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)
30 #define MIN_THROTL_BPS (320 * 1024)
31 #define MIN_THROTL_IOPS (10)
33 #define SKIP_LATENCY (((u64)1) << BLK_STAT_RES_SHIFT)
35 static struct blkcg_policy blkcg_policy_throtl;
37 /* A workqueue to queue throttle related work */
38 static struct workqueue_struct *kthrotld_workqueue;
41 * To implement hierarchical throttling, throtl_grps form a tree and bios
42 * are dispatched upwards level by level until they reach the top and get
43 * issued. When dispatching bios from the children and local group at each
44 * level, if the bios are dispatched into a single bio_list, there's a risk
45 * of a local or child group which can queue many bios at once filling up
46 * the list starving others.
48 * To avoid such starvation, dispatched bios are queued separately
49 * according to where they came from. When they are again dispatched to
50 * the parent, they're popped in round-robin order so that no single source
51 * hogs the dispatch window.
53 * throtl_qnode is used to keep the queued bios separated by their sources.
54 * Bios are queued to throtl_qnode which in turn is queued to
55 * throtl_service_queue and then dispatched in round-robin order.
57 * It's also used to track the reference counts on blkg's. A qnode always
58 * belongs to a throtl_grp and gets queued on itself or the parent, so
59 * incrementing the reference of the associated throtl_grp when a qnode is
60 * queued and decrementing when dequeued is enough to keep the whole blkg
61 * tree pinned while bios are in flight.
64 struct list_head node; /* service_queue->queued[] */
65 struct bio_list bios; /* queued bios */
66 struct throtl_grp *tg; /* tg this qnode belongs to */
69 struct throtl_service_queue {
70 struct throtl_service_queue *parent_sq; /* the parent service_queue */
73 * Bios queued directly to this service_queue or dispatched from
74 * children throtl_grp's.
76 struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */
77 unsigned int nr_queued[2]; /* number of queued bios */
80 * RB tree of active children throtl_grp's, which are sorted by
83 struct rb_root pending_tree; /* RB tree of active tgs */
84 struct rb_node *first_pending; /* first node in the tree */
85 unsigned int nr_pending; /* # queued in the tree */
86 unsigned long first_pending_disptime; /* disptime of the first tg */
87 struct timer_list pending_timer; /* fires on first_pending_disptime */
91 THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */
92 THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */
95 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
104 /* must be the first member */
105 struct blkg_policy_data pd;
107 /* active throtl group service_queue member */
108 struct rb_node rb_node;
110 /* throtl_data this group belongs to */
111 struct throtl_data *td;
113 /* this group's service queue */
114 struct throtl_service_queue service_queue;
117 * qnode_on_self is used when bios are directly queued to this
118 * throtl_grp so that local bios compete fairly with bios
119 * dispatched from children. qnode_on_parent is used when bios are
120 * dispatched from this throtl_grp into its parent and will compete
121 * with the sibling qnode_on_parents and the parent's
124 struct throtl_qnode qnode_on_self[2];
125 struct throtl_qnode qnode_on_parent[2];
128 * Dispatch time in jiffies. This is the estimated time when group
129 * will unthrottle and is ready to dispatch more bio. It is used as
130 * key to sort active groups in service tree.
132 unsigned long disptime;
136 /* are there any throtl rules between this group and td? */
139 /* internally used bytes per second rate limits */
140 uint64_t bps[2][LIMIT_CNT];
141 /* user configured bps limits */
142 uint64_t bps_conf[2][LIMIT_CNT];
144 /* internally used IOPS limits */
145 unsigned int iops[2][LIMIT_CNT];
146 /* user configured IOPS limits */
147 unsigned int iops_conf[2][LIMIT_CNT];
149 /* Number of bytes disptached in current slice */
150 uint64_t bytes_disp[2];
151 /* Number of bio's dispatched in current slice */
152 unsigned int io_disp[2];
154 unsigned long last_low_overflow_time[2];
156 uint64_t last_bytes_disp[2];
157 unsigned int last_io_disp[2];
159 unsigned long last_check_time;
161 unsigned long latency_target; /* us */
162 unsigned long latency_target_conf; /* us */
163 /* When did we start a new slice */
164 unsigned long slice_start[2];
165 unsigned long slice_end[2];
167 unsigned long last_finish_time; /* ns / 1024 */
168 unsigned long checked_last_finish_time; /* ns / 1024 */
169 unsigned long avg_idletime; /* ns / 1024 */
170 unsigned long idletime_threshold; /* us */
171 unsigned long idletime_threshold_conf; /* us */
173 unsigned int bio_cnt; /* total bios */
174 unsigned int bad_bio_cnt; /* bios exceeding latency threshold */
175 unsigned long bio_cnt_reset_time;
178 /* We measure latency for request size from <= 4k to >= 1M */
179 #define LATENCY_BUCKET_SIZE 9
181 struct latency_bucket {
182 unsigned long total_latency; /* ns / 1024 */
186 struct avg_latency_bucket {
187 unsigned long latency; /* ns / 1024 */
193 /* service tree for active throtl groups */
194 struct throtl_service_queue service_queue;
196 struct request_queue *queue;
198 /* Total Number of queued bios on READ and WRITE lists */
199 unsigned int nr_queued[2];
201 unsigned int throtl_slice;
203 /* Work for dispatching throttled bios */
204 struct work_struct dispatch_work;
205 unsigned int limit_index;
206 bool limit_valid[LIMIT_CNT];
208 unsigned long dft_idletime_threshold; /* us */
210 unsigned long low_upgrade_time;
211 unsigned long low_downgrade_time;
215 struct latency_bucket tmp_buckets[LATENCY_BUCKET_SIZE];
216 struct avg_latency_bucket avg_buckets[LATENCY_BUCKET_SIZE];
217 struct latency_bucket __percpu *latency_buckets;
218 unsigned long last_calculate_time;
220 bool track_bio_latency;
223 static void throtl_pending_timer_fn(unsigned long arg);
225 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
227 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
230 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
232 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
235 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
237 return pd_to_blkg(&tg->pd);
241 * sq_to_tg - return the throl_grp the specified service queue belongs to
242 * @sq: the throtl_service_queue of interest
244 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
245 * embedded in throtl_data, %NULL is returned.
247 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
249 if (sq && sq->parent_sq)
250 return container_of(sq, struct throtl_grp, service_queue);
256 * sq_to_td - return throtl_data the specified service queue belongs to
257 * @sq: the throtl_service_queue of interest
259 * A service_queue can be embedded in either a throtl_grp or throtl_data.
260 * Determine the associated throtl_data accordingly and return it.
262 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
264 struct throtl_grp *tg = sq_to_tg(sq);
269 return container_of(sq, struct throtl_data, service_queue);
273 * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
274 * make the IO dispatch more smooth.
275 * Scale up: linearly scale up according to lapsed time since upgrade. For
276 * every throtl_slice, the limit scales up 1/2 .low limit till the
277 * limit hits .max limit
278 * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
280 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
282 /* arbitrary value to avoid too big scale */
283 if (td->scale < 4096 && time_after_eq(jiffies,
284 td->low_upgrade_time + td->scale * td->throtl_slice))
285 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
287 return low + (low >> 1) * td->scale;
290 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
292 struct blkcg_gq *blkg = tg_to_blkg(tg);
293 struct throtl_data *td;
296 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
300 ret = tg->bps[rw][td->limit_index];
301 if (ret == 0 && td->limit_index == LIMIT_LOW) {
302 /* intermediate node or iops isn't 0 */
303 if (!list_empty(&blkg->blkcg->css.children) ||
304 tg->iops[rw][td->limit_index])
307 return MIN_THROTL_BPS;
310 if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
311 tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
314 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
315 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
320 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
322 struct blkcg_gq *blkg = tg_to_blkg(tg);
323 struct throtl_data *td;
326 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
330 ret = tg->iops[rw][td->limit_index];
331 if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
332 /* intermediate node or bps isn't 0 */
333 if (!list_empty(&blkg->blkcg->css.children) ||
334 tg->bps[rw][td->limit_index])
337 return MIN_THROTL_IOPS;
340 if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
341 tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
344 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
345 if (adjusted > UINT_MAX)
347 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
352 #define request_bucket_index(sectors) \
353 clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
356 * throtl_log - log debug message via blktrace
357 * @sq: the service_queue being reported
358 * @fmt: printf format string
361 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
362 * throtl_grp; otherwise, just "throtl".
364 #define throtl_log(sq, fmt, args...) do { \
365 struct throtl_grp *__tg = sq_to_tg((sq)); \
366 struct throtl_data *__td = sq_to_td((sq)); \
369 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
374 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
375 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
377 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
381 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
383 INIT_LIST_HEAD(&qn->node);
384 bio_list_init(&qn->bios);
389 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
390 * @bio: bio being added
391 * @qn: qnode to add bio to
392 * @queued: the service_queue->queued[] list @qn belongs to
394 * Add @bio to @qn and put @qn on @queued if it's not already on.
395 * @qn->tg's reference count is bumped when @qn is activated. See the
396 * comment on top of throtl_qnode definition for details.
398 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
399 struct list_head *queued)
401 bio_list_add(&qn->bios, bio);
402 if (list_empty(&qn->node)) {
403 list_add_tail(&qn->node, queued);
404 blkg_get(tg_to_blkg(qn->tg));
409 * throtl_peek_queued - peek the first bio on a qnode list
410 * @queued: the qnode list to peek
412 static struct bio *throtl_peek_queued(struct list_head *queued)
414 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
417 if (list_empty(queued))
420 bio = bio_list_peek(&qn->bios);
426 * throtl_pop_queued - pop the first bio form a qnode list
427 * @queued: the qnode list to pop a bio from
428 * @tg_to_put: optional out argument for throtl_grp to put
430 * Pop the first bio from the qnode list @queued. After popping, the first
431 * qnode is removed from @queued if empty or moved to the end of @queued so
432 * that the popping order is round-robin.
434 * When the first qnode is removed, its associated throtl_grp should be put
435 * too. If @tg_to_put is NULL, this function automatically puts it;
436 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
437 * responsible for putting it.
439 static struct bio *throtl_pop_queued(struct list_head *queued,
440 struct throtl_grp **tg_to_put)
442 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
445 if (list_empty(queued))
448 bio = bio_list_pop(&qn->bios);
451 if (bio_list_empty(&qn->bios)) {
452 list_del_init(&qn->node);
456 blkg_put(tg_to_blkg(qn->tg));
458 list_move_tail(&qn->node, queued);
464 /* init a service_queue, assumes the caller zeroed it */
465 static void throtl_service_queue_init(struct throtl_service_queue *sq)
467 INIT_LIST_HEAD(&sq->queued[0]);
468 INIT_LIST_HEAD(&sq->queued[1]);
469 sq->pending_tree = RB_ROOT;
470 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
474 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
476 struct throtl_grp *tg;
479 tg = kzalloc_node(sizeof(*tg), gfp, node);
483 throtl_service_queue_init(&tg->service_queue);
485 for (rw = READ; rw <= WRITE; rw++) {
486 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
487 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
490 RB_CLEAR_NODE(&tg->rb_node);
491 tg->bps[READ][LIMIT_MAX] = U64_MAX;
492 tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
493 tg->iops[READ][LIMIT_MAX] = UINT_MAX;
494 tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
495 tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
496 tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
497 tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
498 tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
499 /* LIMIT_LOW will have default value 0 */
501 tg->latency_target = DFL_LATENCY_TARGET;
502 tg->latency_target_conf = DFL_LATENCY_TARGET;
507 static void throtl_pd_init(struct blkg_policy_data *pd)
509 struct throtl_grp *tg = pd_to_tg(pd);
510 struct blkcg_gq *blkg = tg_to_blkg(tg);
511 struct throtl_data *td = blkg->q->td;
512 struct throtl_service_queue *sq = &tg->service_queue;
515 * If on the default hierarchy, we switch to properly hierarchical
516 * behavior where limits on a given throtl_grp are applied to the
517 * whole subtree rather than just the group itself. e.g. If 16M
518 * read_bps limit is set on the root group, the whole system can't
519 * exceed 16M for the device.
521 * If not on the default hierarchy, the broken flat hierarchy
522 * behavior is retained where all throtl_grps are treated as if
523 * they're all separate root groups right below throtl_data.
524 * Limits of a group don't interact with limits of other groups
525 * regardless of the position of the group in the hierarchy.
527 sq->parent_sq = &td->service_queue;
528 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
529 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
532 tg->idletime_threshold = td->dft_idletime_threshold;
533 tg->idletime_threshold_conf = td->dft_idletime_threshold;
537 * Set has_rules[] if @tg or any of its parents have limits configured.
538 * This doesn't require walking up to the top of the hierarchy as the
539 * parent's has_rules[] is guaranteed to be correct.
541 static void tg_update_has_rules(struct throtl_grp *tg)
543 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
544 struct throtl_data *td = tg->td;
547 for (rw = READ; rw <= WRITE; rw++)
548 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
549 (td->limit_valid[td->limit_index] &&
550 (tg_bps_limit(tg, rw) != U64_MAX ||
551 tg_iops_limit(tg, rw) != UINT_MAX));
554 static void throtl_pd_online(struct blkg_policy_data *pd)
556 struct throtl_grp *tg = pd_to_tg(pd);
558 * We don't want new groups to escape the limits of its ancestors.
559 * Update has_rules[] after a new group is brought online.
561 tg_update_has_rules(tg);
564 static void blk_throtl_update_limit_valid(struct throtl_data *td)
566 struct cgroup_subsys_state *pos_css;
567 struct blkcg_gq *blkg;
568 bool low_valid = false;
571 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
572 struct throtl_grp *tg = blkg_to_tg(blkg);
574 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
575 tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
580 td->limit_valid[LIMIT_LOW] = low_valid;
583 static void throtl_upgrade_state(struct throtl_data *td);
584 static void throtl_pd_offline(struct blkg_policy_data *pd)
586 struct throtl_grp *tg = pd_to_tg(pd);
588 tg->bps[READ][LIMIT_LOW] = 0;
589 tg->bps[WRITE][LIMIT_LOW] = 0;
590 tg->iops[READ][LIMIT_LOW] = 0;
591 tg->iops[WRITE][LIMIT_LOW] = 0;
593 blk_throtl_update_limit_valid(tg->td);
595 if (!tg->td->limit_valid[tg->td->limit_index])
596 throtl_upgrade_state(tg->td);
599 static void throtl_pd_free(struct blkg_policy_data *pd)
601 struct throtl_grp *tg = pd_to_tg(pd);
603 del_timer_sync(&tg->service_queue.pending_timer);
607 static struct throtl_grp *
608 throtl_rb_first(struct throtl_service_queue *parent_sq)
610 /* Service tree is empty */
611 if (!parent_sq->nr_pending)
614 if (!parent_sq->first_pending)
615 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
617 if (parent_sq->first_pending)
618 return rb_entry_tg(parent_sq->first_pending);
623 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
629 static void throtl_rb_erase(struct rb_node *n,
630 struct throtl_service_queue *parent_sq)
632 if (parent_sq->first_pending == n)
633 parent_sq->first_pending = NULL;
634 rb_erase_init(n, &parent_sq->pending_tree);
635 --parent_sq->nr_pending;
638 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
640 struct throtl_grp *tg;
642 tg = throtl_rb_first(parent_sq);
646 parent_sq->first_pending_disptime = tg->disptime;
649 static void tg_service_queue_add(struct throtl_grp *tg)
651 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
652 struct rb_node **node = &parent_sq->pending_tree.rb_node;
653 struct rb_node *parent = NULL;
654 struct throtl_grp *__tg;
655 unsigned long key = tg->disptime;
658 while (*node != NULL) {
660 __tg = rb_entry_tg(parent);
662 if (time_before(key, __tg->disptime))
663 node = &parent->rb_left;
665 node = &parent->rb_right;
671 parent_sq->first_pending = &tg->rb_node;
673 rb_link_node(&tg->rb_node, parent, node);
674 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
677 static void __throtl_enqueue_tg(struct throtl_grp *tg)
679 tg_service_queue_add(tg);
680 tg->flags |= THROTL_TG_PENDING;
681 tg->service_queue.parent_sq->nr_pending++;
684 static void throtl_enqueue_tg(struct throtl_grp *tg)
686 if (!(tg->flags & THROTL_TG_PENDING))
687 __throtl_enqueue_tg(tg);
690 static void __throtl_dequeue_tg(struct throtl_grp *tg)
692 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
693 tg->flags &= ~THROTL_TG_PENDING;
696 static void throtl_dequeue_tg(struct throtl_grp *tg)
698 if (tg->flags & THROTL_TG_PENDING)
699 __throtl_dequeue_tg(tg);
702 /* Call with queue lock held */
703 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
704 unsigned long expires)
706 unsigned long max_expire = jiffies + 8 * sq_to_tg(sq)->td->throtl_slice;
709 * Since we are adjusting the throttle limit dynamically, the sleep
710 * time calculated according to previous limit might be invalid. It's
711 * possible the cgroup sleep time is very long and no other cgroups
712 * have IO running so notify the limit changes. Make sure the cgroup
713 * doesn't sleep too long to avoid the missed notification.
715 if (time_after(expires, max_expire))
716 expires = max_expire;
717 mod_timer(&sq->pending_timer, expires);
718 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
719 expires - jiffies, jiffies);
723 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
724 * @sq: the service_queue to schedule dispatch for
725 * @force: force scheduling
727 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
728 * dispatch time of the first pending child. Returns %true if either timer
729 * is armed or there's no pending child left. %false if the current
730 * dispatch window is still open and the caller should continue
733 * If @force is %true, the dispatch timer is always scheduled and this
734 * function is guaranteed to return %true. This is to be used when the
735 * caller can't dispatch itself and needs to invoke pending_timer
736 * unconditionally. Note that forced scheduling is likely to induce short
737 * delay before dispatch starts even if @sq->first_pending_disptime is not
738 * in the future and thus shouldn't be used in hot paths.
740 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
743 /* any pending children left? */
747 update_min_dispatch_time(sq);
749 /* is the next dispatch time in the future? */
750 if (force || time_after(sq->first_pending_disptime, jiffies)) {
751 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
755 /* tell the caller to continue dispatching */
759 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
760 bool rw, unsigned long start)
762 tg->bytes_disp[rw] = 0;
766 * Previous slice has expired. We must have trimmed it after last
767 * bio dispatch. That means since start of last slice, we never used
768 * that bandwidth. Do try to make use of that bandwidth while giving
771 if (time_after_eq(start, tg->slice_start[rw]))
772 tg->slice_start[rw] = start;
774 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
775 throtl_log(&tg->service_queue,
776 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
777 rw == READ ? 'R' : 'W', tg->slice_start[rw],
778 tg->slice_end[rw], jiffies);
781 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
783 tg->bytes_disp[rw] = 0;
785 tg->slice_start[rw] = jiffies;
786 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
787 throtl_log(&tg->service_queue,
788 "[%c] new slice start=%lu end=%lu jiffies=%lu",
789 rw == READ ? 'R' : 'W', tg->slice_start[rw],
790 tg->slice_end[rw], jiffies);
793 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
794 unsigned long jiffy_end)
796 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
799 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
800 unsigned long jiffy_end)
802 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
803 throtl_log(&tg->service_queue,
804 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
805 rw == READ ? 'R' : 'W', tg->slice_start[rw],
806 tg->slice_end[rw], jiffies);
809 /* Determine if previously allocated or extended slice is complete or not */
810 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
812 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
818 /* Trim the used slices and adjust slice start accordingly */
819 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
821 unsigned long nr_slices, time_elapsed, io_trim;
824 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
827 * If bps are unlimited (-1), then time slice don't get
828 * renewed. Don't try to trim the slice if slice is used. A new
829 * slice will start when appropriate.
831 if (throtl_slice_used(tg, rw))
835 * A bio has been dispatched. Also adjust slice_end. It might happen
836 * that initially cgroup limit was very low resulting in high
837 * slice_end, but later limit was bumped up and bio was dispached
838 * sooner, then we need to reduce slice_end. A high bogus slice_end
839 * is bad because it does not allow new slice to start.
842 throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
844 time_elapsed = jiffies - tg->slice_start[rw];
846 nr_slices = time_elapsed / tg->td->throtl_slice;
850 tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
854 io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
857 if (!bytes_trim && !io_trim)
860 if (tg->bytes_disp[rw] >= bytes_trim)
861 tg->bytes_disp[rw] -= bytes_trim;
863 tg->bytes_disp[rw] = 0;
865 if (tg->io_disp[rw] >= io_trim)
866 tg->io_disp[rw] -= io_trim;
870 tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
872 throtl_log(&tg->service_queue,
873 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
874 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
875 tg->slice_start[rw], tg->slice_end[rw], jiffies);
878 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
881 bool rw = bio_data_dir(bio);
882 unsigned int io_allowed;
883 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
886 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
888 /* Slice has just started. Consider one slice interval */
890 jiffy_elapsed_rnd = tg->td->throtl_slice;
892 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
895 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
896 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
897 * will allow dispatch after 1 second and after that slice should
901 tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd;
905 io_allowed = UINT_MAX;
909 if (tg->io_disp[rw] + 1 <= io_allowed) {
915 /* Calc approx time to dispatch */
916 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ) / tg_iops_limit(tg, rw) + 1;
918 if (jiffy_wait > jiffy_elapsed)
919 jiffy_wait = jiffy_wait - jiffy_elapsed;
928 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
931 bool rw = bio_data_dir(bio);
932 u64 bytes_allowed, extra_bytes, tmp;
933 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
935 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
937 /* Slice has just started. Consider one slice interval */
939 jiffy_elapsed_rnd = tg->td->throtl_slice;
941 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
943 tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd;
947 if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
953 /* Calc approx time to dispatch */
954 extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
955 jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw));
961 * This wait time is without taking into consideration the rounding
962 * up we did. Add that time also.
964 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
971 * Returns whether one can dispatch a bio or not. Also returns approx number
972 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
974 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
977 bool rw = bio_data_dir(bio);
978 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
981 * Currently whole state machine of group depends on first bio
982 * queued in the group bio list. So one should not be calling
983 * this function with a different bio if there are other bios
986 BUG_ON(tg->service_queue.nr_queued[rw] &&
987 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
989 /* If tg->bps = -1, then BW is unlimited */
990 if (tg_bps_limit(tg, rw) == U64_MAX &&
991 tg_iops_limit(tg, rw) == UINT_MAX) {
998 * If previous slice expired, start a new one otherwise renew/extend
999 * existing slice to make sure it is at least throtl_slice interval
1000 * long since now. New slice is started only for empty throttle group.
1001 * If there is queued bio, that means there should be an active
1002 * slice and it should be extended instead.
1004 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
1005 throtl_start_new_slice(tg, rw);
1007 if (time_before(tg->slice_end[rw],
1008 jiffies + tg->td->throtl_slice))
1009 throtl_extend_slice(tg, rw,
1010 jiffies + tg->td->throtl_slice);
1013 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
1014 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
1020 max_wait = max(bps_wait, iops_wait);
1025 if (time_before(tg->slice_end[rw], jiffies + max_wait))
1026 throtl_extend_slice(tg, rw, jiffies + max_wait);
1031 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1033 bool rw = bio_data_dir(bio);
1035 /* Charge the bio to the group */
1036 tg->bytes_disp[rw] += bio->bi_iter.bi_size;
1038 tg->last_bytes_disp[rw] += bio->bi_iter.bi_size;
1039 tg->last_io_disp[rw]++;
1042 * BIO_THROTTLED is used to prevent the same bio to be throttled
1043 * more than once as a throttled bio will go through blk-throtl the
1044 * second time when it eventually gets issued. Set it when a bio
1045 * is being charged to a tg.
1047 if (!bio_flagged(bio, BIO_THROTTLED))
1048 bio_set_flag(bio, BIO_THROTTLED);
1052 * throtl_add_bio_tg - add a bio to the specified throtl_grp
1055 * @tg: the target throtl_grp
1057 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
1058 * tg->qnode_on_self[] is used.
1060 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1061 struct throtl_grp *tg)
1063 struct throtl_service_queue *sq = &tg->service_queue;
1064 bool rw = bio_data_dir(bio);
1067 qn = &tg->qnode_on_self[rw];
1070 * If @tg doesn't currently have any bios queued in the same
1071 * direction, queueing @bio can change when @tg should be
1072 * dispatched. Mark that @tg was empty. This is automatically
1073 * cleaered on the next tg_update_disptime().
1075 if (!sq->nr_queued[rw])
1076 tg->flags |= THROTL_TG_WAS_EMPTY;
1078 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1080 sq->nr_queued[rw]++;
1081 throtl_enqueue_tg(tg);
1084 static void tg_update_disptime(struct throtl_grp *tg)
1086 struct throtl_service_queue *sq = &tg->service_queue;
1087 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1090 bio = throtl_peek_queued(&sq->queued[READ]);
1092 tg_may_dispatch(tg, bio, &read_wait);
1094 bio = throtl_peek_queued(&sq->queued[WRITE]);
1096 tg_may_dispatch(tg, bio, &write_wait);
1098 min_wait = min(read_wait, write_wait);
1099 disptime = jiffies + min_wait;
1101 /* Update dispatch time */
1102 throtl_dequeue_tg(tg);
1103 tg->disptime = disptime;
1104 throtl_enqueue_tg(tg);
1106 /* see throtl_add_bio_tg() */
1107 tg->flags &= ~THROTL_TG_WAS_EMPTY;
1110 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1111 struct throtl_grp *parent_tg, bool rw)
1113 if (throtl_slice_used(parent_tg, rw)) {
1114 throtl_start_new_slice_with_credit(parent_tg, rw,
1115 child_tg->slice_start[rw]);
1120 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1122 struct throtl_service_queue *sq = &tg->service_queue;
1123 struct throtl_service_queue *parent_sq = sq->parent_sq;
1124 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1125 struct throtl_grp *tg_to_put = NULL;
1129 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1130 * from @tg may put its reference and @parent_sq might end up
1131 * getting released prematurely. Remember the tg to put and put it
1132 * after @bio is transferred to @parent_sq.
1134 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1135 sq->nr_queued[rw]--;
1137 throtl_charge_bio(tg, bio);
1140 * If our parent is another tg, we just need to transfer @bio to
1141 * the parent using throtl_add_bio_tg(). If our parent is
1142 * @td->service_queue, @bio is ready to be issued. Put it on its
1143 * bio_lists[] and decrease total number queued. The caller is
1144 * responsible for issuing these bios.
1147 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1148 start_parent_slice_with_credit(tg, parent_tg, rw);
1150 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1151 &parent_sq->queued[rw]);
1152 BUG_ON(tg->td->nr_queued[rw] <= 0);
1153 tg->td->nr_queued[rw]--;
1156 throtl_trim_slice(tg, rw);
1159 blkg_put(tg_to_blkg(tg_to_put));
1162 static int throtl_dispatch_tg(struct throtl_grp *tg)
1164 struct throtl_service_queue *sq = &tg->service_queue;
1165 unsigned int nr_reads = 0, nr_writes = 0;
1166 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1167 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1170 /* Try to dispatch 75% READS and 25% WRITES */
1172 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1173 tg_may_dispatch(tg, bio, NULL)) {
1175 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1178 if (nr_reads >= max_nr_reads)
1182 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1183 tg_may_dispatch(tg, bio, NULL)) {
1185 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1188 if (nr_writes >= max_nr_writes)
1192 return nr_reads + nr_writes;
1195 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1197 unsigned int nr_disp = 0;
1200 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1201 struct throtl_service_queue *sq = &tg->service_queue;
1206 if (time_before(jiffies, tg->disptime))
1209 throtl_dequeue_tg(tg);
1211 nr_disp += throtl_dispatch_tg(tg);
1213 if (sq->nr_queued[0] || sq->nr_queued[1])
1214 tg_update_disptime(tg);
1216 if (nr_disp >= throtl_quantum)
1223 static bool throtl_can_upgrade(struct throtl_data *td,
1224 struct throtl_grp *this_tg);
1226 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1227 * @arg: the throtl_service_queue being serviced
1229 * This timer is armed when a child throtl_grp with active bio's become
1230 * pending and queued on the service_queue's pending_tree and expires when
1231 * the first child throtl_grp should be dispatched. This function
1232 * dispatches bio's from the children throtl_grps to the parent
1235 * If the parent's parent is another throtl_grp, dispatching is propagated
1236 * by either arming its pending_timer or repeating dispatch directly. If
1237 * the top-level service_tree is reached, throtl_data->dispatch_work is
1238 * kicked so that the ready bio's are issued.
1240 static void throtl_pending_timer_fn(unsigned long arg)
1242 struct throtl_service_queue *sq = (void *)arg;
1243 struct throtl_grp *tg = sq_to_tg(sq);
1244 struct throtl_data *td = sq_to_td(sq);
1245 struct request_queue *q = td->queue;
1246 struct throtl_service_queue *parent_sq;
1250 spin_lock_irq(q->queue_lock);
1251 if (throtl_can_upgrade(td, NULL))
1252 throtl_upgrade_state(td);
1255 parent_sq = sq->parent_sq;
1259 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1260 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1261 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1263 ret = throtl_select_dispatch(sq);
1265 throtl_log(sq, "bios disp=%u", ret);
1269 if (throtl_schedule_next_dispatch(sq, false))
1272 /* this dispatch windows is still open, relax and repeat */
1273 spin_unlock_irq(q->queue_lock);
1275 spin_lock_irq(q->queue_lock);
1282 /* @parent_sq is another throl_grp, propagate dispatch */
1283 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1284 tg_update_disptime(tg);
1285 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1286 /* window is already open, repeat dispatching */
1293 /* reached the top-level, queue issueing */
1294 queue_work(kthrotld_workqueue, &td->dispatch_work);
1297 spin_unlock_irq(q->queue_lock);
1301 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1302 * @work: work item being executed
1304 * This function is queued for execution when bio's reach the bio_lists[]
1305 * of throtl_data->service_queue. Those bio's are ready and issued by this
1308 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1310 struct throtl_data *td = container_of(work, struct throtl_data,
1312 struct throtl_service_queue *td_sq = &td->service_queue;
1313 struct request_queue *q = td->queue;
1314 struct bio_list bio_list_on_stack;
1316 struct blk_plug plug;
1319 bio_list_init(&bio_list_on_stack);
1321 spin_lock_irq(q->queue_lock);
1322 for (rw = READ; rw <= WRITE; rw++)
1323 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1324 bio_list_add(&bio_list_on_stack, bio);
1325 spin_unlock_irq(q->queue_lock);
1327 if (!bio_list_empty(&bio_list_on_stack)) {
1328 blk_start_plug(&plug);
1329 while((bio = bio_list_pop(&bio_list_on_stack)))
1330 generic_make_request(bio);
1331 blk_finish_plug(&plug);
1335 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1338 struct throtl_grp *tg = pd_to_tg(pd);
1339 u64 v = *(u64 *)((void *)tg + off);
1343 return __blkg_prfill_u64(sf, pd, v);
1346 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1349 struct throtl_grp *tg = pd_to_tg(pd);
1350 unsigned int v = *(unsigned int *)((void *)tg + off);
1354 return __blkg_prfill_u64(sf, pd, v);
1357 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1359 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1360 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1364 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1366 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1367 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1371 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1373 struct throtl_service_queue *sq = &tg->service_queue;
1374 struct cgroup_subsys_state *pos_css;
1375 struct blkcg_gq *blkg;
1377 throtl_log(&tg->service_queue,
1378 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1379 tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1380 tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1383 * Update has_rules[] flags for the updated tg's subtree. A tg is
1384 * considered to have rules if either the tg itself or any of its
1385 * ancestors has rules. This identifies groups without any
1386 * restrictions in the whole hierarchy and allows them to bypass
1389 blkg_for_each_descendant_pre(blkg, pos_css,
1390 global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1391 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1392 struct throtl_grp *parent_tg;
1394 tg_update_has_rules(this_tg);
1395 /* ignore root/second level */
1396 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1397 !blkg->parent->parent)
1399 parent_tg = blkg_to_tg(blkg->parent);
1401 * make sure all children has lower idle time threshold and
1402 * higher latency target
1404 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1405 parent_tg->idletime_threshold);
1406 this_tg->latency_target = max(this_tg->latency_target,
1407 parent_tg->latency_target);
1411 * We're already holding queue_lock and know @tg is valid. Let's
1412 * apply the new config directly.
1414 * Restart the slices for both READ and WRITES. It might happen
1415 * that a group's limit are dropped suddenly and we don't want to
1416 * account recently dispatched IO with new low rate.
1418 throtl_start_new_slice(tg, 0);
1419 throtl_start_new_slice(tg, 1);
1421 if (tg->flags & THROTL_TG_PENDING) {
1422 tg_update_disptime(tg);
1423 throtl_schedule_next_dispatch(sq->parent_sq, true);
1427 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1428 char *buf, size_t nbytes, loff_t off, bool is_u64)
1430 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1431 struct blkg_conf_ctx ctx;
1432 struct throtl_grp *tg;
1436 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1441 if (sscanf(ctx.body, "%llu", &v) != 1)
1446 tg = blkg_to_tg(ctx.blkg);
1449 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1451 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1453 tg_conf_updated(tg, false);
1456 blkg_conf_finish(&ctx);
1457 return ret ?: nbytes;
1460 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1461 char *buf, size_t nbytes, loff_t off)
1463 return tg_set_conf(of, buf, nbytes, off, true);
1466 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1467 char *buf, size_t nbytes, loff_t off)
1469 return tg_set_conf(of, buf, nbytes, off, false);
1472 static struct cftype throtl_legacy_files[] = {
1474 .name = "throttle.read_bps_device",
1475 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1476 .seq_show = tg_print_conf_u64,
1477 .write = tg_set_conf_u64,
1480 .name = "throttle.write_bps_device",
1481 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1482 .seq_show = tg_print_conf_u64,
1483 .write = tg_set_conf_u64,
1486 .name = "throttle.read_iops_device",
1487 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1488 .seq_show = tg_print_conf_uint,
1489 .write = tg_set_conf_uint,
1492 .name = "throttle.write_iops_device",
1493 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1494 .seq_show = tg_print_conf_uint,
1495 .write = tg_set_conf_uint,
1498 .name = "throttle.io_service_bytes",
1499 .private = (unsigned long)&blkcg_policy_throtl,
1500 .seq_show = blkg_print_stat_bytes,
1503 .name = "throttle.io_serviced",
1504 .private = (unsigned long)&blkcg_policy_throtl,
1505 .seq_show = blkg_print_stat_ios,
1510 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1513 struct throtl_grp *tg = pd_to_tg(pd);
1514 const char *dname = blkg_dev_name(pd->blkg);
1515 char bufs[4][21] = { "max", "max", "max", "max" };
1517 unsigned int iops_dft;
1518 char idle_time[26] = "";
1519 char latency_time[26] = "";
1524 if (off == LIMIT_LOW) {
1529 iops_dft = UINT_MAX;
1532 if (tg->bps_conf[READ][off] == bps_dft &&
1533 tg->bps_conf[WRITE][off] == bps_dft &&
1534 tg->iops_conf[READ][off] == iops_dft &&
1535 tg->iops_conf[WRITE][off] == iops_dft &&
1536 (off != LIMIT_LOW ||
1537 (tg->idletime_threshold_conf == tg->td->dft_idletime_threshold &&
1538 tg->latency_target_conf == DFL_LATENCY_TARGET)))
1541 if (tg->bps_conf[READ][off] != U64_MAX)
1542 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1543 tg->bps_conf[READ][off]);
1544 if (tg->bps_conf[WRITE][off] != U64_MAX)
1545 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1546 tg->bps_conf[WRITE][off]);
1547 if (tg->iops_conf[READ][off] != UINT_MAX)
1548 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1549 tg->iops_conf[READ][off]);
1550 if (tg->iops_conf[WRITE][off] != UINT_MAX)
1551 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1552 tg->iops_conf[WRITE][off]);
1553 if (off == LIMIT_LOW) {
1554 if (tg->idletime_threshold_conf == ULONG_MAX)
1555 strcpy(idle_time, " idle=max");
1557 snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1558 tg->idletime_threshold_conf);
1560 if (tg->latency_target_conf == ULONG_MAX)
1561 strcpy(latency_time, " latency=max");
1563 snprintf(latency_time, sizeof(latency_time),
1564 " latency=%lu", tg->latency_target_conf);
1567 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1568 dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1573 static int tg_print_limit(struct seq_file *sf, void *v)
1575 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1576 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1580 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1581 char *buf, size_t nbytes, loff_t off)
1583 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1584 struct blkg_conf_ctx ctx;
1585 struct throtl_grp *tg;
1587 unsigned long idle_time;
1588 unsigned long latency_time;
1590 int index = of_cft(of)->private;
1592 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1596 tg = blkg_to_tg(ctx.blkg);
1598 v[0] = tg->bps_conf[READ][index];
1599 v[1] = tg->bps_conf[WRITE][index];
1600 v[2] = tg->iops_conf[READ][index];
1601 v[3] = tg->iops_conf[WRITE][index];
1603 idle_time = tg->idletime_threshold_conf;
1604 latency_time = tg->latency_target_conf;
1606 char tok[27]; /* wiops=18446744073709551616 */
1611 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1620 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1628 if (!strcmp(tok, "rbps"))
1630 else if (!strcmp(tok, "wbps"))
1632 else if (!strcmp(tok, "riops"))
1633 v[2] = min_t(u64, val, UINT_MAX);
1634 else if (!strcmp(tok, "wiops"))
1635 v[3] = min_t(u64, val, UINT_MAX);
1636 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1638 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1644 tg->bps_conf[READ][index] = v[0];
1645 tg->bps_conf[WRITE][index] = v[1];
1646 tg->iops_conf[READ][index] = v[2];
1647 tg->iops_conf[WRITE][index] = v[3];
1649 if (index == LIMIT_MAX) {
1650 tg->bps[READ][index] = v[0];
1651 tg->bps[WRITE][index] = v[1];
1652 tg->iops[READ][index] = v[2];
1653 tg->iops[WRITE][index] = v[3];
1655 tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1656 tg->bps_conf[READ][LIMIT_MAX]);
1657 tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1658 tg->bps_conf[WRITE][LIMIT_MAX]);
1659 tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1660 tg->iops_conf[READ][LIMIT_MAX]);
1661 tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1662 tg->iops_conf[WRITE][LIMIT_MAX]);
1664 if (index == LIMIT_LOW) {
1665 blk_throtl_update_limit_valid(tg->td);
1666 if (tg->td->limit_valid[LIMIT_LOW])
1667 tg->td->limit_index = LIMIT_LOW;
1668 tg->idletime_threshold_conf = idle_time;
1669 tg->idletime_threshold = tg->idletime_threshold_conf;
1670 tg->latency_target_conf = latency_time;
1671 tg->latency_target = tg->latency_target_conf;
1673 tg_conf_updated(tg, index == LIMIT_LOW &&
1674 tg->td->limit_valid[LIMIT_LOW]);
1677 blkg_conf_finish(&ctx);
1678 return ret ?: nbytes;
1681 static struct cftype throtl_files[] = {
1682 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1685 .flags = CFTYPE_NOT_ON_ROOT,
1686 .seq_show = tg_print_limit,
1687 .write = tg_set_limit,
1688 .private = LIMIT_LOW,
1693 .flags = CFTYPE_NOT_ON_ROOT,
1694 .seq_show = tg_print_limit,
1695 .write = tg_set_limit,
1696 .private = LIMIT_MAX,
1701 static void throtl_shutdown_wq(struct request_queue *q)
1703 struct throtl_data *td = q->td;
1705 cancel_work_sync(&td->dispatch_work);
1708 static struct blkcg_policy blkcg_policy_throtl = {
1709 .dfl_cftypes = throtl_files,
1710 .legacy_cftypes = throtl_legacy_files,
1712 .pd_alloc_fn = throtl_pd_alloc,
1713 .pd_init_fn = throtl_pd_init,
1714 .pd_online_fn = throtl_pd_online,
1715 .pd_offline_fn = throtl_pd_offline,
1716 .pd_free_fn = throtl_pd_free,
1719 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1721 unsigned long rtime = jiffies, wtime = jiffies;
1723 if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1724 rtime = tg->last_low_overflow_time[READ];
1725 if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1726 wtime = tg->last_low_overflow_time[WRITE];
1727 return min(rtime, wtime);
1730 /* tg should not be an intermediate node */
1731 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1733 struct throtl_service_queue *parent_sq;
1734 struct throtl_grp *parent = tg;
1735 unsigned long ret = __tg_last_low_overflow_time(tg);
1738 parent_sq = parent->service_queue.parent_sq;
1739 parent = sq_to_tg(parent_sq);
1744 * The parent doesn't have low limit, it always reaches low
1745 * limit. Its overflow time is useless for children
1747 if (!parent->bps[READ][LIMIT_LOW] &&
1748 !parent->iops[READ][LIMIT_LOW] &&
1749 !parent->bps[WRITE][LIMIT_LOW] &&
1750 !parent->iops[WRITE][LIMIT_LOW])
1752 if (time_after(__tg_last_low_overflow_time(parent), ret))
1753 ret = __tg_last_low_overflow_time(parent);
1758 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1761 * cgroup is idle if:
1762 * - single idle is too long, longer than a fixed value (in case user
1763 * configure a too big threshold) or 4 times of slice
1764 * - average think time is more than threshold
1765 * - IO latency is largely below threshold
1767 unsigned long time = jiffies_to_usecs(4 * tg->td->throtl_slice);
1770 time = min_t(unsigned long, MAX_IDLE_TIME, time);
1771 ret = (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1772 tg->avg_idletime > tg->idletime_threshold ||
1773 (tg->latency_target && tg->bio_cnt &&
1774 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1775 throtl_log(&tg->service_queue,
1776 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1777 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1778 tg->bio_cnt, ret, tg->td->scale);
1782 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1784 struct throtl_service_queue *sq = &tg->service_queue;
1785 bool read_limit, write_limit;
1788 * if cgroup reaches low limit (if low limit is 0, the cgroup always
1789 * reaches), it's ok to upgrade to next limit
1791 read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1792 write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1793 if (!read_limit && !write_limit)
1795 if (read_limit && sq->nr_queued[READ] &&
1796 (!write_limit || sq->nr_queued[WRITE]))
1798 if (write_limit && sq->nr_queued[WRITE] &&
1799 (!read_limit || sq->nr_queued[READ]))
1802 if (time_after_eq(jiffies,
1803 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1804 throtl_tg_is_idle(tg))
1809 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1812 if (throtl_tg_can_upgrade(tg))
1814 tg = sq_to_tg(tg->service_queue.parent_sq);
1815 if (!tg || !tg_to_blkg(tg)->parent)
1821 static bool throtl_can_upgrade(struct throtl_data *td,
1822 struct throtl_grp *this_tg)
1824 struct cgroup_subsys_state *pos_css;
1825 struct blkcg_gq *blkg;
1827 if (td->limit_index != LIMIT_LOW)
1830 if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1834 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1835 struct throtl_grp *tg = blkg_to_tg(blkg);
1839 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1841 if (!throtl_hierarchy_can_upgrade(tg)) {
1850 static void throtl_upgrade_check(struct throtl_grp *tg)
1852 unsigned long now = jiffies;
1854 if (tg->td->limit_index != LIMIT_LOW)
1857 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1860 tg->last_check_time = now;
1862 if (!time_after_eq(now,
1863 __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1866 if (throtl_can_upgrade(tg->td, NULL))
1867 throtl_upgrade_state(tg->td);
1870 static void throtl_upgrade_state(struct throtl_data *td)
1872 struct cgroup_subsys_state *pos_css;
1873 struct blkcg_gq *blkg;
1875 throtl_log(&td->service_queue, "upgrade to max");
1876 td->limit_index = LIMIT_MAX;
1877 td->low_upgrade_time = jiffies;
1880 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1881 struct throtl_grp *tg = blkg_to_tg(blkg);
1882 struct throtl_service_queue *sq = &tg->service_queue;
1884 tg->disptime = jiffies - 1;
1885 throtl_select_dispatch(sq);
1886 throtl_schedule_next_dispatch(sq, false);
1889 throtl_select_dispatch(&td->service_queue);
1890 throtl_schedule_next_dispatch(&td->service_queue, false);
1891 queue_work(kthrotld_workqueue, &td->dispatch_work);
1894 static void throtl_downgrade_state(struct throtl_data *td, int new)
1898 throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1900 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1904 td->limit_index = new;
1905 td->low_downgrade_time = jiffies;
1908 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1910 struct throtl_data *td = tg->td;
1911 unsigned long now = jiffies;
1914 * If cgroup is below low limit, consider downgrade and throttle other
1917 if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1918 time_after_eq(now, tg_last_low_overflow_time(tg) +
1919 td->throtl_slice) &&
1920 (!throtl_tg_is_idle(tg) ||
1921 !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1926 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1929 if (!throtl_tg_can_downgrade(tg))
1931 tg = sq_to_tg(tg->service_queue.parent_sq);
1932 if (!tg || !tg_to_blkg(tg)->parent)
1938 static void throtl_downgrade_check(struct throtl_grp *tg)
1942 unsigned long elapsed_time;
1943 unsigned long now = jiffies;
1945 if (tg->td->limit_index != LIMIT_MAX ||
1946 !tg->td->limit_valid[LIMIT_LOW])
1948 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1950 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1953 elapsed_time = now - tg->last_check_time;
1954 tg->last_check_time = now;
1956 if (time_before(now, tg_last_low_overflow_time(tg) +
1957 tg->td->throtl_slice))
1960 if (tg->bps[READ][LIMIT_LOW]) {
1961 bps = tg->last_bytes_disp[READ] * HZ;
1962 do_div(bps, elapsed_time);
1963 if (bps >= tg->bps[READ][LIMIT_LOW])
1964 tg->last_low_overflow_time[READ] = now;
1967 if (tg->bps[WRITE][LIMIT_LOW]) {
1968 bps = tg->last_bytes_disp[WRITE] * HZ;
1969 do_div(bps, elapsed_time);
1970 if (bps >= tg->bps[WRITE][LIMIT_LOW])
1971 tg->last_low_overflow_time[WRITE] = now;
1974 if (tg->iops[READ][LIMIT_LOW]) {
1975 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
1976 if (iops >= tg->iops[READ][LIMIT_LOW])
1977 tg->last_low_overflow_time[READ] = now;
1980 if (tg->iops[WRITE][LIMIT_LOW]) {
1981 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
1982 if (iops >= tg->iops[WRITE][LIMIT_LOW])
1983 tg->last_low_overflow_time[WRITE] = now;
1987 * If cgroup is below low limit, consider downgrade and throttle other
1990 if (throtl_hierarchy_can_downgrade(tg))
1991 throtl_downgrade_state(tg->td, LIMIT_LOW);
1993 tg->last_bytes_disp[READ] = 0;
1994 tg->last_bytes_disp[WRITE] = 0;
1995 tg->last_io_disp[READ] = 0;
1996 tg->last_io_disp[WRITE] = 0;
1999 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2001 unsigned long now = ktime_get_ns() >> 10;
2002 unsigned long last_finish_time = tg->last_finish_time;
2004 if (now <= last_finish_time || last_finish_time == 0 ||
2005 last_finish_time == tg->checked_last_finish_time)
2008 tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2009 tg->checked_last_finish_time = last_finish_time;
2012 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2013 static void throtl_update_latency_buckets(struct throtl_data *td)
2015 struct avg_latency_bucket avg_latency[LATENCY_BUCKET_SIZE];
2017 unsigned long last_latency = 0;
2018 unsigned long latency;
2020 if (!blk_queue_nonrot(td->queue))
2022 if (time_before(jiffies, td->last_calculate_time + HZ))
2024 td->last_calculate_time = jiffies;
2026 memset(avg_latency, 0, sizeof(avg_latency));
2027 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2028 struct latency_bucket *tmp = &td->tmp_buckets[i];
2030 for_each_possible_cpu(cpu) {
2031 struct latency_bucket *bucket;
2033 /* this isn't race free, but ok in practice */
2034 bucket = per_cpu_ptr(td->latency_buckets, cpu);
2035 tmp->total_latency += bucket[i].total_latency;
2036 tmp->samples += bucket[i].samples;
2037 bucket[i].total_latency = 0;
2038 bucket[i].samples = 0;
2041 if (tmp->samples >= 32) {
2042 int samples = tmp->samples;
2044 latency = tmp->total_latency;
2046 tmp->total_latency = 0;
2051 avg_latency[i].latency = latency;
2055 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2056 if (!avg_latency[i].latency) {
2057 if (td->avg_buckets[i].latency < last_latency)
2058 td->avg_buckets[i].latency = last_latency;
2062 if (!td->avg_buckets[i].valid)
2063 latency = avg_latency[i].latency;
2065 latency = (td->avg_buckets[i].latency * 7 +
2066 avg_latency[i].latency) >> 3;
2068 td->avg_buckets[i].latency = max(latency, last_latency);
2069 td->avg_buckets[i].valid = true;
2070 last_latency = td->avg_buckets[i].latency;
2073 for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2074 throtl_log(&td->service_queue,
2075 "Latency bucket %d: latency=%ld, valid=%d", i,
2076 td->avg_buckets[i].latency, td->avg_buckets[i].valid);
2079 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2084 static void blk_throtl_assoc_bio(struct throtl_grp *tg, struct bio *bio)
2086 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2089 ret = bio_associate_current(bio);
2090 if (ret == 0 || ret == -EBUSY)
2091 bio->bi_cg_private = tg;
2092 blk_stat_set_issue(&bio->bi_issue_stat, bio_sectors(bio));
2094 bio_associate_current(bio);
2098 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
2101 struct throtl_qnode *qn = NULL;
2102 struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
2103 struct throtl_service_queue *sq;
2104 bool rw = bio_data_dir(bio);
2105 bool throttled = false;
2106 struct throtl_data *td = tg->td;
2108 WARN_ON_ONCE(!rcu_read_lock_held());
2110 /* see throtl_charge_bio() */
2111 if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw])
2114 spin_lock_irq(q->queue_lock);
2116 throtl_update_latency_buckets(td);
2118 if (unlikely(blk_queue_bypass(q)))
2121 blk_throtl_assoc_bio(tg, bio);
2122 blk_throtl_update_idletime(tg);
2124 sq = &tg->service_queue;
2128 if (tg->last_low_overflow_time[rw] == 0)
2129 tg->last_low_overflow_time[rw] = jiffies;
2130 throtl_downgrade_check(tg);
2131 throtl_upgrade_check(tg);
2132 /* throtl is FIFO - if bios are already queued, should queue */
2133 if (sq->nr_queued[rw])
2136 /* if above limits, break to queue */
2137 if (!tg_may_dispatch(tg, bio, NULL)) {
2138 tg->last_low_overflow_time[rw] = jiffies;
2139 if (throtl_can_upgrade(td, tg)) {
2140 throtl_upgrade_state(td);
2146 /* within limits, let's charge and dispatch directly */
2147 throtl_charge_bio(tg, bio);
2150 * We need to trim slice even when bios are not being queued
2151 * otherwise it might happen that a bio is not queued for
2152 * a long time and slice keeps on extending and trim is not
2153 * called for a long time. Now if limits are reduced suddenly
2154 * we take into account all the IO dispatched so far at new
2155 * low rate and * newly queued IO gets a really long dispatch
2158 * So keep on trimming slice even if bio is not queued.
2160 throtl_trim_slice(tg, rw);
2163 * @bio passed through this layer without being throttled.
2164 * Climb up the ladder. If we''re already at the top, it
2165 * can be executed directly.
2167 qn = &tg->qnode_on_parent[rw];
2174 /* out-of-limit, queue to @tg */
2175 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2176 rw == READ ? 'R' : 'W',
2177 tg->bytes_disp[rw], bio->bi_iter.bi_size,
2178 tg_bps_limit(tg, rw),
2179 tg->io_disp[rw], tg_iops_limit(tg, rw),
2180 sq->nr_queued[READ], sq->nr_queued[WRITE]);
2182 tg->last_low_overflow_time[rw] = jiffies;
2184 td->nr_queued[rw]++;
2185 throtl_add_bio_tg(bio, qn, tg);
2189 * Update @tg's dispatch time and force schedule dispatch if @tg
2190 * was empty before @bio. The forced scheduling isn't likely to
2191 * cause undue delay as @bio is likely to be dispatched directly if
2192 * its @tg's disptime is not in the future.
2194 if (tg->flags & THROTL_TG_WAS_EMPTY) {
2195 tg_update_disptime(tg);
2196 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2200 spin_unlock_irq(q->queue_lock);
2203 * As multiple blk-throtls may stack in the same issue path, we
2204 * don't want bios to leave with the flag set. Clear the flag if
2208 bio_clear_flag(bio, BIO_THROTTLED);
2210 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2211 if (throttled || !td->track_bio_latency)
2212 bio->bi_issue_stat.stat |= SKIP_LATENCY;
2217 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2218 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2219 int op, unsigned long time)
2221 struct latency_bucket *latency;
2224 if (!td || td->limit_index != LIMIT_LOW || op != REQ_OP_READ ||
2225 !blk_queue_nonrot(td->queue))
2228 index = request_bucket_index(size);
2230 latency = get_cpu_ptr(td->latency_buckets);
2231 latency[index].total_latency += time;
2232 latency[index].samples++;
2233 put_cpu_ptr(td->latency_buckets);
2236 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2238 struct request_queue *q = rq->q;
2239 struct throtl_data *td = q->td;
2241 throtl_track_latency(td, blk_stat_size(&rq->issue_stat),
2242 req_op(rq), time_ns >> 10);
2245 void blk_throtl_bio_endio(struct bio *bio)
2247 struct throtl_grp *tg;
2249 unsigned long finish_time;
2250 unsigned long start_time;
2253 tg = bio->bi_cg_private;
2256 bio->bi_cg_private = NULL;
2258 finish_time_ns = ktime_get_ns();
2259 tg->last_finish_time = finish_time_ns >> 10;
2261 start_time = blk_stat_time(&bio->bi_issue_stat) >> 10;
2262 finish_time = __blk_stat_time(finish_time_ns) >> 10;
2263 if (!start_time || finish_time <= start_time)
2266 lat = finish_time - start_time;
2267 /* this is only for bio based driver */
2268 if (!(bio->bi_issue_stat.stat & SKIP_LATENCY))
2269 throtl_track_latency(tg->td, blk_stat_size(&bio->bi_issue_stat),
2272 if (tg->latency_target) {
2274 unsigned int threshold;
2276 bucket = request_bucket_index(
2277 blk_stat_size(&bio->bi_issue_stat));
2278 threshold = tg->td->avg_buckets[bucket].latency +
2280 if (lat > threshold)
2283 * Not race free, could get wrong count, which means cgroups
2289 if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2290 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2292 tg->bad_bio_cnt /= 2;
2298 * Dispatch all bios from all children tg's queued on @parent_sq. On
2299 * return, @parent_sq is guaranteed to not have any active children tg's
2300 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
2302 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
2304 struct throtl_grp *tg;
2306 while ((tg = throtl_rb_first(parent_sq))) {
2307 struct throtl_service_queue *sq = &tg->service_queue;
2310 throtl_dequeue_tg(tg);
2312 while ((bio = throtl_peek_queued(&sq->queued[READ])))
2313 tg_dispatch_one_bio(tg, bio_data_dir(bio));
2314 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
2315 tg_dispatch_one_bio(tg, bio_data_dir(bio));
2320 * blk_throtl_drain - drain throttled bios
2321 * @q: request_queue to drain throttled bios for
2323 * Dispatch all currently throttled bios on @q through ->make_request_fn().
2325 void blk_throtl_drain(struct request_queue *q)
2326 __releases(q->queue_lock) __acquires(q->queue_lock)
2328 struct throtl_data *td = q->td;
2329 struct blkcg_gq *blkg;
2330 struct cgroup_subsys_state *pos_css;
2334 queue_lockdep_assert_held(q);
2338 * Drain each tg while doing post-order walk on the blkg tree, so
2339 * that all bios are propagated to td->service_queue. It'd be
2340 * better to walk service_queue tree directly but blkg walk is
2343 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
2344 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
2346 /* finally, transfer bios from top-level tg's into the td */
2347 tg_drain_bios(&td->service_queue);
2350 spin_unlock_irq(q->queue_lock);
2352 /* all bios now should be in td->service_queue, issue them */
2353 for (rw = READ; rw <= WRITE; rw++)
2354 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
2356 generic_make_request(bio);
2358 spin_lock_irq(q->queue_lock);
2361 int blk_throtl_init(struct request_queue *q)
2363 struct throtl_data *td;
2366 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2369 td->latency_buckets = __alloc_percpu(sizeof(struct latency_bucket) *
2370 LATENCY_BUCKET_SIZE, __alignof__(u64));
2371 if (!td->latency_buckets) {
2376 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2377 throtl_service_queue_init(&td->service_queue);
2382 td->limit_valid[LIMIT_MAX] = true;
2383 td->limit_index = LIMIT_MAX;
2384 td->low_upgrade_time = jiffies;
2385 td->low_downgrade_time = jiffies;
2387 /* activate policy */
2388 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2390 free_percpu(td->latency_buckets);
2396 void blk_throtl_exit(struct request_queue *q)
2399 throtl_shutdown_wq(q);
2400 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2401 free_percpu(q->td->latency_buckets);
2405 void blk_throtl_register_queue(struct request_queue *q)
2407 struct throtl_data *td;
2408 struct cgroup_subsys_state *pos_css;
2409 struct blkcg_gq *blkg;
2414 if (blk_queue_nonrot(q)) {
2415 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2416 td->dft_idletime_threshold = DFL_IDLE_THRESHOLD_SSD;
2418 td->throtl_slice = DFL_THROTL_SLICE_HD;
2419 td->dft_idletime_threshold = DFL_IDLE_THRESHOLD_HD;
2421 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2422 /* if no low limit, use previous default */
2423 td->throtl_slice = DFL_THROTL_SLICE_HD;
2426 td->track_bio_latency = !q->mq_ops && !q->request_fn;
2427 if (!td->track_bio_latency)
2428 blk_stat_enable_accounting(q);
2431 * some tg are created before queue is fully initialized, eg, nonrot
2432 * isn't initialized yet
2435 blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) {
2436 struct throtl_grp *tg = blkg_to_tg(blkg);
2438 tg->idletime_threshold = td->dft_idletime_threshold;
2439 tg->idletime_threshold_conf = td->dft_idletime_threshold;
2444 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2445 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2449 return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2452 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2453 const char *page, size_t count)
2460 if (kstrtoul(page, 10, &v))
2462 t = msecs_to_jiffies(v);
2463 if (t == 0 || t > MAX_THROTL_SLICE)
2465 q->td->throtl_slice = t;
2470 static int __init throtl_init(void)
2472 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2473 if (!kthrotld_workqueue)
2474 panic("Failed to create kthrotld\n");
2476 return blkcg_policy_register(&blkcg_policy_throtl);
2479 module_init(throtl_init);