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 /* When did we start a new slice */
161 unsigned long slice_start[2];
162 unsigned long slice_end[2];
164 unsigned long last_finish_time; /* ns / 1024 */
165 unsigned long checked_last_finish_time; /* ns / 1024 */
166 unsigned long avg_idletime; /* ns / 1024 */
167 unsigned long idletime_threshold; /* us */
170 /* We measure latency for request size from <= 4k to >= 1M */
171 #define LATENCY_BUCKET_SIZE 9
173 struct latency_bucket {
174 unsigned long total_latency; /* ns / 1024 */
178 struct avg_latency_bucket {
179 unsigned long latency; /* ns / 1024 */
185 /* service tree for active throtl groups */
186 struct throtl_service_queue service_queue;
188 struct request_queue *queue;
190 /* Total Number of queued bios on READ and WRITE lists */
191 unsigned int nr_queued[2];
193 unsigned int throtl_slice;
195 /* Work for dispatching throttled bios */
196 struct work_struct dispatch_work;
197 unsigned int limit_index;
198 bool limit_valid[LIMIT_CNT];
200 unsigned long dft_idletime_threshold; /* us */
202 unsigned long low_upgrade_time;
203 unsigned long low_downgrade_time;
207 struct latency_bucket tmp_buckets[LATENCY_BUCKET_SIZE];
208 struct avg_latency_bucket avg_buckets[LATENCY_BUCKET_SIZE];
209 struct latency_bucket __percpu *latency_buckets;
210 unsigned long last_calculate_time;
212 bool track_bio_latency;
215 static void throtl_pending_timer_fn(unsigned long arg);
217 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
219 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
222 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
224 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
227 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
229 return pd_to_blkg(&tg->pd);
233 * sq_to_tg - return the throl_grp the specified service queue belongs to
234 * @sq: the throtl_service_queue of interest
236 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
237 * embedded in throtl_data, %NULL is returned.
239 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
241 if (sq && sq->parent_sq)
242 return container_of(sq, struct throtl_grp, service_queue);
248 * sq_to_td - return throtl_data the specified service queue belongs to
249 * @sq: the throtl_service_queue of interest
251 * A service_queue can be embedded in either a throtl_grp or throtl_data.
252 * Determine the associated throtl_data accordingly and return it.
254 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
256 struct throtl_grp *tg = sq_to_tg(sq);
261 return container_of(sq, struct throtl_data, service_queue);
265 * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
266 * make the IO dispatch more smooth.
267 * Scale up: linearly scale up according to lapsed time since upgrade. For
268 * every throtl_slice, the limit scales up 1/2 .low limit till the
269 * limit hits .max limit
270 * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
272 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
274 /* arbitrary value to avoid too big scale */
275 if (td->scale < 4096 && time_after_eq(jiffies,
276 td->low_upgrade_time + td->scale * td->throtl_slice))
277 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
279 return low + (low >> 1) * td->scale;
282 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
284 struct blkcg_gq *blkg = tg_to_blkg(tg);
285 struct throtl_data *td;
288 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
292 ret = tg->bps[rw][td->limit_index];
293 if (ret == 0 && td->limit_index == LIMIT_LOW)
294 return tg->bps[rw][LIMIT_MAX];
296 if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
297 tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
300 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
301 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
306 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
308 struct blkcg_gq *blkg = tg_to_blkg(tg);
309 struct throtl_data *td;
312 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
315 ret = tg->iops[rw][td->limit_index];
316 if (ret == 0 && tg->td->limit_index == LIMIT_LOW)
317 return tg->iops[rw][LIMIT_MAX];
319 if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
320 tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
323 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
324 if (adjusted > UINT_MAX)
326 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
331 #define request_bucket_index(sectors) \
332 clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
335 * throtl_log - log debug message via blktrace
336 * @sq: the service_queue being reported
337 * @fmt: printf format string
340 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
341 * throtl_grp; otherwise, just "throtl".
343 #define throtl_log(sq, fmt, args...) do { \
344 struct throtl_grp *__tg = sq_to_tg((sq)); \
345 struct throtl_data *__td = sq_to_td((sq)); \
348 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
353 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
354 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
356 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
360 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
362 INIT_LIST_HEAD(&qn->node);
363 bio_list_init(&qn->bios);
368 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
369 * @bio: bio being added
370 * @qn: qnode to add bio to
371 * @queued: the service_queue->queued[] list @qn belongs to
373 * Add @bio to @qn and put @qn on @queued if it's not already on.
374 * @qn->tg's reference count is bumped when @qn is activated. See the
375 * comment on top of throtl_qnode definition for details.
377 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
378 struct list_head *queued)
380 bio_list_add(&qn->bios, bio);
381 if (list_empty(&qn->node)) {
382 list_add_tail(&qn->node, queued);
383 blkg_get(tg_to_blkg(qn->tg));
388 * throtl_peek_queued - peek the first bio on a qnode list
389 * @queued: the qnode list to peek
391 static struct bio *throtl_peek_queued(struct list_head *queued)
393 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
396 if (list_empty(queued))
399 bio = bio_list_peek(&qn->bios);
405 * throtl_pop_queued - pop the first bio form a qnode list
406 * @queued: the qnode list to pop a bio from
407 * @tg_to_put: optional out argument for throtl_grp to put
409 * Pop the first bio from the qnode list @queued. After popping, the first
410 * qnode is removed from @queued if empty or moved to the end of @queued so
411 * that the popping order is round-robin.
413 * When the first qnode is removed, its associated throtl_grp should be put
414 * too. If @tg_to_put is NULL, this function automatically puts it;
415 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
416 * responsible for putting it.
418 static struct bio *throtl_pop_queued(struct list_head *queued,
419 struct throtl_grp **tg_to_put)
421 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
424 if (list_empty(queued))
427 bio = bio_list_pop(&qn->bios);
430 if (bio_list_empty(&qn->bios)) {
431 list_del_init(&qn->node);
435 blkg_put(tg_to_blkg(qn->tg));
437 list_move_tail(&qn->node, queued);
443 /* init a service_queue, assumes the caller zeroed it */
444 static void throtl_service_queue_init(struct throtl_service_queue *sq)
446 INIT_LIST_HEAD(&sq->queued[0]);
447 INIT_LIST_HEAD(&sq->queued[1]);
448 sq->pending_tree = RB_ROOT;
449 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
453 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
455 struct throtl_grp *tg;
458 tg = kzalloc_node(sizeof(*tg), gfp, node);
462 throtl_service_queue_init(&tg->service_queue);
464 for (rw = READ; rw <= WRITE; rw++) {
465 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
466 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
469 RB_CLEAR_NODE(&tg->rb_node);
470 tg->bps[READ][LIMIT_MAX] = U64_MAX;
471 tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
472 tg->iops[READ][LIMIT_MAX] = UINT_MAX;
473 tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
474 tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
475 tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
476 tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
477 tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
478 /* LIMIT_LOW will have default value 0 */
480 tg->latency_target = DFL_LATENCY_TARGET;
485 static void throtl_pd_init(struct blkg_policy_data *pd)
487 struct throtl_grp *tg = pd_to_tg(pd);
488 struct blkcg_gq *blkg = tg_to_blkg(tg);
489 struct throtl_data *td = blkg->q->td;
490 struct throtl_service_queue *sq = &tg->service_queue;
493 * If on the default hierarchy, we switch to properly hierarchical
494 * behavior where limits on a given throtl_grp are applied to the
495 * whole subtree rather than just the group itself. e.g. If 16M
496 * read_bps limit is set on the root group, the whole system can't
497 * exceed 16M for the device.
499 * If not on the default hierarchy, the broken flat hierarchy
500 * behavior is retained where all throtl_grps are treated as if
501 * they're all separate root groups right below throtl_data.
502 * Limits of a group don't interact with limits of other groups
503 * regardless of the position of the group in the hierarchy.
505 sq->parent_sq = &td->service_queue;
506 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
507 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
510 tg->idletime_threshold = td->dft_idletime_threshold;
514 * Set has_rules[] if @tg or any of its parents have limits configured.
515 * This doesn't require walking up to the top of the hierarchy as the
516 * parent's has_rules[] is guaranteed to be correct.
518 static void tg_update_has_rules(struct throtl_grp *tg)
520 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
521 struct throtl_data *td = tg->td;
524 for (rw = READ; rw <= WRITE; rw++)
525 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
526 (td->limit_valid[td->limit_index] &&
527 (tg_bps_limit(tg, rw) != U64_MAX ||
528 tg_iops_limit(tg, rw) != UINT_MAX));
531 static void throtl_pd_online(struct blkg_policy_data *pd)
533 struct throtl_grp *tg = pd_to_tg(pd);
535 * We don't want new groups to escape the limits of its ancestors.
536 * Update has_rules[] after a new group is brought online.
538 tg_update_has_rules(tg);
541 static void blk_throtl_update_limit_valid(struct throtl_data *td)
543 struct cgroup_subsys_state *pos_css;
544 struct blkcg_gq *blkg;
545 bool low_valid = false;
548 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
549 struct throtl_grp *tg = blkg_to_tg(blkg);
551 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
552 tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
557 td->limit_valid[LIMIT_LOW] = low_valid;
560 static void throtl_upgrade_state(struct throtl_data *td);
561 static void throtl_pd_offline(struct blkg_policy_data *pd)
563 struct throtl_grp *tg = pd_to_tg(pd);
565 tg->bps[READ][LIMIT_LOW] = 0;
566 tg->bps[WRITE][LIMIT_LOW] = 0;
567 tg->iops[READ][LIMIT_LOW] = 0;
568 tg->iops[WRITE][LIMIT_LOW] = 0;
570 blk_throtl_update_limit_valid(tg->td);
572 if (!tg->td->limit_valid[tg->td->limit_index])
573 throtl_upgrade_state(tg->td);
576 static void throtl_pd_free(struct blkg_policy_data *pd)
578 struct throtl_grp *tg = pd_to_tg(pd);
580 del_timer_sync(&tg->service_queue.pending_timer);
584 static struct throtl_grp *
585 throtl_rb_first(struct throtl_service_queue *parent_sq)
587 /* Service tree is empty */
588 if (!parent_sq->nr_pending)
591 if (!parent_sq->first_pending)
592 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
594 if (parent_sq->first_pending)
595 return rb_entry_tg(parent_sq->first_pending);
600 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
606 static void throtl_rb_erase(struct rb_node *n,
607 struct throtl_service_queue *parent_sq)
609 if (parent_sq->first_pending == n)
610 parent_sq->first_pending = NULL;
611 rb_erase_init(n, &parent_sq->pending_tree);
612 --parent_sq->nr_pending;
615 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
617 struct throtl_grp *tg;
619 tg = throtl_rb_first(parent_sq);
623 parent_sq->first_pending_disptime = tg->disptime;
626 static void tg_service_queue_add(struct throtl_grp *tg)
628 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
629 struct rb_node **node = &parent_sq->pending_tree.rb_node;
630 struct rb_node *parent = NULL;
631 struct throtl_grp *__tg;
632 unsigned long key = tg->disptime;
635 while (*node != NULL) {
637 __tg = rb_entry_tg(parent);
639 if (time_before(key, __tg->disptime))
640 node = &parent->rb_left;
642 node = &parent->rb_right;
648 parent_sq->first_pending = &tg->rb_node;
650 rb_link_node(&tg->rb_node, parent, node);
651 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
654 static void __throtl_enqueue_tg(struct throtl_grp *tg)
656 tg_service_queue_add(tg);
657 tg->flags |= THROTL_TG_PENDING;
658 tg->service_queue.parent_sq->nr_pending++;
661 static void throtl_enqueue_tg(struct throtl_grp *tg)
663 if (!(tg->flags & THROTL_TG_PENDING))
664 __throtl_enqueue_tg(tg);
667 static void __throtl_dequeue_tg(struct throtl_grp *tg)
669 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
670 tg->flags &= ~THROTL_TG_PENDING;
673 static void throtl_dequeue_tg(struct throtl_grp *tg)
675 if (tg->flags & THROTL_TG_PENDING)
676 __throtl_dequeue_tg(tg);
679 /* Call with queue lock held */
680 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
681 unsigned long expires)
683 unsigned long max_expire = jiffies + 8 * sq_to_tg(sq)->td->throtl_slice;
686 * Since we are adjusting the throttle limit dynamically, the sleep
687 * time calculated according to previous limit might be invalid. It's
688 * possible the cgroup sleep time is very long and no other cgroups
689 * have IO running so notify the limit changes. Make sure the cgroup
690 * doesn't sleep too long to avoid the missed notification.
692 if (time_after(expires, max_expire))
693 expires = max_expire;
694 mod_timer(&sq->pending_timer, expires);
695 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
696 expires - jiffies, jiffies);
700 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
701 * @sq: the service_queue to schedule dispatch for
702 * @force: force scheduling
704 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
705 * dispatch time of the first pending child. Returns %true if either timer
706 * is armed or there's no pending child left. %false if the current
707 * dispatch window is still open and the caller should continue
710 * If @force is %true, the dispatch timer is always scheduled and this
711 * function is guaranteed to return %true. This is to be used when the
712 * caller can't dispatch itself and needs to invoke pending_timer
713 * unconditionally. Note that forced scheduling is likely to induce short
714 * delay before dispatch starts even if @sq->first_pending_disptime is not
715 * in the future and thus shouldn't be used in hot paths.
717 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
720 /* any pending children left? */
724 update_min_dispatch_time(sq);
726 /* is the next dispatch time in the future? */
727 if (force || time_after(sq->first_pending_disptime, jiffies)) {
728 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
732 /* tell the caller to continue dispatching */
736 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
737 bool rw, unsigned long start)
739 tg->bytes_disp[rw] = 0;
743 * Previous slice has expired. We must have trimmed it after last
744 * bio dispatch. That means since start of last slice, we never used
745 * that bandwidth. Do try to make use of that bandwidth while giving
748 if (time_after_eq(start, tg->slice_start[rw]))
749 tg->slice_start[rw] = start;
751 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
752 throtl_log(&tg->service_queue,
753 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
754 rw == READ ? 'R' : 'W', tg->slice_start[rw],
755 tg->slice_end[rw], jiffies);
758 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
760 tg->bytes_disp[rw] = 0;
762 tg->slice_start[rw] = jiffies;
763 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
764 throtl_log(&tg->service_queue,
765 "[%c] new slice start=%lu end=%lu jiffies=%lu",
766 rw == READ ? 'R' : 'W', tg->slice_start[rw],
767 tg->slice_end[rw], jiffies);
770 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
771 unsigned long jiffy_end)
773 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
776 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
777 unsigned long jiffy_end)
779 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
780 throtl_log(&tg->service_queue,
781 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
782 rw == READ ? 'R' : 'W', tg->slice_start[rw],
783 tg->slice_end[rw], jiffies);
786 /* Determine if previously allocated or extended slice is complete or not */
787 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
789 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
795 /* Trim the used slices and adjust slice start accordingly */
796 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
798 unsigned long nr_slices, time_elapsed, io_trim;
801 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
804 * If bps are unlimited (-1), then time slice don't get
805 * renewed. Don't try to trim the slice if slice is used. A new
806 * slice will start when appropriate.
808 if (throtl_slice_used(tg, rw))
812 * A bio has been dispatched. Also adjust slice_end. It might happen
813 * that initially cgroup limit was very low resulting in high
814 * slice_end, but later limit was bumped up and bio was dispached
815 * sooner, then we need to reduce slice_end. A high bogus slice_end
816 * is bad because it does not allow new slice to start.
819 throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
821 time_elapsed = jiffies - tg->slice_start[rw];
823 nr_slices = time_elapsed / tg->td->throtl_slice;
827 tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
831 io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
834 if (!bytes_trim && !io_trim)
837 if (tg->bytes_disp[rw] >= bytes_trim)
838 tg->bytes_disp[rw] -= bytes_trim;
840 tg->bytes_disp[rw] = 0;
842 if (tg->io_disp[rw] >= io_trim)
843 tg->io_disp[rw] -= io_trim;
847 tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
849 throtl_log(&tg->service_queue,
850 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
851 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
852 tg->slice_start[rw], tg->slice_end[rw], jiffies);
855 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
858 bool rw = bio_data_dir(bio);
859 unsigned int io_allowed;
860 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
863 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
865 /* Slice has just started. Consider one slice interval */
867 jiffy_elapsed_rnd = tg->td->throtl_slice;
869 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
872 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
873 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
874 * will allow dispatch after 1 second and after that slice should
878 tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd;
882 io_allowed = UINT_MAX;
886 if (tg->io_disp[rw] + 1 <= io_allowed) {
892 /* Calc approx time to dispatch */
893 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ) / tg_iops_limit(tg, rw) + 1;
895 if (jiffy_wait > jiffy_elapsed)
896 jiffy_wait = jiffy_wait - jiffy_elapsed;
905 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
908 bool rw = bio_data_dir(bio);
909 u64 bytes_allowed, extra_bytes, tmp;
910 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
912 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
914 /* Slice has just started. Consider one slice interval */
916 jiffy_elapsed_rnd = tg->td->throtl_slice;
918 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
920 tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd;
924 if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
930 /* Calc approx time to dispatch */
931 extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
932 jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw));
938 * This wait time is without taking into consideration the rounding
939 * up we did. Add that time also.
941 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
948 * Returns whether one can dispatch a bio or not. Also returns approx number
949 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
951 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
954 bool rw = bio_data_dir(bio);
955 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
958 * Currently whole state machine of group depends on first bio
959 * queued in the group bio list. So one should not be calling
960 * this function with a different bio if there are other bios
963 BUG_ON(tg->service_queue.nr_queued[rw] &&
964 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
966 /* If tg->bps = -1, then BW is unlimited */
967 if (tg_bps_limit(tg, rw) == U64_MAX &&
968 tg_iops_limit(tg, rw) == UINT_MAX) {
975 * If previous slice expired, start a new one otherwise renew/extend
976 * existing slice to make sure it is at least throtl_slice interval
977 * long since now. New slice is started only for empty throttle group.
978 * If there is queued bio, that means there should be an active
979 * slice and it should be extended instead.
981 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
982 throtl_start_new_slice(tg, rw);
984 if (time_before(tg->slice_end[rw],
985 jiffies + tg->td->throtl_slice))
986 throtl_extend_slice(tg, rw,
987 jiffies + tg->td->throtl_slice);
990 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
991 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
997 max_wait = max(bps_wait, iops_wait);
1002 if (time_before(tg->slice_end[rw], jiffies + max_wait))
1003 throtl_extend_slice(tg, rw, jiffies + max_wait);
1008 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1010 bool rw = bio_data_dir(bio);
1012 /* Charge the bio to the group */
1013 tg->bytes_disp[rw] += bio->bi_iter.bi_size;
1015 tg->last_bytes_disp[rw] += bio->bi_iter.bi_size;
1016 tg->last_io_disp[rw]++;
1019 * BIO_THROTTLED is used to prevent the same bio to be throttled
1020 * more than once as a throttled bio will go through blk-throtl the
1021 * second time when it eventually gets issued. Set it when a bio
1022 * is being charged to a tg.
1024 if (!bio_flagged(bio, BIO_THROTTLED))
1025 bio_set_flag(bio, BIO_THROTTLED);
1029 * throtl_add_bio_tg - add a bio to the specified throtl_grp
1032 * @tg: the target throtl_grp
1034 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
1035 * tg->qnode_on_self[] is used.
1037 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1038 struct throtl_grp *tg)
1040 struct throtl_service_queue *sq = &tg->service_queue;
1041 bool rw = bio_data_dir(bio);
1044 qn = &tg->qnode_on_self[rw];
1047 * If @tg doesn't currently have any bios queued in the same
1048 * direction, queueing @bio can change when @tg should be
1049 * dispatched. Mark that @tg was empty. This is automatically
1050 * cleaered on the next tg_update_disptime().
1052 if (!sq->nr_queued[rw])
1053 tg->flags |= THROTL_TG_WAS_EMPTY;
1055 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1057 sq->nr_queued[rw]++;
1058 throtl_enqueue_tg(tg);
1061 static void tg_update_disptime(struct throtl_grp *tg)
1063 struct throtl_service_queue *sq = &tg->service_queue;
1064 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1067 bio = throtl_peek_queued(&sq->queued[READ]);
1069 tg_may_dispatch(tg, bio, &read_wait);
1071 bio = throtl_peek_queued(&sq->queued[WRITE]);
1073 tg_may_dispatch(tg, bio, &write_wait);
1075 min_wait = min(read_wait, write_wait);
1076 disptime = jiffies + min_wait;
1078 /* Update dispatch time */
1079 throtl_dequeue_tg(tg);
1080 tg->disptime = disptime;
1081 throtl_enqueue_tg(tg);
1083 /* see throtl_add_bio_tg() */
1084 tg->flags &= ~THROTL_TG_WAS_EMPTY;
1087 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1088 struct throtl_grp *parent_tg, bool rw)
1090 if (throtl_slice_used(parent_tg, rw)) {
1091 throtl_start_new_slice_with_credit(parent_tg, rw,
1092 child_tg->slice_start[rw]);
1097 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1099 struct throtl_service_queue *sq = &tg->service_queue;
1100 struct throtl_service_queue *parent_sq = sq->parent_sq;
1101 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1102 struct throtl_grp *tg_to_put = NULL;
1106 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1107 * from @tg may put its reference and @parent_sq might end up
1108 * getting released prematurely. Remember the tg to put and put it
1109 * after @bio is transferred to @parent_sq.
1111 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1112 sq->nr_queued[rw]--;
1114 throtl_charge_bio(tg, bio);
1117 * If our parent is another tg, we just need to transfer @bio to
1118 * the parent using throtl_add_bio_tg(). If our parent is
1119 * @td->service_queue, @bio is ready to be issued. Put it on its
1120 * bio_lists[] and decrease total number queued. The caller is
1121 * responsible for issuing these bios.
1124 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1125 start_parent_slice_with_credit(tg, parent_tg, rw);
1127 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1128 &parent_sq->queued[rw]);
1129 BUG_ON(tg->td->nr_queued[rw] <= 0);
1130 tg->td->nr_queued[rw]--;
1133 throtl_trim_slice(tg, rw);
1136 blkg_put(tg_to_blkg(tg_to_put));
1139 static int throtl_dispatch_tg(struct throtl_grp *tg)
1141 struct throtl_service_queue *sq = &tg->service_queue;
1142 unsigned int nr_reads = 0, nr_writes = 0;
1143 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1144 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1147 /* Try to dispatch 75% READS and 25% WRITES */
1149 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1150 tg_may_dispatch(tg, bio, NULL)) {
1152 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1155 if (nr_reads >= max_nr_reads)
1159 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1160 tg_may_dispatch(tg, bio, NULL)) {
1162 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1165 if (nr_writes >= max_nr_writes)
1169 return nr_reads + nr_writes;
1172 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1174 unsigned int nr_disp = 0;
1177 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1178 struct throtl_service_queue *sq = &tg->service_queue;
1183 if (time_before(jiffies, tg->disptime))
1186 throtl_dequeue_tg(tg);
1188 nr_disp += throtl_dispatch_tg(tg);
1190 if (sq->nr_queued[0] || sq->nr_queued[1])
1191 tg_update_disptime(tg);
1193 if (nr_disp >= throtl_quantum)
1200 static bool throtl_can_upgrade(struct throtl_data *td,
1201 struct throtl_grp *this_tg);
1203 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1204 * @arg: the throtl_service_queue being serviced
1206 * This timer is armed when a child throtl_grp with active bio's become
1207 * pending and queued on the service_queue's pending_tree and expires when
1208 * the first child throtl_grp should be dispatched. This function
1209 * dispatches bio's from the children throtl_grps to the parent
1212 * If the parent's parent is another throtl_grp, dispatching is propagated
1213 * by either arming its pending_timer or repeating dispatch directly. If
1214 * the top-level service_tree is reached, throtl_data->dispatch_work is
1215 * kicked so that the ready bio's are issued.
1217 static void throtl_pending_timer_fn(unsigned long arg)
1219 struct throtl_service_queue *sq = (void *)arg;
1220 struct throtl_grp *tg = sq_to_tg(sq);
1221 struct throtl_data *td = sq_to_td(sq);
1222 struct request_queue *q = td->queue;
1223 struct throtl_service_queue *parent_sq;
1227 spin_lock_irq(q->queue_lock);
1228 if (throtl_can_upgrade(td, NULL))
1229 throtl_upgrade_state(td);
1232 parent_sq = sq->parent_sq;
1236 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1237 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1238 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1240 ret = throtl_select_dispatch(sq);
1242 throtl_log(sq, "bios disp=%u", ret);
1246 if (throtl_schedule_next_dispatch(sq, false))
1249 /* this dispatch windows is still open, relax and repeat */
1250 spin_unlock_irq(q->queue_lock);
1252 spin_lock_irq(q->queue_lock);
1259 /* @parent_sq is another throl_grp, propagate dispatch */
1260 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1261 tg_update_disptime(tg);
1262 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1263 /* window is already open, repeat dispatching */
1270 /* reached the top-level, queue issueing */
1271 queue_work(kthrotld_workqueue, &td->dispatch_work);
1274 spin_unlock_irq(q->queue_lock);
1278 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1279 * @work: work item being executed
1281 * This function is queued for execution when bio's reach the bio_lists[]
1282 * of throtl_data->service_queue. Those bio's are ready and issued by this
1285 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1287 struct throtl_data *td = container_of(work, struct throtl_data,
1289 struct throtl_service_queue *td_sq = &td->service_queue;
1290 struct request_queue *q = td->queue;
1291 struct bio_list bio_list_on_stack;
1293 struct blk_plug plug;
1296 bio_list_init(&bio_list_on_stack);
1298 spin_lock_irq(q->queue_lock);
1299 for (rw = READ; rw <= WRITE; rw++)
1300 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1301 bio_list_add(&bio_list_on_stack, bio);
1302 spin_unlock_irq(q->queue_lock);
1304 if (!bio_list_empty(&bio_list_on_stack)) {
1305 blk_start_plug(&plug);
1306 while((bio = bio_list_pop(&bio_list_on_stack)))
1307 generic_make_request(bio);
1308 blk_finish_plug(&plug);
1312 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1315 struct throtl_grp *tg = pd_to_tg(pd);
1316 u64 v = *(u64 *)((void *)tg + off);
1320 return __blkg_prfill_u64(sf, pd, v);
1323 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1326 struct throtl_grp *tg = pd_to_tg(pd);
1327 unsigned int v = *(unsigned int *)((void *)tg + off);
1331 return __blkg_prfill_u64(sf, pd, v);
1334 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1336 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1337 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1341 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1343 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1344 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1348 static void tg_conf_updated(struct throtl_grp *tg)
1350 struct throtl_service_queue *sq = &tg->service_queue;
1351 struct cgroup_subsys_state *pos_css;
1352 struct blkcg_gq *blkg;
1354 throtl_log(&tg->service_queue,
1355 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1356 tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1357 tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1360 * Update has_rules[] flags for the updated tg's subtree. A tg is
1361 * considered to have rules if either the tg itself or any of its
1362 * ancestors has rules. This identifies groups without any
1363 * restrictions in the whole hierarchy and allows them to bypass
1366 blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg))
1367 tg_update_has_rules(blkg_to_tg(blkg));
1370 * We're already holding queue_lock and know @tg is valid. Let's
1371 * apply the new config directly.
1373 * Restart the slices for both READ and WRITES. It might happen
1374 * that a group's limit are dropped suddenly and we don't want to
1375 * account recently dispatched IO with new low rate.
1377 throtl_start_new_slice(tg, 0);
1378 throtl_start_new_slice(tg, 1);
1380 if (tg->flags & THROTL_TG_PENDING) {
1381 tg_update_disptime(tg);
1382 throtl_schedule_next_dispatch(sq->parent_sq, true);
1386 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1387 char *buf, size_t nbytes, loff_t off, bool is_u64)
1389 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1390 struct blkg_conf_ctx ctx;
1391 struct throtl_grp *tg;
1395 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1400 if (sscanf(ctx.body, "%llu", &v) != 1)
1405 tg = blkg_to_tg(ctx.blkg);
1408 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1410 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1412 tg_conf_updated(tg);
1415 blkg_conf_finish(&ctx);
1416 return ret ?: nbytes;
1419 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1420 char *buf, size_t nbytes, loff_t off)
1422 return tg_set_conf(of, buf, nbytes, off, true);
1425 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1426 char *buf, size_t nbytes, loff_t off)
1428 return tg_set_conf(of, buf, nbytes, off, false);
1431 static struct cftype throtl_legacy_files[] = {
1433 .name = "throttle.read_bps_device",
1434 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1435 .seq_show = tg_print_conf_u64,
1436 .write = tg_set_conf_u64,
1439 .name = "throttle.write_bps_device",
1440 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1441 .seq_show = tg_print_conf_u64,
1442 .write = tg_set_conf_u64,
1445 .name = "throttle.read_iops_device",
1446 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1447 .seq_show = tg_print_conf_uint,
1448 .write = tg_set_conf_uint,
1451 .name = "throttle.write_iops_device",
1452 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1453 .seq_show = tg_print_conf_uint,
1454 .write = tg_set_conf_uint,
1457 .name = "throttle.io_service_bytes",
1458 .private = (unsigned long)&blkcg_policy_throtl,
1459 .seq_show = blkg_print_stat_bytes,
1462 .name = "throttle.io_serviced",
1463 .private = (unsigned long)&blkcg_policy_throtl,
1464 .seq_show = blkg_print_stat_ios,
1469 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1472 struct throtl_grp *tg = pd_to_tg(pd);
1473 const char *dname = blkg_dev_name(pd->blkg);
1474 char bufs[4][21] = { "max", "max", "max", "max" };
1476 unsigned int iops_dft;
1477 char idle_time[26] = "";
1478 char latency_time[26] = "";
1483 if (off == LIMIT_LOW) {
1488 iops_dft = UINT_MAX;
1491 if (tg->bps_conf[READ][off] == bps_dft &&
1492 tg->bps_conf[WRITE][off] == bps_dft &&
1493 tg->iops_conf[READ][off] == iops_dft &&
1494 tg->iops_conf[WRITE][off] == iops_dft &&
1495 (off != LIMIT_LOW ||
1496 (tg->idletime_threshold == tg->td->dft_idletime_threshold &&
1497 tg->latency_target == DFL_LATENCY_TARGET)))
1500 if (tg->bps_conf[READ][off] != bps_dft)
1501 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1502 tg->bps_conf[READ][off]);
1503 if (tg->bps_conf[WRITE][off] != bps_dft)
1504 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1505 tg->bps_conf[WRITE][off]);
1506 if (tg->iops_conf[READ][off] != iops_dft)
1507 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1508 tg->iops_conf[READ][off]);
1509 if (tg->iops_conf[WRITE][off] != iops_dft)
1510 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1511 tg->iops_conf[WRITE][off]);
1512 if (off == LIMIT_LOW) {
1513 if (tg->idletime_threshold == ULONG_MAX)
1514 strcpy(idle_time, " idle=max");
1516 snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1517 tg->idletime_threshold);
1519 if (tg->latency_target == ULONG_MAX)
1520 strcpy(latency_time, " latency=max");
1522 snprintf(latency_time, sizeof(latency_time),
1523 " latency=%lu", tg->latency_target);
1526 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1527 dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1532 static int tg_print_limit(struct seq_file *sf, void *v)
1534 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1535 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1539 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1540 char *buf, size_t nbytes, loff_t off)
1542 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1543 struct blkg_conf_ctx ctx;
1544 struct throtl_grp *tg;
1546 unsigned long idle_time;
1547 unsigned long latency_time;
1549 int index = of_cft(of)->private;
1551 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1555 tg = blkg_to_tg(ctx.blkg);
1557 v[0] = tg->bps_conf[READ][index];
1558 v[1] = tg->bps_conf[WRITE][index];
1559 v[2] = tg->iops_conf[READ][index];
1560 v[3] = tg->iops_conf[WRITE][index];
1562 idle_time = tg->idletime_threshold;
1563 latency_time = tg->latency_target;
1565 char tok[27]; /* wiops=18446744073709551616 */
1570 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1579 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1587 if (!strcmp(tok, "rbps"))
1589 else if (!strcmp(tok, "wbps"))
1591 else if (!strcmp(tok, "riops"))
1592 v[2] = min_t(u64, val, UINT_MAX);
1593 else if (!strcmp(tok, "wiops"))
1594 v[3] = min_t(u64, val, UINT_MAX);
1595 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1597 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1603 tg->bps_conf[READ][index] = v[0];
1604 tg->bps_conf[WRITE][index] = v[1];
1605 tg->iops_conf[READ][index] = v[2];
1606 tg->iops_conf[WRITE][index] = v[3];
1608 if (index == LIMIT_MAX) {
1609 tg->bps[READ][index] = v[0];
1610 tg->bps[WRITE][index] = v[1];
1611 tg->iops[READ][index] = v[2];
1612 tg->iops[WRITE][index] = v[3];
1614 tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1615 tg->bps_conf[READ][LIMIT_MAX]);
1616 tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1617 tg->bps_conf[WRITE][LIMIT_MAX]);
1618 tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1619 tg->iops_conf[READ][LIMIT_MAX]);
1620 tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1621 tg->iops_conf[WRITE][LIMIT_MAX]);
1623 if (index == LIMIT_LOW) {
1624 blk_throtl_update_limit_valid(tg->td);
1625 if (tg->td->limit_valid[LIMIT_LOW])
1626 tg->td->limit_index = LIMIT_LOW;
1627 tg->idletime_threshold = (idle_time == ULONG_MAX) ?
1628 ULONG_MAX : idle_time;
1629 tg->latency_target = (latency_time == ULONG_MAX) ?
1630 ULONG_MAX : latency_time;
1632 tg_conf_updated(tg);
1635 blkg_conf_finish(&ctx);
1636 return ret ?: nbytes;
1639 static struct cftype throtl_files[] = {
1640 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1643 .flags = CFTYPE_NOT_ON_ROOT,
1644 .seq_show = tg_print_limit,
1645 .write = tg_set_limit,
1646 .private = LIMIT_LOW,
1651 .flags = CFTYPE_NOT_ON_ROOT,
1652 .seq_show = tg_print_limit,
1653 .write = tg_set_limit,
1654 .private = LIMIT_MAX,
1659 static void throtl_shutdown_wq(struct request_queue *q)
1661 struct throtl_data *td = q->td;
1663 cancel_work_sync(&td->dispatch_work);
1666 static struct blkcg_policy blkcg_policy_throtl = {
1667 .dfl_cftypes = throtl_files,
1668 .legacy_cftypes = throtl_legacy_files,
1670 .pd_alloc_fn = throtl_pd_alloc,
1671 .pd_init_fn = throtl_pd_init,
1672 .pd_online_fn = throtl_pd_online,
1673 .pd_offline_fn = throtl_pd_offline,
1674 .pd_free_fn = throtl_pd_free,
1677 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1679 unsigned long rtime = jiffies, wtime = jiffies;
1681 if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1682 rtime = tg->last_low_overflow_time[READ];
1683 if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1684 wtime = tg->last_low_overflow_time[WRITE];
1685 return min(rtime, wtime);
1688 /* tg should not be an intermediate node */
1689 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1691 struct throtl_service_queue *parent_sq;
1692 struct throtl_grp *parent = tg;
1693 unsigned long ret = __tg_last_low_overflow_time(tg);
1696 parent_sq = parent->service_queue.parent_sq;
1697 parent = sq_to_tg(parent_sq);
1702 * The parent doesn't have low limit, it always reaches low
1703 * limit. Its overflow time is useless for children
1705 if (!parent->bps[READ][LIMIT_LOW] &&
1706 !parent->iops[READ][LIMIT_LOW] &&
1707 !parent->bps[WRITE][LIMIT_LOW] &&
1708 !parent->iops[WRITE][LIMIT_LOW])
1710 if (time_after(__tg_last_low_overflow_time(parent), ret))
1711 ret = __tg_last_low_overflow_time(parent);
1716 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1719 * cgroup is idle if:
1720 * - single idle is too long, longer than a fixed value (in case user
1721 * configure a too big threshold) or 4 times of slice
1722 * - average think time is more than threshold
1724 unsigned long time = jiffies_to_usecs(4 * tg->td->throtl_slice);
1726 time = min_t(unsigned long, MAX_IDLE_TIME, time);
1727 return (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1728 tg->avg_idletime > tg->idletime_threshold;
1731 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1733 struct throtl_service_queue *sq = &tg->service_queue;
1734 bool read_limit, write_limit;
1737 * if cgroup reaches low limit (if low limit is 0, the cgroup always
1738 * reaches), it's ok to upgrade to next limit
1740 read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1741 write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1742 if (!read_limit && !write_limit)
1744 if (read_limit && sq->nr_queued[READ] &&
1745 (!write_limit || sq->nr_queued[WRITE]))
1747 if (write_limit && sq->nr_queued[WRITE] &&
1748 (!read_limit || sq->nr_queued[READ]))
1751 if (time_after_eq(jiffies,
1752 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1753 throtl_tg_is_idle(tg))
1758 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1761 if (throtl_tg_can_upgrade(tg))
1763 tg = sq_to_tg(tg->service_queue.parent_sq);
1764 if (!tg || !tg_to_blkg(tg)->parent)
1770 static bool throtl_can_upgrade(struct throtl_data *td,
1771 struct throtl_grp *this_tg)
1773 struct cgroup_subsys_state *pos_css;
1774 struct blkcg_gq *blkg;
1776 if (td->limit_index != LIMIT_LOW)
1779 if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1783 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1784 struct throtl_grp *tg = blkg_to_tg(blkg);
1788 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1790 if (!throtl_hierarchy_can_upgrade(tg)) {
1799 static void throtl_upgrade_check(struct throtl_grp *tg)
1801 unsigned long now = jiffies;
1803 if (tg->td->limit_index != LIMIT_LOW)
1806 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1809 tg->last_check_time = now;
1811 if (!time_after_eq(now,
1812 __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1815 if (throtl_can_upgrade(tg->td, NULL))
1816 throtl_upgrade_state(tg->td);
1819 static void throtl_upgrade_state(struct throtl_data *td)
1821 struct cgroup_subsys_state *pos_css;
1822 struct blkcg_gq *blkg;
1824 td->limit_index = LIMIT_MAX;
1825 td->low_upgrade_time = jiffies;
1828 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1829 struct throtl_grp *tg = blkg_to_tg(blkg);
1830 struct throtl_service_queue *sq = &tg->service_queue;
1832 tg->disptime = jiffies - 1;
1833 throtl_select_dispatch(sq);
1834 throtl_schedule_next_dispatch(sq, false);
1837 throtl_select_dispatch(&td->service_queue);
1838 throtl_schedule_next_dispatch(&td->service_queue, false);
1839 queue_work(kthrotld_workqueue, &td->dispatch_work);
1842 static void throtl_downgrade_state(struct throtl_data *td, int new)
1847 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1851 td->limit_index = new;
1852 td->low_downgrade_time = jiffies;
1855 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1857 struct throtl_data *td = tg->td;
1858 unsigned long now = jiffies;
1861 * If cgroup is below low limit, consider downgrade and throttle other
1864 if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1865 time_after_eq(now, tg_last_low_overflow_time(tg) +
1866 td->throtl_slice) &&
1867 (!throtl_tg_is_idle(tg) ||
1868 !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1873 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1876 if (!throtl_tg_can_downgrade(tg))
1878 tg = sq_to_tg(tg->service_queue.parent_sq);
1879 if (!tg || !tg_to_blkg(tg)->parent)
1885 static void throtl_downgrade_check(struct throtl_grp *tg)
1889 unsigned long elapsed_time;
1890 unsigned long now = jiffies;
1892 if (tg->td->limit_index != LIMIT_MAX ||
1893 !tg->td->limit_valid[LIMIT_LOW])
1895 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1897 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1900 elapsed_time = now - tg->last_check_time;
1901 tg->last_check_time = now;
1903 if (time_before(now, tg_last_low_overflow_time(tg) +
1904 tg->td->throtl_slice))
1907 if (tg->bps[READ][LIMIT_LOW]) {
1908 bps = tg->last_bytes_disp[READ] * HZ;
1909 do_div(bps, elapsed_time);
1910 if (bps >= tg->bps[READ][LIMIT_LOW])
1911 tg->last_low_overflow_time[READ] = now;
1914 if (tg->bps[WRITE][LIMIT_LOW]) {
1915 bps = tg->last_bytes_disp[WRITE] * HZ;
1916 do_div(bps, elapsed_time);
1917 if (bps >= tg->bps[WRITE][LIMIT_LOW])
1918 tg->last_low_overflow_time[WRITE] = now;
1921 if (tg->iops[READ][LIMIT_LOW]) {
1922 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
1923 if (iops >= tg->iops[READ][LIMIT_LOW])
1924 tg->last_low_overflow_time[READ] = now;
1927 if (tg->iops[WRITE][LIMIT_LOW]) {
1928 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
1929 if (iops >= tg->iops[WRITE][LIMIT_LOW])
1930 tg->last_low_overflow_time[WRITE] = now;
1934 * If cgroup is below low limit, consider downgrade and throttle other
1937 if (throtl_hierarchy_can_downgrade(tg))
1938 throtl_downgrade_state(tg->td, LIMIT_LOW);
1940 tg->last_bytes_disp[READ] = 0;
1941 tg->last_bytes_disp[WRITE] = 0;
1942 tg->last_io_disp[READ] = 0;
1943 tg->last_io_disp[WRITE] = 0;
1946 static void blk_throtl_update_idletime(struct throtl_grp *tg)
1948 unsigned long now = ktime_get_ns() >> 10;
1949 unsigned long last_finish_time = tg->last_finish_time;
1951 if (now <= last_finish_time || last_finish_time == 0 ||
1952 last_finish_time == tg->checked_last_finish_time)
1955 tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
1956 tg->checked_last_finish_time = last_finish_time;
1959 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1960 static void throtl_update_latency_buckets(struct throtl_data *td)
1962 struct avg_latency_bucket avg_latency[LATENCY_BUCKET_SIZE];
1964 unsigned long last_latency = 0;
1965 unsigned long latency;
1967 if (!blk_queue_nonrot(td->queue))
1969 if (time_before(jiffies, td->last_calculate_time + HZ))
1971 td->last_calculate_time = jiffies;
1973 memset(avg_latency, 0, sizeof(avg_latency));
1974 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
1975 struct latency_bucket *tmp = &td->tmp_buckets[i];
1977 for_each_possible_cpu(cpu) {
1978 struct latency_bucket *bucket;
1980 /* this isn't race free, but ok in practice */
1981 bucket = per_cpu_ptr(td->latency_buckets, cpu);
1982 tmp->total_latency += bucket[i].total_latency;
1983 tmp->samples += bucket[i].samples;
1984 bucket[i].total_latency = 0;
1985 bucket[i].samples = 0;
1988 if (tmp->samples >= 32) {
1989 int samples = tmp->samples;
1991 latency = tmp->total_latency;
1993 tmp->total_latency = 0;
1998 avg_latency[i].latency = latency;
2002 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2003 if (!avg_latency[i].latency) {
2004 if (td->avg_buckets[i].latency < last_latency)
2005 td->avg_buckets[i].latency = last_latency;
2009 if (!td->avg_buckets[i].valid)
2010 latency = avg_latency[i].latency;
2012 latency = (td->avg_buckets[i].latency * 7 +
2013 avg_latency[i].latency) >> 3;
2015 td->avg_buckets[i].latency = max(latency, last_latency);
2016 td->avg_buckets[i].valid = true;
2017 last_latency = td->avg_buckets[i].latency;
2021 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2026 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
2029 struct throtl_qnode *qn = NULL;
2030 struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
2031 struct throtl_service_queue *sq;
2032 bool rw = bio_data_dir(bio);
2033 bool throttled = false;
2034 struct throtl_data *td = tg->td;
2037 WARN_ON_ONCE(!rcu_read_lock_held());
2039 /* see throtl_charge_bio() */
2040 if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw])
2043 spin_lock_irq(q->queue_lock);
2045 throtl_update_latency_buckets(td);
2047 if (unlikely(blk_queue_bypass(q)))
2050 ret = bio_associate_current(bio);
2051 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2052 if (ret == 0 || ret == -EBUSY)
2053 bio->bi_cg_private = tg;
2054 blk_stat_set_issue(&bio->bi_issue_stat, bio_sectors(bio));
2056 blk_throtl_update_idletime(tg);
2058 sq = &tg->service_queue;
2062 if (tg->last_low_overflow_time[rw] == 0)
2063 tg->last_low_overflow_time[rw] = jiffies;
2064 throtl_downgrade_check(tg);
2065 throtl_upgrade_check(tg);
2066 /* throtl is FIFO - if bios are already queued, should queue */
2067 if (sq->nr_queued[rw])
2070 /* if above limits, break to queue */
2071 if (!tg_may_dispatch(tg, bio, NULL)) {
2072 tg->last_low_overflow_time[rw] = jiffies;
2073 if (throtl_can_upgrade(td, tg)) {
2074 throtl_upgrade_state(td);
2080 /* within limits, let's charge and dispatch directly */
2081 throtl_charge_bio(tg, bio);
2084 * We need to trim slice even when bios are not being queued
2085 * otherwise it might happen that a bio is not queued for
2086 * a long time and slice keeps on extending and trim is not
2087 * called for a long time. Now if limits are reduced suddenly
2088 * we take into account all the IO dispatched so far at new
2089 * low rate and * newly queued IO gets a really long dispatch
2092 * So keep on trimming slice even if bio is not queued.
2094 throtl_trim_slice(tg, rw);
2097 * @bio passed through this layer without being throttled.
2098 * Climb up the ladder. If we''re already at the top, it
2099 * can be executed directly.
2101 qn = &tg->qnode_on_parent[rw];
2108 /* out-of-limit, queue to @tg */
2109 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2110 rw == READ ? 'R' : 'W',
2111 tg->bytes_disp[rw], bio->bi_iter.bi_size,
2112 tg_bps_limit(tg, rw),
2113 tg->io_disp[rw], tg_iops_limit(tg, rw),
2114 sq->nr_queued[READ], sq->nr_queued[WRITE]);
2116 tg->last_low_overflow_time[rw] = jiffies;
2118 td->nr_queued[rw]++;
2119 throtl_add_bio_tg(bio, qn, tg);
2123 * Update @tg's dispatch time and force schedule dispatch if @tg
2124 * was empty before @bio. The forced scheduling isn't likely to
2125 * cause undue delay as @bio is likely to be dispatched directly if
2126 * its @tg's disptime is not in the future.
2128 if (tg->flags & THROTL_TG_WAS_EMPTY) {
2129 tg_update_disptime(tg);
2130 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2134 spin_unlock_irq(q->queue_lock);
2137 * As multiple blk-throtls may stack in the same issue path, we
2138 * don't want bios to leave with the flag set. Clear the flag if
2142 bio_clear_flag(bio, BIO_THROTTLED);
2144 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2145 if (throttled || !td->track_bio_latency)
2146 bio->bi_issue_stat.stat |= SKIP_LATENCY;
2151 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2152 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2153 int op, unsigned long time)
2155 struct latency_bucket *latency;
2158 if (!td || td->limit_index != LIMIT_LOW || op != REQ_OP_READ ||
2159 !blk_queue_nonrot(td->queue))
2162 index = request_bucket_index(size);
2164 latency = get_cpu_ptr(td->latency_buckets);
2165 latency[index].total_latency += time;
2166 latency[index].samples++;
2167 put_cpu_ptr(td->latency_buckets);
2170 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2172 struct request_queue *q = rq->q;
2173 struct throtl_data *td = q->td;
2175 throtl_track_latency(td, blk_stat_size(&rq->issue_stat),
2176 req_op(rq), time_ns >> 10);
2179 void blk_throtl_bio_endio(struct bio *bio)
2181 struct throtl_grp *tg;
2183 unsigned long finish_time;
2184 unsigned long start_time;
2187 tg = bio->bi_cg_private;
2190 bio->bi_cg_private = NULL;
2192 finish_time_ns = ktime_get_ns();
2193 tg->last_finish_time = finish_time_ns >> 10;
2195 start_time = blk_stat_time(&bio->bi_issue_stat) >> 10;
2196 finish_time = __blk_stat_time(finish_time_ns) >> 10;
2197 /* this is only for bio based driver */
2198 if (start_time && finish_time > start_time &&
2199 !(bio->bi_issue_stat.stat & SKIP_LATENCY)) {
2200 lat = finish_time - start_time;
2201 throtl_track_latency(tg->td, blk_stat_size(&bio->bi_issue_stat),
2208 * Dispatch all bios from all children tg's queued on @parent_sq. On
2209 * return, @parent_sq is guaranteed to not have any active children tg's
2210 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
2212 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
2214 struct throtl_grp *tg;
2216 while ((tg = throtl_rb_first(parent_sq))) {
2217 struct throtl_service_queue *sq = &tg->service_queue;
2220 throtl_dequeue_tg(tg);
2222 while ((bio = throtl_peek_queued(&sq->queued[READ])))
2223 tg_dispatch_one_bio(tg, bio_data_dir(bio));
2224 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
2225 tg_dispatch_one_bio(tg, bio_data_dir(bio));
2230 * blk_throtl_drain - drain throttled bios
2231 * @q: request_queue to drain throttled bios for
2233 * Dispatch all currently throttled bios on @q through ->make_request_fn().
2235 void blk_throtl_drain(struct request_queue *q)
2236 __releases(q->queue_lock) __acquires(q->queue_lock)
2238 struct throtl_data *td = q->td;
2239 struct blkcg_gq *blkg;
2240 struct cgroup_subsys_state *pos_css;
2244 queue_lockdep_assert_held(q);
2248 * Drain each tg while doing post-order walk on the blkg tree, so
2249 * that all bios are propagated to td->service_queue. It'd be
2250 * better to walk service_queue tree directly but blkg walk is
2253 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
2254 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
2256 /* finally, transfer bios from top-level tg's into the td */
2257 tg_drain_bios(&td->service_queue);
2260 spin_unlock_irq(q->queue_lock);
2262 /* all bios now should be in td->service_queue, issue them */
2263 for (rw = READ; rw <= WRITE; rw++)
2264 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
2266 generic_make_request(bio);
2268 spin_lock_irq(q->queue_lock);
2271 int blk_throtl_init(struct request_queue *q)
2273 struct throtl_data *td;
2276 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2279 td->latency_buckets = __alloc_percpu(sizeof(struct latency_bucket) *
2280 LATENCY_BUCKET_SIZE, __alignof__(u64));
2281 if (!td->latency_buckets) {
2286 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2287 throtl_service_queue_init(&td->service_queue);
2292 td->limit_valid[LIMIT_MAX] = true;
2293 td->limit_index = LIMIT_MAX;
2294 td->low_upgrade_time = jiffies;
2295 td->low_downgrade_time = jiffies;
2297 /* activate policy */
2298 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2300 free_percpu(td->latency_buckets);
2306 void blk_throtl_exit(struct request_queue *q)
2309 throtl_shutdown_wq(q);
2310 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2311 free_percpu(q->td->latency_buckets);
2315 void blk_throtl_register_queue(struct request_queue *q)
2317 struct throtl_data *td;
2318 struct cgroup_subsys_state *pos_css;
2319 struct blkcg_gq *blkg;
2324 if (blk_queue_nonrot(q)) {
2325 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2326 td->dft_idletime_threshold = DFL_IDLE_THRESHOLD_SSD;
2328 td->throtl_slice = DFL_THROTL_SLICE_HD;
2329 td->dft_idletime_threshold = DFL_IDLE_THRESHOLD_HD;
2331 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2332 /* if no low limit, use previous default */
2333 td->throtl_slice = DFL_THROTL_SLICE_HD;
2336 td->track_bio_latency = !q->mq_ops && !q->request_fn;
2337 if (!td->track_bio_latency)
2338 blk_stat_enable_accounting(q);
2341 * some tg are created before queue is fully initialized, eg, nonrot
2342 * isn't initialized yet
2345 blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) {
2346 struct throtl_grp *tg = blkg_to_tg(blkg);
2348 tg->idletime_threshold = td->dft_idletime_threshold;
2353 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2354 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2358 return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2361 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2362 const char *page, size_t count)
2369 if (kstrtoul(page, 10, &v))
2371 t = msecs_to_jiffies(v);
2372 if (t == 0 || t > MAX_THROTL_SLICE)
2374 q->td->throtl_slice = t;
2379 static int __init throtl_init(void)
2381 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2382 if (!kthrotld_workqueue)
2383 panic("Failed to create kthrotld\n");
2385 return blkcg_policy_register(&blkcg_policy_throtl);
2388 module_init(throtl_init);