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 */
50 THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */
51 THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */
54 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
56 /* Per-cpu group stats */
58 /* total bytes transferred */
59 struct blkg_rwstat service_bytes;
60 /* total IOs serviced, post merge */
61 struct blkg_rwstat serviced;
65 /* must be the first member */
66 struct blkg_policy_data pd;
68 /* active throtl group service_queue member */
69 struct rb_node rb_node;
71 /* throtl_data this group belongs to */
72 struct throtl_data *td;
74 /* this group's service queue */
75 struct throtl_service_queue service_queue;
78 * Dispatch time in jiffies. This is the estimated time when group
79 * will unthrottle and is ready to dispatch more bio. It is used as
80 * key to sort active groups in service tree.
82 unsigned long disptime;
86 /* bytes per second rate limits */
92 /* Number of bytes disptached in current slice */
93 uint64_t bytes_disp[2];
94 /* Number of bio's dispatched in current slice */
95 unsigned int io_disp[2];
97 /* When did we start a new slice */
98 unsigned long slice_start[2];
99 unsigned long slice_end[2];
101 /* Per cpu stats pointer */
102 struct tg_stats_cpu __percpu *stats_cpu;
104 /* List of tgs waiting for per cpu stats memory to be allocated */
105 struct list_head stats_alloc_node;
110 /* service tree for active throtl groups */
111 struct throtl_service_queue service_queue;
113 struct request_queue *queue;
115 /* Total Number of queued bios on READ and WRITE lists */
116 unsigned int nr_queued[2];
119 * number of total undestroyed groups
121 unsigned int nr_undestroyed_grps;
123 /* Work for dispatching throttled bios */
124 struct delayed_work dispatch_work;
127 /* list and work item to allocate percpu group stats */
128 static DEFINE_SPINLOCK(tg_stats_alloc_lock);
129 static LIST_HEAD(tg_stats_alloc_list);
131 static void tg_stats_alloc_fn(struct work_struct *);
132 static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn);
134 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
136 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
139 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
141 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
144 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
146 return pd_to_blkg(&tg->pd);
149 static inline struct throtl_grp *td_root_tg(struct throtl_data *td)
151 return blkg_to_tg(td->queue->root_blkg);
155 * sq_to_tg - return the throl_grp the specified service queue belongs to
156 * @sq: the throtl_service_queue of interest
158 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
159 * embedded in throtl_data, %NULL is returned.
161 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
163 if (sq && sq->parent_sq)
164 return container_of(sq, struct throtl_grp, service_queue);
170 * sq_to_td - return throtl_data the specified service queue belongs to
171 * @sq: the throtl_service_queue of interest
173 * A service_queue can be embeded in either a throtl_grp or throtl_data.
174 * Determine the associated throtl_data accordingly and return it.
176 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
178 struct throtl_grp *tg = sq_to_tg(sq);
183 return container_of(sq, struct throtl_data, service_queue);
187 * throtl_log - log debug message via blktrace
188 * @sq: the service_queue being reported
189 * @fmt: printf format string
192 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
193 * throtl_grp; otherwise, just "throtl".
195 * TODO: this should be made a function and name formatting should happen
196 * after testing whether blktrace is enabled.
198 #define throtl_log(sq, fmt, args...) do { \
199 struct throtl_grp *__tg = sq_to_tg((sq)); \
200 struct throtl_data *__td = sq_to_td((sq)); \
206 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
207 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
209 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
214 * Worker for allocating per cpu stat for tgs. This is scheduled on the
215 * system_wq once there are some groups on the alloc_list waiting for
218 static void tg_stats_alloc_fn(struct work_struct *work)
220 static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */
221 struct delayed_work *dwork = to_delayed_work(work);
226 stats_cpu = alloc_percpu(struct tg_stats_cpu);
228 /* allocation failed, try again after some time */
229 schedule_delayed_work(dwork, msecs_to_jiffies(10));
234 spin_lock_irq(&tg_stats_alloc_lock);
236 if (!list_empty(&tg_stats_alloc_list)) {
237 struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list,
240 swap(tg->stats_cpu, stats_cpu);
241 list_del_init(&tg->stats_alloc_node);
244 empty = list_empty(&tg_stats_alloc_list);
245 spin_unlock_irq(&tg_stats_alloc_lock);
250 /* init a service_queue, assumes the caller zeroed it */
251 static void throtl_service_queue_init(struct throtl_service_queue *sq,
252 struct throtl_service_queue *parent_sq)
254 bio_list_init(&sq->bio_lists[0]);
255 bio_list_init(&sq->bio_lists[1]);
256 sq->pending_tree = RB_ROOT;
257 sq->parent_sq = parent_sq;
260 static void throtl_pd_init(struct blkcg_gq *blkg)
262 struct throtl_grp *tg = blkg_to_tg(blkg);
263 struct throtl_data *td = blkg->q->td;
266 throtl_service_queue_init(&tg->service_queue, &td->service_queue);
267 RB_CLEAR_NODE(&tg->rb_node);
273 tg->iops[WRITE] = -1;
276 * Ugh... We need to perform per-cpu allocation for tg->stats_cpu
277 * but percpu allocator can't be called from IO path. Queue tg on
278 * tg_stats_alloc_list and allocate from work item.
280 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
281 list_add(&tg->stats_alloc_node, &tg_stats_alloc_list);
282 schedule_delayed_work(&tg_stats_alloc_work, 0);
283 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
286 static void throtl_pd_exit(struct blkcg_gq *blkg)
288 struct throtl_grp *tg = blkg_to_tg(blkg);
291 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
292 list_del_init(&tg->stats_alloc_node);
293 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
295 free_percpu(tg->stats_cpu);
298 static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
300 struct throtl_grp *tg = blkg_to_tg(blkg);
303 if (tg->stats_cpu == NULL)
306 for_each_possible_cpu(cpu) {
307 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
309 blkg_rwstat_reset(&sc->service_bytes);
310 blkg_rwstat_reset(&sc->serviced);
314 static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
318 * This is the common case when there are no blkcgs. Avoid lookup
321 if (blkcg == &blkcg_root)
322 return td_root_tg(td);
324 return blkg_to_tg(blkg_lookup(blkcg, td->queue));
327 static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
330 struct request_queue *q = td->queue;
331 struct throtl_grp *tg = NULL;
334 * This is the common case when there are no blkcgs. Avoid lookup
337 if (blkcg == &blkcg_root) {
340 struct blkcg_gq *blkg;
342 blkg = blkg_lookup_create(blkcg, q);
344 /* if %NULL and @q is alive, fall back to root_tg */
346 tg = blkg_to_tg(blkg);
347 else if (!blk_queue_dying(q))
354 static struct throtl_grp *
355 throtl_rb_first(struct throtl_service_queue *parent_sq)
357 /* Service tree is empty */
358 if (!parent_sq->nr_pending)
361 if (!parent_sq->first_pending)
362 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
364 if (parent_sq->first_pending)
365 return rb_entry_tg(parent_sq->first_pending);
370 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
376 static void throtl_rb_erase(struct rb_node *n,
377 struct throtl_service_queue *parent_sq)
379 if (parent_sq->first_pending == n)
380 parent_sq->first_pending = NULL;
381 rb_erase_init(n, &parent_sq->pending_tree);
382 --parent_sq->nr_pending;
385 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
387 struct throtl_grp *tg;
389 tg = throtl_rb_first(parent_sq);
393 parent_sq->first_pending_disptime = tg->disptime;
396 static void tg_service_queue_add(struct throtl_grp *tg)
398 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
399 struct rb_node **node = &parent_sq->pending_tree.rb_node;
400 struct rb_node *parent = NULL;
401 struct throtl_grp *__tg;
402 unsigned long key = tg->disptime;
405 while (*node != NULL) {
407 __tg = rb_entry_tg(parent);
409 if (time_before(key, __tg->disptime))
410 node = &parent->rb_left;
412 node = &parent->rb_right;
418 parent_sq->first_pending = &tg->rb_node;
420 rb_link_node(&tg->rb_node, parent, node);
421 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
424 static void __throtl_enqueue_tg(struct throtl_grp *tg)
426 tg_service_queue_add(tg);
427 tg->flags |= THROTL_TG_PENDING;
428 tg->service_queue.parent_sq->nr_pending++;
431 static void throtl_enqueue_tg(struct throtl_grp *tg)
433 if (!(tg->flags & THROTL_TG_PENDING))
434 __throtl_enqueue_tg(tg);
437 static void __throtl_dequeue_tg(struct throtl_grp *tg)
439 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
440 tg->flags &= ~THROTL_TG_PENDING;
443 static void throtl_dequeue_tg(struct throtl_grp *tg)
445 if (tg->flags & THROTL_TG_PENDING)
446 __throtl_dequeue_tg(tg);
449 /* Call with queue lock held */
450 static void throtl_schedule_delayed_work(struct throtl_data *td,
453 struct delayed_work *dwork = &td->dispatch_work;
454 struct throtl_service_queue *sq = &td->service_queue;
456 mod_delayed_work(kthrotld_workqueue, dwork, delay);
457 throtl_log(sq, "schedule work. delay=%lu jiffies=%lu", delay, jiffies);
460 static void throtl_schedule_next_dispatch(struct throtl_data *td)
462 struct throtl_service_queue *sq = &td->service_queue;
464 /* any pending children left? */
468 update_min_dispatch_time(sq);
470 if (time_before_eq(sq->first_pending_disptime, jiffies))
471 throtl_schedule_delayed_work(td, 0);
473 throtl_schedule_delayed_work(td, sq->first_pending_disptime - jiffies);
476 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
478 tg->bytes_disp[rw] = 0;
480 tg->slice_start[rw] = jiffies;
481 tg->slice_end[rw] = jiffies + throtl_slice;
482 throtl_log(&tg->service_queue,
483 "[%c] new slice start=%lu end=%lu jiffies=%lu",
484 rw == READ ? 'R' : 'W', tg->slice_start[rw],
485 tg->slice_end[rw], jiffies);
488 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
489 unsigned long jiffy_end)
491 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
494 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
495 unsigned long jiffy_end)
497 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
498 throtl_log(&tg->service_queue,
499 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
500 rw == READ ? 'R' : 'W', tg->slice_start[rw],
501 tg->slice_end[rw], jiffies);
504 /* Determine if previously allocated or extended slice is complete or not */
505 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
507 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
513 /* Trim the used slices and adjust slice start accordingly */
514 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
516 unsigned long nr_slices, time_elapsed, io_trim;
519 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
522 * If bps are unlimited (-1), then time slice don't get
523 * renewed. Don't try to trim the slice if slice is used. A new
524 * slice will start when appropriate.
526 if (throtl_slice_used(tg, rw))
530 * A bio has been dispatched. Also adjust slice_end. It might happen
531 * that initially cgroup limit was very low resulting in high
532 * slice_end, but later limit was bumped up and bio was dispached
533 * sooner, then we need to reduce slice_end. A high bogus slice_end
534 * is bad because it does not allow new slice to start.
537 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
539 time_elapsed = jiffies - tg->slice_start[rw];
541 nr_slices = time_elapsed / throtl_slice;
545 tmp = tg->bps[rw] * throtl_slice * nr_slices;
549 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
551 if (!bytes_trim && !io_trim)
554 if (tg->bytes_disp[rw] >= bytes_trim)
555 tg->bytes_disp[rw] -= bytes_trim;
557 tg->bytes_disp[rw] = 0;
559 if (tg->io_disp[rw] >= io_trim)
560 tg->io_disp[rw] -= io_trim;
564 tg->slice_start[rw] += nr_slices * throtl_slice;
566 throtl_log(&tg->service_queue,
567 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
568 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
569 tg->slice_start[rw], tg->slice_end[rw], jiffies);
572 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
575 bool rw = bio_data_dir(bio);
576 unsigned int io_allowed;
577 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
580 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
582 /* Slice has just started. Consider one slice interval */
584 jiffy_elapsed_rnd = throtl_slice;
586 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
589 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
590 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
591 * will allow dispatch after 1 second and after that slice should
595 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
599 io_allowed = UINT_MAX;
603 if (tg->io_disp[rw] + 1 <= io_allowed) {
609 /* Calc approx time to dispatch */
610 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
612 if (jiffy_wait > jiffy_elapsed)
613 jiffy_wait = jiffy_wait - jiffy_elapsed;
622 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
625 bool rw = bio_data_dir(bio);
626 u64 bytes_allowed, extra_bytes, tmp;
627 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
629 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
631 /* Slice has just started. Consider one slice interval */
633 jiffy_elapsed_rnd = throtl_slice;
635 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
637 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
641 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
647 /* Calc approx time to dispatch */
648 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
649 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
655 * This wait time is without taking into consideration the rounding
656 * up we did. Add that time also.
658 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
664 static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
665 if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
671 * Returns whether one can dispatch a bio or not. Also returns approx number
672 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
674 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
677 bool rw = bio_data_dir(bio);
678 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
681 * Currently whole state machine of group depends on first bio
682 * queued in the group bio list. So one should not be calling
683 * this function with a different bio if there are other bios
686 BUG_ON(tg->service_queue.nr_queued[rw] &&
687 bio != bio_list_peek(&tg->service_queue.bio_lists[rw]));
689 /* If tg->bps = -1, then BW is unlimited */
690 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
697 * If previous slice expired, start a new one otherwise renew/extend
698 * existing slice to make sure it is at least throtl_slice interval
701 if (throtl_slice_used(tg, rw))
702 throtl_start_new_slice(tg, rw);
704 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
705 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
708 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
709 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
715 max_wait = max(bps_wait, iops_wait);
720 if (time_before(tg->slice_end[rw], jiffies + max_wait))
721 throtl_extend_slice(tg, rw, jiffies + max_wait);
726 static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
729 struct throtl_grp *tg = blkg_to_tg(blkg);
730 struct tg_stats_cpu *stats_cpu;
733 /* If per cpu stats are not allocated yet, don't do any accounting. */
734 if (tg->stats_cpu == NULL)
738 * Disabling interrupts to provide mutual exclusion between two
739 * writes on same cpu. It probably is not needed for 64bit. Not
740 * optimizing that case yet.
742 local_irq_save(flags);
744 stats_cpu = this_cpu_ptr(tg->stats_cpu);
746 blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
747 blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
749 local_irq_restore(flags);
752 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
754 bool rw = bio_data_dir(bio);
756 /* Charge the bio to the group */
757 tg->bytes_disp[rw] += bio->bi_size;
761 * REQ_THROTTLED is used to prevent the same bio to be throttled
762 * more than once as a throttled bio will go through blk-throtl the
763 * second time when it eventually gets issued. Set it when a bio
764 * is being charged to a tg.
766 * Dispatch stats aren't recursive and each @bio should only be
767 * accounted by the @tg it was originally associated with. Let's
768 * update the stats when setting REQ_THROTTLED for the first time
769 * which is guaranteed to be for the @bio's original tg.
771 if (!(bio->bi_rw & REQ_THROTTLED)) {
772 bio->bi_rw |= REQ_THROTTLED;
773 throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size,
778 static void throtl_add_bio_tg(struct bio *bio, struct throtl_grp *tg)
780 struct throtl_service_queue *sq = &tg->service_queue;
781 bool rw = bio_data_dir(bio);
784 * If @tg doesn't currently have any bios queued in the same
785 * direction, queueing @bio can change when @tg should be
786 * dispatched. Mark that @tg was empty. This is automatically
787 * cleaered on the next tg_update_disptime().
789 if (!sq->nr_queued[rw])
790 tg->flags |= THROTL_TG_WAS_EMPTY;
792 bio_list_add(&sq->bio_lists[rw], bio);
793 /* Take a bio reference on tg */
794 blkg_get(tg_to_blkg(tg));
796 tg->td->nr_queued[rw]++;
797 throtl_enqueue_tg(tg);
800 static void tg_update_disptime(struct throtl_grp *tg)
802 struct throtl_service_queue *sq = &tg->service_queue;
803 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
806 if ((bio = bio_list_peek(&sq->bio_lists[READ])))
807 tg_may_dispatch(tg, bio, &read_wait);
809 if ((bio = bio_list_peek(&sq->bio_lists[WRITE])))
810 tg_may_dispatch(tg, bio, &write_wait);
812 min_wait = min(read_wait, write_wait);
813 disptime = jiffies + min_wait;
815 /* Update dispatch time */
816 throtl_dequeue_tg(tg);
817 tg->disptime = disptime;
818 throtl_enqueue_tg(tg);
820 /* see throtl_add_bio_tg() */
821 tg->flags &= ~THROTL_TG_WAS_EMPTY;
824 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
826 struct throtl_service_queue *sq = &tg->service_queue;
829 bio = bio_list_pop(&sq->bio_lists[rw]);
831 /* Drop bio reference on blkg */
832 blkg_put(tg_to_blkg(tg));
834 BUG_ON(tg->td->nr_queued[rw] <= 0);
835 tg->td->nr_queued[rw]--;
837 throtl_charge_bio(tg, bio);
838 bio_list_add(&sq->parent_sq->bio_lists[rw], bio);
840 throtl_trim_slice(tg, rw);
843 static int throtl_dispatch_tg(struct throtl_grp *tg)
845 struct throtl_service_queue *sq = &tg->service_queue;
846 unsigned int nr_reads = 0, nr_writes = 0;
847 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
848 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
851 /* Try to dispatch 75% READS and 25% WRITES */
853 while ((bio = bio_list_peek(&sq->bio_lists[READ])) &&
854 tg_may_dispatch(tg, bio, NULL)) {
856 tg_dispatch_one_bio(tg, bio_data_dir(bio));
859 if (nr_reads >= max_nr_reads)
863 while ((bio = bio_list_peek(&sq->bio_lists[WRITE])) &&
864 tg_may_dispatch(tg, bio, NULL)) {
866 tg_dispatch_one_bio(tg, bio_data_dir(bio));
869 if (nr_writes >= max_nr_writes)
873 return nr_reads + nr_writes;
876 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
878 unsigned int nr_disp = 0;
881 struct throtl_grp *tg = throtl_rb_first(parent_sq);
882 struct throtl_service_queue *sq = &tg->service_queue;
887 if (time_before(jiffies, tg->disptime))
890 throtl_dequeue_tg(tg);
892 nr_disp += throtl_dispatch_tg(tg);
894 if (sq->nr_queued[0] || sq->nr_queued[1])
895 tg_update_disptime(tg);
897 if (nr_disp >= throtl_quantum)
904 /* work function to dispatch throttled bios */
905 void blk_throtl_dispatch_work_fn(struct work_struct *work)
907 struct throtl_data *td = container_of(to_delayed_work(work),
908 struct throtl_data, dispatch_work);
909 struct throtl_service_queue *sq = &td->service_queue;
910 struct request_queue *q = td->queue;
911 unsigned int nr_disp = 0;
912 struct bio_list bio_list_on_stack;
914 struct blk_plug plug;
917 spin_lock_irq(q->queue_lock);
919 bio_list_init(&bio_list_on_stack);
921 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
922 td->nr_queued[READ] + td->nr_queued[WRITE],
923 td->nr_queued[READ], td->nr_queued[WRITE]);
925 nr_disp = throtl_select_dispatch(sq);
928 for (rw = READ; rw <= WRITE; rw++) {
929 bio_list_merge(&bio_list_on_stack, &sq->bio_lists[rw]);
930 bio_list_init(&sq->bio_lists[rw]);
932 throtl_log(sq, "bios disp=%u", nr_disp);
935 throtl_schedule_next_dispatch(td);
937 spin_unlock_irq(q->queue_lock);
940 * If we dispatched some requests, unplug the queue to make sure
944 blk_start_plug(&plug);
945 while((bio = bio_list_pop(&bio_list_on_stack)))
946 generic_make_request(bio);
947 blk_finish_plug(&plug);
951 static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
952 struct blkg_policy_data *pd, int off)
954 struct throtl_grp *tg = pd_to_tg(pd);
955 struct blkg_rwstat rwstat = { }, tmp;
958 for_each_possible_cpu(cpu) {
959 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
961 tmp = blkg_rwstat_read((void *)sc + off);
962 for (i = 0; i < BLKG_RWSTAT_NR; i++)
963 rwstat.cnt[i] += tmp.cnt[i];
966 return __blkg_prfill_rwstat(sf, pd, &rwstat);
969 static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft,
972 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
974 blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl,
979 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
982 struct throtl_grp *tg = pd_to_tg(pd);
983 u64 v = *(u64 *)((void *)tg + off);
987 return __blkg_prfill_u64(sf, pd, v);
990 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
993 struct throtl_grp *tg = pd_to_tg(pd);
994 unsigned int v = *(unsigned int *)((void *)tg + off);
998 return __blkg_prfill_u64(sf, pd, v);
1001 static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft,
1002 struct seq_file *sf)
1004 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64,
1005 &blkcg_policy_throtl, cft->private, false);
1009 static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1010 struct seq_file *sf)
1012 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint,
1013 &blkcg_policy_throtl, cft->private, false);
1017 static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf,
1020 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1021 struct blkg_conf_ctx ctx;
1022 struct throtl_grp *tg;
1023 struct throtl_data *td;
1026 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1030 tg = blkg_to_tg(ctx.blkg);
1031 td = ctx.blkg->q->td;
1037 *(u64 *)((void *)tg + cft->private) = ctx.v;
1039 *(unsigned int *)((void *)tg + cft->private) = ctx.v;
1041 throtl_log(&tg->service_queue,
1042 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1043 tg->bps[READ], tg->bps[WRITE],
1044 tg->iops[READ], tg->iops[WRITE]);
1047 * We're already holding queue_lock and know @tg is valid. Let's
1048 * apply the new config directly.
1050 * Restart the slices for both READ and WRITES. It might happen
1051 * that a group's limit are dropped suddenly and we don't want to
1052 * account recently dispatched IO with new low rate.
1054 throtl_start_new_slice(tg, 0);
1055 throtl_start_new_slice(tg, 1);
1057 if (tg->flags & THROTL_TG_PENDING) {
1058 tg_update_disptime(tg);
1059 throtl_schedule_next_dispatch(td);
1062 blkg_conf_finish(&ctx);
1066 static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft,
1069 return tg_set_conf(cgrp, cft, buf, true);
1072 static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1075 return tg_set_conf(cgrp, cft, buf, false);
1078 static struct cftype throtl_files[] = {
1080 .name = "throttle.read_bps_device",
1081 .private = offsetof(struct throtl_grp, bps[READ]),
1082 .read_seq_string = tg_print_conf_u64,
1083 .write_string = tg_set_conf_u64,
1084 .max_write_len = 256,
1087 .name = "throttle.write_bps_device",
1088 .private = offsetof(struct throtl_grp, bps[WRITE]),
1089 .read_seq_string = tg_print_conf_u64,
1090 .write_string = tg_set_conf_u64,
1091 .max_write_len = 256,
1094 .name = "throttle.read_iops_device",
1095 .private = offsetof(struct throtl_grp, iops[READ]),
1096 .read_seq_string = tg_print_conf_uint,
1097 .write_string = tg_set_conf_uint,
1098 .max_write_len = 256,
1101 .name = "throttle.write_iops_device",
1102 .private = offsetof(struct throtl_grp, iops[WRITE]),
1103 .read_seq_string = tg_print_conf_uint,
1104 .write_string = tg_set_conf_uint,
1105 .max_write_len = 256,
1108 .name = "throttle.io_service_bytes",
1109 .private = offsetof(struct tg_stats_cpu, service_bytes),
1110 .read_seq_string = tg_print_cpu_rwstat,
1113 .name = "throttle.io_serviced",
1114 .private = offsetof(struct tg_stats_cpu, serviced),
1115 .read_seq_string = tg_print_cpu_rwstat,
1120 static void throtl_shutdown_wq(struct request_queue *q)
1122 struct throtl_data *td = q->td;
1124 cancel_delayed_work_sync(&td->dispatch_work);
1127 static struct blkcg_policy blkcg_policy_throtl = {
1128 .pd_size = sizeof(struct throtl_grp),
1129 .cftypes = throtl_files,
1131 .pd_init_fn = throtl_pd_init,
1132 .pd_exit_fn = throtl_pd_exit,
1133 .pd_reset_stats_fn = throtl_pd_reset_stats,
1136 bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
1138 struct throtl_data *td = q->td;
1139 struct throtl_grp *tg;
1140 struct throtl_service_queue *sq;
1141 bool rw = bio_data_dir(bio);
1142 struct blkcg *blkcg;
1143 bool throttled = false;
1145 /* see throtl_charge_bio() */
1146 if (bio->bi_rw & REQ_THROTTLED)
1150 * A throtl_grp pointer retrieved under rcu can be used to access
1151 * basic fields like stats and io rates. If a group has no rules,
1152 * just update the dispatch stats in lockless manner and return.
1155 blkcg = bio_blkcg(bio);
1156 tg = throtl_lookup_tg(td, blkcg);
1158 if (tg_no_rule_group(tg, rw)) {
1159 throtl_update_dispatch_stats(tg_to_blkg(tg),
1160 bio->bi_size, bio->bi_rw);
1161 goto out_unlock_rcu;
1166 * Either group has not been allocated yet or it is not an unlimited
1169 spin_lock_irq(q->queue_lock);
1170 tg = throtl_lookup_create_tg(td, blkcg);
1174 sq = &tg->service_queue;
1176 /* throtl is FIFO - if other bios are already queued, should queue */
1177 if (sq->nr_queued[rw])
1180 /* Bio is with-in rate limit of group */
1181 if (tg_may_dispatch(tg, bio, NULL)) {
1182 throtl_charge_bio(tg, bio);
1185 * We need to trim slice even when bios are not being queued
1186 * otherwise it might happen that a bio is not queued for
1187 * a long time and slice keeps on extending and trim is not
1188 * called for a long time. Now if limits are reduced suddenly
1189 * we take into account all the IO dispatched so far at new
1190 * low rate and * newly queued IO gets a really long dispatch
1193 * So keep on trimming slice even if bio is not queued.
1195 throtl_trim_slice(tg, rw);
1200 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1201 rw == READ ? 'R' : 'W',
1202 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1203 tg->io_disp[rw], tg->iops[rw],
1204 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1206 bio_associate_current(bio);
1207 throtl_add_bio_tg(bio, tg);
1210 /* update @tg's dispatch time if @tg was empty before @bio */
1211 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1212 tg_update_disptime(tg);
1213 throtl_schedule_next_dispatch(td);
1217 spin_unlock_irq(q->queue_lock);
1222 * As multiple blk-throtls may stack in the same issue path, we
1223 * don't want bios to leave with the flag set. Clear the flag if
1227 bio->bi_rw &= ~REQ_THROTTLED;
1232 * blk_throtl_drain - drain throttled bios
1233 * @q: request_queue to drain throttled bios for
1235 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1237 void blk_throtl_drain(struct request_queue *q)
1238 __releases(q->queue_lock) __acquires(q->queue_lock)
1240 struct throtl_data *td = q->td;
1241 struct throtl_service_queue *parent_sq = &td->service_queue;
1242 struct throtl_grp *tg;
1246 queue_lockdep_assert_held(q);
1248 while ((tg = throtl_rb_first(parent_sq))) {
1249 struct throtl_service_queue *sq = &tg->service_queue;
1251 throtl_dequeue_tg(tg);
1253 while ((bio = bio_list_peek(&sq->bio_lists[READ])))
1254 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1255 while ((bio = bio_list_peek(&sq->bio_lists[WRITE])))
1256 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1258 spin_unlock_irq(q->queue_lock);
1260 for (rw = READ; rw <= WRITE; rw++)
1261 while ((bio = bio_list_pop(&parent_sq->bio_lists[rw])))
1262 generic_make_request(bio);
1264 spin_lock_irq(q->queue_lock);
1267 int blk_throtl_init(struct request_queue *q)
1269 struct throtl_data *td;
1272 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1276 INIT_DELAYED_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1277 throtl_service_queue_init(&td->service_queue, NULL);
1282 /* activate policy */
1283 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1289 void blk_throtl_exit(struct request_queue *q)
1292 throtl_shutdown_wq(q);
1293 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1297 static int __init throtl_init(void)
1299 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1300 if (!kthrotld_workqueue)
1301 panic("Failed to create kthrotld\n");
1303 return blkcg_policy_register(&blkcg_policy_throtl);
1306 module_init(throtl_init);