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