]>
Commit | Line | Data |
---|---|---|
e43473b7 VG |
1 | /* |
2 | * Interface for controlling IO bandwidth on a request queue | |
3 | * | |
4 | * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com> | |
5 | */ | |
6 | ||
7 | #include <linux/module.h> | |
8 | #include <linux/slab.h> | |
9 | #include <linux/blkdev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blktrace_api.h> | |
12 | #include "blk-cgroup.h" | |
bc9fcbf9 | 13 | #include "blk.h" |
e43473b7 VG |
14 | |
15 | /* Max dispatch from a group in 1 round */ | |
16 | static int throtl_grp_quantum = 8; | |
17 | ||
18 | /* Total max dispatch from all groups in one round */ | |
19 | static int throtl_quantum = 32; | |
20 | ||
21 | /* Throttling is performed over 100ms slice and after that slice is renewed */ | |
22 | static unsigned long throtl_slice = HZ/10; /* 100 ms */ | |
23 | ||
3c798398 | 24 | static struct blkcg_policy blkcg_policy_throtl; |
0381411e | 25 | |
450adcbe VG |
26 | /* A workqueue to queue throttle related work */ |
27 | static struct workqueue_struct *kthrotld_workqueue; | |
450adcbe | 28 | |
c9e0332e | 29 | struct throtl_service_queue { |
77216b04 TH |
30 | struct throtl_service_queue *parent_sq; /* the parent service_queue */ |
31 | ||
73f0d49a TH |
32 | /* |
33 | * Bios queued directly to this service_queue or dispatched from | |
34 | * children throtl_grp's. | |
35 | */ | |
36 | struct bio_list bio_lists[2]; /* queued bios [READ/WRITE] */ | |
37 | unsigned int nr_queued[2]; /* number of queued bios */ | |
38 | ||
39 | /* | |
40 | * RB tree of active children throtl_grp's, which are sorted by | |
41 | * their ->disptime. | |
42 | */ | |
c9e0332e TH |
43 | struct rb_root pending_tree; /* RB tree of active tgs */ |
44 | struct rb_node *first_pending; /* first node in the tree */ | |
45 | unsigned int nr_pending; /* # queued in the tree */ | |
46 | unsigned long first_pending_disptime; /* disptime of the first tg */ | |
69df0ab0 | 47 | struct timer_list pending_timer; /* fires on first_pending_disptime */ |
e43473b7 VG |
48 | }; |
49 | ||
5b2c16aa TH |
50 | enum tg_state_flags { |
51 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
0e9f4164 | 52 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ |
5b2c16aa TH |
53 | }; |
54 | ||
e43473b7 VG |
55 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
56 | ||
8a3d2615 TH |
57 | /* Per-cpu group stats */ |
58 | struct tg_stats_cpu { | |
59 | /* total bytes transferred */ | |
60 | struct blkg_rwstat service_bytes; | |
61 | /* total IOs serviced, post merge */ | |
62 | struct blkg_rwstat serviced; | |
63 | }; | |
64 | ||
e43473b7 | 65 | struct throtl_grp { |
f95a04af TH |
66 | /* must be the first member */ |
67 | struct blkg_policy_data pd; | |
68 | ||
c9e0332e | 69 | /* active throtl group service_queue member */ |
e43473b7 VG |
70 | struct rb_node rb_node; |
71 | ||
0f3457f6 TH |
72 | /* throtl_data this group belongs to */ |
73 | struct throtl_data *td; | |
74 | ||
49a2f1e3 TH |
75 | /* this group's service queue */ |
76 | struct throtl_service_queue service_queue; | |
77 | ||
e43473b7 VG |
78 | /* |
79 | * Dispatch time in jiffies. This is the estimated time when group | |
80 | * will unthrottle and is ready to dispatch more bio. It is used as | |
81 | * key to sort active groups in service tree. | |
82 | */ | |
83 | unsigned long disptime; | |
84 | ||
e43473b7 VG |
85 | unsigned int flags; |
86 | ||
e43473b7 VG |
87 | /* bytes per second rate limits */ |
88 | uint64_t bps[2]; | |
89 | ||
8e89d13f VG |
90 | /* IOPS limits */ |
91 | unsigned int iops[2]; | |
92 | ||
e43473b7 VG |
93 | /* Number of bytes disptached in current slice */ |
94 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
95 | /* Number of bio's dispatched in current slice */ |
96 | unsigned int io_disp[2]; | |
e43473b7 VG |
97 | |
98 | /* When did we start a new slice */ | |
99 | unsigned long slice_start[2]; | |
100 | unsigned long slice_end[2]; | |
fe071437 | 101 | |
8a3d2615 TH |
102 | /* Per cpu stats pointer */ |
103 | struct tg_stats_cpu __percpu *stats_cpu; | |
104 | ||
105 | /* List of tgs waiting for per cpu stats memory to be allocated */ | |
106 | struct list_head stats_alloc_node; | |
e43473b7 VG |
107 | }; |
108 | ||
109 | struct throtl_data | |
110 | { | |
e43473b7 | 111 | /* service tree for active throtl groups */ |
c9e0332e | 112 | struct throtl_service_queue service_queue; |
e43473b7 | 113 | |
e43473b7 VG |
114 | struct request_queue *queue; |
115 | ||
116 | /* Total Number of queued bios on READ and WRITE lists */ | |
117 | unsigned int nr_queued[2]; | |
118 | ||
119 | /* | |
02977e4a | 120 | * number of total undestroyed groups |
e43473b7 VG |
121 | */ |
122 | unsigned int nr_undestroyed_grps; | |
123 | ||
124 | /* Work for dispatching throttled bios */ | |
69df0ab0 | 125 | struct work_struct dispatch_work; |
e43473b7 VG |
126 | }; |
127 | ||
8a3d2615 TH |
128 | /* list and work item to allocate percpu group stats */ |
129 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
130 | static LIST_HEAD(tg_stats_alloc_list); | |
131 | ||
132 | static void tg_stats_alloc_fn(struct work_struct *); | |
133 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
134 | ||
69df0ab0 TH |
135 | static void throtl_pending_timer_fn(unsigned long arg); |
136 | ||
f95a04af TH |
137 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
138 | { | |
139 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
140 | } | |
141 | ||
3c798398 | 142 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 143 | { |
f95a04af | 144 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
145 | } |
146 | ||
3c798398 | 147 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 148 | { |
f95a04af | 149 | return pd_to_blkg(&tg->pd); |
0381411e TH |
150 | } |
151 | ||
03d8e111 TH |
152 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) |
153 | { | |
154 | return blkg_to_tg(td->queue->root_blkg); | |
155 | } | |
156 | ||
fda6f272 TH |
157 | /** |
158 | * sq_to_tg - return the throl_grp the specified service queue belongs to | |
159 | * @sq: the throtl_service_queue of interest | |
160 | * | |
161 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one | |
162 | * embedded in throtl_data, %NULL is returned. | |
163 | */ | |
164 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) | |
165 | { | |
166 | if (sq && sq->parent_sq) | |
167 | return container_of(sq, struct throtl_grp, service_queue); | |
168 | else | |
169 | return NULL; | |
170 | } | |
171 | ||
172 | /** | |
173 | * sq_to_td - return throtl_data the specified service queue belongs to | |
174 | * @sq: the throtl_service_queue of interest | |
175 | * | |
176 | * A service_queue can be embeded in either a throtl_grp or throtl_data. | |
177 | * Determine the associated throtl_data accordingly and return it. | |
178 | */ | |
179 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) | |
180 | { | |
181 | struct throtl_grp *tg = sq_to_tg(sq); | |
182 | ||
183 | if (tg) | |
184 | return tg->td; | |
185 | else | |
186 | return container_of(sq, struct throtl_data, service_queue); | |
187 | } | |
188 | ||
189 | /** | |
190 | * throtl_log - log debug message via blktrace | |
191 | * @sq: the service_queue being reported | |
192 | * @fmt: printf format string | |
193 | * @args: printf args | |
194 | * | |
195 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a | |
196 | * throtl_grp; otherwise, just "throtl". | |
197 | * | |
198 | * TODO: this should be made a function and name formatting should happen | |
199 | * after testing whether blktrace is enabled. | |
200 | */ | |
201 | #define throtl_log(sq, fmt, args...) do { \ | |
202 | struct throtl_grp *__tg = sq_to_tg((sq)); \ | |
203 | struct throtl_data *__td = sq_to_td((sq)); \ | |
204 | \ | |
205 | (void)__td; \ | |
206 | if ((__tg)) { \ | |
207 | char __pbuf[128]; \ | |
54e7ed12 | 208 | \ |
fda6f272 TH |
209 | blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \ |
210 | blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
211 | } else { \ | |
212 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
213 | } \ | |
54e7ed12 | 214 | } while (0) |
e43473b7 | 215 | |
8a3d2615 TH |
216 | /* |
217 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
3b07e9ca | 218 | * system_wq once there are some groups on the alloc_list waiting for |
8a3d2615 TH |
219 | * allocation. |
220 | */ | |
221 | static void tg_stats_alloc_fn(struct work_struct *work) | |
222 | { | |
223 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
224 | struct delayed_work *dwork = to_delayed_work(work); | |
225 | bool empty = false; | |
226 | ||
227 | alloc_stats: | |
228 | if (!stats_cpu) { | |
229 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
230 | if (!stats_cpu) { | |
231 | /* allocation failed, try again after some time */ | |
3b07e9ca | 232 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); |
8a3d2615 TH |
233 | return; |
234 | } | |
235 | } | |
236 | ||
237 | spin_lock_irq(&tg_stats_alloc_lock); | |
238 | ||
239 | if (!list_empty(&tg_stats_alloc_list)) { | |
240 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
241 | struct throtl_grp, | |
242 | stats_alloc_node); | |
243 | swap(tg->stats_cpu, stats_cpu); | |
244 | list_del_init(&tg->stats_alloc_node); | |
245 | } | |
246 | ||
247 | empty = list_empty(&tg_stats_alloc_list); | |
248 | spin_unlock_irq(&tg_stats_alloc_lock); | |
249 | if (!empty) | |
250 | goto alloc_stats; | |
251 | } | |
252 | ||
49a2f1e3 | 253 | /* init a service_queue, assumes the caller zeroed it */ |
77216b04 TH |
254 | static void throtl_service_queue_init(struct throtl_service_queue *sq, |
255 | struct throtl_service_queue *parent_sq) | |
49a2f1e3 | 256 | { |
73f0d49a TH |
257 | bio_list_init(&sq->bio_lists[0]); |
258 | bio_list_init(&sq->bio_lists[1]); | |
49a2f1e3 | 259 | sq->pending_tree = RB_ROOT; |
77216b04 | 260 | sq->parent_sq = parent_sq; |
69df0ab0 TH |
261 | setup_timer(&sq->pending_timer, throtl_pending_timer_fn, |
262 | (unsigned long)sq); | |
263 | } | |
264 | ||
265 | static void throtl_service_queue_exit(struct throtl_service_queue *sq) | |
266 | { | |
267 | del_timer_sync(&sq->pending_timer); | |
49a2f1e3 TH |
268 | } |
269 | ||
3c798398 | 270 | static void throtl_pd_init(struct blkcg_gq *blkg) |
a29a171e | 271 | { |
0381411e | 272 | struct throtl_grp *tg = blkg_to_tg(blkg); |
77216b04 | 273 | struct throtl_data *td = blkg->q->td; |
ff26eaad | 274 | unsigned long flags; |
cd1604fa | 275 | |
77216b04 | 276 | throtl_service_queue_init(&tg->service_queue, &td->service_queue); |
a29a171e | 277 | RB_CLEAR_NODE(&tg->rb_node); |
77216b04 | 278 | tg->td = td; |
a29a171e | 279 | |
e56da7e2 TH |
280 | tg->bps[READ] = -1; |
281 | tg->bps[WRITE] = -1; | |
282 | tg->iops[READ] = -1; | |
283 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
284 | |
285 | /* | |
286 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
287 | * but percpu allocator can't be called from IO path. Queue tg on | |
288 | * tg_stats_alloc_list and allocate from work item. | |
289 | */ | |
ff26eaad | 290 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 291 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); |
3b07e9ca | 292 | schedule_delayed_work(&tg_stats_alloc_work, 0); |
ff26eaad | 293 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
294 | } |
295 | ||
3c798398 | 296 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
8a3d2615 TH |
297 | { |
298 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
ff26eaad | 299 | unsigned long flags; |
8a3d2615 | 300 | |
ff26eaad | 301 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 302 | list_del_init(&tg->stats_alloc_node); |
ff26eaad | 303 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
304 | |
305 | free_percpu(tg->stats_cpu); | |
69df0ab0 TH |
306 | |
307 | throtl_service_queue_exit(&tg->service_queue); | |
8a3d2615 TH |
308 | } |
309 | ||
3c798398 | 310 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
8a3d2615 TH |
311 | { |
312 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
313 | int cpu; | |
314 | ||
315 | if (tg->stats_cpu == NULL) | |
316 | return; | |
317 | ||
318 | for_each_possible_cpu(cpu) { | |
319 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
320 | ||
321 | blkg_rwstat_reset(&sc->service_bytes); | |
322 | blkg_rwstat_reset(&sc->serviced); | |
323 | } | |
a29a171e VG |
324 | } |
325 | ||
3c798398 TH |
326 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
327 | struct blkcg *blkcg) | |
e43473b7 | 328 | { |
be2c6b19 | 329 | /* |
3c798398 TH |
330 | * This is the common case when there are no blkcgs. Avoid lookup |
331 | * in this case | |
cd1604fa | 332 | */ |
3c798398 | 333 | if (blkcg == &blkcg_root) |
03d8e111 | 334 | return td_root_tg(td); |
e43473b7 | 335 | |
e8989fae | 336 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
337 | } |
338 | ||
cd1604fa | 339 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
3c798398 | 340 | struct blkcg *blkcg) |
e43473b7 | 341 | { |
f469a7b4 | 342 | struct request_queue *q = td->queue; |
cd1604fa | 343 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 344 | |
f469a7b4 | 345 | /* |
3c798398 TH |
346 | * This is the common case when there are no blkcgs. Avoid lookup |
347 | * in this case | |
f469a7b4 | 348 | */ |
3c798398 | 349 | if (blkcg == &blkcg_root) { |
03d8e111 | 350 | tg = td_root_tg(td); |
cd1604fa | 351 | } else { |
3c798398 | 352 | struct blkcg_gq *blkg; |
f469a7b4 | 353 | |
3c96cb32 | 354 | blkg = blkg_lookup_create(blkcg, q); |
f469a7b4 | 355 | |
cd1604fa TH |
356 | /* if %NULL and @q is alive, fall back to root_tg */ |
357 | if (!IS_ERR(blkg)) | |
0381411e | 358 | tg = blkg_to_tg(blkg); |
3f3299d5 | 359 | else if (!blk_queue_dying(q)) |
03d8e111 | 360 | tg = td_root_tg(td); |
f469a7b4 VG |
361 | } |
362 | ||
e43473b7 VG |
363 | return tg; |
364 | } | |
365 | ||
0049af73 TH |
366 | static struct throtl_grp * |
367 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
368 | { |
369 | /* Service tree is empty */ | |
0049af73 | 370 | if (!parent_sq->nr_pending) |
e43473b7 VG |
371 | return NULL; |
372 | ||
0049af73 TH |
373 | if (!parent_sq->first_pending) |
374 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
e43473b7 | 375 | |
0049af73 TH |
376 | if (parent_sq->first_pending) |
377 | return rb_entry_tg(parent_sq->first_pending); | |
e43473b7 VG |
378 | |
379 | return NULL; | |
380 | } | |
381 | ||
382 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
383 | { | |
384 | rb_erase(n, root); | |
385 | RB_CLEAR_NODE(n); | |
386 | } | |
387 | ||
0049af73 TH |
388 | static void throtl_rb_erase(struct rb_node *n, |
389 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 390 | { |
0049af73 TH |
391 | if (parent_sq->first_pending == n) |
392 | parent_sq->first_pending = NULL; | |
393 | rb_erase_init(n, &parent_sq->pending_tree); | |
394 | --parent_sq->nr_pending; | |
e43473b7 VG |
395 | } |
396 | ||
0049af73 | 397 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
398 | { |
399 | struct throtl_grp *tg; | |
400 | ||
0049af73 | 401 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
402 | if (!tg) |
403 | return; | |
404 | ||
0049af73 | 405 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
406 | } |
407 | ||
77216b04 | 408 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 409 | { |
77216b04 | 410 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
0049af73 | 411 | struct rb_node **node = &parent_sq->pending_tree.rb_node; |
e43473b7 VG |
412 | struct rb_node *parent = NULL; |
413 | struct throtl_grp *__tg; | |
414 | unsigned long key = tg->disptime; | |
415 | int left = 1; | |
416 | ||
417 | while (*node != NULL) { | |
418 | parent = *node; | |
419 | __tg = rb_entry_tg(parent); | |
420 | ||
421 | if (time_before(key, __tg->disptime)) | |
422 | node = &parent->rb_left; | |
423 | else { | |
424 | node = &parent->rb_right; | |
425 | left = 0; | |
426 | } | |
427 | } | |
428 | ||
429 | if (left) | |
0049af73 | 430 | parent_sq->first_pending = &tg->rb_node; |
e43473b7 VG |
431 | |
432 | rb_link_node(&tg->rb_node, parent, node); | |
0049af73 | 433 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); |
e43473b7 VG |
434 | } |
435 | ||
77216b04 | 436 | static void __throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 437 | { |
77216b04 | 438 | tg_service_queue_add(tg); |
5b2c16aa | 439 | tg->flags |= THROTL_TG_PENDING; |
77216b04 | 440 | tg->service_queue.parent_sq->nr_pending++; |
e43473b7 VG |
441 | } |
442 | ||
77216b04 | 443 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 444 | { |
5b2c16aa | 445 | if (!(tg->flags & THROTL_TG_PENDING)) |
77216b04 | 446 | __throtl_enqueue_tg(tg); |
e43473b7 VG |
447 | } |
448 | ||
77216b04 | 449 | static void __throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 450 | { |
77216b04 | 451 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
5b2c16aa | 452 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
453 | } |
454 | ||
77216b04 | 455 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 456 | { |
5b2c16aa | 457 | if (tg->flags & THROTL_TG_PENDING) |
77216b04 | 458 | __throtl_dequeue_tg(tg); |
e43473b7 VG |
459 | } |
460 | ||
a9131a27 | 461 | /* Call with queue lock held */ |
69df0ab0 TH |
462 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, |
463 | unsigned long expires) | |
a9131a27 | 464 | { |
69df0ab0 TH |
465 | mod_timer(&sq->pending_timer, expires); |
466 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
467 | expires - jiffies, jiffies); | |
a9131a27 TH |
468 | } |
469 | ||
7f52f98c TH |
470 | /** |
471 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle | |
472 | * @sq: the service_queue to schedule dispatch for | |
473 | * @force: force scheduling | |
474 | * | |
475 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the | |
476 | * dispatch time of the first pending child. Returns %true if either timer | |
477 | * is armed or there's no pending child left. %false if the current | |
478 | * dispatch window is still open and the caller should continue | |
479 | * dispatching. | |
480 | * | |
481 | * If @force is %true, the dispatch timer is always scheduled and this | |
482 | * function is guaranteed to return %true. This is to be used when the | |
483 | * caller can't dispatch itself and needs to invoke pending_timer | |
484 | * unconditionally. Note that forced scheduling is likely to induce short | |
485 | * delay before dispatch starts even if @sq->first_pending_disptime is not | |
486 | * in the future and thus shouldn't be used in hot paths. | |
487 | */ | |
488 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
489 | bool force) | |
e43473b7 | 490 | { |
6a525600 | 491 | /* any pending children left? */ |
c9e0332e | 492 | if (!sq->nr_pending) |
7f52f98c | 493 | return true; |
e43473b7 | 494 | |
c9e0332e | 495 | update_min_dispatch_time(sq); |
e43473b7 | 496 | |
69df0ab0 | 497 | /* is the next dispatch time in the future? */ |
7f52f98c | 498 | if (force || time_after(sq->first_pending_disptime, jiffies)) { |
69df0ab0 | 499 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); |
7f52f98c | 500 | return true; |
69df0ab0 TH |
501 | } |
502 | ||
7f52f98c TH |
503 | /* tell the caller to continue dispatching */ |
504 | return false; | |
e43473b7 VG |
505 | } |
506 | ||
0f3457f6 | 507 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
508 | { |
509 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 510 | tg->io_disp[rw] = 0; |
e43473b7 VG |
511 | tg->slice_start[rw] = jiffies; |
512 | tg->slice_end[rw] = jiffies + throtl_slice; | |
fda6f272 TH |
513 | throtl_log(&tg->service_queue, |
514 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
515 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
516 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
517 | } |
518 | ||
0f3457f6 TH |
519 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
520 | unsigned long jiffy_end) | |
d1ae8ffd VG |
521 | { |
522 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
523 | } | |
524 | ||
0f3457f6 TH |
525 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
526 | unsigned long jiffy_end) | |
e43473b7 VG |
527 | { |
528 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
fda6f272 TH |
529 | throtl_log(&tg->service_queue, |
530 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
531 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
532 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
533 | } |
534 | ||
535 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 536 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
537 | { |
538 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
539 | return 0; | |
540 | ||
541 | return 1; | |
542 | } | |
543 | ||
544 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 545 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 546 | { |
3aad5d3e VG |
547 | unsigned long nr_slices, time_elapsed, io_trim; |
548 | u64 bytes_trim, tmp; | |
e43473b7 VG |
549 | |
550 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
551 | ||
552 | /* | |
553 | * If bps are unlimited (-1), then time slice don't get | |
554 | * renewed. Don't try to trim the slice if slice is used. A new | |
555 | * slice will start when appropriate. | |
556 | */ | |
0f3457f6 | 557 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
558 | return; |
559 | ||
d1ae8ffd VG |
560 | /* |
561 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
562 | * that initially cgroup limit was very low resulting in high | |
563 | * slice_end, but later limit was bumped up and bio was dispached | |
564 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
565 | * is bad because it does not allow new slice to start. | |
566 | */ | |
567 | ||
0f3457f6 | 568 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); |
d1ae8ffd | 569 | |
e43473b7 VG |
570 | time_elapsed = jiffies - tg->slice_start[rw]; |
571 | ||
572 | nr_slices = time_elapsed / throtl_slice; | |
573 | ||
574 | if (!nr_slices) | |
575 | return; | |
3aad5d3e VG |
576 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
577 | do_div(tmp, HZ); | |
578 | bytes_trim = tmp; | |
e43473b7 | 579 | |
8e89d13f | 580 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 581 | |
8e89d13f | 582 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
583 | return; |
584 | ||
585 | if (tg->bytes_disp[rw] >= bytes_trim) | |
586 | tg->bytes_disp[rw] -= bytes_trim; | |
587 | else | |
588 | tg->bytes_disp[rw] = 0; | |
589 | ||
8e89d13f VG |
590 | if (tg->io_disp[rw] >= io_trim) |
591 | tg->io_disp[rw] -= io_trim; | |
592 | else | |
593 | tg->io_disp[rw] = 0; | |
594 | ||
e43473b7 VG |
595 | tg->slice_start[rw] += nr_slices * throtl_slice; |
596 | ||
fda6f272 TH |
597 | throtl_log(&tg->service_queue, |
598 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
599 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
600 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
e43473b7 VG |
601 | } |
602 | ||
0f3457f6 TH |
603 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
604 | unsigned long *wait) | |
e43473b7 VG |
605 | { |
606 | bool rw = bio_data_dir(bio); | |
8e89d13f | 607 | unsigned int io_allowed; |
e43473b7 | 608 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 609 | u64 tmp; |
e43473b7 | 610 | |
8e89d13f | 611 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 612 | |
8e89d13f VG |
613 | /* Slice has just started. Consider one slice interval */ |
614 | if (!jiffy_elapsed) | |
615 | jiffy_elapsed_rnd = throtl_slice; | |
616 | ||
617 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
618 | ||
c49c06e4 VG |
619 | /* |
620 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
621 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
622 | * will allow dispatch after 1 second and after that slice should | |
623 | * have been trimmed. | |
624 | */ | |
625 | ||
626 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
627 | do_div(tmp, HZ); | |
628 | ||
629 | if (tmp > UINT_MAX) | |
630 | io_allowed = UINT_MAX; | |
631 | else | |
632 | io_allowed = tmp; | |
8e89d13f VG |
633 | |
634 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
635 | if (wait) |
636 | *wait = 0; | |
637 | return 1; | |
638 | } | |
639 | ||
8e89d13f VG |
640 | /* Calc approx time to dispatch */ |
641 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
642 | ||
643 | if (jiffy_wait > jiffy_elapsed) | |
644 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
645 | else | |
646 | jiffy_wait = 1; | |
647 | ||
648 | if (wait) | |
649 | *wait = jiffy_wait; | |
650 | return 0; | |
651 | } | |
652 | ||
0f3457f6 TH |
653 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
654 | unsigned long *wait) | |
8e89d13f VG |
655 | { |
656 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 657 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 658 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
659 | |
660 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
661 | ||
662 | /* Slice has just started. Consider one slice interval */ | |
663 | if (!jiffy_elapsed) | |
664 | jiffy_elapsed_rnd = throtl_slice; | |
665 | ||
666 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
667 | ||
5e901a2b VG |
668 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
669 | do_div(tmp, HZ); | |
3aad5d3e | 670 | bytes_allowed = tmp; |
e43473b7 VG |
671 | |
672 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
673 | if (wait) | |
674 | *wait = 0; | |
675 | return 1; | |
676 | } | |
677 | ||
678 | /* Calc approx time to dispatch */ | |
679 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
680 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
681 | ||
682 | if (!jiffy_wait) | |
683 | jiffy_wait = 1; | |
684 | ||
685 | /* | |
686 | * This wait time is without taking into consideration the rounding | |
687 | * up we did. Add that time also. | |
688 | */ | |
689 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
690 | if (wait) |
691 | *wait = jiffy_wait; | |
8e89d13f VG |
692 | return 0; |
693 | } | |
694 | ||
af75cd3c VG |
695 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
696 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
697 | return 1; | |
698 | return 0; | |
699 | } | |
700 | ||
8e89d13f VG |
701 | /* |
702 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
703 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
704 | */ | |
0f3457f6 TH |
705 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
706 | unsigned long *wait) | |
8e89d13f VG |
707 | { |
708 | bool rw = bio_data_dir(bio); | |
709 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
710 | ||
711 | /* | |
712 | * Currently whole state machine of group depends on first bio | |
713 | * queued in the group bio list. So one should not be calling | |
714 | * this function with a different bio if there are other bios | |
715 | * queued. | |
716 | */ | |
73f0d49a TH |
717 | BUG_ON(tg->service_queue.nr_queued[rw] && |
718 | bio != bio_list_peek(&tg->service_queue.bio_lists[rw])); | |
e43473b7 | 719 | |
8e89d13f VG |
720 | /* If tg->bps = -1, then BW is unlimited */ |
721 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
722 | if (wait) | |
723 | *wait = 0; | |
724 | return 1; | |
725 | } | |
726 | ||
727 | /* | |
728 | * If previous slice expired, start a new one otherwise renew/extend | |
729 | * existing slice to make sure it is at least throtl_slice interval | |
730 | * long since now. | |
731 | */ | |
0f3457f6 TH |
732 | if (throtl_slice_used(tg, rw)) |
733 | throtl_start_new_slice(tg, rw); | |
8e89d13f VG |
734 | else { |
735 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
0f3457f6 | 736 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); |
8e89d13f VG |
737 | } |
738 | ||
0f3457f6 TH |
739 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
740 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
741 | if (wait) |
742 | *wait = 0; | |
743 | return 1; | |
744 | } | |
745 | ||
746 | max_wait = max(bps_wait, iops_wait); | |
747 | ||
748 | if (wait) | |
749 | *wait = max_wait; | |
750 | ||
751 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 752 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
753 | |
754 | return 0; | |
755 | } | |
756 | ||
3c798398 | 757 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
758 | int rw) |
759 | { | |
8a3d2615 TH |
760 | struct throtl_grp *tg = blkg_to_tg(blkg); |
761 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
762 | unsigned long flags; |
763 | ||
764 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 765 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
766 | return; |
767 | ||
768 | /* | |
769 | * Disabling interrupts to provide mutual exclusion between two | |
770 | * writes on same cpu. It probably is not needed for 64bit. Not | |
771 | * optimizing that case yet. | |
772 | */ | |
773 | local_irq_save(flags); | |
774 | ||
8a3d2615 | 775 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 776 | |
629ed0b1 TH |
777 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
778 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
779 | ||
780 | local_irq_restore(flags); | |
781 | } | |
782 | ||
e43473b7 VG |
783 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
784 | { | |
785 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
786 | |
787 | /* Charge the bio to the group */ | |
788 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 789 | tg->io_disp[rw]++; |
e43473b7 | 790 | |
2a0f61e6 TH |
791 | /* |
792 | * REQ_THROTTLED is used to prevent the same bio to be throttled | |
793 | * more than once as a throttled bio will go through blk-throtl the | |
794 | * second time when it eventually gets issued. Set it when a bio | |
795 | * is being charged to a tg. | |
796 | * | |
797 | * Dispatch stats aren't recursive and each @bio should only be | |
798 | * accounted by the @tg it was originally associated with. Let's | |
799 | * update the stats when setting REQ_THROTTLED for the first time | |
800 | * which is guaranteed to be for the @bio's original tg. | |
801 | */ | |
802 | if (!(bio->bi_rw & REQ_THROTTLED)) { | |
803 | bio->bi_rw |= REQ_THROTTLED; | |
804 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, | |
805 | bio->bi_rw); | |
806 | } | |
e43473b7 VG |
807 | } |
808 | ||
77216b04 | 809 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_grp *tg) |
e43473b7 | 810 | { |
73f0d49a | 811 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
812 | bool rw = bio_data_dir(bio); |
813 | ||
0e9f4164 TH |
814 | /* |
815 | * If @tg doesn't currently have any bios queued in the same | |
816 | * direction, queueing @bio can change when @tg should be | |
817 | * dispatched. Mark that @tg was empty. This is automatically | |
818 | * cleaered on the next tg_update_disptime(). | |
819 | */ | |
820 | if (!sq->nr_queued[rw]) | |
821 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
822 | ||
73f0d49a | 823 | bio_list_add(&sq->bio_lists[rw], bio); |
e43473b7 | 824 | /* Take a bio reference on tg */ |
1adaf3dd | 825 | blkg_get(tg_to_blkg(tg)); |
73f0d49a | 826 | sq->nr_queued[rw]++; |
77216b04 | 827 | throtl_enqueue_tg(tg); |
e43473b7 VG |
828 | } |
829 | ||
77216b04 | 830 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 831 | { |
73f0d49a | 832 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
833 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
834 | struct bio *bio; | |
835 | ||
73f0d49a | 836 | if ((bio = bio_list_peek(&sq->bio_lists[READ]))) |
0f3457f6 | 837 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 838 | |
73f0d49a | 839 | if ((bio = bio_list_peek(&sq->bio_lists[WRITE]))) |
0f3457f6 | 840 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
841 | |
842 | min_wait = min(read_wait, write_wait); | |
843 | disptime = jiffies + min_wait; | |
844 | ||
e43473b7 | 845 | /* Update dispatch time */ |
77216b04 | 846 | throtl_dequeue_tg(tg); |
e43473b7 | 847 | tg->disptime = disptime; |
77216b04 | 848 | throtl_enqueue_tg(tg); |
0e9f4164 TH |
849 | |
850 | /* see throtl_add_bio_tg() */ | |
851 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
852 | } |
853 | ||
77216b04 | 854 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 855 | { |
73f0d49a | 856 | struct throtl_service_queue *sq = &tg->service_queue; |
6bc9c2b4 TH |
857 | struct throtl_service_queue *parent_sq = sq->parent_sq; |
858 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
e43473b7 VG |
859 | struct bio *bio; |
860 | ||
73f0d49a TH |
861 | bio = bio_list_pop(&sq->bio_lists[rw]); |
862 | sq->nr_queued[rw]--; | |
e43473b7 VG |
863 | |
864 | throtl_charge_bio(tg, bio); | |
6bc9c2b4 TH |
865 | |
866 | /* | |
867 | * If our parent is another tg, we just need to transfer @bio to | |
868 | * the parent using throtl_add_bio_tg(). If our parent is | |
869 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
870 | * bio_lists[] and decrease total number queued. The caller is | |
871 | * responsible for issuing these bios. | |
872 | */ | |
873 | if (parent_tg) { | |
874 | throtl_add_bio_tg(bio, parent_tg); | |
875 | } else { | |
876 | bio_list_add(&parent_sq->bio_lists[rw], bio); | |
877 | BUG_ON(tg->td->nr_queued[rw] <= 0); | |
878 | tg->td->nr_queued[rw]--; | |
879 | } | |
e43473b7 | 880 | |
0f3457f6 | 881 | throtl_trim_slice(tg, rw); |
6bc9c2b4 TH |
882 | |
883 | /* @bio is transferred to parent, drop its blkg reference */ | |
884 | blkg_put(tg_to_blkg(tg)); | |
e43473b7 VG |
885 | } |
886 | ||
77216b04 | 887 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 888 | { |
73f0d49a | 889 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
890 | unsigned int nr_reads = 0, nr_writes = 0; |
891 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 892 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
893 | struct bio *bio; |
894 | ||
895 | /* Try to dispatch 75% READS and 25% WRITES */ | |
896 | ||
73f0d49a | 897 | while ((bio = bio_list_peek(&sq->bio_lists[READ])) && |
0f3457f6 | 898 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 899 | |
77216b04 | 900 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
901 | nr_reads++; |
902 | ||
903 | if (nr_reads >= max_nr_reads) | |
904 | break; | |
905 | } | |
906 | ||
73f0d49a | 907 | while ((bio = bio_list_peek(&sq->bio_lists[WRITE])) && |
0f3457f6 | 908 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 909 | |
77216b04 | 910 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
911 | nr_writes++; |
912 | ||
913 | if (nr_writes >= max_nr_writes) | |
914 | break; | |
915 | } | |
916 | ||
917 | return nr_reads + nr_writes; | |
918 | } | |
919 | ||
651930bc | 920 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
921 | { |
922 | unsigned int nr_disp = 0; | |
e43473b7 VG |
923 | |
924 | while (1) { | |
73f0d49a TH |
925 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
926 | struct throtl_service_queue *sq = &tg->service_queue; | |
e43473b7 VG |
927 | |
928 | if (!tg) | |
929 | break; | |
930 | ||
931 | if (time_before(jiffies, tg->disptime)) | |
932 | break; | |
933 | ||
77216b04 | 934 | throtl_dequeue_tg(tg); |
e43473b7 | 935 | |
77216b04 | 936 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 937 | |
73f0d49a | 938 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
77216b04 | 939 | tg_update_disptime(tg); |
e43473b7 VG |
940 | |
941 | if (nr_disp >= throtl_quantum) | |
942 | break; | |
943 | } | |
944 | ||
945 | return nr_disp; | |
946 | } | |
947 | ||
6e1a5704 TH |
948 | /** |
949 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
950 | * @arg: the throtl_service_queue being serviced | |
951 | * | |
952 | * This timer is armed when a child throtl_grp with active bio's become | |
953 | * pending and queued on the service_queue's pending_tree and expires when | |
954 | * the first child throtl_grp should be dispatched. This function | |
955 | * dispatches bio's from the children throtl_grps and kicks | |
956 | * throtl_data->dispatch_work if there are bio's ready to be issued. | |
957 | */ | |
69df0ab0 TH |
958 | static void throtl_pending_timer_fn(unsigned long arg) |
959 | { | |
960 | struct throtl_service_queue *sq = (void *)arg; | |
961 | struct throtl_data *td = sq_to_td(sq); | |
cb76199c | 962 | struct request_queue *q = td->queue; |
7f52f98c | 963 | bool dispatched = false; |
6e1a5704 | 964 | int ret; |
e43473b7 VG |
965 | |
966 | spin_lock_irq(q->queue_lock); | |
967 | ||
7f52f98c TH |
968 | while (true) { |
969 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
970 | td->nr_queued[READ] + td->nr_queued[WRITE], | |
971 | td->nr_queued[READ], td->nr_queued[WRITE]); | |
972 | ||
973 | ret = throtl_select_dispatch(sq); | |
974 | if (ret) { | |
7f52f98c TH |
975 | throtl_log(sq, "bios disp=%u", ret); |
976 | dispatched = true; | |
977 | } | |
e43473b7 | 978 | |
7f52f98c TH |
979 | if (throtl_schedule_next_dispatch(sq, false)) |
980 | break; | |
e43473b7 | 981 | |
7f52f98c TH |
982 | /* this dispatch windows is still open, relax and repeat */ |
983 | spin_unlock_irq(q->queue_lock); | |
984 | cpu_relax(); | |
985 | spin_lock_irq(q->queue_lock); | |
651930bc | 986 | } |
e43473b7 | 987 | |
6e1a5704 TH |
988 | if (dispatched) |
989 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
990 | ||
e43473b7 | 991 | spin_unlock_irq(q->queue_lock); |
6e1a5704 | 992 | } |
e43473b7 | 993 | |
6e1a5704 TH |
994 | /** |
995 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
996 | * @work: work item being executed | |
997 | * | |
998 | * This function is queued for execution when bio's reach the bio_lists[] | |
999 | * of throtl_data->service_queue. Those bio's are ready and issued by this | |
1000 | * function. | |
1001 | */ | |
1002 | void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
1003 | { | |
1004 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1005 | dispatch_work); | |
1006 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1007 | struct request_queue *q = td->queue; | |
1008 | struct bio_list bio_list_on_stack; | |
1009 | struct bio *bio; | |
1010 | struct blk_plug plug; | |
1011 | int rw; | |
1012 | ||
1013 | bio_list_init(&bio_list_on_stack); | |
1014 | ||
1015 | spin_lock_irq(q->queue_lock); | |
1016 | for (rw = READ; rw <= WRITE; rw++) { | |
1017 | bio_list_merge(&bio_list_on_stack, &td_sq->bio_lists[rw]); | |
1018 | bio_list_init(&td_sq->bio_lists[rw]); | |
1019 | } | |
1020 | spin_unlock_irq(q->queue_lock); | |
1021 | ||
1022 | if (!bio_list_empty(&bio_list_on_stack)) { | |
69d60eb9 | 1023 | blk_start_plug(&plug); |
e43473b7 VG |
1024 | while((bio = bio_list_pop(&bio_list_on_stack))) |
1025 | generic_make_request(bio); | |
69d60eb9 | 1026 | blk_finish_plug(&plug); |
e43473b7 | 1027 | } |
e43473b7 VG |
1028 | } |
1029 | ||
f95a04af TH |
1030 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
1031 | struct blkg_policy_data *pd, int off) | |
41b38b6d | 1032 | { |
f95a04af | 1033 | struct throtl_grp *tg = pd_to_tg(pd); |
41b38b6d TH |
1034 | struct blkg_rwstat rwstat = { }, tmp; |
1035 | int i, cpu; | |
1036 | ||
1037 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 1038 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
1039 | |
1040 | tmp = blkg_rwstat_read((void *)sc + off); | |
1041 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
1042 | rwstat.cnt[i] += tmp.cnt[i]; | |
1043 | } | |
1044 | ||
f95a04af | 1045 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
41b38b6d TH |
1046 | } |
1047 | ||
8a3d2615 TH |
1048 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
1049 | struct seq_file *sf) | |
41b38b6d | 1050 | { |
3c798398 | 1051 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
41b38b6d | 1052 | |
3c798398 | 1053 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 1054 | cft->private, true); |
41b38b6d TH |
1055 | return 0; |
1056 | } | |
1057 | ||
f95a04af TH |
1058 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
1059 | int off) | |
60c2bc2d | 1060 | { |
f95a04af TH |
1061 | struct throtl_grp *tg = pd_to_tg(pd); |
1062 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 1063 | |
af133ceb | 1064 | if (v == -1) |
60c2bc2d | 1065 | return 0; |
f95a04af | 1066 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
1067 | } |
1068 | ||
f95a04af TH |
1069 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
1070 | int off) | |
e43473b7 | 1071 | { |
f95a04af TH |
1072 | struct throtl_grp *tg = pd_to_tg(pd); |
1073 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 1074 | |
af133ceb TH |
1075 | if (v == -1) |
1076 | return 0; | |
f95a04af | 1077 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
1078 | } |
1079 | ||
af133ceb TH |
1080 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1081 | struct seq_file *sf) | |
8e89d13f | 1082 | { |
3c798398 TH |
1083 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, |
1084 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 1085 | return 0; |
8e89d13f VG |
1086 | } |
1087 | ||
af133ceb TH |
1088 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1089 | struct seq_file *sf) | |
8e89d13f | 1090 | { |
3c798398 TH |
1091 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, |
1092 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 1093 | return 0; |
60c2bc2d TH |
1094 | } |
1095 | ||
af133ceb TH |
1096 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
1097 | bool is_u64) | |
60c2bc2d | 1098 | { |
3c798398 | 1099 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
60c2bc2d | 1100 | struct blkg_conf_ctx ctx; |
af133ceb | 1101 | struct throtl_grp *tg; |
69df0ab0 | 1102 | struct throtl_service_queue *sq; |
60c2bc2d TH |
1103 | int ret; |
1104 | ||
3c798398 | 1105 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
1106 | if (ret) |
1107 | return ret; | |
1108 | ||
af133ceb | 1109 | tg = blkg_to_tg(ctx.blkg); |
69df0ab0 | 1110 | sq = &tg->service_queue; |
af133ceb | 1111 | |
a2b1693b TH |
1112 | if (!ctx.v) |
1113 | ctx.v = -1; | |
af133ceb | 1114 | |
a2b1693b TH |
1115 | if (is_u64) |
1116 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
1117 | else | |
1118 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 1119 | |
fda6f272 TH |
1120 | throtl_log(&tg->service_queue, |
1121 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
1122 | tg->bps[READ], tg->bps[WRITE], | |
1123 | tg->iops[READ], tg->iops[WRITE]); | |
632b4493 TH |
1124 | |
1125 | /* | |
1126 | * We're already holding queue_lock and know @tg is valid. Let's | |
1127 | * apply the new config directly. | |
1128 | * | |
1129 | * Restart the slices for both READ and WRITES. It might happen | |
1130 | * that a group's limit are dropped suddenly and we don't want to | |
1131 | * account recently dispatched IO with new low rate. | |
1132 | */ | |
0f3457f6 TH |
1133 | throtl_start_new_slice(tg, 0); |
1134 | throtl_start_new_slice(tg, 1); | |
632b4493 | 1135 | |
5b2c16aa | 1136 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 1137 | tg_update_disptime(tg); |
7f52f98c | 1138 | throtl_schedule_next_dispatch(sq->parent_sq, true); |
632b4493 | 1139 | } |
60c2bc2d TH |
1140 | |
1141 | blkg_conf_finish(&ctx); | |
a2b1693b | 1142 | return 0; |
8e89d13f VG |
1143 | } |
1144 | ||
af133ceb TH |
1145 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1146 | const char *buf) | |
60c2bc2d | 1147 | { |
af133ceb | 1148 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
1149 | } |
1150 | ||
af133ceb TH |
1151 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1152 | const char *buf) | |
60c2bc2d | 1153 | { |
af133ceb | 1154 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
1155 | } |
1156 | ||
1157 | static struct cftype throtl_files[] = { | |
1158 | { | |
1159 | .name = "throttle.read_bps_device", | |
af133ceb TH |
1160 | .private = offsetof(struct throtl_grp, bps[READ]), |
1161 | .read_seq_string = tg_print_conf_u64, | |
1162 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1163 | .max_write_len = 256, |
1164 | }, | |
1165 | { | |
1166 | .name = "throttle.write_bps_device", | |
af133ceb TH |
1167 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
1168 | .read_seq_string = tg_print_conf_u64, | |
1169 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1170 | .max_write_len = 256, |
1171 | }, | |
1172 | { | |
1173 | .name = "throttle.read_iops_device", | |
af133ceb TH |
1174 | .private = offsetof(struct throtl_grp, iops[READ]), |
1175 | .read_seq_string = tg_print_conf_uint, | |
1176 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1177 | .max_write_len = 256, |
1178 | }, | |
1179 | { | |
1180 | .name = "throttle.write_iops_device", | |
af133ceb TH |
1181 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
1182 | .read_seq_string = tg_print_conf_uint, | |
1183 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1184 | .max_write_len = 256, |
1185 | }, | |
1186 | { | |
1187 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1188 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1189 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1190 | }, |
1191 | { | |
1192 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1193 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1194 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1195 | }, |
1196 | { } /* terminate */ | |
1197 | }; | |
1198 | ||
da527770 | 1199 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1200 | { |
1201 | struct throtl_data *td = q->td; | |
1202 | ||
69df0ab0 | 1203 | cancel_work_sync(&td->dispatch_work); |
e43473b7 VG |
1204 | } |
1205 | ||
3c798398 | 1206 | static struct blkcg_policy blkcg_policy_throtl = { |
f9fcc2d3 TH |
1207 | .pd_size = sizeof(struct throtl_grp), |
1208 | .cftypes = throtl_files, | |
1209 | ||
1210 | .pd_init_fn = throtl_pd_init, | |
1211 | .pd_exit_fn = throtl_pd_exit, | |
1212 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 VG |
1213 | }; |
1214 | ||
bc16a4f9 | 1215 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1216 | { |
1217 | struct throtl_data *td = q->td; | |
1218 | struct throtl_grp *tg; | |
73f0d49a | 1219 | struct throtl_service_queue *sq; |
0e9f4164 | 1220 | bool rw = bio_data_dir(bio); |
3c798398 | 1221 | struct blkcg *blkcg; |
bc16a4f9 | 1222 | bool throttled = false; |
e43473b7 | 1223 | |
2a0f61e6 TH |
1224 | /* see throtl_charge_bio() */ |
1225 | if (bio->bi_rw & REQ_THROTTLED) | |
bc16a4f9 | 1226 | goto out; |
e43473b7 | 1227 | |
af75cd3c VG |
1228 | /* |
1229 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1230 | * basic fields like stats and io rates. If a group has no rules, | |
1231 | * just update the dispatch stats in lockless manner and return. | |
1232 | */ | |
af75cd3c | 1233 | rcu_read_lock(); |
3c798398 | 1234 | blkcg = bio_blkcg(bio); |
cd1604fa | 1235 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1236 | if (tg) { |
af75cd3c | 1237 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1238 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1239 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1240 | goto out_unlock_rcu; |
af75cd3c VG |
1241 | } |
1242 | } | |
af75cd3c VG |
1243 | |
1244 | /* | |
1245 | * Either group has not been allocated yet or it is not an unlimited | |
1246 | * IO group | |
1247 | */ | |
e43473b7 | 1248 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1249 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1250 | if (unlikely(!tg)) |
1251 | goto out_unlock; | |
f469a7b4 | 1252 | |
73f0d49a TH |
1253 | sq = &tg->service_queue; |
1254 | ||
9e660acf TH |
1255 | while (true) { |
1256 | /* throtl is FIFO - if bios are already queued, should queue */ | |
1257 | if (sq->nr_queued[rw]) | |
1258 | break; | |
de701c74 | 1259 | |
9e660acf TH |
1260 | /* if above limits, break to queue */ |
1261 | if (!tg_may_dispatch(tg, bio, NULL)) | |
1262 | break; | |
1263 | ||
1264 | /* within limits, let's charge and dispatch directly */ | |
e43473b7 | 1265 | throtl_charge_bio(tg, bio); |
04521db0 VG |
1266 | |
1267 | /* | |
1268 | * We need to trim slice even when bios are not being queued | |
1269 | * otherwise it might happen that a bio is not queued for | |
1270 | * a long time and slice keeps on extending and trim is not | |
1271 | * called for a long time. Now if limits are reduced suddenly | |
1272 | * we take into account all the IO dispatched so far at new | |
1273 | * low rate and * newly queued IO gets a really long dispatch | |
1274 | * time. | |
1275 | * | |
1276 | * So keep on trimming slice even if bio is not queued. | |
1277 | */ | |
0f3457f6 | 1278 | throtl_trim_slice(tg, rw); |
9e660acf TH |
1279 | |
1280 | /* | |
1281 | * @bio passed through this layer without being throttled. | |
1282 | * Climb up the ladder. If we''re already at the top, it | |
1283 | * can be executed directly. | |
1284 | */ | |
1285 | sq = sq->parent_sq; | |
1286 | tg = sq_to_tg(sq); | |
1287 | if (!tg) | |
1288 | goto out_unlock; | |
e43473b7 VG |
1289 | } |
1290 | ||
9e660acf | 1291 | /* out-of-limit, queue to @tg */ |
fda6f272 TH |
1292 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
1293 | rw == READ ? 'R' : 'W', | |
1294 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], | |
1295 | tg->io_disp[rw], tg->iops[rw], | |
1296 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
e43473b7 | 1297 | |
671058fb | 1298 | bio_associate_current(bio); |
6bc9c2b4 | 1299 | tg->td->nr_queued[rw]++; |
77216b04 | 1300 | throtl_add_bio_tg(bio, tg); |
bc16a4f9 | 1301 | throttled = true; |
e43473b7 | 1302 | |
7f52f98c TH |
1303 | /* |
1304 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
1305 | * was empty before @bio. The forced scheduling isn't likely to | |
1306 | * cause undue delay as @bio is likely to be dispatched directly if | |
1307 | * its @tg's disptime is not in the future. | |
1308 | */ | |
0e9f4164 | 1309 | if (tg->flags & THROTL_TG_WAS_EMPTY) { |
77216b04 | 1310 | tg_update_disptime(tg); |
7f52f98c | 1311 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); |
e43473b7 VG |
1312 | } |
1313 | ||
bc16a4f9 | 1314 | out_unlock: |
e43473b7 | 1315 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1316 | out_unlock_rcu: |
1317 | rcu_read_unlock(); | |
bc16a4f9 | 1318 | out: |
2a0f61e6 TH |
1319 | /* |
1320 | * As multiple blk-throtls may stack in the same issue path, we | |
1321 | * don't want bios to leave with the flag set. Clear the flag if | |
1322 | * being issued. | |
1323 | */ | |
1324 | if (!throttled) | |
1325 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1326 | return throttled; |
e43473b7 VG |
1327 | } |
1328 | ||
2a12f0dc TH |
1329 | /* |
1330 | * Dispatch all bios from all children tg's queued on @parent_sq. On | |
1331 | * return, @parent_sq is guaranteed to not have any active children tg's | |
1332 | * and all bios from previously active tg's are on @parent_sq->bio_lists[]. | |
1333 | */ | |
1334 | static void tg_drain_bios(struct throtl_service_queue *parent_sq) | |
1335 | { | |
1336 | struct throtl_grp *tg; | |
1337 | ||
1338 | while ((tg = throtl_rb_first(parent_sq))) { | |
1339 | struct throtl_service_queue *sq = &tg->service_queue; | |
1340 | struct bio *bio; | |
1341 | ||
1342 | throtl_dequeue_tg(tg); | |
1343 | ||
1344 | while ((bio = bio_list_peek(&sq->bio_lists[READ]))) | |
1345 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); | |
1346 | while ((bio = bio_list_peek(&sq->bio_lists[WRITE]))) | |
1347 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); | |
1348 | } | |
1349 | } | |
1350 | ||
c9a929dd TH |
1351 | /** |
1352 | * blk_throtl_drain - drain throttled bios | |
1353 | * @q: request_queue to drain throttled bios for | |
1354 | * | |
1355 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1356 | */ | |
1357 | void blk_throtl_drain(struct request_queue *q) | |
1358 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1359 | { | |
1360 | struct throtl_data *td = q->td; | |
2a12f0dc TH |
1361 | struct blkcg_gq *blkg; |
1362 | struct cgroup *pos_cgrp; | |
c9a929dd | 1363 | struct bio *bio; |
651930bc | 1364 | int rw; |
c9a929dd | 1365 | |
8bcb6c7d | 1366 | queue_lockdep_assert_held(q); |
2a12f0dc | 1367 | rcu_read_lock(); |
c9a929dd | 1368 | |
2a12f0dc TH |
1369 | /* |
1370 | * Drain each tg while doing post-order walk on the blkg tree, so | |
1371 | * that all bios are propagated to td->service_queue. It'd be | |
1372 | * better to walk service_queue tree directly but blkg walk is | |
1373 | * easier. | |
1374 | */ | |
1375 | blkg_for_each_descendant_post(blkg, pos_cgrp, td->queue->root_blkg) | |
1376 | tg_drain_bios(&blkg_to_tg(blkg)->service_queue); | |
73f0d49a | 1377 | |
2a12f0dc | 1378 | tg_drain_bios(&td_root_tg(td)->service_queue); |
c9a929dd | 1379 | |
2a12f0dc TH |
1380 | /* finally, transfer bios from top-level tg's into the td */ |
1381 | tg_drain_bios(&td->service_queue); | |
1382 | ||
1383 | rcu_read_unlock(); | |
c9a929dd TH |
1384 | spin_unlock_irq(q->queue_lock); |
1385 | ||
2a12f0dc | 1386 | /* all bios now should be in td->service_queue, issue them */ |
651930bc | 1387 | for (rw = READ; rw <= WRITE; rw++) |
2a12f0dc | 1388 | while ((bio = bio_list_pop(&td->service_queue.bio_lists[rw]))) |
651930bc | 1389 | generic_make_request(bio); |
c9a929dd TH |
1390 | |
1391 | spin_lock_irq(q->queue_lock); | |
1392 | } | |
1393 | ||
e43473b7 VG |
1394 | int blk_throtl_init(struct request_queue *q) |
1395 | { | |
1396 | struct throtl_data *td; | |
a2b1693b | 1397 | int ret; |
e43473b7 VG |
1398 | |
1399 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1400 | if (!td) | |
1401 | return -ENOMEM; | |
1402 | ||
69df0ab0 | 1403 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
77216b04 | 1404 | throtl_service_queue_init(&td->service_queue, NULL); |
e43473b7 | 1405 | |
cd1604fa | 1406 | q->td = td; |
29b12589 | 1407 | td->queue = q; |
02977e4a | 1408 | |
a2b1693b | 1409 | /* activate policy */ |
3c798398 | 1410 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1411 | if (ret) |
f51b802c | 1412 | kfree(td); |
a2b1693b | 1413 | return ret; |
e43473b7 VG |
1414 | } |
1415 | ||
1416 | void blk_throtl_exit(struct request_queue *q) | |
1417 | { | |
c875f4d0 | 1418 | BUG_ON(!q->td); |
da527770 | 1419 | throtl_shutdown_wq(q); |
3c798398 | 1420 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1421 | kfree(q->td); |
e43473b7 VG |
1422 | } |
1423 | ||
1424 | static int __init throtl_init(void) | |
1425 | { | |
450adcbe VG |
1426 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1427 | if (!kthrotld_workqueue) | |
1428 | panic("Failed to create kthrotld\n"); | |
1429 | ||
3c798398 | 1430 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1431 | } |
1432 | ||
1433 | module_init(throtl_init); |