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];
149 * number of total undestroyed groups
151 unsigned int nr_undestroyed_grps;
153 /* Work for dispatching throttled bios */
154 struct work_struct dispatch_work;
157 static void throtl_pending_timer_fn(unsigned long arg);
159 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
161 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
164 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
166 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
169 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
171 return pd_to_blkg(&tg->pd);
175 * sq_to_tg - return the throl_grp the specified service queue belongs to
176 * @sq: the throtl_service_queue of interest
178 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
179 * embedded in throtl_data, %NULL is returned.
181 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
183 if (sq && sq->parent_sq)
184 return container_of(sq, struct throtl_grp, service_queue);
190 * sq_to_td - return throtl_data the specified service queue belongs to
191 * @sq: the throtl_service_queue of interest
193 * A service_queue can be embeded in either a throtl_grp or throtl_data.
194 * Determine the associated throtl_data accordingly and return it.
196 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
198 struct throtl_grp *tg = sq_to_tg(sq);
203 return container_of(sq, struct throtl_data, service_queue);
207 * throtl_log - log debug message via blktrace
208 * @sq: the service_queue being reported
209 * @fmt: printf format string
212 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
213 * throtl_grp; otherwise, just "throtl".
215 * TODO: this should be made a function and name formatting should happen
216 * after testing whether blktrace is enabled.
218 #define throtl_log(sq, fmt, args...) do { \
219 struct throtl_grp *__tg = sq_to_tg((sq)); \
220 struct throtl_data *__td = sq_to_td((sq)); \
226 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
227 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
229 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
233 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
235 INIT_LIST_HEAD(&qn->node);
236 bio_list_init(&qn->bios);
241 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
242 * @bio: bio being added
243 * @qn: qnode to add bio to
244 * @queued: the service_queue->queued[] list @qn belongs to
246 * Add @bio to @qn and put @qn on @queued if it's not already on.
247 * @qn->tg's reference count is bumped when @qn is activated. See the
248 * comment on top of throtl_qnode definition for details.
250 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
251 struct list_head *queued)
253 bio_list_add(&qn->bios, bio);
254 if (list_empty(&qn->node)) {
255 list_add_tail(&qn->node, queued);
256 blkg_get(tg_to_blkg(qn->tg));
261 * throtl_peek_queued - peek the first bio on a qnode list
262 * @queued: the qnode list to peek
264 static struct bio *throtl_peek_queued(struct list_head *queued)
266 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
269 if (list_empty(queued))
272 bio = bio_list_peek(&qn->bios);
278 * throtl_pop_queued - pop the first bio form a qnode list
279 * @queued: the qnode list to pop a bio from
280 * @tg_to_put: optional out argument for throtl_grp to put
282 * Pop the first bio from the qnode list @queued. After popping, the first
283 * qnode is removed from @queued if empty or moved to the end of @queued so
284 * that the popping order is round-robin.
286 * When the first qnode is removed, its associated throtl_grp should be put
287 * too. If @tg_to_put is NULL, this function automatically puts it;
288 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
289 * responsible for putting it.
291 static struct bio *throtl_pop_queued(struct list_head *queued,
292 struct throtl_grp **tg_to_put)
294 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
297 if (list_empty(queued))
300 bio = bio_list_pop(&qn->bios);
303 if (bio_list_empty(&qn->bios)) {
304 list_del_init(&qn->node);
308 blkg_put(tg_to_blkg(qn->tg));
310 list_move_tail(&qn->node, queued);
316 /* init a service_queue, assumes the caller zeroed it */
317 static void throtl_service_queue_init(struct throtl_service_queue *sq)
319 INIT_LIST_HEAD(&sq->queued[0]);
320 INIT_LIST_HEAD(&sq->queued[1]);
321 sq->pending_tree = RB_ROOT;
322 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
326 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
328 struct throtl_grp *tg;
331 tg = kzalloc_node(sizeof(*tg), gfp, node);
335 throtl_service_queue_init(&tg->service_queue);
337 for (rw = READ; rw <= WRITE; rw++) {
338 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
339 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
342 RB_CLEAR_NODE(&tg->rb_node);
346 tg->iops[WRITE] = -1;
351 static void throtl_pd_init(struct blkg_policy_data *pd)
353 struct throtl_grp *tg = pd_to_tg(pd);
354 struct blkcg_gq *blkg = tg_to_blkg(tg);
355 struct throtl_data *td = blkg->q->td;
356 struct throtl_service_queue *sq = &tg->service_queue;
359 * If on the default hierarchy, we switch to properly hierarchical
360 * behavior where limits on a given throtl_grp are applied to the
361 * whole subtree rather than just the group itself. e.g. If 16M
362 * read_bps limit is set on the root group, the whole system can't
363 * exceed 16M for the device.
365 * If not on the default hierarchy, the broken flat hierarchy
366 * behavior is retained where all throtl_grps are treated as if
367 * they're all separate root groups right below throtl_data.
368 * Limits of a group don't interact with limits of other groups
369 * regardless of the position of the group in the hierarchy.
371 sq->parent_sq = &td->service_queue;
372 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
373 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
378 * Set has_rules[] if @tg or any of its parents have limits configured.
379 * This doesn't require walking up to the top of the hierarchy as the
380 * parent's has_rules[] is guaranteed to be correct.
382 static void tg_update_has_rules(struct throtl_grp *tg)
384 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
387 for (rw = READ; rw <= WRITE; rw++)
388 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
389 (tg->bps[rw] != -1 || tg->iops[rw] != -1);
392 static void throtl_pd_online(struct blkg_policy_data *pd)
395 * We don't want new groups to escape the limits of its ancestors.
396 * Update has_rules[] after a new group is brought online.
398 tg_update_has_rules(pd_to_tg(pd));
401 static void throtl_pd_free(struct blkg_policy_data *pd)
403 struct throtl_grp *tg = pd_to_tg(pd);
405 del_timer_sync(&tg->service_queue.pending_timer);
409 static struct throtl_grp *
410 throtl_rb_first(struct throtl_service_queue *parent_sq)
412 /* Service tree is empty */
413 if (!parent_sq->nr_pending)
416 if (!parent_sq->first_pending)
417 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
419 if (parent_sq->first_pending)
420 return rb_entry_tg(parent_sq->first_pending);
425 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
431 static void throtl_rb_erase(struct rb_node *n,
432 struct throtl_service_queue *parent_sq)
434 if (parent_sq->first_pending == n)
435 parent_sq->first_pending = NULL;
436 rb_erase_init(n, &parent_sq->pending_tree);
437 --parent_sq->nr_pending;
440 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
442 struct throtl_grp *tg;
444 tg = throtl_rb_first(parent_sq);
448 parent_sq->first_pending_disptime = tg->disptime;
451 static void tg_service_queue_add(struct throtl_grp *tg)
453 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
454 struct rb_node **node = &parent_sq->pending_tree.rb_node;
455 struct rb_node *parent = NULL;
456 struct throtl_grp *__tg;
457 unsigned long key = tg->disptime;
460 while (*node != NULL) {
462 __tg = rb_entry_tg(parent);
464 if (time_before(key, __tg->disptime))
465 node = &parent->rb_left;
467 node = &parent->rb_right;
473 parent_sq->first_pending = &tg->rb_node;
475 rb_link_node(&tg->rb_node, parent, node);
476 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
479 static void __throtl_enqueue_tg(struct throtl_grp *tg)
481 tg_service_queue_add(tg);
482 tg->flags |= THROTL_TG_PENDING;
483 tg->service_queue.parent_sq->nr_pending++;
486 static void throtl_enqueue_tg(struct throtl_grp *tg)
488 if (!(tg->flags & THROTL_TG_PENDING))
489 __throtl_enqueue_tg(tg);
492 static void __throtl_dequeue_tg(struct throtl_grp *tg)
494 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
495 tg->flags &= ~THROTL_TG_PENDING;
498 static void throtl_dequeue_tg(struct throtl_grp *tg)
500 if (tg->flags & THROTL_TG_PENDING)
501 __throtl_dequeue_tg(tg);
504 /* Call with queue lock held */
505 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
506 unsigned long expires)
508 mod_timer(&sq->pending_timer, expires);
509 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
510 expires - jiffies, jiffies);
514 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
515 * @sq: the service_queue to schedule dispatch for
516 * @force: force scheduling
518 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
519 * dispatch time of the first pending child. Returns %true if either timer
520 * is armed or there's no pending child left. %false if the current
521 * dispatch window is still open and the caller should continue
524 * If @force is %true, the dispatch timer is always scheduled and this
525 * function is guaranteed to return %true. This is to be used when the
526 * caller can't dispatch itself and needs to invoke pending_timer
527 * unconditionally. Note that forced scheduling is likely to induce short
528 * delay before dispatch starts even if @sq->first_pending_disptime is not
529 * in the future and thus shouldn't be used in hot paths.
531 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
534 /* any pending children left? */
538 update_min_dispatch_time(sq);
540 /* is the next dispatch time in the future? */
541 if (force || time_after(sq->first_pending_disptime, jiffies)) {
542 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
546 /* tell the caller to continue dispatching */
550 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
551 bool rw, unsigned long start)
553 tg->bytes_disp[rw] = 0;
557 * Previous slice has expired. We must have trimmed it after last
558 * bio dispatch. That means since start of last slice, we never used
559 * that bandwidth. Do try to make use of that bandwidth while giving
562 if (time_after_eq(start, tg->slice_start[rw]))
563 tg->slice_start[rw] = start;
565 tg->slice_end[rw] = jiffies + throtl_slice;
566 throtl_log(&tg->service_queue,
567 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
568 rw == READ ? 'R' : 'W', tg->slice_start[rw],
569 tg->slice_end[rw], jiffies);
572 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
574 tg->bytes_disp[rw] = 0;
576 tg->slice_start[rw] = jiffies;
577 tg->slice_end[rw] = jiffies + throtl_slice;
578 throtl_log(&tg->service_queue,
579 "[%c] new slice start=%lu end=%lu jiffies=%lu",
580 rw == READ ? 'R' : 'W', tg->slice_start[rw],
581 tg->slice_end[rw], jiffies);
584 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
585 unsigned long jiffy_end)
587 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
590 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
591 unsigned long jiffy_end)
593 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
594 throtl_log(&tg->service_queue,
595 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
596 rw == READ ? 'R' : 'W', tg->slice_start[rw],
597 tg->slice_end[rw], jiffies);
600 /* Determine if previously allocated or extended slice is complete or not */
601 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
603 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
609 /* Trim the used slices and adjust slice start accordingly */
610 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
612 unsigned long nr_slices, time_elapsed, io_trim;
615 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
618 * If bps are unlimited (-1), then time slice don't get
619 * renewed. Don't try to trim the slice if slice is used. A new
620 * slice will start when appropriate.
622 if (throtl_slice_used(tg, rw))
626 * A bio has been dispatched. Also adjust slice_end. It might happen
627 * that initially cgroup limit was very low resulting in high
628 * slice_end, but later limit was bumped up and bio was dispached
629 * sooner, then we need to reduce slice_end. A high bogus slice_end
630 * is bad because it does not allow new slice to start.
633 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
635 time_elapsed = jiffies - tg->slice_start[rw];
637 nr_slices = time_elapsed / throtl_slice;
641 tmp = tg->bps[rw] * throtl_slice * nr_slices;
645 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
647 if (!bytes_trim && !io_trim)
650 if (tg->bytes_disp[rw] >= bytes_trim)
651 tg->bytes_disp[rw] -= bytes_trim;
653 tg->bytes_disp[rw] = 0;
655 if (tg->io_disp[rw] >= io_trim)
656 tg->io_disp[rw] -= io_trim;
660 tg->slice_start[rw] += nr_slices * throtl_slice;
662 throtl_log(&tg->service_queue,
663 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
664 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
665 tg->slice_start[rw], tg->slice_end[rw], jiffies);
668 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
671 bool rw = bio_data_dir(bio);
672 unsigned int io_allowed;
673 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
676 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
678 /* Slice has just started. Consider one slice interval */
680 jiffy_elapsed_rnd = throtl_slice;
682 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
685 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
686 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
687 * will allow dispatch after 1 second and after that slice should
691 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
695 io_allowed = UINT_MAX;
699 if (tg->io_disp[rw] + 1 <= io_allowed) {
705 /* Calc approx time to dispatch */
706 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
708 if (jiffy_wait > jiffy_elapsed)
709 jiffy_wait = jiffy_wait - jiffy_elapsed;
718 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
721 bool rw = bio_data_dir(bio);
722 u64 bytes_allowed, extra_bytes, tmp;
723 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
725 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
727 /* Slice has just started. Consider one slice interval */
729 jiffy_elapsed_rnd = throtl_slice;
731 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
733 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
737 if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
743 /* Calc approx time to dispatch */
744 extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
745 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
751 * This wait time is without taking into consideration the rounding
752 * up we did. Add that time also.
754 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
761 * Returns whether one can dispatch a bio or not. Also returns approx number
762 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
764 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
767 bool rw = bio_data_dir(bio);
768 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
771 * Currently whole state machine of group depends on first bio
772 * queued in the group bio list. So one should not be calling
773 * this function with a different bio if there are other bios
776 BUG_ON(tg->service_queue.nr_queued[rw] &&
777 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
779 /* If tg->bps = -1, then BW is unlimited */
780 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
787 * If previous slice expired, start a new one otherwise renew/extend
788 * existing slice to make sure it is at least throtl_slice interval
791 if (throtl_slice_used(tg, rw))
792 throtl_start_new_slice(tg, rw);
794 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
795 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
798 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
799 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
805 max_wait = max(bps_wait, iops_wait);
810 if (time_before(tg->slice_end[rw], jiffies + max_wait))
811 throtl_extend_slice(tg, rw, jiffies + max_wait);
816 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
818 bool rw = bio_data_dir(bio);
820 /* Charge the bio to the group */
821 tg->bytes_disp[rw] += bio->bi_iter.bi_size;
825 * REQ_THROTTLED is used to prevent the same bio to be throttled
826 * more than once as a throttled bio will go through blk-throtl the
827 * second time when it eventually gets issued. Set it when a bio
828 * is being charged to a tg.
830 if (!(bio->bi_rw & REQ_THROTTLED))
831 bio->bi_rw |= REQ_THROTTLED;
835 * throtl_add_bio_tg - add a bio to the specified throtl_grp
838 * @tg: the target throtl_grp
840 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
841 * tg->qnode_on_self[] is used.
843 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
844 struct throtl_grp *tg)
846 struct throtl_service_queue *sq = &tg->service_queue;
847 bool rw = bio_data_dir(bio);
850 qn = &tg->qnode_on_self[rw];
853 * If @tg doesn't currently have any bios queued in the same
854 * direction, queueing @bio can change when @tg should be
855 * dispatched. Mark that @tg was empty. This is automatically
856 * cleaered on the next tg_update_disptime().
858 if (!sq->nr_queued[rw])
859 tg->flags |= THROTL_TG_WAS_EMPTY;
861 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
864 throtl_enqueue_tg(tg);
867 static void tg_update_disptime(struct throtl_grp *tg)
869 struct throtl_service_queue *sq = &tg->service_queue;
870 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
873 if ((bio = throtl_peek_queued(&sq->queued[READ])))
874 tg_may_dispatch(tg, bio, &read_wait);
876 if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
877 tg_may_dispatch(tg, bio, &write_wait);
879 min_wait = min(read_wait, write_wait);
880 disptime = jiffies + min_wait;
882 /* Update dispatch time */
883 throtl_dequeue_tg(tg);
884 tg->disptime = disptime;
885 throtl_enqueue_tg(tg);
887 /* see throtl_add_bio_tg() */
888 tg->flags &= ~THROTL_TG_WAS_EMPTY;
891 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
892 struct throtl_grp *parent_tg, bool rw)
894 if (throtl_slice_used(parent_tg, rw)) {
895 throtl_start_new_slice_with_credit(parent_tg, rw,
896 child_tg->slice_start[rw]);
901 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
903 struct throtl_service_queue *sq = &tg->service_queue;
904 struct throtl_service_queue *parent_sq = sq->parent_sq;
905 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
906 struct throtl_grp *tg_to_put = NULL;
910 * @bio is being transferred from @tg to @parent_sq. Popping a bio
911 * from @tg may put its reference and @parent_sq might end up
912 * getting released prematurely. Remember the tg to put and put it
913 * after @bio is transferred to @parent_sq.
915 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
918 throtl_charge_bio(tg, bio);
921 * If our parent is another tg, we just need to transfer @bio to
922 * the parent using throtl_add_bio_tg(). If our parent is
923 * @td->service_queue, @bio is ready to be issued. Put it on its
924 * bio_lists[] and decrease total number queued. The caller is
925 * responsible for issuing these bios.
928 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
929 start_parent_slice_with_credit(tg, parent_tg, rw);
931 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
932 &parent_sq->queued[rw]);
933 BUG_ON(tg->td->nr_queued[rw] <= 0);
934 tg->td->nr_queued[rw]--;
937 throtl_trim_slice(tg, rw);
940 blkg_put(tg_to_blkg(tg_to_put));
943 static int throtl_dispatch_tg(struct throtl_grp *tg)
945 struct throtl_service_queue *sq = &tg->service_queue;
946 unsigned int nr_reads = 0, nr_writes = 0;
947 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
948 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
951 /* Try to dispatch 75% READS and 25% WRITES */
953 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
954 tg_may_dispatch(tg, bio, NULL)) {
956 tg_dispatch_one_bio(tg, bio_data_dir(bio));
959 if (nr_reads >= max_nr_reads)
963 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
964 tg_may_dispatch(tg, bio, NULL)) {
966 tg_dispatch_one_bio(tg, bio_data_dir(bio));
969 if (nr_writes >= max_nr_writes)
973 return nr_reads + nr_writes;
976 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
978 unsigned int nr_disp = 0;
981 struct throtl_grp *tg = throtl_rb_first(parent_sq);
982 struct throtl_service_queue *sq = &tg->service_queue;
987 if (time_before(jiffies, tg->disptime))
990 throtl_dequeue_tg(tg);
992 nr_disp += throtl_dispatch_tg(tg);
994 if (sq->nr_queued[0] || sq->nr_queued[1])
995 tg_update_disptime(tg);
997 if (nr_disp >= throtl_quantum)
1005 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1006 * @arg: the throtl_service_queue being serviced
1008 * This timer is armed when a child throtl_grp with active bio's become
1009 * pending and queued on the service_queue's pending_tree and expires when
1010 * the first child throtl_grp should be dispatched. This function
1011 * dispatches bio's from the children throtl_grps to the parent
1014 * If the parent's parent is another throtl_grp, dispatching is propagated
1015 * by either arming its pending_timer or repeating dispatch directly. If
1016 * the top-level service_tree is reached, throtl_data->dispatch_work is
1017 * kicked so that the ready bio's are issued.
1019 static void throtl_pending_timer_fn(unsigned long arg)
1021 struct throtl_service_queue *sq = (void *)arg;
1022 struct throtl_grp *tg = sq_to_tg(sq);
1023 struct throtl_data *td = sq_to_td(sq);
1024 struct request_queue *q = td->queue;
1025 struct throtl_service_queue *parent_sq;
1029 spin_lock_irq(q->queue_lock);
1031 parent_sq = sq->parent_sq;
1035 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1036 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1037 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1039 ret = throtl_select_dispatch(sq);
1041 throtl_log(sq, "bios disp=%u", ret);
1045 if (throtl_schedule_next_dispatch(sq, false))
1048 /* this dispatch windows is still open, relax and repeat */
1049 spin_unlock_irq(q->queue_lock);
1051 spin_lock_irq(q->queue_lock);
1058 /* @parent_sq is another throl_grp, propagate dispatch */
1059 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1060 tg_update_disptime(tg);
1061 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1062 /* window is already open, repeat dispatching */
1069 /* reached the top-level, queue issueing */
1070 queue_work(kthrotld_workqueue, &td->dispatch_work);
1073 spin_unlock_irq(q->queue_lock);
1077 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1078 * @work: work item being executed
1080 * This function is queued for execution when bio's reach the bio_lists[]
1081 * of throtl_data->service_queue. Those bio's are ready and issued by this
1084 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1086 struct throtl_data *td = container_of(work, struct throtl_data,
1088 struct throtl_service_queue *td_sq = &td->service_queue;
1089 struct request_queue *q = td->queue;
1090 struct bio_list bio_list_on_stack;
1092 struct blk_plug plug;
1095 bio_list_init(&bio_list_on_stack);
1097 spin_lock_irq(q->queue_lock);
1098 for (rw = READ; rw <= WRITE; rw++)
1099 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1100 bio_list_add(&bio_list_on_stack, bio);
1101 spin_unlock_irq(q->queue_lock);
1103 if (!bio_list_empty(&bio_list_on_stack)) {
1104 blk_start_plug(&plug);
1105 while((bio = bio_list_pop(&bio_list_on_stack)))
1106 generic_make_request(bio);
1107 blk_finish_plug(&plug);
1111 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1114 struct throtl_grp *tg = pd_to_tg(pd);
1115 u64 v = *(u64 *)((void *)tg + off);
1119 return __blkg_prfill_u64(sf, pd, v);
1122 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1125 struct throtl_grp *tg = pd_to_tg(pd);
1126 unsigned int v = *(unsigned int *)((void *)tg + off);
1130 return __blkg_prfill_u64(sf, pd, v);
1133 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1135 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1136 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1140 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1142 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1143 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1147 static void tg_conf_updated(struct throtl_grp *tg)
1149 struct throtl_service_queue *sq = &tg->service_queue;
1150 struct cgroup_subsys_state *pos_css;
1151 struct blkcg_gq *blkg;
1153 throtl_log(&tg->service_queue,
1154 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1155 tg->bps[READ], tg->bps[WRITE],
1156 tg->iops[READ], tg->iops[WRITE]);
1159 * Update has_rules[] flags for the updated tg's subtree. A tg is
1160 * considered to have rules if either the tg itself or any of its
1161 * ancestors has rules. This identifies groups without any
1162 * restrictions in the whole hierarchy and allows them to bypass
1165 blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg))
1166 tg_update_has_rules(blkg_to_tg(blkg));
1169 * We're already holding queue_lock and know @tg is valid. Let's
1170 * apply the new config directly.
1172 * Restart the slices for both READ and WRITES. It might happen
1173 * that a group's limit are dropped suddenly and we don't want to
1174 * account recently dispatched IO with new low rate.
1176 throtl_start_new_slice(tg, 0);
1177 throtl_start_new_slice(tg, 1);
1179 if (tg->flags & THROTL_TG_PENDING) {
1180 tg_update_disptime(tg);
1181 throtl_schedule_next_dispatch(sq->parent_sq, true);
1185 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1186 char *buf, size_t nbytes, loff_t off, bool is_u64)
1188 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1189 struct blkg_conf_ctx ctx;
1190 struct throtl_grp *tg;
1194 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1199 if (sscanf(ctx.body, "%llu", &v) != 1)
1204 tg = blkg_to_tg(ctx.blkg);
1207 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1209 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1211 tg_conf_updated(tg);
1214 blkg_conf_finish(&ctx);
1215 return ret ?: nbytes;
1218 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1219 char *buf, size_t nbytes, loff_t off)
1221 return tg_set_conf(of, buf, nbytes, off, true);
1224 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1225 char *buf, size_t nbytes, loff_t off)
1227 return tg_set_conf(of, buf, nbytes, off, false);
1230 static struct cftype throtl_legacy_files[] = {
1232 .name = "throttle.read_bps_device",
1233 .private = offsetof(struct throtl_grp, bps[READ]),
1234 .seq_show = tg_print_conf_u64,
1235 .write = tg_set_conf_u64,
1238 .name = "throttle.write_bps_device",
1239 .private = offsetof(struct throtl_grp, bps[WRITE]),
1240 .seq_show = tg_print_conf_u64,
1241 .write = tg_set_conf_u64,
1244 .name = "throttle.read_iops_device",
1245 .private = offsetof(struct throtl_grp, iops[READ]),
1246 .seq_show = tg_print_conf_uint,
1247 .write = tg_set_conf_uint,
1250 .name = "throttle.write_iops_device",
1251 .private = offsetof(struct throtl_grp, iops[WRITE]),
1252 .seq_show = tg_print_conf_uint,
1253 .write = tg_set_conf_uint,
1256 .name = "throttle.io_service_bytes",
1257 .private = (unsigned long)&blkcg_policy_throtl,
1258 .seq_show = blkg_print_stat_bytes,
1261 .name = "throttle.io_serviced",
1262 .private = (unsigned long)&blkcg_policy_throtl,
1263 .seq_show = blkg_print_stat_ios,
1268 static u64 tg_prfill_max(struct seq_file *sf, struct blkg_policy_data *pd,
1271 struct throtl_grp *tg = pd_to_tg(pd);
1272 const char *dname = blkg_dev_name(pd->blkg);
1273 char bufs[4][21] = { "max", "max", "max", "max" };
1277 if (tg->bps[READ] == -1 && tg->bps[WRITE] == -1 &&
1278 tg->iops[READ] == -1 && tg->iops[WRITE] == -1)
1281 if (tg->bps[READ] != -1)
1282 snprintf(bufs[0], sizeof(bufs[0]), "%llu", tg->bps[READ]);
1283 if (tg->bps[WRITE] != -1)
1284 snprintf(bufs[1], sizeof(bufs[1]), "%llu", tg->bps[WRITE]);
1285 if (tg->iops[READ] != -1)
1286 snprintf(bufs[2], sizeof(bufs[2]), "%u", tg->iops[READ]);
1287 if (tg->iops[WRITE] != -1)
1288 snprintf(bufs[3], sizeof(bufs[3]), "%u", tg->iops[WRITE]);
1290 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s\n",
1291 dname, bufs[0], bufs[1], bufs[2], bufs[3]);
1295 static int tg_print_max(struct seq_file *sf, void *v)
1297 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_max,
1298 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1302 static ssize_t tg_set_max(struct kernfs_open_file *of,
1303 char *buf, size_t nbytes, loff_t off)
1305 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1306 struct blkg_conf_ctx ctx;
1307 struct throtl_grp *tg;
1311 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1315 tg = blkg_to_tg(ctx.blkg);
1317 v[0] = tg->bps[READ];
1318 v[1] = tg->bps[WRITE];
1319 v[2] = tg->iops[READ];
1320 v[3] = tg->iops[WRITE];
1323 char tok[27]; /* wiops=18446744073709551616 */
1328 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1337 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1345 if (!strcmp(tok, "rbps"))
1347 else if (!strcmp(tok, "wbps"))
1349 else if (!strcmp(tok, "riops"))
1350 v[2] = min_t(u64, val, UINT_MAX);
1351 else if (!strcmp(tok, "wiops"))
1352 v[3] = min_t(u64, val, UINT_MAX);
1357 tg->bps[READ] = v[0];
1358 tg->bps[WRITE] = v[1];
1359 tg->iops[READ] = v[2];
1360 tg->iops[WRITE] = v[3];
1362 tg_conf_updated(tg);
1365 blkg_conf_finish(&ctx);
1366 return ret ?: nbytes;
1369 static struct cftype throtl_files[] = {
1372 .flags = CFTYPE_NOT_ON_ROOT,
1373 .seq_show = tg_print_max,
1374 .write = tg_set_max,
1379 static void throtl_shutdown_wq(struct request_queue *q)
1381 struct throtl_data *td = q->td;
1383 cancel_work_sync(&td->dispatch_work);
1386 static struct blkcg_policy blkcg_policy_throtl = {
1387 .dfl_cftypes = throtl_files,
1388 .legacy_cftypes = throtl_legacy_files,
1390 .pd_alloc_fn = throtl_pd_alloc,
1391 .pd_init_fn = throtl_pd_init,
1392 .pd_online_fn = throtl_pd_online,
1393 .pd_free_fn = throtl_pd_free,
1396 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
1399 struct throtl_qnode *qn = NULL;
1400 struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
1401 struct throtl_service_queue *sq;
1402 bool rw = bio_data_dir(bio);
1403 bool throttled = false;
1405 WARN_ON_ONCE(!rcu_read_lock_held());
1407 /* see throtl_charge_bio() */
1408 if ((bio->bi_rw & REQ_THROTTLED) || !tg->has_rules[rw])
1411 spin_lock_irq(q->queue_lock);
1413 if (unlikely(blk_queue_bypass(q)))
1416 sq = &tg->service_queue;
1419 /* throtl is FIFO - if bios are already queued, should queue */
1420 if (sq->nr_queued[rw])
1423 /* if above limits, break to queue */
1424 if (!tg_may_dispatch(tg, bio, NULL))
1427 /* within limits, let's charge and dispatch directly */
1428 throtl_charge_bio(tg, bio);
1431 * We need to trim slice even when bios are not being queued
1432 * otherwise it might happen that a bio is not queued for
1433 * a long time and slice keeps on extending and trim is not
1434 * called for a long time. Now if limits are reduced suddenly
1435 * we take into account all the IO dispatched so far at new
1436 * low rate and * newly queued IO gets a really long dispatch
1439 * So keep on trimming slice even if bio is not queued.
1441 throtl_trim_slice(tg, rw);
1444 * @bio passed through this layer without being throttled.
1445 * Climb up the ladder. If we''re already at the top, it
1446 * can be executed directly.
1448 qn = &tg->qnode_on_parent[rw];
1455 /* out-of-limit, queue to @tg */
1456 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1457 rw == READ ? 'R' : 'W',
1458 tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw],
1459 tg->io_disp[rw], tg->iops[rw],
1460 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1462 bio_associate_current(bio);
1463 tg->td->nr_queued[rw]++;
1464 throtl_add_bio_tg(bio, qn, tg);
1468 * Update @tg's dispatch time and force schedule dispatch if @tg
1469 * was empty before @bio. The forced scheduling isn't likely to
1470 * cause undue delay as @bio is likely to be dispatched directly if
1471 * its @tg's disptime is not in the future.
1473 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1474 tg_update_disptime(tg);
1475 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1479 spin_unlock_irq(q->queue_lock);
1482 * As multiple blk-throtls may stack in the same issue path, we
1483 * don't want bios to leave with the flag set. Clear the flag if
1487 bio->bi_rw &= ~REQ_THROTTLED;
1492 * Dispatch all bios from all children tg's queued on @parent_sq. On
1493 * return, @parent_sq is guaranteed to not have any active children tg's
1494 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1496 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1498 struct throtl_grp *tg;
1500 while ((tg = throtl_rb_first(parent_sq))) {
1501 struct throtl_service_queue *sq = &tg->service_queue;
1504 throtl_dequeue_tg(tg);
1506 while ((bio = throtl_peek_queued(&sq->queued[READ])))
1507 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1508 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1509 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1514 * blk_throtl_drain - drain throttled bios
1515 * @q: request_queue to drain throttled bios for
1517 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1519 void blk_throtl_drain(struct request_queue *q)
1520 __releases(q->queue_lock) __acquires(q->queue_lock)
1522 struct throtl_data *td = q->td;
1523 struct blkcg_gq *blkg;
1524 struct cgroup_subsys_state *pos_css;
1528 queue_lockdep_assert_held(q);
1532 * Drain each tg while doing post-order walk on the blkg tree, so
1533 * that all bios are propagated to td->service_queue. It'd be
1534 * better to walk service_queue tree directly but blkg walk is
1537 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
1538 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1540 /* finally, transfer bios from top-level tg's into the td */
1541 tg_drain_bios(&td->service_queue);
1544 spin_unlock_irq(q->queue_lock);
1546 /* all bios now should be in td->service_queue, issue them */
1547 for (rw = READ; rw <= WRITE; rw++)
1548 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1550 generic_make_request(bio);
1552 spin_lock_irq(q->queue_lock);
1555 int blk_throtl_init(struct request_queue *q)
1557 struct throtl_data *td;
1560 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1564 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1565 throtl_service_queue_init(&td->service_queue);
1570 /* activate policy */
1571 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1577 void blk_throtl_exit(struct request_queue *q)
1580 throtl_shutdown_wq(q);
1581 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1585 static int __init throtl_init(void)
1587 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1588 if (!kthrotld_workqueue)
1589 panic("Failed to create kthrotld\n");
1591 return blkcg_policy_register(&blkcg_policy_throtl);
1594 module_init(throtl_init);