1 // SPDX-License-Identifier: GPL-2.0
3 * Interface for controlling IO bandwidth on a request queue
5 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
8 #include <linux/module.h>
9 #include <linux/slab.h>
10 #include <linux/blkdev.h>
11 #include <linux/bio.h>
12 #include <linux/blktrace_api.h>
13 #include <linux/blk-cgroup.h>
15 #include "blk-cgroup-rwstat.h"
17 /* Max dispatch from a group in 1 round */
18 #define THROTL_GRP_QUANTUM 8
20 /* Total max dispatch from all groups in one round */
21 #define THROTL_QUANTUM 32
23 /* Throttling is performed over a slice and after that slice is renewed */
24 #define DFL_THROTL_SLICE_HD (HZ / 10)
25 #define DFL_THROTL_SLICE_SSD (HZ / 50)
26 #define MAX_THROTL_SLICE (HZ)
27 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
28 #define MIN_THROTL_BPS (320 * 1024)
29 #define MIN_THROTL_IOPS (10)
30 #define DFL_LATENCY_TARGET (-1L)
31 #define DFL_IDLE_THRESHOLD (0)
32 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
33 #define LATENCY_FILTERED_SSD (0)
35 * For HD, very small latency comes from sequential IO. Such IO is helpless to
36 * help determine if its IO is impacted by others, hence we ignore the IO
38 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
40 static struct blkcg_policy blkcg_policy_throtl;
42 /* A workqueue to queue throttle related work */
43 static struct workqueue_struct *kthrotld_workqueue;
46 * To implement hierarchical throttling, throtl_grps form a tree and bios
47 * are dispatched upwards level by level until they reach the top and get
48 * issued. When dispatching bios from the children and local group at each
49 * level, if the bios are dispatched into a single bio_list, there's a risk
50 * of a local or child group which can queue many bios at once filling up
51 * the list starving others.
53 * To avoid such starvation, dispatched bios are queued separately
54 * according to where they came from. When they are again dispatched to
55 * the parent, they're popped in round-robin order so that no single source
56 * hogs the dispatch window.
58 * throtl_qnode is used to keep the queued bios separated by their sources.
59 * Bios are queued to throtl_qnode which in turn is queued to
60 * throtl_service_queue and then dispatched in round-robin order.
62 * It's also used to track the reference counts on blkg's. A qnode always
63 * belongs to a throtl_grp and gets queued on itself or the parent, so
64 * incrementing the reference of the associated throtl_grp when a qnode is
65 * queued and decrementing when dequeued is enough to keep the whole blkg
66 * tree pinned while bios are in flight.
69 struct list_head node; /* service_queue->queued[] */
70 struct bio_list bios; /* queued bios */
71 struct throtl_grp *tg; /* tg this qnode belongs to */
74 struct throtl_service_queue {
75 struct throtl_service_queue *parent_sq; /* the parent service_queue */
78 * Bios queued directly to this service_queue or dispatched from
79 * children throtl_grp's.
81 struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */
82 unsigned int nr_queued[2]; /* number of queued bios */
85 * RB tree of active children throtl_grp's, which are sorted by
88 struct rb_root_cached pending_tree; /* RB tree of active tgs */
89 unsigned int nr_pending; /* # queued in the tree */
90 unsigned long first_pending_disptime; /* disptime of the first tg */
91 struct timer_list pending_timer; /* fires on first_pending_disptime */
95 THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */
96 THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */
99 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
108 /* must be the first member */
109 struct blkg_policy_data pd;
111 /* active throtl group service_queue member */
112 struct rb_node rb_node;
114 /* throtl_data this group belongs to */
115 struct throtl_data *td;
117 /* this group's service queue */
118 struct throtl_service_queue service_queue;
121 * qnode_on_self is used when bios are directly queued to this
122 * throtl_grp so that local bios compete fairly with bios
123 * dispatched from children. qnode_on_parent is used when bios are
124 * dispatched from this throtl_grp into its parent and will compete
125 * with the sibling qnode_on_parents and the parent's
128 struct throtl_qnode qnode_on_self[2];
129 struct throtl_qnode qnode_on_parent[2];
132 * Dispatch time in jiffies. This is the estimated time when group
133 * will unthrottle and is ready to dispatch more bio. It is used as
134 * key to sort active groups in service tree.
136 unsigned long disptime;
140 /* are there any throtl rules between this group and td? */
143 /* internally used bytes per second rate limits */
144 uint64_t bps[2][LIMIT_CNT];
145 /* user configured bps limits */
146 uint64_t bps_conf[2][LIMIT_CNT];
148 /* internally used IOPS limits */
149 unsigned int iops[2][LIMIT_CNT];
150 /* user configured IOPS limits */
151 unsigned int iops_conf[2][LIMIT_CNT];
153 /* Number of bytes dispatched in current slice */
154 uint64_t bytes_disp[2];
155 /* Number of bio's dispatched in current slice */
156 unsigned int io_disp[2];
158 unsigned long last_low_overflow_time[2];
160 uint64_t last_bytes_disp[2];
161 unsigned int last_io_disp[2];
163 unsigned long last_check_time;
165 unsigned long latency_target; /* us */
166 unsigned long latency_target_conf; /* us */
167 /* When did we start a new slice */
168 unsigned long slice_start[2];
169 unsigned long slice_end[2];
171 unsigned long last_finish_time; /* ns / 1024 */
172 unsigned long checked_last_finish_time; /* ns / 1024 */
173 unsigned long avg_idletime; /* ns / 1024 */
174 unsigned long idletime_threshold; /* us */
175 unsigned long idletime_threshold_conf; /* us */
177 unsigned int bio_cnt; /* total bios */
178 unsigned int bad_bio_cnt; /* bios exceeding latency threshold */
179 unsigned long bio_cnt_reset_time;
181 struct blkg_rwstat stat_bytes;
182 struct blkg_rwstat stat_ios;
185 /* We measure latency for request size from <= 4k to >= 1M */
186 #define LATENCY_BUCKET_SIZE 9
188 struct latency_bucket {
189 unsigned long total_latency; /* ns / 1024 */
193 struct avg_latency_bucket {
194 unsigned long latency; /* ns / 1024 */
200 /* service tree for active throtl groups */
201 struct throtl_service_queue service_queue;
203 struct request_queue *queue;
205 /* Total Number of queued bios on READ and WRITE lists */
206 unsigned int nr_queued[2];
208 unsigned int throtl_slice;
210 /* Work for dispatching throttled bios */
211 struct work_struct dispatch_work;
212 unsigned int limit_index;
213 bool limit_valid[LIMIT_CNT];
215 unsigned long low_upgrade_time;
216 unsigned long low_downgrade_time;
220 struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
221 struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
222 struct latency_bucket __percpu *latency_buckets[2];
223 unsigned long last_calculate_time;
224 unsigned long filtered_latency;
226 bool track_bio_latency;
229 static void throtl_pending_timer_fn(struct timer_list *t);
231 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
233 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
236 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
238 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
241 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
243 return pd_to_blkg(&tg->pd);
247 * sq_to_tg - return the throl_grp the specified service queue belongs to
248 * @sq: the throtl_service_queue of interest
250 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
251 * embedded in throtl_data, %NULL is returned.
253 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
255 if (sq && sq->parent_sq)
256 return container_of(sq, struct throtl_grp, service_queue);
262 * sq_to_td - return throtl_data the specified service queue belongs to
263 * @sq: the throtl_service_queue of interest
265 * A service_queue can be embedded in either a throtl_grp or throtl_data.
266 * Determine the associated throtl_data accordingly and return it.
268 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
270 struct throtl_grp *tg = sq_to_tg(sq);
275 return container_of(sq, struct throtl_data, service_queue);
279 * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
280 * make the IO dispatch more smooth.
281 * Scale up: linearly scale up according to lapsed time since upgrade. For
282 * every throtl_slice, the limit scales up 1/2 .low limit till the
283 * limit hits .max limit
284 * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
286 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
288 /* arbitrary value to avoid too big scale */
289 if (td->scale < 4096 && time_after_eq(jiffies,
290 td->low_upgrade_time + td->scale * td->throtl_slice))
291 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
293 return low + (low >> 1) * td->scale;
296 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
298 struct blkcg_gq *blkg = tg_to_blkg(tg);
299 struct throtl_data *td;
302 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
306 ret = tg->bps[rw][td->limit_index];
307 if (ret == 0 && td->limit_index == LIMIT_LOW) {
308 /* intermediate node or iops isn't 0 */
309 if (!list_empty(&blkg->blkcg->css.children) ||
310 tg->iops[rw][td->limit_index])
313 return MIN_THROTL_BPS;
316 if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
317 tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
320 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
321 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
326 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
328 struct blkcg_gq *blkg = tg_to_blkg(tg);
329 struct throtl_data *td;
332 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
336 ret = tg->iops[rw][td->limit_index];
337 if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
338 /* intermediate node or bps isn't 0 */
339 if (!list_empty(&blkg->blkcg->css.children) ||
340 tg->bps[rw][td->limit_index])
343 return MIN_THROTL_IOPS;
346 if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
347 tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
350 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
351 if (adjusted > UINT_MAX)
353 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
358 #define request_bucket_index(sectors) \
359 clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
362 * throtl_log - log debug message via blktrace
363 * @sq: the service_queue being reported
364 * @fmt: printf format string
367 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
368 * throtl_grp; otherwise, just "throtl".
370 #define throtl_log(sq, fmt, args...) do { \
371 struct throtl_grp *__tg = sq_to_tg((sq)); \
372 struct throtl_data *__td = sq_to_td((sq)); \
375 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
378 blk_add_cgroup_trace_msg(__td->queue, \
379 tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\
381 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
385 static inline unsigned int throtl_bio_data_size(struct bio *bio)
387 /* assume it's one sector */
388 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
390 return bio->bi_iter.bi_size;
393 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
395 INIT_LIST_HEAD(&qn->node);
396 bio_list_init(&qn->bios);
401 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
402 * @bio: bio being added
403 * @qn: qnode to add bio to
404 * @queued: the service_queue->queued[] list @qn belongs to
406 * Add @bio to @qn and put @qn on @queued if it's not already on.
407 * @qn->tg's reference count is bumped when @qn is activated. See the
408 * comment on top of throtl_qnode definition for details.
410 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
411 struct list_head *queued)
413 bio_list_add(&qn->bios, bio);
414 if (list_empty(&qn->node)) {
415 list_add_tail(&qn->node, queued);
416 blkg_get(tg_to_blkg(qn->tg));
421 * throtl_peek_queued - peek the first bio on a qnode list
422 * @queued: the qnode list to peek
424 static struct bio *throtl_peek_queued(struct list_head *queued)
426 struct throtl_qnode *qn;
429 if (list_empty(queued))
432 qn = list_first_entry(queued, struct throtl_qnode, node);
433 bio = bio_list_peek(&qn->bios);
439 * throtl_pop_queued - pop the first bio form a qnode list
440 * @queued: the qnode list to pop a bio from
441 * @tg_to_put: optional out argument for throtl_grp to put
443 * Pop the first bio from the qnode list @queued. After popping, the first
444 * qnode is removed from @queued if empty or moved to the end of @queued so
445 * that the popping order is round-robin.
447 * When the first qnode is removed, its associated throtl_grp should be put
448 * too. If @tg_to_put is NULL, this function automatically puts it;
449 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
450 * responsible for putting it.
452 static struct bio *throtl_pop_queued(struct list_head *queued,
453 struct throtl_grp **tg_to_put)
455 struct throtl_qnode *qn;
458 if (list_empty(queued))
461 qn = list_first_entry(queued, struct throtl_qnode, node);
462 bio = bio_list_pop(&qn->bios);
465 if (bio_list_empty(&qn->bios)) {
466 list_del_init(&qn->node);
470 blkg_put(tg_to_blkg(qn->tg));
472 list_move_tail(&qn->node, queued);
478 /* init a service_queue, assumes the caller zeroed it */
479 static void throtl_service_queue_init(struct throtl_service_queue *sq)
481 INIT_LIST_HEAD(&sq->queued[0]);
482 INIT_LIST_HEAD(&sq->queued[1]);
483 sq->pending_tree = RB_ROOT_CACHED;
484 timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
487 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp,
488 struct request_queue *q,
491 struct throtl_grp *tg;
494 tg = kzalloc_node(sizeof(*tg), gfp, q->node);
498 if (blkg_rwstat_init(&tg->stat_bytes, gfp))
501 if (blkg_rwstat_init(&tg->stat_ios, gfp))
502 goto err_exit_stat_bytes;
504 throtl_service_queue_init(&tg->service_queue);
506 for (rw = READ; rw <= WRITE; rw++) {
507 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
508 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
511 RB_CLEAR_NODE(&tg->rb_node);
512 tg->bps[READ][LIMIT_MAX] = U64_MAX;
513 tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
514 tg->iops[READ][LIMIT_MAX] = UINT_MAX;
515 tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
516 tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
517 tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
518 tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
519 tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
520 /* LIMIT_LOW will have default value 0 */
522 tg->latency_target = DFL_LATENCY_TARGET;
523 tg->latency_target_conf = DFL_LATENCY_TARGET;
524 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
525 tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
530 blkg_rwstat_exit(&tg->stat_bytes);
536 static void throtl_pd_init(struct blkg_policy_data *pd)
538 struct throtl_grp *tg = pd_to_tg(pd);
539 struct blkcg_gq *blkg = tg_to_blkg(tg);
540 struct throtl_data *td = blkg->q->td;
541 struct throtl_service_queue *sq = &tg->service_queue;
544 * If on the default hierarchy, we switch to properly hierarchical
545 * behavior where limits on a given throtl_grp are applied to the
546 * whole subtree rather than just the group itself. e.g. If 16M
547 * read_bps limit is set on the root group, the whole system can't
548 * exceed 16M for the device.
550 * If not on the default hierarchy, the broken flat hierarchy
551 * behavior is retained where all throtl_grps are treated as if
552 * they're all separate root groups right below throtl_data.
553 * Limits of a group don't interact with limits of other groups
554 * regardless of the position of the group in the hierarchy.
556 sq->parent_sq = &td->service_queue;
557 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
558 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
563 * Set has_rules[] if @tg or any of its parents have limits configured.
564 * This doesn't require walking up to the top of the hierarchy as the
565 * parent's has_rules[] is guaranteed to be correct.
567 static void tg_update_has_rules(struct throtl_grp *tg)
569 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
570 struct throtl_data *td = tg->td;
573 for (rw = READ; rw <= WRITE; rw++)
574 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
575 (td->limit_valid[td->limit_index] &&
576 (tg_bps_limit(tg, rw) != U64_MAX ||
577 tg_iops_limit(tg, rw) != UINT_MAX));
580 static void throtl_pd_online(struct blkg_policy_data *pd)
582 struct throtl_grp *tg = pd_to_tg(pd);
584 * We don't want new groups to escape the limits of its ancestors.
585 * Update has_rules[] after a new group is brought online.
587 tg_update_has_rules(tg);
590 static void blk_throtl_update_limit_valid(struct throtl_data *td)
592 struct cgroup_subsys_state *pos_css;
593 struct blkcg_gq *blkg;
594 bool low_valid = false;
597 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
598 struct throtl_grp *tg = blkg_to_tg(blkg);
600 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
601 tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
608 td->limit_valid[LIMIT_LOW] = low_valid;
611 static void throtl_upgrade_state(struct throtl_data *td);
612 static void throtl_pd_offline(struct blkg_policy_data *pd)
614 struct throtl_grp *tg = pd_to_tg(pd);
616 tg->bps[READ][LIMIT_LOW] = 0;
617 tg->bps[WRITE][LIMIT_LOW] = 0;
618 tg->iops[READ][LIMIT_LOW] = 0;
619 tg->iops[WRITE][LIMIT_LOW] = 0;
621 blk_throtl_update_limit_valid(tg->td);
623 if (!tg->td->limit_valid[tg->td->limit_index])
624 throtl_upgrade_state(tg->td);
627 static void throtl_pd_free(struct blkg_policy_data *pd)
629 struct throtl_grp *tg = pd_to_tg(pd);
631 del_timer_sync(&tg->service_queue.pending_timer);
632 blkg_rwstat_exit(&tg->stat_bytes);
633 blkg_rwstat_exit(&tg->stat_ios);
637 static struct throtl_grp *
638 throtl_rb_first(struct throtl_service_queue *parent_sq)
642 n = rb_first_cached(&parent_sq->pending_tree);
646 return rb_entry_tg(n);
649 static void throtl_rb_erase(struct rb_node *n,
650 struct throtl_service_queue *parent_sq)
652 rb_erase_cached(n, &parent_sq->pending_tree);
654 --parent_sq->nr_pending;
657 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
659 struct throtl_grp *tg;
661 tg = throtl_rb_first(parent_sq);
665 parent_sq->first_pending_disptime = tg->disptime;
668 static void tg_service_queue_add(struct throtl_grp *tg)
670 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
671 struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
672 struct rb_node *parent = NULL;
673 struct throtl_grp *__tg;
674 unsigned long key = tg->disptime;
675 bool leftmost = true;
677 while (*node != NULL) {
679 __tg = rb_entry_tg(parent);
681 if (time_before(key, __tg->disptime))
682 node = &parent->rb_left;
684 node = &parent->rb_right;
689 rb_link_node(&tg->rb_node, parent, node);
690 rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
694 static void __throtl_enqueue_tg(struct throtl_grp *tg)
696 tg_service_queue_add(tg);
697 tg->flags |= THROTL_TG_PENDING;
698 tg->service_queue.parent_sq->nr_pending++;
701 static void throtl_enqueue_tg(struct throtl_grp *tg)
703 if (!(tg->flags & THROTL_TG_PENDING))
704 __throtl_enqueue_tg(tg);
707 static void __throtl_dequeue_tg(struct throtl_grp *tg)
709 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
710 tg->flags &= ~THROTL_TG_PENDING;
713 static void throtl_dequeue_tg(struct throtl_grp *tg)
715 if (tg->flags & THROTL_TG_PENDING)
716 __throtl_dequeue_tg(tg);
719 /* Call with queue lock held */
720 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
721 unsigned long expires)
723 unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
726 * Since we are adjusting the throttle limit dynamically, the sleep
727 * time calculated according to previous limit might be invalid. It's
728 * possible the cgroup sleep time is very long and no other cgroups
729 * have IO running so notify the limit changes. Make sure the cgroup
730 * doesn't sleep too long to avoid the missed notification.
732 if (time_after(expires, max_expire))
733 expires = max_expire;
734 mod_timer(&sq->pending_timer, expires);
735 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
736 expires - jiffies, jiffies);
740 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
741 * @sq: the service_queue to schedule dispatch for
742 * @force: force scheduling
744 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
745 * dispatch time of the first pending child. Returns %true if either timer
746 * is armed or there's no pending child left. %false if the current
747 * dispatch window is still open and the caller should continue
750 * If @force is %true, the dispatch timer is always scheduled and this
751 * function is guaranteed to return %true. This is to be used when the
752 * caller can't dispatch itself and needs to invoke pending_timer
753 * unconditionally. Note that forced scheduling is likely to induce short
754 * delay before dispatch starts even if @sq->first_pending_disptime is not
755 * in the future and thus shouldn't be used in hot paths.
757 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
760 /* any pending children left? */
764 update_min_dispatch_time(sq);
766 /* is the next dispatch time in the future? */
767 if (force || time_after(sq->first_pending_disptime, jiffies)) {
768 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
772 /* tell the caller to continue dispatching */
776 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
777 bool rw, unsigned long start)
779 tg->bytes_disp[rw] = 0;
783 * Previous slice has expired. We must have trimmed it after last
784 * bio dispatch. That means since start of last slice, we never used
785 * that bandwidth. Do try to make use of that bandwidth while giving
788 if (time_after_eq(start, tg->slice_start[rw]))
789 tg->slice_start[rw] = start;
791 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
792 throtl_log(&tg->service_queue,
793 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
794 rw == READ ? 'R' : 'W', tg->slice_start[rw],
795 tg->slice_end[rw], jiffies);
798 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
800 tg->bytes_disp[rw] = 0;
802 tg->slice_start[rw] = jiffies;
803 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
804 throtl_log(&tg->service_queue,
805 "[%c] new slice start=%lu end=%lu jiffies=%lu",
806 rw == READ ? 'R' : 'W', tg->slice_start[rw],
807 tg->slice_end[rw], jiffies);
810 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
811 unsigned long jiffy_end)
813 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
816 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
817 unsigned long jiffy_end)
819 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
820 throtl_log(&tg->service_queue,
821 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
822 rw == READ ? 'R' : 'W', tg->slice_start[rw],
823 tg->slice_end[rw], jiffies);
826 /* Determine if previously allocated or extended slice is complete or not */
827 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
829 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
835 /* Trim the used slices and adjust slice start accordingly */
836 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
838 unsigned long nr_slices, time_elapsed, io_trim;
841 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
844 * If bps are unlimited (-1), then time slice don't get
845 * renewed. Don't try to trim the slice if slice is used. A new
846 * slice will start when appropriate.
848 if (throtl_slice_used(tg, rw))
852 * A bio has been dispatched. Also adjust slice_end. It might happen
853 * that initially cgroup limit was very low resulting in high
854 * slice_end, but later limit was bumped up and bio was dispatched
855 * sooner, then we need to reduce slice_end. A high bogus slice_end
856 * is bad because it does not allow new slice to start.
859 throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
861 time_elapsed = jiffies - tg->slice_start[rw];
863 nr_slices = time_elapsed / tg->td->throtl_slice;
867 tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
871 io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
874 if (!bytes_trim && !io_trim)
877 if (tg->bytes_disp[rw] >= bytes_trim)
878 tg->bytes_disp[rw] -= bytes_trim;
880 tg->bytes_disp[rw] = 0;
882 if (tg->io_disp[rw] >= io_trim)
883 tg->io_disp[rw] -= io_trim;
887 tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
889 throtl_log(&tg->service_queue,
890 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
891 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
892 tg->slice_start[rw], tg->slice_end[rw], jiffies);
895 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
896 u32 iops_limit, unsigned long *wait)
898 bool rw = bio_data_dir(bio);
899 unsigned int io_allowed;
900 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
903 if (iops_limit == UINT_MAX) {
909 jiffy_elapsed = jiffies - tg->slice_start[rw];
911 /* Round up to the next throttle slice, wait time must be nonzero */
912 jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
915 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
916 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
917 * will allow dispatch after 1 second and after that slice should
921 tmp = (u64)iops_limit * jiffy_elapsed_rnd;
925 io_allowed = UINT_MAX;
929 if (tg->io_disp[rw] + 1 <= io_allowed) {
935 /* Calc approx time to dispatch */
936 jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
943 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
944 u64 bps_limit, unsigned long *wait)
946 bool rw = bio_data_dir(bio);
947 u64 bytes_allowed, extra_bytes, tmp;
948 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
949 unsigned int bio_size = throtl_bio_data_size(bio);
951 if (bps_limit == U64_MAX) {
957 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
959 /* Slice has just started. Consider one slice interval */
961 jiffy_elapsed_rnd = tg->td->throtl_slice;
963 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
965 tmp = bps_limit * jiffy_elapsed_rnd;
969 if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
975 /* Calc approx time to dispatch */
976 extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
977 jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
983 * This wait time is without taking into consideration the rounding
984 * up we did. Add that time also.
986 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
993 * Returns whether one can dispatch a bio or not. Also returns approx number
994 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
996 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
999 bool rw = bio_data_dir(bio);
1000 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
1001 u64 bps_limit = tg_bps_limit(tg, rw);
1002 u32 iops_limit = tg_iops_limit(tg, rw);
1005 * Currently whole state machine of group depends on first bio
1006 * queued in the group bio list. So one should not be calling
1007 * this function with a different bio if there are other bios
1010 BUG_ON(tg->service_queue.nr_queued[rw] &&
1011 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
1013 /* If tg->bps = -1, then BW is unlimited */
1014 if (bps_limit == U64_MAX && iops_limit == UINT_MAX) {
1021 * If previous slice expired, start a new one otherwise renew/extend
1022 * existing slice to make sure it is at least throtl_slice interval
1023 * long since now. New slice is started only for empty throttle group.
1024 * If there is queued bio, that means there should be an active
1025 * slice and it should be extended instead.
1027 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
1028 throtl_start_new_slice(tg, rw);
1030 if (time_before(tg->slice_end[rw],
1031 jiffies + tg->td->throtl_slice))
1032 throtl_extend_slice(tg, rw,
1033 jiffies + tg->td->throtl_slice);
1036 if (tg_with_in_bps_limit(tg, bio, bps_limit, &bps_wait) &&
1037 tg_with_in_iops_limit(tg, bio, iops_limit, &iops_wait)) {
1043 max_wait = max(bps_wait, iops_wait);
1048 if (time_before(tg->slice_end[rw], jiffies + max_wait))
1049 throtl_extend_slice(tg, rw, jiffies + max_wait);
1054 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1056 bool rw = bio_data_dir(bio);
1057 unsigned int bio_size = throtl_bio_data_size(bio);
1059 /* Charge the bio to the group */
1060 tg->bytes_disp[rw] += bio_size;
1062 tg->last_bytes_disp[rw] += bio_size;
1063 tg->last_io_disp[rw]++;
1066 * BIO_THROTTLED is used to prevent the same bio to be throttled
1067 * more than once as a throttled bio will go through blk-throtl the
1068 * second time when it eventually gets issued. Set it when a bio
1069 * is being charged to a tg.
1071 if (!bio_flagged(bio, BIO_THROTTLED))
1072 bio_set_flag(bio, BIO_THROTTLED);
1076 * throtl_add_bio_tg - add a bio to the specified throtl_grp
1079 * @tg: the target throtl_grp
1081 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
1082 * tg->qnode_on_self[] is used.
1084 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1085 struct throtl_grp *tg)
1087 struct throtl_service_queue *sq = &tg->service_queue;
1088 bool rw = bio_data_dir(bio);
1091 qn = &tg->qnode_on_self[rw];
1094 * If @tg doesn't currently have any bios queued in the same
1095 * direction, queueing @bio can change when @tg should be
1096 * dispatched. Mark that @tg was empty. This is automatically
1097 * cleared on the next tg_update_disptime().
1099 if (!sq->nr_queued[rw])
1100 tg->flags |= THROTL_TG_WAS_EMPTY;
1102 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1104 sq->nr_queued[rw]++;
1105 throtl_enqueue_tg(tg);
1108 static void tg_update_disptime(struct throtl_grp *tg)
1110 struct throtl_service_queue *sq = &tg->service_queue;
1111 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1114 bio = throtl_peek_queued(&sq->queued[READ]);
1116 tg_may_dispatch(tg, bio, &read_wait);
1118 bio = throtl_peek_queued(&sq->queued[WRITE]);
1120 tg_may_dispatch(tg, bio, &write_wait);
1122 min_wait = min(read_wait, write_wait);
1123 disptime = jiffies + min_wait;
1125 /* Update dispatch time */
1126 throtl_dequeue_tg(tg);
1127 tg->disptime = disptime;
1128 throtl_enqueue_tg(tg);
1130 /* see throtl_add_bio_tg() */
1131 tg->flags &= ~THROTL_TG_WAS_EMPTY;
1134 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1135 struct throtl_grp *parent_tg, bool rw)
1137 if (throtl_slice_used(parent_tg, rw)) {
1138 throtl_start_new_slice_with_credit(parent_tg, rw,
1139 child_tg->slice_start[rw]);
1144 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1146 struct throtl_service_queue *sq = &tg->service_queue;
1147 struct throtl_service_queue *parent_sq = sq->parent_sq;
1148 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1149 struct throtl_grp *tg_to_put = NULL;
1153 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1154 * from @tg may put its reference and @parent_sq might end up
1155 * getting released prematurely. Remember the tg to put and put it
1156 * after @bio is transferred to @parent_sq.
1158 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1159 sq->nr_queued[rw]--;
1161 throtl_charge_bio(tg, bio);
1164 * If our parent is another tg, we just need to transfer @bio to
1165 * the parent using throtl_add_bio_tg(). If our parent is
1166 * @td->service_queue, @bio is ready to be issued. Put it on its
1167 * bio_lists[] and decrease total number queued. The caller is
1168 * responsible for issuing these bios.
1171 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1172 start_parent_slice_with_credit(tg, parent_tg, rw);
1174 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1175 &parent_sq->queued[rw]);
1176 BUG_ON(tg->td->nr_queued[rw] <= 0);
1177 tg->td->nr_queued[rw]--;
1180 throtl_trim_slice(tg, rw);
1183 blkg_put(tg_to_blkg(tg_to_put));
1186 static int throtl_dispatch_tg(struct throtl_grp *tg)
1188 struct throtl_service_queue *sq = &tg->service_queue;
1189 unsigned int nr_reads = 0, nr_writes = 0;
1190 unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1191 unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1194 /* Try to dispatch 75% READS and 25% WRITES */
1196 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1197 tg_may_dispatch(tg, bio, NULL)) {
1199 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1202 if (nr_reads >= max_nr_reads)
1206 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1207 tg_may_dispatch(tg, bio, NULL)) {
1209 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1212 if (nr_writes >= max_nr_writes)
1216 return nr_reads + nr_writes;
1219 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1221 unsigned int nr_disp = 0;
1224 struct throtl_grp *tg;
1225 struct throtl_service_queue *sq;
1227 if (!parent_sq->nr_pending)
1230 tg = throtl_rb_first(parent_sq);
1234 if (time_before(jiffies, tg->disptime))
1237 throtl_dequeue_tg(tg);
1239 nr_disp += throtl_dispatch_tg(tg);
1241 sq = &tg->service_queue;
1242 if (sq->nr_queued[0] || sq->nr_queued[1])
1243 tg_update_disptime(tg);
1245 if (nr_disp >= THROTL_QUANTUM)
1252 static bool throtl_can_upgrade(struct throtl_data *td,
1253 struct throtl_grp *this_tg);
1255 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1256 * @t: the pending_timer member of the throtl_service_queue being serviced
1258 * This timer is armed when a child throtl_grp with active bio's become
1259 * pending and queued on the service_queue's pending_tree and expires when
1260 * the first child throtl_grp should be dispatched. This function
1261 * dispatches bio's from the children throtl_grps to the parent
1264 * If the parent's parent is another throtl_grp, dispatching is propagated
1265 * by either arming its pending_timer or repeating dispatch directly. If
1266 * the top-level service_tree is reached, throtl_data->dispatch_work is
1267 * kicked so that the ready bio's are issued.
1269 static void throtl_pending_timer_fn(struct timer_list *t)
1271 struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1272 struct throtl_grp *tg = sq_to_tg(sq);
1273 struct throtl_data *td = sq_to_td(sq);
1274 struct request_queue *q = td->queue;
1275 struct throtl_service_queue *parent_sq;
1279 spin_lock_irq(&q->queue_lock);
1280 if (throtl_can_upgrade(td, NULL))
1281 throtl_upgrade_state(td);
1284 parent_sq = sq->parent_sq;
1288 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1289 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1290 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1292 ret = throtl_select_dispatch(sq);
1294 throtl_log(sq, "bios disp=%u", ret);
1298 if (throtl_schedule_next_dispatch(sq, false))
1301 /* this dispatch windows is still open, relax and repeat */
1302 spin_unlock_irq(&q->queue_lock);
1304 spin_lock_irq(&q->queue_lock);
1311 /* @parent_sq is another throl_grp, propagate dispatch */
1312 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1313 tg_update_disptime(tg);
1314 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1315 /* window is already open, repeat dispatching */
1322 /* reached the top-level, queue issuing */
1323 queue_work(kthrotld_workqueue, &td->dispatch_work);
1326 spin_unlock_irq(&q->queue_lock);
1330 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1331 * @work: work item being executed
1333 * This function is queued for execution when bios reach the bio_lists[]
1334 * of throtl_data->service_queue. Those bios are ready and issued by this
1337 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1339 struct throtl_data *td = container_of(work, struct throtl_data,
1341 struct throtl_service_queue *td_sq = &td->service_queue;
1342 struct request_queue *q = td->queue;
1343 struct bio_list bio_list_on_stack;
1345 struct blk_plug plug;
1348 bio_list_init(&bio_list_on_stack);
1350 spin_lock_irq(&q->queue_lock);
1351 for (rw = READ; rw <= WRITE; rw++)
1352 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1353 bio_list_add(&bio_list_on_stack, bio);
1354 spin_unlock_irq(&q->queue_lock);
1356 if (!bio_list_empty(&bio_list_on_stack)) {
1357 blk_start_plug(&plug);
1358 while ((bio = bio_list_pop(&bio_list_on_stack)))
1359 submit_bio_noacct(bio);
1360 blk_finish_plug(&plug);
1364 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1367 struct throtl_grp *tg = pd_to_tg(pd);
1368 u64 v = *(u64 *)((void *)tg + off);
1372 return __blkg_prfill_u64(sf, pd, v);
1375 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1378 struct throtl_grp *tg = pd_to_tg(pd);
1379 unsigned int v = *(unsigned int *)((void *)tg + off);
1383 return __blkg_prfill_u64(sf, pd, v);
1386 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1388 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1389 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1393 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1395 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1396 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1400 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1402 struct throtl_service_queue *sq = &tg->service_queue;
1403 struct cgroup_subsys_state *pos_css;
1404 struct blkcg_gq *blkg;
1406 throtl_log(&tg->service_queue,
1407 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1408 tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1409 tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1412 * Update has_rules[] flags for the updated tg's subtree. A tg is
1413 * considered to have rules if either the tg itself or any of its
1414 * ancestors has rules. This identifies groups without any
1415 * restrictions in the whole hierarchy and allows them to bypass
1418 blkg_for_each_descendant_pre(blkg, pos_css,
1419 global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1420 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1421 struct throtl_grp *parent_tg;
1423 tg_update_has_rules(this_tg);
1424 /* ignore root/second level */
1425 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1426 !blkg->parent->parent)
1428 parent_tg = blkg_to_tg(blkg->parent);
1430 * make sure all children has lower idle time threshold and
1431 * higher latency target
1433 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1434 parent_tg->idletime_threshold);
1435 this_tg->latency_target = max(this_tg->latency_target,
1436 parent_tg->latency_target);
1440 * We're already holding queue_lock and know @tg is valid. Let's
1441 * apply the new config directly.
1443 * Restart the slices for both READ and WRITES. It might happen
1444 * that a group's limit are dropped suddenly and we don't want to
1445 * account recently dispatched IO with new low rate.
1447 throtl_start_new_slice(tg, READ);
1448 throtl_start_new_slice(tg, WRITE);
1450 if (tg->flags & THROTL_TG_PENDING) {
1451 tg_update_disptime(tg);
1452 throtl_schedule_next_dispatch(sq->parent_sq, true);
1456 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1457 char *buf, size_t nbytes, loff_t off, bool is_u64)
1459 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1460 struct blkg_conf_ctx ctx;
1461 struct throtl_grp *tg;
1465 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1470 if (sscanf(ctx.body, "%llu", &v) != 1)
1475 tg = blkg_to_tg(ctx.blkg);
1478 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1480 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1482 tg_conf_updated(tg, false);
1485 blkg_conf_finish(&ctx);
1486 return ret ?: nbytes;
1489 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1490 char *buf, size_t nbytes, loff_t off)
1492 return tg_set_conf(of, buf, nbytes, off, true);
1495 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1496 char *buf, size_t nbytes, loff_t off)
1498 return tg_set_conf(of, buf, nbytes, off, false);
1501 static int tg_print_rwstat(struct seq_file *sf, void *v)
1503 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1504 blkg_prfill_rwstat, &blkcg_policy_throtl,
1505 seq_cft(sf)->private, true);
1509 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1510 struct blkg_policy_data *pd, int off)
1512 struct blkg_rwstat_sample sum;
1514 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1516 return __blkg_prfill_rwstat(sf, pd, &sum);
1519 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1521 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1522 tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1523 seq_cft(sf)->private, true);
1527 static struct cftype throtl_legacy_files[] = {
1529 .name = "throttle.read_bps_device",
1530 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1531 .seq_show = tg_print_conf_u64,
1532 .write = tg_set_conf_u64,
1535 .name = "throttle.write_bps_device",
1536 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1537 .seq_show = tg_print_conf_u64,
1538 .write = tg_set_conf_u64,
1541 .name = "throttle.read_iops_device",
1542 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1543 .seq_show = tg_print_conf_uint,
1544 .write = tg_set_conf_uint,
1547 .name = "throttle.write_iops_device",
1548 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1549 .seq_show = tg_print_conf_uint,
1550 .write = tg_set_conf_uint,
1553 .name = "throttle.io_service_bytes",
1554 .private = offsetof(struct throtl_grp, stat_bytes),
1555 .seq_show = tg_print_rwstat,
1558 .name = "throttle.io_service_bytes_recursive",
1559 .private = offsetof(struct throtl_grp, stat_bytes),
1560 .seq_show = tg_print_rwstat_recursive,
1563 .name = "throttle.io_serviced",
1564 .private = offsetof(struct throtl_grp, stat_ios),
1565 .seq_show = tg_print_rwstat,
1568 .name = "throttle.io_serviced_recursive",
1569 .private = offsetof(struct throtl_grp, stat_ios),
1570 .seq_show = tg_print_rwstat_recursive,
1575 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1578 struct throtl_grp *tg = pd_to_tg(pd);
1579 const char *dname = blkg_dev_name(pd->blkg);
1580 char bufs[4][21] = { "max", "max", "max", "max" };
1582 unsigned int iops_dft;
1583 char idle_time[26] = "";
1584 char latency_time[26] = "";
1589 if (off == LIMIT_LOW) {
1594 iops_dft = UINT_MAX;
1597 if (tg->bps_conf[READ][off] == bps_dft &&
1598 tg->bps_conf[WRITE][off] == bps_dft &&
1599 tg->iops_conf[READ][off] == iops_dft &&
1600 tg->iops_conf[WRITE][off] == iops_dft &&
1601 (off != LIMIT_LOW ||
1602 (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1603 tg->latency_target_conf == DFL_LATENCY_TARGET)))
1606 if (tg->bps_conf[READ][off] != U64_MAX)
1607 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1608 tg->bps_conf[READ][off]);
1609 if (tg->bps_conf[WRITE][off] != U64_MAX)
1610 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1611 tg->bps_conf[WRITE][off]);
1612 if (tg->iops_conf[READ][off] != UINT_MAX)
1613 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1614 tg->iops_conf[READ][off]);
1615 if (tg->iops_conf[WRITE][off] != UINT_MAX)
1616 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1617 tg->iops_conf[WRITE][off]);
1618 if (off == LIMIT_LOW) {
1619 if (tg->idletime_threshold_conf == ULONG_MAX)
1620 strcpy(idle_time, " idle=max");
1622 snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1623 tg->idletime_threshold_conf);
1625 if (tg->latency_target_conf == ULONG_MAX)
1626 strcpy(latency_time, " latency=max");
1628 snprintf(latency_time, sizeof(latency_time),
1629 " latency=%lu", tg->latency_target_conf);
1632 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1633 dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1638 static int tg_print_limit(struct seq_file *sf, void *v)
1640 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1641 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1645 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1646 char *buf, size_t nbytes, loff_t off)
1648 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1649 struct blkg_conf_ctx ctx;
1650 struct throtl_grp *tg;
1652 unsigned long idle_time;
1653 unsigned long latency_time;
1655 int index = of_cft(of)->private;
1657 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1661 tg = blkg_to_tg(ctx.blkg);
1663 v[0] = tg->bps_conf[READ][index];
1664 v[1] = tg->bps_conf[WRITE][index];
1665 v[2] = tg->iops_conf[READ][index];
1666 v[3] = tg->iops_conf[WRITE][index];
1668 idle_time = tg->idletime_threshold_conf;
1669 latency_time = tg->latency_target_conf;
1671 char tok[27]; /* wiops=18446744073709551616 */
1676 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1685 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1693 if (!strcmp(tok, "rbps") && val > 1)
1695 else if (!strcmp(tok, "wbps") && val > 1)
1697 else if (!strcmp(tok, "riops") && val > 1)
1698 v[2] = min_t(u64, val, UINT_MAX);
1699 else if (!strcmp(tok, "wiops") && val > 1)
1700 v[3] = min_t(u64, val, UINT_MAX);
1701 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1703 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1709 tg->bps_conf[READ][index] = v[0];
1710 tg->bps_conf[WRITE][index] = v[1];
1711 tg->iops_conf[READ][index] = v[2];
1712 tg->iops_conf[WRITE][index] = v[3];
1714 if (index == LIMIT_MAX) {
1715 tg->bps[READ][index] = v[0];
1716 tg->bps[WRITE][index] = v[1];
1717 tg->iops[READ][index] = v[2];
1718 tg->iops[WRITE][index] = v[3];
1720 tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1721 tg->bps_conf[READ][LIMIT_MAX]);
1722 tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1723 tg->bps_conf[WRITE][LIMIT_MAX]);
1724 tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1725 tg->iops_conf[READ][LIMIT_MAX]);
1726 tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1727 tg->iops_conf[WRITE][LIMIT_MAX]);
1728 tg->idletime_threshold_conf = idle_time;
1729 tg->latency_target_conf = latency_time;
1731 /* force user to configure all settings for low limit */
1732 if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1733 tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1734 tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1735 tg->latency_target_conf == DFL_LATENCY_TARGET) {
1736 tg->bps[READ][LIMIT_LOW] = 0;
1737 tg->bps[WRITE][LIMIT_LOW] = 0;
1738 tg->iops[READ][LIMIT_LOW] = 0;
1739 tg->iops[WRITE][LIMIT_LOW] = 0;
1740 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1741 tg->latency_target = DFL_LATENCY_TARGET;
1742 } else if (index == LIMIT_LOW) {
1743 tg->idletime_threshold = tg->idletime_threshold_conf;
1744 tg->latency_target = tg->latency_target_conf;
1747 blk_throtl_update_limit_valid(tg->td);
1748 if (tg->td->limit_valid[LIMIT_LOW]) {
1749 if (index == LIMIT_LOW)
1750 tg->td->limit_index = LIMIT_LOW;
1752 tg->td->limit_index = LIMIT_MAX;
1753 tg_conf_updated(tg, index == LIMIT_LOW &&
1754 tg->td->limit_valid[LIMIT_LOW]);
1757 blkg_conf_finish(&ctx);
1758 return ret ?: nbytes;
1761 static struct cftype throtl_files[] = {
1762 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1765 .flags = CFTYPE_NOT_ON_ROOT,
1766 .seq_show = tg_print_limit,
1767 .write = tg_set_limit,
1768 .private = LIMIT_LOW,
1773 .flags = CFTYPE_NOT_ON_ROOT,
1774 .seq_show = tg_print_limit,
1775 .write = tg_set_limit,
1776 .private = LIMIT_MAX,
1781 static void throtl_shutdown_wq(struct request_queue *q)
1783 struct throtl_data *td = q->td;
1785 cancel_work_sync(&td->dispatch_work);
1788 static struct blkcg_policy blkcg_policy_throtl = {
1789 .dfl_cftypes = throtl_files,
1790 .legacy_cftypes = throtl_legacy_files,
1792 .pd_alloc_fn = throtl_pd_alloc,
1793 .pd_init_fn = throtl_pd_init,
1794 .pd_online_fn = throtl_pd_online,
1795 .pd_offline_fn = throtl_pd_offline,
1796 .pd_free_fn = throtl_pd_free,
1799 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1801 unsigned long rtime = jiffies, wtime = jiffies;
1803 if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1804 rtime = tg->last_low_overflow_time[READ];
1805 if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1806 wtime = tg->last_low_overflow_time[WRITE];
1807 return min(rtime, wtime);
1810 /* tg should not be an intermediate node */
1811 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1813 struct throtl_service_queue *parent_sq;
1814 struct throtl_grp *parent = tg;
1815 unsigned long ret = __tg_last_low_overflow_time(tg);
1818 parent_sq = parent->service_queue.parent_sq;
1819 parent = sq_to_tg(parent_sq);
1824 * The parent doesn't have low limit, it always reaches low
1825 * limit. Its overflow time is useless for children
1827 if (!parent->bps[READ][LIMIT_LOW] &&
1828 !parent->iops[READ][LIMIT_LOW] &&
1829 !parent->bps[WRITE][LIMIT_LOW] &&
1830 !parent->iops[WRITE][LIMIT_LOW])
1832 if (time_after(__tg_last_low_overflow_time(parent), ret))
1833 ret = __tg_last_low_overflow_time(parent);
1838 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1841 * cgroup is idle if:
1842 * - single idle is too long, longer than a fixed value (in case user
1843 * configure a too big threshold) or 4 times of idletime threshold
1844 * - average think time is more than threshold
1845 * - IO latency is largely below threshold
1850 time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1851 ret = tg->latency_target == DFL_LATENCY_TARGET ||
1852 tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1853 (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1854 tg->avg_idletime > tg->idletime_threshold ||
1855 (tg->latency_target && tg->bio_cnt &&
1856 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1857 throtl_log(&tg->service_queue,
1858 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1859 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1860 tg->bio_cnt, ret, tg->td->scale);
1864 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1866 struct throtl_service_queue *sq = &tg->service_queue;
1867 bool read_limit, write_limit;
1870 * if cgroup reaches low limit (if low limit is 0, the cgroup always
1871 * reaches), it's ok to upgrade to next limit
1873 read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1874 write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1875 if (!read_limit && !write_limit)
1877 if (read_limit && sq->nr_queued[READ] &&
1878 (!write_limit || sq->nr_queued[WRITE]))
1880 if (write_limit && sq->nr_queued[WRITE] &&
1881 (!read_limit || sq->nr_queued[READ]))
1884 if (time_after_eq(jiffies,
1885 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1886 throtl_tg_is_idle(tg))
1891 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1894 if (throtl_tg_can_upgrade(tg))
1896 tg = sq_to_tg(tg->service_queue.parent_sq);
1897 if (!tg || !tg_to_blkg(tg)->parent)
1903 static bool throtl_can_upgrade(struct throtl_data *td,
1904 struct throtl_grp *this_tg)
1906 struct cgroup_subsys_state *pos_css;
1907 struct blkcg_gq *blkg;
1909 if (td->limit_index != LIMIT_LOW)
1912 if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1916 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1917 struct throtl_grp *tg = blkg_to_tg(blkg);
1921 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1923 if (!throtl_hierarchy_can_upgrade(tg)) {
1932 static void throtl_upgrade_check(struct throtl_grp *tg)
1934 unsigned long now = jiffies;
1936 if (tg->td->limit_index != LIMIT_LOW)
1939 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1942 tg->last_check_time = now;
1944 if (!time_after_eq(now,
1945 __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1948 if (throtl_can_upgrade(tg->td, NULL))
1949 throtl_upgrade_state(tg->td);
1952 static void throtl_upgrade_state(struct throtl_data *td)
1954 struct cgroup_subsys_state *pos_css;
1955 struct blkcg_gq *blkg;
1957 throtl_log(&td->service_queue, "upgrade to max");
1958 td->limit_index = LIMIT_MAX;
1959 td->low_upgrade_time = jiffies;
1962 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1963 struct throtl_grp *tg = blkg_to_tg(blkg);
1964 struct throtl_service_queue *sq = &tg->service_queue;
1966 tg->disptime = jiffies - 1;
1967 throtl_select_dispatch(sq);
1968 throtl_schedule_next_dispatch(sq, true);
1971 throtl_select_dispatch(&td->service_queue);
1972 throtl_schedule_next_dispatch(&td->service_queue, true);
1973 queue_work(kthrotld_workqueue, &td->dispatch_work);
1976 static void throtl_downgrade_state(struct throtl_data *td)
1980 throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1982 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1986 td->limit_index = LIMIT_LOW;
1987 td->low_downgrade_time = jiffies;
1990 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1992 struct throtl_data *td = tg->td;
1993 unsigned long now = jiffies;
1996 * If cgroup is below low limit, consider downgrade and throttle other
1999 if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
2000 time_after_eq(now, tg_last_low_overflow_time(tg) +
2001 td->throtl_slice) &&
2002 (!throtl_tg_is_idle(tg) ||
2003 !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
2008 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
2011 if (!throtl_tg_can_downgrade(tg))
2013 tg = sq_to_tg(tg->service_queue.parent_sq);
2014 if (!tg || !tg_to_blkg(tg)->parent)
2020 static void throtl_downgrade_check(struct throtl_grp *tg)
2024 unsigned long elapsed_time;
2025 unsigned long now = jiffies;
2027 if (tg->td->limit_index != LIMIT_MAX ||
2028 !tg->td->limit_valid[LIMIT_LOW])
2030 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
2032 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
2035 elapsed_time = now - tg->last_check_time;
2036 tg->last_check_time = now;
2038 if (time_before(now, tg_last_low_overflow_time(tg) +
2039 tg->td->throtl_slice))
2042 if (tg->bps[READ][LIMIT_LOW]) {
2043 bps = tg->last_bytes_disp[READ] * HZ;
2044 do_div(bps, elapsed_time);
2045 if (bps >= tg->bps[READ][LIMIT_LOW])
2046 tg->last_low_overflow_time[READ] = now;
2049 if (tg->bps[WRITE][LIMIT_LOW]) {
2050 bps = tg->last_bytes_disp[WRITE] * HZ;
2051 do_div(bps, elapsed_time);
2052 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2053 tg->last_low_overflow_time[WRITE] = now;
2056 if (tg->iops[READ][LIMIT_LOW]) {
2057 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2058 if (iops >= tg->iops[READ][LIMIT_LOW])
2059 tg->last_low_overflow_time[READ] = now;
2062 if (tg->iops[WRITE][LIMIT_LOW]) {
2063 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2064 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2065 tg->last_low_overflow_time[WRITE] = now;
2069 * If cgroup is below low limit, consider downgrade and throttle other
2072 if (throtl_hierarchy_can_downgrade(tg))
2073 throtl_downgrade_state(tg->td);
2075 tg->last_bytes_disp[READ] = 0;
2076 tg->last_bytes_disp[WRITE] = 0;
2077 tg->last_io_disp[READ] = 0;
2078 tg->last_io_disp[WRITE] = 0;
2081 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2084 unsigned long last_finish_time = tg->last_finish_time;
2086 if (last_finish_time == 0)
2089 now = ktime_get_ns() >> 10;
2090 if (now <= last_finish_time ||
2091 last_finish_time == tg->checked_last_finish_time)
2094 tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2095 tg->checked_last_finish_time = last_finish_time;
2098 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2099 static void throtl_update_latency_buckets(struct throtl_data *td)
2101 struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2103 unsigned long last_latency[2] = { 0 };
2104 unsigned long latency[2];
2106 if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
2108 if (time_before(jiffies, td->last_calculate_time + HZ))
2110 td->last_calculate_time = jiffies;
2112 memset(avg_latency, 0, sizeof(avg_latency));
2113 for (rw = READ; rw <= WRITE; rw++) {
2114 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2115 struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2117 for_each_possible_cpu(cpu) {
2118 struct latency_bucket *bucket;
2120 /* this isn't race free, but ok in practice */
2121 bucket = per_cpu_ptr(td->latency_buckets[rw],
2123 tmp->total_latency += bucket[i].total_latency;
2124 tmp->samples += bucket[i].samples;
2125 bucket[i].total_latency = 0;
2126 bucket[i].samples = 0;
2129 if (tmp->samples >= 32) {
2130 int samples = tmp->samples;
2132 latency[rw] = tmp->total_latency;
2134 tmp->total_latency = 0;
2136 latency[rw] /= samples;
2137 if (latency[rw] == 0)
2139 avg_latency[rw][i].latency = latency[rw];
2144 for (rw = READ; rw <= WRITE; rw++) {
2145 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2146 if (!avg_latency[rw][i].latency) {
2147 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2148 td->avg_buckets[rw][i].latency =
2153 if (!td->avg_buckets[rw][i].valid)
2154 latency[rw] = avg_latency[rw][i].latency;
2156 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2157 avg_latency[rw][i].latency) >> 3;
2159 td->avg_buckets[rw][i].latency = max(latency[rw],
2161 td->avg_buckets[rw][i].valid = true;
2162 last_latency[rw] = td->avg_buckets[rw][i].latency;
2166 for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2167 throtl_log(&td->service_queue,
2168 "Latency bucket %d: read latency=%ld, read valid=%d, "
2169 "write latency=%ld, write valid=%d", i,
2170 td->avg_buckets[READ][i].latency,
2171 td->avg_buckets[READ][i].valid,
2172 td->avg_buckets[WRITE][i].latency,
2173 td->avg_buckets[WRITE][i].valid);
2176 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2181 bool blk_throtl_bio(struct bio *bio)
2183 struct request_queue *q = bio->bi_disk->queue;
2184 struct blkcg_gq *blkg = bio->bi_blkg;
2185 struct throtl_qnode *qn = NULL;
2186 struct throtl_grp *tg = blkg_to_tg(blkg);
2187 struct throtl_service_queue *sq;
2188 bool rw = bio_data_dir(bio);
2189 bool throttled = false;
2190 struct throtl_data *td = tg->td;
2194 /* see throtl_charge_bio() */
2195 if (bio_flagged(bio, BIO_THROTTLED))
2198 if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
2199 blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
2200 bio->bi_iter.bi_size);
2201 blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
2204 if (!tg->has_rules[rw])
2207 spin_lock_irq(&q->queue_lock);
2209 throtl_update_latency_buckets(td);
2211 blk_throtl_update_idletime(tg);
2213 sq = &tg->service_queue;
2217 if (tg->last_low_overflow_time[rw] == 0)
2218 tg->last_low_overflow_time[rw] = jiffies;
2219 throtl_downgrade_check(tg);
2220 throtl_upgrade_check(tg);
2221 /* throtl is FIFO - if bios are already queued, should queue */
2222 if (sq->nr_queued[rw])
2225 /* if above limits, break to queue */
2226 if (!tg_may_dispatch(tg, bio, NULL)) {
2227 tg->last_low_overflow_time[rw] = jiffies;
2228 if (throtl_can_upgrade(td, tg)) {
2229 throtl_upgrade_state(td);
2235 /* within limits, let's charge and dispatch directly */
2236 throtl_charge_bio(tg, bio);
2239 * We need to trim slice even when bios are not being queued
2240 * otherwise it might happen that a bio is not queued for
2241 * a long time and slice keeps on extending and trim is not
2242 * called for a long time. Now if limits are reduced suddenly
2243 * we take into account all the IO dispatched so far at new
2244 * low rate and * newly queued IO gets a really long dispatch
2247 * So keep on trimming slice even if bio is not queued.
2249 throtl_trim_slice(tg, rw);
2252 * @bio passed through this layer without being throttled.
2253 * Climb up the ladder. If we're already at the top, it
2254 * can be executed directly.
2256 qn = &tg->qnode_on_parent[rw];
2263 /* out-of-limit, queue to @tg */
2264 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2265 rw == READ ? 'R' : 'W',
2266 tg->bytes_disp[rw], bio->bi_iter.bi_size,
2267 tg_bps_limit(tg, rw),
2268 tg->io_disp[rw], tg_iops_limit(tg, rw),
2269 sq->nr_queued[READ], sq->nr_queued[WRITE]);
2271 tg->last_low_overflow_time[rw] = jiffies;
2273 td->nr_queued[rw]++;
2274 throtl_add_bio_tg(bio, qn, tg);
2278 * Update @tg's dispatch time and force schedule dispatch if @tg
2279 * was empty before @bio. The forced scheduling isn't likely to
2280 * cause undue delay as @bio is likely to be dispatched directly if
2281 * its @tg's disptime is not in the future.
2283 if (tg->flags & THROTL_TG_WAS_EMPTY) {
2284 tg_update_disptime(tg);
2285 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2289 spin_unlock_irq(&q->queue_lock);
2291 bio_set_flag(bio, BIO_THROTTLED);
2293 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2294 if (throttled || !td->track_bio_latency)
2295 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2301 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2302 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2303 int op, unsigned long time)
2305 struct latency_bucket *latency;
2308 if (!td || td->limit_index != LIMIT_LOW ||
2309 !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2310 !blk_queue_nonrot(td->queue))
2313 index = request_bucket_index(size);
2315 latency = get_cpu_ptr(td->latency_buckets[op]);
2316 latency[index].total_latency += time;
2317 latency[index].samples++;
2318 put_cpu_ptr(td->latency_buckets[op]);
2321 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2323 struct request_queue *q = rq->q;
2324 struct throtl_data *td = q->td;
2326 throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2330 void blk_throtl_bio_endio(struct bio *bio)
2332 struct blkcg_gq *blkg;
2333 struct throtl_grp *tg;
2335 unsigned long finish_time;
2336 unsigned long start_time;
2338 int rw = bio_data_dir(bio);
2340 blkg = bio->bi_blkg;
2343 tg = blkg_to_tg(blkg);
2344 if (!tg->td->limit_valid[LIMIT_LOW])
2347 finish_time_ns = ktime_get_ns();
2348 tg->last_finish_time = finish_time_ns >> 10;
2350 start_time = bio_issue_time(&bio->bi_issue) >> 10;
2351 finish_time = __bio_issue_time(finish_time_ns) >> 10;
2352 if (!start_time || finish_time <= start_time)
2355 lat = finish_time - start_time;
2356 /* this is only for bio based driver */
2357 if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2358 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2361 if (tg->latency_target && lat >= tg->td->filtered_latency) {
2363 unsigned int threshold;
2365 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2366 threshold = tg->td->avg_buckets[rw][bucket].latency +
2368 if (lat > threshold)
2371 * Not race free, could get wrong count, which means cgroups
2377 if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2378 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2380 tg->bad_bio_cnt /= 2;
2385 int blk_throtl_init(struct request_queue *q)
2387 struct throtl_data *td;
2390 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2393 td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2394 LATENCY_BUCKET_SIZE, __alignof__(u64));
2395 if (!td->latency_buckets[READ]) {
2399 td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2400 LATENCY_BUCKET_SIZE, __alignof__(u64));
2401 if (!td->latency_buckets[WRITE]) {
2402 free_percpu(td->latency_buckets[READ]);
2407 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2408 throtl_service_queue_init(&td->service_queue);
2413 td->limit_valid[LIMIT_MAX] = true;
2414 td->limit_index = LIMIT_MAX;
2415 td->low_upgrade_time = jiffies;
2416 td->low_downgrade_time = jiffies;
2418 /* activate policy */
2419 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2421 free_percpu(td->latency_buckets[READ]);
2422 free_percpu(td->latency_buckets[WRITE]);
2428 void blk_throtl_exit(struct request_queue *q)
2431 throtl_shutdown_wq(q);
2432 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2433 free_percpu(q->td->latency_buckets[READ]);
2434 free_percpu(q->td->latency_buckets[WRITE]);
2438 void blk_throtl_register_queue(struct request_queue *q)
2440 struct throtl_data *td;
2446 if (blk_queue_nonrot(q)) {
2447 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2448 td->filtered_latency = LATENCY_FILTERED_SSD;
2450 td->throtl_slice = DFL_THROTL_SLICE_HD;
2451 td->filtered_latency = LATENCY_FILTERED_HD;
2452 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2453 td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2454 td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2457 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2458 /* if no low limit, use previous default */
2459 td->throtl_slice = DFL_THROTL_SLICE_HD;
2462 td->track_bio_latency = !queue_is_mq(q);
2463 if (!td->track_bio_latency)
2464 blk_stat_enable_accounting(q);
2467 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2468 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2472 return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2475 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2476 const char *page, size_t count)
2483 if (kstrtoul(page, 10, &v))
2485 t = msecs_to_jiffies(v);
2486 if (t == 0 || t > MAX_THROTL_SLICE)
2488 q->td->throtl_slice = t;
2493 static int __init throtl_init(void)
2495 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2496 if (!kthrotld_workqueue)
2497 panic("Failed to create kthrotld\n");
2499 return blkcg_policy_register(&blkcg_policy_throtl);
2502 module_init(throtl_init);