]>
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 blkio_policy_type blkio_policy_throtl; | |
25 | ||
26 | /* A workqueue to queue throttle related work */ | |
27 | static struct workqueue_struct *kthrotld_workqueue; | |
28 | static void throtl_schedule_delayed_work(struct throtl_data *td, | |
29 | unsigned long delay); | |
30 | ||
31 | struct throtl_rb_root { | |
32 | struct rb_root rb; | |
33 | struct rb_node *left; | |
34 | unsigned int count; | |
35 | unsigned long min_disptime; | |
36 | }; | |
37 | ||
38 | #define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \ | |
39 | .count = 0, .min_disptime = 0} | |
40 | ||
41 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) | |
42 | ||
43 | /* Per-cpu group stats */ | |
44 | struct tg_stats_cpu { | |
45 | /* total bytes transferred */ | |
46 | struct blkg_rwstat service_bytes; | |
47 | /* total IOs serviced, post merge */ | |
48 | struct blkg_rwstat serviced; | |
49 | }; | |
50 | ||
51 | struct throtl_grp { | |
52 | /* active throtl group service_tree member */ | |
53 | struct rb_node rb_node; | |
54 | ||
55 | /* | |
56 | * Dispatch time in jiffies. This is the estimated time when group | |
57 | * will unthrottle and is ready to dispatch more bio. It is used as | |
58 | * key to sort active groups in service tree. | |
59 | */ | |
60 | unsigned long disptime; | |
61 | ||
62 | unsigned int flags; | |
63 | ||
64 | /* Two lists for READ and WRITE */ | |
65 | struct bio_list bio_lists[2]; | |
66 | ||
67 | /* Number of queued bios on READ and WRITE lists */ | |
68 | unsigned int nr_queued[2]; | |
69 | ||
70 | /* bytes per second rate limits */ | |
71 | uint64_t bps[2]; | |
72 | ||
73 | /* IOPS limits */ | |
74 | unsigned int iops[2]; | |
75 | ||
76 | /* Number of bytes disptached in current slice */ | |
77 | uint64_t bytes_disp[2]; | |
78 | /* Number of bio's dispatched in current slice */ | |
79 | unsigned int io_disp[2]; | |
80 | ||
81 | /* When did we start a new slice */ | |
82 | unsigned long slice_start[2]; | |
83 | unsigned long slice_end[2]; | |
84 | ||
85 | /* Some throttle limits got updated for the group */ | |
86 | int limits_changed; | |
87 | ||
88 | /* Per cpu stats pointer */ | |
89 | struct tg_stats_cpu __percpu *stats_cpu; | |
90 | ||
91 | /* List of tgs waiting for per cpu stats memory to be allocated */ | |
92 | struct list_head stats_alloc_node; | |
93 | }; | |
94 | ||
95 | struct throtl_data | |
96 | { | |
97 | /* service tree for active throtl groups */ | |
98 | struct throtl_rb_root tg_service_tree; | |
99 | ||
100 | struct throtl_grp *root_tg; | |
101 | struct request_queue *queue; | |
102 | ||
103 | /* Total Number of queued bios on READ and WRITE lists */ | |
104 | unsigned int nr_queued[2]; | |
105 | ||
106 | /* | |
107 | * number of total undestroyed groups | |
108 | */ | |
109 | unsigned int nr_undestroyed_grps; | |
110 | ||
111 | /* Work for dispatching throttled bios */ | |
112 | struct delayed_work throtl_work; | |
113 | ||
114 | int limits_changed; | |
115 | }; | |
116 | ||
117 | /* list and work item to allocate percpu group stats */ | |
118 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
119 | static LIST_HEAD(tg_stats_alloc_list); | |
120 | ||
121 | static void tg_stats_alloc_fn(struct work_struct *); | |
122 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
123 | ||
124 | static inline struct throtl_grp *blkg_to_tg(struct blkio_group *blkg) | |
125 | { | |
126 | return blkg_to_pdata(blkg, &blkio_policy_throtl); | |
127 | } | |
128 | ||
129 | static inline struct blkio_group *tg_to_blkg(struct throtl_grp *tg) | |
130 | { | |
131 | return pdata_to_blkg(tg); | |
132 | } | |
133 | ||
134 | enum tg_state_flags { | |
135 | THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */ | |
136 | }; | |
137 | ||
138 | #define THROTL_TG_FNS(name) \ | |
139 | static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \ | |
140 | { \ | |
141 | (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \ | |
142 | } \ | |
143 | static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \ | |
144 | { \ | |
145 | (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \ | |
146 | } \ | |
147 | static inline int throtl_tg_##name(const struct throtl_grp *tg) \ | |
148 | { \ | |
149 | return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \ | |
150 | } | |
151 | ||
152 | THROTL_TG_FNS(on_rr); | |
153 | ||
154 | #define throtl_log_tg(td, tg, fmt, args...) \ | |
155 | blk_add_trace_msg((td)->queue, "throtl %s " fmt, \ | |
156 | blkg_path(tg_to_blkg(tg)), ##args); \ | |
157 | ||
158 | #define throtl_log(td, fmt, args...) \ | |
159 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) | |
160 | ||
161 | static inline unsigned int total_nr_queued(struct throtl_data *td) | |
162 | { | |
163 | return td->nr_queued[0] + td->nr_queued[1]; | |
164 | } | |
165 | ||
166 | /* | |
167 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
168 | * system_nrt_wq once there are some groups on the alloc_list waiting for | |
169 | * allocation. | |
170 | */ | |
171 | static void tg_stats_alloc_fn(struct work_struct *work) | |
172 | { | |
173 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
174 | struct delayed_work *dwork = to_delayed_work(work); | |
175 | bool empty = false; | |
176 | ||
177 | alloc_stats: | |
178 | if (!stats_cpu) { | |
179 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
180 | if (!stats_cpu) { | |
181 | /* allocation failed, try again after some time */ | |
182 | queue_delayed_work(system_nrt_wq, dwork, | |
183 | msecs_to_jiffies(10)); | |
184 | return; | |
185 | } | |
186 | } | |
187 | ||
188 | spin_lock_irq(&tg_stats_alloc_lock); | |
189 | ||
190 | if (!list_empty(&tg_stats_alloc_list)) { | |
191 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
192 | struct throtl_grp, | |
193 | stats_alloc_node); | |
194 | swap(tg->stats_cpu, stats_cpu); | |
195 | list_del_init(&tg->stats_alloc_node); | |
196 | } | |
197 | ||
198 | empty = list_empty(&tg_stats_alloc_list); | |
199 | spin_unlock_irq(&tg_stats_alloc_lock); | |
200 | if (!empty) | |
201 | goto alloc_stats; | |
202 | } | |
203 | ||
204 | static void throtl_init_blkio_group(struct blkio_group *blkg) | |
205 | { | |
206 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
207 | ||
208 | RB_CLEAR_NODE(&tg->rb_node); | |
209 | bio_list_init(&tg->bio_lists[0]); | |
210 | bio_list_init(&tg->bio_lists[1]); | |
211 | tg->limits_changed = false; | |
212 | ||
213 | tg->bps[READ] = -1; | |
214 | tg->bps[WRITE] = -1; | |
215 | tg->iops[READ] = -1; | |
216 | tg->iops[WRITE] = -1; | |
217 | ||
218 | /* | |
219 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
220 | * but percpu allocator can't be called from IO path. Queue tg on | |
221 | * tg_stats_alloc_list and allocate from work item. | |
222 | */ | |
223 | spin_lock(&tg_stats_alloc_lock); | |
224 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); | |
225 | queue_delayed_work(system_nrt_wq, &tg_stats_alloc_work, 0); | |
226 | spin_unlock(&tg_stats_alloc_lock); | |
227 | } | |
228 | ||
229 | static void throtl_exit_blkio_group(struct blkio_group *blkg) | |
230 | { | |
231 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
232 | ||
233 | spin_lock(&tg_stats_alloc_lock); | |
234 | list_del_init(&tg->stats_alloc_node); | |
235 | spin_unlock(&tg_stats_alloc_lock); | |
236 | ||
237 | free_percpu(tg->stats_cpu); | |
238 | } | |
239 | ||
240 | static void throtl_reset_group_stats(struct blkio_group *blkg) | |
241 | { | |
242 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
243 | int cpu; | |
244 | ||
245 | if (tg->stats_cpu == NULL) | |
246 | return; | |
247 | ||
248 | for_each_possible_cpu(cpu) { | |
249 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
250 | ||
251 | blkg_rwstat_reset(&sc->service_bytes); | |
252 | blkg_rwstat_reset(&sc->serviced); | |
253 | } | |
254 | } | |
255 | ||
256 | static struct | |
257 | throtl_grp *throtl_lookup_tg(struct throtl_data *td, struct blkio_cgroup *blkcg) | |
258 | { | |
259 | /* | |
260 | * This is the common case when there are no blkio cgroups. | |
261 | * Avoid lookup in this case | |
262 | */ | |
263 | if (blkcg == &blkio_root_cgroup) | |
264 | return td->root_tg; | |
265 | ||
266 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); | |
267 | } | |
268 | ||
269 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, | |
270 | struct blkio_cgroup *blkcg) | |
271 | { | |
272 | struct request_queue *q = td->queue; | |
273 | struct throtl_grp *tg = NULL; | |
274 | ||
275 | /* | |
276 | * This is the common case when there are no blkio cgroups. | |
277 | * Avoid lookup in this case | |
278 | */ | |
279 | if (blkcg == &blkio_root_cgroup) { | |
280 | tg = td->root_tg; | |
281 | } else { | |
282 | struct blkio_group *blkg; | |
283 | ||
284 | blkg = blkg_lookup_create(blkcg, q, false); | |
285 | ||
286 | /* if %NULL and @q is alive, fall back to root_tg */ | |
287 | if (!IS_ERR(blkg)) | |
288 | tg = blkg_to_tg(blkg); | |
289 | else if (!blk_queue_dead(q)) | |
290 | tg = td->root_tg; | |
291 | } | |
292 | ||
293 | return tg; | |
294 | } | |
295 | ||
296 | static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root) | |
297 | { | |
298 | /* Service tree is empty */ | |
299 | if (!root->count) | |
300 | return NULL; | |
301 | ||
302 | if (!root->left) | |
303 | root->left = rb_first(&root->rb); | |
304 | ||
305 | if (root->left) | |
306 | return rb_entry_tg(root->left); | |
307 | ||
308 | return NULL; | |
309 | } | |
310 | ||
311 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
312 | { | |
313 | rb_erase(n, root); | |
314 | RB_CLEAR_NODE(n); | |
315 | } | |
316 | ||
317 | static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root) | |
318 | { | |
319 | if (root->left == n) | |
320 | root->left = NULL; | |
321 | rb_erase_init(n, &root->rb); | |
322 | --root->count; | |
323 | } | |
324 | ||
325 | static void update_min_dispatch_time(struct throtl_rb_root *st) | |
326 | { | |
327 | struct throtl_grp *tg; | |
328 | ||
329 | tg = throtl_rb_first(st); | |
330 | if (!tg) | |
331 | return; | |
332 | ||
333 | st->min_disptime = tg->disptime; | |
334 | } | |
335 | ||
336 | static void | |
337 | tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg) | |
338 | { | |
339 | struct rb_node **node = &st->rb.rb_node; | |
340 | struct rb_node *parent = NULL; | |
341 | struct throtl_grp *__tg; | |
342 | unsigned long key = tg->disptime; | |
343 | int left = 1; | |
344 | ||
345 | while (*node != NULL) { | |
346 | parent = *node; | |
347 | __tg = rb_entry_tg(parent); | |
348 | ||
349 | if (time_before(key, __tg->disptime)) | |
350 | node = &parent->rb_left; | |
351 | else { | |
352 | node = &parent->rb_right; | |
353 | left = 0; | |
354 | } | |
355 | } | |
356 | ||
357 | if (left) | |
358 | st->left = &tg->rb_node; | |
359 | ||
360 | rb_link_node(&tg->rb_node, parent, node); | |
361 | rb_insert_color(&tg->rb_node, &st->rb); | |
362 | } | |
363 | ||
364 | static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
365 | { | |
366 | struct throtl_rb_root *st = &td->tg_service_tree; | |
367 | ||
368 | tg_service_tree_add(st, tg); | |
369 | throtl_mark_tg_on_rr(tg); | |
370 | st->count++; | |
371 | } | |
372 | ||
373 | static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
374 | { | |
375 | if (!throtl_tg_on_rr(tg)) | |
376 | __throtl_enqueue_tg(td, tg); | |
377 | } | |
378 | ||
379 | static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
380 | { | |
381 | throtl_rb_erase(&tg->rb_node, &td->tg_service_tree); | |
382 | throtl_clear_tg_on_rr(tg); | |
383 | } | |
384 | ||
385 | static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
386 | { | |
387 | if (throtl_tg_on_rr(tg)) | |
388 | __throtl_dequeue_tg(td, tg); | |
389 | } | |
390 | ||
391 | static void throtl_schedule_next_dispatch(struct throtl_data *td) | |
392 | { | |
393 | struct throtl_rb_root *st = &td->tg_service_tree; | |
394 | ||
395 | /* | |
396 | * If there are more bios pending, schedule more work. | |
397 | */ | |
398 | if (!total_nr_queued(td)) | |
399 | return; | |
400 | ||
401 | BUG_ON(!st->count); | |
402 | ||
403 | update_min_dispatch_time(st); | |
404 | ||
405 | if (time_before_eq(st->min_disptime, jiffies)) | |
406 | throtl_schedule_delayed_work(td, 0); | |
407 | else | |
408 | throtl_schedule_delayed_work(td, (st->min_disptime - jiffies)); | |
409 | } | |
410 | ||
411 | static inline void | |
412 | throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
413 | { | |
414 | tg->bytes_disp[rw] = 0; | |
415 | tg->io_disp[rw] = 0; | |
416 | tg->slice_start[rw] = jiffies; | |
417 | tg->slice_end[rw] = jiffies + throtl_slice; | |
418 | throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
419 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
420 | tg->slice_end[rw], jiffies); | |
421 | } | |
422 | ||
423 | static inline void throtl_set_slice_end(struct throtl_data *td, | |
424 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
425 | { | |
426 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
427 | } | |
428 | ||
429 | static inline void throtl_extend_slice(struct throtl_data *td, | |
430 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
431 | { | |
432 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
433 | throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
434 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
435 | tg->slice_end[rw], jiffies); | |
436 | } | |
437 | ||
438 | /* Determine if previously allocated or extended slice is complete or not */ | |
439 | static bool | |
440 | throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
441 | { | |
442 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
443 | return 0; | |
444 | ||
445 | return 1; | |
446 | } | |
447 | ||
448 | /* Trim the used slices and adjust slice start accordingly */ | |
449 | static inline void | |
450 | throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
451 | { | |
452 | unsigned long nr_slices, time_elapsed, io_trim; | |
453 | u64 bytes_trim, tmp; | |
454 | ||
455 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
456 | ||
457 | /* | |
458 | * If bps are unlimited (-1), then time slice don't get | |
459 | * renewed. Don't try to trim the slice if slice is used. A new | |
460 | * slice will start when appropriate. | |
461 | */ | |
462 | if (throtl_slice_used(td, tg, rw)) | |
463 | return; | |
464 | ||
465 | /* | |
466 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
467 | * that initially cgroup limit was very low resulting in high | |
468 | * slice_end, but later limit was bumped up and bio was dispached | |
469 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
470 | * is bad because it does not allow new slice to start. | |
471 | */ | |
472 | ||
473 | throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice); | |
474 | ||
475 | time_elapsed = jiffies - tg->slice_start[rw]; | |
476 | ||
477 | nr_slices = time_elapsed / throtl_slice; | |
478 | ||
479 | if (!nr_slices) | |
480 | return; | |
481 | tmp = tg->bps[rw] * throtl_slice * nr_slices; | |
482 | do_div(tmp, HZ); | |
483 | bytes_trim = tmp; | |
484 | ||
485 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; | |
486 | ||
487 | if (!bytes_trim && !io_trim) | |
488 | return; | |
489 | ||
490 | if (tg->bytes_disp[rw] >= bytes_trim) | |
491 | tg->bytes_disp[rw] -= bytes_trim; | |
492 | else | |
493 | tg->bytes_disp[rw] = 0; | |
494 | ||
495 | if (tg->io_disp[rw] >= io_trim) | |
496 | tg->io_disp[rw] -= io_trim; | |
497 | else | |
498 | tg->io_disp[rw] = 0; | |
499 | ||
500 | tg->slice_start[rw] += nr_slices * throtl_slice; | |
501 | ||
502 | throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" | |
503 | " start=%lu end=%lu jiffies=%lu", | |
504 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
505 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
506 | } | |
507 | ||
508 | static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg, | |
509 | struct bio *bio, unsigned long *wait) | |
510 | { | |
511 | bool rw = bio_data_dir(bio); | |
512 | unsigned int io_allowed; | |
513 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; | |
514 | u64 tmp; | |
515 | ||
516 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
517 | ||
518 | /* Slice has just started. Consider one slice interval */ | |
519 | if (!jiffy_elapsed) | |
520 | jiffy_elapsed_rnd = throtl_slice; | |
521 | ||
522 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
523 | ||
524 | /* | |
525 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
526 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
527 | * will allow dispatch after 1 second and after that slice should | |
528 | * have been trimmed. | |
529 | */ | |
530 | ||
531 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
532 | do_div(tmp, HZ); | |
533 | ||
534 | if (tmp > UINT_MAX) | |
535 | io_allowed = UINT_MAX; | |
536 | else | |
537 | io_allowed = tmp; | |
538 | ||
539 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
540 | if (wait) | |
541 | *wait = 0; | |
542 | return 1; | |
543 | } | |
544 | ||
545 | /* Calc approx time to dispatch */ | |
546 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
547 | ||
548 | if (jiffy_wait > jiffy_elapsed) | |
549 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
550 | else | |
551 | jiffy_wait = 1; | |
552 | ||
553 | if (wait) | |
554 | *wait = jiffy_wait; | |
555 | return 0; | |
556 | } | |
557 | ||
558 | static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg, | |
559 | struct bio *bio, unsigned long *wait) | |
560 | { | |
561 | bool rw = bio_data_dir(bio); | |
562 | u64 bytes_allowed, extra_bytes, tmp; | |
563 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; | |
564 | ||
565 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
566 | ||
567 | /* Slice has just started. Consider one slice interval */ | |
568 | if (!jiffy_elapsed) | |
569 | jiffy_elapsed_rnd = throtl_slice; | |
570 | ||
571 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
572 | ||
573 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; | |
574 | do_div(tmp, HZ); | |
575 | bytes_allowed = tmp; | |
576 | ||
577 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
578 | if (wait) | |
579 | *wait = 0; | |
580 | return 1; | |
581 | } | |
582 | ||
583 | /* Calc approx time to dispatch */ | |
584 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
585 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
586 | ||
587 | if (!jiffy_wait) | |
588 | jiffy_wait = 1; | |
589 | ||
590 | /* | |
591 | * This wait time is without taking into consideration the rounding | |
592 | * up we did. Add that time also. | |
593 | */ | |
594 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
595 | if (wait) | |
596 | *wait = jiffy_wait; | |
597 | return 0; | |
598 | } | |
599 | ||
600 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { | |
601 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
602 | return 1; | |
603 | return 0; | |
604 | } | |
605 | ||
606 | /* | |
607 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
608 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
609 | */ | |
610 | static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg, | |
611 | struct bio *bio, unsigned long *wait) | |
612 | { | |
613 | bool rw = bio_data_dir(bio); | |
614 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
615 | ||
616 | /* | |
617 | * Currently whole state machine of group depends on first bio | |
618 | * queued in the group bio list. So one should not be calling | |
619 | * this function with a different bio if there are other bios | |
620 | * queued. | |
621 | */ | |
622 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); | |
623 | ||
624 | /* If tg->bps = -1, then BW is unlimited */ | |
625 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
626 | if (wait) | |
627 | *wait = 0; | |
628 | return 1; | |
629 | } | |
630 | ||
631 | /* | |
632 | * If previous slice expired, start a new one otherwise renew/extend | |
633 | * existing slice to make sure it is at least throtl_slice interval | |
634 | * long since now. | |
635 | */ | |
636 | if (throtl_slice_used(td, tg, rw)) | |
637 | throtl_start_new_slice(td, tg, rw); | |
638 | else { | |
639 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
640 | throtl_extend_slice(td, tg, rw, jiffies + throtl_slice); | |
641 | } | |
642 | ||
643 | if (tg_with_in_bps_limit(td, tg, bio, &bps_wait) | |
644 | && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) { | |
645 | if (wait) | |
646 | *wait = 0; | |
647 | return 1; | |
648 | } | |
649 | ||
650 | max_wait = max(bps_wait, iops_wait); | |
651 | ||
652 | if (wait) | |
653 | *wait = max_wait; | |
654 | ||
655 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
656 | throtl_extend_slice(td, tg, rw, jiffies + max_wait); | |
657 | ||
658 | return 0; | |
659 | } | |
660 | ||
661 | static void throtl_update_dispatch_stats(struct blkio_group *blkg, u64 bytes, | |
662 | int rw) | |
663 | { | |
664 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
665 | struct tg_stats_cpu *stats_cpu; | |
666 | unsigned long flags; | |
667 | ||
668 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
669 | if (tg->stats_cpu == NULL) | |
670 | return; | |
671 | ||
672 | /* | |
673 | * Disabling interrupts to provide mutual exclusion between two | |
674 | * writes on same cpu. It probably is not needed for 64bit. Not | |
675 | * optimizing that case yet. | |
676 | */ | |
677 | local_irq_save(flags); | |
678 | ||
679 | stats_cpu = this_cpu_ptr(tg->stats_cpu); | |
680 | ||
681 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); | |
682 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
683 | ||
684 | local_irq_restore(flags); | |
685 | } | |
686 | ||
687 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
688 | { | |
689 | bool rw = bio_data_dir(bio); | |
690 | ||
691 | /* Charge the bio to the group */ | |
692 | tg->bytes_disp[rw] += bio->bi_size; | |
693 | tg->io_disp[rw]++; | |
694 | ||
695 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw); | |
696 | } | |
697 | ||
698 | static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg, | |
699 | struct bio *bio) | |
700 | { | |
701 | bool rw = bio_data_dir(bio); | |
702 | ||
703 | bio_list_add(&tg->bio_lists[rw], bio); | |
704 | /* Take a bio reference on tg */ | |
705 | blkg_get(tg_to_blkg(tg)); | |
706 | tg->nr_queued[rw]++; | |
707 | td->nr_queued[rw]++; | |
708 | throtl_enqueue_tg(td, tg); | |
709 | } | |
710 | ||
711 | static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg) | |
712 | { | |
713 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
714 | struct bio *bio; | |
715 | ||
716 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
717 | tg_may_dispatch(td, tg, bio, &read_wait); | |
718 | ||
719 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
720 | tg_may_dispatch(td, tg, bio, &write_wait); | |
721 | ||
722 | min_wait = min(read_wait, write_wait); | |
723 | disptime = jiffies + min_wait; | |
724 | ||
725 | /* Update dispatch time */ | |
726 | throtl_dequeue_tg(td, tg); | |
727 | tg->disptime = disptime; | |
728 | throtl_enqueue_tg(td, tg); | |
729 | } | |
730 | ||
731 | static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg, | |
732 | bool rw, struct bio_list *bl) | |
733 | { | |
734 | struct bio *bio; | |
735 | ||
736 | bio = bio_list_pop(&tg->bio_lists[rw]); | |
737 | tg->nr_queued[rw]--; | |
738 | /* Drop bio reference on blkg */ | |
739 | blkg_put(tg_to_blkg(tg)); | |
740 | ||
741 | BUG_ON(td->nr_queued[rw] <= 0); | |
742 | td->nr_queued[rw]--; | |
743 | ||
744 | throtl_charge_bio(tg, bio); | |
745 | bio_list_add(bl, bio); | |
746 | bio->bi_rw |= REQ_THROTTLED; | |
747 | ||
748 | throtl_trim_slice(td, tg, rw); | |
749 | } | |
750 | ||
751 | static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg, | |
752 | struct bio_list *bl) | |
753 | { | |
754 | unsigned int nr_reads = 0, nr_writes = 0; | |
755 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
756 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; | |
757 | struct bio *bio; | |
758 | ||
759 | /* Try to dispatch 75% READS and 25% WRITES */ | |
760 | ||
761 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) | |
762 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
763 | ||
764 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
765 | nr_reads++; | |
766 | ||
767 | if (nr_reads >= max_nr_reads) | |
768 | break; | |
769 | } | |
770 | ||
771 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) | |
772 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
773 | ||
774 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
775 | nr_writes++; | |
776 | ||
777 | if (nr_writes >= max_nr_writes) | |
778 | break; | |
779 | } | |
780 | ||
781 | return nr_reads + nr_writes; | |
782 | } | |
783 | ||
784 | static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl) | |
785 | { | |
786 | unsigned int nr_disp = 0; | |
787 | struct throtl_grp *tg; | |
788 | struct throtl_rb_root *st = &td->tg_service_tree; | |
789 | ||
790 | while (1) { | |
791 | tg = throtl_rb_first(st); | |
792 | ||
793 | if (!tg) | |
794 | break; | |
795 | ||
796 | if (time_before(jiffies, tg->disptime)) | |
797 | break; | |
798 | ||
799 | throtl_dequeue_tg(td, tg); | |
800 | ||
801 | nr_disp += throtl_dispatch_tg(td, tg, bl); | |
802 | ||
803 | if (tg->nr_queued[0] || tg->nr_queued[1]) { | |
804 | tg_update_disptime(td, tg); | |
805 | throtl_enqueue_tg(td, tg); | |
806 | } | |
807 | ||
808 | if (nr_disp >= throtl_quantum) | |
809 | break; | |
810 | } | |
811 | ||
812 | return nr_disp; | |
813 | } | |
814 | ||
815 | static void throtl_process_limit_change(struct throtl_data *td) | |
816 | { | |
817 | struct request_queue *q = td->queue; | |
818 | struct blkio_group *blkg, *n; | |
819 | ||
820 | if (!td->limits_changed) | |
821 | return; | |
822 | ||
823 | xchg(&td->limits_changed, false); | |
824 | ||
825 | throtl_log(td, "limits changed"); | |
826 | ||
827 | list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) { | |
828 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
829 | ||
830 | if (!tg->limits_changed) | |
831 | continue; | |
832 | ||
833 | if (!xchg(&tg->limits_changed, false)) | |
834 | continue; | |
835 | ||
836 | throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu" | |
837 | " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE], | |
838 | tg->iops[READ], tg->iops[WRITE]); | |
839 | ||
840 | /* | |
841 | * Restart the slices for both READ and WRITES. It | |
842 | * might happen that a group's limit are dropped | |
843 | * suddenly and we don't want to account recently | |
844 | * dispatched IO with new low rate | |
845 | */ | |
846 | throtl_start_new_slice(td, tg, 0); | |
847 | throtl_start_new_slice(td, tg, 1); | |
848 | ||
849 | if (throtl_tg_on_rr(tg)) | |
850 | tg_update_disptime(td, tg); | |
851 | } | |
852 | } | |
853 | ||
854 | /* Dispatch throttled bios. Should be called without queue lock held. */ | |
855 | static int throtl_dispatch(struct request_queue *q) | |
856 | { | |
857 | struct throtl_data *td = q->td; | |
858 | unsigned int nr_disp = 0; | |
859 | struct bio_list bio_list_on_stack; | |
860 | struct bio *bio; | |
861 | struct blk_plug plug; | |
862 | ||
863 | spin_lock_irq(q->queue_lock); | |
864 | ||
865 | throtl_process_limit_change(td); | |
866 | ||
867 | if (!total_nr_queued(td)) | |
868 | goto out; | |
869 | ||
870 | bio_list_init(&bio_list_on_stack); | |
871 | ||
872 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", | |
873 | total_nr_queued(td), td->nr_queued[READ], | |
874 | td->nr_queued[WRITE]); | |
875 | ||
876 | nr_disp = throtl_select_dispatch(td, &bio_list_on_stack); | |
877 | ||
878 | if (nr_disp) | |
879 | throtl_log(td, "bios disp=%u", nr_disp); | |
880 | ||
881 | throtl_schedule_next_dispatch(td); | |
882 | out: | |
883 | spin_unlock_irq(q->queue_lock); | |
884 | ||
885 | /* | |
886 | * If we dispatched some requests, unplug the queue to make sure | |
887 | * immediate dispatch | |
888 | */ | |
889 | if (nr_disp) { | |
890 | blk_start_plug(&plug); | |
891 | while((bio = bio_list_pop(&bio_list_on_stack))) | |
892 | generic_make_request(bio); | |
893 | blk_finish_plug(&plug); | |
894 | } | |
895 | return nr_disp; | |
896 | } | |
897 | ||
898 | void blk_throtl_work(struct work_struct *work) | |
899 | { | |
900 | struct throtl_data *td = container_of(work, struct throtl_data, | |
901 | throtl_work.work); | |
902 | struct request_queue *q = td->queue; | |
903 | ||
904 | throtl_dispatch(q); | |
905 | } | |
906 | ||
907 | /* Call with queue lock held */ | |
908 | static void | |
909 | throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay) | |
910 | { | |
911 | ||
912 | struct delayed_work *dwork = &td->throtl_work; | |
913 | ||
914 | /* schedule work if limits changed even if no bio is queued */ | |
915 | if (total_nr_queued(td) || td->limits_changed) { | |
916 | /* | |
917 | * We might have a work scheduled to be executed in future. | |
918 | * Cancel that and schedule a new one. | |
919 | */ | |
920 | __cancel_delayed_work(dwork); | |
921 | queue_delayed_work(kthrotld_workqueue, dwork, delay); | |
922 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", | |
923 | delay, jiffies); | |
924 | } | |
925 | } | |
926 | ||
927 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, void *pdata, int off) | |
928 | { | |
929 | struct throtl_grp *tg = pdata; | |
930 | struct blkg_rwstat rwstat = { }, tmp; | |
931 | int i, cpu; | |
932 | ||
933 | for_each_possible_cpu(cpu) { | |
934 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
935 | ||
936 | tmp = blkg_rwstat_read((void *)sc + off); | |
937 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
938 | rwstat.cnt[i] += tmp.cnt[i]; | |
939 | } | |
940 | ||
941 | return __blkg_prfill_rwstat(sf, pdata, &rwstat); | |
942 | } | |
943 | ||
944 | /* print per-cpu blkg_rwstat specified by BLKCG_STAT_PRIV() */ | |
945 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, | |
946 | struct seq_file *sf) | |
947 | { | |
948 | struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgrp); | |
949 | ||
950 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, | |
951 | BLKCG_STAT_POL(cft->private), | |
952 | BLKCG_STAT_OFF(cft->private), true); | |
953 | return 0; | |
954 | } | |
955 | ||
956 | static u64 tg_prfill_conf_u64(struct seq_file *sf, void *pdata, int off) | |
957 | { | |
958 | u64 v = *(u64 *)(pdata + off); | |
959 | ||
960 | if (v == -1) | |
961 | return 0; | |
962 | return __blkg_prfill_u64(sf, pdata, v); | |
963 | } | |
964 | ||
965 | static u64 tg_prfill_conf_uint(struct seq_file *sf, void *pdata, int off) | |
966 | { | |
967 | unsigned int v = *(unsigned int *)(pdata + off); | |
968 | ||
969 | if (v == -1) | |
970 | return 0; | |
971 | return __blkg_prfill_u64(sf, pdata, v); | |
972 | } | |
973 | ||
974 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, | |
975 | struct seq_file *sf) | |
976 | { | |
977 | blkcg_print_blkgs(sf, cgroup_to_blkio_cgroup(cgrp), tg_prfill_conf_u64, | |
978 | BLKIO_POLICY_THROTL, cft->private, false); | |
979 | return 0; | |
980 | } | |
981 | ||
982 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, | |
983 | struct seq_file *sf) | |
984 | { | |
985 | blkcg_print_blkgs(sf, cgroup_to_blkio_cgroup(cgrp), tg_prfill_conf_uint, | |
986 | BLKIO_POLICY_THROTL, cft->private, false); | |
987 | return 0; | |
988 | } | |
989 | ||
990 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, | |
991 | bool is_u64) | |
992 | { | |
993 | struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgrp); | |
994 | struct blkg_conf_ctx ctx; | |
995 | struct throtl_grp *tg; | |
996 | int ret; | |
997 | ||
998 | ret = blkg_conf_prep(blkcg, buf, &ctx); | |
999 | if (ret) | |
1000 | return ret; | |
1001 | ||
1002 | ret = -EINVAL; | |
1003 | tg = blkg_to_tg(ctx.blkg); | |
1004 | if (tg) { | |
1005 | struct throtl_data *td = ctx.blkg->q->td; | |
1006 | ||
1007 | if (!ctx.v) | |
1008 | ctx.v = -1; | |
1009 | ||
1010 | if (is_u64) | |
1011 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
1012 | else | |
1013 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
1014 | ||
1015 | /* XXX: we don't need the following deferred processing */ | |
1016 | xchg(&tg->limits_changed, true); | |
1017 | xchg(&td->limits_changed, true); | |
1018 | throtl_schedule_delayed_work(td, 0); | |
1019 | ||
1020 | ret = 0; | |
1021 | } | |
1022 | ||
1023 | blkg_conf_finish(&ctx); | |
1024 | return ret; | |
1025 | } | |
1026 | ||
1027 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, | |
1028 | const char *buf) | |
1029 | { | |
1030 | return tg_set_conf(cgrp, cft, buf, true); | |
1031 | } | |
1032 | ||
1033 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, | |
1034 | const char *buf) | |
1035 | { | |
1036 | return tg_set_conf(cgrp, cft, buf, false); | |
1037 | } | |
1038 | ||
1039 | static struct cftype throtl_files[] = { | |
1040 | { | |
1041 | .name = "throttle.read_bps_device", | |
1042 | .private = offsetof(struct throtl_grp, bps[READ]), | |
1043 | .read_seq_string = tg_print_conf_u64, | |
1044 | .write_string = tg_set_conf_u64, | |
1045 | .max_write_len = 256, | |
1046 | }, | |
1047 | { | |
1048 | .name = "throttle.write_bps_device", | |
1049 | .private = offsetof(struct throtl_grp, bps[WRITE]), | |
1050 | .read_seq_string = tg_print_conf_u64, | |
1051 | .write_string = tg_set_conf_u64, | |
1052 | .max_write_len = 256, | |
1053 | }, | |
1054 | { | |
1055 | .name = "throttle.read_iops_device", | |
1056 | .private = offsetof(struct throtl_grp, iops[READ]), | |
1057 | .read_seq_string = tg_print_conf_uint, | |
1058 | .write_string = tg_set_conf_uint, | |
1059 | .max_write_len = 256, | |
1060 | }, | |
1061 | { | |
1062 | .name = "throttle.write_iops_device", | |
1063 | .private = offsetof(struct throtl_grp, iops[WRITE]), | |
1064 | .read_seq_string = tg_print_conf_uint, | |
1065 | .write_string = tg_set_conf_uint, | |
1066 | .max_write_len = 256, | |
1067 | }, | |
1068 | { | |
1069 | .name = "throttle.io_service_bytes", | |
1070 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_THROTL, | |
1071 | offsetof(struct tg_stats_cpu, service_bytes)), | |
1072 | .read_seq_string = tg_print_cpu_rwstat, | |
1073 | }, | |
1074 | { | |
1075 | .name = "throttle.io_serviced", | |
1076 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_THROTL, | |
1077 | offsetof(struct tg_stats_cpu, serviced)), | |
1078 | .read_seq_string = tg_print_cpu_rwstat, | |
1079 | }, | |
1080 | { } /* terminate */ | |
1081 | }; | |
1082 | ||
1083 | static void throtl_shutdown_wq(struct request_queue *q) | |
1084 | { | |
1085 | struct throtl_data *td = q->td; | |
1086 | ||
1087 | cancel_delayed_work_sync(&td->throtl_work); | |
1088 | } | |
1089 | ||
1090 | static struct blkio_policy_type blkio_policy_throtl = { | |
1091 | .ops = { | |
1092 | .blkio_init_group_fn = throtl_init_blkio_group, | |
1093 | .blkio_exit_group_fn = throtl_exit_blkio_group, | |
1094 | .blkio_reset_group_stats_fn = throtl_reset_group_stats, | |
1095 | }, | |
1096 | .plid = BLKIO_POLICY_THROTL, | |
1097 | .pdata_size = sizeof(struct throtl_grp), | |
1098 | .cftypes = throtl_files, | |
1099 | }; | |
1100 | ||
1101 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) | |
1102 | { | |
1103 | struct throtl_data *td = q->td; | |
1104 | struct throtl_grp *tg; | |
1105 | bool rw = bio_data_dir(bio), update_disptime = true; | |
1106 | struct blkio_cgroup *blkcg; | |
1107 | bool throttled = false; | |
1108 | ||
1109 | if (bio->bi_rw & REQ_THROTTLED) { | |
1110 | bio->bi_rw &= ~REQ_THROTTLED; | |
1111 | goto out; | |
1112 | } | |
1113 | ||
1114 | /* bio_associate_current() needs ioc, try creating */ | |
1115 | create_io_context(GFP_ATOMIC, q->node); | |
1116 | ||
1117 | /* | |
1118 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1119 | * basic fields like stats and io rates. If a group has no rules, | |
1120 | * just update the dispatch stats in lockless manner and return. | |
1121 | */ | |
1122 | rcu_read_lock(); | |
1123 | blkcg = bio_blkio_cgroup(bio); | |
1124 | tg = throtl_lookup_tg(td, blkcg); | |
1125 | if (tg) { | |
1126 | if (tg_no_rule_group(tg, rw)) { | |
1127 | throtl_update_dispatch_stats(tg_to_blkg(tg), | |
1128 | bio->bi_size, bio->bi_rw); | |
1129 | goto out_unlock_rcu; | |
1130 | } | |
1131 | } | |
1132 | ||
1133 | /* | |
1134 | * Either group has not been allocated yet or it is not an unlimited | |
1135 | * IO group | |
1136 | */ | |
1137 | spin_lock_irq(q->queue_lock); | |
1138 | tg = throtl_lookup_create_tg(td, blkcg); | |
1139 | if (unlikely(!tg)) | |
1140 | goto out_unlock; | |
1141 | ||
1142 | if (tg->nr_queued[rw]) { | |
1143 | /* | |
1144 | * There is already another bio queued in same dir. No | |
1145 | * need to update dispatch time. | |
1146 | */ | |
1147 | update_disptime = false; | |
1148 | goto queue_bio; | |
1149 | ||
1150 | } | |
1151 | ||
1152 | /* Bio is with-in rate limit of group */ | |
1153 | if (tg_may_dispatch(td, tg, bio, NULL)) { | |
1154 | throtl_charge_bio(tg, bio); | |
1155 | ||
1156 | /* | |
1157 | * We need to trim slice even when bios are not being queued | |
1158 | * otherwise it might happen that a bio is not queued for | |
1159 | * a long time and slice keeps on extending and trim is not | |
1160 | * called for a long time. Now if limits are reduced suddenly | |
1161 | * we take into account all the IO dispatched so far at new | |
1162 | * low rate and * newly queued IO gets a really long dispatch | |
1163 | * time. | |
1164 | * | |
1165 | * So keep on trimming slice even if bio is not queued. | |
1166 | */ | |
1167 | throtl_trim_slice(td, tg, rw); | |
1168 | goto out_unlock; | |
1169 | } | |
1170 | ||
1171 | queue_bio: | |
1172 | throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" | |
1173 | " iodisp=%u iops=%u queued=%d/%d", | |
1174 | rw == READ ? 'R' : 'W', | |
1175 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], | |
1176 | tg->io_disp[rw], tg->iops[rw], | |
1177 | tg->nr_queued[READ], tg->nr_queued[WRITE]); | |
1178 | ||
1179 | bio_associate_current(bio); | |
1180 | throtl_add_bio_tg(q->td, tg, bio); | |
1181 | throttled = true; | |
1182 | ||
1183 | if (update_disptime) { | |
1184 | tg_update_disptime(td, tg); | |
1185 | throtl_schedule_next_dispatch(td); | |
1186 | } | |
1187 | ||
1188 | out_unlock: | |
1189 | spin_unlock_irq(q->queue_lock); | |
1190 | out_unlock_rcu: | |
1191 | rcu_read_unlock(); | |
1192 | out: | |
1193 | return throttled; | |
1194 | } | |
1195 | ||
1196 | /** | |
1197 | * blk_throtl_drain - drain throttled bios | |
1198 | * @q: request_queue to drain throttled bios for | |
1199 | * | |
1200 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1201 | */ | |
1202 | void blk_throtl_drain(struct request_queue *q) | |
1203 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1204 | { | |
1205 | struct throtl_data *td = q->td; | |
1206 | struct throtl_rb_root *st = &td->tg_service_tree; | |
1207 | struct throtl_grp *tg; | |
1208 | struct bio_list bl; | |
1209 | struct bio *bio; | |
1210 | ||
1211 | WARN_ON_ONCE(!queue_is_locked(q)); | |
1212 | ||
1213 | bio_list_init(&bl); | |
1214 | ||
1215 | while ((tg = throtl_rb_first(st))) { | |
1216 | throtl_dequeue_tg(td, tg); | |
1217 | ||
1218 | while ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
1219 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1220 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
1221 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1222 | } | |
1223 | spin_unlock_irq(q->queue_lock); | |
1224 | ||
1225 | while ((bio = bio_list_pop(&bl))) | |
1226 | generic_make_request(bio); | |
1227 | ||
1228 | spin_lock_irq(q->queue_lock); | |
1229 | } | |
1230 | ||
1231 | int blk_throtl_init(struct request_queue *q) | |
1232 | { | |
1233 | struct throtl_data *td; | |
1234 | struct blkio_group *blkg; | |
1235 | ||
1236 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1237 | if (!td) | |
1238 | return -ENOMEM; | |
1239 | ||
1240 | td->tg_service_tree = THROTL_RB_ROOT; | |
1241 | td->limits_changed = false; | |
1242 | INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work); | |
1243 | ||
1244 | q->td = td; | |
1245 | td->queue = q; | |
1246 | ||
1247 | /* alloc and init root group. */ | |
1248 | rcu_read_lock(); | |
1249 | spin_lock_irq(q->queue_lock); | |
1250 | ||
1251 | blkg = blkg_lookup_create(&blkio_root_cgroup, q, true); | |
1252 | if (!IS_ERR(blkg)) | |
1253 | td->root_tg = blkg_to_tg(blkg); | |
1254 | ||
1255 | spin_unlock_irq(q->queue_lock); | |
1256 | rcu_read_unlock(); | |
1257 | ||
1258 | if (!td->root_tg) { | |
1259 | kfree(td); | |
1260 | return -ENOMEM; | |
1261 | } | |
1262 | return 0; | |
1263 | } | |
1264 | ||
1265 | void blk_throtl_exit(struct request_queue *q) | |
1266 | { | |
1267 | BUG_ON(!q->td); | |
1268 | throtl_shutdown_wq(q); | |
1269 | kfree(q->td); | |
1270 | } | |
1271 | ||
1272 | static int __init throtl_init(void) | |
1273 | { | |
1274 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); | |
1275 | if (!kthrotld_workqueue) | |
1276 | panic("Failed to create kthrotld\n"); | |
1277 | ||
1278 | blkio_policy_register(&blkio_policy_throtl); | |
1279 | return 0; | |
1280 | } | |
1281 | ||
1282 | module_init(throtl_init); |