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 VG |
87 | |
88 | /* Some throttle limits got updated for the group */ | |
6f037937 | 89 | int limits_changed; |
8a3d2615 TH |
90 | |
91 | /* Per cpu stats pointer */ | |
92 | struct tg_stats_cpu __percpu *stats_cpu; | |
93 | ||
94 | /* List of tgs waiting for per cpu stats memory to be allocated */ | |
95 | struct list_head stats_alloc_node; | |
e43473b7 VG |
96 | }; |
97 | ||
98 | struct throtl_data | |
99 | { | |
e43473b7 VG |
100 | /* service tree for active throtl groups */ |
101 | struct throtl_rb_root tg_service_tree; | |
102 | ||
e43473b7 VG |
103 | struct request_queue *queue; |
104 | ||
105 | /* Total Number of queued bios on READ and WRITE lists */ | |
106 | unsigned int nr_queued[2]; | |
107 | ||
108 | /* | |
02977e4a | 109 | * number of total undestroyed groups |
e43473b7 VG |
110 | */ |
111 | unsigned int nr_undestroyed_grps; | |
112 | ||
113 | /* Work for dispatching throttled bios */ | |
114 | struct delayed_work throtl_work; | |
fe071437 | 115 | |
6f037937 | 116 | int limits_changed; |
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 | ||
e43473b7 VG |
146 | enum tg_state_flags { |
147 | THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */ | |
148 | }; | |
149 | ||
150 | #define THROTL_TG_FNS(name) \ | |
151 | static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \ | |
152 | { \ | |
153 | (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \ | |
154 | } \ | |
155 | static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \ | |
156 | { \ | |
157 | (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \ | |
158 | } \ | |
159 | static inline int throtl_tg_##name(const struct throtl_grp *tg) \ | |
160 | { \ | |
161 | return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \ | |
162 | } | |
163 | ||
164 | THROTL_TG_FNS(on_rr); | |
165 | ||
54e7ed12 TH |
166 | #define throtl_log_tg(td, tg, fmt, args...) do { \ |
167 | char __pbuf[128]; \ | |
168 | \ | |
169 | blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \ | |
170 | blk_add_trace_msg((td)->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
171 | } while (0) | |
e43473b7 VG |
172 | |
173 | #define throtl_log(td, fmt, args...) \ | |
174 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) | |
175 | ||
d2f31a5f | 176 | static inline unsigned int total_nr_queued(struct throtl_data *td) |
e43473b7 | 177 | { |
d2f31a5f | 178 | return td->nr_queued[0] + td->nr_queued[1]; |
e43473b7 VG |
179 | } |
180 | ||
8a3d2615 TH |
181 | /* |
182 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
183 | * system_nrt_wq once there are some groups on the alloc_list waiting for | |
184 | * allocation. | |
185 | */ | |
186 | static void tg_stats_alloc_fn(struct work_struct *work) | |
187 | { | |
188 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
189 | struct delayed_work *dwork = to_delayed_work(work); | |
190 | bool empty = false; | |
191 | ||
192 | alloc_stats: | |
193 | if (!stats_cpu) { | |
194 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
195 | if (!stats_cpu) { | |
196 | /* allocation failed, try again after some time */ | |
197 | queue_delayed_work(system_nrt_wq, dwork, | |
198 | msecs_to_jiffies(10)); | |
199 | return; | |
200 | } | |
201 | } | |
202 | ||
203 | spin_lock_irq(&tg_stats_alloc_lock); | |
204 | ||
205 | if (!list_empty(&tg_stats_alloc_list)) { | |
206 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
207 | struct throtl_grp, | |
208 | stats_alloc_node); | |
209 | swap(tg->stats_cpu, stats_cpu); | |
210 | list_del_init(&tg->stats_alloc_node); | |
211 | } | |
212 | ||
213 | empty = list_empty(&tg_stats_alloc_list); | |
214 | spin_unlock_irq(&tg_stats_alloc_lock); | |
215 | if (!empty) | |
216 | goto alloc_stats; | |
217 | } | |
218 | ||
3c798398 | 219 | static void throtl_pd_init(struct blkcg_gq *blkg) |
a29a171e | 220 | { |
0381411e | 221 | struct throtl_grp *tg = blkg_to_tg(blkg); |
ff26eaad | 222 | unsigned long flags; |
cd1604fa | 223 | |
a29a171e VG |
224 | RB_CLEAR_NODE(&tg->rb_node); |
225 | bio_list_init(&tg->bio_lists[0]); | |
226 | bio_list_init(&tg->bio_lists[1]); | |
227 | tg->limits_changed = false; | |
228 | ||
e56da7e2 TH |
229 | tg->bps[READ] = -1; |
230 | tg->bps[WRITE] = -1; | |
231 | tg->iops[READ] = -1; | |
232 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
233 | |
234 | /* | |
235 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
236 | * but percpu allocator can't be called from IO path. Queue tg on | |
237 | * tg_stats_alloc_list and allocate from work item. | |
238 | */ | |
ff26eaad | 239 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
240 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); |
241 | queue_delayed_work(system_nrt_wq, &tg_stats_alloc_work, 0); | |
ff26eaad | 242 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
243 | } |
244 | ||
3c798398 | 245 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
8a3d2615 TH |
246 | { |
247 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
ff26eaad | 248 | unsigned long flags; |
8a3d2615 | 249 | |
ff26eaad | 250 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 251 | list_del_init(&tg->stats_alloc_node); |
ff26eaad | 252 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
253 | |
254 | free_percpu(tg->stats_cpu); | |
255 | } | |
256 | ||
3c798398 | 257 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
8a3d2615 TH |
258 | { |
259 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
260 | int cpu; | |
261 | ||
262 | if (tg->stats_cpu == NULL) | |
263 | return; | |
264 | ||
265 | for_each_possible_cpu(cpu) { | |
266 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
267 | ||
268 | blkg_rwstat_reset(&sc->service_bytes); | |
269 | blkg_rwstat_reset(&sc->serviced); | |
270 | } | |
a29a171e VG |
271 | } |
272 | ||
3c798398 TH |
273 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
274 | struct blkcg *blkcg) | |
e43473b7 | 275 | { |
be2c6b19 | 276 | /* |
3c798398 TH |
277 | * This is the common case when there are no blkcgs. Avoid lookup |
278 | * in this case | |
cd1604fa | 279 | */ |
3c798398 | 280 | if (blkcg == &blkcg_root) |
03d8e111 | 281 | return td_root_tg(td); |
e43473b7 | 282 | |
e8989fae | 283 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
284 | } |
285 | ||
cd1604fa | 286 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
3c798398 | 287 | struct blkcg *blkcg) |
e43473b7 | 288 | { |
f469a7b4 | 289 | struct request_queue *q = td->queue; |
cd1604fa | 290 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 291 | |
f469a7b4 | 292 | /* |
3c798398 TH |
293 | * This is the common case when there are no blkcgs. Avoid lookup |
294 | * in this case | |
f469a7b4 | 295 | */ |
3c798398 | 296 | if (blkcg == &blkcg_root) { |
03d8e111 | 297 | tg = td_root_tg(td); |
cd1604fa | 298 | } else { |
3c798398 | 299 | struct blkcg_gq *blkg; |
f469a7b4 | 300 | |
3c96cb32 | 301 | blkg = blkg_lookup_create(blkcg, q); |
f469a7b4 | 302 | |
cd1604fa TH |
303 | /* if %NULL and @q is alive, fall back to root_tg */ |
304 | if (!IS_ERR(blkg)) | |
0381411e | 305 | tg = blkg_to_tg(blkg); |
cd1604fa | 306 | else if (!blk_queue_dead(q)) |
03d8e111 | 307 | tg = td_root_tg(td); |
f469a7b4 VG |
308 | } |
309 | ||
e43473b7 VG |
310 | return tg; |
311 | } | |
312 | ||
313 | static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root) | |
314 | { | |
315 | /* Service tree is empty */ | |
316 | if (!root->count) | |
317 | return NULL; | |
318 | ||
319 | if (!root->left) | |
320 | root->left = rb_first(&root->rb); | |
321 | ||
322 | if (root->left) | |
323 | return rb_entry_tg(root->left); | |
324 | ||
325 | return NULL; | |
326 | } | |
327 | ||
328 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
329 | { | |
330 | rb_erase(n, root); | |
331 | RB_CLEAR_NODE(n); | |
332 | } | |
333 | ||
334 | static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root) | |
335 | { | |
336 | if (root->left == n) | |
337 | root->left = NULL; | |
338 | rb_erase_init(n, &root->rb); | |
339 | --root->count; | |
340 | } | |
341 | ||
342 | static void update_min_dispatch_time(struct throtl_rb_root *st) | |
343 | { | |
344 | struct throtl_grp *tg; | |
345 | ||
346 | tg = throtl_rb_first(st); | |
347 | if (!tg) | |
348 | return; | |
349 | ||
350 | st->min_disptime = tg->disptime; | |
351 | } | |
352 | ||
353 | static void | |
354 | tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg) | |
355 | { | |
356 | struct rb_node **node = &st->rb.rb_node; | |
357 | struct rb_node *parent = NULL; | |
358 | struct throtl_grp *__tg; | |
359 | unsigned long key = tg->disptime; | |
360 | int left = 1; | |
361 | ||
362 | while (*node != NULL) { | |
363 | parent = *node; | |
364 | __tg = rb_entry_tg(parent); | |
365 | ||
366 | if (time_before(key, __tg->disptime)) | |
367 | node = &parent->rb_left; | |
368 | else { | |
369 | node = &parent->rb_right; | |
370 | left = 0; | |
371 | } | |
372 | } | |
373 | ||
374 | if (left) | |
375 | st->left = &tg->rb_node; | |
376 | ||
377 | rb_link_node(&tg->rb_node, parent, node); | |
378 | rb_insert_color(&tg->rb_node, &st->rb); | |
379 | } | |
380 | ||
381 | static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
382 | { | |
383 | struct throtl_rb_root *st = &td->tg_service_tree; | |
384 | ||
385 | tg_service_tree_add(st, tg); | |
386 | throtl_mark_tg_on_rr(tg); | |
387 | st->count++; | |
388 | } | |
389 | ||
390 | static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
391 | { | |
392 | if (!throtl_tg_on_rr(tg)) | |
393 | __throtl_enqueue_tg(td, tg); | |
394 | } | |
395 | ||
396 | static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
397 | { | |
398 | throtl_rb_erase(&tg->rb_node, &td->tg_service_tree); | |
399 | throtl_clear_tg_on_rr(tg); | |
400 | } | |
401 | ||
402 | static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
403 | { | |
404 | if (throtl_tg_on_rr(tg)) | |
405 | __throtl_dequeue_tg(td, tg); | |
406 | } | |
407 | ||
408 | static void throtl_schedule_next_dispatch(struct throtl_data *td) | |
409 | { | |
410 | struct throtl_rb_root *st = &td->tg_service_tree; | |
411 | ||
412 | /* | |
413 | * If there are more bios pending, schedule more work. | |
414 | */ | |
415 | if (!total_nr_queued(td)) | |
416 | return; | |
417 | ||
418 | BUG_ON(!st->count); | |
419 | ||
420 | update_min_dispatch_time(st); | |
421 | ||
422 | if (time_before_eq(st->min_disptime, jiffies)) | |
450adcbe | 423 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 424 | else |
450adcbe | 425 | throtl_schedule_delayed_work(td, (st->min_disptime - jiffies)); |
e43473b7 VG |
426 | } |
427 | ||
428 | static inline void | |
429 | throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
430 | { | |
431 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 432 | tg->io_disp[rw] = 0; |
e43473b7 VG |
433 | tg->slice_start[rw] = jiffies; |
434 | tg->slice_end[rw] = jiffies + throtl_slice; | |
435 | throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
436 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
437 | tg->slice_end[rw], jiffies); | |
438 | } | |
439 | ||
d1ae8ffd VG |
440 | static inline void throtl_set_slice_end(struct throtl_data *td, |
441 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
442 | { | |
443 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
444 | } | |
445 | ||
e43473b7 VG |
446 | static inline void throtl_extend_slice(struct throtl_data *td, |
447 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
448 | { | |
449 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
450 | throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
451 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
452 | tg->slice_end[rw], jiffies); | |
453 | } | |
454 | ||
455 | /* Determine if previously allocated or extended slice is complete or not */ | |
456 | static bool | |
457 | throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
458 | { | |
459 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
460 | return 0; | |
461 | ||
462 | return 1; | |
463 | } | |
464 | ||
465 | /* Trim the used slices and adjust slice start accordingly */ | |
466 | static inline void | |
467 | throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
468 | { | |
3aad5d3e VG |
469 | unsigned long nr_slices, time_elapsed, io_trim; |
470 | u64 bytes_trim, tmp; | |
e43473b7 VG |
471 | |
472 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
473 | ||
474 | /* | |
475 | * If bps are unlimited (-1), then time slice don't get | |
476 | * renewed. Don't try to trim the slice if slice is used. A new | |
477 | * slice will start when appropriate. | |
478 | */ | |
479 | if (throtl_slice_used(td, tg, rw)) | |
480 | return; | |
481 | ||
d1ae8ffd VG |
482 | /* |
483 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
484 | * that initially cgroup limit was very low resulting in high | |
485 | * slice_end, but later limit was bumped up and bio was dispached | |
486 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
487 | * is bad because it does not allow new slice to start. | |
488 | */ | |
489 | ||
490 | throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice); | |
491 | ||
e43473b7 VG |
492 | time_elapsed = jiffies - tg->slice_start[rw]; |
493 | ||
494 | nr_slices = time_elapsed / throtl_slice; | |
495 | ||
496 | if (!nr_slices) | |
497 | return; | |
3aad5d3e VG |
498 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
499 | do_div(tmp, HZ); | |
500 | bytes_trim = tmp; | |
e43473b7 | 501 | |
8e89d13f | 502 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 503 | |
8e89d13f | 504 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
505 | return; |
506 | ||
507 | if (tg->bytes_disp[rw] >= bytes_trim) | |
508 | tg->bytes_disp[rw] -= bytes_trim; | |
509 | else | |
510 | tg->bytes_disp[rw] = 0; | |
511 | ||
8e89d13f VG |
512 | if (tg->io_disp[rw] >= io_trim) |
513 | tg->io_disp[rw] -= io_trim; | |
514 | else | |
515 | tg->io_disp[rw] = 0; | |
516 | ||
e43473b7 VG |
517 | tg->slice_start[rw] += nr_slices * throtl_slice; |
518 | ||
3aad5d3e | 519 | throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 520 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 521 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
522 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
523 | } | |
524 | ||
8e89d13f VG |
525 | static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg, |
526 | struct bio *bio, unsigned long *wait) | |
e43473b7 VG |
527 | { |
528 | bool rw = bio_data_dir(bio); | |
8e89d13f | 529 | unsigned int io_allowed; |
e43473b7 | 530 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 531 | u64 tmp; |
e43473b7 | 532 | |
8e89d13f | 533 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 534 | |
8e89d13f VG |
535 | /* Slice has just started. Consider one slice interval */ |
536 | if (!jiffy_elapsed) | |
537 | jiffy_elapsed_rnd = throtl_slice; | |
538 | ||
539 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
540 | ||
c49c06e4 VG |
541 | /* |
542 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
543 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
544 | * will allow dispatch after 1 second and after that slice should | |
545 | * have been trimmed. | |
546 | */ | |
547 | ||
548 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
549 | do_div(tmp, HZ); | |
550 | ||
551 | if (tmp > UINT_MAX) | |
552 | io_allowed = UINT_MAX; | |
553 | else | |
554 | io_allowed = tmp; | |
8e89d13f VG |
555 | |
556 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
557 | if (wait) |
558 | *wait = 0; | |
559 | return 1; | |
560 | } | |
561 | ||
8e89d13f VG |
562 | /* Calc approx time to dispatch */ |
563 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
564 | ||
565 | if (jiffy_wait > jiffy_elapsed) | |
566 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
567 | else | |
568 | jiffy_wait = 1; | |
569 | ||
570 | if (wait) | |
571 | *wait = jiffy_wait; | |
572 | return 0; | |
573 | } | |
574 | ||
575 | static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg, | |
576 | struct bio *bio, unsigned long *wait) | |
577 | { | |
578 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 579 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 580 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
581 | |
582 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
583 | ||
584 | /* Slice has just started. Consider one slice interval */ | |
585 | if (!jiffy_elapsed) | |
586 | jiffy_elapsed_rnd = throtl_slice; | |
587 | ||
588 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
589 | ||
5e901a2b VG |
590 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
591 | do_div(tmp, HZ); | |
3aad5d3e | 592 | bytes_allowed = tmp; |
e43473b7 VG |
593 | |
594 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
595 | if (wait) | |
596 | *wait = 0; | |
597 | return 1; | |
598 | } | |
599 | ||
600 | /* Calc approx time to dispatch */ | |
601 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
602 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
603 | ||
604 | if (!jiffy_wait) | |
605 | jiffy_wait = 1; | |
606 | ||
607 | /* | |
608 | * This wait time is without taking into consideration the rounding | |
609 | * up we did. Add that time also. | |
610 | */ | |
611 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
612 | if (wait) |
613 | *wait = jiffy_wait; | |
8e89d13f VG |
614 | return 0; |
615 | } | |
616 | ||
af75cd3c VG |
617 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
618 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
619 | return 1; | |
620 | return 0; | |
621 | } | |
622 | ||
8e89d13f VG |
623 | /* |
624 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
625 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
626 | */ | |
627 | static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg, | |
628 | struct bio *bio, unsigned long *wait) | |
629 | { | |
630 | bool rw = bio_data_dir(bio); | |
631 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
632 | ||
633 | /* | |
634 | * Currently whole state machine of group depends on first bio | |
635 | * queued in the group bio list. So one should not be calling | |
636 | * this function with a different bio if there are other bios | |
637 | * queued. | |
638 | */ | |
639 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); | |
e43473b7 | 640 | |
8e89d13f VG |
641 | /* If tg->bps = -1, then BW is unlimited */ |
642 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
643 | if (wait) | |
644 | *wait = 0; | |
645 | return 1; | |
646 | } | |
647 | ||
648 | /* | |
649 | * If previous slice expired, start a new one otherwise renew/extend | |
650 | * existing slice to make sure it is at least throtl_slice interval | |
651 | * long since now. | |
652 | */ | |
653 | if (throtl_slice_used(td, tg, rw)) | |
654 | throtl_start_new_slice(td, tg, rw); | |
655 | else { | |
656 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
657 | throtl_extend_slice(td, tg, rw, jiffies + throtl_slice); | |
658 | } | |
659 | ||
660 | if (tg_with_in_bps_limit(td, tg, bio, &bps_wait) | |
661 | && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) { | |
662 | if (wait) | |
663 | *wait = 0; | |
664 | return 1; | |
665 | } | |
666 | ||
667 | max_wait = max(bps_wait, iops_wait); | |
668 | ||
669 | if (wait) | |
670 | *wait = max_wait; | |
671 | ||
672 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
673 | throtl_extend_slice(td, tg, rw, jiffies + max_wait); | |
e43473b7 VG |
674 | |
675 | return 0; | |
676 | } | |
677 | ||
3c798398 | 678 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
679 | int rw) |
680 | { | |
8a3d2615 TH |
681 | struct throtl_grp *tg = blkg_to_tg(blkg); |
682 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
683 | unsigned long flags; |
684 | ||
685 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 686 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
687 | return; |
688 | ||
689 | /* | |
690 | * Disabling interrupts to provide mutual exclusion between two | |
691 | * writes on same cpu. It probably is not needed for 64bit. Not | |
692 | * optimizing that case yet. | |
693 | */ | |
694 | local_irq_save(flags); | |
695 | ||
8a3d2615 | 696 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 697 | |
629ed0b1 TH |
698 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
699 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
700 | ||
701 | local_irq_restore(flags); | |
702 | } | |
703 | ||
e43473b7 VG |
704 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
705 | { | |
706 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
707 | |
708 | /* Charge the bio to the group */ | |
709 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 710 | tg->io_disp[rw]++; |
e43473b7 | 711 | |
629ed0b1 | 712 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw); |
e43473b7 VG |
713 | } |
714 | ||
715 | static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg, | |
716 | struct bio *bio) | |
717 | { | |
718 | bool rw = bio_data_dir(bio); | |
719 | ||
720 | bio_list_add(&tg->bio_lists[rw], bio); | |
721 | /* Take a bio reference on tg */ | |
1adaf3dd | 722 | blkg_get(tg_to_blkg(tg)); |
e43473b7 VG |
723 | tg->nr_queued[rw]++; |
724 | td->nr_queued[rw]++; | |
725 | throtl_enqueue_tg(td, tg); | |
726 | } | |
727 | ||
728 | static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg) | |
729 | { | |
730 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
731 | struct bio *bio; | |
732 | ||
733 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
734 | tg_may_dispatch(td, tg, bio, &read_wait); | |
735 | ||
736 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
737 | tg_may_dispatch(td, tg, bio, &write_wait); | |
738 | ||
739 | min_wait = min(read_wait, write_wait); | |
740 | disptime = jiffies + min_wait; | |
741 | ||
e43473b7 VG |
742 | /* Update dispatch time */ |
743 | throtl_dequeue_tg(td, tg); | |
744 | tg->disptime = disptime; | |
745 | throtl_enqueue_tg(td, tg); | |
746 | } | |
747 | ||
748 | static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg, | |
749 | bool rw, struct bio_list *bl) | |
750 | { | |
751 | struct bio *bio; | |
752 | ||
753 | bio = bio_list_pop(&tg->bio_lists[rw]); | |
754 | tg->nr_queued[rw]--; | |
1adaf3dd TH |
755 | /* Drop bio reference on blkg */ |
756 | blkg_put(tg_to_blkg(tg)); | |
e43473b7 VG |
757 | |
758 | BUG_ON(td->nr_queued[rw] <= 0); | |
759 | td->nr_queued[rw]--; | |
760 | ||
761 | throtl_charge_bio(tg, bio); | |
762 | bio_list_add(bl, bio); | |
763 | bio->bi_rw |= REQ_THROTTLED; | |
764 | ||
765 | throtl_trim_slice(td, tg, rw); | |
766 | } | |
767 | ||
768 | static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg, | |
769 | struct bio_list *bl) | |
770 | { | |
771 | unsigned int nr_reads = 0, nr_writes = 0; | |
772 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 773 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
774 | struct bio *bio; |
775 | ||
776 | /* Try to dispatch 75% READS and 25% WRITES */ | |
777 | ||
778 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) | |
779 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
780 | ||
781 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
782 | nr_reads++; | |
783 | ||
784 | if (nr_reads >= max_nr_reads) | |
785 | break; | |
786 | } | |
787 | ||
788 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) | |
789 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
790 | ||
791 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
792 | nr_writes++; | |
793 | ||
794 | if (nr_writes >= max_nr_writes) | |
795 | break; | |
796 | } | |
797 | ||
798 | return nr_reads + nr_writes; | |
799 | } | |
800 | ||
801 | static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl) | |
802 | { | |
803 | unsigned int nr_disp = 0; | |
804 | struct throtl_grp *tg; | |
805 | struct throtl_rb_root *st = &td->tg_service_tree; | |
806 | ||
807 | while (1) { | |
808 | tg = throtl_rb_first(st); | |
809 | ||
810 | if (!tg) | |
811 | break; | |
812 | ||
813 | if (time_before(jiffies, tg->disptime)) | |
814 | break; | |
815 | ||
816 | throtl_dequeue_tg(td, tg); | |
817 | ||
818 | nr_disp += throtl_dispatch_tg(td, tg, bl); | |
819 | ||
820 | if (tg->nr_queued[0] || tg->nr_queued[1]) { | |
821 | tg_update_disptime(td, tg); | |
822 | throtl_enqueue_tg(td, tg); | |
823 | } | |
824 | ||
825 | if (nr_disp >= throtl_quantum) | |
826 | break; | |
827 | } | |
828 | ||
829 | return nr_disp; | |
830 | } | |
831 | ||
fe071437 VG |
832 | static void throtl_process_limit_change(struct throtl_data *td) |
833 | { | |
4eef3049 | 834 | struct request_queue *q = td->queue; |
3c798398 | 835 | struct blkcg_gq *blkg, *n; |
fe071437 | 836 | |
de701c74 | 837 | if (!td->limits_changed) |
fe071437 VG |
838 | return; |
839 | ||
de701c74 | 840 | xchg(&td->limits_changed, false); |
fe071437 | 841 | |
de701c74 | 842 | throtl_log(td, "limits changed"); |
fe071437 | 843 | |
e8989fae | 844 | list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) { |
4eef3049 TH |
845 | struct throtl_grp *tg = blkg_to_tg(blkg); |
846 | ||
de701c74 VG |
847 | if (!tg->limits_changed) |
848 | continue; | |
849 | ||
850 | if (!xchg(&tg->limits_changed, false)) | |
851 | continue; | |
852 | ||
853 | throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu" | |
854 | " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE], | |
855 | tg->iops[READ], tg->iops[WRITE]); | |
856 | ||
04521db0 VG |
857 | /* |
858 | * Restart the slices for both READ and WRITES. It | |
859 | * might happen that a group's limit are dropped | |
860 | * suddenly and we don't want to account recently | |
861 | * dispatched IO with new low rate | |
862 | */ | |
863 | throtl_start_new_slice(td, tg, 0); | |
864 | throtl_start_new_slice(td, tg, 1); | |
865 | ||
de701c74 | 866 | if (throtl_tg_on_rr(tg)) |
fe071437 | 867 | tg_update_disptime(td, tg); |
fe071437 | 868 | } |
fe071437 VG |
869 | } |
870 | ||
e43473b7 VG |
871 | /* Dispatch throttled bios. Should be called without queue lock held. */ |
872 | static int throtl_dispatch(struct request_queue *q) | |
873 | { | |
874 | struct throtl_data *td = q->td; | |
875 | unsigned int nr_disp = 0; | |
876 | struct bio_list bio_list_on_stack; | |
877 | struct bio *bio; | |
69d60eb9 | 878 | struct blk_plug plug; |
e43473b7 VG |
879 | |
880 | spin_lock_irq(q->queue_lock); | |
881 | ||
fe071437 VG |
882 | throtl_process_limit_change(td); |
883 | ||
e43473b7 VG |
884 | if (!total_nr_queued(td)) |
885 | goto out; | |
886 | ||
887 | bio_list_init(&bio_list_on_stack); | |
888 | ||
d2f31a5f | 889 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
e43473b7 VG |
890 | total_nr_queued(td), td->nr_queued[READ], |
891 | td->nr_queued[WRITE]); | |
892 | ||
893 | nr_disp = throtl_select_dispatch(td, &bio_list_on_stack); | |
894 | ||
895 | if (nr_disp) | |
896 | throtl_log(td, "bios disp=%u", nr_disp); | |
897 | ||
898 | throtl_schedule_next_dispatch(td); | |
899 | out: | |
900 | spin_unlock_irq(q->queue_lock); | |
901 | ||
902 | /* | |
903 | * If we dispatched some requests, unplug the queue to make sure | |
904 | * immediate dispatch | |
905 | */ | |
906 | if (nr_disp) { | |
69d60eb9 | 907 | blk_start_plug(&plug); |
e43473b7 VG |
908 | while((bio = bio_list_pop(&bio_list_on_stack))) |
909 | generic_make_request(bio); | |
69d60eb9 | 910 | blk_finish_plug(&plug); |
e43473b7 VG |
911 | } |
912 | return nr_disp; | |
913 | } | |
914 | ||
915 | void blk_throtl_work(struct work_struct *work) | |
916 | { | |
917 | struct throtl_data *td = container_of(work, struct throtl_data, | |
918 | throtl_work.work); | |
919 | struct request_queue *q = td->queue; | |
920 | ||
921 | throtl_dispatch(q); | |
922 | } | |
923 | ||
924 | /* Call with queue lock held */ | |
450adcbe VG |
925 | static void |
926 | throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay) | |
e43473b7 VG |
927 | { |
928 | ||
e43473b7 VG |
929 | struct delayed_work *dwork = &td->throtl_work; |
930 | ||
04521db0 | 931 | /* schedule work if limits changed even if no bio is queued */ |
d2f31a5f | 932 | if (total_nr_queued(td) || td->limits_changed) { |
e43473b7 VG |
933 | /* |
934 | * We might have a work scheduled to be executed in future. | |
935 | * Cancel that and schedule a new one. | |
936 | */ | |
937 | __cancel_delayed_work(dwork); | |
450adcbe | 938 | queue_delayed_work(kthrotld_workqueue, dwork, delay); |
e43473b7 VG |
939 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", |
940 | delay, jiffies); | |
941 | } | |
942 | } | |
e43473b7 | 943 | |
f95a04af TH |
944 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
945 | struct blkg_policy_data *pd, int off) | |
41b38b6d | 946 | { |
f95a04af | 947 | struct throtl_grp *tg = pd_to_tg(pd); |
41b38b6d TH |
948 | struct blkg_rwstat rwstat = { }, tmp; |
949 | int i, cpu; | |
950 | ||
951 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 952 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
953 | |
954 | tmp = blkg_rwstat_read((void *)sc + off); | |
955 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
956 | rwstat.cnt[i] += tmp.cnt[i]; | |
957 | } | |
958 | ||
f95a04af | 959 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
41b38b6d TH |
960 | } |
961 | ||
8a3d2615 TH |
962 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
963 | struct seq_file *sf) | |
41b38b6d | 964 | { |
3c798398 | 965 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
41b38b6d | 966 | |
3c798398 | 967 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 968 | cft->private, true); |
41b38b6d TH |
969 | return 0; |
970 | } | |
971 | ||
f95a04af TH |
972 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
973 | int off) | |
60c2bc2d | 974 | { |
f95a04af TH |
975 | struct throtl_grp *tg = pd_to_tg(pd); |
976 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 977 | |
af133ceb | 978 | if (v == -1) |
60c2bc2d | 979 | return 0; |
f95a04af | 980 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
981 | } |
982 | ||
f95a04af TH |
983 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
984 | int off) | |
e43473b7 | 985 | { |
f95a04af TH |
986 | struct throtl_grp *tg = pd_to_tg(pd); |
987 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 988 | |
af133ceb TH |
989 | if (v == -1) |
990 | return 0; | |
f95a04af | 991 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
992 | } |
993 | ||
af133ceb TH |
994 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
995 | struct seq_file *sf) | |
8e89d13f | 996 | { |
3c798398 TH |
997 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, |
998 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 999 | return 0; |
8e89d13f VG |
1000 | } |
1001 | ||
af133ceb TH |
1002 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1003 | struct seq_file *sf) | |
8e89d13f | 1004 | { |
3c798398 TH |
1005 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, |
1006 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 1007 | return 0; |
60c2bc2d TH |
1008 | } |
1009 | ||
af133ceb TH |
1010 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
1011 | bool is_u64) | |
60c2bc2d | 1012 | { |
3c798398 | 1013 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
60c2bc2d | 1014 | struct blkg_conf_ctx ctx; |
af133ceb | 1015 | struct throtl_grp *tg; |
a2b1693b | 1016 | struct throtl_data *td; |
60c2bc2d TH |
1017 | int ret; |
1018 | ||
3c798398 | 1019 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
1020 | if (ret) |
1021 | return ret; | |
1022 | ||
af133ceb | 1023 | tg = blkg_to_tg(ctx.blkg); |
a2b1693b | 1024 | td = ctx.blkg->q->td; |
af133ceb | 1025 | |
a2b1693b TH |
1026 | if (!ctx.v) |
1027 | ctx.v = -1; | |
af133ceb | 1028 | |
a2b1693b TH |
1029 | if (is_u64) |
1030 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
1031 | else | |
1032 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 1033 | |
a2b1693b TH |
1034 | /* XXX: we don't need the following deferred processing */ |
1035 | xchg(&tg->limits_changed, true); | |
1036 | xchg(&td->limits_changed, true); | |
1037 | throtl_schedule_delayed_work(td, 0); | |
60c2bc2d TH |
1038 | |
1039 | blkg_conf_finish(&ctx); | |
a2b1693b | 1040 | return 0; |
8e89d13f VG |
1041 | } |
1042 | ||
af133ceb TH |
1043 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1044 | const char *buf) | |
60c2bc2d | 1045 | { |
af133ceb | 1046 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
1047 | } |
1048 | ||
af133ceb TH |
1049 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1050 | const char *buf) | |
60c2bc2d | 1051 | { |
af133ceb | 1052 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
1053 | } |
1054 | ||
1055 | static struct cftype throtl_files[] = { | |
1056 | { | |
1057 | .name = "throttle.read_bps_device", | |
af133ceb TH |
1058 | .private = offsetof(struct throtl_grp, bps[READ]), |
1059 | .read_seq_string = tg_print_conf_u64, | |
1060 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1061 | .max_write_len = 256, |
1062 | }, | |
1063 | { | |
1064 | .name = "throttle.write_bps_device", | |
af133ceb TH |
1065 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
1066 | .read_seq_string = tg_print_conf_u64, | |
1067 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1068 | .max_write_len = 256, |
1069 | }, | |
1070 | { | |
1071 | .name = "throttle.read_iops_device", | |
af133ceb TH |
1072 | .private = offsetof(struct throtl_grp, iops[READ]), |
1073 | .read_seq_string = tg_print_conf_uint, | |
1074 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1075 | .max_write_len = 256, |
1076 | }, | |
1077 | { | |
1078 | .name = "throttle.write_iops_device", | |
af133ceb TH |
1079 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
1080 | .read_seq_string = tg_print_conf_uint, | |
1081 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1082 | .max_write_len = 256, |
1083 | }, | |
1084 | { | |
1085 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1086 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1087 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1088 | }, |
1089 | { | |
1090 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1091 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1092 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1093 | }, |
1094 | { } /* terminate */ | |
1095 | }; | |
1096 | ||
da527770 | 1097 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1098 | { |
1099 | struct throtl_data *td = q->td; | |
1100 | ||
1101 | cancel_delayed_work_sync(&td->throtl_work); | |
1102 | } | |
1103 | ||
3c798398 | 1104 | static struct blkcg_policy blkcg_policy_throtl = { |
f9fcc2d3 TH |
1105 | .pd_size = sizeof(struct throtl_grp), |
1106 | .cftypes = throtl_files, | |
1107 | ||
1108 | .pd_init_fn = throtl_pd_init, | |
1109 | .pd_exit_fn = throtl_pd_exit, | |
1110 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 VG |
1111 | }; |
1112 | ||
bc16a4f9 | 1113 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1114 | { |
1115 | struct throtl_data *td = q->td; | |
1116 | struct throtl_grp *tg; | |
e43473b7 | 1117 | bool rw = bio_data_dir(bio), update_disptime = true; |
3c798398 | 1118 | struct blkcg *blkcg; |
bc16a4f9 | 1119 | bool throttled = false; |
e43473b7 VG |
1120 | |
1121 | if (bio->bi_rw & REQ_THROTTLED) { | |
1122 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1123 | goto out; |
e43473b7 VG |
1124 | } |
1125 | ||
671058fb TH |
1126 | /* bio_associate_current() needs ioc, try creating */ |
1127 | create_io_context(GFP_ATOMIC, q->node); | |
1128 | ||
af75cd3c VG |
1129 | /* |
1130 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1131 | * basic fields like stats and io rates. If a group has no rules, | |
1132 | * just update the dispatch stats in lockless manner and return. | |
1133 | */ | |
af75cd3c | 1134 | rcu_read_lock(); |
3c798398 | 1135 | blkcg = bio_blkcg(bio); |
cd1604fa | 1136 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1137 | if (tg) { |
af75cd3c | 1138 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1139 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1140 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1141 | goto out_unlock_rcu; |
af75cd3c VG |
1142 | } |
1143 | } | |
af75cd3c VG |
1144 | |
1145 | /* | |
1146 | * Either group has not been allocated yet or it is not an unlimited | |
1147 | * IO group | |
1148 | */ | |
e43473b7 | 1149 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1150 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1151 | if (unlikely(!tg)) |
1152 | goto out_unlock; | |
f469a7b4 | 1153 | |
e43473b7 VG |
1154 | if (tg->nr_queued[rw]) { |
1155 | /* | |
1156 | * There is already another bio queued in same dir. No | |
1157 | * need to update dispatch time. | |
1158 | */ | |
231d704b | 1159 | update_disptime = false; |
e43473b7 | 1160 | goto queue_bio; |
de701c74 | 1161 | |
e43473b7 VG |
1162 | } |
1163 | ||
1164 | /* Bio is with-in rate limit of group */ | |
1165 | if (tg_may_dispatch(td, tg, bio, NULL)) { | |
1166 | throtl_charge_bio(tg, bio); | |
04521db0 VG |
1167 | |
1168 | /* | |
1169 | * We need to trim slice even when bios are not being queued | |
1170 | * otherwise it might happen that a bio is not queued for | |
1171 | * a long time and slice keeps on extending and trim is not | |
1172 | * called for a long time. Now if limits are reduced suddenly | |
1173 | * we take into account all the IO dispatched so far at new | |
1174 | * low rate and * newly queued IO gets a really long dispatch | |
1175 | * time. | |
1176 | * | |
1177 | * So keep on trimming slice even if bio is not queued. | |
1178 | */ | |
1179 | throtl_trim_slice(td, tg, rw); | |
bc16a4f9 | 1180 | goto out_unlock; |
e43473b7 VG |
1181 | } |
1182 | ||
1183 | queue_bio: | |
fd16d263 | 1184 | throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
8e89d13f VG |
1185 | " iodisp=%u iops=%u queued=%d/%d", |
1186 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1187 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1188 | tg->io_disp[rw], tg->iops[rw], |
e43473b7 VG |
1189 | tg->nr_queued[READ], tg->nr_queued[WRITE]); |
1190 | ||
671058fb | 1191 | bio_associate_current(bio); |
e43473b7 | 1192 | throtl_add_bio_tg(q->td, tg, bio); |
bc16a4f9 | 1193 | throttled = true; |
e43473b7 VG |
1194 | |
1195 | if (update_disptime) { | |
1196 | tg_update_disptime(td, tg); | |
1197 | throtl_schedule_next_dispatch(td); | |
1198 | } | |
1199 | ||
bc16a4f9 | 1200 | out_unlock: |
e43473b7 | 1201 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1202 | out_unlock_rcu: |
1203 | rcu_read_unlock(); | |
bc16a4f9 TH |
1204 | out: |
1205 | return throttled; | |
e43473b7 VG |
1206 | } |
1207 | ||
c9a929dd TH |
1208 | /** |
1209 | * blk_throtl_drain - drain throttled bios | |
1210 | * @q: request_queue to drain throttled bios for | |
1211 | * | |
1212 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1213 | */ | |
1214 | void blk_throtl_drain(struct request_queue *q) | |
1215 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1216 | { | |
1217 | struct throtl_data *td = q->td; | |
1218 | struct throtl_rb_root *st = &td->tg_service_tree; | |
1219 | struct throtl_grp *tg; | |
1220 | struct bio_list bl; | |
1221 | struct bio *bio; | |
1222 | ||
8bcb6c7d | 1223 | queue_lockdep_assert_held(q); |
c9a929dd TH |
1224 | |
1225 | bio_list_init(&bl); | |
1226 | ||
1227 | while ((tg = throtl_rb_first(st))) { | |
1228 | throtl_dequeue_tg(td, tg); | |
1229 | ||
1230 | while ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
1231 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1232 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
1233 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1234 | } | |
1235 | spin_unlock_irq(q->queue_lock); | |
1236 | ||
1237 | while ((bio = bio_list_pop(&bl))) | |
1238 | generic_make_request(bio); | |
1239 | ||
1240 | spin_lock_irq(q->queue_lock); | |
1241 | } | |
1242 | ||
e43473b7 VG |
1243 | int blk_throtl_init(struct request_queue *q) |
1244 | { | |
1245 | struct throtl_data *td; | |
a2b1693b | 1246 | int ret; |
e43473b7 VG |
1247 | |
1248 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1249 | if (!td) | |
1250 | return -ENOMEM; | |
1251 | ||
e43473b7 | 1252 | td->tg_service_tree = THROTL_RB_ROOT; |
de701c74 | 1253 | td->limits_changed = false; |
a29a171e | 1254 | INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work); |
e43473b7 | 1255 | |
cd1604fa | 1256 | q->td = td; |
29b12589 | 1257 | td->queue = q; |
02977e4a | 1258 | |
a2b1693b | 1259 | /* activate policy */ |
3c798398 | 1260 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1261 | if (ret) |
f51b802c | 1262 | kfree(td); |
a2b1693b | 1263 | return ret; |
e43473b7 VG |
1264 | } |
1265 | ||
1266 | void blk_throtl_exit(struct request_queue *q) | |
1267 | { | |
c875f4d0 | 1268 | BUG_ON(!q->td); |
da527770 | 1269 | throtl_shutdown_wq(q); |
3c798398 | 1270 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1271 | kfree(q->td); |
e43473b7 VG |
1272 | } |
1273 | ||
1274 | static int __init throtl_init(void) | |
1275 | { | |
450adcbe VG |
1276 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1277 | if (!kthrotld_workqueue) | |
1278 | panic("Failed to create kthrotld\n"); | |
1279 | ||
3c798398 | 1280 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1281 | } |
1282 | ||
1283 | module_init(throtl_init); |