blob: a8bb2fd8f5230a512665cf0d7d220f807e2d2e7d [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>
Tejun Heoeea8f412015-05-22 17:13:17 -040012#include <linux/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
86struct throtl_grp {
Tejun Heof95a04a2012-04-16 13:57:26 -070087 /* must be the first member */
88 struct blkg_policy_data pd;
89
Tejun Heoc9e03322013-05-14 13:52:32 -070090 /* active throtl group service_queue member */
Vivek Goyale43473b2010-09-15 17:06:35 -040091 struct rb_node rb_node;
92
Tejun Heo0f3457f2013-05-14 13:52:32 -070093 /* throtl_data this group belongs to */
94 struct throtl_data *td;
95
Tejun Heo49a2f1e2013-05-14 13:52:34 -070096 /* this group's service queue */
97 struct throtl_service_queue service_queue;
98
Vivek Goyale43473b2010-09-15 17:06:35 -040099 /*
Tejun Heoc5cc2072013-05-14 13:52:38 -0700100 * qnode_on_self is used when bios are directly queued to this
101 * throtl_grp so that local bios compete fairly with bios
102 * dispatched from children. qnode_on_parent is used when bios are
103 * dispatched from this throtl_grp into its parent and will compete
104 * with the sibling qnode_on_parents and the parent's
105 * qnode_on_self.
106 */
107 struct throtl_qnode qnode_on_self[2];
108 struct throtl_qnode qnode_on_parent[2];
109
110 /*
Vivek Goyale43473b2010-09-15 17:06:35 -0400111 * Dispatch time in jiffies. This is the estimated time when group
112 * will unthrottle and is ready to dispatch more bio. It is used as
113 * key to sort active groups in service tree.
114 */
115 unsigned long disptime;
116
Vivek Goyale43473b2010-09-15 17:06:35 -0400117 unsigned int flags;
118
Tejun Heo693e7512013-05-14 13:52:38 -0700119 /* are there any throtl rules between this group and td? */
120 bool has_rules[2];
121
Vivek Goyale43473b2010-09-15 17:06:35 -0400122 /* bytes per second rate limits */
123 uint64_t bps[2];
124
Vivek Goyal8e89d132010-09-15 17:06:37 -0400125 /* IOPS limits */
126 unsigned int iops[2];
127
Vivek Goyale43473b2010-09-15 17:06:35 -0400128 /* Number of bytes disptached in current slice */
129 uint64_t bytes_disp[2];
Vivek Goyal8e89d132010-09-15 17:06:37 -0400130 /* Number of bio's dispatched in current slice */
131 unsigned int io_disp[2];
Vivek Goyale43473b2010-09-15 17:06:35 -0400132
133 /* When did we start a new slice */
134 unsigned long slice_start[2];
135 unsigned long slice_end[2];
136};
137
138struct throtl_data
139{
Vivek Goyale43473b2010-09-15 17:06:35 -0400140 /* service tree for active throtl groups */
Tejun Heoc9e03322013-05-14 13:52:32 -0700141 struct throtl_service_queue service_queue;
Vivek Goyale43473b2010-09-15 17:06:35 -0400142
Vivek Goyale43473b2010-09-15 17:06:35 -0400143 struct request_queue *queue;
144
145 /* Total Number of queued bios on READ and WRITE lists */
146 unsigned int nr_queued[2];
147
148 /*
Vivek Goyal02977e42010-10-01 14:49:48 +0200149 * number of total undestroyed groups
Vivek Goyale43473b2010-09-15 17:06:35 -0400150 */
151 unsigned int nr_undestroyed_grps;
152
153 /* Work for dispatching throttled bios */
Tejun Heo69df0ab2013-05-14 13:52:36 -0700154 struct work_struct dispatch_work;
Vivek Goyale43473b2010-09-15 17:06:35 -0400155};
156
Tejun Heo69df0ab2013-05-14 13:52:36 -0700157static void throtl_pending_timer_fn(unsigned long arg);
158
Tejun Heof95a04a2012-04-16 13:57:26 -0700159static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
160{
161 return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
162}
163
Tejun Heo3c798392012-04-16 13:57:25 -0700164static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
Tejun Heo03814112012-03-05 13:15:14 -0800165{
Tejun Heof95a04a2012-04-16 13:57:26 -0700166 return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
Tejun Heo03814112012-03-05 13:15:14 -0800167}
168
Tejun Heo3c798392012-04-16 13:57:25 -0700169static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
Tejun Heo03814112012-03-05 13:15:14 -0800170{
Tejun Heof95a04a2012-04-16 13:57:26 -0700171 return pd_to_blkg(&tg->pd);
Tejun Heo03814112012-03-05 13:15:14 -0800172}
173
Tejun Heofda6f272013-05-14 13:52:36 -0700174/**
175 * sq_to_tg - return the throl_grp the specified service queue belongs to
176 * @sq: the throtl_service_queue of interest
177 *
178 * Return the throtl_grp @sq belongs to. If @sq is the top-level one
179 * embedded in throtl_data, %NULL is returned.
180 */
181static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
182{
183 if (sq && sq->parent_sq)
184 return container_of(sq, struct throtl_grp, service_queue);
185 else
186 return NULL;
187}
Vivek Goyale43473b2010-09-15 17:06:35 -0400188
Tejun Heofda6f272013-05-14 13:52:36 -0700189/**
190 * sq_to_td - return throtl_data the specified service queue belongs to
191 * @sq: the throtl_service_queue of interest
192 *
193 * A service_queue can be embeded in either a throtl_grp or throtl_data.
194 * Determine the associated throtl_data accordingly and return it.
195 */
196static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
197{
198 struct throtl_grp *tg = sq_to_tg(sq);
199
200 if (tg)
201 return tg->td;
202 else
203 return container_of(sq, struct throtl_data, service_queue);
204}
205
206/**
207 * throtl_log - log debug message via blktrace
208 * @sq: the service_queue being reported
209 * @fmt: printf format string
210 * @args: printf args
211 *
212 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
213 * throtl_grp; otherwise, just "throtl".
214 *
215 * TODO: this should be made a function and name formatting should happen
216 * after testing whether blktrace is enabled.
217 */
218#define throtl_log(sq, fmt, args...) do { \
219 struct throtl_grp *__tg = sq_to_tg((sq)); \
220 struct throtl_data *__td = sq_to_td((sq)); \
221 \
222 (void)__td; \
223 if ((__tg)) { \
224 char __pbuf[128]; \
225 \
226 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \
227 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
228 } else { \
229 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \
230 } \
231} while (0)
Vivek Goyale43473b2010-09-15 17:06:35 -0400232
Tejun Heoc5cc2072013-05-14 13:52:38 -0700233static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
234{
235 INIT_LIST_HEAD(&qn->node);
236 bio_list_init(&qn->bios);
237 qn->tg = tg;
238}
239
240/**
241 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
242 * @bio: bio being added
243 * @qn: qnode to add bio to
244 * @queued: the service_queue->queued[] list @qn belongs to
245 *
246 * Add @bio to @qn and put @qn on @queued if it's not already on.
247 * @qn->tg's reference count is bumped when @qn is activated. See the
248 * comment on top of throtl_qnode definition for details.
249 */
250static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
251 struct list_head *queued)
252{
253 bio_list_add(&qn->bios, bio);
254 if (list_empty(&qn->node)) {
255 list_add_tail(&qn->node, queued);
256 blkg_get(tg_to_blkg(qn->tg));
257 }
258}
259
260/**
261 * throtl_peek_queued - peek the first bio on a qnode list
262 * @queued: the qnode list to peek
263 */
264static struct bio *throtl_peek_queued(struct list_head *queued)
265{
266 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
267 struct bio *bio;
268
269 if (list_empty(queued))
270 return NULL;
271
272 bio = bio_list_peek(&qn->bios);
273 WARN_ON_ONCE(!bio);
274 return bio;
275}
276
277/**
278 * throtl_pop_queued - pop the first bio form a qnode list
279 * @queued: the qnode list to pop a bio from
280 * @tg_to_put: optional out argument for throtl_grp to put
281 *
282 * Pop the first bio from the qnode list @queued. After popping, the first
283 * qnode is removed from @queued if empty or moved to the end of @queued so
284 * that the popping order is round-robin.
285 *
286 * When the first qnode is removed, its associated throtl_grp should be put
287 * too. If @tg_to_put is NULL, this function automatically puts it;
288 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
289 * responsible for putting it.
290 */
291static struct bio *throtl_pop_queued(struct list_head *queued,
292 struct throtl_grp **tg_to_put)
293{
294 struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
295 struct bio *bio;
296
297 if (list_empty(queued))
298 return NULL;
299
300 bio = bio_list_pop(&qn->bios);
301 WARN_ON_ONCE(!bio);
302
303 if (bio_list_empty(&qn->bios)) {
304 list_del_init(&qn->node);
305 if (tg_to_put)
306 *tg_to_put = qn->tg;
307 else
308 blkg_put(tg_to_blkg(qn->tg));
309 } else {
310 list_move_tail(&qn->node, queued);
311 }
312
313 return bio;
314}
315
Tejun Heo49a2f1e2013-05-14 13:52:34 -0700316/* init a service_queue, assumes the caller zeroed it */
Tejun Heob2ce2642015-08-18 14:55:13 -0700317static void throtl_service_queue_init(struct throtl_service_queue *sq)
Tejun Heo49a2f1e2013-05-14 13:52:34 -0700318{
Tejun Heoc5cc2072013-05-14 13:52:38 -0700319 INIT_LIST_HEAD(&sq->queued[0]);
320 INIT_LIST_HEAD(&sq->queued[1]);
Tejun Heo49a2f1e2013-05-14 13:52:34 -0700321 sq->pending_tree = RB_ROOT;
Tejun Heo69df0ab2013-05-14 13:52:36 -0700322 setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
323 (unsigned long)sq);
324}
325
Tejun Heo001bea72015-08-18 14:55:11 -0700326static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
327{
Tejun Heo4fb72032015-08-18 14:55:12 -0700328 struct throtl_grp *tg;
Tejun Heo24bdb8e2015-08-18 14:55:22 -0700329 int rw;
Tejun Heo4fb72032015-08-18 14:55:12 -0700330
331 tg = kzalloc_node(sizeof(*tg), gfp, node);
332 if (!tg)
Tejun Heo77ea7332015-08-18 14:55:24 -0700333 return NULL;
Tejun Heo4fb72032015-08-18 14:55:12 -0700334
Tejun Heob2ce2642015-08-18 14:55:13 -0700335 throtl_service_queue_init(&tg->service_queue);
336
337 for (rw = READ; rw <= WRITE; rw++) {
338 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
339 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
340 }
341
342 RB_CLEAR_NODE(&tg->rb_node);
343 tg->bps[READ] = -1;
344 tg->bps[WRITE] = -1;
345 tg->iops[READ] = -1;
346 tg->iops[WRITE] = -1;
347
Tejun Heo4fb72032015-08-18 14:55:12 -0700348 return &tg->pd;
Tejun Heo001bea72015-08-18 14:55:11 -0700349}
350
Tejun Heoa9520cd2015-08-18 14:55:14 -0700351static void throtl_pd_init(struct blkg_policy_data *pd)
Vivek Goyala29a1712011-05-19 15:38:19 -0400352{
Tejun Heoa9520cd2015-08-18 14:55:14 -0700353 struct throtl_grp *tg = pd_to_tg(pd);
354 struct blkcg_gq *blkg = tg_to_blkg(tg);
Tejun Heo77216b02013-05-14 13:52:36 -0700355 struct throtl_data *td = blkg->q->td;
Tejun Heob2ce2642015-08-18 14:55:13 -0700356 struct throtl_service_queue *sq = &tg->service_queue;
Tejun Heocd1604f2012-03-05 13:15:06 -0800357
Tejun Heo91381252013-05-14 13:52:38 -0700358 /*
Tejun Heoaa6ec292014-07-09 10:08:08 -0400359 * If on the default hierarchy, we switch to properly hierarchical
Tejun Heo91381252013-05-14 13:52:38 -0700360 * behavior where limits on a given throtl_grp are applied to the
361 * whole subtree rather than just the group itself. e.g. If 16M
362 * read_bps limit is set on the root group, the whole system can't
363 * exceed 16M for the device.
364 *
Tejun Heoaa6ec292014-07-09 10:08:08 -0400365 * If not on the default hierarchy, the broken flat hierarchy
Tejun Heo91381252013-05-14 13:52:38 -0700366 * behavior is retained where all throtl_grps are treated as if
367 * they're all separate root groups right below throtl_data.
368 * Limits of a group don't interact with limits of other groups
369 * regardless of the position of the group in the hierarchy.
370 */
Tejun Heob2ce2642015-08-18 14:55:13 -0700371 sq->parent_sq = &td->service_queue;
Tejun Heoaa6ec292014-07-09 10:08:08 -0400372 if (cgroup_on_dfl(blkg->blkcg->css.cgroup) && blkg->parent)
Tejun Heob2ce2642015-08-18 14:55:13 -0700373 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
Tejun Heo77216b02013-05-14 13:52:36 -0700374 tg->td = td;
Tejun Heo8a3d2612012-04-01 14:38:44 -0700375}
376
Tejun Heo693e7512013-05-14 13:52:38 -0700377/*
378 * Set has_rules[] if @tg or any of its parents have limits configured.
379 * This doesn't require walking up to the top of the hierarchy as the
380 * parent's has_rules[] is guaranteed to be correct.
381 */
382static void tg_update_has_rules(struct throtl_grp *tg)
383{
384 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
385 int rw;
386
387 for (rw = READ; rw <= WRITE; rw++)
388 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
389 (tg->bps[rw] != -1 || tg->iops[rw] != -1);
390}
391
Tejun Heoa9520cd2015-08-18 14:55:14 -0700392static void throtl_pd_online(struct blkg_policy_data *pd)
Tejun Heo693e7512013-05-14 13:52:38 -0700393{
394 /*
395 * We don't want new groups to escape the limits of its ancestors.
396 * Update has_rules[] after a new group is brought online.
397 */
Tejun Heoa9520cd2015-08-18 14:55:14 -0700398 tg_update_has_rules(pd_to_tg(pd));
Tejun Heo693e7512013-05-14 13:52:38 -0700399}
400
Tejun Heo001bea72015-08-18 14:55:11 -0700401static void throtl_pd_free(struct blkg_policy_data *pd)
402{
Tejun Heo4fb72032015-08-18 14:55:12 -0700403 struct throtl_grp *tg = pd_to_tg(pd);
404
Tejun Heob2ce2642015-08-18 14:55:13 -0700405 del_timer_sync(&tg->service_queue.pending_timer);
Tejun Heo4fb72032015-08-18 14:55:12 -0700406 kfree(tg);
Tejun Heo001bea72015-08-18 14:55:11 -0700407}
408
Tejun Heo0049af72013-05-14 13:52:33 -0700409static struct throtl_grp *
410throtl_rb_first(struct throtl_service_queue *parent_sq)
Vivek Goyale43473b2010-09-15 17:06:35 -0400411{
412 /* Service tree is empty */
Tejun Heo0049af72013-05-14 13:52:33 -0700413 if (!parent_sq->nr_pending)
Vivek Goyale43473b2010-09-15 17:06:35 -0400414 return NULL;
415
Tejun Heo0049af72013-05-14 13:52:33 -0700416 if (!parent_sq->first_pending)
417 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
Vivek Goyale43473b2010-09-15 17:06:35 -0400418
Tejun Heo0049af72013-05-14 13:52:33 -0700419 if (parent_sq->first_pending)
420 return rb_entry_tg(parent_sq->first_pending);
Vivek Goyale43473b2010-09-15 17:06:35 -0400421
422 return NULL;
423}
424
425static void rb_erase_init(struct rb_node *n, struct rb_root *root)
426{
427 rb_erase(n, root);
428 RB_CLEAR_NODE(n);
429}
430
Tejun Heo0049af72013-05-14 13:52:33 -0700431static void throtl_rb_erase(struct rb_node *n,
432 struct throtl_service_queue *parent_sq)
Vivek Goyale43473b2010-09-15 17:06:35 -0400433{
Tejun Heo0049af72013-05-14 13:52:33 -0700434 if (parent_sq->first_pending == n)
435 parent_sq->first_pending = NULL;
436 rb_erase_init(n, &parent_sq->pending_tree);
437 --parent_sq->nr_pending;
Vivek Goyale43473b2010-09-15 17:06:35 -0400438}
439
Tejun Heo0049af72013-05-14 13:52:33 -0700440static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
Vivek Goyale43473b2010-09-15 17:06:35 -0400441{
442 struct throtl_grp *tg;
443
Tejun Heo0049af72013-05-14 13:52:33 -0700444 tg = throtl_rb_first(parent_sq);
Vivek Goyale43473b2010-09-15 17:06:35 -0400445 if (!tg)
446 return;
447
Tejun Heo0049af72013-05-14 13:52:33 -0700448 parent_sq->first_pending_disptime = tg->disptime;
Vivek Goyale43473b2010-09-15 17:06:35 -0400449}
450
Tejun Heo77216b02013-05-14 13:52:36 -0700451static void tg_service_queue_add(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400452{
Tejun Heo77216b02013-05-14 13:52:36 -0700453 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
Tejun Heo0049af72013-05-14 13:52:33 -0700454 struct rb_node **node = &parent_sq->pending_tree.rb_node;
Vivek Goyale43473b2010-09-15 17:06:35 -0400455 struct rb_node *parent = NULL;
456 struct throtl_grp *__tg;
457 unsigned long key = tg->disptime;
458 int left = 1;
459
460 while (*node != NULL) {
461 parent = *node;
462 __tg = rb_entry_tg(parent);
463
464 if (time_before(key, __tg->disptime))
465 node = &parent->rb_left;
466 else {
467 node = &parent->rb_right;
468 left = 0;
469 }
470 }
471
472 if (left)
Tejun Heo0049af72013-05-14 13:52:33 -0700473 parent_sq->first_pending = &tg->rb_node;
Vivek Goyale43473b2010-09-15 17:06:35 -0400474
475 rb_link_node(&tg->rb_node, parent, node);
Tejun Heo0049af72013-05-14 13:52:33 -0700476 rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
Vivek Goyale43473b2010-09-15 17:06:35 -0400477}
478
Tejun Heo77216b02013-05-14 13:52:36 -0700479static void __throtl_enqueue_tg(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400480{
Tejun Heo77216b02013-05-14 13:52:36 -0700481 tg_service_queue_add(tg);
Tejun Heo5b2c16a2013-05-14 13:52:32 -0700482 tg->flags |= THROTL_TG_PENDING;
Tejun Heo77216b02013-05-14 13:52:36 -0700483 tg->service_queue.parent_sq->nr_pending++;
Vivek Goyale43473b2010-09-15 17:06:35 -0400484}
485
Tejun Heo77216b02013-05-14 13:52:36 -0700486static void throtl_enqueue_tg(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400487{
Tejun Heo5b2c16a2013-05-14 13:52:32 -0700488 if (!(tg->flags & THROTL_TG_PENDING))
Tejun Heo77216b02013-05-14 13:52:36 -0700489 __throtl_enqueue_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -0400490}
491
Tejun Heo77216b02013-05-14 13:52:36 -0700492static void __throtl_dequeue_tg(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400493{
Tejun Heo77216b02013-05-14 13:52:36 -0700494 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
Tejun Heo5b2c16a2013-05-14 13:52:32 -0700495 tg->flags &= ~THROTL_TG_PENDING;
Vivek Goyale43473b2010-09-15 17:06:35 -0400496}
497
Tejun Heo77216b02013-05-14 13:52:36 -0700498static void throtl_dequeue_tg(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400499{
Tejun Heo5b2c16a2013-05-14 13:52:32 -0700500 if (tg->flags & THROTL_TG_PENDING)
Tejun Heo77216b02013-05-14 13:52:36 -0700501 __throtl_dequeue_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -0400502}
503
Tejun Heoa9131a22013-05-14 13:52:31 -0700504/* Call with queue lock held */
Tejun Heo69df0ab2013-05-14 13:52:36 -0700505static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
506 unsigned long expires)
Tejun Heoa9131a22013-05-14 13:52:31 -0700507{
Tejun Heo69df0ab2013-05-14 13:52:36 -0700508 mod_timer(&sq->pending_timer, expires);
509 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
510 expires - jiffies, jiffies);
Tejun Heoa9131a22013-05-14 13:52:31 -0700511}
512
Tejun Heo7f52f982013-05-14 13:52:37 -0700513/**
514 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
515 * @sq: the service_queue to schedule dispatch for
516 * @force: force scheduling
517 *
518 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
519 * dispatch time of the first pending child. Returns %true if either timer
520 * is armed or there's no pending child left. %false if the current
521 * dispatch window is still open and the caller should continue
522 * dispatching.
523 *
524 * If @force is %true, the dispatch timer is always scheduled and this
525 * function is guaranteed to return %true. This is to be used when the
526 * caller can't dispatch itself and needs to invoke pending_timer
527 * unconditionally. Note that forced scheduling is likely to induce short
528 * delay before dispatch starts even if @sq->first_pending_disptime is not
529 * in the future and thus shouldn't be used in hot paths.
530 */
531static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
532 bool force)
Vivek Goyale43473b2010-09-15 17:06:35 -0400533{
Tejun Heo6a525602013-05-14 13:52:32 -0700534 /* any pending children left? */
Tejun Heoc9e03322013-05-14 13:52:32 -0700535 if (!sq->nr_pending)
Tejun Heo7f52f982013-05-14 13:52:37 -0700536 return true;
Vivek Goyale43473b2010-09-15 17:06:35 -0400537
Tejun Heoc9e03322013-05-14 13:52:32 -0700538 update_min_dispatch_time(sq);
Vivek Goyale43473b2010-09-15 17:06:35 -0400539
Tejun Heo69df0ab2013-05-14 13:52:36 -0700540 /* is the next dispatch time in the future? */
Tejun Heo7f52f982013-05-14 13:52:37 -0700541 if (force || time_after(sq->first_pending_disptime, jiffies)) {
Tejun Heo69df0ab2013-05-14 13:52:36 -0700542 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
Tejun Heo7f52f982013-05-14 13:52:37 -0700543 return true;
Tejun Heo69df0ab2013-05-14 13:52:36 -0700544 }
545
Tejun Heo7f52f982013-05-14 13:52:37 -0700546 /* tell the caller to continue dispatching */
547 return false;
Vivek Goyale43473b2010-09-15 17:06:35 -0400548}
549
Vivek Goyal32ee5bc2013-05-14 13:52:38 -0700550static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
551 bool rw, unsigned long start)
552{
553 tg->bytes_disp[rw] = 0;
554 tg->io_disp[rw] = 0;
555
556 /*
557 * Previous slice has expired. We must have trimmed it after last
558 * bio dispatch. That means since start of last slice, we never used
559 * that bandwidth. Do try to make use of that bandwidth while giving
560 * credit.
561 */
562 if (time_after_eq(start, tg->slice_start[rw]))
563 tg->slice_start[rw] = start;
564
565 tg->slice_end[rw] = jiffies + throtl_slice;
566 throtl_log(&tg->service_queue,
567 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
568 rw == READ ? 'R' : 'W', tg->slice_start[rw],
569 tg->slice_end[rw], jiffies);
570}
571
Tejun Heo0f3457f2013-05-14 13:52:32 -0700572static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
Vivek Goyale43473b2010-09-15 17:06:35 -0400573{
574 tg->bytes_disp[rw] = 0;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400575 tg->io_disp[rw] = 0;
Vivek Goyale43473b2010-09-15 17:06:35 -0400576 tg->slice_start[rw] = jiffies;
577 tg->slice_end[rw] = jiffies + throtl_slice;
Tejun Heofda6f272013-05-14 13:52:36 -0700578 throtl_log(&tg->service_queue,
579 "[%c] new slice start=%lu end=%lu jiffies=%lu",
580 rw == READ ? 'R' : 'W', tg->slice_start[rw],
581 tg->slice_end[rw], jiffies);
Vivek Goyale43473b2010-09-15 17:06:35 -0400582}
583
Tejun Heo0f3457f2013-05-14 13:52:32 -0700584static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
585 unsigned long jiffy_end)
Vivek Goyald1ae8ff2010-12-01 19:34:46 +0100586{
587 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
588}
589
Tejun Heo0f3457f2013-05-14 13:52:32 -0700590static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
591 unsigned long jiffy_end)
Vivek Goyale43473b2010-09-15 17:06:35 -0400592{
593 tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
Tejun Heofda6f272013-05-14 13:52:36 -0700594 throtl_log(&tg->service_queue,
595 "[%c] extend slice start=%lu end=%lu jiffies=%lu",
596 rw == READ ? 'R' : 'W', tg->slice_start[rw],
597 tg->slice_end[rw], jiffies);
Vivek Goyale43473b2010-09-15 17:06:35 -0400598}
599
600/* Determine if previously allocated or extended slice is complete or not */
Tejun Heo0f3457f2013-05-14 13:52:32 -0700601static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
Vivek Goyale43473b2010-09-15 17:06:35 -0400602{
603 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
Fabian Frederick5cf8c222014-05-02 18:28:17 +0200604 return false;
Vivek Goyale43473b2010-09-15 17:06:35 -0400605
606 return 1;
607}
608
609/* Trim the used slices and adjust slice start accordingly */
Tejun Heo0f3457f2013-05-14 13:52:32 -0700610static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
Vivek Goyale43473b2010-09-15 17:06:35 -0400611{
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200612 unsigned long nr_slices, time_elapsed, io_trim;
613 u64 bytes_trim, tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400614
615 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
616
617 /*
618 * If bps are unlimited (-1), then time slice don't get
619 * renewed. Don't try to trim the slice if slice is used. A new
620 * slice will start when appropriate.
621 */
Tejun Heo0f3457f2013-05-14 13:52:32 -0700622 if (throtl_slice_used(tg, rw))
Vivek Goyale43473b2010-09-15 17:06:35 -0400623 return;
624
Vivek Goyald1ae8ff2010-12-01 19:34:46 +0100625 /*
626 * A bio has been dispatched. Also adjust slice_end. It might happen
627 * that initially cgroup limit was very low resulting in high
628 * slice_end, but later limit was bumped up and bio was dispached
629 * sooner, then we need to reduce slice_end. A high bogus slice_end
630 * is bad because it does not allow new slice to start.
631 */
632
Tejun Heo0f3457f2013-05-14 13:52:32 -0700633 throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
Vivek Goyald1ae8ff2010-12-01 19:34:46 +0100634
Vivek Goyale43473b2010-09-15 17:06:35 -0400635 time_elapsed = jiffies - tg->slice_start[rw];
636
637 nr_slices = time_elapsed / throtl_slice;
638
639 if (!nr_slices)
640 return;
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200641 tmp = tg->bps[rw] * throtl_slice * nr_slices;
642 do_div(tmp, HZ);
643 bytes_trim = tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400644
Vivek Goyal8e89d132010-09-15 17:06:37 -0400645 io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
Vivek Goyale43473b2010-09-15 17:06:35 -0400646
Vivek Goyal8e89d132010-09-15 17:06:37 -0400647 if (!bytes_trim && !io_trim)
Vivek Goyale43473b2010-09-15 17:06:35 -0400648 return;
649
650 if (tg->bytes_disp[rw] >= bytes_trim)
651 tg->bytes_disp[rw] -= bytes_trim;
652 else
653 tg->bytes_disp[rw] = 0;
654
Vivek Goyal8e89d132010-09-15 17:06:37 -0400655 if (tg->io_disp[rw] >= io_trim)
656 tg->io_disp[rw] -= io_trim;
657 else
658 tg->io_disp[rw] = 0;
659
Vivek Goyale43473b2010-09-15 17:06:35 -0400660 tg->slice_start[rw] += nr_slices * throtl_slice;
661
Tejun Heofda6f272013-05-14 13:52:36 -0700662 throtl_log(&tg->service_queue,
663 "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
664 rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
665 tg->slice_start[rw], tg->slice_end[rw], jiffies);
Vivek Goyale43473b2010-09-15 17:06:35 -0400666}
667
Tejun Heo0f3457f2013-05-14 13:52:32 -0700668static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
669 unsigned long *wait)
Vivek Goyale43473b2010-09-15 17:06:35 -0400670{
671 bool rw = bio_data_dir(bio);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400672 unsigned int io_allowed;
Vivek Goyale43473b2010-09-15 17:06:35 -0400673 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
Vivek Goyalc49c06e2010-10-01 21:16:42 +0200674 u64 tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400675
Vivek Goyal8e89d132010-09-15 17:06:37 -0400676 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
Vivek Goyale43473b2010-09-15 17:06:35 -0400677
Vivek Goyal8e89d132010-09-15 17:06:37 -0400678 /* Slice has just started. Consider one slice interval */
679 if (!jiffy_elapsed)
680 jiffy_elapsed_rnd = throtl_slice;
681
682 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
683
Vivek Goyalc49c06e2010-10-01 21:16:42 +0200684 /*
685 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
686 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
687 * will allow dispatch after 1 second and after that slice should
688 * have been trimmed.
689 */
690
691 tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
692 do_div(tmp, HZ);
693
694 if (tmp > UINT_MAX)
695 io_allowed = UINT_MAX;
696 else
697 io_allowed = tmp;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400698
699 if (tg->io_disp[rw] + 1 <= io_allowed) {
Vivek Goyale43473b2010-09-15 17:06:35 -0400700 if (wait)
701 *wait = 0;
Fabian Frederick5cf8c222014-05-02 18:28:17 +0200702 return true;
Vivek Goyale43473b2010-09-15 17:06:35 -0400703 }
704
Vivek Goyal8e89d132010-09-15 17:06:37 -0400705 /* Calc approx time to dispatch */
706 jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
707
708 if (jiffy_wait > jiffy_elapsed)
709 jiffy_wait = jiffy_wait - jiffy_elapsed;
710 else
711 jiffy_wait = 1;
712
713 if (wait)
714 *wait = jiffy_wait;
715 return 0;
716}
717
Tejun Heo0f3457f2013-05-14 13:52:32 -0700718static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
719 unsigned long *wait)
Vivek Goyal8e89d132010-09-15 17:06:37 -0400720{
721 bool rw = bio_data_dir(bio);
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200722 u64 bytes_allowed, extra_bytes, tmp;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400723 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
Vivek Goyale43473b2010-09-15 17:06:35 -0400724
725 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
726
727 /* Slice has just started. Consider one slice interval */
728 if (!jiffy_elapsed)
729 jiffy_elapsed_rnd = throtl_slice;
730
731 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
732
Vivek Goyal5e901a22010-10-01 21:16:38 +0200733 tmp = tg->bps[rw] * jiffy_elapsed_rnd;
734 do_div(tmp, HZ);
Vivek Goyal3aad5d32010-10-01 14:51:14 +0200735 bytes_allowed = tmp;
Vivek Goyale43473b2010-09-15 17:06:35 -0400736
Kent Overstreet4f024f32013-10-11 15:44:27 -0700737 if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
Vivek Goyale43473b2010-09-15 17:06:35 -0400738 if (wait)
739 *wait = 0;
Fabian Frederick5cf8c222014-05-02 18:28:17 +0200740 return true;
Vivek Goyale43473b2010-09-15 17:06:35 -0400741 }
742
743 /* Calc approx time to dispatch */
Kent Overstreet4f024f32013-10-11 15:44:27 -0700744 extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
Vivek Goyale43473b2010-09-15 17:06:35 -0400745 jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
746
747 if (!jiffy_wait)
748 jiffy_wait = 1;
749
750 /*
751 * This wait time is without taking into consideration the rounding
752 * up we did. Add that time also.
753 */
754 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
Vivek Goyale43473b2010-09-15 17:06:35 -0400755 if (wait)
756 *wait = jiffy_wait;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400757 return 0;
758}
Vivek Goyale43473b2010-09-15 17:06:35 -0400759
Vivek Goyal8e89d132010-09-15 17:06:37 -0400760/*
761 * Returns whether one can dispatch a bio or not. Also returns approx number
762 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
763 */
Tejun Heo0f3457f2013-05-14 13:52:32 -0700764static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
765 unsigned long *wait)
Vivek Goyal8e89d132010-09-15 17:06:37 -0400766{
767 bool rw = bio_data_dir(bio);
768 unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
769
770 /*
771 * Currently whole state machine of group depends on first bio
772 * queued in the group bio list. So one should not be calling
773 * this function with a different bio if there are other bios
774 * queued.
775 */
Tejun Heo73f0d492013-05-14 13:52:35 -0700776 BUG_ON(tg->service_queue.nr_queued[rw] &&
Tejun Heoc5cc2072013-05-14 13:52:38 -0700777 bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
Vivek Goyal8e89d132010-09-15 17:06:37 -0400778
779 /* If tg->bps = -1, then BW is unlimited */
780 if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
781 if (wait)
782 *wait = 0;
Fabian Frederick5cf8c222014-05-02 18:28:17 +0200783 return true;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400784 }
785
786 /*
787 * If previous slice expired, start a new one otherwise renew/extend
788 * existing slice to make sure it is at least throtl_slice interval
789 * long since now.
790 */
Tejun Heo0f3457f2013-05-14 13:52:32 -0700791 if (throtl_slice_used(tg, rw))
792 throtl_start_new_slice(tg, rw);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400793 else {
794 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
Tejun Heo0f3457f2013-05-14 13:52:32 -0700795 throtl_extend_slice(tg, rw, jiffies + throtl_slice);
Vivek Goyal8e89d132010-09-15 17:06:37 -0400796 }
797
Tejun Heo0f3457f2013-05-14 13:52:32 -0700798 if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
799 tg_with_in_iops_limit(tg, bio, &iops_wait)) {
Vivek Goyal8e89d132010-09-15 17:06:37 -0400800 if (wait)
801 *wait = 0;
802 return 1;
803 }
804
805 max_wait = max(bps_wait, iops_wait);
806
807 if (wait)
808 *wait = max_wait;
809
810 if (time_before(tg->slice_end[rw], jiffies + max_wait))
Tejun Heo0f3457f2013-05-14 13:52:32 -0700811 throtl_extend_slice(tg, rw, jiffies + max_wait);
Vivek Goyale43473b2010-09-15 17:06:35 -0400812
813 return 0;
814}
815
816static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
817{
818 bool rw = bio_data_dir(bio);
Vivek Goyale43473b2010-09-15 17:06:35 -0400819
820 /* Charge the bio to the group */
Kent Overstreet4f024f32013-10-11 15:44:27 -0700821 tg->bytes_disp[rw] += bio->bi_iter.bi_size;
Vivek Goyal8e89d132010-09-15 17:06:37 -0400822 tg->io_disp[rw]++;
Vivek Goyale43473b2010-09-15 17:06:35 -0400823
Tejun Heo2a0f61e2013-05-14 13:52:36 -0700824 /*
825 * REQ_THROTTLED is used to prevent the same bio to be throttled
826 * more than once as a throttled bio will go through blk-throtl the
827 * second time when it eventually gets issued. Set it when a bio
828 * is being charged to a tg.
Tejun Heo2a0f61e2013-05-14 13:52:36 -0700829 */
Tejun Heo77ea7332015-08-18 14:55:24 -0700830 if (!(bio->bi_rw & REQ_THROTTLED))
Tejun Heo2a0f61e2013-05-14 13:52:36 -0700831 bio->bi_rw |= REQ_THROTTLED;
Vivek Goyale43473b2010-09-15 17:06:35 -0400832}
833
Tejun Heoc5cc2072013-05-14 13:52:38 -0700834/**
835 * throtl_add_bio_tg - add a bio to the specified throtl_grp
836 * @bio: bio to add
837 * @qn: qnode to use
838 * @tg: the target throtl_grp
839 *
840 * Add @bio to @tg's service_queue using @qn. If @qn is not specified,
841 * tg->qnode_on_self[] is used.
842 */
843static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
844 struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400845{
Tejun Heo73f0d492013-05-14 13:52:35 -0700846 struct throtl_service_queue *sq = &tg->service_queue;
Vivek Goyale43473b2010-09-15 17:06:35 -0400847 bool rw = bio_data_dir(bio);
848
Tejun Heoc5cc2072013-05-14 13:52:38 -0700849 if (!qn)
850 qn = &tg->qnode_on_self[rw];
851
Tejun Heo0e9f4162013-05-14 13:52:35 -0700852 /*
853 * If @tg doesn't currently have any bios queued in the same
854 * direction, queueing @bio can change when @tg should be
855 * dispatched. Mark that @tg was empty. This is automatically
856 * cleaered on the next tg_update_disptime().
857 */
858 if (!sq->nr_queued[rw])
859 tg->flags |= THROTL_TG_WAS_EMPTY;
860
Tejun Heoc5cc2072013-05-14 13:52:38 -0700861 throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
862
Tejun Heo73f0d492013-05-14 13:52:35 -0700863 sq->nr_queued[rw]++;
Tejun Heo77216b02013-05-14 13:52:36 -0700864 throtl_enqueue_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -0400865}
866
Tejun Heo77216b02013-05-14 13:52:36 -0700867static void tg_update_disptime(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400868{
Tejun Heo73f0d492013-05-14 13:52:35 -0700869 struct throtl_service_queue *sq = &tg->service_queue;
Vivek Goyale43473b2010-09-15 17:06:35 -0400870 unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
871 struct bio *bio;
872
Tejun Heoc5cc2072013-05-14 13:52:38 -0700873 if ((bio = throtl_peek_queued(&sq->queued[READ])))
Tejun Heo0f3457f2013-05-14 13:52:32 -0700874 tg_may_dispatch(tg, bio, &read_wait);
Vivek Goyale43473b2010-09-15 17:06:35 -0400875
Tejun Heoc5cc2072013-05-14 13:52:38 -0700876 if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
Tejun Heo0f3457f2013-05-14 13:52:32 -0700877 tg_may_dispatch(tg, bio, &write_wait);
Vivek Goyale43473b2010-09-15 17:06:35 -0400878
879 min_wait = min(read_wait, write_wait);
880 disptime = jiffies + min_wait;
881
Vivek Goyale43473b2010-09-15 17:06:35 -0400882 /* Update dispatch time */
Tejun Heo77216b02013-05-14 13:52:36 -0700883 throtl_dequeue_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -0400884 tg->disptime = disptime;
Tejun Heo77216b02013-05-14 13:52:36 -0700885 throtl_enqueue_tg(tg);
Tejun Heo0e9f4162013-05-14 13:52:35 -0700886
887 /* see throtl_add_bio_tg() */
888 tg->flags &= ~THROTL_TG_WAS_EMPTY;
Vivek Goyale43473b2010-09-15 17:06:35 -0400889}
890
Vivek Goyal32ee5bc2013-05-14 13:52:38 -0700891static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
892 struct throtl_grp *parent_tg, bool rw)
893{
894 if (throtl_slice_used(parent_tg, rw)) {
895 throtl_start_new_slice_with_credit(parent_tg, rw,
896 child_tg->slice_start[rw]);
897 }
898
899}
900
Tejun Heo77216b02013-05-14 13:52:36 -0700901static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
Vivek Goyale43473b2010-09-15 17:06:35 -0400902{
Tejun Heo73f0d492013-05-14 13:52:35 -0700903 struct throtl_service_queue *sq = &tg->service_queue;
Tejun Heo6bc9c2b2013-05-14 13:52:38 -0700904 struct throtl_service_queue *parent_sq = sq->parent_sq;
905 struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
Tejun Heoc5cc2072013-05-14 13:52:38 -0700906 struct throtl_grp *tg_to_put = NULL;
Vivek Goyale43473b2010-09-15 17:06:35 -0400907 struct bio *bio;
908
Tejun Heoc5cc2072013-05-14 13:52:38 -0700909 /*
910 * @bio is being transferred from @tg to @parent_sq. Popping a bio
911 * from @tg may put its reference and @parent_sq might end up
912 * getting released prematurely. Remember the tg to put and put it
913 * after @bio is transferred to @parent_sq.
914 */
915 bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
Tejun Heo73f0d492013-05-14 13:52:35 -0700916 sq->nr_queued[rw]--;
Vivek Goyale43473b2010-09-15 17:06:35 -0400917
918 throtl_charge_bio(tg, bio);
Tejun Heo6bc9c2b2013-05-14 13:52:38 -0700919
920 /*
921 * If our parent is another tg, we just need to transfer @bio to
922 * the parent using throtl_add_bio_tg(). If our parent is
923 * @td->service_queue, @bio is ready to be issued. Put it on its
924 * bio_lists[] and decrease total number queued. The caller is
925 * responsible for issuing these bios.
926 */
927 if (parent_tg) {
Tejun Heoc5cc2072013-05-14 13:52:38 -0700928 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
Vivek Goyal32ee5bc2013-05-14 13:52:38 -0700929 start_parent_slice_with_credit(tg, parent_tg, rw);
Tejun Heo6bc9c2b2013-05-14 13:52:38 -0700930 } else {
Tejun Heoc5cc2072013-05-14 13:52:38 -0700931 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
932 &parent_sq->queued[rw]);
Tejun Heo6bc9c2b2013-05-14 13:52:38 -0700933 BUG_ON(tg->td->nr_queued[rw] <= 0);
934 tg->td->nr_queued[rw]--;
935 }
Vivek Goyale43473b2010-09-15 17:06:35 -0400936
Tejun Heo0f3457f2013-05-14 13:52:32 -0700937 throtl_trim_slice(tg, rw);
Tejun Heo6bc9c2b2013-05-14 13:52:38 -0700938
Tejun Heoc5cc2072013-05-14 13:52:38 -0700939 if (tg_to_put)
940 blkg_put(tg_to_blkg(tg_to_put));
Vivek Goyale43473b2010-09-15 17:06:35 -0400941}
942
Tejun Heo77216b02013-05-14 13:52:36 -0700943static int throtl_dispatch_tg(struct throtl_grp *tg)
Vivek Goyale43473b2010-09-15 17:06:35 -0400944{
Tejun Heo73f0d492013-05-14 13:52:35 -0700945 struct throtl_service_queue *sq = &tg->service_queue;
Vivek Goyale43473b2010-09-15 17:06:35 -0400946 unsigned int nr_reads = 0, nr_writes = 0;
947 unsigned int max_nr_reads = throtl_grp_quantum*3/4;
Vivek Goyalc2f68052010-11-15 19:32:42 +0100948 unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
Vivek Goyale43473b2010-09-15 17:06:35 -0400949 struct bio *bio;
950
951 /* Try to dispatch 75% READS and 25% WRITES */
952
Tejun Heoc5cc2072013-05-14 13:52:38 -0700953 while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
Tejun Heo0f3457f2013-05-14 13:52:32 -0700954 tg_may_dispatch(tg, bio, NULL)) {
Vivek Goyale43473b2010-09-15 17:06:35 -0400955
Tejun Heo77216b02013-05-14 13:52:36 -0700956 tg_dispatch_one_bio(tg, bio_data_dir(bio));
Vivek Goyale43473b2010-09-15 17:06:35 -0400957 nr_reads++;
958
959 if (nr_reads >= max_nr_reads)
960 break;
961 }
962
Tejun Heoc5cc2072013-05-14 13:52:38 -0700963 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
Tejun Heo0f3457f2013-05-14 13:52:32 -0700964 tg_may_dispatch(tg, bio, NULL)) {
Vivek Goyale43473b2010-09-15 17:06:35 -0400965
Tejun Heo77216b02013-05-14 13:52:36 -0700966 tg_dispatch_one_bio(tg, bio_data_dir(bio));
Vivek Goyale43473b2010-09-15 17:06:35 -0400967 nr_writes++;
968
969 if (nr_writes >= max_nr_writes)
970 break;
971 }
972
973 return nr_reads + nr_writes;
974}
975
Tejun Heo651930b2013-05-14 13:52:35 -0700976static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
Vivek Goyale43473b2010-09-15 17:06:35 -0400977{
978 unsigned int nr_disp = 0;
Vivek Goyale43473b2010-09-15 17:06:35 -0400979
980 while (1) {
Tejun Heo73f0d492013-05-14 13:52:35 -0700981 struct throtl_grp *tg = throtl_rb_first(parent_sq);
982 struct throtl_service_queue *sq = &tg->service_queue;
Vivek Goyale43473b2010-09-15 17:06:35 -0400983
984 if (!tg)
985 break;
986
987 if (time_before(jiffies, tg->disptime))
988 break;
989
Tejun Heo77216b02013-05-14 13:52:36 -0700990 throtl_dequeue_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -0400991
Tejun Heo77216b02013-05-14 13:52:36 -0700992 nr_disp += throtl_dispatch_tg(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -0400993
Tejun Heo73f0d492013-05-14 13:52:35 -0700994 if (sq->nr_queued[0] || sq->nr_queued[1])
Tejun Heo77216b02013-05-14 13:52:36 -0700995 tg_update_disptime(tg);
Vivek Goyale43473b2010-09-15 17:06:35 -0400996
997 if (nr_disp >= throtl_quantum)
998 break;
999 }
1000
1001 return nr_disp;
1002}
1003
Tejun Heo6e1a5702013-05-14 13:52:37 -07001004/**
1005 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1006 * @arg: the throtl_service_queue being serviced
1007 *
1008 * This timer is armed when a child throtl_grp with active bio's become
1009 * pending and queued on the service_queue's pending_tree and expires when
1010 * the first child throtl_grp should be dispatched. This function
Tejun Heo2e48a532013-05-14 13:52:38 -07001011 * dispatches bio's from the children throtl_grps to the parent
1012 * service_queue.
1013 *
1014 * If the parent's parent is another throtl_grp, dispatching is propagated
1015 * by either arming its pending_timer or repeating dispatch directly. If
1016 * the top-level service_tree is reached, throtl_data->dispatch_work is
1017 * kicked so that the ready bio's are issued.
Tejun Heo6e1a5702013-05-14 13:52:37 -07001018 */
Tejun Heo69df0ab2013-05-14 13:52:36 -07001019static void throtl_pending_timer_fn(unsigned long arg)
1020{
1021 struct throtl_service_queue *sq = (void *)arg;
Tejun Heo2e48a532013-05-14 13:52:38 -07001022 struct throtl_grp *tg = sq_to_tg(sq);
Tejun Heo69df0ab2013-05-14 13:52:36 -07001023 struct throtl_data *td = sq_to_td(sq);
Tejun Heocb761992013-05-14 13:52:31 -07001024 struct request_queue *q = td->queue;
Tejun Heo2e48a532013-05-14 13:52:38 -07001025 struct throtl_service_queue *parent_sq;
1026 bool dispatched;
Tejun Heo6e1a5702013-05-14 13:52:37 -07001027 int ret;
Vivek Goyale43473b2010-09-15 17:06:35 -04001028
1029 spin_lock_irq(q->queue_lock);
Tejun Heo2e48a532013-05-14 13:52:38 -07001030again:
1031 parent_sq = sq->parent_sq;
1032 dispatched = false;
Vivek Goyale43473b2010-09-15 17:06:35 -04001033
Tejun Heo7f52f982013-05-14 13:52:37 -07001034 while (true) {
1035 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
Tejun Heo2e48a532013-05-14 13:52:38 -07001036 sq->nr_queued[READ] + sq->nr_queued[WRITE],
1037 sq->nr_queued[READ], sq->nr_queued[WRITE]);
Vivek Goyale43473b2010-09-15 17:06:35 -04001038
Tejun Heo7f52f982013-05-14 13:52:37 -07001039 ret = throtl_select_dispatch(sq);
1040 if (ret) {
Tejun Heo7f52f982013-05-14 13:52:37 -07001041 throtl_log(sq, "bios disp=%u", ret);
1042 dispatched = true;
Tejun Heo651930b2013-05-14 13:52:35 -07001043 }
Vivek Goyale43473b2010-09-15 17:06:35 -04001044
Tejun Heo7f52f982013-05-14 13:52:37 -07001045 if (throtl_schedule_next_dispatch(sq, false))
1046 break;
1047
1048 /* this dispatch windows is still open, relax and repeat */
1049 spin_unlock_irq(q->queue_lock);
1050 cpu_relax();
1051 spin_lock_irq(q->queue_lock);
1052 }
Tejun Heo6a525602013-05-14 13:52:32 -07001053
Tejun Heo2e48a532013-05-14 13:52:38 -07001054 if (!dispatched)
1055 goto out_unlock;
Tejun Heo6e1a5702013-05-14 13:52:37 -07001056
Tejun Heo2e48a532013-05-14 13:52:38 -07001057 if (parent_sq) {
1058 /* @parent_sq is another throl_grp, propagate dispatch */
1059 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1060 tg_update_disptime(tg);
1061 if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1062 /* window is already open, repeat dispatching */
1063 sq = parent_sq;
1064 tg = sq_to_tg(sq);
1065 goto again;
1066 }
1067 }
1068 } else {
1069 /* reached the top-level, queue issueing */
1070 queue_work(kthrotld_workqueue, &td->dispatch_work);
1071 }
1072out_unlock:
Tejun Heo6e1a5702013-05-14 13:52:37 -07001073 spin_unlock_irq(q->queue_lock);
1074}
1075
1076/**
1077 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1078 * @work: work item being executed
1079 *
1080 * This function is queued for execution when bio's reach the bio_lists[]
1081 * of throtl_data->service_queue. Those bio's are ready and issued by this
1082 * function.
1083 */
Fabian Frederick8876e142014-04-17 21:41:16 +02001084static void blk_throtl_dispatch_work_fn(struct work_struct *work)
Tejun Heo6e1a5702013-05-14 13:52:37 -07001085{
1086 struct throtl_data *td = container_of(work, struct throtl_data,
1087 dispatch_work);
1088 struct throtl_service_queue *td_sq = &td->service_queue;
1089 struct request_queue *q = td->queue;
1090 struct bio_list bio_list_on_stack;
1091 struct bio *bio;
1092 struct blk_plug plug;
1093 int rw;
1094
1095 bio_list_init(&bio_list_on_stack);
1096
1097 spin_lock_irq(q->queue_lock);
Tejun Heoc5cc2072013-05-14 13:52:38 -07001098 for (rw = READ; rw <= WRITE; rw++)
1099 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1100 bio_list_add(&bio_list_on_stack, bio);
Vivek Goyale43473b2010-09-15 17:06:35 -04001101 spin_unlock_irq(q->queue_lock);
1102
Tejun Heo6e1a5702013-05-14 13:52:37 -07001103 if (!bio_list_empty(&bio_list_on_stack)) {
Vivek Goyal69d60eb2011-03-09 08:27:37 +01001104 blk_start_plug(&plug);
Vivek Goyale43473b2010-09-15 17:06:35 -04001105 while((bio = bio_list_pop(&bio_list_on_stack)))
1106 generic_make_request(bio);
Vivek Goyal69d60eb2011-03-09 08:27:37 +01001107 blk_finish_plug(&plug);
Vivek Goyale43473b2010-09-15 17:06:35 -04001108 }
Vivek Goyale43473b2010-09-15 17:06:35 -04001109}
1110
Tejun Heof95a04a2012-04-16 13:57:26 -07001111static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1112 int off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001113{
Tejun Heof95a04a2012-04-16 13:57:26 -07001114 struct throtl_grp *tg = pd_to_tg(pd);
1115 u64 v = *(u64 *)((void *)tg + off);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001116
Tejun Heoaf133ce2012-04-01 14:38:44 -07001117 if (v == -1)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001118 return 0;
Tejun Heof95a04a2012-04-16 13:57:26 -07001119 return __blkg_prfill_u64(sf, pd, v);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001120}
1121
Tejun Heof95a04a2012-04-16 13:57:26 -07001122static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1123 int off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001124{
Tejun Heof95a04a2012-04-16 13:57:26 -07001125 struct throtl_grp *tg = pd_to_tg(pd);
1126 unsigned int v = *(unsigned int *)((void *)tg + off);
Tejun Heoaf133ce2012-04-01 14:38:44 -07001127
1128 if (v == -1)
1129 return 0;
Tejun Heof95a04a2012-04-16 13:57:26 -07001130 return __blkg_prfill_u64(sf, pd, v);
Tejun Heoaf133ce2012-04-01 14:38:44 -07001131}
1132
Tejun Heo2da8ca82013-12-05 12:28:04 -05001133static int tg_print_conf_u64(struct seq_file *sf, void *v)
Tejun Heoaf133ce2012-04-01 14:38:44 -07001134{
Tejun Heo2da8ca82013-12-05 12:28:04 -05001135 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1136 &blkcg_policy_throtl, seq_cft(sf)->private, false);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001137 return 0;
1138}
1139
Tejun Heo2da8ca82013-12-05 12:28:04 -05001140static int tg_print_conf_uint(struct seq_file *sf, void *v)
Vivek Goyale43473b2010-09-15 17:06:35 -04001141{
Tejun Heo2da8ca82013-12-05 12:28:04 -05001142 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1143 &blkcg_policy_throtl, seq_cft(sf)->private, false);
Tejun Heoaf133ce2012-04-01 14:38:44 -07001144 return 0;
Vivek Goyale43473b2010-09-15 17:06:35 -04001145}
1146
Tejun Heo69948b02015-08-18 14:55:32 -07001147static void tg_conf_updated(struct throtl_grp *tg)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001148{
Tejun Heo69948b02015-08-18 14:55:32 -07001149 struct throtl_service_queue *sq = &tg->service_queue;
Tejun Heo492eb212013-08-08 20:11:25 -04001150 struct cgroup_subsys_state *pos_css;
Tejun Heo69948b02015-08-18 14:55:32 -07001151 struct blkcg_gq *blkg;
Tejun Heoaf133ce2012-04-01 14:38:44 -07001152
Tejun Heofda6f272013-05-14 13:52:36 -07001153 throtl_log(&tg->service_queue,
1154 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1155 tg->bps[READ], tg->bps[WRITE],
1156 tg->iops[READ], tg->iops[WRITE]);
Tejun Heo632b4492013-05-14 13:52:31 -07001157
1158 /*
Tejun Heo693e7512013-05-14 13:52:38 -07001159 * Update has_rules[] flags for the updated tg's subtree. A tg is
1160 * considered to have rules if either the tg itself or any of its
1161 * ancestors has rules. This identifies groups without any
1162 * restrictions in the whole hierarchy and allows them to bypass
1163 * blk-throttle.
1164 */
Tejun Heo69948b02015-08-18 14:55:32 -07001165 blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg))
Tejun Heo693e7512013-05-14 13:52:38 -07001166 tg_update_has_rules(blkg_to_tg(blkg));
1167
1168 /*
Tejun Heo632b4492013-05-14 13:52:31 -07001169 * We're already holding queue_lock and know @tg is valid. Let's
1170 * apply the new config directly.
1171 *
1172 * Restart the slices for both READ and WRITES. It might happen
1173 * that a group's limit are dropped suddenly and we don't want to
1174 * account recently dispatched IO with new low rate.
1175 */
Tejun Heo0f3457f2013-05-14 13:52:32 -07001176 throtl_start_new_slice(tg, 0);
1177 throtl_start_new_slice(tg, 1);
Tejun Heo632b4492013-05-14 13:52:31 -07001178
Tejun Heo5b2c16a2013-05-14 13:52:32 -07001179 if (tg->flags & THROTL_TG_PENDING) {
Tejun Heo77216b02013-05-14 13:52:36 -07001180 tg_update_disptime(tg);
Tejun Heo7f52f982013-05-14 13:52:37 -07001181 throtl_schedule_next_dispatch(sq->parent_sq, true);
Tejun Heo632b4492013-05-14 13:52:31 -07001182 }
Tejun Heo69948b02015-08-18 14:55:32 -07001183}
Tejun Heo60c2bc22012-04-01 14:38:43 -07001184
Tejun Heo69948b02015-08-18 14:55:32 -07001185static ssize_t tg_set_conf(struct kernfs_open_file *of,
1186 char *buf, size_t nbytes, loff_t off, bool is_u64)
1187{
1188 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1189 struct blkg_conf_ctx ctx;
1190 struct throtl_grp *tg;
1191 int ret;
1192 u64 v;
1193
1194 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1195 if (ret)
1196 return ret;
1197
1198 ret = -EINVAL;
1199 if (sscanf(ctx.body, "%llu", &v) != 1)
1200 goto out_finish;
1201 if (!v)
1202 v = -1;
1203
1204 tg = blkg_to_tg(ctx.blkg);
1205
1206 if (is_u64)
1207 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1208 else
1209 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1210
1211 tg_conf_updated(tg);
Tejun Heo36aa9e52015-08-18 14:55:31 -07001212 ret = 0;
1213out_finish:
Tejun Heo60c2bc22012-04-01 14:38:43 -07001214 blkg_conf_finish(&ctx);
Tejun Heo36aa9e52015-08-18 14:55:31 -07001215 return ret ?: nbytes;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001216}
1217
Tejun Heo451af502014-05-13 12:16:21 -04001218static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1219 char *buf, size_t nbytes, loff_t off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001220{
Tejun Heo451af502014-05-13 12:16:21 -04001221 return tg_set_conf(of, buf, nbytes, off, true);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001222}
1223
Tejun Heo451af502014-05-13 12:16:21 -04001224static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1225 char *buf, size_t nbytes, loff_t off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001226{
Tejun Heo451af502014-05-13 12:16:21 -04001227 return tg_set_conf(of, buf, nbytes, off, false);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001228}
1229
Tejun Heo880f50e2015-08-18 14:55:30 -07001230static struct cftype throtl_legacy_files[] = {
Tejun Heo60c2bc22012-04-01 14:38:43 -07001231 {
1232 .name = "throttle.read_bps_device",
Tejun Heoaf133ce2012-04-01 14:38:44 -07001233 .private = offsetof(struct throtl_grp, bps[READ]),
Tejun Heo2da8ca82013-12-05 12:28:04 -05001234 .seq_show = tg_print_conf_u64,
Tejun Heo451af502014-05-13 12:16:21 -04001235 .write = tg_set_conf_u64,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001236 },
1237 {
1238 .name = "throttle.write_bps_device",
Tejun Heoaf133ce2012-04-01 14:38:44 -07001239 .private = offsetof(struct throtl_grp, bps[WRITE]),
Tejun Heo2da8ca82013-12-05 12:28:04 -05001240 .seq_show = tg_print_conf_u64,
Tejun Heo451af502014-05-13 12:16:21 -04001241 .write = tg_set_conf_u64,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001242 },
1243 {
1244 .name = "throttle.read_iops_device",
Tejun Heoaf133ce2012-04-01 14:38:44 -07001245 .private = offsetof(struct throtl_grp, iops[READ]),
Tejun Heo2da8ca82013-12-05 12:28:04 -05001246 .seq_show = tg_print_conf_uint,
Tejun Heo451af502014-05-13 12:16:21 -04001247 .write = tg_set_conf_uint,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001248 },
1249 {
1250 .name = "throttle.write_iops_device",
Tejun Heoaf133ce2012-04-01 14:38:44 -07001251 .private = offsetof(struct throtl_grp, iops[WRITE]),
Tejun Heo2da8ca82013-12-05 12:28:04 -05001252 .seq_show = tg_print_conf_uint,
Tejun Heo451af502014-05-13 12:16:21 -04001253 .write = tg_set_conf_uint,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001254 },
1255 {
1256 .name = "throttle.io_service_bytes",
Tejun Heo77ea7332015-08-18 14:55:24 -07001257 .private = (unsigned long)&blkcg_policy_throtl,
1258 .seq_show = blkg_print_stat_bytes,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001259 },
1260 {
1261 .name = "throttle.io_serviced",
Tejun Heo77ea7332015-08-18 14:55:24 -07001262 .private = (unsigned long)&blkcg_policy_throtl,
1263 .seq_show = blkg_print_stat_ios,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001264 },
1265 { } /* terminate */
1266};
1267
Vivek Goyalda527772011-03-02 19:05:33 -05001268static void throtl_shutdown_wq(struct request_queue *q)
Vivek Goyale43473b2010-09-15 17:06:35 -04001269{
1270 struct throtl_data *td = q->td;
1271
Tejun Heo69df0ab2013-05-14 13:52:36 -07001272 cancel_work_sync(&td->dispatch_work);
Vivek Goyale43473b2010-09-15 17:06:35 -04001273}
1274
Tejun Heo3c798392012-04-16 13:57:25 -07001275static struct blkcg_policy blkcg_policy_throtl = {
Tejun Heo880f50e2015-08-18 14:55:30 -07001276 .legacy_cftypes = throtl_legacy_files,
Tejun Heof9fcc2d2012-04-16 13:57:27 -07001277
Tejun Heo001bea72015-08-18 14:55:11 -07001278 .pd_alloc_fn = throtl_pd_alloc,
Tejun Heof9fcc2d2012-04-16 13:57:27 -07001279 .pd_init_fn = throtl_pd_init,
Tejun Heo693e7512013-05-14 13:52:38 -07001280 .pd_online_fn = throtl_pd_online,
Tejun Heo001bea72015-08-18 14:55:11 -07001281 .pd_free_fn = throtl_pd_free,
Vivek Goyale43473b2010-09-15 17:06:35 -04001282};
1283
Tejun Heoae118892015-08-18 14:55:20 -07001284bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
1285 struct bio *bio)
Vivek Goyale43473b2010-09-15 17:06:35 -04001286{
Tejun Heoc5cc2072013-05-14 13:52:38 -07001287 struct throtl_qnode *qn = NULL;
Tejun Heoae118892015-08-18 14:55:20 -07001288 struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
Tejun Heo73f0d492013-05-14 13:52:35 -07001289 struct throtl_service_queue *sq;
Tejun Heo0e9f4162013-05-14 13:52:35 -07001290 bool rw = bio_data_dir(bio);
Tejun Heobc16a4f2011-10-19 14:33:01 +02001291 bool throttled = false;
Vivek Goyale43473b2010-09-15 17:06:35 -04001292
Tejun Heoae118892015-08-18 14:55:20 -07001293 WARN_ON_ONCE(!rcu_read_lock_held());
1294
Tejun Heo2a0f61e2013-05-14 13:52:36 -07001295 /* see throtl_charge_bio() */
Tejun Heoae118892015-08-18 14:55:20 -07001296 if ((bio->bi_rw & REQ_THROTTLED) || !tg->has_rules[rw])
Tejun Heobc16a4f2011-10-19 14:33:01 +02001297 goto out;
Vivek Goyale43473b2010-09-15 17:06:35 -04001298
1299 spin_lock_irq(q->queue_lock);
Tejun Heoc9589f02015-08-18 14:55:19 -07001300
1301 if (unlikely(blk_queue_bypass(q)))
Tejun Heobc16a4f2011-10-19 14:33:01 +02001302 goto out_unlock;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001303
Tejun Heo73f0d492013-05-14 13:52:35 -07001304 sq = &tg->service_queue;
1305
Tejun Heo9e660ac2013-05-14 13:52:38 -07001306 while (true) {
1307 /* throtl is FIFO - if bios are already queued, should queue */
1308 if (sq->nr_queued[rw])
1309 break;
Vivek Goyalde701c72011-03-07 21:09:32 +01001310
Tejun Heo9e660ac2013-05-14 13:52:38 -07001311 /* if above limits, break to queue */
1312 if (!tg_may_dispatch(tg, bio, NULL))
1313 break;
1314
1315 /* within limits, let's charge and dispatch directly */
Vivek Goyale43473b2010-09-15 17:06:35 -04001316 throtl_charge_bio(tg, bio);
Vivek Goyal04521db2011-03-22 21:54:29 +01001317
1318 /*
1319 * We need to trim slice even when bios are not being queued
1320 * otherwise it might happen that a bio is not queued for
1321 * a long time and slice keeps on extending and trim is not
1322 * called for a long time. Now if limits are reduced suddenly
1323 * we take into account all the IO dispatched so far at new
1324 * low rate and * newly queued IO gets a really long dispatch
1325 * time.
1326 *
1327 * So keep on trimming slice even if bio is not queued.
1328 */
Tejun Heo0f3457f2013-05-14 13:52:32 -07001329 throtl_trim_slice(tg, rw);
Tejun Heo9e660ac2013-05-14 13:52:38 -07001330
1331 /*
1332 * @bio passed through this layer without being throttled.
1333 * Climb up the ladder. If we''re already at the top, it
1334 * can be executed directly.
1335 */
Tejun Heoc5cc2072013-05-14 13:52:38 -07001336 qn = &tg->qnode_on_parent[rw];
Tejun Heo9e660ac2013-05-14 13:52:38 -07001337 sq = sq->parent_sq;
1338 tg = sq_to_tg(sq);
1339 if (!tg)
1340 goto out_unlock;
Vivek Goyale43473b2010-09-15 17:06:35 -04001341 }
1342
Tejun Heo9e660ac2013-05-14 13:52:38 -07001343 /* out-of-limit, queue to @tg */
Tejun Heofda6f272013-05-14 13:52:36 -07001344 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1345 rw == READ ? 'R' : 'W',
Kent Overstreet4f024f32013-10-11 15:44:27 -07001346 tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw],
Tejun Heofda6f272013-05-14 13:52:36 -07001347 tg->io_disp[rw], tg->iops[rw],
1348 sq->nr_queued[READ], sq->nr_queued[WRITE]);
Vivek Goyale43473b2010-09-15 17:06:35 -04001349
Tejun Heo671058f2012-03-05 13:15:29 -08001350 bio_associate_current(bio);
Tejun Heo6bc9c2b2013-05-14 13:52:38 -07001351 tg->td->nr_queued[rw]++;
Tejun Heoc5cc2072013-05-14 13:52:38 -07001352 throtl_add_bio_tg(bio, qn, tg);
Tejun Heobc16a4f2011-10-19 14:33:01 +02001353 throttled = true;
Vivek Goyale43473b2010-09-15 17:06:35 -04001354
Tejun Heo7f52f982013-05-14 13:52:37 -07001355 /*
1356 * Update @tg's dispatch time and force schedule dispatch if @tg
1357 * was empty before @bio. The forced scheduling isn't likely to
1358 * cause undue delay as @bio is likely to be dispatched directly if
1359 * its @tg's disptime is not in the future.
1360 */
Tejun Heo0e9f4162013-05-14 13:52:35 -07001361 if (tg->flags & THROTL_TG_WAS_EMPTY) {
Tejun Heo77216b02013-05-14 13:52:36 -07001362 tg_update_disptime(tg);
Tejun Heo7f52f982013-05-14 13:52:37 -07001363 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
Vivek Goyale43473b2010-09-15 17:06:35 -04001364 }
1365
Tejun Heobc16a4f2011-10-19 14:33:01 +02001366out_unlock:
Vivek Goyale43473b2010-09-15 17:06:35 -04001367 spin_unlock_irq(q->queue_lock);
Tejun Heobc16a4f2011-10-19 14:33:01 +02001368out:
Tejun Heo2a0f61e2013-05-14 13:52:36 -07001369 /*
1370 * As multiple blk-throtls may stack in the same issue path, we
1371 * don't want bios to leave with the flag set. Clear the flag if
1372 * being issued.
1373 */
1374 if (!throttled)
1375 bio->bi_rw &= ~REQ_THROTTLED;
Tejun Heobc16a4f2011-10-19 14:33:01 +02001376 return throttled;
Vivek Goyale43473b2010-09-15 17:06:35 -04001377}
1378
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001379/*
1380 * Dispatch all bios from all children tg's queued on @parent_sq. On
1381 * return, @parent_sq is guaranteed to not have any active children tg's
1382 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1383 */
1384static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1385{
1386 struct throtl_grp *tg;
1387
1388 while ((tg = throtl_rb_first(parent_sq))) {
1389 struct throtl_service_queue *sq = &tg->service_queue;
1390 struct bio *bio;
1391
1392 throtl_dequeue_tg(tg);
1393
Tejun Heoc5cc2072013-05-14 13:52:38 -07001394 while ((bio = throtl_peek_queued(&sq->queued[READ])))
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001395 tg_dispatch_one_bio(tg, bio_data_dir(bio));
Tejun Heoc5cc2072013-05-14 13:52:38 -07001396 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001397 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1398 }
1399}
1400
Tejun Heoc9a929d2011-10-19 14:42:16 +02001401/**
1402 * blk_throtl_drain - drain throttled bios
1403 * @q: request_queue to drain throttled bios for
1404 *
1405 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1406 */
1407void blk_throtl_drain(struct request_queue *q)
1408 __releases(q->queue_lock) __acquires(q->queue_lock)
1409{
1410 struct throtl_data *td = q->td;
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001411 struct blkcg_gq *blkg;
Tejun Heo492eb212013-08-08 20:11:25 -04001412 struct cgroup_subsys_state *pos_css;
Tejun Heoc9a929d2011-10-19 14:42:16 +02001413 struct bio *bio;
Tejun Heo651930b2013-05-14 13:52:35 -07001414 int rw;
Tejun Heoc9a929d2011-10-19 14:42:16 +02001415
Andi Kleen8bcb6c72012-03-30 12:33:28 +02001416 queue_lockdep_assert_held(q);
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001417 rcu_read_lock();
Tejun Heoc9a929d2011-10-19 14:42:16 +02001418
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001419 /*
1420 * Drain each tg while doing post-order walk on the blkg tree, so
1421 * that all bios are propagated to td->service_queue. It'd be
1422 * better to walk service_queue tree directly but blkg walk is
1423 * easier.
1424 */
Tejun Heo492eb212013-08-08 20:11:25 -04001425 blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001426 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
Tejun Heo73f0d492013-05-14 13:52:35 -07001427
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001428 /* finally, transfer bios from top-level tg's into the td */
1429 tg_drain_bios(&td->service_queue);
1430
1431 rcu_read_unlock();
Tejun Heoc9a929d2011-10-19 14:42:16 +02001432 spin_unlock_irq(q->queue_lock);
1433
Tejun Heo2a12f0d2013-05-14 13:52:37 -07001434 /* all bios now should be in td->service_queue, issue them */
Tejun Heo651930b2013-05-14 13:52:35 -07001435 for (rw = READ; rw <= WRITE; rw++)
Tejun Heoc5cc2072013-05-14 13:52:38 -07001436 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1437 NULL)))
Tejun Heo651930b2013-05-14 13:52:35 -07001438 generic_make_request(bio);
Tejun Heoc9a929d2011-10-19 14:42:16 +02001439
1440 spin_lock_irq(q->queue_lock);
1441}
1442
Vivek Goyale43473b2010-09-15 17:06:35 -04001443int blk_throtl_init(struct request_queue *q)
1444{
1445 struct throtl_data *td;
Tejun Heoa2b16932012-04-13 13:11:33 -07001446 int ret;
Vivek Goyale43473b2010-09-15 17:06:35 -04001447
1448 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1449 if (!td)
1450 return -ENOMEM;
1451
Tejun Heo69df0ab2013-05-14 13:52:36 -07001452 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
Tejun Heob2ce2642015-08-18 14:55:13 -07001453 throtl_service_queue_init(&td->service_queue);
Vivek Goyale43473b2010-09-15 17:06:35 -04001454
Tejun Heocd1604f2012-03-05 13:15:06 -08001455 q->td = td;
Vivek Goyal29b12582011-05-19 15:38:24 -04001456 td->queue = q;
Vivek Goyal02977e42010-10-01 14:49:48 +02001457
Tejun Heoa2b16932012-04-13 13:11:33 -07001458 /* activate policy */
Tejun Heo3c798392012-04-16 13:57:25 -07001459 ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
Tejun Heoa2b16932012-04-13 13:11:33 -07001460 if (ret)
Vivek Goyal29b12582011-05-19 15:38:24 -04001461 kfree(td);
Tejun Heoa2b16932012-04-13 13:11:33 -07001462 return ret;
Vivek Goyale43473b2010-09-15 17:06:35 -04001463}
1464
1465void blk_throtl_exit(struct request_queue *q)
1466{
Tejun Heoc875f4d2012-03-05 13:15:22 -08001467 BUG_ON(!q->td);
Vivek Goyalda527772011-03-02 19:05:33 -05001468 throtl_shutdown_wq(q);
Tejun Heo3c798392012-04-16 13:57:25 -07001469 blkcg_deactivate_policy(q, &blkcg_policy_throtl);
Tejun Heoc9a929d2011-10-19 14:42:16 +02001470 kfree(q->td);
Vivek Goyale43473b2010-09-15 17:06:35 -04001471}
1472
1473static int __init throtl_init(void)
1474{
Vivek Goyal450adcb2011-03-01 13:40:54 -05001475 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1476 if (!kthrotld_workqueue)
1477 panic("Failed to create kthrotld\n");
1478
Tejun Heo3c798392012-04-16 13:57:25 -07001479 return blkcg_policy_register(&blkcg_policy_throtl);
Vivek Goyale43473b2010-09-15 17:06:35 -04001480}
1481
1482module_init(throtl_init);