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)
641 /* Service tree is empty */
642 if (!parent_sq->nr_pending)
645 n = rb_first_cached(&parent_sq->pending_tree);
649 return rb_entry_tg(n);
652 static void throtl_rb_erase(struct rb_node *n,
653 struct throtl_service_queue *parent_sq)
655 rb_erase_cached(n, &parent_sq->pending_tree);
657 --parent_sq->nr_pending;
660 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
662 struct throtl_grp *tg;
664 tg = throtl_rb_first(parent_sq);
668 parent_sq->first_pending_disptime = tg->disptime;
671 static void tg_service_queue_add(struct throtl_grp *tg)
673 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
674 struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
675 struct rb_node *parent = NULL;
676 struct throtl_grp *__tg;
677 unsigned long key = tg->disptime;
678 bool leftmost = true;
680 while (*node != NULL) {
682 __tg = rb_entry_tg(parent);
684 if (time_before(key, __tg->disptime))
685 node = &parent->rb_left;
687 node = &parent->rb_right;
692 rb_link_node(&tg->rb_node, parent, node);
693 rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
697 static void __throtl_enqueue_tg(struct throtl_grp *tg)
699 tg_service_queue_add(tg);
700 tg->flags |= THROTL_TG_PENDING;
701 tg->service_queue.parent_sq->nr_pending++;
704 static void throtl_enqueue_tg(struct throtl_grp *tg)
706 if (!(tg->flags & THROTL_TG_PENDING))
707 __throtl_enqueue_tg(tg);
710 static void __throtl_dequeue_tg(struct throtl_grp *tg)
712 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
713 tg->flags &= ~THROTL_TG_PENDING;
716 static void throtl_dequeue_tg(struct throtl_grp *tg)
718 if (tg->flags & THROTL_TG_PENDING)
719 __throtl_dequeue_tg(tg);
722 /* Call with queue lock held */
723 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
724 unsigned long expires)
726 unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
729 * Since we are adjusting the throttle limit dynamically, the sleep
730 * time calculated according to previous limit might be invalid. It's
731 * possible the cgroup sleep time is very long and no other cgroups
732 * have IO running so notify the limit changes. Make sure the cgroup
733 * doesn't sleep too long to avoid the missed notification.
735 if (time_after(expires, max_expire))
736 expires = max_expire;
737 mod_timer(&sq->pending_timer, expires);
738 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
739 expires - jiffies, jiffies);
743 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
744 * @sq: the service_queue to schedule dispatch for
745 * @force: force scheduling
747 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
748 * dispatch time of the first pending child. Returns %true if either timer
749 * is armed or there's no pending child left. %false if the current
750 * dispatch window is still open and the caller should continue
753 * If @force is %true, the dispatch timer is always scheduled and this
754 * function is guaranteed to return %true. This is to be used when the
755 * caller can't dispatch itself and needs to invoke pending_timer
756 * unconditionally. Note that forced scheduling is likely to induce short
757 * delay before dispatch starts even if @sq->first_pending_disptime is not
758 * in the future and thus shouldn't be used in hot paths.
760 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
763 /* any pending children left? */
767 update_min_dispatch_time(sq);
769 /* is the next dispatch time in the future? */
770 if (force || time_after(sq->first_pending_disptime, jiffies)) {
771 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
775 /* tell the caller to continue dispatching */
779 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
780 bool rw, unsigned long start)
782 tg->bytes_disp[rw] = 0;
786 * Previous slice has expired. We must have trimmed it after last
787 * bio dispatch. That means since start of last slice, we never used
788 * that bandwidth. Do try to make use of that bandwidth while giving
791 if (time_after_eq(start, tg->slice_start[rw]))
792 tg->slice_start[rw] = start;
794 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
795 throtl_log(&tg->service_queue,
796 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
797 rw == READ ? 'R' : 'W', tg->slice_start[rw],
798 tg->slice_end[rw], jiffies);
801 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
803 tg->bytes_disp[rw] = 0;
805 tg->slice_start[rw] = jiffies;
806 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
807 throtl_log(&tg->service_queue,
808 "[%c] new slice start=%lu end=%lu jiffies=%lu",
809 rw == READ ? 'R' : 'W', tg->slice_start[rw],
810 tg->slice_end[rw], jiffies);
813 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
814 unsigned long jiffy_end)
816 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
819 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
820 unsigned long jiffy_end)
822 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
823 throtl_log(&tg->service_queue,
824 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
825 rw == READ ? 'R' : 'W', tg->slice_start[rw],
826 tg->slice_end[rw], jiffies);
829 /* Determine if previously allocated or extended slice is complete or not */
830 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
832 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
838 /* Trim the used slices and adjust slice start accordingly */
839 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
841 unsigned long nr_slices, time_elapsed, io_trim;
844 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
847 * If bps are unlimited (-1), then time slice don't get
848 * renewed. Don't try to trim the slice if slice is used. A new
849 * slice will start when appropriate.
851 if (throtl_slice_used(tg, rw))
855 * A bio has been dispatched. Also adjust slice_end. It might happen
856 * that initially cgroup limit was very low resulting in high
857 * slice_end, but later limit was bumped up and bio was dispatched
858 * sooner, then we need to reduce slice_end. A high bogus slice_end
859 * is bad because it does not allow new slice to start.
862 throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
864 time_elapsed = jiffies - tg->slice_start[rw];
866 nr_slices = time_elapsed / tg->td->throtl_slice;
870 tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
874 io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
877 if (!bytes_trim && !io_trim)
880 if (tg->bytes_disp[rw] >= bytes_trim)
881 tg->bytes_disp[rw] -= bytes_trim;
883 tg->bytes_disp[rw] = 0;
885 if (tg->io_disp[rw] >= io_trim)
886 tg->io_disp[rw] -= io_trim;
890 tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
892 throtl_log(&tg->service_queue,
893 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
894 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
895 tg->slice_start[rw], tg->slice_end[rw], jiffies);
898 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
899 u32 iops_limit, unsigned long *wait)
901 bool rw = bio_data_dir(bio);
902 unsigned int io_allowed;
903 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
906 if (iops_limit == UINT_MAX) {
912 jiffy_elapsed = jiffies - tg->slice_start[rw];
914 /* Round up to the next throttle slice, wait time must be nonzero */
915 jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
918 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
919 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
920 * will allow dispatch after 1 second and after that slice should
924 tmp = (u64)iops_limit * jiffy_elapsed_rnd;
928 io_allowed = UINT_MAX;
932 if (tg->io_disp[rw] + 1 <= io_allowed) {
938 /* Calc approx time to dispatch */
939 jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
946 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
947 u64 bps_limit, unsigned long *wait)
949 bool rw = bio_data_dir(bio);
950 u64 bytes_allowed, extra_bytes, tmp;
951 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
952 unsigned int bio_size = throtl_bio_data_size(bio);
954 if (bps_limit == U64_MAX) {
960 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
962 /* Slice has just started. Consider one slice interval */
964 jiffy_elapsed_rnd = tg->td->throtl_slice;
966 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
968 tmp = bps_limit * jiffy_elapsed_rnd;
972 if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
978 /* Calc approx time to dispatch */
979 extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
980 jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
986 * This wait time is without taking into consideration the rounding
987 * up we did. Add that time also.
989 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
996 * Returns whether one can dispatch a bio or not. Also returns approx number
997 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
999 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
1000 unsigned long *wait)
1002 bool rw = bio_data_dir(bio);
1003 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
1004 u64 bps_limit = tg_bps_limit(tg, rw);
1005 u32 iops_limit = tg_iops_limit(tg, rw);
1008 * Currently whole state machine of group depends on first bio
1009 * queued in the group bio list. So one should not be calling
1010 * this function with a different bio if there are other bios
1013 BUG_ON(tg->service_queue.nr_queued[rw] &&
1014 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
1016 /* If tg->bps = -1, then BW is unlimited */
1017 if (bps_limit == U64_MAX && iops_limit == UINT_MAX) {
1024 * If previous slice expired, start a new one otherwise renew/extend
1025 * existing slice to make sure it is at least throtl_slice interval
1026 * long since now. New slice is started only for empty throttle group.
1027 * If there is queued bio, that means there should be an active
1028 * slice and it should be extended instead.
1030 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
1031 throtl_start_new_slice(tg, rw);
1033 if (time_before(tg->slice_end[rw],
1034 jiffies + tg->td->throtl_slice))
1035 throtl_extend_slice(tg, rw,
1036 jiffies + tg->td->throtl_slice);
1039 if (tg_with_in_bps_limit(tg, bio, bps_limit, &bps_wait) &&
1040 tg_with_in_iops_limit(tg, bio, iops_limit, &iops_wait)) {
1046 max_wait = max(bps_wait, iops_wait);
1051 if (time_before(tg->slice_end[rw], jiffies + max_wait))
1052 throtl_extend_slice(tg, rw, jiffies + max_wait);
1057 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1059 bool rw = bio_data_dir(bio);
1060 unsigned int bio_size = throtl_bio_data_size(bio);
1062 /* Charge the bio to the group */
1063 tg->bytes_disp[rw] += bio_size;
1065 tg->last_bytes_disp[rw] += bio_size;
1066 tg->last_io_disp[rw]++;
1069 * BIO_THROTTLED is used to prevent the same bio to be throttled
1070 * more than once as a throttled bio will go through blk-throtl the
1071 * second time when it eventually gets issued. Set it when a bio
1072 * is being charged to a tg.
1074 if (!bio_flagged(bio, BIO_THROTTLED))
1075 bio_set_flag(bio, BIO_THROTTLED);
1079 * throtl_add_bio_tg - add a bio to the specified throtl_grp
1082 * @tg: the target throtl_grp
1084 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
1085 * tg->qnode_on_self[] is used.
1087 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1088 struct throtl_grp *tg)
1090 struct throtl_service_queue *sq = &tg->service_queue;
1091 bool rw = bio_data_dir(bio);
1094 qn = &tg->qnode_on_self[rw];
1097 * If @tg doesn't currently have any bios queued in the same
1098 * direction, queueing @bio can change when @tg should be
1099 * dispatched. Mark that @tg was empty. This is automatically
1100 * cleared on the next tg_update_disptime().
1102 if (!sq->nr_queued[rw])
1103 tg->flags |= THROTL_TG_WAS_EMPTY;
1105 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1107 sq->nr_queued[rw]++;
1108 throtl_enqueue_tg(tg);
1111 static void tg_update_disptime(struct throtl_grp *tg)
1113 struct throtl_service_queue *sq = &tg->service_queue;
1114 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1117 bio = throtl_peek_queued(&sq->queued[READ]);
1119 tg_may_dispatch(tg, bio, &read_wait);
1121 bio = throtl_peek_queued(&sq->queued[WRITE]);
1123 tg_may_dispatch(tg, bio, &write_wait);
1125 min_wait = min(read_wait, write_wait);
1126 disptime = jiffies + min_wait;
1128 /* Update dispatch time */
1129 throtl_dequeue_tg(tg);
1130 tg->disptime = disptime;
1131 throtl_enqueue_tg(tg);
1133 /* see throtl_add_bio_tg() */
1134 tg->flags &= ~THROTL_TG_WAS_EMPTY;
1137 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1138 struct throtl_grp *parent_tg, bool rw)
1140 if (throtl_slice_used(parent_tg, rw)) {
1141 throtl_start_new_slice_with_credit(parent_tg, rw,
1142 child_tg->slice_start[rw]);
1147 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1149 struct throtl_service_queue *sq = &tg->service_queue;
1150 struct throtl_service_queue *parent_sq = sq->parent_sq;
1151 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1152 struct throtl_grp *tg_to_put = NULL;
1156 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1157 * from @tg may put its reference and @parent_sq might end up
1158 * getting released prematurely. Remember the tg to put and put it
1159 * after @bio is transferred to @parent_sq.
1161 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1162 sq->nr_queued[rw]--;
1164 throtl_charge_bio(tg, bio);
1167 * If our parent is another tg, we just need to transfer @bio to
1168 * the parent using throtl_add_bio_tg(). If our parent is
1169 * @td->service_queue, @bio is ready to be issued. Put it on its
1170 * bio_lists[] and decrease total number queued. The caller is
1171 * responsible for issuing these bios.
1174 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1175 start_parent_slice_with_credit(tg, parent_tg, rw);
1177 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1178 &parent_sq->queued[rw]);
1179 BUG_ON(tg->td->nr_queued[rw] <= 0);
1180 tg->td->nr_queued[rw]--;
1183 throtl_trim_slice(tg, rw);
1186 blkg_put(tg_to_blkg(tg_to_put));
1189 static int throtl_dispatch_tg(struct throtl_grp *tg)
1191 struct throtl_service_queue *sq = &tg->service_queue;
1192 unsigned int nr_reads = 0, nr_writes = 0;
1193 unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1194 unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1197 /* Try to dispatch 75% READS and 25% WRITES */
1199 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1200 tg_may_dispatch(tg, bio, NULL)) {
1202 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1205 if (nr_reads >= max_nr_reads)
1209 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1210 tg_may_dispatch(tg, bio, NULL)) {
1212 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1215 if (nr_writes >= max_nr_writes)
1219 return nr_reads + nr_writes;
1222 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1224 unsigned int nr_disp = 0;
1227 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1228 struct throtl_service_queue *sq;
1233 if (time_before(jiffies, tg->disptime))
1236 throtl_dequeue_tg(tg);
1238 nr_disp += throtl_dispatch_tg(tg);
1240 sq = &tg->service_queue;
1241 if (sq->nr_queued[0] || sq->nr_queued[1])
1242 tg_update_disptime(tg);
1244 if (nr_disp >= THROTL_QUANTUM)
1251 static bool throtl_can_upgrade(struct throtl_data *td,
1252 struct throtl_grp *this_tg);
1254 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1255 * @t: the pending_timer member of the throtl_service_queue being serviced
1257 * This timer is armed when a child throtl_grp with active bio's become
1258 * pending and queued on the service_queue's pending_tree and expires when
1259 * the first child throtl_grp should be dispatched. This function
1260 * dispatches bio's from the children throtl_grps to the parent
1263 * If the parent's parent is another throtl_grp, dispatching is propagated
1264 * by either arming its pending_timer or repeating dispatch directly. If
1265 * the top-level service_tree is reached, throtl_data->dispatch_work is
1266 * kicked so that the ready bio's are issued.
1268 static void throtl_pending_timer_fn(struct timer_list *t)
1270 struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1271 struct throtl_grp *tg = sq_to_tg(sq);
1272 struct throtl_data *td = sq_to_td(sq);
1273 struct request_queue *q = td->queue;
1274 struct throtl_service_queue *parent_sq;
1278 spin_lock_irq(&q->queue_lock);
1279 if (throtl_can_upgrade(td, NULL))
1280 throtl_upgrade_state(td);
1283 parent_sq = sq->parent_sq;
1287 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1288 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1289 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1291 ret = throtl_select_dispatch(sq);
1293 throtl_log(sq, "bios disp=%u", ret);
1297 if (throtl_schedule_next_dispatch(sq, false))
1300 /* this dispatch windows is still open, relax and repeat */
1301 spin_unlock_irq(&q->queue_lock);
1303 spin_lock_irq(&q->queue_lock);
1310 /* @parent_sq is another throl_grp, propagate dispatch */
1311 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1312 tg_update_disptime(tg);
1313 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1314 /* window is already open, repeat dispatching */
1321 /* reached the top-level, queue issuing */
1322 queue_work(kthrotld_workqueue, &td->dispatch_work);
1325 spin_unlock_irq(&q->queue_lock);
1329 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1330 * @work: work item being executed
1332 * This function is queued for execution when bios reach the bio_lists[]
1333 * of throtl_data->service_queue. Those bios are ready and issued by this
1336 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1338 struct throtl_data *td = container_of(work, struct throtl_data,
1340 struct throtl_service_queue *td_sq = &td->service_queue;
1341 struct request_queue *q = td->queue;
1342 struct bio_list bio_list_on_stack;
1344 struct blk_plug plug;
1347 bio_list_init(&bio_list_on_stack);
1349 spin_lock_irq(&q->queue_lock);
1350 for (rw = READ; rw <= WRITE; rw++)
1351 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1352 bio_list_add(&bio_list_on_stack, bio);
1353 spin_unlock_irq(&q->queue_lock);
1355 if (!bio_list_empty(&bio_list_on_stack)) {
1356 blk_start_plug(&plug);
1357 while ((bio = bio_list_pop(&bio_list_on_stack)))
1358 submit_bio_noacct(bio);
1359 blk_finish_plug(&plug);
1363 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1366 struct throtl_grp *tg = pd_to_tg(pd);
1367 u64 v = *(u64 *)((void *)tg + off);
1371 return __blkg_prfill_u64(sf, pd, v);
1374 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1377 struct throtl_grp *tg = pd_to_tg(pd);
1378 unsigned int v = *(unsigned int *)((void *)tg + off);
1382 return __blkg_prfill_u64(sf, pd, v);
1385 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1387 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1388 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1392 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1394 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1395 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1399 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1401 struct throtl_service_queue *sq = &tg->service_queue;
1402 struct cgroup_subsys_state *pos_css;
1403 struct blkcg_gq *blkg;
1405 throtl_log(&tg->service_queue,
1406 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1407 tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1408 tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1411 * Update has_rules[] flags for the updated tg's subtree. A tg is
1412 * considered to have rules if either the tg itself or any of its
1413 * ancestors has rules. This identifies groups without any
1414 * restrictions in the whole hierarchy and allows them to bypass
1417 blkg_for_each_descendant_pre(blkg, pos_css,
1418 global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1419 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1420 struct throtl_grp *parent_tg;
1422 tg_update_has_rules(this_tg);
1423 /* ignore root/second level */
1424 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1425 !blkg->parent->parent)
1427 parent_tg = blkg_to_tg(blkg->parent);
1429 * make sure all children has lower idle time threshold and
1430 * higher latency target
1432 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1433 parent_tg->idletime_threshold);
1434 this_tg->latency_target = max(this_tg->latency_target,
1435 parent_tg->latency_target);
1439 * We're already holding queue_lock and know @tg is valid. Let's
1440 * apply the new config directly.
1442 * Restart the slices for both READ and WRITES. It might happen
1443 * that a group's limit are dropped suddenly and we don't want to
1444 * account recently dispatched IO with new low rate.
1446 throtl_start_new_slice(tg, READ);
1447 throtl_start_new_slice(tg, WRITE);
1449 if (tg->flags & THROTL_TG_PENDING) {
1450 tg_update_disptime(tg);
1451 throtl_schedule_next_dispatch(sq->parent_sq, true);
1455 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1456 char *buf, size_t nbytes, loff_t off, bool is_u64)
1458 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1459 struct blkg_conf_ctx ctx;
1460 struct throtl_grp *tg;
1464 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1469 if (sscanf(ctx.body, "%llu", &v) != 1)
1474 tg = blkg_to_tg(ctx.blkg);
1477 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1479 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1481 tg_conf_updated(tg, false);
1484 blkg_conf_finish(&ctx);
1485 return ret ?: nbytes;
1488 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1489 char *buf, size_t nbytes, loff_t off)
1491 return tg_set_conf(of, buf, nbytes, off, true);
1494 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1495 char *buf, size_t nbytes, loff_t off)
1497 return tg_set_conf(of, buf, nbytes, off, false);
1500 static int tg_print_rwstat(struct seq_file *sf, void *v)
1502 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1503 blkg_prfill_rwstat, &blkcg_policy_throtl,
1504 seq_cft(sf)->private, true);
1508 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1509 struct blkg_policy_data *pd, int off)
1511 struct blkg_rwstat_sample sum;
1513 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1515 return __blkg_prfill_rwstat(sf, pd, &sum);
1518 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1520 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1521 tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1522 seq_cft(sf)->private, true);
1526 static struct cftype throtl_legacy_files[] = {
1528 .name = "throttle.read_bps_device",
1529 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1530 .seq_show = tg_print_conf_u64,
1531 .write = tg_set_conf_u64,
1534 .name = "throttle.write_bps_device",
1535 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1536 .seq_show = tg_print_conf_u64,
1537 .write = tg_set_conf_u64,
1540 .name = "throttle.read_iops_device",
1541 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1542 .seq_show = tg_print_conf_uint,
1543 .write = tg_set_conf_uint,
1546 .name = "throttle.write_iops_device",
1547 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1548 .seq_show = tg_print_conf_uint,
1549 .write = tg_set_conf_uint,
1552 .name = "throttle.io_service_bytes",
1553 .private = offsetof(struct throtl_grp, stat_bytes),
1554 .seq_show = tg_print_rwstat,
1557 .name = "throttle.io_service_bytes_recursive",
1558 .private = offsetof(struct throtl_grp, stat_bytes),
1559 .seq_show = tg_print_rwstat_recursive,
1562 .name = "throttle.io_serviced",
1563 .private = offsetof(struct throtl_grp, stat_ios),
1564 .seq_show = tg_print_rwstat,
1567 .name = "throttle.io_serviced_recursive",
1568 .private = offsetof(struct throtl_grp, stat_ios),
1569 .seq_show = tg_print_rwstat_recursive,
1574 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1577 struct throtl_grp *tg = pd_to_tg(pd);
1578 const char *dname = blkg_dev_name(pd->blkg);
1579 char bufs[4][21] = { "max", "max", "max", "max" };
1581 unsigned int iops_dft;
1582 char idle_time[26] = "";
1583 char latency_time[26] = "";
1588 if (off == LIMIT_LOW) {
1593 iops_dft = UINT_MAX;
1596 if (tg->bps_conf[READ][off] == bps_dft &&
1597 tg->bps_conf[WRITE][off] == bps_dft &&
1598 tg->iops_conf[READ][off] == iops_dft &&
1599 tg->iops_conf[WRITE][off] == iops_dft &&
1600 (off != LIMIT_LOW ||
1601 (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1602 tg->latency_target_conf == DFL_LATENCY_TARGET)))
1605 if (tg->bps_conf[READ][off] != U64_MAX)
1606 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1607 tg->bps_conf[READ][off]);
1608 if (tg->bps_conf[WRITE][off] != U64_MAX)
1609 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1610 tg->bps_conf[WRITE][off]);
1611 if (tg->iops_conf[READ][off] != UINT_MAX)
1612 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1613 tg->iops_conf[READ][off]);
1614 if (tg->iops_conf[WRITE][off] != UINT_MAX)
1615 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1616 tg->iops_conf[WRITE][off]);
1617 if (off == LIMIT_LOW) {
1618 if (tg->idletime_threshold_conf == ULONG_MAX)
1619 strcpy(idle_time, " idle=max");
1621 snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1622 tg->idletime_threshold_conf);
1624 if (tg->latency_target_conf == ULONG_MAX)
1625 strcpy(latency_time, " latency=max");
1627 snprintf(latency_time, sizeof(latency_time),
1628 " latency=%lu", tg->latency_target_conf);
1631 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1632 dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1637 static int tg_print_limit(struct seq_file *sf, void *v)
1639 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1640 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1644 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1645 char *buf, size_t nbytes, loff_t off)
1647 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1648 struct blkg_conf_ctx ctx;
1649 struct throtl_grp *tg;
1651 unsigned long idle_time;
1652 unsigned long latency_time;
1654 int index = of_cft(of)->private;
1656 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1660 tg = blkg_to_tg(ctx.blkg);
1662 v[0] = tg->bps_conf[READ][index];
1663 v[1] = tg->bps_conf[WRITE][index];
1664 v[2] = tg->iops_conf[READ][index];
1665 v[3] = tg->iops_conf[WRITE][index];
1667 idle_time = tg->idletime_threshold_conf;
1668 latency_time = tg->latency_target_conf;
1670 char tok[27]; /* wiops=18446744073709551616 */
1675 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1684 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1692 if (!strcmp(tok, "rbps") && val > 1)
1694 else if (!strcmp(tok, "wbps") && val > 1)
1696 else if (!strcmp(tok, "riops") && val > 1)
1697 v[2] = min_t(u64, val, UINT_MAX);
1698 else if (!strcmp(tok, "wiops") && val > 1)
1699 v[3] = min_t(u64, val, UINT_MAX);
1700 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1702 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1708 tg->bps_conf[READ][index] = v[0];
1709 tg->bps_conf[WRITE][index] = v[1];
1710 tg->iops_conf[READ][index] = v[2];
1711 tg->iops_conf[WRITE][index] = v[3];
1713 if (index == LIMIT_MAX) {
1714 tg->bps[READ][index] = v[0];
1715 tg->bps[WRITE][index] = v[1];
1716 tg->iops[READ][index] = v[2];
1717 tg->iops[WRITE][index] = v[3];
1719 tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1720 tg->bps_conf[READ][LIMIT_MAX]);
1721 tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1722 tg->bps_conf[WRITE][LIMIT_MAX]);
1723 tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1724 tg->iops_conf[READ][LIMIT_MAX]);
1725 tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1726 tg->iops_conf[WRITE][LIMIT_MAX]);
1727 tg->idletime_threshold_conf = idle_time;
1728 tg->latency_target_conf = latency_time;
1730 /* force user to configure all settings for low limit */
1731 if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1732 tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1733 tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1734 tg->latency_target_conf == DFL_LATENCY_TARGET) {
1735 tg->bps[READ][LIMIT_LOW] = 0;
1736 tg->bps[WRITE][LIMIT_LOW] = 0;
1737 tg->iops[READ][LIMIT_LOW] = 0;
1738 tg->iops[WRITE][LIMIT_LOW] = 0;
1739 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1740 tg->latency_target = DFL_LATENCY_TARGET;
1741 } else if (index == LIMIT_LOW) {
1742 tg->idletime_threshold = tg->idletime_threshold_conf;
1743 tg->latency_target = tg->latency_target_conf;
1746 blk_throtl_update_limit_valid(tg->td);
1747 if (tg->td->limit_valid[LIMIT_LOW]) {
1748 if (index == LIMIT_LOW)
1749 tg->td->limit_index = LIMIT_LOW;
1751 tg->td->limit_index = LIMIT_MAX;
1752 tg_conf_updated(tg, index == LIMIT_LOW &&
1753 tg->td->limit_valid[LIMIT_LOW]);
1756 blkg_conf_finish(&ctx);
1757 return ret ?: nbytes;
1760 static struct cftype throtl_files[] = {
1761 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1764 .flags = CFTYPE_NOT_ON_ROOT,
1765 .seq_show = tg_print_limit,
1766 .write = tg_set_limit,
1767 .private = LIMIT_LOW,
1772 .flags = CFTYPE_NOT_ON_ROOT,
1773 .seq_show = tg_print_limit,
1774 .write = tg_set_limit,
1775 .private = LIMIT_MAX,
1780 static void throtl_shutdown_wq(struct request_queue *q)
1782 struct throtl_data *td = q->td;
1784 cancel_work_sync(&td->dispatch_work);
1787 static struct blkcg_policy blkcg_policy_throtl = {
1788 .dfl_cftypes = throtl_files,
1789 .legacy_cftypes = throtl_legacy_files,
1791 .pd_alloc_fn = throtl_pd_alloc,
1792 .pd_init_fn = throtl_pd_init,
1793 .pd_online_fn = throtl_pd_online,
1794 .pd_offline_fn = throtl_pd_offline,
1795 .pd_free_fn = throtl_pd_free,
1798 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1800 unsigned long rtime = jiffies, wtime = jiffies;
1802 if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1803 rtime = tg->last_low_overflow_time[READ];
1804 if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1805 wtime = tg->last_low_overflow_time[WRITE];
1806 return min(rtime, wtime);
1809 /* tg should not be an intermediate node */
1810 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1812 struct throtl_service_queue *parent_sq;
1813 struct throtl_grp *parent = tg;
1814 unsigned long ret = __tg_last_low_overflow_time(tg);
1817 parent_sq = parent->service_queue.parent_sq;
1818 parent = sq_to_tg(parent_sq);
1823 * The parent doesn't have low limit, it always reaches low
1824 * limit. Its overflow time is useless for children
1826 if (!parent->bps[READ][LIMIT_LOW] &&
1827 !parent->iops[READ][LIMIT_LOW] &&
1828 !parent->bps[WRITE][LIMIT_LOW] &&
1829 !parent->iops[WRITE][LIMIT_LOW])
1831 if (time_after(__tg_last_low_overflow_time(parent), ret))
1832 ret = __tg_last_low_overflow_time(parent);
1837 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1840 * cgroup is idle if:
1841 * - single idle is too long, longer than a fixed value (in case user
1842 * configure a too big threshold) or 4 times of idletime threshold
1843 * - average think time is more than threshold
1844 * - IO latency is largely below threshold
1849 time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1850 ret = tg->latency_target == DFL_LATENCY_TARGET ||
1851 tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1852 (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1853 tg->avg_idletime > tg->idletime_threshold ||
1854 (tg->latency_target && tg->bio_cnt &&
1855 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1856 throtl_log(&tg->service_queue,
1857 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1858 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1859 tg->bio_cnt, ret, tg->td->scale);
1863 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1865 struct throtl_service_queue *sq = &tg->service_queue;
1866 bool read_limit, write_limit;
1869 * if cgroup reaches low limit (if low limit is 0, the cgroup always
1870 * reaches), it's ok to upgrade to next limit
1872 read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1873 write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1874 if (!read_limit && !write_limit)
1876 if (read_limit && sq->nr_queued[READ] &&
1877 (!write_limit || sq->nr_queued[WRITE]))
1879 if (write_limit && sq->nr_queued[WRITE] &&
1880 (!read_limit || sq->nr_queued[READ]))
1883 if (time_after_eq(jiffies,
1884 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1885 throtl_tg_is_idle(tg))
1890 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1893 if (throtl_tg_can_upgrade(tg))
1895 tg = sq_to_tg(tg->service_queue.parent_sq);
1896 if (!tg || !tg_to_blkg(tg)->parent)
1902 static bool throtl_can_upgrade(struct throtl_data *td,
1903 struct throtl_grp *this_tg)
1905 struct cgroup_subsys_state *pos_css;
1906 struct blkcg_gq *blkg;
1908 if (td->limit_index != LIMIT_LOW)
1911 if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1915 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1916 struct throtl_grp *tg = blkg_to_tg(blkg);
1920 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1922 if (!throtl_hierarchy_can_upgrade(tg)) {
1931 static void throtl_upgrade_check(struct throtl_grp *tg)
1933 unsigned long now = jiffies;
1935 if (tg->td->limit_index != LIMIT_LOW)
1938 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1941 tg->last_check_time = now;
1943 if (!time_after_eq(now,
1944 __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1947 if (throtl_can_upgrade(tg->td, NULL))
1948 throtl_upgrade_state(tg->td);
1951 static void throtl_upgrade_state(struct throtl_data *td)
1953 struct cgroup_subsys_state *pos_css;
1954 struct blkcg_gq *blkg;
1956 throtl_log(&td->service_queue, "upgrade to max");
1957 td->limit_index = LIMIT_MAX;
1958 td->low_upgrade_time = jiffies;
1961 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1962 struct throtl_grp *tg = blkg_to_tg(blkg);
1963 struct throtl_service_queue *sq = &tg->service_queue;
1965 tg->disptime = jiffies - 1;
1966 throtl_select_dispatch(sq);
1967 throtl_schedule_next_dispatch(sq, true);
1970 throtl_select_dispatch(&td->service_queue);
1971 throtl_schedule_next_dispatch(&td->service_queue, true);
1972 queue_work(kthrotld_workqueue, &td->dispatch_work);
1975 static void throtl_downgrade_state(struct throtl_data *td)
1979 throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1981 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1985 td->limit_index = LIMIT_LOW;
1986 td->low_downgrade_time = jiffies;
1989 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1991 struct throtl_data *td = tg->td;
1992 unsigned long now = jiffies;
1995 * If cgroup is below low limit, consider downgrade and throttle other
1998 if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1999 time_after_eq(now, tg_last_low_overflow_time(tg) +
2000 td->throtl_slice) &&
2001 (!throtl_tg_is_idle(tg) ||
2002 !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
2007 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
2010 if (!throtl_tg_can_downgrade(tg))
2012 tg = sq_to_tg(tg->service_queue.parent_sq);
2013 if (!tg || !tg_to_blkg(tg)->parent)
2019 static void throtl_downgrade_check(struct throtl_grp *tg)
2023 unsigned long elapsed_time;
2024 unsigned long now = jiffies;
2026 if (tg->td->limit_index != LIMIT_MAX ||
2027 !tg->td->limit_valid[LIMIT_LOW])
2029 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
2031 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
2034 elapsed_time = now - tg->last_check_time;
2035 tg->last_check_time = now;
2037 if (time_before(now, tg_last_low_overflow_time(tg) +
2038 tg->td->throtl_slice))
2041 if (tg->bps[READ][LIMIT_LOW]) {
2042 bps = tg->last_bytes_disp[READ] * HZ;
2043 do_div(bps, elapsed_time);
2044 if (bps >= tg->bps[READ][LIMIT_LOW])
2045 tg->last_low_overflow_time[READ] = now;
2048 if (tg->bps[WRITE][LIMIT_LOW]) {
2049 bps = tg->last_bytes_disp[WRITE] * HZ;
2050 do_div(bps, elapsed_time);
2051 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2052 tg->last_low_overflow_time[WRITE] = now;
2055 if (tg->iops[READ][LIMIT_LOW]) {
2056 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2057 if (iops >= tg->iops[READ][LIMIT_LOW])
2058 tg->last_low_overflow_time[READ] = now;
2061 if (tg->iops[WRITE][LIMIT_LOW]) {
2062 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2063 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2064 tg->last_low_overflow_time[WRITE] = now;
2068 * If cgroup is below low limit, consider downgrade and throttle other
2071 if (throtl_hierarchy_can_downgrade(tg))
2072 throtl_downgrade_state(tg->td);
2074 tg->last_bytes_disp[READ] = 0;
2075 tg->last_bytes_disp[WRITE] = 0;
2076 tg->last_io_disp[READ] = 0;
2077 tg->last_io_disp[WRITE] = 0;
2080 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2083 unsigned long last_finish_time = tg->last_finish_time;
2085 if (last_finish_time == 0)
2088 now = ktime_get_ns() >> 10;
2089 if (now <= last_finish_time ||
2090 last_finish_time == tg->checked_last_finish_time)
2093 tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2094 tg->checked_last_finish_time = last_finish_time;
2097 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2098 static void throtl_update_latency_buckets(struct throtl_data *td)
2100 struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2102 unsigned long last_latency[2] = { 0 };
2103 unsigned long latency[2];
2105 if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
2107 if (time_before(jiffies, td->last_calculate_time + HZ))
2109 td->last_calculate_time = jiffies;
2111 memset(avg_latency, 0, sizeof(avg_latency));
2112 for (rw = READ; rw <= WRITE; rw++) {
2113 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2114 struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2116 for_each_possible_cpu(cpu) {
2117 struct latency_bucket *bucket;
2119 /* this isn't race free, but ok in practice */
2120 bucket = per_cpu_ptr(td->latency_buckets[rw],
2122 tmp->total_latency += bucket[i].total_latency;
2123 tmp->samples += bucket[i].samples;
2124 bucket[i].total_latency = 0;
2125 bucket[i].samples = 0;
2128 if (tmp->samples >= 32) {
2129 int samples = tmp->samples;
2131 latency[rw] = tmp->total_latency;
2133 tmp->total_latency = 0;
2135 latency[rw] /= samples;
2136 if (latency[rw] == 0)
2138 avg_latency[rw][i].latency = latency[rw];
2143 for (rw = READ; rw <= WRITE; rw++) {
2144 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2145 if (!avg_latency[rw][i].latency) {
2146 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2147 td->avg_buckets[rw][i].latency =
2152 if (!td->avg_buckets[rw][i].valid)
2153 latency[rw] = avg_latency[rw][i].latency;
2155 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2156 avg_latency[rw][i].latency) >> 3;
2158 td->avg_buckets[rw][i].latency = max(latency[rw],
2160 td->avg_buckets[rw][i].valid = true;
2161 last_latency[rw] = td->avg_buckets[rw][i].latency;
2165 for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2166 throtl_log(&td->service_queue,
2167 "Latency bucket %d: read latency=%ld, read valid=%d, "
2168 "write latency=%ld, write valid=%d", i,
2169 td->avg_buckets[READ][i].latency,
2170 td->avg_buckets[READ][i].valid,
2171 td->avg_buckets[WRITE][i].latency,
2172 td->avg_buckets[WRITE][i].valid);
2175 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2180 bool blk_throtl_bio(struct bio *bio)
2182 struct request_queue *q = bio->bi_disk->queue;
2183 struct blkcg_gq *blkg = bio->bi_blkg;
2184 struct throtl_qnode *qn = NULL;
2185 struct throtl_grp *tg = blkg_to_tg(blkg);
2186 struct throtl_service_queue *sq;
2187 bool rw = bio_data_dir(bio);
2188 bool throttled = false;
2189 struct throtl_data *td = tg->td;
2193 /* see throtl_charge_bio() */
2194 if (bio_flagged(bio, BIO_THROTTLED))
2197 if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
2198 blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
2199 bio->bi_iter.bi_size);
2200 blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
2203 if (!tg->has_rules[rw])
2206 spin_lock_irq(&q->queue_lock);
2208 throtl_update_latency_buckets(td);
2210 blk_throtl_update_idletime(tg);
2212 sq = &tg->service_queue;
2216 if (tg->last_low_overflow_time[rw] == 0)
2217 tg->last_low_overflow_time[rw] = jiffies;
2218 throtl_downgrade_check(tg);
2219 throtl_upgrade_check(tg);
2220 /* throtl is FIFO - if bios are already queued, should queue */
2221 if (sq->nr_queued[rw])
2224 /* if above limits, break to queue */
2225 if (!tg_may_dispatch(tg, bio, NULL)) {
2226 tg->last_low_overflow_time[rw] = jiffies;
2227 if (throtl_can_upgrade(td, tg)) {
2228 throtl_upgrade_state(td);
2234 /* within limits, let's charge and dispatch directly */
2235 throtl_charge_bio(tg, bio);
2238 * We need to trim slice even when bios are not being queued
2239 * otherwise it might happen that a bio is not queued for
2240 * a long time and slice keeps on extending and trim is not
2241 * called for a long time. Now if limits are reduced suddenly
2242 * we take into account all the IO dispatched so far at new
2243 * low rate and * newly queued IO gets a really long dispatch
2246 * So keep on trimming slice even if bio is not queued.
2248 throtl_trim_slice(tg, rw);
2251 * @bio passed through this layer without being throttled.
2252 * Climb up the ladder. If we're already at the top, it
2253 * can be executed directly.
2255 qn = &tg->qnode_on_parent[rw];
2262 /* out-of-limit, queue to @tg */
2263 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2264 rw == READ ? 'R' : 'W',
2265 tg->bytes_disp[rw], bio->bi_iter.bi_size,
2266 tg_bps_limit(tg, rw),
2267 tg->io_disp[rw], tg_iops_limit(tg, rw),
2268 sq->nr_queued[READ], sq->nr_queued[WRITE]);
2270 tg->last_low_overflow_time[rw] = jiffies;
2272 td->nr_queued[rw]++;
2273 throtl_add_bio_tg(bio, qn, tg);
2277 * Update @tg's dispatch time and force schedule dispatch if @tg
2278 * was empty before @bio. The forced scheduling isn't likely to
2279 * cause undue delay as @bio is likely to be dispatched directly if
2280 * its @tg's disptime is not in the future.
2282 if (tg->flags & THROTL_TG_WAS_EMPTY) {
2283 tg_update_disptime(tg);
2284 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2288 spin_unlock_irq(&q->queue_lock);
2290 bio_set_flag(bio, BIO_THROTTLED);
2292 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2293 if (throttled || !td->track_bio_latency)
2294 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2300 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2301 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2302 int op, unsigned long time)
2304 struct latency_bucket *latency;
2307 if (!td || td->limit_index != LIMIT_LOW ||
2308 !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2309 !blk_queue_nonrot(td->queue))
2312 index = request_bucket_index(size);
2314 latency = get_cpu_ptr(td->latency_buckets[op]);
2315 latency[index].total_latency += time;
2316 latency[index].samples++;
2317 put_cpu_ptr(td->latency_buckets[op]);
2320 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2322 struct request_queue *q = rq->q;
2323 struct throtl_data *td = q->td;
2325 throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2329 void blk_throtl_bio_endio(struct bio *bio)
2331 struct blkcg_gq *blkg;
2332 struct throtl_grp *tg;
2334 unsigned long finish_time;
2335 unsigned long start_time;
2337 int rw = bio_data_dir(bio);
2339 blkg = bio->bi_blkg;
2342 tg = blkg_to_tg(blkg);
2343 if (!tg->td->limit_valid[LIMIT_LOW])
2346 finish_time_ns = ktime_get_ns();
2347 tg->last_finish_time = finish_time_ns >> 10;
2349 start_time = bio_issue_time(&bio->bi_issue) >> 10;
2350 finish_time = __bio_issue_time(finish_time_ns) >> 10;
2351 if (!start_time || finish_time <= start_time)
2354 lat = finish_time - start_time;
2355 /* this is only for bio based driver */
2356 if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2357 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2360 if (tg->latency_target && lat >= tg->td->filtered_latency) {
2362 unsigned int threshold;
2364 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2365 threshold = tg->td->avg_buckets[rw][bucket].latency +
2367 if (lat > threshold)
2370 * Not race free, could get wrong count, which means cgroups
2376 if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2377 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2379 tg->bad_bio_cnt /= 2;
2384 int blk_throtl_init(struct request_queue *q)
2386 struct throtl_data *td;
2389 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2392 td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2393 LATENCY_BUCKET_SIZE, __alignof__(u64));
2394 if (!td->latency_buckets[READ]) {
2398 td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2399 LATENCY_BUCKET_SIZE, __alignof__(u64));
2400 if (!td->latency_buckets[WRITE]) {
2401 free_percpu(td->latency_buckets[READ]);
2406 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2407 throtl_service_queue_init(&td->service_queue);
2412 td->limit_valid[LIMIT_MAX] = true;
2413 td->limit_index = LIMIT_MAX;
2414 td->low_upgrade_time = jiffies;
2415 td->low_downgrade_time = jiffies;
2417 /* activate policy */
2418 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2420 free_percpu(td->latency_buckets[READ]);
2421 free_percpu(td->latency_buckets[WRITE]);
2427 void blk_throtl_exit(struct request_queue *q)
2430 throtl_shutdown_wq(q);
2431 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2432 free_percpu(q->td->latency_buckets[READ]);
2433 free_percpu(q->td->latency_buckets[WRITE]);
2437 void blk_throtl_register_queue(struct request_queue *q)
2439 struct throtl_data *td;
2445 if (blk_queue_nonrot(q)) {
2446 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2447 td->filtered_latency = LATENCY_FILTERED_SSD;
2449 td->throtl_slice = DFL_THROTL_SLICE_HD;
2450 td->filtered_latency = LATENCY_FILTERED_HD;
2451 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2452 td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2453 td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2456 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2457 /* if no low limit, use previous default */
2458 td->throtl_slice = DFL_THROTL_SLICE_HD;
2461 td->track_bio_latency = !queue_is_mq(q);
2462 if (!td->track_bio_latency)
2463 blk_stat_enable_accounting(q);
2466 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2467 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2471 return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2474 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2475 const char *page, size_t count)
2482 if (kstrtoul(page, 10, &v))
2484 t = msecs_to_jiffies(v);
2485 if (t == 0 || t > MAX_THROTL_SLICE)
2487 q->td->throtl_slice = t;
2492 static int __init throtl_init(void)
2494 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2495 if (!kthrotld_workqueue)
2496 panic("Failed to create kthrotld\n");
2498 return blkcg_policy_register(&blkcg_policy_throtl);
2501 module_init(throtl_init);