]>
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" | |
13 | ||
14 | /* Max dispatch from a group in 1 round */ | |
15 | static int throtl_grp_quantum = 8; | |
16 | ||
17 | /* Total max dispatch from all groups in one round */ | |
18 | static int throtl_quantum = 32; | |
19 | ||
20 | /* Throttling is performed over 100ms slice and after that slice is renewed */ | |
21 | static unsigned long throtl_slice = HZ/10; /* 100 ms */ | |
22 | ||
450adcbe VG |
23 | /* A workqueue to queue throttle related work */ |
24 | static struct workqueue_struct *kthrotld_workqueue; | |
25 | static void throtl_schedule_delayed_work(struct throtl_data *td, | |
26 | unsigned long delay); | |
27 | ||
e43473b7 VG |
28 | struct throtl_rb_root { |
29 | struct rb_root rb; | |
30 | struct rb_node *left; | |
31 | unsigned int count; | |
32 | unsigned long min_disptime; | |
33 | }; | |
34 | ||
35 | #define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \ | |
36 | .count = 0, .min_disptime = 0} | |
37 | ||
38 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) | |
39 | ||
40 | struct throtl_grp { | |
41 | /* List of throtl groups on the request queue*/ | |
42 | struct hlist_node tg_node; | |
43 | ||
44 | /* active throtl group service_tree member */ | |
45 | struct rb_node rb_node; | |
46 | ||
47 | /* | |
48 | * Dispatch time in jiffies. This is the estimated time when group | |
49 | * will unthrottle and is ready to dispatch more bio. It is used as | |
50 | * key to sort active groups in service tree. | |
51 | */ | |
52 | unsigned long disptime; | |
53 | ||
54 | struct blkio_group blkg; | |
55 | atomic_t ref; | |
56 | unsigned int flags; | |
57 | ||
58 | /* Two lists for READ and WRITE */ | |
59 | struct bio_list bio_lists[2]; | |
60 | ||
61 | /* Number of queued bios on READ and WRITE lists */ | |
62 | unsigned int nr_queued[2]; | |
63 | ||
64 | /* bytes per second rate limits */ | |
65 | uint64_t bps[2]; | |
66 | ||
8e89d13f VG |
67 | /* IOPS limits */ |
68 | unsigned int iops[2]; | |
69 | ||
e43473b7 VG |
70 | /* Number of bytes disptached in current slice */ |
71 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
72 | /* Number of bio's dispatched in current slice */ |
73 | unsigned int io_disp[2]; | |
e43473b7 VG |
74 | |
75 | /* When did we start a new slice */ | |
76 | unsigned long slice_start[2]; | |
77 | unsigned long slice_end[2]; | |
fe071437 VG |
78 | |
79 | /* Some throttle limits got updated for the group */ | |
6f037937 | 80 | int limits_changed; |
e43473b7 VG |
81 | }; |
82 | ||
83 | struct throtl_data | |
84 | { | |
85 | /* List of throtl groups */ | |
86 | struct hlist_head tg_list; | |
87 | ||
88 | /* service tree for active throtl groups */ | |
89 | struct throtl_rb_root tg_service_tree; | |
90 | ||
91 | struct throtl_grp root_tg; | |
92 | struct request_queue *queue; | |
93 | ||
94 | /* Total Number of queued bios on READ and WRITE lists */ | |
95 | unsigned int nr_queued[2]; | |
96 | ||
97 | /* | |
02977e4a | 98 | * number of total undestroyed groups |
e43473b7 VG |
99 | */ |
100 | unsigned int nr_undestroyed_grps; | |
101 | ||
102 | /* Work for dispatching throttled bios */ | |
103 | struct delayed_work throtl_work; | |
fe071437 | 104 | |
6f037937 | 105 | int limits_changed; |
e43473b7 VG |
106 | }; |
107 | ||
108 | enum tg_state_flags { | |
109 | THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */ | |
110 | }; | |
111 | ||
112 | #define THROTL_TG_FNS(name) \ | |
113 | static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \ | |
114 | { \ | |
115 | (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \ | |
116 | } \ | |
117 | static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \ | |
118 | { \ | |
119 | (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \ | |
120 | } \ | |
121 | static inline int throtl_tg_##name(const struct throtl_grp *tg) \ | |
122 | { \ | |
123 | return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \ | |
124 | } | |
125 | ||
126 | THROTL_TG_FNS(on_rr); | |
127 | ||
128 | #define throtl_log_tg(td, tg, fmt, args...) \ | |
129 | blk_add_trace_msg((td)->queue, "throtl %s " fmt, \ | |
130 | blkg_path(&(tg)->blkg), ##args); \ | |
131 | ||
132 | #define throtl_log(td, fmt, args...) \ | |
133 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) | |
134 | ||
135 | static inline struct throtl_grp *tg_of_blkg(struct blkio_group *blkg) | |
136 | { | |
137 | if (blkg) | |
138 | return container_of(blkg, struct throtl_grp, blkg); | |
139 | ||
140 | return NULL; | |
141 | } | |
142 | ||
143 | static inline int total_nr_queued(struct throtl_data *td) | |
144 | { | |
145 | return (td->nr_queued[0] + td->nr_queued[1]); | |
146 | } | |
147 | ||
148 | static inline struct throtl_grp *throtl_ref_get_tg(struct throtl_grp *tg) | |
149 | { | |
150 | atomic_inc(&tg->ref); | |
151 | return tg; | |
152 | } | |
153 | ||
154 | static void throtl_put_tg(struct throtl_grp *tg) | |
155 | { | |
156 | BUG_ON(atomic_read(&tg->ref) <= 0); | |
157 | if (!atomic_dec_and_test(&tg->ref)) | |
158 | return; | |
159 | kfree(tg); | |
160 | } | |
161 | ||
a29a171e VG |
162 | static void throtl_init_group(struct throtl_grp *tg) |
163 | { | |
164 | INIT_HLIST_NODE(&tg->tg_node); | |
165 | RB_CLEAR_NODE(&tg->rb_node); | |
166 | bio_list_init(&tg->bio_lists[0]); | |
167 | bio_list_init(&tg->bio_lists[1]); | |
168 | tg->limits_changed = false; | |
169 | ||
170 | /* Practically unlimited BW */ | |
171 | tg->bps[0] = tg->bps[1] = -1; | |
172 | tg->iops[0] = tg->iops[1] = -1; | |
173 | ||
174 | /* | |
175 | * Take the initial reference that will be released on destroy | |
176 | * This can be thought of a joint reference by cgroup and | |
177 | * request queue which will be dropped by either request queue | |
178 | * exit or cgroup deletion path depending on who is exiting first. | |
179 | */ | |
180 | atomic_set(&tg->ref, 1); | |
181 | } | |
182 | ||
183 | /* Should be called with rcu read lock held (needed for blkcg) */ | |
184 | static void | |
185 | throtl_add_group_to_td_list(struct throtl_data *td, struct throtl_grp *tg) | |
186 | { | |
187 | hlist_add_head(&tg->tg_node, &td->tg_list); | |
188 | td->nr_undestroyed_grps++; | |
189 | } | |
190 | ||
f469a7b4 VG |
191 | static void throtl_init_add_tg_lists(struct throtl_data *td, |
192 | struct throtl_grp *tg, struct blkio_cgroup *blkcg) | |
193 | { | |
194 | struct backing_dev_info *bdi = &td->queue->backing_dev_info; | |
195 | unsigned int major, minor; | |
196 | ||
197 | /* Add group onto cgroup list */ | |
198 | sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); | |
199 | blkiocg_add_blkio_group(blkcg, &tg->blkg, (void *)td, | |
200 | MKDEV(major, minor), BLKIO_POLICY_THROTL); | |
201 | ||
202 | tg->bps[READ] = blkcg_get_read_bps(blkcg, tg->blkg.dev); | |
203 | tg->bps[WRITE] = blkcg_get_write_bps(blkcg, tg->blkg.dev); | |
204 | tg->iops[READ] = blkcg_get_read_iops(blkcg, tg->blkg.dev); | |
205 | tg->iops[WRITE] = blkcg_get_write_iops(blkcg, tg->blkg.dev); | |
206 | ||
207 | throtl_add_group_to_td_list(td, tg); | |
208 | } | |
209 | ||
210 | /* Should be called without queue lock and outside of rcu period */ | |
211 | static struct throtl_grp *throtl_alloc_tg(struct throtl_data *td) | |
212 | { | |
213 | struct throtl_grp *tg = NULL; | |
214 | ||
215 | tg = kzalloc_node(sizeof(*tg), GFP_ATOMIC, td->queue->node); | |
216 | if (!tg) | |
217 | return NULL; | |
218 | ||
219 | throtl_init_group(tg); | |
220 | return tg; | |
221 | } | |
222 | ||
223 | static struct | |
224 | throtl_grp *throtl_find_tg(struct throtl_data *td, struct blkio_cgroup *blkcg) | |
e43473b7 | 225 | { |
e43473b7 VG |
226 | struct throtl_grp *tg = NULL; |
227 | void *key = td; | |
228 | struct backing_dev_info *bdi = &td->queue->backing_dev_info; | |
229 | unsigned int major, minor; | |
230 | ||
be2c6b19 VG |
231 | /* |
232 | * This is the common case when there are no blkio cgroups. | |
233 | * Avoid lookup in this case | |
234 | */ | |
235 | if (blkcg == &blkio_root_cgroup) | |
236 | tg = &td->root_tg; | |
237 | else | |
238 | tg = tg_of_blkg(blkiocg_lookup_group(blkcg, key)); | |
e43473b7 VG |
239 | |
240 | /* Fill in device details for root group */ | |
241 | if (tg && !tg->blkg.dev && bdi->dev && dev_name(bdi->dev)) { | |
242 | sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); | |
243 | tg->blkg.dev = MKDEV(major, minor); | |
e43473b7 VG |
244 | } |
245 | ||
e43473b7 VG |
246 | return tg; |
247 | } | |
248 | ||
f469a7b4 VG |
249 | /* |
250 | * This function returns with queue lock unlocked in case of error, like | |
251 | * request queue is no more | |
252 | */ | |
e43473b7 VG |
253 | static struct throtl_grp * throtl_get_tg(struct throtl_data *td) |
254 | { | |
f469a7b4 | 255 | struct throtl_grp *tg = NULL, *__tg = NULL; |
70087dc3 | 256 | struct blkio_cgroup *blkcg; |
f469a7b4 | 257 | struct request_queue *q = td->queue; |
e43473b7 VG |
258 | |
259 | rcu_read_lock(); | |
70087dc3 | 260 | blkcg = task_blkio_cgroup(current); |
f469a7b4 VG |
261 | tg = throtl_find_tg(td, blkcg); |
262 | if (tg) { | |
263 | rcu_read_unlock(); | |
264 | return tg; | |
265 | } | |
266 | ||
267 | /* | |
268 | * Need to allocate a group. Allocation of group also needs allocation | |
269 | * of per cpu stats which in-turn takes a mutex() and can block. Hence | |
270 | * we need to drop rcu lock and queue_lock before we call alloc | |
271 | * | |
272 | * Take the request queue reference to make sure queue does not | |
273 | * go away once we return from allocation. | |
274 | */ | |
275 | blk_get_queue(q); | |
276 | rcu_read_unlock(); | |
277 | spin_unlock_irq(q->queue_lock); | |
278 | ||
279 | tg = throtl_alloc_tg(td); | |
280 | /* | |
281 | * We might have slept in group allocation. Make sure queue is not | |
282 | * dead | |
283 | */ | |
284 | if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) { | |
285 | blk_put_queue(q); | |
286 | if (tg) | |
287 | kfree(tg); | |
288 | ||
289 | return ERR_PTR(-ENODEV); | |
290 | } | |
291 | blk_put_queue(q); | |
292 | ||
293 | /* Group allocated and queue is still alive. take the lock */ | |
294 | spin_lock_irq(q->queue_lock); | |
295 | ||
296 | /* | |
297 | * Initialize the new group. After sleeping, read the blkcg again. | |
298 | */ | |
299 | rcu_read_lock(); | |
300 | blkcg = task_blkio_cgroup(current); | |
301 | ||
302 | /* | |
303 | * If some other thread already allocated the group while we were | |
304 | * not holding queue lock, free up the group | |
305 | */ | |
306 | __tg = throtl_find_tg(td, blkcg); | |
307 | ||
308 | if (__tg) { | |
309 | kfree(tg); | |
310 | rcu_read_unlock(); | |
311 | return __tg; | |
312 | } | |
313 | ||
314 | /* Group allocation failed. Account the IO to root group */ | |
315 | if (!tg) { | |
e43473b7 | 316 | tg = &td->root_tg; |
f469a7b4 VG |
317 | return tg; |
318 | } | |
319 | ||
320 | throtl_init_add_tg_lists(td, tg, blkcg); | |
e43473b7 VG |
321 | rcu_read_unlock(); |
322 | return tg; | |
323 | } | |
324 | ||
325 | static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root) | |
326 | { | |
327 | /* Service tree is empty */ | |
328 | if (!root->count) | |
329 | return NULL; | |
330 | ||
331 | if (!root->left) | |
332 | root->left = rb_first(&root->rb); | |
333 | ||
334 | if (root->left) | |
335 | return rb_entry_tg(root->left); | |
336 | ||
337 | return NULL; | |
338 | } | |
339 | ||
340 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
341 | { | |
342 | rb_erase(n, root); | |
343 | RB_CLEAR_NODE(n); | |
344 | } | |
345 | ||
346 | static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root) | |
347 | { | |
348 | if (root->left == n) | |
349 | root->left = NULL; | |
350 | rb_erase_init(n, &root->rb); | |
351 | --root->count; | |
352 | } | |
353 | ||
354 | static void update_min_dispatch_time(struct throtl_rb_root *st) | |
355 | { | |
356 | struct throtl_grp *tg; | |
357 | ||
358 | tg = throtl_rb_first(st); | |
359 | if (!tg) | |
360 | return; | |
361 | ||
362 | st->min_disptime = tg->disptime; | |
363 | } | |
364 | ||
365 | static void | |
366 | tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg) | |
367 | { | |
368 | struct rb_node **node = &st->rb.rb_node; | |
369 | struct rb_node *parent = NULL; | |
370 | struct throtl_grp *__tg; | |
371 | unsigned long key = tg->disptime; | |
372 | int left = 1; | |
373 | ||
374 | while (*node != NULL) { | |
375 | parent = *node; | |
376 | __tg = rb_entry_tg(parent); | |
377 | ||
378 | if (time_before(key, __tg->disptime)) | |
379 | node = &parent->rb_left; | |
380 | else { | |
381 | node = &parent->rb_right; | |
382 | left = 0; | |
383 | } | |
384 | } | |
385 | ||
386 | if (left) | |
387 | st->left = &tg->rb_node; | |
388 | ||
389 | rb_link_node(&tg->rb_node, parent, node); | |
390 | rb_insert_color(&tg->rb_node, &st->rb); | |
391 | } | |
392 | ||
393 | static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
394 | { | |
395 | struct throtl_rb_root *st = &td->tg_service_tree; | |
396 | ||
397 | tg_service_tree_add(st, tg); | |
398 | throtl_mark_tg_on_rr(tg); | |
399 | st->count++; | |
400 | } | |
401 | ||
402 | static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
403 | { | |
404 | if (!throtl_tg_on_rr(tg)) | |
405 | __throtl_enqueue_tg(td, tg); | |
406 | } | |
407 | ||
408 | static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
409 | { | |
410 | throtl_rb_erase(&tg->rb_node, &td->tg_service_tree); | |
411 | throtl_clear_tg_on_rr(tg); | |
412 | } | |
413 | ||
414 | static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
415 | { | |
416 | if (throtl_tg_on_rr(tg)) | |
417 | __throtl_dequeue_tg(td, tg); | |
418 | } | |
419 | ||
420 | static void throtl_schedule_next_dispatch(struct throtl_data *td) | |
421 | { | |
422 | struct throtl_rb_root *st = &td->tg_service_tree; | |
423 | ||
424 | /* | |
425 | * If there are more bios pending, schedule more work. | |
426 | */ | |
427 | if (!total_nr_queued(td)) | |
428 | return; | |
429 | ||
430 | BUG_ON(!st->count); | |
431 | ||
432 | update_min_dispatch_time(st); | |
433 | ||
434 | if (time_before_eq(st->min_disptime, jiffies)) | |
450adcbe | 435 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 436 | else |
450adcbe | 437 | throtl_schedule_delayed_work(td, (st->min_disptime - jiffies)); |
e43473b7 VG |
438 | } |
439 | ||
440 | static inline void | |
441 | throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
442 | { | |
443 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 444 | tg->io_disp[rw] = 0; |
e43473b7 VG |
445 | tg->slice_start[rw] = jiffies; |
446 | tg->slice_end[rw] = jiffies + throtl_slice; | |
447 | throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
448 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
449 | tg->slice_end[rw], jiffies); | |
450 | } | |
451 | ||
d1ae8ffd VG |
452 | static inline void throtl_set_slice_end(struct throtl_data *td, |
453 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
454 | { | |
455 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
456 | } | |
457 | ||
e43473b7 VG |
458 | static inline void throtl_extend_slice(struct throtl_data *td, |
459 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
460 | { | |
461 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
462 | throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
463 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
464 | tg->slice_end[rw], jiffies); | |
465 | } | |
466 | ||
467 | /* Determine if previously allocated or extended slice is complete or not */ | |
468 | static bool | |
469 | throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
470 | { | |
471 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
472 | return 0; | |
473 | ||
474 | return 1; | |
475 | } | |
476 | ||
477 | /* Trim the used slices and adjust slice start accordingly */ | |
478 | static inline void | |
479 | throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
480 | { | |
3aad5d3e VG |
481 | unsigned long nr_slices, time_elapsed, io_trim; |
482 | u64 bytes_trim, tmp; | |
e43473b7 VG |
483 | |
484 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
485 | ||
486 | /* | |
487 | * If bps are unlimited (-1), then time slice don't get | |
488 | * renewed. Don't try to trim the slice if slice is used. A new | |
489 | * slice will start when appropriate. | |
490 | */ | |
491 | if (throtl_slice_used(td, tg, rw)) | |
492 | return; | |
493 | ||
d1ae8ffd VG |
494 | /* |
495 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
496 | * that initially cgroup limit was very low resulting in high | |
497 | * slice_end, but later limit was bumped up and bio was dispached | |
498 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
499 | * is bad because it does not allow new slice to start. | |
500 | */ | |
501 | ||
502 | throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice); | |
503 | ||
e43473b7 VG |
504 | time_elapsed = jiffies - tg->slice_start[rw]; |
505 | ||
506 | nr_slices = time_elapsed / throtl_slice; | |
507 | ||
508 | if (!nr_slices) | |
509 | return; | |
3aad5d3e VG |
510 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
511 | do_div(tmp, HZ); | |
512 | bytes_trim = tmp; | |
e43473b7 | 513 | |
8e89d13f | 514 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 515 | |
8e89d13f | 516 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
517 | return; |
518 | ||
519 | if (tg->bytes_disp[rw] >= bytes_trim) | |
520 | tg->bytes_disp[rw] -= bytes_trim; | |
521 | else | |
522 | tg->bytes_disp[rw] = 0; | |
523 | ||
8e89d13f VG |
524 | if (tg->io_disp[rw] >= io_trim) |
525 | tg->io_disp[rw] -= io_trim; | |
526 | else | |
527 | tg->io_disp[rw] = 0; | |
528 | ||
e43473b7 VG |
529 | tg->slice_start[rw] += nr_slices * throtl_slice; |
530 | ||
3aad5d3e | 531 | throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 532 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 533 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
534 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
535 | } | |
536 | ||
8e89d13f VG |
537 | static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg, |
538 | struct bio *bio, unsigned long *wait) | |
e43473b7 VG |
539 | { |
540 | bool rw = bio_data_dir(bio); | |
8e89d13f | 541 | unsigned int io_allowed; |
e43473b7 | 542 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 543 | u64 tmp; |
e43473b7 | 544 | |
8e89d13f | 545 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 546 | |
8e89d13f VG |
547 | /* Slice has just started. Consider one slice interval */ |
548 | if (!jiffy_elapsed) | |
549 | jiffy_elapsed_rnd = throtl_slice; | |
550 | ||
551 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
552 | ||
c49c06e4 VG |
553 | /* |
554 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
555 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
556 | * will allow dispatch after 1 second and after that slice should | |
557 | * have been trimmed. | |
558 | */ | |
559 | ||
560 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
561 | do_div(tmp, HZ); | |
562 | ||
563 | if (tmp > UINT_MAX) | |
564 | io_allowed = UINT_MAX; | |
565 | else | |
566 | io_allowed = tmp; | |
8e89d13f VG |
567 | |
568 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
569 | if (wait) |
570 | *wait = 0; | |
571 | return 1; | |
572 | } | |
573 | ||
8e89d13f VG |
574 | /* Calc approx time to dispatch */ |
575 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
576 | ||
577 | if (jiffy_wait > jiffy_elapsed) | |
578 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
579 | else | |
580 | jiffy_wait = 1; | |
581 | ||
582 | if (wait) | |
583 | *wait = jiffy_wait; | |
584 | return 0; | |
585 | } | |
586 | ||
587 | static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg, | |
588 | struct bio *bio, unsigned long *wait) | |
589 | { | |
590 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 591 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 592 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
593 | |
594 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
595 | ||
596 | /* Slice has just started. Consider one slice interval */ | |
597 | if (!jiffy_elapsed) | |
598 | jiffy_elapsed_rnd = throtl_slice; | |
599 | ||
600 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
601 | ||
5e901a2b VG |
602 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
603 | do_div(tmp, HZ); | |
3aad5d3e | 604 | bytes_allowed = tmp; |
e43473b7 VG |
605 | |
606 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
607 | if (wait) | |
608 | *wait = 0; | |
609 | return 1; | |
610 | } | |
611 | ||
612 | /* Calc approx time to dispatch */ | |
613 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
614 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
615 | ||
616 | if (!jiffy_wait) | |
617 | jiffy_wait = 1; | |
618 | ||
619 | /* | |
620 | * This wait time is without taking into consideration the rounding | |
621 | * up we did. Add that time also. | |
622 | */ | |
623 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
624 | if (wait) |
625 | *wait = jiffy_wait; | |
8e89d13f VG |
626 | return 0; |
627 | } | |
628 | ||
629 | /* | |
630 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
631 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
632 | */ | |
633 | static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg, | |
634 | struct bio *bio, unsigned long *wait) | |
635 | { | |
636 | bool rw = bio_data_dir(bio); | |
637 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
638 | ||
639 | /* | |
640 | * Currently whole state machine of group depends on first bio | |
641 | * queued in the group bio list. So one should not be calling | |
642 | * this function with a different bio if there are other bios | |
643 | * queued. | |
644 | */ | |
645 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); | |
e43473b7 | 646 | |
8e89d13f VG |
647 | /* If tg->bps = -1, then BW is unlimited */ |
648 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
649 | if (wait) | |
650 | *wait = 0; | |
651 | return 1; | |
652 | } | |
653 | ||
654 | /* | |
655 | * If previous slice expired, start a new one otherwise renew/extend | |
656 | * existing slice to make sure it is at least throtl_slice interval | |
657 | * long since now. | |
658 | */ | |
659 | if (throtl_slice_used(td, tg, rw)) | |
660 | throtl_start_new_slice(td, tg, rw); | |
661 | else { | |
662 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
663 | throtl_extend_slice(td, tg, rw, jiffies + throtl_slice); | |
664 | } | |
665 | ||
666 | if (tg_with_in_bps_limit(td, tg, bio, &bps_wait) | |
667 | && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) { | |
668 | if (wait) | |
669 | *wait = 0; | |
670 | return 1; | |
671 | } | |
672 | ||
673 | max_wait = max(bps_wait, iops_wait); | |
674 | ||
675 | if (wait) | |
676 | *wait = max_wait; | |
677 | ||
678 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
679 | throtl_extend_slice(td, tg, rw, jiffies + max_wait); | |
e43473b7 VG |
680 | |
681 | return 0; | |
682 | } | |
683 | ||
684 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
685 | { | |
686 | bool rw = bio_data_dir(bio); | |
687 | bool sync = bio->bi_rw & REQ_SYNC; | |
688 | ||
689 | /* Charge the bio to the group */ | |
690 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 691 | tg->io_disp[rw]++; |
e43473b7 VG |
692 | |
693 | /* | |
694 | * TODO: This will take blkg->stats_lock. Figure out a way | |
695 | * to avoid this cost. | |
696 | */ | |
697 | blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync); | |
e43473b7 VG |
698 | } |
699 | ||
700 | static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg, | |
701 | struct bio *bio) | |
702 | { | |
703 | bool rw = bio_data_dir(bio); | |
704 | ||
705 | bio_list_add(&tg->bio_lists[rw], bio); | |
706 | /* Take a bio reference on tg */ | |
707 | throtl_ref_get_tg(tg); | |
708 | tg->nr_queued[rw]++; | |
709 | td->nr_queued[rw]++; | |
710 | throtl_enqueue_tg(td, tg); | |
711 | } | |
712 | ||
713 | static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg) | |
714 | { | |
715 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
716 | struct bio *bio; | |
717 | ||
718 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
719 | tg_may_dispatch(td, tg, bio, &read_wait); | |
720 | ||
721 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
722 | tg_may_dispatch(td, tg, bio, &write_wait); | |
723 | ||
724 | min_wait = min(read_wait, write_wait); | |
725 | disptime = jiffies + min_wait; | |
726 | ||
e43473b7 VG |
727 | /* Update dispatch time */ |
728 | throtl_dequeue_tg(td, tg); | |
729 | tg->disptime = disptime; | |
730 | throtl_enqueue_tg(td, tg); | |
731 | } | |
732 | ||
733 | static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg, | |
734 | bool rw, struct bio_list *bl) | |
735 | { | |
736 | struct bio *bio; | |
737 | ||
738 | bio = bio_list_pop(&tg->bio_lists[rw]); | |
739 | tg->nr_queued[rw]--; | |
740 | /* Drop bio reference on tg */ | |
741 | throtl_put_tg(tg); | |
742 | ||
743 | BUG_ON(td->nr_queued[rw] <= 0); | |
744 | td->nr_queued[rw]--; | |
745 | ||
746 | throtl_charge_bio(tg, bio); | |
747 | bio_list_add(bl, bio); | |
748 | bio->bi_rw |= REQ_THROTTLED; | |
749 | ||
750 | throtl_trim_slice(td, tg, rw); | |
751 | } | |
752 | ||
753 | static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg, | |
754 | struct bio_list *bl) | |
755 | { | |
756 | unsigned int nr_reads = 0, nr_writes = 0; | |
757 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 758 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
759 | struct bio *bio; |
760 | ||
761 | /* Try to dispatch 75% READS and 25% WRITES */ | |
762 | ||
763 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) | |
764 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
765 | ||
766 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
767 | nr_reads++; | |
768 | ||
769 | if (nr_reads >= max_nr_reads) | |
770 | break; | |
771 | } | |
772 | ||
773 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) | |
774 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
775 | ||
776 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
777 | nr_writes++; | |
778 | ||
779 | if (nr_writes >= max_nr_writes) | |
780 | break; | |
781 | } | |
782 | ||
783 | return nr_reads + nr_writes; | |
784 | } | |
785 | ||
786 | static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl) | |
787 | { | |
788 | unsigned int nr_disp = 0; | |
789 | struct throtl_grp *tg; | |
790 | struct throtl_rb_root *st = &td->tg_service_tree; | |
791 | ||
792 | while (1) { | |
793 | tg = throtl_rb_first(st); | |
794 | ||
795 | if (!tg) | |
796 | break; | |
797 | ||
798 | if (time_before(jiffies, tg->disptime)) | |
799 | break; | |
800 | ||
801 | throtl_dequeue_tg(td, tg); | |
802 | ||
803 | nr_disp += throtl_dispatch_tg(td, tg, bl); | |
804 | ||
805 | if (tg->nr_queued[0] || tg->nr_queued[1]) { | |
806 | tg_update_disptime(td, tg); | |
807 | throtl_enqueue_tg(td, tg); | |
808 | } | |
809 | ||
810 | if (nr_disp >= throtl_quantum) | |
811 | break; | |
812 | } | |
813 | ||
814 | return nr_disp; | |
815 | } | |
816 | ||
fe071437 VG |
817 | static void throtl_process_limit_change(struct throtl_data *td) |
818 | { | |
819 | struct throtl_grp *tg; | |
820 | struct hlist_node *pos, *n; | |
821 | ||
de701c74 | 822 | if (!td->limits_changed) |
fe071437 VG |
823 | return; |
824 | ||
de701c74 | 825 | xchg(&td->limits_changed, false); |
fe071437 | 826 | |
de701c74 | 827 | throtl_log(td, "limits changed"); |
fe071437 | 828 | |
04a6b516 | 829 | hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) { |
de701c74 VG |
830 | if (!tg->limits_changed) |
831 | continue; | |
832 | ||
833 | if (!xchg(&tg->limits_changed, false)) | |
834 | continue; | |
835 | ||
836 | throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu" | |
837 | " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE], | |
838 | tg->iops[READ], tg->iops[WRITE]); | |
839 | ||
04521db0 VG |
840 | /* |
841 | * Restart the slices for both READ and WRITES. It | |
842 | * might happen that a group's limit are dropped | |
843 | * suddenly and we don't want to account recently | |
844 | * dispatched IO with new low rate | |
845 | */ | |
846 | throtl_start_new_slice(td, tg, 0); | |
847 | throtl_start_new_slice(td, tg, 1); | |
848 | ||
de701c74 | 849 | if (throtl_tg_on_rr(tg)) |
fe071437 | 850 | tg_update_disptime(td, tg); |
fe071437 | 851 | } |
fe071437 VG |
852 | } |
853 | ||
e43473b7 VG |
854 | /* Dispatch throttled bios. Should be called without queue lock held. */ |
855 | static int throtl_dispatch(struct request_queue *q) | |
856 | { | |
857 | struct throtl_data *td = q->td; | |
858 | unsigned int nr_disp = 0; | |
859 | struct bio_list bio_list_on_stack; | |
860 | struct bio *bio; | |
69d60eb9 | 861 | struct blk_plug plug; |
e43473b7 VG |
862 | |
863 | spin_lock_irq(q->queue_lock); | |
864 | ||
fe071437 VG |
865 | throtl_process_limit_change(td); |
866 | ||
e43473b7 VG |
867 | if (!total_nr_queued(td)) |
868 | goto out; | |
869 | ||
870 | bio_list_init(&bio_list_on_stack); | |
871 | ||
872 | throtl_log(td, "dispatch nr_queued=%lu read=%u write=%u", | |
873 | total_nr_queued(td), td->nr_queued[READ], | |
874 | td->nr_queued[WRITE]); | |
875 | ||
876 | nr_disp = throtl_select_dispatch(td, &bio_list_on_stack); | |
877 | ||
878 | if (nr_disp) | |
879 | throtl_log(td, "bios disp=%u", nr_disp); | |
880 | ||
881 | throtl_schedule_next_dispatch(td); | |
882 | out: | |
883 | spin_unlock_irq(q->queue_lock); | |
884 | ||
885 | /* | |
886 | * If we dispatched some requests, unplug the queue to make sure | |
887 | * immediate dispatch | |
888 | */ | |
889 | if (nr_disp) { | |
69d60eb9 | 890 | blk_start_plug(&plug); |
e43473b7 VG |
891 | while((bio = bio_list_pop(&bio_list_on_stack))) |
892 | generic_make_request(bio); | |
69d60eb9 | 893 | blk_finish_plug(&plug); |
e43473b7 VG |
894 | } |
895 | return nr_disp; | |
896 | } | |
897 | ||
898 | void blk_throtl_work(struct work_struct *work) | |
899 | { | |
900 | struct throtl_data *td = container_of(work, struct throtl_data, | |
901 | throtl_work.work); | |
902 | struct request_queue *q = td->queue; | |
903 | ||
904 | throtl_dispatch(q); | |
905 | } | |
906 | ||
907 | /* Call with queue lock held */ | |
450adcbe VG |
908 | static void |
909 | throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay) | |
e43473b7 VG |
910 | { |
911 | ||
e43473b7 VG |
912 | struct delayed_work *dwork = &td->throtl_work; |
913 | ||
04521db0 VG |
914 | /* schedule work if limits changed even if no bio is queued */ |
915 | if (total_nr_queued(td) > 0 || td->limits_changed) { | |
e43473b7 VG |
916 | /* |
917 | * We might have a work scheduled to be executed in future. | |
918 | * Cancel that and schedule a new one. | |
919 | */ | |
920 | __cancel_delayed_work(dwork); | |
450adcbe | 921 | queue_delayed_work(kthrotld_workqueue, dwork, delay); |
e43473b7 VG |
922 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", |
923 | delay, jiffies); | |
924 | } | |
925 | } | |
e43473b7 VG |
926 | |
927 | static void | |
928 | throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg) | |
929 | { | |
930 | /* Something wrong if we are trying to remove same group twice */ | |
931 | BUG_ON(hlist_unhashed(&tg->tg_node)); | |
932 | ||
933 | hlist_del_init(&tg->tg_node); | |
934 | ||
935 | /* | |
936 | * Put the reference taken at the time of creation so that when all | |
937 | * queues are gone, group can be destroyed. | |
938 | */ | |
939 | throtl_put_tg(tg); | |
940 | td->nr_undestroyed_grps--; | |
941 | } | |
942 | ||
943 | static void throtl_release_tgs(struct throtl_data *td) | |
944 | { | |
945 | struct hlist_node *pos, *n; | |
946 | struct throtl_grp *tg; | |
947 | ||
948 | hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) { | |
949 | /* | |
950 | * If cgroup removal path got to blk_group first and removed | |
951 | * it from cgroup list, then it will take care of destroying | |
952 | * cfqg also. | |
953 | */ | |
954 | if (!blkiocg_del_blkio_group(&tg->blkg)) | |
955 | throtl_destroy_tg(td, tg); | |
956 | } | |
957 | } | |
958 | ||
959 | static void throtl_td_free(struct throtl_data *td) | |
960 | { | |
961 | kfree(td); | |
962 | } | |
963 | ||
964 | /* | |
965 | * Blk cgroup controller notification saying that blkio_group object is being | |
966 | * delinked as associated cgroup object is going away. That also means that | |
967 | * no new IO will come in this group. So get rid of this group as soon as | |
968 | * any pending IO in the group is finished. | |
969 | * | |
970 | * This function is called under rcu_read_lock(). key is the rcu protected | |
971 | * pointer. That means "key" is a valid throtl_data pointer as long as we are | |
972 | * rcu read lock. | |
973 | * | |
974 | * "key" was fetched from blkio_group under blkio_cgroup->lock. That means | |
975 | * it should not be NULL as even if queue was going away, cgroup deltion | |
976 | * path got to it first. | |
977 | */ | |
978 | void throtl_unlink_blkio_group(void *key, struct blkio_group *blkg) | |
979 | { | |
980 | unsigned long flags; | |
981 | struct throtl_data *td = key; | |
982 | ||
983 | spin_lock_irqsave(td->queue->queue_lock, flags); | |
984 | throtl_destroy_tg(td, tg_of_blkg(blkg)); | |
985 | spin_unlock_irqrestore(td->queue->queue_lock, flags); | |
986 | } | |
987 | ||
de701c74 VG |
988 | static void throtl_update_blkio_group_common(struct throtl_data *td, |
989 | struct throtl_grp *tg) | |
990 | { | |
991 | xchg(&tg->limits_changed, true); | |
992 | xchg(&td->limits_changed, true); | |
993 | /* Schedule a work now to process the limit change */ | |
994 | throtl_schedule_delayed_work(td, 0); | |
995 | } | |
996 | ||
fe071437 VG |
997 | /* |
998 | * For all update functions, key should be a valid pointer because these | |
999 | * update functions are called under blkcg_lock, that means, blkg is | |
25985edc | 1000 | * valid and in turn key is valid. queue exit path can not race because |
fe071437 VG |
1001 | * of blkcg_lock |
1002 | * | |
1003 | * Can not take queue lock in update functions as queue lock under blkcg_lock | |
1004 | * is not allowed. Under other paths we take blkcg_lock under queue_lock. | |
1005 | */ | |
1006 | static void throtl_update_blkio_group_read_bps(void *key, | |
1007 | struct blkio_group *blkg, u64 read_bps) | |
e43473b7 | 1008 | { |
fe071437 | 1009 | struct throtl_data *td = key; |
de701c74 | 1010 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1011 | |
de701c74 VG |
1012 | tg->bps[READ] = read_bps; |
1013 | throtl_update_blkio_group_common(td, tg); | |
e43473b7 VG |
1014 | } |
1015 | ||
fe071437 VG |
1016 | static void throtl_update_blkio_group_write_bps(void *key, |
1017 | struct blkio_group *blkg, u64 write_bps) | |
e43473b7 | 1018 | { |
fe071437 | 1019 | struct throtl_data *td = key; |
de701c74 | 1020 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1021 | |
de701c74 VG |
1022 | tg->bps[WRITE] = write_bps; |
1023 | throtl_update_blkio_group_common(td, tg); | |
e43473b7 VG |
1024 | } |
1025 | ||
fe071437 VG |
1026 | static void throtl_update_blkio_group_read_iops(void *key, |
1027 | struct blkio_group *blkg, unsigned int read_iops) | |
8e89d13f | 1028 | { |
fe071437 | 1029 | struct throtl_data *td = key; |
de701c74 | 1030 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1031 | |
de701c74 VG |
1032 | tg->iops[READ] = read_iops; |
1033 | throtl_update_blkio_group_common(td, tg); | |
8e89d13f VG |
1034 | } |
1035 | ||
fe071437 VG |
1036 | static void throtl_update_blkio_group_write_iops(void *key, |
1037 | struct blkio_group *blkg, unsigned int write_iops) | |
8e89d13f | 1038 | { |
fe071437 | 1039 | struct throtl_data *td = key; |
de701c74 | 1040 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1041 | |
de701c74 VG |
1042 | tg->iops[WRITE] = write_iops; |
1043 | throtl_update_blkio_group_common(td, tg); | |
8e89d13f VG |
1044 | } |
1045 | ||
da527770 | 1046 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1047 | { |
1048 | struct throtl_data *td = q->td; | |
1049 | ||
1050 | cancel_delayed_work_sync(&td->throtl_work); | |
1051 | } | |
1052 | ||
1053 | static struct blkio_policy_type blkio_policy_throtl = { | |
1054 | .ops = { | |
1055 | .blkio_unlink_group_fn = throtl_unlink_blkio_group, | |
1056 | .blkio_update_group_read_bps_fn = | |
1057 | throtl_update_blkio_group_read_bps, | |
1058 | .blkio_update_group_write_bps_fn = | |
1059 | throtl_update_blkio_group_write_bps, | |
8e89d13f VG |
1060 | .blkio_update_group_read_iops_fn = |
1061 | throtl_update_blkio_group_read_iops, | |
1062 | .blkio_update_group_write_iops_fn = | |
1063 | throtl_update_blkio_group_write_iops, | |
e43473b7 | 1064 | }, |
8e89d13f | 1065 | .plid = BLKIO_POLICY_THROTL, |
e43473b7 VG |
1066 | }; |
1067 | ||
1068 | int blk_throtl_bio(struct request_queue *q, struct bio **biop) | |
1069 | { | |
1070 | struct throtl_data *td = q->td; | |
1071 | struct throtl_grp *tg; | |
1072 | struct bio *bio = *biop; | |
1073 | bool rw = bio_data_dir(bio), update_disptime = true; | |
1074 | ||
1075 | if (bio->bi_rw & REQ_THROTTLED) { | |
1076 | bio->bi_rw &= ~REQ_THROTTLED; | |
1077 | return 0; | |
1078 | } | |
1079 | ||
1080 | spin_lock_irq(q->queue_lock); | |
1081 | tg = throtl_get_tg(td); | |
1082 | ||
f469a7b4 VG |
1083 | if (IS_ERR(tg)) { |
1084 | if (PTR_ERR(tg) == -ENODEV) { | |
1085 | /* | |
1086 | * Queue is gone. No queue lock held here. | |
1087 | */ | |
1088 | return -ENODEV; | |
1089 | } | |
1090 | } | |
1091 | ||
e43473b7 VG |
1092 | if (tg->nr_queued[rw]) { |
1093 | /* | |
1094 | * There is already another bio queued in same dir. No | |
1095 | * need to update dispatch time. | |
1096 | */ | |
231d704b | 1097 | update_disptime = false; |
e43473b7 | 1098 | goto queue_bio; |
de701c74 | 1099 | |
e43473b7 VG |
1100 | } |
1101 | ||
1102 | /* Bio is with-in rate limit of group */ | |
1103 | if (tg_may_dispatch(td, tg, bio, NULL)) { | |
1104 | throtl_charge_bio(tg, bio); | |
04521db0 VG |
1105 | |
1106 | /* | |
1107 | * We need to trim slice even when bios are not being queued | |
1108 | * otherwise it might happen that a bio is not queued for | |
1109 | * a long time and slice keeps on extending and trim is not | |
1110 | * called for a long time. Now if limits are reduced suddenly | |
1111 | * we take into account all the IO dispatched so far at new | |
1112 | * low rate and * newly queued IO gets a really long dispatch | |
1113 | * time. | |
1114 | * | |
1115 | * So keep on trimming slice even if bio is not queued. | |
1116 | */ | |
1117 | throtl_trim_slice(td, tg, rw); | |
e43473b7 VG |
1118 | goto out; |
1119 | } | |
1120 | ||
1121 | queue_bio: | |
8e89d13f VG |
1122 | throtl_log_tg(td, tg, "[%c] bio. bdisp=%u sz=%u bps=%llu" |
1123 | " iodisp=%u iops=%u queued=%d/%d", | |
1124 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1125 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1126 | tg->io_disp[rw], tg->iops[rw], |
e43473b7 VG |
1127 | tg->nr_queued[READ], tg->nr_queued[WRITE]); |
1128 | ||
1129 | throtl_add_bio_tg(q->td, tg, bio); | |
1130 | *biop = NULL; | |
1131 | ||
1132 | if (update_disptime) { | |
1133 | tg_update_disptime(td, tg); | |
1134 | throtl_schedule_next_dispatch(td); | |
1135 | } | |
1136 | ||
1137 | out: | |
1138 | spin_unlock_irq(q->queue_lock); | |
1139 | return 0; | |
1140 | } | |
1141 | ||
1142 | int blk_throtl_init(struct request_queue *q) | |
1143 | { | |
1144 | struct throtl_data *td; | |
1145 | struct throtl_grp *tg; | |
1146 | ||
1147 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1148 | if (!td) | |
1149 | return -ENOMEM; | |
1150 | ||
1151 | INIT_HLIST_HEAD(&td->tg_list); | |
1152 | td->tg_service_tree = THROTL_RB_ROOT; | |
de701c74 | 1153 | td->limits_changed = false; |
a29a171e | 1154 | INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work); |
e43473b7 VG |
1155 | |
1156 | /* Init root group */ | |
1157 | tg = &td->root_tg; | |
a29a171e | 1158 | throtl_init_group(tg); |
02977e4a VG |
1159 | |
1160 | /* | |
1161 | * Set root group reference to 2. One reference will be dropped when | |
1162 | * all groups on tg_list are being deleted during queue exit. Other | |
1163 | * reference will remain there as we don't want to delete this group | |
1164 | * as it is statically allocated and gets destroyed when throtl_data | |
1165 | * goes away. | |
1166 | */ | |
a29a171e | 1167 | atomic_inc(&tg->ref); |
e43473b7 VG |
1168 | |
1169 | rcu_read_lock(); | |
1170 | blkiocg_add_blkio_group(&blkio_root_cgroup, &tg->blkg, (void *)td, | |
1171 | 0, BLKIO_POLICY_THROTL); | |
1172 | rcu_read_unlock(); | |
a29a171e | 1173 | throtl_add_group_to_td_list(td, tg); |
e43473b7 VG |
1174 | |
1175 | /* Attach throtl data to request queue */ | |
1176 | td->queue = q; | |
1177 | q->td = td; | |
1178 | return 0; | |
1179 | } | |
1180 | ||
1181 | void blk_throtl_exit(struct request_queue *q) | |
1182 | { | |
1183 | struct throtl_data *td = q->td; | |
1184 | bool wait = false; | |
1185 | ||
1186 | BUG_ON(!td); | |
1187 | ||
da527770 | 1188 | throtl_shutdown_wq(q); |
e43473b7 VG |
1189 | |
1190 | spin_lock_irq(q->queue_lock); | |
1191 | throtl_release_tgs(td); | |
e43473b7 VG |
1192 | |
1193 | /* If there are other groups */ | |
02977e4a | 1194 | if (td->nr_undestroyed_grps > 0) |
e43473b7 VG |
1195 | wait = true; |
1196 | ||
1197 | spin_unlock_irq(q->queue_lock); | |
1198 | ||
1199 | /* | |
1200 | * Wait for tg->blkg->key accessors to exit their grace periods. | |
1201 | * Do this wait only if there are other undestroyed groups out | |
1202 | * there (other than root group). This can happen if cgroup deletion | |
1203 | * path claimed the responsibility of cleaning up a group before | |
1204 | * queue cleanup code get to the group. | |
1205 | * | |
1206 | * Do not call synchronize_rcu() unconditionally as there are drivers | |
1207 | * which create/delete request queue hundreds of times during scan/boot | |
1208 | * and synchronize_rcu() can take significant time and slow down boot. | |
1209 | */ | |
1210 | if (wait) | |
1211 | synchronize_rcu(); | |
fe071437 VG |
1212 | |
1213 | /* | |
1214 | * Just being safe to make sure after previous flush if some body did | |
1215 | * update limits through cgroup and another work got queued, cancel | |
1216 | * it. | |
1217 | */ | |
da527770 | 1218 | throtl_shutdown_wq(q); |
e43473b7 VG |
1219 | throtl_td_free(td); |
1220 | } | |
1221 | ||
1222 | static int __init throtl_init(void) | |
1223 | { | |
450adcbe VG |
1224 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1225 | if (!kthrotld_workqueue) | |
1226 | panic("Failed to create kthrotld\n"); | |
1227 | ||
e43473b7 VG |
1228 | blkio_policy_register(&blkio_policy_throtl); |
1229 | return 0; | |
1230 | } | |
1231 | ||
1232 | module_init(throtl_init); |