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 DFL_IDLE_THRESHOLD_SSD (1000L) /* 1 ms */
26 #define DFL_IDLE_THRESHOLD_HD (100L * 1000) /* 100 ms */
27 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
28 /* default latency target is 0, eg, guarantee IO latency by default */
29 #define DFL_LATENCY_TARGET (0)
31 #define SKIP_LATENCY (((u64)1) << BLK_STAT_RES_SHIFT)
33 static struct blkcg_policy blkcg_policy_throtl;
35 /* A workqueue to queue throttle related work */
36 static struct workqueue_struct *kthrotld_workqueue;
39 * To implement hierarchical throttling, throtl_grps form a tree and bios
40 * are dispatched upwards level by level until they reach the top and get
41 * issued. When dispatching bios from the children and local group at each
42 * level, if the bios are dispatched into a single bio_list, there's a risk
43 * of a local or child group which can queue many bios at once filling up
44 * the list starving others.
46 * To avoid such starvation, dispatched bios are queued separately
47 * according to where they came from. When they are again dispatched to
48 * the parent, they're popped in round-robin order so that no single source
49 * hogs the dispatch window.
51 * throtl_qnode is used to keep the queued bios separated by their sources.
52 * Bios are queued to throtl_qnode which in turn is queued to
53 * throtl_service_queue and then dispatched in round-robin order.
55 * It's also used to track the reference counts on blkg's. A qnode always
56 * belongs to a throtl_grp and gets queued on itself or the parent, so
57 * incrementing the reference of the associated throtl_grp when a qnode is
58 * queued and decrementing when dequeued is enough to keep the whole blkg
59 * tree pinned while bios are in flight.
62 struct list_head node; /* service_queue->queued[] */
63 struct bio_list bios; /* queued bios */
64 struct throtl_grp *tg; /* tg this qnode belongs to */
67 struct throtl_service_queue {
68 struct throtl_service_queue *parent_sq; /* the parent service_queue */
71 * Bios queued directly to this service_queue or dispatched from
72 * children throtl_grp's.
74 struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */
75 unsigned int nr_queued[2]; /* number of queued bios */
78 * RB tree of active children throtl_grp's, which are sorted by
81 struct rb_root pending_tree; /* RB tree of active tgs */
82 struct rb_node *first_pending; /* first node in the tree */
83 unsigned int nr_pending; /* # queued in the tree */
84 unsigned long first_pending_disptime; /* disptime of the first tg */
85 struct timer_list pending_timer; /* fires on first_pending_disptime */
89 THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */
90 THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */
93 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
102 /* must be the first member */
103 struct blkg_policy_data pd;
105 /* active throtl group service_queue member */
106 struct rb_node rb_node;
108 /* throtl_data this group belongs to */
109 struct throtl_data *td;
111 /* this group's service queue */
112 struct throtl_service_queue service_queue;
115 * qnode_on_self is used when bios are directly queued to this
116 * throtl_grp so that local bios compete fairly with bios
117 * dispatched from children. qnode_on_parent is used when bios are
118 * dispatched from this throtl_grp into its parent and will compete
119 * with the sibling qnode_on_parents and the parent's
122 struct throtl_qnode qnode_on_self[2];
123 struct throtl_qnode qnode_on_parent[2];
126 * Dispatch time in jiffies. This is the estimated time when group
127 * will unthrottle and is ready to dispatch more bio. It is used as
128 * key to sort active groups in service tree.
130 unsigned long disptime;
134 /* are there any throtl rules between this group and td? */
137 /* internally used bytes per second rate limits */
138 uint64_t bps[2][LIMIT_CNT];
139 /* user configured bps limits */
140 uint64_t bps_conf[2][LIMIT_CNT];
142 /* internally used IOPS limits */
143 unsigned int iops[2][LIMIT_CNT];
144 /* user configured IOPS limits */
145 unsigned int iops_conf[2][LIMIT_CNT];
147 /* Number of bytes disptached in current slice */
148 uint64_t bytes_disp[2];
149 /* Number of bio's dispatched in current slice */
150 unsigned int io_disp[2];
152 unsigned long last_low_overflow_time[2];
154 uint64_t last_bytes_disp[2];
155 unsigned int last_io_disp[2];
157 unsigned long last_check_time;
159 unsigned long latency_target; /* us */
160 unsigned long latency_target_conf; /* us */
161 /* When did we start a new slice */
162 unsigned long slice_start[2];
163 unsigned long slice_end[2];
165 unsigned long last_finish_time; /* ns / 1024 */
166 unsigned long checked_last_finish_time; /* ns / 1024 */
167 unsigned long avg_idletime; /* ns / 1024 */
168 unsigned long idletime_threshold; /* us */
169 unsigned long idletime_threshold_conf; /* us */
171 unsigned int bio_cnt; /* total bios */
172 unsigned int bad_bio_cnt; /* bios exceeding latency threshold */
173 unsigned long bio_cnt_reset_time;
176 /* We measure latency for request size from <= 4k to >= 1M */
177 #define LATENCY_BUCKET_SIZE 9
179 struct latency_bucket {
180 unsigned long total_latency; /* ns / 1024 */
184 struct avg_latency_bucket {
185 unsigned long latency; /* ns / 1024 */
191 /* service tree for active throtl groups */
192 struct throtl_service_queue service_queue;
194 struct request_queue *queue;
196 /* Total Number of queued bios on READ and WRITE lists */
197 unsigned int nr_queued[2];
199 unsigned int throtl_slice;
201 /* Work for dispatching throttled bios */
202 struct work_struct dispatch_work;
203 unsigned int limit_index;
204 bool limit_valid[LIMIT_CNT];
206 unsigned long dft_idletime_threshold; /* us */
208 unsigned long low_upgrade_time;
209 unsigned long low_downgrade_time;
213 struct latency_bucket tmp_buckets[LATENCY_BUCKET_SIZE];
214 struct avg_latency_bucket avg_buckets[LATENCY_BUCKET_SIZE];
215 struct latency_bucket __percpu *latency_buckets;
216 unsigned long last_calculate_time;
218 bool track_bio_latency;
221 static void throtl_pending_timer_fn(unsigned long arg);
223 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
225 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
228 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
230 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
233 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
235 return pd_to_blkg(&tg->pd);
239 * sq_to_tg - return the throl_grp the specified service queue belongs to
240 * @sq: the throtl_service_queue of interest
242 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
243 * embedded in throtl_data, %NULL is returned.
245 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
247 if (sq && sq->parent_sq)
248 return container_of(sq, struct throtl_grp, service_queue);
254 * sq_to_td - return throtl_data the specified service queue belongs to
255 * @sq: the throtl_service_queue of interest
257 * A service_queue can be embedded in either a throtl_grp or throtl_data.
258 * Determine the associated throtl_data accordingly and return it.
260 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
262 struct throtl_grp *tg = sq_to_tg(sq);
267 return container_of(sq, struct throtl_data, service_queue);
271 * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
272 * make the IO dispatch more smooth.
273 * Scale up: linearly scale up according to lapsed time since upgrade. For
274 * every throtl_slice, the limit scales up 1/2 .low limit till the
275 * limit hits .max limit
276 * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
278 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
280 /* arbitrary value to avoid too big scale */
281 if (td->scale < 4096 && time_after_eq(jiffies,
282 td->low_upgrade_time + td->scale * td->throtl_slice))
283 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
285 return low + (low >> 1) * td->scale;
288 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
290 struct blkcg_gq *blkg = tg_to_blkg(tg);
291 struct throtl_data *td;
294 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
298 ret = tg->bps[rw][td->limit_index];
299 if (ret == 0 && td->limit_index == LIMIT_LOW)
300 return tg->bps[rw][LIMIT_MAX];
302 if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
303 tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
306 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
307 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
312 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
314 struct blkcg_gq *blkg = tg_to_blkg(tg);
315 struct throtl_data *td;
318 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
321 ret = tg->iops[rw][td->limit_index];
322 if (ret == 0 && tg->td->limit_index == LIMIT_LOW)
323 return tg->iops[rw][LIMIT_MAX];
325 if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
326 tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
329 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
330 if (adjusted > UINT_MAX)
332 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
337 #define request_bucket_index(sectors) \
338 clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
341 * throtl_log - log debug message via blktrace
342 * @sq: the service_queue being reported
343 * @fmt: printf format string
346 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
347 * throtl_grp; otherwise, just "throtl".
349 #define throtl_log(sq, fmt, args...) do { \
350 struct throtl_grp *__tg = sq_to_tg((sq)); \
351 struct throtl_data *__td = sq_to_td((sq)); \
354 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
359 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
360 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
362 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
366 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
368 INIT_LIST_HEAD(&qn->node);
369 bio_list_init(&qn->bios);
374 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
375 * @bio: bio being added
376 * @qn: qnode to add bio to
377 * @queued: the service_queue->queued[] list @qn belongs to
379 * Add @bio to @qn and put @qn on @queued if it's not already on.
380 * @qn->tg's reference count is bumped when @qn is activated. See the
381 * comment on top of throtl_qnode definition for details.
383 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
384 struct list_head *queued)
386 bio_list_add(&qn->bios, bio);
387 if (list_empty(&qn->node)) {
388 list_add_tail(&qn->node, queued);
389 blkg_get(tg_to_blkg(qn->tg));
394 * throtl_peek_queued - peek the first bio on a qnode list
395 * @queued: the qnode list to peek
397 static struct bio *throtl_peek_queued(struct list_head *queued)
399 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
402 if (list_empty(queued))
405 bio = bio_list_peek(&qn->bios);
411 * throtl_pop_queued - pop the first bio form a qnode list
412 * @queued: the qnode list to pop a bio from
413 * @tg_to_put: optional out argument for throtl_grp to put
415 * Pop the first bio from the qnode list @queued. After popping, the first
416 * qnode is removed from @queued if empty or moved to the end of @queued so
417 * that the popping order is round-robin.
419 * When the first qnode is removed, its associated throtl_grp should be put
420 * too. If @tg_to_put is NULL, this function automatically puts it;
421 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
422 * responsible for putting it.
424 static struct bio *throtl_pop_queued(struct list_head *queued,
425 struct throtl_grp **tg_to_put)
427 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
430 if (list_empty(queued))
433 bio = bio_list_pop(&qn->bios);
436 if (bio_list_empty(&qn->bios)) {
437 list_del_init(&qn->node);
441 blkg_put(tg_to_blkg(qn->tg));
443 list_move_tail(&qn->node, queued);
449 /* init a service_queue, assumes the caller zeroed it */
450 static void throtl_service_queue_init(struct throtl_service_queue *sq)
452 INIT_LIST_HEAD(&sq->queued[0]);
453 INIT_LIST_HEAD(&sq->queued[1]);
454 sq->pending_tree = RB_ROOT;
455 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
459 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
461 struct throtl_grp *tg;
464 tg = kzalloc_node(sizeof(*tg), gfp, node);
468 throtl_service_queue_init(&tg->service_queue);
470 for (rw = READ; rw <= WRITE; rw++) {
471 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
472 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
475 RB_CLEAR_NODE(&tg->rb_node);
476 tg->bps[READ][LIMIT_MAX] = U64_MAX;
477 tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
478 tg->iops[READ][LIMIT_MAX] = UINT_MAX;
479 tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
480 tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
481 tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
482 tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
483 tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
484 /* LIMIT_LOW will have default value 0 */
486 tg->latency_target = DFL_LATENCY_TARGET;
487 tg->latency_target_conf = DFL_LATENCY_TARGET;
492 static void throtl_pd_init(struct blkg_policy_data *pd)
494 struct throtl_grp *tg = pd_to_tg(pd);
495 struct blkcg_gq *blkg = tg_to_blkg(tg);
496 struct throtl_data *td = blkg->q->td;
497 struct throtl_service_queue *sq = &tg->service_queue;
500 * If on the default hierarchy, we switch to properly hierarchical
501 * behavior where limits on a given throtl_grp are applied to the
502 * whole subtree rather than just the group itself. e.g. If 16M
503 * read_bps limit is set on the root group, the whole system can't
504 * exceed 16M for the device.
506 * If not on the default hierarchy, the broken flat hierarchy
507 * behavior is retained where all throtl_grps are treated as if
508 * they're all separate root groups right below throtl_data.
509 * Limits of a group don't interact with limits of other groups
510 * regardless of the position of the group in the hierarchy.
512 sq->parent_sq = &td->service_queue;
513 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
514 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
517 tg->idletime_threshold = td->dft_idletime_threshold;
518 tg->idletime_threshold_conf = td->dft_idletime_threshold;
522 * Set has_rules[] if @tg or any of its parents have limits configured.
523 * This doesn't require walking up to the top of the hierarchy as the
524 * parent's has_rules[] is guaranteed to be correct.
526 static void tg_update_has_rules(struct throtl_grp *tg)
528 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
529 struct throtl_data *td = tg->td;
532 for (rw = READ; rw <= WRITE; rw++)
533 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
534 (td->limit_valid[td->limit_index] &&
535 (tg_bps_limit(tg, rw) != U64_MAX ||
536 tg_iops_limit(tg, rw) != UINT_MAX));
539 static void throtl_pd_online(struct blkg_policy_data *pd)
541 struct throtl_grp *tg = pd_to_tg(pd);
543 * We don't want new groups to escape the limits of its ancestors.
544 * Update has_rules[] after a new group is brought online.
546 tg_update_has_rules(tg);
549 static void blk_throtl_update_limit_valid(struct throtl_data *td)
551 struct cgroup_subsys_state *pos_css;
552 struct blkcg_gq *blkg;
553 bool low_valid = false;
556 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
557 struct throtl_grp *tg = blkg_to_tg(blkg);
559 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
560 tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
565 td->limit_valid[LIMIT_LOW] = low_valid;
568 static void throtl_upgrade_state(struct throtl_data *td);
569 static void throtl_pd_offline(struct blkg_policy_data *pd)
571 struct throtl_grp *tg = pd_to_tg(pd);
573 tg->bps[READ][LIMIT_LOW] = 0;
574 tg->bps[WRITE][LIMIT_LOW] = 0;
575 tg->iops[READ][LIMIT_LOW] = 0;
576 tg->iops[WRITE][LIMIT_LOW] = 0;
578 blk_throtl_update_limit_valid(tg->td);
580 if (!tg->td->limit_valid[tg->td->limit_index])
581 throtl_upgrade_state(tg->td);
584 static void throtl_pd_free(struct blkg_policy_data *pd)
586 struct throtl_grp *tg = pd_to_tg(pd);
588 del_timer_sync(&tg->service_queue.pending_timer);
592 static struct throtl_grp *
593 throtl_rb_first(struct throtl_service_queue *parent_sq)
595 /* Service tree is empty */
596 if (!parent_sq->nr_pending)
599 if (!parent_sq->first_pending)
600 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
602 if (parent_sq->first_pending)
603 return rb_entry_tg(parent_sq->first_pending);
608 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
614 static void throtl_rb_erase(struct rb_node *n,
615 struct throtl_service_queue *parent_sq)
617 if (parent_sq->first_pending == n)
618 parent_sq->first_pending = NULL;
619 rb_erase_init(n, &parent_sq->pending_tree);
620 --parent_sq->nr_pending;
623 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
625 struct throtl_grp *tg;
627 tg = throtl_rb_first(parent_sq);
631 parent_sq->first_pending_disptime = tg->disptime;
634 static void tg_service_queue_add(struct throtl_grp *tg)
636 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
637 struct rb_node **node = &parent_sq->pending_tree.rb_node;
638 struct rb_node *parent = NULL;
639 struct throtl_grp *__tg;
640 unsigned long key = tg->disptime;
643 while (*node != NULL) {
645 __tg = rb_entry_tg(parent);
647 if (time_before(key, __tg->disptime))
648 node = &parent->rb_left;
650 node = &parent->rb_right;
656 parent_sq->first_pending = &tg->rb_node;
658 rb_link_node(&tg->rb_node, parent, node);
659 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
662 static void __throtl_enqueue_tg(struct throtl_grp *tg)
664 tg_service_queue_add(tg);
665 tg->flags |= THROTL_TG_PENDING;
666 tg->service_queue.parent_sq->nr_pending++;
669 static void throtl_enqueue_tg(struct throtl_grp *tg)
671 if (!(tg->flags & THROTL_TG_PENDING))
672 __throtl_enqueue_tg(tg);
675 static void __throtl_dequeue_tg(struct throtl_grp *tg)
677 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
678 tg->flags &= ~THROTL_TG_PENDING;
681 static void throtl_dequeue_tg(struct throtl_grp *tg)
683 if (tg->flags & THROTL_TG_PENDING)
684 __throtl_dequeue_tg(tg);
687 /* Call with queue lock held */
688 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
689 unsigned long expires)
691 unsigned long max_expire = jiffies + 8 * sq_to_tg(sq)->td->throtl_slice;
694 * Since we are adjusting the throttle limit dynamically, the sleep
695 * time calculated according to previous limit might be invalid. It's
696 * possible the cgroup sleep time is very long and no other cgroups
697 * have IO running so notify the limit changes. Make sure the cgroup
698 * doesn't sleep too long to avoid the missed notification.
700 if (time_after(expires, max_expire))
701 expires = max_expire;
702 mod_timer(&sq->pending_timer, expires);
703 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
704 expires - jiffies, jiffies);
708 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
709 * @sq: the service_queue to schedule dispatch for
710 * @force: force scheduling
712 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
713 * dispatch time of the first pending child. Returns %true if either timer
714 * is armed or there's no pending child left. %false if the current
715 * dispatch window is still open and the caller should continue
718 * If @force is %true, the dispatch timer is always scheduled and this
719 * function is guaranteed to return %true. This is to be used when the
720 * caller can't dispatch itself and needs to invoke pending_timer
721 * unconditionally. Note that forced scheduling is likely to induce short
722 * delay before dispatch starts even if @sq->first_pending_disptime is not
723 * in the future and thus shouldn't be used in hot paths.
725 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
728 /* any pending children left? */
732 update_min_dispatch_time(sq);
734 /* is the next dispatch time in the future? */
735 if (force || time_after(sq->first_pending_disptime, jiffies)) {
736 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
740 /* tell the caller to continue dispatching */
744 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
745 bool rw, unsigned long start)
747 tg->bytes_disp[rw] = 0;
751 * Previous slice has expired. We must have trimmed it after last
752 * bio dispatch. That means since start of last slice, we never used
753 * that bandwidth. Do try to make use of that bandwidth while giving
756 if (time_after_eq(start, tg->slice_start[rw]))
757 tg->slice_start[rw] = start;
759 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
760 throtl_log(&tg->service_queue,
761 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
762 rw == READ ? 'R' : 'W', tg->slice_start[rw],
763 tg->slice_end[rw], jiffies);
766 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
768 tg->bytes_disp[rw] = 0;
770 tg->slice_start[rw] = jiffies;
771 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
772 throtl_log(&tg->service_queue,
773 "[%c] new slice start=%lu end=%lu jiffies=%lu",
774 rw == READ ? 'R' : 'W', tg->slice_start[rw],
775 tg->slice_end[rw], jiffies);
778 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
779 unsigned long jiffy_end)
781 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
784 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
785 unsigned long jiffy_end)
787 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
788 throtl_log(&tg->service_queue,
789 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
790 rw == READ ? 'R' : 'W', tg->slice_start[rw],
791 tg->slice_end[rw], jiffies);
794 /* Determine if previously allocated or extended slice is complete or not */
795 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
797 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
803 /* Trim the used slices and adjust slice start accordingly */
804 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
806 unsigned long nr_slices, time_elapsed, io_trim;
809 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
812 * If bps are unlimited (-1), then time slice don't get
813 * renewed. Don't try to trim the slice if slice is used. A new
814 * slice will start when appropriate.
816 if (throtl_slice_used(tg, rw))
820 * A bio has been dispatched. Also adjust slice_end. It might happen
821 * that initially cgroup limit was very low resulting in high
822 * slice_end, but later limit was bumped up and bio was dispached
823 * sooner, then we need to reduce slice_end. A high bogus slice_end
824 * is bad because it does not allow new slice to start.
827 throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
829 time_elapsed = jiffies - tg->slice_start[rw];
831 nr_slices = time_elapsed / tg->td->throtl_slice;
835 tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
839 io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
842 if (!bytes_trim && !io_trim)
845 if (tg->bytes_disp[rw] >= bytes_trim)
846 tg->bytes_disp[rw] -= bytes_trim;
848 tg->bytes_disp[rw] = 0;
850 if (tg->io_disp[rw] >= io_trim)
851 tg->io_disp[rw] -= io_trim;
855 tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
857 throtl_log(&tg->service_queue,
858 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
859 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
860 tg->slice_start[rw], tg->slice_end[rw], jiffies);
863 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
866 bool rw = bio_data_dir(bio);
867 unsigned int io_allowed;
868 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
871 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
873 /* Slice has just started. Consider one slice interval */
875 jiffy_elapsed_rnd = tg->td->throtl_slice;
877 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
880 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
881 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
882 * will allow dispatch after 1 second and after that slice should
886 tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd;
890 io_allowed = UINT_MAX;
894 if (tg->io_disp[rw] + 1 <= io_allowed) {
900 /* Calc approx time to dispatch */
901 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ) / tg_iops_limit(tg, rw) + 1;
903 if (jiffy_wait > jiffy_elapsed)
904 jiffy_wait = jiffy_wait - jiffy_elapsed;
913 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
916 bool rw = bio_data_dir(bio);
917 u64 bytes_allowed, extra_bytes, tmp;
918 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
920 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
922 /* Slice has just started. Consider one slice interval */
924 jiffy_elapsed_rnd = tg->td->throtl_slice;
926 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
928 tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd;
932 if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
938 /* Calc approx time to dispatch */
939 extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
940 jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw));
946 * This wait time is without taking into consideration the rounding
947 * up we did. Add that time also.
949 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
956 * Returns whether one can dispatch a bio or not. Also returns approx number
957 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
959 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
962 bool rw = bio_data_dir(bio);
963 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
966 * Currently whole state machine of group depends on first bio
967 * queued in the group bio list. So one should not be calling
968 * this function with a different bio if there are other bios
971 BUG_ON(tg->service_queue.nr_queued[rw] &&
972 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
974 /* If tg->bps = -1, then BW is unlimited */
975 if (tg_bps_limit(tg, rw) == U64_MAX &&
976 tg_iops_limit(tg, rw) == UINT_MAX) {
983 * If previous slice expired, start a new one otherwise renew/extend
984 * existing slice to make sure it is at least throtl_slice interval
985 * long since now. New slice is started only for empty throttle group.
986 * If there is queued bio, that means there should be an active
987 * slice and it should be extended instead.
989 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
990 throtl_start_new_slice(tg, rw);
992 if (time_before(tg->slice_end[rw],
993 jiffies + tg->td->throtl_slice))
994 throtl_extend_slice(tg, rw,
995 jiffies + tg->td->throtl_slice);
998 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
999 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
1005 max_wait = max(bps_wait, iops_wait);
1010 if (time_before(tg->slice_end[rw], jiffies + max_wait))
1011 throtl_extend_slice(tg, rw, jiffies + max_wait);
1016 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1018 bool rw = bio_data_dir(bio);
1020 /* Charge the bio to the group */
1021 tg->bytes_disp[rw] += bio->bi_iter.bi_size;
1023 tg->last_bytes_disp[rw] += bio->bi_iter.bi_size;
1024 tg->last_io_disp[rw]++;
1027 * BIO_THROTTLED is used to prevent the same bio to be throttled
1028 * more than once as a throttled bio will go through blk-throtl the
1029 * second time when it eventually gets issued. Set it when a bio
1030 * is being charged to a tg.
1032 if (!bio_flagged(bio, BIO_THROTTLED))
1033 bio_set_flag(bio, BIO_THROTTLED);
1037 * throtl_add_bio_tg - add a bio to the specified throtl_grp
1040 * @tg: the target throtl_grp
1042 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
1043 * tg->qnode_on_self[] is used.
1045 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1046 struct throtl_grp *tg)
1048 struct throtl_service_queue *sq = &tg->service_queue;
1049 bool rw = bio_data_dir(bio);
1052 qn = &tg->qnode_on_self[rw];
1055 * If @tg doesn't currently have any bios queued in the same
1056 * direction, queueing @bio can change when @tg should be
1057 * dispatched. Mark that @tg was empty. This is automatically
1058 * cleaered on the next tg_update_disptime().
1060 if (!sq->nr_queued[rw])
1061 tg->flags |= THROTL_TG_WAS_EMPTY;
1063 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1065 sq->nr_queued[rw]++;
1066 throtl_enqueue_tg(tg);
1069 static void tg_update_disptime(struct throtl_grp *tg)
1071 struct throtl_service_queue *sq = &tg->service_queue;
1072 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1075 bio = throtl_peek_queued(&sq->queued[READ]);
1077 tg_may_dispatch(tg, bio, &read_wait);
1079 bio = throtl_peek_queued(&sq->queued[WRITE]);
1081 tg_may_dispatch(tg, bio, &write_wait);
1083 min_wait = min(read_wait, write_wait);
1084 disptime = jiffies + min_wait;
1086 /* Update dispatch time */
1087 throtl_dequeue_tg(tg);
1088 tg->disptime = disptime;
1089 throtl_enqueue_tg(tg);
1091 /* see throtl_add_bio_tg() */
1092 tg->flags &= ~THROTL_TG_WAS_EMPTY;
1095 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1096 struct throtl_grp *parent_tg, bool rw)
1098 if (throtl_slice_used(parent_tg, rw)) {
1099 throtl_start_new_slice_with_credit(parent_tg, rw,
1100 child_tg->slice_start[rw]);
1105 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1107 struct throtl_service_queue *sq = &tg->service_queue;
1108 struct throtl_service_queue *parent_sq = sq->parent_sq;
1109 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1110 struct throtl_grp *tg_to_put = NULL;
1114 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1115 * from @tg may put its reference and @parent_sq might end up
1116 * getting released prematurely. Remember the tg to put and put it
1117 * after @bio is transferred to @parent_sq.
1119 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1120 sq->nr_queued[rw]--;
1122 throtl_charge_bio(tg, bio);
1125 * If our parent is another tg, we just need to transfer @bio to
1126 * the parent using throtl_add_bio_tg(). If our parent is
1127 * @td->service_queue, @bio is ready to be issued. Put it on its
1128 * bio_lists[] and decrease total number queued. The caller is
1129 * responsible for issuing these bios.
1132 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1133 start_parent_slice_with_credit(tg, parent_tg, rw);
1135 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1136 &parent_sq->queued[rw]);
1137 BUG_ON(tg->td->nr_queued[rw] <= 0);
1138 tg->td->nr_queued[rw]--;
1141 throtl_trim_slice(tg, rw);
1144 blkg_put(tg_to_blkg(tg_to_put));
1147 static int throtl_dispatch_tg(struct throtl_grp *tg)
1149 struct throtl_service_queue *sq = &tg->service_queue;
1150 unsigned int nr_reads = 0, nr_writes = 0;
1151 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1152 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1155 /* Try to dispatch 75% READS and 25% WRITES */
1157 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1158 tg_may_dispatch(tg, bio, NULL)) {
1160 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1163 if (nr_reads >= max_nr_reads)
1167 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1168 tg_may_dispatch(tg, bio, NULL)) {
1170 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1173 if (nr_writes >= max_nr_writes)
1177 return nr_reads + nr_writes;
1180 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1182 unsigned int nr_disp = 0;
1185 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1186 struct throtl_service_queue *sq = &tg->service_queue;
1191 if (time_before(jiffies, tg->disptime))
1194 throtl_dequeue_tg(tg);
1196 nr_disp += throtl_dispatch_tg(tg);
1198 if (sq->nr_queued[0] || sq->nr_queued[1])
1199 tg_update_disptime(tg);
1201 if (nr_disp >= throtl_quantum)
1208 static bool throtl_can_upgrade(struct throtl_data *td,
1209 struct throtl_grp *this_tg);
1211 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1212 * @arg: the throtl_service_queue being serviced
1214 * This timer is armed when a child throtl_grp with active bio's become
1215 * pending and queued on the service_queue's pending_tree and expires when
1216 * the first child throtl_grp should be dispatched. This function
1217 * dispatches bio's from the children throtl_grps to the parent
1220 * If the parent's parent is another throtl_grp, dispatching is propagated
1221 * by either arming its pending_timer or repeating dispatch directly. If
1222 * the top-level service_tree is reached, throtl_data->dispatch_work is
1223 * kicked so that the ready bio's are issued.
1225 static void throtl_pending_timer_fn(unsigned long arg)
1227 struct throtl_service_queue *sq = (void *)arg;
1228 struct throtl_grp *tg = sq_to_tg(sq);
1229 struct throtl_data *td = sq_to_td(sq);
1230 struct request_queue *q = td->queue;
1231 struct throtl_service_queue *parent_sq;
1235 spin_lock_irq(q->queue_lock);
1236 if (throtl_can_upgrade(td, NULL))
1237 throtl_upgrade_state(td);
1240 parent_sq = sq->parent_sq;
1244 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1245 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1246 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1248 ret = throtl_select_dispatch(sq);
1250 throtl_log(sq, "bios disp=%u", ret);
1254 if (throtl_schedule_next_dispatch(sq, false))
1257 /* this dispatch windows is still open, relax and repeat */
1258 spin_unlock_irq(q->queue_lock);
1260 spin_lock_irq(q->queue_lock);
1267 /* @parent_sq is another throl_grp, propagate dispatch */
1268 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1269 tg_update_disptime(tg);
1270 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1271 /* window is already open, repeat dispatching */
1278 /* reached the top-level, queue issueing */
1279 queue_work(kthrotld_workqueue, &td->dispatch_work);
1282 spin_unlock_irq(q->queue_lock);
1286 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1287 * @work: work item being executed
1289 * This function is queued for execution when bio's reach the bio_lists[]
1290 * of throtl_data->service_queue. Those bio's are ready and issued by this
1293 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1295 struct throtl_data *td = container_of(work, struct throtl_data,
1297 struct throtl_service_queue *td_sq = &td->service_queue;
1298 struct request_queue *q = td->queue;
1299 struct bio_list bio_list_on_stack;
1301 struct blk_plug plug;
1304 bio_list_init(&bio_list_on_stack);
1306 spin_lock_irq(q->queue_lock);
1307 for (rw = READ; rw <= WRITE; rw++)
1308 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1309 bio_list_add(&bio_list_on_stack, bio);
1310 spin_unlock_irq(q->queue_lock);
1312 if (!bio_list_empty(&bio_list_on_stack)) {
1313 blk_start_plug(&plug);
1314 while((bio = bio_list_pop(&bio_list_on_stack)))
1315 generic_make_request(bio);
1316 blk_finish_plug(&plug);
1320 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1323 struct throtl_grp *tg = pd_to_tg(pd);
1324 u64 v = *(u64 *)((void *)tg + off);
1328 return __blkg_prfill_u64(sf, pd, v);
1331 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1334 struct throtl_grp *tg = pd_to_tg(pd);
1335 unsigned int v = *(unsigned int *)((void *)tg + off);
1339 return __blkg_prfill_u64(sf, pd, v);
1342 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1344 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1345 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1349 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1351 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1352 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1356 static void tg_conf_updated(struct throtl_grp *tg)
1358 struct throtl_service_queue *sq = &tg->service_queue;
1359 struct cgroup_subsys_state *pos_css;
1360 struct blkcg_gq *blkg;
1362 throtl_log(&tg->service_queue,
1363 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1364 tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1365 tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1368 * Update has_rules[] flags for the updated tg's subtree. A tg is
1369 * considered to have rules if either the tg itself or any of its
1370 * ancestors has rules. This identifies groups without any
1371 * restrictions in the whole hierarchy and allows them to bypass
1374 blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg)) {
1375 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1376 struct throtl_grp *parent_tg;
1378 tg_update_has_rules(this_tg);
1379 /* ignore root/second level */
1380 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1381 !blkg->parent->parent)
1383 parent_tg = blkg_to_tg(blkg->parent);
1385 * make sure all children has lower idle time threshold and
1386 * higher latency target
1388 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1389 parent_tg->idletime_threshold);
1390 this_tg->latency_target = max(this_tg->latency_target,
1391 parent_tg->latency_target);
1395 * We're already holding queue_lock and know @tg is valid. Let's
1396 * apply the new config directly.
1398 * Restart the slices for both READ and WRITES. It might happen
1399 * that a group's limit are dropped suddenly and we don't want to
1400 * account recently dispatched IO with new low rate.
1402 throtl_start_new_slice(tg, 0);
1403 throtl_start_new_slice(tg, 1);
1405 if (tg->flags & THROTL_TG_PENDING) {
1406 tg_update_disptime(tg);
1407 throtl_schedule_next_dispatch(sq->parent_sq, true);
1411 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1412 char *buf, size_t nbytes, loff_t off, bool is_u64)
1414 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1415 struct blkg_conf_ctx ctx;
1416 struct throtl_grp *tg;
1420 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1425 if (sscanf(ctx.body, "%llu", &v) != 1)
1430 tg = blkg_to_tg(ctx.blkg);
1433 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1435 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1437 tg_conf_updated(tg);
1440 blkg_conf_finish(&ctx);
1441 return ret ?: nbytes;
1444 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1445 char *buf, size_t nbytes, loff_t off)
1447 return tg_set_conf(of, buf, nbytes, off, true);
1450 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1451 char *buf, size_t nbytes, loff_t off)
1453 return tg_set_conf(of, buf, nbytes, off, false);
1456 static struct cftype throtl_legacy_files[] = {
1458 .name = "throttle.read_bps_device",
1459 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1460 .seq_show = tg_print_conf_u64,
1461 .write = tg_set_conf_u64,
1464 .name = "throttle.write_bps_device",
1465 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1466 .seq_show = tg_print_conf_u64,
1467 .write = tg_set_conf_u64,
1470 .name = "throttle.read_iops_device",
1471 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1472 .seq_show = tg_print_conf_uint,
1473 .write = tg_set_conf_uint,
1476 .name = "throttle.write_iops_device",
1477 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1478 .seq_show = tg_print_conf_uint,
1479 .write = tg_set_conf_uint,
1482 .name = "throttle.io_service_bytes",
1483 .private = (unsigned long)&blkcg_policy_throtl,
1484 .seq_show = blkg_print_stat_bytes,
1487 .name = "throttle.io_serviced",
1488 .private = (unsigned long)&blkcg_policy_throtl,
1489 .seq_show = blkg_print_stat_ios,
1494 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1497 struct throtl_grp *tg = pd_to_tg(pd);
1498 const char *dname = blkg_dev_name(pd->blkg);
1499 char bufs[4][21] = { "max", "max", "max", "max" };
1501 unsigned int iops_dft;
1502 char idle_time[26] = "";
1503 char latency_time[26] = "";
1508 if (off == LIMIT_LOW) {
1513 iops_dft = UINT_MAX;
1516 if (tg->bps_conf[READ][off] == bps_dft &&
1517 tg->bps_conf[WRITE][off] == bps_dft &&
1518 tg->iops_conf[READ][off] == iops_dft &&
1519 tg->iops_conf[WRITE][off] == iops_dft &&
1520 (off != LIMIT_LOW ||
1521 (tg->idletime_threshold_conf == tg->td->dft_idletime_threshold &&
1522 tg->latency_target_conf == DFL_LATENCY_TARGET)))
1525 if (tg->bps_conf[READ][off] != bps_dft)
1526 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1527 tg->bps_conf[READ][off]);
1528 if (tg->bps_conf[WRITE][off] != bps_dft)
1529 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1530 tg->bps_conf[WRITE][off]);
1531 if (tg->iops_conf[READ][off] != iops_dft)
1532 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1533 tg->iops_conf[READ][off]);
1534 if (tg->iops_conf[WRITE][off] != iops_dft)
1535 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1536 tg->iops_conf[WRITE][off]);
1537 if (off == LIMIT_LOW) {
1538 if (tg->idletime_threshold_conf == ULONG_MAX)
1539 strcpy(idle_time, " idle=max");
1541 snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1542 tg->idletime_threshold_conf);
1544 if (tg->latency_target_conf == ULONG_MAX)
1545 strcpy(latency_time, " latency=max");
1547 snprintf(latency_time, sizeof(latency_time),
1548 " latency=%lu", tg->latency_target_conf);
1551 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1552 dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1557 static int tg_print_limit(struct seq_file *sf, void *v)
1559 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1560 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1564 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1565 char *buf, size_t nbytes, loff_t off)
1567 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1568 struct blkg_conf_ctx ctx;
1569 struct throtl_grp *tg;
1571 unsigned long idle_time;
1572 unsigned long latency_time;
1574 int index = of_cft(of)->private;
1576 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1580 tg = blkg_to_tg(ctx.blkg);
1582 v[0] = tg->bps_conf[READ][index];
1583 v[1] = tg->bps_conf[WRITE][index];
1584 v[2] = tg->iops_conf[READ][index];
1585 v[3] = tg->iops_conf[WRITE][index];
1587 idle_time = tg->idletime_threshold_conf;
1588 latency_time = tg->latency_target_conf;
1590 char tok[27]; /* wiops=18446744073709551616 */
1595 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1604 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1612 if (!strcmp(tok, "rbps"))
1614 else if (!strcmp(tok, "wbps"))
1616 else if (!strcmp(tok, "riops"))
1617 v[2] = min_t(u64, val, UINT_MAX);
1618 else if (!strcmp(tok, "wiops"))
1619 v[3] = min_t(u64, val, UINT_MAX);
1620 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1622 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1628 tg->bps_conf[READ][index] = v[0];
1629 tg->bps_conf[WRITE][index] = v[1];
1630 tg->iops_conf[READ][index] = v[2];
1631 tg->iops_conf[WRITE][index] = v[3];
1633 if (index == LIMIT_MAX) {
1634 tg->bps[READ][index] = v[0];
1635 tg->bps[WRITE][index] = v[1];
1636 tg->iops[READ][index] = v[2];
1637 tg->iops[WRITE][index] = v[3];
1639 tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1640 tg->bps_conf[READ][LIMIT_MAX]);
1641 tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1642 tg->bps_conf[WRITE][LIMIT_MAX]);
1643 tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1644 tg->iops_conf[READ][LIMIT_MAX]);
1645 tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1646 tg->iops_conf[WRITE][LIMIT_MAX]);
1648 if (index == LIMIT_LOW) {
1649 blk_throtl_update_limit_valid(tg->td);
1650 if (tg->td->limit_valid[LIMIT_LOW])
1651 tg->td->limit_index = LIMIT_LOW;
1652 tg->idletime_threshold_conf = idle_time;
1653 tg->idletime_threshold = tg->idletime_threshold_conf;
1654 tg->latency_target_conf = latency_time;
1655 tg->latency_target = tg->latency_target_conf;
1657 tg_conf_updated(tg);
1660 blkg_conf_finish(&ctx);
1661 return ret ?: nbytes;
1664 static struct cftype throtl_files[] = {
1665 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1668 .flags = CFTYPE_NOT_ON_ROOT,
1669 .seq_show = tg_print_limit,
1670 .write = tg_set_limit,
1671 .private = LIMIT_LOW,
1676 .flags = CFTYPE_NOT_ON_ROOT,
1677 .seq_show = tg_print_limit,
1678 .write = tg_set_limit,
1679 .private = LIMIT_MAX,
1684 static void throtl_shutdown_wq(struct request_queue *q)
1686 struct throtl_data *td = q->td;
1688 cancel_work_sync(&td->dispatch_work);
1691 static struct blkcg_policy blkcg_policy_throtl = {
1692 .dfl_cftypes = throtl_files,
1693 .legacy_cftypes = throtl_legacy_files,
1695 .pd_alloc_fn = throtl_pd_alloc,
1696 .pd_init_fn = throtl_pd_init,
1697 .pd_online_fn = throtl_pd_online,
1698 .pd_offline_fn = throtl_pd_offline,
1699 .pd_free_fn = throtl_pd_free,
1702 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1704 unsigned long rtime = jiffies, wtime = jiffies;
1706 if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1707 rtime = tg->last_low_overflow_time[READ];
1708 if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1709 wtime = tg->last_low_overflow_time[WRITE];
1710 return min(rtime, wtime);
1713 /* tg should not be an intermediate node */
1714 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1716 struct throtl_service_queue *parent_sq;
1717 struct throtl_grp *parent = tg;
1718 unsigned long ret = __tg_last_low_overflow_time(tg);
1721 parent_sq = parent->service_queue.parent_sq;
1722 parent = sq_to_tg(parent_sq);
1727 * The parent doesn't have low limit, it always reaches low
1728 * limit. Its overflow time is useless for children
1730 if (!parent->bps[READ][LIMIT_LOW] &&
1731 !parent->iops[READ][LIMIT_LOW] &&
1732 !parent->bps[WRITE][LIMIT_LOW] &&
1733 !parent->iops[WRITE][LIMIT_LOW])
1735 if (time_after(__tg_last_low_overflow_time(parent), ret))
1736 ret = __tg_last_low_overflow_time(parent);
1741 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1744 * cgroup is idle if:
1745 * - single idle is too long, longer than a fixed value (in case user
1746 * configure a too big threshold) or 4 times of slice
1747 * - average think time is more than threshold
1748 * - IO latency is largely below threshold
1750 unsigned long time = jiffies_to_usecs(4 * tg->td->throtl_slice);
1753 time = min_t(unsigned long, MAX_IDLE_TIME, time);
1754 ret = (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1755 tg->avg_idletime > tg->idletime_threshold ||
1756 (tg->latency_target && tg->bio_cnt &&
1757 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1758 throtl_log(&tg->service_queue,
1759 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1760 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1761 tg->bio_cnt, ret, tg->td->scale);
1765 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1767 struct throtl_service_queue *sq = &tg->service_queue;
1768 bool read_limit, write_limit;
1771 * if cgroup reaches low limit (if low limit is 0, the cgroup always
1772 * reaches), it's ok to upgrade to next limit
1774 read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1775 write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1776 if (!read_limit && !write_limit)
1778 if (read_limit && sq->nr_queued[READ] &&
1779 (!write_limit || sq->nr_queued[WRITE]))
1781 if (write_limit && sq->nr_queued[WRITE] &&
1782 (!read_limit || sq->nr_queued[READ]))
1785 if (time_after_eq(jiffies,
1786 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1787 throtl_tg_is_idle(tg))
1792 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1795 if (throtl_tg_can_upgrade(tg))
1797 tg = sq_to_tg(tg->service_queue.parent_sq);
1798 if (!tg || !tg_to_blkg(tg)->parent)
1804 static bool throtl_can_upgrade(struct throtl_data *td,
1805 struct throtl_grp *this_tg)
1807 struct cgroup_subsys_state *pos_css;
1808 struct blkcg_gq *blkg;
1810 if (td->limit_index != LIMIT_LOW)
1813 if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1817 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1818 struct throtl_grp *tg = blkg_to_tg(blkg);
1822 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1824 if (!throtl_hierarchy_can_upgrade(tg)) {
1833 static void throtl_upgrade_check(struct throtl_grp *tg)
1835 unsigned long now = jiffies;
1837 if (tg->td->limit_index != LIMIT_LOW)
1840 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1843 tg->last_check_time = now;
1845 if (!time_after_eq(now,
1846 __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1849 if (throtl_can_upgrade(tg->td, NULL))
1850 throtl_upgrade_state(tg->td);
1853 static void throtl_upgrade_state(struct throtl_data *td)
1855 struct cgroup_subsys_state *pos_css;
1856 struct blkcg_gq *blkg;
1858 throtl_log(&td->service_queue, "upgrade to max");
1859 td->limit_index = LIMIT_MAX;
1860 td->low_upgrade_time = jiffies;
1863 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1864 struct throtl_grp *tg = blkg_to_tg(blkg);
1865 struct throtl_service_queue *sq = &tg->service_queue;
1867 tg->disptime = jiffies - 1;
1868 throtl_select_dispatch(sq);
1869 throtl_schedule_next_dispatch(sq, false);
1872 throtl_select_dispatch(&td->service_queue);
1873 throtl_schedule_next_dispatch(&td->service_queue, false);
1874 queue_work(kthrotld_workqueue, &td->dispatch_work);
1877 static void throtl_downgrade_state(struct throtl_data *td, int new)
1881 throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1883 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1887 td->limit_index = new;
1888 td->low_downgrade_time = jiffies;
1891 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1893 struct throtl_data *td = tg->td;
1894 unsigned long now = jiffies;
1897 * If cgroup is below low limit, consider downgrade and throttle other
1900 if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1901 time_after_eq(now, tg_last_low_overflow_time(tg) +
1902 td->throtl_slice) &&
1903 (!throtl_tg_is_idle(tg) ||
1904 !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1909 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1912 if (!throtl_tg_can_downgrade(tg))
1914 tg = sq_to_tg(tg->service_queue.parent_sq);
1915 if (!tg || !tg_to_blkg(tg)->parent)
1921 static void throtl_downgrade_check(struct throtl_grp *tg)
1925 unsigned long elapsed_time;
1926 unsigned long now = jiffies;
1928 if (tg->td->limit_index != LIMIT_MAX ||
1929 !tg->td->limit_valid[LIMIT_LOW])
1931 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1933 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1936 elapsed_time = now - tg->last_check_time;
1937 tg->last_check_time = now;
1939 if (time_before(now, tg_last_low_overflow_time(tg) +
1940 tg->td->throtl_slice))
1943 if (tg->bps[READ][LIMIT_LOW]) {
1944 bps = tg->last_bytes_disp[READ] * HZ;
1945 do_div(bps, elapsed_time);
1946 if (bps >= tg->bps[READ][LIMIT_LOW])
1947 tg->last_low_overflow_time[READ] = now;
1950 if (tg->bps[WRITE][LIMIT_LOW]) {
1951 bps = tg->last_bytes_disp[WRITE] * HZ;
1952 do_div(bps, elapsed_time);
1953 if (bps >= tg->bps[WRITE][LIMIT_LOW])
1954 tg->last_low_overflow_time[WRITE] = now;
1957 if (tg->iops[READ][LIMIT_LOW]) {
1958 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
1959 if (iops >= tg->iops[READ][LIMIT_LOW])
1960 tg->last_low_overflow_time[READ] = now;
1963 if (tg->iops[WRITE][LIMIT_LOW]) {
1964 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
1965 if (iops >= tg->iops[WRITE][LIMIT_LOW])
1966 tg->last_low_overflow_time[WRITE] = now;
1970 * If cgroup is below low limit, consider downgrade and throttle other
1973 if (throtl_hierarchy_can_downgrade(tg))
1974 throtl_downgrade_state(tg->td, LIMIT_LOW);
1976 tg->last_bytes_disp[READ] = 0;
1977 tg->last_bytes_disp[WRITE] = 0;
1978 tg->last_io_disp[READ] = 0;
1979 tg->last_io_disp[WRITE] = 0;
1982 static void blk_throtl_update_idletime(struct throtl_grp *tg)
1984 unsigned long now = ktime_get_ns() >> 10;
1985 unsigned long last_finish_time = tg->last_finish_time;
1987 if (now <= last_finish_time || last_finish_time == 0 ||
1988 last_finish_time == tg->checked_last_finish_time)
1991 tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
1992 tg->checked_last_finish_time = last_finish_time;
1995 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1996 static void throtl_update_latency_buckets(struct throtl_data *td)
1998 struct avg_latency_bucket avg_latency[LATENCY_BUCKET_SIZE];
2000 unsigned long last_latency = 0;
2001 unsigned long latency;
2003 if (!blk_queue_nonrot(td->queue))
2005 if (time_before(jiffies, td->last_calculate_time + HZ))
2007 td->last_calculate_time = jiffies;
2009 memset(avg_latency, 0, sizeof(avg_latency));
2010 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2011 struct latency_bucket *tmp = &td->tmp_buckets[i];
2013 for_each_possible_cpu(cpu) {
2014 struct latency_bucket *bucket;
2016 /* this isn't race free, but ok in practice */
2017 bucket = per_cpu_ptr(td->latency_buckets, cpu);
2018 tmp->total_latency += bucket[i].total_latency;
2019 tmp->samples += bucket[i].samples;
2020 bucket[i].total_latency = 0;
2021 bucket[i].samples = 0;
2024 if (tmp->samples >= 32) {
2025 int samples = tmp->samples;
2027 latency = tmp->total_latency;
2029 tmp->total_latency = 0;
2034 avg_latency[i].latency = latency;
2038 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2039 if (!avg_latency[i].latency) {
2040 if (td->avg_buckets[i].latency < last_latency)
2041 td->avg_buckets[i].latency = last_latency;
2045 if (!td->avg_buckets[i].valid)
2046 latency = avg_latency[i].latency;
2048 latency = (td->avg_buckets[i].latency * 7 +
2049 avg_latency[i].latency) >> 3;
2051 td->avg_buckets[i].latency = max(latency, last_latency);
2052 td->avg_buckets[i].valid = true;
2053 last_latency = td->avg_buckets[i].latency;
2056 for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2057 throtl_log(&td->service_queue,
2058 "Latency bucket %d: latency=%ld, valid=%d", i,
2059 td->avg_buckets[i].latency, td->avg_buckets[i].valid);
2062 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2067 static void blk_throtl_assoc_bio(struct throtl_grp *tg, struct bio *bio)
2069 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2072 ret = bio_associate_current(bio);
2073 if (ret == 0 || ret == -EBUSY)
2074 bio->bi_cg_private = tg;
2075 blk_stat_set_issue(&bio->bi_issue_stat, bio_sectors(bio));
2077 bio_associate_current(bio);
2081 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
2084 struct throtl_qnode *qn = NULL;
2085 struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
2086 struct throtl_service_queue *sq;
2087 bool rw = bio_data_dir(bio);
2088 bool throttled = false;
2089 struct throtl_data *td = tg->td;
2091 WARN_ON_ONCE(!rcu_read_lock_held());
2093 /* see throtl_charge_bio() */
2094 if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw])
2097 spin_lock_irq(q->queue_lock);
2099 throtl_update_latency_buckets(td);
2101 if (unlikely(blk_queue_bypass(q)))
2104 blk_throtl_assoc_bio(tg, bio);
2105 blk_throtl_update_idletime(tg);
2107 sq = &tg->service_queue;
2111 if (tg->last_low_overflow_time[rw] == 0)
2112 tg->last_low_overflow_time[rw] = jiffies;
2113 throtl_downgrade_check(tg);
2114 throtl_upgrade_check(tg);
2115 /* throtl is FIFO - if bios are already queued, should queue */
2116 if (sq->nr_queued[rw])
2119 /* if above limits, break to queue */
2120 if (!tg_may_dispatch(tg, bio, NULL)) {
2121 tg->last_low_overflow_time[rw] = jiffies;
2122 if (throtl_can_upgrade(td, tg)) {
2123 throtl_upgrade_state(td);
2129 /* within limits, let's charge and dispatch directly */
2130 throtl_charge_bio(tg, bio);
2133 * We need to trim slice even when bios are not being queued
2134 * otherwise it might happen that a bio is not queued for
2135 * a long time and slice keeps on extending and trim is not
2136 * called for a long time. Now if limits are reduced suddenly
2137 * we take into account all the IO dispatched so far at new
2138 * low rate and * newly queued IO gets a really long dispatch
2141 * So keep on trimming slice even if bio is not queued.
2143 throtl_trim_slice(tg, rw);
2146 * @bio passed through this layer without being throttled.
2147 * Climb up the ladder. If we''re already at the top, it
2148 * can be executed directly.
2150 qn = &tg->qnode_on_parent[rw];
2157 /* out-of-limit, queue to @tg */
2158 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2159 rw == READ ? 'R' : 'W',
2160 tg->bytes_disp[rw], bio->bi_iter.bi_size,
2161 tg_bps_limit(tg, rw),
2162 tg->io_disp[rw], tg_iops_limit(tg, rw),
2163 sq->nr_queued[READ], sq->nr_queued[WRITE]);
2165 tg->last_low_overflow_time[rw] = jiffies;
2167 td->nr_queued[rw]++;
2168 throtl_add_bio_tg(bio, qn, tg);
2172 * Update @tg's dispatch time and force schedule dispatch if @tg
2173 * was empty before @bio. The forced scheduling isn't likely to
2174 * cause undue delay as @bio is likely to be dispatched directly if
2175 * its @tg's disptime is not in the future.
2177 if (tg->flags & THROTL_TG_WAS_EMPTY) {
2178 tg_update_disptime(tg);
2179 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2183 spin_unlock_irq(q->queue_lock);
2186 * As multiple blk-throtls may stack in the same issue path, we
2187 * don't want bios to leave with the flag set. Clear the flag if
2191 bio_clear_flag(bio, BIO_THROTTLED);
2193 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2194 if (throttled || !td->track_bio_latency)
2195 bio->bi_issue_stat.stat |= SKIP_LATENCY;
2200 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2201 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2202 int op, unsigned long time)
2204 struct latency_bucket *latency;
2207 if (!td || td->limit_index != LIMIT_LOW || op != REQ_OP_READ ||
2208 !blk_queue_nonrot(td->queue))
2211 index = request_bucket_index(size);
2213 latency = get_cpu_ptr(td->latency_buckets);
2214 latency[index].total_latency += time;
2215 latency[index].samples++;
2216 put_cpu_ptr(td->latency_buckets);
2219 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2221 struct request_queue *q = rq->q;
2222 struct throtl_data *td = q->td;
2224 throtl_track_latency(td, blk_stat_size(&rq->issue_stat),
2225 req_op(rq), time_ns >> 10);
2228 void blk_throtl_bio_endio(struct bio *bio)
2230 struct throtl_grp *tg;
2232 unsigned long finish_time;
2233 unsigned long start_time;
2236 tg = bio->bi_cg_private;
2239 bio->bi_cg_private = NULL;
2241 finish_time_ns = ktime_get_ns();
2242 tg->last_finish_time = finish_time_ns >> 10;
2244 start_time = blk_stat_time(&bio->bi_issue_stat) >> 10;
2245 finish_time = __blk_stat_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_stat.stat & SKIP_LATENCY))
2252 throtl_track_latency(tg->td, blk_stat_size(&bio->bi_issue_stat),
2255 if (tg->latency_target) {
2257 unsigned int threshold;
2259 bucket = request_bucket_index(
2260 blk_stat_size(&bio->bi_issue_stat));
2261 threshold = tg->td->avg_buckets[bucket].latency +
2263 if (lat > threshold)
2266 * Not race free, could get wrong count, which means cgroups
2272 if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2273 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2275 tg->bad_bio_cnt /= 2;
2281 * Dispatch all bios from all children tg's queued on @parent_sq. On
2282 * return, @parent_sq is guaranteed to not have any active children tg's
2283 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
2285 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
2287 struct throtl_grp *tg;
2289 while ((tg = throtl_rb_first(parent_sq))) {
2290 struct throtl_service_queue *sq = &tg->service_queue;
2293 throtl_dequeue_tg(tg);
2295 while ((bio = throtl_peek_queued(&sq->queued[READ])))
2296 tg_dispatch_one_bio(tg, bio_data_dir(bio));
2297 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
2298 tg_dispatch_one_bio(tg, bio_data_dir(bio));
2303 * blk_throtl_drain - drain throttled bios
2304 * @q: request_queue to drain throttled bios for
2306 * Dispatch all currently throttled bios on @q through ->make_request_fn().
2308 void blk_throtl_drain(struct request_queue *q)
2309 __releases(q->queue_lock) __acquires(q->queue_lock)
2311 struct throtl_data *td = q->td;
2312 struct blkcg_gq *blkg;
2313 struct cgroup_subsys_state *pos_css;
2317 queue_lockdep_assert_held(q);
2321 * Drain each tg while doing post-order walk on the blkg tree, so
2322 * that all bios are propagated to td->service_queue. It'd be
2323 * better to walk service_queue tree directly but blkg walk is
2326 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
2327 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
2329 /* finally, transfer bios from top-level tg's into the td */
2330 tg_drain_bios(&td->service_queue);
2333 spin_unlock_irq(q->queue_lock);
2335 /* all bios now should be in td->service_queue, issue them */
2336 for (rw = READ; rw <= WRITE; rw++)
2337 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
2339 generic_make_request(bio);
2341 spin_lock_irq(q->queue_lock);
2344 int blk_throtl_init(struct request_queue *q)
2346 struct throtl_data *td;
2349 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2352 td->latency_buckets = __alloc_percpu(sizeof(struct latency_bucket) *
2353 LATENCY_BUCKET_SIZE, __alignof__(u64));
2354 if (!td->latency_buckets) {
2359 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2360 throtl_service_queue_init(&td->service_queue);
2365 td->limit_valid[LIMIT_MAX] = true;
2366 td->limit_index = LIMIT_MAX;
2367 td->low_upgrade_time = jiffies;
2368 td->low_downgrade_time = jiffies;
2370 /* activate policy */
2371 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2373 free_percpu(td->latency_buckets);
2379 void blk_throtl_exit(struct request_queue *q)
2382 throtl_shutdown_wq(q);
2383 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2384 free_percpu(q->td->latency_buckets);
2388 void blk_throtl_register_queue(struct request_queue *q)
2390 struct throtl_data *td;
2391 struct cgroup_subsys_state *pos_css;
2392 struct blkcg_gq *blkg;
2397 if (blk_queue_nonrot(q)) {
2398 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2399 td->dft_idletime_threshold = DFL_IDLE_THRESHOLD_SSD;
2401 td->throtl_slice = DFL_THROTL_SLICE_HD;
2402 td->dft_idletime_threshold = DFL_IDLE_THRESHOLD_HD;
2404 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2405 /* if no low limit, use previous default */
2406 td->throtl_slice = DFL_THROTL_SLICE_HD;
2409 td->track_bio_latency = !q->mq_ops && !q->request_fn;
2410 if (!td->track_bio_latency)
2411 blk_stat_enable_accounting(q);
2414 * some tg are created before queue is fully initialized, eg, nonrot
2415 * isn't initialized yet
2418 blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) {
2419 struct throtl_grp *tg = blkg_to_tg(blkg);
2421 tg->idletime_threshold = td->dft_idletime_threshold;
2422 tg->idletime_threshold_conf = td->dft_idletime_threshold;
2427 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2428 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2432 return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2435 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2436 const char *page, size_t count)
2443 if (kstrtoul(page, 10, &v))
2445 t = msecs_to_jiffies(v);
2446 if (t == 0 || t > MAX_THROTL_SLICE)
2448 q->td->throtl_slice = t;
2453 static int __init throtl_init(void)
2455 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2456 if (!kthrotld_workqueue)
2457 panic("Failed to create kthrotld\n");
2459 return blkcg_policy_register(&blkcg_policy_throtl);
2462 module_init(throtl_init);