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;
30 * To implement hierarchical throttling, throtl_grps form a tree and bios
31 * are dispatched upwards level by level until they reach the top and get
32 * issued. When dispatching bios from the children and local group at each
33 * level, if the bios are dispatched into a single bio_list, there's a risk
34 * of a local or child group which can queue many bios at once filling up
35 * the list starving others.
37 * To avoid such starvation, dispatched bios are queued separately
38 * according to where they came from. When they are again dispatched to
39 * the parent, they're popped in round-robin order so that no single source
40 * hogs the dispatch window.
42 * throtl_qnode is used to keep the queued bios separated by their sources.
43 * Bios are queued to throtl_qnode which in turn is queued to
44 * throtl_service_queue and then dispatched in round-robin order.
46 * It's also used to track the reference counts on blkg's. A qnode always
47 * belongs to a throtl_grp and gets queued on itself or the parent, so
48 * incrementing the reference of the associated throtl_grp when a qnode is
49 * queued and decrementing when dequeued is enough to keep the whole blkg
50 * tree pinned while bios are in flight.
53 struct list_head node; /* service_queue->queued[] */
54 struct bio_list bios; /* queued bios */
55 struct throtl_grp *tg; /* tg this qnode belongs to */
58 struct throtl_service_queue {
59 struct throtl_service_queue *parent_sq; /* the parent service_queue */
62 * Bios queued directly to this service_queue or dispatched from
63 * children throtl_grp's.
65 struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */
66 unsigned int nr_queued[2]; /* number of queued bios */
69 * RB tree of active children throtl_grp's, which are sorted by
72 struct rb_root pending_tree; /* RB tree of active tgs */
73 struct rb_node *first_pending; /* first node in the tree */
74 unsigned int nr_pending; /* # queued in the tree */
75 unsigned long first_pending_disptime; /* disptime of the first tg */
76 struct timer_list pending_timer; /* fires on first_pending_disptime */
80 THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */
81 THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */
84 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
86 /* Per-cpu group stats */
88 /* total bytes transferred */
89 struct blkg_rwstat service_bytes;
90 /* total IOs serviced, post merge */
91 struct blkg_rwstat serviced;
95 /* must be the first member */
96 struct blkg_policy_data pd;
98 /* active throtl group service_queue member */
99 struct rb_node rb_node;
101 /* throtl_data this group belongs to */
102 struct throtl_data *td;
104 /* this group's service queue */
105 struct throtl_service_queue service_queue;
108 * qnode_on_self is used when bios are directly queued to this
109 * throtl_grp so that local bios compete fairly with bios
110 * dispatched from children. qnode_on_parent is used when bios are
111 * dispatched from this throtl_grp into its parent and will compete
112 * with the sibling qnode_on_parents and the parent's
115 struct throtl_qnode qnode_on_self[2];
116 struct throtl_qnode qnode_on_parent[2];
119 * Dispatch time in jiffies. This is the estimated time when group
120 * will unthrottle and is ready to dispatch more bio. It is used as
121 * key to sort active groups in service tree.
123 unsigned long disptime;
127 /* bytes per second rate limits */
131 unsigned int iops[2];
133 /* Number of bytes disptached in current slice */
134 uint64_t bytes_disp[2];
135 /* Number of bio's dispatched in current slice */
136 unsigned int io_disp[2];
138 /* When did we start a new slice */
139 unsigned long slice_start[2];
140 unsigned long slice_end[2];
142 /* Per cpu stats pointer */
143 struct tg_stats_cpu __percpu *stats_cpu;
145 /* List of tgs waiting for per cpu stats memory to be allocated */
146 struct list_head stats_alloc_node;
151 /* service tree for active throtl groups */
152 struct throtl_service_queue service_queue;
154 struct request_queue *queue;
156 /* Total Number of queued bios on READ and WRITE lists */
157 unsigned int nr_queued[2];
160 * number of total undestroyed groups
162 unsigned int nr_undestroyed_grps;
164 /* Work for dispatching throttled bios */
165 struct work_struct dispatch_work;
168 /* list and work item to allocate percpu group stats */
169 static DEFINE_SPINLOCK(tg_stats_alloc_lock);
170 static LIST_HEAD(tg_stats_alloc_list);
172 static void tg_stats_alloc_fn(struct work_struct *);
173 static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn);
175 static void throtl_pending_timer_fn(unsigned long arg);
177 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
179 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
182 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
184 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
187 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
189 return pd_to_blkg(&tg->pd);
192 static inline struct throtl_grp *td_root_tg(struct throtl_data *td)
194 return blkg_to_tg(td->queue->root_blkg);
198 * sq_to_tg - return the throl_grp the specified service queue belongs to
199 * @sq: the throtl_service_queue of interest
201 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
202 * embedded in throtl_data, %NULL is returned.
204 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
206 if (sq && sq->parent_sq)
207 return container_of(sq, struct throtl_grp, service_queue);
213 * sq_to_td - return throtl_data the specified service queue belongs to
214 * @sq: the throtl_service_queue of interest
216 * A service_queue can be embeded in either a throtl_grp or throtl_data.
217 * Determine the associated throtl_data accordingly and return it.
219 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
221 struct throtl_grp *tg = sq_to_tg(sq);
226 return container_of(sq, struct throtl_data, service_queue);
230 * throtl_log - log debug message via blktrace
231 * @sq: the service_queue being reported
232 * @fmt: printf format string
235 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
236 * throtl_grp; otherwise, just "throtl".
238 * TODO: this should be made a function and name formatting should happen
239 * after testing whether blktrace is enabled.
241 #define throtl_log(sq, fmt, args...) do { \
242 struct throtl_grp *__tg = sq_to_tg((sq)); \
243 struct throtl_data *__td = sq_to_td((sq)); \
249 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
250 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
252 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
257 * Worker for allocating per cpu stat for tgs. This is scheduled on the
258 * system_wq once there are some groups on the alloc_list waiting for
261 static void tg_stats_alloc_fn(struct work_struct *work)
263 static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */
264 struct delayed_work *dwork = to_delayed_work(work);
269 stats_cpu = alloc_percpu(struct tg_stats_cpu);
271 /* allocation failed, try again after some time */
272 schedule_delayed_work(dwork, msecs_to_jiffies(10));
277 spin_lock_irq(&tg_stats_alloc_lock);
279 if (!list_empty(&tg_stats_alloc_list)) {
280 struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list,
283 swap(tg->stats_cpu, stats_cpu);
284 list_del_init(&tg->stats_alloc_node);
287 empty = list_empty(&tg_stats_alloc_list);
288 spin_unlock_irq(&tg_stats_alloc_lock);
293 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
295 INIT_LIST_HEAD(&qn->node);
296 bio_list_init(&qn->bios);
301 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
302 * @bio: bio being added
303 * @qn: qnode to add bio to
304 * @queued: the service_queue->queued[] list @qn belongs to
306 * Add @bio to @qn and put @qn on @queued if it's not already on.
307 * @qn->tg's reference count is bumped when @qn is activated. See the
308 * comment on top of throtl_qnode definition for details.
310 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
311 struct list_head *queued)
313 bio_list_add(&qn->bios, bio);
314 if (list_empty(&qn->node)) {
315 list_add_tail(&qn->node, queued);
316 blkg_get(tg_to_blkg(qn->tg));
321 * throtl_peek_queued - peek the first bio on a qnode list
322 * @queued: the qnode list to peek
324 static struct bio *throtl_peek_queued(struct list_head *queued)
326 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
329 if (list_empty(queued))
332 bio = bio_list_peek(&qn->bios);
338 * throtl_pop_queued - pop the first bio form a qnode list
339 * @queued: the qnode list to pop a bio from
340 * @tg_to_put: optional out argument for throtl_grp to put
342 * Pop the first bio from the qnode list @queued. After popping, the first
343 * qnode is removed from @queued if empty or moved to the end of @queued so
344 * that the popping order is round-robin.
346 * When the first qnode is removed, its associated throtl_grp should be put
347 * too. If @tg_to_put is NULL, this function automatically puts it;
348 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
349 * responsible for putting it.
351 static struct bio *throtl_pop_queued(struct list_head *queued,
352 struct throtl_grp **tg_to_put)
354 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
357 if (list_empty(queued))
360 bio = bio_list_pop(&qn->bios);
363 if (bio_list_empty(&qn->bios)) {
364 list_del_init(&qn->node);
368 blkg_put(tg_to_blkg(qn->tg));
370 list_move_tail(&qn->node, queued);
376 /* init a service_queue, assumes the caller zeroed it */
377 static void throtl_service_queue_init(struct throtl_service_queue *sq,
378 struct throtl_service_queue *parent_sq)
380 INIT_LIST_HEAD(&sq->queued[0]);
381 INIT_LIST_HEAD(&sq->queued[1]);
382 sq->pending_tree = RB_ROOT;
383 sq->parent_sq = parent_sq;
384 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
388 static void throtl_service_queue_exit(struct throtl_service_queue *sq)
390 del_timer_sync(&sq->pending_timer);
393 static void throtl_pd_init(struct blkcg_gq *blkg)
395 struct throtl_grp *tg = blkg_to_tg(blkg);
396 struct throtl_data *td = blkg->q->td;
400 throtl_service_queue_init(&tg->service_queue, &td->service_queue);
401 for (rw = READ; rw <= WRITE; rw++) {
402 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
403 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
406 RB_CLEAR_NODE(&tg->rb_node);
412 tg->iops[WRITE] = -1;
415 * Ugh... We need to perform per-cpu allocation for tg->stats_cpu
416 * but percpu allocator can't be called from IO path. Queue tg on
417 * tg_stats_alloc_list and allocate from work item.
419 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
420 list_add(&tg->stats_alloc_node, &tg_stats_alloc_list);
421 schedule_delayed_work(&tg_stats_alloc_work, 0);
422 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
425 static void throtl_pd_exit(struct blkcg_gq *blkg)
427 struct throtl_grp *tg = blkg_to_tg(blkg);
430 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
431 list_del_init(&tg->stats_alloc_node);
432 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
434 free_percpu(tg->stats_cpu);
436 throtl_service_queue_exit(&tg->service_queue);
439 static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
441 struct throtl_grp *tg = blkg_to_tg(blkg);
444 if (tg->stats_cpu == NULL)
447 for_each_possible_cpu(cpu) {
448 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
450 blkg_rwstat_reset(&sc->service_bytes);
451 blkg_rwstat_reset(&sc->serviced);
455 static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
459 * This is the common case when there are no blkcgs. Avoid lookup
462 if (blkcg == &blkcg_root)
463 return td_root_tg(td);
465 return blkg_to_tg(blkg_lookup(blkcg, td->queue));
468 static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
471 struct request_queue *q = td->queue;
472 struct throtl_grp *tg = NULL;
475 * This is the common case when there are no blkcgs. Avoid lookup
478 if (blkcg == &blkcg_root) {
481 struct blkcg_gq *blkg;
483 blkg = blkg_lookup_create(blkcg, q);
485 /* if %NULL and @q is alive, fall back to root_tg */
487 tg = blkg_to_tg(blkg);
488 else if (!blk_queue_dying(q))
495 static struct throtl_grp *
496 throtl_rb_first(struct throtl_service_queue *parent_sq)
498 /* Service tree is empty */
499 if (!parent_sq->nr_pending)
502 if (!parent_sq->first_pending)
503 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
505 if (parent_sq->first_pending)
506 return rb_entry_tg(parent_sq->first_pending);
511 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
517 static void throtl_rb_erase(struct rb_node *n,
518 struct throtl_service_queue *parent_sq)
520 if (parent_sq->first_pending == n)
521 parent_sq->first_pending = NULL;
522 rb_erase_init(n, &parent_sq->pending_tree);
523 --parent_sq->nr_pending;
526 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
528 struct throtl_grp *tg;
530 tg = throtl_rb_first(parent_sq);
534 parent_sq->first_pending_disptime = tg->disptime;
537 static void tg_service_queue_add(struct throtl_grp *tg)
539 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
540 struct rb_node **node = &parent_sq->pending_tree.rb_node;
541 struct rb_node *parent = NULL;
542 struct throtl_grp *__tg;
543 unsigned long key = tg->disptime;
546 while (*node != NULL) {
548 __tg = rb_entry_tg(parent);
550 if (time_before(key, __tg->disptime))
551 node = &parent->rb_left;
553 node = &parent->rb_right;
559 parent_sq->first_pending = &tg->rb_node;
561 rb_link_node(&tg->rb_node, parent, node);
562 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
565 static void __throtl_enqueue_tg(struct throtl_grp *tg)
567 tg_service_queue_add(tg);
568 tg->flags |= THROTL_TG_PENDING;
569 tg->service_queue.parent_sq->nr_pending++;
572 static void throtl_enqueue_tg(struct throtl_grp *tg)
574 if (!(tg->flags & THROTL_TG_PENDING))
575 __throtl_enqueue_tg(tg);
578 static void __throtl_dequeue_tg(struct throtl_grp *tg)
580 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
581 tg->flags &= ~THROTL_TG_PENDING;
584 static void throtl_dequeue_tg(struct throtl_grp *tg)
586 if (tg->flags & THROTL_TG_PENDING)
587 __throtl_dequeue_tg(tg);
590 /* Call with queue lock held */
591 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
592 unsigned long expires)
594 mod_timer(&sq->pending_timer, expires);
595 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
596 expires - jiffies, jiffies);
600 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
601 * @sq: the service_queue to schedule dispatch for
602 * @force: force scheduling
604 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
605 * dispatch time of the first pending child. Returns %true if either timer
606 * is armed or there's no pending child left. %false if the current
607 * dispatch window is still open and the caller should continue
610 * If @force is %true, the dispatch timer is always scheduled and this
611 * function is guaranteed to return %true. This is to be used when the
612 * caller can't dispatch itself and needs to invoke pending_timer
613 * unconditionally. Note that forced scheduling is likely to induce short
614 * delay before dispatch starts even if @sq->first_pending_disptime is not
615 * in the future and thus shouldn't be used in hot paths.
617 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
620 /* any pending children left? */
624 update_min_dispatch_time(sq);
626 /* is the next dispatch time in the future? */
627 if (force || time_after(sq->first_pending_disptime, jiffies)) {
628 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
632 /* tell the caller to continue dispatching */
636 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
637 bool rw, unsigned long start)
639 tg->bytes_disp[rw] = 0;
643 * Previous slice has expired. We must have trimmed it after last
644 * bio dispatch. That means since start of last slice, we never used
645 * that bandwidth. Do try to make use of that bandwidth while giving
648 if (time_after_eq(start, tg->slice_start[rw]))
649 tg->slice_start[rw] = start;
651 tg->slice_end[rw] = jiffies + throtl_slice;
652 throtl_log(&tg->service_queue,
653 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
654 rw == READ ? 'R' : 'W', tg->slice_start[rw],
655 tg->slice_end[rw], jiffies);
658 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
660 tg->bytes_disp[rw] = 0;
662 tg->slice_start[rw] = jiffies;
663 tg->slice_end[rw] = jiffies + throtl_slice;
664 throtl_log(&tg->service_queue,
665 "[%c] new slice start=%lu end=%lu jiffies=%lu",
666 rw == READ ? 'R' : 'W', tg->slice_start[rw],
667 tg->slice_end[rw], jiffies);
670 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
671 unsigned long jiffy_end)
673 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
676 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
677 unsigned long jiffy_end)
679 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
680 throtl_log(&tg->service_queue,
681 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
682 rw == READ ? 'R' : 'W', tg->slice_start[rw],
683 tg->slice_end[rw], jiffies);
686 /* Determine if previously allocated or extended slice is complete or not */
687 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
689 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
695 /* Trim the used slices and adjust slice start accordingly */
696 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
698 unsigned long nr_slices, time_elapsed, io_trim;
701 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
704 * If bps are unlimited (-1), then time slice don't get
705 * renewed. Don't try to trim the slice if slice is used. A new
706 * slice will start when appropriate.
708 if (throtl_slice_used(tg, rw))
712 * A bio has been dispatched. Also adjust slice_end. It might happen
713 * that initially cgroup limit was very low resulting in high
714 * slice_end, but later limit was bumped up and bio was dispached
715 * sooner, then we need to reduce slice_end. A high bogus slice_end
716 * is bad because it does not allow new slice to start.
719 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
721 time_elapsed = jiffies - tg->slice_start[rw];
723 nr_slices = time_elapsed / throtl_slice;
727 tmp = tg->bps[rw] * throtl_slice * nr_slices;
731 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
733 if (!bytes_trim && !io_trim)
736 if (tg->bytes_disp[rw] >= bytes_trim)
737 tg->bytes_disp[rw] -= bytes_trim;
739 tg->bytes_disp[rw] = 0;
741 if (tg->io_disp[rw] >= io_trim)
742 tg->io_disp[rw] -= io_trim;
746 tg->slice_start[rw] += nr_slices * throtl_slice;
748 throtl_log(&tg->service_queue,
749 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
750 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
751 tg->slice_start[rw], tg->slice_end[rw], jiffies);
754 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
757 bool rw = bio_data_dir(bio);
758 unsigned int io_allowed;
759 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
762 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
764 /* Slice has just started. Consider one slice interval */
766 jiffy_elapsed_rnd = throtl_slice;
768 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
771 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
772 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
773 * will allow dispatch after 1 second and after that slice should
777 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
781 io_allowed = UINT_MAX;
785 if (tg->io_disp[rw] + 1 <= io_allowed) {
791 /* Calc approx time to dispatch */
792 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
794 if (jiffy_wait > jiffy_elapsed)
795 jiffy_wait = jiffy_wait - jiffy_elapsed;
804 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
807 bool rw = bio_data_dir(bio);
808 u64 bytes_allowed, extra_bytes, tmp;
809 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
811 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
813 /* Slice has just started. Consider one slice interval */
815 jiffy_elapsed_rnd = throtl_slice;
817 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
819 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
823 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
829 /* Calc approx time to dispatch */
830 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
831 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
837 * This wait time is without taking into consideration the rounding
838 * up we did. Add that time also.
840 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
846 static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
847 if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
853 * Returns whether one can dispatch a bio or not. Also returns approx number
854 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
856 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
859 bool rw = bio_data_dir(bio);
860 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
863 * Currently whole state machine of group depends on first bio
864 * queued in the group bio list. So one should not be calling
865 * this function with a different bio if there are other bios
868 BUG_ON(tg->service_queue.nr_queued[rw] &&
869 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
871 /* If tg->bps = -1, then BW is unlimited */
872 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
879 * If previous slice expired, start a new one otherwise renew/extend
880 * existing slice to make sure it is at least throtl_slice interval
883 if (throtl_slice_used(tg, rw))
884 throtl_start_new_slice(tg, rw);
886 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
887 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
890 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
891 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
897 max_wait = max(bps_wait, iops_wait);
902 if (time_before(tg->slice_end[rw], jiffies + max_wait))
903 throtl_extend_slice(tg, rw, jiffies + max_wait);
908 static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
911 struct throtl_grp *tg = blkg_to_tg(blkg);
912 struct tg_stats_cpu *stats_cpu;
915 /* If per cpu stats are not allocated yet, don't do any accounting. */
916 if (tg->stats_cpu == NULL)
920 * Disabling interrupts to provide mutual exclusion between two
921 * writes on same cpu. It probably is not needed for 64bit. Not
922 * optimizing that case yet.
924 local_irq_save(flags);
926 stats_cpu = this_cpu_ptr(tg->stats_cpu);
928 blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
929 blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
931 local_irq_restore(flags);
934 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
936 bool rw = bio_data_dir(bio);
938 /* Charge the bio to the group */
939 tg->bytes_disp[rw] += bio->bi_size;
943 * REQ_THROTTLED is used to prevent the same bio to be throttled
944 * more than once as a throttled bio will go through blk-throtl the
945 * second time when it eventually gets issued. Set it when a bio
946 * is being charged to a tg.
948 * Dispatch stats aren't recursive and each @bio should only be
949 * accounted by the @tg it was originally associated with. Let's
950 * update the stats when setting REQ_THROTTLED for the first time
951 * which is guaranteed to be for the @bio's original tg.
953 if (!(bio->bi_rw & REQ_THROTTLED)) {
954 bio->bi_rw |= REQ_THROTTLED;
955 throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size,
961 * throtl_add_bio_tg - add a bio to the specified throtl_grp
964 * @tg: the target throtl_grp
966 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
967 * tg->qnode_on_self[] is used.
969 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
970 struct throtl_grp *tg)
972 struct throtl_service_queue *sq = &tg->service_queue;
973 bool rw = bio_data_dir(bio);
976 qn = &tg->qnode_on_self[rw];
979 * If @tg doesn't currently have any bios queued in the same
980 * direction, queueing @bio can change when @tg should be
981 * dispatched. Mark that @tg was empty. This is automatically
982 * cleaered on the next tg_update_disptime().
984 if (!sq->nr_queued[rw])
985 tg->flags |= THROTL_TG_WAS_EMPTY;
987 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
990 throtl_enqueue_tg(tg);
993 static void tg_update_disptime(struct throtl_grp *tg)
995 struct throtl_service_queue *sq = &tg->service_queue;
996 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
999 if ((bio = throtl_peek_queued(&sq->queued[READ])))
1000 tg_may_dispatch(tg, bio, &read_wait);
1002 if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1003 tg_may_dispatch(tg, bio, &write_wait);
1005 min_wait = min(read_wait, write_wait);
1006 disptime = jiffies + min_wait;
1008 /* Update dispatch time */
1009 throtl_dequeue_tg(tg);
1010 tg->disptime = disptime;
1011 throtl_enqueue_tg(tg);
1013 /* see throtl_add_bio_tg() */
1014 tg->flags &= ~THROTL_TG_WAS_EMPTY;
1017 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1018 struct throtl_grp *parent_tg, bool rw)
1020 if (throtl_slice_used(parent_tg, rw)) {
1021 throtl_start_new_slice_with_credit(parent_tg, rw,
1022 child_tg->slice_start[rw]);
1027 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1029 struct throtl_service_queue *sq = &tg->service_queue;
1030 struct throtl_service_queue *parent_sq = sq->parent_sq;
1031 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1032 struct throtl_grp *tg_to_put = NULL;
1036 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1037 * from @tg may put its reference and @parent_sq might end up
1038 * getting released prematurely. Remember the tg to put and put it
1039 * after @bio is transferred to @parent_sq.
1041 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1042 sq->nr_queued[rw]--;
1044 throtl_charge_bio(tg, bio);
1047 * If our parent is another tg, we just need to transfer @bio to
1048 * the parent using throtl_add_bio_tg(). If our parent is
1049 * @td->service_queue, @bio is ready to be issued. Put it on its
1050 * bio_lists[] and decrease total number queued. The caller is
1051 * responsible for issuing these bios.
1054 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1055 start_parent_slice_with_credit(tg, parent_tg, rw);
1057 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1058 &parent_sq->queued[rw]);
1059 BUG_ON(tg->td->nr_queued[rw] <= 0);
1060 tg->td->nr_queued[rw]--;
1063 throtl_trim_slice(tg, rw);
1066 blkg_put(tg_to_blkg(tg_to_put));
1069 static int throtl_dispatch_tg(struct throtl_grp *tg)
1071 struct throtl_service_queue *sq = &tg->service_queue;
1072 unsigned int nr_reads = 0, nr_writes = 0;
1073 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1074 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1077 /* Try to dispatch 75% READS and 25% WRITES */
1079 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1080 tg_may_dispatch(tg, bio, NULL)) {
1082 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1085 if (nr_reads >= max_nr_reads)
1089 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1090 tg_may_dispatch(tg, bio, NULL)) {
1092 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1095 if (nr_writes >= max_nr_writes)
1099 return nr_reads + nr_writes;
1102 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1104 unsigned int nr_disp = 0;
1107 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1108 struct throtl_service_queue *sq = &tg->service_queue;
1113 if (time_before(jiffies, tg->disptime))
1116 throtl_dequeue_tg(tg);
1118 nr_disp += throtl_dispatch_tg(tg);
1120 if (sq->nr_queued[0] || sq->nr_queued[1])
1121 tg_update_disptime(tg);
1123 if (nr_disp >= throtl_quantum)
1131 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1132 * @arg: the throtl_service_queue being serviced
1134 * This timer is armed when a child throtl_grp with active bio's become
1135 * pending and queued on the service_queue's pending_tree and expires when
1136 * the first child throtl_grp should be dispatched. This function
1137 * dispatches bio's from the children throtl_grps to the parent
1140 * If the parent's parent is another throtl_grp, dispatching is propagated
1141 * by either arming its pending_timer or repeating dispatch directly. If
1142 * the top-level service_tree is reached, throtl_data->dispatch_work is
1143 * kicked so that the ready bio's are issued.
1145 static void throtl_pending_timer_fn(unsigned long arg)
1147 struct throtl_service_queue *sq = (void *)arg;
1148 struct throtl_grp *tg = sq_to_tg(sq);
1149 struct throtl_data *td = sq_to_td(sq);
1150 struct request_queue *q = td->queue;
1151 struct throtl_service_queue *parent_sq;
1155 spin_lock_irq(q->queue_lock);
1157 parent_sq = sq->parent_sq;
1161 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1162 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1163 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1165 ret = throtl_select_dispatch(sq);
1167 throtl_log(sq, "bios disp=%u", ret);
1171 if (throtl_schedule_next_dispatch(sq, false))
1174 /* this dispatch windows is still open, relax and repeat */
1175 spin_unlock_irq(q->queue_lock);
1177 spin_lock_irq(q->queue_lock);
1184 /* @parent_sq is another throl_grp, propagate dispatch */
1185 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1186 tg_update_disptime(tg);
1187 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1188 /* window is already open, repeat dispatching */
1195 /* reached the top-level, queue issueing */
1196 queue_work(kthrotld_workqueue, &td->dispatch_work);
1199 spin_unlock_irq(q->queue_lock);
1203 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1204 * @work: work item being executed
1206 * This function is queued for execution when bio's reach the bio_lists[]
1207 * of throtl_data->service_queue. Those bio's are ready and issued by this
1210 void blk_throtl_dispatch_work_fn(struct work_struct *work)
1212 struct throtl_data *td = container_of(work, struct throtl_data,
1214 struct throtl_service_queue *td_sq = &td->service_queue;
1215 struct request_queue *q = td->queue;
1216 struct bio_list bio_list_on_stack;
1218 struct blk_plug plug;
1221 bio_list_init(&bio_list_on_stack);
1223 spin_lock_irq(q->queue_lock);
1224 for (rw = READ; rw <= WRITE; rw++)
1225 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1226 bio_list_add(&bio_list_on_stack, bio);
1227 spin_unlock_irq(q->queue_lock);
1229 if (!bio_list_empty(&bio_list_on_stack)) {
1230 blk_start_plug(&plug);
1231 while((bio = bio_list_pop(&bio_list_on_stack)))
1232 generic_make_request(bio);
1233 blk_finish_plug(&plug);
1237 static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
1238 struct blkg_policy_data *pd, int off)
1240 struct throtl_grp *tg = pd_to_tg(pd);
1241 struct blkg_rwstat rwstat = { }, tmp;
1244 for_each_possible_cpu(cpu) {
1245 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
1247 tmp = blkg_rwstat_read((void *)sc + off);
1248 for (i = 0; i < BLKG_RWSTAT_NR; i++)
1249 rwstat.cnt[i] += tmp.cnt[i];
1252 return __blkg_prfill_rwstat(sf, pd, &rwstat);
1255 static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft,
1256 struct seq_file *sf)
1258 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1260 blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl,
1261 cft->private, true);
1265 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1268 struct throtl_grp *tg = pd_to_tg(pd);
1269 u64 v = *(u64 *)((void *)tg + off);
1273 return __blkg_prfill_u64(sf, pd, v);
1276 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1279 struct throtl_grp *tg = pd_to_tg(pd);
1280 unsigned int v = *(unsigned int *)((void *)tg + off);
1284 return __blkg_prfill_u64(sf, pd, v);
1287 static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft,
1288 struct seq_file *sf)
1290 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64,
1291 &blkcg_policy_throtl, cft->private, false);
1295 static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1296 struct seq_file *sf)
1298 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint,
1299 &blkcg_policy_throtl, cft->private, false);
1303 static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf,
1306 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1307 struct blkg_conf_ctx ctx;
1308 struct throtl_grp *tg;
1309 struct throtl_service_queue *sq;
1312 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1316 tg = blkg_to_tg(ctx.blkg);
1317 sq = &tg->service_queue;
1323 *(u64 *)((void *)tg + cft->private) = ctx.v;
1325 *(unsigned int *)((void *)tg + cft->private) = ctx.v;
1327 throtl_log(&tg->service_queue,
1328 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1329 tg->bps[READ], tg->bps[WRITE],
1330 tg->iops[READ], tg->iops[WRITE]);
1333 * We're already holding queue_lock and know @tg is valid. Let's
1334 * apply the new config directly.
1336 * Restart the slices for both READ and WRITES. It might happen
1337 * that a group's limit are dropped suddenly and we don't want to
1338 * account recently dispatched IO with new low rate.
1340 throtl_start_new_slice(tg, 0);
1341 throtl_start_new_slice(tg, 1);
1343 if (tg->flags & THROTL_TG_PENDING) {
1344 tg_update_disptime(tg);
1345 throtl_schedule_next_dispatch(sq->parent_sq, true);
1348 blkg_conf_finish(&ctx);
1352 static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft,
1355 return tg_set_conf(cgrp, cft, buf, true);
1358 static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1361 return tg_set_conf(cgrp, cft, buf, false);
1364 static struct cftype throtl_files[] = {
1366 .name = "throttle.read_bps_device",
1367 .private = offsetof(struct throtl_grp, bps[READ]),
1368 .read_seq_string = tg_print_conf_u64,
1369 .write_string = tg_set_conf_u64,
1370 .max_write_len = 256,
1373 .name = "throttle.write_bps_device",
1374 .private = offsetof(struct throtl_grp, bps[WRITE]),
1375 .read_seq_string = tg_print_conf_u64,
1376 .write_string = tg_set_conf_u64,
1377 .max_write_len = 256,
1380 .name = "throttle.read_iops_device",
1381 .private = offsetof(struct throtl_grp, iops[READ]),
1382 .read_seq_string = tg_print_conf_uint,
1383 .write_string = tg_set_conf_uint,
1384 .max_write_len = 256,
1387 .name = "throttle.write_iops_device",
1388 .private = offsetof(struct throtl_grp, iops[WRITE]),
1389 .read_seq_string = tg_print_conf_uint,
1390 .write_string = tg_set_conf_uint,
1391 .max_write_len = 256,
1394 .name = "throttle.io_service_bytes",
1395 .private = offsetof(struct tg_stats_cpu, service_bytes),
1396 .read_seq_string = tg_print_cpu_rwstat,
1399 .name = "throttle.io_serviced",
1400 .private = offsetof(struct tg_stats_cpu, serviced),
1401 .read_seq_string = tg_print_cpu_rwstat,
1406 static void throtl_shutdown_wq(struct request_queue *q)
1408 struct throtl_data *td = q->td;
1410 cancel_work_sync(&td->dispatch_work);
1413 static struct blkcg_policy blkcg_policy_throtl = {
1414 .pd_size = sizeof(struct throtl_grp),
1415 .cftypes = throtl_files,
1417 .pd_init_fn = throtl_pd_init,
1418 .pd_exit_fn = throtl_pd_exit,
1419 .pd_reset_stats_fn = throtl_pd_reset_stats,
1422 bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
1424 struct throtl_data *td = q->td;
1425 struct throtl_qnode *qn = NULL;
1426 struct throtl_grp *tg;
1427 struct throtl_service_queue *sq;
1428 bool rw = bio_data_dir(bio);
1429 struct blkcg *blkcg;
1430 bool throttled = false;
1432 /* see throtl_charge_bio() */
1433 if (bio->bi_rw & REQ_THROTTLED)
1437 * A throtl_grp pointer retrieved under rcu can be used to access
1438 * basic fields like stats and io rates. If a group has no rules,
1439 * just update the dispatch stats in lockless manner and return.
1442 blkcg = bio_blkcg(bio);
1443 tg = throtl_lookup_tg(td, blkcg);
1445 if (tg_no_rule_group(tg, rw)) {
1446 throtl_update_dispatch_stats(tg_to_blkg(tg),
1447 bio->bi_size, bio->bi_rw);
1448 goto out_unlock_rcu;
1453 * Either group has not been allocated yet or it is not an unlimited
1456 spin_lock_irq(q->queue_lock);
1457 tg = throtl_lookup_create_tg(td, blkcg);
1461 sq = &tg->service_queue;
1464 /* throtl is FIFO - if bios are already queued, should queue */
1465 if (sq->nr_queued[rw])
1468 /* if above limits, break to queue */
1469 if (!tg_may_dispatch(tg, bio, NULL))
1472 /* within limits, let's charge and dispatch directly */
1473 throtl_charge_bio(tg, bio);
1476 * We need to trim slice even when bios are not being queued
1477 * otherwise it might happen that a bio is not queued for
1478 * a long time and slice keeps on extending and trim is not
1479 * called for a long time. Now if limits are reduced suddenly
1480 * we take into account all the IO dispatched so far at new
1481 * low rate and * newly queued IO gets a really long dispatch
1484 * So keep on trimming slice even if bio is not queued.
1486 throtl_trim_slice(tg, rw);
1489 * @bio passed through this layer without being throttled.
1490 * Climb up the ladder. If we''re already at the top, it
1491 * can be executed directly.
1493 qn = &tg->qnode_on_parent[rw];
1500 /* out-of-limit, queue to @tg */
1501 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1502 rw == READ ? 'R' : 'W',
1503 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1504 tg->io_disp[rw], tg->iops[rw],
1505 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1507 bio_associate_current(bio);
1508 tg->td->nr_queued[rw]++;
1509 throtl_add_bio_tg(bio, qn, tg);
1513 * Update @tg's dispatch time and force schedule dispatch if @tg
1514 * was empty before @bio. The forced scheduling isn't likely to
1515 * cause undue delay as @bio is likely to be dispatched directly if
1516 * its @tg's disptime is not in the future.
1518 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1519 tg_update_disptime(tg);
1520 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1524 spin_unlock_irq(q->queue_lock);
1529 * As multiple blk-throtls may stack in the same issue path, we
1530 * don't want bios to leave with the flag set. Clear the flag if
1534 bio->bi_rw &= ~REQ_THROTTLED;
1539 * Dispatch all bios from all children tg's queued on @parent_sq. On
1540 * return, @parent_sq is guaranteed to not have any active children tg's
1541 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1543 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1545 struct throtl_grp *tg;
1547 while ((tg = throtl_rb_first(parent_sq))) {
1548 struct throtl_service_queue *sq = &tg->service_queue;
1551 throtl_dequeue_tg(tg);
1553 while ((bio = throtl_peek_queued(&sq->queued[READ])))
1554 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1555 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1556 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1561 * blk_throtl_drain - drain throttled bios
1562 * @q: request_queue to drain throttled bios for
1564 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1566 void blk_throtl_drain(struct request_queue *q)
1567 __releases(q->queue_lock) __acquires(q->queue_lock)
1569 struct throtl_data *td = q->td;
1570 struct blkcg_gq *blkg;
1571 struct cgroup *pos_cgrp;
1575 queue_lockdep_assert_held(q);
1579 * Drain each tg while doing post-order walk on the blkg tree, so
1580 * that all bios are propagated to td->service_queue. It'd be
1581 * better to walk service_queue tree directly but blkg walk is
1584 blkg_for_each_descendant_post(blkg, pos_cgrp, td->queue->root_blkg)
1585 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1587 tg_drain_bios(&td_root_tg(td)->service_queue);
1589 /* finally, transfer bios from top-level tg's into the td */
1590 tg_drain_bios(&td->service_queue);
1593 spin_unlock_irq(q->queue_lock);
1595 /* all bios now should be in td->service_queue, issue them */
1596 for (rw = READ; rw <= WRITE; rw++)
1597 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1599 generic_make_request(bio);
1601 spin_lock_irq(q->queue_lock);
1604 int blk_throtl_init(struct request_queue *q)
1606 struct throtl_data *td;
1609 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1613 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1614 throtl_service_queue_init(&td->service_queue, NULL);
1619 /* activate policy */
1620 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1626 void blk_throtl_exit(struct request_queue *q)
1629 throtl_shutdown_wq(q);
1630 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1634 static int __init throtl_init(void)
1636 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1637 if (!kthrotld_workqueue)
1638 panic("Failed to create kthrotld\n");
1640 return blkcg_policy_register(&blkcg_policy_throtl);
1643 module_init(throtl_init);