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