Loading...
1/*
2 * Interface for controlling IO bandwidth on a request queue
3 *
4 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
5 */
6
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>
13#include "blk.h"
14
15/* Max dispatch from a group in 1 round */
16static int throtl_grp_quantum = 8;
17
18/* Total max dispatch from all groups in one round */
19static int throtl_quantum = 32;
20
21/* Throttling is performed over 100ms slice and after that slice is renewed */
22static unsigned long throtl_slice = HZ/10; /* 100 ms */
23
24static struct blkcg_policy blkcg_policy_throtl;
25
26/* A workqueue to queue throttle related work */
27static struct workqueue_struct *kthrotld_workqueue;
28
29/*
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.
36 *
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.
41 *
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.
45 *
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.
51 */
52struct throtl_qnode {
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 */
56};
57
58struct throtl_service_queue {
59 struct throtl_service_queue *parent_sq; /* the parent service_queue */
60
61 /*
62 * Bios queued directly to this service_queue or dispatched from
63 * children throtl_grp's.
64 */
65 struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */
66 unsigned int nr_queued[2]; /* number of queued bios */
67
68 /*
69 * RB tree of active children throtl_grp's, which are sorted by
70 * their ->disptime.
71 */
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 */
77};
78
79enum tg_state_flags {
80 THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */
81 THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */
82};
83
84#define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
85
86struct throtl_grp {
87 /* must be the first member */
88 struct blkg_policy_data pd;
89
90 /* active throtl group service_queue member */
91 struct rb_node rb_node;
92
93 /* throtl_data this group belongs to */
94 struct throtl_data *td;
95
96 /* this group's service queue */
97 struct throtl_service_queue service_queue;
98
99 /*
100 * qnode_on_self is used when bios are directly queued to this
101 * throtl_grp so that local bios compete fairly with bios
102 * dispatched from children. qnode_on_parent is used when bios are
103 * dispatched from this throtl_grp into its parent and will compete
104 * with the sibling qnode_on_parents and the parent's
105 * qnode_on_self.
106 */
107 struct throtl_qnode qnode_on_self[2];
108 struct throtl_qnode qnode_on_parent[2];
109
110 /*
111 * Dispatch time in jiffies. This is the estimated time when group
112 * will unthrottle and is ready to dispatch more bio. It is used as
113 * key to sort active groups in service tree.
114 */
115 unsigned long disptime;
116
117 unsigned int flags;
118
119 /* are there any throtl rules between this group and td? */
120 bool has_rules[2];
121
122 /* bytes per second rate limits */
123 uint64_t bps[2];
124
125 /* IOPS limits */
126 unsigned int iops[2];
127
128 /* Number of bytes disptached in current slice */
129 uint64_t bytes_disp[2];
130 /* Number of bio's dispatched in current slice */
131 unsigned int io_disp[2];
132
133 /* When did we start a new slice */
134 unsigned long slice_start[2];
135 unsigned long slice_end[2];
136};
137
138struct throtl_data
139{
140 /* service tree for active throtl groups */
141 struct throtl_service_queue service_queue;
142
143 struct request_queue *queue;
144
145 /* Total Number of queued bios on READ and WRITE lists */
146 unsigned int nr_queued[2];
147
148 /*
149 * number of total undestroyed groups
150 */
151 unsigned int nr_undestroyed_grps;
152
153 /* Work for dispatching throttled bios */
154 struct work_struct dispatch_work;
155};
156
157static void throtl_pending_timer_fn(unsigned long arg);
158
159static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
160{
161 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
162}
163
164static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
165{
166 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
167}
168
169static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
170{
171 return pd_to_blkg(&tg->pd);
172}
173
174/**
175 * sq_to_tg - return the throl_grp the specified service queue belongs to
176 * @sq: the throtl_service_queue of interest
177 *
178 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
179 * embedded in throtl_data, %NULL is returned.
180 */
181static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
182{
183 if (sq && sq->parent_sq)
184 return container_of(sq, struct throtl_grp, service_queue);
185 else
186 return NULL;
187}
188
189/**
190 * sq_to_td - return throtl_data the specified service queue belongs to
191 * @sq: the throtl_service_queue of interest
192 *
193 * A service_queue can be embeded in either a throtl_grp or throtl_data.
194 * Determine the associated throtl_data accordingly and return it.
195 */
196static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
197{
198 struct throtl_grp *tg = sq_to_tg(sq);
199
200 if (tg)
201 return tg->td;
202 else
203 return container_of(sq, struct throtl_data, service_queue);
204}
205
206/**
207 * throtl_log - log debug message via blktrace
208 * @sq: the service_queue being reported
209 * @fmt: printf format string
210 * @args: printf args
211 *
212 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
213 * throtl_grp; otherwise, just "throtl".
214 *
215 * TODO: this should be made a function and name formatting should happen
216 * after testing whether blktrace is enabled.
217 */
218#define throtl_log(sq, fmt, args...) do { \
219 struct throtl_grp *__tg = sq_to_tg((sq)); \
220 struct throtl_data *__td = sq_to_td((sq)); \
221 \
222 (void)__td; \
223 if ((__tg)) { \
224 char __pbuf[128]; \
225 \
226 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
227 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
228 } else { \
229 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
230 } \
231} while (0)
232
233static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
234{
235 INIT_LIST_HEAD(&qn->node);
236 bio_list_init(&qn->bios);
237 qn->tg = tg;
238}
239
240/**
241 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
242 * @bio: bio being added
243 * @qn: qnode to add bio to
244 * @queued: the service_queue->queued[] list @qn belongs to
245 *
246 * Add @bio to @qn and put @qn on @queued if it's not already on.
247 * @qn->tg's reference count is bumped when @qn is activated. See the
248 * comment on top of throtl_qnode definition for details.
249 */
250static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
251 struct list_head *queued)
252{
253 bio_list_add(&qn->bios, bio);
254 if (list_empty(&qn->node)) {
255 list_add_tail(&qn->node, queued);
256 blkg_get(tg_to_blkg(qn->tg));
257 }
258}
259
260/**
261 * throtl_peek_queued - peek the first bio on a qnode list
262 * @queued: the qnode list to peek
263 */
264static struct bio *throtl_peek_queued(struct list_head *queued)
265{
266 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
267 struct bio *bio;
268
269 if (list_empty(queued))
270 return NULL;
271
272 bio = bio_list_peek(&qn->bios);
273 WARN_ON_ONCE(!bio);
274 return bio;
275}
276
277/**
278 * throtl_pop_queued - pop the first bio form a qnode list
279 * @queued: the qnode list to pop a bio from
280 * @tg_to_put: optional out argument for throtl_grp to put
281 *
282 * Pop the first bio from the qnode list @queued. After popping, the first
283 * qnode is removed from @queued if empty or moved to the end of @queued so
284 * that the popping order is round-robin.
285 *
286 * When the first qnode is removed, its associated throtl_grp should be put
287 * too. If @tg_to_put is NULL, this function automatically puts it;
288 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
289 * responsible for putting it.
290 */
291static struct bio *throtl_pop_queued(struct list_head *queued,
292 struct throtl_grp **tg_to_put)
293{
294 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
295 struct bio *bio;
296
297 if (list_empty(queued))
298 return NULL;
299
300 bio = bio_list_pop(&qn->bios);
301 WARN_ON_ONCE(!bio);
302
303 if (bio_list_empty(&qn->bios)) {
304 list_del_init(&qn->node);
305 if (tg_to_put)
306 *tg_to_put = qn->tg;
307 else
308 blkg_put(tg_to_blkg(qn->tg));
309 } else {
310 list_move_tail(&qn->node, queued);
311 }
312
313 return bio;
314}
315
316/* init a service_queue, assumes the caller zeroed it */
317static void throtl_service_queue_init(struct throtl_service_queue *sq)
318{
319 INIT_LIST_HEAD(&sq->queued[0]);
320 INIT_LIST_HEAD(&sq->queued[1]);
321 sq->pending_tree = RB_ROOT;
322 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
323 (unsigned long)sq);
324}
325
326static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
327{
328 struct throtl_grp *tg;
329 int rw;
330
331 tg = kzalloc_node(sizeof(*tg), gfp, node);
332 if (!tg)
333 return NULL;
334
335 throtl_service_queue_init(&tg->service_queue);
336
337 for (rw = READ; rw <= WRITE; rw++) {
338 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
339 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
340 }
341
342 RB_CLEAR_NODE(&tg->rb_node);
343 tg->bps[READ] = -1;
344 tg->bps[WRITE] = -1;
345 tg->iops[READ] = -1;
346 tg->iops[WRITE] = -1;
347
348 return &tg->pd;
349}
350
351static void throtl_pd_init(struct blkg_policy_data *pd)
352{
353 struct throtl_grp *tg = pd_to_tg(pd);
354 struct blkcg_gq *blkg = tg_to_blkg(tg);
355 struct throtl_data *td = blkg->q->td;
356 struct throtl_service_queue *sq = &tg->service_queue;
357
358 /*
359 * If on the default hierarchy, we switch to properly hierarchical
360 * behavior where limits on a given throtl_grp are applied to the
361 * whole subtree rather than just the group itself. e.g. If 16M
362 * read_bps limit is set on the root group, the whole system can't
363 * exceed 16M for the device.
364 *
365 * If not on the default hierarchy, the broken flat hierarchy
366 * behavior is retained where all throtl_grps are treated as if
367 * they're all separate root groups right below throtl_data.
368 * Limits of a group don't interact with limits of other groups
369 * regardless of the position of the group in the hierarchy.
370 */
371 sq->parent_sq = &td->service_queue;
372 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
373 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
374 tg->td = td;
375}
376
377/*
378 * Set has_rules[] if @tg or any of its parents have limits configured.
379 * This doesn't require walking up to the top of the hierarchy as the
380 * parent's has_rules[] is guaranteed to be correct.
381 */
382static void tg_update_has_rules(struct throtl_grp *tg)
383{
384 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
385 int rw;
386
387 for (rw = READ; rw <= WRITE; rw++)
388 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
389 (tg->bps[rw] != -1 || tg->iops[rw] != -1);
390}
391
392static void throtl_pd_online(struct blkg_policy_data *pd)
393{
394 /*
395 * We don't want new groups to escape the limits of its ancestors.
396 * Update has_rules[] after a new group is brought online.
397 */
398 tg_update_has_rules(pd_to_tg(pd));
399}
400
401static void throtl_pd_free(struct blkg_policy_data *pd)
402{
403 struct throtl_grp *tg = pd_to_tg(pd);
404
405 del_timer_sync(&tg->service_queue.pending_timer);
406 kfree(tg);
407}
408
409static struct throtl_grp *
410throtl_rb_first(struct throtl_service_queue *parent_sq)
411{
412 /* Service tree is empty */
413 if (!parent_sq->nr_pending)
414 return NULL;
415
416 if (!parent_sq->first_pending)
417 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
418
419 if (parent_sq->first_pending)
420 return rb_entry_tg(parent_sq->first_pending);
421
422 return NULL;
423}
424
425static void rb_erase_init(struct rb_node *n, struct rb_root *root)
426{
427 rb_erase(n, root);
428 RB_CLEAR_NODE(n);
429}
430
431static void throtl_rb_erase(struct rb_node *n,
432 struct throtl_service_queue *parent_sq)
433{
434 if (parent_sq->first_pending == n)
435 parent_sq->first_pending = NULL;
436 rb_erase_init(n, &parent_sq->pending_tree);
437 --parent_sq->nr_pending;
438}
439
440static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
441{
442 struct throtl_grp *tg;
443
444 tg = throtl_rb_first(parent_sq);
445 if (!tg)
446 return;
447
448 parent_sq->first_pending_disptime = tg->disptime;
449}
450
451static void tg_service_queue_add(struct throtl_grp *tg)
452{
453 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
454 struct rb_node **node = &parent_sq->pending_tree.rb_node;
455 struct rb_node *parent = NULL;
456 struct throtl_grp *__tg;
457 unsigned long key = tg->disptime;
458 int left = 1;
459
460 while (*node != NULL) {
461 parent = *node;
462 __tg = rb_entry_tg(parent);
463
464 if (time_before(key, __tg->disptime))
465 node = &parent->rb_left;
466 else {
467 node = &parent->rb_right;
468 left = 0;
469 }
470 }
471
472 if (left)
473 parent_sq->first_pending = &tg->rb_node;
474
475 rb_link_node(&tg->rb_node, parent, node);
476 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
477}
478
479static void __throtl_enqueue_tg(struct throtl_grp *tg)
480{
481 tg_service_queue_add(tg);
482 tg->flags |= THROTL_TG_PENDING;
483 tg->service_queue.parent_sq->nr_pending++;
484}
485
486static void throtl_enqueue_tg(struct throtl_grp *tg)
487{
488 if (!(tg->flags & THROTL_TG_PENDING))
489 __throtl_enqueue_tg(tg);
490}
491
492static void __throtl_dequeue_tg(struct throtl_grp *tg)
493{
494 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
495 tg->flags &= ~THROTL_TG_PENDING;
496}
497
498static void throtl_dequeue_tg(struct throtl_grp *tg)
499{
500 if (tg->flags & THROTL_TG_PENDING)
501 __throtl_dequeue_tg(tg);
502}
503
504/* Call with queue lock held */
505static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
506 unsigned long expires)
507{
508 mod_timer(&sq->pending_timer, expires);
509 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
510 expires - jiffies, jiffies);
511}
512
513/**
514 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
515 * @sq: the service_queue to schedule dispatch for
516 * @force: force scheduling
517 *
518 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
519 * dispatch time of the first pending child. Returns %true if either timer
520 * is armed or there's no pending child left. %false if the current
521 * dispatch window is still open and the caller should continue
522 * dispatching.
523 *
524 * If @force is %true, the dispatch timer is always scheduled and this
525 * function is guaranteed to return %true. This is to be used when the
526 * caller can't dispatch itself and needs to invoke pending_timer
527 * unconditionally. Note that forced scheduling is likely to induce short
528 * delay before dispatch starts even if @sq->first_pending_disptime is not
529 * in the future and thus shouldn't be used in hot paths.
530 */
531static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
532 bool force)
533{
534 /* any pending children left? */
535 if (!sq->nr_pending)
536 return true;
537
538 update_min_dispatch_time(sq);
539
540 /* is the next dispatch time in the future? */
541 if (force || time_after(sq->first_pending_disptime, jiffies)) {
542 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
543 return true;
544 }
545
546 /* tell the caller to continue dispatching */
547 return false;
548}
549
550static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
551 bool rw, unsigned long start)
552{
553 tg->bytes_disp[rw] = 0;
554 tg->io_disp[rw] = 0;
555
556 /*
557 * Previous slice has expired. We must have trimmed it after last
558 * bio dispatch. That means since start of last slice, we never used
559 * that bandwidth. Do try to make use of that bandwidth while giving
560 * credit.
561 */
562 if (time_after_eq(start, tg->slice_start[rw]))
563 tg->slice_start[rw] = start;
564
565 tg->slice_end[rw] = jiffies + throtl_slice;
566 throtl_log(&tg->service_queue,
567 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
568 rw == READ ? 'R' : 'W', tg->slice_start[rw],
569 tg->slice_end[rw], jiffies);
570}
571
572static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
573{
574 tg->bytes_disp[rw] = 0;
575 tg->io_disp[rw] = 0;
576 tg->slice_start[rw] = jiffies;
577 tg->slice_end[rw] = jiffies + throtl_slice;
578 throtl_log(&tg->service_queue,
579 "[%c] new slice start=%lu end=%lu jiffies=%lu",
580 rw == READ ? 'R' : 'W', tg->slice_start[rw],
581 tg->slice_end[rw], jiffies);
582}
583
584static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
585 unsigned long jiffy_end)
586{
587 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
588}
589
590static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
591 unsigned long jiffy_end)
592{
593 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
594 throtl_log(&tg->service_queue,
595 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
596 rw == READ ? 'R' : 'W', tg->slice_start[rw],
597 tg->slice_end[rw], jiffies);
598}
599
600/* Determine if previously allocated or extended slice is complete or not */
601static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
602{
603 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
604 return false;
605
606 return 1;
607}
608
609/* Trim the used slices and adjust slice start accordingly */
610static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
611{
612 unsigned long nr_slices, time_elapsed, io_trim;
613 u64 bytes_trim, tmp;
614
615 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
616
617 /*
618 * If bps are unlimited (-1), then time slice don't get
619 * renewed. Don't try to trim the slice if slice is used. A new
620 * slice will start when appropriate.
621 */
622 if (throtl_slice_used(tg, rw))
623 return;
624
625 /*
626 * A bio has been dispatched. Also adjust slice_end. It might happen
627 * that initially cgroup limit was very low resulting in high
628 * slice_end, but later limit was bumped up and bio was dispached
629 * sooner, then we need to reduce slice_end. A high bogus slice_end
630 * is bad because it does not allow new slice to start.
631 */
632
633 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
634
635 time_elapsed = jiffies - tg->slice_start[rw];
636
637 nr_slices = time_elapsed / throtl_slice;
638
639 if (!nr_slices)
640 return;
641 tmp = tg->bps[rw] * throtl_slice * nr_slices;
642 do_div(tmp, HZ);
643 bytes_trim = tmp;
644
645 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
646
647 if (!bytes_trim && !io_trim)
648 return;
649
650 if (tg->bytes_disp[rw] >= bytes_trim)
651 tg->bytes_disp[rw] -= bytes_trim;
652 else
653 tg->bytes_disp[rw] = 0;
654
655 if (tg->io_disp[rw] >= io_trim)
656 tg->io_disp[rw] -= io_trim;
657 else
658 tg->io_disp[rw] = 0;
659
660 tg->slice_start[rw] += nr_slices * throtl_slice;
661
662 throtl_log(&tg->service_queue,
663 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
664 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
665 tg->slice_start[rw], tg->slice_end[rw], jiffies);
666}
667
668static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
669 unsigned long *wait)
670{
671 bool rw = bio_data_dir(bio);
672 unsigned int io_allowed;
673 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
674 u64 tmp;
675
676 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
677
678 /* Slice has just started. Consider one slice interval */
679 if (!jiffy_elapsed)
680 jiffy_elapsed_rnd = throtl_slice;
681
682 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
683
684 /*
685 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
686 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
687 * will allow dispatch after 1 second and after that slice should
688 * have been trimmed.
689 */
690
691 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
692 do_div(tmp, HZ);
693
694 if (tmp > UINT_MAX)
695 io_allowed = UINT_MAX;
696 else
697 io_allowed = tmp;
698
699 if (tg->io_disp[rw] + 1 <= io_allowed) {
700 if (wait)
701 *wait = 0;
702 return true;
703 }
704
705 /* Calc approx time to dispatch */
706 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
707
708 if (jiffy_wait > jiffy_elapsed)
709 jiffy_wait = jiffy_wait - jiffy_elapsed;
710 else
711 jiffy_wait = 1;
712
713 if (wait)
714 *wait = jiffy_wait;
715 return 0;
716}
717
718static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
719 unsigned long *wait)
720{
721 bool rw = bio_data_dir(bio);
722 u64 bytes_allowed, extra_bytes, tmp;
723 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
724
725 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
726
727 /* Slice has just started. Consider one slice interval */
728 if (!jiffy_elapsed)
729 jiffy_elapsed_rnd = throtl_slice;
730
731 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
732
733 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
734 do_div(tmp, HZ);
735 bytes_allowed = tmp;
736
737 if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
738 if (wait)
739 *wait = 0;
740 return true;
741 }
742
743 /* Calc approx time to dispatch */
744 extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
745 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
746
747 if (!jiffy_wait)
748 jiffy_wait = 1;
749
750 /*
751 * This wait time is without taking into consideration the rounding
752 * up we did. Add that time also.
753 */
754 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
755 if (wait)
756 *wait = jiffy_wait;
757 return 0;
758}
759
760/*
761 * Returns whether one can dispatch a bio or not. Also returns approx number
762 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
763 */
764static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
765 unsigned long *wait)
766{
767 bool rw = bio_data_dir(bio);
768 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
769
770 /*
771 * Currently whole state machine of group depends on first bio
772 * queued in the group bio list. So one should not be calling
773 * this function with a different bio if there are other bios
774 * queued.
775 */
776 BUG_ON(tg->service_queue.nr_queued[rw] &&
777 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
778
779 /* If tg->bps = -1, then BW is unlimited */
780 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
781 if (wait)
782 *wait = 0;
783 return true;
784 }
785
786 /*
787 * If previous slice expired, start a new one otherwise renew/extend
788 * existing slice to make sure it is at least throtl_slice interval
789 * long since now.
790 */
791 if (throtl_slice_used(tg, rw))
792 throtl_start_new_slice(tg, rw);
793 else {
794 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
795 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
796 }
797
798 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
799 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
800 if (wait)
801 *wait = 0;
802 return 1;
803 }
804
805 max_wait = max(bps_wait, iops_wait);
806
807 if (wait)
808 *wait = max_wait;
809
810 if (time_before(tg->slice_end[rw], jiffies + max_wait))
811 throtl_extend_slice(tg, rw, jiffies + max_wait);
812
813 return 0;
814}
815
816static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
817{
818 bool rw = bio_data_dir(bio);
819
820 /* Charge the bio to the group */
821 tg->bytes_disp[rw] += bio->bi_iter.bi_size;
822 tg->io_disp[rw]++;
823
824 /*
825 * REQ_THROTTLED is used to prevent the same bio to be throttled
826 * more than once as a throttled bio will go through blk-throtl the
827 * second time when it eventually gets issued. Set it when a bio
828 * is being charged to a tg.
829 */
830 if (!(bio->bi_rw & REQ_THROTTLED))
831 bio->bi_rw |= REQ_THROTTLED;
832}
833
834/**
835 * throtl_add_bio_tg - add a bio to the specified throtl_grp
836 * @bio: bio to add
837 * @qn: qnode to use
838 * @tg: the target throtl_grp
839 *
840 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
841 * tg->qnode_on_self[] is used.
842 */
843static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
844 struct throtl_grp *tg)
845{
846 struct throtl_service_queue *sq = &tg->service_queue;
847 bool rw = bio_data_dir(bio);
848
849 if (!qn)
850 qn = &tg->qnode_on_self[rw];
851
852 /*
853 * If @tg doesn't currently have any bios queued in the same
854 * direction, queueing @bio can change when @tg should be
855 * dispatched. Mark that @tg was empty. This is automatically
856 * cleaered on the next tg_update_disptime().
857 */
858 if (!sq->nr_queued[rw])
859 tg->flags |= THROTL_TG_WAS_EMPTY;
860
861 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
862
863 sq->nr_queued[rw]++;
864 throtl_enqueue_tg(tg);
865}
866
867static void tg_update_disptime(struct throtl_grp *tg)
868{
869 struct throtl_service_queue *sq = &tg->service_queue;
870 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
871 struct bio *bio;
872
873 if ((bio = throtl_peek_queued(&sq->queued[READ])))
874 tg_may_dispatch(tg, bio, &read_wait);
875
876 if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
877 tg_may_dispatch(tg, bio, &write_wait);
878
879 min_wait = min(read_wait, write_wait);
880 disptime = jiffies + min_wait;
881
882 /* Update dispatch time */
883 throtl_dequeue_tg(tg);
884 tg->disptime = disptime;
885 throtl_enqueue_tg(tg);
886
887 /* see throtl_add_bio_tg() */
888 tg->flags &= ~THROTL_TG_WAS_EMPTY;
889}
890
891static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
892 struct throtl_grp *parent_tg, bool rw)
893{
894 if (throtl_slice_used(parent_tg, rw)) {
895 throtl_start_new_slice_with_credit(parent_tg, rw,
896 child_tg->slice_start[rw]);
897 }
898
899}
900
901static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
902{
903 struct throtl_service_queue *sq = &tg->service_queue;
904 struct throtl_service_queue *parent_sq = sq->parent_sq;
905 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
906 struct throtl_grp *tg_to_put = NULL;
907 struct bio *bio;
908
909 /*
910 * @bio is being transferred from @tg to @parent_sq. Popping a bio
911 * from @tg may put its reference and @parent_sq might end up
912 * getting released prematurely. Remember the tg to put and put it
913 * after @bio is transferred to @parent_sq.
914 */
915 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
916 sq->nr_queued[rw]--;
917
918 throtl_charge_bio(tg, bio);
919
920 /*
921 * If our parent is another tg, we just need to transfer @bio to
922 * the parent using throtl_add_bio_tg(). If our parent is
923 * @td->service_queue, @bio is ready to be issued. Put it on its
924 * bio_lists[] and decrease total number queued. The caller is
925 * responsible for issuing these bios.
926 */
927 if (parent_tg) {
928 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
929 start_parent_slice_with_credit(tg, parent_tg, rw);
930 } else {
931 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
932 &parent_sq->queued[rw]);
933 BUG_ON(tg->td->nr_queued[rw] <= 0);
934 tg->td->nr_queued[rw]--;
935 }
936
937 throtl_trim_slice(tg, rw);
938
939 if (tg_to_put)
940 blkg_put(tg_to_blkg(tg_to_put));
941}
942
943static int throtl_dispatch_tg(struct throtl_grp *tg)
944{
945 struct throtl_service_queue *sq = &tg->service_queue;
946 unsigned int nr_reads = 0, nr_writes = 0;
947 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
948 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
949 struct bio *bio;
950
951 /* Try to dispatch 75% READS and 25% WRITES */
952
953 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
954 tg_may_dispatch(tg, bio, NULL)) {
955
956 tg_dispatch_one_bio(tg, bio_data_dir(bio));
957 nr_reads++;
958
959 if (nr_reads >= max_nr_reads)
960 break;
961 }
962
963 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
964 tg_may_dispatch(tg, bio, NULL)) {
965
966 tg_dispatch_one_bio(tg, bio_data_dir(bio));
967 nr_writes++;
968
969 if (nr_writes >= max_nr_writes)
970 break;
971 }
972
973 return nr_reads + nr_writes;
974}
975
976static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
977{
978 unsigned int nr_disp = 0;
979
980 while (1) {
981 struct throtl_grp *tg = throtl_rb_first(parent_sq);
982 struct throtl_service_queue *sq = &tg->service_queue;
983
984 if (!tg)
985 break;
986
987 if (time_before(jiffies, tg->disptime))
988 break;
989
990 throtl_dequeue_tg(tg);
991
992 nr_disp += throtl_dispatch_tg(tg);
993
994 if (sq->nr_queued[0] || sq->nr_queued[1])
995 tg_update_disptime(tg);
996
997 if (nr_disp >= throtl_quantum)
998 break;
999 }
1000
1001 return nr_disp;
1002}
1003
1004/**
1005 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1006 * @arg: the throtl_service_queue being serviced
1007 *
1008 * This timer is armed when a child throtl_grp with active bio's become
1009 * pending and queued on the service_queue's pending_tree and expires when
1010 * the first child throtl_grp should be dispatched. This function
1011 * dispatches bio's from the children throtl_grps to the parent
1012 * service_queue.
1013 *
1014 * If the parent's parent is another throtl_grp, dispatching is propagated
1015 * by either arming its pending_timer or repeating dispatch directly. If
1016 * the top-level service_tree is reached, throtl_data->dispatch_work is
1017 * kicked so that the ready bio's are issued.
1018 */
1019static void throtl_pending_timer_fn(unsigned long arg)
1020{
1021 struct throtl_service_queue *sq = (void *)arg;
1022 struct throtl_grp *tg = sq_to_tg(sq);
1023 struct throtl_data *td = sq_to_td(sq);
1024 struct request_queue *q = td->queue;
1025 struct throtl_service_queue *parent_sq;
1026 bool dispatched;
1027 int ret;
1028
1029 spin_lock_irq(q->queue_lock);
1030again:
1031 parent_sq = sq->parent_sq;
1032 dispatched = false;
1033
1034 while (true) {
1035 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1036 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1037 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1038
1039 ret = throtl_select_dispatch(sq);
1040 if (ret) {
1041 throtl_log(sq, "bios disp=%u", ret);
1042 dispatched = true;
1043 }
1044
1045 if (throtl_schedule_next_dispatch(sq, false))
1046 break;
1047
1048 /* this dispatch windows is still open, relax and repeat */
1049 spin_unlock_irq(q->queue_lock);
1050 cpu_relax();
1051 spin_lock_irq(q->queue_lock);
1052 }
1053
1054 if (!dispatched)
1055 goto out_unlock;
1056
1057 if (parent_sq) {
1058 /* @parent_sq is another throl_grp, propagate dispatch */
1059 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1060 tg_update_disptime(tg);
1061 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1062 /* window is already open, repeat dispatching */
1063 sq = parent_sq;
1064 tg = sq_to_tg(sq);
1065 goto again;
1066 }
1067 }
1068 } else {
1069 /* reached the top-level, queue issueing */
1070 queue_work(kthrotld_workqueue, &td->dispatch_work);
1071 }
1072out_unlock:
1073 spin_unlock_irq(q->queue_lock);
1074}
1075
1076/**
1077 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1078 * @work: work item being executed
1079 *
1080 * This function is queued for execution when bio's reach the bio_lists[]
1081 * of throtl_data->service_queue. Those bio's are ready and issued by this
1082 * function.
1083 */
1084static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1085{
1086 struct throtl_data *td = container_of(work, struct throtl_data,
1087 dispatch_work);
1088 struct throtl_service_queue *td_sq = &td->service_queue;
1089 struct request_queue *q = td->queue;
1090 struct bio_list bio_list_on_stack;
1091 struct bio *bio;
1092 struct blk_plug plug;
1093 int rw;
1094
1095 bio_list_init(&bio_list_on_stack);
1096
1097 spin_lock_irq(q->queue_lock);
1098 for (rw = READ; rw <= WRITE; rw++)
1099 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1100 bio_list_add(&bio_list_on_stack, bio);
1101 spin_unlock_irq(q->queue_lock);
1102
1103 if (!bio_list_empty(&bio_list_on_stack)) {
1104 blk_start_plug(&plug);
1105 while((bio = bio_list_pop(&bio_list_on_stack)))
1106 generic_make_request(bio);
1107 blk_finish_plug(&plug);
1108 }
1109}
1110
1111static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1112 int off)
1113{
1114 struct throtl_grp *tg = pd_to_tg(pd);
1115 u64 v = *(u64 *)((void *)tg + off);
1116
1117 if (v == -1)
1118 return 0;
1119 return __blkg_prfill_u64(sf, pd, v);
1120}
1121
1122static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1123 int off)
1124{
1125 struct throtl_grp *tg = pd_to_tg(pd);
1126 unsigned int v = *(unsigned int *)((void *)tg + off);
1127
1128 if (v == -1)
1129 return 0;
1130 return __blkg_prfill_u64(sf, pd, v);
1131}
1132
1133static int tg_print_conf_u64(struct seq_file *sf, void *v)
1134{
1135 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1136 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1137 return 0;
1138}
1139
1140static int tg_print_conf_uint(struct seq_file *sf, void *v)
1141{
1142 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1143 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1144 return 0;
1145}
1146
1147static void tg_conf_updated(struct throtl_grp *tg)
1148{
1149 struct throtl_service_queue *sq = &tg->service_queue;
1150 struct cgroup_subsys_state *pos_css;
1151 struct blkcg_gq *blkg;
1152
1153 throtl_log(&tg->service_queue,
1154 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1155 tg->bps[READ], tg->bps[WRITE],
1156 tg->iops[READ], tg->iops[WRITE]);
1157
1158 /*
1159 * Update has_rules[] flags for the updated tg's subtree. A tg is
1160 * considered to have rules if either the tg itself or any of its
1161 * ancestors has rules. This identifies groups without any
1162 * restrictions in the whole hierarchy and allows them to bypass
1163 * blk-throttle.
1164 */
1165 blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg))
1166 tg_update_has_rules(blkg_to_tg(blkg));
1167
1168 /*
1169 * We're already holding queue_lock and know @tg is valid. Let's
1170 * apply the new config directly.
1171 *
1172 * Restart the slices for both READ and WRITES. It might happen
1173 * that a group's limit are dropped suddenly and we don't want to
1174 * account recently dispatched IO with new low rate.
1175 */
1176 throtl_start_new_slice(tg, 0);
1177 throtl_start_new_slice(tg, 1);
1178
1179 if (tg->flags & THROTL_TG_PENDING) {
1180 tg_update_disptime(tg);
1181 throtl_schedule_next_dispatch(sq->parent_sq, true);
1182 }
1183}
1184
1185static ssize_t tg_set_conf(struct kernfs_open_file *of,
1186 char *buf, size_t nbytes, loff_t off, bool is_u64)
1187{
1188 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1189 struct blkg_conf_ctx ctx;
1190 struct throtl_grp *tg;
1191 int ret;
1192 u64 v;
1193
1194 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1195 if (ret)
1196 return ret;
1197
1198 ret = -EINVAL;
1199 if (sscanf(ctx.body, "%llu", &v) != 1)
1200 goto out_finish;
1201 if (!v)
1202 v = -1;
1203
1204 tg = blkg_to_tg(ctx.blkg);
1205
1206 if (is_u64)
1207 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1208 else
1209 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1210
1211 tg_conf_updated(tg);
1212 ret = 0;
1213out_finish:
1214 blkg_conf_finish(&ctx);
1215 return ret ?: nbytes;
1216}
1217
1218static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1219 char *buf, size_t nbytes, loff_t off)
1220{
1221 return tg_set_conf(of, buf, nbytes, off, true);
1222}
1223
1224static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1225 char *buf, size_t nbytes, loff_t off)
1226{
1227 return tg_set_conf(of, buf, nbytes, off, false);
1228}
1229
1230static struct cftype throtl_legacy_files[] = {
1231 {
1232 .name = "throttle.read_bps_device",
1233 .private = offsetof(struct throtl_grp, bps[READ]),
1234 .seq_show = tg_print_conf_u64,
1235 .write = tg_set_conf_u64,
1236 },
1237 {
1238 .name = "throttle.write_bps_device",
1239 .private = offsetof(struct throtl_grp, bps[WRITE]),
1240 .seq_show = tg_print_conf_u64,
1241 .write = tg_set_conf_u64,
1242 },
1243 {
1244 .name = "throttle.read_iops_device",
1245 .private = offsetof(struct throtl_grp, iops[READ]),
1246 .seq_show = tg_print_conf_uint,
1247 .write = tg_set_conf_uint,
1248 },
1249 {
1250 .name = "throttle.write_iops_device",
1251 .private = offsetof(struct throtl_grp, iops[WRITE]),
1252 .seq_show = tg_print_conf_uint,
1253 .write = tg_set_conf_uint,
1254 },
1255 {
1256 .name = "throttle.io_service_bytes",
1257 .private = (unsigned long)&blkcg_policy_throtl,
1258 .seq_show = blkg_print_stat_bytes,
1259 },
1260 {
1261 .name = "throttle.io_serviced",
1262 .private = (unsigned long)&blkcg_policy_throtl,
1263 .seq_show = blkg_print_stat_ios,
1264 },
1265 { } /* terminate */
1266};
1267
1268static u64 tg_prfill_max(struct seq_file *sf, struct blkg_policy_data *pd,
1269 int off)
1270{
1271 struct throtl_grp *tg = pd_to_tg(pd);
1272 const char *dname = blkg_dev_name(pd->blkg);
1273 char bufs[4][21] = { "max", "max", "max", "max" };
1274
1275 if (!dname)
1276 return 0;
1277 if (tg->bps[READ] == -1 && tg->bps[WRITE] == -1 &&
1278 tg->iops[READ] == -1 && tg->iops[WRITE] == -1)
1279 return 0;
1280
1281 if (tg->bps[READ] != -1)
1282 snprintf(bufs[0], sizeof(bufs[0]), "%llu", tg->bps[READ]);
1283 if (tg->bps[WRITE] != -1)
1284 snprintf(bufs[1], sizeof(bufs[1]), "%llu", tg->bps[WRITE]);
1285 if (tg->iops[READ] != -1)
1286 snprintf(bufs[2], sizeof(bufs[2]), "%u", tg->iops[READ]);
1287 if (tg->iops[WRITE] != -1)
1288 snprintf(bufs[3], sizeof(bufs[3]), "%u", tg->iops[WRITE]);
1289
1290 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s\n",
1291 dname, bufs[0], bufs[1], bufs[2], bufs[3]);
1292 return 0;
1293}
1294
1295static int tg_print_max(struct seq_file *sf, void *v)
1296{
1297 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_max,
1298 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1299 return 0;
1300}
1301
1302static ssize_t tg_set_max(struct kernfs_open_file *of,
1303 char *buf, size_t nbytes, loff_t off)
1304{
1305 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1306 struct blkg_conf_ctx ctx;
1307 struct throtl_grp *tg;
1308 u64 v[4];
1309 int ret;
1310
1311 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1312 if (ret)
1313 return ret;
1314
1315 tg = blkg_to_tg(ctx.blkg);
1316
1317 v[0] = tg->bps[READ];
1318 v[1] = tg->bps[WRITE];
1319 v[2] = tg->iops[READ];
1320 v[3] = tg->iops[WRITE];
1321
1322 while (true) {
1323 char tok[27]; /* wiops=18446744073709551616 */
1324 char *p;
1325 u64 val = -1;
1326 int len;
1327
1328 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1329 break;
1330 if (tok[0] == '\0')
1331 break;
1332 ctx.body += len;
1333
1334 ret = -EINVAL;
1335 p = tok;
1336 strsep(&p, "=");
1337 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1338 goto out_finish;
1339
1340 ret = -ERANGE;
1341 if (!val)
1342 goto out_finish;
1343
1344 ret = -EINVAL;
1345 if (!strcmp(tok, "rbps"))
1346 v[0] = val;
1347 else if (!strcmp(tok, "wbps"))
1348 v[1] = val;
1349 else if (!strcmp(tok, "riops"))
1350 v[2] = min_t(u64, val, UINT_MAX);
1351 else if (!strcmp(tok, "wiops"))
1352 v[3] = min_t(u64, val, UINT_MAX);
1353 else
1354 goto out_finish;
1355 }
1356
1357 tg->bps[READ] = v[0];
1358 tg->bps[WRITE] = v[1];
1359 tg->iops[READ] = v[2];
1360 tg->iops[WRITE] = v[3];
1361
1362 tg_conf_updated(tg);
1363 ret = 0;
1364out_finish:
1365 blkg_conf_finish(&ctx);
1366 return ret ?: nbytes;
1367}
1368
1369static struct cftype throtl_files[] = {
1370 {
1371 .name = "max",
1372 .flags = CFTYPE_NOT_ON_ROOT,
1373 .seq_show = tg_print_max,
1374 .write = tg_set_max,
1375 },
1376 { } /* terminate */
1377};
1378
1379static void throtl_shutdown_wq(struct request_queue *q)
1380{
1381 struct throtl_data *td = q->td;
1382
1383 cancel_work_sync(&td->dispatch_work);
1384}
1385
1386static struct blkcg_policy blkcg_policy_throtl = {
1387 .dfl_cftypes = throtl_files,
1388 .legacy_cftypes = throtl_legacy_files,
1389
1390 .pd_alloc_fn = throtl_pd_alloc,
1391 .pd_init_fn = throtl_pd_init,
1392 .pd_online_fn = throtl_pd_online,
1393 .pd_free_fn = throtl_pd_free,
1394};
1395
1396bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
1397 struct bio *bio)
1398{
1399 struct throtl_qnode *qn = NULL;
1400 struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
1401 struct throtl_service_queue *sq;
1402 bool rw = bio_data_dir(bio);
1403 bool throttled = false;
1404
1405 WARN_ON_ONCE(!rcu_read_lock_held());
1406
1407 /* see throtl_charge_bio() */
1408 if ((bio->bi_rw & REQ_THROTTLED) || !tg->has_rules[rw])
1409 goto out;
1410
1411 spin_lock_irq(q->queue_lock);
1412
1413 if (unlikely(blk_queue_bypass(q)))
1414 goto out_unlock;
1415
1416 sq = &tg->service_queue;
1417
1418 while (true) {
1419 /* throtl is FIFO - if bios are already queued, should queue */
1420 if (sq->nr_queued[rw])
1421 break;
1422
1423 /* if above limits, break to queue */
1424 if (!tg_may_dispatch(tg, bio, NULL))
1425 break;
1426
1427 /* within limits, let's charge and dispatch directly */
1428 throtl_charge_bio(tg, bio);
1429
1430 /*
1431 * We need to trim slice even when bios are not being queued
1432 * otherwise it might happen that a bio is not queued for
1433 * a long time and slice keeps on extending and trim is not
1434 * called for a long time. Now if limits are reduced suddenly
1435 * we take into account all the IO dispatched so far at new
1436 * low rate and * newly queued IO gets a really long dispatch
1437 * time.
1438 *
1439 * So keep on trimming slice even if bio is not queued.
1440 */
1441 throtl_trim_slice(tg, rw);
1442
1443 /*
1444 * @bio passed through this layer without being throttled.
1445 * Climb up the ladder. If we''re already at the top, it
1446 * can be executed directly.
1447 */
1448 qn = &tg->qnode_on_parent[rw];
1449 sq = sq->parent_sq;
1450 tg = sq_to_tg(sq);
1451 if (!tg)
1452 goto out_unlock;
1453 }
1454
1455 /* out-of-limit, queue to @tg */
1456 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1457 rw == READ ? 'R' : 'W',
1458 tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw],
1459 tg->io_disp[rw], tg->iops[rw],
1460 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1461
1462 bio_associate_current(bio);
1463 tg->td->nr_queued[rw]++;
1464 throtl_add_bio_tg(bio, qn, tg);
1465 throttled = true;
1466
1467 /*
1468 * Update @tg's dispatch time and force schedule dispatch if @tg
1469 * was empty before @bio. The forced scheduling isn't likely to
1470 * cause undue delay as @bio is likely to be dispatched directly if
1471 * its @tg's disptime is not in the future.
1472 */
1473 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1474 tg_update_disptime(tg);
1475 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1476 }
1477
1478out_unlock:
1479 spin_unlock_irq(q->queue_lock);
1480out:
1481 /*
1482 * As multiple blk-throtls may stack in the same issue path, we
1483 * don't want bios to leave with the flag set. Clear the flag if
1484 * being issued.
1485 */
1486 if (!throttled)
1487 bio->bi_rw &= ~REQ_THROTTLED;
1488 return throttled;
1489}
1490
1491/*
1492 * Dispatch all bios from all children tg's queued on @parent_sq. On
1493 * return, @parent_sq is guaranteed to not have any active children tg's
1494 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1495 */
1496static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1497{
1498 struct throtl_grp *tg;
1499
1500 while ((tg = throtl_rb_first(parent_sq))) {
1501 struct throtl_service_queue *sq = &tg->service_queue;
1502 struct bio *bio;
1503
1504 throtl_dequeue_tg(tg);
1505
1506 while ((bio = throtl_peek_queued(&sq->queued[READ])))
1507 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1508 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1509 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1510 }
1511}
1512
1513/**
1514 * blk_throtl_drain - drain throttled bios
1515 * @q: request_queue to drain throttled bios for
1516 *
1517 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1518 */
1519void blk_throtl_drain(struct request_queue *q)
1520 __releases(q->queue_lock) __acquires(q->queue_lock)
1521{
1522 struct throtl_data *td = q->td;
1523 struct blkcg_gq *blkg;
1524 struct cgroup_subsys_state *pos_css;
1525 struct bio *bio;
1526 int rw;
1527
1528 queue_lockdep_assert_held(q);
1529 rcu_read_lock();
1530
1531 /*
1532 * Drain each tg while doing post-order walk on the blkg tree, so
1533 * that all bios are propagated to td->service_queue. It'd be
1534 * better to walk service_queue tree directly but blkg walk is
1535 * easier.
1536 */
1537 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
1538 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1539
1540 /* finally, transfer bios from top-level tg's into the td */
1541 tg_drain_bios(&td->service_queue);
1542
1543 rcu_read_unlock();
1544 spin_unlock_irq(q->queue_lock);
1545
1546 /* all bios now should be in td->service_queue, issue them */
1547 for (rw = READ; rw <= WRITE; rw++)
1548 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1549 NULL)))
1550 generic_make_request(bio);
1551
1552 spin_lock_irq(q->queue_lock);
1553}
1554
1555int blk_throtl_init(struct request_queue *q)
1556{
1557 struct throtl_data *td;
1558 int ret;
1559
1560 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1561 if (!td)
1562 return -ENOMEM;
1563
1564 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1565 throtl_service_queue_init(&td->service_queue);
1566
1567 q->td = td;
1568 td->queue = q;
1569
1570 /* activate policy */
1571 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1572 if (ret)
1573 kfree(td);
1574 return ret;
1575}
1576
1577void blk_throtl_exit(struct request_queue *q)
1578{
1579 BUG_ON(!q->td);
1580 throtl_shutdown_wq(q);
1581 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1582 kfree(q->td);
1583}
1584
1585static int __init throtl_init(void)
1586{
1587 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1588 if (!kthrotld_workqueue)
1589 panic("Failed to create kthrotld\n");
1590
1591 return blkcg_policy_register(&blkcg_policy_throtl);
1592}
1593
1594module_init(throtl_init);
1/*
2 * Interface for controlling IO bandwidth on a request queue
3 *
4 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
5 */
6
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 "blk-cgroup.h"
13
14/* Max dispatch from a group in 1 round */
15static int throtl_grp_quantum = 8;
16
17/* Total max dispatch from all groups in one round */
18static int throtl_quantum = 32;
19
20/* Throttling is performed over 100ms slice and after that slice is renewed */
21static unsigned long throtl_slice = HZ/10; /* 100 ms */
22
23/* A workqueue to queue throttle related work */
24static struct workqueue_struct *kthrotld_workqueue;
25static void throtl_schedule_delayed_work(struct throtl_data *td,
26 unsigned long delay);
27
28struct throtl_rb_root {
29 struct rb_root rb;
30 struct rb_node *left;
31 unsigned int count;
32 unsigned long min_disptime;
33};
34
35#define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \
36 .count = 0, .min_disptime = 0}
37
38#define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
39
40struct throtl_grp {
41 /* List of throtl groups on the request queue*/
42 struct hlist_node tg_node;
43
44 /* active throtl group service_tree member */
45 struct rb_node rb_node;
46
47 /*
48 * Dispatch time in jiffies. This is the estimated time when group
49 * will unthrottle and is ready to dispatch more bio. It is used as
50 * key to sort active groups in service tree.
51 */
52 unsigned long disptime;
53
54 struct blkio_group blkg;
55 atomic_t ref;
56 unsigned int flags;
57
58 /* Two lists for READ and WRITE */
59 struct bio_list bio_lists[2];
60
61 /* Number of queued bios on READ and WRITE lists */
62 unsigned int nr_queued[2];
63
64 /* bytes per second rate limits */
65 uint64_t bps[2];
66
67 /* IOPS limits */
68 unsigned int iops[2];
69
70 /* Number of bytes disptached in current slice */
71 uint64_t bytes_disp[2];
72 /* Number of bio's dispatched in current slice */
73 unsigned int io_disp[2];
74
75 /* When did we start a new slice */
76 unsigned long slice_start[2];
77 unsigned long slice_end[2];
78
79 /* Some throttle limits got updated for the group */
80 int limits_changed;
81
82 struct rcu_head rcu_head;
83};
84
85struct throtl_data
86{
87 /* List of throtl groups */
88 struct hlist_head tg_list;
89
90 /* service tree for active throtl groups */
91 struct throtl_rb_root tg_service_tree;
92
93 struct throtl_grp *root_tg;
94 struct request_queue *queue;
95
96 /* Total Number of queued bios on READ and WRITE lists */
97 unsigned int nr_queued[2];
98
99 /*
100 * number of total undestroyed groups
101 */
102 unsigned int nr_undestroyed_grps;
103
104 /* Work for dispatching throttled bios */
105 struct delayed_work throtl_work;
106
107 int limits_changed;
108};
109
110enum tg_state_flags {
111 THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */
112};
113
114#define THROTL_TG_FNS(name) \
115static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \
116{ \
117 (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \
118} \
119static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \
120{ \
121 (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \
122} \
123static inline int throtl_tg_##name(const struct throtl_grp *tg) \
124{ \
125 return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \
126}
127
128THROTL_TG_FNS(on_rr);
129
130#define throtl_log_tg(td, tg, fmt, args...) \
131 blk_add_trace_msg((td)->queue, "throtl %s " fmt, \
132 blkg_path(&(tg)->blkg), ##args); \
133
134#define throtl_log(td, fmt, args...) \
135 blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
136
137static inline struct throtl_grp *tg_of_blkg(struct blkio_group *blkg)
138{
139 if (blkg)
140 return container_of(blkg, struct throtl_grp, blkg);
141
142 return NULL;
143}
144
145static inline unsigned int total_nr_queued(struct throtl_data *td)
146{
147 return td->nr_queued[0] + td->nr_queued[1];
148}
149
150static inline struct throtl_grp *throtl_ref_get_tg(struct throtl_grp *tg)
151{
152 atomic_inc(&tg->ref);
153 return tg;
154}
155
156static void throtl_free_tg(struct rcu_head *head)
157{
158 struct throtl_grp *tg;
159
160 tg = container_of(head, struct throtl_grp, rcu_head);
161 free_percpu(tg->blkg.stats_cpu);
162 kfree(tg);
163}
164
165static void throtl_put_tg(struct throtl_grp *tg)
166{
167 BUG_ON(atomic_read(&tg->ref) <= 0);
168 if (!atomic_dec_and_test(&tg->ref))
169 return;
170
171 /*
172 * A group is freed in rcu manner. But having an rcu lock does not
173 * mean that one can access all the fields of blkg and assume these
174 * are valid. For example, don't try to follow throtl_data and
175 * request queue links.
176 *
177 * Having a reference to blkg under an rcu allows acess to only
178 * values local to groups like group stats and group rate limits
179 */
180 call_rcu(&tg->rcu_head, throtl_free_tg);
181}
182
183static void throtl_init_group(struct throtl_grp *tg)
184{
185 INIT_HLIST_NODE(&tg->tg_node);
186 RB_CLEAR_NODE(&tg->rb_node);
187 bio_list_init(&tg->bio_lists[0]);
188 bio_list_init(&tg->bio_lists[1]);
189 tg->limits_changed = false;
190
191 /* Practically unlimited BW */
192 tg->bps[0] = tg->bps[1] = -1;
193 tg->iops[0] = tg->iops[1] = -1;
194
195 /*
196 * Take the initial reference that will be released on destroy
197 * This can be thought of a joint reference by cgroup and
198 * request queue which will be dropped by either request queue
199 * exit or cgroup deletion path depending on who is exiting first.
200 */
201 atomic_set(&tg->ref, 1);
202}
203
204/* Should be called with rcu read lock held (needed for blkcg) */
205static void
206throtl_add_group_to_td_list(struct throtl_data *td, struct throtl_grp *tg)
207{
208 hlist_add_head(&tg->tg_node, &td->tg_list);
209 td->nr_undestroyed_grps++;
210}
211
212static void
213__throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg)
214{
215 struct backing_dev_info *bdi = &td->queue->backing_dev_info;
216 unsigned int major, minor;
217
218 if (!tg || tg->blkg.dev)
219 return;
220
221 /*
222 * Fill in device details for a group which might not have been
223 * filled at group creation time as queue was being instantiated
224 * and driver had not attached a device yet
225 */
226 if (bdi->dev && dev_name(bdi->dev)) {
227 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
228 tg->blkg.dev = MKDEV(major, minor);
229 }
230}
231
232/*
233 * Should be called with without queue lock held. Here queue lock will be
234 * taken rarely. It will be taken only once during life time of a group
235 * if need be
236 */
237static void
238throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg)
239{
240 if (!tg || tg->blkg.dev)
241 return;
242
243 spin_lock_irq(td->queue->queue_lock);
244 __throtl_tg_fill_dev_details(td, tg);
245 spin_unlock_irq(td->queue->queue_lock);
246}
247
248static void throtl_init_add_tg_lists(struct throtl_data *td,
249 struct throtl_grp *tg, struct blkio_cgroup *blkcg)
250{
251 __throtl_tg_fill_dev_details(td, tg);
252
253 /* Add group onto cgroup list */
254 blkiocg_add_blkio_group(blkcg, &tg->blkg, (void *)td,
255 tg->blkg.dev, BLKIO_POLICY_THROTL);
256
257 tg->bps[READ] = blkcg_get_read_bps(blkcg, tg->blkg.dev);
258 tg->bps[WRITE] = blkcg_get_write_bps(blkcg, tg->blkg.dev);
259 tg->iops[READ] = blkcg_get_read_iops(blkcg, tg->blkg.dev);
260 tg->iops[WRITE] = blkcg_get_write_iops(blkcg, tg->blkg.dev);
261
262 throtl_add_group_to_td_list(td, tg);
263}
264
265/* Should be called without queue lock and outside of rcu period */
266static struct throtl_grp *throtl_alloc_tg(struct throtl_data *td)
267{
268 struct throtl_grp *tg = NULL;
269 int ret;
270
271 tg = kzalloc_node(sizeof(*tg), GFP_ATOMIC, td->queue->node);
272 if (!tg)
273 return NULL;
274
275 ret = blkio_alloc_blkg_stats(&tg->blkg);
276
277 if (ret) {
278 kfree(tg);
279 return NULL;
280 }
281
282 throtl_init_group(tg);
283 return tg;
284}
285
286static struct
287throtl_grp *throtl_find_tg(struct throtl_data *td, struct blkio_cgroup *blkcg)
288{
289 struct throtl_grp *tg = NULL;
290 void *key = td;
291
292 /*
293 * This is the common case when there are no blkio cgroups.
294 * Avoid lookup in this case
295 */
296 if (blkcg == &blkio_root_cgroup)
297 tg = td->root_tg;
298 else
299 tg = tg_of_blkg(blkiocg_lookup_group(blkcg, key));
300
301 __throtl_tg_fill_dev_details(td, tg);
302 return tg;
303}
304
305/*
306 * This function returns with queue lock unlocked in case of error, like
307 * request queue is no more
308 */
309static struct throtl_grp * throtl_get_tg(struct throtl_data *td)
310{
311 struct throtl_grp *tg = NULL, *__tg = NULL;
312 struct blkio_cgroup *blkcg;
313 struct request_queue *q = td->queue;
314
315 rcu_read_lock();
316 blkcg = task_blkio_cgroup(current);
317 tg = throtl_find_tg(td, blkcg);
318 if (tg) {
319 rcu_read_unlock();
320 return tg;
321 }
322
323 /*
324 * Need to allocate a group. Allocation of group also needs allocation
325 * of per cpu stats which in-turn takes a mutex() and can block. Hence
326 * we need to drop rcu lock and queue_lock before we call alloc
327 *
328 * Take the request queue reference to make sure queue does not
329 * go away once we return from allocation.
330 */
331 blk_get_queue(q);
332 rcu_read_unlock();
333 spin_unlock_irq(q->queue_lock);
334
335 tg = throtl_alloc_tg(td);
336 /*
337 * We might have slept in group allocation. Make sure queue is not
338 * dead
339 */
340 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
341 blk_put_queue(q);
342 if (tg)
343 kfree(tg);
344
345 return ERR_PTR(-ENODEV);
346 }
347 blk_put_queue(q);
348
349 /* Group allocated and queue is still alive. take the lock */
350 spin_lock_irq(q->queue_lock);
351
352 /*
353 * Initialize the new group. After sleeping, read the blkcg again.
354 */
355 rcu_read_lock();
356 blkcg = task_blkio_cgroup(current);
357
358 /*
359 * If some other thread already allocated the group while we were
360 * not holding queue lock, free up the group
361 */
362 __tg = throtl_find_tg(td, blkcg);
363
364 if (__tg) {
365 kfree(tg);
366 rcu_read_unlock();
367 return __tg;
368 }
369
370 /* Group allocation failed. Account the IO to root group */
371 if (!tg) {
372 tg = td->root_tg;
373 return tg;
374 }
375
376 throtl_init_add_tg_lists(td, tg, blkcg);
377 rcu_read_unlock();
378 return tg;
379}
380
381static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root)
382{
383 /* Service tree is empty */
384 if (!root->count)
385 return NULL;
386
387 if (!root->left)
388 root->left = rb_first(&root->rb);
389
390 if (root->left)
391 return rb_entry_tg(root->left);
392
393 return NULL;
394}
395
396static void rb_erase_init(struct rb_node *n, struct rb_root *root)
397{
398 rb_erase(n, root);
399 RB_CLEAR_NODE(n);
400}
401
402static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root)
403{
404 if (root->left == n)
405 root->left = NULL;
406 rb_erase_init(n, &root->rb);
407 --root->count;
408}
409
410static void update_min_dispatch_time(struct throtl_rb_root *st)
411{
412 struct throtl_grp *tg;
413
414 tg = throtl_rb_first(st);
415 if (!tg)
416 return;
417
418 st->min_disptime = tg->disptime;
419}
420
421static void
422tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg)
423{
424 struct rb_node **node = &st->rb.rb_node;
425 struct rb_node *parent = NULL;
426 struct throtl_grp *__tg;
427 unsigned long key = tg->disptime;
428 int left = 1;
429
430 while (*node != NULL) {
431 parent = *node;
432 __tg = rb_entry_tg(parent);
433
434 if (time_before(key, __tg->disptime))
435 node = &parent->rb_left;
436 else {
437 node = &parent->rb_right;
438 left = 0;
439 }
440 }
441
442 if (left)
443 st->left = &tg->rb_node;
444
445 rb_link_node(&tg->rb_node, parent, node);
446 rb_insert_color(&tg->rb_node, &st->rb);
447}
448
449static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
450{
451 struct throtl_rb_root *st = &td->tg_service_tree;
452
453 tg_service_tree_add(st, tg);
454 throtl_mark_tg_on_rr(tg);
455 st->count++;
456}
457
458static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
459{
460 if (!throtl_tg_on_rr(tg))
461 __throtl_enqueue_tg(td, tg);
462}
463
464static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
465{
466 throtl_rb_erase(&tg->rb_node, &td->tg_service_tree);
467 throtl_clear_tg_on_rr(tg);
468}
469
470static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
471{
472 if (throtl_tg_on_rr(tg))
473 __throtl_dequeue_tg(td, tg);
474}
475
476static void throtl_schedule_next_dispatch(struct throtl_data *td)
477{
478 struct throtl_rb_root *st = &td->tg_service_tree;
479
480 /*
481 * If there are more bios pending, schedule more work.
482 */
483 if (!total_nr_queued(td))
484 return;
485
486 BUG_ON(!st->count);
487
488 update_min_dispatch_time(st);
489
490 if (time_before_eq(st->min_disptime, jiffies))
491 throtl_schedule_delayed_work(td, 0);
492 else
493 throtl_schedule_delayed_work(td, (st->min_disptime - jiffies));
494}
495
496static inline void
497throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
498{
499 tg->bytes_disp[rw] = 0;
500 tg->io_disp[rw] = 0;
501 tg->slice_start[rw] = jiffies;
502 tg->slice_end[rw] = jiffies + throtl_slice;
503 throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
504 rw == READ ? 'R' : 'W', tg->slice_start[rw],
505 tg->slice_end[rw], jiffies);
506}
507
508static inline void throtl_set_slice_end(struct throtl_data *td,
509 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
510{
511 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
512}
513
514static inline void throtl_extend_slice(struct throtl_data *td,
515 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
516{
517 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
518 throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
519 rw == READ ? 'R' : 'W', tg->slice_start[rw],
520 tg->slice_end[rw], jiffies);
521}
522
523/* Determine if previously allocated or extended slice is complete or not */
524static bool
525throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
526{
527 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
528 return 0;
529
530 return 1;
531}
532
533/* Trim the used slices and adjust slice start accordingly */
534static inline void
535throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
536{
537 unsigned long nr_slices, time_elapsed, io_trim;
538 u64 bytes_trim, tmp;
539
540 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
541
542 /*
543 * If bps are unlimited (-1), then time slice don't get
544 * renewed. Don't try to trim the slice if slice is used. A new
545 * slice will start when appropriate.
546 */
547 if (throtl_slice_used(td, tg, rw))
548 return;
549
550 /*
551 * A bio has been dispatched. Also adjust slice_end. It might happen
552 * that initially cgroup limit was very low resulting in high
553 * slice_end, but later limit was bumped up and bio was dispached
554 * sooner, then we need to reduce slice_end. A high bogus slice_end
555 * is bad because it does not allow new slice to start.
556 */
557
558 throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice);
559
560 time_elapsed = jiffies - tg->slice_start[rw];
561
562 nr_slices = time_elapsed / throtl_slice;
563
564 if (!nr_slices)
565 return;
566 tmp = tg->bps[rw] * throtl_slice * nr_slices;
567 do_div(tmp, HZ);
568 bytes_trim = tmp;
569
570 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
571
572 if (!bytes_trim && !io_trim)
573 return;
574
575 if (tg->bytes_disp[rw] >= bytes_trim)
576 tg->bytes_disp[rw] -= bytes_trim;
577 else
578 tg->bytes_disp[rw] = 0;
579
580 if (tg->io_disp[rw] >= io_trim)
581 tg->io_disp[rw] -= io_trim;
582 else
583 tg->io_disp[rw] = 0;
584
585 tg->slice_start[rw] += nr_slices * throtl_slice;
586
587 throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
588 " start=%lu end=%lu jiffies=%lu",
589 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
590 tg->slice_start[rw], tg->slice_end[rw], jiffies);
591}
592
593static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
594 struct bio *bio, unsigned long *wait)
595{
596 bool rw = bio_data_dir(bio);
597 unsigned int io_allowed;
598 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
599 u64 tmp;
600
601 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
602
603 /* Slice has just started. Consider one slice interval */
604 if (!jiffy_elapsed)
605 jiffy_elapsed_rnd = throtl_slice;
606
607 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
608
609 /*
610 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
611 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
612 * will allow dispatch after 1 second and after that slice should
613 * have been trimmed.
614 */
615
616 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
617 do_div(tmp, HZ);
618
619 if (tmp > UINT_MAX)
620 io_allowed = UINT_MAX;
621 else
622 io_allowed = tmp;
623
624 if (tg->io_disp[rw] + 1 <= io_allowed) {
625 if (wait)
626 *wait = 0;
627 return 1;
628 }
629
630 /* Calc approx time to dispatch */
631 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
632
633 if (jiffy_wait > jiffy_elapsed)
634 jiffy_wait = jiffy_wait - jiffy_elapsed;
635 else
636 jiffy_wait = 1;
637
638 if (wait)
639 *wait = jiffy_wait;
640 return 0;
641}
642
643static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
644 struct bio *bio, unsigned long *wait)
645{
646 bool rw = bio_data_dir(bio);
647 u64 bytes_allowed, extra_bytes, tmp;
648 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
649
650 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
651
652 /* Slice has just started. Consider one slice interval */
653 if (!jiffy_elapsed)
654 jiffy_elapsed_rnd = throtl_slice;
655
656 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
657
658 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
659 do_div(tmp, HZ);
660 bytes_allowed = tmp;
661
662 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
663 if (wait)
664 *wait = 0;
665 return 1;
666 }
667
668 /* Calc approx time to dispatch */
669 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
670 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
671
672 if (!jiffy_wait)
673 jiffy_wait = 1;
674
675 /*
676 * This wait time is without taking into consideration the rounding
677 * up we did. Add that time also.
678 */
679 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
680 if (wait)
681 *wait = jiffy_wait;
682 return 0;
683}
684
685static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
686 if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
687 return 1;
688 return 0;
689}
690
691/*
692 * Returns whether one can dispatch a bio or not. Also returns approx number
693 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
694 */
695static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
696 struct bio *bio, unsigned long *wait)
697{
698 bool rw = bio_data_dir(bio);
699 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
700
701 /*
702 * Currently whole state machine of group depends on first bio
703 * queued in the group bio list. So one should not be calling
704 * this function with a different bio if there are other bios
705 * queued.
706 */
707 BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
708
709 /* If tg->bps = -1, then BW is unlimited */
710 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
711 if (wait)
712 *wait = 0;
713 return 1;
714 }
715
716 /*
717 * If previous slice expired, start a new one otherwise renew/extend
718 * existing slice to make sure it is at least throtl_slice interval
719 * long since now.
720 */
721 if (throtl_slice_used(td, tg, rw))
722 throtl_start_new_slice(td, tg, rw);
723 else {
724 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
725 throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
726 }
727
728 if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
729 && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
730 if (wait)
731 *wait = 0;
732 return 1;
733 }
734
735 max_wait = max(bps_wait, iops_wait);
736
737 if (wait)
738 *wait = max_wait;
739
740 if (time_before(tg->slice_end[rw], jiffies + max_wait))
741 throtl_extend_slice(td, tg, rw, jiffies + max_wait);
742
743 return 0;
744}
745
746static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
747{
748 bool rw = bio_data_dir(bio);
749 bool sync = rw_is_sync(bio->bi_rw);
750
751 /* Charge the bio to the group */
752 tg->bytes_disp[rw] += bio->bi_size;
753 tg->io_disp[rw]++;
754
755 blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync);
756}
757
758static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
759 struct bio *bio)
760{
761 bool rw = bio_data_dir(bio);
762
763 bio_list_add(&tg->bio_lists[rw], bio);
764 /* Take a bio reference on tg */
765 throtl_ref_get_tg(tg);
766 tg->nr_queued[rw]++;
767 td->nr_queued[rw]++;
768 throtl_enqueue_tg(td, tg);
769}
770
771static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
772{
773 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
774 struct bio *bio;
775
776 if ((bio = bio_list_peek(&tg->bio_lists[READ])))
777 tg_may_dispatch(td, tg, bio, &read_wait);
778
779 if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
780 tg_may_dispatch(td, tg, bio, &write_wait);
781
782 min_wait = min(read_wait, write_wait);
783 disptime = jiffies + min_wait;
784
785 /* Update dispatch time */
786 throtl_dequeue_tg(td, tg);
787 tg->disptime = disptime;
788 throtl_enqueue_tg(td, tg);
789}
790
791static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
792 bool rw, struct bio_list *bl)
793{
794 struct bio *bio;
795
796 bio = bio_list_pop(&tg->bio_lists[rw]);
797 tg->nr_queued[rw]--;
798 /* Drop bio reference on tg */
799 throtl_put_tg(tg);
800
801 BUG_ON(td->nr_queued[rw] <= 0);
802 td->nr_queued[rw]--;
803
804 throtl_charge_bio(tg, bio);
805 bio_list_add(bl, bio);
806 bio->bi_rw |= REQ_THROTTLED;
807
808 throtl_trim_slice(td, tg, rw);
809}
810
811static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
812 struct bio_list *bl)
813{
814 unsigned int nr_reads = 0, nr_writes = 0;
815 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
816 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
817 struct bio *bio;
818
819 /* Try to dispatch 75% READS and 25% WRITES */
820
821 while ((bio = bio_list_peek(&tg->bio_lists[READ]))
822 && tg_may_dispatch(td, tg, bio, NULL)) {
823
824 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
825 nr_reads++;
826
827 if (nr_reads >= max_nr_reads)
828 break;
829 }
830
831 while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
832 && tg_may_dispatch(td, tg, bio, NULL)) {
833
834 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
835 nr_writes++;
836
837 if (nr_writes >= max_nr_writes)
838 break;
839 }
840
841 return nr_reads + nr_writes;
842}
843
844static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
845{
846 unsigned int nr_disp = 0;
847 struct throtl_grp *tg;
848 struct throtl_rb_root *st = &td->tg_service_tree;
849
850 while (1) {
851 tg = throtl_rb_first(st);
852
853 if (!tg)
854 break;
855
856 if (time_before(jiffies, tg->disptime))
857 break;
858
859 throtl_dequeue_tg(td, tg);
860
861 nr_disp += throtl_dispatch_tg(td, tg, bl);
862
863 if (tg->nr_queued[0] || tg->nr_queued[1]) {
864 tg_update_disptime(td, tg);
865 throtl_enqueue_tg(td, tg);
866 }
867
868 if (nr_disp >= throtl_quantum)
869 break;
870 }
871
872 return nr_disp;
873}
874
875static void throtl_process_limit_change(struct throtl_data *td)
876{
877 struct throtl_grp *tg;
878 struct hlist_node *pos, *n;
879
880 if (!td->limits_changed)
881 return;
882
883 xchg(&td->limits_changed, false);
884
885 throtl_log(td, "limits changed");
886
887 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
888 if (!tg->limits_changed)
889 continue;
890
891 if (!xchg(&tg->limits_changed, false))
892 continue;
893
894 throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
895 " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
896 tg->iops[READ], tg->iops[WRITE]);
897
898 /*
899 * Restart the slices for both READ and WRITES. It
900 * might happen that a group's limit are dropped
901 * suddenly and we don't want to account recently
902 * dispatched IO with new low rate
903 */
904 throtl_start_new_slice(td, tg, 0);
905 throtl_start_new_slice(td, tg, 1);
906
907 if (throtl_tg_on_rr(tg))
908 tg_update_disptime(td, tg);
909 }
910}
911
912/* Dispatch throttled bios. Should be called without queue lock held. */
913static int throtl_dispatch(struct request_queue *q)
914{
915 struct throtl_data *td = q->td;
916 unsigned int nr_disp = 0;
917 struct bio_list bio_list_on_stack;
918 struct bio *bio;
919 struct blk_plug plug;
920
921 spin_lock_irq(q->queue_lock);
922
923 throtl_process_limit_change(td);
924
925 if (!total_nr_queued(td))
926 goto out;
927
928 bio_list_init(&bio_list_on_stack);
929
930 throtl_log(td, "dispatch nr_queued=%u read=%u write=%u",
931 total_nr_queued(td), td->nr_queued[READ],
932 td->nr_queued[WRITE]);
933
934 nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
935
936 if (nr_disp)
937 throtl_log(td, "bios disp=%u", nr_disp);
938
939 throtl_schedule_next_dispatch(td);
940out:
941 spin_unlock_irq(q->queue_lock);
942
943 /*
944 * If we dispatched some requests, unplug the queue to make sure
945 * immediate dispatch
946 */
947 if (nr_disp) {
948 blk_start_plug(&plug);
949 while((bio = bio_list_pop(&bio_list_on_stack)))
950 generic_make_request(bio);
951 blk_finish_plug(&plug);
952 }
953 return nr_disp;
954}
955
956void blk_throtl_work(struct work_struct *work)
957{
958 struct throtl_data *td = container_of(work, struct throtl_data,
959 throtl_work.work);
960 struct request_queue *q = td->queue;
961
962 throtl_dispatch(q);
963}
964
965/* Call with queue lock held */
966static void
967throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay)
968{
969
970 struct delayed_work *dwork = &td->throtl_work;
971
972 /* schedule work if limits changed even if no bio is queued */
973 if (total_nr_queued(td) || td->limits_changed) {
974 /*
975 * We might have a work scheduled to be executed in future.
976 * Cancel that and schedule a new one.
977 */
978 __cancel_delayed_work(dwork);
979 queue_delayed_work(kthrotld_workqueue, dwork, delay);
980 throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
981 delay, jiffies);
982 }
983}
984
985static void
986throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg)
987{
988 /* Something wrong if we are trying to remove same group twice */
989 BUG_ON(hlist_unhashed(&tg->tg_node));
990
991 hlist_del_init(&tg->tg_node);
992
993 /*
994 * Put the reference taken at the time of creation so that when all
995 * queues are gone, group can be destroyed.
996 */
997 throtl_put_tg(tg);
998 td->nr_undestroyed_grps--;
999}
1000
1001static void throtl_release_tgs(struct throtl_data *td)
1002{
1003 struct hlist_node *pos, *n;
1004 struct throtl_grp *tg;
1005
1006 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
1007 /*
1008 * If cgroup removal path got to blk_group first and removed
1009 * it from cgroup list, then it will take care of destroying
1010 * cfqg also.
1011 */
1012 if (!blkiocg_del_blkio_group(&tg->blkg))
1013 throtl_destroy_tg(td, tg);
1014 }
1015}
1016
1017static void throtl_td_free(struct throtl_data *td)
1018{
1019 kfree(td);
1020}
1021
1022/*
1023 * Blk cgroup controller notification saying that blkio_group object is being
1024 * delinked as associated cgroup object is going away. That also means that
1025 * no new IO will come in this group. So get rid of this group as soon as
1026 * any pending IO in the group is finished.
1027 *
1028 * This function is called under rcu_read_lock(). key is the rcu protected
1029 * pointer. That means "key" is a valid throtl_data pointer as long as we are
1030 * rcu read lock.
1031 *
1032 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1033 * it should not be NULL as even if queue was going away, cgroup deltion
1034 * path got to it first.
1035 */
1036void throtl_unlink_blkio_group(void *key, struct blkio_group *blkg)
1037{
1038 unsigned long flags;
1039 struct throtl_data *td = key;
1040
1041 spin_lock_irqsave(td->queue->queue_lock, flags);
1042 throtl_destroy_tg(td, tg_of_blkg(blkg));
1043 spin_unlock_irqrestore(td->queue->queue_lock, flags);
1044}
1045
1046static void throtl_update_blkio_group_common(struct throtl_data *td,
1047 struct throtl_grp *tg)
1048{
1049 xchg(&tg->limits_changed, true);
1050 xchg(&td->limits_changed, true);
1051 /* Schedule a work now to process the limit change */
1052 throtl_schedule_delayed_work(td, 0);
1053}
1054
1055/*
1056 * For all update functions, key should be a valid pointer because these
1057 * update functions are called under blkcg_lock, that means, blkg is
1058 * valid and in turn key is valid. queue exit path can not race because
1059 * of blkcg_lock
1060 *
1061 * Can not take queue lock in update functions as queue lock under blkcg_lock
1062 * is not allowed. Under other paths we take blkcg_lock under queue_lock.
1063 */
1064static void throtl_update_blkio_group_read_bps(void *key,
1065 struct blkio_group *blkg, u64 read_bps)
1066{
1067 struct throtl_data *td = key;
1068 struct throtl_grp *tg = tg_of_blkg(blkg);
1069
1070 tg->bps[READ] = read_bps;
1071 throtl_update_blkio_group_common(td, tg);
1072}
1073
1074static void throtl_update_blkio_group_write_bps(void *key,
1075 struct blkio_group *blkg, u64 write_bps)
1076{
1077 struct throtl_data *td = key;
1078 struct throtl_grp *tg = tg_of_blkg(blkg);
1079
1080 tg->bps[WRITE] = write_bps;
1081 throtl_update_blkio_group_common(td, tg);
1082}
1083
1084static void throtl_update_blkio_group_read_iops(void *key,
1085 struct blkio_group *blkg, unsigned int read_iops)
1086{
1087 struct throtl_data *td = key;
1088 struct throtl_grp *tg = tg_of_blkg(blkg);
1089
1090 tg->iops[READ] = read_iops;
1091 throtl_update_blkio_group_common(td, tg);
1092}
1093
1094static void throtl_update_blkio_group_write_iops(void *key,
1095 struct blkio_group *blkg, unsigned int write_iops)
1096{
1097 struct throtl_data *td = key;
1098 struct throtl_grp *tg = tg_of_blkg(blkg);
1099
1100 tg->iops[WRITE] = write_iops;
1101 throtl_update_blkio_group_common(td, tg);
1102}
1103
1104static void throtl_shutdown_wq(struct request_queue *q)
1105{
1106 struct throtl_data *td = q->td;
1107
1108 cancel_delayed_work_sync(&td->throtl_work);
1109}
1110
1111static struct blkio_policy_type blkio_policy_throtl = {
1112 .ops = {
1113 .blkio_unlink_group_fn = throtl_unlink_blkio_group,
1114 .blkio_update_group_read_bps_fn =
1115 throtl_update_blkio_group_read_bps,
1116 .blkio_update_group_write_bps_fn =
1117 throtl_update_blkio_group_write_bps,
1118 .blkio_update_group_read_iops_fn =
1119 throtl_update_blkio_group_read_iops,
1120 .blkio_update_group_write_iops_fn =
1121 throtl_update_blkio_group_write_iops,
1122 },
1123 .plid = BLKIO_POLICY_THROTL,
1124};
1125
1126int blk_throtl_bio(struct request_queue *q, struct bio **biop)
1127{
1128 struct throtl_data *td = q->td;
1129 struct throtl_grp *tg;
1130 struct bio *bio = *biop;
1131 bool rw = bio_data_dir(bio), update_disptime = true;
1132 struct blkio_cgroup *blkcg;
1133
1134 if (bio->bi_rw & REQ_THROTTLED) {
1135 bio->bi_rw &= ~REQ_THROTTLED;
1136 return 0;
1137 }
1138
1139 /*
1140 * A throtl_grp pointer retrieved under rcu can be used to access
1141 * basic fields like stats and io rates. If a group has no rules,
1142 * just update the dispatch stats in lockless manner and return.
1143 */
1144
1145 rcu_read_lock();
1146 blkcg = task_blkio_cgroup(current);
1147 tg = throtl_find_tg(td, blkcg);
1148 if (tg) {
1149 throtl_tg_fill_dev_details(td, tg);
1150
1151 if (tg_no_rule_group(tg, rw)) {
1152 blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size,
1153 rw, rw_is_sync(bio->bi_rw));
1154 rcu_read_unlock();
1155 return 0;
1156 }
1157 }
1158 rcu_read_unlock();
1159
1160 /*
1161 * Either group has not been allocated yet or it is not an unlimited
1162 * IO group
1163 */
1164
1165 spin_lock_irq(q->queue_lock);
1166 tg = throtl_get_tg(td);
1167
1168 if (IS_ERR(tg)) {
1169 if (PTR_ERR(tg) == -ENODEV) {
1170 /*
1171 * Queue is gone. No queue lock held here.
1172 */
1173 return -ENODEV;
1174 }
1175 }
1176
1177 if (tg->nr_queued[rw]) {
1178 /*
1179 * There is already another bio queued in same dir. No
1180 * need to update dispatch time.
1181 */
1182 update_disptime = false;
1183 goto queue_bio;
1184
1185 }
1186
1187 /* Bio is with-in rate limit of group */
1188 if (tg_may_dispatch(td, tg, bio, NULL)) {
1189 throtl_charge_bio(tg, bio);
1190
1191 /*
1192 * We need to trim slice even when bios are not being queued
1193 * otherwise it might happen that a bio is not queued for
1194 * a long time and slice keeps on extending and trim is not
1195 * called for a long time. Now if limits are reduced suddenly
1196 * we take into account all the IO dispatched so far at new
1197 * low rate and * newly queued IO gets a really long dispatch
1198 * time.
1199 *
1200 * So keep on trimming slice even if bio is not queued.
1201 */
1202 throtl_trim_slice(td, tg, rw);
1203 goto out;
1204 }
1205
1206queue_bio:
1207 throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
1208 " iodisp=%u iops=%u queued=%d/%d",
1209 rw == READ ? 'R' : 'W',
1210 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1211 tg->io_disp[rw], tg->iops[rw],
1212 tg->nr_queued[READ], tg->nr_queued[WRITE]);
1213
1214 throtl_add_bio_tg(q->td, tg, bio);
1215 *biop = NULL;
1216
1217 if (update_disptime) {
1218 tg_update_disptime(td, tg);
1219 throtl_schedule_next_dispatch(td);
1220 }
1221
1222out:
1223 spin_unlock_irq(q->queue_lock);
1224 return 0;
1225}
1226
1227int blk_throtl_init(struct request_queue *q)
1228{
1229 struct throtl_data *td;
1230 struct throtl_grp *tg;
1231
1232 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1233 if (!td)
1234 return -ENOMEM;
1235
1236 INIT_HLIST_HEAD(&td->tg_list);
1237 td->tg_service_tree = THROTL_RB_ROOT;
1238 td->limits_changed = false;
1239 INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);
1240
1241 /* alloc and Init root group. */
1242 td->queue = q;
1243 tg = throtl_alloc_tg(td);
1244
1245 if (!tg) {
1246 kfree(td);
1247 return -ENOMEM;
1248 }
1249
1250 td->root_tg = tg;
1251
1252 rcu_read_lock();
1253 throtl_init_add_tg_lists(td, tg, &blkio_root_cgroup);
1254 rcu_read_unlock();
1255
1256 /* Attach throtl data to request queue */
1257 q->td = td;
1258 return 0;
1259}
1260
1261void blk_throtl_exit(struct request_queue *q)
1262{
1263 struct throtl_data *td = q->td;
1264 bool wait = false;
1265
1266 BUG_ON(!td);
1267
1268 throtl_shutdown_wq(q);
1269
1270 spin_lock_irq(q->queue_lock);
1271 throtl_release_tgs(td);
1272
1273 /* If there are other groups */
1274 if (td->nr_undestroyed_grps > 0)
1275 wait = true;
1276
1277 spin_unlock_irq(q->queue_lock);
1278
1279 /*
1280 * Wait for tg->blkg->key accessors to exit their grace periods.
1281 * Do this wait only if there are other undestroyed groups out
1282 * there (other than root group). This can happen if cgroup deletion
1283 * path claimed the responsibility of cleaning up a group before
1284 * queue cleanup code get to the group.
1285 *
1286 * Do not call synchronize_rcu() unconditionally as there are drivers
1287 * which create/delete request queue hundreds of times during scan/boot
1288 * and synchronize_rcu() can take significant time and slow down boot.
1289 */
1290 if (wait)
1291 synchronize_rcu();
1292
1293 /*
1294 * Just being safe to make sure after previous flush if some body did
1295 * update limits through cgroup and another work got queued, cancel
1296 * it.
1297 */
1298 throtl_shutdown_wq(q);
1299 throtl_td_free(td);
1300}
1301
1302static int __init throtl_init(void)
1303{
1304 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1305 if (!kthrotld_workqueue)
1306 panic("Failed to create kthrotld\n");
1307
1308 blkio_policy_register(&blkio_policy_throtl);
1309 return 0;
1310}
1311
1312module_init(throtl_init);