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 "blk.h"
14#include "blk-cgroup-rwstat.h"
15#include "blk-stat.h"
16#include "blk-throttle.h"
17
18/* Max dispatch from a group in 1 round */
19#define THROTL_GRP_QUANTUM 8
20
21/* Total max dispatch from all groups in one round */
22#define THROTL_QUANTUM 32
23
24/* Throttling is performed over a slice and after that slice is renewed */
25#define DFL_THROTL_SLICE_HD (HZ / 10)
26#define DFL_THROTL_SLICE_SSD (HZ / 50)
27#define MAX_THROTL_SLICE (HZ)
28#define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
29#define MIN_THROTL_BPS (320 * 1024)
30#define MIN_THROTL_IOPS (10)
31#define DFL_LATENCY_TARGET (-1L)
32#define DFL_IDLE_THRESHOLD (0)
33#define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
34#define LATENCY_FILTERED_SSD (0)
35/*
36 * For HD, very small latency comes from sequential IO. Such IO is helpless to
37 * help determine if its IO is impacted by others, hence we ignore the IO
38 */
39#define LATENCY_FILTERED_HD (1000L) /* 1ms */
40
41/* A workqueue to queue throttle related work */
42static struct workqueue_struct *kthrotld_workqueue;
43
44#define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
45
46/* We measure latency for request size from <= 4k to >= 1M */
47#define LATENCY_BUCKET_SIZE 9
48
49struct latency_bucket {
50 unsigned long total_latency; /* ns / 1024 */
51 int samples;
52};
53
54struct avg_latency_bucket {
55 unsigned long latency; /* ns / 1024 */
56 bool valid;
57};
58
59struct throtl_data
60{
61 /* service tree for active throtl groups */
62 struct throtl_service_queue service_queue;
63
64 struct request_queue *queue;
65
66 /* Total Number of queued bios on READ and WRITE lists */
67 unsigned int nr_queued[2];
68
69 unsigned int throtl_slice;
70
71 /* Work for dispatching throttled bios */
72 struct work_struct dispatch_work;
73 unsigned int limit_index;
74 bool limit_valid[LIMIT_CNT];
75
76 unsigned long low_upgrade_time;
77 unsigned long low_downgrade_time;
78
79 unsigned int scale;
80
81 struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
82 struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
83 struct latency_bucket __percpu *latency_buckets[2];
84 unsigned long last_calculate_time;
85 unsigned long filtered_latency;
86
87 bool track_bio_latency;
88};
89
90static void throtl_pending_timer_fn(struct timer_list *t);
91
92static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
93{
94 return pd_to_blkg(&tg->pd);
95}
96
97/**
98 * sq_to_tg - return the throl_grp the specified service queue belongs to
99 * @sq: the throtl_service_queue of interest
100 *
101 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
102 * embedded in throtl_data, %NULL is returned.
103 */
104static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
105{
106 if (sq && sq->parent_sq)
107 return container_of(sq, struct throtl_grp, service_queue);
108 else
109 return NULL;
110}
111
112/**
113 * sq_to_td - return throtl_data the specified service queue belongs to
114 * @sq: the throtl_service_queue of interest
115 *
116 * A service_queue can be embedded in either a throtl_grp or throtl_data.
117 * Determine the associated throtl_data accordingly and return it.
118 */
119static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
120{
121 struct throtl_grp *tg = sq_to_tg(sq);
122
123 if (tg)
124 return tg->td;
125 else
126 return container_of(sq, struct throtl_data, service_queue);
127}
128
129/*
130 * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
131 * make the IO dispatch more smooth.
132 * Scale up: linearly scale up according to elapsed time since upgrade. For
133 * every throtl_slice, the limit scales up 1/2 .low limit till the
134 * limit hits .max limit
135 * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
136 */
137static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
138{
139 /* arbitrary value to avoid too big scale */
140 if (td->scale < 4096 && time_after_eq(jiffies,
141 td->low_upgrade_time + td->scale * td->throtl_slice))
142 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
143
144 return low + (low >> 1) * td->scale;
145}
146
147static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
148{
149 struct blkcg_gq *blkg = tg_to_blkg(tg);
150 struct throtl_data *td;
151 uint64_t ret;
152
153 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
154 return U64_MAX;
155
156 td = tg->td;
157 ret = tg->bps[rw][td->limit_index];
158 if (ret == 0 && td->limit_index == LIMIT_LOW) {
159 /* intermediate node or iops isn't 0 */
160 if (!list_empty(&blkg->blkcg->css.children) ||
161 tg->iops[rw][td->limit_index])
162 return U64_MAX;
163 else
164 return MIN_THROTL_BPS;
165 }
166
167 if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
168 tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
169 uint64_t adjusted;
170
171 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
172 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
173 }
174 return ret;
175}
176
177static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
178{
179 struct blkcg_gq *blkg = tg_to_blkg(tg);
180 struct throtl_data *td;
181 unsigned int ret;
182
183 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
184 return UINT_MAX;
185
186 td = tg->td;
187 ret = tg->iops[rw][td->limit_index];
188 if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
189 /* intermediate node or bps isn't 0 */
190 if (!list_empty(&blkg->blkcg->css.children) ||
191 tg->bps[rw][td->limit_index])
192 return UINT_MAX;
193 else
194 return MIN_THROTL_IOPS;
195 }
196
197 if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
198 tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
199 uint64_t adjusted;
200
201 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
202 if (adjusted > UINT_MAX)
203 adjusted = UINT_MAX;
204 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
205 }
206 return ret;
207}
208
209#define request_bucket_index(sectors) \
210 clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
211
212/**
213 * throtl_log - log debug message via blktrace
214 * @sq: the service_queue being reported
215 * @fmt: printf format string
216 * @args: printf args
217 *
218 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
219 * throtl_grp; otherwise, just "throtl".
220 */
221#define throtl_log(sq, fmt, args...) do { \
222 struct throtl_grp *__tg = sq_to_tg((sq)); \
223 struct throtl_data *__td = sq_to_td((sq)); \
224 \
225 (void)__td; \
226 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
227 break; \
228 if ((__tg)) { \
229 blk_add_cgroup_trace_msg(__td->queue, \
230 &tg_to_blkg(__tg)->blkcg->css, "throtl " fmt, ##args);\
231 } else { \
232 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
233 } \
234} while (0)
235
236static inline unsigned int throtl_bio_data_size(struct bio *bio)
237{
238 /* assume it's one sector */
239 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
240 return 512;
241 return bio->bi_iter.bi_size;
242}
243
244static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
245{
246 INIT_LIST_HEAD(&qn->node);
247 bio_list_init(&qn->bios);
248 qn->tg = tg;
249}
250
251/**
252 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
253 * @bio: bio being added
254 * @qn: qnode to add bio to
255 * @queued: the service_queue->queued[] list @qn belongs to
256 *
257 * Add @bio to @qn and put @qn on @queued if it's not already on.
258 * @qn->tg's reference count is bumped when @qn is activated. See the
259 * comment on top of throtl_qnode definition for details.
260 */
261static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
262 struct list_head *queued)
263{
264 bio_list_add(&qn->bios, bio);
265 if (list_empty(&qn->node)) {
266 list_add_tail(&qn->node, queued);
267 blkg_get(tg_to_blkg(qn->tg));
268 }
269}
270
271/**
272 * throtl_peek_queued - peek the first bio on a qnode list
273 * @queued: the qnode list to peek
274 */
275static struct bio *throtl_peek_queued(struct list_head *queued)
276{
277 struct throtl_qnode *qn;
278 struct bio *bio;
279
280 if (list_empty(queued))
281 return NULL;
282
283 qn = list_first_entry(queued, struct throtl_qnode, node);
284 bio = bio_list_peek(&qn->bios);
285 WARN_ON_ONCE(!bio);
286 return bio;
287}
288
289/**
290 * throtl_pop_queued - pop the first bio form a qnode list
291 * @queued: the qnode list to pop a bio from
292 * @tg_to_put: optional out argument for throtl_grp to put
293 *
294 * Pop the first bio from the qnode list @queued. After popping, the first
295 * qnode is removed from @queued if empty or moved to the end of @queued so
296 * that the popping order is round-robin.
297 *
298 * When the first qnode is removed, its associated throtl_grp should be put
299 * too. If @tg_to_put is NULL, this function automatically puts it;
300 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
301 * responsible for putting it.
302 */
303static struct bio *throtl_pop_queued(struct list_head *queued,
304 struct throtl_grp **tg_to_put)
305{
306 struct throtl_qnode *qn;
307 struct bio *bio;
308
309 if (list_empty(queued))
310 return NULL;
311
312 qn = list_first_entry(queued, struct throtl_qnode, node);
313 bio = bio_list_pop(&qn->bios);
314 WARN_ON_ONCE(!bio);
315
316 if (bio_list_empty(&qn->bios)) {
317 list_del_init(&qn->node);
318 if (tg_to_put)
319 *tg_to_put = qn->tg;
320 else
321 blkg_put(tg_to_blkg(qn->tg));
322 } else {
323 list_move_tail(&qn->node, queued);
324 }
325
326 return bio;
327}
328
329/* init a service_queue, assumes the caller zeroed it */
330static void throtl_service_queue_init(struct throtl_service_queue *sq)
331{
332 INIT_LIST_HEAD(&sq->queued[READ]);
333 INIT_LIST_HEAD(&sq->queued[WRITE]);
334 sq->pending_tree = RB_ROOT_CACHED;
335 timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
336}
337
338static struct blkg_policy_data *throtl_pd_alloc(struct gendisk *disk,
339 struct blkcg *blkcg, gfp_t gfp)
340{
341 struct throtl_grp *tg;
342 int rw;
343
344 tg = kzalloc_node(sizeof(*tg), gfp, disk->node_id);
345 if (!tg)
346 return NULL;
347
348 if (blkg_rwstat_init(&tg->stat_bytes, gfp))
349 goto err_free_tg;
350
351 if (blkg_rwstat_init(&tg->stat_ios, gfp))
352 goto err_exit_stat_bytes;
353
354 throtl_service_queue_init(&tg->service_queue);
355
356 for (rw = READ; rw <= WRITE; rw++) {
357 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
358 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
359 }
360
361 RB_CLEAR_NODE(&tg->rb_node);
362 tg->bps[READ][LIMIT_MAX] = U64_MAX;
363 tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
364 tg->iops[READ][LIMIT_MAX] = UINT_MAX;
365 tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
366 tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
367 tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
368 tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
369 tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
370 /* LIMIT_LOW will have default value 0 */
371
372 tg->latency_target = DFL_LATENCY_TARGET;
373 tg->latency_target_conf = DFL_LATENCY_TARGET;
374 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
375 tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
376
377 return &tg->pd;
378
379err_exit_stat_bytes:
380 blkg_rwstat_exit(&tg->stat_bytes);
381err_free_tg:
382 kfree(tg);
383 return NULL;
384}
385
386static void throtl_pd_init(struct blkg_policy_data *pd)
387{
388 struct throtl_grp *tg = pd_to_tg(pd);
389 struct blkcg_gq *blkg = tg_to_blkg(tg);
390 struct throtl_data *td = blkg->q->td;
391 struct throtl_service_queue *sq = &tg->service_queue;
392
393 /*
394 * If on the default hierarchy, we switch to properly hierarchical
395 * behavior where limits on a given throtl_grp are applied to the
396 * whole subtree rather than just the group itself. e.g. If 16M
397 * read_bps limit is set on a parent group, summary bps of
398 * parent group and its subtree groups can't exceed 16M for the
399 * device.
400 *
401 * If not on the default hierarchy, the broken flat hierarchy
402 * behavior is retained where all throtl_grps are treated as if
403 * they're all separate root groups right below throtl_data.
404 * Limits of a group don't interact with limits of other groups
405 * regardless of the position of the group in the hierarchy.
406 */
407 sq->parent_sq = &td->service_queue;
408 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
409 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
410 tg->td = td;
411}
412
413/*
414 * Set has_rules[] if @tg or any of its parents have limits configured.
415 * This doesn't require walking up to the top of the hierarchy as the
416 * parent's has_rules[] is guaranteed to be correct.
417 */
418static void tg_update_has_rules(struct throtl_grp *tg)
419{
420 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
421 struct throtl_data *td = tg->td;
422 int rw;
423
424 for (rw = READ; rw <= WRITE; rw++) {
425 tg->has_rules_iops[rw] =
426 (parent_tg && parent_tg->has_rules_iops[rw]) ||
427 (td->limit_valid[td->limit_index] &&
428 tg_iops_limit(tg, rw) != UINT_MAX);
429 tg->has_rules_bps[rw] =
430 (parent_tg && parent_tg->has_rules_bps[rw]) ||
431 (td->limit_valid[td->limit_index] &&
432 (tg_bps_limit(tg, rw) != U64_MAX));
433 }
434}
435
436static void throtl_pd_online(struct blkg_policy_data *pd)
437{
438 struct throtl_grp *tg = pd_to_tg(pd);
439 /*
440 * We don't want new groups to escape the limits of its ancestors.
441 * Update has_rules[] after a new group is brought online.
442 */
443 tg_update_has_rules(tg);
444}
445
446#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
447static void blk_throtl_update_limit_valid(struct throtl_data *td)
448{
449 struct cgroup_subsys_state *pos_css;
450 struct blkcg_gq *blkg;
451 bool low_valid = false;
452
453 rcu_read_lock();
454 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
455 struct throtl_grp *tg = blkg_to_tg(blkg);
456
457 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
458 tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
459 low_valid = true;
460 break;
461 }
462 }
463 rcu_read_unlock();
464
465 td->limit_valid[LIMIT_LOW] = low_valid;
466}
467#else
468static inline void blk_throtl_update_limit_valid(struct throtl_data *td)
469{
470}
471#endif
472
473static void throtl_upgrade_state(struct throtl_data *td);
474static void throtl_pd_offline(struct blkg_policy_data *pd)
475{
476 struct throtl_grp *tg = pd_to_tg(pd);
477
478 tg->bps[READ][LIMIT_LOW] = 0;
479 tg->bps[WRITE][LIMIT_LOW] = 0;
480 tg->iops[READ][LIMIT_LOW] = 0;
481 tg->iops[WRITE][LIMIT_LOW] = 0;
482
483 blk_throtl_update_limit_valid(tg->td);
484
485 if (!tg->td->limit_valid[tg->td->limit_index])
486 throtl_upgrade_state(tg->td);
487}
488
489static void throtl_pd_free(struct blkg_policy_data *pd)
490{
491 struct throtl_grp *tg = pd_to_tg(pd);
492
493 del_timer_sync(&tg->service_queue.pending_timer);
494 blkg_rwstat_exit(&tg->stat_bytes);
495 blkg_rwstat_exit(&tg->stat_ios);
496 kfree(tg);
497}
498
499static struct throtl_grp *
500throtl_rb_first(struct throtl_service_queue *parent_sq)
501{
502 struct rb_node *n;
503
504 n = rb_first_cached(&parent_sq->pending_tree);
505 WARN_ON_ONCE(!n);
506 if (!n)
507 return NULL;
508 return rb_entry_tg(n);
509}
510
511static void throtl_rb_erase(struct rb_node *n,
512 struct throtl_service_queue *parent_sq)
513{
514 rb_erase_cached(n, &parent_sq->pending_tree);
515 RB_CLEAR_NODE(n);
516}
517
518static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
519{
520 struct throtl_grp *tg;
521
522 tg = throtl_rb_first(parent_sq);
523 if (!tg)
524 return;
525
526 parent_sq->first_pending_disptime = tg->disptime;
527}
528
529static void tg_service_queue_add(struct throtl_grp *tg)
530{
531 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
532 struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
533 struct rb_node *parent = NULL;
534 struct throtl_grp *__tg;
535 unsigned long key = tg->disptime;
536 bool leftmost = true;
537
538 while (*node != NULL) {
539 parent = *node;
540 __tg = rb_entry_tg(parent);
541
542 if (time_before(key, __tg->disptime))
543 node = &parent->rb_left;
544 else {
545 node = &parent->rb_right;
546 leftmost = false;
547 }
548 }
549
550 rb_link_node(&tg->rb_node, parent, node);
551 rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
552 leftmost);
553}
554
555static void throtl_enqueue_tg(struct throtl_grp *tg)
556{
557 if (!(tg->flags & THROTL_TG_PENDING)) {
558 tg_service_queue_add(tg);
559 tg->flags |= THROTL_TG_PENDING;
560 tg->service_queue.parent_sq->nr_pending++;
561 }
562}
563
564static void throtl_dequeue_tg(struct throtl_grp *tg)
565{
566 if (tg->flags & THROTL_TG_PENDING) {
567 struct throtl_service_queue *parent_sq =
568 tg->service_queue.parent_sq;
569
570 throtl_rb_erase(&tg->rb_node, parent_sq);
571 --parent_sq->nr_pending;
572 tg->flags &= ~THROTL_TG_PENDING;
573 }
574}
575
576/* Call with queue lock held */
577static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
578 unsigned long expires)
579{
580 unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
581
582 /*
583 * Since we are adjusting the throttle limit dynamically, the sleep
584 * time calculated according to previous limit might be invalid. It's
585 * possible the cgroup sleep time is very long and no other cgroups
586 * have IO running so notify the limit changes. Make sure the cgroup
587 * doesn't sleep too long to avoid the missed notification.
588 */
589 if (time_after(expires, max_expire))
590 expires = max_expire;
591 mod_timer(&sq->pending_timer, expires);
592 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
593 expires - jiffies, jiffies);
594}
595
596/**
597 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
598 * @sq: the service_queue to schedule dispatch for
599 * @force: force scheduling
600 *
601 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
602 * dispatch time of the first pending child. Returns %true if either timer
603 * is armed or there's no pending child left. %false if the current
604 * dispatch window is still open and the caller should continue
605 * dispatching.
606 *
607 * If @force is %true, the dispatch timer is always scheduled and this
608 * function is guaranteed to return %true. This is to be used when the
609 * caller can't dispatch itself and needs to invoke pending_timer
610 * unconditionally. Note that forced scheduling is likely to induce short
611 * delay before dispatch starts even if @sq->first_pending_disptime is not
612 * in the future and thus shouldn't be used in hot paths.
613 */
614static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
615 bool force)
616{
617 /* any pending children left? */
618 if (!sq->nr_pending)
619 return true;
620
621 update_min_dispatch_time(sq);
622
623 /* is the next dispatch time in the future? */
624 if (force || time_after(sq->first_pending_disptime, jiffies)) {
625 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
626 return true;
627 }
628
629 /* tell the caller to continue dispatching */
630 return false;
631}
632
633static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
634 bool rw, unsigned long start)
635{
636 tg->bytes_disp[rw] = 0;
637 tg->io_disp[rw] = 0;
638 tg->carryover_bytes[rw] = 0;
639 tg->carryover_ios[rw] = 0;
640
641 /*
642 * Previous slice has expired. We must have trimmed it after last
643 * bio dispatch. That means since start of last slice, we never used
644 * that bandwidth. Do try to make use of that bandwidth while giving
645 * credit.
646 */
647 if (time_after(start, tg->slice_start[rw]))
648 tg->slice_start[rw] = start;
649
650 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
651 throtl_log(&tg->service_queue,
652 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
653 rw == READ ? 'R' : 'W', tg->slice_start[rw],
654 tg->slice_end[rw], jiffies);
655}
656
657static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw,
658 bool clear_carryover)
659{
660 tg->bytes_disp[rw] = 0;
661 tg->io_disp[rw] = 0;
662 tg->slice_start[rw] = jiffies;
663 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
664 if (clear_carryover) {
665 tg->carryover_bytes[rw] = 0;
666 tg->carryover_ios[rw] = 0;
667 }
668
669 throtl_log(&tg->service_queue,
670 "[%c] new slice start=%lu end=%lu jiffies=%lu",
671 rw == READ ? 'R' : 'W', tg->slice_start[rw],
672 tg->slice_end[rw], jiffies);
673}
674
675static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
676 unsigned long jiffy_end)
677{
678 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
679}
680
681static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
682 unsigned long jiffy_end)
683{
684 throtl_set_slice_end(tg, rw, jiffy_end);
685 throtl_log(&tg->service_queue,
686 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
687 rw == READ ? 'R' : 'W', tg->slice_start[rw],
688 tg->slice_end[rw], jiffies);
689}
690
691/* Determine if previously allocated or extended slice is complete or not */
692static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
693{
694 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
695 return false;
696
697 return true;
698}
699
700static unsigned int calculate_io_allowed(u32 iops_limit,
701 unsigned long jiffy_elapsed)
702{
703 unsigned int io_allowed;
704 u64 tmp;
705
706 /*
707 * jiffy_elapsed should not be a big value as minimum iops can be
708 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
709 * will allow dispatch after 1 second and after that slice should
710 * have been trimmed.
711 */
712
713 tmp = (u64)iops_limit * jiffy_elapsed;
714 do_div(tmp, HZ);
715
716 if (tmp > UINT_MAX)
717 io_allowed = UINT_MAX;
718 else
719 io_allowed = tmp;
720
721 return io_allowed;
722}
723
724static u64 calculate_bytes_allowed(u64 bps_limit, unsigned long jiffy_elapsed)
725{
726 /*
727 * Can result be wider than 64 bits?
728 * We check against 62, not 64, due to ilog2 truncation.
729 */
730 if (ilog2(bps_limit) + ilog2(jiffy_elapsed) - ilog2(HZ) > 62)
731 return U64_MAX;
732 return mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed, (u64)HZ);
733}
734
735/* Trim the used slices and adjust slice start accordingly */
736static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
737{
738 unsigned long time_elapsed;
739 long long bytes_trim;
740 int io_trim;
741
742 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
743
744 /*
745 * If bps are unlimited (-1), then time slice don't get
746 * renewed. Don't try to trim the slice if slice is used. A new
747 * slice will start when appropriate.
748 */
749 if (throtl_slice_used(tg, rw))
750 return;
751
752 /*
753 * A bio has been dispatched. Also adjust slice_end. It might happen
754 * that initially cgroup limit was very low resulting in high
755 * slice_end, but later limit was bumped up and bio was dispatched
756 * sooner, then we need to reduce slice_end. A high bogus slice_end
757 * is bad because it does not allow new slice to start.
758 */
759
760 throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
761
762 time_elapsed = rounddown(jiffies - tg->slice_start[rw],
763 tg->td->throtl_slice);
764 if (!time_elapsed)
765 return;
766
767 bytes_trim = calculate_bytes_allowed(tg_bps_limit(tg, rw),
768 time_elapsed) +
769 tg->carryover_bytes[rw];
770 io_trim = calculate_io_allowed(tg_iops_limit(tg, rw), time_elapsed) +
771 tg->carryover_ios[rw];
772 if (bytes_trim <= 0 && io_trim <= 0)
773 return;
774
775 tg->carryover_bytes[rw] = 0;
776 if ((long long)tg->bytes_disp[rw] >= bytes_trim)
777 tg->bytes_disp[rw] -= bytes_trim;
778 else
779 tg->bytes_disp[rw] = 0;
780
781 tg->carryover_ios[rw] = 0;
782 if ((int)tg->io_disp[rw] >= io_trim)
783 tg->io_disp[rw] -= io_trim;
784 else
785 tg->io_disp[rw] = 0;
786
787 tg->slice_start[rw] += time_elapsed;
788
789 throtl_log(&tg->service_queue,
790 "[%c] trim slice nr=%lu bytes=%lld io=%d start=%lu end=%lu jiffies=%lu",
791 rw == READ ? 'R' : 'W', time_elapsed / tg->td->throtl_slice,
792 bytes_trim, io_trim, tg->slice_start[rw], tg->slice_end[rw],
793 jiffies);
794}
795
796static void __tg_update_carryover(struct throtl_grp *tg, bool rw)
797{
798 unsigned long jiffy_elapsed = jiffies - tg->slice_start[rw];
799 u64 bps_limit = tg_bps_limit(tg, rw);
800 u32 iops_limit = tg_iops_limit(tg, rw);
801
802 /*
803 * If config is updated while bios are still throttled, calculate and
804 * accumulate how many bytes/ios are waited across changes. And
805 * carryover_bytes/ios will be used to calculate new wait time under new
806 * configuration.
807 */
808 if (bps_limit != U64_MAX)
809 tg->carryover_bytes[rw] +=
810 calculate_bytes_allowed(bps_limit, jiffy_elapsed) -
811 tg->bytes_disp[rw];
812 if (iops_limit != UINT_MAX)
813 tg->carryover_ios[rw] +=
814 calculate_io_allowed(iops_limit, jiffy_elapsed) -
815 tg->io_disp[rw];
816}
817
818static void tg_update_carryover(struct throtl_grp *tg)
819{
820 if (tg->service_queue.nr_queued[READ])
821 __tg_update_carryover(tg, READ);
822 if (tg->service_queue.nr_queued[WRITE])
823 __tg_update_carryover(tg, WRITE);
824
825 /* see comments in struct throtl_grp for meaning of these fields. */
826 throtl_log(&tg->service_queue, "%s: %lld %lld %d %d\n", __func__,
827 tg->carryover_bytes[READ], tg->carryover_bytes[WRITE],
828 tg->carryover_ios[READ], tg->carryover_ios[WRITE]);
829}
830
831static unsigned long tg_within_iops_limit(struct throtl_grp *tg, struct bio *bio,
832 u32 iops_limit)
833{
834 bool rw = bio_data_dir(bio);
835 int io_allowed;
836 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
837
838 if (iops_limit == UINT_MAX) {
839 return 0;
840 }
841
842 jiffy_elapsed = jiffies - tg->slice_start[rw];
843
844 /* Round up to the next throttle slice, wait time must be nonzero */
845 jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
846 io_allowed = calculate_io_allowed(iops_limit, jiffy_elapsed_rnd) +
847 tg->carryover_ios[rw];
848 if (io_allowed > 0 && tg->io_disp[rw] + 1 <= io_allowed)
849 return 0;
850
851 /* Calc approx time to dispatch */
852 jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
853 return jiffy_wait;
854}
855
856static unsigned long tg_within_bps_limit(struct throtl_grp *tg, struct bio *bio,
857 u64 bps_limit)
858{
859 bool rw = bio_data_dir(bio);
860 long long bytes_allowed;
861 u64 extra_bytes;
862 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
863 unsigned int bio_size = throtl_bio_data_size(bio);
864
865 /* no need to throttle if this bio's bytes have been accounted */
866 if (bps_limit == U64_MAX || bio_flagged(bio, BIO_BPS_THROTTLED)) {
867 return 0;
868 }
869
870 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
871
872 /* Slice has just started. Consider one slice interval */
873 if (!jiffy_elapsed)
874 jiffy_elapsed_rnd = tg->td->throtl_slice;
875
876 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
877 bytes_allowed = calculate_bytes_allowed(bps_limit, jiffy_elapsed_rnd) +
878 tg->carryover_bytes[rw];
879 if (bytes_allowed > 0 && tg->bytes_disp[rw] + bio_size <= bytes_allowed)
880 return 0;
881
882 /* Calc approx time to dispatch */
883 extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
884 jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
885
886 if (!jiffy_wait)
887 jiffy_wait = 1;
888
889 /*
890 * This wait time is without taking into consideration the rounding
891 * up we did. Add that time also.
892 */
893 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
894 return jiffy_wait;
895}
896
897/*
898 * Returns whether one can dispatch a bio or not. Also returns approx number
899 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
900 */
901static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
902 unsigned long *wait)
903{
904 bool rw = bio_data_dir(bio);
905 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
906 u64 bps_limit = tg_bps_limit(tg, rw);
907 u32 iops_limit = tg_iops_limit(tg, rw);
908
909 /*
910 * Currently whole state machine of group depends on first bio
911 * queued in the group bio list. So one should not be calling
912 * this function with a different bio if there are other bios
913 * queued.
914 */
915 BUG_ON(tg->service_queue.nr_queued[rw] &&
916 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
917
918 /* If tg->bps = -1, then BW is unlimited */
919 if ((bps_limit == U64_MAX && iops_limit == UINT_MAX) ||
920 tg->flags & THROTL_TG_CANCELING) {
921 if (wait)
922 *wait = 0;
923 return true;
924 }
925
926 /*
927 * If previous slice expired, start a new one otherwise renew/extend
928 * existing slice to make sure it is at least throtl_slice interval
929 * long since now. New slice is started only for empty throttle group.
930 * If there is queued bio, that means there should be an active
931 * slice and it should be extended instead.
932 */
933 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
934 throtl_start_new_slice(tg, rw, true);
935 else {
936 if (time_before(tg->slice_end[rw],
937 jiffies + tg->td->throtl_slice))
938 throtl_extend_slice(tg, rw,
939 jiffies + tg->td->throtl_slice);
940 }
941
942 bps_wait = tg_within_bps_limit(tg, bio, bps_limit);
943 iops_wait = tg_within_iops_limit(tg, bio, iops_limit);
944 if (bps_wait + iops_wait == 0) {
945 if (wait)
946 *wait = 0;
947 return true;
948 }
949
950 max_wait = max(bps_wait, iops_wait);
951
952 if (wait)
953 *wait = max_wait;
954
955 if (time_before(tg->slice_end[rw], jiffies + max_wait))
956 throtl_extend_slice(tg, rw, jiffies + max_wait);
957
958 return false;
959}
960
961static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
962{
963 bool rw = bio_data_dir(bio);
964 unsigned int bio_size = throtl_bio_data_size(bio);
965
966 /* Charge the bio to the group */
967 if (!bio_flagged(bio, BIO_BPS_THROTTLED)) {
968 tg->bytes_disp[rw] += bio_size;
969 tg->last_bytes_disp[rw] += bio_size;
970 }
971
972 tg->io_disp[rw]++;
973 tg->last_io_disp[rw]++;
974}
975
976/**
977 * throtl_add_bio_tg - add a bio to the specified throtl_grp
978 * @bio: bio to add
979 * @qn: qnode to use
980 * @tg: the target throtl_grp
981 *
982 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
983 * tg->qnode_on_self[] is used.
984 */
985static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
986 struct throtl_grp *tg)
987{
988 struct throtl_service_queue *sq = &tg->service_queue;
989 bool rw = bio_data_dir(bio);
990
991 if (!qn)
992 qn = &tg->qnode_on_self[rw];
993
994 /*
995 * If @tg doesn't currently have any bios queued in the same
996 * direction, queueing @bio can change when @tg should be
997 * dispatched. Mark that @tg was empty. This is automatically
998 * cleared on the next tg_update_disptime().
999 */
1000 if (!sq->nr_queued[rw])
1001 tg->flags |= THROTL_TG_WAS_EMPTY;
1002
1003 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1004
1005 sq->nr_queued[rw]++;
1006 throtl_enqueue_tg(tg);
1007}
1008
1009static void tg_update_disptime(struct throtl_grp *tg)
1010{
1011 struct throtl_service_queue *sq = &tg->service_queue;
1012 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1013 struct bio *bio;
1014
1015 bio = throtl_peek_queued(&sq->queued[READ]);
1016 if (bio)
1017 tg_may_dispatch(tg, bio, &read_wait);
1018
1019 bio = throtl_peek_queued(&sq->queued[WRITE]);
1020 if (bio)
1021 tg_may_dispatch(tg, bio, &write_wait);
1022
1023 min_wait = min(read_wait, write_wait);
1024 disptime = jiffies + min_wait;
1025
1026 /* Update dispatch time */
1027 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
1028 tg->disptime = disptime;
1029 tg_service_queue_add(tg);
1030
1031 /* see throtl_add_bio_tg() */
1032 tg->flags &= ~THROTL_TG_WAS_EMPTY;
1033}
1034
1035static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1036 struct throtl_grp *parent_tg, bool rw)
1037{
1038 if (throtl_slice_used(parent_tg, rw)) {
1039 throtl_start_new_slice_with_credit(parent_tg, rw,
1040 child_tg->slice_start[rw]);
1041 }
1042
1043}
1044
1045static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1046{
1047 struct throtl_service_queue *sq = &tg->service_queue;
1048 struct throtl_service_queue *parent_sq = sq->parent_sq;
1049 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1050 struct throtl_grp *tg_to_put = NULL;
1051 struct bio *bio;
1052
1053 /*
1054 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1055 * from @tg may put its reference and @parent_sq might end up
1056 * getting released prematurely. Remember the tg to put and put it
1057 * after @bio is transferred to @parent_sq.
1058 */
1059 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1060 sq->nr_queued[rw]--;
1061
1062 throtl_charge_bio(tg, bio);
1063
1064 /*
1065 * If our parent is another tg, we just need to transfer @bio to
1066 * the parent using throtl_add_bio_tg(). If our parent is
1067 * @td->service_queue, @bio is ready to be issued. Put it on its
1068 * bio_lists[] and decrease total number queued. The caller is
1069 * responsible for issuing these bios.
1070 */
1071 if (parent_tg) {
1072 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1073 start_parent_slice_with_credit(tg, parent_tg, rw);
1074 } else {
1075 bio_set_flag(bio, BIO_BPS_THROTTLED);
1076 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1077 &parent_sq->queued[rw]);
1078 BUG_ON(tg->td->nr_queued[rw] <= 0);
1079 tg->td->nr_queued[rw]--;
1080 }
1081
1082 throtl_trim_slice(tg, rw);
1083
1084 if (tg_to_put)
1085 blkg_put(tg_to_blkg(tg_to_put));
1086}
1087
1088static int throtl_dispatch_tg(struct throtl_grp *tg)
1089{
1090 struct throtl_service_queue *sq = &tg->service_queue;
1091 unsigned int nr_reads = 0, nr_writes = 0;
1092 unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1093 unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1094 struct bio *bio;
1095
1096 /* Try to dispatch 75% READS and 25% WRITES */
1097
1098 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1099 tg_may_dispatch(tg, bio, NULL)) {
1100
1101 tg_dispatch_one_bio(tg, READ);
1102 nr_reads++;
1103
1104 if (nr_reads >= max_nr_reads)
1105 break;
1106 }
1107
1108 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1109 tg_may_dispatch(tg, bio, NULL)) {
1110
1111 tg_dispatch_one_bio(tg, WRITE);
1112 nr_writes++;
1113
1114 if (nr_writes >= max_nr_writes)
1115 break;
1116 }
1117
1118 return nr_reads + nr_writes;
1119}
1120
1121static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1122{
1123 unsigned int nr_disp = 0;
1124
1125 while (1) {
1126 struct throtl_grp *tg;
1127 struct throtl_service_queue *sq;
1128
1129 if (!parent_sq->nr_pending)
1130 break;
1131
1132 tg = throtl_rb_first(parent_sq);
1133 if (!tg)
1134 break;
1135
1136 if (time_before(jiffies, tg->disptime))
1137 break;
1138
1139 nr_disp += throtl_dispatch_tg(tg);
1140
1141 sq = &tg->service_queue;
1142 if (sq->nr_queued[READ] || sq->nr_queued[WRITE])
1143 tg_update_disptime(tg);
1144 else
1145 throtl_dequeue_tg(tg);
1146
1147 if (nr_disp >= THROTL_QUANTUM)
1148 break;
1149 }
1150
1151 return nr_disp;
1152}
1153
1154static bool throtl_can_upgrade(struct throtl_data *td,
1155 struct throtl_grp *this_tg);
1156/**
1157 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1158 * @t: the pending_timer member of the throtl_service_queue being serviced
1159 *
1160 * This timer is armed when a child throtl_grp with active bio's become
1161 * pending and queued on the service_queue's pending_tree and expires when
1162 * the first child throtl_grp should be dispatched. This function
1163 * dispatches bio's from the children throtl_grps to the parent
1164 * service_queue.
1165 *
1166 * If the parent's parent is another throtl_grp, dispatching is propagated
1167 * by either arming its pending_timer or repeating dispatch directly. If
1168 * the top-level service_tree is reached, throtl_data->dispatch_work is
1169 * kicked so that the ready bio's are issued.
1170 */
1171static void throtl_pending_timer_fn(struct timer_list *t)
1172{
1173 struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1174 struct throtl_grp *tg = sq_to_tg(sq);
1175 struct throtl_data *td = sq_to_td(sq);
1176 struct throtl_service_queue *parent_sq;
1177 struct request_queue *q;
1178 bool dispatched;
1179 int ret;
1180
1181 /* throtl_data may be gone, so figure out request queue by blkg */
1182 if (tg)
1183 q = tg->pd.blkg->q;
1184 else
1185 q = td->queue;
1186
1187 spin_lock_irq(&q->queue_lock);
1188
1189 if (!q->root_blkg)
1190 goto out_unlock;
1191
1192 if (throtl_can_upgrade(td, NULL))
1193 throtl_upgrade_state(td);
1194
1195again:
1196 parent_sq = sq->parent_sq;
1197 dispatched = false;
1198
1199 while (true) {
1200 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1201 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1202 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1203
1204 ret = throtl_select_dispatch(sq);
1205 if (ret) {
1206 throtl_log(sq, "bios disp=%u", ret);
1207 dispatched = true;
1208 }
1209
1210 if (throtl_schedule_next_dispatch(sq, false))
1211 break;
1212
1213 /* this dispatch windows is still open, relax and repeat */
1214 spin_unlock_irq(&q->queue_lock);
1215 cpu_relax();
1216 spin_lock_irq(&q->queue_lock);
1217 }
1218
1219 if (!dispatched)
1220 goto out_unlock;
1221
1222 if (parent_sq) {
1223 /* @parent_sq is another throl_grp, propagate dispatch */
1224 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1225 tg_update_disptime(tg);
1226 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1227 /* window is already open, repeat dispatching */
1228 sq = parent_sq;
1229 tg = sq_to_tg(sq);
1230 goto again;
1231 }
1232 }
1233 } else {
1234 /* reached the top-level, queue issuing */
1235 queue_work(kthrotld_workqueue, &td->dispatch_work);
1236 }
1237out_unlock:
1238 spin_unlock_irq(&q->queue_lock);
1239}
1240
1241/**
1242 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1243 * @work: work item being executed
1244 *
1245 * This function is queued for execution when bios reach the bio_lists[]
1246 * of throtl_data->service_queue. Those bios are ready and issued by this
1247 * function.
1248 */
1249static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1250{
1251 struct throtl_data *td = container_of(work, struct throtl_data,
1252 dispatch_work);
1253 struct throtl_service_queue *td_sq = &td->service_queue;
1254 struct request_queue *q = td->queue;
1255 struct bio_list bio_list_on_stack;
1256 struct bio *bio;
1257 struct blk_plug plug;
1258 int rw;
1259
1260 bio_list_init(&bio_list_on_stack);
1261
1262 spin_lock_irq(&q->queue_lock);
1263 for (rw = READ; rw <= WRITE; rw++)
1264 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1265 bio_list_add(&bio_list_on_stack, bio);
1266 spin_unlock_irq(&q->queue_lock);
1267
1268 if (!bio_list_empty(&bio_list_on_stack)) {
1269 blk_start_plug(&plug);
1270 while ((bio = bio_list_pop(&bio_list_on_stack)))
1271 submit_bio_noacct_nocheck(bio);
1272 blk_finish_plug(&plug);
1273 }
1274}
1275
1276static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1277 int off)
1278{
1279 struct throtl_grp *tg = pd_to_tg(pd);
1280 u64 v = *(u64 *)((void *)tg + off);
1281
1282 if (v == U64_MAX)
1283 return 0;
1284 return __blkg_prfill_u64(sf, pd, v);
1285}
1286
1287static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1288 int off)
1289{
1290 struct throtl_grp *tg = pd_to_tg(pd);
1291 unsigned int v = *(unsigned int *)((void *)tg + off);
1292
1293 if (v == UINT_MAX)
1294 return 0;
1295 return __blkg_prfill_u64(sf, pd, v);
1296}
1297
1298static int tg_print_conf_u64(struct seq_file *sf, void *v)
1299{
1300 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1301 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1302 return 0;
1303}
1304
1305static int tg_print_conf_uint(struct seq_file *sf, void *v)
1306{
1307 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1308 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1309 return 0;
1310}
1311
1312static void tg_conf_updated(struct throtl_grp *tg, bool global)
1313{
1314 struct throtl_service_queue *sq = &tg->service_queue;
1315 struct cgroup_subsys_state *pos_css;
1316 struct blkcg_gq *blkg;
1317
1318 throtl_log(&tg->service_queue,
1319 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1320 tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1321 tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1322
1323 rcu_read_lock();
1324 /*
1325 * Update has_rules[] flags for the updated tg's subtree. A tg is
1326 * considered to have rules if either the tg itself or any of its
1327 * ancestors has rules. This identifies groups without any
1328 * restrictions in the whole hierarchy and allows them to bypass
1329 * blk-throttle.
1330 */
1331 blkg_for_each_descendant_pre(blkg, pos_css,
1332 global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1333 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1334 struct throtl_grp *parent_tg;
1335
1336 tg_update_has_rules(this_tg);
1337 /* ignore root/second level */
1338 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1339 !blkg->parent->parent)
1340 continue;
1341 parent_tg = blkg_to_tg(blkg->parent);
1342 /*
1343 * make sure all children has lower idle time threshold and
1344 * higher latency target
1345 */
1346 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1347 parent_tg->idletime_threshold);
1348 this_tg->latency_target = max(this_tg->latency_target,
1349 parent_tg->latency_target);
1350 }
1351 rcu_read_unlock();
1352
1353 /*
1354 * We're already holding queue_lock and know @tg is valid. Let's
1355 * apply the new config directly.
1356 *
1357 * Restart the slices for both READ and WRITES. It might happen
1358 * that a group's limit are dropped suddenly and we don't want to
1359 * account recently dispatched IO with new low rate.
1360 */
1361 throtl_start_new_slice(tg, READ, false);
1362 throtl_start_new_slice(tg, WRITE, false);
1363
1364 if (tg->flags & THROTL_TG_PENDING) {
1365 tg_update_disptime(tg);
1366 throtl_schedule_next_dispatch(sq->parent_sq, true);
1367 }
1368}
1369
1370static ssize_t tg_set_conf(struct kernfs_open_file *of,
1371 char *buf, size_t nbytes, loff_t off, bool is_u64)
1372{
1373 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1374 struct blkg_conf_ctx ctx;
1375 struct throtl_grp *tg;
1376 int ret;
1377 u64 v;
1378
1379 blkg_conf_init(&ctx, buf);
1380
1381 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
1382 if (ret)
1383 goto out_finish;
1384
1385 ret = -EINVAL;
1386 if (sscanf(ctx.body, "%llu", &v) != 1)
1387 goto out_finish;
1388 if (!v)
1389 v = U64_MAX;
1390
1391 tg = blkg_to_tg(ctx.blkg);
1392 tg_update_carryover(tg);
1393
1394 if (is_u64)
1395 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1396 else
1397 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1398
1399 tg_conf_updated(tg, false);
1400 ret = 0;
1401out_finish:
1402 blkg_conf_exit(&ctx);
1403 return ret ?: nbytes;
1404}
1405
1406static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1407 char *buf, size_t nbytes, loff_t off)
1408{
1409 return tg_set_conf(of, buf, nbytes, off, true);
1410}
1411
1412static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1413 char *buf, size_t nbytes, loff_t off)
1414{
1415 return tg_set_conf(of, buf, nbytes, off, false);
1416}
1417
1418static int tg_print_rwstat(struct seq_file *sf, void *v)
1419{
1420 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1421 blkg_prfill_rwstat, &blkcg_policy_throtl,
1422 seq_cft(sf)->private, true);
1423 return 0;
1424}
1425
1426static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1427 struct blkg_policy_data *pd, int off)
1428{
1429 struct blkg_rwstat_sample sum;
1430
1431 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1432 &sum);
1433 return __blkg_prfill_rwstat(sf, pd, &sum);
1434}
1435
1436static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1437{
1438 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1439 tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1440 seq_cft(sf)->private, true);
1441 return 0;
1442}
1443
1444static struct cftype throtl_legacy_files[] = {
1445 {
1446 .name = "throttle.read_bps_device",
1447 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1448 .seq_show = tg_print_conf_u64,
1449 .write = tg_set_conf_u64,
1450 },
1451 {
1452 .name = "throttle.write_bps_device",
1453 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1454 .seq_show = tg_print_conf_u64,
1455 .write = tg_set_conf_u64,
1456 },
1457 {
1458 .name = "throttle.read_iops_device",
1459 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1460 .seq_show = tg_print_conf_uint,
1461 .write = tg_set_conf_uint,
1462 },
1463 {
1464 .name = "throttle.write_iops_device",
1465 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1466 .seq_show = tg_print_conf_uint,
1467 .write = tg_set_conf_uint,
1468 },
1469 {
1470 .name = "throttle.io_service_bytes",
1471 .private = offsetof(struct throtl_grp, stat_bytes),
1472 .seq_show = tg_print_rwstat,
1473 },
1474 {
1475 .name = "throttle.io_service_bytes_recursive",
1476 .private = offsetof(struct throtl_grp, stat_bytes),
1477 .seq_show = tg_print_rwstat_recursive,
1478 },
1479 {
1480 .name = "throttle.io_serviced",
1481 .private = offsetof(struct throtl_grp, stat_ios),
1482 .seq_show = tg_print_rwstat,
1483 },
1484 {
1485 .name = "throttle.io_serviced_recursive",
1486 .private = offsetof(struct throtl_grp, stat_ios),
1487 .seq_show = tg_print_rwstat_recursive,
1488 },
1489 { } /* terminate */
1490};
1491
1492static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1493 int off)
1494{
1495 struct throtl_grp *tg = pd_to_tg(pd);
1496 const char *dname = blkg_dev_name(pd->blkg);
1497 char bufs[4][21] = { "max", "max", "max", "max" };
1498 u64 bps_dft;
1499 unsigned int iops_dft;
1500 char idle_time[26] = "";
1501 char latency_time[26] = "";
1502
1503 if (!dname)
1504 return 0;
1505
1506 if (off == LIMIT_LOW) {
1507 bps_dft = 0;
1508 iops_dft = 0;
1509 } else {
1510 bps_dft = U64_MAX;
1511 iops_dft = UINT_MAX;
1512 }
1513
1514 if (tg->bps_conf[READ][off] == bps_dft &&
1515 tg->bps_conf[WRITE][off] == bps_dft &&
1516 tg->iops_conf[READ][off] == iops_dft &&
1517 tg->iops_conf[WRITE][off] == iops_dft &&
1518 (off != LIMIT_LOW ||
1519 (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1520 tg->latency_target_conf == DFL_LATENCY_TARGET)))
1521 return 0;
1522
1523 if (tg->bps_conf[READ][off] != U64_MAX)
1524 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1525 tg->bps_conf[READ][off]);
1526 if (tg->bps_conf[WRITE][off] != U64_MAX)
1527 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1528 tg->bps_conf[WRITE][off]);
1529 if (tg->iops_conf[READ][off] != UINT_MAX)
1530 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1531 tg->iops_conf[READ][off]);
1532 if (tg->iops_conf[WRITE][off] != UINT_MAX)
1533 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1534 tg->iops_conf[WRITE][off]);
1535 if (off == LIMIT_LOW) {
1536 if (tg->idletime_threshold_conf == ULONG_MAX)
1537 strcpy(idle_time, " idle=max");
1538 else
1539 snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1540 tg->idletime_threshold_conf);
1541
1542 if (tg->latency_target_conf == ULONG_MAX)
1543 strcpy(latency_time, " latency=max");
1544 else
1545 snprintf(latency_time, sizeof(latency_time),
1546 " latency=%lu", tg->latency_target_conf);
1547 }
1548
1549 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1550 dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1551 latency_time);
1552 return 0;
1553}
1554
1555static int tg_print_limit(struct seq_file *sf, void *v)
1556{
1557 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1558 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1559 return 0;
1560}
1561
1562static ssize_t tg_set_limit(struct kernfs_open_file *of,
1563 char *buf, size_t nbytes, loff_t off)
1564{
1565 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1566 struct blkg_conf_ctx ctx;
1567 struct throtl_grp *tg;
1568 u64 v[4];
1569 unsigned long idle_time;
1570 unsigned long latency_time;
1571 int ret;
1572 int index = of_cft(of)->private;
1573
1574 blkg_conf_init(&ctx, buf);
1575
1576 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
1577 if (ret)
1578 goto out_finish;
1579
1580 tg = blkg_to_tg(ctx.blkg);
1581 tg_update_carryover(tg);
1582
1583 v[0] = tg->bps_conf[READ][index];
1584 v[1] = tg->bps_conf[WRITE][index];
1585 v[2] = tg->iops_conf[READ][index];
1586 v[3] = tg->iops_conf[WRITE][index];
1587
1588 idle_time = tg->idletime_threshold_conf;
1589 latency_time = tg->latency_target_conf;
1590 while (true) {
1591 char tok[27]; /* wiops=18446744073709551616 */
1592 char *p;
1593 u64 val = U64_MAX;
1594 int len;
1595
1596 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1597 break;
1598 if (tok[0] == '\0')
1599 break;
1600 ctx.body += len;
1601
1602 ret = -EINVAL;
1603 p = tok;
1604 strsep(&p, "=");
1605 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1606 goto out_finish;
1607
1608 ret = -ERANGE;
1609 if (!val)
1610 goto out_finish;
1611
1612 ret = -EINVAL;
1613 if (!strcmp(tok, "rbps") && val > 1)
1614 v[0] = val;
1615 else if (!strcmp(tok, "wbps") && val > 1)
1616 v[1] = val;
1617 else if (!strcmp(tok, "riops") && val > 1)
1618 v[2] = min_t(u64, val, UINT_MAX);
1619 else if (!strcmp(tok, "wiops") && val > 1)
1620 v[3] = min_t(u64, val, UINT_MAX);
1621 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1622 idle_time = val;
1623 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1624 latency_time = val;
1625 else
1626 goto out_finish;
1627 }
1628
1629 tg->bps_conf[READ][index] = v[0];
1630 tg->bps_conf[WRITE][index] = v[1];
1631 tg->iops_conf[READ][index] = v[2];
1632 tg->iops_conf[WRITE][index] = v[3];
1633
1634 if (index == LIMIT_MAX) {
1635 tg->bps[READ][index] = v[0];
1636 tg->bps[WRITE][index] = v[1];
1637 tg->iops[READ][index] = v[2];
1638 tg->iops[WRITE][index] = v[3];
1639 }
1640 tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1641 tg->bps_conf[READ][LIMIT_MAX]);
1642 tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1643 tg->bps_conf[WRITE][LIMIT_MAX]);
1644 tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1645 tg->iops_conf[READ][LIMIT_MAX]);
1646 tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1647 tg->iops_conf[WRITE][LIMIT_MAX]);
1648 tg->idletime_threshold_conf = idle_time;
1649 tg->latency_target_conf = latency_time;
1650
1651 /* force user to configure all settings for low limit */
1652 if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1653 tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1654 tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1655 tg->latency_target_conf == DFL_LATENCY_TARGET) {
1656 tg->bps[READ][LIMIT_LOW] = 0;
1657 tg->bps[WRITE][LIMIT_LOW] = 0;
1658 tg->iops[READ][LIMIT_LOW] = 0;
1659 tg->iops[WRITE][LIMIT_LOW] = 0;
1660 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1661 tg->latency_target = DFL_LATENCY_TARGET;
1662 } else if (index == LIMIT_LOW) {
1663 tg->idletime_threshold = tg->idletime_threshold_conf;
1664 tg->latency_target = tg->latency_target_conf;
1665 }
1666
1667 blk_throtl_update_limit_valid(tg->td);
1668 if (tg->td->limit_valid[LIMIT_LOW]) {
1669 if (index == LIMIT_LOW)
1670 tg->td->limit_index = LIMIT_LOW;
1671 } else
1672 tg->td->limit_index = LIMIT_MAX;
1673 tg_conf_updated(tg, index == LIMIT_LOW &&
1674 tg->td->limit_valid[LIMIT_LOW]);
1675 ret = 0;
1676out_finish:
1677 blkg_conf_exit(&ctx);
1678 return ret ?: nbytes;
1679}
1680
1681static struct cftype throtl_files[] = {
1682#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1683 {
1684 .name = "low",
1685 .flags = CFTYPE_NOT_ON_ROOT,
1686 .seq_show = tg_print_limit,
1687 .write = tg_set_limit,
1688 .private = LIMIT_LOW,
1689 },
1690#endif
1691 {
1692 .name = "max",
1693 .flags = CFTYPE_NOT_ON_ROOT,
1694 .seq_show = tg_print_limit,
1695 .write = tg_set_limit,
1696 .private = LIMIT_MAX,
1697 },
1698 { } /* terminate */
1699};
1700
1701static void throtl_shutdown_wq(struct request_queue *q)
1702{
1703 struct throtl_data *td = q->td;
1704
1705 cancel_work_sync(&td->dispatch_work);
1706}
1707
1708struct blkcg_policy blkcg_policy_throtl = {
1709 .dfl_cftypes = throtl_files,
1710 .legacy_cftypes = throtl_legacy_files,
1711
1712 .pd_alloc_fn = throtl_pd_alloc,
1713 .pd_init_fn = throtl_pd_init,
1714 .pd_online_fn = throtl_pd_online,
1715 .pd_offline_fn = throtl_pd_offline,
1716 .pd_free_fn = throtl_pd_free,
1717};
1718
1719void blk_throtl_cancel_bios(struct gendisk *disk)
1720{
1721 struct request_queue *q = disk->queue;
1722 struct cgroup_subsys_state *pos_css;
1723 struct blkcg_gq *blkg;
1724
1725 spin_lock_irq(&q->queue_lock);
1726 /*
1727 * queue_lock is held, rcu lock is not needed here technically.
1728 * However, rcu lock is still held to emphasize that following
1729 * path need RCU protection and to prevent warning from lockdep.
1730 */
1731 rcu_read_lock();
1732 blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) {
1733 struct throtl_grp *tg = blkg_to_tg(blkg);
1734 struct throtl_service_queue *sq = &tg->service_queue;
1735
1736 /*
1737 * Set the flag to make sure throtl_pending_timer_fn() won't
1738 * stop until all throttled bios are dispatched.
1739 */
1740 tg->flags |= THROTL_TG_CANCELING;
1741
1742 /*
1743 * Do not dispatch cgroup without THROTL_TG_PENDING or cgroup
1744 * will be inserted to service queue without THROTL_TG_PENDING
1745 * set in tg_update_disptime below. Then IO dispatched from
1746 * child in tg_dispatch_one_bio will trigger double insertion
1747 * and corrupt the tree.
1748 */
1749 if (!(tg->flags & THROTL_TG_PENDING))
1750 continue;
1751
1752 /*
1753 * Update disptime after setting the above flag to make sure
1754 * throtl_select_dispatch() won't exit without dispatching.
1755 */
1756 tg_update_disptime(tg);
1757
1758 throtl_schedule_pending_timer(sq, jiffies + 1);
1759 }
1760 rcu_read_unlock();
1761 spin_unlock_irq(&q->queue_lock);
1762}
1763
1764#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1765static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1766{
1767 unsigned long rtime = jiffies, wtime = jiffies;
1768
1769 if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1770 rtime = tg->last_low_overflow_time[READ];
1771 if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1772 wtime = tg->last_low_overflow_time[WRITE];
1773 return min(rtime, wtime);
1774}
1775
1776static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1777{
1778 struct throtl_service_queue *parent_sq;
1779 struct throtl_grp *parent = tg;
1780 unsigned long ret = __tg_last_low_overflow_time(tg);
1781
1782 while (true) {
1783 parent_sq = parent->service_queue.parent_sq;
1784 parent = sq_to_tg(parent_sq);
1785 if (!parent)
1786 break;
1787
1788 /*
1789 * The parent doesn't have low limit, it always reaches low
1790 * limit. Its overflow time is useless for children
1791 */
1792 if (!parent->bps[READ][LIMIT_LOW] &&
1793 !parent->iops[READ][LIMIT_LOW] &&
1794 !parent->bps[WRITE][LIMIT_LOW] &&
1795 !parent->iops[WRITE][LIMIT_LOW])
1796 continue;
1797 if (time_after(__tg_last_low_overflow_time(parent), ret))
1798 ret = __tg_last_low_overflow_time(parent);
1799 }
1800 return ret;
1801}
1802
1803static bool throtl_tg_is_idle(struct throtl_grp *tg)
1804{
1805 /*
1806 * cgroup is idle if:
1807 * - single idle is too long, longer than a fixed value (in case user
1808 * configure a too big threshold) or 4 times of idletime threshold
1809 * - average think time is more than threshold
1810 * - IO latency is largely below threshold
1811 */
1812 unsigned long time;
1813 bool ret;
1814
1815 time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1816 ret = tg->latency_target == DFL_LATENCY_TARGET ||
1817 tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1818 (blk_time_get_ns() >> 10) - tg->last_finish_time > time ||
1819 tg->avg_idletime > tg->idletime_threshold ||
1820 (tg->latency_target && tg->bio_cnt &&
1821 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1822 throtl_log(&tg->service_queue,
1823 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1824 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1825 tg->bio_cnt, ret, tg->td->scale);
1826 return ret;
1827}
1828
1829static bool throtl_low_limit_reached(struct throtl_grp *tg, int rw)
1830{
1831 struct throtl_service_queue *sq = &tg->service_queue;
1832 bool limit = tg->bps[rw][LIMIT_LOW] || tg->iops[rw][LIMIT_LOW];
1833
1834 /*
1835 * if low limit is zero, low limit is always reached.
1836 * if low limit is non-zero, we can check if there is any request
1837 * is queued to determine if low limit is reached as we throttle
1838 * request according to limit.
1839 */
1840 return !limit || sq->nr_queued[rw];
1841}
1842
1843static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1844{
1845 /*
1846 * cgroup reaches low limit when low limit of READ and WRITE are
1847 * both reached, it's ok to upgrade to next limit if cgroup reaches
1848 * low limit
1849 */
1850 if (throtl_low_limit_reached(tg, READ) &&
1851 throtl_low_limit_reached(tg, WRITE))
1852 return true;
1853
1854 if (time_after_eq(jiffies,
1855 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1856 throtl_tg_is_idle(tg))
1857 return true;
1858 return false;
1859}
1860
1861static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1862{
1863 while (true) {
1864 if (throtl_tg_can_upgrade(tg))
1865 return true;
1866 tg = sq_to_tg(tg->service_queue.parent_sq);
1867 if (!tg || !tg_to_blkg(tg)->parent)
1868 return false;
1869 }
1870 return false;
1871}
1872
1873static bool throtl_can_upgrade(struct throtl_data *td,
1874 struct throtl_grp *this_tg)
1875{
1876 struct cgroup_subsys_state *pos_css;
1877 struct blkcg_gq *blkg;
1878
1879 if (td->limit_index != LIMIT_LOW)
1880 return false;
1881
1882 if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1883 return false;
1884
1885 rcu_read_lock();
1886 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1887 struct throtl_grp *tg = blkg_to_tg(blkg);
1888
1889 if (tg == this_tg)
1890 continue;
1891 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1892 continue;
1893 if (!throtl_hierarchy_can_upgrade(tg)) {
1894 rcu_read_unlock();
1895 return false;
1896 }
1897 }
1898 rcu_read_unlock();
1899 return true;
1900}
1901
1902static void throtl_upgrade_check(struct throtl_grp *tg)
1903{
1904 unsigned long now = jiffies;
1905
1906 if (tg->td->limit_index != LIMIT_LOW)
1907 return;
1908
1909 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1910 return;
1911
1912 tg->last_check_time = now;
1913
1914 if (!time_after_eq(now,
1915 __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1916 return;
1917
1918 if (throtl_can_upgrade(tg->td, NULL))
1919 throtl_upgrade_state(tg->td);
1920}
1921
1922static void throtl_upgrade_state(struct throtl_data *td)
1923{
1924 struct cgroup_subsys_state *pos_css;
1925 struct blkcg_gq *blkg;
1926
1927 throtl_log(&td->service_queue, "upgrade to max");
1928 td->limit_index = LIMIT_MAX;
1929 td->low_upgrade_time = jiffies;
1930 td->scale = 0;
1931 rcu_read_lock();
1932 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1933 struct throtl_grp *tg = blkg_to_tg(blkg);
1934 struct throtl_service_queue *sq = &tg->service_queue;
1935
1936 tg->disptime = jiffies - 1;
1937 throtl_select_dispatch(sq);
1938 throtl_schedule_next_dispatch(sq, true);
1939 }
1940 rcu_read_unlock();
1941 throtl_select_dispatch(&td->service_queue);
1942 throtl_schedule_next_dispatch(&td->service_queue, true);
1943 queue_work(kthrotld_workqueue, &td->dispatch_work);
1944}
1945
1946static void throtl_downgrade_state(struct throtl_data *td)
1947{
1948 td->scale /= 2;
1949
1950 throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1951 if (td->scale) {
1952 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1953 return;
1954 }
1955
1956 td->limit_index = LIMIT_LOW;
1957 td->low_downgrade_time = jiffies;
1958}
1959
1960static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1961{
1962 struct throtl_data *td = tg->td;
1963 unsigned long now = jiffies;
1964
1965 /*
1966 * If cgroup is below low limit, consider downgrade and throttle other
1967 * cgroups
1968 */
1969 if (time_after_eq(now, tg_last_low_overflow_time(tg) +
1970 td->throtl_slice) &&
1971 (!throtl_tg_is_idle(tg) ||
1972 !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1973 return true;
1974 return false;
1975}
1976
1977static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1978{
1979 struct throtl_data *td = tg->td;
1980
1981 if (time_before(jiffies, td->low_upgrade_time + td->throtl_slice))
1982 return false;
1983
1984 while (true) {
1985 if (!throtl_tg_can_downgrade(tg))
1986 return false;
1987 tg = sq_to_tg(tg->service_queue.parent_sq);
1988 if (!tg || !tg_to_blkg(tg)->parent)
1989 break;
1990 }
1991 return true;
1992}
1993
1994static void throtl_downgrade_check(struct throtl_grp *tg)
1995{
1996 uint64_t bps;
1997 unsigned int iops;
1998 unsigned long elapsed_time;
1999 unsigned long now = jiffies;
2000
2001 if (tg->td->limit_index != LIMIT_MAX ||
2002 !tg->td->limit_valid[LIMIT_LOW])
2003 return;
2004 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
2005 return;
2006 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
2007 return;
2008
2009 elapsed_time = now - tg->last_check_time;
2010 tg->last_check_time = now;
2011
2012 if (time_before(now, tg_last_low_overflow_time(tg) +
2013 tg->td->throtl_slice))
2014 return;
2015
2016 if (tg->bps[READ][LIMIT_LOW]) {
2017 bps = tg->last_bytes_disp[READ] * HZ;
2018 do_div(bps, elapsed_time);
2019 if (bps >= tg->bps[READ][LIMIT_LOW])
2020 tg->last_low_overflow_time[READ] = now;
2021 }
2022
2023 if (tg->bps[WRITE][LIMIT_LOW]) {
2024 bps = tg->last_bytes_disp[WRITE] * HZ;
2025 do_div(bps, elapsed_time);
2026 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2027 tg->last_low_overflow_time[WRITE] = now;
2028 }
2029
2030 if (tg->iops[READ][LIMIT_LOW]) {
2031 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2032 if (iops >= tg->iops[READ][LIMIT_LOW])
2033 tg->last_low_overflow_time[READ] = now;
2034 }
2035
2036 if (tg->iops[WRITE][LIMIT_LOW]) {
2037 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2038 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2039 tg->last_low_overflow_time[WRITE] = now;
2040 }
2041
2042 /*
2043 * If cgroup is below low limit, consider downgrade and throttle other
2044 * cgroups
2045 */
2046 if (throtl_hierarchy_can_downgrade(tg))
2047 throtl_downgrade_state(tg->td);
2048
2049 tg->last_bytes_disp[READ] = 0;
2050 tg->last_bytes_disp[WRITE] = 0;
2051 tg->last_io_disp[READ] = 0;
2052 tg->last_io_disp[WRITE] = 0;
2053}
2054
2055static void blk_throtl_update_idletime(struct throtl_grp *tg)
2056{
2057 unsigned long now;
2058 unsigned long last_finish_time = tg->last_finish_time;
2059
2060 if (last_finish_time == 0)
2061 return;
2062
2063 now = blk_time_get_ns() >> 10;
2064 if (now <= last_finish_time ||
2065 last_finish_time == tg->checked_last_finish_time)
2066 return;
2067
2068 tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2069 tg->checked_last_finish_time = last_finish_time;
2070}
2071
2072static void throtl_update_latency_buckets(struct throtl_data *td)
2073{
2074 struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2075 int i, cpu, rw;
2076 unsigned long last_latency[2] = { 0 };
2077 unsigned long latency[2];
2078
2079 if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
2080 return;
2081 if (time_before(jiffies, td->last_calculate_time + HZ))
2082 return;
2083 td->last_calculate_time = jiffies;
2084
2085 memset(avg_latency, 0, sizeof(avg_latency));
2086 for (rw = READ; rw <= WRITE; rw++) {
2087 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2088 struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2089
2090 for_each_possible_cpu(cpu) {
2091 struct latency_bucket *bucket;
2092
2093 /* this isn't race free, but ok in practice */
2094 bucket = per_cpu_ptr(td->latency_buckets[rw],
2095 cpu);
2096 tmp->total_latency += bucket[i].total_latency;
2097 tmp->samples += bucket[i].samples;
2098 bucket[i].total_latency = 0;
2099 bucket[i].samples = 0;
2100 }
2101
2102 if (tmp->samples >= 32) {
2103 int samples = tmp->samples;
2104
2105 latency[rw] = tmp->total_latency;
2106
2107 tmp->total_latency = 0;
2108 tmp->samples = 0;
2109 latency[rw] /= samples;
2110 if (latency[rw] == 0)
2111 continue;
2112 avg_latency[rw][i].latency = latency[rw];
2113 }
2114 }
2115 }
2116
2117 for (rw = READ; rw <= WRITE; rw++) {
2118 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2119 if (!avg_latency[rw][i].latency) {
2120 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2121 td->avg_buckets[rw][i].latency =
2122 last_latency[rw];
2123 continue;
2124 }
2125
2126 if (!td->avg_buckets[rw][i].valid)
2127 latency[rw] = avg_latency[rw][i].latency;
2128 else
2129 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2130 avg_latency[rw][i].latency) >> 3;
2131
2132 td->avg_buckets[rw][i].latency = max(latency[rw],
2133 last_latency[rw]);
2134 td->avg_buckets[rw][i].valid = true;
2135 last_latency[rw] = td->avg_buckets[rw][i].latency;
2136 }
2137 }
2138
2139 for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2140 throtl_log(&td->service_queue,
2141 "Latency bucket %d: read latency=%ld, read valid=%d, "
2142 "write latency=%ld, write valid=%d", i,
2143 td->avg_buckets[READ][i].latency,
2144 td->avg_buckets[READ][i].valid,
2145 td->avg_buckets[WRITE][i].latency,
2146 td->avg_buckets[WRITE][i].valid);
2147}
2148#else
2149static inline void throtl_update_latency_buckets(struct throtl_data *td)
2150{
2151}
2152
2153static void blk_throtl_update_idletime(struct throtl_grp *tg)
2154{
2155}
2156
2157static void throtl_downgrade_check(struct throtl_grp *tg)
2158{
2159}
2160
2161static void throtl_upgrade_check(struct throtl_grp *tg)
2162{
2163}
2164
2165static bool throtl_can_upgrade(struct throtl_data *td,
2166 struct throtl_grp *this_tg)
2167{
2168 return false;
2169}
2170
2171static void throtl_upgrade_state(struct throtl_data *td)
2172{
2173}
2174#endif
2175
2176bool __blk_throtl_bio(struct bio *bio)
2177{
2178 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2179 struct blkcg_gq *blkg = bio->bi_blkg;
2180 struct throtl_qnode *qn = NULL;
2181 struct throtl_grp *tg = blkg_to_tg(blkg);
2182 struct throtl_service_queue *sq;
2183 bool rw = bio_data_dir(bio);
2184 bool throttled = false;
2185 struct throtl_data *td = tg->td;
2186
2187 rcu_read_lock();
2188
2189 spin_lock_irq(&q->queue_lock);
2190
2191 throtl_update_latency_buckets(td);
2192
2193 blk_throtl_update_idletime(tg);
2194
2195 sq = &tg->service_queue;
2196
2197again:
2198 while (true) {
2199 if (tg->last_low_overflow_time[rw] == 0)
2200 tg->last_low_overflow_time[rw] = jiffies;
2201 throtl_downgrade_check(tg);
2202 throtl_upgrade_check(tg);
2203 /* throtl is FIFO - if bios are already queued, should queue */
2204 if (sq->nr_queued[rw])
2205 break;
2206
2207 /* if above limits, break to queue */
2208 if (!tg_may_dispatch(tg, bio, NULL)) {
2209 tg->last_low_overflow_time[rw] = jiffies;
2210 if (throtl_can_upgrade(td, tg)) {
2211 throtl_upgrade_state(td);
2212 goto again;
2213 }
2214 break;
2215 }
2216
2217 /* within limits, let's charge and dispatch directly */
2218 throtl_charge_bio(tg, bio);
2219
2220 /*
2221 * We need to trim slice even when bios are not being queued
2222 * otherwise it might happen that a bio is not queued for
2223 * a long time and slice keeps on extending and trim is not
2224 * called for a long time. Now if limits are reduced suddenly
2225 * we take into account all the IO dispatched so far at new
2226 * low rate and * newly queued IO gets a really long dispatch
2227 * time.
2228 *
2229 * So keep on trimming slice even if bio is not queued.
2230 */
2231 throtl_trim_slice(tg, rw);
2232
2233 /*
2234 * @bio passed through this layer without being throttled.
2235 * Climb up the ladder. If we're already at the top, it
2236 * can be executed directly.
2237 */
2238 qn = &tg->qnode_on_parent[rw];
2239 sq = sq->parent_sq;
2240 tg = sq_to_tg(sq);
2241 if (!tg) {
2242 bio_set_flag(bio, BIO_BPS_THROTTLED);
2243 goto out_unlock;
2244 }
2245 }
2246
2247 /* out-of-limit, queue to @tg */
2248 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2249 rw == READ ? 'R' : 'W',
2250 tg->bytes_disp[rw], bio->bi_iter.bi_size,
2251 tg_bps_limit(tg, rw),
2252 tg->io_disp[rw], tg_iops_limit(tg, rw),
2253 sq->nr_queued[READ], sq->nr_queued[WRITE]);
2254
2255 tg->last_low_overflow_time[rw] = jiffies;
2256
2257 td->nr_queued[rw]++;
2258 throtl_add_bio_tg(bio, qn, tg);
2259 throttled = true;
2260
2261 /*
2262 * Update @tg's dispatch time and force schedule dispatch if @tg
2263 * was empty before @bio. The forced scheduling isn't likely to
2264 * cause undue delay as @bio is likely to be dispatched directly if
2265 * its @tg's disptime is not in the future.
2266 */
2267 if (tg->flags & THROTL_TG_WAS_EMPTY) {
2268 tg_update_disptime(tg);
2269 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2270 }
2271
2272out_unlock:
2273#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2274 if (throttled || !td->track_bio_latency)
2275 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2276#endif
2277 spin_unlock_irq(&q->queue_lock);
2278
2279 rcu_read_unlock();
2280 return throttled;
2281}
2282
2283#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2284static void throtl_track_latency(struct throtl_data *td, sector_t size,
2285 enum req_op op, unsigned long time)
2286{
2287 const bool rw = op_is_write(op);
2288 struct latency_bucket *latency;
2289 int index;
2290
2291 if (!td || td->limit_index != LIMIT_LOW ||
2292 !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2293 !blk_queue_nonrot(td->queue))
2294 return;
2295
2296 index = request_bucket_index(size);
2297
2298 latency = get_cpu_ptr(td->latency_buckets[rw]);
2299 latency[index].total_latency += time;
2300 latency[index].samples++;
2301 put_cpu_ptr(td->latency_buckets[rw]);
2302}
2303
2304void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2305{
2306 struct request_queue *q = rq->q;
2307 struct throtl_data *td = q->td;
2308
2309 throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2310 time_ns >> 10);
2311}
2312
2313void blk_throtl_bio_endio(struct bio *bio)
2314{
2315 struct blkcg_gq *blkg;
2316 struct throtl_grp *tg;
2317 u64 finish_time_ns;
2318 unsigned long finish_time;
2319 unsigned long start_time;
2320 unsigned long lat;
2321 int rw = bio_data_dir(bio);
2322
2323 blkg = bio->bi_blkg;
2324 if (!blkg)
2325 return;
2326 tg = blkg_to_tg(blkg);
2327 if (!tg->td->limit_valid[LIMIT_LOW])
2328 return;
2329
2330 finish_time_ns = blk_time_get_ns();
2331 tg->last_finish_time = finish_time_ns >> 10;
2332
2333 start_time = bio_issue_time(&bio->bi_issue) >> 10;
2334 finish_time = __bio_issue_time(finish_time_ns) >> 10;
2335 if (!start_time || finish_time <= start_time)
2336 return;
2337
2338 lat = finish_time - start_time;
2339 /* this is only for bio based driver */
2340 if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2341 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2342 bio_op(bio), lat);
2343
2344 if (tg->latency_target && lat >= tg->td->filtered_latency) {
2345 int bucket;
2346 unsigned int threshold;
2347
2348 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2349 threshold = tg->td->avg_buckets[rw][bucket].latency +
2350 tg->latency_target;
2351 if (lat > threshold)
2352 tg->bad_bio_cnt++;
2353 /*
2354 * Not race free, could get wrong count, which means cgroups
2355 * will be throttled
2356 */
2357 tg->bio_cnt++;
2358 }
2359
2360 if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2361 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2362 tg->bio_cnt /= 2;
2363 tg->bad_bio_cnt /= 2;
2364 }
2365}
2366#endif
2367
2368int blk_throtl_init(struct gendisk *disk)
2369{
2370 struct request_queue *q = disk->queue;
2371 struct throtl_data *td;
2372 int ret;
2373
2374 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2375 if (!td)
2376 return -ENOMEM;
2377 td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2378 LATENCY_BUCKET_SIZE, __alignof__(u64));
2379 if (!td->latency_buckets[READ]) {
2380 kfree(td);
2381 return -ENOMEM;
2382 }
2383 td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2384 LATENCY_BUCKET_SIZE, __alignof__(u64));
2385 if (!td->latency_buckets[WRITE]) {
2386 free_percpu(td->latency_buckets[READ]);
2387 kfree(td);
2388 return -ENOMEM;
2389 }
2390
2391 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2392 throtl_service_queue_init(&td->service_queue);
2393
2394 q->td = td;
2395 td->queue = q;
2396
2397 td->limit_valid[LIMIT_MAX] = true;
2398 td->limit_index = LIMIT_MAX;
2399 td->low_upgrade_time = jiffies;
2400 td->low_downgrade_time = jiffies;
2401
2402 /* activate policy */
2403 ret = blkcg_activate_policy(disk, &blkcg_policy_throtl);
2404 if (ret) {
2405 free_percpu(td->latency_buckets[READ]);
2406 free_percpu(td->latency_buckets[WRITE]);
2407 kfree(td);
2408 }
2409 return ret;
2410}
2411
2412void blk_throtl_exit(struct gendisk *disk)
2413{
2414 struct request_queue *q = disk->queue;
2415
2416 BUG_ON(!q->td);
2417 del_timer_sync(&q->td->service_queue.pending_timer);
2418 throtl_shutdown_wq(q);
2419 blkcg_deactivate_policy(disk, &blkcg_policy_throtl);
2420 free_percpu(q->td->latency_buckets[READ]);
2421 free_percpu(q->td->latency_buckets[WRITE]);
2422 kfree(q->td);
2423}
2424
2425void blk_throtl_register(struct gendisk *disk)
2426{
2427 struct request_queue *q = disk->queue;
2428 struct throtl_data *td;
2429 int i;
2430
2431 td = q->td;
2432 BUG_ON(!td);
2433
2434 if (blk_queue_nonrot(q)) {
2435 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2436 td->filtered_latency = LATENCY_FILTERED_SSD;
2437 } else {
2438 td->throtl_slice = DFL_THROTL_SLICE_HD;
2439 td->filtered_latency = LATENCY_FILTERED_HD;
2440 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2441 td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2442 td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2443 }
2444 }
2445#ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2446 /* if no low limit, use previous default */
2447 td->throtl_slice = DFL_THROTL_SLICE_HD;
2448
2449#else
2450 td->track_bio_latency = !queue_is_mq(q);
2451 if (!td->track_bio_latency)
2452 blk_stat_enable_accounting(q);
2453#endif
2454}
2455
2456#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2457ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2458{
2459 if (!q->td)
2460 return -EINVAL;
2461 return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2462}
2463
2464ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2465 const char *page, size_t count)
2466{
2467 unsigned long v;
2468 unsigned long t;
2469
2470 if (!q->td)
2471 return -EINVAL;
2472 if (kstrtoul(page, 10, &v))
2473 return -EINVAL;
2474 t = msecs_to_jiffies(v);
2475 if (t == 0 || t > MAX_THROTL_SLICE)
2476 return -EINVAL;
2477 q->td->throtl_slice = t;
2478 return count;
2479}
2480#endif
2481
2482static int __init throtl_init(void)
2483{
2484 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2485 if (!kthrotld_workqueue)
2486 panic("Failed to create kthrotld\n");
2487
2488 return blkcg_policy_register(&blkcg_policy_throtl);
2489}
2490
2491module_init(throtl_init);