blob: 0a0bc00059df4793c862241db7918e877ceb3569 [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"
Tejun Heobc9fcbf2011-10-19 14:31:18 +020013#include "blk.h"
Vivek Goyale43473b2010-09-15 17:06:35 -040014
15/* Max dispatch from a group in 1 round */
16static int throtl_grp_quantum = 8;
17
18/* Total max dispatch from all groups in one round */
19static int throtl_quantum = 32;
20
21/* Throttling is performed over 100ms slice and after that slice is renewed */
22static unsigned long throtl_slice = HZ/10; /* 100 ms */
23
Tejun Heo3c798392012-04-16 13:57:25 -070024static struct blkcg_policy blkcg_policy_throtl;
Tejun Heo03814112012-03-05 13:15:14 -080025
Vivek Goyal450adcb2011-03-01 13:40:54 -050026/* A workqueue to queue throttle related work */
27static struct workqueue_struct *kthrotld_workqueue;
28static void throtl_schedule_delayed_work(struct throtl_data *td,
29 unsigned long delay);
30
Vivek Goyale43473b2010-09-15 17:06:35 -040031struct throtl_rb_root {
32 struct rb_root rb;
33 struct rb_node *left;
34 unsigned int count;
35 unsigned long min_disptime;
36};
37
38#define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \
39 .count = 0, .min_disptime = 0}
40
41#define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
42
Tejun Heo8a3d2612012-04-01 14:38:44 -070043/* Per-cpu group stats */
44struct tg_stats_cpu {
45 /* total bytes transferred */
46 struct blkg_rwstat service_bytes;
47 /* total IOs serviced, post merge */
48 struct blkg_rwstat serviced;
49};
50
Vivek Goyale43473b2010-09-15 17:06:35 -040051struct throtl_grp {
Tejun Heof95a04a2012-04-16 13:57:26 -070052 /* must be the first member */
53 struct blkg_policy_data pd;
54
Vivek Goyale43473b2010-09-15 17:06:35 -040055 /* active throtl group service_tree member */
56 struct rb_node rb_node;
57
58 /*
59 * Dispatch time in jiffies. This is the estimated time when group
60 * will unthrottle and is ready to dispatch more bio. It is used as
61 * key to sort active groups in service tree.
62 */
63 unsigned long disptime;
64
Vivek Goyale43473b2010-09-15 17:06:35 -040065 unsigned int flags;
66
67 /* Two lists for READ and WRITE */
68 struct bio_list bio_lists[2];
69
70 /* Number of queued bios on READ and WRITE lists */
71 unsigned int nr_queued[2];
72
73 /* bytes per second rate limits */
74 uint64_t bps[2];
75
Vivek Goyal8e89d132010-09-15 17:06:37 -040076 /* IOPS limits */
77 unsigned int iops[2];
78
Vivek Goyale43473b2010-09-15 17:06:35 -040079 /* Number of bytes disptached in current slice */
80 uint64_t bytes_disp[2];
Vivek Goyal8e89d132010-09-15 17:06:37 -040081 /* Number of bio's dispatched in current slice */
82 unsigned int io_disp[2];
Vivek Goyale43473b2010-09-15 17:06:35 -040083
84 /* When did we start a new slice */
85 unsigned long slice_start[2];
86 unsigned long slice_end[2];
Vivek Goyalfe071432010-10-01 14:49:49 +020087
Tejun Heo8a3d2612012-04-01 14:38:44 -070088 /* Per cpu stats pointer */
89 struct tg_stats_cpu __percpu *stats_cpu;
90
91 /* List of tgs waiting for per cpu stats memory to be allocated */
92 struct list_head stats_alloc_node;
Vivek Goyale43473b2010-09-15 17:06:35 -040093};
94
95struct throtl_data
96{
Vivek Goyale43473b2010-09-15 17:06:35 -040097 /* service tree for active throtl groups */
98 struct throtl_rb_root tg_service_tree;
99
Vivek Goyale43473b2010-09-15 17:06:35 -0400100 struct request_queue *queue;
101
102 /* Total Number of queued bios on READ and WRITE lists */
103 unsigned int nr_queued[2];
104
105 /*
Vivek Goyal02977e42010-10-01 14:49:48 +0200106 * number of total undestroyed groups
Vivek Goyale43473b2010-09-15 17:06:35 -0400107 */
108 unsigned int nr_undestroyed_grps;
109
110 /* Work for dispatching throttled bios */
Tejun Heocb761992013-05-14 13:52:31 -0700111 struct delayed_work dispatch_work;
Vivek Goyale43473b2010-09-15 17:06:35 -0400112};
113
Tejun Heo8a3d2612012-04-01 14:38:44 -0700114/* list and work item to allocate percpu group stats */
115static DEFINE_SPINLOCK(tg_stats_alloc_lock);
116static LIST_HEAD(tg_stats_alloc_list);
117
118static void tg_stats_alloc_fn(struct work_struct *);
119static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn);
120
Tejun Heof95a04a2012-04-16 13:57:26 -0700121static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
122{
123 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
124}
125
Tejun Heo3c798392012-04-16 13:57:25 -0700126static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
Tejun Heo03814112012-03-05 13:15:14 -0800127{
Tejun Heof95a04a2012-04-16 13:57:26 -0700128 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
Tejun Heo03814112012-03-05 13:15:14 -0800129}
130
Tejun Heo3c798392012-04-16 13:57:25 -0700131static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
Tejun Heo03814112012-03-05 13:15:14 -0800132{
Tejun Heof95a04a2012-04-16 13:57:26 -0700133 return pd_to_blkg(&tg->pd);
Tejun Heo03814112012-03-05 13:15:14 -0800134}
135
Tejun Heo03d8e112012-04-13 13:11:32 -0700136static inline struct throtl_grp *td_root_tg(struct throtl_data *td)
137{
138 return blkg_to_tg(td->queue->root_blkg);
139}
140
Vivek Goyale43473b2010-09-15 17:06:35 -0400141enum tg_state_flags {
142 THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */
143};
144
145#define THROTL_TG_FNS(name) \
146static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \
147{ \
148 (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \
149} \
150static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \
151{ \
152 (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \
153} \
154static inline int throtl_tg_##name(const struct throtl_grp *tg) \
155{ \
156 return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \
157}
158
159THROTL_TG_FNS(on_rr);
160
Tejun Heo54e7ed12012-04-16 13:57:23 -0700161#define throtl_log_tg(td, tg, fmt, args...) do { \
162 char __pbuf[128]; \
163 \
164 blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \
165 blk_add_trace_msg((td)->queue, "throtl %s " fmt, __pbuf, ##args); \
166} while (0)
Vivek Goyale43473b2010-09-15 17:06:35 -0400167
168#define throtl_log(td, fmt, args...) \
169 blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
170
Joe Perchesd2f31a52011-06-13 20:19:27 +0200171static inline unsigned int total_nr_queued(struct throtl_data *td)
Vivek Goyale43473b2010-09-15 17:06:35 -0400172{
Joe Perchesd2f31a52011-06-13 20:19:27 +0200173 return td->nr_queued[0] + td->nr_queued[1];
Vivek Goyale43473b2010-09-15 17:06:35 -0400174}
175
Tejun Heo8a3d2612012-04-01 14:38:44 -0700176/*
177 * Worker for allocating per cpu stat for tgs. This is scheduled on the
Tejun Heo3b07e9c2012-08-20 14:51:24 -0700178 * system_wq once there are some groups on the alloc_list waiting for
Tejun Heo8a3d2612012-04-01 14:38:44 -0700179 * allocation.
180 */
181static void tg_stats_alloc_fn(struct work_struct *work)
182{
183 static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */
184 struct delayed_work *dwork = to_delayed_work(work);
185 bool empty = false;
186
187alloc_stats:
188 if (!stats_cpu) {
189 stats_cpu = alloc_percpu(struct tg_stats_cpu);
190 if (!stats_cpu) {
191 /* allocation failed, try again after some time */
Tejun Heo3b07e9c2012-08-20 14:51:24 -0700192 schedule_delayed_work(dwork, msecs_to_jiffies(10));
Tejun Heo8a3d2612012-04-01 14:38:44 -0700193 return;
194 }
195 }
196
197 spin_lock_irq(&tg_stats_alloc_lock);
198
199 if (!list_empty(&tg_stats_alloc_list)) {
200 struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list,
201 struct throtl_grp,
202 stats_alloc_node);
203 swap(tg->stats_cpu, stats_cpu);
204 list_del_init(&tg->stats_alloc_node);
205 }
206
207 empty = list_empty(&tg_stats_alloc_list);
208 spin_unlock_irq(&tg_stats_alloc_lock);
209 if (!empty)
210 goto alloc_stats;
211}
212
Tejun Heo3c798392012-04-16 13:57:25 -0700213static void throtl_pd_init(struct blkcg_gq *blkg)
Vivek Goyala29a1712011-05-19 15:38:19 -0400214{
Tejun Heo03814112012-03-05 13:15:14 -0800215 struct throtl_grp *tg = blkg_to_tg(blkg);
Tejun Heoff26eaa2012-05-23 12:16:21 +0200216 unsigned long flags;
Tejun Heocd1604f2012-03-05 13:15:06 -0800217
Vivek Goyala29a1712011-05-19 15:38:19 -0400218 RB_CLEAR_NODE(&tg->rb_node);
219 bio_list_init(&tg->bio_lists[0]);
220 bio_list_init(&tg->bio_lists[1]);
Vivek Goyala29a1712011-05-19 15:38:19 -0400221
Tejun Heoe56da7e2012-03-05 13:15:07 -0800222 tg->bps[READ] = -1;
223 tg->bps[WRITE] = -1;
224 tg->iops[READ] = -1;
225 tg->iops[WRITE] = -1;
Tejun Heo8a3d2612012-04-01 14:38:44 -0700226
227 /*
228 * Ugh... We need to perform per-cpu allocation for tg->stats_cpu
229 * but percpu allocator can't be called from IO path. Queue tg on
230 * tg_stats_alloc_list and allocate from work item.
231 */
Tejun Heoff26eaa2012-05-23 12:16:21 +0200232 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
Tejun Heo8a3d2612012-04-01 14:38:44 -0700233 list_add(&tg->stats_alloc_node, &tg_stats_alloc_list);
Tejun Heo3b07e9c2012-08-20 14:51:24 -0700234 schedule_delayed_work(&tg_stats_alloc_work, 0);
Tejun Heoff26eaa2012-05-23 12:16:21 +0200235 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
Tejun Heo8a3d2612012-04-01 14:38:44 -0700236}
237
Tejun Heo3c798392012-04-16 13:57:25 -0700238static void throtl_pd_exit(struct blkcg_gq *blkg)
Tejun Heo8a3d2612012-04-01 14:38:44 -0700239{
240 struct throtl_grp *tg = blkg_to_tg(blkg);
Tejun Heoff26eaa2012-05-23 12:16:21 +0200241 unsigned long flags;
Tejun Heo8a3d2612012-04-01 14:38:44 -0700242
Tejun Heoff26eaa2012-05-23 12:16:21 +0200243 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
Tejun Heo8a3d2612012-04-01 14:38:44 -0700244 list_del_init(&tg->stats_alloc_node);
Tejun Heoff26eaa2012-05-23 12:16:21 +0200245 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
Tejun Heo8a3d2612012-04-01 14:38:44 -0700246
247 free_percpu(tg->stats_cpu);
248}
249
Tejun Heo3c798392012-04-16 13:57:25 -0700250static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
Tejun Heo8a3d2612012-04-01 14:38:44 -0700251{
252 struct throtl_grp *tg = blkg_to_tg(blkg);
253 int cpu;
254
255 if (tg->stats_cpu == NULL)
256 return;
257
258 for_each_possible_cpu(cpu) {
259 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
260
261 blkg_rwstat_reset(&sc->service_bytes);
262 blkg_rwstat_reset(&sc->serviced);
263 }
Vivek Goyala29a1712011-05-19 15:38:19 -0400264}
265
Tejun Heo3c798392012-04-16 13:57:25 -0700266static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
267 struct blkcg *blkcg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400268{
Vivek Goyale43473b2010-09-15 17:06:35 -0400269 /*
Tejun Heo3c798392012-04-16 13:57:25 -0700270 * This is the common case when there are no blkcgs. Avoid lookup
271 * in this case
Tejun Heocd1604f2012-03-05 13:15:06 -0800272 */
Tejun Heo3c798392012-04-16 13:57:25 -0700273 if (blkcg == &blkcg_root)
Tejun Heo03d8e112012-04-13 13:11:32 -0700274 return td_root_tg(td);
Vivek Goyale43473b2010-09-15 17:06:35 -0400275
Tejun Heoe8989fa2012-03-05 13:15:20 -0800276 return blkg_to_tg(blkg_lookup(blkcg, td->queue));
Vivek Goyale43473b2010-09-15 17:06:35 -0400277}
278
Tejun Heocd1604f2012-03-05 13:15:06 -0800279static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
Tejun Heo3c798392012-04-16 13:57:25 -0700280 struct blkcg *blkcg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400281{
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400282 struct request_queue *q = td->queue;
Tejun Heocd1604f2012-03-05 13:15:06 -0800283 struct throtl_grp *tg = NULL;
Tejun Heo0a5a7d02012-03-05 13:15:02 -0800284
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400285 /*
Tejun Heo3c798392012-04-16 13:57:25 -0700286 * This is the common case when there are no blkcgs. Avoid lookup
287 * in this case
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400288 */
Tejun Heo3c798392012-04-16 13:57:25 -0700289 if (blkcg == &blkcg_root) {
Tejun Heo03d8e112012-04-13 13:11:32 -0700290 tg = td_root_tg(td);
Tejun Heocd1604f2012-03-05 13:15:06 -0800291 } else {
Tejun Heo3c798392012-04-16 13:57:25 -0700292 struct blkcg_gq *blkg;
Tejun Heocd1604f2012-03-05 13:15:06 -0800293
Tejun Heo3c96cb32012-04-13 13:11:34 -0700294 blkg = blkg_lookup_create(blkcg, q);
Tejun Heocd1604f2012-03-05 13:15:06 -0800295
296 /* if %NULL and @q is alive, fall back to root_tg */
297 if (!IS_ERR(blkg))
Tejun Heo03814112012-03-05 13:15:14 -0800298 tg = blkg_to_tg(blkg);
Bart Van Assche3f3299d2012-11-28 13:42:38 +0100299 else if (!blk_queue_dying(q))
Tejun Heo03d8e112012-04-13 13:11:32 -0700300 tg = td_root_tg(td);
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400301 }
302
Vivek Goyale43473b2010-09-15 17:06:35 -0400303 return tg;
304}
305
306static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root)
307{
308 /* Service tree is empty */
309 if (!root->count)
310 return NULL;
311
312 if (!root->left)
313 root->left = rb_first(&root->rb);
314
315 if (root->left)
316 return rb_entry_tg(root->left);
317
318 return NULL;
319}
320
321static void rb_erase_init(struct rb_node *n, struct rb_root *root)
322{
323 rb_erase(n, root);
324 RB_CLEAR_NODE(n);
325}
326
327static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root)
328{
329 if (root->left == n)
330 root->left = NULL;
331 rb_erase_init(n, &root->rb);
332 --root->count;
333}
334
335static void update_min_dispatch_time(struct throtl_rb_root *st)
336{
337 struct throtl_grp *tg;
338
339 tg = throtl_rb_first(st);
340 if (!tg)
341 return;
342
343 st->min_disptime = tg->disptime;
344}
345
346static void
347tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg)
348{
349 struct rb_node **node = &st->rb.rb_node;
350 struct rb_node *parent = NULL;
351 struct throtl_grp *__tg;
352 unsigned long key = tg->disptime;
353 int left = 1;
354
355 while (*node != NULL) {
356 parent = *node;
357 __tg = rb_entry_tg(parent);
358
359 if (time_before(key, __tg->disptime))
360 node = &parent->rb_left;
361 else {
362 node = &parent->rb_right;
363 left = 0;
364 }
365 }
366
367 if (left)
368 st->left = &tg->rb_node;
369
370 rb_link_node(&tg->rb_node, parent, node);
371 rb_insert_color(&tg->rb_node, &st->rb);
372}
373
374static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
375{
376 struct throtl_rb_root *st = &td->tg_service_tree;
377
378 tg_service_tree_add(st, tg);
379 throtl_mark_tg_on_rr(tg);
380 st->count++;
381}
382
383static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
384{
385 if (!throtl_tg_on_rr(tg))
386 __throtl_enqueue_tg(td, tg);
387}
388
389static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
390{
391 throtl_rb_erase(&tg->rb_node, &td->tg_service_tree);
392 throtl_clear_tg_on_rr(tg);
393}
394
395static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
396{
397 if (throtl_tg_on_rr(tg))
398 __throtl_dequeue_tg(td, tg);
399}
400
401static void throtl_schedule_next_dispatch(struct throtl_data *td)
402{
403 struct throtl_rb_root *st = &td->tg_service_tree;
404
405 /*
406 * If there are more bios pending, schedule more work.
407 */
408 if (!total_nr_queued(td))
409 return;
410
411 BUG_ON(!st->count);
412
413 update_min_dispatch_time(st);
414
415 if (time_before_eq(st->min_disptime, jiffies))
Vivek Goyal450adcb2011-03-01 13:40:54 -0500416 throtl_schedule_delayed_work(td, 0);
Vivek Goyale43473b2010-09-15 17:06:35 -0400417 else
Vivek Goyal450adcb2011-03-01 13:40:54 -0500418 throtl_schedule_delayed_work(td, (st->min_disptime - jiffies));
Vivek Goyale43473b2010-09-15 17:06:35 -0400419}
420
421static inline void
422throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
423{
424 tg->bytes_disp[rw] = 0;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400425 tg->io_disp[rw] = 0;
Vivek Goyale43473b2010-09-15 17:06:35 -0400426 tg->slice_start[rw] = jiffies;
427 tg->slice_end[rw] = jiffies + throtl_slice;
428 throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
429 rw == READ ? 'R' : 'W', tg->slice_start[rw],
430 tg->slice_end[rw], jiffies);
431}
432
Vivek Goyald1ae8ff2010-12-01 19:34:46 +0100433static inline void throtl_set_slice_end(struct throtl_data *td,
434 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
435{
436 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
437}
438
Vivek Goyale43473b2010-09-15 17:06:35 -0400439static inline void throtl_extend_slice(struct throtl_data *td,
440 struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
441{
442 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
443 throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
444 rw == READ ? 'R' : 'W', tg->slice_start[rw],
445 tg->slice_end[rw], jiffies);
446}
447
448/* Determine if previously allocated or extended slice is complete or not */
449static bool
450throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
451{
452 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
453 return 0;
454
455 return 1;
456}
457
458/* Trim the used slices and adjust slice start accordingly */
459static inline void
460throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
461{
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200462 unsigned long nr_slices, time_elapsed, io_trim;
463 u64 bytes_trim, tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400464
465 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
466
467 /*
468 * If bps are unlimited (-1), then time slice don't get
469 * renewed. Don't try to trim the slice if slice is used. A new
470 * slice will start when appropriate.
471 */
472 if (throtl_slice_used(td, tg, rw))
473 return;
474
Vivek Goyald1ae8ff2010-12-01 19:34:46 +0100475 /*
476 * A bio has been dispatched. Also adjust slice_end. It might happen
477 * that initially cgroup limit was very low resulting in high
478 * slice_end, but later limit was bumped up and bio was dispached
479 * sooner, then we need to reduce slice_end. A high bogus slice_end
480 * is bad because it does not allow new slice to start.
481 */
482
483 throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice);
484
Vivek Goyale43473b2010-09-15 17:06:35 -0400485 time_elapsed = jiffies - tg->slice_start[rw];
486
487 nr_slices = time_elapsed / throtl_slice;
488
489 if (!nr_slices)
490 return;
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200491 tmp = tg->bps[rw] * throtl_slice * nr_slices;
492 do_div(tmp, HZ);
493 bytes_trim = tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400494
Vivek Goyal8e89d132010-09-15 17:06:37 -0400495 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
Vivek Goyale43473b2010-09-15 17:06:35 -0400496
Vivek Goyal8e89d132010-09-15 17:06:37 -0400497 if (!bytes_trim && !io_trim)
Vivek Goyale43473b2010-09-15 17:06:35 -0400498 return;
499
500 if (tg->bytes_disp[rw] >= bytes_trim)
501 tg->bytes_disp[rw] -= bytes_trim;
502 else
503 tg->bytes_disp[rw] = 0;
504
Vivek Goyal8e89d132010-09-15 17:06:37 -0400505 if (tg->io_disp[rw] >= io_trim)
506 tg->io_disp[rw] -= io_trim;
507 else
508 tg->io_disp[rw] = 0;
509
Vivek Goyale43473b2010-09-15 17:06:35 -0400510 tg->slice_start[rw] += nr_slices * throtl_slice;
511
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200512 throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
Vivek Goyale43473b2010-09-15 17:06:35 -0400513 " start=%lu end=%lu jiffies=%lu",
Vivek Goyal8e89d132010-09-15 17:06:37 -0400514 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
Vivek Goyale43473b2010-09-15 17:06:35 -0400515 tg->slice_start[rw], tg->slice_end[rw], jiffies);
516}
517
Vivek Goyal8e89d132010-09-15 17:06:37 -0400518static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
519 struct bio *bio, unsigned long *wait)
Vivek Goyale43473b2010-09-15 17:06:35 -0400520{
521 bool rw = bio_data_dir(bio);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400522 unsigned int io_allowed;
Vivek Goyale43473b2010-09-15 17:06:35 -0400523 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
Vivek Goyalc49c06e2010-10-01 21:16:42 +0200524 u64 tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400525
Vivek Goyal8e89d132010-09-15 17:06:37 -0400526 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
Vivek Goyale43473b2010-09-15 17:06:35 -0400527
Vivek Goyal8e89d132010-09-15 17:06:37 -0400528 /* Slice has just started. Consider one slice interval */
529 if (!jiffy_elapsed)
530 jiffy_elapsed_rnd = throtl_slice;
531
532 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
533
Vivek Goyalc49c06e2010-10-01 21:16:42 +0200534 /*
535 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
536 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
537 * will allow dispatch after 1 second and after that slice should
538 * have been trimmed.
539 */
540
541 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
542 do_div(tmp, HZ);
543
544 if (tmp > UINT_MAX)
545 io_allowed = UINT_MAX;
546 else
547 io_allowed = tmp;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400548
549 if (tg->io_disp[rw] + 1 <= io_allowed) {
Vivek Goyale43473b2010-09-15 17:06:35 -0400550 if (wait)
551 *wait = 0;
552 return 1;
553 }
554
Vivek Goyal8e89d132010-09-15 17:06:37 -0400555 /* Calc approx time to dispatch */
556 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
557
558 if (jiffy_wait > jiffy_elapsed)
559 jiffy_wait = jiffy_wait - jiffy_elapsed;
560 else
561 jiffy_wait = 1;
562
563 if (wait)
564 *wait = jiffy_wait;
565 return 0;
566}
567
568static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
569 struct bio *bio, unsigned long *wait)
570{
571 bool rw = bio_data_dir(bio);
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200572 u64 bytes_allowed, extra_bytes, tmp;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400573 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
Vivek Goyale43473b2010-09-15 17:06:35 -0400574
575 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
576
577 /* Slice has just started. Consider one slice interval */
578 if (!jiffy_elapsed)
579 jiffy_elapsed_rnd = throtl_slice;
580
581 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
582
Vivek Goyal5e901a22010-10-01 21:16:38 +0200583 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
584 do_div(tmp, HZ);
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200585 bytes_allowed = tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400586
587 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
588 if (wait)
589 *wait = 0;
590 return 1;
591 }
592
593 /* Calc approx time to dispatch */
594 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
595 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
596
597 if (!jiffy_wait)
598 jiffy_wait = 1;
599
600 /*
601 * This wait time is without taking into consideration the rounding
602 * up we did. Add that time also.
603 */
604 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
Vivek Goyale43473b2010-09-15 17:06:35 -0400605 if (wait)
606 *wait = jiffy_wait;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400607 return 0;
608}
Vivek Goyale43473b2010-09-15 17:06:35 -0400609
Vivek Goyalaf75cd32011-05-19 15:38:31 -0400610static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
611 if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
612 return 1;
613 return 0;
614}
615
Vivek Goyal8e89d132010-09-15 17:06:37 -0400616/*
617 * Returns whether one can dispatch a bio or not. Also returns approx number
618 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
619 */
620static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
621 struct bio *bio, unsigned long *wait)
622{
623 bool rw = bio_data_dir(bio);
624 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
625
626 /*
627 * Currently whole state machine of group depends on first bio
628 * queued in the group bio list. So one should not be calling
629 * this function with a different bio if there are other bios
630 * queued.
631 */
632 BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
633
634 /* If tg->bps = -1, then BW is unlimited */
635 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
636 if (wait)
637 *wait = 0;
638 return 1;
639 }
640
641 /*
642 * If previous slice expired, start a new one otherwise renew/extend
643 * existing slice to make sure it is at least throtl_slice interval
644 * long since now.
645 */
646 if (throtl_slice_used(td, tg, rw))
647 throtl_start_new_slice(td, tg, rw);
648 else {
649 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
650 throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
651 }
652
653 if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
654 && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
655 if (wait)
656 *wait = 0;
657 return 1;
658 }
659
660 max_wait = max(bps_wait, iops_wait);
661
662 if (wait)
663 *wait = max_wait;
664
665 if (time_before(tg->slice_end[rw], jiffies + max_wait))
666 throtl_extend_slice(td, tg, rw, jiffies + max_wait);
Vivek Goyale43473b2010-09-15 17:06:35 -0400667
668 return 0;
669}
670
Tejun Heo3c798392012-04-16 13:57:25 -0700671static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
Tejun Heo629ed0b2012-04-01 14:38:44 -0700672 int rw)
673{
Tejun Heo8a3d2612012-04-01 14:38:44 -0700674 struct throtl_grp *tg = blkg_to_tg(blkg);
675 struct tg_stats_cpu *stats_cpu;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700676 unsigned long flags;
677
678 /* If per cpu stats are not allocated yet, don't do any accounting. */
Tejun Heo8a3d2612012-04-01 14:38:44 -0700679 if (tg->stats_cpu == NULL)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700680 return;
681
682 /*
683 * Disabling interrupts to provide mutual exclusion between two
684 * writes on same cpu. It probably is not needed for 64bit. Not
685 * optimizing that case yet.
686 */
687 local_irq_save(flags);
688
Tejun Heo8a3d2612012-04-01 14:38:44 -0700689 stats_cpu = this_cpu_ptr(tg->stats_cpu);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700690
Tejun Heo629ed0b2012-04-01 14:38:44 -0700691 blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
692 blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
693
694 local_irq_restore(flags);
695}
696
Vivek Goyale43473b2010-09-15 17:06:35 -0400697static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
698{
699 bool rw = bio_data_dir(bio);
Vivek Goyale43473b2010-09-15 17:06:35 -0400700
701 /* Charge the bio to the group */
702 tg->bytes_disp[rw] += bio->bi_size;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400703 tg->io_disp[rw]++;
Vivek Goyale43473b2010-09-15 17:06:35 -0400704
Tejun Heo629ed0b2012-04-01 14:38:44 -0700705 throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw);
Vivek Goyale43473b2010-09-15 17:06:35 -0400706}
707
708static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
709 struct bio *bio)
710{
711 bool rw = bio_data_dir(bio);
712
713 bio_list_add(&tg->bio_lists[rw], bio);
714 /* Take a bio reference on tg */
Tejun Heo1adaf3d2012-03-05 13:15:15 -0800715 blkg_get(tg_to_blkg(tg));
Vivek Goyale43473b2010-09-15 17:06:35 -0400716 tg->nr_queued[rw]++;
717 td->nr_queued[rw]++;
718 throtl_enqueue_tg(td, tg);
719}
720
721static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
722{
723 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
724 struct bio *bio;
725
726 if ((bio = bio_list_peek(&tg->bio_lists[READ])))
727 tg_may_dispatch(td, tg, bio, &read_wait);
728
729 if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
730 tg_may_dispatch(td, tg, bio, &write_wait);
731
732 min_wait = min(read_wait, write_wait);
733 disptime = jiffies + min_wait;
734
Vivek Goyale43473b2010-09-15 17:06:35 -0400735 /* Update dispatch time */
736 throtl_dequeue_tg(td, tg);
737 tg->disptime = disptime;
738 throtl_enqueue_tg(td, tg);
739}
740
741static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
742 bool rw, struct bio_list *bl)
743{
744 struct bio *bio;
745
746 bio = bio_list_pop(&tg->bio_lists[rw]);
747 tg->nr_queued[rw]--;
Tejun Heo1adaf3d2012-03-05 13:15:15 -0800748 /* Drop bio reference on blkg */
749 blkg_put(tg_to_blkg(tg));
Vivek Goyale43473b2010-09-15 17:06:35 -0400750
751 BUG_ON(td->nr_queued[rw] <= 0);
752 td->nr_queued[rw]--;
753
754 throtl_charge_bio(tg, bio);
755 bio_list_add(bl, bio);
756 bio->bi_rw |= REQ_THROTTLED;
757
758 throtl_trim_slice(td, tg, rw);
759}
760
761static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
762 struct bio_list *bl)
763{
764 unsigned int nr_reads = 0, nr_writes = 0;
765 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
Vivek Goyalc2f68052010-11-15 19:32:42 +0100766 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
Vivek Goyale43473b2010-09-15 17:06:35 -0400767 struct bio *bio;
768
769 /* Try to dispatch 75% READS and 25% WRITES */
770
771 while ((bio = bio_list_peek(&tg->bio_lists[READ]))
772 && tg_may_dispatch(td, tg, bio, NULL)) {
773
774 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
775 nr_reads++;
776
777 if (nr_reads >= max_nr_reads)
778 break;
779 }
780
781 while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
782 && tg_may_dispatch(td, tg, bio, NULL)) {
783
784 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
785 nr_writes++;
786
787 if (nr_writes >= max_nr_writes)
788 break;
789 }
790
791 return nr_reads + nr_writes;
792}
793
794static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
795{
796 unsigned int nr_disp = 0;
797 struct throtl_grp *tg;
798 struct throtl_rb_root *st = &td->tg_service_tree;
799
800 while (1) {
801 tg = throtl_rb_first(st);
802
803 if (!tg)
804 break;
805
806 if (time_before(jiffies, tg->disptime))
807 break;
808
809 throtl_dequeue_tg(td, tg);
810
811 nr_disp += throtl_dispatch_tg(td, tg, bl);
812
Tejun Heo2db63142013-05-14 13:52:31 -0700813 if (tg->nr_queued[0] || tg->nr_queued[1])
Vivek Goyale43473b2010-09-15 17:06:35 -0400814 tg_update_disptime(td, tg);
Vivek Goyale43473b2010-09-15 17:06:35 -0400815
816 if (nr_disp >= throtl_quantum)
817 break;
818 }
819
820 return nr_disp;
821}
822
Tejun Heocb761992013-05-14 13:52:31 -0700823/* work function to dispatch throttled bios */
824void blk_throtl_dispatch_work_fn(struct work_struct *work)
Vivek Goyale43473b2010-09-15 17:06:35 -0400825{
Tejun Heocb761992013-05-14 13:52:31 -0700826 struct throtl_data *td = container_of(to_delayed_work(work),
827 struct throtl_data, dispatch_work);
828 struct request_queue *q = td->queue;
Vivek Goyale43473b2010-09-15 17:06:35 -0400829 unsigned int nr_disp = 0;
830 struct bio_list bio_list_on_stack;
831 struct bio *bio;
Vivek Goyal69d60eb2011-03-09 08:27:37 +0100832 struct blk_plug plug;
Vivek Goyale43473b2010-09-15 17:06:35 -0400833
834 spin_lock_irq(q->queue_lock);
835
836 if (!total_nr_queued(td))
837 goto out;
838
839 bio_list_init(&bio_list_on_stack);
840
Joe Perchesd2f31a52011-06-13 20:19:27 +0200841 throtl_log(td, "dispatch nr_queued=%u read=%u write=%u",
Vivek Goyale43473b2010-09-15 17:06:35 -0400842 total_nr_queued(td), td->nr_queued[READ],
843 td->nr_queued[WRITE]);
844
845 nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
846
847 if (nr_disp)
848 throtl_log(td, "bios disp=%u", nr_disp);
849
850 throtl_schedule_next_dispatch(td);
851out:
852 spin_unlock_irq(q->queue_lock);
853
854 /*
855 * If we dispatched some requests, unplug the queue to make sure
856 * immediate dispatch
857 */
858 if (nr_disp) {
Vivek Goyal69d60eb2011-03-09 08:27:37 +0100859 blk_start_plug(&plug);
Vivek Goyale43473b2010-09-15 17:06:35 -0400860 while((bio = bio_list_pop(&bio_list_on_stack)))
861 generic_make_request(bio);
Vivek Goyal69d60eb2011-03-09 08:27:37 +0100862 blk_finish_plug(&plug);
Vivek Goyale43473b2010-09-15 17:06:35 -0400863 }
Vivek Goyale43473b2010-09-15 17:06:35 -0400864}
865
866/* Call with queue lock held */
Vivek Goyal450adcb2011-03-01 13:40:54 -0500867static void
868throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay)
Vivek Goyale43473b2010-09-15 17:06:35 -0400869{
870
Tejun Heocb761992013-05-14 13:52:31 -0700871 struct delayed_work *dwork = &td->dispatch_work;
Vivek Goyale43473b2010-09-15 17:06:35 -0400872
Tejun Heo632b4492013-05-14 13:52:31 -0700873 if (total_nr_queued(td)) {
Tejun Heoe7c2f962012-08-21 13:18:24 -0700874 mod_delayed_work(kthrotld_workqueue, dwork, delay);
Vivek Goyale43473b2010-09-15 17:06:35 -0400875 throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
876 delay, jiffies);
877 }
878}
Vivek Goyale43473b2010-09-15 17:06:35 -0400879
Tejun Heof95a04a2012-04-16 13:57:26 -0700880static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
881 struct blkg_policy_data *pd, int off)
Tejun Heo41b38b62012-04-01 14:38:44 -0700882{
Tejun Heof95a04a2012-04-16 13:57:26 -0700883 struct throtl_grp *tg = pd_to_tg(pd);
Tejun Heo41b38b62012-04-01 14:38:44 -0700884 struct blkg_rwstat rwstat = { }, tmp;
885 int i, cpu;
886
887 for_each_possible_cpu(cpu) {
Tejun Heo8a3d2612012-04-01 14:38:44 -0700888 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
Tejun Heo41b38b62012-04-01 14:38:44 -0700889
890 tmp = blkg_rwstat_read((void *)sc + off);
891 for (i = 0; i < BLKG_RWSTAT_NR; i++)
892 rwstat.cnt[i] += tmp.cnt[i];
893 }
894
Tejun Heof95a04a2012-04-16 13:57:26 -0700895 return __blkg_prfill_rwstat(sf, pd, &rwstat);
Tejun Heo41b38b62012-04-01 14:38:44 -0700896}
897
Tejun Heo8a3d2612012-04-01 14:38:44 -0700898static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft,
899 struct seq_file *sf)
Tejun Heo41b38b62012-04-01 14:38:44 -0700900{
Tejun Heo3c798392012-04-16 13:57:25 -0700901 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
Tejun Heo41b38b62012-04-01 14:38:44 -0700902
Tejun Heo3c798392012-04-16 13:57:25 -0700903 blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl,
Tejun Heo5bc4afb12012-04-01 14:38:45 -0700904 cft->private, true);
Tejun Heo41b38b62012-04-01 14:38:44 -0700905 return 0;
906}
907
Tejun Heof95a04a2012-04-16 13:57:26 -0700908static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
909 int off)
Tejun Heo60c2bc22012-04-01 14:38:43 -0700910{
Tejun Heof95a04a2012-04-16 13:57:26 -0700911 struct throtl_grp *tg = pd_to_tg(pd);
912 u64 v = *(u64 *)((void *)tg + off);
Tejun Heo60c2bc22012-04-01 14:38:43 -0700913
Tejun Heoaf133ce2012-04-01 14:38:44 -0700914 if (v == -1)
Tejun Heo60c2bc22012-04-01 14:38:43 -0700915 return 0;
Tejun Heof95a04a2012-04-16 13:57:26 -0700916 return __blkg_prfill_u64(sf, pd, v);
Tejun Heo60c2bc22012-04-01 14:38:43 -0700917}
918
Tejun Heof95a04a2012-04-16 13:57:26 -0700919static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
920 int off)
Tejun Heo60c2bc22012-04-01 14:38:43 -0700921{
Tejun Heof95a04a2012-04-16 13:57:26 -0700922 struct throtl_grp *tg = pd_to_tg(pd);
923 unsigned int v = *(unsigned int *)((void *)tg + off);
Tejun Heoaf133ce2012-04-01 14:38:44 -0700924
925 if (v == -1)
926 return 0;
Tejun Heof95a04a2012-04-16 13:57:26 -0700927 return __blkg_prfill_u64(sf, pd, v);
Tejun Heoaf133ce2012-04-01 14:38:44 -0700928}
929
930static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft,
931 struct seq_file *sf)
932{
Tejun Heo3c798392012-04-16 13:57:25 -0700933 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64,
934 &blkcg_policy_throtl, cft->private, false);
Tejun Heo60c2bc22012-04-01 14:38:43 -0700935 return 0;
936}
937
Tejun Heoaf133ce2012-04-01 14:38:44 -0700938static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft,
939 struct seq_file *sf)
Vivek Goyale43473b2010-09-15 17:06:35 -0400940{
Tejun Heo3c798392012-04-16 13:57:25 -0700941 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint,
942 &blkcg_policy_throtl, cft->private, false);
Tejun Heoaf133ce2012-04-01 14:38:44 -0700943 return 0;
Vivek Goyale43473b2010-09-15 17:06:35 -0400944}
945
Tejun Heoaf133ce2012-04-01 14:38:44 -0700946static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf,
947 bool is_u64)
Tejun Heo60c2bc22012-04-01 14:38:43 -0700948{
Tejun Heo3c798392012-04-16 13:57:25 -0700949 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
Tejun Heo60c2bc22012-04-01 14:38:43 -0700950 struct blkg_conf_ctx ctx;
Tejun Heoaf133ce2012-04-01 14:38:44 -0700951 struct throtl_grp *tg;
Tejun Heoa2b16932012-04-13 13:11:33 -0700952 struct throtl_data *td;
Tejun Heo60c2bc22012-04-01 14:38:43 -0700953 int ret;
954
Tejun Heo3c798392012-04-16 13:57:25 -0700955 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
Tejun Heo60c2bc22012-04-01 14:38:43 -0700956 if (ret)
957 return ret;
958
Tejun Heoaf133ce2012-04-01 14:38:44 -0700959 tg = blkg_to_tg(ctx.blkg);
Tejun Heoa2b16932012-04-13 13:11:33 -0700960 td = ctx.blkg->q->td;
Tejun Heoaf133ce2012-04-01 14:38:44 -0700961
Tejun Heoa2b16932012-04-13 13:11:33 -0700962 if (!ctx.v)
963 ctx.v = -1;
Tejun Heoaf133ce2012-04-01 14:38:44 -0700964
Tejun Heoa2b16932012-04-13 13:11:33 -0700965 if (is_u64)
966 *(u64 *)((void *)tg + cft->private) = ctx.v;
967 else
968 *(unsigned int *)((void *)tg + cft->private) = ctx.v;
Tejun Heoaf133ce2012-04-01 14:38:44 -0700969
Tejun Heo632b4492013-05-14 13:52:31 -0700970 throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
971 tg->bps[READ], tg->bps[WRITE],
972 tg->iops[READ], tg->iops[WRITE]);
973
974 /*
975 * We're already holding queue_lock and know @tg is valid. Let's
976 * apply the new config directly.
977 *
978 * Restart the slices for both READ and WRITES. It might happen
979 * that a group's limit are dropped suddenly and we don't want to
980 * account recently dispatched IO with new low rate.
981 */
982 throtl_start_new_slice(td, tg, 0);
983 throtl_start_new_slice(td, tg, 1);
984
985 if (throtl_tg_on_rr(tg)) {
986 tg_update_disptime(td, tg);
987 throtl_schedule_next_dispatch(td);
988 }
Tejun Heo60c2bc22012-04-01 14:38:43 -0700989
990 blkg_conf_finish(&ctx);
Tejun Heoa2b16932012-04-13 13:11:33 -0700991 return 0;
Tejun Heo60c2bc22012-04-01 14:38:43 -0700992}
993
Tejun Heoaf133ce2012-04-01 14:38:44 -0700994static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft,
995 const char *buf)
Tejun Heo60c2bc22012-04-01 14:38:43 -0700996{
Tejun Heoaf133ce2012-04-01 14:38:44 -0700997 return tg_set_conf(cgrp, cft, buf, true);
Tejun Heo60c2bc22012-04-01 14:38:43 -0700998}
999
Tejun Heoaf133ce2012-04-01 14:38:44 -07001000static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1001 const char *buf)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001002{
Tejun Heoaf133ce2012-04-01 14:38:44 -07001003 return tg_set_conf(cgrp, cft, buf, false);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001004}
1005
1006static struct cftype throtl_files[] = {
1007 {
1008 .name = "throttle.read_bps_device",
Tejun Heoaf133ce2012-04-01 14:38:44 -07001009 .private = offsetof(struct throtl_grp, bps[READ]),
1010 .read_seq_string = tg_print_conf_u64,
1011 .write_string = tg_set_conf_u64,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001012 .max_write_len = 256,
1013 },
1014 {
1015 .name = "throttle.write_bps_device",
Tejun Heoaf133ce2012-04-01 14:38:44 -07001016 .private = offsetof(struct throtl_grp, bps[WRITE]),
1017 .read_seq_string = tg_print_conf_u64,
1018 .write_string = tg_set_conf_u64,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001019 .max_write_len = 256,
1020 },
1021 {
1022 .name = "throttle.read_iops_device",
Tejun Heoaf133ce2012-04-01 14:38:44 -07001023 .private = offsetof(struct throtl_grp, iops[READ]),
1024 .read_seq_string = tg_print_conf_uint,
1025 .write_string = tg_set_conf_uint,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001026 .max_write_len = 256,
1027 },
1028 {
1029 .name = "throttle.write_iops_device",
Tejun Heoaf133ce2012-04-01 14:38:44 -07001030 .private = offsetof(struct throtl_grp, iops[WRITE]),
1031 .read_seq_string = tg_print_conf_uint,
1032 .write_string = tg_set_conf_uint,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001033 .max_write_len = 256,
1034 },
1035 {
1036 .name = "throttle.io_service_bytes",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001037 .private = offsetof(struct tg_stats_cpu, service_bytes),
Tejun Heo8a3d2612012-04-01 14:38:44 -07001038 .read_seq_string = tg_print_cpu_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001039 },
1040 {
1041 .name = "throttle.io_serviced",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001042 .private = offsetof(struct tg_stats_cpu, serviced),
Tejun Heo8a3d2612012-04-01 14:38:44 -07001043 .read_seq_string = tg_print_cpu_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001044 },
1045 { } /* terminate */
1046};
1047
Vivek Goyalda527772011-03-02 19:05:33 -05001048static void throtl_shutdown_wq(struct request_queue *q)
Vivek Goyale43473b2010-09-15 17:06:35 -04001049{
1050 struct throtl_data *td = q->td;
1051
Tejun Heocb761992013-05-14 13:52:31 -07001052 cancel_delayed_work_sync(&td->dispatch_work);
Vivek Goyale43473b2010-09-15 17:06:35 -04001053}
1054
Tejun Heo3c798392012-04-16 13:57:25 -07001055static struct blkcg_policy blkcg_policy_throtl = {
Tejun Heof9fcc2d2012-04-16 13:57:27 -07001056 .pd_size = sizeof(struct throtl_grp),
1057 .cftypes = throtl_files,
1058
1059 .pd_init_fn = throtl_pd_init,
1060 .pd_exit_fn = throtl_pd_exit,
1061 .pd_reset_stats_fn = throtl_pd_reset_stats,
Vivek Goyale43473b2010-09-15 17:06:35 -04001062};
1063
Tejun Heobc16a4f2011-10-19 14:33:01 +02001064bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
Vivek Goyale43473b2010-09-15 17:06:35 -04001065{
1066 struct throtl_data *td = q->td;
1067 struct throtl_grp *tg;
Vivek Goyale43473b2010-09-15 17:06:35 -04001068 bool rw = bio_data_dir(bio), update_disptime = true;
Tejun Heo3c798392012-04-16 13:57:25 -07001069 struct blkcg *blkcg;
Tejun Heobc16a4f2011-10-19 14:33:01 +02001070 bool throttled = false;
Vivek Goyale43473b2010-09-15 17:06:35 -04001071
1072 if (bio->bi_rw & REQ_THROTTLED) {
1073 bio->bi_rw &= ~REQ_THROTTLED;
Tejun Heobc16a4f2011-10-19 14:33:01 +02001074 goto out;
Vivek Goyale43473b2010-09-15 17:06:35 -04001075 }
1076
Vivek Goyalaf75cd32011-05-19 15:38:31 -04001077 /*
1078 * A throtl_grp pointer retrieved under rcu can be used to access
1079 * basic fields like stats and io rates. If a group has no rules,
1080 * just update the dispatch stats in lockless manner and return.
1081 */
Vivek Goyalaf75cd32011-05-19 15:38:31 -04001082 rcu_read_lock();
Tejun Heo3c798392012-04-16 13:57:25 -07001083 blkcg = bio_blkcg(bio);
Tejun Heocd1604f2012-03-05 13:15:06 -08001084 tg = throtl_lookup_tg(td, blkcg);
Vivek Goyalaf75cd32011-05-19 15:38:31 -04001085 if (tg) {
Vivek Goyalaf75cd32011-05-19 15:38:31 -04001086 if (tg_no_rule_group(tg, rw)) {
Tejun Heo629ed0b2012-04-01 14:38:44 -07001087 throtl_update_dispatch_stats(tg_to_blkg(tg),
1088 bio->bi_size, bio->bi_rw);
Tejun Heo2a7f1242012-03-05 13:15:01 -08001089 goto out_unlock_rcu;
Vivek Goyalaf75cd32011-05-19 15:38:31 -04001090 }
1091 }
Vivek Goyalaf75cd32011-05-19 15:38:31 -04001092
1093 /*
1094 * Either group has not been allocated yet or it is not an unlimited
1095 * IO group
1096 */
Vivek Goyale43473b2010-09-15 17:06:35 -04001097 spin_lock_irq(q->queue_lock);
Tejun Heocd1604f2012-03-05 13:15:06 -08001098 tg = throtl_lookup_create_tg(td, blkcg);
Tejun Heobc16a4f2011-10-19 14:33:01 +02001099 if (unlikely(!tg))
1100 goto out_unlock;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001101
Vivek Goyale43473b2010-09-15 17:06:35 -04001102 if (tg->nr_queued[rw]) {
1103 /*
1104 * There is already another bio queued in same dir. No
1105 * need to update dispatch time.
1106 */
Vivek Goyal231d7042011-03-07 21:05:14 +01001107 update_disptime = false;
Vivek Goyale43473b2010-09-15 17:06:35 -04001108 goto queue_bio;
Vivek Goyalde701c72011-03-07 21:09:32 +01001109
Vivek Goyale43473b2010-09-15 17:06:35 -04001110 }
1111
1112 /* Bio is with-in rate limit of group */
1113 if (tg_may_dispatch(td, tg, bio, NULL)) {
1114 throtl_charge_bio(tg, bio);
Vivek Goyal04521db2011-03-22 21:54:29 +01001115
1116 /*
1117 * We need to trim slice even when bios are not being queued
1118 * otherwise it might happen that a bio is not queued for
1119 * a long time and slice keeps on extending and trim is not
1120 * called for a long time. Now if limits are reduced suddenly
1121 * we take into account all the IO dispatched so far at new
1122 * low rate and * newly queued IO gets a really long dispatch
1123 * time.
1124 *
1125 * So keep on trimming slice even if bio is not queued.
1126 */
1127 throtl_trim_slice(td, tg, rw);
Tejun Heobc16a4f2011-10-19 14:33:01 +02001128 goto out_unlock;
Vivek Goyale43473b2010-09-15 17:06:35 -04001129 }
1130
1131queue_bio:
Joe Perchesfd16d262011-06-13 10:42:49 +02001132 throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
Vivek Goyal8e89d132010-09-15 17:06:37 -04001133 " iodisp=%u iops=%u queued=%d/%d",
1134 rw == READ ? 'R' : 'W',
Vivek Goyale43473b2010-09-15 17:06:35 -04001135 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
Vivek Goyal8e89d132010-09-15 17:06:37 -04001136 tg->io_disp[rw], tg->iops[rw],
Vivek Goyale43473b2010-09-15 17:06:35 -04001137 tg->nr_queued[READ], tg->nr_queued[WRITE]);
1138
Tejun Heo671058f2012-03-05 13:15:29 -08001139 bio_associate_current(bio);
Vivek Goyale43473b2010-09-15 17:06:35 -04001140 throtl_add_bio_tg(q->td, tg, bio);
Tejun Heobc16a4f2011-10-19 14:33:01 +02001141 throttled = true;
Vivek Goyale43473b2010-09-15 17:06:35 -04001142
1143 if (update_disptime) {
1144 tg_update_disptime(td, tg);
1145 throtl_schedule_next_dispatch(td);
1146 }
1147
Tejun Heobc16a4f2011-10-19 14:33:01 +02001148out_unlock:
Vivek Goyale43473b2010-09-15 17:06:35 -04001149 spin_unlock_irq(q->queue_lock);
Tejun Heo2a7f1242012-03-05 13:15:01 -08001150out_unlock_rcu:
1151 rcu_read_unlock();
Tejun Heobc16a4f2011-10-19 14:33:01 +02001152out:
1153 return throttled;
Vivek Goyale43473b2010-09-15 17:06:35 -04001154}
1155
Tejun Heoc9a929d2011-10-19 14:42:16 +02001156/**
1157 * blk_throtl_drain - drain throttled bios
1158 * @q: request_queue to drain throttled bios for
1159 *
1160 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1161 */
1162void blk_throtl_drain(struct request_queue *q)
1163 __releases(q->queue_lock) __acquires(q->queue_lock)
1164{
1165 struct throtl_data *td = q->td;
1166 struct throtl_rb_root *st = &td->tg_service_tree;
1167 struct throtl_grp *tg;
1168 struct bio_list bl;
1169 struct bio *bio;
1170
Andi Kleen8bcb6c72012-03-30 12:33:28 +02001171 queue_lockdep_assert_held(q);
Tejun Heoc9a929d2011-10-19 14:42:16 +02001172
1173 bio_list_init(&bl);
1174
1175 while ((tg = throtl_rb_first(st))) {
1176 throtl_dequeue_tg(td, tg);
1177
1178 while ((bio = bio_list_peek(&tg->bio_lists[READ])))
1179 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl);
1180 while ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
1181 tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl);
1182 }
1183 spin_unlock_irq(q->queue_lock);
1184
1185 while ((bio = bio_list_pop(&bl)))
1186 generic_make_request(bio);
1187
1188 spin_lock_irq(q->queue_lock);
1189}
1190
Vivek Goyale43473b2010-09-15 17:06:35 -04001191int blk_throtl_init(struct request_queue *q)
1192{
1193 struct throtl_data *td;
Tejun Heoa2b16932012-04-13 13:11:33 -07001194 int ret;
Vivek Goyale43473b2010-09-15 17:06:35 -04001195
1196 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1197 if (!td)
1198 return -ENOMEM;
1199
Vivek Goyale43473b2010-09-15 17:06:35 -04001200 td->tg_service_tree = THROTL_RB_ROOT;
Tejun Heocb761992013-05-14 13:52:31 -07001201 INIT_DELAYED_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
Vivek Goyale43473b2010-09-15 17:06:35 -04001202
Tejun Heocd1604f2012-03-05 13:15:06 -08001203 q->td = td;
Vivek Goyal29b12582011-05-19 15:38:24 -04001204 td->queue = q;
Vivek Goyal02977e42010-10-01 14:49:48 +02001205
Tejun Heoa2b16932012-04-13 13:11:33 -07001206 /* activate policy */
Tejun Heo3c798392012-04-16 13:57:25 -07001207 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
Tejun Heoa2b16932012-04-13 13:11:33 -07001208 if (ret)
Vivek Goyal29b12582011-05-19 15:38:24 -04001209 kfree(td);
Tejun Heoa2b16932012-04-13 13:11:33 -07001210 return ret;
Vivek Goyale43473b2010-09-15 17:06:35 -04001211}
1212
1213void blk_throtl_exit(struct request_queue *q)
1214{
Tejun Heoc875f4d2012-03-05 13:15:22 -08001215 BUG_ON(!q->td);
Vivek Goyalda527772011-03-02 19:05:33 -05001216 throtl_shutdown_wq(q);
Tejun Heo3c798392012-04-16 13:57:25 -07001217 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
Tejun Heoc9a929d2011-10-19 14:42:16 +02001218 kfree(q->td);
Vivek Goyale43473b2010-09-15 17:06:35 -04001219}
1220
1221static int __init throtl_init(void)
1222{
Vivek Goyal450adcb2011-03-01 13:40:54 -05001223 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1224 if (!kthrotld_workqueue)
1225 panic("Failed to create kthrotld\n");
1226
Tejun Heo3c798392012-04-16 13:57:25 -07001227 return blkcg_policy_register(&blkcg_policy_throtl);
Vivek Goyale43473b2010-09-15 17:06:35 -04001228}
1229
1230module_init(throtl_init);