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>
14 #include "blk-cgroup-rwstat.h"
16 #include "blk-throttle.h"
18 /* Max dispatch from a group in 1 round */
19 #define THROTL_GRP_QUANTUM 8
21 /* Total max dispatch from all groups in one round */
22 #define THROTL_QUANTUM 32
24 /* Throttling is performed over a slice and after that slice is renewed */
25 #define DFL_THROTL_SLICE_HD (HZ / 10)
26 #define DFL_THROTL_SLICE_SSD (HZ / 50)
27 #define MAX_THROTL_SLICE (HZ)
28 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
29 #define MIN_THROTL_BPS (320 * 1024)
30 #define MIN_THROTL_IOPS (10)
31 #define DFL_LATENCY_TARGET (-1L)
32 #define DFL_IDLE_THRESHOLD (0)
33 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
34 #define LATENCY_FILTERED_SSD (0)
36 * For HD, very small latency comes from sequential IO. Such IO is helpless to
37 * help determine if its IO is impacted by others, hence we ignore the IO
39 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
41 /* A workqueue to queue throttle related work */
42 static struct workqueue_struct *kthrotld_workqueue;
44 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
46 /* We measure latency for request size from <= 4k to >= 1M */
47 #define LATENCY_BUCKET_SIZE 9
49 struct latency_bucket {
50 unsigned long total_latency; /* ns / 1024 */
54 struct avg_latency_bucket {
55 unsigned long latency; /* ns / 1024 */
61 /* service tree for active throtl groups */
62 struct throtl_service_queue service_queue;
64 struct request_queue *queue;
66 /* Total Number of queued bios on READ and WRITE lists */
67 unsigned int nr_queued[2];
69 unsigned int throtl_slice;
71 /* Work for dispatching throttled bios */
72 struct work_struct dispatch_work;
73 unsigned int limit_index;
74 bool limit_valid[LIMIT_CNT];
76 unsigned long low_upgrade_time;
77 unsigned long low_downgrade_time;
81 struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
82 struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
83 struct latency_bucket __percpu *latency_buckets[2];
84 unsigned long last_calculate_time;
85 unsigned long filtered_latency;
87 bool track_bio_latency;
90 static void throtl_pending_timer_fn(struct timer_list *t);
92 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
94 return pd_to_blkg(&tg->pd);
98 * sq_to_tg - return the throl_grp the specified service queue belongs to
99 * @sq: the throtl_service_queue of interest
101 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
102 * embedded in throtl_data, %NULL is returned.
104 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
106 if (sq && sq->parent_sq)
107 return container_of(sq, struct throtl_grp, service_queue);
113 * sq_to_td - return throtl_data the specified service queue belongs to
114 * @sq: the throtl_service_queue of interest
116 * A service_queue can be embedded in either a throtl_grp or throtl_data.
117 * Determine the associated throtl_data accordingly and return it.
119 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
121 struct throtl_grp *tg = sq_to_tg(sq);
126 return container_of(sq, struct throtl_data, service_queue);
130 * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
131 * make the IO dispatch more smooth.
132 * Scale up: linearly scale up according to lapsed time since upgrade. For
133 * every throtl_slice, the limit scales up 1/2 .low limit till the
134 * limit hits .max limit
135 * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
137 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
139 /* arbitrary value to avoid too big scale */
140 if (td->scale < 4096 && time_after_eq(jiffies,
141 td->low_upgrade_time + td->scale * td->throtl_slice))
142 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
144 return low + (low >> 1) * td->scale;
147 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
149 struct blkcg_gq *blkg = tg_to_blkg(tg);
150 struct throtl_data *td;
153 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
157 ret = tg->bps[rw][td->limit_index];
158 if (ret == 0 && td->limit_index == LIMIT_LOW) {
159 /* intermediate node or iops isn't 0 */
160 if (!list_empty(&blkg->blkcg->css.children) ||
161 tg->iops[rw][td->limit_index])
164 return MIN_THROTL_BPS;
167 if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
168 tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
171 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
172 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
177 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
179 struct blkcg_gq *blkg = tg_to_blkg(tg);
180 struct throtl_data *td;
183 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
187 ret = tg->iops[rw][td->limit_index];
188 if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
189 /* intermediate node or bps isn't 0 */
190 if (!list_empty(&blkg->blkcg->css.children) ||
191 tg->bps[rw][td->limit_index])
194 return MIN_THROTL_IOPS;
197 if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
198 tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
201 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
202 if (adjusted > UINT_MAX)
204 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
209 #define request_bucket_index(sectors) \
210 clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
213 * throtl_log - log debug message via blktrace
214 * @sq: the service_queue being reported
215 * @fmt: printf format string
218 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
219 * throtl_grp; otherwise, just "throtl".
221 #define throtl_log(sq, fmt, args...) do { \
222 struct throtl_grp *__tg = sq_to_tg((sq)); \
223 struct throtl_data *__td = sq_to_td((sq)); \
226 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
229 blk_add_cgroup_trace_msg(__td->queue, \
230 &tg_to_blkg(__tg)->blkcg->css, "throtl " fmt, ##args);\
232 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
236 static inline unsigned int throtl_bio_data_size(struct bio *bio)
238 /* assume it's one sector */
239 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
241 return bio->bi_iter.bi_size;
244 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
246 INIT_LIST_HEAD(&qn->node);
247 bio_list_init(&qn->bios);
252 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
253 * @bio: bio being added
254 * @qn: qnode to add bio to
255 * @queued: the service_queue->queued[] list @qn belongs to
257 * Add @bio to @qn and put @qn on @queued if it's not already on.
258 * @qn->tg's reference count is bumped when @qn is activated. See the
259 * comment on top of throtl_qnode definition for details.
261 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
262 struct list_head *queued)
264 bio_list_add(&qn->bios, bio);
265 if (list_empty(&qn->node)) {
266 list_add_tail(&qn->node, queued);
267 blkg_get(tg_to_blkg(qn->tg));
272 * throtl_peek_queued - peek the first bio on a qnode list
273 * @queued: the qnode list to peek
275 static struct bio *throtl_peek_queued(struct list_head *queued)
277 struct throtl_qnode *qn;
280 if (list_empty(queued))
283 qn = list_first_entry(queued, struct throtl_qnode, node);
284 bio = bio_list_peek(&qn->bios);
290 * throtl_pop_queued - pop the first bio form a qnode list
291 * @queued: the qnode list to pop a bio from
292 * @tg_to_put: optional out argument for throtl_grp to put
294 * Pop the first bio from the qnode list @queued. After popping, the first
295 * qnode is removed from @queued if empty or moved to the end of @queued so
296 * that the popping order is round-robin.
298 * When the first qnode is removed, its associated throtl_grp should be put
299 * too. If @tg_to_put is NULL, this function automatically puts it;
300 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
301 * responsible for putting it.
303 static struct bio *throtl_pop_queued(struct list_head *queued,
304 struct throtl_grp **tg_to_put)
306 struct throtl_qnode *qn;
309 if (list_empty(queued))
312 qn = list_first_entry(queued, struct throtl_qnode, node);
313 bio = bio_list_pop(&qn->bios);
316 if (bio_list_empty(&qn->bios)) {
317 list_del_init(&qn->node);
321 blkg_put(tg_to_blkg(qn->tg));
323 list_move_tail(&qn->node, queued);
329 /* init a service_queue, assumes the caller zeroed it */
330 static void throtl_service_queue_init(struct throtl_service_queue *sq)
332 INIT_LIST_HEAD(&sq->queued[0]);
333 INIT_LIST_HEAD(&sq->queued[1]);
334 sq->pending_tree = RB_ROOT_CACHED;
335 timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
338 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp,
339 struct request_queue *q,
342 struct throtl_grp *tg;
345 tg = kzalloc_node(sizeof(*tg), gfp, q->node);
349 if (blkg_rwstat_init(&tg->stat_bytes, gfp))
352 if (blkg_rwstat_init(&tg->stat_ios, gfp))
353 goto err_exit_stat_bytes;
355 throtl_service_queue_init(&tg->service_queue);
357 for (rw = READ; rw <= WRITE; rw++) {
358 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
359 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
362 RB_CLEAR_NODE(&tg->rb_node);
363 tg->bps[READ][LIMIT_MAX] = U64_MAX;
364 tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
365 tg->iops[READ][LIMIT_MAX] = UINT_MAX;
366 tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
367 tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
368 tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
369 tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
370 tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
371 /* LIMIT_LOW will have default value 0 */
373 tg->latency_target = DFL_LATENCY_TARGET;
374 tg->latency_target_conf = DFL_LATENCY_TARGET;
375 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
376 tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
381 blkg_rwstat_exit(&tg->stat_bytes);
387 static void throtl_pd_init(struct blkg_policy_data *pd)
389 struct throtl_grp *tg = pd_to_tg(pd);
390 struct blkcg_gq *blkg = tg_to_blkg(tg);
391 struct throtl_data *td = blkg->q->td;
392 struct throtl_service_queue *sq = &tg->service_queue;
395 * If on the default hierarchy, we switch to properly hierarchical
396 * behavior where limits on a given throtl_grp are applied to the
397 * whole subtree rather than just the group itself. e.g. If 16M
398 * read_bps limit is set on the root group, the whole system can't
399 * exceed 16M for the device.
401 * If not on the default hierarchy, the broken flat hierarchy
402 * behavior is retained where all throtl_grps are treated as if
403 * they're all separate root groups right below throtl_data.
404 * Limits of a group don't interact with limits of other groups
405 * regardless of the position of the group in the hierarchy.
407 sq->parent_sq = &td->service_queue;
408 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
409 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
414 * Set has_rules[] if @tg or any of its parents have limits configured.
415 * This doesn't require walking up to the top of the hierarchy as the
416 * parent's has_rules[] is guaranteed to be correct.
418 static void tg_update_has_rules(struct throtl_grp *tg)
420 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
421 struct throtl_data *td = tg->td;
423 int has_iops_limit = 0;
425 for (rw = READ; rw <= WRITE; rw++) {
426 unsigned int iops_limit = tg_iops_limit(tg, rw);
428 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
429 (td->limit_valid[td->limit_index] &&
430 (tg_bps_limit(tg, rw) != U64_MAX ||
431 iops_limit != UINT_MAX));
433 if (iops_limit != UINT_MAX)
438 tg->flags |= THROTL_TG_HAS_IOPS_LIMIT;
440 tg->flags &= ~THROTL_TG_HAS_IOPS_LIMIT;
443 static void throtl_pd_online(struct blkg_policy_data *pd)
445 struct throtl_grp *tg = pd_to_tg(pd);
447 * We don't want new groups to escape the limits of its ancestors.
448 * Update has_rules[] after a new group is brought online.
450 tg_update_has_rules(tg);
453 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
454 static void blk_throtl_update_limit_valid(struct throtl_data *td)
456 struct cgroup_subsys_state *pos_css;
457 struct blkcg_gq *blkg;
458 bool low_valid = false;
461 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
462 struct throtl_grp *tg = blkg_to_tg(blkg);
464 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
465 tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
472 td->limit_valid[LIMIT_LOW] = low_valid;
475 static inline void blk_throtl_update_limit_valid(struct throtl_data *td)
480 static void throtl_upgrade_state(struct throtl_data *td);
481 static void throtl_pd_offline(struct blkg_policy_data *pd)
483 struct throtl_grp *tg = pd_to_tg(pd);
485 tg->bps[READ][LIMIT_LOW] = 0;
486 tg->bps[WRITE][LIMIT_LOW] = 0;
487 tg->iops[READ][LIMIT_LOW] = 0;
488 tg->iops[WRITE][LIMIT_LOW] = 0;
490 blk_throtl_update_limit_valid(tg->td);
492 if (!tg->td->limit_valid[tg->td->limit_index])
493 throtl_upgrade_state(tg->td);
496 static void throtl_pd_free(struct blkg_policy_data *pd)
498 struct throtl_grp *tg = pd_to_tg(pd);
500 del_timer_sync(&tg->service_queue.pending_timer);
501 blkg_rwstat_exit(&tg->stat_bytes);
502 blkg_rwstat_exit(&tg->stat_ios);
506 static struct throtl_grp *
507 throtl_rb_first(struct throtl_service_queue *parent_sq)
511 n = rb_first_cached(&parent_sq->pending_tree);
515 return rb_entry_tg(n);
518 static void throtl_rb_erase(struct rb_node *n,
519 struct throtl_service_queue *parent_sq)
521 rb_erase_cached(n, &parent_sq->pending_tree);
523 --parent_sq->nr_pending;
526 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
528 struct throtl_grp *tg;
530 tg = throtl_rb_first(parent_sq);
534 parent_sq->first_pending_disptime = tg->disptime;
537 static void tg_service_queue_add(struct throtl_grp *tg)
539 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
540 struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
541 struct rb_node *parent = NULL;
542 struct throtl_grp *__tg;
543 unsigned long key = tg->disptime;
544 bool leftmost = true;
546 while (*node != NULL) {
548 __tg = rb_entry_tg(parent);
550 if (time_before(key, __tg->disptime))
551 node = &parent->rb_left;
553 node = &parent->rb_right;
558 rb_link_node(&tg->rb_node, parent, node);
559 rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
563 static void throtl_enqueue_tg(struct throtl_grp *tg)
565 if (!(tg->flags & THROTL_TG_PENDING)) {
566 tg_service_queue_add(tg);
567 tg->flags |= THROTL_TG_PENDING;
568 tg->service_queue.parent_sq->nr_pending++;
572 static void throtl_dequeue_tg(struct throtl_grp *tg)
574 if (tg->flags & THROTL_TG_PENDING) {
575 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
576 tg->flags &= ~THROTL_TG_PENDING;
580 /* Call with queue lock held */
581 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
582 unsigned long expires)
584 unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
587 * Since we are adjusting the throttle limit dynamically, the sleep
588 * time calculated according to previous limit might be invalid. It's
589 * possible the cgroup sleep time is very long and no other cgroups
590 * have IO running so notify the limit changes. Make sure the cgroup
591 * doesn't sleep too long to avoid the missed notification.
593 if (time_after(expires, max_expire))
594 expires = max_expire;
595 mod_timer(&sq->pending_timer, expires);
596 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
597 expires - jiffies, jiffies);
601 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
602 * @sq: the service_queue to schedule dispatch for
603 * @force: force scheduling
605 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
606 * dispatch time of the first pending child. Returns %true if either timer
607 * is armed or there's no pending child left. %false if the current
608 * dispatch window is still open and the caller should continue
611 * If @force is %true, the dispatch timer is always scheduled and this
612 * function is guaranteed to return %true. This is to be used when the
613 * caller can't dispatch itself and needs to invoke pending_timer
614 * unconditionally. Note that forced scheduling is likely to induce short
615 * delay before dispatch starts even if @sq->first_pending_disptime is not
616 * in the future and thus shouldn't be used in hot paths.
618 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
621 /* any pending children left? */
625 update_min_dispatch_time(sq);
627 /* is the next dispatch time in the future? */
628 if (force || time_after(sq->first_pending_disptime, jiffies)) {
629 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
633 /* tell the caller to continue dispatching */
637 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
638 bool rw, unsigned long start)
640 tg->bytes_disp[rw] = 0;
644 * Previous slice has expired. We must have trimmed it after last
645 * bio dispatch. That means since start of last slice, we never used
646 * that bandwidth. Do try to make use of that bandwidth while giving
649 if (time_after_eq(start, tg->slice_start[rw]))
650 tg->slice_start[rw] = start;
652 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
653 throtl_log(&tg->service_queue,
654 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
655 rw == READ ? 'R' : 'W', tg->slice_start[rw],
656 tg->slice_end[rw], jiffies);
659 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
661 tg->bytes_disp[rw] = 0;
663 tg->slice_start[rw] = jiffies;
664 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
666 throtl_log(&tg->service_queue,
667 "[%c] new slice start=%lu end=%lu jiffies=%lu",
668 rw == READ ? 'R' : 'W', tg->slice_start[rw],
669 tg->slice_end[rw], jiffies);
672 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
673 unsigned long jiffy_end)
675 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
678 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
679 unsigned long jiffy_end)
681 throtl_set_slice_end(tg, rw, jiffy_end);
682 throtl_log(&tg->service_queue,
683 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
684 rw == READ ? 'R' : 'W', tg->slice_start[rw],
685 tg->slice_end[rw], jiffies);
688 /* Determine if previously allocated or extended slice is complete or not */
689 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
691 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
697 /* Trim the used slices and adjust slice start accordingly */
698 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
700 unsigned long nr_slices, time_elapsed, io_trim;
703 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
706 * If bps are unlimited (-1), then time slice don't get
707 * renewed. Don't try to trim the slice if slice is used. A new
708 * slice will start when appropriate.
710 if (throtl_slice_used(tg, rw))
714 * A bio has been dispatched. Also adjust slice_end. It might happen
715 * that initially cgroup limit was very low resulting in high
716 * slice_end, but later limit was bumped up and bio was dispatched
717 * sooner, then we need to reduce slice_end. A high bogus slice_end
718 * is bad because it does not allow new slice to start.
721 throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
723 time_elapsed = jiffies - tg->slice_start[rw];
725 nr_slices = time_elapsed / tg->td->throtl_slice;
729 tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
733 io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
736 if (!bytes_trim && !io_trim)
739 if (tg->bytes_disp[rw] >= bytes_trim)
740 tg->bytes_disp[rw] -= bytes_trim;
742 tg->bytes_disp[rw] = 0;
744 if (tg->io_disp[rw] >= io_trim)
745 tg->io_disp[rw] -= io_trim;
749 tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
751 throtl_log(&tg->service_queue,
752 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
753 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
754 tg->slice_start[rw], tg->slice_end[rw], jiffies);
757 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
758 u32 iops_limit, unsigned long *wait)
760 bool rw = bio_data_dir(bio);
761 unsigned int io_allowed;
762 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
765 if (iops_limit == UINT_MAX) {
771 jiffy_elapsed = jiffies - tg->slice_start[rw];
773 /* Round up to the next throttle slice, wait time must be nonzero */
774 jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
777 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
778 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
779 * will allow dispatch after 1 second and after that slice should
783 tmp = (u64)iops_limit * jiffy_elapsed_rnd;
787 io_allowed = UINT_MAX;
791 if (tg->io_disp[rw] + 1 <= io_allowed) {
797 /* Calc approx time to dispatch */
798 jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
805 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
806 u64 bps_limit, unsigned long *wait)
808 bool rw = bio_data_dir(bio);
809 u64 bytes_allowed, extra_bytes;
810 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
811 unsigned int bio_size = throtl_bio_data_size(bio);
813 /* no need to throttle if this bio's bytes have been accounted */
814 if (bps_limit == U64_MAX || bio_flagged(bio, BIO_BPS_THROTTLED)) {
820 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
822 /* Slice has just started. Consider one slice interval */
824 jiffy_elapsed_rnd = tg->td->throtl_slice;
826 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
827 bytes_allowed = mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed_rnd,
830 if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
836 /* Calc approx time to dispatch */
837 extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
838 jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
844 * This wait time is without taking into consideration the rounding
845 * up we did. Add that time also.
847 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
854 * Returns whether one can dispatch a bio or not. Also returns approx number
855 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
857 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
860 bool rw = bio_data_dir(bio);
861 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
862 u64 bps_limit = tg_bps_limit(tg, rw);
863 u32 iops_limit = tg_iops_limit(tg, rw);
866 * Currently whole state machine of group depends on first bio
867 * queued in the group bio list. So one should not be calling
868 * this function with a different bio if there are other bios
871 BUG_ON(tg->service_queue.nr_queued[rw] &&
872 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
874 /* If tg->bps = -1, then BW is unlimited */
875 if ((bps_limit == U64_MAX && iops_limit == UINT_MAX) ||
876 tg->flags & THROTL_TG_CANCELING) {
883 * If previous slice expired, start a new one otherwise renew/extend
884 * existing slice to make sure it is at least throtl_slice interval
885 * long since now. New slice is started only for empty throttle group.
886 * If there is queued bio, that means there should be an active
887 * slice and it should be extended instead.
889 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
890 throtl_start_new_slice(tg, rw);
892 if (time_before(tg->slice_end[rw],
893 jiffies + tg->td->throtl_slice))
894 throtl_extend_slice(tg, rw,
895 jiffies + tg->td->throtl_slice);
898 if (tg_with_in_bps_limit(tg, bio, bps_limit, &bps_wait) &&
899 tg_with_in_iops_limit(tg, bio, iops_limit, &iops_wait)) {
905 max_wait = max(bps_wait, iops_wait);
910 if (time_before(tg->slice_end[rw], jiffies + max_wait))
911 throtl_extend_slice(tg, rw, jiffies + max_wait);
916 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
918 bool rw = bio_data_dir(bio);
919 unsigned int bio_size = throtl_bio_data_size(bio);
921 /* Charge the bio to the group */
922 if (!bio_flagged(bio, BIO_BPS_THROTTLED)) {
923 tg->bytes_disp[rw] += bio_size;
924 tg->last_bytes_disp[rw] += bio_size;
928 tg->last_io_disp[rw]++;
932 * throtl_add_bio_tg - add a bio to the specified throtl_grp
935 * @tg: the target throtl_grp
937 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
938 * tg->qnode_on_self[] is used.
940 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
941 struct throtl_grp *tg)
943 struct throtl_service_queue *sq = &tg->service_queue;
944 bool rw = bio_data_dir(bio);
947 qn = &tg->qnode_on_self[rw];
950 * If @tg doesn't currently have any bios queued in the same
951 * direction, queueing @bio can change when @tg should be
952 * dispatched. Mark that @tg was empty. This is automatically
953 * cleared on the next tg_update_disptime().
955 if (!sq->nr_queued[rw])
956 tg->flags |= THROTL_TG_WAS_EMPTY;
958 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
961 throtl_enqueue_tg(tg);
964 static void tg_update_disptime(struct throtl_grp *tg)
966 struct throtl_service_queue *sq = &tg->service_queue;
967 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
970 bio = throtl_peek_queued(&sq->queued[READ]);
972 tg_may_dispatch(tg, bio, &read_wait);
974 bio = throtl_peek_queued(&sq->queued[WRITE]);
976 tg_may_dispatch(tg, bio, &write_wait);
978 min_wait = min(read_wait, write_wait);
979 disptime = jiffies + min_wait;
981 /* Update dispatch time */
982 throtl_dequeue_tg(tg);
983 tg->disptime = disptime;
984 throtl_enqueue_tg(tg);
986 /* see throtl_add_bio_tg() */
987 tg->flags &= ~THROTL_TG_WAS_EMPTY;
990 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
991 struct throtl_grp *parent_tg, bool rw)
993 if (throtl_slice_used(parent_tg, rw)) {
994 throtl_start_new_slice_with_credit(parent_tg, rw,
995 child_tg->slice_start[rw]);
1000 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1002 struct throtl_service_queue *sq = &tg->service_queue;
1003 struct throtl_service_queue *parent_sq = sq->parent_sq;
1004 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1005 struct throtl_grp *tg_to_put = NULL;
1009 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1010 * from @tg may put its reference and @parent_sq might end up
1011 * getting released prematurely. Remember the tg to put and put it
1012 * after @bio is transferred to @parent_sq.
1014 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1015 sq->nr_queued[rw]--;
1017 throtl_charge_bio(tg, bio);
1018 bio_set_flag(bio, BIO_BPS_THROTTLED);
1021 * If our parent is another tg, we just need to transfer @bio to
1022 * the parent using throtl_add_bio_tg(). If our parent is
1023 * @td->service_queue, @bio is ready to be issued. Put it on its
1024 * bio_lists[] and decrease total number queued. The caller is
1025 * responsible for issuing these bios.
1028 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1029 start_parent_slice_with_credit(tg, parent_tg, rw);
1031 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1032 &parent_sq->queued[rw]);
1033 BUG_ON(tg->td->nr_queued[rw] <= 0);
1034 tg->td->nr_queued[rw]--;
1037 throtl_trim_slice(tg, rw);
1040 blkg_put(tg_to_blkg(tg_to_put));
1043 static int throtl_dispatch_tg(struct throtl_grp *tg)
1045 struct throtl_service_queue *sq = &tg->service_queue;
1046 unsigned int nr_reads = 0, nr_writes = 0;
1047 unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1048 unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1051 /* Try to dispatch 75% READS and 25% WRITES */
1053 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1054 tg_may_dispatch(tg, bio, NULL)) {
1056 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1059 if (nr_reads >= max_nr_reads)
1063 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1064 tg_may_dispatch(tg, bio, NULL)) {
1066 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1069 if (nr_writes >= max_nr_writes)
1073 return nr_reads + nr_writes;
1076 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1078 unsigned int nr_disp = 0;
1081 struct throtl_grp *tg;
1082 struct throtl_service_queue *sq;
1084 if (!parent_sq->nr_pending)
1087 tg = throtl_rb_first(parent_sq);
1091 if (time_before(jiffies, tg->disptime))
1094 throtl_dequeue_tg(tg);
1096 nr_disp += throtl_dispatch_tg(tg);
1098 sq = &tg->service_queue;
1099 if (sq->nr_queued[0] || sq->nr_queued[1])
1100 tg_update_disptime(tg);
1102 if (nr_disp >= THROTL_QUANTUM)
1109 static bool throtl_can_upgrade(struct throtl_data *td,
1110 struct throtl_grp *this_tg);
1112 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1113 * @t: the pending_timer member of the throtl_service_queue being serviced
1115 * This timer is armed when a child throtl_grp with active bio's become
1116 * pending and queued on the service_queue's pending_tree and expires when
1117 * the first child throtl_grp should be dispatched. This function
1118 * dispatches bio's from the children throtl_grps to the parent
1121 * If the parent's parent is another throtl_grp, dispatching is propagated
1122 * by either arming its pending_timer or repeating dispatch directly. If
1123 * the top-level service_tree is reached, throtl_data->dispatch_work is
1124 * kicked so that the ready bio's are issued.
1126 static void throtl_pending_timer_fn(struct timer_list *t)
1128 struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1129 struct throtl_grp *tg = sq_to_tg(sq);
1130 struct throtl_data *td = sq_to_td(sq);
1131 struct throtl_service_queue *parent_sq;
1132 struct request_queue *q;
1136 /* throtl_data may be gone, so figure out request queue by blkg */
1142 spin_lock_irq(&q->queue_lock);
1147 if (throtl_can_upgrade(td, NULL))
1148 throtl_upgrade_state(td);
1151 parent_sq = sq->parent_sq;
1155 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1156 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1157 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1159 ret = throtl_select_dispatch(sq);
1161 throtl_log(sq, "bios disp=%u", ret);
1165 if (throtl_schedule_next_dispatch(sq, false))
1168 /* this dispatch windows is still open, relax and repeat */
1169 spin_unlock_irq(&q->queue_lock);
1171 spin_lock_irq(&q->queue_lock);
1178 /* @parent_sq is another throl_grp, propagate dispatch */
1179 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1180 tg_update_disptime(tg);
1181 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1182 /* window is already open, repeat dispatching */
1189 /* reached the top-level, queue issuing */
1190 queue_work(kthrotld_workqueue, &td->dispatch_work);
1193 spin_unlock_irq(&q->queue_lock);
1197 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1198 * @work: work item being executed
1200 * This function is queued for execution when bios reach the bio_lists[]
1201 * of throtl_data->service_queue. Those bios are ready and issued by this
1204 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1206 struct throtl_data *td = container_of(work, struct throtl_data,
1208 struct throtl_service_queue *td_sq = &td->service_queue;
1209 struct request_queue *q = td->queue;
1210 struct bio_list bio_list_on_stack;
1212 struct blk_plug plug;
1215 bio_list_init(&bio_list_on_stack);
1217 spin_lock_irq(&q->queue_lock);
1218 for (rw = READ; rw <= WRITE; rw++)
1219 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1220 bio_list_add(&bio_list_on_stack, bio);
1221 spin_unlock_irq(&q->queue_lock);
1223 if (!bio_list_empty(&bio_list_on_stack)) {
1224 blk_start_plug(&plug);
1225 while ((bio = bio_list_pop(&bio_list_on_stack)))
1226 submit_bio_noacct_nocheck(bio);
1227 blk_finish_plug(&plug);
1231 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1234 struct throtl_grp *tg = pd_to_tg(pd);
1235 u64 v = *(u64 *)((void *)tg + off);
1239 return __blkg_prfill_u64(sf, pd, v);
1242 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1245 struct throtl_grp *tg = pd_to_tg(pd);
1246 unsigned int v = *(unsigned int *)((void *)tg + off);
1250 return __blkg_prfill_u64(sf, pd, v);
1253 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1255 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1256 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1260 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1262 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1263 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1267 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1269 struct throtl_service_queue *sq = &tg->service_queue;
1270 struct cgroup_subsys_state *pos_css;
1271 struct blkcg_gq *blkg;
1273 throtl_log(&tg->service_queue,
1274 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1275 tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1276 tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1279 * Update has_rules[] flags for the updated tg's subtree. A tg is
1280 * considered to have rules if either the tg itself or any of its
1281 * ancestors has rules. This identifies groups without any
1282 * restrictions in the whole hierarchy and allows them to bypass
1285 blkg_for_each_descendant_pre(blkg, pos_css,
1286 global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1287 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1288 struct throtl_grp *parent_tg;
1290 tg_update_has_rules(this_tg);
1291 /* ignore root/second level */
1292 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1293 !blkg->parent->parent)
1295 parent_tg = blkg_to_tg(blkg->parent);
1297 * make sure all children has lower idle time threshold and
1298 * higher latency target
1300 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1301 parent_tg->idletime_threshold);
1302 this_tg->latency_target = max(this_tg->latency_target,
1303 parent_tg->latency_target);
1307 * We're already holding queue_lock and know @tg is valid. Let's
1308 * apply the new config directly.
1310 * Restart the slices for both READ and WRITES. It might happen
1311 * that a group's limit are dropped suddenly and we don't want to
1312 * account recently dispatched IO with new low rate.
1314 throtl_start_new_slice(tg, READ);
1315 throtl_start_new_slice(tg, WRITE);
1317 if (tg->flags & THROTL_TG_PENDING) {
1318 tg_update_disptime(tg);
1319 throtl_schedule_next_dispatch(sq->parent_sq, true);
1323 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1324 char *buf, size_t nbytes, loff_t off, bool is_u64)
1326 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1327 struct blkg_conf_ctx ctx;
1328 struct throtl_grp *tg;
1332 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1337 if (sscanf(ctx.body, "%llu", &v) != 1)
1342 tg = blkg_to_tg(ctx.blkg);
1345 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1347 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1349 tg_conf_updated(tg, false);
1352 blkg_conf_finish(&ctx);
1353 return ret ?: nbytes;
1356 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1357 char *buf, size_t nbytes, loff_t off)
1359 return tg_set_conf(of, buf, nbytes, off, true);
1362 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1363 char *buf, size_t nbytes, loff_t off)
1365 return tg_set_conf(of, buf, nbytes, off, false);
1368 static int tg_print_rwstat(struct seq_file *sf, void *v)
1370 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1371 blkg_prfill_rwstat, &blkcg_policy_throtl,
1372 seq_cft(sf)->private, true);
1376 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1377 struct blkg_policy_data *pd, int off)
1379 struct blkg_rwstat_sample sum;
1381 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1383 return __blkg_prfill_rwstat(sf, pd, &sum);
1386 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1388 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1389 tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1390 seq_cft(sf)->private, true);
1394 static struct cftype throtl_legacy_files[] = {
1396 .name = "throttle.read_bps_device",
1397 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1398 .seq_show = tg_print_conf_u64,
1399 .write = tg_set_conf_u64,
1402 .name = "throttle.write_bps_device",
1403 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1404 .seq_show = tg_print_conf_u64,
1405 .write = tg_set_conf_u64,
1408 .name = "throttle.read_iops_device",
1409 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1410 .seq_show = tg_print_conf_uint,
1411 .write = tg_set_conf_uint,
1414 .name = "throttle.write_iops_device",
1415 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1416 .seq_show = tg_print_conf_uint,
1417 .write = tg_set_conf_uint,
1420 .name = "throttle.io_service_bytes",
1421 .private = offsetof(struct throtl_grp, stat_bytes),
1422 .seq_show = tg_print_rwstat,
1425 .name = "throttle.io_service_bytes_recursive",
1426 .private = offsetof(struct throtl_grp, stat_bytes),
1427 .seq_show = tg_print_rwstat_recursive,
1430 .name = "throttle.io_serviced",
1431 .private = offsetof(struct throtl_grp, stat_ios),
1432 .seq_show = tg_print_rwstat,
1435 .name = "throttle.io_serviced_recursive",
1436 .private = offsetof(struct throtl_grp, stat_ios),
1437 .seq_show = tg_print_rwstat_recursive,
1442 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1445 struct throtl_grp *tg = pd_to_tg(pd);
1446 const char *dname = blkg_dev_name(pd->blkg);
1447 char bufs[4][21] = { "max", "max", "max", "max" };
1449 unsigned int iops_dft;
1450 char idle_time[26] = "";
1451 char latency_time[26] = "";
1456 if (off == LIMIT_LOW) {
1461 iops_dft = UINT_MAX;
1464 if (tg->bps_conf[READ][off] == bps_dft &&
1465 tg->bps_conf[WRITE][off] == bps_dft &&
1466 tg->iops_conf[READ][off] == iops_dft &&
1467 tg->iops_conf[WRITE][off] == iops_dft &&
1468 (off != LIMIT_LOW ||
1469 (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1470 tg->latency_target_conf == DFL_LATENCY_TARGET)))
1473 if (tg->bps_conf[READ][off] != U64_MAX)
1474 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1475 tg->bps_conf[READ][off]);
1476 if (tg->bps_conf[WRITE][off] != U64_MAX)
1477 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1478 tg->bps_conf[WRITE][off]);
1479 if (tg->iops_conf[READ][off] != UINT_MAX)
1480 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1481 tg->iops_conf[READ][off]);
1482 if (tg->iops_conf[WRITE][off] != UINT_MAX)
1483 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1484 tg->iops_conf[WRITE][off]);
1485 if (off == LIMIT_LOW) {
1486 if (tg->idletime_threshold_conf == ULONG_MAX)
1487 strcpy(idle_time, " idle=max");
1489 snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1490 tg->idletime_threshold_conf);
1492 if (tg->latency_target_conf == ULONG_MAX)
1493 strcpy(latency_time, " latency=max");
1495 snprintf(latency_time, sizeof(latency_time),
1496 " latency=%lu", tg->latency_target_conf);
1499 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1500 dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1505 static int tg_print_limit(struct seq_file *sf, void *v)
1507 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1508 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1512 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1513 char *buf, size_t nbytes, loff_t off)
1515 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1516 struct blkg_conf_ctx ctx;
1517 struct throtl_grp *tg;
1519 unsigned long idle_time;
1520 unsigned long latency_time;
1522 int index = of_cft(of)->private;
1524 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1528 tg = blkg_to_tg(ctx.blkg);
1530 v[0] = tg->bps_conf[READ][index];
1531 v[1] = tg->bps_conf[WRITE][index];
1532 v[2] = tg->iops_conf[READ][index];
1533 v[3] = tg->iops_conf[WRITE][index];
1535 idle_time = tg->idletime_threshold_conf;
1536 latency_time = tg->latency_target_conf;
1538 char tok[27]; /* wiops=18446744073709551616 */
1543 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1552 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1560 if (!strcmp(tok, "rbps") && val > 1)
1562 else if (!strcmp(tok, "wbps") && val > 1)
1564 else if (!strcmp(tok, "riops") && val > 1)
1565 v[2] = min_t(u64, val, UINT_MAX);
1566 else if (!strcmp(tok, "wiops") && val > 1)
1567 v[3] = min_t(u64, val, UINT_MAX);
1568 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1570 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1576 tg->bps_conf[READ][index] = v[0];
1577 tg->bps_conf[WRITE][index] = v[1];
1578 tg->iops_conf[READ][index] = v[2];
1579 tg->iops_conf[WRITE][index] = v[3];
1581 if (index == LIMIT_MAX) {
1582 tg->bps[READ][index] = v[0];
1583 tg->bps[WRITE][index] = v[1];
1584 tg->iops[READ][index] = v[2];
1585 tg->iops[WRITE][index] = v[3];
1587 tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1588 tg->bps_conf[READ][LIMIT_MAX]);
1589 tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1590 tg->bps_conf[WRITE][LIMIT_MAX]);
1591 tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1592 tg->iops_conf[READ][LIMIT_MAX]);
1593 tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1594 tg->iops_conf[WRITE][LIMIT_MAX]);
1595 tg->idletime_threshold_conf = idle_time;
1596 tg->latency_target_conf = latency_time;
1598 /* force user to configure all settings for low limit */
1599 if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1600 tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1601 tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1602 tg->latency_target_conf == DFL_LATENCY_TARGET) {
1603 tg->bps[READ][LIMIT_LOW] = 0;
1604 tg->bps[WRITE][LIMIT_LOW] = 0;
1605 tg->iops[READ][LIMIT_LOW] = 0;
1606 tg->iops[WRITE][LIMIT_LOW] = 0;
1607 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1608 tg->latency_target = DFL_LATENCY_TARGET;
1609 } else if (index == LIMIT_LOW) {
1610 tg->idletime_threshold = tg->idletime_threshold_conf;
1611 tg->latency_target = tg->latency_target_conf;
1614 blk_throtl_update_limit_valid(tg->td);
1615 if (tg->td->limit_valid[LIMIT_LOW]) {
1616 if (index == LIMIT_LOW)
1617 tg->td->limit_index = LIMIT_LOW;
1619 tg->td->limit_index = LIMIT_MAX;
1620 tg_conf_updated(tg, index == LIMIT_LOW &&
1621 tg->td->limit_valid[LIMIT_LOW]);
1624 blkg_conf_finish(&ctx);
1625 return ret ?: nbytes;
1628 static struct cftype throtl_files[] = {
1629 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1632 .flags = CFTYPE_NOT_ON_ROOT,
1633 .seq_show = tg_print_limit,
1634 .write = tg_set_limit,
1635 .private = LIMIT_LOW,
1640 .flags = CFTYPE_NOT_ON_ROOT,
1641 .seq_show = tg_print_limit,
1642 .write = tg_set_limit,
1643 .private = LIMIT_MAX,
1648 static void throtl_shutdown_wq(struct request_queue *q)
1650 struct throtl_data *td = q->td;
1652 cancel_work_sync(&td->dispatch_work);
1655 struct blkcg_policy blkcg_policy_throtl = {
1656 .dfl_cftypes = throtl_files,
1657 .legacy_cftypes = throtl_legacy_files,
1659 .pd_alloc_fn = throtl_pd_alloc,
1660 .pd_init_fn = throtl_pd_init,
1661 .pd_online_fn = throtl_pd_online,
1662 .pd_offline_fn = throtl_pd_offline,
1663 .pd_free_fn = throtl_pd_free,
1666 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1668 unsigned long rtime = jiffies, wtime = jiffies;
1670 if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1671 rtime = tg->last_low_overflow_time[READ];
1672 if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1673 wtime = tg->last_low_overflow_time[WRITE];
1674 return min(rtime, wtime);
1677 /* tg should not be an intermediate node */
1678 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1680 struct throtl_service_queue *parent_sq;
1681 struct throtl_grp *parent = tg;
1682 unsigned long ret = __tg_last_low_overflow_time(tg);
1685 parent_sq = parent->service_queue.parent_sq;
1686 parent = sq_to_tg(parent_sq);
1691 * The parent doesn't have low limit, it always reaches low
1692 * limit. Its overflow time is useless for children
1694 if (!parent->bps[READ][LIMIT_LOW] &&
1695 !parent->iops[READ][LIMIT_LOW] &&
1696 !parent->bps[WRITE][LIMIT_LOW] &&
1697 !parent->iops[WRITE][LIMIT_LOW])
1699 if (time_after(__tg_last_low_overflow_time(parent), ret))
1700 ret = __tg_last_low_overflow_time(parent);
1705 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1708 * cgroup is idle if:
1709 * - single idle is too long, longer than a fixed value (in case user
1710 * configure a too big threshold) or 4 times of idletime threshold
1711 * - average think time is more than threshold
1712 * - IO latency is largely below threshold
1717 time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1718 ret = tg->latency_target == DFL_LATENCY_TARGET ||
1719 tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1720 (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1721 tg->avg_idletime > tg->idletime_threshold ||
1722 (tg->latency_target && tg->bio_cnt &&
1723 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1724 throtl_log(&tg->service_queue,
1725 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1726 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1727 tg->bio_cnt, ret, tg->td->scale);
1731 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1733 struct throtl_service_queue *sq = &tg->service_queue;
1734 bool read_limit, write_limit;
1737 * if cgroup reaches low limit (if low limit is 0, the cgroup always
1738 * reaches), it's ok to upgrade to next limit
1740 read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1741 write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1742 if (!read_limit && !write_limit)
1744 if (read_limit && sq->nr_queued[READ] &&
1745 (!write_limit || sq->nr_queued[WRITE]))
1747 if (write_limit && sq->nr_queued[WRITE] &&
1748 (!read_limit || sq->nr_queued[READ]))
1751 if (time_after_eq(jiffies,
1752 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1753 throtl_tg_is_idle(tg))
1758 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1761 if (throtl_tg_can_upgrade(tg))
1763 tg = sq_to_tg(tg->service_queue.parent_sq);
1764 if (!tg || !tg_to_blkg(tg)->parent)
1770 void blk_throtl_cancel_bios(struct request_queue *q)
1772 struct cgroup_subsys_state *pos_css;
1773 struct blkcg_gq *blkg;
1775 spin_lock_irq(&q->queue_lock);
1777 * queue_lock is held, rcu lock is not needed here technically.
1778 * However, rcu lock is still held to emphasize that following
1779 * path need RCU protection and to prevent warning from lockdep.
1782 blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) {
1783 struct throtl_grp *tg = blkg_to_tg(blkg);
1784 struct throtl_service_queue *sq = &tg->service_queue;
1787 * Set the flag to make sure throtl_pending_timer_fn() won't
1788 * stop until all throttled bios are dispatched.
1790 blkg_to_tg(blkg)->flags |= THROTL_TG_CANCELING;
1792 * Update disptime after setting the above flag to make sure
1793 * throtl_select_dispatch() won't exit without dispatching.
1795 tg_update_disptime(tg);
1797 throtl_schedule_pending_timer(sq, jiffies + 1);
1800 spin_unlock_irq(&q->queue_lock);
1803 static bool throtl_can_upgrade(struct throtl_data *td,
1804 struct throtl_grp *this_tg)
1806 struct cgroup_subsys_state *pos_css;
1807 struct blkcg_gq *blkg;
1809 if (td->limit_index != LIMIT_LOW)
1812 if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1816 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1817 struct throtl_grp *tg = blkg_to_tg(blkg);
1821 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1823 if (!throtl_hierarchy_can_upgrade(tg)) {
1832 static void throtl_upgrade_check(struct throtl_grp *tg)
1834 unsigned long now = jiffies;
1836 if (tg->td->limit_index != LIMIT_LOW)
1839 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1842 tg->last_check_time = now;
1844 if (!time_after_eq(now,
1845 __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1848 if (throtl_can_upgrade(tg->td, NULL))
1849 throtl_upgrade_state(tg->td);
1852 static void throtl_upgrade_state(struct throtl_data *td)
1854 struct cgroup_subsys_state *pos_css;
1855 struct blkcg_gq *blkg;
1857 throtl_log(&td->service_queue, "upgrade to max");
1858 td->limit_index = LIMIT_MAX;
1859 td->low_upgrade_time = jiffies;
1862 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1863 struct throtl_grp *tg = blkg_to_tg(blkg);
1864 struct throtl_service_queue *sq = &tg->service_queue;
1866 tg->disptime = jiffies - 1;
1867 throtl_select_dispatch(sq);
1868 throtl_schedule_next_dispatch(sq, true);
1871 throtl_select_dispatch(&td->service_queue);
1872 throtl_schedule_next_dispatch(&td->service_queue, true);
1873 queue_work(kthrotld_workqueue, &td->dispatch_work);
1876 static void throtl_downgrade_state(struct throtl_data *td)
1880 throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1882 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1886 td->limit_index = LIMIT_LOW;
1887 td->low_downgrade_time = jiffies;
1890 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1892 struct throtl_data *td = tg->td;
1893 unsigned long now = jiffies;
1896 * If cgroup is below low limit, consider downgrade and throttle other
1899 if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1900 time_after_eq(now, tg_last_low_overflow_time(tg) +
1901 td->throtl_slice) &&
1902 (!throtl_tg_is_idle(tg) ||
1903 !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1908 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1911 if (!throtl_tg_can_downgrade(tg))
1913 tg = sq_to_tg(tg->service_queue.parent_sq);
1914 if (!tg || !tg_to_blkg(tg)->parent)
1920 static void throtl_downgrade_check(struct throtl_grp *tg)
1924 unsigned long elapsed_time;
1925 unsigned long now = jiffies;
1927 if (tg->td->limit_index != LIMIT_MAX ||
1928 !tg->td->limit_valid[LIMIT_LOW])
1930 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1932 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1935 elapsed_time = now - tg->last_check_time;
1936 tg->last_check_time = now;
1938 if (time_before(now, tg_last_low_overflow_time(tg) +
1939 tg->td->throtl_slice))
1942 if (tg->bps[READ][LIMIT_LOW]) {
1943 bps = tg->last_bytes_disp[READ] * HZ;
1944 do_div(bps, elapsed_time);
1945 if (bps >= tg->bps[READ][LIMIT_LOW])
1946 tg->last_low_overflow_time[READ] = now;
1949 if (tg->bps[WRITE][LIMIT_LOW]) {
1950 bps = tg->last_bytes_disp[WRITE] * HZ;
1951 do_div(bps, elapsed_time);
1952 if (bps >= tg->bps[WRITE][LIMIT_LOW])
1953 tg->last_low_overflow_time[WRITE] = now;
1956 if (tg->iops[READ][LIMIT_LOW]) {
1957 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
1958 if (iops >= tg->iops[READ][LIMIT_LOW])
1959 tg->last_low_overflow_time[READ] = now;
1962 if (tg->iops[WRITE][LIMIT_LOW]) {
1963 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
1964 if (iops >= tg->iops[WRITE][LIMIT_LOW])
1965 tg->last_low_overflow_time[WRITE] = now;
1969 * If cgroup is below low limit, consider downgrade and throttle other
1972 if (throtl_hierarchy_can_downgrade(tg))
1973 throtl_downgrade_state(tg->td);
1975 tg->last_bytes_disp[READ] = 0;
1976 tg->last_bytes_disp[WRITE] = 0;
1977 tg->last_io_disp[READ] = 0;
1978 tg->last_io_disp[WRITE] = 0;
1981 static void blk_throtl_update_idletime(struct throtl_grp *tg)
1984 unsigned long last_finish_time = tg->last_finish_time;
1986 if (last_finish_time == 0)
1989 now = ktime_get_ns() >> 10;
1990 if (now <= last_finish_time ||
1991 last_finish_time == tg->checked_last_finish_time)
1994 tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
1995 tg->checked_last_finish_time = last_finish_time;
1998 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1999 static void throtl_update_latency_buckets(struct throtl_data *td)
2001 struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2003 unsigned long last_latency[2] = { 0 };
2004 unsigned long latency[2];
2006 if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
2008 if (time_before(jiffies, td->last_calculate_time + HZ))
2010 td->last_calculate_time = jiffies;
2012 memset(avg_latency, 0, sizeof(avg_latency));
2013 for (rw = READ; rw <= WRITE; rw++) {
2014 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2015 struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2017 for_each_possible_cpu(cpu) {
2018 struct latency_bucket *bucket;
2020 /* this isn't race free, but ok in practice */
2021 bucket = per_cpu_ptr(td->latency_buckets[rw],
2023 tmp->total_latency += bucket[i].total_latency;
2024 tmp->samples += bucket[i].samples;
2025 bucket[i].total_latency = 0;
2026 bucket[i].samples = 0;
2029 if (tmp->samples >= 32) {
2030 int samples = tmp->samples;
2032 latency[rw] = tmp->total_latency;
2034 tmp->total_latency = 0;
2036 latency[rw] /= samples;
2037 if (latency[rw] == 0)
2039 avg_latency[rw][i].latency = latency[rw];
2044 for (rw = READ; rw <= WRITE; rw++) {
2045 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2046 if (!avg_latency[rw][i].latency) {
2047 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2048 td->avg_buckets[rw][i].latency =
2053 if (!td->avg_buckets[rw][i].valid)
2054 latency[rw] = avg_latency[rw][i].latency;
2056 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2057 avg_latency[rw][i].latency) >> 3;
2059 td->avg_buckets[rw][i].latency = max(latency[rw],
2061 td->avg_buckets[rw][i].valid = true;
2062 last_latency[rw] = td->avg_buckets[rw][i].latency;
2066 for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2067 throtl_log(&td->service_queue,
2068 "Latency bucket %d: read latency=%ld, read valid=%d, "
2069 "write latency=%ld, write valid=%d", i,
2070 td->avg_buckets[READ][i].latency,
2071 td->avg_buckets[READ][i].valid,
2072 td->avg_buckets[WRITE][i].latency,
2073 td->avg_buckets[WRITE][i].valid);
2076 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2081 bool __blk_throtl_bio(struct bio *bio)
2083 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2084 struct blkcg_gq *blkg = bio->bi_blkg;
2085 struct throtl_qnode *qn = NULL;
2086 struct throtl_grp *tg = blkg_to_tg(blkg);
2087 struct throtl_service_queue *sq;
2088 bool rw = bio_data_dir(bio);
2089 bool throttled = false;
2090 struct throtl_data *td = tg->td;
2094 if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
2095 blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
2096 bio->bi_iter.bi_size);
2097 blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
2100 spin_lock_irq(&q->queue_lock);
2102 throtl_update_latency_buckets(td);
2104 blk_throtl_update_idletime(tg);
2106 sq = &tg->service_queue;
2110 if (tg->last_low_overflow_time[rw] == 0)
2111 tg->last_low_overflow_time[rw] = jiffies;
2112 throtl_downgrade_check(tg);
2113 throtl_upgrade_check(tg);
2114 /* throtl is FIFO - if bios are already queued, should queue */
2115 if (sq->nr_queued[rw])
2118 /* if above limits, break to queue */
2119 if (!tg_may_dispatch(tg, bio, NULL)) {
2120 tg->last_low_overflow_time[rw] = jiffies;
2121 if (throtl_can_upgrade(td, tg)) {
2122 throtl_upgrade_state(td);
2128 /* within limits, let's charge and dispatch directly */
2129 throtl_charge_bio(tg, bio);
2132 * We need to trim slice even when bios are not being queued
2133 * otherwise it might happen that a bio is not queued for
2134 * a long time and slice keeps on extending and trim is not
2135 * called for a long time. Now if limits are reduced suddenly
2136 * we take into account all the IO dispatched so far at new
2137 * low rate and * newly queued IO gets a really long dispatch
2140 * So keep on trimming slice even if bio is not queued.
2142 throtl_trim_slice(tg, rw);
2145 * @bio passed through this layer without being throttled.
2146 * Climb up the ladder. If we're already at the top, it
2147 * can be executed directly.
2149 qn = &tg->qnode_on_parent[rw];
2153 bio_set_flag(bio, BIO_BPS_THROTTLED);
2158 /* out-of-limit, queue to @tg */
2159 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2160 rw == READ ? 'R' : 'W',
2161 tg->bytes_disp[rw], bio->bi_iter.bi_size,
2162 tg_bps_limit(tg, rw),
2163 tg->io_disp[rw], tg_iops_limit(tg, rw),
2164 sq->nr_queued[READ], sq->nr_queued[WRITE]);
2166 tg->last_low_overflow_time[rw] = jiffies;
2168 td->nr_queued[rw]++;
2169 throtl_add_bio_tg(bio, qn, tg);
2173 * Update @tg's dispatch time and force schedule dispatch if @tg
2174 * was empty before @bio. The forced scheduling isn't likely to
2175 * cause undue delay as @bio is likely to be dispatched directly if
2176 * its @tg's disptime is not in the future.
2178 if (tg->flags & THROTL_TG_WAS_EMPTY) {
2179 tg_update_disptime(tg);
2180 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2184 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2185 if (throttled || !td->track_bio_latency)
2186 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2188 spin_unlock_irq(&q->queue_lock);
2194 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2195 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2196 enum req_op op, unsigned long time)
2198 const bool rw = op_is_write(op);
2199 struct latency_bucket *latency;
2202 if (!td || td->limit_index != LIMIT_LOW ||
2203 !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2204 !blk_queue_nonrot(td->queue))
2207 index = request_bucket_index(size);
2209 latency = get_cpu_ptr(td->latency_buckets[rw]);
2210 latency[index].total_latency += time;
2211 latency[index].samples++;
2212 put_cpu_ptr(td->latency_buckets[rw]);
2215 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2217 struct request_queue *q = rq->q;
2218 struct throtl_data *td = q->td;
2220 throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2224 void blk_throtl_bio_endio(struct bio *bio)
2226 struct blkcg_gq *blkg;
2227 struct throtl_grp *tg;
2229 unsigned long finish_time;
2230 unsigned long start_time;
2232 int rw = bio_data_dir(bio);
2234 blkg = bio->bi_blkg;
2237 tg = blkg_to_tg(blkg);
2238 if (!tg->td->limit_valid[LIMIT_LOW])
2241 finish_time_ns = ktime_get_ns();
2242 tg->last_finish_time = finish_time_ns >> 10;
2244 start_time = bio_issue_time(&bio->bi_issue) >> 10;
2245 finish_time = __bio_issue_time(finish_time_ns) >> 10;
2246 if (!start_time || finish_time <= start_time)
2249 lat = finish_time - start_time;
2250 /* this is only for bio based driver */
2251 if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2252 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2255 if (tg->latency_target && lat >= tg->td->filtered_latency) {
2257 unsigned int threshold;
2259 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2260 threshold = tg->td->avg_buckets[rw][bucket].latency +
2262 if (lat > threshold)
2265 * Not race free, could get wrong count, which means cgroups
2271 if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2272 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2274 tg->bad_bio_cnt /= 2;
2279 int blk_throtl_init(struct request_queue *q)
2281 struct throtl_data *td;
2284 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2287 td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2288 LATENCY_BUCKET_SIZE, __alignof__(u64));
2289 if (!td->latency_buckets[READ]) {
2293 td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2294 LATENCY_BUCKET_SIZE, __alignof__(u64));
2295 if (!td->latency_buckets[WRITE]) {
2296 free_percpu(td->latency_buckets[READ]);
2301 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2302 throtl_service_queue_init(&td->service_queue);
2307 td->limit_valid[LIMIT_MAX] = true;
2308 td->limit_index = LIMIT_MAX;
2309 td->low_upgrade_time = jiffies;
2310 td->low_downgrade_time = jiffies;
2312 /* activate policy */
2313 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2315 free_percpu(td->latency_buckets[READ]);
2316 free_percpu(td->latency_buckets[WRITE]);
2322 void blk_throtl_exit(struct request_queue *q)
2325 del_timer_sync(&q->td->service_queue.pending_timer);
2326 throtl_shutdown_wq(q);
2327 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2328 free_percpu(q->td->latency_buckets[READ]);
2329 free_percpu(q->td->latency_buckets[WRITE]);
2333 void blk_throtl_register_queue(struct request_queue *q)
2335 struct throtl_data *td;
2341 if (blk_queue_nonrot(q)) {
2342 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2343 td->filtered_latency = LATENCY_FILTERED_SSD;
2345 td->throtl_slice = DFL_THROTL_SLICE_HD;
2346 td->filtered_latency = LATENCY_FILTERED_HD;
2347 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2348 td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2349 td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2352 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2353 /* if no low limit, use previous default */
2354 td->throtl_slice = DFL_THROTL_SLICE_HD;
2357 td->track_bio_latency = !queue_is_mq(q);
2358 if (!td->track_bio_latency)
2359 blk_stat_enable_accounting(q);
2362 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2363 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2367 return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2370 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2371 const char *page, size_t count)
2378 if (kstrtoul(page, 10, &v))
2380 t = msecs_to_jiffies(v);
2381 if (t == 0 || t > MAX_THROTL_SLICE)
2383 q->td->throtl_slice = t;
2388 static int __init throtl_init(void)
2390 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2391 if (!kthrotld_workqueue)
2392 panic("Failed to create kthrotld\n");
2394 return blkcg_policy_register(&blkcg_policy_throtl);
2397 module_init(throtl_init);