]>
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 */ | |
69df0ab0 | 47 | struct timer_list pending_timer; /* fires on first_pending_disptime */ |
e43473b7 VG |
48 | }; |
49 | ||
5b2c16aa TH |
50 | enum tg_state_flags { |
51 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
0e9f4164 | 52 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ |
5b2c16aa TH |
53 | }; |
54 | ||
e43473b7 VG |
55 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
56 | ||
8a3d2615 TH |
57 | /* Per-cpu group stats */ |
58 | struct tg_stats_cpu { | |
59 | /* total bytes transferred */ | |
60 | struct blkg_rwstat service_bytes; | |
61 | /* total IOs serviced, post merge */ | |
62 | struct blkg_rwstat serviced; | |
63 | }; | |
64 | ||
e43473b7 | 65 | struct throtl_grp { |
f95a04af TH |
66 | /* must be the first member */ |
67 | struct blkg_policy_data pd; | |
68 | ||
c9e0332e | 69 | /* active throtl group service_queue member */ |
e43473b7 VG |
70 | struct rb_node rb_node; |
71 | ||
0f3457f6 TH |
72 | /* throtl_data this group belongs to */ |
73 | struct throtl_data *td; | |
74 | ||
49a2f1e3 TH |
75 | /* this group's service queue */ |
76 | struct throtl_service_queue service_queue; | |
77 | ||
e43473b7 VG |
78 | /* |
79 | * Dispatch time in jiffies. This is the estimated time when group | |
80 | * will unthrottle and is ready to dispatch more bio. It is used as | |
81 | * key to sort active groups in service tree. | |
82 | */ | |
83 | unsigned long disptime; | |
84 | ||
e43473b7 VG |
85 | unsigned int flags; |
86 | ||
e43473b7 VG |
87 | /* bytes per second rate limits */ |
88 | uint64_t bps[2]; | |
89 | ||
8e89d13f VG |
90 | /* IOPS limits */ |
91 | unsigned int iops[2]; | |
92 | ||
e43473b7 VG |
93 | /* Number of bytes disptached in current slice */ |
94 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
95 | /* Number of bio's dispatched in current slice */ |
96 | unsigned int io_disp[2]; | |
e43473b7 VG |
97 | |
98 | /* When did we start a new slice */ | |
99 | unsigned long slice_start[2]; | |
100 | unsigned long slice_end[2]; | |
fe071437 | 101 | |
8a3d2615 TH |
102 | /* Per cpu stats pointer */ |
103 | struct tg_stats_cpu __percpu *stats_cpu; | |
104 | ||
105 | /* List of tgs waiting for per cpu stats memory to be allocated */ | |
106 | struct list_head stats_alloc_node; | |
e43473b7 VG |
107 | }; |
108 | ||
109 | struct throtl_data | |
110 | { | |
e43473b7 | 111 | /* service tree for active throtl groups */ |
c9e0332e | 112 | struct throtl_service_queue service_queue; |
e43473b7 | 113 | |
e43473b7 VG |
114 | struct request_queue *queue; |
115 | ||
116 | /* Total Number of queued bios on READ and WRITE lists */ | |
117 | unsigned int nr_queued[2]; | |
118 | ||
119 | /* | |
02977e4a | 120 | * number of total undestroyed groups |
e43473b7 VG |
121 | */ |
122 | unsigned int nr_undestroyed_grps; | |
123 | ||
124 | /* Work for dispatching throttled bios */ | |
69df0ab0 | 125 | struct work_struct dispatch_work; |
e43473b7 VG |
126 | }; |
127 | ||
8a3d2615 TH |
128 | /* list and work item to allocate percpu group stats */ |
129 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
130 | static LIST_HEAD(tg_stats_alloc_list); | |
131 | ||
132 | static void tg_stats_alloc_fn(struct work_struct *); | |
133 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
134 | ||
69df0ab0 TH |
135 | static void throtl_pending_timer_fn(unsigned long arg); |
136 | ||
f95a04af TH |
137 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
138 | { | |
139 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
140 | } | |
141 | ||
3c798398 | 142 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 143 | { |
f95a04af | 144 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
145 | } |
146 | ||
3c798398 | 147 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 148 | { |
f95a04af | 149 | return pd_to_blkg(&tg->pd); |
0381411e TH |
150 | } |
151 | ||
03d8e111 TH |
152 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) |
153 | { | |
154 | return blkg_to_tg(td->queue->root_blkg); | |
155 | } | |
156 | ||
fda6f272 TH |
157 | /** |
158 | * sq_to_tg - return the throl_grp the specified service queue belongs to | |
159 | * @sq: the throtl_service_queue of interest | |
160 | * | |
161 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one | |
162 | * embedded in throtl_data, %NULL is returned. | |
163 | */ | |
164 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) | |
165 | { | |
166 | if (sq && sq->parent_sq) | |
167 | return container_of(sq, struct throtl_grp, service_queue); | |
168 | else | |
169 | return NULL; | |
170 | } | |
171 | ||
172 | /** | |
173 | * sq_to_td - return throtl_data the specified service queue belongs to | |
174 | * @sq: the throtl_service_queue of interest | |
175 | * | |
176 | * A service_queue can be embeded in either a throtl_grp or throtl_data. | |
177 | * Determine the associated throtl_data accordingly and return it. | |
178 | */ | |
179 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) | |
180 | { | |
181 | struct throtl_grp *tg = sq_to_tg(sq); | |
182 | ||
183 | if (tg) | |
184 | return tg->td; | |
185 | else | |
186 | return container_of(sq, struct throtl_data, service_queue); | |
187 | } | |
188 | ||
189 | /** | |
190 | * throtl_log - log debug message via blktrace | |
191 | * @sq: the service_queue being reported | |
192 | * @fmt: printf format string | |
193 | * @args: printf args | |
194 | * | |
195 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a | |
196 | * throtl_grp; otherwise, just "throtl". | |
197 | * | |
198 | * TODO: this should be made a function and name formatting should happen | |
199 | * after testing whether blktrace is enabled. | |
200 | */ | |
201 | #define throtl_log(sq, fmt, args...) do { \ | |
202 | struct throtl_grp *__tg = sq_to_tg((sq)); \ | |
203 | struct throtl_data *__td = sq_to_td((sq)); \ | |
204 | \ | |
205 | (void)__td; \ | |
206 | if ((__tg)) { \ | |
207 | char __pbuf[128]; \ | |
54e7ed12 | 208 | \ |
fda6f272 TH |
209 | blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \ |
210 | blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
211 | } else { \ | |
212 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
213 | } \ | |
54e7ed12 | 214 | } while (0) |
e43473b7 | 215 | |
8a3d2615 TH |
216 | /* |
217 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
3b07e9ca | 218 | * system_wq once there are some groups on the alloc_list waiting for |
8a3d2615 TH |
219 | * allocation. |
220 | */ | |
221 | static void tg_stats_alloc_fn(struct work_struct *work) | |
222 | { | |
223 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
224 | struct delayed_work *dwork = to_delayed_work(work); | |
225 | bool empty = false; | |
226 | ||
227 | alloc_stats: | |
228 | if (!stats_cpu) { | |
229 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
230 | if (!stats_cpu) { | |
231 | /* allocation failed, try again after some time */ | |
3b07e9ca | 232 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); |
8a3d2615 TH |
233 | return; |
234 | } | |
235 | } | |
236 | ||
237 | spin_lock_irq(&tg_stats_alloc_lock); | |
238 | ||
239 | if (!list_empty(&tg_stats_alloc_list)) { | |
240 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
241 | struct throtl_grp, | |
242 | stats_alloc_node); | |
243 | swap(tg->stats_cpu, stats_cpu); | |
244 | list_del_init(&tg->stats_alloc_node); | |
245 | } | |
246 | ||
247 | empty = list_empty(&tg_stats_alloc_list); | |
248 | spin_unlock_irq(&tg_stats_alloc_lock); | |
249 | if (!empty) | |
250 | goto alloc_stats; | |
251 | } | |
252 | ||
49a2f1e3 | 253 | /* init a service_queue, assumes the caller zeroed it */ |
77216b04 TH |
254 | static void throtl_service_queue_init(struct throtl_service_queue *sq, |
255 | struct throtl_service_queue *parent_sq) | |
49a2f1e3 | 256 | { |
73f0d49a TH |
257 | bio_list_init(&sq->bio_lists[0]); |
258 | bio_list_init(&sq->bio_lists[1]); | |
49a2f1e3 | 259 | sq->pending_tree = RB_ROOT; |
77216b04 | 260 | sq->parent_sq = parent_sq; |
69df0ab0 TH |
261 | setup_timer(&sq->pending_timer, throtl_pending_timer_fn, |
262 | (unsigned long)sq); | |
263 | } | |
264 | ||
265 | static void throtl_service_queue_exit(struct throtl_service_queue *sq) | |
266 | { | |
267 | del_timer_sync(&sq->pending_timer); | |
49a2f1e3 TH |
268 | } |
269 | ||
3c798398 | 270 | static void throtl_pd_init(struct blkcg_gq *blkg) |
a29a171e | 271 | { |
0381411e | 272 | struct throtl_grp *tg = blkg_to_tg(blkg); |
77216b04 | 273 | struct throtl_data *td = blkg->q->td; |
ff26eaad | 274 | unsigned long flags; |
cd1604fa | 275 | |
77216b04 | 276 | throtl_service_queue_init(&tg->service_queue, &td->service_queue); |
a29a171e | 277 | RB_CLEAR_NODE(&tg->rb_node); |
77216b04 | 278 | tg->td = td; |
a29a171e | 279 | |
e56da7e2 TH |
280 | tg->bps[READ] = -1; |
281 | tg->bps[WRITE] = -1; | |
282 | tg->iops[READ] = -1; | |
283 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
284 | |
285 | /* | |
286 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
287 | * but percpu allocator can't be called from IO path. Queue tg on | |
288 | * tg_stats_alloc_list and allocate from work item. | |
289 | */ | |
ff26eaad | 290 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 291 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); |
3b07e9ca | 292 | schedule_delayed_work(&tg_stats_alloc_work, 0); |
ff26eaad | 293 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
294 | } |
295 | ||
3c798398 | 296 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
8a3d2615 TH |
297 | { |
298 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
ff26eaad | 299 | unsigned long flags; |
8a3d2615 | 300 | |
ff26eaad | 301 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 302 | list_del_init(&tg->stats_alloc_node); |
ff26eaad | 303 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
304 | |
305 | free_percpu(tg->stats_cpu); | |
69df0ab0 TH |
306 | |
307 | throtl_service_queue_exit(&tg->service_queue); | |
8a3d2615 TH |
308 | } |
309 | ||
3c798398 | 310 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
8a3d2615 TH |
311 | { |
312 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
313 | int cpu; | |
314 | ||
315 | if (tg->stats_cpu == NULL) | |
316 | return; | |
317 | ||
318 | for_each_possible_cpu(cpu) { | |
319 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
320 | ||
321 | blkg_rwstat_reset(&sc->service_bytes); | |
322 | blkg_rwstat_reset(&sc->serviced); | |
323 | } | |
a29a171e VG |
324 | } |
325 | ||
3c798398 TH |
326 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
327 | struct blkcg *blkcg) | |
e43473b7 | 328 | { |
be2c6b19 | 329 | /* |
3c798398 TH |
330 | * This is the common case when there are no blkcgs. Avoid lookup |
331 | * in this case | |
cd1604fa | 332 | */ |
3c798398 | 333 | if (blkcg == &blkcg_root) |
03d8e111 | 334 | return td_root_tg(td); |
e43473b7 | 335 | |
e8989fae | 336 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
337 | } |
338 | ||
cd1604fa | 339 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
3c798398 | 340 | struct blkcg *blkcg) |
e43473b7 | 341 | { |
f469a7b4 | 342 | struct request_queue *q = td->queue; |
cd1604fa | 343 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 344 | |
f469a7b4 | 345 | /* |
3c798398 TH |
346 | * This is the common case when there are no blkcgs. Avoid lookup |
347 | * in this case | |
f469a7b4 | 348 | */ |
3c798398 | 349 | if (blkcg == &blkcg_root) { |
03d8e111 | 350 | tg = td_root_tg(td); |
cd1604fa | 351 | } else { |
3c798398 | 352 | struct blkcg_gq *blkg; |
f469a7b4 | 353 | |
3c96cb32 | 354 | blkg = blkg_lookup_create(blkcg, q); |
f469a7b4 | 355 | |
cd1604fa TH |
356 | /* if %NULL and @q is alive, fall back to root_tg */ |
357 | if (!IS_ERR(blkg)) | |
0381411e | 358 | tg = blkg_to_tg(blkg); |
3f3299d5 | 359 | else if (!blk_queue_dying(q)) |
03d8e111 | 360 | tg = td_root_tg(td); |
f469a7b4 VG |
361 | } |
362 | ||
e43473b7 VG |
363 | return tg; |
364 | } | |
365 | ||
0049af73 TH |
366 | static struct throtl_grp * |
367 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
368 | { |
369 | /* Service tree is empty */ | |
0049af73 | 370 | if (!parent_sq->nr_pending) |
e43473b7 VG |
371 | return NULL; |
372 | ||
0049af73 TH |
373 | if (!parent_sq->first_pending) |
374 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
e43473b7 | 375 | |
0049af73 TH |
376 | if (parent_sq->first_pending) |
377 | return rb_entry_tg(parent_sq->first_pending); | |
e43473b7 VG |
378 | |
379 | return NULL; | |
380 | } | |
381 | ||
382 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
383 | { | |
384 | rb_erase(n, root); | |
385 | RB_CLEAR_NODE(n); | |
386 | } | |
387 | ||
0049af73 TH |
388 | static void throtl_rb_erase(struct rb_node *n, |
389 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 390 | { |
0049af73 TH |
391 | if (parent_sq->first_pending == n) |
392 | parent_sq->first_pending = NULL; | |
393 | rb_erase_init(n, &parent_sq->pending_tree); | |
394 | --parent_sq->nr_pending; | |
e43473b7 VG |
395 | } |
396 | ||
0049af73 | 397 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
398 | { |
399 | struct throtl_grp *tg; | |
400 | ||
0049af73 | 401 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
402 | if (!tg) |
403 | return; | |
404 | ||
0049af73 | 405 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
406 | } |
407 | ||
77216b04 | 408 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 409 | { |
77216b04 | 410 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
0049af73 | 411 | struct rb_node **node = &parent_sq->pending_tree.rb_node; |
e43473b7 VG |
412 | struct rb_node *parent = NULL; |
413 | struct throtl_grp *__tg; | |
414 | unsigned long key = tg->disptime; | |
415 | int left = 1; | |
416 | ||
417 | while (*node != NULL) { | |
418 | parent = *node; | |
419 | __tg = rb_entry_tg(parent); | |
420 | ||
421 | if (time_before(key, __tg->disptime)) | |
422 | node = &parent->rb_left; | |
423 | else { | |
424 | node = &parent->rb_right; | |
425 | left = 0; | |
426 | } | |
427 | } | |
428 | ||
429 | if (left) | |
0049af73 | 430 | parent_sq->first_pending = &tg->rb_node; |
e43473b7 VG |
431 | |
432 | rb_link_node(&tg->rb_node, parent, node); | |
0049af73 | 433 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); |
e43473b7 VG |
434 | } |
435 | ||
77216b04 | 436 | static void __throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 437 | { |
77216b04 | 438 | tg_service_queue_add(tg); |
5b2c16aa | 439 | tg->flags |= THROTL_TG_PENDING; |
77216b04 | 440 | tg->service_queue.parent_sq->nr_pending++; |
e43473b7 VG |
441 | } |
442 | ||
77216b04 | 443 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 444 | { |
5b2c16aa | 445 | if (!(tg->flags & THROTL_TG_PENDING)) |
77216b04 | 446 | __throtl_enqueue_tg(tg); |
e43473b7 VG |
447 | } |
448 | ||
77216b04 | 449 | static void __throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 450 | { |
77216b04 | 451 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
5b2c16aa | 452 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
453 | } |
454 | ||
77216b04 | 455 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 456 | { |
5b2c16aa | 457 | if (tg->flags & THROTL_TG_PENDING) |
77216b04 | 458 | __throtl_dequeue_tg(tg); |
e43473b7 VG |
459 | } |
460 | ||
a9131a27 | 461 | /* Call with queue lock held */ |
69df0ab0 TH |
462 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, |
463 | unsigned long expires) | |
a9131a27 | 464 | { |
69df0ab0 TH |
465 | mod_timer(&sq->pending_timer, expires); |
466 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
467 | expires - jiffies, jiffies); | |
a9131a27 TH |
468 | } |
469 | ||
69df0ab0 | 470 | static void throtl_schedule_next_dispatch(struct throtl_service_queue *sq) |
e43473b7 | 471 | { |
69df0ab0 | 472 | struct throtl_data *td = sq_to_td(sq); |
e43473b7 | 473 | |
6a525600 | 474 | /* any pending children left? */ |
c9e0332e | 475 | if (!sq->nr_pending) |
e43473b7 VG |
476 | return; |
477 | ||
c9e0332e | 478 | update_min_dispatch_time(sq); |
e43473b7 | 479 | |
69df0ab0 TH |
480 | /* is the next dispatch time in the future? */ |
481 | if (time_after(sq->first_pending_disptime, jiffies)) { | |
482 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); | |
483 | return; | |
484 | } | |
485 | ||
486 | /* kick immediate execution */ | |
487 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
e43473b7 VG |
488 | } |
489 | ||
0f3457f6 | 490 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
491 | { |
492 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 493 | tg->io_disp[rw] = 0; |
e43473b7 VG |
494 | tg->slice_start[rw] = jiffies; |
495 | tg->slice_end[rw] = jiffies + throtl_slice; | |
fda6f272 TH |
496 | throtl_log(&tg->service_queue, |
497 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
498 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
499 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
500 | } |
501 | ||
0f3457f6 TH |
502 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
503 | unsigned long jiffy_end) | |
d1ae8ffd VG |
504 | { |
505 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
506 | } | |
507 | ||
0f3457f6 TH |
508 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
509 | unsigned long jiffy_end) | |
e43473b7 VG |
510 | { |
511 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
fda6f272 TH |
512 | throtl_log(&tg->service_queue, |
513 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
514 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
515 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
516 | } |
517 | ||
518 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 519 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
520 | { |
521 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
522 | return 0; | |
523 | ||
524 | return 1; | |
525 | } | |
526 | ||
527 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 528 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 529 | { |
3aad5d3e VG |
530 | unsigned long nr_slices, time_elapsed, io_trim; |
531 | u64 bytes_trim, tmp; | |
e43473b7 VG |
532 | |
533 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
534 | ||
535 | /* | |
536 | * If bps are unlimited (-1), then time slice don't get | |
537 | * renewed. Don't try to trim the slice if slice is used. A new | |
538 | * slice will start when appropriate. | |
539 | */ | |
0f3457f6 | 540 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
541 | return; |
542 | ||
d1ae8ffd VG |
543 | /* |
544 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
545 | * that initially cgroup limit was very low resulting in high | |
546 | * slice_end, but later limit was bumped up and bio was dispached | |
547 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
548 | * is bad because it does not allow new slice to start. | |
549 | */ | |
550 | ||
0f3457f6 | 551 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); |
d1ae8ffd | 552 | |
e43473b7 VG |
553 | time_elapsed = jiffies - tg->slice_start[rw]; |
554 | ||
555 | nr_slices = time_elapsed / throtl_slice; | |
556 | ||
557 | if (!nr_slices) | |
558 | return; | |
3aad5d3e VG |
559 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
560 | do_div(tmp, HZ); | |
561 | bytes_trim = tmp; | |
e43473b7 | 562 | |
8e89d13f | 563 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 564 | |
8e89d13f | 565 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
566 | return; |
567 | ||
568 | if (tg->bytes_disp[rw] >= bytes_trim) | |
569 | tg->bytes_disp[rw] -= bytes_trim; | |
570 | else | |
571 | tg->bytes_disp[rw] = 0; | |
572 | ||
8e89d13f VG |
573 | if (tg->io_disp[rw] >= io_trim) |
574 | tg->io_disp[rw] -= io_trim; | |
575 | else | |
576 | tg->io_disp[rw] = 0; | |
577 | ||
e43473b7 VG |
578 | tg->slice_start[rw] += nr_slices * throtl_slice; |
579 | ||
fda6f272 TH |
580 | throtl_log(&tg->service_queue, |
581 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
582 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
583 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
e43473b7 VG |
584 | } |
585 | ||
0f3457f6 TH |
586 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
587 | unsigned long *wait) | |
e43473b7 VG |
588 | { |
589 | bool rw = bio_data_dir(bio); | |
8e89d13f | 590 | unsigned int io_allowed; |
e43473b7 | 591 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 592 | u64 tmp; |
e43473b7 | 593 | |
8e89d13f | 594 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 595 | |
8e89d13f VG |
596 | /* Slice has just started. Consider one slice interval */ |
597 | if (!jiffy_elapsed) | |
598 | jiffy_elapsed_rnd = throtl_slice; | |
599 | ||
600 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
601 | ||
c49c06e4 VG |
602 | /* |
603 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
604 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
605 | * will allow dispatch after 1 second and after that slice should | |
606 | * have been trimmed. | |
607 | */ | |
608 | ||
609 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
610 | do_div(tmp, HZ); | |
611 | ||
612 | if (tmp > UINT_MAX) | |
613 | io_allowed = UINT_MAX; | |
614 | else | |
615 | io_allowed = tmp; | |
8e89d13f VG |
616 | |
617 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
618 | if (wait) |
619 | *wait = 0; | |
620 | return 1; | |
621 | } | |
622 | ||
8e89d13f VG |
623 | /* Calc approx time to dispatch */ |
624 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
625 | ||
626 | if (jiffy_wait > jiffy_elapsed) | |
627 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
628 | else | |
629 | jiffy_wait = 1; | |
630 | ||
631 | if (wait) | |
632 | *wait = jiffy_wait; | |
633 | return 0; | |
634 | } | |
635 | ||
0f3457f6 TH |
636 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
637 | unsigned long *wait) | |
8e89d13f VG |
638 | { |
639 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 640 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 641 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
642 | |
643 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
644 | ||
645 | /* Slice has just started. Consider one slice interval */ | |
646 | if (!jiffy_elapsed) | |
647 | jiffy_elapsed_rnd = throtl_slice; | |
648 | ||
649 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
650 | ||
5e901a2b VG |
651 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
652 | do_div(tmp, HZ); | |
3aad5d3e | 653 | bytes_allowed = tmp; |
e43473b7 VG |
654 | |
655 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
656 | if (wait) | |
657 | *wait = 0; | |
658 | return 1; | |
659 | } | |
660 | ||
661 | /* Calc approx time to dispatch */ | |
662 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
663 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
664 | ||
665 | if (!jiffy_wait) | |
666 | jiffy_wait = 1; | |
667 | ||
668 | /* | |
669 | * This wait time is without taking into consideration the rounding | |
670 | * up we did. Add that time also. | |
671 | */ | |
672 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
673 | if (wait) |
674 | *wait = jiffy_wait; | |
8e89d13f VG |
675 | return 0; |
676 | } | |
677 | ||
af75cd3c VG |
678 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
679 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
680 | return 1; | |
681 | return 0; | |
682 | } | |
683 | ||
8e89d13f VG |
684 | /* |
685 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
686 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
687 | */ | |
0f3457f6 TH |
688 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
689 | unsigned long *wait) | |
8e89d13f VG |
690 | { |
691 | bool rw = bio_data_dir(bio); | |
692 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
693 | ||
694 | /* | |
695 | * Currently whole state machine of group depends on first bio | |
696 | * queued in the group bio list. So one should not be calling | |
697 | * this function with a different bio if there are other bios | |
698 | * queued. | |
699 | */ | |
73f0d49a TH |
700 | BUG_ON(tg->service_queue.nr_queued[rw] && |
701 | bio != bio_list_peek(&tg->service_queue.bio_lists[rw])); | |
e43473b7 | 702 | |
8e89d13f VG |
703 | /* If tg->bps = -1, then BW is unlimited */ |
704 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
705 | if (wait) | |
706 | *wait = 0; | |
707 | return 1; | |
708 | } | |
709 | ||
710 | /* | |
711 | * If previous slice expired, start a new one otherwise renew/extend | |
712 | * existing slice to make sure it is at least throtl_slice interval | |
713 | * long since now. | |
714 | */ | |
0f3457f6 TH |
715 | if (throtl_slice_used(tg, rw)) |
716 | throtl_start_new_slice(tg, rw); | |
8e89d13f VG |
717 | else { |
718 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
0f3457f6 | 719 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); |
8e89d13f VG |
720 | } |
721 | ||
0f3457f6 TH |
722 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
723 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
724 | if (wait) |
725 | *wait = 0; | |
726 | return 1; | |
727 | } | |
728 | ||
729 | max_wait = max(bps_wait, iops_wait); | |
730 | ||
731 | if (wait) | |
732 | *wait = max_wait; | |
733 | ||
734 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 735 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
736 | |
737 | return 0; | |
738 | } | |
739 | ||
3c798398 | 740 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
741 | int rw) |
742 | { | |
8a3d2615 TH |
743 | struct throtl_grp *tg = blkg_to_tg(blkg); |
744 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
745 | unsigned long flags; |
746 | ||
747 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 748 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
749 | return; |
750 | ||
751 | /* | |
752 | * Disabling interrupts to provide mutual exclusion between two | |
753 | * writes on same cpu. It probably is not needed for 64bit. Not | |
754 | * optimizing that case yet. | |
755 | */ | |
756 | local_irq_save(flags); | |
757 | ||
8a3d2615 | 758 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 759 | |
629ed0b1 TH |
760 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
761 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
762 | ||
763 | local_irq_restore(flags); | |
764 | } | |
765 | ||
e43473b7 VG |
766 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
767 | { | |
768 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
769 | |
770 | /* Charge the bio to the group */ | |
771 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 772 | tg->io_disp[rw]++; |
e43473b7 | 773 | |
2a0f61e6 TH |
774 | /* |
775 | * REQ_THROTTLED is used to prevent the same bio to be throttled | |
776 | * more than once as a throttled bio will go through blk-throtl the | |
777 | * second time when it eventually gets issued. Set it when a bio | |
778 | * is being charged to a tg. | |
779 | * | |
780 | * Dispatch stats aren't recursive and each @bio should only be | |
781 | * accounted by the @tg it was originally associated with. Let's | |
782 | * update the stats when setting REQ_THROTTLED for the first time | |
783 | * which is guaranteed to be for the @bio's original tg. | |
784 | */ | |
785 | if (!(bio->bi_rw & REQ_THROTTLED)) { | |
786 | bio->bi_rw |= REQ_THROTTLED; | |
787 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, | |
788 | bio->bi_rw); | |
789 | } | |
e43473b7 VG |
790 | } |
791 | ||
77216b04 | 792 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_grp *tg) |
e43473b7 | 793 | { |
73f0d49a | 794 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
795 | bool rw = bio_data_dir(bio); |
796 | ||
0e9f4164 TH |
797 | /* |
798 | * If @tg doesn't currently have any bios queued in the same | |
799 | * direction, queueing @bio can change when @tg should be | |
800 | * dispatched. Mark that @tg was empty. This is automatically | |
801 | * cleaered on the next tg_update_disptime(). | |
802 | */ | |
803 | if (!sq->nr_queued[rw]) | |
804 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
805 | ||
73f0d49a | 806 | bio_list_add(&sq->bio_lists[rw], bio); |
e43473b7 | 807 | /* Take a bio reference on tg */ |
1adaf3dd | 808 | blkg_get(tg_to_blkg(tg)); |
73f0d49a | 809 | sq->nr_queued[rw]++; |
e2d57e60 | 810 | tg->td->nr_queued[rw]++; |
77216b04 | 811 | throtl_enqueue_tg(tg); |
e43473b7 VG |
812 | } |
813 | ||
77216b04 | 814 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 815 | { |
73f0d49a | 816 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
817 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
818 | struct bio *bio; | |
819 | ||
73f0d49a | 820 | if ((bio = bio_list_peek(&sq->bio_lists[READ]))) |
0f3457f6 | 821 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 822 | |
73f0d49a | 823 | if ((bio = bio_list_peek(&sq->bio_lists[WRITE]))) |
0f3457f6 | 824 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
825 | |
826 | min_wait = min(read_wait, write_wait); | |
827 | disptime = jiffies + min_wait; | |
828 | ||
e43473b7 | 829 | /* Update dispatch time */ |
77216b04 | 830 | throtl_dequeue_tg(tg); |
e43473b7 | 831 | tg->disptime = disptime; |
77216b04 | 832 | throtl_enqueue_tg(tg); |
0e9f4164 TH |
833 | |
834 | /* see throtl_add_bio_tg() */ | |
835 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
836 | } |
837 | ||
77216b04 | 838 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 839 | { |
73f0d49a | 840 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
841 | struct bio *bio; |
842 | ||
73f0d49a TH |
843 | bio = bio_list_pop(&sq->bio_lists[rw]); |
844 | sq->nr_queued[rw]--; | |
1adaf3dd TH |
845 | /* Drop bio reference on blkg */ |
846 | blkg_put(tg_to_blkg(tg)); | |
e43473b7 | 847 | |
0f3457f6 TH |
848 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
849 | tg->td->nr_queued[rw]--; | |
e43473b7 VG |
850 | |
851 | throtl_charge_bio(tg, bio); | |
77216b04 | 852 | bio_list_add(&sq->parent_sq->bio_lists[rw], bio); |
e43473b7 | 853 | |
0f3457f6 | 854 | throtl_trim_slice(tg, rw); |
e43473b7 VG |
855 | } |
856 | ||
77216b04 | 857 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 858 | { |
73f0d49a | 859 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
860 | unsigned int nr_reads = 0, nr_writes = 0; |
861 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 862 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
863 | struct bio *bio; |
864 | ||
865 | /* Try to dispatch 75% READS and 25% WRITES */ | |
866 | ||
73f0d49a | 867 | while ((bio = bio_list_peek(&sq->bio_lists[READ])) && |
0f3457f6 | 868 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 869 | |
77216b04 | 870 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
871 | nr_reads++; |
872 | ||
873 | if (nr_reads >= max_nr_reads) | |
874 | break; | |
875 | } | |
876 | ||
73f0d49a | 877 | while ((bio = bio_list_peek(&sq->bio_lists[WRITE])) && |
0f3457f6 | 878 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 879 | |
77216b04 | 880 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
881 | nr_writes++; |
882 | ||
883 | if (nr_writes >= max_nr_writes) | |
884 | break; | |
885 | } | |
886 | ||
887 | return nr_reads + nr_writes; | |
888 | } | |
889 | ||
651930bc | 890 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
891 | { |
892 | unsigned int nr_disp = 0; | |
e43473b7 VG |
893 | |
894 | while (1) { | |
73f0d49a TH |
895 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
896 | struct throtl_service_queue *sq = &tg->service_queue; | |
e43473b7 VG |
897 | |
898 | if (!tg) | |
899 | break; | |
900 | ||
901 | if (time_before(jiffies, tg->disptime)) | |
902 | break; | |
903 | ||
77216b04 | 904 | throtl_dequeue_tg(tg); |
e43473b7 | 905 | |
77216b04 | 906 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 907 | |
73f0d49a | 908 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
77216b04 | 909 | tg_update_disptime(tg); |
e43473b7 VG |
910 | |
911 | if (nr_disp >= throtl_quantum) | |
912 | break; | |
913 | } | |
914 | ||
915 | return nr_disp; | |
916 | } | |
917 | ||
69df0ab0 TH |
918 | static void throtl_pending_timer_fn(unsigned long arg) |
919 | { | |
920 | struct throtl_service_queue *sq = (void *)arg; | |
921 | struct throtl_data *td = sq_to_td(sq); | |
922 | ||
923 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
924 | } | |
925 | ||
cb76199c TH |
926 | /* work function to dispatch throttled bios */ |
927 | void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
e43473b7 | 928 | { |
69df0ab0 TH |
929 | struct throtl_data *td = container_of(work, struct throtl_data, |
930 | dispatch_work); | |
651930bc | 931 | struct throtl_service_queue *sq = &td->service_queue; |
cb76199c | 932 | struct request_queue *q = td->queue; |
e43473b7 VG |
933 | unsigned int nr_disp = 0; |
934 | struct bio_list bio_list_on_stack; | |
935 | struct bio *bio; | |
69d60eb9 | 936 | struct blk_plug plug; |
651930bc | 937 | int rw; |
e43473b7 VG |
938 | |
939 | spin_lock_irq(q->queue_lock); | |
940 | ||
e43473b7 VG |
941 | bio_list_init(&bio_list_on_stack); |
942 | ||
fda6f272 | 943 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", |
6a525600 TH |
944 | td->nr_queued[READ] + td->nr_queued[WRITE], |
945 | td->nr_queued[READ], td->nr_queued[WRITE]); | |
e43473b7 | 946 | |
651930bc | 947 | nr_disp = throtl_select_dispatch(sq); |
e43473b7 | 948 | |
651930bc TH |
949 | if (nr_disp) { |
950 | for (rw = READ; rw <= WRITE; rw++) { | |
951 | bio_list_merge(&bio_list_on_stack, &sq->bio_lists[rw]); | |
952 | bio_list_init(&sq->bio_lists[rw]); | |
953 | } | |
fda6f272 | 954 | throtl_log(sq, "bios disp=%u", nr_disp); |
651930bc | 955 | } |
e43473b7 | 956 | |
69df0ab0 | 957 | throtl_schedule_next_dispatch(sq); |
6a525600 | 958 | |
e43473b7 VG |
959 | spin_unlock_irq(q->queue_lock); |
960 | ||
961 | /* | |
962 | * If we dispatched some requests, unplug the queue to make sure | |
963 | * immediate dispatch | |
964 | */ | |
965 | if (nr_disp) { | |
69d60eb9 | 966 | blk_start_plug(&plug); |
e43473b7 VG |
967 | while((bio = bio_list_pop(&bio_list_on_stack))) |
968 | generic_make_request(bio); | |
69d60eb9 | 969 | blk_finish_plug(&plug); |
e43473b7 | 970 | } |
e43473b7 VG |
971 | } |
972 | ||
f95a04af TH |
973 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
974 | struct blkg_policy_data *pd, int off) | |
41b38b6d | 975 | { |
f95a04af | 976 | struct throtl_grp *tg = pd_to_tg(pd); |
41b38b6d TH |
977 | struct blkg_rwstat rwstat = { }, tmp; |
978 | int i, cpu; | |
979 | ||
980 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 981 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
982 | |
983 | tmp = blkg_rwstat_read((void *)sc + off); | |
984 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
985 | rwstat.cnt[i] += tmp.cnt[i]; | |
986 | } | |
987 | ||
f95a04af | 988 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
41b38b6d TH |
989 | } |
990 | ||
8a3d2615 TH |
991 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
992 | struct seq_file *sf) | |
41b38b6d | 993 | { |
3c798398 | 994 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
41b38b6d | 995 | |
3c798398 | 996 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 997 | cft->private, true); |
41b38b6d TH |
998 | return 0; |
999 | } | |
1000 | ||
f95a04af TH |
1001 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
1002 | int off) | |
60c2bc2d | 1003 | { |
f95a04af TH |
1004 | struct throtl_grp *tg = pd_to_tg(pd); |
1005 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 1006 | |
af133ceb | 1007 | if (v == -1) |
60c2bc2d | 1008 | return 0; |
f95a04af | 1009 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
1010 | } |
1011 | ||
f95a04af TH |
1012 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
1013 | int off) | |
e43473b7 | 1014 | { |
f95a04af TH |
1015 | struct throtl_grp *tg = pd_to_tg(pd); |
1016 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 1017 | |
af133ceb TH |
1018 | if (v == -1) |
1019 | return 0; | |
f95a04af | 1020 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
1021 | } |
1022 | ||
af133ceb TH |
1023 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1024 | struct seq_file *sf) | |
8e89d13f | 1025 | { |
3c798398 TH |
1026 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, |
1027 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 1028 | return 0; |
8e89d13f VG |
1029 | } |
1030 | ||
af133ceb TH |
1031 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1032 | struct seq_file *sf) | |
8e89d13f | 1033 | { |
3c798398 TH |
1034 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, |
1035 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 1036 | return 0; |
60c2bc2d TH |
1037 | } |
1038 | ||
af133ceb TH |
1039 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
1040 | bool is_u64) | |
60c2bc2d | 1041 | { |
3c798398 | 1042 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
60c2bc2d | 1043 | struct blkg_conf_ctx ctx; |
af133ceb | 1044 | struct throtl_grp *tg; |
69df0ab0 | 1045 | struct throtl_service_queue *sq; |
60c2bc2d TH |
1046 | int ret; |
1047 | ||
3c798398 | 1048 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
1049 | if (ret) |
1050 | return ret; | |
1051 | ||
af133ceb | 1052 | tg = blkg_to_tg(ctx.blkg); |
69df0ab0 | 1053 | sq = &tg->service_queue; |
af133ceb | 1054 | |
a2b1693b TH |
1055 | if (!ctx.v) |
1056 | ctx.v = -1; | |
af133ceb | 1057 | |
a2b1693b TH |
1058 | if (is_u64) |
1059 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
1060 | else | |
1061 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 1062 | |
fda6f272 TH |
1063 | throtl_log(&tg->service_queue, |
1064 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
1065 | tg->bps[READ], tg->bps[WRITE], | |
1066 | tg->iops[READ], tg->iops[WRITE]); | |
632b4493 TH |
1067 | |
1068 | /* | |
1069 | * We're already holding queue_lock and know @tg is valid. Let's | |
1070 | * apply the new config directly. | |
1071 | * | |
1072 | * Restart the slices for both READ and WRITES. It might happen | |
1073 | * that a group's limit are dropped suddenly and we don't want to | |
1074 | * account recently dispatched IO with new low rate. | |
1075 | */ | |
0f3457f6 TH |
1076 | throtl_start_new_slice(tg, 0); |
1077 | throtl_start_new_slice(tg, 1); | |
632b4493 | 1078 | |
5b2c16aa | 1079 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 1080 | tg_update_disptime(tg); |
69df0ab0 | 1081 | throtl_schedule_next_dispatch(sq->parent_sq); |
632b4493 | 1082 | } |
60c2bc2d TH |
1083 | |
1084 | blkg_conf_finish(&ctx); | |
a2b1693b | 1085 | return 0; |
8e89d13f VG |
1086 | } |
1087 | ||
af133ceb TH |
1088 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1089 | const char *buf) | |
60c2bc2d | 1090 | { |
af133ceb | 1091 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
1092 | } |
1093 | ||
af133ceb TH |
1094 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1095 | const char *buf) | |
60c2bc2d | 1096 | { |
af133ceb | 1097 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
1098 | } |
1099 | ||
1100 | static struct cftype throtl_files[] = { | |
1101 | { | |
1102 | .name = "throttle.read_bps_device", | |
af133ceb TH |
1103 | .private = offsetof(struct throtl_grp, bps[READ]), |
1104 | .read_seq_string = tg_print_conf_u64, | |
1105 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1106 | .max_write_len = 256, |
1107 | }, | |
1108 | { | |
1109 | .name = "throttle.write_bps_device", | |
af133ceb TH |
1110 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
1111 | .read_seq_string = tg_print_conf_u64, | |
1112 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1113 | .max_write_len = 256, |
1114 | }, | |
1115 | { | |
1116 | .name = "throttle.read_iops_device", | |
af133ceb TH |
1117 | .private = offsetof(struct throtl_grp, iops[READ]), |
1118 | .read_seq_string = tg_print_conf_uint, | |
1119 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1120 | .max_write_len = 256, |
1121 | }, | |
1122 | { | |
1123 | .name = "throttle.write_iops_device", | |
af133ceb TH |
1124 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
1125 | .read_seq_string = tg_print_conf_uint, | |
1126 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1127 | .max_write_len = 256, |
1128 | }, | |
1129 | { | |
1130 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1131 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1132 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1133 | }, |
1134 | { | |
1135 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1136 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1137 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1138 | }, |
1139 | { } /* terminate */ | |
1140 | }; | |
1141 | ||
da527770 | 1142 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1143 | { |
1144 | struct throtl_data *td = q->td; | |
1145 | ||
69df0ab0 | 1146 | cancel_work_sync(&td->dispatch_work); |
e43473b7 VG |
1147 | } |
1148 | ||
3c798398 | 1149 | static struct blkcg_policy blkcg_policy_throtl = { |
f9fcc2d3 TH |
1150 | .pd_size = sizeof(struct throtl_grp), |
1151 | .cftypes = throtl_files, | |
1152 | ||
1153 | .pd_init_fn = throtl_pd_init, | |
1154 | .pd_exit_fn = throtl_pd_exit, | |
1155 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 VG |
1156 | }; |
1157 | ||
bc16a4f9 | 1158 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1159 | { |
1160 | struct throtl_data *td = q->td; | |
1161 | struct throtl_grp *tg; | |
73f0d49a | 1162 | struct throtl_service_queue *sq; |
0e9f4164 | 1163 | bool rw = bio_data_dir(bio); |
3c798398 | 1164 | struct blkcg *blkcg; |
bc16a4f9 | 1165 | bool throttled = false; |
e43473b7 | 1166 | |
2a0f61e6 TH |
1167 | /* see throtl_charge_bio() */ |
1168 | if (bio->bi_rw & REQ_THROTTLED) | |
bc16a4f9 | 1169 | goto out; |
e43473b7 | 1170 | |
af75cd3c VG |
1171 | /* |
1172 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1173 | * basic fields like stats and io rates. If a group has no rules, | |
1174 | * just update the dispatch stats in lockless manner and return. | |
1175 | */ | |
af75cd3c | 1176 | rcu_read_lock(); |
3c798398 | 1177 | blkcg = bio_blkcg(bio); |
cd1604fa | 1178 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1179 | if (tg) { |
af75cd3c | 1180 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1181 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1182 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1183 | goto out_unlock_rcu; |
af75cd3c VG |
1184 | } |
1185 | } | |
af75cd3c VG |
1186 | |
1187 | /* | |
1188 | * Either group has not been allocated yet or it is not an unlimited | |
1189 | * IO group | |
1190 | */ | |
e43473b7 | 1191 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1192 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1193 | if (unlikely(!tg)) |
1194 | goto out_unlock; | |
f469a7b4 | 1195 | |
73f0d49a TH |
1196 | sq = &tg->service_queue; |
1197 | ||
0e9f4164 TH |
1198 | /* throtl is FIFO - if other bios are already queued, should queue */ |
1199 | if (sq->nr_queued[rw]) | |
e43473b7 | 1200 | goto queue_bio; |
de701c74 | 1201 | |
e43473b7 | 1202 | /* Bio is with-in rate limit of group */ |
0f3457f6 | 1203 | if (tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1204 | throtl_charge_bio(tg, bio); |
04521db0 VG |
1205 | |
1206 | /* | |
1207 | * We need to trim slice even when bios are not being queued | |
1208 | * otherwise it might happen that a bio is not queued for | |
1209 | * a long time and slice keeps on extending and trim is not | |
1210 | * called for a long time. Now if limits are reduced suddenly | |
1211 | * we take into account all the IO dispatched so far at new | |
1212 | * low rate and * newly queued IO gets a really long dispatch | |
1213 | * time. | |
1214 | * | |
1215 | * So keep on trimming slice even if bio is not queued. | |
1216 | */ | |
0f3457f6 | 1217 | throtl_trim_slice(tg, rw); |
bc16a4f9 | 1218 | goto out_unlock; |
e43473b7 VG |
1219 | } |
1220 | ||
1221 | queue_bio: | |
fda6f272 TH |
1222 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
1223 | rw == READ ? 'R' : 'W', | |
1224 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], | |
1225 | tg->io_disp[rw], tg->iops[rw], | |
1226 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
e43473b7 | 1227 | |
671058fb | 1228 | bio_associate_current(bio); |
77216b04 | 1229 | throtl_add_bio_tg(bio, tg); |
bc16a4f9 | 1230 | throttled = true; |
e43473b7 | 1231 | |
0e9f4164 TH |
1232 | /* update @tg's dispatch time if @tg was empty before @bio */ |
1233 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
77216b04 | 1234 | tg_update_disptime(tg); |
69df0ab0 | 1235 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq); |
e43473b7 VG |
1236 | } |
1237 | ||
bc16a4f9 | 1238 | out_unlock: |
e43473b7 | 1239 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1240 | out_unlock_rcu: |
1241 | rcu_read_unlock(); | |
bc16a4f9 | 1242 | out: |
2a0f61e6 TH |
1243 | /* |
1244 | * As multiple blk-throtls may stack in the same issue path, we | |
1245 | * don't want bios to leave with the flag set. Clear the flag if | |
1246 | * being issued. | |
1247 | */ | |
1248 | if (!throttled) | |
1249 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1250 | return throttled; |
e43473b7 VG |
1251 | } |
1252 | ||
c9a929dd TH |
1253 | /** |
1254 | * blk_throtl_drain - drain throttled bios | |
1255 | * @q: request_queue to drain throttled bios for | |
1256 | * | |
1257 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1258 | */ | |
1259 | void blk_throtl_drain(struct request_queue *q) | |
1260 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1261 | { | |
1262 | struct throtl_data *td = q->td; | |
0049af73 | 1263 | struct throtl_service_queue *parent_sq = &td->service_queue; |
c9a929dd | 1264 | struct throtl_grp *tg; |
c9a929dd | 1265 | struct bio *bio; |
651930bc | 1266 | int rw; |
c9a929dd | 1267 | |
8bcb6c7d | 1268 | queue_lockdep_assert_held(q); |
c9a929dd | 1269 | |
0049af73 | 1270 | while ((tg = throtl_rb_first(parent_sq))) { |
73f0d49a TH |
1271 | struct throtl_service_queue *sq = &tg->service_queue; |
1272 | ||
77216b04 | 1273 | throtl_dequeue_tg(tg); |
c9a929dd | 1274 | |
73f0d49a | 1275 | while ((bio = bio_list_peek(&sq->bio_lists[READ]))) |
77216b04 | 1276 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
73f0d49a | 1277 | while ((bio = bio_list_peek(&sq->bio_lists[WRITE]))) |
77216b04 | 1278 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
c9a929dd TH |
1279 | } |
1280 | spin_unlock_irq(q->queue_lock); | |
1281 | ||
651930bc TH |
1282 | for (rw = READ; rw <= WRITE; rw++) |
1283 | while ((bio = bio_list_pop(&parent_sq->bio_lists[rw]))) | |
1284 | generic_make_request(bio); | |
c9a929dd TH |
1285 | |
1286 | spin_lock_irq(q->queue_lock); | |
1287 | } | |
1288 | ||
e43473b7 VG |
1289 | int blk_throtl_init(struct request_queue *q) |
1290 | { | |
1291 | struct throtl_data *td; | |
a2b1693b | 1292 | int ret; |
e43473b7 VG |
1293 | |
1294 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1295 | if (!td) | |
1296 | return -ENOMEM; | |
1297 | ||
69df0ab0 | 1298 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
77216b04 | 1299 | throtl_service_queue_init(&td->service_queue, NULL); |
e43473b7 | 1300 | |
cd1604fa | 1301 | q->td = td; |
29b12589 | 1302 | td->queue = q; |
02977e4a | 1303 | |
a2b1693b | 1304 | /* activate policy */ |
3c798398 | 1305 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1306 | if (ret) |
f51b802c | 1307 | kfree(td); |
a2b1693b | 1308 | return ret; |
e43473b7 VG |
1309 | } |
1310 | ||
1311 | void blk_throtl_exit(struct request_queue *q) | |
1312 | { | |
c875f4d0 | 1313 | BUG_ON(!q->td); |
da527770 | 1314 | throtl_shutdown_wq(q); |
3c798398 | 1315 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1316 | kfree(q->td); |
e43473b7 VG |
1317 | } |
1318 | ||
1319 | static int __init throtl_init(void) | |
1320 | { | |
450adcbe VG |
1321 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1322 | if (!kthrotld_workqueue) | |
1323 | panic("Failed to create kthrotld\n"); | |
1324 | ||
3c798398 | 1325 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1326 | } |
1327 | ||
1328 | module_init(throtl_init); |