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