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 #define throtl_log(sq, fmt, args...) do { \
216 struct throtl_grp *__tg = sq_to_tg((sq)); \
217 struct throtl_data *__td = sq_to_td((sq)); \
220 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
225 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
226 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
228 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
232 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
234 INIT_LIST_HEAD(&qn->node);
235 bio_list_init(&qn->bios);
240 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
241 * @bio: bio being added
242 * @qn: qnode to add bio to
243 * @queued: the service_queue->queued[] list @qn belongs to
245 * Add @bio to @qn and put @qn on @queued if it's not already on.
246 * @qn->tg's reference count is bumped when @qn is activated. See the
247 * comment on top of throtl_qnode definition for details.
249 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
250 struct list_head *queued)
252 bio_list_add(&qn->bios, bio);
253 if (list_empty(&qn->node)) {
254 list_add_tail(&qn->node, queued);
255 blkg_get(tg_to_blkg(qn->tg));
260 * throtl_peek_queued - peek the first bio on a qnode list
261 * @queued: the qnode list to peek
263 static struct bio *throtl_peek_queued(struct list_head *queued)
265 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
268 if (list_empty(queued))
271 bio = bio_list_peek(&qn->bios);
277 * throtl_pop_queued - pop the first bio form a qnode list
278 * @queued: the qnode list to pop a bio from
279 * @tg_to_put: optional out argument for throtl_grp to put
281 * Pop the first bio from the qnode list @queued. After popping, the first
282 * qnode is removed from @queued if empty or moved to the end of @queued so
283 * that the popping order is round-robin.
285 * When the first qnode is removed, its associated throtl_grp should be put
286 * too. If @tg_to_put is NULL, this function automatically puts it;
287 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
288 * responsible for putting it.
290 static struct bio *throtl_pop_queued(struct list_head *queued,
291 struct throtl_grp **tg_to_put)
293 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
296 if (list_empty(queued))
299 bio = bio_list_pop(&qn->bios);
302 if (bio_list_empty(&qn->bios)) {
303 list_del_init(&qn->node);
307 blkg_put(tg_to_blkg(qn->tg));
309 list_move_tail(&qn->node, queued);
315 /* init a service_queue, assumes the caller zeroed it */
316 static void throtl_service_queue_init(struct throtl_service_queue *sq)
318 INIT_LIST_HEAD(&sq->queued[0]);
319 INIT_LIST_HEAD(&sq->queued[1]);
320 sq->pending_tree = RB_ROOT;
321 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
325 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
327 struct throtl_grp *tg;
330 tg = kzalloc_node(sizeof(*tg), gfp, node);
334 throtl_service_queue_init(&tg->service_queue);
336 for (rw = READ; rw <= WRITE; rw++) {
337 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
338 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
341 RB_CLEAR_NODE(&tg->rb_node);
345 tg->iops[WRITE] = -1;
350 static void throtl_pd_init(struct blkg_policy_data *pd)
352 struct throtl_grp *tg = pd_to_tg(pd);
353 struct blkcg_gq *blkg = tg_to_blkg(tg);
354 struct throtl_data *td = blkg->q->td;
355 struct throtl_service_queue *sq = &tg->service_queue;
358 * If on the default hierarchy, we switch to properly hierarchical
359 * behavior where limits on a given throtl_grp are applied to the
360 * whole subtree rather than just the group itself. e.g. If 16M
361 * read_bps limit is set on the root group, the whole system can't
362 * exceed 16M for the device.
364 * If not on the default hierarchy, the broken flat hierarchy
365 * behavior is retained where all throtl_grps are treated as if
366 * they're all separate root groups right below throtl_data.
367 * Limits of a group don't interact with limits of other groups
368 * regardless of the position of the group in the hierarchy.
370 sq->parent_sq = &td->service_queue;
371 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
372 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
377 * Set has_rules[] if @tg or any of its parents have limits configured.
378 * This doesn't require walking up to the top of the hierarchy as the
379 * parent's has_rules[] is guaranteed to be correct.
381 static void tg_update_has_rules(struct throtl_grp *tg)
383 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
386 for (rw = READ; rw <= WRITE; rw++)
387 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
388 (tg->bps[rw] != -1 || tg->iops[rw] != -1);
391 static void throtl_pd_online(struct blkg_policy_data *pd)
394 * We don't want new groups to escape the limits of its ancestors.
395 * Update has_rules[] after a new group is brought online.
397 tg_update_has_rules(pd_to_tg(pd));
400 static void throtl_pd_free(struct blkg_policy_data *pd)
402 struct throtl_grp *tg = pd_to_tg(pd);
404 del_timer_sync(&tg->service_queue.pending_timer);
408 static struct throtl_grp *
409 throtl_rb_first(struct throtl_service_queue *parent_sq)
411 /* Service tree is empty */
412 if (!parent_sq->nr_pending)
415 if (!parent_sq->first_pending)
416 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
418 if (parent_sq->first_pending)
419 return rb_entry_tg(parent_sq->first_pending);
424 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
430 static void throtl_rb_erase(struct rb_node *n,
431 struct throtl_service_queue *parent_sq)
433 if (parent_sq->first_pending == n)
434 parent_sq->first_pending = NULL;
435 rb_erase_init(n, &parent_sq->pending_tree);
436 --parent_sq->nr_pending;
439 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
441 struct throtl_grp *tg;
443 tg = throtl_rb_first(parent_sq);
447 parent_sq->first_pending_disptime = tg->disptime;
450 static void tg_service_queue_add(struct throtl_grp *tg)
452 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
453 struct rb_node **node = &parent_sq->pending_tree.rb_node;
454 struct rb_node *parent = NULL;
455 struct throtl_grp *__tg;
456 unsigned long key = tg->disptime;
459 while (*node != NULL) {
461 __tg = rb_entry_tg(parent);
463 if (time_before(key, __tg->disptime))
464 node = &parent->rb_left;
466 node = &parent->rb_right;
472 parent_sq->first_pending = &tg->rb_node;
474 rb_link_node(&tg->rb_node, parent, node);
475 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
478 static void __throtl_enqueue_tg(struct throtl_grp *tg)
480 tg_service_queue_add(tg);
481 tg->flags |= THROTL_TG_PENDING;
482 tg->service_queue.parent_sq->nr_pending++;
485 static void throtl_enqueue_tg(struct throtl_grp *tg)
487 if (!(tg->flags & THROTL_TG_PENDING))
488 __throtl_enqueue_tg(tg);
491 static void __throtl_dequeue_tg(struct throtl_grp *tg)
493 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
494 tg->flags &= ~THROTL_TG_PENDING;
497 static void throtl_dequeue_tg(struct throtl_grp *tg)
499 if (tg->flags & THROTL_TG_PENDING)
500 __throtl_dequeue_tg(tg);
503 /* Call with queue lock held */
504 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
505 unsigned long expires)
507 mod_timer(&sq->pending_timer, expires);
508 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
509 expires - jiffies, jiffies);
513 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
514 * @sq: the service_queue to schedule dispatch for
515 * @force: force scheduling
517 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
518 * dispatch time of the first pending child. Returns %true if either timer
519 * is armed or there's no pending child left. %false if the current
520 * dispatch window is still open and the caller should continue
523 * If @force is %true, the dispatch timer is always scheduled and this
524 * function is guaranteed to return %true. This is to be used when the
525 * caller can't dispatch itself and needs to invoke pending_timer
526 * unconditionally. Note that forced scheduling is likely to induce short
527 * delay before dispatch starts even if @sq->first_pending_disptime is not
528 * in the future and thus shouldn't be used in hot paths.
530 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
533 /* any pending children left? */
537 update_min_dispatch_time(sq);
539 /* is the next dispatch time in the future? */
540 if (force || time_after(sq->first_pending_disptime, jiffies)) {
541 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
545 /* tell the caller to continue dispatching */
549 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
550 bool rw, unsigned long start)
552 tg->bytes_disp[rw] = 0;
556 * Previous slice has expired. We must have trimmed it after last
557 * bio dispatch. That means since start of last slice, we never used
558 * that bandwidth. Do try to make use of that bandwidth while giving
561 if (time_after_eq(start, tg->slice_start[rw]))
562 tg->slice_start[rw] = start;
564 tg->slice_end[rw] = jiffies + throtl_slice;
565 throtl_log(&tg->service_queue,
566 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
567 rw == READ ? 'R' : 'W', tg->slice_start[rw],
568 tg->slice_end[rw], jiffies);
571 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
573 tg->bytes_disp[rw] = 0;
575 tg->slice_start[rw] = jiffies;
576 tg->slice_end[rw] = jiffies + throtl_slice;
577 throtl_log(&tg->service_queue,
578 "[%c] new slice start=%lu end=%lu jiffies=%lu",
579 rw == READ ? 'R' : 'W', tg->slice_start[rw],
580 tg->slice_end[rw], jiffies);
583 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
584 unsigned long jiffy_end)
586 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
589 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
590 unsigned long jiffy_end)
592 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
593 throtl_log(&tg->service_queue,
594 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
595 rw == READ ? 'R' : 'W', tg->slice_start[rw],
596 tg->slice_end[rw], jiffies);
599 /* Determine if previously allocated or extended slice is complete or not */
600 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
602 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
608 /* Trim the used slices and adjust slice start accordingly */
609 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
611 unsigned long nr_slices, time_elapsed, io_trim;
614 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
617 * If bps are unlimited (-1), then time slice don't get
618 * renewed. Don't try to trim the slice if slice is used. A new
619 * slice will start when appropriate.
621 if (throtl_slice_used(tg, rw))
625 * A bio has been dispatched. Also adjust slice_end. It might happen
626 * that initially cgroup limit was very low resulting in high
627 * slice_end, but later limit was bumped up and bio was dispached
628 * sooner, then we need to reduce slice_end. A high bogus slice_end
629 * is bad because it does not allow new slice to start.
632 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
634 time_elapsed = jiffies - tg->slice_start[rw];
636 nr_slices = time_elapsed / throtl_slice;
640 tmp = tg->bps[rw] * throtl_slice * nr_slices;
644 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
646 if (!bytes_trim && !io_trim)
649 if (tg->bytes_disp[rw] >= bytes_trim)
650 tg->bytes_disp[rw] -= bytes_trim;
652 tg->bytes_disp[rw] = 0;
654 if (tg->io_disp[rw] >= io_trim)
655 tg->io_disp[rw] -= io_trim;
659 tg->slice_start[rw] += nr_slices * throtl_slice;
661 throtl_log(&tg->service_queue,
662 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
663 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
664 tg->slice_start[rw], tg->slice_end[rw], jiffies);
667 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
670 bool rw = bio_data_dir(bio);
671 unsigned int io_allowed;
672 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
675 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
677 /* Slice has just started. Consider one slice interval */
679 jiffy_elapsed_rnd = throtl_slice;
681 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
684 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
685 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
686 * will allow dispatch after 1 second and after that slice should
690 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
694 io_allowed = UINT_MAX;
698 if (tg->io_disp[rw] + 1 <= io_allowed) {
704 /* Calc approx time to dispatch */
705 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
707 if (jiffy_wait > jiffy_elapsed)
708 jiffy_wait = jiffy_wait - jiffy_elapsed;
717 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
720 bool rw = bio_data_dir(bio);
721 u64 bytes_allowed, extra_bytes, tmp;
722 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
724 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
726 /* Slice has just started. Consider one slice interval */
728 jiffy_elapsed_rnd = throtl_slice;
730 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
732 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
736 if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
742 /* Calc approx time to dispatch */
743 extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
744 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
750 * This wait time is without taking into consideration the rounding
751 * up we did. Add that time also.
753 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
760 * Returns whether one can dispatch a bio or not. Also returns approx number
761 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
763 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
766 bool rw = bio_data_dir(bio);
767 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
770 * Currently whole state machine of group depends on first bio
771 * queued in the group bio list. So one should not be calling
772 * this function with a different bio if there are other bios
775 BUG_ON(tg->service_queue.nr_queued[rw] &&
776 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
778 /* If tg->bps = -1, then BW is unlimited */
779 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
786 * If previous slice expired, start a new one otherwise renew/extend
787 * existing slice to make sure it is at least throtl_slice interval
790 if (throtl_slice_used(tg, rw))
791 throtl_start_new_slice(tg, rw);
793 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
794 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
797 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
798 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
804 max_wait = max(bps_wait, iops_wait);
809 if (time_before(tg->slice_end[rw], jiffies + max_wait))
810 throtl_extend_slice(tg, rw, jiffies + max_wait);
815 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
817 bool rw = bio_data_dir(bio);
819 /* Charge the bio to the group */
820 tg->bytes_disp[rw] += bio->bi_iter.bi_size;
824 * REQ_THROTTLED is used to prevent the same bio to be throttled
825 * more than once as a throttled bio will go through blk-throtl the
826 * second time when it eventually gets issued. Set it when a bio
827 * is being charged to a tg.
829 if (!(bio->bi_rw & REQ_THROTTLED))
830 bio->bi_rw |= REQ_THROTTLED;
834 * throtl_add_bio_tg - add a bio to the specified throtl_grp
837 * @tg: the target throtl_grp
839 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
840 * tg->qnode_on_self[] is used.
842 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
843 struct throtl_grp *tg)
845 struct throtl_service_queue *sq = &tg->service_queue;
846 bool rw = bio_data_dir(bio);
849 qn = &tg->qnode_on_self[rw];
852 * If @tg doesn't currently have any bios queued in the same
853 * direction, queueing @bio can change when @tg should be
854 * dispatched. Mark that @tg was empty. This is automatically
855 * cleaered on the next tg_update_disptime().
857 if (!sq->nr_queued[rw])
858 tg->flags |= THROTL_TG_WAS_EMPTY;
860 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
863 throtl_enqueue_tg(tg);
866 static void tg_update_disptime(struct throtl_grp *tg)
868 struct throtl_service_queue *sq = &tg->service_queue;
869 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
872 if ((bio = throtl_peek_queued(&sq->queued[READ])))
873 tg_may_dispatch(tg, bio, &read_wait);
875 if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
876 tg_may_dispatch(tg, bio, &write_wait);
878 min_wait = min(read_wait, write_wait);
879 disptime = jiffies + min_wait;
881 /* Update dispatch time */
882 throtl_dequeue_tg(tg);
883 tg->disptime = disptime;
884 throtl_enqueue_tg(tg);
886 /* see throtl_add_bio_tg() */
887 tg->flags &= ~THROTL_TG_WAS_EMPTY;
890 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
891 struct throtl_grp *parent_tg, bool rw)
893 if (throtl_slice_used(parent_tg, rw)) {
894 throtl_start_new_slice_with_credit(parent_tg, rw,
895 child_tg->slice_start[rw]);
900 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
902 struct throtl_service_queue *sq = &tg->service_queue;
903 struct throtl_service_queue *parent_sq = sq->parent_sq;
904 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
905 struct throtl_grp *tg_to_put = NULL;
909 * @bio is being transferred from @tg to @parent_sq. Popping a bio
910 * from @tg may put its reference and @parent_sq might end up
911 * getting released prematurely. Remember the tg to put and put it
912 * after @bio is transferred to @parent_sq.
914 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
917 throtl_charge_bio(tg, bio);
920 * If our parent is another tg, we just need to transfer @bio to
921 * the parent using throtl_add_bio_tg(). If our parent is
922 * @td->service_queue, @bio is ready to be issued. Put it on its
923 * bio_lists[] and decrease total number queued. The caller is
924 * responsible for issuing these bios.
927 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
928 start_parent_slice_with_credit(tg, parent_tg, rw);
930 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
931 &parent_sq->queued[rw]);
932 BUG_ON(tg->td->nr_queued[rw] <= 0);
933 tg->td->nr_queued[rw]--;
936 throtl_trim_slice(tg, rw);
939 blkg_put(tg_to_blkg(tg_to_put));
942 static int throtl_dispatch_tg(struct throtl_grp *tg)
944 struct throtl_service_queue *sq = &tg->service_queue;
945 unsigned int nr_reads = 0, nr_writes = 0;
946 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
947 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
950 /* Try to dispatch 75% READS and 25% WRITES */
952 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
953 tg_may_dispatch(tg, bio, NULL)) {
955 tg_dispatch_one_bio(tg, bio_data_dir(bio));
958 if (nr_reads >= max_nr_reads)
962 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
963 tg_may_dispatch(tg, bio, NULL)) {
965 tg_dispatch_one_bio(tg, bio_data_dir(bio));
968 if (nr_writes >= max_nr_writes)
972 return nr_reads + nr_writes;
975 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
977 unsigned int nr_disp = 0;
980 struct throtl_grp *tg = throtl_rb_first(parent_sq);
981 struct throtl_service_queue *sq = &tg->service_queue;
986 if (time_before(jiffies, tg->disptime))
989 throtl_dequeue_tg(tg);
991 nr_disp += throtl_dispatch_tg(tg);
993 if (sq->nr_queued[0] || sq->nr_queued[1])
994 tg_update_disptime(tg);
996 if (nr_disp >= throtl_quantum)
1004 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1005 * @arg: the throtl_service_queue being serviced
1007 * This timer is armed when a child throtl_grp with active bio's become
1008 * pending and queued on the service_queue's pending_tree and expires when
1009 * the first child throtl_grp should be dispatched. This function
1010 * dispatches bio's from the children throtl_grps to the parent
1013 * If the parent's parent is another throtl_grp, dispatching is propagated
1014 * by either arming its pending_timer or repeating dispatch directly. If
1015 * the top-level service_tree is reached, throtl_data->dispatch_work is
1016 * kicked so that the ready bio's are issued.
1018 static void throtl_pending_timer_fn(unsigned long arg)
1020 struct throtl_service_queue *sq = (void *)arg;
1021 struct throtl_grp *tg = sq_to_tg(sq);
1022 struct throtl_data *td = sq_to_td(sq);
1023 struct request_queue *q = td->queue;
1024 struct throtl_service_queue *parent_sq;
1028 spin_lock_irq(q->queue_lock);
1030 parent_sq = sq->parent_sq;
1034 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1035 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1036 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1038 ret = throtl_select_dispatch(sq);
1040 throtl_log(sq, "bios disp=%u", ret);
1044 if (throtl_schedule_next_dispatch(sq, false))
1047 /* this dispatch windows is still open, relax and repeat */
1048 spin_unlock_irq(q->queue_lock);
1050 spin_lock_irq(q->queue_lock);
1057 /* @parent_sq is another throl_grp, propagate dispatch */
1058 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1059 tg_update_disptime(tg);
1060 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1061 /* window is already open, repeat dispatching */
1068 /* reached the top-level, queue issueing */
1069 queue_work(kthrotld_workqueue, &td->dispatch_work);
1072 spin_unlock_irq(q->queue_lock);
1076 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1077 * @work: work item being executed
1079 * This function is queued for execution when bio's reach the bio_lists[]
1080 * of throtl_data->service_queue. Those bio's are ready and issued by this
1083 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1085 struct throtl_data *td = container_of(work, struct throtl_data,
1087 struct throtl_service_queue *td_sq = &td->service_queue;
1088 struct request_queue *q = td->queue;
1089 struct bio_list bio_list_on_stack;
1091 struct blk_plug plug;
1094 bio_list_init(&bio_list_on_stack);
1096 spin_lock_irq(q->queue_lock);
1097 for (rw = READ; rw <= WRITE; rw++)
1098 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1099 bio_list_add(&bio_list_on_stack, bio);
1100 spin_unlock_irq(q->queue_lock);
1102 if (!bio_list_empty(&bio_list_on_stack)) {
1103 blk_start_plug(&plug);
1104 while((bio = bio_list_pop(&bio_list_on_stack)))
1105 generic_make_request(bio);
1106 blk_finish_plug(&plug);
1110 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1113 struct throtl_grp *tg = pd_to_tg(pd);
1114 u64 v = *(u64 *)((void *)tg + off);
1118 return __blkg_prfill_u64(sf, pd, v);
1121 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1124 struct throtl_grp *tg = pd_to_tg(pd);
1125 unsigned int v = *(unsigned int *)((void *)tg + off);
1129 return __blkg_prfill_u64(sf, pd, v);
1132 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1134 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1135 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1139 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1141 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1142 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1146 static void tg_conf_updated(struct throtl_grp *tg)
1148 struct throtl_service_queue *sq = &tg->service_queue;
1149 struct cgroup_subsys_state *pos_css;
1150 struct blkcg_gq *blkg;
1152 throtl_log(&tg->service_queue,
1153 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1154 tg->bps[READ], tg->bps[WRITE],
1155 tg->iops[READ], tg->iops[WRITE]);
1158 * Update has_rules[] flags for the updated tg's subtree. A tg is
1159 * considered to have rules if either the tg itself or any of its
1160 * ancestors has rules. This identifies groups without any
1161 * restrictions in the whole hierarchy and allows them to bypass
1164 blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg))
1165 tg_update_has_rules(blkg_to_tg(blkg));
1168 * We're already holding queue_lock and know @tg is valid. Let's
1169 * apply the new config directly.
1171 * Restart the slices for both READ and WRITES. It might happen
1172 * that a group's limit are dropped suddenly and we don't want to
1173 * account recently dispatched IO with new low rate.
1175 throtl_start_new_slice(tg, 0);
1176 throtl_start_new_slice(tg, 1);
1178 if (tg->flags & THROTL_TG_PENDING) {
1179 tg_update_disptime(tg);
1180 throtl_schedule_next_dispatch(sq->parent_sq, true);
1184 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1185 char *buf, size_t nbytes, loff_t off, bool is_u64)
1187 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1188 struct blkg_conf_ctx ctx;
1189 struct throtl_grp *tg;
1193 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1198 if (sscanf(ctx.body, "%llu", &v) != 1)
1203 tg = blkg_to_tg(ctx.blkg);
1206 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1208 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1210 tg_conf_updated(tg);
1213 blkg_conf_finish(&ctx);
1214 return ret ?: nbytes;
1217 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1218 char *buf, size_t nbytes, loff_t off)
1220 return tg_set_conf(of, buf, nbytes, off, true);
1223 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1224 char *buf, size_t nbytes, loff_t off)
1226 return tg_set_conf(of, buf, nbytes, off, false);
1229 static struct cftype throtl_legacy_files[] = {
1231 .name = "throttle.read_bps_device",
1232 .private = offsetof(struct throtl_grp, bps[READ]),
1233 .seq_show = tg_print_conf_u64,
1234 .write = tg_set_conf_u64,
1237 .name = "throttle.write_bps_device",
1238 .private = offsetof(struct throtl_grp, bps[WRITE]),
1239 .seq_show = tg_print_conf_u64,
1240 .write = tg_set_conf_u64,
1243 .name = "throttle.read_iops_device",
1244 .private = offsetof(struct throtl_grp, iops[READ]),
1245 .seq_show = tg_print_conf_uint,
1246 .write = tg_set_conf_uint,
1249 .name = "throttle.write_iops_device",
1250 .private = offsetof(struct throtl_grp, iops[WRITE]),
1251 .seq_show = tg_print_conf_uint,
1252 .write = tg_set_conf_uint,
1255 .name = "throttle.io_service_bytes",
1256 .private = (unsigned long)&blkcg_policy_throtl,
1257 .seq_show = blkg_print_stat_bytes,
1260 .name = "throttle.io_serviced",
1261 .private = (unsigned long)&blkcg_policy_throtl,
1262 .seq_show = blkg_print_stat_ios,
1267 static u64 tg_prfill_max(struct seq_file *sf, struct blkg_policy_data *pd,
1270 struct throtl_grp *tg = pd_to_tg(pd);
1271 const char *dname = blkg_dev_name(pd->blkg);
1272 char bufs[4][21] = { "max", "max", "max", "max" };
1276 if (tg->bps[READ] == -1 && tg->bps[WRITE] == -1 &&
1277 tg->iops[READ] == -1 && tg->iops[WRITE] == -1)
1280 if (tg->bps[READ] != -1)
1281 snprintf(bufs[0], sizeof(bufs[0]), "%llu", tg->bps[READ]);
1282 if (tg->bps[WRITE] != -1)
1283 snprintf(bufs[1], sizeof(bufs[1]), "%llu", tg->bps[WRITE]);
1284 if (tg->iops[READ] != -1)
1285 snprintf(bufs[2], sizeof(bufs[2]), "%u", tg->iops[READ]);
1286 if (tg->iops[WRITE] != -1)
1287 snprintf(bufs[3], sizeof(bufs[3]), "%u", tg->iops[WRITE]);
1289 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s\n",
1290 dname, bufs[0], bufs[1], bufs[2], bufs[3]);
1294 static int tg_print_max(struct seq_file *sf, void *v)
1296 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_max,
1297 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1301 static ssize_t tg_set_max(struct kernfs_open_file *of,
1302 char *buf, size_t nbytes, loff_t off)
1304 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1305 struct blkg_conf_ctx ctx;
1306 struct throtl_grp *tg;
1310 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1314 tg = blkg_to_tg(ctx.blkg);
1316 v[0] = tg->bps[READ];
1317 v[1] = tg->bps[WRITE];
1318 v[2] = tg->iops[READ];
1319 v[3] = tg->iops[WRITE];
1322 char tok[27]; /* wiops=18446744073709551616 */
1327 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1336 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1344 if (!strcmp(tok, "rbps"))
1346 else if (!strcmp(tok, "wbps"))
1348 else if (!strcmp(tok, "riops"))
1349 v[2] = min_t(u64, val, UINT_MAX);
1350 else if (!strcmp(tok, "wiops"))
1351 v[3] = min_t(u64, val, UINT_MAX);
1356 tg->bps[READ] = v[0];
1357 tg->bps[WRITE] = v[1];
1358 tg->iops[READ] = v[2];
1359 tg->iops[WRITE] = v[3];
1361 tg_conf_updated(tg);
1364 blkg_conf_finish(&ctx);
1365 return ret ?: nbytes;
1368 static struct cftype throtl_files[] = {
1371 .flags = CFTYPE_NOT_ON_ROOT,
1372 .seq_show = tg_print_max,
1373 .write = tg_set_max,
1378 static void throtl_shutdown_wq(struct request_queue *q)
1380 struct throtl_data *td = q->td;
1382 cancel_work_sync(&td->dispatch_work);
1385 static struct blkcg_policy blkcg_policy_throtl = {
1386 .dfl_cftypes = throtl_files,
1387 .legacy_cftypes = throtl_legacy_files,
1389 .pd_alloc_fn = throtl_pd_alloc,
1390 .pd_init_fn = throtl_pd_init,
1391 .pd_online_fn = throtl_pd_online,
1392 .pd_free_fn = throtl_pd_free,
1395 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
1398 struct throtl_qnode *qn = NULL;
1399 struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
1400 struct throtl_service_queue *sq;
1401 bool rw = bio_data_dir(bio);
1402 bool throttled = false;
1404 WARN_ON_ONCE(!rcu_read_lock_held());
1406 /* see throtl_charge_bio() */
1407 if ((bio->bi_rw & REQ_THROTTLED) || !tg->has_rules[rw])
1410 spin_lock_irq(q->queue_lock);
1412 if (unlikely(blk_queue_bypass(q)))
1415 sq = &tg->service_queue;
1418 /* throtl is FIFO - if bios are already queued, should queue */
1419 if (sq->nr_queued[rw])
1422 /* if above limits, break to queue */
1423 if (!tg_may_dispatch(tg, bio, NULL))
1426 /* within limits, let's charge and dispatch directly */
1427 throtl_charge_bio(tg, bio);
1430 * We need to trim slice even when bios are not being queued
1431 * otherwise it might happen that a bio is not queued for
1432 * a long time and slice keeps on extending and trim is not
1433 * called for a long time. Now if limits are reduced suddenly
1434 * we take into account all the IO dispatched so far at new
1435 * low rate and * newly queued IO gets a really long dispatch
1438 * So keep on trimming slice even if bio is not queued.
1440 throtl_trim_slice(tg, rw);
1443 * @bio passed through this layer without being throttled.
1444 * Climb up the ladder. If we''re already at the top, it
1445 * can be executed directly.
1447 qn = &tg->qnode_on_parent[rw];
1454 /* out-of-limit, queue to @tg */
1455 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1456 rw == READ ? 'R' : 'W',
1457 tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw],
1458 tg->io_disp[rw], tg->iops[rw],
1459 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1461 bio_associate_current(bio);
1462 tg->td->nr_queued[rw]++;
1463 throtl_add_bio_tg(bio, qn, tg);
1467 * Update @tg's dispatch time and force schedule dispatch if @tg
1468 * was empty before @bio. The forced scheduling isn't likely to
1469 * cause undue delay as @bio is likely to be dispatched directly if
1470 * its @tg's disptime is not in the future.
1472 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1473 tg_update_disptime(tg);
1474 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1478 spin_unlock_irq(q->queue_lock);
1481 * As multiple blk-throtls may stack in the same issue path, we
1482 * don't want bios to leave with the flag set. Clear the flag if
1486 bio->bi_rw &= ~REQ_THROTTLED;
1491 * Dispatch all bios from all children tg's queued on @parent_sq. On
1492 * return, @parent_sq is guaranteed to not have any active children tg's
1493 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1495 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1497 struct throtl_grp *tg;
1499 while ((tg = throtl_rb_first(parent_sq))) {
1500 struct throtl_service_queue *sq = &tg->service_queue;
1503 throtl_dequeue_tg(tg);
1505 while ((bio = throtl_peek_queued(&sq->queued[READ])))
1506 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1507 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1508 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1513 * blk_throtl_drain - drain throttled bios
1514 * @q: request_queue to drain throttled bios for
1516 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1518 void blk_throtl_drain(struct request_queue *q)
1519 __releases(q->queue_lock) __acquires(q->queue_lock)
1521 struct throtl_data *td = q->td;
1522 struct blkcg_gq *blkg;
1523 struct cgroup_subsys_state *pos_css;
1527 queue_lockdep_assert_held(q);
1531 * Drain each tg while doing post-order walk on the blkg tree, so
1532 * that all bios are propagated to td->service_queue. It'd be
1533 * better to walk service_queue tree directly but blkg walk is
1536 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
1537 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1539 /* finally, transfer bios from top-level tg's into the td */
1540 tg_drain_bios(&td->service_queue);
1543 spin_unlock_irq(q->queue_lock);
1545 /* all bios now should be in td->service_queue, issue them */
1546 for (rw = READ; rw <= WRITE; rw++)
1547 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1549 generic_make_request(bio);
1551 spin_lock_irq(q->queue_lock);
1554 int blk_throtl_init(struct request_queue *q)
1556 struct throtl_data *td;
1559 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1563 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1564 throtl_service_queue_init(&td->service_queue);
1569 /* activate policy */
1570 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1576 void blk_throtl_exit(struct request_queue *q)
1579 throtl_shutdown_wq(q);
1580 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1584 static int __init throtl_init(void)
1586 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1587 if (!kthrotld_workqueue)
1588 panic("Failed to create kthrotld\n");
1590 return blkcg_policy_register(&blkcg_policy_throtl);
1593 module_init(throtl_init);