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 VG |
311 | |
312 | rcu_read_lock(); | |
70087dc3 | 313 | blkcg = task_blkio_cgroup(current); |
f469a7b4 VG |
314 | tg = throtl_find_tg(td, blkcg); |
315 | if (tg) { | |
316 | rcu_read_unlock(); | |
317 | return tg; | |
318 | } | |
319 | ||
320 | /* | |
321 | * Need to allocate a group. Allocation of group also needs allocation | |
322 | * of per cpu stats which in-turn takes a mutex() and can block. Hence | |
315fceee | 323 | * we need to drop rcu lock and queue_lock before we call alloc. |
f469a7b4 | 324 | */ |
f469a7b4 VG |
325 | rcu_read_unlock(); |
326 | spin_unlock_irq(q->queue_lock); | |
327 | ||
328 | tg = throtl_alloc_tg(td); | |
f469a7b4 VG |
329 | |
330 | /* Group allocated and queue is still alive. take the lock */ | |
331 | spin_lock_irq(q->queue_lock); | |
332 | ||
bc16a4f9 TH |
333 | /* Make sure @q is still alive */ |
334 | if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) { | |
335 | kfree(tg); | |
336 | return NULL; | |
337 | } | |
338 | ||
f469a7b4 VG |
339 | /* |
340 | * Initialize the new group. After sleeping, read the blkcg again. | |
341 | */ | |
342 | rcu_read_lock(); | |
343 | blkcg = task_blkio_cgroup(current); | |
344 | ||
345 | /* | |
346 | * If some other thread already allocated the group while we were | |
347 | * not holding queue lock, free up the group | |
348 | */ | |
349 | __tg = throtl_find_tg(td, blkcg); | |
350 | ||
351 | if (__tg) { | |
352 | kfree(tg); | |
353 | rcu_read_unlock(); | |
354 | return __tg; | |
355 | } | |
356 | ||
357 | /* Group allocation failed. Account the IO to root group */ | |
358 | if (!tg) { | |
29b12589 | 359 | tg = td->root_tg; |
f469a7b4 VG |
360 | return tg; |
361 | } | |
362 | ||
363 | throtl_init_add_tg_lists(td, tg, blkcg); | |
e43473b7 VG |
364 | rcu_read_unlock(); |
365 | return tg; | |
366 | } | |
367 | ||
368 | static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root) | |
369 | { | |
370 | /* Service tree is empty */ | |
371 | if (!root->count) | |
372 | return NULL; | |
373 | ||
374 | if (!root->left) | |
375 | root->left = rb_first(&root->rb); | |
376 | ||
377 | if (root->left) | |
378 | return rb_entry_tg(root->left); | |
379 | ||
380 | return NULL; | |
381 | } | |
382 | ||
383 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
384 | { | |
385 | rb_erase(n, root); | |
386 | RB_CLEAR_NODE(n); | |
387 | } | |
388 | ||
389 | static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root) | |
390 | { | |
391 | if (root->left == n) | |
392 | root->left = NULL; | |
393 | rb_erase_init(n, &root->rb); | |
394 | --root->count; | |
395 | } | |
396 | ||
397 | static void update_min_dispatch_time(struct throtl_rb_root *st) | |
398 | { | |
399 | struct throtl_grp *tg; | |
400 | ||
401 | tg = throtl_rb_first(st); | |
402 | if (!tg) | |
403 | return; | |
404 | ||
405 | st->min_disptime = tg->disptime; | |
406 | } | |
407 | ||
408 | static void | |
409 | tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg) | |
410 | { | |
411 | struct rb_node **node = &st->rb.rb_node; | |
412 | struct rb_node *parent = NULL; | |
413 | struct throtl_grp *__tg; | |
414 | unsigned long key = tg->disptime; | |
415 | int left = 1; | |
416 | ||
417 | while (*node != NULL) { | |
418 | parent = *node; | |
419 | __tg = rb_entry_tg(parent); | |
420 | ||
421 | if (time_before(key, __tg->disptime)) | |
422 | node = &parent->rb_left; | |
423 | else { | |
424 | node = &parent->rb_right; | |
425 | left = 0; | |
426 | } | |
427 | } | |
428 | ||
429 | if (left) | |
430 | st->left = &tg->rb_node; | |
431 | ||
432 | rb_link_node(&tg->rb_node, parent, node); | |
433 | rb_insert_color(&tg->rb_node, &st->rb); | |
434 | } | |
435 | ||
436 | static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
437 | { | |
438 | struct throtl_rb_root *st = &td->tg_service_tree; | |
439 | ||
440 | tg_service_tree_add(st, tg); | |
441 | throtl_mark_tg_on_rr(tg); | |
442 | st->count++; | |
443 | } | |
444 | ||
445 | static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
446 | { | |
447 | if (!throtl_tg_on_rr(tg)) | |
448 | __throtl_enqueue_tg(td, tg); | |
449 | } | |
450 | ||
451 | static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
452 | { | |
453 | throtl_rb_erase(&tg->rb_node, &td->tg_service_tree); | |
454 | throtl_clear_tg_on_rr(tg); | |
455 | } | |
456 | ||
457 | static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
458 | { | |
459 | if (throtl_tg_on_rr(tg)) | |
460 | __throtl_dequeue_tg(td, tg); | |
461 | } | |
462 | ||
463 | static void throtl_schedule_next_dispatch(struct throtl_data *td) | |
464 | { | |
465 | struct throtl_rb_root *st = &td->tg_service_tree; | |
466 | ||
467 | /* | |
468 | * If there are more bios pending, schedule more work. | |
469 | */ | |
470 | if (!total_nr_queued(td)) | |
471 | return; | |
472 | ||
473 | BUG_ON(!st->count); | |
474 | ||
475 | update_min_dispatch_time(st); | |
476 | ||
477 | if (time_before_eq(st->min_disptime, jiffies)) | |
450adcbe | 478 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 479 | else |
450adcbe | 480 | throtl_schedule_delayed_work(td, (st->min_disptime - jiffies)); |
e43473b7 VG |
481 | } |
482 | ||
483 | static inline void | |
484 | throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
485 | { | |
486 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 487 | tg->io_disp[rw] = 0; |
e43473b7 VG |
488 | tg->slice_start[rw] = jiffies; |
489 | tg->slice_end[rw] = jiffies + throtl_slice; | |
490 | throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
491 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
492 | tg->slice_end[rw], jiffies); | |
493 | } | |
494 | ||
d1ae8ffd VG |
495 | static inline void throtl_set_slice_end(struct throtl_data *td, |
496 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
497 | { | |
498 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
499 | } | |
500 | ||
e43473b7 VG |
501 | static inline void throtl_extend_slice(struct throtl_data *td, |
502 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
503 | { | |
504 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
505 | throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
506 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
507 | tg->slice_end[rw], jiffies); | |
508 | } | |
509 | ||
510 | /* Determine if previously allocated or extended slice is complete or not */ | |
511 | static bool | |
512 | throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
513 | { | |
514 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
515 | return 0; | |
516 | ||
517 | return 1; | |
518 | } | |
519 | ||
520 | /* Trim the used slices and adjust slice start accordingly */ | |
521 | static inline void | |
522 | throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
523 | { | |
3aad5d3e VG |
524 | unsigned long nr_slices, time_elapsed, io_trim; |
525 | u64 bytes_trim, tmp; | |
e43473b7 VG |
526 | |
527 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
528 | ||
529 | /* | |
530 | * If bps are unlimited (-1), then time slice don't get | |
531 | * renewed. Don't try to trim the slice if slice is used. A new | |
532 | * slice will start when appropriate. | |
533 | */ | |
534 | if (throtl_slice_used(td, tg, rw)) | |
535 | return; | |
536 | ||
d1ae8ffd VG |
537 | /* |
538 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
539 | * that initially cgroup limit was very low resulting in high | |
540 | * slice_end, but later limit was bumped up and bio was dispached | |
541 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
542 | * is bad because it does not allow new slice to start. | |
543 | */ | |
544 | ||
545 | throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice); | |
546 | ||
e43473b7 VG |
547 | time_elapsed = jiffies - tg->slice_start[rw]; |
548 | ||
549 | nr_slices = time_elapsed / throtl_slice; | |
550 | ||
551 | if (!nr_slices) | |
552 | return; | |
3aad5d3e VG |
553 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
554 | do_div(tmp, HZ); | |
555 | bytes_trim = tmp; | |
e43473b7 | 556 | |
8e89d13f | 557 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 558 | |
8e89d13f | 559 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
560 | return; |
561 | ||
562 | if (tg->bytes_disp[rw] >= bytes_trim) | |
563 | tg->bytes_disp[rw] -= bytes_trim; | |
564 | else | |
565 | tg->bytes_disp[rw] = 0; | |
566 | ||
8e89d13f VG |
567 | if (tg->io_disp[rw] >= io_trim) |
568 | tg->io_disp[rw] -= io_trim; | |
569 | else | |
570 | tg->io_disp[rw] = 0; | |
571 | ||
e43473b7 VG |
572 | tg->slice_start[rw] += nr_slices * throtl_slice; |
573 | ||
3aad5d3e | 574 | throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 575 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 576 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
577 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
578 | } | |
579 | ||
8e89d13f VG |
580 | static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg, |
581 | struct bio *bio, unsigned long *wait) | |
e43473b7 VG |
582 | { |
583 | bool rw = bio_data_dir(bio); | |
8e89d13f | 584 | unsigned int io_allowed; |
e43473b7 | 585 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 586 | u64 tmp; |
e43473b7 | 587 | |
8e89d13f | 588 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 589 | |
8e89d13f VG |
590 | /* Slice has just started. Consider one slice interval */ |
591 | if (!jiffy_elapsed) | |
592 | jiffy_elapsed_rnd = throtl_slice; | |
593 | ||
594 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
595 | ||
c49c06e4 VG |
596 | /* |
597 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
598 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
599 | * will allow dispatch after 1 second and after that slice should | |
600 | * have been trimmed. | |
601 | */ | |
602 | ||
603 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
604 | do_div(tmp, HZ); | |
605 | ||
606 | if (tmp > UINT_MAX) | |
607 | io_allowed = UINT_MAX; | |
608 | else | |
609 | io_allowed = tmp; | |
8e89d13f VG |
610 | |
611 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
612 | if (wait) |
613 | *wait = 0; | |
614 | return 1; | |
615 | } | |
616 | ||
8e89d13f VG |
617 | /* Calc approx time to dispatch */ |
618 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
619 | ||
620 | if (jiffy_wait > jiffy_elapsed) | |
621 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
622 | else | |
623 | jiffy_wait = 1; | |
624 | ||
625 | if (wait) | |
626 | *wait = jiffy_wait; | |
627 | return 0; | |
628 | } | |
629 | ||
630 | static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg, | |
631 | struct bio *bio, unsigned long *wait) | |
632 | { | |
633 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 634 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 635 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
636 | |
637 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
638 | ||
639 | /* Slice has just started. Consider one slice interval */ | |
640 | if (!jiffy_elapsed) | |
641 | jiffy_elapsed_rnd = throtl_slice; | |
642 | ||
643 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
644 | ||
5e901a2b VG |
645 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
646 | do_div(tmp, HZ); | |
3aad5d3e | 647 | bytes_allowed = tmp; |
e43473b7 VG |
648 | |
649 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
650 | if (wait) | |
651 | *wait = 0; | |
652 | return 1; | |
653 | } | |
654 | ||
655 | /* Calc approx time to dispatch */ | |
656 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
657 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
658 | ||
659 | if (!jiffy_wait) | |
660 | jiffy_wait = 1; | |
661 | ||
662 | /* | |
663 | * This wait time is without taking into consideration the rounding | |
664 | * up we did. Add that time also. | |
665 | */ | |
666 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
667 | if (wait) |
668 | *wait = jiffy_wait; | |
8e89d13f VG |
669 | return 0; |
670 | } | |
671 | ||
af75cd3c VG |
672 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
673 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
674 | return 1; | |
675 | return 0; | |
676 | } | |
677 | ||
8e89d13f VG |
678 | /* |
679 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
680 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
681 | */ | |
682 | static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg, | |
683 | struct bio *bio, unsigned long *wait) | |
684 | { | |
685 | bool rw = bio_data_dir(bio); | |
686 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
687 | ||
688 | /* | |
689 | * Currently whole state machine of group depends on first bio | |
690 | * queued in the group bio list. So one should not be calling | |
691 | * this function with a different bio if there are other bios | |
692 | * queued. | |
693 | */ | |
694 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); | |
e43473b7 | 695 | |
8e89d13f VG |
696 | /* If tg->bps = -1, then BW is unlimited */ |
697 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
698 | if (wait) | |
699 | *wait = 0; | |
700 | return 1; | |
701 | } | |
702 | ||
703 | /* | |
704 | * If previous slice expired, start a new one otherwise renew/extend | |
705 | * existing slice to make sure it is at least throtl_slice interval | |
706 | * long since now. | |
707 | */ | |
708 | if (throtl_slice_used(td, tg, rw)) | |
709 | throtl_start_new_slice(td, tg, rw); | |
710 | else { | |
711 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
712 | throtl_extend_slice(td, tg, rw, jiffies + throtl_slice); | |
713 | } | |
714 | ||
715 | if (tg_with_in_bps_limit(td, tg, bio, &bps_wait) | |
716 | && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) { | |
717 | if (wait) | |
718 | *wait = 0; | |
719 | return 1; | |
720 | } | |
721 | ||
722 | max_wait = max(bps_wait, iops_wait); | |
723 | ||
724 | if (wait) | |
725 | *wait = max_wait; | |
726 | ||
727 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
728 | throtl_extend_slice(td, tg, rw, jiffies + max_wait); | |
e43473b7 VG |
729 | |
730 | return 0; | |
731 | } | |
732 | ||
733 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
734 | { | |
735 | bool rw = bio_data_dir(bio); | |
e5a94f56 | 736 | bool sync = rw_is_sync(bio->bi_rw); |
e43473b7 VG |
737 | |
738 | /* Charge the bio to the group */ | |
739 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 740 | tg->io_disp[rw]++; |
e43473b7 | 741 | |
e43473b7 | 742 | blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync); |
e43473b7 VG |
743 | } |
744 | ||
745 | static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg, | |
746 | struct bio *bio) | |
747 | { | |
748 | bool rw = bio_data_dir(bio); | |
749 | ||
750 | bio_list_add(&tg->bio_lists[rw], bio); | |
751 | /* Take a bio reference on tg */ | |
752 | throtl_ref_get_tg(tg); | |
753 | tg->nr_queued[rw]++; | |
754 | td->nr_queued[rw]++; | |
755 | throtl_enqueue_tg(td, tg); | |
756 | } | |
757 | ||
758 | static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg) | |
759 | { | |
760 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
761 | struct bio *bio; | |
762 | ||
763 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
764 | tg_may_dispatch(td, tg, bio, &read_wait); | |
765 | ||
766 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
767 | tg_may_dispatch(td, tg, bio, &write_wait); | |
768 | ||
769 | min_wait = min(read_wait, write_wait); | |
770 | disptime = jiffies + min_wait; | |
771 | ||
e43473b7 VG |
772 | /* Update dispatch time */ |
773 | throtl_dequeue_tg(td, tg); | |
774 | tg->disptime = disptime; | |
775 | throtl_enqueue_tg(td, tg); | |
776 | } | |
777 | ||
778 | static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg, | |
779 | bool rw, struct bio_list *bl) | |
780 | { | |
781 | struct bio *bio; | |
782 | ||
783 | bio = bio_list_pop(&tg->bio_lists[rw]); | |
784 | tg->nr_queued[rw]--; | |
785 | /* Drop bio reference on tg */ | |
786 | throtl_put_tg(tg); | |
787 | ||
788 | BUG_ON(td->nr_queued[rw] <= 0); | |
789 | td->nr_queued[rw]--; | |
790 | ||
791 | throtl_charge_bio(tg, bio); | |
792 | bio_list_add(bl, bio); | |
793 | bio->bi_rw |= REQ_THROTTLED; | |
794 | ||
795 | throtl_trim_slice(td, tg, rw); | |
796 | } | |
797 | ||
798 | static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg, | |
799 | struct bio_list *bl) | |
800 | { | |
801 | unsigned int nr_reads = 0, nr_writes = 0; | |
802 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 803 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
804 | struct bio *bio; |
805 | ||
806 | /* Try to dispatch 75% READS and 25% WRITES */ | |
807 | ||
808 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) | |
809 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
810 | ||
811 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
812 | nr_reads++; | |
813 | ||
814 | if (nr_reads >= max_nr_reads) | |
815 | break; | |
816 | } | |
817 | ||
818 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) | |
819 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
820 | ||
821 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
822 | nr_writes++; | |
823 | ||
824 | if (nr_writes >= max_nr_writes) | |
825 | break; | |
826 | } | |
827 | ||
828 | return nr_reads + nr_writes; | |
829 | } | |
830 | ||
831 | static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl) | |
832 | { | |
833 | unsigned int nr_disp = 0; | |
834 | struct throtl_grp *tg; | |
835 | struct throtl_rb_root *st = &td->tg_service_tree; | |
836 | ||
837 | while (1) { | |
838 | tg = throtl_rb_first(st); | |
839 | ||
840 | if (!tg) | |
841 | break; | |
842 | ||
843 | if (time_before(jiffies, tg->disptime)) | |
844 | break; | |
845 | ||
846 | throtl_dequeue_tg(td, tg); | |
847 | ||
848 | nr_disp += throtl_dispatch_tg(td, tg, bl); | |
849 | ||
850 | if (tg->nr_queued[0] || tg->nr_queued[1]) { | |
851 | tg_update_disptime(td, tg); | |
852 | throtl_enqueue_tg(td, tg); | |
853 | } | |
854 | ||
855 | if (nr_disp >= throtl_quantum) | |
856 | break; | |
857 | } | |
858 | ||
859 | return nr_disp; | |
860 | } | |
861 | ||
fe071437 VG |
862 | static void throtl_process_limit_change(struct throtl_data *td) |
863 | { | |
864 | struct throtl_grp *tg; | |
865 | struct hlist_node *pos, *n; | |
866 | ||
de701c74 | 867 | if (!td->limits_changed) |
fe071437 VG |
868 | return; |
869 | ||
de701c74 | 870 | xchg(&td->limits_changed, false); |
fe071437 | 871 | |
de701c74 | 872 | throtl_log(td, "limits changed"); |
fe071437 | 873 | |
04a6b516 | 874 | hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) { |
de701c74 VG |
875 | if (!tg->limits_changed) |
876 | continue; | |
877 | ||
878 | if (!xchg(&tg->limits_changed, false)) | |
879 | continue; | |
880 | ||
881 | throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu" | |
882 | " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE], | |
883 | tg->iops[READ], tg->iops[WRITE]); | |
884 | ||
04521db0 VG |
885 | /* |
886 | * Restart the slices for both READ and WRITES. It | |
887 | * might happen that a group's limit are dropped | |
888 | * suddenly and we don't want to account recently | |
889 | * dispatched IO with new low rate | |
890 | */ | |
891 | throtl_start_new_slice(td, tg, 0); | |
892 | throtl_start_new_slice(td, tg, 1); | |
893 | ||
de701c74 | 894 | if (throtl_tg_on_rr(tg)) |
fe071437 | 895 | tg_update_disptime(td, tg); |
fe071437 | 896 | } |
fe071437 VG |
897 | } |
898 | ||
e43473b7 VG |
899 | /* Dispatch throttled bios. Should be called without queue lock held. */ |
900 | static int throtl_dispatch(struct request_queue *q) | |
901 | { | |
902 | struct throtl_data *td = q->td; | |
903 | unsigned int nr_disp = 0; | |
904 | struct bio_list bio_list_on_stack; | |
905 | struct bio *bio; | |
69d60eb9 | 906 | struct blk_plug plug; |
e43473b7 VG |
907 | |
908 | spin_lock_irq(q->queue_lock); | |
909 | ||
fe071437 VG |
910 | throtl_process_limit_change(td); |
911 | ||
e43473b7 VG |
912 | if (!total_nr_queued(td)) |
913 | goto out; | |
914 | ||
915 | bio_list_init(&bio_list_on_stack); | |
916 | ||
d2f31a5f | 917 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
e43473b7 VG |
918 | total_nr_queued(td), td->nr_queued[READ], |
919 | td->nr_queued[WRITE]); | |
920 | ||
921 | nr_disp = throtl_select_dispatch(td, &bio_list_on_stack); | |
922 | ||
923 | if (nr_disp) | |
924 | throtl_log(td, "bios disp=%u", nr_disp); | |
925 | ||
926 | throtl_schedule_next_dispatch(td); | |
927 | out: | |
928 | spin_unlock_irq(q->queue_lock); | |
929 | ||
930 | /* | |
931 | * If we dispatched some requests, unplug the queue to make sure | |
932 | * immediate dispatch | |
933 | */ | |
934 | if (nr_disp) { | |
69d60eb9 | 935 | blk_start_plug(&plug); |
e43473b7 VG |
936 | while((bio = bio_list_pop(&bio_list_on_stack))) |
937 | generic_make_request(bio); | |
69d60eb9 | 938 | blk_finish_plug(&plug); |
e43473b7 VG |
939 | } |
940 | return nr_disp; | |
941 | } | |
942 | ||
943 | void blk_throtl_work(struct work_struct *work) | |
944 | { | |
945 | struct throtl_data *td = container_of(work, struct throtl_data, | |
946 | throtl_work.work); | |
947 | struct request_queue *q = td->queue; | |
948 | ||
949 | throtl_dispatch(q); | |
950 | } | |
951 | ||
952 | /* Call with queue lock held */ | |
450adcbe VG |
953 | static void |
954 | throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay) | |
e43473b7 VG |
955 | { |
956 | ||
e43473b7 VG |
957 | struct delayed_work *dwork = &td->throtl_work; |
958 | ||
04521db0 | 959 | /* schedule work if limits changed even if no bio is queued */ |
d2f31a5f | 960 | if (total_nr_queued(td) || td->limits_changed) { |
e43473b7 VG |
961 | /* |
962 | * We might have a work scheduled to be executed in future. | |
963 | * Cancel that and schedule a new one. | |
964 | */ | |
965 | __cancel_delayed_work(dwork); | |
450adcbe | 966 | queue_delayed_work(kthrotld_workqueue, dwork, delay); |
e43473b7 VG |
967 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", |
968 | delay, jiffies); | |
969 | } | |
970 | } | |
e43473b7 VG |
971 | |
972 | static void | |
973 | throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg) | |
974 | { | |
975 | /* Something wrong if we are trying to remove same group twice */ | |
976 | BUG_ON(hlist_unhashed(&tg->tg_node)); | |
977 | ||
978 | hlist_del_init(&tg->tg_node); | |
979 | ||
980 | /* | |
981 | * Put the reference taken at the time of creation so that when all | |
982 | * queues are gone, group can be destroyed. | |
983 | */ | |
984 | throtl_put_tg(tg); | |
985 | td->nr_undestroyed_grps--; | |
986 | } | |
987 | ||
988 | static void throtl_release_tgs(struct throtl_data *td) | |
989 | { | |
990 | struct hlist_node *pos, *n; | |
991 | struct throtl_grp *tg; | |
992 | ||
993 | hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) { | |
994 | /* | |
995 | * If cgroup removal path got to blk_group first and removed | |
996 | * it from cgroup list, then it will take care of destroying | |
997 | * cfqg also. | |
998 | */ | |
999 | if (!blkiocg_del_blkio_group(&tg->blkg)) | |
1000 | throtl_destroy_tg(td, tg); | |
1001 | } | |
1002 | } | |
1003 | ||
1004 | static void throtl_td_free(struct throtl_data *td) | |
1005 | { | |
1006 | kfree(td); | |
1007 | } | |
1008 | ||
1009 | /* | |
1010 | * Blk cgroup controller notification saying that blkio_group object is being | |
1011 | * delinked as associated cgroup object is going away. That also means that | |
1012 | * no new IO will come in this group. So get rid of this group as soon as | |
1013 | * any pending IO in the group is finished. | |
1014 | * | |
1015 | * This function is called under rcu_read_lock(). key is the rcu protected | |
1016 | * pointer. That means "key" is a valid throtl_data pointer as long as we are | |
1017 | * rcu read lock. | |
1018 | * | |
1019 | * "key" was fetched from blkio_group under blkio_cgroup->lock. That means | |
1020 | * it should not be NULL as even if queue was going away, cgroup deltion | |
1021 | * path got to it first. | |
1022 | */ | |
1023 | void throtl_unlink_blkio_group(void *key, struct blkio_group *blkg) | |
1024 | { | |
1025 | unsigned long flags; | |
1026 | struct throtl_data *td = key; | |
1027 | ||
1028 | spin_lock_irqsave(td->queue->queue_lock, flags); | |
1029 | throtl_destroy_tg(td, tg_of_blkg(blkg)); | |
1030 | spin_unlock_irqrestore(td->queue->queue_lock, flags); | |
1031 | } | |
1032 | ||
de701c74 VG |
1033 | static void throtl_update_blkio_group_common(struct throtl_data *td, |
1034 | struct throtl_grp *tg) | |
1035 | { | |
1036 | xchg(&tg->limits_changed, true); | |
1037 | xchg(&td->limits_changed, true); | |
1038 | /* Schedule a work now to process the limit change */ | |
1039 | throtl_schedule_delayed_work(td, 0); | |
1040 | } | |
1041 | ||
fe071437 VG |
1042 | /* |
1043 | * For all update functions, key should be a valid pointer because these | |
1044 | * update functions are called under blkcg_lock, that means, blkg is | |
25985edc | 1045 | * valid and in turn key is valid. queue exit path can not race because |
fe071437 VG |
1046 | * of blkcg_lock |
1047 | * | |
1048 | * Can not take queue lock in update functions as queue lock under blkcg_lock | |
1049 | * is not allowed. Under other paths we take blkcg_lock under queue_lock. | |
1050 | */ | |
1051 | static void throtl_update_blkio_group_read_bps(void *key, | |
1052 | struct blkio_group *blkg, u64 read_bps) | |
e43473b7 | 1053 | { |
fe071437 | 1054 | struct throtl_data *td = key; |
de701c74 | 1055 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1056 | |
de701c74 VG |
1057 | tg->bps[READ] = read_bps; |
1058 | throtl_update_blkio_group_common(td, tg); | |
e43473b7 VG |
1059 | } |
1060 | ||
fe071437 VG |
1061 | static void throtl_update_blkio_group_write_bps(void *key, |
1062 | struct blkio_group *blkg, u64 write_bps) | |
e43473b7 | 1063 | { |
fe071437 | 1064 | struct throtl_data *td = key; |
de701c74 | 1065 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1066 | |
de701c74 VG |
1067 | tg->bps[WRITE] = write_bps; |
1068 | throtl_update_blkio_group_common(td, tg); | |
e43473b7 VG |
1069 | } |
1070 | ||
fe071437 VG |
1071 | static void throtl_update_blkio_group_read_iops(void *key, |
1072 | struct blkio_group *blkg, unsigned int read_iops) | |
8e89d13f | 1073 | { |
fe071437 | 1074 | struct throtl_data *td = key; |
de701c74 | 1075 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1076 | |
de701c74 VG |
1077 | tg->iops[READ] = read_iops; |
1078 | throtl_update_blkio_group_common(td, tg); | |
8e89d13f VG |
1079 | } |
1080 | ||
fe071437 VG |
1081 | static void throtl_update_blkio_group_write_iops(void *key, |
1082 | struct blkio_group *blkg, unsigned int write_iops) | |
8e89d13f | 1083 | { |
fe071437 | 1084 | struct throtl_data *td = key; |
de701c74 | 1085 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1086 | |
de701c74 VG |
1087 | tg->iops[WRITE] = write_iops; |
1088 | throtl_update_blkio_group_common(td, tg); | |
8e89d13f VG |
1089 | } |
1090 | ||
da527770 | 1091 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1092 | { |
1093 | struct throtl_data *td = q->td; | |
1094 | ||
1095 | cancel_delayed_work_sync(&td->throtl_work); | |
1096 | } | |
1097 | ||
1098 | static struct blkio_policy_type blkio_policy_throtl = { | |
1099 | .ops = { | |
1100 | .blkio_unlink_group_fn = throtl_unlink_blkio_group, | |
1101 | .blkio_update_group_read_bps_fn = | |
1102 | throtl_update_blkio_group_read_bps, | |
1103 | .blkio_update_group_write_bps_fn = | |
1104 | throtl_update_blkio_group_write_bps, | |
8e89d13f VG |
1105 | .blkio_update_group_read_iops_fn = |
1106 | throtl_update_blkio_group_read_iops, | |
1107 | .blkio_update_group_write_iops_fn = | |
1108 | throtl_update_blkio_group_write_iops, | |
e43473b7 | 1109 | }, |
8e89d13f | 1110 | .plid = BLKIO_POLICY_THROTL, |
e43473b7 VG |
1111 | }; |
1112 | ||
bc16a4f9 | 1113 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1114 | { |
1115 | struct throtl_data *td = q->td; | |
1116 | struct throtl_grp *tg; | |
e43473b7 | 1117 | bool rw = bio_data_dir(bio), update_disptime = true; |
af75cd3c | 1118 | struct blkio_cgroup *blkcg; |
bc16a4f9 | 1119 | bool throttled = false; |
e43473b7 VG |
1120 | |
1121 | if (bio->bi_rw & REQ_THROTTLED) { | |
1122 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1123 | goto out; |
e43473b7 VG |
1124 | } |
1125 | ||
af75cd3c VG |
1126 | /* |
1127 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1128 | * basic fields like stats and io rates. If a group has no rules, | |
1129 | * just update the dispatch stats in lockless manner and return. | |
1130 | */ | |
1131 | ||
1132 | rcu_read_lock(); | |
1133 | blkcg = task_blkio_cgroup(current); | |
1134 | tg = throtl_find_tg(td, blkcg); | |
1135 | if (tg) { | |
1136 | throtl_tg_fill_dev_details(td, tg); | |
1137 | ||
1138 | if (tg_no_rule_group(tg, rw)) { | |
1139 | blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, | |
e5a94f56 | 1140 | rw, rw_is_sync(bio->bi_rw)); |
af75cd3c | 1141 | rcu_read_unlock(); |
bc16a4f9 | 1142 | goto out; |
af75cd3c VG |
1143 | } |
1144 | } | |
1145 | rcu_read_unlock(); | |
1146 | ||
1147 | /* | |
1148 | * Either group has not been allocated yet or it is not an unlimited | |
1149 | * IO group | |
1150 | */ | |
e43473b7 VG |
1151 | spin_lock_irq(q->queue_lock); |
1152 | tg = throtl_get_tg(td); | |
bc16a4f9 TH |
1153 | if (unlikely(!tg)) |
1154 | goto out_unlock; | |
f469a7b4 | 1155 | |
e43473b7 VG |
1156 | if (tg->nr_queued[rw]) { |
1157 | /* | |
1158 | * There is already another bio queued in same dir. No | |
1159 | * need to update dispatch time. | |
1160 | */ | |
231d704b | 1161 | update_disptime = false; |
e43473b7 | 1162 | goto queue_bio; |
de701c74 | 1163 | |
e43473b7 VG |
1164 | } |
1165 | ||
1166 | /* Bio is with-in rate limit of group */ | |
1167 | if (tg_may_dispatch(td, tg, bio, NULL)) { | |
1168 | throtl_charge_bio(tg, bio); | |
04521db0 VG |
1169 | |
1170 | /* | |
1171 | * We need to trim slice even when bios are not being queued | |
1172 | * otherwise it might happen that a bio is not queued for | |
1173 | * a long time and slice keeps on extending and trim is not | |
1174 | * called for a long time. Now if limits are reduced suddenly | |
1175 | * we take into account all the IO dispatched so far at new | |
1176 | * low rate and * newly queued IO gets a really long dispatch | |
1177 | * time. | |
1178 | * | |
1179 | * So keep on trimming slice even if bio is not queued. | |
1180 | */ | |
1181 | throtl_trim_slice(td, tg, rw); | |
bc16a4f9 | 1182 | goto out_unlock; |
e43473b7 VG |
1183 | } |
1184 | ||
1185 | queue_bio: | |
fd16d263 | 1186 | throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
8e89d13f VG |
1187 | " iodisp=%u iops=%u queued=%d/%d", |
1188 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1189 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1190 | tg->io_disp[rw], tg->iops[rw], |
e43473b7 VG |
1191 | tg->nr_queued[READ], tg->nr_queued[WRITE]); |
1192 | ||
1193 | throtl_add_bio_tg(q->td, tg, bio); | |
bc16a4f9 | 1194 | throttled = true; |
e43473b7 VG |
1195 | |
1196 | if (update_disptime) { | |
1197 | tg_update_disptime(td, tg); | |
1198 | throtl_schedule_next_dispatch(td); | |
1199 | } | |
1200 | ||
bc16a4f9 | 1201 | out_unlock: |
e43473b7 | 1202 | spin_unlock_irq(q->queue_lock); |
bc16a4f9 TH |
1203 | out: |
1204 | return throttled; | |
e43473b7 VG |
1205 | } |
1206 | ||
1207 | int blk_throtl_init(struct request_queue *q) | |
1208 | { | |
1209 | struct throtl_data *td; | |
1210 | struct throtl_grp *tg; | |
1211 | ||
1212 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1213 | if (!td) | |
1214 | return -ENOMEM; | |
1215 | ||
1216 | INIT_HLIST_HEAD(&td->tg_list); | |
1217 | td->tg_service_tree = THROTL_RB_ROOT; | |
de701c74 | 1218 | td->limits_changed = false; |
a29a171e | 1219 | INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work); |
e43473b7 | 1220 | |
29b12589 VG |
1221 | /* alloc and Init root group. */ |
1222 | td->queue = q; | |
1223 | tg = throtl_alloc_tg(td); | |
02977e4a | 1224 | |
29b12589 VG |
1225 | if (!tg) { |
1226 | kfree(td); | |
1227 | return -ENOMEM; | |
1228 | } | |
1229 | ||
1230 | td->root_tg = tg; | |
e43473b7 VG |
1231 | |
1232 | rcu_read_lock(); | |
5617cbef | 1233 | throtl_init_add_tg_lists(td, tg, &blkio_root_cgroup); |
e43473b7 VG |
1234 | rcu_read_unlock(); |
1235 | ||
1236 | /* Attach throtl data to request queue */ | |
e43473b7 VG |
1237 | q->td = td; |
1238 | return 0; | |
1239 | } | |
1240 | ||
1241 | void blk_throtl_exit(struct request_queue *q) | |
1242 | { | |
1243 | struct throtl_data *td = q->td; | |
1244 | bool wait = false; | |
1245 | ||
1246 | BUG_ON(!td); | |
1247 | ||
da527770 | 1248 | throtl_shutdown_wq(q); |
e43473b7 VG |
1249 | |
1250 | spin_lock_irq(q->queue_lock); | |
1251 | throtl_release_tgs(td); | |
e43473b7 VG |
1252 | |
1253 | /* If there are other groups */ | |
02977e4a | 1254 | if (td->nr_undestroyed_grps > 0) |
e43473b7 VG |
1255 | wait = true; |
1256 | ||
1257 | spin_unlock_irq(q->queue_lock); | |
1258 | ||
1259 | /* | |
1260 | * Wait for tg->blkg->key accessors to exit their grace periods. | |
1261 | * Do this wait only if there are other undestroyed groups out | |
1262 | * there (other than root group). This can happen if cgroup deletion | |
1263 | * path claimed the responsibility of cleaning up a group before | |
1264 | * queue cleanup code get to the group. | |
1265 | * | |
1266 | * Do not call synchronize_rcu() unconditionally as there are drivers | |
1267 | * which create/delete request queue hundreds of times during scan/boot | |
1268 | * and synchronize_rcu() can take significant time and slow down boot. | |
1269 | */ | |
1270 | if (wait) | |
1271 | synchronize_rcu(); | |
fe071437 VG |
1272 | |
1273 | /* | |
1274 | * Just being safe to make sure after previous flush if some body did | |
1275 | * update limits through cgroup and another work got queued, cancel | |
1276 | * it. | |
1277 | */ | |
da527770 | 1278 | throtl_shutdown_wq(q); |
e43473b7 VG |
1279 | throtl_td_free(td); |
1280 | } | |
1281 | ||
1282 | static int __init throtl_init(void) | |
1283 | { | |
450adcbe VG |
1284 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1285 | if (!kthrotld_workqueue) | |
1286 | panic("Failed to create kthrotld\n"); | |
1287 | ||
e43473b7 VG |
1288 | blkio_policy_register(&blkio_policy_throtl); |
1289 | return 0; | |
1290 | } | |
1291 | ||
1292 | module_init(throtl_init); |