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; | |
28 | static void throtl_schedule_delayed_work(struct throtl_data *td, | |
29 | unsigned long delay); | |
30 | ||
e43473b7 VG |
31 | struct throtl_rb_root { |
32 | struct rb_root rb; | |
33 | struct rb_node *left; | |
34 | unsigned int count; | |
35 | unsigned long min_disptime; | |
36 | }; | |
37 | ||
38 | #define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \ | |
39 | .count = 0, .min_disptime = 0} | |
40 | ||
41 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) | |
42 | ||
8a3d2615 TH |
43 | /* Per-cpu group stats */ |
44 | struct tg_stats_cpu { | |
45 | /* total bytes transferred */ | |
46 | struct blkg_rwstat service_bytes; | |
47 | /* total IOs serviced, post merge */ | |
48 | struct blkg_rwstat serviced; | |
49 | }; | |
50 | ||
e43473b7 | 51 | struct throtl_grp { |
f95a04af TH |
52 | /* must be the first member */ |
53 | struct blkg_policy_data pd; | |
54 | ||
e43473b7 VG |
55 | /* active throtl group service_tree member */ |
56 | struct rb_node rb_node; | |
57 | ||
58 | /* | |
59 | * Dispatch time in jiffies. This is the estimated time when group | |
60 | * will unthrottle and is ready to dispatch more bio. It is used as | |
61 | * key to sort active groups in service tree. | |
62 | */ | |
63 | unsigned long disptime; | |
64 | ||
e43473b7 VG |
65 | unsigned int flags; |
66 | ||
67 | /* Two lists for READ and WRITE */ | |
68 | struct bio_list bio_lists[2]; | |
69 | ||
70 | /* Number of queued bios on READ and WRITE lists */ | |
71 | unsigned int nr_queued[2]; | |
72 | ||
73 | /* bytes per second rate limits */ | |
74 | uint64_t bps[2]; | |
75 | ||
8e89d13f VG |
76 | /* IOPS limits */ |
77 | unsigned int iops[2]; | |
78 | ||
e43473b7 VG |
79 | /* Number of bytes disptached in current slice */ |
80 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
81 | /* Number of bio's dispatched in current slice */ |
82 | unsigned int io_disp[2]; | |
e43473b7 VG |
83 | |
84 | /* When did we start a new slice */ | |
85 | unsigned long slice_start[2]; | |
86 | unsigned long slice_end[2]; | |
fe071437 | 87 | |
8a3d2615 TH |
88 | /* Per cpu stats pointer */ |
89 | struct tg_stats_cpu __percpu *stats_cpu; | |
90 | ||
91 | /* List of tgs waiting for per cpu stats memory to be allocated */ | |
92 | struct list_head stats_alloc_node; | |
e43473b7 VG |
93 | }; |
94 | ||
95 | struct throtl_data | |
96 | { | |
e43473b7 VG |
97 | /* service tree for active throtl groups */ |
98 | struct throtl_rb_root tg_service_tree; | |
99 | ||
e43473b7 VG |
100 | struct request_queue *queue; |
101 | ||
102 | /* Total Number of queued bios on READ and WRITE lists */ | |
103 | unsigned int nr_queued[2]; | |
104 | ||
105 | /* | |
02977e4a | 106 | * number of total undestroyed groups |
e43473b7 VG |
107 | */ |
108 | unsigned int nr_undestroyed_grps; | |
109 | ||
110 | /* Work for dispatching throttled bios */ | |
111 | struct delayed_work throtl_work; | |
112 | }; | |
113 | ||
8a3d2615 TH |
114 | /* list and work item to allocate percpu group stats */ |
115 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
116 | static LIST_HEAD(tg_stats_alloc_list); | |
117 | ||
118 | static void tg_stats_alloc_fn(struct work_struct *); | |
119 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
120 | ||
f95a04af TH |
121 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
122 | { | |
123 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
124 | } | |
125 | ||
3c798398 | 126 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 127 | { |
f95a04af | 128 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
129 | } |
130 | ||
3c798398 | 131 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 132 | { |
f95a04af | 133 | return pd_to_blkg(&tg->pd); |
0381411e TH |
134 | } |
135 | ||
03d8e111 TH |
136 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) |
137 | { | |
138 | return blkg_to_tg(td->queue->root_blkg); | |
139 | } | |
140 | ||
e43473b7 VG |
141 | enum tg_state_flags { |
142 | THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */ | |
143 | }; | |
144 | ||
145 | #define THROTL_TG_FNS(name) \ | |
146 | static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \ | |
147 | { \ | |
148 | (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \ | |
149 | } \ | |
150 | static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \ | |
151 | { \ | |
152 | (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \ | |
153 | } \ | |
154 | static inline int throtl_tg_##name(const struct throtl_grp *tg) \ | |
155 | { \ | |
156 | return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \ | |
157 | } | |
158 | ||
159 | THROTL_TG_FNS(on_rr); | |
160 | ||
54e7ed12 TH |
161 | #define throtl_log_tg(td, tg, fmt, args...) do { \ |
162 | char __pbuf[128]; \ | |
163 | \ | |
164 | blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \ | |
165 | blk_add_trace_msg((td)->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
166 | } while (0) | |
e43473b7 VG |
167 | |
168 | #define throtl_log(td, fmt, args...) \ | |
169 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) | |
170 | ||
d2f31a5f | 171 | static inline unsigned int total_nr_queued(struct throtl_data *td) |
e43473b7 | 172 | { |
d2f31a5f | 173 | return td->nr_queued[0] + td->nr_queued[1]; |
e43473b7 VG |
174 | } |
175 | ||
8a3d2615 TH |
176 | /* |
177 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
3b07e9ca | 178 | * system_wq once there are some groups on the alloc_list waiting for |
8a3d2615 TH |
179 | * allocation. |
180 | */ | |
181 | static void tg_stats_alloc_fn(struct work_struct *work) | |
182 | { | |
183 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
184 | struct delayed_work *dwork = to_delayed_work(work); | |
185 | bool empty = false; | |
186 | ||
187 | alloc_stats: | |
188 | if (!stats_cpu) { | |
189 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
190 | if (!stats_cpu) { | |
191 | /* allocation failed, try again after some time */ | |
3b07e9ca | 192 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); |
8a3d2615 TH |
193 | return; |
194 | } | |
195 | } | |
196 | ||
197 | spin_lock_irq(&tg_stats_alloc_lock); | |
198 | ||
199 | if (!list_empty(&tg_stats_alloc_list)) { | |
200 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
201 | struct throtl_grp, | |
202 | stats_alloc_node); | |
203 | swap(tg->stats_cpu, stats_cpu); | |
204 | list_del_init(&tg->stats_alloc_node); | |
205 | } | |
206 | ||
207 | empty = list_empty(&tg_stats_alloc_list); | |
208 | spin_unlock_irq(&tg_stats_alloc_lock); | |
209 | if (!empty) | |
210 | goto alloc_stats; | |
211 | } | |
212 | ||
3c798398 | 213 | static void throtl_pd_init(struct blkcg_gq *blkg) |
a29a171e | 214 | { |
0381411e | 215 | struct throtl_grp *tg = blkg_to_tg(blkg); |
ff26eaad | 216 | unsigned long flags; |
cd1604fa | 217 | |
a29a171e VG |
218 | RB_CLEAR_NODE(&tg->rb_node); |
219 | bio_list_init(&tg->bio_lists[0]); | |
220 | bio_list_init(&tg->bio_lists[1]); | |
a29a171e | 221 | |
e56da7e2 TH |
222 | tg->bps[READ] = -1; |
223 | tg->bps[WRITE] = -1; | |
224 | tg->iops[READ] = -1; | |
225 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
226 | |
227 | /* | |
228 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
229 | * but percpu allocator can't be called from IO path. Queue tg on | |
230 | * tg_stats_alloc_list and allocate from work item. | |
231 | */ | |
ff26eaad | 232 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 233 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); |
3b07e9ca | 234 | schedule_delayed_work(&tg_stats_alloc_work, 0); |
ff26eaad | 235 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
236 | } |
237 | ||
3c798398 | 238 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
8a3d2615 TH |
239 | { |
240 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
ff26eaad | 241 | unsigned long flags; |
8a3d2615 | 242 | |
ff26eaad | 243 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 244 | list_del_init(&tg->stats_alloc_node); |
ff26eaad | 245 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
246 | |
247 | free_percpu(tg->stats_cpu); | |
248 | } | |
249 | ||
3c798398 | 250 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
8a3d2615 TH |
251 | { |
252 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
253 | int cpu; | |
254 | ||
255 | if (tg->stats_cpu == NULL) | |
256 | return; | |
257 | ||
258 | for_each_possible_cpu(cpu) { | |
259 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
260 | ||
261 | blkg_rwstat_reset(&sc->service_bytes); | |
262 | blkg_rwstat_reset(&sc->serviced); | |
263 | } | |
a29a171e VG |
264 | } |
265 | ||
3c798398 TH |
266 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
267 | struct blkcg *blkcg) | |
e43473b7 | 268 | { |
be2c6b19 | 269 | /* |
3c798398 TH |
270 | * This is the common case when there are no blkcgs. Avoid lookup |
271 | * in this case | |
cd1604fa | 272 | */ |
3c798398 | 273 | if (blkcg == &blkcg_root) |
03d8e111 | 274 | return td_root_tg(td); |
e43473b7 | 275 | |
e8989fae | 276 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
277 | } |
278 | ||
cd1604fa | 279 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
3c798398 | 280 | struct blkcg *blkcg) |
e43473b7 | 281 | { |
f469a7b4 | 282 | struct request_queue *q = td->queue; |
cd1604fa | 283 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 284 | |
f469a7b4 | 285 | /* |
3c798398 TH |
286 | * This is the common case when there are no blkcgs. Avoid lookup |
287 | * in this case | |
f469a7b4 | 288 | */ |
3c798398 | 289 | if (blkcg == &blkcg_root) { |
03d8e111 | 290 | tg = td_root_tg(td); |
cd1604fa | 291 | } else { |
3c798398 | 292 | struct blkcg_gq *blkg; |
f469a7b4 | 293 | |
3c96cb32 | 294 | blkg = blkg_lookup_create(blkcg, q); |
f469a7b4 | 295 | |
cd1604fa TH |
296 | /* if %NULL and @q is alive, fall back to root_tg */ |
297 | if (!IS_ERR(blkg)) | |
0381411e | 298 | tg = blkg_to_tg(blkg); |
3f3299d5 | 299 | else if (!blk_queue_dying(q)) |
03d8e111 | 300 | tg = td_root_tg(td); |
f469a7b4 VG |
301 | } |
302 | ||
e43473b7 VG |
303 | return tg; |
304 | } | |
305 | ||
306 | static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root) | |
307 | { | |
308 | /* Service tree is empty */ | |
309 | if (!root->count) | |
310 | return NULL; | |
311 | ||
312 | if (!root->left) | |
313 | root->left = rb_first(&root->rb); | |
314 | ||
315 | if (root->left) | |
316 | return rb_entry_tg(root->left); | |
317 | ||
318 | return NULL; | |
319 | } | |
320 | ||
321 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
322 | { | |
323 | rb_erase(n, root); | |
324 | RB_CLEAR_NODE(n); | |
325 | } | |
326 | ||
327 | static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root) | |
328 | { | |
329 | if (root->left == n) | |
330 | root->left = NULL; | |
331 | rb_erase_init(n, &root->rb); | |
332 | --root->count; | |
333 | } | |
334 | ||
335 | static void update_min_dispatch_time(struct throtl_rb_root *st) | |
336 | { | |
337 | struct throtl_grp *tg; | |
338 | ||
339 | tg = throtl_rb_first(st); | |
340 | if (!tg) | |
341 | return; | |
342 | ||
343 | st->min_disptime = tg->disptime; | |
344 | } | |
345 | ||
346 | static void | |
347 | tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg) | |
348 | { | |
349 | struct rb_node **node = &st->rb.rb_node; | |
350 | struct rb_node *parent = NULL; | |
351 | struct throtl_grp *__tg; | |
352 | unsigned long key = tg->disptime; | |
353 | int left = 1; | |
354 | ||
355 | while (*node != NULL) { | |
356 | parent = *node; | |
357 | __tg = rb_entry_tg(parent); | |
358 | ||
359 | if (time_before(key, __tg->disptime)) | |
360 | node = &parent->rb_left; | |
361 | else { | |
362 | node = &parent->rb_right; | |
363 | left = 0; | |
364 | } | |
365 | } | |
366 | ||
367 | if (left) | |
368 | st->left = &tg->rb_node; | |
369 | ||
370 | rb_link_node(&tg->rb_node, parent, node); | |
371 | rb_insert_color(&tg->rb_node, &st->rb); | |
372 | } | |
373 | ||
374 | static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
375 | { | |
376 | struct throtl_rb_root *st = &td->tg_service_tree; | |
377 | ||
378 | tg_service_tree_add(st, tg); | |
379 | throtl_mark_tg_on_rr(tg); | |
380 | st->count++; | |
381 | } | |
382 | ||
383 | static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
384 | { | |
385 | if (!throtl_tg_on_rr(tg)) | |
386 | __throtl_enqueue_tg(td, tg); | |
387 | } | |
388 | ||
389 | static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
390 | { | |
391 | throtl_rb_erase(&tg->rb_node, &td->tg_service_tree); | |
392 | throtl_clear_tg_on_rr(tg); | |
393 | } | |
394 | ||
395 | static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
396 | { | |
397 | if (throtl_tg_on_rr(tg)) | |
398 | __throtl_dequeue_tg(td, tg); | |
399 | } | |
400 | ||
401 | static void throtl_schedule_next_dispatch(struct throtl_data *td) | |
402 | { | |
403 | struct throtl_rb_root *st = &td->tg_service_tree; | |
404 | ||
405 | /* | |
406 | * If there are more bios pending, schedule more work. | |
407 | */ | |
408 | if (!total_nr_queued(td)) | |
409 | return; | |
410 | ||
411 | BUG_ON(!st->count); | |
412 | ||
413 | update_min_dispatch_time(st); | |
414 | ||
415 | if (time_before_eq(st->min_disptime, jiffies)) | |
450adcbe | 416 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 417 | else |
450adcbe | 418 | throtl_schedule_delayed_work(td, (st->min_disptime - jiffies)); |
e43473b7 VG |
419 | } |
420 | ||
421 | static inline void | |
422 | throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
423 | { | |
424 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 425 | tg->io_disp[rw] = 0; |
e43473b7 VG |
426 | tg->slice_start[rw] = jiffies; |
427 | tg->slice_end[rw] = jiffies + throtl_slice; | |
428 | throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
429 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
430 | tg->slice_end[rw], jiffies); | |
431 | } | |
432 | ||
d1ae8ffd VG |
433 | static inline void throtl_set_slice_end(struct throtl_data *td, |
434 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
435 | { | |
436 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
437 | } | |
438 | ||
e43473b7 VG |
439 | static inline void throtl_extend_slice(struct throtl_data *td, |
440 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
441 | { | |
442 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
443 | throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
444 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
445 | tg->slice_end[rw], jiffies); | |
446 | } | |
447 | ||
448 | /* Determine if previously allocated or extended slice is complete or not */ | |
449 | static bool | |
450 | throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
451 | { | |
452 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
453 | return 0; | |
454 | ||
455 | return 1; | |
456 | } | |
457 | ||
458 | /* Trim the used slices and adjust slice start accordingly */ | |
459 | static inline void | |
460 | throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
461 | { | |
3aad5d3e VG |
462 | unsigned long nr_slices, time_elapsed, io_trim; |
463 | u64 bytes_trim, tmp; | |
e43473b7 VG |
464 | |
465 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
466 | ||
467 | /* | |
468 | * If bps are unlimited (-1), then time slice don't get | |
469 | * renewed. Don't try to trim the slice if slice is used. A new | |
470 | * slice will start when appropriate. | |
471 | */ | |
472 | if (throtl_slice_used(td, tg, rw)) | |
473 | return; | |
474 | ||
d1ae8ffd VG |
475 | /* |
476 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
477 | * that initially cgroup limit was very low resulting in high | |
478 | * slice_end, but later limit was bumped up and bio was dispached | |
479 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
480 | * is bad because it does not allow new slice to start. | |
481 | */ | |
482 | ||
483 | throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice); | |
484 | ||
e43473b7 VG |
485 | time_elapsed = jiffies - tg->slice_start[rw]; |
486 | ||
487 | nr_slices = time_elapsed / throtl_slice; | |
488 | ||
489 | if (!nr_slices) | |
490 | return; | |
3aad5d3e VG |
491 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
492 | do_div(tmp, HZ); | |
493 | bytes_trim = tmp; | |
e43473b7 | 494 | |
8e89d13f | 495 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 496 | |
8e89d13f | 497 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
498 | return; |
499 | ||
500 | if (tg->bytes_disp[rw] >= bytes_trim) | |
501 | tg->bytes_disp[rw] -= bytes_trim; | |
502 | else | |
503 | tg->bytes_disp[rw] = 0; | |
504 | ||
8e89d13f VG |
505 | if (tg->io_disp[rw] >= io_trim) |
506 | tg->io_disp[rw] -= io_trim; | |
507 | else | |
508 | tg->io_disp[rw] = 0; | |
509 | ||
e43473b7 VG |
510 | tg->slice_start[rw] += nr_slices * throtl_slice; |
511 | ||
3aad5d3e | 512 | throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 513 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 514 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
515 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
516 | } | |
517 | ||
8e89d13f VG |
518 | static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg, |
519 | struct bio *bio, unsigned long *wait) | |
e43473b7 VG |
520 | { |
521 | bool rw = bio_data_dir(bio); | |
8e89d13f | 522 | unsigned int io_allowed; |
e43473b7 | 523 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 524 | u64 tmp; |
e43473b7 | 525 | |
8e89d13f | 526 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 527 | |
8e89d13f VG |
528 | /* Slice has just started. Consider one slice interval */ |
529 | if (!jiffy_elapsed) | |
530 | jiffy_elapsed_rnd = throtl_slice; | |
531 | ||
532 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
533 | ||
c49c06e4 VG |
534 | /* |
535 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
536 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
537 | * will allow dispatch after 1 second and after that slice should | |
538 | * have been trimmed. | |
539 | */ | |
540 | ||
541 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
542 | do_div(tmp, HZ); | |
543 | ||
544 | if (tmp > UINT_MAX) | |
545 | io_allowed = UINT_MAX; | |
546 | else | |
547 | io_allowed = tmp; | |
8e89d13f VG |
548 | |
549 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
550 | if (wait) |
551 | *wait = 0; | |
552 | return 1; | |
553 | } | |
554 | ||
8e89d13f VG |
555 | /* Calc approx time to dispatch */ |
556 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
557 | ||
558 | if (jiffy_wait > jiffy_elapsed) | |
559 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
560 | else | |
561 | jiffy_wait = 1; | |
562 | ||
563 | if (wait) | |
564 | *wait = jiffy_wait; | |
565 | return 0; | |
566 | } | |
567 | ||
568 | static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg, | |
569 | struct bio *bio, unsigned long *wait) | |
570 | { | |
571 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 572 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 573 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
574 | |
575 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
576 | ||
577 | /* Slice has just started. Consider one slice interval */ | |
578 | if (!jiffy_elapsed) | |
579 | jiffy_elapsed_rnd = throtl_slice; | |
580 | ||
581 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
582 | ||
5e901a2b VG |
583 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
584 | do_div(tmp, HZ); | |
3aad5d3e | 585 | bytes_allowed = tmp; |
e43473b7 VG |
586 | |
587 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
588 | if (wait) | |
589 | *wait = 0; | |
590 | return 1; | |
591 | } | |
592 | ||
593 | /* Calc approx time to dispatch */ | |
594 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
595 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
596 | ||
597 | if (!jiffy_wait) | |
598 | jiffy_wait = 1; | |
599 | ||
600 | /* | |
601 | * This wait time is without taking into consideration the rounding | |
602 | * up we did. Add that time also. | |
603 | */ | |
604 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
605 | if (wait) |
606 | *wait = jiffy_wait; | |
8e89d13f VG |
607 | return 0; |
608 | } | |
609 | ||
af75cd3c VG |
610 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
611 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
612 | return 1; | |
613 | return 0; | |
614 | } | |
615 | ||
8e89d13f VG |
616 | /* |
617 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
618 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
619 | */ | |
620 | static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg, | |
621 | struct bio *bio, unsigned long *wait) | |
622 | { | |
623 | bool rw = bio_data_dir(bio); | |
624 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
625 | ||
626 | /* | |
627 | * Currently whole state machine of group depends on first bio | |
628 | * queued in the group bio list. So one should not be calling | |
629 | * this function with a different bio if there are other bios | |
630 | * queued. | |
631 | */ | |
632 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); | |
e43473b7 | 633 | |
8e89d13f VG |
634 | /* If tg->bps = -1, then BW is unlimited */ |
635 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
636 | if (wait) | |
637 | *wait = 0; | |
638 | return 1; | |
639 | } | |
640 | ||
641 | /* | |
642 | * If previous slice expired, start a new one otherwise renew/extend | |
643 | * existing slice to make sure it is at least throtl_slice interval | |
644 | * long since now. | |
645 | */ | |
646 | if (throtl_slice_used(td, tg, rw)) | |
647 | throtl_start_new_slice(td, tg, rw); | |
648 | else { | |
649 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
650 | throtl_extend_slice(td, tg, rw, jiffies + throtl_slice); | |
651 | } | |
652 | ||
653 | if (tg_with_in_bps_limit(td, tg, bio, &bps_wait) | |
654 | && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) { | |
655 | if (wait) | |
656 | *wait = 0; | |
657 | return 1; | |
658 | } | |
659 | ||
660 | max_wait = max(bps_wait, iops_wait); | |
661 | ||
662 | if (wait) | |
663 | *wait = max_wait; | |
664 | ||
665 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
666 | throtl_extend_slice(td, tg, rw, jiffies + max_wait); | |
e43473b7 VG |
667 | |
668 | return 0; | |
669 | } | |
670 | ||
3c798398 | 671 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
672 | int rw) |
673 | { | |
8a3d2615 TH |
674 | struct throtl_grp *tg = blkg_to_tg(blkg); |
675 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
676 | unsigned long flags; |
677 | ||
678 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 679 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
680 | return; |
681 | ||
682 | /* | |
683 | * Disabling interrupts to provide mutual exclusion between two | |
684 | * writes on same cpu. It probably is not needed for 64bit. Not | |
685 | * optimizing that case yet. | |
686 | */ | |
687 | local_irq_save(flags); | |
688 | ||
8a3d2615 | 689 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 690 | |
629ed0b1 TH |
691 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
692 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
693 | ||
694 | local_irq_restore(flags); | |
695 | } | |
696 | ||
e43473b7 VG |
697 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
698 | { | |
699 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
700 | |
701 | /* Charge the bio to the group */ | |
702 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 703 | tg->io_disp[rw]++; |
e43473b7 | 704 | |
629ed0b1 | 705 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw); |
e43473b7 VG |
706 | } |
707 | ||
708 | static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg, | |
709 | struct bio *bio) | |
710 | { | |
711 | bool rw = bio_data_dir(bio); | |
712 | ||
713 | bio_list_add(&tg->bio_lists[rw], bio); | |
714 | /* Take a bio reference on tg */ | |
1adaf3dd | 715 | blkg_get(tg_to_blkg(tg)); |
e43473b7 VG |
716 | tg->nr_queued[rw]++; |
717 | td->nr_queued[rw]++; | |
718 | throtl_enqueue_tg(td, tg); | |
719 | } | |
720 | ||
721 | static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg) | |
722 | { | |
723 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
724 | struct bio *bio; | |
725 | ||
726 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
727 | tg_may_dispatch(td, tg, bio, &read_wait); | |
728 | ||
729 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
730 | tg_may_dispatch(td, tg, bio, &write_wait); | |
731 | ||
732 | min_wait = min(read_wait, write_wait); | |
733 | disptime = jiffies + min_wait; | |
734 | ||
e43473b7 VG |
735 | /* Update dispatch time */ |
736 | throtl_dequeue_tg(td, tg); | |
737 | tg->disptime = disptime; | |
738 | throtl_enqueue_tg(td, tg); | |
739 | } | |
740 | ||
741 | static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg, | |
742 | bool rw, struct bio_list *bl) | |
743 | { | |
744 | struct bio *bio; | |
745 | ||
746 | bio = bio_list_pop(&tg->bio_lists[rw]); | |
747 | tg->nr_queued[rw]--; | |
1adaf3dd TH |
748 | /* Drop bio reference on blkg */ |
749 | blkg_put(tg_to_blkg(tg)); | |
e43473b7 VG |
750 | |
751 | BUG_ON(td->nr_queued[rw] <= 0); | |
752 | td->nr_queued[rw]--; | |
753 | ||
754 | throtl_charge_bio(tg, bio); | |
755 | bio_list_add(bl, bio); | |
756 | bio->bi_rw |= REQ_THROTTLED; | |
757 | ||
758 | throtl_trim_slice(td, tg, rw); | |
759 | } | |
760 | ||
761 | static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg, | |
762 | struct bio_list *bl) | |
763 | { | |
764 | unsigned int nr_reads = 0, nr_writes = 0; | |
765 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 766 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
767 | struct bio *bio; |
768 | ||
769 | /* Try to dispatch 75% READS and 25% WRITES */ | |
770 | ||
771 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) | |
772 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
773 | ||
774 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
775 | nr_reads++; | |
776 | ||
777 | if (nr_reads >= max_nr_reads) | |
778 | break; | |
779 | } | |
780 | ||
781 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) | |
782 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
783 | ||
784 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
785 | nr_writes++; | |
786 | ||
787 | if (nr_writes >= max_nr_writes) | |
788 | break; | |
789 | } | |
790 | ||
791 | return nr_reads + nr_writes; | |
792 | } | |
793 | ||
794 | static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl) | |
795 | { | |
796 | unsigned int nr_disp = 0; | |
797 | struct throtl_grp *tg; | |
798 | struct throtl_rb_root *st = &td->tg_service_tree; | |
799 | ||
800 | while (1) { | |
801 | tg = throtl_rb_first(st); | |
802 | ||
803 | if (!tg) | |
804 | break; | |
805 | ||
806 | if (time_before(jiffies, tg->disptime)) | |
807 | break; | |
808 | ||
809 | throtl_dequeue_tg(td, tg); | |
810 | ||
811 | nr_disp += throtl_dispatch_tg(td, tg, bl); | |
812 | ||
2db6314c | 813 | if (tg->nr_queued[0] || tg->nr_queued[1]) |
e43473b7 | 814 | tg_update_disptime(td, tg); |
e43473b7 VG |
815 | |
816 | if (nr_disp >= throtl_quantum) | |
817 | break; | |
818 | } | |
819 | ||
820 | return nr_disp; | |
821 | } | |
822 | ||
823 | /* Dispatch throttled bios. Should be called without queue lock held. */ | |
824 | static int throtl_dispatch(struct request_queue *q) | |
825 | { | |
826 | struct throtl_data *td = q->td; | |
827 | unsigned int nr_disp = 0; | |
828 | struct bio_list bio_list_on_stack; | |
829 | struct bio *bio; | |
69d60eb9 | 830 | struct blk_plug plug; |
e43473b7 VG |
831 | |
832 | spin_lock_irq(q->queue_lock); | |
833 | ||
834 | if (!total_nr_queued(td)) | |
835 | goto out; | |
836 | ||
837 | bio_list_init(&bio_list_on_stack); | |
838 | ||
d2f31a5f | 839 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
e43473b7 VG |
840 | total_nr_queued(td), td->nr_queued[READ], |
841 | td->nr_queued[WRITE]); | |
842 | ||
843 | nr_disp = throtl_select_dispatch(td, &bio_list_on_stack); | |
844 | ||
845 | if (nr_disp) | |
846 | throtl_log(td, "bios disp=%u", nr_disp); | |
847 | ||
848 | throtl_schedule_next_dispatch(td); | |
849 | out: | |
850 | spin_unlock_irq(q->queue_lock); | |
851 | ||
852 | /* | |
853 | * If we dispatched some requests, unplug the queue to make sure | |
854 | * immediate dispatch | |
855 | */ | |
856 | if (nr_disp) { | |
69d60eb9 | 857 | blk_start_plug(&plug); |
e43473b7 VG |
858 | while((bio = bio_list_pop(&bio_list_on_stack))) |
859 | generic_make_request(bio); | |
69d60eb9 | 860 | blk_finish_plug(&plug); |
e43473b7 VG |
861 | } |
862 | return nr_disp; | |
863 | } | |
864 | ||
865 | void blk_throtl_work(struct work_struct *work) | |
866 | { | |
867 | struct throtl_data *td = container_of(work, struct throtl_data, | |
868 | throtl_work.work); | |
869 | struct request_queue *q = td->queue; | |
870 | ||
871 | throtl_dispatch(q); | |
872 | } | |
873 | ||
874 | /* Call with queue lock held */ | |
450adcbe VG |
875 | static void |
876 | throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay) | |
e43473b7 VG |
877 | { |
878 | ||
e43473b7 VG |
879 | struct delayed_work *dwork = &td->throtl_work; |
880 | ||
632b4493 | 881 | if (total_nr_queued(td)) { |
e7c2f967 | 882 | mod_delayed_work(kthrotld_workqueue, dwork, delay); |
e43473b7 VG |
883 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", |
884 | delay, jiffies); | |
885 | } | |
886 | } | |
e43473b7 | 887 | |
f95a04af TH |
888 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
889 | struct blkg_policy_data *pd, int off) | |
41b38b6d | 890 | { |
f95a04af | 891 | struct throtl_grp *tg = pd_to_tg(pd); |
41b38b6d TH |
892 | struct blkg_rwstat rwstat = { }, tmp; |
893 | int i, cpu; | |
894 | ||
895 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 896 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
897 | |
898 | tmp = blkg_rwstat_read((void *)sc + off); | |
899 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
900 | rwstat.cnt[i] += tmp.cnt[i]; | |
901 | } | |
902 | ||
f95a04af | 903 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
41b38b6d TH |
904 | } |
905 | ||
8a3d2615 TH |
906 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
907 | struct seq_file *sf) | |
41b38b6d | 908 | { |
3c798398 | 909 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
41b38b6d | 910 | |
3c798398 | 911 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 912 | cft->private, true); |
41b38b6d TH |
913 | return 0; |
914 | } | |
915 | ||
f95a04af TH |
916 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
917 | int off) | |
60c2bc2d | 918 | { |
f95a04af TH |
919 | struct throtl_grp *tg = pd_to_tg(pd); |
920 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 921 | |
af133ceb | 922 | if (v == -1) |
60c2bc2d | 923 | return 0; |
f95a04af | 924 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
925 | } |
926 | ||
f95a04af TH |
927 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
928 | int off) | |
e43473b7 | 929 | { |
f95a04af TH |
930 | struct throtl_grp *tg = pd_to_tg(pd); |
931 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 932 | |
af133ceb TH |
933 | if (v == -1) |
934 | return 0; | |
f95a04af | 935 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
936 | } |
937 | ||
af133ceb TH |
938 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
939 | struct seq_file *sf) | |
8e89d13f | 940 | { |
3c798398 TH |
941 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, |
942 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 943 | return 0; |
8e89d13f VG |
944 | } |
945 | ||
af133ceb TH |
946 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
947 | struct seq_file *sf) | |
8e89d13f | 948 | { |
3c798398 TH |
949 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, |
950 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 951 | return 0; |
60c2bc2d TH |
952 | } |
953 | ||
af133ceb TH |
954 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
955 | bool is_u64) | |
60c2bc2d | 956 | { |
3c798398 | 957 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
60c2bc2d | 958 | struct blkg_conf_ctx ctx; |
af133ceb | 959 | struct throtl_grp *tg; |
a2b1693b | 960 | struct throtl_data *td; |
60c2bc2d TH |
961 | int ret; |
962 | ||
3c798398 | 963 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
964 | if (ret) |
965 | return ret; | |
966 | ||
af133ceb | 967 | tg = blkg_to_tg(ctx.blkg); |
a2b1693b | 968 | td = ctx.blkg->q->td; |
af133ceb | 969 | |
a2b1693b TH |
970 | if (!ctx.v) |
971 | ctx.v = -1; | |
af133ceb | 972 | |
a2b1693b TH |
973 | if (is_u64) |
974 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
975 | else | |
976 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 977 | |
632b4493 TH |
978 | throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", |
979 | tg->bps[READ], tg->bps[WRITE], | |
980 | tg->iops[READ], tg->iops[WRITE]); | |
981 | ||
982 | /* | |
983 | * We're already holding queue_lock and know @tg is valid. Let's | |
984 | * apply the new config directly. | |
985 | * | |
986 | * Restart the slices for both READ and WRITES. It might happen | |
987 | * that a group's limit are dropped suddenly and we don't want to | |
988 | * account recently dispatched IO with new low rate. | |
989 | */ | |
990 | throtl_start_new_slice(td, tg, 0); | |
991 | throtl_start_new_slice(td, tg, 1); | |
992 | ||
993 | if (throtl_tg_on_rr(tg)) { | |
994 | tg_update_disptime(td, tg); | |
995 | throtl_schedule_next_dispatch(td); | |
996 | } | |
60c2bc2d TH |
997 | |
998 | blkg_conf_finish(&ctx); | |
a2b1693b | 999 | return 0; |
8e89d13f VG |
1000 | } |
1001 | ||
af133ceb TH |
1002 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1003 | const char *buf) | |
60c2bc2d | 1004 | { |
af133ceb | 1005 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
1006 | } |
1007 | ||
af133ceb TH |
1008 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1009 | const char *buf) | |
60c2bc2d | 1010 | { |
af133ceb | 1011 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
1012 | } |
1013 | ||
1014 | static struct cftype throtl_files[] = { | |
1015 | { | |
1016 | .name = "throttle.read_bps_device", | |
af133ceb TH |
1017 | .private = offsetof(struct throtl_grp, bps[READ]), |
1018 | .read_seq_string = tg_print_conf_u64, | |
1019 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1020 | .max_write_len = 256, |
1021 | }, | |
1022 | { | |
1023 | .name = "throttle.write_bps_device", | |
af133ceb TH |
1024 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
1025 | .read_seq_string = tg_print_conf_u64, | |
1026 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1027 | .max_write_len = 256, |
1028 | }, | |
1029 | { | |
1030 | .name = "throttle.read_iops_device", | |
af133ceb TH |
1031 | .private = offsetof(struct throtl_grp, iops[READ]), |
1032 | .read_seq_string = tg_print_conf_uint, | |
1033 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1034 | .max_write_len = 256, |
1035 | }, | |
1036 | { | |
1037 | .name = "throttle.write_iops_device", | |
af133ceb TH |
1038 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
1039 | .read_seq_string = tg_print_conf_uint, | |
1040 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1041 | .max_write_len = 256, |
1042 | }, | |
1043 | { | |
1044 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1045 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1046 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1047 | }, |
1048 | { | |
1049 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1050 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1051 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1052 | }, |
1053 | { } /* terminate */ | |
1054 | }; | |
1055 | ||
da527770 | 1056 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1057 | { |
1058 | struct throtl_data *td = q->td; | |
1059 | ||
1060 | cancel_delayed_work_sync(&td->throtl_work); | |
1061 | } | |
1062 | ||
3c798398 | 1063 | static struct blkcg_policy blkcg_policy_throtl = { |
f9fcc2d3 TH |
1064 | .pd_size = sizeof(struct throtl_grp), |
1065 | .cftypes = throtl_files, | |
1066 | ||
1067 | .pd_init_fn = throtl_pd_init, | |
1068 | .pd_exit_fn = throtl_pd_exit, | |
1069 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 VG |
1070 | }; |
1071 | ||
bc16a4f9 | 1072 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1073 | { |
1074 | struct throtl_data *td = q->td; | |
1075 | struct throtl_grp *tg; | |
e43473b7 | 1076 | bool rw = bio_data_dir(bio), update_disptime = true; |
3c798398 | 1077 | struct blkcg *blkcg; |
bc16a4f9 | 1078 | bool throttled = false; |
e43473b7 VG |
1079 | |
1080 | if (bio->bi_rw & REQ_THROTTLED) { | |
1081 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1082 | goto out; |
e43473b7 VG |
1083 | } |
1084 | ||
af75cd3c VG |
1085 | /* |
1086 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1087 | * basic fields like stats and io rates. If a group has no rules, | |
1088 | * just update the dispatch stats in lockless manner and return. | |
1089 | */ | |
af75cd3c | 1090 | rcu_read_lock(); |
3c798398 | 1091 | blkcg = bio_blkcg(bio); |
cd1604fa | 1092 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1093 | if (tg) { |
af75cd3c | 1094 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1095 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1096 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1097 | goto out_unlock_rcu; |
af75cd3c VG |
1098 | } |
1099 | } | |
af75cd3c VG |
1100 | |
1101 | /* | |
1102 | * Either group has not been allocated yet or it is not an unlimited | |
1103 | * IO group | |
1104 | */ | |
e43473b7 | 1105 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1106 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1107 | if (unlikely(!tg)) |
1108 | goto out_unlock; | |
f469a7b4 | 1109 | |
e43473b7 VG |
1110 | if (tg->nr_queued[rw]) { |
1111 | /* | |
1112 | * There is already another bio queued in same dir. No | |
1113 | * need to update dispatch time. | |
1114 | */ | |
231d704b | 1115 | update_disptime = false; |
e43473b7 | 1116 | goto queue_bio; |
de701c74 | 1117 | |
e43473b7 VG |
1118 | } |
1119 | ||
1120 | /* Bio is with-in rate limit of group */ | |
1121 | if (tg_may_dispatch(td, tg, bio, NULL)) { | |
1122 | throtl_charge_bio(tg, bio); | |
04521db0 VG |
1123 | |
1124 | /* | |
1125 | * We need to trim slice even when bios are not being queued | |
1126 | * otherwise it might happen that a bio is not queued for | |
1127 | * a long time and slice keeps on extending and trim is not | |
1128 | * called for a long time. Now if limits are reduced suddenly | |
1129 | * we take into account all the IO dispatched so far at new | |
1130 | * low rate and * newly queued IO gets a really long dispatch | |
1131 | * time. | |
1132 | * | |
1133 | * So keep on trimming slice even if bio is not queued. | |
1134 | */ | |
1135 | throtl_trim_slice(td, tg, rw); | |
bc16a4f9 | 1136 | goto out_unlock; |
e43473b7 VG |
1137 | } |
1138 | ||
1139 | queue_bio: | |
fd16d263 | 1140 | throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
8e89d13f VG |
1141 | " iodisp=%u iops=%u queued=%d/%d", |
1142 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1143 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1144 | tg->io_disp[rw], tg->iops[rw], |
e43473b7 VG |
1145 | tg->nr_queued[READ], tg->nr_queued[WRITE]); |
1146 | ||
671058fb | 1147 | bio_associate_current(bio); |
e43473b7 | 1148 | throtl_add_bio_tg(q->td, tg, bio); |
bc16a4f9 | 1149 | throttled = true; |
e43473b7 VG |
1150 | |
1151 | if (update_disptime) { | |
1152 | tg_update_disptime(td, tg); | |
1153 | throtl_schedule_next_dispatch(td); | |
1154 | } | |
1155 | ||
bc16a4f9 | 1156 | out_unlock: |
e43473b7 | 1157 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1158 | out_unlock_rcu: |
1159 | rcu_read_unlock(); | |
bc16a4f9 TH |
1160 | out: |
1161 | return throttled; | |
e43473b7 VG |
1162 | } |
1163 | ||
c9a929dd TH |
1164 | /** |
1165 | * blk_throtl_drain - drain throttled bios | |
1166 | * @q: request_queue to drain throttled bios for | |
1167 | * | |
1168 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1169 | */ | |
1170 | void blk_throtl_drain(struct request_queue *q) | |
1171 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1172 | { | |
1173 | struct throtl_data *td = q->td; | |
1174 | struct throtl_rb_root *st = &td->tg_service_tree; | |
1175 | struct throtl_grp *tg; | |
1176 | struct bio_list bl; | |
1177 | struct bio *bio; | |
1178 | ||
8bcb6c7d | 1179 | queue_lockdep_assert_held(q); |
c9a929dd TH |
1180 | |
1181 | bio_list_init(&bl); | |
1182 | ||
1183 | while ((tg = throtl_rb_first(st))) { | |
1184 | throtl_dequeue_tg(td, tg); | |
1185 | ||
1186 | while ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
1187 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1188 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
1189 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1190 | } | |
1191 | spin_unlock_irq(q->queue_lock); | |
1192 | ||
1193 | while ((bio = bio_list_pop(&bl))) | |
1194 | generic_make_request(bio); | |
1195 | ||
1196 | spin_lock_irq(q->queue_lock); | |
1197 | } | |
1198 | ||
e43473b7 VG |
1199 | int blk_throtl_init(struct request_queue *q) |
1200 | { | |
1201 | struct throtl_data *td; | |
a2b1693b | 1202 | int ret; |
e43473b7 VG |
1203 | |
1204 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1205 | if (!td) | |
1206 | return -ENOMEM; | |
1207 | ||
e43473b7 | 1208 | td->tg_service_tree = THROTL_RB_ROOT; |
a29a171e | 1209 | INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work); |
e43473b7 | 1210 | |
cd1604fa | 1211 | q->td = td; |
29b12589 | 1212 | td->queue = q; |
02977e4a | 1213 | |
a2b1693b | 1214 | /* activate policy */ |
3c798398 | 1215 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1216 | if (ret) |
f51b802c | 1217 | kfree(td); |
a2b1693b | 1218 | return ret; |
e43473b7 VG |
1219 | } |
1220 | ||
1221 | void blk_throtl_exit(struct request_queue *q) | |
1222 | { | |
c875f4d0 | 1223 | BUG_ON(!q->td); |
da527770 | 1224 | throtl_shutdown_wq(q); |
3c798398 | 1225 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1226 | kfree(q->td); |
e43473b7 VG |
1227 | } |
1228 | ||
1229 | static int __init throtl_init(void) | |
1230 | { | |
450adcbe VG |
1231 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1232 | if (!kthrotld_workqueue) | |
1233 | panic("Failed to create kthrotld\n"); | |
1234 | ||
3c798398 | 1235 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1236 | } |
1237 | ||
1238 | module_init(throtl_init); |