]>
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 | ||
cd1604fa TH |
184 | static struct blkio_group *throtl_alloc_blkio_group(struct request_queue *q, |
185 | struct blkio_cgroup *blkcg) | |
a29a171e | 186 | { |
cd1604fa TH |
187 | struct throtl_grp *tg; |
188 | ||
189 | tg = kzalloc_node(sizeof(*tg), GFP_ATOMIC, q->node); | |
190 | if (!tg) | |
191 | return NULL; | |
192 | ||
a29a171e VG |
193 | INIT_HLIST_NODE(&tg->tg_node); |
194 | RB_CLEAR_NODE(&tg->rb_node); | |
195 | bio_list_init(&tg->bio_lists[0]); | |
196 | bio_list_init(&tg->bio_lists[1]); | |
197 | tg->limits_changed = false; | |
198 | ||
cd1604fa TH |
199 | tg->bps[READ] = blkcg_get_read_bps(blkcg, tg->blkg.dev); |
200 | tg->bps[WRITE] = blkcg_get_write_bps(blkcg, tg->blkg.dev); | |
201 | tg->iops[READ] = blkcg_get_read_iops(blkcg, tg->blkg.dev); | |
202 | tg->iops[WRITE] = blkcg_get_write_iops(blkcg, tg->blkg.dev); | |
a29a171e VG |
203 | |
204 | /* | |
205 | * Take the initial reference that will be released on destroy | |
206 | * This can be thought of a joint reference by cgroup and | |
207 | * request queue which will be dropped by either request queue | |
208 | * exit or cgroup deletion path depending on who is exiting first. | |
209 | */ | |
210 | atomic_set(&tg->ref, 1); | |
a29a171e | 211 | |
cd1604fa | 212 | return &tg->blkg; |
a29a171e VG |
213 | } |
214 | ||
269f5415 VG |
215 | static void |
216 | __throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg) | |
f469a7b4 VG |
217 | { |
218 | struct backing_dev_info *bdi = &td->queue->backing_dev_info; | |
219 | unsigned int major, minor; | |
220 | ||
269f5415 VG |
221 | if (!tg || tg->blkg.dev) |
222 | return; | |
223 | ||
224 | /* | |
225 | * Fill in device details for a group which might not have been | |
226 | * filled at group creation time as queue was being instantiated | |
227 | * and driver had not attached a device yet | |
228 | */ | |
229 | if (bdi->dev && dev_name(bdi->dev)) { | |
230 | sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); | |
231 | tg->blkg.dev = MKDEV(major, minor); | |
232 | } | |
233 | } | |
234 | ||
af75cd3c VG |
235 | /* |
236 | * Should be called with without queue lock held. Here queue lock will be | |
237 | * taken rarely. It will be taken only once during life time of a group | |
238 | * if need be | |
239 | */ | |
240 | static void | |
241 | throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg) | |
242 | { | |
243 | if (!tg || tg->blkg.dev) | |
244 | return; | |
245 | ||
246 | spin_lock_irq(td->queue->queue_lock); | |
247 | __throtl_tg_fill_dev_details(td, tg); | |
248 | spin_unlock_irq(td->queue->queue_lock); | |
249 | } | |
250 | ||
cd1604fa TH |
251 | static void throtl_link_blkio_group(struct request_queue *q, |
252 | struct blkio_group *blkg) | |
269f5415 | 253 | { |
cd1604fa TH |
254 | struct throtl_data *td = q->td; |
255 | struct throtl_grp *tg = tg_of_blkg(blkg); | |
5624a4e4 | 256 | |
cd1604fa | 257 | __throtl_tg_fill_dev_details(td, tg); |
5624a4e4 | 258 | |
cd1604fa TH |
259 | hlist_add_head(&tg->tg_node, &td->tg_list); |
260 | td->nr_undestroyed_grps++; | |
f469a7b4 VG |
261 | } |
262 | ||
263 | static struct | |
cd1604fa | 264 | throtl_grp *throtl_lookup_tg(struct throtl_data *td, struct blkio_cgroup *blkcg) |
e43473b7 | 265 | { |
e43473b7 | 266 | struct throtl_grp *tg = NULL; |
e43473b7 | 267 | |
be2c6b19 VG |
268 | /* |
269 | * This is the common case when there are no blkio cgroups. | |
cd1604fa TH |
270 | * Avoid lookup in this case |
271 | */ | |
be2c6b19 | 272 | if (blkcg == &blkio_root_cgroup) |
29b12589 | 273 | tg = td->root_tg; |
be2c6b19 | 274 | else |
cd1604fa TH |
275 | tg = tg_of_blkg(blkg_lookup(blkcg, td->queue, |
276 | BLKIO_POLICY_THROTL)); | |
e43473b7 | 277 | |
269f5415 | 278 | __throtl_tg_fill_dev_details(td, tg); |
e43473b7 VG |
279 | return tg; |
280 | } | |
281 | ||
cd1604fa TH |
282 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
283 | struct blkio_cgroup *blkcg) | |
e43473b7 | 284 | { |
f469a7b4 | 285 | struct request_queue *q = td->queue; |
cd1604fa | 286 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 287 | |
f469a7b4 | 288 | /* |
cd1604fa TH |
289 | * This is the common case when there are no blkio cgroups. |
290 | * Avoid lookup in this case | |
f469a7b4 | 291 | */ |
cd1604fa TH |
292 | if (blkcg == &blkio_root_cgroup) { |
293 | tg = td->root_tg; | |
294 | } else { | |
295 | struct blkio_group *blkg; | |
f469a7b4 | 296 | |
cd1604fa | 297 | blkg = blkg_lookup_create(blkcg, q, BLKIO_POLICY_THROTL, false); |
f469a7b4 | 298 | |
cd1604fa TH |
299 | /* if %NULL and @q is alive, fall back to root_tg */ |
300 | if (!IS_ERR(blkg)) | |
301 | tg = tg_of_blkg(blkg); | |
302 | else if (!blk_queue_dead(q)) | |
303 | tg = td->root_tg; | |
f469a7b4 VG |
304 | } |
305 | ||
cd1604fa | 306 | __throtl_tg_fill_dev_details(td, tg); |
e43473b7 VG |
307 | return tg; |
308 | } | |
309 | ||
310 | static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root) | |
311 | { | |
312 | /* Service tree is empty */ | |
313 | if (!root->count) | |
314 | return NULL; | |
315 | ||
316 | if (!root->left) | |
317 | root->left = rb_first(&root->rb); | |
318 | ||
319 | if (root->left) | |
320 | return rb_entry_tg(root->left); | |
321 | ||
322 | return NULL; | |
323 | } | |
324 | ||
325 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
326 | { | |
327 | rb_erase(n, root); | |
328 | RB_CLEAR_NODE(n); | |
329 | } | |
330 | ||
331 | static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root) | |
332 | { | |
333 | if (root->left == n) | |
334 | root->left = NULL; | |
335 | rb_erase_init(n, &root->rb); | |
336 | --root->count; | |
337 | } | |
338 | ||
339 | static void update_min_dispatch_time(struct throtl_rb_root *st) | |
340 | { | |
341 | struct throtl_grp *tg; | |
342 | ||
343 | tg = throtl_rb_first(st); | |
344 | if (!tg) | |
345 | return; | |
346 | ||
347 | st->min_disptime = tg->disptime; | |
348 | } | |
349 | ||
350 | static void | |
351 | tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg) | |
352 | { | |
353 | struct rb_node **node = &st->rb.rb_node; | |
354 | struct rb_node *parent = NULL; | |
355 | struct throtl_grp *__tg; | |
356 | unsigned long key = tg->disptime; | |
357 | int left = 1; | |
358 | ||
359 | while (*node != NULL) { | |
360 | parent = *node; | |
361 | __tg = rb_entry_tg(parent); | |
362 | ||
363 | if (time_before(key, __tg->disptime)) | |
364 | node = &parent->rb_left; | |
365 | else { | |
366 | node = &parent->rb_right; | |
367 | left = 0; | |
368 | } | |
369 | } | |
370 | ||
371 | if (left) | |
372 | st->left = &tg->rb_node; | |
373 | ||
374 | rb_link_node(&tg->rb_node, parent, node); | |
375 | rb_insert_color(&tg->rb_node, &st->rb); | |
376 | } | |
377 | ||
378 | static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
379 | { | |
380 | struct throtl_rb_root *st = &td->tg_service_tree; | |
381 | ||
382 | tg_service_tree_add(st, tg); | |
383 | throtl_mark_tg_on_rr(tg); | |
384 | st->count++; | |
385 | } | |
386 | ||
387 | static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
388 | { | |
389 | if (!throtl_tg_on_rr(tg)) | |
390 | __throtl_enqueue_tg(td, tg); | |
391 | } | |
392 | ||
393 | static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
394 | { | |
395 | throtl_rb_erase(&tg->rb_node, &td->tg_service_tree); | |
396 | throtl_clear_tg_on_rr(tg); | |
397 | } | |
398 | ||
399 | static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
400 | { | |
401 | if (throtl_tg_on_rr(tg)) | |
402 | __throtl_dequeue_tg(td, tg); | |
403 | } | |
404 | ||
405 | static void throtl_schedule_next_dispatch(struct throtl_data *td) | |
406 | { | |
407 | struct throtl_rb_root *st = &td->tg_service_tree; | |
408 | ||
409 | /* | |
410 | * If there are more bios pending, schedule more work. | |
411 | */ | |
412 | if (!total_nr_queued(td)) | |
413 | return; | |
414 | ||
415 | BUG_ON(!st->count); | |
416 | ||
417 | update_min_dispatch_time(st); | |
418 | ||
419 | if (time_before_eq(st->min_disptime, jiffies)) | |
450adcbe | 420 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 421 | else |
450adcbe | 422 | throtl_schedule_delayed_work(td, (st->min_disptime - jiffies)); |
e43473b7 VG |
423 | } |
424 | ||
425 | static inline void | |
426 | throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
427 | { | |
428 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 429 | tg->io_disp[rw] = 0; |
e43473b7 VG |
430 | tg->slice_start[rw] = jiffies; |
431 | tg->slice_end[rw] = jiffies + throtl_slice; | |
432 | throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
433 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
434 | tg->slice_end[rw], jiffies); | |
435 | } | |
436 | ||
d1ae8ffd VG |
437 | static inline void throtl_set_slice_end(struct throtl_data *td, |
438 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
439 | { | |
440 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
441 | } | |
442 | ||
e43473b7 VG |
443 | static inline void throtl_extend_slice(struct throtl_data *td, |
444 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
445 | { | |
446 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
447 | throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
448 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
449 | tg->slice_end[rw], jiffies); | |
450 | } | |
451 | ||
452 | /* Determine if previously allocated or extended slice is complete or not */ | |
453 | static bool | |
454 | throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
455 | { | |
456 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
457 | return 0; | |
458 | ||
459 | return 1; | |
460 | } | |
461 | ||
462 | /* Trim the used slices and adjust slice start accordingly */ | |
463 | static inline void | |
464 | throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
465 | { | |
3aad5d3e VG |
466 | unsigned long nr_slices, time_elapsed, io_trim; |
467 | u64 bytes_trim, tmp; | |
e43473b7 VG |
468 | |
469 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
470 | ||
471 | /* | |
472 | * If bps are unlimited (-1), then time slice don't get | |
473 | * renewed. Don't try to trim the slice if slice is used. A new | |
474 | * slice will start when appropriate. | |
475 | */ | |
476 | if (throtl_slice_used(td, tg, rw)) | |
477 | return; | |
478 | ||
d1ae8ffd VG |
479 | /* |
480 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
481 | * that initially cgroup limit was very low resulting in high | |
482 | * slice_end, but later limit was bumped up and bio was dispached | |
483 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
484 | * is bad because it does not allow new slice to start. | |
485 | */ | |
486 | ||
487 | throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice); | |
488 | ||
e43473b7 VG |
489 | time_elapsed = jiffies - tg->slice_start[rw]; |
490 | ||
491 | nr_slices = time_elapsed / throtl_slice; | |
492 | ||
493 | if (!nr_slices) | |
494 | return; | |
3aad5d3e VG |
495 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
496 | do_div(tmp, HZ); | |
497 | bytes_trim = tmp; | |
e43473b7 | 498 | |
8e89d13f | 499 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 500 | |
8e89d13f | 501 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
502 | return; |
503 | ||
504 | if (tg->bytes_disp[rw] >= bytes_trim) | |
505 | tg->bytes_disp[rw] -= bytes_trim; | |
506 | else | |
507 | tg->bytes_disp[rw] = 0; | |
508 | ||
8e89d13f VG |
509 | if (tg->io_disp[rw] >= io_trim) |
510 | tg->io_disp[rw] -= io_trim; | |
511 | else | |
512 | tg->io_disp[rw] = 0; | |
513 | ||
e43473b7 VG |
514 | tg->slice_start[rw] += nr_slices * throtl_slice; |
515 | ||
3aad5d3e | 516 | throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 517 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 518 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
519 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
520 | } | |
521 | ||
8e89d13f VG |
522 | static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg, |
523 | struct bio *bio, unsigned long *wait) | |
e43473b7 VG |
524 | { |
525 | bool rw = bio_data_dir(bio); | |
8e89d13f | 526 | unsigned int io_allowed; |
e43473b7 | 527 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 528 | u64 tmp; |
e43473b7 | 529 | |
8e89d13f | 530 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 531 | |
8e89d13f VG |
532 | /* Slice has just started. Consider one slice interval */ |
533 | if (!jiffy_elapsed) | |
534 | jiffy_elapsed_rnd = throtl_slice; | |
535 | ||
536 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
537 | ||
c49c06e4 VG |
538 | /* |
539 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
540 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
541 | * will allow dispatch after 1 second and after that slice should | |
542 | * have been trimmed. | |
543 | */ | |
544 | ||
545 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
546 | do_div(tmp, HZ); | |
547 | ||
548 | if (tmp > UINT_MAX) | |
549 | io_allowed = UINT_MAX; | |
550 | else | |
551 | io_allowed = tmp; | |
8e89d13f VG |
552 | |
553 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
554 | if (wait) |
555 | *wait = 0; | |
556 | return 1; | |
557 | } | |
558 | ||
8e89d13f VG |
559 | /* Calc approx time to dispatch */ |
560 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
561 | ||
562 | if (jiffy_wait > jiffy_elapsed) | |
563 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
564 | else | |
565 | jiffy_wait = 1; | |
566 | ||
567 | if (wait) | |
568 | *wait = jiffy_wait; | |
569 | return 0; | |
570 | } | |
571 | ||
572 | static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg, | |
573 | struct bio *bio, unsigned long *wait) | |
574 | { | |
575 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 576 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 577 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
578 | |
579 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
580 | ||
581 | /* Slice has just started. Consider one slice interval */ | |
582 | if (!jiffy_elapsed) | |
583 | jiffy_elapsed_rnd = throtl_slice; | |
584 | ||
585 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
586 | ||
5e901a2b VG |
587 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
588 | do_div(tmp, HZ); | |
3aad5d3e | 589 | bytes_allowed = tmp; |
e43473b7 VG |
590 | |
591 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
592 | if (wait) | |
593 | *wait = 0; | |
594 | return 1; | |
595 | } | |
596 | ||
597 | /* Calc approx time to dispatch */ | |
598 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
599 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
600 | ||
601 | if (!jiffy_wait) | |
602 | jiffy_wait = 1; | |
603 | ||
604 | /* | |
605 | * This wait time is without taking into consideration the rounding | |
606 | * up we did. Add that time also. | |
607 | */ | |
608 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
609 | if (wait) |
610 | *wait = jiffy_wait; | |
8e89d13f VG |
611 | return 0; |
612 | } | |
613 | ||
af75cd3c VG |
614 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
615 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
616 | return 1; | |
617 | return 0; | |
618 | } | |
619 | ||
8e89d13f VG |
620 | /* |
621 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
622 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
623 | */ | |
624 | static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg, | |
625 | struct bio *bio, unsigned long *wait) | |
626 | { | |
627 | bool rw = bio_data_dir(bio); | |
628 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
629 | ||
630 | /* | |
631 | * Currently whole state machine of group depends on first bio | |
632 | * queued in the group bio list. So one should not be calling | |
633 | * this function with a different bio if there are other bios | |
634 | * queued. | |
635 | */ | |
636 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); | |
e43473b7 | 637 | |
8e89d13f VG |
638 | /* If tg->bps = -1, then BW is unlimited */ |
639 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
640 | if (wait) | |
641 | *wait = 0; | |
642 | return 1; | |
643 | } | |
644 | ||
645 | /* | |
646 | * If previous slice expired, start a new one otherwise renew/extend | |
647 | * existing slice to make sure it is at least throtl_slice interval | |
648 | * long since now. | |
649 | */ | |
650 | if (throtl_slice_used(td, tg, rw)) | |
651 | throtl_start_new_slice(td, tg, rw); | |
652 | else { | |
653 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
654 | throtl_extend_slice(td, tg, rw, jiffies + throtl_slice); | |
655 | } | |
656 | ||
657 | if (tg_with_in_bps_limit(td, tg, bio, &bps_wait) | |
658 | && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) { | |
659 | if (wait) | |
660 | *wait = 0; | |
661 | return 1; | |
662 | } | |
663 | ||
664 | max_wait = max(bps_wait, iops_wait); | |
665 | ||
666 | if (wait) | |
667 | *wait = max_wait; | |
668 | ||
669 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
670 | throtl_extend_slice(td, tg, rw, jiffies + max_wait); | |
e43473b7 VG |
671 | |
672 | return 0; | |
673 | } | |
674 | ||
675 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
676 | { | |
677 | bool rw = bio_data_dir(bio); | |
e5a94f56 | 678 | bool sync = rw_is_sync(bio->bi_rw); |
e43473b7 VG |
679 | |
680 | /* Charge the bio to the group */ | |
681 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 682 | tg->io_disp[rw]++; |
e43473b7 | 683 | |
e43473b7 | 684 | blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync); |
e43473b7 VG |
685 | } |
686 | ||
687 | static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg, | |
688 | struct bio *bio) | |
689 | { | |
690 | bool rw = bio_data_dir(bio); | |
691 | ||
692 | bio_list_add(&tg->bio_lists[rw], bio); | |
693 | /* Take a bio reference on tg */ | |
694 | throtl_ref_get_tg(tg); | |
695 | tg->nr_queued[rw]++; | |
696 | td->nr_queued[rw]++; | |
697 | throtl_enqueue_tg(td, tg); | |
698 | } | |
699 | ||
700 | static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg) | |
701 | { | |
702 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
703 | struct bio *bio; | |
704 | ||
705 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
706 | tg_may_dispatch(td, tg, bio, &read_wait); | |
707 | ||
708 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
709 | tg_may_dispatch(td, tg, bio, &write_wait); | |
710 | ||
711 | min_wait = min(read_wait, write_wait); | |
712 | disptime = jiffies + min_wait; | |
713 | ||
e43473b7 VG |
714 | /* Update dispatch time */ |
715 | throtl_dequeue_tg(td, tg); | |
716 | tg->disptime = disptime; | |
717 | throtl_enqueue_tg(td, tg); | |
718 | } | |
719 | ||
720 | static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg, | |
721 | bool rw, struct bio_list *bl) | |
722 | { | |
723 | struct bio *bio; | |
724 | ||
725 | bio = bio_list_pop(&tg->bio_lists[rw]); | |
726 | tg->nr_queued[rw]--; | |
727 | /* Drop bio reference on tg */ | |
728 | throtl_put_tg(tg); | |
729 | ||
730 | BUG_ON(td->nr_queued[rw] <= 0); | |
731 | td->nr_queued[rw]--; | |
732 | ||
733 | throtl_charge_bio(tg, bio); | |
734 | bio_list_add(bl, bio); | |
735 | bio->bi_rw |= REQ_THROTTLED; | |
736 | ||
737 | throtl_trim_slice(td, tg, rw); | |
738 | } | |
739 | ||
740 | static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg, | |
741 | struct bio_list *bl) | |
742 | { | |
743 | unsigned int nr_reads = 0, nr_writes = 0; | |
744 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 745 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
746 | struct bio *bio; |
747 | ||
748 | /* Try to dispatch 75% READS and 25% WRITES */ | |
749 | ||
750 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) | |
751 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
752 | ||
753 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
754 | nr_reads++; | |
755 | ||
756 | if (nr_reads >= max_nr_reads) | |
757 | break; | |
758 | } | |
759 | ||
760 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) | |
761 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
762 | ||
763 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
764 | nr_writes++; | |
765 | ||
766 | if (nr_writes >= max_nr_writes) | |
767 | break; | |
768 | } | |
769 | ||
770 | return nr_reads + nr_writes; | |
771 | } | |
772 | ||
773 | static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl) | |
774 | { | |
775 | unsigned int nr_disp = 0; | |
776 | struct throtl_grp *tg; | |
777 | struct throtl_rb_root *st = &td->tg_service_tree; | |
778 | ||
779 | while (1) { | |
780 | tg = throtl_rb_first(st); | |
781 | ||
782 | if (!tg) | |
783 | break; | |
784 | ||
785 | if (time_before(jiffies, tg->disptime)) | |
786 | break; | |
787 | ||
788 | throtl_dequeue_tg(td, tg); | |
789 | ||
790 | nr_disp += throtl_dispatch_tg(td, tg, bl); | |
791 | ||
792 | if (tg->nr_queued[0] || tg->nr_queued[1]) { | |
793 | tg_update_disptime(td, tg); | |
794 | throtl_enqueue_tg(td, tg); | |
795 | } | |
796 | ||
797 | if (nr_disp >= throtl_quantum) | |
798 | break; | |
799 | } | |
800 | ||
801 | return nr_disp; | |
802 | } | |
803 | ||
fe071437 VG |
804 | static void throtl_process_limit_change(struct throtl_data *td) |
805 | { | |
806 | struct throtl_grp *tg; | |
807 | struct hlist_node *pos, *n; | |
808 | ||
de701c74 | 809 | if (!td->limits_changed) |
fe071437 VG |
810 | return; |
811 | ||
de701c74 | 812 | xchg(&td->limits_changed, false); |
fe071437 | 813 | |
de701c74 | 814 | throtl_log(td, "limits changed"); |
fe071437 | 815 | |
04a6b516 | 816 | hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) { |
de701c74 VG |
817 | if (!tg->limits_changed) |
818 | continue; | |
819 | ||
820 | if (!xchg(&tg->limits_changed, false)) | |
821 | continue; | |
822 | ||
823 | throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu" | |
824 | " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE], | |
825 | tg->iops[READ], tg->iops[WRITE]); | |
826 | ||
04521db0 VG |
827 | /* |
828 | * Restart the slices for both READ and WRITES. It | |
829 | * might happen that a group's limit are dropped | |
830 | * suddenly and we don't want to account recently | |
831 | * dispatched IO with new low rate | |
832 | */ | |
833 | throtl_start_new_slice(td, tg, 0); | |
834 | throtl_start_new_slice(td, tg, 1); | |
835 | ||
de701c74 | 836 | if (throtl_tg_on_rr(tg)) |
fe071437 | 837 | tg_update_disptime(td, tg); |
fe071437 | 838 | } |
fe071437 VG |
839 | } |
840 | ||
e43473b7 VG |
841 | /* Dispatch throttled bios. Should be called without queue lock held. */ |
842 | static int throtl_dispatch(struct request_queue *q) | |
843 | { | |
844 | struct throtl_data *td = q->td; | |
845 | unsigned int nr_disp = 0; | |
846 | struct bio_list bio_list_on_stack; | |
847 | struct bio *bio; | |
69d60eb9 | 848 | struct blk_plug plug; |
e43473b7 VG |
849 | |
850 | spin_lock_irq(q->queue_lock); | |
851 | ||
fe071437 VG |
852 | throtl_process_limit_change(td); |
853 | ||
e43473b7 VG |
854 | if (!total_nr_queued(td)) |
855 | goto out; | |
856 | ||
857 | bio_list_init(&bio_list_on_stack); | |
858 | ||
d2f31a5f | 859 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
e43473b7 VG |
860 | total_nr_queued(td), td->nr_queued[READ], |
861 | td->nr_queued[WRITE]); | |
862 | ||
863 | nr_disp = throtl_select_dispatch(td, &bio_list_on_stack); | |
864 | ||
865 | if (nr_disp) | |
866 | throtl_log(td, "bios disp=%u", nr_disp); | |
867 | ||
868 | throtl_schedule_next_dispatch(td); | |
869 | out: | |
870 | spin_unlock_irq(q->queue_lock); | |
871 | ||
872 | /* | |
873 | * If we dispatched some requests, unplug the queue to make sure | |
874 | * immediate dispatch | |
875 | */ | |
876 | if (nr_disp) { | |
69d60eb9 | 877 | blk_start_plug(&plug); |
e43473b7 VG |
878 | while((bio = bio_list_pop(&bio_list_on_stack))) |
879 | generic_make_request(bio); | |
69d60eb9 | 880 | blk_finish_plug(&plug); |
e43473b7 VG |
881 | } |
882 | return nr_disp; | |
883 | } | |
884 | ||
885 | void blk_throtl_work(struct work_struct *work) | |
886 | { | |
887 | struct throtl_data *td = container_of(work, struct throtl_data, | |
888 | throtl_work.work); | |
889 | struct request_queue *q = td->queue; | |
890 | ||
891 | throtl_dispatch(q); | |
892 | } | |
893 | ||
894 | /* Call with queue lock held */ | |
450adcbe VG |
895 | static void |
896 | throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay) | |
e43473b7 VG |
897 | { |
898 | ||
e43473b7 VG |
899 | struct delayed_work *dwork = &td->throtl_work; |
900 | ||
04521db0 | 901 | /* schedule work if limits changed even if no bio is queued */ |
d2f31a5f | 902 | if (total_nr_queued(td) || td->limits_changed) { |
e43473b7 VG |
903 | /* |
904 | * We might have a work scheduled to be executed in future. | |
905 | * Cancel that and schedule a new one. | |
906 | */ | |
907 | __cancel_delayed_work(dwork); | |
450adcbe | 908 | queue_delayed_work(kthrotld_workqueue, dwork, delay); |
e43473b7 VG |
909 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", |
910 | delay, jiffies); | |
911 | } | |
912 | } | |
e43473b7 VG |
913 | |
914 | static void | |
915 | throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg) | |
916 | { | |
917 | /* Something wrong if we are trying to remove same group twice */ | |
918 | BUG_ON(hlist_unhashed(&tg->tg_node)); | |
919 | ||
920 | hlist_del_init(&tg->tg_node); | |
921 | ||
922 | /* | |
923 | * Put the reference taken at the time of creation so that when all | |
924 | * queues are gone, group can be destroyed. | |
925 | */ | |
926 | throtl_put_tg(tg); | |
927 | td->nr_undestroyed_grps--; | |
928 | } | |
929 | ||
72e06c25 | 930 | static bool throtl_release_tgs(struct throtl_data *td, bool release_root) |
e43473b7 VG |
931 | { |
932 | struct hlist_node *pos, *n; | |
933 | struct throtl_grp *tg; | |
72e06c25 | 934 | bool empty = true; |
e43473b7 VG |
935 | |
936 | hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) { | |
72e06c25 TH |
937 | /* skip root? */ |
938 | if (!release_root && tg == td->root_tg) | |
939 | continue; | |
940 | ||
e43473b7 VG |
941 | /* |
942 | * If cgroup removal path got to blk_group first and removed | |
943 | * it from cgroup list, then it will take care of destroying | |
944 | * cfqg also. | |
945 | */ | |
946 | if (!blkiocg_del_blkio_group(&tg->blkg)) | |
947 | throtl_destroy_tg(td, tg); | |
72e06c25 TH |
948 | else |
949 | empty = false; | |
e43473b7 | 950 | } |
72e06c25 | 951 | return empty; |
e43473b7 VG |
952 | } |
953 | ||
e43473b7 VG |
954 | /* |
955 | * Blk cgroup controller notification saying that blkio_group object is being | |
956 | * delinked as associated cgroup object is going away. That also means that | |
957 | * no new IO will come in this group. So get rid of this group as soon as | |
958 | * any pending IO in the group is finished. | |
959 | * | |
ca32aefc TH |
960 | * This function is called under rcu_read_lock(). @q is the rcu protected |
961 | * pointer. That means @q is a valid request_queue pointer as long as we | |
962 | * are rcu read lock. | |
e43473b7 | 963 | * |
ca32aefc | 964 | * @q was fetched from blkio_group under blkio_cgroup->lock. That means |
e43473b7 VG |
965 | * it should not be NULL as even if queue was going away, cgroup deltion |
966 | * path got to it first. | |
967 | */ | |
ca32aefc TH |
968 | void throtl_unlink_blkio_group(struct request_queue *q, |
969 | struct blkio_group *blkg) | |
e43473b7 VG |
970 | { |
971 | unsigned long flags; | |
e43473b7 | 972 | |
ca32aefc TH |
973 | spin_lock_irqsave(q->queue_lock, flags); |
974 | throtl_destroy_tg(q->td, tg_of_blkg(blkg)); | |
975 | spin_unlock_irqrestore(q->queue_lock, flags); | |
e43473b7 VG |
976 | } |
977 | ||
72e06c25 TH |
978 | static bool throtl_clear_queue(struct request_queue *q) |
979 | { | |
980 | lockdep_assert_held(q->queue_lock); | |
981 | ||
982 | /* | |
983 | * Clear tgs but leave the root one alone. This is necessary | |
984 | * because root_tg is expected to be persistent and safe because | |
985 | * blk-throtl can never be disabled while @q is alive. This is a | |
986 | * kludge to prepare for unified blkg. This whole function will be | |
987 | * removed soon. | |
988 | */ | |
989 | return throtl_release_tgs(q->td, false); | |
990 | } | |
991 | ||
de701c74 VG |
992 | static void throtl_update_blkio_group_common(struct throtl_data *td, |
993 | struct throtl_grp *tg) | |
994 | { | |
995 | xchg(&tg->limits_changed, true); | |
996 | xchg(&td->limits_changed, true); | |
997 | /* Schedule a work now to process the limit change */ | |
998 | throtl_schedule_delayed_work(td, 0); | |
999 | } | |
1000 | ||
fe071437 | 1001 | /* |
ca32aefc | 1002 | * For all update functions, @q should be a valid pointer because these |
fe071437 | 1003 | * update functions are called under blkcg_lock, that means, blkg is |
ca32aefc | 1004 | * valid and in turn @q is valid. queue exit path can not race because |
fe071437 VG |
1005 | * of blkcg_lock |
1006 | * | |
1007 | * Can not take queue lock in update functions as queue lock under blkcg_lock | |
1008 | * is not allowed. Under other paths we take blkcg_lock under queue_lock. | |
1009 | */ | |
ca32aefc | 1010 | static void throtl_update_blkio_group_read_bps(struct request_queue *q, |
fe071437 | 1011 | struct blkio_group *blkg, u64 read_bps) |
e43473b7 | 1012 | { |
de701c74 | 1013 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1014 | |
de701c74 | 1015 | tg->bps[READ] = read_bps; |
ca32aefc | 1016 | throtl_update_blkio_group_common(q->td, tg); |
e43473b7 VG |
1017 | } |
1018 | ||
ca32aefc | 1019 | static void throtl_update_blkio_group_write_bps(struct request_queue *q, |
fe071437 | 1020 | struct blkio_group *blkg, u64 write_bps) |
e43473b7 | 1021 | { |
de701c74 | 1022 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1023 | |
de701c74 | 1024 | tg->bps[WRITE] = write_bps; |
ca32aefc | 1025 | throtl_update_blkio_group_common(q->td, tg); |
e43473b7 VG |
1026 | } |
1027 | ||
ca32aefc | 1028 | static void throtl_update_blkio_group_read_iops(struct request_queue *q, |
fe071437 | 1029 | struct blkio_group *blkg, unsigned int read_iops) |
8e89d13f | 1030 | { |
de701c74 | 1031 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1032 | |
de701c74 | 1033 | tg->iops[READ] = read_iops; |
ca32aefc | 1034 | throtl_update_blkio_group_common(q->td, tg); |
8e89d13f VG |
1035 | } |
1036 | ||
ca32aefc | 1037 | static void throtl_update_blkio_group_write_iops(struct request_queue *q, |
fe071437 | 1038 | struct blkio_group *blkg, unsigned int write_iops) |
8e89d13f | 1039 | { |
de701c74 | 1040 | struct throtl_grp *tg = tg_of_blkg(blkg); |
fe071437 | 1041 | |
de701c74 | 1042 | tg->iops[WRITE] = write_iops; |
ca32aefc | 1043 | throtl_update_blkio_group_common(q->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 = { | |
cd1604fa TH |
1055 | .blkio_alloc_group_fn = throtl_alloc_blkio_group, |
1056 | .blkio_link_group_fn = throtl_link_blkio_group, | |
e43473b7 | 1057 | .blkio_unlink_group_fn = throtl_unlink_blkio_group, |
72e06c25 | 1058 | .blkio_clear_queue_fn = throtl_clear_queue, |
e43473b7 VG |
1059 | .blkio_update_group_read_bps_fn = |
1060 | throtl_update_blkio_group_read_bps, | |
1061 | .blkio_update_group_write_bps_fn = | |
1062 | throtl_update_blkio_group_write_bps, | |
8e89d13f VG |
1063 | .blkio_update_group_read_iops_fn = |
1064 | throtl_update_blkio_group_read_iops, | |
1065 | .blkio_update_group_write_iops_fn = | |
1066 | throtl_update_blkio_group_write_iops, | |
e43473b7 | 1067 | }, |
8e89d13f | 1068 | .plid = BLKIO_POLICY_THROTL, |
e43473b7 VG |
1069 | }; |
1070 | ||
bc16a4f9 | 1071 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1072 | { |
1073 | struct throtl_data *td = q->td; | |
1074 | struct throtl_grp *tg; | |
e43473b7 | 1075 | bool rw = bio_data_dir(bio), update_disptime = true; |
af75cd3c | 1076 | struct blkio_cgroup *blkcg; |
bc16a4f9 | 1077 | bool throttled = false; |
e43473b7 VG |
1078 | |
1079 | if (bio->bi_rw & REQ_THROTTLED) { | |
1080 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1081 | goto out; |
e43473b7 VG |
1082 | } |
1083 | ||
af75cd3c VG |
1084 | /* |
1085 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1086 | * basic fields like stats and io rates. If a group has no rules, | |
1087 | * just update the dispatch stats in lockless manner and return. | |
1088 | */ | |
af75cd3c VG |
1089 | rcu_read_lock(); |
1090 | blkcg = task_blkio_cgroup(current); | |
cd1604fa | 1091 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c VG |
1092 | if (tg) { |
1093 | throtl_tg_fill_dev_details(td, tg); | |
1094 | ||
1095 | if (tg_no_rule_group(tg, rw)) { | |
1096 | blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, | |
e5a94f56 | 1097 | rw, rw_is_sync(bio->bi_rw)); |
2a7f1244 | 1098 | goto out_unlock_rcu; |
af75cd3c VG |
1099 | } |
1100 | } | |
af75cd3c VG |
1101 | |
1102 | /* | |
1103 | * Either group has not been allocated yet or it is not an unlimited | |
1104 | * IO group | |
1105 | */ | |
e43473b7 | 1106 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1107 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1108 | if (unlikely(!tg)) |
1109 | goto out_unlock; | |
f469a7b4 | 1110 | |
e43473b7 VG |
1111 | if (tg->nr_queued[rw]) { |
1112 | /* | |
1113 | * There is already another bio queued in same dir. No | |
1114 | * need to update dispatch time. | |
1115 | */ | |
231d704b | 1116 | update_disptime = false; |
e43473b7 | 1117 | goto queue_bio; |
de701c74 | 1118 | |
e43473b7 VG |
1119 | } |
1120 | ||
1121 | /* Bio is with-in rate limit of group */ | |
1122 | if (tg_may_dispatch(td, tg, bio, NULL)) { | |
1123 | throtl_charge_bio(tg, bio); | |
04521db0 VG |
1124 | |
1125 | /* | |
1126 | * We need to trim slice even when bios are not being queued | |
1127 | * otherwise it might happen that a bio is not queued for | |
1128 | * a long time and slice keeps on extending and trim is not | |
1129 | * called for a long time. Now if limits are reduced suddenly | |
1130 | * we take into account all the IO dispatched so far at new | |
1131 | * low rate and * newly queued IO gets a really long dispatch | |
1132 | * time. | |
1133 | * | |
1134 | * So keep on trimming slice even if bio is not queued. | |
1135 | */ | |
1136 | throtl_trim_slice(td, tg, rw); | |
bc16a4f9 | 1137 | goto out_unlock; |
e43473b7 VG |
1138 | } |
1139 | ||
1140 | queue_bio: | |
fd16d263 | 1141 | throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
8e89d13f VG |
1142 | " iodisp=%u iops=%u queued=%d/%d", |
1143 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1144 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1145 | tg->io_disp[rw], tg->iops[rw], |
e43473b7 VG |
1146 | tg->nr_queued[READ], tg->nr_queued[WRITE]); |
1147 | ||
1148 | throtl_add_bio_tg(q->td, tg, bio); | |
bc16a4f9 | 1149 | throttled = true; |
e43473b7 VG |
1150 | |
1151 | if (update_disptime) { | |
1152 | tg_update_disptime(td, tg); | |
1153 | throtl_schedule_next_dispatch(td); | |
1154 | } | |
1155 | ||
bc16a4f9 | 1156 | out_unlock: |
e43473b7 | 1157 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1158 | out_unlock_rcu: |
1159 | rcu_read_unlock(); | |
bc16a4f9 TH |
1160 | out: |
1161 | return throttled; | |
e43473b7 VG |
1162 | } |
1163 | ||
c9a929dd TH |
1164 | /** |
1165 | * blk_throtl_drain - drain throttled bios | |
1166 | * @q: request_queue to drain throttled bios for | |
1167 | * | |
1168 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1169 | */ | |
1170 | void blk_throtl_drain(struct request_queue *q) | |
1171 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1172 | { | |
1173 | struct throtl_data *td = q->td; | |
1174 | struct throtl_rb_root *st = &td->tg_service_tree; | |
1175 | struct throtl_grp *tg; | |
1176 | struct bio_list bl; | |
1177 | struct bio *bio; | |
1178 | ||
334c2b0b | 1179 | WARN_ON_ONCE(!queue_is_locked(q)); |
c9a929dd TH |
1180 | |
1181 | bio_list_init(&bl); | |
1182 | ||
1183 | while ((tg = throtl_rb_first(st))) { | |
1184 | throtl_dequeue_tg(td, tg); | |
1185 | ||
1186 | while ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
1187 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1188 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
1189 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1190 | } | |
1191 | spin_unlock_irq(q->queue_lock); | |
1192 | ||
1193 | while ((bio = bio_list_pop(&bl))) | |
1194 | generic_make_request(bio); | |
1195 | ||
1196 | spin_lock_irq(q->queue_lock); | |
1197 | } | |
1198 | ||
e43473b7 VG |
1199 | int blk_throtl_init(struct request_queue *q) |
1200 | { | |
1201 | struct throtl_data *td; | |
cd1604fa | 1202 | struct blkio_group *blkg; |
e43473b7 VG |
1203 | |
1204 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1205 | if (!td) | |
1206 | return -ENOMEM; | |
1207 | ||
1208 | INIT_HLIST_HEAD(&td->tg_list); | |
1209 | td->tg_service_tree = THROTL_RB_ROOT; | |
de701c74 | 1210 | td->limits_changed = false; |
a29a171e | 1211 | INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work); |
e43473b7 | 1212 | |
cd1604fa | 1213 | q->td = td; |
29b12589 | 1214 | td->queue = q; |
02977e4a | 1215 | |
cd1604fa | 1216 | /* alloc and init root group. */ |
f51b802c TH |
1217 | rcu_read_lock(); |
1218 | spin_lock_irq(q->queue_lock); | |
29b12589 | 1219 | |
cd1604fa TH |
1220 | blkg = blkg_lookup_create(&blkio_root_cgroup, q, BLKIO_POLICY_THROTL, |
1221 | true); | |
1222 | if (!IS_ERR(blkg)) | |
1223 | td->root_tg = tg_of_blkg(blkg); | |
e43473b7 | 1224 | |
f51b802c | 1225 | spin_unlock_irq(q->queue_lock); |
e43473b7 VG |
1226 | rcu_read_unlock(); |
1227 | ||
f51b802c TH |
1228 | if (!td->root_tg) { |
1229 | kfree(td); | |
1230 | return -ENOMEM; | |
1231 | } | |
e43473b7 VG |
1232 | return 0; |
1233 | } | |
1234 | ||
1235 | void blk_throtl_exit(struct request_queue *q) | |
1236 | { | |
1237 | struct throtl_data *td = q->td; | |
1238 | bool wait = false; | |
1239 | ||
1240 | BUG_ON(!td); | |
1241 | ||
da527770 | 1242 | throtl_shutdown_wq(q); |
e43473b7 VG |
1243 | |
1244 | spin_lock_irq(q->queue_lock); | |
72e06c25 | 1245 | throtl_release_tgs(td, true); |
e43473b7 VG |
1246 | |
1247 | /* If there are other groups */ | |
02977e4a | 1248 | if (td->nr_undestroyed_grps > 0) |
e43473b7 VG |
1249 | wait = true; |
1250 | ||
1251 | spin_unlock_irq(q->queue_lock); | |
1252 | ||
1253 | /* | |
ca32aefc | 1254 | * Wait for tg->blkg->q accessors to exit their grace periods. |
e43473b7 VG |
1255 | * Do this wait only if there are other undestroyed groups out |
1256 | * there (other than root group). This can happen if cgroup deletion | |
1257 | * path claimed the responsibility of cleaning up a group before | |
1258 | * queue cleanup code get to the group. | |
1259 | * | |
1260 | * Do not call synchronize_rcu() unconditionally as there are drivers | |
1261 | * which create/delete request queue hundreds of times during scan/boot | |
1262 | * and synchronize_rcu() can take significant time and slow down boot. | |
1263 | */ | |
1264 | if (wait) | |
1265 | synchronize_rcu(); | |
fe071437 VG |
1266 | |
1267 | /* | |
1268 | * Just being safe to make sure after previous flush if some body did | |
1269 | * update limits through cgroup and another work got queued, cancel | |
1270 | * it. | |
1271 | */ | |
da527770 | 1272 | throtl_shutdown_wq(q); |
c9a929dd TH |
1273 | } |
1274 | ||
1275 | void blk_throtl_release(struct request_queue *q) | |
1276 | { | |
1277 | kfree(q->td); | |
e43473b7 VG |
1278 | } |
1279 | ||
1280 | static int __init throtl_init(void) | |
1281 | { | |
450adcbe VG |
1282 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1283 | if (!kthrotld_workqueue) | |
1284 | panic("Failed to create kthrotld\n"); | |
1285 | ||
e43473b7 VG |
1286 | blkio_policy_register(&blkio_policy_throtl); |
1287 | return 0; | |
1288 | } | |
1289 | ||
1290 | module_init(throtl_init); |