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 | ||
450adcbe VG |
24 | /* A workqueue to queue throttle related work */ |
25 | static struct workqueue_struct *kthrotld_workqueue; | |
26 | static void throtl_schedule_delayed_work(struct throtl_data *td, | |
27 | unsigned long delay); | |
28 | ||
e43473b7 VG |
29 | struct throtl_rb_root { |
30 | struct rb_root rb; | |
31 | struct rb_node *left; | |
32 | unsigned int count; | |
33 | unsigned long min_disptime; | |
34 | }; | |
35 | ||
36 | #define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \ | |
37 | .count = 0, .min_disptime = 0} | |
38 | ||
39 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) | |
40 | ||
41 | struct throtl_grp { | |
42 | /* List of throtl groups on the request queue*/ | |
43 | struct hlist_node tg_node; | |
44 | ||
45 | /* active throtl group service_tree member */ | |
46 | struct rb_node rb_node; | |
47 | ||
48 | /* | |
49 | * Dispatch time in jiffies. This is the estimated time when group | |
50 | * will unthrottle and is ready to dispatch more bio. It is used as | |
51 | * key to sort active groups in service tree. | |
52 | */ | |
53 | unsigned long disptime; | |
54 | ||
55 | struct blkio_group blkg; | |
56 | atomic_t ref; | |
57 | unsigned int flags; | |
58 | ||
59 | /* Two lists for READ and WRITE */ | |
60 | struct bio_list bio_lists[2]; | |
61 | ||
62 | /* Number of queued bios on READ and WRITE lists */ | |
63 | unsigned int nr_queued[2]; | |
64 | ||
65 | /* bytes per second rate limits */ | |
66 | uint64_t bps[2]; | |
67 | ||
8e89d13f VG |
68 | /* IOPS limits */ |
69 | unsigned int iops[2]; | |
70 | ||
e43473b7 VG |
71 | /* Number of bytes disptached in current slice */ |
72 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
73 | /* Number of bio's dispatched in current slice */ |
74 | unsigned int io_disp[2]; | |
e43473b7 VG |
75 | |
76 | /* When did we start a new slice */ | |
77 | unsigned long slice_start[2]; | |
78 | unsigned long slice_end[2]; | |
fe071437 VG |
79 | |
80 | /* Some throttle limits got updated for the group */ | |
6f037937 | 81 | int limits_changed; |
4843c69d VG |
82 | |
83 | struct rcu_head rcu_head; | |
e43473b7 VG |
84 | }; |
85 | ||
86 | struct throtl_data | |
87 | { | |
88 | /* List of throtl groups */ | |
89 | struct hlist_head tg_list; | |
90 | ||
91 | /* service tree for active throtl groups */ | |
92 | struct throtl_rb_root tg_service_tree; | |
93 | ||
29b12589 | 94 | struct throtl_grp *root_tg; |
e43473b7 VG |
95 | struct request_queue *queue; |
96 | ||
97 | /* Total Number of queued bios on READ and WRITE lists */ | |
98 | unsigned int nr_queued[2]; | |
99 | ||
100 | /* | |
02977e4a | 101 | * number of total undestroyed groups |
e43473b7 VG |
102 | */ |
103 | unsigned int nr_undestroyed_grps; | |
104 | ||
105 | /* Work for dispatching throttled bios */ | |
106 | struct delayed_work throtl_work; | |
fe071437 | 107 | |
6f037937 | 108 | int limits_changed; |
e43473b7 VG |
109 | }; |
110 | ||
111 | enum tg_state_flags { | |
112 | THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */ | |
113 | }; | |
114 | ||
115 | #define THROTL_TG_FNS(name) \ | |
116 | static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \ | |
117 | { \ | |
118 | (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \ | |
119 | } \ | |
120 | static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \ | |
121 | { \ | |
122 | (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \ | |
123 | } \ | |
124 | static inline int throtl_tg_##name(const struct throtl_grp *tg) \ | |
125 | { \ | |
126 | return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \ | |
127 | } | |
128 | ||
129 | THROTL_TG_FNS(on_rr); | |
130 | ||
131 | #define throtl_log_tg(td, tg, fmt, args...) \ | |
132 | blk_add_trace_msg((td)->queue, "throtl %s " fmt, \ | |
133 | blkg_path(&(tg)->blkg), ##args); \ | |
134 | ||
135 | #define throtl_log(td, fmt, args...) \ | |
136 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) | |
137 | ||
138 | static inline struct throtl_grp *tg_of_blkg(struct blkio_group *blkg) | |
139 | { | |
140 | if (blkg) | |
141 | return container_of(blkg, struct throtl_grp, blkg); | |
142 | ||
143 | return NULL; | |
144 | } | |
145 | ||
d2f31a5f | 146 | static inline unsigned int total_nr_queued(struct throtl_data *td) |
e43473b7 | 147 | { |
d2f31a5f | 148 | return td->nr_queued[0] + td->nr_queued[1]; |
e43473b7 VG |
149 | } |
150 | ||
151 | static inline struct throtl_grp *throtl_ref_get_tg(struct throtl_grp *tg) | |
152 | { | |
153 | atomic_inc(&tg->ref); | |
154 | return tg; | |
155 | } | |
156 | ||
4843c69d VG |
157 | static void throtl_free_tg(struct rcu_head *head) |
158 | { | |
159 | struct throtl_grp *tg; | |
160 | ||
161 | tg = container_of(head, struct throtl_grp, rcu_head); | |
5624a4e4 | 162 | free_percpu(tg->blkg.stats_cpu); |
4843c69d VG |
163 | kfree(tg); |
164 | } | |
165 | ||
e43473b7 VG |
166 | static void throtl_put_tg(struct throtl_grp *tg) |
167 | { | |
168 | BUG_ON(atomic_read(&tg->ref) <= 0); | |
169 | if (!atomic_dec_and_test(&tg->ref)) | |
170 | return; | |
4843c69d VG |
171 | |
172 | /* | |
173 | * A group is freed in rcu manner. But having an rcu lock does not | |
174 | * mean that one can access all the fields of blkg and assume these | |
175 | * are valid. For example, don't try to follow throtl_data and | |
176 | * request queue links. | |
177 | * | |
178 | * Having a reference to blkg under an rcu allows acess to only | |
179 | * values local to groups like group stats and group rate limits | |
180 | */ | |
181 | call_rcu(&tg->rcu_head, throtl_free_tg); | |
e43473b7 VG |
182 | } |
183 | ||
a29a171e VG |
184 | static void throtl_init_group(struct throtl_grp *tg) |
185 | { | |
186 | INIT_HLIST_NODE(&tg->tg_node); | |
187 | RB_CLEAR_NODE(&tg->rb_node); | |
188 | bio_list_init(&tg->bio_lists[0]); | |
189 | bio_list_init(&tg->bio_lists[1]); | |
190 | tg->limits_changed = false; | |
191 | ||
192 | /* Practically unlimited BW */ | |
193 | tg->bps[0] = tg->bps[1] = -1; | |
194 | tg->iops[0] = tg->iops[1] = -1; | |
195 | ||
196 | /* | |
197 | * Take the initial reference that will be released on destroy | |
198 | * This can be thought of a joint reference by cgroup and | |
199 | * request queue which will be dropped by either request queue | |
200 | * exit or cgroup deletion path depending on who is exiting first. | |
201 | */ | |
202 | atomic_set(&tg->ref, 1); | |
203 | } | |
204 | ||
205 | /* Should be called with rcu read lock held (needed for blkcg) */ | |
206 | static void | |
207 | throtl_add_group_to_td_list(struct throtl_data *td, struct throtl_grp *tg) | |
208 | { | |
209 | hlist_add_head(&tg->tg_node, &td->tg_list); | |
210 | td->nr_undestroyed_grps++; | |
211 | } | |
212 | ||
269f5415 VG |
213 | static void |
214 | __throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg) | |
f469a7b4 VG |
215 | { |
216 | struct backing_dev_info *bdi = &td->queue->backing_dev_info; | |
217 | unsigned int major, minor; | |
218 | ||
269f5415 VG |
219 | if (!tg || tg->blkg.dev) |
220 | return; | |
221 | ||
222 | /* | |
223 | * Fill in device details for a group which might not have been | |
224 | * filled at group creation time as queue was being instantiated | |
225 | * and driver had not attached a device yet | |
226 | */ | |
227 | if (bdi->dev && dev_name(bdi->dev)) { | |
228 | sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); | |
229 | tg->blkg.dev = MKDEV(major, minor); | |
230 | } | |
231 | } | |
232 | ||
af75cd3c VG |
233 | /* |
234 | * Should be called with without queue lock held. Here queue lock will be | |
235 | * taken rarely. It will be taken only once during life time of a group | |
236 | * if need be | |
237 | */ | |
238 | static void | |
239 | throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg) | |
240 | { | |
241 | if (!tg || tg->blkg.dev) | |
242 | return; | |
243 | ||
244 | spin_lock_irq(td->queue->queue_lock); | |
245 | __throtl_tg_fill_dev_details(td, tg); | |
246 | spin_unlock_irq(td->queue->queue_lock); | |
247 | } | |
248 | ||
269f5415 VG |
249 | static void throtl_init_add_tg_lists(struct throtl_data *td, |
250 | struct throtl_grp *tg, struct blkio_cgroup *blkcg) | |
251 | { | |
252 | __throtl_tg_fill_dev_details(td, tg); | |
253 | ||
f469a7b4 | 254 | /* Add group onto cgroup list */ |
f469a7b4 | 255 | blkiocg_add_blkio_group(blkcg, &tg->blkg, (void *)td, |
269f5415 | 256 | tg->blkg.dev, BLKIO_POLICY_THROTL); |
f469a7b4 VG |
257 | |
258 | tg->bps[READ] = blkcg_get_read_bps(blkcg, tg->blkg.dev); | |
259 | tg->bps[WRITE] = blkcg_get_write_bps(blkcg, tg->blkg.dev); | |
260 | tg->iops[READ] = blkcg_get_read_iops(blkcg, tg->blkg.dev); | |
261 | tg->iops[WRITE] = blkcg_get_write_iops(blkcg, tg->blkg.dev); | |
262 | ||
263 | throtl_add_group_to_td_list(td, tg); | |
264 | } | |
265 | ||
266 | /* Should be called without queue lock and outside of rcu period */ | |
267 | static struct throtl_grp *throtl_alloc_tg(struct throtl_data *td) | |
268 | { | |
269 | struct throtl_grp *tg = NULL; | |
5624a4e4 | 270 | int ret; |
f469a7b4 VG |
271 | |
272 | tg = kzalloc_node(sizeof(*tg), GFP_ATOMIC, td->queue->node); | |
273 | if (!tg) | |
274 | return NULL; | |
275 | ||
5624a4e4 VG |
276 | ret = blkio_alloc_blkg_stats(&tg->blkg); |
277 | ||
278 | if (ret) { | |
279 | kfree(tg); | |
280 | return NULL; | |
281 | } | |
282 | ||
f469a7b4 VG |
283 | throtl_init_group(tg); |
284 | return tg; | |
285 | } | |
286 | ||
287 | static struct | |
288 | throtl_grp *throtl_find_tg(struct throtl_data *td, struct blkio_cgroup *blkcg) | |
e43473b7 | 289 | { |
e43473b7 VG |
290 | struct throtl_grp *tg = NULL; |
291 | void *key = td; | |
e43473b7 | 292 | |
be2c6b19 VG |
293 | /* |
294 | * This is the common case when there are no blkio cgroups. | |
295 | * Avoid lookup in this case | |
296 | */ | |
297 | if (blkcg == &blkio_root_cgroup) | |
29b12589 | 298 | tg = td->root_tg; |
be2c6b19 VG |
299 | else |
300 | tg = tg_of_blkg(blkiocg_lookup_group(blkcg, key)); | |
e43473b7 | 301 | |
269f5415 | 302 | __throtl_tg_fill_dev_details(td, tg); |
e43473b7 VG |
303 | return tg; |
304 | } | |
305 | ||
306 | static struct throtl_grp * throtl_get_tg(struct throtl_data *td) | |
307 | { | |
f469a7b4 | 308 | struct throtl_grp *tg = NULL, *__tg = NULL; |
70087dc3 | 309 | struct blkio_cgroup *blkcg; |
f469a7b4 | 310 | struct request_queue *q = td->queue; |
e43473b7 | 311 | |
c9a929dd | 312 | /* no throttling for dead queue */ |
34f6055c | 313 | if (unlikely(blk_queue_dead(q))) |
c9a929dd TH |
314 | return NULL; |
315 | ||
e43473b7 | 316 | rcu_read_lock(); |
70087dc3 | 317 | blkcg = task_blkio_cgroup(current); |
f469a7b4 VG |
318 | tg = throtl_find_tg(td, blkcg); |
319 | if (tg) { | |
320 | rcu_read_unlock(); | |
321 | return tg; | |
322 | } | |
323 | ||
324 | /* | |
325 | * Need to allocate a group. Allocation of group also needs allocation | |
326 | * of per cpu stats which in-turn takes a mutex() and can block. Hence | |
315fceee | 327 | * we need to drop rcu lock and queue_lock before we call alloc. |
f469a7b4 | 328 | */ |
f469a7b4 VG |
329 | rcu_read_unlock(); |
330 | spin_unlock_irq(q->queue_lock); | |
331 | ||
332 | tg = throtl_alloc_tg(td); | |
f469a7b4 VG |
333 | |
334 | /* Group allocated and queue is still alive. take the lock */ | |
335 | spin_lock_irq(q->queue_lock); | |
336 | ||
bc16a4f9 | 337 | /* Make sure @q is still alive */ |
34f6055c | 338 | if (unlikely(blk_queue_dead(q))) { |
bc16a4f9 TH |
339 | kfree(tg); |
340 | return NULL; | |
341 | } | |
342 | ||
f469a7b4 VG |
343 | /* |
344 | * Initialize the new group. After sleeping, read the blkcg again. | |
345 | */ | |
346 | rcu_read_lock(); | |
347 | blkcg = task_blkio_cgroup(current); | |
348 | ||
349 | /* | |
350 | * If some other thread already allocated the group while we were | |
351 | * not holding queue lock, free up the group | |
352 | */ | |
353 | __tg = throtl_find_tg(td, blkcg); | |
354 | ||
355 | if (__tg) { | |
356 | kfree(tg); | |
357 | rcu_read_unlock(); | |
358 | return __tg; | |
359 | } | |
360 | ||
361 | /* Group allocation failed. Account the IO to root group */ | |
362 | if (!tg) { | |
29b12589 | 363 | tg = td->root_tg; |
f469a7b4 VG |
364 | return tg; |
365 | } | |
366 | ||
367 | throtl_init_add_tg_lists(td, tg, blkcg); | |
e43473b7 VG |
368 | rcu_read_unlock(); |
369 | return tg; | |
370 | } | |
371 | ||
372 | static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root) | |
373 | { | |
374 | /* Service tree is empty */ | |
375 | if (!root->count) | |
376 | return NULL; | |
377 | ||
378 | if (!root->left) | |
379 | root->left = rb_first(&root->rb); | |
380 | ||
381 | if (root->left) | |
382 | return rb_entry_tg(root->left); | |
383 | ||
384 | return NULL; | |
385 | } | |
386 | ||
387 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
388 | { | |
389 | rb_erase(n, root); | |
390 | RB_CLEAR_NODE(n); | |
391 | } | |
392 | ||
393 | static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root) | |
394 | { | |
395 | if (root->left == n) | |
396 | root->left = NULL; | |
397 | rb_erase_init(n, &root->rb); | |
398 | --root->count; | |
399 | } | |
400 | ||
401 | static void update_min_dispatch_time(struct throtl_rb_root *st) | |
402 | { | |
403 | struct throtl_grp *tg; | |
404 | ||
405 | tg = throtl_rb_first(st); | |
406 | if (!tg) | |
407 | return; | |
408 | ||
409 | st->min_disptime = tg->disptime; | |
410 | } | |
411 | ||
412 | static void | |
413 | tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg) | |
414 | { | |
415 | struct rb_node **node = &st->rb.rb_node; | |
416 | struct rb_node *parent = NULL; | |
417 | struct throtl_grp *__tg; | |
418 | unsigned long key = tg->disptime; | |
419 | int left = 1; | |
420 | ||
421 | while (*node != NULL) { | |
422 | parent = *node; | |
423 | __tg = rb_entry_tg(parent); | |
424 | ||
425 | if (time_before(key, __tg->disptime)) | |
426 | node = &parent->rb_left; | |
427 | else { | |
428 | node = &parent->rb_right; | |
429 | left = 0; | |
430 | } | |
431 | } | |
432 | ||
433 | if (left) | |
434 | st->left = &tg->rb_node; | |
435 | ||
436 | rb_link_node(&tg->rb_node, parent, node); | |
437 | rb_insert_color(&tg->rb_node, &st->rb); | |
438 | } | |
439 | ||
440 | static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
441 | { | |
442 | struct throtl_rb_root *st = &td->tg_service_tree; | |
443 | ||
444 | tg_service_tree_add(st, tg); | |
445 | throtl_mark_tg_on_rr(tg); | |
446 | st->count++; | |
447 | } | |
448 | ||
449 | static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
450 | { | |
451 | if (!throtl_tg_on_rr(tg)) | |
452 | __throtl_enqueue_tg(td, tg); | |
453 | } | |
454 | ||
455 | static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
456 | { | |
457 | throtl_rb_erase(&tg->rb_node, &td->tg_service_tree); | |
458 | throtl_clear_tg_on_rr(tg); | |
459 | } | |
460 | ||
461 | static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
462 | { | |
463 | if (throtl_tg_on_rr(tg)) | |
464 | __throtl_dequeue_tg(td, tg); | |
465 | } | |
466 | ||
467 | static void throtl_schedule_next_dispatch(struct throtl_data *td) | |
468 | { | |
469 | struct throtl_rb_root *st = &td->tg_service_tree; | |
470 | ||
471 | /* | |
472 | * If there are more bios pending, schedule more work. | |
473 | */ | |
474 | if (!total_nr_queued(td)) | |
475 | return; | |
476 | ||
477 | BUG_ON(!st->count); | |
478 | ||
479 | update_min_dispatch_time(st); | |
480 | ||
481 | if (time_before_eq(st->min_disptime, jiffies)) | |
450adcbe | 482 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 483 | else |
450adcbe | 484 | throtl_schedule_delayed_work(td, (st->min_disptime - jiffies)); |
e43473b7 VG |
485 | } |
486 | ||
487 | static inline void | |
488 | throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
489 | { | |
490 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 491 | tg->io_disp[rw] = 0; |
e43473b7 VG |
492 | tg->slice_start[rw] = jiffies; |
493 | tg->slice_end[rw] = jiffies + throtl_slice; | |
494 | throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
495 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
496 | tg->slice_end[rw], jiffies); | |
497 | } | |
498 | ||
d1ae8ffd VG |
499 | static inline void throtl_set_slice_end(struct throtl_data *td, |
500 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
501 | { | |
502 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
503 | } | |
504 | ||
e43473b7 VG |
505 | static inline void throtl_extend_slice(struct throtl_data *td, |
506 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
507 | { | |
508 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
509 | throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
510 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
511 | tg->slice_end[rw], jiffies); | |
512 | } | |
513 | ||
514 | /* Determine if previously allocated or extended slice is complete or not */ | |
515 | static bool | |
516 | throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
517 | { | |
518 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
519 | return 0; | |
520 | ||
521 | return 1; | |
522 | } | |
523 | ||
524 | /* Trim the used slices and adjust slice start accordingly */ | |
525 | static inline void | |
526 | throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
527 | { | |
3aad5d3e VG |
528 | unsigned long nr_slices, time_elapsed, io_trim; |
529 | u64 bytes_trim, tmp; | |
e43473b7 VG |
530 | |
531 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
532 | ||
533 | /* | |
534 | * If bps are unlimited (-1), then time slice don't get | |
535 | * renewed. Don't try to trim the slice if slice is used. A new | |
536 | * slice will start when appropriate. | |
537 | */ | |
538 | if (throtl_slice_used(td, tg, rw)) | |
539 | return; | |
540 | ||
d1ae8ffd VG |
541 | /* |
542 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
543 | * that initially cgroup limit was very low resulting in high | |
544 | * slice_end, but later limit was bumped up and bio was dispached | |
545 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
546 | * is bad because it does not allow new slice to start. | |
547 | */ | |
548 | ||
549 | throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice); | |
550 | ||
e43473b7 VG |
551 | time_elapsed = jiffies - tg->slice_start[rw]; |
552 | ||
553 | nr_slices = time_elapsed / throtl_slice; | |
554 | ||
555 | if (!nr_slices) | |
556 | return; | |
3aad5d3e VG |
557 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
558 | do_div(tmp, HZ); | |
559 | bytes_trim = tmp; | |
e43473b7 | 560 | |
8e89d13f | 561 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 562 | |
8e89d13f | 563 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
564 | return; |
565 | ||
566 | if (tg->bytes_disp[rw] >= bytes_trim) | |
567 | tg->bytes_disp[rw] -= bytes_trim; | |
568 | else | |
569 | tg->bytes_disp[rw] = 0; | |
570 | ||
8e89d13f VG |
571 | if (tg->io_disp[rw] >= io_trim) |
572 | tg->io_disp[rw] -= io_trim; | |
573 | else | |
574 | tg->io_disp[rw] = 0; | |
575 | ||
e43473b7 VG |
576 | tg->slice_start[rw] += nr_slices * throtl_slice; |
577 | ||
3aad5d3e | 578 | throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 579 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 580 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
581 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
582 | } | |
583 | ||
8e89d13f VG |
584 | static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg, |
585 | struct bio *bio, unsigned long *wait) | |
e43473b7 VG |
586 | { |
587 | bool rw = bio_data_dir(bio); | |
8e89d13f | 588 | unsigned int io_allowed; |
e43473b7 | 589 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 590 | u64 tmp; |
e43473b7 | 591 | |
8e89d13f | 592 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 593 | |
8e89d13f VG |
594 | /* Slice has just started. Consider one slice interval */ |
595 | if (!jiffy_elapsed) | |
596 | jiffy_elapsed_rnd = throtl_slice; | |
597 | ||
598 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
599 | ||
c49c06e4 VG |
600 | /* |
601 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
602 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
603 | * will allow dispatch after 1 second and after that slice should | |
604 | * have been trimmed. | |
605 | */ | |
606 | ||
607 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
608 | do_div(tmp, HZ); | |
609 | ||
610 | if (tmp > UINT_MAX) | |
611 | io_allowed = UINT_MAX; | |
612 | else | |
613 | io_allowed = tmp; | |
8e89d13f VG |
614 | |
615 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
616 | if (wait) |
617 | *wait = 0; | |
618 | return 1; | |
619 | } | |
620 | ||
8e89d13f VG |
621 | /* Calc approx time to dispatch */ |
622 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
623 | ||
624 | if (jiffy_wait > jiffy_elapsed) | |
625 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
626 | else | |
627 | jiffy_wait = 1; | |
628 | ||
629 | if (wait) | |
630 | *wait = jiffy_wait; | |
631 | return 0; | |
632 | } | |
633 | ||
634 | static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg, | |
635 | struct bio *bio, unsigned long *wait) | |
636 | { | |
637 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 638 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 639 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
640 | |
641 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
642 | ||
643 | /* Slice has just started. Consider one slice interval */ | |
644 | if (!jiffy_elapsed) | |
645 | jiffy_elapsed_rnd = throtl_slice; | |
646 | ||
647 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
648 | ||
5e901a2b VG |
649 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
650 | do_div(tmp, HZ); | |
3aad5d3e | 651 | bytes_allowed = tmp; |
e43473b7 VG |
652 | |
653 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
654 | if (wait) | |
655 | *wait = 0; | |
656 | return 1; | |
657 | } | |
658 | ||
659 | /* Calc approx time to dispatch */ | |
660 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
661 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
662 | ||
663 | if (!jiffy_wait) | |
664 | jiffy_wait = 1; | |
665 | ||
666 | /* | |
667 | * This wait time is without taking into consideration the rounding | |
668 | * up we did. Add that time also. | |
669 | */ | |
670 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
671 | if (wait) |
672 | *wait = jiffy_wait; | |
8e89d13f VG |
673 | return 0; |
674 | } | |
675 | ||
af75cd3c VG |
676 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
677 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
678 | return 1; | |
679 | return 0; | |
680 | } | |
681 | ||
8e89d13f VG |
682 | /* |
683 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
684 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
685 | */ | |
686 | static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg, | |
687 | struct bio *bio, unsigned long *wait) | |
688 | { | |
689 | bool rw = bio_data_dir(bio); | |
690 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
691 | ||
692 | /* | |
693 | * Currently whole state machine of group depends on first bio | |
694 | * queued in the group bio list. So one should not be calling | |
695 | * this function with a different bio if there are other bios | |
696 | * queued. | |
697 | */ | |
698 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); | |
e43473b7 | 699 | |
8e89d13f VG |
700 | /* If tg->bps = -1, then BW is unlimited */ |
701 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
702 | if (wait) | |
703 | *wait = 0; | |
704 | return 1; | |
705 | } | |
706 | ||
707 | /* | |
708 | * If previous slice expired, start a new one otherwise renew/extend | |
709 | * existing slice to make sure it is at least throtl_slice interval | |
710 | * long since now. | |
711 | */ | |
712 | if (throtl_slice_used(td, tg, rw)) | |
713 | throtl_start_new_slice(td, tg, rw); | |
714 | else { | |
715 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
716 | throtl_extend_slice(td, tg, rw, jiffies + throtl_slice); | |
717 | } | |
718 | ||
719 | if (tg_with_in_bps_limit(td, tg, bio, &bps_wait) | |
720 | && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) { | |
721 | if (wait) | |
722 | *wait = 0; | |
723 | return 1; | |
724 | } | |
725 | ||
726 | max_wait = max(bps_wait, iops_wait); | |
727 | ||
728 | if (wait) | |
729 | *wait = max_wait; | |
730 | ||
731 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
732 | throtl_extend_slice(td, tg, rw, jiffies + max_wait); | |
e43473b7 VG |
733 | |
734 | return 0; | |
735 | } | |
736 | ||
737 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
738 | { | |
739 | bool rw = bio_data_dir(bio); | |
e5a94f56 | 740 | bool sync = rw_is_sync(bio->bi_rw); |
e43473b7 VG |
741 | |
742 | /* Charge the bio to the group */ | |
743 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 744 | tg->io_disp[rw]++; |
e43473b7 | 745 | |
e43473b7 | 746 | blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync); |
e43473b7 VG |
747 | } |
748 | ||
749 | static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg, | |
750 | struct bio *bio) | |
751 | { | |
752 | bool rw = bio_data_dir(bio); | |
753 | ||
754 | bio_list_add(&tg->bio_lists[rw], bio); | |
755 | /* Take a bio reference on tg */ | |
756 | throtl_ref_get_tg(tg); | |
757 | tg->nr_queued[rw]++; | |
758 | td->nr_queued[rw]++; | |
759 | throtl_enqueue_tg(td, tg); | |
760 | } | |
761 | ||
762 | static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg) | |
763 | { | |
764 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
765 | struct bio *bio; | |
766 | ||
767 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
768 | tg_may_dispatch(td, tg, bio, &read_wait); | |
769 | ||
770 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
771 | tg_may_dispatch(td, tg, bio, &write_wait); | |
772 | ||
773 | min_wait = min(read_wait, write_wait); | |
774 | disptime = jiffies + min_wait; | |
775 | ||
e43473b7 VG |
776 | /* Update dispatch time */ |
777 | throtl_dequeue_tg(td, tg); | |
778 | tg->disptime = disptime; | |
779 | throtl_enqueue_tg(td, tg); | |
780 | } | |
781 | ||
782 | static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg, | |
783 | bool rw, struct bio_list *bl) | |
784 | { | |
785 | struct bio *bio; | |
786 | ||
787 | bio = bio_list_pop(&tg->bio_lists[rw]); | |
788 | tg->nr_queued[rw]--; | |
789 | /* Drop bio reference on tg */ | |
790 | throtl_put_tg(tg); | |
791 | ||
792 | BUG_ON(td->nr_queued[rw] <= 0); | |
793 | td->nr_queued[rw]--; | |
794 | ||
795 | throtl_charge_bio(tg, bio); | |
796 | bio_list_add(bl, bio); | |
797 | bio->bi_rw |= REQ_THROTTLED; | |
798 | ||
799 | throtl_trim_slice(td, tg, rw); | |
800 | } | |
801 | ||
802 | static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg, | |
803 | struct bio_list *bl) | |
804 | { | |
805 | unsigned int nr_reads = 0, nr_writes = 0; | |
806 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 807 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
808 | struct bio *bio; |
809 | ||
810 | /* Try to dispatch 75% READS and 25% WRITES */ | |
811 | ||
812 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) | |
813 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
814 | ||
815 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
816 | nr_reads++; | |
817 | ||
818 | if (nr_reads >= max_nr_reads) | |
819 | break; | |
820 | } | |
821 | ||
822 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) | |
823 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
824 | ||
825 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
826 | nr_writes++; | |
827 | ||
828 | if (nr_writes >= max_nr_writes) | |
829 | break; | |
830 | } | |
831 | ||
832 | return nr_reads + nr_writes; | |
833 | } | |
834 | ||
835 | static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl) | |
836 | { | |
837 | unsigned int nr_disp = 0; | |
838 | struct throtl_grp *tg; | |
839 | struct throtl_rb_root *st = &td->tg_service_tree; | |
840 | ||
841 | while (1) { | |
842 | tg = throtl_rb_first(st); | |
843 | ||
844 | if (!tg) | |
845 | break; | |
846 | ||
847 | if (time_before(jiffies, tg->disptime)) | |
848 | break; | |
849 | ||
850 | throtl_dequeue_tg(td, tg); | |
851 | ||
852 | nr_disp += throtl_dispatch_tg(td, tg, bl); | |
853 | ||
854 | if (tg->nr_queued[0] || tg->nr_queued[1]) { | |
855 | tg_update_disptime(td, tg); | |
856 | throtl_enqueue_tg(td, tg); | |
857 | } | |
858 | ||
859 | if (nr_disp >= throtl_quantum) | |
860 | break; | |
861 | } | |
862 | ||
863 | return nr_disp; | |
864 | } | |
865 | ||
fe071437 VG |
866 | static void throtl_process_limit_change(struct throtl_data *td) |
867 | { | |
868 | struct throtl_grp *tg; | |
869 | struct hlist_node *pos, *n; | |
870 | ||
de701c74 | 871 | if (!td->limits_changed) |
fe071437 VG |
872 | return; |
873 | ||
de701c74 | 874 | xchg(&td->limits_changed, false); |
fe071437 | 875 | |
de701c74 | 876 | throtl_log(td, "limits changed"); |
fe071437 | 877 | |
04a6b516 | 878 | hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) { |
de701c74 VG |
879 | if (!tg->limits_changed) |
880 | continue; | |
881 | ||
882 | if (!xchg(&tg->limits_changed, false)) | |
883 | continue; | |
884 | ||
885 | throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu" | |
886 | " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE], | |
887 | tg->iops[READ], tg->iops[WRITE]); | |
888 | ||
04521db0 VG |
889 | /* |
890 | * Restart the slices for both READ and WRITES. It | |
891 | * might happen that a group's limit are dropped | |
892 | * suddenly and we don't want to account recently | |
893 | * dispatched IO with new low rate | |
894 | */ | |
895 | throtl_start_new_slice(td, tg, 0); | |
896 | throtl_start_new_slice(td, tg, 1); | |
897 | ||
de701c74 | 898 | if (throtl_tg_on_rr(tg)) |
fe071437 | 899 | tg_update_disptime(td, tg); |
fe071437 | 900 | } |
fe071437 VG |
901 | } |
902 | ||
e43473b7 VG |
903 | /* Dispatch throttled bios. Should be called without queue lock held. */ |
904 | static int throtl_dispatch(struct request_queue *q) | |
905 | { | |
906 | struct throtl_data *td = q->td; | |
907 | unsigned int nr_disp = 0; | |
908 | struct bio_list bio_list_on_stack; | |
909 | struct bio *bio; | |
69d60eb9 | 910 | struct blk_plug plug; |
e43473b7 VG |
911 | |
912 | spin_lock_irq(q->queue_lock); | |
913 | ||
fe071437 VG |
914 | throtl_process_limit_change(td); |
915 | ||
e43473b7 VG |
916 | if (!total_nr_queued(td)) |
917 | goto out; | |
918 | ||
919 | bio_list_init(&bio_list_on_stack); | |
920 | ||
d2f31a5f | 921 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
e43473b7 VG |
922 | total_nr_queued(td), td->nr_queued[READ], |
923 | td->nr_queued[WRITE]); | |
924 | ||
925 | nr_disp = throtl_select_dispatch(td, &bio_list_on_stack); | |
926 | ||
927 | if (nr_disp) | |
928 | throtl_log(td, "bios disp=%u", nr_disp); | |
929 | ||
930 | throtl_schedule_next_dispatch(td); | |
931 | out: | |
932 | spin_unlock_irq(q->queue_lock); | |
933 | ||
934 | /* | |
935 | * If we dispatched some requests, unplug the queue to make sure | |
936 | * immediate dispatch | |
937 | */ | |
938 | if (nr_disp) { | |
69d60eb9 | 939 | blk_start_plug(&plug); |
e43473b7 VG |
940 | while((bio = bio_list_pop(&bio_list_on_stack))) |
941 | generic_make_request(bio); | |
69d60eb9 | 942 | blk_finish_plug(&plug); |
e43473b7 VG |
943 | } |
944 | return nr_disp; | |
945 | } | |
946 | ||
947 | void blk_throtl_work(struct work_struct *work) | |
948 | { | |
949 | struct throtl_data *td = container_of(work, struct throtl_data, | |
950 | throtl_work.work); | |
951 | struct request_queue *q = td->queue; | |
952 | ||
953 | throtl_dispatch(q); | |
954 | } | |
955 | ||
956 | /* Call with queue lock held */ | |
450adcbe VG |
957 | static void |
958 | throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay) | |
e43473b7 VG |
959 | { |
960 | ||
e43473b7 VG |
961 | struct delayed_work *dwork = &td->throtl_work; |
962 | ||
04521db0 | 963 | /* schedule work if limits changed even if no bio is queued */ |
d2f31a5f | 964 | if (total_nr_queued(td) || td->limits_changed) { |
e43473b7 VG |
965 | /* |
966 | * We might have a work scheduled to be executed in future. | |
967 | * Cancel that and schedule a new one. | |
968 | */ | |
969 | __cancel_delayed_work(dwork); | |
450adcbe | 970 | queue_delayed_work(kthrotld_workqueue, dwork, delay); |
e43473b7 VG |
971 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", |
972 | delay, jiffies); | |
973 | } | |
974 | } | |
e43473b7 VG |
975 | |
976 | static void | |
977 | throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg) | |
978 | { | |
979 | /* Something wrong if we are trying to remove same group twice */ | |
980 | BUG_ON(hlist_unhashed(&tg->tg_node)); | |
981 | ||
982 | hlist_del_init(&tg->tg_node); | |
983 | ||
984 | /* | |
985 | * Put the reference taken at the time of creation so that when all | |
986 | * queues are gone, group can be destroyed. | |
987 | */ | |
988 | throtl_put_tg(tg); | |
989 | td->nr_undestroyed_grps--; | |
990 | } | |
991 | ||
992 | static void throtl_release_tgs(struct throtl_data *td) | |
993 | { | |
994 | struct hlist_node *pos, *n; | |
995 | struct throtl_grp *tg; | |
996 | ||
997 | hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) { | |
998 | /* | |
999 | * If cgroup removal path got to blk_group first and removed | |
1000 | * it from cgroup list, then it will take care of destroying | |
1001 | * cfqg also. | |
1002 | */ | |
1003 | if (!blkiocg_del_blkio_group(&tg->blkg)) | |
1004 | throtl_destroy_tg(td, tg); | |
1005 | } | |
1006 | } | |
1007 | ||
e43473b7 VG |
1008 | /* |
1009 | * Blk cgroup controller notification saying that blkio_group object is being | |
1010 | * delinked as associated cgroup object is going away. That also means that | |
1011 | * no new IO will come in this group. So get rid of this group as soon as | |
1012 | * any pending IO in the group is finished. | |
1013 | * | |
1014 | * This function is called under rcu_read_lock(). key is the rcu protected | |
1015 | * pointer. That means "key" is a valid throtl_data pointer as long as we are | |
1016 | * rcu read lock. | |
1017 | * | |
1018 | * "key" was fetched from blkio_group under blkio_cgroup->lock. That means | |
1019 | * it should not be NULL as even if queue was going away, cgroup deltion | |
1020 | * path got to it first. | |
1021 | */ | |
1022 | void throtl_unlink_blkio_group(void *key, struct blkio_group *blkg) | |
1023 | { | |
1024 | unsigned long flags; | |
1025 | struct throtl_data *td = key; | |
1026 | ||
1027 | spin_lock_irqsave(td->queue->queue_lock, flags); | |
1028 | throtl_destroy_tg(td, tg_of_blkg(blkg)); | |
1029 | spin_unlock_irqrestore(td->queue->queue_lock, flags); | |
1030 | } | |
1031 | ||
de701c74 VG |
1032 | static void throtl_update_blkio_group_common(struct throtl_data *td, |
1033 | struct throtl_grp *tg) | |
1034 | { | |
1035 | xchg(&tg->limits_changed, true); | |
1036 | xchg(&td->limits_changed, true); | |
1037 | /* Schedule a work now to process the limit change */ | |
1038 | throtl_schedule_delayed_work(td, 0); | |
1039 | } | |
1040 | ||
fe071437 VG |
1041 | /* |
1042 | * For all update functions, key should be a valid pointer because these | |
1043 | * update functions are called under blkcg_lock, that means, blkg is | |
25985edc | 1044 | * valid and in turn key is valid. queue exit path can not race because |
fe071437 VG |
1045 | * of blkcg_lock |
1046 | * | |
1047 | * Can not take queue lock in update functions as queue lock under blkcg_lock | |
1048 | * is not allowed. Under other paths we take blkcg_lock under queue_lock. | |
1049 | */ | |
1050 | static void throtl_update_blkio_group_read_bps(void *key, | |
1051 | struct blkio_group *blkg, u64 read_bps) | |
e43473b7 | 1052 | { |
fe071437 | 1053 | struct throtl_data *td = key; |
de701c74 | 1054 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1055 | |
de701c74 VG |
1056 | tg->bps[READ] = read_bps; |
1057 | throtl_update_blkio_group_common(td, tg); | |
e43473b7 VG |
1058 | } |
1059 | ||
fe071437 VG |
1060 | static void throtl_update_blkio_group_write_bps(void *key, |
1061 | struct blkio_group *blkg, u64 write_bps) | |
e43473b7 | 1062 | { |
fe071437 | 1063 | struct throtl_data *td = key; |
de701c74 | 1064 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1065 | |
de701c74 VG |
1066 | tg->bps[WRITE] = write_bps; |
1067 | throtl_update_blkio_group_common(td, tg); | |
e43473b7 VG |
1068 | } |
1069 | ||
fe071437 VG |
1070 | static void throtl_update_blkio_group_read_iops(void *key, |
1071 | struct blkio_group *blkg, unsigned int read_iops) | |
8e89d13f | 1072 | { |
fe071437 | 1073 | struct throtl_data *td = key; |
de701c74 | 1074 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1075 | |
de701c74 VG |
1076 | tg->iops[READ] = read_iops; |
1077 | throtl_update_blkio_group_common(td, tg); | |
8e89d13f VG |
1078 | } |
1079 | ||
fe071437 VG |
1080 | static void throtl_update_blkio_group_write_iops(void *key, |
1081 | struct blkio_group *blkg, unsigned int write_iops) | |
8e89d13f | 1082 | { |
fe071437 | 1083 | struct throtl_data *td = key; |
de701c74 | 1084 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1085 | |
de701c74 VG |
1086 | tg->iops[WRITE] = write_iops; |
1087 | throtl_update_blkio_group_common(td, tg); | |
8e89d13f VG |
1088 | } |
1089 | ||
da527770 | 1090 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1091 | { |
1092 | struct throtl_data *td = q->td; | |
1093 | ||
1094 | cancel_delayed_work_sync(&td->throtl_work); | |
1095 | } | |
1096 | ||
1097 | static struct blkio_policy_type blkio_policy_throtl = { | |
1098 | .ops = { | |
1099 | .blkio_unlink_group_fn = throtl_unlink_blkio_group, | |
1100 | .blkio_update_group_read_bps_fn = | |
1101 | throtl_update_blkio_group_read_bps, | |
1102 | .blkio_update_group_write_bps_fn = | |
1103 | throtl_update_blkio_group_write_bps, | |
8e89d13f VG |
1104 | .blkio_update_group_read_iops_fn = |
1105 | throtl_update_blkio_group_read_iops, | |
1106 | .blkio_update_group_write_iops_fn = | |
1107 | throtl_update_blkio_group_write_iops, | |
e43473b7 | 1108 | }, |
8e89d13f | 1109 | .plid = BLKIO_POLICY_THROTL, |
e43473b7 VG |
1110 | }; |
1111 | ||
bc16a4f9 | 1112 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1113 | { |
1114 | struct throtl_data *td = q->td; | |
1115 | struct throtl_grp *tg; | |
e43473b7 | 1116 | bool rw = bio_data_dir(bio), update_disptime = true; |
af75cd3c | 1117 | struct blkio_cgroup *blkcg; |
bc16a4f9 | 1118 | bool throttled = false; |
e43473b7 VG |
1119 | |
1120 | if (bio->bi_rw & REQ_THROTTLED) { | |
1121 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1122 | goto out; |
e43473b7 VG |
1123 | } |
1124 | ||
af75cd3c VG |
1125 | /* |
1126 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1127 | * basic fields like stats and io rates. If a group has no rules, | |
1128 | * just update the dispatch stats in lockless manner and return. | |
1129 | */ | |
1130 | ||
1131 | rcu_read_lock(); | |
1132 | blkcg = task_blkio_cgroup(current); | |
1133 | tg = throtl_find_tg(td, blkcg); | |
1134 | if (tg) { | |
1135 | throtl_tg_fill_dev_details(td, tg); | |
1136 | ||
1137 | if (tg_no_rule_group(tg, rw)) { | |
1138 | blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, | |
e5a94f56 | 1139 | rw, rw_is_sync(bio->bi_rw)); |
af75cd3c | 1140 | rcu_read_unlock(); |
bc16a4f9 | 1141 | goto out; |
af75cd3c VG |
1142 | } |
1143 | } | |
1144 | rcu_read_unlock(); | |
1145 | ||
1146 | /* | |
1147 | * Either group has not been allocated yet or it is not an unlimited | |
1148 | * IO group | |
1149 | */ | |
e43473b7 VG |
1150 | spin_lock_irq(q->queue_lock); |
1151 | tg = throtl_get_tg(td); | |
bc16a4f9 TH |
1152 | if (unlikely(!tg)) |
1153 | goto out_unlock; | |
f469a7b4 | 1154 | |
e43473b7 VG |
1155 | if (tg->nr_queued[rw]) { |
1156 | /* | |
1157 | * There is already another bio queued in same dir. No | |
1158 | * need to update dispatch time. | |
1159 | */ | |
231d704b | 1160 | update_disptime = false; |
e43473b7 | 1161 | goto queue_bio; |
de701c74 | 1162 | |
e43473b7 VG |
1163 | } |
1164 | ||
1165 | /* Bio is with-in rate limit of group */ | |
1166 | if (tg_may_dispatch(td, tg, bio, NULL)) { | |
1167 | throtl_charge_bio(tg, bio); | |
04521db0 VG |
1168 | |
1169 | /* | |
1170 | * We need to trim slice even when bios are not being queued | |
1171 | * otherwise it might happen that a bio is not queued for | |
1172 | * a long time and slice keeps on extending and trim is not | |
1173 | * called for a long time. Now if limits are reduced suddenly | |
1174 | * we take into account all the IO dispatched so far at new | |
1175 | * low rate and * newly queued IO gets a really long dispatch | |
1176 | * time. | |
1177 | * | |
1178 | * So keep on trimming slice even if bio is not queued. | |
1179 | */ | |
1180 | throtl_trim_slice(td, tg, rw); | |
bc16a4f9 | 1181 | goto out_unlock; |
e43473b7 VG |
1182 | } |
1183 | ||
1184 | queue_bio: | |
fd16d263 | 1185 | throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
8e89d13f VG |
1186 | " iodisp=%u iops=%u queued=%d/%d", |
1187 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1188 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1189 | tg->io_disp[rw], tg->iops[rw], |
e43473b7 VG |
1190 | tg->nr_queued[READ], tg->nr_queued[WRITE]); |
1191 | ||
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); |
bc16a4f9 TH |
1202 | out: |
1203 | return throttled; | |
e43473b7 VG |
1204 | } |
1205 | ||
c9a929dd TH |
1206 | /** |
1207 | * blk_throtl_drain - drain throttled bios | |
1208 | * @q: request_queue to drain throttled bios for | |
1209 | * | |
1210 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1211 | */ | |
1212 | void blk_throtl_drain(struct request_queue *q) | |
1213 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1214 | { | |
1215 | struct throtl_data *td = q->td; | |
1216 | struct throtl_rb_root *st = &td->tg_service_tree; | |
1217 | struct throtl_grp *tg; | |
1218 | struct bio_list bl; | |
1219 | struct bio *bio; | |
1220 | ||
334c2b0b | 1221 | WARN_ON_ONCE(!queue_is_locked(q)); |
c9a929dd TH |
1222 | |
1223 | bio_list_init(&bl); | |
1224 | ||
1225 | while ((tg = throtl_rb_first(st))) { | |
1226 | throtl_dequeue_tg(td, tg); | |
1227 | ||
1228 | while ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
1229 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1230 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
1231 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1232 | } | |
1233 | spin_unlock_irq(q->queue_lock); | |
1234 | ||
1235 | while ((bio = bio_list_pop(&bl))) | |
1236 | generic_make_request(bio); | |
1237 | ||
1238 | spin_lock_irq(q->queue_lock); | |
1239 | } | |
1240 | ||
e43473b7 VG |
1241 | int blk_throtl_init(struct request_queue *q) |
1242 | { | |
1243 | struct throtl_data *td; | |
1244 | struct throtl_grp *tg; | |
1245 | ||
1246 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1247 | if (!td) | |
1248 | return -ENOMEM; | |
1249 | ||
1250 | INIT_HLIST_HEAD(&td->tg_list); | |
1251 | td->tg_service_tree = THROTL_RB_ROOT; | |
de701c74 | 1252 | td->limits_changed = false; |
a29a171e | 1253 | INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work); |
e43473b7 | 1254 | |
29b12589 VG |
1255 | /* alloc and Init root group. */ |
1256 | td->queue = q; | |
1257 | tg = throtl_alloc_tg(td); | |
02977e4a | 1258 | |
29b12589 VG |
1259 | if (!tg) { |
1260 | kfree(td); | |
1261 | return -ENOMEM; | |
1262 | } | |
1263 | ||
1264 | td->root_tg = tg; | |
e43473b7 VG |
1265 | |
1266 | rcu_read_lock(); | |
5617cbef | 1267 | throtl_init_add_tg_lists(td, tg, &blkio_root_cgroup); |
e43473b7 VG |
1268 | rcu_read_unlock(); |
1269 | ||
1270 | /* Attach throtl data to request queue */ | |
e43473b7 VG |
1271 | q->td = td; |
1272 | return 0; | |
1273 | } | |
1274 | ||
1275 | void blk_throtl_exit(struct request_queue *q) | |
1276 | { | |
1277 | struct throtl_data *td = q->td; | |
1278 | bool wait = false; | |
1279 | ||
1280 | BUG_ON(!td); | |
1281 | ||
da527770 | 1282 | throtl_shutdown_wq(q); |
e43473b7 VG |
1283 | |
1284 | spin_lock_irq(q->queue_lock); | |
1285 | throtl_release_tgs(td); | |
e43473b7 VG |
1286 | |
1287 | /* If there are other groups */ | |
02977e4a | 1288 | if (td->nr_undestroyed_grps > 0) |
e43473b7 VG |
1289 | wait = true; |
1290 | ||
1291 | spin_unlock_irq(q->queue_lock); | |
1292 | ||
1293 | /* | |
1294 | * Wait for tg->blkg->key accessors to exit their grace periods. | |
1295 | * Do this wait only if there are other undestroyed groups out | |
1296 | * there (other than root group). This can happen if cgroup deletion | |
1297 | * path claimed the responsibility of cleaning up a group before | |
1298 | * queue cleanup code get to the group. | |
1299 | * | |
1300 | * Do not call synchronize_rcu() unconditionally as there are drivers | |
1301 | * which create/delete request queue hundreds of times during scan/boot | |
1302 | * and synchronize_rcu() can take significant time and slow down boot. | |
1303 | */ | |
1304 | if (wait) | |
1305 | synchronize_rcu(); | |
fe071437 VG |
1306 | |
1307 | /* | |
1308 | * Just being safe to make sure after previous flush if some body did | |
1309 | * update limits through cgroup and another work got queued, cancel | |
1310 | * it. | |
1311 | */ | |
da527770 | 1312 | throtl_shutdown_wq(q); |
c9a929dd TH |
1313 | } |
1314 | ||
1315 | void blk_throtl_release(struct request_queue *q) | |
1316 | { | |
1317 | kfree(q->td); | |
e43473b7 VG |
1318 | } |
1319 | ||
1320 | static int __init throtl_init(void) | |
1321 | { | |
450adcbe VG |
1322 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1323 | if (!kthrotld_workqueue) | |
1324 | panic("Failed to create kthrotld\n"); | |
1325 | ||
e43473b7 VG |
1326 | blkio_policy_register(&blkio_policy_throtl); |
1327 | return 0; | |
1328 | } | |
1329 | ||
1330 | module_init(throtl_init); |