MAINTAINERS: remove Xavier as maintainer of HISILICON ROCE DRIVER
[linux-2.6-microblaze.git] / block / blk-throttle.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Interface for controlling IO bandwidth on a request queue
4  *
5  * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
6  */
7
8 #include <linux/module.h>
9 #include <linux/slab.h>
10 #include <linux/blkdev.h>
11 #include <linux/bio.h>
12 #include <linux/blktrace_api.h>
13 #include <linux/blk-cgroup.h>
14 #include "blk.h"
15 #include "blk-cgroup-rwstat.h"
16
17 /* Max dispatch from a group in 1 round */
18 #define THROTL_GRP_QUANTUM 8
19
20 /* Total max dispatch from all groups in one round */
21 #define THROTL_QUANTUM 32
22
23 /* Throttling is performed over a slice and after that slice is renewed */
24 #define DFL_THROTL_SLICE_HD (HZ / 10)
25 #define DFL_THROTL_SLICE_SSD (HZ / 50)
26 #define MAX_THROTL_SLICE (HZ)
27 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
28 #define MIN_THROTL_BPS (320 * 1024)
29 #define MIN_THROTL_IOPS (10)
30 #define DFL_LATENCY_TARGET (-1L)
31 #define DFL_IDLE_THRESHOLD (0)
32 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
33 #define LATENCY_FILTERED_SSD (0)
34 /*
35  * For HD, very small latency comes from sequential IO. Such IO is helpless to
36  * help determine if its IO is impacted by others, hence we ignore the IO
37  */
38 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
39
40 static struct blkcg_policy blkcg_policy_throtl;
41
42 /* A workqueue to queue throttle related work */
43 static struct workqueue_struct *kthrotld_workqueue;
44
45 /*
46  * To implement hierarchical throttling, throtl_grps form a tree and bios
47  * are dispatched upwards level by level until they reach the top and get
48  * issued.  When dispatching bios from the children and local group at each
49  * level, if the bios are dispatched into a single bio_list, there's a risk
50  * of a local or child group which can queue many bios at once filling up
51  * the list starving others.
52  *
53  * To avoid such starvation, dispatched bios are queued separately
54  * according to where they came from.  When they are again dispatched to
55  * the parent, they're popped in round-robin order so that no single source
56  * hogs the dispatch window.
57  *
58  * throtl_qnode is used to keep the queued bios separated by their sources.
59  * Bios are queued to throtl_qnode which in turn is queued to
60  * throtl_service_queue and then dispatched in round-robin order.
61  *
62  * It's also used to track the reference counts on blkg's.  A qnode always
63  * belongs to a throtl_grp and gets queued on itself or the parent, so
64  * incrementing the reference of the associated throtl_grp when a qnode is
65  * queued and decrementing when dequeued is enough to keep the whole blkg
66  * tree pinned while bios are in flight.
67  */
68 struct throtl_qnode {
69         struct list_head        node;           /* service_queue->queued[] */
70         struct bio_list         bios;           /* queued bios */
71         struct throtl_grp       *tg;            /* tg this qnode belongs to */
72 };
73
74 struct throtl_service_queue {
75         struct throtl_service_queue *parent_sq; /* the parent service_queue */
76
77         /*
78          * Bios queued directly to this service_queue or dispatched from
79          * children throtl_grp's.
80          */
81         struct list_head        queued[2];      /* throtl_qnode [READ/WRITE] */
82         unsigned int            nr_queued[2];   /* number of queued bios */
83
84         /*
85          * RB tree of active children throtl_grp's, which are sorted by
86          * their ->disptime.
87          */
88         struct rb_root_cached   pending_tree;   /* RB tree of active tgs */
89         unsigned int            nr_pending;     /* # queued in the tree */
90         unsigned long           first_pending_disptime; /* disptime of the first tg */
91         struct timer_list       pending_timer;  /* fires on first_pending_disptime */
92 };
93
94 enum tg_state_flags {
95         THROTL_TG_PENDING       = 1 << 0,       /* on parent's pending tree */
96         THROTL_TG_WAS_EMPTY     = 1 << 1,       /* bio_lists[] became non-empty */
97 };
98
99 #define rb_entry_tg(node)       rb_entry((node), struct throtl_grp, rb_node)
100
101 enum {
102         LIMIT_LOW,
103         LIMIT_MAX,
104         LIMIT_CNT,
105 };
106
107 struct throtl_grp {
108         /* must be the first member */
109         struct blkg_policy_data pd;
110
111         /* active throtl group service_queue member */
112         struct rb_node rb_node;
113
114         /* throtl_data this group belongs to */
115         struct throtl_data *td;
116
117         /* this group's service queue */
118         struct throtl_service_queue service_queue;
119
120         /*
121          * qnode_on_self is used when bios are directly queued to this
122          * throtl_grp so that local bios compete fairly with bios
123          * dispatched from children.  qnode_on_parent is used when bios are
124          * dispatched from this throtl_grp into its parent and will compete
125          * with the sibling qnode_on_parents and the parent's
126          * qnode_on_self.
127          */
128         struct throtl_qnode qnode_on_self[2];
129         struct throtl_qnode qnode_on_parent[2];
130
131         /*
132          * Dispatch time in jiffies. This is the estimated time when group
133          * will unthrottle and is ready to dispatch more bio. It is used as
134          * key to sort active groups in service tree.
135          */
136         unsigned long disptime;
137
138         unsigned int flags;
139
140         /* are there any throtl rules between this group and td? */
141         bool has_rules[2];
142
143         /* internally used bytes per second rate limits */
144         uint64_t bps[2][LIMIT_CNT];
145         /* user configured bps limits */
146         uint64_t bps_conf[2][LIMIT_CNT];
147
148         /* internally used IOPS limits */
149         unsigned int iops[2][LIMIT_CNT];
150         /* user configured IOPS limits */
151         unsigned int iops_conf[2][LIMIT_CNT];
152
153         /* Number of bytes dispatched in current slice */
154         uint64_t bytes_disp[2];
155         /* Number of bio's dispatched in current slice */
156         unsigned int io_disp[2];
157
158         unsigned long last_low_overflow_time[2];
159
160         uint64_t last_bytes_disp[2];
161         unsigned int last_io_disp[2];
162
163         unsigned long last_check_time;
164
165         unsigned long latency_target; /* us */
166         unsigned long latency_target_conf; /* us */
167         /* When did we start a new slice */
168         unsigned long slice_start[2];
169         unsigned long slice_end[2];
170
171         unsigned long last_finish_time; /* ns / 1024 */
172         unsigned long checked_last_finish_time; /* ns / 1024 */
173         unsigned long avg_idletime; /* ns / 1024 */
174         unsigned long idletime_threshold; /* us */
175         unsigned long idletime_threshold_conf; /* us */
176
177         unsigned int bio_cnt; /* total bios */
178         unsigned int bad_bio_cnt; /* bios exceeding latency threshold */
179         unsigned long bio_cnt_reset_time;
180
181         struct blkg_rwstat stat_bytes;
182         struct blkg_rwstat stat_ios;
183 };
184
185 /* We measure latency for request size from <= 4k to >= 1M */
186 #define LATENCY_BUCKET_SIZE 9
187
188 struct latency_bucket {
189         unsigned long total_latency; /* ns / 1024 */
190         int samples;
191 };
192
193 struct avg_latency_bucket {
194         unsigned long latency; /* ns / 1024 */
195         bool valid;
196 };
197
198 struct throtl_data
199 {
200         /* service tree for active throtl groups */
201         struct throtl_service_queue service_queue;
202
203         struct request_queue *queue;
204
205         /* Total Number of queued bios on READ and WRITE lists */
206         unsigned int nr_queued[2];
207
208         unsigned int throtl_slice;
209
210         /* Work for dispatching throttled bios */
211         struct work_struct dispatch_work;
212         unsigned int limit_index;
213         bool limit_valid[LIMIT_CNT];
214
215         unsigned long low_upgrade_time;
216         unsigned long low_downgrade_time;
217
218         unsigned int scale;
219
220         struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
221         struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
222         struct latency_bucket __percpu *latency_buckets[2];
223         unsigned long last_calculate_time;
224         unsigned long filtered_latency;
225
226         bool track_bio_latency;
227 };
228
229 static void throtl_pending_timer_fn(struct timer_list *t);
230
231 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
232 {
233         return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
234 }
235
236 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
237 {
238         return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
239 }
240
241 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
242 {
243         return pd_to_blkg(&tg->pd);
244 }
245
246 /**
247  * sq_to_tg - return the throl_grp the specified service queue belongs to
248  * @sq: the throtl_service_queue of interest
249  *
250  * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
251  * embedded in throtl_data, %NULL is returned.
252  */
253 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
254 {
255         if (sq && sq->parent_sq)
256                 return container_of(sq, struct throtl_grp, service_queue);
257         else
258                 return NULL;
259 }
260
261 /**
262  * sq_to_td - return throtl_data the specified service queue belongs to
263  * @sq: the throtl_service_queue of interest
264  *
265  * A service_queue can be embedded in either a throtl_grp or throtl_data.
266  * Determine the associated throtl_data accordingly and return it.
267  */
268 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
269 {
270         struct throtl_grp *tg = sq_to_tg(sq);
271
272         if (tg)
273                 return tg->td;
274         else
275                 return container_of(sq, struct throtl_data, service_queue);
276 }
277
278 /*
279  * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
280  * make the IO dispatch more smooth.
281  * Scale up: linearly scale up according to lapsed time since upgrade. For
282  *           every throtl_slice, the limit scales up 1/2 .low limit till the
283  *           limit hits .max limit
284  * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
285  */
286 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
287 {
288         /* arbitrary value to avoid too big scale */
289         if (td->scale < 4096 && time_after_eq(jiffies,
290             td->low_upgrade_time + td->scale * td->throtl_slice))
291                 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
292
293         return low + (low >> 1) * td->scale;
294 }
295
296 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
297 {
298         struct blkcg_gq *blkg = tg_to_blkg(tg);
299         struct throtl_data *td;
300         uint64_t ret;
301
302         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
303                 return U64_MAX;
304
305         td = tg->td;
306         ret = tg->bps[rw][td->limit_index];
307         if (ret == 0 && td->limit_index == LIMIT_LOW) {
308                 /* intermediate node or iops isn't 0 */
309                 if (!list_empty(&blkg->blkcg->css.children) ||
310                     tg->iops[rw][td->limit_index])
311                         return U64_MAX;
312                 else
313                         return MIN_THROTL_BPS;
314         }
315
316         if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
317             tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
318                 uint64_t adjusted;
319
320                 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
321                 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
322         }
323         return ret;
324 }
325
326 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
327 {
328         struct blkcg_gq *blkg = tg_to_blkg(tg);
329         struct throtl_data *td;
330         unsigned int ret;
331
332         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
333                 return UINT_MAX;
334
335         td = tg->td;
336         ret = tg->iops[rw][td->limit_index];
337         if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
338                 /* intermediate node or bps isn't 0 */
339                 if (!list_empty(&blkg->blkcg->css.children) ||
340                     tg->bps[rw][td->limit_index])
341                         return UINT_MAX;
342                 else
343                         return MIN_THROTL_IOPS;
344         }
345
346         if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
347             tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
348                 uint64_t adjusted;
349
350                 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
351                 if (adjusted > UINT_MAX)
352                         adjusted = UINT_MAX;
353                 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
354         }
355         return ret;
356 }
357
358 #define request_bucket_index(sectors) \
359         clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
360
361 /**
362  * throtl_log - log debug message via blktrace
363  * @sq: the service_queue being reported
364  * @fmt: printf format string
365  * @args: printf args
366  *
367  * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
368  * throtl_grp; otherwise, just "throtl".
369  */
370 #define throtl_log(sq, fmt, args...)    do {                            \
371         struct throtl_grp *__tg = sq_to_tg((sq));                       \
372         struct throtl_data *__td = sq_to_td((sq));                      \
373                                                                         \
374         (void)__td;                                                     \
375         if (likely(!blk_trace_note_message_enabled(__td->queue)))       \
376                 break;                                                  \
377         if ((__tg)) {                                                   \
378                 blk_add_cgroup_trace_msg(__td->queue,                   \
379                         tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\
380         } else {                                                        \
381                 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);  \
382         }                                                               \
383 } while (0)
384
385 static inline unsigned int throtl_bio_data_size(struct bio *bio)
386 {
387         /* assume it's one sector */
388         if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
389                 return 512;
390         return bio->bi_iter.bi_size;
391 }
392
393 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
394 {
395         INIT_LIST_HEAD(&qn->node);
396         bio_list_init(&qn->bios);
397         qn->tg = tg;
398 }
399
400 /**
401  * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
402  * @bio: bio being added
403  * @qn: qnode to add bio to
404  * @queued: the service_queue->queued[] list @qn belongs to
405  *
406  * Add @bio to @qn and put @qn on @queued if it's not already on.
407  * @qn->tg's reference count is bumped when @qn is activated.  See the
408  * comment on top of throtl_qnode definition for details.
409  */
410 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
411                                  struct list_head *queued)
412 {
413         bio_list_add(&qn->bios, bio);
414         if (list_empty(&qn->node)) {
415                 list_add_tail(&qn->node, queued);
416                 blkg_get(tg_to_blkg(qn->tg));
417         }
418 }
419
420 /**
421  * throtl_peek_queued - peek the first bio on a qnode list
422  * @queued: the qnode list to peek
423  */
424 static struct bio *throtl_peek_queued(struct list_head *queued)
425 {
426         struct throtl_qnode *qn;
427         struct bio *bio;
428
429         if (list_empty(queued))
430                 return NULL;
431
432         qn = list_first_entry(queued, struct throtl_qnode, node);
433         bio = bio_list_peek(&qn->bios);
434         WARN_ON_ONCE(!bio);
435         return bio;
436 }
437
438 /**
439  * throtl_pop_queued - pop the first bio form a qnode list
440  * @queued: the qnode list to pop a bio from
441  * @tg_to_put: optional out argument for throtl_grp to put
442  *
443  * Pop the first bio from the qnode list @queued.  After popping, the first
444  * qnode is removed from @queued if empty or moved to the end of @queued so
445  * that the popping order is round-robin.
446  *
447  * When the first qnode is removed, its associated throtl_grp should be put
448  * too.  If @tg_to_put is NULL, this function automatically puts it;
449  * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
450  * responsible for putting it.
451  */
452 static struct bio *throtl_pop_queued(struct list_head *queued,
453                                      struct throtl_grp **tg_to_put)
454 {
455         struct throtl_qnode *qn;
456         struct bio *bio;
457
458         if (list_empty(queued))
459                 return NULL;
460
461         qn = list_first_entry(queued, struct throtl_qnode, node);
462         bio = bio_list_pop(&qn->bios);
463         WARN_ON_ONCE(!bio);
464
465         if (bio_list_empty(&qn->bios)) {
466                 list_del_init(&qn->node);
467                 if (tg_to_put)
468                         *tg_to_put = qn->tg;
469                 else
470                         blkg_put(tg_to_blkg(qn->tg));
471         } else {
472                 list_move_tail(&qn->node, queued);
473         }
474
475         return bio;
476 }
477
478 /* init a service_queue, assumes the caller zeroed it */
479 static void throtl_service_queue_init(struct throtl_service_queue *sq)
480 {
481         INIT_LIST_HEAD(&sq->queued[0]);
482         INIT_LIST_HEAD(&sq->queued[1]);
483         sq->pending_tree = RB_ROOT_CACHED;
484         timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
485 }
486
487 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp,
488                                                 struct request_queue *q,
489                                                 struct blkcg *blkcg)
490 {
491         struct throtl_grp *tg;
492         int rw;
493
494         tg = kzalloc_node(sizeof(*tg), gfp, q->node);
495         if (!tg)
496                 return NULL;
497
498         if (blkg_rwstat_init(&tg->stat_bytes, gfp))
499                 goto err_free_tg;
500
501         if (blkg_rwstat_init(&tg->stat_ios, gfp))
502                 goto err_exit_stat_bytes;
503
504         throtl_service_queue_init(&tg->service_queue);
505
506         for (rw = READ; rw <= WRITE; rw++) {
507                 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
508                 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
509         }
510
511         RB_CLEAR_NODE(&tg->rb_node);
512         tg->bps[READ][LIMIT_MAX] = U64_MAX;
513         tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
514         tg->iops[READ][LIMIT_MAX] = UINT_MAX;
515         tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
516         tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
517         tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
518         tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
519         tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
520         /* LIMIT_LOW will have default value 0 */
521
522         tg->latency_target = DFL_LATENCY_TARGET;
523         tg->latency_target_conf = DFL_LATENCY_TARGET;
524         tg->idletime_threshold = DFL_IDLE_THRESHOLD;
525         tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
526
527         return &tg->pd;
528
529 err_exit_stat_bytes:
530         blkg_rwstat_exit(&tg->stat_bytes);
531 err_free_tg:
532         kfree(tg);
533         return NULL;
534 }
535
536 static void throtl_pd_init(struct blkg_policy_data *pd)
537 {
538         struct throtl_grp *tg = pd_to_tg(pd);
539         struct blkcg_gq *blkg = tg_to_blkg(tg);
540         struct throtl_data *td = blkg->q->td;
541         struct throtl_service_queue *sq = &tg->service_queue;
542
543         /*
544          * If on the default hierarchy, we switch to properly hierarchical
545          * behavior where limits on a given throtl_grp are applied to the
546          * whole subtree rather than just the group itself.  e.g. If 16M
547          * read_bps limit is set on the root group, the whole system can't
548          * exceed 16M for the device.
549          *
550          * If not on the default hierarchy, the broken flat hierarchy
551          * behavior is retained where all throtl_grps are treated as if
552          * they're all separate root groups right below throtl_data.
553          * Limits of a group don't interact with limits of other groups
554          * regardless of the position of the group in the hierarchy.
555          */
556         sq->parent_sq = &td->service_queue;
557         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
558                 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
559         tg->td = td;
560 }
561
562 /*
563  * Set has_rules[] if @tg or any of its parents have limits configured.
564  * This doesn't require walking up to the top of the hierarchy as the
565  * parent's has_rules[] is guaranteed to be correct.
566  */
567 static void tg_update_has_rules(struct throtl_grp *tg)
568 {
569         struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
570         struct throtl_data *td = tg->td;
571         int rw;
572
573         for (rw = READ; rw <= WRITE; rw++)
574                 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
575                         (td->limit_valid[td->limit_index] &&
576                          (tg_bps_limit(tg, rw) != U64_MAX ||
577                           tg_iops_limit(tg, rw) != UINT_MAX));
578 }
579
580 static void throtl_pd_online(struct blkg_policy_data *pd)
581 {
582         struct throtl_grp *tg = pd_to_tg(pd);
583         /*
584          * We don't want new groups to escape the limits of its ancestors.
585          * Update has_rules[] after a new group is brought online.
586          */
587         tg_update_has_rules(tg);
588 }
589
590 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
591 static void blk_throtl_update_limit_valid(struct throtl_data *td)
592 {
593         struct cgroup_subsys_state *pos_css;
594         struct blkcg_gq *blkg;
595         bool low_valid = false;
596
597         rcu_read_lock();
598         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
599                 struct throtl_grp *tg = blkg_to_tg(blkg);
600
601                 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
602                     tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
603                         low_valid = true;
604                         break;
605                 }
606         }
607         rcu_read_unlock();
608
609         td->limit_valid[LIMIT_LOW] = low_valid;
610 }
611 #else
612 static inline void blk_throtl_update_limit_valid(struct throtl_data *td)
613 {
614 }
615 #endif
616
617 static void throtl_upgrade_state(struct throtl_data *td);
618 static void throtl_pd_offline(struct blkg_policy_data *pd)
619 {
620         struct throtl_grp *tg = pd_to_tg(pd);
621
622         tg->bps[READ][LIMIT_LOW] = 0;
623         tg->bps[WRITE][LIMIT_LOW] = 0;
624         tg->iops[READ][LIMIT_LOW] = 0;
625         tg->iops[WRITE][LIMIT_LOW] = 0;
626
627         blk_throtl_update_limit_valid(tg->td);
628
629         if (!tg->td->limit_valid[tg->td->limit_index])
630                 throtl_upgrade_state(tg->td);
631 }
632
633 static void throtl_pd_free(struct blkg_policy_data *pd)
634 {
635         struct throtl_grp *tg = pd_to_tg(pd);
636
637         del_timer_sync(&tg->service_queue.pending_timer);
638         blkg_rwstat_exit(&tg->stat_bytes);
639         blkg_rwstat_exit(&tg->stat_ios);
640         kfree(tg);
641 }
642
643 static struct throtl_grp *
644 throtl_rb_first(struct throtl_service_queue *parent_sq)
645 {
646         struct rb_node *n;
647
648         n = rb_first_cached(&parent_sq->pending_tree);
649         WARN_ON_ONCE(!n);
650         if (!n)
651                 return NULL;
652         return rb_entry_tg(n);
653 }
654
655 static void throtl_rb_erase(struct rb_node *n,
656                             struct throtl_service_queue *parent_sq)
657 {
658         rb_erase_cached(n, &parent_sq->pending_tree);
659         RB_CLEAR_NODE(n);
660         --parent_sq->nr_pending;
661 }
662
663 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
664 {
665         struct throtl_grp *tg;
666
667         tg = throtl_rb_first(parent_sq);
668         if (!tg)
669                 return;
670
671         parent_sq->first_pending_disptime = tg->disptime;
672 }
673
674 static void tg_service_queue_add(struct throtl_grp *tg)
675 {
676         struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
677         struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
678         struct rb_node *parent = NULL;
679         struct throtl_grp *__tg;
680         unsigned long key = tg->disptime;
681         bool leftmost = true;
682
683         while (*node != NULL) {
684                 parent = *node;
685                 __tg = rb_entry_tg(parent);
686
687                 if (time_before(key, __tg->disptime))
688                         node = &parent->rb_left;
689                 else {
690                         node = &parent->rb_right;
691                         leftmost = false;
692                 }
693         }
694
695         rb_link_node(&tg->rb_node, parent, node);
696         rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
697                                leftmost);
698 }
699
700 static void throtl_enqueue_tg(struct throtl_grp *tg)
701 {
702         if (!(tg->flags & THROTL_TG_PENDING)) {
703                 tg_service_queue_add(tg);
704                 tg->flags |= THROTL_TG_PENDING;
705                 tg->service_queue.parent_sq->nr_pending++;
706         }
707 }
708
709 static void throtl_dequeue_tg(struct throtl_grp *tg)
710 {
711         if (tg->flags & THROTL_TG_PENDING) {
712                 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
713                 tg->flags &= ~THROTL_TG_PENDING;
714         }
715 }
716
717 /* Call with queue lock held */
718 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
719                                           unsigned long expires)
720 {
721         unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
722
723         /*
724          * Since we are adjusting the throttle limit dynamically, the sleep
725          * time calculated according to previous limit might be invalid. It's
726          * possible the cgroup sleep time is very long and no other cgroups
727          * have IO running so notify the limit changes. Make sure the cgroup
728          * doesn't sleep too long to avoid the missed notification.
729          */
730         if (time_after(expires, max_expire))
731                 expires = max_expire;
732         mod_timer(&sq->pending_timer, expires);
733         throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
734                    expires - jiffies, jiffies);
735 }
736
737 /**
738  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
739  * @sq: the service_queue to schedule dispatch for
740  * @force: force scheduling
741  *
742  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
743  * dispatch time of the first pending child.  Returns %true if either timer
744  * is armed or there's no pending child left.  %false if the current
745  * dispatch window is still open and the caller should continue
746  * dispatching.
747  *
748  * If @force is %true, the dispatch timer is always scheduled and this
749  * function is guaranteed to return %true.  This is to be used when the
750  * caller can't dispatch itself and needs to invoke pending_timer
751  * unconditionally.  Note that forced scheduling is likely to induce short
752  * delay before dispatch starts even if @sq->first_pending_disptime is not
753  * in the future and thus shouldn't be used in hot paths.
754  */
755 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
756                                           bool force)
757 {
758         /* any pending children left? */
759         if (!sq->nr_pending)
760                 return true;
761
762         update_min_dispatch_time(sq);
763
764         /* is the next dispatch time in the future? */
765         if (force || time_after(sq->first_pending_disptime, jiffies)) {
766                 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
767                 return true;
768         }
769
770         /* tell the caller to continue dispatching */
771         return false;
772 }
773
774 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
775                 bool rw, unsigned long start)
776 {
777         tg->bytes_disp[rw] = 0;
778         tg->io_disp[rw] = 0;
779
780         /*
781          * Previous slice has expired. We must have trimmed it after last
782          * bio dispatch. That means since start of last slice, we never used
783          * that bandwidth. Do try to make use of that bandwidth while giving
784          * credit.
785          */
786         if (time_after_eq(start, tg->slice_start[rw]))
787                 tg->slice_start[rw] = start;
788
789         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
790         throtl_log(&tg->service_queue,
791                    "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
792                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
793                    tg->slice_end[rw], jiffies);
794 }
795
796 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
797 {
798         tg->bytes_disp[rw] = 0;
799         tg->io_disp[rw] = 0;
800         tg->slice_start[rw] = jiffies;
801         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
802         throtl_log(&tg->service_queue,
803                    "[%c] new slice start=%lu end=%lu jiffies=%lu",
804                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
805                    tg->slice_end[rw], jiffies);
806 }
807
808 static inline void throtl_set_slice_end(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 }
813
814 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
815                                        unsigned long jiffy_end)
816 {
817         throtl_set_slice_end(tg, rw, jiffy_end);
818         throtl_log(&tg->service_queue,
819                    "[%c] extend slice start=%lu end=%lu jiffies=%lu",
820                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
821                    tg->slice_end[rw], jiffies);
822 }
823
824 /* Determine if previously allocated or extended slice is complete or not */
825 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
826 {
827         if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
828                 return false;
829
830         return true;
831 }
832
833 /* Trim the used slices and adjust slice start accordingly */
834 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
835 {
836         unsigned long nr_slices, time_elapsed, io_trim;
837         u64 bytes_trim, tmp;
838
839         BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
840
841         /*
842          * If bps are unlimited (-1), then time slice don't get
843          * renewed. Don't try to trim the slice if slice is used. A new
844          * slice will start when appropriate.
845          */
846         if (throtl_slice_used(tg, rw))
847                 return;
848
849         /*
850          * A bio has been dispatched. Also adjust slice_end. It might happen
851          * that initially cgroup limit was very low resulting in high
852          * slice_end, but later limit was bumped up and bio was dispatched
853          * sooner, then we need to reduce slice_end. A high bogus slice_end
854          * is bad because it does not allow new slice to start.
855          */
856
857         throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
858
859         time_elapsed = jiffies - tg->slice_start[rw];
860
861         nr_slices = time_elapsed / tg->td->throtl_slice;
862
863         if (!nr_slices)
864                 return;
865         tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
866         do_div(tmp, HZ);
867         bytes_trim = tmp;
868
869         io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
870                 HZ;
871
872         if (!bytes_trim && !io_trim)
873                 return;
874
875         if (tg->bytes_disp[rw] >= bytes_trim)
876                 tg->bytes_disp[rw] -= bytes_trim;
877         else
878                 tg->bytes_disp[rw] = 0;
879
880         if (tg->io_disp[rw] >= io_trim)
881                 tg->io_disp[rw] -= io_trim;
882         else
883                 tg->io_disp[rw] = 0;
884
885         tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
886
887         throtl_log(&tg->service_queue,
888                    "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
889                    rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
890                    tg->slice_start[rw], tg->slice_end[rw], jiffies);
891 }
892
893 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
894                                   u32 iops_limit, unsigned long *wait)
895 {
896         bool rw = bio_data_dir(bio);
897         unsigned int io_allowed;
898         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
899         u64 tmp;
900
901         if (iops_limit == UINT_MAX) {
902                 if (wait)
903                         *wait = 0;
904                 return true;
905         }
906
907         jiffy_elapsed = jiffies - tg->slice_start[rw];
908
909         /* Round up to the next throttle slice, wait time must be nonzero */
910         jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
911
912         /*
913          * jiffy_elapsed_rnd should not be a big value as minimum iops can be
914          * 1 then at max jiffy elapsed should be equivalent of 1 second as we
915          * will allow dispatch after 1 second and after that slice should
916          * have been trimmed.
917          */
918
919         tmp = (u64)iops_limit * jiffy_elapsed_rnd;
920         do_div(tmp, HZ);
921
922         if (tmp > UINT_MAX)
923                 io_allowed = UINT_MAX;
924         else
925                 io_allowed = tmp;
926
927         if (tg->io_disp[rw] + 1 <= io_allowed) {
928                 if (wait)
929                         *wait = 0;
930                 return true;
931         }
932
933         /* Calc approx time to dispatch */
934         jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
935
936         if (wait)
937                 *wait = jiffy_wait;
938         return false;
939 }
940
941 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
942                                  u64 bps_limit, unsigned long *wait)
943 {
944         bool rw = bio_data_dir(bio);
945         u64 bytes_allowed, extra_bytes, tmp;
946         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
947         unsigned int bio_size = throtl_bio_data_size(bio);
948
949         if (bps_limit == U64_MAX) {
950                 if (wait)
951                         *wait = 0;
952                 return true;
953         }
954
955         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
956
957         /* Slice has just started. Consider one slice interval */
958         if (!jiffy_elapsed)
959                 jiffy_elapsed_rnd = tg->td->throtl_slice;
960
961         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
962
963         tmp = bps_limit * jiffy_elapsed_rnd;
964         do_div(tmp, HZ);
965         bytes_allowed = tmp;
966
967         if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
968                 if (wait)
969                         *wait = 0;
970                 return true;
971         }
972
973         /* Calc approx time to dispatch */
974         extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
975         jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
976
977         if (!jiffy_wait)
978                 jiffy_wait = 1;
979
980         /*
981          * This wait time is without taking into consideration the rounding
982          * up we did. Add that time also.
983          */
984         jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
985         if (wait)
986                 *wait = jiffy_wait;
987         return false;
988 }
989
990 /*
991  * Returns whether one can dispatch a bio or not. Also returns approx number
992  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
993  */
994 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
995                             unsigned long *wait)
996 {
997         bool rw = bio_data_dir(bio);
998         unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
999         u64 bps_limit = tg_bps_limit(tg, rw);
1000         u32 iops_limit = tg_iops_limit(tg, rw);
1001
1002         /*
1003          * Currently whole state machine of group depends on first bio
1004          * queued in the group bio list. So one should not be calling
1005          * this function with a different bio if there are other bios
1006          * queued.
1007          */
1008         BUG_ON(tg->service_queue.nr_queued[rw] &&
1009                bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
1010
1011         /* If tg->bps = -1, then BW is unlimited */
1012         if (bps_limit == U64_MAX && iops_limit == UINT_MAX) {
1013                 if (wait)
1014                         *wait = 0;
1015                 return true;
1016         }
1017
1018         /*
1019          * If previous slice expired, start a new one otherwise renew/extend
1020          * existing slice to make sure it is at least throtl_slice interval
1021          * long since now. New slice is started only for empty throttle group.
1022          * If there is queued bio, that means there should be an active
1023          * slice and it should be extended instead.
1024          */
1025         if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
1026                 throtl_start_new_slice(tg, rw);
1027         else {
1028                 if (time_before(tg->slice_end[rw],
1029                     jiffies + tg->td->throtl_slice))
1030                         throtl_extend_slice(tg, rw,
1031                                 jiffies + tg->td->throtl_slice);
1032         }
1033
1034         if (tg_with_in_bps_limit(tg, bio, bps_limit, &bps_wait) &&
1035             tg_with_in_iops_limit(tg, bio, iops_limit, &iops_wait)) {
1036                 if (wait)
1037                         *wait = 0;
1038                 return true;
1039         }
1040
1041         max_wait = max(bps_wait, iops_wait);
1042
1043         if (wait)
1044                 *wait = max_wait;
1045
1046         if (time_before(tg->slice_end[rw], jiffies + max_wait))
1047                 throtl_extend_slice(tg, rw, jiffies + max_wait);
1048
1049         return false;
1050 }
1051
1052 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1053 {
1054         bool rw = bio_data_dir(bio);
1055         unsigned int bio_size = throtl_bio_data_size(bio);
1056
1057         /* Charge the bio to the group */
1058         tg->bytes_disp[rw] += bio_size;
1059         tg->io_disp[rw]++;
1060         tg->last_bytes_disp[rw] += bio_size;
1061         tg->last_io_disp[rw]++;
1062
1063         /*
1064          * BIO_THROTTLED is used to prevent the same bio to be throttled
1065          * more than once as a throttled bio will go through blk-throtl the
1066          * second time when it eventually gets issued.  Set it when a bio
1067          * is being charged to a tg.
1068          */
1069         if (!bio_flagged(bio, BIO_THROTTLED))
1070                 bio_set_flag(bio, BIO_THROTTLED);
1071 }
1072
1073 /**
1074  * throtl_add_bio_tg - add a bio to the specified throtl_grp
1075  * @bio: bio to add
1076  * @qn: qnode to use
1077  * @tg: the target throtl_grp
1078  *
1079  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
1080  * tg->qnode_on_self[] is used.
1081  */
1082 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1083                               struct throtl_grp *tg)
1084 {
1085         struct throtl_service_queue *sq = &tg->service_queue;
1086         bool rw = bio_data_dir(bio);
1087
1088         if (!qn)
1089                 qn = &tg->qnode_on_self[rw];
1090
1091         /*
1092          * If @tg doesn't currently have any bios queued in the same
1093          * direction, queueing @bio can change when @tg should be
1094          * dispatched.  Mark that @tg was empty.  This is automatically
1095          * cleared on the next tg_update_disptime().
1096          */
1097         if (!sq->nr_queued[rw])
1098                 tg->flags |= THROTL_TG_WAS_EMPTY;
1099
1100         throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1101
1102         sq->nr_queued[rw]++;
1103         throtl_enqueue_tg(tg);
1104 }
1105
1106 static void tg_update_disptime(struct throtl_grp *tg)
1107 {
1108         struct throtl_service_queue *sq = &tg->service_queue;
1109         unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1110         struct bio *bio;
1111
1112         bio = throtl_peek_queued(&sq->queued[READ]);
1113         if (bio)
1114                 tg_may_dispatch(tg, bio, &read_wait);
1115
1116         bio = throtl_peek_queued(&sq->queued[WRITE]);
1117         if (bio)
1118                 tg_may_dispatch(tg, bio, &write_wait);
1119
1120         min_wait = min(read_wait, write_wait);
1121         disptime = jiffies + min_wait;
1122
1123         /* Update dispatch time */
1124         throtl_dequeue_tg(tg);
1125         tg->disptime = disptime;
1126         throtl_enqueue_tg(tg);
1127
1128         /* see throtl_add_bio_tg() */
1129         tg->flags &= ~THROTL_TG_WAS_EMPTY;
1130 }
1131
1132 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1133                                         struct throtl_grp *parent_tg, bool rw)
1134 {
1135         if (throtl_slice_used(parent_tg, rw)) {
1136                 throtl_start_new_slice_with_credit(parent_tg, rw,
1137                                 child_tg->slice_start[rw]);
1138         }
1139
1140 }
1141
1142 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1143 {
1144         struct throtl_service_queue *sq = &tg->service_queue;
1145         struct throtl_service_queue *parent_sq = sq->parent_sq;
1146         struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1147         struct throtl_grp *tg_to_put = NULL;
1148         struct bio *bio;
1149
1150         /*
1151          * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1152          * from @tg may put its reference and @parent_sq might end up
1153          * getting released prematurely.  Remember the tg to put and put it
1154          * after @bio is transferred to @parent_sq.
1155          */
1156         bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1157         sq->nr_queued[rw]--;
1158
1159         throtl_charge_bio(tg, bio);
1160
1161         /*
1162          * If our parent is another tg, we just need to transfer @bio to
1163          * the parent using throtl_add_bio_tg().  If our parent is
1164          * @td->service_queue, @bio is ready to be issued.  Put it on its
1165          * bio_lists[] and decrease total number queued.  The caller is
1166          * responsible for issuing these bios.
1167          */
1168         if (parent_tg) {
1169                 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1170                 start_parent_slice_with_credit(tg, parent_tg, rw);
1171         } else {
1172                 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1173                                      &parent_sq->queued[rw]);
1174                 BUG_ON(tg->td->nr_queued[rw] <= 0);
1175                 tg->td->nr_queued[rw]--;
1176         }
1177
1178         throtl_trim_slice(tg, rw);
1179
1180         if (tg_to_put)
1181                 blkg_put(tg_to_blkg(tg_to_put));
1182 }
1183
1184 static int throtl_dispatch_tg(struct throtl_grp *tg)
1185 {
1186         struct throtl_service_queue *sq = &tg->service_queue;
1187         unsigned int nr_reads = 0, nr_writes = 0;
1188         unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1189         unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1190         struct bio *bio;
1191
1192         /* Try to dispatch 75% READS and 25% WRITES */
1193
1194         while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1195                tg_may_dispatch(tg, bio, NULL)) {
1196
1197                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1198                 nr_reads++;
1199
1200                 if (nr_reads >= max_nr_reads)
1201                         break;
1202         }
1203
1204         while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1205                tg_may_dispatch(tg, bio, NULL)) {
1206
1207                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1208                 nr_writes++;
1209
1210                 if (nr_writes >= max_nr_writes)
1211                         break;
1212         }
1213
1214         return nr_reads + nr_writes;
1215 }
1216
1217 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1218 {
1219         unsigned int nr_disp = 0;
1220
1221         while (1) {
1222                 struct throtl_grp *tg;
1223                 struct throtl_service_queue *sq;
1224
1225                 if (!parent_sq->nr_pending)
1226                         break;
1227
1228                 tg = throtl_rb_first(parent_sq);
1229                 if (!tg)
1230                         break;
1231
1232                 if (time_before(jiffies, tg->disptime))
1233                         break;
1234
1235                 throtl_dequeue_tg(tg);
1236
1237                 nr_disp += throtl_dispatch_tg(tg);
1238
1239                 sq = &tg->service_queue;
1240                 if (sq->nr_queued[0] || sq->nr_queued[1])
1241                         tg_update_disptime(tg);
1242
1243                 if (nr_disp >= THROTL_QUANTUM)
1244                         break;
1245         }
1246
1247         return nr_disp;
1248 }
1249
1250 static bool throtl_can_upgrade(struct throtl_data *td,
1251         struct throtl_grp *this_tg);
1252 /**
1253  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1254  * @t: the pending_timer member of the throtl_service_queue being serviced
1255  *
1256  * This timer is armed when a child throtl_grp with active bio's become
1257  * pending and queued on the service_queue's pending_tree and expires when
1258  * the first child throtl_grp should be dispatched.  This function
1259  * dispatches bio's from the children throtl_grps to the parent
1260  * service_queue.
1261  *
1262  * If the parent's parent is another throtl_grp, dispatching is propagated
1263  * by either arming its pending_timer or repeating dispatch directly.  If
1264  * the top-level service_tree is reached, throtl_data->dispatch_work is
1265  * kicked so that the ready bio's are issued.
1266  */
1267 static void throtl_pending_timer_fn(struct timer_list *t)
1268 {
1269         struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1270         struct throtl_grp *tg = sq_to_tg(sq);
1271         struct throtl_data *td = sq_to_td(sq);
1272         struct request_queue *q = td->queue;
1273         struct throtl_service_queue *parent_sq;
1274         bool dispatched;
1275         int ret;
1276
1277         spin_lock_irq(&q->queue_lock);
1278         if (throtl_can_upgrade(td, NULL))
1279                 throtl_upgrade_state(td);
1280
1281 again:
1282         parent_sq = sq->parent_sq;
1283         dispatched = false;
1284
1285         while (true) {
1286                 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1287                            sq->nr_queued[READ] + sq->nr_queued[WRITE],
1288                            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1289
1290                 ret = throtl_select_dispatch(sq);
1291                 if (ret) {
1292                         throtl_log(sq, "bios disp=%u", ret);
1293                         dispatched = true;
1294                 }
1295
1296                 if (throtl_schedule_next_dispatch(sq, false))
1297                         break;
1298
1299                 /* this dispatch windows is still open, relax and repeat */
1300                 spin_unlock_irq(&q->queue_lock);
1301                 cpu_relax();
1302                 spin_lock_irq(&q->queue_lock);
1303         }
1304
1305         if (!dispatched)
1306                 goto out_unlock;
1307
1308         if (parent_sq) {
1309                 /* @parent_sq is another throl_grp, propagate dispatch */
1310                 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1311                         tg_update_disptime(tg);
1312                         if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1313                                 /* window is already open, repeat dispatching */
1314                                 sq = parent_sq;
1315                                 tg = sq_to_tg(sq);
1316                                 goto again;
1317                         }
1318                 }
1319         } else {
1320                 /* reached the top-level, queue issuing */
1321                 queue_work(kthrotld_workqueue, &td->dispatch_work);
1322         }
1323 out_unlock:
1324         spin_unlock_irq(&q->queue_lock);
1325 }
1326
1327 /**
1328  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1329  * @work: work item being executed
1330  *
1331  * This function is queued for execution when bios reach the bio_lists[]
1332  * of throtl_data->service_queue.  Those bios are ready and issued by this
1333  * function.
1334  */
1335 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1336 {
1337         struct throtl_data *td = container_of(work, struct throtl_data,
1338                                               dispatch_work);
1339         struct throtl_service_queue *td_sq = &td->service_queue;
1340         struct request_queue *q = td->queue;
1341         struct bio_list bio_list_on_stack;
1342         struct bio *bio;
1343         struct blk_plug plug;
1344         int rw;
1345
1346         bio_list_init(&bio_list_on_stack);
1347
1348         spin_lock_irq(&q->queue_lock);
1349         for (rw = READ; rw <= WRITE; rw++)
1350                 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1351                         bio_list_add(&bio_list_on_stack, bio);
1352         spin_unlock_irq(&q->queue_lock);
1353
1354         if (!bio_list_empty(&bio_list_on_stack)) {
1355                 blk_start_plug(&plug);
1356                 while ((bio = bio_list_pop(&bio_list_on_stack)))
1357                         submit_bio_noacct(bio);
1358                 blk_finish_plug(&plug);
1359         }
1360 }
1361
1362 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1363                               int off)
1364 {
1365         struct throtl_grp *tg = pd_to_tg(pd);
1366         u64 v = *(u64 *)((void *)tg + off);
1367
1368         if (v == U64_MAX)
1369                 return 0;
1370         return __blkg_prfill_u64(sf, pd, v);
1371 }
1372
1373 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1374                                int off)
1375 {
1376         struct throtl_grp *tg = pd_to_tg(pd);
1377         unsigned int v = *(unsigned int *)((void *)tg + off);
1378
1379         if (v == UINT_MAX)
1380                 return 0;
1381         return __blkg_prfill_u64(sf, pd, v);
1382 }
1383
1384 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1385 {
1386         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1387                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1388         return 0;
1389 }
1390
1391 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1392 {
1393         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1394                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1395         return 0;
1396 }
1397
1398 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1399 {
1400         struct throtl_service_queue *sq = &tg->service_queue;
1401         struct cgroup_subsys_state *pos_css;
1402         struct blkcg_gq *blkg;
1403
1404         throtl_log(&tg->service_queue,
1405                    "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1406                    tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1407                    tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1408
1409         /*
1410          * Update has_rules[] flags for the updated tg's subtree.  A tg is
1411          * considered to have rules if either the tg itself or any of its
1412          * ancestors has rules.  This identifies groups without any
1413          * restrictions in the whole hierarchy and allows them to bypass
1414          * blk-throttle.
1415          */
1416         blkg_for_each_descendant_pre(blkg, pos_css,
1417                         global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1418                 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1419                 struct throtl_grp *parent_tg;
1420
1421                 tg_update_has_rules(this_tg);
1422                 /* ignore root/second level */
1423                 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1424                     !blkg->parent->parent)
1425                         continue;
1426                 parent_tg = blkg_to_tg(blkg->parent);
1427                 /*
1428                  * make sure all children has lower idle time threshold and
1429                  * higher latency target
1430                  */
1431                 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1432                                 parent_tg->idletime_threshold);
1433                 this_tg->latency_target = max(this_tg->latency_target,
1434                                 parent_tg->latency_target);
1435         }
1436
1437         /*
1438          * We're already holding queue_lock and know @tg is valid.  Let's
1439          * apply the new config directly.
1440          *
1441          * Restart the slices for both READ and WRITES. It might happen
1442          * that a group's limit are dropped suddenly and we don't want to
1443          * account recently dispatched IO with new low rate.
1444          */
1445         throtl_start_new_slice(tg, READ);
1446         throtl_start_new_slice(tg, WRITE);
1447
1448         if (tg->flags & THROTL_TG_PENDING) {
1449                 tg_update_disptime(tg);
1450                 throtl_schedule_next_dispatch(sq->parent_sq, true);
1451         }
1452 }
1453
1454 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1455                            char *buf, size_t nbytes, loff_t off, bool is_u64)
1456 {
1457         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1458         struct blkg_conf_ctx ctx;
1459         struct throtl_grp *tg;
1460         int ret;
1461         u64 v;
1462
1463         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1464         if (ret)
1465                 return ret;
1466
1467         ret = -EINVAL;
1468         if (sscanf(ctx.body, "%llu", &v) != 1)
1469                 goto out_finish;
1470         if (!v)
1471                 v = U64_MAX;
1472
1473         tg = blkg_to_tg(ctx.blkg);
1474
1475         if (is_u64)
1476                 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1477         else
1478                 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1479
1480         tg_conf_updated(tg, false);
1481         ret = 0;
1482 out_finish:
1483         blkg_conf_finish(&ctx);
1484         return ret ?: nbytes;
1485 }
1486
1487 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1488                                char *buf, size_t nbytes, loff_t off)
1489 {
1490         return tg_set_conf(of, buf, nbytes, off, true);
1491 }
1492
1493 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1494                                 char *buf, size_t nbytes, loff_t off)
1495 {
1496         return tg_set_conf(of, buf, nbytes, off, false);
1497 }
1498
1499 static int tg_print_rwstat(struct seq_file *sf, void *v)
1500 {
1501         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1502                           blkg_prfill_rwstat, &blkcg_policy_throtl,
1503                           seq_cft(sf)->private, true);
1504         return 0;
1505 }
1506
1507 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1508                                       struct blkg_policy_data *pd, int off)
1509 {
1510         struct blkg_rwstat_sample sum;
1511
1512         blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1513                                   &sum);
1514         return __blkg_prfill_rwstat(sf, pd, &sum);
1515 }
1516
1517 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1518 {
1519         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1520                           tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1521                           seq_cft(sf)->private, true);
1522         return 0;
1523 }
1524
1525 static struct cftype throtl_legacy_files[] = {
1526         {
1527                 .name = "throttle.read_bps_device",
1528                 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1529                 .seq_show = tg_print_conf_u64,
1530                 .write = tg_set_conf_u64,
1531         },
1532         {
1533                 .name = "throttle.write_bps_device",
1534                 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1535                 .seq_show = tg_print_conf_u64,
1536                 .write = tg_set_conf_u64,
1537         },
1538         {
1539                 .name = "throttle.read_iops_device",
1540                 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1541                 .seq_show = tg_print_conf_uint,
1542                 .write = tg_set_conf_uint,
1543         },
1544         {
1545                 .name = "throttle.write_iops_device",
1546                 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1547                 .seq_show = tg_print_conf_uint,
1548                 .write = tg_set_conf_uint,
1549         },
1550         {
1551                 .name = "throttle.io_service_bytes",
1552                 .private = offsetof(struct throtl_grp, stat_bytes),
1553                 .seq_show = tg_print_rwstat,
1554         },
1555         {
1556                 .name = "throttle.io_service_bytes_recursive",
1557                 .private = offsetof(struct throtl_grp, stat_bytes),
1558                 .seq_show = tg_print_rwstat_recursive,
1559         },
1560         {
1561                 .name = "throttle.io_serviced",
1562                 .private = offsetof(struct throtl_grp, stat_ios),
1563                 .seq_show = tg_print_rwstat,
1564         },
1565         {
1566                 .name = "throttle.io_serviced_recursive",
1567                 .private = offsetof(struct throtl_grp, stat_ios),
1568                 .seq_show = tg_print_rwstat_recursive,
1569         },
1570         { }     /* terminate */
1571 };
1572
1573 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1574                          int off)
1575 {
1576         struct throtl_grp *tg = pd_to_tg(pd);
1577         const char *dname = blkg_dev_name(pd->blkg);
1578         char bufs[4][21] = { "max", "max", "max", "max" };
1579         u64 bps_dft;
1580         unsigned int iops_dft;
1581         char idle_time[26] = "";
1582         char latency_time[26] = "";
1583
1584         if (!dname)
1585                 return 0;
1586
1587         if (off == LIMIT_LOW) {
1588                 bps_dft = 0;
1589                 iops_dft = 0;
1590         } else {
1591                 bps_dft = U64_MAX;
1592                 iops_dft = UINT_MAX;
1593         }
1594
1595         if (tg->bps_conf[READ][off] == bps_dft &&
1596             tg->bps_conf[WRITE][off] == bps_dft &&
1597             tg->iops_conf[READ][off] == iops_dft &&
1598             tg->iops_conf[WRITE][off] == iops_dft &&
1599             (off != LIMIT_LOW ||
1600              (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1601               tg->latency_target_conf == DFL_LATENCY_TARGET)))
1602                 return 0;
1603
1604         if (tg->bps_conf[READ][off] != U64_MAX)
1605                 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1606                         tg->bps_conf[READ][off]);
1607         if (tg->bps_conf[WRITE][off] != U64_MAX)
1608                 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1609                         tg->bps_conf[WRITE][off]);
1610         if (tg->iops_conf[READ][off] != UINT_MAX)
1611                 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1612                         tg->iops_conf[READ][off]);
1613         if (tg->iops_conf[WRITE][off] != UINT_MAX)
1614                 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1615                         tg->iops_conf[WRITE][off]);
1616         if (off == LIMIT_LOW) {
1617                 if (tg->idletime_threshold_conf == ULONG_MAX)
1618                         strcpy(idle_time, " idle=max");
1619                 else
1620                         snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1621                                 tg->idletime_threshold_conf);
1622
1623                 if (tg->latency_target_conf == ULONG_MAX)
1624                         strcpy(latency_time, " latency=max");
1625                 else
1626                         snprintf(latency_time, sizeof(latency_time),
1627                                 " latency=%lu", tg->latency_target_conf);
1628         }
1629
1630         seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1631                    dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1632                    latency_time);
1633         return 0;
1634 }
1635
1636 static int tg_print_limit(struct seq_file *sf, void *v)
1637 {
1638         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1639                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1640         return 0;
1641 }
1642
1643 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1644                           char *buf, size_t nbytes, loff_t off)
1645 {
1646         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1647         struct blkg_conf_ctx ctx;
1648         struct throtl_grp *tg;
1649         u64 v[4];
1650         unsigned long idle_time;
1651         unsigned long latency_time;
1652         int ret;
1653         int index = of_cft(of)->private;
1654
1655         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1656         if (ret)
1657                 return ret;
1658
1659         tg = blkg_to_tg(ctx.blkg);
1660
1661         v[0] = tg->bps_conf[READ][index];
1662         v[1] = tg->bps_conf[WRITE][index];
1663         v[2] = tg->iops_conf[READ][index];
1664         v[3] = tg->iops_conf[WRITE][index];
1665
1666         idle_time = tg->idletime_threshold_conf;
1667         latency_time = tg->latency_target_conf;
1668         while (true) {
1669                 char tok[27];   /* wiops=18446744073709551616 */
1670                 char *p;
1671                 u64 val = U64_MAX;
1672                 int len;
1673
1674                 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1675                         break;
1676                 if (tok[0] == '\0')
1677                         break;
1678                 ctx.body += len;
1679
1680                 ret = -EINVAL;
1681                 p = tok;
1682                 strsep(&p, "=");
1683                 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1684                         goto out_finish;
1685
1686                 ret = -ERANGE;
1687                 if (!val)
1688                         goto out_finish;
1689
1690                 ret = -EINVAL;
1691                 if (!strcmp(tok, "rbps") && val > 1)
1692                         v[0] = val;
1693                 else if (!strcmp(tok, "wbps") && val > 1)
1694                         v[1] = val;
1695                 else if (!strcmp(tok, "riops") && val > 1)
1696                         v[2] = min_t(u64, val, UINT_MAX);
1697                 else if (!strcmp(tok, "wiops") && val > 1)
1698                         v[3] = min_t(u64, val, UINT_MAX);
1699                 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1700                         idle_time = val;
1701                 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1702                         latency_time = val;
1703                 else
1704                         goto out_finish;
1705         }
1706
1707         tg->bps_conf[READ][index] = v[0];
1708         tg->bps_conf[WRITE][index] = v[1];
1709         tg->iops_conf[READ][index] = v[2];
1710         tg->iops_conf[WRITE][index] = v[3];
1711
1712         if (index == LIMIT_MAX) {
1713                 tg->bps[READ][index] = v[0];
1714                 tg->bps[WRITE][index] = v[1];
1715                 tg->iops[READ][index] = v[2];
1716                 tg->iops[WRITE][index] = v[3];
1717         }
1718         tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1719                 tg->bps_conf[READ][LIMIT_MAX]);
1720         tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1721                 tg->bps_conf[WRITE][LIMIT_MAX]);
1722         tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1723                 tg->iops_conf[READ][LIMIT_MAX]);
1724         tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1725                 tg->iops_conf[WRITE][LIMIT_MAX]);
1726         tg->idletime_threshold_conf = idle_time;
1727         tg->latency_target_conf = latency_time;
1728
1729         /* force user to configure all settings for low limit  */
1730         if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1731               tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1732             tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1733             tg->latency_target_conf == DFL_LATENCY_TARGET) {
1734                 tg->bps[READ][LIMIT_LOW] = 0;
1735                 tg->bps[WRITE][LIMIT_LOW] = 0;
1736                 tg->iops[READ][LIMIT_LOW] = 0;
1737                 tg->iops[WRITE][LIMIT_LOW] = 0;
1738                 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1739                 tg->latency_target = DFL_LATENCY_TARGET;
1740         } else if (index == LIMIT_LOW) {
1741                 tg->idletime_threshold = tg->idletime_threshold_conf;
1742                 tg->latency_target = tg->latency_target_conf;
1743         }
1744
1745         blk_throtl_update_limit_valid(tg->td);
1746         if (tg->td->limit_valid[LIMIT_LOW]) {
1747                 if (index == LIMIT_LOW)
1748                         tg->td->limit_index = LIMIT_LOW;
1749         } else
1750                 tg->td->limit_index = LIMIT_MAX;
1751         tg_conf_updated(tg, index == LIMIT_LOW &&
1752                 tg->td->limit_valid[LIMIT_LOW]);
1753         ret = 0;
1754 out_finish:
1755         blkg_conf_finish(&ctx);
1756         return ret ?: nbytes;
1757 }
1758
1759 static struct cftype throtl_files[] = {
1760 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1761         {
1762                 .name = "low",
1763                 .flags = CFTYPE_NOT_ON_ROOT,
1764                 .seq_show = tg_print_limit,
1765                 .write = tg_set_limit,
1766                 .private = LIMIT_LOW,
1767         },
1768 #endif
1769         {
1770                 .name = "max",
1771                 .flags = CFTYPE_NOT_ON_ROOT,
1772                 .seq_show = tg_print_limit,
1773                 .write = tg_set_limit,
1774                 .private = LIMIT_MAX,
1775         },
1776         { }     /* terminate */
1777 };
1778
1779 static void throtl_shutdown_wq(struct request_queue *q)
1780 {
1781         struct throtl_data *td = q->td;
1782
1783         cancel_work_sync(&td->dispatch_work);
1784 }
1785
1786 static struct blkcg_policy blkcg_policy_throtl = {
1787         .dfl_cftypes            = throtl_files,
1788         .legacy_cftypes         = throtl_legacy_files,
1789
1790         .pd_alloc_fn            = throtl_pd_alloc,
1791         .pd_init_fn             = throtl_pd_init,
1792         .pd_online_fn           = throtl_pd_online,
1793         .pd_offline_fn          = throtl_pd_offline,
1794         .pd_free_fn             = throtl_pd_free,
1795 };
1796
1797 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1798 {
1799         unsigned long rtime = jiffies, wtime = jiffies;
1800
1801         if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1802                 rtime = tg->last_low_overflow_time[READ];
1803         if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1804                 wtime = tg->last_low_overflow_time[WRITE];
1805         return min(rtime, wtime);
1806 }
1807
1808 /* tg should not be an intermediate node */
1809 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1810 {
1811         struct throtl_service_queue *parent_sq;
1812         struct throtl_grp *parent = tg;
1813         unsigned long ret = __tg_last_low_overflow_time(tg);
1814
1815         while (true) {
1816                 parent_sq = parent->service_queue.parent_sq;
1817                 parent = sq_to_tg(parent_sq);
1818                 if (!parent)
1819                         break;
1820
1821                 /*
1822                  * The parent doesn't have low limit, it always reaches low
1823                  * limit. Its overflow time is useless for children
1824                  */
1825                 if (!parent->bps[READ][LIMIT_LOW] &&
1826                     !parent->iops[READ][LIMIT_LOW] &&
1827                     !parent->bps[WRITE][LIMIT_LOW] &&
1828                     !parent->iops[WRITE][LIMIT_LOW])
1829                         continue;
1830                 if (time_after(__tg_last_low_overflow_time(parent), ret))
1831                         ret = __tg_last_low_overflow_time(parent);
1832         }
1833         return ret;
1834 }
1835
1836 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1837 {
1838         /*
1839          * cgroup is idle if:
1840          * - single idle is too long, longer than a fixed value (in case user
1841          *   configure a too big threshold) or 4 times of idletime threshold
1842          * - average think time is more than threshold
1843          * - IO latency is largely below threshold
1844          */
1845         unsigned long time;
1846         bool ret;
1847
1848         time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1849         ret = tg->latency_target == DFL_LATENCY_TARGET ||
1850               tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1851               (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1852               tg->avg_idletime > tg->idletime_threshold ||
1853               (tg->latency_target && tg->bio_cnt &&
1854                 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1855         throtl_log(&tg->service_queue,
1856                 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1857                 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1858                 tg->bio_cnt, ret, tg->td->scale);
1859         return ret;
1860 }
1861
1862 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1863 {
1864         struct throtl_service_queue *sq = &tg->service_queue;
1865         bool read_limit, write_limit;
1866
1867         /*
1868          * if cgroup reaches low limit (if low limit is 0, the cgroup always
1869          * reaches), it's ok to upgrade to next limit
1870          */
1871         read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1872         write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1873         if (!read_limit && !write_limit)
1874                 return true;
1875         if (read_limit && sq->nr_queued[READ] &&
1876             (!write_limit || sq->nr_queued[WRITE]))
1877                 return true;
1878         if (write_limit && sq->nr_queued[WRITE] &&
1879             (!read_limit || sq->nr_queued[READ]))
1880                 return true;
1881
1882         if (time_after_eq(jiffies,
1883                 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1884             throtl_tg_is_idle(tg))
1885                 return true;
1886         return false;
1887 }
1888
1889 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1890 {
1891         while (true) {
1892                 if (throtl_tg_can_upgrade(tg))
1893                         return true;
1894                 tg = sq_to_tg(tg->service_queue.parent_sq);
1895                 if (!tg || !tg_to_blkg(tg)->parent)
1896                         return false;
1897         }
1898         return false;
1899 }
1900
1901 static bool throtl_can_upgrade(struct throtl_data *td,
1902         struct throtl_grp *this_tg)
1903 {
1904         struct cgroup_subsys_state *pos_css;
1905         struct blkcg_gq *blkg;
1906
1907         if (td->limit_index != LIMIT_LOW)
1908                 return false;
1909
1910         if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1911                 return false;
1912
1913         rcu_read_lock();
1914         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1915                 struct throtl_grp *tg = blkg_to_tg(blkg);
1916
1917                 if (tg == this_tg)
1918                         continue;
1919                 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1920                         continue;
1921                 if (!throtl_hierarchy_can_upgrade(tg)) {
1922                         rcu_read_unlock();
1923                         return false;
1924                 }
1925         }
1926         rcu_read_unlock();
1927         return true;
1928 }
1929
1930 static void throtl_upgrade_check(struct throtl_grp *tg)
1931 {
1932         unsigned long now = jiffies;
1933
1934         if (tg->td->limit_index != LIMIT_LOW)
1935                 return;
1936
1937         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1938                 return;
1939
1940         tg->last_check_time = now;
1941
1942         if (!time_after_eq(now,
1943              __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1944                 return;
1945
1946         if (throtl_can_upgrade(tg->td, NULL))
1947                 throtl_upgrade_state(tg->td);
1948 }
1949
1950 static void throtl_upgrade_state(struct throtl_data *td)
1951 {
1952         struct cgroup_subsys_state *pos_css;
1953         struct blkcg_gq *blkg;
1954
1955         throtl_log(&td->service_queue, "upgrade to max");
1956         td->limit_index = LIMIT_MAX;
1957         td->low_upgrade_time = jiffies;
1958         td->scale = 0;
1959         rcu_read_lock();
1960         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1961                 struct throtl_grp *tg = blkg_to_tg(blkg);
1962                 struct throtl_service_queue *sq = &tg->service_queue;
1963
1964                 tg->disptime = jiffies - 1;
1965                 throtl_select_dispatch(sq);
1966                 throtl_schedule_next_dispatch(sq, true);
1967         }
1968         rcu_read_unlock();
1969         throtl_select_dispatch(&td->service_queue);
1970         throtl_schedule_next_dispatch(&td->service_queue, true);
1971         queue_work(kthrotld_workqueue, &td->dispatch_work);
1972 }
1973
1974 static void throtl_downgrade_state(struct throtl_data *td)
1975 {
1976         td->scale /= 2;
1977
1978         throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1979         if (td->scale) {
1980                 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1981                 return;
1982         }
1983
1984         td->limit_index = LIMIT_LOW;
1985         td->low_downgrade_time = jiffies;
1986 }
1987
1988 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1989 {
1990         struct throtl_data *td = tg->td;
1991         unsigned long now = jiffies;
1992
1993         /*
1994          * If cgroup is below low limit, consider downgrade and throttle other
1995          * cgroups
1996          */
1997         if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1998             time_after_eq(now, tg_last_low_overflow_time(tg) +
1999                                         td->throtl_slice) &&
2000             (!throtl_tg_is_idle(tg) ||
2001              !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
2002                 return true;
2003         return false;
2004 }
2005
2006 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
2007 {
2008         while (true) {
2009                 if (!throtl_tg_can_downgrade(tg))
2010                         return false;
2011                 tg = sq_to_tg(tg->service_queue.parent_sq);
2012                 if (!tg || !tg_to_blkg(tg)->parent)
2013                         break;
2014         }
2015         return true;
2016 }
2017
2018 static void throtl_downgrade_check(struct throtl_grp *tg)
2019 {
2020         uint64_t bps;
2021         unsigned int iops;
2022         unsigned long elapsed_time;
2023         unsigned long now = jiffies;
2024
2025         if (tg->td->limit_index != LIMIT_MAX ||
2026             !tg->td->limit_valid[LIMIT_LOW])
2027                 return;
2028         if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
2029                 return;
2030         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
2031                 return;
2032
2033         elapsed_time = now - tg->last_check_time;
2034         tg->last_check_time = now;
2035
2036         if (time_before(now, tg_last_low_overflow_time(tg) +
2037                         tg->td->throtl_slice))
2038                 return;
2039
2040         if (tg->bps[READ][LIMIT_LOW]) {
2041                 bps = tg->last_bytes_disp[READ] * HZ;
2042                 do_div(bps, elapsed_time);
2043                 if (bps >= tg->bps[READ][LIMIT_LOW])
2044                         tg->last_low_overflow_time[READ] = now;
2045         }
2046
2047         if (tg->bps[WRITE][LIMIT_LOW]) {
2048                 bps = tg->last_bytes_disp[WRITE] * HZ;
2049                 do_div(bps, elapsed_time);
2050                 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2051                         tg->last_low_overflow_time[WRITE] = now;
2052         }
2053
2054         if (tg->iops[READ][LIMIT_LOW]) {
2055                 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2056                 if (iops >= tg->iops[READ][LIMIT_LOW])
2057                         tg->last_low_overflow_time[READ] = now;
2058         }
2059
2060         if (tg->iops[WRITE][LIMIT_LOW]) {
2061                 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2062                 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2063                         tg->last_low_overflow_time[WRITE] = now;
2064         }
2065
2066         /*
2067          * If cgroup is below low limit, consider downgrade and throttle other
2068          * cgroups
2069          */
2070         if (throtl_hierarchy_can_downgrade(tg))
2071                 throtl_downgrade_state(tg->td);
2072
2073         tg->last_bytes_disp[READ] = 0;
2074         tg->last_bytes_disp[WRITE] = 0;
2075         tg->last_io_disp[READ] = 0;
2076         tg->last_io_disp[WRITE] = 0;
2077 }
2078
2079 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2080 {
2081         unsigned long now;
2082         unsigned long last_finish_time = tg->last_finish_time;
2083
2084         if (last_finish_time == 0)
2085                 return;
2086
2087         now = ktime_get_ns() >> 10;
2088         if (now <= last_finish_time ||
2089             last_finish_time == tg->checked_last_finish_time)
2090                 return;
2091
2092         tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2093         tg->checked_last_finish_time = last_finish_time;
2094 }
2095
2096 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2097 static void throtl_update_latency_buckets(struct throtl_data *td)
2098 {
2099         struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2100         int i, cpu, rw;
2101         unsigned long last_latency[2] = { 0 };
2102         unsigned long latency[2];
2103
2104         if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
2105                 return;
2106         if (time_before(jiffies, td->last_calculate_time + HZ))
2107                 return;
2108         td->last_calculate_time = jiffies;
2109
2110         memset(avg_latency, 0, sizeof(avg_latency));
2111         for (rw = READ; rw <= WRITE; rw++) {
2112                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2113                         struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2114
2115                         for_each_possible_cpu(cpu) {
2116                                 struct latency_bucket *bucket;
2117
2118                                 /* this isn't race free, but ok in practice */
2119                                 bucket = per_cpu_ptr(td->latency_buckets[rw],
2120                                         cpu);
2121                                 tmp->total_latency += bucket[i].total_latency;
2122                                 tmp->samples += bucket[i].samples;
2123                                 bucket[i].total_latency = 0;
2124                                 bucket[i].samples = 0;
2125                         }
2126
2127                         if (tmp->samples >= 32) {
2128                                 int samples = tmp->samples;
2129
2130                                 latency[rw] = tmp->total_latency;
2131
2132                                 tmp->total_latency = 0;
2133                                 tmp->samples = 0;
2134                                 latency[rw] /= samples;
2135                                 if (latency[rw] == 0)
2136                                         continue;
2137                                 avg_latency[rw][i].latency = latency[rw];
2138                         }
2139                 }
2140         }
2141
2142         for (rw = READ; rw <= WRITE; rw++) {
2143                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2144                         if (!avg_latency[rw][i].latency) {
2145                                 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2146                                         td->avg_buckets[rw][i].latency =
2147                                                 last_latency[rw];
2148                                 continue;
2149                         }
2150
2151                         if (!td->avg_buckets[rw][i].valid)
2152                                 latency[rw] = avg_latency[rw][i].latency;
2153                         else
2154                                 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2155                                         avg_latency[rw][i].latency) >> 3;
2156
2157                         td->avg_buckets[rw][i].latency = max(latency[rw],
2158                                 last_latency[rw]);
2159                         td->avg_buckets[rw][i].valid = true;
2160                         last_latency[rw] = td->avg_buckets[rw][i].latency;
2161                 }
2162         }
2163
2164         for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2165                 throtl_log(&td->service_queue,
2166                         "Latency bucket %d: read latency=%ld, read valid=%d, "
2167                         "write latency=%ld, write valid=%d", i,
2168                         td->avg_buckets[READ][i].latency,
2169                         td->avg_buckets[READ][i].valid,
2170                         td->avg_buckets[WRITE][i].latency,
2171                         td->avg_buckets[WRITE][i].valid);
2172 }
2173 #else
2174 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2175 {
2176 }
2177 #endif
2178
2179 bool blk_throtl_bio(struct bio *bio)
2180 {
2181         struct request_queue *q = bio->bi_bdev->bd_disk->queue;
2182         struct blkcg_gq *blkg = bio->bi_blkg;
2183         struct throtl_qnode *qn = NULL;
2184         struct throtl_grp *tg = blkg_to_tg(blkg);
2185         struct throtl_service_queue *sq;
2186         bool rw = bio_data_dir(bio);
2187         bool throttled = false;
2188         struct throtl_data *td = tg->td;
2189
2190         rcu_read_lock();
2191
2192         /* see throtl_charge_bio() */
2193         if (bio_flagged(bio, BIO_THROTTLED))
2194                 goto out;
2195
2196         if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
2197                 blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
2198                                 bio->bi_iter.bi_size);
2199                 blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
2200         }
2201
2202         if (!tg->has_rules[rw])
2203                 goto out;
2204
2205         spin_lock_irq(&q->queue_lock);
2206
2207         throtl_update_latency_buckets(td);
2208
2209         blk_throtl_update_idletime(tg);
2210
2211         sq = &tg->service_queue;
2212
2213 again:
2214         while (true) {
2215                 if (tg->last_low_overflow_time[rw] == 0)
2216                         tg->last_low_overflow_time[rw] = jiffies;
2217                 throtl_downgrade_check(tg);
2218                 throtl_upgrade_check(tg);
2219                 /* throtl is FIFO - if bios are already queued, should queue */
2220                 if (sq->nr_queued[rw])
2221                         break;
2222
2223                 /* if above limits, break to queue */
2224                 if (!tg_may_dispatch(tg, bio, NULL)) {
2225                         tg->last_low_overflow_time[rw] = jiffies;
2226                         if (throtl_can_upgrade(td, tg)) {
2227                                 throtl_upgrade_state(td);
2228                                 goto again;
2229                         }
2230                         break;
2231                 }
2232
2233                 /* within limits, let's charge and dispatch directly */
2234                 throtl_charge_bio(tg, bio);
2235
2236                 /*
2237                  * We need to trim slice even when bios are not being queued
2238                  * otherwise it might happen that a bio is not queued for
2239                  * a long time and slice keeps on extending and trim is not
2240                  * called for a long time. Now if limits are reduced suddenly
2241                  * we take into account all the IO dispatched so far at new
2242                  * low rate and * newly queued IO gets a really long dispatch
2243                  * time.
2244                  *
2245                  * So keep on trimming slice even if bio is not queued.
2246                  */
2247                 throtl_trim_slice(tg, rw);
2248
2249                 /*
2250                  * @bio passed through this layer without being throttled.
2251                  * Climb up the ladder.  If we're already at the top, it
2252                  * can be executed directly.
2253                  */
2254                 qn = &tg->qnode_on_parent[rw];
2255                 sq = sq->parent_sq;
2256                 tg = sq_to_tg(sq);
2257                 if (!tg)
2258                         goto out_unlock;
2259         }
2260
2261         /* out-of-limit, queue to @tg */
2262         throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2263                    rw == READ ? 'R' : 'W',
2264                    tg->bytes_disp[rw], bio->bi_iter.bi_size,
2265                    tg_bps_limit(tg, rw),
2266                    tg->io_disp[rw], tg_iops_limit(tg, rw),
2267                    sq->nr_queued[READ], sq->nr_queued[WRITE]);
2268
2269         tg->last_low_overflow_time[rw] = jiffies;
2270
2271         td->nr_queued[rw]++;
2272         throtl_add_bio_tg(bio, qn, tg);
2273         throttled = true;
2274
2275         /*
2276          * Update @tg's dispatch time and force schedule dispatch if @tg
2277          * was empty before @bio.  The forced scheduling isn't likely to
2278          * cause undue delay as @bio is likely to be dispatched directly if
2279          * its @tg's disptime is not in the future.
2280          */
2281         if (tg->flags & THROTL_TG_WAS_EMPTY) {
2282                 tg_update_disptime(tg);
2283                 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2284         }
2285
2286 out_unlock:
2287         spin_unlock_irq(&q->queue_lock);
2288 out:
2289         bio_set_flag(bio, BIO_THROTTLED);
2290
2291 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2292         if (throttled || !td->track_bio_latency)
2293                 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2294 #endif
2295         rcu_read_unlock();
2296         return throttled;
2297 }
2298
2299 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2300 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2301         int op, unsigned long time)
2302 {
2303         struct latency_bucket *latency;
2304         int index;
2305
2306         if (!td || td->limit_index != LIMIT_LOW ||
2307             !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2308             !blk_queue_nonrot(td->queue))
2309                 return;
2310
2311         index = request_bucket_index(size);
2312
2313         latency = get_cpu_ptr(td->latency_buckets[op]);
2314         latency[index].total_latency += time;
2315         latency[index].samples++;
2316         put_cpu_ptr(td->latency_buckets[op]);
2317 }
2318
2319 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2320 {
2321         struct request_queue *q = rq->q;
2322         struct throtl_data *td = q->td;
2323
2324         throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2325                              time_ns >> 10);
2326 }
2327
2328 void blk_throtl_bio_endio(struct bio *bio)
2329 {
2330         struct blkcg_gq *blkg;
2331         struct throtl_grp *tg;
2332         u64 finish_time_ns;
2333         unsigned long finish_time;
2334         unsigned long start_time;
2335         unsigned long lat;
2336         int rw = bio_data_dir(bio);
2337
2338         blkg = bio->bi_blkg;
2339         if (!blkg)
2340                 return;
2341         tg = blkg_to_tg(blkg);
2342         if (!tg->td->limit_valid[LIMIT_LOW])
2343                 return;
2344
2345         finish_time_ns = ktime_get_ns();
2346         tg->last_finish_time = finish_time_ns >> 10;
2347
2348         start_time = bio_issue_time(&bio->bi_issue) >> 10;
2349         finish_time = __bio_issue_time(finish_time_ns) >> 10;
2350         if (!start_time || finish_time <= start_time)
2351                 return;
2352
2353         lat = finish_time - start_time;
2354         /* this is only for bio based driver */
2355         if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2356                 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2357                                      bio_op(bio), lat);
2358
2359         if (tg->latency_target && lat >= tg->td->filtered_latency) {
2360                 int bucket;
2361                 unsigned int threshold;
2362
2363                 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2364                 threshold = tg->td->avg_buckets[rw][bucket].latency +
2365                         tg->latency_target;
2366                 if (lat > threshold)
2367                         tg->bad_bio_cnt++;
2368                 /*
2369                  * Not race free, could get wrong count, which means cgroups
2370                  * will be throttled
2371                  */
2372                 tg->bio_cnt++;
2373         }
2374
2375         if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2376                 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2377                 tg->bio_cnt /= 2;
2378                 tg->bad_bio_cnt /= 2;
2379         }
2380 }
2381 #endif
2382
2383 int blk_throtl_init(struct request_queue *q)
2384 {
2385         struct throtl_data *td;
2386         int ret;
2387
2388         td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2389         if (!td)
2390                 return -ENOMEM;
2391         td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2392                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2393         if (!td->latency_buckets[READ]) {
2394                 kfree(td);
2395                 return -ENOMEM;
2396         }
2397         td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2398                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2399         if (!td->latency_buckets[WRITE]) {
2400                 free_percpu(td->latency_buckets[READ]);
2401                 kfree(td);
2402                 return -ENOMEM;
2403         }
2404
2405         INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2406         throtl_service_queue_init(&td->service_queue);
2407
2408         q->td = td;
2409         td->queue = q;
2410
2411         td->limit_valid[LIMIT_MAX] = true;
2412         td->limit_index = LIMIT_MAX;
2413         td->low_upgrade_time = jiffies;
2414         td->low_downgrade_time = jiffies;
2415
2416         /* activate policy */
2417         ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2418         if (ret) {
2419                 free_percpu(td->latency_buckets[READ]);
2420                 free_percpu(td->latency_buckets[WRITE]);
2421                 kfree(td);
2422         }
2423         return ret;
2424 }
2425
2426 void blk_throtl_exit(struct request_queue *q)
2427 {
2428         BUG_ON(!q->td);
2429         throtl_shutdown_wq(q);
2430         blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2431         free_percpu(q->td->latency_buckets[READ]);
2432         free_percpu(q->td->latency_buckets[WRITE]);
2433         kfree(q->td);
2434 }
2435
2436 void blk_throtl_register_queue(struct request_queue *q)
2437 {
2438         struct throtl_data *td;
2439         int i;
2440
2441         td = q->td;
2442         BUG_ON(!td);
2443
2444         if (blk_queue_nonrot(q)) {
2445                 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2446                 td->filtered_latency = LATENCY_FILTERED_SSD;
2447         } else {
2448                 td->throtl_slice = DFL_THROTL_SLICE_HD;
2449                 td->filtered_latency = LATENCY_FILTERED_HD;
2450                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2451                         td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2452                         td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2453                 }
2454         }
2455 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2456         /* if no low limit, use previous default */
2457         td->throtl_slice = DFL_THROTL_SLICE_HD;
2458 #endif
2459
2460         td->track_bio_latency = !queue_is_mq(q);
2461         if (!td->track_bio_latency)
2462                 blk_stat_enable_accounting(q);
2463 }
2464
2465 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2466 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2467 {
2468         if (!q->td)
2469                 return -EINVAL;
2470         return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2471 }
2472
2473 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2474         const char *page, size_t count)
2475 {
2476         unsigned long v;
2477         unsigned long t;
2478
2479         if (!q->td)
2480                 return -EINVAL;
2481         if (kstrtoul(page, 10, &v))
2482                 return -EINVAL;
2483         t = msecs_to_jiffies(v);
2484         if (t == 0 || t > MAX_THROTL_SLICE)
2485                 return -EINVAL;
2486         q->td->throtl_slice = t;
2487         return count;
2488 }
2489 #endif
2490
2491 static int __init throtl_init(void)
2492 {
2493         kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2494         if (!kthrotld_workqueue)
2495                 panic("Failed to create kthrotld\n");
2496
2497         return blkcg_policy_register(&blkcg_policy_throtl);
2498 }
2499
2500 module_init(throtl_init);