blob: 004be80fd89402efbdc3f4a89a894570b7f27e59 [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
23struct throtl_rb_root {
24 struct rb_root rb;
25 struct rb_node *left;
26 unsigned int count;
27 unsigned long min_disptime;
28};
29
30#define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \
31 .count = 0, .min_disptime = 0}
32
33#define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
34
35struct throtl_grp {
36 /* List of throtl groups on the request queue*/
37 struct hlist_node tg_node;
38
39 /* active throtl group service_tree member */
40 struct rb_node rb_node;
41
42 /*
43 * Dispatch time in jiffies. This is the estimated time when group
44 * will unthrottle and is ready to dispatch more bio. It is used as
45 * key to sort active groups in service tree.
46 */
47 unsigned long disptime;
48
49 struct blkio_group blkg;
50 atomic_t ref;
51 unsigned int flags;
52
53 /* Two lists for READ and WRITE */
54 struct bio_list bio_lists[2];
55
56 /* Number of queued bios on READ and WRITE lists */
57 unsigned int nr_queued[2];
58
59 /* bytes per second rate limits */
60 uint64_t bps[2];
61
Vivek Goyal8e89d132010-09-15 17:06:37 -040062 /* IOPS limits */
63 unsigned int iops[2];
64
Vivek Goyale43473b2010-09-15 17:06:35 -040065 /* Number of bytes disptached in current slice */
66 uint64_t bytes_disp[2];
Vivek Goyal8e89d132010-09-15 17:06:37 -040067 /* Number of bio's dispatched in current slice */
68 unsigned int io_disp[2];
Vivek Goyale43473b2010-09-15 17:06:35 -040069
70 /* When did we start a new slice */
71 unsigned long slice_start[2];
72 unsigned long slice_end[2];
Vivek Goyalfe071432010-10-01 14:49:49 +020073
74 /* Some throttle limits got updated for the group */
75 bool limits_changed;
Vivek Goyale43473b2010-09-15 17:06:35 -040076};
77
78struct throtl_data
79{
80 /* List of throtl groups */
81 struct hlist_head tg_list;
82
83 /* service tree for active throtl groups */
84 struct throtl_rb_root tg_service_tree;
85
86 struct throtl_grp root_tg;
87 struct request_queue *queue;
88
89 /* Total Number of queued bios on READ and WRITE lists */
90 unsigned int nr_queued[2];
91
92 /*
Vivek Goyal02977e42010-10-01 14:49:48 +020093 * number of total undestroyed groups
Vivek Goyale43473b2010-09-15 17:06:35 -040094 */
95 unsigned int nr_undestroyed_grps;
96
97 /* Work for dispatching throttled bios */
98 struct delayed_work throtl_work;
Vivek Goyalfe071432010-10-01 14:49:49 +020099
100 atomic_t limits_changed;
Vivek Goyale43473b2010-09-15 17:06:35 -0400101};
102
103enum tg_state_flags {
104 THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */
105};
106
107#define THROTL_TG_FNS(name) \
108static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \
109{ \
110 (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \
111} \
112static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \
113{ \
114 (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \
115} \
116static inline int throtl_tg_##name(const struct throtl_grp *tg) \
117{ \
118 return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \
119}
120
121THROTL_TG_FNS(on_rr);
122
123#define throtl_log_tg(td, tg, fmt, args...) \
124 blk_add_trace_msg((td)->queue, "throtl %s " fmt, \
125 blkg_path(&(tg)->blkg), ##args); \
126
127#define throtl_log(td, fmt, args...) \
128 blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
129
130static inline struct throtl_grp *tg_of_blkg(struct blkio_group *blkg)
131{
132 if (blkg)
133 return container_of(blkg, struct throtl_grp, blkg);
134
135 return NULL;
136}
137
138static inline int total_nr_queued(struct throtl_data *td)
139{
140 return (td->nr_queued[0] + td->nr_queued[1]);
141}
142
143static inline struct throtl_grp *throtl_ref_get_tg(struct throtl_grp *tg)
144{
145 atomic_inc(&tg->ref);
146 return tg;
147}
148
149static void throtl_put_tg(struct throtl_grp *tg)
150{
151 BUG_ON(atomic_read(&tg->ref) <= 0);
152 if (!atomic_dec_and_test(&tg->ref))
153 return;
154 kfree(tg);
155}
156
157static struct throtl_grp * throtl_find_alloc_tg(struct throtl_data *td,
158 struct cgroup *cgroup)
159{
160 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
161 struct throtl_grp *tg = NULL;
162 void *key = td;
163 struct backing_dev_info *bdi = &td->queue->backing_dev_info;
164 unsigned int major, minor;
165
166 /*
167 * TODO: Speed up blkiocg_lookup_group() by maintaining a radix
168 * tree of blkg (instead of traversing through hash list all
169 * the time.
170 */
171 tg = tg_of_blkg(blkiocg_lookup_group(blkcg, key));
172
173 /* Fill in device details for root group */
174 if (tg && !tg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
175 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
176 tg->blkg.dev = MKDEV(major, minor);
177 goto done;
178 }
179
180 if (tg)
181 goto done;
182
183 tg = kzalloc_node(sizeof(*tg), GFP_ATOMIC, td->queue->node);
184 if (!tg)
185 goto done;
186
187 INIT_HLIST_NODE(&tg->tg_node);
188 RB_CLEAR_NODE(&tg->rb_node);
189 bio_list_init(&tg->bio_lists[0]);
190 bio_list_init(&tg->bio_lists[1]);
191
192 /*
193 * Take the initial reference that will be released on destroy
194 * This can be thought of a joint reference by cgroup and
195 * request queue which will be dropped by either request queue
196 * exit or cgroup deletion path depending on who is exiting first.
197 */
198 atomic_set(&tg->ref, 1);
199
200 /* Add group onto cgroup list */
201 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
202 blkiocg_add_blkio_group(blkcg, &tg->blkg, (void *)td,
203 MKDEV(major, minor), BLKIO_POLICY_THROTL);
204
205 tg->bps[READ] = blkcg_get_read_bps(blkcg, tg->blkg.dev);
206 tg->bps[WRITE] = blkcg_get_write_bps(blkcg, tg->blkg.dev);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400207 tg->iops[READ] = blkcg_get_read_iops(blkcg, tg->blkg.dev);
208 tg->iops[WRITE] = blkcg_get_write_iops(blkcg, tg->blkg.dev);
Vivek Goyale43473b2010-09-15 17:06:35 -0400209
210 hlist_add_head(&tg->tg_node, &td->tg_list);
211 td->nr_undestroyed_grps++;
212done:
213 return tg;
214}
215
216static struct throtl_grp * throtl_get_tg(struct throtl_data *td)
217{
218 struct cgroup *cgroup;
219 struct throtl_grp *tg = NULL;
220
221 rcu_read_lock();
222 cgroup = task_cgroup(current, blkio_subsys_id);
223 tg = throtl_find_alloc_tg(td, cgroup);
224 if (!tg)
225 tg = &td->root_tg;
226 rcu_read_unlock();
227 return tg;
228}
229
230static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root)
231{
232 /* Service tree is empty */
233 if (!root->count)
234 return NULL;
235
236 if (!root->left)
237 root->left = rb_first(&root->rb);
238
239 if (root->left)
240 return rb_entry_tg(root->left);
241
242 return NULL;
243}
244
245static void rb_erase_init(struct rb_node *n, struct rb_root *root)
246{
247 rb_erase(n, root);
248 RB_CLEAR_NODE(n);
249}
250
251static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root)
252{
253 if (root->left == n)
254 root->left = NULL;
255 rb_erase_init(n, &root->rb);
256 --root->count;
257}
258
259static void update_min_dispatch_time(struct throtl_rb_root *st)
260{
261 struct throtl_grp *tg;
262
263 tg = throtl_rb_first(st);
264 if (!tg)
265 return;
266
267 st->min_disptime = tg->disptime;
268}
269
270static void
271tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg)
272{
273 struct rb_node **node = &st->rb.rb_node;
274 struct rb_node *parent = NULL;
275 struct throtl_grp *__tg;
276 unsigned long key = tg->disptime;
277 int left = 1;
278
279 while (*node != NULL) {
280 parent = *node;
281 __tg = rb_entry_tg(parent);
282
283 if (time_before(key, __tg->disptime))
284 node = &parent->rb_left;
285 else {
286 node = &parent->rb_right;
287 left = 0;
288 }
289 }
290
291 if (left)
292 st->left = &tg->rb_node;
293
294 rb_link_node(&tg->rb_node, parent, node);
295 rb_insert_color(&tg->rb_node, &st->rb);
296}
297
298static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
299{
300 struct throtl_rb_root *st = &td->tg_service_tree;
301
302 tg_service_tree_add(st, tg);
303 throtl_mark_tg_on_rr(tg);
304 st->count++;
305}
306
307static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
308{
309 if (!throtl_tg_on_rr(tg))
310 __throtl_enqueue_tg(td, tg);
311}
312
313static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
314{
315 throtl_rb_erase(&tg->rb_node, &td->tg_service_tree);
316 throtl_clear_tg_on_rr(tg);
317}
318
319static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
320{
321 if (throtl_tg_on_rr(tg))
322 __throtl_dequeue_tg(td, tg);
323}
324
325static void throtl_schedule_next_dispatch(struct throtl_data *td)
326{
327 struct throtl_rb_root *st = &td->tg_service_tree;
328
329 /*
330 * If there are more bios pending, schedule more work.
331 */
332 if (!total_nr_queued(td))
333 return;
334
335 BUG_ON(!st->count);
336
337 update_min_dispatch_time(st);
338
339 if (time_before_eq(st->min_disptime, jiffies))
340 throtl_schedule_delayed_work(td->queue, 0);
341 else
342 throtl_schedule_delayed_work(td->queue,
343 (st->min_disptime - jiffies));
344}
345
346static inline void
347throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
348{
349 tg->bytes_disp[rw] = 0;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400350 tg->io_disp[rw] = 0;
Vivek Goyale43473b2010-09-15 17:06:35 -0400351 tg->slice_start[rw] = jiffies;
352 tg->slice_end[rw] = jiffies + throtl_slice;
353 throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
354 rw == READ ? 'R' : 'W', tg->slice_start[rw],
355 tg->slice_end[rw], jiffies);
356}
357
358static inline void throtl_extend_slice(struct throtl_data *td,
359 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
360{
361 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
362 throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
363 rw == READ ? 'R' : 'W', tg->slice_start[rw],
364 tg->slice_end[rw], jiffies);
365}
366
367/* Determine if previously allocated or extended slice is complete or not */
368static bool
369throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
370{
371 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
372 return 0;
373
374 return 1;
375}
376
377/* Trim the used slices and adjust slice start accordingly */
378static inline void
379throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
380{
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200381 unsigned long nr_slices, time_elapsed, io_trim;
382 u64 bytes_trim, tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400383
384 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
385
386 /*
387 * If bps are unlimited (-1), then time slice don't get
388 * renewed. Don't try to trim the slice if slice is used. A new
389 * slice will start when appropriate.
390 */
391 if (throtl_slice_used(td, tg, rw))
392 return;
393
394 time_elapsed = jiffies - tg->slice_start[rw];
395
396 nr_slices = time_elapsed / throtl_slice;
397
398 if (!nr_slices)
399 return;
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200400 tmp = tg->bps[rw] * throtl_slice * nr_slices;
401 do_div(tmp, HZ);
402 bytes_trim = tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400403
Vivek Goyal8e89d132010-09-15 17:06:37 -0400404 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
Vivek Goyale43473b2010-09-15 17:06:35 -0400405
Vivek Goyal8e89d132010-09-15 17:06:37 -0400406 if (!bytes_trim && !io_trim)
Vivek Goyale43473b2010-09-15 17:06:35 -0400407 return;
408
409 if (tg->bytes_disp[rw] >= bytes_trim)
410 tg->bytes_disp[rw] -= bytes_trim;
411 else
412 tg->bytes_disp[rw] = 0;
413
Vivek Goyal8e89d132010-09-15 17:06:37 -0400414 if (tg->io_disp[rw] >= io_trim)
415 tg->io_disp[rw] -= io_trim;
416 else
417 tg->io_disp[rw] = 0;
418
Vivek Goyale43473b2010-09-15 17:06:35 -0400419 tg->slice_start[rw] += nr_slices * throtl_slice;
420
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200421 throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
Vivek Goyale43473b2010-09-15 17:06:35 -0400422 " start=%lu end=%lu jiffies=%lu",
Vivek Goyal8e89d132010-09-15 17:06:37 -0400423 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
Vivek Goyale43473b2010-09-15 17:06:35 -0400424 tg->slice_start[rw], tg->slice_end[rw], jiffies);
425}
426
Vivek Goyal8e89d132010-09-15 17:06:37 -0400427static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
428 struct bio *bio, unsigned long *wait)
Vivek Goyale43473b2010-09-15 17:06:35 -0400429{
430 bool rw = bio_data_dir(bio);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400431 unsigned int io_allowed;
Vivek Goyale43473b2010-09-15 17:06:35 -0400432 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
Vivek Goyalc49c06e2010-10-01 21:16:42 +0200433 u64 tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400434
Vivek Goyal8e89d132010-09-15 17:06:37 -0400435 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
Vivek Goyale43473b2010-09-15 17:06:35 -0400436
Vivek Goyal8e89d132010-09-15 17:06:37 -0400437 /* Slice has just started. Consider one slice interval */
438 if (!jiffy_elapsed)
439 jiffy_elapsed_rnd = throtl_slice;
440
441 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
442
Vivek Goyalc49c06e2010-10-01 21:16:42 +0200443 /*
444 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
445 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
446 * will allow dispatch after 1 second and after that slice should
447 * have been trimmed.
448 */
449
450 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
451 do_div(tmp, HZ);
452
453 if (tmp > UINT_MAX)
454 io_allowed = UINT_MAX;
455 else
456 io_allowed = tmp;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400457
458 if (tg->io_disp[rw] + 1 <= io_allowed) {
Vivek Goyale43473b2010-09-15 17:06:35 -0400459 if (wait)
460 *wait = 0;
461 return 1;
462 }
463
Vivek Goyal8e89d132010-09-15 17:06:37 -0400464 /* Calc approx time to dispatch */
465 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
466
467 if (jiffy_wait > jiffy_elapsed)
468 jiffy_wait = jiffy_wait - jiffy_elapsed;
469 else
470 jiffy_wait = 1;
471
472 if (wait)
473 *wait = jiffy_wait;
474 return 0;
475}
476
477static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
478 struct bio *bio, unsigned long *wait)
479{
480 bool rw = bio_data_dir(bio);
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200481 u64 bytes_allowed, extra_bytes, tmp;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400482 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
Vivek Goyale43473b2010-09-15 17:06:35 -0400483
484 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
485
486 /* Slice has just started. Consider one slice interval */
487 if (!jiffy_elapsed)
488 jiffy_elapsed_rnd = throtl_slice;
489
490 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
491
Vivek Goyal5e901a22010-10-01 21:16:38 +0200492 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
493 do_div(tmp, HZ);
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200494 bytes_allowed = tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400495
496 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
497 if (wait)
498 *wait = 0;
499 return 1;
500 }
501
502 /* Calc approx time to dispatch */
503 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
504 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
505
506 if (!jiffy_wait)
507 jiffy_wait = 1;
508
509 /*
510 * This wait time is without taking into consideration the rounding
511 * up we did. Add that time also.
512 */
513 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
Vivek Goyale43473b2010-09-15 17:06:35 -0400514 if (wait)
515 *wait = jiffy_wait;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400516 return 0;
517}
Vivek Goyale43473b2010-09-15 17:06:35 -0400518
Vivek Goyal8e89d132010-09-15 17:06:37 -0400519/*
520 * Returns whether one can dispatch a bio or not. Also returns approx number
521 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
522 */
523static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
524 struct bio *bio, unsigned long *wait)
525{
526 bool rw = bio_data_dir(bio);
527 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
528
529 /*
530 * Currently whole state machine of group depends on first bio
531 * queued in the group bio list. So one should not be calling
532 * this function with a different bio if there are other bios
533 * queued.
534 */
535 BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
536
537 /* If tg->bps = -1, then BW is unlimited */
538 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
539 if (wait)
540 *wait = 0;
541 return 1;
542 }
543
544 /*
545 * If previous slice expired, start a new one otherwise renew/extend
546 * existing slice to make sure it is at least throtl_slice interval
547 * long since now.
548 */
549 if (throtl_slice_used(td, tg, rw))
550 throtl_start_new_slice(td, tg, rw);
551 else {
552 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
553 throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
554 }
555
556 if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
557 && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
558 if (wait)
559 *wait = 0;
560 return 1;
561 }
562
563 max_wait = max(bps_wait, iops_wait);
564
565 if (wait)
566 *wait = max_wait;
567
568 if (time_before(tg->slice_end[rw], jiffies + max_wait))
569 throtl_extend_slice(td, tg, rw, jiffies + max_wait);
Vivek Goyale43473b2010-09-15 17:06:35 -0400570
571 return 0;
572}
573
574static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
575{
576 bool rw = bio_data_dir(bio);
577 bool sync = bio->bi_rw & REQ_SYNC;
578
579 /* Charge the bio to the group */
580 tg->bytes_disp[rw] += bio->bi_size;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400581 tg->io_disp[rw]++;
Vivek Goyale43473b2010-09-15 17:06:35 -0400582
583 /*
584 * TODO: This will take blkg->stats_lock. Figure out a way
585 * to avoid this cost.
586 */
587 blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync);
Vivek Goyale43473b2010-09-15 17:06:35 -0400588}
589
590static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
591 struct bio *bio)
592{
593 bool rw = bio_data_dir(bio);
594
595 bio_list_add(&tg->bio_lists[rw], bio);
596 /* Take a bio reference on tg */
597 throtl_ref_get_tg(tg);
598 tg->nr_queued[rw]++;
599 td->nr_queued[rw]++;
600 throtl_enqueue_tg(td, tg);
601}
602
603static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
604{
605 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
606 struct bio *bio;
607
608 if ((bio = bio_list_peek(&tg->bio_lists[READ])))
609 tg_may_dispatch(td, tg, bio, &read_wait);
610
611 if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
612 tg_may_dispatch(td, tg, bio, &write_wait);
613
614 min_wait = min(read_wait, write_wait);
615 disptime = jiffies + min_wait;
616
Vivek Goyale43473b2010-09-15 17:06:35 -0400617 /* Update dispatch time */
618 throtl_dequeue_tg(td, tg);
619 tg->disptime = disptime;
620 throtl_enqueue_tg(td, tg);
621}
622
623static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
624 bool rw, struct bio_list *bl)
625{
626 struct bio *bio;
627
628 bio = bio_list_pop(&tg->bio_lists[rw]);
629 tg->nr_queued[rw]--;
630 /* Drop bio reference on tg */
631 throtl_put_tg(tg);
632
633 BUG_ON(td->nr_queued[rw] <= 0);
634 td->nr_queued[rw]--;
635
636 throtl_charge_bio(tg, bio);
637 bio_list_add(bl, bio);
638 bio->bi_rw |= REQ_THROTTLED;
639
640 throtl_trim_slice(td, tg, rw);
641}
642
643static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
644 struct bio_list *bl)
645{
646 unsigned int nr_reads = 0, nr_writes = 0;
647 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
Vivek Goyalc2f68052010-11-15 19:32:42 +0100648 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
Vivek Goyale43473b2010-09-15 17:06:35 -0400649 struct bio *bio;
650
651 /* Try to dispatch 75% READS and 25% WRITES */
652
653 while ((bio = bio_list_peek(&tg->bio_lists[READ]))
654 && tg_may_dispatch(td, tg, bio, NULL)) {
655
656 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
657 nr_reads++;
658
659 if (nr_reads >= max_nr_reads)
660 break;
661 }
662
663 while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
664 && tg_may_dispatch(td, tg, bio, NULL)) {
665
666 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
667 nr_writes++;
668
669 if (nr_writes >= max_nr_writes)
670 break;
671 }
672
673 return nr_reads + nr_writes;
674}
675
676static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
677{
678 unsigned int nr_disp = 0;
679 struct throtl_grp *tg;
680 struct throtl_rb_root *st = &td->tg_service_tree;
681
682 while (1) {
683 tg = throtl_rb_first(st);
684
685 if (!tg)
686 break;
687
688 if (time_before(jiffies, tg->disptime))
689 break;
690
691 throtl_dequeue_tg(td, tg);
692
693 nr_disp += throtl_dispatch_tg(td, tg, bl);
694
695 if (tg->nr_queued[0] || tg->nr_queued[1]) {
696 tg_update_disptime(td, tg);
697 throtl_enqueue_tg(td, tg);
698 }
699
700 if (nr_disp >= throtl_quantum)
701 break;
702 }
703
704 return nr_disp;
705}
706
Vivek Goyalfe071432010-10-01 14:49:49 +0200707static void throtl_process_limit_change(struct throtl_data *td)
708{
709 struct throtl_grp *tg;
710 struct hlist_node *pos, *n;
711
712 /*
713 * Make sure atomic_inc() effects from
714 * throtl_update_blkio_group_read_bps(), group of functions are
715 * visible.
716 * Is this required or smp_mb__after_atomic_inc() was suffcient
717 * after the atomic_inc().
718 */
719 smp_rmb();
720 if (!atomic_read(&td->limits_changed))
721 return;
722
723 throtl_log(td, "limit changed =%d", atomic_read(&td->limits_changed));
724
725 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
726 /*
727 * Do I need an smp_rmb() here to make sure tg->limits_changed
728 * update is visible. I am relying on smp_rmb() at the
729 * beginning of function and not putting a new one here.
730 */
731
732 if (throtl_tg_on_rr(tg) && tg->limits_changed) {
733 throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
734 " riops=%u wiops=%u", tg->bps[READ],
735 tg->bps[WRITE], tg->iops[READ],
736 tg->iops[WRITE]);
737 tg_update_disptime(td, tg);
738 tg->limits_changed = false;
739 }
740 }
741
742 smp_mb__before_atomic_dec();
743 atomic_dec(&td->limits_changed);
744 smp_mb__after_atomic_dec();
745}
746
Vivek Goyale43473b2010-09-15 17:06:35 -0400747/* Dispatch throttled bios. Should be called without queue lock held. */
748static int throtl_dispatch(struct request_queue *q)
749{
750 struct throtl_data *td = q->td;
751 unsigned int nr_disp = 0;
752 struct bio_list bio_list_on_stack;
753 struct bio *bio;
754
755 spin_lock_irq(q->queue_lock);
756
Vivek Goyalfe071432010-10-01 14:49:49 +0200757 throtl_process_limit_change(td);
758
Vivek Goyale43473b2010-09-15 17:06:35 -0400759 if (!total_nr_queued(td))
760 goto out;
761
762 bio_list_init(&bio_list_on_stack);
763
764 throtl_log(td, "dispatch nr_queued=%lu read=%u write=%u",
765 total_nr_queued(td), td->nr_queued[READ],
766 td->nr_queued[WRITE]);
767
768 nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
769
770 if (nr_disp)
771 throtl_log(td, "bios disp=%u", nr_disp);
772
773 throtl_schedule_next_dispatch(td);
774out:
775 spin_unlock_irq(q->queue_lock);
776
777 /*
778 * If we dispatched some requests, unplug the queue to make sure
779 * immediate dispatch
780 */
781 if (nr_disp) {
782 while((bio = bio_list_pop(&bio_list_on_stack)))
783 generic_make_request(bio);
784 blk_unplug(q);
785 }
786 return nr_disp;
787}
788
789void blk_throtl_work(struct work_struct *work)
790{
791 struct throtl_data *td = container_of(work, struct throtl_data,
792 throtl_work.work);
793 struct request_queue *q = td->queue;
794
795 throtl_dispatch(q);
796}
797
798/* Call with queue lock held */
799void throtl_schedule_delayed_work(struct request_queue *q, unsigned long delay)
800{
801
802 struct throtl_data *td = q->td;
803 struct delayed_work *dwork = &td->throtl_work;
804
805 if (total_nr_queued(td) > 0) {
806 /*
807 * We might have a work scheduled to be executed in future.
808 * Cancel that and schedule a new one.
809 */
810 __cancel_delayed_work(dwork);
811 kblockd_schedule_delayed_work(q, dwork, delay);
812 throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
813 delay, jiffies);
814 }
815}
816EXPORT_SYMBOL(throtl_schedule_delayed_work);
817
818static void
819throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg)
820{
821 /* Something wrong if we are trying to remove same group twice */
822 BUG_ON(hlist_unhashed(&tg->tg_node));
823
824 hlist_del_init(&tg->tg_node);
825
826 /*
827 * Put the reference taken at the time of creation so that when all
828 * queues are gone, group can be destroyed.
829 */
830 throtl_put_tg(tg);
831 td->nr_undestroyed_grps--;
832}
833
834static void throtl_release_tgs(struct throtl_data *td)
835{
836 struct hlist_node *pos, *n;
837 struct throtl_grp *tg;
838
839 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
840 /*
841 * If cgroup removal path got to blk_group first and removed
842 * it from cgroup list, then it will take care of destroying
843 * cfqg also.
844 */
845 if (!blkiocg_del_blkio_group(&tg->blkg))
846 throtl_destroy_tg(td, tg);
847 }
848}
849
850static void throtl_td_free(struct throtl_data *td)
851{
852 kfree(td);
853}
854
855/*
856 * Blk cgroup controller notification saying that blkio_group object is being
857 * delinked as associated cgroup object is going away. That also means that
858 * no new IO will come in this group. So get rid of this group as soon as
859 * any pending IO in the group is finished.
860 *
861 * This function is called under rcu_read_lock(). key is the rcu protected
862 * pointer. That means "key" is a valid throtl_data pointer as long as we are
863 * rcu read lock.
864 *
865 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
866 * it should not be NULL as even if queue was going away, cgroup deltion
867 * path got to it first.
868 */
869void throtl_unlink_blkio_group(void *key, struct blkio_group *blkg)
870{
871 unsigned long flags;
872 struct throtl_data *td = key;
873
874 spin_lock_irqsave(td->queue->queue_lock, flags);
875 throtl_destroy_tg(td, tg_of_blkg(blkg));
876 spin_unlock_irqrestore(td->queue->queue_lock, flags);
877}
878
Vivek Goyalfe071432010-10-01 14:49:49 +0200879/*
880 * For all update functions, key should be a valid pointer because these
881 * update functions are called under blkcg_lock, that means, blkg is
882 * valid and in turn key is valid. queue exit path can not race becuase
883 * of blkcg_lock
884 *
885 * Can not take queue lock in update functions as queue lock under blkcg_lock
886 * is not allowed. Under other paths we take blkcg_lock under queue_lock.
887 */
888static void throtl_update_blkio_group_read_bps(void *key,
889 struct blkio_group *blkg, u64 read_bps)
Vivek Goyale43473b2010-09-15 17:06:35 -0400890{
Vivek Goyalfe071432010-10-01 14:49:49 +0200891 struct throtl_data *td = key;
892
Vivek Goyale43473b2010-09-15 17:06:35 -0400893 tg_of_blkg(blkg)->bps[READ] = read_bps;
Vivek Goyalfe071432010-10-01 14:49:49 +0200894 /* Make sure read_bps is updated before setting limits_changed */
895 smp_wmb();
896 tg_of_blkg(blkg)->limits_changed = true;
897
898 /* Make sure tg->limits_changed is updated before td->limits_changed */
899 smp_mb__before_atomic_inc();
900 atomic_inc(&td->limits_changed);
901 smp_mb__after_atomic_inc();
902
903 /* Schedule a work now to process the limit change */
904 throtl_schedule_delayed_work(td->queue, 0);
Vivek Goyale43473b2010-09-15 17:06:35 -0400905}
906
Vivek Goyalfe071432010-10-01 14:49:49 +0200907static void throtl_update_blkio_group_write_bps(void *key,
908 struct blkio_group *blkg, u64 write_bps)
Vivek Goyale43473b2010-09-15 17:06:35 -0400909{
Vivek Goyalfe071432010-10-01 14:49:49 +0200910 struct throtl_data *td = key;
911
Vivek Goyale43473b2010-09-15 17:06:35 -0400912 tg_of_blkg(blkg)->bps[WRITE] = write_bps;
Vivek Goyalfe071432010-10-01 14:49:49 +0200913 smp_wmb();
914 tg_of_blkg(blkg)->limits_changed = true;
915 smp_mb__before_atomic_inc();
916 atomic_inc(&td->limits_changed);
917 smp_mb__after_atomic_inc();
918 throtl_schedule_delayed_work(td->queue, 0);
Vivek Goyale43473b2010-09-15 17:06:35 -0400919}
920
Vivek Goyalfe071432010-10-01 14:49:49 +0200921static void throtl_update_blkio_group_read_iops(void *key,
922 struct blkio_group *blkg, unsigned int read_iops)
Vivek Goyal8e89d132010-09-15 17:06:37 -0400923{
Vivek Goyalfe071432010-10-01 14:49:49 +0200924 struct throtl_data *td = key;
925
Vivek Goyal8e89d132010-09-15 17:06:37 -0400926 tg_of_blkg(blkg)->iops[READ] = read_iops;
Vivek Goyalfe071432010-10-01 14:49:49 +0200927 smp_wmb();
928 tg_of_blkg(blkg)->limits_changed = true;
929 smp_mb__before_atomic_inc();
930 atomic_inc(&td->limits_changed);
931 smp_mb__after_atomic_inc();
932 throtl_schedule_delayed_work(td->queue, 0);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400933}
934
Vivek Goyalfe071432010-10-01 14:49:49 +0200935static void throtl_update_blkio_group_write_iops(void *key,
936 struct blkio_group *blkg, unsigned int write_iops)
Vivek Goyal8e89d132010-09-15 17:06:37 -0400937{
Vivek Goyalfe071432010-10-01 14:49:49 +0200938 struct throtl_data *td = key;
939
Vivek Goyal8e89d132010-09-15 17:06:37 -0400940 tg_of_blkg(blkg)->iops[WRITE] = write_iops;
Vivek Goyalfe071432010-10-01 14:49:49 +0200941 smp_wmb();
942 tg_of_blkg(blkg)->limits_changed = true;
943 smp_mb__before_atomic_inc();
944 atomic_inc(&td->limits_changed);
945 smp_mb__after_atomic_inc();
946 throtl_schedule_delayed_work(td->queue, 0);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400947}
948
Vivek Goyale43473b2010-09-15 17:06:35 -0400949void throtl_shutdown_timer_wq(struct request_queue *q)
950{
951 struct throtl_data *td = q->td;
952
953 cancel_delayed_work_sync(&td->throtl_work);
954}
955
956static struct blkio_policy_type blkio_policy_throtl = {
957 .ops = {
958 .blkio_unlink_group_fn = throtl_unlink_blkio_group,
959 .blkio_update_group_read_bps_fn =
960 throtl_update_blkio_group_read_bps,
961 .blkio_update_group_write_bps_fn =
962 throtl_update_blkio_group_write_bps,
Vivek Goyal8e89d132010-09-15 17:06:37 -0400963 .blkio_update_group_read_iops_fn =
964 throtl_update_blkio_group_read_iops,
965 .blkio_update_group_write_iops_fn =
966 throtl_update_blkio_group_write_iops,
Vivek Goyale43473b2010-09-15 17:06:35 -0400967 },
Vivek Goyal8e89d132010-09-15 17:06:37 -0400968 .plid = BLKIO_POLICY_THROTL,
Vivek Goyale43473b2010-09-15 17:06:35 -0400969};
970
971int blk_throtl_bio(struct request_queue *q, struct bio **biop)
972{
973 struct throtl_data *td = q->td;
974 struct throtl_grp *tg;
975 struct bio *bio = *biop;
976 bool rw = bio_data_dir(bio), update_disptime = true;
977
978 if (bio->bi_rw & REQ_THROTTLED) {
979 bio->bi_rw &= ~REQ_THROTTLED;
980 return 0;
981 }
982
983 spin_lock_irq(q->queue_lock);
984 tg = throtl_get_tg(td);
985
986 if (tg->nr_queued[rw]) {
987 /*
988 * There is already another bio queued in same dir. No
989 * need to update dispatch time.
Vivek Goyalfe071432010-10-01 14:49:49 +0200990 * Still update the disptime if rate limits on this group
991 * were changed.
Vivek Goyale43473b2010-09-15 17:06:35 -0400992 */
Vivek Goyalfe071432010-10-01 14:49:49 +0200993 if (!tg->limits_changed)
994 update_disptime = false;
995 else
996 tg->limits_changed = false;
997
Vivek Goyale43473b2010-09-15 17:06:35 -0400998 goto queue_bio;
999 }
1000
1001 /* Bio is with-in rate limit of group */
1002 if (tg_may_dispatch(td, tg, bio, NULL)) {
1003 throtl_charge_bio(tg, bio);
1004 goto out;
1005 }
1006
1007queue_bio:
Vivek Goyal8e89d132010-09-15 17:06:37 -04001008 throtl_log_tg(td, tg, "[%c] bio. bdisp=%u sz=%u bps=%llu"
1009 " iodisp=%u iops=%u queued=%d/%d",
1010 rw == READ ? 'R' : 'W',
Vivek Goyale43473b2010-09-15 17:06:35 -04001011 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
Vivek Goyal8e89d132010-09-15 17:06:37 -04001012 tg->io_disp[rw], tg->iops[rw],
Vivek Goyale43473b2010-09-15 17:06:35 -04001013 tg->nr_queued[READ], tg->nr_queued[WRITE]);
1014
1015 throtl_add_bio_tg(q->td, tg, bio);
1016 *biop = NULL;
1017
1018 if (update_disptime) {
1019 tg_update_disptime(td, tg);
1020 throtl_schedule_next_dispatch(td);
1021 }
1022
1023out:
1024 spin_unlock_irq(q->queue_lock);
1025 return 0;
1026}
1027
1028int blk_throtl_init(struct request_queue *q)
1029{
1030 struct throtl_data *td;
1031 struct throtl_grp *tg;
1032
1033 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1034 if (!td)
1035 return -ENOMEM;
1036
1037 INIT_HLIST_HEAD(&td->tg_list);
1038 td->tg_service_tree = THROTL_RB_ROOT;
Vivek Goyalfe071432010-10-01 14:49:49 +02001039 atomic_set(&td->limits_changed, 0);
Vivek Goyale43473b2010-09-15 17:06:35 -04001040
1041 /* Init root group */
1042 tg = &td->root_tg;
1043 INIT_HLIST_NODE(&tg->tg_node);
1044 RB_CLEAR_NODE(&tg->rb_node);
1045 bio_list_init(&tg->bio_lists[0]);
1046 bio_list_init(&tg->bio_lists[1]);
1047
1048 /* Practically unlimited BW */
1049 tg->bps[0] = tg->bps[1] = -1;
Vivek Goyal8e89d132010-09-15 17:06:37 -04001050 tg->iops[0] = tg->iops[1] = -1;
Vivek Goyal02977e42010-10-01 14:49:48 +02001051
1052 /*
1053 * Set root group reference to 2. One reference will be dropped when
1054 * all groups on tg_list are being deleted during queue exit. Other
1055 * reference will remain there as we don't want to delete this group
1056 * as it is statically allocated and gets destroyed when throtl_data
1057 * goes away.
1058 */
1059 atomic_set(&tg->ref, 2);
1060 hlist_add_head(&tg->tg_node, &td->tg_list);
1061 td->nr_undestroyed_grps++;
Vivek Goyale43473b2010-09-15 17:06:35 -04001062
1063 INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);
1064
1065 rcu_read_lock();
1066 blkiocg_add_blkio_group(&blkio_root_cgroup, &tg->blkg, (void *)td,
1067 0, BLKIO_POLICY_THROTL);
1068 rcu_read_unlock();
1069
1070 /* Attach throtl data to request queue */
1071 td->queue = q;
1072 q->td = td;
1073 return 0;
1074}
1075
1076void blk_throtl_exit(struct request_queue *q)
1077{
1078 struct throtl_data *td = q->td;
1079 bool wait = false;
1080
1081 BUG_ON(!td);
1082
1083 throtl_shutdown_timer_wq(q);
1084
1085 spin_lock_irq(q->queue_lock);
1086 throtl_release_tgs(td);
Vivek Goyale43473b2010-09-15 17:06:35 -04001087
1088 /* If there are other groups */
Vivek Goyal02977e42010-10-01 14:49:48 +02001089 if (td->nr_undestroyed_grps > 0)
Vivek Goyale43473b2010-09-15 17:06:35 -04001090 wait = true;
1091
1092 spin_unlock_irq(q->queue_lock);
1093
1094 /*
1095 * Wait for tg->blkg->key accessors to exit their grace periods.
1096 * Do this wait only if there are other undestroyed groups out
1097 * there (other than root group). This can happen if cgroup deletion
1098 * path claimed the responsibility of cleaning up a group before
1099 * queue cleanup code get to the group.
1100 *
1101 * Do not call synchronize_rcu() unconditionally as there are drivers
1102 * which create/delete request queue hundreds of times during scan/boot
1103 * and synchronize_rcu() can take significant time and slow down boot.
1104 */
1105 if (wait)
1106 synchronize_rcu();
Vivek Goyalfe071432010-10-01 14:49:49 +02001107
1108 /*
1109 * Just being safe to make sure after previous flush if some body did
1110 * update limits through cgroup and another work got queued, cancel
1111 * it.
1112 */
1113 throtl_shutdown_timer_wq(q);
Vivek Goyale43473b2010-09-15 17:06:35 -04001114 throtl_td_free(td);
1115}
1116
1117static int __init throtl_init(void)
1118{
1119 blkio_policy_register(&blkio_policy_throtl);
1120 return 0;
1121}
1122
1123module_init(throtl_init);