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 | ||
e2d57e60 TH |
355 | static void __throtl_enqueue_tg(struct throtl_service_queue *sq, |
356 | struct throtl_grp *tg) | |
e43473b7 | 357 | { |
c9e0332e | 358 | tg_service_queue_add(sq, tg); |
5b2c16aa | 359 | tg->flags |= THROTL_TG_PENDING; |
c9e0332e | 360 | sq->nr_pending++; |
e43473b7 VG |
361 | } |
362 | ||
e2d57e60 TH |
363 | static void throtl_enqueue_tg(struct throtl_service_queue *sq, |
364 | struct throtl_grp *tg) | |
e43473b7 | 365 | { |
5b2c16aa | 366 | if (!(tg->flags & THROTL_TG_PENDING)) |
e2d57e60 | 367 | __throtl_enqueue_tg(sq, tg); |
e43473b7 VG |
368 | } |
369 | ||
e2d57e60 TH |
370 | static void __throtl_dequeue_tg(struct throtl_service_queue *sq, |
371 | struct throtl_grp *tg) | |
e43473b7 | 372 | { |
e2d57e60 | 373 | throtl_rb_erase(&tg->rb_node, sq); |
5b2c16aa | 374 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
375 | } |
376 | ||
e2d57e60 TH |
377 | static void throtl_dequeue_tg(struct throtl_service_queue *sq, |
378 | struct throtl_grp *tg) | |
e43473b7 | 379 | { |
5b2c16aa | 380 | if (tg->flags & THROTL_TG_PENDING) |
e2d57e60 | 381 | __throtl_dequeue_tg(sq, tg); |
e43473b7 VG |
382 | } |
383 | ||
a9131a27 TH |
384 | /* Call with queue lock held */ |
385 | static void throtl_schedule_delayed_work(struct throtl_data *td, | |
386 | unsigned long delay) | |
387 | { | |
388 | struct delayed_work *dwork = &td->dispatch_work; | |
389 | ||
6a525600 TH |
390 | mod_delayed_work(kthrotld_workqueue, dwork, delay); |
391 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", delay, jiffies); | |
a9131a27 TH |
392 | } |
393 | ||
e43473b7 VG |
394 | static void throtl_schedule_next_dispatch(struct throtl_data *td) |
395 | { | |
c9e0332e | 396 | struct throtl_service_queue *sq = &td->service_queue; |
e43473b7 | 397 | |
6a525600 | 398 | /* any pending children left? */ |
c9e0332e | 399 | if (!sq->nr_pending) |
e43473b7 VG |
400 | return; |
401 | ||
c9e0332e | 402 | update_min_dispatch_time(sq); |
e43473b7 | 403 | |
c9e0332e | 404 | if (time_before_eq(sq->first_pending_disptime, jiffies)) |
450adcbe | 405 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 406 | else |
c9e0332e | 407 | throtl_schedule_delayed_work(td, sq->first_pending_disptime - jiffies); |
e43473b7 VG |
408 | } |
409 | ||
0f3457f6 | 410 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
411 | { |
412 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 413 | tg->io_disp[rw] = 0; |
e43473b7 VG |
414 | tg->slice_start[rw] = jiffies; |
415 | tg->slice_end[rw] = jiffies + throtl_slice; | |
0f3457f6 | 416 | throtl_log_tg(tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", |
e43473b7 VG |
417 | rw == READ ? 'R' : 'W', tg->slice_start[rw], |
418 | tg->slice_end[rw], jiffies); | |
419 | } | |
420 | ||
0f3457f6 TH |
421 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
422 | unsigned long jiffy_end) | |
d1ae8ffd VG |
423 | { |
424 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
425 | } | |
426 | ||
0f3457f6 TH |
427 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
428 | unsigned long jiffy_end) | |
e43473b7 VG |
429 | { |
430 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
0f3457f6 | 431 | throtl_log_tg(tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", |
e43473b7 VG |
432 | rw == READ ? 'R' : 'W', tg->slice_start[rw], |
433 | tg->slice_end[rw], jiffies); | |
434 | } | |
435 | ||
436 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 437 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
438 | { |
439 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
440 | return 0; | |
441 | ||
442 | return 1; | |
443 | } | |
444 | ||
445 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 446 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 447 | { |
3aad5d3e VG |
448 | unsigned long nr_slices, time_elapsed, io_trim; |
449 | u64 bytes_trim, tmp; | |
e43473b7 VG |
450 | |
451 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
452 | ||
453 | /* | |
454 | * If bps are unlimited (-1), then time slice don't get | |
455 | * renewed. Don't try to trim the slice if slice is used. A new | |
456 | * slice will start when appropriate. | |
457 | */ | |
0f3457f6 | 458 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
459 | return; |
460 | ||
d1ae8ffd VG |
461 | /* |
462 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
463 | * that initially cgroup limit was very low resulting in high | |
464 | * slice_end, but later limit was bumped up and bio was dispached | |
465 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
466 | * is bad because it does not allow new slice to start. | |
467 | */ | |
468 | ||
0f3457f6 | 469 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); |
d1ae8ffd | 470 | |
e43473b7 VG |
471 | time_elapsed = jiffies - tg->slice_start[rw]; |
472 | ||
473 | nr_slices = time_elapsed / throtl_slice; | |
474 | ||
475 | if (!nr_slices) | |
476 | return; | |
3aad5d3e VG |
477 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
478 | do_div(tmp, HZ); | |
479 | bytes_trim = tmp; | |
e43473b7 | 480 | |
8e89d13f | 481 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 482 | |
8e89d13f | 483 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
484 | return; |
485 | ||
486 | if (tg->bytes_disp[rw] >= bytes_trim) | |
487 | tg->bytes_disp[rw] -= bytes_trim; | |
488 | else | |
489 | tg->bytes_disp[rw] = 0; | |
490 | ||
8e89d13f VG |
491 | if (tg->io_disp[rw] >= io_trim) |
492 | tg->io_disp[rw] -= io_trim; | |
493 | else | |
494 | tg->io_disp[rw] = 0; | |
495 | ||
e43473b7 VG |
496 | tg->slice_start[rw] += nr_slices * throtl_slice; |
497 | ||
0f3457f6 | 498 | throtl_log_tg(tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 499 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 500 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
501 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
502 | } | |
503 | ||
0f3457f6 TH |
504 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
505 | unsigned long *wait) | |
e43473b7 VG |
506 | { |
507 | bool rw = bio_data_dir(bio); | |
8e89d13f | 508 | unsigned int io_allowed; |
e43473b7 | 509 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 510 | u64 tmp; |
e43473b7 | 511 | |
8e89d13f | 512 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 513 | |
8e89d13f VG |
514 | /* Slice has just started. Consider one slice interval */ |
515 | if (!jiffy_elapsed) | |
516 | jiffy_elapsed_rnd = throtl_slice; | |
517 | ||
518 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
519 | ||
c49c06e4 VG |
520 | /* |
521 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
522 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
523 | * will allow dispatch after 1 second and after that slice should | |
524 | * have been trimmed. | |
525 | */ | |
526 | ||
527 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
528 | do_div(tmp, HZ); | |
529 | ||
530 | if (tmp > UINT_MAX) | |
531 | io_allowed = UINT_MAX; | |
532 | else | |
533 | io_allowed = tmp; | |
8e89d13f VG |
534 | |
535 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
536 | if (wait) |
537 | *wait = 0; | |
538 | return 1; | |
539 | } | |
540 | ||
8e89d13f VG |
541 | /* Calc approx time to dispatch */ |
542 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
543 | ||
544 | if (jiffy_wait > jiffy_elapsed) | |
545 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
546 | else | |
547 | jiffy_wait = 1; | |
548 | ||
549 | if (wait) | |
550 | *wait = jiffy_wait; | |
551 | return 0; | |
552 | } | |
553 | ||
0f3457f6 TH |
554 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
555 | unsigned long *wait) | |
8e89d13f VG |
556 | { |
557 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 558 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 559 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
560 | |
561 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
562 | ||
563 | /* Slice has just started. Consider one slice interval */ | |
564 | if (!jiffy_elapsed) | |
565 | jiffy_elapsed_rnd = throtl_slice; | |
566 | ||
567 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
568 | ||
5e901a2b VG |
569 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
570 | do_div(tmp, HZ); | |
3aad5d3e | 571 | bytes_allowed = tmp; |
e43473b7 VG |
572 | |
573 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
574 | if (wait) | |
575 | *wait = 0; | |
576 | return 1; | |
577 | } | |
578 | ||
579 | /* Calc approx time to dispatch */ | |
580 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
581 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
582 | ||
583 | if (!jiffy_wait) | |
584 | jiffy_wait = 1; | |
585 | ||
586 | /* | |
587 | * This wait time is without taking into consideration the rounding | |
588 | * up we did. Add that time also. | |
589 | */ | |
590 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
591 | if (wait) |
592 | *wait = jiffy_wait; | |
8e89d13f VG |
593 | return 0; |
594 | } | |
595 | ||
af75cd3c VG |
596 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
597 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
598 | return 1; | |
599 | return 0; | |
600 | } | |
601 | ||
8e89d13f VG |
602 | /* |
603 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
604 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
605 | */ | |
0f3457f6 TH |
606 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
607 | unsigned long *wait) | |
8e89d13f VG |
608 | { |
609 | bool rw = bio_data_dir(bio); | |
610 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
611 | ||
612 | /* | |
613 | * Currently whole state machine of group depends on first bio | |
614 | * queued in the group bio list. So one should not be calling | |
615 | * this function with a different bio if there are other bios | |
616 | * queued. | |
617 | */ | |
618 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); | |
e43473b7 | 619 | |
8e89d13f VG |
620 | /* If tg->bps = -1, then BW is unlimited */ |
621 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
622 | if (wait) | |
623 | *wait = 0; | |
624 | return 1; | |
625 | } | |
626 | ||
627 | /* | |
628 | * If previous slice expired, start a new one otherwise renew/extend | |
629 | * existing slice to make sure it is at least throtl_slice interval | |
630 | * long since now. | |
631 | */ | |
0f3457f6 TH |
632 | if (throtl_slice_used(tg, rw)) |
633 | throtl_start_new_slice(tg, rw); | |
8e89d13f VG |
634 | else { |
635 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
0f3457f6 | 636 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); |
8e89d13f VG |
637 | } |
638 | ||
0f3457f6 TH |
639 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
640 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
641 | if (wait) |
642 | *wait = 0; | |
643 | return 1; | |
644 | } | |
645 | ||
646 | max_wait = max(bps_wait, iops_wait); | |
647 | ||
648 | if (wait) | |
649 | *wait = max_wait; | |
650 | ||
651 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 652 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
653 | |
654 | return 0; | |
655 | } | |
656 | ||
3c798398 | 657 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
658 | int rw) |
659 | { | |
8a3d2615 TH |
660 | struct throtl_grp *tg = blkg_to_tg(blkg); |
661 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
662 | unsigned long flags; |
663 | ||
664 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 665 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
666 | return; |
667 | ||
668 | /* | |
669 | * Disabling interrupts to provide mutual exclusion between two | |
670 | * writes on same cpu. It probably is not needed for 64bit. Not | |
671 | * optimizing that case yet. | |
672 | */ | |
673 | local_irq_save(flags); | |
674 | ||
8a3d2615 | 675 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 676 | |
629ed0b1 TH |
677 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
678 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
679 | ||
680 | local_irq_restore(flags); | |
681 | } | |
682 | ||
e43473b7 VG |
683 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
684 | { | |
685 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
686 | |
687 | /* Charge the bio to the group */ | |
688 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 689 | tg->io_disp[rw]++; |
e43473b7 | 690 | |
629ed0b1 | 691 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw); |
e43473b7 VG |
692 | } |
693 | ||
e2d57e60 TH |
694 | static void throtl_add_bio_tg(struct throtl_service_queue *sq, |
695 | struct throtl_grp *tg, struct bio *bio) | |
e43473b7 VG |
696 | { |
697 | bool rw = bio_data_dir(bio); | |
698 | ||
699 | bio_list_add(&tg->bio_lists[rw], bio); | |
700 | /* Take a bio reference on tg */ | |
1adaf3dd | 701 | blkg_get(tg_to_blkg(tg)); |
e43473b7 | 702 | tg->nr_queued[rw]++; |
e2d57e60 TH |
703 | tg->td->nr_queued[rw]++; |
704 | throtl_enqueue_tg(sq, tg); | |
e43473b7 VG |
705 | } |
706 | ||
e2d57e60 TH |
707 | static void tg_update_disptime(struct throtl_service_queue *sq, |
708 | struct throtl_grp *tg) | |
e43473b7 VG |
709 | { |
710 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
711 | struct bio *bio; | |
712 | ||
713 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
0f3457f6 | 714 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 VG |
715 | |
716 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
0f3457f6 | 717 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
718 | |
719 | min_wait = min(read_wait, write_wait); | |
720 | disptime = jiffies + min_wait; | |
721 | ||
e43473b7 | 722 | /* Update dispatch time */ |
e2d57e60 | 723 | throtl_dequeue_tg(sq, tg); |
e43473b7 | 724 | tg->disptime = disptime; |
e2d57e60 | 725 | throtl_enqueue_tg(sq, tg); |
e43473b7 VG |
726 | } |
727 | ||
0f3457f6 TH |
728 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw, |
729 | struct bio_list *bl) | |
e43473b7 VG |
730 | { |
731 | struct bio *bio; | |
732 | ||
733 | bio = bio_list_pop(&tg->bio_lists[rw]); | |
734 | tg->nr_queued[rw]--; | |
1adaf3dd TH |
735 | /* Drop bio reference on blkg */ |
736 | blkg_put(tg_to_blkg(tg)); | |
e43473b7 | 737 | |
0f3457f6 TH |
738 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
739 | tg->td->nr_queued[rw]--; | |
e43473b7 VG |
740 | |
741 | throtl_charge_bio(tg, bio); | |
742 | bio_list_add(bl, bio); | |
743 | bio->bi_rw |= REQ_THROTTLED; | |
744 | ||
0f3457f6 | 745 | throtl_trim_slice(tg, rw); |
e43473b7 VG |
746 | } |
747 | ||
0f3457f6 | 748 | static int throtl_dispatch_tg(struct throtl_grp *tg, struct bio_list *bl) |
e43473b7 VG |
749 | { |
750 | unsigned int nr_reads = 0, nr_writes = 0; | |
751 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 752 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
753 | struct bio *bio; |
754 | ||
755 | /* Try to dispatch 75% READS and 25% WRITES */ | |
756 | ||
0f3457f6 TH |
757 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) && |
758 | tg_may_dispatch(tg, bio, NULL)) { | |
e43473b7 | 759 | |
0f3457f6 | 760 | tg_dispatch_one_bio(tg, bio_data_dir(bio), bl); |
e43473b7 VG |
761 | nr_reads++; |
762 | ||
763 | if (nr_reads >= max_nr_reads) | |
764 | break; | |
765 | } | |
766 | ||
0f3457f6 TH |
767 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) && |
768 | tg_may_dispatch(tg, bio, NULL)) { | |
e43473b7 | 769 | |
0f3457f6 | 770 | tg_dispatch_one_bio(tg, bio_data_dir(bio), bl); |
e43473b7 VG |
771 | nr_writes++; |
772 | ||
773 | if (nr_writes >= max_nr_writes) | |
774 | break; | |
775 | } | |
776 | ||
777 | return nr_reads + nr_writes; | |
778 | } | |
779 | ||
e2d57e60 TH |
780 | static int throtl_select_dispatch(struct throtl_service_queue *sq, |
781 | struct bio_list *bl) | |
e43473b7 VG |
782 | { |
783 | unsigned int nr_disp = 0; | |
784 | struct throtl_grp *tg; | |
e43473b7 VG |
785 | |
786 | while (1) { | |
c9e0332e | 787 | tg = throtl_rb_first(sq); |
e43473b7 VG |
788 | |
789 | if (!tg) | |
790 | break; | |
791 | ||
792 | if (time_before(jiffies, tg->disptime)) | |
793 | break; | |
794 | ||
e2d57e60 | 795 | throtl_dequeue_tg(sq, tg); |
e43473b7 | 796 | |
0f3457f6 | 797 | nr_disp += throtl_dispatch_tg(tg, bl); |
e43473b7 | 798 | |
2db6314c | 799 | if (tg->nr_queued[0] || tg->nr_queued[1]) |
e2d57e60 | 800 | tg_update_disptime(sq, tg); |
e43473b7 VG |
801 | |
802 | if (nr_disp >= throtl_quantum) | |
803 | break; | |
804 | } | |
805 | ||
806 | return nr_disp; | |
807 | } | |
808 | ||
cb76199c TH |
809 | /* work function to dispatch throttled bios */ |
810 | void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
e43473b7 | 811 | { |
cb76199c TH |
812 | struct throtl_data *td = container_of(to_delayed_work(work), |
813 | struct throtl_data, dispatch_work); | |
814 | struct request_queue *q = td->queue; | |
e43473b7 VG |
815 | unsigned int nr_disp = 0; |
816 | struct bio_list bio_list_on_stack; | |
817 | struct bio *bio; | |
69d60eb9 | 818 | struct blk_plug plug; |
e43473b7 VG |
819 | |
820 | spin_lock_irq(q->queue_lock); | |
821 | ||
e43473b7 VG |
822 | bio_list_init(&bio_list_on_stack); |
823 | ||
d2f31a5f | 824 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
6a525600 TH |
825 | td->nr_queued[READ] + td->nr_queued[WRITE], |
826 | td->nr_queued[READ], td->nr_queued[WRITE]); | |
e43473b7 | 827 | |
e2d57e60 | 828 | nr_disp = throtl_select_dispatch(&td->service_queue, &bio_list_on_stack); |
e43473b7 VG |
829 | |
830 | if (nr_disp) | |
831 | throtl_log(td, "bios disp=%u", nr_disp); | |
832 | ||
833 | throtl_schedule_next_dispatch(td); | |
6a525600 | 834 | |
e43473b7 VG |
835 | spin_unlock_irq(q->queue_lock); |
836 | ||
837 | /* | |
838 | * If we dispatched some requests, unplug the queue to make sure | |
839 | * immediate dispatch | |
840 | */ | |
841 | if (nr_disp) { | |
69d60eb9 | 842 | blk_start_plug(&plug); |
e43473b7 VG |
843 | while((bio = bio_list_pop(&bio_list_on_stack))) |
844 | generic_make_request(bio); | |
69d60eb9 | 845 | blk_finish_plug(&plug); |
e43473b7 | 846 | } |
e43473b7 VG |
847 | } |
848 | ||
f95a04af TH |
849 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
850 | struct blkg_policy_data *pd, int off) | |
41b38b6d | 851 | { |
f95a04af | 852 | struct throtl_grp *tg = pd_to_tg(pd); |
41b38b6d TH |
853 | struct blkg_rwstat rwstat = { }, tmp; |
854 | int i, cpu; | |
855 | ||
856 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 857 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
858 | |
859 | tmp = blkg_rwstat_read((void *)sc + off); | |
860 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
861 | rwstat.cnt[i] += tmp.cnt[i]; | |
862 | } | |
863 | ||
f95a04af | 864 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
41b38b6d TH |
865 | } |
866 | ||
8a3d2615 TH |
867 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
868 | struct seq_file *sf) | |
41b38b6d | 869 | { |
3c798398 | 870 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
41b38b6d | 871 | |
3c798398 | 872 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 873 | cft->private, true); |
41b38b6d TH |
874 | return 0; |
875 | } | |
876 | ||
f95a04af TH |
877 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
878 | int off) | |
60c2bc2d | 879 | { |
f95a04af TH |
880 | struct throtl_grp *tg = pd_to_tg(pd); |
881 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 882 | |
af133ceb | 883 | if (v == -1) |
60c2bc2d | 884 | return 0; |
f95a04af | 885 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
886 | } |
887 | ||
f95a04af TH |
888 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
889 | int off) | |
e43473b7 | 890 | { |
f95a04af TH |
891 | struct throtl_grp *tg = pd_to_tg(pd); |
892 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 893 | |
af133ceb TH |
894 | if (v == -1) |
895 | return 0; | |
f95a04af | 896 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
897 | } |
898 | ||
af133ceb TH |
899 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
900 | struct seq_file *sf) | |
8e89d13f | 901 | { |
3c798398 TH |
902 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, |
903 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 904 | return 0; |
8e89d13f VG |
905 | } |
906 | ||
af133ceb TH |
907 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
908 | struct seq_file *sf) | |
8e89d13f | 909 | { |
3c798398 TH |
910 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, |
911 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 912 | return 0; |
60c2bc2d TH |
913 | } |
914 | ||
af133ceb TH |
915 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
916 | bool is_u64) | |
60c2bc2d | 917 | { |
3c798398 | 918 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
60c2bc2d | 919 | struct blkg_conf_ctx ctx; |
af133ceb | 920 | struct throtl_grp *tg; |
a2b1693b | 921 | struct throtl_data *td; |
60c2bc2d TH |
922 | int ret; |
923 | ||
3c798398 | 924 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
925 | if (ret) |
926 | return ret; | |
927 | ||
af133ceb | 928 | tg = blkg_to_tg(ctx.blkg); |
a2b1693b | 929 | td = ctx.blkg->q->td; |
af133ceb | 930 | |
a2b1693b TH |
931 | if (!ctx.v) |
932 | ctx.v = -1; | |
af133ceb | 933 | |
a2b1693b TH |
934 | if (is_u64) |
935 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
936 | else | |
937 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 938 | |
0f3457f6 | 939 | throtl_log_tg(tg, "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", |
632b4493 TH |
940 | tg->bps[READ], tg->bps[WRITE], |
941 | tg->iops[READ], tg->iops[WRITE]); | |
942 | ||
943 | /* | |
944 | * We're already holding queue_lock and know @tg is valid. Let's | |
945 | * apply the new config directly. | |
946 | * | |
947 | * Restart the slices for both READ and WRITES. It might happen | |
948 | * that a group's limit are dropped suddenly and we don't want to | |
949 | * account recently dispatched IO with new low rate. | |
950 | */ | |
0f3457f6 TH |
951 | throtl_start_new_slice(tg, 0); |
952 | throtl_start_new_slice(tg, 1); | |
632b4493 | 953 | |
5b2c16aa | 954 | if (tg->flags & THROTL_TG_PENDING) { |
e2d57e60 | 955 | tg_update_disptime(&td->service_queue, tg); |
632b4493 TH |
956 | throtl_schedule_next_dispatch(td); |
957 | } | |
60c2bc2d TH |
958 | |
959 | blkg_conf_finish(&ctx); | |
a2b1693b | 960 | return 0; |
8e89d13f VG |
961 | } |
962 | ||
af133ceb TH |
963 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
964 | const char *buf) | |
60c2bc2d | 965 | { |
af133ceb | 966 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
967 | } |
968 | ||
af133ceb TH |
969 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
970 | const char *buf) | |
60c2bc2d | 971 | { |
af133ceb | 972 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
973 | } |
974 | ||
975 | static struct cftype throtl_files[] = { | |
976 | { | |
977 | .name = "throttle.read_bps_device", | |
af133ceb TH |
978 | .private = offsetof(struct throtl_grp, bps[READ]), |
979 | .read_seq_string = tg_print_conf_u64, | |
980 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
981 | .max_write_len = 256, |
982 | }, | |
983 | { | |
984 | .name = "throttle.write_bps_device", | |
af133ceb TH |
985 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
986 | .read_seq_string = tg_print_conf_u64, | |
987 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
988 | .max_write_len = 256, |
989 | }, | |
990 | { | |
991 | .name = "throttle.read_iops_device", | |
af133ceb TH |
992 | .private = offsetof(struct throtl_grp, iops[READ]), |
993 | .read_seq_string = tg_print_conf_uint, | |
994 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
995 | .max_write_len = 256, |
996 | }, | |
997 | { | |
998 | .name = "throttle.write_iops_device", | |
af133ceb TH |
999 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
1000 | .read_seq_string = tg_print_conf_uint, | |
1001 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1002 | .max_write_len = 256, |
1003 | }, | |
1004 | { | |
1005 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1006 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1007 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1008 | }, |
1009 | { | |
1010 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1011 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1012 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1013 | }, |
1014 | { } /* terminate */ | |
1015 | }; | |
1016 | ||
da527770 | 1017 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1018 | { |
1019 | struct throtl_data *td = q->td; | |
1020 | ||
cb76199c | 1021 | cancel_delayed_work_sync(&td->dispatch_work); |
e43473b7 VG |
1022 | } |
1023 | ||
3c798398 | 1024 | static struct blkcg_policy blkcg_policy_throtl = { |
f9fcc2d3 TH |
1025 | .pd_size = sizeof(struct throtl_grp), |
1026 | .cftypes = throtl_files, | |
1027 | ||
1028 | .pd_init_fn = throtl_pd_init, | |
1029 | .pd_exit_fn = throtl_pd_exit, | |
1030 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 VG |
1031 | }; |
1032 | ||
bc16a4f9 | 1033 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1034 | { |
1035 | struct throtl_data *td = q->td; | |
1036 | struct throtl_grp *tg; | |
e43473b7 | 1037 | bool rw = bio_data_dir(bio), update_disptime = true; |
3c798398 | 1038 | struct blkcg *blkcg; |
bc16a4f9 | 1039 | bool throttled = false; |
e43473b7 VG |
1040 | |
1041 | if (bio->bi_rw & REQ_THROTTLED) { | |
1042 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1043 | goto out; |
e43473b7 VG |
1044 | } |
1045 | ||
af75cd3c VG |
1046 | /* |
1047 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1048 | * basic fields like stats and io rates. If a group has no rules, | |
1049 | * just update the dispatch stats in lockless manner and return. | |
1050 | */ | |
af75cd3c | 1051 | rcu_read_lock(); |
3c798398 | 1052 | blkcg = bio_blkcg(bio); |
cd1604fa | 1053 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1054 | if (tg) { |
af75cd3c | 1055 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1056 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1057 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1058 | goto out_unlock_rcu; |
af75cd3c VG |
1059 | } |
1060 | } | |
af75cd3c VG |
1061 | |
1062 | /* | |
1063 | * Either group has not been allocated yet or it is not an unlimited | |
1064 | * IO group | |
1065 | */ | |
e43473b7 | 1066 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1067 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1068 | if (unlikely(!tg)) |
1069 | goto out_unlock; | |
f469a7b4 | 1070 | |
e43473b7 VG |
1071 | if (tg->nr_queued[rw]) { |
1072 | /* | |
1073 | * There is already another bio queued in same dir. No | |
1074 | * need to update dispatch time. | |
1075 | */ | |
231d704b | 1076 | update_disptime = false; |
e43473b7 | 1077 | goto queue_bio; |
de701c74 | 1078 | |
e43473b7 VG |
1079 | } |
1080 | ||
1081 | /* Bio is with-in rate limit of group */ | |
0f3457f6 | 1082 | if (tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1083 | throtl_charge_bio(tg, bio); |
04521db0 VG |
1084 | |
1085 | /* | |
1086 | * We need to trim slice even when bios are not being queued | |
1087 | * otherwise it might happen that a bio is not queued for | |
1088 | * a long time and slice keeps on extending and trim is not | |
1089 | * called for a long time. Now if limits are reduced suddenly | |
1090 | * we take into account all the IO dispatched so far at new | |
1091 | * low rate and * newly queued IO gets a really long dispatch | |
1092 | * time. | |
1093 | * | |
1094 | * So keep on trimming slice even if bio is not queued. | |
1095 | */ | |
0f3457f6 | 1096 | throtl_trim_slice(tg, rw); |
bc16a4f9 | 1097 | goto out_unlock; |
e43473b7 VG |
1098 | } |
1099 | ||
1100 | queue_bio: | |
0f3457f6 | 1101 | throtl_log_tg(tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
8e89d13f VG |
1102 | " iodisp=%u iops=%u queued=%d/%d", |
1103 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1104 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1105 | tg->io_disp[rw], tg->iops[rw], |
e43473b7 VG |
1106 | tg->nr_queued[READ], tg->nr_queued[WRITE]); |
1107 | ||
671058fb | 1108 | bio_associate_current(bio); |
e2d57e60 | 1109 | throtl_add_bio_tg(&q->td->service_queue, tg, bio); |
bc16a4f9 | 1110 | throttled = true; |
e43473b7 VG |
1111 | |
1112 | if (update_disptime) { | |
e2d57e60 | 1113 | tg_update_disptime(&td->service_queue, tg); |
e43473b7 VG |
1114 | throtl_schedule_next_dispatch(td); |
1115 | } | |
1116 | ||
bc16a4f9 | 1117 | out_unlock: |
e43473b7 | 1118 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1119 | out_unlock_rcu: |
1120 | rcu_read_unlock(); | |
bc16a4f9 TH |
1121 | out: |
1122 | return throttled; | |
e43473b7 VG |
1123 | } |
1124 | ||
c9a929dd TH |
1125 | /** |
1126 | * blk_throtl_drain - drain throttled bios | |
1127 | * @q: request_queue to drain throttled bios for | |
1128 | * | |
1129 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1130 | */ | |
1131 | void blk_throtl_drain(struct request_queue *q) | |
1132 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1133 | { | |
1134 | struct throtl_data *td = q->td; | |
c9e0332e | 1135 | struct throtl_service_queue *sq = &td->service_queue; |
c9a929dd TH |
1136 | struct throtl_grp *tg; |
1137 | struct bio_list bl; | |
1138 | struct bio *bio; | |
1139 | ||
8bcb6c7d | 1140 | queue_lockdep_assert_held(q); |
c9a929dd TH |
1141 | |
1142 | bio_list_init(&bl); | |
1143 | ||
c9e0332e | 1144 | while ((tg = throtl_rb_first(sq))) { |
e2d57e60 | 1145 | throtl_dequeue_tg(sq, tg); |
c9a929dd TH |
1146 | |
1147 | while ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
0f3457f6 | 1148 | tg_dispatch_one_bio(tg, bio_data_dir(bio), &bl); |
c9a929dd | 1149 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) |
0f3457f6 | 1150 | tg_dispatch_one_bio(tg, bio_data_dir(bio), &bl); |
c9a929dd TH |
1151 | } |
1152 | spin_unlock_irq(q->queue_lock); | |
1153 | ||
1154 | while ((bio = bio_list_pop(&bl))) | |
1155 | generic_make_request(bio); | |
1156 | ||
1157 | spin_lock_irq(q->queue_lock); | |
1158 | } | |
1159 | ||
e43473b7 VG |
1160 | int blk_throtl_init(struct request_queue *q) |
1161 | { | |
1162 | struct throtl_data *td; | |
a2b1693b | 1163 | int ret; |
e43473b7 VG |
1164 | |
1165 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1166 | if (!td) | |
1167 | return -ENOMEM; | |
1168 | ||
c9e0332e | 1169 | td->service_queue = THROTL_SERVICE_QUEUE_INITIALIZER; |
cb76199c | 1170 | INIT_DELAYED_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
e43473b7 | 1171 | |
cd1604fa | 1172 | q->td = td; |
29b12589 | 1173 | td->queue = q; |
02977e4a | 1174 | |
a2b1693b | 1175 | /* activate policy */ |
3c798398 | 1176 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1177 | if (ret) |
f51b802c | 1178 | kfree(td); |
a2b1693b | 1179 | return ret; |
e43473b7 VG |
1180 | } |
1181 | ||
1182 | void blk_throtl_exit(struct request_queue *q) | |
1183 | { | |
c875f4d0 | 1184 | BUG_ON(!q->td); |
da527770 | 1185 | throtl_shutdown_wq(q); |
3c798398 | 1186 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1187 | kfree(q->td); |
e43473b7 VG |
1188 | } |
1189 | ||
1190 | static int __init throtl_init(void) | |
1191 | { | |
450adcbe VG |
1192 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1193 | if (!kthrotld_workqueue) | |
1194 | panic("Failed to create kthrotld\n"); | |
1195 | ||
3c798398 | 1196 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1197 | } |
1198 | ||
1199 | module_init(throtl_init); |