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 100ms slice and after that slice is renewed */
22 static unsigned long throtl_slice = HZ/10; /* 100 ms */
24 static struct blkcg_policy blkcg_policy_throtl;
26 /* A workqueue to queue throttle related work */
27 static struct workqueue_struct *kthrotld_workqueue;
30 * To implement hierarchical throttling, throtl_grps form a tree and bios
31 * are dispatched upwards level by level until they reach the top and get
32 * issued. When dispatching bios from the children and local group at each
33 * level, if the bios are dispatched into a single bio_list, there's a risk
34 * of a local or child group which can queue many bios at once filling up
35 * the list starving others.
37 * To avoid such starvation, dispatched bios are queued separately
38 * according to where they came from. When they are again dispatched to
39 * the parent, they're popped in round-robin order so that no single source
40 * hogs the dispatch window.
42 * throtl_qnode is used to keep the queued bios separated by their sources.
43 * Bios are queued to throtl_qnode which in turn is queued to
44 * throtl_service_queue and then dispatched in round-robin order.
46 * It's also used to track the reference counts on blkg's. A qnode always
47 * belongs to a throtl_grp and gets queued on itself or the parent, so
48 * incrementing the reference of the associated throtl_grp when a qnode is
49 * queued and decrementing when dequeued is enough to keep the whole blkg
50 * tree pinned while bios are in flight.
53 struct list_head node; /* service_queue->queued[] */
54 struct bio_list bios; /* queued bios */
55 struct throtl_grp *tg; /* tg this qnode belongs to */
58 struct throtl_service_queue {
59 struct throtl_service_queue *parent_sq; /* the parent service_queue */
62 * Bios queued directly to this service_queue or dispatched from
63 * children throtl_grp's.
65 struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */
66 unsigned int nr_queued[2]; /* number of queued bios */
69 * RB tree of active children throtl_grp's, which are sorted by
72 struct rb_root pending_tree; /* RB tree of active tgs */
73 struct rb_node *first_pending; /* first node in the tree */
74 unsigned int nr_pending; /* # queued in the tree */
75 unsigned long first_pending_disptime; /* disptime of the first tg */
76 struct timer_list pending_timer; /* fires on first_pending_disptime */
80 THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */
81 THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */
84 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
86 /* Per-cpu group stats */
88 /* total bytes transferred */
89 struct blkg_rwstat service_bytes;
90 /* total IOs serviced, post merge */
91 struct blkg_rwstat serviced;
95 /* must be the first member */
96 struct blkg_policy_data pd;
98 /* active throtl group service_queue member */
99 struct rb_node rb_node;
101 /* throtl_data this group belongs to */
102 struct throtl_data *td;
104 /* this group's service queue */
105 struct throtl_service_queue service_queue;
108 * qnode_on_self is used when bios are directly queued to this
109 * throtl_grp so that local bios compete fairly with bios
110 * dispatched from children. qnode_on_parent is used when bios are
111 * dispatched from this throtl_grp into its parent and will compete
112 * with the sibling qnode_on_parents and the parent's
115 struct throtl_qnode qnode_on_self[2];
116 struct throtl_qnode qnode_on_parent[2];
119 * Dispatch time in jiffies. This is the estimated time when group
120 * will unthrottle and is ready to dispatch more bio. It is used as
121 * key to sort active groups in service tree.
123 unsigned long disptime;
127 /* are there any throtl rules between this group and td? */
130 /* bytes per second rate limits */
134 unsigned int iops[2];
136 /* Number of bytes disptached in current slice */
137 uint64_t bytes_disp[2];
138 /* Number of bio's dispatched in current slice */
139 unsigned int io_disp[2];
141 /* When did we start a new slice */
142 unsigned long slice_start[2];
143 unsigned long slice_end[2];
145 /* Per cpu stats pointer */
146 struct tg_stats_cpu __percpu *stats_cpu;
151 /* service tree for active throtl groups */
152 struct throtl_service_queue service_queue;
154 struct request_queue *queue;
156 /* Total Number of queued bios on READ and WRITE lists */
157 unsigned int nr_queued[2];
160 * number of total undestroyed groups
162 unsigned int nr_undestroyed_grps;
164 /* Work for dispatching throttled bios */
165 struct work_struct dispatch_work;
168 static void throtl_pending_timer_fn(unsigned long arg);
170 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
172 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
175 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
177 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
180 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
182 return pd_to_blkg(&tg->pd);
185 static inline struct throtl_grp *td_root_tg(struct throtl_data *td)
187 return blkg_to_tg(td->queue->root_blkg);
191 * sq_to_tg - return the throl_grp the specified service queue belongs to
192 * @sq: the throtl_service_queue of interest
194 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
195 * embedded in throtl_data, %NULL is returned.
197 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
199 if (sq && sq->parent_sq)
200 return container_of(sq, struct throtl_grp, service_queue);
206 * sq_to_td - return throtl_data the specified service queue belongs to
207 * @sq: the throtl_service_queue of interest
209 * A service_queue can be embeded in either a throtl_grp or throtl_data.
210 * Determine the associated throtl_data accordingly and return it.
212 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
214 struct throtl_grp *tg = sq_to_tg(sq);
219 return container_of(sq, struct throtl_data, service_queue);
223 * throtl_log - log debug message via blktrace
224 * @sq: the service_queue being reported
225 * @fmt: printf format string
228 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
229 * throtl_grp; otherwise, just "throtl".
231 * TODO: this should be made a function and name formatting should happen
232 * after testing whether blktrace is enabled.
234 #define throtl_log(sq, fmt, args...) do { \
235 struct throtl_grp *__tg = sq_to_tg((sq)); \
236 struct throtl_data *__td = sq_to_td((sq)); \
242 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
243 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
245 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
249 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
251 INIT_LIST_HEAD(&qn->node);
252 bio_list_init(&qn->bios);
257 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
258 * @bio: bio being added
259 * @qn: qnode to add bio to
260 * @queued: the service_queue->queued[] list @qn belongs to
262 * Add @bio to @qn and put @qn on @queued if it's not already on.
263 * @qn->tg's reference count is bumped when @qn is activated. See the
264 * comment on top of throtl_qnode definition for details.
266 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
267 struct list_head *queued)
269 bio_list_add(&qn->bios, bio);
270 if (list_empty(&qn->node)) {
271 list_add_tail(&qn->node, queued);
272 blkg_get(tg_to_blkg(qn->tg));
277 * throtl_peek_queued - peek the first bio on a qnode list
278 * @queued: the qnode list to peek
280 static struct bio *throtl_peek_queued(struct list_head *queued)
282 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
285 if (list_empty(queued))
288 bio = bio_list_peek(&qn->bios);
294 * throtl_pop_queued - pop the first bio form a qnode list
295 * @queued: the qnode list to pop a bio from
296 * @tg_to_put: optional out argument for throtl_grp to put
298 * Pop the first bio from the qnode list @queued. After popping, the first
299 * qnode is removed from @queued if empty or moved to the end of @queued so
300 * that the popping order is round-robin.
302 * When the first qnode is removed, its associated throtl_grp should be put
303 * too. If @tg_to_put is NULL, this function automatically puts it;
304 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
305 * responsible for putting it.
307 static struct bio *throtl_pop_queued(struct list_head *queued,
308 struct throtl_grp **tg_to_put)
310 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
313 if (list_empty(queued))
316 bio = bio_list_pop(&qn->bios);
319 if (bio_list_empty(&qn->bios)) {
320 list_del_init(&qn->node);
324 blkg_put(tg_to_blkg(qn->tg));
326 list_move_tail(&qn->node, queued);
332 /* init a service_queue, assumes the caller zeroed it */
333 static void throtl_service_queue_init(struct throtl_service_queue *sq,
334 struct throtl_service_queue *parent_sq)
336 INIT_LIST_HEAD(&sq->queued[0]);
337 INIT_LIST_HEAD(&sq->queued[1]);
338 sq->pending_tree = RB_ROOT;
339 sq->parent_sq = parent_sq;
340 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
344 static void throtl_service_queue_exit(struct throtl_service_queue *sq)
346 del_timer_sync(&sq->pending_timer);
349 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
351 struct throtl_grp *tg;
354 tg = kzalloc_node(sizeof(*tg), gfp, node);
358 tg->stats_cpu = alloc_percpu_gfp(struct tg_stats_cpu, gfp);
359 if (!tg->stats_cpu) {
364 for_each_possible_cpu(cpu) {
365 struct tg_stats_cpu *stats_cpu = per_cpu_ptr(tg->stats_cpu, cpu);
367 blkg_rwstat_init(&stats_cpu->service_bytes);
368 blkg_rwstat_init(&stats_cpu->serviced);
374 static void throtl_pd_init(struct blkcg_gq *blkg)
376 struct throtl_grp *tg = blkg_to_tg(blkg);
377 struct throtl_data *td = blkg->q->td;
378 struct throtl_service_queue *parent_sq;
382 * If on the default hierarchy, we switch to properly hierarchical
383 * behavior where limits on a given throtl_grp are applied to the
384 * whole subtree rather than just the group itself. e.g. If 16M
385 * read_bps limit is set on the root group, the whole system can't
386 * exceed 16M for the device.
388 * If not on the default hierarchy, the broken flat hierarchy
389 * behavior is retained where all throtl_grps are treated as if
390 * they're all separate root groups right below throtl_data.
391 * Limits of a group don't interact with limits of other groups
392 * regardless of the position of the group in the hierarchy.
394 parent_sq = &td->service_queue;
396 if (cgroup_on_dfl(blkg->blkcg->css.cgroup) && blkg->parent)
397 parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
399 throtl_service_queue_init(&tg->service_queue, parent_sq);
401 for (rw = READ; rw <= WRITE; rw++) {
402 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
403 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
406 RB_CLEAR_NODE(&tg->rb_node);
412 tg->iops[WRITE] = -1;
416 * Set has_rules[] if @tg or any of its parents have limits configured.
417 * This doesn't require walking up to the top of the hierarchy as the
418 * parent's has_rules[] is guaranteed to be correct.
420 static void tg_update_has_rules(struct throtl_grp *tg)
422 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
425 for (rw = READ; rw <= WRITE; rw++)
426 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
427 (tg->bps[rw] != -1 || tg->iops[rw] != -1);
430 static void throtl_pd_online(struct blkcg_gq *blkg)
433 * We don't want new groups to escape the limits of its ancestors.
434 * Update has_rules[] after a new group is brought online.
436 tg_update_has_rules(blkg_to_tg(blkg));
439 static void throtl_pd_exit(struct blkcg_gq *blkg)
441 struct throtl_grp *tg = blkg_to_tg(blkg);
443 throtl_service_queue_exit(&tg->service_queue);
446 static void throtl_pd_free(struct blkg_policy_data *pd)
448 struct throtl_grp *tg = pd_to_tg(pd);
450 free_percpu(tg->stats_cpu);
454 static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
456 struct throtl_grp *tg = blkg_to_tg(blkg);
459 for_each_possible_cpu(cpu) {
460 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
462 blkg_rwstat_reset(&sc->service_bytes);
463 blkg_rwstat_reset(&sc->serviced);
467 static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
471 * This is the common case when there are no blkcgs. Avoid lookup
474 if (blkcg == &blkcg_root)
475 return td_root_tg(td);
477 return blkg_to_tg(blkg_lookup(blkcg, td->queue));
480 static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
483 struct request_queue *q = td->queue;
484 struct throtl_grp *tg = NULL;
487 * This is the common case when there are no blkcgs. Avoid lookup
490 if (blkcg == &blkcg_root) {
493 struct blkcg_gq *blkg;
495 blkg = blkg_lookup_create(blkcg, q);
497 /* if %NULL and @q is alive, fall back to root_tg */
499 tg = blkg_to_tg(blkg);
500 else if (!blk_queue_dying(q))
507 static struct throtl_grp *
508 throtl_rb_first(struct throtl_service_queue *parent_sq)
510 /* Service tree is empty */
511 if (!parent_sq->nr_pending)
514 if (!parent_sq->first_pending)
515 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
517 if (parent_sq->first_pending)
518 return rb_entry_tg(parent_sq->first_pending);
523 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
529 static void throtl_rb_erase(struct rb_node *n,
530 struct throtl_service_queue *parent_sq)
532 if (parent_sq->first_pending == n)
533 parent_sq->first_pending = NULL;
534 rb_erase_init(n, &parent_sq->pending_tree);
535 --parent_sq->nr_pending;
538 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
540 struct throtl_grp *tg;
542 tg = throtl_rb_first(parent_sq);
546 parent_sq->first_pending_disptime = tg->disptime;
549 static void tg_service_queue_add(struct throtl_grp *tg)
551 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
552 struct rb_node **node = &parent_sq->pending_tree.rb_node;
553 struct rb_node *parent = NULL;
554 struct throtl_grp *__tg;
555 unsigned long key = tg->disptime;
558 while (*node != NULL) {
560 __tg = rb_entry_tg(parent);
562 if (time_before(key, __tg->disptime))
563 node = &parent->rb_left;
565 node = &parent->rb_right;
571 parent_sq->first_pending = &tg->rb_node;
573 rb_link_node(&tg->rb_node, parent, node);
574 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
577 static void __throtl_enqueue_tg(struct throtl_grp *tg)
579 tg_service_queue_add(tg);
580 tg->flags |= THROTL_TG_PENDING;
581 tg->service_queue.parent_sq->nr_pending++;
584 static void throtl_enqueue_tg(struct throtl_grp *tg)
586 if (!(tg->flags & THROTL_TG_PENDING))
587 __throtl_enqueue_tg(tg);
590 static void __throtl_dequeue_tg(struct throtl_grp *tg)
592 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
593 tg->flags &= ~THROTL_TG_PENDING;
596 static void throtl_dequeue_tg(struct throtl_grp *tg)
598 if (tg->flags & THROTL_TG_PENDING)
599 __throtl_dequeue_tg(tg);
602 /* Call with queue lock held */
603 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
604 unsigned long expires)
606 mod_timer(&sq->pending_timer, expires);
607 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
608 expires - jiffies, jiffies);
612 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
613 * @sq: the service_queue to schedule dispatch for
614 * @force: force scheduling
616 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
617 * dispatch time of the first pending child. Returns %true if either timer
618 * is armed or there's no pending child left. %false if the current
619 * dispatch window is still open and the caller should continue
622 * If @force is %true, the dispatch timer is always scheduled and this
623 * function is guaranteed to return %true. This is to be used when the
624 * caller can't dispatch itself and needs to invoke pending_timer
625 * unconditionally. Note that forced scheduling is likely to induce short
626 * delay before dispatch starts even if @sq->first_pending_disptime is not
627 * in the future and thus shouldn't be used in hot paths.
629 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
632 /* any pending children left? */
636 update_min_dispatch_time(sq);
638 /* is the next dispatch time in the future? */
639 if (force || time_after(sq->first_pending_disptime, jiffies)) {
640 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
644 /* tell the caller to continue dispatching */
648 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
649 bool rw, unsigned long start)
651 tg->bytes_disp[rw] = 0;
655 * Previous slice has expired. We must have trimmed it after last
656 * bio dispatch. That means since start of last slice, we never used
657 * that bandwidth. Do try to make use of that bandwidth while giving
660 if (time_after_eq(start, tg->slice_start[rw]))
661 tg->slice_start[rw] = start;
663 tg->slice_end[rw] = jiffies + throtl_slice;
664 throtl_log(&tg->service_queue,
665 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
666 rw == READ ? 'R' : 'W', tg->slice_start[rw],
667 tg->slice_end[rw], jiffies);
670 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
672 tg->bytes_disp[rw] = 0;
674 tg->slice_start[rw] = jiffies;
675 tg->slice_end[rw] = jiffies + throtl_slice;
676 throtl_log(&tg->service_queue,
677 "[%c] new slice start=%lu end=%lu jiffies=%lu",
678 rw == READ ? 'R' : 'W', tg->slice_start[rw],
679 tg->slice_end[rw], jiffies);
682 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
683 unsigned long jiffy_end)
685 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
688 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
689 unsigned long jiffy_end)
691 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
692 throtl_log(&tg->service_queue,
693 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
694 rw == READ ? 'R' : 'W', tg->slice_start[rw],
695 tg->slice_end[rw], jiffies);
698 /* Determine if previously allocated or extended slice is complete or not */
699 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
701 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
707 /* Trim the used slices and adjust slice start accordingly */
708 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
710 unsigned long nr_slices, time_elapsed, io_trim;
713 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
716 * If bps are unlimited (-1), then time slice don't get
717 * renewed. Don't try to trim the slice if slice is used. A new
718 * slice will start when appropriate.
720 if (throtl_slice_used(tg, rw))
724 * A bio has been dispatched. Also adjust slice_end. It might happen
725 * that initially cgroup limit was very low resulting in high
726 * slice_end, but later limit was bumped up and bio was dispached
727 * sooner, then we need to reduce slice_end. A high bogus slice_end
728 * is bad because it does not allow new slice to start.
731 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
733 time_elapsed = jiffies - tg->slice_start[rw];
735 nr_slices = time_elapsed / throtl_slice;
739 tmp = tg->bps[rw] * throtl_slice * nr_slices;
743 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
745 if (!bytes_trim && !io_trim)
748 if (tg->bytes_disp[rw] >= bytes_trim)
749 tg->bytes_disp[rw] -= bytes_trim;
751 tg->bytes_disp[rw] = 0;
753 if (tg->io_disp[rw] >= io_trim)
754 tg->io_disp[rw] -= io_trim;
758 tg->slice_start[rw] += nr_slices * throtl_slice;
760 throtl_log(&tg->service_queue,
761 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
762 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
763 tg->slice_start[rw], tg->slice_end[rw], jiffies);
766 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
769 bool rw = bio_data_dir(bio);
770 unsigned int io_allowed;
771 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
774 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
776 /* Slice has just started. Consider one slice interval */
778 jiffy_elapsed_rnd = throtl_slice;
780 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
783 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
784 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
785 * will allow dispatch after 1 second and after that slice should
789 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
793 io_allowed = UINT_MAX;
797 if (tg->io_disp[rw] + 1 <= io_allowed) {
803 /* Calc approx time to dispatch */
804 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
806 if (jiffy_wait > jiffy_elapsed)
807 jiffy_wait = jiffy_wait - jiffy_elapsed;
816 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
819 bool rw = bio_data_dir(bio);
820 u64 bytes_allowed, extra_bytes, tmp;
821 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
823 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
825 /* Slice has just started. Consider one slice interval */
827 jiffy_elapsed_rnd = throtl_slice;
829 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
831 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
835 if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
841 /* Calc approx time to dispatch */
842 extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
843 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
849 * This wait time is without taking into consideration the rounding
850 * up we did. Add that time also.
852 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
859 * Returns whether one can dispatch a bio or not. Also returns approx number
860 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
862 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
865 bool rw = bio_data_dir(bio);
866 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
869 * Currently whole state machine of group depends on first bio
870 * queued in the group bio list. So one should not be calling
871 * this function with a different bio if there are other bios
874 BUG_ON(tg->service_queue.nr_queued[rw] &&
875 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
877 /* If tg->bps = -1, then BW is unlimited */
878 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
885 * If previous slice expired, start a new one otherwise renew/extend
886 * existing slice to make sure it is at least throtl_slice interval
889 if (throtl_slice_used(tg, rw))
890 throtl_start_new_slice(tg, rw);
892 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
893 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
896 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
897 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
903 max_wait = max(bps_wait, iops_wait);
908 if (time_before(tg->slice_end[rw], jiffies + max_wait))
909 throtl_extend_slice(tg, rw, jiffies + max_wait);
914 static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
917 struct throtl_grp *tg = blkg_to_tg(blkg);
918 struct tg_stats_cpu *stats_cpu;
922 * Disabling interrupts to provide mutual exclusion between two
923 * writes on same cpu. It probably is not needed for 64bit. Not
924 * optimizing that case yet.
926 local_irq_save(flags);
928 stats_cpu = this_cpu_ptr(tg->stats_cpu);
930 blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
931 blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
933 local_irq_restore(flags);
936 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
938 bool rw = bio_data_dir(bio);
940 /* Charge the bio to the group */
941 tg->bytes_disp[rw] += bio->bi_iter.bi_size;
945 * REQ_THROTTLED is used to prevent the same bio to be throttled
946 * more than once as a throttled bio will go through blk-throtl the
947 * second time when it eventually gets issued. Set it when a bio
948 * is being charged to a tg.
950 * Dispatch stats aren't recursive and each @bio should only be
951 * accounted by the @tg it was originally associated with. Let's
952 * update the stats when setting REQ_THROTTLED for the first time
953 * which is guaranteed to be for the @bio's original tg.
955 if (!(bio->bi_rw & REQ_THROTTLED)) {
956 bio->bi_rw |= REQ_THROTTLED;
957 throtl_update_dispatch_stats(tg_to_blkg(tg),
958 bio->bi_iter.bi_size, bio->bi_rw);
963 * throtl_add_bio_tg - add a bio to the specified throtl_grp
966 * @tg: the target throtl_grp
968 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
969 * tg->qnode_on_self[] is used.
971 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
972 struct throtl_grp *tg)
974 struct throtl_service_queue *sq = &tg->service_queue;
975 bool rw = bio_data_dir(bio);
978 qn = &tg->qnode_on_self[rw];
981 * If @tg doesn't currently have any bios queued in the same
982 * direction, queueing @bio can change when @tg should be
983 * dispatched. Mark that @tg was empty. This is automatically
984 * cleaered on the next tg_update_disptime().
986 if (!sq->nr_queued[rw])
987 tg->flags |= THROTL_TG_WAS_EMPTY;
989 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
992 throtl_enqueue_tg(tg);
995 static void tg_update_disptime(struct throtl_grp *tg)
997 struct throtl_service_queue *sq = &tg->service_queue;
998 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1001 if ((bio = throtl_peek_queued(&sq->queued[READ])))
1002 tg_may_dispatch(tg, bio, &read_wait);
1004 if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1005 tg_may_dispatch(tg, bio, &write_wait);
1007 min_wait = min(read_wait, write_wait);
1008 disptime = jiffies + min_wait;
1010 /* Update dispatch time */
1011 throtl_dequeue_tg(tg);
1012 tg->disptime = disptime;
1013 throtl_enqueue_tg(tg);
1015 /* see throtl_add_bio_tg() */
1016 tg->flags &= ~THROTL_TG_WAS_EMPTY;
1019 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1020 struct throtl_grp *parent_tg, bool rw)
1022 if (throtl_slice_used(parent_tg, rw)) {
1023 throtl_start_new_slice_with_credit(parent_tg, rw,
1024 child_tg->slice_start[rw]);
1029 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1031 struct throtl_service_queue *sq = &tg->service_queue;
1032 struct throtl_service_queue *parent_sq = sq->parent_sq;
1033 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1034 struct throtl_grp *tg_to_put = NULL;
1038 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1039 * from @tg may put its reference and @parent_sq might end up
1040 * getting released prematurely. Remember the tg to put and put it
1041 * after @bio is transferred to @parent_sq.
1043 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1044 sq->nr_queued[rw]--;
1046 throtl_charge_bio(tg, bio);
1049 * If our parent is another tg, we just need to transfer @bio to
1050 * the parent using throtl_add_bio_tg(). If our parent is
1051 * @td->service_queue, @bio is ready to be issued. Put it on its
1052 * bio_lists[] and decrease total number queued. The caller is
1053 * responsible for issuing these bios.
1056 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1057 start_parent_slice_with_credit(tg, parent_tg, rw);
1059 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1060 &parent_sq->queued[rw]);
1061 BUG_ON(tg->td->nr_queued[rw] <= 0);
1062 tg->td->nr_queued[rw]--;
1065 throtl_trim_slice(tg, rw);
1068 blkg_put(tg_to_blkg(tg_to_put));
1071 static int throtl_dispatch_tg(struct throtl_grp *tg)
1073 struct throtl_service_queue *sq = &tg->service_queue;
1074 unsigned int nr_reads = 0, nr_writes = 0;
1075 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1076 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1079 /* Try to dispatch 75% READS and 25% WRITES */
1081 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1082 tg_may_dispatch(tg, bio, NULL)) {
1084 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1087 if (nr_reads >= max_nr_reads)
1091 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1092 tg_may_dispatch(tg, bio, NULL)) {
1094 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1097 if (nr_writes >= max_nr_writes)
1101 return nr_reads + nr_writes;
1104 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1106 unsigned int nr_disp = 0;
1109 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1110 struct throtl_service_queue *sq = &tg->service_queue;
1115 if (time_before(jiffies, tg->disptime))
1118 throtl_dequeue_tg(tg);
1120 nr_disp += throtl_dispatch_tg(tg);
1122 if (sq->nr_queued[0] || sq->nr_queued[1])
1123 tg_update_disptime(tg);
1125 if (nr_disp >= throtl_quantum)
1133 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1134 * @arg: the throtl_service_queue being serviced
1136 * This timer is armed when a child throtl_grp with active bio's become
1137 * pending and queued on the service_queue's pending_tree and expires when
1138 * the first child throtl_grp should be dispatched. This function
1139 * dispatches bio's from the children throtl_grps to the parent
1142 * If the parent's parent is another throtl_grp, dispatching is propagated
1143 * by either arming its pending_timer or repeating dispatch directly. If
1144 * the top-level service_tree is reached, throtl_data->dispatch_work is
1145 * kicked so that the ready bio's are issued.
1147 static void throtl_pending_timer_fn(unsigned long arg)
1149 struct throtl_service_queue *sq = (void *)arg;
1150 struct throtl_grp *tg = sq_to_tg(sq);
1151 struct throtl_data *td = sq_to_td(sq);
1152 struct request_queue *q = td->queue;
1153 struct throtl_service_queue *parent_sq;
1157 spin_lock_irq(q->queue_lock);
1159 parent_sq = sq->parent_sq;
1163 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1164 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1165 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1167 ret = throtl_select_dispatch(sq);
1169 throtl_log(sq, "bios disp=%u", ret);
1173 if (throtl_schedule_next_dispatch(sq, false))
1176 /* this dispatch windows is still open, relax and repeat */
1177 spin_unlock_irq(q->queue_lock);
1179 spin_lock_irq(q->queue_lock);
1186 /* @parent_sq is another throl_grp, propagate dispatch */
1187 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1188 tg_update_disptime(tg);
1189 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1190 /* window is already open, repeat dispatching */
1197 /* reached the top-level, queue issueing */
1198 queue_work(kthrotld_workqueue, &td->dispatch_work);
1201 spin_unlock_irq(q->queue_lock);
1205 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1206 * @work: work item being executed
1208 * This function is queued for execution when bio's reach the bio_lists[]
1209 * of throtl_data->service_queue. Those bio's are ready and issued by this
1212 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1214 struct throtl_data *td = container_of(work, struct throtl_data,
1216 struct throtl_service_queue *td_sq = &td->service_queue;
1217 struct request_queue *q = td->queue;
1218 struct bio_list bio_list_on_stack;
1220 struct blk_plug plug;
1223 bio_list_init(&bio_list_on_stack);
1225 spin_lock_irq(q->queue_lock);
1226 for (rw = READ; rw <= WRITE; rw++)
1227 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1228 bio_list_add(&bio_list_on_stack, bio);
1229 spin_unlock_irq(q->queue_lock);
1231 if (!bio_list_empty(&bio_list_on_stack)) {
1232 blk_start_plug(&plug);
1233 while((bio = bio_list_pop(&bio_list_on_stack)))
1234 generic_make_request(bio);
1235 blk_finish_plug(&plug);
1239 static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
1240 struct blkg_policy_data *pd, int off)
1242 struct throtl_grp *tg = pd_to_tg(pd);
1243 struct blkg_rwstat rwstat = { }, tmp;
1246 for_each_possible_cpu(cpu) {
1247 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
1249 tmp = blkg_rwstat_read((void *)sc + off);
1250 for (i = 0; i < BLKG_RWSTAT_NR; i++)
1251 rwstat.cnt[i] += tmp.cnt[i];
1254 return __blkg_prfill_rwstat(sf, pd, &rwstat);
1257 static int tg_print_cpu_rwstat(struct seq_file *sf, void *v)
1259 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_cpu_rwstat,
1260 &blkcg_policy_throtl, seq_cft(sf)->private, true);
1264 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1267 struct throtl_grp *tg = pd_to_tg(pd);
1268 u64 v = *(u64 *)((void *)tg + off);
1272 return __blkg_prfill_u64(sf, pd, v);
1275 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1278 struct throtl_grp *tg = pd_to_tg(pd);
1279 unsigned int v = *(unsigned int *)((void *)tg + off);
1283 return __blkg_prfill_u64(sf, pd, v);
1286 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1288 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1289 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1293 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1295 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1296 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1300 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1301 char *buf, size_t nbytes, loff_t off, bool is_u64)
1303 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1304 struct blkg_conf_ctx ctx;
1305 struct throtl_grp *tg;
1306 struct throtl_service_queue *sq;
1307 struct blkcg_gq *blkg;
1308 struct cgroup_subsys_state *pos_css;
1311 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1315 tg = blkg_to_tg(ctx.blkg);
1316 sq = &tg->service_queue;
1322 *(u64 *)((void *)tg + of_cft(of)->private) = ctx.v;
1324 *(unsigned int *)((void *)tg + of_cft(of)->private) = ctx.v;
1326 throtl_log(&tg->service_queue,
1327 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1328 tg->bps[READ], tg->bps[WRITE],
1329 tg->iops[READ], tg->iops[WRITE]);
1332 * Update has_rules[] flags for the updated tg's subtree. A tg is
1333 * considered to have rules if either the tg itself or any of its
1334 * ancestors has rules. This identifies groups without any
1335 * restrictions in the whole hierarchy and allows them to bypass
1338 blkg_for_each_descendant_pre(blkg, pos_css, ctx.blkg)
1339 tg_update_has_rules(blkg_to_tg(blkg));
1342 * We're already holding queue_lock and know @tg is valid. Let's
1343 * apply the new config directly.
1345 * Restart the slices for both READ and WRITES. It might happen
1346 * that a group's limit are dropped suddenly and we don't want to
1347 * account recently dispatched IO with new low rate.
1349 throtl_start_new_slice(tg, 0);
1350 throtl_start_new_slice(tg, 1);
1352 if (tg->flags & THROTL_TG_PENDING) {
1353 tg_update_disptime(tg);
1354 throtl_schedule_next_dispatch(sq->parent_sq, true);
1357 blkg_conf_finish(&ctx);
1361 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1362 char *buf, size_t nbytes, loff_t off)
1364 return tg_set_conf(of, buf, nbytes, off, true);
1367 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1368 char *buf, size_t nbytes, loff_t off)
1370 return tg_set_conf(of, buf, nbytes, off, false);
1373 static struct cftype throtl_files[] = {
1375 .name = "throttle.read_bps_device",
1376 .private = offsetof(struct throtl_grp, bps[READ]),
1377 .seq_show = tg_print_conf_u64,
1378 .write = tg_set_conf_u64,
1381 .name = "throttle.write_bps_device",
1382 .private = offsetof(struct throtl_grp, bps[WRITE]),
1383 .seq_show = tg_print_conf_u64,
1384 .write = tg_set_conf_u64,
1387 .name = "throttle.read_iops_device",
1388 .private = offsetof(struct throtl_grp, iops[READ]),
1389 .seq_show = tg_print_conf_uint,
1390 .write = tg_set_conf_uint,
1393 .name = "throttle.write_iops_device",
1394 .private = offsetof(struct throtl_grp, iops[WRITE]),
1395 .seq_show = tg_print_conf_uint,
1396 .write = tg_set_conf_uint,
1399 .name = "throttle.io_service_bytes",
1400 .private = offsetof(struct tg_stats_cpu, service_bytes),
1401 .seq_show = tg_print_cpu_rwstat,
1404 .name = "throttle.io_serviced",
1405 .private = offsetof(struct tg_stats_cpu, serviced),
1406 .seq_show = tg_print_cpu_rwstat,
1411 static void throtl_shutdown_wq(struct request_queue *q)
1413 struct throtl_data *td = q->td;
1415 cancel_work_sync(&td->dispatch_work);
1418 static struct blkcg_policy blkcg_policy_throtl = {
1419 .cftypes = throtl_files,
1421 .pd_alloc_fn = throtl_pd_alloc,
1422 .pd_init_fn = throtl_pd_init,
1423 .pd_online_fn = throtl_pd_online,
1424 .pd_exit_fn = throtl_pd_exit,
1425 .pd_free_fn = throtl_pd_free,
1426 .pd_reset_stats_fn = throtl_pd_reset_stats,
1429 bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
1431 struct throtl_data *td = q->td;
1432 struct throtl_qnode *qn = NULL;
1433 struct throtl_grp *tg;
1434 struct throtl_service_queue *sq;
1435 bool rw = bio_data_dir(bio);
1436 struct blkcg *blkcg;
1437 bool throttled = false;
1439 /* see throtl_charge_bio() */
1440 if (bio->bi_rw & REQ_THROTTLED)
1444 * A throtl_grp pointer retrieved under rcu can be used to access
1445 * basic fields like stats and io rates. If a group has no rules,
1446 * just update the dispatch stats in lockless manner and return.
1449 blkcg = bio_blkcg(bio);
1450 tg = throtl_lookup_tg(td, blkcg);
1452 if (!tg->has_rules[rw]) {
1453 throtl_update_dispatch_stats(tg_to_blkg(tg),
1454 bio->bi_iter.bi_size, bio->bi_rw);
1455 goto out_unlock_rcu;
1460 * Either group has not been allocated yet or it is not an unlimited
1463 spin_lock_irq(q->queue_lock);
1464 tg = throtl_lookup_create_tg(td, blkcg);
1468 sq = &tg->service_queue;
1471 /* throtl is FIFO - if bios are already queued, should queue */
1472 if (sq->nr_queued[rw])
1475 /* if above limits, break to queue */
1476 if (!tg_may_dispatch(tg, bio, NULL))
1479 /* within limits, let's charge and dispatch directly */
1480 throtl_charge_bio(tg, bio);
1483 * We need to trim slice even when bios are not being queued
1484 * otherwise it might happen that a bio is not queued for
1485 * a long time and slice keeps on extending and trim is not
1486 * called for a long time. Now if limits are reduced suddenly
1487 * we take into account all the IO dispatched so far at new
1488 * low rate and * newly queued IO gets a really long dispatch
1491 * So keep on trimming slice even if bio is not queued.
1493 throtl_trim_slice(tg, rw);
1496 * @bio passed through this layer without being throttled.
1497 * Climb up the ladder. If we''re already at the top, it
1498 * can be executed directly.
1500 qn = &tg->qnode_on_parent[rw];
1507 /* out-of-limit, queue to @tg */
1508 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1509 rw == READ ? 'R' : 'W',
1510 tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw],
1511 tg->io_disp[rw], tg->iops[rw],
1512 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1514 bio_associate_current(bio);
1515 tg->td->nr_queued[rw]++;
1516 throtl_add_bio_tg(bio, qn, tg);
1520 * Update @tg's dispatch time and force schedule dispatch if @tg
1521 * was empty before @bio. The forced scheduling isn't likely to
1522 * cause undue delay as @bio is likely to be dispatched directly if
1523 * its @tg's disptime is not in the future.
1525 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1526 tg_update_disptime(tg);
1527 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1531 spin_unlock_irq(q->queue_lock);
1536 * As multiple blk-throtls may stack in the same issue path, we
1537 * don't want bios to leave with the flag set. Clear the flag if
1541 bio->bi_rw &= ~REQ_THROTTLED;
1546 * Dispatch all bios from all children tg's queued on @parent_sq. On
1547 * return, @parent_sq is guaranteed to not have any active children tg's
1548 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1550 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1552 struct throtl_grp *tg;
1554 while ((tg = throtl_rb_first(parent_sq))) {
1555 struct throtl_service_queue *sq = &tg->service_queue;
1558 throtl_dequeue_tg(tg);
1560 while ((bio = throtl_peek_queued(&sq->queued[READ])))
1561 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1562 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1563 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1568 * blk_throtl_drain - drain throttled bios
1569 * @q: request_queue to drain throttled bios for
1571 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1573 void blk_throtl_drain(struct request_queue *q)
1574 __releases(q->queue_lock) __acquires(q->queue_lock)
1576 struct throtl_data *td = q->td;
1577 struct blkcg_gq *blkg;
1578 struct cgroup_subsys_state *pos_css;
1582 queue_lockdep_assert_held(q);
1586 * Drain each tg while doing post-order walk on the blkg tree, so
1587 * that all bios are propagated to td->service_queue. It'd be
1588 * better to walk service_queue tree directly but blkg walk is
1591 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
1592 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1594 /* finally, transfer bios from top-level tg's into the td */
1595 tg_drain_bios(&td->service_queue);
1598 spin_unlock_irq(q->queue_lock);
1600 /* all bios now should be in td->service_queue, issue them */
1601 for (rw = READ; rw <= WRITE; rw++)
1602 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1604 generic_make_request(bio);
1606 spin_lock_irq(q->queue_lock);
1609 int blk_throtl_init(struct request_queue *q)
1611 struct throtl_data *td;
1614 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1618 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1619 throtl_service_queue_init(&td->service_queue, NULL);
1624 /* activate policy */
1625 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1631 void blk_throtl_exit(struct request_queue *q)
1634 throtl_shutdown_wq(q);
1635 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1639 static int __init throtl_init(void)
1641 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1642 if (!kthrotld_workqueue)
1643 panic("Failed to create kthrotld\n");
1645 return blkcg_policy_register(&blkcg_policy_throtl);
1648 module_init(throtl_init);