]>
Commit | Line | Data |
---|---|---|
e43473b7 VG |
1 | /* |
2 | * Interface for controlling IO bandwidth on a request queue | |
3 | * | |
4 | * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com> | |
5 | */ | |
6 | ||
7 | #include <linux/module.h> | |
8 | #include <linux/slab.h> | |
9 | #include <linux/blkdev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blktrace_api.h> | |
12 | #include "blk-cgroup.h" | |
bc9fcbf9 | 13 | #include "blk.h" |
e43473b7 VG |
14 | |
15 | /* Max dispatch from a group in 1 round */ | |
16 | static int throtl_grp_quantum = 8; | |
17 | ||
18 | /* Total max dispatch from all groups in one round */ | |
19 | static int throtl_quantum = 32; | |
20 | ||
21 | /* Throttling is performed over 100ms slice and after that slice is renewed */ | |
22 | static unsigned long throtl_slice = HZ/10; /* 100 ms */ | |
23 | ||
3c798398 | 24 | static struct blkcg_policy blkcg_policy_throtl; |
0381411e | 25 | |
450adcbe VG |
26 | /* A workqueue to queue throttle related work */ |
27 | static struct workqueue_struct *kthrotld_workqueue; | |
450adcbe | 28 | |
c9e0332e TH |
29 | struct throtl_service_queue { |
30 | struct rb_root pending_tree; /* RB tree of active tgs */ | |
31 | struct rb_node *first_pending; /* first node in the tree */ | |
32 | unsigned int nr_pending; /* # queued in the tree */ | |
33 | unsigned long first_pending_disptime; /* disptime of the first tg */ | |
e43473b7 VG |
34 | }; |
35 | ||
c9e0332e TH |
36 | #define THROTL_SERVICE_QUEUE_INITIALIZER \ |
37 | (struct throtl_service_queue){ .pending_tree = RB_ROOT } | |
e43473b7 | 38 | |
5b2c16aa TH |
39 | enum tg_state_flags { |
40 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
41 | }; | |
42 | ||
e43473b7 VG |
43 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
44 | ||
8a3d2615 TH |
45 | /* Per-cpu group stats */ |
46 | struct tg_stats_cpu { | |
47 | /* total bytes transferred */ | |
48 | struct blkg_rwstat service_bytes; | |
49 | /* total IOs serviced, post merge */ | |
50 | struct blkg_rwstat serviced; | |
51 | }; | |
52 | ||
e43473b7 | 53 | struct throtl_grp { |
f95a04af TH |
54 | /* must be the first member */ |
55 | struct blkg_policy_data pd; | |
56 | ||
c9e0332e | 57 | /* active throtl group service_queue member */ |
e43473b7 VG |
58 | struct rb_node rb_node; |
59 | ||
0f3457f6 TH |
60 | /* throtl_data this group belongs to */ |
61 | struct throtl_data *td; | |
62 | ||
e43473b7 VG |
63 | /* |
64 | * Dispatch time in jiffies. This is the estimated time when group | |
65 | * will unthrottle and is ready to dispatch more bio. It is used as | |
66 | * key to sort active groups in service tree. | |
67 | */ | |
68 | unsigned long disptime; | |
69 | ||
e43473b7 VG |
70 | unsigned int flags; |
71 | ||
72 | /* Two lists for READ and WRITE */ | |
73 | struct bio_list bio_lists[2]; | |
74 | ||
75 | /* Number of queued bios on READ and WRITE lists */ | |
76 | unsigned int nr_queued[2]; | |
77 | ||
78 | /* bytes per second rate limits */ | |
79 | uint64_t bps[2]; | |
80 | ||
8e89d13f VG |
81 | /* IOPS limits */ |
82 | unsigned int iops[2]; | |
83 | ||
e43473b7 VG |
84 | /* Number of bytes disptached in current slice */ |
85 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
86 | /* Number of bio's dispatched in current slice */ |
87 | unsigned int io_disp[2]; | |
e43473b7 VG |
88 | |
89 | /* When did we start a new slice */ | |
90 | unsigned long slice_start[2]; | |
91 | unsigned long slice_end[2]; | |
fe071437 | 92 | |
8a3d2615 TH |
93 | /* Per cpu stats pointer */ |
94 | struct tg_stats_cpu __percpu *stats_cpu; | |
95 | ||
96 | /* List of tgs waiting for per cpu stats memory to be allocated */ | |
97 | struct list_head stats_alloc_node; | |
e43473b7 VG |
98 | }; |
99 | ||
100 | struct throtl_data | |
101 | { | |
e43473b7 | 102 | /* service tree for active throtl groups */ |
c9e0332e | 103 | struct throtl_service_queue service_queue; |
e43473b7 | 104 | |
e43473b7 VG |
105 | struct request_queue *queue; |
106 | ||
107 | /* Total Number of queued bios on READ and WRITE lists */ | |
108 | unsigned int nr_queued[2]; | |
109 | ||
110 | /* | |
02977e4a | 111 | * number of total undestroyed groups |
e43473b7 VG |
112 | */ |
113 | unsigned int nr_undestroyed_grps; | |
114 | ||
115 | /* Work for dispatching throttled bios */ | |
cb76199c | 116 | struct delayed_work dispatch_work; |
e43473b7 VG |
117 | }; |
118 | ||
8a3d2615 TH |
119 | /* list and work item to allocate percpu group stats */ |
120 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
121 | static LIST_HEAD(tg_stats_alloc_list); | |
122 | ||
123 | static void tg_stats_alloc_fn(struct work_struct *); | |
124 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
125 | ||
f95a04af TH |
126 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
127 | { | |
128 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
129 | } | |
130 | ||
3c798398 | 131 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 132 | { |
f95a04af | 133 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
134 | } |
135 | ||
3c798398 | 136 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 137 | { |
f95a04af | 138 | return pd_to_blkg(&tg->pd); |
0381411e TH |
139 | } |
140 | ||
03d8e111 TH |
141 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) |
142 | { | |
143 | return blkg_to_tg(td->queue->root_blkg); | |
144 | } | |
145 | ||
0f3457f6 | 146 | #define throtl_log_tg(tg, fmt, args...) do { \ |
54e7ed12 TH |
147 | char __pbuf[128]; \ |
148 | \ | |
149 | blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \ | |
0f3457f6 | 150 | blk_add_trace_msg((tg)->td->queue, "throtl %s " fmt, __pbuf, ##args); \ |
54e7ed12 | 151 | } while (0) |
e43473b7 VG |
152 | |
153 | #define throtl_log(td, fmt, args...) \ | |
154 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) | |
155 | ||
8a3d2615 TH |
156 | /* |
157 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
3b07e9ca | 158 | * system_wq once there are some groups on the alloc_list waiting for |
8a3d2615 TH |
159 | * allocation. |
160 | */ | |
161 | static void tg_stats_alloc_fn(struct work_struct *work) | |
162 | { | |
163 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
164 | struct delayed_work *dwork = to_delayed_work(work); | |
165 | bool empty = false; | |
166 | ||
167 | alloc_stats: | |
168 | if (!stats_cpu) { | |
169 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
170 | if (!stats_cpu) { | |
171 | /* allocation failed, try again after some time */ | |
3b07e9ca | 172 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); |
8a3d2615 TH |
173 | return; |
174 | } | |
175 | } | |
176 | ||
177 | spin_lock_irq(&tg_stats_alloc_lock); | |
178 | ||
179 | if (!list_empty(&tg_stats_alloc_list)) { | |
180 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
181 | struct throtl_grp, | |
182 | stats_alloc_node); | |
183 | swap(tg->stats_cpu, stats_cpu); | |
184 | list_del_init(&tg->stats_alloc_node); | |
185 | } | |
186 | ||
187 | empty = list_empty(&tg_stats_alloc_list); | |
188 | spin_unlock_irq(&tg_stats_alloc_lock); | |
189 | if (!empty) | |
190 | goto alloc_stats; | |
191 | } | |
192 | ||
3c798398 | 193 | static void throtl_pd_init(struct blkcg_gq *blkg) |
a29a171e | 194 | { |
0381411e | 195 | struct throtl_grp *tg = blkg_to_tg(blkg); |
ff26eaad | 196 | unsigned long flags; |
cd1604fa | 197 | |
a29a171e | 198 | RB_CLEAR_NODE(&tg->rb_node); |
0f3457f6 | 199 | tg->td = blkg->q->td; |
a29a171e VG |
200 | bio_list_init(&tg->bio_lists[0]); |
201 | bio_list_init(&tg->bio_lists[1]); | |
a29a171e | 202 | |
e56da7e2 TH |
203 | tg->bps[READ] = -1; |
204 | tg->bps[WRITE] = -1; | |
205 | tg->iops[READ] = -1; | |
206 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
207 | |
208 | /* | |
209 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
210 | * but percpu allocator can't be called from IO path. Queue tg on | |
211 | * tg_stats_alloc_list and allocate from work item. | |
212 | */ | |
ff26eaad | 213 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 214 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); |
3b07e9ca | 215 | schedule_delayed_work(&tg_stats_alloc_work, 0); |
ff26eaad | 216 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
217 | } |
218 | ||
3c798398 | 219 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
8a3d2615 TH |
220 | { |
221 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
ff26eaad | 222 | unsigned long flags; |
8a3d2615 | 223 | |
ff26eaad | 224 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 225 | list_del_init(&tg->stats_alloc_node); |
ff26eaad | 226 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
227 | |
228 | free_percpu(tg->stats_cpu); | |
229 | } | |
230 | ||
3c798398 | 231 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
8a3d2615 TH |
232 | { |
233 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
234 | int cpu; | |
235 | ||
236 | if (tg->stats_cpu == NULL) | |
237 | return; | |
238 | ||
239 | for_each_possible_cpu(cpu) { | |
240 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
241 | ||
242 | blkg_rwstat_reset(&sc->service_bytes); | |
243 | blkg_rwstat_reset(&sc->serviced); | |
244 | } | |
a29a171e VG |
245 | } |
246 | ||
3c798398 TH |
247 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
248 | struct blkcg *blkcg) | |
e43473b7 | 249 | { |
be2c6b19 | 250 | /* |
3c798398 TH |
251 | * This is the common case when there are no blkcgs. Avoid lookup |
252 | * in this case | |
cd1604fa | 253 | */ |
3c798398 | 254 | if (blkcg == &blkcg_root) |
03d8e111 | 255 | return td_root_tg(td); |
e43473b7 | 256 | |
e8989fae | 257 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
258 | } |
259 | ||
cd1604fa | 260 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
3c798398 | 261 | struct blkcg *blkcg) |
e43473b7 | 262 | { |
f469a7b4 | 263 | struct request_queue *q = td->queue; |
cd1604fa | 264 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 265 | |
f469a7b4 | 266 | /* |
3c798398 TH |
267 | * This is the common case when there are no blkcgs. Avoid lookup |
268 | * in this case | |
f469a7b4 | 269 | */ |
3c798398 | 270 | if (blkcg == &blkcg_root) { |
03d8e111 | 271 | tg = td_root_tg(td); |
cd1604fa | 272 | } else { |
3c798398 | 273 | struct blkcg_gq *blkg; |
f469a7b4 | 274 | |
3c96cb32 | 275 | blkg = blkg_lookup_create(blkcg, q); |
f469a7b4 | 276 | |
cd1604fa TH |
277 | /* if %NULL and @q is alive, fall back to root_tg */ |
278 | if (!IS_ERR(blkg)) | |
0381411e | 279 | tg = blkg_to_tg(blkg); |
3f3299d5 | 280 | else if (!blk_queue_dying(q)) |
03d8e111 | 281 | tg = td_root_tg(td); |
f469a7b4 VG |
282 | } |
283 | ||
e43473b7 VG |
284 | return tg; |
285 | } | |
286 | ||
0049af73 TH |
287 | static struct throtl_grp * |
288 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
289 | { |
290 | /* Service tree is empty */ | |
0049af73 | 291 | if (!parent_sq->nr_pending) |
e43473b7 VG |
292 | return NULL; |
293 | ||
0049af73 TH |
294 | if (!parent_sq->first_pending) |
295 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
e43473b7 | 296 | |
0049af73 TH |
297 | if (parent_sq->first_pending) |
298 | return rb_entry_tg(parent_sq->first_pending); | |
e43473b7 VG |
299 | |
300 | return NULL; | |
301 | } | |
302 | ||
303 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
304 | { | |
305 | rb_erase(n, root); | |
306 | RB_CLEAR_NODE(n); | |
307 | } | |
308 | ||
0049af73 TH |
309 | static void throtl_rb_erase(struct rb_node *n, |
310 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 311 | { |
0049af73 TH |
312 | if (parent_sq->first_pending == n) |
313 | parent_sq->first_pending = NULL; | |
314 | rb_erase_init(n, &parent_sq->pending_tree); | |
315 | --parent_sq->nr_pending; | |
e43473b7 VG |
316 | } |
317 | ||
0049af73 | 318 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
319 | { |
320 | struct throtl_grp *tg; | |
321 | ||
0049af73 | 322 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
323 | if (!tg) |
324 | return; | |
325 | ||
0049af73 | 326 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
327 | } |
328 | ||
0049af73 TH |
329 | static void tg_service_queue_add(struct throtl_grp *tg, |
330 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 331 | { |
0049af73 | 332 | struct rb_node **node = &parent_sq->pending_tree.rb_node; |
e43473b7 VG |
333 | struct rb_node *parent = NULL; |
334 | struct throtl_grp *__tg; | |
335 | unsigned long key = tg->disptime; | |
336 | int left = 1; | |
337 | ||
338 | while (*node != NULL) { | |
339 | parent = *node; | |
340 | __tg = rb_entry_tg(parent); | |
341 | ||
342 | if (time_before(key, __tg->disptime)) | |
343 | node = &parent->rb_left; | |
344 | else { | |
345 | node = &parent->rb_right; | |
346 | left = 0; | |
347 | } | |
348 | } | |
349 | ||
350 | if (left) | |
0049af73 | 351 | parent_sq->first_pending = &tg->rb_node; |
e43473b7 VG |
352 | |
353 | rb_link_node(&tg->rb_node, parent, node); | |
0049af73 | 354 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); |
e43473b7 VG |
355 | } |
356 | ||
0049af73 TH |
357 | static void __throtl_enqueue_tg(struct throtl_grp *tg, |
358 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 359 | { |
0049af73 | 360 | tg_service_queue_add(tg, parent_sq); |
5b2c16aa | 361 | tg->flags |= THROTL_TG_PENDING; |
0049af73 | 362 | parent_sq->nr_pending++; |
e43473b7 VG |
363 | } |
364 | ||
0049af73 TH |
365 | static void throtl_enqueue_tg(struct throtl_grp *tg, |
366 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 367 | { |
5b2c16aa | 368 | if (!(tg->flags & THROTL_TG_PENDING)) |
0049af73 | 369 | __throtl_enqueue_tg(tg, parent_sq); |
e43473b7 VG |
370 | } |
371 | ||
0049af73 TH |
372 | static void __throtl_dequeue_tg(struct throtl_grp *tg, |
373 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 374 | { |
0049af73 | 375 | throtl_rb_erase(&tg->rb_node, parent_sq); |
5b2c16aa | 376 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
377 | } |
378 | ||
0049af73 TH |
379 | static void throtl_dequeue_tg(struct throtl_grp *tg, |
380 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 381 | { |
5b2c16aa | 382 | if (tg->flags & THROTL_TG_PENDING) |
0049af73 | 383 | __throtl_dequeue_tg(tg, parent_sq); |
e43473b7 VG |
384 | } |
385 | ||
a9131a27 TH |
386 | /* Call with queue lock held */ |
387 | static void throtl_schedule_delayed_work(struct throtl_data *td, | |
388 | unsigned long delay) | |
389 | { | |
390 | struct delayed_work *dwork = &td->dispatch_work; | |
391 | ||
6a525600 TH |
392 | mod_delayed_work(kthrotld_workqueue, dwork, delay); |
393 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", delay, jiffies); | |
a9131a27 TH |
394 | } |
395 | ||
e43473b7 VG |
396 | static void throtl_schedule_next_dispatch(struct throtl_data *td) |
397 | { | |
c9e0332e | 398 | struct throtl_service_queue *sq = &td->service_queue; |
e43473b7 | 399 | |
6a525600 | 400 | /* any pending children left? */ |
c9e0332e | 401 | if (!sq->nr_pending) |
e43473b7 VG |
402 | return; |
403 | ||
c9e0332e | 404 | update_min_dispatch_time(sq); |
e43473b7 | 405 | |
c9e0332e | 406 | if (time_before_eq(sq->first_pending_disptime, jiffies)) |
450adcbe | 407 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 408 | else |
c9e0332e | 409 | throtl_schedule_delayed_work(td, sq->first_pending_disptime - jiffies); |
e43473b7 VG |
410 | } |
411 | ||
0f3457f6 | 412 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
413 | { |
414 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 415 | tg->io_disp[rw] = 0; |
e43473b7 VG |
416 | tg->slice_start[rw] = jiffies; |
417 | tg->slice_end[rw] = jiffies + throtl_slice; | |
0f3457f6 | 418 | throtl_log_tg(tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", |
e43473b7 VG |
419 | rw == READ ? 'R' : 'W', tg->slice_start[rw], |
420 | tg->slice_end[rw], jiffies); | |
421 | } | |
422 | ||
0f3457f6 TH |
423 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
424 | unsigned long jiffy_end) | |
d1ae8ffd VG |
425 | { |
426 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
427 | } | |
428 | ||
0f3457f6 TH |
429 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
430 | unsigned long jiffy_end) | |
e43473b7 VG |
431 | { |
432 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
0f3457f6 | 433 | throtl_log_tg(tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", |
e43473b7 VG |
434 | rw == READ ? 'R' : 'W', tg->slice_start[rw], |
435 | tg->slice_end[rw], jiffies); | |
436 | } | |
437 | ||
438 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 439 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
440 | { |
441 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
442 | return 0; | |
443 | ||
444 | return 1; | |
445 | } | |
446 | ||
447 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 448 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 449 | { |
3aad5d3e VG |
450 | unsigned long nr_slices, time_elapsed, io_trim; |
451 | u64 bytes_trim, tmp; | |
e43473b7 VG |
452 | |
453 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
454 | ||
455 | /* | |
456 | * If bps are unlimited (-1), then time slice don't get | |
457 | * renewed. Don't try to trim the slice if slice is used. A new | |
458 | * slice will start when appropriate. | |
459 | */ | |
0f3457f6 | 460 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
461 | return; |
462 | ||
d1ae8ffd VG |
463 | /* |
464 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
465 | * that initially cgroup limit was very low resulting in high | |
466 | * slice_end, but later limit was bumped up and bio was dispached | |
467 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
468 | * is bad because it does not allow new slice to start. | |
469 | */ | |
470 | ||
0f3457f6 | 471 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); |
d1ae8ffd | 472 | |
e43473b7 VG |
473 | time_elapsed = jiffies - tg->slice_start[rw]; |
474 | ||
475 | nr_slices = time_elapsed / throtl_slice; | |
476 | ||
477 | if (!nr_slices) | |
478 | return; | |
3aad5d3e VG |
479 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
480 | do_div(tmp, HZ); | |
481 | bytes_trim = tmp; | |
e43473b7 | 482 | |
8e89d13f | 483 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 484 | |
8e89d13f | 485 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
486 | return; |
487 | ||
488 | if (tg->bytes_disp[rw] >= bytes_trim) | |
489 | tg->bytes_disp[rw] -= bytes_trim; | |
490 | else | |
491 | tg->bytes_disp[rw] = 0; | |
492 | ||
8e89d13f VG |
493 | if (tg->io_disp[rw] >= io_trim) |
494 | tg->io_disp[rw] -= io_trim; | |
495 | else | |
496 | tg->io_disp[rw] = 0; | |
497 | ||
e43473b7 VG |
498 | tg->slice_start[rw] += nr_slices * throtl_slice; |
499 | ||
0f3457f6 | 500 | throtl_log_tg(tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 501 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 502 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
503 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
504 | } | |
505 | ||
0f3457f6 TH |
506 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
507 | unsigned long *wait) | |
e43473b7 VG |
508 | { |
509 | bool rw = bio_data_dir(bio); | |
8e89d13f | 510 | unsigned int io_allowed; |
e43473b7 | 511 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 512 | u64 tmp; |
e43473b7 | 513 | |
8e89d13f | 514 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 515 | |
8e89d13f VG |
516 | /* Slice has just started. Consider one slice interval */ |
517 | if (!jiffy_elapsed) | |
518 | jiffy_elapsed_rnd = throtl_slice; | |
519 | ||
520 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
521 | ||
c49c06e4 VG |
522 | /* |
523 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
524 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
525 | * will allow dispatch after 1 second and after that slice should | |
526 | * have been trimmed. | |
527 | */ | |
528 | ||
529 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
530 | do_div(tmp, HZ); | |
531 | ||
532 | if (tmp > UINT_MAX) | |
533 | io_allowed = UINT_MAX; | |
534 | else | |
535 | io_allowed = tmp; | |
8e89d13f VG |
536 | |
537 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
538 | if (wait) |
539 | *wait = 0; | |
540 | return 1; | |
541 | } | |
542 | ||
8e89d13f VG |
543 | /* Calc approx time to dispatch */ |
544 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
545 | ||
546 | if (jiffy_wait > jiffy_elapsed) | |
547 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
548 | else | |
549 | jiffy_wait = 1; | |
550 | ||
551 | if (wait) | |
552 | *wait = jiffy_wait; | |
553 | return 0; | |
554 | } | |
555 | ||
0f3457f6 TH |
556 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
557 | unsigned long *wait) | |
8e89d13f VG |
558 | { |
559 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 560 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 561 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
562 | |
563 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
564 | ||
565 | /* Slice has just started. Consider one slice interval */ | |
566 | if (!jiffy_elapsed) | |
567 | jiffy_elapsed_rnd = throtl_slice; | |
568 | ||
569 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
570 | ||
5e901a2b VG |
571 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
572 | do_div(tmp, HZ); | |
3aad5d3e | 573 | bytes_allowed = tmp; |
e43473b7 VG |
574 | |
575 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
576 | if (wait) | |
577 | *wait = 0; | |
578 | return 1; | |
579 | } | |
580 | ||
581 | /* Calc approx time to dispatch */ | |
582 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
583 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
584 | ||
585 | if (!jiffy_wait) | |
586 | jiffy_wait = 1; | |
587 | ||
588 | /* | |
589 | * This wait time is without taking into consideration the rounding | |
590 | * up we did. Add that time also. | |
591 | */ | |
592 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
593 | if (wait) |
594 | *wait = jiffy_wait; | |
8e89d13f VG |
595 | return 0; |
596 | } | |
597 | ||
af75cd3c VG |
598 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
599 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
600 | return 1; | |
601 | return 0; | |
602 | } | |
603 | ||
8e89d13f VG |
604 | /* |
605 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
606 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
607 | */ | |
0f3457f6 TH |
608 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
609 | unsigned long *wait) | |
8e89d13f VG |
610 | { |
611 | bool rw = bio_data_dir(bio); | |
612 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
613 | ||
614 | /* | |
615 | * Currently whole state machine of group depends on first bio | |
616 | * queued in the group bio list. So one should not be calling | |
617 | * this function with a different bio if there are other bios | |
618 | * queued. | |
619 | */ | |
620 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); | |
e43473b7 | 621 | |
8e89d13f VG |
622 | /* If tg->bps = -1, then BW is unlimited */ |
623 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
624 | if (wait) | |
625 | *wait = 0; | |
626 | return 1; | |
627 | } | |
628 | ||
629 | /* | |
630 | * If previous slice expired, start a new one otherwise renew/extend | |
631 | * existing slice to make sure it is at least throtl_slice interval | |
632 | * long since now. | |
633 | */ | |
0f3457f6 TH |
634 | if (throtl_slice_used(tg, rw)) |
635 | throtl_start_new_slice(tg, rw); | |
8e89d13f VG |
636 | else { |
637 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
0f3457f6 | 638 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); |
8e89d13f VG |
639 | } |
640 | ||
0f3457f6 TH |
641 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
642 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
643 | if (wait) |
644 | *wait = 0; | |
645 | return 1; | |
646 | } | |
647 | ||
648 | max_wait = max(bps_wait, iops_wait); | |
649 | ||
650 | if (wait) | |
651 | *wait = max_wait; | |
652 | ||
653 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 654 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
655 | |
656 | return 0; | |
657 | } | |
658 | ||
3c798398 | 659 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
660 | int rw) |
661 | { | |
8a3d2615 TH |
662 | struct throtl_grp *tg = blkg_to_tg(blkg); |
663 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
664 | unsigned long flags; |
665 | ||
666 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 667 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
668 | return; |
669 | ||
670 | /* | |
671 | * Disabling interrupts to provide mutual exclusion between two | |
672 | * writes on same cpu. It probably is not needed for 64bit. Not | |
673 | * optimizing that case yet. | |
674 | */ | |
675 | local_irq_save(flags); | |
676 | ||
8a3d2615 | 677 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 678 | |
629ed0b1 TH |
679 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
680 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
681 | ||
682 | local_irq_restore(flags); | |
683 | } | |
684 | ||
e43473b7 VG |
685 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
686 | { | |
687 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
688 | |
689 | /* Charge the bio to the group */ | |
690 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 691 | tg->io_disp[rw]++; |
e43473b7 | 692 | |
629ed0b1 | 693 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw); |
e43473b7 VG |
694 | } |
695 | ||
0049af73 TH |
696 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_grp *tg, |
697 | struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
698 | { |
699 | bool rw = bio_data_dir(bio); | |
700 | ||
701 | bio_list_add(&tg->bio_lists[rw], bio); | |
702 | /* Take a bio reference on tg */ | |
1adaf3dd | 703 | blkg_get(tg_to_blkg(tg)); |
e43473b7 | 704 | tg->nr_queued[rw]++; |
e2d57e60 | 705 | tg->td->nr_queued[rw]++; |
0049af73 | 706 | throtl_enqueue_tg(tg, parent_sq); |
e43473b7 VG |
707 | } |
708 | ||
0049af73 TH |
709 | static void tg_update_disptime(struct throtl_grp *tg, |
710 | struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
711 | { |
712 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
713 | struct bio *bio; | |
714 | ||
715 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
0f3457f6 | 716 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 VG |
717 | |
718 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
0f3457f6 | 719 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
720 | |
721 | min_wait = min(read_wait, write_wait); | |
722 | disptime = jiffies + min_wait; | |
723 | ||
e43473b7 | 724 | /* Update dispatch time */ |
0049af73 | 725 | throtl_dequeue_tg(tg, parent_sq); |
e43473b7 | 726 | tg->disptime = disptime; |
0049af73 | 727 | throtl_enqueue_tg(tg, parent_sq); |
e43473b7 VG |
728 | } |
729 | ||
0f3457f6 TH |
730 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw, |
731 | struct bio_list *bl) | |
e43473b7 VG |
732 | { |
733 | struct bio *bio; | |
734 | ||
735 | bio = bio_list_pop(&tg->bio_lists[rw]); | |
736 | tg->nr_queued[rw]--; | |
1adaf3dd TH |
737 | /* Drop bio reference on blkg */ |
738 | blkg_put(tg_to_blkg(tg)); | |
e43473b7 | 739 | |
0f3457f6 TH |
740 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
741 | tg->td->nr_queued[rw]--; | |
e43473b7 VG |
742 | |
743 | throtl_charge_bio(tg, bio); | |
744 | bio_list_add(bl, bio); | |
745 | bio->bi_rw |= REQ_THROTTLED; | |
746 | ||
0f3457f6 | 747 | throtl_trim_slice(tg, rw); |
e43473b7 VG |
748 | } |
749 | ||
0f3457f6 | 750 | static int throtl_dispatch_tg(struct throtl_grp *tg, struct bio_list *bl) |
e43473b7 VG |
751 | { |
752 | unsigned int nr_reads = 0, nr_writes = 0; | |
753 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 754 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
755 | struct bio *bio; |
756 | ||
757 | /* Try to dispatch 75% READS and 25% WRITES */ | |
758 | ||
0f3457f6 TH |
759 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) && |
760 | tg_may_dispatch(tg, bio, NULL)) { | |
e43473b7 | 761 | |
0f3457f6 | 762 | tg_dispatch_one_bio(tg, bio_data_dir(bio), bl); |
e43473b7 VG |
763 | nr_reads++; |
764 | ||
765 | if (nr_reads >= max_nr_reads) | |
766 | break; | |
767 | } | |
768 | ||
0f3457f6 TH |
769 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) && |
770 | tg_may_dispatch(tg, bio, NULL)) { | |
e43473b7 | 771 | |
0f3457f6 | 772 | tg_dispatch_one_bio(tg, bio_data_dir(bio), bl); |
e43473b7 VG |
773 | nr_writes++; |
774 | ||
775 | if (nr_writes >= max_nr_writes) | |
776 | break; | |
777 | } | |
778 | ||
779 | return nr_reads + nr_writes; | |
780 | } | |
781 | ||
0049af73 | 782 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq, |
e2d57e60 | 783 | struct bio_list *bl) |
e43473b7 VG |
784 | { |
785 | unsigned int nr_disp = 0; | |
786 | struct throtl_grp *tg; | |
e43473b7 VG |
787 | |
788 | while (1) { | |
0049af73 | 789 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
790 | |
791 | if (!tg) | |
792 | break; | |
793 | ||
794 | if (time_before(jiffies, tg->disptime)) | |
795 | break; | |
796 | ||
0049af73 | 797 | throtl_dequeue_tg(tg, parent_sq); |
e43473b7 | 798 | |
0f3457f6 | 799 | nr_disp += throtl_dispatch_tg(tg, bl); |
e43473b7 | 800 | |
2db6314c | 801 | if (tg->nr_queued[0] || tg->nr_queued[1]) |
0049af73 | 802 | tg_update_disptime(tg, parent_sq); |
e43473b7 VG |
803 | |
804 | if (nr_disp >= throtl_quantum) | |
805 | break; | |
806 | } | |
807 | ||
808 | return nr_disp; | |
809 | } | |
810 | ||
cb76199c TH |
811 | /* work function to dispatch throttled bios */ |
812 | void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
e43473b7 | 813 | { |
cb76199c TH |
814 | struct throtl_data *td = container_of(to_delayed_work(work), |
815 | struct throtl_data, dispatch_work); | |
816 | struct request_queue *q = td->queue; | |
e43473b7 VG |
817 | unsigned int nr_disp = 0; |
818 | struct bio_list bio_list_on_stack; | |
819 | struct bio *bio; | |
69d60eb9 | 820 | struct blk_plug plug; |
e43473b7 VG |
821 | |
822 | spin_lock_irq(q->queue_lock); | |
823 | ||
e43473b7 VG |
824 | bio_list_init(&bio_list_on_stack); |
825 | ||
d2f31a5f | 826 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
6a525600 TH |
827 | td->nr_queued[READ] + td->nr_queued[WRITE], |
828 | td->nr_queued[READ], td->nr_queued[WRITE]); | |
e43473b7 | 829 | |
e2d57e60 | 830 | nr_disp = throtl_select_dispatch(&td->service_queue, &bio_list_on_stack); |
e43473b7 VG |
831 | |
832 | if (nr_disp) | |
833 | throtl_log(td, "bios disp=%u", nr_disp); | |
834 | ||
835 | throtl_schedule_next_dispatch(td); | |
6a525600 | 836 | |
e43473b7 VG |
837 | spin_unlock_irq(q->queue_lock); |
838 | ||
839 | /* | |
840 | * If we dispatched some requests, unplug the queue to make sure | |
841 | * immediate dispatch | |
842 | */ | |
843 | if (nr_disp) { | |
69d60eb9 | 844 | blk_start_plug(&plug); |
e43473b7 VG |
845 | while((bio = bio_list_pop(&bio_list_on_stack))) |
846 | generic_make_request(bio); | |
69d60eb9 | 847 | blk_finish_plug(&plug); |
e43473b7 | 848 | } |
e43473b7 VG |
849 | } |
850 | ||
f95a04af TH |
851 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
852 | struct blkg_policy_data *pd, int off) | |
41b38b6d | 853 | { |
f95a04af | 854 | struct throtl_grp *tg = pd_to_tg(pd); |
41b38b6d TH |
855 | struct blkg_rwstat rwstat = { }, tmp; |
856 | int i, cpu; | |
857 | ||
858 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 859 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
860 | |
861 | tmp = blkg_rwstat_read((void *)sc + off); | |
862 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
863 | rwstat.cnt[i] += tmp.cnt[i]; | |
864 | } | |
865 | ||
f95a04af | 866 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
41b38b6d TH |
867 | } |
868 | ||
8a3d2615 TH |
869 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
870 | struct seq_file *sf) | |
41b38b6d | 871 | { |
3c798398 | 872 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
41b38b6d | 873 | |
3c798398 | 874 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 875 | cft->private, true); |
41b38b6d TH |
876 | return 0; |
877 | } | |
878 | ||
f95a04af TH |
879 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
880 | int off) | |
60c2bc2d | 881 | { |
f95a04af TH |
882 | struct throtl_grp *tg = pd_to_tg(pd); |
883 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 884 | |
af133ceb | 885 | if (v == -1) |
60c2bc2d | 886 | return 0; |
f95a04af | 887 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
888 | } |
889 | ||
f95a04af TH |
890 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
891 | int off) | |
e43473b7 | 892 | { |
f95a04af TH |
893 | struct throtl_grp *tg = pd_to_tg(pd); |
894 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 895 | |
af133ceb TH |
896 | if (v == -1) |
897 | return 0; | |
f95a04af | 898 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
899 | } |
900 | ||
af133ceb TH |
901 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
902 | struct seq_file *sf) | |
8e89d13f | 903 | { |
3c798398 TH |
904 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, |
905 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 906 | return 0; |
8e89d13f VG |
907 | } |
908 | ||
af133ceb TH |
909 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
910 | struct seq_file *sf) | |
8e89d13f | 911 | { |
3c798398 TH |
912 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, |
913 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 914 | return 0; |
60c2bc2d TH |
915 | } |
916 | ||
af133ceb TH |
917 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
918 | bool is_u64) | |
60c2bc2d | 919 | { |
3c798398 | 920 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
60c2bc2d | 921 | struct blkg_conf_ctx ctx; |
af133ceb | 922 | struct throtl_grp *tg; |
a2b1693b | 923 | struct throtl_data *td; |
60c2bc2d TH |
924 | int ret; |
925 | ||
3c798398 | 926 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
927 | if (ret) |
928 | return ret; | |
929 | ||
af133ceb | 930 | tg = blkg_to_tg(ctx.blkg); |
a2b1693b | 931 | td = ctx.blkg->q->td; |
af133ceb | 932 | |
a2b1693b TH |
933 | if (!ctx.v) |
934 | ctx.v = -1; | |
af133ceb | 935 | |
a2b1693b TH |
936 | if (is_u64) |
937 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
938 | else | |
939 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 940 | |
0f3457f6 | 941 | throtl_log_tg(tg, "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", |
632b4493 TH |
942 | tg->bps[READ], tg->bps[WRITE], |
943 | tg->iops[READ], tg->iops[WRITE]); | |
944 | ||
945 | /* | |
946 | * We're already holding queue_lock and know @tg is valid. Let's | |
947 | * apply the new config directly. | |
948 | * | |
949 | * Restart the slices for both READ and WRITES. It might happen | |
950 | * that a group's limit are dropped suddenly and we don't want to | |
951 | * account recently dispatched IO with new low rate. | |
952 | */ | |
0f3457f6 TH |
953 | throtl_start_new_slice(tg, 0); |
954 | throtl_start_new_slice(tg, 1); | |
632b4493 | 955 | |
5b2c16aa | 956 | if (tg->flags & THROTL_TG_PENDING) { |
0049af73 | 957 | tg_update_disptime(tg, &td->service_queue); |
632b4493 TH |
958 | throtl_schedule_next_dispatch(td); |
959 | } | |
60c2bc2d TH |
960 | |
961 | blkg_conf_finish(&ctx); | |
a2b1693b | 962 | return 0; |
8e89d13f VG |
963 | } |
964 | ||
af133ceb TH |
965 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
966 | const char *buf) | |
60c2bc2d | 967 | { |
af133ceb | 968 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
969 | } |
970 | ||
af133ceb TH |
971 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
972 | const char *buf) | |
60c2bc2d | 973 | { |
af133ceb | 974 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
975 | } |
976 | ||
977 | static struct cftype throtl_files[] = { | |
978 | { | |
979 | .name = "throttle.read_bps_device", | |
af133ceb TH |
980 | .private = offsetof(struct throtl_grp, bps[READ]), |
981 | .read_seq_string = tg_print_conf_u64, | |
982 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
983 | .max_write_len = 256, |
984 | }, | |
985 | { | |
986 | .name = "throttle.write_bps_device", | |
af133ceb TH |
987 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
988 | .read_seq_string = tg_print_conf_u64, | |
989 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
990 | .max_write_len = 256, |
991 | }, | |
992 | { | |
993 | .name = "throttle.read_iops_device", | |
af133ceb TH |
994 | .private = offsetof(struct throtl_grp, iops[READ]), |
995 | .read_seq_string = tg_print_conf_uint, | |
996 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
997 | .max_write_len = 256, |
998 | }, | |
999 | { | |
1000 | .name = "throttle.write_iops_device", | |
af133ceb TH |
1001 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
1002 | .read_seq_string = tg_print_conf_uint, | |
1003 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1004 | .max_write_len = 256, |
1005 | }, | |
1006 | { | |
1007 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1008 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1009 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1010 | }, |
1011 | { | |
1012 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1013 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1014 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1015 | }, |
1016 | { } /* terminate */ | |
1017 | }; | |
1018 | ||
da527770 | 1019 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1020 | { |
1021 | struct throtl_data *td = q->td; | |
1022 | ||
cb76199c | 1023 | cancel_delayed_work_sync(&td->dispatch_work); |
e43473b7 VG |
1024 | } |
1025 | ||
3c798398 | 1026 | static struct blkcg_policy blkcg_policy_throtl = { |
f9fcc2d3 TH |
1027 | .pd_size = sizeof(struct throtl_grp), |
1028 | .cftypes = throtl_files, | |
1029 | ||
1030 | .pd_init_fn = throtl_pd_init, | |
1031 | .pd_exit_fn = throtl_pd_exit, | |
1032 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 VG |
1033 | }; |
1034 | ||
bc16a4f9 | 1035 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1036 | { |
1037 | struct throtl_data *td = q->td; | |
1038 | struct throtl_grp *tg; | |
e43473b7 | 1039 | bool rw = bio_data_dir(bio), update_disptime = true; |
3c798398 | 1040 | struct blkcg *blkcg; |
bc16a4f9 | 1041 | bool throttled = false; |
e43473b7 VG |
1042 | |
1043 | if (bio->bi_rw & REQ_THROTTLED) { | |
1044 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1045 | goto out; |
e43473b7 VG |
1046 | } |
1047 | ||
af75cd3c VG |
1048 | /* |
1049 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1050 | * basic fields like stats and io rates. If a group has no rules, | |
1051 | * just update the dispatch stats in lockless manner and return. | |
1052 | */ | |
af75cd3c | 1053 | rcu_read_lock(); |
3c798398 | 1054 | blkcg = bio_blkcg(bio); |
cd1604fa | 1055 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1056 | if (tg) { |
af75cd3c | 1057 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1058 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1059 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1060 | goto out_unlock_rcu; |
af75cd3c VG |
1061 | } |
1062 | } | |
af75cd3c VG |
1063 | |
1064 | /* | |
1065 | * Either group has not been allocated yet or it is not an unlimited | |
1066 | * IO group | |
1067 | */ | |
e43473b7 | 1068 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1069 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1070 | if (unlikely(!tg)) |
1071 | goto out_unlock; | |
f469a7b4 | 1072 | |
e43473b7 VG |
1073 | if (tg->nr_queued[rw]) { |
1074 | /* | |
1075 | * There is already another bio queued in same dir. No | |
1076 | * need to update dispatch time. | |
1077 | */ | |
231d704b | 1078 | update_disptime = false; |
e43473b7 | 1079 | goto queue_bio; |
de701c74 | 1080 | |
e43473b7 VG |
1081 | } |
1082 | ||
1083 | /* Bio is with-in rate limit of group */ | |
0f3457f6 | 1084 | if (tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1085 | throtl_charge_bio(tg, bio); |
04521db0 VG |
1086 | |
1087 | /* | |
1088 | * We need to trim slice even when bios are not being queued | |
1089 | * otherwise it might happen that a bio is not queued for | |
1090 | * a long time and slice keeps on extending and trim is not | |
1091 | * called for a long time. Now if limits are reduced suddenly | |
1092 | * we take into account all the IO dispatched so far at new | |
1093 | * low rate and * newly queued IO gets a really long dispatch | |
1094 | * time. | |
1095 | * | |
1096 | * So keep on trimming slice even if bio is not queued. | |
1097 | */ | |
0f3457f6 | 1098 | throtl_trim_slice(tg, rw); |
bc16a4f9 | 1099 | goto out_unlock; |
e43473b7 VG |
1100 | } |
1101 | ||
1102 | queue_bio: | |
0f3457f6 | 1103 | throtl_log_tg(tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
8e89d13f VG |
1104 | " iodisp=%u iops=%u queued=%d/%d", |
1105 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1106 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1107 | tg->io_disp[rw], tg->iops[rw], |
e43473b7 VG |
1108 | tg->nr_queued[READ], tg->nr_queued[WRITE]); |
1109 | ||
671058fb | 1110 | bio_associate_current(bio); |
0049af73 | 1111 | throtl_add_bio_tg(bio, tg, &q->td->service_queue); |
bc16a4f9 | 1112 | throttled = true; |
e43473b7 VG |
1113 | |
1114 | if (update_disptime) { | |
0049af73 | 1115 | tg_update_disptime(tg, &td->service_queue); |
e43473b7 VG |
1116 | throtl_schedule_next_dispatch(td); |
1117 | } | |
1118 | ||
bc16a4f9 | 1119 | out_unlock: |
e43473b7 | 1120 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1121 | out_unlock_rcu: |
1122 | rcu_read_unlock(); | |
bc16a4f9 TH |
1123 | out: |
1124 | return throttled; | |
e43473b7 VG |
1125 | } |
1126 | ||
c9a929dd TH |
1127 | /** |
1128 | * blk_throtl_drain - drain throttled bios | |
1129 | * @q: request_queue to drain throttled bios for | |
1130 | * | |
1131 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1132 | */ | |
1133 | void blk_throtl_drain(struct request_queue *q) | |
1134 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1135 | { | |
1136 | struct throtl_data *td = q->td; | |
0049af73 | 1137 | struct throtl_service_queue *parent_sq = &td->service_queue; |
c9a929dd TH |
1138 | struct throtl_grp *tg; |
1139 | struct bio_list bl; | |
1140 | struct bio *bio; | |
1141 | ||
8bcb6c7d | 1142 | queue_lockdep_assert_held(q); |
c9a929dd TH |
1143 | |
1144 | bio_list_init(&bl); | |
1145 | ||
0049af73 TH |
1146 | while ((tg = throtl_rb_first(parent_sq))) { |
1147 | throtl_dequeue_tg(tg, parent_sq); | |
c9a929dd TH |
1148 | |
1149 | while ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
0f3457f6 | 1150 | tg_dispatch_one_bio(tg, bio_data_dir(bio), &bl); |
c9a929dd | 1151 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) |
0f3457f6 | 1152 | tg_dispatch_one_bio(tg, bio_data_dir(bio), &bl); |
c9a929dd TH |
1153 | } |
1154 | spin_unlock_irq(q->queue_lock); | |
1155 | ||
1156 | while ((bio = bio_list_pop(&bl))) | |
1157 | generic_make_request(bio); | |
1158 | ||
1159 | spin_lock_irq(q->queue_lock); | |
1160 | } | |
1161 | ||
e43473b7 VG |
1162 | int blk_throtl_init(struct request_queue *q) |
1163 | { | |
1164 | struct throtl_data *td; | |
a2b1693b | 1165 | int ret; |
e43473b7 VG |
1166 | |
1167 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1168 | if (!td) | |
1169 | return -ENOMEM; | |
1170 | ||
c9e0332e | 1171 | td->service_queue = THROTL_SERVICE_QUEUE_INITIALIZER; |
cb76199c | 1172 | INIT_DELAYED_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
e43473b7 | 1173 | |
cd1604fa | 1174 | q->td = td; |
29b12589 | 1175 | td->queue = q; |
02977e4a | 1176 | |
a2b1693b | 1177 | /* activate policy */ |
3c798398 | 1178 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1179 | if (ret) |
f51b802c | 1180 | kfree(td); |
a2b1693b | 1181 | return ret; |
e43473b7 VG |
1182 | } |
1183 | ||
1184 | void blk_throtl_exit(struct request_queue *q) | |
1185 | { | |
c875f4d0 | 1186 | BUG_ON(!q->td); |
da527770 | 1187 | throtl_shutdown_wq(q); |
3c798398 | 1188 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1189 | kfree(q->td); |
e43473b7 VG |
1190 | } |
1191 | ||
1192 | static int __init throtl_init(void) | |
1193 | { | |
450adcbe VG |
1194 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1195 | if (!kthrotld_workqueue) | |
1196 | panic("Failed to create kthrotld\n"); | |
1197 | ||
3c798398 | 1198 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1199 | } |
1200 | ||
1201 | module_init(throtl_init); |