]>
Commit | Line | Data |
---|---|---|
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" | |
13 | #include "blk.h" | |
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 | ||
24 | static struct blkcg_policy blkcg_policy_throtl; | |
25 | ||
26 | /* A workqueue to queue throttle related work */ | |
27 | static struct workqueue_struct *kthrotld_workqueue; | |
28 | ||
29 | struct throtl_service_queue { | |
30 | /* | |
31 | * Bios queued directly to this service_queue or dispatched from | |
32 | * children throtl_grp's. | |
33 | */ | |
34 | struct bio_list bio_lists[2]; /* queued bios [READ/WRITE] */ | |
35 | unsigned int nr_queued[2]; /* number of queued bios */ | |
36 | ||
37 | /* | |
38 | * RB tree of active children throtl_grp's, which are sorted by | |
39 | * their ->disptime. | |
40 | */ | |
41 | struct rb_root pending_tree; /* RB tree of active tgs */ | |
42 | struct rb_node *first_pending; /* first node in the tree */ | |
43 | unsigned int nr_pending; /* # queued in the tree */ | |
44 | unsigned long first_pending_disptime; /* disptime of the first tg */ | |
45 | }; | |
46 | ||
47 | enum tg_state_flags { | |
48 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
49 | }; | |
50 | ||
51 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) | |
52 | ||
53 | /* Per-cpu group stats */ | |
54 | struct tg_stats_cpu { | |
55 | /* total bytes transferred */ | |
56 | struct blkg_rwstat service_bytes; | |
57 | /* total IOs serviced, post merge */ | |
58 | struct blkg_rwstat serviced; | |
59 | }; | |
60 | ||
61 | struct throtl_grp { | |
62 | /* must be the first member */ | |
63 | struct blkg_policy_data pd; | |
64 | ||
65 | /* active throtl group service_queue member */ | |
66 | struct rb_node rb_node; | |
67 | ||
68 | /* throtl_data this group belongs to */ | |
69 | struct throtl_data *td; | |
70 | ||
71 | /* this group's service queue */ | |
72 | struct throtl_service_queue service_queue; | |
73 | ||
74 | /* | |
75 | * Dispatch time in jiffies. This is the estimated time when group | |
76 | * will unthrottle and is ready to dispatch more bio. It is used as | |
77 | * key to sort active groups in service tree. | |
78 | */ | |
79 | unsigned long disptime; | |
80 | ||
81 | unsigned int flags; | |
82 | ||
83 | /* bytes per second rate limits */ | |
84 | uint64_t bps[2]; | |
85 | ||
86 | /* IOPS limits */ | |
87 | unsigned int iops[2]; | |
88 | ||
89 | /* Number of bytes disptached in current slice */ | |
90 | uint64_t bytes_disp[2]; | |
91 | /* Number of bio's dispatched in current slice */ | |
92 | unsigned int io_disp[2]; | |
93 | ||
94 | /* When did we start a new slice */ | |
95 | unsigned long slice_start[2]; | |
96 | unsigned long slice_end[2]; | |
97 | ||
98 | /* Per cpu stats pointer */ | |
99 | struct tg_stats_cpu __percpu *stats_cpu; | |
100 | ||
101 | /* List of tgs waiting for per cpu stats memory to be allocated */ | |
102 | struct list_head stats_alloc_node; | |
103 | }; | |
104 | ||
105 | struct throtl_data | |
106 | { | |
107 | /* service tree for active throtl groups */ | |
108 | struct throtl_service_queue service_queue; | |
109 | ||
110 | struct request_queue *queue; | |
111 | ||
112 | /* Total Number of queued bios on READ and WRITE lists */ | |
113 | unsigned int nr_queued[2]; | |
114 | ||
115 | /* | |
116 | * number of total undestroyed groups | |
117 | */ | |
118 | unsigned int nr_undestroyed_grps; | |
119 | ||
120 | /* Work for dispatching throttled bios */ | |
121 | struct delayed_work dispatch_work; | |
122 | }; | |
123 | ||
124 | /* list and work item to allocate percpu group stats */ | |
125 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
126 | static LIST_HEAD(tg_stats_alloc_list); | |
127 | ||
128 | static void tg_stats_alloc_fn(struct work_struct *); | |
129 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
130 | ||
131 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) | |
132 | { | |
133 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
134 | } | |
135 | ||
136 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) | |
137 | { | |
138 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); | |
139 | } | |
140 | ||
141 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) | |
142 | { | |
143 | return pd_to_blkg(&tg->pd); | |
144 | } | |
145 | ||
146 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) | |
147 | { | |
148 | return blkg_to_tg(td->queue->root_blkg); | |
149 | } | |
150 | ||
151 | #define throtl_log_tg(tg, fmt, args...) do { \ | |
152 | char __pbuf[128]; \ | |
153 | \ | |
154 | blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \ | |
155 | blk_add_trace_msg((tg)->td->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
156 | } while (0) | |
157 | ||
158 | #define throtl_log(td, fmt, args...) \ | |
159 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) | |
160 | ||
161 | /* | |
162 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
163 | * system_wq once there are some groups on the alloc_list waiting for | |
164 | * allocation. | |
165 | */ | |
166 | static void tg_stats_alloc_fn(struct work_struct *work) | |
167 | { | |
168 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
169 | struct delayed_work *dwork = to_delayed_work(work); | |
170 | bool empty = false; | |
171 | ||
172 | alloc_stats: | |
173 | if (!stats_cpu) { | |
174 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
175 | if (!stats_cpu) { | |
176 | /* allocation failed, try again after some time */ | |
177 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); | |
178 | return; | |
179 | } | |
180 | } | |
181 | ||
182 | spin_lock_irq(&tg_stats_alloc_lock); | |
183 | ||
184 | if (!list_empty(&tg_stats_alloc_list)) { | |
185 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
186 | struct throtl_grp, | |
187 | stats_alloc_node); | |
188 | swap(tg->stats_cpu, stats_cpu); | |
189 | list_del_init(&tg->stats_alloc_node); | |
190 | } | |
191 | ||
192 | empty = list_empty(&tg_stats_alloc_list); | |
193 | spin_unlock_irq(&tg_stats_alloc_lock); | |
194 | if (!empty) | |
195 | goto alloc_stats; | |
196 | } | |
197 | ||
198 | /* init a service_queue, assumes the caller zeroed it */ | |
199 | static void throtl_service_queue_init(struct throtl_service_queue *sq) | |
200 | { | |
201 | bio_list_init(&sq->bio_lists[0]); | |
202 | bio_list_init(&sq->bio_lists[1]); | |
203 | sq->pending_tree = RB_ROOT; | |
204 | } | |
205 | ||
206 | static void throtl_pd_init(struct blkcg_gq *blkg) | |
207 | { | |
208 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
209 | unsigned long flags; | |
210 | ||
211 | throtl_service_queue_init(&tg->service_queue); | |
212 | RB_CLEAR_NODE(&tg->rb_node); | |
213 | tg->td = blkg->q->td; | |
214 | ||
215 | tg->bps[READ] = -1; | |
216 | tg->bps[WRITE] = -1; | |
217 | tg->iops[READ] = -1; | |
218 | tg->iops[WRITE] = -1; | |
219 | ||
220 | /* | |
221 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
222 | * but percpu allocator can't be called from IO path. Queue tg on | |
223 | * tg_stats_alloc_list and allocate from work item. | |
224 | */ | |
225 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); | |
226 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); | |
227 | schedule_delayed_work(&tg_stats_alloc_work, 0); | |
228 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); | |
229 | } | |
230 | ||
231 | static void throtl_pd_exit(struct blkcg_gq *blkg) | |
232 | { | |
233 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
234 | unsigned long flags; | |
235 | ||
236 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); | |
237 | list_del_init(&tg->stats_alloc_node); | |
238 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); | |
239 | ||
240 | free_percpu(tg->stats_cpu); | |
241 | } | |
242 | ||
243 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) | |
244 | { | |
245 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
246 | int cpu; | |
247 | ||
248 | if (tg->stats_cpu == NULL) | |
249 | return; | |
250 | ||
251 | for_each_possible_cpu(cpu) { | |
252 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
253 | ||
254 | blkg_rwstat_reset(&sc->service_bytes); | |
255 | blkg_rwstat_reset(&sc->serviced); | |
256 | } | |
257 | } | |
258 | ||
259 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, | |
260 | struct blkcg *blkcg) | |
261 | { | |
262 | /* | |
263 | * This is the common case when there are no blkcgs. Avoid lookup | |
264 | * in this case | |
265 | */ | |
266 | if (blkcg == &blkcg_root) | |
267 | return td_root_tg(td); | |
268 | ||
269 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); | |
270 | } | |
271 | ||
272 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, | |
273 | struct blkcg *blkcg) | |
274 | { | |
275 | struct request_queue *q = td->queue; | |
276 | struct throtl_grp *tg = NULL; | |
277 | ||
278 | /* | |
279 | * This is the common case when there are no blkcgs. Avoid lookup | |
280 | * in this case | |
281 | */ | |
282 | if (blkcg == &blkcg_root) { | |
283 | tg = td_root_tg(td); | |
284 | } else { | |
285 | struct blkcg_gq *blkg; | |
286 | ||
287 | blkg = blkg_lookup_create(blkcg, q); | |
288 | ||
289 | /* if %NULL and @q is alive, fall back to root_tg */ | |
290 | if (!IS_ERR(blkg)) | |
291 | tg = blkg_to_tg(blkg); | |
292 | else if (!blk_queue_dying(q)) | |
293 | tg = td_root_tg(td); | |
294 | } | |
295 | ||
296 | return tg; | |
297 | } | |
298 | ||
299 | static struct throtl_grp * | |
300 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
301 | { | |
302 | /* Service tree is empty */ | |
303 | if (!parent_sq->nr_pending) | |
304 | return NULL; | |
305 | ||
306 | if (!parent_sq->first_pending) | |
307 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
308 | ||
309 | if (parent_sq->first_pending) | |
310 | return rb_entry_tg(parent_sq->first_pending); | |
311 | ||
312 | return NULL; | |
313 | } | |
314 | ||
315 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
316 | { | |
317 | rb_erase(n, root); | |
318 | RB_CLEAR_NODE(n); | |
319 | } | |
320 | ||
321 | static void throtl_rb_erase(struct rb_node *n, | |
322 | struct throtl_service_queue *parent_sq) | |
323 | { | |
324 | if (parent_sq->first_pending == n) | |
325 | parent_sq->first_pending = NULL; | |
326 | rb_erase_init(n, &parent_sq->pending_tree); | |
327 | --parent_sq->nr_pending; | |
328 | } | |
329 | ||
330 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) | |
331 | { | |
332 | struct throtl_grp *tg; | |
333 | ||
334 | tg = throtl_rb_first(parent_sq); | |
335 | if (!tg) | |
336 | return; | |
337 | ||
338 | parent_sq->first_pending_disptime = tg->disptime; | |
339 | } | |
340 | ||
341 | static void tg_service_queue_add(struct throtl_grp *tg, | |
342 | struct throtl_service_queue *parent_sq) | |
343 | { | |
344 | struct rb_node **node = &parent_sq->pending_tree.rb_node; | |
345 | struct rb_node *parent = NULL; | |
346 | struct throtl_grp *__tg; | |
347 | unsigned long key = tg->disptime; | |
348 | int left = 1; | |
349 | ||
350 | while (*node != NULL) { | |
351 | parent = *node; | |
352 | __tg = rb_entry_tg(parent); | |
353 | ||
354 | if (time_before(key, __tg->disptime)) | |
355 | node = &parent->rb_left; | |
356 | else { | |
357 | node = &parent->rb_right; | |
358 | left = 0; | |
359 | } | |
360 | } | |
361 | ||
362 | if (left) | |
363 | parent_sq->first_pending = &tg->rb_node; | |
364 | ||
365 | rb_link_node(&tg->rb_node, parent, node); | |
366 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); | |
367 | } | |
368 | ||
369 | static void __throtl_enqueue_tg(struct throtl_grp *tg, | |
370 | struct throtl_service_queue *parent_sq) | |
371 | { | |
372 | tg_service_queue_add(tg, parent_sq); | |
373 | tg->flags |= THROTL_TG_PENDING; | |
374 | parent_sq->nr_pending++; | |
375 | } | |
376 | ||
377 | static void throtl_enqueue_tg(struct throtl_grp *tg, | |
378 | struct throtl_service_queue *parent_sq) | |
379 | { | |
380 | if (!(tg->flags & THROTL_TG_PENDING)) | |
381 | __throtl_enqueue_tg(tg, parent_sq); | |
382 | } | |
383 | ||
384 | static void __throtl_dequeue_tg(struct throtl_grp *tg, | |
385 | struct throtl_service_queue *parent_sq) | |
386 | { | |
387 | throtl_rb_erase(&tg->rb_node, parent_sq); | |
388 | tg->flags &= ~THROTL_TG_PENDING; | |
389 | } | |
390 | ||
391 | static void throtl_dequeue_tg(struct throtl_grp *tg, | |
392 | struct throtl_service_queue *parent_sq) | |
393 | { | |
394 | if (tg->flags & THROTL_TG_PENDING) | |
395 | __throtl_dequeue_tg(tg, parent_sq); | |
396 | } | |
397 | ||
398 | /* Call with queue lock held */ | |
399 | static void throtl_schedule_delayed_work(struct throtl_data *td, | |
400 | unsigned long delay) | |
401 | { | |
402 | struct delayed_work *dwork = &td->dispatch_work; | |
403 | ||
404 | mod_delayed_work(kthrotld_workqueue, dwork, delay); | |
405 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", delay, jiffies); | |
406 | } | |
407 | ||
408 | static void throtl_schedule_next_dispatch(struct throtl_data *td) | |
409 | { | |
410 | struct throtl_service_queue *sq = &td->service_queue; | |
411 | ||
412 | /* any pending children left? */ | |
413 | if (!sq->nr_pending) | |
414 | return; | |
415 | ||
416 | update_min_dispatch_time(sq); | |
417 | ||
418 | if (time_before_eq(sq->first_pending_disptime, jiffies)) | |
419 | throtl_schedule_delayed_work(td, 0); | |
420 | else | |
421 | throtl_schedule_delayed_work(td, sq->first_pending_disptime - jiffies); | |
422 | } | |
423 | ||
424 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) | |
425 | { | |
426 | tg->bytes_disp[rw] = 0; | |
427 | tg->io_disp[rw] = 0; | |
428 | tg->slice_start[rw] = jiffies; | |
429 | tg->slice_end[rw] = jiffies + throtl_slice; | |
430 | throtl_log_tg(tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
431 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
432 | tg->slice_end[rw], jiffies); | |
433 | } | |
434 | ||
435 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, | |
436 | unsigned long jiffy_end) | |
437 | { | |
438 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
439 | } | |
440 | ||
441 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, | |
442 | unsigned long jiffy_end) | |
443 | { | |
444 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
445 | throtl_log_tg(tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
446 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
447 | tg->slice_end[rw], jiffies); | |
448 | } | |
449 | ||
450 | /* Determine if previously allocated or extended slice is complete or not */ | |
451 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) | |
452 | { | |
453 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
454 | return 0; | |
455 | ||
456 | return 1; | |
457 | } | |
458 | ||
459 | /* Trim the used slices and adjust slice start accordingly */ | |
460 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) | |
461 | { | |
462 | unsigned long nr_slices, time_elapsed, io_trim; | |
463 | u64 bytes_trim, tmp; | |
464 | ||
465 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
466 | ||
467 | /* | |
468 | * If bps are unlimited (-1), then time slice don't get | |
469 | * renewed. Don't try to trim the slice if slice is used. A new | |
470 | * slice will start when appropriate. | |
471 | */ | |
472 | if (throtl_slice_used(tg, rw)) | |
473 | return; | |
474 | ||
475 | /* | |
476 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
477 | * that initially cgroup limit was very low resulting in high | |
478 | * slice_end, but later limit was bumped up and bio was dispached | |
479 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
480 | * is bad because it does not allow new slice to start. | |
481 | */ | |
482 | ||
483 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); | |
484 | ||
485 | time_elapsed = jiffies - tg->slice_start[rw]; | |
486 | ||
487 | nr_slices = time_elapsed / throtl_slice; | |
488 | ||
489 | if (!nr_slices) | |
490 | return; | |
491 | tmp = tg->bps[rw] * throtl_slice * nr_slices; | |
492 | do_div(tmp, HZ); | |
493 | bytes_trim = tmp; | |
494 | ||
495 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; | |
496 | ||
497 | if (!bytes_trim && !io_trim) | |
498 | return; | |
499 | ||
500 | if (tg->bytes_disp[rw] >= bytes_trim) | |
501 | tg->bytes_disp[rw] -= bytes_trim; | |
502 | else | |
503 | tg->bytes_disp[rw] = 0; | |
504 | ||
505 | if (tg->io_disp[rw] >= io_trim) | |
506 | tg->io_disp[rw] -= io_trim; | |
507 | else | |
508 | tg->io_disp[rw] = 0; | |
509 | ||
510 | tg->slice_start[rw] += nr_slices * throtl_slice; | |
511 | ||
512 | throtl_log_tg(tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" | |
513 | " start=%lu end=%lu jiffies=%lu", | |
514 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
515 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
516 | } | |
517 | ||
518 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, | |
519 | unsigned long *wait) | |
520 | { | |
521 | bool rw = bio_data_dir(bio); | |
522 | unsigned int io_allowed; | |
523 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; | |
524 | u64 tmp; | |
525 | ||
526 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
527 | ||
528 | /* Slice has just started. Consider one slice interval */ | |
529 | if (!jiffy_elapsed) | |
530 | jiffy_elapsed_rnd = throtl_slice; | |
531 | ||
532 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
533 | ||
534 | /* | |
535 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
536 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
537 | * will allow dispatch after 1 second and after that slice should | |
538 | * have been trimmed. | |
539 | */ | |
540 | ||
541 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
542 | do_div(tmp, HZ); | |
543 | ||
544 | if (tmp > UINT_MAX) | |
545 | io_allowed = UINT_MAX; | |
546 | else | |
547 | io_allowed = tmp; | |
548 | ||
549 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
550 | if (wait) | |
551 | *wait = 0; | |
552 | return 1; | |
553 | } | |
554 | ||
555 | /* Calc approx time to dispatch */ | |
556 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
557 | ||
558 | if (jiffy_wait > jiffy_elapsed) | |
559 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
560 | else | |
561 | jiffy_wait = 1; | |
562 | ||
563 | if (wait) | |
564 | *wait = jiffy_wait; | |
565 | return 0; | |
566 | } | |
567 | ||
568 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, | |
569 | unsigned long *wait) | |
570 | { | |
571 | bool rw = bio_data_dir(bio); | |
572 | u64 bytes_allowed, extra_bytes, tmp; | |
573 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; | |
574 | ||
575 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
576 | ||
577 | /* Slice has just started. Consider one slice interval */ | |
578 | if (!jiffy_elapsed) | |
579 | jiffy_elapsed_rnd = throtl_slice; | |
580 | ||
581 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
582 | ||
583 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; | |
584 | do_div(tmp, HZ); | |
585 | bytes_allowed = tmp; | |
586 | ||
587 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
588 | if (wait) | |
589 | *wait = 0; | |
590 | return 1; | |
591 | } | |
592 | ||
593 | /* Calc approx time to dispatch */ | |
594 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
595 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
596 | ||
597 | if (!jiffy_wait) | |
598 | jiffy_wait = 1; | |
599 | ||
600 | /* | |
601 | * This wait time is without taking into consideration the rounding | |
602 | * up we did. Add that time also. | |
603 | */ | |
604 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
605 | if (wait) | |
606 | *wait = jiffy_wait; | |
607 | return 0; | |
608 | } | |
609 | ||
610 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { | |
611 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
612 | return 1; | |
613 | return 0; | |
614 | } | |
615 | ||
616 | /* | |
617 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
618 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
619 | */ | |
620 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, | |
621 | unsigned long *wait) | |
622 | { | |
623 | bool rw = bio_data_dir(bio); | |
624 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
625 | ||
626 | /* | |
627 | * Currently whole state machine of group depends on first bio | |
628 | * queued in the group bio list. So one should not be calling | |
629 | * this function with a different bio if there are other bios | |
630 | * queued. | |
631 | */ | |
632 | BUG_ON(tg->service_queue.nr_queued[rw] && | |
633 | bio != bio_list_peek(&tg->service_queue.bio_lists[rw])); | |
634 | ||
635 | /* If tg->bps = -1, then BW is unlimited */ | |
636 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
637 | if (wait) | |
638 | *wait = 0; | |
639 | return 1; | |
640 | } | |
641 | ||
642 | /* | |
643 | * If previous slice expired, start a new one otherwise renew/extend | |
644 | * existing slice to make sure it is at least throtl_slice interval | |
645 | * long since now. | |
646 | */ | |
647 | if (throtl_slice_used(tg, rw)) | |
648 | throtl_start_new_slice(tg, rw); | |
649 | else { | |
650 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
651 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); | |
652 | } | |
653 | ||
654 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && | |
655 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
656 | if (wait) | |
657 | *wait = 0; | |
658 | return 1; | |
659 | } | |
660 | ||
661 | max_wait = max(bps_wait, iops_wait); | |
662 | ||
663 | if (wait) | |
664 | *wait = max_wait; | |
665 | ||
666 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
667 | throtl_extend_slice(tg, rw, jiffies + max_wait); | |
668 | ||
669 | return 0; | |
670 | } | |
671 | ||
672 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, | |
673 | int rw) | |
674 | { | |
675 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
676 | struct tg_stats_cpu *stats_cpu; | |
677 | unsigned long flags; | |
678 | ||
679 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
680 | if (tg->stats_cpu == NULL) | |
681 | return; | |
682 | ||
683 | /* | |
684 | * Disabling interrupts to provide mutual exclusion between two | |
685 | * writes on same cpu. It probably is not needed for 64bit. Not | |
686 | * optimizing that case yet. | |
687 | */ | |
688 | local_irq_save(flags); | |
689 | ||
690 | stats_cpu = this_cpu_ptr(tg->stats_cpu); | |
691 | ||
692 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); | |
693 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
694 | ||
695 | local_irq_restore(flags); | |
696 | } | |
697 | ||
698 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
699 | { | |
700 | bool rw = bio_data_dir(bio); | |
701 | ||
702 | /* Charge the bio to the group */ | |
703 | tg->bytes_disp[rw] += bio->bi_size; | |
704 | tg->io_disp[rw]++; | |
705 | ||
706 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw); | |
707 | } | |
708 | ||
709 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_grp *tg, | |
710 | struct throtl_service_queue *parent_sq) | |
711 | { | |
712 | struct throtl_service_queue *sq = &tg->service_queue; | |
713 | bool rw = bio_data_dir(bio); | |
714 | ||
715 | bio_list_add(&sq->bio_lists[rw], bio); | |
716 | /* Take a bio reference on tg */ | |
717 | blkg_get(tg_to_blkg(tg)); | |
718 | sq->nr_queued[rw]++; | |
719 | tg->td->nr_queued[rw]++; | |
720 | throtl_enqueue_tg(tg, parent_sq); | |
721 | } | |
722 | ||
723 | static void tg_update_disptime(struct throtl_grp *tg, | |
724 | struct throtl_service_queue *parent_sq) | |
725 | { | |
726 | struct throtl_service_queue *sq = &tg->service_queue; | |
727 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
728 | struct bio *bio; | |
729 | ||
730 | if ((bio = bio_list_peek(&sq->bio_lists[READ]))) | |
731 | tg_may_dispatch(tg, bio, &read_wait); | |
732 | ||
733 | if ((bio = bio_list_peek(&sq->bio_lists[WRITE]))) | |
734 | tg_may_dispatch(tg, bio, &write_wait); | |
735 | ||
736 | min_wait = min(read_wait, write_wait); | |
737 | disptime = jiffies + min_wait; | |
738 | ||
739 | /* Update dispatch time */ | |
740 | throtl_dequeue_tg(tg, parent_sq); | |
741 | tg->disptime = disptime; | |
742 | throtl_enqueue_tg(tg, parent_sq); | |
743 | } | |
744 | ||
745 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw, | |
746 | struct bio_list *bl) | |
747 | { | |
748 | struct throtl_service_queue *sq = &tg->service_queue; | |
749 | struct bio *bio; | |
750 | ||
751 | bio = bio_list_pop(&sq->bio_lists[rw]); | |
752 | sq->nr_queued[rw]--; | |
753 | /* Drop bio reference on blkg */ | |
754 | blkg_put(tg_to_blkg(tg)); | |
755 | ||
756 | BUG_ON(tg->td->nr_queued[rw] <= 0); | |
757 | tg->td->nr_queued[rw]--; | |
758 | ||
759 | throtl_charge_bio(tg, bio); | |
760 | bio_list_add(bl, bio); | |
761 | bio->bi_rw |= REQ_THROTTLED; | |
762 | ||
763 | throtl_trim_slice(tg, rw); | |
764 | } | |
765 | ||
766 | static int throtl_dispatch_tg(struct throtl_grp *tg, struct bio_list *bl) | |
767 | { | |
768 | struct throtl_service_queue *sq = &tg->service_queue; | |
769 | unsigned int nr_reads = 0, nr_writes = 0; | |
770 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
771 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; | |
772 | struct bio *bio; | |
773 | ||
774 | /* Try to dispatch 75% READS and 25% WRITES */ | |
775 | ||
776 | while ((bio = bio_list_peek(&sq->bio_lists[READ])) && | |
777 | tg_may_dispatch(tg, bio, NULL)) { | |
778 | ||
779 | tg_dispatch_one_bio(tg, bio_data_dir(bio), bl); | |
780 | nr_reads++; | |
781 | ||
782 | if (nr_reads >= max_nr_reads) | |
783 | break; | |
784 | } | |
785 | ||
786 | while ((bio = bio_list_peek(&sq->bio_lists[WRITE])) && | |
787 | tg_may_dispatch(tg, bio, NULL)) { | |
788 | ||
789 | tg_dispatch_one_bio(tg, bio_data_dir(bio), bl); | |
790 | nr_writes++; | |
791 | ||
792 | if (nr_writes >= max_nr_writes) | |
793 | break; | |
794 | } | |
795 | ||
796 | return nr_reads + nr_writes; | |
797 | } | |
798 | ||
799 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq, | |
800 | struct bio_list *bl) | |
801 | { | |
802 | unsigned int nr_disp = 0; | |
803 | ||
804 | while (1) { | |
805 | struct throtl_grp *tg = throtl_rb_first(parent_sq); | |
806 | struct throtl_service_queue *sq = &tg->service_queue; | |
807 | ||
808 | if (!tg) | |
809 | break; | |
810 | ||
811 | if (time_before(jiffies, tg->disptime)) | |
812 | break; | |
813 | ||
814 | throtl_dequeue_tg(tg, parent_sq); | |
815 | ||
816 | nr_disp += throtl_dispatch_tg(tg, bl); | |
817 | ||
818 | if (sq->nr_queued[0] || sq->nr_queued[1]) | |
819 | tg_update_disptime(tg, parent_sq); | |
820 | ||
821 | if (nr_disp >= throtl_quantum) | |
822 | break; | |
823 | } | |
824 | ||
825 | return nr_disp; | |
826 | } | |
827 | ||
828 | /* work function to dispatch throttled bios */ | |
829 | void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
830 | { | |
831 | struct throtl_data *td = container_of(to_delayed_work(work), | |
832 | struct throtl_data, dispatch_work); | |
833 | struct request_queue *q = td->queue; | |
834 | unsigned int nr_disp = 0; | |
835 | struct bio_list bio_list_on_stack; | |
836 | struct bio *bio; | |
837 | struct blk_plug plug; | |
838 | ||
839 | spin_lock_irq(q->queue_lock); | |
840 | ||
841 | bio_list_init(&bio_list_on_stack); | |
842 | ||
843 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", | |
844 | td->nr_queued[READ] + td->nr_queued[WRITE], | |
845 | td->nr_queued[READ], td->nr_queued[WRITE]); | |
846 | ||
847 | nr_disp = throtl_select_dispatch(&td->service_queue, &bio_list_on_stack); | |
848 | ||
849 | if (nr_disp) | |
850 | throtl_log(td, "bios disp=%u", nr_disp); | |
851 | ||
852 | throtl_schedule_next_dispatch(td); | |
853 | ||
854 | spin_unlock_irq(q->queue_lock); | |
855 | ||
856 | /* | |
857 | * If we dispatched some requests, unplug the queue to make sure | |
858 | * immediate dispatch | |
859 | */ | |
860 | if (nr_disp) { | |
861 | blk_start_plug(&plug); | |
862 | while((bio = bio_list_pop(&bio_list_on_stack))) | |
863 | generic_make_request(bio); | |
864 | blk_finish_plug(&plug); | |
865 | } | |
866 | } | |
867 | ||
868 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, | |
869 | struct blkg_policy_data *pd, int off) | |
870 | { | |
871 | struct throtl_grp *tg = pd_to_tg(pd); | |
872 | struct blkg_rwstat rwstat = { }, tmp; | |
873 | int i, cpu; | |
874 | ||
875 | for_each_possible_cpu(cpu) { | |
876 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
877 | ||
878 | tmp = blkg_rwstat_read((void *)sc + off); | |
879 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
880 | rwstat.cnt[i] += tmp.cnt[i]; | |
881 | } | |
882 | ||
883 | return __blkg_prfill_rwstat(sf, pd, &rwstat); | |
884 | } | |
885 | ||
886 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, | |
887 | struct seq_file *sf) | |
888 | { | |
889 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); | |
890 | ||
891 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, | |
892 | cft->private, true); | |
893 | return 0; | |
894 | } | |
895 | ||
896 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, | |
897 | int off) | |
898 | { | |
899 | struct throtl_grp *tg = pd_to_tg(pd); | |
900 | u64 v = *(u64 *)((void *)tg + off); | |
901 | ||
902 | if (v == -1) | |
903 | return 0; | |
904 | return __blkg_prfill_u64(sf, pd, v); | |
905 | } | |
906 | ||
907 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, | |
908 | int off) | |
909 | { | |
910 | struct throtl_grp *tg = pd_to_tg(pd); | |
911 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
912 | ||
913 | if (v == -1) | |
914 | return 0; | |
915 | return __blkg_prfill_u64(sf, pd, v); | |
916 | } | |
917 | ||
918 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, | |
919 | struct seq_file *sf) | |
920 | { | |
921 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, | |
922 | &blkcg_policy_throtl, cft->private, false); | |
923 | return 0; | |
924 | } | |
925 | ||
926 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, | |
927 | struct seq_file *sf) | |
928 | { | |
929 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, | |
930 | &blkcg_policy_throtl, cft->private, false); | |
931 | return 0; | |
932 | } | |
933 | ||
934 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, | |
935 | bool is_u64) | |
936 | { | |
937 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); | |
938 | struct blkg_conf_ctx ctx; | |
939 | struct throtl_grp *tg; | |
940 | struct throtl_data *td; | |
941 | int ret; | |
942 | ||
943 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
944 | if (ret) | |
945 | return ret; | |
946 | ||
947 | tg = blkg_to_tg(ctx.blkg); | |
948 | td = ctx.blkg->q->td; | |
949 | ||
950 | if (!ctx.v) | |
951 | ctx.v = -1; | |
952 | ||
953 | if (is_u64) | |
954 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
955 | else | |
956 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
957 | ||
958 | throtl_log_tg(tg, "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
959 | tg->bps[READ], tg->bps[WRITE], | |
960 | tg->iops[READ], tg->iops[WRITE]); | |
961 | ||
962 | /* | |
963 | * We're already holding queue_lock and know @tg is valid. Let's | |
964 | * apply the new config directly. | |
965 | * | |
966 | * Restart the slices for both READ and WRITES. It might happen | |
967 | * that a group's limit are dropped suddenly and we don't want to | |
968 | * account recently dispatched IO with new low rate. | |
969 | */ | |
970 | throtl_start_new_slice(tg, 0); | |
971 | throtl_start_new_slice(tg, 1); | |
972 | ||
973 | if (tg->flags & THROTL_TG_PENDING) { | |
974 | tg_update_disptime(tg, &td->service_queue); | |
975 | throtl_schedule_next_dispatch(td); | |
976 | } | |
977 | ||
978 | blkg_conf_finish(&ctx); | |
979 | return 0; | |
980 | } | |
981 | ||
982 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, | |
983 | const char *buf) | |
984 | { | |
985 | return tg_set_conf(cgrp, cft, buf, true); | |
986 | } | |
987 | ||
988 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, | |
989 | const char *buf) | |
990 | { | |
991 | return tg_set_conf(cgrp, cft, buf, false); | |
992 | } | |
993 | ||
994 | static struct cftype throtl_files[] = { | |
995 | { | |
996 | .name = "throttle.read_bps_device", | |
997 | .private = offsetof(struct throtl_grp, bps[READ]), | |
998 | .read_seq_string = tg_print_conf_u64, | |
999 | .write_string = tg_set_conf_u64, | |
1000 | .max_write_len = 256, | |
1001 | }, | |
1002 | { | |
1003 | .name = "throttle.write_bps_device", | |
1004 | .private = offsetof(struct throtl_grp, bps[WRITE]), | |
1005 | .read_seq_string = tg_print_conf_u64, | |
1006 | .write_string = tg_set_conf_u64, | |
1007 | .max_write_len = 256, | |
1008 | }, | |
1009 | { | |
1010 | .name = "throttle.read_iops_device", | |
1011 | .private = offsetof(struct throtl_grp, iops[READ]), | |
1012 | .read_seq_string = tg_print_conf_uint, | |
1013 | .write_string = tg_set_conf_uint, | |
1014 | .max_write_len = 256, | |
1015 | }, | |
1016 | { | |
1017 | .name = "throttle.write_iops_device", | |
1018 | .private = offsetof(struct throtl_grp, iops[WRITE]), | |
1019 | .read_seq_string = tg_print_conf_uint, | |
1020 | .write_string = tg_set_conf_uint, | |
1021 | .max_write_len = 256, | |
1022 | }, | |
1023 | { | |
1024 | .name = "throttle.io_service_bytes", | |
1025 | .private = offsetof(struct tg_stats_cpu, service_bytes), | |
1026 | .read_seq_string = tg_print_cpu_rwstat, | |
1027 | }, | |
1028 | { | |
1029 | .name = "throttle.io_serviced", | |
1030 | .private = offsetof(struct tg_stats_cpu, serviced), | |
1031 | .read_seq_string = tg_print_cpu_rwstat, | |
1032 | }, | |
1033 | { } /* terminate */ | |
1034 | }; | |
1035 | ||
1036 | static void throtl_shutdown_wq(struct request_queue *q) | |
1037 | { | |
1038 | struct throtl_data *td = q->td; | |
1039 | ||
1040 | cancel_delayed_work_sync(&td->dispatch_work); | |
1041 | } | |
1042 | ||
1043 | static struct blkcg_policy blkcg_policy_throtl = { | |
1044 | .pd_size = sizeof(struct throtl_grp), | |
1045 | .cftypes = throtl_files, | |
1046 | ||
1047 | .pd_init_fn = throtl_pd_init, | |
1048 | .pd_exit_fn = throtl_pd_exit, | |
1049 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
1050 | }; | |
1051 | ||
1052 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) | |
1053 | { | |
1054 | struct throtl_data *td = q->td; | |
1055 | struct throtl_grp *tg; | |
1056 | struct throtl_service_queue *sq; | |
1057 | bool rw = bio_data_dir(bio), update_disptime = true; | |
1058 | struct blkcg *blkcg; | |
1059 | bool throttled = false; | |
1060 | ||
1061 | if (bio->bi_rw & REQ_THROTTLED) { | |
1062 | bio->bi_rw &= ~REQ_THROTTLED; | |
1063 | goto out; | |
1064 | } | |
1065 | ||
1066 | /* | |
1067 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1068 | * basic fields like stats and io rates. If a group has no rules, | |
1069 | * just update the dispatch stats in lockless manner and return. | |
1070 | */ | |
1071 | rcu_read_lock(); | |
1072 | blkcg = bio_blkcg(bio); | |
1073 | tg = throtl_lookup_tg(td, blkcg); | |
1074 | if (tg) { | |
1075 | if (tg_no_rule_group(tg, rw)) { | |
1076 | throtl_update_dispatch_stats(tg_to_blkg(tg), | |
1077 | bio->bi_size, bio->bi_rw); | |
1078 | goto out_unlock_rcu; | |
1079 | } | |
1080 | } | |
1081 | ||
1082 | /* | |
1083 | * Either group has not been allocated yet or it is not an unlimited | |
1084 | * IO group | |
1085 | */ | |
1086 | spin_lock_irq(q->queue_lock); | |
1087 | tg = throtl_lookup_create_tg(td, blkcg); | |
1088 | if (unlikely(!tg)) | |
1089 | goto out_unlock; | |
1090 | ||
1091 | sq = &tg->service_queue; | |
1092 | ||
1093 | if (sq->nr_queued[rw]) { | |
1094 | /* | |
1095 | * There is already another bio queued in same dir. No | |
1096 | * need to update dispatch time. | |
1097 | */ | |
1098 | update_disptime = false; | |
1099 | goto queue_bio; | |
1100 | ||
1101 | } | |
1102 | ||
1103 | /* Bio is with-in rate limit of group */ | |
1104 | if (tg_may_dispatch(tg, bio, NULL)) { | |
1105 | throtl_charge_bio(tg, bio); | |
1106 | ||
1107 | /* | |
1108 | * We need to trim slice even when bios are not being queued | |
1109 | * otherwise it might happen that a bio is not queued for | |
1110 | * a long time and slice keeps on extending and trim is not | |
1111 | * called for a long time. Now if limits are reduced suddenly | |
1112 | * we take into account all the IO dispatched so far at new | |
1113 | * low rate and * newly queued IO gets a really long dispatch | |
1114 | * time. | |
1115 | * | |
1116 | * So keep on trimming slice even if bio is not queued. | |
1117 | */ | |
1118 | throtl_trim_slice(tg, rw); | |
1119 | goto out_unlock; | |
1120 | } | |
1121 | ||
1122 | queue_bio: | |
1123 | throtl_log_tg(tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" | |
1124 | " iodisp=%u iops=%u queued=%d/%d", | |
1125 | rw == READ ? 'R' : 'W', | |
1126 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], | |
1127 | tg->io_disp[rw], tg->iops[rw], | |
1128 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
1129 | ||
1130 | bio_associate_current(bio); | |
1131 | throtl_add_bio_tg(bio, tg, &q->td->service_queue); | |
1132 | throttled = true; | |
1133 | ||
1134 | if (update_disptime) { | |
1135 | tg_update_disptime(tg, &td->service_queue); | |
1136 | throtl_schedule_next_dispatch(td); | |
1137 | } | |
1138 | ||
1139 | out_unlock: | |
1140 | spin_unlock_irq(q->queue_lock); | |
1141 | out_unlock_rcu: | |
1142 | rcu_read_unlock(); | |
1143 | out: | |
1144 | return throttled; | |
1145 | } | |
1146 | ||
1147 | /** | |
1148 | * blk_throtl_drain - drain throttled bios | |
1149 | * @q: request_queue to drain throttled bios for | |
1150 | * | |
1151 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1152 | */ | |
1153 | void blk_throtl_drain(struct request_queue *q) | |
1154 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1155 | { | |
1156 | struct throtl_data *td = q->td; | |
1157 | struct throtl_service_queue *parent_sq = &td->service_queue; | |
1158 | struct throtl_grp *tg; | |
1159 | struct bio_list bl; | |
1160 | struct bio *bio; | |
1161 | ||
1162 | queue_lockdep_assert_held(q); | |
1163 | ||
1164 | bio_list_init(&bl); | |
1165 | ||
1166 | while ((tg = throtl_rb_first(parent_sq))) { | |
1167 | struct throtl_service_queue *sq = &tg->service_queue; | |
1168 | ||
1169 | throtl_dequeue_tg(tg, parent_sq); | |
1170 | ||
1171 | while ((bio = bio_list_peek(&sq->bio_lists[READ]))) | |
1172 | tg_dispatch_one_bio(tg, bio_data_dir(bio), &bl); | |
1173 | while ((bio = bio_list_peek(&sq->bio_lists[WRITE]))) | |
1174 | tg_dispatch_one_bio(tg, bio_data_dir(bio), &bl); | |
1175 | } | |
1176 | spin_unlock_irq(q->queue_lock); | |
1177 | ||
1178 | while ((bio = bio_list_pop(&bl))) | |
1179 | generic_make_request(bio); | |
1180 | ||
1181 | spin_lock_irq(q->queue_lock); | |
1182 | } | |
1183 | ||
1184 | int blk_throtl_init(struct request_queue *q) | |
1185 | { | |
1186 | struct throtl_data *td; | |
1187 | int ret; | |
1188 | ||
1189 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1190 | if (!td) | |
1191 | return -ENOMEM; | |
1192 | ||
1193 | INIT_DELAYED_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); | |
1194 | throtl_service_queue_init(&td->service_queue); | |
1195 | ||
1196 | q->td = td; | |
1197 | td->queue = q; | |
1198 | ||
1199 | /* activate policy */ | |
1200 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); | |
1201 | if (ret) | |
1202 | kfree(td); | |
1203 | return ret; | |
1204 | } | |
1205 | ||
1206 | void blk_throtl_exit(struct request_queue *q) | |
1207 | { | |
1208 | BUG_ON(!q->td); | |
1209 | throtl_shutdown_wq(q); | |
1210 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); | |
1211 | kfree(q->td); | |
1212 | } | |
1213 | ||
1214 | static int __init throtl_init(void) | |
1215 | { | |
1216 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); | |
1217 | if (!kthrotld_workqueue) | |
1218 | panic("Failed to create kthrotld\n"); | |
1219 | ||
1220 | return blkcg_policy_register(&blkcg_policy_throtl); | |
1221 | } | |
1222 | ||
1223 | module_init(throtl_init); |