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)
335 INIT_LIST_HEAD(&sq->queued[0]);
336 INIT_LIST_HEAD(&sq->queued[1]);
337 sq->pending_tree = RB_ROOT;
338 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
342 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
344 struct throtl_grp *tg;
347 tg = kzalloc_node(sizeof(*tg), gfp, node);
351 tg->stats_cpu = alloc_percpu_gfp(struct tg_stats_cpu, gfp);
352 if (!tg->stats_cpu) {
357 throtl_service_queue_init(&tg->service_queue);
359 for (rw = READ; rw <= WRITE; rw++) {
360 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
361 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
364 RB_CLEAR_NODE(&tg->rb_node);
368 tg->iops[WRITE] = -1;
370 for_each_possible_cpu(cpu) {
371 struct tg_stats_cpu *stats_cpu = per_cpu_ptr(tg->stats_cpu, cpu);
373 blkg_rwstat_init(&stats_cpu->service_bytes);
374 blkg_rwstat_init(&stats_cpu->serviced);
380 static void throtl_pd_init(struct blkcg_gq *blkg)
382 struct throtl_grp *tg = blkg_to_tg(blkg);
383 struct throtl_data *td = blkg->q->td;
384 struct throtl_service_queue *sq = &tg->service_queue;
387 * If on the default hierarchy, we switch to properly hierarchical
388 * behavior where limits on a given throtl_grp are applied to the
389 * whole subtree rather than just the group itself. e.g. If 16M
390 * read_bps limit is set on the root group, the whole system can't
391 * exceed 16M for the device.
393 * If not on the default hierarchy, the broken flat hierarchy
394 * behavior is retained where all throtl_grps are treated as if
395 * they're all separate root groups right below throtl_data.
396 * Limits of a group don't interact with limits of other groups
397 * regardless of the position of the group in the hierarchy.
399 sq->parent_sq = &td->service_queue;
400 if (cgroup_on_dfl(blkg->blkcg->css.cgroup) && blkg->parent)
401 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
406 * Set has_rules[] if @tg or any of its parents have limits configured.
407 * This doesn't require walking up to the top of the hierarchy as the
408 * parent's has_rules[] is guaranteed to be correct.
410 static void tg_update_has_rules(struct throtl_grp *tg)
412 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
415 for (rw = READ; rw <= WRITE; rw++)
416 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
417 (tg->bps[rw] != -1 || tg->iops[rw] != -1);
420 static void throtl_pd_online(struct blkcg_gq *blkg)
423 * We don't want new groups to escape the limits of its ancestors.
424 * Update has_rules[] after a new group is brought online.
426 tg_update_has_rules(blkg_to_tg(blkg));
429 static void throtl_pd_free(struct blkg_policy_data *pd)
431 struct throtl_grp *tg = pd_to_tg(pd);
433 del_timer_sync(&tg->service_queue.pending_timer);
434 free_percpu(tg->stats_cpu);
438 static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
440 struct throtl_grp *tg = blkg_to_tg(blkg);
443 for_each_possible_cpu(cpu) {
444 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
446 blkg_rwstat_reset(&sc->service_bytes);
447 blkg_rwstat_reset(&sc->serviced);
451 static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
455 * This is the common case when there are no blkcgs. Avoid lookup
458 if (blkcg == &blkcg_root)
459 return td_root_tg(td);
461 return blkg_to_tg(blkg_lookup(blkcg, td->queue));
464 static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
467 struct request_queue *q = td->queue;
468 struct throtl_grp *tg = NULL;
471 * This is the common case when there are no blkcgs. Avoid lookup
474 if (blkcg == &blkcg_root) {
477 struct blkcg_gq *blkg;
479 blkg = blkg_lookup_create(blkcg, q);
481 /* if %NULL and @q is alive, fall back to root_tg */
483 tg = blkg_to_tg(blkg);
484 else if (!blk_queue_dying(q))
491 static struct throtl_grp *
492 throtl_rb_first(struct throtl_service_queue *parent_sq)
494 /* Service tree is empty */
495 if (!parent_sq->nr_pending)
498 if (!parent_sq->first_pending)
499 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
501 if (parent_sq->first_pending)
502 return rb_entry_tg(parent_sq->first_pending);
507 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
513 static void throtl_rb_erase(struct rb_node *n,
514 struct throtl_service_queue *parent_sq)
516 if (parent_sq->first_pending == n)
517 parent_sq->first_pending = NULL;
518 rb_erase_init(n, &parent_sq->pending_tree);
519 --parent_sq->nr_pending;
522 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
524 struct throtl_grp *tg;
526 tg = throtl_rb_first(parent_sq);
530 parent_sq->first_pending_disptime = tg->disptime;
533 static void tg_service_queue_add(struct throtl_grp *tg)
535 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
536 struct rb_node **node = &parent_sq->pending_tree.rb_node;
537 struct rb_node *parent = NULL;
538 struct throtl_grp *__tg;
539 unsigned long key = tg->disptime;
542 while (*node != NULL) {
544 __tg = rb_entry_tg(parent);
546 if (time_before(key, __tg->disptime))
547 node = &parent->rb_left;
549 node = &parent->rb_right;
555 parent_sq->first_pending = &tg->rb_node;
557 rb_link_node(&tg->rb_node, parent, node);
558 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
561 static void __throtl_enqueue_tg(struct throtl_grp *tg)
563 tg_service_queue_add(tg);
564 tg->flags |= THROTL_TG_PENDING;
565 tg->service_queue.parent_sq->nr_pending++;
568 static void throtl_enqueue_tg(struct throtl_grp *tg)
570 if (!(tg->flags & THROTL_TG_PENDING))
571 __throtl_enqueue_tg(tg);
574 static void __throtl_dequeue_tg(struct throtl_grp *tg)
576 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
577 tg->flags &= ~THROTL_TG_PENDING;
580 static void throtl_dequeue_tg(struct throtl_grp *tg)
582 if (tg->flags & THROTL_TG_PENDING)
583 __throtl_dequeue_tg(tg);
586 /* Call with queue lock held */
587 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
588 unsigned long expires)
590 mod_timer(&sq->pending_timer, expires);
591 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
592 expires - jiffies, jiffies);
596 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
597 * @sq: the service_queue to schedule dispatch for
598 * @force: force scheduling
600 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
601 * dispatch time of the first pending child. Returns %true if either timer
602 * is armed or there's no pending child left. %false if the current
603 * dispatch window is still open and the caller should continue
606 * If @force is %true, the dispatch timer is always scheduled and this
607 * function is guaranteed to return %true. This is to be used when the
608 * caller can't dispatch itself and needs to invoke pending_timer
609 * unconditionally. Note that forced scheduling is likely to induce short
610 * delay before dispatch starts even if @sq->first_pending_disptime is not
611 * in the future and thus shouldn't be used in hot paths.
613 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
616 /* any pending children left? */
620 update_min_dispatch_time(sq);
622 /* is the next dispatch time in the future? */
623 if (force || time_after(sq->first_pending_disptime, jiffies)) {
624 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
628 /* tell the caller to continue dispatching */
632 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
633 bool rw, unsigned long start)
635 tg->bytes_disp[rw] = 0;
639 * Previous slice has expired. We must have trimmed it after last
640 * bio dispatch. That means since start of last slice, we never used
641 * that bandwidth. Do try to make use of that bandwidth while giving
644 if (time_after_eq(start, tg->slice_start[rw]))
645 tg->slice_start[rw] = start;
647 tg->slice_end[rw] = jiffies + throtl_slice;
648 throtl_log(&tg->service_queue,
649 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
650 rw == READ ? 'R' : 'W', tg->slice_start[rw],
651 tg->slice_end[rw], jiffies);
654 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
656 tg->bytes_disp[rw] = 0;
658 tg->slice_start[rw] = jiffies;
659 tg->slice_end[rw] = jiffies + throtl_slice;
660 throtl_log(&tg->service_queue,
661 "[%c] new slice start=%lu end=%lu jiffies=%lu",
662 rw == READ ? 'R' : 'W', tg->slice_start[rw],
663 tg->slice_end[rw], jiffies);
666 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
667 unsigned long jiffy_end)
669 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
672 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
673 unsigned long jiffy_end)
675 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
676 throtl_log(&tg->service_queue,
677 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
678 rw == READ ? 'R' : 'W', tg->slice_start[rw],
679 tg->slice_end[rw], jiffies);
682 /* Determine if previously allocated or extended slice is complete or not */
683 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
685 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
691 /* Trim the used slices and adjust slice start accordingly */
692 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
694 unsigned long nr_slices, time_elapsed, io_trim;
697 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
700 * If bps are unlimited (-1), then time slice don't get
701 * renewed. Don't try to trim the slice if slice is used. A new
702 * slice will start when appropriate.
704 if (throtl_slice_used(tg, rw))
708 * A bio has been dispatched. Also adjust slice_end. It might happen
709 * that initially cgroup limit was very low resulting in high
710 * slice_end, but later limit was bumped up and bio was dispached
711 * sooner, then we need to reduce slice_end. A high bogus slice_end
712 * is bad because it does not allow new slice to start.
715 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
717 time_elapsed = jiffies - tg->slice_start[rw];
719 nr_slices = time_elapsed / throtl_slice;
723 tmp = tg->bps[rw] * throtl_slice * nr_slices;
727 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
729 if (!bytes_trim && !io_trim)
732 if (tg->bytes_disp[rw] >= bytes_trim)
733 tg->bytes_disp[rw] -= bytes_trim;
735 tg->bytes_disp[rw] = 0;
737 if (tg->io_disp[rw] >= io_trim)
738 tg->io_disp[rw] -= io_trim;
742 tg->slice_start[rw] += nr_slices * throtl_slice;
744 throtl_log(&tg->service_queue,
745 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
746 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
747 tg->slice_start[rw], tg->slice_end[rw], jiffies);
750 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
753 bool rw = bio_data_dir(bio);
754 unsigned int io_allowed;
755 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
758 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
760 /* Slice has just started. Consider one slice interval */
762 jiffy_elapsed_rnd = throtl_slice;
764 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
767 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
768 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
769 * will allow dispatch after 1 second and after that slice should
773 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
777 io_allowed = UINT_MAX;
781 if (tg->io_disp[rw] + 1 <= io_allowed) {
787 /* Calc approx time to dispatch */
788 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
790 if (jiffy_wait > jiffy_elapsed)
791 jiffy_wait = jiffy_wait - jiffy_elapsed;
800 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
803 bool rw = bio_data_dir(bio);
804 u64 bytes_allowed, extra_bytes, tmp;
805 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
807 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
809 /* Slice has just started. Consider one slice interval */
811 jiffy_elapsed_rnd = throtl_slice;
813 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
815 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
819 if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
825 /* Calc approx time to dispatch */
826 extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
827 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
833 * This wait time is without taking into consideration the rounding
834 * up we did. Add that time also.
836 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
843 * Returns whether one can dispatch a bio or not. Also returns approx number
844 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
846 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
849 bool rw = bio_data_dir(bio);
850 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
853 * Currently whole state machine of group depends on first bio
854 * queued in the group bio list. So one should not be calling
855 * this function with a different bio if there are other bios
858 BUG_ON(tg->service_queue.nr_queued[rw] &&
859 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
861 /* If tg->bps = -1, then BW is unlimited */
862 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
869 * If previous slice expired, start a new one otherwise renew/extend
870 * existing slice to make sure it is at least throtl_slice interval
873 if (throtl_slice_used(tg, rw))
874 throtl_start_new_slice(tg, rw);
876 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
877 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
880 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
881 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
887 max_wait = max(bps_wait, iops_wait);
892 if (time_before(tg->slice_end[rw], jiffies + max_wait))
893 throtl_extend_slice(tg, rw, jiffies + max_wait);
898 static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
901 struct throtl_grp *tg = blkg_to_tg(blkg);
902 struct tg_stats_cpu *stats_cpu;
906 * Disabling interrupts to provide mutual exclusion between two
907 * writes on same cpu. It probably is not needed for 64bit. Not
908 * optimizing that case yet.
910 local_irq_save(flags);
912 stats_cpu = this_cpu_ptr(tg->stats_cpu);
914 blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
915 blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
917 local_irq_restore(flags);
920 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
922 bool rw = bio_data_dir(bio);
924 /* Charge the bio to the group */
925 tg->bytes_disp[rw] += bio->bi_iter.bi_size;
929 * REQ_THROTTLED is used to prevent the same bio to be throttled
930 * more than once as a throttled bio will go through blk-throtl the
931 * second time when it eventually gets issued. Set it when a bio
932 * is being charged to a tg.
934 * Dispatch stats aren't recursive and each @bio should only be
935 * accounted by the @tg it was originally associated with. Let's
936 * update the stats when setting REQ_THROTTLED for the first time
937 * which is guaranteed to be for the @bio's original tg.
939 if (!(bio->bi_rw & REQ_THROTTLED)) {
940 bio->bi_rw |= REQ_THROTTLED;
941 throtl_update_dispatch_stats(tg_to_blkg(tg),
942 bio->bi_iter.bi_size, bio->bi_rw);
947 * throtl_add_bio_tg - add a bio to the specified throtl_grp
950 * @tg: the target throtl_grp
952 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
953 * tg->qnode_on_self[] is used.
955 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
956 struct throtl_grp *tg)
958 struct throtl_service_queue *sq = &tg->service_queue;
959 bool rw = bio_data_dir(bio);
962 qn = &tg->qnode_on_self[rw];
965 * If @tg doesn't currently have any bios queued in the same
966 * direction, queueing @bio can change when @tg should be
967 * dispatched. Mark that @tg was empty. This is automatically
968 * cleaered on the next tg_update_disptime().
970 if (!sq->nr_queued[rw])
971 tg->flags |= THROTL_TG_WAS_EMPTY;
973 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
976 throtl_enqueue_tg(tg);
979 static void tg_update_disptime(struct throtl_grp *tg)
981 struct throtl_service_queue *sq = &tg->service_queue;
982 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
985 if ((bio = throtl_peek_queued(&sq->queued[READ])))
986 tg_may_dispatch(tg, bio, &read_wait);
988 if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
989 tg_may_dispatch(tg, bio, &write_wait);
991 min_wait = min(read_wait, write_wait);
992 disptime = jiffies + min_wait;
994 /* Update dispatch time */
995 throtl_dequeue_tg(tg);
996 tg->disptime = disptime;
997 throtl_enqueue_tg(tg);
999 /* see throtl_add_bio_tg() */
1000 tg->flags &= ~THROTL_TG_WAS_EMPTY;
1003 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1004 struct throtl_grp *parent_tg, bool rw)
1006 if (throtl_slice_used(parent_tg, rw)) {
1007 throtl_start_new_slice_with_credit(parent_tg, rw,
1008 child_tg->slice_start[rw]);
1013 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1015 struct throtl_service_queue *sq = &tg->service_queue;
1016 struct throtl_service_queue *parent_sq = sq->parent_sq;
1017 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1018 struct throtl_grp *tg_to_put = NULL;
1022 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1023 * from @tg may put its reference and @parent_sq might end up
1024 * getting released prematurely. Remember the tg to put and put it
1025 * after @bio is transferred to @parent_sq.
1027 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1028 sq->nr_queued[rw]--;
1030 throtl_charge_bio(tg, bio);
1033 * If our parent is another tg, we just need to transfer @bio to
1034 * the parent using throtl_add_bio_tg(). If our parent is
1035 * @td->service_queue, @bio is ready to be issued. Put it on its
1036 * bio_lists[] and decrease total number queued. The caller is
1037 * responsible for issuing these bios.
1040 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1041 start_parent_slice_with_credit(tg, parent_tg, rw);
1043 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1044 &parent_sq->queued[rw]);
1045 BUG_ON(tg->td->nr_queued[rw] <= 0);
1046 tg->td->nr_queued[rw]--;
1049 throtl_trim_slice(tg, rw);
1052 blkg_put(tg_to_blkg(tg_to_put));
1055 static int throtl_dispatch_tg(struct throtl_grp *tg)
1057 struct throtl_service_queue *sq = &tg->service_queue;
1058 unsigned int nr_reads = 0, nr_writes = 0;
1059 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1060 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1063 /* Try to dispatch 75% READS and 25% WRITES */
1065 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1066 tg_may_dispatch(tg, bio, NULL)) {
1068 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1071 if (nr_reads >= max_nr_reads)
1075 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1076 tg_may_dispatch(tg, bio, NULL)) {
1078 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1081 if (nr_writes >= max_nr_writes)
1085 return nr_reads + nr_writes;
1088 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1090 unsigned int nr_disp = 0;
1093 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1094 struct throtl_service_queue *sq = &tg->service_queue;
1099 if (time_before(jiffies, tg->disptime))
1102 throtl_dequeue_tg(tg);
1104 nr_disp += throtl_dispatch_tg(tg);
1106 if (sq->nr_queued[0] || sq->nr_queued[1])
1107 tg_update_disptime(tg);
1109 if (nr_disp >= throtl_quantum)
1117 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1118 * @arg: the throtl_service_queue being serviced
1120 * This timer is armed when a child throtl_grp with active bio's become
1121 * pending and queued on the service_queue's pending_tree and expires when
1122 * the first child throtl_grp should be dispatched. This function
1123 * dispatches bio's from the children throtl_grps to the parent
1126 * If the parent's parent is another throtl_grp, dispatching is propagated
1127 * by either arming its pending_timer or repeating dispatch directly. If
1128 * the top-level service_tree is reached, throtl_data->dispatch_work is
1129 * kicked so that the ready bio's are issued.
1131 static void throtl_pending_timer_fn(unsigned long arg)
1133 struct throtl_service_queue *sq = (void *)arg;
1134 struct throtl_grp *tg = sq_to_tg(sq);
1135 struct throtl_data *td = sq_to_td(sq);
1136 struct request_queue *q = td->queue;
1137 struct throtl_service_queue *parent_sq;
1141 spin_lock_irq(q->queue_lock);
1143 parent_sq = sq->parent_sq;
1147 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1148 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1149 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1151 ret = throtl_select_dispatch(sq);
1153 throtl_log(sq, "bios disp=%u", ret);
1157 if (throtl_schedule_next_dispatch(sq, false))
1160 /* this dispatch windows is still open, relax and repeat */
1161 spin_unlock_irq(q->queue_lock);
1163 spin_lock_irq(q->queue_lock);
1170 /* @parent_sq is another throl_grp, propagate dispatch */
1171 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1172 tg_update_disptime(tg);
1173 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1174 /* window is already open, repeat dispatching */
1181 /* reached the top-level, queue issueing */
1182 queue_work(kthrotld_workqueue, &td->dispatch_work);
1185 spin_unlock_irq(q->queue_lock);
1189 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1190 * @work: work item being executed
1192 * This function is queued for execution when bio's reach the bio_lists[]
1193 * of throtl_data->service_queue. Those bio's are ready and issued by this
1196 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1198 struct throtl_data *td = container_of(work, struct throtl_data,
1200 struct throtl_service_queue *td_sq = &td->service_queue;
1201 struct request_queue *q = td->queue;
1202 struct bio_list bio_list_on_stack;
1204 struct blk_plug plug;
1207 bio_list_init(&bio_list_on_stack);
1209 spin_lock_irq(q->queue_lock);
1210 for (rw = READ; rw <= WRITE; rw++)
1211 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1212 bio_list_add(&bio_list_on_stack, bio);
1213 spin_unlock_irq(q->queue_lock);
1215 if (!bio_list_empty(&bio_list_on_stack)) {
1216 blk_start_plug(&plug);
1217 while((bio = bio_list_pop(&bio_list_on_stack)))
1218 generic_make_request(bio);
1219 blk_finish_plug(&plug);
1223 static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
1224 struct blkg_policy_data *pd, int off)
1226 struct throtl_grp *tg = pd_to_tg(pd);
1227 struct blkg_rwstat rwstat = { }, tmp;
1230 for_each_possible_cpu(cpu) {
1231 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
1233 tmp = blkg_rwstat_read((void *)sc + off);
1234 for (i = 0; i < BLKG_RWSTAT_NR; i++)
1235 rwstat.cnt[i] += tmp.cnt[i];
1238 return __blkg_prfill_rwstat(sf, pd, &rwstat);
1241 static int tg_print_cpu_rwstat(struct seq_file *sf, void *v)
1243 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_cpu_rwstat,
1244 &blkcg_policy_throtl, seq_cft(sf)->private, true);
1248 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1251 struct throtl_grp *tg = pd_to_tg(pd);
1252 u64 v = *(u64 *)((void *)tg + off);
1256 return __blkg_prfill_u64(sf, pd, v);
1259 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1262 struct throtl_grp *tg = pd_to_tg(pd);
1263 unsigned int v = *(unsigned int *)((void *)tg + off);
1267 return __blkg_prfill_u64(sf, pd, v);
1270 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1272 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1273 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1277 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1279 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1280 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1284 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1285 char *buf, size_t nbytes, loff_t off, bool is_u64)
1287 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1288 struct blkg_conf_ctx ctx;
1289 struct throtl_grp *tg;
1290 struct throtl_service_queue *sq;
1291 struct blkcg_gq *blkg;
1292 struct cgroup_subsys_state *pos_css;
1295 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1299 tg = blkg_to_tg(ctx.blkg);
1300 sq = &tg->service_queue;
1306 *(u64 *)((void *)tg + of_cft(of)->private) = ctx.v;
1308 *(unsigned int *)((void *)tg + of_cft(of)->private) = ctx.v;
1310 throtl_log(&tg->service_queue,
1311 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1312 tg->bps[READ], tg->bps[WRITE],
1313 tg->iops[READ], tg->iops[WRITE]);
1316 * Update has_rules[] flags for the updated tg's subtree. A tg is
1317 * considered to have rules if either the tg itself or any of its
1318 * ancestors has rules. This identifies groups without any
1319 * restrictions in the whole hierarchy and allows them to bypass
1322 blkg_for_each_descendant_pre(blkg, pos_css, ctx.blkg)
1323 tg_update_has_rules(blkg_to_tg(blkg));
1326 * We're already holding queue_lock and know @tg is valid. Let's
1327 * apply the new config directly.
1329 * Restart the slices for both READ and WRITES. It might happen
1330 * that a group's limit are dropped suddenly and we don't want to
1331 * account recently dispatched IO with new low rate.
1333 throtl_start_new_slice(tg, 0);
1334 throtl_start_new_slice(tg, 1);
1336 if (tg->flags & THROTL_TG_PENDING) {
1337 tg_update_disptime(tg);
1338 throtl_schedule_next_dispatch(sq->parent_sq, true);
1341 blkg_conf_finish(&ctx);
1345 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1346 char *buf, size_t nbytes, loff_t off)
1348 return tg_set_conf(of, buf, nbytes, off, true);
1351 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1352 char *buf, size_t nbytes, loff_t off)
1354 return tg_set_conf(of, buf, nbytes, off, false);
1357 static struct cftype throtl_files[] = {
1359 .name = "throttle.read_bps_device",
1360 .private = offsetof(struct throtl_grp, bps[READ]),
1361 .seq_show = tg_print_conf_u64,
1362 .write = tg_set_conf_u64,
1365 .name = "throttle.write_bps_device",
1366 .private = offsetof(struct throtl_grp, bps[WRITE]),
1367 .seq_show = tg_print_conf_u64,
1368 .write = tg_set_conf_u64,
1371 .name = "throttle.read_iops_device",
1372 .private = offsetof(struct throtl_grp, iops[READ]),
1373 .seq_show = tg_print_conf_uint,
1374 .write = tg_set_conf_uint,
1377 .name = "throttle.write_iops_device",
1378 .private = offsetof(struct throtl_grp, iops[WRITE]),
1379 .seq_show = tg_print_conf_uint,
1380 .write = tg_set_conf_uint,
1383 .name = "throttle.io_service_bytes",
1384 .private = offsetof(struct tg_stats_cpu, service_bytes),
1385 .seq_show = tg_print_cpu_rwstat,
1388 .name = "throttle.io_serviced",
1389 .private = offsetof(struct tg_stats_cpu, serviced),
1390 .seq_show = tg_print_cpu_rwstat,
1395 static void throtl_shutdown_wq(struct request_queue *q)
1397 struct throtl_data *td = q->td;
1399 cancel_work_sync(&td->dispatch_work);
1402 static struct blkcg_policy blkcg_policy_throtl = {
1403 .cftypes = throtl_files,
1405 .pd_alloc_fn = throtl_pd_alloc,
1406 .pd_init_fn = throtl_pd_init,
1407 .pd_online_fn = throtl_pd_online,
1408 .pd_free_fn = throtl_pd_free,
1409 .pd_reset_stats_fn = throtl_pd_reset_stats,
1412 bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
1414 struct throtl_data *td = q->td;
1415 struct throtl_qnode *qn = NULL;
1416 struct throtl_grp *tg;
1417 struct throtl_service_queue *sq;
1418 bool rw = bio_data_dir(bio);
1419 struct blkcg *blkcg;
1420 bool throttled = false;
1422 /* see throtl_charge_bio() */
1423 if (bio->bi_rw & REQ_THROTTLED)
1427 * A throtl_grp pointer retrieved under rcu can be used to access
1428 * basic fields like stats and io rates. If a group has no rules,
1429 * just update the dispatch stats in lockless manner and return.
1432 blkcg = bio_blkcg(bio);
1433 tg = throtl_lookup_tg(td, blkcg);
1435 if (!tg->has_rules[rw]) {
1436 throtl_update_dispatch_stats(tg_to_blkg(tg),
1437 bio->bi_iter.bi_size, bio->bi_rw);
1438 goto out_unlock_rcu;
1443 * Either group has not been allocated yet or it is not an unlimited
1446 spin_lock_irq(q->queue_lock);
1447 tg = throtl_lookup_create_tg(td, blkcg);
1451 sq = &tg->service_queue;
1454 /* throtl is FIFO - if bios are already queued, should queue */
1455 if (sq->nr_queued[rw])
1458 /* if above limits, break to queue */
1459 if (!tg_may_dispatch(tg, bio, NULL))
1462 /* within limits, let's charge and dispatch directly */
1463 throtl_charge_bio(tg, bio);
1466 * We need to trim slice even when bios are not being queued
1467 * otherwise it might happen that a bio is not queued for
1468 * a long time and slice keeps on extending and trim is not
1469 * called for a long time. Now if limits are reduced suddenly
1470 * we take into account all the IO dispatched so far at new
1471 * low rate and * newly queued IO gets a really long dispatch
1474 * So keep on trimming slice even if bio is not queued.
1476 throtl_trim_slice(tg, rw);
1479 * @bio passed through this layer without being throttled.
1480 * Climb up the ladder. If we''re already at the top, it
1481 * can be executed directly.
1483 qn = &tg->qnode_on_parent[rw];
1490 /* out-of-limit, queue to @tg */
1491 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1492 rw == READ ? 'R' : 'W',
1493 tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw],
1494 tg->io_disp[rw], tg->iops[rw],
1495 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1497 bio_associate_current(bio);
1498 tg->td->nr_queued[rw]++;
1499 throtl_add_bio_tg(bio, qn, tg);
1503 * Update @tg's dispatch time and force schedule dispatch if @tg
1504 * was empty before @bio. The forced scheduling isn't likely to
1505 * cause undue delay as @bio is likely to be dispatched directly if
1506 * its @tg's disptime is not in the future.
1508 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1509 tg_update_disptime(tg);
1510 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1514 spin_unlock_irq(q->queue_lock);
1519 * As multiple blk-throtls may stack in the same issue path, we
1520 * don't want bios to leave with the flag set. Clear the flag if
1524 bio->bi_rw &= ~REQ_THROTTLED;
1529 * Dispatch all bios from all children tg's queued on @parent_sq. On
1530 * return, @parent_sq is guaranteed to not have any active children tg's
1531 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1533 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1535 struct throtl_grp *tg;
1537 while ((tg = throtl_rb_first(parent_sq))) {
1538 struct throtl_service_queue *sq = &tg->service_queue;
1541 throtl_dequeue_tg(tg);
1543 while ((bio = throtl_peek_queued(&sq->queued[READ])))
1544 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1545 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1546 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1551 * blk_throtl_drain - drain throttled bios
1552 * @q: request_queue to drain throttled bios for
1554 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1556 void blk_throtl_drain(struct request_queue *q)
1557 __releases(q->queue_lock) __acquires(q->queue_lock)
1559 struct throtl_data *td = q->td;
1560 struct blkcg_gq *blkg;
1561 struct cgroup_subsys_state *pos_css;
1565 queue_lockdep_assert_held(q);
1569 * Drain each tg while doing post-order walk on the blkg tree, so
1570 * that all bios are propagated to td->service_queue. It'd be
1571 * better to walk service_queue tree directly but blkg walk is
1574 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
1575 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1577 /* finally, transfer bios from top-level tg's into the td */
1578 tg_drain_bios(&td->service_queue);
1581 spin_unlock_irq(q->queue_lock);
1583 /* all bios now should be in td->service_queue, issue them */
1584 for (rw = READ; rw <= WRITE; rw++)
1585 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1587 generic_make_request(bio);
1589 spin_lock_irq(q->queue_lock);
1592 int blk_throtl_init(struct request_queue *q)
1594 struct throtl_data *td;
1597 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1601 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1602 throtl_service_queue_init(&td->service_queue);
1607 /* activate policy */
1608 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1614 void blk_throtl_exit(struct request_queue *q)
1617 throtl_shutdown_wq(q);
1618 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1622 static int __init throtl_init(void)
1624 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1625 if (!kthrotld_workqueue)
1626 panic("Failed to create kthrotld\n");
1628 return blkcg_policy_register(&blkcg_policy_throtl);
1631 module_init(throtl_init);