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"
14 /* Max dispatch from a group in 1 round */
15 static int throtl_grp_quantum = 8;
17 /* Total max dispatch from all groups in one round */
18 static int throtl_quantum = 32;
20 /* Throttling is performed over 100ms slice and after that slice is renewed */
21 static unsigned long throtl_slice = HZ/10; /* 100 ms */
23 /* A workqueue to queue throttle related work */
24 static struct workqueue_struct *kthrotld_workqueue;
25 static void throtl_schedule_delayed_work(struct throtl_data *td,
28 struct throtl_rb_root {
32 unsigned long min_disptime;
35 #define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \
36 .count = 0, .min_disptime = 0}
38 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
41 /* List of throtl groups on the request queue*/
42 struct hlist_node tg_node;
44 /* active throtl group service_tree member */
45 struct rb_node rb_node;
48 * Dispatch time in jiffies. This is the estimated time when group
49 * will unthrottle and is ready to dispatch more bio. It is used as
50 * key to sort active groups in service tree.
52 unsigned long disptime;
54 struct blkio_group blkg;
58 /* Two lists for READ and WRITE */
59 struct bio_list bio_lists[2];
61 /* Number of queued bios on READ and WRITE lists */
62 unsigned int nr_queued[2];
64 /* bytes per second rate limits */
70 /* Number of bytes disptached in current slice */
71 uint64_t bytes_disp[2];
72 /* Number of bio's dispatched in current slice */
73 unsigned int io_disp[2];
75 /* When did we start a new slice */
76 unsigned long slice_start[2];
77 unsigned long slice_end[2];
79 /* Some throttle limits got updated for the group */
82 struct rcu_head rcu_head;
87 /* List of throtl groups */
88 struct hlist_head tg_list;
90 /* service tree for active throtl groups */
91 struct throtl_rb_root tg_service_tree;
93 struct throtl_grp *root_tg;
94 struct request_queue *queue;
96 /* Total Number of queued bios on READ and WRITE lists */
97 unsigned int nr_queued[2];
100 * number of total undestroyed groups
102 unsigned int nr_undestroyed_grps;
104 /* Work for dispatching throttled bios */
105 struct delayed_work throtl_work;
110 enum tg_state_flags {
111 THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */
114 #define THROTL_TG_FNS(name) \
115 static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \
117 (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \
119 static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \
121 (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \
123 static inline int throtl_tg_##name(const struct throtl_grp *tg) \
125 return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \
128 THROTL_TG_FNS(on_rr);
130 #define throtl_log_tg(td, tg, fmt, args...) \
131 blk_add_trace_msg((td)->queue, "throtl %s " fmt, \
132 blkg_path(&(tg)->blkg), ##args); \
134 #define throtl_log(td, fmt, args...) \
135 blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
137 static inline struct throtl_grp *tg_of_blkg(struct blkio_group *blkg)
140 return container_of(blkg, struct throtl_grp, blkg);
145 static inline int total_nr_queued(struct throtl_data *td)
147 return (td->nr_queued[0] + td->nr_queued[1]);
150 static inline struct throtl_grp *throtl_ref_get_tg(struct throtl_grp *tg)
152 atomic_inc(&tg->ref);
156 static void throtl_free_tg(struct rcu_head *head)
158 struct throtl_grp *tg;
160 tg = container_of(head, struct throtl_grp, rcu_head);
161 free_percpu(tg->blkg.stats_cpu);
165 static void throtl_put_tg(struct throtl_grp *tg)
167 BUG_ON(atomic_read(&tg->ref) <= 0);
168 if (!atomic_dec_and_test(&tg->ref))
172 * A group is freed in rcu manner. But having an rcu lock does not
173 * mean that one can access all the fields of blkg and assume these
174 * are valid. For example, don't try to follow throtl_data and
175 * request queue links.
177 * Having a reference to blkg under an rcu allows acess to only
178 * values local to groups like group stats and group rate limits
180 call_rcu(&tg->rcu_head, throtl_free_tg);
183 static void throtl_init_group(struct throtl_grp *tg)
185 INIT_HLIST_NODE(&tg->tg_node);
186 RB_CLEAR_NODE(&tg->rb_node);
187 bio_list_init(&tg->bio_lists[0]);
188 bio_list_init(&tg->bio_lists[1]);
189 tg->limits_changed = false;
191 /* Practically unlimited BW */
192 tg->bps[0] = tg->bps[1] = -1;
193 tg->iops[0] = tg->iops[1] = -1;
196 * Take the initial reference that will be released on destroy
197 * This can be thought of a joint reference by cgroup and
198 * request queue which will be dropped by either request queue
199 * exit or cgroup deletion path depending on who is exiting first.
201 atomic_set(&tg->ref, 1);
204 /* Should be called with rcu read lock held (needed for blkcg) */
206 throtl_add_group_to_td_list(struct throtl_data *td, struct throtl_grp *tg)
208 hlist_add_head(&tg->tg_node, &td->tg_list);
209 td->nr_undestroyed_grps++;
213 __throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg)
215 struct backing_dev_info *bdi = &td->queue->backing_dev_info;
216 unsigned int major, minor;
218 if (!tg || tg->blkg.dev)
222 * Fill in device details for a group which might not have been
223 * filled at group creation time as queue was being instantiated
224 * and driver had not attached a device yet
226 if (bdi->dev && dev_name(bdi->dev)) {
227 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
228 tg->blkg.dev = MKDEV(major, minor);
232 static void throtl_init_add_tg_lists(struct throtl_data *td,
233 struct throtl_grp *tg, struct blkio_cgroup *blkcg)
235 __throtl_tg_fill_dev_details(td, tg);
237 /* Add group onto cgroup list */
238 blkiocg_add_blkio_group(blkcg, &tg->blkg, (void *)td,
239 tg->blkg.dev, BLKIO_POLICY_THROTL);
241 tg->bps[READ] = blkcg_get_read_bps(blkcg, tg->blkg.dev);
242 tg->bps[WRITE] = blkcg_get_write_bps(blkcg, tg->blkg.dev);
243 tg->iops[READ] = blkcg_get_read_iops(blkcg, tg->blkg.dev);
244 tg->iops[WRITE] = blkcg_get_write_iops(blkcg, tg->blkg.dev);
246 throtl_add_group_to_td_list(td, tg);
249 /* Should be called without queue lock and outside of rcu period */
250 static struct throtl_grp *throtl_alloc_tg(struct throtl_data *td)
252 struct throtl_grp *tg = NULL;
255 tg = kzalloc_node(sizeof(*tg), GFP_ATOMIC, td->queue->node);
259 ret = blkio_alloc_blkg_stats(&tg->blkg);
266 throtl_init_group(tg);
271 throtl_grp *throtl_find_tg(struct throtl_data *td, struct blkio_cgroup *blkcg)
273 struct throtl_grp *tg = NULL;
277 * This is the common case when there are no blkio cgroups.
278 * Avoid lookup in this case
280 if (blkcg == &blkio_root_cgroup)
283 tg = tg_of_blkg(blkiocg_lookup_group(blkcg, key));
285 __throtl_tg_fill_dev_details(td, tg);
290 * This function returns with queue lock unlocked in case of error, like
291 * request queue is no more
293 static struct throtl_grp * throtl_get_tg(struct throtl_data *td)
295 struct throtl_grp *tg = NULL, *__tg = NULL;
296 struct blkio_cgroup *blkcg;
297 struct request_queue *q = td->queue;
300 blkcg = task_blkio_cgroup(current);
301 tg = throtl_find_tg(td, blkcg);
308 * Need to allocate a group. Allocation of group also needs allocation
309 * of per cpu stats which in-turn takes a mutex() and can block. Hence
310 * we need to drop rcu lock and queue_lock before we call alloc
312 * Take the request queue reference to make sure queue does not
313 * go away once we return from allocation.
317 spin_unlock_irq(q->queue_lock);
319 tg = throtl_alloc_tg(td);
321 * We might have slept in group allocation. Make sure queue is not
324 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
329 return ERR_PTR(-ENODEV);
333 /* Group allocated and queue is still alive. take the lock */
334 spin_lock_irq(q->queue_lock);
337 * Initialize the new group. After sleeping, read the blkcg again.
340 blkcg = task_blkio_cgroup(current);
343 * If some other thread already allocated the group while we were
344 * not holding queue lock, free up the group
346 __tg = throtl_find_tg(td, blkcg);
354 /* Group allocation failed. Account the IO to root group */
360 throtl_init_add_tg_lists(td, tg, blkcg);
365 static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root)
367 /* Service tree is empty */
372 root->left = rb_first(&root->rb);
375 return rb_entry_tg(root->left);
380 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
386 static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root)
390 rb_erase_init(n, &root->rb);
394 static void update_min_dispatch_time(struct throtl_rb_root *st)
396 struct throtl_grp *tg;
398 tg = throtl_rb_first(st);
402 st->min_disptime = tg->disptime;
406 tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg)
408 struct rb_node **node = &st->rb.rb_node;
409 struct rb_node *parent = NULL;
410 struct throtl_grp *__tg;
411 unsigned long key = tg->disptime;
414 while (*node != NULL) {
416 __tg = rb_entry_tg(parent);
418 if (time_before(key, __tg->disptime))
419 node = &parent->rb_left;
421 node = &parent->rb_right;
427 st->left = &tg->rb_node;
429 rb_link_node(&tg->rb_node, parent, node);
430 rb_insert_color(&tg->rb_node, &st->rb);
433 static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
435 struct throtl_rb_root *st = &td->tg_service_tree;
437 tg_service_tree_add(st, tg);
438 throtl_mark_tg_on_rr(tg);
442 static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
444 if (!throtl_tg_on_rr(tg))
445 __throtl_enqueue_tg(td, tg);
448 static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
450 throtl_rb_erase(&tg->rb_node, &td->tg_service_tree);
451 throtl_clear_tg_on_rr(tg);
454 static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
456 if (throtl_tg_on_rr(tg))
457 __throtl_dequeue_tg(td, tg);
460 static void throtl_schedule_next_dispatch(struct throtl_data *td)
462 struct throtl_rb_root *st = &td->tg_service_tree;
465 * If there are more bios pending, schedule more work.
467 if (!total_nr_queued(td))
472 update_min_dispatch_time(st);
474 if (time_before_eq(st->min_disptime, jiffies))
475 throtl_schedule_delayed_work(td, 0);
477 throtl_schedule_delayed_work(td, (st->min_disptime - jiffies));
481 throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
483 tg->bytes_disp[rw] = 0;
485 tg->slice_start[rw] = jiffies;
486 tg->slice_end[rw] = jiffies + throtl_slice;
487 throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
488 rw == READ ? 'R' : 'W', tg->slice_start[rw],
489 tg->slice_end[rw], jiffies);
492 static inline void throtl_set_slice_end(struct throtl_data *td,
493 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
495 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
498 static inline void throtl_extend_slice(struct throtl_data *td,
499 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
501 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
502 throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
503 rw == READ ? 'R' : 'W', tg->slice_start[rw],
504 tg->slice_end[rw], jiffies);
507 /* Determine if previously allocated or extended slice is complete or not */
509 throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
511 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
517 /* Trim the used slices and adjust slice start accordingly */
519 throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
521 unsigned long nr_slices, time_elapsed, io_trim;
524 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
527 * If bps are unlimited (-1), then time slice don't get
528 * renewed. Don't try to trim the slice if slice is used. A new
529 * slice will start when appropriate.
531 if (throtl_slice_used(td, tg, rw))
535 * A bio has been dispatched. Also adjust slice_end. It might happen
536 * that initially cgroup limit was very low resulting in high
537 * slice_end, but later limit was bumped up and bio was dispached
538 * sooner, then we need to reduce slice_end. A high bogus slice_end
539 * is bad because it does not allow new slice to start.
542 throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice);
544 time_elapsed = jiffies - tg->slice_start[rw];
546 nr_slices = time_elapsed / throtl_slice;
550 tmp = tg->bps[rw] * throtl_slice * nr_slices;
554 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
556 if (!bytes_trim && !io_trim)
559 if (tg->bytes_disp[rw] >= bytes_trim)
560 tg->bytes_disp[rw] -= bytes_trim;
562 tg->bytes_disp[rw] = 0;
564 if (tg->io_disp[rw] >= io_trim)
565 tg->io_disp[rw] -= io_trim;
569 tg->slice_start[rw] += nr_slices * throtl_slice;
571 throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
572 " start=%lu end=%lu jiffies=%lu",
573 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
574 tg->slice_start[rw], tg->slice_end[rw], jiffies);
577 static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
578 struct bio *bio, unsigned long *wait)
580 bool rw = bio_data_dir(bio);
581 unsigned int io_allowed;
582 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
585 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
587 /* Slice has just started. Consider one slice interval */
589 jiffy_elapsed_rnd = throtl_slice;
591 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
594 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
595 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
596 * will allow dispatch after 1 second and after that slice should
600 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
604 io_allowed = UINT_MAX;
608 if (tg->io_disp[rw] + 1 <= io_allowed) {
614 /* Calc approx time to dispatch */
615 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
617 if (jiffy_wait > jiffy_elapsed)
618 jiffy_wait = jiffy_wait - jiffy_elapsed;
627 static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
628 struct bio *bio, unsigned long *wait)
630 bool rw = bio_data_dir(bio);
631 u64 bytes_allowed, extra_bytes, tmp;
632 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
634 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
636 /* Slice has just started. Consider one slice interval */
638 jiffy_elapsed_rnd = throtl_slice;
640 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
642 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
646 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
652 /* Calc approx time to dispatch */
653 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
654 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
660 * This wait time is without taking into consideration the rounding
661 * up we did. Add that time also.
663 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
670 * Returns whether one can dispatch a bio or not. Also returns approx number
671 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
673 static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
674 struct bio *bio, unsigned long *wait)
676 bool rw = bio_data_dir(bio);
677 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
680 * Currently whole state machine of group depends on first bio
681 * queued in the group bio list. So one should not be calling
682 * this function with a different bio if there are other bios
685 BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
687 /* If tg->bps = -1, then BW is unlimited */
688 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
695 * If previous slice expired, start a new one otherwise renew/extend
696 * existing slice to make sure it is at least throtl_slice interval
699 if (throtl_slice_used(td, tg, rw))
700 throtl_start_new_slice(td, tg, rw);
702 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
703 throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
706 if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
707 && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
713 max_wait = max(bps_wait, iops_wait);
718 if (time_before(tg->slice_end[rw], jiffies + max_wait))
719 throtl_extend_slice(td, tg, rw, jiffies + max_wait);
724 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
726 bool rw = bio_data_dir(bio);
727 bool sync = bio->bi_rw & REQ_SYNC;
729 /* Charge the bio to the group */
730 tg->bytes_disp[rw] += bio->bi_size;
734 * TODO: This will take blkg->stats_lock. Figure out a way
735 * to avoid this cost.
737 blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync);
740 static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
743 bool rw = bio_data_dir(bio);
745 bio_list_add(&tg->bio_lists[rw], bio);
746 /* Take a bio reference on tg */
747 throtl_ref_get_tg(tg);
750 throtl_enqueue_tg(td, tg);
753 static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
755 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
758 if ((bio = bio_list_peek(&tg->bio_lists[READ])))
759 tg_may_dispatch(td, tg, bio, &read_wait);
761 if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
762 tg_may_dispatch(td, tg, bio, &write_wait);
764 min_wait = min(read_wait, write_wait);
765 disptime = jiffies + min_wait;
767 /* Update dispatch time */
768 throtl_dequeue_tg(td, tg);
769 tg->disptime = disptime;
770 throtl_enqueue_tg(td, tg);
773 static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
774 bool rw, struct bio_list *bl)
778 bio = bio_list_pop(&tg->bio_lists[rw]);
780 /* Drop bio reference on tg */
783 BUG_ON(td->nr_queued[rw] <= 0);
786 throtl_charge_bio(tg, bio);
787 bio_list_add(bl, bio);
788 bio->bi_rw |= REQ_THROTTLED;
790 throtl_trim_slice(td, tg, rw);
793 static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
796 unsigned int nr_reads = 0, nr_writes = 0;
797 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
798 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
801 /* Try to dispatch 75% READS and 25% WRITES */
803 while ((bio = bio_list_peek(&tg->bio_lists[READ]))
804 && tg_may_dispatch(td, tg, bio, NULL)) {
806 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
809 if (nr_reads >= max_nr_reads)
813 while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
814 && tg_may_dispatch(td, tg, bio, NULL)) {
816 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
819 if (nr_writes >= max_nr_writes)
823 return nr_reads + nr_writes;
826 static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
828 unsigned int nr_disp = 0;
829 struct throtl_grp *tg;
830 struct throtl_rb_root *st = &td->tg_service_tree;
833 tg = throtl_rb_first(st);
838 if (time_before(jiffies, tg->disptime))
841 throtl_dequeue_tg(td, tg);
843 nr_disp += throtl_dispatch_tg(td, tg, bl);
845 if (tg->nr_queued[0] || tg->nr_queued[1]) {
846 tg_update_disptime(td, tg);
847 throtl_enqueue_tg(td, tg);
850 if (nr_disp >= throtl_quantum)
857 static void throtl_process_limit_change(struct throtl_data *td)
859 struct throtl_grp *tg;
860 struct hlist_node *pos, *n;
862 if (!td->limits_changed)
865 xchg(&td->limits_changed, false);
867 throtl_log(td, "limits changed");
869 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
870 if (!tg->limits_changed)
873 if (!xchg(&tg->limits_changed, false))
876 throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
877 " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
878 tg->iops[READ], tg->iops[WRITE]);
881 * Restart the slices for both READ and WRITES. It
882 * might happen that a group's limit are dropped
883 * suddenly and we don't want to account recently
884 * dispatched IO with new low rate
886 throtl_start_new_slice(td, tg, 0);
887 throtl_start_new_slice(td, tg, 1);
889 if (throtl_tg_on_rr(tg))
890 tg_update_disptime(td, tg);
894 /* Dispatch throttled bios. Should be called without queue lock held. */
895 static int throtl_dispatch(struct request_queue *q)
897 struct throtl_data *td = q->td;
898 unsigned int nr_disp = 0;
899 struct bio_list bio_list_on_stack;
901 struct blk_plug plug;
903 spin_lock_irq(q->queue_lock);
905 throtl_process_limit_change(td);
907 if (!total_nr_queued(td))
910 bio_list_init(&bio_list_on_stack);
912 throtl_log(td, "dispatch nr_queued=%lu read=%u write=%u",
913 total_nr_queued(td), td->nr_queued[READ],
914 td->nr_queued[WRITE]);
916 nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
919 throtl_log(td, "bios disp=%u", nr_disp);
921 throtl_schedule_next_dispatch(td);
923 spin_unlock_irq(q->queue_lock);
926 * If we dispatched some requests, unplug the queue to make sure
930 blk_start_plug(&plug);
931 while((bio = bio_list_pop(&bio_list_on_stack)))
932 generic_make_request(bio);
933 blk_finish_plug(&plug);
938 void blk_throtl_work(struct work_struct *work)
940 struct throtl_data *td = container_of(work, struct throtl_data,
942 struct request_queue *q = td->queue;
947 /* Call with queue lock held */
949 throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay)
952 struct delayed_work *dwork = &td->throtl_work;
954 /* schedule work if limits changed even if no bio is queued */
955 if (total_nr_queued(td) > 0 || td->limits_changed) {
957 * We might have a work scheduled to be executed in future.
958 * Cancel that and schedule a new one.
960 __cancel_delayed_work(dwork);
961 queue_delayed_work(kthrotld_workqueue, dwork, delay);
962 throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
968 throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg)
970 /* Something wrong if we are trying to remove same group twice */
971 BUG_ON(hlist_unhashed(&tg->tg_node));
973 hlist_del_init(&tg->tg_node);
976 * Put the reference taken at the time of creation so that when all
977 * queues are gone, group can be destroyed.
980 td->nr_undestroyed_grps--;
983 static void throtl_release_tgs(struct throtl_data *td)
985 struct hlist_node *pos, *n;
986 struct throtl_grp *tg;
988 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
990 * If cgroup removal path got to blk_group first and removed
991 * it from cgroup list, then it will take care of destroying
994 if (!blkiocg_del_blkio_group(&tg->blkg))
995 throtl_destroy_tg(td, tg);
999 static void throtl_td_free(struct throtl_data *td)
1005 * Blk cgroup controller notification saying that blkio_group object is being
1006 * delinked as associated cgroup object is going away. That also means that
1007 * no new IO will come in this group. So get rid of this group as soon as
1008 * any pending IO in the group is finished.
1010 * This function is called under rcu_read_lock(). key is the rcu protected
1011 * pointer. That means "key" is a valid throtl_data pointer as long as we are
1014 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1015 * it should not be NULL as even if queue was going away, cgroup deltion
1016 * path got to it first.
1018 void throtl_unlink_blkio_group(void *key, struct blkio_group *blkg)
1020 unsigned long flags;
1021 struct throtl_data *td = key;
1023 spin_lock_irqsave(td->queue->queue_lock, flags);
1024 throtl_destroy_tg(td, tg_of_blkg(blkg));
1025 spin_unlock_irqrestore(td->queue->queue_lock, flags);
1028 static void throtl_update_blkio_group_common(struct throtl_data *td,
1029 struct throtl_grp *tg)
1031 xchg(&tg->limits_changed, true);
1032 xchg(&td->limits_changed, true);
1033 /* Schedule a work now to process the limit change */
1034 throtl_schedule_delayed_work(td, 0);
1038 * For all update functions, key should be a valid pointer because these
1039 * update functions are called under blkcg_lock, that means, blkg is
1040 * valid and in turn key is valid. queue exit path can not race because
1043 * Can not take queue lock in update functions as queue lock under blkcg_lock
1044 * is not allowed. Under other paths we take blkcg_lock under queue_lock.
1046 static void throtl_update_blkio_group_read_bps(void *key,
1047 struct blkio_group *blkg, u64 read_bps)
1049 struct throtl_data *td = key;
1050 struct throtl_grp *tg = tg_of_blkg(blkg);
1052 tg->bps[READ] = read_bps;
1053 throtl_update_blkio_group_common(td, tg);
1056 static void throtl_update_blkio_group_write_bps(void *key,
1057 struct blkio_group *blkg, u64 write_bps)
1059 struct throtl_data *td = key;
1060 struct throtl_grp *tg = tg_of_blkg(blkg);
1062 tg->bps[WRITE] = write_bps;
1063 throtl_update_blkio_group_common(td, tg);
1066 static void throtl_update_blkio_group_read_iops(void *key,
1067 struct blkio_group *blkg, unsigned int read_iops)
1069 struct throtl_data *td = key;
1070 struct throtl_grp *tg = tg_of_blkg(blkg);
1072 tg->iops[READ] = read_iops;
1073 throtl_update_blkio_group_common(td, tg);
1076 static void throtl_update_blkio_group_write_iops(void *key,
1077 struct blkio_group *blkg, unsigned int write_iops)
1079 struct throtl_data *td = key;
1080 struct throtl_grp *tg = tg_of_blkg(blkg);
1082 tg->iops[WRITE] = write_iops;
1083 throtl_update_blkio_group_common(td, tg);
1086 static void throtl_shutdown_wq(struct request_queue *q)
1088 struct throtl_data *td = q->td;
1090 cancel_delayed_work_sync(&td->throtl_work);
1093 static struct blkio_policy_type blkio_policy_throtl = {
1095 .blkio_unlink_group_fn = throtl_unlink_blkio_group,
1096 .blkio_update_group_read_bps_fn =
1097 throtl_update_blkio_group_read_bps,
1098 .blkio_update_group_write_bps_fn =
1099 throtl_update_blkio_group_write_bps,
1100 .blkio_update_group_read_iops_fn =
1101 throtl_update_blkio_group_read_iops,
1102 .blkio_update_group_write_iops_fn =
1103 throtl_update_blkio_group_write_iops,
1105 .plid = BLKIO_POLICY_THROTL,
1108 int blk_throtl_bio(struct request_queue *q, struct bio **biop)
1110 struct throtl_data *td = q->td;
1111 struct throtl_grp *tg;
1112 struct bio *bio = *biop;
1113 bool rw = bio_data_dir(bio), update_disptime = true;
1115 if (bio->bi_rw & REQ_THROTTLED) {
1116 bio->bi_rw &= ~REQ_THROTTLED;
1120 spin_lock_irq(q->queue_lock);
1121 tg = throtl_get_tg(td);
1124 if (PTR_ERR(tg) == -ENODEV) {
1126 * Queue is gone. No queue lock held here.
1132 if (tg->nr_queued[rw]) {
1134 * There is already another bio queued in same dir. No
1135 * need to update dispatch time.
1137 update_disptime = false;
1142 /* Bio is with-in rate limit of group */
1143 if (tg_may_dispatch(td, tg, bio, NULL)) {
1144 throtl_charge_bio(tg, bio);
1147 * We need to trim slice even when bios are not being queued
1148 * otherwise it might happen that a bio is not queued for
1149 * a long time and slice keeps on extending and trim is not
1150 * called for a long time. Now if limits are reduced suddenly
1151 * we take into account all the IO dispatched so far at new
1152 * low rate and * newly queued IO gets a really long dispatch
1155 * So keep on trimming slice even if bio is not queued.
1157 throtl_trim_slice(td, tg, rw);
1162 throtl_log_tg(td, tg, "[%c] bio. bdisp=%u sz=%u bps=%llu"
1163 " iodisp=%u iops=%u queued=%d/%d",
1164 rw == READ ? 'R' : 'W',
1165 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1166 tg->io_disp[rw], tg->iops[rw],
1167 tg->nr_queued[READ], tg->nr_queued[WRITE]);
1169 throtl_add_bio_tg(q->td, tg, bio);
1172 if (update_disptime) {
1173 tg_update_disptime(td, tg);
1174 throtl_schedule_next_dispatch(td);
1178 spin_unlock_irq(q->queue_lock);
1182 int blk_throtl_init(struct request_queue *q)
1184 struct throtl_data *td;
1185 struct throtl_grp *tg;
1187 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1191 INIT_HLIST_HEAD(&td->tg_list);
1192 td->tg_service_tree = THROTL_RB_ROOT;
1193 td->limits_changed = false;
1194 INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);
1196 /* alloc and Init root group. */
1198 tg = throtl_alloc_tg(td);
1208 throtl_init_add_tg_lists(td, tg, &blkio_root_cgroup);
1211 /* Attach throtl data to request queue */
1216 void blk_throtl_exit(struct request_queue *q)
1218 struct throtl_data *td = q->td;
1223 throtl_shutdown_wq(q);
1225 spin_lock_irq(q->queue_lock);
1226 throtl_release_tgs(td);
1228 /* If there are other groups */
1229 if (td->nr_undestroyed_grps > 0)
1232 spin_unlock_irq(q->queue_lock);
1235 * Wait for tg->blkg->key accessors to exit their grace periods.
1236 * Do this wait only if there are other undestroyed groups out
1237 * there (other than root group). This can happen if cgroup deletion
1238 * path claimed the responsibility of cleaning up a group before
1239 * queue cleanup code get to the group.
1241 * Do not call synchronize_rcu() unconditionally as there are drivers
1242 * which create/delete request queue hundreds of times during scan/boot
1243 * and synchronize_rcu() can take significant time and slow down boot.
1249 * Just being safe to make sure after previous flush if some body did
1250 * update limits through cgroup and another work got queued, cancel
1253 throtl_shutdown_wq(q);
1257 static int __init throtl_init(void)
1259 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1260 if (!kthrotld_workqueue)
1261 panic("Failed to create kthrotld\n");
1263 blkio_policy_register(&blkio_policy_throtl);
1267 module_init(throtl_init);