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 */ | |
e43473b7 VG |
47 | }; |
48 | ||
5b2c16aa TH |
49 | enum tg_state_flags { |
50 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
0e9f4164 | 51 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ |
5b2c16aa TH |
52 | }; |
53 | ||
e43473b7 VG |
54 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
55 | ||
8a3d2615 TH |
56 | /* Per-cpu group stats */ |
57 | struct tg_stats_cpu { | |
58 | /* total bytes transferred */ | |
59 | struct blkg_rwstat service_bytes; | |
60 | /* total IOs serviced, post merge */ | |
61 | struct blkg_rwstat serviced; | |
62 | }; | |
63 | ||
e43473b7 | 64 | struct throtl_grp { |
f95a04af TH |
65 | /* must be the first member */ |
66 | struct blkg_policy_data pd; | |
67 | ||
c9e0332e | 68 | /* active throtl group service_queue member */ |
e43473b7 VG |
69 | struct rb_node rb_node; |
70 | ||
0f3457f6 TH |
71 | /* throtl_data this group belongs to */ |
72 | struct throtl_data *td; | |
73 | ||
49a2f1e3 TH |
74 | /* this group's service queue */ |
75 | struct throtl_service_queue service_queue; | |
76 | ||
e43473b7 VG |
77 | /* |
78 | * Dispatch time in jiffies. This is the estimated time when group | |
79 | * will unthrottle and is ready to dispatch more bio. It is used as | |
80 | * key to sort active groups in service tree. | |
81 | */ | |
82 | unsigned long disptime; | |
83 | ||
e43473b7 VG |
84 | unsigned int flags; |
85 | ||
e43473b7 VG |
86 | /* bytes per second rate limits */ |
87 | uint64_t bps[2]; | |
88 | ||
8e89d13f VG |
89 | /* IOPS limits */ |
90 | unsigned int iops[2]; | |
91 | ||
e43473b7 VG |
92 | /* Number of bytes disptached in current slice */ |
93 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
94 | /* Number of bio's dispatched in current slice */ |
95 | unsigned int io_disp[2]; | |
e43473b7 VG |
96 | |
97 | /* When did we start a new slice */ | |
98 | unsigned long slice_start[2]; | |
99 | unsigned long slice_end[2]; | |
fe071437 | 100 | |
8a3d2615 TH |
101 | /* Per cpu stats pointer */ |
102 | struct tg_stats_cpu __percpu *stats_cpu; | |
103 | ||
104 | /* List of tgs waiting for per cpu stats memory to be allocated */ | |
105 | struct list_head stats_alloc_node; | |
e43473b7 VG |
106 | }; |
107 | ||
108 | struct throtl_data | |
109 | { | |
e43473b7 | 110 | /* service tree for active throtl groups */ |
c9e0332e | 111 | struct throtl_service_queue service_queue; |
e43473b7 | 112 | |
e43473b7 VG |
113 | struct request_queue *queue; |
114 | ||
115 | /* Total Number of queued bios on READ and WRITE lists */ | |
116 | unsigned int nr_queued[2]; | |
117 | ||
118 | /* | |
02977e4a | 119 | * number of total undestroyed groups |
e43473b7 VG |
120 | */ |
121 | unsigned int nr_undestroyed_grps; | |
122 | ||
123 | /* Work for dispatching throttled bios */ | |
cb76199c | 124 | struct delayed_work dispatch_work; |
e43473b7 VG |
125 | }; |
126 | ||
8a3d2615 TH |
127 | /* list and work item to allocate percpu group stats */ |
128 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
129 | static LIST_HEAD(tg_stats_alloc_list); | |
130 | ||
131 | static void tg_stats_alloc_fn(struct work_struct *); | |
132 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
133 | ||
f95a04af TH |
134 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
135 | { | |
136 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
137 | } | |
138 | ||
3c798398 | 139 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 140 | { |
f95a04af | 141 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
142 | } |
143 | ||
3c798398 | 144 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 145 | { |
f95a04af | 146 | return pd_to_blkg(&tg->pd); |
0381411e TH |
147 | } |
148 | ||
03d8e111 TH |
149 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) |
150 | { | |
151 | return blkg_to_tg(td->queue->root_blkg); | |
152 | } | |
153 | ||
0f3457f6 | 154 | #define throtl_log_tg(tg, fmt, args...) do { \ |
54e7ed12 TH |
155 | char __pbuf[128]; \ |
156 | \ | |
157 | blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \ | |
0f3457f6 | 158 | blk_add_trace_msg((tg)->td->queue, "throtl %s " fmt, __pbuf, ##args); \ |
54e7ed12 | 159 | } while (0) |
e43473b7 VG |
160 | |
161 | #define throtl_log(td, fmt, args...) \ | |
162 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) | |
163 | ||
8a3d2615 TH |
164 | /* |
165 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
3b07e9ca | 166 | * system_wq once there are some groups on the alloc_list waiting for |
8a3d2615 TH |
167 | * allocation. |
168 | */ | |
169 | static void tg_stats_alloc_fn(struct work_struct *work) | |
170 | { | |
171 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
172 | struct delayed_work *dwork = to_delayed_work(work); | |
173 | bool empty = false; | |
174 | ||
175 | alloc_stats: | |
176 | if (!stats_cpu) { | |
177 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
178 | if (!stats_cpu) { | |
179 | /* allocation failed, try again after some time */ | |
3b07e9ca | 180 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); |
8a3d2615 TH |
181 | return; |
182 | } | |
183 | } | |
184 | ||
185 | spin_lock_irq(&tg_stats_alloc_lock); | |
186 | ||
187 | if (!list_empty(&tg_stats_alloc_list)) { | |
188 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
189 | struct throtl_grp, | |
190 | stats_alloc_node); | |
191 | swap(tg->stats_cpu, stats_cpu); | |
192 | list_del_init(&tg->stats_alloc_node); | |
193 | } | |
194 | ||
195 | empty = list_empty(&tg_stats_alloc_list); | |
196 | spin_unlock_irq(&tg_stats_alloc_lock); | |
197 | if (!empty) | |
198 | goto alloc_stats; | |
199 | } | |
200 | ||
49a2f1e3 | 201 | /* init a service_queue, assumes the caller zeroed it */ |
77216b04 TH |
202 | static void throtl_service_queue_init(struct throtl_service_queue *sq, |
203 | struct throtl_service_queue *parent_sq) | |
49a2f1e3 | 204 | { |
73f0d49a TH |
205 | bio_list_init(&sq->bio_lists[0]); |
206 | bio_list_init(&sq->bio_lists[1]); | |
49a2f1e3 | 207 | sq->pending_tree = RB_ROOT; |
77216b04 | 208 | sq->parent_sq = parent_sq; |
49a2f1e3 TH |
209 | } |
210 | ||
3c798398 | 211 | static void throtl_pd_init(struct blkcg_gq *blkg) |
a29a171e | 212 | { |
0381411e | 213 | struct throtl_grp *tg = blkg_to_tg(blkg); |
77216b04 | 214 | struct throtl_data *td = blkg->q->td; |
ff26eaad | 215 | unsigned long flags; |
cd1604fa | 216 | |
77216b04 | 217 | throtl_service_queue_init(&tg->service_queue, &td->service_queue); |
a29a171e | 218 | RB_CLEAR_NODE(&tg->rb_node); |
77216b04 | 219 | tg->td = td; |
a29a171e | 220 | |
e56da7e2 TH |
221 | tg->bps[READ] = -1; |
222 | tg->bps[WRITE] = -1; | |
223 | tg->iops[READ] = -1; | |
224 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
225 | |
226 | /* | |
227 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
228 | * but percpu allocator can't be called from IO path. Queue tg on | |
229 | * tg_stats_alloc_list and allocate from work item. | |
230 | */ | |
ff26eaad | 231 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 232 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); |
3b07e9ca | 233 | schedule_delayed_work(&tg_stats_alloc_work, 0); |
ff26eaad | 234 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
235 | } |
236 | ||
3c798398 | 237 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
8a3d2615 TH |
238 | { |
239 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
ff26eaad | 240 | unsigned long flags; |
8a3d2615 | 241 | |
ff26eaad | 242 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 243 | list_del_init(&tg->stats_alloc_node); |
ff26eaad | 244 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
245 | |
246 | free_percpu(tg->stats_cpu); | |
247 | } | |
248 | ||
3c798398 | 249 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
8a3d2615 TH |
250 | { |
251 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
252 | int cpu; | |
253 | ||
254 | if (tg->stats_cpu == NULL) | |
255 | return; | |
256 | ||
257 | for_each_possible_cpu(cpu) { | |
258 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
259 | ||
260 | blkg_rwstat_reset(&sc->service_bytes); | |
261 | blkg_rwstat_reset(&sc->serviced); | |
262 | } | |
a29a171e VG |
263 | } |
264 | ||
3c798398 TH |
265 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
266 | struct blkcg *blkcg) | |
e43473b7 | 267 | { |
be2c6b19 | 268 | /* |
3c798398 TH |
269 | * This is the common case when there are no blkcgs. Avoid lookup |
270 | * in this case | |
cd1604fa | 271 | */ |
3c798398 | 272 | if (blkcg == &blkcg_root) |
03d8e111 | 273 | return td_root_tg(td); |
e43473b7 | 274 | |
e8989fae | 275 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
276 | } |
277 | ||
cd1604fa | 278 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
3c798398 | 279 | struct blkcg *blkcg) |
e43473b7 | 280 | { |
f469a7b4 | 281 | struct request_queue *q = td->queue; |
cd1604fa | 282 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 283 | |
f469a7b4 | 284 | /* |
3c798398 TH |
285 | * This is the common case when there are no blkcgs. Avoid lookup |
286 | * in this case | |
f469a7b4 | 287 | */ |
3c798398 | 288 | if (blkcg == &blkcg_root) { |
03d8e111 | 289 | tg = td_root_tg(td); |
cd1604fa | 290 | } else { |
3c798398 | 291 | struct blkcg_gq *blkg; |
f469a7b4 | 292 | |
3c96cb32 | 293 | blkg = blkg_lookup_create(blkcg, q); |
f469a7b4 | 294 | |
cd1604fa TH |
295 | /* if %NULL and @q is alive, fall back to root_tg */ |
296 | if (!IS_ERR(blkg)) | |
0381411e | 297 | tg = blkg_to_tg(blkg); |
3f3299d5 | 298 | else if (!blk_queue_dying(q)) |
03d8e111 | 299 | tg = td_root_tg(td); |
f469a7b4 VG |
300 | } |
301 | ||
e43473b7 VG |
302 | return tg; |
303 | } | |
304 | ||
0049af73 TH |
305 | static struct throtl_grp * |
306 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
307 | { |
308 | /* Service tree is empty */ | |
0049af73 | 309 | if (!parent_sq->nr_pending) |
e43473b7 VG |
310 | return NULL; |
311 | ||
0049af73 TH |
312 | if (!parent_sq->first_pending) |
313 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
e43473b7 | 314 | |
0049af73 TH |
315 | if (parent_sq->first_pending) |
316 | return rb_entry_tg(parent_sq->first_pending); | |
e43473b7 VG |
317 | |
318 | return NULL; | |
319 | } | |
320 | ||
321 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
322 | { | |
323 | rb_erase(n, root); | |
324 | RB_CLEAR_NODE(n); | |
325 | } | |
326 | ||
0049af73 TH |
327 | static void throtl_rb_erase(struct rb_node *n, |
328 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 329 | { |
0049af73 TH |
330 | if (parent_sq->first_pending == n) |
331 | parent_sq->first_pending = NULL; | |
332 | rb_erase_init(n, &parent_sq->pending_tree); | |
333 | --parent_sq->nr_pending; | |
e43473b7 VG |
334 | } |
335 | ||
0049af73 | 336 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
337 | { |
338 | struct throtl_grp *tg; | |
339 | ||
0049af73 | 340 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
341 | if (!tg) |
342 | return; | |
343 | ||
0049af73 | 344 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
345 | } |
346 | ||
77216b04 | 347 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 348 | { |
77216b04 | 349 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
0049af73 | 350 | struct rb_node **node = &parent_sq->pending_tree.rb_node; |
e43473b7 VG |
351 | struct rb_node *parent = NULL; |
352 | struct throtl_grp *__tg; | |
353 | unsigned long key = tg->disptime; | |
354 | int left = 1; | |
355 | ||
356 | while (*node != NULL) { | |
357 | parent = *node; | |
358 | __tg = rb_entry_tg(parent); | |
359 | ||
360 | if (time_before(key, __tg->disptime)) | |
361 | node = &parent->rb_left; | |
362 | else { | |
363 | node = &parent->rb_right; | |
364 | left = 0; | |
365 | } | |
366 | } | |
367 | ||
368 | if (left) | |
0049af73 | 369 | parent_sq->first_pending = &tg->rb_node; |
e43473b7 VG |
370 | |
371 | rb_link_node(&tg->rb_node, parent, node); | |
0049af73 | 372 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); |
e43473b7 VG |
373 | } |
374 | ||
77216b04 | 375 | static void __throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 376 | { |
77216b04 | 377 | tg_service_queue_add(tg); |
5b2c16aa | 378 | tg->flags |= THROTL_TG_PENDING; |
77216b04 | 379 | tg->service_queue.parent_sq->nr_pending++; |
e43473b7 VG |
380 | } |
381 | ||
77216b04 | 382 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 383 | { |
5b2c16aa | 384 | if (!(tg->flags & THROTL_TG_PENDING)) |
77216b04 | 385 | __throtl_enqueue_tg(tg); |
e43473b7 VG |
386 | } |
387 | ||
77216b04 | 388 | static void __throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 389 | { |
77216b04 | 390 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
5b2c16aa | 391 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
392 | } |
393 | ||
77216b04 | 394 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 395 | { |
5b2c16aa | 396 | if (tg->flags & THROTL_TG_PENDING) |
77216b04 | 397 | __throtl_dequeue_tg(tg); |
e43473b7 VG |
398 | } |
399 | ||
a9131a27 TH |
400 | /* Call with queue lock held */ |
401 | static void throtl_schedule_delayed_work(struct throtl_data *td, | |
402 | unsigned long delay) | |
403 | { | |
404 | struct delayed_work *dwork = &td->dispatch_work; | |
405 | ||
6a525600 TH |
406 | mod_delayed_work(kthrotld_workqueue, dwork, delay); |
407 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", delay, jiffies); | |
a9131a27 TH |
408 | } |
409 | ||
e43473b7 VG |
410 | static void throtl_schedule_next_dispatch(struct throtl_data *td) |
411 | { | |
c9e0332e | 412 | struct throtl_service_queue *sq = &td->service_queue; |
e43473b7 | 413 | |
6a525600 | 414 | /* any pending children left? */ |
c9e0332e | 415 | if (!sq->nr_pending) |
e43473b7 VG |
416 | return; |
417 | ||
c9e0332e | 418 | update_min_dispatch_time(sq); |
e43473b7 | 419 | |
c9e0332e | 420 | if (time_before_eq(sq->first_pending_disptime, jiffies)) |
450adcbe | 421 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 422 | else |
c9e0332e | 423 | throtl_schedule_delayed_work(td, sq->first_pending_disptime - jiffies); |
e43473b7 VG |
424 | } |
425 | ||
0f3457f6 | 426 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
427 | { |
428 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 429 | tg->io_disp[rw] = 0; |
e43473b7 VG |
430 | tg->slice_start[rw] = jiffies; |
431 | tg->slice_end[rw] = jiffies + throtl_slice; | |
0f3457f6 | 432 | throtl_log_tg(tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", |
e43473b7 VG |
433 | rw == READ ? 'R' : 'W', tg->slice_start[rw], |
434 | tg->slice_end[rw], jiffies); | |
435 | } | |
436 | ||
0f3457f6 TH |
437 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
438 | unsigned long jiffy_end) | |
d1ae8ffd VG |
439 | { |
440 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
441 | } | |
442 | ||
0f3457f6 TH |
443 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
444 | unsigned long jiffy_end) | |
e43473b7 VG |
445 | { |
446 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
0f3457f6 | 447 | throtl_log_tg(tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", |
e43473b7 VG |
448 | rw == READ ? 'R' : 'W', tg->slice_start[rw], |
449 | tg->slice_end[rw], jiffies); | |
450 | } | |
451 | ||
452 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 453 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
454 | { |
455 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
456 | return 0; | |
457 | ||
458 | return 1; | |
459 | } | |
460 | ||
461 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 462 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 463 | { |
3aad5d3e VG |
464 | unsigned long nr_slices, time_elapsed, io_trim; |
465 | u64 bytes_trim, tmp; | |
e43473b7 VG |
466 | |
467 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
468 | ||
469 | /* | |
470 | * If bps are unlimited (-1), then time slice don't get | |
471 | * renewed. Don't try to trim the slice if slice is used. A new | |
472 | * slice will start when appropriate. | |
473 | */ | |
0f3457f6 | 474 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
475 | return; |
476 | ||
d1ae8ffd VG |
477 | /* |
478 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
479 | * that initially cgroup limit was very low resulting in high | |
480 | * slice_end, but later limit was bumped up and bio was dispached | |
481 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
482 | * is bad because it does not allow new slice to start. | |
483 | */ | |
484 | ||
0f3457f6 | 485 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); |
d1ae8ffd | 486 | |
e43473b7 VG |
487 | time_elapsed = jiffies - tg->slice_start[rw]; |
488 | ||
489 | nr_slices = time_elapsed / throtl_slice; | |
490 | ||
491 | if (!nr_slices) | |
492 | return; | |
3aad5d3e VG |
493 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
494 | do_div(tmp, HZ); | |
495 | bytes_trim = tmp; | |
e43473b7 | 496 | |
8e89d13f | 497 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 498 | |
8e89d13f | 499 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
500 | return; |
501 | ||
502 | if (tg->bytes_disp[rw] >= bytes_trim) | |
503 | tg->bytes_disp[rw] -= bytes_trim; | |
504 | else | |
505 | tg->bytes_disp[rw] = 0; | |
506 | ||
8e89d13f VG |
507 | if (tg->io_disp[rw] >= io_trim) |
508 | tg->io_disp[rw] -= io_trim; | |
509 | else | |
510 | tg->io_disp[rw] = 0; | |
511 | ||
e43473b7 VG |
512 | tg->slice_start[rw] += nr_slices * throtl_slice; |
513 | ||
0f3457f6 | 514 | throtl_log_tg(tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 515 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 516 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
517 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
518 | } | |
519 | ||
0f3457f6 TH |
520 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
521 | unsigned long *wait) | |
e43473b7 VG |
522 | { |
523 | bool rw = bio_data_dir(bio); | |
8e89d13f | 524 | unsigned int io_allowed; |
e43473b7 | 525 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 526 | u64 tmp; |
e43473b7 | 527 | |
8e89d13f | 528 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 529 | |
8e89d13f VG |
530 | /* Slice has just started. Consider one slice interval */ |
531 | if (!jiffy_elapsed) | |
532 | jiffy_elapsed_rnd = throtl_slice; | |
533 | ||
534 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
535 | ||
c49c06e4 VG |
536 | /* |
537 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
538 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
539 | * will allow dispatch after 1 second and after that slice should | |
540 | * have been trimmed. | |
541 | */ | |
542 | ||
543 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
544 | do_div(tmp, HZ); | |
545 | ||
546 | if (tmp > UINT_MAX) | |
547 | io_allowed = UINT_MAX; | |
548 | else | |
549 | io_allowed = tmp; | |
8e89d13f VG |
550 | |
551 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
552 | if (wait) |
553 | *wait = 0; | |
554 | return 1; | |
555 | } | |
556 | ||
8e89d13f VG |
557 | /* Calc approx time to dispatch */ |
558 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
559 | ||
560 | if (jiffy_wait > jiffy_elapsed) | |
561 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
562 | else | |
563 | jiffy_wait = 1; | |
564 | ||
565 | if (wait) | |
566 | *wait = jiffy_wait; | |
567 | return 0; | |
568 | } | |
569 | ||
0f3457f6 TH |
570 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
571 | unsigned long *wait) | |
8e89d13f VG |
572 | { |
573 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 574 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 575 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
576 | |
577 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
578 | ||
579 | /* Slice has just started. Consider one slice interval */ | |
580 | if (!jiffy_elapsed) | |
581 | jiffy_elapsed_rnd = throtl_slice; | |
582 | ||
583 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
584 | ||
5e901a2b VG |
585 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
586 | do_div(tmp, HZ); | |
3aad5d3e | 587 | bytes_allowed = tmp; |
e43473b7 VG |
588 | |
589 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
590 | if (wait) | |
591 | *wait = 0; | |
592 | return 1; | |
593 | } | |
594 | ||
595 | /* Calc approx time to dispatch */ | |
596 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
597 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
598 | ||
599 | if (!jiffy_wait) | |
600 | jiffy_wait = 1; | |
601 | ||
602 | /* | |
603 | * This wait time is without taking into consideration the rounding | |
604 | * up we did. Add that time also. | |
605 | */ | |
606 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
607 | if (wait) |
608 | *wait = jiffy_wait; | |
8e89d13f VG |
609 | return 0; |
610 | } | |
611 | ||
af75cd3c VG |
612 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
613 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
614 | return 1; | |
615 | return 0; | |
616 | } | |
617 | ||
8e89d13f VG |
618 | /* |
619 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
620 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
621 | */ | |
0f3457f6 TH |
622 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
623 | unsigned long *wait) | |
8e89d13f VG |
624 | { |
625 | bool rw = bio_data_dir(bio); | |
626 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
627 | ||
628 | /* | |
629 | * Currently whole state machine of group depends on first bio | |
630 | * queued in the group bio list. So one should not be calling | |
631 | * this function with a different bio if there are other bios | |
632 | * queued. | |
633 | */ | |
73f0d49a TH |
634 | BUG_ON(tg->service_queue.nr_queued[rw] && |
635 | bio != bio_list_peek(&tg->service_queue.bio_lists[rw])); | |
e43473b7 | 636 | |
8e89d13f VG |
637 | /* If tg->bps = -1, then BW is unlimited */ |
638 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
639 | if (wait) | |
640 | *wait = 0; | |
641 | return 1; | |
642 | } | |
643 | ||
644 | /* | |
645 | * If previous slice expired, start a new one otherwise renew/extend | |
646 | * existing slice to make sure it is at least throtl_slice interval | |
647 | * long since now. | |
648 | */ | |
0f3457f6 TH |
649 | if (throtl_slice_used(tg, rw)) |
650 | throtl_start_new_slice(tg, rw); | |
8e89d13f VG |
651 | else { |
652 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
0f3457f6 | 653 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); |
8e89d13f VG |
654 | } |
655 | ||
0f3457f6 TH |
656 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
657 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
658 | if (wait) |
659 | *wait = 0; | |
660 | return 1; | |
661 | } | |
662 | ||
663 | max_wait = max(bps_wait, iops_wait); | |
664 | ||
665 | if (wait) | |
666 | *wait = max_wait; | |
667 | ||
668 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 669 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
670 | |
671 | return 0; | |
672 | } | |
673 | ||
3c798398 | 674 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
675 | int rw) |
676 | { | |
8a3d2615 TH |
677 | struct throtl_grp *tg = blkg_to_tg(blkg); |
678 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
679 | unsigned long flags; |
680 | ||
681 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 682 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
683 | return; |
684 | ||
685 | /* | |
686 | * Disabling interrupts to provide mutual exclusion between two | |
687 | * writes on same cpu. It probably is not needed for 64bit. Not | |
688 | * optimizing that case yet. | |
689 | */ | |
690 | local_irq_save(flags); | |
691 | ||
8a3d2615 | 692 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 693 | |
629ed0b1 TH |
694 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
695 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
696 | ||
697 | local_irq_restore(flags); | |
698 | } | |
699 | ||
e43473b7 VG |
700 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
701 | { | |
702 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
703 | |
704 | /* Charge the bio to the group */ | |
705 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 706 | tg->io_disp[rw]++; |
e43473b7 | 707 | |
629ed0b1 | 708 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw); |
e43473b7 VG |
709 | } |
710 | ||
77216b04 | 711 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_grp *tg) |
e43473b7 | 712 | { |
73f0d49a | 713 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
714 | bool rw = bio_data_dir(bio); |
715 | ||
0e9f4164 TH |
716 | /* |
717 | * If @tg doesn't currently have any bios queued in the same | |
718 | * direction, queueing @bio can change when @tg should be | |
719 | * dispatched. Mark that @tg was empty. This is automatically | |
720 | * cleaered on the next tg_update_disptime(). | |
721 | */ | |
722 | if (!sq->nr_queued[rw]) | |
723 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
724 | ||
73f0d49a | 725 | bio_list_add(&sq->bio_lists[rw], bio); |
e43473b7 | 726 | /* Take a bio reference on tg */ |
1adaf3dd | 727 | blkg_get(tg_to_blkg(tg)); |
73f0d49a | 728 | sq->nr_queued[rw]++; |
e2d57e60 | 729 | tg->td->nr_queued[rw]++; |
77216b04 | 730 | throtl_enqueue_tg(tg); |
e43473b7 VG |
731 | } |
732 | ||
77216b04 | 733 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 734 | { |
73f0d49a | 735 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
736 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
737 | struct bio *bio; | |
738 | ||
73f0d49a | 739 | if ((bio = bio_list_peek(&sq->bio_lists[READ]))) |
0f3457f6 | 740 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 741 | |
73f0d49a | 742 | if ((bio = bio_list_peek(&sq->bio_lists[WRITE]))) |
0f3457f6 | 743 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
744 | |
745 | min_wait = min(read_wait, write_wait); | |
746 | disptime = jiffies + min_wait; | |
747 | ||
e43473b7 | 748 | /* Update dispatch time */ |
77216b04 | 749 | throtl_dequeue_tg(tg); |
e43473b7 | 750 | tg->disptime = disptime; |
77216b04 | 751 | throtl_enqueue_tg(tg); |
0e9f4164 TH |
752 | |
753 | /* see throtl_add_bio_tg() */ | |
754 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
755 | } |
756 | ||
77216b04 | 757 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 758 | { |
73f0d49a | 759 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
760 | struct bio *bio; |
761 | ||
73f0d49a TH |
762 | bio = bio_list_pop(&sq->bio_lists[rw]); |
763 | sq->nr_queued[rw]--; | |
1adaf3dd TH |
764 | /* Drop bio reference on blkg */ |
765 | blkg_put(tg_to_blkg(tg)); | |
e43473b7 | 766 | |
0f3457f6 TH |
767 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
768 | tg->td->nr_queued[rw]--; | |
e43473b7 VG |
769 | |
770 | throtl_charge_bio(tg, bio); | |
77216b04 | 771 | bio_list_add(&sq->parent_sq->bio_lists[rw], bio); |
e43473b7 VG |
772 | bio->bi_rw |= REQ_THROTTLED; |
773 | ||
0f3457f6 | 774 | throtl_trim_slice(tg, rw); |
e43473b7 VG |
775 | } |
776 | ||
77216b04 | 777 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 778 | { |
73f0d49a | 779 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
780 | unsigned int nr_reads = 0, nr_writes = 0; |
781 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 782 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
783 | struct bio *bio; |
784 | ||
785 | /* Try to dispatch 75% READS and 25% WRITES */ | |
786 | ||
73f0d49a | 787 | while ((bio = bio_list_peek(&sq->bio_lists[READ])) && |
0f3457f6 | 788 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 789 | |
77216b04 | 790 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
791 | nr_reads++; |
792 | ||
793 | if (nr_reads >= max_nr_reads) | |
794 | break; | |
795 | } | |
796 | ||
73f0d49a | 797 | while ((bio = bio_list_peek(&sq->bio_lists[WRITE])) && |
0f3457f6 | 798 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 799 | |
77216b04 | 800 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
801 | nr_writes++; |
802 | ||
803 | if (nr_writes >= max_nr_writes) | |
804 | break; | |
805 | } | |
806 | ||
807 | return nr_reads + nr_writes; | |
808 | } | |
809 | ||
651930bc | 810 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
811 | { |
812 | unsigned int nr_disp = 0; | |
e43473b7 VG |
813 | |
814 | while (1) { | |
73f0d49a TH |
815 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
816 | struct throtl_service_queue *sq = &tg->service_queue; | |
e43473b7 VG |
817 | |
818 | if (!tg) | |
819 | break; | |
820 | ||
821 | if (time_before(jiffies, tg->disptime)) | |
822 | break; | |
823 | ||
77216b04 | 824 | throtl_dequeue_tg(tg); |
e43473b7 | 825 | |
77216b04 | 826 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 827 | |
73f0d49a | 828 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
77216b04 | 829 | tg_update_disptime(tg); |
e43473b7 VG |
830 | |
831 | if (nr_disp >= throtl_quantum) | |
832 | break; | |
833 | } | |
834 | ||
835 | return nr_disp; | |
836 | } | |
837 | ||
cb76199c TH |
838 | /* work function to dispatch throttled bios */ |
839 | void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
e43473b7 | 840 | { |
cb76199c TH |
841 | struct throtl_data *td = container_of(to_delayed_work(work), |
842 | struct throtl_data, dispatch_work); | |
651930bc | 843 | struct throtl_service_queue *sq = &td->service_queue; |
cb76199c | 844 | struct request_queue *q = td->queue; |
e43473b7 VG |
845 | unsigned int nr_disp = 0; |
846 | struct bio_list bio_list_on_stack; | |
847 | struct bio *bio; | |
69d60eb9 | 848 | struct blk_plug plug; |
651930bc | 849 | int rw; |
e43473b7 VG |
850 | |
851 | spin_lock_irq(q->queue_lock); | |
852 | ||
e43473b7 VG |
853 | bio_list_init(&bio_list_on_stack); |
854 | ||
d2f31a5f | 855 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
6a525600 TH |
856 | td->nr_queued[READ] + td->nr_queued[WRITE], |
857 | td->nr_queued[READ], td->nr_queued[WRITE]); | |
e43473b7 | 858 | |
651930bc | 859 | nr_disp = throtl_select_dispatch(sq); |
e43473b7 | 860 | |
651930bc TH |
861 | if (nr_disp) { |
862 | for (rw = READ; rw <= WRITE; rw++) { | |
863 | bio_list_merge(&bio_list_on_stack, &sq->bio_lists[rw]); | |
864 | bio_list_init(&sq->bio_lists[rw]); | |
865 | } | |
e43473b7 | 866 | throtl_log(td, "bios disp=%u", nr_disp); |
651930bc | 867 | } |
e43473b7 VG |
868 | |
869 | throtl_schedule_next_dispatch(td); | |
6a525600 | 870 | |
e43473b7 VG |
871 | spin_unlock_irq(q->queue_lock); |
872 | ||
873 | /* | |
874 | * If we dispatched some requests, unplug the queue to make sure | |
875 | * immediate dispatch | |
876 | */ | |
877 | if (nr_disp) { | |
69d60eb9 | 878 | blk_start_plug(&plug); |
e43473b7 VG |
879 | while((bio = bio_list_pop(&bio_list_on_stack))) |
880 | generic_make_request(bio); | |
69d60eb9 | 881 | blk_finish_plug(&plug); |
e43473b7 | 882 | } |
e43473b7 VG |
883 | } |
884 | ||
f95a04af TH |
885 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
886 | struct blkg_policy_data *pd, int off) | |
41b38b6d | 887 | { |
f95a04af | 888 | struct throtl_grp *tg = pd_to_tg(pd); |
41b38b6d TH |
889 | struct blkg_rwstat rwstat = { }, tmp; |
890 | int i, cpu; | |
891 | ||
892 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 893 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
894 | |
895 | tmp = blkg_rwstat_read((void *)sc + off); | |
896 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
897 | rwstat.cnt[i] += tmp.cnt[i]; | |
898 | } | |
899 | ||
f95a04af | 900 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
41b38b6d TH |
901 | } |
902 | ||
8a3d2615 TH |
903 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
904 | struct seq_file *sf) | |
41b38b6d | 905 | { |
3c798398 | 906 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
41b38b6d | 907 | |
3c798398 | 908 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 909 | cft->private, true); |
41b38b6d TH |
910 | return 0; |
911 | } | |
912 | ||
f95a04af TH |
913 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
914 | int off) | |
60c2bc2d | 915 | { |
f95a04af TH |
916 | struct throtl_grp *tg = pd_to_tg(pd); |
917 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 918 | |
af133ceb | 919 | if (v == -1) |
60c2bc2d | 920 | return 0; |
f95a04af | 921 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
922 | } |
923 | ||
f95a04af TH |
924 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
925 | int off) | |
e43473b7 | 926 | { |
f95a04af TH |
927 | struct throtl_grp *tg = pd_to_tg(pd); |
928 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 929 | |
af133ceb TH |
930 | if (v == -1) |
931 | return 0; | |
f95a04af | 932 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
933 | } |
934 | ||
af133ceb TH |
935 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
936 | struct seq_file *sf) | |
8e89d13f | 937 | { |
3c798398 TH |
938 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, |
939 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 940 | return 0; |
8e89d13f VG |
941 | } |
942 | ||
af133ceb TH |
943 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
944 | struct seq_file *sf) | |
8e89d13f | 945 | { |
3c798398 TH |
946 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, |
947 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 948 | return 0; |
60c2bc2d TH |
949 | } |
950 | ||
af133ceb TH |
951 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
952 | bool is_u64) | |
60c2bc2d | 953 | { |
3c798398 | 954 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
60c2bc2d | 955 | struct blkg_conf_ctx ctx; |
af133ceb | 956 | struct throtl_grp *tg; |
a2b1693b | 957 | struct throtl_data *td; |
60c2bc2d TH |
958 | int ret; |
959 | ||
3c798398 | 960 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
961 | if (ret) |
962 | return ret; | |
963 | ||
af133ceb | 964 | tg = blkg_to_tg(ctx.blkg); |
a2b1693b | 965 | td = ctx.blkg->q->td; |
af133ceb | 966 | |
a2b1693b TH |
967 | if (!ctx.v) |
968 | ctx.v = -1; | |
af133ceb | 969 | |
a2b1693b TH |
970 | if (is_u64) |
971 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
972 | else | |
973 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 974 | |
0f3457f6 | 975 | throtl_log_tg(tg, "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", |
632b4493 TH |
976 | tg->bps[READ], tg->bps[WRITE], |
977 | tg->iops[READ], tg->iops[WRITE]); | |
978 | ||
979 | /* | |
980 | * We're already holding queue_lock and know @tg is valid. Let's | |
981 | * apply the new config directly. | |
982 | * | |
983 | * Restart the slices for both READ and WRITES. It might happen | |
984 | * that a group's limit are dropped suddenly and we don't want to | |
985 | * account recently dispatched IO with new low rate. | |
986 | */ | |
0f3457f6 TH |
987 | throtl_start_new_slice(tg, 0); |
988 | throtl_start_new_slice(tg, 1); | |
632b4493 | 989 | |
5b2c16aa | 990 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 991 | tg_update_disptime(tg); |
632b4493 TH |
992 | throtl_schedule_next_dispatch(td); |
993 | } | |
60c2bc2d TH |
994 | |
995 | blkg_conf_finish(&ctx); | |
a2b1693b | 996 | return 0; |
8e89d13f VG |
997 | } |
998 | ||
af133ceb TH |
999 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1000 | const char *buf) | |
60c2bc2d | 1001 | { |
af133ceb | 1002 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
1003 | } |
1004 | ||
af133ceb TH |
1005 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1006 | const char *buf) | |
60c2bc2d | 1007 | { |
af133ceb | 1008 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
1009 | } |
1010 | ||
1011 | static struct cftype throtl_files[] = { | |
1012 | { | |
1013 | .name = "throttle.read_bps_device", | |
af133ceb TH |
1014 | .private = offsetof(struct throtl_grp, bps[READ]), |
1015 | .read_seq_string = tg_print_conf_u64, | |
1016 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1017 | .max_write_len = 256, |
1018 | }, | |
1019 | { | |
1020 | .name = "throttle.write_bps_device", | |
af133ceb TH |
1021 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
1022 | .read_seq_string = tg_print_conf_u64, | |
1023 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1024 | .max_write_len = 256, |
1025 | }, | |
1026 | { | |
1027 | .name = "throttle.read_iops_device", | |
af133ceb TH |
1028 | .private = offsetof(struct throtl_grp, iops[READ]), |
1029 | .read_seq_string = tg_print_conf_uint, | |
1030 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1031 | .max_write_len = 256, |
1032 | }, | |
1033 | { | |
1034 | .name = "throttle.write_iops_device", | |
af133ceb TH |
1035 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
1036 | .read_seq_string = tg_print_conf_uint, | |
1037 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1038 | .max_write_len = 256, |
1039 | }, | |
1040 | { | |
1041 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1042 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1043 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1044 | }, |
1045 | { | |
1046 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1047 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1048 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1049 | }, |
1050 | { } /* terminate */ | |
1051 | }; | |
1052 | ||
da527770 | 1053 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1054 | { |
1055 | struct throtl_data *td = q->td; | |
1056 | ||
cb76199c | 1057 | cancel_delayed_work_sync(&td->dispatch_work); |
e43473b7 VG |
1058 | } |
1059 | ||
3c798398 | 1060 | static struct blkcg_policy blkcg_policy_throtl = { |
f9fcc2d3 TH |
1061 | .pd_size = sizeof(struct throtl_grp), |
1062 | .cftypes = throtl_files, | |
1063 | ||
1064 | .pd_init_fn = throtl_pd_init, | |
1065 | .pd_exit_fn = throtl_pd_exit, | |
1066 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 VG |
1067 | }; |
1068 | ||
bc16a4f9 | 1069 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1070 | { |
1071 | struct throtl_data *td = q->td; | |
1072 | struct throtl_grp *tg; | |
73f0d49a | 1073 | struct throtl_service_queue *sq; |
0e9f4164 | 1074 | bool rw = bio_data_dir(bio); |
3c798398 | 1075 | struct blkcg *blkcg; |
bc16a4f9 | 1076 | bool throttled = false; |
e43473b7 VG |
1077 | |
1078 | if (bio->bi_rw & REQ_THROTTLED) { | |
1079 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1080 | goto out; |
e43473b7 VG |
1081 | } |
1082 | ||
af75cd3c VG |
1083 | /* |
1084 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1085 | * basic fields like stats and io rates. If a group has no rules, | |
1086 | * just update the dispatch stats in lockless manner and return. | |
1087 | */ | |
af75cd3c | 1088 | rcu_read_lock(); |
3c798398 | 1089 | blkcg = bio_blkcg(bio); |
cd1604fa | 1090 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1091 | if (tg) { |
af75cd3c | 1092 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1093 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1094 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1095 | goto out_unlock_rcu; |
af75cd3c VG |
1096 | } |
1097 | } | |
af75cd3c VG |
1098 | |
1099 | /* | |
1100 | * Either group has not been allocated yet or it is not an unlimited | |
1101 | * IO group | |
1102 | */ | |
e43473b7 | 1103 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1104 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1105 | if (unlikely(!tg)) |
1106 | goto out_unlock; | |
f469a7b4 | 1107 | |
73f0d49a TH |
1108 | sq = &tg->service_queue; |
1109 | ||
0e9f4164 TH |
1110 | /* throtl is FIFO - if other bios are already queued, should queue */ |
1111 | if (sq->nr_queued[rw]) | |
e43473b7 | 1112 | goto queue_bio; |
de701c74 | 1113 | |
e43473b7 | 1114 | /* Bio is with-in rate limit of group */ |
0f3457f6 | 1115 | if (tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1116 | throtl_charge_bio(tg, bio); |
04521db0 VG |
1117 | |
1118 | /* | |
1119 | * We need to trim slice even when bios are not being queued | |
1120 | * otherwise it might happen that a bio is not queued for | |
1121 | * a long time and slice keeps on extending and trim is not | |
1122 | * called for a long time. Now if limits are reduced suddenly | |
1123 | * we take into account all the IO dispatched so far at new | |
1124 | * low rate and * newly queued IO gets a really long dispatch | |
1125 | * time. | |
1126 | * | |
1127 | * So keep on trimming slice even if bio is not queued. | |
1128 | */ | |
0f3457f6 | 1129 | throtl_trim_slice(tg, rw); |
bc16a4f9 | 1130 | goto out_unlock; |
e43473b7 VG |
1131 | } |
1132 | ||
1133 | queue_bio: | |
0f3457f6 | 1134 | throtl_log_tg(tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
8e89d13f VG |
1135 | " iodisp=%u iops=%u queued=%d/%d", |
1136 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1137 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1138 | tg->io_disp[rw], tg->iops[rw], |
73f0d49a | 1139 | sq->nr_queued[READ], sq->nr_queued[WRITE]); |
e43473b7 | 1140 | |
671058fb | 1141 | bio_associate_current(bio); |
77216b04 | 1142 | throtl_add_bio_tg(bio, tg); |
bc16a4f9 | 1143 | throttled = true; |
e43473b7 | 1144 | |
0e9f4164 TH |
1145 | /* update @tg's dispatch time if @tg was empty before @bio */ |
1146 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
77216b04 | 1147 | tg_update_disptime(tg); |
e43473b7 VG |
1148 | throtl_schedule_next_dispatch(td); |
1149 | } | |
1150 | ||
bc16a4f9 | 1151 | out_unlock: |
e43473b7 | 1152 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1153 | out_unlock_rcu: |
1154 | rcu_read_unlock(); | |
bc16a4f9 TH |
1155 | out: |
1156 | return throttled; | |
e43473b7 VG |
1157 | } |
1158 | ||
c9a929dd TH |
1159 | /** |
1160 | * blk_throtl_drain - drain throttled bios | |
1161 | * @q: request_queue to drain throttled bios for | |
1162 | * | |
1163 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1164 | */ | |
1165 | void blk_throtl_drain(struct request_queue *q) | |
1166 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1167 | { | |
1168 | struct throtl_data *td = q->td; | |
0049af73 | 1169 | struct throtl_service_queue *parent_sq = &td->service_queue; |
c9a929dd | 1170 | struct throtl_grp *tg; |
c9a929dd | 1171 | struct bio *bio; |
651930bc | 1172 | int rw; |
c9a929dd | 1173 | |
8bcb6c7d | 1174 | queue_lockdep_assert_held(q); |
c9a929dd | 1175 | |
0049af73 | 1176 | while ((tg = throtl_rb_first(parent_sq))) { |
73f0d49a TH |
1177 | struct throtl_service_queue *sq = &tg->service_queue; |
1178 | ||
77216b04 | 1179 | throtl_dequeue_tg(tg); |
c9a929dd | 1180 | |
73f0d49a | 1181 | while ((bio = bio_list_peek(&sq->bio_lists[READ]))) |
77216b04 | 1182 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
73f0d49a | 1183 | while ((bio = bio_list_peek(&sq->bio_lists[WRITE]))) |
77216b04 | 1184 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
c9a929dd TH |
1185 | } |
1186 | spin_unlock_irq(q->queue_lock); | |
1187 | ||
651930bc TH |
1188 | for (rw = READ; rw <= WRITE; rw++) |
1189 | while ((bio = bio_list_pop(&parent_sq->bio_lists[rw]))) | |
1190 | generic_make_request(bio); | |
c9a929dd TH |
1191 | |
1192 | spin_lock_irq(q->queue_lock); | |
1193 | } | |
1194 | ||
e43473b7 VG |
1195 | int blk_throtl_init(struct request_queue *q) |
1196 | { | |
1197 | struct throtl_data *td; | |
a2b1693b | 1198 | int ret; |
e43473b7 VG |
1199 | |
1200 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1201 | if (!td) | |
1202 | return -ENOMEM; | |
1203 | ||
cb76199c | 1204 | INIT_DELAYED_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
77216b04 | 1205 | throtl_service_queue_init(&td->service_queue, NULL); |
e43473b7 | 1206 | |
cd1604fa | 1207 | q->td = td; |
29b12589 | 1208 | td->queue = q; |
02977e4a | 1209 | |
a2b1693b | 1210 | /* activate policy */ |
3c798398 | 1211 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1212 | if (ret) |
f51b802c | 1213 | kfree(td); |
a2b1693b | 1214 | return ret; |
e43473b7 VG |
1215 | } |
1216 | ||
1217 | void blk_throtl_exit(struct request_queue *q) | |
1218 | { | |
c875f4d0 | 1219 | BUG_ON(!q->td); |
da527770 | 1220 | throtl_shutdown_wq(q); |
3c798398 | 1221 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1222 | kfree(q->td); |
e43473b7 VG |
1223 | } |
1224 | ||
1225 | static int __init throtl_init(void) | |
1226 | { | |
450adcbe VG |
1227 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1228 | if (!kthrotld_workqueue) | |
1229 | panic("Failed to create kthrotld\n"); | |
1230 | ||
3c798398 | 1231 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1232 | } |
1233 | ||
1234 | module_init(throtl_init); |