2 * Interface for controlling IO bandwidth on a request queue
4 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
7 #include <linux/module.h>
8 #include <linux/slab.h>
9 #include <linux/blkdev.h>
10 #include <linux/bio.h>
11 #include <linux/blktrace_api.h>
12 #include <linux/blk-cgroup.h>
15 /* Max dispatch from a group in 1 round */
16 static int throtl_grp_quantum = 8;
18 /* Total max dispatch from all groups in one round */
19 static int throtl_quantum = 32;
21 /* Throttling is performed over a slice and after that slice is renewed */
22 #define DFL_THROTL_SLICE_HD (HZ / 10)
23 #define DFL_THROTL_SLICE_SSD (HZ / 50)
24 #define MAX_THROTL_SLICE (HZ)
25 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
26 #define MIN_THROTL_BPS (320 * 1024)
27 #define MIN_THROTL_IOPS (10)
28 #define DFL_LATENCY_TARGET (-1L)
29 #define DFL_IDLE_THRESHOLD (0)
30 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
31 #define LATENCY_FILTERED_SSD (0)
33 * For HD, very small latency comes from sequential IO. Such IO is helpless to
34 * help determine if its IO is impacted by others, hence we ignore the IO
36 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
38 #define SKIP_LATENCY (((u64)1) << BLK_STAT_RES_SHIFT)
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 pending_tree; /* RB tree of active tgs */
89 struct rb_node *first_pending; /* first node in the tree */
90 unsigned int nr_pending; /* # queued in the tree */
91 unsigned long first_pending_disptime; /* disptime of the first tg */
92 struct timer_list pending_timer; /* fires on first_pending_disptime */
96 THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */
97 THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */
100 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
109 /* must be the first member */
110 struct blkg_policy_data pd;
112 /* active throtl group service_queue member */
113 struct rb_node rb_node;
115 /* throtl_data this group belongs to */
116 struct throtl_data *td;
118 /* this group's service queue */
119 struct throtl_service_queue service_queue;
122 * qnode_on_self is used when bios are directly queued to this
123 * throtl_grp so that local bios compete fairly with bios
124 * dispatched from children. qnode_on_parent is used when bios are
125 * dispatched from this throtl_grp into its parent and will compete
126 * with the sibling qnode_on_parents and the parent's
129 struct throtl_qnode qnode_on_self[2];
130 struct throtl_qnode qnode_on_parent[2];
133 * Dispatch time in jiffies. This is the estimated time when group
134 * will unthrottle and is ready to dispatch more bio. It is used as
135 * key to sort active groups in service tree.
137 unsigned long disptime;
141 /* are there any throtl rules between this group and td? */
144 /* internally used bytes per second rate limits */
145 uint64_t bps[2][LIMIT_CNT];
146 /* user configured bps limits */
147 uint64_t bps_conf[2][LIMIT_CNT];
149 /* internally used IOPS limits */
150 unsigned int iops[2][LIMIT_CNT];
151 /* user configured IOPS limits */
152 unsigned int iops_conf[2][LIMIT_CNT];
154 /* Number of bytes disptached in current slice */
155 uint64_t bytes_disp[2];
156 /* Number of bio's dispatched in current slice */
157 unsigned int io_disp[2];
159 unsigned long last_low_overflow_time[2];
161 uint64_t last_bytes_disp[2];
162 unsigned int last_io_disp[2];
164 unsigned long last_check_time;
166 unsigned long latency_target; /* us */
167 unsigned long latency_target_conf; /* us */
168 /* When did we start a new slice */
169 unsigned long slice_start[2];
170 unsigned long slice_end[2];
172 unsigned long last_finish_time; /* ns / 1024 */
173 unsigned long checked_last_finish_time; /* ns / 1024 */
174 unsigned long avg_idletime; /* ns / 1024 */
175 unsigned long idletime_threshold; /* us */
176 unsigned long idletime_threshold_conf; /* us */
178 unsigned int bio_cnt; /* total bios */
179 unsigned int bad_bio_cnt; /* bios exceeding latency threshold */
180 unsigned long bio_cnt_reset_time;
183 /* We measure latency for request size from <= 4k to >= 1M */
184 #define LATENCY_BUCKET_SIZE 9
186 struct latency_bucket {
187 unsigned long total_latency; /* ns / 1024 */
191 struct avg_latency_bucket {
192 unsigned long latency; /* ns / 1024 */
198 /* service tree for active throtl groups */
199 struct throtl_service_queue service_queue;
201 struct request_queue *queue;
203 /* Total Number of queued bios on READ and WRITE lists */
204 unsigned int nr_queued[2];
206 unsigned int throtl_slice;
208 /* Work for dispatching throttled bios */
209 struct work_struct dispatch_work;
210 unsigned int limit_index;
211 bool limit_valid[LIMIT_CNT];
213 unsigned long low_upgrade_time;
214 unsigned long low_downgrade_time;
218 struct latency_bucket tmp_buckets[LATENCY_BUCKET_SIZE];
219 struct avg_latency_bucket avg_buckets[LATENCY_BUCKET_SIZE];
220 struct latency_bucket __percpu *latency_buckets;
221 unsigned long last_calculate_time;
222 unsigned long filtered_latency;
224 bool track_bio_latency;
227 static void throtl_pending_timer_fn(unsigned long arg);
229 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
231 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
234 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
236 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
239 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
241 return pd_to_blkg(&tg->pd);
245 * sq_to_tg - return the throl_grp the specified service queue belongs to
246 * @sq: the throtl_service_queue of interest
248 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
249 * embedded in throtl_data, %NULL is returned.
251 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
253 if (sq && sq->parent_sq)
254 return container_of(sq, struct throtl_grp, service_queue);
260 * sq_to_td - return throtl_data the specified service queue belongs to
261 * @sq: the throtl_service_queue of interest
263 * A service_queue can be embedded in either a throtl_grp or throtl_data.
264 * Determine the associated throtl_data accordingly and return it.
266 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
268 struct throtl_grp *tg = sq_to_tg(sq);
273 return container_of(sq, struct throtl_data, service_queue);
277 * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
278 * make the IO dispatch more smooth.
279 * Scale up: linearly scale up according to lapsed time since upgrade. For
280 * every throtl_slice, the limit scales up 1/2 .low limit till the
281 * limit hits .max limit
282 * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
284 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
286 /* arbitrary value to avoid too big scale */
287 if (td->scale < 4096 && time_after_eq(jiffies,
288 td->low_upgrade_time + td->scale * td->throtl_slice))
289 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
291 return low + (low >> 1) * td->scale;
294 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
296 struct blkcg_gq *blkg = tg_to_blkg(tg);
297 struct throtl_data *td;
300 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
304 ret = tg->bps[rw][td->limit_index];
305 if (ret == 0 && td->limit_index == LIMIT_LOW) {
306 /* intermediate node or iops isn't 0 */
307 if (!list_empty(&blkg->blkcg->css.children) ||
308 tg->iops[rw][td->limit_index])
311 return MIN_THROTL_BPS;
314 if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
315 tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
318 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
319 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
324 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
326 struct blkcg_gq *blkg = tg_to_blkg(tg);
327 struct throtl_data *td;
330 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
334 ret = tg->iops[rw][td->limit_index];
335 if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
336 /* intermediate node or bps isn't 0 */
337 if (!list_empty(&blkg->blkcg->css.children) ||
338 tg->bps[rw][td->limit_index])
341 return MIN_THROTL_IOPS;
344 if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
345 tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
348 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
349 if (adjusted > UINT_MAX)
351 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
356 #define request_bucket_index(sectors) \
357 clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
360 * throtl_log - log debug message via blktrace
361 * @sq: the service_queue being reported
362 * @fmt: printf format string
365 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
366 * throtl_grp; otherwise, just "throtl".
368 #define throtl_log(sq, fmt, args...) do { \
369 struct throtl_grp *__tg = sq_to_tg((sq)); \
370 struct throtl_data *__td = sq_to_td((sq)); \
373 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
378 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
379 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
381 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
385 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
387 INIT_LIST_HEAD(&qn->node);
388 bio_list_init(&qn->bios);
393 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
394 * @bio: bio being added
395 * @qn: qnode to add bio to
396 * @queued: the service_queue->queued[] list @qn belongs to
398 * Add @bio to @qn and put @qn on @queued if it's not already on.
399 * @qn->tg's reference count is bumped when @qn is activated. See the
400 * comment on top of throtl_qnode definition for details.
402 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
403 struct list_head *queued)
405 bio_list_add(&qn->bios, bio);
406 if (list_empty(&qn->node)) {
407 list_add_tail(&qn->node, queued);
408 blkg_get(tg_to_blkg(qn->tg));
413 * throtl_peek_queued - peek the first bio on a qnode list
414 * @queued: the qnode list to peek
416 static struct bio *throtl_peek_queued(struct list_head *queued)
418 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
421 if (list_empty(queued))
424 bio = bio_list_peek(&qn->bios);
430 * throtl_pop_queued - pop the first bio form a qnode list
431 * @queued: the qnode list to pop a bio from
432 * @tg_to_put: optional out argument for throtl_grp to put
434 * Pop the first bio from the qnode list @queued. After popping, the first
435 * qnode is removed from @queued if empty or moved to the end of @queued so
436 * that the popping order is round-robin.
438 * When the first qnode is removed, its associated throtl_grp should be put
439 * too. If @tg_to_put is NULL, this function automatically puts it;
440 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
441 * responsible for putting it.
443 static struct bio *throtl_pop_queued(struct list_head *queued,
444 struct throtl_grp **tg_to_put)
446 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
449 if (list_empty(queued))
452 bio = bio_list_pop(&qn->bios);
455 if (bio_list_empty(&qn->bios)) {
456 list_del_init(&qn->node);
460 blkg_put(tg_to_blkg(qn->tg));
462 list_move_tail(&qn->node, queued);
468 /* init a service_queue, assumes the caller zeroed it */
469 static void throtl_service_queue_init(struct throtl_service_queue *sq)
471 INIT_LIST_HEAD(&sq->queued[0]);
472 INIT_LIST_HEAD(&sq->queued[1]);
473 sq->pending_tree = RB_ROOT;
474 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
478 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
480 struct throtl_grp *tg;
483 tg = kzalloc_node(sizeof(*tg), gfp, node);
487 throtl_service_queue_init(&tg->service_queue);
489 for (rw = READ; rw <= WRITE; rw++) {
490 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
491 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
494 RB_CLEAR_NODE(&tg->rb_node);
495 tg->bps[READ][LIMIT_MAX] = U64_MAX;
496 tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
497 tg->iops[READ][LIMIT_MAX] = UINT_MAX;
498 tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
499 tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
500 tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
501 tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
502 tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
503 /* LIMIT_LOW will have default value 0 */
505 tg->latency_target = DFL_LATENCY_TARGET;
506 tg->latency_target_conf = DFL_LATENCY_TARGET;
507 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
508 tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
513 static void throtl_pd_init(struct blkg_policy_data *pd)
515 struct throtl_grp *tg = pd_to_tg(pd);
516 struct blkcg_gq *blkg = tg_to_blkg(tg);
517 struct throtl_data *td = blkg->q->td;
518 struct throtl_service_queue *sq = &tg->service_queue;
521 * If on the default hierarchy, we switch to properly hierarchical
522 * behavior where limits on a given throtl_grp are applied to the
523 * whole subtree rather than just the group itself. e.g. If 16M
524 * read_bps limit is set on the root group, the whole system can't
525 * exceed 16M for the device.
527 * If not on the default hierarchy, the broken flat hierarchy
528 * behavior is retained where all throtl_grps are treated as if
529 * they're all separate root groups right below throtl_data.
530 * Limits of a group don't interact with limits of other groups
531 * regardless of the position of the group in the hierarchy.
533 sq->parent_sq = &td->service_queue;
534 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
535 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
540 * Set has_rules[] if @tg or any of its parents have limits configured.
541 * This doesn't require walking up to the top of the hierarchy as the
542 * parent's has_rules[] is guaranteed to be correct.
544 static void tg_update_has_rules(struct throtl_grp *tg)
546 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
547 struct throtl_data *td = tg->td;
550 for (rw = READ; rw <= WRITE; rw++)
551 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
552 (td->limit_valid[td->limit_index] &&
553 (tg_bps_limit(tg, rw) != U64_MAX ||
554 tg_iops_limit(tg, rw) != UINT_MAX));
557 static void throtl_pd_online(struct blkg_policy_data *pd)
559 struct throtl_grp *tg = pd_to_tg(pd);
561 * We don't want new groups to escape the limits of its ancestors.
562 * Update has_rules[] after a new group is brought online.
564 tg_update_has_rules(tg);
567 static void blk_throtl_update_limit_valid(struct throtl_data *td)
569 struct cgroup_subsys_state *pos_css;
570 struct blkcg_gq *blkg;
571 bool low_valid = false;
574 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
575 struct throtl_grp *tg = blkg_to_tg(blkg);
577 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
578 tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
583 td->limit_valid[LIMIT_LOW] = low_valid;
586 static void throtl_upgrade_state(struct throtl_data *td);
587 static void throtl_pd_offline(struct blkg_policy_data *pd)
589 struct throtl_grp *tg = pd_to_tg(pd);
591 tg->bps[READ][LIMIT_LOW] = 0;
592 tg->bps[WRITE][LIMIT_LOW] = 0;
593 tg->iops[READ][LIMIT_LOW] = 0;
594 tg->iops[WRITE][LIMIT_LOW] = 0;
596 blk_throtl_update_limit_valid(tg->td);
598 if (!tg->td->limit_valid[tg->td->limit_index])
599 throtl_upgrade_state(tg->td);
602 static void throtl_pd_free(struct blkg_policy_data *pd)
604 struct throtl_grp *tg = pd_to_tg(pd);
606 del_timer_sync(&tg->service_queue.pending_timer);
610 static struct throtl_grp *
611 throtl_rb_first(struct throtl_service_queue *parent_sq)
613 /* Service tree is empty */
614 if (!parent_sq->nr_pending)
617 if (!parent_sq->first_pending)
618 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
620 if (parent_sq->first_pending)
621 return rb_entry_tg(parent_sq->first_pending);
626 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
632 static void throtl_rb_erase(struct rb_node *n,
633 struct throtl_service_queue *parent_sq)
635 if (parent_sq->first_pending == n)
636 parent_sq->first_pending = NULL;
637 rb_erase_init(n, &parent_sq->pending_tree);
638 --parent_sq->nr_pending;
641 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
643 struct throtl_grp *tg;
645 tg = throtl_rb_first(parent_sq);
649 parent_sq->first_pending_disptime = tg->disptime;
652 static void tg_service_queue_add(struct throtl_grp *tg)
654 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
655 struct rb_node **node = &parent_sq->pending_tree.rb_node;
656 struct rb_node *parent = NULL;
657 struct throtl_grp *__tg;
658 unsigned long key = tg->disptime;
661 while (*node != NULL) {
663 __tg = rb_entry_tg(parent);
665 if (time_before(key, __tg->disptime))
666 node = &parent->rb_left;
668 node = &parent->rb_right;
674 parent_sq->first_pending = &tg->rb_node;
676 rb_link_node(&tg->rb_node, parent, node);
677 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
680 static void __throtl_enqueue_tg(struct throtl_grp *tg)
682 tg_service_queue_add(tg);
683 tg->flags |= THROTL_TG_PENDING;
684 tg->service_queue.parent_sq->nr_pending++;
687 static void throtl_enqueue_tg(struct throtl_grp *tg)
689 if (!(tg->flags & THROTL_TG_PENDING))
690 __throtl_enqueue_tg(tg);
693 static void __throtl_dequeue_tg(struct throtl_grp *tg)
695 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
696 tg->flags &= ~THROTL_TG_PENDING;
699 static void throtl_dequeue_tg(struct throtl_grp *tg)
701 if (tg->flags & THROTL_TG_PENDING)
702 __throtl_dequeue_tg(tg);
705 /* Call with queue lock held */
706 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
707 unsigned long expires)
709 unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
712 * Since we are adjusting the throttle limit dynamically, the sleep
713 * time calculated according to previous limit might be invalid. It's
714 * possible the cgroup sleep time is very long and no other cgroups
715 * have IO running so notify the limit changes. Make sure the cgroup
716 * doesn't sleep too long to avoid the missed notification.
718 if (time_after(expires, max_expire))
719 expires = max_expire;
720 mod_timer(&sq->pending_timer, expires);
721 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
722 expires - jiffies, jiffies);
726 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
727 * @sq: the service_queue to schedule dispatch for
728 * @force: force scheduling
730 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
731 * dispatch time of the first pending child. Returns %true if either timer
732 * is armed or there's no pending child left. %false if the current
733 * dispatch window is still open and the caller should continue
736 * If @force is %true, the dispatch timer is always scheduled and this
737 * function is guaranteed to return %true. This is to be used when the
738 * caller can't dispatch itself and needs to invoke pending_timer
739 * unconditionally. Note that forced scheduling is likely to induce short
740 * delay before dispatch starts even if @sq->first_pending_disptime is not
741 * in the future and thus shouldn't be used in hot paths.
743 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
746 /* any pending children left? */
750 update_min_dispatch_time(sq);
752 /* is the next dispatch time in the future? */
753 if (force || time_after(sq->first_pending_disptime, jiffies)) {
754 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
758 /* tell the caller to continue dispatching */
762 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
763 bool rw, unsigned long start)
765 tg->bytes_disp[rw] = 0;
769 * Previous slice has expired. We must have trimmed it after last
770 * bio dispatch. That means since start of last slice, we never used
771 * that bandwidth. Do try to make use of that bandwidth while giving
774 if (time_after_eq(start, tg->slice_start[rw]))
775 tg->slice_start[rw] = start;
777 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
778 throtl_log(&tg->service_queue,
779 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
780 rw == READ ? 'R' : 'W', tg->slice_start[rw],
781 tg->slice_end[rw], jiffies);
784 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
786 tg->bytes_disp[rw] = 0;
788 tg->slice_start[rw] = jiffies;
789 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
790 throtl_log(&tg->service_queue,
791 "[%c] new slice start=%lu end=%lu jiffies=%lu",
792 rw == READ ? 'R' : 'W', tg->slice_start[rw],
793 tg->slice_end[rw], jiffies);
796 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
797 unsigned long jiffy_end)
799 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
802 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
803 unsigned long jiffy_end)
805 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
806 throtl_log(&tg->service_queue,
807 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
808 rw == READ ? 'R' : 'W', tg->slice_start[rw],
809 tg->slice_end[rw], jiffies);
812 /* Determine if previously allocated or extended slice is complete or not */
813 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
815 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
821 /* Trim the used slices and adjust slice start accordingly */
822 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
824 unsigned long nr_slices, time_elapsed, io_trim;
827 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
830 * If bps are unlimited (-1), then time slice don't get
831 * renewed. Don't try to trim the slice if slice is used. A new
832 * slice will start when appropriate.
834 if (throtl_slice_used(tg, rw))
838 * A bio has been dispatched. Also adjust slice_end. It might happen
839 * that initially cgroup limit was very low resulting in high
840 * slice_end, but later limit was bumped up and bio was dispached
841 * sooner, then we need to reduce slice_end. A high bogus slice_end
842 * is bad because it does not allow new slice to start.
845 throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
847 time_elapsed = jiffies - tg->slice_start[rw];
849 nr_slices = time_elapsed / tg->td->throtl_slice;
853 tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
857 io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
860 if (!bytes_trim && !io_trim)
863 if (tg->bytes_disp[rw] >= bytes_trim)
864 tg->bytes_disp[rw] -= bytes_trim;
866 tg->bytes_disp[rw] = 0;
868 if (tg->io_disp[rw] >= io_trim)
869 tg->io_disp[rw] -= io_trim;
873 tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
875 throtl_log(&tg->service_queue,
876 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
877 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
878 tg->slice_start[rw], tg->slice_end[rw], jiffies);
881 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
884 bool rw = bio_data_dir(bio);
885 unsigned int io_allowed;
886 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
889 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
891 /* Slice has just started. Consider one slice interval */
893 jiffy_elapsed_rnd = tg->td->throtl_slice;
895 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
898 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
899 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
900 * will allow dispatch after 1 second and after that slice should
904 tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd;
908 io_allowed = UINT_MAX;
912 if (tg->io_disp[rw] + 1 <= io_allowed) {
918 /* Calc approx time to dispatch */
919 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ) / tg_iops_limit(tg, rw) + 1;
921 if (jiffy_wait > jiffy_elapsed)
922 jiffy_wait = jiffy_wait - jiffy_elapsed;
931 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
934 bool rw = bio_data_dir(bio);
935 u64 bytes_allowed, extra_bytes, tmp;
936 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
938 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
940 /* Slice has just started. Consider one slice interval */
942 jiffy_elapsed_rnd = tg->td->throtl_slice;
944 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
946 tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd;
950 if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
956 /* Calc approx time to dispatch */
957 extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
958 jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw));
964 * This wait time is without taking into consideration the rounding
965 * up we did. Add that time also.
967 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
974 * Returns whether one can dispatch a bio or not. Also returns approx number
975 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
977 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
980 bool rw = bio_data_dir(bio);
981 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
984 * Currently whole state machine of group depends on first bio
985 * queued in the group bio list. So one should not be calling
986 * this function with a different bio if there are other bios
989 BUG_ON(tg->service_queue.nr_queued[rw] &&
990 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
992 /* If tg->bps = -1, then BW is unlimited */
993 if (tg_bps_limit(tg, rw) == U64_MAX &&
994 tg_iops_limit(tg, rw) == UINT_MAX) {
1001 * If previous slice expired, start a new one otherwise renew/extend
1002 * existing slice to make sure it is at least throtl_slice interval
1003 * long since now. New slice is started only for empty throttle group.
1004 * If there is queued bio, that means there should be an active
1005 * slice and it should be extended instead.
1007 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
1008 throtl_start_new_slice(tg, rw);
1010 if (time_before(tg->slice_end[rw],
1011 jiffies + tg->td->throtl_slice))
1012 throtl_extend_slice(tg, rw,
1013 jiffies + tg->td->throtl_slice);
1016 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
1017 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
1023 max_wait = max(bps_wait, iops_wait);
1028 if (time_before(tg->slice_end[rw], jiffies + max_wait))
1029 throtl_extend_slice(tg, rw, jiffies + max_wait);
1034 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1036 bool rw = bio_data_dir(bio);
1038 /* Charge the bio to the group */
1039 tg->bytes_disp[rw] += bio->bi_iter.bi_size;
1041 tg->last_bytes_disp[rw] += bio->bi_iter.bi_size;
1042 tg->last_io_disp[rw]++;
1045 * BIO_THROTTLED is used to prevent the same bio to be throttled
1046 * more than once as a throttled bio will go through blk-throtl the
1047 * second time when it eventually gets issued. Set it when a bio
1048 * is being charged to a tg.
1050 if (!bio_flagged(bio, BIO_THROTTLED))
1051 bio_set_flag(bio, BIO_THROTTLED);
1055 * throtl_add_bio_tg - add a bio to the specified throtl_grp
1058 * @tg: the target throtl_grp
1060 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
1061 * tg->qnode_on_self[] is used.
1063 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1064 struct throtl_grp *tg)
1066 struct throtl_service_queue *sq = &tg->service_queue;
1067 bool rw = bio_data_dir(bio);
1070 qn = &tg->qnode_on_self[rw];
1073 * If @tg doesn't currently have any bios queued in the same
1074 * direction, queueing @bio can change when @tg should be
1075 * dispatched. Mark that @tg was empty. This is automatically
1076 * cleaered on the next tg_update_disptime().
1078 if (!sq->nr_queued[rw])
1079 tg->flags |= THROTL_TG_WAS_EMPTY;
1081 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1083 sq->nr_queued[rw]++;
1084 throtl_enqueue_tg(tg);
1087 static void tg_update_disptime(struct throtl_grp *tg)
1089 struct throtl_service_queue *sq = &tg->service_queue;
1090 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1093 bio = throtl_peek_queued(&sq->queued[READ]);
1095 tg_may_dispatch(tg, bio, &read_wait);
1097 bio = throtl_peek_queued(&sq->queued[WRITE]);
1099 tg_may_dispatch(tg, bio, &write_wait);
1101 min_wait = min(read_wait, write_wait);
1102 disptime = jiffies + min_wait;
1104 /* Update dispatch time */
1105 throtl_dequeue_tg(tg);
1106 tg->disptime = disptime;
1107 throtl_enqueue_tg(tg);
1109 /* see throtl_add_bio_tg() */
1110 tg->flags &= ~THROTL_TG_WAS_EMPTY;
1113 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1114 struct throtl_grp *parent_tg, bool rw)
1116 if (throtl_slice_used(parent_tg, rw)) {
1117 throtl_start_new_slice_with_credit(parent_tg, rw,
1118 child_tg->slice_start[rw]);
1123 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1125 struct throtl_service_queue *sq = &tg->service_queue;
1126 struct throtl_service_queue *parent_sq = sq->parent_sq;
1127 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1128 struct throtl_grp *tg_to_put = NULL;
1132 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1133 * from @tg may put its reference and @parent_sq might end up
1134 * getting released prematurely. Remember the tg to put and put it
1135 * after @bio is transferred to @parent_sq.
1137 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1138 sq->nr_queued[rw]--;
1140 throtl_charge_bio(tg, bio);
1143 * If our parent is another tg, we just need to transfer @bio to
1144 * the parent using throtl_add_bio_tg(). If our parent is
1145 * @td->service_queue, @bio is ready to be issued. Put it on its
1146 * bio_lists[] and decrease total number queued. The caller is
1147 * responsible for issuing these bios.
1150 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1151 start_parent_slice_with_credit(tg, parent_tg, rw);
1153 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1154 &parent_sq->queued[rw]);
1155 BUG_ON(tg->td->nr_queued[rw] <= 0);
1156 tg->td->nr_queued[rw]--;
1159 throtl_trim_slice(tg, rw);
1162 blkg_put(tg_to_blkg(tg_to_put));
1165 static int throtl_dispatch_tg(struct throtl_grp *tg)
1167 struct throtl_service_queue *sq = &tg->service_queue;
1168 unsigned int nr_reads = 0, nr_writes = 0;
1169 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1170 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1173 /* Try to dispatch 75% READS and 25% WRITES */
1175 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1176 tg_may_dispatch(tg, bio, NULL)) {
1178 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1181 if (nr_reads >= max_nr_reads)
1185 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1186 tg_may_dispatch(tg, bio, NULL)) {
1188 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1191 if (nr_writes >= max_nr_writes)
1195 return nr_reads + nr_writes;
1198 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1200 unsigned int nr_disp = 0;
1203 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1204 struct throtl_service_queue *sq = &tg->service_queue;
1209 if (time_before(jiffies, tg->disptime))
1212 throtl_dequeue_tg(tg);
1214 nr_disp += throtl_dispatch_tg(tg);
1216 if (sq->nr_queued[0] || sq->nr_queued[1])
1217 tg_update_disptime(tg);
1219 if (nr_disp >= throtl_quantum)
1226 static bool throtl_can_upgrade(struct throtl_data *td,
1227 struct throtl_grp *this_tg);
1229 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1230 * @arg: the throtl_service_queue being serviced
1232 * This timer is armed when a child throtl_grp with active bio's become
1233 * pending and queued on the service_queue's pending_tree and expires when
1234 * the first child throtl_grp should be dispatched. This function
1235 * dispatches bio's from the children throtl_grps to the parent
1238 * If the parent's parent is another throtl_grp, dispatching is propagated
1239 * by either arming its pending_timer or repeating dispatch directly. If
1240 * the top-level service_tree is reached, throtl_data->dispatch_work is
1241 * kicked so that the ready bio's are issued.
1243 static void throtl_pending_timer_fn(unsigned long arg)
1245 struct throtl_service_queue *sq = (void *)arg;
1246 struct throtl_grp *tg = sq_to_tg(sq);
1247 struct throtl_data *td = sq_to_td(sq);
1248 struct request_queue *q = td->queue;
1249 struct throtl_service_queue *parent_sq;
1253 spin_lock_irq(q->queue_lock);
1254 if (throtl_can_upgrade(td, NULL))
1255 throtl_upgrade_state(td);
1258 parent_sq = sq->parent_sq;
1262 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1263 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1264 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1266 ret = throtl_select_dispatch(sq);
1268 throtl_log(sq, "bios disp=%u", ret);
1272 if (throtl_schedule_next_dispatch(sq, false))
1275 /* this dispatch windows is still open, relax and repeat */
1276 spin_unlock_irq(q->queue_lock);
1278 spin_lock_irq(q->queue_lock);
1285 /* @parent_sq is another throl_grp, propagate dispatch */
1286 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1287 tg_update_disptime(tg);
1288 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1289 /* window is already open, repeat dispatching */
1296 /* reached the top-level, queue issueing */
1297 queue_work(kthrotld_workqueue, &td->dispatch_work);
1300 spin_unlock_irq(q->queue_lock);
1304 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1305 * @work: work item being executed
1307 * This function is queued for execution when bio's reach the bio_lists[]
1308 * of throtl_data->service_queue. Those bio's are ready and issued by this
1311 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1313 struct throtl_data *td = container_of(work, struct throtl_data,
1315 struct throtl_service_queue *td_sq = &td->service_queue;
1316 struct request_queue *q = td->queue;
1317 struct bio_list bio_list_on_stack;
1319 struct blk_plug plug;
1322 bio_list_init(&bio_list_on_stack);
1324 spin_lock_irq(q->queue_lock);
1325 for (rw = READ; rw <= WRITE; rw++)
1326 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1327 bio_list_add(&bio_list_on_stack, bio);
1328 spin_unlock_irq(q->queue_lock);
1330 if (!bio_list_empty(&bio_list_on_stack)) {
1331 blk_start_plug(&plug);
1332 while((bio = bio_list_pop(&bio_list_on_stack)))
1333 generic_make_request(bio);
1334 blk_finish_plug(&plug);
1338 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1341 struct throtl_grp *tg = pd_to_tg(pd);
1342 u64 v = *(u64 *)((void *)tg + off);
1346 return __blkg_prfill_u64(sf, pd, v);
1349 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1352 struct throtl_grp *tg = pd_to_tg(pd);
1353 unsigned int v = *(unsigned int *)((void *)tg + off);
1357 return __blkg_prfill_u64(sf, pd, v);
1360 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1362 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1363 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1367 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1369 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1370 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1374 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1376 struct throtl_service_queue *sq = &tg->service_queue;
1377 struct cgroup_subsys_state *pos_css;
1378 struct blkcg_gq *blkg;
1380 throtl_log(&tg->service_queue,
1381 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1382 tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1383 tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1386 * Update has_rules[] flags for the updated tg's subtree. A tg is
1387 * considered to have rules if either the tg itself or any of its
1388 * ancestors has rules. This identifies groups without any
1389 * restrictions in the whole hierarchy and allows them to bypass
1392 blkg_for_each_descendant_pre(blkg, pos_css,
1393 global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1394 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1395 struct throtl_grp *parent_tg;
1397 tg_update_has_rules(this_tg);
1398 /* ignore root/second level */
1399 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1400 !blkg->parent->parent)
1402 parent_tg = blkg_to_tg(blkg->parent);
1404 * make sure all children has lower idle time threshold and
1405 * higher latency target
1407 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1408 parent_tg->idletime_threshold);
1409 this_tg->latency_target = max(this_tg->latency_target,
1410 parent_tg->latency_target);
1414 * We're already holding queue_lock and know @tg is valid. Let's
1415 * apply the new config directly.
1417 * Restart the slices for both READ and WRITES. It might happen
1418 * that a group's limit are dropped suddenly and we don't want to
1419 * account recently dispatched IO with new low rate.
1421 throtl_start_new_slice(tg, 0);
1422 throtl_start_new_slice(tg, 1);
1424 if (tg->flags & THROTL_TG_PENDING) {
1425 tg_update_disptime(tg);
1426 throtl_schedule_next_dispatch(sq->parent_sq, true);
1430 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1431 char *buf, size_t nbytes, loff_t off, bool is_u64)
1433 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1434 struct blkg_conf_ctx ctx;
1435 struct throtl_grp *tg;
1439 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1444 if (sscanf(ctx.body, "%llu", &v) != 1)
1449 tg = blkg_to_tg(ctx.blkg);
1452 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1454 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1456 tg_conf_updated(tg, false);
1459 blkg_conf_finish(&ctx);
1460 return ret ?: nbytes;
1463 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1464 char *buf, size_t nbytes, loff_t off)
1466 return tg_set_conf(of, buf, nbytes, off, true);
1469 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1470 char *buf, size_t nbytes, loff_t off)
1472 return tg_set_conf(of, buf, nbytes, off, false);
1475 static struct cftype throtl_legacy_files[] = {
1477 .name = "throttle.read_bps_device",
1478 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1479 .seq_show = tg_print_conf_u64,
1480 .write = tg_set_conf_u64,
1483 .name = "throttle.write_bps_device",
1484 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1485 .seq_show = tg_print_conf_u64,
1486 .write = tg_set_conf_u64,
1489 .name = "throttle.read_iops_device",
1490 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1491 .seq_show = tg_print_conf_uint,
1492 .write = tg_set_conf_uint,
1495 .name = "throttle.write_iops_device",
1496 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1497 .seq_show = tg_print_conf_uint,
1498 .write = tg_set_conf_uint,
1501 .name = "throttle.io_service_bytes",
1502 .private = (unsigned long)&blkcg_policy_throtl,
1503 .seq_show = blkg_print_stat_bytes,
1506 .name = "throttle.io_serviced",
1507 .private = (unsigned long)&blkcg_policy_throtl,
1508 .seq_show = blkg_print_stat_ios,
1513 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1516 struct throtl_grp *tg = pd_to_tg(pd);
1517 const char *dname = blkg_dev_name(pd->blkg);
1518 char bufs[4][21] = { "max", "max", "max", "max" };
1520 unsigned int iops_dft;
1521 char idle_time[26] = "";
1522 char latency_time[26] = "";
1527 if (off == LIMIT_LOW) {
1532 iops_dft = UINT_MAX;
1535 if (tg->bps_conf[READ][off] == bps_dft &&
1536 tg->bps_conf[WRITE][off] == bps_dft &&
1537 tg->iops_conf[READ][off] == iops_dft &&
1538 tg->iops_conf[WRITE][off] == iops_dft &&
1539 (off != LIMIT_LOW ||
1540 (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1541 tg->latency_target_conf == DFL_LATENCY_TARGET)))
1544 if (tg->bps_conf[READ][off] != U64_MAX)
1545 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1546 tg->bps_conf[READ][off]);
1547 if (tg->bps_conf[WRITE][off] != U64_MAX)
1548 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1549 tg->bps_conf[WRITE][off]);
1550 if (tg->iops_conf[READ][off] != UINT_MAX)
1551 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1552 tg->iops_conf[READ][off]);
1553 if (tg->iops_conf[WRITE][off] != UINT_MAX)
1554 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1555 tg->iops_conf[WRITE][off]);
1556 if (off == LIMIT_LOW) {
1557 if (tg->idletime_threshold_conf == ULONG_MAX)
1558 strcpy(idle_time, " idle=max");
1560 snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1561 tg->idletime_threshold_conf);
1563 if (tg->latency_target_conf == ULONG_MAX)
1564 strcpy(latency_time, " latency=max");
1566 snprintf(latency_time, sizeof(latency_time),
1567 " latency=%lu", tg->latency_target_conf);
1570 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1571 dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1576 static int tg_print_limit(struct seq_file *sf, void *v)
1578 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1579 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1583 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1584 char *buf, size_t nbytes, loff_t off)
1586 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1587 struct blkg_conf_ctx ctx;
1588 struct throtl_grp *tg;
1590 unsigned long idle_time;
1591 unsigned long latency_time;
1593 int index = of_cft(of)->private;
1595 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1599 tg = blkg_to_tg(ctx.blkg);
1601 v[0] = tg->bps_conf[READ][index];
1602 v[1] = tg->bps_conf[WRITE][index];
1603 v[2] = tg->iops_conf[READ][index];
1604 v[3] = tg->iops_conf[WRITE][index];
1606 idle_time = tg->idletime_threshold_conf;
1607 latency_time = tg->latency_target_conf;
1609 char tok[27]; /* wiops=18446744073709551616 */
1614 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1623 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1631 if (!strcmp(tok, "rbps"))
1633 else if (!strcmp(tok, "wbps"))
1635 else if (!strcmp(tok, "riops"))
1636 v[2] = min_t(u64, val, UINT_MAX);
1637 else if (!strcmp(tok, "wiops"))
1638 v[3] = min_t(u64, val, UINT_MAX);
1639 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1641 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1647 tg->bps_conf[READ][index] = v[0];
1648 tg->bps_conf[WRITE][index] = v[1];
1649 tg->iops_conf[READ][index] = v[2];
1650 tg->iops_conf[WRITE][index] = v[3];
1652 if (index == LIMIT_MAX) {
1653 tg->bps[READ][index] = v[0];
1654 tg->bps[WRITE][index] = v[1];
1655 tg->iops[READ][index] = v[2];
1656 tg->iops[WRITE][index] = v[3];
1658 tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1659 tg->bps_conf[READ][LIMIT_MAX]);
1660 tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1661 tg->bps_conf[WRITE][LIMIT_MAX]);
1662 tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1663 tg->iops_conf[READ][LIMIT_MAX]);
1664 tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1665 tg->iops_conf[WRITE][LIMIT_MAX]);
1666 tg->idletime_threshold_conf = idle_time;
1667 tg->latency_target_conf = latency_time;
1669 /* force user to configure all settings for low limit */
1670 if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1671 tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1672 tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1673 tg->latency_target_conf == DFL_LATENCY_TARGET) {
1674 tg->bps[READ][LIMIT_LOW] = 0;
1675 tg->bps[WRITE][LIMIT_LOW] = 0;
1676 tg->iops[READ][LIMIT_LOW] = 0;
1677 tg->iops[WRITE][LIMIT_LOW] = 0;
1678 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1679 tg->latency_target = DFL_LATENCY_TARGET;
1680 } else if (index == LIMIT_LOW) {
1681 tg->idletime_threshold = tg->idletime_threshold_conf;
1682 tg->latency_target = tg->latency_target_conf;
1685 blk_throtl_update_limit_valid(tg->td);
1686 if (tg->td->limit_valid[LIMIT_LOW]) {
1687 if (index == LIMIT_LOW)
1688 tg->td->limit_index = LIMIT_LOW;
1690 tg->td->limit_index = LIMIT_MAX;
1691 tg_conf_updated(tg, index == LIMIT_LOW &&
1692 tg->td->limit_valid[LIMIT_LOW]);
1695 blkg_conf_finish(&ctx);
1696 return ret ?: nbytes;
1699 static struct cftype throtl_files[] = {
1700 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1703 .flags = CFTYPE_NOT_ON_ROOT,
1704 .seq_show = tg_print_limit,
1705 .write = tg_set_limit,
1706 .private = LIMIT_LOW,
1711 .flags = CFTYPE_NOT_ON_ROOT,
1712 .seq_show = tg_print_limit,
1713 .write = tg_set_limit,
1714 .private = LIMIT_MAX,
1719 static void throtl_shutdown_wq(struct request_queue *q)
1721 struct throtl_data *td = q->td;
1723 cancel_work_sync(&td->dispatch_work);
1726 static struct blkcg_policy blkcg_policy_throtl = {
1727 .dfl_cftypes = throtl_files,
1728 .legacy_cftypes = throtl_legacy_files,
1730 .pd_alloc_fn = throtl_pd_alloc,
1731 .pd_init_fn = throtl_pd_init,
1732 .pd_online_fn = throtl_pd_online,
1733 .pd_offline_fn = throtl_pd_offline,
1734 .pd_free_fn = throtl_pd_free,
1737 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1739 unsigned long rtime = jiffies, wtime = jiffies;
1741 if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1742 rtime = tg->last_low_overflow_time[READ];
1743 if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1744 wtime = tg->last_low_overflow_time[WRITE];
1745 return min(rtime, wtime);
1748 /* tg should not be an intermediate node */
1749 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1751 struct throtl_service_queue *parent_sq;
1752 struct throtl_grp *parent = tg;
1753 unsigned long ret = __tg_last_low_overflow_time(tg);
1756 parent_sq = parent->service_queue.parent_sq;
1757 parent = sq_to_tg(parent_sq);
1762 * The parent doesn't have low limit, it always reaches low
1763 * limit. Its overflow time is useless for children
1765 if (!parent->bps[READ][LIMIT_LOW] &&
1766 !parent->iops[READ][LIMIT_LOW] &&
1767 !parent->bps[WRITE][LIMIT_LOW] &&
1768 !parent->iops[WRITE][LIMIT_LOW])
1770 if (time_after(__tg_last_low_overflow_time(parent), ret))
1771 ret = __tg_last_low_overflow_time(parent);
1776 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1779 * cgroup is idle if:
1780 * - single idle is too long, longer than a fixed value (in case user
1781 * configure a too big threshold) or 4 times of idletime threshold
1782 * - average think time is more than threshold
1783 * - IO latency is largely below threshold
1788 time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1789 ret = tg->latency_target == DFL_LATENCY_TARGET ||
1790 tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1791 (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1792 tg->avg_idletime > tg->idletime_threshold ||
1793 (tg->latency_target && tg->bio_cnt &&
1794 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1795 throtl_log(&tg->service_queue,
1796 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1797 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1798 tg->bio_cnt, ret, tg->td->scale);
1802 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1804 struct throtl_service_queue *sq = &tg->service_queue;
1805 bool read_limit, write_limit;
1808 * if cgroup reaches low limit (if low limit is 0, the cgroup always
1809 * reaches), it's ok to upgrade to next limit
1811 read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1812 write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1813 if (!read_limit && !write_limit)
1815 if (read_limit && sq->nr_queued[READ] &&
1816 (!write_limit || sq->nr_queued[WRITE]))
1818 if (write_limit && sq->nr_queued[WRITE] &&
1819 (!read_limit || sq->nr_queued[READ]))
1822 if (time_after_eq(jiffies,
1823 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1824 throtl_tg_is_idle(tg))
1829 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1832 if (throtl_tg_can_upgrade(tg))
1834 tg = sq_to_tg(tg->service_queue.parent_sq);
1835 if (!tg || !tg_to_blkg(tg)->parent)
1841 static bool throtl_can_upgrade(struct throtl_data *td,
1842 struct throtl_grp *this_tg)
1844 struct cgroup_subsys_state *pos_css;
1845 struct blkcg_gq *blkg;
1847 if (td->limit_index != LIMIT_LOW)
1850 if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1854 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1855 struct throtl_grp *tg = blkg_to_tg(blkg);
1859 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1861 if (!throtl_hierarchy_can_upgrade(tg)) {
1870 static void throtl_upgrade_check(struct throtl_grp *tg)
1872 unsigned long now = jiffies;
1874 if (tg->td->limit_index != LIMIT_LOW)
1877 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1880 tg->last_check_time = now;
1882 if (!time_after_eq(now,
1883 __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1886 if (throtl_can_upgrade(tg->td, NULL))
1887 throtl_upgrade_state(tg->td);
1890 static void throtl_upgrade_state(struct throtl_data *td)
1892 struct cgroup_subsys_state *pos_css;
1893 struct blkcg_gq *blkg;
1895 throtl_log(&td->service_queue, "upgrade to max");
1896 td->limit_index = LIMIT_MAX;
1897 td->low_upgrade_time = jiffies;
1900 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1901 struct throtl_grp *tg = blkg_to_tg(blkg);
1902 struct throtl_service_queue *sq = &tg->service_queue;
1904 tg->disptime = jiffies - 1;
1905 throtl_select_dispatch(sq);
1906 throtl_schedule_next_dispatch(sq, false);
1909 throtl_select_dispatch(&td->service_queue);
1910 throtl_schedule_next_dispatch(&td->service_queue, false);
1911 queue_work(kthrotld_workqueue, &td->dispatch_work);
1914 static void throtl_downgrade_state(struct throtl_data *td, int new)
1918 throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1920 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1924 td->limit_index = new;
1925 td->low_downgrade_time = jiffies;
1928 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1930 struct throtl_data *td = tg->td;
1931 unsigned long now = jiffies;
1934 * If cgroup is below low limit, consider downgrade and throttle other
1937 if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1938 time_after_eq(now, tg_last_low_overflow_time(tg) +
1939 td->throtl_slice) &&
1940 (!throtl_tg_is_idle(tg) ||
1941 !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1946 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1949 if (!throtl_tg_can_downgrade(tg))
1951 tg = sq_to_tg(tg->service_queue.parent_sq);
1952 if (!tg || !tg_to_blkg(tg)->parent)
1958 static void throtl_downgrade_check(struct throtl_grp *tg)
1962 unsigned long elapsed_time;
1963 unsigned long now = jiffies;
1965 if (tg->td->limit_index != LIMIT_MAX ||
1966 !tg->td->limit_valid[LIMIT_LOW])
1968 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1970 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1973 elapsed_time = now - tg->last_check_time;
1974 tg->last_check_time = now;
1976 if (time_before(now, tg_last_low_overflow_time(tg) +
1977 tg->td->throtl_slice))
1980 if (tg->bps[READ][LIMIT_LOW]) {
1981 bps = tg->last_bytes_disp[READ] * HZ;
1982 do_div(bps, elapsed_time);
1983 if (bps >= tg->bps[READ][LIMIT_LOW])
1984 tg->last_low_overflow_time[READ] = now;
1987 if (tg->bps[WRITE][LIMIT_LOW]) {
1988 bps = tg->last_bytes_disp[WRITE] * HZ;
1989 do_div(bps, elapsed_time);
1990 if (bps >= tg->bps[WRITE][LIMIT_LOW])
1991 tg->last_low_overflow_time[WRITE] = now;
1994 if (tg->iops[READ][LIMIT_LOW]) {
1995 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
1996 if (iops >= tg->iops[READ][LIMIT_LOW])
1997 tg->last_low_overflow_time[READ] = now;
2000 if (tg->iops[WRITE][LIMIT_LOW]) {
2001 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2002 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2003 tg->last_low_overflow_time[WRITE] = now;
2007 * If cgroup is below low limit, consider downgrade and throttle other
2010 if (throtl_hierarchy_can_downgrade(tg))
2011 throtl_downgrade_state(tg->td, LIMIT_LOW);
2013 tg->last_bytes_disp[READ] = 0;
2014 tg->last_bytes_disp[WRITE] = 0;
2015 tg->last_io_disp[READ] = 0;
2016 tg->last_io_disp[WRITE] = 0;
2019 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2021 unsigned long now = ktime_get_ns() >> 10;
2022 unsigned long last_finish_time = tg->last_finish_time;
2024 if (now <= last_finish_time || last_finish_time == 0 ||
2025 last_finish_time == tg->checked_last_finish_time)
2028 tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2029 tg->checked_last_finish_time = last_finish_time;
2032 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2033 static void throtl_update_latency_buckets(struct throtl_data *td)
2035 struct avg_latency_bucket avg_latency[LATENCY_BUCKET_SIZE];
2037 unsigned long last_latency = 0;
2038 unsigned long latency;
2040 if (!blk_queue_nonrot(td->queue))
2042 if (time_before(jiffies, td->last_calculate_time + HZ))
2044 td->last_calculate_time = jiffies;
2046 memset(avg_latency, 0, sizeof(avg_latency));
2047 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2048 struct latency_bucket *tmp = &td->tmp_buckets[i];
2050 for_each_possible_cpu(cpu) {
2051 struct latency_bucket *bucket;
2053 /* this isn't race free, but ok in practice */
2054 bucket = per_cpu_ptr(td->latency_buckets, cpu);
2055 tmp->total_latency += bucket[i].total_latency;
2056 tmp->samples += bucket[i].samples;
2057 bucket[i].total_latency = 0;
2058 bucket[i].samples = 0;
2061 if (tmp->samples >= 32) {
2062 int samples = tmp->samples;
2064 latency = tmp->total_latency;
2066 tmp->total_latency = 0;
2071 avg_latency[i].latency = latency;
2075 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2076 if (!avg_latency[i].latency) {
2077 if (td->avg_buckets[i].latency < last_latency)
2078 td->avg_buckets[i].latency = last_latency;
2082 if (!td->avg_buckets[i].valid)
2083 latency = avg_latency[i].latency;
2085 latency = (td->avg_buckets[i].latency * 7 +
2086 avg_latency[i].latency) >> 3;
2088 td->avg_buckets[i].latency = max(latency, last_latency);
2089 td->avg_buckets[i].valid = true;
2090 last_latency = td->avg_buckets[i].latency;
2093 for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2094 throtl_log(&td->service_queue,
2095 "Latency bucket %d: latency=%ld, valid=%d", i,
2096 td->avg_buckets[i].latency, td->avg_buckets[i].valid);
2099 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2104 static void blk_throtl_assoc_bio(struct throtl_grp *tg, struct bio *bio)
2106 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2109 ret = bio_associate_current(bio);
2110 if (ret == 0 || ret == -EBUSY)
2111 bio->bi_cg_private = tg;
2112 blk_stat_set_issue(&bio->bi_issue_stat, bio_sectors(bio));
2114 bio_associate_current(bio);
2118 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
2121 struct throtl_qnode *qn = NULL;
2122 struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
2123 struct throtl_service_queue *sq;
2124 bool rw = bio_data_dir(bio);
2125 bool throttled = false;
2126 struct throtl_data *td = tg->td;
2128 WARN_ON_ONCE(!rcu_read_lock_held());
2130 /* see throtl_charge_bio() */
2131 if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw])
2134 spin_lock_irq(q->queue_lock);
2136 throtl_update_latency_buckets(td);
2138 if (unlikely(blk_queue_bypass(q)))
2141 blk_throtl_assoc_bio(tg, bio);
2142 blk_throtl_update_idletime(tg);
2144 sq = &tg->service_queue;
2148 if (tg->last_low_overflow_time[rw] == 0)
2149 tg->last_low_overflow_time[rw] = jiffies;
2150 throtl_downgrade_check(tg);
2151 throtl_upgrade_check(tg);
2152 /* throtl is FIFO - if bios are already queued, should queue */
2153 if (sq->nr_queued[rw])
2156 /* if above limits, break to queue */
2157 if (!tg_may_dispatch(tg, bio, NULL)) {
2158 tg->last_low_overflow_time[rw] = jiffies;
2159 if (throtl_can_upgrade(td, tg)) {
2160 throtl_upgrade_state(td);
2166 /* within limits, let's charge and dispatch directly */
2167 throtl_charge_bio(tg, bio);
2170 * We need to trim slice even when bios are not being queued
2171 * otherwise it might happen that a bio is not queued for
2172 * a long time and slice keeps on extending and trim is not
2173 * called for a long time. Now if limits are reduced suddenly
2174 * we take into account all the IO dispatched so far at new
2175 * low rate and * newly queued IO gets a really long dispatch
2178 * So keep on trimming slice even if bio is not queued.
2180 throtl_trim_slice(tg, rw);
2183 * @bio passed through this layer without being throttled.
2184 * Climb up the ladder. If we''re already at the top, it
2185 * can be executed directly.
2187 qn = &tg->qnode_on_parent[rw];
2194 /* out-of-limit, queue to @tg */
2195 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2196 rw == READ ? 'R' : 'W',
2197 tg->bytes_disp[rw], bio->bi_iter.bi_size,
2198 tg_bps_limit(tg, rw),
2199 tg->io_disp[rw], tg_iops_limit(tg, rw),
2200 sq->nr_queued[READ], sq->nr_queued[WRITE]);
2202 tg->last_low_overflow_time[rw] = jiffies;
2204 td->nr_queued[rw]++;
2205 throtl_add_bio_tg(bio, qn, tg);
2209 * Update @tg's dispatch time and force schedule dispatch if @tg
2210 * was empty before @bio. The forced scheduling isn't likely to
2211 * cause undue delay as @bio is likely to be dispatched directly if
2212 * its @tg's disptime is not in the future.
2214 if (tg->flags & THROTL_TG_WAS_EMPTY) {
2215 tg_update_disptime(tg);
2216 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2220 spin_unlock_irq(q->queue_lock);
2223 * As multiple blk-throtls may stack in the same issue path, we
2224 * don't want bios to leave with the flag set. Clear the flag if
2228 bio_clear_flag(bio, BIO_THROTTLED);
2230 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2231 if (throttled || !td->track_bio_latency)
2232 bio->bi_issue_stat.stat |= SKIP_LATENCY;
2237 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2238 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2239 int op, unsigned long time)
2241 struct latency_bucket *latency;
2244 if (!td || td->limit_index != LIMIT_LOW || op != REQ_OP_READ ||
2245 !blk_queue_nonrot(td->queue))
2248 index = request_bucket_index(size);
2250 latency = get_cpu_ptr(td->latency_buckets);
2251 latency[index].total_latency += time;
2252 latency[index].samples++;
2253 put_cpu_ptr(td->latency_buckets);
2256 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2258 struct request_queue *q = rq->q;
2259 struct throtl_data *td = q->td;
2261 throtl_track_latency(td, blk_stat_size(&rq->issue_stat),
2262 req_op(rq), time_ns >> 10);
2265 void blk_throtl_bio_endio(struct bio *bio)
2267 struct throtl_grp *tg;
2269 unsigned long finish_time;
2270 unsigned long start_time;
2273 tg = bio->bi_cg_private;
2276 bio->bi_cg_private = NULL;
2278 finish_time_ns = ktime_get_ns();
2279 tg->last_finish_time = finish_time_ns >> 10;
2281 start_time = blk_stat_time(&bio->bi_issue_stat) >> 10;
2282 finish_time = __blk_stat_time(finish_time_ns) >> 10;
2283 if (!start_time || finish_time <= start_time)
2286 lat = finish_time - start_time;
2287 /* this is only for bio based driver */
2288 if (!(bio->bi_issue_stat.stat & SKIP_LATENCY))
2289 throtl_track_latency(tg->td, blk_stat_size(&bio->bi_issue_stat),
2292 if (tg->latency_target && lat >= tg->td->filtered_latency) {
2294 unsigned int threshold;
2296 bucket = request_bucket_index(
2297 blk_stat_size(&bio->bi_issue_stat));
2298 threshold = tg->td->avg_buckets[bucket].latency +
2300 if (lat > threshold)
2303 * Not race free, could get wrong count, which means cgroups
2309 if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2310 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2312 tg->bad_bio_cnt /= 2;
2318 * Dispatch all bios from all children tg's queued on @parent_sq. On
2319 * return, @parent_sq is guaranteed to not have any active children tg's
2320 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
2322 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
2324 struct throtl_grp *tg;
2326 while ((tg = throtl_rb_first(parent_sq))) {
2327 struct throtl_service_queue *sq = &tg->service_queue;
2330 throtl_dequeue_tg(tg);
2332 while ((bio = throtl_peek_queued(&sq->queued[READ])))
2333 tg_dispatch_one_bio(tg, bio_data_dir(bio));
2334 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
2335 tg_dispatch_one_bio(tg, bio_data_dir(bio));
2340 * blk_throtl_drain - drain throttled bios
2341 * @q: request_queue to drain throttled bios for
2343 * Dispatch all currently throttled bios on @q through ->make_request_fn().
2345 void blk_throtl_drain(struct request_queue *q)
2346 __releases(q->queue_lock) __acquires(q->queue_lock)
2348 struct throtl_data *td = q->td;
2349 struct blkcg_gq *blkg;
2350 struct cgroup_subsys_state *pos_css;
2354 queue_lockdep_assert_held(q);
2358 * Drain each tg while doing post-order walk on the blkg tree, so
2359 * that all bios are propagated to td->service_queue. It'd be
2360 * better to walk service_queue tree directly but blkg walk is
2363 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
2364 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
2366 /* finally, transfer bios from top-level tg's into the td */
2367 tg_drain_bios(&td->service_queue);
2370 spin_unlock_irq(q->queue_lock);
2372 /* all bios now should be in td->service_queue, issue them */
2373 for (rw = READ; rw <= WRITE; rw++)
2374 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
2376 generic_make_request(bio);
2378 spin_lock_irq(q->queue_lock);
2381 int blk_throtl_init(struct request_queue *q)
2383 struct throtl_data *td;
2386 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2389 td->latency_buckets = __alloc_percpu(sizeof(struct latency_bucket) *
2390 LATENCY_BUCKET_SIZE, __alignof__(u64));
2391 if (!td->latency_buckets) {
2396 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2397 throtl_service_queue_init(&td->service_queue);
2402 td->limit_valid[LIMIT_MAX] = true;
2403 td->limit_index = LIMIT_MAX;
2404 td->low_upgrade_time = jiffies;
2405 td->low_downgrade_time = jiffies;
2407 /* activate policy */
2408 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2410 free_percpu(td->latency_buckets);
2416 void blk_throtl_exit(struct request_queue *q)
2419 throtl_shutdown_wq(q);
2420 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2421 free_percpu(q->td->latency_buckets);
2425 void blk_throtl_register_queue(struct request_queue *q)
2427 struct throtl_data *td;
2433 if (blk_queue_nonrot(q)) {
2434 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2435 td->filtered_latency = LATENCY_FILTERED_SSD;
2437 td->throtl_slice = DFL_THROTL_SLICE_HD;
2438 td->filtered_latency = LATENCY_FILTERED_HD;
2439 for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2440 td->avg_buckets[i].latency = DFL_HD_BASELINE_LATENCY;
2442 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2443 /* if no low limit, use previous default */
2444 td->throtl_slice = DFL_THROTL_SLICE_HD;
2447 td->track_bio_latency = !q->mq_ops && !q->request_fn;
2448 if (!td->track_bio_latency)
2449 blk_stat_enable_accounting(q);
2452 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2453 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2457 return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2460 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2461 const char *page, size_t count)
2468 if (kstrtoul(page, 10, &v))
2470 t = msecs_to_jiffies(v);
2471 if (t == 0 || t > MAX_THROTL_SLICE)
2473 q->td->throtl_slice = t;
2478 static int __init throtl_init(void)
2480 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2481 if (!kthrotld_workqueue)
2482 panic("Failed to create kthrotld\n");
2484 return blkcg_policy_register(&blkcg_policy_throtl);
2487 module_init(throtl_init);