• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Interface for controlling IO bandwidth on a request queue
4  *
5  * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
6  */
7 
8 #include <linux/module.h>
9 #include <linux/slab.h>
10 #include <linux/blkdev.h>
11 #include <linux/bio.h>
12 #include <linux/blktrace_api.h>
13 #include <linux/blk-cgroup.h>
14 #include "blk.h"
15 #include "blk-cgroup-rwstat.h"
16 
17 /* Max dispatch from a group in 1 round */
18 #define THROTL_GRP_QUANTUM 8
19 
20 /* Total max dispatch from all groups in one round */
21 #define THROTL_QUANTUM 32
22 
23 /* Throttling is performed over a slice and after that slice is renewed */
24 #define DFL_THROTL_SLICE_HD (HZ / 10)
25 #define DFL_THROTL_SLICE_SSD (HZ / 50)
26 #define MAX_THROTL_SLICE (HZ)
27 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
28 #define MIN_THROTL_BPS (320 * 1024)
29 #define MIN_THROTL_IOPS (10)
30 #define DFL_LATENCY_TARGET (-1L)
31 #define DFL_IDLE_THRESHOLD (0)
32 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
33 #define LATENCY_FILTERED_SSD (0)
34 /*
35  * For HD, very small latency comes from sequential IO. Such IO is helpless to
36  * help determine if its IO is impacted by others, hence we ignore the IO
37  */
38 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
39 
40 static struct blkcg_policy blkcg_policy_throtl;
41 
42 /* A workqueue to queue throttle related work */
43 static struct workqueue_struct *kthrotld_workqueue;
44 
45 /*
46  * To implement hierarchical throttling, throtl_grps form a tree and bios
47  * are dispatched upwards level by level until they reach the top and get
48  * issued.  When dispatching bios from the children and local group at each
49  * level, if the bios are dispatched into a single bio_list, there's a risk
50  * of a local or child group which can queue many bios at once filling up
51  * the list starving others.
52  *
53  * To avoid such starvation, dispatched bios are queued separately
54  * according to where they came from.  When they are again dispatched to
55  * the parent, they're popped in round-robin order so that no single source
56  * hogs the dispatch window.
57  *
58  * throtl_qnode is used to keep the queued bios separated by their sources.
59  * Bios are queued to throtl_qnode which in turn is queued to
60  * throtl_service_queue and then dispatched in round-robin order.
61  *
62  * It's also used to track the reference counts on blkg's.  A qnode always
63  * belongs to a throtl_grp and gets queued on itself or the parent, so
64  * incrementing the reference of the associated throtl_grp when a qnode is
65  * queued and decrementing when dequeued is enough to keep the whole blkg
66  * tree pinned while bios are in flight.
67  */
68 struct throtl_qnode {
69 	struct list_head	node;		/* service_queue->queued[] */
70 	struct bio_list		bios;		/* queued bios */
71 	struct throtl_grp	*tg;		/* tg this qnode belongs to */
72 };
73 
74 struct throtl_service_queue {
75 	struct throtl_service_queue *parent_sq;	/* the parent service_queue */
76 
77 	/*
78 	 * Bios queued directly to this service_queue or dispatched from
79 	 * children throtl_grp's.
80 	 */
81 	struct list_head	queued[2];	/* throtl_qnode [READ/WRITE] */
82 	unsigned int		nr_queued[2];	/* number of queued bios */
83 
84 	/*
85 	 * RB tree of active children throtl_grp's, which are sorted by
86 	 * their ->disptime.
87 	 */
88 	struct rb_root_cached	pending_tree;	/* RB tree of active tgs */
89 	unsigned int		nr_pending;	/* # queued in the tree */
90 	unsigned long		first_pending_disptime;	/* disptime of the first tg */
91 	struct timer_list	pending_timer;	/* fires on first_pending_disptime */
92 };
93 
94 enum tg_state_flags {
95 	THROTL_TG_PENDING	= 1 << 0,	/* on parent's pending tree */
96 	THROTL_TG_WAS_EMPTY	= 1 << 1,	/* bio_lists[] became non-empty */
97 };
98 
99 #define rb_entry_tg(node)	rb_entry((node), struct throtl_grp, rb_node)
100 
101 enum {
102 	LIMIT_LOW,
103 	LIMIT_MAX,
104 	LIMIT_CNT,
105 };
106 
107 struct throtl_grp {
108 	/* must be the first member */
109 	struct blkg_policy_data pd;
110 
111 	/* active throtl group service_queue member */
112 	struct rb_node rb_node;
113 
114 	/* throtl_data this group belongs to */
115 	struct throtl_data *td;
116 
117 	/* this group's service queue */
118 	struct throtl_service_queue service_queue;
119 
120 	/*
121 	 * qnode_on_self is used when bios are directly queued to this
122 	 * throtl_grp so that local bios compete fairly with bios
123 	 * dispatched from children.  qnode_on_parent is used when bios are
124 	 * dispatched from this throtl_grp into its parent and will compete
125 	 * with the sibling qnode_on_parents and the parent's
126 	 * qnode_on_self.
127 	 */
128 	struct throtl_qnode qnode_on_self[2];
129 	struct throtl_qnode qnode_on_parent[2];
130 
131 	/*
132 	 * Dispatch time in jiffies. This is the estimated time when group
133 	 * will unthrottle and is ready to dispatch more bio. It is used as
134 	 * key to sort active groups in service tree.
135 	 */
136 	unsigned long disptime;
137 
138 	unsigned int flags;
139 
140 	/* are there any throtl rules between this group and td? */
141 	bool has_rules[2];
142 
143 	/* internally used bytes per second rate limits */
144 	uint64_t bps[2][LIMIT_CNT];
145 	/* user configured bps limits */
146 	uint64_t bps_conf[2][LIMIT_CNT];
147 
148 	/* internally used IOPS limits */
149 	unsigned int iops[2][LIMIT_CNT];
150 	/* user configured IOPS limits */
151 	unsigned int iops_conf[2][LIMIT_CNT];
152 
153 	/* Number of bytes dispatched in current slice */
154 	uint64_t bytes_disp[2];
155 	/* Number of bio's dispatched in current slice */
156 	unsigned int io_disp[2];
157 
158 	unsigned long last_low_overflow_time[2];
159 
160 	uint64_t last_bytes_disp[2];
161 	unsigned int last_io_disp[2];
162 
163 	unsigned long last_check_time;
164 
165 	unsigned long latency_target; /* us */
166 	unsigned long latency_target_conf; /* us */
167 	/* When did we start a new slice */
168 	unsigned long slice_start[2];
169 	unsigned long slice_end[2];
170 
171 	unsigned long last_finish_time; /* ns / 1024 */
172 	unsigned long checked_last_finish_time; /* ns / 1024 */
173 	unsigned long avg_idletime; /* ns / 1024 */
174 	unsigned long idletime_threshold; /* us */
175 	unsigned long idletime_threshold_conf; /* us */
176 
177 	unsigned int bio_cnt; /* total bios */
178 	unsigned int bad_bio_cnt; /* bios exceeding latency threshold */
179 	unsigned long bio_cnt_reset_time;
180 
181 	atomic_t io_split_cnt[2];
182 	atomic_t last_io_split_cnt[2];
183 
184 	struct blkg_rwstat stat_bytes;
185 	struct blkg_rwstat stat_ios;
186 };
187 
188 /* We measure latency for request size from <= 4k to >= 1M */
189 #define LATENCY_BUCKET_SIZE 9
190 
191 struct latency_bucket {
192 	unsigned long total_latency; /* ns / 1024 */
193 	int samples;
194 };
195 
196 struct avg_latency_bucket {
197 	unsigned long latency; /* ns / 1024 */
198 	bool valid;
199 };
200 
201 struct throtl_data
202 {
203 	/* service tree for active throtl groups */
204 	struct throtl_service_queue service_queue;
205 
206 	struct request_queue *queue;
207 
208 	/* Total Number of queued bios on READ and WRITE lists */
209 	unsigned int nr_queued[2];
210 
211 	unsigned int throtl_slice;
212 
213 	/* Work for dispatching throttled bios */
214 	struct work_struct dispatch_work;
215 	unsigned int limit_index;
216 	bool limit_valid[LIMIT_CNT];
217 
218 	unsigned long low_upgrade_time;
219 	unsigned long low_downgrade_time;
220 
221 	unsigned int scale;
222 
223 	struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
224 	struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
225 	struct latency_bucket __percpu *latency_buckets[2];
226 	unsigned long last_calculate_time;
227 	unsigned long filtered_latency;
228 
229 	bool track_bio_latency;
230 };
231 
232 static void throtl_pending_timer_fn(struct timer_list *t);
233 
pd_to_tg(struct blkg_policy_data * pd)234 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
235 {
236 	return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
237 }
238 
blkg_to_tg(struct blkcg_gq * blkg)239 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
240 {
241 	return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
242 }
243 
tg_to_blkg(struct throtl_grp * tg)244 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
245 {
246 	return pd_to_blkg(&tg->pd);
247 }
248 
249 /**
250  * sq_to_tg - return the throl_grp the specified service queue belongs to
251  * @sq: the throtl_service_queue of interest
252  *
253  * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
254  * embedded in throtl_data, %NULL is returned.
255  */
sq_to_tg(struct throtl_service_queue * sq)256 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
257 {
258 	if (sq && sq->parent_sq)
259 		return container_of(sq, struct throtl_grp, service_queue);
260 	else
261 		return NULL;
262 }
263 
264 /**
265  * sq_to_td - return throtl_data the specified service queue belongs to
266  * @sq: the throtl_service_queue of interest
267  *
268  * A service_queue can be embedded in either a throtl_grp or throtl_data.
269  * Determine the associated throtl_data accordingly and return it.
270  */
sq_to_td(struct throtl_service_queue * sq)271 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
272 {
273 	struct throtl_grp *tg = sq_to_tg(sq);
274 
275 	if (tg)
276 		return tg->td;
277 	else
278 		return container_of(sq, struct throtl_data, service_queue);
279 }
280 
281 /*
282  * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
283  * make the IO dispatch more smooth.
284  * Scale up: linearly scale up according to lapsed time since upgrade. For
285  *           every throtl_slice, the limit scales up 1/2 .low limit till the
286  *           limit hits .max limit
287  * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
288  */
throtl_adjusted_limit(uint64_t low,struct throtl_data * td)289 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
290 {
291 	/* arbitrary value to avoid too big scale */
292 	if (td->scale < 4096 && time_after_eq(jiffies,
293 	    td->low_upgrade_time + td->scale * td->throtl_slice))
294 		td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
295 
296 	return low + (low >> 1) * td->scale;
297 }
298 
tg_bps_limit(struct throtl_grp * tg,int rw)299 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
300 {
301 	struct blkcg_gq *blkg = tg_to_blkg(tg);
302 	struct throtl_data *td;
303 	uint64_t ret;
304 
305 	if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
306 		return U64_MAX;
307 
308 	td = tg->td;
309 	ret = tg->bps[rw][td->limit_index];
310 	if (ret == 0 && td->limit_index == LIMIT_LOW) {
311 		/* intermediate node or iops isn't 0 */
312 		if (!list_empty(&blkg->blkcg->css.children) ||
313 		    tg->iops[rw][td->limit_index])
314 			return U64_MAX;
315 		else
316 			return MIN_THROTL_BPS;
317 	}
318 
319 	if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
320 	    tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
321 		uint64_t adjusted;
322 
323 		adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
324 		ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
325 	}
326 	return ret;
327 }
328 
tg_iops_limit(struct throtl_grp * tg,int rw)329 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
330 {
331 	struct blkcg_gq *blkg = tg_to_blkg(tg);
332 	struct throtl_data *td;
333 	unsigned int ret;
334 
335 	if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
336 		return UINT_MAX;
337 
338 	td = tg->td;
339 	ret = tg->iops[rw][td->limit_index];
340 	if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
341 		/* intermediate node or bps isn't 0 */
342 		if (!list_empty(&blkg->blkcg->css.children) ||
343 		    tg->bps[rw][td->limit_index])
344 			return UINT_MAX;
345 		else
346 			return MIN_THROTL_IOPS;
347 	}
348 
349 	if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
350 	    tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
351 		uint64_t adjusted;
352 
353 		adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
354 		if (adjusted > UINT_MAX)
355 			adjusted = UINT_MAX;
356 		ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
357 	}
358 	return ret;
359 }
360 
361 #define request_bucket_index(sectors) \
362 	clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
363 
364 /**
365  * throtl_log - log debug message via blktrace
366  * @sq: the service_queue being reported
367  * @fmt: printf format string
368  * @args: printf args
369  *
370  * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
371  * throtl_grp; otherwise, just "throtl".
372  */
373 #define throtl_log(sq, fmt, args...)	do {				\
374 	struct throtl_grp *__tg = sq_to_tg((sq));			\
375 	struct throtl_data *__td = sq_to_td((sq));			\
376 									\
377 	(void)__td;							\
378 	if (likely(!blk_trace_note_message_enabled(__td->queue)))	\
379 		break;							\
380 	if ((__tg)) {							\
381 		blk_add_cgroup_trace_msg(__td->queue,			\
382 			tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\
383 	} else {							\
384 		blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);	\
385 	}								\
386 } while (0)
387 
throtl_bio_data_size(struct bio * bio)388 static inline unsigned int throtl_bio_data_size(struct bio *bio)
389 {
390 	/* assume it's one sector */
391 	if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
392 		return 512;
393 	return bio->bi_iter.bi_size;
394 }
395 
throtl_qnode_init(struct throtl_qnode * qn,struct throtl_grp * tg)396 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
397 {
398 	INIT_LIST_HEAD(&qn->node);
399 	bio_list_init(&qn->bios);
400 	qn->tg = tg;
401 }
402 
403 /**
404  * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
405  * @bio: bio being added
406  * @qn: qnode to add bio to
407  * @queued: the service_queue->queued[] list @qn belongs to
408  *
409  * Add @bio to @qn and put @qn on @queued if it's not already on.
410  * @qn->tg's reference count is bumped when @qn is activated.  See the
411  * comment on top of throtl_qnode definition for details.
412  */
throtl_qnode_add_bio(struct bio * bio,struct throtl_qnode * qn,struct list_head * queued)413 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
414 				 struct list_head *queued)
415 {
416 	bio_list_add(&qn->bios, bio);
417 	if (list_empty(&qn->node)) {
418 		list_add_tail(&qn->node, queued);
419 		blkg_get(tg_to_blkg(qn->tg));
420 	}
421 }
422 
423 /**
424  * throtl_peek_queued - peek the first bio on a qnode list
425  * @queued: the qnode list to peek
426  */
throtl_peek_queued(struct list_head * queued)427 static struct bio *throtl_peek_queued(struct list_head *queued)
428 {
429 	struct throtl_qnode *qn;
430 	struct bio *bio;
431 
432 	if (list_empty(queued))
433 		return NULL;
434 
435 	qn = list_first_entry(queued, struct throtl_qnode, node);
436 	bio = bio_list_peek(&qn->bios);
437 	WARN_ON_ONCE(!bio);
438 	return bio;
439 }
440 
441 /**
442  * throtl_pop_queued - pop the first bio form a qnode list
443  * @queued: the qnode list to pop a bio from
444  * @tg_to_put: optional out argument for throtl_grp to put
445  *
446  * Pop the first bio from the qnode list @queued.  After popping, the first
447  * qnode is removed from @queued if empty or moved to the end of @queued so
448  * that the popping order is round-robin.
449  *
450  * When the first qnode is removed, its associated throtl_grp should be put
451  * too.  If @tg_to_put is NULL, this function automatically puts it;
452  * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
453  * responsible for putting it.
454  */
throtl_pop_queued(struct list_head * queued,struct throtl_grp ** tg_to_put)455 static struct bio *throtl_pop_queued(struct list_head *queued,
456 				     struct throtl_grp **tg_to_put)
457 {
458 	struct throtl_qnode *qn;
459 	struct bio *bio;
460 
461 	if (list_empty(queued))
462 		return NULL;
463 
464 	qn = list_first_entry(queued, struct throtl_qnode, node);
465 	bio = bio_list_pop(&qn->bios);
466 	WARN_ON_ONCE(!bio);
467 
468 	if (bio_list_empty(&qn->bios)) {
469 		list_del_init(&qn->node);
470 		if (tg_to_put)
471 			*tg_to_put = qn->tg;
472 		else
473 			blkg_put(tg_to_blkg(qn->tg));
474 	} else {
475 		list_move_tail(&qn->node, queued);
476 	}
477 
478 	return bio;
479 }
480 
481 /* init a service_queue, assumes the caller zeroed it */
throtl_service_queue_init(struct throtl_service_queue * sq)482 static void throtl_service_queue_init(struct throtl_service_queue *sq)
483 {
484 	INIT_LIST_HEAD(&sq->queued[0]);
485 	INIT_LIST_HEAD(&sq->queued[1]);
486 	sq->pending_tree = RB_ROOT_CACHED;
487 	timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
488 }
489 
throtl_pd_alloc(gfp_t gfp,struct request_queue * q,struct blkcg * blkcg)490 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp,
491 						struct request_queue *q,
492 						struct blkcg *blkcg)
493 {
494 	struct throtl_grp *tg;
495 	int rw;
496 
497 	tg = kzalloc_node(sizeof(*tg), gfp, q->node);
498 	if (!tg)
499 		return NULL;
500 
501 	if (blkg_rwstat_init(&tg->stat_bytes, gfp))
502 		goto err_free_tg;
503 
504 	if (blkg_rwstat_init(&tg->stat_ios, gfp))
505 		goto err_exit_stat_bytes;
506 
507 	throtl_service_queue_init(&tg->service_queue);
508 
509 	for (rw = READ; rw <= WRITE; rw++) {
510 		throtl_qnode_init(&tg->qnode_on_self[rw], tg);
511 		throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
512 	}
513 
514 	RB_CLEAR_NODE(&tg->rb_node);
515 	tg->bps[READ][LIMIT_MAX] = U64_MAX;
516 	tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
517 	tg->iops[READ][LIMIT_MAX] = UINT_MAX;
518 	tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
519 	tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
520 	tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
521 	tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
522 	tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
523 	/* LIMIT_LOW will have default value 0 */
524 
525 	tg->latency_target = DFL_LATENCY_TARGET;
526 	tg->latency_target_conf = DFL_LATENCY_TARGET;
527 	tg->idletime_threshold = DFL_IDLE_THRESHOLD;
528 	tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
529 
530 	return &tg->pd;
531 
532 err_exit_stat_bytes:
533 	blkg_rwstat_exit(&tg->stat_bytes);
534 err_free_tg:
535 	kfree(tg);
536 	return NULL;
537 }
538 
throtl_pd_init(struct blkg_policy_data * pd)539 static void throtl_pd_init(struct blkg_policy_data *pd)
540 {
541 	struct throtl_grp *tg = pd_to_tg(pd);
542 	struct blkcg_gq *blkg = tg_to_blkg(tg);
543 	struct throtl_data *td = blkg->q->td;
544 	struct throtl_service_queue *sq = &tg->service_queue;
545 
546 	/*
547 	 * If on the default hierarchy, we switch to properly hierarchical
548 	 * behavior where limits on a given throtl_grp are applied to the
549 	 * whole subtree rather than just the group itself.  e.g. If 16M
550 	 * read_bps limit is set on the root group, the whole system can't
551 	 * exceed 16M for the device.
552 	 *
553 	 * If not on the default hierarchy, the broken flat hierarchy
554 	 * behavior is retained where all throtl_grps are treated as if
555 	 * they're all separate root groups right below throtl_data.
556 	 * Limits of a group don't interact with limits of other groups
557 	 * regardless of the position of the group in the hierarchy.
558 	 */
559 	sq->parent_sq = &td->service_queue;
560 	if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
561 		sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
562 	tg->td = td;
563 }
564 
565 /*
566  * Set has_rules[] if @tg or any of its parents have limits configured.
567  * This doesn't require walking up to the top of the hierarchy as the
568  * parent's has_rules[] is guaranteed to be correct.
569  */
tg_update_has_rules(struct throtl_grp * tg)570 static void tg_update_has_rules(struct throtl_grp *tg)
571 {
572 	struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
573 	struct throtl_data *td = tg->td;
574 	int rw;
575 
576 	for (rw = READ; rw <= WRITE; rw++)
577 		tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
578 			(td->limit_valid[td->limit_index] &&
579 			 (tg_bps_limit(tg, rw) != U64_MAX ||
580 			  tg_iops_limit(tg, rw) != UINT_MAX));
581 }
582 
throtl_pd_online(struct blkg_policy_data * pd)583 static void throtl_pd_online(struct blkg_policy_data *pd)
584 {
585 	struct throtl_grp *tg = pd_to_tg(pd);
586 	/*
587 	 * We don't want new groups to escape the limits of its ancestors.
588 	 * Update has_rules[] after a new group is brought online.
589 	 */
590 	tg_update_has_rules(tg);
591 }
592 
blk_throtl_update_limit_valid(struct throtl_data * td)593 static void blk_throtl_update_limit_valid(struct throtl_data *td)
594 {
595 	struct cgroup_subsys_state *pos_css;
596 	struct blkcg_gq *blkg;
597 	bool low_valid = false;
598 
599 	rcu_read_lock();
600 	blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
601 		struct throtl_grp *tg = blkg_to_tg(blkg);
602 
603 		if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
604 		    tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
605 			low_valid = true;
606 			break;
607 		}
608 	}
609 	rcu_read_unlock();
610 
611 	td->limit_valid[LIMIT_LOW] = low_valid;
612 }
613 
614 static void throtl_upgrade_state(struct throtl_data *td);
throtl_pd_offline(struct blkg_policy_data * pd)615 static void throtl_pd_offline(struct blkg_policy_data *pd)
616 {
617 	struct throtl_grp *tg = pd_to_tg(pd);
618 
619 	tg->bps[READ][LIMIT_LOW] = 0;
620 	tg->bps[WRITE][LIMIT_LOW] = 0;
621 	tg->iops[READ][LIMIT_LOW] = 0;
622 	tg->iops[WRITE][LIMIT_LOW] = 0;
623 
624 	blk_throtl_update_limit_valid(tg->td);
625 
626 	if (!tg->td->limit_valid[tg->td->limit_index])
627 		throtl_upgrade_state(tg->td);
628 }
629 
throtl_pd_free(struct blkg_policy_data * pd)630 static void throtl_pd_free(struct blkg_policy_data *pd)
631 {
632 	struct throtl_grp *tg = pd_to_tg(pd);
633 
634 	del_timer_sync(&tg->service_queue.pending_timer);
635 	blkg_rwstat_exit(&tg->stat_bytes);
636 	blkg_rwstat_exit(&tg->stat_ios);
637 	kfree(tg);
638 }
639 
640 static struct throtl_grp *
throtl_rb_first(struct throtl_service_queue * parent_sq)641 throtl_rb_first(struct throtl_service_queue *parent_sq)
642 {
643 	struct rb_node *n;
644 
645 	n = rb_first_cached(&parent_sq->pending_tree);
646 	WARN_ON_ONCE(!n);
647 	if (!n)
648 		return NULL;
649 	return rb_entry_tg(n);
650 }
651 
throtl_rb_erase(struct rb_node * n,struct throtl_service_queue * parent_sq)652 static void throtl_rb_erase(struct rb_node *n,
653 			    struct throtl_service_queue *parent_sq)
654 {
655 	rb_erase_cached(n, &parent_sq->pending_tree);
656 	RB_CLEAR_NODE(n);
657 	--parent_sq->nr_pending;
658 }
659 
update_min_dispatch_time(struct throtl_service_queue * parent_sq)660 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
661 {
662 	struct throtl_grp *tg;
663 
664 	tg = throtl_rb_first(parent_sq);
665 	if (!tg)
666 		return;
667 
668 	parent_sq->first_pending_disptime = tg->disptime;
669 }
670 
tg_service_queue_add(struct throtl_grp * tg)671 static void tg_service_queue_add(struct throtl_grp *tg)
672 {
673 	struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
674 	struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
675 	struct rb_node *parent = NULL;
676 	struct throtl_grp *__tg;
677 	unsigned long key = tg->disptime;
678 	bool leftmost = true;
679 
680 	while (*node != NULL) {
681 		parent = *node;
682 		__tg = rb_entry_tg(parent);
683 
684 		if (time_before(key, __tg->disptime))
685 			node = &parent->rb_left;
686 		else {
687 			node = &parent->rb_right;
688 			leftmost = false;
689 		}
690 	}
691 
692 	rb_link_node(&tg->rb_node, parent, node);
693 	rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
694 			       leftmost);
695 }
696 
throtl_enqueue_tg(struct throtl_grp * tg)697 static void throtl_enqueue_tg(struct throtl_grp *tg)
698 {
699 	if (!(tg->flags & THROTL_TG_PENDING)) {
700 		tg_service_queue_add(tg);
701 		tg->flags |= THROTL_TG_PENDING;
702 		tg->service_queue.parent_sq->nr_pending++;
703 	}
704 }
705 
throtl_dequeue_tg(struct throtl_grp * tg)706 static void throtl_dequeue_tg(struct throtl_grp *tg)
707 {
708 	if (tg->flags & THROTL_TG_PENDING) {
709 		throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
710 		tg->flags &= ~THROTL_TG_PENDING;
711 	}
712 }
713 
714 /* Call with queue lock held */
throtl_schedule_pending_timer(struct throtl_service_queue * sq,unsigned long expires)715 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
716 					  unsigned long expires)
717 {
718 	unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
719 
720 	/*
721 	 * Since we are adjusting the throttle limit dynamically, the sleep
722 	 * time calculated according to previous limit might be invalid. It's
723 	 * possible the cgroup sleep time is very long and no other cgroups
724 	 * have IO running so notify the limit changes. Make sure the cgroup
725 	 * doesn't sleep too long to avoid the missed notification.
726 	 */
727 	if (time_after(expires, max_expire))
728 		expires = max_expire;
729 	mod_timer(&sq->pending_timer, expires);
730 	throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
731 		   expires - jiffies, jiffies);
732 }
733 
734 /**
735  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
736  * @sq: the service_queue to schedule dispatch for
737  * @force: force scheduling
738  *
739  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
740  * dispatch time of the first pending child.  Returns %true if either timer
741  * is armed or there's no pending child left.  %false if the current
742  * dispatch window is still open and the caller should continue
743  * dispatching.
744  *
745  * If @force is %true, the dispatch timer is always scheduled and this
746  * function is guaranteed to return %true.  This is to be used when the
747  * caller can't dispatch itself and needs to invoke pending_timer
748  * unconditionally.  Note that forced scheduling is likely to induce short
749  * delay before dispatch starts even if @sq->first_pending_disptime is not
750  * in the future and thus shouldn't be used in hot paths.
751  */
throtl_schedule_next_dispatch(struct throtl_service_queue * sq,bool force)752 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
753 					  bool force)
754 {
755 	/* any pending children left? */
756 	if (!sq->nr_pending)
757 		return true;
758 
759 	update_min_dispatch_time(sq);
760 
761 	/* is the next dispatch time in the future? */
762 	if (force || time_after(sq->first_pending_disptime, jiffies)) {
763 		throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
764 		return true;
765 	}
766 
767 	/* tell the caller to continue dispatching */
768 	return false;
769 }
770 
throtl_start_new_slice_with_credit(struct throtl_grp * tg,bool rw,unsigned long start)771 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
772 		bool rw, unsigned long start)
773 {
774 	tg->bytes_disp[rw] = 0;
775 	tg->io_disp[rw] = 0;
776 
777 	atomic_set(&tg->io_split_cnt[rw], 0);
778 
779 	/*
780 	 * Previous slice has expired. We must have trimmed it after last
781 	 * bio dispatch. That means since start of last slice, we never used
782 	 * that bandwidth. Do try to make use of that bandwidth while giving
783 	 * credit.
784 	 */
785 	if (time_after_eq(start, tg->slice_start[rw]))
786 		tg->slice_start[rw] = start;
787 
788 	tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
789 	throtl_log(&tg->service_queue,
790 		   "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
791 		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
792 		   tg->slice_end[rw], jiffies);
793 }
794 
throtl_start_new_slice(struct throtl_grp * tg,bool rw)795 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
796 {
797 	tg->bytes_disp[rw] = 0;
798 	tg->io_disp[rw] = 0;
799 	tg->slice_start[rw] = jiffies;
800 	tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
801 
802 	atomic_set(&tg->io_split_cnt[rw], 0);
803 
804 	throtl_log(&tg->service_queue,
805 		   "[%c] new slice start=%lu end=%lu jiffies=%lu",
806 		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
807 		   tg->slice_end[rw], jiffies);
808 }
809 
throtl_set_slice_end(struct throtl_grp * tg,bool rw,unsigned long jiffy_end)810 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
811 					unsigned long jiffy_end)
812 {
813 	tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
814 }
815 
throtl_extend_slice(struct throtl_grp * tg,bool rw,unsigned long jiffy_end)816 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
817 				       unsigned long jiffy_end)
818 {
819 	throtl_set_slice_end(tg, rw, jiffy_end);
820 	throtl_log(&tg->service_queue,
821 		   "[%c] extend slice start=%lu end=%lu jiffies=%lu",
822 		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
823 		   tg->slice_end[rw], jiffies);
824 }
825 
826 /* Determine if previously allocated or extended slice is complete or not */
throtl_slice_used(struct throtl_grp * tg,bool rw)827 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
828 {
829 	if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
830 		return false;
831 
832 	return true;
833 }
834 
835 /* Trim the used slices and adjust slice start accordingly */
throtl_trim_slice(struct throtl_grp * tg,bool rw)836 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
837 {
838 	unsigned long nr_slices, time_elapsed, io_trim;
839 	u64 bytes_trim, tmp;
840 
841 	BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
842 
843 	/*
844 	 * If bps are unlimited (-1), then time slice don't get
845 	 * renewed. Don't try to trim the slice if slice is used. A new
846 	 * slice will start when appropriate.
847 	 */
848 	if (throtl_slice_used(tg, rw))
849 		return;
850 
851 	/*
852 	 * A bio has been dispatched. Also adjust slice_end. It might happen
853 	 * that initially cgroup limit was very low resulting in high
854 	 * slice_end, but later limit was bumped up and bio was dispatched
855 	 * sooner, then we need to reduce slice_end. A high bogus slice_end
856 	 * is bad because it does not allow new slice to start.
857 	 */
858 
859 	throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
860 
861 	time_elapsed = jiffies - tg->slice_start[rw];
862 
863 	nr_slices = time_elapsed / tg->td->throtl_slice;
864 
865 	if (!nr_slices)
866 		return;
867 	tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
868 	do_div(tmp, HZ);
869 	bytes_trim = tmp;
870 
871 	io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
872 		HZ;
873 
874 	if (!bytes_trim && !io_trim)
875 		return;
876 
877 	if (tg->bytes_disp[rw] >= bytes_trim)
878 		tg->bytes_disp[rw] -= bytes_trim;
879 	else
880 		tg->bytes_disp[rw] = 0;
881 
882 	if (tg->io_disp[rw] >= io_trim)
883 		tg->io_disp[rw] -= io_trim;
884 	else
885 		tg->io_disp[rw] = 0;
886 
887 	tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
888 
889 	throtl_log(&tg->service_queue,
890 		   "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
891 		   rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
892 		   tg->slice_start[rw], tg->slice_end[rw], jiffies);
893 }
894 
tg_with_in_iops_limit(struct throtl_grp * tg,struct bio * bio,u32 iops_limit,unsigned long * wait)895 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
896 				  u32 iops_limit, unsigned long *wait)
897 {
898 	bool rw = bio_data_dir(bio);
899 	unsigned int io_allowed;
900 	unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
901 	u64 tmp;
902 
903 	if (iops_limit == UINT_MAX) {
904 		if (wait)
905 			*wait = 0;
906 		return true;
907 	}
908 
909 	jiffy_elapsed = jiffies - tg->slice_start[rw];
910 
911 	/* Round up to the next throttle slice, wait time must be nonzero */
912 	jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
913 
914 	/*
915 	 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
916 	 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
917 	 * will allow dispatch after 1 second and after that slice should
918 	 * have been trimmed.
919 	 */
920 
921 	tmp = (u64)iops_limit * jiffy_elapsed_rnd;
922 	do_div(tmp, HZ);
923 
924 	if (tmp > UINT_MAX)
925 		io_allowed = UINT_MAX;
926 	else
927 		io_allowed = tmp;
928 
929 	if (tg->io_disp[rw] + 1 <= io_allowed) {
930 		if (wait)
931 			*wait = 0;
932 		return true;
933 	}
934 
935 	/* Calc approx time to dispatch */
936 	jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
937 
938 	if (wait)
939 		*wait = jiffy_wait;
940 	return false;
941 }
942 
tg_with_in_bps_limit(struct throtl_grp * tg,struct bio * bio,u64 bps_limit,unsigned long * wait)943 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
944 				 u64 bps_limit, unsigned long *wait)
945 {
946 	bool rw = bio_data_dir(bio);
947 	u64 bytes_allowed, extra_bytes;
948 	unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
949 	unsigned int bio_size = throtl_bio_data_size(bio);
950 
951 	if (bps_limit == U64_MAX) {
952 		if (wait)
953 			*wait = 0;
954 		return true;
955 	}
956 
957 	jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
958 
959 	/* Slice has just started. Consider one slice interval */
960 	if (!jiffy_elapsed)
961 		jiffy_elapsed_rnd = tg->td->throtl_slice;
962 
963 	jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
964 	bytes_allowed = mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed_rnd,
965 					    (u64)HZ);
966 
967 	if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
968 		if (wait)
969 			*wait = 0;
970 		return true;
971 	}
972 
973 	/* Calc approx time to dispatch */
974 	extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
975 	jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
976 
977 	if (!jiffy_wait)
978 		jiffy_wait = 1;
979 
980 	/*
981 	 * This wait time is without taking into consideration the rounding
982 	 * up we did. Add that time also.
983 	 */
984 	jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
985 	if (wait)
986 		*wait = jiffy_wait;
987 	return false;
988 }
989 
990 /*
991  * Returns whether one can dispatch a bio or not. Also returns approx number
992  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
993  */
tg_may_dispatch(struct throtl_grp * tg,struct bio * bio,unsigned long * wait)994 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
995 			    unsigned long *wait)
996 {
997 	bool rw = bio_data_dir(bio);
998 	unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
999 	u64 bps_limit = tg_bps_limit(tg, rw);
1000 	u32 iops_limit = tg_iops_limit(tg, rw);
1001 
1002 	/*
1003  	 * Currently whole state machine of group depends on first bio
1004 	 * queued in the group bio list. So one should not be calling
1005 	 * this function with a different bio if there are other bios
1006 	 * queued.
1007 	 */
1008 	BUG_ON(tg->service_queue.nr_queued[rw] &&
1009 	       bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
1010 
1011 	/* If tg->bps = -1, then BW is unlimited */
1012 	if (bps_limit == U64_MAX && iops_limit == UINT_MAX) {
1013 		if (wait)
1014 			*wait = 0;
1015 		return true;
1016 	}
1017 
1018 	/*
1019 	 * If previous slice expired, start a new one otherwise renew/extend
1020 	 * existing slice to make sure it is at least throtl_slice interval
1021 	 * long since now. New slice is started only for empty throttle group.
1022 	 * If there is queued bio, that means there should be an active
1023 	 * slice and it should be extended instead.
1024 	 */
1025 	if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
1026 		throtl_start_new_slice(tg, rw);
1027 	else {
1028 		if (time_before(tg->slice_end[rw],
1029 		    jiffies + tg->td->throtl_slice))
1030 			throtl_extend_slice(tg, rw,
1031 				jiffies + tg->td->throtl_slice);
1032 	}
1033 
1034 	if (iops_limit != UINT_MAX)
1035 		tg->io_disp[rw] += atomic_xchg(&tg->io_split_cnt[rw], 0);
1036 
1037 	if (tg_with_in_bps_limit(tg, bio, bps_limit, &bps_wait) &&
1038 	    tg_with_in_iops_limit(tg, bio, iops_limit, &iops_wait)) {
1039 		if (wait)
1040 			*wait = 0;
1041 		return true;
1042 	}
1043 
1044 	max_wait = max(bps_wait, iops_wait);
1045 
1046 	if (wait)
1047 		*wait = max_wait;
1048 
1049 	if (time_before(tg->slice_end[rw], jiffies + max_wait))
1050 		throtl_extend_slice(tg, rw, jiffies + max_wait);
1051 
1052 	return false;
1053 }
1054 
throtl_charge_bio(struct throtl_grp * tg,struct bio * bio)1055 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1056 {
1057 	bool rw = bio_data_dir(bio);
1058 	unsigned int bio_size = throtl_bio_data_size(bio);
1059 
1060 	/* Charge the bio to the group */
1061 	tg->bytes_disp[rw] += bio_size;
1062 	tg->io_disp[rw]++;
1063 	tg->last_bytes_disp[rw] += bio_size;
1064 	tg->last_io_disp[rw]++;
1065 
1066 	/*
1067 	 * BIO_THROTTLED is used to prevent the same bio to be throttled
1068 	 * more than once as a throttled bio will go through blk-throtl the
1069 	 * second time when it eventually gets issued.  Set it when a bio
1070 	 * is being charged to a tg.
1071 	 */
1072 	if (!bio_flagged(bio, BIO_THROTTLED))
1073 		bio_set_flag(bio, BIO_THROTTLED);
1074 }
1075 
1076 /**
1077  * throtl_add_bio_tg - add a bio to the specified throtl_grp
1078  * @bio: bio to add
1079  * @qn: qnode to use
1080  * @tg: the target throtl_grp
1081  *
1082  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
1083  * tg->qnode_on_self[] is used.
1084  */
throtl_add_bio_tg(struct bio * bio,struct throtl_qnode * qn,struct throtl_grp * tg)1085 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1086 			      struct throtl_grp *tg)
1087 {
1088 	struct throtl_service_queue *sq = &tg->service_queue;
1089 	bool rw = bio_data_dir(bio);
1090 
1091 	if (!qn)
1092 		qn = &tg->qnode_on_self[rw];
1093 
1094 	/*
1095 	 * If @tg doesn't currently have any bios queued in the same
1096 	 * direction, queueing @bio can change when @tg should be
1097 	 * dispatched.  Mark that @tg was empty.  This is automatically
1098 	 * cleared on the next tg_update_disptime().
1099 	 */
1100 	if (!sq->nr_queued[rw])
1101 		tg->flags |= THROTL_TG_WAS_EMPTY;
1102 
1103 	throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1104 
1105 	sq->nr_queued[rw]++;
1106 	throtl_enqueue_tg(tg);
1107 }
1108 
tg_update_disptime(struct throtl_grp * tg)1109 static void tg_update_disptime(struct throtl_grp *tg)
1110 {
1111 	struct throtl_service_queue *sq = &tg->service_queue;
1112 	unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1113 	struct bio *bio;
1114 
1115 	bio = throtl_peek_queued(&sq->queued[READ]);
1116 	if (bio)
1117 		tg_may_dispatch(tg, bio, &read_wait);
1118 
1119 	bio = throtl_peek_queued(&sq->queued[WRITE]);
1120 	if (bio)
1121 		tg_may_dispatch(tg, bio, &write_wait);
1122 
1123 	min_wait = min(read_wait, write_wait);
1124 	disptime = jiffies + min_wait;
1125 
1126 	/* Update dispatch time */
1127 	throtl_dequeue_tg(tg);
1128 	tg->disptime = disptime;
1129 	throtl_enqueue_tg(tg);
1130 
1131 	/* see throtl_add_bio_tg() */
1132 	tg->flags &= ~THROTL_TG_WAS_EMPTY;
1133 }
1134 
start_parent_slice_with_credit(struct throtl_grp * child_tg,struct throtl_grp * parent_tg,bool rw)1135 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1136 					struct throtl_grp *parent_tg, bool rw)
1137 {
1138 	if (throtl_slice_used(parent_tg, rw)) {
1139 		throtl_start_new_slice_with_credit(parent_tg, rw,
1140 				child_tg->slice_start[rw]);
1141 	}
1142 
1143 }
1144 
tg_dispatch_one_bio(struct throtl_grp * tg,bool rw)1145 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1146 {
1147 	struct throtl_service_queue *sq = &tg->service_queue;
1148 	struct throtl_service_queue *parent_sq = sq->parent_sq;
1149 	struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1150 	struct throtl_grp *tg_to_put = NULL;
1151 	struct bio *bio;
1152 
1153 	/*
1154 	 * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1155 	 * from @tg may put its reference and @parent_sq might end up
1156 	 * getting released prematurely.  Remember the tg to put and put it
1157 	 * after @bio is transferred to @parent_sq.
1158 	 */
1159 	bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1160 	sq->nr_queued[rw]--;
1161 
1162 	throtl_charge_bio(tg, bio);
1163 
1164 	/*
1165 	 * If our parent is another tg, we just need to transfer @bio to
1166 	 * the parent using throtl_add_bio_tg().  If our parent is
1167 	 * @td->service_queue, @bio is ready to be issued.  Put it on its
1168 	 * bio_lists[] and decrease total number queued.  The caller is
1169 	 * responsible for issuing these bios.
1170 	 */
1171 	if (parent_tg) {
1172 		throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1173 		start_parent_slice_with_credit(tg, parent_tg, rw);
1174 	} else {
1175 		throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1176 				     &parent_sq->queued[rw]);
1177 		BUG_ON(tg->td->nr_queued[rw] <= 0);
1178 		tg->td->nr_queued[rw]--;
1179 	}
1180 
1181 	throtl_trim_slice(tg, rw);
1182 
1183 	if (tg_to_put)
1184 		blkg_put(tg_to_blkg(tg_to_put));
1185 }
1186 
throtl_dispatch_tg(struct throtl_grp * tg)1187 static int throtl_dispatch_tg(struct throtl_grp *tg)
1188 {
1189 	struct throtl_service_queue *sq = &tg->service_queue;
1190 	unsigned int nr_reads = 0, nr_writes = 0;
1191 	unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1192 	unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1193 	struct bio *bio;
1194 
1195 	/* Try to dispatch 75% READS and 25% WRITES */
1196 
1197 	while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1198 	       tg_may_dispatch(tg, bio, NULL)) {
1199 
1200 		tg_dispatch_one_bio(tg, bio_data_dir(bio));
1201 		nr_reads++;
1202 
1203 		if (nr_reads >= max_nr_reads)
1204 			break;
1205 	}
1206 
1207 	while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1208 	       tg_may_dispatch(tg, bio, NULL)) {
1209 
1210 		tg_dispatch_one_bio(tg, bio_data_dir(bio));
1211 		nr_writes++;
1212 
1213 		if (nr_writes >= max_nr_writes)
1214 			break;
1215 	}
1216 
1217 	return nr_reads + nr_writes;
1218 }
1219 
throtl_select_dispatch(struct throtl_service_queue * parent_sq)1220 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1221 {
1222 	unsigned int nr_disp = 0;
1223 
1224 	while (1) {
1225 		struct throtl_grp *tg;
1226 		struct throtl_service_queue *sq;
1227 
1228 		if (!parent_sq->nr_pending)
1229 			break;
1230 
1231 		tg = throtl_rb_first(parent_sq);
1232 		if (!tg)
1233 			break;
1234 
1235 		if (time_before(jiffies, tg->disptime))
1236 			break;
1237 
1238 		throtl_dequeue_tg(tg);
1239 
1240 		nr_disp += throtl_dispatch_tg(tg);
1241 
1242 		sq = &tg->service_queue;
1243 		if (sq->nr_queued[0] || sq->nr_queued[1])
1244 			tg_update_disptime(tg);
1245 
1246 		if (nr_disp >= THROTL_QUANTUM)
1247 			break;
1248 	}
1249 
1250 	return nr_disp;
1251 }
1252 
1253 static bool throtl_can_upgrade(struct throtl_data *td,
1254 	struct throtl_grp *this_tg);
1255 /**
1256  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1257  * @t: the pending_timer member of the throtl_service_queue being serviced
1258  *
1259  * This timer is armed when a child throtl_grp with active bio's become
1260  * pending and queued on the service_queue's pending_tree and expires when
1261  * the first child throtl_grp should be dispatched.  This function
1262  * dispatches bio's from the children throtl_grps to the parent
1263  * service_queue.
1264  *
1265  * If the parent's parent is another throtl_grp, dispatching is propagated
1266  * by either arming its pending_timer or repeating dispatch directly.  If
1267  * the top-level service_tree is reached, throtl_data->dispatch_work is
1268  * kicked so that the ready bio's are issued.
1269  */
throtl_pending_timer_fn(struct timer_list * t)1270 static void throtl_pending_timer_fn(struct timer_list *t)
1271 {
1272 	struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1273 	struct throtl_grp *tg = sq_to_tg(sq);
1274 	struct throtl_data *td = sq_to_td(sq);
1275 	struct request_queue *q = td->queue;
1276 	struct throtl_service_queue *parent_sq;
1277 	bool dispatched;
1278 	int ret;
1279 
1280 	spin_lock_irq(&q->queue_lock);
1281 	if (throtl_can_upgrade(td, NULL))
1282 		throtl_upgrade_state(td);
1283 
1284 again:
1285 	parent_sq = sq->parent_sq;
1286 	dispatched = false;
1287 
1288 	while (true) {
1289 		throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1290 			   sq->nr_queued[READ] + sq->nr_queued[WRITE],
1291 			   sq->nr_queued[READ], sq->nr_queued[WRITE]);
1292 
1293 		ret = throtl_select_dispatch(sq);
1294 		if (ret) {
1295 			throtl_log(sq, "bios disp=%u", ret);
1296 			dispatched = true;
1297 		}
1298 
1299 		if (throtl_schedule_next_dispatch(sq, false))
1300 			break;
1301 
1302 		/* this dispatch windows is still open, relax and repeat */
1303 		spin_unlock_irq(&q->queue_lock);
1304 		cpu_relax();
1305 		spin_lock_irq(&q->queue_lock);
1306 	}
1307 
1308 	if (!dispatched)
1309 		goto out_unlock;
1310 
1311 	if (parent_sq) {
1312 		/* @parent_sq is another throl_grp, propagate dispatch */
1313 		if (tg->flags & THROTL_TG_WAS_EMPTY) {
1314 			tg_update_disptime(tg);
1315 			if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1316 				/* window is already open, repeat dispatching */
1317 				sq = parent_sq;
1318 				tg = sq_to_tg(sq);
1319 				goto again;
1320 			}
1321 		}
1322 	} else {
1323 		/* reached the top-level, queue issuing */
1324 		queue_work(kthrotld_workqueue, &td->dispatch_work);
1325 	}
1326 out_unlock:
1327 	spin_unlock_irq(&q->queue_lock);
1328 }
1329 
1330 /**
1331  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1332  * @work: work item being executed
1333  *
1334  * This function is queued for execution when bios reach the bio_lists[]
1335  * of throtl_data->service_queue.  Those bios are ready and issued by this
1336  * function.
1337  */
blk_throtl_dispatch_work_fn(struct work_struct * work)1338 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1339 {
1340 	struct throtl_data *td = container_of(work, struct throtl_data,
1341 					      dispatch_work);
1342 	struct throtl_service_queue *td_sq = &td->service_queue;
1343 	struct request_queue *q = td->queue;
1344 	struct bio_list bio_list_on_stack;
1345 	struct bio *bio;
1346 	struct blk_plug plug;
1347 	int rw;
1348 
1349 	bio_list_init(&bio_list_on_stack);
1350 
1351 	spin_lock_irq(&q->queue_lock);
1352 	for (rw = READ; rw <= WRITE; rw++)
1353 		while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1354 			bio_list_add(&bio_list_on_stack, bio);
1355 	spin_unlock_irq(&q->queue_lock);
1356 
1357 	if (!bio_list_empty(&bio_list_on_stack)) {
1358 		blk_start_plug(&plug);
1359 		while ((bio = bio_list_pop(&bio_list_on_stack)))
1360 			submit_bio_noacct(bio);
1361 		blk_finish_plug(&plug);
1362 	}
1363 }
1364 
tg_prfill_conf_u64(struct seq_file * sf,struct blkg_policy_data * pd,int off)1365 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1366 			      int off)
1367 {
1368 	struct throtl_grp *tg = pd_to_tg(pd);
1369 	u64 v = *(u64 *)((void *)tg + off);
1370 
1371 	if (v == U64_MAX)
1372 		return 0;
1373 	return __blkg_prfill_u64(sf, pd, v);
1374 }
1375 
tg_prfill_conf_uint(struct seq_file * sf,struct blkg_policy_data * pd,int off)1376 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1377 			       int off)
1378 {
1379 	struct throtl_grp *tg = pd_to_tg(pd);
1380 	unsigned int v = *(unsigned int *)((void *)tg + off);
1381 
1382 	if (v == UINT_MAX)
1383 		return 0;
1384 	return __blkg_prfill_u64(sf, pd, v);
1385 }
1386 
tg_print_conf_u64(struct seq_file * sf,void * v)1387 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1388 {
1389 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1390 			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1391 	return 0;
1392 }
1393 
tg_print_conf_uint(struct seq_file * sf,void * v)1394 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1395 {
1396 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1397 			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1398 	return 0;
1399 }
1400 
tg_conf_updated(struct throtl_grp * tg,bool global)1401 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1402 {
1403 	struct throtl_service_queue *sq = &tg->service_queue;
1404 	struct cgroup_subsys_state *pos_css;
1405 	struct blkcg_gq *blkg;
1406 
1407 	throtl_log(&tg->service_queue,
1408 		   "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1409 		   tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1410 		   tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1411 
1412 	rcu_read_lock();
1413 	/*
1414 	 * Update has_rules[] flags for the updated tg's subtree.  A tg is
1415 	 * considered to have rules if either the tg itself or any of its
1416 	 * ancestors has rules.  This identifies groups without any
1417 	 * restrictions in the whole hierarchy and allows them to bypass
1418 	 * blk-throttle.
1419 	 */
1420 	blkg_for_each_descendant_pre(blkg, pos_css,
1421 			global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1422 		struct throtl_grp *this_tg = blkg_to_tg(blkg);
1423 		struct throtl_grp *parent_tg;
1424 
1425 		tg_update_has_rules(this_tg);
1426 		/* ignore root/second level */
1427 		if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1428 		    !blkg->parent->parent)
1429 			continue;
1430 		parent_tg = blkg_to_tg(blkg->parent);
1431 		/*
1432 		 * make sure all children has lower idle time threshold and
1433 		 * higher latency target
1434 		 */
1435 		this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1436 				parent_tg->idletime_threshold);
1437 		this_tg->latency_target = max(this_tg->latency_target,
1438 				parent_tg->latency_target);
1439 	}
1440 	rcu_read_unlock();
1441 
1442 	/*
1443 	 * We're already holding queue_lock and know @tg is valid.  Let's
1444 	 * apply the new config directly.
1445 	 *
1446 	 * Restart the slices for both READ and WRITES. It might happen
1447 	 * that a group's limit are dropped suddenly and we don't want to
1448 	 * account recently dispatched IO with new low rate.
1449 	 */
1450 	throtl_start_new_slice(tg, READ);
1451 	throtl_start_new_slice(tg, WRITE);
1452 
1453 	if (tg->flags & THROTL_TG_PENDING) {
1454 		tg_update_disptime(tg);
1455 		throtl_schedule_next_dispatch(sq->parent_sq, true);
1456 	}
1457 }
1458 
tg_set_conf(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off,bool is_u64)1459 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1460 			   char *buf, size_t nbytes, loff_t off, bool is_u64)
1461 {
1462 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1463 	struct blkg_conf_ctx ctx;
1464 	struct throtl_grp *tg;
1465 	int ret;
1466 	u64 v;
1467 
1468 	ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1469 	if (ret)
1470 		return ret;
1471 
1472 	ret = -EINVAL;
1473 	if (sscanf(ctx.body, "%llu", &v) != 1)
1474 		goto out_finish;
1475 	if (!v)
1476 		v = U64_MAX;
1477 
1478 	tg = blkg_to_tg(ctx.blkg);
1479 
1480 	if (is_u64)
1481 		*(u64 *)((void *)tg + of_cft(of)->private) = v;
1482 	else
1483 		*(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1484 
1485 	tg_conf_updated(tg, false);
1486 	ret = 0;
1487 out_finish:
1488 	blkg_conf_finish(&ctx);
1489 	return ret ?: nbytes;
1490 }
1491 
tg_set_conf_u64(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)1492 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1493 			       char *buf, size_t nbytes, loff_t off)
1494 {
1495 	return tg_set_conf(of, buf, nbytes, off, true);
1496 }
1497 
tg_set_conf_uint(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)1498 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1499 				char *buf, size_t nbytes, loff_t off)
1500 {
1501 	return tg_set_conf(of, buf, nbytes, off, false);
1502 }
1503 
tg_print_rwstat(struct seq_file * sf,void * v)1504 static int tg_print_rwstat(struct seq_file *sf, void *v)
1505 {
1506 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1507 			  blkg_prfill_rwstat, &blkcg_policy_throtl,
1508 			  seq_cft(sf)->private, true);
1509 	return 0;
1510 }
1511 
tg_prfill_rwstat_recursive(struct seq_file * sf,struct blkg_policy_data * pd,int off)1512 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1513 				      struct blkg_policy_data *pd, int off)
1514 {
1515 	struct blkg_rwstat_sample sum;
1516 
1517 	blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1518 				  &sum);
1519 	return __blkg_prfill_rwstat(sf, pd, &sum);
1520 }
1521 
tg_print_rwstat_recursive(struct seq_file * sf,void * v)1522 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1523 {
1524 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1525 			  tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1526 			  seq_cft(sf)->private, true);
1527 	return 0;
1528 }
1529 
1530 static struct cftype throtl_legacy_files[] = {
1531 	{
1532 		.name = "throttle.read_bps_device",
1533 		.private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1534 		.seq_show = tg_print_conf_u64,
1535 		.write = tg_set_conf_u64,
1536 	},
1537 	{
1538 		.name = "throttle.write_bps_device",
1539 		.private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1540 		.seq_show = tg_print_conf_u64,
1541 		.write = tg_set_conf_u64,
1542 	},
1543 	{
1544 		.name = "throttle.read_iops_device",
1545 		.private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1546 		.seq_show = tg_print_conf_uint,
1547 		.write = tg_set_conf_uint,
1548 	},
1549 	{
1550 		.name = "throttle.write_iops_device",
1551 		.private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1552 		.seq_show = tg_print_conf_uint,
1553 		.write = tg_set_conf_uint,
1554 	},
1555 	{
1556 		.name = "throttle.io_service_bytes",
1557 		.private = offsetof(struct throtl_grp, stat_bytes),
1558 		.seq_show = tg_print_rwstat,
1559 	},
1560 	{
1561 		.name = "throttle.io_service_bytes_recursive",
1562 		.private = offsetof(struct throtl_grp, stat_bytes),
1563 		.seq_show = tg_print_rwstat_recursive,
1564 	},
1565 	{
1566 		.name = "throttle.io_serviced",
1567 		.private = offsetof(struct throtl_grp, stat_ios),
1568 		.seq_show = tg_print_rwstat,
1569 	},
1570 	{
1571 		.name = "throttle.io_serviced_recursive",
1572 		.private = offsetof(struct throtl_grp, stat_ios),
1573 		.seq_show = tg_print_rwstat_recursive,
1574 	},
1575 	{ }	/* terminate */
1576 };
1577 
tg_prfill_limit(struct seq_file * sf,struct blkg_policy_data * pd,int off)1578 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1579 			 int off)
1580 {
1581 	struct throtl_grp *tg = pd_to_tg(pd);
1582 	const char *dname = blkg_dev_name(pd->blkg);
1583 	char bufs[4][21] = { "max", "max", "max", "max" };
1584 	u64 bps_dft;
1585 	unsigned int iops_dft;
1586 	char idle_time[26] = "";
1587 	char latency_time[26] = "";
1588 
1589 	if (!dname)
1590 		return 0;
1591 
1592 	if (off == LIMIT_LOW) {
1593 		bps_dft = 0;
1594 		iops_dft = 0;
1595 	} else {
1596 		bps_dft = U64_MAX;
1597 		iops_dft = UINT_MAX;
1598 	}
1599 
1600 	if (tg->bps_conf[READ][off] == bps_dft &&
1601 	    tg->bps_conf[WRITE][off] == bps_dft &&
1602 	    tg->iops_conf[READ][off] == iops_dft &&
1603 	    tg->iops_conf[WRITE][off] == iops_dft &&
1604 	    (off != LIMIT_LOW ||
1605 	     (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1606 	      tg->latency_target_conf == DFL_LATENCY_TARGET)))
1607 		return 0;
1608 
1609 	if (tg->bps_conf[READ][off] != U64_MAX)
1610 		snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1611 			tg->bps_conf[READ][off]);
1612 	if (tg->bps_conf[WRITE][off] != U64_MAX)
1613 		snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1614 			tg->bps_conf[WRITE][off]);
1615 	if (tg->iops_conf[READ][off] != UINT_MAX)
1616 		snprintf(bufs[2], sizeof(bufs[2]), "%u",
1617 			tg->iops_conf[READ][off]);
1618 	if (tg->iops_conf[WRITE][off] != UINT_MAX)
1619 		snprintf(bufs[3], sizeof(bufs[3]), "%u",
1620 			tg->iops_conf[WRITE][off]);
1621 	if (off == LIMIT_LOW) {
1622 		if (tg->idletime_threshold_conf == ULONG_MAX)
1623 			strcpy(idle_time, " idle=max");
1624 		else
1625 			snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1626 				tg->idletime_threshold_conf);
1627 
1628 		if (tg->latency_target_conf == ULONG_MAX)
1629 			strcpy(latency_time, " latency=max");
1630 		else
1631 			snprintf(latency_time, sizeof(latency_time),
1632 				" latency=%lu", tg->latency_target_conf);
1633 	}
1634 
1635 	seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1636 		   dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1637 		   latency_time);
1638 	return 0;
1639 }
1640 
tg_print_limit(struct seq_file * sf,void * v)1641 static int tg_print_limit(struct seq_file *sf, void *v)
1642 {
1643 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1644 			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1645 	return 0;
1646 }
1647 
tg_set_limit(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)1648 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1649 			  char *buf, size_t nbytes, loff_t off)
1650 {
1651 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1652 	struct blkg_conf_ctx ctx;
1653 	struct throtl_grp *tg;
1654 	u64 v[4];
1655 	unsigned long idle_time;
1656 	unsigned long latency_time;
1657 	int ret;
1658 	int index = of_cft(of)->private;
1659 
1660 	ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1661 	if (ret)
1662 		return ret;
1663 
1664 	tg = blkg_to_tg(ctx.blkg);
1665 
1666 	v[0] = tg->bps_conf[READ][index];
1667 	v[1] = tg->bps_conf[WRITE][index];
1668 	v[2] = tg->iops_conf[READ][index];
1669 	v[3] = tg->iops_conf[WRITE][index];
1670 
1671 	idle_time = tg->idletime_threshold_conf;
1672 	latency_time = tg->latency_target_conf;
1673 	while (true) {
1674 		char tok[27];	/* wiops=18446744073709551616 */
1675 		char *p;
1676 		u64 val = U64_MAX;
1677 		int len;
1678 
1679 		if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1680 			break;
1681 		if (tok[0] == '\0')
1682 			break;
1683 		ctx.body += len;
1684 
1685 		ret = -EINVAL;
1686 		p = tok;
1687 		strsep(&p, "=");
1688 		if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1689 			goto out_finish;
1690 
1691 		ret = -ERANGE;
1692 		if (!val)
1693 			goto out_finish;
1694 
1695 		ret = -EINVAL;
1696 		if (!strcmp(tok, "rbps") && val > 1)
1697 			v[0] = val;
1698 		else if (!strcmp(tok, "wbps") && val > 1)
1699 			v[1] = val;
1700 		else if (!strcmp(tok, "riops") && val > 1)
1701 			v[2] = min_t(u64, val, UINT_MAX);
1702 		else if (!strcmp(tok, "wiops") && val > 1)
1703 			v[3] = min_t(u64, val, UINT_MAX);
1704 		else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1705 			idle_time = val;
1706 		else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1707 			latency_time = val;
1708 		else
1709 			goto out_finish;
1710 	}
1711 
1712 	tg->bps_conf[READ][index] = v[0];
1713 	tg->bps_conf[WRITE][index] = v[1];
1714 	tg->iops_conf[READ][index] = v[2];
1715 	tg->iops_conf[WRITE][index] = v[3];
1716 
1717 	if (index == LIMIT_MAX) {
1718 		tg->bps[READ][index] = v[0];
1719 		tg->bps[WRITE][index] = v[1];
1720 		tg->iops[READ][index] = v[2];
1721 		tg->iops[WRITE][index] = v[3];
1722 	}
1723 	tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1724 		tg->bps_conf[READ][LIMIT_MAX]);
1725 	tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1726 		tg->bps_conf[WRITE][LIMIT_MAX]);
1727 	tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1728 		tg->iops_conf[READ][LIMIT_MAX]);
1729 	tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1730 		tg->iops_conf[WRITE][LIMIT_MAX]);
1731 	tg->idletime_threshold_conf = idle_time;
1732 	tg->latency_target_conf = latency_time;
1733 
1734 	/* force user to configure all settings for low limit  */
1735 	if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1736 	      tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1737 	    tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1738 	    tg->latency_target_conf == DFL_LATENCY_TARGET) {
1739 		tg->bps[READ][LIMIT_LOW] = 0;
1740 		tg->bps[WRITE][LIMIT_LOW] = 0;
1741 		tg->iops[READ][LIMIT_LOW] = 0;
1742 		tg->iops[WRITE][LIMIT_LOW] = 0;
1743 		tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1744 		tg->latency_target = DFL_LATENCY_TARGET;
1745 	} else if (index == LIMIT_LOW) {
1746 		tg->idletime_threshold = tg->idletime_threshold_conf;
1747 		tg->latency_target = tg->latency_target_conf;
1748 	}
1749 
1750 	blk_throtl_update_limit_valid(tg->td);
1751 	if (tg->td->limit_valid[LIMIT_LOW]) {
1752 		if (index == LIMIT_LOW)
1753 			tg->td->limit_index = LIMIT_LOW;
1754 	} else
1755 		tg->td->limit_index = LIMIT_MAX;
1756 	tg_conf_updated(tg, index == LIMIT_LOW &&
1757 		tg->td->limit_valid[LIMIT_LOW]);
1758 	ret = 0;
1759 out_finish:
1760 	blkg_conf_finish(&ctx);
1761 	return ret ?: nbytes;
1762 }
1763 
1764 static struct cftype throtl_files[] = {
1765 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1766 	{
1767 		.name = "low",
1768 		.flags = CFTYPE_NOT_ON_ROOT,
1769 		.seq_show = tg_print_limit,
1770 		.write = tg_set_limit,
1771 		.private = LIMIT_LOW,
1772 	},
1773 #endif
1774 	{
1775 		.name = "max",
1776 		.flags = CFTYPE_NOT_ON_ROOT,
1777 		.seq_show = tg_print_limit,
1778 		.write = tg_set_limit,
1779 		.private = LIMIT_MAX,
1780 	},
1781 	{ }	/* terminate */
1782 };
1783 
throtl_shutdown_wq(struct request_queue * q)1784 static void throtl_shutdown_wq(struct request_queue *q)
1785 {
1786 	struct throtl_data *td = q->td;
1787 
1788 	cancel_work_sync(&td->dispatch_work);
1789 }
1790 
1791 static struct blkcg_policy blkcg_policy_throtl = {
1792 	.dfl_cftypes		= throtl_files,
1793 	.legacy_cftypes		= throtl_legacy_files,
1794 
1795 	.pd_alloc_fn		= throtl_pd_alloc,
1796 	.pd_init_fn		= throtl_pd_init,
1797 	.pd_online_fn		= throtl_pd_online,
1798 	.pd_offline_fn		= throtl_pd_offline,
1799 	.pd_free_fn		= throtl_pd_free,
1800 };
1801 
__tg_last_low_overflow_time(struct throtl_grp * tg)1802 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1803 {
1804 	unsigned long rtime = jiffies, wtime = jiffies;
1805 
1806 	if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1807 		rtime = tg->last_low_overflow_time[READ];
1808 	if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1809 		wtime = tg->last_low_overflow_time[WRITE];
1810 	return min(rtime, wtime);
1811 }
1812 
1813 /* tg should not be an intermediate node */
tg_last_low_overflow_time(struct throtl_grp * tg)1814 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1815 {
1816 	struct throtl_service_queue *parent_sq;
1817 	struct throtl_grp *parent = tg;
1818 	unsigned long ret = __tg_last_low_overflow_time(tg);
1819 
1820 	while (true) {
1821 		parent_sq = parent->service_queue.parent_sq;
1822 		parent = sq_to_tg(parent_sq);
1823 		if (!parent)
1824 			break;
1825 
1826 		/*
1827 		 * The parent doesn't have low limit, it always reaches low
1828 		 * limit. Its overflow time is useless for children
1829 		 */
1830 		if (!parent->bps[READ][LIMIT_LOW] &&
1831 		    !parent->iops[READ][LIMIT_LOW] &&
1832 		    !parent->bps[WRITE][LIMIT_LOW] &&
1833 		    !parent->iops[WRITE][LIMIT_LOW])
1834 			continue;
1835 		if (time_after(__tg_last_low_overflow_time(parent), ret))
1836 			ret = __tg_last_low_overflow_time(parent);
1837 	}
1838 	return ret;
1839 }
1840 
throtl_tg_is_idle(struct throtl_grp * tg)1841 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1842 {
1843 	/*
1844 	 * cgroup is idle if:
1845 	 * - single idle is too long, longer than a fixed value (in case user
1846 	 *   configure a too big threshold) or 4 times of idletime threshold
1847 	 * - average think time is more than threshold
1848 	 * - IO latency is largely below threshold
1849 	 */
1850 	unsigned long time;
1851 	bool ret;
1852 
1853 	time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1854 	ret = tg->latency_target == DFL_LATENCY_TARGET ||
1855 	      tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1856 	      (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1857 	      tg->avg_idletime > tg->idletime_threshold ||
1858 	      (tg->latency_target && tg->bio_cnt &&
1859 		tg->bad_bio_cnt * 5 < tg->bio_cnt);
1860 	throtl_log(&tg->service_queue,
1861 		"avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1862 		tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1863 		tg->bio_cnt, ret, tg->td->scale);
1864 	return ret;
1865 }
1866 
throtl_tg_can_upgrade(struct throtl_grp * tg)1867 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1868 {
1869 	struct throtl_service_queue *sq = &tg->service_queue;
1870 	bool read_limit, write_limit;
1871 
1872 	/*
1873 	 * if cgroup reaches low limit (if low limit is 0, the cgroup always
1874 	 * reaches), it's ok to upgrade to next limit
1875 	 */
1876 	read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1877 	write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1878 	if (!read_limit && !write_limit)
1879 		return true;
1880 	if (read_limit && sq->nr_queued[READ] &&
1881 	    (!write_limit || sq->nr_queued[WRITE]))
1882 		return true;
1883 	if (write_limit && sq->nr_queued[WRITE] &&
1884 	    (!read_limit || sq->nr_queued[READ]))
1885 		return true;
1886 
1887 	if (time_after_eq(jiffies,
1888 		tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1889 	    throtl_tg_is_idle(tg))
1890 		return true;
1891 	return false;
1892 }
1893 
throtl_hierarchy_can_upgrade(struct throtl_grp * tg)1894 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1895 {
1896 	while (true) {
1897 		if (throtl_tg_can_upgrade(tg))
1898 			return true;
1899 		tg = sq_to_tg(tg->service_queue.parent_sq);
1900 		if (!tg || !tg_to_blkg(tg)->parent)
1901 			return false;
1902 	}
1903 	return false;
1904 }
1905 
throtl_can_upgrade(struct throtl_data * td,struct throtl_grp * this_tg)1906 static bool throtl_can_upgrade(struct throtl_data *td,
1907 	struct throtl_grp *this_tg)
1908 {
1909 	struct cgroup_subsys_state *pos_css;
1910 	struct blkcg_gq *blkg;
1911 
1912 	if (td->limit_index != LIMIT_LOW)
1913 		return false;
1914 
1915 	if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1916 		return false;
1917 
1918 	rcu_read_lock();
1919 	blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1920 		struct throtl_grp *tg = blkg_to_tg(blkg);
1921 
1922 		if (tg == this_tg)
1923 			continue;
1924 		if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1925 			continue;
1926 		if (!throtl_hierarchy_can_upgrade(tg)) {
1927 			rcu_read_unlock();
1928 			return false;
1929 		}
1930 	}
1931 	rcu_read_unlock();
1932 	return true;
1933 }
1934 
throtl_upgrade_check(struct throtl_grp * tg)1935 static void throtl_upgrade_check(struct throtl_grp *tg)
1936 {
1937 	unsigned long now = jiffies;
1938 
1939 	if (tg->td->limit_index != LIMIT_LOW)
1940 		return;
1941 
1942 	if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1943 		return;
1944 
1945 	tg->last_check_time = now;
1946 
1947 	if (!time_after_eq(now,
1948 	     __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1949 		return;
1950 
1951 	if (throtl_can_upgrade(tg->td, NULL))
1952 		throtl_upgrade_state(tg->td);
1953 }
1954 
throtl_upgrade_state(struct throtl_data * td)1955 static void throtl_upgrade_state(struct throtl_data *td)
1956 {
1957 	struct cgroup_subsys_state *pos_css;
1958 	struct blkcg_gq *blkg;
1959 
1960 	throtl_log(&td->service_queue, "upgrade to max");
1961 	td->limit_index = LIMIT_MAX;
1962 	td->low_upgrade_time = jiffies;
1963 	td->scale = 0;
1964 	rcu_read_lock();
1965 	blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1966 		struct throtl_grp *tg = blkg_to_tg(blkg);
1967 		struct throtl_service_queue *sq = &tg->service_queue;
1968 
1969 		tg->disptime = jiffies - 1;
1970 		throtl_select_dispatch(sq);
1971 		throtl_schedule_next_dispatch(sq, true);
1972 	}
1973 	rcu_read_unlock();
1974 	throtl_select_dispatch(&td->service_queue);
1975 	throtl_schedule_next_dispatch(&td->service_queue, true);
1976 	queue_work(kthrotld_workqueue, &td->dispatch_work);
1977 }
1978 
throtl_downgrade_state(struct throtl_data * td)1979 static void throtl_downgrade_state(struct throtl_data *td)
1980 {
1981 	td->scale /= 2;
1982 
1983 	throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1984 	if (td->scale) {
1985 		td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1986 		return;
1987 	}
1988 
1989 	td->limit_index = LIMIT_LOW;
1990 	td->low_downgrade_time = jiffies;
1991 }
1992 
throtl_tg_can_downgrade(struct throtl_grp * tg)1993 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1994 {
1995 	struct throtl_data *td = tg->td;
1996 	unsigned long now = jiffies;
1997 
1998 	/*
1999 	 * If cgroup is below low limit, consider downgrade and throttle other
2000 	 * cgroups
2001 	 */
2002 	if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
2003 	    time_after_eq(now, tg_last_low_overflow_time(tg) +
2004 					td->throtl_slice) &&
2005 	    (!throtl_tg_is_idle(tg) ||
2006 	     !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
2007 		return true;
2008 	return false;
2009 }
2010 
throtl_hierarchy_can_downgrade(struct throtl_grp * tg)2011 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
2012 {
2013 	while (true) {
2014 		if (!throtl_tg_can_downgrade(tg))
2015 			return false;
2016 		tg = sq_to_tg(tg->service_queue.parent_sq);
2017 		if (!tg || !tg_to_blkg(tg)->parent)
2018 			break;
2019 	}
2020 	return true;
2021 }
2022 
throtl_downgrade_check(struct throtl_grp * tg)2023 static void throtl_downgrade_check(struct throtl_grp *tg)
2024 {
2025 	uint64_t bps;
2026 	unsigned int iops;
2027 	unsigned long elapsed_time;
2028 	unsigned long now = jiffies;
2029 
2030 	if (tg->td->limit_index != LIMIT_MAX ||
2031 	    !tg->td->limit_valid[LIMIT_LOW])
2032 		return;
2033 	if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
2034 		return;
2035 	if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
2036 		return;
2037 
2038 	elapsed_time = now - tg->last_check_time;
2039 	tg->last_check_time = now;
2040 
2041 	if (time_before(now, tg_last_low_overflow_time(tg) +
2042 			tg->td->throtl_slice))
2043 		return;
2044 
2045 	if (tg->bps[READ][LIMIT_LOW]) {
2046 		bps = tg->last_bytes_disp[READ] * HZ;
2047 		do_div(bps, elapsed_time);
2048 		if (bps >= tg->bps[READ][LIMIT_LOW])
2049 			tg->last_low_overflow_time[READ] = now;
2050 	}
2051 
2052 	if (tg->bps[WRITE][LIMIT_LOW]) {
2053 		bps = tg->last_bytes_disp[WRITE] * HZ;
2054 		do_div(bps, elapsed_time);
2055 		if (bps >= tg->bps[WRITE][LIMIT_LOW])
2056 			tg->last_low_overflow_time[WRITE] = now;
2057 	}
2058 
2059 	if (tg->iops[READ][LIMIT_LOW]) {
2060 		tg->last_io_disp[READ] += atomic_xchg(&tg->last_io_split_cnt[READ], 0);
2061 		iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2062 		if (iops >= tg->iops[READ][LIMIT_LOW])
2063 			tg->last_low_overflow_time[READ] = now;
2064 	}
2065 
2066 	if (tg->iops[WRITE][LIMIT_LOW]) {
2067 		tg->last_io_disp[WRITE] += atomic_xchg(&tg->last_io_split_cnt[WRITE], 0);
2068 		iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2069 		if (iops >= tg->iops[WRITE][LIMIT_LOW])
2070 			tg->last_low_overflow_time[WRITE] = now;
2071 	}
2072 
2073 	/*
2074 	 * If cgroup is below low limit, consider downgrade and throttle other
2075 	 * cgroups
2076 	 */
2077 	if (throtl_hierarchy_can_downgrade(tg))
2078 		throtl_downgrade_state(tg->td);
2079 
2080 	tg->last_bytes_disp[READ] = 0;
2081 	tg->last_bytes_disp[WRITE] = 0;
2082 	tg->last_io_disp[READ] = 0;
2083 	tg->last_io_disp[WRITE] = 0;
2084 }
2085 
blk_throtl_update_idletime(struct throtl_grp * tg)2086 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2087 {
2088 	unsigned long now;
2089 	unsigned long last_finish_time = tg->last_finish_time;
2090 
2091 	if (last_finish_time == 0)
2092 		return;
2093 
2094 	now = ktime_get_ns() >> 10;
2095 	if (now <= last_finish_time ||
2096 	    last_finish_time == tg->checked_last_finish_time)
2097 		return;
2098 
2099 	tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2100 	tg->checked_last_finish_time = last_finish_time;
2101 }
2102 
2103 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
throtl_update_latency_buckets(struct throtl_data * td)2104 static void throtl_update_latency_buckets(struct throtl_data *td)
2105 {
2106 	struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2107 	int i, cpu, rw;
2108 	unsigned long last_latency[2] = { 0 };
2109 	unsigned long latency[2];
2110 
2111 	if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
2112 		return;
2113 	if (time_before(jiffies, td->last_calculate_time + HZ))
2114 		return;
2115 	td->last_calculate_time = jiffies;
2116 
2117 	memset(avg_latency, 0, sizeof(avg_latency));
2118 	for (rw = READ; rw <= WRITE; rw++) {
2119 		for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2120 			struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2121 
2122 			for_each_possible_cpu(cpu) {
2123 				struct latency_bucket *bucket;
2124 
2125 				/* this isn't race free, but ok in practice */
2126 				bucket = per_cpu_ptr(td->latency_buckets[rw],
2127 					cpu);
2128 				tmp->total_latency += bucket[i].total_latency;
2129 				tmp->samples += bucket[i].samples;
2130 				bucket[i].total_latency = 0;
2131 				bucket[i].samples = 0;
2132 			}
2133 
2134 			if (tmp->samples >= 32) {
2135 				int samples = tmp->samples;
2136 
2137 				latency[rw] = tmp->total_latency;
2138 
2139 				tmp->total_latency = 0;
2140 				tmp->samples = 0;
2141 				latency[rw] /= samples;
2142 				if (latency[rw] == 0)
2143 					continue;
2144 				avg_latency[rw][i].latency = latency[rw];
2145 			}
2146 		}
2147 	}
2148 
2149 	for (rw = READ; rw <= WRITE; rw++) {
2150 		for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2151 			if (!avg_latency[rw][i].latency) {
2152 				if (td->avg_buckets[rw][i].latency < last_latency[rw])
2153 					td->avg_buckets[rw][i].latency =
2154 						last_latency[rw];
2155 				continue;
2156 			}
2157 
2158 			if (!td->avg_buckets[rw][i].valid)
2159 				latency[rw] = avg_latency[rw][i].latency;
2160 			else
2161 				latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2162 					avg_latency[rw][i].latency) >> 3;
2163 
2164 			td->avg_buckets[rw][i].latency = max(latency[rw],
2165 				last_latency[rw]);
2166 			td->avg_buckets[rw][i].valid = true;
2167 			last_latency[rw] = td->avg_buckets[rw][i].latency;
2168 		}
2169 	}
2170 
2171 	for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2172 		throtl_log(&td->service_queue,
2173 			"Latency bucket %d: read latency=%ld, read valid=%d, "
2174 			"write latency=%ld, write valid=%d", i,
2175 			td->avg_buckets[READ][i].latency,
2176 			td->avg_buckets[READ][i].valid,
2177 			td->avg_buckets[WRITE][i].latency,
2178 			td->avg_buckets[WRITE][i].valid);
2179 }
2180 #else
throtl_update_latency_buckets(struct throtl_data * td)2181 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2182 {
2183 }
2184 #endif
2185 
blk_throtl_charge_bio_split(struct bio * bio)2186 void blk_throtl_charge_bio_split(struct bio *bio)
2187 {
2188 	struct blkcg_gq *blkg = bio->bi_blkg;
2189 	struct throtl_grp *parent = blkg_to_tg(blkg);
2190 	struct throtl_service_queue *parent_sq;
2191 	bool rw = bio_data_dir(bio);
2192 
2193 	do {
2194 		if (!parent->has_rules[rw])
2195 			break;
2196 
2197 		atomic_inc(&parent->io_split_cnt[rw]);
2198 		atomic_inc(&parent->last_io_split_cnt[rw]);
2199 
2200 		parent_sq = parent->service_queue.parent_sq;
2201 		parent = sq_to_tg(parent_sq);
2202 	} while (parent);
2203 }
2204 
blk_throtl_bio(struct bio * bio)2205 bool blk_throtl_bio(struct bio *bio)
2206 {
2207 	struct request_queue *q = bio->bi_disk->queue;
2208 	struct blkcg_gq *blkg = bio->bi_blkg;
2209 	struct throtl_qnode *qn = NULL;
2210 	struct throtl_grp *tg = blkg_to_tg(blkg);
2211 	struct throtl_service_queue *sq;
2212 	bool rw = bio_data_dir(bio);
2213 	bool throttled = false;
2214 	struct throtl_data *td = tg->td;
2215 
2216 	rcu_read_lock();
2217 
2218 	/* see throtl_charge_bio() */
2219 	if (bio_flagged(bio, BIO_THROTTLED))
2220 		goto out;
2221 
2222 	if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
2223 		blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
2224 				bio->bi_iter.bi_size);
2225 		blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
2226 	}
2227 
2228 	if (!tg->has_rules[rw])
2229 		goto out;
2230 
2231 	spin_lock_irq(&q->queue_lock);
2232 
2233 	throtl_update_latency_buckets(td);
2234 
2235 	blk_throtl_update_idletime(tg);
2236 
2237 	sq = &tg->service_queue;
2238 
2239 again:
2240 	while (true) {
2241 		if (tg->last_low_overflow_time[rw] == 0)
2242 			tg->last_low_overflow_time[rw] = jiffies;
2243 		throtl_downgrade_check(tg);
2244 		throtl_upgrade_check(tg);
2245 		/* throtl is FIFO - if bios are already queued, should queue */
2246 		if (sq->nr_queued[rw])
2247 			break;
2248 
2249 		/* if above limits, break to queue */
2250 		if (!tg_may_dispatch(tg, bio, NULL)) {
2251 			tg->last_low_overflow_time[rw] = jiffies;
2252 			if (throtl_can_upgrade(td, tg)) {
2253 				throtl_upgrade_state(td);
2254 				goto again;
2255 			}
2256 			break;
2257 		}
2258 
2259 		/* within limits, let's charge and dispatch directly */
2260 		throtl_charge_bio(tg, bio);
2261 
2262 		/*
2263 		 * We need to trim slice even when bios are not being queued
2264 		 * otherwise it might happen that a bio is not queued for
2265 		 * a long time and slice keeps on extending and trim is not
2266 		 * called for a long time. Now if limits are reduced suddenly
2267 		 * we take into account all the IO dispatched so far at new
2268 		 * low rate and * newly queued IO gets a really long dispatch
2269 		 * time.
2270 		 *
2271 		 * So keep on trimming slice even if bio is not queued.
2272 		 */
2273 		throtl_trim_slice(tg, rw);
2274 
2275 		/*
2276 		 * @bio passed through this layer without being throttled.
2277 		 * Climb up the ladder.  If we're already at the top, it
2278 		 * can be executed directly.
2279 		 */
2280 		qn = &tg->qnode_on_parent[rw];
2281 		sq = sq->parent_sq;
2282 		tg = sq_to_tg(sq);
2283 		if (!tg)
2284 			goto out_unlock;
2285 	}
2286 
2287 	/* out-of-limit, queue to @tg */
2288 	throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2289 		   rw == READ ? 'R' : 'W',
2290 		   tg->bytes_disp[rw], bio->bi_iter.bi_size,
2291 		   tg_bps_limit(tg, rw),
2292 		   tg->io_disp[rw], tg_iops_limit(tg, rw),
2293 		   sq->nr_queued[READ], sq->nr_queued[WRITE]);
2294 
2295 	tg->last_low_overflow_time[rw] = jiffies;
2296 
2297 	td->nr_queued[rw]++;
2298 	throtl_add_bio_tg(bio, qn, tg);
2299 	throttled = true;
2300 
2301 	/*
2302 	 * Update @tg's dispatch time and force schedule dispatch if @tg
2303 	 * was empty before @bio.  The forced scheduling isn't likely to
2304 	 * cause undue delay as @bio is likely to be dispatched directly if
2305 	 * its @tg's disptime is not in the future.
2306 	 */
2307 	if (tg->flags & THROTL_TG_WAS_EMPTY) {
2308 		tg_update_disptime(tg);
2309 		throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2310 	}
2311 
2312 out_unlock:
2313 	spin_unlock_irq(&q->queue_lock);
2314 out:
2315 	bio_set_flag(bio, BIO_THROTTLED);
2316 
2317 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2318 	if (throttled || !td->track_bio_latency)
2319 		bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2320 #endif
2321 	rcu_read_unlock();
2322 	return throttled;
2323 }
2324 
2325 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
throtl_track_latency(struct throtl_data * td,sector_t size,int op,unsigned long time)2326 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2327 	int op, unsigned long time)
2328 {
2329 	struct latency_bucket *latency;
2330 	int index;
2331 
2332 	if (!td || td->limit_index != LIMIT_LOW ||
2333 	    !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2334 	    !blk_queue_nonrot(td->queue))
2335 		return;
2336 
2337 	index = request_bucket_index(size);
2338 
2339 	latency = get_cpu_ptr(td->latency_buckets[op]);
2340 	latency[index].total_latency += time;
2341 	latency[index].samples++;
2342 	put_cpu_ptr(td->latency_buckets[op]);
2343 }
2344 
blk_throtl_stat_add(struct request * rq,u64 time_ns)2345 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2346 {
2347 	struct request_queue *q = rq->q;
2348 	struct throtl_data *td = q->td;
2349 
2350 	throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2351 			     time_ns >> 10);
2352 }
2353 
blk_throtl_bio_endio(struct bio * bio)2354 void blk_throtl_bio_endio(struct bio *bio)
2355 {
2356 	struct blkcg_gq *blkg;
2357 	struct throtl_grp *tg;
2358 	u64 finish_time_ns;
2359 	unsigned long finish_time;
2360 	unsigned long start_time;
2361 	unsigned long lat;
2362 	int rw = bio_data_dir(bio);
2363 
2364 	blkg = bio->bi_blkg;
2365 	if (!blkg)
2366 		return;
2367 	tg = blkg_to_tg(blkg);
2368 	if (!tg->td->limit_valid[LIMIT_LOW])
2369 		return;
2370 
2371 	finish_time_ns = ktime_get_ns();
2372 	tg->last_finish_time = finish_time_ns >> 10;
2373 
2374 	start_time = bio_issue_time(&bio->bi_issue) >> 10;
2375 	finish_time = __bio_issue_time(finish_time_ns) >> 10;
2376 	if (!start_time || finish_time <= start_time)
2377 		return;
2378 
2379 	lat = finish_time - start_time;
2380 	/* this is only for bio based driver */
2381 	if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2382 		throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2383 				     bio_op(bio), lat);
2384 
2385 	if (tg->latency_target && lat >= tg->td->filtered_latency) {
2386 		int bucket;
2387 		unsigned int threshold;
2388 
2389 		bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2390 		threshold = tg->td->avg_buckets[rw][bucket].latency +
2391 			tg->latency_target;
2392 		if (lat > threshold)
2393 			tg->bad_bio_cnt++;
2394 		/*
2395 		 * Not race free, could get wrong count, which means cgroups
2396 		 * will be throttled
2397 		 */
2398 		tg->bio_cnt++;
2399 	}
2400 
2401 	if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2402 		tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2403 		tg->bio_cnt /= 2;
2404 		tg->bad_bio_cnt /= 2;
2405 	}
2406 }
2407 #endif
2408 
blk_throtl_init(struct request_queue * q)2409 int blk_throtl_init(struct request_queue *q)
2410 {
2411 	struct throtl_data *td;
2412 	int ret;
2413 
2414 	td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2415 	if (!td)
2416 		return -ENOMEM;
2417 	td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2418 		LATENCY_BUCKET_SIZE, __alignof__(u64));
2419 	if (!td->latency_buckets[READ]) {
2420 		kfree(td);
2421 		return -ENOMEM;
2422 	}
2423 	td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2424 		LATENCY_BUCKET_SIZE, __alignof__(u64));
2425 	if (!td->latency_buckets[WRITE]) {
2426 		free_percpu(td->latency_buckets[READ]);
2427 		kfree(td);
2428 		return -ENOMEM;
2429 	}
2430 
2431 	INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2432 	throtl_service_queue_init(&td->service_queue);
2433 
2434 	q->td = td;
2435 	td->queue = q;
2436 
2437 	td->limit_valid[LIMIT_MAX] = true;
2438 	td->limit_index = LIMIT_MAX;
2439 	td->low_upgrade_time = jiffies;
2440 	td->low_downgrade_time = jiffies;
2441 
2442 	/* activate policy */
2443 	ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2444 	if (ret) {
2445 		free_percpu(td->latency_buckets[READ]);
2446 		free_percpu(td->latency_buckets[WRITE]);
2447 		kfree(td);
2448 	}
2449 	return ret;
2450 }
2451 
blk_throtl_exit(struct request_queue * q)2452 void blk_throtl_exit(struct request_queue *q)
2453 {
2454 	BUG_ON(!q->td);
2455 	del_timer_sync(&q->td->service_queue.pending_timer);
2456 	throtl_shutdown_wq(q);
2457 	blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2458 	free_percpu(q->td->latency_buckets[READ]);
2459 	free_percpu(q->td->latency_buckets[WRITE]);
2460 	kfree(q->td);
2461 }
2462 
blk_throtl_register_queue(struct request_queue * q)2463 void blk_throtl_register_queue(struct request_queue *q)
2464 {
2465 	struct throtl_data *td;
2466 	int i;
2467 
2468 	td = q->td;
2469 	BUG_ON(!td);
2470 
2471 	if (blk_queue_nonrot(q)) {
2472 		td->throtl_slice = DFL_THROTL_SLICE_SSD;
2473 		td->filtered_latency = LATENCY_FILTERED_SSD;
2474 	} else {
2475 		td->throtl_slice = DFL_THROTL_SLICE_HD;
2476 		td->filtered_latency = LATENCY_FILTERED_HD;
2477 		for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2478 			td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2479 			td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2480 		}
2481 	}
2482 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2483 	/* if no low limit, use previous default */
2484 	td->throtl_slice = DFL_THROTL_SLICE_HD;
2485 #endif
2486 
2487 	td->track_bio_latency = !queue_is_mq(q);
2488 	if (!td->track_bio_latency)
2489 		blk_stat_enable_accounting(q);
2490 }
2491 
2492 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
blk_throtl_sample_time_show(struct request_queue * q,char * page)2493 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2494 {
2495 	if (!q->td)
2496 		return -EINVAL;
2497 	return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2498 }
2499 
blk_throtl_sample_time_store(struct request_queue * q,const char * page,size_t count)2500 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2501 	const char *page, size_t count)
2502 {
2503 	unsigned long v;
2504 	unsigned long t;
2505 
2506 	if (!q->td)
2507 		return -EINVAL;
2508 	if (kstrtoul(page, 10, &v))
2509 		return -EINVAL;
2510 	t = msecs_to_jiffies(v);
2511 	if (t == 0 || t > MAX_THROTL_SLICE)
2512 		return -EINVAL;
2513 	q->td->throtl_slice = t;
2514 	return count;
2515 }
2516 #endif
2517 
throtl_init(void)2518 static int __init throtl_init(void)
2519 {
2520 	kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2521 	if (!kthrotld_workqueue)
2522 		panic("Failed to create kthrotld\n");
2523 
2524 	return blkcg_policy_register(&blkcg_policy_throtl);
2525 }
2526 
2527 module_init(throtl_init);
2528