blob: f6a794120505668084929afdacd0b5e17f8fca8d [file] [log] [blame]
Vivek Goyale43473b2010-09-15 17:06:35 -04001/*
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 */
15static int throtl_grp_quantum = 8;
16
17/* Total max dispatch from all groups in one round */
18static int throtl_quantum = 32;
19
20/* Throttling is performed over 100ms slice and after that slice is renewed */
21static unsigned long throtl_slice = HZ/10; /* 100 ms */
22
Vivek Goyal450adcb2011-03-01 13:40:54 -050023/* A workqueue to queue throttle related work */
24static struct workqueue_struct *kthrotld_workqueue;
25static void throtl_schedule_delayed_work(struct throtl_data *td,
26 unsigned long delay);
27
Vivek Goyale43473b2010-09-15 17:06:35 -040028struct 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
40struct 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
Vivek Goyal8e89d132010-09-15 17:06:37 -040067 /* IOPS limits */
68 unsigned int iops[2];
69
Vivek Goyale43473b2010-09-15 17:06:35 -040070 /* Number of bytes disptached in current slice */
71 uint64_t bytes_disp[2];
Vivek Goyal8e89d132010-09-15 17:06:37 -040072 /* Number of bio's dispatched in current slice */
73 unsigned int io_disp[2];
Vivek Goyale43473b2010-09-15 17:06:35 -040074
75 /* When did we start a new slice */
76 unsigned long slice_start[2];
77 unsigned long slice_end[2];
Vivek Goyalfe071432010-10-01 14:49:49 +020078
79 /* Some throttle limits got updated for the group */
Andreas Schwab6f037932011-03-30 12:21:56 +020080 int limits_changed;
Vivek Goyal4843c692011-05-19 15:38:27 -040081
82 struct rcu_head rcu_head;
Vivek Goyale43473b2010-09-15 17:06:35 -040083};
84
85struct throtl_data
86{
87 /* List of throtl groups */
88 struct hlist_head tg_list;
89
90 /* service tree for active throtl groups */
91 struct throtl_rb_root tg_service_tree;
92
Vivek Goyal29b12582011-05-19 15:38:24 -040093 struct throtl_grp *root_tg;
Vivek Goyale43473b2010-09-15 17:06:35 -040094 struct request_queue *queue;
95
96 /* Total Number of queued bios on READ and WRITE lists */
97 unsigned int nr_queued[2];
98
99 /*
Vivek Goyal02977e42010-10-01 14:49:48 +0200100 * number of total undestroyed groups
Vivek Goyale43473b2010-09-15 17:06:35 -0400101 */
102 unsigned int nr_undestroyed_grps;
103
104 /* Work for dispatching throttled bios */
105 struct delayed_work throtl_work;
Vivek Goyalfe071432010-10-01 14:49:49 +0200106
Andreas Schwab6f037932011-03-30 12:21:56 +0200107 int limits_changed;
Vivek Goyale43473b2010-09-15 17:06:35 -0400108};
109
110enum tg_state_flags {
111 THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */
112};
113
114#define THROTL_TG_FNS(name) \
115static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \
116{ \
117 (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \
118} \
119static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \
120{ \
121 (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \
122} \
123static inline int throtl_tg_##name(const struct throtl_grp *tg) \
124{ \
125 return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \
126}
127
128THROTL_TG_FNS(on_rr);
129
130#define throtl_log_tg(td, tg, fmt, args...) \
131 blk_add_trace_msg((td)->queue, "throtl %s " fmt, \
132 blkg_path(&(tg)->blkg), ##args); \
133
134#define throtl_log(td, fmt, args...) \
135 blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
136
137static inline struct throtl_grp *tg_of_blkg(struct blkio_group *blkg)
138{
139 if (blkg)
140 return container_of(blkg, struct throtl_grp, blkg);
141
142 return NULL;
143}
144
Joe Perchesd2f31a52011-06-13 20:19:27 +0200145static inline unsigned int total_nr_queued(struct throtl_data *td)
Vivek Goyale43473b2010-09-15 17:06:35 -0400146{
Joe Perchesd2f31a52011-06-13 20:19:27 +0200147 return td->nr_queued[0] + td->nr_queued[1];
Vivek Goyale43473b2010-09-15 17:06:35 -0400148}
149
150static inline struct throtl_grp *throtl_ref_get_tg(struct throtl_grp *tg)
151{
152 atomic_inc(&tg->ref);
153 return tg;
154}
155
Vivek Goyal4843c692011-05-19 15:38:27 -0400156static void throtl_free_tg(struct rcu_head *head)
157{
158 struct throtl_grp *tg;
159
160 tg = container_of(head, struct throtl_grp, rcu_head);
Vivek Goyal5624a4e2011-05-19 15:38:28 -0400161 free_percpu(tg->blkg.stats_cpu);
Vivek Goyal4843c692011-05-19 15:38:27 -0400162 kfree(tg);
163}
164
Vivek Goyale43473b2010-09-15 17:06:35 -0400165static void throtl_put_tg(struct throtl_grp *tg)
166{
167 BUG_ON(atomic_read(&tg->ref) <= 0);
168 if (!atomic_dec_and_test(&tg->ref))
169 return;
Vivek Goyal4843c692011-05-19 15:38:27 -0400170
171 /*
172 * A group is freed in rcu manner. But having an rcu lock does not
173 * mean that one can access all the fields of blkg and assume these
174 * are valid. For example, don't try to follow throtl_data and
175 * request queue links.
176 *
177 * Having a reference to blkg under an rcu allows acess to only
178 * values local to groups like group stats and group rate limits
179 */
180 call_rcu(&tg->rcu_head, throtl_free_tg);
Vivek Goyale43473b2010-09-15 17:06:35 -0400181}
182
Vivek Goyala29a1712011-05-19 15:38:19 -0400183static void throtl_init_group(struct throtl_grp *tg)
184{
185 INIT_HLIST_NODE(&tg->tg_node);
186 RB_CLEAR_NODE(&tg->rb_node);
187 bio_list_init(&tg->bio_lists[0]);
188 bio_list_init(&tg->bio_lists[1]);
189 tg->limits_changed = false;
190
191 /* Practically unlimited BW */
192 tg->bps[0] = tg->bps[1] = -1;
193 tg->iops[0] = tg->iops[1] = -1;
194
195 /*
196 * Take the initial reference that will be released on destroy
197 * This can be thought of a joint reference by cgroup and
198 * request queue which will be dropped by either request queue
199 * exit or cgroup deletion path depending on who is exiting first.
200 */
201 atomic_set(&tg->ref, 1);
202}
203
204/* Should be called with rcu read lock held (needed for blkcg) */
205static void
206throtl_add_group_to_td_list(struct throtl_data *td, struct throtl_grp *tg)
207{
208 hlist_add_head(&tg->tg_node, &td->tg_list);
209 td->nr_undestroyed_grps++;
210}
211
Vivek Goyal269f5412011-05-19 15:38:25 -0400212static void
213__throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg)
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400214{
215 struct backing_dev_info *bdi = &td->queue->backing_dev_info;
216 unsigned int major, minor;
217
Vivek Goyal269f5412011-05-19 15:38:25 -0400218 if (!tg || tg->blkg.dev)
219 return;
220
221 /*
222 * Fill in device details for a group which might not have been
223 * filled at group creation time as queue was being instantiated
224 * and driver had not attached a device yet
225 */
226 if (bdi->dev && dev_name(bdi->dev)) {
227 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
228 tg->blkg.dev = MKDEV(major, minor);
229 }
230}
231
Vivek Goyalaf75cd32011-05-19 15:38:31 -0400232/*
233 * Should be called with without queue lock held. Here queue lock will be
234 * taken rarely. It will be taken only once during life time of a group
235 * if need be
236 */
237static void
238throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg)
239{
240 if (!tg || tg->blkg.dev)
241 return;
242
243 spin_lock_irq(td->queue->queue_lock);
244 __throtl_tg_fill_dev_details(td, tg);
245 spin_unlock_irq(td->queue->queue_lock);
246}
247
Vivek Goyal269f5412011-05-19 15:38:25 -0400248static void throtl_init_add_tg_lists(struct throtl_data *td,
249 struct throtl_grp *tg, struct blkio_cgroup *blkcg)
250{
251 __throtl_tg_fill_dev_details(td, tg);
252
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400253 /* Add group onto cgroup list */
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400254 blkiocg_add_blkio_group(blkcg, &tg->blkg, (void *)td,
Vivek Goyal269f5412011-05-19 15:38:25 -0400255 tg->blkg.dev, BLKIO_POLICY_THROTL);
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400256
257 tg->bps[READ] = blkcg_get_read_bps(blkcg, tg->blkg.dev);
258 tg->bps[WRITE] = blkcg_get_write_bps(blkcg, tg->blkg.dev);
259 tg->iops[READ] = blkcg_get_read_iops(blkcg, tg->blkg.dev);
260 tg->iops[WRITE] = blkcg_get_write_iops(blkcg, tg->blkg.dev);
261
262 throtl_add_group_to_td_list(td, tg);
263}
264
265/* Should be called without queue lock and outside of rcu period */
266static struct throtl_grp *throtl_alloc_tg(struct throtl_data *td)
267{
268 struct throtl_grp *tg = NULL;
Vivek Goyal5624a4e2011-05-19 15:38:28 -0400269 int ret;
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400270
271 tg = kzalloc_node(sizeof(*tg), GFP_ATOMIC, td->queue->node);
272 if (!tg)
273 return NULL;
274
Vivek Goyal5624a4e2011-05-19 15:38:28 -0400275 ret = blkio_alloc_blkg_stats(&tg->blkg);
276
277 if (ret) {
278 kfree(tg);
279 return NULL;
280 }
281
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400282 throtl_init_group(tg);
283 return tg;
284}
285
286static struct
287throtl_grp *throtl_find_tg(struct throtl_data *td, struct blkio_cgroup *blkcg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400288{
Vivek Goyale43473b2010-09-15 17:06:35 -0400289 struct throtl_grp *tg = NULL;
290 void *key = td;
Vivek Goyale43473b2010-09-15 17:06:35 -0400291
292 /*
Vivek Goyalbe2c6b12011-01-19 08:25:02 -0700293 * This is the common case when there are no blkio cgroups.
294 * Avoid lookup in this case
295 */
296 if (blkcg == &blkio_root_cgroup)
Vivek Goyal29b12582011-05-19 15:38:24 -0400297 tg = td->root_tg;
Vivek Goyalbe2c6b12011-01-19 08:25:02 -0700298 else
299 tg = tg_of_blkg(blkiocg_lookup_group(blkcg, key));
Vivek Goyale43473b2010-09-15 17:06:35 -0400300
Vivek Goyal269f5412011-05-19 15:38:25 -0400301 __throtl_tg_fill_dev_details(td, tg);
Vivek Goyale43473b2010-09-15 17:06:35 -0400302 return tg;
303}
304
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400305/*
306 * This function returns with queue lock unlocked in case of error, like
307 * request queue is no more
308 */
Vivek Goyale43473b2010-09-15 17:06:35 -0400309static struct throtl_grp * throtl_get_tg(struct throtl_data *td)
310{
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400311 struct throtl_grp *tg = NULL, *__tg = NULL;
Vivek Goyal70087dc2011-05-16 15:24:08 +0200312 struct blkio_cgroup *blkcg;
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400313 struct request_queue *q = td->queue;
Vivek Goyale43473b2010-09-15 17:06:35 -0400314
315 rcu_read_lock();
Vivek Goyal70087dc2011-05-16 15:24:08 +0200316 blkcg = task_blkio_cgroup(current);
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400317 tg = throtl_find_tg(td, blkcg);
318 if (tg) {
319 rcu_read_unlock();
320 return tg;
321 }
322
323 /*
324 * Need to allocate a group. Allocation of group also needs allocation
325 * of per cpu stats which in-turn takes a mutex() and can block. Hence
326 * we need to drop rcu lock and queue_lock before we call alloc
327 *
328 * Take the request queue reference to make sure queue does not
329 * go away once we return from allocation.
330 */
331 blk_get_queue(q);
332 rcu_read_unlock();
333 spin_unlock_irq(q->queue_lock);
334
335 tg = throtl_alloc_tg(td);
336 /*
337 * We might have slept in group allocation. Make sure queue is not
338 * dead
339 */
340 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
341 blk_put_queue(q);
342 if (tg)
343 kfree(tg);
344
345 return ERR_PTR(-ENODEV);
346 }
347 blk_put_queue(q);
348
349 /* Group allocated and queue is still alive. take the lock */
350 spin_lock_irq(q->queue_lock);
351
352 /*
353 * Initialize the new group. After sleeping, read the blkcg again.
354 */
355 rcu_read_lock();
356 blkcg = task_blkio_cgroup(current);
357
358 /*
359 * If some other thread already allocated the group while we were
360 * not holding queue lock, free up the group
361 */
362 __tg = throtl_find_tg(td, blkcg);
363
364 if (__tg) {
365 kfree(tg);
366 rcu_read_unlock();
367 return __tg;
368 }
369
370 /* Group allocation failed. Account the IO to root group */
371 if (!tg) {
Vivek Goyal29b12582011-05-19 15:38:24 -0400372 tg = td->root_tg;
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400373 return tg;
374 }
375
376 throtl_init_add_tg_lists(td, tg, blkcg);
Vivek Goyale43473b2010-09-15 17:06:35 -0400377 rcu_read_unlock();
378 return tg;
379}
380
381static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root)
382{
383 /* Service tree is empty */
384 if (!root->count)
385 return NULL;
386
387 if (!root->left)
388 root->left = rb_first(&root->rb);
389
390 if (root->left)
391 return rb_entry_tg(root->left);
392
393 return NULL;
394}
395
396static void rb_erase_init(struct rb_node *n, struct rb_root *root)
397{
398 rb_erase(n, root);
399 RB_CLEAR_NODE(n);
400}
401
402static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root)
403{
404 if (root->left == n)
405 root->left = NULL;
406 rb_erase_init(n, &root->rb);
407 --root->count;
408}
409
410static void update_min_dispatch_time(struct throtl_rb_root *st)
411{
412 struct throtl_grp *tg;
413
414 tg = throtl_rb_first(st);
415 if (!tg)
416 return;
417
418 st->min_disptime = tg->disptime;
419}
420
421static void
422tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg)
423{
424 struct rb_node **node = &st->rb.rb_node;
425 struct rb_node *parent = NULL;
426 struct throtl_grp *__tg;
427 unsigned long key = tg->disptime;
428 int left = 1;
429
430 while (*node != NULL) {
431 parent = *node;
432 __tg = rb_entry_tg(parent);
433
434 if (time_before(key, __tg->disptime))
435 node = &parent->rb_left;
436 else {
437 node = &parent->rb_right;
438 left = 0;
439 }
440 }
441
442 if (left)
443 st->left = &tg->rb_node;
444
445 rb_link_node(&tg->rb_node, parent, node);
446 rb_insert_color(&tg->rb_node, &st->rb);
447}
448
449static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
450{
451 struct throtl_rb_root *st = &td->tg_service_tree;
452
453 tg_service_tree_add(st, tg);
454 throtl_mark_tg_on_rr(tg);
455 st->count++;
456}
457
458static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
459{
460 if (!throtl_tg_on_rr(tg))
461 __throtl_enqueue_tg(td, tg);
462}
463
464static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
465{
466 throtl_rb_erase(&tg->rb_node, &td->tg_service_tree);
467 throtl_clear_tg_on_rr(tg);
468}
469
470static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
471{
472 if (throtl_tg_on_rr(tg))
473 __throtl_dequeue_tg(td, tg);
474}
475
476static void throtl_schedule_next_dispatch(struct throtl_data *td)
477{
478 struct throtl_rb_root *st = &td->tg_service_tree;
479
480 /*
481 * If there are more bios pending, schedule more work.
482 */
483 if (!total_nr_queued(td))
484 return;
485
486 BUG_ON(!st->count);
487
488 update_min_dispatch_time(st);
489
490 if (time_before_eq(st->min_disptime, jiffies))
Vivek Goyal450adcb2011-03-01 13:40:54 -0500491 throtl_schedule_delayed_work(td, 0);
Vivek Goyale43473b2010-09-15 17:06:35 -0400492 else
Vivek Goyal450adcb2011-03-01 13:40:54 -0500493 throtl_schedule_delayed_work(td, (st->min_disptime - jiffies));
Vivek Goyale43473b2010-09-15 17:06:35 -0400494}
495
496static inline void
497throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
498{
499 tg->bytes_disp[rw] = 0;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400500 tg->io_disp[rw] = 0;
Vivek Goyale43473b2010-09-15 17:06:35 -0400501 tg->slice_start[rw] = jiffies;
502 tg->slice_end[rw] = jiffies + throtl_slice;
503 throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
504 rw == READ ? 'R' : 'W', tg->slice_start[rw],
505 tg->slice_end[rw], jiffies);
506}
507
Vivek Goyald1ae8ff2010-12-01 19:34:46 +0100508static inline void throtl_set_slice_end(struct throtl_data *td,
509 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
510{
511 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
512}
513
Vivek Goyale43473b2010-09-15 17:06:35 -0400514static inline void throtl_extend_slice(struct throtl_data *td,
515 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
516{
517 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
518 throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
519 rw == READ ? 'R' : 'W', tg->slice_start[rw],
520 tg->slice_end[rw], jiffies);
521}
522
523/* Determine if previously allocated or extended slice is complete or not */
524static bool
525throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
526{
527 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
528 return 0;
529
530 return 1;
531}
532
533/* Trim the used slices and adjust slice start accordingly */
534static inline void
535throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
536{
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200537 unsigned long nr_slices, time_elapsed, io_trim;
538 u64 bytes_trim, tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400539
540 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
541
542 /*
543 * If bps are unlimited (-1), then time slice don't get
544 * renewed. Don't try to trim the slice if slice is used. A new
545 * slice will start when appropriate.
546 */
547 if (throtl_slice_used(td, tg, rw))
548 return;
549
Vivek Goyald1ae8ff2010-12-01 19:34:46 +0100550 /*
551 * A bio has been dispatched. Also adjust slice_end. It might happen
552 * that initially cgroup limit was very low resulting in high
553 * slice_end, but later limit was bumped up and bio was dispached
554 * sooner, then we need to reduce slice_end. A high bogus slice_end
555 * is bad because it does not allow new slice to start.
556 */
557
558 throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice);
559
Vivek Goyale43473b2010-09-15 17:06:35 -0400560 time_elapsed = jiffies - tg->slice_start[rw];
561
562 nr_slices = time_elapsed / throtl_slice;
563
564 if (!nr_slices)
565 return;
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200566 tmp = tg->bps[rw] * throtl_slice * nr_slices;
567 do_div(tmp, HZ);
568 bytes_trim = tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400569
Vivek Goyal8e89d132010-09-15 17:06:37 -0400570 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
Vivek Goyale43473b2010-09-15 17:06:35 -0400571
Vivek Goyal8e89d132010-09-15 17:06:37 -0400572 if (!bytes_trim && !io_trim)
Vivek Goyale43473b2010-09-15 17:06:35 -0400573 return;
574
575 if (tg->bytes_disp[rw] >= bytes_trim)
576 tg->bytes_disp[rw] -= bytes_trim;
577 else
578 tg->bytes_disp[rw] = 0;
579
Vivek Goyal8e89d132010-09-15 17:06:37 -0400580 if (tg->io_disp[rw] >= io_trim)
581 tg->io_disp[rw] -= io_trim;
582 else
583 tg->io_disp[rw] = 0;
584
Vivek Goyale43473b2010-09-15 17:06:35 -0400585 tg->slice_start[rw] += nr_slices * throtl_slice;
586
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200587 throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
Vivek Goyale43473b2010-09-15 17:06:35 -0400588 " start=%lu end=%lu jiffies=%lu",
Vivek Goyal8e89d132010-09-15 17:06:37 -0400589 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
Vivek Goyale43473b2010-09-15 17:06:35 -0400590 tg->slice_start[rw], tg->slice_end[rw], jiffies);
591}
592
Vivek Goyal8e89d132010-09-15 17:06:37 -0400593static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
594 struct bio *bio, unsigned long *wait)
Vivek Goyale43473b2010-09-15 17:06:35 -0400595{
596 bool rw = bio_data_dir(bio);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400597 unsigned int io_allowed;
Vivek Goyale43473b2010-09-15 17:06:35 -0400598 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
Vivek Goyalc49c06e2010-10-01 21:16:42 +0200599 u64 tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400600
Vivek Goyal8e89d132010-09-15 17:06:37 -0400601 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
Vivek Goyale43473b2010-09-15 17:06:35 -0400602
Vivek Goyal8e89d132010-09-15 17:06:37 -0400603 /* Slice has just started. Consider one slice interval */
604 if (!jiffy_elapsed)
605 jiffy_elapsed_rnd = throtl_slice;
606
607 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
608
Vivek Goyalc49c06e2010-10-01 21:16:42 +0200609 /*
610 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
611 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
612 * will allow dispatch after 1 second and after that slice should
613 * have been trimmed.
614 */
615
616 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
617 do_div(tmp, HZ);
618
619 if (tmp > UINT_MAX)
620 io_allowed = UINT_MAX;
621 else
622 io_allowed = tmp;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400623
624 if (tg->io_disp[rw] + 1 <= io_allowed) {
Vivek Goyale43473b2010-09-15 17:06:35 -0400625 if (wait)
626 *wait = 0;
627 return 1;
628 }
629
Vivek Goyal8e89d132010-09-15 17:06:37 -0400630 /* Calc approx time to dispatch */
631 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
632
633 if (jiffy_wait > jiffy_elapsed)
634 jiffy_wait = jiffy_wait - jiffy_elapsed;
635 else
636 jiffy_wait = 1;
637
638 if (wait)
639 *wait = jiffy_wait;
640 return 0;
641}
642
643static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
644 struct bio *bio, unsigned long *wait)
645{
646 bool rw = bio_data_dir(bio);
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200647 u64 bytes_allowed, extra_bytes, tmp;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400648 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
Vivek Goyale43473b2010-09-15 17:06:35 -0400649
650 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
651
652 /* Slice has just started. Consider one slice interval */
653 if (!jiffy_elapsed)
654 jiffy_elapsed_rnd = throtl_slice;
655
656 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
657
Vivek Goyal5e901a22010-10-01 21:16:38 +0200658 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
659 do_div(tmp, HZ);
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200660 bytes_allowed = tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400661
662 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
663 if (wait)
664 *wait = 0;
665 return 1;
666 }
667
668 /* Calc approx time to dispatch */
669 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
670 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
671
672 if (!jiffy_wait)
673 jiffy_wait = 1;
674
675 /*
676 * This wait time is without taking into consideration the rounding
677 * up we did. Add that time also.
678 */
679 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
Vivek Goyale43473b2010-09-15 17:06:35 -0400680 if (wait)
681 *wait = jiffy_wait;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400682 return 0;
683}
Vivek Goyale43473b2010-09-15 17:06:35 -0400684
Vivek Goyalaf75cd32011-05-19 15:38:31 -0400685static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
686 if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
687 return 1;
688 return 0;
689}
690
Vivek Goyal8e89d132010-09-15 17:06:37 -0400691/*
692 * Returns whether one can dispatch a bio or not. Also returns approx number
693 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
694 */
695static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
696 struct bio *bio, unsigned long *wait)
697{
698 bool rw = bio_data_dir(bio);
699 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
700
701 /*
702 * Currently whole state machine of group depends on first bio
703 * queued in the group bio list. So one should not be calling
704 * this function with a different bio if there are other bios
705 * queued.
706 */
707 BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
708
709 /* If tg->bps = -1, then BW is unlimited */
710 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
711 if (wait)
712 *wait = 0;
713 return 1;
714 }
715
716 /*
717 * If previous slice expired, start a new one otherwise renew/extend
718 * existing slice to make sure it is at least throtl_slice interval
719 * long since now.
720 */
721 if (throtl_slice_used(td, tg, rw))
722 throtl_start_new_slice(td, tg, rw);
723 else {
724 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
725 throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
726 }
727
728 if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
729 && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
730 if (wait)
731 *wait = 0;
732 return 1;
733 }
734
735 max_wait = max(bps_wait, iops_wait);
736
737 if (wait)
738 *wait = max_wait;
739
740 if (time_before(tg->slice_end[rw], jiffies + max_wait))
741 throtl_extend_slice(td, tg, rw, jiffies + max_wait);
Vivek Goyale43473b2010-09-15 17:06:35 -0400742
743 return 0;
744}
745
746static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
747{
748 bool rw = bio_data_dir(bio);
749 bool sync = bio->bi_rw & REQ_SYNC;
750
751 /* Charge the bio to the group */
752 tg->bytes_disp[rw] += bio->bi_size;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400753 tg->io_disp[rw]++;
Vivek Goyale43473b2010-09-15 17:06:35 -0400754
Vivek Goyale43473b2010-09-15 17:06:35 -0400755 blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync);
Vivek Goyale43473b2010-09-15 17:06:35 -0400756}
757
758static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
759 struct bio *bio)
760{
761 bool rw = bio_data_dir(bio);
762
763 bio_list_add(&tg->bio_lists[rw], bio);
764 /* Take a bio reference on tg */
765 throtl_ref_get_tg(tg);
766 tg->nr_queued[rw]++;
767 td->nr_queued[rw]++;
768 throtl_enqueue_tg(td, tg);
769}
770
771static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
772{
773 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
774 struct bio *bio;
775
776 if ((bio = bio_list_peek(&tg->bio_lists[READ])))
777 tg_may_dispatch(td, tg, bio, &read_wait);
778
779 if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
780 tg_may_dispatch(td, tg, bio, &write_wait);
781
782 min_wait = min(read_wait, write_wait);
783 disptime = jiffies + min_wait;
784
Vivek Goyale43473b2010-09-15 17:06:35 -0400785 /* Update dispatch time */
786 throtl_dequeue_tg(td, tg);
787 tg->disptime = disptime;
788 throtl_enqueue_tg(td, tg);
789}
790
791static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
792 bool rw, struct bio_list *bl)
793{
794 struct bio *bio;
795
796 bio = bio_list_pop(&tg->bio_lists[rw]);
797 tg->nr_queued[rw]--;
798 /* Drop bio reference on tg */
799 throtl_put_tg(tg);
800
801 BUG_ON(td->nr_queued[rw] <= 0);
802 td->nr_queued[rw]--;
803
804 throtl_charge_bio(tg, bio);
805 bio_list_add(bl, bio);
806 bio->bi_rw |= REQ_THROTTLED;
807
808 throtl_trim_slice(td, tg, rw);
809}
810
811static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
812 struct bio_list *bl)
813{
814 unsigned int nr_reads = 0, nr_writes = 0;
815 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
Vivek Goyalc2f68052010-11-15 19:32:42 +0100816 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
Vivek Goyale43473b2010-09-15 17:06:35 -0400817 struct bio *bio;
818
819 /* Try to dispatch 75% READS and 25% WRITES */
820
821 while ((bio = bio_list_peek(&tg->bio_lists[READ]))
822 && tg_may_dispatch(td, tg, bio, NULL)) {
823
824 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
825 nr_reads++;
826
827 if (nr_reads >= max_nr_reads)
828 break;
829 }
830
831 while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
832 && tg_may_dispatch(td, tg, bio, NULL)) {
833
834 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
835 nr_writes++;
836
837 if (nr_writes >= max_nr_writes)
838 break;
839 }
840
841 return nr_reads + nr_writes;
842}
843
844static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
845{
846 unsigned int nr_disp = 0;
847 struct throtl_grp *tg;
848 struct throtl_rb_root *st = &td->tg_service_tree;
849
850 while (1) {
851 tg = throtl_rb_first(st);
852
853 if (!tg)
854 break;
855
856 if (time_before(jiffies, tg->disptime))
857 break;
858
859 throtl_dequeue_tg(td, tg);
860
861 nr_disp += throtl_dispatch_tg(td, tg, bl);
862
863 if (tg->nr_queued[0] || tg->nr_queued[1]) {
864 tg_update_disptime(td, tg);
865 throtl_enqueue_tg(td, tg);
866 }
867
868 if (nr_disp >= throtl_quantum)
869 break;
870 }
871
872 return nr_disp;
873}
874
Vivek Goyalfe071432010-10-01 14:49:49 +0200875static void throtl_process_limit_change(struct throtl_data *td)
876{
877 struct throtl_grp *tg;
878 struct hlist_node *pos, *n;
879
Vivek Goyalde701c72011-03-07 21:09:32 +0100880 if (!td->limits_changed)
Vivek Goyalfe071432010-10-01 14:49:49 +0200881 return;
882
Vivek Goyalde701c72011-03-07 21:09:32 +0100883 xchg(&td->limits_changed, false);
Vivek Goyalfe071432010-10-01 14:49:49 +0200884
Vivek Goyalde701c72011-03-07 21:09:32 +0100885 throtl_log(td, "limits changed");
Vivek Goyalfe071432010-10-01 14:49:49 +0200886
Vivek Goyal04a6b512010-12-01 19:34:52 +0100887 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
Vivek Goyalde701c72011-03-07 21:09:32 +0100888 if (!tg->limits_changed)
889 continue;
Vivek Goyalfe071432010-10-01 14:49:49 +0200890
Vivek Goyalde701c72011-03-07 21:09:32 +0100891 if (!xchg(&tg->limits_changed, false))
892 continue;
893
894 throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
895 " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
896 tg->iops[READ], tg->iops[WRITE]);
897
Vivek Goyal04521db2011-03-22 21:54:29 +0100898 /*
899 * Restart the slices for both READ and WRITES. It
900 * might happen that a group's limit are dropped
901 * suddenly and we don't want to account recently
902 * dispatched IO with new low rate
903 */
904 throtl_start_new_slice(td, tg, 0);
905 throtl_start_new_slice(td, tg, 1);
906
Vivek Goyalde701c72011-03-07 21:09:32 +0100907 if (throtl_tg_on_rr(tg))
908 tg_update_disptime(td, tg);
909 }
Vivek Goyalfe071432010-10-01 14:49:49 +0200910}
911
Vivek Goyale43473b2010-09-15 17:06:35 -0400912/* Dispatch throttled bios. Should be called without queue lock held. */
913static int throtl_dispatch(struct request_queue *q)
914{
915 struct throtl_data *td = q->td;
916 unsigned int nr_disp = 0;
917 struct bio_list bio_list_on_stack;
918 struct bio *bio;
Vivek Goyal69d60eb2011-03-09 08:27:37 +0100919 struct blk_plug plug;
Vivek Goyale43473b2010-09-15 17:06:35 -0400920
921 spin_lock_irq(q->queue_lock);
922
Vivek Goyalfe071432010-10-01 14:49:49 +0200923 throtl_process_limit_change(td);
924
Vivek Goyale43473b2010-09-15 17:06:35 -0400925 if (!total_nr_queued(td))
926 goto out;
927
928 bio_list_init(&bio_list_on_stack);
929
Joe Perchesd2f31a52011-06-13 20:19:27 +0200930 throtl_log(td, "dispatch nr_queued=%u read=%u write=%u",
Vivek Goyale43473b2010-09-15 17:06:35 -0400931 total_nr_queued(td), td->nr_queued[READ],
932 td->nr_queued[WRITE]);
933
934 nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
935
936 if (nr_disp)
937 throtl_log(td, "bios disp=%u", nr_disp);
938
939 throtl_schedule_next_dispatch(td);
940out:
941 spin_unlock_irq(q->queue_lock);
942
943 /*
944 * If we dispatched some requests, unplug the queue to make sure
945 * immediate dispatch
946 */
947 if (nr_disp) {
Vivek Goyal69d60eb2011-03-09 08:27:37 +0100948 blk_start_plug(&plug);
Vivek Goyale43473b2010-09-15 17:06:35 -0400949 while((bio = bio_list_pop(&bio_list_on_stack)))
950 generic_make_request(bio);
Vivek Goyal69d60eb2011-03-09 08:27:37 +0100951 blk_finish_plug(&plug);
Vivek Goyale43473b2010-09-15 17:06:35 -0400952 }
953 return nr_disp;
954}
955
956void blk_throtl_work(struct work_struct *work)
957{
958 struct throtl_data *td = container_of(work, struct throtl_data,
959 throtl_work.work);
960 struct request_queue *q = td->queue;
961
962 throtl_dispatch(q);
963}
964
965/* Call with queue lock held */
Vivek Goyal450adcb2011-03-01 13:40:54 -0500966static void
967throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay)
Vivek Goyale43473b2010-09-15 17:06:35 -0400968{
969
Vivek Goyale43473b2010-09-15 17:06:35 -0400970 struct delayed_work *dwork = &td->throtl_work;
971
Vivek Goyal04521db2011-03-22 21:54:29 +0100972 /* schedule work if limits changed even if no bio is queued */
Joe Perchesd2f31a52011-06-13 20:19:27 +0200973 if (total_nr_queued(td) || td->limits_changed) {
Vivek Goyale43473b2010-09-15 17:06:35 -0400974 /*
975 * We might have a work scheduled to be executed in future.
976 * Cancel that and schedule a new one.
977 */
978 __cancel_delayed_work(dwork);
Vivek Goyal450adcb2011-03-01 13:40:54 -0500979 queue_delayed_work(kthrotld_workqueue, dwork, delay);
Vivek Goyale43473b2010-09-15 17:06:35 -0400980 throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
981 delay, jiffies);
982 }
983}
Vivek Goyale43473b2010-09-15 17:06:35 -0400984
985static void
986throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg)
987{
988 /* Something wrong if we are trying to remove same group twice */
989 BUG_ON(hlist_unhashed(&tg->tg_node));
990
991 hlist_del_init(&tg->tg_node);
992
993 /*
994 * Put the reference taken at the time of creation so that when all
995 * queues are gone, group can be destroyed.
996 */
997 throtl_put_tg(tg);
998 td->nr_undestroyed_grps--;
999}
1000
1001static void throtl_release_tgs(struct throtl_data *td)
1002{
1003 struct hlist_node *pos, *n;
1004 struct throtl_grp *tg;
1005
1006 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
1007 /*
1008 * If cgroup removal path got to blk_group first and removed
1009 * it from cgroup list, then it will take care of destroying
1010 * cfqg also.
1011 */
1012 if (!blkiocg_del_blkio_group(&tg->blkg))
1013 throtl_destroy_tg(td, tg);
1014 }
1015}
1016
1017static void throtl_td_free(struct throtl_data *td)
1018{
1019 kfree(td);
1020}
1021
1022/*
1023 * Blk cgroup controller notification saying that blkio_group object is being
1024 * delinked as associated cgroup object is going away. That also means that
1025 * no new IO will come in this group. So get rid of this group as soon as
1026 * any pending IO in the group is finished.
1027 *
1028 * This function is called under rcu_read_lock(). key is the rcu protected
1029 * pointer. That means "key" is a valid throtl_data pointer as long as we are
1030 * rcu read lock.
1031 *
1032 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1033 * it should not be NULL as even if queue was going away, cgroup deltion
1034 * path got to it first.
1035 */
1036void throtl_unlink_blkio_group(void *key, struct blkio_group *blkg)
1037{
1038 unsigned long flags;
1039 struct throtl_data *td = key;
1040
1041 spin_lock_irqsave(td->queue->queue_lock, flags);
1042 throtl_destroy_tg(td, tg_of_blkg(blkg));
1043 spin_unlock_irqrestore(td->queue->queue_lock, flags);
1044}
1045
Vivek Goyalde701c72011-03-07 21:09:32 +01001046static void throtl_update_blkio_group_common(struct throtl_data *td,
1047 struct throtl_grp *tg)
1048{
1049 xchg(&tg->limits_changed, true);
1050 xchg(&td->limits_changed, true);
1051 /* Schedule a work now to process the limit change */
1052 throtl_schedule_delayed_work(td, 0);
1053}
1054
Vivek Goyalfe071432010-10-01 14:49:49 +02001055/*
1056 * For all update functions, key should be a valid pointer because these
1057 * update functions are called under blkcg_lock, that means, blkg is
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001058 * valid and in turn key is valid. queue exit path can not race because
Vivek Goyalfe071432010-10-01 14:49:49 +02001059 * of blkcg_lock
1060 *
1061 * Can not take queue lock in update functions as queue lock under blkcg_lock
1062 * is not allowed. Under other paths we take blkcg_lock under queue_lock.
1063 */
1064static void throtl_update_blkio_group_read_bps(void *key,
1065 struct blkio_group *blkg, u64 read_bps)
Vivek Goyale43473b2010-09-15 17:06:35 -04001066{
Vivek Goyalfe071432010-10-01 14:49:49 +02001067 struct throtl_data *td = key;
Vivek Goyalde701c72011-03-07 21:09:32 +01001068 struct throtl_grp *tg = tg_of_blkg(blkg);
Vivek Goyalfe071432010-10-01 14:49:49 +02001069
Vivek Goyalde701c72011-03-07 21:09:32 +01001070 tg->bps[READ] = read_bps;
1071 throtl_update_blkio_group_common(td, tg);
Vivek Goyale43473b2010-09-15 17:06:35 -04001072}
1073
Vivek Goyalfe071432010-10-01 14:49:49 +02001074static void throtl_update_blkio_group_write_bps(void *key,
1075 struct blkio_group *blkg, u64 write_bps)
Vivek Goyale43473b2010-09-15 17:06:35 -04001076{
Vivek Goyalfe071432010-10-01 14:49:49 +02001077 struct throtl_data *td = key;
Vivek Goyalde701c72011-03-07 21:09:32 +01001078 struct throtl_grp *tg = tg_of_blkg(blkg);
Vivek Goyalfe071432010-10-01 14:49:49 +02001079
Vivek Goyalde701c72011-03-07 21:09:32 +01001080 tg->bps[WRITE] = write_bps;
1081 throtl_update_blkio_group_common(td, tg);
Vivek Goyale43473b2010-09-15 17:06:35 -04001082}
1083
Vivek Goyalfe071432010-10-01 14:49:49 +02001084static void throtl_update_blkio_group_read_iops(void *key,
1085 struct blkio_group *blkg, unsigned int read_iops)
Vivek Goyal8e89d132010-09-15 17:06:37 -04001086{
Vivek Goyalfe071432010-10-01 14:49:49 +02001087 struct throtl_data *td = key;
Vivek Goyalde701c72011-03-07 21:09:32 +01001088 struct throtl_grp *tg = tg_of_blkg(blkg);
Vivek Goyalfe071432010-10-01 14:49:49 +02001089
Vivek Goyalde701c72011-03-07 21:09:32 +01001090 tg->iops[READ] = read_iops;
1091 throtl_update_blkio_group_common(td, tg);
Vivek Goyal8e89d132010-09-15 17:06:37 -04001092}
1093
Vivek Goyalfe071432010-10-01 14:49:49 +02001094static void throtl_update_blkio_group_write_iops(void *key,
1095 struct blkio_group *blkg, unsigned int write_iops)
Vivek Goyal8e89d132010-09-15 17:06:37 -04001096{
Vivek Goyalfe071432010-10-01 14:49:49 +02001097 struct throtl_data *td = key;
Vivek Goyalde701c72011-03-07 21:09:32 +01001098 struct throtl_grp *tg = tg_of_blkg(blkg);
Vivek Goyalfe071432010-10-01 14:49:49 +02001099
Vivek Goyalde701c72011-03-07 21:09:32 +01001100 tg->iops[WRITE] = write_iops;
1101 throtl_update_blkio_group_common(td, tg);
Vivek Goyal8e89d132010-09-15 17:06:37 -04001102}
1103
Vivek Goyalda527772011-03-02 19:05:33 -05001104static void throtl_shutdown_wq(struct request_queue *q)
Vivek Goyale43473b2010-09-15 17:06:35 -04001105{
1106 struct throtl_data *td = q->td;
1107
1108 cancel_delayed_work_sync(&td->throtl_work);
1109}
1110
1111static struct blkio_policy_type blkio_policy_throtl = {
1112 .ops = {
1113 .blkio_unlink_group_fn = throtl_unlink_blkio_group,
1114 .blkio_update_group_read_bps_fn =
1115 throtl_update_blkio_group_read_bps,
1116 .blkio_update_group_write_bps_fn =
1117 throtl_update_blkio_group_write_bps,
Vivek Goyal8e89d132010-09-15 17:06:37 -04001118 .blkio_update_group_read_iops_fn =
1119 throtl_update_blkio_group_read_iops,
1120 .blkio_update_group_write_iops_fn =
1121 throtl_update_blkio_group_write_iops,
Vivek Goyale43473b2010-09-15 17:06:35 -04001122 },
Vivek Goyal8e89d132010-09-15 17:06:37 -04001123 .plid = BLKIO_POLICY_THROTL,
Vivek Goyale43473b2010-09-15 17:06:35 -04001124};
1125
1126int blk_throtl_bio(struct request_queue *q, struct bio **biop)
1127{
1128 struct throtl_data *td = q->td;
1129 struct throtl_grp *tg;
1130 struct bio *bio = *biop;
1131 bool rw = bio_data_dir(bio), update_disptime = true;
Vivek Goyalaf75cd32011-05-19 15:38:31 -04001132 struct blkio_cgroup *blkcg;
Vivek Goyale43473b2010-09-15 17:06:35 -04001133
1134 if (bio->bi_rw & REQ_THROTTLED) {
1135 bio->bi_rw &= ~REQ_THROTTLED;
1136 return 0;
1137 }
1138
Vivek Goyalaf75cd32011-05-19 15:38:31 -04001139 /*
1140 * A throtl_grp pointer retrieved under rcu can be used to access
1141 * basic fields like stats and io rates. If a group has no rules,
1142 * just update the dispatch stats in lockless manner and return.
1143 */
1144
1145 rcu_read_lock();
1146 blkcg = task_blkio_cgroup(current);
1147 tg = throtl_find_tg(td, blkcg);
1148 if (tg) {
1149 throtl_tg_fill_dev_details(td, tg);
1150
1151 if (tg_no_rule_group(tg, rw)) {
1152 blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size,
1153 rw, bio->bi_rw & REQ_SYNC);
1154 rcu_read_unlock();
1155 return 0;
1156 }
1157 }
1158 rcu_read_unlock();
1159
1160 /*
1161 * Either group has not been allocated yet or it is not an unlimited
1162 * IO group
1163 */
1164
Vivek Goyale43473b2010-09-15 17:06:35 -04001165 spin_lock_irq(q->queue_lock);
1166 tg = throtl_get_tg(td);
1167
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001168 if (IS_ERR(tg)) {
1169 if (PTR_ERR(tg) == -ENODEV) {
1170 /*
1171 * Queue is gone. No queue lock held here.
1172 */
1173 return -ENODEV;
1174 }
1175 }
1176
Vivek Goyale43473b2010-09-15 17:06:35 -04001177 if (tg->nr_queued[rw]) {
1178 /*
1179 * There is already another bio queued in same dir. No
1180 * need to update dispatch time.
1181 */
Vivek Goyal231d7042011-03-07 21:05:14 +01001182 update_disptime = false;
Vivek Goyale43473b2010-09-15 17:06:35 -04001183 goto queue_bio;
Vivek Goyalde701c72011-03-07 21:09:32 +01001184
Vivek Goyale43473b2010-09-15 17:06:35 -04001185 }
1186
1187 /* Bio is with-in rate limit of group */
1188 if (tg_may_dispatch(td, tg, bio, NULL)) {
1189 throtl_charge_bio(tg, bio);
Vivek Goyal04521db2011-03-22 21:54:29 +01001190
1191 /*
1192 * We need to trim slice even when bios are not being queued
1193 * otherwise it might happen that a bio is not queued for
1194 * a long time and slice keeps on extending and trim is not
1195 * called for a long time. Now if limits are reduced suddenly
1196 * we take into account all the IO dispatched so far at new
1197 * low rate and * newly queued IO gets a really long dispatch
1198 * time.
1199 *
1200 * So keep on trimming slice even if bio is not queued.
1201 */
1202 throtl_trim_slice(td, tg, rw);
Vivek Goyale43473b2010-09-15 17:06:35 -04001203 goto out;
1204 }
1205
1206queue_bio:
Joe Perchesfd16d262011-06-13 10:42:49 +02001207 throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
Vivek Goyal8e89d132010-09-15 17:06:37 -04001208 " iodisp=%u iops=%u queued=%d/%d",
1209 rw == READ ? 'R' : 'W',
Vivek Goyale43473b2010-09-15 17:06:35 -04001210 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
Vivek Goyal8e89d132010-09-15 17:06:37 -04001211 tg->io_disp[rw], tg->iops[rw],
Vivek Goyale43473b2010-09-15 17:06:35 -04001212 tg->nr_queued[READ], tg->nr_queued[WRITE]);
1213
1214 throtl_add_bio_tg(q->td, tg, bio);
1215 *biop = NULL;
1216
1217 if (update_disptime) {
1218 tg_update_disptime(td, tg);
1219 throtl_schedule_next_dispatch(td);
1220 }
1221
1222out:
1223 spin_unlock_irq(q->queue_lock);
1224 return 0;
1225}
1226
1227int blk_throtl_init(struct request_queue *q)
1228{
1229 struct throtl_data *td;
1230 struct throtl_grp *tg;
1231
1232 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1233 if (!td)
1234 return -ENOMEM;
1235
1236 INIT_HLIST_HEAD(&td->tg_list);
1237 td->tg_service_tree = THROTL_RB_ROOT;
Vivek Goyalde701c72011-03-07 21:09:32 +01001238 td->limits_changed = false;
Vivek Goyala29a1712011-05-19 15:38:19 -04001239 INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);
Vivek Goyale43473b2010-09-15 17:06:35 -04001240
Vivek Goyal29b12582011-05-19 15:38:24 -04001241 /* alloc and Init root group. */
1242 td->queue = q;
1243 tg = throtl_alloc_tg(td);
Vivek Goyal02977e42010-10-01 14:49:48 +02001244
Vivek Goyal29b12582011-05-19 15:38:24 -04001245 if (!tg) {
1246 kfree(td);
1247 return -ENOMEM;
1248 }
1249
1250 td->root_tg = tg;
Vivek Goyale43473b2010-09-15 17:06:35 -04001251
1252 rcu_read_lock();
Vivek Goyal5617cbe2011-05-19 15:38:26 -04001253 throtl_init_add_tg_lists(td, tg, &blkio_root_cgroup);
Vivek Goyale43473b2010-09-15 17:06:35 -04001254 rcu_read_unlock();
1255
1256 /* Attach throtl data to request queue */
Vivek Goyale43473b2010-09-15 17:06:35 -04001257 q->td = td;
1258 return 0;
1259}
1260
1261void blk_throtl_exit(struct request_queue *q)
1262{
1263 struct throtl_data *td = q->td;
1264 bool wait = false;
1265
1266 BUG_ON(!td);
1267
Vivek Goyalda527772011-03-02 19:05:33 -05001268 throtl_shutdown_wq(q);
Vivek Goyale43473b2010-09-15 17:06:35 -04001269
1270 spin_lock_irq(q->queue_lock);
1271 throtl_release_tgs(td);
Vivek Goyale43473b2010-09-15 17:06:35 -04001272
1273 /* If there are other groups */
Vivek Goyal02977e42010-10-01 14:49:48 +02001274 if (td->nr_undestroyed_grps > 0)
Vivek Goyale43473b2010-09-15 17:06:35 -04001275 wait = true;
1276
1277 spin_unlock_irq(q->queue_lock);
1278
1279 /*
1280 * Wait for tg->blkg->key accessors to exit their grace periods.
1281 * Do this wait only if there are other undestroyed groups out
1282 * there (other than root group). This can happen if cgroup deletion
1283 * path claimed the responsibility of cleaning up a group before
1284 * queue cleanup code get to the group.
1285 *
1286 * Do not call synchronize_rcu() unconditionally as there are drivers
1287 * which create/delete request queue hundreds of times during scan/boot
1288 * and synchronize_rcu() can take significant time and slow down boot.
1289 */
1290 if (wait)
1291 synchronize_rcu();
Vivek Goyalfe071432010-10-01 14:49:49 +02001292
1293 /*
1294 * Just being safe to make sure after previous flush if some body did
1295 * update limits through cgroup and another work got queued, cancel
1296 * it.
1297 */
Vivek Goyalda527772011-03-02 19:05:33 -05001298 throtl_shutdown_wq(q);
Vivek Goyale43473b2010-09-15 17:06:35 -04001299 throtl_td_free(td);
1300}
1301
1302static int __init throtl_init(void)
1303{
Vivek Goyal450adcb2011-03-01 13:40:54 -05001304 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1305 if (!kthrotld_workqueue)
1306 panic("Failed to create kthrotld\n");
1307
Vivek Goyale43473b2010-09-15 17:06:35 -04001308 blkio_policy_register(&blkio_policy_throtl);
1309 return 0;
1310}
1311
1312module_init(throtl_init);