blob: a46700255719d6693abe2900d173f8bf45cd7ae5 [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;
433
Vivek Goyal8e89d132010-09-15 17:06:37 -0400434 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
Vivek Goyale43473b2010-09-15 17:06:35 -0400435
Vivek Goyal8e89d132010-09-15 17:06:37 -0400436 /* Slice has just started. Consider one slice interval */
437 if (!jiffy_elapsed)
438 jiffy_elapsed_rnd = throtl_slice;
439
440 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
441
442 io_allowed = (tg->iops[rw] * jiffies_to_msecs(jiffy_elapsed_rnd))
443 / MSEC_PER_SEC;
444
445 if (tg->io_disp[rw] + 1 <= io_allowed) {
Vivek Goyale43473b2010-09-15 17:06:35 -0400446 if (wait)
447 *wait = 0;
448 return 1;
449 }
450
Vivek Goyal8e89d132010-09-15 17:06:37 -0400451 /* Calc approx time to dispatch */
452 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
453
454 if (jiffy_wait > jiffy_elapsed)
455 jiffy_wait = jiffy_wait - jiffy_elapsed;
456 else
457 jiffy_wait = 1;
458
459 if (wait)
460 *wait = jiffy_wait;
461 return 0;
462}
463
464static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
465 struct bio *bio, unsigned long *wait)
466{
467 bool rw = bio_data_dir(bio);
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200468 u64 bytes_allowed, extra_bytes, tmp;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400469 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
Vivek Goyale43473b2010-09-15 17:06:35 -0400470
471 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
472
473 /* Slice has just started. Consider one slice interval */
474 if (!jiffy_elapsed)
475 jiffy_elapsed_rnd = throtl_slice;
476
477 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
478
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200479 tmp = tg->bps[rw] * jiffies_to_msecs(jiffy_elapsed_rnd);
480 do_div(tmp, MSEC_PER_SEC);
481 bytes_allowed = tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400482
483 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
484 if (wait)
485 *wait = 0;
486 return 1;
487 }
488
489 /* Calc approx time to dispatch */
490 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
491 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
492
493 if (!jiffy_wait)
494 jiffy_wait = 1;
495
496 /*
497 * This wait time is without taking into consideration the rounding
498 * up we did. Add that time also.
499 */
500 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
Vivek Goyale43473b2010-09-15 17:06:35 -0400501 if (wait)
502 *wait = jiffy_wait;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400503 return 0;
504}
Vivek Goyale43473b2010-09-15 17:06:35 -0400505
Vivek Goyal8e89d132010-09-15 17:06:37 -0400506/*
507 * Returns whether one can dispatch a bio or not. Also returns approx number
508 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
509 */
510static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
511 struct bio *bio, unsigned long *wait)
512{
513 bool rw = bio_data_dir(bio);
514 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
515
516 /*
517 * Currently whole state machine of group depends on first bio
518 * queued in the group bio list. So one should not be calling
519 * this function with a different bio if there are other bios
520 * queued.
521 */
522 BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
523
524 /* If tg->bps = -1, then BW is unlimited */
525 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
526 if (wait)
527 *wait = 0;
528 return 1;
529 }
530
531 /*
532 * If previous slice expired, start a new one otherwise renew/extend
533 * existing slice to make sure it is at least throtl_slice interval
534 * long since now.
535 */
536 if (throtl_slice_used(td, tg, rw))
537 throtl_start_new_slice(td, tg, rw);
538 else {
539 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
540 throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
541 }
542
543 if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
544 && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
545 if (wait)
546 *wait = 0;
547 return 1;
548 }
549
550 max_wait = max(bps_wait, iops_wait);
551
552 if (wait)
553 *wait = max_wait;
554
555 if (time_before(tg->slice_end[rw], jiffies + max_wait))
556 throtl_extend_slice(td, tg, rw, jiffies + max_wait);
Vivek Goyale43473b2010-09-15 17:06:35 -0400557
558 return 0;
559}
560
561static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
562{
563 bool rw = bio_data_dir(bio);
564 bool sync = bio->bi_rw & REQ_SYNC;
565
566 /* Charge the bio to the group */
567 tg->bytes_disp[rw] += bio->bi_size;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400568 tg->io_disp[rw]++;
Vivek Goyale43473b2010-09-15 17:06:35 -0400569
570 /*
571 * TODO: This will take blkg->stats_lock. Figure out a way
572 * to avoid this cost.
573 */
574 blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync);
Vivek Goyale43473b2010-09-15 17:06:35 -0400575}
576
577static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
578 struct bio *bio)
579{
580 bool rw = bio_data_dir(bio);
581
582 bio_list_add(&tg->bio_lists[rw], bio);
583 /* Take a bio reference on tg */
584 throtl_ref_get_tg(tg);
585 tg->nr_queued[rw]++;
586 td->nr_queued[rw]++;
587 throtl_enqueue_tg(td, tg);
588}
589
590static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
591{
592 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
593 struct bio *bio;
594
595 if ((bio = bio_list_peek(&tg->bio_lists[READ])))
596 tg_may_dispatch(td, tg, bio, &read_wait);
597
598 if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
599 tg_may_dispatch(td, tg, bio, &write_wait);
600
601 min_wait = min(read_wait, write_wait);
602 disptime = jiffies + min_wait;
603
Vivek Goyale43473b2010-09-15 17:06:35 -0400604 /* Update dispatch time */
605 throtl_dequeue_tg(td, tg);
606 tg->disptime = disptime;
607 throtl_enqueue_tg(td, tg);
608}
609
610static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
611 bool rw, struct bio_list *bl)
612{
613 struct bio *bio;
614
615 bio = bio_list_pop(&tg->bio_lists[rw]);
616 tg->nr_queued[rw]--;
617 /* Drop bio reference on tg */
618 throtl_put_tg(tg);
619
620 BUG_ON(td->nr_queued[rw] <= 0);
621 td->nr_queued[rw]--;
622
623 throtl_charge_bio(tg, bio);
624 bio_list_add(bl, bio);
625 bio->bi_rw |= REQ_THROTTLED;
626
627 throtl_trim_slice(td, tg, rw);
628}
629
630static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
631 struct bio_list *bl)
632{
633 unsigned int nr_reads = 0, nr_writes = 0;
634 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
635 unsigned int max_nr_writes = throtl_grp_quantum - nr_reads;
636 struct bio *bio;
637
638 /* Try to dispatch 75% READS and 25% WRITES */
639
640 while ((bio = bio_list_peek(&tg->bio_lists[READ]))
641 && tg_may_dispatch(td, tg, bio, NULL)) {
642
643 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
644 nr_reads++;
645
646 if (nr_reads >= max_nr_reads)
647 break;
648 }
649
650 while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
651 && tg_may_dispatch(td, tg, bio, NULL)) {
652
653 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
654 nr_writes++;
655
656 if (nr_writes >= max_nr_writes)
657 break;
658 }
659
660 return nr_reads + nr_writes;
661}
662
663static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
664{
665 unsigned int nr_disp = 0;
666 struct throtl_grp *tg;
667 struct throtl_rb_root *st = &td->tg_service_tree;
668
669 while (1) {
670 tg = throtl_rb_first(st);
671
672 if (!tg)
673 break;
674
675 if (time_before(jiffies, tg->disptime))
676 break;
677
678 throtl_dequeue_tg(td, tg);
679
680 nr_disp += throtl_dispatch_tg(td, tg, bl);
681
682 if (tg->nr_queued[0] || tg->nr_queued[1]) {
683 tg_update_disptime(td, tg);
684 throtl_enqueue_tg(td, tg);
685 }
686
687 if (nr_disp >= throtl_quantum)
688 break;
689 }
690
691 return nr_disp;
692}
693
Vivek Goyalfe071432010-10-01 14:49:49 +0200694static void throtl_process_limit_change(struct throtl_data *td)
695{
696 struct throtl_grp *tg;
697 struct hlist_node *pos, *n;
698
699 /*
700 * Make sure atomic_inc() effects from
701 * throtl_update_blkio_group_read_bps(), group of functions are
702 * visible.
703 * Is this required or smp_mb__after_atomic_inc() was suffcient
704 * after the atomic_inc().
705 */
706 smp_rmb();
707 if (!atomic_read(&td->limits_changed))
708 return;
709
710 throtl_log(td, "limit changed =%d", atomic_read(&td->limits_changed));
711
712 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
713 /*
714 * Do I need an smp_rmb() here to make sure tg->limits_changed
715 * update is visible. I am relying on smp_rmb() at the
716 * beginning of function and not putting a new one here.
717 */
718
719 if (throtl_tg_on_rr(tg) && tg->limits_changed) {
720 throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
721 " riops=%u wiops=%u", tg->bps[READ],
722 tg->bps[WRITE], tg->iops[READ],
723 tg->iops[WRITE]);
724 tg_update_disptime(td, tg);
725 tg->limits_changed = false;
726 }
727 }
728
729 smp_mb__before_atomic_dec();
730 atomic_dec(&td->limits_changed);
731 smp_mb__after_atomic_dec();
732}
733
Vivek Goyale43473b2010-09-15 17:06:35 -0400734/* Dispatch throttled bios. Should be called without queue lock held. */
735static int throtl_dispatch(struct request_queue *q)
736{
737 struct throtl_data *td = q->td;
738 unsigned int nr_disp = 0;
739 struct bio_list bio_list_on_stack;
740 struct bio *bio;
741
742 spin_lock_irq(q->queue_lock);
743
Vivek Goyalfe071432010-10-01 14:49:49 +0200744 throtl_process_limit_change(td);
745
Vivek Goyale43473b2010-09-15 17:06:35 -0400746 if (!total_nr_queued(td))
747 goto out;
748
749 bio_list_init(&bio_list_on_stack);
750
751 throtl_log(td, "dispatch nr_queued=%lu read=%u write=%u",
752 total_nr_queued(td), td->nr_queued[READ],
753 td->nr_queued[WRITE]);
754
755 nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
756
757 if (nr_disp)
758 throtl_log(td, "bios disp=%u", nr_disp);
759
760 throtl_schedule_next_dispatch(td);
761out:
762 spin_unlock_irq(q->queue_lock);
763
764 /*
765 * If we dispatched some requests, unplug the queue to make sure
766 * immediate dispatch
767 */
768 if (nr_disp) {
769 while((bio = bio_list_pop(&bio_list_on_stack)))
770 generic_make_request(bio);
771 blk_unplug(q);
772 }
773 return nr_disp;
774}
775
776void blk_throtl_work(struct work_struct *work)
777{
778 struct throtl_data *td = container_of(work, struct throtl_data,
779 throtl_work.work);
780 struct request_queue *q = td->queue;
781
782 throtl_dispatch(q);
783}
784
785/* Call with queue lock held */
786void throtl_schedule_delayed_work(struct request_queue *q, unsigned long delay)
787{
788
789 struct throtl_data *td = q->td;
790 struct delayed_work *dwork = &td->throtl_work;
791
792 if (total_nr_queued(td) > 0) {
793 /*
794 * We might have a work scheduled to be executed in future.
795 * Cancel that and schedule a new one.
796 */
797 __cancel_delayed_work(dwork);
798 kblockd_schedule_delayed_work(q, dwork, delay);
799 throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
800 delay, jiffies);
801 }
802}
803EXPORT_SYMBOL(throtl_schedule_delayed_work);
804
805static void
806throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg)
807{
808 /* Something wrong if we are trying to remove same group twice */
809 BUG_ON(hlist_unhashed(&tg->tg_node));
810
811 hlist_del_init(&tg->tg_node);
812
813 /*
814 * Put the reference taken at the time of creation so that when all
815 * queues are gone, group can be destroyed.
816 */
817 throtl_put_tg(tg);
818 td->nr_undestroyed_grps--;
819}
820
821static void throtl_release_tgs(struct throtl_data *td)
822{
823 struct hlist_node *pos, *n;
824 struct throtl_grp *tg;
825
826 hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
827 /*
828 * If cgroup removal path got to blk_group first and removed
829 * it from cgroup list, then it will take care of destroying
830 * cfqg also.
831 */
832 if (!blkiocg_del_blkio_group(&tg->blkg))
833 throtl_destroy_tg(td, tg);
834 }
835}
836
837static void throtl_td_free(struct throtl_data *td)
838{
839 kfree(td);
840}
841
842/*
843 * Blk cgroup controller notification saying that blkio_group object is being
844 * delinked as associated cgroup object is going away. That also means that
845 * no new IO will come in this group. So get rid of this group as soon as
846 * any pending IO in the group is finished.
847 *
848 * This function is called under rcu_read_lock(). key is the rcu protected
849 * pointer. That means "key" is a valid throtl_data pointer as long as we are
850 * rcu read lock.
851 *
852 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
853 * it should not be NULL as even if queue was going away, cgroup deltion
854 * path got to it first.
855 */
856void throtl_unlink_blkio_group(void *key, struct blkio_group *blkg)
857{
858 unsigned long flags;
859 struct throtl_data *td = key;
860
861 spin_lock_irqsave(td->queue->queue_lock, flags);
862 throtl_destroy_tg(td, tg_of_blkg(blkg));
863 spin_unlock_irqrestore(td->queue->queue_lock, flags);
864}
865
Vivek Goyalfe071432010-10-01 14:49:49 +0200866/*
867 * For all update functions, key should be a valid pointer because these
868 * update functions are called under blkcg_lock, that means, blkg is
869 * valid and in turn key is valid. queue exit path can not race becuase
870 * of blkcg_lock
871 *
872 * Can not take queue lock in update functions as queue lock under blkcg_lock
873 * is not allowed. Under other paths we take blkcg_lock under queue_lock.
874 */
875static void throtl_update_blkio_group_read_bps(void *key,
876 struct blkio_group *blkg, u64 read_bps)
Vivek Goyale43473b2010-09-15 17:06:35 -0400877{
Vivek Goyalfe071432010-10-01 14:49:49 +0200878 struct throtl_data *td = key;
879
Vivek Goyale43473b2010-09-15 17:06:35 -0400880 tg_of_blkg(blkg)->bps[READ] = read_bps;
Vivek Goyalfe071432010-10-01 14:49:49 +0200881 /* Make sure read_bps is updated before setting limits_changed */
882 smp_wmb();
883 tg_of_blkg(blkg)->limits_changed = true;
884
885 /* Make sure tg->limits_changed is updated before td->limits_changed */
886 smp_mb__before_atomic_inc();
887 atomic_inc(&td->limits_changed);
888 smp_mb__after_atomic_inc();
889
890 /* Schedule a work now to process the limit change */
891 throtl_schedule_delayed_work(td->queue, 0);
Vivek Goyale43473b2010-09-15 17:06:35 -0400892}
893
Vivek Goyalfe071432010-10-01 14:49:49 +0200894static void throtl_update_blkio_group_write_bps(void *key,
895 struct blkio_group *blkg, u64 write_bps)
Vivek Goyale43473b2010-09-15 17:06:35 -0400896{
Vivek Goyalfe071432010-10-01 14:49:49 +0200897 struct throtl_data *td = key;
898
Vivek Goyale43473b2010-09-15 17:06:35 -0400899 tg_of_blkg(blkg)->bps[WRITE] = write_bps;
Vivek Goyalfe071432010-10-01 14:49:49 +0200900 smp_wmb();
901 tg_of_blkg(blkg)->limits_changed = true;
902 smp_mb__before_atomic_inc();
903 atomic_inc(&td->limits_changed);
904 smp_mb__after_atomic_inc();
905 throtl_schedule_delayed_work(td->queue, 0);
Vivek Goyale43473b2010-09-15 17:06:35 -0400906}
907
Vivek Goyalfe071432010-10-01 14:49:49 +0200908static void throtl_update_blkio_group_read_iops(void *key,
909 struct blkio_group *blkg, unsigned int read_iops)
Vivek Goyal8e89d132010-09-15 17:06:37 -0400910{
Vivek Goyalfe071432010-10-01 14:49:49 +0200911 struct throtl_data *td = key;
912
Vivek Goyal8e89d132010-09-15 17:06:37 -0400913 tg_of_blkg(blkg)->iops[READ] = read_iops;
Vivek Goyalfe071432010-10-01 14:49:49 +0200914 smp_wmb();
915 tg_of_blkg(blkg)->limits_changed = true;
916 smp_mb__before_atomic_inc();
917 atomic_inc(&td->limits_changed);
918 smp_mb__after_atomic_inc();
919 throtl_schedule_delayed_work(td->queue, 0);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400920}
921
Vivek Goyalfe071432010-10-01 14:49:49 +0200922static void throtl_update_blkio_group_write_iops(void *key,
923 struct blkio_group *blkg, unsigned int write_iops)
Vivek Goyal8e89d132010-09-15 17:06:37 -0400924{
Vivek Goyalfe071432010-10-01 14:49:49 +0200925 struct throtl_data *td = key;
926
Vivek Goyal8e89d132010-09-15 17:06:37 -0400927 tg_of_blkg(blkg)->iops[WRITE] = write_iops;
Vivek Goyalfe071432010-10-01 14:49:49 +0200928 smp_wmb();
929 tg_of_blkg(blkg)->limits_changed = true;
930 smp_mb__before_atomic_inc();
931 atomic_inc(&td->limits_changed);
932 smp_mb__after_atomic_inc();
933 throtl_schedule_delayed_work(td->queue, 0);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400934}
935
Vivek Goyale43473b2010-09-15 17:06:35 -0400936void throtl_shutdown_timer_wq(struct request_queue *q)
937{
938 struct throtl_data *td = q->td;
939
940 cancel_delayed_work_sync(&td->throtl_work);
941}
942
943static struct blkio_policy_type blkio_policy_throtl = {
944 .ops = {
945 .blkio_unlink_group_fn = throtl_unlink_blkio_group,
946 .blkio_update_group_read_bps_fn =
947 throtl_update_blkio_group_read_bps,
948 .blkio_update_group_write_bps_fn =
949 throtl_update_blkio_group_write_bps,
Vivek Goyal8e89d132010-09-15 17:06:37 -0400950 .blkio_update_group_read_iops_fn =
951 throtl_update_blkio_group_read_iops,
952 .blkio_update_group_write_iops_fn =
953 throtl_update_blkio_group_write_iops,
Vivek Goyale43473b2010-09-15 17:06:35 -0400954 },
Vivek Goyal8e89d132010-09-15 17:06:37 -0400955 .plid = BLKIO_POLICY_THROTL,
Vivek Goyale43473b2010-09-15 17:06:35 -0400956};
957
958int blk_throtl_bio(struct request_queue *q, struct bio **biop)
959{
960 struct throtl_data *td = q->td;
961 struct throtl_grp *tg;
962 struct bio *bio = *biop;
963 bool rw = bio_data_dir(bio), update_disptime = true;
964
965 if (bio->bi_rw & REQ_THROTTLED) {
966 bio->bi_rw &= ~REQ_THROTTLED;
967 return 0;
968 }
969
970 spin_lock_irq(q->queue_lock);
971 tg = throtl_get_tg(td);
972
973 if (tg->nr_queued[rw]) {
974 /*
975 * There is already another bio queued in same dir. No
976 * need to update dispatch time.
Vivek Goyalfe071432010-10-01 14:49:49 +0200977 * Still update the disptime if rate limits on this group
978 * were changed.
Vivek Goyale43473b2010-09-15 17:06:35 -0400979 */
Vivek Goyalfe071432010-10-01 14:49:49 +0200980 if (!tg->limits_changed)
981 update_disptime = false;
982 else
983 tg->limits_changed = false;
984
Vivek Goyale43473b2010-09-15 17:06:35 -0400985 goto queue_bio;
986 }
987
988 /* Bio is with-in rate limit of group */
989 if (tg_may_dispatch(td, tg, bio, NULL)) {
990 throtl_charge_bio(tg, bio);
991 goto out;
992 }
993
994queue_bio:
Vivek Goyal8e89d132010-09-15 17:06:37 -0400995 throtl_log_tg(td, tg, "[%c] bio. bdisp=%u sz=%u bps=%llu"
996 " iodisp=%u iops=%u queued=%d/%d",
997 rw == READ ? 'R' : 'W',
Vivek Goyale43473b2010-09-15 17:06:35 -0400998 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
Vivek Goyal8e89d132010-09-15 17:06:37 -0400999 tg->io_disp[rw], tg->iops[rw],
Vivek Goyale43473b2010-09-15 17:06:35 -04001000 tg->nr_queued[READ], tg->nr_queued[WRITE]);
1001
1002 throtl_add_bio_tg(q->td, tg, bio);
1003 *biop = NULL;
1004
1005 if (update_disptime) {
1006 tg_update_disptime(td, tg);
1007 throtl_schedule_next_dispatch(td);
1008 }
1009
1010out:
1011 spin_unlock_irq(q->queue_lock);
1012 return 0;
1013}
1014
1015int blk_throtl_init(struct request_queue *q)
1016{
1017 struct throtl_data *td;
1018 struct throtl_grp *tg;
1019
1020 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1021 if (!td)
1022 return -ENOMEM;
1023
1024 INIT_HLIST_HEAD(&td->tg_list);
1025 td->tg_service_tree = THROTL_RB_ROOT;
Vivek Goyalfe071432010-10-01 14:49:49 +02001026 atomic_set(&td->limits_changed, 0);
Vivek Goyale43473b2010-09-15 17:06:35 -04001027
1028 /* Init root group */
1029 tg = &td->root_tg;
1030 INIT_HLIST_NODE(&tg->tg_node);
1031 RB_CLEAR_NODE(&tg->rb_node);
1032 bio_list_init(&tg->bio_lists[0]);
1033 bio_list_init(&tg->bio_lists[1]);
1034
1035 /* Practically unlimited BW */
1036 tg->bps[0] = tg->bps[1] = -1;
Vivek Goyal8e89d132010-09-15 17:06:37 -04001037 tg->iops[0] = tg->iops[1] = -1;
Vivek Goyal02977e42010-10-01 14:49:48 +02001038
1039 /*
1040 * Set root group reference to 2. One reference will be dropped when
1041 * all groups on tg_list are being deleted during queue exit. Other
1042 * reference will remain there as we don't want to delete this group
1043 * as it is statically allocated and gets destroyed when throtl_data
1044 * goes away.
1045 */
1046 atomic_set(&tg->ref, 2);
1047 hlist_add_head(&tg->tg_node, &td->tg_list);
1048 td->nr_undestroyed_grps++;
Vivek Goyale43473b2010-09-15 17:06:35 -04001049
1050 INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);
1051
1052 rcu_read_lock();
1053 blkiocg_add_blkio_group(&blkio_root_cgroup, &tg->blkg, (void *)td,
1054 0, BLKIO_POLICY_THROTL);
1055 rcu_read_unlock();
1056
1057 /* Attach throtl data to request queue */
1058 td->queue = q;
1059 q->td = td;
1060 return 0;
1061}
1062
1063void blk_throtl_exit(struct request_queue *q)
1064{
1065 struct throtl_data *td = q->td;
1066 bool wait = false;
1067
1068 BUG_ON(!td);
1069
1070 throtl_shutdown_timer_wq(q);
1071
1072 spin_lock_irq(q->queue_lock);
1073 throtl_release_tgs(td);
Vivek Goyale43473b2010-09-15 17:06:35 -04001074
1075 /* If there are other groups */
Vivek Goyal02977e42010-10-01 14:49:48 +02001076 if (td->nr_undestroyed_grps > 0)
Vivek Goyale43473b2010-09-15 17:06:35 -04001077 wait = true;
1078
1079 spin_unlock_irq(q->queue_lock);
1080
1081 /*
1082 * Wait for tg->blkg->key accessors to exit their grace periods.
1083 * Do this wait only if there are other undestroyed groups out
1084 * there (other than root group). This can happen if cgroup deletion
1085 * path claimed the responsibility of cleaning up a group before
1086 * queue cleanup code get to the group.
1087 *
1088 * Do not call synchronize_rcu() unconditionally as there are drivers
1089 * which create/delete request queue hundreds of times during scan/boot
1090 * and synchronize_rcu() can take significant time and slow down boot.
1091 */
1092 if (wait)
1093 synchronize_rcu();
Vivek Goyalfe071432010-10-01 14:49:49 +02001094
1095 /*
1096 * Just being safe to make sure after previous flush if some body did
1097 * update limits through cgroup and another work got queued, cancel
1098 * it.
1099 */
1100 throtl_shutdown_timer_wq(q);
Vivek Goyale43473b2010-09-15 17:06:35 -04001101 throtl_td_free(td);
1102}
1103
1104static int __init throtl_init(void)
1105{
1106 blkio_policy_register(&blkio_policy_throtl);
1107 return 0;
1108}
1109
1110module_init(throtl_init);