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 /* A workqueue to queue throttle related work */
25 static struct workqueue_struct *kthrotld_workqueue;
26 static void throtl_schedule_delayed_work(struct throtl_data *td,
29 struct throtl_rb_root {
33 unsigned long min_disptime;
36 #define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \
37 .count = 0, .min_disptime = 0}
39 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
42 /* List of throtl groups on the request queue*/
43 struct hlist_node tg_node;
45 /* active throtl group service_tree member */
46 struct rb_node rb_node;
49 * Dispatch time in jiffies. This is the estimated time when group
50 * will unthrottle and is ready to dispatch more bio. It is used as
51 * key to sort active groups in service tree.
53 unsigned long disptime;
55 struct blkio_group blkg;
59 /* Two lists for READ and WRITE */
60 struct bio_list bio_lists[2];
62 /* Number of queued bios on READ and WRITE lists */
63 unsigned int nr_queued[2];
65 /* bytes per second rate limits */
71 /* Number of bytes disptached in current slice */
72 uint64_t bytes_disp[2];
73 /* Number of bio's dispatched in current slice */
74 unsigned int io_disp[2];
76 /* When did we start a new slice */
77 unsigned long slice_start[2];
78 unsigned long slice_end[2];
80 /* Some throttle limits got updated for the group */
83 struct rcu_head rcu_head;
88 /* List of throtl groups */
89 struct hlist_head tg_list;
91 /* service tree for active throtl groups */
92 struct throtl_rb_root tg_service_tree;
94 struct throtl_grp *root_tg;
95 struct request_queue *queue;
97 /* Total Number of queued bios on READ and WRITE lists */
98 unsigned int nr_queued[2];
101 * number of total undestroyed groups
103 unsigned int nr_undestroyed_grps;
105 /* Work for dispatching throttled bios */
106 struct delayed_work throtl_work;
111 enum tg_state_flags {
112 THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */
115 #define THROTL_TG_FNS(name) \
116 static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \
118 (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \
120 static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \
122 (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \
124 static inline int throtl_tg_##name(const struct throtl_grp *tg) \
126 return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \
129 THROTL_TG_FNS(on_rr);
131 #define throtl_log_tg(td, tg, fmt, args...) \
132 blk_add_trace_msg((td)->queue, "throtl %s " fmt, \
133 blkg_path(&(tg)->blkg), ##args); \
135 #define throtl_log(td, fmt, args...) \
136 blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
138 static inline struct throtl_grp *tg_of_blkg(struct blkio_group *blkg)
141 return container_of(blkg, struct throtl_grp, blkg);
146 static inline unsigned int total_nr_queued(struct throtl_data *td)
148 return td->nr_queued[0] + td->nr_queued[1];
151 static inline struct throtl_grp *throtl_ref_get_tg(struct throtl_grp *tg)
153 atomic_inc(&tg->ref);
157 static void throtl_free_tg(struct rcu_head *head)
159 struct throtl_grp *tg;
161 tg = container_of(head, struct throtl_grp, rcu_head);
162 free_percpu(tg->blkg.stats_cpu);
166 static void throtl_put_tg(struct throtl_grp *tg)
168 BUG_ON(atomic_read(&tg->ref) <= 0);
169 if (!atomic_dec_and_test(&tg->ref))
173 * A group is freed in rcu manner. But having an rcu lock does not
174 * mean that one can access all the fields of blkg and assume these
175 * are valid. For example, don't try to follow throtl_data and
176 * request queue links.
178 * Having a reference to blkg under an rcu allows acess to only
179 * values local to groups like group stats and group rate limits
181 call_rcu(&tg->rcu_head, throtl_free_tg);
184 static struct blkio_group *throtl_alloc_blkio_group(struct request_queue *q,
185 struct blkio_cgroup *blkcg)
187 struct throtl_grp *tg;
189 tg = kzalloc_node(sizeof(*tg), GFP_ATOMIC, q->node);
193 INIT_HLIST_NODE(&tg->tg_node);
194 RB_CLEAR_NODE(&tg->rb_node);
195 bio_list_init(&tg->bio_lists[0]);
196 bio_list_init(&tg->bio_lists[1]);
197 tg->limits_changed = false;
199 tg->bps[READ] = blkcg_get_read_bps(blkcg, tg->blkg.dev);
200 tg->bps[WRITE] = blkcg_get_write_bps(blkcg, tg->blkg.dev);
201 tg->iops[READ] = blkcg_get_read_iops(blkcg, tg->blkg.dev);
202 tg->iops[WRITE] = blkcg_get_write_iops(blkcg, tg->blkg.dev);
205 * Take the initial reference that will be released on destroy
206 * This can be thought of a joint reference by cgroup and
207 * request queue which will be dropped by either request queue
208 * exit or cgroup deletion path depending on who is exiting first.
210 atomic_set(&tg->ref, 1);
216 __throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg)
218 struct backing_dev_info *bdi = &td->queue->backing_dev_info;
219 unsigned int major, minor;
221 if (!tg || tg->blkg.dev)
225 * Fill in device details for a group which might not have been
226 * filled at group creation time as queue was being instantiated
227 * and driver had not attached a device yet
229 if (bdi->dev && dev_name(bdi->dev)) {
230 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
231 tg->blkg.dev = MKDEV(major, minor);
236 * Should be called with without queue lock held. Here queue lock will be
237 * taken rarely. It will be taken only once during life time of a group
241 throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg)
243 if (!tg || tg->blkg.dev)
246 spin_lock_irq(td->queue->queue_lock);
247 __throtl_tg_fill_dev_details(td, tg);
248 spin_unlock_irq(td->queue->queue_lock);
251 static void throtl_link_blkio_group(struct request_queue *q,
252 struct blkio_group *blkg)
254 struct throtl_data *td = q->td;
255 struct throtl_grp *tg = tg_of_blkg(blkg);
257 __throtl_tg_fill_dev_details(td, tg);
259 hlist_add_head(&tg->tg_node, &td->tg_list);
260 td->nr_undestroyed_grps++;
264 throtl_grp *throtl_lookup_tg(struct throtl_data *td, struct blkio_cgroup *blkcg)
266 struct throtl_grp *tg = NULL;
269 * This is the common case when there are no blkio cgroups.
270 * Avoid lookup in this case
272 if (blkcg == &blkio_root_cgroup)
275 tg = tg_of_blkg(blkg_lookup(blkcg, td->queue,
276 BLKIO_POLICY_THROTL));
278 __throtl_tg_fill_dev_details(td, tg);
282 static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
283 struct blkio_cgroup *blkcg)
285 struct request_queue *q = td->queue;
286 struct throtl_grp *tg = NULL;
289 * This is the common case when there are no blkio cgroups.
290 * Avoid lookup in this case
292 if (blkcg == &blkio_root_cgroup) {
295 struct blkio_group *blkg;
297 blkg = blkg_lookup_create(blkcg, q, BLKIO_POLICY_THROTL, false);
299 /* if %NULL and @q is alive, fall back to root_tg */
301 tg = tg_of_blkg(blkg);
302 else if (!blk_queue_dead(q))
306 __throtl_tg_fill_dev_details(td, tg);
310 static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root)
312 /* Service tree is empty */
317 root->left = rb_first(&root->rb);
320 return rb_entry_tg(root->left);
325 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
331 static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root)
335 rb_erase_init(n, &root->rb);
339 static void update_min_dispatch_time(struct throtl_rb_root *st)
341 struct throtl_grp *tg;
343 tg = throtl_rb_first(st);
347 st->min_disptime = tg->disptime;
351 tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg)
353 struct rb_node **node = &st->rb.rb_node;
354 struct rb_node *parent = NULL;
355 struct throtl_grp *__tg;
356 unsigned long key = tg->disptime;
359 while (*node != NULL) {
361 __tg = rb_entry_tg(parent);
363 if (time_before(key, __tg->disptime))
364 node = &parent->rb_left;
366 node = &parent->rb_right;
372 st->left = &tg->rb_node;
374 rb_link_node(&tg->rb_node, parent, node);
375 rb_insert_color(&tg->rb_node, &st->rb);
378 static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
380 struct throtl_rb_root *st = &td->tg_service_tree;
382 tg_service_tree_add(st, tg);
383 throtl_mark_tg_on_rr(tg);
387 static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
389 if (!throtl_tg_on_rr(tg))
390 __throtl_enqueue_tg(td, tg);
393 static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
395 throtl_rb_erase(&tg->rb_node, &td->tg_service_tree);
396 throtl_clear_tg_on_rr(tg);
399 static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
401 if (throtl_tg_on_rr(tg))
402 __throtl_dequeue_tg(td, tg);
405 static void throtl_schedule_next_dispatch(struct throtl_data *td)
407 struct throtl_rb_root *st = &td->tg_service_tree;
410 * If there are more bios pending, schedule more work.
412 if (!total_nr_queued(td))
417 update_min_dispatch_time(st);
419 if (time_before_eq(st->min_disptime, jiffies))
420 throtl_schedule_delayed_work(td, 0);
422 throtl_schedule_delayed_work(td, (st->min_disptime - jiffies));
426 throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
428 tg->bytes_disp[rw] = 0;
430 tg->slice_start[rw] = jiffies;
431 tg->slice_end[rw] = jiffies + throtl_slice;
432 throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
433 rw == READ ? 'R' : 'W', tg->slice_start[rw],
434 tg->slice_end[rw], jiffies);
437 static inline void throtl_set_slice_end(struct throtl_data *td,
438 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
440 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
443 static inline void throtl_extend_slice(struct throtl_data *td,
444 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
446 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
447 throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
448 rw == READ ? 'R' : 'W', tg->slice_start[rw],
449 tg->slice_end[rw], jiffies);
452 /* Determine if previously allocated or extended slice is complete or not */
454 throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
456 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
462 /* Trim the used slices and adjust slice start accordingly */
464 throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
466 unsigned long nr_slices, time_elapsed, io_trim;
469 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
472 * If bps are unlimited (-1), then time slice don't get
473 * renewed. Don't try to trim the slice if slice is used. A new
474 * slice will start when appropriate.
476 if (throtl_slice_used(td, tg, rw))
480 * A bio has been dispatched. Also adjust slice_end. It might happen
481 * that initially cgroup limit was very low resulting in high
482 * slice_end, but later limit was bumped up and bio was dispached
483 * sooner, then we need to reduce slice_end. A high bogus slice_end
484 * is bad because it does not allow new slice to start.
487 throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice);
489 time_elapsed = jiffies - tg->slice_start[rw];
491 nr_slices = time_elapsed / throtl_slice;
495 tmp = tg->bps[rw] * throtl_slice * nr_slices;
499 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
501 if (!bytes_trim && !io_trim)
504 if (tg->bytes_disp[rw] >= bytes_trim)
505 tg->bytes_disp[rw] -= bytes_trim;
507 tg->bytes_disp[rw] = 0;
509 if (tg->io_disp[rw] >= io_trim)
510 tg->io_disp[rw] -= io_trim;
514 tg->slice_start[rw] += nr_slices * throtl_slice;
516 throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
517 " start=%lu end=%lu jiffies=%lu",
518 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
519 tg->slice_start[rw], tg->slice_end[rw], jiffies);
522 static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
523 struct bio *bio, unsigned long *wait)
525 bool rw = bio_data_dir(bio);
526 unsigned int io_allowed;
527 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
530 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
532 /* Slice has just started. Consider one slice interval */
534 jiffy_elapsed_rnd = throtl_slice;
536 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
539 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
540 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
541 * will allow dispatch after 1 second and after that slice should
545 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
549 io_allowed = UINT_MAX;
553 if (tg->io_disp[rw] + 1 <= io_allowed) {
559 /* Calc approx time to dispatch */
560 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
562 if (jiffy_wait > jiffy_elapsed)
563 jiffy_wait = jiffy_wait - jiffy_elapsed;
572 static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
573 struct bio *bio, unsigned long *wait)
575 bool rw = bio_data_dir(bio);
576 u64 bytes_allowed, extra_bytes, tmp;
577 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
579 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
581 /* Slice has just started. Consider one slice interval */
583 jiffy_elapsed_rnd = throtl_slice;
585 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
587 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
591 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
597 /* Calc approx time to dispatch */
598 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
599 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
605 * This wait time is without taking into consideration the rounding
606 * up we did. Add that time also.
608 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
614 static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
615 if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
621 * Returns whether one can dispatch a bio or not. Also returns approx number
622 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
624 static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
625 struct bio *bio, unsigned long *wait)
627 bool rw = bio_data_dir(bio);
628 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
631 * Currently whole state machine of group depends on first bio
632 * queued in the group bio list. So one should not be calling
633 * this function with a different bio if there are other bios
636 BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
638 /* If tg->bps = -1, then BW is unlimited */
639 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
646 * If previous slice expired, start a new one otherwise renew/extend
647 * existing slice to make sure it is at least throtl_slice interval
650 if (throtl_slice_used(td, tg, rw))
651 throtl_start_new_slice(td, tg, rw);
653 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
654 throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
657 if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
658 && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
664 max_wait = max(bps_wait, iops_wait);
669 if (time_before(tg->slice_end[rw], jiffies + max_wait))
670 throtl_extend_slice(td, tg, rw, jiffies + max_wait);
675 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
677 bool rw = bio_data_dir(bio);
678 bool sync = rw_is_sync(bio->bi_rw);
680 /* Charge the bio to the group */
681 tg->bytes_disp[rw] += bio->bi_size;
684 blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync);
687 static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
690 bool rw = bio_data_dir(bio);
692 bio_list_add(&tg->bio_lists[rw], bio);
693 /* Take a bio reference on tg */
694 throtl_ref_get_tg(tg);
697 throtl_enqueue_tg(td, tg);
700 static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
702 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
705 if ((bio = bio_list_peek(&tg->bio_lists[READ])))
706 tg_may_dispatch(td, tg, bio, &read_wait);
708 if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
709 tg_may_dispatch(td, tg, bio, &write_wait);
711 min_wait = min(read_wait, write_wait);
712 disptime = jiffies + min_wait;
714 /* Update dispatch time */
715 throtl_dequeue_tg(td, tg);
716 tg->disptime = disptime;
717 throtl_enqueue_tg(td, tg);
720 static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
721 bool rw, struct bio_list *bl)
725 bio = bio_list_pop(&tg->bio_lists[rw]);
727 /* Drop bio reference on tg */
730 BUG_ON(td->nr_queued[rw] <= 0);
733 throtl_charge_bio(tg, bio);
734 bio_list_add(bl, bio);
735 bio->bi_rw |= REQ_THROTTLED;
737 throtl_trim_slice(td, tg, rw);
740 static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
743 unsigned int nr_reads = 0, nr_writes = 0;
744 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
745 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
748 /* Try to dispatch 75% READS and 25% WRITES */
750 while ((bio = bio_list_peek(&tg->bio_lists[READ]))
751 && tg_may_dispatch(td, tg, bio, NULL)) {
753 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
756 if (nr_reads >= max_nr_reads)
760 while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
761 && tg_may_dispatch(td, tg, bio, NULL)) {
763 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
766 if (nr_writes >= max_nr_writes)
770 return nr_reads + nr_writes;
773 static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
775 unsigned int nr_disp = 0;
776 struct throtl_grp *tg;
777 struct throtl_rb_root *st = &td->tg_service_tree;
780 tg = throtl_rb_first(st);
785 if (time_before(jiffies, tg->disptime))
788 throtl_dequeue_tg(td, tg);
790 nr_disp += throtl_dispatch_tg(td, tg, bl);
792 if (tg->nr_queued[0] || tg->nr_queued[1]) {
793 tg_update_disptime(td, tg);
794 throtl_enqueue_tg(td, tg);
797 if (nr_disp >= throtl_quantum)
804 static void throtl_process_limit_change(struct throtl_data *td)
806 struct throtl_grp *tg;
807 struct hlist_node *pos, *n;
809 if (!td->limits_changed)
812 xchg(&td->limits_changed, false);
814 throtl_log(td, "limits changed");
816 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
817 if (!tg->limits_changed)
820 if (!xchg(&tg->limits_changed, false))
823 throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
824 " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
825 tg->iops[READ], tg->iops[WRITE]);
828 * Restart the slices for both READ and WRITES. It
829 * might happen that a group's limit are dropped
830 * suddenly and we don't want to account recently
831 * dispatched IO with new low rate
833 throtl_start_new_slice(td, tg, 0);
834 throtl_start_new_slice(td, tg, 1);
836 if (throtl_tg_on_rr(tg))
837 tg_update_disptime(td, tg);
841 /* Dispatch throttled bios. Should be called without queue lock held. */
842 static int throtl_dispatch(struct request_queue *q)
844 struct throtl_data *td = q->td;
845 unsigned int nr_disp = 0;
846 struct bio_list bio_list_on_stack;
848 struct blk_plug plug;
850 spin_lock_irq(q->queue_lock);
852 throtl_process_limit_change(td);
854 if (!total_nr_queued(td))
857 bio_list_init(&bio_list_on_stack);
859 throtl_log(td, "dispatch nr_queued=%u read=%u write=%u",
860 total_nr_queued(td), td->nr_queued[READ],
861 td->nr_queued[WRITE]);
863 nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
866 throtl_log(td, "bios disp=%u", nr_disp);
868 throtl_schedule_next_dispatch(td);
870 spin_unlock_irq(q->queue_lock);
873 * If we dispatched some requests, unplug the queue to make sure
877 blk_start_plug(&plug);
878 while((bio = bio_list_pop(&bio_list_on_stack)))
879 generic_make_request(bio);
880 blk_finish_plug(&plug);
885 void blk_throtl_work(struct work_struct *work)
887 struct throtl_data *td = container_of(work, struct throtl_data,
889 struct request_queue *q = td->queue;
894 /* Call with queue lock held */
896 throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay)
899 struct delayed_work *dwork = &td->throtl_work;
901 /* schedule work if limits changed even if no bio is queued */
902 if (total_nr_queued(td) || td->limits_changed) {
904 * We might have a work scheduled to be executed in future.
905 * Cancel that and schedule a new one.
907 __cancel_delayed_work(dwork);
908 queue_delayed_work(kthrotld_workqueue, dwork, delay);
909 throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
915 throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg)
917 /* Something wrong if we are trying to remove same group twice */
918 BUG_ON(hlist_unhashed(&tg->tg_node));
920 hlist_del_init(&tg->tg_node);
923 * Put the reference taken at the time of creation so that when all
924 * queues are gone, group can be destroyed.
927 td->nr_undestroyed_grps--;
930 static bool throtl_release_tgs(struct throtl_data *td, bool release_root)
932 struct hlist_node *pos, *n;
933 struct throtl_grp *tg;
936 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
938 if (!release_root && tg == td->root_tg)
942 * If cgroup removal path got to blk_group first and removed
943 * it from cgroup list, then it will take care of destroying
946 if (!blkiocg_del_blkio_group(&tg->blkg))
947 throtl_destroy_tg(td, tg);
955 * Blk cgroup controller notification saying that blkio_group object is being
956 * delinked as associated cgroup object is going away. That also means that
957 * no new IO will come in this group. So get rid of this group as soon as
958 * any pending IO in the group is finished.
960 * This function is called under rcu_read_lock(). @q is the rcu protected
961 * pointer. That means @q is a valid request_queue pointer as long as we
964 * @q was fetched from blkio_group under blkio_cgroup->lock. That means
965 * it should not be NULL as even if queue was going away, cgroup deltion
966 * path got to it first.
968 void throtl_unlink_blkio_group(struct request_queue *q,
969 struct blkio_group *blkg)
973 spin_lock_irqsave(q->queue_lock, flags);
974 throtl_destroy_tg(q->td, tg_of_blkg(blkg));
975 spin_unlock_irqrestore(q->queue_lock, flags);
978 static bool throtl_clear_queue(struct request_queue *q)
980 lockdep_assert_held(q->queue_lock);
983 * Clear tgs but leave the root one alone. This is necessary
984 * because root_tg is expected to be persistent and safe because
985 * blk-throtl can never be disabled while @q is alive. This is a
986 * kludge to prepare for unified blkg. This whole function will be
989 return throtl_release_tgs(q->td, false);
992 static void throtl_update_blkio_group_common(struct throtl_data *td,
993 struct throtl_grp *tg)
995 xchg(&tg->limits_changed, true);
996 xchg(&td->limits_changed, true);
997 /* Schedule a work now to process the limit change */
998 throtl_schedule_delayed_work(td, 0);
1002 * For all update functions, @q should be a valid pointer because these
1003 * update functions are called under blkcg_lock, that means, blkg is
1004 * valid and in turn @q is valid. queue exit path can not race because
1007 * Can not take queue lock in update functions as queue lock under blkcg_lock
1008 * is not allowed. Under other paths we take blkcg_lock under queue_lock.
1010 static void throtl_update_blkio_group_read_bps(struct request_queue *q,
1011 struct blkio_group *blkg, u64 read_bps)
1013 struct throtl_grp *tg = tg_of_blkg(blkg);
1015 tg->bps[READ] = read_bps;
1016 throtl_update_blkio_group_common(q->td, tg);
1019 static void throtl_update_blkio_group_write_bps(struct request_queue *q,
1020 struct blkio_group *blkg, u64 write_bps)
1022 struct throtl_grp *tg = tg_of_blkg(blkg);
1024 tg->bps[WRITE] = write_bps;
1025 throtl_update_blkio_group_common(q->td, tg);
1028 static void throtl_update_blkio_group_read_iops(struct request_queue *q,
1029 struct blkio_group *blkg, unsigned int read_iops)
1031 struct throtl_grp *tg = tg_of_blkg(blkg);
1033 tg->iops[READ] = read_iops;
1034 throtl_update_blkio_group_common(q->td, tg);
1037 static void throtl_update_blkio_group_write_iops(struct request_queue *q,
1038 struct blkio_group *blkg, unsigned int write_iops)
1040 struct throtl_grp *tg = tg_of_blkg(blkg);
1042 tg->iops[WRITE] = write_iops;
1043 throtl_update_blkio_group_common(q->td, tg);
1046 static void throtl_shutdown_wq(struct request_queue *q)
1048 struct throtl_data *td = q->td;
1050 cancel_delayed_work_sync(&td->throtl_work);
1053 static struct blkio_policy_type blkio_policy_throtl = {
1055 .blkio_alloc_group_fn = throtl_alloc_blkio_group,
1056 .blkio_link_group_fn = throtl_link_blkio_group,
1057 .blkio_unlink_group_fn = throtl_unlink_blkio_group,
1058 .blkio_clear_queue_fn = throtl_clear_queue,
1059 .blkio_update_group_read_bps_fn =
1060 throtl_update_blkio_group_read_bps,
1061 .blkio_update_group_write_bps_fn =
1062 throtl_update_blkio_group_write_bps,
1063 .blkio_update_group_read_iops_fn =
1064 throtl_update_blkio_group_read_iops,
1065 .blkio_update_group_write_iops_fn =
1066 throtl_update_blkio_group_write_iops,
1068 .plid = BLKIO_POLICY_THROTL,
1071 bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
1073 struct throtl_data *td = q->td;
1074 struct throtl_grp *tg;
1075 bool rw = bio_data_dir(bio), update_disptime = true;
1076 struct blkio_cgroup *blkcg;
1077 bool throttled = false;
1079 if (bio->bi_rw & REQ_THROTTLED) {
1080 bio->bi_rw &= ~REQ_THROTTLED;
1085 * A throtl_grp pointer retrieved under rcu can be used to access
1086 * basic fields like stats and io rates. If a group has no rules,
1087 * just update the dispatch stats in lockless manner and return.
1090 blkcg = task_blkio_cgroup(current);
1091 tg = throtl_lookup_tg(td, blkcg);
1093 throtl_tg_fill_dev_details(td, tg);
1095 if (tg_no_rule_group(tg, rw)) {
1096 blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size,
1097 rw, rw_is_sync(bio->bi_rw));
1098 goto out_unlock_rcu;
1103 * Either group has not been allocated yet or it is not an unlimited
1106 spin_lock_irq(q->queue_lock);
1107 tg = throtl_lookup_create_tg(td, blkcg);
1111 if (tg->nr_queued[rw]) {
1113 * There is already another bio queued in same dir. No
1114 * need to update dispatch time.
1116 update_disptime = false;
1121 /* Bio is with-in rate limit of group */
1122 if (tg_may_dispatch(td, tg, bio, NULL)) {
1123 throtl_charge_bio(tg, bio);
1126 * We need to trim slice even when bios are not being queued
1127 * otherwise it might happen that a bio is not queued for
1128 * a long time and slice keeps on extending and trim is not
1129 * called for a long time. Now if limits are reduced suddenly
1130 * we take into account all the IO dispatched so far at new
1131 * low rate and * newly queued IO gets a really long dispatch
1134 * So keep on trimming slice even if bio is not queued.
1136 throtl_trim_slice(td, tg, rw);
1141 throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
1142 " iodisp=%u iops=%u queued=%d/%d",
1143 rw == READ ? 'R' : 'W',
1144 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1145 tg->io_disp[rw], tg->iops[rw],
1146 tg->nr_queued[READ], tg->nr_queued[WRITE]);
1148 throtl_add_bio_tg(q->td, tg, bio);
1151 if (update_disptime) {
1152 tg_update_disptime(td, tg);
1153 throtl_schedule_next_dispatch(td);
1157 spin_unlock_irq(q->queue_lock);
1165 * blk_throtl_drain - drain throttled bios
1166 * @q: request_queue to drain throttled bios for
1168 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1170 void blk_throtl_drain(struct request_queue *q)
1171 __releases(q->queue_lock) __acquires(q->queue_lock)
1173 struct throtl_data *td = q->td;
1174 struct throtl_rb_root *st = &td->tg_service_tree;
1175 struct throtl_grp *tg;
1179 WARN_ON_ONCE(!queue_is_locked(q));
1183 while ((tg = throtl_rb_first(st))) {
1184 throtl_dequeue_tg(td, tg);
1186 while ((bio = bio_list_peek(&tg->bio_lists[READ])))
1187 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl);
1188 while ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
1189 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl);
1191 spin_unlock_irq(q->queue_lock);
1193 while ((bio = bio_list_pop(&bl)))
1194 generic_make_request(bio);
1196 spin_lock_irq(q->queue_lock);
1199 int blk_throtl_init(struct request_queue *q)
1201 struct throtl_data *td;
1202 struct blkio_group *blkg;
1204 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1208 INIT_HLIST_HEAD(&td->tg_list);
1209 td->tg_service_tree = THROTL_RB_ROOT;
1210 td->limits_changed = false;
1211 INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);
1216 /* alloc and init root group. */
1218 spin_lock_irq(q->queue_lock);
1220 blkg = blkg_lookup_create(&blkio_root_cgroup, q, BLKIO_POLICY_THROTL,
1223 td->root_tg = tg_of_blkg(blkg);
1225 spin_unlock_irq(q->queue_lock);
1235 void blk_throtl_exit(struct request_queue *q)
1237 struct throtl_data *td = q->td;
1242 throtl_shutdown_wq(q);
1244 spin_lock_irq(q->queue_lock);
1245 throtl_release_tgs(td, true);
1247 /* If there are other groups */
1248 if (td->nr_undestroyed_grps > 0)
1251 spin_unlock_irq(q->queue_lock);
1254 * Wait for tg->blkg->q accessors to exit their grace periods.
1255 * Do this wait only if there are other undestroyed groups out
1256 * there (other than root group). This can happen if cgroup deletion
1257 * path claimed the responsibility of cleaning up a group before
1258 * queue cleanup code get to the group.
1260 * Do not call synchronize_rcu() unconditionally as there are drivers
1261 * which create/delete request queue hundreds of times during scan/boot
1262 * and synchronize_rcu() can take significant time and slow down boot.
1268 * Just being safe to make sure after previous flush if some body did
1269 * update limits through cgroup and another work got queued, cancel
1272 throtl_shutdown_wq(q);
1275 void blk_throtl_release(struct request_queue *q)
1280 static int __init throtl_init(void)
1282 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1283 if (!kthrotld_workqueue)
1284 panic("Failed to create kthrotld\n");
1286 blkio_policy_register(&blkio_policy_throtl);
1290 module_init(throtl_init);