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)
93 /* must be the first member */
94 struct blkg_policy_data pd;
96 /* active throtl group service_queue member */
97 struct rb_node rb_node;
99 /* throtl_data this group belongs to */
100 struct throtl_data *td;
102 /* this group's service queue */
103 struct throtl_service_queue service_queue;
106 * qnode_on_self is used when bios are directly queued to this
107 * throtl_grp so that local bios compete fairly with bios
108 * dispatched from children. qnode_on_parent is used when bios are
109 * dispatched from this throtl_grp into its parent and will compete
110 * with the sibling qnode_on_parents and the parent's
113 struct throtl_qnode qnode_on_self[2];
114 struct throtl_qnode qnode_on_parent[2];
117 * Dispatch time in jiffies. This is the estimated time when group
118 * will unthrottle and is ready to dispatch more bio. It is used as
119 * key to sort active groups in service tree.
121 unsigned long disptime;
125 /* are there any throtl rules between this group and td? */
128 /* internally used bytes per second rate limits */
129 uint64_t bps[2][LIMIT_CNT];
130 /* user configured bps limits */
131 uint64_t bps_conf[2][LIMIT_CNT];
133 /* internally used IOPS limits */
134 unsigned int iops[2][LIMIT_CNT];
135 /* user configured IOPS limits */
136 unsigned int iops_conf[2][LIMIT_CNT];
138 /* Number of bytes disptached in current slice */
139 uint64_t bytes_disp[2];
140 /* Number of bio's dispatched in current slice */
141 unsigned int io_disp[2];
143 /* When did we start a new slice */
144 unsigned long slice_start[2];
145 unsigned long slice_end[2];
150 /* service tree for active throtl groups */
151 struct throtl_service_queue service_queue;
153 struct request_queue *queue;
155 /* Total Number of queued bios on READ and WRITE lists */
156 unsigned int nr_queued[2];
158 /* Work for dispatching throttled bios */
159 struct work_struct dispatch_work;
160 unsigned int limit_index;
161 bool limit_valid[LIMIT_CNT];
164 static void throtl_pending_timer_fn(unsigned long arg);
166 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
168 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
171 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
173 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
176 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
178 return pd_to_blkg(&tg->pd);
182 * sq_to_tg - return the throl_grp the specified service queue belongs to
183 * @sq: the throtl_service_queue of interest
185 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
186 * embedded in throtl_data, %NULL is returned.
188 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
190 if (sq && sq->parent_sq)
191 return container_of(sq, struct throtl_grp, service_queue);
197 * sq_to_td - return throtl_data the specified service queue belongs to
198 * @sq: the throtl_service_queue of interest
200 * A service_queue can be embedded in either a throtl_grp or throtl_data.
201 * Determine the associated throtl_data accordingly and return it.
203 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
205 struct throtl_grp *tg = sq_to_tg(sq);
210 return container_of(sq, struct throtl_data, service_queue);
213 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
215 return tg->bps[rw][tg->td->limit_index];
218 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
220 return tg->iops[rw][tg->td->limit_index];
224 * throtl_log - log debug message via blktrace
225 * @sq: the service_queue being reported
226 * @fmt: printf format string
229 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
230 * throtl_grp; otherwise, just "throtl".
232 #define throtl_log(sq, fmt, args...) do { \
233 struct throtl_grp *__tg = sq_to_tg((sq)); \
234 struct throtl_data *__td = sq_to_td((sq)); \
237 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
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 throtl_service_queue_init(&tg->service_queue);
353 for (rw = READ; rw <= WRITE; rw++) {
354 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
355 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
358 RB_CLEAR_NODE(&tg->rb_node);
359 tg->bps[READ][LIMIT_MAX] = U64_MAX;
360 tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
361 tg->iops[READ][LIMIT_MAX] = UINT_MAX;
362 tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
363 tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
364 tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
365 tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
366 tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
367 /* LIMIT_LOW will have default value 0 */
372 static void throtl_pd_init(struct blkg_policy_data *pd)
374 struct throtl_grp *tg = pd_to_tg(pd);
375 struct blkcg_gq *blkg = tg_to_blkg(tg);
376 struct throtl_data *td = blkg->q->td;
377 struct throtl_service_queue *sq = &tg->service_queue;
380 * If on the default hierarchy, we switch to properly hierarchical
381 * behavior where limits on a given throtl_grp are applied to the
382 * whole subtree rather than just the group itself. e.g. If 16M
383 * read_bps limit is set on the root group, the whole system can't
384 * exceed 16M for the device.
386 * If not on the default hierarchy, the broken flat hierarchy
387 * behavior is retained where all throtl_grps are treated as if
388 * they're all separate root groups right below throtl_data.
389 * Limits of a group don't interact with limits of other groups
390 * regardless of the position of the group in the hierarchy.
392 sq->parent_sq = &td->service_queue;
393 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
394 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
399 * Set has_rules[] if @tg or any of its parents have limits configured.
400 * This doesn't require walking up to the top of the hierarchy as the
401 * parent's has_rules[] is guaranteed to be correct.
403 static void tg_update_has_rules(struct throtl_grp *tg)
405 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
406 struct throtl_data *td = tg->td;
409 for (rw = READ; rw <= WRITE; rw++)
410 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
411 (td->limit_valid[td->limit_index] &&
412 (tg_bps_limit(tg, rw) != U64_MAX ||
413 tg_iops_limit(tg, rw) != UINT_MAX));
416 static void throtl_pd_online(struct blkg_policy_data *pd)
419 * We don't want new groups to escape the limits of its ancestors.
420 * Update has_rules[] after a new group is brought online.
422 tg_update_has_rules(pd_to_tg(pd));
425 static void blk_throtl_update_limit_valid(struct throtl_data *td)
427 struct cgroup_subsys_state *pos_css;
428 struct blkcg_gq *blkg;
429 bool low_valid = false;
432 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
433 struct throtl_grp *tg = blkg_to_tg(blkg);
435 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
436 tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
441 td->limit_valid[LIMIT_LOW] = low_valid;
444 static void throtl_pd_offline(struct blkg_policy_data *pd)
446 struct throtl_grp *tg = pd_to_tg(pd);
448 tg->bps[READ][LIMIT_LOW] = 0;
449 tg->bps[WRITE][LIMIT_LOW] = 0;
450 tg->iops[READ][LIMIT_LOW] = 0;
451 tg->iops[WRITE][LIMIT_LOW] = 0;
453 blk_throtl_update_limit_valid(tg->td);
455 if (tg->td->limit_index == LIMIT_LOW &&
456 !tg->td->limit_valid[LIMIT_LOW])
457 tg->td->limit_index = LIMIT_MAX;
460 static void throtl_pd_free(struct blkg_policy_data *pd)
462 struct throtl_grp *tg = pd_to_tg(pd);
464 del_timer_sync(&tg->service_queue.pending_timer);
468 static struct throtl_grp *
469 throtl_rb_first(struct throtl_service_queue *parent_sq)
471 /* Service tree is empty */
472 if (!parent_sq->nr_pending)
475 if (!parent_sq->first_pending)
476 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
478 if (parent_sq->first_pending)
479 return rb_entry_tg(parent_sq->first_pending);
484 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
490 static void throtl_rb_erase(struct rb_node *n,
491 struct throtl_service_queue *parent_sq)
493 if (parent_sq->first_pending == n)
494 parent_sq->first_pending = NULL;
495 rb_erase_init(n, &parent_sq->pending_tree);
496 --parent_sq->nr_pending;
499 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
501 struct throtl_grp *tg;
503 tg = throtl_rb_first(parent_sq);
507 parent_sq->first_pending_disptime = tg->disptime;
510 static void tg_service_queue_add(struct throtl_grp *tg)
512 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
513 struct rb_node **node = &parent_sq->pending_tree.rb_node;
514 struct rb_node *parent = NULL;
515 struct throtl_grp *__tg;
516 unsigned long key = tg->disptime;
519 while (*node != NULL) {
521 __tg = rb_entry_tg(parent);
523 if (time_before(key, __tg->disptime))
524 node = &parent->rb_left;
526 node = &parent->rb_right;
532 parent_sq->first_pending = &tg->rb_node;
534 rb_link_node(&tg->rb_node, parent, node);
535 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
538 static void __throtl_enqueue_tg(struct throtl_grp *tg)
540 tg_service_queue_add(tg);
541 tg->flags |= THROTL_TG_PENDING;
542 tg->service_queue.parent_sq->nr_pending++;
545 static void throtl_enqueue_tg(struct throtl_grp *tg)
547 if (!(tg->flags & THROTL_TG_PENDING))
548 __throtl_enqueue_tg(tg);
551 static void __throtl_dequeue_tg(struct throtl_grp *tg)
553 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
554 tg->flags &= ~THROTL_TG_PENDING;
557 static void throtl_dequeue_tg(struct throtl_grp *tg)
559 if (tg->flags & THROTL_TG_PENDING)
560 __throtl_dequeue_tg(tg);
563 /* Call with queue lock held */
564 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
565 unsigned long expires)
567 mod_timer(&sq->pending_timer, expires);
568 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
569 expires - jiffies, jiffies);
573 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
574 * @sq: the service_queue to schedule dispatch for
575 * @force: force scheduling
577 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
578 * dispatch time of the first pending child. Returns %true if either timer
579 * is armed or there's no pending child left. %false if the current
580 * dispatch window is still open and the caller should continue
583 * If @force is %true, the dispatch timer is always scheduled and this
584 * function is guaranteed to return %true. This is to be used when the
585 * caller can't dispatch itself and needs to invoke pending_timer
586 * unconditionally. Note that forced scheduling is likely to induce short
587 * delay before dispatch starts even if @sq->first_pending_disptime is not
588 * in the future and thus shouldn't be used in hot paths.
590 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
593 /* any pending children left? */
597 update_min_dispatch_time(sq);
599 /* is the next dispatch time in the future? */
600 if (force || time_after(sq->first_pending_disptime, jiffies)) {
601 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
605 /* tell the caller to continue dispatching */
609 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
610 bool rw, unsigned long start)
612 tg->bytes_disp[rw] = 0;
616 * Previous slice has expired. We must have trimmed it after last
617 * bio dispatch. That means since start of last slice, we never used
618 * that bandwidth. Do try to make use of that bandwidth while giving
621 if (time_after_eq(start, tg->slice_start[rw]))
622 tg->slice_start[rw] = start;
624 tg->slice_end[rw] = jiffies + throtl_slice;
625 throtl_log(&tg->service_queue,
626 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
627 rw == READ ? 'R' : 'W', tg->slice_start[rw],
628 tg->slice_end[rw], jiffies);
631 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
633 tg->bytes_disp[rw] = 0;
635 tg->slice_start[rw] = jiffies;
636 tg->slice_end[rw] = jiffies + throtl_slice;
637 throtl_log(&tg->service_queue,
638 "[%c] new slice start=%lu end=%lu jiffies=%lu",
639 rw == READ ? 'R' : 'W', tg->slice_start[rw],
640 tg->slice_end[rw], jiffies);
643 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
644 unsigned long jiffy_end)
646 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
649 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
650 unsigned long jiffy_end)
652 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
653 throtl_log(&tg->service_queue,
654 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
655 rw == READ ? 'R' : 'W', tg->slice_start[rw],
656 tg->slice_end[rw], jiffies);
659 /* Determine if previously allocated or extended slice is complete or not */
660 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
662 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
668 /* Trim the used slices and adjust slice start accordingly */
669 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
671 unsigned long nr_slices, time_elapsed, io_trim;
674 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
677 * If bps are unlimited (-1), then time slice don't get
678 * renewed. Don't try to trim the slice if slice is used. A new
679 * slice will start when appropriate.
681 if (throtl_slice_used(tg, rw))
685 * A bio has been dispatched. Also adjust slice_end. It might happen
686 * that initially cgroup limit was very low resulting in high
687 * slice_end, but later limit was bumped up and bio was dispached
688 * sooner, then we need to reduce slice_end. A high bogus slice_end
689 * is bad because it does not allow new slice to start.
692 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
694 time_elapsed = jiffies - tg->slice_start[rw];
696 nr_slices = time_elapsed / throtl_slice;
700 tmp = tg_bps_limit(tg, rw) * throtl_slice * nr_slices;
704 io_trim = (tg_iops_limit(tg, rw) * throtl_slice * nr_slices) / HZ;
706 if (!bytes_trim && !io_trim)
709 if (tg->bytes_disp[rw] >= bytes_trim)
710 tg->bytes_disp[rw] -= bytes_trim;
712 tg->bytes_disp[rw] = 0;
714 if (tg->io_disp[rw] >= io_trim)
715 tg->io_disp[rw] -= io_trim;
719 tg->slice_start[rw] += nr_slices * throtl_slice;
721 throtl_log(&tg->service_queue,
722 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
723 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
724 tg->slice_start[rw], tg->slice_end[rw], jiffies);
727 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
730 bool rw = bio_data_dir(bio);
731 unsigned int io_allowed;
732 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
735 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
737 /* Slice has just started. Consider one slice interval */
739 jiffy_elapsed_rnd = throtl_slice;
741 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
744 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
745 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
746 * will allow dispatch after 1 second and after that slice should
750 tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd;
754 io_allowed = UINT_MAX;
758 if (tg->io_disp[rw] + 1 <= io_allowed) {
764 /* Calc approx time to dispatch */
765 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ) / tg_iops_limit(tg, rw) + 1;
767 if (jiffy_wait > jiffy_elapsed)
768 jiffy_wait = jiffy_wait - jiffy_elapsed;
777 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
780 bool rw = bio_data_dir(bio);
781 u64 bytes_allowed, extra_bytes, tmp;
782 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
784 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
786 /* Slice has just started. Consider one slice interval */
788 jiffy_elapsed_rnd = throtl_slice;
790 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
792 tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd;
796 if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
802 /* Calc approx time to dispatch */
803 extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
804 jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw));
810 * This wait time is without taking into consideration the rounding
811 * up we did. Add that time also.
813 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
820 * Returns whether one can dispatch a bio or not. Also returns approx number
821 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
823 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
826 bool rw = bio_data_dir(bio);
827 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
830 * Currently whole state machine of group depends on first bio
831 * queued in the group bio list. So one should not be calling
832 * this function with a different bio if there are other bios
835 BUG_ON(tg->service_queue.nr_queued[rw] &&
836 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
838 /* If tg->bps = -1, then BW is unlimited */
839 if (tg_bps_limit(tg, rw) == U64_MAX &&
840 tg_iops_limit(tg, rw) == UINT_MAX) {
847 * If previous slice expired, start a new one otherwise renew/extend
848 * existing slice to make sure it is at least throtl_slice interval
849 * long since now. New slice is started only for empty throttle group.
850 * If there is queued bio, that means there should be an active
851 * slice and it should be extended instead.
853 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
854 throtl_start_new_slice(tg, rw);
856 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
857 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
860 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
861 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
867 max_wait = max(bps_wait, iops_wait);
872 if (time_before(tg->slice_end[rw], jiffies + max_wait))
873 throtl_extend_slice(tg, rw, jiffies + max_wait);
878 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
880 bool rw = bio_data_dir(bio);
882 /* Charge the bio to the group */
883 tg->bytes_disp[rw] += bio->bi_iter.bi_size;
887 * BIO_THROTTLED is used to prevent the same bio to be throttled
888 * more than once as a throttled bio will go through blk-throtl the
889 * second time when it eventually gets issued. Set it when a bio
890 * is being charged to a tg.
892 if (!bio_flagged(bio, BIO_THROTTLED))
893 bio_set_flag(bio, BIO_THROTTLED);
897 * throtl_add_bio_tg - add a bio to the specified throtl_grp
900 * @tg: the target throtl_grp
902 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
903 * tg->qnode_on_self[] is used.
905 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
906 struct throtl_grp *tg)
908 struct throtl_service_queue *sq = &tg->service_queue;
909 bool rw = bio_data_dir(bio);
912 qn = &tg->qnode_on_self[rw];
915 * If @tg doesn't currently have any bios queued in the same
916 * direction, queueing @bio can change when @tg should be
917 * dispatched. Mark that @tg was empty. This is automatically
918 * cleaered on the next tg_update_disptime().
920 if (!sq->nr_queued[rw])
921 tg->flags |= THROTL_TG_WAS_EMPTY;
923 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
926 throtl_enqueue_tg(tg);
929 static void tg_update_disptime(struct throtl_grp *tg)
931 struct throtl_service_queue *sq = &tg->service_queue;
932 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
935 bio = throtl_peek_queued(&sq->queued[READ]);
937 tg_may_dispatch(tg, bio, &read_wait);
939 bio = throtl_peek_queued(&sq->queued[WRITE]);
941 tg_may_dispatch(tg, bio, &write_wait);
943 min_wait = min(read_wait, write_wait);
944 disptime = jiffies + min_wait;
946 /* Update dispatch time */
947 throtl_dequeue_tg(tg);
948 tg->disptime = disptime;
949 throtl_enqueue_tg(tg);
951 /* see throtl_add_bio_tg() */
952 tg->flags &= ~THROTL_TG_WAS_EMPTY;
955 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
956 struct throtl_grp *parent_tg, bool rw)
958 if (throtl_slice_used(parent_tg, rw)) {
959 throtl_start_new_slice_with_credit(parent_tg, rw,
960 child_tg->slice_start[rw]);
965 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
967 struct throtl_service_queue *sq = &tg->service_queue;
968 struct throtl_service_queue *parent_sq = sq->parent_sq;
969 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
970 struct throtl_grp *tg_to_put = NULL;
974 * @bio is being transferred from @tg to @parent_sq. Popping a bio
975 * from @tg may put its reference and @parent_sq might end up
976 * getting released prematurely. Remember the tg to put and put it
977 * after @bio is transferred to @parent_sq.
979 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
982 throtl_charge_bio(tg, bio);
985 * If our parent is another tg, we just need to transfer @bio to
986 * the parent using throtl_add_bio_tg(). If our parent is
987 * @td->service_queue, @bio is ready to be issued. Put it on its
988 * bio_lists[] and decrease total number queued. The caller is
989 * responsible for issuing these bios.
992 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
993 start_parent_slice_with_credit(tg, parent_tg, rw);
995 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
996 &parent_sq->queued[rw]);
997 BUG_ON(tg->td->nr_queued[rw] <= 0);
998 tg->td->nr_queued[rw]--;
1001 throtl_trim_slice(tg, rw);
1004 blkg_put(tg_to_blkg(tg_to_put));
1007 static int throtl_dispatch_tg(struct throtl_grp *tg)
1009 struct throtl_service_queue *sq = &tg->service_queue;
1010 unsigned int nr_reads = 0, nr_writes = 0;
1011 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1012 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1015 /* Try to dispatch 75% READS and 25% WRITES */
1017 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1018 tg_may_dispatch(tg, bio, NULL)) {
1020 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1023 if (nr_reads >= max_nr_reads)
1027 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1028 tg_may_dispatch(tg, bio, NULL)) {
1030 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1033 if (nr_writes >= max_nr_writes)
1037 return nr_reads + nr_writes;
1040 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1042 unsigned int nr_disp = 0;
1045 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1046 struct throtl_service_queue *sq = &tg->service_queue;
1051 if (time_before(jiffies, tg->disptime))
1054 throtl_dequeue_tg(tg);
1056 nr_disp += throtl_dispatch_tg(tg);
1058 if (sq->nr_queued[0] || sq->nr_queued[1])
1059 tg_update_disptime(tg);
1061 if (nr_disp >= throtl_quantum)
1069 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1070 * @arg: the throtl_service_queue being serviced
1072 * This timer is armed when a child throtl_grp with active bio's become
1073 * pending and queued on the service_queue's pending_tree and expires when
1074 * the first child throtl_grp should be dispatched. This function
1075 * dispatches bio's from the children throtl_grps to the parent
1078 * If the parent's parent is another throtl_grp, dispatching is propagated
1079 * by either arming its pending_timer or repeating dispatch directly. If
1080 * the top-level service_tree is reached, throtl_data->dispatch_work is
1081 * kicked so that the ready bio's are issued.
1083 static void throtl_pending_timer_fn(unsigned long arg)
1085 struct throtl_service_queue *sq = (void *)arg;
1086 struct throtl_grp *tg = sq_to_tg(sq);
1087 struct throtl_data *td = sq_to_td(sq);
1088 struct request_queue *q = td->queue;
1089 struct throtl_service_queue *parent_sq;
1093 spin_lock_irq(q->queue_lock);
1095 parent_sq = sq->parent_sq;
1099 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1100 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1101 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1103 ret = throtl_select_dispatch(sq);
1105 throtl_log(sq, "bios disp=%u", ret);
1109 if (throtl_schedule_next_dispatch(sq, false))
1112 /* this dispatch windows is still open, relax and repeat */
1113 spin_unlock_irq(q->queue_lock);
1115 spin_lock_irq(q->queue_lock);
1122 /* @parent_sq is another throl_grp, propagate dispatch */
1123 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1124 tg_update_disptime(tg);
1125 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1126 /* window is already open, repeat dispatching */
1133 /* reached the top-level, queue issueing */
1134 queue_work(kthrotld_workqueue, &td->dispatch_work);
1137 spin_unlock_irq(q->queue_lock);
1141 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1142 * @work: work item being executed
1144 * This function is queued for execution when bio's reach the bio_lists[]
1145 * of throtl_data->service_queue. Those bio's are ready and issued by this
1148 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1150 struct throtl_data *td = container_of(work, struct throtl_data,
1152 struct throtl_service_queue *td_sq = &td->service_queue;
1153 struct request_queue *q = td->queue;
1154 struct bio_list bio_list_on_stack;
1156 struct blk_plug plug;
1159 bio_list_init(&bio_list_on_stack);
1161 spin_lock_irq(q->queue_lock);
1162 for (rw = READ; rw <= WRITE; rw++)
1163 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1164 bio_list_add(&bio_list_on_stack, bio);
1165 spin_unlock_irq(q->queue_lock);
1167 if (!bio_list_empty(&bio_list_on_stack)) {
1168 blk_start_plug(&plug);
1169 while((bio = bio_list_pop(&bio_list_on_stack)))
1170 generic_make_request(bio);
1171 blk_finish_plug(&plug);
1175 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1178 struct throtl_grp *tg = pd_to_tg(pd);
1179 u64 v = *(u64 *)((void *)tg + off);
1183 return __blkg_prfill_u64(sf, pd, v);
1186 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1189 struct throtl_grp *tg = pd_to_tg(pd);
1190 unsigned int v = *(unsigned int *)((void *)tg + off);
1194 return __blkg_prfill_u64(sf, pd, v);
1197 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1199 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1200 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1204 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1206 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1207 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1211 static void tg_conf_updated(struct throtl_grp *tg)
1213 struct throtl_service_queue *sq = &tg->service_queue;
1214 struct cgroup_subsys_state *pos_css;
1215 struct blkcg_gq *blkg;
1217 throtl_log(&tg->service_queue,
1218 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1219 tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1220 tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1223 * Update has_rules[] flags for the updated tg's subtree. A tg is
1224 * considered to have rules if either the tg itself or any of its
1225 * ancestors has rules. This identifies groups without any
1226 * restrictions in the whole hierarchy and allows them to bypass
1229 blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg))
1230 tg_update_has_rules(blkg_to_tg(blkg));
1233 * We're already holding queue_lock and know @tg is valid. Let's
1234 * apply the new config directly.
1236 * Restart the slices for both READ and WRITES. It might happen
1237 * that a group's limit are dropped suddenly and we don't want to
1238 * account recently dispatched IO with new low rate.
1240 throtl_start_new_slice(tg, 0);
1241 throtl_start_new_slice(tg, 1);
1243 if (tg->flags & THROTL_TG_PENDING) {
1244 tg_update_disptime(tg);
1245 throtl_schedule_next_dispatch(sq->parent_sq, true);
1249 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1250 char *buf, size_t nbytes, loff_t off, bool is_u64)
1252 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1253 struct blkg_conf_ctx ctx;
1254 struct throtl_grp *tg;
1258 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1263 if (sscanf(ctx.body, "%llu", &v) != 1)
1268 tg = blkg_to_tg(ctx.blkg);
1271 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1273 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1275 tg_conf_updated(tg);
1278 blkg_conf_finish(&ctx);
1279 return ret ?: nbytes;
1282 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1283 char *buf, size_t nbytes, loff_t off)
1285 return tg_set_conf(of, buf, nbytes, off, true);
1288 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1289 char *buf, size_t nbytes, loff_t off)
1291 return tg_set_conf(of, buf, nbytes, off, false);
1294 static struct cftype throtl_legacy_files[] = {
1296 .name = "throttle.read_bps_device",
1297 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1298 .seq_show = tg_print_conf_u64,
1299 .write = tg_set_conf_u64,
1302 .name = "throttle.write_bps_device",
1303 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1304 .seq_show = tg_print_conf_u64,
1305 .write = tg_set_conf_u64,
1308 .name = "throttle.read_iops_device",
1309 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1310 .seq_show = tg_print_conf_uint,
1311 .write = tg_set_conf_uint,
1314 .name = "throttle.write_iops_device",
1315 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1316 .seq_show = tg_print_conf_uint,
1317 .write = tg_set_conf_uint,
1320 .name = "throttle.io_service_bytes",
1321 .private = (unsigned long)&blkcg_policy_throtl,
1322 .seq_show = blkg_print_stat_bytes,
1325 .name = "throttle.io_serviced",
1326 .private = (unsigned long)&blkcg_policy_throtl,
1327 .seq_show = blkg_print_stat_ios,
1332 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1335 struct throtl_grp *tg = pd_to_tg(pd);
1336 const char *dname = blkg_dev_name(pd->blkg);
1337 char bufs[4][21] = { "max", "max", "max", "max" };
1339 unsigned int iops_dft;
1344 if (off == LIMIT_LOW) {
1349 iops_dft = UINT_MAX;
1352 if (tg->bps_conf[READ][off] == bps_dft &&
1353 tg->bps_conf[WRITE][off] == bps_dft &&
1354 tg->iops_conf[READ][off] == iops_dft &&
1355 tg->iops_conf[WRITE][off] == iops_dft)
1358 if (tg->bps_conf[READ][off] != bps_dft)
1359 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1360 tg->bps_conf[READ][off]);
1361 if (tg->bps_conf[WRITE][off] != bps_dft)
1362 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1363 tg->bps_conf[WRITE][off]);
1364 if (tg->iops_conf[READ][off] != iops_dft)
1365 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1366 tg->iops_conf[READ][off]);
1367 if (tg->iops_conf[WRITE][off] != iops_dft)
1368 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1369 tg->iops_conf[WRITE][off]);
1371 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s\n",
1372 dname, bufs[0], bufs[1], bufs[2], bufs[3]);
1376 static int tg_print_limit(struct seq_file *sf, void *v)
1378 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1379 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1383 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1384 char *buf, size_t nbytes, loff_t off)
1386 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1387 struct blkg_conf_ctx ctx;
1388 struct throtl_grp *tg;
1391 int index = of_cft(of)->private;
1393 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1397 tg = blkg_to_tg(ctx.blkg);
1399 v[0] = tg->bps_conf[READ][index];
1400 v[1] = tg->bps_conf[WRITE][index];
1401 v[2] = tg->iops_conf[READ][index];
1402 v[3] = tg->iops_conf[WRITE][index];
1405 char tok[27]; /* wiops=18446744073709551616 */
1410 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1419 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1427 if (!strcmp(tok, "rbps"))
1429 else if (!strcmp(tok, "wbps"))
1431 else if (!strcmp(tok, "riops"))
1432 v[2] = min_t(u64, val, UINT_MAX);
1433 else if (!strcmp(tok, "wiops"))
1434 v[3] = min_t(u64, val, UINT_MAX);
1439 tg->bps_conf[READ][index] = v[0];
1440 tg->bps_conf[WRITE][index] = v[1];
1441 tg->iops_conf[READ][index] = v[2];
1442 tg->iops_conf[WRITE][index] = v[3];
1444 if (index == LIMIT_MAX) {
1445 tg->bps[READ][index] = v[0];
1446 tg->bps[WRITE][index] = v[1];
1447 tg->iops[READ][index] = v[2];
1448 tg->iops[WRITE][index] = v[3];
1450 tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1451 tg->bps_conf[READ][LIMIT_MAX]);
1452 tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1453 tg->bps_conf[WRITE][LIMIT_MAX]);
1454 tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1455 tg->iops_conf[READ][LIMIT_MAX]);
1456 tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1457 tg->iops_conf[WRITE][LIMIT_MAX]);
1459 if (index == LIMIT_LOW) {
1460 blk_throtl_update_limit_valid(tg->td);
1461 if (tg->td->limit_valid[LIMIT_LOW])
1462 tg->td->limit_index = LIMIT_LOW;
1464 tg_conf_updated(tg);
1467 blkg_conf_finish(&ctx);
1468 return ret ?: nbytes;
1471 static struct cftype throtl_files[] = {
1472 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1475 .flags = CFTYPE_NOT_ON_ROOT,
1476 .seq_show = tg_print_limit,
1477 .write = tg_set_limit,
1478 .private = LIMIT_LOW,
1483 .flags = CFTYPE_NOT_ON_ROOT,
1484 .seq_show = tg_print_limit,
1485 .write = tg_set_limit,
1486 .private = LIMIT_MAX,
1491 static void throtl_shutdown_wq(struct request_queue *q)
1493 struct throtl_data *td = q->td;
1495 cancel_work_sync(&td->dispatch_work);
1498 static struct blkcg_policy blkcg_policy_throtl = {
1499 .dfl_cftypes = throtl_files,
1500 .legacy_cftypes = throtl_legacy_files,
1502 .pd_alloc_fn = throtl_pd_alloc,
1503 .pd_init_fn = throtl_pd_init,
1504 .pd_online_fn = throtl_pd_online,
1505 .pd_offline_fn = throtl_pd_offline,
1506 .pd_free_fn = throtl_pd_free,
1509 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
1512 struct throtl_qnode *qn = NULL;
1513 struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
1514 struct throtl_service_queue *sq;
1515 bool rw = bio_data_dir(bio);
1516 bool throttled = false;
1518 WARN_ON_ONCE(!rcu_read_lock_held());
1520 /* see throtl_charge_bio() */
1521 if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw])
1524 spin_lock_irq(q->queue_lock);
1526 if (unlikely(blk_queue_bypass(q)))
1529 sq = &tg->service_queue;
1532 /* throtl is FIFO - if bios are already queued, should queue */
1533 if (sq->nr_queued[rw])
1536 /* if above limits, break to queue */
1537 if (!tg_may_dispatch(tg, bio, NULL))
1540 /* within limits, let's charge and dispatch directly */
1541 throtl_charge_bio(tg, bio);
1544 * We need to trim slice even when bios are not being queued
1545 * otherwise it might happen that a bio is not queued for
1546 * a long time and slice keeps on extending and trim is not
1547 * called for a long time. Now if limits are reduced suddenly
1548 * we take into account all the IO dispatched so far at new
1549 * low rate and * newly queued IO gets a really long dispatch
1552 * So keep on trimming slice even if bio is not queued.
1554 throtl_trim_slice(tg, rw);
1557 * @bio passed through this layer without being throttled.
1558 * Climb up the ladder. If we''re already at the top, it
1559 * can be executed directly.
1561 qn = &tg->qnode_on_parent[rw];
1568 /* out-of-limit, queue to @tg */
1569 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1570 rw == READ ? 'R' : 'W',
1571 tg->bytes_disp[rw], bio->bi_iter.bi_size,
1572 tg_bps_limit(tg, rw),
1573 tg->io_disp[rw], tg_iops_limit(tg, rw),
1574 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1576 bio_associate_current(bio);
1577 tg->td->nr_queued[rw]++;
1578 throtl_add_bio_tg(bio, qn, tg);
1582 * Update @tg's dispatch time and force schedule dispatch if @tg
1583 * was empty before @bio. The forced scheduling isn't likely to
1584 * cause undue delay as @bio is likely to be dispatched directly if
1585 * its @tg's disptime is not in the future.
1587 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1588 tg_update_disptime(tg);
1589 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1593 spin_unlock_irq(q->queue_lock);
1596 * As multiple blk-throtls may stack in the same issue path, we
1597 * don't want bios to leave with the flag set. Clear the flag if
1601 bio_clear_flag(bio, BIO_THROTTLED);
1606 * Dispatch all bios from all children tg's queued on @parent_sq. On
1607 * return, @parent_sq is guaranteed to not have any active children tg's
1608 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1610 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1612 struct throtl_grp *tg;
1614 while ((tg = throtl_rb_first(parent_sq))) {
1615 struct throtl_service_queue *sq = &tg->service_queue;
1618 throtl_dequeue_tg(tg);
1620 while ((bio = throtl_peek_queued(&sq->queued[READ])))
1621 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1622 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1623 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1628 * blk_throtl_drain - drain throttled bios
1629 * @q: request_queue to drain throttled bios for
1631 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1633 void blk_throtl_drain(struct request_queue *q)
1634 __releases(q->queue_lock) __acquires(q->queue_lock)
1636 struct throtl_data *td = q->td;
1637 struct blkcg_gq *blkg;
1638 struct cgroup_subsys_state *pos_css;
1642 queue_lockdep_assert_held(q);
1646 * Drain each tg while doing post-order walk on the blkg tree, so
1647 * that all bios are propagated to td->service_queue. It'd be
1648 * better to walk service_queue tree directly but blkg walk is
1651 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
1652 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1654 /* finally, transfer bios from top-level tg's into the td */
1655 tg_drain_bios(&td->service_queue);
1658 spin_unlock_irq(q->queue_lock);
1660 /* all bios now should be in td->service_queue, issue them */
1661 for (rw = READ; rw <= WRITE; rw++)
1662 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1664 generic_make_request(bio);
1666 spin_lock_irq(q->queue_lock);
1669 int blk_throtl_init(struct request_queue *q)
1671 struct throtl_data *td;
1674 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1678 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1679 throtl_service_queue_init(&td->service_queue);
1684 td->limit_valid[LIMIT_MAX] = true;
1685 td->limit_index = LIMIT_MAX;
1686 /* activate policy */
1687 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1693 void blk_throtl_exit(struct request_queue *q)
1696 throtl_shutdown_wq(q);
1697 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1701 static int __init throtl_init(void)
1703 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1704 if (!kthrotld_workqueue)
1705 panic("Failed to create kthrotld\n");
1707 return blkcg_policy_register(&blkcg_policy_throtl);
1710 module_init(throtl_init);