xref: /linux/block/blk-throttle.c (revision d87c25e8f4051f813762da6a182c57f246b17441)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Interface for controlling IO bandwidth on a request queue
4  *
5  * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
6  */
7 
8 #include <linux/module.h>
9 #include <linux/slab.h>
10 #include <linux/blkdev.h>
11 #include <linux/bio.h>
12 #include <linux/blktrace_api.h>
13 #include <linux/blk-cgroup.h>
14 #include "blk.h"
15 #include "blk-cgroup-rwstat.h"
16 #include "blk-stat.h"
17 #include "blk-throttle.h"
18 
19 /* Max dispatch from a group in 1 round */
20 #define THROTL_GRP_QUANTUM 8
21 
22 /* Total max dispatch from all groups in one round */
23 #define THROTL_QUANTUM 32
24 
25 /* Throttling is performed over a slice and after that slice is renewed */
26 #define DFL_THROTL_SLICE_HD (HZ / 10)
27 #define DFL_THROTL_SLICE_SSD (HZ / 50)
28 #define MAX_THROTL_SLICE (HZ)
29 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
30 #define MIN_THROTL_BPS (320 * 1024)
31 #define MIN_THROTL_IOPS (10)
32 #define DFL_LATENCY_TARGET (-1L)
33 #define DFL_IDLE_THRESHOLD (0)
34 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
35 #define LATENCY_FILTERED_SSD (0)
36 /*
37  * For HD, very small latency comes from sequential IO. Such IO is helpless to
38  * help determine if its IO is impacted by others, hence we ignore the IO
39  */
40 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
41 
42 /* A workqueue to queue throttle related work */
43 static struct workqueue_struct *kthrotld_workqueue;
44 
45 enum tg_state_flags {
46 	THROTL_TG_PENDING	= 1 << 0,	/* on parent's pending tree */
47 	THROTL_TG_WAS_EMPTY	= 1 << 1,	/* bio_lists[] became non-empty */
48 };
49 
50 #define rb_entry_tg(node)	rb_entry((node), struct throtl_grp, rb_node)
51 
52 /* We measure latency for request size from <= 4k to >= 1M */
53 #define LATENCY_BUCKET_SIZE 9
54 
55 struct latency_bucket {
56 	unsigned long total_latency; /* ns / 1024 */
57 	int samples;
58 };
59 
60 struct avg_latency_bucket {
61 	unsigned long latency; /* ns / 1024 */
62 	bool valid;
63 };
64 
65 struct throtl_data
66 {
67 	/* service tree for active throtl groups */
68 	struct throtl_service_queue service_queue;
69 
70 	struct request_queue *queue;
71 
72 	/* Total Number of queued bios on READ and WRITE lists */
73 	unsigned int nr_queued[2];
74 
75 	unsigned int throtl_slice;
76 
77 	/* Work for dispatching throttled bios */
78 	struct work_struct dispatch_work;
79 	unsigned int limit_index;
80 	bool limit_valid[LIMIT_CNT];
81 
82 	unsigned long low_upgrade_time;
83 	unsigned long low_downgrade_time;
84 
85 	unsigned int scale;
86 
87 	struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
88 	struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
89 	struct latency_bucket __percpu *latency_buckets[2];
90 	unsigned long last_calculate_time;
91 	unsigned long filtered_latency;
92 
93 	bool track_bio_latency;
94 };
95 
96 static void throtl_pending_timer_fn(struct timer_list *t);
97 
98 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
99 {
100 	return pd_to_blkg(&tg->pd);
101 }
102 
103 /**
104  * sq_to_tg - return the throl_grp the specified service queue belongs to
105  * @sq: the throtl_service_queue of interest
106  *
107  * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
108  * embedded in throtl_data, %NULL is returned.
109  */
110 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
111 {
112 	if (sq && sq->parent_sq)
113 		return container_of(sq, struct throtl_grp, service_queue);
114 	else
115 		return NULL;
116 }
117 
118 /**
119  * sq_to_td - return throtl_data the specified service queue belongs to
120  * @sq: the throtl_service_queue of interest
121  *
122  * A service_queue can be embedded in either a throtl_grp or throtl_data.
123  * Determine the associated throtl_data accordingly and return it.
124  */
125 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
126 {
127 	struct throtl_grp *tg = sq_to_tg(sq);
128 
129 	if (tg)
130 		return tg->td;
131 	else
132 		return container_of(sq, struct throtl_data, service_queue);
133 }
134 
135 /*
136  * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
137  * make the IO dispatch more smooth.
138  * Scale up: linearly scale up according to lapsed time since upgrade. For
139  *           every throtl_slice, the limit scales up 1/2 .low limit till the
140  *           limit hits .max limit
141  * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
142  */
143 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
144 {
145 	/* arbitrary value to avoid too big scale */
146 	if (td->scale < 4096 && time_after_eq(jiffies,
147 	    td->low_upgrade_time + td->scale * td->throtl_slice))
148 		td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
149 
150 	return low + (low >> 1) * td->scale;
151 }
152 
153 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
154 {
155 	struct blkcg_gq *blkg = tg_to_blkg(tg);
156 	struct throtl_data *td;
157 	uint64_t ret;
158 
159 	if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
160 		return U64_MAX;
161 
162 	td = tg->td;
163 	ret = tg->bps[rw][td->limit_index];
164 	if (ret == 0 && td->limit_index == LIMIT_LOW) {
165 		/* intermediate node or iops isn't 0 */
166 		if (!list_empty(&blkg->blkcg->css.children) ||
167 		    tg->iops[rw][td->limit_index])
168 			return U64_MAX;
169 		else
170 			return MIN_THROTL_BPS;
171 	}
172 
173 	if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
174 	    tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
175 		uint64_t adjusted;
176 
177 		adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
178 		ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
179 	}
180 	return ret;
181 }
182 
183 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
184 {
185 	struct blkcg_gq *blkg = tg_to_blkg(tg);
186 	struct throtl_data *td;
187 	unsigned int ret;
188 
189 	if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
190 		return UINT_MAX;
191 
192 	td = tg->td;
193 	ret = tg->iops[rw][td->limit_index];
194 	if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
195 		/* intermediate node or bps isn't 0 */
196 		if (!list_empty(&blkg->blkcg->css.children) ||
197 		    tg->bps[rw][td->limit_index])
198 			return UINT_MAX;
199 		else
200 			return MIN_THROTL_IOPS;
201 	}
202 
203 	if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
204 	    tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
205 		uint64_t adjusted;
206 
207 		adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
208 		if (adjusted > UINT_MAX)
209 			adjusted = UINT_MAX;
210 		ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
211 	}
212 	return ret;
213 }
214 
215 #define request_bucket_index(sectors) \
216 	clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
217 
218 /**
219  * throtl_log - log debug message via blktrace
220  * @sq: the service_queue being reported
221  * @fmt: printf format string
222  * @args: printf args
223  *
224  * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
225  * throtl_grp; otherwise, just "throtl".
226  */
227 #define throtl_log(sq, fmt, args...)	do {				\
228 	struct throtl_grp *__tg = sq_to_tg((sq));			\
229 	struct throtl_data *__td = sq_to_td((sq));			\
230 									\
231 	(void)__td;							\
232 	if (likely(!blk_trace_note_message_enabled(__td->queue)))	\
233 		break;							\
234 	if ((__tg)) {							\
235 		blk_add_cgroup_trace_msg(__td->queue,			\
236 			tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\
237 	} else {							\
238 		blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);	\
239 	}								\
240 } while (0)
241 
242 static inline unsigned int throtl_bio_data_size(struct bio *bio)
243 {
244 	/* assume it's one sector */
245 	if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
246 		return 512;
247 	return bio->bi_iter.bi_size;
248 }
249 
250 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
251 {
252 	INIT_LIST_HEAD(&qn->node);
253 	bio_list_init(&qn->bios);
254 	qn->tg = tg;
255 }
256 
257 /**
258  * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
259  * @bio: bio being added
260  * @qn: qnode to add bio to
261  * @queued: the service_queue->queued[] list @qn belongs to
262  *
263  * Add @bio to @qn and put @qn on @queued if it's not already on.
264  * @qn->tg's reference count is bumped when @qn is activated.  See the
265  * comment on top of throtl_qnode definition for details.
266  */
267 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
268 				 struct list_head *queued)
269 {
270 	bio_list_add(&qn->bios, bio);
271 	if (list_empty(&qn->node)) {
272 		list_add_tail(&qn->node, queued);
273 		blkg_get(tg_to_blkg(qn->tg));
274 	}
275 }
276 
277 /**
278  * throtl_peek_queued - peek the first bio on a qnode list
279  * @queued: the qnode list to peek
280  */
281 static struct bio *throtl_peek_queued(struct list_head *queued)
282 {
283 	struct throtl_qnode *qn;
284 	struct bio *bio;
285 
286 	if (list_empty(queued))
287 		return NULL;
288 
289 	qn = list_first_entry(queued, struct throtl_qnode, node);
290 	bio = bio_list_peek(&qn->bios);
291 	WARN_ON_ONCE(!bio);
292 	return bio;
293 }
294 
295 /**
296  * throtl_pop_queued - pop the first bio form a qnode list
297  * @queued: the qnode list to pop a bio from
298  * @tg_to_put: optional out argument for throtl_grp to put
299  *
300  * Pop the first bio from the qnode list @queued.  After popping, the first
301  * qnode is removed from @queued if empty or moved to the end of @queued so
302  * that the popping order is round-robin.
303  *
304  * When the first qnode is removed, its associated throtl_grp should be put
305  * too.  If @tg_to_put is NULL, this function automatically puts it;
306  * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
307  * responsible for putting it.
308  */
309 static struct bio *throtl_pop_queued(struct list_head *queued,
310 				     struct throtl_grp **tg_to_put)
311 {
312 	struct throtl_qnode *qn;
313 	struct bio *bio;
314 
315 	if (list_empty(queued))
316 		return NULL;
317 
318 	qn = list_first_entry(queued, struct throtl_qnode, node);
319 	bio = bio_list_pop(&qn->bios);
320 	WARN_ON_ONCE(!bio);
321 
322 	if (bio_list_empty(&qn->bios)) {
323 		list_del_init(&qn->node);
324 		if (tg_to_put)
325 			*tg_to_put = qn->tg;
326 		else
327 			blkg_put(tg_to_blkg(qn->tg));
328 	} else {
329 		list_move_tail(&qn->node, queued);
330 	}
331 
332 	return bio;
333 }
334 
335 /* init a service_queue, assumes the caller zeroed it */
336 static void throtl_service_queue_init(struct throtl_service_queue *sq)
337 {
338 	INIT_LIST_HEAD(&sq->queued[0]);
339 	INIT_LIST_HEAD(&sq->queued[1]);
340 	sq->pending_tree = RB_ROOT_CACHED;
341 	timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
342 }
343 
344 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp,
345 						struct request_queue *q,
346 						struct blkcg *blkcg)
347 {
348 	struct throtl_grp *tg;
349 	int rw;
350 
351 	tg = kzalloc_node(sizeof(*tg), gfp, q->node);
352 	if (!tg)
353 		return NULL;
354 
355 	if (blkg_rwstat_init(&tg->stat_bytes, gfp))
356 		goto err_free_tg;
357 
358 	if (blkg_rwstat_init(&tg->stat_ios, gfp))
359 		goto err_exit_stat_bytes;
360 
361 	throtl_service_queue_init(&tg->service_queue);
362 
363 	for (rw = READ; rw <= WRITE; rw++) {
364 		throtl_qnode_init(&tg->qnode_on_self[rw], tg);
365 		throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
366 	}
367 
368 	RB_CLEAR_NODE(&tg->rb_node);
369 	tg->bps[READ][LIMIT_MAX] = U64_MAX;
370 	tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
371 	tg->iops[READ][LIMIT_MAX] = UINT_MAX;
372 	tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
373 	tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
374 	tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
375 	tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
376 	tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
377 	/* LIMIT_LOW will have default value 0 */
378 
379 	tg->latency_target = DFL_LATENCY_TARGET;
380 	tg->latency_target_conf = DFL_LATENCY_TARGET;
381 	tg->idletime_threshold = DFL_IDLE_THRESHOLD;
382 	tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
383 
384 	return &tg->pd;
385 
386 err_exit_stat_bytes:
387 	blkg_rwstat_exit(&tg->stat_bytes);
388 err_free_tg:
389 	kfree(tg);
390 	return NULL;
391 }
392 
393 static void throtl_pd_init(struct blkg_policy_data *pd)
394 {
395 	struct throtl_grp *tg = pd_to_tg(pd);
396 	struct blkcg_gq *blkg = tg_to_blkg(tg);
397 	struct throtl_data *td = blkg->q->td;
398 	struct throtl_service_queue *sq = &tg->service_queue;
399 
400 	/*
401 	 * If on the default hierarchy, we switch to properly hierarchical
402 	 * behavior where limits on a given throtl_grp are applied to the
403 	 * whole subtree rather than just the group itself.  e.g. If 16M
404 	 * read_bps limit is set on the root group, the whole system can't
405 	 * exceed 16M for the device.
406 	 *
407 	 * If not on the default hierarchy, the broken flat hierarchy
408 	 * behavior is retained where all throtl_grps are treated as if
409 	 * they're all separate root groups right below throtl_data.
410 	 * Limits of a group don't interact with limits of other groups
411 	 * regardless of the position of the group in the hierarchy.
412 	 */
413 	sq->parent_sq = &td->service_queue;
414 	if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
415 		sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
416 	tg->td = td;
417 }
418 
419 /*
420  * Set has_rules[] if @tg or any of its parents have limits configured.
421  * This doesn't require walking up to the top of the hierarchy as the
422  * parent's has_rules[] is guaranteed to be correct.
423  */
424 static void tg_update_has_rules(struct throtl_grp *tg)
425 {
426 	struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
427 	struct throtl_data *td = tg->td;
428 	int rw;
429 
430 	for (rw = READ; rw <= WRITE; rw++)
431 		tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
432 			(td->limit_valid[td->limit_index] &&
433 			 (tg_bps_limit(tg, rw) != U64_MAX ||
434 			  tg_iops_limit(tg, rw) != UINT_MAX));
435 }
436 
437 static void throtl_pd_online(struct blkg_policy_data *pd)
438 {
439 	struct throtl_grp *tg = pd_to_tg(pd);
440 	/*
441 	 * We don't want new groups to escape the limits of its ancestors.
442 	 * Update has_rules[] after a new group is brought online.
443 	 */
444 	tg_update_has_rules(tg);
445 }
446 
447 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
448 static void blk_throtl_update_limit_valid(struct throtl_data *td)
449 {
450 	struct cgroup_subsys_state *pos_css;
451 	struct blkcg_gq *blkg;
452 	bool low_valid = false;
453 
454 	rcu_read_lock();
455 	blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
456 		struct throtl_grp *tg = blkg_to_tg(blkg);
457 
458 		if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
459 		    tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
460 			low_valid = true;
461 			break;
462 		}
463 	}
464 	rcu_read_unlock();
465 
466 	td->limit_valid[LIMIT_LOW] = low_valid;
467 }
468 #else
469 static inline void blk_throtl_update_limit_valid(struct throtl_data *td)
470 {
471 }
472 #endif
473 
474 static void throtl_upgrade_state(struct throtl_data *td);
475 static void throtl_pd_offline(struct blkg_policy_data *pd)
476 {
477 	struct throtl_grp *tg = pd_to_tg(pd);
478 
479 	tg->bps[READ][LIMIT_LOW] = 0;
480 	tg->bps[WRITE][LIMIT_LOW] = 0;
481 	tg->iops[READ][LIMIT_LOW] = 0;
482 	tg->iops[WRITE][LIMIT_LOW] = 0;
483 
484 	blk_throtl_update_limit_valid(tg->td);
485 
486 	if (!tg->td->limit_valid[tg->td->limit_index])
487 		throtl_upgrade_state(tg->td);
488 }
489 
490 static void throtl_pd_free(struct blkg_policy_data *pd)
491 {
492 	struct throtl_grp *tg = pd_to_tg(pd);
493 
494 	del_timer_sync(&tg->service_queue.pending_timer);
495 	blkg_rwstat_exit(&tg->stat_bytes);
496 	blkg_rwstat_exit(&tg->stat_ios);
497 	kfree(tg);
498 }
499 
500 static struct throtl_grp *
501 throtl_rb_first(struct throtl_service_queue *parent_sq)
502 {
503 	struct rb_node *n;
504 
505 	n = rb_first_cached(&parent_sq->pending_tree);
506 	WARN_ON_ONCE(!n);
507 	if (!n)
508 		return NULL;
509 	return rb_entry_tg(n);
510 }
511 
512 static void throtl_rb_erase(struct rb_node *n,
513 			    struct throtl_service_queue *parent_sq)
514 {
515 	rb_erase_cached(n, &parent_sq->pending_tree);
516 	RB_CLEAR_NODE(n);
517 	--parent_sq->nr_pending;
518 }
519 
520 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
521 {
522 	struct throtl_grp *tg;
523 
524 	tg = throtl_rb_first(parent_sq);
525 	if (!tg)
526 		return;
527 
528 	parent_sq->first_pending_disptime = tg->disptime;
529 }
530 
531 static void tg_service_queue_add(struct throtl_grp *tg)
532 {
533 	struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
534 	struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
535 	struct rb_node *parent = NULL;
536 	struct throtl_grp *__tg;
537 	unsigned long key = tg->disptime;
538 	bool leftmost = true;
539 
540 	while (*node != NULL) {
541 		parent = *node;
542 		__tg = rb_entry_tg(parent);
543 
544 		if (time_before(key, __tg->disptime))
545 			node = &parent->rb_left;
546 		else {
547 			node = &parent->rb_right;
548 			leftmost = false;
549 		}
550 	}
551 
552 	rb_link_node(&tg->rb_node, parent, node);
553 	rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
554 			       leftmost);
555 }
556 
557 static void throtl_enqueue_tg(struct throtl_grp *tg)
558 {
559 	if (!(tg->flags & THROTL_TG_PENDING)) {
560 		tg_service_queue_add(tg);
561 		tg->flags |= THROTL_TG_PENDING;
562 		tg->service_queue.parent_sq->nr_pending++;
563 	}
564 }
565 
566 static void throtl_dequeue_tg(struct throtl_grp *tg)
567 {
568 	if (tg->flags & THROTL_TG_PENDING) {
569 		throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
570 		tg->flags &= ~THROTL_TG_PENDING;
571 	}
572 }
573 
574 /* Call with queue lock held */
575 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
576 					  unsigned long expires)
577 {
578 	unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
579 
580 	/*
581 	 * Since we are adjusting the throttle limit dynamically, the sleep
582 	 * time calculated according to previous limit might be invalid. It's
583 	 * possible the cgroup sleep time is very long and no other cgroups
584 	 * have IO running so notify the limit changes. Make sure the cgroup
585 	 * doesn't sleep too long to avoid the missed notification.
586 	 */
587 	if (time_after(expires, max_expire))
588 		expires = max_expire;
589 	mod_timer(&sq->pending_timer, expires);
590 	throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
591 		   expires - jiffies, jiffies);
592 }
593 
594 /**
595  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
596  * @sq: the service_queue to schedule dispatch for
597  * @force: force scheduling
598  *
599  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
600  * dispatch time of the first pending child.  Returns %true if either timer
601  * is armed or there's no pending child left.  %false if the current
602  * dispatch window is still open and the caller should continue
603  * dispatching.
604  *
605  * If @force is %true, the dispatch timer is always scheduled and this
606  * function is guaranteed to return %true.  This is to be used when the
607  * caller can't dispatch itself and needs to invoke pending_timer
608  * unconditionally.  Note that forced scheduling is likely to induce short
609  * delay before dispatch starts even if @sq->first_pending_disptime is not
610  * in the future and thus shouldn't be used in hot paths.
611  */
612 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
613 					  bool force)
614 {
615 	/* any pending children left? */
616 	if (!sq->nr_pending)
617 		return true;
618 
619 	update_min_dispatch_time(sq);
620 
621 	/* is the next dispatch time in the future? */
622 	if (force || time_after(sq->first_pending_disptime, jiffies)) {
623 		throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
624 		return true;
625 	}
626 
627 	/* tell the caller to continue dispatching */
628 	return false;
629 }
630 
631 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
632 		bool rw, unsigned long start)
633 {
634 	tg->bytes_disp[rw] = 0;
635 	tg->io_disp[rw] = 0;
636 
637 	atomic_set(&tg->io_split_cnt[rw], 0);
638 
639 	/*
640 	 * Previous slice has expired. We must have trimmed it after last
641 	 * bio dispatch. That means since start of last slice, we never used
642 	 * that bandwidth. Do try to make use of that bandwidth while giving
643 	 * credit.
644 	 */
645 	if (time_after_eq(start, tg->slice_start[rw]))
646 		tg->slice_start[rw] = start;
647 
648 	tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
649 	throtl_log(&tg->service_queue,
650 		   "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
651 		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
652 		   tg->slice_end[rw], jiffies);
653 }
654 
655 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
656 {
657 	tg->bytes_disp[rw] = 0;
658 	tg->io_disp[rw] = 0;
659 	tg->slice_start[rw] = jiffies;
660 	tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
661 
662 	atomic_set(&tg->io_split_cnt[rw], 0);
663 
664 	throtl_log(&tg->service_queue,
665 		   "[%c] new slice start=%lu end=%lu jiffies=%lu",
666 		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
667 		   tg->slice_end[rw], jiffies);
668 }
669 
670 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
671 					unsigned long jiffy_end)
672 {
673 	tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
674 }
675 
676 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
677 				       unsigned long jiffy_end)
678 {
679 	throtl_set_slice_end(tg, rw, jiffy_end);
680 	throtl_log(&tg->service_queue,
681 		   "[%c] extend slice start=%lu end=%lu jiffies=%lu",
682 		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
683 		   tg->slice_end[rw], jiffies);
684 }
685 
686 /* Determine if previously allocated or extended slice is complete or not */
687 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
688 {
689 	if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
690 		return false;
691 
692 	return true;
693 }
694 
695 /* Trim the used slices and adjust slice start accordingly */
696 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
697 {
698 	unsigned long nr_slices, time_elapsed, io_trim;
699 	u64 bytes_trim, tmp;
700 
701 	BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
702 
703 	/*
704 	 * If bps are unlimited (-1), then time slice don't get
705 	 * renewed. Don't try to trim the slice if slice is used. A new
706 	 * slice will start when appropriate.
707 	 */
708 	if (throtl_slice_used(tg, rw))
709 		return;
710 
711 	/*
712 	 * A bio has been dispatched. Also adjust slice_end. It might happen
713 	 * that initially cgroup limit was very low resulting in high
714 	 * slice_end, but later limit was bumped up and bio was dispatched
715 	 * sooner, then we need to reduce slice_end. A high bogus slice_end
716 	 * is bad because it does not allow new slice to start.
717 	 */
718 
719 	throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
720 
721 	time_elapsed = jiffies - tg->slice_start[rw];
722 
723 	nr_slices = time_elapsed / tg->td->throtl_slice;
724 
725 	if (!nr_slices)
726 		return;
727 	tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
728 	do_div(tmp, HZ);
729 	bytes_trim = tmp;
730 
731 	io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
732 		HZ;
733 
734 	if (!bytes_trim && !io_trim)
735 		return;
736 
737 	if (tg->bytes_disp[rw] >= bytes_trim)
738 		tg->bytes_disp[rw] -= bytes_trim;
739 	else
740 		tg->bytes_disp[rw] = 0;
741 
742 	if (tg->io_disp[rw] >= io_trim)
743 		tg->io_disp[rw] -= io_trim;
744 	else
745 		tg->io_disp[rw] = 0;
746 
747 	tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
748 
749 	throtl_log(&tg->service_queue,
750 		   "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
751 		   rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
752 		   tg->slice_start[rw], tg->slice_end[rw], jiffies);
753 }
754 
755 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
756 				  u32 iops_limit, unsigned long *wait)
757 {
758 	bool rw = bio_data_dir(bio);
759 	unsigned int io_allowed;
760 	unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
761 	u64 tmp;
762 
763 	if (iops_limit == UINT_MAX) {
764 		if (wait)
765 			*wait = 0;
766 		return true;
767 	}
768 
769 	jiffy_elapsed = jiffies - tg->slice_start[rw];
770 
771 	/* Round up to the next throttle slice, wait time must be nonzero */
772 	jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
773 
774 	/*
775 	 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
776 	 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
777 	 * will allow dispatch after 1 second and after that slice should
778 	 * have been trimmed.
779 	 */
780 
781 	tmp = (u64)iops_limit * jiffy_elapsed_rnd;
782 	do_div(tmp, HZ);
783 
784 	if (tmp > UINT_MAX)
785 		io_allowed = UINT_MAX;
786 	else
787 		io_allowed = tmp;
788 
789 	if (tg->io_disp[rw] + 1 <= io_allowed) {
790 		if (wait)
791 			*wait = 0;
792 		return true;
793 	}
794 
795 	/* Calc approx time to dispatch */
796 	jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
797 
798 	if (wait)
799 		*wait = jiffy_wait;
800 	return false;
801 }
802 
803 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
804 				 u64 bps_limit, unsigned long *wait)
805 {
806 	bool rw = bio_data_dir(bio);
807 	u64 bytes_allowed, extra_bytes, tmp;
808 	unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
809 	unsigned int bio_size = throtl_bio_data_size(bio);
810 
811 	if (bps_limit == U64_MAX) {
812 		if (wait)
813 			*wait = 0;
814 		return true;
815 	}
816 
817 	jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
818 
819 	/* Slice has just started. Consider one slice interval */
820 	if (!jiffy_elapsed)
821 		jiffy_elapsed_rnd = tg->td->throtl_slice;
822 
823 	jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
824 
825 	tmp = bps_limit * jiffy_elapsed_rnd;
826 	do_div(tmp, HZ);
827 	bytes_allowed = tmp;
828 
829 	if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
830 		if (wait)
831 			*wait = 0;
832 		return true;
833 	}
834 
835 	/* Calc approx time to dispatch */
836 	extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
837 	jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
838 
839 	if (!jiffy_wait)
840 		jiffy_wait = 1;
841 
842 	/*
843 	 * This wait time is without taking into consideration the rounding
844 	 * up we did. Add that time also.
845 	 */
846 	jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
847 	if (wait)
848 		*wait = jiffy_wait;
849 	return false;
850 }
851 
852 /*
853  * Returns whether one can dispatch a bio or not. Also returns approx number
854  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
855  */
856 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
857 			    unsigned long *wait)
858 {
859 	bool rw = bio_data_dir(bio);
860 	unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
861 	u64 bps_limit = tg_bps_limit(tg, rw);
862 	u32 iops_limit = tg_iops_limit(tg, rw);
863 
864 	/*
865  	 * Currently whole state machine of group depends on first bio
866 	 * queued in the group bio list. So one should not be calling
867 	 * this function with a different bio if there are other bios
868 	 * queued.
869 	 */
870 	BUG_ON(tg->service_queue.nr_queued[rw] &&
871 	       bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
872 
873 	/* If tg->bps = -1, then BW is unlimited */
874 	if (bps_limit == U64_MAX && iops_limit == UINT_MAX) {
875 		if (wait)
876 			*wait = 0;
877 		return true;
878 	}
879 
880 	/*
881 	 * If previous slice expired, start a new one otherwise renew/extend
882 	 * existing slice to make sure it is at least throtl_slice interval
883 	 * long since now. New slice is started only for empty throttle group.
884 	 * If there is queued bio, that means there should be an active
885 	 * slice and it should be extended instead.
886 	 */
887 	if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
888 		throtl_start_new_slice(tg, rw);
889 	else {
890 		if (time_before(tg->slice_end[rw],
891 		    jiffies + tg->td->throtl_slice))
892 			throtl_extend_slice(tg, rw,
893 				jiffies + tg->td->throtl_slice);
894 	}
895 
896 	if (iops_limit != UINT_MAX)
897 		tg->io_disp[rw] += atomic_xchg(&tg->io_split_cnt[rw], 0);
898 
899 	if (tg_with_in_bps_limit(tg, bio, bps_limit, &bps_wait) &&
900 	    tg_with_in_iops_limit(tg, bio, iops_limit, &iops_wait)) {
901 		if (wait)
902 			*wait = 0;
903 		return true;
904 	}
905 
906 	max_wait = max(bps_wait, iops_wait);
907 
908 	if (wait)
909 		*wait = max_wait;
910 
911 	if (time_before(tg->slice_end[rw], jiffies + max_wait))
912 		throtl_extend_slice(tg, rw, jiffies + max_wait);
913 
914 	return false;
915 }
916 
917 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
918 {
919 	bool rw = bio_data_dir(bio);
920 	unsigned int bio_size = throtl_bio_data_size(bio);
921 
922 	/* Charge the bio to the group */
923 	tg->bytes_disp[rw] += bio_size;
924 	tg->io_disp[rw]++;
925 	tg->last_bytes_disp[rw] += bio_size;
926 	tg->last_io_disp[rw]++;
927 
928 	/*
929 	 * BIO_THROTTLED is used to prevent the same bio to be throttled
930 	 * more than once as a throttled bio will go through blk-throtl the
931 	 * second time when it eventually gets issued.  Set it when a bio
932 	 * is being charged to a tg.
933 	 */
934 	if (!bio_flagged(bio, BIO_THROTTLED))
935 		bio_set_flag(bio, BIO_THROTTLED);
936 }
937 
938 /**
939  * throtl_add_bio_tg - add a bio to the specified throtl_grp
940  * @bio: bio to add
941  * @qn: qnode to use
942  * @tg: the target throtl_grp
943  *
944  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
945  * tg->qnode_on_self[] is used.
946  */
947 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
948 			      struct throtl_grp *tg)
949 {
950 	struct throtl_service_queue *sq = &tg->service_queue;
951 	bool rw = bio_data_dir(bio);
952 
953 	if (!qn)
954 		qn = &tg->qnode_on_self[rw];
955 
956 	/*
957 	 * If @tg doesn't currently have any bios queued in the same
958 	 * direction, queueing @bio can change when @tg should be
959 	 * dispatched.  Mark that @tg was empty.  This is automatically
960 	 * cleared on the next tg_update_disptime().
961 	 */
962 	if (!sq->nr_queued[rw])
963 		tg->flags |= THROTL_TG_WAS_EMPTY;
964 
965 	throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
966 
967 	sq->nr_queued[rw]++;
968 	throtl_enqueue_tg(tg);
969 }
970 
971 static void tg_update_disptime(struct throtl_grp *tg)
972 {
973 	struct throtl_service_queue *sq = &tg->service_queue;
974 	unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
975 	struct bio *bio;
976 
977 	bio = throtl_peek_queued(&sq->queued[READ]);
978 	if (bio)
979 		tg_may_dispatch(tg, bio, &read_wait);
980 
981 	bio = throtl_peek_queued(&sq->queued[WRITE]);
982 	if (bio)
983 		tg_may_dispatch(tg, bio, &write_wait);
984 
985 	min_wait = min(read_wait, write_wait);
986 	disptime = jiffies + min_wait;
987 
988 	/* Update dispatch time */
989 	throtl_dequeue_tg(tg);
990 	tg->disptime = disptime;
991 	throtl_enqueue_tg(tg);
992 
993 	/* see throtl_add_bio_tg() */
994 	tg->flags &= ~THROTL_TG_WAS_EMPTY;
995 }
996 
997 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
998 					struct throtl_grp *parent_tg, bool rw)
999 {
1000 	if (throtl_slice_used(parent_tg, rw)) {
1001 		throtl_start_new_slice_with_credit(parent_tg, rw,
1002 				child_tg->slice_start[rw]);
1003 	}
1004 
1005 }
1006 
1007 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1008 {
1009 	struct throtl_service_queue *sq = &tg->service_queue;
1010 	struct throtl_service_queue *parent_sq = sq->parent_sq;
1011 	struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1012 	struct throtl_grp *tg_to_put = NULL;
1013 	struct bio *bio;
1014 
1015 	/*
1016 	 * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1017 	 * from @tg may put its reference and @parent_sq might end up
1018 	 * getting released prematurely.  Remember the tg to put and put it
1019 	 * after @bio is transferred to @parent_sq.
1020 	 */
1021 	bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1022 	sq->nr_queued[rw]--;
1023 
1024 	throtl_charge_bio(tg, bio);
1025 
1026 	/*
1027 	 * If our parent is another tg, we just need to transfer @bio to
1028 	 * the parent using throtl_add_bio_tg().  If our parent is
1029 	 * @td->service_queue, @bio is ready to be issued.  Put it on its
1030 	 * bio_lists[] and decrease total number queued.  The caller is
1031 	 * responsible for issuing these bios.
1032 	 */
1033 	if (parent_tg) {
1034 		throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1035 		start_parent_slice_with_credit(tg, parent_tg, rw);
1036 	} else {
1037 		throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1038 				     &parent_sq->queued[rw]);
1039 		BUG_ON(tg->td->nr_queued[rw] <= 0);
1040 		tg->td->nr_queued[rw]--;
1041 	}
1042 
1043 	throtl_trim_slice(tg, rw);
1044 
1045 	if (tg_to_put)
1046 		blkg_put(tg_to_blkg(tg_to_put));
1047 }
1048 
1049 static int throtl_dispatch_tg(struct throtl_grp *tg)
1050 {
1051 	struct throtl_service_queue *sq = &tg->service_queue;
1052 	unsigned int nr_reads = 0, nr_writes = 0;
1053 	unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1054 	unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1055 	struct bio *bio;
1056 
1057 	/* Try to dispatch 75% READS and 25% WRITES */
1058 
1059 	while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1060 	       tg_may_dispatch(tg, bio, NULL)) {
1061 
1062 		tg_dispatch_one_bio(tg, bio_data_dir(bio));
1063 		nr_reads++;
1064 
1065 		if (nr_reads >= max_nr_reads)
1066 			break;
1067 	}
1068 
1069 	while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1070 	       tg_may_dispatch(tg, bio, NULL)) {
1071 
1072 		tg_dispatch_one_bio(tg, bio_data_dir(bio));
1073 		nr_writes++;
1074 
1075 		if (nr_writes >= max_nr_writes)
1076 			break;
1077 	}
1078 
1079 	return nr_reads + nr_writes;
1080 }
1081 
1082 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1083 {
1084 	unsigned int nr_disp = 0;
1085 
1086 	while (1) {
1087 		struct throtl_grp *tg;
1088 		struct throtl_service_queue *sq;
1089 
1090 		if (!parent_sq->nr_pending)
1091 			break;
1092 
1093 		tg = throtl_rb_first(parent_sq);
1094 		if (!tg)
1095 			break;
1096 
1097 		if (time_before(jiffies, tg->disptime))
1098 			break;
1099 
1100 		throtl_dequeue_tg(tg);
1101 
1102 		nr_disp += throtl_dispatch_tg(tg);
1103 
1104 		sq = &tg->service_queue;
1105 		if (sq->nr_queued[0] || sq->nr_queued[1])
1106 			tg_update_disptime(tg);
1107 
1108 		if (nr_disp >= THROTL_QUANTUM)
1109 			break;
1110 	}
1111 
1112 	return nr_disp;
1113 }
1114 
1115 static bool throtl_can_upgrade(struct throtl_data *td,
1116 	struct throtl_grp *this_tg);
1117 /**
1118  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1119  * @t: the pending_timer member of the throtl_service_queue being serviced
1120  *
1121  * This timer is armed when a child throtl_grp with active bio's become
1122  * pending and queued on the service_queue's pending_tree and expires when
1123  * the first child throtl_grp should be dispatched.  This function
1124  * dispatches bio's from the children throtl_grps to the parent
1125  * service_queue.
1126  *
1127  * If the parent's parent is another throtl_grp, dispatching is propagated
1128  * by either arming its pending_timer or repeating dispatch directly.  If
1129  * the top-level service_tree is reached, throtl_data->dispatch_work is
1130  * kicked so that the ready bio's are issued.
1131  */
1132 static void throtl_pending_timer_fn(struct timer_list *t)
1133 {
1134 	struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1135 	struct throtl_grp *tg = sq_to_tg(sq);
1136 	struct throtl_data *td = sq_to_td(sq);
1137 	struct request_queue *q = td->queue;
1138 	struct throtl_service_queue *parent_sq;
1139 	bool dispatched;
1140 	int ret;
1141 
1142 	spin_lock_irq(&q->queue_lock);
1143 	if (throtl_can_upgrade(td, NULL))
1144 		throtl_upgrade_state(td);
1145 
1146 again:
1147 	parent_sq = sq->parent_sq;
1148 	dispatched = false;
1149 
1150 	while (true) {
1151 		throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1152 			   sq->nr_queued[READ] + sq->nr_queued[WRITE],
1153 			   sq->nr_queued[READ], sq->nr_queued[WRITE]);
1154 
1155 		ret = throtl_select_dispatch(sq);
1156 		if (ret) {
1157 			throtl_log(sq, "bios disp=%u", ret);
1158 			dispatched = true;
1159 		}
1160 
1161 		if (throtl_schedule_next_dispatch(sq, false))
1162 			break;
1163 
1164 		/* this dispatch windows is still open, relax and repeat */
1165 		spin_unlock_irq(&q->queue_lock);
1166 		cpu_relax();
1167 		spin_lock_irq(&q->queue_lock);
1168 	}
1169 
1170 	if (!dispatched)
1171 		goto out_unlock;
1172 
1173 	if (parent_sq) {
1174 		/* @parent_sq is another throl_grp, propagate dispatch */
1175 		if (tg->flags & THROTL_TG_WAS_EMPTY) {
1176 			tg_update_disptime(tg);
1177 			if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1178 				/* window is already open, repeat dispatching */
1179 				sq = parent_sq;
1180 				tg = sq_to_tg(sq);
1181 				goto again;
1182 			}
1183 		}
1184 	} else {
1185 		/* reached the top-level, queue issuing */
1186 		queue_work(kthrotld_workqueue, &td->dispatch_work);
1187 	}
1188 out_unlock:
1189 	spin_unlock_irq(&q->queue_lock);
1190 }
1191 
1192 /**
1193  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1194  * @work: work item being executed
1195  *
1196  * This function is queued for execution when bios reach the bio_lists[]
1197  * of throtl_data->service_queue.  Those bios are ready and issued by this
1198  * function.
1199  */
1200 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1201 {
1202 	struct throtl_data *td = container_of(work, struct throtl_data,
1203 					      dispatch_work);
1204 	struct throtl_service_queue *td_sq = &td->service_queue;
1205 	struct request_queue *q = td->queue;
1206 	struct bio_list bio_list_on_stack;
1207 	struct bio *bio;
1208 	struct blk_plug plug;
1209 	int rw;
1210 
1211 	bio_list_init(&bio_list_on_stack);
1212 
1213 	spin_lock_irq(&q->queue_lock);
1214 	for (rw = READ; rw <= WRITE; rw++)
1215 		while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1216 			bio_list_add(&bio_list_on_stack, bio);
1217 	spin_unlock_irq(&q->queue_lock);
1218 
1219 	if (!bio_list_empty(&bio_list_on_stack)) {
1220 		blk_start_plug(&plug);
1221 		while ((bio = bio_list_pop(&bio_list_on_stack)))
1222 			submit_bio_noacct(bio);
1223 		blk_finish_plug(&plug);
1224 	}
1225 }
1226 
1227 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1228 			      int off)
1229 {
1230 	struct throtl_grp *tg = pd_to_tg(pd);
1231 	u64 v = *(u64 *)((void *)tg + off);
1232 
1233 	if (v == U64_MAX)
1234 		return 0;
1235 	return __blkg_prfill_u64(sf, pd, v);
1236 }
1237 
1238 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1239 			       int off)
1240 {
1241 	struct throtl_grp *tg = pd_to_tg(pd);
1242 	unsigned int v = *(unsigned int *)((void *)tg + off);
1243 
1244 	if (v == UINT_MAX)
1245 		return 0;
1246 	return __blkg_prfill_u64(sf, pd, v);
1247 }
1248 
1249 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1250 {
1251 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1252 			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1253 	return 0;
1254 }
1255 
1256 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1257 {
1258 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1259 			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1260 	return 0;
1261 }
1262 
1263 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1264 {
1265 	struct throtl_service_queue *sq = &tg->service_queue;
1266 	struct cgroup_subsys_state *pos_css;
1267 	struct blkcg_gq *blkg;
1268 
1269 	throtl_log(&tg->service_queue,
1270 		   "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1271 		   tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1272 		   tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1273 
1274 	/*
1275 	 * Update has_rules[] flags for the updated tg's subtree.  A tg is
1276 	 * considered to have rules if either the tg itself or any of its
1277 	 * ancestors has rules.  This identifies groups without any
1278 	 * restrictions in the whole hierarchy and allows them to bypass
1279 	 * blk-throttle.
1280 	 */
1281 	blkg_for_each_descendant_pre(blkg, pos_css,
1282 			global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1283 		struct throtl_grp *this_tg = blkg_to_tg(blkg);
1284 		struct throtl_grp *parent_tg;
1285 
1286 		tg_update_has_rules(this_tg);
1287 		/* ignore root/second level */
1288 		if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1289 		    !blkg->parent->parent)
1290 			continue;
1291 		parent_tg = blkg_to_tg(blkg->parent);
1292 		/*
1293 		 * make sure all children has lower idle time threshold and
1294 		 * higher latency target
1295 		 */
1296 		this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1297 				parent_tg->idletime_threshold);
1298 		this_tg->latency_target = max(this_tg->latency_target,
1299 				parent_tg->latency_target);
1300 	}
1301 
1302 	/*
1303 	 * We're already holding queue_lock and know @tg is valid.  Let's
1304 	 * apply the new config directly.
1305 	 *
1306 	 * Restart the slices for both READ and WRITES. It might happen
1307 	 * that a group's limit are dropped suddenly and we don't want to
1308 	 * account recently dispatched IO with new low rate.
1309 	 */
1310 	throtl_start_new_slice(tg, READ);
1311 	throtl_start_new_slice(tg, WRITE);
1312 
1313 	if (tg->flags & THROTL_TG_PENDING) {
1314 		tg_update_disptime(tg);
1315 		throtl_schedule_next_dispatch(sq->parent_sq, true);
1316 	}
1317 }
1318 
1319 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1320 			   char *buf, size_t nbytes, loff_t off, bool is_u64)
1321 {
1322 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1323 	struct blkg_conf_ctx ctx;
1324 	struct throtl_grp *tg;
1325 	int ret;
1326 	u64 v;
1327 
1328 	ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1329 	if (ret)
1330 		return ret;
1331 
1332 	ret = -EINVAL;
1333 	if (sscanf(ctx.body, "%llu", &v) != 1)
1334 		goto out_finish;
1335 	if (!v)
1336 		v = U64_MAX;
1337 
1338 	tg = blkg_to_tg(ctx.blkg);
1339 
1340 	if (is_u64)
1341 		*(u64 *)((void *)tg + of_cft(of)->private) = v;
1342 	else
1343 		*(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1344 
1345 	tg_conf_updated(tg, false);
1346 	ret = 0;
1347 out_finish:
1348 	blkg_conf_finish(&ctx);
1349 	return ret ?: nbytes;
1350 }
1351 
1352 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1353 			       char *buf, size_t nbytes, loff_t off)
1354 {
1355 	return tg_set_conf(of, buf, nbytes, off, true);
1356 }
1357 
1358 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1359 				char *buf, size_t nbytes, loff_t off)
1360 {
1361 	return tg_set_conf(of, buf, nbytes, off, false);
1362 }
1363 
1364 static int tg_print_rwstat(struct seq_file *sf, void *v)
1365 {
1366 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1367 			  blkg_prfill_rwstat, &blkcg_policy_throtl,
1368 			  seq_cft(sf)->private, true);
1369 	return 0;
1370 }
1371 
1372 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1373 				      struct blkg_policy_data *pd, int off)
1374 {
1375 	struct blkg_rwstat_sample sum;
1376 
1377 	blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1378 				  &sum);
1379 	return __blkg_prfill_rwstat(sf, pd, &sum);
1380 }
1381 
1382 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1383 {
1384 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1385 			  tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1386 			  seq_cft(sf)->private, true);
1387 	return 0;
1388 }
1389 
1390 static struct cftype throtl_legacy_files[] = {
1391 	{
1392 		.name = "throttle.read_bps_device",
1393 		.private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1394 		.seq_show = tg_print_conf_u64,
1395 		.write = tg_set_conf_u64,
1396 	},
1397 	{
1398 		.name = "throttle.write_bps_device",
1399 		.private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1400 		.seq_show = tg_print_conf_u64,
1401 		.write = tg_set_conf_u64,
1402 	},
1403 	{
1404 		.name = "throttle.read_iops_device",
1405 		.private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1406 		.seq_show = tg_print_conf_uint,
1407 		.write = tg_set_conf_uint,
1408 	},
1409 	{
1410 		.name = "throttle.write_iops_device",
1411 		.private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1412 		.seq_show = tg_print_conf_uint,
1413 		.write = tg_set_conf_uint,
1414 	},
1415 	{
1416 		.name = "throttle.io_service_bytes",
1417 		.private = offsetof(struct throtl_grp, stat_bytes),
1418 		.seq_show = tg_print_rwstat,
1419 	},
1420 	{
1421 		.name = "throttle.io_service_bytes_recursive",
1422 		.private = offsetof(struct throtl_grp, stat_bytes),
1423 		.seq_show = tg_print_rwstat_recursive,
1424 	},
1425 	{
1426 		.name = "throttle.io_serviced",
1427 		.private = offsetof(struct throtl_grp, stat_ios),
1428 		.seq_show = tg_print_rwstat,
1429 	},
1430 	{
1431 		.name = "throttle.io_serviced_recursive",
1432 		.private = offsetof(struct throtl_grp, stat_ios),
1433 		.seq_show = tg_print_rwstat_recursive,
1434 	},
1435 	{ }	/* terminate */
1436 };
1437 
1438 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1439 			 int off)
1440 {
1441 	struct throtl_grp *tg = pd_to_tg(pd);
1442 	const char *dname = blkg_dev_name(pd->blkg);
1443 	char bufs[4][21] = { "max", "max", "max", "max" };
1444 	u64 bps_dft;
1445 	unsigned int iops_dft;
1446 	char idle_time[26] = "";
1447 	char latency_time[26] = "";
1448 
1449 	if (!dname)
1450 		return 0;
1451 
1452 	if (off == LIMIT_LOW) {
1453 		bps_dft = 0;
1454 		iops_dft = 0;
1455 	} else {
1456 		bps_dft = U64_MAX;
1457 		iops_dft = UINT_MAX;
1458 	}
1459 
1460 	if (tg->bps_conf[READ][off] == bps_dft &&
1461 	    tg->bps_conf[WRITE][off] == bps_dft &&
1462 	    tg->iops_conf[READ][off] == iops_dft &&
1463 	    tg->iops_conf[WRITE][off] == iops_dft &&
1464 	    (off != LIMIT_LOW ||
1465 	     (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1466 	      tg->latency_target_conf == DFL_LATENCY_TARGET)))
1467 		return 0;
1468 
1469 	if (tg->bps_conf[READ][off] != U64_MAX)
1470 		snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1471 			tg->bps_conf[READ][off]);
1472 	if (tg->bps_conf[WRITE][off] != U64_MAX)
1473 		snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1474 			tg->bps_conf[WRITE][off]);
1475 	if (tg->iops_conf[READ][off] != UINT_MAX)
1476 		snprintf(bufs[2], sizeof(bufs[2]), "%u",
1477 			tg->iops_conf[READ][off]);
1478 	if (tg->iops_conf[WRITE][off] != UINT_MAX)
1479 		snprintf(bufs[3], sizeof(bufs[3]), "%u",
1480 			tg->iops_conf[WRITE][off]);
1481 	if (off == LIMIT_LOW) {
1482 		if (tg->idletime_threshold_conf == ULONG_MAX)
1483 			strcpy(idle_time, " idle=max");
1484 		else
1485 			snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1486 				tg->idletime_threshold_conf);
1487 
1488 		if (tg->latency_target_conf == ULONG_MAX)
1489 			strcpy(latency_time, " latency=max");
1490 		else
1491 			snprintf(latency_time, sizeof(latency_time),
1492 				" latency=%lu", tg->latency_target_conf);
1493 	}
1494 
1495 	seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1496 		   dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1497 		   latency_time);
1498 	return 0;
1499 }
1500 
1501 static int tg_print_limit(struct seq_file *sf, void *v)
1502 {
1503 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1504 			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1505 	return 0;
1506 }
1507 
1508 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1509 			  char *buf, size_t nbytes, loff_t off)
1510 {
1511 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1512 	struct blkg_conf_ctx ctx;
1513 	struct throtl_grp *tg;
1514 	u64 v[4];
1515 	unsigned long idle_time;
1516 	unsigned long latency_time;
1517 	int ret;
1518 	int index = of_cft(of)->private;
1519 
1520 	ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1521 	if (ret)
1522 		return ret;
1523 
1524 	tg = blkg_to_tg(ctx.blkg);
1525 
1526 	v[0] = tg->bps_conf[READ][index];
1527 	v[1] = tg->bps_conf[WRITE][index];
1528 	v[2] = tg->iops_conf[READ][index];
1529 	v[3] = tg->iops_conf[WRITE][index];
1530 
1531 	idle_time = tg->idletime_threshold_conf;
1532 	latency_time = tg->latency_target_conf;
1533 	while (true) {
1534 		char tok[27];	/* wiops=18446744073709551616 */
1535 		char *p;
1536 		u64 val = U64_MAX;
1537 		int len;
1538 
1539 		if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1540 			break;
1541 		if (tok[0] == '\0')
1542 			break;
1543 		ctx.body += len;
1544 
1545 		ret = -EINVAL;
1546 		p = tok;
1547 		strsep(&p, "=");
1548 		if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1549 			goto out_finish;
1550 
1551 		ret = -ERANGE;
1552 		if (!val)
1553 			goto out_finish;
1554 
1555 		ret = -EINVAL;
1556 		if (!strcmp(tok, "rbps") && val > 1)
1557 			v[0] = val;
1558 		else if (!strcmp(tok, "wbps") && val > 1)
1559 			v[1] = val;
1560 		else if (!strcmp(tok, "riops") && val > 1)
1561 			v[2] = min_t(u64, val, UINT_MAX);
1562 		else if (!strcmp(tok, "wiops") && val > 1)
1563 			v[3] = min_t(u64, val, UINT_MAX);
1564 		else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1565 			idle_time = val;
1566 		else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1567 			latency_time = val;
1568 		else
1569 			goto out_finish;
1570 	}
1571 
1572 	tg->bps_conf[READ][index] = v[0];
1573 	tg->bps_conf[WRITE][index] = v[1];
1574 	tg->iops_conf[READ][index] = v[2];
1575 	tg->iops_conf[WRITE][index] = v[3];
1576 
1577 	if (index == LIMIT_MAX) {
1578 		tg->bps[READ][index] = v[0];
1579 		tg->bps[WRITE][index] = v[1];
1580 		tg->iops[READ][index] = v[2];
1581 		tg->iops[WRITE][index] = v[3];
1582 	}
1583 	tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1584 		tg->bps_conf[READ][LIMIT_MAX]);
1585 	tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1586 		tg->bps_conf[WRITE][LIMIT_MAX]);
1587 	tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1588 		tg->iops_conf[READ][LIMIT_MAX]);
1589 	tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1590 		tg->iops_conf[WRITE][LIMIT_MAX]);
1591 	tg->idletime_threshold_conf = idle_time;
1592 	tg->latency_target_conf = latency_time;
1593 
1594 	/* force user to configure all settings for low limit  */
1595 	if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1596 	      tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1597 	    tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1598 	    tg->latency_target_conf == DFL_LATENCY_TARGET) {
1599 		tg->bps[READ][LIMIT_LOW] = 0;
1600 		tg->bps[WRITE][LIMIT_LOW] = 0;
1601 		tg->iops[READ][LIMIT_LOW] = 0;
1602 		tg->iops[WRITE][LIMIT_LOW] = 0;
1603 		tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1604 		tg->latency_target = DFL_LATENCY_TARGET;
1605 	} else if (index == LIMIT_LOW) {
1606 		tg->idletime_threshold = tg->idletime_threshold_conf;
1607 		tg->latency_target = tg->latency_target_conf;
1608 	}
1609 
1610 	blk_throtl_update_limit_valid(tg->td);
1611 	if (tg->td->limit_valid[LIMIT_LOW]) {
1612 		if (index == LIMIT_LOW)
1613 			tg->td->limit_index = LIMIT_LOW;
1614 	} else
1615 		tg->td->limit_index = LIMIT_MAX;
1616 	tg_conf_updated(tg, index == LIMIT_LOW &&
1617 		tg->td->limit_valid[LIMIT_LOW]);
1618 	ret = 0;
1619 out_finish:
1620 	blkg_conf_finish(&ctx);
1621 	return ret ?: nbytes;
1622 }
1623 
1624 static struct cftype throtl_files[] = {
1625 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1626 	{
1627 		.name = "low",
1628 		.flags = CFTYPE_NOT_ON_ROOT,
1629 		.seq_show = tg_print_limit,
1630 		.write = tg_set_limit,
1631 		.private = LIMIT_LOW,
1632 	},
1633 #endif
1634 	{
1635 		.name = "max",
1636 		.flags = CFTYPE_NOT_ON_ROOT,
1637 		.seq_show = tg_print_limit,
1638 		.write = tg_set_limit,
1639 		.private = LIMIT_MAX,
1640 	},
1641 	{ }	/* terminate */
1642 };
1643 
1644 static void throtl_shutdown_wq(struct request_queue *q)
1645 {
1646 	struct throtl_data *td = q->td;
1647 
1648 	cancel_work_sync(&td->dispatch_work);
1649 }
1650 
1651 struct blkcg_policy blkcg_policy_throtl = {
1652 	.dfl_cftypes		= throtl_files,
1653 	.legacy_cftypes		= throtl_legacy_files,
1654 
1655 	.pd_alloc_fn		= throtl_pd_alloc,
1656 	.pd_init_fn		= throtl_pd_init,
1657 	.pd_online_fn		= throtl_pd_online,
1658 	.pd_offline_fn		= throtl_pd_offline,
1659 	.pd_free_fn		= throtl_pd_free,
1660 };
1661 
1662 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1663 {
1664 	unsigned long rtime = jiffies, wtime = jiffies;
1665 
1666 	if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1667 		rtime = tg->last_low_overflow_time[READ];
1668 	if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1669 		wtime = tg->last_low_overflow_time[WRITE];
1670 	return min(rtime, wtime);
1671 }
1672 
1673 /* tg should not be an intermediate node */
1674 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1675 {
1676 	struct throtl_service_queue *parent_sq;
1677 	struct throtl_grp *parent = tg;
1678 	unsigned long ret = __tg_last_low_overflow_time(tg);
1679 
1680 	while (true) {
1681 		parent_sq = parent->service_queue.parent_sq;
1682 		parent = sq_to_tg(parent_sq);
1683 		if (!parent)
1684 			break;
1685 
1686 		/*
1687 		 * The parent doesn't have low limit, it always reaches low
1688 		 * limit. Its overflow time is useless for children
1689 		 */
1690 		if (!parent->bps[READ][LIMIT_LOW] &&
1691 		    !parent->iops[READ][LIMIT_LOW] &&
1692 		    !parent->bps[WRITE][LIMIT_LOW] &&
1693 		    !parent->iops[WRITE][LIMIT_LOW])
1694 			continue;
1695 		if (time_after(__tg_last_low_overflow_time(parent), ret))
1696 			ret = __tg_last_low_overflow_time(parent);
1697 	}
1698 	return ret;
1699 }
1700 
1701 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1702 {
1703 	/*
1704 	 * cgroup is idle if:
1705 	 * - single idle is too long, longer than a fixed value (in case user
1706 	 *   configure a too big threshold) or 4 times of idletime threshold
1707 	 * - average think time is more than threshold
1708 	 * - IO latency is largely below threshold
1709 	 */
1710 	unsigned long time;
1711 	bool ret;
1712 
1713 	time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1714 	ret = tg->latency_target == DFL_LATENCY_TARGET ||
1715 	      tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1716 	      (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1717 	      tg->avg_idletime > tg->idletime_threshold ||
1718 	      (tg->latency_target && tg->bio_cnt &&
1719 		tg->bad_bio_cnt * 5 < tg->bio_cnt);
1720 	throtl_log(&tg->service_queue,
1721 		"avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1722 		tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1723 		tg->bio_cnt, ret, tg->td->scale);
1724 	return ret;
1725 }
1726 
1727 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1728 {
1729 	struct throtl_service_queue *sq = &tg->service_queue;
1730 	bool read_limit, write_limit;
1731 
1732 	/*
1733 	 * if cgroup reaches low limit (if low limit is 0, the cgroup always
1734 	 * reaches), it's ok to upgrade to next limit
1735 	 */
1736 	read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1737 	write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1738 	if (!read_limit && !write_limit)
1739 		return true;
1740 	if (read_limit && sq->nr_queued[READ] &&
1741 	    (!write_limit || sq->nr_queued[WRITE]))
1742 		return true;
1743 	if (write_limit && sq->nr_queued[WRITE] &&
1744 	    (!read_limit || sq->nr_queued[READ]))
1745 		return true;
1746 
1747 	if (time_after_eq(jiffies,
1748 		tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1749 	    throtl_tg_is_idle(tg))
1750 		return true;
1751 	return false;
1752 }
1753 
1754 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1755 {
1756 	while (true) {
1757 		if (throtl_tg_can_upgrade(tg))
1758 			return true;
1759 		tg = sq_to_tg(tg->service_queue.parent_sq);
1760 		if (!tg || !tg_to_blkg(tg)->parent)
1761 			return false;
1762 	}
1763 	return false;
1764 }
1765 
1766 static bool throtl_can_upgrade(struct throtl_data *td,
1767 	struct throtl_grp *this_tg)
1768 {
1769 	struct cgroup_subsys_state *pos_css;
1770 	struct blkcg_gq *blkg;
1771 
1772 	if (td->limit_index != LIMIT_LOW)
1773 		return false;
1774 
1775 	if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1776 		return false;
1777 
1778 	rcu_read_lock();
1779 	blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1780 		struct throtl_grp *tg = blkg_to_tg(blkg);
1781 
1782 		if (tg == this_tg)
1783 			continue;
1784 		if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1785 			continue;
1786 		if (!throtl_hierarchy_can_upgrade(tg)) {
1787 			rcu_read_unlock();
1788 			return false;
1789 		}
1790 	}
1791 	rcu_read_unlock();
1792 	return true;
1793 }
1794 
1795 static void throtl_upgrade_check(struct throtl_grp *tg)
1796 {
1797 	unsigned long now = jiffies;
1798 
1799 	if (tg->td->limit_index != LIMIT_LOW)
1800 		return;
1801 
1802 	if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1803 		return;
1804 
1805 	tg->last_check_time = now;
1806 
1807 	if (!time_after_eq(now,
1808 	     __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1809 		return;
1810 
1811 	if (throtl_can_upgrade(tg->td, NULL))
1812 		throtl_upgrade_state(tg->td);
1813 }
1814 
1815 static void throtl_upgrade_state(struct throtl_data *td)
1816 {
1817 	struct cgroup_subsys_state *pos_css;
1818 	struct blkcg_gq *blkg;
1819 
1820 	throtl_log(&td->service_queue, "upgrade to max");
1821 	td->limit_index = LIMIT_MAX;
1822 	td->low_upgrade_time = jiffies;
1823 	td->scale = 0;
1824 	rcu_read_lock();
1825 	blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1826 		struct throtl_grp *tg = blkg_to_tg(blkg);
1827 		struct throtl_service_queue *sq = &tg->service_queue;
1828 
1829 		tg->disptime = jiffies - 1;
1830 		throtl_select_dispatch(sq);
1831 		throtl_schedule_next_dispatch(sq, true);
1832 	}
1833 	rcu_read_unlock();
1834 	throtl_select_dispatch(&td->service_queue);
1835 	throtl_schedule_next_dispatch(&td->service_queue, true);
1836 	queue_work(kthrotld_workqueue, &td->dispatch_work);
1837 }
1838 
1839 static void throtl_downgrade_state(struct throtl_data *td)
1840 {
1841 	td->scale /= 2;
1842 
1843 	throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1844 	if (td->scale) {
1845 		td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1846 		return;
1847 	}
1848 
1849 	td->limit_index = LIMIT_LOW;
1850 	td->low_downgrade_time = jiffies;
1851 }
1852 
1853 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1854 {
1855 	struct throtl_data *td = tg->td;
1856 	unsigned long now = jiffies;
1857 
1858 	/*
1859 	 * If cgroup is below low limit, consider downgrade and throttle other
1860 	 * cgroups
1861 	 */
1862 	if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1863 	    time_after_eq(now, tg_last_low_overflow_time(tg) +
1864 					td->throtl_slice) &&
1865 	    (!throtl_tg_is_idle(tg) ||
1866 	     !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1867 		return true;
1868 	return false;
1869 }
1870 
1871 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1872 {
1873 	while (true) {
1874 		if (!throtl_tg_can_downgrade(tg))
1875 			return false;
1876 		tg = sq_to_tg(tg->service_queue.parent_sq);
1877 		if (!tg || !tg_to_blkg(tg)->parent)
1878 			break;
1879 	}
1880 	return true;
1881 }
1882 
1883 static void throtl_downgrade_check(struct throtl_grp *tg)
1884 {
1885 	uint64_t bps;
1886 	unsigned int iops;
1887 	unsigned long elapsed_time;
1888 	unsigned long now = jiffies;
1889 
1890 	if (tg->td->limit_index != LIMIT_MAX ||
1891 	    !tg->td->limit_valid[LIMIT_LOW])
1892 		return;
1893 	if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1894 		return;
1895 	if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1896 		return;
1897 
1898 	elapsed_time = now - tg->last_check_time;
1899 	tg->last_check_time = now;
1900 
1901 	if (time_before(now, tg_last_low_overflow_time(tg) +
1902 			tg->td->throtl_slice))
1903 		return;
1904 
1905 	if (tg->bps[READ][LIMIT_LOW]) {
1906 		bps = tg->last_bytes_disp[READ] * HZ;
1907 		do_div(bps, elapsed_time);
1908 		if (bps >= tg->bps[READ][LIMIT_LOW])
1909 			tg->last_low_overflow_time[READ] = now;
1910 	}
1911 
1912 	if (tg->bps[WRITE][LIMIT_LOW]) {
1913 		bps = tg->last_bytes_disp[WRITE] * HZ;
1914 		do_div(bps, elapsed_time);
1915 		if (bps >= tg->bps[WRITE][LIMIT_LOW])
1916 			tg->last_low_overflow_time[WRITE] = now;
1917 	}
1918 
1919 	if (tg->iops[READ][LIMIT_LOW]) {
1920 		tg->last_io_disp[READ] += atomic_xchg(&tg->last_io_split_cnt[READ], 0);
1921 		iops = tg->last_io_disp[READ] * HZ / elapsed_time;
1922 		if (iops >= tg->iops[READ][LIMIT_LOW])
1923 			tg->last_low_overflow_time[READ] = now;
1924 	}
1925 
1926 	if (tg->iops[WRITE][LIMIT_LOW]) {
1927 		tg->last_io_disp[WRITE] += atomic_xchg(&tg->last_io_split_cnt[WRITE], 0);
1928 		iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
1929 		if (iops >= tg->iops[WRITE][LIMIT_LOW])
1930 			tg->last_low_overflow_time[WRITE] = now;
1931 	}
1932 
1933 	/*
1934 	 * If cgroup is below low limit, consider downgrade and throttle other
1935 	 * cgroups
1936 	 */
1937 	if (throtl_hierarchy_can_downgrade(tg))
1938 		throtl_downgrade_state(tg->td);
1939 
1940 	tg->last_bytes_disp[READ] = 0;
1941 	tg->last_bytes_disp[WRITE] = 0;
1942 	tg->last_io_disp[READ] = 0;
1943 	tg->last_io_disp[WRITE] = 0;
1944 }
1945 
1946 static void blk_throtl_update_idletime(struct throtl_grp *tg)
1947 {
1948 	unsigned long now;
1949 	unsigned long last_finish_time = tg->last_finish_time;
1950 
1951 	if (last_finish_time == 0)
1952 		return;
1953 
1954 	now = ktime_get_ns() >> 10;
1955 	if (now <= last_finish_time ||
1956 	    last_finish_time == tg->checked_last_finish_time)
1957 		return;
1958 
1959 	tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
1960 	tg->checked_last_finish_time = last_finish_time;
1961 }
1962 
1963 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1964 static void throtl_update_latency_buckets(struct throtl_data *td)
1965 {
1966 	struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
1967 	int i, cpu, rw;
1968 	unsigned long last_latency[2] = { 0 };
1969 	unsigned long latency[2];
1970 
1971 	if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
1972 		return;
1973 	if (time_before(jiffies, td->last_calculate_time + HZ))
1974 		return;
1975 	td->last_calculate_time = jiffies;
1976 
1977 	memset(avg_latency, 0, sizeof(avg_latency));
1978 	for (rw = READ; rw <= WRITE; rw++) {
1979 		for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
1980 			struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
1981 
1982 			for_each_possible_cpu(cpu) {
1983 				struct latency_bucket *bucket;
1984 
1985 				/* this isn't race free, but ok in practice */
1986 				bucket = per_cpu_ptr(td->latency_buckets[rw],
1987 					cpu);
1988 				tmp->total_latency += bucket[i].total_latency;
1989 				tmp->samples += bucket[i].samples;
1990 				bucket[i].total_latency = 0;
1991 				bucket[i].samples = 0;
1992 			}
1993 
1994 			if (tmp->samples >= 32) {
1995 				int samples = tmp->samples;
1996 
1997 				latency[rw] = tmp->total_latency;
1998 
1999 				tmp->total_latency = 0;
2000 				tmp->samples = 0;
2001 				latency[rw] /= samples;
2002 				if (latency[rw] == 0)
2003 					continue;
2004 				avg_latency[rw][i].latency = latency[rw];
2005 			}
2006 		}
2007 	}
2008 
2009 	for (rw = READ; rw <= WRITE; rw++) {
2010 		for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2011 			if (!avg_latency[rw][i].latency) {
2012 				if (td->avg_buckets[rw][i].latency < last_latency[rw])
2013 					td->avg_buckets[rw][i].latency =
2014 						last_latency[rw];
2015 				continue;
2016 			}
2017 
2018 			if (!td->avg_buckets[rw][i].valid)
2019 				latency[rw] = avg_latency[rw][i].latency;
2020 			else
2021 				latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2022 					avg_latency[rw][i].latency) >> 3;
2023 
2024 			td->avg_buckets[rw][i].latency = max(latency[rw],
2025 				last_latency[rw]);
2026 			td->avg_buckets[rw][i].valid = true;
2027 			last_latency[rw] = td->avg_buckets[rw][i].latency;
2028 		}
2029 	}
2030 
2031 	for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2032 		throtl_log(&td->service_queue,
2033 			"Latency bucket %d: read latency=%ld, read valid=%d, "
2034 			"write latency=%ld, write valid=%d", i,
2035 			td->avg_buckets[READ][i].latency,
2036 			td->avg_buckets[READ][i].valid,
2037 			td->avg_buckets[WRITE][i].latency,
2038 			td->avg_buckets[WRITE][i].valid);
2039 }
2040 #else
2041 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2042 {
2043 }
2044 #endif
2045 
2046 void blk_throtl_charge_bio_split(struct bio *bio)
2047 {
2048 	struct blkcg_gq *blkg = bio->bi_blkg;
2049 	struct throtl_grp *parent = blkg_to_tg(blkg);
2050 	struct throtl_service_queue *parent_sq;
2051 	bool rw = bio_data_dir(bio);
2052 
2053 	do {
2054 		if (!parent->has_rules[rw])
2055 			break;
2056 
2057 		atomic_inc(&parent->io_split_cnt[rw]);
2058 		atomic_inc(&parent->last_io_split_cnt[rw]);
2059 
2060 		parent_sq = parent->service_queue.parent_sq;
2061 		parent = sq_to_tg(parent_sq);
2062 	} while (parent);
2063 }
2064 
2065 bool __blk_throtl_bio(struct bio *bio)
2066 {
2067 	struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2068 	struct blkcg_gq *blkg = bio->bi_blkg;
2069 	struct throtl_qnode *qn = NULL;
2070 	struct throtl_grp *tg = blkg_to_tg(blkg);
2071 	struct throtl_service_queue *sq;
2072 	bool rw = bio_data_dir(bio);
2073 	bool throttled = false;
2074 	struct throtl_data *td = tg->td;
2075 
2076 	rcu_read_lock();
2077 
2078 	if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
2079 		blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
2080 				bio->bi_iter.bi_size);
2081 		blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
2082 	}
2083 
2084 	spin_lock_irq(&q->queue_lock);
2085 
2086 	throtl_update_latency_buckets(td);
2087 
2088 	blk_throtl_update_idletime(tg);
2089 
2090 	sq = &tg->service_queue;
2091 
2092 again:
2093 	while (true) {
2094 		if (tg->last_low_overflow_time[rw] == 0)
2095 			tg->last_low_overflow_time[rw] = jiffies;
2096 		throtl_downgrade_check(tg);
2097 		throtl_upgrade_check(tg);
2098 		/* throtl is FIFO - if bios are already queued, should queue */
2099 		if (sq->nr_queued[rw])
2100 			break;
2101 
2102 		/* if above limits, break to queue */
2103 		if (!tg_may_dispatch(tg, bio, NULL)) {
2104 			tg->last_low_overflow_time[rw] = jiffies;
2105 			if (throtl_can_upgrade(td, tg)) {
2106 				throtl_upgrade_state(td);
2107 				goto again;
2108 			}
2109 			break;
2110 		}
2111 
2112 		/* within limits, let's charge and dispatch directly */
2113 		throtl_charge_bio(tg, bio);
2114 
2115 		/*
2116 		 * We need to trim slice even when bios are not being queued
2117 		 * otherwise it might happen that a bio is not queued for
2118 		 * a long time and slice keeps on extending and trim is not
2119 		 * called for a long time. Now if limits are reduced suddenly
2120 		 * we take into account all the IO dispatched so far at new
2121 		 * low rate and * newly queued IO gets a really long dispatch
2122 		 * time.
2123 		 *
2124 		 * So keep on trimming slice even if bio is not queued.
2125 		 */
2126 		throtl_trim_slice(tg, rw);
2127 
2128 		/*
2129 		 * @bio passed through this layer without being throttled.
2130 		 * Climb up the ladder.  If we're already at the top, it
2131 		 * can be executed directly.
2132 		 */
2133 		qn = &tg->qnode_on_parent[rw];
2134 		sq = sq->parent_sq;
2135 		tg = sq_to_tg(sq);
2136 		if (!tg)
2137 			goto out_unlock;
2138 	}
2139 
2140 	/* out-of-limit, queue to @tg */
2141 	throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2142 		   rw == READ ? 'R' : 'W',
2143 		   tg->bytes_disp[rw], bio->bi_iter.bi_size,
2144 		   tg_bps_limit(tg, rw),
2145 		   tg->io_disp[rw], tg_iops_limit(tg, rw),
2146 		   sq->nr_queued[READ], sq->nr_queued[WRITE]);
2147 
2148 	tg->last_low_overflow_time[rw] = jiffies;
2149 
2150 	td->nr_queued[rw]++;
2151 	throtl_add_bio_tg(bio, qn, tg);
2152 	throttled = true;
2153 
2154 	/*
2155 	 * Update @tg's dispatch time and force schedule dispatch if @tg
2156 	 * was empty before @bio.  The forced scheduling isn't likely to
2157 	 * cause undue delay as @bio is likely to be dispatched directly if
2158 	 * its @tg's disptime is not in the future.
2159 	 */
2160 	if (tg->flags & THROTL_TG_WAS_EMPTY) {
2161 		tg_update_disptime(tg);
2162 		throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2163 	}
2164 
2165 out_unlock:
2166 	spin_unlock_irq(&q->queue_lock);
2167 	bio_set_flag(bio, BIO_THROTTLED);
2168 
2169 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2170 	if (throttled || !td->track_bio_latency)
2171 		bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2172 #endif
2173 	rcu_read_unlock();
2174 	return throttled;
2175 }
2176 
2177 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2178 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2179 	int op, unsigned long time)
2180 {
2181 	struct latency_bucket *latency;
2182 	int index;
2183 
2184 	if (!td || td->limit_index != LIMIT_LOW ||
2185 	    !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2186 	    !blk_queue_nonrot(td->queue))
2187 		return;
2188 
2189 	index = request_bucket_index(size);
2190 
2191 	latency = get_cpu_ptr(td->latency_buckets[op]);
2192 	latency[index].total_latency += time;
2193 	latency[index].samples++;
2194 	put_cpu_ptr(td->latency_buckets[op]);
2195 }
2196 
2197 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2198 {
2199 	struct request_queue *q = rq->q;
2200 	struct throtl_data *td = q->td;
2201 
2202 	throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2203 			     time_ns >> 10);
2204 }
2205 
2206 void blk_throtl_bio_endio(struct bio *bio)
2207 {
2208 	struct blkcg_gq *blkg;
2209 	struct throtl_grp *tg;
2210 	u64 finish_time_ns;
2211 	unsigned long finish_time;
2212 	unsigned long start_time;
2213 	unsigned long lat;
2214 	int rw = bio_data_dir(bio);
2215 
2216 	blkg = bio->bi_blkg;
2217 	if (!blkg)
2218 		return;
2219 	tg = blkg_to_tg(blkg);
2220 	if (!tg->td->limit_valid[LIMIT_LOW])
2221 		return;
2222 
2223 	finish_time_ns = ktime_get_ns();
2224 	tg->last_finish_time = finish_time_ns >> 10;
2225 
2226 	start_time = bio_issue_time(&bio->bi_issue) >> 10;
2227 	finish_time = __bio_issue_time(finish_time_ns) >> 10;
2228 	if (!start_time || finish_time <= start_time)
2229 		return;
2230 
2231 	lat = finish_time - start_time;
2232 	/* this is only for bio based driver */
2233 	if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2234 		throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2235 				     bio_op(bio), lat);
2236 
2237 	if (tg->latency_target && lat >= tg->td->filtered_latency) {
2238 		int bucket;
2239 		unsigned int threshold;
2240 
2241 		bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2242 		threshold = tg->td->avg_buckets[rw][bucket].latency +
2243 			tg->latency_target;
2244 		if (lat > threshold)
2245 			tg->bad_bio_cnt++;
2246 		/*
2247 		 * Not race free, could get wrong count, which means cgroups
2248 		 * will be throttled
2249 		 */
2250 		tg->bio_cnt++;
2251 	}
2252 
2253 	if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2254 		tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2255 		tg->bio_cnt /= 2;
2256 		tg->bad_bio_cnt /= 2;
2257 	}
2258 }
2259 #endif
2260 
2261 int blk_throtl_init(struct request_queue *q)
2262 {
2263 	struct throtl_data *td;
2264 	int ret;
2265 
2266 	td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2267 	if (!td)
2268 		return -ENOMEM;
2269 	td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2270 		LATENCY_BUCKET_SIZE, __alignof__(u64));
2271 	if (!td->latency_buckets[READ]) {
2272 		kfree(td);
2273 		return -ENOMEM;
2274 	}
2275 	td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2276 		LATENCY_BUCKET_SIZE, __alignof__(u64));
2277 	if (!td->latency_buckets[WRITE]) {
2278 		free_percpu(td->latency_buckets[READ]);
2279 		kfree(td);
2280 		return -ENOMEM;
2281 	}
2282 
2283 	INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2284 	throtl_service_queue_init(&td->service_queue);
2285 
2286 	q->td = td;
2287 	td->queue = q;
2288 
2289 	td->limit_valid[LIMIT_MAX] = true;
2290 	td->limit_index = LIMIT_MAX;
2291 	td->low_upgrade_time = jiffies;
2292 	td->low_downgrade_time = jiffies;
2293 
2294 	/* activate policy */
2295 	ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2296 	if (ret) {
2297 		free_percpu(td->latency_buckets[READ]);
2298 		free_percpu(td->latency_buckets[WRITE]);
2299 		kfree(td);
2300 	}
2301 	return ret;
2302 }
2303 
2304 void blk_throtl_exit(struct request_queue *q)
2305 {
2306 	BUG_ON(!q->td);
2307 	del_timer_sync(&q->td->service_queue.pending_timer);
2308 	throtl_shutdown_wq(q);
2309 	blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2310 	free_percpu(q->td->latency_buckets[READ]);
2311 	free_percpu(q->td->latency_buckets[WRITE]);
2312 	kfree(q->td);
2313 }
2314 
2315 void blk_throtl_register_queue(struct request_queue *q)
2316 {
2317 	struct throtl_data *td;
2318 	int i;
2319 
2320 	td = q->td;
2321 	BUG_ON(!td);
2322 
2323 	if (blk_queue_nonrot(q)) {
2324 		td->throtl_slice = DFL_THROTL_SLICE_SSD;
2325 		td->filtered_latency = LATENCY_FILTERED_SSD;
2326 	} else {
2327 		td->throtl_slice = DFL_THROTL_SLICE_HD;
2328 		td->filtered_latency = LATENCY_FILTERED_HD;
2329 		for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2330 			td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2331 			td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2332 		}
2333 	}
2334 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2335 	/* if no low limit, use previous default */
2336 	td->throtl_slice = DFL_THROTL_SLICE_HD;
2337 #endif
2338 
2339 	td->track_bio_latency = !queue_is_mq(q);
2340 	if (!td->track_bio_latency)
2341 		blk_stat_enable_accounting(q);
2342 }
2343 
2344 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2345 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2346 {
2347 	if (!q->td)
2348 		return -EINVAL;
2349 	return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2350 }
2351 
2352 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2353 	const char *page, size_t count)
2354 {
2355 	unsigned long v;
2356 	unsigned long t;
2357 
2358 	if (!q->td)
2359 		return -EINVAL;
2360 	if (kstrtoul(page, 10, &v))
2361 		return -EINVAL;
2362 	t = msecs_to_jiffies(v);
2363 	if (t == 0 || t > MAX_THROTL_SLICE)
2364 		return -EINVAL;
2365 	q->td->throtl_slice = t;
2366 	return count;
2367 }
2368 #endif
2369 
2370 static int __init throtl_init(void)
2371 {
2372 	kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2373 	if (!kthrotld_workqueue)
2374 		panic("Failed to create kthrotld\n");
2375 
2376 	return blkcg_policy_register(&blkcg_policy_throtl);
2377 }
2378 
2379 module_init(throtl_init);
2380