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 <linux/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)
87 /* must be the first member */
88 struct blkg_policy_data pd;
90 /* active throtl group service_queue member */
91 struct rb_node rb_node;
93 /* throtl_data this group belongs to */
94 struct throtl_data *td;
96 /* this group's service queue */
97 struct throtl_service_queue service_queue;
100 * qnode_on_self is used when bios are directly queued to this
101 * throtl_grp so that local bios compete fairly with bios
102 * dispatched from children. qnode_on_parent is used when bios are
103 * dispatched from this throtl_grp into its parent and will compete
104 * with the sibling qnode_on_parents and the parent's
107 struct throtl_qnode qnode_on_self[2];
108 struct throtl_qnode qnode_on_parent[2];
111 * Dispatch time in jiffies. This is the estimated time when group
112 * will unthrottle and is ready to dispatch more bio. It is used as
113 * key to sort active groups in service tree.
115 unsigned long disptime;
119 /* are there any throtl rules between this group and td? */
122 /* bytes per second rate limits */
126 unsigned int iops[2];
128 /* Number of bytes disptached in current slice */
129 uint64_t bytes_disp[2];
130 /* Number of bio's dispatched in current slice */
131 unsigned int io_disp[2];
133 /* When did we start a new slice */
134 unsigned long slice_start[2];
135 unsigned long slice_end[2];
140 /* service tree for active throtl groups */
141 struct throtl_service_queue service_queue;
143 struct request_queue *queue;
145 /* Total Number of queued bios on READ and WRITE lists */
146 unsigned int nr_queued[2];
148 /* Work for dispatching throttled bios */
149 struct work_struct dispatch_work;
152 static void throtl_pending_timer_fn(unsigned long arg);
154 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
156 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
159 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
161 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
164 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
166 return pd_to_blkg(&tg->pd);
170 * sq_to_tg - return the throl_grp the specified service queue belongs to
171 * @sq: the throtl_service_queue of interest
173 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
174 * embedded in throtl_data, %NULL is returned.
176 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
178 if (sq && sq->parent_sq)
179 return container_of(sq, struct throtl_grp, service_queue);
185 * sq_to_td - return throtl_data the specified service queue belongs to
186 * @sq: the throtl_service_queue of interest
188 * A service_queue can be embeded in either a throtl_grp or throtl_data.
189 * Determine the associated throtl_data accordingly and return it.
191 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
193 struct throtl_grp *tg = sq_to_tg(sq);
198 return container_of(sq, struct throtl_data, service_queue);
202 * throtl_log - log debug message via blktrace
203 * @sq: the service_queue being reported
204 * @fmt: printf format string
207 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
208 * throtl_grp; otherwise, just "throtl".
210 #define throtl_log(sq, fmt, args...) do { \
211 struct throtl_grp *__tg = sq_to_tg((sq)); \
212 struct throtl_data *__td = sq_to_td((sq)); \
215 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
220 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
221 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
223 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
227 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
229 INIT_LIST_HEAD(&qn->node);
230 bio_list_init(&qn->bios);
235 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
236 * @bio: bio being added
237 * @qn: qnode to add bio to
238 * @queued: the service_queue->queued[] list @qn belongs to
240 * Add @bio to @qn and put @qn on @queued if it's not already on.
241 * @qn->tg's reference count is bumped when @qn is activated. See the
242 * comment on top of throtl_qnode definition for details.
244 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
245 struct list_head *queued)
247 bio_list_add(&qn->bios, bio);
248 if (list_empty(&qn->node)) {
249 list_add_tail(&qn->node, queued);
250 blkg_get(tg_to_blkg(qn->tg));
255 * throtl_peek_queued - peek the first bio on a qnode list
256 * @queued: the qnode list to peek
258 static struct bio *throtl_peek_queued(struct list_head *queued)
260 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
263 if (list_empty(queued))
266 bio = bio_list_peek(&qn->bios);
272 * throtl_pop_queued - pop the first bio form a qnode list
273 * @queued: the qnode list to pop a bio from
274 * @tg_to_put: optional out argument for throtl_grp to put
276 * Pop the first bio from the qnode list @queued. After popping, the first
277 * qnode is removed from @queued if empty or moved to the end of @queued so
278 * that the popping order is round-robin.
280 * When the first qnode is removed, its associated throtl_grp should be put
281 * too. If @tg_to_put is NULL, this function automatically puts it;
282 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
283 * responsible for putting it.
285 static struct bio *throtl_pop_queued(struct list_head *queued,
286 struct throtl_grp **tg_to_put)
288 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
291 if (list_empty(queued))
294 bio = bio_list_pop(&qn->bios);
297 if (bio_list_empty(&qn->bios)) {
298 list_del_init(&qn->node);
302 blkg_put(tg_to_blkg(qn->tg));
304 list_move_tail(&qn->node, queued);
310 /* init a service_queue, assumes the caller zeroed it */
311 static void throtl_service_queue_init(struct throtl_service_queue *sq)
313 INIT_LIST_HEAD(&sq->queued[0]);
314 INIT_LIST_HEAD(&sq->queued[1]);
315 sq->pending_tree = RB_ROOT;
316 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
320 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
322 struct throtl_grp *tg;
325 tg = kzalloc_node(sizeof(*tg), gfp, node);
329 throtl_service_queue_init(&tg->service_queue);
331 for (rw = READ; rw <= WRITE; rw++) {
332 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
333 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
336 RB_CLEAR_NODE(&tg->rb_node);
340 tg->iops[WRITE] = -1;
345 static void throtl_pd_init(struct blkg_policy_data *pd)
347 struct throtl_grp *tg = pd_to_tg(pd);
348 struct blkcg_gq *blkg = tg_to_blkg(tg);
349 struct throtl_data *td = blkg->q->td;
350 struct throtl_service_queue *sq = &tg->service_queue;
353 * If on the default hierarchy, we switch to properly hierarchical
354 * behavior where limits on a given throtl_grp are applied to the
355 * whole subtree rather than just the group itself. e.g. If 16M
356 * read_bps limit is set on the root group, the whole system can't
357 * exceed 16M for the device.
359 * If not on the default hierarchy, the broken flat hierarchy
360 * behavior is retained where all throtl_grps are treated as if
361 * they're all separate root groups right below throtl_data.
362 * Limits of a group don't interact with limits of other groups
363 * regardless of the position of the group in the hierarchy.
365 sq->parent_sq = &td->service_queue;
366 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
367 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
372 * Set has_rules[] if @tg or any of its parents have limits configured.
373 * This doesn't require walking up to the top of the hierarchy as the
374 * parent's has_rules[] is guaranteed to be correct.
376 static void tg_update_has_rules(struct throtl_grp *tg)
378 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
381 for (rw = READ; rw <= WRITE; rw++)
382 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
383 (tg->bps[rw] != -1 || tg->iops[rw] != -1);
386 static void throtl_pd_online(struct blkg_policy_data *pd)
389 * We don't want new groups to escape the limits of its ancestors.
390 * Update has_rules[] after a new group is brought online.
392 tg_update_has_rules(pd_to_tg(pd));
395 static void throtl_pd_free(struct blkg_policy_data *pd)
397 struct throtl_grp *tg = pd_to_tg(pd);
399 del_timer_sync(&tg->service_queue.pending_timer);
403 static struct throtl_grp *
404 throtl_rb_first(struct throtl_service_queue *parent_sq)
406 /* Service tree is empty */
407 if (!parent_sq->nr_pending)
410 if (!parent_sq->first_pending)
411 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
413 if (parent_sq->first_pending)
414 return rb_entry_tg(parent_sq->first_pending);
419 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
425 static void throtl_rb_erase(struct rb_node *n,
426 struct throtl_service_queue *parent_sq)
428 if (parent_sq->first_pending == n)
429 parent_sq->first_pending = NULL;
430 rb_erase_init(n, &parent_sq->pending_tree);
431 --parent_sq->nr_pending;
434 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
436 struct throtl_grp *tg;
438 tg = throtl_rb_first(parent_sq);
442 parent_sq->first_pending_disptime = tg->disptime;
445 static void tg_service_queue_add(struct throtl_grp *tg)
447 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
448 struct rb_node **node = &parent_sq->pending_tree.rb_node;
449 struct rb_node *parent = NULL;
450 struct throtl_grp *__tg;
451 unsigned long key = tg->disptime;
454 while (*node != NULL) {
456 __tg = rb_entry_tg(parent);
458 if (time_before(key, __tg->disptime))
459 node = &parent->rb_left;
461 node = &parent->rb_right;
467 parent_sq->first_pending = &tg->rb_node;
469 rb_link_node(&tg->rb_node, parent, node);
470 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
473 static void __throtl_enqueue_tg(struct throtl_grp *tg)
475 tg_service_queue_add(tg);
476 tg->flags |= THROTL_TG_PENDING;
477 tg->service_queue.parent_sq->nr_pending++;
480 static void throtl_enqueue_tg(struct throtl_grp *tg)
482 if (!(tg->flags & THROTL_TG_PENDING))
483 __throtl_enqueue_tg(tg);
486 static void __throtl_dequeue_tg(struct throtl_grp *tg)
488 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
489 tg->flags &= ~THROTL_TG_PENDING;
492 static void throtl_dequeue_tg(struct throtl_grp *tg)
494 if (tg->flags & THROTL_TG_PENDING)
495 __throtl_dequeue_tg(tg);
498 /* Call with queue lock held */
499 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
500 unsigned long expires)
502 mod_timer(&sq->pending_timer, expires);
503 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
504 expires - jiffies, jiffies);
508 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
509 * @sq: the service_queue to schedule dispatch for
510 * @force: force scheduling
512 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
513 * dispatch time of the first pending child. Returns %true if either timer
514 * is armed or there's no pending child left. %false if the current
515 * dispatch window is still open and the caller should continue
518 * If @force is %true, the dispatch timer is always scheduled and this
519 * function is guaranteed to return %true. This is to be used when the
520 * caller can't dispatch itself and needs to invoke pending_timer
521 * unconditionally. Note that forced scheduling is likely to induce short
522 * delay before dispatch starts even if @sq->first_pending_disptime is not
523 * in the future and thus shouldn't be used in hot paths.
525 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
528 /* any pending children left? */
532 update_min_dispatch_time(sq);
534 /* is the next dispatch time in the future? */
535 if (force || time_after(sq->first_pending_disptime, jiffies)) {
536 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
540 /* tell the caller to continue dispatching */
544 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
545 bool rw, unsigned long start)
547 tg->bytes_disp[rw] = 0;
551 * Previous slice has expired. We must have trimmed it after last
552 * bio dispatch. That means since start of last slice, we never used
553 * that bandwidth. Do try to make use of that bandwidth while giving
556 if (time_after_eq(start, tg->slice_start[rw]))
557 tg->slice_start[rw] = start;
559 tg->slice_end[rw] = jiffies + throtl_slice;
560 throtl_log(&tg->service_queue,
561 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
562 rw == READ ? 'R' : 'W', tg->slice_start[rw],
563 tg->slice_end[rw], jiffies);
566 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
568 tg->bytes_disp[rw] = 0;
570 tg->slice_start[rw] = jiffies;
571 tg->slice_end[rw] = jiffies + throtl_slice;
572 throtl_log(&tg->service_queue,
573 "[%c] new slice start=%lu end=%lu jiffies=%lu",
574 rw == READ ? 'R' : 'W', tg->slice_start[rw],
575 tg->slice_end[rw], jiffies);
578 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
579 unsigned long jiffy_end)
581 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
584 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
585 unsigned long jiffy_end)
587 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
588 throtl_log(&tg->service_queue,
589 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
590 rw == READ ? 'R' : 'W', tg->slice_start[rw],
591 tg->slice_end[rw], jiffies);
594 /* Determine if previously allocated or extended slice is complete or not */
595 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
597 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
603 /* Trim the used slices and adjust slice start accordingly */
604 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
606 unsigned long nr_slices, time_elapsed, io_trim;
609 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
612 * If bps are unlimited (-1), then time slice don't get
613 * renewed. Don't try to trim the slice if slice is used. A new
614 * slice will start when appropriate.
616 if (throtl_slice_used(tg, rw))
620 * A bio has been dispatched. Also adjust slice_end. It might happen
621 * that initially cgroup limit was very low resulting in high
622 * slice_end, but later limit was bumped up and bio was dispached
623 * sooner, then we need to reduce slice_end. A high bogus slice_end
624 * is bad because it does not allow new slice to start.
627 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
629 time_elapsed = jiffies - tg->slice_start[rw];
631 nr_slices = time_elapsed / throtl_slice;
635 tmp = tg->bps[rw] * throtl_slice * nr_slices;
639 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
641 if (!bytes_trim && !io_trim)
644 if (tg->bytes_disp[rw] >= bytes_trim)
645 tg->bytes_disp[rw] -= bytes_trim;
647 tg->bytes_disp[rw] = 0;
649 if (tg->io_disp[rw] >= io_trim)
650 tg->io_disp[rw] -= io_trim;
654 tg->slice_start[rw] += nr_slices * throtl_slice;
656 throtl_log(&tg->service_queue,
657 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
658 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
659 tg->slice_start[rw], tg->slice_end[rw], jiffies);
662 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
665 bool rw = bio_data_dir(bio);
666 unsigned int io_allowed;
667 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
670 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
672 /* Slice has just started. Consider one slice interval */
674 jiffy_elapsed_rnd = throtl_slice;
676 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
679 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
680 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
681 * will allow dispatch after 1 second and after that slice should
685 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
689 io_allowed = UINT_MAX;
693 if (tg->io_disp[rw] + 1 <= io_allowed) {
699 /* Calc approx time to dispatch */
700 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
702 if (jiffy_wait > jiffy_elapsed)
703 jiffy_wait = jiffy_wait - jiffy_elapsed;
712 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
715 bool rw = bio_data_dir(bio);
716 u64 bytes_allowed, extra_bytes, tmp;
717 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
719 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
721 /* Slice has just started. Consider one slice interval */
723 jiffy_elapsed_rnd = throtl_slice;
725 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
727 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
731 if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
737 /* Calc approx time to dispatch */
738 extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
739 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
745 * This wait time is without taking into consideration the rounding
746 * up we did. Add that time also.
748 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
755 * Returns whether one can dispatch a bio or not. Also returns approx number
756 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
758 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
761 bool rw = bio_data_dir(bio);
762 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
765 * Currently whole state machine of group depends on first bio
766 * queued in the group bio list. So one should not be calling
767 * this function with a different bio if there are other bios
770 BUG_ON(tg->service_queue.nr_queued[rw] &&
771 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
773 /* If tg->bps = -1, then BW is unlimited */
774 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
781 * If previous slice expired, start a new one otherwise renew/extend
782 * existing slice to make sure it is at least throtl_slice interval
783 * long since now. New slice is started only for empty throttle group.
784 * If there is queued bio, that means there should be an active
785 * slice and it should be extended instead.
787 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
788 throtl_start_new_slice(tg, rw);
790 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
791 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
794 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
795 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
801 max_wait = max(bps_wait, iops_wait);
806 if (time_before(tg->slice_end[rw], jiffies + max_wait))
807 throtl_extend_slice(tg, rw, jiffies + max_wait);
812 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
814 bool rw = bio_data_dir(bio);
816 /* Charge the bio to the group */
817 tg->bytes_disp[rw] += bio->bi_iter.bi_size;
821 * BIO_THROTTLED is used to prevent the same bio to be throttled
822 * more than once as a throttled bio will go through blk-throtl the
823 * second time when it eventually gets issued. Set it when a bio
824 * is being charged to a tg.
826 if (!bio_flagged(bio, BIO_THROTTLED))
827 bio_set_flag(bio, BIO_THROTTLED);
831 * throtl_add_bio_tg - add a bio to the specified throtl_grp
834 * @tg: the target throtl_grp
836 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
837 * tg->qnode_on_self[] is used.
839 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
840 struct throtl_grp *tg)
842 struct throtl_service_queue *sq = &tg->service_queue;
843 bool rw = bio_data_dir(bio);
846 qn = &tg->qnode_on_self[rw];
849 * If @tg doesn't currently have any bios queued in the same
850 * direction, queueing @bio can change when @tg should be
851 * dispatched. Mark that @tg was empty. This is automatically
852 * cleaered on the next tg_update_disptime().
854 if (!sq->nr_queued[rw])
855 tg->flags |= THROTL_TG_WAS_EMPTY;
857 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
860 throtl_enqueue_tg(tg);
863 static void tg_update_disptime(struct throtl_grp *tg)
865 struct throtl_service_queue *sq = &tg->service_queue;
866 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
869 if ((bio = throtl_peek_queued(&sq->queued[READ])))
870 tg_may_dispatch(tg, bio, &read_wait);
872 if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
873 tg_may_dispatch(tg, bio, &write_wait);
875 min_wait = min(read_wait, write_wait);
876 disptime = jiffies + min_wait;
878 /* Update dispatch time */
879 throtl_dequeue_tg(tg);
880 tg->disptime = disptime;
881 throtl_enqueue_tg(tg);
883 /* see throtl_add_bio_tg() */
884 tg->flags &= ~THROTL_TG_WAS_EMPTY;
887 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
888 struct throtl_grp *parent_tg, bool rw)
890 if (throtl_slice_used(parent_tg, rw)) {
891 throtl_start_new_slice_with_credit(parent_tg, rw,
892 child_tg->slice_start[rw]);
897 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
899 struct throtl_service_queue *sq = &tg->service_queue;
900 struct throtl_service_queue *parent_sq = sq->parent_sq;
901 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
902 struct throtl_grp *tg_to_put = NULL;
906 * @bio is being transferred from @tg to @parent_sq. Popping a bio
907 * from @tg may put its reference and @parent_sq might end up
908 * getting released prematurely. Remember the tg to put and put it
909 * after @bio is transferred to @parent_sq.
911 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
914 throtl_charge_bio(tg, bio);
917 * If our parent is another tg, we just need to transfer @bio to
918 * the parent using throtl_add_bio_tg(). If our parent is
919 * @td->service_queue, @bio is ready to be issued. Put it on its
920 * bio_lists[] and decrease total number queued. The caller is
921 * responsible for issuing these bios.
924 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
925 start_parent_slice_with_credit(tg, parent_tg, rw);
927 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
928 &parent_sq->queued[rw]);
929 BUG_ON(tg->td->nr_queued[rw] <= 0);
930 tg->td->nr_queued[rw]--;
933 throtl_trim_slice(tg, rw);
936 blkg_put(tg_to_blkg(tg_to_put));
939 static int throtl_dispatch_tg(struct throtl_grp *tg)
941 struct throtl_service_queue *sq = &tg->service_queue;
942 unsigned int nr_reads = 0, nr_writes = 0;
943 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
944 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
947 /* Try to dispatch 75% READS and 25% WRITES */
949 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
950 tg_may_dispatch(tg, bio, NULL)) {
952 tg_dispatch_one_bio(tg, bio_data_dir(bio));
955 if (nr_reads >= max_nr_reads)
959 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
960 tg_may_dispatch(tg, bio, NULL)) {
962 tg_dispatch_one_bio(tg, bio_data_dir(bio));
965 if (nr_writes >= max_nr_writes)
969 return nr_reads + nr_writes;
972 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
974 unsigned int nr_disp = 0;
977 struct throtl_grp *tg = throtl_rb_first(parent_sq);
978 struct throtl_service_queue *sq = &tg->service_queue;
983 if (time_before(jiffies, tg->disptime))
986 throtl_dequeue_tg(tg);
988 nr_disp += throtl_dispatch_tg(tg);
990 if (sq->nr_queued[0] || sq->nr_queued[1])
991 tg_update_disptime(tg);
993 if (nr_disp >= throtl_quantum)
1001 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1002 * @arg: the throtl_service_queue being serviced
1004 * This timer is armed when a child throtl_grp with active bio's become
1005 * pending and queued on the service_queue's pending_tree and expires when
1006 * the first child throtl_grp should be dispatched. This function
1007 * dispatches bio's from the children throtl_grps to the parent
1010 * If the parent's parent is another throtl_grp, dispatching is propagated
1011 * by either arming its pending_timer or repeating dispatch directly. If
1012 * the top-level service_tree is reached, throtl_data->dispatch_work is
1013 * kicked so that the ready bio's are issued.
1015 static void throtl_pending_timer_fn(unsigned long arg)
1017 struct throtl_service_queue *sq = (void *)arg;
1018 struct throtl_grp *tg = sq_to_tg(sq);
1019 struct throtl_data *td = sq_to_td(sq);
1020 struct request_queue *q = td->queue;
1021 struct throtl_service_queue *parent_sq;
1025 spin_lock_irq(q->queue_lock);
1027 parent_sq = sq->parent_sq;
1031 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1032 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1033 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1035 ret = throtl_select_dispatch(sq);
1037 throtl_log(sq, "bios disp=%u", ret);
1041 if (throtl_schedule_next_dispatch(sq, false))
1044 /* this dispatch windows is still open, relax and repeat */
1045 spin_unlock_irq(q->queue_lock);
1047 spin_lock_irq(q->queue_lock);
1054 /* @parent_sq is another throl_grp, propagate dispatch */
1055 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1056 tg_update_disptime(tg);
1057 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1058 /* window is already open, repeat dispatching */
1065 /* reached the top-level, queue issueing */
1066 queue_work(kthrotld_workqueue, &td->dispatch_work);
1069 spin_unlock_irq(q->queue_lock);
1073 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1074 * @work: work item being executed
1076 * This function is queued for execution when bio's reach the bio_lists[]
1077 * of throtl_data->service_queue. Those bio's are ready and issued by this
1080 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1082 struct throtl_data *td = container_of(work, struct throtl_data,
1084 struct throtl_service_queue *td_sq = &td->service_queue;
1085 struct request_queue *q = td->queue;
1086 struct bio_list bio_list_on_stack;
1088 struct blk_plug plug;
1091 bio_list_init(&bio_list_on_stack);
1093 spin_lock_irq(q->queue_lock);
1094 for (rw = READ; rw <= WRITE; rw++)
1095 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1096 bio_list_add(&bio_list_on_stack, bio);
1097 spin_unlock_irq(q->queue_lock);
1099 if (!bio_list_empty(&bio_list_on_stack)) {
1100 blk_start_plug(&plug);
1101 while((bio = bio_list_pop(&bio_list_on_stack)))
1102 generic_make_request(bio);
1103 blk_finish_plug(&plug);
1107 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1110 struct throtl_grp *tg = pd_to_tg(pd);
1111 u64 v = *(u64 *)((void *)tg + off);
1115 return __blkg_prfill_u64(sf, pd, v);
1118 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1121 struct throtl_grp *tg = pd_to_tg(pd);
1122 unsigned int v = *(unsigned int *)((void *)tg + off);
1126 return __blkg_prfill_u64(sf, pd, v);
1129 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1131 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1132 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1136 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1138 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1139 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1143 static void tg_conf_updated(struct throtl_grp *tg)
1145 struct throtl_service_queue *sq = &tg->service_queue;
1146 struct cgroup_subsys_state *pos_css;
1147 struct blkcg_gq *blkg;
1149 throtl_log(&tg->service_queue,
1150 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1151 tg->bps[READ], tg->bps[WRITE],
1152 tg->iops[READ], tg->iops[WRITE]);
1155 * Update has_rules[] flags for the updated tg's subtree. A tg is
1156 * considered to have rules if either the tg itself or any of its
1157 * ancestors has rules. This identifies groups without any
1158 * restrictions in the whole hierarchy and allows them to bypass
1161 blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg))
1162 tg_update_has_rules(blkg_to_tg(blkg));
1165 * We're already holding queue_lock and know @tg is valid. Let's
1166 * apply the new config directly.
1168 * Restart the slices for both READ and WRITES. It might happen
1169 * that a group's limit are dropped suddenly and we don't want to
1170 * account recently dispatched IO with new low rate.
1172 throtl_start_new_slice(tg, 0);
1173 throtl_start_new_slice(tg, 1);
1175 if (tg->flags & THROTL_TG_PENDING) {
1176 tg_update_disptime(tg);
1177 throtl_schedule_next_dispatch(sq->parent_sq, true);
1181 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1182 char *buf, size_t nbytes, loff_t off, bool is_u64)
1184 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1185 struct blkg_conf_ctx ctx;
1186 struct throtl_grp *tg;
1190 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1195 if (sscanf(ctx.body, "%llu", &v) != 1)
1200 tg = blkg_to_tg(ctx.blkg);
1203 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1205 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1207 tg_conf_updated(tg);
1210 blkg_conf_finish(&ctx);
1211 return ret ?: nbytes;
1214 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1215 char *buf, size_t nbytes, loff_t off)
1217 return tg_set_conf(of, buf, nbytes, off, true);
1220 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1221 char *buf, size_t nbytes, loff_t off)
1223 return tg_set_conf(of, buf, nbytes, off, false);
1226 static struct cftype throtl_legacy_files[] = {
1228 .name = "throttle.read_bps_device",
1229 .private = offsetof(struct throtl_grp, bps[READ]),
1230 .seq_show = tg_print_conf_u64,
1231 .write = tg_set_conf_u64,
1234 .name = "throttle.write_bps_device",
1235 .private = offsetof(struct throtl_grp, bps[WRITE]),
1236 .seq_show = tg_print_conf_u64,
1237 .write = tg_set_conf_u64,
1240 .name = "throttle.read_iops_device",
1241 .private = offsetof(struct throtl_grp, iops[READ]),
1242 .seq_show = tg_print_conf_uint,
1243 .write = tg_set_conf_uint,
1246 .name = "throttle.write_iops_device",
1247 .private = offsetof(struct throtl_grp, iops[WRITE]),
1248 .seq_show = tg_print_conf_uint,
1249 .write = tg_set_conf_uint,
1252 .name = "throttle.io_service_bytes",
1253 .private = (unsigned long)&blkcg_policy_throtl,
1254 .seq_show = blkg_print_stat_bytes,
1257 .name = "throttle.io_serviced",
1258 .private = (unsigned long)&blkcg_policy_throtl,
1259 .seq_show = blkg_print_stat_ios,
1264 static u64 tg_prfill_max(struct seq_file *sf, struct blkg_policy_data *pd,
1267 struct throtl_grp *tg = pd_to_tg(pd);
1268 const char *dname = blkg_dev_name(pd->blkg);
1269 char bufs[4][21] = { "max", "max", "max", "max" };
1273 if (tg->bps[READ] == -1 && tg->bps[WRITE] == -1 &&
1274 tg->iops[READ] == -1 && tg->iops[WRITE] == -1)
1277 if (tg->bps[READ] != -1)
1278 snprintf(bufs[0], sizeof(bufs[0]), "%llu", tg->bps[READ]);
1279 if (tg->bps[WRITE] != -1)
1280 snprintf(bufs[1], sizeof(bufs[1]), "%llu", tg->bps[WRITE]);
1281 if (tg->iops[READ] != -1)
1282 snprintf(bufs[2], sizeof(bufs[2]), "%u", tg->iops[READ]);
1283 if (tg->iops[WRITE] != -1)
1284 snprintf(bufs[3], sizeof(bufs[3]), "%u", tg->iops[WRITE]);
1286 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s\n",
1287 dname, bufs[0], bufs[1], bufs[2], bufs[3]);
1291 static int tg_print_max(struct seq_file *sf, void *v)
1293 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_max,
1294 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1298 static ssize_t tg_set_max(struct kernfs_open_file *of,
1299 char *buf, size_t nbytes, loff_t off)
1301 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1302 struct blkg_conf_ctx ctx;
1303 struct throtl_grp *tg;
1307 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1311 tg = blkg_to_tg(ctx.blkg);
1313 v[0] = tg->bps[READ];
1314 v[1] = tg->bps[WRITE];
1315 v[2] = tg->iops[READ];
1316 v[3] = tg->iops[WRITE];
1319 char tok[27]; /* wiops=18446744073709551616 */
1324 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1333 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1341 if (!strcmp(tok, "rbps"))
1343 else if (!strcmp(tok, "wbps"))
1345 else if (!strcmp(tok, "riops"))
1346 v[2] = min_t(u64, val, UINT_MAX);
1347 else if (!strcmp(tok, "wiops"))
1348 v[3] = min_t(u64, val, UINT_MAX);
1353 tg->bps[READ] = v[0];
1354 tg->bps[WRITE] = v[1];
1355 tg->iops[READ] = v[2];
1356 tg->iops[WRITE] = v[3];
1358 tg_conf_updated(tg);
1361 blkg_conf_finish(&ctx);
1362 return ret ?: nbytes;
1365 static struct cftype throtl_files[] = {
1368 .flags = CFTYPE_NOT_ON_ROOT,
1369 .seq_show = tg_print_max,
1370 .write = tg_set_max,
1375 static void throtl_shutdown_wq(struct request_queue *q)
1377 struct throtl_data *td = q->td;
1379 cancel_work_sync(&td->dispatch_work);
1382 static struct blkcg_policy blkcg_policy_throtl = {
1383 .dfl_cftypes = throtl_files,
1384 .legacy_cftypes = throtl_legacy_files,
1386 .pd_alloc_fn = throtl_pd_alloc,
1387 .pd_init_fn = throtl_pd_init,
1388 .pd_online_fn = throtl_pd_online,
1389 .pd_free_fn = throtl_pd_free,
1392 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
1395 struct throtl_qnode *qn = NULL;
1396 struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
1397 struct throtl_service_queue *sq;
1398 bool rw = bio_data_dir(bio);
1399 bool throttled = false;
1401 WARN_ON_ONCE(!rcu_read_lock_held());
1403 /* see throtl_charge_bio() */
1404 if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw])
1407 spin_lock_irq(q->queue_lock);
1409 if (unlikely(blk_queue_bypass(q)))
1412 sq = &tg->service_queue;
1415 /* throtl is FIFO - if bios are already queued, should queue */
1416 if (sq->nr_queued[rw])
1419 /* if above limits, break to queue */
1420 if (!tg_may_dispatch(tg, bio, NULL))
1423 /* within limits, let's charge and dispatch directly */
1424 throtl_charge_bio(tg, bio);
1427 * We need to trim slice even when bios are not being queued
1428 * otherwise it might happen that a bio is not queued for
1429 * a long time and slice keeps on extending and trim is not
1430 * called for a long time. Now if limits are reduced suddenly
1431 * we take into account all the IO dispatched so far at new
1432 * low rate and * newly queued IO gets a really long dispatch
1435 * So keep on trimming slice even if bio is not queued.
1437 throtl_trim_slice(tg, rw);
1440 * @bio passed through this layer without being throttled.
1441 * Climb up the ladder. If we''re already at the top, it
1442 * can be executed directly.
1444 qn = &tg->qnode_on_parent[rw];
1451 /* out-of-limit, queue to @tg */
1452 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1453 rw == READ ? 'R' : 'W',
1454 tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw],
1455 tg->io_disp[rw], tg->iops[rw],
1456 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1458 bio_associate_current(bio);
1459 tg->td->nr_queued[rw]++;
1460 throtl_add_bio_tg(bio, qn, tg);
1464 * Update @tg's dispatch time and force schedule dispatch if @tg
1465 * was empty before @bio. The forced scheduling isn't likely to
1466 * cause undue delay as @bio is likely to be dispatched directly if
1467 * its @tg's disptime is not in the future.
1469 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1470 tg_update_disptime(tg);
1471 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1475 spin_unlock_irq(q->queue_lock);
1478 * As multiple blk-throtls may stack in the same issue path, we
1479 * don't want bios to leave with the flag set. Clear the flag if
1483 bio_clear_flag(bio, BIO_THROTTLED);
1488 * Dispatch all bios from all children tg's queued on @parent_sq. On
1489 * return, @parent_sq is guaranteed to not have any active children tg's
1490 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1492 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1494 struct throtl_grp *tg;
1496 while ((tg = throtl_rb_first(parent_sq))) {
1497 struct throtl_service_queue *sq = &tg->service_queue;
1500 throtl_dequeue_tg(tg);
1502 while ((bio = throtl_peek_queued(&sq->queued[READ])))
1503 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1504 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1505 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1510 * blk_throtl_drain - drain throttled bios
1511 * @q: request_queue to drain throttled bios for
1513 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1515 void blk_throtl_drain(struct request_queue *q)
1516 __releases(q->queue_lock) __acquires(q->queue_lock)
1518 struct throtl_data *td = q->td;
1519 struct blkcg_gq *blkg;
1520 struct cgroup_subsys_state *pos_css;
1524 queue_lockdep_assert_held(q);
1528 * Drain each tg while doing post-order walk on the blkg tree, so
1529 * that all bios are propagated to td->service_queue. It'd be
1530 * better to walk service_queue tree directly but blkg walk is
1533 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
1534 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1536 /* finally, transfer bios from top-level tg's into the td */
1537 tg_drain_bios(&td->service_queue);
1540 spin_unlock_irq(q->queue_lock);
1542 /* all bios now should be in td->service_queue, issue them */
1543 for (rw = READ; rw <= WRITE; rw++)
1544 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1546 generic_make_request(bio);
1548 spin_lock_irq(q->queue_lock);
1551 int blk_throtl_init(struct request_queue *q)
1553 struct throtl_data *td;
1556 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1560 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1561 throtl_service_queue_init(&td->service_queue);
1566 /* activate policy */
1567 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1573 void blk_throtl_exit(struct request_queue *q)
1576 throtl_shutdown_wq(q);
1577 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1581 static int __init throtl_init(void)
1583 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1584 if (!kthrotld_workqueue)
1585 panic("Failed to create kthrotld\n");
1587 return blkcg_policy_register(&blkcg_policy_throtl);
1590 module_init(throtl_init);