]>
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) |
e43473b7 | 28 | |
450adcbe VG |
29 | /* A workqueue to queue throttle related work */ |
30 | static struct workqueue_struct *kthrotld_workqueue; | |
450adcbe | 31 | |
e43473b7 VG |
32 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
33 | ||
e43473b7 VG |
34 | struct throtl_data |
35 | { | |
e43473b7 | 36 | /* service tree for active throtl groups */ |
c9e0332e | 37 | struct throtl_service_queue service_queue; |
e43473b7 | 38 | |
e43473b7 VG |
39 | struct request_queue *queue; |
40 | ||
41 | /* Total Number of queued bios on READ and WRITE lists */ | |
42 | unsigned int nr_queued[2]; | |
43 | ||
297e3d85 SL |
44 | unsigned int throtl_slice; |
45 | ||
e43473b7 | 46 | /* Work for dispatching throttled bios */ |
69df0ab0 | 47 | struct work_struct dispatch_work; |
b9147dd1 SL |
48 | |
49 | bool track_bio_latency; | |
e43473b7 VG |
50 | }; |
51 | ||
e99e88a9 | 52 | static void throtl_pending_timer_fn(struct timer_list *t); |
69df0ab0 | 53 | |
3c798398 | 54 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 55 | { |
f95a04af | 56 | return pd_to_blkg(&tg->pd); |
0381411e TH |
57 | } |
58 | ||
fda6f272 TH |
59 | /** |
60 | * sq_to_tg - return the throl_grp the specified service queue belongs to | |
61 | * @sq: the throtl_service_queue of interest | |
62 | * | |
63 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one | |
64 | * embedded in throtl_data, %NULL is returned. | |
65 | */ | |
66 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) | |
67 | { | |
68 | if (sq && sq->parent_sq) | |
69 | return container_of(sq, struct throtl_grp, service_queue); | |
70 | else | |
71 | return NULL; | |
72 | } | |
73 | ||
74 | /** | |
75 | * sq_to_td - return throtl_data the specified service queue belongs to | |
76 | * @sq: the throtl_service_queue of interest | |
77 | * | |
b43daedc | 78 | * A service_queue can be embedded in either a throtl_grp or throtl_data. |
fda6f272 TH |
79 | * Determine the associated throtl_data accordingly and return it. |
80 | */ | |
81 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) | |
82 | { | |
83 | struct throtl_grp *tg = sq_to_tg(sq); | |
84 | ||
85 | if (tg) | |
86 | return tg->td; | |
87 | else | |
88 | return container_of(sq, struct throtl_data, service_queue); | |
89 | } | |
90 | ||
9f626e37 SL |
91 | static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw) |
92 | { | |
b22c417c | 93 | struct blkcg_gq *blkg = tg_to_blkg(tg); |
b22c417c SL |
94 | |
95 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
96 | return U64_MAX; | |
7394e31f | 97 | |
bf20ab53 | 98 | return tg->bps[rw]; |
9f626e37 SL |
99 | } |
100 | ||
101 | static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw) | |
102 | { | |
b22c417c | 103 | struct blkcg_gq *blkg = tg_to_blkg(tg); |
b22c417c SL |
104 | |
105 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
106 | return UINT_MAX; | |
9bb67aeb | 107 | |
bf20ab53 | 108 | return tg->iops[rw]; |
9f626e37 SL |
109 | } |
110 | ||
fda6f272 TH |
111 | /** |
112 | * throtl_log - log debug message via blktrace | |
113 | * @sq: the service_queue being reported | |
114 | * @fmt: printf format string | |
115 | * @args: printf args | |
116 | * | |
117 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a | |
118 | * throtl_grp; otherwise, just "throtl". | |
fda6f272 TH |
119 | */ |
120 | #define throtl_log(sq, fmt, args...) do { \ | |
121 | struct throtl_grp *__tg = sq_to_tg((sq)); \ | |
122 | struct throtl_data *__td = sq_to_td((sq)); \ | |
123 | \ | |
124 | (void)__td; \ | |
59fa0224 SL |
125 | if (likely(!blk_trace_note_message_enabled(__td->queue))) \ |
126 | break; \ | |
fda6f272 | 127 | if ((__tg)) { \ |
35fe6d76 | 128 | blk_add_cgroup_trace_msg(__td->queue, \ |
f4a6a61c | 129 | &tg_to_blkg(__tg)->blkcg->css, "throtl " fmt, ##args);\ |
fda6f272 TH |
130 | } else { \ |
131 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
132 | } \ | |
54e7ed12 | 133 | } while (0) |
e43473b7 | 134 | |
ea0ea2bc SL |
135 | static inline unsigned int throtl_bio_data_size(struct bio *bio) |
136 | { | |
137 | /* assume it's one sector */ | |
138 | if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) | |
139 | return 512; | |
140 | return bio->bi_iter.bi_size; | |
141 | } | |
142 | ||
c5cc2070 TH |
143 | static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg) |
144 | { | |
145 | INIT_LIST_HEAD(&qn->node); | |
146 | bio_list_init(&qn->bios); | |
147 | qn->tg = tg; | |
148 | } | |
149 | ||
150 | /** | |
151 | * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it | |
152 | * @bio: bio being added | |
153 | * @qn: qnode to add bio to | |
154 | * @queued: the service_queue->queued[] list @qn belongs to | |
155 | * | |
156 | * Add @bio to @qn and put @qn on @queued if it's not already on. | |
157 | * @qn->tg's reference count is bumped when @qn is activated. See the | |
158 | * comment on top of throtl_qnode definition for details. | |
159 | */ | |
160 | static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn, | |
161 | struct list_head *queued) | |
162 | { | |
163 | bio_list_add(&qn->bios, bio); | |
164 | if (list_empty(&qn->node)) { | |
165 | list_add_tail(&qn->node, queued); | |
166 | blkg_get(tg_to_blkg(qn->tg)); | |
167 | } | |
168 | } | |
169 | ||
170 | /** | |
171 | * throtl_peek_queued - peek the first bio on a qnode list | |
172 | * @queued: the qnode list to peek | |
173 | */ | |
174 | static struct bio *throtl_peek_queued(struct list_head *queued) | |
175 | { | |
b7b609de | 176 | struct throtl_qnode *qn; |
c5cc2070 TH |
177 | struct bio *bio; |
178 | ||
179 | if (list_empty(queued)) | |
180 | return NULL; | |
181 | ||
b7b609de | 182 | qn = list_first_entry(queued, struct throtl_qnode, node); |
c5cc2070 TH |
183 | bio = bio_list_peek(&qn->bios); |
184 | WARN_ON_ONCE(!bio); | |
185 | return bio; | |
186 | } | |
187 | ||
188 | /** | |
189 | * throtl_pop_queued - pop the first bio form a qnode list | |
190 | * @queued: the qnode list to pop a bio from | |
191 | * @tg_to_put: optional out argument for throtl_grp to put | |
192 | * | |
193 | * Pop the first bio from the qnode list @queued. After popping, the first | |
194 | * qnode is removed from @queued if empty or moved to the end of @queued so | |
195 | * that the popping order is round-robin. | |
196 | * | |
197 | * When the first qnode is removed, its associated throtl_grp should be put | |
198 | * too. If @tg_to_put is NULL, this function automatically puts it; | |
199 | * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is | |
200 | * responsible for putting it. | |
201 | */ | |
202 | static struct bio *throtl_pop_queued(struct list_head *queued, | |
203 | struct throtl_grp **tg_to_put) | |
204 | { | |
b7b609de | 205 | struct throtl_qnode *qn; |
c5cc2070 TH |
206 | struct bio *bio; |
207 | ||
208 | if (list_empty(queued)) | |
209 | return NULL; | |
210 | ||
b7b609de | 211 | qn = list_first_entry(queued, struct throtl_qnode, node); |
c5cc2070 TH |
212 | bio = bio_list_pop(&qn->bios); |
213 | WARN_ON_ONCE(!bio); | |
214 | ||
215 | if (bio_list_empty(&qn->bios)) { | |
216 | list_del_init(&qn->node); | |
217 | if (tg_to_put) | |
218 | *tg_to_put = qn->tg; | |
219 | else | |
220 | blkg_put(tg_to_blkg(qn->tg)); | |
221 | } else { | |
222 | list_move_tail(&qn->node, queued); | |
223 | } | |
224 | ||
225 | return bio; | |
226 | } | |
227 | ||
49a2f1e3 | 228 | /* init a service_queue, assumes the caller zeroed it */ |
b2ce2643 | 229 | static void throtl_service_queue_init(struct throtl_service_queue *sq) |
49a2f1e3 | 230 | { |
7e9c5c54 YK |
231 | INIT_LIST_HEAD(&sq->queued[READ]); |
232 | INIT_LIST_HEAD(&sq->queued[WRITE]); | |
9ff01255 | 233 | sq->pending_tree = RB_ROOT_CACHED; |
e99e88a9 | 234 | timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0); |
69df0ab0 TH |
235 | } |
236 | ||
0a0b4f79 CH |
237 | static struct blkg_policy_data *throtl_pd_alloc(struct gendisk *disk, |
238 | struct blkcg *blkcg, gfp_t gfp) | |
001bea73 | 239 | { |
4fb72036 | 240 | struct throtl_grp *tg; |
24bdb8ef | 241 | int rw; |
4fb72036 | 242 | |
0a0b4f79 | 243 | tg = kzalloc_node(sizeof(*tg), gfp, disk->node_id); |
4fb72036 | 244 | if (!tg) |
77ea7338 | 245 | return NULL; |
4fb72036 | 246 | |
7ca46438 TH |
247 | if (blkg_rwstat_init(&tg->stat_bytes, gfp)) |
248 | goto err_free_tg; | |
249 | ||
250 | if (blkg_rwstat_init(&tg->stat_ios, gfp)) | |
251 | goto err_exit_stat_bytes; | |
252 | ||
b2ce2643 TH |
253 | throtl_service_queue_init(&tg->service_queue); |
254 | ||
255 | for (rw = READ; rw <= WRITE; rw++) { | |
256 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); | |
257 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); | |
258 | } | |
259 | ||
260 | RB_CLEAR_NODE(&tg->rb_node); | |
bf20ab53 YK |
261 | tg->bps[READ] = U64_MAX; |
262 | tg->bps[WRITE] = U64_MAX; | |
263 | tg->iops[READ] = UINT_MAX; | |
264 | tg->iops[WRITE] = UINT_MAX; | |
ec80991d | 265 | |
4fb72036 | 266 | return &tg->pd; |
7ca46438 TH |
267 | |
268 | err_exit_stat_bytes: | |
269 | blkg_rwstat_exit(&tg->stat_bytes); | |
270 | err_free_tg: | |
271 | kfree(tg); | |
272 | return NULL; | |
001bea73 TH |
273 | } |
274 | ||
a9520cd6 | 275 | static void throtl_pd_init(struct blkg_policy_data *pd) |
a29a171e | 276 | { |
a9520cd6 TH |
277 | struct throtl_grp *tg = pd_to_tg(pd); |
278 | struct blkcg_gq *blkg = tg_to_blkg(tg); | |
a06377c5 | 279 | struct throtl_data *td = blkg->q->td; |
b2ce2643 | 280 | struct throtl_service_queue *sq = &tg->service_queue; |
cd1604fa | 281 | |
9138125b | 282 | /* |
aa6ec29b | 283 | * If on the default hierarchy, we switch to properly hierarchical |
9138125b TH |
284 | * behavior where limits on a given throtl_grp are applied to the |
285 | * whole subtree rather than just the group itself. e.g. If 16M | |
f56019ae KS |
286 | * read_bps limit is set on a parent group, summary bps of |
287 | * parent group and its subtree groups can't exceed 16M for the | |
288 | * device. | |
9138125b | 289 | * |
aa6ec29b | 290 | * If not on the default hierarchy, the broken flat hierarchy |
9138125b TH |
291 | * behavior is retained where all throtl_grps are treated as if |
292 | * they're all separate root groups right below throtl_data. | |
293 | * Limits of a group don't interact with limits of other groups | |
294 | * regardless of the position of the group in the hierarchy. | |
295 | */ | |
b2ce2643 | 296 | sq->parent_sq = &td->service_queue; |
9e10a130 | 297 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent) |
b2ce2643 | 298 | sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue; |
77216b04 | 299 | tg->td = td; |
8a3d2615 TH |
300 | } |
301 | ||
693e751e TH |
302 | /* |
303 | * Set has_rules[] if @tg or any of its parents have limits configured. | |
304 | * This doesn't require walking up to the top of the hierarchy as the | |
305 | * parent's has_rules[] is guaranteed to be correct. | |
306 | */ | |
307 | static void tg_update_has_rules(struct throtl_grp *tg) | |
308 | { | |
309 | struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); | |
310 | int rw; | |
311 | ||
81c7a63a YK |
312 | for (rw = READ; rw <= WRITE; rw++) { |
313 | tg->has_rules_iops[rw] = | |
314 | (parent_tg && parent_tg->has_rules_iops[rw]) || | |
bf20ab53 | 315 | tg_iops_limit(tg, rw) != UINT_MAX; |
81c7a63a YK |
316 | tg->has_rules_bps[rw] = |
317 | (parent_tg && parent_tg->has_rules_bps[rw]) || | |
bf20ab53 | 318 | tg_bps_limit(tg, rw) != U64_MAX; |
81c7a63a | 319 | } |
693e751e TH |
320 | } |
321 | ||
a9520cd6 | 322 | static void throtl_pd_online(struct blkg_policy_data *pd) |
693e751e | 323 | { |
aec24246 | 324 | struct throtl_grp *tg = pd_to_tg(pd); |
693e751e TH |
325 | /* |
326 | * We don't want new groups to escape the limits of its ancestors. | |
327 | * Update has_rules[] after a new group is brought online. | |
328 | */ | |
aec24246 | 329 | tg_update_has_rules(tg); |
693e751e TH |
330 | } |
331 | ||
001bea73 TH |
332 | static void throtl_pd_free(struct blkg_policy_data *pd) |
333 | { | |
4fb72036 TH |
334 | struct throtl_grp *tg = pd_to_tg(pd); |
335 | ||
b2ce2643 | 336 | del_timer_sync(&tg->service_queue.pending_timer); |
7ca46438 TH |
337 | blkg_rwstat_exit(&tg->stat_bytes); |
338 | blkg_rwstat_exit(&tg->stat_ios); | |
4fb72036 | 339 | kfree(tg); |
001bea73 TH |
340 | } |
341 | ||
0049af73 TH |
342 | static struct throtl_grp * |
343 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 | 344 | { |
9ff01255 | 345 | struct rb_node *n; |
e43473b7 | 346 | |
9ff01255 LB |
347 | n = rb_first_cached(&parent_sq->pending_tree); |
348 | WARN_ON_ONCE(!n); | |
349 | if (!n) | |
350 | return NULL; | |
351 | return rb_entry_tg(n); | |
e43473b7 VG |
352 | } |
353 | ||
0049af73 TH |
354 | static void throtl_rb_erase(struct rb_node *n, |
355 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 356 | { |
9ff01255 LB |
357 | rb_erase_cached(n, &parent_sq->pending_tree); |
358 | RB_CLEAR_NODE(n); | |
e43473b7 VG |
359 | } |
360 | ||
0049af73 | 361 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
362 | { |
363 | struct throtl_grp *tg; | |
364 | ||
0049af73 | 365 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
366 | if (!tg) |
367 | return; | |
368 | ||
0049af73 | 369 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
370 | } |
371 | ||
77216b04 | 372 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 373 | { |
77216b04 | 374 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
9ff01255 | 375 | struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node; |
e43473b7 VG |
376 | struct rb_node *parent = NULL; |
377 | struct throtl_grp *__tg; | |
378 | unsigned long key = tg->disptime; | |
9ff01255 | 379 | bool leftmost = true; |
e43473b7 VG |
380 | |
381 | while (*node != NULL) { | |
382 | parent = *node; | |
383 | __tg = rb_entry_tg(parent); | |
384 | ||
385 | if (time_before(key, __tg->disptime)) | |
386 | node = &parent->rb_left; | |
387 | else { | |
388 | node = &parent->rb_right; | |
9ff01255 | 389 | leftmost = false; |
e43473b7 VG |
390 | } |
391 | } | |
392 | ||
e43473b7 | 393 | rb_link_node(&tg->rb_node, parent, node); |
9ff01255 LB |
394 | rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree, |
395 | leftmost); | |
e43473b7 VG |
396 | } |
397 | ||
77216b04 | 398 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 399 | { |
29379674 BW |
400 | if (!(tg->flags & THROTL_TG_PENDING)) { |
401 | tg_service_queue_add(tg); | |
402 | tg->flags |= THROTL_TG_PENDING; | |
403 | tg->service_queue.parent_sq->nr_pending++; | |
404 | } | |
e43473b7 VG |
405 | } |
406 | ||
77216b04 | 407 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 408 | { |
29379674 | 409 | if (tg->flags & THROTL_TG_PENDING) { |
c013710e YK |
410 | struct throtl_service_queue *parent_sq = |
411 | tg->service_queue.parent_sq; | |
412 | ||
413 | throtl_rb_erase(&tg->rb_node, parent_sq); | |
414 | --parent_sq->nr_pending; | |
29379674 BW |
415 | tg->flags &= ~THROTL_TG_PENDING; |
416 | } | |
e43473b7 VG |
417 | } |
418 | ||
a9131a27 | 419 | /* Call with queue lock held */ |
69df0ab0 TH |
420 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, |
421 | unsigned long expires) | |
a9131a27 | 422 | { |
a41b816c | 423 | unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice; |
06cceedc SL |
424 | |
425 | /* | |
426 | * Since we are adjusting the throttle limit dynamically, the sleep | |
427 | * time calculated according to previous limit might be invalid. It's | |
428 | * possible the cgroup sleep time is very long and no other cgroups | |
429 | * have IO running so notify the limit changes. Make sure the cgroup | |
430 | * doesn't sleep too long to avoid the missed notification. | |
431 | */ | |
432 | if (time_after(expires, max_expire)) | |
433 | expires = max_expire; | |
69df0ab0 TH |
434 | mod_timer(&sq->pending_timer, expires); |
435 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
436 | expires - jiffies, jiffies); | |
a9131a27 TH |
437 | } |
438 | ||
7f52f98c TH |
439 | /** |
440 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle | |
441 | * @sq: the service_queue to schedule dispatch for | |
442 | * @force: force scheduling | |
443 | * | |
444 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the | |
445 | * dispatch time of the first pending child. Returns %true if either timer | |
446 | * is armed or there's no pending child left. %false if the current | |
447 | * dispatch window is still open and the caller should continue | |
448 | * dispatching. | |
449 | * | |
450 | * If @force is %true, the dispatch timer is always scheduled and this | |
451 | * function is guaranteed to return %true. This is to be used when the | |
452 | * caller can't dispatch itself and needs to invoke pending_timer | |
453 | * unconditionally. Note that forced scheduling is likely to induce short | |
454 | * delay before dispatch starts even if @sq->first_pending_disptime is not | |
455 | * in the future and thus shouldn't be used in hot paths. | |
456 | */ | |
457 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
458 | bool force) | |
e43473b7 | 459 | { |
6a525600 | 460 | /* any pending children left? */ |
c9e0332e | 461 | if (!sq->nr_pending) |
7f52f98c | 462 | return true; |
e43473b7 | 463 | |
c9e0332e | 464 | update_min_dispatch_time(sq); |
e43473b7 | 465 | |
69df0ab0 | 466 | /* is the next dispatch time in the future? */ |
7f52f98c | 467 | if (force || time_after(sq->first_pending_disptime, jiffies)) { |
69df0ab0 | 468 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); |
7f52f98c | 469 | return true; |
69df0ab0 TH |
470 | } |
471 | ||
7f52f98c TH |
472 | /* tell the caller to continue dispatching */ |
473 | return false; | |
e43473b7 VG |
474 | } |
475 | ||
32ee5bc4 VG |
476 | static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, |
477 | bool rw, unsigned long start) | |
478 | { | |
479 | tg->bytes_disp[rw] = 0; | |
480 | tg->io_disp[rw] = 0; | |
a880ae93 YK |
481 | tg->carryover_bytes[rw] = 0; |
482 | tg->carryover_ios[rw] = 0; | |
32ee5bc4 VG |
483 | |
484 | /* | |
485 | * Previous slice has expired. We must have trimmed it after last | |
486 | * bio dispatch. That means since start of last slice, we never used | |
487 | * that bandwidth. Do try to make use of that bandwidth while giving | |
488 | * credit. | |
489 | */ | |
eea3e8b7 | 490 | if (time_after(start, tg->slice_start[rw])) |
32ee5bc4 VG |
491 | tg->slice_start[rw] = start; |
492 | ||
297e3d85 | 493 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; |
32ee5bc4 VG |
494 | throtl_log(&tg->service_queue, |
495 | "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", | |
496 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
497 | tg->slice_end[rw], jiffies); | |
498 | } | |
499 | ||
a880ae93 YK |
500 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw, |
501 | bool clear_carryover) | |
e43473b7 VG |
502 | { |
503 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 504 | tg->io_disp[rw] = 0; |
e43473b7 | 505 | tg->slice_start[rw] = jiffies; |
297e3d85 | 506 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; |
a880ae93 YK |
507 | if (clear_carryover) { |
508 | tg->carryover_bytes[rw] = 0; | |
509 | tg->carryover_ios[rw] = 0; | |
510 | } | |
4f1e9630 | 511 | |
fda6f272 TH |
512 | throtl_log(&tg->service_queue, |
513 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
514 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
515 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
516 | } |
517 | ||
0f3457f6 TH |
518 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
519 | unsigned long jiffy_end) | |
d1ae8ffd | 520 | { |
297e3d85 | 521 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); |
d1ae8ffd VG |
522 | } |
523 | ||
0f3457f6 TH |
524 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
525 | unsigned long jiffy_end) | |
e43473b7 | 526 | { |
1da30f95 | 527 | throtl_set_slice_end(tg, rw, jiffy_end); |
fda6f272 TH |
528 | throtl_log(&tg->service_queue, |
529 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
530 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
531 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
532 | } |
533 | ||
534 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 535 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
536 | { |
537 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
5cf8c227 | 538 | return false; |
e43473b7 | 539 | |
0b6bad7d | 540 | return true; |
e43473b7 VG |
541 | } |
542 | ||
e8368b57 YK |
543 | static unsigned int calculate_io_allowed(u32 iops_limit, |
544 | unsigned long jiffy_elapsed) | |
545 | { | |
546 | unsigned int io_allowed; | |
547 | u64 tmp; | |
548 | ||
549 | /* | |
550 | * jiffy_elapsed should not be a big value as minimum iops can be | |
551 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
552 | * will allow dispatch after 1 second and after that slice should | |
553 | * have been trimmed. | |
554 | */ | |
555 | ||
556 | tmp = (u64)iops_limit * jiffy_elapsed; | |
557 | do_div(tmp, HZ); | |
558 | ||
559 | if (tmp > UINT_MAX) | |
560 | io_allowed = UINT_MAX; | |
561 | else | |
562 | io_allowed = tmp; | |
563 | ||
564 | return io_allowed; | |
565 | } | |
566 | ||
567 | static u64 calculate_bytes_allowed(u64 bps_limit, unsigned long jiffy_elapsed) | |
568 | { | |
2dd710d4 KK |
569 | /* |
570 | * Can result be wider than 64 bits? | |
571 | * We check against 62, not 64, due to ilog2 truncation. | |
572 | */ | |
573 | if (ilog2(bps_limit) + ilog2(jiffy_elapsed) - ilog2(HZ) > 62) | |
574 | return U64_MAX; | |
e8368b57 YK |
575 | return mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed, (u64)HZ); |
576 | } | |
577 | ||
e43473b7 | 578 | /* Trim the used slices and adjust slice start accordingly */ |
0f3457f6 | 579 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 580 | { |
eead0056 YK |
581 | unsigned long time_elapsed; |
582 | long long bytes_trim; | |
583 | int io_trim; | |
e43473b7 VG |
584 | |
585 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
586 | ||
587 | /* | |
588 | * If bps are unlimited (-1), then time slice don't get | |
589 | * renewed. Don't try to trim the slice if slice is used. A new | |
590 | * slice will start when appropriate. | |
591 | */ | |
0f3457f6 | 592 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
593 | return; |
594 | ||
d1ae8ffd VG |
595 | /* |
596 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
597 | * that initially cgroup limit was very low resulting in high | |
b53b072c | 598 | * slice_end, but later limit was bumped up and bio was dispatched |
d1ae8ffd VG |
599 | * sooner, then we need to reduce slice_end. A high bogus slice_end |
600 | * is bad because it does not allow new slice to start. | |
601 | */ | |
602 | ||
297e3d85 | 603 | throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice); |
d1ae8ffd | 604 | |
e8368b57 YK |
605 | time_elapsed = rounddown(jiffies - tg->slice_start[rw], |
606 | tg->td->throtl_slice); | |
607 | if (!time_elapsed) | |
e43473b7 | 608 | return; |
e43473b7 | 609 | |
e8368b57 | 610 | bytes_trim = calculate_bytes_allowed(tg_bps_limit(tg, rw), |
eead0056 YK |
611 | time_elapsed) + |
612 | tg->carryover_bytes[rw]; | |
613 | io_trim = calculate_io_allowed(tg_iops_limit(tg, rw), time_elapsed) + | |
614 | tg->carryover_ios[rw]; | |
615 | if (bytes_trim <= 0 && io_trim <= 0) | |
e43473b7 VG |
616 | return; |
617 | ||
eead0056 YK |
618 | tg->carryover_bytes[rw] = 0; |
619 | if ((long long)tg->bytes_disp[rw] >= bytes_trim) | |
e43473b7 VG |
620 | tg->bytes_disp[rw] -= bytes_trim; |
621 | else | |
622 | tg->bytes_disp[rw] = 0; | |
623 | ||
eead0056 YK |
624 | tg->carryover_ios[rw] = 0; |
625 | if ((int)tg->io_disp[rw] >= io_trim) | |
8e89d13f VG |
626 | tg->io_disp[rw] -= io_trim; |
627 | else | |
628 | tg->io_disp[rw] = 0; | |
629 | ||
e8368b57 | 630 | tg->slice_start[rw] += time_elapsed; |
e43473b7 | 631 | |
fda6f272 | 632 | throtl_log(&tg->service_queue, |
eead0056 | 633 | "[%c] trim slice nr=%lu bytes=%lld io=%d start=%lu end=%lu jiffies=%lu", |
e8368b57 YK |
634 | rw == READ ? 'R' : 'W', time_elapsed / tg->td->throtl_slice, |
635 | bytes_trim, io_trim, tg->slice_start[rw], tg->slice_end[rw], | |
636 | jiffies); | |
681cd46f YK |
637 | } |
638 | ||
a880ae93 YK |
639 | static void __tg_update_carryover(struct throtl_grp *tg, bool rw) |
640 | { | |
641 | unsigned long jiffy_elapsed = jiffies - tg->slice_start[rw]; | |
642 | u64 bps_limit = tg_bps_limit(tg, rw); | |
643 | u32 iops_limit = tg_iops_limit(tg, rw); | |
644 | ||
645 | /* | |
646 | * If config is updated while bios are still throttled, calculate and | |
647 | * accumulate how many bytes/ios are waited across changes. And | |
648 | * carryover_bytes/ios will be used to calculate new wait time under new | |
649 | * configuration. | |
650 | */ | |
651 | if (bps_limit != U64_MAX) | |
652 | tg->carryover_bytes[rw] += | |
653 | calculate_bytes_allowed(bps_limit, jiffy_elapsed) - | |
654 | tg->bytes_disp[rw]; | |
655 | if (iops_limit != UINT_MAX) | |
656 | tg->carryover_ios[rw] += | |
657 | calculate_io_allowed(iops_limit, jiffy_elapsed) - | |
658 | tg->io_disp[rw]; | |
659 | } | |
660 | ||
661 | static void tg_update_carryover(struct throtl_grp *tg) | |
662 | { | |
663 | if (tg->service_queue.nr_queued[READ]) | |
664 | __tg_update_carryover(tg, READ); | |
665 | if (tg->service_queue.nr_queued[WRITE]) | |
666 | __tg_update_carryover(tg, WRITE); | |
667 | ||
668 | /* see comments in struct throtl_grp for meaning of these fields. */ | |
ef100397 | 669 | throtl_log(&tg->service_queue, "%s: %lld %lld %d %d\n", __func__, |
a880ae93 YK |
670 | tg->carryover_bytes[READ], tg->carryover_bytes[WRITE], |
671 | tg->carryover_ios[READ], tg->carryover_ios[WRITE]); | |
672 | } | |
673 | ||
183daeb1 KS |
674 | static unsigned long tg_within_iops_limit(struct throtl_grp *tg, struct bio *bio, |
675 | u32 iops_limit) | |
681cd46f YK |
676 | { |
677 | bool rw = bio_data_dir(bio); | |
bb8d5587 | 678 | int io_allowed; |
681cd46f YK |
679 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
680 | ||
681 | if (iops_limit == UINT_MAX) { | |
183daeb1 | 682 | return 0; |
681cd46f YK |
683 | } |
684 | ||
685 | jiffy_elapsed = jiffies - tg->slice_start[rw]; | |
686 | ||
687 | /* Round up to the next throttle slice, wait time must be nonzero */ | |
688 | jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice); | |
a880ae93 YK |
689 | io_allowed = calculate_io_allowed(iops_limit, jiffy_elapsed_rnd) + |
690 | tg->carryover_ios[rw]; | |
bb8d5587 | 691 | if (io_allowed > 0 && tg->io_disp[rw] + 1 <= io_allowed) |
183daeb1 | 692 | return 0; |
e43473b7 | 693 | |
8e89d13f | 694 | /* Calc approx time to dispatch */ |
991f61fe | 695 | jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed; |
1beabab8 YK |
696 | |
697 | /* make sure at least one io can be dispatched after waiting */ | |
698 | jiffy_wait = max(jiffy_wait, HZ / iops_limit + 1); | |
183daeb1 | 699 | return jiffy_wait; |
8e89d13f VG |
700 | } |
701 | ||
183daeb1 KS |
702 | static unsigned long tg_within_bps_limit(struct throtl_grp *tg, struct bio *bio, |
703 | u64 bps_limit) | |
8e89d13f VG |
704 | { |
705 | bool rw = bio_data_dir(bio); | |
bb8d5587 YK |
706 | long long bytes_allowed; |
707 | u64 extra_bytes; | |
8e89d13f | 708 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
ea0ea2bc | 709 | unsigned int bio_size = throtl_bio_data_size(bio); |
e43473b7 | 710 | |
9f5ede3c | 711 | /* no need to throttle if this bio's bytes have been accounted */ |
320fb0f9 | 712 | if (bps_limit == U64_MAX || bio_flagged(bio, BIO_BPS_THROTTLED)) { |
183daeb1 | 713 | return 0; |
87fbeb88 BW |
714 | } |
715 | ||
e43473b7 VG |
716 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
717 | ||
718 | /* Slice has just started. Consider one slice interval */ | |
719 | if (!jiffy_elapsed) | |
297e3d85 | 720 | jiffy_elapsed_rnd = tg->td->throtl_slice; |
e43473b7 | 721 | |
297e3d85 | 722 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); |
a880ae93 YK |
723 | bytes_allowed = calculate_bytes_allowed(bps_limit, jiffy_elapsed_rnd) + |
724 | tg->carryover_bytes[rw]; | |
bb8d5587 | 725 | if (bytes_allowed > 0 && tg->bytes_disp[rw] + bio_size <= bytes_allowed) |
183daeb1 | 726 | return 0; |
e43473b7 VG |
727 | |
728 | /* Calc approx time to dispatch */ | |
ea0ea2bc | 729 | extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed; |
4599ea49 | 730 | jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit); |
e43473b7 VG |
731 | |
732 | if (!jiffy_wait) | |
733 | jiffy_wait = 1; | |
734 | ||
735 | /* | |
736 | * This wait time is without taking into consideration the rounding | |
737 | * up we did. Add that time also. | |
738 | */ | |
739 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
183daeb1 | 740 | return jiffy_wait; |
8e89d13f VG |
741 | } |
742 | ||
743 | /* | |
744 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
745 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
746 | */ | |
0f3457f6 TH |
747 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
748 | unsigned long *wait) | |
8e89d13f VG |
749 | { |
750 | bool rw = bio_data_dir(bio); | |
751 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
4599ea49 BW |
752 | u64 bps_limit = tg_bps_limit(tg, rw); |
753 | u32 iops_limit = tg_iops_limit(tg, rw); | |
8e89d13f VG |
754 | |
755 | /* | |
756 | * Currently whole state machine of group depends on first bio | |
757 | * queued in the group bio list. So one should not be calling | |
758 | * this function with a different bio if there are other bios | |
759 | * queued. | |
760 | */ | |
73f0d49a | 761 | BUG_ON(tg->service_queue.nr_queued[rw] && |
c5cc2070 | 762 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); |
e43473b7 | 763 | |
8e89d13f | 764 | /* If tg->bps = -1, then BW is unlimited */ |
8f9e7b65 YK |
765 | if ((bps_limit == U64_MAX && iops_limit == UINT_MAX) || |
766 | tg->flags & THROTL_TG_CANCELING) { | |
8e89d13f VG |
767 | if (wait) |
768 | *wait = 0; | |
5cf8c227 | 769 | return true; |
8e89d13f VG |
770 | } |
771 | ||
772 | /* | |
773 | * If previous slice expired, start a new one otherwise renew/extend | |
774 | * existing slice to make sure it is at least throtl_slice interval | |
164c80ed VG |
775 | * long since now. New slice is started only for empty throttle group. |
776 | * If there is queued bio, that means there should be an active | |
777 | * slice and it should be extended instead. | |
8e89d13f | 778 | */ |
164c80ed | 779 | if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw])) |
a880ae93 | 780 | throtl_start_new_slice(tg, rw, true); |
8e89d13f | 781 | else { |
297e3d85 SL |
782 | if (time_before(tg->slice_end[rw], |
783 | jiffies + tg->td->throtl_slice)) | |
784 | throtl_extend_slice(tg, rw, | |
785 | jiffies + tg->td->throtl_slice); | |
8e89d13f VG |
786 | } |
787 | ||
183daeb1 KS |
788 | bps_wait = tg_within_bps_limit(tg, bio, bps_limit); |
789 | iops_wait = tg_within_iops_limit(tg, bio, iops_limit); | |
790 | if (bps_wait + iops_wait == 0) { | |
8e89d13f VG |
791 | if (wait) |
792 | *wait = 0; | |
0b6bad7d | 793 | return true; |
8e89d13f VG |
794 | } |
795 | ||
796 | max_wait = max(bps_wait, iops_wait); | |
797 | ||
798 | if (wait) | |
799 | *wait = max_wait; | |
800 | ||
801 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 802 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 | 803 | |
0b6bad7d | 804 | return false; |
e43473b7 VG |
805 | } |
806 | ||
807 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
808 | { | |
809 | bool rw = bio_data_dir(bio); | |
ea0ea2bc | 810 | unsigned int bio_size = throtl_bio_data_size(bio); |
e43473b7 VG |
811 | |
812 | /* Charge the bio to the group */ | |
320fb0f9 | 813 | if (!bio_flagged(bio, BIO_BPS_THROTTLED)) { |
9f5ede3c ML |
814 | tg->bytes_disp[rw] += bio_size; |
815 | tg->last_bytes_disp[rw] += bio_size; | |
816 | } | |
817 | ||
8e89d13f | 818 | tg->io_disp[rw]++; |
3f0abd80 | 819 | tg->last_io_disp[rw]++; |
e43473b7 VG |
820 | } |
821 | ||
c5cc2070 TH |
822 | /** |
823 | * throtl_add_bio_tg - add a bio to the specified throtl_grp | |
824 | * @bio: bio to add | |
825 | * @qn: qnode to use | |
826 | * @tg: the target throtl_grp | |
827 | * | |
828 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, | |
829 | * tg->qnode_on_self[] is used. | |
830 | */ | |
831 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, | |
832 | struct throtl_grp *tg) | |
e43473b7 | 833 | { |
73f0d49a | 834 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
835 | bool rw = bio_data_dir(bio); |
836 | ||
c5cc2070 TH |
837 | if (!qn) |
838 | qn = &tg->qnode_on_self[rw]; | |
839 | ||
0e9f4164 TH |
840 | /* |
841 | * If @tg doesn't currently have any bios queued in the same | |
842 | * direction, queueing @bio can change when @tg should be | |
843 | * dispatched. Mark that @tg was empty. This is automatically | |
b53b072c | 844 | * cleared on the next tg_update_disptime(). |
0e9f4164 TH |
845 | */ |
846 | if (!sq->nr_queued[rw]) | |
847 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
848 | ||
c5cc2070 TH |
849 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); |
850 | ||
73f0d49a | 851 | sq->nr_queued[rw]++; |
77216b04 | 852 | throtl_enqueue_tg(tg); |
e43473b7 VG |
853 | } |
854 | ||
77216b04 | 855 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 856 | { |
73f0d49a | 857 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
858 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
859 | struct bio *bio; | |
860 | ||
d609af3a ME |
861 | bio = throtl_peek_queued(&sq->queued[READ]); |
862 | if (bio) | |
0f3457f6 | 863 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 864 | |
d609af3a ME |
865 | bio = throtl_peek_queued(&sq->queued[WRITE]); |
866 | if (bio) | |
0f3457f6 | 867 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
868 | |
869 | min_wait = min(read_wait, write_wait); | |
870 | disptime = jiffies + min_wait; | |
871 | ||
e43473b7 | 872 | /* Update dispatch time */ |
c013710e | 873 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
e43473b7 | 874 | tg->disptime = disptime; |
c013710e | 875 | tg_service_queue_add(tg); |
0e9f4164 TH |
876 | |
877 | /* see throtl_add_bio_tg() */ | |
878 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
879 | } |
880 | ||
32ee5bc4 VG |
881 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, |
882 | struct throtl_grp *parent_tg, bool rw) | |
883 | { | |
884 | if (throtl_slice_used(parent_tg, rw)) { | |
885 | throtl_start_new_slice_with_credit(parent_tg, rw, | |
886 | child_tg->slice_start[rw]); | |
887 | } | |
888 | ||
889 | } | |
890 | ||
77216b04 | 891 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 892 | { |
73f0d49a | 893 | struct throtl_service_queue *sq = &tg->service_queue; |
6bc9c2b4 TH |
894 | struct throtl_service_queue *parent_sq = sq->parent_sq; |
895 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
c5cc2070 | 896 | struct throtl_grp *tg_to_put = NULL; |
e43473b7 VG |
897 | struct bio *bio; |
898 | ||
c5cc2070 TH |
899 | /* |
900 | * @bio is being transferred from @tg to @parent_sq. Popping a bio | |
901 | * from @tg may put its reference and @parent_sq might end up | |
902 | * getting released prematurely. Remember the tg to put and put it | |
903 | * after @bio is transferred to @parent_sq. | |
904 | */ | |
905 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); | |
73f0d49a | 906 | sq->nr_queued[rw]--; |
e43473b7 VG |
907 | |
908 | throtl_charge_bio(tg, bio); | |
6bc9c2b4 TH |
909 | |
910 | /* | |
911 | * If our parent is another tg, we just need to transfer @bio to | |
912 | * the parent using throtl_add_bio_tg(). If our parent is | |
913 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
914 | * bio_lists[] and decrease total number queued. The caller is | |
915 | * responsible for issuing these bios. | |
916 | */ | |
917 | if (parent_tg) { | |
c5cc2070 | 918 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); |
32ee5bc4 | 919 | start_parent_slice_with_credit(tg, parent_tg, rw); |
6bc9c2b4 | 920 | } else { |
84aca0a7 | 921 | bio_set_flag(bio, BIO_BPS_THROTTLED); |
c5cc2070 TH |
922 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], |
923 | &parent_sq->queued[rw]); | |
6bc9c2b4 TH |
924 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
925 | tg->td->nr_queued[rw]--; | |
926 | } | |
e43473b7 | 927 | |
0f3457f6 | 928 | throtl_trim_slice(tg, rw); |
6bc9c2b4 | 929 | |
c5cc2070 TH |
930 | if (tg_to_put) |
931 | blkg_put(tg_to_blkg(tg_to_put)); | |
e43473b7 VG |
932 | } |
933 | ||
77216b04 | 934 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 935 | { |
73f0d49a | 936 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 | 937 | unsigned int nr_reads = 0, nr_writes = 0; |
e675df2a BW |
938 | unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4; |
939 | unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads; | |
e43473b7 VG |
940 | struct bio *bio; |
941 | ||
942 | /* Try to dispatch 75% READS and 25% WRITES */ | |
943 | ||
c5cc2070 | 944 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && |
0f3457f6 | 945 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 946 | |
3bca7640 | 947 | tg_dispatch_one_bio(tg, READ); |
e43473b7 VG |
948 | nr_reads++; |
949 | ||
950 | if (nr_reads >= max_nr_reads) | |
951 | break; | |
952 | } | |
953 | ||
c5cc2070 | 954 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && |
0f3457f6 | 955 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 956 | |
3bca7640 | 957 | tg_dispatch_one_bio(tg, WRITE); |
e43473b7 VG |
958 | nr_writes++; |
959 | ||
960 | if (nr_writes >= max_nr_writes) | |
961 | break; | |
962 | } | |
963 | ||
964 | return nr_reads + nr_writes; | |
965 | } | |
966 | ||
651930bc | 967 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
968 | { |
969 | unsigned int nr_disp = 0; | |
e43473b7 VG |
970 | |
971 | while (1) { | |
2397611a | 972 | struct throtl_grp *tg; |
2ab74cd2 | 973 | struct throtl_service_queue *sq; |
e43473b7 | 974 | |
2397611a BW |
975 | if (!parent_sq->nr_pending) |
976 | break; | |
977 | ||
978 | tg = throtl_rb_first(parent_sq); | |
e43473b7 VG |
979 | if (!tg) |
980 | break; | |
981 | ||
982 | if (time_before(jiffies, tg->disptime)) | |
983 | break; | |
984 | ||
77216b04 | 985 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 986 | |
2ab74cd2 | 987 | sq = &tg->service_queue; |
7e9c5c54 | 988 | if (sq->nr_queued[READ] || sq->nr_queued[WRITE]) |
77216b04 | 989 | tg_update_disptime(tg); |
8c25ed0c YK |
990 | else |
991 | throtl_dequeue_tg(tg); | |
e43473b7 | 992 | |
e675df2a | 993 | if (nr_disp >= THROTL_QUANTUM) |
e43473b7 VG |
994 | break; |
995 | } | |
996 | ||
997 | return nr_disp; | |
998 | } | |
999 | ||
6e1a5704 TH |
1000 | /** |
1001 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
216382dc | 1002 | * @t: the pending_timer member of the throtl_service_queue being serviced |
6e1a5704 TH |
1003 | * |
1004 | * This timer is armed when a child throtl_grp with active bio's become | |
1005 | * pending and queued on the service_queue's pending_tree and expires when | |
1006 | * the first child throtl_grp should be dispatched. This function | |
2e48a530 TH |
1007 | * dispatches bio's from the children throtl_grps to the parent |
1008 | * service_queue. | |
1009 | * | |
1010 | * If the parent's parent is another throtl_grp, dispatching is propagated | |
1011 | * by either arming its pending_timer or repeating dispatch directly. If | |
1012 | * the top-level service_tree is reached, throtl_data->dispatch_work is | |
1013 | * kicked so that the ready bio's are issued. | |
6e1a5704 | 1014 | */ |
e99e88a9 | 1015 | static void throtl_pending_timer_fn(struct timer_list *t) |
69df0ab0 | 1016 | { |
e99e88a9 | 1017 | struct throtl_service_queue *sq = from_timer(sq, t, pending_timer); |
2e48a530 | 1018 | struct throtl_grp *tg = sq_to_tg(sq); |
69df0ab0 | 1019 | struct throtl_data *td = sq_to_td(sq); |
2e48a530 | 1020 | struct throtl_service_queue *parent_sq; |
ee37eddb | 1021 | struct request_queue *q; |
2e48a530 | 1022 | bool dispatched; |
6e1a5704 | 1023 | int ret; |
e43473b7 | 1024 | |
ee37eddb ML |
1025 | /* throtl_data may be gone, so figure out request queue by blkg */ |
1026 | if (tg) | |
a06377c5 | 1027 | q = tg->pd.blkg->q; |
ee37eddb ML |
1028 | else |
1029 | q = td->queue; | |
1030 | ||
0d945c1f | 1031 | spin_lock_irq(&q->queue_lock); |
ee37eddb | 1032 | |
1231039d | 1033 | if (!q->root_blkg) |
ee37eddb ML |
1034 | goto out_unlock; |
1035 | ||
2e48a530 TH |
1036 | again: |
1037 | parent_sq = sq->parent_sq; | |
1038 | dispatched = false; | |
e43473b7 | 1039 | |
7f52f98c TH |
1040 | while (true) { |
1041 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
2e48a530 TH |
1042 | sq->nr_queued[READ] + sq->nr_queued[WRITE], |
1043 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
7f52f98c TH |
1044 | |
1045 | ret = throtl_select_dispatch(sq); | |
1046 | if (ret) { | |
7f52f98c TH |
1047 | throtl_log(sq, "bios disp=%u", ret); |
1048 | dispatched = true; | |
1049 | } | |
e43473b7 | 1050 | |
7f52f98c TH |
1051 | if (throtl_schedule_next_dispatch(sq, false)) |
1052 | break; | |
e43473b7 | 1053 | |
7f52f98c | 1054 | /* this dispatch windows is still open, relax and repeat */ |
0d945c1f | 1055 | spin_unlock_irq(&q->queue_lock); |
7f52f98c | 1056 | cpu_relax(); |
0d945c1f | 1057 | spin_lock_irq(&q->queue_lock); |
651930bc | 1058 | } |
e43473b7 | 1059 | |
2e48a530 TH |
1060 | if (!dispatched) |
1061 | goto out_unlock; | |
6e1a5704 | 1062 | |
2e48a530 TH |
1063 | if (parent_sq) { |
1064 | /* @parent_sq is another throl_grp, propagate dispatch */ | |
1065 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1066 | tg_update_disptime(tg); | |
1067 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { | |
1068 | /* window is already open, repeat dispatching */ | |
1069 | sq = parent_sq; | |
1070 | tg = sq_to_tg(sq); | |
1071 | goto again; | |
1072 | } | |
1073 | } | |
1074 | } else { | |
b53b072c | 1075 | /* reached the top-level, queue issuing */ |
2e48a530 TH |
1076 | queue_work(kthrotld_workqueue, &td->dispatch_work); |
1077 | } | |
1078 | out_unlock: | |
0d945c1f | 1079 | spin_unlock_irq(&q->queue_lock); |
6e1a5704 | 1080 | } |
e43473b7 | 1081 | |
6e1a5704 TH |
1082 | /** |
1083 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
1084 | * @work: work item being executed | |
1085 | * | |
b53b072c BW |
1086 | * This function is queued for execution when bios reach the bio_lists[] |
1087 | * of throtl_data->service_queue. Those bios are ready and issued by this | |
6e1a5704 TH |
1088 | * function. |
1089 | */ | |
8876e140 | 1090 | static void blk_throtl_dispatch_work_fn(struct work_struct *work) |
6e1a5704 TH |
1091 | { |
1092 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1093 | dispatch_work); | |
1094 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1095 | struct request_queue *q = td->queue; | |
1096 | struct bio_list bio_list_on_stack; | |
1097 | struct bio *bio; | |
1098 | struct blk_plug plug; | |
1099 | int rw; | |
1100 | ||
1101 | bio_list_init(&bio_list_on_stack); | |
1102 | ||
0d945c1f | 1103 | spin_lock_irq(&q->queue_lock); |
c5cc2070 TH |
1104 | for (rw = READ; rw <= WRITE; rw++) |
1105 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) | |
1106 | bio_list_add(&bio_list_on_stack, bio); | |
0d945c1f | 1107 | spin_unlock_irq(&q->queue_lock); |
6e1a5704 TH |
1108 | |
1109 | if (!bio_list_empty(&bio_list_on_stack)) { | |
69d60eb9 | 1110 | blk_start_plug(&plug); |
ed00aabd | 1111 | while ((bio = bio_list_pop(&bio_list_on_stack))) |
3f98c753 | 1112 | submit_bio_noacct_nocheck(bio); |
69d60eb9 | 1113 | blk_finish_plug(&plug); |
e43473b7 | 1114 | } |
e43473b7 VG |
1115 | } |
1116 | ||
f95a04af TH |
1117 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
1118 | int off) | |
60c2bc2d | 1119 | { |
f95a04af TH |
1120 | struct throtl_grp *tg = pd_to_tg(pd); |
1121 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 1122 | |
2ab5492d | 1123 | if (v == U64_MAX) |
60c2bc2d | 1124 | return 0; |
f95a04af | 1125 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
1126 | } |
1127 | ||
f95a04af TH |
1128 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
1129 | int off) | |
e43473b7 | 1130 | { |
f95a04af TH |
1131 | struct throtl_grp *tg = pd_to_tg(pd); |
1132 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 1133 | |
2ab5492d | 1134 | if (v == UINT_MAX) |
af133ceb | 1135 | return 0; |
f95a04af | 1136 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
1137 | } |
1138 | ||
2da8ca82 | 1139 | static int tg_print_conf_u64(struct seq_file *sf, void *v) |
8e89d13f | 1140 | { |
2da8ca82 TH |
1141 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64, |
1142 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1143 | return 0; |
8e89d13f VG |
1144 | } |
1145 | ||
2da8ca82 | 1146 | static int tg_print_conf_uint(struct seq_file *sf, void *v) |
8e89d13f | 1147 | { |
2da8ca82 TH |
1148 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint, |
1149 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1150 | return 0; |
60c2bc2d TH |
1151 | } |
1152 | ||
9bb67aeb | 1153 | static void tg_conf_updated(struct throtl_grp *tg, bool global) |
60c2bc2d | 1154 | { |
69948b07 | 1155 | struct throtl_service_queue *sq = &tg->service_queue; |
492eb21b | 1156 | struct cgroup_subsys_state *pos_css; |
69948b07 | 1157 | struct blkcg_gq *blkg; |
af133ceb | 1158 | |
fda6f272 TH |
1159 | throtl_log(&tg->service_queue, |
1160 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
9f626e37 SL |
1161 | tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE), |
1162 | tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE)); | |
632b4493 | 1163 | |
27b13e20 | 1164 | rcu_read_lock(); |
693e751e TH |
1165 | /* |
1166 | * Update has_rules[] flags for the updated tg's subtree. A tg is | |
1167 | * considered to have rules if either the tg itself or any of its | |
1168 | * ancestors has rules. This identifies groups without any | |
1169 | * restrictions in the whole hierarchy and allows them to bypass | |
1170 | * blk-throttle. | |
1171 | */ | |
9bb67aeb | 1172 | blkg_for_each_descendant_pre(blkg, pos_css, |
1231039d | 1173 | global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) { |
5b81fc3c | 1174 | struct throtl_grp *this_tg = blkg_to_tg(blkg); |
5b81fc3c SL |
1175 | |
1176 | tg_update_has_rules(this_tg); | |
1177 | /* ignore root/second level */ | |
1178 | if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent || | |
1179 | !blkg->parent->parent) | |
1180 | continue; | |
5b81fc3c | 1181 | } |
27b13e20 | 1182 | rcu_read_unlock(); |
693e751e | 1183 | |
632b4493 TH |
1184 | /* |
1185 | * We're already holding queue_lock and know @tg is valid. Let's | |
1186 | * apply the new config directly. | |
1187 | * | |
1188 | * Restart the slices for both READ and WRITES. It might happen | |
1189 | * that a group's limit are dropped suddenly and we don't want to | |
1190 | * account recently dispatched IO with new low rate. | |
1191 | */ | |
a880ae93 YK |
1192 | throtl_start_new_slice(tg, READ, false); |
1193 | throtl_start_new_slice(tg, WRITE, false); | |
632b4493 | 1194 | |
5b2c16aa | 1195 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 1196 | tg_update_disptime(tg); |
7f52f98c | 1197 | throtl_schedule_next_dispatch(sq->parent_sq, true); |
632b4493 | 1198 | } |
69948b07 TH |
1199 | } |
1200 | ||
a3166c51 YK |
1201 | static int blk_throtl_init(struct gendisk *disk) |
1202 | { | |
1203 | struct request_queue *q = disk->queue; | |
1204 | struct throtl_data *td; | |
1205 | int ret; | |
1206 | ||
1207 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1208 | if (!td) | |
1209 | return -ENOMEM; | |
1210 | ||
1211 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); | |
1212 | throtl_service_queue_init(&td->service_queue); | |
1213 | ||
1214 | /* | |
1215 | * Freeze queue before activating policy, to synchronize with IO path, | |
1216 | * which is protected by 'q_usage_counter'. | |
1217 | */ | |
1218 | blk_mq_freeze_queue(disk->queue); | |
1219 | blk_mq_quiesce_queue(disk->queue); | |
1220 | ||
1221 | q->td = td; | |
1222 | td->queue = q; | |
1223 | ||
1224 | /* activate policy */ | |
1225 | ret = blkcg_activate_policy(disk, &blkcg_policy_throtl); | |
1226 | if (ret) { | |
1227 | q->td = NULL; | |
1228 | kfree(td); | |
1229 | goto out; | |
1230 | } | |
1231 | ||
1232 | if (blk_queue_nonrot(q)) | |
1233 | td->throtl_slice = DFL_THROTL_SLICE_SSD; | |
1234 | else | |
1235 | td->throtl_slice = DFL_THROTL_SLICE_HD; | |
1236 | td->track_bio_latency = !queue_is_mq(q); | |
1237 | if (!td->track_bio_latency) | |
1238 | blk_stat_enable_accounting(q); | |
1239 | ||
1240 | out: | |
1241 | blk_mq_unquiesce_queue(disk->queue); | |
1242 | blk_mq_unfreeze_queue(disk->queue); | |
1243 | ||
1244 | return ret; | |
1245 | } | |
1246 | ||
1247 | ||
69948b07 TH |
1248 | static ssize_t tg_set_conf(struct kernfs_open_file *of, |
1249 | char *buf, size_t nbytes, loff_t off, bool is_u64) | |
1250 | { | |
1251 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1252 | struct blkg_conf_ctx ctx; | |
1253 | struct throtl_grp *tg; | |
1254 | int ret; | |
1255 | u64 v; | |
1256 | ||
faffaab2 TH |
1257 | blkg_conf_init(&ctx, buf); |
1258 | ||
a3166c51 YK |
1259 | ret = blkg_conf_open_bdev(&ctx); |
1260 | if (ret) | |
1261 | goto out_finish; | |
1262 | ||
1263 | if (!blk_throtl_activated(ctx.bdev->bd_queue)) { | |
1264 | ret = blk_throtl_init(ctx.bdev->bd_disk); | |
1265 | if (ret) | |
1266 | goto out_finish; | |
1267 | } | |
1268 | ||
faffaab2 | 1269 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx); |
69948b07 | 1270 | if (ret) |
faffaab2 | 1271 | goto out_finish; |
69948b07 TH |
1272 | |
1273 | ret = -EINVAL; | |
1274 | if (sscanf(ctx.body, "%llu", &v) != 1) | |
1275 | goto out_finish; | |
1276 | if (!v) | |
2ab5492d | 1277 | v = U64_MAX; |
69948b07 TH |
1278 | |
1279 | tg = blkg_to_tg(ctx.blkg); | |
a880ae93 | 1280 | tg_update_carryover(tg); |
69948b07 TH |
1281 | |
1282 | if (is_u64) | |
1283 | *(u64 *)((void *)tg + of_cft(of)->private) = v; | |
1284 | else | |
1285 | *(unsigned int *)((void *)tg + of_cft(of)->private) = v; | |
60c2bc2d | 1286 | |
9bb67aeb | 1287 | tg_conf_updated(tg, false); |
36aa9e5f TH |
1288 | ret = 0; |
1289 | out_finish: | |
faffaab2 | 1290 | blkg_conf_exit(&ctx); |
36aa9e5f | 1291 | return ret ?: nbytes; |
8e89d13f VG |
1292 | } |
1293 | ||
451af504 TH |
1294 | static ssize_t tg_set_conf_u64(struct kernfs_open_file *of, |
1295 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1296 | { |
451af504 | 1297 | return tg_set_conf(of, buf, nbytes, off, true); |
60c2bc2d TH |
1298 | } |
1299 | ||
451af504 TH |
1300 | static ssize_t tg_set_conf_uint(struct kernfs_open_file *of, |
1301 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1302 | { |
451af504 | 1303 | return tg_set_conf(of, buf, nbytes, off, false); |
60c2bc2d TH |
1304 | } |
1305 | ||
7ca46438 TH |
1306 | static int tg_print_rwstat(struct seq_file *sf, void *v) |
1307 | { | |
1308 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
1309 | blkg_prfill_rwstat, &blkcg_policy_throtl, | |
1310 | seq_cft(sf)->private, true); | |
1311 | return 0; | |
1312 | } | |
1313 | ||
1314 | static u64 tg_prfill_rwstat_recursive(struct seq_file *sf, | |
1315 | struct blkg_policy_data *pd, int off) | |
1316 | { | |
1317 | struct blkg_rwstat_sample sum; | |
1318 | ||
1319 | blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off, | |
1320 | &sum); | |
1321 | return __blkg_prfill_rwstat(sf, pd, &sum); | |
1322 | } | |
1323 | ||
1324 | static int tg_print_rwstat_recursive(struct seq_file *sf, void *v) | |
1325 | { | |
1326 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
1327 | tg_prfill_rwstat_recursive, &blkcg_policy_throtl, | |
1328 | seq_cft(sf)->private, true); | |
1329 | return 0; | |
1330 | } | |
1331 | ||
880f50e2 | 1332 | static struct cftype throtl_legacy_files[] = { |
60c2bc2d TH |
1333 | { |
1334 | .name = "throttle.read_bps_device", | |
bf20ab53 | 1335 | .private = offsetof(struct throtl_grp, bps[READ]), |
2da8ca82 | 1336 | .seq_show = tg_print_conf_u64, |
451af504 | 1337 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1338 | }, |
1339 | { | |
1340 | .name = "throttle.write_bps_device", | |
bf20ab53 | 1341 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
2da8ca82 | 1342 | .seq_show = tg_print_conf_u64, |
451af504 | 1343 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1344 | }, |
1345 | { | |
1346 | .name = "throttle.read_iops_device", | |
bf20ab53 | 1347 | .private = offsetof(struct throtl_grp, iops[READ]), |
2da8ca82 | 1348 | .seq_show = tg_print_conf_uint, |
451af504 | 1349 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1350 | }, |
1351 | { | |
1352 | .name = "throttle.write_iops_device", | |
bf20ab53 | 1353 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
2da8ca82 | 1354 | .seq_show = tg_print_conf_uint, |
451af504 | 1355 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1356 | }, |
1357 | { | |
1358 | .name = "throttle.io_service_bytes", | |
7ca46438 TH |
1359 | .private = offsetof(struct throtl_grp, stat_bytes), |
1360 | .seq_show = tg_print_rwstat, | |
60c2bc2d | 1361 | }, |
17534c6f | 1362 | { |
1363 | .name = "throttle.io_service_bytes_recursive", | |
7ca46438 TH |
1364 | .private = offsetof(struct throtl_grp, stat_bytes), |
1365 | .seq_show = tg_print_rwstat_recursive, | |
17534c6f | 1366 | }, |
60c2bc2d TH |
1367 | { |
1368 | .name = "throttle.io_serviced", | |
7ca46438 TH |
1369 | .private = offsetof(struct throtl_grp, stat_ios), |
1370 | .seq_show = tg_print_rwstat, | |
60c2bc2d | 1371 | }, |
17534c6f | 1372 | { |
1373 | .name = "throttle.io_serviced_recursive", | |
7ca46438 TH |
1374 | .private = offsetof(struct throtl_grp, stat_ios), |
1375 | .seq_show = tg_print_rwstat_recursive, | |
17534c6f | 1376 | }, |
60c2bc2d TH |
1377 | { } /* terminate */ |
1378 | }; | |
1379 | ||
cd5ab1b0 | 1380 | static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd, |
2ee867dc TH |
1381 | int off) |
1382 | { | |
1383 | struct throtl_grp *tg = pd_to_tg(pd); | |
1384 | const char *dname = blkg_dev_name(pd->blkg); | |
cd5ab1b0 SL |
1385 | u64 bps_dft; |
1386 | unsigned int iops_dft; | |
2ee867dc TH |
1387 | |
1388 | if (!dname) | |
1389 | return 0; | |
9f626e37 | 1390 | |
bf20ab53 YK |
1391 | bps_dft = U64_MAX; |
1392 | iops_dft = UINT_MAX; | |
cd5ab1b0 | 1393 | |
0a751df4 WL |
1394 | if (tg->bps[READ] == bps_dft && |
1395 | tg->bps[WRITE] == bps_dft && | |
1396 | tg->iops[READ] == iops_dft && | |
1397 | tg->iops[WRITE] == iops_dft) | |
2ee867dc TH |
1398 | return 0; |
1399 | ||
d3a3a086 | 1400 | seq_printf(sf, "%s", dname); |
0a751df4 | 1401 | if (tg->bps[READ] == U64_MAX) |
d3a3a086 JG |
1402 | seq_printf(sf, " rbps=max"); |
1403 | else | |
0a751df4 | 1404 | seq_printf(sf, " rbps=%llu", tg->bps[READ]); |
d3a3a086 | 1405 | |
0a751df4 | 1406 | if (tg->bps[WRITE] == U64_MAX) |
d3a3a086 JG |
1407 | seq_printf(sf, " wbps=max"); |
1408 | else | |
0a751df4 | 1409 | seq_printf(sf, " wbps=%llu", tg->bps[WRITE]); |
d3a3a086 | 1410 | |
0a751df4 | 1411 | if (tg->iops[READ] == UINT_MAX) |
d3a3a086 JG |
1412 | seq_printf(sf, " riops=max"); |
1413 | else | |
0a751df4 | 1414 | seq_printf(sf, " riops=%u", tg->iops[READ]); |
d3a3a086 | 1415 | |
0a751df4 | 1416 | if (tg->iops[WRITE] == UINT_MAX) |
d3a3a086 JG |
1417 | seq_printf(sf, " wiops=max"); |
1418 | else | |
0a751df4 | 1419 | seq_printf(sf, " wiops=%u", tg->iops[WRITE]); |
d3a3a086 | 1420 | |
d3a3a086 | 1421 | seq_printf(sf, "\n"); |
2ee867dc TH |
1422 | return 0; |
1423 | } | |
1424 | ||
cd5ab1b0 | 1425 | static int tg_print_limit(struct seq_file *sf, void *v) |
2ee867dc | 1426 | { |
cd5ab1b0 | 1427 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit, |
2ee867dc TH |
1428 | &blkcg_policy_throtl, seq_cft(sf)->private, false); |
1429 | return 0; | |
1430 | } | |
1431 | ||
cd5ab1b0 | 1432 | static ssize_t tg_set_limit(struct kernfs_open_file *of, |
2ee867dc TH |
1433 | char *buf, size_t nbytes, loff_t off) |
1434 | { | |
1435 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1436 | struct blkg_conf_ctx ctx; | |
1437 | struct throtl_grp *tg; | |
1438 | u64 v[4]; | |
1439 | int ret; | |
1440 | ||
faffaab2 TH |
1441 | blkg_conf_init(&ctx, buf); |
1442 | ||
a3166c51 YK |
1443 | ret = blkg_conf_open_bdev(&ctx); |
1444 | if (ret) | |
1445 | goto out_finish; | |
1446 | ||
1447 | if (!blk_throtl_activated(ctx.bdev->bd_queue)) { | |
1448 | ret = blk_throtl_init(ctx.bdev->bd_disk); | |
1449 | if (ret) | |
1450 | goto out_finish; | |
1451 | } | |
1452 | ||
faffaab2 | 1453 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx); |
2ee867dc | 1454 | if (ret) |
faffaab2 | 1455 | goto out_finish; |
2ee867dc TH |
1456 | |
1457 | tg = blkg_to_tg(ctx.blkg); | |
a880ae93 | 1458 | tg_update_carryover(tg); |
2ee867dc | 1459 | |
bf20ab53 YK |
1460 | v[0] = tg->bps[READ]; |
1461 | v[1] = tg->bps[WRITE]; | |
1462 | v[2] = tg->iops[READ]; | |
1463 | v[3] = tg->iops[WRITE]; | |
2ee867dc TH |
1464 | |
1465 | while (true) { | |
1466 | char tok[27]; /* wiops=18446744073709551616 */ | |
1467 | char *p; | |
2ab5492d | 1468 | u64 val = U64_MAX; |
2ee867dc TH |
1469 | int len; |
1470 | ||
1471 | if (sscanf(ctx.body, "%26s%n", tok, &len) != 1) | |
1472 | break; | |
1473 | if (tok[0] == '\0') | |
1474 | break; | |
1475 | ctx.body += len; | |
1476 | ||
1477 | ret = -EINVAL; | |
1478 | p = tok; | |
1479 | strsep(&p, "="); | |
1480 | if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max"))) | |
1481 | goto out_finish; | |
1482 | ||
1483 | ret = -ERANGE; | |
1484 | if (!val) | |
1485 | goto out_finish; | |
1486 | ||
1487 | ret = -EINVAL; | |
5b7048b8 | 1488 | if (!strcmp(tok, "rbps") && val > 1) |
2ee867dc | 1489 | v[0] = val; |
5b7048b8 | 1490 | else if (!strcmp(tok, "wbps") && val > 1) |
2ee867dc | 1491 | v[1] = val; |
5b7048b8 | 1492 | else if (!strcmp(tok, "riops") && val > 1) |
2ee867dc | 1493 | v[2] = min_t(u64, val, UINT_MAX); |
5b7048b8 | 1494 | else if (!strcmp(tok, "wiops") && val > 1) |
2ee867dc TH |
1495 | v[3] = min_t(u64, val, UINT_MAX); |
1496 | else | |
1497 | goto out_finish; | |
1498 | } | |
1499 | ||
bf20ab53 YK |
1500 | tg->bps[READ] = v[0]; |
1501 | tg->bps[WRITE] = v[1]; | |
1502 | tg->iops[READ] = v[2]; | |
1503 | tg->iops[WRITE] = v[3]; | |
2ee867dc | 1504 | |
bf20ab53 | 1505 | tg_conf_updated(tg, false); |
2ee867dc TH |
1506 | ret = 0; |
1507 | out_finish: | |
faffaab2 | 1508 | blkg_conf_exit(&ctx); |
2ee867dc TH |
1509 | return ret ?: nbytes; |
1510 | } | |
1511 | ||
1512 | static struct cftype throtl_files[] = { | |
1513 | { | |
1514 | .name = "max", | |
1515 | .flags = CFTYPE_NOT_ON_ROOT, | |
cd5ab1b0 SL |
1516 | .seq_show = tg_print_limit, |
1517 | .write = tg_set_limit, | |
2ee867dc TH |
1518 | }, |
1519 | { } /* terminate */ | |
1520 | }; | |
1521 | ||
da527770 | 1522 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1523 | { |
1524 | struct throtl_data *td = q->td; | |
1525 | ||
69df0ab0 | 1526 | cancel_work_sync(&td->dispatch_work); |
e43473b7 VG |
1527 | } |
1528 | ||
a7b36ee6 | 1529 | struct blkcg_policy blkcg_policy_throtl = { |
2ee867dc | 1530 | .dfl_cftypes = throtl_files, |
880f50e2 | 1531 | .legacy_cftypes = throtl_legacy_files, |
f9fcc2d3 | 1532 | |
001bea73 | 1533 | .pd_alloc_fn = throtl_pd_alloc, |
f9fcc2d3 | 1534 | .pd_init_fn = throtl_pd_init, |
693e751e | 1535 | .pd_online_fn = throtl_pd_online, |
001bea73 | 1536 | .pd_free_fn = throtl_pd_free, |
e43473b7 VG |
1537 | }; |
1538 | ||
cad9266a | 1539 | void blk_throtl_cancel_bios(struct gendisk *disk) |
2d8f7a3b | 1540 | { |
cad9266a | 1541 | struct request_queue *q = disk->queue; |
2d8f7a3b YK |
1542 | struct cgroup_subsys_state *pos_css; |
1543 | struct blkcg_gq *blkg; | |
1544 | ||
a3166c51 YK |
1545 | if (!blk_throtl_activated(q)) |
1546 | return; | |
1547 | ||
2d8f7a3b YK |
1548 | spin_lock_irq(&q->queue_lock); |
1549 | /* | |
1550 | * queue_lock is held, rcu lock is not needed here technically. | |
1551 | * However, rcu lock is still held to emphasize that following | |
1552 | * path need RCU protection and to prevent warning from lockdep. | |
1553 | */ | |
1554 | rcu_read_lock(); | |
1231039d | 1555 | blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) { |
2d8f7a3b YK |
1556 | struct throtl_grp *tg = blkg_to_tg(blkg); |
1557 | struct throtl_service_queue *sq = &tg->service_queue; | |
1558 | ||
1559 | /* | |
1560 | * Set the flag to make sure throtl_pending_timer_fn() won't | |
1561 | * stop until all throttled bios are dispatched. | |
1562 | */ | |
eb184791 KS |
1563 | tg->flags |= THROTL_TG_CANCELING; |
1564 | ||
1565 | /* | |
1566 | * Do not dispatch cgroup without THROTL_TG_PENDING or cgroup | |
1567 | * will be inserted to service queue without THROTL_TG_PENDING | |
1568 | * set in tg_update_disptime below. Then IO dispatched from | |
1569 | * child in tg_dispatch_one_bio will trigger double insertion | |
1570 | * and corrupt the tree. | |
1571 | */ | |
1572 | if (!(tg->flags & THROTL_TG_PENDING)) | |
1573 | continue; | |
1574 | ||
2d8f7a3b YK |
1575 | /* |
1576 | * Update disptime after setting the above flag to make sure | |
1577 | * throtl_select_dispatch() won't exit without dispatching. | |
1578 | */ | |
1579 | tg_update_disptime(tg); | |
1580 | ||
1581 | throtl_schedule_pending_timer(sq, jiffies + 1); | |
1582 | } | |
1583 | rcu_read_unlock(); | |
1584 | spin_unlock_irq(&q->queue_lock); | |
1585 | } | |
1586 | ||
29390bb5 YK |
1587 | static bool tg_within_limit(struct throtl_grp *tg, struct bio *bio, bool rw) |
1588 | { | |
1589 | /* throtl is FIFO - if bios are already queued, should queue */ | |
1590 | if (tg->service_queue.nr_queued[rw]) | |
1591 | return false; | |
1592 | ||
1593 | return tg_may_dispatch(tg, bio, NULL); | |
1594 | } | |
1595 | ||
1596 | static void tg_dispatch_in_debt(struct throtl_grp *tg, struct bio *bio, bool rw) | |
1597 | { | |
1598 | if (!bio_flagged(bio, BIO_BPS_THROTTLED)) | |
1599 | tg->carryover_bytes[rw] -= throtl_bio_data_size(bio); | |
1600 | tg->carryover_ios[rw]--; | |
1601 | } | |
1602 | ||
a7b36ee6 | 1603 | bool __blk_throtl_bio(struct bio *bio) |
e43473b7 | 1604 | { |
ed6cddef | 1605 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
db18a53e | 1606 | struct blkcg_gq *blkg = bio->bi_blkg; |
c5cc2070 | 1607 | struct throtl_qnode *qn = NULL; |
a2e83ef9 | 1608 | struct throtl_grp *tg = blkg_to_tg(blkg); |
73f0d49a | 1609 | struct throtl_service_queue *sq; |
0e9f4164 | 1610 | bool rw = bio_data_dir(bio); |
bc16a4f9 | 1611 | bool throttled = false; |
b9147dd1 | 1612 | struct throtl_data *td = tg->td; |
e43473b7 | 1613 | |
93b80638 | 1614 | rcu_read_lock(); |
0d945c1f | 1615 | spin_lock_irq(&q->queue_lock); |
73f0d49a TH |
1616 | sq = &tg->service_queue; |
1617 | ||
9e660acf | 1618 | while (true) { |
29390bb5 YK |
1619 | if (tg_within_limit(tg, bio, rw)) { |
1620 | /* within limits, let's charge and dispatch directly */ | |
1621 | throtl_charge_bio(tg, bio); | |
1622 | ||
1623 | /* | |
1624 | * We need to trim slice even when bios are not being | |
1625 | * queued otherwise it might happen that a bio is not | |
1626 | * queued for a long time and slice keeps on extending | |
1627 | * and trim is not called for a long time. Now if limits | |
1628 | * are reduced suddenly we take into account all the IO | |
1629 | * dispatched so far at new low rate and * newly queued | |
1630 | * IO gets a really long dispatch time. | |
1631 | * | |
1632 | * So keep on trimming slice even if bio is not queued. | |
1633 | */ | |
1634 | throtl_trim_slice(tg, rw); | |
1635 | } else if (bio_issue_as_root_blkg(bio)) { | |
1636 | /* | |
1637 | * IOs which may cause priority inversions are | |
1638 | * dispatched directly, even if they're over limit. | |
1639 | * Debts are handled by carryover_bytes/ios while | |
1640 | * calculating wait time. | |
1641 | */ | |
1642 | tg_dispatch_in_debt(tg, bio, rw); | |
1643 | } else { | |
1644 | /* if above limits, break to queue */ | |
9e660acf | 1645 | break; |
c79892c5 | 1646 | } |
9e660acf | 1647 | |
9e660acf TH |
1648 | /* |
1649 | * @bio passed through this layer without being throttled. | |
b53b072c | 1650 | * Climb up the ladder. If we're already at the top, it |
9e660acf TH |
1651 | * can be executed directly. |
1652 | */ | |
c5cc2070 | 1653 | qn = &tg->qnode_on_parent[rw]; |
9e660acf TH |
1654 | sq = sq->parent_sq; |
1655 | tg = sq_to_tg(sq); | |
320fb0f9 YK |
1656 | if (!tg) { |
1657 | bio_set_flag(bio, BIO_BPS_THROTTLED); | |
9e660acf | 1658 | goto out_unlock; |
320fb0f9 | 1659 | } |
e43473b7 VG |
1660 | } |
1661 | ||
9e660acf | 1662 | /* out-of-limit, queue to @tg */ |
fda6f272 TH |
1663 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
1664 | rw == READ ? 'R' : 'W', | |
9f626e37 SL |
1665 | tg->bytes_disp[rw], bio->bi_iter.bi_size, |
1666 | tg_bps_limit(tg, rw), | |
1667 | tg->io_disp[rw], tg_iops_limit(tg, rw), | |
fda6f272 | 1668 | sq->nr_queued[READ], sq->nr_queued[WRITE]); |
e43473b7 | 1669 | |
b9147dd1 | 1670 | td->nr_queued[rw]++; |
c5cc2070 | 1671 | throtl_add_bio_tg(bio, qn, tg); |
bc16a4f9 | 1672 | throttled = true; |
e43473b7 | 1673 | |
7f52f98c TH |
1674 | /* |
1675 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
1676 | * was empty before @bio. The forced scheduling isn't likely to | |
1677 | * cause undue delay as @bio is likely to be dispatched directly if | |
1678 | * its @tg's disptime is not in the future. | |
1679 | */ | |
0e9f4164 | 1680 | if (tg->flags & THROTL_TG_WAS_EMPTY) { |
77216b04 | 1681 | tg_update_disptime(tg); |
7f52f98c | 1682 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); |
e43473b7 VG |
1683 | } |
1684 | ||
bc16a4f9 | 1685 | out_unlock: |
5a011f88 LQ |
1686 | spin_unlock_irq(&q->queue_lock); |
1687 | ||
93b80638 | 1688 | rcu_read_unlock(); |
bc16a4f9 | 1689 | return throttled; |
e43473b7 VG |
1690 | } |
1691 | ||
e13793ba | 1692 | void blk_throtl_exit(struct gendisk *disk) |
e43473b7 | 1693 | { |
e13793ba CH |
1694 | struct request_queue *q = disk->queue; |
1695 | ||
a3166c51 YK |
1696 | if (!blk_throtl_activated(q)) |
1697 | return; | |
1698 | ||
884f0e84 | 1699 | del_timer_sync(&q->td->service_queue.pending_timer); |
da527770 | 1700 | throtl_shutdown_wq(q); |
40e4996e | 1701 | blkcg_deactivate_policy(disk, &blkcg_policy_throtl); |
c9a929dd | 1702 | kfree(q->td); |
e43473b7 VG |
1703 | } |
1704 | ||
1705 | static int __init throtl_init(void) | |
1706 | { | |
450adcbe VG |
1707 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1708 | if (!kthrotld_workqueue) | |
1709 | panic("Failed to create kthrotld\n"); | |
1710 | ||
3c798398 | 1711 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1712 | } |
1713 | ||
1714 | module_init(throtl_init); |