• 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 
593 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
blk_throtl_update_limit_valid(struct throtl_data * td)594 static void blk_throtl_update_limit_valid(struct throtl_data *td)
595 {
596 	struct cgroup_subsys_state *pos_css;
597 	struct blkcg_gq *blkg;
598 	bool low_valid = false;
599 
600 	rcu_read_lock();
601 	blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
602 		struct throtl_grp *tg = blkg_to_tg(blkg);
603 
604 		if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
605 		    tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
606 			low_valid = true;
607 			break;
608 		}
609 	}
610 	rcu_read_unlock();
611 
612 	td->limit_valid[LIMIT_LOW] = low_valid;
613 }
614 #else
blk_throtl_update_limit_valid(struct throtl_data * td)615 static inline void blk_throtl_update_limit_valid(struct throtl_data *td)
616 {
617 }
618 #endif
619 
620 static void throtl_upgrade_state(struct throtl_data *td);
throtl_pd_offline(struct blkg_policy_data * pd)621 static void throtl_pd_offline(struct blkg_policy_data *pd)
622 {
623 	struct throtl_grp *tg = pd_to_tg(pd);
624 
625 	tg->bps[READ][LIMIT_LOW] = 0;
626 	tg->bps[WRITE][LIMIT_LOW] = 0;
627 	tg->iops[READ][LIMIT_LOW] = 0;
628 	tg->iops[WRITE][LIMIT_LOW] = 0;
629 
630 	blk_throtl_update_limit_valid(tg->td);
631 
632 	if (!tg->td->limit_valid[tg->td->limit_index])
633 		throtl_upgrade_state(tg->td);
634 }
635 
throtl_pd_free(struct blkg_policy_data * pd)636 static void throtl_pd_free(struct blkg_policy_data *pd)
637 {
638 	struct throtl_grp *tg = pd_to_tg(pd);
639 
640 	del_timer_sync(&tg->service_queue.pending_timer);
641 	blkg_rwstat_exit(&tg->stat_bytes);
642 	blkg_rwstat_exit(&tg->stat_ios);
643 	kfree(tg);
644 }
645 
646 static struct throtl_grp *
throtl_rb_first(struct throtl_service_queue * parent_sq)647 throtl_rb_first(struct throtl_service_queue *parent_sq)
648 {
649 	struct rb_node *n;
650 
651 	n = rb_first_cached(&parent_sq->pending_tree);
652 	WARN_ON_ONCE(!n);
653 	if (!n)
654 		return NULL;
655 	return rb_entry_tg(n);
656 }
657 
throtl_rb_erase(struct rb_node * n,struct throtl_service_queue * parent_sq)658 static void throtl_rb_erase(struct rb_node *n,
659 			    struct throtl_service_queue *parent_sq)
660 {
661 	rb_erase_cached(n, &parent_sq->pending_tree);
662 	RB_CLEAR_NODE(n);
663 	--parent_sq->nr_pending;
664 }
665 
update_min_dispatch_time(struct throtl_service_queue * parent_sq)666 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
667 {
668 	struct throtl_grp *tg;
669 
670 	tg = throtl_rb_first(parent_sq);
671 	if (!tg)
672 		return;
673 
674 	parent_sq->first_pending_disptime = tg->disptime;
675 }
676 
tg_service_queue_add(struct throtl_grp * tg)677 static void tg_service_queue_add(struct throtl_grp *tg)
678 {
679 	struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
680 	struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
681 	struct rb_node *parent = NULL;
682 	struct throtl_grp *__tg;
683 	unsigned long key = tg->disptime;
684 	bool leftmost = true;
685 
686 	while (*node != NULL) {
687 		parent = *node;
688 		__tg = rb_entry_tg(parent);
689 
690 		if (time_before(key, __tg->disptime))
691 			node = &parent->rb_left;
692 		else {
693 			node = &parent->rb_right;
694 			leftmost = false;
695 		}
696 	}
697 
698 	rb_link_node(&tg->rb_node, parent, node);
699 	rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
700 			       leftmost);
701 }
702 
throtl_enqueue_tg(struct throtl_grp * tg)703 static void throtl_enqueue_tg(struct throtl_grp *tg)
704 {
705 	if (!(tg->flags & THROTL_TG_PENDING)) {
706 		tg_service_queue_add(tg);
707 		tg->flags |= THROTL_TG_PENDING;
708 		tg->service_queue.parent_sq->nr_pending++;
709 	}
710 }
711 
throtl_dequeue_tg(struct throtl_grp * tg)712 static void throtl_dequeue_tg(struct throtl_grp *tg)
713 {
714 	if (tg->flags & THROTL_TG_PENDING) {
715 		throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
716 		tg->flags &= ~THROTL_TG_PENDING;
717 	}
718 }
719 
720 /* Call with queue lock held */
throtl_schedule_pending_timer(struct throtl_service_queue * sq,unsigned long expires)721 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
722 					  unsigned long expires)
723 {
724 	unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
725 
726 	/*
727 	 * Since we are adjusting the throttle limit dynamically, the sleep
728 	 * time calculated according to previous limit might be invalid. It's
729 	 * possible the cgroup sleep time is very long and no other cgroups
730 	 * have IO running so notify the limit changes. Make sure the cgroup
731 	 * doesn't sleep too long to avoid the missed notification.
732 	 */
733 	if (time_after(expires, max_expire))
734 		expires = max_expire;
735 	mod_timer(&sq->pending_timer, expires);
736 	throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
737 		   expires - jiffies, jiffies);
738 }
739 
740 /**
741  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
742  * @sq: the service_queue to schedule dispatch for
743  * @force: force scheduling
744  *
745  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
746  * dispatch time of the first pending child.  Returns %true if either timer
747  * is armed or there's no pending child left.  %false if the current
748  * dispatch window is still open and the caller should continue
749  * dispatching.
750  *
751  * If @force is %true, the dispatch timer is always scheduled and this
752  * function is guaranteed to return %true.  This is to be used when the
753  * caller can't dispatch itself and needs to invoke pending_timer
754  * unconditionally.  Note that forced scheduling is likely to induce short
755  * delay before dispatch starts even if @sq->first_pending_disptime is not
756  * in the future and thus shouldn't be used in hot paths.
757  */
throtl_schedule_next_dispatch(struct throtl_service_queue * sq,bool force)758 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
759 					  bool force)
760 {
761 	/* any pending children left? */
762 	if (!sq->nr_pending)
763 		return true;
764 
765 	update_min_dispatch_time(sq);
766 
767 	/* is the next dispatch time in the future? */
768 	if (force || time_after(sq->first_pending_disptime, jiffies)) {
769 		throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
770 		return true;
771 	}
772 
773 	/* tell the caller to continue dispatching */
774 	return false;
775 }
776 
throtl_start_new_slice_with_credit(struct throtl_grp * tg,bool rw,unsigned long start)777 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
778 		bool rw, unsigned long start)
779 {
780 	tg->bytes_disp[rw] = 0;
781 	tg->io_disp[rw] = 0;
782 
783 	atomic_set(&tg->io_split_cnt[rw], 0);
784 
785 	/*
786 	 * Previous slice has expired. We must have trimmed it after last
787 	 * bio dispatch. That means since start of last slice, we never used
788 	 * that bandwidth. Do try to make use of that bandwidth while giving
789 	 * credit.
790 	 */
791 	if (time_after_eq(start, tg->slice_start[rw]))
792 		tg->slice_start[rw] = start;
793 
794 	tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
795 	throtl_log(&tg->service_queue,
796 		   "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
797 		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
798 		   tg->slice_end[rw], jiffies);
799 }
800 
throtl_start_new_slice(struct throtl_grp * tg,bool rw)801 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
802 {
803 	tg->bytes_disp[rw] = 0;
804 	tg->io_disp[rw] = 0;
805 	tg->slice_start[rw] = jiffies;
806 	tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
807 
808 	atomic_set(&tg->io_split_cnt[rw], 0);
809 
810 	throtl_log(&tg->service_queue,
811 		   "[%c] new slice start=%lu end=%lu jiffies=%lu",
812 		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
813 		   tg->slice_end[rw], jiffies);
814 }
815 
throtl_set_slice_end(struct throtl_grp * tg,bool rw,unsigned long jiffy_end)816 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
817 					unsigned long jiffy_end)
818 {
819 	tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
820 }
821 
throtl_extend_slice(struct throtl_grp * tg,bool rw,unsigned long jiffy_end)822 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
823 				       unsigned long jiffy_end)
824 {
825 	throtl_set_slice_end(tg, rw, jiffy_end);
826 	throtl_log(&tg->service_queue,
827 		   "[%c] extend slice start=%lu end=%lu jiffies=%lu",
828 		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
829 		   tg->slice_end[rw], jiffies);
830 }
831 
832 /* Determine if previously allocated or extended slice is complete or not */
throtl_slice_used(struct throtl_grp * tg,bool rw)833 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
834 {
835 	if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
836 		return false;
837 
838 	return true;
839 }
840 
841 /* Trim the used slices and adjust slice start accordingly */
throtl_trim_slice(struct throtl_grp * tg,bool rw)842 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
843 {
844 	unsigned long nr_slices, time_elapsed, io_trim;
845 	u64 bytes_trim, tmp;
846 
847 	BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
848 
849 	/*
850 	 * If bps are unlimited (-1), then time slice don't get
851 	 * renewed. Don't try to trim the slice if slice is used. A new
852 	 * slice will start when appropriate.
853 	 */
854 	if (throtl_slice_used(tg, rw))
855 		return;
856 
857 	/*
858 	 * A bio has been dispatched. Also adjust slice_end. It might happen
859 	 * that initially cgroup limit was very low resulting in high
860 	 * slice_end, but later limit was bumped up and bio was dispatched
861 	 * sooner, then we need to reduce slice_end. A high bogus slice_end
862 	 * is bad because it does not allow new slice to start.
863 	 */
864 
865 	throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
866 
867 	time_elapsed = jiffies - tg->slice_start[rw];
868 
869 	nr_slices = time_elapsed / tg->td->throtl_slice;
870 
871 	if (!nr_slices)
872 		return;
873 	tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
874 	do_div(tmp, HZ);
875 	bytes_trim = tmp;
876 
877 	io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
878 		HZ;
879 
880 	if (!bytes_trim && !io_trim)
881 		return;
882 
883 	if (tg->bytes_disp[rw] >= bytes_trim)
884 		tg->bytes_disp[rw] -= bytes_trim;
885 	else
886 		tg->bytes_disp[rw] = 0;
887 
888 	if (tg->io_disp[rw] >= io_trim)
889 		tg->io_disp[rw] -= io_trim;
890 	else
891 		tg->io_disp[rw] = 0;
892 
893 	tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
894 
895 	throtl_log(&tg->service_queue,
896 		   "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
897 		   rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
898 		   tg->slice_start[rw], tg->slice_end[rw], jiffies);
899 }
900 
tg_with_in_iops_limit(struct throtl_grp * tg,struct bio * bio,u32 iops_limit,unsigned long * wait)901 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
902 				  u32 iops_limit, unsigned long *wait)
903 {
904 	bool rw = bio_data_dir(bio);
905 	unsigned int io_allowed;
906 	unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
907 	u64 tmp;
908 
909 	if (iops_limit == UINT_MAX) {
910 		if (wait)
911 			*wait = 0;
912 		return true;
913 	}
914 
915 	jiffy_elapsed = jiffies - tg->slice_start[rw];
916 
917 	/* Round up to the next throttle slice, wait time must be nonzero */
918 	jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
919 
920 	/*
921 	 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
922 	 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
923 	 * will allow dispatch after 1 second and after that slice should
924 	 * have been trimmed.
925 	 */
926 
927 	tmp = (u64)iops_limit * jiffy_elapsed_rnd;
928 	do_div(tmp, HZ);
929 
930 	if (tmp > UINT_MAX)
931 		io_allowed = UINT_MAX;
932 	else
933 		io_allowed = tmp;
934 
935 	if (tg->io_disp[rw] + 1 <= io_allowed) {
936 		if (wait)
937 			*wait = 0;
938 		return true;
939 	}
940 
941 	/* Calc approx time to dispatch */
942 	jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
943 
944 	if (wait)
945 		*wait = jiffy_wait;
946 	return false;
947 }
948 
tg_with_in_bps_limit(struct throtl_grp * tg,struct bio * bio,u64 bps_limit,unsigned long * wait)949 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
950 				 u64 bps_limit, unsigned long *wait)
951 {
952 	bool rw = bio_data_dir(bio);
953 	u64 bytes_allowed, extra_bytes;
954 	unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
955 	unsigned int bio_size = throtl_bio_data_size(bio);
956 
957 	if (bps_limit == U64_MAX) {
958 		if (wait)
959 			*wait = 0;
960 		return true;
961 	}
962 
963 	jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
964 
965 	/* Slice has just started. Consider one slice interval */
966 	if (!jiffy_elapsed)
967 		jiffy_elapsed_rnd = tg->td->throtl_slice;
968 
969 	jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
970 	bytes_allowed = mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed_rnd,
971 					    (u64)HZ);
972 
973 	if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
974 		if (wait)
975 			*wait = 0;
976 		return true;
977 	}
978 
979 	/* Calc approx time to dispatch */
980 	extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
981 	jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
982 
983 	if (!jiffy_wait)
984 		jiffy_wait = 1;
985 
986 	/*
987 	 * This wait time is without taking into consideration the rounding
988 	 * up we did. Add that time also.
989 	 */
990 	jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
991 	if (wait)
992 		*wait = jiffy_wait;
993 	return false;
994 }
995 
996 /*
997  * Returns whether one can dispatch a bio or not. Also returns approx number
998  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
999  */
tg_may_dispatch(struct throtl_grp * tg,struct bio * bio,unsigned long * wait)1000 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
1001 			    unsigned long *wait)
1002 {
1003 	bool rw = bio_data_dir(bio);
1004 	unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
1005 	u64 bps_limit = tg_bps_limit(tg, rw);
1006 	u32 iops_limit = tg_iops_limit(tg, rw);
1007 
1008 	/*
1009  	 * Currently whole state machine of group depends on first bio
1010 	 * queued in the group bio list. So one should not be calling
1011 	 * this function with a different bio if there are other bios
1012 	 * queued.
1013 	 */
1014 	BUG_ON(tg->service_queue.nr_queued[rw] &&
1015 	       bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
1016 
1017 	/* If tg->bps = -1, then BW is unlimited */
1018 	if (bps_limit == U64_MAX && iops_limit == UINT_MAX) {
1019 		if (wait)
1020 			*wait = 0;
1021 		return true;
1022 	}
1023 
1024 	/*
1025 	 * If previous slice expired, start a new one otherwise renew/extend
1026 	 * existing slice to make sure it is at least throtl_slice interval
1027 	 * long since now. New slice is started only for empty throttle group.
1028 	 * If there is queued bio, that means there should be an active
1029 	 * slice and it should be extended instead.
1030 	 */
1031 	if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
1032 		throtl_start_new_slice(tg, rw);
1033 	else {
1034 		if (time_before(tg->slice_end[rw],
1035 		    jiffies + tg->td->throtl_slice))
1036 			throtl_extend_slice(tg, rw,
1037 				jiffies + tg->td->throtl_slice);
1038 	}
1039 
1040 	if (iops_limit != UINT_MAX)
1041 		tg->io_disp[rw] += atomic_xchg(&tg->io_split_cnt[rw], 0);
1042 
1043 	if (tg_with_in_bps_limit(tg, bio, bps_limit, &bps_wait) &&
1044 	    tg_with_in_iops_limit(tg, bio, iops_limit, &iops_wait)) {
1045 		if (wait)
1046 			*wait = 0;
1047 		return true;
1048 	}
1049 
1050 	max_wait = max(bps_wait, iops_wait);
1051 
1052 	if (wait)
1053 		*wait = max_wait;
1054 
1055 	if (time_before(tg->slice_end[rw], jiffies + max_wait))
1056 		throtl_extend_slice(tg, rw, jiffies + max_wait);
1057 
1058 	return false;
1059 }
1060 
throtl_charge_bio(struct throtl_grp * tg,struct bio * bio)1061 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1062 {
1063 	bool rw = bio_data_dir(bio);
1064 	unsigned int bio_size = throtl_bio_data_size(bio);
1065 
1066 	/* Charge the bio to the group */
1067 	tg->bytes_disp[rw] += bio_size;
1068 	tg->io_disp[rw]++;
1069 	tg->last_bytes_disp[rw] += bio_size;
1070 	tg->last_io_disp[rw]++;
1071 
1072 	/*
1073 	 * BIO_THROTTLED is used to prevent the same bio to be throttled
1074 	 * more than once as a throttled bio will go through blk-throtl the
1075 	 * second time when it eventually gets issued.  Set it when a bio
1076 	 * is being charged to a tg.
1077 	 */
1078 	if (!bio_flagged(bio, BIO_THROTTLED))
1079 		bio_set_flag(bio, BIO_THROTTLED);
1080 }
1081 
1082 /**
1083  * throtl_add_bio_tg - add a bio to the specified throtl_grp
1084  * @bio: bio to add
1085  * @qn: qnode to use
1086  * @tg: the target throtl_grp
1087  *
1088  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
1089  * tg->qnode_on_self[] is used.
1090  */
throtl_add_bio_tg(struct bio * bio,struct throtl_qnode * qn,struct throtl_grp * tg)1091 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1092 			      struct throtl_grp *tg)
1093 {
1094 	struct throtl_service_queue *sq = &tg->service_queue;
1095 	bool rw = bio_data_dir(bio);
1096 
1097 	if (!qn)
1098 		qn = &tg->qnode_on_self[rw];
1099 
1100 	/*
1101 	 * If @tg doesn't currently have any bios queued in the same
1102 	 * direction, queueing @bio can change when @tg should be
1103 	 * dispatched.  Mark that @tg was empty.  This is automatically
1104 	 * cleared on the next tg_update_disptime().
1105 	 */
1106 	if (!sq->nr_queued[rw])
1107 		tg->flags |= THROTL_TG_WAS_EMPTY;
1108 
1109 	throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1110 
1111 	sq->nr_queued[rw]++;
1112 	throtl_enqueue_tg(tg);
1113 }
1114 
tg_update_disptime(struct throtl_grp * tg)1115 static void tg_update_disptime(struct throtl_grp *tg)
1116 {
1117 	struct throtl_service_queue *sq = &tg->service_queue;
1118 	unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1119 	struct bio *bio;
1120 
1121 	bio = throtl_peek_queued(&sq->queued[READ]);
1122 	if (bio)
1123 		tg_may_dispatch(tg, bio, &read_wait);
1124 
1125 	bio = throtl_peek_queued(&sq->queued[WRITE]);
1126 	if (bio)
1127 		tg_may_dispatch(tg, bio, &write_wait);
1128 
1129 	min_wait = min(read_wait, write_wait);
1130 	disptime = jiffies + min_wait;
1131 
1132 	/* Update dispatch time */
1133 	throtl_dequeue_tg(tg);
1134 	tg->disptime = disptime;
1135 	throtl_enqueue_tg(tg);
1136 
1137 	/* see throtl_add_bio_tg() */
1138 	tg->flags &= ~THROTL_TG_WAS_EMPTY;
1139 }
1140 
start_parent_slice_with_credit(struct throtl_grp * child_tg,struct throtl_grp * parent_tg,bool rw)1141 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1142 					struct throtl_grp *parent_tg, bool rw)
1143 {
1144 	if (throtl_slice_used(parent_tg, rw)) {
1145 		throtl_start_new_slice_with_credit(parent_tg, rw,
1146 				child_tg->slice_start[rw]);
1147 	}
1148 
1149 }
1150 
tg_dispatch_one_bio(struct throtl_grp * tg,bool rw)1151 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1152 {
1153 	struct throtl_service_queue *sq = &tg->service_queue;
1154 	struct throtl_service_queue *parent_sq = sq->parent_sq;
1155 	struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1156 	struct throtl_grp *tg_to_put = NULL;
1157 	struct bio *bio;
1158 
1159 	/*
1160 	 * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1161 	 * from @tg may put its reference and @parent_sq might end up
1162 	 * getting released prematurely.  Remember the tg to put and put it
1163 	 * after @bio is transferred to @parent_sq.
1164 	 */
1165 	bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1166 	sq->nr_queued[rw]--;
1167 
1168 	throtl_charge_bio(tg, bio);
1169 
1170 	/*
1171 	 * If our parent is another tg, we just need to transfer @bio to
1172 	 * the parent using throtl_add_bio_tg().  If our parent is
1173 	 * @td->service_queue, @bio is ready to be issued.  Put it on its
1174 	 * bio_lists[] and decrease total number queued.  The caller is
1175 	 * responsible for issuing these bios.
1176 	 */
1177 	if (parent_tg) {
1178 		throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1179 		start_parent_slice_with_credit(tg, parent_tg, rw);
1180 	} else {
1181 		throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1182 				     &parent_sq->queued[rw]);
1183 		BUG_ON(tg->td->nr_queued[rw] <= 0);
1184 		tg->td->nr_queued[rw]--;
1185 	}
1186 
1187 	throtl_trim_slice(tg, rw);
1188 
1189 	if (tg_to_put)
1190 		blkg_put(tg_to_blkg(tg_to_put));
1191 }
1192 
throtl_dispatch_tg(struct throtl_grp * tg)1193 static int throtl_dispatch_tg(struct throtl_grp *tg)
1194 {
1195 	struct throtl_service_queue *sq = &tg->service_queue;
1196 	unsigned int nr_reads = 0, nr_writes = 0;
1197 	unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1198 	unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1199 	struct bio *bio;
1200 
1201 	/* Try to dispatch 75% READS and 25% WRITES */
1202 
1203 	while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1204 	       tg_may_dispatch(tg, bio, NULL)) {
1205 
1206 		tg_dispatch_one_bio(tg, bio_data_dir(bio));
1207 		nr_reads++;
1208 
1209 		if (nr_reads >= max_nr_reads)
1210 			break;
1211 	}
1212 
1213 	while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1214 	       tg_may_dispatch(tg, bio, NULL)) {
1215 
1216 		tg_dispatch_one_bio(tg, bio_data_dir(bio));
1217 		nr_writes++;
1218 
1219 		if (nr_writes >= max_nr_writes)
1220 			break;
1221 	}
1222 
1223 	return nr_reads + nr_writes;
1224 }
1225 
throtl_select_dispatch(struct throtl_service_queue * parent_sq)1226 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1227 {
1228 	unsigned int nr_disp = 0;
1229 
1230 	while (1) {
1231 		struct throtl_grp *tg;
1232 		struct throtl_service_queue *sq;
1233 
1234 		if (!parent_sq->nr_pending)
1235 			break;
1236 
1237 		tg = throtl_rb_first(parent_sq);
1238 		if (!tg)
1239 			break;
1240 
1241 		if (time_before(jiffies, tg->disptime))
1242 			break;
1243 
1244 		throtl_dequeue_tg(tg);
1245 
1246 		nr_disp += throtl_dispatch_tg(tg);
1247 
1248 		sq = &tg->service_queue;
1249 		if (sq->nr_queued[0] || sq->nr_queued[1])
1250 			tg_update_disptime(tg);
1251 
1252 		if (nr_disp >= THROTL_QUANTUM)
1253 			break;
1254 	}
1255 
1256 	return nr_disp;
1257 }
1258 
1259 static bool throtl_can_upgrade(struct throtl_data *td,
1260 	struct throtl_grp *this_tg);
1261 /**
1262  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1263  * @t: the pending_timer member of the throtl_service_queue being serviced
1264  *
1265  * This timer is armed when a child throtl_grp with active bio's become
1266  * pending and queued on the service_queue's pending_tree and expires when
1267  * the first child throtl_grp should be dispatched.  This function
1268  * dispatches bio's from the children throtl_grps to the parent
1269  * service_queue.
1270  *
1271  * If the parent's parent is another throtl_grp, dispatching is propagated
1272  * by either arming its pending_timer or repeating dispatch directly.  If
1273  * the top-level service_tree is reached, throtl_data->dispatch_work is
1274  * kicked so that the ready bio's are issued.
1275  */
throtl_pending_timer_fn(struct timer_list * t)1276 static void throtl_pending_timer_fn(struct timer_list *t)
1277 {
1278 	struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1279 	struct throtl_grp *tg = sq_to_tg(sq);
1280 	struct throtl_data *td = sq_to_td(sq);
1281 	struct request_queue *q = td->queue;
1282 	struct throtl_service_queue *parent_sq;
1283 	bool dispatched;
1284 	int ret;
1285 
1286 	spin_lock_irq(&q->queue_lock);
1287 	if (throtl_can_upgrade(td, NULL))
1288 		throtl_upgrade_state(td);
1289 
1290 again:
1291 	parent_sq = sq->parent_sq;
1292 	dispatched = false;
1293 
1294 	while (true) {
1295 		throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1296 			   sq->nr_queued[READ] + sq->nr_queued[WRITE],
1297 			   sq->nr_queued[READ], sq->nr_queued[WRITE]);
1298 
1299 		ret = throtl_select_dispatch(sq);
1300 		if (ret) {
1301 			throtl_log(sq, "bios disp=%u", ret);
1302 			dispatched = true;
1303 		}
1304 
1305 		if (throtl_schedule_next_dispatch(sq, false))
1306 			break;
1307 
1308 		/* this dispatch windows is still open, relax and repeat */
1309 		spin_unlock_irq(&q->queue_lock);
1310 		cpu_relax();
1311 		spin_lock_irq(&q->queue_lock);
1312 	}
1313 
1314 	if (!dispatched)
1315 		goto out_unlock;
1316 
1317 	if (parent_sq) {
1318 		/* @parent_sq is another throl_grp, propagate dispatch */
1319 		if (tg->flags & THROTL_TG_WAS_EMPTY) {
1320 			tg_update_disptime(tg);
1321 			if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1322 				/* window is already open, repeat dispatching */
1323 				sq = parent_sq;
1324 				tg = sq_to_tg(sq);
1325 				goto again;
1326 			}
1327 		}
1328 	} else {
1329 		/* reached the top-level, queue issuing */
1330 		queue_work(kthrotld_workqueue, &td->dispatch_work);
1331 	}
1332 out_unlock:
1333 	spin_unlock_irq(&q->queue_lock);
1334 }
1335 
1336 /**
1337  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1338  * @work: work item being executed
1339  *
1340  * This function is queued for execution when bios reach the bio_lists[]
1341  * of throtl_data->service_queue.  Those bios are ready and issued by this
1342  * function.
1343  */
blk_throtl_dispatch_work_fn(struct work_struct * work)1344 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1345 {
1346 	struct throtl_data *td = container_of(work, struct throtl_data,
1347 					      dispatch_work);
1348 	struct throtl_service_queue *td_sq = &td->service_queue;
1349 	struct request_queue *q = td->queue;
1350 	struct bio_list bio_list_on_stack;
1351 	struct bio *bio;
1352 	struct blk_plug plug;
1353 	int rw;
1354 
1355 	bio_list_init(&bio_list_on_stack);
1356 
1357 	spin_lock_irq(&q->queue_lock);
1358 	for (rw = READ; rw <= WRITE; rw++)
1359 		while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1360 			bio_list_add(&bio_list_on_stack, bio);
1361 	spin_unlock_irq(&q->queue_lock);
1362 
1363 	if (!bio_list_empty(&bio_list_on_stack)) {
1364 		blk_start_plug(&plug);
1365 		while ((bio = bio_list_pop(&bio_list_on_stack)))
1366 			submit_bio_noacct(bio);
1367 		blk_finish_plug(&plug);
1368 	}
1369 }
1370 
tg_prfill_conf_u64(struct seq_file * sf,struct blkg_policy_data * pd,int off)1371 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1372 			      int off)
1373 {
1374 	struct throtl_grp *tg = pd_to_tg(pd);
1375 	u64 v = *(u64 *)((void *)tg + off);
1376 
1377 	if (v == U64_MAX)
1378 		return 0;
1379 	return __blkg_prfill_u64(sf, pd, v);
1380 }
1381 
tg_prfill_conf_uint(struct seq_file * sf,struct blkg_policy_data * pd,int off)1382 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1383 			       int off)
1384 {
1385 	struct throtl_grp *tg = pd_to_tg(pd);
1386 	unsigned int v = *(unsigned int *)((void *)tg + off);
1387 
1388 	if (v == UINT_MAX)
1389 		return 0;
1390 	return __blkg_prfill_u64(sf, pd, v);
1391 }
1392 
tg_print_conf_u64(struct seq_file * sf,void * v)1393 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1394 {
1395 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1396 			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1397 	return 0;
1398 }
1399 
tg_print_conf_uint(struct seq_file * sf,void * v)1400 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1401 {
1402 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1403 			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1404 	return 0;
1405 }
1406 
tg_conf_updated(struct throtl_grp * tg,bool global)1407 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1408 {
1409 	struct throtl_service_queue *sq = &tg->service_queue;
1410 	struct cgroup_subsys_state *pos_css;
1411 	struct blkcg_gq *blkg;
1412 
1413 	throtl_log(&tg->service_queue,
1414 		   "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1415 		   tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1416 		   tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1417 
1418 	rcu_read_lock();
1419 	/*
1420 	 * Update has_rules[] flags for the updated tg's subtree.  A tg is
1421 	 * considered to have rules if either the tg itself or any of its
1422 	 * ancestors has rules.  This identifies groups without any
1423 	 * restrictions in the whole hierarchy and allows them to bypass
1424 	 * blk-throttle.
1425 	 */
1426 	blkg_for_each_descendant_pre(blkg, pos_css,
1427 			global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1428 		struct throtl_grp *this_tg = blkg_to_tg(blkg);
1429 		struct throtl_grp *parent_tg;
1430 
1431 		tg_update_has_rules(this_tg);
1432 		/* ignore root/second level */
1433 		if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1434 		    !blkg->parent->parent)
1435 			continue;
1436 		parent_tg = blkg_to_tg(blkg->parent);
1437 		/*
1438 		 * make sure all children has lower idle time threshold and
1439 		 * higher latency target
1440 		 */
1441 		this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1442 				parent_tg->idletime_threshold);
1443 		this_tg->latency_target = max(this_tg->latency_target,
1444 				parent_tg->latency_target);
1445 	}
1446 	rcu_read_unlock();
1447 
1448 	/*
1449 	 * We're already holding queue_lock and know @tg is valid.  Let's
1450 	 * apply the new config directly.
1451 	 *
1452 	 * Restart the slices for both READ and WRITES. It might happen
1453 	 * that a group's limit are dropped suddenly and we don't want to
1454 	 * account recently dispatched IO with new low rate.
1455 	 */
1456 	throtl_start_new_slice(tg, READ);
1457 	throtl_start_new_slice(tg, WRITE);
1458 
1459 	if (tg->flags & THROTL_TG_PENDING) {
1460 		tg_update_disptime(tg);
1461 		throtl_schedule_next_dispatch(sq->parent_sq, true);
1462 	}
1463 }
1464 
tg_set_conf(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off,bool is_u64)1465 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1466 			   char *buf, size_t nbytes, loff_t off, bool is_u64)
1467 {
1468 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1469 	struct blkg_conf_ctx ctx;
1470 	struct throtl_grp *tg;
1471 	int ret;
1472 	u64 v;
1473 
1474 	ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1475 	if (ret)
1476 		return ret;
1477 
1478 	ret = -EINVAL;
1479 	if (sscanf(ctx.body, "%llu", &v) != 1)
1480 		goto out_finish;
1481 	if (!v)
1482 		v = U64_MAX;
1483 
1484 	tg = blkg_to_tg(ctx.blkg);
1485 
1486 	if (is_u64)
1487 		*(u64 *)((void *)tg + of_cft(of)->private) = v;
1488 	else
1489 		*(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1490 
1491 	tg_conf_updated(tg, false);
1492 	ret = 0;
1493 out_finish:
1494 	blkg_conf_finish(&ctx);
1495 	return ret ?: nbytes;
1496 }
1497 
tg_set_conf_u64(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)1498 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1499 			       char *buf, size_t nbytes, loff_t off)
1500 {
1501 	return tg_set_conf(of, buf, nbytes, off, true);
1502 }
1503 
tg_set_conf_uint(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)1504 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1505 				char *buf, size_t nbytes, loff_t off)
1506 {
1507 	return tg_set_conf(of, buf, nbytes, off, false);
1508 }
1509 
tg_print_rwstat(struct seq_file * sf,void * v)1510 static int tg_print_rwstat(struct seq_file *sf, void *v)
1511 {
1512 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1513 			  blkg_prfill_rwstat, &blkcg_policy_throtl,
1514 			  seq_cft(sf)->private, true);
1515 	return 0;
1516 }
1517 
tg_prfill_rwstat_recursive(struct seq_file * sf,struct blkg_policy_data * pd,int off)1518 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1519 				      struct blkg_policy_data *pd, int off)
1520 {
1521 	struct blkg_rwstat_sample sum;
1522 
1523 	blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1524 				  &sum);
1525 	return __blkg_prfill_rwstat(sf, pd, &sum);
1526 }
1527 
tg_print_rwstat_recursive(struct seq_file * sf,void * v)1528 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1529 {
1530 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1531 			  tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1532 			  seq_cft(sf)->private, true);
1533 	return 0;
1534 }
1535 
1536 static struct cftype throtl_legacy_files[] = {
1537 	{
1538 		.name = "throttle.read_bps_device",
1539 		.private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1540 		.seq_show = tg_print_conf_u64,
1541 		.write = tg_set_conf_u64,
1542 	},
1543 	{
1544 		.name = "throttle.write_bps_device",
1545 		.private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1546 		.seq_show = tg_print_conf_u64,
1547 		.write = tg_set_conf_u64,
1548 	},
1549 	{
1550 		.name = "throttle.read_iops_device",
1551 		.private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1552 		.seq_show = tg_print_conf_uint,
1553 		.write = tg_set_conf_uint,
1554 	},
1555 	{
1556 		.name = "throttle.write_iops_device",
1557 		.private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1558 		.seq_show = tg_print_conf_uint,
1559 		.write = tg_set_conf_uint,
1560 	},
1561 	{
1562 		.name = "throttle.io_service_bytes",
1563 		.private = offsetof(struct throtl_grp, stat_bytes),
1564 		.seq_show = tg_print_rwstat,
1565 	},
1566 	{
1567 		.name = "throttle.io_service_bytes_recursive",
1568 		.private = offsetof(struct throtl_grp, stat_bytes),
1569 		.seq_show = tg_print_rwstat_recursive,
1570 	},
1571 	{
1572 		.name = "throttle.io_serviced",
1573 		.private = offsetof(struct throtl_grp, stat_ios),
1574 		.seq_show = tg_print_rwstat,
1575 	},
1576 	{
1577 		.name = "throttle.io_serviced_recursive",
1578 		.private = offsetof(struct throtl_grp, stat_ios),
1579 		.seq_show = tg_print_rwstat_recursive,
1580 	},
1581 	{ }	/* terminate */
1582 };
1583 
tg_prfill_limit(struct seq_file * sf,struct blkg_policy_data * pd,int off)1584 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1585 			 int off)
1586 {
1587 	struct throtl_grp *tg = pd_to_tg(pd);
1588 	const char *dname = blkg_dev_name(pd->blkg);
1589 	char bufs[4][21] = { "max", "max", "max", "max" };
1590 	u64 bps_dft;
1591 	unsigned int iops_dft;
1592 	char idle_time[26] = "";
1593 	char latency_time[26] = "";
1594 
1595 	if (!dname)
1596 		return 0;
1597 
1598 	if (off == LIMIT_LOW) {
1599 		bps_dft = 0;
1600 		iops_dft = 0;
1601 	} else {
1602 		bps_dft = U64_MAX;
1603 		iops_dft = UINT_MAX;
1604 	}
1605 
1606 	if (tg->bps_conf[READ][off] == bps_dft &&
1607 	    tg->bps_conf[WRITE][off] == bps_dft &&
1608 	    tg->iops_conf[READ][off] == iops_dft &&
1609 	    tg->iops_conf[WRITE][off] == iops_dft &&
1610 	    (off != LIMIT_LOW ||
1611 	     (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1612 	      tg->latency_target_conf == DFL_LATENCY_TARGET)))
1613 		return 0;
1614 
1615 	if (tg->bps_conf[READ][off] != U64_MAX)
1616 		snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1617 			tg->bps_conf[READ][off]);
1618 	if (tg->bps_conf[WRITE][off] != U64_MAX)
1619 		snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1620 			tg->bps_conf[WRITE][off]);
1621 	if (tg->iops_conf[READ][off] != UINT_MAX)
1622 		snprintf(bufs[2], sizeof(bufs[2]), "%u",
1623 			tg->iops_conf[READ][off]);
1624 	if (tg->iops_conf[WRITE][off] != UINT_MAX)
1625 		snprintf(bufs[3], sizeof(bufs[3]), "%u",
1626 			tg->iops_conf[WRITE][off]);
1627 	if (off == LIMIT_LOW) {
1628 		if (tg->idletime_threshold_conf == ULONG_MAX)
1629 			strcpy(idle_time, " idle=max");
1630 		else
1631 			snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1632 				tg->idletime_threshold_conf);
1633 
1634 		if (tg->latency_target_conf == ULONG_MAX)
1635 			strcpy(latency_time, " latency=max");
1636 		else
1637 			snprintf(latency_time, sizeof(latency_time),
1638 				" latency=%lu", tg->latency_target_conf);
1639 	}
1640 
1641 	seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1642 		   dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1643 		   latency_time);
1644 	return 0;
1645 }
1646 
tg_print_limit(struct seq_file * sf,void * v)1647 static int tg_print_limit(struct seq_file *sf, void *v)
1648 {
1649 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1650 			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1651 	return 0;
1652 }
1653 
tg_set_limit(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)1654 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1655 			  char *buf, size_t nbytes, loff_t off)
1656 {
1657 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1658 	struct blkg_conf_ctx ctx;
1659 	struct throtl_grp *tg;
1660 	u64 v[4];
1661 	unsigned long idle_time;
1662 	unsigned long latency_time;
1663 	int ret;
1664 	int index = of_cft(of)->private;
1665 
1666 	ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1667 	if (ret)
1668 		return ret;
1669 
1670 	tg = blkg_to_tg(ctx.blkg);
1671 
1672 	v[0] = tg->bps_conf[READ][index];
1673 	v[1] = tg->bps_conf[WRITE][index];
1674 	v[2] = tg->iops_conf[READ][index];
1675 	v[3] = tg->iops_conf[WRITE][index];
1676 
1677 	idle_time = tg->idletime_threshold_conf;
1678 	latency_time = tg->latency_target_conf;
1679 	while (true) {
1680 		char tok[27];	/* wiops=18446744073709551616 */
1681 		char *p;
1682 		u64 val = U64_MAX;
1683 		int len;
1684 
1685 		if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1686 			break;
1687 		if (tok[0] == '\0')
1688 			break;
1689 		ctx.body += len;
1690 
1691 		ret = -EINVAL;
1692 		p = tok;
1693 		strsep(&p, "=");
1694 		if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1695 			goto out_finish;
1696 
1697 		ret = -ERANGE;
1698 		if (!val)
1699 			goto out_finish;
1700 
1701 		ret = -EINVAL;
1702 		if (!strcmp(tok, "rbps") && val > 1)
1703 			v[0] = val;
1704 		else if (!strcmp(tok, "wbps") && val > 1)
1705 			v[1] = val;
1706 		else if (!strcmp(tok, "riops") && val > 1)
1707 			v[2] = min_t(u64, val, UINT_MAX);
1708 		else if (!strcmp(tok, "wiops") && val > 1)
1709 			v[3] = min_t(u64, val, UINT_MAX);
1710 		else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1711 			idle_time = val;
1712 		else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1713 			latency_time = val;
1714 		else
1715 			goto out_finish;
1716 	}
1717 
1718 	tg->bps_conf[READ][index] = v[0];
1719 	tg->bps_conf[WRITE][index] = v[1];
1720 	tg->iops_conf[READ][index] = v[2];
1721 	tg->iops_conf[WRITE][index] = v[3];
1722 
1723 	if (index == LIMIT_MAX) {
1724 		tg->bps[READ][index] = v[0];
1725 		tg->bps[WRITE][index] = v[1];
1726 		tg->iops[READ][index] = v[2];
1727 		tg->iops[WRITE][index] = v[3];
1728 	}
1729 	tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1730 		tg->bps_conf[READ][LIMIT_MAX]);
1731 	tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1732 		tg->bps_conf[WRITE][LIMIT_MAX]);
1733 	tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1734 		tg->iops_conf[READ][LIMIT_MAX]);
1735 	tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1736 		tg->iops_conf[WRITE][LIMIT_MAX]);
1737 	tg->idletime_threshold_conf = idle_time;
1738 	tg->latency_target_conf = latency_time;
1739 
1740 	/* force user to configure all settings for low limit  */
1741 	if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1742 	      tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1743 	    tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1744 	    tg->latency_target_conf == DFL_LATENCY_TARGET) {
1745 		tg->bps[READ][LIMIT_LOW] = 0;
1746 		tg->bps[WRITE][LIMIT_LOW] = 0;
1747 		tg->iops[READ][LIMIT_LOW] = 0;
1748 		tg->iops[WRITE][LIMIT_LOW] = 0;
1749 		tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1750 		tg->latency_target = DFL_LATENCY_TARGET;
1751 	} else if (index == LIMIT_LOW) {
1752 		tg->idletime_threshold = tg->idletime_threshold_conf;
1753 		tg->latency_target = tg->latency_target_conf;
1754 	}
1755 
1756 	blk_throtl_update_limit_valid(tg->td);
1757 	if (tg->td->limit_valid[LIMIT_LOW]) {
1758 		if (index == LIMIT_LOW)
1759 			tg->td->limit_index = LIMIT_LOW;
1760 	} else
1761 		tg->td->limit_index = LIMIT_MAX;
1762 	tg_conf_updated(tg, index == LIMIT_LOW &&
1763 		tg->td->limit_valid[LIMIT_LOW]);
1764 	ret = 0;
1765 out_finish:
1766 	blkg_conf_finish(&ctx);
1767 	return ret ?: nbytes;
1768 }
1769 
1770 static struct cftype throtl_files[] = {
1771 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1772 	{
1773 		.name = "low",
1774 		.flags = CFTYPE_NOT_ON_ROOT,
1775 		.seq_show = tg_print_limit,
1776 		.write = tg_set_limit,
1777 		.private = LIMIT_LOW,
1778 	},
1779 #endif
1780 	{
1781 		.name = "max",
1782 		.flags = CFTYPE_NOT_ON_ROOT,
1783 		.seq_show = tg_print_limit,
1784 		.write = tg_set_limit,
1785 		.private = LIMIT_MAX,
1786 	},
1787 	{ }	/* terminate */
1788 };
1789 
throtl_shutdown_wq(struct request_queue * q)1790 static void throtl_shutdown_wq(struct request_queue *q)
1791 {
1792 	struct throtl_data *td = q->td;
1793 
1794 	cancel_work_sync(&td->dispatch_work);
1795 }
1796 
1797 static struct blkcg_policy blkcg_policy_throtl = {
1798 	.dfl_cftypes		= throtl_files,
1799 	.legacy_cftypes		= throtl_legacy_files,
1800 
1801 	.pd_alloc_fn		= throtl_pd_alloc,
1802 	.pd_init_fn		= throtl_pd_init,
1803 	.pd_online_fn		= throtl_pd_online,
1804 	.pd_offline_fn		= throtl_pd_offline,
1805 	.pd_free_fn		= throtl_pd_free,
1806 };
1807 
__tg_last_low_overflow_time(struct throtl_grp * tg)1808 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1809 {
1810 	unsigned long rtime = jiffies, wtime = jiffies;
1811 
1812 	if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1813 		rtime = tg->last_low_overflow_time[READ];
1814 	if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1815 		wtime = tg->last_low_overflow_time[WRITE];
1816 	return min(rtime, wtime);
1817 }
1818 
1819 /* tg should not be an intermediate node */
tg_last_low_overflow_time(struct throtl_grp * tg)1820 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1821 {
1822 	struct throtl_service_queue *parent_sq;
1823 	struct throtl_grp *parent = tg;
1824 	unsigned long ret = __tg_last_low_overflow_time(tg);
1825 
1826 	while (true) {
1827 		parent_sq = parent->service_queue.parent_sq;
1828 		parent = sq_to_tg(parent_sq);
1829 		if (!parent)
1830 			break;
1831 
1832 		/*
1833 		 * The parent doesn't have low limit, it always reaches low
1834 		 * limit. Its overflow time is useless for children
1835 		 */
1836 		if (!parent->bps[READ][LIMIT_LOW] &&
1837 		    !parent->iops[READ][LIMIT_LOW] &&
1838 		    !parent->bps[WRITE][LIMIT_LOW] &&
1839 		    !parent->iops[WRITE][LIMIT_LOW])
1840 			continue;
1841 		if (time_after(__tg_last_low_overflow_time(parent), ret))
1842 			ret = __tg_last_low_overflow_time(parent);
1843 	}
1844 	return ret;
1845 }
1846 
throtl_tg_is_idle(struct throtl_grp * tg)1847 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1848 {
1849 	/*
1850 	 * cgroup is idle if:
1851 	 * - single idle is too long, longer than a fixed value (in case user
1852 	 *   configure a too big threshold) or 4 times of idletime threshold
1853 	 * - average think time is more than threshold
1854 	 * - IO latency is largely below threshold
1855 	 */
1856 	unsigned long time;
1857 	bool ret;
1858 
1859 	time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1860 	ret = tg->latency_target == DFL_LATENCY_TARGET ||
1861 	      tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1862 	      (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1863 	      tg->avg_idletime > tg->idletime_threshold ||
1864 	      (tg->latency_target && tg->bio_cnt &&
1865 		tg->bad_bio_cnt * 5 < tg->bio_cnt);
1866 	throtl_log(&tg->service_queue,
1867 		"avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1868 		tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1869 		tg->bio_cnt, ret, tg->td->scale);
1870 	return ret;
1871 }
1872 
throtl_tg_can_upgrade(struct throtl_grp * tg)1873 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1874 {
1875 	struct throtl_service_queue *sq = &tg->service_queue;
1876 	bool read_limit, write_limit;
1877 
1878 	/*
1879 	 * if cgroup reaches low limit (if low limit is 0, the cgroup always
1880 	 * reaches), it's ok to upgrade to next limit
1881 	 */
1882 	read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1883 	write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1884 	if (!read_limit && !write_limit)
1885 		return true;
1886 	if (read_limit && sq->nr_queued[READ] &&
1887 	    (!write_limit || sq->nr_queued[WRITE]))
1888 		return true;
1889 	if (write_limit && sq->nr_queued[WRITE] &&
1890 	    (!read_limit || sq->nr_queued[READ]))
1891 		return true;
1892 
1893 	if (time_after_eq(jiffies,
1894 		tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1895 	    throtl_tg_is_idle(tg))
1896 		return true;
1897 	return false;
1898 }
1899 
throtl_hierarchy_can_upgrade(struct throtl_grp * tg)1900 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1901 {
1902 	while (true) {
1903 		if (throtl_tg_can_upgrade(tg))
1904 			return true;
1905 		tg = sq_to_tg(tg->service_queue.parent_sq);
1906 		if (!tg || !tg_to_blkg(tg)->parent)
1907 			return false;
1908 	}
1909 	return false;
1910 }
1911 
throtl_can_upgrade(struct throtl_data * td,struct throtl_grp * this_tg)1912 static bool throtl_can_upgrade(struct throtl_data *td,
1913 	struct throtl_grp *this_tg)
1914 {
1915 	struct cgroup_subsys_state *pos_css;
1916 	struct blkcg_gq *blkg;
1917 
1918 	if (td->limit_index != LIMIT_LOW)
1919 		return false;
1920 
1921 	if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1922 		return false;
1923 
1924 	rcu_read_lock();
1925 	blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1926 		struct throtl_grp *tg = blkg_to_tg(blkg);
1927 
1928 		if (tg == this_tg)
1929 			continue;
1930 		if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1931 			continue;
1932 		if (!throtl_hierarchy_can_upgrade(tg)) {
1933 			rcu_read_unlock();
1934 			return false;
1935 		}
1936 	}
1937 	rcu_read_unlock();
1938 	return true;
1939 }
1940 
throtl_upgrade_check(struct throtl_grp * tg)1941 static void throtl_upgrade_check(struct throtl_grp *tg)
1942 {
1943 	unsigned long now = jiffies;
1944 
1945 	if (tg->td->limit_index != LIMIT_LOW)
1946 		return;
1947 
1948 	if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1949 		return;
1950 
1951 	tg->last_check_time = now;
1952 
1953 	if (!time_after_eq(now,
1954 	     __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1955 		return;
1956 
1957 	if (throtl_can_upgrade(tg->td, NULL))
1958 		throtl_upgrade_state(tg->td);
1959 }
1960 
throtl_upgrade_state(struct throtl_data * td)1961 static void throtl_upgrade_state(struct throtl_data *td)
1962 {
1963 	struct cgroup_subsys_state *pos_css;
1964 	struct blkcg_gq *blkg;
1965 
1966 	throtl_log(&td->service_queue, "upgrade to max");
1967 	td->limit_index = LIMIT_MAX;
1968 	td->low_upgrade_time = jiffies;
1969 	td->scale = 0;
1970 	rcu_read_lock();
1971 	blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1972 		struct throtl_grp *tg = blkg_to_tg(blkg);
1973 		struct throtl_service_queue *sq = &tg->service_queue;
1974 
1975 		tg->disptime = jiffies - 1;
1976 		throtl_select_dispatch(sq);
1977 		throtl_schedule_next_dispatch(sq, true);
1978 	}
1979 	rcu_read_unlock();
1980 	throtl_select_dispatch(&td->service_queue);
1981 	throtl_schedule_next_dispatch(&td->service_queue, true);
1982 	queue_work(kthrotld_workqueue, &td->dispatch_work);
1983 }
1984 
throtl_downgrade_state(struct throtl_data * td)1985 static void throtl_downgrade_state(struct throtl_data *td)
1986 {
1987 	td->scale /= 2;
1988 
1989 	throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1990 	if (td->scale) {
1991 		td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1992 		return;
1993 	}
1994 
1995 	td->limit_index = LIMIT_LOW;
1996 	td->low_downgrade_time = jiffies;
1997 }
1998 
throtl_tg_can_downgrade(struct throtl_grp * tg)1999 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
2000 {
2001 	struct throtl_data *td = tg->td;
2002 	unsigned long now = jiffies;
2003 
2004 	/*
2005 	 * If cgroup is below low limit, consider downgrade and throttle other
2006 	 * cgroups
2007 	 */
2008 	if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
2009 	    time_after_eq(now, tg_last_low_overflow_time(tg) +
2010 					td->throtl_slice) &&
2011 	    (!throtl_tg_is_idle(tg) ||
2012 	     !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
2013 		return true;
2014 	return false;
2015 }
2016 
throtl_hierarchy_can_downgrade(struct throtl_grp * tg)2017 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
2018 {
2019 	while (true) {
2020 		if (!throtl_tg_can_downgrade(tg))
2021 			return false;
2022 		tg = sq_to_tg(tg->service_queue.parent_sq);
2023 		if (!tg || !tg_to_blkg(tg)->parent)
2024 			break;
2025 	}
2026 	return true;
2027 }
2028 
throtl_downgrade_check(struct throtl_grp * tg)2029 static void throtl_downgrade_check(struct throtl_grp *tg)
2030 {
2031 	uint64_t bps;
2032 	unsigned int iops;
2033 	unsigned long elapsed_time;
2034 	unsigned long now = jiffies;
2035 
2036 	if (tg->td->limit_index != LIMIT_MAX ||
2037 	    !tg->td->limit_valid[LIMIT_LOW])
2038 		return;
2039 	if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
2040 		return;
2041 	if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
2042 		return;
2043 
2044 	elapsed_time = now - tg->last_check_time;
2045 	tg->last_check_time = now;
2046 
2047 	if (time_before(now, tg_last_low_overflow_time(tg) +
2048 			tg->td->throtl_slice))
2049 		return;
2050 
2051 	if (tg->bps[READ][LIMIT_LOW]) {
2052 		bps = tg->last_bytes_disp[READ] * HZ;
2053 		do_div(bps, elapsed_time);
2054 		if (bps >= tg->bps[READ][LIMIT_LOW])
2055 			tg->last_low_overflow_time[READ] = now;
2056 	}
2057 
2058 	if (tg->bps[WRITE][LIMIT_LOW]) {
2059 		bps = tg->last_bytes_disp[WRITE] * HZ;
2060 		do_div(bps, elapsed_time);
2061 		if (bps >= tg->bps[WRITE][LIMIT_LOW])
2062 			tg->last_low_overflow_time[WRITE] = now;
2063 	}
2064 
2065 	if (tg->iops[READ][LIMIT_LOW]) {
2066 		tg->last_io_disp[READ] += atomic_xchg(&tg->last_io_split_cnt[READ], 0);
2067 		iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2068 		if (iops >= tg->iops[READ][LIMIT_LOW])
2069 			tg->last_low_overflow_time[READ] = now;
2070 	}
2071 
2072 	if (tg->iops[WRITE][LIMIT_LOW]) {
2073 		tg->last_io_disp[WRITE] += atomic_xchg(&tg->last_io_split_cnt[WRITE], 0);
2074 		iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2075 		if (iops >= tg->iops[WRITE][LIMIT_LOW])
2076 			tg->last_low_overflow_time[WRITE] = now;
2077 	}
2078 
2079 	/*
2080 	 * If cgroup is below low limit, consider downgrade and throttle other
2081 	 * cgroups
2082 	 */
2083 	if (throtl_hierarchy_can_downgrade(tg))
2084 		throtl_downgrade_state(tg->td);
2085 
2086 	tg->last_bytes_disp[READ] = 0;
2087 	tg->last_bytes_disp[WRITE] = 0;
2088 	tg->last_io_disp[READ] = 0;
2089 	tg->last_io_disp[WRITE] = 0;
2090 }
2091 
blk_throtl_update_idletime(struct throtl_grp * tg)2092 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2093 {
2094 	unsigned long now;
2095 	unsigned long last_finish_time = tg->last_finish_time;
2096 
2097 	if (last_finish_time == 0)
2098 		return;
2099 
2100 	now = ktime_get_ns() >> 10;
2101 	if (now <= last_finish_time ||
2102 	    last_finish_time == tg->checked_last_finish_time)
2103 		return;
2104 
2105 	tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2106 	tg->checked_last_finish_time = last_finish_time;
2107 }
2108 
2109 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
throtl_update_latency_buckets(struct throtl_data * td)2110 static void throtl_update_latency_buckets(struct throtl_data *td)
2111 {
2112 	struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2113 	int i, cpu, rw;
2114 	unsigned long last_latency[2] = { 0 };
2115 	unsigned long latency[2];
2116 
2117 	if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
2118 		return;
2119 	if (time_before(jiffies, td->last_calculate_time + HZ))
2120 		return;
2121 	td->last_calculate_time = jiffies;
2122 
2123 	memset(avg_latency, 0, sizeof(avg_latency));
2124 	for (rw = READ; rw <= WRITE; rw++) {
2125 		for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2126 			struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2127 
2128 			for_each_possible_cpu(cpu) {
2129 				struct latency_bucket *bucket;
2130 
2131 				/* this isn't race free, but ok in practice */
2132 				bucket = per_cpu_ptr(td->latency_buckets[rw],
2133 					cpu);
2134 				tmp->total_latency += bucket[i].total_latency;
2135 				tmp->samples += bucket[i].samples;
2136 				bucket[i].total_latency = 0;
2137 				bucket[i].samples = 0;
2138 			}
2139 
2140 			if (tmp->samples >= 32) {
2141 				int samples = tmp->samples;
2142 
2143 				latency[rw] = tmp->total_latency;
2144 
2145 				tmp->total_latency = 0;
2146 				tmp->samples = 0;
2147 				latency[rw] /= samples;
2148 				if (latency[rw] == 0)
2149 					continue;
2150 				avg_latency[rw][i].latency = latency[rw];
2151 			}
2152 		}
2153 	}
2154 
2155 	for (rw = READ; rw <= WRITE; rw++) {
2156 		for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2157 			if (!avg_latency[rw][i].latency) {
2158 				if (td->avg_buckets[rw][i].latency < last_latency[rw])
2159 					td->avg_buckets[rw][i].latency =
2160 						last_latency[rw];
2161 				continue;
2162 			}
2163 
2164 			if (!td->avg_buckets[rw][i].valid)
2165 				latency[rw] = avg_latency[rw][i].latency;
2166 			else
2167 				latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2168 					avg_latency[rw][i].latency) >> 3;
2169 
2170 			td->avg_buckets[rw][i].latency = max(latency[rw],
2171 				last_latency[rw]);
2172 			td->avg_buckets[rw][i].valid = true;
2173 			last_latency[rw] = td->avg_buckets[rw][i].latency;
2174 		}
2175 	}
2176 
2177 	for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2178 		throtl_log(&td->service_queue,
2179 			"Latency bucket %d: read latency=%ld, read valid=%d, "
2180 			"write latency=%ld, write valid=%d", i,
2181 			td->avg_buckets[READ][i].latency,
2182 			td->avg_buckets[READ][i].valid,
2183 			td->avg_buckets[WRITE][i].latency,
2184 			td->avg_buckets[WRITE][i].valid);
2185 }
2186 #else
throtl_update_latency_buckets(struct throtl_data * td)2187 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2188 {
2189 }
2190 #endif
2191 
blk_throtl_charge_bio_split(struct bio * bio)2192 void blk_throtl_charge_bio_split(struct bio *bio)
2193 {
2194 	struct blkcg_gq *blkg = bio->bi_blkg;
2195 	struct throtl_grp *parent = blkg_to_tg(blkg);
2196 	struct throtl_service_queue *parent_sq;
2197 	bool rw = bio_data_dir(bio);
2198 
2199 	do {
2200 		if (!parent->has_rules[rw])
2201 			break;
2202 
2203 		atomic_inc(&parent->io_split_cnt[rw]);
2204 		atomic_inc(&parent->last_io_split_cnt[rw]);
2205 
2206 		parent_sq = parent->service_queue.parent_sq;
2207 		parent = sq_to_tg(parent_sq);
2208 	} while (parent);
2209 }
2210 
blk_throtl_bio(struct bio * bio)2211 bool blk_throtl_bio(struct bio *bio)
2212 {
2213 	struct request_queue *q = bio->bi_bdev->bd_disk->queue;
2214 	struct blkcg_gq *blkg = bio->bi_blkg;
2215 	struct throtl_qnode *qn = NULL;
2216 	struct throtl_grp *tg = blkg_to_tg(blkg);
2217 	struct throtl_service_queue *sq;
2218 	bool rw = bio_data_dir(bio);
2219 	bool throttled = false;
2220 	struct throtl_data *td = tg->td;
2221 
2222 	rcu_read_lock();
2223 
2224 	/* see throtl_charge_bio() */
2225 	if (bio_flagged(bio, BIO_THROTTLED))
2226 		goto out;
2227 
2228 	if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
2229 		blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
2230 				bio->bi_iter.bi_size);
2231 		blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
2232 	}
2233 
2234 	if (!tg->has_rules[rw])
2235 		goto out;
2236 
2237 	spin_lock_irq(&q->queue_lock);
2238 
2239 	throtl_update_latency_buckets(td);
2240 
2241 	blk_throtl_update_idletime(tg);
2242 
2243 	sq = &tg->service_queue;
2244 
2245 again:
2246 	while (true) {
2247 		if (tg->last_low_overflow_time[rw] == 0)
2248 			tg->last_low_overflow_time[rw] = jiffies;
2249 		throtl_downgrade_check(tg);
2250 		throtl_upgrade_check(tg);
2251 		/* throtl is FIFO - if bios are already queued, should queue */
2252 		if (sq->nr_queued[rw])
2253 			break;
2254 
2255 		/* if above limits, break to queue */
2256 		if (!tg_may_dispatch(tg, bio, NULL)) {
2257 			tg->last_low_overflow_time[rw] = jiffies;
2258 			if (throtl_can_upgrade(td, tg)) {
2259 				throtl_upgrade_state(td);
2260 				goto again;
2261 			}
2262 			break;
2263 		}
2264 
2265 		/* within limits, let's charge and dispatch directly */
2266 		throtl_charge_bio(tg, bio);
2267 
2268 		/*
2269 		 * We need to trim slice even when bios are not being queued
2270 		 * otherwise it might happen that a bio is not queued for
2271 		 * a long time and slice keeps on extending and trim is not
2272 		 * called for a long time. Now if limits are reduced suddenly
2273 		 * we take into account all the IO dispatched so far at new
2274 		 * low rate and * newly queued IO gets a really long dispatch
2275 		 * time.
2276 		 *
2277 		 * So keep on trimming slice even if bio is not queued.
2278 		 */
2279 		throtl_trim_slice(tg, rw);
2280 
2281 		/*
2282 		 * @bio passed through this layer without being throttled.
2283 		 * Climb up the ladder.  If we're already at the top, it
2284 		 * can be executed directly.
2285 		 */
2286 		qn = &tg->qnode_on_parent[rw];
2287 		sq = sq->parent_sq;
2288 		tg = sq_to_tg(sq);
2289 		if (!tg)
2290 			goto out_unlock;
2291 	}
2292 
2293 	/* out-of-limit, queue to @tg */
2294 	throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2295 		   rw == READ ? 'R' : 'W',
2296 		   tg->bytes_disp[rw], bio->bi_iter.bi_size,
2297 		   tg_bps_limit(tg, rw),
2298 		   tg->io_disp[rw], tg_iops_limit(tg, rw),
2299 		   sq->nr_queued[READ], sq->nr_queued[WRITE]);
2300 
2301 	tg->last_low_overflow_time[rw] = jiffies;
2302 
2303 	td->nr_queued[rw]++;
2304 	throtl_add_bio_tg(bio, qn, tg);
2305 	throttled = true;
2306 
2307 	/*
2308 	 * Update @tg's dispatch time and force schedule dispatch if @tg
2309 	 * was empty before @bio.  The forced scheduling isn't likely to
2310 	 * cause undue delay as @bio is likely to be dispatched directly if
2311 	 * its @tg's disptime is not in the future.
2312 	 */
2313 	if (tg->flags & THROTL_TG_WAS_EMPTY) {
2314 		tg_update_disptime(tg);
2315 		throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2316 	}
2317 
2318 out_unlock:
2319 	spin_unlock_irq(&q->queue_lock);
2320 out:
2321 	bio_set_flag(bio, BIO_THROTTLED);
2322 
2323 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2324 	if (throttled || !td->track_bio_latency)
2325 		bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2326 #endif
2327 	rcu_read_unlock();
2328 	return throttled;
2329 }
2330 
2331 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
throtl_track_latency(struct throtl_data * td,sector_t size,int op,unsigned long time)2332 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2333 	int op, unsigned long time)
2334 {
2335 	struct latency_bucket *latency;
2336 	int index;
2337 
2338 	if (!td || td->limit_index != LIMIT_LOW ||
2339 	    !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2340 	    !blk_queue_nonrot(td->queue))
2341 		return;
2342 
2343 	index = request_bucket_index(size);
2344 
2345 	latency = get_cpu_ptr(td->latency_buckets[op]);
2346 	latency[index].total_latency += time;
2347 	latency[index].samples++;
2348 	put_cpu_ptr(td->latency_buckets[op]);
2349 }
2350 
blk_throtl_stat_add(struct request * rq,u64 time_ns)2351 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2352 {
2353 	struct request_queue *q = rq->q;
2354 	struct throtl_data *td = q->td;
2355 
2356 	throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2357 			     time_ns >> 10);
2358 }
2359 
blk_throtl_bio_endio(struct bio * bio)2360 void blk_throtl_bio_endio(struct bio *bio)
2361 {
2362 	struct blkcg_gq *blkg;
2363 	struct throtl_grp *tg;
2364 	u64 finish_time_ns;
2365 	unsigned long finish_time;
2366 	unsigned long start_time;
2367 	unsigned long lat;
2368 	int rw = bio_data_dir(bio);
2369 
2370 	blkg = bio->bi_blkg;
2371 	if (!blkg)
2372 		return;
2373 	tg = blkg_to_tg(blkg);
2374 	if (!tg->td->limit_valid[LIMIT_LOW])
2375 		return;
2376 
2377 	finish_time_ns = ktime_get_ns();
2378 	tg->last_finish_time = finish_time_ns >> 10;
2379 
2380 	start_time = bio_issue_time(&bio->bi_issue) >> 10;
2381 	finish_time = __bio_issue_time(finish_time_ns) >> 10;
2382 	if (!start_time || finish_time <= start_time)
2383 		return;
2384 
2385 	lat = finish_time - start_time;
2386 	/* this is only for bio based driver */
2387 	if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2388 		throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2389 				     bio_op(bio), lat);
2390 
2391 	if (tg->latency_target && lat >= tg->td->filtered_latency) {
2392 		int bucket;
2393 		unsigned int threshold;
2394 
2395 		bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2396 		threshold = tg->td->avg_buckets[rw][bucket].latency +
2397 			tg->latency_target;
2398 		if (lat > threshold)
2399 			tg->bad_bio_cnt++;
2400 		/*
2401 		 * Not race free, could get wrong count, which means cgroups
2402 		 * will be throttled
2403 		 */
2404 		tg->bio_cnt++;
2405 	}
2406 
2407 	if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2408 		tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2409 		tg->bio_cnt /= 2;
2410 		tg->bad_bio_cnt /= 2;
2411 	}
2412 }
2413 #endif
2414 
blk_throtl_init(struct request_queue * q)2415 int blk_throtl_init(struct request_queue *q)
2416 {
2417 	struct throtl_data *td;
2418 	int ret;
2419 
2420 	td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2421 	if (!td)
2422 		return -ENOMEM;
2423 	td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2424 		LATENCY_BUCKET_SIZE, __alignof__(u64));
2425 	if (!td->latency_buckets[READ]) {
2426 		kfree(td);
2427 		return -ENOMEM;
2428 	}
2429 	td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2430 		LATENCY_BUCKET_SIZE, __alignof__(u64));
2431 	if (!td->latency_buckets[WRITE]) {
2432 		free_percpu(td->latency_buckets[READ]);
2433 		kfree(td);
2434 		return -ENOMEM;
2435 	}
2436 
2437 	INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2438 	throtl_service_queue_init(&td->service_queue);
2439 
2440 	q->td = td;
2441 	td->queue = q;
2442 
2443 	td->limit_valid[LIMIT_MAX] = true;
2444 	td->limit_index = LIMIT_MAX;
2445 	td->low_upgrade_time = jiffies;
2446 	td->low_downgrade_time = jiffies;
2447 
2448 	/* activate policy */
2449 	ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2450 	if (ret) {
2451 		free_percpu(td->latency_buckets[READ]);
2452 		free_percpu(td->latency_buckets[WRITE]);
2453 		kfree(td);
2454 	}
2455 	return ret;
2456 }
2457 
blk_throtl_exit(struct request_queue * q)2458 void blk_throtl_exit(struct request_queue *q)
2459 {
2460 	BUG_ON(!q->td);
2461 	del_timer_sync(&q->td->service_queue.pending_timer);
2462 	throtl_shutdown_wq(q);
2463 	blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2464 	free_percpu(q->td->latency_buckets[READ]);
2465 	free_percpu(q->td->latency_buckets[WRITE]);
2466 	kfree(q->td);
2467 }
2468 
blk_throtl_register_queue(struct request_queue * q)2469 void blk_throtl_register_queue(struct request_queue *q)
2470 {
2471 	struct throtl_data *td;
2472 	int i;
2473 
2474 	td = q->td;
2475 	BUG_ON(!td);
2476 
2477 	if (blk_queue_nonrot(q)) {
2478 		td->throtl_slice = DFL_THROTL_SLICE_SSD;
2479 		td->filtered_latency = LATENCY_FILTERED_SSD;
2480 	} else {
2481 		td->throtl_slice = DFL_THROTL_SLICE_HD;
2482 		td->filtered_latency = LATENCY_FILTERED_HD;
2483 		for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2484 			td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2485 			td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2486 		}
2487 	}
2488 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2489 	/* if no low limit, use previous default */
2490 	td->throtl_slice = DFL_THROTL_SLICE_HD;
2491 #endif
2492 
2493 	td->track_bio_latency = !queue_is_mq(q);
2494 	if (!td->track_bio_latency)
2495 		blk_stat_enable_accounting(q);
2496 }
2497 
2498 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
blk_throtl_sample_time_show(struct request_queue * q,char * page)2499 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2500 {
2501 	if (!q->td)
2502 		return -EINVAL;
2503 	return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2504 }
2505 
blk_throtl_sample_time_store(struct request_queue * q,const char * page,size_t count)2506 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2507 	const char *page, size_t count)
2508 {
2509 	unsigned long v;
2510 	unsigned long t;
2511 
2512 	if (!q->td)
2513 		return -EINVAL;
2514 	if (kstrtoul(page, 10, &v))
2515 		return -EINVAL;
2516 	t = msecs_to_jiffies(v);
2517 	if (t == 0 || t > MAX_THROTL_SLICE)
2518 		return -EINVAL;
2519 	q->td->throtl_slice = t;
2520 	return count;
2521 }
2522 #endif
2523 
throtl_init(void)2524 static int __init throtl_init(void)
2525 {
2526 	kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2527 	if (!kthrotld_workqueue)
2528 		panic("Failed to create kthrotld\n");
2529 
2530 	return blkcg_policy_register(&blkcg_policy_throtl);
2531 }
2532 
2533 module_init(throtl_init);
2534