]>
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 | ||
fda6f272 TH |
154 | /** |
155 | * sq_to_tg - return the throl_grp the specified service queue belongs to | |
156 | * @sq: the throtl_service_queue of interest | |
157 | * | |
158 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one | |
159 | * embedded in throtl_data, %NULL is returned. | |
160 | */ | |
161 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) | |
162 | { | |
163 | if (sq && sq->parent_sq) | |
164 | return container_of(sq, struct throtl_grp, service_queue); | |
165 | else | |
166 | return NULL; | |
167 | } | |
168 | ||
169 | /** | |
170 | * sq_to_td - return throtl_data the specified service queue belongs to | |
171 | * @sq: the throtl_service_queue of interest | |
172 | * | |
173 | * A service_queue can be embeded in either a throtl_grp or throtl_data. | |
174 | * Determine the associated throtl_data accordingly and return it. | |
175 | */ | |
176 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) | |
177 | { | |
178 | struct throtl_grp *tg = sq_to_tg(sq); | |
179 | ||
180 | if (tg) | |
181 | return tg->td; | |
182 | else | |
183 | return container_of(sq, struct throtl_data, service_queue); | |
184 | } | |
185 | ||
186 | /** | |
187 | * throtl_log - log debug message via blktrace | |
188 | * @sq: the service_queue being reported | |
189 | * @fmt: printf format string | |
190 | * @args: printf args | |
191 | * | |
192 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a | |
193 | * throtl_grp; otherwise, just "throtl". | |
194 | * | |
195 | * TODO: this should be made a function and name formatting should happen | |
196 | * after testing whether blktrace is enabled. | |
197 | */ | |
198 | #define throtl_log(sq, fmt, args...) do { \ | |
199 | struct throtl_grp *__tg = sq_to_tg((sq)); \ | |
200 | struct throtl_data *__td = sq_to_td((sq)); \ | |
201 | \ | |
202 | (void)__td; \ | |
203 | if ((__tg)) { \ | |
204 | char __pbuf[128]; \ | |
54e7ed12 | 205 | \ |
fda6f272 TH |
206 | blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \ |
207 | blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
208 | } else { \ | |
209 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
210 | } \ | |
54e7ed12 | 211 | } while (0) |
e43473b7 | 212 | |
8a3d2615 TH |
213 | /* |
214 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
3b07e9ca | 215 | * system_wq once there are some groups on the alloc_list waiting for |
8a3d2615 TH |
216 | * allocation. |
217 | */ | |
218 | static void tg_stats_alloc_fn(struct work_struct *work) | |
219 | { | |
220 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
221 | struct delayed_work *dwork = to_delayed_work(work); | |
222 | bool empty = false; | |
223 | ||
224 | alloc_stats: | |
225 | if (!stats_cpu) { | |
226 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
227 | if (!stats_cpu) { | |
228 | /* allocation failed, try again after some time */ | |
3b07e9ca | 229 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); |
8a3d2615 TH |
230 | return; |
231 | } | |
232 | } | |
233 | ||
234 | spin_lock_irq(&tg_stats_alloc_lock); | |
235 | ||
236 | if (!list_empty(&tg_stats_alloc_list)) { | |
237 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
238 | struct throtl_grp, | |
239 | stats_alloc_node); | |
240 | swap(tg->stats_cpu, stats_cpu); | |
241 | list_del_init(&tg->stats_alloc_node); | |
242 | } | |
243 | ||
244 | empty = list_empty(&tg_stats_alloc_list); | |
245 | spin_unlock_irq(&tg_stats_alloc_lock); | |
246 | if (!empty) | |
247 | goto alloc_stats; | |
248 | } | |
249 | ||
49a2f1e3 | 250 | /* init a service_queue, assumes the caller zeroed it */ |
77216b04 TH |
251 | static void throtl_service_queue_init(struct throtl_service_queue *sq, |
252 | struct throtl_service_queue *parent_sq) | |
49a2f1e3 | 253 | { |
73f0d49a TH |
254 | bio_list_init(&sq->bio_lists[0]); |
255 | bio_list_init(&sq->bio_lists[1]); | |
49a2f1e3 | 256 | sq->pending_tree = RB_ROOT; |
77216b04 | 257 | sq->parent_sq = parent_sq; |
49a2f1e3 TH |
258 | } |
259 | ||
3c798398 | 260 | static void throtl_pd_init(struct blkcg_gq *blkg) |
a29a171e | 261 | { |
0381411e | 262 | struct throtl_grp *tg = blkg_to_tg(blkg); |
77216b04 | 263 | struct throtl_data *td = blkg->q->td; |
ff26eaad | 264 | unsigned long flags; |
cd1604fa | 265 | |
77216b04 | 266 | throtl_service_queue_init(&tg->service_queue, &td->service_queue); |
a29a171e | 267 | RB_CLEAR_NODE(&tg->rb_node); |
77216b04 | 268 | tg->td = td; |
a29a171e | 269 | |
e56da7e2 TH |
270 | tg->bps[READ] = -1; |
271 | tg->bps[WRITE] = -1; | |
272 | tg->iops[READ] = -1; | |
273 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
274 | |
275 | /* | |
276 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
277 | * but percpu allocator can't be called from IO path. Queue tg on | |
278 | * tg_stats_alloc_list and allocate from work item. | |
279 | */ | |
ff26eaad | 280 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 281 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); |
3b07e9ca | 282 | schedule_delayed_work(&tg_stats_alloc_work, 0); |
ff26eaad | 283 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
284 | } |
285 | ||
3c798398 | 286 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
8a3d2615 TH |
287 | { |
288 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
ff26eaad | 289 | unsigned long flags; |
8a3d2615 | 290 | |
ff26eaad | 291 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 292 | list_del_init(&tg->stats_alloc_node); |
ff26eaad | 293 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
294 | |
295 | free_percpu(tg->stats_cpu); | |
296 | } | |
297 | ||
3c798398 | 298 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
8a3d2615 TH |
299 | { |
300 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
301 | int cpu; | |
302 | ||
303 | if (tg->stats_cpu == NULL) | |
304 | return; | |
305 | ||
306 | for_each_possible_cpu(cpu) { | |
307 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
308 | ||
309 | blkg_rwstat_reset(&sc->service_bytes); | |
310 | blkg_rwstat_reset(&sc->serviced); | |
311 | } | |
a29a171e VG |
312 | } |
313 | ||
3c798398 TH |
314 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
315 | struct blkcg *blkcg) | |
e43473b7 | 316 | { |
be2c6b19 | 317 | /* |
3c798398 TH |
318 | * This is the common case when there are no blkcgs. Avoid lookup |
319 | * in this case | |
cd1604fa | 320 | */ |
3c798398 | 321 | if (blkcg == &blkcg_root) |
03d8e111 | 322 | return td_root_tg(td); |
e43473b7 | 323 | |
e8989fae | 324 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
325 | } |
326 | ||
cd1604fa | 327 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
3c798398 | 328 | struct blkcg *blkcg) |
e43473b7 | 329 | { |
f469a7b4 | 330 | struct request_queue *q = td->queue; |
cd1604fa | 331 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 332 | |
f469a7b4 | 333 | /* |
3c798398 TH |
334 | * This is the common case when there are no blkcgs. Avoid lookup |
335 | * in this case | |
f469a7b4 | 336 | */ |
3c798398 | 337 | if (blkcg == &blkcg_root) { |
03d8e111 | 338 | tg = td_root_tg(td); |
cd1604fa | 339 | } else { |
3c798398 | 340 | struct blkcg_gq *blkg; |
f469a7b4 | 341 | |
3c96cb32 | 342 | blkg = blkg_lookup_create(blkcg, q); |
f469a7b4 | 343 | |
cd1604fa TH |
344 | /* if %NULL and @q is alive, fall back to root_tg */ |
345 | if (!IS_ERR(blkg)) | |
0381411e | 346 | tg = blkg_to_tg(blkg); |
3f3299d5 | 347 | else if (!blk_queue_dying(q)) |
03d8e111 | 348 | tg = td_root_tg(td); |
f469a7b4 VG |
349 | } |
350 | ||
e43473b7 VG |
351 | return tg; |
352 | } | |
353 | ||
0049af73 TH |
354 | static struct throtl_grp * |
355 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
356 | { |
357 | /* Service tree is empty */ | |
0049af73 | 358 | if (!parent_sq->nr_pending) |
e43473b7 VG |
359 | return NULL; |
360 | ||
0049af73 TH |
361 | if (!parent_sq->first_pending) |
362 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
e43473b7 | 363 | |
0049af73 TH |
364 | if (parent_sq->first_pending) |
365 | return rb_entry_tg(parent_sq->first_pending); | |
e43473b7 VG |
366 | |
367 | return NULL; | |
368 | } | |
369 | ||
370 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
371 | { | |
372 | rb_erase(n, root); | |
373 | RB_CLEAR_NODE(n); | |
374 | } | |
375 | ||
0049af73 TH |
376 | static void throtl_rb_erase(struct rb_node *n, |
377 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 378 | { |
0049af73 TH |
379 | if (parent_sq->first_pending == n) |
380 | parent_sq->first_pending = NULL; | |
381 | rb_erase_init(n, &parent_sq->pending_tree); | |
382 | --parent_sq->nr_pending; | |
e43473b7 VG |
383 | } |
384 | ||
0049af73 | 385 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
386 | { |
387 | struct throtl_grp *tg; | |
388 | ||
0049af73 | 389 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
390 | if (!tg) |
391 | return; | |
392 | ||
0049af73 | 393 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
394 | } |
395 | ||
77216b04 | 396 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 397 | { |
77216b04 | 398 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
0049af73 | 399 | struct rb_node **node = &parent_sq->pending_tree.rb_node; |
e43473b7 VG |
400 | struct rb_node *parent = NULL; |
401 | struct throtl_grp *__tg; | |
402 | unsigned long key = tg->disptime; | |
403 | int left = 1; | |
404 | ||
405 | while (*node != NULL) { | |
406 | parent = *node; | |
407 | __tg = rb_entry_tg(parent); | |
408 | ||
409 | if (time_before(key, __tg->disptime)) | |
410 | node = &parent->rb_left; | |
411 | else { | |
412 | node = &parent->rb_right; | |
413 | left = 0; | |
414 | } | |
415 | } | |
416 | ||
417 | if (left) | |
0049af73 | 418 | parent_sq->first_pending = &tg->rb_node; |
e43473b7 VG |
419 | |
420 | rb_link_node(&tg->rb_node, parent, node); | |
0049af73 | 421 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); |
e43473b7 VG |
422 | } |
423 | ||
77216b04 | 424 | static void __throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 425 | { |
77216b04 | 426 | tg_service_queue_add(tg); |
5b2c16aa | 427 | tg->flags |= THROTL_TG_PENDING; |
77216b04 | 428 | tg->service_queue.parent_sq->nr_pending++; |
e43473b7 VG |
429 | } |
430 | ||
77216b04 | 431 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 432 | { |
5b2c16aa | 433 | if (!(tg->flags & THROTL_TG_PENDING)) |
77216b04 | 434 | __throtl_enqueue_tg(tg); |
e43473b7 VG |
435 | } |
436 | ||
77216b04 | 437 | static void __throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 438 | { |
77216b04 | 439 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
5b2c16aa | 440 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
441 | } |
442 | ||
77216b04 | 443 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 444 | { |
5b2c16aa | 445 | if (tg->flags & THROTL_TG_PENDING) |
77216b04 | 446 | __throtl_dequeue_tg(tg); |
e43473b7 VG |
447 | } |
448 | ||
a9131a27 TH |
449 | /* Call with queue lock held */ |
450 | static void throtl_schedule_delayed_work(struct throtl_data *td, | |
451 | unsigned long delay) | |
452 | { | |
453 | struct delayed_work *dwork = &td->dispatch_work; | |
fda6f272 | 454 | struct throtl_service_queue *sq = &td->service_queue; |
a9131a27 | 455 | |
6a525600 | 456 | mod_delayed_work(kthrotld_workqueue, dwork, delay); |
fda6f272 | 457 | throtl_log(sq, "schedule work. delay=%lu jiffies=%lu", delay, jiffies); |
a9131a27 TH |
458 | } |
459 | ||
e43473b7 VG |
460 | static void throtl_schedule_next_dispatch(struct throtl_data *td) |
461 | { | |
c9e0332e | 462 | struct throtl_service_queue *sq = &td->service_queue; |
e43473b7 | 463 | |
6a525600 | 464 | /* any pending children left? */ |
c9e0332e | 465 | if (!sq->nr_pending) |
e43473b7 VG |
466 | return; |
467 | ||
c9e0332e | 468 | update_min_dispatch_time(sq); |
e43473b7 | 469 | |
c9e0332e | 470 | if (time_before_eq(sq->first_pending_disptime, jiffies)) |
450adcbe | 471 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 472 | else |
c9e0332e | 473 | throtl_schedule_delayed_work(td, sq->first_pending_disptime - jiffies); |
e43473b7 VG |
474 | } |
475 | ||
0f3457f6 | 476 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
477 | { |
478 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 479 | tg->io_disp[rw] = 0; |
e43473b7 VG |
480 | tg->slice_start[rw] = jiffies; |
481 | tg->slice_end[rw] = jiffies + throtl_slice; | |
fda6f272 TH |
482 | throtl_log(&tg->service_queue, |
483 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
484 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
485 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
486 | } |
487 | ||
0f3457f6 TH |
488 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
489 | unsigned long jiffy_end) | |
d1ae8ffd VG |
490 | { |
491 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
492 | } | |
493 | ||
0f3457f6 TH |
494 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
495 | unsigned long jiffy_end) | |
e43473b7 VG |
496 | { |
497 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
fda6f272 TH |
498 | throtl_log(&tg->service_queue, |
499 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
500 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
501 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
502 | } |
503 | ||
504 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 505 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
506 | { |
507 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
508 | return 0; | |
509 | ||
510 | return 1; | |
511 | } | |
512 | ||
513 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 514 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 515 | { |
3aad5d3e VG |
516 | unsigned long nr_slices, time_elapsed, io_trim; |
517 | u64 bytes_trim, tmp; | |
e43473b7 VG |
518 | |
519 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
520 | ||
521 | /* | |
522 | * If bps are unlimited (-1), then time slice don't get | |
523 | * renewed. Don't try to trim the slice if slice is used. A new | |
524 | * slice will start when appropriate. | |
525 | */ | |
0f3457f6 | 526 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
527 | return; |
528 | ||
d1ae8ffd VG |
529 | /* |
530 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
531 | * that initially cgroup limit was very low resulting in high | |
532 | * slice_end, but later limit was bumped up and bio was dispached | |
533 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
534 | * is bad because it does not allow new slice to start. | |
535 | */ | |
536 | ||
0f3457f6 | 537 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); |
d1ae8ffd | 538 | |
e43473b7 VG |
539 | time_elapsed = jiffies - tg->slice_start[rw]; |
540 | ||
541 | nr_slices = time_elapsed / throtl_slice; | |
542 | ||
543 | if (!nr_slices) | |
544 | return; | |
3aad5d3e VG |
545 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
546 | do_div(tmp, HZ); | |
547 | bytes_trim = tmp; | |
e43473b7 | 548 | |
8e89d13f | 549 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 550 | |
8e89d13f | 551 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
552 | return; |
553 | ||
554 | if (tg->bytes_disp[rw] >= bytes_trim) | |
555 | tg->bytes_disp[rw] -= bytes_trim; | |
556 | else | |
557 | tg->bytes_disp[rw] = 0; | |
558 | ||
8e89d13f VG |
559 | if (tg->io_disp[rw] >= io_trim) |
560 | tg->io_disp[rw] -= io_trim; | |
561 | else | |
562 | tg->io_disp[rw] = 0; | |
563 | ||
e43473b7 VG |
564 | tg->slice_start[rw] += nr_slices * throtl_slice; |
565 | ||
fda6f272 TH |
566 | throtl_log(&tg->service_queue, |
567 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
568 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
569 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
e43473b7 VG |
570 | } |
571 | ||
0f3457f6 TH |
572 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
573 | unsigned long *wait) | |
e43473b7 VG |
574 | { |
575 | bool rw = bio_data_dir(bio); | |
8e89d13f | 576 | unsigned int io_allowed; |
e43473b7 | 577 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 578 | u64 tmp; |
e43473b7 | 579 | |
8e89d13f | 580 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 581 | |
8e89d13f VG |
582 | /* Slice has just started. Consider one slice interval */ |
583 | if (!jiffy_elapsed) | |
584 | jiffy_elapsed_rnd = throtl_slice; | |
585 | ||
586 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
587 | ||
c49c06e4 VG |
588 | /* |
589 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
590 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
591 | * will allow dispatch after 1 second and after that slice should | |
592 | * have been trimmed. | |
593 | */ | |
594 | ||
595 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
596 | do_div(tmp, HZ); | |
597 | ||
598 | if (tmp > UINT_MAX) | |
599 | io_allowed = UINT_MAX; | |
600 | else | |
601 | io_allowed = tmp; | |
8e89d13f VG |
602 | |
603 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
604 | if (wait) |
605 | *wait = 0; | |
606 | return 1; | |
607 | } | |
608 | ||
8e89d13f VG |
609 | /* Calc approx time to dispatch */ |
610 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
611 | ||
612 | if (jiffy_wait > jiffy_elapsed) | |
613 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
614 | else | |
615 | jiffy_wait = 1; | |
616 | ||
617 | if (wait) | |
618 | *wait = jiffy_wait; | |
619 | return 0; | |
620 | } | |
621 | ||
0f3457f6 TH |
622 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
623 | unsigned long *wait) | |
8e89d13f VG |
624 | { |
625 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 626 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 627 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
628 | |
629 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
630 | ||
631 | /* Slice has just started. Consider one slice interval */ | |
632 | if (!jiffy_elapsed) | |
633 | jiffy_elapsed_rnd = throtl_slice; | |
634 | ||
635 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
636 | ||
5e901a2b VG |
637 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
638 | do_div(tmp, HZ); | |
3aad5d3e | 639 | bytes_allowed = tmp; |
e43473b7 VG |
640 | |
641 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
642 | if (wait) | |
643 | *wait = 0; | |
644 | return 1; | |
645 | } | |
646 | ||
647 | /* Calc approx time to dispatch */ | |
648 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
649 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
650 | ||
651 | if (!jiffy_wait) | |
652 | jiffy_wait = 1; | |
653 | ||
654 | /* | |
655 | * This wait time is without taking into consideration the rounding | |
656 | * up we did. Add that time also. | |
657 | */ | |
658 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
659 | if (wait) |
660 | *wait = jiffy_wait; | |
8e89d13f VG |
661 | return 0; |
662 | } | |
663 | ||
af75cd3c VG |
664 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
665 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
666 | return 1; | |
667 | return 0; | |
668 | } | |
669 | ||
8e89d13f VG |
670 | /* |
671 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
672 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
673 | */ | |
0f3457f6 TH |
674 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
675 | unsigned long *wait) | |
8e89d13f VG |
676 | { |
677 | bool rw = bio_data_dir(bio); | |
678 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
679 | ||
680 | /* | |
681 | * Currently whole state machine of group depends on first bio | |
682 | * queued in the group bio list. So one should not be calling | |
683 | * this function with a different bio if there are other bios | |
684 | * queued. | |
685 | */ | |
73f0d49a TH |
686 | BUG_ON(tg->service_queue.nr_queued[rw] && |
687 | bio != bio_list_peek(&tg->service_queue.bio_lists[rw])); | |
e43473b7 | 688 | |
8e89d13f VG |
689 | /* If tg->bps = -1, then BW is unlimited */ |
690 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
691 | if (wait) | |
692 | *wait = 0; | |
693 | return 1; | |
694 | } | |
695 | ||
696 | /* | |
697 | * If previous slice expired, start a new one otherwise renew/extend | |
698 | * existing slice to make sure it is at least throtl_slice interval | |
699 | * long since now. | |
700 | */ | |
0f3457f6 TH |
701 | if (throtl_slice_used(tg, rw)) |
702 | throtl_start_new_slice(tg, rw); | |
8e89d13f VG |
703 | else { |
704 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
0f3457f6 | 705 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); |
8e89d13f VG |
706 | } |
707 | ||
0f3457f6 TH |
708 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
709 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
710 | if (wait) |
711 | *wait = 0; | |
712 | return 1; | |
713 | } | |
714 | ||
715 | max_wait = max(bps_wait, iops_wait); | |
716 | ||
717 | if (wait) | |
718 | *wait = max_wait; | |
719 | ||
720 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 721 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
722 | |
723 | return 0; | |
724 | } | |
725 | ||
3c798398 | 726 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
727 | int rw) |
728 | { | |
8a3d2615 TH |
729 | struct throtl_grp *tg = blkg_to_tg(blkg); |
730 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
731 | unsigned long flags; |
732 | ||
733 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 734 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
735 | return; |
736 | ||
737 | /* | |
738 | * Disabling interrupts to provide mutual exclusion between two | |
739 | * writes on same cpu. It probably is not needed for 64bit. Not | |
740 | * optimizing that case yet. | |
741 | */ | |
742 | local_irq_save(flags); | |
743 | ||
8a3d2615 | 744 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 745 | |
629ed0b1 TH |
746 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
747 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
748 | ||
749 | local_irq_restore(flags); | |
750 | } | |
751 | ||
e43473b7 VG |
752 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
753 | { | |
754 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
755 | |
756 | /* Charge the bio to the group */ | |
757 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 758 | tg->io_disp[rw]++; |
e43473b7 | 759 | |
629ed0b1 | 760 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw); |
e43473b7 VG |
761 | } |
762 | ||
77216b04 | 763 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_grp *tg) |
e43473b7 | 764 | { |
73f0d49a | 765 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
766 | bool rw = bio_data_dir(bio); |
767 | ||
0e9f4164 TH |
768 | /* |
769 | * If @tg doesn't currently have any bios queued in the same | |
770 | * direction, queueing @bio can change when @tg should be | |
771 | * dispatched. Mark that @tg was empty. This is automatically | |
772 | * cleaered on the next tg_update_disptime(). | |
773 | */ | |
774 | if (!sq->nr_queued[rw]) | |
775 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
776 | ||
73f0d49a | 777 | bio_list_add(&sq->bio_lists[rw], bio); |
e43473b7 | 778 | /* Take a bio reference on tg */ |
1adaf3dd | 779 | blkg_get(tg_to_blkg(tg)); |
73f0d49a | 780 | sq->nr_queued[rw]++; |
e2d57e60 | 781 | tg->td->nr_queued[rw]++; |
77216b04 | 782 | throtl_enqueue_tg(tg); |
e43473b7 VG |
783 | } |
784 | ||
77216b04 | 785 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 786 | { |
73f0d49a | 787 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
788 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
789 | struct bio *bio; | |
790 | ||
73f0d49a | 791 | if ((bio = bio_list_peek(&sq->bio_lists[READ]))) |
0f3457f6 | 792 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 793 | |
73f0d49a | 794 | if ((bio = bio_list_peek(&sq->bio_lists[WRITE]))) |
0f3457f6 | 795 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
796 | |
797 | min_wait = min(read_wait, write_wait); | |
798 | disptime = jiffies + min_wait; | |
799 | ||
e43473b7 | 800 | /* Update dispatch time */ |
77216b04 | 801 | throtl_dequeue_tg(tg); |
e43473b7 | 802 | tg->disptime = disptime; |
77216b04 | 803 | throtl_enqueue_tg(tg); |
0e9f4164 TH |
804 | |
805 | /* see throtl_add_bio_tg() */ | |
806 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
807 | } |
808 | ||
77216b04 | 809 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 810 | { |
73f0d49a | 811 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
812 | struct bio *bio; |
813 | ||
73f0d49a TH |
814 | bio = bio_list_pop(&sq->bio_lists[rw]); |
815 | sq->nr_queued[rw]--; | |
1adaf3dd TH |
816 | /* Drop bio reference on blkg */ |
817 | blkg_put(tg_to_blkg(tg)); | |
e43473b7 | 818 | |
0f3457f6 TH |
819 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
820 | tg->td->nr_queued[rw]--; | |
e43473b7 VG |
821 | |
822 | throtl_charge_bio(tg, bio); | |
77216b04 | 823 | bio_list_add(&sq->parent_sq->bio_lists[rw], bio); |
e43473b7 VG |
824 | bio->bi_rw |= REQ_THROTTLED; |
825 | ||
0f3457f6 | 826 | throtl_trim_slice(tg, rw); |
e43473b7 VG |
827 | } |
828 | ||
77216b04 | 829 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 830 | { |
73f0d49a | 831 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
832 | unsigned int nr_reads = 0, nr_writes = 0; |
833 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 834 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
835 | struct bio *bio; |
836 | ||
837 | /* Try to dispatch 75% READS and 25% WRITES */ | |
838 | ||
73f0d49a | 839 | while ((bio = bio_list_peek(&sq->bio_lists[READ])) && |
0f3457f6 | 840 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 841 | |
77216b04 | 842 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
843 | nr_reads++; |
844 | ||
845 | if (nr_reads >= max_nr_reads) | |
846 | break; | |
847 | } | |
848 | ||
73f0d49a | 849 | while ((bio = bio_list_peek(&sq->bio_lists[WRITE])) && |
0f3457f6 | 850 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 851 | |
77216b04 | 852 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
853 | nr_writes++; |
854 | ||
855 | if (nr_writes >= max_nr_writes) | |
856 | break; | |
857 | } | |
858 | ||
859 | return nr_reads + nr_writes; | |
860 | } | |
861 | ||
651930bc | 862 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
863 | { |
864 | unsigned int nr_disp = 0; | |
e43473b7 VG |
865 | |
866 | while (1) { | |
73f0d49a TH |
867 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
868 | struct throtl_service_queue *sq = &tg->service_queue; | |
e43473b7 VG |
869 | |
870 | if (!tg) | |
871 | break; | |
872 | ||
873 | if (time_before(jiffies, tg->disptime)) | |
874 | break; | |
875 | ||
77216b04 | 876 | throtl_dequeue_tg(tg); |
e43473b7 | 877 | |
77216b04 | 878 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 879 | |
73f0d49a | 880 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
77216b04 | 881 | tg_update_disptime(tg); |
e43473b7 VG |
882 | |
883 | if (nr_disp >= throtl_quantum) | |
884 | break; | |
885 | } | |
886 | ||
887 | return nr_disp; | |
888 | } | |
889 | ||
cb76199c TH |
890 | /* work function to dispatch throttled bios */ |
891 | void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
e43473b7 | 892 | { |
cb76199c TH |
893 | struct throtl_data *td = container_of(to_delayed_work(work), |
894 | struct throtl_data, dispatch_work); | |
651930bc | 895 | struct throtl_service_queue *sq = &td->service_queue; |
cb76199c | 896 | struct request_queue *q = td->queue; |
e43473b7 VG |
897 | unsigned int nr_disp = 0; |
898 | struct bio_list bio_list_on_stack; | |
899 | struct bio *bio; | |
69d60eb9 | 900 | struct blk_plug plug; |
651930bc | 901 | int rw; |
e43473b7 VG |
902 | |
903 | spin_lock_irq(q->queue_lock); | |
904 | ||
e43473b7 VG |
905 | bio_list_init(&bio_list_on_stack); |
906 | ||
fda6f272 | 907 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", |
6a525600 TH |
908 | td->nr_queued[READ] + td->nr_queued[WRITE], |
909 | td->nr_queued[READ], td->nr_queued[WRITE]); | |
e43473b7 | 910 | |
651930bc | 911 | nr_disp = throtl_select_dispatch(sq); |
e43473b7 | 912 | |
651930bc TH |
913 | if (nr_disp) { |
914 | for (rw = READ; rw <= WRITE; rw++) { | |
915 | bio_list_merge(&bio_list_on_stack, &sq->bio_lists[rw]); | |
916 | bio_list_init(&sq->bio_lists[rw]); | |
917 | } | |
fda6f272 | 918 | throtl_log(sq, "bios disp=%u", nr_disp); |
651930bc | 919 | } |
e43473b7 VG |
920 | |
921 | throtl_schedule_next_dispatch(td); | |
6a525600 | 922 | |
e43473b7 VG |
923 | spin_unlock_irq(q->queue_lock); |
924 | ||
925 | /* | |
926 | * If we dispatched some requests, unplug the queue to make sure | |
927 | * immediate dispatch | |
928 | */ | |
929 | if (nr_disp) { | |
69d60eb9 | 930 | blk_start_plug(&plug); |
e43473b7 VG |
931 | while((bio = bio_list_pop(&bio_list_on_stack))) |
932 | generic_make_request(bio); | |
69d60eb9 | 933 | blk_finish_plug(&plug); |
e43473b7 | 934 | } |
e43473b7 VG |
935 | } |
936 | ||
f95a04af TH |
937 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
938 | struct blkg_policy_data *pd, int off) | |
41b38b6d | 939 | { |
f95a04af | 940 | struct throtl_grp *tg = pd_to_tg(pd); |
41b38b6d TH |
941 | struct blkg_rwstat rwstat = { }, tmp; |
942 | int i, cpu; | |
943 | ||
944 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 945 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
946 | |
947 | tmp = blkg_rwstat_read((void *)sc + off); | |
948 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
949 | rwstat.cnt[i] += tmp.cnt[i]; | |
950 | } | |
951 | ||
f95a04af | 952 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
41b38b6d TH |
953 | } |
954 | ||
8a3d2615 TH |
955 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
956 | struct seq_file *sf) | |
41b38b6d | 957 | { |
3c798398 | 958 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
41b38b6d | 959 | |
3c798398 | 960 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 961 | cft->private, true); |
41b38b6d TH |
962 | return 0; |
963 | } | |
964 | ||
f95a04af TH |
965 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
966 | int off) | |
60c2bc2d | 967 | { |
f95a04af TH |
968 | struct throtl_grp *tg = pd_to_tg(pd); |
969 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 970 | |
af133ceb | 971 | if (v == -1) |
60c2bc2d | 972 | return 0; |
f95a04af | 973 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
974 | } |
975 | ||
f95a04af TH |
976 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
977 | int off) | |
e43473b7 | 978 | { |
f95a04af TH |
979 | struct throtl_grp *tg = pd_to_tg(pd); |
980 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 981 | |
af133ceb TH |
982 | if (v == -1) |
983 | return 0; | |
f95a04af | 984 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
985 | } |
986 | ||
af133ceb TH |
987 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
988 | struct seq_file *sf) | |
8e89d13f | 989 | { |
3c798398 TH |
990 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, |
991 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 992 | return 0; |
8e89d13f VG |
993 | } |
994 | ||
af133ceb TH |
995 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
996 | struct seq_file *sf) | |
8e89d13f | 997 | { |
3c798398 TH |
998 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, |
999 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 1000 | return 0; |
60c2bc2d TH |
1001 | } |
1002 | ||
af133ceb TH |
1003 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
1004 | bool is_u64) | |
60c2bc2d | 1005 | { |
3c798398 | 1006 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
60c2bc2d | 1007 | struct blkg_conf_ctx ctx; |
af133ceb | 1008 | struct throtl_grp *tg; |
a2b1693b | 1009 | struct throtl_data *td; |
60c2bc2d TH |
1010 | int ret; |
1011 | ||
3c798398 | 1012 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
1013 | if (ret) |
1014 | return ret; | |
1015 | ||
af133ceb | 1016 | tg = blkg_to_tg(ctx.blkg); |
a2b1693b | 1017 | td = ctx.blkg->q->td; |
af133ceb | 1018 | |
a2b1693b TH |
1019 | if (!ctx.v) |
1020 | ctx.v = -1; | |
af133ceb | 1021 | |
a2b1693b TH |
1022 | if (is_u64) |
1023 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
1024 | else | |
1025 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 1026 | |
fda6f272 TH |
1027 | throtl_log(&tg->service_queue, |
1028 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
1029 | tg->bps[READ], tg->bps[WRITE], | |
1030 | tg->iops[READ], tg->iops[WRITE]); | |
632b4493 TH |
1031 | |
1032 | /* | |
1033 | * We're already holding queue_lock and know @tg is valid. Let's | |
1034 | * apply the new config directly. | |
1035 | * | |
1036 | * Restart the slices for both READ and WRITES. It might happen | |
1037 | * that a group's limit are dropped suddenly and we don't want to | |
1038 | * account recently dispatched IO with new low rate. | |
1039 | */ | |
0f3457f6 TH |
1040 | throtl_start_new_slice(tg, 0); |
1041 | throtl_start_new_slice(tg, 1); | |
632b4493 | 1042 | |
5b2c16aa | 1043 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 1044 | tg_update_disptime(tg); |
632b4493 TH |
1045 | throtl_schedule_next_dispatch(td); |
1046 | } | |
60c2bc2d TH |
1047 | |
1048 | blkg_conf_finish(&ctx); | |
a2b1693b | 1049 | return 0; |
8e89d13f VG |
1050 | } |
1051 | ||
af133ceb TH |
1052 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1053 | const char *buf) | |
60c2bc2d | 1054 | { |
af133ceb | 1055 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
1056 | } |
1057 | ||
af133ceb TH |
1058 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1059 | const char *buf) | |
60c2bc2d | 1060 | { |
af133ceb | 1061 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
1062 | } |
1063 | ||
1064 | static struct cftype throtl_files[] = { | |
1065 | { | |
1066 | .name = "throttle.read_bps_device", | |
af133ceb TH |
1067 | .private = offsetof(struct throtl_grp, bps[READ]), |
1068 | .read_seq_string = tg_print_conf_u64, | |
1069 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1070 | .max_write_len = 256, |
1071 | }, | |
1072 | { | |
1073 | .name = "throttle.write_bps_device", | |
af133ceb TH |
1074 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
1075 | .read_seq_string = tg_print_conf_u64, | |
1076 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1077 | .max_write_len = 256, |
1078 | }, | |
1079 | { | |
1080 | .name = "throttle.read_iops_device", | |
af133ceb TH |
1081 | .private = offsetof(struct throtl_grp, iops[READ]), |
1082 | .read_seq_string = tg_print_conf_uint, | |
1083 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1084 | .max_write_len = 256, |
1085 | }, | |
1086 | { | |
1087 | .name = "throttle.write_iops_device", | |
af133ceb TH |
1088 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
1089 | .read_seq_string = tg_print_conf_uint, | |
1090 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1091 | .max_write_len = 256, |
1092 | }, | |
1093 | { | |
1094 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1095 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1096 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1097 | }, |
1098 | { | |
1099 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1100 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1101 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1102 | }, |
1103 | { } /* terminate */ | |
1104 | }; | |
1105 | ||
da527770 | 1106 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1107 | { |
1108 | struct throtl_data *td = q->td; | |
1109 | ||
cb76199c | 1110 | cancel_delayed_work_sync(&td->dispatch_work); |
e43473b7 VG |
1111 | } |
1112 | ||
3c798398 | 1113 | static struct blkcg_policy blkcg_policy_throtl = { |
f9fcc2d3 TH |
1114 | .pd_size = sizeof(struct throtl_grp), |
1115 | .cftypes = throtl_files, | |
1116 | ||
1117 | .pd_init_fn = throtl_pd_init, | |
1118 | .pd_exit_fn = throtl_pd_exit, | |
1119 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 VG |
1120 | }; |
1121 | ||
bc16a4f9 | 1122 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1123 | { |
1124 | struct throtl_data *td = q->td; | |
1125 | struct throtl_grp *tg; | |
73f0d49a | 1126 | struct throtl_service_queue *sq; |
0e9f4164 | 1127 | bool rw = bio_data_dir(bio); |
3c798398 | 1128 | struct blkcg *blkcg; |
bc16a4f9 | 1129 | bool throttled = false; |
e43473b7 VG |
1130 | |
1131 | if (bio->bi_rw & REQ_THROTTLED) { | |
1132 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1133 | goto out; |
e43473b7 VG |
1134 | } |
1135 | ||
af75cd3c VG |
1136 | /* |
1137 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1138 | * basic fields like stats and io rates. If a group has no rules, | |
1139 | * just update the dispatch stats in lockless manner and return. | |
1140 | */ | |
af75cd3c | 1141 | rcu_read_lock(); |
3c798398 | 1142 | blkcg = bio_blkcg(bio); |
cd1604fa | 1143 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1144 | if (tg) { |
af75cd3c | 1145 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1146 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1147 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1148 | goto out_unlock_rcu; |
af75cd3c VG |
1149 | } |
1150 | } | |
af75cd3c VG |
1151 | |
1152 | /* | |
1153 | * Either group has not been allocated yet or it is not an unlimited | |
1154 | * IO group | |
1155 | */ | |
e43473b7 | 1156 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1157 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1158 | if (unlikely(!tg)) |
1159 | goto out_unlock; | |
f469a7b4 | 1160 | |
73f0d49a TH |
1161 | sq = &tg->service_queue; |
1162 | ||
0e9f4164 TH |
1163 | /* throtl is FIFO - if other bios are already queued, should queue */ |
1164 | if (sq->nr_queued[rw]) | |
e43473b7 | 1165 | goto queue_bio; |
de701c74 | 1166 | |
e43473b7 | 1167 | /* Bio is with-in rate limit of group */ |
0f3457f6 | 1168 | if (tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1169 | throtl_charge_bio(tg, bio); |
04521db0 VG |
1170 | |
1171 | /* | |
1172 | * We need to trim slice even when bios are not being queued | |
1173 | * otherwise it might happen that a bio is not queued for | |
1174 | * a long time and slice keeps on extending and trim is not | |
1175 | * called for a long time. Now if limits are reduced suddenly | |
1176 | * we take into account all the IO dispatched so far at new | |
1177 | * low rate and * newly queued IO gets a really long dispatch | |
1178 | * time. | |
1179 | * | |
1180 | * So keep on trimming slice even if bio is not queued. | |
1181 | */ | |
0f3457f6 | 1182 | throtl_trim_slice(tg, rw); |
bc16a4f9 | 1183 | goto out_unlock; |
e43473b7 VG |
1184 | } |
1185 | ||
1186 | queue_bio: | |
fda6f272 TH |
1187 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
1188 | rw == READ ? 'R' : 'W', | |
1189 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], | |
1190 | tg->io_disp[rw], tg->iops[rw], | |
1191 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
e43473b7 | 1192 | |
671058fb | 1193 | bio_associate_current(bio); |
77216b04 | 1194 | throtl_add_bio_tg(bio, tg); |
bc16a4f9 | 1195 | throttled = true; |
e43473b7 | 1196 | |
0e9f4164 TH |
1197 | /* update @tg's dispatch time if @tg was empty before @bio */ |
1198 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
77216b04 | 1199 | tg_update_disptime(tg); |
e43473b7 VG |
1200 | throtl_schedule_next_dispatch(td); |
1201 | } | |
1202 | ||
bc16a4f9 | 1203 | out_unlock: |
e43473b7 | 1204 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1205 | out_unlock_rcu: |
1206 | rcu_read_unlock(); | |
bc16a4f9 TH |
1207 | out: |
1208 | return throttled; | |
e43473b7 VG |
1209 | } |
1210 | ||
c9a929dd TH |
1211 | /** |
1212 | * blk_throtl_drain - drain throttled bios | |
1213 | * @q: request_queue to drain throttled bios for | |
1214 | * | |
1215 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1216 | */ | |
1217 | void blk_throtl_drain(struct request_queue *q) | |
1218 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1219 | { | |
1220 | struct throtl_data *td = q->td; | |
0049af73 | 1221 | struct throtl_service_queue *parent_sq = &td->service_queue; |
c9a929dd | 1222 | struct throtl_grp *tg; |
c9a929dd | 1223 | struct bio *bio; |
651930bc | 1224 | int rw; |
c9a929dd | 1225 | |
8bcb6c7d | 1226 | queue_lockdep_assert_held(q); |
c9a929dd | 1227 | |
0049af73 | 1228 | while ((tg = throtl_rb_first(parent_sq))) { |
73f0d49a TH |
1229 | struct throtl_service_queue *sq = &tg->service_queue; |
1230 | ||
77216b04 | 1231 | throtl_dequeue_tg(tg); |
c9a929dd | 1232 | |
73f0d49a | 1233 | while ((bio = bio_list_peek(&sq->bio_lists[READ]))) |
77216b04 | 1234 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
73f0d49a | 1235 | while ((bio = bio_list_peek(&sq->bio_lists[WRITE]))) |
77216b04 | 1236 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
c9a929dd TH |
1237 | } |
1238 | spin_unlock_irq(q->queue_lock); | |
1239 | ||
651930bc TH |
1240 | for (rw = READ; rw <= WRITE; rw++) |
1241 | while ((bio = bio_list_pop(&parent_sq->bio_lists[rw]))) | |
1242 | generic_make_request(bio); | |
c9a929dd TH |
1243 | |
1244 | spin_lock_irq(q->queue_lock); | |
1245 | } | |
1246 | ||
e43473b7 VG |
1247 | int blk_throtl_init(struct request_queue *q) |
1248 | { | |
1249 | struct throtl_data *td; | |
a2b1693b | 1250 | int ret; |
e43473b7 VG |
1251 | |
1252 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1253 | if (!td) | |
1254 | return -ENOMEM; | |
1255 | ||
cb76199c | 1256 | INIT_DELAYED_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
77216b04 | 1257 | throtl_service_queue_init(&td->service_queue, NULL); |
e43473b7 | 1258 | |
cd1604fa | 1259 | q->td = td; |
29b12589 | 1260 | td->queue = q; |
02977e4a | 1261 | |
a2b1693b | 1262 | /* activate policy */ |
3c798398 | 1263 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1264 | if (ret) |
f51b802c | 1265 | kfree(td); |
a2b1693b | 1266 | return ret; |
e43473b7 VG |
1267 | } |
1268 | ||
1269 | void blk_throtl_exit(struct request_queue *q) | |
1270 | { | |
c875f4d0 | 1271 | BUG_ON(!q->td); |
da527770 | 1272 | throtl_shutdown_wq(q); |
3c798398 | 1273 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1274 | kfree(q->td); |
e43473b7 VG |
1275 | } |
1276 | ||
1277 | static int __init throtl_init(void) | |
1278 | { | |
450adcbe VG |
1279 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1280 | if (!kthrotld_workqueue) | |
1281 | panic("Failed to create kthrotld\n"); | |
1282 | ||
3c798398 | 1283 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1284 | } |
1285 | ||
1286 | module_init(throtl_init); |