 * queued in the group bio list. So one should not be calling
697 * this function with a different bio if there are other bios
700 BUG_ON(tg->service_queue.nr_queued[rw] &&
701 bio != bio_list_peek(&tg->service_queue.bio_lists[rw]));
703 /* If tg->bps = -1, then BW is unlimited */
704 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
711 * If previous slice expired, start a new one otherwise renew/extend
712 * existing slice to make sure it is at least throtl_slice interval
715 if (throtl_slice_used(tg, rw))
716 throtl_start_new_slice(tg, rw);
718 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
719 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
722 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
723 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
729 max_wait = max(bps_wait, iops_wait);
734 if (time_before(tg->slice_end[rw], jiffies + max_wait))
735 throtl_extend_slice(tg, rw, jiffies + max_wait);
740 static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
743 struct throtl_grp *tg = blkg_to_tg(blkg);
744 struct tg_stats_cpu *stats_cpu;
747 /* If per cpu stats are not allocated yet, don't do any accounting. */
748 if (tg->stats_cpu == NULL)
752 * Disabling interrupts to provide mutual exclusion between two
753 * writes on same cpu. It probably is not needed for 64bit. Not
754 * optimizing that case yet.
756 local_irq_save(flags);
758 stats_cpu = this_cpu_ptr(tg->stats_cpu);
760 blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
761 blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
763 local_irq_restore(flags);
766 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
768 bool rw = bio_data_dir(bio);
770 /* Charge the bio to the group */
771 tg->bytes_disp[rw] += bio->bi_size;
775 * REQ_THROTTLED is used to prevent the same bio to be throttled
776 * more than once as a throttled bio will go through blk-throtl the
777 * second time when it eventually gets issued. Set it when a bio
778 * is being charged to a tg.
780 * Dispatch stats aren't recursive and each @bio should only be
781 * accounted by the @tg it was originally associated with. Let's
782 * update the stats when setting REQ_THROTTLED for the first time
783 * which is guaranteed to be for the @bio's original tg.
785 if (!(bio->bi_rw & REQ_THROTTLED)) {
786 bio->bi_rw |= REQ_THROTTLED;
787 throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size,
792 static void throtl_add_bio_tg(struct bio *bio, struct throtl_grp *tg)
794 struct throtl_service_queue *sq = &tg->service_queue;
795 bool rw = bio_data_dir(bio);
798 * If @tg doesn't currently have any bios queued in the same
799 * direction, queueing @bio can change when @tg should be
800 * dispatched. Mark that @tg was empty. This is automatically
801 * cleaered on the next tg_update_disptime().
803 if (!sq->nr_queued[rw])
804 tg->flags |= THROTL_TG_WAS_EMPTY;
806 bio_list_add(&sq->bio_lists[rw], bio);
807 /* Take a bio reference on tg */
808 blkg_get(tg_to_blkg(tg));
810 tg->td->nr_queued[rw]++;
811 throtl_enqueue_tg(tg);
814 static void tg_update_disptime(struct throtl_grp *tg)
816 struct throtl_service_queue *sq = &tg->service_queue;
817 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
820 if ((bio = bio_list_peek(&sq->bio_lists[READ])))
821 tg_may_dispatch(tg, bio, &read_wait);
823 if ((bio = bio_list_peek(&sq->bio_lists[WRITE])))
824 tg_may_dispatch(tg, bio, &write_wait);
826 min_wait = min(read_wait, write_wait);
827 disptime = jiffies + min_wait;
829 /* Update dispatch time */
830 throtl_dequeue_tg(tg);
831 tg->disptime = disptime;
832 throtl_enqueue_tg(tg);
834 /* see throtl_add_bio_tg() */
835 tg->flags &= ~THROTL_TG_WAS_EMPTY;
838 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
840 struct throtl_service_queue *sq = &tg->service_queue;
843 bio = bio_list_pop(&sq->bio_lists[rw]);
845 /* Drop bio reference on blkg */
846 blkg_put(tg_to_blkg(tg));
848 BUG_ON(tg->td->nr_queued[rw] <= 0);
849 tg->td->nr_queued[rw]--;
851 throtl_charge_bio(tg, bio);
852 bio_list_add(&sq->parent_sq->bio_lists[rw], bio);
854 throtl_trim_slice(tg, rw);
857 static int throtl_dispatch_tg(struct throtl_grp *tg)
859 struct throtl_service_queue *sq = &tg->service_queue;
860 unsigned int nr_reads = 0, nr_writes = 0;
861 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
862 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
865 /* Try to dispatch 75% READS and 25% WRITES */
867 while ((bio = bio_list_peek(&sq->bio_lists[READ])) &&
868 tg_may_dispatch(tg, bio, NULL)) {
870 tg_dispatch_one_bio(tg, bio_data_dir(bio));
873 if (nr_reads >= max_nr_reads)
877 while ((bio = bio_list_peek(&sq->bio_lists[WRITE])) &&
878 tg_may_dispatch(tg, bio, NULL)) {
880 tg_dispatch_one_bio(tg, bio_data_dir(bio));
883 if (nr_writes >= max_nr_writes)
887 return nr_reads + nr_writes;
890 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
892 unsigned int nr_disp = 0;
895 struct throtl_grp *tg = throtl_rb_first(parent_sq);
896 struct throtl_service_queue *sq = &tg->service_queue;
901 if (time_before(jiffies, tg->disptime))
904 throtl_dequeue_tg(tg);
906 nr_disp += throtl_dispatch_tg(tg);
908 if (sq->nr_queued[0] || sq->nr_queued[1])
909 tg_update_disptime(tg);
911 if (nr_disp >= throtl_quantum)
918 static void throtl_pending_timer_fn(unsigned long arg)
920 struct throtl_service_queue *sq = (void *)arg;
921 struct throtl_data *td = sq_to_td(sq);
923 queue_work(kthrotld_workqueue, &td->dispatch_work);
926 /* work function to dispatch throttled bios */
927 void blk_throtl_dispatch_work_fn(struct work_struct *work)
929 struct throtl_data *td = container_of(work, struct throtl_data,
931 struct throtl_service_queue *sq = &td->service_queue;
932 struct request_queue *q = td->queue;
933 unsigned int nr_disp = 0;
934 struct bio_list bio_list_on_stack;
936 struct blk_plug plug;
939 spin_lock_irq(q->queue_lock);
941 bio_list_init(&bio_list_on_stack);
943 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
944 td->nr_queued[READ] + td->nr_queued[WRITE],
945 td->nr_queued[READ], td->nr_queued[WRITE]);
947 nr_disp = throtl_select_dispatch(sq);
950 for (rw = READ; rw <= WRITE; rw++) {
951 bio_list_merge(&bio_list_on_stack, &sq->bio_lists[rw]);
952 bio_list_init(&sq->bio_lists[rw]);
954 throtl_log(sq, "bios disp=%u", nr_disp);
957 throtl_schedule_next_dispatch(sq);
959 spin_unlock_irq(q->queue_lock);
962 * If we dispatched some requests, unplug the queue to make sure
966 blk_start_plug(&plug);
967 while((bio = bio_list_pop(&bio_list_on_stack)))
968 generic_make_request(bio);
969 blk_finish_plug(&plug);
973 static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
974 struct blkg_policy_data *pd, int off)
976 struct throtl_grp *tg = pd_to_tg(pd);
977 struct blkg_rwstat rwstat = { }, tmp;
980 for_each_possible_cpu(cpu) {
981 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
983 tmp = blkg_rwstat_read((void *)sc + off);
984 for (i = 0; i < BLKG_RWSTAT_NR; i++)
985 rwstat.cnt[i] += tmp.cnt[i];
988 return __blkg_prfill_rwstat(sf, pd, &rwstat);
991 static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft,
994 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
996 blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl,
1001 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1004 struct throtl_grp *tg = pd_to_tg(pd);
1005 u64 v = *(u64 *)((void *)tg + off);
1009 return __blkg_prfill_u64(sf, pd, v);
1012 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1015 struct throtl_grp *tg = pd_to_tg(pd);
1016 unsigned int v = *(unsigned int *)((void *)tg + off);
1020 return __blkg_prfill_u64(sf, pd, v);
1023 static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft,
1024 struct seq_file *sf)
1026 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64,
1027 &blkcg_policy_throtl, cft->private, false);
1031 static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1032 struct seq_file *sf)
1034 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint,
1035 &blkcg_policy_throtl, cft->private, false);
1039 static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf,
1042 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1043 struct blkg_conf_ctx ctx;
1044 struct throtl_grp *tg;
1045 struct throtl_service_queue *sq;
1048 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1052 tg = blkg_to_tg(ctx.blkg);
1053 sq = &tg->service_queue;
1059 *(u64 *)((void *)tg + cft->private) = ctx.v;
1061 *(unsigned int *)((void *)tg + cft->private) = ctx.v;
1063 throtl_log(&tg->service_queue,
1064 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1065 tg->bps[READ], tg->bps[WRITE],
1066 tg->iops[READ], tg->iops[WRITE]);
1069 * We're already holding queue_lock and know @tg is valid. Let's
1070 * apply the new config directly.
1072 * Restart the slices for both READ and WRITES. It might happen
1073 * that a group's limit are dropped suddenly and we don't want to
1074 * account recently dispatched IO with new low rate.
1076 throtl_start_new_slice(tg, 0);
1077 throtl_start_new_slice(tg, 1);
1079 if (tg->flags & THROTL_TG_PENDING) {
1080 tg_update_disptime(tg);
1081 throtl_schedule_next_dispatch(sq->parent_sq);
1084 blkg_conf_finish(&ctx);
1088 static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft,
1091 return tg_set_conf(cgrp, cft, buf, true);
1094 static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1097 return tg_set_conf(cgrp, cft, buf, false);
1100 static struct cftype throtl_files[] = {
1102 .name = "throttle.read_bps_device",
1103 .private = offsetof(struct throtl_grp, bps[READ]),
1104 .read_seq_string = tg_print_conf_u64,
1105 .write_string = tg_set_conf_u64,
1106 .max_write_len = 256,
1109 .name = "throttle.write_bps_device",
1110 .private = offsetof(struct throtl_grp, bps[WRITE]),
1111 .read_seq_string = tg_print_conf_u64,
1112 .write_string = tg_set_conf_u64,
1113 .max_write_len = 256,
1116 .name = "throttle.read_iops_device",
1117 .private = offsetof(struct throtl_grp, iops[READ]),
1118 .read_seq_string = tg_print_conf_uint,
1119 .write_string = tg_set_conf_uint,
1120 .max_write_len = 256,
1123 .name = "throttle.write_iops_device",
1124 .private = offsetof(struct throtl_grp, iops[WRITE]),
1125 .read_seq_string = tg_print_conf_uint,
1126 .write_string = tg_set_conf_uint,
1127 .max_write_len = 256,
1130 .name = "throttle.io_service_bytes",
1131 .private = offsetof(struct tg_stats_cpu, service_bytes),
1132 .read_seq_string = tg_print_cpu_rwstat,
1135 .name = "throttle.io_serviced",
1136 .private = offsetof(struct tg_stats_cpu, serviced),
1137 .read_seq_string = tg_print_cpu_rwstat,
1142 static void throtl_shutdown_wq(struct request_queue *q)
1144 struct throtl_data *td = q->td;
1146 cancel_work_sync(&td->dispatch_work);
1149 static struct blkcg_policy blkcg_policy_throtl = {
1150 .pd_size = sizeof(struct throtl_grp),
1151 .cftypes = throtl_files,
1153 .pd_init_fn = throtl_pd_init,
1154 .pd_exit_fn = throtl_pd_exit,
1155 .pd_reset_stats_fn = throtl_pd_reset_stats,
1158 bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
1160 struct throtl_data *td = q->td;
1161 struct throtl_grp *tg;
1162 struct throtl_service_queue *sq;
1163 bool rw = bio_data_dir(bio);
1164 struct blkcg *blkcg;
1165 bool throttled = false;
1167 /* see throtl_charge_bio() */
1168 if (bio->bi_rw & REQ_THROTTLED)
1172 * A throtl_grp pointer retrieved under rcu can be used to access
1173 * basic fields like stats and io rates. If a group has no rules,
1174 * just update the dispatch stats in lockless manner and return.
1177 blkcg = bio_blkcg(bio);
1178 tg = throtl_lookup_tg(td, blkcg);
1180 if (tg_no_rule_group(tg, rw)) {
1181 throtl_update_dispatch_stats(tg_to_blkg(tg),
1182 bio->bi_size, bio->bi_rw);
1183 goto out_unlock_rcu;
1188 * Either group has not been allocated yet or it is not an unlimited
1191 spin_lock_irq(q->queue_lock);
1192 tg = throtl_lookup_create_tg(td, blkcg);
1196 sq = &tg->service_queue;
1198 /* throtl is FIFO - if other bios are already queued, should queue */
1199 if (sq->nr_queued[rw])
1202 /* Bio is with-in rate limit of group */
1203 if (tg_may_dispatch(tg, bio, NULL)) {
1204 throtl_charge_bio(tg, bio);
1207 * We need to trim slice even when bios are not being queued
1208 * otherwise it might happen that a bio is not queued for
1209 * a long time and slice keeps on extending and trim is not
1210 * called for a long time. Now if limits are reduced suddenly
1211 * we take into account all the IO dispatched so far at new
1212 * low rate and * newly queued IO gets a really long dispatch
1215 * So keep on trimming slice even if bio is not queued.
1217 throtl_trim_slice(tg, rw);
1222 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1223 rw == READ ? 'R' : 'W',
1224 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1225 tg->io_disp[rw], tg->iops[rw],
1226 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1228 bio_associate_current(bio);
1229 throtl_add_bio_tg(bio, tg);
1232 /* update @tg's dispatch time if @tg was empty before @bio */
1233 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1234 tg_update_disptime(tg);
1235 throtl_schedule_next_dispatch(tg->service_queue.parent_sq);
1239 spin_unlock_irq(q->queue_lock);
1244 * As multiple blk-throtls may stack in the same issue path, we
1245 * don't want bios to leave with the flag set. Clear the flag if
1249 bio->bi_rw &= ~REQ_THROTTLED;
1254 * blk_throtl_drain - drain throttled bios
1255 * @q: request_queue to drain throttled bios for
1257 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1259 void blk_throtl_drain(struct request_queue *q)
1260 __releases(q->queue_lock) __acquires(q->queue_lock)
1262 struct throtl_data *td = q->td;
1263 struct throtl_service_queue *parent_sq = &td->service_queue;
1264 struct throtl_grp *tg;
1268 queue_lockdep_assert_held(q);
1270 while ((tg = throtl_rb_first(parent_sq))) {
1271 struct throtl_service_queue *sq = &tg->service_queue;
1273 throtl_dequeue_tg(tg);
1275 while ((bio = bio_list_peek(&sq->bio_lists[READ])))
1276 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1277 while ((bio = bio_list_peek(&sq->bio_lists[WRITE])))
1278 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1280 spin_unlock_irq(q->queue_lock);
1282 for (rw = READ; rw <= WRITE; rw++)
1283 while ((bio = bio_list_pop(&parent_sq->bio_lists[rw])))
1284 generic_make_request(bio);
1286 spin_lock_irq(q->queue_lock);
1289 int blk_throtl_init(struct request_queue *q)
1291 struct throtl_data *td;
1294 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1298 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1299 throtl_service_queue_init(&td->service_queue, NULL);
1304 /* activate policy */
1305 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1311 void blk_throtl_exit(struct request_queue *q)
1314 throtl_shutdown_wq(q);
1315 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1319 static int __init throtl_init(void)
1321 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1322 if (!kthrotld_workqueue)
1323 panic("Failed to create kthrotld\n");
1325 return blkcg_policy_register(&blkcg_policy_throtl);
1328 module_init(throtl_init);