]>
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 TH |
29 | struct throtl_service_queue { |
30 | struct rb_root pending_tree; /* RB tree of active tgs */ | |
31 | struct rb_node *first_pending; /* first node in the tree */ | |
32 | unsigned int nr_pending; /* # queued in the tree */ | |
33 | unsigned long first_pending_disptime; /* disptime of the first tg */ | |
e43473b7 VG |
34 | }; |
35 | ||
c9e0332e TH |
36 | #define THROTL_SERVICE_QUEUE_INITIALIZER \ |
37 | (struct throtl_service_queue){ .pending_tree = RB_ROOT } | |
e43473b7 | 38 | |
5b2c16aa TH |
39 | enum tg_state_flags { |
40 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
41 | }; | |
42 | ||
e43473b7 VG |
43 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
44 | ||
8a3d2615 TH |
45 | /* Per-cpu group stats */ |
46 | struct tg_stats_cpu { | |
47 | /* total bytes transferred */ | |
48 | struct blkg_rwstat service_bytes; | |
49 | /* total IOs serviced, post merge */ | |
50 | struct blkg_rwstat serviced; | |
51 | }; | |
52 | ||
e43473b7 | 53 | struct throtl_grp { |
f95a04af TH |
54 | /* must be the first member */ |
55 | struct blkg_policy_data pd; | |
56 | ||
c9e0332e | 57 | /* active throtl group service_queue member */ |
e43473b7 VG |
58 | struct rb_node rb_node; |
59 | ||
0f3457f6 TH |
60 | /* throtl_data this group belongs to */ |
61 | struct throtl_data *td; | |
62 | ||
e43473b7 VG |
63 | /* |
64 | * Dispatch time in jiffies. This is the estimated time when group | |
65 | * will unthrottle and is ready to dispatch more bio. It is used as | |
66 | * key to sort active groups in service tree. | |
67 | */ | |
68 | unsigned long disptime; | |
69 | ||
e43473b7 VG |
70 | unsigned int flags; |
71 | ||
72 | /* Two lists for READ and WRITE */ | |
73 | struct bio_list bio_lists[2]; | |
74 | ||
75 | /* Number of queued bios on READ and WRITE lists */ | |
76 | unsigned int nr_queued[2]; | |
77 | ||
78 | /* bytes per second rate limits */ | |
79 | uint64_t bps[2]; | |
80 | ||
8e89d13f VG |
81 | /* IOPS limits */ |
82 | unsigned int iops[2]; | |
83 | ||
e43473b7 VG |
84 | /* Number of bytes disptached in current slice */ |
85 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
86 | /* Number of bio's dispatched in current slice */ |
87 | unsigned int io_disp[2]; | |
e43473b7 VG |
88 | |
89 | /* When did we start a new slice */ | |
90 | unsigned long slice_start[2]; | |
91 | unsigned long slice_end[2]; | |
fe071437 | 92 | |
8a3d2615 TH |
93 | /* Per cpu stats pointer */ |
94 | struct tg_stats_cpu __percpu *stats_cpu; | |
95 | ||
96 | /* List of tgs waiting for per cpu stats memory to be allocated */ | |
97 | struct list_head stats_alloc_node; | |
e43473b7 VG |
98 | }; |
99 | ||
100 | struct throtl_data | |
101 | { | |
e43473b7 | 102 | /* service tree for active throtl groups */ |
c9e0332e | 103 | struct throtl_service_queue service_queue; |
e43473b7 | 104 | |
e43473b7 VG |
105 | struct request_queue *queue; |
106 | ||
107 | /* Total Number of queued bios on READ and WRITE lists */ | |
108 | unsigned int nr_queued[2]; | |
109 | ||
110 | /* | |
02977e4a | 111 | * number of total undestroyed groups |
e43473b7 VG |
112 | */ |
113 | unsigned int nr_undestroyed_grps; | |
114 | ||
115 | /* Work for dispatching throttled bios */ | |
cb76199c | 116 | struct delayed_work dispatch_work; |
e43473b7 VG |
117 | }; |
118 | ||
8a3d2615 TH |
119 | /* list and work item to allocate percpu group stats */ |
120 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
121 | static LIST_HEAD(tg_stats_alloc_list); | |
122 | ||
123 | static void tg_stats_alloc_fn(struct work_struct *); | |
124 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
125 | ||
f95a04af TH |
126 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
127 | { | |
128 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
129 | } | |
130 | ||
3c798398 | 131 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 132 | { |
f95a04af | 133 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
134 | } |
135 | ||
3c798398 | 136 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 137 | { |
f95a04af | 138 | return pd_to_blkg(&tg->pd); |
0381411e TH |
139 | } |
140 | ||
03d8e111 TH |
141 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) |
142 | { | |
143 | return blkg_to_tg(td->queue->root_blkg); | |
144 | } | |
145 | ||
0f3457f6 | 146 | #define throtl_log_tg(tg, fmt, args...) do { \ |
54e7ed12 TH |
147 | char __pbuf[128]; \ |
148 | \ | |
149 | blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \ | |
0f3457f6 | 150 | blk_add_trace_msg((tg)->td->queue, "throtl %s " fmt, __pbuf, ##args); \ |
54e7ed12 | 151 | } while (0) |
e43473b7 VG |
152 | |
153 | #define throtl_log(td, fmt, args...) \ | |
154 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) | |
155 | ||
8a3d2615 TH |
156 | /* |
157 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
3b07e9ca | 158 | * system_wq once there are some groups on the alloc_list waiting for |
8a3d2615 TH |
159 | * allocation. |
160 | */ | |
161 | static void tg_stats_alloc_fn(struct work_struct *work) | |
162 | { | |
163 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
164 | struct delayed_work *dwork = to_delayed_work(work); | |
165 | bool empty = false; | |
166 | ||
167 | alloc_stats: | |
168 | if (!stats_cpu) { | |
169 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
170 | if (!stats_cpu) { | |
171 | /* allocation failed, try again after some time */ | |
3b07e9ca | 172 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); |
8a3d2615 TH |
173 | return; |
174 | } | |
175 | } | |
176 | ||
177 | spin_lock_irq(&tg_stats_alloc_lock); | |
178 | ||
179 | if (!list_empty(&tg_stats_alloc_list)) { | |
180 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
181 | struct throtl_grp, | |
182 | stats_alloc_node); | |
183 | swap(tg->stats_cpu, stats_cpu); | |
184 | list_del_init(&tg->stats_alloc_node); | |
185 | } | |
186 | ||
187 | empty = list_empty(&tg_stats_alloc_list); | |
188 | spin_unlock_irq(&tg_stats_alloc_lock); | |
189 | if (!empty) | |
190 | goto alloc_stats; | |
191 | } | |
192 | ||
3c798398 | 193 | static void throtl_pd_init(struct blkcg_gq *blkg) |
a29a171e | 194 | { |
0381411e | 195 | struct throtl_grp *tg = blkg_to_tg(blkg); |
ff26eaad | 196 | unsigned long flags; |
cd1604fa | 197 | |
a29a171e | 198 | RB_CLEAR_NODE(&tg->rb_node); |
0f3457f6 | 199 | tg->td = blkg->q->td; |
a29a171e VG |
200 | bio_list_init(&tg->bio_lists[0]); |
201 | bio_list_init(&tg->bio_lists[1]); | |
a29a171e | 202 | |
e56da7e2 TH |
203 | tg->bps[READ] = -1; |
204 | tg->bps[WRITE] = -1; | |
205 | tg->iops[READ] = -1; | |
206 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
207 | |
208 | /* | |
209 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
210 | * but percpu allocator can't be called from IO path. Queue tg on | |
211 | * tg_stats_alloc_list and allocate from work item. | |
212 | */ | |
ff26eaad | 213 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 214 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); |
3b07e9ca | 215 | schedule_delayed_work(&tg_stats_alloc_work, 0); |
ff26eaad | 216 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
217 | } |
218 | ||
3c798398 | 219 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
8a3d2615 TH |
220 | { |
221 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
ff26eaad | 222 | unsigned long flags; |
8a3d2615 | 223 | |
ff26eaad | 224 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 225 | list_del_init(&tg->stats_alloc_node); |
ff26eaad | 226 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
227 | |
228 | free_percpu(tg->stats_cpu); | |
229 | } | |
230 | ||
3c798398 | 231 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
8a3d2615 TH |
232 | { |
233 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
234 | int cpu; | |
235 | ||
236 | if (tg->stats_cpu == NULL) | |
237 | return; | |
238 | ||
239 | for_each_possible_cpu(cpu) { | |
240 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
241 | ||
242 | blkg_rwstat_reset(&sc->service_bytes); | |
243 | blkg_rwstat_reset(&sc->serviced); | |
244 | } | |
a29a171e VG |
245 | } |
246 | ||
3c798398 TH |
247 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
248 | struct blkcg *blkcg) | |
e43473b7 | 249 | { |
be2c6b19 | 250 | /* |
3c798398 TH |
251 | * This is the common case when there are no blkcgs. Avoid lookup |
252 | * in this case | |
cd1604fa | 253 | */ |
3c798398 | 254 | if (blkcg == &blkcg_root) |
03d8e111 | 255 | return td_root_tg(td); |
e43473b7 | 256 | |
e8989fae | 257 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
258 | } |
259 | ||
cd1604fa | 260 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
3c798398 | 261 | struct blkcg *blkcg) |
e43473b7 | 262 | { |
f469a7b4 | 263 | struct request_queue *q = td->queue; |
cd1604fa | 264 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 265 | |
f469a7b4 | 266 | /* |
3c798398 TH |
267 | * This is the common case when there are no blkcgs. Avoid lookup |
268 | * in this case | |
f469a7b4 | 269 | */ |
3c798398 | 270 | if (blkcg == &blkcg_root) { |
03d8e111 | 271 | tg = td_root_tg(td); |
cd1604fa | 272 | } else { |
3c798398 | 273 | struct blkcg_gq *blkg; |
f469a7b4 | 274 | |
3c96cb32 | 275 | blkg = blkg_lookup_create(blkcg, q); |
f469a7b4 | 276 | |
cd1604fa TH |
277 | /* if %NULL and @q is alive, fall back to root_tg */ |
278 | if (!IS_ERR(blkg)) | |
0381411e | 279 | tg = blkg_to_tg(blkg); |
3f3299d5 | 280 | else if (!blk_queue_dying(q)) |
03d8e111 | 281 | tg = td_root_tg(td); |
f469a7b4 VG |
282 | } |
283 | ||
e43473b7 VG |
284 | return tg; |
285 | } | |
286 | ||
c9e0332e | 287 | static struct throtl_grp *throtl_rb_first(struct throtl_service_queue *sq) |
e43473b7 VG |
288 | { |
289 | /* Service tree is empty */ | |
c9e0332e | 290 | if (!sq->nr_pending) |
e43473b7 VG |
291 | return NULL; |
292 | ||
c9e0332e TH |
293 | if (!sq->first_pending) |
294 | sq->first_pending = rb_first(&sq->pending_tree); | |
e43473b7 | 295 | |
c9e0332e TH |
296 | if (sq->first_pending) |
297 | return rb_entry_tg(sq->first_pending); | |
e43473b7 VG |
298 | |
299 | return NULL; | |
300 | } | |
301 | ||
302 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
303 | { | |
304 | rb_erase(n, root); | |
305 | RB_CLEAR_NODE(n); | |
306 | } | |
307 | ||
c9e0332e | 308 | static void throtl_rb_erase(struct rb_node *n, struct throtl_service_queue *sq) |
e43473b7 | 309 | { |
c9e0332e TH |
310 | if (sq->first_pending == n) |
311 | sq->first_pending = NULL; | |
312 | rb_erase_init(n, &sq->pending_tree); | |
313 | --sq->nr_pending; | |
e43473b7 VG |
314 | } |
315 | ||
c9e0332e | 316 | static void update_min_dispatch_time(struct throtl_service_queue *sq) |
e43473b7 VG |
317 | { |
318 | struct throtl_grp *tg; | |
319 | ||
c9e0332e | 320 | tg = throtl_rb_first(sq); |
e43473b7 VG |
321 | if (!tg) |
322 | return; | |
323 | ||
c9e0332e | 324 | sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
325 | } |
326 | ||
c9e0332e TH |
327 | static void tg_service_queue_add(struct throtl_service_queue *sq, |
328 | struct throtl_grp *tg) | |
e43473b7 | 329 | { |
c9e0332e | 330 | struct rb_node **node = &sq->pending_tree.rb_node; |
e43473b7 VG |
331 | struct rb_node *parent = NULL; |
332 | struct throtl_grp *__tg; | |
333 | unsigned long key = tg->disptime; | |
334 | int left = 1; | |
335 | ||
336 | while (*node != NULL) { | |
337 | parent = *node; | |
338 | __tg = rb_entry_tg(parent); | |
339 | ||
340 | if (time_before(key, __tg->disptime)) | |
341 | node = &parent->rb_left; | |
342 | else { | |
343 | node = &parent->rb_right; | |
344 | left = 0; | |
345 | } | |
346 | } | |
347 | ||
348 | if (left) | |
c9e0332e | 349 | sq->first_pending = &tg->rb_node; |
e43473b7 VG |
350 | |
351 | rb_link_node(&tg->rb_node, parent, node); | |
c9e0332e | 352 | rb_insert_color(&tg->rb_node, &sq->pending_tree); |
e43473b7 VG |
353 | } |
354 | ||
355 | static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
356 | { | |
c9e0332e | 357 | struct throtl_service_queue *sq = &td->service_queue; |
e43473b7 | 358 | |
c9e0332e | 359 | tg_service_queue_add(sq, tg); |
5b2c16aa | 360 | tg->flags |= THROTL_TG_PENDING; |
c9e0332e | 361 | sq->nr_pending++; |
e43473b7 VG |
362 | } |
363 | ||
364 | static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
365 | { | |
5b2c16aa | 366 | if (!(tg->flags & THROTL_TG_PENDING)) |
e43473b7 VG |
367 | __throtl_enqueue_tg(td, tg); |
368 | } | |
369 | ||
370 | static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
371 | { | |
c9e0332e | 372 | throtl_rb_erase(&tg->rb_node, &td->service_queue); |
5b2c16aa | 373 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
374 | } |
375 | ||
376 | static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
377 | { | |
5b2c16aa | 378 | if (tg->flags & THROTL_TG_PENDING) |
e43473b7 VG |
379 | __throtl_dequeue_tg(td, tg); |
380 | } | |
381 | ||
a9131a27 TH |
382 | /* Call with queue lock held */ |
383 | static void throtl_schedule_delayed_work(struct throtl_data *td, | |
384 | unsigned long delay) | |
385 | { | |
386 | struct delayed_work *dwork = &td->dispatch_work; | |
387 | ||
6a525600 TH |
388 | mod_delayed_work(kthrotld_workqueue, dwork, delay); |
389 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", delay, jiffies); | |
a9131a27 TH |
390 | } |
391 | ||
e43473b7 VG |
392 | static void throtl_schedule_next_dispatch(struct throtl_data *td) |
393 | { | |
c9e0332e | 394 | struct throtl_service_queue *sq = &td->service_queue; |
e43473b7 | 395 | |
6a525600 | 396 | /* any pending children left? */ |
c9e0332e | 397 | if (!sq->nr_pending) |
e43473b7 VG |
398 | return; |
399 | ||
c9e0332e | 400 | update_min_dispatch_time(sq); |
e43473b7 | 401 | |
c9e0332e | 402 | if (time_before_eq(sq->first_pending_disptime, jiffies)) |
450adcbe | 403 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 404 | else |
c9e0332e | 405 | throtl_schedule_delayed_work(td, sq->first_pending_disptime - jiffies); |
e43473b7 VG |
406 | } |
407 | ||
0f3457f6 | 408 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
409 | { |
410 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 411 | tg->io_disp[rw] = 0; |
e43473b7 VG |
412 | tg->slice_start[rw] = jiffies; |
413 | tg->slice_end[rw] = jiffies + throtl_slice; | |
0f3457f6 | 414 | throtl_log_tg(tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", |
e43473b7 VG |
415 | rw == READ ? 'R' : 'W', tg->slice_start[rw], |
416 | tg->slice_end[rw], jiffies); | |
417 | } | |
418 | ||
0f3457f6 TH |
419 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
420 | unsigned long jiffy_end) | |
d1ae8ffd VG |
421 | { |
422 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
423 | } | |
424 | ||
0f3457f6 TH |
425 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
426 | unsigned long jiffy_end) | |
e43473b7 VG |
427 | { |
428 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
0f3457f6 | 429 | throtl_log_tg(tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", |
e43473b7 VG |
430 | rw == READ ? 'R' : 'W', tg->slice_start[rw], |
431 | tg->slice_end[rw], jiffies); | |
432 | } | |
433 | ||
434 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 435 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
436 | { |
437 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
438 | return 0; | |
439 | ||
440 | return 1; | |
441 | } | |
442 | ||
443 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 444 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 445 | { |
3aad5d3e VG |
446 | unsigned long nr_slices, time_elapsed, io_trim; |
447 | u64 bytes_trim, tmp; | |
e43473b7 VG |
448 | |
449 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
450 | ||
451 | /* | |
452 | * If bps are unlimited (-1), then time slice don't get | |
453 | * renewed. Don't try to trim the slice if slice is used. A new | |
454 | * slice will start when appropriate. | |
455 | */ | |
0f3457f6 | 456 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
457 | return; |
458 | ||
d1ae8ffd VG |
459 | /* |
460 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
461 | * that initially cgroup limit was very low resulting in high | |
462 | * slice_end, but later limit was bumped up and bio was dispached | |
463 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
464 | * is bad because it does not allow new slice to start. | |
465 | */ | |
466 | ||
0f3457f6 | 467 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); |
d1ae8ffd | 468 | |
e43473b7 VG |
469 | time_elapsed = jiffies - tg->slice_start[rw]; |
470 | ||
471 | nr_slices = time_elapsed / throtl_slice; | |
472 | ||
473 | if (!nr_slices) | |
474 | return; | |
3aad5d3e VG |
475 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
476 | do_div(tmp, HZ); | |
477 | bytes_trim = tmp; | |
e43473b7 | 478 | |
8e89d13f | 479 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 480 | |
8e89d13f | 481 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
482 | return; |
483 | ||
484 | if (tg->bytes_disp[rw] >= bytes_trim) | |
485 | tg->bytes_disp[rw] -= bytes_trim; | |
486 | else | |
487 | tg->bytes_disp[rw] = 0; | |
488 | ||
8e89d13f VG |
489 | if (tg->io_disp[rw] >= io_trim) |
490 | tg->io_disp[rw] -= io_trim; | |
491 | else | |
492 | tg->io_disp[rw] = 0; | |
493 | ||
e43473b7 VG |
494 | tg->slice_start[rw] += nr_slices * throtl_slice; |
495 | ||
0f3457f6 | 496 | throtl_log_tg(tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 497 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 498 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
499 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
500 | } | |
501 | ||
0f3457f6 TH |
502 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
503 | unsigned long *wait) | |
e43473b7 VG |
504 | { |
505 | bool rw = bio_data_dir(bio); | |
8e89d13f | 506 | unsigned int io_allowed; |
e43473b7 | 507 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 508 | u64 tmp; |
e43473b7 | 509 | |
8e89d13f | 510 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 511 | |
8e89d13f VG |
512 | /* Slice has just started. Consider one slice interval */ |
513 | if (!jiffy_elapsed) | |
514 | jiffy_elapsed_rnd = throtl_slice; | |
515 | ||
516 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
517 | ||
c49c06e4 VG |
518 | /* |
519 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
520 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
521 | * will allow dispatch after 1 second and after that slice should | |
522 | * have been trimmed. | |
523 | */ | |
524 | ||
525 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
526 | do_div(tmp, HZ); | |
527 | ||
528 | if (tmp > UINT_MAX) | |
529 | io_allowed = UINT_MAX; | |
530 | else | |
531 | io_allowed = tmp; | |
8e89d13f VG |
532 | |
533 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
534 | if (wait) |
535 | *wait = 0; | |
536 | return 1; | |
537 | } | |
538 | ||
8e89d13f VG |
539 | /* Calc approx time to dispatch */ |
540 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
541 | ||
542 | if (jiffy_wait > jiffy_elapsed) | |
543 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
544 | else | |
545 | jiffy_wait = 1; | |
546 | ||
547 | if (wait) | |
548 | *wait = jiffy_wait; | |
549 | return 0; | |
550 | } | |
551 | ||
0f3457f6 TH |
552 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
553 | unsigned long *wait) | |
8e89d13f VG |
554 | { |
555 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 556 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 557 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
558 | |
559 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
560 | ||
561 | /* Slice has just started. Consider one slice interval */ | |
562 | if (!jiffy_elapsed) | |
563 | jiffy_elapsed_rnd = throtl_slice; | |
564 | ||
565 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
566 | ||
5e901a2b VG |
567 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
568 | do_div(tmp, HZ); | |
3aad5d3e | 569 | bytes_allowed = tmp; |
e43473b7 VG |
570 | |
571 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
572 | if (wait) | |
573 | *wait = 0; | |
574 | return 1; | |
575 | } | |
576 | ||
577 | /* Calc approx time to dispatch */ | |
578 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
579 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
580 | ||
581 | if (!jiffy_wait) | |
582 | jiffy_wait = 1; | |
583 | ||
584 | /* | |
585 | * This wait time is without taking into consideration the rounding | |
586 | * up we did. Add that time also. | |
587 | */ | |
588 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
589 | if (wait) |
590 | *wait = jiffy_wait; | |
8e89d13f VG |
591 | return 0; |
592 | } | |
593 | ||
af75cd3c VG |
594 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
595 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
596 | return 1; | |
597 | return 0; | |
598 | } | |
599 | ||
8e89d13f VG |
600 | /* |
601 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
602 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
603 | */ | |
0f3457f6 TH |
604 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
605 | unsigned long *wait) | |
8e89d13f VG |
606 | { |
607 | bool rw = bio_data_dir(bio); | |
608 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
609 | ||
610 | /* | |
611 | * Currently whole state machine of group depends on first bio | |
612 | * queued in the group bio list. So one should not be calling | |
613 | * this function with a different bio if there are other bios | |
614 | * queued. | |
615 | */ | |
616 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); | |
e43473b7 | 617 | |
8e89d13f VG |
618 | /* If tg->bps = -1, then BW is unlimited */ |
619 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
620 | if (wait) | |
621 | *wait = 0; | |
622 | return 1; | |
623 | } | |
624 | ||
625 | /* | |
626 | * If previous slice expired, start a new one otherwise renew/extend | |
627 | * existing slice to make sure it is at least throtl_slice interval | |
628 | * long since now. | |
629 | */ | |
0f3457f6 TH |
630 | if (throtl_slice_used(tg, rw)) |
631 | throtl_start_new_slice(tg, rw); | |
8e89d13f VG |
632 | else { |
633 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
0f3457f6 | 634 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); |
8e89d13f VG |
635 | } |
636 | ||
0f3457f6 TH |
637 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
638 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
639 | if (wait) |
640 | *wait = 0; | |
641 | return 1; | |
642 | } | |
643 | ||
644 | max_wait = max(bps_wait, iops_wait); | |
645 | ||
646 | if (wait) | |
647 | *wait = max_wait; | |
648 | ||
649 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 650 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
651 | |
652 | return 0; | |
653 | } | |
654 | ||
3c798398 | 655 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
656 | int rw) |
657 | { | |
8a3d2615 TH |
658 | struct throtl_grp *tg = blkg_to_tg(blkg); |
659 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
660 | unsigned long flags; |
661 | ||
662 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 663 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
664 | return; |
665 | ||
666 | /* | |
667 | * Disabling interrupts to provide mutual exclusion between two | |
668 | * writes on same cpu. It probably is not needed for 64bit. Not | |
669 | * optimizing that case yet. | |
670 | */ | |
671 | local_irq_save(flags); | |
672 | ||
8a3d2615 | 673 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 674 | |
629ed0b1 TH |
675 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
676 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
677 | ||
678 | local_irq_restore(flags); | |
679 | } | |
680 | ||
e43473b7 VG |
681 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
682 | { | |
683 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
684 | |
685 | /* Charge the bio to the group */ | |
686 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 687 | tg->io_disp[rw]++; |
e43473b7 | 688 | |
629ed0b1 | 689 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw); |
e43473b7 VG |
690 | } |
691 | ||
692 | static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg, | |
693 | struct bio *bio) | |
694 | { | |
695 | bool rw = bio_data_dir(bio); | |
696 | ||
697 | bio_list_add(&tg->bio_lists[rw], bio); | |
698 | /* Take a bio reference on tg */ | |
1adaf3dd | 699 | blkg_get(tg_to_blkg(tg)); |
e43473b7 VG |
700 | tg->nr_queued[rw]++; |
701 | td->nr_queued[rw]++; | |
702 | throtl_enqueue_tg(td, tg); | |
703 | } | |
704 | ||
705 | static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg) | |
706 | { | |
707 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
708 | struct bio *bio; | |
709 | ||
710 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
0f3457f6 | 711 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 VG |
712 | |
713 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
0f3457f6 | 714 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
715 | |
716 | min_wait = min(read_wait, write_wait); | |
717 | disptime = jiffies + min_wait; | |
718 | ||
e43473b7 VG |
719 | /* Update dispatch time */ |
720 | throtl_dequeue_tg(td, tg); | |
721 | tg->disptime = disptime; | |
722 | throtl_enqueue_tg(td, tg); | |
723 | } | |
724 | ||
0f3457f6 TH |
725 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw, |
726 | struct bio_list *bl) | |
e43473b7 VG |
727 | { |
728 | struct bio *bio; | |
729 | ||
730 | bio = bio_list_pop(&tg->bio_lists[rw]); | |
731 | tg->nr_queued[rw]--; | |
1adaf3dd TH |
732 | /* Drop bio reference on blkg */ |
733 | blkg_put(tg_to_blkg(tg)); | |
e43473b7 | 734 | |
0f3457f6 TH |
735 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
736 | tg->td->nr_queued[rw]--; | |
e43473b7 VG |
737 | |
738 | throtl_charge_bio(tg, bio); | |
739 | bio_list_add(bl, bio); | |
740 | bio->bi_rw |= REQ_THROTTLED; | |
741 | ||
0f3457f6 | 742 | throtl_trim_slice(tg, rw); |
e43473b7 VG |
743 | } |
744 | ||
0f3457f6 | 745 | static int throtl_dispatch_tg(struct throtl_grp *tg, struct bio_list *bl) |
e43473b7 VG |
746 | { |
747 | unsigned int nr_reads = 0, nr_writes = 0; | |
748 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 749 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
750 | struct bio *bio; |
751 | ||
752 | /* Try to dispatch 75% READS and 25% WRITES */ | |
753 | ||
0f3457f6 TH |
754 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) && |
755 | tg_may_dispatch(tg, bio, NULL)) { | |
e43473b7 | 756 | |
0f3457f6 | 757 | tg_dispatch_one_bio(tg, bio_data_dir(bio), bl); |
e43473b7 VG |
758 | nr_reads++; |
759 | ||
760 | if (nr_reads >= max_nr_reads) | |
761 | break; | |
762 | } | |
763 | ||
0f3457f6 TH |
764 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) && |
765 | tg_may_dispatch(tg, bio, NULL)) { | |
e43473b7 | 766 | |
0f3457f6 | 767 | tg_dispatch_one_bio(tg, bio_data_dir(bio), bl); |
e43473b7 VG |
768 | nr_writes++; |
769 | ||
770 | if (nr_writes >= max_nr_writes) | |
771 | break; | |
772 | } | |
773 | ||
774 | return nr_reads + nr_writes; | |
775 | } | |
776 | ||
777 | static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl) | |
778 | { | |
779 | unsigned int nr_disp = 0; | |
780 | struct throtl_grp *tg; | |
c9e0332e | 781 | struct throtl_service_queue *sq = &td->service_queue; |
e43473b7 VG |
782 | |
783 | while (1) { | |
c9e0332e | 784 | tg = throtl_rb_first(sq); |
e43473b7 VG |
785 | |
786 | if (!tg) | |
787 | break; | |
788 | ||
789 | if (time_before(jiffies, tg->disptime)) | |
790 | break; | |
791 | ||
792 | throtl_dequeue_tg(td, tg); | |
793 | ||
0f3457f6 | 794 | nr_disp += throtl_dispatch_tg(tg, bl); |
e43473b7 | 795 | |
2db6314c | 796 | if (tg->nr_queued[0] || tg->nr_queued[1]) |
e43473b7 | 797 | tg_update_disptime(td, tg); |
e43473b7 VG |
798 | |
799 | if (nr_disp >= throtl_quantum) | |
800 | break; | |
801 | } | |
802 | ||
803 | return nr_disp; | |
804 | } | |
805 | ||
cb76199c TH |
806 | /* work function to dispatch throttled bios */ |
807 | void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
e43473b7 | 808 | { |
cb76199c TH |
809 | struct throtl_data *td = container_of(to_delayed_work(work), |
810 | struct throtl_data, dispatch_work); | |
811 | struct request_queue *q = td->queue; | |
e43473b7 VG |
812 | unsigned int nr_disp = 0; |
813 | struct bio_list bio_list_on_stack; | |
814 | struct bio *bio; | |
69d60eb9 | 815 | struct blk_plug plug; |
e43473b7 VG |
816 | |
817 | spin_lock_irq(q->queue_lock); | |
818 | ||
e43473b7 VG |
819 | bio_list_init(&bio_list_on_stack); |
820 | ||
d2f31a5f | 821 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
6a525600 TH |
822 | td->nr_queued[READ] + td->nr_queued[WRITE], |
823 | td->nr_queued[READ], td->nr_queued[WRITE]); | |
e43473b7 VG |
824 | |
825 | nr_disp = throtl_select_dispatch(td, &bio_list_on_stack); | |
826 | ||
827 | if (nr_disp) | |
828 | throtl_log(td, "bios disp=%u", nr_disp); | |
829 | ||
830 | throtl_schedule_next_dispatch(td); | |
6a525600 | 831 | |
e43473b7 VG |
832 | spin_unlock_irq(q->queue_lock); |
833 | ||
834 | /* | |
835 | * If we dispatched some requests, unplug the queue to make sure | |
836 | * immediate dispatch | |
837 | */ | |
838 | if (nr_disp) { | |
69d60eb9 | 839 | blk_start_plug(&plug); |
e43473b7 VG |
840 | while((bio = bio_list_pop(&bio_list_on_stack))) |
841 | generic_make_request(bio); | |
69d60eb9 | 842 | blk_finish_plug(&plug); |
e43473b7 | 843 | } |
e43473b7 VG |
844 | } |
845 | ||
f95a04af TH |
846 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
847 | struct blkg_policy_data *pd, int off) | |
41b38b6d | 848 | { |
f95a04af | 849 | struct throtl_grp *tg = pd_to_tg(pd); |
41b38b6d TH |
850 | struct blkg_rwstat rwstat = { }, tmp; |
851 | int i, cpu; | |
852 | ||
853 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 854 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
855 | |
856 | tmp = blkg_rwstat_read((void *)sc + off); | |
857 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
858 | rwstat.cnt[i] += tmp.cnt[i]; | |
859 | } | |
860 | ||
f95a04af | 861 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
41b38b6d TH |
862 | } |
863 | ||
8a3d2615 TH |
864 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
865 | struct seq_file *sf) | |
41b38b6d | 866 | { |
3c798398 | 867 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
41b38b6d | 868 | |
3c798398 | 869 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 870 | cft->private, true); |
41b38b6d TH |
871 | return 0; |
872 | } | |
873 | ||
f95a04af TH |
874 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
875 | int off) | |
60c2bc2d | 876 | { |
f95a04af TH |
877 | struct throtl_grp *tg = pd_to_tg(pd); |
878 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 879 | |
af133ceb | 880 | if (v == -1) |
60c2bc2d | 881 | return 0; |
f95a04af | 882 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
883 | } |
884 | ||
f95a04af TH |
885 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
886 | int off) | |
e43473b7 | 887 | { |
f95a04af TH |
888 | struct throtl_grp *tg = pd_to_tg(pd); |
889 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 890 | |
af133ceb TH |
891 | if (v == -1) |
892 | return 0; | |
f95a04af | 893 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
894 | } |
895 | ||
af133ceb TH |
896 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
897 | struct seq_file *sf) | |
8e89d13f | 898 | { |
3c798398 TH |
899 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, |
900 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 901 | return 0; |
8e89d13f VG |
902 | } |
903 | ||
af133ceb TH |
904 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
905 | struct seq_file *sf) | |
8e89d13f | 906 | { |
3c798398 TH |
907 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, |
908 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 909 | return 0; |
60c2bc2d TH |
910 | } |
911 | ||
af133ceb TH |
912 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
913 | bool is_u64) | |
60c2bc2d | 914 | { |
3c798398 | 915 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
60c2bc2d | 916 | struct blkg_conf_ctx ctx; |
af133ceb | 917 | struct throtl_grp *tg; |
a2b1693b | 918 | struct throtl_data *td; |
60c2bc2d TH |
919 | int ret; |
920 | ||
3c798398 | 921 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
922 | if (ret) |
923 | return ret; | |
924 | ||
af133ceb | 925 | tg = blkg_to_tg(ctx.blkg); |
a2b1693b | 926 | td = ctx.blkg->q->td; |
af133ceb | 927 | |
a2b1693b TH |
928 | if (!ctx.v) |
929 | ctx.v = -1; | |
af133ceb | 930 | |
a2b1693b TH |
931 | if (is_u64) |
932 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
933 | else | |
934 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 935 | |
0f3457f6 | 936 | throtl_log_tg(tg, "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", |
632b4493 TH |
937 | tg->bps[READ], tg->bps[WRITE], |
938 | tg->iops[READ], tg->iops[WRITE]); | |
939 | ||
940 | /* | |
941 | * We're already holding queue_lock and know @tg is valid. Let's | |
942 | * apply the new config directly. | |
943 | * | |
944 | * Restart the slices for both READ and WRITES. It might happen | |
945 | * that a group's limit are dropped suddenly and we don't want to | |
946 | * account recently dispatched IO with new low rate. | |
947 | */ | |
0f3457f6 TH |
948 | throtl_start_new_slice(tg, 0); |
949 | throtl_start_new_slice(tg, 1); | |
632b4493 | 950 | |
5b2c16aa | 951 | if (tg->flags & THROTL_TG_PENDING) { |
632b4493 TH |
952 | tg_update_disptime(td, tg); |
953 | throtl_schedule_next_dispatch(td); | |
954 | } | |
60c2bc2d TH |
955 | |
956 | blkg_conf_finish(&ctx); | |
a2b1693b | 957 | return 0; |
8e89d13f VG |
958 | } |
959 | ||
af133ceb TH |
960 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
961 | const char *buf) | |
60c2bc2d | 962 | { |
af133ceb | 963 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
964 | } |
965 | ||
af133ceb TH |
966 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
967 | const char *buf) | |
60c2bc2d | 968 | { |
af133ceb | 969 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
970 | } |
971 | ||
972 | static struct cftype throtl_files[] = { | |
973 | { | |
974 | .name = "throttle.read_bps_device", | |
af133ceb TH |
975 | .private = offsetof(struct throtl_grp, bps[READ]), |
976 | .read_seq_string = tg_print_conf_u64, | |
977 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
978 | .max_write_len = 256, |
979 | }, | |
980 | { | |
981 | .name = "throttle.write_bps_device", | |
af133ceb TH |
982 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
983 | .read_seq_string = tg_print_conf_u64, | |
984 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
985 | .max_write_len = 256, |
986 | }, | |
987 | { | |
988 | .name = "throttle.read_iops_device", | |
af133ceb TH |
989 | .private = offsetof(struct throtl_grp, iops[READ]), |
990 | .read_seq_string = tg_print_conf_uint, | |
991 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
992 | .max_write_len = 256, |
993 | }, | |
994 | { | |
995 | .name = "throttle.write_iops_device", | |
af133ceb TH |
996 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
997 | .read_seq_string = tg_print_conf_uint, | |
998 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
999 | .max_write_len = 256, |
1000 | }, | |
1001 | { | |
1002 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1003 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1004 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1005 | }, |
1006 | { | |
1007 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1008 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1009 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1010 | }, |
1011 | { } /* terminate */ | |
1012 | }; | |
1013 | ||
da527770 | 1014 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1015 | { |
1016 | struct throtl_data *td = q->td; | |
1017 | ||
cb76199c | 1018 | cancel_delayed_work_sync(&td->dispatch_work); |
e43473b7 VG |
1019 | } |
1020 | ||
3c798398 | 1021 | static struct blkcg_policy blkcg_policy_throtl = { |
f9fcc2d3 TH |
1022 | .pd_size = sizeof(struct throtl_grp), |
1023 | .cftypes = throtl_files, | |
1024 | ||
1025 | .pd_init_fn = throtl_pd_init, | |
1026 | .pd_exit_fn = throtl_pd_exit, | |
1027 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 VG |
1028 | }; |
1029 | ||
bc16a4f9 | 1030 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1031 | { |
1032 | struct throtl_data *td = q->td; | |
1033 | struct throtl_grp *tg; | |
e43473b7 | 1034 | bool rw = bio_data_dir(bio), update_disptime = true; |
3c798398 | 1035 | struct blkcg *blkcg; |
bc16a4f9 | 1036 | bool throttled = false; |
e43473b7 VG |
1037 | |
1038 | if (bio->bi_rw & REQ_THROTTLED) { | |
1039 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1040 | goto out; |
e43473b7 VG |
1041 | } |
1042 | ||
af75cd3c VG |
1043 | /* |
1044 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1045 | * basic fields like stats and io rates. If a group has no rules, | |
1046 | * just update the dispatch stats in lockless manner and return. | |
1047 | */ | |
af75cd3c | 1048 | rcu_read_lock(); |
3c798398 | 1049 | blkcg = bio_blkcg(bio); |
cd1604fa | 1050 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1051 | if (tg) { |
af75cd3c | 1052 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1053 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1054 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1055 | goto out_unlock_rcu; |
af75cd3c VG |
1056 | } |
1057 | } | |
af75cd3c VG |
1058 | |
1059 | /* | |
1060 | * Either group has not been allocated yet or it is not an unlimited | |
1061 | * IO group | |
1062 | */ | |
e43473b7 | 1063 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1064 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1065 | if (unlikely(!tg)) |
1066 | goto out_unlock; | |
f469a7b4 | 1067 | |
e43473b7 VG |
1068 | if (tg->nr_queued[rw]) { |
1069 | /* | |
1070 | * There is already another bio queued in same dir. No | |
1071 | * need to update dispatch time. | |
1072 | */ | |
231d704b | 1073 | update_disptime = false; |
e43473b7 | 1074 | goto queue_bio; |
de701c74 | 1075 | |
e43473b7 VG |
1076 | } |
1077 | ||
1078 | /* Bio is with-in rate limit of group */ | |
0f3457f6 | 1079 | if (tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1080 | throtl_charge_bio(tg, bio); |
04521db0 VG |
1081 | |
1082 | /* | |
1083 | * We need to trim slice even when bios are not being queued | |
1084 | * otherwise it might happen that a bio is not queued for | |
1085 | * a long time and slice keeps on extending and trim is not | |
1086 | * called for a long time. Now if limits are reduced suddenly | |
1087 | * we take into account all the IO dispatched so far at new | |
1088 | * low rate and * newly queued IO gets a really long dispatch | |
1089 | * time. | |
1090 | * | |
1091 | * So keep on trimming slice even if bio is not queued. | |
1092 | */ | |
0f3457f6 | 1093 | throtl_trim_slice(tg, rw); |
bc16a4f9 | 1094 | goto out_unlock; |
e43473b7 VG |
1095 | } |
1096 | ||
1097 | queue_bio: | |
0f3457f6 | 1098 | throtl_log_tg(tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
8e89d13f VG |
1099 | " iodisp=%u iops=%u queued=%d/%d", |
1100 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1101 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1102 | tg->io_disp[rw], tg->iops[rw], |
e43473b7 VG |
1103 | tg->nr_queued[READ], tg->nr_queued[WRITE]); |
1104 | ||
671058fb | 1105 | bio_associate_current(bio); |
e43473b7 | 1106 | throtl_add_bio_tg(q->td, tg, bio); |
bc16a4f9 | 1107 | throttled = true; |
e43473b7 VG |
1108 | |
1109 | if (update_disptime) { | |
1110 | tg_update_disptime(td, tg); | |
1111 | throtl_schedule_next_dispatch(td); | |
1112 | } | |
1113 | ||
bc16a4f9 | 1114 | out_unlock: |
e43473b7 | 1115 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1116 | out_unlock_rcu: |
1117 | rcu_read_unlock(); | |
bc16a4f9 TH |
1118 | out: |
1119 | return throttled; | |
e43473b7 VG |
1120 | } |
1121 | ||
c9a929dd TH |
1122 | /** |
1123 | * blk_throtl_drain - drain throttled bios | |
1124 | * @q: request_queue to drain throttled bios for | |
1125 | * | |
1126 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1127 | */ | |
1128 | void blk_throtl_drain(struct request_queue *q) | |
1129 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1130 | { | |
1131 | struct throtl_data *td = q->td; | |
c9e0332e | 1132 | struct throtl_service_queue *sq = &td->service_queue; |
c9a929dd TH |
1133 | struct throtl_grp *tg; |
1134 | struct bio_list bl; | |
1135 | struct bio *bio; | |
1136 | ||
8bcb6c7d | 1137 | queue_lockdep_assert_held(q); |
c9a929dd TH |
1138 | |
1139 | bio_list_init(&bl); | |
1140 | ||
c9e0332e | 1141 | while ((tg = throtl_rb_first(sq))) { |
c9a929dd TH |
1142 | throtl_dequeue_tg(td, tg); |
1143 | ||
1144 | while ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
0f3457f6 | 1145 | tg_dispatch_one_bio(tg, bio_data_dir(bio), &bl); |
c9a929dd | 1146 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) |
0f3457f6 | 1147 | tg_dispatch_one_bio(tg, bio_data_dir(bio), &bl); |
c9a929dd TH |
1148 | } |
1149 | spin_unlock_irq(q->queue_lock); | |
1150 | ||
1151 | while ((bio = bio_list_pop(&bl))) | |
1152 | generic_make_request(bio); | |
1153 | ||
1154 | spin_lock_irq(q->queue_lock); | |
1155 | } | |
1156 | ||
e43473b7 VG |
1157 | int blk_throtl_init(struct request_queue *q) |
1158 | { | |
1159 | struct throtl_data *td; | |
a2b1693b | 1160 | int ret; |
e43473b7 VG |
1161 | |
1162 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1163 | if (!td) | |
1164 | return -ENOMEM; | |
1165 | ||
c9e0332e | 1166 | td->service_queue = THROTL_SERVICE_QUEUE_INITIALIZER; |
cb76199c | 1167 | INIT_DELAYED_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
e43473b7 | 1168 | |
cd1604fa | 1169 | q->td = td; |
29b12589 | 1170 | td->queue = q; |
02977e4a | 1171 | |
a2b1693b | 1172 | /* activate policy */ |
3c798398 | 1173 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1174 | if (ret) |
f51b802c | 1175 | kfree(td); |
a2b1693b | 1176 | return ret; |
e43473b7 VG |
1177 | } |
1178 | ||
1179 | void blk_throtl_exit(struct request_queue *q) | |
1180 | { | |
c875f4d0 | 1181 | BUG_ON(!q->td); |
da527770 | 1182 | throtl_shutdown_wq(q); |
3c798398 | 1183 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1184 | kfree(q->td); |
e43473b7 VG |
1185 | } |
1186 | ||
1187 | static int __init throtl_init(void) | |
1188 | { | |
450adcbe VG |
1189 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1190 | if (!kthrotld_workqueue) | |
1191 | panic("Failed to create kthrotld\n"); | |
1192 | ||
3c798398 | 1193 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1194 | } |
1195 | ||
1196 | module_init(throtl_init); |