]>
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 { |
73f0d49a TH |
30 | /* |
31 | * Bios queued directly to this service_queue or dispatched from | |
32 | * children throtl_grp's. | |
33 | */ | |
34 | struct bio_list bio_lists[2]; /* queued bios [READ/WRITE] */ | |
35 | unsigned int nr_queued[2]; /* number of queued bios */ | |
36 | ||
37 | /* | |
38 | * RB tree of active children throtl_grp's, which are sorted by | |
39 | * their ->disptime. | |
40 | */ | |
c9e0332e TH |
41 | struct rb_root pending_tree; /* RB tree of active tgs */ |
42 | struct rb_node *first_pending; /* first node in the tree */ | |
43 | unsigned int nr_pending; /* # queued in the tree */ | |
44 | unsigned long first_pending_disptime; /* disptime of the first tg */ | |
e43473b7 VG |
45 | }; |
46 | ||
5b2c16aa TH |
47 | enum tg_state_flags { |
48 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
0e9f4164 | 49 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ |
5b2c16aa TH |
50 | }; |
51 | ||
e43473b7 VG |
52 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
53 | ||
8a3d2615 TH |
54 | /* Per-cpu group stats */ |
55 | struct tg_stats_cpu { | |
56 | /* total bytes transferred */ | |
57 | struct blkg_rwstat service_bytes; | |
58 | /* total IOs serviced, post merge */ | |
59 | struct blkg_rwstat serviced; | |
60 | }; | |
61 | ||
e43473b7 | 62 | struct throtl_grp { |
f95a04af TH |
63 | /* must be the first member */ |
64 | struct blkg_policy_data pd; | |
65 | ||
c9e0332e | 66 | /* active throtl group service_queue member */ |
e43473b7 VG |
67 | struct rb_node rb_node; |
68 | ||
0f3457f6 TH |
69 | /* throtl_data this group belongs to */ |
70 | struct throtl_data *td; | |
71 | ||
49a2f1e3 TH |
72 | /* this group's service queue */ |
73 | struct throtl_service_queue service_queue; | |
74 | ||
e43473b7 VG |
75 | /* |
76 | * Dispatch time in jiffies. This is the estimated time when group | |
77 | * will unthrottle and is ready to dispatch more bio. It is used as | |
78 | * key to sort active groups in service tree. | |
79 | */ | |
80 | unsigned long disptime; | |
81 | ||
e43473b7 VG |
82 | unsigned int flags; |
83 | ||
e43473b7 VG |
84 | /* bytes per second rate limits */ |
85 | uint64_t bps[2]; | |
86 | ||
8e89d13f VG |
87 | /* IOPS limits */ |
88 | unsigned int iops[2]; | |
89 | ||
e43473b7 VG |
90 | /* Number of bytes disptached in current slice */ |
91 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
92 | /* Number of bio's dispatched in current slice */ |
93 | unsigned int io_disp[2]; | |
e43473b7 VG |
94 | |
95 | /* When did we start a new slice */ | |
96 | unsigned long slice_start[2]; | |
97 | unsigned long slice_end[2]; | |
fe071437 | 98 | |
8a3d2615 TH |
99 | /* Per cpu stats pointer */ |
100 | struct tg_stats_cpu __percpu *stats_cpu; | |
101 | ||
102 | /* List of tgs waiting for per cpu stats memory to be allocated */ | |
103 | struct list_head stats_alloc_node; | |
e43473b7 VG |
104 | }; |
105 | ||
106 | struct throtl_data | |
107 | { | |
e43473b7 | 108 | /* service tree for active throtl groups */ |
c9e0332e | 109 | struct throtl_service_queue service_queue; |
e43473b7 | 110 | |
e43473b7 VG |
111 | struct request_queue *queue; |
112 | ||
113 | /* Total Number of queued bios on READ and WRITE lists */ | |
114 | unsigned int nr_queued[2]; | |
115 | ||
116 | /* | |
02977e4a | 117 | * number of total undestroyed groups |
e43473b7 VG |
118 | */ |
119 | unsigned int nr_undestroyed_grps; | |
120 | ||
121 | /* Work for dispatching throttled bios */ | |
cb76199c | 122 | struct delayed_work dispatch_work; |
e43473b7 VG |
123 | }; |
124 | ||
8a3d2615 TH |
125 | /* list and work item to allocate percpu group stats */ |
126 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
127 | static LIST_HEAD(tg_stats_alloc_list); | |
128 | ||
129 | static void tg_stats_alloc_fn(struct work_struct *); | |
130 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
131 | ||
f95a04af TH |
132 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
133 | { | |
134 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
135 | } | |
136 | ||
3c798398 | 137 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 138 | { |
f95a04af | 139 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
140 | } |
141 | ||
3c798398 | 142 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 143 | { |
f95a04af | 144 | return pd_to_blkg(&tg->pd); |
0381411e TH |
145 | } |
146 | ||
03d8e111 TH |
147 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) |
148 | { | |
149 | return blkg_to_tg(td->queue->root_blkg); | |
150 | } | |
151 | ||
0f3457f6 | 152 | #define throtl_log_tg(tg, fmt, args...) do { \ |
54e7ed12 TH |
153 | char __pbuf[128]; \ |
154 | \ | |
155 | blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \ | |
0f3457f6 | 156 | blk_add_trace_msg((tg)->td->queue, "throtl %s " fmt, __pbuf, ##args); \ |
54e7ed12 | 157 | } while (0) |
e43473b7 VG |
158 | |
159 | #define throtl_log(td, fmt, args...) \ | |
160 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) | |
161 | ||
8a3d2615 TH |
162 | /* |
163 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
3b07e9ca | 164 | * system_wq once there are some groups on the alloc_list waiting for |
8a3d2615 TH |
165 | * allocation. |
166 | */ | |
167 | static void tg_stats_alloc_fn(struct work_struct *work) | |
168 | { | |
169 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
170 | struct delayed_work *dwork = to_delayed_work(work); | |
171 | bool empty = false; | |
172 | ||
173 | alloc_stats: | |
174 | if (!stats_cpu) { | |
175 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
176 | if (!stats_cpu) { | |
177 | /* allocation failed, try again after some time */ | |
3b07e9ca | 178 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); |
8a3d2615 TH |
179 | return; |
180 | } | |
181 | } | |
182 | ||
183 | spin_lock_irq(&tg_stats_alloc_lock); | |
184 | ||
185 | if (!list_empty(&tg_stats_alloc_list)) { | |
186 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
187 | struct throtl_grp, | |
188 | stats_alloc_node); | |
189 | swap(tg->stats_cpu, stats_cpu); | |
190 | list_del_init(&tg->stats_alloc_node); | |
191 | } | |
192 | ||
193 | empty = list_empty(&tg_stats_alloc_list); | |
194 | spin_unlock_irq(&tg_stats_alloc_lock); | |
195 | if (!empty) | |
196 | goto alloc_stats; | |
197 | } | |
198 | ||
49a2f1e3 TH |
199 | /* init a service_queue, assumes the caller zeroed it */ |
200 | static void throtl_service_queue_init(struct throtl_service_queue *sq) | |
201 | { | |
73f0d49a TH |
202 | bio_list_init(&sq->bio_lists[0]); |
203 | bio_list_init(&sq->bio_lists[1]); | |
49a2f1e3 TH |
204 | sq->pending_tree = RB_ROOT; |
205 | } | |
206 | ||
3c798398 | 207 | static void throtl_pd_init(struct blkcg_gq *blkg) |
a29a171e | 208 | { |
0381411e | 209 | struct throtl_grp *tg = blkg_to_tg(blkg); |
ff26eaad | 210 | unsigned long flags; |
cd1604fa | 211 | |
49a2f1e3 | 212 | throtl_service_queue_init(&tg->service_queue); |
a29a171e | 213 | RB_CLEAR_NODE(&tg->rb_node); |
0f3457f6 | 214 | tg->td = blkg->q->td; |
a29a171e | 215 | |
e56da7e2 TH |
216 | tg->bps[READ] = -1; |
217 | tg->bps[WRITE] = -1; | |
218 | tg->iops[READ] = -1; | |
219 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
220 | |
221 | /* | |
222 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
223 | * but percpu allocator can't be called from IO path. Queue tg on | |
224 | * tg_stats_alloc_list and allocate from work item. | |
225 | */ | |
ff26eaad | 226 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 227 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); |
3b07e9ca | 228 | schedule_delayed_work(&tg_stats_alloc_work, 0); |
ff26eaad | 229 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
230 | } |
231 | ||
3c798398 | 232 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
8a3d2615 TH |
233 | { |
234 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
ff26eaad | 235 | unsigned long flags; |
8a3d2615 | 236 | |
ff26eaad | 237 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 238 | list_del_init(&tg->stats_alloc_node); |
ff26eaad | 239 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
240 | |
241 | free_percpu(tg->stats_cpu); | |
242 | } | |
243 | ||
3c798398 | 244 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
8a3d2615 TH |
245 | { |
246 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
247 | int cpu; | |
248 | ||
249 | if (tg->stats_cpu == NULL) | |
250 | return; | |
251 | ||
252 | for_each_possible_cpu(cpu) { | |
253 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
254 | ||
255 | blkg_rwstat_reset(&sc->service_bytes); | |
256 | blkg_rwstat_reset(&sc->serviced); | |
257 | } | |
a29a171e VG |
258 | } |
259 | ||
3c798398 TH |
260 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
261 | struct blkcg *blkcg) | |
e43473b7 | 262 | { |
be2c6b19 | 263 | /* |
3c798398 TH |
264 | * This is the common case when there are no blkcgs. Avoid lookup |
265 | * in this case | |
cd1604fa | 266 | */ |
3c798398 | 267 | if (blkcg == &blkcg_root) |
03d8e111 | 268 | return td_root_tg(td); |
e43473b7 | 269 | |
e8989fae | 270 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
271 | } |
272 | ||
cd1604fa | 273 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
3c798398 | 274 | struct blkcg *blkcg) |
e43473b7 | 275 | { |
f469a7b4 | 276 | struct request_queue *q = td->queue; |
cd1604fa | 277 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 278 | |
f469a7b4 | 279 | /* |
3c798398 TH |
280 | * This is the common case when there are no blkcgs. Avoid lookup |
281 | * in this case | |
f469a7b4 | 282 | */ |
3c798398 | 283 | if (blkcg == &blkcg_root) { |
03d8e111 | 284 | tg = td_root_tg(td); |
cd1604fa | 285 | } else { |
3c798398 | 286 | struct blkcg_gq *blkg; |
f469a7b4 | 287 | |
3c96cb32 | 288 | blkg = blkg_lookup_create(blkcg, q); |
f469a7b4 | 289 | |
cd1604fa TH |
290 | /* if %NULL and @q is alive, fall back to root_tg */ |
291 | if (!IS_ERR(blkg)) | |
0381411e | 292 | tg = blkg_to_tg(blkg); |
3f3299d5 | 293 | else if (!blk_queue_dying(q)) |
03d8e111 | 294 | tg = td_root_tg(td); |
f469a7b4 VG |
295 | } |
296 | ||
e43473b7 VG |
297 | return tg; |
298 | } | |
299 | ||
0049af73 TH |
300 | static struct throtl_grp * |
301 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
302 | { |
303 | /* Service tree is empty */ | |
0049af73 | 304 | if (!parent_sq->nr_pending) |
e43473b7 VG |
305 | return NULL; |
306 | ||
0049af73 TH |
307 | if (!parent_sq->first_pending) |
308 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
e43473b7 | 309 | |
0049af73 TH |
310 | if (parent_sq->first_pending) |
311 | return rb_entry_tg(parent_sq->first_pending); | |
e43473b7 VG |
312 | |
313 | return NULL; | |
314 | } | |
315 | ||
316 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
317 | { | |
318 | rb_erase(n, root); | |
319 | RB_CLEAR_NODE(n); | |
320 | } | |
321 | ||
0049af73 TH |
322 | static void throtl_rb_erase(struct rb_node *n, |
323 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 324 | { |
0049af73 TH |
325 | if (parent_sq->first_pending == n) |
326 | parent_sq->first_pending = NULL; | |
327 | rb_erase_init(n, &parent_sq->pending_tree); | |
328 | --parent_sq->nr_pending; | |
e43473b7 VG |
329 | } |
330 | ||
0049af73 | 331 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
332 | { |
333 | struct throtl_grp *tg; | |
334 | ||
0049af73 | 335 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
336 | if (!tg) |
337 | return; | |
338 | ||
0049af73 | 339 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
340 | } |
341 | ||
0049af73 TH |
342 | static void tg_service_queue_add(struct throtl_grp *tg, |
343 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 344 | { |
0049af73 | 345 | struct rb_node **node = &parent_sq->pending_tree.rb_node; |
e43473b7 VG |
346 | struct rb_node *parent = NULL; |
347 | struct throtl_grp *__tg; | |
348 | unsigned long key = tg->disptime; | |
349 | int left = 1; | |
350 | ||
351 | while (*node != NULL) { | |
352 | parent = *node; | |
353 | __tg = rb_entry_tg(parent); | |
354 | ||
355 | if (time_before(key, __tg->disptime)) | |
356 | node = &parent->rb_left; | |
357 | else { | |
358 | node = &parent->rb_right; | |
359 | left = 0; | |
360 | } | |
361 | } | |
362 | ||
363 | if (left) | |
0049af73 | 364 | parent_sq->first_pending = &tg->rb_node; |
e43473b7 VG |
365 | |
366 | rb_link_node(&tg->rb_node, parent, node); | |
0049af73 | 367 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); |
e43473b7 VG |
368 | } |
369 | ||
0049af73 TH |
370 | static void __throtl_enqueue_tg(struct throtl_grp *tg, |
371 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 372 | { |
0049af73 | 373 | tg_service_queue_add(tg, parent_sq); |
5b2c16aa | 374 | tg->flags |= THROTL_TG_PENDING; |
0049af73 | 375 | parent_sq->nr_pending++; |
e43473b7 VG |
376 | } |
377 | ||
0049af73 TH |
378 | static void throtl_enqueue_tg(struct throtl_grp *tg, |
379 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 380 | { |
5b2c16aa | 381 | if (!(tg->flags & THROTL_TG_PENDING)) |
0049af73 | 382 | __throtl_enqueue_tg(tg, parent_sq); |
e43473b7 VG |
383 | } |
384 | ||
0049af73 TH |
385 | static void __throtl_dequeue_tg(struct throtl_grp *tg, |
386 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 387 | { |
0049af73 | 388 | throtl_rb_erase(&tg->rb_node, parent_sq); |
5b2c16aa | 389 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
390 | } |
391 | ||
0049af73 TH |
392 | static void throtl_dequeue_tg(struct throtl_grp *tg, |
393 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 394 | { |
5b2c16aa | 395 | if (tg->flags & THROTL_TG_PENDING) |
0049af73 | 396 | __throtl_dequeue_tg(tg, parent_sq); |
e43473b7 VG |
397 | } |
398 | ||
a9131a27 TH |
399 | /* Call with queue lock held */ |
400 | static void throtl_schedule_delayed_work(struct throtl_data *td, | |
401 | unsigned long delay) | |
402 | { | |
403 | struct delayed_work *dwork = &td->dispatch_work; | |
404 | ||
6a525600 TH |
405 | mod_delayed_work(kthrotld_workqueue, dwork, delay); |
406 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", delay, jiffies); | |
a9131a27 TH |
407 | } |
408 | ||
e43473b7 VG |
409 | static void throtl_schedule_next_dispatch(struct throtl_data *td) |
410 | { | |
c9e0332e | 411 | struct throtl_service_queue *sq = &td->service_queue; |
e43473b7 | 412 | |
6a525600 | 413 | /* any pending children left? */ |
c9e0332e | 414 | if (!sq->nr_pending) |
e43473b7 VG |
415 | return; |
416 | ||
c9e0332e | 417 | update_min_dispatch_time(sq); |
e43473b7 | 418 | |
c9e0332e | 419 | if (time_before_eq(sq->first_pending_disptime, jiffies)) |
450adcbe | 420 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 421 | else |
c9e0332e | 422 | throtl_schedule_delayed_work(td, sq->first_pending_disptime - jiffies); |
e43473b7 VG |
423 | } |
424 | ||
0f3457f6 | 425 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
426 | { |
427 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 428 | tg->io_disp[rw] = 0; |
e43473b7 VG |
429 | tg->slice_start[rw] = jiffies; |
430 | tg->slice_end[rw] = jiffies + throtl_slice; | |
0f3457f6 | 431 | throtl_log_tg(tg, "[%c] new 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 | ||
0f3457f6 TH |
436 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
437 | unsigned long jiffy_end) | |
d1ae8ffd VG |
438 | { |
439 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
440 | } | |
441 | ||
0f3457f6 TH |
442 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
443 | unsigned long jiffy_end) | |
e43473b7 VG |
444 | { |
445 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
0f3457f6 | 446 | throtl_log_tg(tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", |
e43473b7 VG |
447 | rw == READ ? 'R' : 'W', tg->slice_start[rw], |
448 | tg->slice_end[rw], jiffies); | |
449 | } | |
450 | ||
451 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 452 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
453 | { |
454 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
455 | return 0; | |
456 | ||
457 | return 1; | |
458 | } | |
459 | ||
460 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 461 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 462 | { |
3aad5d3e VG |
463 | unsigned long nr_slices, time_elapsed, io_trim; |
464 | u64 bytes_trim, tmp; | |
e43473b7 VG |
465 | |
466 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
467 | ||
468 | /* | |
469 | * If bps are unlimited (-1), then time slice don't get | |
470 | * renewed. Don't try to trim the slice if slice is used. A new | |
471 | * slice will start when appropriate. | |
472 | */ | |
0f3457f6 | 473 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
474 | return; |
475 | ||
d1ae8ffd VG |
476 | /* |
477 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
478 | * that initially cgroup limit was very low resulting in high | |
479 | * slice_end, but later limit was bumped up and bio was dispached | |
480 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
481 | * is bad because it does not allow new slice to start. | |
482 | */ | |
483 | ||
0f3457f6 | 484 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); |
d1ae8ffd | 485 | |
e43473b7 VG |
486 | time_elapsed = jiffies - tg->slice_start[rw]; |
487 | ||
488 | nr_slices = time_elapsed / throtl_slice; | |
489 | ||
490 | if (!nr_slices) | |
491 | return; | |
3aad5d3e VG |
492 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
493 | do_div(tmp, HZ); | |
494 | bytes_trim = tmp; | |
e43473b7 | 495 | |
8e89d13f | 496 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 497 | |
8e89d13f | 498 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
499 | return; |
500 | ||
501 | if (tg->bytes_disp[rw] >= bytes_trim) | |
502 | tg->bytes_disp[rw] -= bytes_trim; | |
503 | else | |
504 | tg->bytes_disp[rw] = 0; | |
505 | ||
8e89d13f VG |
506 | if (tg->io_disp[rw] >= io_trim) |
507 | tg->io_disp[rw] -= io_trim; | |
508 | else | |
509 | tg->io_disp[rw] = 0; | |
510 | ||
e43473b7 VG |
511 | tg->slice_start[rw] += nr_slices * throtl_slice; |
512 | ||
0f3457f6 | 513 | throtl_log_tg(tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 514 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 515 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
516 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
517 | } | |
518 | ||
0f3457f6 TH |
519 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
520 | unsigned long *wait) | |
e43473b7 VG |
521 | { |
522 | bool rw = bio_data_dir(bio); | |
8e89d13f | 523 | unsigned int io_allowed; |
e43473b7 | 524 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 525 | u64 tmp; |
e43473b7 | 526 | |
8e89d13f | 527 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 528 | |
8e89d13f VG |
529 | /* Slice has just started. Consider one slice interval */ |
530 | if (!jiffy_elapsed) | |
531 | jiffy_elapsed_rnd = throtl_slice; | |
532 | ||
533 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
534 | ||
c49c06e4 VG |
535 | /* |
536 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
537 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
538 | * will allow dispatch after 1 second and after that slice should | |
539 | * have been trimmed. | |
540 | */ | |
541 | ||
542 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
543 | do_div(tmp, HZ); | |
544 | ||
545 | if (tmp > UINT_MAX) | |
546 | io_allowed = UINT_MAX; | |
547 | else | |
548 | io_allowed = tmp; | |
8e89d13f VG |
549 | |
550 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
551 | if (wait) |
552 | *wait = 0; | |
553 | return 1; | |
554 | } | |
555 | ||
8e89d13f VG |
556 | /* Calc approx time to dispatch */ |
557 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
558 | ||
559 | if (jiffy_wait > jiffy_elapsed) | |
560 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
561 | else | |
562 | jiffy_wait = 1; | |
563 | ||
564 | if (wait) | |
565 | *wait = jiffy_wait; | |
566 | return 0; | |
567 | } | |
568 | ||
0f3457f6 TH |
569 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
570 | unsigned long *wait) | |
8e89d13f VG |
571 | { |
572 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 573 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 574 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
575 | |
576 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
577 | ||
578 | /* Slice has just started. Consider one slice interval */ | |
579 | if (!jiffy_elapsed) | |
580 | jiffy_elapsed_rnd = throtl_slice; | |
581 | ||
582 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
583 | ||
5e901a2b VG |
584 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
585 | do_div(tmp, HZ); | |
3aad5d3e | 586 | bytes_allowed = tmp; |
e43473b7 VG |
587 | |
588 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
589 | if (wait) | |
590 | *wait = 0; | |
591 | return 1; | |
592 | } | |
593 | ||
594 | /* Calc approx time to dispatch */ | |
595 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
596 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
597 | ||
598 | if (!jiffy_wait) | |
599 | jiffy_wait = 1; | |
600 | ||
601 | /* | |
602 | * This wait time is without taking into consideration the rounding | |
603 | * up we did. Add that time also. | |
604 | */ | |
605 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
606 | if (wait) |
607 | *wait = jiffy_wait; | |
8e89d13f VG |
608 | return 0; |
609 | } | |
610 | ||
af75cd3c VG |
611 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
612 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
613 | return 1; | |
614 | return 0; | |
615 | } | |
616 | ||
8e89d13f VG |
617 | /* |
618 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
619 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
620 | */ | |
0f3457f6 TH |
621 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
622 | unsigned long *wait) | |
8e89d13f VG |
623 | { |
624 | bool rw = bio_data_dir(bio); | |
625 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
626 | ||
627 | /* | |
628 | * Currently whole state machine of group depends on first bio | |
629 | * queued in the group bio list. So one should not be calling | |
630 | * this function with a different bio if there are other bios | |
631 | * queued. | |
632 | */ | |
73f0d49a TH |
633 | BUG_ON(tg->service_queue.nr_queued[rw] && |
634 | bio != bio_list_peek(&tg->service_queue.bio_lists[rw])); | |
e43473b7 | 635 | |
8e89d13f VG |
636 | /* If tg->bps = -1, then BW is unlimited */ |
637 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
638 | if (wait) | |
639 | *wait = 0; | |
640 | return 1; | |
641 | } | |
642 | ||
643 | /* | |
644 | * If previous slice expired, start a new one otherwise renew/extend | |
645 | * existing slice to make sure it is at least throtl_slice interval | |
646 | * long since now. | |
647 | */ | |
0f3457f6 TH |
648 | if (throtl_slice_used(tg, rw)) |
649 | throtl_start_new_slice(tg, rw); | |
8e89d13f VG |
650 | else { |
651 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
0f3457f6 | 652 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); |
8e89d13f VG |
653 | } |
654 | ||
0f3457f6 TH |
655 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
656 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
657 | if (wait) |
658 | *wait = 0; | |
659 | return 1; | |
660 | } | |
661 | ||
662 | max_wait = max(bps_wait, iops_wait); | |
663 | ||
664 | if (wait) | |
665 | *wait = max_wait; | |
666 | ||
667 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 668 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
669 | |
670 | return 0; | |
671 | } | |
672 | ||
3c798398 | 673 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
674 | int rw) |
675 | { | |
8a3d2615 TH |
676 | struct throtl_grp *tg = blkg_to_tg(blkg); |
677 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
678 | unsigned long flags; |
679 | ||
680 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 681 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
682 | return; |
683 | ||
684 | /* | |
685 | * Disabling interrupts to provide mutual exclusion between two | |
686 | * writes on same cpu. It probably is not needed for 64bit. Not | |
687 | * optimizing that case yet. | |
688 | */ | |
689 | local_irq_save(flags); | |
690 | ||
8a3d2615 | 691 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 692 | |
629ed0b1 TH |
693 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
694 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
695 | ||
696 | local_irq_restore(flags); | |
697 | } | |
698 | ||
e43473b7 VG |
699 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
700 | { | |
701 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
702 | |
703 | /* Charge the bio to the group */ | |
704 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 705 | tg->io_disp[rw]++; |
e43473b7 | 706 | |
629ed0b1 | 707 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw); |
e43473b7 VG |
708 | } |
709 | ||
0049af73 TH |
710 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_grp *tg, |
711 | struct throtl_service_queue *parent_sq) | |
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]++; |
0049af73 | 730 | throtl_enqueue_tg(tg, parent_sq); |
e43473b7 VG |
731 | } |
732 | ||
0049af73 TH |
733 | static void tg_update_disptime(struct throtl_grp *tg, |
734 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 735 | { |
73f0d49a | 736 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
737 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
738 | struct bio *bio; | |
739 | ||
73f0d49a | 740 | if ((bio = bio_list_peek(&sq->bio_lists[READ]))) |
0f3457f6 | 741 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 742 | |
73f0d49a | 743 | if ((bio = bio_list_peek(&sq->bio_lists[WRITE]))) |
0f3457f6 | 744 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
745 | |
746 | min_wait = min(read_wait, write_wait); | |
747 | disptime = jiffies + min_wait; | |
748 | ||
e43473b7 | 749 | /* Update dispatch time */ |
0049af73 | 750 | throtl_dequeue_tg(tg, parent_sq); |
e43473b7 | 751 | tg->disptime = disptime; |
0049af73 | 752 | throtl_enqueue_tg(tg, parent_sq); |
0e9f4164 TH |
753 | |
754 | /* see throtl_add_bio_tg() */ | |
755 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
756 | } |
757 | ||
0f3457f6 | 758 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw, |
651930bc | 759 | struct throtl_service_queue *parent_sq) |
e43473b7 | 760 | { |
73f0d49a | 761 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
762 | struct bio *bio; |
763 | ||
73f0d49a TH |
764 | bio = bio_list_pop(&sq->bio_lists[rw]); |
765 | sq->nr_queued[rw]--; | |
1adaf3dd TH |
766 | /* Drop bio reference on blkg */ |
767 | blkg_put(tg_to_blkg(tg)); | |
e43473b7 | 768 | |
0f3457f6 TH |
769 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
770 | tg->td->nr_queued[rw]--; | |
e43473b7 VG |
771 | |
772 | throtl_charge_bio(tg, bio); | |
651930bc | 773 | bio_list_add(&parent_sq->bio_lists[rw], bio); |
e43473b7 VG |
774 | bio->bi_rw |= REQ_THROTTLED; |
775 | ||
0f3457f6 | 776 | throtl_trim_slice(tg, rw); |
e43473b7 VG |
777 | } |
778 | ||
651930bc TH |
779 | static int throtl_dispatch_tg(struct throtl_grp *tg, |
780 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 781 | { |
73f0d49a | 782 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
783 | unsigned int nr_reads = 0, nr_writes = 0; |
784 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 785 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
786 | struct bio *bio; |
787 | ||
788 | /* Try to dispatch 75% READS and 25% WRITES */ | |
789 | ||
73f0d49a | 790 | while ((bio = bio_list_peek(&sq->bio_lists[READ])) && |
0f3457f6 | 791 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 792 | |
651930bc | 793 | tg_dispatch_one_bio(tg, bio_data_dir(bio), parent_sq); |
e43473b7 VG |
794 | nr_reads++; |
795 | ||
796 | if (nr_reads >= max_nr_reads) | |
797 | break; | |
798 | } | |
799 | ||
73f0d49a | 800 | while ((bio = bio_list_peek(&sq->bio_lists[WRITE])) && |
0f3457f6 | 801 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 802 | |
651930bc | 803 | tg_dispatch_one_bio(tg, bio_data_dir(bio), parent_sq); |
e43473b7 VG |
804 | nr_writes++; |
805 | ||
806 | if (nr_writes >= max_nr_writes) | |
807 | break; | |
808 | } | |
809 | ||
810 | return nr_reads + nr_writes; | |
811 | } | |
812 | ||
651930bc | 813 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
814 | { |
815 | unsigned int nr_disp = 0; | |
e43473b7 VG |
816 | |
817 | while (1) { | |
73f0d49a TH |
818 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
819 | struct throtl_service_queue *sq = &tg->service_queue; | |
e43473b7 VG |
820 | |
821 | if (!tg) | |
822 | break; | |
823 | ||
824 | if (time_before(jiffies, tg->disptime)) | |
825 | break; | |
826 | ||
0049af73 | 827 | throtl_dequeue_tg(tg, parent_sq); |
e43473b7 | 828 | |
651930bc | 829 | nr_disp += throtl_dispatch_tg(tg, parent_sq); |
e43473b7 | 830 | |
73f0d49a | 831 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
0049af73 | 832 | tg_update_disptime(tg, parent_sq); |
e43473b7 VG |
833 | |
834 | if (nr_disp >= throtl_quantum) | |
835 | break; | |
836 | } | |
837 | ||
838 | return nr_disp; | |
839 | } | |
840 | ||
cb76199c TH |
841 | /* work function to dispatch throttled bios */ |
842 | void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
e43473b7 | 843 | { |
cb76199c TH |
844 | struct throtl_data *td = container_of(to_delayed_work(work), |
845 | struct throtl_data, dispatch_work); | |
651930bc | 846 | struct throtl_service_queue *sq = &td->service_queue; |
cb76199c | 847 | struct request_queue *q = td->queue; |
e43473b7 VG |
848 | unsigned int nr_disp = 0; |
849 | struct bio_list bio_list_on_stack; | |
850 | struct bio *bio; | |
69d60eb9 | 851 | struct blk_plug plug; |
651930bc | 852 | int rw; |
e43473b7 VG |
853 | |
854 | spin_lock_irq(q->queue_lock); | |
855 | ||
e43473b7 VG |
856 | bio_list_init(&bio_list_on_stack); |
857 | ||
d2f31a5f | 858 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
6a525600 TH |
859 | td->nr_queued[READ] + td->nr_queued[WRITE], |
860 | td->nr_queued[READ], td->nr_queued[WRITE]); | |
e43473b7 | 861 | |
651930bc | 862 | nr_disp = throtl_select_dispatch(sq); |
e43473b7 | 863 | |
651930bc TH |
864 | if (nr_disp) { |
865 | for (rw = READ; rw <= WRITE; rw++) { | |
866 | bio_list_merge(&bio_list_on_stack, &sq->bio_lists[rw]); | |
867 | bio_list_init(&sq->bio_lists[rw]); | |
868 | } | |
e43473b7 | 869 | throtl_log(td, "bios disp=%u", nr_disp); |
651930bc | 870 | } |
e43473b7 VG |
871 | |
872 | throtl_schedule_next_dispatch(td); | |
6a525600 | 873 | |
e43473b7 VG |
874 | spin_unlock_irq(q->queue_lock); |
875 | ||
876 | /* | |
877 | * If we dispatched some requests, unplug the queue to make sure | |
878 | * immediate dispatch | |
879 | */ | |
880 | if (nr_disp) { | |
69d60eb9 | 881 | blk_start_plug(&plug); |
e43473b7 VG |
882 | while((bio = bio_list_pop(&bio_list_on_stack))) |
883 | generic_make_request(bio); | |
69d60eb9 | 884 | blk_finish_plug(&plug); |
e43473b7 | 885 | } |
e43473b7 VG |
886 | } |
887 | ||
f95a04af TH |
888 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
889 | struct blkg_policy_data *pd, int off) | |
41b38b6d | 890 | { |
f95a04af | 891 | struct throtl_grp *tg = pd_to_tg(pd); |
41b38b6d TH |
892 | struct blkg_rwstat rwstat = { }, tmp; |
893 | int i, cpu; | |
894 | ||
895 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 896 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
897 | |
898 | tmp = blkg_rwstat_read((void *)sc + off); | |
899 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
900 | rwstat.cnt[i] += tmp.cnt[i]; | |
901 | } | |
902 | ||
f95a04af | 903 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
41b38b6d TH |
904 | } |
905 | ||
8a3d2615 TH |
906 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
907 | struct seq_file *sf) | |
41b38b6d | 908 | { |
3c798398 | 909 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
41b38b6d | 910 | |
3c798398 | 911 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 912 | cft->private, true); |
41b38b6d TH |
913 | return 0; |
914 | } | |
915 | ||
f95a04af TH |
916 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
917 | int off) | |
60c2bc2d | 918 | { |
f95a04af TH |
919 | struct throtl_grp *tg = pd_to_tg(pd); |
920 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 921 | |
af133ceb | 922 | if (v == -1) |
60c2bc2d | 923 | return 0; |
f95a04af | 924 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
925 | } |
926 | ||
f95a04af TH |
927 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
928 | int off) | |
e43473b7 | 929 | { |
f95a04af TH |
930 | struct throtl_grp *tg = pd_to_tg(pd); |
931 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 932 | |
af133ceb TH |
933 | if (v == -1) |
934 | return 0; | |
f95a04af | 935 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
936 | } |
937 | ||
af133ceb TH |
938 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
939 | struct seq_file *sf) | |
8e89d13f | 940 | { |
3c798398 TH |
941 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, |
942 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 943 | return 0; |
8e89d13f VG |
944 | } |
945 | ||
af133ceb TH |
946 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
947 | struct seq_file *sf) | |
8e89d13f | 948 | { |
3c798398 TH |
949 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, |
950 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 951 | return 0; |
60c2bc2d TH |
952 | } |
953 | ||
af133ceb TH |
954 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
955 | bool is_u64) | |
60c2bc2d | 956 | { |
3c798398 | 957 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
60c2bc2d | 958 | struct blkg_conf_ctx ctx; |
af133ceb | 959 | struct throtl_grp *tg; |
a2b1693b | 960 | struct throtl_data *td; |
60c2bc2d TH |
961 | int ret; |
962 | ||
3c798398 | 963 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
964 | if (ret) |
965 | return ret; | |
966 | ||
af133ceb | 967 | tg = blkg_to_tg(ctx.blkg); |
a2b1693b | 968 | td = ctx.blkg->q->td; |
af133ceb | 969 | |
a2b1693b TH |
970 | if (!ctx.v) |
971 | ctx.v = -1; | |
af133ceb | 972 | |
a2b1693b TH |
973 | if (is_u64) |
974 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
975 | else | |
976 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 977 | |
0f3457f6 | 978 | throtl_log_tg(tg, "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", |
632b4493 TH |
979 | tg->bps[READ], tg->bps[WRITE], |
980 | tg->iops[READ], tg->iops[WRITE]); | |
981 | ||
982 | /* | |
983 | * We're already holding queue_lock and know @tg is valid. Let's | |
984 | * apply the new config directly. | |
985 | * | |
986 | * Restart the slices for both READ and WRITES. It might happen | |
987 | * that a group's limit are dropped suddenly and we don't want to | |
988 | * account recently dispatched IO with new low rate. | |
989 | */ | |
0f3457f6 TH |
990 | throtl_start_new_slice(tg, 0); |
991 | throtl_start_new_slice(tg, 1); | |
632b4493 | 992 | |
5b2c16aa | 993 | if (tg->flags & THROTL_TG_PENDING) { |
0049af73 | 994 | tg_update_disptime(tg, &td->service_queue); |
632b4493 TH |
995 | throtl_schedule_next_dispatch(td); |
996 | } | |
60c2bc2d TH |
997 | |
998 | blkg_conf_finish(&ctx); | |
a2b1693b | 999 | return 0; |
8e89d13f VG |
1000 | } |
1001 | ||
af133ceb TH |
1002 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1003 | const char *buf) | |
60c2bc2d | 1004 | { |
af133ceb | 1005 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
1006 | } |
1007 | ||
af133ceb TH |
1008 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1009 | const char *buf) | |
60c2bc2d | 1010 | { |
af133ceb | 1011 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
1012 | } |
1013 | ||
1014 | static struct cftype throtl_files[] = { | |
1015 | { | |
1016 | .name = "throttle.read_bps_device", | |
af133ceb TH |
1017 | .private = offsetof(struct throtl_grp, bps[READ]), |
1018 | .read_seq_string = tg_print_conf_u64, | |
1019 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1020 | .max_write_len = 256, |
1021 | }, | |
1022 | { | |
1023 | .name = "throttle.write_bps_device", | |
af133ceb TH |
1024 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
1025 | .read_seq_string = tg_print_conf_u64, | |
1026 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1027 | .max_write_len = 256, |
1028 | }, | |
1029 | { | |
1030 | .name = "throttle.read_iops_device", | |
af133ceb TH |
1031 | .private = offsetof(struct throtl_grp, iops[READ]), |
1032 | .read_seq_string = tg_print_conf_uint, | |
1033 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1034 | .max_write_len = 256, |
1035 | }, | |
1036 | { | |
1037 | .name = "throttle.write_iops_device", | |
af133ceb TH |
1038 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
1039 | .read_seq_string = tg_print_conf_uint, | |
1040 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1041 | .max_write_len = 256, |
1042 | }, | |
1043 | { | |
1044 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1045 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1046 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1047 | }, |
1048 | { | |
1049 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1050 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1051 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1052 | }, |
1053 | { } /* terminate */ | |
1054 | }; | |
1055 | ||
da527770 | 1056 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1057 | { |
1058 | struct throtl_data *td = q->td; | |
1059 | ||
cb76199c | 1060 | cancel_delayed_work_sync(&td->dispatch_work); |
e43473b7 VG |
1061 | } |
1062 | ||
3c798398 | 1063 | static struct blkcg_policy blkcg_policy_throtl = { |
f9fcc2d3 TH |
1064 | .pd_size = sizeof(struct throtl_grp), |
1065 | .cftypes = throtl_files, | |
1066 | ||
1067 | .pd_init_fn = throtl_pd_init, | |
1068 | .pd_exit_fn = throtl_pd_exit, | |
1069 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 VG |
1070 | }; |
1071 | ||
bc16a4f9 | 1072 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1073 | { |
1074 | struct throtl_data *td = q->td; | |
1075 | struct throtl_grp *tg; | |
73f0d49a | 1076 | struct throtl_service_queue *sq; |
0e9f4164 | 1077 | bool rw = bio_data_dir(bio); |
3c798398 | 1078 | struct blkcg *blkcg; |
bc16a4f9 | 1079 | bool throttled = false; |
e43473b7 VG |
1080 | |
1081 | if (bio->bi_rw & REQ_THROTTLED) { | |
1082 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1083 | goto out; |
e43473b7 VG |
1084 | } |
1085 | ||
af75cd3c VG |
1086 | /* |
1087 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1088 | * basic fields like stats and io rates. If a group has no rules, | |
1089 | * just update the dispatch stats in lockless manner and return. | |
1090 | */ | |
af75cd3c | 1091 | rcu_read_lock(); |
3c798398 | 1092 | blkcg = bio_blkcg(bio); |
cd1604fa | 1093 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1094 | if (tg) { |
af75cd3c | 1095 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1096 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1097 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1098 | goto out_unlock_rcu; |
af75cd3c VG |
1099 | } |
1100 | } | |
af75cd3c VG |
1101 | |
1102 | /* | |
1103 | * Either group has not been allocated yet or it is not an unlimited | |
1104 | * IO group | |
1105 | */ | |
e43473b7 | 1106 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1107 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1108 | if (unlikely(!tg)) |
1109 | goto out_unlock; | |
f469a7b4 | 1110 | |
73f0d49a TH |
1111 | sq = &tg->service_queue; |
1112 | ||
0e9f4164 TH |
1113 | /* throtl is FIFO - if other bios are already queued, should queue */ |
1114 | if (sq->nr_queued[rw]) | |
e43473b7 | 1115 | goto queue_bio; |
de701c74 | 1116 | |
e43473b7 | 1117 | /* Bio is with-in rate limit of group */ |
0f3457f6 | 1118 | if (tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1119 | throtl_charge_bio(tg, bio); |
04521db0 VG |
1120 | |
1121 | /* | |
1122 | * We need to trim slice even when bios are not being queued | |
1123 | * otherwise it might happen that a bio is not queued for | |
1124 | * a long time and slice keeps on extending and trim is not | |
1125 | * called for a long time. Now if limits are reduced suddenly | |
1126 | * we take into account all the IO dispatched so far at new | |
1127 | * low rate and * newly queued IO gets a really long dispatch | |
1128 | * time. | |
1129 | * | |
1130 | * So keep on trimming slice even if bio is not queued. | |
1131 | */ | |
0f3457f6 | 1132 | throtl_trim_slice(tg, rw); |
bc16a4f9 | 1133 | goto out_unlock; |
e43473b7 VG |
1134 | } |
1135 | ||
1136 | queue_bio: | |
0f3457f6 | 1137 | throtl_log_tg(tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
8e89d13f VG |
1138 | " iodisp=%u iops=%u queued=%d/%d", |
1139 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1140 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1141 | tg->io_disp[rw], tg->iops[rw], |
73f0d49a | 1142 | sq->nr_queued[READ], sq->nr_queued[WRITE]); |
e43473b7 | 1143 | |
671058fb | 1144 | bio_associate_current(bio); |
0049af73 | 1145 | throtl_add_bio_tg(bio, tg, &q->td->service_queue); |
bc16a4f9 | 1146 | throttled = true; |
e43473b7 | 1147 | |
0e9f4164 TH |
1148 | /* update @tg's dispatch time if @tg was empty before @bio */ |
1149 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
0049af73 | 1150 | tg_update_disptime(tg, &td->service_queue); |
e43473b7 VG |
1151 | throtl_schedule_next_dispatch(td); |
1152 | } | |
1153 | ||
bc16a4f9 | 1154 | out_unlock: |
e43473b7 | 1155 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1156 | out_unlock_rcu: |
1157 | rcu_read_unlock(); | |
bc16a4f9 TH |
1158 | out: |
1159 | return throttled; | |
e43473b7 VG |
1160 | } |
1161 | ||
c9a929dd TH |
1162 | /** |
1163 | * blk_throtl_drain - drain throttled bios | |
1164 | * @q: request_queue to drain throttled bios for | |
1165 | * | |
1166 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1167 | */ | |
1168 | void blk_throtl_drain(struct request_queue *q) | |
1169 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1170 | { | |
1171 | struct throtl_data *td = q->td; | |
0049af73 | 1172 | struct throtl_service_queue *parent_sq = &td->service_queue; |
c9a929dd | 1173 | struct throtl_grp *tg; |
c9a929dd | 1174 | struct bio *bio; |
651930bc | 1175 | int rw; |
c9a929dd | 1176 | |
8bcb6c7d | 1177 | queue_lockdep_assert_held(q); |
c9a929dd | 1178 | |
0049af73 | 1179 | while ((tg = throtl_rb_first(parent_sq))) { |
73f0d49a TH |
1180 | struct throtl_service_queue *sq = &tg->service_queue; |
1181 | ||
0049af73 | 1182 | throtl_dequeue_tg(tg, parent_sq); |
c9a929dd | 1183 | |
73f0d49a | 1184 | while ((bio = bio_list_peek(&sq->bio_lists[READ]))) |
651930bc | 1185 | tg_dispatch_one_bio(tg, bio_data_dir(bio), parent_sq); |
73f0d49a | 1186 | while ((bio = bio_list_peek(&sq->bio_lists[WRITE]))) |
651930bc | 1187 | tg_dispatch_one_bio(tg, bio_data_dir(bio), parent_sq); |
c9a929dd TH |
1188 | } |
1189 | spin_unlock_irq(q->queue_lock); | |
1190 | ||
651930bc TH |
1191 | for (rw = READ; rw <= WRITE; rw++) |
1192 | while ((bio = bio_list_pop(&parent_sq->bio_lists[rw]))) | |
1193 | generic_make_request(bio); | |
c9a929dd TH |
1194 | |
1195 | spin_lock_irq(q->queue_lock); | |
1196 | } | |
1197 | ||
e43473b7 VG |
1198 | int blk_throtl_init(struct request_queue *q) |
1199 | { | |
1200 | struct throtl_data *td; | |
a2b1693b | 1201 | int ret; |
e43473b7 VG |
1202 | |
1203 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1204 | if (!td) | |
1205 | return -ENOMEM; | |
1206 | ||
cb76199c | 1207 | INIT_DELAYED_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
49a2f1e3 | 1208 | throtl_service_queue_init(&td->service_queue); |
e43473b7 | 1209 | |
cd1604fa | 1210 | q->td = td; |
29b12589 | 1211 | td->queue = q; |
02977e4a | 1212 | |
a2b1693b | 1213 | /* activate policy */ |
3c798398 | 1214 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1215 | if (ret) |
f51b802c | 1216 | kfree(td); |
a2b1693b | 1217 | return ret; |
e43473b7 VG |
1218 | } |
1219 | ||
1220 | void blk_throtl_exit(struct request_queue *q) | |
1221 | { | |
c875f4d0 | 1222 | BUG_ON(!q->td); |
da527770 | 1223 | throtl_shutdown_wq(q); |
3c798398 | 1224 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1225 | kfree(q->td); |
e43473b7 VG |
1226 | } |
1227 | ||
1228 | static int __init throtl_init(void) | |
1229 | { | |
450adcbe VG |
1230 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1231 | if (!kthrotld_workqueue) | |
1232 | panic("Failed to create kthrotld\n"); | |
1233 | ||
3c798398 | 1234 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1235 | } |
1236 | ||
1237 | module_init(throtl_init); |