Commit | Line | Data |
---|---|---|
e43473b7 VG |
1 | /* |
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" | |
bc9fcbf9 | 13 | #include "blk.h" |
e43473b7 VG |
14 | |
15 | /* Max dispatch from a group in 1 round */ | |
16 | static int throtl_grp_quantum = 8; | |
17 | ||
18 | /* Total max dispatch from all groups in one round */ | |
19 | static int throtl_quantum = 32; | |
20 | ||
21 | /* Throttling is performed over 100ms slice and after that slice is renewed */ | |
22 | static unsigned long throtl_slice = HZ/10; /* 100 ms */ | |
23 | ||
3c798398 | 24 | static struct blkcg_policy blkcg_policy_throtl; |
0381411e | 25 | |
450adcbe VG |
26 | /* A workqueue to queue throttle related work */ |
27 | static struct workqueue_struct *kthrotld_workqueue; | |
450adcbe | 28 | |
c5cc2070 TH |
29 | /* |
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 | */ | |
52 | struct 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 | ||
c9e0332e | 58 | struct throtl_service_queue { |
77216b04 TH |
59 | struct throtl_service_queue *parent_sq; /* the parent service_queue */ |
60 | ||
73f0d49a TH |
61 | /* |
62 | * Bios queued directly to this service_queue or dispatched from | |
63 | * children throtl_grp's. | |
64 | */ | |
c5cc2070 | 65 | struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */ |
73f0d49a TH |
66 | 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 | */ | |
c9e0332e TH |
72 | 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 */ | |
69df0ab0 | 76 | struct timer_list pending_timer; /* fires on first_pending_disptime */ |
e43473b7 VG |
77 | }; |
78 | ||
5b2c16aa TH |
79 | enum tg_state_flags { |
80 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
0e9f4164 | 81 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ |
5b2c16aa TH |
82 | }; |
83 | ||
e43473b7 VG |
84 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
85 | ||
8a3d2615 TH |
86 | /* Per-cpu group stats */ |
87 | struct 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 | ||
e43473b7 | 94 | struct throtl_grp { |
f95a04af TH |
95 | /* must be the first member */ |
96 | struct blkg_policy_data pd; | |
97 | ||
c9e0332e | 98 | /* active throtl group service_queue member */ |
e43473b7 VG |
99 | struct rb_node rb_node; |
100 | ||
0f3457f6 TH |
101 | /* throtl_data this group belongs to */ |
102 | struct throtl_data *td; | |
103 | ||
49a2f1e3 TH |
104 | /* this group's service queue */ |
105 | struct throtl_service_queue service_queue; | |
106 | ||
c5cc2070 TH |
107 | /* |
108 | * 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 | ||
e43473b7 VG |
118 | /* |
119 | * 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 | ||
e43473b7 VG |
125 | unsigned int flags; |
126 | ||
693e751e TH |
127 | /* are there any throtl rules between this group and td? */ |
128 | bool has_rules[2]; | |
129 | ||
e43473b7 VG |
130 | /* bytes per second rate limits */ |
131 | uint64_t bps[2]; | |
132 | ||
8e89d13f VG |
133 | /* IOPS limits */ |
134 | unsigned int iops[2]; | |
135 | ||
e43473b7 VG |
136 | /* Number of bytes disptached in current slice */ |
137 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
138 | /* Number of bio's dispatched in current slice */ |
139 | unsigned int io_disp[2]; | |
e43473b7 VG |
140 | |
141 | /* When did we start a new slice */ | |
142 | unsigned long slice_start[2]; | |
143 | unsigned long slice_end[2]; | |
fe071437 | 144 | |
8a3d2615 TH |
145 | /* 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; | |
e43473b7 VG |
150 | }; |
151 | ||
152 | struct throtl_data | |
153 | { | |
e43473b7 | 154 | /* service tree for active throtl groups */ |
c9e0332e | 155 | struct throtl_service_queue service_queue; |
e43473b7 | 156 | |
e43473b7 VG |
157 | struct request_queue *queue; |
158 | ||
159 | /* Total Number of queued bios on READ and WRITE lists */ | |
160 | unsigned int nr_queued[2]; | |
161 | ||
162 | /* | |
02977e4a | 163 | * number of total undestroyed groups |
e43473b7 VG |
164 | */ |
165 | unsigned int nr_undestroyed_grps; | |
166 | ||
167 | /* Work for dispatching throttled bios */ | |
69df0ab0 | 168 | struct work_struct dispatch_work; |
e43473b7 VG |
169 | }; |
170 | ||
8a3d2615 TH |
171 | /* list and work item to allocate percpu group stats */ |
172 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
173 | static LIST_HEAD(tg_stats_alloc_list); | |
174 | ||
175 | static void tg_stats_alloc_fn(struct work_struct *); | |
176 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
177 | ||
69df0ab0 TH |
178 | static void throtl_pending_timer_fn(unsigned long arg); |
179 | ||
f95a04af TH |
180 | static 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 | ||
3c798398 | 185 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 186 | { |
f95a04af | 187 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
188 | } |
189 | ||
3c798398 | 190 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 191 | { |
f95a04af | 192 | return pd_to_blkg(&tg->pd); |
0381411e TH |
193 | } |
194 | ||
03d8e111 TH |
195 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) |
196 | { | |
197 | return blkg_to_tg(td->queue->root_blkg); | |
198 | } | |
199 | ||
fda6f272 TH |
200 | /** |
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 | */ | |
207 | static 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 | } | |
214 | ||
215 | /** | |
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 | */ | |
222 | static 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]; \ | |
54e7ed12 | 251 | \ |
fda6f272 TH |
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 | } \ | |
54e7ed12 | 257 | } while (0) |
e43473b7 | 258 | |
8a3d2615 TH |
259 | /* |
260 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
3b07e9ca | 261 | * system_wq once there are some groups on the alloc_list waiting for |
8a3d2615 TH |
262 | * allocation. |
263 | */ | |
264 | static 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 | ||
270 | alloc_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 */ | |
3b07e9ca | 275 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); |
8a3d2615 TH |
276 | 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 | ||
c5cc2070 TH |
296 | static 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 | */ | |
313 | static 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 | */ | |
327 | static 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 | */ | |
354 | static 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 | ||
49a2f1e3 | 379 | /* init a service_queue, assumes the caller zeroed it */ |
77216b04 TH |
380 | static void throtl_service_queue_init(struct throtl_service_queue *sq, |
381 | struct throtl_service_queue *parent_sq) | |
49a2f1e3 | 382 | { |
c5cc2070 TH |
383 | INIT_LIST_HEAD(&sq->queued[0]); |
384 | INIT_LIST_HEAD(&sq->queued[1]); | |
49a2f1e3 | 385 | sq->pending_tree = RB_ROOT; |
77216b04 | 386 | sq->parent_sq = parent_sq; |
69df0ab0 TH |
387 | setup_timer(&sq->pending_timer, throtl_pending_timer_fn, |
388 | (unsigned long)sq); | |
389 | } | |
390 | ||
391 | static void throtl_service_queue_exit(struct throtl_service_queue *sq) | |
392 | { | |
393 | del_timer_sync(&sq->pending_timer); | |
49a2f1e3 TH |
394 | } |
395 | ||
3c798398 | 396 | static void throtl_pd_init(struct blkcg_gq *blkg) |
a29a171e | 397 | { |
0381411e | 398 | struct throtl_grp *tg = blkg_to_tg(blkg); |
77216b04 | 399 | struct throtl_data *td = blkg->q->td; |
ff26eaad | 400 | unsigned long flags; |
c5cc2070 | 401 | int rw; |
cd1604fa | 402 | |
77216b04 | 403 | throtl_service_queue_init(&tg->service_queue, &td->service_queue); |
c5cc2070 TH |
404 | for (rw = READ; rw <= WRITE; rw++) { |
405 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); | |
406 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); | |
407 | } | |
408 | ||
a29a171e | 409 | RB_CLEAR_NODE(&tg->rb_node); |
77216b04 | 410 | tg->td = td; |
a29a171e | 411 | |
e56da7e2 TH |
412 | tg->bps[READ] = -1; |
413 | tg->bps[WRITE] = -1; | |
414 | tg->iops[READ] = -1; | |
415 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
416 | |
417 | /* | |
418 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
419 | * but percpu allocator can't be called from IO path. Queue tg on | |
420 | * tg_stats_alloc_list and allocate from work item. | |
421 | */ | |
ff26eaad | 422 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 423 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); |
3b07e9ca | 424 | schedule_delayed_work(&tg_stats_alloc_work, 0); |
ff26eaad | 425 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
426 | } |
427 | ||
693e751e TH |
428 | /* |
429 | * Set has_rules[] if @tg or any of its parents have limits configured. | |
430 | * This doesn't require walking up to the top of the hierarchy as the | |
431 | * parent's has_rules[] is guaranteed to be correct. | |
432 | */ | |
433 | static void tg_update_has_rules(struct throtl_grp *tg) | |
434 | { | |
435 | struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); | |
436 | int rw; | |
437 | ||
438 | for (rw = READ; rw <= WRITE; rw++) | |
439 | tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) || | |
440 | (tg->bps[rw] != -1 || tg->iops[rw] != -1); | |
441 | } | |
442 | ||
443 | static void throtl_pd_online(struct blkcg_gq *blkg) | |
444 | { | |
445 | /* | |
446 | * We don't want new groups to escape the limits of its ancestors. | |
447 | * Update has_rules[] after a new group is brought online. | |
448 | */ | |
449 | tg_update_has_rules(blkg_to_tg(blkg)); | |
450 | } | |
451 | ||
3c798398 | 452 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
8a3d2615 TH |
453 | { |
454 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
ff26eaad | 455 | unsigned long flags; |
8a3d2615 | 456 | |
ff26eaad | 457 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 458 | list_del_init(&tg->stats_alloc_node); |
ff26eaad | 459 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
460 | |
461 | free_percpu(tg->stats_cpu); | |
69df0ab0 TH |
462 | |
463 | throtl_service_queue_exit(&tg->service_queue); | |
8a3d2615 TH |
464 | } |
465 | ||
3c798398 | 466 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
8a3d2615 TH |
467 | { |
468 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
469 | int cpu; | |
470 | ||
471 | if (tg->stats_cpu == NULL) | |
472 | return; | |
473 | ||
474 | for_each_possible_cpu(cpu) { | |
475 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
476 | ||
477 | blkg_rwstat_reset(&sc->service_bytes); | |
478 | blkg_rwstat_reset(&sc->serviced); | |
479 | } | |
a29a171e VG |
480 | } |
481 | ||
3c798398 TH |
482 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
483 | struct blkcg *blkcg) | |
e43473b7 | 484 | { |
be2c6b19 | 485 | /* |
3c798398 TH |
486 | * This is the common case when there are no blkcgs. Avoid lookup |
487 | * in this case | |
cd1604fa | 488 | */ |
3c798398 | 489 | if (blkcg == &blkcg_root) |
03d8e111 | 490 | return td_root_tg(td); |
e43473b7 | 491 | |
e8989fae | 492 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
493 | } |
494 | ||
cd1604fa | 495 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
3c798398 | 496 | struct blkcg *blkcg) |
e43473b7 | 497 | { |
f469a7b4 | 498 | struct request_queue *q = td->queue; |
cd1604fa | 499 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 500 | |
f469a7b4 | 501 | /* |
3c798398 TH |
502 | * This is the common case when there are no blkcgs. Avoid lookup |
503 | * in this case | |
f469a7b4 | 504 | */ |
3c798398 | 505 | if (blkcg == &blkcg_root) { |
03d8e111 | 506 | tg = td_root_tg(td); |
cd1604fa | 507 | } else { |
3c798398 | 508 | struct blkcg_gq *blkg; |
f469a7b4 | 509 | |
3c96cb32 | 510 | blkg = blkg_lookup_create(blkcg, q); |
f469a7b4 | 511 | |
cd1604fa TH |
512 | /* if %NULL and @q is alive, fall back to root_tg */ |
513 | if (!IS_ERR(blkg)) | |
0381411e | 514 | tg = blkg_to_tg(blkg); |
3f3299d5 | 515 | else if (!blk_queue_dying(q)) |
03d8e111 | 516 | tg = td_root_tg(td); |
f469a7b4 VG |
517 | } |
518 | ||
e43473b7 VG |
519 | return tg; |
520 | } | |
521 | ||
0049af73 TH |
522 | static struct throtl_grp * |
523 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
524 | { |
525 | /* Service tree is empty */ | |
0049af73 | 526 | if (!parent_sq->nr_pending) |
e43473b7 VG |
527 | return NULL; |
528 | ||
0049af73 TH |
529 | if (!parent_sq->first_pending) |
530 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
e43473b7 | 531 | |
0049af73 TH |
532 | if (parent_sq->first_pending) |
533 | return rb_entry_tg(parent_sq->first_pending); | |
e43473b7 VG |
534 | |
535 | return NULL; | |
536 | } | |
537 | ||
538 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
539 | { | |
540 | rb_erase(n, root); | |
541 | RB_CLEAR_NODE(n); | |
542 | } | |
543 | ||
0049af73 TH |
544 | static void throtl_rb_erase(struct rb_node *n, |
545 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 546 | { |
0049af73 TH |
547 | if (parent_sq->first_pending == n) |
548 | parent_sq->first_pending = NULL; | |
549 | rb_erase_init(n, &parent_sq->pending_tree); | |
550 | --parent_sq->nr_pending; | |
e43473b7 VG |
551 | } |
552 | ||
0049af73 | 553 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
554 | { |
555 | struct throtl_grp *tg; | |
556 | ||
0049af73 | 557 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
558 | if (!tg) |
559 | return; | |
560 | ||
0049af73 | 561 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
562 | } |
563 | ||
77216b04 | 564 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 565 | { |
77216b04 | 566 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
0049af73 | 567 | struct rb_node **node = &parent_sq->pending_tree.rb_node; |
e43473b7 VG |
568 | struct rb_node *parent = NULL; |
569 | struct throtl_grp *__tg; | |
570 | unsigned long key = tg->disptime; | |
571 | int left = 1; | |
572 | ||
573 | while (*node != NULL) { | |
574 | parent = *node; | |
575 | __tg = rb_entry_tg(parent); | |
576 | ||
577 | if (time_before(key, __tg->disptime)) | |
578 | node = &parent->rb_left; | |
579 | else { | |
580 | node = &parent->rb_right; | |
581 | left = 0; | |
582 | } | |
583 | } | |
584 | ||
585 | if (left) | |
0049af73 | 586 | parent_sq->first_pending = &tg->rb_node; |
e43473b7 VG |
587 | |
588 | rb_link_node(&tg->rb_node, parent, node); | |
0049af73 | 589 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); |
e43473b7 VG |
590 | } |
591 | ||
77216b04 | 592 | static void __throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 593 | { |
77216b04 | 594 | tg_service_queue_add(tg); |
5b2c16aa | 595 | tg->flags |= THROTL_TG_PENDING; |
77216b04 | 596 | tg->service_queue.parent_sq->nr_pending++; |
e43473b7 VG |
597 | } |
598 | ||
77216b04 | 599 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 600 | { |
5b2c16aa | 601 | if (!(tg->flags & THROTL_TG_PENDING)) |
77216b04 | 602 | __throtl_enqueue_tg(tg); |
e43473b7 VG |
603 | } |
604 | ||
77216b04 | 605 | static void __throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 606 | { |
77216b04 | 607 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
5b2c16aa | 608 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
609 | } |
610 | ||
77216b04 | 611 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 612 | { |
5b2c16aa | 613 | if (tg->flags & THROTL_TG_PENDING) |
77216b04 | 614 | __throtl_dequeue_tg(tg); |
e43473b7 VG |
615 | } |
616 | ||
a9131a27 | 617 | /* Call with queue lock held */ |
69df0ab0 TH |
618 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, |
619 | unsigned long expires) | |
a9131a27 | 620 | { |
69df0ab0 TH |
621 | mod_timer(&sq->pending_timer, expires); |
622 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
623 | expires - jiffies, jiffies); | |
a9131a27 TH |
624 | } |
625 | ||
7f52f98c TH |
626 | /** |
627 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle | |
628 | * @sq: the service_queue to schedule dispatch for | |
629 | * @force: force scheduling | |
630 | * | |
631 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the | |
632 | * dispatch time of the first pending child. Returns %true if either timer | |
633 | * is armed or there's no pending child left. %false if the current | |
634 | * dispatch window is still open and the caller should continue | |
635 | * dispatching. | |
636 | * | |
637 | * If @force is %true, the dispatch timer is always scheduled and this | |
638 | * function is guaranteed to return %true. This is to be used when the | |
639 | * caller can't dispatch itself and needs to invoke pending_timer | |
640 | * unconditionally. Note that forced scheduling is likely to induce short | |
641 | * delay before dispatch starts even if @sq->first_pending_disptime is not | |
642 | * in the future and thus shouldn't be used in hot paths. | |
643 | */ | |
644 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
645 | bool force) | |
e43473b7 | 646 | { |
6a525600 | 647 | /* any pending children left? */ |
c9e0332e | 648 | if (!sq->nr_pending) |
7f52f98c | 649 | return true; |
e43473b7 | 650 | |
c9e0332e | 651 | update_min_dispatch_time(sq); |
e43473b7 | 652 | |
69df0ab0 | 653 | /* is the next dispatch time in the future? */ |
7f52f98c | 654 | if (force || time_after(sq->first_pending_disptime, jiffies)) { |
69df0ab0 | 655 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); |
7f52f98c | 656 | return true; |
69df0ab0 TH |
657 | } |
658 | ||
7f52f98c TH |
659 | /* tell the caller to continue dispatching */ |
660 | return false; | |
e43473b7 VG |
661 | } |
662 | ||
32ee5bc4 VG |
663 | static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, |
664 | bool rw, unsigned long start) | |
665 | { | |
666 | tg->bytes_disp[rw] = 0; | |
667 | tg->io_disp[rw] = 0; | |
668 | ||
669 | /* | |
670 | * Previous slice has expired. We must have trimmed it after last | |
671 | * bio dispatch. That means since start of last slice, we never used | |
672 | * that bandwidth. Do try to make use of that bandwidth while giving | |
673 | * credit. | |
674 | */ | |
675 | if (time_after_eq(start, tg->slice_start[rw])) | |
676 | tg->slice_start[rw] = start; | |
677 | ||
678 | tg->slice_end[rw] = jiffies + throtl_slice; | |
679 | throtl_log(&tg->service_queue, | |
680 | "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", | |
681 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
682 | tg->slice_end[rw], jiffies); | |
683 | } | |
684 | ||
0f3457f6 | 685 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
686 | { |
687 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 688 | tg->io_disp[rw] = 0; |
e43473b7 VG |
689 | tg->slice_start[rw] = jiffies; |
690 | tg->slice_end[rw] = jiffies + throtl_slice; | |
fda6f272 TH |
691 | throtl_log(&tg->service_queue, |
692 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
693 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
694 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
695 | } |
696 | ||
0f3457f6 TH |
697 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
698 | unsigned long jiffy_end) | |
d1ae8ffd VG |
699 | { |
700 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
701 | } | |
702 | ||
0f3457f6 TH |
703 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
704 | unsigned long jiffy_end) | |
e43473b7 VG |
705 | { |
706 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
fda6f272 TH |
707 | throtl_log(&tg->service_queue, |
708 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
709 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
710 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
711 | } |
712 | ||
713 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 714 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
715 | { |
716 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
717 | return 0; | |
718 | ||
719 | return 1; | |
720 | } | |
721 | ||
722 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 723 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 724 | { |
3aad5d3e VG |
725 | unsigned long nr_slices, time_elapsed, io_trim; |
726 | u64 bytes_trim, tmp; | |
e43473b7 VG |
727 | |
728 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
729 | ||
730 | /* | |
731 | * If bps are unlimited (-1), then time slice don't get | |
732 | * renewed. Don't try to trim the slice if slice is used. A new | |
733 | * slice will start when appropriate. | |
734 | */ | |
0f3457f6 | 735 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
736 | return; |
737 | ||
d1ae8ffd VG |
738 | /* |
739 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
740 | * that initially cgroup limit was very low resulting in high | |
741 | * slice_end, but later limit was bumped up and bio was dispached | |
742 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
743 | * is bad because it does not allow new slice to start. | |
744 | */ | |
745 | ||
0f3457f6 | 746 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); |
d1ae8ffd | 747 | |
e43473b7 VG |
748 | time_elapsed = jiffies - tg->slice_start[rw]; |
749 | ||
750 | nr_slices = time_elapsed / throtl_slice; | |
751 | ||
752 | if (!nr_slices) | |
753 | return; | |
3aad5d3e VG |
754 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
755 | do_div(tmp, HZ); | |
756 | bytes_trim = tmp; | |
e43473b7 | 757 | |
8e89d13f | 758 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 759 | |
8e89d13f | 760 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
761 | return; |
762 | ||
763 | if (tg->bytes_disp[rw] >= bytes_trim) | |
764 | tg->bytes_disp[rw] -= bytes_trim; | |
765 | else | |
766 | tg->bytes_disp[rw] = 0; | |
767 | ||
8e89d13f VG |
768 | if (tg->io_disp[rw] >= io_trim) |
769 | tg->io_disp[rw] -= io_trim; | |
770 | else | |
771 | tg->io_disp[rw] = 0; | |
772 | ||
e43473b7 VG |
773 | tg->slice_start[rw] += nr_slices * throtl_slice; |
774 | ||
fda6f272 TH |
775 | throtl_log(&tg->service_queue, |
776 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
777 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
778 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
e43473b7 VG |
779 | } |
780 | ||
0f3457f6 TH |
781 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
782 | unsigned long *wait) | |
e43473b7 VG |
783 | { |
784 | bool rw = bio_data_dir(bio); | |
8e89d13f | 785 | unsigned int io_allowed; |
e43473b7 | 786 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 787 | u64 tmp; |
e43473b7 | 788 | |
8e89d13f | 789 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 790 | |
8e89d13f VG |
791 | /* Slice has just started. Consider one slice interval */ |
792 | if (!jiffy_elapsed) | |
793 | jiffy_elapsed_rnd = throtl_slice; | |
794 | ||
795 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
796 | ||
c49c06e4 VG |
797 | /* |
798 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
799 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
800 | * will allow dispatch after 1 second and after that slice should | |
801 | * have been trimmed. | |
802 | */ | |
803 | ||
804 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
805 | do_div(tmp, HZ); | |
806 | ||
807 | if (tmp > UINT_MAX) | |
808 | io_allowed = UINT_MAX; | |
809 | else | |
810 | io_allowed = tmp; | |
8e89d13f VG |
811 | |
812 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
813 | if (wait) |
814 | *wait = 0; | |
815 | return 1; | |
816 | } | |
817 | ||
8e89d13f VG |
818 | /* Calc approx time to dispatch */ |
819 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
820 | ||
821 | if (jiffy_wait > jiffy_elapsed) | |
822 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
823 | else | |
824 | jiffy_wait = 1; | |
825 | ||
826 | if (wait) | |
827 | *wait = jiffy_wait; | |
828 | return 0; | |
829 | } | |
830 | ||
0f3457f6 TH |
831 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
832 | unsigned long *wait) | |
8e89d13f VG |
833 | { |
834 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 835 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 836 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
837 | |
838 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
839 | ||
840 | /* Slice has just started. Consider one slice interval */ | |
841 | if (!jiffy_elapsed) | |
842 | jiffy_elapsed_rnd = throtl_slice; | |
843 | ||
844 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
845 | ||
5e901a2b VG |
846 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
847 | do_div(tmp, HZ); | |
3aad5d3e | 848 | bytes_allowed = tmp; |
e43473b7 VG |
849 | |
850 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
851 | if (wait) | |
852 | *wait = 0; | |
853 | return 1; | |
854 | } | |
855 | ||
856 | /* Calc approx time to dispatch */ | |
857 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
858 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
859 | ||
860 | if (!jiffy_wait) | |
861 | jiffy_wait = 1; | |
862 | ||
863 | /* | |
864 | * This wait time is without taking into consideration the rounding | |
865 | * up we did. Add that time also. | |
866 | */ | |
867 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
868 | if (wait) |
869 | *wait = jiffy_wait; | |
8e89d13f VG |
870 | return 0; |
871 | } | |
872 | ||
873 | /* | |
874 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
875 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
876 | */ | |
0f3457f6 TH |
877 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
878 | unsigned long *wait) | |
8e89d13f VG |
879 | { |
880 | bool rw = bio_data_dir(bio); | |
881 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
882 | ||
883 | /* | |
884 | * Currently whole state machine of group depends on first bio | |
885 | * queued in the group bio list. So one should not be calling | |
886 | * this function with a different bio if there are other bios | |
887 | * queued. | |
888 | */ | |
73f0d49a | 889 | BUG_ON(tg->service_queue.nr_queued[rw] && |
c5cc2070 | 890 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); |
e43473b7 | 891 | |
8e89d13f VG |
892 | /* If tg->bps = -1, then BW is unlimited */ |
893 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
894 | if (wait) | |
895 | *wait = 0; | |
896 | return 1; | |
897 | } | |
898 | ||
899 | /* | |
900 | * If previous slice expired, start a new one otherwise renew/extend | |
901 | * existing slice to make sure it is at least throtl_slice interval | |
902 | * long since now. | |
903 | */ | |
0f3457f6 TH |
904 | if (throtl_slice_used(tg, rw)) |
905 | throtl_start_new_slice(tg, rw); | |
8e89d13f VG |
906 | else { |
907 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
0f3457f6 | 908 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); |
8e89d13f VG |
909 | } |
910 | ||
0f3457f6 TH |
911 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
912 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
913 | if (wait) |
914 | *wait = 0; | |
915 | return 1; | |
916 | } | |
917 | ||
918 | max_wait = max(bps_wait, iops_wait); | |
919 | ||
920 | if (wait) | |
921 | *wait = max_wait; | |
922 | ||
923 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 924 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
925 | |
926 | return 0; | |
927 | } | |
928 | ||
3c798398 | 929 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
930 | int rw) |
931 | { | |
8a3d2615 TH |
932 | struct throtl_grp *tg = blkg_to_tg(blkg); |
933 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
934 | unsigned long flags; |
935 | ||
936 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 937 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
938 | return; |
939 | ||
940 | /* | |
941 | * Disabling interrupts to provide mutual exclusion between two | |
942 | * writes on same cpu. It probably is not needed for 64bit. Not | |
943 | * optimizing that case yet. | |
944 | */ | |
945 | local_irq_save(flags); | |
946 | ||
8a3d2615 | 947 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 948 | |
629ed0b1 TH |
949 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
950 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
951 | ||
952 | local_irq_restore(flags); | |
953 | } | |
954 | ||
e43473b7 VG |
955 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
956 | { | |
957 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
958 | |
959 | /* Charge the bio to the group */ | |
960 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 961 | tg->io_disp[rw]++; |
e43473b7 | 962 | |
2a0f61e6 TH |
963 | /* |
964 | * REQ_THROTTLED is used to prevent the same bio to be throttled | |
965 | * more than once as a throttled bio will go through blk-throtl the | |
966 | * second time when it eventually gets issued. Set it when a bio | |
967 | * is being charged to a tg. | |
968 | * | |
969 | * Dispatch stats aren't recursive and each @bio should only be | |
970 | * accounted by the @tg it was originally associated with. Let's | |
971 | * update the stats when setting REQ_THROTTLED for the first time | |
972 | * which is guaranteed to be for the @bio's original tg. | |
973 | */ | |
974 | if (!(bio->bi_rw & REQ_THROTTLED)) { | |
975 | bio->bi_rw |= REQ_THROTTLED; | |
976 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, | |
977 | bio->bi_rw); | |
978 | } | |
e43473b7 VG |
979 | } |
980 | ||
c5cc2070 TH |
981 | /** |
982 | * throtl_add_bio_tg - add a bio to the specified throtl_grp | |
983 | * @bio: bio to add | |
984 | * @qn: qnode to use | |
985 | * @tg: the target throtl_grp | |
986 | * | |
987 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, | |
988 | * tg->qnode_on_self[] is used. | |
989 | */ | |
990 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, | |
991 | struct throtl_grp *tg) | |
e43473b7 | 992 | { |
73f0d49a | 993 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
994 | bool rw = bio_data_dir(bio); |
995 | ||
c5cc2070 TH |
996 | if (!qn) |
997 | qn = &tg->qnode_on_self[rw]; | |
998 | ||
0e9f4164 TH |
999 | /* |
1000 | * If @tg doesn't currently have any bios queued in the same | |
1001 | * direction, queueing @bio can change when @tg should be | |
1002 | * dispatched. Mark that @tg was empty. This is automatically | |
1003 | * cleaered on the next tg_update_disptime(). | |
1004 | */ | |
1005 | if (!sq->nr_queued[rw]) | |
1006 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
1007 | ||
c5cc2070 TH |
1008 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); |
1009 | ||
73f0d49a | 1010 | sq->nr_queued[rw]++; |
77216b04 | 1011 | throtl_enqueue_tg(tg); |
e43473b7 VG |
1012 | } |
1013 | ||
77216b04 | 1014 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 1015 | { |
73f0d49a | 1016 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1017 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
1018 | struct bio *bio; | |
1019 | ||
c5cc2070 | 1020 | if ((bio = throtl_peek_queued(&sq->queued[READ]))) |
0f3457f6 | 1021 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 1022 | |
c5cc2070 | 1023 | if ((bio = throtl_peek_queued(&sq->queued[WRITE]))) |
0f3457f6 | 1024 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
1025 | |
1026 | min_wait = min(read_wait, write_wait); | |
1027 | disptime = jiffies + min_wait; | |
1028 | ||
e43473b7 | 1029 | /* Update dispatch time */ |
77216b04 | 1030 | throtl_dequeue_tg(tg); |
e43473b7 | 1031 | tg->disptime = disptime; |
77216b04 | 1032 | throtl_enqueue_tg(tg); |
0e9f4164 TH |
1033 | |
1034 | /* see throtl_add_bio_tg() */ | |
1035 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
1036 | } |
1037 | ||
32ee5bc4 VG |
1038 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, |
1039 | struct throtl_grp *parent_tg, bool rw) | |
1040 | { | |
1041 | if (throtl_slice_used(parent_tg, rw)) { | |
1042 | throtl_start_new_slice_with_credit(parent_tg, rw, | |
1043 | child_tg->slice_start[rw]); | |
1044 | } | |
1045 | ||
1046 | } | |
1047 | ||
77216b04 | 1048 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 1049 | { |
73f0d49a | 1050 | struct throtl_service_queue *sq = &tg->service_queue; |
6bc9c2b4 TH |
1051 | struct throtl_service_queue *parent_sq = sq->parent_sq; |
1052 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
c5cc2070 | 1053 | struct throtl_grp *tg_to_put = NULL; |
e43473b7 VG |
1054 | struct bio *bio; |
1055 | ||
c5cc2070 TH |
1056 | /* |
1057 | * @bio is being transferred from @tg to @parent_sq. Popping a bio | |
1058 | * from @tg may put its reference and @parent_sq might end up | |
1059 | * getting released prematurely. Remember the tg to put and put it | |
1060 | * after @bio is transferred to @parent_sq. | |
1061 | */ | |
1062 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); | |
73f0d49a | 1063 | sq->nr_queued[rw]--; |
e43473b7 VG |
1064 | |
1065 | throtl_charge_bio(tg, bio); | |
6bc9c2b4 TH |
1066 | |
1067 | /* | |
1068 | * If our parent is another tg, we just need to transfer @bio to | |
1069 | * the parent using throtl_add_bio_tg(). If our parent is | |
1070 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
1071 | * bio_lists[] and decrease total number queued. The caller is | |
1072 | * responsible for issuing these bios. | |
1073 | */ | |
1074 | if (parent_tg) { | |
c5cc2070 | 1075 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); |
32ee5bc4 | 1076 | start_parent_slice_with_credit(tg, parent_tg, rw); |
6bc9c2b4 | 1077 | } else { |
c5cc2070 TH |
1078 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], |
1079 | &parent_sq->queued[rw]); | |
6bc9c2b4 TH |
1080 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
1081 | tg->td->nr_queued[rw]--; | |
1082 | } | |
e43473b7 | 1083 | |
0f3457f6 | 1084 | throtl_trim_slice(tg, rw); |
6bc9c2b4 | 1085 | |
c5cc2070 TH |
1086 | if (tg_to_put) |
1087 | blkg_put(tg_to_blkg(tg_to_put)); | |
e43473b7 VG |
1088 | } |
1089 | ||
77216b04 | 1090 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 1091 | { |
73f0d49a | 1092 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1093 | unsigned int nr_reads = 0, nr_writes = 0; |
1094 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 1095 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
1096 | struct bio *bio; |
1097 | ||
1098 | /* Try to dispatch 75% READS and 25% WRITES */ | |
1099 | ||
c5cc2070 | 1100 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && |
0f3457f6 | 1101 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1102 | |
77216b04 | 1103 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1104 | nr_reads++; |
1105 | ||
1106 | if (nr_reads >= max_nr_reads) | |
1107 | break; | |
1108 | } | |
1109 | ||
c5cc2070 | 1110 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && |
0f3457f6 | 1111 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1112 | |
77216b04 | 1113 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1114 | nr_writes++; |
1115 | ||
1116 | if (nr_writes >= max_nr_writes) | |
1117 | break; | |
1118 | } | |
1119 | ||
1120 | return nr_reads + nr_writes; | |
1121 | } | |
1122 | ||
651930bc | 1123 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
1124 | { |
1125 | unsigned int nr_disp = 0; | |
e43473b7 VG |
1126 | |
1127 | while (1) { | |
73f0d49a TH |
1128 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
1129 | struct throtl_service_queue *sq = &tg->service_queue; | |
e43473b7 VG |
1130 | |
1131 | if (!tg) | |
1132 | break; | |
1133 | ||
1134 | if (time_before(jiffies, tg->disptime)) | |
1135 | break; | |
1136 | ||
77216b04 | 1137 | throtl_dequeue_tg(tg); |
e43473b7 | 1138 | |
77216b04 | 1139 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 1140 | |
73f0d49a | 1141 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
77216b04 | 1142 | tg_update_disptime(tg); |
e43473b7 VG |
1143 | |
1144 | if (nr_disp >= throtl_quantum) | |
1145 | break; | |
1146 | } | |
1147 | ||
1148 | return nr_disp; | |
1149 | } | |
1150 | ||
6e1a5704 TH |
1151 | /** |
1152 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
1153 | * @arg: the throtl_service_queue being serviced | |
1154 | * | |
1155 | * This timer is armed when a child throtl_grp with active bio's become | |
1156 | * pending and queued on the service_queue's pending_tree and expires when | |
1157 | * the first child throtl_grp should be dispatched. This function | |
2e48a530 TH |
1158 | * dispatches bio's from the children throtl_grps to the parent |
1159 | * service_queue. | |
1160 | * | |
1161 | * If the parent's parent is another throtl_grp, dispatching is propagated | |
1162 | * by either arming its pending_timer or repeating dispatch directly. If | |
1163 | * the top-level service_tree is reached, throtl_data->dispatch_work is | |
1164 | * kicked so that the ready bio's are issued. | |
6e1a5704 | 1165 | */ |
69df0ab0 TH |
1166 | static void throtl_pending_timer_fn(unsigned long arg) |
1167 | { | |
1168 | struct throtl_service_queue *sq = (void *)arg; | |
2e48a530 | 1169 | struct throtl_grp *tg = sq_to_tg(sq); |
69df0ab0 | 1170 | struct throtl_data *td = sq_to_td(sq); |
cb76199c | 1171 | struct request_queue *q = td->queue; |
2e48a530 TH |
1172 | struct throtl_service_queue *parent_sq; |
1173 | bool dispatched; | |
6e1a5704 | 1174 | int ret; |
e43473b7 VG |
1175 | |
1176 | spin_lock_irq(q->queue_lock); | |
2e48a530 TH |
1177 | again: |
1178 | parent_sq = sq->parent_sq; | |
1179 | dispatched = false; | |
e43473b7 | 1180 | |
7f52f98c TH |
1181 | while (true) { |
1182 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
2e48a530 TH |
1183 | sq->nr_queued[READ] + sq->nr_queued[WRITE], |
1184 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
7f52f98c TH |
1185 | |
1186 | ret = throtl_select_dispatch(sq); | |
1187 | if (ret) { | |
7f52f98c TH |
1188 | throtl_log(sq, "bios disp=%u", ret); |
1189 | dispatched = true; | |
1190 | } | |
e43473b7 | 1191 | |
7f52f98c TH |
1192 | if (throtl_schedule_next_dispatch(sq, false)) |
1193 | break; | |
e43473b7 | 1194 | |
7f52f98c TH |
1195 | /* this dispatch windows is still open, relax and repeat */ |
1196 | spin_unlock_irq(q->queue_lock); | |
1197 | cpu_relax(); | |
1198 | spin_lock_irq(q->queue_lock); | |
651930bc | 1199 | } |
e43473b7 | 1200 | |
2e48a530 TH |
1201 | if (!dispatched) |
1202 | goto out_unlock; | |
6e1a5704 | 1203 | |
2e48a530 TH |
1204 | if (parent_sq) { |
1205 | /* @parent_sq is another throl_grp, propagate dispatch */ | |
1206 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1207 | tg_update_disptime(tg); | |
1208 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { | |
1209 | /* window is already open, repeat dispatching */ | |
1210 | sq = parent_sq; | |
1211 | tg = sq_to_tg(sq); | |
1212 | goto again; | |
1213 | } | |
1214 | } | |
1215 | } else { | |
1216 | /* reached the top-level, queue issueing */ | |
1217 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1218 | } | |
1219 | out_unlock: | |
e43473b7 | 1220 | spin_unlock_irq(q->queue_lock); |
6e1a5704 | 1221 | } |
e43473b7 | 1222 | |
6e1a5704 TH |
1223 | /** |
1224 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
1225 | * @work: work item being executed | |
1226 | * | |
1227 | * This function is queued for execution when bio's reach the bio_lists[] | |
1228 | * of throtl_data->service_queue. Those bio's are ready and issued by this | |
1229 | * function. | |
1230 | */ | |
1231 | void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
1232 | { | |
1233 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1234 | dispatch_work); | |
1235 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1236 | struct request_queue *q = td->queue; | |
1237 | struct bio_list bio_list_on_stack; | |
1238 | struct bio *bio; | |
1239 | struct blk_plug plug; | |
1240 | int rw; | |
1241 | ||
1242 | bio_list_init(&bio_list_on_stack); | |
1243 | ||
1244 | spin_lock_irq(q->queue_lock); | |
c5cc2070 TH |
1245 | for (rw = READ; rw <= WRITE; rw++) |
1246 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) | |
1247 | bio_list_add(&bio_list_on_stack, bio); | |
6e1a5704 TH |
1248 | spin_unlock_irq(q->queue_lock); |
1249 | ||
1250 | if (!bio_list_empty(&bio_list_on_stack)) { | |
69d60eb9 | 1251 | blk_start_plug(&plug); |
e43473b7 VG |
1252 | while((bio = bio_list_pop(&bio_list_on_stack))) |
1253 | generic_make_request(bio); | |
69d60eb9 | 1254 | blk_finish_plug(&plug); |
e43473b7 | 1255 | } |
e43473b7 VG |
1256 | } |
1257 | ||
f95a04af TH |
1258 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
1259 | struct blkg_policy_data *pd, int off) | |
41b38b6d | 1260 | { |
f95a04af | 1261 | struct throtl_grp *tg = pd_to_tg(pd); |
41b38b6d TH |
1262 | struct blkg_rwstat rwstat = { }, tmp; |
1263 | int i, cpu; | |
1264 | ||
1265 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 1266 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
1267 | |
1268 | tmp = blkg_rwstat_read((void *)sc + off); | |
1269 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
1270 | rwstat.cnt[i] += tmp.cnt[i]; | |
1271 | } | |
1272 | ||
f95a04af | 1273 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
41b38b6d TH |
1274 | } |
1275 | ||
8a3d2615 TH |
1276 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
1277 | struct seq_file *sf) | |
41b38b6d | 1278 | { |
3c798398 | 1279 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
41b38b6d | 1280 | |
3c798398 | 1281 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 1282 | cft->private, true); |
41b38b6d TH |
1283 | return 0; |
1284 | } | |
1285 | ||
f95a04af TH |
1286 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
1287 | int off) | |
60c2bc2d | 1288 | { |
f95a04af TH |
1289 | struct throtl_grp *tg = pd_to_tg(pd); |
1290 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 1291 | |
af133ceb | 1292 | if (v == -1) |
60c2bc2d | 1293 | return 0; |
f95a04af | 1294 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
1295 | } |
1296 | ||
f95a04af TH |
1297 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
1298 | int off) | |
e43473b7 | 1299 | { |
f95a04af TH |
1300 | struct throtl_grp *tg = pd_to_tg(pd); |
1301 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 1302 | |
af133ceb TH |
1303 | if (v == -1) |
1304 | return 0; | |
f95a04af | 1305 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
1306 | } |
1307 | ||
af133ceb TH |
1308 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1309 | struct seq_file *sf) | |
8e89d13f | 1310 | { |
3c798398 TH |
1311 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, |
1312 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 1313 | return 0; |
8e89d13f VG |
1314 | } |
1315 | ||
af133ceb TH |
1316 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1317 | struct seq_file *sf) | |
8e89d13f | 1318 | { |
3c798398 TH |
1319 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, |
1320 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 1321 | return 0; |
60c2bc2d TH |
1322 | } |
1323 | ||
af133ceb TH |
1324 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
1325 | bool is_u64) | |
60c2bc2d | 1326 | { |
3c798398 | 1327 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
60c2bc2d | 1328 | struct blkg_conf_ctx ctx; |
af133ceb | 1329 | struct throtl_grp *tg; |
69df0ab0 | 1330 | struct throtl_service_queue *sq; |
693e751e TH |
1331 | struct blkcg_gq *blkg; |
1332 | struct cgroup *pos_cgrp; | |
60c2bc2d TH |
1333 | int ret; |
1334 | ||
3c798398 | 1335 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
1336 | if (ret) |
1337 | return ret; | |
1338 | ||
af133ceb | 1339 | tg = blkg_to_tg(ctx.blkg); |
69df0ab0 | 1340 | sq = &tg->service_queue; |
af133ceb | 1341 | |
a2b1693b TH |
1342 | if (!ctx.v) |
1343 | ctx.v = -1; | |
af133ceb | 1344 | |
a2b1693b TH |
1345 | if (is_u64) |
1346 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
1347 | else | |
1348 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 1349 | |
fda6f272 TH |
1350 | throtl_log(&tg->service_queue, |
1351 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
1352 | tg->bps[READ], tg->bps[WRITE], | |
1353 | tg->iops[READ], tg->iops[WRITE]); | |
632b4493 | 1354 | |
693e751e TH |
1355 | /* |
1356 | * Update has_rules[] flags for the updated tg's subtree. A tg is | |
1357 | * considered to have rules if either the tg itself or any of its | |
1358 | * ancestors has rules. This identifies groups without any | |
1359 | * restrictions in the whole hierarchy and allows them to bypass | |
1360 | * blk-throttle. | |
1361 | */ | |
1362 | tg_update_has_rules(tg); | |
1363 | blkg_for_each_descendant_pre(blkg, pos_cgrp, ctx.blkg) | |
1364 | tg_update_has_rules(blkg_to_tg(blkg)); | |
1365 | ||
632b4493 TH |
1366 | /* |
1367 | * We're already holding queue_lock and know @tg is valid. Let's | |
1368 | * apply the new config directly. | |
1369 | * | |
1370 | * Restart the slices for both READ and WRITES. It might happen | |
1371 | * that a group's limit are dropped suddenly and we don't want to | |
1372 | * account recently dispatched IO with new low rate. | |
1373 | */ | |
0f3457f6 TH |
1374 | throtl_start_new_slice(tg, 0); |
1375 | throtl_start_new_slice(tg, 1); | |
632b4493 | 1376 | |
5b2c16aa | 1377 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 1378 | tg_update_disptime(tg); |
7f52f98c | 1379 | throtl_schedule_next_dispatch(sq->parent_sq, true); |
632b4493 | 1380 | } |
60c2bc2d TH |
1381 | |
1382 | blkg_conf_finish(&ctx); | |
a2b1693b | 1383 | return 0; |
8e89d13f VG |
1384 | } |
1385 | ||
af133ceb TH |
1386 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1387 | const char *buf) | |
60c2bc2d | 1388 | { |
af133ceb | 1389 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
1390 | } |
1391 | ||
af133ceb TH |
1392 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1393 | const char *buf) | |
60c2bc2d | 1394 | { |
af133ceb | 1395 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
1396 | } |
1397 | ||
1398 | static struct cftype throtl_files[] = { | |
1399 | { | |
1400 | .name = "throttle.read_bps_device", | |
af133ceb TH |
1401 | .private = offsetof(struct throtl_grp, bps[READ]), |
1402 | .read_seq_string = tg_print_conf_u64, | |
1403 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1404 | .max_write_len = 256, |
1405 | }, | |
1406 | { | |
1407 | .name = "throttle.write_bps_device", | |
af133ceb TH |
1408 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
1409 | .read_seq_string = tg_print_conf_u64, | |
1410 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1411 | .max_write_len = 256, |
1412 | }, | |
1413 | { | |
1414 | .name = "throttle.read_iops_device", | |
af133ceb TH |
1415 | .private = offsetof(struct throtl_grp, iops[READ]), |
1416 | .read_seq_string = tg_print_conf_uint, | |
1417 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1418 | .max_write_len = 256, |
1419 | }, | |
1420 | { | |
1421 | .name = "throttle.write_iops_device", | |
af133ceb TH |
1422 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
1423 | .read_seq_string = tg_print_conf_uint, | |
1424 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1425 | .max_write_len = 256, |
1426 | }, | |
1427 | { | |
1428 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1429 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1430 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1431 | }, |
1432 | { | |
1433 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1434 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1435 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1436 | }, |
1437 | { } /* terminate */ | |
1438 | }; | |
1439 | ||
da527770 | 1440 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1441 | { |
1442 | struct throtl_data *td = q->td; | |
1443 | ||
69df0ab0 | 1444 | cancel_work_sync(&td->dispatch_work); |
e43473b7 VG |
1445 | } |
1446 | ||
3c798398 | 1447 | static struct blkcg_policy blkcg_policy_throtl = { |
f9fcc2d3 TH |
1448 | .pd_size = sizeof(struct throtl_grp), |
1449 | .cftypes = throtl_files, | |
1450 | ||
1451 | .pd_init_fn = throtl_pd_init, | |
693e751e | 1452 | .pd_online_fn = throtl_pd_online, |
f9fcc2d3 TH |
1453 | .pd_exit_fn = throtl_pd_exit, |
1454 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 VG |
1455 | }; |
1456 | ||
bc16a4f9 | 1457 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1458 | { |
1459 | struct throtl_data *td = q->td; | |
c5cc2070 | 1460 | struct throtl_qnode *qn = NULL; |
e43473b7 | 1461 | struct throtl_grp *tg; |
73f0d49a | 1462 | struct throtl_service_queue *sq; |
0e9f4164 | 1463 | bool rw = bio_data_dir(bio); |
3c798398 | 1464 | struct blkcg *blkcg; |
bc16a4f9 | 1465 | bool throttled = false; |
e43473b7 | 1466 | |
2a0f61e6 TH |
1467 | /* see throtl_charge_bio() */ |
1468 | if (bio->bi_rw & REQ_THROTTLED) | |
bc16a4f9 | 1469 | goto out; |
e43473b7 | 1470 | |
af75cd3c VG |
1471 | /* |
1472 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1473 | * basic fields like stats and io rates. If a group has no rules, | |
1474 | * just update the dispatch stats in lockless manner and return. | |
1475 | */ | |
af75cd3c | 1476 | rcu_read_lock(); |
3c798398 | 1477 | blkcg = bio_blkcg(bio); |
cd1604fa | 1478 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1479 | if (tg) { |
693e751e | 1480 | if (!tg->has_rules[rw]) { |
629ed0b1 TH |
1481 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1482 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1483 | goto out_unlock_rcu; |
af75cd3c VG |
1484 | } |
1485 | } | |
af75cd3c VG |
1486 | |
1487 | /* | |
1488 | * Either group has not been allocated yet or it is not an unlimited | |
1489 | * IO group | |
1490 | */ | |
e43473b7 | 1491 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1492 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1493 | if (unlikely(!tg)) |
1494 | goto out_unlock; | |
f469a7b4 | 1495 | |
73f0d49a TH |
1496 | sq = &tg->service_queue; |
1497 | ||
9e660acf TH |
1498 | while (true) { |
1499 | /* throtl is FIFO - if bios are already queued, should queue */ | |
1500 | if (sq->nr_queued[rw]) | |
1501 | break; | |
de701c74 | 1502 | |
9e660acf TH |
1503 | /* if above limits, break to queue */ |
1504 | if (!tg_may_dispatch(tg, bio, NULL)) | |
1505 | break; | |
1506 | ||
1507 | /* within limits, let's charge and dispatch directly */ | |
e43473b7 | 1508 | throtl_charge_bio(tg, bio); |
04521db0 VG |
1509 | |
1510 | /* | |
1511 | * We need to trim slice even when bios are not being queued | |
1512 | * otherwise it might happen that a bio is not queued for | |
1513 | * a long time and slice keeps on extending and trim is not | |
1514 | * called for a long time. Now if limits are reduced suddenly | |
1515 | * we take into account all the IO dispatched so far at new | |
1516 | * low rate and * newly queued IO gets a really long dispatch | |
1517 | * time. | |
1518 | * | |
1519 | * So keep on trimming slice even if bio is not queued. | |
1520 | */ | |
0f3457f6 | 1521 | throtl_trim_slice(tg, rw); |
9e660acf TH |
1522 | |
1523 | /* | |
1524 | * @bio passed through this layer without being throttled. | |
1525 | * Climb up the ladder. If we''re already at the top, it | |
1526 | * can be executed directly. | |
1527 | */ | |
c5cc2070 | 1528 | qn = &tg->qnode_on_parent[rw]; |
9e660acf TH |
1529 | sq = sq->parent_sq; |
1530 | tg = sq_to_tg(sq); | |
1531 | if (!tg) | |
1532 | goto out_unlock; | |
e43473b7 VG |
1533 | } |
1534 | ||
9e660acf | 1535 | /* out-of-limit, queue to @tg */ |
fda6f272 TH |
1536 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
1537 | rw == READ ? 'R' : 'W', | |
1538 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], | |
1539 | tg->io_disp[rw], tg->iops[rw], | |
1540 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
e43473b7 | 1541 | |
671058fb | 1542 | bio_associate_current(bio); |
6bc9c2b4 | 1543 | tg->td->nr_queued[rw]++; |
c5cc2070 | 1544 | throtl_add_bio_tg(bio, qn, tg); |
bc16a4f9 | 1545 | throttled = true; |
e43473b7 | 1546 | |
7f52f98c TH |
1547 | /* |
1548 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
1549 | * was empty before @bio. The forced scheduling isn't likely to | |
1550 | * cause undue delay as @bio is likely to be dispatched directly if | |
1551 | * its @tg's disptime is not in the future. | |
1552 | */ | |
0e9f4164 | 1553 | if (tg->flags & THROTL_TG_WAS_EMPTY) { |
77216b04 | 1554 | tg_update_disptime(tg); |
7f52f98c | 1555 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); |
e43473b7 VG |
1556 | } |
1557 | ||
bc16a4f9 | 1558 | out_unlock: |
e43473b7 | 1559 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1560 | out_unlock_rcu: |
1561 | rcu_read_unlock(); | |
bc16a4f9 | 1562 | out: |
2a0f61e6 TH |
1563 | /* |
1564 | * As multiple blk-throtls may stack in the same issue path, we | |
1565 | * don't want bios to leave with the flag set. Clear the flag if | |
1566 | * being issued. | |
1567 | */ | |
1568 | if (!throttled) | |
1569 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1570 | return throttled; |
e43473b7 VG |
1571 | } |
1572 | ||
2a12f0dc TH |
1573 | /* |
1574 | * Dispatch all bios from all children tg's queued on @parent_sq. On | |
1575 | * return, @parent_sq is guaranteed to not have any active children tg's | |
1576 | * and all bios from previously active tg's are on @parent_sq->bio_lists[]. | |
1577 | */ | |
1578 | static void tg_drain_bios(struct throtl_service_queue *parent_sq) | |
1579 | { | |
1580 | struct throtl_grp *tg; | |
1581 | ||
1582 | while ((tg = throtl_rb_first(parent_sq))) { | |
1583 | struct throtl_service_queue *sq = &tg->service_queue; | |
1584 | struct bio *bio; | |
1585 | ||
1586 | throtl_dequeue_tg(tg); | |
1587 | ||
c5cc2070 | 1588 | while ((bio = throtl_peek_queued(&sq->queued[READ]))) |
2a12f0dc | 1589 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
c5cc2070 | 1590 | while ((bio = throtl_peek_queued(&sq->queued[WRITE]))) |
2a12f0dc TH |
1591 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
1592 | } | |
1593 | } | |
1594 | ||
c9a929dd TH |
1595 | /** |
1596 | * blk_throtl_drain - drain throttled bios | |
1597 | * @q: request_queue to drain throttled bios for | |
1598 | * | |
1599 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1600 | */ | |
1601 | void blk_throtl_drain(struct request_queue *q) | |
1602 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1603 | { | |
1604 | struct throtl_data *td = q->td; | |
2a12f0dc TH |
1605 | struct blkcg_gq *blkg; |
1606 | struct cgroup *pos_cgrp; | |
c9a929dd | 1607 | struct bio *bio; |
651930bc | 1608 | int rw; |
c9a929dd | 1609 | |
8bcb6c7d | 1610 | queue_lockdep_assert_held(q); |
2a12f0dc | 1611 | rcu_read_lock(); |
c9a929dd | 1612 | |
2a12f0dc TH |
1613 | /* |
1614 | * Drain each tg while doing post-order walk on the blkg tree, so | |
1615 | * that all bios are propagated to td->service_queue. It'd be | |
1616 | * better to walk service_queue tree directly but blkg walk is | |
1617 | * easier. | |
1618 | */ | |
1619 | blkg_for_each_descendant_post(blkg, pos_cgrp, td->queue->root_blkg) | |
1620 | tg_drain_bios(&blkg_to_tg(blkg)->service_queue); | |
73f0d49a | 1621 | |
2a12f0dc | 1622 | tg_drain_bios(&td_root_tg(td)->service_queue); |
c9a929dd | 1623 | |
2a12f0dc TH |
1624 | /* finally, transfer bios from top-level tg's into the td */ |
1625 | tg_drain_bios(&td->service_queue); | |
1626 | ||
1627 | rcu_read_unlock(); | |
c9a929dd TH |
1628 | spin_unlock_irq(q->queue_lock); |
1629 | ||
2a12f0dc | 1630 | /* all bios now should be in td->service_queue, issue them */ |
651930bc | 1631 | for (rw = READ; rw <= WRITE; rw++) |
c5cc2070 TH |
1632 | while ((bio = throtl_pop_queued(&td->service_queue.queued[rw], |
1633 | NULL))) | |
651930bc | 1634 | generic_make_request(bio); |
c9a929dd TH |
1635 | |
1636 | spin_lock_irq(q->queue_lock); | |
1637 | } | |
1638 | ||
e43473b7 VG |
1639 | int blk_throtl_init(struct request_queue *q) |
1640 | { | |
1641 | struct throtl_data *td; | |
a2b1693b | 1642 | int ret; |
e43473b7 VG |
1643 | |
1644 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1645 | if (!td) | |
1646 | return -ENOMEM; | |
1647 | ||
69df0ab0 | 1648 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
77216b04 | 1649 | throtl_service_queue_init(&td->service_queue, NULL); |
e43473b7 | 1650 | |
cd1604fa | 1651 | q->td = td; |
29b12589 | 1652 | td->queue = q; |
02977e4a | 1653 | |
a2b1693b | 1654 | /* activate policy */ |
3c798398 | 1655 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1656 | if (ret) |
f51b802c | 1657 | kfree(td); |
a2b1693b | 1658 | return ret; |
e43473b7 VG |
1659 | } |
1660 | ||
1661 | void blk_throtl_exit(struct request_queue *q) | |
1662 | { | |
c875f4d0 | 1663 | BUG_ON(!q->td); |
da527770 | 1664 | throtl_shutdown_wq(q); |
3c798398 | 1665 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1666 | kfree(q->td); |
e43473b7 VG |
1667 | } |
1668 | ||
1669 | static int __init throtl_init(void) | |
1670 | { | |
450adcbe VG |
1671 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1672 | if (!kthrotld_workqueue) | |
1673 | panic("Failed to create kthrotld\n"); | |
1674 | ||
3c798398 | 1675 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1676 | } |
1677 | ||
1678 | module_init(throtl_init); |