]>
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 { |
e43473b7 VG |
52 | /* active throtl group service_tree member */ |
53 | struct rb_node rb_node; | |
54 | ||
55 | /* | |
56 | * Dispatch time in jiffies. This is the estimated time when group | |
57 | * will unthrottle and is ready to dispatch more bio. It is used as | |
58 | * key to sort active groups in service tree. | |
59 | */ | |
60 | unsigned long disptime; | |
61 | ||
e43473b7 VG |
62 | unsigned int flags; |
63 | ||
64 | /* Two lists for READ and WRITE */ | |
65 | struct bio_list bio_lists[2]; | |
66 | ||
67 | /* Number of queued bios on READ and WRITE lists */ | |
68 | unsigned int nr_queued[2]; | |
69 | ||
70 | /* bytes per second rate limits */ | |
71 | uint64_t bps[2]; | |
72 | ||
8e89d13f VG |
73 | /* IOPS limits */ |
74 | unsigned int iops[2]; | |
75 | ||
e43473b7 VG |
76 | /* Number of bytes disptached in current slice */ |
77 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
78 | /* Number of bio's dispatched in current slice */ |
79 | unsigned int io_disp[2]; | |
e43473b7 VG |
80 | |
81 | /* When did we start a new slice */ | |
82 | unsigned long slice_start[2]; | |
83 | unsigned long slice_end[2]; | |
fe071437 VG |
84 | |
85 | /* Some throttle limits got updated for the group */ | |
6f037937 | 86 | int limits_changed; |
8a3d2615 TH |
87 | |
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; | |
fe071437 | 112 | |
6f037937 | 113 | int limits_changed; |
e43473b7 VG |
114 | }; |
115 | ||
8a3d2615 TH |
116 | /* list and work item to allocate percpu group stats */ |
117 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
118 | static LIST_HEAD(tg_stats_alloc_list); | |
119 | ||
120 | static void tg_stats_alloc_fn(struct work_struct *); | |
121 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
122 | ||
3c798398 | 123 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 124 | { |
3c798398 | 125 | return blkg_to_pdata(blkg, &blkcg_policy_throtl); |
0381411e TH |
126 | } |
127 | ||
3c798398 | 128 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 129 | { |
aaec55a0 | 130 | return pdata_to_blkg(tg); |
0381411e TH |
131 | } |
132 | ||
03d8e111 TH |
133 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) |
134 | { | |
135 | return blkg_to_tg(td->queue->root_blkg); | |
136 | } | |
137 | ||
e43473b7 VG |
138 | enum tg_state_flags { |
139 | THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */ | |
140 | }; | |
141 | ||
142 | #define THROTL_TG_FNS(name) \ | |
143 | static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \ | |
144 | { \ | |
145 | (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \ | |
146 | } \ | |
147 | static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \ | |
148 | { \ | |
149 | (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \ | |
150 | } \ | |
151 | static inline int throtl_tg_##name(const struct throtl_grp *tg) \ | |
152 | { \ | |
153 | return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \ | |
154 | } | |
155 | ||
156 | THROTL_TG_FNS(on_rr); | |
157 | ||
54e7ed12 TH |
158 | #define throtl_log_tg(td, tg, fmt, args...) do { \ |
159 | char __pbuf[128]; \ | |
160 | \ | |
161 | blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \ | |
162 | blk_add_trace_msg((td)->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
163 | } while (0) | |
e43473b7 VG |
164 | |
165 | #define throtl_log(td, fmt, args...) \ | |
166 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) | |
167 | ||
d2f31a5f | 168 | static inline unsigned int total_nr_queued(struct throtl_data *td) |
e43473b7 | 169 | { |
d2f31a5f | 170 | return td->nr_queued[0] + td->nr_queued[1]; |
e43473b7 VG |
171 | } |
172 | ||
8a3d2615 TH |
173 | /* |
174 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
175 | * system_nrt_wq once there are some groups on the alloc_list waiting for | |
176 | * allocation. | |
177 | */ | |
178 | static void tg_stats_alloc_fn(struct work_struct *work) | |
179 | { | |
180 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
181 | struct delayed_work *dwork = to_delayed_work(work); | |
182 | bool empty = false; | |
183 | ||
184 | alloc_stats: | |
185 | if (!stats_cpu) { | |
186 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
187 | if (!stats_cpu) { | |
188 | /* allocation failed, try again after some time */ | |
189 | queue_delayed_work(system_nrt_wq, dwork, | |
190 | msecs_to_jiffies(10)); | |
191 | return; | |
192 | } | |
193 | } | |
194 | ||
195 | spin_lock_irq(&tg_stats_alloc_lock); | |
196 | ||
197 | if (!list_empty(&tg_stats_alloc_list)) { | |
198 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
199 | struct throtl_grp, | |
200 | stats_alloc_node); | |
201 | swap(tg->stats_cpu, stats_cpu); | |
202 | list_del_init(&tg->stats_alloc_node); | |
203 | } | |
204 | ||
205 | empty = list_empty(&tg_stats_alloc_list); | |
206 | spin_unlock_irq(&tg_stats_alloc_lock); | |
207 | if (!empty) | |
208 | goto alloc_stats; | |
209 | } | |
210 | ||
3c798398 | 211 | static void throtl_pd_init(struct blkcg_gq *blkg) |
a29a171e | 212 | { |
0381411e | 213 | struct throtl_grp *tg = blkg_to_tg(blkg); |
cd1604fa | 214 | |
a29a171e VG |
215 | RB_CLEAR_NODE(&tg->rb_node); |
216 | bio_list_init(&tg->bio_lists[0]); | |
217 | bio_list_init(&tg->bio_lists[1]); | |
218 | tg->limits_changed = false; | |
219 | ||
e56da7e2 TH |
220 | tg->bps[READ] = -1; |
221 | tg->bps[WRITE] = -1; | |
222 | tg->iops[READ] = -1; | |
223 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
224 | |
225 | /* | |
226 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
227 | * but percpu allocator can't be called from IO path. Queue tg on | |
228 | * tg_stats_alloc_list and allocate from work item. | |
229 | */ | |
230 | spin_lock(&tg_stats_alloc_lock); | |
231 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); | |
232 | queue_delayed_work(system_nrt_wq, &tg_stats_alloc_work, 0); | |
233 | spin_unlock(&tg_stats_alloc_lock); | |
234 | } | |
235 | ||
3c798398 | 236 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
8a3d2615 TH |
237 | { |
238 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
239 | ||
240 | spin_lock(&tg_stats_alloc_lock); | |
241 | list_del_init(&tg->stats_alloc_node); | |
242 | spin_unlock(&tg_stats_alloc_lock); | |
243 | ||
244 | free_percpu(tg->stats_cpu); | |
245 | } | |
246 | ||
3c798398 | 247 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
8a3d2615 TH |
248 | { |
249 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
250 | int cpu; | |
251 | ||
252 | if (tg->stats_cpu == NULL) | |
253 | return; | |
254 | ||
255 | for_each_possible_cpu(cpu) { | |
256 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
257 | ||
258 | blkg_rwstat_reset(&sc->service_bytes); | |
259 | blkg_rwstat_reset(&sc->serviced); | |
260 | } | |
a29a171e VG |
261 | } |
262 | ||
3c798398 TH |
263 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
264 | struct blkcg *blkcg) | |
e43473b7 | 265 | { |
be2c6b19 | 266 | /* |
3c798398 TH |
267 | * This is the common case when there are no blkcgs. Avoid lookup |
268 | * in this case | |
cd1604fa | 269 | */ |
3c798398 | 270 | if (blkcg == &blkcg_root) |
03d8e111 | 271 | return td_root_tg(td); |
e43473b7 | 272 | |
e8989fae | 273 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
274 | } |
275 | ||
cd1604fa | 276 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
3c798398 | 277 | struct blkcg *blkcg) |
e43473b7 | 278 | { |
f469a7b4 | 279 | struct request_queue *q = td->queue; |
cd1604fa | 280 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 281 | |
f469a7b4 | 282 | /* |
3c798398 TH |
283 | * This is the common case when there are no blkcgs. Avoid lookup |
284 | * in this case | |
f469a7b4 | 285 | */ |
3c798398 | 286 | if (blkcg == &blkcg_root) { |
03d8e111 | 287 | tg = td_root_tg(td); |
cd1604fa | 288 | } else { |
3c798398 | 289 | struct blkcg_gq *blkg; |
f469a7b4 | 290 | |
3c96cb32 | 291 | blkg = blkg_lookup_create(blkcg, q); |
f469a7b4 | 292 | |
cd1604fa TH |
293 | /* if %NULL and @q is alive, fall back to root_tg */ |
294 | if (!IS_ERR(blkg)) | |
0381411e | 295 | tg = blkg_to_tg(blkg); |
cd1604fa | 296 | else if (!blk_queue_dead(q)) |
03d8e111 | 297 | tg = td_root_tg(td); |
f469a7b4 VG |
298 | } |
299 | ||
e43473b7 VG |
300 | return tg; |
301 | } | |
302 | ||
303 | static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root) | |
304 | { | |
305 | /* Service tree is empty */ | |
306 | if (!root->count) | |
307 | return NULL; | |
308 | ||
309 | if (!root->left) | |
310 | root->left = rb_first(&root->rb); | |
311 | ||
312 | if (root->left) | |
313 | return rb_entry_tg(root->left); | |
314 | ||
315 | return NULL; | |
316 | } | |
317 | ||
318 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
319 | { | |
320 | rb_erase(n, root); | |
321 | RB_CLEAR_NODE(n); | |
322 | } | |
323 | ||
324 | static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root) | |
325 | { | |
326 | if (root->left == n) | |
327 | root->left = NULL; | |
328 | rb_erase_init(n, &root->rb); | |
329 | --root->count; | |
330 | } | |
331 | ||
332 | static void update_min_dispatch_time(struct throtl_rb_root *st) | |
333 | { | |
334 | struct throtl_grp *tg; | |
335 | ||
336 | tg = throtl_rb_first(st); | |
337 | if (!tg) | |
338 | return; | |
339 | ||
340 | st->min_disptime = tg->disptime; | |
341 | } | |
342 | ||
343 | static void | |
344 | tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg) | |
345 | { | |
346 | struct rb_node **node = &st->rb.rb_node; | |
347 | struct rb_node *parent = NULL; | |
348 | struct throtl_grp *__tg; | |
349 | unsigned long key = tg->disptime; | |
350 | int left = 1; | |
351 | ||
352 | while (*node != NULL) { | |
353 | parent = *node; | |
354 | __tg = rb_entry_tg(parent); | |
355 | ||
356 | if (time_before(key, __tg->disptime)) | |
357 | node = &parent->rb_left; | |
358 | else { | |
359 | node = &parent->rb_right; | |
360 | left = 0; | |
361 | } | |
362 | } | |
363 | ||
364 | if (left) | |
365 | st->left = &tg->rb_node; | |
366 | ||
367 | rb_link_node(&tg->rb_node, parent, node); | |
368 | rb_insert_color(&tg->rb_node, &st->rb); | |
369 | } | |
370 | ||
371 | static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
372 | { | |
373 | struct throtl_rb_root *st = &td->tg_service_tree; | |
374 | ||
375 | tg_service_tree_add(st, tg); | |
376 | throtl_mark_tg_on_rr(tg); | |
377 | st->count++; | |
378 | } | |
379 | ||
380 | static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
381 | { | |
382 | if (!throtl_tg_on_rr(tg)) | |
383 | __throtl_enqueue_tg(td, tg); | |
384 | } | |
385 | ||
386 | static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
387 | { | |
388 | throtl_rb_erase(&tg->rb_node, &td->tg_service_tree); | |
389 | throtl_clear_tg_on_rr(tg); | |
390 | } | |
391 | ||
392 | static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
393 | { | |
394 | if (throtl_tg_on_rr(tg)) | |
395 | __throtl_dequeue_tg(td, tg); | |
396 | } | |
397 | ||
398 | static void throtl_schedule_next_dispatch(struct throtl_data *td) | |
399 | { | |
400 | struct throtl_rb_root *st = &td->tg_service_tree; | |
401 | ||
402 | /* | |
403 | * If there are more bios pending, schedule more work. | |
404 | */ | |
405 | if (!total_nr_queued(td)) | |
406 | return; | |
407 | ||
408 | BUG_ON(!st->count); | |
409 | ||
410 | update_min_dispatch_time(st); | |
411 | ||
412 | if (time_before_eq(st->min_disptime, jiffies)) | |
450adcbe | 413 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 414 | else |
450adcbe | 415 | throtl_schedule_delayed_work(td, (st->min_disptime - jiffies)); |
e43473b7 VG |
416 | } |
417 | ||
418 | static inline void | |
419 | throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
420 | { | |
421 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 422 | tg->io_disp[rw] = 0; |
e43473b7 VG |
423 | tg->slice_start[rw] = jiffies; |
424 | tg->slice_end[rw] = jiffies + throtl_slice; | |
425 | throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
426 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
427 | tg->slice_end[rw], jiffies); | |
428 | } | |
429 | ||
d1ae8ffd VG |
430 | static inline void throtl_set_slice_end(struct throtl_data *td, |
431 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
432 | { | |
433 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
434 | } | |
435 | ||
e43473b7 VG |
436 | static inline void throtl_extend_slice(struct throtl_data *td, |
437 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
438 | { | |
439 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
440 | throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
441 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
442 | tg->slice_end[rw], jiffies); | |
443 | } | |
444 | ||
445 | /* Determine if previously allocated or extended slice is complete or not */ | |
446 | static bool | |
447 | throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
448 | { | |
449 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
450 | return 0; | |
451 | ||
452 | return 1; | |
453 | } | |
454 | ||
455 | /* Trim the used slices and adjust slice start accordingly */ | |
456 | static inline void | |
457 | throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
458 | { | |
3aad5d3e VG |
459 | unsigned long nr_slices, time_elapsed, io_trim; |
460 | u64 bytes_trim, tmp; | |
e43473b7 VG |
461 | |
462 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
463 | ||
464 | /* | |
465 | * If bps are unlimited (-1), then time slice don't get | |
466 | * renewed. Don't try to trim the slice if slice is used. A new | |
467 | * slice will start when appropriate. | |
468 | */ | |
469 | if (throtl_slice_used(td, tg, rw)) | |
470 | return; | |
471 | ||
d1ae8ffd VG |
472 | /* |
473 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
474 | * that initially cgroup limit was very low resulting in high | |
475 | * slice_end, but later limit was bumped up and bio was dispached | |
476 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
477 | * is bad because it does not allow new slice to start. | |
478 | */ | |
479 | ||
480 | throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice); | |
481 | ||
e43473b7 VG |
482 | time_elapsed = jiffies - tg->slice_start[rw]; |
483 | ||
484 | nr_slices = time_elapsed / throtl_slice; | |
485 | ||
486 | if (!nr_slices) | |
487 | return; | |
3aad5d3e VG |
488 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
489 | do_div(tmp, HZ); | |
490 | bytes_trim = tmp; | |
e43473b7 | 491 | |
8e89d13f | 492 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 493 | |
8e89d13f | 494 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
495 | return; |
496 | ||
497 | if (tg->bytes_disp[rw] >= bytes_trim) | |
498 | tg->bytes_disp[rw] -= bytes_trim; | |
499 | else | |
500 | tg->bytes_disp[rw] = 0; | |
501 | ||
8e89d13f VG |
502 | if (tg->io_disp[rw] >= io_trim) |
503 | tg->io_disp[rw] -= io_trim; | |
504 | else | |
505 | tg->io_disp[rw] = 0; | |
506 | ||
e43473b7 VG |
507 | tg->slice_start[rw] += nr_slices * throtl_slice; |
508 | ||
3aad5d3e | 509 | throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 510 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 511 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
512 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
513 | } | |
514 | ||
8e89d13f VG |
515 | static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg, |
516 | struct bio *bio, unsigned long *wait) | |
e43473b7 VG |
517 | { |
518 | bool rw = bio_data_dir(bio); | |
8e89d13f | 519 | unsigned int io_allowed; |
e43473b7 | 520 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 521 | u64 tmp; |
e43473b7 | 522 | |
8e89d13f | 523 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 524 | |
8e89d13f VG |
525 | /* Slice has just started. Consider one slice interval */ |
526 | if (!jiffy_elapsed) | |
527 | jiffy_elapsed_rnd = throtl_slice; | |
528 | ||
529 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
530 | ||
c49c06e4 VG |
531 | /* |
532 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
533 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
534 | * will allow dispatch after 1 second and after that slice should | |
535 | * have been trimmed. | |
536 | */ | |
537 | ||
538 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
539 | do_div(tmp, HZ); | |
540 | ||
541 | if (tmp > UINT_MAX) | |
542 | io_allowed = UINT_MAX; | |
543 | else | |
544 | io_allowed = tmp; | |
8e89d13f VG |
545 | |
546 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
547 | if (wait) |
548 | *wait = 0; | |
549 | return 1; | |
550 | } | |
551 | ||
8e89d13f VG |
552 | /* Calc approx time to dispatch */ |
553 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
554 | ||
555 | if (jiffy_wait > jiffy_elapsed) | |
556 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
557 | else | |
558 | jiffy_wait = 1; | |
559 | ||
560 | if (wait) | |
561 | *wait = jiffy_wait; | |
562 | return 0; | |
563 | } | |
564 | ||
565 | static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg, | |
566 | struct bio *bio, unsigned long *wait) | |
567 | { | |
568 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 569 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 570 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
571 | |
572 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
573 | ||
574 | /* Slice has just started. Consider one slice interval */ | |
575 | if (!jiffy_elapsed) | |
576 | jiffy_elapsed_rnd = throtl_slice; | |
577 | ||
578 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
579 | ||
5e901a2b VG |
580 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
581 | do_div(tmp, HZ); | |
3aad5d3e | 582 | bytes_allowed = tmp; |
e43473b7 VG |
583 | |
584 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
585 | if (wait) | |
586 | *wait = 0; | |
587 | return 1; | |
588 | } | |
589 | ||
590 | /* Calc approx time to dispatch */ | |
591 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
592 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
593 | ||
594 | if (!jiffy_wait) | |
595 | jiffy_wait = 1; | |
596 | ||
597 | /* | |
598 | * This wait time is without taking into consideration the rounding | |
599 | * up we did. Add that time also. | |
600 | */ | |
601 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
602 | if (wait) |
603 | *wait = jiffy_wait; | |
8e89d13f VG |
604 | return 0; |
605 | } | |
606 | ||
af75cd3c VG |
607 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
608 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
609 | return 1; | |
610 | return 0; | |
611 | } | |
612 | ||
8e89d13f VG |
613 | /* |
614 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
615 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
616 | */ | |
617 | static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg, | |
618 | struct bio *bio, unsigned long *wait) | |
619 | { | |
620 | bool rw = bio_data_dir(bio); | |
621 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
622 | ||
623 | /* | |
624 | * Currently whole state machine of group depends on first bio | |
625 | * queued in the group bio list. So one should not be calling | |
626 | * this function with a different bio if there are other bios | |
627 | * queued. | |
628 | */ | |
629 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); | |
e43473b7 | 630 | |
8e89d13f VG |
631 | /* If tg->bps = -1, then BW is unlimited */ |
632 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
633 | if (wait) | |
634 | *wait = 0; | |
635 | return 1; | |
636 | } | |
637 | ||
638 | /* | |
639 | * If previous slice expired, start a new one otherwise renew/extend | |
640 | * existing slice to make sure it is at least throtl_slice interval | |
641 | * long since now. | |
642 | */ | |
643 | if (throtl_slice_used(td, tg, rw)) | |
644 | throtl_start_new_slice(td, tg, rw); | |
645 | else { | |
646 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
647 | throtl_extend_slice(td, tg, rw, jiffies + throtl_slice); | |
648 | } | |
649 | ||
650 | if (tg_with_in_bps_limit(td, tg, bio, &bps_wait) | |
651 | && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) { | |
652 | if (wait) | |
653 | *wait = 0; | |
654 | return 1; | |
655 | } | |
656 | ||
657 | max_wait = max(bps_wait, iops_wait); | |
658 | ||
659 | if (wait) | |
660 | *wait = max_wait; | |
661 | ||
662 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
663 | throtl_extend_slice(td, tg, rw, jiffies + max_wait); | |
e43473b7 VG |
664 | |
665 | return 0; | |
666 | } | |
667 | ||
3c798398 | 668 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
669 | int rw) |
670 | { | |
8a3d2615 TH |
671 | struct throtl_grp *tg = blkg_to_tg(blkg); |
672 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
673 | unsigned long flags; |
674 | ||
675 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 676 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
677 | return; |
678 | ||
679 | /* | |
680 | * Disabling interrupts to provide mutual exclusion between two | |
681 | * writes on same cpu. It probably is not needed for 64bit. Not | |
682 | * optimizing that case yet. | |
683 | */ | |
684 | local_irq_save(flags); | |
685 | ||
8a3d2615 | 686 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 687 | |
629ed0b1 TH |
688 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
689 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
690 | ||
691 | local_irq_restore(flags); | |
692 | } | |
693 | ||
e43473b7 VG |
694 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
695 | { | |
696 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
697 | |
698 | /* Charge the bio to the group */ | |
699 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 700 | tg->io_disp[rw]++; |
e43473b7 | 701 | |
629ed0b1 | 702 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw); |
e43473b7 VG |
703 | } |
704 | ||
705 | static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg, | |
706 | struct bio *bio) | |
707 | { | |
708 | bool rw = bio_data_dir(bio); | |
709 | ||
710 | bio_list_add(&tg->bio_lists[rw], bio); | |
711 | /* Take a bio reference on tg */ | |
1adaf3dd | 712 | blkg_get(tg_to_blkg(tg)); |
e43473b7 VG |
713 | tg->nr_queued[rw]++; |
714 | td->nr_queued[rw]++; | |
715 | throtl_enqueue_tg(td, tg); | |
716 | } | |
717 | ||
718 | static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg) | |
719 | { | |
720 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
721 | struct bio *bio; | |
722 | ||
723 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
724 | tg_may_dispatch(td, tg, bio, &read_wait); | |
725 | ||
726 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
727 | tg_may_dispatch(td, tg, bio, &write_wait); | |
728 | ||
729 | min_wait = min(read_wait, write_wait); | |
730 | disptime = jiffies + min_wait; | |
731 | ||
e43473b7 VG |
732 | /* Update dispatch time */ |
733 | throtl_dequeue_tg(td, tg); | |
734 | tg->disptime = disptime; | |
735 | throtl_enqueue_tg(td, tg); | |
736 | } | |
737 | ||
738 | static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg, | |
739 | bool rw, struct bio_list *bl) | |
740 | { | |
741 | struct bio *bio; | |
742 | ||
743 | bio = bio_list_pop(&tg->bio_lists[rw]); | |
744 | tg->nr_queued[rw]--; | |
1adaf3dd TH |
745 | /* Drop bio reference on blkg */ |
746 | blkg_put(tg_to_blkg(tg)); | |
e43473b7 VG |
747 | |
748 | BUG_ON(td->nr_queued[rw] <= 0); | |
749 | td->nr_queued[rw]--; | |
750 | ||
751 | throtl_charge_bio(tg, bio); | |
752 | bio_list_add(bl, bio); | |
753 | bio->bi_rw |= REQ_THROTTLED; | |
754 | ||
755 | throtl_trim_slice(td, tg, rw); | |
756 | } | |
757 | ||
758 | static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg, | |
759 | struct bio_list *bl) | |
760 | { | |
761 | unsigned int nr_reads = 0, nr_writes = 0; | |
762 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 763 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
764 | struct bio *bio; |
765 | ||
766 | /* Try to dispatch 75% READS and 25% WRITES */ | |
767 | ||
768 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) | |
769 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
770 | ||
771 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
772 | nr_reads++; | |
773 | ||
774 | if (nr_reads >= max_nr_reads) | |
775 | break; | |
776 | } | |
777 | ||
778 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) | |
779 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
780 | ||
781 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
782 | nr_writes++; | |
783 | ||
784 | if (nr_writes >= max_nr_writes) | |
785 | break; | |
786 | } | |
787 | ||
788 | return nr_reads + nr_writes; | |
789 | } | |
790 | ||
791 | static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl) | |
792 | { | |
793 | unsigned int nr_disp = 0; | |
794 | struct throtl_grp *tg; | |
795 | struct throtl_rb_root *st = &td->tg_service_tree; | |
796 | ||
797 | while (1) { | |
798 | tg = throtl_rb_first(st); | |
799 | ||
800 | if (!tg) | |
801 | break; | |
802 | ||
803 | if (time_before(jiffies, tg->disptime)) | |
804 | break; | |
805 | ||
806 | throtl_dequeue_tg(td, tg); | |
807 | ||
808 | nr_disp += throtl_dispatch_tg(td, tg, bl); | |
809 | ||
810 | if (tg->nr_queued[0] || tg->nr_queued[1]) { | |
811 | tg_update_disptime(td, tg); | |
812 | throtl_enqueue_tg(td, tg); | |
813 | } | |
814 | ||
815 | if (nr_disp >= throtl_quantum) | |
816 | break; | |
817 | } | |
818 | ||
819 | return nr_disp; | |
820 | } | |
821 | ||
fe071437 VG |
822 | static void throtl_process_limit_change(struct throtl_data *td) |
823 | { | |
4eef3049 | 824 | struct request_queue *q = td->queue; |
3c798398 | 825 | struct blkcg_gq *blkg, *n; |
fe071437 | 826 | |
de701c74 | 827 | if (!td->limits_changed) |
fe071437 VG |
828 | return; |
829 | ||
de701c74 | 830 | xchg(&td->limits_changed, false); |
fe071437 | 831 | |
de701c74 | 832 | throtl_log(td, "limits changed"); |
fe071437 | 833 | |
e8989fae | 834 | list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) { |
4eef3049 TH |
835 | struct throtl_grp *tg = blkg_to_tg(blkg); |
836 | ||
de701c74 VG |
837 | if (!tg->limits_changed) |
838 | continue; | |
839 | ||
840 | if (!xchg(&tg->limits_changed, false)) | |
841 | continue; | |
842 | ||
843 | throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu" | |
844 | " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE], | |
845 | tg->iops[READ], tg->iops[WRITE]); | |
846 | ||
04521db0 VG |
847 | /* |
848 | * Restart the slices for both READ and WRITES. It | |
849 | * might happen that a group's limit are dropped | |
850 | * suddenly and we don't want to account recently | |
851 | * dispatched IO with new low rate | |
852 | */ | |
853 | throtl_start_new_slice(td, tg, 0); | |
854 | throtl_start_new_slice(td, tg, 1); | |
855 | ||
de701c74 | 856 | if (throtl_tg_on_rr(tg)) |
fe071437 | 857 | tg_update_disptime(td, tg); |
fe071437 | 858 | } |
fe071437 VG |
859 | } |
860 | ||
e43473b7 VG |
861 | /* Dispatch throttled bios. Should be called without queue lock held. */ |
862 | static int throtl_dispatch(struct request_queue *q) | |
863 | { | |
864 | struct throtl_data *td = q->td; | |
865 | unsigned int nr_disp = 0; | |
866 | struct bio_list bio_list_on_stack; | |
867 | struct bio *bio; | |
69d60eb9 | 868 | struct blk_plug plug; |
e43473b7 VG |
869 | |
870 | spin_lock_irq(q->queue_lock); | |
871 | ||
fe071437 VG |
872 | throtl_process_limit_change(td); |
873 | ||
e43473b7 VG |
874 | if (!total_nr_queued(td)) |
875 | goto out; | |
876 | ||
877 | bio_list_init(&bio_list_on_stack); | |
878 | ||
d2f31a5f | 879 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
e43473b7 VG |
880 | total_nr_queued(td), td->nr_queued[READ], |
881 | td->nr_queued[WRITE]); | |
882 | ||
883 | nr_disp = throtl_select_dispatch(td, &bio_list_on_stack); | |
884 | ||
885 | if (nr_disp) | |
886 | throtl_log(td, "bios disp=%u", nr_disp); | |
887 | ||
888 | throtl_schedule_next_dispatch(td); | |
889 | out: | |
890 | spin_unlock_irq(q->queue_lock); | |
891 | ||
892 | /* | |
893 | * If we dispatched some requests, unplug the queue to make sure | |
894 | * immediate dispatch | |
895 | */ | |
896 | if (nr_disp) { | |
69d60eb9 | 897 | blk_start_plug(&plug); |
e43473b7 VG |
898 | while((bio = bio_list_pop(&bio_list_on_stack))) |
899 | generic_make_request(bio); | |
69d60eb9 | 900 | blk_finish_plug(&plug); |
e43473b7 VG |
901 | } |
902 | return nr_disp; | |
903 | } | |
904 | ||
905 | void blk_throtl_work(struct work_struct *work) | |
906 | { | |
907 | struct throtl_data *td = container_of(work, struct throtl_data, | |
908 | throtl_work.work); | |
909 | struct request_queue *q = td->queue; | |
910 | ||
911 | throtl_dispatch(q); | |
912 | } | |
913 | ||
914 | /* Call with queue lock held */ | |
450adcbe VG |
915 | static void |
916 | throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay) | |
e43473b7 VG |
917 | { |
918 | ||
e43473b7 VG |
919 | struct delayed_work *dwork = &td->throtl_work; |
920 | ||
04521db0 | 921 | /* schedule work if limits changed even if no bio is queued */ |
d2f31a5f | 922 | if (total_nr_queued(td) || td->limits_changed) { |
e43473b7 VG |
923 | /* |
924 | * We might have a work scheduled to be executed in future. | |
925 | * Cancel that and schedule a new one. | |
926 | */ | |
927 | __cancel_delayed_work(dwork); | |
450adcbe | 928 | queue_delayed_work(kthrotld_workqueue, dwork, delay); |
e43473b7 VG |
929 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", |
930 | delay, jiffies); | |
931 | } | |
932 | } | |
e43473b7 | 933 | |
d366e7ec | 934 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, void *pdata, int off) |
41b38b6d | 935 | { |
d366e7ec | 936 | struct throtl_grp *tg = pdata; |
41b38b6d TH |
937 | struct blkg_rwstat rwstat = { }, tmp; |
938 | int i, cpu; | |
939 | ||
940 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 941 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
942 | |
943 | tmp = blkg_rwstat_read((void *)sc + off); | |
944 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
945 | rwstat.cnt[i] += tmp.cnt[i]; | |
946 | } | |
947 | ||
d366e7ec | 948 | return __blkg_prfill_rwstat(sf, pdata, &rwstat); |
41b38b6d TH |
949 | } |
950 | ||
8a3d2615 TH |
951 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
952 | struct seq_file *sf) | |
41b38b6d | 953 | { |
3c798398 | 954 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
41b38b6d | 955 | |
3c798398 | 956 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 957 | cft->private, true); |
41b38b6d TH |
958 | return 0; |
959 | } | |
960 | ||
d366e7ec | 961 | static u64 tg_prfill_conf_u64(struct seq_file *sf, void *pdata, int off) |
60c2bc2d | 962 | { |
d366e7ec | 963 | u64 v = *(u64 *)(pdata + off); |
60c2bc2d | 964 | |
af133ceb | 965 | if (v == -1) |
60c2bc2d | 966 | return 0; |
d366e7ec | 967 | return __blkg_prfill_u64(sf, pdata, v); |
60c2bc2d TH |
968 | } |
969 | ||
d366e7ec | 970 | static u64 tg_prfill_conf_uint(struct seq_file *sf, void *pdata, int off) |
e43473b7 | 971 | { |
d366e7ec | 972 | unsigned int v = *(unsigned int *)(pdata + off); |
fe071437 | 973 | |
af133ceb TH |
974 | if (v == -1) |
975 | return 0; | |
d366e7ec | 976 | return __blkg_prfill_u64(sf, pdata, v); |
e43473b7 VG |
977 | } |
978 | ||
af133ceb TH |
979 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
980 | struct seq_file *sf) | |
8e89d13f | 981 | { |
3c798398 TH |
982 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, |
983 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 984 | return 0; |
8e89d13f VG |
985 | } |
986 | ||
af133ceb TH |
987 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
988 | struct seq_file *sf) | |
8e89d13f | 989 | { |
3c798398 TH |
990 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, |
991 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 992 | return 0; |
60c2bc2d TH |
993 | } |
994 | ||
af133ceb TH |
995 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
996 | bool is_u64) | |
60c2bc2d | 997 | { |
3c798398 | 998 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
60c2bc2d | 999 | struct blkg_conf_ctx ctx; |
af133ceb | 1000 | struct throtl_grp *tg; |
a2b1693b | 1001 | struct throtl_data *td; |
60c2bc2d TH |
1002 | int ret; |
1003 | ||
3c798398 | 1004 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
1005 | if (ret) |
1006 | return ret; | |
1007 | ||
af133ceb | 1008 | tg = blkg_to_tg(ctx.blkg); |
a2b1693b | 1009 | td = ctx.blkg->q->td; |
af133ceb | 1010 | |
a2b1693b TH |
1011 | if (!ctx.v) |
1012 | ctx.v = -1; | |
af133ceb | 1013 | |
a2b1693b TH |
1014 | if (is_u64) |
1015 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
1016 | else | |
1017 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 1018 | |
a2b1693b TH |
1019 | /* XXX: we don't need the following deferred processing */ |
1020 | xchg(&tg->limits_changed, true); | |
1021 | xchg(&td->limits_changed, true); | |
1022 | throtl_schedule_delayed_work(td, 0); | |
60c2bc2d TH |
1023 | |
1024 | blkg_conf_finish(&ctx); | |
a2b1693b | 1025 | return 0; |
8e89d13f VG |
1026 | } |
1027 | ||
af133ceb TH |
1028 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1029 | const char *buf) | |
60c2bc2d | 1030 | { |
af133ceb | 1031 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
1032 | } |
1033 | ||
af133ceb TH |
1034 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1035 | const char *buf) | |
60c2bc2d | 1036 | { |
af133ceb | 1037 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
1038 | } |
1039 | ||
1040 | static struct cftype throtl_files[] = { | |
1041 | { | |
1042 | .name = "throttle.read_bps_device", | |
af133ceb TH |
1043 | .private = offsetof(struct throtl_grp, bps[READ]), |
1044 | .read_seq_string = tg_print_conf_u64, | |
1045 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1046 | .max_write_len = 256, |
1047 | }, | |
1048 | { | |
1049 | .name = "throttle.write_bps_device", | |
af133ceb TH |
1050 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
1051 | .read_seq_string = tg_print_conf_u64, | |
1052 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1053 | .max_write_len = 256, |
1054 | }, | |
1055 | { | |
1056 | .name = "throttle.read_iops_device", | |
af133ceb TH |
1057 | .private = offsetof(struct throtl_grp, iops[READ]), |
1058 | .read_seq_string = tg_print_conf_uint, | |
1059 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1060 | .max_write_len = 256, |
1061 | }, | |
1062 | { | |
1063 | .name = "throttle.write_iops_device", | |
af133ceb TH |
1064 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
1065 | .read_seq_string = tg_print_conf_uint, | |
1066 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1067 | .max_write_len = 256, |
1068 | }, | |
1069 | { | |
1070 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1071 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1072 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1073 | }, |
1074 | { | |
1075 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1076 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1077 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1078 | }, |
1079 | { } /* terminate */ | |
1080 | }; | |
1081 | ||
da527770 | 1082 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1083 | { |
1084 | struct throtl_data *td = q->td; | |
1085 | ||
1086 | cancel_delayed_work_sync(&td->throtl_work); | |
1087 | } | |
1088 | ||
3c798398 | 1089 | static struct blkcg_policy blkcg_policy_throtl = { |
e43473b7 | 1090 | .ops = { |
3c798398 TH |
1091 | .pd_init_fn = throtl_pd_init, |
1092 | .pd_exit_fn = throtl_pd_exit, | |
1093 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 | 1094 | }, |
0381411e | 1095 | .pdata_size = sizeof(struct throtl_grp), |
60c2bc2d | 1096 | .cftypes = throtl_files, |
e43473b7 VG |
1097 | }; |
1098 | ||
bc16a4f9 | 1099 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1100 | { |
1101 | struct throtl_data *td = q->td; | |
1102 | struct throtl_grp *tg; | |
e43473b7 | 1103 | bool rw = bio_data_dir(bio), update_disptime = true; |
3c798398 | 1104 | struct blkcg *blkcg; |
bc16a4f9 | 1105 | bool throttled = false; |
e43473b7 VG |
1106 | |
1107 | if (bio->bi_rw & REQ_THROTTLED) { | |
1108 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1109 | goto out; |
e43473b7 VG |
1110 | } |
1111 | ||
671058fb TH |
1112 | /* bio_associate_current() needs ioc, try creating */ |
1113 | create_io_context(GFP_ATOMIC, q->node); | |
1114 | ||
af75cd3c VG |
1115 | /* |
1116 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1117 | * basic fields like stats and io rates. If a group has no rules, | |
1118 | * just update the dispatch stats in lockless manner and return. | |
1119 | */ | |
af75cd3c | 1120 | rcu_read_lock(); |
3c798398 | 1121 | blkcg = bio_blkcg(bio); |
cd1604fa | 1122 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1123 | if (tg) { |
af75cd3c | 1124 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1125 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1126 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1127 | goto out_unlock_rcu; |
af75cd3c VG |
1128 | } |
1129 | } | |
af75cd3c VG |
1130 | |
1131 | /* | |
1132 | * Either group has not been allocated yet or it is not an unlimited | |
1133 | * IO group | |
1134 | */ | |
e43473b7 | 1135 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1136 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1137 | if (unlikely(!tg)) |
1138 | goto out_unlock; | |
f469a7b4 | 1139 | |
e43473b7 VG |
1140 | if (tg->nr_queued[rw]) { |
1141 | /* | |
1142 | * There is already another bio queued in same dir. No | |
1143 | * need to update dispatch time. | |
1144 | */ | |
231d704b | 1145 | update_disptime = false; |
e43473b7 | 1146 | goto queue_bio; |
de701c74 | 1147 | |
e43473b7 VG |
1148 | } |
1149 | ||
1150 | /* Bio is with-in rate limit of group */ | |
1151 | if (tg_may_dispatch(td, tg, bio, NULL)) { | |
1152 | throtl_charge_bio(tg, bio); | |
04521db0 VG |
1153 | |
1154 | /* | |
1155 | * We need to trim slice even when bios are not being queued | |
1156 | * otherwise it might happen that a bio is not queued for | |
1157 | * a long time and slice keeps on extending and trim is not | |
1158 | * called for a long time. Now if limits are reduced suddenly | |
1159 | * we take into account all the IO dispatched so far at new | |
1160 | * low rate and * newly queued IO gets a really long dispatch | |
1161 | * time. | |
1162 | * | |
1163 | * So keep on trimming slice even if bio is not queued. | |
1164 | */ | |
1165 | throtl_trim_slice(td, tg, rw); | |
bc16a4f9 | 1166 | goto out_unlock; |
e43473b7 VG |
1167 | } |
1168 | ||
1169 | queue_bio: | |
fd16d263 | 1170 | throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
8e89d13f VG |
1171 | " iodisp=%u iops=%u queued=%d/%d", |
1172 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1173 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1174 | tg->io_disp[rw], tg->iops[rw], |
e43473b7 VG |
1175 | tg->nr_queued[READ], tg->nr_queued[WRITE]); |
1176 | ||
671058fb | 1177 | bio_associate_current(bio); |
e43473b7 | 1178 | throtl_add_bio_tg(q->td, tg, bio); |
bc16a4f9 | 1179 | throttled = true; |
e43473b7 VG |
1180 | |
1181 | if (update_disptime) { | |
1182 | tg_update_disptime(td, tg); | |
1183 | throtl_schedule_next_dispatch(td); | |
1184 | } | |
1185 | ||
bc16a4f9 | 1186 | out_unlock: |
e43473b7 | 1187 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1188 | out_unlock_rcu: |
1189 | rcu_read_unlock(); | |
bc16a4f9 TH |
1190 | out: |
1191 | return throttled; | |
e43473b7 VG |
1192 | } |
1193 | ||
c9a929dd TH |
1194 | /** |
1195 | * blk_throtl_drain - drain throttled bios | |
1196 | * @q: request_queue to drain throttled bios for | |
1197 | * | |
1198 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1199 | */ | |
1200 | void blk_throtl_drain(struct request_queue *q) | |
1201 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1202 | { | |
1203 | struct throtl_data *td = q->td; | |
1204 | struct throtl_rb_root *st = &td->tg_service_tree; | |
1205 | struct throtl_grp *tg; | |
1206 | struct bio_list bl; | |
1207 | struct bio *bio; | |
1208 | ||
334c2b0b | 1209 | WARN_ON_ONCE(!queue_is_locked(q)); |
c9a929dd TH |
1210 | |
1211 | bio_list_init(&bl); | |
1212 | ||
1213 | while ((tg = throtl_rb_first(st))) { | |
1214 | throtl_dequeue_tg(td, tg); | |
1215 | ||
1216 | while ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
1217 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1218 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
1219 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1220 | } | |
1221 | spin_unlock_irq(q->queue_lock); | |
1222 | ||
1223 | while ((bio = bio_list_pop(&bl))) | |
1224 | generic_make_request(bio); | |
1225 | ||
1226 | spin_lock_irq(q->queue_lock); | |
1227 | } | |
1228 | ||
e43473b7 VG |
1229 | int blk_throtl_init(struct request_queue *q) |
1230 | { | |
1231 | struct throtl_data *td; | |
a2b1693b | 1232 | int ret; |
e43473b7 VG |
1233 | |
1234 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1235 | if (!td) | |
1236 | return -ENOMEM; | |
1237 | ||
e43473b7 | 1238 | td->tg_service_tree = THROTL_RB_ROOT; |
de701c74 | 1239 | td->limits_changed = false; |
a29a171e | 1240 | INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work); |
e43473b7 | 1241 | |
cd1604fa | 1242 | q->td = td; |
29b12589 | 1243 | td->queue = q; |
02977e4a | 1244 | |
a2b1693b | 1245 | /* activate policy */ |
3c798398 | 1246 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1247 | if (ret) |
f51b802c | 1248 | kfree(td); |
a2b1693b | 1249 | return ret; |
e43473b7 VG |
1250 | } |
1251 | ||
1252 | void blk_throtl_exit(struct request_queue *q) | |
1253 | { | |
c875f4d0 | 1254 | BUG_ON(!q->td); |
da527770 | 1255 | throtl_shutdown_wq(q); |
3c798398 | 1256 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1257 | kfree(q->td); |
e43473b7 VG |
1258 | } |
1259 | ||
1260 | static int __init throtl_init(void) | |
1261 | { | |
450adcbe VG |
1262 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1263 | if (!kthrotld_workqueue) | |
1264 | panic("Failed to create kthrotld\n"); | |
1265 | ||
3c798398 | 1266 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1267 | } |
1268 | ||
1269 | module_init(throtl_init); |