Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
e43473b7 VG |
2 | /* |
3 | * Interface for controlling IO bandwidth on a request queue | |
4 | * | |
5 | * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com> | |
6 | */ | |
7 | ||
8 | #include <linux/module.h> | |
9 | #include <linux/slab.h> | |
10 | #include <linux/blkdev.h> | |
11 | #include <linux/bio.h> | |
12 | #include <linux/blktrace_api.h> | |
bc9fcbf9 | 13 | #include "blk.h" |
1d156646 | 14 | #include "blk-cgroup-rwstat.h" |
e4a19f72 | 15 | #include "blk-stat.h" |
a7b36ee6 | 16 | #include "blk-throttle.h" |
e43473b7 VG |
17 | |
18 | /* Max dispatch from a group in 1 round */ | |
e675df2a | 19 | #define THROTL_GRP_QUANTUM 8 |
e43473b7 VG |
20 | |
21 | /* Total max dispatch from all groups in one round */ | |
e675df2a | 22 | #define THROTL_QUANTUM 32 |
e43473b7 | 23 | |
d61fcfa4 SL |
24 | /* Throttling is performed over a slice and after that slice is renewed */ |
25 | #define DFL_THROTL_SLICE_HD (HZ / 10) | |
26 | #define DFL_THROTL_SLICE_SSD (HZ / 50) | |
297e3d85 | 27 | #define MAX_THROTL_SLICE (HZ) |
9e234eea | 28 | #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */ |
9bb67aeb SL |
29 | #define MIN_THROTL_BPS (320 * 1024) |
30 | #define MIN_THROTL_IOPS (10) | |
b4f428ef SL |
31 | #define DFL_LATENCY_TARGET (-1L) |
32 | #define DFL_IDLE_THRESHOLD (0) | |
6679a90c SL |
33 | #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 */ | |
e43473b7 | 40 | |
450adcbe VG |
41 | /* A workqueue to queue throttle related work */ |
42 | static struct workqueue_struct *kthrotld_workqueue; | |
450adcbe | 43 | |
e43473b7 VG |
44 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
45 | ||
b9147dd1 SL |
46 | /* We measure latency for request size from <= 4k to >= 1M */ |
47 | #define LATENCY_BUCKET_SIZE 9 | |
48 | ||
49 | struct latency_bucket { | |
50 | unsigned long total_latency; /* ns / 1024 */ | |
51 | int samples; | |
52 | }; | |
53 | ||
54 | struct avg_latency_bucket { | |
55 | unsigned long latency; /* ns / 1024 */ | |
56 | bool valid; | |
57 | }; | |
58 | ||
e43473b7 VG |
59 | struct throtl_data |
60 | { | |
e43473b7 | 61 | /* service tree for active throtl groups */ |
c9e0332e | 62 | struct throtl_service_queue service_queue; |
e43473b7 | 63 | |
e43473b7 VG |
64 | struct request_queue *queue; |
65 | ||
66 | /* Total Number of queued bios on READ and WRITE lists */ | |
67 | unsigned int nr_queued[2]; | |
68 | ||
297e3d85 SL |
69 | unsigned int throtl_slice; |
70 | ||
e43473b7 | 71 | /* Work for dispatching throttled bios */ |
69df0ab0 | 72 | struct work_struct dispatch_work; |
9f626e37 SL |
73 | unsigned int limit_index; |
74 | bool limit_valid[LIMIT_CNT]; | |
3f0abd80 SL |
75 | |
76 | unsigned long low_upgrade_time; | |
77 | unsigned long low_downgrade_time; | |
7394e31f SL |
78 | |
79 | unsigned int scale; | |
b9147dd1 | 80 | |
b889bf66 JQ |
81 | 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]; | |
b9147dd1 | 84 | unsigned long last_calculate_time; |
6679a90c | 85 | unsigned long filtered_latency; |
b9147dd1 SL |
86 | |
87 | bool track_bio_latency; | |
e43473b7 VG |
88 | }; |
89 | ||
e99e88a9 | 90 | static void throtl_pending_timer_fn(struct timer_list *t); |
69df0ab0 | 91 | |
3c798398 | 92 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 93 | { |
f95a04af | 94 | return pd_to_blkg(&tg->pd); |
0381411e TH |
95 | } |
96 | ||
fda6f272 TH |
97 | /** |
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 | */ | |
104 | static 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 | } | |
111 | ||
112 | /** | |
113 | * sq_to_td - return throtl_data the specified service queue belongs to | |
114 | * @sq: the throtl_service_queue of interest | |
115 | * | |
b43daedc | 116 | * A service_queue can be embedded in either a throtl_grp or throtl_data. |
fda6f272 TH |
117 | * Determine the associated throtl_data accordingly and return it. |
118 | */ | |
119 | static 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 | ||
7394e31f SL |
129 | /* |
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. | |
132 | * Scale up: linearly scale up according to lapsed time since upgrade. For | |
133 | * 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 | */ | |
137 | static 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 | ||
9f626e37 SL |
147 | static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw) |
148 | { | |
b22c417c | 149 | struct blkcg_gq *blkg = tg_to_blkg(tg); |
7394e31f | 150 | struct throtl_data *td; |
b22c417c SL |
151 | uint64_t ret; |
152 | ||
153 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
154 | return U64_MAX; | |
7394e31f SL |
155 | |
156 | td = tg->td; | |
157 | ret = tg->bps[rw][td->limit_index]; | |
9bb67aeb SL |
158 | 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 | } | |
7394e31f SL |
166 | |
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 | } | |
b22c417c | 174 | return ret; |
9f626e37 SL |
175 | } |
176 | ||
177 | static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw) | |
178 | { | |
b22c417c | 179 | struct blkcg_gq *blkg = tg_to_blkg(tg); |
7394e31f | 180 | struct throtl_data *td; |
b22c417c SL |
181 | unsigned int ret; |
182 | ||
183 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
184 | return UINT_MAX; | |
9bb67aeb | 185 | |
7394e31f SL |
186 | td = tg->td; |
187 | ret = tg->iops[rw][td->limit_index]; | |
9bb67aeb SL |
188 | 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 | } | |
7394e31f SL |
196 | |
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 | } | |
b22c417c | 206 | return ret; |
9f626e37 SL |
207 | } |
208 | ||
b9147dd1 SL |
209 | #define request_bucket_index(sectors) \ |
210 | clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1) | |
211 | ||
fda6f272 TH |
212 | /** |
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". | |
fda6f272 TH |
220 | */ |
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; \ | |
59fa0224 SL |
226 | if (likely(!blk_trace_note_message_enabled(__td->queue))) \ |
227 | break; \ | |
fda6f272 | 228 | if ((__tg)) { \ |
35fe6d76 | 229 | blk_add_cgroup_trace_msg(__td->queue, \ |
f4a6a61c | 230 | &tg_to_blkg(__tg)->blkcg->css, "throtl " fmt, ##args);\ |
fda6f272 TH |
231 | } else { \ |
232 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
233 | } \ | |
54e7ed12 | 234 | } while (0) |
e43473b7 | 235 | |
ea0ea2bc SL |
236 | static 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 | ||
c5cc2070 TH |
244 | static 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 | */ | |
261 | static 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 | */ | |
275 | static struct bio *throtl_peek_queued(struct list_head *queued) | |
276 | { | |
b7b609de | 277 | struct throtl_qnode *qn; |
c5cc2070 TH |
278 | struct bio *bio; |
279 | ||
280 | if (list_empty(queued)) | |
281 | return NULL; | |
282 | ||
b7b609de | 283 | qn = list_first_entry(queued, struct throtl_qnode, node); |
c5cc2070 TH |
284 | 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 | */ | |
303 | static struct bio *throtl_pop_queued(struct list_head *queued, | |
304 | struct throtl_grp **tg_to_put) | |
305 | { | |
b7b609de | 306 | struct throtl_qnode *qn; |
c5cc2070 TH |
307 | struct bio *bio; |
308 | ||
309 | if (list_empty(queued)) | |
310 | return NULL; | |
311 | ||
b7b609de | 312 | qn = list_first_entry(queued, struct throtl_qnode, node); |
c5cc2070 TH |
313 | 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 | ||
49a2f1e3 | 329 | /* init a service_queue, assumes the caller zeroed it */ |
b2ce2643 | 330 | static void throtl_service_queue_init(struct throtl_service_queue *sq) |
49a2f1e3 | 331 | { |
7e9c5c54 YK |
332 | INIT_LIST_HEAD(&sq->queued[READ]); |
333 | INIT_LIST_HEAD(&sq->queued[WRITE]); | |
9ff01255 | 334 | sq->pending_tree = RB_ROOT_CACHED; |
e99e88a9 | 335 | timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0); |
69df0ab0 TH |
336 | } |
337 | ||
cf09a8ee TH |
338 | static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, |
339 | struct request_queue *q, | |
340 | struct blkcg *blkcg) | |
001bea73 | 341 | { |
4fb72036 | 342 | struct throtl_grp *tg; |
24bdb8ef | 343 | int rw; |
4fb72036 | 344 | |
cf09a8ee | 345 | tg = kzalloc_node(sizeof(*tg), gfp, q->node); |
4fb72036 | 346 | if (!tg) |
77ea7338 | 347 | return NULL; |
4fb72036 | 348 | |
7ca46438 TH |
349 | if (blkg_rwstat_init(&tg->stat_bytes, gfp)) |
350 | goto err_free_tg; | |
351 | ||
352 | if (blkg_rwstat_init(&tg->stat_ios, gfp)) | |
353 | goto err_exit_stat_bytes; | |
354 | ||
b2ce2643 TH |
355 | throtl_service_queue_init(&tg->service_queue); |
356 | ||
357 | for (rw = READ; rw <= WRITE; rw++) { | |
358 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); | |
359 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); | |
360 | } | |
361 | ||
362 | RB_CLEAR_NODE(&tg->rb_node); | |
9f626e37 SL |
363 | tg->bps[READ][LIMIT_MAX] = U64_MAX; |
364 | tg->bps[WRITE][LIMIT_MAX] = U64_MAX; | |
365 | tg->iops[READ][LIMIT_MAX] = UINT_MAX; | |
366 | tg->iops[WRITE][LIMIT_MAX] = UINT_MAX; | |
cd5ab1b0 SL |
367 | tg->bps_conf[READ][LIMIT_MAX] = U64_MAX; |
368 | tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX; | |
369 | tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX; | |
370 | tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX; | |
371 | /* LIMIT_LOW will have default value 0 */ | |
b2ce2643 | 372 | |
ec80991d | 373 | tg->latency_target = DFL_LATENCY_TARGET; |
5b81fc3c | 374 | tg->latency_target_conf = DFL_LATENCY_TARGET; |
b4f428ef SL |
375 | tg->idletime_threshold = DFL_IDLE_THRESHOLD; |
376 | tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD; | |
ec80991d | 377 | |
4fb72036 | 378 | return &tg->pd; |
7ca46438 TH |
379 | |
380 | err_exit_stat_bytes: | |
381 | blkg_rwstat_exit(&tg->stat_bytes); | |
382 | err_free_tg: | |
383 | kfree(tg); | |
384 | return NULL; | |
001bea73 TH |
385 | } |
386 | ||
a9520cd6 | 387 | static void throtl_pd_init(struct blkg_policy_data *pd) |
a29a171e | 388 | { |
a9520cd6 TH |
389 | struct throtl_grp *tg = pd_to_tg(pd); |
390 | struct blkcg_gq *blkg = tg_to_blkg(tg); | |
77216b04 | 391 | struct throtl_data *td = blkg->q->td; |
b2ce2643 | 392 | struct throtl_service_queue *sq = &tg->service_queue; |
cd1604fa | 393 | |
9138125b | 394 | /* |
aa6ec29b | 395 | * If on the default hierarchy, we switch to properly hierarchical |
9138125b TH |
396 | * behavior where limits on a given throtl_grp are applied to the |
397 | * whole subtree rather than just the group itself. e.g. If 16M | |
398 | * read_bps limit is set on the root group, the whole system can't | |
399 | * exceed 16M for the device. | |
400 | * | |
aa6ec29b | 401 | * If not on the default hierarchy, the broken flat hierarchy |
9138125b TH |
402 | * 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 | */ | |
b2ce2643 | 407 | sq->parent_sq = &td->service_queue; |
9e10a130 | 408 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent) |
b2ce2643 | 409 | sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue; |
77216b04 | 410 | tg->td = td; |
8a3d2615 TH |
411 | } |
412 | ||
693e751e TH |
413 | /* |
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 | */ | |
418 | static void tg_update_has_rules(struct throtl_grp *tg) | |
419 | { | |
420 | struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); | |
9f626e37 | 421 | struct throtl_data *td = tg->td; |
693e751e TH |
422 | int rw; |
423 | ||
81c7a63a YK |
424 | for (rw = READ; rw <= WRITE; rw++) { |
425 | tg->has_rules_iops[rw] = | |
426 | (parent_tg && parent_tg->has_rules_iops[rw]) || | |
427 | (td->limit_valid[td->limit_index] && | |
428 | tg_iops_limit(tg, rw) != UINT_MAX); | |
429 | tg->has_rules_bps[rw] = | |
430 | (parent_tg && parent_tg->has_rules_bps[rw]) || | |
9f626e37 | 431 | (td->limit_valid[td->limit_index] && |
81c7a63a YK |
432 | (tg_bps_limit(tg, rw) != U64_MAX)); |
433 | } | |
693e751e TH |
434 | } |
435 | ||
a9520cd6 | 436 | static void throtl_pd_online(struct blkg_policy_data *pd) |
693e751e | 437 | { |
aec24246 | 438 | struct throtl_grp *tg = pd_to_tg(pd); |
693e751e TH |
439 | /* |
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 | */ | |
aec24246 | 443 | tg_update_has_rules(tg); |
693e751e TH |
444 | } |
445 | ||
acaf523a | 446 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
cd5ab1b0 SL |
447 | static 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(); | |
454 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
455 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
456 | ||
457 | if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] || | |
43ada787 | 458 | tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) { |
cd5ab1b0 | 459 | low_valid = true; |
43ada787 LB |
460 | break; |
461 | } | |
cd5ab1b0 SL |
462 | } |
463 | rcu_read_unlock(); | |
464 | ||
465 | td->limit_valid[LIMIT_LOW] = low_valid; | |
466 | } | |
acaf523a YK |
467 | #else |
468 | static inline void blk_throtl_update_limit_valid(struct throtl_data *td) | |
469 | { | |
470 | } | |
471 | #endif | |
cd5ab1b0 | 472 | |
c79892c5 | 473 | static void throtl_upgrade_state(struct throtl_data *td); |
cd5ab1b0 SL |
474 | static 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 | ||
c79892c5 SL |
485 | if (!tg->td->limit_valid[tg->td->limit_index]) |
486 | throtl_upgrade_state(tg->td); | |
cd5ab1b0 SL |
487 | } |
488 | ||
001bea73 TH |
489 | static void throtl_pd_free(struct blkg_policy_data *pd) |
490 | { | |
4fb72036 TH |
491 | struct throtl_grp *tg = pd_to_tg(pd); |
492 | ||
b2ce2643 | 493 | del_timer_sync(&tg->service_queue.pending_timer); |
7ca46438 TH |
494 | blkg_rwstat_exit(&tg->stat_bytes); |
495 | blkg_rwstat_exit(&tg->stat_ios); | |
4fb72036 | 496 | kfree(tg); |
001bea73 TH |
497 | } |
498 | ||
0049af73 TH |
499 | static struct throtl_grp * |
500 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 | 501 | { |
9ff01255 | 502 | struct rb_node *n; |
e43473b7 | 503 | |
9ff01255 LB |
504 | 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); | |
e43473b7 VG |
509 | } |
510 | ||
0049af73 TH |
511 | static void throtl_rb_erase(struct rb_node *n, |
512 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 513 | { |
9ff01255 LB |
514 | rb_erase_cached(n, &parent_sq->pending_tree); |
515 | RB_CLEAR_NODE(n); | |
e43473b7 VG |
516 | } |
517 | ||
0049af73 | 518 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
519 | { |
520 | struct throtl_grp *tg; | |
521 | ||
0049af73 | 522 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
523 | if (!tg) |
524 | return; | |
525 | ||
0049af73 | 526 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
527 | } |
528 | ||
77216b04 | 529 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 530 | { |
77216b04 | 531 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
9ff01255 | 532 | struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node; |
e43473b7 VG |
533 | struct rb_node *parent = NULL; |
534 | struct throtl_grp *__tg; | |
535 | unsigned long key = tg->disptime; | |
9ff01255 | 536 | bool leftmost = true; |
e43473b7 VG |
537 | |
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; | |
9ff01255 | 546 | leftmost = false; |
e43473b7 VG |
547 | } |
548 | } | |
549 | ||
e43473b7 | 550 | rb_link_node(&tg->rb_node, parent, node); |
9ff01255 LB |
551 | rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree, |
552 | leftmost); | |
e43473b7 VG |
553 | } |
554 | ||
77216b04 | 555 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 556 | { |
29379674 BW |
557 | 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 | } | |
e43473b7 VG |
562 | } |
563 | ||
77216b04 | 564 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 565 | { |
29379674 | 566 | if (tg->flags & THROTL_TG_PENDING) { |
c013710e YK |
567 | 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; | |
29379674 BW |
572 | tg->flags &= ~THROTL_TG_PENDING; |
573 | } | |
e43473b7 VG |
574 | } |
575 | ||
a9131a27 | 576 | /* Call with queue lock held */ |
69df0ab0 TH |
577 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, |
578 | unsigned long expires) | |
a9131a27 | 579 | { |
a41b816c | 580 | unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice; |
06cceedc SL |
581 | |
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; | |
69df0ab0 TH |
591 | mod_timer(&sq->pending_timer, expires); |
592 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
593 | expires - jiffies, jiffies); | |
a9131a27 TH |
594 | } |
595 | ||
7f52f98c TH |
596 | /** |
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 | */ | |
614 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
615 | bool force) | |
e43473b7 | 616 | { |
6a525600 | 617 | /* any pending children left? */ |
c9e0332e | 618 | if (!sq->nr_pending) |
7f52f98c | 619 | return true; |
e43473b7 | 620 | |
c9e0332e | 621 | update_min_dispatch_time(sq); |
e43473b7 | 622 | |
69df0ab0 | 623 | /* is the next dispatch time in the future? */ |
7f52f98c | 624 | if (force || time_after(sq->first_pending_disptime, jiffies)) { |
69df0ab0 | 625 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); |
7f52f98c | 626 | return true; |
69df0ab0 TH |
627 | } |
628 | ||
7f52f98c TH |
629 | /* tell the caller to continue dispatching */ |
630 | return false; | |
e43473b7 VG |
631 | } |
632 | ||
32ee5bc4 VG |
633 | static 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; | |
a880ae93 YK |
638 | tg->carryover_bytes[rw] = 0; |
639 | tg->carryover_ios[rw] = 0; | |
32ee5bc4 VG |
640 | |
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 | */ | |
647 | if (time_after_eq(start, tg->slice_start[rw])) | |
648 | tg->slice_start[rw] = start; | |
649 | ||
297e3d85 | 650 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; |
32ee5bc4 VG |
651 | 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 | ||
a880ae93 YK |
657 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw, |
658 | bool clear_carryover) | |
e43473b7 VG |
659 | { |
660 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 661 | tg->io_disp[rw] = 0; |
e43473b7 | 662 | tg->slice_start[rw] = jiffies; |
297e3d85 | 663 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; |
a880ae93 YK |
664 | if (clear_carryover) { |
665 | tg->carryover_bytes[rw] = 0; | |
666 | tg->carryover_ios[rw] = 0; | |
667 | } | |
4f1e9630 | 668 | |
fda6f272 TH |
669 | 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); | |
e43473b7 VG |
673 | } |
674 | ||
0f3457f6 TH |
675 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
676 | unsigned long jiffy_end) | |
d1ae8ffd | 677 | { |
297e3d85 | 678 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); |
d1ae8ffd VG |
679 | } |
680 | ||
0f3457f6 TH |
681 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
682 | unsigned long jiffy_end) | |
e43473b7 | 683 | { |
1da30f95 | 684 | throtl_set_slice_end(tg, rw, jiffy_end); |
fda6f272 TH |
685 | 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); | |
e43473b7 VG |
689 | } |
690 | ||
691 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 692 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
693 | { |
694 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
5cf8c227 | 695 | return false; |
e43473b7 | 696 | |
0b6bad7d | 697 | return true; |
e43473b7 VG |
698 | } |
699 | ||
700 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 701 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 702 | { |
3aad5d3e VG |
703 | unsigned long nr_slices, time_elapsed, io_trim; |
704 | u64 bytes_trim, tmp; | |
e43473b7 VG |
705 | |
706 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
707 | ||
708 | /* | |
709 | * If bps are unlimited (-1), then time slice don't get | |
710 | * renewed. Don't try to trim the slice if slice is used. A new | |
711 | * slice will start when appropriate. | |
712 | */ | |
0f3457f6 | 713 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
714 | return; |
715 | ||
d1ae8ffd VG |
716 | /* |
717 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
718 | * that initially cgroup limit was very low resulting in high | |
b53b072c | 719 | * slice_end, but later limit was bumped up and bio was dispatched |
d1ae8ffd VG |
720 | * sooner, then we need to reduce slice_end. A high bogus slice_end |
721 | * is bad because it does not allow new slice to start. | |
722 | */ | |
723 | ||
297e3d85 | 724 | throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice); |
d1ae8ffd | 725 | |
e43473b7 VG |
726 | time_elapsed = jiffies - tg->slice_start[rw]; |
727 | ||
297e3d85 | 728 | nr_slices = time_elapsed / tg->td->throtl_slice; |
e43473b7 VG |
729 | |
730 | if (!nr_slices) | |
731 | return; | |
297e3d85 | 732 | tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices; |
3aad5d3e VG |
733 | do_div(tmp, HZ); |
734 | bytes_trim = tmp; | |
e43473b7 | 735 | |
297e3d85 SL |
736 | io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) / |
737 | HZ; | |
e43473b7 | 738 | |
8e89d13f | 739 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
740 | return; |
741 | ||
742 | if (tg->bytes_disp[rw] >= bytes_trim) | |
743 | tg->bytes_disp[rw] -= bytes_trim; | |
744 | else | |
745 | tg->bytes_disp[rw] = 0; | |
746 | ||
8e89d13f VG |
747 | if (tg->io_disp[rw] >= io_trim) |
748 | tg->io_disp[rw] -= io_trim; | |
749 | else | |
750 | tg->io_disp[rw] = 0; | |
751 | ||
297e3d85 | 752 | tg->slice_start[rw] += nr_slices * tg->td->throtl_slice; |
e43473b7 | 753 | |
fda6f272 TH |
754 | throtl_log(&tg->service_queue, |
755 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
756 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
757 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
e43473b7 VG |
758 | } |
759 | ||
681cd46f YK |
760 | static unsigned int calculate_io_allowed(u32 iops_limit, |
761 | unsigned long jiffy_elapsed) | |
e43473b7 | 762 | { |
8e89d13f | 763 | unsigned int io_allowed; |
c49c06e4 | 764 | u64 tmp; |
e43473b7 | 765 | |
c49c06e4 | 766 | /* |
681cd46f | 767 | * jiffy_elapsed should not be a big value as minimum iops can be |
c49c06e4 VG |
768 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we |
769 | * will allow dispatch after 1 second and after that slice should | |
770 | * have been trimmed. | |
771 | */ | |
772 | ||
681cd46f | 773 | tmp = (u64)iops_limit * jiffy_elapsed; |
c49c06e4 VG |
774 | do_div(tmp, HZ); |
775 | ||
776 | if (tmp > UINT_MAX) | |
777 | io_allowed = UINT_MAX; | |
778 | else | |
779 | io_allowed = tmp; | |
8e89d13f | 780 | |
681cd46f YK |
781 | return io_allowed; |
782 | } | |
783 | ||
784 | static u64 calculate_bytes_allowed(u64 bps_limit, unsigned long jiffy_elapsed) | |
785 | { | |
786 | return mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed, (u64)HZ); | |
787 | } | |
788 | ||
a880ae93 YK |
789 | static void __tg_update_carryover(struct throtl_grp *tg, bool rw) |
790 | { | |
791 | unsigned long jiffy_elapsed = jiffies - tg->slice_start[rw]; | |
792 | u64 bps_limit = tg_bps_limit(tg, rw); | |
793 | u32 iops_limit = tg_iops_limit(tg, rw); | |
794 | ||
795 | /* | |
796 | * If config is updated while bios are still throttled, calculate and | |
797 | * accumulate how many bytes/ios are waited across changes. And | |
798 | * carryover_bytes/ios will be used to calculate new wait time under new | |
799 | * configuration. | |
800 | */ | |
801 | if (bps_limit != U64_MAX) | |
802 | tg->carryover_bytes[rw] += | |
803 | calculate_bytes_allowed(bps_limit, jiffy_elapsed) - | |
804 | tg->bytes_disp[rw]; | |
805 | if (iops_limit != UINT_MAX) | |
806 | tg->carryover_ios[rw] += | |
807 | calculate_io_allowed(iops_limit, jiffy_elapsed) - | |
808 | tg->io_disp[rw]; | |
809 | } | |
810 | ||
811 | static void tg_update_carryover(struct throtl_grp *tg) | |
812 | { | |
813 | if (tg->service_queue.nr_queued[READ]) | |
814 | __tg_update_carryover(tg, READ); | |
815 | if (tg->service_queue.nr_queued[WRITE]) | |
816 | __tg_update_carryover(tg, WRITE); | |
817 | ||
818 | /* see comments in struct throtl_grp for meaning of these fields. */ | |
819 | throtl_log(&tg->service_queue, "%s: %llu %llu %u %u\n", __func__, | |
820 | tg->carryover_bytes[READ], tg->carryover_bytes[WRITE], | |
821 | tg->carryover_ios[READ], tg->carryover_ios[WRITE]); | |
822 | } | |
823 | ||
681cd46f YK |
824 | static bool tg_within_iops_limit(struct throtl_grp *tg, struct bio *bio, |
825 | u32 iops_limit, unsigned long *wait) | |
826 | { | |
827 | bool rw = bio_data_dir(bio); | |
828 | unsigned int io_allowed; | |
829 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; | |
830 | ||
831 | if (iops_limit == UINT_MAX) { | |
832 | if (wait) | |
833 | *wait = 0; | |
834 | return true; | |
835 | } | |
836 | ||
837 | jiffy_elapsed = jiffies - tg->slice_start[rw]; | |
838 | ||
839 | /* Round up to the next throttle slice, wait time must be nonzero */ | |
840 | jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice); | |
a880ae93 YK |
841 | io_allowed = calculate_io_allowed(iops_limit, jiffy_elapsed_rnd) + |
842 | tg->carryover_ios[rw]; | |
8e89d13f | 843 | if (tg->io_disp[rw] + 1 <= io_allowed) { |
e43473b7 VG |
844 | if (wait) |
845 | *wait = 0; | |
5cf8c227 | 846 | return true; |
e43473b7 VG |
847 | } |
848 | ||
8e89d13f | 849 | /* Calc approx time to dispatch */ |
991f61fe | 850 | jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed; |
8e89d13f VG |
851 | |
852 | if (wait) | |
853 | *wait = jiffy_wait; | |
0b6bad7d | 854 | return false; |
8e89d13f VG |
855 | } |
856 | ||
681cd46f YK |
857 | static bool tg_within_bps_limit(struct throtl_grp *tg, struct bio *bio, |
858 | u64 bps_limit, unsigned long *wait) | |
8e89d13f VG |
859 | { |
860 | bool rw = bio_data_dir(bio); | |
8d6bbaad | 861 | u64 bytes_allowed, extra_bytes; |
8e89d13f | 862 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
ea0ea2bc | 863 | unsigned int bio_size = throtl_bio_data_size(bio); |
e43473b7 | 864 | |
9f5ede3c | 865 | /* no need to throttle if this bio's bytes have been accounted */ |
320fb0f9 | 866 | if (bps_limit == U64_MAX || bio_flagged(bio, BIO_BPS_THROTTLED)) { |
87fbeb88 BW |
867 | if (wait) |
868 | *wait = 0; | |
869 | return true; | |
870 | } | |
871 | ||
e43473b7 VG |
872 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
873 | ||
874 | /* Slice has just started. Consider one slice interval */ | |
875 | if (!jiffy_elapsed) | |
297e3d85 | 876 | jiffy_elapsed_rnd = tg->td->throtl_slice; |
e43473b7 | 877 | |
297e3d85 | 878 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); |
a880ae93 YK |
879 | bytes_allowed = calculate_bytes_allowed(bps_limit, jiffy_elapsed_rnd) + |
880 | tg->carryover_bytes[rw]; | |
ea0ea2bc | 881 | if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) { |
e43473b7 VG |
882 | if (wait) |
883 | *wait = 0; | |
5cf8c227 | 884 | return true; |
e43473b7 VG |
885 | } |
886 | ||
887 | /* Calc approx time to dispatch */ | |
ea0ea2bc | 888 | extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed; |
4599ea49 | 889 | jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit); |
e43473b7 VG |
890 | |
891 | if (!jiffy_wait) | |
892 | jiffy_wait = 1; | |
893 | ||
894 | /* | |
895 | * This wait time is without taking into consideration the rounding | |
896 | * up we did. Add that time also. | |
897 | */ | |
898 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
899 | if (wait) |
900 | *wait = jiffy_wait; | |
0b6bad7d | 901 | return false; |
8e89d13f VG |
902 | } |
903 | ||
904 | /* | |
905 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
906 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
907 | */ | |
0f3457f6 TH |
908 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
909 | unsigned long *wait) | |
8e89d13f VG |
910 | { |
911 | bool rw = bio_data_dir(bio); | |
912 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
4599ea49 BW |
913 | u64 bps_limit = tg_bps_limit(tg, rw); |
914 | u32 iops_limit = tg_iops_limit(tg, rw); | |
8e89d13f VG |
915 | |
916 | /* | |
917 | * Currently whole state machine of group depends on first bio | |
918 | * queued in the group bio list. So one should not be calling | |
919 | * this function with a different bio if there are other bios | |
920 | * queued. | |
921 | */ | |
73f0d49a | 922 | BUG_ON(tg->service_queue.nr_queued[rw] && |
c5cc2070 | 923 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); |
e43473b7 | 924 | |
8e89d13f | 925 | /* If tg->bps = -1, then BW is unlimited */ |
8f9e7b65 YK |
926 | if ((bps_limit == U64_MAX && iops_limit == UINT_MAX) || |
927 | tg->flags & THROTL_TG_CANCELING) { | |
8e89d13f VG |
928 | if (wait) |
929 | *wait = 0; | |
5cf8c227 | 930 | return true; |
8e89d13f VG |
931 | } |
932 | ||
933 | /* | |
934 | * If previous slice expired, start a new one otherwise renew/extend | |
935 | * existing slice to make sure it is at least throtl_slice interval | |
164c80ed VG |
936 | * long since now. New slice is started only for empty throttle group. |
937 | * If there is queued bio, that means there should be an active | |
938 | * slice and it should be extended instead. | |
8e89d13f | 939 | */ |
164c80ed | 940 | if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw])) |
a880ae93 | 941 | throtl_start_new_slice(tg, rw, true); |
8e89d13f | 942 | else { |
297e3d85 SL |
943 | if (time_before(tg->slice_end[rw], |
944 | jiffies + tg->td->throtl_slice)) | |
945 | throtl_extend_slice(tg, rw, | |
946 | jiffies + tg->td->throtl_slice); | |
8e89d13f VG |
947 | } |
948 | ||
681cd46f YK |
949 | if (tg_within_bps_limit(tg, bio, bps_limit, &bps_wait) && |
950 | tg_within_iops_limit(tg, bio, iops_limit, &iops_wait)) { | |
8e89d13f VG |
951 | if (wait) |
952 | *wait = 0; | |
0b6bad7d | 953 | return true; |
8e89d13f VG |
954 | } |
955 | ||
956 | max_wait = max(bps_wait, iops_wait); | |
957 | ||
958 | if (wait) | |
959 | *wait = max_wait; | |
960 | ||
961 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 962 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 | 963 | |
0b6bad7d | 964 | return false; |
e43473b7 VG |
965 | } |
966 | ||
967 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
968 | { | |
969 | bool rw = bio_data_dir(bio); | |
ea0ea2bc | 970 | unsigned int bio_size = throtl_bio_data_size(bio); |
e43473b7 VG |
971 | |
972 | /* Charge the bio to the group */ | |
320fb0f9 | 973 | if (!bio_flagged(bio, BIO_BPS_THROTTLED)) { |
9f5ede3c ML |
974 | tg->bytes_disp[rw] += bio_size; |
975 | tg->last_bytes_disp[rw] += bio_size; | |
976 | } | |
977 | ||
8e89d13f | 978 | tg->io_disp[rw]++; |
3f0abd80 | 979 | tg->last_io_disp[rw]++; |
e43473b7 VG |
980 | } |
981 | ||
c5cc2070 TH |
982 | /** |
983 | * throtl_add_bio_tg - add a bio to the specified throtl_grp | |
984 | * @bio: bio to add | |
985 | * @qn: qnode to use | |
986 | * @tg: the target throtl_grp | |
987 | * | |
988 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, | |
989 | * tg->qnode_on_self[] is used. | |
990 | */ | |
991 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, | |
992 | struct throtl_grp *tg) | |
e43473b7 | 993 | { |
73f0d49a | 994 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
995 | bool rw = bio_data_dir(bio); |
996 | ||
c5cc2070 TH |
997 | if (!qn) |
998 | qn = &tg->qnode_on_self[rw]; | |
999 | ||
0e9f4164 TH |
1000 | /* |
1001 | * If @tg doesn't currently have any bios queued in the same | |
1002 | * direction, queueing @bio can change when @tg should be | |
1003 | * dispatched. Mark that @tg was empty. This is automatically | |
b53b072c | 1004 | * cleared on the next tg_update_disptime(). |
0e9f4164 TH |
1005 | */ |
1006 | if (!sq->nr_queued[rw]) | |
1007 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
1008 | ||
c5cc2070 TH |
1009 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); |
1010 | ||
73f0d49a | 1011 | sq->nr_queued[rw]++; |
77216b04 | 1012 | throtl_enqueue_tg(tg); |
e43473b7 VG |
1013 | } |
1014 | ||
77216b04 | 1015 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 1016 | { |
73f0d49a | 1017 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1018 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
1019 | struct bio *bio; | |
1020 | ||
d609af3a ME |
1021 | bio = throtl_peek_queued(&sq->queued[READ]); |
1022 | if (bio) | |
0f3457f6 | 1023 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 1024 | |
d609af3a ME |
1025 | bio = throtl_peek_queued(&sq->queued[WRITE]); |
1026 | if (bio) | |
0f3457f6 | 1027 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
1028 | |
1029 | min_wait = min(read_wait, write_wait); | |
1030 | disptime = jiffies + min_wait; | |
1031 | ||
e43473b7 | 1032 | /* Update dispatch time */ |
c013710e | 1033 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
e43473b7 | 1034 | tg->disptime = disptime; |
c013710e | 1035 | tg_service_queue_add(tg); |
0e9f4164 TH |
1036 | |
1037 | /* see throtl_add_bio_tg() */ | |
1038 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
1039 | } |
1040 | ||
32ee5bc4 VG |
1041 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, |
1042 | struct throtl_grp *parent_tg, bool rw) | |
1043 | { | |
1044 | if (throtl_slice_used(parent_tg, rw)) { | |
1045 | throtl_start_new_slice_with_credit(parent_tg, rw, | |
1046 | child_tg->slice_start[rw]); | |
1047 | } | |
1048 | ||
1049 | } | |
1050 | ||
77216b04 | 1051 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 1052 | { |
73f0d49a | 1053 | struct throtl_service_queue *sq = &tg->service_queue; |
6bc9c2b4 TH |
1054 | struct throtl_service_queue *parent_sq = sq->parent_sq; |
1055 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
c5cc2070 | 1056 | struct throtl_grp *tg_to_put = NULL; |
e43473b7 VG |
1057 | struct bio *bio; |
1058 | ||
c5cc2070 TH |
1059 | /* |
1060 | * @bio is being transferred from @tg to @parent_sq. Popping a bio | |
1061 | * from @tg may put its reference and @parent_sq might end up | |
1062 | * getting released prematurely. Remember the tg to put and put it | |
1063 | * after @bio is transferred to @parent_sq. | |
1064 | */ | |
1065 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); | |
73f0d49a | 1066 | sq->nr_queued[rw]--; |
e43473b7 VG |
1067 | |
1068 | throtl_charge_bio(tg, bio); | |
320fb0f9 | 1069 | bio_set_flag(bio, BIO_BPS_THROTTLED); |
6bc9c2b4 TH |
1070 | |
1071 | /* | |
1072 | * If our parent is another tg, we just need to transfer @bio to | |
1073 | * the parent using throtl_add_bio_tg(). If our parent is | |
1074 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
1075 | * bio_lists[] and decrease total number queued. The caller is | |
1076 | * responsible for issuing these bios. | |
1077 | */ | |
1078 | if (parent_tg) { | |
c5cc2070 | 1079 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); |
32ee5bc4 | 1080 | start_parent_slice_with_credit(tg, parent_tg, rw); |
6bc9c2b4 | 1081 | } else { |
c5cc2070 TH |
1082 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], |
1083 | &parent_sq->queued[rw]); | |
6bc9c2b4 TH |
1084 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
1085 | tg->td->nr_queued[rw]--; | |
1086 | } | |
e43473b7 | 1087 | |
0f3457f6 | 1088 | throtl_trim_slice(tg, rw); |
6bc9c2b4 | 1089 | |
c5cc2070 TH |
1090 | if (tg_to_put) |
1091 | blkg_put(tg_to_blkg(tg_to_put)); | |
e43473b7 VG |
1092 | } |
1093 | ||
77216b04 | 1094 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 1095 | { |
73f0d49a | 1096 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 | 1097 | unsigned int nr_reads = 0, nr_writes = 0; |
e675df2a BW |
1098 | unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4; |
1099 | unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads; | |
e43473b7 VG |
1100 | struct bio *bio; |
1101 | ||
1102 | /* Try to dispatch 75% READS and 25% WRITES */ | |
1103 | ||
c5cc2070 | 1104 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && |
0f3457f6 | 1105 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1106 | |
77216b04 | 1107 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1108 | nr_reads++; |
1109 | ||
1110 | if (nr_reads >= max_nr_reads) | |
1111 | break; | |
1112 | } | |
1113 | ||
c5cc2070 | 1114 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && |
0f3457f6 | 1115 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1116 | |
77216b04 | 1117 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1118 | nr_writes++; |
1119 | ||
1120 | if (nr_writes >= max_nr_writes) | |
1121 | break; | |
1122 | } | |
1123 | ||
1124 | return nr_reads + nr_writes; | |
1125 | } | |
1126 | ||
651930bc | 1127 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
1128 | { |
1129 | unsigned int nr_disp = 0; | |
e43473b7 VG |
1130 | |
1131 | while (1) { | |
2397611a | 1132 | struct throtl_grp *tg; |
2ab74cd2 | 1133 | struct throtl_service_queue *sq; |
e43473b7 | 1134 | |
2397611a BW |
1135 | if (!parent_sq->nr_pending) |
1136 | break; | |
1137 | ||
1138 | tg = throtl_rb_first(parent_sq); | |
e43473b7 VG |
1139 | if (!tg) |
1140 | break; | |
1141 | ||
1142 | if (time_before(jiffies, tg->disptime)) | |
1143 | break; | |
1144 | ||
77216b04 | 1145 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 1146 | |
2ab74cd2 | 1147 | sq = &tg->service_queue; |
7e9c5c54 | 1148 | if (sq->nr_queued[READ] || sq->nr_queued[WRITE]) |
77216b04 | 1149 | tg_update_disptime(tg); |
8c25ed0c YK |
1150 | else |
1151 | throtl_dequeue_tg(tg); | |
e43473b7 | 1152 | |
e675df2a | 1153 | if (nr_disp >= THROTL_QUANTUM) |
e43473b7 VG |
1154 | break; |
1155 | } | |
1156 | ||
1157 | return nr_disp; | |
1158 | } | |
1159 | ||
c79892c5 SL |
1160 | static bool throtl_can_upgrade(struct throtl_data *td, |
1161 | struct throtl_grp *this_tg); | |
6e1a5704 TH |
1162 | /** |
1163 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
216382dc | 1164 | * @t: the pending_timer member of the throtl_service_queue being serviced |
6e1a5704 TH |
1165 | * |
1166 | * This timer is armed when a child throtl_grp with active bio's become | |
1167 | * pending and queued on the service_queue's pending_tree and expires when | |
1168 | * the first child throtl_grp should be dispatched. This function | |
2e48a530 TH |
1169 | * dispatches bio's from the children throtl_grps to the parent |
1170 | * service_queue. | |
1171 | * | |
1172 | * If the parent's parent is another throtl_grp, dispatching is propagated | |
1173 | * by either arming its pending_timer or repeating dispatch directly. If | |
1174 | * the top-level service_tree is reached, throtl_data->dispatch_work is | |
1175 | * kicked so that the ready bio's are issued. | |
6e1a5704 | 1176 | */ |
e99e88a9 | 1177 | static void throtl_pending_timer_fn(struct timer_list *t) |
69df0ab0 | 1178 | { |
e99e88a9 | 1179 | struct throtl_service_queue *sq = from_timer(sq, t, pending_timer); |
2e48a530 | 1180 | struct throtl_grp *tg = sq_to_tg(sq); |
69df0ab0 | 1181 | struct throtl_data *td = sq_to_td(sq); |
2e48a530 | 1182 | struct throtl_service_queue *parent_sq; |
ee37eddb | 1183 | struct request_queue *q; |
2e48a530 | 1184 | bool dispatched; |
6e1a5704 | 1185 | int ret; |
e43473b7 | 1186 | |
ee37eddb ML |
1187 | /* throtl_data may be gone, so figure out request queue by blkg */ |
1188 | if (tg) | |
1189 | q = tg->pd.blkg->q; | |
1190 | else | |
1191 | q = td->queue; | |
1192 | ||
0d945c1f | 1193 | spin_lock_irq(&q->queue_lock); |
ee37eddb ML |
1194 | |
1195 | if (!q->root_blkg) | |
1196 | goto out_unlock; | |
1197 | ||
c79892c5 SL |
1198 | if (throtl_can_upgrade(td, NULL)) |
1199 | throtl_upgrade_state(td); | |
1200 | ||
2e48a530 TH |
1201 | again: |
1202 | parent_sq = sq->parent_sq; | |
1203 | dispatched = false; | |
e43473b7 | 1204 | |
7f52f98c TH |
1205 | while (true) { |
1206 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
2e48a530 TH |
1207 | sq->nr_queued[READ] + sq->nr_queued[WRITE], |
1208 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
7f52f98c TH |
1209 | |
1210 | ret = throtl_select_dispatch(sq); | |
1211 | if (ret) { | |
7f52f98c TH |
1212 | throtl_log(sq, "bios disp=%u", ret); |
1213 | dispatched = true; | |
1214 | } | |
e43473b7 | 1215 | |
7f52f98c TH |
1216 | if (throtl_schedule_next_dispatch(sq, false)) |
1217 | break; | |
e43473b7 | 1218 | |
7f52f98c | 1219 | /* this dispatch windows is still open, relax and repeat */ |
0d945c1f | 1220 | spin_unlock_irq(&q->queue_lock); |
7f52f98c | 1221 | cpu_relax(); |
0d945c1f | 1222 | spin_lock_irq(&q->queue_lock); |
651930bc | 1223 | } |
e43473b7 | 1224 | |
2e48a530 TH |
1225 | if (!dispatched) |
1226 | goto out_unlock; | |
6e1a5704 | 1227 | |
2e48a530 TH |
1228 | if (parent_sq) { |
1229 | /* @parent_sq is another throl_grp, propagate dispatch */ | |
1230 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1231 | tg_update_disptime(tg); | |
1232 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { | |
1233 | /* window is already open, repeat dispatching */ | |
1234 | sq = parent_sq; | |
1235 | tg = sq_to_tg(sq); | |
1236 | goto again; | |
1237 | } | |
1238 | } | |
1239 | } else { | |
b53b072c | 1240 | /* reached the top-level, queue issuing */ |
2e48a530 TH |
1241 | queue_work(kthrotld_workqueue, &td->dispatch_work); |
1242 | } | |
1243 | out_unlock: | |
0d945c1f | 1244 | spin_unlock_irq(&q->queue_lock); |
6e1a5704 | 1245 | } |
e43473b7 | 1246 | |
6e1a5704 TH |
1247 | /** |
1248 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
1249 | * @work: work item being executed | |
1250 | * | |
b53b072c BW |
1251 | * This function is queued for execution when bios reach the bio_lists[] |
1252 | * of throtl_data->service_queue. Those bios are ready and issued by this | |
6e1a5704 TH |
1253 | * function. |
1254 | */ | |
8876e140 | 1255 | static void blk_throtl_dispatch_work_fn(struct work_struct *work) |
6e1a5704 TH |
1256 | { |
1257 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1258 | dispatch_work); | |
1259 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1260 | struct request_queue *q = td->queue; | |
1261 | struct bio_list bio_list_on_stack; | |
1262 | struct bio *bio; | |
1263 | struct blk_plug plug; | |
1264 | int rw; | |
1265 | ||
1266 | bio_list_init(&bio_list_on_stack); | |
1267 | ||
0d945c1f | 1268 | spin_lock_irq(&q->queue_lock); |
c5cc2070 TH |
1269 | for (rw = READ; rw <= WRITE; rw++) |
1270 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) | |
1271 | bio_list_add(&bio_list_on_stack, bio); | |
0d945c1f | 1272 | spin_unlock_irq(&q->queue_lock); |
6e1a5704 TH |
1273 | |
1274 | if (!bio_list_empty(&bio_list_on_stack)) { | |
69d60eb9 | 1275 | blk_start_plug(&plug); |
ed00aabd | 1276 | while ((bio = bio_list_pop(&bio_list_on_stack))) |
3f98c753 | 1277 | submit_bio_noacct_nocheck(bio); |
69d60eb9 | 1278 | blk_finish_plug(&plug); |
e43473b7 | 1279 | } |
e43473b7 VG |
1280 | } |
1281 | ||
f95a04af TH |
1282 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
1283 | int off) | |
60c2bc2d | 1284 | { |
f95a04af TH |
1285 | struct throtl_grp *tg = pd_to_tg(pd); |
1286 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 1287 | |
2ab5492d | 1288 | if (v == U64_MAX) |
60c2bc2d | 1289 | return 0; |
f95a04af | 1290 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
1291 | } |
1292 | ||
f95a04af TH |
1293 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
1294 | int off) | |
e43473b7 | 1295 | { |
f95a04af TH |
1296 | struct throtl_grp *tg = pd_to_tg(pd); |
1297 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 1298 | |
2ab5492d | 1299 | if (v == UINT_MAX) |
af133ceb | 1300 | return 0; |
f95a04af | 1301 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
1302 | } |
1303 | ||
2da8ca82 | 1304 | static int tg_print_conf_u64(struct seq_file *sf, void *v) |
8e89d13f | 1305 | { |
2da8ca82 TH |
1306 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64, |
1307 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1308 | return 0; |
8e89d13f VG |
1309 | } |
1310 | ||
2da8ca82 | 1311 | static int tg_print_conf_uint(struct seq_file *sf, void *v) |
8e89d13f | 1312 | { |
2da8ca82 TH |
1313 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint, |
1314 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1315 | return 0; |
60c2bc2d TH |
1316 | } |
1317 | ||
9bb67aeb | 1318 | static void tg_conf_updated(struct throtl_grp *tg, bool global) |
60c2bc2d | 1319 | { |
69948b07 | 1320 | struct throtl_service_queue *sq = &tg->service_queue; |
492eb21b | 1321 | struct cgroup_subsys_state *pos_css; |
69948b07 | 1322 | struct blkcg_gq *blkg; |
af133ceb | 1323 | |
fda6f272 TH |
1324 | throtl_log(&tg->service_queue, |
1325 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
9f626e37 SL |
1326 | tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE), |
1327 | tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE)); | |
632b4493 | 1328 | |
693e751e TH |
1329 | /* |
1330 | * Update has_rules[] flags for the updated tg's subtree. A tg is | |
1331 | * considered to have rules if either the tg itself or any of its | |
1332 | * ancestors has rules. This identifies groups without any | |
1333 | * restrictions in the whole hierarchy and allows them to bypass | |
1334 | * blk-throttle. | |
1335 | */ | |
9bb67aeb SL |
1336 | blkg_for_each_descendant_pre(blkg, pos_css, |
1337 | global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) { | |
5b81fc3c SL |
1338 | struct throtl_grp *this_tg = blkg_to_tg(blkg); |
1339 | struct throtl_grp *parent_tg; | |
1340 | ||
1341 | tg_update_has_rules(this_tg); | |
1342 | /* ignore root/second level */ | |
1343 | if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent || | |
1344 | !blkg->parent->parent) | |
1345 | continue; | |
1346 | parent_tg = blkg_to_tg(blkg->parent); | |
1347 | /* | |
1348 | * make sure all children has lower idle time threshold and | |
1349 | * higher latency target | |
1350 | */ | |
1351 | this_tg->idletime_threshold = min(this_tg->idletime_threshold, | |
1352 | parent_tg->idletime_threshold); | |
1353 | this_tg->latency_target = max(this_tg->latency_target, | |
1354 | parent_tg->latency_target); | |
1355 | } | |
693e751e | 1356 | |
632b4493 TH |
1357 | /* |
1358 | * We're already holding queue_lock and know @tg is valid. Let's | |
1359 | * apply the new config directly. | |
1360 | * | |
1361 | * Restart the slices for both READ and WRITES. It might happen | |
1362 | * that a group's limit are dropped suddenly and we don't want to | |
1363 | * account recently dispatched IO with new low rate. | |
1364 | */ | |
a880ae93 YK |
1365 | throtl_start_new_slice(tg, READ, false); |
1366 | throtl_start_new_slice(tg, WRITE, false); | |
632b4493 | 1367 | |
5b2c16aa | 1368 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 1369 | tg_update_disptime(tg); |
7f52f98c | 1370 | throtl_schedule_next_dispatch(sq->parent_sq, true); |
632b4493 | 1371 | } |
69948b07 TH |
1372 | } |
1373 | ||
1374 | static ssize_t tg_set_conf(struct kernfs_open_file *of, | |
1375 | char *buf, size_t nbytes, loff_t off, bool is_u64) | |
1376 | { | |
1377 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1378 | struct blkg_conf_ctx ctx; | |
1379 | struct throtl_grp *tg; | |
1380 | int ret; | |
1381 | u64 v; | |
1382 | ||
1383 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1384 | if (ret) | |
1385 | return ret; | |
1386 | ||
1387 | ret = -EINVAL; | |
1388 | if (sscanf(ctx.body, "%llu", &v) != 1) | |
1389 | goto out_finish; | |
1390 | if (!v) | |
2ab5492d | 1391 | v = U64_MAX; |
69948b07 TH |
1392 | |
1393 | tg = blkg_to_tg(ctx.blkg); | |
a880ae93 | 1394 | tg_update_carryover(tg); |
69948b07 TH |
1395 | |
1396 | if (is_u64) | |
1397 | *(u64 *)((void *)tg + of_cft(of)->private) = v; | |
1398 | else | |
1399 | *(unsigned int *)((void *)tg + of_cft(of)->private) = v; | |
60c2bc2d | 1400 | |
9bb67aeb | 1401 | tg_conf_updated(tg, false); |
36aa9e5f TH |
1402 | ret = 0; |
1403 | out_finish: | |
60c2bc2d | 1404 | blkg_conf_finish(&ctx); |
36aa9e5f | 1405 | return ret ?: nbytes; |
8e89d13f VG |
1406 | } |
1407 | ||
451af504 TH |
1408 | static ssize_t tg_set_conf_u64(struct kernfs_open_file *of, |
1409 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1410 | { |
451af504 | 1411 | return tg_set_conf(of, buf, nbytes, off, true); |
60c2bc2d TH |
1412 | } |
1413 | ||
451af504 TH |
1414 | static ssize_t tg_set_conf_uint(struct kernfs_open_file *of, |
1415 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1416 | { |
451af504 | 1417 | return tg_set_conf(of, buf, nbytes, off, false); |
60c2bc2d TH |
1418 | } |
1419 | ||
7ca46438 TH |
1420 | static int tg_print_rwstat(struct seq_file *sf, void *v) |
1421 | { | |
1422 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
1423 | blkg_prfill_rwstat, &blkcg_policy_throtl, | |
1424 | seq_cft(sf)->private, true); | |
1425 | return 0; | |
1426 | } | |
1427 | ||
1428 | static u64 tg_prfill_rwstat_recursive(struct seq_file *sf, | |
1429 | struct blkg_policy_data *pd, int off) | |
1430 | { | |
1431 | struct blkg_rwstat_sample sum; | |
1432 | ||
1433 | blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off, | |
1434 | &sum); | |
1435 | return __blkg_prfill_rwstat(sf, pd, &sum); | |
1436 | } | |
1437 | ||
1438 | static int tg_print_rwstat_recursive(struct seq_file *sf, void *v) | |
1439 | { | |
1440 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
1441 | tg_prfill_rwstat_recursive, &blkcg_policy_throtl, | |
1442 | seq_cft(sf)->private, true); | |
1443 | return 0; | |
1444 | } | |
1445 | ||
880f50e2 | 1446 | static struct cftype throtl_legacy_files[] = { |
60c2bc2d TH |
1447 | { |
1448 | .name = "throttle.read_bps_device", | |
9f626e37 | 1449 | .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]), |
2da8ca82 | 1450 | .seq_show = tg_print_conf_u64, |
451af504 | 1451 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1452 | }, |
1453 | { | |
1454 | .name = "throttle.write_bps_device", | |
9f626e37 | 1455 | .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]), |
2da8ca82 | 1456 | .seq_show = tg_print_conf_u64, |
451af504 | 1457 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1458 | }, |
1459 | { | |
1460 | .name = "throttle.read_iops_device", | |
9f626e37 | 1461 | .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]), |
2da8ca82 | 1462 | .seq_show = tg_print_conf_uint, |
451af504 | 1463 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1464 | }, |
1465 | { | |
1466 | .name = "throttle.write_iops_device", | |
9f626e37 | 1467 | .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]), |
2da8ca82 | 1468 | .seq_show = tg_print_conf_uint, |
451af504 | 1469 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1470 | }, |
1471 | { | |
1472 | .name = "throttle.io_service_bytes", | |
7ca46438 TH |
1473 | .private = offsetof(struct throtl_grp, stat_bytes), |
1474 | .seq_show = tg_print_rwstat, | |
60c2bc2d | 1475 | }, |
17534c6f | 1476 | { |
1477 | .name = "throttle.io_service_bytes_recursive", | |
7ca46438 TH |
1478 | .private = offsetof(struct throtl_grp, stat_bytes), |
1479 | .seq_show = tg_print_rwstat_recursive, | |
17534c6f | 1480 | }, |
60c2bc2d TH |
1481 | { |
1482 | .name = "throttle.io_serviced", | |
7ca46438 TH |
1483 | .private = offsetof(struct throtl_grp, stat_ios), |
1484 | .seq_show = tg_print_rwstat, | |
60c2bc2d | 1485 | }, |
17534c6f | 1486 | { |
1487 | .name = "throttle.io_serviced_recursive", | |
7ca46438 TH |
1488 | .private = offsetof(struct throtl_grp, stat_ios), |
1489 | .seq_show = tg_print_rwstat_recursive, | |
17534c6f | 1490 | }, |
60c2bc2d TH |
1491 | { } /* terminate */ |
1492 | }; | |
1493 | ||
cd5ab1b0 | 1494 | static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd, |
2ee867dc TH |
1495 | int off) |
1496 | { | |
1497 | struct throtl_grp *tg = pd_to_tg(pd); | |
1498 | const char *dname = blkg_dev_name(pd->blkg); | |
1499 | char bufs[4][21] = { "max", "max", "max", "max" }; | |
cd5ab1b0 SL |
1500 | u64 bps_dft; |
1501 | unsigned int iops_dft; | |
ada75b6e | 1502 | char idle_time[26] = ""; |
ec80991d | 1503 | char latency_time[26] = ""; |
2ee867dc TH |
1504 | |
1505 | if (!dname) | |
1506 | return 0; | |
9f626e37 | 1507 | |
cd5ab1b0 SL |
1508 | if (off == LIMIT_LOW) { |
1509 | bps_dft = 0; | |
1510 | iops_dft = 0; | |
1511 | } else { | |
1512 | bps_dft = U64_MAX; | |
1513 | iops_dft = UINT_MAX; | |
1514 | } | |
1515 | ||
1516 | if (tg->bps_conf[READ][off] == bps_dft && | |
1517 | tg->bps_conf[WRITE][off] == bps_dft && | |
1518 | tg->iops_conf[READ][off] == iops_dft && | |
ada75b6e | 1519 | tg->iops_conf[WRITE][off] == iops_dft && |
ec80991d | 1520 | (off != LIMIT_LOW || |
b4f428ef | 1521 | (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD && |
5b81fc3c | 1522 | tg->latency_target_conf == DFL_LATENCY_TARGET))) |
2ee867dc TH |
1523 | return 0; |
1524 | ||
9bb67aeb | 1525 | if (tg->bps_conf[READ][off] != U64_MAX) |
9f626e37 | 1526 | snprintf(bufs[0], sizeof(bufs[0]), "%llu", |
cd5ab1b0 | 1527 | tg->bps_conf[READ][off]); |
9bb67aeb | 1528 | if (tg->bps_conf[WRITE][off] != U64_MAX) |
9f626e37 | 1529 | snprintf(bufs[1], sizeof(bufs[1]), "%llu", |
cd5ab1b0 | 1530 | tg->bps_conf[WRITE][off]); |
9bb67aeb | 1531 | if (tg->iops_conf[READ][off] != UINT_MAX) |
9f626e37 | 1532 | snprintf(bufs[2], sizeof(bufs[2]), "%u", |
cd5ab1b0 | 1533 | tg->iops_conf[READ][off]); |
9bb67aeb | 1534 | if (tg->iops_conf[WRITE][off] != UINT_MAX) |
9f626e37 | 1535 | snprintf(bufs[3], sizeof(bufs[3]), "%u", |
cd5ab1b0 | 1536 | tg->iops_conf[WRITE][off]); |
ada75b6e | 1537 | if (off == LIMIT_LOW) { |
5b81fc3c | 1538 | if (tg->idletime_threshold_conf == ULONG_MAX) |
ada75b6e SL |
1539 | strcpy(idle_time, " idle=max"); |
1540 | else | |
1541 | snprintf(idle_time, sizeof(idle_time), " idle=%lu", | |
5b81fc3c | 1542 | tg->idletime_threshold_conf); |
ec80991d | 1543 | |
5b81fc3c | 1544 | if (tg->latency_target_conf == ULONG_MAX) |
ec80991d SL |
1545 | strcpy(latency_time, " latency=max"); |
1546 | else | |
1547 | snprintf(latency_time, sizeof(latency_time), | |
5b81fc3c | 1548 | " latency=%lu", tg->latency_target_conf); |
ada75b6e | 1549 | } |
2ee867dc | 1550 | |
ec80991d SL |
1551 | seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n", |
1552 | dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time, | |
1553 | latency_time); | |
2ee867dc TH |
1554 | return 0; |
1555 | } | |
1556 | ||
cd5ab1b0 | 1557 | static int tg_print_limit(struct seq_file *sf, void *v) |
2ee867dc | 1558 | { |
cd5ab1b0 | 1559 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit, |
2ee867dc TH |
1560 | &blkcg_policy_throtl, seq_cft(sf)->private, false); |
1561 | return 0; | |
1562 | } | |
1563 | ||
cd5ab1b0 | 1564 | static ssize_t tg_set_limit(struct kernfs_open_file *of, |
2ee867dc TH |
1565 | char *buf, size_t nbytes, loff_t off) |
1566 | { | |
1567 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1568 | struct blkg_conf_ctx ctx; | |
1569 | struct throtl_grp *tg; | |
1570 | u64 v[4]; | |
ada75b6e | 1571 | unsigned long idle_time; |
ec80991d | 1572 | unsigned long latency_time; |
2ee867dc | 1573 | int ret; |
cd5ab1b0 | 1574 | int index = of_cft(of)->private; |
2ee867dc TH |
1575 | |
1576 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1577 | if (ret) | |
1578 | return ret; | |
1579 | ||
1580 | tg = blkg_to_tg(ctx.blkg); | |
a880ae93 | 1581 | tg_update_carryover(tg); |
2ee867dc | 1582 | |
cd5ab1b0 SL |
1583 | 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]; | |
2ee867dc | 1587 | |
5b81fc3c SL |
1588 | idle_time = tg->idletime_threshold_conf; |
1589 | latency_time = tg->latency_target_conf; | |
2ee867dc TH |
1590 | while (true) { |
1591 | char tok[27]; /* wiops=18446744073709551616 */ | |
1592 | char *p; | |
2ab5492d | 1593 | u64 val = U64_MAX; |
2ee867dc TH |
1594 | 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; | |
5b7048b8 | 1613 | if (!strcmp(tok, "rbps") && val > 1) |
2ee867dc | 1614 | v[0] = val; |
5b7048b8 | 1615 | else if (!strcmp(tok, "wbps") && val > 1) |
2ee867dc | 1616 | v[1] = val; |
5b7048b8 | 1617 | else if (!strcmp(tok, "riops") && val > 1) |
2ee867dc | 1618 | v[2] = min_t(u64, val, UINT_MAX); |
5b7048b8 | 1619 | else if (!strcmp(tok, "wiops") && val > 1) |
2ee867dc | 1620 | v[3] = min_t(u64, val, UINT_MAX); |
ada75b6e SL |
1621 | else if (off == LIMIT_LOW && !strcmp(tok, "idle")) |
1622 | idle_time = val; | |
ec80991d SL |
1623 | else if (off == LIMIT_LOW && !strcmp(tok, "latency")) |
1624 | latency_time = val; | |
2ee867dc TH |
1625 | else |
1626 | goto out_finish; | |
1627 | } | |
1628 | ||
cd5ab1b0 SL |
1629 | 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]; | |
2ee867dc | 1633 | |
cd5ab1b0 SL |
1634 | 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]); | |
b4f428ef SL |
1648 | tg->idletime_threshold_conf = idle_time; |
1649 | tg->latency_target_conf = latency_time; | |
1650 | ||
1651 | /* 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) { | |
5b81fc3c | 1663 | tg->idletime_threshold = tg->idletime_threshold_conf; |
5b81fc3c | 1664 | tg->latency_target = tg->latency_target_conf; |
cd5ab1b0 | 1665 | } |
b4f428ef SL |
1666 | |
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; | |
9bb67aeb SL |
1673 | tg_conf_updated(tg, index == LIMIT_LOW && |
1674 | tg->td->limit_valid[LIMIT_LOW]); | |
2ee867dc TH |
1675 | ret = 0; |
1676 | out_finish: | |
1677 | blkg_conf_finish(&ctx); | |
1678 | return ret ?: nbytes; | |
1679 | } | |
1680 | ||
1681 | static struct cftype throtl_files[] = { | |
cd5ab1b0 SL |
1682 | #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 | |
2ee867dc TH |
1691 | { |
1692 | .name = "max", | |
1693 | .flags = CFTYPE_NOT_ON_ROOT, | |
cd5ab1b0 SL |
1694 | .seq_show = tg_print_limit, |
1695 | .write = tg_set_limit, | |
1696 | .private = LIMIT_MAX, | |
2ee867dc TH |
1697 | }, |
1698 | { } /* terminate */ | |
1699 | }; | |
1700 | ||
da527770 | 1701 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1702 | { |
1703 | struct throtl_data *td = q->td; | |
1704 | ||
69df0ab0 | 1705 | cancel_work_sync(&td->dispatch_work); |
e43473b7 VG |
1706 | } |
1707 | ||
a7b36ee6 | 1708 | struct blkcg_policy blkcg_policy_throtl = { |
2ee867dc | 1709 | .dfl_cftypes = throtl_files, |
880f50e2 | 1710 | .legacy_cftypes = throtl_legacy_files, |
f9fcc2d3 | 1711 | |
001bea73 | 1712 | .pd_alloc_fn = throtl_pd_alloc, |
f9fcc2d3 | 1713 | .pd_init_fn = throtl_pd_init, |
693e751e | 1714 | .pd_online_fn = throtl_pd_online, |
cd5ab1b0 | 1715 | .pd_offline_fn = throtl_pd_offline, |
001bea73 | 1716 | .pd_free_fn = throtl_pd_free, |
e43473b7 VG |
1717 | }; |
1718 | ||
2d8f7a3b YK |
1719 | void blk_throtl_cancel_bios(struct request_queue *q) |
1720 | { | |
1721 | struct cgroup_subsys_state *pos_css; | |
1722 | struct blkcg_gq *blkg; | |
1723 | ||
1724 | spin_lock_irq(&q->queue_lock); | |
1725 | /* | |
1726 | * queue_lock is held, rcu lock is not needed here technically. | |
1727 | * However, rcu lock is still held to emphasize that following | |
1728 | * path need RCU protection and to prevent warning from lockdep. | |
1729 | */ | |
1730 | rcu_read_lock(); | |
1731 | blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) { | |
1732 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1733 | struct throtl_service_queue *sq = &tg->service_queue; | |
1734 | ||
1735 | /* | |
1736 | * Set the flag to make sure throtl_pending_timer_fn() won't | |
1737 | * stop until all throttled bios are dispatched. | |
1738 | */ | |
1739 | blkg_to_tg(blkg)->flags |= THROTL_TG_CANCELING; | |
1740 | /* | |
1741 | * Update disptime after setting the above flag to make sure | |
1742 | * throtl_select_dispatch() won't exit without dispatching. | |
1743 | */ | |
1744 | tg_update_disptime(tg); | |
1745 | ||
1746 | throtl_schedule_pending_timer(sq, jiffies + 1); | |
1747 | } | |
1748 | rcu_read_unlock(); | |
1749 | spin_unlock_irq(&q->queue_lock); | |
1750 | } | |
1751 | ||
1752 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
3f0abd80 SL |
1753 | static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg) |
1754 | { | |
1755 | unsigned long rtime = jiffies, wtime = jiffies; | |
1756 | ||
1757 | if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]) | |
1758 | rtime = tg->last_low_overflow_time[READ]; | |
1759 | if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) | |
1760 | wtime = tg->last_low_overflow_time[WRITE]; | |
1761 | return min(rtime, wtime); | |
1762 | } | |
1763 | ||
1764 | /* tg should not be an intermediate node */ | |
1765 | static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg) | |
1766 | { | |
1767 | struct throtl_service_queue *parent_sq; | |
1768 | struct throtl_grp *parent = tg; | |
1769 | unsigned long ret = __tg_last_low_overflow_time(tg); | |
1770 | ||
1771 | while (true) { | |
1772 | parent_sq = parent->service_queue.parent_sq; | |
1773 | parent = sq_to_tg(parent_sq); | |
1774 | if (!parent) | |
1775 | break; | |
1776 | ||
1777 | /* | |
1778 | * The parent doesn't have low limit, it always reaches low | |
1779 | * limit. Its overflow time is useless for children | |
1780 | */ | |
1781 | if (!parent->bps[READ][LIMIT_LOW] && | |
1782 | !parent->iops[READ][LIMIT_LOW] && | |
1783 | !parent->bps[WRITE][LIMIT_LOW] && | |
1784 | !parent->iops[WRITE][LIMIT_LOW]) | |
1785 | continue; | |
1786 | if (time_after(__tg_last_low_overflow_time(parent), ret)) | |
1787 | ret = __tg_last_low_overflow_time(parent); | |
1788 | } | |
1789 | return ret; | |
1790 | } | |
1791 | ||
9e234eea SL |
1792 | static bool throtl_tg_is_idle(struct throtl_grp *tg) |
1793 | { | |
1794 | /* | |
1795 | * cgroup is idle if: | |
1796 | * - single idle is too long, longer than a fixed value (in case user | |
b4f428ef | 1797 | * configure a too big threshold) or 4 times of idletime threshold |
9e234eea | 1798 | * - average think time is more than threshold |
53696b8d | 1799 | * - IO latency is largely below threshold |
9e234eea | 1800 | */ |
b4f428ef | 1801 | unsigned long time; |
4cff729f | 1802 | bool ret; |
9e234eea | 1803 | |
b4f428ef SL |
1804 | time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold); |
1805 | ret = tg->latency_target == DFL_LATENCY_TARGET || | |
1806 | tg->idletime_threshold == DFL_IDLE_THRESHOLD || | |
1807 | (ktime_get_ns() >> 10) - tg->last_finish_time > time || | |
1808 | tg->avg_idletime > tg->idletime_threshold || | |
1809 | (tg->latency_target && tg->bio_cnt && | |
53696b8d | 1810 | tg->bad_bio_cnt * 5 < tg->bio_cnt); |
4cff729f SL |
1811 | throtl_log(&tg->service_queue, |
1812 | "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d", | |
1813 | tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt, | |
1814 | tg->bio_cnt, ret, tg->td->scale); | |
1815 | return ret; | |
9e234eea SL |
1816 | } |
1817 | ||
c79892c5 SL |
1818 | static bool throtl_tg_can_upgrade(struct throtl_grp *tg) |
1819 | { | |
1820 | struct throtl_service_queue *sq = &tg->service_queue; | |
1821 | bool read_limit, write_limit; | |
1822 | ||
1823 | /* | |
1824 | * if cgroup reaches low limit (if low limit is 0, the cgroup always | |
1825 | * reaches), it's ok to upgrade to next limit | |
1826 | */ | |
1827 | read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]; | |
1828 | write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]; | |
1829 | if (!read_limit && !write_limit) | |
1830 | return true; | |
1831 | if (read_limit && sq->nr_queued[READ] && | |
1832 | (!write_limit || sq->nr_queued[WRITE])) | |
1833 | return true; | |
1834 | if (write_limit && sq->nr_queued[WRITE] && | |
1835 | (!read_limit || sq->nr_queued[READ])) | |
1836 | return true; | |
aec24246 SL |
1837 | |
1838 | if (time_after_eq(jiffies, | |
fa6fb5aa SL |
1839 | tg_last_low_overflow_time(tg) + tg->td->throtl_slice) && |
1840 | throtl_tg_is_idle(tg)) | |
aec24246 | 1841 | return true; |
c79892c5 SL |
1842 | return false; |
1843 | } | |
1844 | ||
1845 | static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg) | |
1846 | { | |
1847 | while (true) { | |
1848 | if (throtl_tg_can_upgrade(tg)) | |
1849 | return true; | |
1850 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1851 | if (!tg || !tg_to_blkg(tg)->parent) | |
1852 | return false; | |
1853 | } | |
1854 | return false; | |
1855 | } | |
1856 | ||
1857 | static bool throtl_can_upgrade(struct throtl_data *td, | |
1858 | struct throtl_grp *this_tg) | |
1859 | { | |
1860 | struct cgroup_subsys_state *pos_css; | |
1861 | struct blkcg_gq *blkg; | |
1862 | ||
1863 | if (td->limit_index != LIMIT_LOW) | |
1864 | return false; | |
1865 | ||
297e3d85 | 1866 | if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice)) |
3f0abd80 SL |
1867 | return false; |
1868 | ||
c79892c5 SL |
1869 | rcu_read_lock(); |
1870 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1871 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1872 | ||
1873 | if (tg == this_tg) | |
1874 | continue; | |
1875 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1876 | continue; | |
1877 | if (!throtl_hierarchy_can_upgrade(tg)) { | |
1878 | rcu_read_unlock(); | |
1879 | return false; | |
1880 | } | |
1881 | } | |
1882 | rcu_read_unlock(); | |
1883 | return true; | |
1884 | } | |
1885 | ||
fa6fb5aa SL |
1886 | static void throtl_upgrade_check(struct throtl_grp *tg) |
1887 | { | |
1888 | unsigned long now = jiffies; | |
1889 | ||
1890 | if (tg->td->limit_index != LIMIT_LOW) | |
1891 | return; | |
1892 | ||
1893 | if (time_after(tg->last_check_time + tg->td->throtl_slice, now)) | |
1894 | return; | |
1895 | ||
1896 | tg->last_check_time = now; | |
1897 | ||
1898 | if (!time_after_eq(now, | |
1899 | __tg_last_low_overflow_time(tg) + tg->td->throtl_slice)) | |
1900 | return; | |
1901 | ||
1902 | if (throtl_can_upgrade(tg->td, NULL)) | |
1903 | throtl_upgrade_state(tg->td); | |
1904 | } | |
1905 | ||
c79892c5 SL |
1906 | static void throtl_upgrade_state(struct throtl_data *td) |
1907 | { | |
1908 | struct cgroup_subsys_state *pos_css; | |
1909 | struct blkcg_gq *blkg; | |
1910 | ||
4cff729f | 1911 | throtl_log(&td->service_queue, "upgrade to max"); |
c79892c5 | 1912 | td->limit_index = LIMIT_MAX; |
3f0abd80 | 1913 | td->low_upgrade_time = jiffies; |
7394e31f | 1914 | td->scale = 0; |
c79892c5 SL |
1915 | rcu_read_lock(); |
1916 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1917 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1918 | struct throtl_service_queue *sq = &tg->service_queue; | |
1919 | ||
1920 | tg->disptime = jiffies - 1; | |
1921 | throtl_select_dispatch(sq); | |
4f02fb76 | 1922 | throtl_schedule_next_dispatch(sq, true); |
c79892c5 SL |
1923 | } |
1924 | rcu_read_unlock(); | |
1925 | throtl_select_dispatch(&td->service_queue); | |
4f02fb76 | 1926 | throtl_schedule_next_dispatch(&td->service_queue, true); |
c79892c5 SL |
1927 | queue_work(kthrotld_workqueue, &td->dispatch_work); |
1928 | } | |
1929 | ||
4247d9c8 | 1930 | static void throtl_downgrade_state(struct throtl_data *td) |
3f0abd80 | 1931 | { |
7394e31f SL |
1932 | td->scale /= 2; |
1933 | ||
4cff729f | 1934 | throtl_log(&td->service_queue, "downgrade, scale %d", td->scale); |
7394e31f SL |
1935 | if (td->scale) { |
1936 | td->low_upgrade_time = jiffies - td->scale * td->throtl_slice; | |
1937 | return; | |
1938 | } | |
1939 | ||
4247d9c8 | 1940 | td->limit_index = LIMIT_LOW; |
3f0abd80 SL |
1941 | td->low_downgrade_time = jiffies; |
1942 | } | |
1943 | ||
1944 | static bool throtl_tg_can_downgrade(struct throtl_grp *tg) | |
1945 | { | |
1946 | struct throtl_data *td = tg->td; | |
1947 | unsigned long now = jiffies; | |
1948 | ||
1949 | /* | |
1950 | * If cgroup is below low limit, consider downgrade and throttle other | |
1951 | * cgroups | |
1952 | */ | |
297e3d85 SL |
1953 | if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) && |
1954 | time_after_eq(now, tg_last_low_overflow_time(tg) + | |
fa6fb5aa SL |
1955 | td->throtl_slice) && |
1956 | (!throtl_tg_is_idle(tg) || | |
1957 | !list_empty(&tg_to_blkg(tg)->blkcg->css.children))) | |
3f0abd80 SL |
1958 | return true; |
1959 | return false; | |
1960 | } | |
1961 | ||
1962 | static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg) | |
1963 | { | |
1964 | while (true) { | |
1965 | if (!throtl_tg_can_downgrade(tg)) | |
1966 | return false; | |
1967 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1968 | if (!tg || !tg_to_blkg(tg)->parent) | |
1969 | break; | |
1970 | } | |
1971 | return true; | |
1972 | } | |
1973 | ||
1974 | static void throtl_downgrade_check(struct throtl_grp *tg) | |
1975 | { | |
1976 | uint64_t bps; | |
1977 | unsigned int iops; | |
1978 | unsigned long elapsed_time; | |
1979 | unsigned long now = jiffies; | |
1980 | ||
1981 | if (tg->td->limit_index != LIMIT_MAX || | |
1982 | !tg->td->limit_valid[LIMIT_LOW]) | |
1983 | return; | |
1984 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1985 | return; | |
297e3d85 | 1986 | if (time_after(tg->last_check_time + tg->td->throtl_slice, now)) |
3f0abd80 SL |
1987 | return; |
1988 | ||
1989 | elapsed_time = now - tg->last_check_time; | |
1990 | tg->last_check_time = now; | |
1991 | ||
297e3d85 SL |
1992 | if (time_before(now, tg_last_low_overflow_time(tg) + |
1993 | tg->td->throtl_slice)) | |
3f0abd80 SL |
1994 | return; |
1995 | ||
1996 | if (tg->bps[READ][LIMIT_LOW]) { | |
1997 | bps = tg->last_bytes_disp[READ] * HZ; | |
1998 | do_div(bps, elapsed_time); | |
1999 | if (bps >= tg->bps[READ][LIMIT_LOW]) | |
2000 | tg->last_low_overflow_time[READ] = now; | |
2001 | } | |
2002 | ||
2003 | if (tg->bps[WRITE][LIMIT_LOW]) { | |
2004 | bps = tg->last_bytes_disp[WRITE] * HZ; | |
2005 | do_div(bps, elapsed_time); | |
2006 | if (bps >= tg->bps[WRITE][LIMIT_LOW]) | |
2007 | tg->last_low_overflow_time[WRITE] = now; | |
2008 | } | |
2009 | ||
2010 | if (tg->iops[READ][LIMIT_LOW]) { | |
2011 | iops = tg->last_io_disp[READ] * HZ / elapsed_time; | |
2012 | if (iops >= tg->iops[READ][LIMIT_LOW]) | |
2013 | tg->last_low_overflow_time[READ] = now; | |
2014 | } | |
2015 | ||
2016 | if (tg->iops[WRITE][LIMIT_LOW]) { | |
2017 | iops = tg->last_io_disp[WRITE] * HZ / elapsed_time; | |
2018 | if (iops >= tg->iops[WRITE][LIMIT_LOW]) | |
2019 | tg->last_low_overflow_time[WRITE] = now; | |
2020 | } | |
2021 | ||
2022 | /* | |
2023 | * If cgroup is below low limit, consider downgrade and throttle other | |
2024 | * cgroups | |
2025 | */ | |
2026 | if (throtl_hierarchy_can_downgrade(tg)) | |
4247d9c8 | 2027 | throtl_downgrade_state(tg->td); |
3f0abd80 SL |
2028 | |
2029 | tg->last_bytes_disp[READ] = 0; | |
2030 | tg->last_bytes_disp[WRITE] = 0; | |
2031 | tg->last_io_disp[READ] = 0; | |
2032 | tg->last_io_disp[WRITE] = 0; | |
2033 | } | |
2034 | ||
9e234eea SL |
2035 | static void blk_throtl_update_idletime(struct throtl_grp *tg) |
2036 | { | |
7901601a | 2037 | unsigned long now; |
9e234eea SL |
2038 | unsigned long last_finish_time = tg->last_finish_time; |
2039 | ||
7901601a BW |
2040 | if (last_finish_time == 0) |
2041 | return; | |
2042 | ||
2043 | now = ktime_get_ns() >> 10; | |
2044 | if (now <= last_finish_time || | |
9e234eea SL |
2045 | last_finish_time == tg->checked_last_finish_time) |
2046 | return; | |
2047 | ||
2048 | tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3; | |
2049 | tg->checked_last_finish_time = last_finish_time; | |
2050 | } | |
2051 | ||
b9147dd1 SL |
2052 | static void throtl_update_latency_buckets(struct throtl_data *td) |
2053 | { | |
b889bf66 JQ |
2054 | struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE]; |
2055 | int i, cpu, rw; | |
2056 | unsigned long last_latency[2] = { 0 }; | |
2057 | unsigned long latency[2]; | |
b9147dd1 | 2058 | |
b185efa7 | 2059 | if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW]) |
b9147dd1 SL |
2060 | return; |
2061 | if (time_before(jiffies, td->last_calculate_time + HZ)) | |
2062 | return; | |
2063 | td->last_calculate_time = jiffies; | |
2064 | ||
2065 | memset(avg_latency, 0, sizeof(avg_latency)); | |
b889bf66 JQ |
2066 | for (rw = READ; rw <= WRITE; rw++) { |
2067 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) { | |
2068 | struct latency_bucket *tmp = &td->tmp_buckets[rw][i]; | |
2069 | ||
2070 | for_each_possible_cpu(cpu) { | |
2071 | struct latency_bucket *bucket; | |
2072 | ||
2073 | /* this isn't race free, but ok in practice */ | |
2074 | bucket = per_cpu_ptr(td->latency_buckets[rw], | |
2075 | cpu); | |
2076 | tmp->total_latency += bucket[i].total_latency; | |
2077 | tmp->samples += bucket[i].samples; | |
2078 | bucket[i].total_latency = 0; | |
2079 | bucket[i].samples = 0; | |
2080 | } | |
b9147dd1 | 2081 | |
b889bf66 JQ |
2082 | if (tmp->samples >= 32) { |
2083 | int samples = tmp->samples; | |
b9147dd1 | 2084 | |
b889bf66 | 2085 | latency[rw] = tmp->total_latency; |
b9147dd1 | 2086 | |
b889bf66 JQ |
2087 | tmp->total_latency = 0; |
2088 | tmp->samples = 0; | |
2089 | latency[rw] /= samples; | |
2090 | if (latency[rw] == 0) | |
2091 | continue; | |
2092 | avg_latency[rw][i].latency = latency[rw]; | |
2093 | } | |
b9147dd1 SL |
2094 | } |
2095 | } | |
2096 | ||
b889bf66 JQ |
2097 | for (rw = READ; rw <= WRITE; rw++) { |
2098 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) { | |
2099 | if (!avg_latency[rw][i].latency) { | |
2100 | if (td->avg_buckets[rw][i].latency < last_latency[rw]) | |
2101 | td->avg_buckets[rw][i].latency = | |
2102 | last_latency[rw]; | |
2103 | continue; | |
2104 | } | |
b9147dd1 | 2105 | |
b889bf66 JQ |
2106 | if (!td->avg_buckets[rw][i].valid) |
2107 | latency[rw] = avg_latency[rw][i].latency; | |
2108 | else | |
2109 | latency[rw] = (td->avg_buckets[rw][i].latency * 7 + | |
2110 | avg_latency[rw][i].latency) >> 3; | |
b9147dd1 | 2111 | |
b889bf66 JQ |
2112 | td->avg_buckets[rw][i].latency = max(latency[rw], |
2113 | last_latency[rw]); | |
2114 | td->avg_buckets[rw][i].valid = true; | |
2115 | last_latency[rw] = td->avg_buckets[rw][i].latency; | |
2116 | } | |
b9147dd1 | 2117 | } |
4cff729f SL |
2118 | |
2119 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) | |
2120 | throtl_log(&td->service_queue, | |
b889bf66 JQ |
2121 | "Latency bucket %d: read latency=%ld, read valid=%d, " |
2122 | "write latency=%ld, write valid=%d", i, | |
2123 | td->avg_buckets[READ][i].latency, | |
2124 | td->avg_buckets[READ][i].valid, | |
2125 | td->avg_buckets[WRITE][i].latency, | |
2126 | td->avg_buckets[WRITE][i].valid); | |
b9147dd1 SL |
2127 | } |
2128 | #else | |
2129 | static inline void throtl_update_latency_buckets(struct throtl_data *td) | |
2130 | { | |
2131 | } | |
2d8f7a3b YK |
2132 | |
2133 | static void blk_throtl_update_idletime(struct throtl_grp *tg) | |
2134 | { | |
2135 | } | |
2136 | ||
2137 | static void throtl_downgrade_check(struct throtl_grp *tg) | |
2138 | { | |
2139 | } | |
2140 | ||
2141 | static void throtl_upgrade_check(struct throtl_grp *tg) | |
2142 | { | |
2143 | } | |
2144 | ||
2145 | static bool throtl_can_upgrade(struct throtl_data *td, | |
2146 | struct throtl_grp *this_tg) | |
2147 | { | |
2148 | return false; | |
2149 | } | |
2150 | ||
2151 | static void throtl_upgrade_state(struct throtl_data *td) | |
2152 | { | |
2153 | } | |
b9147dd1 SL |
2154 | #endif |
2155 | ||
a7b36ee6 | 2156 | bool __blk_throtl_bio(struct bio *bio) |
e43473b7 | 2157 | { |
ed6cddef | 2158 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
db18a53e | 2159 | struct blkcg_gq *blkg = bio->bi_blkg; |
c5cc2070 | 2160 | struct throtl_qnode *qn = NULL; |
a2e83ef9 | 2161 | struct throtl_grp *tg = blkg_to_tg(blkg); |
73f0d49a | 2162 | struct throtl_service_queue *sq; |
0e9f4164 | 2163 | bool rw = bio_data_dir(bio); |
bc16a4f9 | 2164 | bool throttled = false; |
b9147dd1 | 2165 | struct throtl_data *td = tg->td; |
e43473b7 | 2166 | |
93b80638 | 2167 | rcu_read_lock(); |
ae118896 | 2168 | |
7ca46438 TH |
2169 | if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) { |
2170 | blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf, | |
2171 | bio->bi_iter.bi_size); | |
2172 | blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1); | |
2173 | } | |
2174 | ||
0d945c1f | 2175 | spin_lock_irq(&q->queue_lock); |
c9589f03 | 2176 | |
b9147dd1 SL |
2177 | throtl_update_latency_buckets(td); |
2178 | ||
9e234eea SL |
2179 | blk_throtl_update_idletime(tg); |
2180 | ||
73f0d49a TH |
2181 | sq = &tg->service_queue; |
2182 | ||
c79892c5 | 2183 | again: |
9e660acf | 2184 | while (true) { |
3f0abd80 SL |
2185 | if (tg->last_low_overflow_time[rw] == 0) |
2186 | tg->last_low_overflow_time[rw] = jiffies; | |
2187 | throtl_downgrade_check(tg); | |
fa6fb5aa | 2188 | throtl_upgrade_check(tg); |
9e660acf TH |
2189 | /* throtl is FIFO - if bios are already queued, should queue */ |
2190 | if (sq->nr_queued[rw]) | |
2191 | break; | |
de701c74 | 2192 | |
9e660acf | 2193 | /* if above limits, break to queue */ |
c79892c5 | 2194 | if (!tg_may_dispatch(tg, bio, NULL)) { |
3f0abd80 | 2195 | tg->last_low_overflow_time[rw] = jiffies; |
b9147dd1 SL |
2196 | if (throtl_can_upgrade(td, tg)) { |
2197 | throtl_upgrade_state(td); | |
c79892c5 SL |
2198 | goto again; |
2199 | } | |
9e660acf | 2200 | break; |
c79892c5 | 2201 | } |
9e660acf TH |
2202 | |
2203 | /* within limits, let's charge and dispatch directly */ | |
e43473b7 | 2204 | throtl_charge_bio(tg, bio); |
04521db0 VG |
2205 | |
2206 | /* | |
2207 | * We need to trim slice even when bios are not being queued | |
2208 | * otherwise it might happen that a bio is not queued for | |
2209 | * a long time and slice keeps on extending and trim is not | |
2210 | * called for a long time. Now if limits are reduced suddenly | |
2211 | * we take into account all the IO dispatched so far at new | |
2212 | * low rate and * newly queued IO gets a really long dispatch | |
2213 | * time. | |
2214 | * | |
2215 | * So keep on trimming slice even if bio is not queued. | |
2216 | */ | |
0f3457f6 | 2217 | throtl_trim_slice(tg, rw); |
9e660acf TH |
2218 | |
2219 | /* | |
2220 | * @bio passed through this layer without being throttled. | |
b53b072c | 2221 | * Climb up the ladder. If we're already at the top, it |
9e660acf TH |
2222 | * can be executed directly. |
2223 | */ | |
c5cc2070 | 2224 | qn = &tg->qnode_on_parent[rw]; |
9e660acf TH |
2225 | sq = sq->parent_sq; |
2226 | tg = sq_to_tg(sq); | |
320fb0f9 YK |
2227 | if (!tg) { |
2228 | bio_set_flag(bio, BIO_BPS_THROTTLED); | |
9e660acf | 2229 | goto out_unlock; |
320fb0f9 | 2230 | } |
e43473b7 VG |
2231 | } |
2232 | ||
9e660acf | 2233 | /* out-of-limit, queue to @tg */ |
fda6f272 TH |
2234 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
2235 | rw == READ ? 'R' : 'W', | |
9f626e37 SL |
2236 | tg->bytes_disp[rw], bio->bi_iter.bi_size, |
2237 | tg_bps_limit(tg, rw), | |
2238 | tg->io_disp[rw], tg_iops_limit(tg, rw), | |
fda6f272 | 2239 | sq->nr_queued[READ], sq->nr_queued[WRITE]); |
e43473b7 | 2240 | |
3f0abd80 SL |
2241 | tg->last_low_overflow_time[rw] = jiffies; |
2242 | ||
b9147dd1 | 2243 | td->nr_queued[rw]++; |
c5cc2070 | 2244 | throtl_add_bio_tg(bio, qn, tg); |
bc16a4f9 | 2245 | throttled = true; |
e43473b7 | 2246 | |
7f52f98c TH |
2247 | /* |
2248 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
2249 | * was empty before @bio. The forced scheduling isn't likely to | |
2250 | * cause undue delay as @bio is likely to be dispatched directly if | |
2251 | * its @tg's disptime is not in the future. | |
2252 | */ | |
0e9f4164 | 2253 | if (tg->flags & THROTL_TG_WAS_EMPTY) { |
77216b04 | 2254 | tg_update_disptime(tg); |
7f52f98c | 2255 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); |
e43473b7 VG |
2256 | } |
2257 | ||
bc16a4f9 | 2258 | out_unlock: |
b9147dd1 SL |
2259 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
2260 | if (throttled || !td->track_bio_latency) | |
5238dcf4 | 2261 | bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY; |
b9147dd1 | 2262 | #endif |
5a011f88 LQ |
2263 | spin_unlock_irq(&q->queue_lock); |
2264 | ||
93b80638 | 2265 | rcu_read_unlock(); |
bc16a4f9 | 2266 | return throttled; |
e43473b7 VG |
2267 | } |
2268 | ||
9e234eea | 2269 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
b9147dd1 | 2270 | static void throtl_track_latency(struct throtl_data *td, sector_t size, |
77e7ffd7 | 2271 | enum req_op op, unsigned long time) |
b9147dd1 | 2272 | { |
77e7ffd7 | 2273 | const bool rw = op_is_write(op); |
b9147dd1 SL |
2274 | struct latency_bucket *latency; |
2275 | int index; | |
2276 | ||
b889bf66 JQ |
2277 | if (!td || td->limit_index != LIMIT_LOW || |
2278 | !(op == REQ_OP_READ || op == REQ_OP_WRITE) || | |
b9147dd1 SL |
2279 | !blk_queue_nonrot(td->queue)) |
2280 | return; | |
2281 | ||
2282 | index = request_bucket_index(size); | |
2283 | ||
77e7ffd7 | 2284 | latency = get_cpu_ptr(td->latency_buckets[rw]); |
b9147dd1 SL |
2285 | latency[index].total_latency += time; |
2286 | latency[index].samples++; | |
77e7ffd7 | 2287 | put_cpu_ptr(td->latency_buckets[rw]); |
b9147dd1 SL |
2288 | } |
2289 | ||
2290 | void blk_throtl_stat_add(struct request *rq, u64 time_ns) | |
2291 | { | |
2292 | struct request_queue *q = rq->q; | |
2293 | struct throtl_data *td = q->td; | |
2294 | ||
3d244306 HT |
2295 | throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq), |
2296 | time_ns >> 10); | |
b9147dd1 SL |
2297 | } |
2298 | ||
9e234eea SL |
2299 | void blk_throtl_bio_endio(struct bio *bio) |
2300 | { | |
08e18eab | 2301 | struct blkcg_gq *blkg; |
9e234eea | 2302 | struct throtl_grp *tg; |
b9147dd1 SL |
2303 | u64 finish_time_ns; |
2304 | unsigned long finish_time; | |
2305 | unsigned long start_time; | |
2306 | unsigned long lat; | |
b889bf66 | 2307 | int rw = bio_data_dir(bio); |
9e234eea | 2308 | |
08e18eab JB |
2309 | blkg = bio->bi_blkg; |
2310 | if (!blkg) | |
9e234eea | 2311 | return; |
08e18eab | 2312 | tg = blkg_to_tg(blkg); |
b185efa7 BW |
2313 | if (!tg->td->limit_valid[LIMIT_LOW]) |
2314 | return; | |
9e234eea | 2315 | |
b9147dd1 SL |
2316 | finish_time_ns = ktime_get_ns(); |
2317 | tg->last_finish_time = finish_time_ns >> 10; | |
2318 | ||
5238dcf4 OS |
2319 | start_time = bio_issue_time(&bio->bi_issue) >> 10; |
2320 | finish_time = __bio_issue_time(finish_time_ns) >> 10; | |
08e18eab | 2321 | if (!start_time || finish_time <= start_time) |
53696b8d SL |
2322 | return; |
2323 | ||
2324 | lat = finish_time - start_time; | |
b9147dd1 | 2325 | /* this is only for bio based driver */ |
5238dcf4 OS |
2326 | if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY)) |
2327 | throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue), | |
2328 | bio_op(bio), lat); | |
53696b8d | 2329 | |
6679a90c | 2330 | if (tg->latency_target && lat >= tg->td->filtered_latency) { |
53696b8d SL |
2331 | int bucket; |
2332 | unsigned int threshold; | |
2333 | ||
5238dcf4 | 2334 | bucket = request_bucket_index(bio_issue_size(&bio->bi_issue)); |
b889bf66 | 2335 | threshold = tg->td->avg_buckets[rw][bucket].latency + |
53696b8d SL |
2336 | tg->latency_target; |
2337 | if (lat > threshold) | |
2338 | tg->bad_bio_cnt++; | |
2339 | /* | |
2340 | * Not race free, could get wrong count, which means cgroups | |
2341 | * will be throttled | |
2342 | */ | |
2343 | tg->bio_cnt++; | |
2344 | } | |
2345 | ||
2346 | if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) { | |
2347 | tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies; | |
2348 | tg->bio_cnt /= 2; | |
2349 | tg->bad_bio_cnt /= 2; | |
b9147dd1 | 2350 | } |
9e234eea SL |
2351 | } |
2352 | #endif | |
2353 | ||
e13793ba | 2354 | int blk_throtl_init(struct gendisk *disk) |
e43473b7 | 2355 | { |
e13793ba | 2356 | struct request_queue *q = disk->queue; |
e43473b7 | 2357 | struct throtl_data *td; |
a2b1693b | 2358 | int ret; |
e43473b7 VG |
2359 | |
2360 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
2361 | if (!td) | |
2362 | return -ENOMEM; | |
b889bf66 | 2363 | td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) * |
b9147dd1 | 2364 | LATENCY_BUCKET_SIZE, __alignof__(u64)); |
b889bf66 JQ |
2365 | if (!td->latency_buckets[READ]) { |
2366 | kfree(td); | |
2367 | return -ENOMEM; | |
2368 | } | |
2369 | td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) * | |
b9147dd1 | 2370 | LATENCY_BUCKET_SIZE, __alignof__(u64)); |
b889bf66 JQ |
2371 | if (!td->latency_buckets[WRITE]) { |
2372 | free_percpu(td->latency_buckets[READ]); | |
b9147dd1 SL |
2373 | kfree(td); |
2374 | return -ENOMEM; | |
2375 | } | |
e43473b7 | 2376 | |
69df0ab0 | 2377 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
b2ce2643 | 2378 | throtl_service_queue_init(&td->service_queue); |
e43473b7 | 2379 | |
cd1604fa | 2380 | q->td = td; |
29b12589 | 2381 | td->queue = q; |
02977e4a | 2382 | |
9f626e37 | 2383 | td->limit_valid[LIMIT_MAX] = true; |
cd5ab1b0 | 2384 | td->limit_index = LIMIT_MAX; |
3f0abd80 SL |
2385 | td->low_upgrade_time = jiffies; |
2386 | td->low_downgrade_time = jiffies; | |
9e234eea | 2387 | |
a2b1693b | 2388 | /* activate policy */ |
3c798398 | 2389 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
b9147dd1 | 2390 | if (ret) { |
b889bf66 JQ |
2391 | free_percpu(td->latency_buckets[READ]); |
2392 | free_percpu(td->latency_buckets[WRITE]); | |
f51b802c | 2393 | kfree(td); |
b9147dd1 | 2394 | } |
a2b1693b | 2395 | return ret; |
e43473b7 VG |
2396 | } |
2397 | ||
e13793ba | 2398 | void blk_throtl_exit(struct gendisk *disk) |
e43473b7 | 2399 | { |
e13793ba CH |
2400 | struct request_queue *q = disk->queue; |
2401 | ||
c875f4d0 | 2402 | BUG_ON(!q->td); |
884f0e84 | 2403 | del_timer_sync(&q->td->service_queue.pending_timer); |
da527770 | 2404 | throtl_shutdown_wq(q); |
3c798398 | 2405 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
b889bf66 JQ |
2406 | free_percpu(q->td->latency_buckets[READ]); |
2407 | free_percpu(q->td->latency_buckets[WRITE]); | |
c9a929dd | 2408 | kfree(q->td); |
e43473b7 VG |
2409 | } |
2410 | ||
d61fcfa4 SL |
2411 | void blk_throtl_register_queue(struct request_queue *q) |
2412 | { | |
2413 | struct throtl_data *td; | |
6679a90c | 2414 | int i; |
d61fcfa4 SL |
2415 | |
2416 | td = q->td; | |
2417 | BUG_ON(!td); | |
2418 | ||
6679a90c | 2419 | if (blk_queue_nonrot(q)) { |
d61fcfa4 | 2420 | td->throtl_slice = DFL_THROTL_SLICE_SSD; |
6679a90c SL |
2421 | td->filtered_latency = LATENCY_FILTERED_SSD; |
2422 | } else { | |
d61fcfa4 | 2423 | td->throtl_slice = DFL_THROTL_SLICE_HD; |
6679a90c | 2424 | td->filtered_latency = LATENCY_FILTERED_HD; |
b889bf66 JQ |
2425 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) { |
2426 | td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY; | |
2427 | td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY; | |
2428 | } | |
6679a90c | 2429 | } |
d61fcfa4 SL |
2430 | #ifndef CONFIG_BLK_DEV_THROTTLING_LOW |
2431 | /* if no low limit, use previous default */ | |
2432 | td->throtl_slice = DFL_THROTL_SLICE_HD; | |
2433 | #endif | |
9e234eea | 2434 | |
344e9ffc | 2435 | td->track_bio_latency = !queue_is_mq(q); |
b9147dd1 SL |
2436 | if (!td->track_bio_latency) |
2437 | blk_stat_enable_accounting(q); | |
d61fcfa4 SL |
2438 | } |
2439 | ||
297e3d85 SL |
2440 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
2441 | ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page) | |
2442 | { | |
2443 | if (!q->td) | |
2444 | return -EINVAL; | |
2445 | return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice)); | |
2446 | } | |
2447 | ||
2448 | ssize_t blk_throtl_sample_time_store(struct request_queue *q, | |
2449 | const char *page, size_t count) | |
2450 | { | |
2451 | unsigned long v; | |
2452 | unsigned long t; | |
2453 | ||
2454 | if (!q->td) | |
2455 | return -EINVAL; | |
2456 | if (kstrtoul(page, 10, &v)) | |
2457 | return -EINVAL; | |
2458 | t = msecs_to_jiffies(v); | |
2459 | if (t == 0 || t > MAX_THROTL_SLICE) | |
2460 | return -EINVAL; | |
2461 | q->td->throtl_slice = t; | |
2462 | return count; | |
2463 | } | |
2464 | #endif | |
2465 | ||
e43473b7 VG |
2466 | static int __init throtl_init(void) |
2467 | { | |
450adcbe VG |
2468 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
2469 | if (!kthrotld_workqueue) | |
2470 | panic("Failed to create kthrotld\n"); | |
2471 | ||
3c798398 | 2472 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
2473 | } |
2474 | ||
2475 | module_init(throtl_init); |