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 "blk.h"
14 #include "blk-cgroup-rwstat.h"
15 #include "blk-stat.h"
16 #include "blk-throttle.h"
17
18 /* Max dispatch from a group in 1 round */
19 #define THROTL_GRP_QUANTUM 8
20
21 /* Total max dispatch from all groups in one round */
22 #define THROTL_QUANTUM 32
23
24 /* Throttling is performed over a slice and after that slice is renewed */
25 #define DFL_THROTL_SLICE_HD (HZ / 10)
26 #define DFL_THROTL_SLICE_SSD (HZ / 50)
27 #define MAX_THROTL_SLICE (HZ)
28
29 /* A workqueue to queue throttle related work */
30 static struct workqueue_struct *kthrotld_workqueue;
31
32 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
33
34 /* We measure latency for request size from <= 4k to >= 1M */
35 #define LATENCY_BUCKET_SIZE 9
36
37 struct latency_bucket {
38 unsigned long total_latency; /* ns / 1024 */
39 int samples;
40 };
41
42 struct avg_latency_bucket {
43 unsigned long latency; /* ns / 1024 */
44 bool valid;
45 };
46
47 struct throtl_data
48 {
49 /* service tree for active throtl groups */
50 struct throtl_service_queue service_queue;
51
52 struct request_queue *queue;
53
54 /* Total Number of queued bios on READ and WRITE lists */
55 unsigned int nr_queued[2];
56
57 unsigned int throtl_slice;
58
59 /* Work for dispatching throttled bios */
60 struct work_struct dispatch_work;
61
62 bool track_bio_latency;
63 };
64
65 static void throtl_pending_timer_fn(struct timer_list *t);
66
tg_to_blkg(struct throtl_grp * tg)67 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
68 {
69 return pd_to_blkg(&tg->pd);
70 }
71
72 /**
73 * sq_to_tg - return the throl_grp the specified service queue belongs to
74 * @sq: the throtl_service_queue of interest
75 *
76 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
77 * embedded in throtl_data, %NULL is returned.
78 */
sq_to_tg(struct throtl_service_queue * sq)79 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
80 {
81 if (sq && sq->parent_sq)
82 return container_of(sq, struct throtl_grp, service_queue);
83 else
84 return NULL;
85 }
86
87 /**
88 * sq_to_td - return throtl_data the specified service queue belongs to
89 * @sq: the throtl_service_queue of interest
90 *
91 * A service_queue can be embedded in either a throtl_grp or throtl_data.
92 * Determine the associated throtl_data accordingly and return it.
93 */
sq_to_td(struct throtl_service_queue * sq)94 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
95 {
96 struct throtl_grp *tg = sq_to_tg(sq);
97
98 if (tg)
99 return tg->td;
100 else
101 return container_of(sq, struct throtl_data, service_queue);
102 }
103
tg_bps_limit(struct throtl_grp * tg,int rw)104 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
105 {
106 struct blkcg_gq *blkg = tg_to_blkg(tg);
107
108 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
109 return U64_MAX;
110
111 return tg->bps[rw];
112 }
113
tg_iops_limit(struct throtl_grp * tg,int rw)114 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
115 {
116 struct blkcg_gq *blkg = tg_to_blkg(tg);
117
118 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
119 return UINT_MAX;
120
121 return tg->iops[rw];
122 }
123
124 #define request_bucket_index(sectors) \
125 clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
126
127 /**
128 * throtl_log - log debug message via blktrace
129 * @sq: the service_queue being reported
130 * @fmt: printf format string
131 * @args: printf args
132 *
133 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
134 * throtl_grp; otherwise, just "throtl".
135 */
136 #define throtl_log(sq, fmt, args...) do { \
137 struct throtl_grp *__tg = sq_to_tg((sq)); \
138 struct throtl_data *__td = sq_to_td((sq)); \
139 \
140 (void)__td; \
141 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
142 break; \
143 if ((__tg)) { \
144 blk_add_cgroup_trace_msg(__td->queue, \
145 &tg_to_blkg(__tg)->blkcg->css, "throtl " fmt, ##args);\
146 } else { \
147 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
148 } \
149 } while (0)
150
throtl_bio_data_size(struct bio * bio)151 static inline unsigned int throtl_bio_data_size(struct bio *bio)
152 {
153 /* assume it's one sector */
154 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
155 return 512;
156 return bio->bi_iter.bi_size;
157 }
158
throtl_qnode_init(struct throtl_qnode * qn,struct throtl_grp * tg)159 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
160 {
161 INIT_LIST_HEAD(&qn->node);
162 bio_list_init(&qn->bios);
163 qn->tg = tg;
164 }
165
166 /**
167 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
168 * @bio: bio being added
169 * @qn: qnode to add bio to
170 * @queued: the service_queue->queued[] list @qn belongs to
171 *
172 * Add @bio to @qn and put @qn on @queued if it's not already on.
173 * @qn->tg's reference count is bumped when @qn is activated. See the
174 * comment on top of throtl_qnode definition for details.
175 */
throtl_qnode_add_bio(struct bio * bio,struct throtl_qnode * qn,struct list_head * queued)176 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
177 struct list_head *queued)
178 {
179 bio_list_add(&qn->bios, bio);
180 if (list_empty(&qn->node)) {
181 list_add_tail(&qn->node, queued);
182 blkg_get(tg_to_blkg(qn->tg));
183 }
184 }
185
186 /**
187 * throtl_peek_queued - peek the first bio on a qnode list
188 * @queued: the qnode list to peek
189 */
throtl_peek_queued(struct list_head * queued)190 static struct bio *throtl_peek_queued(struct list_head *queued)
191 {
192 struct throtl_qnode *qn;
193 struct bio *bio;
194
195 if (list_empty(queued))
196 return NULL;
197
198 qn = list_first_entry(queued, struct throtl_qnode, node);
199 bio = bio_list_peek(&qn->bios);
200 WARN_ON_ONCE(!bio);
201 return bio;
202 }
203
204 /**
205 * throtl_pop_queued - pop the first bio form a qnode list
206 * @queued: the qnode list to pop a bio from
207 * @tg_to_put: optional out argument for throtl_grp to put
208 *
209 * Pop the first bio from the qnode list @queued. After popping, the first
210 * qnode is removed from @queued if empty or moved to the end of @queued so
211 * that the popping order is round-robin.
212 *
213 * When the first qnode is removed, its associated throtl_grp should be put
214 * too. If @tg_to_put is NULL, this function automatically puts it;
215 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
216 * responsible for putting it.
217 */
throtl_pop_queued(struct list_head * queued,struct throtl_grp ** tg_to_put)218 static struct bio *throtl_pop_queued(struct list_head *queued,
219 struct throtl_grp **tg_to_put)
220 {
221 struct throtl_qnode *qn;
222 struct bio *bio;
223
224 if (list_empty(queued))
225 return NULL;
226
227 qn = list_first_entry(queued, struct throtl_qnode, node);
228 bio = bio_list_pop(&qn->bios);
229 WARN_ON_ONCE(!bio);
230
231 if (bio_list_empty(&qn->bios)) {
232 list_del_init(&qn->node);
233 if (tg_to_put)
234 *tg_to_put = qn->tg;
235 else
236 blkg_put(tg_to_blkg(qn->tg));
237 } else {
238 list_move_tail(&qn->node, queued);
239 }
240
241 return bio;
242 }
243
244 /* init a service_queue, assumes the caller zeroed it */
throtl_service_queue_init(struct throtl_service_queue * sq)245 static void throtl_service_queue_init(struct throtl_service_queue *sq)
246 {
247 INIT_LIST_HEAD(&sq->queued[READ]);
248 INIT_LIST_HEAD(&sq->queued[WRITE]);
249 sq->pending_tree = RB_ROOT_CACHED;
250 timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
251 }
252
throtl_pd_alloc(struct gendisk * disk,struct blkcg * blkcg,gfp_t gfp)253 static struct blkg_policy_data *throtl_pd_alloc(struct gendisk *disk,
254 struct blkcg *blkcg, gfp_t gfp)
255 {
256 struct throtl_grp *tg;
257 int rw;
258
259 tg = kzalloc_node(sizeof(*tg), gfp, disk->node_id);
260 if (!tg)
261 return NULL;
262
263 if (blkg_rwstat_init(&tg->stat_bytes, gfp))
264 goto err_free_tg;
265
266 if (blkg_rwstat_init(&tg->stat_ios, gfp))
267 goto err_exit_stat_bytes;
268
269 throtl_service_queue_init(&tg->service_queue);
270
271 for (rw = READ; rw <= WRITE; rw++) {
272 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
273 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
274 }
275
276 RB_CLEAR_NODE(&tg->rb_node);
277 tg->bps[READ] = U64_MAX;
278 tg->bps[WRITE] = U64_MAX;
279 tg->iops[READ] = UINT_MAX;
280 tg->iops[WRITE] = UINT_MAX;
281
282 return &tg->pd;
283
284 err_exit_stat_bytes:
285 blkg_rwstat_exit(&tg->stat_bytes);
286 err_free_tg:
287 kfree(tg);
288 return NULL;
289 }
290
throtl_pd_init(struct blkg_policy_data * pd)291 static void throtl_pd_init(struct blkg_policy_data *pd)
292 {
293 struct throtl_grp *tg = pd_to_tg(pd);
294 struct blkcg_gq *blkg = tg_to_blkg(tg);
295 struct throtl_data *td = blkg->q->td;
296 struct throtl_service_queue *sq = &tg->service_queue;
297
298 /*
299 * If on the default hierarchy, we switch to properly hierarchical
300 * behavior where limits on a given throtl_grp are applied to the
301 * whole subtree rather than just the group itself. e.g. If 16M
302 * read_bps limit is set on a parent group, summary bps of
303 * parent group and its subtree groups can't exceed 16M for the
304 * device.
305 *
306 * If not on the default hierarchy, the broken flat hierarchy
307 * behavior is retained where all throtl_grps are treated as if
308 * they're all separate root groups right below throtl_data.
309 * Limits of a group don't interact with limits of other groups
310 * regardless of the position of the group in the hierarchy.
311 */
312 sq->parent_sq = &td->service_queue;
313 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
314 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
315 tg->td = td;
316 }
317
318 /*
319 * Set has_rules[] if @tg or any of its parents have limits configured.
320 * This doesn't require walking up to the top of the hierarchy as the
321 * parent's has_rules[] is guaranteed to be correct.
322 */
tg_update_has_rules(struct throtl_grp * tg)323 static void tg_update_has_rules(struct throtl_grp *tg)
324 {
325 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
326 int rw;
327
328 for (rw = READ; rw <= WRITE; rw++) {
329 tg->has_rules_iops[rw] =
330 (parent_tg && parent_tg->has_rules_iops[rw]) ||
331 tg_iops_limit(tg, rw) != UINT_MAX;
332 tg->has_rules_bps[rw] =
333 (parent_tg && parent_tg->has_rules_bps[rw]) ||
334 tg_bps_limit(tg, rw) != U64_MAX;
335 }
336 }
337
throtl_pd_online(struct blkg_policy_data * pd)338 static void throtl_pd_online(struct blkg_policy_data *pd)
339 {
340 struct throtl_grp *tg = pd_to_tg(pd);
341 /*
342 * We don't want new groups to escape the limits of its ancestors.
343 * Update has_rules[] after a new group is brought online.
344 */
345 tg_update_has_rules(tg);
346 }
347
throtl_pd_free(struct blkg_policy_data * pd)348 static void throtl_pd_free(struct blkg_policy_data *pd)
349 {
350 struct throtl_grp *tg = pd_to_tg(pd);
351
352 del_timer_sync(&tg->service_queue.pending_timer);
353 blkg_rwstat_exit(&tg->stat_bytes);
354 blkg_rwstat_exit(&tg->stat_ios);
355 kfree(tg);
356 }
357
358 static struct throtl_grp *
throtl_rb_first(struct throtl_service_queue * parent_sq)359 throtl_rb_first(struct throtl_service_queue *parent_sq)
360 {
361 struct rb_node *n;
362
363 n = rb_first_cached(&parent_sq->pending_tree);
364 WARN_ON_ONCE(!n);
365 if (!n)
366 return NULL;
367 return rb_entry_tg(n);
368 }
369
throtl_rb_erase(struct rb_node * n,struct throtl_service_queue * parent_sq)370 static void throtl_rb_erase(struct rb_node *n,
371 struct throtl_service_queue *parent_sq)
372 {
373 rb_erase_cached(n, &parent_sq->pending_tree);
374 RB_CLEAR_NODE(n);
375 }
376
update_min_dispatch_time(struct throtl_service_queue * parent_sq)377 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
378 {
379 struct throtl_grp *tg;
380
381 tg = throtl_rb_first(parent_sq);
382 if (!tg)
383 return;
384
385 parent_sq->first_pending_disptime = tg->disptime;
386 }
387
tg_service_queue_add(struct throtl_grp * tg)388 static void tg_service_queue_add(struct throtl_grp *tg)
389 {
390 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
391 struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
392 struct rb_node *parent = NULL;
393 struct throtl_grp *__tg;
394 unsigned long key = tg->disptime;
395 bool leftmost = true;
396
397 while (*node != NULL) {
398 parent = *node;
399 __tg = rb_entry_tg(parent);
400
401 if (time_before(key, __tg->disptime))
402 node = &parent->rb_left;
403 else {
404 node = &parent->rb_right;
405 leftmost = false;
406 }
407 }
408
409 rb_link_node(&tg->rb_node, parent, node);
410 rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
411 leftmost);
412 }
413
throtl_enqueue_tg(struct throtl_grp * tg)414 static void throtl_enqueue_tg(struct throtl_grp *tg)
415 {
416 if (!(tg->flags & THROTL_TG_PENDING)) {
417 tg_service_queue_add(tg);
418 tg->flags |= THROTL_TG_PENDING;
419 tg->service_queue.parent_sq->nr_pending++;
420 }
421 }
422
throtl_dequeue_tg(struct throtl_grp * tg)423 static void throtl_dequeue_tg(struct throtl_grp *tg)
424 {
425 if (tg->flags & THROTL_TG_PENDING) {
426 struct throtl_service_queue *parent_sq =
427 tg->service_queue.parent_sq;
428
429 throtl_rb_erase(&tg->rb_node, parent_sq);
430 --parent_sq->nr_pending;
431 tg->flags &= ~THROTL_TG_PENDING;
432 }
433 }
434
435 /* Call with queue lock held */
throtl_schedule_pending_timer(struct throtl_service_queue * sq,unsigned long expires)436 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
437 unsigned long expires)
438 {
439 unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
440
441 /*
442 * Since we are adjusting the throttle limit dynamically, the sleep
443 * time calculated according to previous limit might be invalid. It's
444 * possible the cgroup sleep time is very long and no other cgroups
445 * have IO running so notify the limit changes. Make sure the cgroup
446 * doesn't sleep too long to avoid the missed notification.
447 */
448 if (time_after(expires, max_expire))
449 expires = max_expire;
450 mod_timer(&sq->pending_timer, expires);
451 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
452 expires - jiffies, jiffies);
453 }
454
455 /**
456 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
457 * @sq: the service_queue to schedule dispatch for
458 * @force: force scheduling
459 *
460 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
461 * dispatch time of the first pending child. Returns %true if either timer
462 * is armed or there's no pending child left. %false if the current
463 * dispatch window is still open and the caller should continue
464 * dispatching.
465 *
466 * If @force is %true, the dispatch timer is always scheduled and this
467 * function is guaranteed to return %true. This is to be used when the
468 * caller can't dispatch itself and needs to invoke pending_timer
469 * unconditionally. Note that forced scheduling is likely to induce short
470 * delay before dispatch starts even if @sq->first_pending_disptime is not
471 * in the future and thus shouldn't be used in hot paths.
472 */
throtl_schedule_next_dispatch(struct throtl_service_queue * sq,bool force)473 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
474 bool force)
475 {
476 /* any pending children left? */
477 if (!sq->nr_pending)
478 return true;
479
480 update_min_dispatch_time(sq);
481
482 /* is the next dispatch time in the future? */
483 if (force || time_after(sq->first_pending_disptime, jiffies)) {
484 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
485 return true;
486 }
487
488 /* tell the caller to continue dispatching */
489 return false;
490 }
491
throtl_start_new_slice_with_credit(struct throtl_grp * tg,bool rw,unsigned long start)492 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
493 bool rw, unsigned long start)
494 {
495 tg->bytes_disp[rw] = 0;
496 tg->io_disp[rw] = 0;
497 tg->carryover_bytes[rw] = 0;
498 tg->carryover_ios[rw] = 0;
499
500 /*
501 * Previous slice has expired. We must have trimmed it after last
502 * bio dispatch. That means since start of last slice, we never used
503 * that bandwidth. Do try to make use of that bandwidth while giving
504 * credit.
505 */
506 if (time_after(start, tg->slice_start[rw]))
507 tg->slice_start[rw] = start;
508
509 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
510 throtl_log(&tg->service_queue,
511 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
512 rw == READ ? 'R' : 'W', tg->slice_start[rw],
513 tg->slice_end[rw], jiffies);
514 }
515
throtl_start_new_slice(struct throtl_grp * tg,bool rw,bool clear_carryover)516 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw,
517 bool clear_carryover)
518 {
519 tg->bytes_disp[rw] = 0;
520 tg->io_disp[rw] = 0;
521 tg->slice_start[rw] = jiffies;
522 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
523 if (clear_carryover) {
524 tg->carryover_bytes[rw] = 0;
525 tg->carryover_ios[rw] = 0;
526 }
527
528 throtl_log(&tg->service_queue,
529 "[%c] new slice start=%lu end=%lu jiffies=%lu",
530 rw == READ ? 'R' : 'W', tg->slice_start[rw],
531 tg->slice_end[rw], jiffies);
532 }
533
throtl_set_slice_end(struct throtl_grp * tg,bool rw,unsigned long jiffy_end)534 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
535 unsigned long jiffy_end)
536 {
537 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
538 }
539
throtl_extend_slice(struct throtl_grp * tg,bool rw,unsigned long jiffy_end)540 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
541 unsigned long jiffy_end)
542 {
543 throtl_set_slice_end(tg, rw, jiffy_end);
544 throtl_log(&tg->service_queue,
545 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
546 rw == READ ? 'R' : 'W', tg->slice_start[rw],
547 tg->slice_end[rw], jiffies);
548 }
549
550 /* Determine if previously allocated or extended slice is complete or not */
throtl_slice_used(struct throtl_grp * tg,bool rw)551 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
552 {
553 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
554 return false;
555
556 return true;
557 }
558
calculate_io_allowed(u32 iops_limit,unsigned long jiffy_elapsed)559 static unsigned int calculate_io_allowed(u32 iops_limit,
560 unsigned long jiffy_elapsed)
561 {
562 unsigned int io_allowed;
563 u64 tmp;
564
565 /*
566 * jiffy_elapsed should not be a big value as minimum iops can be
567 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
568 * will allow dispatch after 1 second and after that slice should
569 * have been trimmed.
570 */
571
572 tmp = (u64)iops_limit * jiffy_elapsed;
573 do_div(tmp, HZ);
574
575 if (tmp > UINT_MAX)
576 io_allowed = UINT_MAX;
577 else
578 io_allowed = tmp;
579
580 return io_allowed;
581 }
582
calculate_bytes_allowed(u64 bps_limit,unsigned long jiffy_elapsed)583 static u64 calculate_bytes_allowed(u64 bps_limit, unsigned long jiffy_elapsed)
584 {
585 /*
586 * Can result be wider than 64 bits?
587 * We check against 62, not 64, due to ilog2 truncation.
588 */
589 if (ilog2(bps_limit) + ilog2(jiffy_elapsed) - ilog2(HZ) > 62)
590 return U64_MAX;
591 return mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed, (u64)HZ);
592 }
593
594 /* Trim the used slices and adjust slice start accordingly */
throtl_trim_slice(struct throtl_grp * tg,bool rw)595 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
596 {
597 unsigned long time_elapsed;
598 long long bytes_trim;
599 int io_trim;
600
601 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
602
603 /*
604 * If bps are unlimited (-1), then time slice don't get
605 * renewed. Don't try to trim the slice if slice is used. A new
606 * slice will start when appropriate.
607 */
608 if (throtl_slice_used(tg, rw))
609 return;
610
611 /*
612 * A bio has been dispatched. Also adjust slice_end. It might happen
613 * that initially cgroup limit was very low resulting in high
614 * slice_end, but later limit was bumped up and bio was dispatched
615 * sooner, then we need to reduce slice_end. A high bogus slice_end
616 * is bad because it does not allow new slice to start.
617 */
618
619 throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
620
621 time_elapsed = rounddown(jiffies - tg->slice_start[rw],
622 tg->td->throtl_slice);
623 if (!time_elapsed)
624 return;
625
626 bytes_trim = calculate_bytes_allowed(tg_bps_limit(tg, rw),
627 time_elapsed) +
628 tg->carryover_bytes[rw];
629 io_trim = calculate_io_allowed(tg_iops_limit(tg, rw), time_elapsed) +
630 tg->carryover_ios[rw];
631 if (bytes_trim <= 0 && io_trim <= 0)
632 return;
633
634 tg->carryover_bytes[rw] = 0;
635 if ((long long)tg->bytes_disp[rw] >= bytes_trim)
636 tg->bytes_disp[rw] -= bytes_trim;
637 else
638 tg->bytes_disp[rw] = 0;
639
640 tg->carryover_ios[rw] = 0;
641 if ((int)tg->io_disp[rw] >= io_trim)
642 tg->io_disp[rw] -= io_trim;
643 else
644 tg->io_disp[rw] = 0;
645
646 tg->slice_start[rw] += time_elapsed;
647
648 throtl_log(&tg->service_queue,
649 "[%c] trim slice nr=%lu bytes=%lld io=%d start=%lu end=%lu jiffies=%lu",
650 rw == READ ? 'R' : 'W', time_elapsed / tg->td->throtl_slice,
651 bytes_trim, io_trim, tg->slice_start[rw], tg->slice_end[rw],
652 jiffies);
653 }
654
__tg_update_carryover(struct throtl_grp * tg,bool rw)655 static void __tg_update_carryover(struct throtl_grp *tg, bool rw)
656 {
657 unsigned long jiffy_elapsed = jiffies - tg->slice_start[rw];
658 u64 bps_limit = tg_bps_limit(tg, rw);
659 u32 iops_limit = tg_iops_limit(tg, rw);
660
661 /*
662 * If config is updated while bios are still throttled, calculate and
663 * accumulate how many bytes/ios are waited across changes. And
664 * carryover_bytes/ios will be used to calculate new wait time under new
665 * configuration.
666 */
667 if (bps_limit != U64_MAX)
668 tg->carryover_bytes[rw] +=
669 calculate_bytes_allowed(bps_limit, jiffy_elapsed) -
670 tg->bytes_disp[rw];
671 if (iops_limit != UINT_MAX)
672 tg->carryover_ios[rw] +=
673 calculate_io_allowed(iops_limit, jiffy_elapsed) -
674 tg->io_disp[rw];
675 }
676
tg_update_carryover(struct throtl_grp * tg)677 static void tg_update_carryover(struct throtl_grp *tg)
678 {
679 if (tg->service_queue.nr_queued[READ])
680 __tg_update_carryover(tg, READ);
681 if (tg->service_queue.nr_queued[WRITE])
682 __tg_update_carryover(tg, WRITE);
683
684 /* see comments in struct throtl_grp for meaning of these fields. */
685 throtl_log(&tg->service_queue, "%s: %lld %lld %d %d\n", __func__,
686 tg->carryover_bytes[READ], tg->carryover_bytes[WRITE],
687 tg->carryover_ios[READ], tg->carryover_ios[WRITE]);
688 }
689
tg_within_iops_limit(struct throtl_grp * tg,struct bio * bio,u32 iops_limit)690 static unsigned long tg_within_iops_limit(struct throtl_grp *tg, struct bio *bio,
691 u32 iops_limit)
692 {
693 bool rw = bio_data_dir(bio);
694 int io_allowed;
695 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
696
697 if (iops_limit == UINT_MAX) {
698 return 0;
699 }
700
701 jiffy_elapsed = jiffies - tg->slice_start[rw];
702
703 /* Round up to the next throttle slice, wait time must be nonzero */
704 jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
705 io_allowed = calculate_io_allowed(iops_limit, jiffy_elapsed_rnd) +
706 tg->carryover_ios[rw];
707 if (io_allowed > 0 && tg->io_disp[rw] + 1 <= io_allowed)
708 return 0;
709
710 /* Calc approx time to dispatch */
711 jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
712 return jiffy_wait;
713 }
714
tg_within_bps_limit(struct throtl_grp * tg,struct bio * bio,u64 bps_limit)715 static unsigned long tg_within_bps_limit(struct throtl_grp *tg, struct bio *bio,
716 u64 bps_limit)
717 {
718 bool rw = bio_data_dir(bio);
719 long long bytes_allowed;
720 u64 extra_bytes;
721 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
722 unsigned int bio_size = throtl_bio_data_size(bio);
723
724 /* no need to throttle if this bio's bytes have been accounted */
725 if (bps_limit == U64_MAX || bio_flagged(bio, BIO_BPS_THROTTLED)) {
726 return 0;
727 }
728
729 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
730
731 /* Slice has just started. Consider one slice interval */
732 if (!jiffy_elapsed)
733 jiffy_elapsed_rnd = tg->td->throtl_slice;
734
735 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
736 bytes_allowed = calculate_bytes_allowed(bps_limit, jiffy_elapsed_rnd) +
737 tg->carryover_bytes[rw];
738 if (bytes_allowed > 0 && tg->bytes_disp[rw] + bio_size <= bytes_allowed)
739 return 0;
740
741 /* Calc approx time to dispatch */
742 extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
743 jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
744
745 if (!jiffy_wait)
746 jiffy_wait = 1;
747
748 /*
749 * This wait time is without taking into consideration the rounding
750 * up we did. Add that time also.
751 */
752 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
753 return jiffy_wait;
754 }
755
756 /*
757 * Returns whether one can dispatch a bio or not. Also returns approx number
758 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
759 */
tg_may_dispatch(struct throtl_grp * tg,struct bio * bio,unsigned long * wait)760 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
761 unsigned long *wait)
762 {
763 bool rw = bio_data_dir(bio);
764 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
765 u64 bps_limit = tg_bps_limit(tg, rw);
766 u32 iops_limit = tg_iops_limit(tg, rw);
767
768 /*
769 * Currently whole state machine of group depends on first bio
770 * queued in the group bio list. So one should not be calling
771 * this function with a different bio if there are other bios
772 * queued.
773 */
774 BUG_ON(tg->service_queue.nr_queued[rw] &&
775 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
776
777 /* If tg->bps = -1, then BW is unlimited */
778 if ((bps_limit == U64_MAX && iops_limit == UINT_MAX) ||
779 tg->flags & THROTL_TG_CANCELING) {
780 if (wait)
781 *wait = 0;
782 return true;
783 }
784
785 /*
786 * If previous slice expired, start a new one otherwise renew/extend
787 * existing slice to make sure it is at least throtl_slice interval
788 * long since now. New slice is started only for empty throttle group.
789 * If there is queued bio, that means there should be an active
790 * slice and it should be extended instead.
791 */
792 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
793 throtl_start_new_slice(tg, rw, true);
794 else {
795 if (time_before(tg->slice_end[rw],
796 jiffies + tg->td->throtl_slice))
797 throtl_extend_slice(tg, rw,
798 jiffies + tg->td->throtl_slice);
799 }
800
801 bps_wait = tg_within_bps_limit(tg, bio, bps_limit);
802 iops_wait = tg_within_iops_limit(tg, bio, iops_limit);
803 if (bps_wait + iops_wait == 0) {
804 if (wait)
805 *wait = 0;
806 return true;
807 }
808
809 max_wait = max(bps_wait, iops_wait);
810
811 if (wait)
812 *wait = max_wait;
813
814 if (time_before(tg->slice_end[rw], jiffies + max_wait))
815 throtl_extend_slice(tg, rw, jiffies + max_wait);
816
817 return false;
818 }
819
throtl_charge_bio(struct throtl_grp * tg,struct bio * bio)820 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
821 {
822 bool rw = bio_data_dir(bio);
823 unsigned int bio_size = throtl_bio_data_size(bio);
824
825 /* Charge the bio to the group */
826 if (!bio_flagged(bio, BIO_BPS_THROTTLED)) {
827 tg->bytes_disp[rw] += bio_size;
828 tg->last_bytes_disp[rw] += bio_size;
829 }
830
831 tg->io_disp[rw]++;
832 tg->last_io_disp[rw]++;
833 }
834
835 /**
836 * throtl_add_bio_tg - add a bio to the specified throtl_grp
837 * @bio: bio to add
838 * @qn: qnode to use
839 * @tg: the target throtl_grp
840 *
841 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
842 * tg->qnode_on_self[] is used.
843 */
throtl_add_bio_tg(struct bio * bio,struct throtl_qnode * qn,struct throtl_grp * tg)844 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
845 struct throtl_grp *tg)
846 {
847 struct throtl_service_queue *sq = &tg->service_queue;
848 bool rw = bio_data_dir(bio);
849
850 if (!qn)
851 qn = &tg->qnode_on_self[rw];
852
853 /*
854 * If @tg doesn't currently have any bios queued in the same
855 * direction, queueing @bio can change when @tg should be
856 * dispatched. Mark that @tg was empty. This is automatically
857 * cleared on the next tg_update_disptime().
858 */
859 if (!sq->nr_queued[rw])
860 tg->flags |= THROTL_TG_WAS_EMPTY;
861
862 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
863
864 sq->nr_queued[rw]++;
865 throtl_enqueue_tg(tg);
866 }
867
tg_update_disptime(struct throtl_grp * tg)868 static void tg_update_disptime(struct throtl_grp *tg)
869 {
870 struct throtl_service_queue *sq = &tg->service_queue;
871 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
872 struct bio *bio;
873
874 bio = throtl_peek_queued(&sq->queued[READ]);
875 if (bio)
876 tg_may_dispatch(tg, bio, &read_wait);
877
878 bio = throtl_peek_queued(&sq->queued[WRITE]);
879 if (bio)
880 tg_may_dispatch(tg, bio, &write_wait);
881
882 min_wait = min(read_wait, write_wait);
883 disptime = jiffies + min_wait;
884
885 /* Update dispatch time */
886 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
887 tg->disptime = disptime;
888 tg_service_queue_add(tg);
889
890 /* see throtl_add_bio_tg() */
891 tg->flags &= ~THROTL_TG_WAS_EMPTY;
892 }
893
start_parent_slice_with_credit(struct throtl_grp * child_tg,struct throtl_grp * parent_tg,bool rw)894 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
895 struct throtl_grp *parent_tg, bool rw)
896 {
897 if (throtl_slice_used(parent_tg, rw)) {
898 throtl_start_new_slice_with_credit(parent_tg, rw,
899 child_tg->slice_start[rw]);
900 }
901
902 }
903
tg_dispatch_one_bio(struct throtl_grp * tg,bool rw)904 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
905 {
906 struct throtl_service_queue *sq = &tg->service_queue;
907 struct throtl_service_queue *parent_sq = sq->parent_sq;
908 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
909 struct throtl_grp *tg_to_put = NULL;
910 struct bio *bio;
911
912 /*
913 * @bio is being transferred from @tg to @parent_sq. Popping a bio
914 * from @tg may put its reference and @parent_sq might end up
915 * getting released prematurely. Remember the tg to put and put it
916 * after @bio is transferred to @parent_sq.
917 */
918 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
919 sq->nr_queued[rw]--;
920
921 throtl_charge_bio(tg, bio);
922
923 /*
924 * If our parent is another tg, we just need to transfer @bio to
925 * the parent using throtl_add_bio_tg(). If our parent is
926 * @td->service_queue, @bio is ready to be issued. Put it on its
927 * bio_lists[] and decrease total number queued. The caller is
928 * responsible for issuing these bios.
929 */
930 if (parent_tg) {
931 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
932 start_parent_slice_with_credit(tg, parent_tg, rw);
933 } else {
934 bio_set_flag(bio, BIO_BPS_THROTTLED);
935 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
936 &parent_sq->queued[rw]);
937 BUG_ON(tg->td->nr_queued[rw] <= 0);
938 tg->td->nr_queued[rw]--;
939 }
940
941 throtl_trim_slice(tg, rw);
942
943 if (tg_to_put)
944 blkg_put(tg_to_blkg(tg_to_put));
945 }
946
throtl_dispatch_tg(struct throtl_grp * tg)947 static int throtl_dispatch_tg(struct throtl_grp *tg)
948 {
949 struct throtl_service_queue *sq = &tg->service_queue;
950 unsigned int nr_reads = 0, nr_writes = 0;
951 unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
952 unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
953 struct bio *bio;
954
955 /* Try to dispatch 75% READS and 25% WRITES */
956
957 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
958 tg_may_dispatch(tg, bio, NULL)) {
959
960 tg_dispatch_one_bio(tg, READ);
961 nr_reads++;
962
963 if (nr_reads >= max_nr_reads)
964 break;
965 }
966
967 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
968 tg_may_dispatch(tg, bio, NULL)) {
969
970 tg_dispatch_one_bio(tg, WRITE);
971 nr_writes++;
972
973 if (nr_writes >= max_nr_writes)
974 break;
975 }
976
977 return nr_reads + nr_writes;
978 }
979
throtl_select_dispatch(struct throtl_service_queue * parent_sq)980 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
981 {
982 unsigned int nr_disp = 0;
983
984 while (1) {
985 struct throtl_grp *tg;
986 struct throtl_service_queue *sq;
987
988 if (!parent_sq->nr_pending)
989 break;
990
991 tg = throtl_rb_first(parent_sq);
992 if (!tg)
993 break;
994
995 if (time_before(jiffies, tg->disptime))
996 break;
997
998 nr_disp += throtl_dispatch_tg(tg);
999
1000 sq = &tg->service_queue;
1001 if (sq->nr_queued[READ] || sq->nr_queued[WRITE])
1002 tg_update_disptime(tg);
1003 else
1004 throtl_dequeue_tg(tg);
1005
1006 if (nr_disp >= THROTL_QUANTUM)
1007 break;
1008 }
1009
1010 return nr_disp;
1011 }
1012
1013 /**
1014 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1015 * @t: the pending_timer member of the throtl_service_queue being serviced
1016 *
1017 * This timer is armed when a child throtl_grp with active bio's become
1018 * pending and queued on the service_queue's pending_tree and expires when
1019 * the first child throtl_grp should be dispatched. This function
1020 * dispatches bio's from the children throtl_grps to the parent
1021 * service_queue.
1022 *
1023 * If the parent's parent is another throtl_grp, dispatching is propagated
1024 * by either arming its pending_timer or repeating dispatch directly. If
1025 * the top-level service_tree is reached, throtl_data->dispatch_work is
1026 * kicked so that the ready bio's are issued.
1027 */
throtl_pending_timer_fn(struct timer_list * t)1028 static void throtl_pending_timer_fn(struct timer_list *t)
1029 {
1030 struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1031 struct throtl_grp *tg = sq_to_tg(sq);
1032 struct throtl_data *td = sq_to_td(sq);
1033 struct throtl_service_queue *parent_sq;
1034 struct request_queue *q;
1035 bool dispatched;
1036 int ret;
1037
1038 /* throtl_data may be gone, so figure out request queue by blkg */
1039 if (tg)
1040 q = tg->pd.blkg->q;
1041 else
1042 q = td->queue;
1043
1044 spin_lock_irq(&q->queue_lock);
1045
1046 if (!q->root_blkg)
1047 goto out_unlock;
1048
1049 again:
1050 parent_sq = sq->parent_sq;
1051 dispatched = false;
1052
1053 while (true) {
1054 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1055 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1056 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1057
1058 ret = throtl_select_dispatch(sq);
1059 if (ret) {
1060 throtl_log(sq, "bios disp=%u", ret);
1061 dispatched = true;
1062 }
1063
1064 if (throtl_schedule_next_dispatch(sq, false))
1065 break;
1066
1067 /* this dispatch windows is still open, relax and repeat */
1068 spin_unlock_irq(&q->queue_lock);
1069 cpu_relax();
1070 spin_lock_irq(&q->queue_lock);
1071 }
1072
1073 if (!dispatched)
1074 goto out_unlock;
1075
1076 if (parent_sq) {
1077 /* @parent_sq is another throl_grp, propagate dispatch */
1078 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1079 tg_update_disptime(tg);
1080 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1081 /* window is already open, repeat dispatching */
1082 sq = parent_sq;
1083 tg = sq_to_tg(sq);
1084 goto again;
1085 }
1086 }
1087 } else {
1088 /* reached the top-level, queue issuing */
1089 queue_work(kthrotld_workqueue, &td->dispatch_work);
1090 }
1091 out_unlock:
1092 spin_unlock_irq(&q->queue_lock);
1093 }
1094
1095 /**
1096 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1097 * @work: work item being executed
1098 *
1099 * This function is queued for execution when bios reach the bio_lists[]
1100 * of throtl_data->service_queue. Those bios are ready and issued by this
1101 * function.
1102 */
blk_throtl_dispatch_work_fn(struct work_struct * work)1103 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1104 {
1105 struct throtl_data *td = container_of(work, struct throtl_data,
1106 dispatch_work);
1107 struct throtl_service_queue *td_sq = &td->service_queue;
1108 struct request_queue *q = td->queue;
1109 struct bio_list bio_list_on_stack;
1110 struct bio *bio;
1111 struct blk_plug plug;
1112 int rw;
1113
1114 bio_list_init(&bio_list_on_stack);
1115
1116 spin_lock_irq(&q->queue_lock);
1117 for (rw = READ; rw <= WRITE; rw++)
1118 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1119 bio_list_add(&bio_list_on_stack, bio);
1120 spin_unlock_irq(&q->queue_lock);
1121
1122 if (!bio_list_empty(&bio_list_on_stack)) {
1123 blk_start_plug(&plug);
1124 while ((bio = bio_list_pop(&bio_list_on_stack)))
1125 submit_bio_noacct_nocheck(bio);
1126 blk_finish_plug(&plug);
1127 }
1128 }
1129
tg_prfill_conf_u64(struct seq_file * sf,struct blkg_policy_data * pd,int off)1130 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1131 int off)
1132 {
1133 struct throtl_grp *tg = pd_to_tg(pd);
1134 u64 v = *(u64 *)((void *)tg + off);
1135
1136 if (v == U64_MAX)
1137 return 0;
1138 return __blkg_prfill_u64(sf, pd, v);
1139 }
1140
tg_prfill_conf_uint(struct seq_file * sf,struct blkg_policy_data * pd,int off)1141 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1142 int off)
1143 {
1144 struct throtl_grp *tg = pd_to_tg(pd);
1145 unsigned int v = *(unsigned int *)((void *)tg + off);
1146
1147 if (v == UINT_MAX)
1148 return 0;
1149 return __blkg_prfill_u64(sf, pd, v);
1150 }
1151
tg_print_conf_u64(struct seq_file * sf,void * v)1152 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1153 {
1154 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1155 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1156 return 0;
1157 }
1158
tg_print_conf_uint(struct seq_file * sf,void * v)1159 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1160 {
1161 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1162 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1163 return 0;
1164 }
1165
tg_conf_updated(struct throtl_grp * tg,bool global)1166 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1167 {
1168 struct throtl_service_queue *sq = &tg->service_queue;
1169 struct cgroup_subsys_state *pos_css;
1170 struct blkcg_gq *blkg;
1171
1172 throtl_log(&tg->service_queue,
1173 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1174 tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1175 tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1176
1177 rcu_read_lock();
1178 /*
1179 * Update has_rules[] flags for the updated tg's subtree. A tg is
1180 * considered to have rules if either the tg itself or any of its
1181 * ancestors has rules. This identifies groups without any
1182 * restrictions in the whole hierarchy and allows them to bypass
1183 * blk-throttle.
1184 */
1185 blkg_for_each_descendant_pre(blkg, pos_css,
1186 global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1187 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1188
1189 tg_update_has_rules(this_tg);
1190 /* ignore root/second level */
1191 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1192 !blkg->parent->parent)
1193 continue;
1194 }
1195 rcu_read_unlock();
1196
1197 /*
1198 * We're already holding queue_lock and know @tg is valid. Let's
1199 * apply the new config directly.
1200 *
1201 * Restart the slices for both READ and WRITES. It might happen
1202 * that a group's limit are dropped suddenly and we don't want to
1203 * account recently dispatched IO with new low rate.
1204 */
1205 throtl_start_new_slice(tg, READ, false);
1206 throtl_start_new_slice(tg, WRITE, false);
1207
1208 if (tg->flags & THROTL_TG_PENDING) {
1209 tg_update_disptime(tg);
1210 throtl_schedule_next_dispatch(sq->parent_sq, true);
1211 }
1212 }
1213
blk_throtl_init(struct gendisk * disk)1214 static int blk_throtl_init(struct gendisk *disk)
1215 {
1216 struct request_queue *q = disk->queue;
1217 struct throtl_data *td;
1218 int ret;
1219
1220 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1221 if (!td)
1222 return -ENOMEM;
1223
1224 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1225 throtl_service_queue_init(&td->service_queue);
1226
1227 /*
1228 * Freeze queue before activating policy, to synchronize with IO path,
1229 * which is protected by 'q_usage_counter'.
1230 */
1231 blk_mq_freeze_queue(disk->queue);
1232 blk_mq_quiesce_queue(disk->queue);
1233
1234 q->td = td;
1235 td->queue = q;
1236
1237 /* activate policy */
1238 ret = blkcg_activate_policy(disk, &blkcg_policy_throtl);
1239 if (ret) {
1240 q->td = NULL;
1241 kfree(td);
1242 goto out;
1243 }
1244
1245 if (blk_queue_nonrot(q))
1246 td->throtl_slice = DFL_THROTL_SLICE_SSD;
1247 else
1248 td->throtl_slice = DFL_THROTL_SLICE_HD;
1249 td->track_bio_latency = !queue_is_mq(q);
1250 if (!td->track_bio_latency)
1251 blk_stat_enable_accounting(q);
1252
1253 out:
1254 blk_mq_unquiesce_queue(disk->queue);
1255 blk_mq_unfreeze_queue(disk->queue);
1256
1257 return ret;
1258 }
1259
1260
tg_set_conf(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off,bool is_u64)1261 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1262 char *buf, size_t nbytes, loff_t off, bool is_u64)
1263 {
1264 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1265 struct blkg_conf_ctx ctx;
1266 struct throtl_grp *tg;
1267 int ret;
1268 u64 v;
1269
1270 blkg_conf_init(&ctx, buf);
1271
1272 ret = blkg_conf_open_bdev(&ctx);
1273 if (ret)
1274 goto out_finish;
1275
1276 if (!blk_throtl_activated(ctx.bdev->bd_queue)) {
1277 ret = blk_throtl_init(ctx.bdev->bd_disk);
1278 if (ret)
1279 goto out_finish;
1280 }
1281
1282 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
1283 if (ret)
1284 goto out_finish;
1285
1286 ret = -EINVAL;
1287 if (sscanf(ctx.body, "%llu", &v) != 1)
1288 goto out_finish;
1289 if (!v)
1290 v = U64_MAX;
1291
1292 tg = blkg_to_tg(ctx.blkg);
1293 tg_update_carryover(tg);
1294
1295 if (is_u64)
1296 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1297 else
1298 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1299
1300 tg_conf_updated(tg, false);
1301 ret = 0;
1302 out_finish:
1303 blkg_conf_exit(&ctx);
1304 return ret ?: nbytes;
1305 }
1306
tg_set_conf_u64(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)1307 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1308 char *buf, size_t nbytes, loff_t off)
1309 {
1310 return tg_set_conf(of, buf, nbytes, off, true);
1311 }
1312
tg_set_conf_uint(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)1313 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1314 char *buf, size_t nbytes, loff_t off)
1315 {
1316 return tg_set_conf(of, buf, nbytes, off, false);
1317 }
1318
tg_print_rwstat(struct seq_file * sf,void * v)1319 static int tg_print_rwstat(struct seq_file *sf, void *v)
1320 {
1321 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1322 blkg_prfill_rwstat, &blkcg_policy_throtl,
1323 seq_cft(sf)->private, true);
1324 return 0;
1325 }
1326
tg_prfill_rwstat_recursive(struct seq_file * sf,struct blkg_policy_data * pd,int off)1327 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1328 struct blkg_policy_data *pd, int off)
1329 {
1330 struct blkg_rwstat_sample sum;
1331
1332 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1333 &sum);
1334 return __blkg_prfill_rwstat(sf, pd, &sum);
1335 }
1336
tg_print_rwstat_recursive(struct seq_file * sf,void * v)1337 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1338 {
1339 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1340 tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1341 seq_cft(sf)->private, true);
1342 return 0;
1343 }
1344
1345 static struct cftype throtl_legacy_files[] = {
1346 {
1347 .name = "throttle.read_bps_device",
1348 .private = offsetof(struct throtl_grp, bps[READ]),
1349 .seq_show = tg_print_conf_u64,
1350 .write = tg_set_conf_u64,
1351 },
1352 {
1353 .name = "throttle.write_bps_device",
1354 .private = offsetof(struct throtl_grp, bps[WRITE]),
1355 .seq_show = tg_print_conf_u64,
1356 .write = tg_set_conf_u64,
1357 },
1358 {
1359 .name = "throttle.read_iops_device",
1360 .private = offsetof(struct throtl_grp, iops[READ]),
1361 .seq_show = tg_print_conf_uint,
1362 .write = tg_set_conf_uint,
1363 },
1364 {
1365 .name = "throttle.write_iops_device",
1366 .private = offsetof(struct throtl_grp, iops[WRITE]),
1367 .seq_show = tg_print_conf_uint,
1368 .write = tg_set_conf_uint,
1369 },
1370 {
1371 .name = "throttle.io_service_bytes",
1372 .private = offsetof(struct throtl_grp, stat_bytes),
1373 .seq_show = tg_print_rwstat,
1374 },
1375 {
1376 .name = "throttle.io_service_bytes_recursive",
1377 .private = offsetof(struct throtl_grp, stat_bytes),
1378 .seq_show = tg_print_rwstat_recursive,
1379 },
1380 {
1381 .name = "throttle.io_serviced",
1382 .private = offsetof(struct throtl_grp, stat_ios),
1383 .seq_show = tg_print_rwstat,
1384 },
1385 {
1386 .name = "throttle.io_serviced_recursive",
1387 .private = offsetof(struct throtl_grp, stat_ios),
1388 .seq_show = tg_print_rwstat_recursive,
1389 },
1390 { } /* terminate */
1391 };
1392
tg_prfill_limit(struct seq_file * sf,struct blkg_policy_data * pd,int off)1393 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1394 int off)
1395 {
1396 struct throtl_grp *tg = pd_to_tg(pd);
1397 const char *dname = blkg_dev_name(pd->blkg);
1398 u64 bps_dft;
1399 unsigned int iops_dft;
1400
1401 if (!dname)
1402 return 0;
1403
1404 bps_dft = U64_MAX;
1405 iops_dft = UINT_MAX;
1406
1407 if (tg->bps_conf[READ] == bps_dft &&
1408 tg->bps_conf[WRITE] == bps_dft &&
1409 tg->iops_conf[READ] == iops_dft &&
1410 tg->iops_conf[WRITE] == iops_dft)
1411 return 0;
1412
1413 seq_printf(sf, "%s", dname);
1414 if (tg->bps_conf[READ] == U64_MAX)
1415 seq_printf(sf, " rbps=max");
1416 else
1417 seq_printf(sf, " rbps=%llu", tg->bps_conf[READ]);
1418
1419 if (tg->bps_conf[WRITE] == U64_MAX)
1420 seq_printf(sf, " wbps=max");
1421 else
1422 seq_printf(sf, " wbps=%llu", tg->bps_conf[WRITE]);
1423
1424 if (tg->iops_conf[READ] == UINT_MAX)
1425 seq_printf(sf, " riops=max");
1426 else
1427 seq_printf(sf, " riops=%u", tg->iops_conf[READ]);
1428
1429 if (tg->iops_conf[WRITE] == UINT_MAX)
1430 seq_printf(sf, " wiops=max");
1431 else
1432 seq_printf(sf, " wiops=%u", tg->iops_conf[WRITE]);
1433
1434 seq_printf(sf, "\n");
1435 return 0;
1436 }
1437
tg_print_limit(struct seq_file * sf,void * v)1438 static int tg_print_limit(struct seq_file *sf, void *v)
1439 {
1440 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1441 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1442 return 0;
1443 }
1444
tg_set_limit(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)1445 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1446 char *buf, size_t nbytes, loff_t off)
1447 {
1448 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1449 struct blkg_conf_ctx ctx;
1450 struct throtl_grp *tg;
1451 u64 v[4];
1452 int ret;
1453
1454 blkg_conf_init(&ctx, buf);
1455
1456 ret = blkg_conf_open_bdev(&ctx);
1457 if (ret)
1458 goto out_finish;
1459
1460 if (!blk_throtl_activated(ctx.bdev->bd_queue)) {
1461 ret = blk_throtl_init(ctx.bdev->bd_disk);
1462 if (ret)
1463 goto out_finish;
1464 }
1465
1466 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
1467 if (ret)
1468 goto out_finish;
1469
1470 tg = blkg_to_tg(ctx.blkg);
1471 tg_update_carryover(tg);
1472
1473 v[0] = tg->bps[READ];
1474 v[1] = tg->bps[WRITE];
1475 v[2] = tg->iops[READ];
1476 v[3] = tg->iops[WRITE];
1477
1478 while (true) {
1479 char tok[27]; /* wiops=18446744073709551616 */
1480 char *p;
1481 u64 val = U64_MAX;
1482 int len;
1483
1484 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1485 break;
1486 if (tok[0] == '\0')
1487 break;
1488 ctx.body += len;
1489
1490 ret = -EINVAL;
1491 p = tok;
1492 strsep(&p, "=");
1493 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1494 goto out_finish;
1495
1496 ret = -ERANGE;
1497 if (!val)
1498 goto out_finish;
1499
1500 ret = -EINVAL;
1501 if (!strcmp(tok, "rbps") && val > 1)
1502 v[0] = val;
1503 else if (!strcmp(tok, "wbps") && val > 1)
1504 v[1] = val;
1505 else if (!strcmp(tok, "riops") && val > 1)
1506 v[2] = min_t(u64, val, UINT_MAX);
1507 else if (!strcmp(tok, "wiops") && val > 1)
1508 v[3] = min_t(u64, val, UINT_MAX);
1509 else
1510 goto out_finish;
1511 }
1512
1513 tg->bps[READ] = v[0];
1514 tg->bps[WRITE] = v[1];
1515 tg->iops[READ] = v[2];
1516 tg->iops[WRITE] = v[3];
1517
1518 tg_conf_updated(tg, false);
1519 ret = 0;
1520 out_finish:
1521 blkg_conf_exit(&ctx);
1522 return ret ?: nbytes;
1523 }
1524
1525 static struct cftype throtl_files[] = {
1526 {
1527 .name = "max",
1528 .flags = CFTYPE_NOT_ON_ROOT,
1529 .seq_show = tg_print_limit,
1530 .write = tg_set_limit,
1531 },
1532 { } /* terminate */
1533 };
1534
throtl_shutdown_wq(struct request_queue * q)1535 static void throtl_shutdown_wq(struct request_queue *q)
1536 {
1537 struct throtl_data *td = q->td;
1538
1539 cancel_work_sync(&td->dispatch_work);
1540 }
1541
1542 struct blkcg_policy blkcg_policy_throtl = {
1543 .dfl_cftypes = throtl_files,
1544 .legacy_cftypes = throtl_legacy_files,
1545
1546 .pd_alloc_fn = throtl_pd_alloc,
1547 .pd_init_fn = throtl_pd_init,
1548 .pd_online_fn = throtl_pd_online,
1549 .pd_free_fn = throtl_pd_free,
1550 };
1551
blk_throtl_cancel_bios(struct gendisk * disk)1552 void blk_throtl_cancel_bios(struct gendisk *disk)
1553 {
1554 struct request_queue *q = disk->queue;
1555 struct cgroup_subsys_state *pos_css;
1556 struct blkcg_gq *blkg;
1557
1558 if (!blk_throtl_activated(q))
1559 return;
1560
1561 spin_lock_irq(&q->queue_lock);
1562 /*
1563 * queue_lock is held, rcu lock is not needed here technically.
1564 * However, rcu lock is still held to emphasize that following
1565 * path need RCU protection and to prevent warning from lockdep.
1566 */
1567 rcu_read_lock();
1568 blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) {
1569 struct throtl_grp *tg = blkg_to_tg(blkg);
1570 struct throtl_service_queue *sq = &tg->service_queue;
1571
1572 /*
1573 * Set the flag to make sure throtl_pending_timer_fn() won't
1574 * stop until all throttled bios are dispatched.
1575 */
1576 tg->flags |= THROTL_TG_CANCELING;
1577
1578 /*
1579 * Do not dispatch cgroup without THROTL_TG_PENDING or cgroup
1580 * will be inserted to service queue without THROTL_TG_PENDING
1581 * set in tg_update_disptime below. Then IO dispatched from
1582 * child in tg_dispatch_one_bio will trigger double insertion
1583 * and corrupt the tree.
1584 */
1585 if (!(tg->flags & THROTL_TG_PENDING))
1586 continue;
1587
1588 /*
1589 * Update disptime after setting the above flag to make sure
1590 * throtl_select_dispatch() won't exit without dispatching.
1591 */
1592 tg_update_disptime(tg);
1593
1594 throtl_schedule_pending_timer(sq, jiffies + 1);
1595 }
1596 rcu_read_unlock();
1597 spin_unlock_irq(&q->queue_lock);
1598 }
1599
__blk_throtl_bio(struct bio * bio)1600 bool __blk_throtl_bio(struct bio *bio)
1601 {
1602 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
1603 struct blkcg_gq *blkg = bio->bi_blkg;
1604 struct throtl_qnode *qn = NULL;
1605 struct throtl_grp *tg = blkg_to_tg(blkg);
1606 struct throtl_service_queue *sq;
1607 bool rw = bio_data_dir(bio);
1608 bool throttled = false;
1609 struct throtl_data *td = tg->td;
1610
1611 rcu_read_lock();
1612 spin_lock_irq(&q->queue_lock);
1613 sq = &tg->service_queue;
1614
1615 while (true) {
1616 if (tg->last_low_overflow_time[rw] == 0)
1617 tg->last_low_overflow_time[rw] = jiffies;
1618 /* throtl is FIFO - if bios are already queued, should queue */
1619 if (sq->nr_queued[rw])
1620 break;
1621
1622 /* if above limits, break to queue */
1623 if (!tg_may_dispatch(tg, bio, NULL)) {
1624 tg->last_low_overflow_time[rw] = jiffies;
1625 break;
1626 }
1627
1628 /* within limits, let's charge and dispatch directly */
1629 throtl_charge_bio(tg, bio);
1630
1631 /*
1632 * We need to trim slice even when bios are not being queued
1633 * otherwise it might happen that a bio is not queued for
1634 * a long time and slice keeps on extending and trim is not
1635 * called for a long time. Now if limits are reduced suddenly
1636 * we take into account all the IO dispatched so far at new
1637 * low rate and * newly queued IO gets a really long dispatch
1638 * time.
1639 *
1640 * So keep on trimming slice even if bio is not queued.
1641 */
1642 throtl_trim_slice(tg, rw);
1643
1644 /*
1645 * @bio passed through this layer without being throttled.
1646 * Climb up the ladder. If we're already at the top, it
1647 * can be executed directly.
1648 */
1649 qn = &tg->qnode_on_parent[rw];
1650 sq = sq->parent_sq;
1651 tg = sq_to_tg(sq);
1652 if (!tg) {
1653 bio_set_flag(bio, BIO_BPS_THROTTLED);
1654 goto out_unlock;
1655 }
1656 }
1657
1658 /* out-of-limit, queue to @tg */
1659 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1660 rw == READ ? 'R' : 'W',
1661 tg->bytes_disp[rw], bio->bi_iter.bi_size,
1662 tg_bps_limit(tg, rw),
1663 tg->io_disp[rw], tg_iops_limit(tg, rw),
1664 sq->nr_queued[READ], sq->nr_queued[WRITE]);
1665
1666 tg->last_low_overflow_time[rw] = jiffies;
1667
1668 td->nr_queued[rw]++;
1669 throtl_add_bio_tg(bio, qn, tg);
1670 throttled = true;
1671
1672 /*
1673 * Update @tg's dispatch time and force schedule dispatch if @tg
1674 * was empty before @bio. The forced scheduling isn't likely to
1675 * cause undue delay as @bio is likely to be dispatched directly if
1676 * its @tg's disptime is not in the future.
1677 */
1678 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1679 tg_update_disptime(tg);
1680 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1681 }
1682
1683 out_unlock:
1684 spin_unlock_irq(&q->queue_lock);
1685
1686 rcu_read_unlock();
1687 return throttled;
1688 }
1689
blk_throtl_exit(struct gendisk * disk)1690 void blk_throtl_exit(struct gendisk *disk)
1691 {
1692 struct request_queue *q = disk->queue;
1693
1694 if (!blk_throtl_activated(q))
1695 return;
1696
1697 del_timer_sync(&q->td->service_queue.pending_timer);
1698 throtl_shutdown_wq(q);
1699 blkcg_deactivate_policy(disk, &blkcg_policy_throtl);
1700 kfree(q->td);
1701 }
1702
throtl_init(void)1703 static int __init throtl_init(void)
1704 {
1705 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1706 if (!kthrotld_workqueue)
1707 panic("Failed to create kthrotld\n");
1708
1709 return blkcg_policy_register(&blkcg_policy_throtl);
1710 }
1711
1712 module_init(throtl_init);
1713