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/*
2 * Interface for controlling IO bandwidth on a request queue
3 *
4 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
5 */
6
7#include <linux/module.h>
8#include <linux/slab.h>
9#include <linux/blkdev.h>
10#include <linux/bio.h>
11#include <linux/blktrace_api.h>
12#include "blk-cgroup.h"
13
14/* Max dispatch from a group in 1 round */
15static int throtl_grp_quantum = 8;
16
17/* Total max dispatch from all groups in one round */
18static int throtl_quantum = 32;
19
20/* Throttling is performed over 100ms slice and after that slice is renewed */
21static unsigned long throtl_slice = HZ/10; /* 100 ms */
22
23/* A workqueue to queue throttle related work */
24static struct workqueue_struct *kthrotld_workqueue;
25static void throtl_schedule_delayed_work(struct throtl_data *td,
26 unsigned long delay);
27
28struct throtl_rb_root {
29 struct rb_root rb;
30 struct rb_node *left;
31 unsigned int count;
32 unsigned long min_disptime;
33};
34
35#define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \
36 .count = 0, .min_disptime = 0}
37
38#define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
39
40struct throtl_grp {
41 /* List of throtl groups on the request queue*/
42 struct hlist_node tg_node;
43
44 /* active throtl group service_tree member */
45 struct rb_node rb_node;
46
47 /*
48 * Dispatch time in jiffies. This is the estimated time when group
49 * will unthrottle and is ready to dispatch more bio. It is used as
50 * key to sort active groups in service tree.
51 */
52 unsigned long disptime;
53
54 struct blkio_group blkg;
55 atomic_t ref;
56 unsigned int flags;
57
58 /* Two lists for READ and WRITE */
59 struct bio_list bio_lists[2];
60
61 /* Number of queued bios on READ and WRITE lists */
62 unsigned int nr_queued[2];
63
64 /* bytes per second rate limits */
65 uint64_t bps[2];
66
67 /* IOPS limits */
68 unsigned int iops[2];
69
70 /* Number of bytes disptached in current slice */
71 uint64_t bytes_disp[2];
72 /* Number of bio's dispatched in current slice */
73 unsigned int io_disp[2];
74
75 /* When did we start a new slice */
76 unsigned long slice_start[2];
77 unsigned long slice_end[2];
78
79 /* Some throttle limits got updated for the group */
80 int limits_changed;
81
82 struct rcu_head rcu_head;
83};
84
85struct throtl_data
86{
87 /* List of throtl groups */
88 struct hlist_head tg_list;
89
90 /* service tree for active throtl groups */
91 struct throtl_rb_root tg_service_tree;
92
93 struct throtl_grp *root_tg;
94 struct request_queue *queue;
95
96 /* Total Number of queued bios on READ and WRITE lists */
97 unsigned int nr_queued[2];
98
99 /*
100 * number of total undestroyed groups
101 */
102 unsigned int nr_undestroyed_grps;
103
104 /* Work for dispatching throttled bios */
105 struct delayed_work throtl_work;
106
107 int limits_changed;
108};
109
110enum tg_state_flags {
111 THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */
112};
113
114#define THROTL_TG_FNS(name) \
115static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \
116{ \
117 (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \
118} \
119static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \
120{ \
121 (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \
122} \
123static inline int throtl_tg_##name(const struct throtl_grp *tg) \
124{ \
125 return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \
126}
127
128THROTL_TG_FNS(on_rr);
129
130#define throtl_log_tg(td, tg, fmt, args...) \
131 blk_add_trace_msg((td)->queue, "throtl %s " fmt, \
132 blkg_path(&(tg)->blkg), ##args); \
133
134#define throtl_log(td, fmt, args...) \
135 blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
136
137static inline struct throtl_grp *tg_of_blkg(struct blkio_group *blkg)
138{
139 if (blkg)
140 return container_of(blkg, struct throtl_grp, blkg);
141
142 return NULL;
143}
144
145static inline unsigned int total_nr_queued(struct throtl_data *td)
146{
147 return td->nr_queued[0] + td->nr_queued[1];
148}
149
150static inline struct throtl_grp *throtl_ref_get_tg(struct throtl_grp *tg)
151{
152 atomic_inc(&tg->ref);
153 return tg;
154}
155
156static void throtl_free_tg(struct rcu_head *head)
157{
158 struct throtl_grp *tg;
159
160 tg = container_of(head, struct throtl_grp, rcu_head);
161 free_percpu(tg->blkg.stats_cpu);
162 kfree(tg);
163}
164
165static void throtl_put_tg(struct throtl_grp *tg)
166{
167 BUG_ON(atomic_read(&tg->ref) <= 0);
168 if (!atomic_dec_and_test(&tg->ref))
169 return;
170
171 /*
172 * A group is freed in rcu manner. But having an rcu lock does not
173 * mean that one can access all the fields of blkg and assume these
174 * are valid. For example, don't try to follow throtl_data and
175 * request queue links.
176 *
177 * Having a reference to blkg under an rcu allows acess to only
178 * values local to groups like group stats and group rate limits
179 */
180 call_rcu(&tg->rcu_head, throtl_free_tg);
181}
182
183static void throtl_init_group(struct throtl_grp *tg)
184{
185 INIT_HLIST_NODE(&tg->tg_node);
186 RB_CLEAR_NODE(&tg->rb_node);
187 bio_list_init(&tg->bio_lists[0]);
188 bio_list_init(&tg->bio_lists[1]);
189 tg->limits_changed = false;
190
191 /* Practically unlimited BW */
192 tg->bps[0] = tg->bps[1] = -1;
193 tg->iops[0] = tg->iops[1] = -1;
194
195 /*
196 * Take the initial reference that will be released on destroy
197 * This can be thought of a joint reference by cgroup and
198 * request queue which will be dropped by either request queue
199 * exit or cgroup deletion path depending on who is exiting first.
200 */
201 atomic_set(&tg->ref, 1);
202}
203
204/* Should be called with rcu read lock held (needed for blkcg) */
205static void
206throtl_add_group_to_td_list(struct throtl_data *td, struct throtl_grp *tg)
207{
208 hlist_add_head(&tg->tg_node, &td->tg_list);
209 td->nr_undestroyed_grps++;
210}
211
212static void
213__throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg)
214{
215 struct backing_dev_info *bdi = &td->queue->backing_dev_info;
216 unsigned int major, minor;
217
218 if (!tg || tg->blkg.dev)
219 return;
220
221 /*
222 * Fill in device details for a group which might not have been
223 * filled at group creation time as queue was being instantiated
224 * and driver had not attached a device yet
225 */
226 if (bdi->dev && dev_name(bdi->dev)) {
227 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
228 tg->blkg.dev = MKDEV(major, minor);
229 }
230}
231
232/*
233 * Should be called with without queue lock held. Here queue lock will be
234 * taken rarely. It will be taken only once during life time of a group
235 * if need be
236 */
237static void
238throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg)
239{
240 if (!tg || tg->blkg.dev)
241 return;
242
243 spin_lock_irq(td->queue->queue_lock);
244 __throtl_tg_fill_dev_details(td, tg);
245 spin_unlock_irq(td->queue->queue_lock);
246}
247
248static void throtl_init_add_tg_lists(struct throtl_data *td,
249 struct throtl_grp *tg, struct blkio_cgroup *blkcg)
250{
251 __throtl_tg_fill_dev_details(td, tg);
252
253 /* Add group onto cgroup list */
254 blkiocg_add_blkio_group(blkcg, &tg->blkg, (void *)td,
255 tg->blkg.dev, BLKIO_POLICY_THROTL);
256
257 tg->bps[READ] = blkcg_get_read_bps(blkcg, tg->blkg.dev);
258 tg->bps[WRITE] = blkcg_get_write_bps(blkcg, tg->blkg.dev);
259 tg->iops[READ] = blkcg_get_read_iops(blkcg, tg->blkg.dev);
260 tg->iops[WRITE] = blkcg_get_write_iops(blkcg, tg->blkg.dev);
261
262 throtl_add_group_to_td_list(td, tg);
263}
264
265/* Should be called without queue lock and outside of rcu period */
266static struct throtl_grp *throtl_alloc_tg(struct throtl_data *td)
267{
268 struct throtl_grp *tg = NULL;
269 int ret;
270
271 tg = kzalloc_node(sizeof(*tg), GFP_ATOMIC, td->queue->node);
272 if (!tg)
273 return NULL;
274
275 ret = blkio_alloc_blkg_stats(&tg->blkg);
276
277 if (ret) {
278 kfree(tg);
279 return NULL;
280 }
281
282 throtl_init_group(tg);
283 return tg;
284}
285
286static struct
287throtl_grp *throtl_find_tg(struct throtl_data *td, struct blkio_cgroup *blkcg)
288{
289 struct throtl_grp *tg = NULL;
290 void *key = td;
291
292 /*
293 * This is the common case when there are no blkio cgroups.
294 * Avoid lookup in this case
295 */
296 if (blkcg == &blkio_root_cgroup)
297 tg = td->root_tg;
298 else
299 tg = tg_of_blkg(blkiocg_lookup_group(blkcg, key));
300
301 __throtl_tg_fill_dev_details(td, tg);
302 return tg;
303}
304
305/*
306 * This function returns with queue lock unlocked in case of error, like
307 * request queue is no more
308 */
309static struct throtl_grp * throtl_get_tg(struct throtl_data *td)
310{
311 struct throtl_grp *tg = NULL, *__tg = NULL;
312 struct blkio_cgroup *blkcg;
313 struct request_queue *q = td->queue;
314
315 rcu_read_lock();
316 blkcg = task_blkio_cgroup(current);
317 tg = throtl_find_tg(td, blkcg);
318 if (tg) {
319 rcu_read_unlock();
320 return tg;
321 }
322
323 /*
324 * Need to allocate a group. Allocation of group also needs allocation
325 * of per cpu stats which in-turn takes a mutex() and can block. Hence
326 * we need to drop rcu lock and queue_lock before we call alloc
327 *
328 * Take the request queue reference to make sure queue does not
329 * go away once we return from allocation.
330 */
331 blk_get_queue(q);
332 rcu_read_unlock();
333 spin_unlock_irq(q->queue_lock);
334
335 tg = throtl_alloc_tg(td);
336 /*
337 * We might have slept in group allocation. Make sure queue is not
338 * dead
339 */
340 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
341 blk_put_queue(q);
342 if (tg)
343 kfree(tg);
344
345 return ERR_PTR(-ENODEV);
346 }
347 blk_put_queue(q);
348
349 /* Group allocated and queue is still alive. take the lock */
350 spin_lock_irq(q->queue_lock);
351
352 /*
353 * Initialize the new group. After sleeping, read the blkcg again.
354 */
355 rcu_read_lock();
356 blkcg = task_blkio_cgroup(current);
357
358 /*
359 * If some other thread already allocated the group while we were
360 * not holding queue lock, free up the group
361 */
362 __tg = throtl_find_tg(td, blkcg);
363
364 if (__tg) {
365 kfree(tg);
366 rcu_read_unlock();
367 return __tg;
368 }
369
370 /* Group allocation failed. Account the IO to root group */
371 if (!tg) {
372 tg = td->root_tg;
373 return tg;
374 }
375
376 throtl_init_add_tg_lists(td, tg, blkcg);
377 rcu_read_unlock();
378 return tg;
379}
380
381static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root)
382{
383 /* Service tree is empty */
384 if (!root->count)
385 return NULL;
386
387 if (!root->left)
388 root->left = rb_first(&root->rb);
389
390 if (root->left)
391 return rb_entry_tg(root->left);
392
393 return NULL;
394}
395
396static void rb_erase_init(struct rb_node *n, struct rb_root *root)
397{
398 rb_erase(n, root);
399 RB_CLEAR_NODE(n);
400}
401
402static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root)
403{
404 if (root->left == n)
405 root->left = NULL;
406 rb_erase_init(n, &root->rb);
407 --root->count;
408}
409
410static void update_min_dispatch_time(struct throtl_rb_root *st)
411{
412 struct throtl_grp *tg;
413
414 tg = throtl_rb_first(st);
415 if (!tg)
416 return;
417
418 st->min_disptime = tg->disptime;
419}
420
421static void
422tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg)
423{
424 struct rb_node **node = &st->rb.rb_node;
425 struct rb_node *parent = NULL;
426 struct throtl_grp *__tg;
427 unsigned long key = tg->disptime;
428 int left = 1;
429
430 while (*node != NULL) {
431 parent = *node;
432 __tg = rb_entry_tg(parent);
433
434 if (time_before(key, __tg->disptime))
435 node = &parent->rb_left;
436 else {
437 node = &parent->rb_right;
438 left = 0;
439 }
440 }
441
442 if (left)
443 st->left = &tg->rb_node;
444
445 rb_link_node(&tg->rb_node, parent, node);
446 rb_insert_color(&tg->rb_node, &st->rb);
447}
448
449static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
450{
451 struct throtl_rb_root *st = &td->tg_service_tree;
452
453 tg_service_tree_add(st, tg);
454 throtl_mark_tg_on_rr(tg);
455 st->count++;
456}
457
458static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
459{
460 if (!throtl_tg_on_rr(tg))
461 __throtl_enqueue_tg(td, tg);
462}
463
464static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
465{
466 throtl_rb_erase(&tg->rb_node, &td->tg_service_tree);
467 throtl_clear_tg_on_rr(tg);
468}
469
470static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
471{
472 if (throtl_tg_on_rr(tg))
473 __throtl_dequeue_tg(td, tg);
474}
475
476static void throtl_schedule_next_dispatch(struct throtl_data *td)
477{
478 struct throtl_rb_root *st = &td->tg_service_tree;
479
480 /*
481 * If there are more bios pending, schedule more work.
482 */
483 if (!total_nr_queued(td))
484 return;
485
486 BUG_ON(!st->count);
487
488 update_min_dispatch_time(st);
489
490 if (time_before_eq(st->min_disptime, jiffies))
491 throtl_schedule_delayed_work(td, 0);
492 else
493 throtl_schedule_delayed_work(td, (st->min_disptime - jiffies));
494}
495
496static inline void
497throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
498{
499 tg->bytes_disp[rw] = 0;
500 tg->io_disp[rw] = 0;
501 tg->slice_start[rw] = jiffies;
502 tg->slice_end[rw] = jiffies + throtl_slice;
503 throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
504 rw == READ ? 'R' : 'W', tg->slice_start[rw],
505 tg->slice_end[rw], jiffies);
506}
507
508static inline void throtl_set_slice_end(struct throtl_data *td,
509 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
510{
511 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
512}
513
514static inline void throtl_extend_slice(struct throtl_data *td,
515 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
516{
517 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
518 throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
519 rw == READ ? 'R' : 'W', tg->slice_start[rw],
520 tg->slice_end[rw], jiffies);
521}
522
523/* Determine if previously allocated or extended slice is complete or not */
524static bool
525throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
526{
527 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
528 return 0;
529
530 return 1;
531}
532
533/* Trim the used slices and adjust slice start accordingly */
534static inline void
535throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
536{
537 unsigned long nr_slices, time_elapsed, io_trim;
538 u64 bytes_trim, tmp;
539
540 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
541
542 /*
543 * If bps are unlimited (-1), then time slice don't get
544 * renewed. Don't try to trim the slice if slice is used. A new
545 * slice will start when appropriate.
546 */
547 if (throtl_slice_used(td, tg, rw))
548 return;
549
550 /*
551 * A bio has been dispatched. Also adjust slice_end. It might happen
552 * that initially cgroup limit was very low resulting in high
553 * slice_end, but later limit was bumped up and bio was dispached
554 * sooner, then we need to reduce slice_end. A high bogus slice_end
555 * is bad because it does not allow new slice to start.
556 */
557
558 throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice);
559
560 time_elapsed = jiffies - tg->slice_start[rw];
561
562 nr_slices = time_elapsed / throtl_slice;
563
564 if (!nr_slices)
565 return;
566 tmp = tg->bps[rw] * throtl_slice * nr_slices;
567 do_div(tmp, HZ);
568 bytes_trim = tmp;
569
570 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
571
572 if (!bytes_trim && !io_trim)
573 return;
574
575 if (tg->bytes_disp[rw] >= bytes_trim)
576 tg->bytes_disp[rw] -= bytes_trim;
577 else
578 tg->bytes_disp[rw] = 0;
579
580 if (tg->io_disp[rw] >= io_trim)
581 tg->io_disp[rw] -= io_trim;
582 else
583 tg->io_disp[rw] = 0;
584
585 tg->slice_start[rw] += nr_slices * throtl_slice;
586
587 throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
588 " start=%lu end=%lu jiffies=%lu",
589 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
590 tg->slice_start[rw], tg->slice_end[rw], jiffies);
591}
592
593static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
594 struct bio *bio, unsigned long *wait)
595{
596 bool rw = bio_data_dir(bio);
597 unsigned int io_allowed;
598 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
599 u64 tmp;
600
601 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
602
603 /* Slice has just started. Consider one slice interval */
604 if (!jiffy_elapsed)
605 jiffy_elapsed_rnd = throtl_slice;
606
607 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
608
609 /*
610 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
611 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
612 * will allow dispatch after 1 second and after that slice should
613 * have been trimmed.
614 */
615
616 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
617 do_div(tmp, HZ);
618
619 if (tmp > UINT_MAX)
620 io_allowed = UINT_MAX;
621 else
622 io_allowed = tmp;
623
624 if (tg->io_disp[rw] + 1 <= io_allowed) {
625 if (wait)
626 *wait = 0;
627 return 1;
628 }
629
630 /* Calc approx time to dispatch */
631 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
632
633 if (jiffy_wait > jiffy_elapsed)
634 jiffy_wait = jiffy_wait - jiffy_elapsed;
635 else
636 jiffy_wait = 1;
637
638 if (wait)
639 *wait = jiffy_wait;
640 return 0;
641}
642
643static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
644 struct bio *bio, unsigned long *wait)
645{
646 bool rw = bio_data_dir(bio);
647 u64 bytes_allowed, extra_bytes, tmp;
648 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
649
650 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
651
652 /* Slice has just started. Consider one slice interval */
653 if (!jiffy_elapsed)
654 jiffy_elapsed_rnd = throtl_slice;
655
656 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
657
658 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
659 do_div(tmp, HZ);
660 bytes_allowed = tmp;
661
662 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
663 if (wait)
664 *wait = 0;
665 return 1;
666 }
667
668 /* Calc approx time to dispatch */
669 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
670 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
671
672 if (!jiffy_wait)
673 jiffy_wait = 1;
674
675 /*
676 * This wait time is without taking into consideration the rounding
677 * up we did. Add that time also.
678 */
679 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
680 if (wait)
681 *wait = jiffy_wait;
682 return 0;
683}
684
685static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
686 if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
687 return 1;
688 return 0;
689}
690
691/*
692 * Returns whether one can dispatch a bio or not. Also returns approx number
693 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
694 */
695static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
696 struct bio *bio, unsigned long *wait)
697{
698 bool rw = bio_data_dir(bio);
699 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
700
701 /*
702 * Currently whole state machine of group depends on first bio
703 * queued in the group bio list. So one should not be calling
704 * this function with a different bio if there are other bios
705 * queued.
706 */
707 BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
708
709 /* If tg->bps = -1, then BW is unlimited */
710 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
711 if (wait)
712 *wait = 0;
713 return 1;
714 }
715
716 /*
717 * If previous slice expired, start a new one otherwise renew/extend
718 * existing slice to make sure it is at least throtl_slice interval
719 * long since now.
720 */
721 if (throtl_slice_used(td, tg, rw))
722 throtl_start_new_slice(td, tg, rw);
723 else {
724 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
725 throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
726 }
727
728 if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
729 && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
730 if (wait)
731 *wait = 0;
732 return 1;
733 }
734
735 max_wait = max(bps_wait, iops_wait);
736
737 if (wait)
738 *wait = max_wait;
739
740 if (time_before(tg->slice_end[rw], jiffies + max_wait))
741 throtl_extend_slice(td, tg, rw, jiffies + max_wait);
742
743 return 0;
744}
745
746static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
747{
748 bool rw = bio_data_dir(bio);
749 bool sync = rw_is_sync(bio->bi_rw);
750
751 /* Charge the bio to the group */
752 tg->bytes_disp[rw] += bio->bi_size;
753 tg->io_disp[rw]++;
754
755 blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync);
756}
757
758static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
759 struct bio *bio)
760{
761 bool rw = bio_data_dir(bio);
762
763 bio_list_add(&tg->bio_lists[rw], bio);
764 /* Take a bio reference on tg */
765 throtl_ref_get_tg(tg);
766 tg->nr_queued[rw]++;
767 td->nr_queued[rw]++;
768 throtl_enqueue_tg(td, tg);
769}
770
771static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
772{
773 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
774 struct bio *bio;
775
776 if ((bio = bio_list_peek(&tg->bio_lists[READ])))
777 tg_may_dispatch(td, tg, bio, &read_wait);
778
779 if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
780 tg_may_dispatch(td, tg, bio, &write_wait);
781
782 min_wait = min(read_wait, write_wait);
783 disptime = jiffies + min_wait;
784
785 /* Update dispatch time */
786 throtl_dequeue_tg(td, tg);
787 tg->disptime = disptime;
788 throtl_enqueue_tg(td, tg);
789}
790
791static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
792 bool rw, struct bio_list *bl)
793{
794 struct bio *bio;
795
796 bio = bio_list_pop(&tg->bio_lists[rw]);
797 tg->nr_queued[rw]--;
798 /* Drop bio reference on tg */
799 throtl_put_tg(tg);
800
801 BUG_ON(td->nr_queued[rw] <= 0);
802 td->nr_queued[rw]--;
803
804 throtl_charge_bio(tg, bio);
805 bio_list_add(bl, bio);
806 bio->bi_rw |= REQ_THROTTLED;
807
808 throtl_trim_slice(td, tg, rw);
809}
810
811static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
812 struct bio_list *bl)
813{
814 unsigned int nr_reads = 0, nr_writes = 0;
815 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
816 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
817 struct bio *bio;
818
819 /* Try to dispatch 75% READS and 25% WRITES */
820
821 while ((bio = bio_list_peek(&tg->bio_lists[READ]))
822 && tg_may_dispatch(td, tg, bio, NULL)) {
823
824 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
825 nr_reads++;
826
827 if (nr_reads >= max_nr_reads)
828 break;
829 }
830
831 while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
832 && tg_may_dispatch(td, tg, bio, NULL)) {
833
834 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
835 nr_writes++;
836
837 if (nr_writes >= max_nr_writes)
838 break;
839 }
840
841 return nr_reads + nr_writes;
842}
843
844static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
845{
846 unsigned int nr_disp = 0;
847 struct throtl_grp *tg;
848 struct throtl_rb_root *st = &td->tg_service_tree;
849
850 while (1) {
851 tg = throtl_rb_first(st);
852
853 if (!tg)
854 break;
855
856 if (time_before(jiffies, tg->disptime))
857 break;
858
859 throtl_dequeue_tg(td, tg);
860
861 nr_disp += throtl_dispatch_tg(td, tg, bl);
862
863 if (tg->nr_queued[0] || tg->nr_queued[1]) {
864 tg_update_disptime(td, tg);
865 throtl_enqueue_tg(td, tg);
866 }
867
868 if (nr_disp >= throtl_quantum)
869 break;
870 }
871
872 return nr_disp;
873}
874
875static void throtl_process_limit_change(struct throtl_data *td)
876{
877 struct throtl_grp *tg;
878 struct hlist_node *pos, *n;
879
880 if (!td->limits_changed)
881 return;
882
883 xchg(&td->limits_changed, false);
884
885 throtl_log(td, "limits changed");
886
887 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
888 if (!tg->limits_changed)
889 continue;
890
891 if (!xchg(&tg->limits_changed, false))
892 continue;
893
894 throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
895 " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
896 tg->iops[READ], tg->iops[WRITE]);
897
898 /*
899 * Restart the slices for both READ and WRITES. It
900 * might happen that a group's limit are dropped
901 * suddenly and we don't want to account recently
902 * dispatched IO with new low rate
903 */
904 throtl_start_new_slice(td, tg, 0);
905 throtl_start_new_slice(td, tg, 1);
906
907 if (throtl_tg_on_rr(tg))
908 tg_update_disptime(td, tg);
909 }
910}
911
912/* Dispatch throttled bios. Should be called without queue lock held. */
913static int throtl_dispatch(struct request_queue *q)
914{
915 struct throtl_data *td = q->td;
916 unsigned int nr_disp = 0;
917 struct bio_list bio_list_on_stack;
918 struct bio *bio;
919 struct blk_plug plug;
920
921 spin_lock_irq(q->queue_lock);
922
923 throtl_process_limit_change(td);
924
925 if (!total_nr_queued(td))
926 goto out;
927
928 bio_list_init(&bio_list_on_stack);
929
930 throtl_log(td, "dispatch nr_queued=%u read=%u write=%u",
931 total_nr_queued(td), td->nr_queued[READ],
932 td->nr_queued[WRITE]);
933
934 nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
935
936 if (nr_disp)
937 throtl_log(td, "bios disp=%u", nr_disp);
938
939 throtl_schedule_next_dispatch(td);
940out:
941 spin_unlock_irq(q->queue_lock);
942
943 /*
944 * If we dispatched some requests, unplug the queue to make sure
945 * immediate dispatch
946 */
947 if (nr_disp) {
948 blk_start_plug(&plug);
949 while((bio = bio_list_pop(&bio_list_on_stack)))
950 generic_make_request(bio);
951 blk_finish_plug(&plug);
952 }
953 return nr_disp;
954}
955
956void blk_throtl_work(struct work_struct *work)
957{
958 struct throtl_data *td = container_of(work, struct throtl_data,
959 throtl_work.work);
960 struct request_queue *q = td->queue;
961
962 throtl_dispatch(q);
963}
964
965/* Call with queue lock held */
966static void
967throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay)
968{
969
970 struct delayed_work *dwork = &td->throtl_work;
971
972 /* schedule work if limits changed even if no bio is queued */
973 if (total_nr_queued(td) || td->limits_changed) {
974 /*
975 * We might have a work scheduled to be executed in future.
976 * Cancel that and schedule a new one.
977 */
978 __cancel_delayed_work(dwork);
979 queue_delayed_work(kthrotld_workqueue, dwork, delay);
980 throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
981 delay, jiffies);
982 }
983}
984
985static void
986throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg)
987{
988 /* Something wrong if we are trying to remove same group twice */
989 BUG_ON(hlist_unhashed(&tg->tg_node));
990
991 hlist_del_init(&tg->tg_node);
992
993 /*
994 * Put the reference taken at the time of creation so that when all
995 * queues are gone, group can be destroyed.
996 */
997 throtl_put_tg(tg);
998 td->nr_undestroyed_grps--;
999}
1000
1001static void throtl_release_tgs(struct throtl_data *td)
1002{
1003 struct hlist_node *pos, *n;
1004 struct throtl_grp *tg;
1005
1006 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
1007 /*
1008 * If cgroup removal path got to blk_group first and removed
1009 * it from cgroup list, then it will take care of destroying
1010 * cfqg also.
1011 */
1012 if (!blkiocg_del_blkio_group(&tg->blkg))
1013 throtl_destroy_tg(td, tg);
1014 }
1015}
1016
1017static void throtl_td_free(struct throtl_data *td)
1018{
1019 kfree(td);
1020}
1021
1022/*
1023 * Blk cgroup controller notification saying that blkio_group object is being
1024 * delinked as associated cgroup object is going away. That also means that
1025 * no new IO will come in this group. So get rid of this group as soon as
1026 * any pending IO in the group is finished.
1027 *
1028 * This function is called under rcu_read_lock(). key is the rcu protected
1029 * pointer. That means "key" is a valid throtl_data pointer as long as we are
1030 * rcu read lock.
1031 *
1032 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1033 * it should not be NULL as even if queue was going away, cgroup deltion
1034 * path got to it first.
1035 */
1036void throtl_unlink_blkio_group(void *key, struct blkio_group *blkg)
1037{
1038 unsigned long flags;
1039 struct throtl_data *td = key;
1040
1041 spin_lock_irqsave(td->queue->queue_lock, flags);
1042 throtl_destroy_tg(td, tg_of_blkg(blkg));
1043 spin_unlock_irqrestore(td->queue->queue_lock, flags);
1044}
1045
1046static void throtl_update_blkio_group_common(struct throtl_data *td,
1047 struct throtl_grp *tg)
1048{
1049 xchg(&tg->limits_changed, true);
1050 xchg(&td->limits_changed, true);
1051 /* Schedule a work now to process the limit change */
1052 throtl_schedule_delayed_work(td, 0);
1053}
1054
1055/*
1056 * For all update functions, key should be a valid pointer because these
1057 * update functions are called under blkcg_lock, that means, blkg is
1058 * valid and in turn key is valid. queue exit path can not race because
1059 * of blkcg_lock
1060 *
1061 * Can not take queue lock in update functions as queue lock under blkcg_lock
1062 * is not allowed. Under other paths we take blkcg_lock under queue_lock.
1063 */
1064static void throtl_update_blkio_group_read_bps(void *key,
1065 struct blkio_group *blkg, u64 read_bps)
1066{
1067 struct throtl_data *td = key;
1068 struct throtl_grp *tg = tg_of_blkg(blkg);
1069
1070 tg->bps[READ] = read_bps;
1071 throtl_update_blkio_group_common(td, tg);
1072}
1073
1074static void throtl_update_blkio_group_write_bps(void *key,
1075 struct blkio_group *blkg, u64 write_bps)
1076{
1077 struct throtl_data *td = key;
1078 struct throtl_grp *tg = tg_of_blkg(blkg);
1079
1080 tg->bps[WRITE] = write_bps;
1081 throtl_update_blkio_group_common(td, tg);
1082}
1083
1084static void throtl_update_blkio_group_read_iops(void *key,
1085 struct blkio_group *blkg, unsigned int read_iops)
1086{
1087 struct throtl_data *td = key;
1088 struct throtl_grp *tg = tg_of_blkg(blkg);
1089
1090 tg->iops[READ] = read_iops;
1091 throtl_update_blkio_group_common(td, tg);
1092}
1093
1094static void throtl_update_blkio_group_write_iops(void *key,
1095 struct blkio_group *blkg, unsigned int write_iops)
1096{
1097 struct throtl_data *td = key;
1098 struct throtl_grp *tg = tg_of_blkg(blkg);
1099
1100 tg->iops[WRITE] = write_iops;
1101 throtl_update_blkio_group_common(td, tg);
1102}
1103
1104static void throtl_shutdown_wq(struct request_queue *q)
1105{
1106 struct throtl_data *td = q->td;
1107
1108 cancel_delayed_work_sync(&td->throtl_work);
1109}
1110
1111static struct blkio_policy_type blkio_policy_throtl = {
1112 .ops = {
1113 .blkio_unlink_group_fn = throtl_unlink_blkio_group,
1114 .blkio_update_group_read_bps_fn =
1115 throtl_update_blkio_group_read_bps,
1116 .blkio_update_group_write_bps_fn =
1117 throtl_update_blkio_group_write_bps,
1118 .blkio_update_group_read_iops_fn =
1119 throtl_update_blkio_group_read_iops,
1120 .blkio_update_group_write_iops_fn =
1121 throtl_update_blkio_group_write_iops,
1122 },
1123 .plid = BLKIO_POLICY_THROTL,
1124};
1125
1126int blk_throtl_bio(struct request_queue *q, struct bio **biop)
1127{
1128 struct throtl_data *td = q->td;
1129 struct throtl_grp *tg;
1130 struct bio *bio = *biop;
1131 bool rw = bio_data_dir(bio), update_disptime = true;
1132 struct blkio_cgroup *blkcg;
1133
1134 if (bio->bi_rw & REQ_THROTTLED) {
1135 bio->bi_rw &= ~REQ_THROTTLED;
1136 return 0;
1137 }
1138
1139 /*
1140 * A throtl_grp pointer retrieved under rcu can be used to access
1141 * basic fields like stats and io rates. If a group has no rules,
1142 * just update the dispatch stats in lockless manner and return.
1143 */
1144
1145 rcu_read_lock();
1146 blkcg = task_blkio_cgroup(current);
1147 tg = throtl_find_tg(td, blkcg);
1148 if (tg) {
1149 throtl_tg_fill_dev_details(td, tg);
1150
1151 if (tg_no_rule_group(tg, rw)) {
1152 blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size,
1153 rw, rw_is_sync(bio->bi_rw));
1154 rcu_read_unlock();
1155 return 0;
1156 }
1157 }
1158 rcu_read_unlock();
1159
1160 /*
1161 * Either group has not been allocated yet or it is not an unlimited
1162 * IO group
1163 */
1164
1165 spin_lock_irq(q->queue_lock);
1166 tg = throtl_get_tg(td);
1167
1168 if (IS_ERR(tg)) {
1169 if (PTR_ERR(tg) == -ENODEV) {
1170 /*
1171 * Queue is gone. No queue lock held here.
1172 */
1173 return -ENODEV;
1174 }
1175 }
1176
1177 if (tg->nr_queued[rw]) {
1178 /*
1179 * There is already another bio queued in same dir. No
1180 * need to update dispatch time.
1181 */
1182 update_disptime = false;
1183 goto queue_bio;
1184
1185 }
1186
1187 /* Bio is with-in rate limit of group */
1188 if (tg_may_dispatch(td, tg, bio, NULL)) {
1189 throtl_charge_bio(tg, bio);
1190
1191 /*
1192 * We need to trim slice even when bios are not being queued
1193 * otherwise it might happen that a bio is not queued for
1194 * a long time and slice keeps on extending and trim is not
1195 * called for a long time. Now if limits are reduced suddenly
1196 * we take into account all the IO dispatched so far at new
1197 * low rate and * newly queued IO gets a really long dispatch
1198 * time.
1199 *
1200 * So keep on trimming slice even if bio is not queued.
1201 */
1202 throtl_trim_slice(td, tg, rw);
1203 goto out;
1204 }
1205
1206queue_bio:
1207 throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
1208 " iodisp=%u iops=%u queued=%d/%d",
1209 rw == READ ? 'R' : 'W',
1210 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1211 tg->io_disp[rw], tg->iops[rw],
1212 tg->nr_queued[READ], tg->nr_queued[WRITE]);
1213
1214 throtl_add_bio_tg(q->td, tg, bio);
1215 *biop = NULL;
1216
1217 if (update_disptime) {
1218 tg_update_disptime(td, tg);
1219 throtl_schedule_next_dispatch(td);
1220 }
1221
1222out:
1223 spin_unlock_irq(q->queue_lock);
1224 return 0;
1225}
1226
1227int blk_throtl_init(struct request_queue *q)
1228{
1229 struct throtl_data *td;
1230 struct throtl_grp *tg;
1231
1232 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1233 if (!td)
1234 return -ENOMEM;
1235
1236 INIT_HLIST_HEAD(&td->tg_list);
1237 td->tg_service_tree = THROTL_RB_ROOT;
1238 td->limits_changed = false;
1239 INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);
1240
1241 /* alloc and Init root group. */
1242 td->queue = q;
1243 tg = throtl_alloc_tg(td);
1244
1245 if (!tg) {
1246 kfree(td);
1247 return -ENOMEM;
1248 }
1249
1250 td->root_tg = tg;
1251
1252 rcu_read_lock();
1253 throtl_init_add_tg_lists(td, tg, &blkio_root_cgroup);
1254 rcu_read_unlock();
1255
1256 /* Attach throtl data to request queue */
1257 q->td = td;
1258 return 0;
1259}
1260
1261void blk_throtl_exit(struct request_queue *q)
1262{
1263 struct throtl_data *td = q->td;
1264 bool wait = false;
1265
1266 BUG_ON(!td);
1267
1268 throtl_shutdown_wq(q);
1269
1270 spin_lock_irq(q->queue_lock);
1271 throtl_release_tgs(td);
1272
1273 /* If there are other groups */
1274 if (td->nr_undestroyed_grps > 0)
1275 wait = true;
1276
1277 spin_unlock_irq(q->queue_lock);
1278
1279 /*
1280 * Wait for tg->blkg->key accessors to exit their grace periods.
1281 * Do this wait only if there are other undestroyed groups out
1282 * there (other than root group). This can happen if cgroup deletion
1283 * path claimed the responsibility of cleaning up a group before
1284 * queue cleanup code get to the group.
1285 *
1286 * Do not call synchronize_rcu() unconditionally as there are drivers
1287 * which create/delete request queue hundreds of times during scan/boot
1288 * and synchronize_rcu() can take significant time and slow down boot.
1289 */
1290 if (wait)
1291 synchronize_rcu();
1292
1293 /*
1294 * Just being safe to make sure after previous flush if some body did
1295 * update limits through cgroup and another work got queued, cancel
1296 * it.
1297 */
1298 throtl_shutdown_wq(q);
1299 throtl_td_free(td);
1300}
1301
1302static int __init throtl_init(void)
1303{
1304 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1305 if (!kthrotld_workqueue)
1306 panic("Failed to create kthrotld\n");
1307
1308 blkio_policy_register(&blkio_policy_throtl);
1309 return 0;
1310}
1311
1312module_init(throtl_init);