blob: 08a32dfd3844cfeeacae4d38164c96f05f1b0c83 [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;
Vivek Goyal450adcb2011-03-01 13:40:54 -050028
Tejun Heoc5cc2072013-05-14 13:52:38 -070029/*
30 * To implement hierarchical throttling, throtl_grps form a tree and bios
31 * are dispatched upwards level by level until they reach the top and get
32 * issued. When dispatching bios from the children and local group at each
33 * level, if the bios are dispatched into a single bio_list, there's a risk
34 * of a local or child group which can queue many bios at once filling up
35 * the list starving others.
36 *
37 * To avoid such starvation, dispatched bios are queued separately
38 * according to where they came from. When they are again dispatched to
39 * the parent, they're popped in round-robin order so that no single source
40 * hogs the dispatch window.
41 *
42 * throtl_qnode is used to keep the queued bios separated by their sources.
43 * Bios are queued to throtl_qnode which in turn is queued to
44 * throtl_service_queue and then dispatched in round-robin order.
45 *
46 * It's also used to track the reference counts on blkg's. A qnode always
47 * belongs to a throtl_grp and gets queued on itself or the parent, so
48 * incrementing the reference of the associated throtl_grp when a qnode is
49 * queued and decrementing when dequeued is enough to keep the whole blkg
50 * tree pinned while bios are in flight.
51 */
52struct throtl_qnode {
53 struct list_head node; /* service_queue->queued[] */
54 struct bio_list bios; /* queued bios */
55 struct throtl_grp *tg; /* tg this qnode belongs to */
56};
57
Tejun Heoc9e03322013-05-14 13:52:32 -070058struct throtl_service_queue {
Tejun Heo77216b02013-05-14 13:52:36 -070059 struct throtl_service_queue *parent_sq; /* the parent service_queue */
60
Tejun Heo73f0d492013-05-14 13:52:35 -070061 /*
62 * Bios queued directly to this service_queue or dispatched from
63 * children throtl_grp's.
64 */
Tejun Heoc5cc2072013-05-14 13:52:38 -070065 struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */
Tejun Heo73f0d492013-05-14 13:52:35 -070066 unsigned int nr_queued[2]; /* number of queued bios */
67
68 /*
69 * RB tree of active children throtl_grp's, which are sorted by
70 * their ->disptime.
71 */
Tejun Heoc9e03322013-05-14 13:52:32 -070072 struct rb_root pending_tree; /* RB tree of active tgs */
73 struct rb_node *first_pending; /* first node in the tree */
74 unsigned int nr_pending; /* # queued in the tree */
75 unsigned long first_pending_disptime; /* disptime of the first tg */
Tejun Heo69df0ab2013-05-14 13:52:36 -070076 struct timer_list pending_timer; /* fires on first_pending_disptime */
Vivek Goyale43473b2010-09-15 17:06:35 -040077};
78
Tejun Heo5b2c16a2013-05-14 13:52:32 -070079enum tg_state_flags {
80 THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */
Tejun Heo0e9f4162013-05-14 13:52:35 -070081 THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */
Tejun Heo5b2c16a2013-05-14 13:52:32 -070082};
83
Vivek Goyale43473b2010-09-15 17:06:35 -040084#define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node)
85
Tejun Heo8a3d2612012-04-01 14:38:44 -070086/* Per-cpu group stats */
87struct tg_stats_cpu {
88 /* total bytes transferred */
89 struct blkg_rwstat service_bytes;
90 /* total IOs serviced, post merge */
91 struct blkg_rwstat serviced;
92};
93
Vivek Goyale43473b2010-09-15 17:06:35 -040094struct throtl_grp {
Tejun Heof95a04a2012-04-16 13:57:26 -070095 /* must be the first member */
96 struct blkg_policy_data pd;
97
Tejun Heoc9e03322013-05-14 13:52:32 -070098 /* active throtl group service_queue member */
Vivek Goyale43473b2010-09-15 17:06:35 -040099 struct rb_node rb_node;
100
Tejun Heo0f3457f2013-05-14 13:52:32 -0700101 /* throtl_data this group belongs to */
102 struct throtl_data *td;
103
Tejun Heo49a2f1e2013-05-14 13:52:34 -0700104 /* this group's service queue */
105 struct throtl_service_queue service_queue;
106
Vivek Goyale43473b2010-09-15 17:06:35 -0400107 /*
Tejun Heoc5cc2072013-05-14 13:52:38 -0700108 * qnode_on_self is used when bios are directly queued to this
109 * throtl_grp so that local bios compete fairly with bios
110 * dispatched from children. qnode_on_parent is used when bios are
111 * dispatched from this throtl_grp into its parent and will compete
112 * with the sibling qnode_on_parents and the parent's
113 * qnode_on_self.
114 */
115 struct throtl_qnode qnode_on_self[2];
116 struct throtl_qnode qnode_on_parent[2];
117
118 /*
Vivek Goyale43473b2010-09-15 17:06:35 -0400119 * Dispatch time in jiffies. This is the estimated time when group
120 * will unthrottle and is ready to dispatch more bio. It is used as
121 * key to sort active groups in service tree.
122 */
123 unsigned long disptime;
124
Vivek Goyale43473b2010-09-15 17:06:35 -0400125 unsigned int flags;
126
Tejun Heo693e7512013-05-14 13:52:38 -0700127 /* are there any throtl rules between this group and td? */
128 bool has_rules[2];
129
Vivek Goyale43473b2010-09-15 17:06:35 -0400130 /* bytes per second rate limits */
131 uint64_t bps[2];
132
Vivek Goyal8e89d132010-09-15 17:06:37 -0400133 /* IOPS limits */
134 unsigned int iops[2];
135
Vivek Goyale43473b2010-09-15 17:06:35 -0400136 /* Number of bytes disptached in current slice */
137 uint64_t bytes_disp[2];
Vivek Goyal8e89d132010-09-15 17:06:37 -0400138 /* Number of bio's dispatched in current slice */
139 unsigned int io_disp[2];
Vivek Goyale43473b2010-09-15 17:06:35 -0400140
141 /* When did we start a new slice */
142 unsigned long slice_start[2];
143 unsigned long slice_end[2];
Vivek Goyalfe071432010-10-01 14:49:49 +0200144
Tejun Heo8a3d2612012-04-01 14:38:44 -0700145 /* Per cpu stats pointer */
146 struct tg_stats_cpu __percpu *stats_cpu;
147
148 /* List of tgs waiting for per cpu stats memory to be allocated */
149 struct list_head stats_alloc_node;
Vivek Goyale43473b2010-09-15 17:06:35 -0400150};
151
152struct throtl_data
153{
Vivek Goyale43473b2010-09-15 17:06:35 -0400154 /* service tree for active throtl groups */
Tejun Heoc9e03322013-05-14 13:52:32 -0700155 struct throtl_service_queue service_queue;
Vivek Goyale43473b2010-09-15 17:06:35 -0400156
Vivek Goyale43473b2010-09-15 17:06:35 -0400157 struct request_queue *queue;
158
159 /* Total Number of queued bios on READ and WRITE lists */
160 unsigned int nr_queued[2];
161
162 /*
Vivek Goyal02977e42010-10-01 14:49:48 +0200163 * number of total undestroyed groups
Vivek Goyale43473b2010-09-15 17:06:35 -0400164 */
165 unsigned int nr_undestroyed_grps;
166
167 /* Work for dispatching throttled bios */
Tejun Heo69df0ab2013-05-14 13:52:36 -0700168 struct work_struct dispatch_work;
Vivek Goyale43473b2010-09-15 17:06:35 -0400169};
170
Tejun Heo8a3d2612012-04-01 14:38:44 -0700171/* list and work item to allocate percpu group stats */
172static DEFINE_SPINLOCK(tg_stats_alloc_lock);
173static LIST_HEAD(tg_stats_alloc_list);
174
175static void tg_stats_alloc_fn(struct work_struct *);
176static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn);
177
Tejun Heo69df0ab2013-05-14 13:52:36 -0700178static void throtl_pending_timer_fn(unsigned long arg);
179
Tejun Heof95a04a2012-04-16 13:57:26 -0700180static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
181{
182 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
183}
184
Tejun Heo3c798392012-04-16 13:57:25 -0700185static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
Tejun Heo03814112012-03-05 13:15:14 -0800186{
Tejun Heof95a04a2012-04-16 13:57:26 -0700187 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
Tejun Heo03814112012-03-05 13:15:14 -0800188}
189
Tejun Heo3c798392012-04-16 13:57:25 -0700190static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
Tejun Heo03814112012-03-05 13:15:14 -0800191{
Tejun Heof95a04a2012-04-16 13:57:26 -0700192 return pd_to_blkg(&tg->pd);
Tejun Heo03814112012-03-05 13:15:14 -0800193}
194
Tejun Heo03d8e112012-04-13 13:11:32 -0700195static inline struct throtl_grp *td_root_tg(struct throtl_data *td)
196{
197 return blkg_to_tg(td->queue->root_blkg);
198}
199
Tejun Heofda6f272013-05-14 13:52:36 -0700200/**
201 * sq_to_tg - return the throl_grp the specified service queue belongs to
202 * @sq: the throtl_service_queue of interest
203 *
204 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
205 * embedded in throtl_data, %NULL is returned.
206 */
207static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
208{
209 if (sq && sq->parent_sq)
210 return container_of(sq, struct throtl_grp, service_queue);
211 else
212 return NULL;
213}
Vivek Goyale43473b2010-09-15 17:06:35 -0400214
Tejun Heofda6f272013-05-14 13:52:36 -0700215/**
216 * sq_to_td - return throtl_data the specified service queue belongs to
217 * @sq: the throtl_service_queue of interest
218 *
219 * A service_queue can be embeded in either a throtl_grp or throtl_data.
220 * Determine the associated throtl_data accordingly and return it.
221 */
222static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
223{
224 struct throtl_grp *tg = sq_to_tg(sq);
225
226 if (tg)
227 return tg->td;
228 else
229 return container_of(sq, struct throtl_data, service_queue);
230}
231
232/**
233 * throtl_log - log debug message via blktrace
234 * @sq: the service_queue being reported
235 * @fmt: printf format string
236 * @args: printf args
237 *
238 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
239 * throtl_grp; otherwise, just "throtl".
240 *
241 * TODO: this should be made a function and name formatting should happen
242 * after testing whether blktrace is enabled.
243 */
244#define throtl_log(sq, fmt, args...) do { \
245 struct throtl_grp *__tg = sq_to_tg((sq)); \
246 struct throtl_data *__td = sq_to_td((sq)); \
247 \
248 (void)__td; \
249 if ((__tg)) { \
250 char __pbuf[128]; \
251 \
252 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
253 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
254 } else { \
255 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
256 } \
257} while (0)
Vivek Goyale43473b2010-09-15 17:06:35 -0400258
Tejun Heo8a3d2612012-04-01 14:38:44 -0700259/*
260 * Worker for allocating per cpu stat for tgs. This is scheduled on the
Tejun Heo3b07e9c2012-08-20 14:51:24 -0700261 * system_wq once there are some groups on the alloc_list waiting for
Tejun Heo8a3d2612012-04-01 14:38:44 -0700262 * allocation.
263 */
264static void tg_stats_alloc_fn(struct work_struct *work)
265{
266 static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */
267 struct delayed_work *dwork = to_delayed_work(work);
268 bool empty = false;
269
270alloc_stats:
271 if (!stats_cpu) {
272 stats_cpu = alloc_percpu(struct tg_stats_cpu);
273 if (!stats_cpu) {
274 /* allocation failed, try again after some time */
Tejun Heo3b07e9c2012-08-20 14:51:24 -0700275 schedule_delayed_work(dwork, msecs_to_jiffies(10));
Tejun Heo8a3d2612012-04-01 14:38:44 -0700276 return;
277 }
278 }
279
280 spin_lock_irq(&tg_stats_alloc_lock);
281
282 if (!list_empty(&tg_stats_alloc_list)) {
283 struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list,
284 struct throtl_grp,
285 stats_alloc_node);
286 swap(tg->stats_cpu, stats_cpu);
287 list_del_init(&tg->stats_alloc_node);
288 }
289
290 empty = list_empty(&tg_stats_alloc_list);
291 spin_unlock_irq(&tg_stats_alloc_lock);
292 if (!empty)
293 goto alloc_stats;
294}
295
Tejun Heoc5cc2072013-05-14 13:52:38 -0700296static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
297{
298 INIT_LIST_HEAD(&qn->node);
299 bio_list_init(&qn->bios);
300 qn->tg = tg;
301}
302
303/**
304 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
305 * @bio: bio being added
306 * @qn: qnode to add bio to
307 * @queued: the service_queue->queued[] list @qn belongs to
308 *
309 * Add @bio to @qn and put @qn on @queued if it's not already on.
310 * @qn->tg's reference count is bumped when @qn is activated. See the
311 * comment on top of throtl_qnode definition for details.
312 */
313static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
314 struct list_head *queued)
315{
316 bio_list_add(&qn->bios, bio);
317 if (list_empty(&qn->node)) {
318 list_add_tail(&qn->node, queued);
319 blkg_get(tg_to_blkg(qn->tg));
320 }
321}
322
323/**
324 * throtl_peek_queued - peek the first bio on a qnode list
325 * @queued: the qnode list to peek
326 */
327static struct bio *throtl_peek_queued(struct list_head *queued)
328{
329 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
330 struct bio *bio;
331
332 if (list_empty(queued))
333 return NULL;
334
335 bio = bio_list_peek(&qn->bios);
336 WARN_ON_ONCE(!bio);
337 return bio;
338}
339
340/**
341 * throtl_pop_queued - pop the first bio form a qnode list
342 * @queued: the qnode list to pop a bio from
343 * @tg_to_put: optional out argument for throtl_grp to put
344 *
345 * Pop the first bio from the qnode list @queued. After popping, the first
346 * qnode is removed from @queued if empty or moved to the end of @queued so
347 * that the popping order is round-robin.
348 *
349 * When the first qnode is removed, its associated throtl_grp should be put
350 * too. If @tg_to_put is NULL, this function automatically puts it;
351 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
352 * responsible for putting it.
353 */
354static struct bio *throtl_pop_queued(struct list_head *queued,
355 struct throtl_grp **tg_to_put)
356{
357 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
358 struct bio *bio;
359
360 if (list_empty(queued))
361 return NULL;
362
363 bio = bio_list_pop(&qn->bios);
364 WARN_ON_ONCE(!bio);
365
366 if (bio_list_empty(&qn->bios)) {
367 list_del_init(&qn->node);
368 if (tg_to_put)
369 *tg_to_put = qn->tg;
370 else
371 blkg_put(tg_to_blkg(qn->tg));
372 } else {
373 list_move_tail(&qn->node, queued);
374 }
375
376 return bio;
377}
378
Tejun Heo49a2f1e2013-05-14 13:52:34 -0700379/* init a service_queue, assumes the caller zeroed it */
Tejun Heo77216b02013-05-14 13:52:36 -0700380static void throtl_service_queue_init(struct throtl_service_queue *sq,
381 struct throtl_service_queue *parent_sq)
Tejun Heo49a2f1e2013-05-14 13:52:34 -0700382{
Tejun Heoc5cc2072013-05-14 13:52:38 -0700383 INIT_LIST_HEAD(&sq->queued[0]);
384 INIT_LIST_HEAD(&sq->queued[1]);
Tejun Heo49a2f1e2013-05-14 13:52:34 -0700385 sq->pending_tree = RB_ROOT;
Tejun Heo77216b02013-05-14 13:52:36 -0700386 sq->parent_sq = parent_sq;
Tejun Heo69df0ab2013-05-14 13:52:36 -0700387 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
388 (unsigned long)sq);
389}
390
391static void throtl_service_queue_exit(struct throtl_service_queue *sq)
392{
393 del_timer_sync(&sq->pending_timer);
Tejun Heo49a2f1e2013-05-14 13:52:34 -0700394}
395
Tejun Heo3c798392012-04-16 13:57:25 -0700396static void throtl_pd_init(struct blkcg_gq *blkg)
Vivek Goyala29a1712011-05-19 15:38:19 -0400397{
Tejun Heo03814112012-03-05 13:15:14 -0800398 struct throtl_grp *tg = blkg_to_tg(blkg);
Tejun Heo77216b02013-05-14 13:52:36 -0700399 struct throtl_data *td = blkg->q->td;
Tejun Heo91381252013-05-14 13:52:38 -0700400 struct throtl_service_queue *parent_sq;
Tejun Heoff26eaa2012-05-23 12:16:21 +0200401 unsigned long flags;
Tejun Heoc5cc2072013-05-14 13:52:38 -0700402 int rw;
Tejun Heocd1604f2012-03-05 13:15:06 -0800403
Tejun Heo91381252013-05-14 13:52:38 -0700404 /*
405 * If sane_hierarchy is enabled, we switch to properly hierarchical
406 * behavior where limits on a given throtl_grp are applied to the
407 * whole subtree rather than just the group itself. e.g. If 16M
408 * read_bps limit is set on the root group, the whole system can't
409 * exceed 16M for the device.
410 *
411 * If sane_hierarchy is not enabled, the broken flat hierarchy
412 * behavior is retained where all throtl_grps are treated as if
413 * they're all separate root groups right below throtl_data.
414 * Limits of a group don't interact with limits of other groups
415 * regardless of the position of the group in the hierarchy.
416 */
417 parent_sq = &td->service_queue;
418
419 if (cgroup_sane_behavior(blkg->blkcg->css.cgroup) && blkg->parent)
420 parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
421
422 throtl_service_queue_init(&tg->service_queue, parent_sq);
423
Tejun Heoc5cc2072013-05-14 13:52:38 -0700424 for (rw = READ; rw <= WRITE; rw++) {
425 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
426 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
427 }
428
Vivek Goyala29a1712011-05-19 15:38:19 -0400429 RB_CLEAR_NODE(&tg->rb_node);
Tejun Heo77216b02013-05-14 13:52:36 -0700430 tg->td = td;
Vivek Goyala29a1712011-05-19 15:38:19 -0400431
Tejun Heoe56da7e2012-03-05 13:15:07 -0800432 tg->bps[READ] = -1;
433 tg->bps[WRITE] = -1;
434 tg->iops[READ] = -1;
435 tg->iops[WRITE] = -1;
Tejun Heo8a3d2612012-04-01 14:38:44 -0700436
437 /*
438 * Ugh... We need to perform per-cpu allocation for tg->stats_cpu
439 * but percpu allocator can't be called from IO path. Queue tg on
440 * tg_stats_alloc_list and allocate from work item.
441 */
Tejun Heoff26eaa2012-05-23 12:16:21 +0200442 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
Tejun Heo8a3d2612012-04-01 14:38:44 -0700443 list_add(&tg->stats_alloc_node, &tg_stats_alloc_list);
Tejun Heo3b07e9c2012-08-20 14:51:24 -0700444 schedule_delayed_work(&tg_stats_alloc_work, 0);
Tejun Heoff26eaa2012-05-23 12:16:21 +0200445 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
Tejun Heo8a3d2612012-04-01 14:38:44 -0700446}
447
Tejun Heo693e7512013-05-14 13:52:38 -0700448/*
449 * Set has_rules[] if @tg or any of its parents have limits configured.
450 * This doesn't require walking up to the top of the hierarchy as the
451 * parent's has_rules[] is guaranteed to be correct.
452 */
453static void tg_update_has_rules(struct throtl_grp *tg)
454{
455 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
456 int rw;
457
458 for (rw = READ; rw <= WRITE; rw++)
459 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
460 (tg->bps[rw] != -1 || tg->iops[rw] != -1);
461}
462
463static void throtl_pd_online(struct blkcg_gq *blkg)
464{
465 /*
466 * We don't want new groups to escape the limits of its ancestors.
467 * Update has_rules[] after a new group is brought online.
468 */
469 tg_update_has_rules(blkg_to_tg(blkg));
470}
471
Tejun Heo3c798392012-04-16 13:57:25 -0700472static void throtl_pd_exit(struct blkcg_gq *blkg)
Tejun Heo8a3d2612012-04-01 14:38:44 -0700473{
474 struct throtl_grp *tg = blkg_to_tg(blkg);
Tejun Heoff26eaa2012-05-23 12:16:21 +0200475 unsigned long flags;
Tejun Heo8a3d2612012-04-01 14:38:44 -0700476
Tejun Heoff26eaa2012-05-23 12:16:21 +0200477 spin_lock_irqsave(&tg_stats_alloc_lock, flags);
Tejun Heo8a3d2612012-04-01 14:38:44 -0700478 list_del_init(&tg->stats_alloc_node);
Tejun Heoff26eaa2012-05-23 12:16:21 +0200479 spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
Tejun Heo8a3d2612012-04-01 14:38:44 -0700480
481 free_percpu(tg->stats_cpu);
Tejun Heo69df0ab2013-05-14 13:52:36 -0700482
483 throtl_service_queue_exit(&tg->service_queue);
Tejun Heo8a3d2612012-04-01 14:38:44 -0700484}
485
Tejun Heo3c798392012-04-16 13:57:25 -0700486static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
Tejun Heo8a3d2612012-04-01 14:38:44 -0700487{
488 struct throtl_grp *tg = blkg_to_tg(blkg);
489 int cpu;
490
491 if (tg->stats_cpu == NULL)
492 return;
493
494 for_each_possible_cpu(cpu) {
495 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
496
497 blkg_rwstat_reset(&sc->service_bytes);
498 blkg_rwstat_reset(&sc->serviced);
499 }
Vivek Goyala29a1712011-05-19 15:38:19 -0400500}
501
Tejun Heo3c798392012-04-16 13:57:25 -0700502static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
503 struct blkcg *blkcg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400504{
Vivek Goyale43473b2010-09-15 17:06:35 -0400505 /*
Tejun Heo3c798392012-04-16 13:57:25 -0700506 * This is the common case when there are no blkcgs. Avoid lookup
507 * in this case
Tejun Heocd1604f2012-03-05 13:15:06 -0800508 */
Tejun Heo3c798392012-04-16 13:57:25 -0700509 if (blkcg == &blkcg_root)
Tejun Heo03d8e112012-04-13 13:11:32 -0700510 return td_root_tg(td);
Vivek Goyale43473b2010-09-15 17:06:35 -0400511
Tejun Heoe8989fa2012-03-05 13:15:20 -0800512 return blkg_to_tg(blkg_lookup(blkcg, td->queue));
Vivek Goyale43473b2010-09-15 17:06:35 -0400513}
514
Tejun Heocd1604f2012-03-05 13:15:06 -0800515static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
Tejun Heo3c798392012-04-16 13:57:25 -0700516 struct blkcg *blkcg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400517{
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400518 struct request_queue *q = td->queue;
Tejun Heocd1604f2012-03-05 13:15:06 -0800519 struct throtl_grp *tg = NULL;
Tejun Heo0a5a7d02012-03-05 13:15:02 -0800520
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400521 /*
Tejun Heo3c798392012-04-16 13:57:25 -0700522 * This is the common case when there are no blkcgs. Avoid lookup
523 * in this case
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400524 */
Tejun Heo3c798392012-04-16 13:57:25 -0700525 if (blkcg == &blkcg_root) {
Tejun Heo03d8e112012-04-13 13:11:32 -0700526 tg = td_root_tg(td);
Tejun Heocd1604f2012-03-05 13:15:06 -0800527 } else {
Tejun Heo3c798392012-04-16 13:57:25 -0700528 struct blkcg_gq *blkg;
Tejun Heocd1604f2012-03-05 13:15:06 -0800529
Tejun Heo3c96cb32012-04-13 13:11:34 -0700530 blkg = blkg_lookup_create(blkcg, q);
Tejun Heocd1604f2012-03-05 13:15:06 -0800531
532 /* if %NULL and @q is alive, fall back to root_tg */
533 if (!IS_ERR(blkg))
Tejun Heo03814112012-03-05 13:15:14 -0800534 tg = blkg_to_tg(blkg);
Bart Van Assche3f3299d2012-11-28 13:42:38 +0100535 else if (!blk_queue_dying(q))
Tejun Heo03d8e112012-04-13 13:11:32 -0700536 tg = td_root_tg(td);
Vivek Goyalf469a7b2011-05-19 15:38:23 -0400537 }
538
Vivek Goyale43473b2010-09-15 17:06:35 -0400539 return tg;
540}
541
Tejun Heo0049af72013-05-14 13:52:33 -0700542static struct throtl_grp *
543throtl_rb_first(struct throtl_service_queue *parent_sq)
Vivek Goyale43473b2010-09-15 17:06:35 -0400544{
545 /* Service tree is empty */
Tejun Heo0049af72013-05-14 13:52:33 -0700546 if (!parent_sq->nr_pending)
Vivek Goyale43473b2010-09-15 17:06:35 -0400547 return NULL;
548
Tejun Heo0049af72013-05-14 13:52:33 -0700549 if (!parent_sq->first_pending)
550 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
Vivek Goyale43473b2010-09-15 17:06:35 -0400551
Tejun Heo0049af72013-05-14 13:52:33 -0700552 if (parent_sq->first_pending)
553 return rb_entry_tg(parent_sq->first_pending);
Vivek Goyale43473b2010-09-15 17:06:35 -0400554
555 return NULL;
556}
557
558static void rb_erase_init(struct rb_node *n, struct rb_root *root)
559{
560 rb_erase(n, root);
561 RB_CLEAR_NODE(n);
562}
563
Tejun Heo0049af72013-05-14 13:52:33 -0700564static void throtl_rb_erase(struct rb_node *n,
565 struct throtl_service_queue *parent_sq)
Vivek Goyale43473b2010-09-15 17:06:35 -0400566{
Tejun Heo0049af72013-05-14 13:52:33 -0700567 if (parent_sq->first_pending == n)
568 parent_sq->first_pending = NULL;
569 rb_erase_init(n, &parent_sq->pending_tree);
570 --parent_sq->nr_pending;
Vivek Goyale43473b2010-09-15 17:06:35 -0400571}
572
Tejun Heo0049af72013-05-14 13:52:33 -0700573static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
Vivek Goyale43473b2010-09-15 17:06:35 -0400574{
575 struct throtl_grp *tg;
576
Tejun Heo0049af72013-05-14 13:52:33 -0700577 tg = throtl_rb_first(parent_sq);
Vivek Goyale43473b2010-09-15 17:06:35 -0400578 if (!tg)
579 return;
580
Tejun Heo0049af72013-05-14 13:52:33 -0700581 parent_sq->first_pending_disptime = tg->disptime;
Vivek Goyale43473b2010-09-15 17:06:35 -0400582}
583
Tejun Heo77216b02013-05-14 13:52:36 -0700584static void tg_service_queue_add(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400585{
Tejun Heo77216b02013-05-14 13:52:36 -0700586 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
Tejun Heo0049af72013-05-14 13:52:33 -0700587 struct rb_node **node = &parent_sq->pending_tree.rb_node;
Vivek Goyale43473b2010-09-15 17:06:35 -0400588 struct rb_node *parent = NULL;
589 struct throtl_grp *__tg;
590 unsigned long key = tg->disptime;
591 int left = 1;
592
593 while (*node != NULL) {
594 parent = *node;
595 __tg = rb_entry_tg(parent);
596
597 if (time_before(key, __tg->disptime))
598 node = &parent->rb_left;
599 else {
600 node = &parent->rb_right;
601 left = 0;
602 }
603 }
604
605 if (left)
Tejun Heo0049af72013-05-14 13:52:33 -0700606 parent_sq->first_pending = &tg->rb_node;
Vivek Goyale43473b2010-09-15 17:06:35 -0400607
608 rb_link_node(&tg->rb_node, parent, node);
Tejun Heo0049af72013-05-14 13:52:33 -0700609 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
Vivek Goyale43473b2010-09-15 17:06:35 -0400610}
611
Tejun Heo77216b02013-05-14 13:52:36 -0700612static void __throtl_enqueue_tg(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400613{
Tejun Heo77216b02013-05-14 13:52:36 -0700614 tg_service_queue_add(tg);
Tejun Heo5b2c16a2013-05-14 13:52:32 -0700615 tg->flags |= THROTL_TG_PENDING;
Tejun Heo77216b02013-05-14 13:52:36 -0700616 tg->service_queue.parent_sq->nr_pending++;
Vivek Goyale43473b2010-09-15 17:06:35 -0400617}
618
Tejun Heo77216b02013-05-14 13:52:36 -0700619static void throtl_enqueue_tg(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400620{
Tejun Heo5b2c16a2013-05-14 13:52:32 -0700621 if (!(tg->flags & THROTL_TG_PENDING))
Tejun Heo77216b02013-05-14 13:52:36 -0700622 __throtl_enqueue_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -0400623}
624
Tejun Heo77216b02013-05-14 13:52:36 -0700625static void __throtl_dequeue_tg(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400626{
Tejun Heo77216b02013-05-14 13:52:36 -0700627 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
Tejun Heo5b2c16a2013-05-14 13:52:32 -0700628 tg->flags &= ~THROTL_TG_PENDING;
Vivek Goyale43473b2010-09-15 17:06:35 -0400629}
630
Tejun Heo77216b02013-05-14 13:52:36 -0700631static void throtl_dequeue_tg(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400632{
Tejun Heo5b2c16a2013-05-14 13:52:32 -0700633 if (tg->flags & THROTL_TG_PENDING)
Tejun Heo77216b02013-05-14 13:52:36 -0700634 __throtl_dequeue_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -0400635}
636
Tejun Heoa9131a22013-05-14 13:52:31 -0700637/* Call with queue lock held */
Tejun Heo69df0ab2013-05-14 13:52:36 -0700638static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
639 unsigned long expires)
Tejun Heoa9131a22013-05-14 13:52:31 -0700640{
Tejun Heo69df0ab2013-05-14 13:52:36 -0700641 mod_timer(&sq->pending_timer, expires);
642 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
643 expires - jiffies, jiffies);
Tejun Heoa9131a22013-05-14 13:52:31 -0700644}
645
Tejun Heo7f52f982013-05-14 13:52:37 -0700646/**
647 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
648 * @sq: the service_queue to schedule dispatch for
649 * @force: force scheduling
650 *
651 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
652 * dispatch time of the first pending child. Returns %true if either timer
653 * is armed or there's no pending child left. %false if the current
654 * dispatch window is still open and the caller should continue
655 * dispatching.
656 *
657 * If @force is %true, the dispatch timer is always scheduled and this
658 * function is guaranteed to return %true. This is to be used when the
659 * caller can't dispatch itself and needs to invoke pending_timer
660 * unconditionally. Note that forced scheduling is likely to induce short
661 * delay before dispatch starts even if @sq->first_pending_disptime is not
662 * in the future and thus shouldn't be used in hot paths.
663 */
664static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
665 bool force)
Vivek Goyale43473b2010-09-15 17:06:35 -0400666{
Tejun Heo6a525602013-05-14 13:52:32 -0700667 /* any pending children left? */
Tejun Heoc9e03322013-05-14 13:52:32 -0700668 if (!sq->nr_pending)
Tejun Heo7f52f982013-05-14 13:52:37 -0700669 return true;
Vivek Goyale43473b2010-09-15 17:06:35 -0400670
Tejun Heoc9e03322013-05-14 13:52:32 -0700671 update_min_dispatch_time(sq);
Vivek Goyale43473b2010-09-15 17:06:35 -0400672
Tejun Heo69df0ab2013-05-14 13:52:36 -0700673 /* is the next dispatch time in the future? */
Tejun Heo7f52f982013-05-14 13:52:37 -0700674 if (force || time_after(sq->first_pending_disptime, jiffies)) {
Tejun Heo69df0ab2013-05-14 13:52:36 -0700675 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
Tejun Heo7f52f982013-05-14 13:52:37 -0700676 return true;
Tejun Heo69df0ab2013-05-14 13:52:36 -0700677 }
678
Tejun Heo7f52f982013-05-14 13:52:37 -0700679 /* tell the caller to continue dispatching */
680 return false;
Vivek Goyale43473b2010-09-15 17:06:35 -0400681}
682
Vivek Goyal32ee5bc2013-05-14 13:52:38 -0700683static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
684 bool rw, unsigned long start)
685{
686 tg->bytes_disp[rw] = 0;
687 tg->io_disp[rw] = 0;
688
689 /*
690 * Previous slice has expired. We must have trimmed it after last
691 * bio dispatch. That means since start of last slice, we never used
692 * that bandwidth. Do try to make use of that bandwidth while giving
693 * credit.
694 */
695 if (time_after_eq(start, tg->slice_start[rw]))
696 tg->slice_start[rw] = start;
697
698 tg->slice_end[rw] = jiffies + throtl_slice;
699 throtl_log(&tg->service_queue,
700 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
701 rw == READ ? 'R' : 'W', tg->slice_start[rw],
702 tg->slice_end[rw], jiffies);
703}
704
Tejun Heo0f3457f2013-05-14 13:52:32 -0700705static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
Vivek Goyale43473b2010-09-15 17:06:35 -0400706{
707 tg->bytes_disp[rw] = 0;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400708 tg->io_disp[rw] = 0;
Vivek Goyale43473b2010-09-15 17:06:35 -0400709 tg->slice_start[rw] = jiffies;
710 tg->slice_end[rw] = jiffies + throtl_slice;
Tejun Heofda6f272013-05-14 13:52:36 -0700711 throtl_log(&tg->service_queue,
712 "[%c] new slice start=%lu end=%lu jiffies=%lu",
713 rw == READ ? 'R' : 'W', tg->slice_start[rw],
714 tg->slice_end[rw], jiffies);
Vivek Goyale43473b2010-09-15 17:06:35 -0400715}
716
Tejun Heo0f3457f2013-05-14 13:52:32 -0700717static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
718 unsigned long jiffy_end)
Vivek Goyald1ae8ff2010-12-01 19:34:46 +0100719{
720 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
721}
722
Tejun Heo0f3457f2013-05-14 13:52:32 -0700723static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
724 unsigned long jiffy_end)
Vivek Goyale43473b2010-09-15 17:06:35 -0400725{
726 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
Tejun Heofda6f272013-05-14 13:52:36 -0700727 throtl_log(&tg->service_queue,
728 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
729 rw == READ ? 'R' : 'W', tg->slice_start[rw],
730 tg->slice_end[rw], jiffies);
Vivek Goyale43473b2010-09-15 17:06:35 -0400731}
732
733/* Determine if previously allocated or extended slice is complete or not */
Tejun Heo0f3457f2013-05-14 13:52:32 -0700734static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
Vivek Goyale43473b2010-09-15 17:06:35 -0400735{
736 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
737 return 0;
738
739 return 1;
740}
741
742/* Trim the used slices and adjust slice start accordingly */
Tejun Heo0f3457f2013-05-14 13:52:32 -0700743static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
Vivek Goyale43473b2010-09-15 17:06:35 -0400744{
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200745 unsigned long nr_slices, time_elapsed, io_trim;
746 u64 bytes_trim, tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400747
748 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
749
750 /*
751 * If bps are unlimited (-1), then time slice don't get
752 * renewed. Don't try to trim the slice if slice is used. A new
753 * slice will start when appropriate.
754 */
Tejun Heo0f3457f2013-05-14 13:52:32 -0700755 if (throtl_slice_used(tg, rw))
Vivek Goyale43473b2010-09-15 17:06:35 -0400756 return;
757
Vivek Goyald1ae8ff2010-12-01 19:34:46 +0100758 /*
759 * A bio has been dispatched. Also adjust slice_end. It might happen
760 * that initially cgroup limit was very low resulting in high
761 * slice_end, but later limit was bumped up and bio was dispached
762 * sooner, then we need to reduce slice_end. A high bogus slice_end
763 * is bad because it does not allow new slice to start.
764 */
765
Tejun Heo0f3457f2013-05-14 13:52:32 -0700766 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
Vivek Goyald1ae8ff2010-12-01 19:34:46 +0100767
Vivek Goyale43473b2010-09-15 17:06:35 -0400768 time_elapsed = jiffies - tg->slice_start[rw];
769
770 nr_slices = time_elapsed / throtl_slice;
771
772 if (!nr_slices)
773 return;
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200774 tmp = tg->bps[rw] * throtl_slice * nr_slices;
775 do_div(tmp, HZ);
776 bytes_trim = tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400777
Vivek Goyal8e89d132010-09-15 17:06:37 -0400778 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
Vivek Goyale43473b2010-09-15 17:06:35 -0400779
Vivek Goyal8e89d132010-09-15 17:06:37 -0400780 if (!bytes_trim && !io_trim)
Vivek Goyale43473b2010-09-15 17:06:35 -0400781 return;
782
783 if (tg->bytes_disp[rw] >= bytes_trim)
784 tg->bytes_disp[rw] -= bytes_trim;
785 else
786 tg->bytes_disp[rw] = 0;
787
Vivek Goyal8e89d132010-09-15 17:06:37 -0400788 if (tg->io_disp[rw] >= io_trim)
789 tg->io_disp[rw] -= io_trim;
790 else
791 tg->io_disp[rw] = 0;
792
Vivek Goyale43473b2010-09-15 17:06:35 -0400793 tg->slice_start[rw] += nr_slices * throtl_slice;
794
Tejun Heofda6f272013-05-14 13:52:36 -0700795 throtl_log(&tg->service_queue,
796 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
797 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
798 tg->slice_start[rw], tg->slice_end[rw], jiffies);
Vivek Goyale43473b2010-09-15 17:06:35 -0400799}
800
Tejun Heo0f3457f2013-05-14 13:52:32 -0700801static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
802 unsigned long *wait)
Vivek Goyale43473b2010-09-15 17:06:35 -0400803{
804 bool rw = bio_data_dir(bio);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400805 unsigned int io_allowed;
Vivek Goyale43473b2010-09-15 17:06:35 -0400806 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
Vivek Goyalc49c06e2010-10-01 21:16:42 +0200807 u64 tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400808
Vivek Goyal8e89d132010-09-15 17:06:37 -0400809 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
Vivek Goyale43473b2010-09-15 17:06:35 -0400810
Vivek Goyal8e89d132010-09-15 17:06:37 -0400811 /* Slice has just started. Consider one slice interval */
812 if (!jiffy_elapsed)
813 jiffy_elapsed_rnd = throtl_slice;
814
815 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
816
Vivek Goyalc49c06e2010-10-01 21:16:42 +0200817 /*
818 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
819 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
820 * will allow dispatch after 1 second and after that slice should
821 * have been trimmed.
822 */
823
824 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
825 do_div(tmp, HZ);
826
827 if (tmp > UINT_MAX)
828 io_allowed = UINT_MAX;
829 else
830 io_allowed = tmp;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400831
832 if (tg->io_disp[rw] + 1 <= io_allowed) {
Vivek Goyale43473b2010-09-15 17:06:35 -0400833 if (wait)
834 *wait = 0;
835 return 1;
836 }
837
Vivek Goyal8e89d132010-09-15 17:06:37 -0400838 /* Calc approx time to dispatch */
839 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
840
841 if (jiffy_wait > jiffy_elapsed)
842 jiffy_wait = jiffy_wait - jiffy_elapsed;
843 else
844 jiffy_wait = 1;
845
846 if (wait)
847 *wait = jiffy_wait;
848 return 0;
849}
850
Tejun Heo0f3457f2013-05-14 13:52:32 -0700851static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
852 unsigned long *wait)
Vivek Goyal8e89d132010-09-15 17:06:37 -0400853{
854 bool rw = bio_data_dir(bio);
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200855 u64 bytes_allowed, extra_bytes, tmp;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400856 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
Vivek Goyale43473b2010-09-15 17:06:35 -0400857
858 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
859
860 /* Slice has just started. Consider one slice interval */
861 if (!jiffy_elapsed)
862 jiffy_elapsed_rnd = throtl_slice;
863
864 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
865
Vivek Goyal5e901a22010-10-01 21:16:38 +0200866 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
867 do_div(tmp, HZ);
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200868 bytes_allowed = tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400869
870 if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
871 if (wait)
872 *wait = 0;
873 return 1;
874 }
875
876 /* Calc approx time to dispatch */
877 extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
878 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
879
880 if (!jiffy_wait)
881 jiffy_wait = 1;
882
883 /*
884 * This wait time is without taking into consideration the rounding
885 * up we did. Add that time also.
886 */
887 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
Vivek Goyale43473b2010-09-15 17:06:35 -0400888 if (wait)
889 *wait = jiffy_wait;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400890 return 0;
891}
Vivek Goyale43473b2010-09-15 17:06:35 -0400892
Vivek Goyal8e89d132010-09-15 17:06:37 -0400893/*
894 * Returns whether one can dispatch a bio or not. Also returns approx number
895 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
896 */
Tejun Heo0f3457f2013-05-14 13:52:32 -0700897static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
898 unsigned long *wait)
Vivek Goyal8e89d132010-09-15 17:06:37 -0400899{
900 bool rw = bio_data_dir(bio);
901 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
902
903 /*
904 * Currently whole state machine of group depends on first bio
905 * queued in the group bio list. So one should not be calling
906 * this function with a different bio if there are other bios
907 * queued.
908 */
Tejun Heo73f0d492013-05-14 13:52:35 -0700909 BUG_ON(tg->service_queue.nr_queued[rw] &&
Tejun Heoc5cc2072013-05-14 13:52:38 -0700910 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
Vivek Goyal8e89d132010-09-15 17:06:37 -0400911
912 /* If tg->bps = -1, then BW is unlimited */
913 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
914 if (wait)
915 *wait = 0;
916 return 1;
917 }
918
919 /*
920 * If previous slice expired, start a new one otherwise renew/extend
921 * existing slice to make sure it is at least throtl_slice interval
922 * long since now.
923 */
Tejun Heo0f3457f2013-05-14 13:52:32 -0700924 if (throtl_slice_used(tg, rw))
925 throtl_start_new_slice(tg, rw);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400926 else {
927 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
Tejun Heo0f3457f2013-05-14 13:52:32 -0700928 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400929 }
930
Tejun Heo0f3457f2013-05-14 13:52:32 -0700931 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
932 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
Vivek Goyal8e89d132010-09-15 17:06:37 -0400933 if (wait)
934 *wait = 0;
935 return 1;
936 }
937
938 max_wait = max(bps_wait, iops_wait);
939
940 if (wait)
941 *wait = max_wait;
942
943 if (time_before(tg->slice_end[rw], jiffies + max_wait))
Tejun Heo0f3457f2013-05-14 13:52:32 -0700944 throtl_extend_slice(tg, rw, jiffies + max_wait);
Vivek Goyale43473b2010-09-15 17:06:35 -0400945
946 return 0;
947}
948
Tejun Heo3c798392012-04-16 13:57:25 -0700949static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
Tejun Heo629ed0b2012-04-01 14:38:44 -0700950 int rw)
951{
Tejun Heo8a3d2612012-04-01 14:38:44 -0700952 struct throtl_grp *tg = blkg_to_tg(blkg);
953 struct tg_stats_cpu *stats_cpu;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700954 unsigned long flags;
955
956 /* If per cpu stats are not allocated yet, don't do any accounting. */
Tejun Heo8a3d2612012-04-01 14:38:44 -0700957 if (tg->stats_cpu == NULL)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700958 return;
959
960 /*
961 * Disabling interrupts to provide mutual exclusion between two
962 * writes on same cpu. It probably is not needed for 64bit. Not
963 * optimizing that case yet.
964 */
965 local_irq_save(flags);
966
Tejun Heo8a3d2612012-04-01 14:38:44 -0700967 stats_cpu = this_cpu_ptr(tg->stats_cpu);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700968
Tejun Heo629ed0b2012-04-01 14:38:44 -0700969 blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
970 blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
971
972 local_irq_restore(flags);
973}
974
Vivek Goyale43473b2010-09-15 17:06:35 -0400975static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
976{
977 bool rw = bio_data_dir(bio);
Vivek Goyale43473b2010-09-15 17:06:35 -0400978
979 /* Charge the bio to the group */
980 tg->bytes_disp[rw] += bio->bi_size;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400981 tg->io_disp[rw]++;
Vivek Goyale43473b2010-09-15 17:06:35 -0400982
Tejun Heo2a0f61e2013-05-14 13:52:36 -0700983 /*
984 * REQ_THROTTLED is used to prevent the same bio to be throttled
985 * more than once as a throttled bio will go through blk-throtl the
986 * second time when it eventually gets issued. Set it when a bio
987 * is being charged to a tg.
988 *
989 * Dispatch stats aren't recursive and each @bio should only be
990 * accounted by the @tg it was originally associated with. Let's
991 * update the stats when setting REQ_THROTTLED for the first time
992 * which is guaranteed to be for the @bio's original tg.
993 */
994 if (!(bio->bi_rw & REQ_THROTTLED)) {
995 bio->bi_rw |= REQ_THROTTLED;
996 throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size,
997 bio->bi_rw);
998 }
Vivek Goyale43473b2010-09-15 17:06:35 -0400999}
1000
Tejun Heoc5cc2072013-05-14 13:52:38 -07001001/**
1002 * throtl_add_bio_tg - add a bio to the specified throtl_grp
1003 * @bio: bio to add
1004 * @qn: qnode to use
1005 * @tg: the target throtl_grp
1006 *
1007 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
1008 * tg->qnode_on_self[] is used.
1009 */
1010static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1011 struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -04001012{
Tejun Heo73f0d492013-05-14 13:52:35 -07001013 struct throtl_service_queue *sq = &tg->service_queue;
Vivek Goyale43473b2010-09-15 17:06:35 -04001014 bool rw = bio_data_dir(bio);
1015
Tejun Heoc5cc2072013-05-14 13:52:38 -07001016 if (!qn)
1017 qn = &tg->qnode_on_self[rw];
1018
Tejun Heo0e9f4162013-05-14 13:52:35 -07001019 /*
1020 * If @tg doesn't currently have any bios queued in the same
1021 * direction, queueing @bio can change when @tg should be
1022 * dispatched. Mark that @tg was empty. This is automatically
1023 * cleaered on the next tg_update_disptime().
1024 */
1025 if (!sq->nr_queued[rw])
1026 tg->flags |= THROTL_TG_WAS_EMPTY;
1027
Tejun Heoc5cc2072013-05-14 13:52:38 -07001028 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1029
Tejun Heo73f0d492013-05-14 13:52:35 -07001030 sq->nr_queued[rw]++;
Tejun Heo77216b02013-05-14 13:52:36 -07001031 throtl_enqueue_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -04001032}
1033
Tejun Heo77216b02013-05-14 13:52:36 -07001034static void tg_update_disptime(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -04001035{
Tejun Heo73f0d492013-05-14 13:52:35 -07001036 struct throtl_service_queue *sq = &tg->service_queue;
Vivek Goyale43473b2010-09-15 17:06:35 -04001037 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1038 struct bio *bio;
1039
Tejun Heoc5cc2072013-05-14 13:52:38 -07001040 if ((bio = throtl_peek_queued(&sq->queued[READ])))
Tejun Heo0f3457f2013-05-14 13:52:32 -07001041 tg_may_dispatch(tg, bio, &read_wait);
Vivek Goyale43473b2010-09-15 17:06:35 -04001042
Tejun Heoc5cc2072013-05-14 13:52:38 -07001043 if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
Tejun Heo0f3457f2013-05-14 13:52:32 -07001044 tg_may_dispatch(tg, bio, &write_wait);
Vivek Goyale43473b2010-09-15 17:06:35 -04001045
1046 min_wait = min(read_wait, write_wait);
1047 disptime = jiffies + min_wait;
1048
Vivek Goyale43473b2010-09-15 17:06:35 -04001049 /* Update dispatch time */
Tejun Heo77216b02013-05-14 13:52:36 -07001050 throtl_dequeue_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -04001051 tg->disptime = disptime;
Tejun Heo77216b02013-05-14 13:52:36 -07001052 throtl_enqueue_tg(tg);
Tejun Heo0e9f4162013-05-14 13:52:35 -07001053
1054 /* see throtl_add_bio_tg() */
1055 tg->flags &= ~THROTL_TG_WAS_EMPTY;
Vivek Goyale43473b2010-09-15 17:06:35 -04001056}
1057
Vivek Goyal32ee5bc2013-05-14 13:52:38 -07001058static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1059 struct throtl_grp *parent_tg, bool rw)
1060{
1061 if (throtl_slice_used(parent_tg, rw)) {
1062 throtl_start_new_slice_with_credit(parent_tg, rw,
1063 child_tg->slice_start[rw]);
1064 }
1065
1066}
1067
Tejun Heo77216b02013-05-14 13:52:36 -07001068static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
Vivek Goyale43473b2010-09-15 17:06:35 -04001069{
Tejun Heo73f0d492013-05-14 13:52:35 -07001070 struct throtl_service_queue *sq = &tg->service_queue;
Tejun Heo6bc9c2b2013-05-14 13:52:38 -07001071 struct throtl_service_queue *parent_sq = sq->parent_sq;
1072 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
Tejun Heoc5cc2072013-05-14 13:52:38 -07001073 struct throtl_grp *tg_to_put = NULL;
Vivek Goyale43473b2010-09-15 17:06:35 -04001074 struct bio *bio;
1075
Tejun Heoc5cc2072013-05-14 13:52:38 -07001076 /*
1077 * @bio is being transferred from @tg to @parent_sq. Popping a bio
1078 * from @tg may put its reference and @parent_sq might end up
1079 * getting released prematurely. Remember the tg to put and put it
1080 * after @bio is transferred to @parent_sq.
1081 */
1082 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
Tejun Heo73f0d492013-05-14 13:52:35 -07001083 sq->nr_queued[rw]--;
Vivek Goyale43473b2010-09-15 17:06:35 -04001084
1085 throtl_charge_bio(tg, bio);
Tejun Heo6bc9c2b2013-05-14 13:52:38 -07001086
1087 /*
1088 * If our parent is another tg, we just need to transfer @bio to
1089 * the parent using throtl_add_bio_tg(). If our parent is
1090 * @td->service_queue, @bio is ready to be issued. Put it on its
1091 * bio_lists[] and decrease total number queued. The caller is
1092 * responsible for issuing these bios.
1093 */
1094 if (parent_tg) {
Tejun Heoc5cc2072013-05-14 13:52:38 -07001095 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
Vivek Goyal32ee5bc2013-05-14 13:52:38 -07001096 start_parent_slice_with_credit(tg, parent_tg, rw);
Tejun Heo6bc9c2b2013-05-14 13:52:38 -07001097 } else {
Tejun Heoc5cc2072013-05-14 13:52:38 -07001098 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1099 &parent_sq->queued[rw]);
Tejun Heo6bc9c2b2013-05-14 13:52:38 -07001100 BUG_ON(tg->td->nr_queued[rw] <= 0);
1101 tg->td->nr_queued[rw]--;
1102 }
Vivek Goyale43473b2010-09-15 17:06:35 -04001103
Tejun Heo0f3457f2013-05-14 13:52:32 -07001104 throtl_trim_slice(tg, rw);
Tejun Heo6bc9c2b2013-05-14 13:52:38 -07001105
Tejun Heoc5cc2072013-05-14 13:52:38 -07001106 if (tg_to_put)
1107 blkg_put(tg_to_blkg(tg_to_put));
Vivek Goyale43473b2010-09-15 17:06:35 -04001108}
1109
Tejun Heo77216b02013-05-14 13:52:36 -07001110static int throtl_dispatch_tg(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -04001111{
Tejun Heo73f0d492013-05-14 13:52:35 -07001112 struct throtl_service_queue *sq = &tg->service_queue;
Vivek Goyale43473b2010-09-15 17:06:35 -04001113 unsigned int nr_reads = 0, nr_writes = 0;
1114 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
Vivek Goyalc2f68052010-11-15 19:32:42 +01001115 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
Vivek Goyale43473b2010-09-15 17:06:35 -04001116 struct bio *bio;
1117
1118 /* Try to dispatch 75% READS and 25% WRITES */
1119
Tejun Heoc5cc2072013-05-14 13:52:38 -07001120 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
Tejun Heo0f3457f2013-05-14 13:52:32 -07001121 tg_may_dispatch(tg, bio, NULL)) {
Vivek Goyale43473b2010-09-15 17:06:35 -04001122
Tejun Heo77216b02013-05-14 13:52:36 -07001123 tg_dispatch_one_bio(tg, bio_data_dir(bio));
Vivek Goyale43473b2010-09-15 17:06:35 -04001124 nr_reads++;
1125
1126 if (nr_reads >= max_nr_reads)
1127 break;
1128 }
1129
Tejun Heoc5cc2072013-05-14 13:52:38 -07001130 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
Tejun Heo0f3457f2013-05-14 13:52:32 -07001131 tg_may_dispatch(tg, bio, NULL)) {
Vivek Goyale43473b2010-09-15 17:06:35 -04001132
Tejun Heo77216b02013-05-14 13:52:36 -07001133 tg_dispatch_one_bio(tg, bio_data_dir(bio));
Vivek Goyale43473b2010-09-15 17:06:35 -04001134 nr_writes++;
1135
1136 if (nr_writes >= max_nr_writes)
1137 break;
1138 }
1139
1140 return nr_reads + nr_writes;
1141}
1142
Tejun Heo651930b2013-05-14 13:52:35 -07001143static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
Vivek Goyale43473b2010-09-15 17:06:35 -04001144{
1145 unsigned int nr_disp = 0;
Vivek Goyale43473b2010-09-15 17:06:35 -04001146
1147 while (1) {
Tejun Heo73f0d492013-05-14 13:52:35 -07001148 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1149 struct throtl_service_queue *sq = &tg->service_queue;
Vivek Goyale43473b2010-09-15 17:06:35 -04001150
1151 if (!tg)
1152 break;
1153
1154 if (time_before(jiffies, tg->disptime))
1155 break;
1156
Tejun Heo77216b02013-05-14 13:52:36 -07001157 throtl_dequeue_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -04001158
Tejun Heo77216b02013-05-14 13:52:36 -07001159 nr_disp += throtl_dispatch_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -04001160
Tejun Heo73f0d492013-05-14 13:52:35 -07001161 if (sq->nr_queued[0] || sq->nr_queued[1])
Tejun Heo77216b02013-05-14 13:52:36 -07001162 tg_update_disptime(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -04001163
1164 if (nr_disp >= throtl_quantum)
1165 break;
1166 }
1167
1168 return nr_disp;
1169}
1170
Tejun Heo6e1a5702013-05-14 13:52:37 -07001171/**
1172 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1173 * @arg: the throtl_service_queue being serviced
1174 *
1175 * This timer is armed when a child throtl_grp with active bio's become
1176 * pending and queued on the service_queue's pending_tree and expires when
1177 * the first child throtl_grp should be dispatched. This function
Tejun Heo2e48a532013-05-14 13:52:38 -07001178 * dispatches bio's from the children throtl_grps to the parent
1179 * service_queue.
1180 *
1181 * If the parent's parent is another throtl_grp, dispatching is propagated
1182 * by either arming its pending_timer or repeating dispatch directly. If
1183 * the top-level service_tree is reached, throtl_data->dispatch_work is
1184 * kicked so that the ready bio's are issued.
Tejun Heo6e1a5702013-05-14 13:52:37 -07001185 */
Tejun Heo69df0ab2013-05-14 13:52:36 -07001186static void throtl_pending_timer_fn(unsigned long arg)
1187{
1188 struct throtl_service_queue *sq = (void *)arg;
Tejun Heo2e48a532013-05-14 13:52:38 -07001189 struct throtl_grp *tg = sq_to_tg(sq);
Tejun Heo69df0ab2013-05-14 13:52:36 -07001190 struct throtl_data *td = sq_to_td(sq);
Tejun Heocb761992013-05-14 13:52:31 -07001191 struct request_queue *q = td->queue;
Tejun Heo2e48a532013-05-14 13:52:38 -07001192 struct throtl_service_queue *parent_sq;
1193 bool dispatched;
Tejun Heo6e1a5702013-05-14 13:52:37 -07001194 int ret;
Vivek Goyale43473b2010-09-15 17:06:35 -04001195
1196 spin_lock_irq(q->queue_lock);
Tejun Heo2e48a532013-05-14 13:52:38 -07001197again:
1198 parent_sq = sq->parent_sq;
1199 dispatched = false;
Vivek Goyale43473b2010-09-15 17:06:35 -04001200
Tejun Heo7f52f982013-05-14 13:52:37 -07001201 while (true) {
1202 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
Tejun Heo2e48a532013-05-14 13:52:38 -07001203 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1204 sq->nr_queued[READ], sq->nr_queued[WRITE]);
Vivek Goyale43473b2010-09-15 17:06:35 -04001205
Tejun Heo7f52f982013-05-14 13:52:37 -07001206 ret = throtl_select_dispatch(sq);
1207 if (ret) {
Tejun Heo7f52f982013-05-14 13:52:37 -07001208 throtl_log(sq, "bios disp=%u", ret);
1209 dispatched = true;
Tejun Heo651930b2013-05-14 13:52:35 -07001210 }
Vivek Goyale43473b2010-09-15 17:06:35 -04001211
Tejun Heo7f52f982013-05-14 13:52:37 -07001212 if (throtl_schedule_next_dispatch(sq, false))
1213 break;
1214
1215 /* this dispatch windows is still open, relax and repeat */
1216 spin_unlock_irq(q->queue_lock);
1217 cpu_relax();
1218 spin_lock_irq(q->queue_lock);
1219 }
Tejun Heo6a525602013-05-14 13:52:32 -07001220
Tejun Heo2e48a532013-05-14 13:52:38 -07001221 if (!dispatched)
1222 goto out_unlock;
Tejun Heo6e1a5702013-05-14 13:52:37 -07001223
Tejun Heo2e48a532013-05-14 13:52:38 -07001224 if (parent_sq) {
1225 /* @parent_sq is another throl_grp, propagate dispatch */
1226 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1227 tg_update_disptime(tg);
1228 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1229 /* window is already open, repeat dispatching */
1230 sq = parent_sq;
1231 tg = sq_to_tg(sq);
1232 goto again;
1233 }
1234 }
1235 } else {
1236 /* reached the top-level, queue issueing */
1237 queue_work(kthrotld_workqueue, &td->dispatch_work);
1238 }
1239out_unlock:
Tejun Heo6e1a5702013-05-14 13:52:37 -07001240 spin_unlock_irq(q->queue_lock);
1241}
1242
1243/**
1244 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1245 * @work: work item being executed
1246 *
1247 * This function is queued for execution when bio's reach the bio_lists[]
1248 * of throtl_data->service_queue. Those bio's are ready and issued by this
1249 * function.
1250 */
1251void blk_throtl_dispatch_work_fn(struct work_struct *work)
1252{
1253 struct throtl_data *td = container_of(work, struct throtl_data,
1254 dispatch_work);
1255 struct throtl_service_queue *td_sq = &td->service_queue;
1256 struct request_queue *q = td->queue;
1257 struct bio_list bio_list_on_stack;
1258 struct bio *bio;
1259 struct blk_plug plug;
1260 int rw;
1261
1262 bio_list_init(&bio_list_on_stack);
1263
1264 spin_lock_irq(q->queue_lock);
Tejun Heoc5cc2072013-05-14 13:52:38 -07001265 for (rw = READ; rw <= WRITE; rw++)
1266 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1267 bio_list_add(&bio_list_on_stack, bio);
Vivek Goyale43473b2010-09-15 17:06:35 -04001268 spin_unlock_irq(q->queue_lock);
1269
Tejun Heo6e1a5702013-05-14 13:52:37 -07001270 if (!bio_list_empty(&bio_list_on_stack)) {
Vivek Goyal69d60eb2011-03-09 08:27:37 +01001271 blk_start_plug(&plug);
Vivek Goyale43473b2010-09-15 17:06:35 -04001272 while((bio = bio_list_pop(&bio_list_on_stack)))
1273 generic_make_request(bio);
Vivek Goyal69d60eb2011-03-09 08:27:37 +01001274 blk_finish_plug(&plug);
Vivek Goyale43473b2010-09-15 17:06:35 -04001275 }
Vivek Goyale43473b2010-09-15 17:06:35 -04001276}
1277
Tejun Heof95a04a2012-04-16 13:57:26 -07001278static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
1279 struct blkg_policy_data *pd, int off)
Tejun Heo41b38b62012-04-01 14:38:44 -07001280{
Tejun Heof95a04a2012-04-16 13:57:26 -07001281 struct throtl_grp *tg = pd_to_tg(pd);
Tejun Heo41b38b62012-04-01 14:38:44 -07001282 struct blkg_rwstat rwstat = { }, tmp;
1283 int i, cpu;
1284
1285 for_each_possible_cpu(cpu) {
Tejun Heo8a3d2612012-04-01 14:38:44 -07001286 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
Tejun Heo41b38b62012-04-01 14:38:44 -07001287
1288 tmp = blkg_rwstat_read((void *)sc + off);
1289 for (i = 0; i < BLKG_RWSTAT_NR; i++)
1290 rwstat.cnt[i] += tmp.cnt[i];
1291 }
1292
Tejun Heof95a04a2012-04-16 13:57:26 -07001293 return __blkg_prfill_rwstat(sf, pd, &rwstat);
Tejun Heo41b38b62012-04-01 14:38:44 -07001294}
1295
Tejun Heo8a3d2612012-04-01 14:38:44 -07001296static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft,
1297 struct seq_file *sf)
Tejun Heo41b38b62012-04-01 14:38:44 -07001298{
Tejun Heo3c798392012-04-16 13:57:25 -07001299 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
Tejun Heo41b38b62012-04-01 14:38:44 -07001300
Tejun Heo3c798392012-04-16 13:57:25 -07001301 blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl,
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001302 cft->private, true);
Tejun Heo41b38b62012-04-01 14:38:44 -07001303 return 0;
1304}
1305
Tejun Heof95a04a2012-04-16 13:57:26 -07001306static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1307 int off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001308{
Tejun Heof95a04a2012-04-16 13:57:26 -07001309 struct throtl_grp *tg = pd_to_tg(pd);
1310 u64 v = *(u64 *)((void *)tg + off);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001311
Tejun Heoaf133ce2012-04-01 14:38:44 -07001312 if (v == -1)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001313 return 0;
Tejun Heof95a04a2012-04-16 13:57:26 -07001314 return __blkg_prfill_u64(sf, pd, v);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001315}
1316
Tejun Heof95a04a2012-04-16 13:57:26 -07001317static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1318 int off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001319{
Tejun Heof95a04a2012-04-16 13:57:26 -07001320 struct throtl_grp *tg = pd_to_tg(pd);
1321 unsigned int v = *(unsigned int *)((void *)tg + off);
Tejun Heoaf133ce2012-04-01 14:38:44 -07001322
1323 if (v == -1)
1324 return 0;
Tejun Heof95a04a2012-04-16 13:57:26 -07001325 return __blkg_prfill_u64(sf, pd, v);
Tejun Heoaf133ce2012-04-01 14:38:44 -07001326}
1327
1328static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft,
1329 struct seq_file *sf)
1330{
Tejun Heo3c798392012-04-16 13:57:25 -07001331 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64,
1332 &blkcg_policy_throtl, cft->private, false);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001333 return 0;
1334}
1335
Tejun Heoaf133ce2012-04-01 14:38:44 -07001336static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1337 struct seq_file *sf)
Vivek Goyale43473b2010-09-15 17:06:35 -04001338{
Tejun Heo3c798392012-04-16 13:57:25 -07001339 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint,
1340 &blkcg_policy_throtl, cft->private, false);
Tejun Heoaf133ce2012-04-01 14:38:44 -07001341 return 0;
Vivek Goyale43473b2010-09-15 17:06:35 -04001342}
1343
Tejun Heoaf133ce2012-04-01 14:38:44 -07001344static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf,
1345 bool is_u64)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001346{
Tejun Heo3c798392012-04-16 13:57:25 -07001347 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001348 struct blkg_conf_ctx ctx;
Tejun Heoaf133ce2012-04-01 14:38:44 -07001349 struct throtl_grp *tg;
Tejun Heo69df0ab2013-05-14 13:52:36 -07001350 struct throtl_service_queue *sq;
Tejun Heo693e7512013-05-14 13:52:38 -07001351 struct blkcg_gq *blkg;
1352 struct cgroup *pos_cgrp;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001353 int ret;
1354
Tejun Heo3c798392012-04-16 13:57:25 -07001355 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001356 if (ret)
1357 return ret;
1358
Tejun Heoaf133ce2012-04-01 14:38:44 -07001359 tg = blkg_to_tg(ctx.blkg);
Tejun Heo69df0ab2013-05-14 13:52:36 -07001360 sq = &tg->service_queue;
Tejun Heoaf133ce2012-04-01 14:38:44 -07001361
Tejun Heoa2b16932012-04-13 13:11:33 -07001362 if (!ctx.v)
1363 ctx.v = -1;
Tejun Heoaf133ce2012-04-01 14:38:44 -07001364
Tejun Heoa2b16932012-04-13 13:11:33 -07001365 if (is_u64)
1366 *(u64 *)((void *)tg + cft->private) = ctx.v;
1367 else
1368 *(unsigned int *)((void *)tg + cft->private) = ctx.v;
Tejun Heoaf133ce2012-04-01 14:38:44 -07001369
Tejun Heofda6f272013-05-14 13:52:36 -07001370 throtl_log(&tg->service_queue,
1371 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1372 tg->bps[READ], tg->bps[WRITE],
1373 tg->iops[READ], tg->iops[WRITE]);
Tejun Heo632b4492013-05-14 13:52:31 -07001374
1375 /*
Tejun Heo693e7512013-05-14 13:52:38 -07001376 * Update has_rules[] flags for the updated tg's subtree. A tg is
1377 * considered to have rules if either the tg itself or any of its
1378 * ancestors has rules. This identifies groups without any
1379 * restrictions in the whole hierarchy and allows them to bypass
1380 * blk-throttle.
1381 */
1382 tg_update_has_rules(tg);
1383 blkg_for_each_descendant_pre(blkg, pos_cgrp, ctx.blkg)
1384 tg_update_has_rules(blkg_to_tg(blkg));
1385
1386 /*
Tejun Heo632b4492013-05-14 13:52:31 -07001387 * We're already holding queue_lock and know @tg is valid. Let's
1388 * apply the new config directly.
1389 *
1390 * Restart the slices for both READ and WRITES. It might happen
1391 * that a group's limit are dropped suddenly and we don't want to
1392 * account recently dispatched IO with new low rate.
1393 */
Tejun Heo0f3457f2013-05-14 13:52:32 -07001394 throtl_start_new_slice(tg, 0);
1395 throtl_start_new_slice(tg, 1);
Tejun Heo632b4492013-05-14 13:52:31 -07001396
Tejun Heo5b2c16a2013-05-14 13:52:32 -07001397 if (tg->flags & THROTL_TG_PENDING) {
Tejun Heo77216b02013-05-14 13:52:36 -07001398 tg_update_disptime(tg);
Tejun Heo7f52f982013-05-14 13:52:37 -07001399 throtl_schedule_next_dispatch(sq->parent_sq, true);
Tejun Heo632b4492013-05-14 13:52:31 -07001400 }
Tejun Heo60c2bc22012-04-01 14:38:43 -07001401
1402 blkg_conf_finish(&ctx);
Tejun Heoa2b16932012-04-13 13:11:33 -07001403 return 0;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001404}
1405
Tejun Heoaf133ce2012-04-01 14:38:44 -07001406static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft,
1407 const char *buf)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001408{
Tejun Heoaf133ce2012-04-01 14:38:44 -07001409 return tg_set_conf(cgrp, cft, buf, true);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001410}
1411
Tejun Heoaf133ce2012-04-01 14:38:44 -07001412static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1413 const char *buf)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001414{
Tejun Heoaf133ce2012-04-01 14:38:44 -07001415 return tg_set_conf(cgrp, cft, buf, false);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001416}
1417
1418static struct cftype throtl_files[] = {
1419 {
1420 .name = "throttle.read_bps_device",
Tejun Heoaf133ce2012-04-01 14:38:44 -07001421 .private = offsetof(struct throtl_grp, bps[READ]),
1422 .read_seq_string = tg_print_conf_u64,
1423 .write_string = tg_set_conf_u64,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001424 .max_write_len = 256,
1425 },
1426 {
1427 .name = "throttle.write_bps_device",
Tejun Heoaf133ce2012-04-01 14:38:44 -07001428 .private = offsetof(struct throtl_grp, bps[WRITE]),
1429 .read_seq_string = tg_print_conf_u64,
1430 .write_string = tg_set_conf_u64,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001431 .max_write_len = 256,
1432 },
1433 {
1434 .name = "throttle.read_iops_device",
Tejun Heoaf133ce2012-04-01 14:38:44 -07001435 .private = offsetof(struct throtl_grp, iops[READ]),
1436 .read_seq_string = tg_print_conf_uint,
1437 .write_string = tg_set_conf_uint,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001438 .max_write_len = 256,
1439 },
1440 {
1441 .name = "throttle.write_iops_device",
Tejun Heoaf133ce2012-04-01 14:38:44 -07001442 .private = offsetof(struct throtl_grp, iops[WRITE]),
1443 .read_seq_string = tg_print_conf_uint,
1444 .write_string = tg_set_conf_uint,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001445 .max_write_len = 256,
1446 },
1447 {
1448 .name = "throttle.io_service_bytes",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001449 .private = offsetof(struct tg_stats_cpu, service_bytes),
Tejun Heo8a3d2612012-04-01 14:38:44 -07001450 .read_seq_string = tg_print_cpu_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001451 },
1452 {
1453 .name = "throttle.io_serviced",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001454 .private = offsetof(struct tg_stats_cpu, serviced),
Tejun Heo8a3d2612012-04-01 14:38:44 -07001455 .read_seq_string = tg_print_cpu_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001456 },
1457 { } /* terminate */
1458};
1459
Vivek Goyalda527772011-03-02 19:05:33 -05001460static void throtl_shutdown_wq(struct request_queue *q)
Vivek Goyale43473b2010-09-15 17:06:35 -04001461{
1462 struct throtl_data *td = q->td;
1463
Tejun Heo69df0ab2013-05-14 13:52:36 -07001464 cancel_work_sync(&td->dispatch_work);
Vivek Goyale43473b2010-09-15 17:06:35 -04001465}
1466
Tejun Heo3c798392012-04-16 13:57:25 -07001467static struct blkcg_policy blkcg_policy_throtl = {
Tejun Heof9fcc2d2012-04-16 13:57:27 -07001468 .pd_size = sizeof(struct throtl_grp),
1469 .cftypes = throtl_files,
1470
1471 .pd_init_fn = throtl_pd_init,
Tejun Heo693e7512013-05-14 13:52:38 -07001472 .pd_online_fn = throtl_pd_online,
Tejun Heof9fcc2d2012-04-16 13:57:27 -07001473 .pd_exit_fn = throtl_pd_exit,
1474 .pd_reset_stats_fn = throtl_pd_reset_stats,
Vivek Goyale43473b2010-09-15 17:06:35 -04001475};
1476
Tejun Heobc16a4f2011-10-19 14:33:01 +02001477bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
Vivek Goyale43473b2010-09-15 17:06:35 -04001478{
1479 struct throtl_data *td = q->td;
Tejun Heoc5cc2072013-05-14 13:52:38 -07001480 struct throtl_qnode *qn = NULL;
Vivek Goyale43473b2010-09-15 17:06:35 -04001481 struct throtl_grp *tg;
Tejun Heo73f0d492013-05-14 13:52:35 -07001482 struct throtl_service_queue *sq;
Tejun Heo0e9f4162013-05-14 13:52:35 -07001483 bool rw = bio_data_dir(bio);
Tejun Heo3c798392012-04-16 13:57:25 -07001484 struct blkcg *blkcg;
Tejun Heobc16a4f2011-10-19 14:33:01 +02001485 bool throttled = false;
Vivek Goyale43473b2010-09-15 17:06:35 -04001486
Tejun Heo2a0f61e2013-05-14 13:52:36 -07001487 /* see throtl_charge_bio() */
1488 if (bio->bi_rw & REQ_THROTTLED)
Tejun Heobc16a4f2011-10-19 14:33:01 +02001489 goto out;
Vivek Goyale43473b2010-09-15 17:06:35 -04001490
Vivek Goyalaf75cd32011-05-19 15:38:31 -04001491 /*
1492 * A throtl_grp pointer retrieved under rcu can be used to access
1493 * basic fields like stats and io rates. If a group has no rules,
1494 * just update the dispatch stats in lockless manner and return.
1495 */
Vivek Goyalaf75cd32011-05-19 15:38:31 -04001496 rcu_read_lock();
Tejun Heo3c798392012-04-16 13:57:25 -07001497 blkcg = bio_blkcg(bio);
Tejun Heocd1604f2012-03-05 13:15:06 -08001498 tg = throtl_lookup_tg(td, blkcg);
Vivek Goyalaf75cd32011-05-19 15:38:31 -04001499 if (tg) {
Tejun Heo693e7512013-05-14 13:52:38 -07001500 if (!tg->has_rules[rw]) {
Tejun Heo629ed0b2012-04-01 14:38:44 -07001501 throtl_update_dispatch_stats(tg_to_blkg(tg),
1502 bio->bi_size, bio->bi_rw);
Tejun Heo2a7f1242012-03-05 13:15:01 -08001503 goto out_unlock_rcu;
Vivek Goyalaf75cd32011-05-19 15:38:31 -04001504 }
1505 }
Vivek Goyalaf75cd32011-05-19 15:38:31 -04001506
1507 /*
1508 * Either group has not been allocated yet or it is not an unlimited
1509 * IO group
1510 */
Vivek Goyale43473b2010-09-15 17:06:35 -04001511 spin_lock_irq(q->queue_lock);
Tejun Heocd1604f2012-03-05 13:15:06 -08001512 tg = throtl_lookup_create_tg(td, blkcg);
Tejun Heobc16a4f2011-10-19 14:33:01 +02001513 if (unlikely(!tg))
1514 goto out_unlock;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001515
Tejun Heo73f0d492013-05-14 13:52:35 -07001516 sq = &tg->service_queue;
1517
Tejun Heo9e660ac2013-05-14 13:52:38 -07001518 while (true) {
1519 /* throtl is FIFO - if bios are already queued, should queue */
1520 if (sq->nr_queued[rw])
1521 break;
Vivek Goyalde701c72011-03-07 21:09:32 +01001522
Tejun Heo9e660ac2013-05-14 13:52:38 -07001523 /* if above limits, break to queue */
1524 if (!tg_may_dispatch(tg, bio, NULL))
1525 break;
1526
1527 /* within limits, let's charge and dispatch directly */
Vivek Goyale43473b2010-09-15 17:06:35 -04001528 throtl_charge_bio(tg, bio);
Vivek Goyal04521db2011-03-22 21:54:29 +01001529
1530 /*
1531 * We need to trim slice even when bios are not being queued
1532 * otherwise it might happen that a bio is not queued for
1533 * a long time and slice keeps on extending and trim is not
1534 * called for a long time. Now if limits are reduced suddenly
1535 * we take into account all the IO dispatched so far at new
1536 * low rate and * newly queued IO gets a really long dispatch
1537 * time.
1538 *
1539 * So keep on trimming slice even if bio is not queued.
1540 */
Tejun Heo0f3457f2013-05-14 13:52:32 -07001541 throtl_trim_slice(tg, rw);
Tejun Heo9e660ac2013-05-14 13:52:38 -07001542
1543 /*
1544 * @bio passed through this layer without being throttled.
1545 * Climb up the ladder. If we''re already at the top, it
1546 * can be executed directly.
1547 */
Tejun Heoc5cc2072013-05-14 13:52:38 -07001548 qn = &tg->qnode_on_parent[rw];
Tejun Heo9e660ac2013-05-14 13:52:38 -07001549 sq = sq->parent_sq;
1550 tg = sq_to_tg(sq);
1551 if (!tg)
1552 goto out_unlock;
Vivek Goyale43473b2010-09-15 17:06:35 -04001553 }
1554
Tejun Heo9e660ac2013-05-14 13:52:38 -07001555 /* out-of-limit, queue to @tg */
Tejun Heofda6f272013-05-14 13:52:36 -07001556 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1557 rw == READ ? 'R' : 'W',
1558 tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1559 tg->io_disp[rw], tg->iops[rw],
1560 sq->nr_queued[READ], sq->nr_queued[WRITE]);
Vivek Goyale43473b2010-09-15 17:06:35 -04001561
Tejun Heo671058f2012-03-05 13:15:29 -08001562 bio_associate_current(bio);
Tejun Heo6bc9c2b2013-05-14 13:52:38 -07001563 tg->td->nr_queued[rw]++;
Tejun Heoc5cc2072013-05-14 13:52:38 -07001564 throtl_add_bio_tg(bio, qn, tg);
Tejun Heobc16a4f2011-10-19 14:33:01 +02001565 throttled = true;
Vivek Goyale43473b2010-09-15 17:06:35 -04001566
Tejun Heo7f52f982013-05-14 13:52:37 -07001567 /*
1568 * Update @tg's dispatch time and force schedule dispatch if @tg
1569 * was empty before @bio. The forced scheduling isn't likely to
1570 * cause undue delay as @bio is likely to be dispatched directly if
1571 * its @tg's disptime is not in the future.
1572 */
Tejun Heo0e9f4162013-05-14 13:52:35 -07001573 if (tg->flags & THROTL_TG_WAS_EMPTY) {
Tejun Heo77216b02013-05-14 13:52:36 -07001574 tg_update_disptime(tg);
Tejun Heo7f52f982013-05-14 13:52:37 -07001575 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
Vivek Goyale43473b2010-09-15 17:06:35 -04001576 }
1577
Tejun Heobc16a4f2011-10-19 14:33:01 +02001578out_unlock:
Vivek Goyale43473b2010-09-15 17:06:35 -04001579 spin_unlock_irq(q->queue_lock);
Tejun Heo2a7f1242012-03-05 13:15:01 -08001580out_unlock_rcu:
1581 rcu_read_unlock();
Tejun Heobc16a4f2011-10-19 14:33:01 +02001582out:
Tejun Heo2a0f61e2013-05-14 13:52:36 -07001583 /*
1584 * As multiple blk-throtls may stack in the same issue path, we
1585 * don't want bios to leave with the flag set. Clear the flag if
1586 * being issued.
1587 */
1588 if (!throttled)
1589 bio->bi_rw &= ~REQ_THROTTLED;
Tejun Heobc16a4f2011-10-19 14:33:01 +02001590 return throttled;
Vivek Goyale43473b2010-09-15 17:06:35 -04001591}
1592
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001593/*
1594 * Dispatch all bios from all children tg's queued on @parent_sq. On
1595 * return, @parent_sq is guaranteed to not have any active children tg's
1596 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1597 */
1598static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1599{
1600 struct throtl_grp *tg;
1601
1602 while ((tg = throtl_rb_first(parent_sq))) {
1603 struct throtl_service_queue *sq = &tg->service_queue;
1604 struct bio *bio;
1605
1606 throtl_dequeue_tg(tg);
1607
Tejun Heoc5cc2072013-05-14 13:52:38 -07001608 while ((bio = throtl_peek_queued(&sq->queued[READ])))
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001609 tg_dispatch_one_bio(tg, bio_data_dir(bio));
Tejun Heoc5cc2072013-05-14 13:52:38 -07001610 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001611 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1612 }
1613}
1614
Tejun Heoc9a929d2011-10-19 14:42:16 +02001615/**
1616 * blk_throtl_drain - drain throttled bios
1617 * @q: request_queue to drain throttled bios for
1618 *
1619 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1620 */
1621void blk_throtl_drain(struct request_queue *q)
1622 __releases(q->queue_lock) __acquires(q->queue_lock)
1623{
1624 struct throtl_data *td = q->td;
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001625 struct blkcg_gq *blkg;
1626 struct cgroup *pos_cgrp;
Tejun Heoc9a929d2011-10-19 14:42:16 +02001627 struct bio *bio;
Tejun Heo651930b2013-05-14 13:52:35 -07001628 int rw;
Tejun Heoc9a929d2011-10-19 14:42:16 +02001629
Andi Kleen8bcb6c72012-03-30 12:33:28 +02001630 queue_lockdep_assert_held(q);
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001631 rcu_read_lock();
Tejun Heoc9a929d2011-10-19 14:42:16 +02001632
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001633 /*
1634 * Drain each tg while doing post-order walk on the blkg tree, so
1635 * that all bios are propagated to td->service_queue. It'd be
1636 * better to walk service_queue tree directly but blkg walk is
1637 * easier.
1638 */
1639 blkg_for_each_descendant_post(blkg, pos_cgrp, td->queue->root_blkg)
1640 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
Tejun Heo73f0d492013-05-14 13:52:35 -07001641
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001642 tg_drain_bios(&td_root_tg(td)->service_queue);
Tejun Heoc9a929d2011-10-19 14:42:16 +02001643
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001644 /* finally, transfer bios from top-level tg's into the td */
1645 tg_drain_bios(&td->service_queue);
1646
1647 rcu_read_unlock();
Tejun Heoc9a929d2011-10-19 14:42:16 +02001648 spin_unlock_irq(q->queue_lock);
1649
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001650 /* all bios now should be in td->service_queue, issue them */
Tejun Heo651930b2013-05-14 13:52:35 -07001651 for (rw = READ; rw <= WRITE; rw++)
Tejun Heoc5cc2072013-05-14 13:52:38 -07001652 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1653 NULL)))
Tejun Heo651930b2013-05-14 13:52:35 -07001654 generic_make_request(bio);
Tejun Heoc9a929d2011-10-19 14:42:16 +02001655
1656 spin_lock_irq(q->queue_lock);
1657}
1658
Vivek Goyale43473b2010-09-15 17:06:35 -04001659int blk_throtl_init(struct request_queue *q)
1660{
1661 struct throtl_data *td;
Tejun Heoa2b16932012-04-13 13:11:33 -07001662 int ret;
Vivek Goyale43473b2010-09-15 17:06:35 -04001663
1664 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1665 if (!td)
1666 return -ENOMEM;
1667
Tejun Heo69df0ab2013-05-14 13:52:36 -07001668 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
Tejun Heo77216b02013-05-14 13:52:36 -07001669 throtl_service_queue_init(&td->service_queue, NULL);
Vivek Goyale43473b2010-09-15 17:06:35 -04001670
Tejun Heocd1604f2012-03-05 13:15:06 -08001671 q->td = td;
Vivek Goyal29b12582011-05-19 15:38:24 -04001672 td->queue = q;
Vivek Goyal02977e42010-10-01 14:49:48 +02001673
Tejun Heoa2b16932012-04-13 13:11:33 -07001674 /* activate policy */
Tejun Heo3c798392012-04-16 13:57:25 -07001675 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
Tejun Heoa2b16932012-04-13 13:11:33 -07001676 if (ret)
Vivek Goyal29b12582011-05-19 15:38:24 -04001677 kfree(td);
Tejun Heoa2b16932012-04-13 13:11:33 -07001678 return ret;
Vivek Goyale43473b2010-09-15 17:06:35 -04001679}
1680
1681void blk_throtl_exit(struct request_queue *q)
1682{
Tejun Heoc875f4d2012-03-05 13:15:22 -08001683 BUG_ON(!q->td);
Vivek Goyalda527772011-03-02 19:05:33 -05001684 throtl_shutdown_wq(q);
Tejun Heo3c798392012-04-16 13:57:25 -07001685 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
Tejun Heoc9a929d2011-10-19 14:42:16 +02001686 kfree(q->td);
Vivek Goyale43473b2010-09-15 17:06:35 -04001687}
1688
1689static int __init throtl_init(void)
1690{
Vivek Goyal450adcb2011-03-01 13:40:54 -05001691 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1692 if (!kthrotld_workqueue)
1693 panic("Failed to create kthrotld\n");
1694
Tejun Heo3c798392012-04-16 13:57:25 -07001695 return blkcg_policy_register(&blkcg_policy_throtl);
Vivek Goyale43473b2010-09-15 17:06:35 -04001696}
1697
1698module_init(throtl_init);