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 "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);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Interface for controlling IO bandwidth on a request queue
4 *
5 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
6 */
7
8#include <linux/module.h>
9#include <linux/slab.h>
10#include <linux/blkdev.h>
11#include <linux/bio.h>
12#include <linux/blktrace_api.h>
13#include <linux/blk-cgroup.h>
14#include "blk.h"
15#include "blk-cgroup-rwstat.h"
16
17/* Max dispatch from a group in 1 round */
18#define THROTL_GRP_QUANTUM 8
19
20/* Total max dispatch from all groups in one round */
21#define THROTL_QUANTUM 32
22
23/* Throttling is performed over a slice and after that slice is renewed */
24#define DFL_THROTL_SLICE_HD (HZ / 10)
25#define DFL_THROTL_SLICE_SSD (HZ / 50)
26#define MAX_THROTL_SLICE (HZ)
27#define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
28#define MIN_THROTL_BPS (320 * 1024)
29#define MIN_THROTL_IOPS (10)
30#define DFL_LATENCY_TARGET (-1L)
31#define DFL_IDLE_THRESHOLD (0)
32#define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
33#define LATENCY_FILTERED_SSD (0)
34/*
35 * For HD, very small latency comes from sequential IO. Such IO is helpless to
36 * help determine if its IO is impacted by others, hence we ignore the IO
37 */
38#define LATENCY_FILTERED_HD (1000L) /* 1ms */
39
40static struct blkcg_policy blkcg_policy_throtl;
41
42/* A workqueue to queue throttle related work */
43static struct workqueue_struct *kthrotld_workqueue;
44
45/*
46 * To implement hierarchical throttling, throtl_grps form a tree and bios
47 * are dispatched upwards level by level until they reach the top and get
48 * issued. When dispatching bios from the children and local group at each
49 * level, if the bios are dispatched into a single bio_list, there's a risk
50 * of a local or child group which can queue many bios at once filling up
51 * the list starving others.
52 *
53 * To avoid such starvation, dispatched bios are queued separately
54 * according to where they came from. When they are again dispatched to
55 * the parent, they're popped in round-robin order so that no single source
56 * hogs the dispatch window.
57 *
58 * throtl_qnode is used to keep the queued bios separated by their sources.
59 * Bios are queued to throtl_qnode which in turn is queued to
60 * throtl_service_queue and then dispatched in round-robin order.
61 *
62 * It's also used to track the reference counts on blkg's. A qnode always
63 * belongs to a throtl_grp and gets queued on itself or the parent, so
64 * incrementing the reference of the associated throtl_grp when a qnode is
65 * queued and decrementing when dequeued is enough to keep the whole blkg
66 * tree pinned while bios are in flight.
67 */
68struct throtl_qnode {
69 struct list_head node; /* service_queue->queued[] */
70 struct bio_list bios; /* queued bios */
71 struct throtl_grp *tg; /* tg this qnode belongs to */
72};
73
74struct throtl_service_queue {
75 struct throtl_service_queue *parent_sq; /* the parent service_queue */
76
77 /*
78 * Bios queued directly to this service_queue or dispatched from
79 * children throtl_grp's.
80 */
81 struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */
82 unsigned int nr_queued[2]; /* number of queued bios */
83
84 /*
85 * RB tree of active children throtl_grp's, which are sorted by
86 * their ->disptime.
87 */
88 struct rb_root_cached pending_tree; /* RB tree of active tgs */
89 unsigned int nr_pending; /* # queued in the tree */
90 unsigned long first_pending_disptime; /* disptime of the first tg */
91 struct timer_list pending_timer; /* fires on first_pending_disptime */
92};
93
94enum tg_state_flags {
95 THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */
96 THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */
97};
98
99#define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
100
101enum {
102 LIMIT_LOW,
103 LIMIT_MAX,
104 LIMIT_CNT,
105};
106
107struct throtl_grp {
108 /* must be the first member */
109 struct blkg_policy_data pd;
110
111 /* active throtl group service_queue member */
112 struct rb_node rb_node;
113
114 /* throtl_data this group belongs to */
115 struct throtl_data *td;
116
117 /* this group's service queue */
118 struct throtl_service_queue service_queue;
119
120 /*
121 * qnode_on_self is used when bios are directly queued to this
122 * throtl_grp so that local bios compete fairly with bios
123 * dispatched from children. qnode_on_parent is used when bios are
124 * dispatched from this throtl_grp into its parent and will compete
125 * with the sibling qnode_on_parents and the parent's
126 * qnode_on_self.
127 */
128 struct throtl_qnode qnode_on_self[2];
129 struct throtl_qnode qnode_on_parent[2];
130
131 /*
132 * Dispatch time in jiffies. This is the estimated time when group
133 * will unthrottle and is ready to dispatch more bio. It is used as
134 * key to sort active groups in service tree.
135 */
136 unsigned long disptime;
137
138 unsigned int flags;
139
140 /* are there any throtl rules between this group and td? */
141 bool has_rules[2];
142
143 /* internally used bytes per second rate limits */
144 uint64_t bps[2][LIMIT_CNT];
145 /* user configured bps limits */
146 uint64_t bps_conf[2][LIMIT_CNT];
147
148 /* internally used IOPS limits */
149 unsigned int iops[2][LIMIT_CNT];
150 /* user configured IOPS limits */
151 unsigned int iops_conf[2][LIMIT_CNT];
152
153 /* Number of bytes dispatched in current slice */
154 uint64_t bytes_disp[2];
155 /* Number of bio's dispatched in current slice */
156 unsigned int io_disp[2];
157
158 unsigned long last_low_overflow_time[2];
159
160 uint64_t last_bytes_disp[2];
161 unsigned int last_io_disp[2];
162
163 unsigned long last_check_time;
164
165 unsigned long latency_target; /* us */
166 unsigned long latency_target_conf; /* us */
167 /* When did we start a new slice */
168 unsigned long slice_start[2];
169 unsigned long slice_end[2];
170
171 unsigned long last_finish_time; /* ns / 1024 */
172 unsigned long checked_last_finish_time; /* ns / 1024 */
173 unsigned long avg_idletime; /* ns / 1024 */
174 unsigned long idletime_threshold; /* us */
175 unsigned long idletime_threshold_conf; /* us */
176
177 unsigned int bio_cnt; /* total bios */
178 unsigned int bad_bio_cnt; /* bios exceeding latency threshold */
179 unsigned long bio_cnt_reset_time;
180
181 atomic_t io_split_cnt[2];
182 atomic_t last_io_split_cnt[2];
183
184 struct blkg_rwstat stat_bytes;
185 struct blkg_rwstat stat_ios;
186};
187
188/* We measure latency for request size from <= 4k to >= 1M */
189#define LATENCY_BUCKET_SIZE 9
190
191struct latency_bucket {
192 unsigned long total_latency; /* ns / 1024 */
193 int samples;
194};
195
196struct avg_latency_bucket {
197 unsigned long latency; /* ns / 1024 */
198 bool valid;
199};
200
201struct throtl_data
202{
203 /* service tree for active throtl groups */
204 struct throtl_service_queue service_queue;
205
206 struct request_queue *queue;
207
208 /* Total Number of queued bios on READ and WRITE lists */
209 unsigned int nr_queued[2];
210
211 unsigned int throtl_slice;
212
213 /* Work for dispatching throttled bios */
214 struct work_struct dispatch_work;
215 unsigned int limit_index;
216 bool limit_valid[LIMIT_CNT];
217
218 unsigned long low_upgrade_time;
219 unsigned long low_downgrade_time;
220
221 unsigned int scale;
222
223 struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
224 struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
225 struct latency_bucket __percpu *latency_buckets[2];
226 unsigned long last_calculate_time;
227 unsigned long filtered_latency;
228
229 bool track_bio_latency;
230};
231
232static void throtl_pending_timer_fn(struct timer_list *t);
233
234static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
235{
236 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
237}
238
239static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
240{
241 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
242}
243
244static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
245{
246 return pd_to_blkg(&tg->pd);
247}
248
249/**
250 * sq_to_tg - return the throl_grp the specified service queue belongs to
251 * @sq: the throtl_service_queue of interest
252 *
253 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
254 * embedded in throtl_data, %NULL is returned.
255 */
256static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
257{
258 if (sq && sq->parent_sq)
259 return container_of(sq, struct throtl_grp, service_queue);
260 else
261 return NULL;
262}
263
264/**
265 * sq_to_td - return throtl_data the specified service queue belongs to
266 * @sq: the throtl_service_queue of interest
267 *
268 * A service_queue can be embedded in either a throtl_grp or throtl_data.
269 * Determine the associated throtl_data accordingly and return it.
270 */
271static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
272{
273 struct throtl_grp *tg = sq_to_tg(sq);
274
275 if (tg)
276 return tg->td;
277 else
278 return container_of(sq, struct throtl_data, service_queue);
279}
280
281/*
282 * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
283 * make the IO dispatch more smooth.
284 * Scale up: linearly scale up according to lapsed time since upgrade. For
285 * every throtl_slice, the limit scales up 1/2 .low limit till the
286 * limit hits .max limit
287 * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
288 */
289static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
290{
291 /* arbitrary value to avoid too big scale */
292 if (td->scale < 4096 && time_after_eq(jiffies,
293 td->low_upgrade_time + td->scale * td->throtl_slice))
294 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
295
296 return low + (low >> 1) * td->scale;
297}
298
299static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
300{
301 struct blkcg_gq *blkg = tg_to_blkg(tg);
302 struct throtl_data *td;
303 uint64_t ret;
304
305 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
306 return U64_MAX;
307
308 td = tg->td;
309 ret = tg->bps[rw][td->limit_index];
310 if (ret == 0 && td->limit_index == LIMIT_LOW) {
311 /* intermediate node or iops isn't 0 */
312 if (!list_empty(&blkg->blkcg->css.children) ||
313 tg->iops[rw][td->limit_index])
314 return U64_MAX;
315 else
316 return MIN_THROTL_BPS;
317 }
318
319 if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
320 tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
321 uint64_t adjusted;
322
323 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
324 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
325 }
326 return ret;
327}
328
329static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
330{
331 struct blkcg_gq *blkg = tg_to_blkg(tg);
332 struct throtl_data *td;
333 unsigned int ret;
334
335 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
336 return UINT_MAX;
337
338 td = tg->td;
339 ret = tg->iops[rw][td->limit_index];
340 if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
341 /* intermediate node or bps isn't 0 */
342 if (!list_empty(&blkg->blkcg->css.children) ||
343 tg->bps[rw][td->limit_index])
344 return UINT_MAX;
345 else
346 return MIN_THROTL_IOPS;
347 }
348
349 if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
350 tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
351 uint64_t adjusted;
352
353 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
354 if (adjusted > UINT_MAX)
355 adjusted = UINT_MAX;
356 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
357 }
358 return ret;
359}
360
361#define request_bucket_index(sectors) \
362 clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
363
364/**
365 * throtl_log - log debug message via blktrace
366 * @sq: the service_queue being reported
367 * @fmt: printf format string
368 * @args: printf args
369 *
370 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
371 * throtl_grp; otherwise, just "throtl".
372 */
373#define throtl_log(sq, fmt, args...) do { \
374 struct throtl_grp *__tg = sq_to_tg((sq)); \
375 struct throtl_data *__td = sq_to_td((sq)); \
376 \
377 (void)__td; \
378 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
379 break; \
380 if ((__tg)) { \
381 blk_add_cgroup_trace_msg(__td->queue, \
382 tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\
383 } else { \
384 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
385 } \
386} while (0)
387
388static inline unsigned int throtl_bio_data_size(struct bio *bio)
389{
390 /* assume it's one sector */
391 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
392 return 512;
393 return bio->bi_iter.bi_size;
394}
395
396static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
397{
398 INIT_LIST_HEAD(&qn->node);
399 bio_list_init(&qn->bios);
400 qn->tg = tg;
401}
402
403/**
404 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
405 * @bio: bio being added
406 * @qn: qnode to add bio to
407 * @queued: the service_queue->queued[] list @qn belongs to
408 *
409 * Add @bio to @qn and put @qn on @queued if it's not already on.
410 * @qn->tg's reference count is bumped when @qn is activated. See the
411 * comment on top of throtl_qnode definition for details.
412 */
413static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
414 struct list_head *queued)
415{
416 bio_list_add(&qn->bios, bio);
417 if (list_empty(&qn->node)) {
418 list_add_tail(&qn->node, queued);
419 blkg_get(tg_to_blkg(qn->tg));
420 }
421}
422
423/**
424 * throtl_peek_queued - peek the first bio on a qnode list
425 * @queued: the qnode list to peek
426 */
427static struct bio *throtl_peek_queued(struct list_head *queued)
428{
429 struct throtl_qnode *qn;
430 struct bio *bio;
431
432 if (list_empty(queued))
433 return NULL;
434
435 qn = list_first_entry(queued, struct throtl_qnode, node);
436 bio = bio_list_peek(&qn->bios);
437 WARN_ON_ONCE(!bio);
438 return bio;
439}
440
441/**
442 * throtl_pop_queued - pop the first bio form a qnode list
443 * @queued: the qnode list to pop a bio from
444 * @tg_to_put: optional out argument for throtl_grp to put
445 *
446 * Pop the first bio from the qnode list @queued. After popping, the first
447 * qnode is removed from @queued if empty or moved to the end of @queued so
448 * that the popping order is round-robin.
449 *
450 * When the first qnode is removed, its associated throtl_grp should be put
451 * too. If @tg_to_put is NULL, this function automatically puts it;
452 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
453 * responsible for putting it.
454 */
455static struct bio *throtl_pop_queued(struct list_head *queued,
456 struct throtl_grp **tg_to_put)
457{
458 struct throtl_qnode *qn;
459 struct bio *bio;
460
461 if (list_empty(queued))
462 return NULL;
463
464 qn = list_first_entry(queued, struct throtl_qnode, node);
465 bio = bio_list_pop(&qn->bios);
466 WARN_ON_ONCE(!bio);
467
468 if (bio_list_empty(&qn->bios)) {
469 list_del_init(&qn->node);
470 if (tg_to_put)
471 *tg_to_put = qn->tg;
472 else
473 blkg_put(tg_to_blkg(qn->tg));
474 } else {
475 list_move_tail(&qn->node, queued);
476 }
477
478 return bio;
479}
480
481/* init a service_queue, assumes the caller zeroed it */
482static void throtl_service_queue_init(struct throtl_service_queue *sq)
483{
484 INIT_LIST_HEAD(&sq->queued[0]);
485 INIT_LIST_HEAD(&sq->queued[1]);
486 sq->pending_tree = RB_ROOT_CACHED;
487 timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
488}
489
490static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp,
491 struct request_queue *q,
492 struct blkcg *blkcg)
493{
494 struct throtl_grp *tg;
495 int rw;
496
497 tg = kzalloc_node(sizeof(*tg), gfp, q->node);
498 if (!tg)
499 return NULL;
500
501 if (blkg_rwstat_init(&tg->stat_bytes, gfp))
502 goto err_free_tg;
503
504 if (blkg_rwstat_init(&tg->stat_ios, gfp))
505 goto err_exit_stat_bytes;
506
507 throtl_service_queue_init(&tg->service_queue);
508
509 for (rw = READ; rw <= WRITE; rw++) {
510 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
511 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
512 }
513
514 RB_CLEAR_NODE(&tg->rb_node);
515 tg->bps[READ][LIMIT_MAX] = U64_MAX;
516 tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
517 tg->iops[READ][LIMIT_MAX] = UINT_MAX;
518 tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
519 tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
520 tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
521 tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
522 tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
523 /* LIMIT_LOW will have default value 0 */
524
525 tg->latency_target = DFL_LATENCY_TARGET;
526 tg->latency_target_conf = DFL_LATENCY_TARGET;
527 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
528 tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
529
530 return &tg->pd;
531
532err_exit_stat_bytes:
533 blkg_rwstat_exit(&tg->stat_bytes);
534err_free_tg:
535 kfree(tg);
536 return NULL;
537}
538
539static void throtl_pd_init(struct blkg_policy_data *pd)
540{
541 struct throtl_grp *tg = pd_to_tg(pd);
542 struct blkcg_gq *blkg = tg_to_blkg(tg);
543 struct throtl_data *td = blkg->q->td;
544 struct throtl_service_queue *sq = &tg->service_queue;
545
546 /*
547 * If on the default hierarchy, we switch to properly hierarchical
548 * behavior where limits on a given throtl_grp are applied to the
549 * whole subtree rather than just the group itself. e.g. If 16M
550 * read_bps limit is set on the root group, the whole system can't
551 * exceed 16M for the device.
552 *
553 * If not on the default hierarchy, the broken flat hierarchy
554 * behavior is retained where all throtl_grps are treated as if
555 * they're all separate root groups right below throtl_data.
556 * Limits of a group don't interact with limits of other groups
557 * regardless of the position of the group in the hierarchy.
558 */
559 sq->parent_sq = &td->service_queue;
560 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
561 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
562 tg->td = td;
563}
564
565/*
566 * Set has_rules[] if @tg or any of its parents have limits configured.
567 * This doesn't require walking up to the top of the hierarchy as the
568 * parent's has_rules[] is guaranteed to be correct.
569 */
570static void tg_update_has_rules(struct throtl_grp *tg)
571{
572 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
573 struct throtl_data *td = tg->td;
574 int rw;
575
576 for (rw = READ; rw <= WRITE; rw++)
577 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
578 (td->limit_valid[td->limit_index] &&
579 (tg_bps_limit(tg, rw) != U64_MAX ||
580 tg_iops_limit(tg, rw) != UINT_MAX));
581}
582
583static void throtl_pd_online(struct blkg_policy_data *pd)
584{
585 struct throtl_grp *tg = pd_to_tg(pd);
586 /*
587 * We don't want new groups to escape the limits of its ancestors.
588 * Update has_rules[] after a new group is brought online.
589 */
590 tg_update_has_rules(tg);
591}
592
593#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
594static void blk_throtl_update_limit_valid(struct throtl_data *td)
595{
596 struct cgroup_subsys_state *pos_css;
597 struct blkcg_gq *blkg;
598 bool low_valid = false;
599
600 rcu_read_lock();
601 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
602 struct throtl_grp *tg = blkg_to_tg(blkg);
603
604 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
605 tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
606 low_valid = true;
607 break;
608 }
609 }
610 rcu_read_unlock();
611
612 td->limit_valid[LIMIT_LOW] = low_valid;
613}
614#else
615static inline void blk_throtl_update_limit_valid(struct throtl_data *td)
616{
617}
618#endif
619
620static void throtl_upgrade_state(struct throtl_data *td);
621static void throtl_pd_offline(struct blkg_policy_data *pd)
622{
623 struct throtl_grp *tg = pd_to_tg(pd);
624
625 tg->bps[READ][LIMIT_LOW] = 0;
626 tg->bps[WRITE][LIMIT_LOW] = 0;
627 tg->iops[READ][LIMIT_LOW] = 0;
628 tg->iops[WRITE][LIMIT_LOW] = 0;
629
630 blk_throtl_update_limit_valid(tg->td);
631
632 if (!tg->td->limit_valid[tg->td->limit_index])
633 throtl_upgrade_state(tg->td);
634}
635
636static void throtl_pd_free(struct blkg_policy_data *pd)
637{
638 struct throtl_grp *tg = pd_to_tg(pd);
639
640 del_timer_sync(&tg->service_queue.pending_timer);
641 blkg_rwstat_exit(&tg->stat_bytes);
642 blkg_rwstat_exit(&tg->stat_ios);
643 kfree(tg);
644}
645
646static struct throtl_grp *
647throtl_rb_first(struct throtl_service_queue *parent_sq)
648{
649 struct rb_node *n;
650
651 n = rb_first_cached(&parent_sq->pending_tree);
652 WARN_ON_ONCE(!n);
653 if (!n)
654 return NULL;
655 return rb_entry_tg(n);
656}
657
658static void throtl_rb_erase(struct rb_node *n,
659 struct throtl_service_queue *parent_sq)
660{
661 rb_erase_cached(n, &parent_sq->pending_tree);
662 RB_CLEAR_NODE(n);
663 --parent_sq->nr_pending;
664}
665
666static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
667{
668 struct throtl_grp *tg;
669
670 tg = throtl_rb_first(parent_sq);
671 if (!tg)
672 return;
673
674 parent_sq->first_pending_disptime = tg->disptime;
675}
676
677static void tg_service_queue_add(struct throtl_grp *tg)
678{
679 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
680 struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
681 struct rb_node *parent = NULL;
682 struct throtl_grp *__tg;
683 unsigned long key = tg->disptime;
684 bool leftmost = true;
685
686 while (*node != NULL) {
687 parent = *node;
688 __tg = rb_entry_tg(parent);
689
690 if (time_before(key, __tg->disptime))
691 node = &parent->rb_left;
692 else {
693 node = &parent->rb_right;
694 leftmost = false;
695 }
696 }
697
698 rb_link_node(&tg->rb_node, parent, node);
699 rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
700 leftmost);
701}
702
703static void throtl_enqueue_tg(struct throtl_grp *tg)
704{
705 if (!(tg->flags & THROTL_TG_PENDING)) {
706 tg_service_queue_add(tg);
707 tg->flags |= THROTL_TG_PENDING;
708 tg->service_queue.parent_sq->nr_pending++;
709 }
710}
711
712static void throtl_dequeue_tg(struct throtl_grp *tg)
713{
714 if (tg->flags & THROTL_TG_PENDING) {
715 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
716 tg->flags &= ~THROTL_TG_PENDING;
717 }
718}
719
720/* Call with queue lock held */
721static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
722 unsigned long expires)
723{
724 unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
725
726 /*
727 * Since we are adjusting the throttle limit dynamically, the sleep
728 * time calculated according to previous limit might be invalid. It's
729 * possible the cgroup sleep time is very long and no other cgroups
730 * have IO running so notify the limit changes. Make sure the cgroup
731 * doesn't sleep too long to avoid the missed notification.
732 */
733 if (time_after(expires, max_expire))
734 expires = max_expire;
735 mod_timer(&sq->pending_timer, expires);
736 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
737 expires - jiffies, jiffies);
738}
739
740/**
741 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
742 * @sq: the service_queue to schedule dispatch for
743 * @force: force scheduling
744 *
745 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
746 * dispatch time of the first pending child. Returns %true if either timer
747 * is armed or there's no pending child left. %false if the current
748 * dispatch window is still open and the caller should continue
749 * dispatching.
750 *
751 * If @force is %true, the dispatch timer is always scheduled and this
752 * function is guaranteed to return %true. This is to be used when the
753 * caller can't dispatch itself and needs to invoke pending_timer
754 * unconditionally. Note that forced scheduling is likely to induce short
755 * delay before dispatch starts even if @sq->first_pending_disptime is not
756 * in the future and thus shouldn't be used in hot paths.
757 */
758static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
759 bool force)
760{
761 /* any pending children left? */
762 if (!sq->nr_pending)
763 return true;
764
765 update_min_dispatch_time(sq);
766
767 /* is the next dispatch time in the future? */
768 if (force || time_after(sq->first_pending_disptime, jiffies)) {
769 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
770 return true;
771 }
772
773 /* tell the caller to continue dispatching */
774 return false;
775}
776
777static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
778 bool rw, unsigned long start)
779{
780 tg->bytes_disp[rw] = 0;
781 tg->io_disp[rw] = 0;
782
783 atomic_set(&tg->io_split_cnt[rw], 0);
784
785 /*
786 * Previous slice has expired. We must have trimmed it after last
787 * bio dispatch. That means since start of last slice, we never used
788 * that bandwidth. Do try to make use of that bandwidth while giving
789 * credit.
790 */
791 if (time_after_eq(start, tg->slice_start[rw]))
792 tg->slice_start[rw] = start;
793
794 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
795 throtl_log(&tg->service_queue,
796 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
797 rw == READ ? 'R' : 'W', tg->slice_start[rw],
798 tg->slice_end[rw], jiffies);
799}
800
801static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
802{
803 tg->bytes_disp[rw] = 0;
804 tg->io_disp[rw] = 0;
805 tg->slice_start[rw] = jiffies;
806 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
807
808 atomic_set(&tg->io_split_cnt[rw], 0);
809
810 throtl_log(&tg->service_queue,
811 "[%c] new slice start=%lu end=%lu jiffies=%lu",
812 rw == READ ? 'R' : 'W', tg->slice_start[rw],
813 tg->slice_end[rw], jiffies);
814}
815
816static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
817 unsigned long jiffy_end)
818{
819 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
820}
821
822static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
823 unsigned long jiffy_end)
824{
825 throtl_set_slice_end(tg, rw, jiffy_end);
826 throtl_log(&tg->service_queue,
827 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
828 rw == READ ? 'R' : 'W', tg->slice_start[rw],
829 tg->slice_end[rw], jiffies);
830}
831
832/* Determine if previously allocated or extended slice is complete or not */
833static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
834{
835 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
836 return false;
837
838 return true;
839}
840
841/* Trim the used slices and adjust slice start accordingly */
842static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
843{
844 unsigned long nr_slices, time_elapsed, io_trim;
845 u64 bytes_trim, tmp;
846
847 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
848
849 /*
850 * If bps are unlimited (-1), then time slice don't get
851 * renewed. Don't try to trim the slice if slice is used. A new
852 * slice will start when appropriate.
853 */
854 if (throtl_slice_used(tg, rw))
855 return;
856
857 /*
858 * A bio has been dispatched. Also adjust slice_end. It might happen
859 * that initially cgroup limit was very low resulting in high
860 * slice_end, but later limit was bumped up and bio was dispatched
861 * sooner, then we need to reduce slice_end. A high bogus slice_end
862 * is bad because it does not allow new slice to start.
863 */
864
865 throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
866
867 time_elapsed = jiffies - tg->slice_start[rw];
868
869 nr_slices = time_elapsed / tg->td->throtl_slice;
870
871 if (!nr_slices)
872 return;
873 tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
874 do_div(tmp, HZ);
875 bytes_trim = tmp;
876
877 io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
878 HZ;
879
880 if (!bytes_trim && !io_trim)
881 return;
882
883 if (tg->bytes_disp[rw] >= bytes_trim)
884 tg->bytes_disp[rw] -= bytes_trim;
885 else
886 tg->bytes_disp[rw] = 0;
887
888 if (tg->io_disp[rw] >= io_trim)
889 tg->io_disp[rw] -= io_trim;
890 else
891 tg->io_disp[rw] = 0;
892
893 tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
894
895 throtl_log(&tg->service_queue,
896 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
897 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
898 tg->slice_start[rw], tg->slice_end[rw], jiffies);
899}
900
901static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
902 u32 iops_limit, unsigned long *wait)
903{
904 bool rw = bio_data_dir(bio);
905 unsigned int io_allowed;
906 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
907 u64 tmp;
908
909 if (iops_limit == UINT_MAX) {
910 if (wait)
911 *wait = 0;
912 return true;
913 }
914
915 jiffy_elapsed = jiffies - tg->slice_start[rw];
916
917 /* Round up to the next throttle slice, wait time must be nonzero */
918 jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
919
920 /*
921 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
922 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
923 * will allow dispatch after 1 second and after that slice should
924 * have been trimmed.
925 */
926
927 tmp = (u64)iops_limit * jiffy_elapsed_rnd;
928 do_div(tmp, HZ);
929
930 if (tmp > UINT_MAX)
931 io_allowed = UINT_MAX;
932 else
933 io_allowed = tmp;
934
935 if (tg->io_disp[rw] + 1 <= io_allowed) {
936 if (wait)
937 *wait = 0;
938 return true;
939 }
940
941 /* Calc approx time to dispatch */
942 jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
943
944 if (wait)
945 *wait = jiffy_wait;
946 return false;
947}
948
949static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
950 u64 bps_limit, unsigned long *wait)
951{
952 bool rw = bio_data_dir(bio);
953 u64 bytes_allowed, extra_bytes, tmp;
954 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
955 unsigned int bio_size = throtl_bio_data_size(bio);
956
957 if (bps_limit == U64_MAX) {
958 if (wait)
959 *wait = 0;
960 return true;
961 }
962
963 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
964
965 /* Slice has just started. Consider one slice interval */
966 if (!jiffy_elapsed)
967 jiffy_elapsed_rnd = tg->td->throtl_slice;
968
969 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
970
971 tmp = bps_limit * jiffy_elapsed_rnd;
972 do_div(tmp, HZ);
973 bytes_allowed = tmp;
974
975 if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
976 if (wait)
977 *wait = 0;
978 return true;
979 }
980
981 /* Calc approx time to dispatch */
982 extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
983 jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
984
985 if (!jiffy_wait)
986 jiffy_wait = 1;
987
988 /*
989 * This wait time is without taking into consideration the rounding
990 * up we did. Add that time also.
991 */
992 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
993 if (wait)
994 *wait = jiffy_wait;
995 return false;
996}
997
998/*
999 * Returns whether one can dispatch a bio or not. Also returns approx number
1000 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
1001 */
1002static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
1003 unsigned long *wait)
1004{
1005 bool rw = bio_data_dir(bio);
1006 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
1007 u64 bps_limit = tg_bps_limit(tg, rw);
1008 u32 iops_limit = tg_iops_limit(tg, rw);
1009
1010 /*
1011 * Currently whole state machine of group depends on first bio
1012 * queued in the group bio list. So one should not be calling
1013 * this function with a different bio if there are other bios
1014 * queued.
1015 */
1016 BUG_ON(tg->service_queue.nr_queued[rw] &&
1017 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
1018
1019 /* If tg->bps = -1, then BW is unlimited */
1020 if (bps_limit == U64_MAX && iops_limit == UINT_MAX) {
1021 if (wait)
1022 *wait = 0;
1023 return true;
1024 }
1025
1026 /*
1027 * If previous slice expired, start a new one otherwise renew/extend
1028 * existing slice to make sure it is at least throtl_slice interval
1029 * long since now. New slice is started only for empty throttle group.
1030 * If there is queued bio, that means there should be an active
1031 * slice and it should be extended instead.
1032 */
1033 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
1034 throtl_start_new_slice(tg, rw);
1035 else {
1036 if (time_before(tg->slice_end[rw],
1037 jiffies + tg->td->throtl_slice))
1038 throtl_extend_slice(tg, rw,
1039 jiffies + tg->td->throtl_slice);
1040 }
1041
1042 if (iops_limit != UINT_MAX)
1043 tg->io_disp[rw] += atomic_xchg(&tg->io_split_cnt[rw], 0);
1044
1045 if (tg_with_in_bps_limit(tg, bio, bps_limit, &bps_wait) &&
1046 tg_with_in_iops_limit(tg, bio, iops_limit, &iops_wait)) {
1047 if (wait)
1048 *wait = 0;
1049 return true;
1050 }
1051
1052 max_wait = max(bps_wait, iops_wait);
1053
1054 if (wait)
1055 *wait = max_wait;
1056
1057 if (time_before(tg->slice_end[rw], jiffies + max_wait))
1058 throtl_extend_slice(tg, rw, jiffies + max_wait);
1059
1060 return false;
1061}
1062
1063static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1064{
1065 bool rw = bio_data_dir(bio);
1066 unsigned int bio_size = throtl_bio_data_size(bio);
1067
1068 /* Charge the bio to the group */
1069 tg->bytes_disp[rw] += bio_size;
1070 tg->io_disp[rw]++;
1071 tg->last_bytes_disp[rw] += bio_size;
1072 tg->last_io_disp[rw]++;
1073
1074 /*
1075 * BIO_THROTTLED is used to prevent the same bio to be throttled
1076 * more than once as a throttled bio will go through blk-throtl the
1077 * second time when it eventually gets issued. Set it when a bio
1078 * is being charged to a tg.
1079 */
1080 if (!bio_flagged(bio, BIO_THROTTLED))
1081 bio_set_flag(bio, BIO_THROTTLED);
1082}
1083
1084/**
1085 * throtl_add_bio_tg - add a bio to the specified throtl_grp
1086 * @bio: bio to add
1087 * @qn: qnode to use
1088 * @tg: the target throtl_grp
1089 *
1090 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
1091 * tg->qnode_on_self[] is used.
1092 */
1093static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1094 struct throtl_grp *tg)
1095{
1096 struct throtl_service_queue *sq = &tg->service_queue;
1097 bool rw = bio_data_dir(bio);
1098
1099 if (!qn)
1100 qn = &tg->qnode_on_self[rw];
1101
1102 /*
1103 * If @tg doesn't currently have any bios queued in the same
1104 * direction, queueing @bio can change when @tg should be
1105 * dispatched. Mark that @tg was empty. This is automatically
1106 * cleared on the next tg_update_disptime().
1107 */
1108 if (!sq->nr_queued[rw])
1109 tg->flags |= THROTL_TG_WAS_EMPTY;
1110
1111 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1112
1113 sq->nr_queued[rw]++;
1114 throtl_enqueue_tg(tg);
1115}
1116
1117static void tg_update_disptime(struct throtl_grp *tg)
1118{
1119 struct throtl_service_queue *sq = &tg->service_queue;
1120 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1121 struct bio *bio;
1122
1123 bio = throtl_peek_queued(&sq->queued[READ]);
1124 if (bio)
1125 tg_may_dispatch(tg, bio, &read_wait);
1126
1127 bio = throtl_peek_queued(&sq->queued[WRITE]);
1128 if (bio)
1129 tg_may_dispatch(tg, bio, &write_wait);
1130
1131 min_wait = min(read_wait, write_wait);
1132 disptime = jiffies + min_wait;
1133
1134 /* Update dispatch time */
1135 throtl_dequeue_tg(tg);
1136 tg->disptime = disptime;
1137 throtl_enqueue_tg(tg);
1138
1139 /* see throtl_add_bio_tg() */
1140 tg->flags &= ~THROTL_TG_WAS_EMPTY;
1141}
1142
1143static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1144 struct throtl_grp *parent_tg, bool rw)
1145{
1146 if (throtl_slice_used(parent_tg, rw)) {
1147 throtl_start_new_slice_with_credit(parent_tg, rw,
1148 child_tg->slice_start[rw]);
1149 }
1150
1151}
1152
1153static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1154{
1155 struct throtl_service_queue *sq = &tg->service_queue;
1156 struct throtl_service_queue *parent_sq = sq->parent_sq;
1157 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1158 struct throtl_grp *tg_to_put = NULL;
1159 struct bio *bio;
1160
1161 /*
1162 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1163 * from @tg may put its reference and @parent_sq might end up
1164 * getting released prematurely. Remember the tg to put and put it
1165 * after @bio is transferred to @parent_sq.
1166 */
1167 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1168 sq->nr_queued[rw]--;
1169
1170 throtl_charge_bio(tg, bio);
1171
1172 /*
1173 * If our parent is another tg, we just need to transfer @bio to
1174 * the parent using throtl_add_bio_tg(). If our parent is
1175 * @td->service_queue, @bio is ready to be issued. Put it on its
1176 * bio_lists[] and decrease total number queued. The caller is
1177 * responsible for issuing these bios.
1178 */
1179 if (parent_tg) {
1180 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1181 start_parent_slice_with_credit(tg, parent_tg, rw);
1182 } else {
1183 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1184 &parent_sq->queued[rw]);
1185 BUG_ON(tg->td->nr_queued[rw] <= 0);
1186 tg->td->nr_queued[rw]--;
1187 }
1188
1189 throtl_trim_slice(tg, rw);
1190
1191 if (tg_to_put)
1192 blkg_put(tg_to_blkg(tg_to_put));
1193}
1194
1195static int throtl_dispatch_tg(struct throtl_grp *tg)
1196{
1197 struct throtl_service_queue *sq = &tg->service_queue;
1198 unsigned int nr_reads = 0, nr_writes = 0;
1199 unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1200 unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1201 struct bio *bio;
1202
1203 /* Try to dispatch 75% READS and 25% WRITES */
1204
1205 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1206 tg_may_dispatch(tg, bio, NULL)) {
1207
1208 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1209 nr_reads++;
1210
1211 if (nr_reads >= max_nr_reads)
1212 break;
1213 }
1214
1215 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1216 tg_may_dispatch(tg, bio, NULL)) {
1217
1218 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1219 nr_writes++;
1220
1221 if (nr_writes >= max_nr_writes)
1222 break;
1223 }
1224
1225 return nr_reads + nr_writes;
1226}
1227
1228static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1229{
1230 unsigned int nr_disp = 0;
1231
1232 while (1) {
1233 struct throtl_grp *tg;
1234 struct throtl_service_queue *sq;
1235
1236 if (!parent_sq->nr_pending)
1237 break;
1238
1239 tg = throtl_rb_first(parent_sq);
1240 if (!tg)
1241 break;
1242
1243 if (time_before(jiffies, tg->disptime))
1244 break;
1245
1246 throtl_dequeue_tg(tg);
1247
1248 nr_disp += throtl_dispatch_tg(tg);
1249
1250 sq = &tg->service_queue;
1251 if (sq->nr_queued[0] || sq->nr_queued[1])
1252 tg_update_disptime(tg);
1253
1254 if (nr_disp >= THROTL_QUANTUM)
1255 break;
1256 }
1257
1258 return nr_disp;
1259}
1260
1261static bool throtl_can_upgrade(struct throtl_data *td,
1262 struct throtl_grp *this_tg);
1263/**
1264 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1265 * @t: the pending_timer member of the throtl_service_queue being serviced
1266 *
1267 * This timer is armed when a child throtl_grp with active bio's become
1268 * pending and queued on the service_queue's pending_tree and expires when
1269 * the first child throtl_grp should be dispatched. This function
1270 * dispatches bio's from the children throtl_grps to the parent
1271 * service_queue.
1272 *
1273 * If the parent's parent is another throtl_grp, dispatching is propagated
1274 * by either arming its pending_timer or repeating dispatch directly. If
1275 * the top-level service_tree is reached, throtl_data->dispatch_work is
1276 * kicked so that the ready bio's are issued.
1277 */
1278static void throtl_pending_timer_fn(struct timer_list *t)
1279{
1280 struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1281 struct throtl_grp *tg = sq_to_tg(sq);
1282 struct throtl_data *td = sq_to_td(sq);
1283 struct request_queue *q = td->queue;
1284 struct throtl_service_queue *parent_sq;
1285 bool dispatched;
1286 int ret;
1287
1288 spin_lock_irq(&q->queue_lock);
1289 if (throtl_can_upgrade(td, NULL))
1290 throtl_upgrade_state(td);
1291
1292again:
1293 parent_sq = sq->parent_sq;
1294 dispatched = false;
1295
1296 while (true) {
1297 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1298 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1299 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1300
1301 ret = throtl_select_dispatch(sq);
1302 if (ret) {
1303 throtl_log(sq, "bios disp=%u", ret);
1304 dispatched = true;
1305 }
1306
1307 if (throtl_schedule_next_dispatch(sq, false))
1308 break;
1309
1310 /* this dispatch windows is still open, relax and repeat */
1311 spin_unlock_irq(&q->queue_lock);
1312 cpu_relax();
1313 spin_lock_irq(&q->queue_lock);
1314 }
1315
1316 if (!dispatched)
1317 goto out_unlock;
1318
1319 if (parent_sq) {
1320 /* @parent_sq is another throl_grp, propagate dispatch */
1321 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1322 tg_update_disptime(tg);
1323 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1324 /* window is already open, repeat dispatching */
1325 sq = parent_sq;
1326 tg = sq_to_tg(sq);
1327 goto again;
1328 }
1329 }
1330 } else {
1331 /* reached the top-level, queue issuing */
1332 queue_work(kthrotld_workqueue, &td->dispatch_work);
1333 }
1334out_unlock:
1335 spin_unlock_irq(&q->queue_lock);
1336}
1337
1338/**
1339 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1340 * @work: work item being executed
1341 *
1342 * This function is queued for execution when bios reach the bio_lists[]
1343 * of throtl_data->service_queue. Those bios are ready and issued by this
1344 * function.
1345 */
1346static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1347{
1348 struct throtl_data *td = container_of(work, struct throtl_data,
1349 dispatch_work);
1350 struct throtl_service_queue *td_sq = &td->service_queue;
1351 struct request_queue *q = td->queue;
1352 struct bio_list bio_list_on_stack;
1353 struct bio *bio;
1354 struct blk_plug plug;
1355 int rw;
1356
1357 bio_list_init(&bio_list_on_stack);
1358
1359 spin_lock_irq(&q->queue_lock);
1360 for (rw = READ; rw <= WRITE; rw++)
1361 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1362 bio_list_add(&bio_list_on_stack, bio);
1363 spin_unlock_irq(&q->queue_lock);
1364
1365 if (!bio_list_empty(&bio_list_on_stack)) {
1366 blk_start_plug(&plug);
1367 while ((bio = bio_list_pop(&bio_list_on_stack)))
1368 submit_bio_noacct(bio);
1369 blk_finish_plug(&plug);
1370 }
1371}
1372
1373static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1374 int off)
1375{
1376 struct throtl_grp *tg = pd_to_tg(pd);
1377 u64 v = *(u64 *)((void *)tg + off);
1378
1379 if (v == U64_MAX)
1380 return 0;
1381 return __blkg_prfill_u64(sf, pd, v);
1382}
1383
1384static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1385 int off)
1386{
1387 struct throtl_grp *tg = pd_to_tg(pd);
1388 unsigned int v = *(unsigned int *)((void *)tg + off);
1389
1390 if (v == UINT_MAX)
1391 return 0;
1392 return __blkg_prfill_u64(sf, pd, v);
1393}
1394
1395static int tg_print_conf_u64(struct seq_file *sf, void *v)
1396{
1397 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1398 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1399 return 0;
1400}
1401
1402static int tg_print_conf_uint(struct seq_file *sf, void *v)
1403{
1404 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1405 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1406 return 0;
1407}
1408
1409static void tg_conf_updated(struct throtl_grp *tg, bool global)
1410{
1411 struct throtl_service_queue *sq = &tg->service_queue;
1412 struct cgroup_subsys_state *pos_css;
1413 struct blkcg_gq *blkg;
1414
1415 throtl_log(&tg->service_queue,
1416 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1417 tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1418 tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1419
1420 /*
1421 * Update has_rules[] flags for the updated tg's subtree. A tg is
1422 * considered to have rules if either the tg itself or any of its
1423 * ancestors has rules. This identifies groups without any
1424 * restrictions in the whole hierarchy and allows them to bypass
1425 * blk-throttle.
1426 */
1427 blkg_for_each_descendant_pre(blkg, pos_css,
1428 global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1429 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1430 struct throtl_grp *parent_tg;
1431
1432 tg_update_has_rules(this_tg);
1433 /* ignore root/second level */
1434 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1435 !blkg->parent->parent)
1436 continue;
1437 parent_tg = blkg_to_tg(blkg->parent);
1438 /*
1439 * make sure all children has lower idle time threshold and
1440 * higher latency target
1441 */
1442 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1443 parent_tg->idletime_threshold);
1444 this_tg->latency_target = max(this_tg->latency_target,
1445 parent_tg->latency_target);
1446 }
1447
1448 /*
1449 * We're already holding queue_lock and know @tg is valid. Let's
1450 * apply the new config directly.
1451 *
1452 * Restart the slices for both READ and WRITES. It might happen
1453 * that a group's limit are dropped suddenly and we don't want to
1454 * account recently dispatched IO with new low rate.
1455 */
1456 throtl_start_new_slice(tg, READ);
1457 throtl_start_new_slice(tg, WRITE);
1458
1459 if (tg->flags & THROTL_TG_PENDING) {
1460 tg_update_disptime(tg);
1461 throtl_schedule_next_dispatch(sq->parent_sq, true);
1462 }
1463}
1464
1465static ssize_t tg_set_conf(struct kernfs_open_file *of,
1466 char *buf, size_t nbytes, loff_t off, bool is_u64)
1467{
1468 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1469 struct blkg_conf_ctx ctx;
1470 struct throtl_grp *tg;
1471 int ret;
1472 u64 v;
1473
1474 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1475 if (ret)
1476 return ret;
1477
1478 ret = -EINVAL;
1479 if (sscanf(ctx.body, "%llu", &v) != 1)
1480 goto out_finish;
1481 if (!v)
1482 v = U64_MAX;
1483
1484 tg = blkg_to_tg(ctx.blkg);
1485
1486 if (is_u64)
1487 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1488 else
1489 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1490
1491 tg_conf_updated(tg, false);
1492 ret = 0;
1493out_finish:
1494 blkg_conf_finish(&ctx);
1495 return ret ?: nbytes;
1496}
1497
1498static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1499 char *buf, size_t nbytes, loff_t off)
1500{
1501 return tg_set_conf(of, buf, nbytes, off, true);
1502}
1503
1504static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1505 char *buf, size_t nbytes, loff_t off)
1506{
1507 return tg_set_conf(of, buf, nbytes, off, false);
1508}
1509
1510static int tg_print_rwstat(struct seq_file *sf, void *v)
1511{
1512 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1513 blkg_prfill_rwstat, &blkcg_policy_throtl,
1514 seq_cft(sf)->private, true);
1515 return 0;
1516}
1517
1518static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1519 struct blkg_policy_data *pd, int off)
1520{
1521 struct blkg_rwstat_sample sum;
1522
1523 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1524 &sum);
1525 return __blkg_prfill_rwstat(sf, pd, &sum);
1526}
1527
1528static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1529{
1530 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1531 tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1532 seq_cft(sf)->private, true);
1533 return 0;
1534}
1535
1536static struct cftype throtl_legacy_files[] = {
1537 {
1538 .name = "throttle.read_bps_device",
1539 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1540 .seq_show = tg_print_conf_u64,
1541 .write = tg_set_conf_u64,
1542 },
1543 {
1544 .name = "throttle.write_bps_device",
1545 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1546 .seq_show = tg_print_conf_u64,
1547 .write = tg_set_conf_u64,
1548 },
1549 {
1550 .name = "throttle.read_iops_device",
1551 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1552 .seq_show = tg_print_conf_uint,
1553 .write = tg_set_conf_uint,
1554 },
1555 {
1556 .name = "throttle.write_iops_device",
1557 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1558 .seq_show = tg_print_conf_uint,
1559 .write = tg_set_conf_uint,
1560 },
1561 {
1562 .name = "throttle.io_service_bytes",
1563 .private = offsetof(struct throtl_grp, stat_bytes),
1564 .seq_show = tg_print_rwstat,
1565 },
1566 {
1567 .name = "throttle.io_service_bytes_recursive",
1568 .private = offsetof(struct throtl_grp, stat_bytes),
1569 .seq_show = tg_print_rwstat_recursive,
1570 },
1571 {
1572 .name = "throttle.io_serviced",
1573 .private = offsetof(struct throtl_grp, stat_ios),
1574 .seq_show = tg_print_rwstat,
1575 },
1576 {
1577 .name = "throttle.io_serviced_recursive",
1578 .private = offsetof(struct throtl_grp, stat_ios),
1579 .seq_show = tg_print_rwstat_recursive,
1580 },
1581 { } /* terminate */
1582};
1583
1584static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1585 int off)
1586{
1587 struct throtl_grp *tg = pd_to_tg(pd);
1588 const char *dname = blkg_dev_name(pd->blkg);
1589 char bufs[4][21] = { "max", "max", "max", "max" };
1590 u64 bps_dft;
1591 unsigned int iops_dft;
1592 char idle_time[26] = "";
1593 char latency_time[26] = "";
1594
1595 if (!dname)
1596 return 0;
1597
1598 if (off == LIMIT_LOW) {
1599 bps_dft = 0;
1600 iops_dft = 0;
1601 } else {
1602 bps_dft = U64_MAX;
1603 iops_dft = UINT_MAX;
1604 }
1605
1606 if (tg->bps_conf[READ][off] == bps_dft &&
1607 tg->bps_conf[WRITE][off] == bps_dft &&
1608 tg->iops_conf[READ][off] == iops_dft &&
1609 tg->iops_conf[WRITE][off] == iops_dft &&
1610 (off != LIMIT_LOW ||
1611 (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1612 tg->latency_target_conf == DFL_LATENCY_TARGET)))
1613 return 0;
1614
1615 if (tg->bps_conf[READ][off] != U64_MAX)
1616 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1617 tg->bps_conf[READ][off]);
1618 if (tg->bps_conf[WRITE][off] != U64_MAX)
1619 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1620 tg->bps_conf[WRITE][off]);
1621 if (tg->iops_conf[READ][off] != UINT_MAX)
1622 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1623 tg->iops_conf[READ][off]);
1624 if (tg->iops_conf[WRITE][off] != UINT_MAX)
1625 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1626 tg->iops_conf[WRITE][off]);
1627 if (off == LIMIT_LOW) {
1628 if (tg->idletime_threshold_conf == ULONG_MAX)
1629 strcpy(idle_time, " idle=max");
1630 else
1631 snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1632 tg->idletime_threshold_conf);
1633
1634 if (tg->latency_target_conf == ULONG_MAX)
1635 strcpy(latency_time, " latency=max");
1636 else
1637 snprintf(latency_time, sizeof(latency_time),
1638 " latency=%lu", tg->latency_target_conf);
1639 }
1640
1641 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1642 dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1643 latency_time);
1644 return 0;
1645}
1646
1647static int tg_print_limit(struct seq_file *sf, void *v)
1648{
1649 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1650 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1651 return 0;
1652}
1653
1654static ssize_t tg_set_limit(struct kernfs_open_file *of,
1655 char *buf, size_t nbytes, loff_t off)
1656{
1657 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1658 struct blkg_conf_ctx ctx;
1659 struct throtl_grp *tg;
1660 u64 v[4];
1661 unsigned long idle_time;
1662 unsigned long latency_time;
1663 int ret;
1664 int index = of_cft(of)->private;
1665
1666 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1667 if (ret)
1668 return ret;
1669
1670 tg = blkg_to_tg(ctx.blkg);
1671
1672 v[0] = tg->bps_conf[READ][index];
1673 v[1] = tg->bps_conf[WRITE][index];
1674 v[2] = tg->iops_conf[READ][index];
1675 v[3] = tg->iops_conf[WRITE][index];
1676
1677 idle_time = tg->idletime_threshold_conf;
1678 latency_time = tg->latency_target_conf;
1679 while (true) {
1680 char tok[27]; /* wiops=18446744073709551616 */
1681 char *p;
1682 u64 val = U64_MAX;
1683 int len;
1684
1685 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1686 break;
1687 if (tok[0] == '\0')
1688 break;
1689 ctx.body += len;
1690
1691 ret = -EINVAL;
1692 p = tok;
1693 strsep(&p, "=");
1694 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1695 goto out_finish;
1696
1697 ret = -ERANGE;
1698 if (!val)
1699 goto out_finish;
1700
1701 ret = -EINVAL;
1702 if (!strcmp(tok, "rbps") && val > 1)
1703 v[0] = val;
1704 else if (!strcmp(tok, "wbps") && val > 1)
1705 v[1] = val;
1706 else if (!strcmp(tok, "riops") && val > 1)
1707 v[2] = min_t(u64, val, UINT_MAX);
1708 else if (!strcmp(tok, "wiops") && val > 1)
1709 v[3] = min_t(u64, val, UINT_MAX);
1710 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1711 idle_time = val;
1712 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1713 latency_time = val;
1714 else
1715 goto out_finish;
1716 }
1717
1718 tg->bps_conf[READ][index] = v[0];
1719 tg->bps_conf[WRITE][index] = v[1];
1720 tg->iops_conf[READ][index] = v[2];
1721 tg->iops_conf[WRITE][index] = v[3];
1722
1723 if (index == LIMIT_MAX) {
1724 tg->bps[READ][index] = v[0];
1725 tg->bps[WRITE][index] = v[1];
1726 tg->iops[READ][index] = v[2];
1727 tg->iops[WRITE][index] = v[3];
1728 }
1729 tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1730 tg->bps_conf[READ][LIMIT_MAX]);
1731 tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1732 tg->bps_conf[WRITE][LIMIT_MAX]);
1733 tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1734 tg->iops_conf[READ][LIMIT_MAX]);
1735 tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1736 tg->iops_conf[WRITE][LIMIT_MAX]);
1737 tg->idletime_threshold_conf = idle_time;
1738 tg->latency_target_conf = latency_time;
1739
1740 /* force user to configure all settings for low limit */
1741 if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1742 tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1743 tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1744 tg->latency_target_conf == DFL_LATENCY_TARGET) {
1745 tg->bps[READ][LIMIT_LOW] = 0;
1746 tg->bps[WRITE][LIMIT_LOW] = 0;
1747 tg->iops[READ][LIMIT_LOW] = 0;
1748 tg->iops[WRITE][LIMIT_LOW] = 0;
1749 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1750 tg->latency_target = DFL_LATENCY_TARGET;
1751 } else if (index == LIMIT_LOW) {
1752 tg->idletime_threshold = tg->idletime_threshold_conf;
1753 tg->latency_target = tg->latency_target_conf;
1754 }
1755
1756 blk_throtl_update_limit_valid(tg->td);
1757 if (tg->td->limit_valid[LIMIT_LOW]) {
1758 if (index == LIMIT_LOW)
1759 tg->td->limit_index = LIMIT_LOW;
1760 } else
1761 tg->td->limit_index = LIMIT_MAX;
1762 tg_conf_updated(tg, index == LIMIT_LOW &&
1763 tg->td->limit_valid[LIMIT_LOW]);
1764 ret = 0;
1765out_finish:
1766 blkg_conf_finish(&ctx);
1767 return ret ?: nbytes;
1768}
1769
1770static struct cftype throtl_files[] = {
1771#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1772 {
1773 .name = "low",
1774 .flags = CFTYPE_NOT_ON_ROOT,
1775 .seq_show = tg_print_limit,
1776 .write = tg_set_limit,
1777 .private = LIMIT_LOW,
1778 },
1779#endif
1780 {
1781 .name = "max",
1782 .flags = CFTYPE_NOT_ON_ROOT,
1783 .seq_show = tg_print_limit,
1784 .write = tg_set_limit,
1785 .private = LIMIT_MAX,
1786 },
1787 { } /* terminate */
1788};
1789
1790static void throtl_shutdown_wq(struct request_queue *q)
1791{
1792 struct throtl_data *td = q->td;
1793
1794 cancel_work_sync(&td->dispatch_work);
1795}
1796
1797static struct blkcg_policy blkcg_policy_throtl = {
1798 .dfl_cftypes = throtl_files,
1799 .legacy_cftypes = throtl_legacy_files,
1800
1801 .pd_alloc_fn = throtl_pd_alloc,
1802 .pd_init_fn = throtl_pd_init,
1803 .pd_online_fn = throtl_pd_online,
1804 .pd_offline_fn = throtl_pd_offline,
1805 .pd_free_fn = throtl_pd_free,
1806};
1807
1808static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1809{
1810 unsigned long rtime = jiffies, wtime = jiffies;
1811
1812 if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1813 rtime = tg->last_low_overflow_time[READ];
1814 if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1815 wtime = tg->last_low_overflow_time[WRITE];
1816 return min(rtime, wtime);
1817}
1818
1819/* tg should not be an intermediate node */
1820static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1821{
1822 struct throtl_service_queue *parent_sq;
1823 struct throtl_grp *parent = tg;
1824 unsigned long ret = __tg_last_low_overflow_time(tg);
1825
1826 while (true) {
1827 parent_sq = parent->service_queue.parent_sq;
1828 parent = sq_to_tg(parent_sq);
1829 if (!parent)
1830 break;
1831
1832 /*
1833 * The parent doesn't have low limit, it always reaches low
1834 * limit. Its overflow time is useless for children
1835 */
1836 if (!parent->bps[READ][LIMIT_LOW] &&
1837 !parent->iops[READ][LIMIT_LOW] &&
1838 !parent->bps[WRITE][LIMIT_LOW] &&
1839 !parent->iops[WRITE][LIMIT_LOW])
1840 continue;
1841 if (time_after(__tg_last_low_overflow_time(parent), ret))
1842 ret = __tg_last_low_overflow_time(parent);
1843 }
1844 return ret;
1845}
1846
1847static bool throtl_tg_is_idle(struct throtl_grp *tg)
1848{
1849 /*
1850 * cgroup is idle if:
1851 * - single idle is too long, longer than a fixed value (in case user
1852 * configure a too big threshold) or 4 times of idletime threshold
1853 * - average think time is more than threshold
1854 * - IO latency is largely below threshold
1855 */
1856 unsigned long time;
1857 bool ret;
1858
1859 time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1860 ret = tg->latency_target == DFL_LATENCY_TARGET ||
1861 tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1862 (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1863 tg->avg_idletime > tg->idletime_threshold ||
1864 (tg->latency_target && tg->bio_cnt &&
1865 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1866 throtl_log(&tg->service_queue,
1867 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1868 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1869 tg->bio_cnt, ret, tg->td->scale);
1870 return ret;
1871}
1872
1873static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1874{
1875 struct throtl_service_queue *sq = &tg->service_queue;
1876 bool read_limit, write_limit;
1877
1878 /*
1879 * if cgroup reaches low limit (if low limit is 0, the cgroup always
1880 * reaches), it's ok to upgrade to next limit
1881 */
1882 read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1883 write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1884 if (!read_limit && !write_limit)
1885 return true;
1886 if (read_limit && sq->nr_queued[READ] &&
1887 (!write_limit || sq->nr_queued[WRITE]))
1888 return true;
1889 if (write_limit && sq->nr_queued[WRITE] &&
1890 (!read_limit || sq->nr_queued[READ]))
1891 return true;
1892
1893 if (time_after_eq(jiffies,
1894 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1895 throtl_tg_is_idle(tg))
1896 return true;
1897 return false;
1898}
1899
1900static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1901{
1902 while (true) {
1903 if (throtl_tg_can_upgrade(tg))
1904 return true;
1905 tg = sq_to_tg(tg->service_queue.parent_sq);
1906 if (!tg || !tg_to_blkg(tg)->parent)
1907 return false;
1908 }
1909 return false;
1910}
1911
1912static bool throtl_can_upgrade(struct throtl_data *td,
1913 struct throtl_grp *this_tg)
1914{
1915 struct cgroup_subsys_state *pos_css;
1916 struct blkcg_gq *blkg;
1917
1918 if (td->limit_index != LIMIT_LOW)
1919 return false;
1920
1921 if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1922 return false;
1923
1924 rcu_read_lock();
1925 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1926 struct throtl_grp *tg = blkg_to_tg(blkg);
1927
1928 if (tg == this_tg)
1929 continue;
1930 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1931 continue;
1932 if (!throtl_hierarchy_can_upgrade(tg)) {
1933 rcu_read_unlock();
1934 return false;
1935 }
1936 }
1937 rcu_read_unlock();
1938 return true;
1939}
1940
1941static void throtl_upgrade_check(struct throtl_grp *tg)
1942{
1943 unsigned long now = jiffies;
1944
1945 if (tg->td->limit_index != LIMIT_LOW)
1946 return;
1947
1948 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1949 return;
1950
1951 tg->last_check_time = now;
1952
1953 if (!time_after_eq(now,
1954 __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1955 return;
1956
1957 if (throtl_can_upgrade(tg->td, NULL))
1958 throtl_upgrade_state(tg->td);
1959}
1960
1961static void throtl_upgrade_state(struct throtl_data *td)
1962{
1963 struct cgroup_subsys_state *pos_css;
1964 struct blkcg_gq *blkg;
1965
1966 throtl_log(&td->service_queue, "upgrade to max");
1967 td->limit_index = LIMIT_MAX;
1968 td->low_upgrade_time = jiffies;
1969 td->scale = 0;
1970 rcu_read_lock();
1971 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1972 struct throtl_grp *tg = blkg_to_tg(blkg);
1973 struct throtl_service_queue *sq = &tg->service_queue;
1974
1975 tg->disptime = jiffies - 1;
1976 throtl_select_dispatch(sq);
1977 throtl_schedule_next_dispatch(sq, true);
1978 }
1979 rcu_read_unlock();
1980 throtl_select_dispatch(&td->service_queue);
1981 throtl_schedule_next_dispatch(&td->service_queue, true);
1982 queue_work(kthrotld_workqueue, &td->dispatch_work);
1983}
1984
1985static void throtl_downgrade_state(struct throtl_data *td)
1986{
1987 td->scale /= 2;
1988
1989 throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1990 if (td->scale) {
1991 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1992 return;
1993 }
1994
1995 td->limit_index = LIMIT_LOW;
1996 td->low_downgrade_time = jiffies;
1997}
1998
1999static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
2000{
2001 struct throtl_data *td = tg->td;
2002 unsigned long now = jiffies;
2003
2004 /*
2005 * If cgroup is below low limit, consider downgrade and throttle other
2006 * cgroups
2007 */
2008 if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
2009 time_after_eq(now, tg_last_low_overflow_time(tg) +
2010 td->throtl_slice) &&
2011 (!throtl_tg_is_idle(tg) ||
2012 !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
2013 return true;
2014 return false;
2015}
2016
2017static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
2018{
2019 while (true) {
2020 if (!throtl_tg_can_downgrade(tg))
2021 return false;
2022 tg = sq_to_tg(tg->service_queue.parent_sq);
2023 if (!tg || !tg_to_blkg(tg)->parent)
2024 break;
2025 }
2026 return true;
2027}
2028
2029static void throtl_downgrade_check(struct throtl_grp *tg)
2030{
2031 uint64_t bps;
2032 unsigned int iops;
2033 unsigned long elapsed_time;
2034 unsigned long now = jiffies;
2035
2036 if (tg->td->limit_index != LIMIT_MAX ||
2037 !tg->td->limit_valid[LIMIT_LOW])
2038 return;
2039 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
2040 return;
2041 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
2042 return;
2043
2044 elapsed_time = now - tg->last_check_time;
2045 tg->last_check_time = now;
2046
2047 if (time_before(now, tg_last_low_overflow_time(tg) +
2048 tg->td->throtl_slice))
2049 return;
2050
2051 if (tg->bps[READ][LIMIT_LOW]) {
2052 bps = tg->last_bytes_disp[READ] * HZ;
2053 do_div(bps, elapsed_time);
2054 if (bps >= tg->bps[READ][LIMIT_LOW])
2055 tg->last_low_overflow_time[READ] = now;
2056 }
2057
2058 if (tg->bps[WRITE][LIMIT_LOW]) {
2059 bps = tg->last_bytes_disp[WRITE] * HZ;
2060 do_div(bps, elapsed_time);
2061 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2062 tg->last_low_overflow_time[WRITE] = now;
2063 }
2064
2065 if (tg->iops[READ][LIMIT_LOW]) {
2066 tg->last_io_disp[READ] += atomic_xchg(&tg->last_io_split_cnt[READ], 0);
2067 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2068 if (iops >= tg->iops[READ][LIMIT_LOW])
2069 tg->last_low_overflow_time[READ] = now;
2070 }
2071
2072 if (tg->iops[WRITE][LIMIT_LOW]) {
2073 tg->last_io_disp[WRITE] += atomic_xchg(&tg->last_io_split_cnt[WRITE], 0);
2074 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2075 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2076 tg->last_low_overflow_time[WRITE] = now;
2077 }
2078
2079 /*
2080 * If cgroup is below low limit, consider downgrade and throttle other
2081 * cgroups
2082 */
2083 if (throtl_hierarchy_can_downgrade(tg))
2084 throtl_downgrade_state(tg->td);
2085
2086 tg->last_bytes_disp[READ] = 0;
2087 tg->last_bytes_disp[WRITE] = 0;
2088 tg->last_io_disp[READ] = 0;
2089 tg->last_io_disp[WRITE] = 0;
2090}
2091
2092static void blk_throtl_update_idletime(struct throtl_grp *tg)
2093{
2094 unsigned long now;
2095 unsigned long last_finish_time = tg->last_finish_time;
2096
2097 if (last_finish_time == 0)
2098 return;
2099
2100 now = ktime_get_ns() >> 10;
2101 if (now <= last_finish_time ||
2102 last_finish_time == tg->checked_last_finish_time)
2103 return;
2104
2105 tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2106 tg->checked_last_finish_time = last_finish_time;
2107}
2108
2109#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2110static void throtl_update_latency_buckets(struct throtl_data *td)
2111{
2112 struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2113 int i, cpu, rw;
2114 unsigned long last_latency[2] = { 0 };
2115 unsigned long latency[2];
2116
2117 if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
2118 return;
2119 if (time_before(jiffies, td->last_calculate_time + HZ))
2120 return;
2121 td->last_calculate_time = jiffies;
2122
2123 memset(avg_latency, 0, sizeof(avg_latency));
2124 for (rw = READ; rw <= WRITE; rw++) {
2125 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2126 struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2127
2128 for_each_possible_cpu(cpu) {
2129 struct latency_bucket *bucket;
2130
2131 /* this isn't race free, but ok in practice */
2132 bucket = per_cpu_ptr(td->latency_buckets[rw],
2133 cpu);
2134 tmp->total_latency += bucket[i].total_latency;
2135 tmp->samples += bucket[i].samples;
2136 bucket[i].total_latency = 0;
2137 bucket[i].samples = 0;
2138 }
2139
2140 if (tmp->samples >= 32) {
2141 int samples = tmp->samples;
2142
2143 latency[rw] = tmp->total_latency;
2144
2145 tmp->total_latency = 0;
2146 tmp->samples = 0;
2147 latency[rw] /= samples;
2148 if (latency[rw] == 0)
2149 continue;
2150 avg_latency[rw][i].latency = latency[rw];
2151 }
2152 }
2153 }
2154
2155 for (rw = READ; rw <= WRITE; rw++) {
2156 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2157 if (!avg_latency[rw][i].latency) {
2158 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2159 td->avg_buckets[rw][i].latency =
2160 last_latency[rw];
2161 continue;
2162 }
2163
2164 if (!td->avg_buckets[rw][i].valid)
2165 latency[rw] = avg_latency[rw][i].latency;
2166 else
2167 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2168 avg_latency[rw][i].latency) >> 3;
2169
2170 td->avg_buckets[rw][i].latency = max(latency[rw],
2171 last_latency[rw]);
2172 td->avg_buckets[rw][i].valid = true;
2173 last_latency[rw] = td->avg_buckets[rw][i].latency;
2174 }
2175 }
2176
2177 for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2178 throtl_log(&td->service_queue,
2179 "Latency bucket %d: read latency=%ld, read valid=%d, "
2180 "write latency=%ld, write valid=%d", i,
2181 td->avg_buckets[READ][i].latency,
2182 td->avg_buckets[READ][i].valid,
2183 td->avg_buckets[WRITE][i].latency,
2184 td->avg_buckets[WRITE][i].valid);
2185}
2186#else
2187static inline void throtl_update_latency_buckets(struct throtl_data *td)
2188{
2189}
2190#endif
2191
2192void blk_throtl_charge_bio_split(struct bio *bio)
2193{
2194 struct blkcg_gq *blkg = bio->bi_blkg;
2195 struct throtl_grp *parent = blkg_to_tg(blkg);
2196 struct throtl_service_queue *parent_sq;
2197 bool rw = bio_data_dir(bio);
2198
2199 do {
2200 if (!parent->has_rules[rw])
2201 break;
2202
2203 atomic_inc(&parent->io_split_cnt[rw]);
2204 atomic_inc(&parent->last_io_split_cnt[rw]);
2205
2206 parent_sq = parent->service_queue.parent_sq;
2207 parent = sq_to_tg(parent_sq);
2208 } while (parent);
2209}
2210
2211bool blk_throtl_bio(struct bio *bio)
2212{
2213 struct request_queue *q = bio->bi_bdev->bd_disk->queue;
2214 struct blkcg_gq *blkg = bio->bi_blkg;
2215 struct throtl_qnode *qn = NULL;
2216 struct throtl_grp *tg = blkg_to_tg(blkg);
2217 struct throtl_service_queue *sq;
2218 bool rw = bio_data_dir(bio);
2219 bool throttled = false;
2220 struct throtl_data *td = tg->td;
2221
2222 rcu_read_lock();
2223
2224 /* see throtl_charge_bio() */
2225 if (bio_flagged(bio, BIO_THROTTLED))
2226 goto out;
2227
2228 if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
2229 blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
2230 bio->bi_iter.bi_size);
2231 blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
2232 }
2233
2234 if (!tg->has_rules[rw])
2235 goto out;
2236
2237 spin_lock_irq(&q->queue_lock);
2238
2239 throtl_update_latency_buckets(td);
2240
2241 blk_throtl_update_idletime(tg);
2242
2243 sq = &tg->service_queue;
2244
2245again:
2246 while (true) {
2247 if (tg->last_low_overflow_time[rw] == 0)
2248 tg->last_low_overflow_time[rw] = jiffies;
2249 throtl_downgrade_check(tg);
2250 throtl_upgrade_check(tg);
2251 /* throtl is FIFO - if bios are already queued, should queue */
2252 if (sq->nr_queued[rw])
2253 break;
2254
2255 /* if above limits, break to queue */
2256 if (!tg_may_dispatch(tg, bio, NULL)) {
2257 tg->last_low_overflow_time[rw] = jiffies;
2258 if (throtl_can_upgrade(td, tg)) {
2259 throtl_upgrade_state(td);
2260 goto again;
2261 }
2262 break;
2263 }
2264
2265 /* within limits, let's charge and dispatch directly */
2266 throtl_charge_bio(tg, bio);
2267
2268 /*
2269 * We need to trim slice even when bios are not being queued
2270 * otherwise it might happen that a bio is not queued for
2271 * a long time and slice keeps on extending and trim is not
2272 * called for a long time. Now if limits are reduced suddenly
2273 * we take into account all the IO dispatched so far at new
2274 * low rate and * newly queued IO gets a really long dispatch
2275 * time.
2276 *
2277 * So keep on trimming slice even if bio is not queued.
2278 */
2279 throtl_trim_slice(tg, rw);
2280
2281 /*
2282 * @bio passed through this layer without being throttled.
2283 * Climb up the ladder. If we're already at the top, it
2284 * can be executed directly.
2285 */
2286 qn = &tg->qnode_on_parent[rw];
2287 sq = sq->parent_sq;
2288 tg = sq_to_tg(sq);
2289 if (!tg)
2290 goto out_unlock;
2291 }
2292
2293 /* out-of-limit, queue to @tg */
2294 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2295 rw == READ ? 'R' : 'W',
2296 tg->bytes_disp[rw], bio->bi_iter.bi_size,
2297 tg_bps_limit(tg, rw),
2298 tg->io_disp[rw], tg_iops_limit(tg, rw),
2299 sq->nr_queued[READ], sq->nr_queued[WRITE]);
2300
2301 tg->last_low_overflow_time[rw] = jiffies;
2302
2303 td->nr_queued[rw]++;
2304 throtl_add_bio_tg(bio, qn, tg);
2305 throttled = true;
2306
2307 /*
2308 * Update @tg's dispatch time and force schedule dispatch if @tg
2309 * was empty before @bio. The forced scheduling isn't likely to
2310 * cause undue delay as @bio is likely to be dispatched directly if
2311 * its @tg's disptime is not in the future.
2312 */
2313 if (tg->flags & THROTL_TG_WAS_EMPTY) {
2314 tg_update_disptime(tg);
2315 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2316 }
2317
2318out_unlock:
2319 spin_unlock_irq(&q->queue_lock);
2320out:
2321 bio_set_flag(bio, BIO_THROTTLED);
2322
2323#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2324 if (throttled || !td->track_bio_latency)
2325 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2326#endif
2327 rcu_read_unlock();
2328 return throttled;
2329}
2330
2331#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2332static void throtl_track_latency(struct throtl_data *td, sector_t size,
2333 int op, unsigned long time)
2334{
2335 struct latency_bucket *latency;
2336 int index;
2337
2338 if (!td || td->limit_index != LIMIT_LOW ||
2339 !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2340 !blk_queue_nonrot(td->queue))
2341 return;
2342
2343 index = request_bucket_index(size);
2344
2345 latency = get_cpu_ptr(td->latency_buckets[op]);
2346 latency[index].total_latency += time;
2347 latency[index].samples++;
2348 put_cpu_ptr(td->latency_buckets[op]);
2349}
2350
2351void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2352{
2353 struct request_queue *q = rq->q;
2354 struct throtl_data *td = q->td;
2355
2356 throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2357 time_ns >> 10);
2358}
2359
2360void blk_throtl_bio_endio(struct bio *bio)
2361{
2362 struct blkcg_gq *blkg;
2363 struct throtl_grp *tg;
2364 u64 finish_time_ns;
2365 unsigned long finish_time;
2366 unsigned long start_time;
2367 unsigned long lat;
2368 int rw = bio_data_dir(bio);
2369
2370 blkg = bio->bi_blkg;
2371 if (!blkg)
2372 return;
2373 tg = blkg_to_tg(blkg);
2374 if (!tg->td->limit_valid[LIMIT_LOW])
2375 return;
2376
2377 finish_time_ns = ktime_get_ns();
2378 tg->last_finish_time = finish_time_ns >> 10;
2379
2380 start_time = bio_issue_time(&bio->bi_issue) >> 10;
2381 finish_time = __bio_issue_time(finish_time_ns) >> 10;
2382 if (!start_time || finish_time <= start_time)
2383 return;
2384
2385 lat = finish_time - start_time;
2386 /* this is only for bio based driver */
2387 if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2388 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2389 bio_op(bio), lat);
2390
2391 if (tg->latency_target && lat >= tg->td->filtered_latency) {
2392 int bucket;
2393 unsigned int threshold;
2394
2395 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2396 threshold = tg->td->avg_buckets[rw][bucket].latency +
2397 tg->latency_target;
2398 if (lat > threshold)
2399 tg->bad_bio_cnt++;
2400 /*
2401 * Not race free, could get wrong count, which means cgroups
2402 * will be throttled
2403 */
2404 tg->bio_cnt++;
2405 }
2406
2407 if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2408 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2409 tg->bio_cnt /= 2;
2410 tg->bad_bio_cnt /= 2;
2411 }
2412}
2413#endif
2414
2415int blk_throtl_init(struct request_queue *q)
2416{
2417 struct throtl_data *td;
2418 int ret;
2419
2420 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2421 if (!td)
2422 return -ENOMEM;
2423 td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2424 LATENCY_BUCKET_SIZE, __alignof__(u64));
2425 if (!td->latency_buckets[READ]) {
2426 kfree(td);
2427 return -ENOMEM;
2428 }
2429 td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2430 LATENCY_BUCKET_SIZE, __alignof__(u64));
2431 if (!td->latency_buckets[WRITE]) {
2432 free_percpu(td->latency_buckets[READ]);
2433 kfree(td);
2434 return -ENOMEM;
2435 }
2436
2437 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2438 throtl_service_queue_init(&td->service_queue);
2439
2440 q->td = td;
2441 td->queue = q;
2442
2443 td->limit_valid[LIMIT_MAX] = true;
2444 td->limit_index = LIMIT_MAX;
2445 td->low_upgrade_time = jiffies;
2446 td->low_downgrade_time = jiffies;
2447
2448 /* activate policy */
2449 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2450 if (ret) {
2451 free_percpu(td->latency_buckets[READ]);
2452 free_percpu(td->latency_buckets[WRITE]);
2453 kfree(td);
2454 }
2455 return ret;
2456}
2457
2458void blk_throtl_exit(struct request_queue *q)
2459{
2460 BUG_ON(!q->td);
2461 del_timer_sync(&q->td->service_queue.pending_timer);
2462 throtl_shutdown_wq(q);
2463 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2464 free_percpu(q->td->latency_buckets[READ]);
2465 free_percpu(q->td->latency_buckets[WRITE]);
2466 kfree(q->td);
2467}
2468
2469void blk_throtl_register_queue(struct request_queue *q)
2470{
2471 struct throtl_data *td;
2472 int i;
2473
2474 td = q->td;
2475 BUG_ON(!td);
2476
2477 if (blk_queue_nonrot(q)) {
2478 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2479 td->filtered_latency = LATENCY_FILTERED_SSD;
2480 } else {
2481 td->throtl_slice = DFL_THROTL_SLICE_HD;
2482 td->filtered_latency = LATENCY_FILTERED_HD;
2483 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2484 td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2485 td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2486 }
2487 }
2488#ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2489 /* if no low limit, use previous default */
2490 td->throtl_slice = DFL_THROTL_SLICE_HD;
2491#endif
2492
2493 td->track_bio_latency = !queue_is_mq(q);
2494 if (!td->track_bio_latency)
2495 blk_stat_enable_accounting(q);
2496}
2497
2498#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2499ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2500{
2501 if (!q->td)
2502 return -EINVAL;
2503 return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2504}
2505
2506ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2507 const char *page, size_t count)
2508{
2509 unsigned long v;
2510 unsigned long t;
2511
2512 if (!q->td)
2513 return -EINVAL;
2514 if (kstrtoul(page, 10, &v))
2515 return -EINVAL;
2516 t = msecs_to_jiffies(v);
2517 if (t == 0 || t > MAX_THROTL_SLICE)
2518 return -EINVAL;
2519 q->td->throtl_slice = t;
2520 return count;
2521}
2522#endif
2523
2524static int __init throtl_init(void)
2525{
2526 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2527 if (!kthrotld_workqueue)
2528 panic("Failed to create kthrotld\n");
2529
2530 return blkcg_policy_register(&blkcg_policy_throtl);
2531}
2532
2533module_init(throtl_init);