blob: 16f5766620a41043645756c51d441f4488af9edf [file] [log] [blame]
Greg Kroah-Hartmanb2441312017-11-01 15:07:57 +01001// SPDX-License-Identifier: GPL-2.0
Vivek Goyale43473b2010-09-15 17:06:35 -04002/*
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>
Tejun Heobc9fcbf2011-10-19 14:31:18 +020013#include "blk.h"
Tejun Heo1d1566462019-11-07 11:18:04 -080014#include "blk-cgroup-rwstat.h"
Christoph Hellwige4a19f72021-11-23 19:53:09 +010015#include "blk-stat.h"
Jens Axboea7b36ee2021-10-05 09:11:56 -060016#include "blk-throttle.h"
Vivek Goyale43473b2010-09-15 17:06:35 -040017
18/* Max dispatch from a group in 1 round */
Baolin Wange675df22020-09-07 16:10:15 +080019#define THROTL_GRP_QUANTUM 8
Vivek Goyale43473b2010-09-15 17:06:35 -040020
21/* Total max dispatch from all groups in one round */
Baolin Wange675df22020-09-07 16:10:15 +080022#define THROTL_QUANTUM 32
Vivek Goyale43473b2010-09-15 17:06:35 -040023
Shaohua Lid61fcfa2017-03-27 10:51:38 -070024/* 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)
Shaohua Li297e3d82017-03-27 10:51:37 -070027#define MAX_THROTL_SLICE (HZ)
Shaohua Li9e234ee2017-03-27 10:51:41 -070028#define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
Shaohua Li9bb67ae2017-05-17 13:07:26 -070029#define MIN_THROTL_BPS (320 * 1024)
30#define MIN_THROTL_IOPS (10)
Shaohua Lib4f428e2017-05-17 13:07:27 -070031#define DFL_LATENCY_TARGET (-1L)
32#define DFL_IDLE_THRESHOLD (0)
Shaohua Li6679a902017-06-06 12:40:43 -070033#define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
34#define LATENCY_FILTERED_SSD (0)
35/*
36 * For HD, very small latency comes from sequential IO. Such IO is helpless to
37 * help determine if its IO is impacted by others, hence we ignore the IO
38 */
39#define LATENCY_FILTERED_HD (1000L) /* 1ms */
Vivek Goyale43473b2010-09-15 17:06:35 -040040
Vivek Goyal450adcb2011-03-01 13:40:54 -050041/* A workqueue to queue throttle related work */
42static struct workqueue_struct *kthrotld_workqueue;
Vivek Goyal450adcb2011-03-01 13:40:54 -050043
Vivek Goyale43473b2010-09-15 17:06:35 -040044#define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
45
Shaohua Lib9147dd2017-03-27 15:19:42 -070046/* We measure latency for request size from <= 4k to >= 1M */
47#define LATENCY_BUCKET_SIZE 9
48
49struct latency_bucket {
50 unsigned long total_latency; /* ns / 1024 */
51 int samples;
52};
53
54struct avg_latency_bucket {
55 unsigned long latency; /* ns / 1024 */
56 bool valid;
57};
58
Vivek Goyale43473b2010-09-15 17:06:35 -040059struct throtl_data
60{
Vivek Goyale43473b2010-09-15 17:06:35 -040061 /* service tree for active throtl groups */
Tejun Heoc9e03322013-05-14 13:52:32 -070062 struct throtl_service_queue service_queue;
Vivek Goyale43473b2010-09-15 17:06:35 -040063
Vivek Goyale43473b2010-09-15 17:06:35 -040064 struct request_queue *queue;
65
66 /* Total Number of queued bios on READ and WRITE lists */
67 unsigned int nr_queued[2];
68
Shaohua Li297e3d82017-03-27 10:51:37 -070069 unsigned int throtl_slice;
70
Vivek Goyale43473b2010-09-15 17:06:35 -040071 /* Work for dispatching throttled bios */
Tejun Heo69df0ab2013-05-14 13:52:36 -070072 struct work_struct dispatch_work;
Shaohua Li9f626e32017-03-27 10:51:30 -070073 unsigned int limit_index;
74 bool limit_valid[LIMIT_CNT];
Shaohua Li3f0abd82017-03-27 10:51:35 -070075
76 unsigned long low_upgrade_time;
77 unsigned long low_downgrade_time;
Shaohua Li7394e312017-03-27 10:51:40 -070078
79 unsigned int scale;
Shaohua Lib9147dd2017-03-27 15:19:42 -070080
Joseph Qib889bf62017-11-21 09:38:30 +080081 struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
82 struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
83 struct latency_bucket __percpu *latency_buckets[2];
Shaohua Lib9147dd2017-03-27 15:19:42 -070084 unsigned long last_calculate_time;
Shaohua Li6679a902017-06-06 12:40:43 -070085 unsigned long filtered_latency;
Shaohua Lib9147dd2017-03-27 15:19:42 -070086
87 bool track_bio_latency;
Vivek Goyale43473b2010-09-15 17:06:35 -040088};
89
Kees Cooke99e88a2017-10-16 14:43:17 -070090static void throtl_pending_timer_fn(struct timer_list *t);
Tejun Heo69df0ab2013-05-14 13:52:36 -070091
Tejun Heo3c798392012-04-16 13:57:25 -070092static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
Tejun Heo03814112012-03-05 13:15:14 -080093{
Tejun Heof95a04a2012-04-16 13:57:26 -070094 return pd_to_blkg(&tg->pd);
Tejun Heo03814112012-03-05 13:15:14 -080095}
96
Tejun Heofda6f272013-05-14 13:52:36 -070097/**
98 * sq_to_tg - return the throl_grp the specified service queue belongs to
99 * @sq: the throtl_service_queue of interest
100 *
101 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
102 * embedded in throtl_data, %NULL is returned.
103 */
104static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
105{
106 if (sq && sq->parent_sq)
107 return container_of(sq, struct throtl_grp, service_queue);
108 else
109 return NULL;
110}
Vivek Goyale43473b2010-09-15 17:06:35 -0400111
Tejun Heofda6f272013-05-14 13:52:36 -0700112/**
113 * sq_to_td - return throtl_data the specified service queue belongs to
114 * @sq: the throtl_service_queue of interest
115 *
Masahiro Yamadab43daedc2017-02-27 14:29:09 -0800116 * A service_queue can be embedded in either a throtl_grp or throtl_data.
Tejun Heofda6f272013-05-14 13:52:36 -0700117 * Determine the associated throtl_data accordingly and return it.
118 */
119static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
120{
121 struct throtl_grp *tg = sq_to_tg(sq);
122
123 if (tg)
124 return tg->td;
125 else
126 return container_of(sq, struct throtl_data, service_queue);
127}
128
Shaohua Li7394e312017-03-27 10:51:40 -0700129/*
130 * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
131 * make the IO dispatch more smooth.
Kemeng Shi009df342022-12-05 19:57:06 +0800132 * Scale up: linearly scale up according to elapsed time since upgrade. For
Shaohua Li7394e312017-03-27 10:51:40 -0700133 * every throtl_slice, the limit scales up 1/2 .low limit till the
134 * limit hits .max limit
135 * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
136 */
137static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
138{
139 /* arbitrary value to avoid too big scale */
140 if (td->scale < 4096 && time_after_eq(jiffies,
141 td->low_upgrade_time + td->scale * td->throtl_slice))
142 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
143
144 return low + (low >> 1) * td->scale;
145}
146
Shaohua Li9f626e32017-03-27 10:51:30 -0700147static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
148{
Shaohua Lib22c4172017-03-27 10:51:33 -0700149 struct blkcg_gq *blkg = tg_to_blkg(tg);
Shaohua Li7394e312017-03-27 10:51:40 -0700150 struct throtl_data *td;
Shaohua Lib22c4172017-03-27 10:51:33 -0700151 uint64_t ret;
152
153 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
154 return U64_MAX;
Shaohua Li7394e312017-03-27 10:51:40 -0700155
156 td = tg->td;
157 ret = tg->bps[rw][td->limit_index];
Shaohua Li9bb67ae2017-05-17 13:07:26 -0700158 if (ret == 0 && td->limit_index == LIMIT_LOW) {
159 /* intermediate node or iops isn't 0 */
160 if (!list_empty(&blkg->blkcg->css.children) ||
161 tg->iops[rw][td->limit_index])
162 return U64_MAX;
163 else
164 return MIN_THROTL_BPS;
165 }
Shaohua Li7394e312017-03-27 10:51:40 -0700166
167 if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
168 tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
169 uint64_t adjusted;
170
171 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
172 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
173 }
Shaohua Lib22c4172017-03-27 10:51:33 -0700174 return ret;
Shaohua Li9f626e32017-03-27 10:51:30 -0700175}
176
177static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
178{
Shaohua Lib22c4172017-03-27 10:51:33 -0700179 struct blkcg_gq *blkg = tg_to_blkg(tg);
Shaohua Li7394e312017-03-27 10:51:40 -0700180 struct throtl_data *td;
Shaohua Lib22c4172017-03-27 10:51:33 -0700181 unsigned int ret;
182
183 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
184 return UINT_MAX;
Shaohua Li9bb67ae2017-05-17 13:07:26 -0700185
Shaohua Li7394e312017-03-27 10:51:40 -0700186 td = tg->td;
187 ret = tg->iops[rw][td->limit_index];
Shaohua Li9bb67ae2017-05-17 13:07:26 -0700188 if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
189 /* intermediate node or bps isn't 0 */
190 if (!list_empty(&blkg->blkcg->css.children) ||
191 tg->bps[rw][td->limit_index])
192 return UINT_MAX;
193 else
194 return MIN_THROTL_IOPS;
195 }
Shaohua Li7394e312017-03-27 10:51:40 -0700196
197 if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
198 tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
199 uint64_t adjusted;
200
201 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
202 if (adjusted > UINT_MAX)
203 adjusted = UINT_MAX;
204 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
205 }
Shaohua Lib22c4172017-03-27 10:51:33 -0700206 return ret;
Shaohua Li9f626e32017-03-27 10:51:30 -0700207}
208
Shaohua Lib9147dd2017-03-27 15:19:42 -0700209#define request_bucket_index(sectors) \
210 clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
211
Tejun Heofda6f272013-05-14 13:52:36 -0700212/**
213 * throtl_log - log debug message via blktrace
214 * @sq: the service_queue being reported
215 * @fmt: printf format string
216 * @args: printf args
217 *
218 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
219 * throtl_grp; otherwise, just "throtl".
Tejun Heofda6f272013-05-14 13:52:36 -0700220 */
221#define throtl_log(sq, fmt, args...) do { \
222 struct throtl_grp *__tg = sq_to_tg((sq)); \
223 struct throtl_data *__td = sq_to_td((sq)); \
224 \
225 (void)__td; \
Shaohua Li59fa0222016-05-09 17:22:15 -0700226 if (likely(!blk_trace_note_message_enabled(__td->queue))) \
227 break; \
Tejun Heofda6f272013-05-14 13:52:36 -0700228 if ((__tg)) { \
Shaohua Li35fe6d72017-07-12 11:49:56 -0700229 blk_add_cgroup_trace_msg(__td->queue, \
Christoph Hellwigf4a6a612022-04-20 06:27:16 +0200230 &tg_to_blkg(__tg)->blkcg->css, "throtl " fmt, ##args);\
Tejun Heofda6f272013-05-14 13:52:36 -0700231 } else { \
232 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
233 } \
234} while (0)
Vivek Goyale43473b2010-09-15 17:06:35 -0400235
Shaohua Liea0ea2b2017-08-18 16:08:13 -0700236static inline unsigned int throtl_bio_data_size(struct bio *bio)
237{
238 /* assume it's one sector */
239 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
240 return 512;
241 return bio->bi_iter.bi_size;
242}
243
Tejun Heoc5cc2072013-05-14 13:52:38 -0700244static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
245{
246 INIT_LIST_HEAD(&qn->node);
247 bio_list_init(&qn->bios);
248 qn->tg = tg;
249}
250
251/**
252 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
253 * @bio: bio being added
254 * @qn: qnode to add bio to
255 * @queued: the service_queue->queued[] list @qn belongs to
256 *
257 * Add @bio to @qn and put @qn on @queued if it's not already on.
258 * @qn->tg's reference count is bumped when @qn is activated. See the
259 * comment on top of throtl_qnode definition for details.
260 */
261static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
262 struct list_head *queued)
263{
264 bio_list_add(&qn->bios, bio);
265 if (list_empty(&qn->node)) {
266 list_add_tail(&qn->node, queued);
267 blkg_get(tg_to_blkg(qn->tg));
268 }
269}
270
271/**
272 * throtl_peek_queued - peek the first bio on a qnode list
273 * @queued: the qnode list to peek
274 */
275static struct bio *throtl_peek_queued(struct list_head *queued)
276{
Baolin Wangb7b609d2020-10-08 11:52:26 +0800277 struct throtl_qnode *qn;
Tejun Heoc5cc2072013-05-14 13:52:38 -0700278 struct bio *bio;
279
280 if (list_empty(queued))
281 return NULL;
282
Baolin Wangb7b609d2020-10-08 11:52:26 +0800283 qn = list_first_entry(queued, struct throtl_qnode, node);
Tejun Heoc5cc2072013-05-14 13:52:38 -0700284 bio = bio_list_peek(&qn->bios);
285 WARN_ON_ONCE(!bio);
286 return bio;
287}
288
289/**
290 * throtl_pop_queued - pop the first bio form a qnode list
291 * @queued: the qnode list to pop a bio from
292 * @tg_to_put: optional out argument for throtl_grp to put
293 *
294 * Pop the first bio from the qnode list @queued. After popping, the first
295 * qnode is removed from @queued if empty or moved to the end of @queued so
296 * that the popping order is round-robin.
297 *
298 * When the first qnode is removed, its associated throtl_grp should be put
299 * too. If @tg_to_put is NULL, this function automatically puts it;
300 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
301 * responsible for putting it.
302 */
303static struct bio *throtl_pop_queued(struct list_head *queued,
304 struct throtl_grp **tg_to_put)
305{
Baolin Wangb7b609d2020-10-08 11:52:26 +0800306 struct throtl_qnode *qn;
Tejun Heoc5cc2072013-05-14 13:52:38 -0700307 struct bio *bio;
308
309 if (list_empty(queued))
310 return NULL;
311
Baolin Wangb7b609d2020-10-08 11:52:26 +0800312 qn = list_first_entry(queued, struct throtl_qnode, node);
Tejun Heoc5cc2072013-05-14 13:52:38 -0700313 bio = bio_list_pop(&qn->bios);
314 WARN_ON_ONCE(!bio);
315
316 if (bio_list_empty(&qn->bios)) {
317 list_del_init(&qn->node);
318 if (tg_to_put)
319 *tg_to_put = qn->tg;
320 else
321 blkg_put(tg_to_blkg(qn->tg));
322 } else {
323 list_move_tail(&qn->node, queued);
324 }
325
326 return bio;
327}
328
Tejun Heo49a2f1e2013-05-14 13:52:34 -0700329/* init a service_queue, assumes the caller zeroed it */
Tejun Heob2ce2642015-08-18 14:55:13 -0700330static void throtl_service_queue_init(struct throtl_service_queue *sq)
Tejun Heo49a2f1e2013-05-14 13:52:34 -0700331{
Yu Kuai7e9c5c52022-08-27 18:16:35 +0800332 INIT_LIST_HEAD(&sq->queued[READ]);
333 INIT_LIST_HEAD(&sq->queued[WRITE]);
Liu Bo9ff01252018-08-21 05:21:15 +0800334 sq->pending_tree = RB_ROOT_CACHED;
Kees Cooke99e88a2017-10-16 14:43:17 -0700335 timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
Tejun Heo69df0ab2013-05-14 13:52:36 -0700336}
337
Christoph Hellwig0a0b4f72023-02-03 16:03:58 +0100338static struct blkg_policy_data *throtl_pd_alloc(struct gendisk *disk,
339 struct blkcg *blkcg, gfp_t gfp)
Tejun Heo001bea72015-08-18 14:55:11 -0700340{
Tejun Heo4fb72032015-08-18 14:55:12 -0700341 struct throtl_grp *tg;
Tejun Heo24bdb8e2015-08-18 14:55:22 -0700342 int rw;
Tejun Heo4fb72032015-08-18 14:55:12 -0700343
Christoph Hellwig0a0b4f72023-02-03 16:03:58 +0100344 tg = kzalloc_node(sizeof(*tg), gfp, disk->node_id);
Tejun Heo4fb72032015-08-18 14:55:12 -0700345 if (!tg)
Tejun Heo77ea7332015-08-18 14:55:24 -0700346 return NULL;
Tejun Heo4fb72032015-08-18 14:55:12 -0700347
Tejun Heo7ca46432019-11-07 11:18:01 -0800348 if (blkg_rwstat_init(&tg->stat_bytes, gfp))
349 goto err_free_tg;
350
351 if (blkg_rwstat_init(&tg->stat_ios, gfp))
352 goto err_exit_stat_bytes;
353
Tejun Heob2ce2642015-08-18 14:55:13 -0700354 throtl_service_queue_init(&tg->service_queue);
355
356 for (rw = READ; rw <= WRITE; rw++) {
357 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
358 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
359 }
360
361 RB_CLEAR_NODE(&tg->rb_node);
Shaohua Li9f626e32017-03-27 10:51:30 -0700362 tg->bps[READ][LIMIT_MAX] = U64_MAX;
363 tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
364 tg->iops[READ][LIMIT_MAX] = UINT_MAX;
365 tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
Shaohua Licd5ab1b2017-03-27 10:51:32 -0700366 tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
367 tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
368 tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
369 tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
370 /* LIMIT_LOW will have default value 0 */
Tejun Heob2ce2642015-08-18 14:55:13 -0700371
Shaohua Liec809912017-03-27 10:51:44 -0700372 tg->latency_target = DFL_LATENCY_TARGET;
Shaohua Li5b81fc32017-05-17 13:07:24 -0700373 tg->latency_target_conf = DFL_LATENCY_TARGET;
Shaohua Lib4f428e2017-05-17 13:07:27 -0700374 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
375 tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
Shaohua Liec809912017-03-27 10:51:44 -0700376
Tejun Heo4fb72032015-08-18 14:55:12 -0700377 return &tg->pd;
Tejun Heo7ca46432019-11-07 11:18:01 -0800378
379err_exit_stat_bytes:
380 blkg_rwstat_exit(&tg->stat_bytes);
381err_free_tg:
382 kfree(tg);
383 return NULL;
Tejun Heo001bea72015-08-18 14:55:11 -0700384}
385
Tejun Heoa9520cd2015-08-18 14:55:14 -0700386static void throtl_pd_init(struct blkg_policy_data *pd)
Vivek Goyala29a1712011-05-19 15:38:19 -0400387{
Tejun Heoa9520cd2015-08-18 14:55:14 -0700388 struct throtl_grp *tg = pd_to_tg(pd);
389 struct blkcg_gq *blkg = tg_to_blkg(tg);
Christoph Hellwiga06377c2023-02-14 19:33:08 +0100390 struct throtl_data *td = blkg->q->td;
Tejun Heob2ce2642015-08-18 14:55:13 -0700391 struct throtl_service_queue *sq = &tg->service_queue;
Tejun Heocd1604f2012-03-05 13:15:06 -0800392
Tejun Heo91381252013-05-14 13:52:38 -0700393 /*
Tejun Heoaa6ec292014-07-09 10:08:08 -0400394 * If on the default hierarchy, we switch to properly hierarchical
Tejun Heo91381252013-05-14 13:52:38 -0700395 * behavior where limits on a given throtl_grp are applied to the
396 * whole subtree rather than just the group itself. e.g. If 16M
Kemeng Shif56019a2022-12-05 19:57:01 +0800397 * read_bps limit is set on a parent group, summary bps of
398 * parent group and its subtree groups can't exceed 16M for the
399 * device.
Tejun Heo91381252013-05-14 13:52:38 -0700400 *
Tejun Heoaa6ec292014-07-09 10:08:08 -0400401 * If not on the default hierarchy, the broken flat hierarchy
Tejun Heo91381252013-05-14 13:52:38 -0700402 * behavior is retained where all throtl_grps are treated as if
403 * they're all separate root groups right below throtl_data.
404 * Limits of a group don't interact with limits of other groups
405 * regardless of the position of the group in the hierarchy.
406 */
Tejun Heob2ce2642015-08-18 14:55:13 -0700407 sq->parent_sq = &td->service_queue;
Tejun Heo9e10a132015-09-18 11:56:28 -0400408 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
Tejun Heob2ce2642015-08-18 14:55:13 -0700409 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
Tejun Heo77216b02013-05-14 13:52:36 -0700410 tg->td = td;
Tejun Heo8a3d2612012-04-01 14:38:44 -0700411}
412
Tejun Heo693e7512013-05-14 13:52:38 -0700413/*
414 * Set has_rules[] if @tg or any of its parents have limits configured.
415 * This doesn't require walking up to the top of the hierarchy as the
416 * parent's has_rules[] is guaranteed to be correct.
417 */
418static void tg_update_has_rules(struct throtl_grp *tg)
419{
420 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
Shaohua Li9f626e32017-03-27 10:51:30 -0700421 struct throtl_data *td = tg->td;
Tejun Heo693e7512013-05-14 13:52:38 -0700422 int rw;
423
Yu Kuai81c7a632022-09-21 17:53:09 +0800424 for (rw = READ; rw <= WRITE; rw++) {
425 tg->has_rules_iops[rw] =
426 (parent_tg && parent_tg->has_rules_iops[rw]) ||
Shaohua Li9f626e32017-03-27 10:51:30 -0700427 (td->limit_valid[td->limit_index] &&
Yu Kuai81c7a632022-09-21 17:53:09 +0800428 tg_iops_limit(tg, rw) != UINT_MAX);
429 tg->has_rules_bps[rw] =
430 (parent_tg && parent_tg->has_rules_bps[rw]) ||
431 (td->limit_valid[td->limit_index] &&
432 (tg_bps_limit(tg, rw) != U64_MAX));
433 }
Tejun Heo693e7512013-05-14 13:52:38 -0700434}
435
Tejun Heoa9520cd2015-08-18 14:55:14 -0700436static void throtl_pd_online(struct blkg_policy_data *pd)
Tejun Heo693e7512013-05-14 13:52:38 -0700437{
Shaohua Liaec24242017-03-27 10:51:39 -0700438 struct throtl_grp *tg = pd_to_tg(pd);
Tejun Heo693e7512013-05-14 13:52:38 -0700439 /*
440 * We don't want new groups to escape the limits of its ancestors.
441 * Update has_rules[] after a new group is brought online.
442 */
Shaohua Liaec24242017-03-27 10:51:39 -0700443 tg_update_has_rules(tg);
Tejun Heo693e7512013-05-14 13:52:38 -0700444}
445
Yu Kuaiacaf5232020-11-26 11:18:34 +0800446#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
Shaohua Licd5ab1b2017-03-27 10:51:32 -0700447static void blk_throtl_update_limit_valid(struct throtl_data *td)
448{
449 struct cgroup_subsys_state *pos_css;
450 struct blkcg_gq *blkg;
451 bool low_valid = false;
452
453 rcu_read_lock();
Christoph Hellwig12310392023-02-14 19:33:04 +0100454 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
Shaohua Licd5ab1b2017-03-27 10:51:32 -0700455 struct throtl_grp *tg = blkg_to_tg(blkg);
456
457 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
Liu Bo43ada782018-06-29 09:56:56 +0800458 tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
Shaohua Licd5ab1b2017-03-27 10:51:32 -0700459 low_valid = true;
Liu Bo43ada782018-06-29 09:56:56 +0800460 break;
461 }
Shaohua Licd5ab1b2017-03-27 10:51:32 -0700462 }
463 rcu_read_unlock();
464
465 td->limit_valid[LIMIT_LOW] = low_valid;
466}
Yu Kuaiacaf5232020-11-26 11:18:34 +0800467#else
468static inline void blk_throtl_update_limit_valid(struct throtl_data *td)
469{
470}
471#endif
Shaohua Licd5ab1b2017-03-27 10:51:32 -0700472
Shaohua Lic79892c2017-03-27 10:51:34 -0700473static void throtl_upgrade_state(struct throtl_data *td);
Shaohua Licd5ab1b2017-03-27 10:51:32 -0700474static void throtl_pd_offline(struct blkg_policy_data *pd)
475{
476 struct throtl_grp *tg = pd_to_tg(pd);
477
478 tg->bps[READ][LIMIT_LOW] = 0;
479 tg->bps[WRITE][LIMIT_LOW] = 0;
480 tg->iops[READ][LIMIT_LOW] = 0;
481 tg->iops[WRITE][LIMIT_LOW] = 0;
482
483 blk_throtl_update_limit_valid(tg->td);
484
Shaohua Lic79892c2017-03-27 10:51:34 -0700485 if (!tg->td->limit_valid[tg->td->limit_index])
486 throtl_upgrade_state(tg->td);
Shaohua Licd5ab1b2017-03-27 10:51:32 -0700487}
488
Tejun Heo001bea72015-08-18 14:55:11 -0700489static void throtl_pd_free(struct blkg_policy_data *pd)
490{
Tejun Heo4fb72032015-08-18 14:55:12 -0700491 struct throtl_grp *tg = pd_to_tg(pd);
492
Tejun Heob2ce2642015-08-18 14:55:13 -0700493 del_timer_sync(&tg->service_queue.pending_timer);
Tejun Heo7ca46432019-11-07 11:18:01 -0800494 blkg_rwstat_exit(&tg->stat_bytes);
495 blkg_rwstat_exit(&tg->stat_ios);
Tejun Heo4fb72032015-08-18 14:55:12 -0700496 kfree(tg);
Tejun Heo001bea72015-08-18 14:55:11 -0700497}
498
Tejun Heo0049af72013-05-14 13:52:33 -0700499static struct throtl_grp *
500throtl_rb_first(struct throtl_service_queue *parent_sq)
Vivek Goyale43473b2010-09-15 17:06:35 -0400501{
Liu Bo9ff01252018-08-21 05:21:15 +0800502 struct rb_node *n;
Vivek Goyale43473b2010-09-15 17:06:35 -0400503
Liu Bo9ff01252018-08-21 05:21:15 +0800504 n = rb_first_cached(&parent_sq->pending_tree);
505 WARN_ON_ONCE(!n);
506 if (!n)
507 return NULL;
508 return rb_entry_tg(n);
Vivek Goyale43473b2010-09-15 17:06:35 -0400509}
510
Tejun Heo0049af72013-05-14 13:52:33 -0700511static void throtl_rb_erase(struct rb_node *n,
512 struct throtl_service_queue *parent_sq)
Vivek Goyale43473b2010-09-15 17:06:35 -0400513{
Liu Bo9ff01252018-08-21 05:21:15 +0800514 rb_erase_cached(n, &parent_sq->pending_tree);
515 RB_CLEAR_NODE(n);
Vivek Goyale43473b2010-09-15 17:06:35 -0400516}
517
Tejun Heo0049af72013-05-14 13:52:33 -0700518static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
Vivek Goyale43473b2010-09-15 17:06:35 -0400519{
520 struct throtl_grp *tg;
521
Tejun Heo0049af72013-05-14 13:52:33 -0700522 tg = throtl_rb_first(parent_sq);
Vivek Goyale43473b2010-09-15 17:06:35 -0400523 if (!tg)
524 return;
525
Tejun Heo0049af72013-05-14 13:52:33 -0700526 parent_sq->first_pending_disptime = tg->disptime;
Vivek Goyale43473b2010-09-15 17:06:35 -0400527}
528
Tejun Heo77216b02013-05-14 13:52:36 -0700529static void tg_service_queue_add(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400530{
Tejun Heo77216b02013-05-14 13:52:36 -0700531 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
Liu Bo9ff01252018-08-21 05:21:15 +0800532 struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
Vivek Goyale43473b2010-09-15 17:06:35 -0400533 struct rb_node *parent = NULL;
534 struct throtl_grp *__tg;
535 unsigned long key = tg->disptime;
Liu Bo9ff01252018-08-21 05:21:15 +0800536 bool leftmost = true;
Vivek Goyale43473b2010-09-15 17:06:35 -0400537
538 while (*node != NULL) {
539 parent = *node;
540 __tg = rb_entry_tg(parent);
541
542 if (time_before(key, __tg->disptime))
543 node = &parent->rb_left;
544 else {
545 node = &parent->rb_right;
Liu Bo9ff01252018-08-21 05:21:15 +0800546 leftmost = false;
Vivek Goyale43473b2010-09-15 17:06:35 -0400547 }
548 }
549
Vivek Goyale43473b2010-09-15 17:06:35 -0400550 rb_link_node(&tg->rb_node, parent, node);
Liu Bo9ff01252018-08-21 05:21:15 +0800551 rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
552 leftmost);
Vivek Goyale43473b2010-09-15 17:06:35 -0400553}
554
Tejun Heo77216b02013-05-14 13:52:36 -0700555static void throtl_enqueue_tg(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400556{
Baolin Wang29379672020-10-08 11:52:28 +0800557 if (!(tg->flags & THROTL_TG_PENDING)) {
558 tg_service_queue_add(tg);
559 tg->flags |= THROTL_TG_PENDING;
560 tg->service_queue.parent_sq->nr_pending++;
561 }
Vivek Goyale43473b2010-09-15 17:06:35 -0400562}
563
Tejun Heo77216b02013-05-14 13:52:36 -0700564static void throtl_dequeue_tg(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400565{
Baolin Wang29379672020-10-08 11:52:28 +0800566 if (tg->flags & THROTL_TG_PENDING) {
Yu Kuaic0137102022-08-27 18:16:37 +0800567 struct throtl_service_queue *parent_sq =
568 tg->service_queue.parent_sq;
569
570 throtl_rb_erase(&tg->rb_node, parent_sq);
571 --parent_sq->nr_pending;
Baolin Wang29379672020-10-08 11:52:28 +0800572 tg->flags &= ~THROTL_TG_PENDING;
573 }
Vivek Goyale43473b2010-09-15 17:06:35 -0400574}
575
Tejun Heoa9131a22013-05-14 13:52:31 -0700576/* Call with queue lock held */
Tejun Heo69df0ab2013-05-14 13:52:36 -0700577static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
578 unsigned long expires)
Tejun Heoa9131a22013-05-14 13:52:31 -0700579{
Joseph Qia41b8162017-06-07 11:36:14 +0800580 unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
Shaohua Li06cceed2017-03-27 10:51:36 -0700581
582 /*
583 * Since we are adjusting the throttle limit dynamically, the sleep
584 * time calculated according to previous limit might be invalid. It's
585 * possible the cgroup sleep time is very long and no other cgroups
586 * have IO running so notify the limit changes. Make sure the cgroup
587 * doesn't sleep too long to avoid the missed notification.
588 */
589 if (time_after(expires, max_expire))
590 expires = max_expire;
Tejun Heo69df0ab2013-05-14 13:52:36 -0700591 mod_timer(&sq->pending_timer, expires);
592 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
593 expires - jiffies, jiffies);
Tejun Heoa9131a22013-05-14 13:52:31 -0700594}
595
Tejun Heo7f52f982013-05-14 13:52:37 -0700596/**
597 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
598 * @sq: the service_queue to schedule dispatch for
599 * @force: force scheduling
600 *
601 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
602 * dispatch time of the first pending child. Returns %true if either timer
603 * is armed or there's no pending child left. %false if the current
604 * dispatch window is still open and the caller should continue
605 * dispatching.
606 *
607 * If @force is %true, the dispatch timer is always scheduled and this
608 * function is guaranteed to return %true. This is to be used when the
609 * caller can't dispatch itself and needs to invoke pending_timer
610 * unconditionally. Note that forced scheduling is likely to induce short
611 * delay before dispatch starts even if @sq->first_pending_disptime is not
612 * in the future and thus shouldn't be used in hot paths.
613 */
614static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
615 bool force)
Vivek Goyale43473b2010-09-15 17:06:35 -0400616{
Tejun Heo6a525602013-05-14 13:52:32 -0700617 /* any pending children left? */
Tejun Heoc9e03322013-05-14 13:52:32 -0700618 if (!sq->nr_pending)
Tejun Heo7f52f982013-05-14 13:52:37 -0700619 return true;
Vivek Goyale43473b2010-09-15 17:06:35 -0400620
Tejun Heoc9e03322013-05-14 13:52:32 -0700621 update_min_dispatch_time(sq);
Vivek Goyale43473b2010-09-15 17:06:35 -0400622
Tejun Heo69df0ab2013-05-14 13:52:36 -0700623 /* is the next dispatch time in the future? */
Tejun Heo7f52f982013-05-14 13:52:37 -0700624 if (force || time_after(sq->first_pending_disptime, jiffies)) {
Tejun Heo69df0ab2013-05-14 13:52:36 -0700625 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
Tejun Heo7f52f982013-05-14 13:52:37 -0700626 return true;
Tejun Heo69df0ab2013-05-14 13:52:36 -0700627 }
628
Tejun Heo7f52f982013-05-14 13:52:37 -0700629 /* tell the caller to continue dispatching */
630 return false;
Vivek Goyale43473b2010-09-15 17:06:35 -0400631}
632
Vivek Goyal32ee5bc2013-05-14 13:52:38 -0700633static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
634 bool rw, unsigned long start)
635{
636 tg->bytes_disp[rw] = 0;
637 tg->io_disp[rw] = 0;
Yu Kuaia880ae92022-08-29 10:22:40 +0800638 tg->carryover_bytes[rw] = 0;
639 tg->carryover_ios[rw] = 0;
Vivek Goyal32ee5bc2013-05-14 13:52:38 -0700640
641 /*
642 * Previous slice has expired. We must have trimmed it after last
643 * bio dispatch. That means since start of last slice, we never used
644 * that bandwidth. Do try to make use of that bandwidth while giving
645 * credit.
646 */
Kemeng Shieea3e8b2022-12-05 19:57:09 +0800647 if (time_after(start, tg->slice_start[rw]))
Vivek Goyal32ee5bc2013-05-14 13:52:38 -0700648 tg->slice_start[rw] = start;
649
Shaohua Li297e3d82017-03-27 10:51:37 -0700650 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
Vivek Goyal32ee5bc2013-05-14 13:52:38 -0700651 throtl_log(&tg->service_queue,
652 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
653 rw == READ ? 'R' : 'W', tg->slice_start[rw],
654 tg->slice_end[rw], jiffies);
655}
656
Yu Kuaia880ae92022-08-29 10:22:40 +0800657static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw,
658 bool clear_carryover)
Vivek Goyale43473b2010-09-15 17:06:35 -0400659{
660 tg->bytes_disp[rw] = 0;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400661 tg->io_disp[rw] = 0;
Vivek Goyale43473b2010-09-15 17:06:35 -0400662 tg->slice_start[rw] = jiffies;
Shaohua Li297e3d82017-03-27 10:51:37 -0700663 tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
Yu Kuaia880ae92022-08-29 10:22:40 +0800664 if (clear_carryover) {
665 tg->carryover_bytes[rw] = 0;
666 tg->carryover_ios[rw] = 0;
667 }
Chunguang Xu4f1e9632021-08-02 11:51:56 +0800668
Tejun Heofda6f272013-05-14 13:52:36 -0700669 throtl_log(&tg->service_queue,
670 "[%c] new slice start=%lu end=%lu jiffies=%lu",
671 rw == READ ? 'R' : 'W', tg->slice_start[rw],
672 tg->slice_end[rw], jiffies);
Vivek Goyale43473b2010-09-15 17:06:35 -0400673}
674
Tejun Heo0f3457f2013-05-14 13:52:32 -0700675static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
676 unsigned long jiffy_end)
Vivek Goyald1ae8ff2010-12-01 19:34:46 +0100677{
Shaohua Li297e3d82017-03-27 10:51:37 -0700678 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
Vivek Goyald1ae8ff2010-12-01 19:34:46 +0100679}
680
Tejun Heo0f3457f2013-05-14 13:52:32 -0700681static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
682 unsigned long jiffy_end)
Vivek Goyale43473b2010-09-15 17:06:35 -0400683{
Baolin Wang1da30f92020-10-08 11:52:29 +0800684 throtl_set_slice_end(tg, rw, jiffy_end);
Tejun Heofda6f272013-05-14 13:52:36 -0700685 throtl_log(&tg->service_queue,
686 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
687 rw == READ ? 'R' : 'W', tg->slice_start[rw],
688 tg->slice_end[rw], jiffies);
Vivek Goyale43473b2010-09-15 17:06:35 -0400689}
690
691/* Determine if previously allocated or extended slice is complete or not */
Tejun Heo0f3457f2013-05-14 13:52:32 -0700692static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
Vivek Goyale43473b2010-09-15 17:06:35 -0400693{
694 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
Fabian Frederick5cf8c222014-05-02 18:28:17 +0200695 return false;
Vivek Goyale43473b2010-09-15 17:06:35 -0400696
Chengguang Xu0b6bad72018-05-29 18:32:44 +0800697 return true;
Vivek Goyale43473b2010-09-15 17:06:35 -0400698}
699
Yu Kuai681cd462022-08-29 10:22:39 +0800700static unsigned int calculate_io_allowed(u32 iops_limit,
701 unsigned long jiffy_elapsed)
702{
703 unsigned int io_allowed;
704 u64 tmp;
705
706 /*
707 * jiffy_elapsed should not be a big value as minimum iops can be
708 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
709 * will allow dispatch after 1 second and after that slice should
710 * have been trimmed.
711 */
712
713 tmp = (u64)iops_limit * jiffy_elapsed;
714 do_div(tmp, HZ);
715
716 if (tmp > UINT_MAX)
717 io_allowed = UINT_MAX;
718 else
719 io_allowed = tmp;
720
721 return io_allowed;
722}
723
724static u64 calculate_bytes_allowed(u64 bps_limit, unsigned long jiffy_elapsed)
725{
Khazhismel Kumykov2dd710d2023-10-20 15:36:17 -0700726 /*
727 * Can result be wider than 64 bits?
728 * We check against 62, not 64, due to ilog2 truncation.
729 */
730 if (ilog2(bps_limit) + ilog2(jiffy_elapsed) - ilog2(HZ) > 62)
731 return U64_MAX;
Yu Kuai681cd462022-08-29 10:22:39 +0800732 return mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed, (u64)HZ);
733}
734
Yu Kuaie8368b52023-08-16 09:27:07 +0800735/* Trim the used slices and adjust slice start accordingly */
736static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
737{
Yu Kuaieead0052023-08-16 09:27:08 +0800738 unsigned long time_elapsed;
739 long long bytes_trim;
740 int io_trim;
Yu Kuaie8368b52023-08-16 09:27:07 +0800741
742 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
743
744 /*
745 * If bps are unlimited (-1), then time slice don't get
746 * renewed. Don't try to trim the slice if slice is used. A new
747 * slice will start when appropriate.
748 */
749 if (throtl_slice_used(tg, rw))
750 return;
751
752 /*
753 * A bio has been dispatched. Also adjust slice_end. It might happen
754 * that initially cgroup limit was very low resulting in high
755 * slice_end, but later limit was bumped up and bio was dispatched
756 * sooner, then we need to reduce slice_end. A high bogus slice_end
757 * is bad because it does not allow new slice to start.
758 */
759
760 throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
761
762 time_elapsed = rounddown(jiffies - tg->slice_start[rw],
763 tg->td->throtl_slice);
764 if (!time_elapsed)
765 return;
766
767 bytes_trim = calculate_bytes_allowed(tg_bps_limit(tg, rw),
Yu Kuaieead0052023-08-16 09:27:08 +0800768 time_elapsed) +
769 tg->carryover_bytes[rw];
770 io_trim = calculate_io_allowed(tg_iops_limit(tg, rw), time_elapsed) +
771 tg->carryover_ios[rw];
772 if (bytes_trim <= 0 && io_trim <= 0)
Yu Kuaie8368b52023-08-16 09:27:07 +0800773 return;
774
Yu Kuaieead0052023-08-16 09:27:08 +0800775 tg->carryover_bytes[rw] = 0;
776 if ((long long)tg->bytes_disp[rw] >= bytes_trim)
Yu Kuaie8368b52023-08-16 09:27:07 +0800777 tg->bytes_disp[rw] -= bytes_trim;
778 else
779 tg->bytes_disp[rw] = 0;
780
Yu Kuaieead0052023-08-16 09:27:08 +0800781 tg->carryover_ios[rw] = 0;
782 if ((int)tg->io_disp[rw] >= io_trim)
Yu Kuaie8368b52023-08-16 09:27:07 +0800783 tg->io_disp[rw] -= io_trim;
784 else
785 tg->io_disp[rw] = 0;
786
787 tg->slice_start[rw] += time_elapsed;
788
789 throtl_log(&tg->service_queue,
Yu Kuaieead0052023-08-16 09:27:08 +0800790 "[%c] trim slice nr=%lu bytes=%lld io=%d start=%lu end=%lu jiffies=%lu",
Yu Kuaie8368b52023-08-16 09:27:07 +0800791 rw == READ ? 'R' : 'W', time_elapsed / tg->td->throtl_slice,
792 bytes_trim, io_trim, tg->slice_start[rw], tg->slice_end[rw],
793 jiffies);
794}
795
Yu Kuaia880ae92022-08-29 10:22:40 +0800796static void __tg_update_carryover(struct throtl_grp *tg, bool rw)
797{
798 unsigned long jiffy_elapsed = jiffies - tg->slice_start[rw];
799 u64 bps_limit = tg_bps_limit(tg, rw);
800 u32 iops_limit = tg_iops_limit(tg, rw);
801
802 /*
803 * If config is updated while bios are still throttled, calculate and
804 * accumulate how many bytes/ios are waited across changes. And
805 * carryover_bytes/ios will be used to calculate new wait time under new
806 * configuration.
807 */
808 if (bps_limit != U64_MAX)
809 tg->carryover_bytes[rw] +=
810 calculate_bytes_allowed(bps_limit, jiffy_elapsed) -
811 tg->bytes_disp[rw];
812 if (iops_limit != UINT_MAX)
813 tg->carryover_ios[rw] +=
814 calculate_io_allowed(iops_limit, jiffy_elapsed) -
815 tg->io_disp[rw];
816}
817
818static void tg_update_carryover(struct throtl_grp *tg)
819{
820 if (tg->service_queue.nr_queued[READ])
821 __tg_update_carryover(tg, READ);
822 if (tg->service_queue.nr_queued[WRITE])
823 __tg_update_carryover(tg, WRITE);
824
825 /* see comments in struct throtl_grp for meaning of these fields. */
Yu Kuaief100392023-08-16 09:27:05 +0800826 throtl_log(&tg->service_queue, "%s: %lld %lld %d %d\n", __func__,
Yu Kuaia880ae92022-08-29 10:22:40 +0800827 tg->carryover_bytes[READ], tg->carryover_bytes[WRITE],
828 tg->carryover_ios[READ], tg->carryover_ios[WRITE]);
829}
830
Kemeng Shi183daeb2022-12-05 19:57:04 +0800831static unsigned long tg_within_iops_limit(struct throtl_grp *tg, struct bio *bio,
832 u32 iops_limit)
Vivek Goyale43473b2010-09-15 17:06:35 -0400833{
834 bool rw = bio_data_dir(bio);
Yu Kuaibb8d5582023-08-16 09:27:06 +0800835 int io_allowed;
Vivek Goyale43473b2010-09-15 17:06:35 -0400836 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
837
Baolin Wang87fbeb82020-09-07 16:10:17 +0800838 if (iops_limit == UINT_MAX) {
Kemeng Shi183daeb2022-12-05 19:57:04 +0800839 return 0;
Baolin Wang87fbeb82020-09-07 16:10:17 +0800840 }
841
Konstantin Khlebnikov3a10f992019-07-08 18:29:57 +0300842 jiffy_elapsed = jiffies - tg->slice_start[rw];
Vivek Goyale43473b2010-09-15 17:06:35 -0400843
Konstantin Khlebnikov3a10f992019-07-08 18:29:57 +0300844 /* Round up to the next throttle slice, wait time must be nonzero */
845 jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
Yu Kuaia880ae92022-08-29 10:22:40 +0800846 io_allowed = calculate_io_allowed(iops_limit, jiffy_elapsed_rnd) +
847 tg->carryover_ios[rw];
Yu Kuaibb8d5582023-08-16 09:27:06 +0800848 if (io_allowed > 0 && tg->io_disp[rw] + 1 <= io_allowed)
Kemeng Shi183daeb2022-12-05 19:57:04 +0800849 return 0;
Vivek Goyale43473b2010-09-15 17:06:35 -0400850
Vivek Goyal8e89d132010-09-15 17:06:37 -0400851 /* Calc approx time to dispatch */
Liu Bo991f61f2018-08-10 01:47:02 +0800852 jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
Kemeng Shi183daeb2022-12-05 19:57:04 +0800853 return jiffy_wait;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400854}
855
Kemeng Shi183daeb2022-12-05 19:57:04 +0800856static unsigned long tg_within_bps_limit(struct throtl_grp *tg, struct bio *bio,
857 u64 bps_limit)
Vivek Goyal8e89d132010-09-15 17:06:37 -0400858{
859 bool rw = bio_data_dir(bio);
Yu Kuaibb8d5582023-08-16 09:27:06 +0800860 long long bytes_allowed;
861 u64 extra_bytes;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400862 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
Shaohua Liea0ea2b2017-08-18 16:08:13 -0700863 unsigned int bio_size = throtl_bio_data_size(bio);
Vivek Goyale43473b2010-09-15 17:06:35 -0400864
Ming Lei9f5ede32022-02-16 12:45:12 +0800865 /* no need to throttle if this bio's bytes have been accounted */
Yu Kuai320fb0f2022-08-29 10:22:37 +0800866 if (bps_limit == U64_MAX || bio_flagged(bio, BIO_BPS_THROTTLED)) {
Kemeng Shi183daeb2022-12-05 19:57:04 +0800867 return 0;
Baolin Wang87fbeb82020-09-07 16:10:17 +0800868 }
869
Vivek Goyale43473b2010-09-15 17:06:35 -0400870 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
871
872 /* Slice has just started. Consider one slice interval */
873 if (!jiffy_elapsed)
Shaohua Li297e3d82017-03-27 10:51:37 -0700874 jiffy_elapsed_rnd = tg->td->throtl_slice;
Vivek Goyale43473b2010-09-15 17:06:35 -0400875
Shaohua Li297e3d82017-03-27 10:51:37 -0700876 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
Yu Kuaia880ae92022-08-29 10:22:40 +0800877 bytes_allowed = calculate_bytes_allowed(bps_limit, jiffy_elapsed_rnd) +
878 tg->carryover_bytes[rw];
Yu Kuaibb8d5582023-08-16 09:27:06 +0800879 if (bytes_allowed > 0 && tg->bytes_disp[rw] + bio_size <= bytes_allowed)
Kemeng Shi183daeb2022-12-05 19:57:04 +0800880 return 0;
Vivek Goyale43473b2010-09-15 17:06:35 -0400881
882 /* Calc approx time to dispatch */
Shaohua Liea0ea2b2017-08-18 16:08:13 -0700883 extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
Baolin Wang4599ea42020-09-07 16:10:16 +0800884 jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
Vivek Goyale43473b2010-09-15 17:06:35 -0400885
886 if (!jiffy_wait)
887 jiffy_wait = 1;
888
889 /*
890 * This wait time is without taking into consideration the rounding
891 * up we did. Add that time also.
892 */
893 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
Kemeng Shi183daeb2022-12-05 19:57:04 +0800894 return jiffy_wait;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400895}
Vivek Goyale43473b2010-09-15 17:06:35 -0400896
Vivek Goyal8e89d132010-09-15 17:06:37 -0400897/*
898 * Returns whether one can dispatch a bio or not. Also returns approx number
899 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
900 */
Tejun Heo0f3457f2013-05-14 13:52:32 -0700901static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
902 unsigned long *wait)
Vivek Goyal8e89d132010-09-15 17:06:37 -0400903{
904 bool rw = bio_data_dir(bio);
905 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
Baolin Wang4599ea42020-09-07 16:10:16 +0800906 u64 bps_limit = tg_bps_limit(tg, rw);
907 u32 iops_limit = tg_iops_limit(tg, rw);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400908
909 /*
910 * Currently whole state machine of group depends on first bio
911 * queued in the group bio list. So one should not be calling
912 * this function with a different bio if there are other bios
913 * queued.
914 */
Tejun Heo73f0d492013-05-14 13:52:35 -0700915 BUG_ON(tg->service_queue.nr_queued[rw] &&
Tejun Heoc5cc2072013-05-14 13:52:38 -0700916 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
Vivek Goyal8e89d132010-09-15 17:06:37 -0400917
918 /* If tg->bps = -1, then BW is unlimited */
Yu Kuai8f9e7b652022-03-18 21:01:44 +0800919 if ((bps_limit == U64_MAX && iops_limit == UINT_MAX) ||
920 tg->flags & THROTL_TG_CANCELING) {
Vivek Goyal8e89d132010-09-15 17:06:37 -0400921 if (wait)
922 *wait = 0;
Fabian Frederick5cf8c222014-05-02 18:28:17 +0200923 return true;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400924 }
925
926 /*
927 * If previous slice expired, start a new one otherwise renew/extend
928 * existing slice to make sure it is at least throtl_slice interval
Vivek Goyal164c80e2016-09-19 15:12:41 -0600929 * long since now. New slice is started only for empty throttle group.
930 * If there is queued bio, that means there should be an active
931 * slice and it should be extended instead.
Vivek Goyal8e89d132010-09-15 17:06:37 -0400932 */
Vivek Goyal164c80e2016-09-19 15:12:41 -0600933 if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
Yu Kuaia880ae92022-08-29 10:22:40 +0800934 throtl_start_new_slice(tg, rw, true);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400935 else {
Shaohua Li297e3d82017-03-27 10:51:37 -0700936 if (time_before(tg->slice_end[rw],
937 jiffies + tg->td->throtl_slice))
938 throtl_extend_slice(tg, rw,
939 jiffies + tg->td->throtl_slice);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400940 }
941
Kemeng Shi183daeb2022-12-05 19:57:04 +0800942 bps_wait = tg_within_bps_limit(tg, bio, bps_limit);
943 iops_wait = tg_within_iops_limit(tg, bio, iops_limit);
944 if (bps_wait + iops_wait == 0) {
Vivek Goyal8e89d132010-09-15 17:06:37 -0400945 if (wait)
946 *wait = 0;
Chengguang Xu0b6bad72018-05-29 18:32:44 +0800947 return true;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400948 }
949
950 max_wait = max(bps_wait, iops_wait);
951
952 if (wait)
953 *wait = max_wait;
954
955 if (time_before(tg->slice_end[rw], jiffies + max_wait))
Tejun Heo0f3457f2013-05-14 13:52:32 -0700956 throtl_extend_slice(tg, rw, jiffies + max_wait);
Vivek Goyale43473b2010-09-15 17:06:35 -0400957
Chengguang Xu0b6bad72018-05-29 18:32:44 +0800958 return false;
Vivek Goyale43473b2010-09-15 17:06:35 -0400959}
960
961static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
962{
963 bool rw = bio_data_dir(bio);
Shaohua Liea0ea2b2017-08-18 16:08:13 -0700964 unsigned int bio_size = throtl_bio_data_size(bio);
Vivek Goyale43473b2010-09-15 17:06:35 -0400965
966 /* Charge the bio to the group */
Yu Kuai320fb0f2022-08-29 10:22:37 +0800967 if (!bio_flagged(bio, BIO_BPS_THROTTLED)) {
Ming Lei9f5ede32022-02-16 12:45:12 +0800968 tg->bytes_disp[rw] += bio_size;
969 tg->last_bytes_disp[rw] += bio_size;
970 }
971
Vivek Goyal8e89d132010-09-15 17:06:37 -0400972 tg->io_disp[rw]++;
Shaohua Li3f0abd82017-03-27 10:51:35 -0700973 tg->last_io_disp[rw]++;
Vivek Goyale43473b2010-09-15 17:06:35 -0400974}
975
Tejun Heoc5cc2072013-05-14 13:52:38 -0700976/**
977 * throtl_add_bio_tg - add a bio to the specified throtl_grp
978 * @bio: bio to add
979 * @qn: qnode to use
980 * @tg: the target throtl_grp
981 *
982 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
983 * tg->qnode_on_self[] is used.
984 */
985static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
986 struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400987{
Tejun Heo73f0d492013-05-14 13:52:35 -0700988 struct throtl_service_queue *sq = &tg->service_queue;
Vivek Goyale43473b2010-09-15 17:06:35 -0400989 bool rw = bio_data_dir(bio);
990
Tejun Heoc5cc2072013-05-14 13:52:38 -0700991 if (!qn)
992 qn = &tg->qnode_on_self[rw];
993
Tejun Heo0e9f4162013-05-14 13:52:35 -0700994 /*
995 * If @tg doesn't currently have any bios queued in the same
996 * direction, queueing @bio can change when @tg should be
997 * dispatched. Mark that @tg was empty. This is automatically
Baolin Wangb53b0722020-09-07 16:10:13 +0800998 * cleared on the next tg_update_disptime().
Tejun Heo0e9f4162013-05-14 13:52:35 -0700999 */
1000 if (!sq->nr_queued[rw])
1001 tg->flags |= THROTL_TG_WAS_EMPTY;
1002
Tejun Heoc5cc2072013-05-14 13:52:38 -07001003 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1004
Tejun Heo73f0d492013-05-14 13:52:35 -07001005 sq->nr_queued[rw]++;
Tejun Heo77216b02013-05-14 13:52:36 -07001006 throtl_enqueue_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -04001007}
1008
Tejun Heo77216b02013-05-14 13:52:36 -07001009static void tg_update_disptime(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -04001010{
Tejun Heo73f0d492013-05-14 13:52:35 -07001011 struct throtl_service_queue *sq = &tg->service_queue;
Vivek Goyale43473b2010-09-15 17:06:35 -04001012 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1013 struct bio *bio;
1014
Markus Elfringd609af32017-01-21 22:15:33 +01001015 bio = throtl_peek_queued(&sq->queued[READ]);
1016 if (bio)
Tejun Heo0f3457f2013-05-14 13:52:32 -07001017 tg_may_dispatch(tg, bio, &read_wait);
Vivek Goyale43473b2010-09-15 17:06:35 -04001018
Markus Elfringd609af32017-01-21 22:15:33 +01001019 bio = throtl_peek_queued(&sq->queued[WRITE]);
1020 if (bio)
Tejun Heo0f3457f2013-05-14 13:52:32 -07001021 tg_may_dispatch(tg, bio, &write_wait);
Vivek Goyale43473b2010-09-15 17:06:35 -04001022
1023 min_wait = min(read_wait, write_wait);
1024 disptime = jiffies + min_wait;
1025
Vivek Goyale43473b2010-09-15 17:06:35 -04001026 /* Update dispatch time */
Yu Kuaic0137102022-08-27 18:16:37 +08001027 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
Vivek Goyale43473b2010-09-15 17:06:35 -04001028 tg->disptime = disptime;
Yu Kuaic0137102022-08-27 18:16:37 +08001029 tg_service_queue_add(tg);
Tejun Heo0e9f4162013-05-14 13:52:35 -07001030
1031 /* see throtl_add_bio_tg() */
1032 tg->flags &= ~THROTL_TG_WAS_EMPTY;
Vivek Goyale43473b2010-09-15 17:06:35 -04001033}
1034
Vivek Goyal32ee5bc2013-05-14 13:52:38 -07001035static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1036 struct throtl_grp *parent_tg, bool rw)
1037{
1038 if (throtl_slice_used(parent_tg, rw)) {
1039 throtl_start_new_slice_with_credit(parent_tg, rw,
1040 child_tg->slice_start[rw]);
1041 }
1042
1043}
1044
Tejun Heo77216b02013-05-14 13:52:36 -07001045static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
Vivek Goyale43473b2010-09-15 17:06:35 -04001046{
Tejun Heo73f0d492013-05-14 13:52:35 -07001047 struct throtl_service_queue *sq = &tg->service_queue;
Tejun Heo6bc9c2b2013-05-14 13:52:38 -07001048 struct throtl_service_queue *parent_sq = sq->parent_sq;
1049 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
Tejun Heoc5cc2072013-05-14 13:52:38 -07001050 struct throtl_grp *tg_to_put = NULL;
Vivek Goyale43473b2010-09-15 17:06:35 -04001051 struct bio *bio;
1052
Tejun Heoc5cc2072013-05-14 13:52:38 -07001053 /*
1054 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1055 * from @tg may put its reference and @parent_sq might end up
1056 * getting released prematurely. Remember the tg to put and put it
1057 * after @bio is transferred to @parent_sq.
1058 */
1059 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
Tejun Heo73f0d492013-05-14 13:52:35 -07001060 sq->nr_queued[rw]--;
Vivek Goyale43473b2010-09-15 17:06:35 -04001061
1062 throtl_charge_bio(tg, bio);
Tejun Heo6bc9c2b2013-05-14 13:52:38 -07001063
1064 /*
1065 * If our parent is another tg, we just need to transfer @bio to
1066 * the parent using throtl_add_bio_tg(). If our parent is
1067 * @td->service_queue, @bio is ready to be issued. Put it on its
1068 * bio_lists[] and decrease total number queued. The caller is
1069 * responsible for issuing these bios.
1070 */
1071 if (parent_tg) {
Tejun Heoc5cc2072013-05-14 13:52:38 -07001072 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
Vivek Goyal32ee5bc2013-05-14 13:52:38 -07001073 start_parent_slice_with_credit(tg, parent_tg, rw);
Tejun Heo6bc9c2b2013-05-14 13:52:38 -07001074 } else {
Kemeng Shi84aca0a2022-12-05 19:57:02 +08001075 bio_set_flag(bio, BIO_BPS_THROTTLED);
Tejun Heoc5cc2072013-05-14 13:52:38 -07001076 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1077 &parent_sq->queued[rw]);
Tejun Heo6bc9c2b2013-05-14 13:52:38 -07001078 BUG_ON(tg->td->nr_queued[rw] <= 0);
1079 tg->td->nr_queued[rw]--;
1080 }
Vivek Goyale43473b2010-09-15 17:06:35 -04001081
Tejun Heo0f3457f2013-05-14 13:52:32 -07001082 throtl_trim_slice(tg, rw);
Tejun Heo6bc9c2b2013-05-14 13:52:38 -07001083
Tejun Heoc5cc2072013-05-14 13:52:38 -07001084 if (tg_to_put)
1085 blkg_put(tg_to_blkg(tg_to_put));
Vivek Goyale43473b2010-09-15 17:06:35 -04001086}
1087
Tejun Heo77216b02013-05-14 13:52:36 -07001088static int throtl_dispatch_tg(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -04001089{
Tejun Heo73f0d492013-05-14 13:52:35 -07001090 struct throtl_service_queue *sq = &tg->service_queue;
Vivek Goyale43473b2010-09-15 17:06:35 -04001091 unsigned int nr_reads = 0, nr_writes = 0;
Baolin Wange675df22020-09-07 16:10:15 +08001092 unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1093 unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
Vivek Goyale43473b2010-09-15 17:06:35 -04001094 struct bio *bio;
1095
1096 /* Try to dispatch 75% READS and 25% WRITES */
1097
Tejun Heoc5cc2072013-05-14 13:52:38 -07001098 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
Tejun Heo0f3457f2013-05-14 13:52:32 -07001099 tg_may_dispatch(tg, bio, NULL)) {
Vivek Goyale43473b2010-09-15 17:06:35 -04001100
Tejun Heo77216b02013-05-14 13:52:36 -07001101 tg_dispatch_one_bio(tg, bio_data_dir(bio));
Vivek Goyale43473b2010-09-15 17:06:35 -04001102 nr_reads++;
1103
1104 if (nr_reads >= max_nr_reads)
1105 break;
1106 }
1107
Tejun Heoc5cc2072013-05-14 13:52:38 -07001108 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
Tejun Heo0f3457f2013-05-14 13:52:32 -07001109 tg_may_dispatch(tg, bio, NULL)) {
Vivek Goyale43473b2010-09-15 17:06:35 -04001110
Tejun Heo77216b02013-05-14 13:52:36 -07001111 tg_dispatch_one_bio(tg, bio_data_dir(bio));
Vivek Goyale43473b2010-09-15 17:06:35 -04001112 nr_writes++;
1113
1114 if (nr_writes >= max_nr_writes)
1115 break;
1116 }
1117
1118 return nr_reads + nr_writes;
1119}
1120
Tejun Heo651930b2013-05-14 13:52:35 -07001121static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
Vivek Goyale43473b2010-09-15 17:06:35 -04001122{
1123 unsigned int nr_disp = 0;
Vivek Goyale43473b2010-09-15 17:06:35 -04001124
1125 while (1) {
Baolin Wang23976112020-10-08 11:52:27 +08001126 struct throtl_grp *tg;
Liu Bo2ab74cd2018-05-29 16:29:12 +08001127 struct throtl_service_queue *sq;
Vivek Goyale43473b2010-09-15 17:06:35 -04001128
Baolin Wang23976112020-10-08 11:52:27 +08001129 if (!parent_sq->nr_pending)
1130 break;
1131
1132 tg = throtl_rb_first(parent_sq);
Vivek Goyale43473b2010-09-15 17:06:35 -04001133 if (!tg)
1134 break;
1135
1136 if (time_before(jiffies, tg->disptime))
1137 break;
1138
Tejun Heo77216b02013-05-14 13:52:36 -07001139 nr_disp += throtl_dispatch_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -04001140
Liu Bo2ab74cd2018-05-29 16:29:12 +08001141 sq = &tg->service_queue;
Yu Kuai7e9c5c52022-08-27 18:16:35 +08001142 if (sq->nr_queued[READ] || sq->nr_queued[WRITE])
Tejun Heo77216b02013-05-14 13:52:36 -07001143 tg_update_disptime(tg);
Yu Kuai8c25ed02022-08-27 18:16:36 +08001144 else
1145 throtl_dequeue_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -04001146
Baolin Wange675df22020-09-07 16:10:15 +08001147 if (nr_disp >= THROTL_QUANTUM)
Vivek Goyale43473b2010-09-15 17:06:35 -04001148 break;
1149 }
1150
1151 return nr_disp;
1152}
1153
Shaohua Lic79892c2017-03-27 10:51:34 -07001154static bool throtl_can_upgrade(struct throtl_data *td,
1155 struct throtl_grp *this_tg);
Tejun Heo6e1a5702013-05-14 13:52:37 -07001156/**
1157 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
Bart Van Assche216382d2019-05-30 17:00:48 -07001158 * @t: the pending_timer member of the throtl_service_queue being serviced
Tejun Heo6e1a5702013-05-14 13:52:37 -07001159 *
1160 * This timer is armed when a child throtl_grp with active bio's become
1161 * pending and queued on the service_queue's pending_tree and expires when
1162 * the first child throtl_grp should be dispatched. This function
Tejun Heo2e48a532013-05-14 13:52:38 -07001163 * dispatches bio's from the children throtl_grps to the parent
1164 * service_queue.
1165 *
1166 * If the parent's parent is another throtl_grp, dispatching is propagated
1167 * by either arming its pending_timer or repeating dispatch directly. If
1168 * the top-level service_tree is reached, throtl_data->dispatch_work is
1169 * kicked so that the ready bio's are issued.
Tejun Heo6e1a5702013-05-14 13:52:37 -07001170 */
Kees Cooke99e88a2017-10-16 14:43:17 -07001171static void throtl_pending_timer_fn(struct timer_list *t)
Tejun Heo69df0ab2013-05-14 13:52:36 -07001172{
Kees Cooke99e88a2017-10-16 14:43:17 -07001173 struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
Tejun Heo2e48a532013-05-14 13:52:38 -07001174 struct throtl_grp *tg = sq_to_tg(sq);
Tejun Heo69df0ab2013-05-14 13:52:36 -07001175 struct throtl_data *td = sq_to_td(sq);
Tejun Heo2e48a532013-05-14 13:52:38 -07001176 struct throtl_service_queue *parent_sq;
Ming Leiee37edd2022-03-18 21:01:42 +08001177 struct request_queue *q;
Tejun Heo2e48a532013-05-14 13:52:38 -07001178 bool dispatched;
Tejun Heo6e1a5702013-05-14 13:52:37 -07001179 int ret;
Vivek Goyale43473b2010-09-15 17:06:35 -04001180
Ming Leiee37edd2022-03-18 21:01:42 +08001181 /* throtl_data may be gone, so figure out request queue by blkg */
1182 if (tg)
Christoph Hellwiga06377c2023-02-14 19:33:08 +01001183 q = tg->pd.blkg->q;
Ming Leiee37edd2022-03-18 21:01:42 +08001184 else
1185 q = td->queue;
1186
Christoph Hellwig0d945c12018-11-15 12:17:28 -07001187 spin_lock_irq(&q->queue_lock);
Ming Leiee37edd2022-03-18 21:01:42 +08001188
Christoph Hellwig12310392023-02-14 19:33:04 +01001189 if (!q->root_blkg)
Ming Leiee37edd2022-03-18 21:01:42 +08001190 goto out_unlock;
1191
Shaohua Lic79892c2017-03-27 10:51:34 -07001192 if (throtl_can_upgrade(td, NULL))
1193 throtl_upgrade_state(td);
1194
Tejun Heo2e48a532013-05-14 13:52:38 -07001195again:
1196 parent_sq = sq->parent_sq;
1197 dispatched = false;
Vivek Goyale43473b2010-09-15 17:06:35 -04001198
Tejun Heo7f52f982013-05-14 13:52:37 -07001199 while (true) {
1200 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
Tejun Heo2e48a532013-05-14 13:52:38 -07001201 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1202 sq->nr_queued[READ], sq->nr_queued[WRITE]);
Vivek Goyale43473b2010-09-15 17:06:35 -04001203
Tejun Heo7f52f982013-05-14 13:52:37 -07001204 ret = throtl_select_dispatch(sq);
1205 if (ret) {
Tejun Heo7f52f982013-05-14 13:52:37 -07001206 throtl_log(sq, "bios disp=%u", ret);
1207 dispatched = true;
Tejun Heo651930b2013-05-14 13:52:35 -07001208 }
Vivek Goyale43473b2010-09-15 17:06:35 -04001209
Tejun Heo7f52f982013-05-14 13:52:37 -07001210 if (throtl_schedule_next_dispatch(sq, false))
1211 break;
1212
1213 /* this dispatch windows is still open, relax and repeat */
Christoph Hellwig0d945c12018-11-15 12:17:28 -07001214 spin_unlock_irq(&q->queue_lock);
Tejun Heo7f52f982013-05-14 13:52:37 -07001215 cpu_relax();
Christoph Hellwig0d945c12018-11-15 12:17:28 -07001216 spin_lock_irq(&q->queue_lock);
Tejun Heo7f52f982013-05-14 13:52:37 -07001217 }
Tejun Heo6a525602013-05-14 13:52:32 -07001218
Tejun Heo2e48a532013-05-14 13:52:38 -07001219 if (!dispatched)
1220 goto out_unlock;
Tejun Heo6e1a5702013-05-14 13:52:37 -07001221
Tejun Heo2e48a532013-05-14 13:52:38 -07001222 if (parent_sq) {
1223 /* @parent_sq is another throl_grp, propagate dispatch */
1224 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1225 tg_update_disptime(tg);
1226 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1227 /* window is already open, repeat dispatching */
1228 sq = parent_sq;
1229 tg = sq_to_tg(sq);
1230 goto again;
1231 }
1232 }
1233 } else {
Baolin Wangb53b0722020-09-07 16:10:13 +08001234 /* reached the top-level, queue issuing */
Tejun Heo2e48a532013-05-14 13:52:38 -07001235 queue_work(kthrotld_workqueue, &td->dispatch_work);
1236 }
1237out_unlock:
Christoph Hellwig0d945c12018-11-15 12:17:28 -07001238 spin_unlock_irq(&q->queue_lock);
Tejun Heo6e1a5702013-05-14 13:52:37 -07001239}
1240
1241/**
1242 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1243 * @work: work item being executed
1244 *
Baolin Wangb53b0722020-09-07 16:10:13 +08001245 * This function is queued for execution when bios reach the bio_lists[]
1246 * of throtl_data->service_queue. Those bios are ready and issued by this
Tejun Heo6e1a5702013-05-14 13:52:37 -07001247 * function.
1248 */
Fabian Frederick8876e1402014-04-17 21:41:16 +02001249static void blk_throtl_dispatch_work_fn(struct work_struct *work)
Tejun Heo6e1a5702013-05-14 13:52:37 -07001250{
1251 struct throtl_data *td = container_of(work, struct throtl_data,
1252 dispatch_work);
1253 struct throtl_service_queue *td_sq = &td->service_queue;
1254 struct request_queue *q = td->queue;
1255 struct bio_list bio_list_on_stack;
1256 struct bio *bio;
1257 struct blk_plug plug;
1258 int rw;
1259
1260 bio_list_init(&bio_list_on_stack);
1261
Christoph Hellwig0d945c12018-11-15 12:17:28 -07001262 spin_lock_irq(&q->queue_lock);
Tejun Heoc5cc2072013-05-14 13:52:38 -07001263 for (rw = READ; rw <= WRITE; rw++)
1264 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1265 bio_list_add(&bio_list_on_stack, bio);
Christoph Hellwig0d945c12018-11-15 12:17:28 -07001266 spin_unlock_irq(&q->queue_lock);
Vivek Goyale43473b2010-09-15 17:06:35 -04001267
Tejun Heo6e1a5702013-05-14 13:52:37 -07001268 if (!bio_list_empty(&bio_list_on_stack)) {
Vivek Goyal69d60eb2011-03-09 08:27:37 +01001269 blk_start_plug(&plug);
Christoph Hellwiged00aab2020-07-01 10:59:44 +02001270 while ((bio = bio_list_pop(&bio_list_on_stack)))
Ming Lei3f98c752022-02-16 12:45:10 +08001271 submit_bio_noacct_nocheck(bio);
Vivek Goyal69d60eb2011-03-09 08:27:37 +01001272 blk_finish_plug(&plug);
Vivek Goyale43473b2010-09-15 17:06:35 -04001273 }
Vivek Goyale43473b2010-09-15 17:06:35 -04001274}
1275
Tejun Heof95a04a2012-04-16 13:57:26 -07001276static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1277 int off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001278{
Tejun Heof95a04a2012-04-16 13:57:26 -07001279 struct throtl_grp *tg = pd_to_tg(pd);
1280 u64 v = *(u64 *)((void *)tg + off);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001281
Shaohua Li2ab54922017-03-27 10:51:29 -07001282 if (v == U64_MAX)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001283 return 0;
Tejun Heof95a04a2012-04-16 13:57:26 -07001284 return __blkg_prfill_u64(sf, pd, v);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001285}
1286
Tejun Heof95a04a2012-04-16 13:57:26 -07001287static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1288 int off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001289{
Tejun Heof95a04a2012-04-16 13:57:26 -07001290 struct throtl_grp *tg = pd_to_tg(pd);
1291 unsigned int v = *(unsigned int *)((void *)tg + off);
Tejun Heoaf133ce2012-04-01 14:38:44 -07001292
Shaohua Li2ab54922017-03-27 10:51:29 -07001293 if (v == UINT_MAX)
Tejun Heoaf133ce2012-04-01 14:38:44 -07001294 return 0;
Tejun Heof95a04a2012-04-16 13:57:26 -07001295 return __blkg_prfill_u64(sf, pd, v);
Tejun Heoaf133ce2012-04-01 14:38:44 -07001296}
1297
Tejun Heo2da8ca82013-12-05 12:28:04 -05001298static int tg_print_conf_u64(struct seq_file *sf, void *v)
Tejun Heoaf133ce2012-04-01 14:38:44 -07001299{
Tejun Heo2da8ca82013-12-05 12:28:04 -05001300 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1301 &blkcg_policy_throtl, seq_cft(sf)->private, false);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001302 return 0;
1303}
1304
Tejun Heo2da8ca82013-12-05 12:28:04 -05001305static int tg_print_conf_uint(struct seq_file *sf, void *v)
Vivek Goyale43473b2010-09-15 17:06:35 -04001306{
Tejun Heo2da8ca82013-12-05 12:28:04 -05001307 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1308 &blkcg_policy_throtl, seq_cft(sf)->private, false);
Tejun Heoaf133ce2012-04-01 14:38:44 -07001309 return 0;
Vivek Goyale43473b2010-09-15 17:06:35 -04001310}
1311
Shaohua Li9bb67ae2017-05-17 13:07:26 -07001312static void tg_conf_updated(struct throtl_grp *tg, bool global)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001313{
Tejun Heo69948b02015-08-18 14:55:32 -07001314 struct throtl_service_queue *sq = &tg->service_queue;
Tejun Heo492eb212013-08-08 20:11:25 -04001315 struct cgroup_subsys_state *pos_css;
Tejun Heo69948b02015-08-18 14:55:32 -07001316 struct blkcg_gq *blkg;
Tejun Heoaf133ce2012-04-01 14:38:44 -07001317
Tejun Heofda6f272013-05-14 13:52:36 -07001318 throtl_log(&tg->service_queue,
1319 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
Shaohua Li9f626e32017-03-27 10:51:30 -07001320 tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1321 tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
Tejun Heo632b4492013-05-14 13:52:31 -07001322
Ming Lei27b13e22023-11-17 10:35:22 +08001323 rcu_read_lock();
Tejun Heo632b4492013-05-14 13:52:31 -07001324 /*
Tejun Heo693e7512013-05-14 13:52:38 -07001325 * Update has_rules[] flags for the updated tg's subtree. A tg is
1326 * considered to have rules if either the tg itself or any of its
1327 * ancestors has rules. This identifies groups without any
1328 * restrictions in the whole hierarchy and allows them to bypass
1329 * blk-throttle.
1330 */
Shaohua Li9bb67ae2017-05-17 13:07:26 -07001331 blkg_for_each_descendant_pre(blkg, pos_css,
Christoph Hellwig12310392023-02-14 19:33:04 +01001332 global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
Shaohua Li5b81fc32017-05-17 13:07:24 -07001333 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1334 struct throtl_grp *parent_tg;
1335
1336 tg_update_has_rules(this_tg);
1337 /* ignore root/second level */
1338 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1339 !blkg->parent->parent)
1340 continue;
1341 parent_tg = blkg_to_tg(blkg->parent);
1342 /*
1343 * make sure all children has lower idle time threshold and
1344 * higher latency target
1345 */
1346 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1347 parent_tg->idletime_threshold);
1348 this_tg->latency_target = max(this_tg->latency_target,
1349 parent_tg->latency_target);
1350 }
Ming Lei27b13e22023-11-17 10:35:22 +08001351 rcu_read_unlock();
Tejun Heo693e7512013-05-14 13:52:38 -07001352
1353 /*
Tejun Heo632b4492013-05-14 13:52:31 -07001354 * We're already holding queue_lock and know @tg is valid. Let's
1355 * apply the new config directly.
1356 *
1357 * Restart the slices for both READ and WRITES. It might happen
1358 * that a group's limit are dropped suddenly and we don't want to
1359 * account recently dispatched IO with new low rate.
1360 */
Yu Kuaia880ae92022-08-29 10:22:40 +08001361 throtl_start_new_slice(tg, READ, false);
1362 throtl_start_new_slice(tg, WRITE, false);
Tejun Heo632b4492013-05-14 13:52:31 -07001363
Tejun Heo5b2c16a2013-05-14 13:52:32 -07001364 if (tg->flags & THROTL_TG_PENDING) {
Tejun Heo77216b02013-05-14 13:52:36 -07001365 tg_update_disptime(tg);
Tejun Heo7f52f982013-05-14 13:52:37 -07001366 throtl_schedule_next_dispatch(sq->parent_sq, true);
Tejun Heo632b4492013-05-14 13:52:31 -07001367 }
Tejun Heo69948b02015-08-18 14:55:32 -07001368}
Tejun Heo60c2bc22012-04-01 14:38:43 -07001369
Tejun Heo69948b02015-08-18 14:55:32 -07001370static ssize_t tg_set_conf(struct kernfs_open_file *of,
1371 char *buf, size_t nbytes, loff_t off, bool is_u64)
1372{
1373 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1374 struct blkg_conf_ctx ctx;
1375 struct throtl_grp *tg;
1376 int ret;
1377 u64 v;
1378
Tejun Heofaffaab2023-04-12 14:06:47 -10001379 blkg_conf_init(&ctx, buf);
1380
1381 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
Tejun Heo69948b02015-08-18 14:55:32 -07001382 if (ret)
Tejun Heofaffaab2023-04-12 14:06:47 -10001383 goto out_finish;
Tejun Heo69948b02015-08-18 14:55:32 -07001384
1385 ret = -EINVAL;
1386 if (sscanf(ctx.body, "%llu", &v) != 1)
1387 goto out_finish;
1388 if (!v)
Shaohua Li2ab54922017-03-27 10:51:29 -07001389 v = U64_MAX;
Tejun Heo69948b02015-08-18 14:55:32 -07001390
1391 tg = blkg_to_tg(ctx.blkg);
Yu Kuaia880ae92022-08-29 10:22:40 +08001392 tg_update_carryover(tg);
Tejun Heo69948b02015-08-18 14:55:32 -07001393
1394 if (is_u64)
1395 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1396 else
1397 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1398
Shaohua Li9bb67ae2017-05-17 13:07:26 -07001399 tg_conf_updated(tg, false);
Tejun Heo36aa9e52015-08-18 14:55:31 -07001400 ret = 0;
1401out_finish:
Tejun Heofaffaab2023-04-12 14:06:47 -10001402 blkg_conf_exit(&ctx);
Tejun Heo36aa9e52015-08-18 14:55:31 -07001403 return ret ?: nbytes;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001404}
1405
Tejun Heo451af5042014-05-13 12:16:21 -04001406static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1407 char *buf, size_t nbytes, loff_t off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001408{
Tejun Heo451af5042014-05-13 12:16:21 -04001409 return tg_set_conf(of, buf, nbytes, off, true);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001410}
1411
Tejun Heo451af5042014-05-13 12:16:21 -04001412static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1413 char *buf, size_t nbytes, loff_t off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001414{
Tejun Heo451af5042014-05-13 12:16:21 -04001415 return tg_set_conf(of, buf, nbytes, off, false);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001416}
1417
Tejun Heo7ca46432019-11-07 11:18:01 -08001418static int tg_print_rwstat(struct seq_file *sf, void *v)
1419{
1420 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1421 blkg_prfill_rwstat, &blkcg_policy_throtl,
1422 seq_cft(sf)->private, true);
1423 return 0;
1424}
1425
1426static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1427 struct blkg_policy_data *pd, int off)
1428{
1429 struct blkg_rwstat_sample sum;
1430
1431 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1432 &sum);
1433 return __blkg_prfill_rwstat(sf, pd, &sum);
1434}
1435
1436static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1437{
1438 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1439 tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1440 seq_cft(sf)->private, true);
1441 return 0;
1442}
1443
Tejun Heo880f50e22015-08-18 14:55:30 -07001444static struct cftype throtl_legacy_files[] = {
Tejun Heo60c2bc22012-04-01 14:38:43 -07001445 {
1446 .name = "throttle.read_bps_device",
Shaohua Li9f626e32017-03-27 10:51:30 -07001447 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
Tejun Heo2da8ca82013-12-05 12:28:04 -05001448 .seq_show = tg_print_conf_u64,
Tejun Heo451af5042014-05-13 12:16:21 -04001449 .write = tg_set_conf_u64,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001450 },
1451 {
1452 .name = "throttle.write_bps_device",
Shaohua Li9f626e32017-03-27 10:51:30 -07001453 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
Tejun Heo2da8ca82013-12-05 12:28:04 -05001454 .seq_show = tg_print_conf_u64,
Tejun Heo451af5042014-05-13 12:16:21 -04001455 .write = tg_set_conf_u64,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001456 },
1457 {
1458 .name = "throttle.read_iops_device",
Shaohua Li9f626e32017-03-27 10:51:30 -07001459 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
Tejun Heo2da8ca82013-12-05 12:28:04 -05001460 .seq_show = tg_print_conf_uint,
Tejun Heo451af5042014-05-13 12:16:21 -04001461 .write = tg_set_conf_uint,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001462 },
1463 {
1464 .name = "throttle.write_iops_device",
Shaohua Li9f626e32017-03-27 10:51:30 -07001465 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
Tejun Heo2da8ca82013-12-05 12:28:04 -05001466 .seq_show = tg_print_conf_uint,
Tejun Heo451af5042014-05-13 12:16:21 -04001467 .write = tg_set_conf_uint,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001468 },
1469 {
1470 .name = "throttle.io_service_bytes",
Tejun Heo7ca46432019-11-07 11:18:01 -08001471 .private = offsetof(struct throtl_grp, stat_bytes),
1472 .seq_show = tg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001473 },
1474 {
weiping zhang17534c62017-12-11 22:56:25 +08001475 .name = "throttle.io_service_bytes_recursive",
Tejun Heo7ca46432019-11-07 11:18:01 -08001476 .private = offsetof(struct throtl_grp, stat_bytes),
1477 .seq_show = tg_print_rwstat_recursive,
weiping zhang17534c62017-12-11 22:56:25 +08001478 },
1479 {
Tejun Heo60c2bc22012-04-01 14:38:43 -07001480 .name = "throttle.io_serviced",
Tejun Heo7ca46432019-11-07 11:18:01 -08001481 .private = offsetof(struct throtl_grp, stat_ios),
1482 .seq_show = tg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001483 },
weiping zhang17534c62017-12-11 22:56:25 +08001484 {
1485 .name = "throttle.io_serviced_recursive",
Tejun Heo7ca46432019-11-07 11:18:01 -08001486 .private = offsetof(struct throtl_grp, stat_ios),
1487 .seq_show = tg_print_rwstat_recursive,
weiping zhang17534c62017-12-11 22:56:25 +08001488 },
Tejun Heo60c2bc22012-04-01 14:38:43 -07001489 { } /* terminate */
1490};
1491
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001492static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001493 int off)
1494{
1495 struct throtl_grp *tg = pd_to_tg(pd);
1496 const char *dname = blkg_dev_name(pd->blkg);
1497 char bufs[4][21] = { "max", "max", "max", "max" };
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001498 u64 bps_dft;
1499 unsigned int iops_dft;
Shaohua Liada75b62017-03-27 10:51:42 -07001500 char idle_time[26] = "";
Shaohua Liec809912017-03-27 10:51:44 -07001501 char latency_time[26] = "";
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001502
1503 if (!dname)
1504 return 0;
Shaohua Li9f626e32017-03-27 10:51:30 -07001505
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001506 if (off == LIMIT_LOW) {
1507 bps_dft = 0;
1508 iops_dft = 0;
1509 } else {
1510 bps_dft = U64_MAX;
1511 iops_dft = UINT_MAX;
1512 }
1513
1514 if (tg->bps_conf[READ][off] == bps_dft &&
1515 tg->bps_conf[WRITE][off] == bps_dft &&
1516 tg->iops_conf[READ][off] == iops_dft &&
Shaohua Liada75b62017-03-27 10:51:42 -07001517 tg->iops_conf[WRITE][off] == iops_dft &&
Shaohua Liec809912017-03-27 10:51:44 -07001518 (off != LIMIT_LOW ||
Shaohua Lib4f428e2017-05-17 13:07:27 -07001519 (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
Shaohua Li5b81fc32017-05-17 13:07:24 -07001520 tg->latency_target_conf == DFL_LATENCY_TARGET)))
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001521 return 0;
1522
Shaohua Li9bb67ae2017-05-17 13:07:26 -07001523 if (tg->bps_conf[READ][off] != U64_MAX)
Shaohua Li9f626e32017-03-27 10:51:30 -07001524 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001525 tg->bps_conf[READ][off]);
Shaohua Li9bb67ae2017-05-17 13:07:26 -07001526 if (tg->bps_conf[WRITE][off] != U64_MAX)
Shaohua Li9f626e32017-03-27 10:51:30 -07001527 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001528 tg->bps_conf[WRITE][off]);
Shaohua Li9bb67ae2017-05-17 13:07:26 -07001529 if (tg->iops_conf[READ][off] != UINT_MAX)
Shaohua Li9f626e32017-03-27 10:51:30 -07001530 snprintf(bufs[2], sizeof(bufs[2]), "%u",
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001531 tg->iops_conf[READ][off]);
Shaohua Li9bb67ae2017-05-17 13:07:26 -07001532 if (tg->iops_conf[WRITE][off] != UINT_MAX)
Shaohua Li9f626e32017-03-27 10:51:30 -07001533 snprintf(bufs[3], sizeof(bufs[3]), "%u",
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001534 tg->iops_conf[WRITE][off]);
Shaohua Liada75b62017-03-27 10:51:42 -07001535 if (off == LIMIT_LOW) {
Shaohua Li5b81fc32017-05-17 13:07:24 -07001536 if (tg->idletime_threshold_conf == ULONG_MAX)
Shaohua Liada75b62017-03-27 10:51:42 -07001537 strcpy(idle_time, " idle=max");
1538 else
1539 snprintf(idle_time, sizeof(idle_time), " idle=%lu",
Shaohua Li5b81fc32017-05-17 13:07:24 -07001540 tg->idletime_threshold_conf);
Shaohua Liec809912017-03-27 10:51:44 -07001541
Shaohua Li5b81fc32017-05-17 13:07:24 -07001542 if (tg->latency_target_conf == ULONG_MAX)
Shaohua Liec809912017-03-27 10:51:44 -07001543 strcpy(latency_time, " latency=max");
1544 else
1545 snprintf(latency_time, sizeof(latency_time),
Shaohua Li5b81fc32017-05-17 13:07:24 -07001546 " latency=%lu", tg->latency_target_conf);
Shaohua Liada75b62017-03-27 10:51:42 -07001547 }
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001548
Shaohua Liec809912017-03-27 10:51:44 -07001549 seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1550 dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1551 latency_time);
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001552 return 0;
1553}
1554
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001555static int tg_print_limit(struct seq_file *sf, void *v)
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001556{
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001557 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001558 &blkcg_policy_throtl, seq_cft(sf)->private, false);
1559 return 0;
1560}
1561
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001562static ssize_t tg_set_limit(struct kernfs_open_file *of,
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001563 char *buf, size_t nbytes, loff_t off)
1564{
1565 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1566 struct blkg_conf_ctx ctx;
1567 struct throtl_grp *tg;
1568 u64 v[4];
Shaohua Liada75b62017-03-27 10:51:42 -07001569 unsigned long idle_time;
Shaohua Liec809912017-03-27 10:51:44 -07001570 unsigned long latency_time;
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001571 int ret;
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001572 int index = of_cft(of)->private;
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001573
Tejun Heofaffaab2023-04-12 14:06:47 -10001574 blkg_conf_init(&ctx, buf);
1575
1576 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001577 if (ret)
Tejun Heofaffaab2023-04-12 14:06:47 -10001578 goto out_finish;
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001579
1580 tg = blkg_to_tg(ctx.blkg);
Yu Kuaia880ae92022-08-29 10:22:40 +08001581 tg_update_carryover(tg);
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001582
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001583 v[0] = tg->bps_conf[READ][index];
1584 v[1] = tg->bps_conf[WRITE][index];
1585 v[2] = tg->iops_conf[READ][index];
1586 v[3] = tg->iops_conf[WRITE][index];
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001587
Shaohua Li5b81fc32017-05-17 13:07:24 -07001588 idle_time = tg->idletime_threshold_conf;
1589 latency_time = tg->latency_target_conf;
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001590 while (true) {
1591 char tok[27]; /* wiops=18446744073709551616 */
1592 char *p;
Shaohua Li2ab54922017-03-27 10:51:29 -07001593 u64 val = U64_MAX;
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001594 int len;
1595
1596 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1597 break;
1598 if (tok[0] == '\0')
1599 break;
1600 ctx.body += len;
1601
1602 ret = -EINVAL;
1603 p = tok;
1604 strsep(&p, "=");
1605 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1606 goto out_finish;
1607
1608 ret = -ERANGE;
1609 if (!val)
1610 goto out_finish;
1611
1612 ret = -EINVAL;
Baolin Wang5b7048b2020-10-08 11:52:25 +08001613 if (!strcmp(tok, "rbps") && val > 1)
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001614 v[0] = val;
Baolin Wang5b7048b2020-10-08 11:52:25 +08001615 else if (!strcmp(tok, "wbps") && val > 1)
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001616 v[1] = val;
Baolin Wang5b7048b2020-10-08 11:52:25 +08001617 else if (!strcmp(tok, "riops") && val > 1)
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001618 v[2] = min_t(u64, val, UINT_MAX);
Baolin Wang5b7048b2020-10-08 11:52:25 +08001619 else if (!strcmp(tok, "wiops") && val > 1)
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001620 v[3] = min_t(u64, val, UINT_MAX);
Shaohua Liada75b62017-03-27 10:51:42 -07001621 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1622 idle_time = val;
Shaohua Liec809912017-03-27 10:51:44 -07001623 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1624 latency_time = val;
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001625 else
1626 goto out_finish;
1627 }
1628
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001629 tg->bps_conf[READ][index] = v[0];
1630 tg->bps_conf[WRITE][index] = v[1];
1631 tg->iops_conf[READ][index] = v[2];
1632 tg->iops_conf[WRITE][index] = v[3];
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001633
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001634 if (index == LIMIT_MAX) {
1635 tg->bps[READ][index] = v[0];
1636 tg->bps[WRITE][index] = v[1];
1637 tg->iops[READ][index] = v[2];
1638 tg->iops[WRITE][index] = v[3];
1639 }
1640 tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1641 tg->bps_conf[READ][LIMIT_MAX]);
1642 tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1643 tg->bps_conf[WRITE][LIMIT_MAX]);
1644 tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1645 tg->iops_conf[READ][LIMIT_MAX]);
1646 tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1647 tg->iops_conf[WRITE][LIMIT_MAX]);
Shaohua Lib4f428e2017-05-17 13:07:27 -07001648 tg->idletime_threshold_conf = idle_time;
1649 tg->latency_target_conf = latency_time;
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001650
Shaohua Lib4f428e2017-05-17 13:07:27 -07001651 /* force user to configure all settings for low limit */
1652 if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1653 tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1654 tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1655 tg->latency_target_conf == DFL_LATENCY_TARGET) {
1656 tg->bps[READ][LIMIT_LOW] = 0;
1657 tg->bps[WRITE][LIMIT_LOW] = 0;
1658 tg->iops[READ][LIMIT_LOW] = 0;
1659 tg->iops[WRITE][LIMIT_LOW] = 0;
1660 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1661 tg->latency_target = DFL_LATENCY_TARGET;
1662 } else if (index == LIMIT_LOW) {
Shaohua Li5b81fc32017-05-17 13:07:24 -07001663 tg->idletime_threshold = tg->idletime_threshold_conf;
Shaohua Li5b81fc32017-05-17 13:07:24 -07001664 tg->latency_target = tg->latency_target_conf;
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001665 }
Shaohua Lib4f428e2017-05-17 13:07:27 -07001666
1667 blk_throtl_update_limit_valid(tg->td);
1668 if (tg->td->limit_valid[LIMIT_LOW]) {
1669 if (index == LIMIT_LOW)
1670 tg->td->limit_index = LIMIT_LOW;
1671 } else
1672 tg->td->limit_index = LIMIT_MAX;
Shaohua Li9bb67ae2017-05-17 13:07:26 -07001673 tg_conf_updated(tg, index == LIMIT_LOW &&
1674 tg->td->limit_valid[LIMIT_LOW]);
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001675 ret = 0;
1676out_finish:
Tejun Heofaffaab2023-04-12 14:06:47 -10001677 blkg_conf_exit(&ctx);
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001678 return ret ?: nbytes;
1679}
1680
1681static struct cftype throtl_files[] = {
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001682#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1683 {
1684 .name = "low",
1685 .flags = CFTYPE_NOT_ON_ROOT,
1686 .seq_show = tg_print_limit,
1687 .write = tg_set_limit,
1688 .private = LIMIT_LOW,
1689 },
1690#endif
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001691 {
1692 .name = "max",
1693 .flags = CFTYPE_NOT_ON_ROOT,
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001694 .seq_show = tg_print_limit,
1695 .write = tg_set_limit,
1696 .private = LIMIT_MAX,
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001697 },
1698 { } /* terminate */
1699};
1700
Vivek Goyalda527772011-03-02 19:05:33 -05001701static void throtl_shutdown_wq(struct request_queue *q)
Vivek Goyale43473b2010-09-15 17:06:35 -04001702{
1703 struct throtl_data *td = q->td;
1704
Tejun Heo69df0ab2013-05-14 13:52:36 -07001705 cancel_work_sync(&td->dispatch_work);
Vivek Goyale43473b2010-09-15 17:06:35 -04001706}
1707
Jens Axboea7b36ee2021-10-05 09:11:56 -06001708struct blkcg_policy blkcg_policy_throtl = {
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001709 .dfl_cftypes = throtl_files,
Tejun Heo880f50e22015-08-18 14:55:30 -07001710 .legacy_cftypes = throtl_legacy_files,
Tejun Heof9fcc2d2012-04-16 13:57:27 -07001711
Tejun Heo001bea72015-08-18 14:55:11 -07001712 .pd_alloc_fn = throtl_pd_alloc,
Tejun Heof9fcc2d2012-04-16 13:57:27 -07001713 .pd_init_fn = throtl_pd_init,
Tejun Heo693e7512013-05-14 13:52:38 -07001714 .pd_online_fn = throtl_pd_online,
Shaohua Licd5ab1b2017-03-27 10:51:32 -07001715 .pd_offline_fn = throtl_pd_offline,
Tejun Heo001bea72015-08-18 14:55:11 -07001716 .pd_free_fn = throtl_pd_free,
Vivek Goyale43473b2010-09-15 17:06:35 -04001717};
1718
Christoph Hellwigcad92662022-09-21 20:04:58 +02001719void blk_throtl_cancel_bios(struct gendisk *disk)
Yu Kuai2d8f7a32022-09-03 14:28:26 +08001720{
Christoph Hellwigcad92662022-09-21 20:04:58 +02001721 struct request_queue *q = disk->queue;
Yu Kuai2d8f7a32022-09-03 14:28:26 +08001722 struct cgroup_subsys_state *pos_css;
1723 struct blkcg_gq *blkg;
1724
1725 spin_lock_irq(&q->queue_lock);
1726 /*
1727 * queue_lock is held, rcu lock is not needed here technically.
1728 * However, rcu lock is still held to emphasize that following
1729 * path need RCU protection and to prevent warning from lockdep.
1730 */
1731 rcu_read_lock();
Christoph Hellwig12310392023-02-14 19:33:04 +01001732 blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) {
Yu Kuai2d8f7a32022-09-03 14:28:26 +08001733 struct throtl_grp *tg = blkg_to_tg(blkg);
1734 struct throtl_service_queue *sq = &tg->service_queue;
1735
1736 /*
1737 * Set the flag to make sure throtl_pending_timer_fn() won't
1738 * stop until all throttled bios are dispatched.
1739 */
Kemeng Shieb184792022-12-05 19:57:03 +08001740 tg->flags |= THROTL_TG_CANCELING;
1741
1742 /*
1743 * Do not dispatch cgroup without THROTL_TG_PENDING or cgroup
1744 * will be inserted to service queue without THROTL_TG_PENDING
1745 * set in tg_update_disptime below. Then IO dispatched from
1746 * child in tg_dispatch_one_bio will trigger double insertion
1747 * and corrupt the tree.
1748 */
1749 if (!(tg->flags & THROTL_TG_PENDING))
1750 continue;
1751
Yu Kuai2d8f7a32022-09-03 14:28:26 +08001752 /*
1753 * Update disptime after setting the above flag to make sure
1754 * throtl_select_dispatch() won't exit without dispatching.
1755 */
1756 tg_update_disptime(tg);
1757
1758 throtl_schedule_pending_timer(sq, jiffies + 1);
1759 }
1760 rcu_read_unlock();
1761 spin_unlock_irq(&q->queue_lock);
1762}
1763
1764#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
Shaohua Li3f0abd82017-03-27 10:51:35 -07001765static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1766{
1767 unsigned long rtime = jiffies, wtime = jiffies;
1768
1769 if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1770 rtime = tg->last_low_overflow_time[READ];
1771 if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1772 wtime = tg->last_low_overflow_time[WRITE];
1773 return min(rtime, wtime);
1774}
1775
Shaohua Li3f0abd82017-03-27 10:51:35 -07001776static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1777{
1778 struct throtl_service_queue *parent_sq;
1779 struct throtl_grp *parent = tg;
1780 unsigned long ret = __tg_last_low_overflow_time(tg);
1781
1782 while (true) {
1783 parent_sq = parent->service_queue.parent_sq;
1784 parent = sq_to_tg(parent_sq);
1785 if (!parent)
1786 break;
1787
1788 /*
1789 * The parent doesn't have low limit, it always reaches low
1790 * limit. Its overflow time is useless for children
1791 */
1792 if (!parent->bps[READ][LIMIT_LOW] &&
1793 !parent->iops[READ][LIMIT_LOW] &&
1794 !parent->bps[WRITE][LIMIT_LOW] &&
1795 !parent->iops[WRITE][LIMIT_LOW])
1796 continue;
1797 if (time_after(__tg_last_low_overflow_time(parent), ret))
1798 ret = __tg_last_low_overflow_time(parent);
1799 }
1800 return ret;
1801}
1802
Shaohua Li9e234ee2017-03-27 10:51:41 -07001803static bool throtl_tg_is_idle(struct throtl_grp *tg)
1804{
1805 /*
1806 * cgroup is idle if:
1807 * - single idle is too long, longer than a fixed value (in case user
Shaohua Lib4f428e2017-05-17 13:07:27 -07001808 * configure a too big threshold) or 4 times of idletime threshold
Shaohua Li9e234ee2017-03-27 10:51:41 -07001809 * - average think time is more than threshold
Shaohua Li53696b82017-03-27 15:19:43 -07001810 * - IO latency is largely below threshold
Shaohua Li9e234ee2017-03-27 10:51:41 -07001811 */
Shaohua Lib4f428e2017-05-17 13:07:27 -07001812 unsigned long time;
Shaohua Li4cff7292017-05-17 13:07:25 -07001813 bool ret;
Shaohua Li9e234ee2017-03-27 10:51:41 -07001814
Shaohua Lib4f428e2017-05-17 13:07:27 -07001815 time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1816 ret = tg->latency_target == DFL_LATENCY_TARGET ||
1817 tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1818 (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1819 tg->avg_idletime > tg->idletime_threshold ||
1820 (tg->latency_target && tg->bio_cnt &&
Shaohua Li53696b82017-03-27 15:19:43 -07001821 tg->bad_bio_cnt * 5 < tg->bio_cnt);
Shaohua Li4cff7292017-05-17 13:07:25 -07001822 throtl_log(&tg->service_queue,
1823 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1824 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1825 tg->bio_cnt, ret, tg->td->scale);
1826 return ret;
Shaohua Li9e234ee2017-03-27 10:51:41 -07001827}
1828
Kemeng Shia4d508e2022-12-05 19:57:05 +08001829static bool throtl_low_limit_reached(struct throtl_grp *tg, int rw)
Shaohua Lic79892c2017-03-27 10:51:34 -07001830{
1831 struct throtl_service_queue *sq = &tg->service_queue;
Kemeng Shia4d508e2022-12-05 19:57:05 +08001832 bool limit = tg->bps[rw][LIMIT_LOW] || tg->iops[rw][LIMIT_LOW];
Shaohua Lic79892c2017-03-27 10:51:34 -07001833
1834 /*
Kemeng Shia4d508e2022-12-05 19:57:05 +08001835 * if low limit is zero, low limit is always reached.
1836 * if low limit is non-zero, we can check if there is any request
1837 * is queued to determine if low limit is reached as we throttle
1838 * request according to limit.
Shaohua Lic79892c2017-03-27 10:51:34 -07001839 */
Kemeng Shia4d508e2022-12-05 19:57:05 +08001840 return !limit || sq->nr_queued[rw];
1841}
1842
1843static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1844{
1845 /*
1846 * cgroup reaches low limit when low limit of READ and WRITE are
1847 * both reached, it's ok to upgrade to next limit if cgroup reaches
1848 * low limit
1849 */
1850 if (throtl_low_limit_reached(tg, READ) &&
1851 throtl_low_limit_reached(tg, WRITE))
Shaohua Lic79892c2017-03-27 10:51:34 -07001852 return true;
Shaohua Liaec24242017-03-27 10:51:39 -07001853
1854 if (time_after_eq(jiffies,
Shaohua Lifa6fb5a2017-03-27 10:51:43 -07001855 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1856 throtl_tg_is_idle(tg))
Shaohua Liaec24242017-03-27 10:51:39 -07001857 return true;
Shaohua Lic79892c2017-03-27 10:51:34 -07001858 return false;
1859}
1860
1861static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1862{
1863 while (true) {
1864 if (throtl_tg_can_upgrade(tg))
1865 return true;
1866 tg = sq_to_tg(tg->service_queue.parent_sq);
1867 if (!tg || !tg_to_blkg(tg)->parent)
1868 return false;
1869 }
1870 return false;
1871}
1872
1873static bool throtl_can_upgrade(struct throtl_data *td,
1874 struct throtl_grp *this_tg)
1875{
1876 struct cgroup_subsys_state *pos_css;
1877 struct blkcg_gq *blkg;
1878
1879 if (td->limit_index != LIMIT_LOW)
1880 return false;
1881
Shaohua Li297e3d82017-03-27 10:51:37 -07001882 if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
Shaohua Li3f0abd82017-03-27 10:51:35 -07001883 return false;
1884
Shaohua Lic79892c2017-03-27 10:51:34 -07001885 rcu_read_lock();
Christoph Hellwig12310392023-02-14 19:33:04 +01001886 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
Shaohua Lic79892c2017-03-27 10:51:34 -07001887 struct throtl_grp *tg = blkg_to_tg(blkg);
1888
1889 if (tg == this_tg)
1890 continue;
1891 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1892 continue;
1893 if (!throtl_hierarchy_can_upgrade(tg)) {
1894 rcu_read_unlock();
1895 return false;
1896 }
1897 }
1898 rcu_read_unlock();
1899 return true;
1900}
1901
Shaohua Lifa6fb5a2017-03-27 10:51:43 -07001902static void throtl_upgrade_check(struct throtl_grp *tg)
1903{
1904 unsigned long now = jiffies;
1905
1906 if (tg->td->limit_index != LIMIT_LOW)
1907 return;
1908
1909 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1910 return;
1911
1912 tg->last_check_time = now;
1913
1914 if (!time_after_eq(now,
1915 __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1916 return;
1917
1918 if (throtl_can_upgrade(tg->td, NULL))
1919 throtl_upgrade_state(tg->td);
1920}
1921
Shaohua Lic79892c2017-03-27 10:51:34 -07001922static void throtl_upgrade_state(struct throtl_data *td)
1923{
1924 struct cgroup_subsys_state *pos_css;
1925 struct blkcg_gq *blkg;
1926
Shaohua Li4cff7292017-05-17 13:07:25 -07001927 throtl_log(&td->service_queue, "upgrade to max");
Shaohua Lic79892c2017-03-27 10:51:34 -07001928 td->limit_index = LIMIT_MAX;
Shaohua Li3f0abd82017-03-27 10:51:35 -07001929 td->low_upgrade_time = jiffies;
Shaohua Li7394e312017-03-27 10:51:40 -07001930 td->scale = 0;
Shaohua Lic79892c2017-03-27 10:51:34 -07001931 rcu_read_lock();
Christoph Hellwig12310392023-02-14 19:33:04 +01001932 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
Shaohua Lic79892c2017-03-27 10:51:34 -07001933 struct throtl_grp *tg = blkg_to_tg(blkg);
1934 struct throtl_service_queue *sq = &tg->service_queue;
1935
1936 tg->disptime = jiffies - 1;
1937 throtl_select_dispatch(sq);
Joseph Qi4f02fb72017-09-30 14:38:49 +08001938 throtl_schedule_next_dispatch(sq, true);
Shaohua Lic79892c2017-03-27 10:51:34 -07001939 }
1940 rcu_read_unlock();
1941 throtl_select_dispatch(&td->service_queue);
Joseph Qi4f02fb72017-09-30 14:38:49 +08001942 throtl_schedule_next_dispatch(&td->service_queue, true);
Shaohua Lic79892c2017-03-27 10:51:34 -07001943 queue_work(kthrotld_workqueue, &td->dispatch_work);
1944}
1945
Baolin Wang4247d9c82020-10-08 11:52:22 +08001946static void throtl_downgrade_state(struct throtl_data *td)
Shaohua Li3f0abd82017-03-27 10:51:35 -07001947{
Shaohua Li7394e312017-03-27 10:51:40 -07001948 td->scale /= 2;
1949
Shaohua Li4cff7292017-05-17 13:07:25 -07001950 throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
Shaohua Li7394e312017-03-27 10:51:40 -07001951 if (td->scale) {
1952 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1953 return;
1954 }
1955
Baolin Wang4247d9c82020-10-08 11:52:22 +08001956 td->limit_index = LIMIT_LOW;
Shaohua Li3f0abd82017-03-27 10:51:35 -07001957 td->low_downgrade_time = jiffies;
1958}
1959
1960static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1961{
1962 struct throtl_data *td = tg->td;
1963 unsigned long now = jiffies;
1964
1965 /*
1966 * If cgroup is below low limit, consider downgrade and throttle other
1967 * cgroups
1968 */
Kemeng Shi9c9f209d2022-12-05 19:57:08 +08001969 if (time_after_eq(now, tg_last_low_overflow_time(tg) +
Shaohua Lifa6fb5a2017-03-27 10:51:43 -07001970 td->throtl_slice) &&
1971 (!throtl_tg_is_idle(tg) ||
1972 !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
Shaohua Li3f0abd82017-03-27 10:51:35 -07001973 return true;
1974 return false;
1975}
1976
1977static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1978{
Kemeng Shi9c9f209d2022-12-05 19:57:08 +08001979 struct throtl_data *td = tg->td;
1980
1981 if (time_before(jiffies, td->low_upgrade_time + td->throtl_slice))
1982 return false;
1983
Shaohua Li3f0abd82017-03-27 10:51:35 -07001984 while (true) {
1985 if (!throtl_tg_can_downgrade(tg))
1986 return false;
1987 tg = sq_to_tg(tg->service_queue.parent_sq);
1988 if (!tg || !tg_to_blkg(tg)->parent)
1989 break;
1990 }
1991 return true;
1992}
1993
1994static void throtl_downgrade_check(struct throtl_grp *tg)
1995{
1996 uint64_t bps;
1997 unsigned int iops;
1998 unsigned long elapsed_time;
1999 unsigned long now = jiffies;
2000
2001 if (tg->td->limit_index != LIMIT_MAX ||
2002 !tg->td->limit_valid[LIMIT_LOW])
2003 return;
2004 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
2005 return;
Shaohua Li297e3d82017-03-27 10:51:37 -07002006 if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
Shaohua Li3f0abd82017-03-27 10:51:35 -07002007 return;
2008
2009 elapsed_time = now - tg->last_check_time;
2010 tg->last_check_time = now;
2011
Shaohua Li297e3d82017-03-27 10:51:37 -07002012 if (time_before(now, tg_last_low_overflow_time(tg) +
2013 tg->td->throtl_slice))
Shaohua Li3f0abd82017-03-27 10:51:35 -07002014 return;
2015
2016 if (tg->bps[READ][LIMIT_LOW]) {
2017 bps = tg->last_bytes_disp[READ] * HZ;
2018 do_div(bps, elapsed_time);
2019 if (bps >= tg->bps[READ][LIMIT_LOW])
2020 tg->last_low_overflow_time[READ] = now;
2021 }
2022
2023 if (tg->bps[WRITE][LIMIT_LOW]) {
2024 bps = tg->last_bytes_disp[WRITE] * HZ;
2025 do_div(bps, elapsed_time);
2026 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2027 tg->last_low_overflow_time[WRITE] = now;
2028 }
2029
2030 if (tg->iops[READ][LIMIT_LOW]) {
2031 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2032 if (iops >= tg->iops[READ][LIMIT_LOW])
2033 tg->last_low_overflow_time[READ] = now;
2034 }
2035
2036 if (tg->iops[WRITE][LIMIT_LOW]) {
2037 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2038 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2039 tg->last_low_overflow_time[WRITE] = now;
2040 }
2041
2042 /*
2043 * If cgroup is below low limit, consider downgrade and throttle other
2044 * cgroups
2045 */
2046 if (throtl_hierarchy_can_downgrade(tg))
Baolin Wang4247d9c82020-10-08 11:52:22 +08002047 throtl_downgrade_state(tg->td);
Shaohua Li3f0abd82017-03-27 10:51:35 -07002048
2049 tg->last_bytes_disp[READ] = 0;
2050 tg->last_bytes_disp[WRITE] = 0;
2051 tg->last_io_disp[READ] = 0;
2052 tg->last_io_disp[WRITE] = 0;
2053}
2054
Shaohua Li9e234ee2017-03-27 10:51:41 -07002055static void blk_throtl_update_idletime(struct throtl_grp *tg)
2056{
Baolin Wang79016012020-10-08 11:52:23 +08002057 unsigned long now;
Shaohua Li9e234ee2017-03-27 10:51:41 -07002058 unsigned long last_finish_time = tg->last_finish_time;
2059
Baolin Wang79016012020-10-08 11:52:23 +08002060 if (last_finish_time == 0)
2061 return;
2062
2063 now = ktime_get_ns() >> 10;
2064 if (now <= last_finish_time ||
Shaohua Li9e234ee2017-03-27 10:51:41 -07002065 last_finish_time == tg->checked_last_finish_time)
2066 return;
2067
2068 tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2069 tg->checked_last_finish_time = last_finish_time;
2070}
2071
Shaohua Lib9147dd2017-03-27 15:19:42 -07002072static void throtl_update_latency_buckets(struct throtl_data *td)
2073{
Joseph Qib889bf62017-11-21 09:38:30 +08002074 struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2075 int i, cpu, rw;
2076 unsigned long last_latency[2] = { 0 };
2077 unsigned long latency[2];
Shaohua Lib9147dd2017-03-27 15:19:42 -07002078
Baolin Wangb185efa2020-10-08 11:52:24 +08002079 if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
Shaohua Lib9147dd2017-03-27 15:19:42 -07002080 return;
2081 if (time_before(jiffies, td->last_calculate_time + HZ))
2082 return;
2083 td->last_calculate_time = jiffies;
2084
2085 memset(avg_latency, 0, sizeof(avg_latency));
Joseph Qib889bf62017-11-21 09:38:30 +08002086 for (rw = READ; rw <= WRITE; rw++) {
2087 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2088 struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
Shaohua Lib9147dd2017-03-27 15:19:42 -07002089
Joseph Qib889bf62017-11-21 09:38:30 +08002090 for_each_possible_cpu(cpu) {
2091 struct latency_bucket *bucket;
Shaohua Lib9147dd2017-03-27 15:19:42 -07002092
Joseph Qib889bf62017-11-21 09:38:30 +08002093 /* this isn't race free, but ok in practice */
2094 bucket = per_cpu_ptr(td->latency_buckets[rw],
2095 cpu);
2096 tmp->total_latency += bucket[i].total_latency;
2097 tmp->samples += bucket[i].samples;
2098 bucket[i].total_latency = 0;
2099 bucket[i].samples = 0;
2100 }
Shaohua Lib9147dd2017-03-27 15:19:42 -07002101
Joseph Qib889bf62017-11-21 09:38:30 +08002102 if (tmp->samples >= 32) {
2103 int samples = tmp->samples;
Shaohua Lib9147dd2017-03-27 15:19:42 -07002104
Joseph Qib889bf62017-11-21 09:38:30 +08002105 latency[rw] = tmp->total_latency;
Shaohua Lib9147dd2017-03-27 15:19:42 -07002106
Joseph Qib889bf62017-11-21 09:38:30 +08002107 tmp->total_latency = 0;
2108 tmp->samples = 0;
2109 latency[rw] /= samples;
2110 if (latency[rw] == 0)
2111 continue;
2112 avg_latency[rw][i].latency = latency[rw];
2113 }
Shaohua Lib9147dd2017-03-27 15:19:42 -07002114 }
2115 }
2116
Joseph Qib889bf62017-11-21 09:38:30 +08002117 for (rw = READ; rw <= WRITE; rw++) {
2118 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2119 if (!avg_latency[rw][i].latency) {
2120 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2121 td->avg_buckets[rw][i].latency =
2122 last_latency[rw];
2123 continue;
2124 }
2125
2126 if (!td->avg_buckets[rw][i].valid)
2127 latency[rw] = avg_latency[rw][i].latency;
2128 else
2129 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2130 avg_latency[rw][i].latency) >> 3;
2131
2132 td->avg_buckets[rw][i].latency = max(latency[rw],
2133 last_latency[rw]);
2134 td->avg_buckets[rw][i].valid = true;
2135 last_latency[rw] = td->avg_buckets[rw][i].latency;
Shaohua Lib9147dd2017-03-27 15:19:42 -07002136 }
Shaohua Lib9147dd2017-03-27 15:19:42 -07002137 }
Shaohua Li4cff7292017-05-17 13:07:25 -07002138
2139 for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2140 throtl_log(&td->service_queue,
Joseph Qib889bf62017-11-21 09:38:30 +08002141 "Latency bucket %d: read latency=%ld, read valid=%d, "
2142 "write latency=%ld, write valid=%d", i,
2143 td->avg_buckets[READ][i].latency,
2144 td->avg_buckets[READ][i].valid,
2145 td->avg_buckets[WRITE][i].latency,
2146 td->avg_buckets[WRITE][i].valid);
Shaohua Lib9147dd2017-03-27 15:19:42 -07002147}
2148#else
2149static inline void throtl_update_latency_buckets(struct throtl_data *td)
2150{
2151}
Yu Kuai2d8f7a32022-09-03 14:28:26 +08002152
2153static void blk_throtl_update_idletime(struct throtl_grp *tg)
2154{
2155}
2156
2157static void throtl_downgrade_check(struct throtl_grp *tg)
2158{
2159}
2160
2161static void throtl_upgrade_check(struct throtl_grp *tg)
2162{
2163}
2164
2165static bool throtl_can_upgrade(struct throtl_data *td,
2166 struct throtl_grp *this_tg)
2167{
2168 return false;
2169}
2170
2171static void throtl_upgrade_state(struct throtl_data *td)
2172{
2173}
Shaohua Lib9147dd2017-03-27 15:19:42 -07002174#endif
2175
Jens Axboea7b36ee2021-10-05 09:11:56 -06002176bool __blk_throtl_bio(struct bio *bio)
Vivek Goyale43473b2010-09-15 17:06:35 -04002177{
Pavel Begunkoved6cdde2021-10-14 15:03:30 +01002178 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
Christoph Hellwigdb18a532020-06-27 09:31:58 +02002179 struct blkcg_gq *blkg = bio->bi_blkg;
Tejun Heoc5cc2072013-05-14 13:52:38 -07002180 struct throtl_qnode *qn = NULL;
Christoph Hellwiga2e83ef2020-06-27 09:31:59 +02002181 struct throtl_grp *tg = blkg_to_tg(blkg);
Tejun Heo73f0d492013-05-14 13:52:35 -07002182 struct throtl_service_queue *sq;
Tejun Heo0e9f4162013-05-14 13:52:35 -07002183 bool rw = bio_data_dir(bio);
Tejun Heobc16a4f2011-10-19 14:33:01 +02002184 bool throttled = false;
Shaohua Lib9147dd2017-03-27 15:19:42 -07002185 struct throtl_data *td = tg->td;
Vivek Goyale43473b2010-09-15 17:06:35 -04002186
Christoph Hellwig93b80632020-06-27 09:31:57 +02002187 rcu_read_lock();
Tejun Heoae118892015-08-18 14:55:20 -07002188
Christoph Hellwig0d945c12018-11-15 12:17:28 -07002189 spin_lock_irq(&q->queue_lock);
Tejun Heoc9589f02015-08-18 14:55:19 -07002190
Shaohua Lib9147dd2017-03-27 15:19:42 -07002191 throtl_update_latency_buckets(td);
2192
Shaohua Li9e234ee2017-03-27 10:51:41 -07002193 blk_throtl_update_idletime(tg);
2194
Tejun Heo73f0d492013-05-14 13:52:35 -07002195 sq = &tg->service_queue;
2196
Shaohua Lic79892c2017-03-27 10:51:34 -07002197again:
Tejun Heo9e660ac2013-05-14 13:52:38 -07002198 while (true) {
Shaohua Li3f0abd82017-03-27 10:51:35 -07002199 if (tg->last_low_overflow_time[rw] == 0)
2200 tg->last_low_overflow_time[rw] = jiffies;
2201 throtl_downgrade_check(tg);
Shaohua Lifa6fb5a2017-03-27 10:51:43 -07002202 throtl_upgrade_check(tg);
Tejun Heo9e660ac2013-05-14 13:52:38 -07002203 /* throtl is FIFO - if bios are already queued, should queue */
2204 if (sq->nr_queued[rw])
2205 break;
Vivek Goyalde701c72011-03-07 21:09:32 +01002206
Tejun Heo9e660ac2013-05-14 13:52:38 -07002207 /* if above limits, break to queue */
Shaohua Lic79892c2017-03-27 10:51:34 -07002208 if (!tg_may_dispatch(tg, bio, NULL)) {
Shaohua Li3f0abd82017-03-27 10:51:35 -07002209 tg->last_low_overflow_time[rw] = jiffies;
Shaohua Lib9147dd2017-03-27 15:19:42 -07002210 if (throtl_can_upgrade(td, tg)) {
2211 throtl_upgrade_state(td);
Shaohua Lic79892c2017-03-27 10:51:34 -07002212 goto again;
2213 }
Tejun Heo9e660ac2013-05-14 13:52:38 -07002214 break;
Shaohua Lic79892c2017-03-27 10:51:34 -07002215 }
Tejun Heo9e660ac2013-05-14 13:52:38 -07002216
2217 /* within limits, let's charge and dispatch directly */
Vivek Goyale43473b2010-09-15 17:06:35 -04002218 throtl_charge_bio(tg, bio);
Vivek Goyal04521db2011-03-22 21:54:29 +01002219
2220 /*
2221 * We need to trim slice even when bios are not being queued
2222 * otherwise it might happen that a bio is not queued for
2223 * a long time and slice keeps on extending and trim is not
2224 * called for a long time. Now if limits are reduced suddenly
2225 * we take into account all the IO dispatched so far at new
2226 * low rate and * newly queued IO gets a really long dispatch
2227 * time.
2228 *
2229 * So keep on trimming slice even if bio is not queued.
2230 */
Tejun Heo0f3457f2013-05-14 13:52:32 -07002231 throtl_trim_slice(tg, rw);
Tejun Heo9e660ac2013-05-14 13:52:38 -07002232
2233 /*
2234 * @bio passed through this layer without being throttled.
Baolin Wangb53b0722020-09-07 16:10:13 +08002235 * Climb up the ladder. If we're already at the top, it
Tejun Heo9e660ac2013-05-14 13:52:38 -07002236 * can be executed directly.
2237 */
Tejun Heoc5cc2072013-05-14 13:52:38 -07002238 qn = &tg->qnode_on_parent[rw];
Tejun Heo9e660ac2013-05-14 13:52:38 -07002239 sq = sq->parent_sq;
2240 tg = sq_to_tg(sq);
Yu Kuai320fb0f2022-08-29 10:22:37 +08002241 if (!tg) {
2242 bio_set_flag(bio, BIO_BPS_THROTTLED);
Tejun Heo9e660ac2013-05-14 13:52:38 -07002243 goto out_unlock;
Yu Kuai320fb0f2022-08-29 10:22:37 +08002244 }
Vivek Goyale43473b2010-09-15 17:06:35 -04002245 }
2246
Tejun Heo9e660ac2013-05-14 13:52:38 -07002247 /* out-of-limit, queue to @tg */
Tejun Heofda6f272013-05-14 13:52:36 -07002248 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2249 rw == READ ? 'R' : 'W',
Shaohua Li9f626e32017-03-27 10:51:30 -07002250 tg->bytes_disp[rw], bio->bi_iter.bi_size,
2251 tg_bps_limit(tg, rw),
2252 tg->io_disp[rw], tg_iops_limit(tg, rw),
Tejun Heofda6f272013-05-14 13:52:36 -07002253 sq->nr_queued[READ], sq->nr_queued[WRITE]);
Vivek Goyale43473b2010-09-15 17:06:35 -04002254
Shaohua Li3f0abd82017-03-27 10:51:35 -07002255 tg->last_low_overflow_time[rw] = jiffies;
2256
Shaohua Lib9147dd2017-03-27 15:19:42 -07002257 td->nr_queued[rw]++;
Tejun Heoc5cc2072013-05-14 13:52:38 -07002258 throtl_add_bio_tg(bio, qn, tg);
Tejun Heobc16a4f2011-10-19 14:33:01 +02002259 throttled = true;
Vivek Goyale43473b2010-09-15 17:06:35 -04002260
Tejun Heo7f52f982013-05-14 13:52:37 -07002261 /*
2262 * Update @tg's dispatch time and force schedule dispatch if @tg
2263 * was empty before @bio. The forced scheduling isn't likely to
2264 * cause undue delay as @bio is likely to be dispatched directly if
2265 * its @tg's disptime is not in the future.
2266 */
Tejun Heo0e9f4162013-05-14 13:52:35 -07002267 if (tg->flags & THROTL_TG_WAS_EMPTY) {
Tejun Heo77216b02013-05-14 13:52:36 -07002268 tg_update_disptime(tg);
Tejun Heo7f52f982013-05-14 13:52:37 -07002269 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
Vivek Goyale43473b2010-09-15 17:06:35 -04002270 }
2271
Tejun Heobc16a4f2011-10-19 14:33:01 +02002272out_unlock:
Shaohua Lib9147dd2017-03-27 15:19:42 -07002273#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2274 if (throttled || !td->track_bio_latency)
Omar Sandoval5238dcf2018-05-09 02:08:49 -07002275 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
Shaohua Lib9147dd2017-03-27 15:19:42 -07002276#endif
Laibin Qiu5a011f882022-03-01 20:39:19 +08002277 spin_unlock_irq(&q->queue_lock);
2278
Christoph Hellwig93b80632020-06-27 09:31:57 +02002279 rcu_read_unlock();
Tejun Heobc16a4f2011-10-19 14:33:01 +02002280 return throttled;
Vivek Goyale43473b2010-09-15 17:06:35 -04002281}
2282
Shaohua Li9e234ee2017-03-27 10:51:41 -07002283#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
Shaohua Lib9147dd2017-03-27 15:19:42 -07002284static void throtl_track_latency(struct throtl_data *td, sector_t size,
Bart Van Assche77e7ffd2022-07-14 11:06:28 -07002285 enum req_op op, unsigned long time)
Shaohua Lib9147dd2017-03-27 15:19:42 -07002286{
Bart Van Assche77e7ffd2022-07-14 11:06:28 -07002287 const bool rw = op_is_write(op);
Shaohua Lib9147dd2017-03-27 15:19:42 -07002288 struct latency_bucket *latency;
2289 int index;
2290
Joseph Qib889bf62017-11-21 09:38:30 +08002291 if (!td || td->limit_index != LIMIT_LOW ||
2292 !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
Shaohua Lib9147dd2017-03-27 15:19:42 -07002293 !blk_queue_nonrot(td->queue))
2294 return;
2295
2296 index = request_bucket_index(size);
2297
Bart Van Assche77e7ffd2022-07-14 11:06:28 -07002298 latency = get_cpu_ptr(td->latency_buckets[rw]);
Shaohua Lib9147dd2017-03-27 15:19:42 -07002299 latency[index].total_latency += time;
2300 latency[index].samples++;
Bart Van Assche77e7ffd2022-07-14 11:06:28 -07002301 put_cpu_ptr(td->latency_buckets[rw]);
Shaohua Lib9147dd2017-03-27 15:19:42 -07002302}
2303
2304void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2305{
2306 struct request_queue *q = rq->q;
2307 struct throtl_data *td = q->td;
2308
Hou Tao3d244302019-05-21 15:59:03 +08002309 throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2310 time_ns >> 10);
Shaohua Lib9147dd2017-03-27 15:19:42 -07002311}
2312
Shaohua Li9e234ee2017-03-27 10:51:41 -07002313void blk_throtl_bio_endio(struct bio *bio)
2314{
Josef Bacik08e18ea2018-07-03 11:14:50 -04002315 struct blkcg_gq *blkg;
Shaohua Li9e234ee2017-03-27 10:51:41 -07002316 struct throtl_grp *tg;
Shaohua Lib9147dd2017-03-27 15:19:42 -07002317 u64 finish_time_ns;
2318 unsigned long finish_time;
2319 unsigned long start_time;
2320 unsigned long lat;
Joseph Qib889bf62017-11-21 09:38:30 +08002321 int rw = bio_data_dir(bio);
Shaohua Li9e234ee2017-03-27 10:51:41 -07002322
Josef Bacik08e18ea2018-07-03 11:14:50 -04002323 blkg = bio->bi_blkg;
2324 if (!blkg)
Shaohua Li9e234ee2017-03-27 10:51:41 -07002325 return;
Josef Bacik08e18ea2018-07-03 11:14:50 -04002326 tg = blkg_to_tg(blkg);
Baolin Wangb185efa2020-10-08 11:52:24 +08002327 if (!tg->td->limit_valid[LIMIT_LOW])
2328 return;
Shaohua Li9e234ee2017-03-27 10:51:41 -07002329
Shaohua Lib9147dd2017-03-27 15:19:42 -07002330 finish_time_ns = ktime_get_ns();
2331 tg->last_finish_time = finish_time_ns >> 10;
2332
Omar Sandoval5238dcf2018-05-09 02:08:49 -07002333 start_time = bio_issue_time(&bio->bi_issue) >> 10;
2334 finish_time = __bio_issue_time(finish_time_ns) >> 10;
Josef Bacik08e18ea2018-07-03 11:14:50 -04002335 if (!start_time || finish_time <= start_time)
Shaohua Li53696b82017-03-27 15:19:43 -07002336 return;
2337
2338 lat = finish_time - start_time;
Shaohua Lib9147dd2017-03-27 15:19:42 -07002339 /* this is only for bio based driver */
Omar Sandoval5238dcf2018-05-09 02:08:49 -07002340 if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2341 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2342 bio_op(bio), lat);
Shaohua Li53696b82017-03-27 15:19:43 -07002343
Shaohua Li6679a902017-06-06 12:40:43 -07002344 if (tg->latency_target && lat >= tg->td->filtered_latency) {
Shaohua Li53696b82017-03-27 15:19:43 -07002345 int bucket;
2346 unsigned int threshold;
2347
Omar Sandoval5238dcf2018-05-09 02:08:49 -07002348 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
Joseph Qib889bf62017-11-21 09:38:30 +08002349 threshold = tg->td->avg_buckets[rw][bucket].latency +
Shaohua Li53696b82017-03-27 15:19:43 -07002350 tg->latency_target;
2351 if (lat > threshold)
2352 tg->bad_bio_cnt++;
2353 /*
2354 * Not race free, could get wrong count, which means cgroups
2355 * will be throttled
2356 */
2357 tg->bio_cnt++;
2358 }
2359
2360 if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2361 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2362 tg->bio_cnt /= 2;
2363 tg->bad_bio_cnt /= 2;
Shaohua Lib9147dd2017-03-27 15:19:42 -07002364 }
Shaohua Li9e234ee2017-03-27 10:51:41 -07002365}
2366#endif
2367
Christoph Hellwige13793b2022-09-21 20:04:56 +02002368int blk_throtl_init(struct gendisk *disk)
Vivek Goyale43473b2010-09-15 17:06:35 -04002369{
Christoph Hellwige13793b2022-09-21 20:04:56 +02002370 struct request_queue *q = disk->queue;
Vivek Goyale43473b2010-09-15 17:06:35 -04002371 struct throtl_data *td;
Tejun Heoa2b16932012-04-13 13:11:33 -07002372 int ret;
Vivek Goyale43473b2010-09-15 17:06:35 -04002373
2374 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2375 if (!td)
2376 return -ENOMEM;
Joseph Qib889bf62017-11-21 09:38:30 +08002377 td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
Shaohua Lib9147dd2017-03-27 15:19:42 -07002378 LATENCY_BUCKET_SIZE, __alignof__(u64));
Joseph Qib889bf62017-11-21 09:38:30 +08002379 if (!td->latency_buckets[READ]) {
2380 kfree(td);
2381 return -ENOMEM;
2382 }
2383 td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2384 LATENCY_BUCKET_SIZE, __alignof__(u64));
2385 if (!td->latency_buckets[WRITE]) {
2386 free_percpu(td->latency_buckets[READ]);
Shaohua Lib9147dd2017-03-27 15:19:42 -07002387 kfree(td);
2388 return -ENOMEM;
2389 }
Vivek Goyale43473b2010-09-15 17:06:35 -04002390
Tejun Heo69df0ab2013-05-14 13:52:36 -07002391 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
Tejun Heob2ce2642015-08-18 14:55:13 -07002392 throtl_service_queue_init(&td->service_queue);
Vivek Goyale43473b2010-09-15 17:06:35 -04002393
Tejun Heocd1604f2012-03-05 13:15:06 -08002394 q->td = td;
Vivek Goyal29b12582011-05-19 15:38:24 -04002395 td->queue = q;
Vivek Goyal02977e42010-10-01 14:49:48 +02002396
Shaohua Li9f626e32017-03-27 10:51:30 -07002397 td->limit_valid[LIMIT_MAX] = true;
Shaohua Licd5ab1b2017-03-27 10:51:32 -07002398 td->limit_index = LIMIT_MAX;
Shaohua Li3f0abd82017-03-27 10:51:35 -07002399 td->low_upgrade_time = jiffies;
2400 td->low_downgrade_time = jiffies;
Shaohua Li9e234ee2017-03-27 10:51:41 -07002401
Tejun Heoa2b16932012-04-13 13:11:33 -07002402 /* activate policy */
Christoph Hellwig40e49962023-02-03 16:03:57 +01002403 ret = blkcg_activate_policy(disk, &blkcg_policy_throtl);
Shaohua Lib9147dd2017-03-27 15:19:42 -07002404 if (ret) {
Joseph Qib889bf62017-11-21 09:38:30 +08002405 free_percpu(td->latency_buckets[READ]);
2406 free_percpu(td->latency_buckets[WRITE]);
Vivek Goyal29b12582011-05-19 15:38:24 -04002407 kfree(td);
Shaohua Lib9147dd2017-03-27 15:19:42 -07002408 }
Tejun Heoa2b16932012-04-13 13:11:33 -07002409 return ret;
Vivek Goyale43473b2010-09-15 17:06:35 -04002410}
2411
Christoph Hellwige13793b2022-09-21 20:04:56 +02002412void blk_throtl_exit(struct gendisk *disk)
Vivek Goyale43473b2010-09-15 17:06:35 -04002413{
Christoph Hellwige13793b2022-09-21 20:04:56 +02002414 struct request_queue *q = disk->queue;
2415
Christoph Hellwigb4e94f92023-02-14 19:33:05 +01002416 BUG_ON(!q->td);
Li Jinlin884f0e82021-09-07 20:12:42 +08002417 del_timer_sync(&q->td->service_queue.pending_timer);
Vivek Goyalda527772011-03-02 19:05:33 -05002418 throtl_shutdown_wq(q);
Christoph Hellwig40e49962023-02-03 16:03:57 +01002419 blkcg_deactivate_policy(disk, &blkcg_policy_throtl);
Joseph Qib889bf62017-11-21 09:38:30 +08002420 free_percpu(q->td->latency_buckets[READ]);
2421 free_percpu(q->td->latency_buckets[WRITE]);
Tejun Heoc9a929d2011-10-19 14:42:16 +02002422 kfree(q->td);
Vivek Goyale43473b2010-09-15 17:06:35 -04002423}
2424
Christoph Hellwig5f6dc752022-09-21 20:04:57 +02002425void blk_throtl_register(struct gendisk *disk)
Shaohua Lid61fcfa2017-03-27 10:51:38 -07002426{
Christoph Hellwig5f6dc752022-09-21 20:04:57 +02002427 struct request_queue *q = disk->queue;
Shaohua Lid61fcfa2017-03-27 10:51:38 -07002428 struct throtl_data *td;
Shaohua Li6679a902017-06-06 12:40:43 -07002429 int i;
Shaohua Lid61fcfa2017-03-27 10:51:38 -07002430
2431 td = q->td;
2432 BUG_ON(!td);
2433
Shaohua Li6679a902017-06-06 12:40:43 -07002434 if (blk_queue_nonrot(q)) {
Shaohua Lid61fcfa2017-03-27 10:51:38 -07002435 td->throtl_slice = DFL_THROTL_SLICE_SSD;
Shaohua Li6679a902017-06-06 12:40:43 -07002436 td->filtered_latency = LATENCY_FILTERED_SSD;
2437 } else {
Shaohua Lid61fcfa2017-03-27 10:51:38 -07002438 td->throtl_slice = DFL_THROTL_SLICE_HD;
Shaohua Li6679a902017-06-06 12:40:43 -07002439 td->filtered_latency = LATENCY_FILTERED_HD;
Joseph Qib889bf62017-11-21 09:38:30 +08002440 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2441 td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2442 td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2443 }
Shaohua Li6679a902017-06-06 12:40:43 -07002444 }
Shaohua Lid61fcfa2017-03-27 10:51:38 -07002445#ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2446 /* if no low limit, use previous default */
2447 td->throtl_slice = DFL_THROTL_SLICE_HD;
Shaohua Li9e234ee2017-03-27 10:51:41 -07002448
Chengming Zhou8e15dfb2023-04-13 14:28:05 +08002449#else
Jens Axboe344e9ff2018-11-15 12:22:51 -07002450 td->track_bio_latency = !queue_is_mq(q);
Shaohua Lib9147dd2017-03-27 15:19:42 -07002451 if (!td->track_bio_latency)
2452 blk_stat_enable_accounting(q);
Chengming Zhou8e15dfb2023-04-13 14:28:05 +08002453#endif
Shaohua Lid61fcfa2017-03-27 10:51:38 -07002454}
2455
Shaohua Li297e3d82017-03-27 10:51:37 -07002456#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2457ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2458{
2459 if (!q->td)
2460 return -EINVAL;
2461 return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2462}
2463
2464ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2465 const char *page, size_t count)
2466{
2467 unsigned long v;
2468 unsigned long t;
2469
2470 if (!q->td)
2471 return -EINVAL;
2472 if (kstrtoul(page, 10, &v))
2473 return -EINVAL;
2474 t = msecs_to_jiffies(v);
2475 if (t == 0 || t > MAX_THROTL_SLICE)
2476 return -EINVAL;
2477 q->td->throtl_slice = t;
2478 return count;
2479}
2480#endif
2481
Vivek Goyale43473b2010-09-15 17:06:35 -04002482static int __init throtl_init(void)
2483{
Vivek Goyal450adcb2011-03-01 13:40:54 -05002484 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2485 if (!kthrotld_workqueue)
2486 panic("Failed to create kthrotld\n");
2487
Tejun Heo3c798392012-04-16 13:57:25 -07002488 return blkcg_policy_register(&blkcg_policy_throtl);
Vivek Goyale43473b2010-09-15 17:06:35 -04002489}
2490
2491module_init(throtl_init);