]>
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 <linux/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 | /* | |
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 | ||
58 | struct throtl_service_queue { | |
59 | struct throtl_service_queue *parent_sq; /* the parent service_queue */ | |
60 | ||
61 | /* | |
62 | * Bios queued directly to this service_queue or dispatched from | |
63 | * children throtl_grp's. | |
64 | */ | |
65 | struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */ | |
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 | */ | |
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 */ | |
76 | struct timer_list pending_timer; /* fires on first_pending_disptime */ | |
77 | }; | |
78 | ||
79 | enum tg_state_flags { | |
80 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
81 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ | |
82 | }; | |
83 | ||
84 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) | |
85 | ||
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 | ||
94 | struct throtl_grp { | |
95 | /* must be the first member */ | |
96 | struct blkg_policy_data pd; | |
97 | ||
98 | /* active throtl group service_queue member */ | |
99 | struct rb_node rb_node; | |
100 | ||
101 | /* throtl_data this group belongs to */ | |
102 | struct throtl_data *td; | |
103 | ||
104 | /* this group's service queue */ | |
105 | struct throtl_service_queue service_queue; | |
106 | ||
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 | ||
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 | ||
125 | unsigned int flags; | |
126 | ||
127 | /* are there any throtl rules between this group and td? */ | |
128 | bool has_rules[2]; | |
129 | ||
130 | /* bytes per second rate limits */ | |
131 | uint64_t bps[2]; | |
132 | ||
133 | /* IOPS limits */ | |
134 | unsigned int iops[2]; | |
135 | ||
136 | /* Number of bytes disptached in current slice */ | |
137 | uint64_t bytes_disp[2]; | |
138 | /* Number of bio's dispatched in current slice */ | |
139 | unsigned int io_disp[2]; | |
140 | ||
141 | /* When did we start a new slice */ | |
142 | unsigned long slice_start[2]; | |
143 | unsigned long slice_end[2]; | |
144 | ||
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; | |
150 | }; | |
151 | ||
152 | struct throtl_data | |
153 | { | |
154 | /* service tree for active throtl groups */ | |
155 | struct throtl_service_queue service_queue; | |
156 | ||
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 | /* | |
163 | * number of total undestroyed groups | |
164 | */ | |
165 | unsigned int nr_undestroyed_grps; | |
166 | ||
167 | /* Work for dispatching throttled bios */ | |
168 | struct work_struct dispatch_work; | |
169 | }; | |
170 | ||
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 | ||
178 | static void throtl_pending_timer_fn(unsigned long arg); | |
179 | ||
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 | ||
185 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) | |
186 | { | |
187 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); | |
188 | } | |
189 | ||
190 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) | |
191 | { | |
192 | return pd_to_blkg(&tg->pd); | |
193 | } | |
194 | ||
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 | ||
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]; \ | |
251 | \ | |
252 | blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \ | |
253 | blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
254 | } else { \ | |
255 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
256 | } \ | |
257 | } while (0) | |
258 | ||
259 | static void tg_stats_init(struct tg_stats_cpu *tg_stats) | |
260 | { | |
261 | blkg_rwstat_init(&tg_stats->service_bytes); | |
262 | blkg_rwstat_init(&tg_stats->serviced); | |
263 | } | |
264 | ||
265 | /* | |
266 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
267 | * system_wq once there are some groups on the alloc_list waiting for | |
268 | * allocation. | |
269 | */ | |
270 | static void tg_stats_alloc_fn(struct work_struct *work) | |
271 | { | |
272 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
273 | struct delayed_work *dwork = to_delayed_work(work); | |
274 | bool empty = false; | |
275 | ||
276 | alloc_stats: | |
277 | if (!stats_cpu) { | |
278 | int cpu; | |
279 | ||
280 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
281 | if (!stats_cpu) { | |
282 | /* allocation failed, try again after some time */ | |
283 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); | |
284 | return; | |
285 | } | |
286 | for_each_possible_cpu(cpu) | |
287 | tg_stats_init(per_cpu_ptr(stats_cpu, cpu)); | |
288 | } | |
289 | ||
290 | spin_lock_irq(&tg_stats_alloc_lock); | |
291 | ||
292 | if (!list_empty(&tg_stats_alloc_list)) { | |
293 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
294 | struct throtl_grp, | |
295 | stats_alloc_node); | |
296 | swap(tg->stats_cpu, stats_cpu); | |
297 | list_del_init(&tg->stats_alloc_node); | |
298 | } | |
299 | ||
300 | empty = list_empty(&tg_stats_alloc_list); | |
301 | spin_unlock_irq(&tg_stats_alloc_lock); | |
302 | if (!empty) | |
303 | goto alloc_stats; | |
304 | } | |
305 | ||
306 | static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg) | |
307 | { | |
308 | INIT_LIST_HEAD(&qn->node); | |
309 | bio_list_init(&qn->bios); | |
310 | qn->tg = tg; | |
311 | } | |
312 | ||
313 | /** | |
314 | * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it | |
315 | * @bio: bio being added | |
316 | * @qn: qnode to add bio to | |
317 | * @queued: the service_queue->queued[] list @qn belongs to | |
318 | * | |
319 | * Add @bio to @qn and put @qn on @queued if it's not already on. | |
320 | * @qn->tg's reference count is bumped when @qn is activated. See the | |
321 | * comment on top of throtl_qnode definition for details. | |
322 | */ | |
323 | static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn, | |
324 | struct list_head *queued) | |
325 | { | |
326 | bio_list_add(&qn->bios, bio); | |
327 | if (list_empty(&qn->node)) { | |
328 | list_add_tail(&qn->node, queued); | |
329 | blkg_get(tg_to_blkg(qn->tg)); | |
330 | } | |
331 | } | |
332 | ||
333 | /** | |
334 | * throtl_peek_queued - peek the first bio on a qnode list | |
335 | * @queued: the qnode list to peek | |
336 | */ | |
337 | static struct bio *throtl_peek_queued(struct list_head *queued) | |
338 | { | |
339 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
340 | struct bio *bio; | |
341 | ||
342 | if (list_empty(queued)) | |
343 | return NULL; | |
344 | ||
345 | bio = bio_list_peek(&qn->bios); | |
346 | WARN_ON_ONCE(!bio); | |
347 | return bio; | |
348 | } | |
349 | ||
350 | /** | |
351 | * throtl_pop_queued - pop the first bio form a qnode list | |
352 | * @queued: the qnode list to pop a bio from | |
353 | * @tg_to_put: optional out argument for throtl_grp to put | |
354 | * | |
355 | * Pop the first bio from the qnode list @queued. After popping, the first | |
356 | * qnode is removed from @queued if empty or moved to the end of @queued so | |
357 | * that the popping order is round-robin. | |
358 | * | |
359 | * When the first qnode is removed, its associated throtl_grp should be put | |
360 | * too. If @tg_to_put is NULL, this function automatically puts it; | |
361 | * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is | |
362 | * responsible for putting it. | |
363 | */ | |
364 | static struct bio *throtl_pop_queued(struct list_head *queued, | |
365 | struct throtl_grp **tg_to_put) | |
366 | { | |
367 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
368 | struct bio *bio; | |
369 | ||
370 | if (list_empty(queued)) | |
371 | return NULL; | |
372 | ||
373 | bio = bio_list_pop(&qn->bios); | |
374 | WARN_ON_ONCE(!bio); | |
375 | ||
376 | if (bio_list_empty(&qn->bios)) { | |
377 | list_del_init(&qn->node); | |
378 | if (tg_to_put) | |
379 | *tg_to_put = qn->tg; | |
380 | else | |
381 | blkg_put(tg_to_blkg(qn->tg)); | |
382 | } else { | |
383 | list_move_tail(&qn->node, queued); | |
384 | } | |
385 | ||
386 | return bio; | |
387 | } | |
388 | ||
389 | /* init a service_queue, assumes the caller zeroed it */ | |
390 | static void throtl_service_queue_init(struct throtl_service_queue *sq, | |
391 | struct throtl_service_queue *parent_sq) | |
392 | { | |
393 | INIT_LIST_HEAD(&sq->queued[0]); | |
394 | INIT_LIST_HEAD(&sq->queued[1]); | |
395 | sq->pending_tree = RB_ROOT; | |
396 | sq->parent_sq = parent_sq; | |
397 | setup_timer(&sq->pending_timer, throtl_pending_timer_fn, | |
398 | (unsigned long)sq); | |
399 | } | |
400 | ||
401 | static void throtl_service_queue_exit(struct throtl_service_queue *sq) | |
402 | { | |
403 | del_timer_sync(&sq->pending_timer); | |
404 | } | |
405 | ||
406 | static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node) | |
407 | { | |
408 | return kzalloc_node(sizeof(struct throtl_grp), gfp, node); | |
409 | } | |
410 | ||
411 | static void throtl_pd_init(struct blkcg_gq *blkg) | |
412 | { | |
413 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
414 | struct throtl_data *td = blkg->q->td; | |
415 | struct throtl_service_queue *parent_sq; | |
416 | unsigned long flags; | |
417 | int rw; | |
418 | ||
419 | /* | |
420 | * If on the default hierarchy, we switch to properly hierarchical | |
421 | * behavior where limits on a given throtl_grp are applied to the | |
422 | * whole subtree rather than just the group itself. e.g. If 16M | |
423 | * read_bps limit is set on the root group, the whole system can't | |
424 | * exceed 16M for the device. | |
425 | * | |
426 | * If not on the default hierarchy, the broken flat hierarchy | |
427 | * behavior is retained where all throtl_grps are treated as if | |
428 | * they're all separate root groups right below throtl_data. | |
429 | * Limits of a group don't interact with limits of other groups | |
430 | * regardless of the position of the group in the hierarchy. | |
431 | */ | |
432 | parent_sq = &td->service_queue; | |
433 | ||
434 | if (cgroup_on_dfl(blkg->blkcg->css.cgroup) && blkg->parent) | |
435 | parent_sq = &blkg_to_tg(blkg->parent)->service_queue; | |
436 | ||
437 | throtl_service_queue_init(&tg->service_queue, parent_sq); | |
438 | ||
439 | for (rw = READ; rw <= WRITE; rw++) { | |
440 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); | |
441 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); | |
442 | } | |
443 | ||
444 | RB_CLEAR_NODE(&tg->rb_node); | |
445 | tg->td = td; | |
446 | ||
447 | tg->bps[READ] = -1; | |
448 | tg->bps[WRITE] = -1; | |
449 | tg->iops[READ] = -1; | |
450 | tg->iops[WRITE] = -1; | |
451 | ||
452 | /* | |
453 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
454 | * but percpu allocator can't be called from IO path. Queue tg on | |
455 | * tg_stats_alloc_list and allocate from work item. | |
456 | */ | |
457 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); | |
458 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); | |
459 | schedule_delayed_work(&tg_stats_alloc_work, 0); | |
460 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); | |
461 | } | |
462 | ||
463 | /* | |
464 | * Set has_rules[] if @tg or any of its parents have limits configured. | |
465 | * This doesn't require walking up to the top of the hierarchy as the | |
466 | * parent's has_rules[] is guaranteed to be correct. | |
467 | */ | |
468 | static void tg_update_has_rules(struct throtl_grp *tg) | |
469 | { | |
470 | struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); | |
471 | int rw; | |
472 | ||
473 | for (rw = READ; rw <= WRITE; rw++) | |
474 | tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) || | |
475 | (tg->bps[rw] != -1 || tg->iops[rw] != -1); | |
476 | } | |
477 | ||
478 | static void throtl_pd_online(struct blkcg_gq *blkg) | |
479 | { | |
480 | /* | |
481 | * We don't want new groups to escape the limits of its ancestors. | |
482 | * Update has_rules[] after a new group is brought online. | |
483 | */ | |
484 | tg_update_has_rules(blkg_to_tg(blkg)); | |
485 | } | |
486 | ||
487 | static void throtl_pd_exit(struct blkcg_gq *blkg) | |
488 | { | |
489 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
490 | unsigned long flags; | |
491 | ||
492 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); | |
493 | list_del_init(&tg->stats_alloc_node); | |
494 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); | |
495 | ||
496 | free_percpu(tg->stats_cpu); | |
497 | ||
498 | throtl_service_queue_exit(&tg->service_queue); | |
499 | } | |
500 | ||
501 | static void throtl_pd_free(struct blkg_policy_data *pd) | |
502 | { | |
503 | kfree(pd); | |
504 | } | |
505 | ||
506 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) | |
507 | { | |
508 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
509 | int cpu; | |
510 | ||
511 | if (tg->stats_cpu == NULL) | |
512 | return; | |
513 | ||
514 | for_each_possible_cpu(cpu) { | |
515 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
516 | ||
517 | blkg_rwstat_reset(&sc->service_bytes); | |
518 | blkg_rwstat_reset(&sc->serviced); | |
519 | } | |
520 | } | |
521 | ||
522 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, | |
523 | struct blkcg *blkcg) | |
524 | { | |
525 | /* | |
526 | * This is the common case when there are no blkcgs. Avoid lookup | |
527 | * in this case | |
528 | */ | |
529 | if (blkcg == &blkcg_root) | |
530 | return td_root_tg(td); | |
531 | ||
532 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); | |
533 | } | |
534 | ||
535 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, | |
536 | struct blkcg *blkcg) | |
537 | { | |
538 | struct request_queue *q = td->queue; | |
539 | struct throtl_grp *tg = NULL; | |
540 | ||
541 | /* | |
542 | * This is the common case when there are no blkcgs. Avoid lookup | |
543 | * in this case | |
544 | */ | |
545 | if (blkcg == &blkcg_root) { | |
546 | tg = td_root_tg(td); | |
547 | } else { | |
548 | struct blkcg_gq *blkg; | |
549 | ||
550 | blkg = blkg_lookup_create(blkcg, q); | |
551 | ||
552 | /* if %NULL and @q is alive, fall back to root_tg */ | |
553 | if (!IS_ERR(blkg)) | |
554 | tg = blkg_to_tg(blkg); | |
555 | else if (!blk_queue_dying(q)) | |
556 | tg = td_root_tg(td); | |
557 | } | |
558 | ||
559 | return tg; | |
560 | } | |
561 | ||
562 | static struct throtl_grp * | |
563 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
564 | { | |
565 | /* Service tree is empty */ | |
566 | if (!parent_sq->nr_pending) | |
567 | return NULL; | |
568 | ||
569 | if (!parent_sq->first_pending) | |
570 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
571 | ||
572 | if (parent_sq->first_pending) | |
573 | return rb_entry_tg(parent_sq->first_pending); | |
574 | ||
575 | return NULL; | |
576 | } | |
577 | ||
578 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
579 | { | |
580 | rb_erase(n, root); | |
581 | RB_CLEAR_NODE(n); | |
582 | } | |
583 | ||
584 | static void throtl_rb_erase(struct rb_node *n, | |
585 | struct throtl_service_queue *parent_sq) | |
586 | { | |
587 | if (parent_sq->first_pending == n) | |
588 | parent_sq->first_pending = NULL; | |
589 | rb_erase_init(n, &parent_sq->pending_tree); | |
590 | --parent_sq->nr_pending; | |
591 | } | |
592 | ||
593 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) | |
594 | { | |
595 | struct throtl_grp *tg; | |
596 | ||
597 | tg = throtl_rb_first(parent_sq); | |
598 | if (!tg) | |
599 | return; | |
600 | ||
601 | parent_sq->first_pending_disptime = tg->disptime; | |
602 | } | |
603 | ||
604 | static void tg_service_queue_add(struct throtl_grp *tg) | |
605 | { | |
606 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; | |
607 | struct rb_node **node = &parent_sq->pending_tree.rb_node; | |
608 | struct rb_node *parent = NULL; | |
609 | struct throtl_grp *__tg; | |
610 | unsigned long key = tg->disptime; | |
611 | int left = 1; | |
612 | ||
613 | while (*node != NULL) { | |
614 | parent = *node; | |
615 | __tg = rb_entry_tg(parent); | |
616 | ||
617 | if (time_before(key, __tg->disptime)) | |
618 | node = &parent->rb_left; | |
619 | else { | |
620 | node = &parent->rb_right; | |
621 | left = 0; | |
622 | } | |
623 | } | |
624 | ||
625 | if (left) | |
626 | parent_sq->first_pending = &tg->rb_node; | |
627 | ||
628 | rb_link_node(&tg->rb_node, parent, node); | |
629 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); | |
630 | } | |
631 | ||
632 | static void __throtl_enqueue_tg(struct throtl_grp *tg) | |
633 | { | |
634 | tg_service_queue_add(tg); | |
635 | tg->flags |= THROTL_TG_PENDING; | |
636 | tg->service_queue.parent_sq->nr_pending++; | |
637 | } | |
638 | ||
639 | static void throtl_enqueue_tg(struct throtl_grp *tg) | |
640 | { | |
641 | if (!(tg->flags & THROTL_TG_PENDING)) | |
642 | __throtl_enqueue_tg(tg); | |
643 | } | |
644 | ||
645 | static void __throtl_dequeue_tg(struct throtl_grp *tg) | |
646 | { | |
647 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); | |
648 | tg->flags &= ~THROTL_TG_PENDING; | |
649 | } | |
650 | ||
651 | static void throtl_dequeue_tg(struct throtl_grp *tg) | |
652 | { | |
653 | if (tg->flags & THROTL_TG_PENDING) | |
654 | __throtl_dequeue_tg(tg); | |
655 | } | |
656 | ||
657 | /* Call with queue lock held */ | |
658 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, | |
659 | unsigned long expires) | |
660 | { | |
661 | mod_timer(&sq->pending_timer, expires); | |
662 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
663 | expires - jiffies, jiffies); | |
664 | } | |
665 | ||
666 | /** | |
667 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle | |
668 | * @sq: the service_queue to schedule dispatch for | |
669 | * @force: force scheduling | |
670 | * | |
671 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the | |
672 | * dispatch time of the first pending child. Returns %true if either timer | |
673 | * is armed or there's no pending child left. %false if the current | |
674 | * dispatch window is still open and the caller should continue | |
675 | * dispatching. | |
676 | * | |
677 | * If @force is %true, the dispatch timer is always scheduled and this | |
678 | * function is guaranteed to return %true. This is to be used when the | |
679 | * caller can't dispatch itself and needs to invoke pending_timer | |
680 | * unconditionally. Note that forced scheduling is likely to induce short | |
681 | * delay before dispatch starts even if @sq->first_pending_disptime is not | |
682 | * in the future and thus shouldn't be used in hot paths. | |
683 | */ | |
684 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
685 | bool force) | |
686 | { | |
687 | /* any pending children left? */ | |
688 | if (!sq->nr_pending) | |
689 | return true; | |
690 | ||
691 | update_min_dispatch_time(sq); | |
692 | ||
693 | /* is the next dispatch time in the future? */ | |
694 | if (force || time_after(sq->first_pending_disptime, jiffies)) { | |
695 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); | |
696 | return true; | |
697 | } | |
698 | ||
699 | /* tell the caller to continue dispatching */ | |
700 | return false; | |
701 | } | |
702 | ||
703 | static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, | |
704 | bool rw, unsigned long start) | |
705 | { | |
706 | tg->bytes_disp[rw] = 0; | |
707 | tg->io_disp[rw] = 0; | |
708 | ||
709 | /* | |
710 | * Previous slice has expired. We must have trimmed it after last | |
711 | * bio dispatch. That means since start of last slice, we never used | |
712 | * that bandwidth. Do try to make use of that bandwidth while giving | |
713 | * credit. | |
714 | */ | |
715 | if (time_after_eq(start, tg->slice_start[rw])) | |
716 | tg->slice_start[rw] = start; | |
717 | ||
718 | tg->slice_end[rw] = jiffies + throtl_slice; | |
719 | throtl_log(&tg->service_queue, | |
720 | "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", | |
721 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
722 | tg->slice_end[rw], jiffies); | |
723 | } | |
724 | ||
725 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) | |
726 | { | |
727 | tg->bytes_disp[rw] = 0; | |
728 | tg->io_disp[rw] = 0; | |
729 | tg->slice_start[rw] = jiffies; | |
730 | tg->slice_end[rw] = jiffies + throtl_slice; | |
731 | throtl_log(&tg->service_queue, | |
732 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
733 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
734 | tg->slice_end[rw], jiffies); | |
735 | } | |
736 | ||
737 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, | |
738 | unsigned long jiffy_end) | |
739 | { | |
740 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
741 | } | |
742 | ||
743 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, | |
744 | unsigned long jiffy_end) | |
745 | { | |
746 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
747 | throtl_log(&tg->service_queue, | |
748 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
749 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
750 | tg->slice_end[rw], jiffies); | |
751 | } | |
752 | ||
753 | /* Determine if previously allocated or extended slice is complete or not */ | |
754 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) | |
755 | { | |
756 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
757 | return false; | |
758 | ||
759 | return 1; | |
760 | } | |
761 | ||
762 | /* Trim the used slices and adjust slice start accordingly */ | |
763 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) | |
764 | { | |
765 | unsigned long nr_slices, time_elapsed, io_trim; | |
766 | u64 bytes_trim, tmp; | |
767 | ||
768 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
769 | ||
770 | /* | |
771 | * If bps are unlimited (-1), then time slice don't get | |
772 | * renewed. Don't try to trim the slice if slice is used. A new | |
773 | * slice will start when appropriate. | |
774 | */ | |
775 | if (throtl_slice_used(tg, rw)) | |
776 | return; | |
777 | ||
778 | /* | |
779 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
780 | * that initially cgroup limit was very low resulting in high | |
781 | * slice_end, but later limit was bumped up and bio was dispached | |
782 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
783 | * is bad because it does not allow new slice to start. | |
784 | */ | |
785 | ||
786 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); | |
787 | ||
788 | time_elapsed = jiffies - tg->slice_start[rw]; | |
789 | ||
790 | nr_slices = time_elapsed / throtl_slice; | |
791 | ||
792 | if (!nr_slices) | |
793 | return; | |
794 | tmp = tg->bps[rw] * throtl_slice * nr_slices; | |
795 | do_div(tmp, HZ); | |
796 | bytes_trim = tmp; | |
797 | ||
798 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; | |
799 | ||
800 | if (!bytes_trim && !io_trim) | |
801 | return; | |
802 | ||
803 | if (tg->bytes_disp[rw] >= bytes_trim) | |
804 | tg->bytes_disp[rw] -= bytes_trim; | |
805 | else | |
806 | tg->bytes_disp[rw] = 0; | |
807 | ||
808 | if (tg->io_disp[rw] >= io_trim) | |
809 | tg->io_disp[rw] -= io_trim; | |
810 | else | |
811 | tg->io_disp[rw] = 0; | |
812 | ||
813 | tg->slice_start[rw] += nr_slices * throtl_slice; | |
814 | ||
815 | throtl_log(&tg->service_queue, | |
816 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
817 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
818 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
819 | } | |
820 | ||
821 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, | |
822 | unsigned long *wait) | |
823 | { | |
824 | bool rw = bio_data_dir(bio); | |
825 | unsigned int io_allowed; | |
826 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; | |
827 | u64 tmp; | |
828 | ||
829 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
830 | ||
831 | /* Slice has just started. Consider one slice interval */ | |
832 | if (!jiffy_elapsed) | |
833 | jiffy_elapsed_rnd = throtl_slice; | |
834 | ||
835 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
836 | ||
837 | /* | |
838 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
839 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
840 | * will allow dispatch after 1 second and after that slice should | |
841 | * have been trimmed. | |
842 | */ | |
843 | ||
844 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
845 | do_div(tmp, HZ); | |
846 | ||
847 | if (tmp > UINT_MAX) | |
848 | io_allowed = UINT_MAX; | |
849 | else | |
850 | io_allowed = tmp; | |
851 | ||
852 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
853 | if (wait) | |
854 | *wait = 0; | |
855 | return true; | |
856 | } | |
857 | ||
858 | /* Calc approx time to dispatch */ | |
859 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
860 | ||
861 | if (jiffy_wait > jiffy_elapsed) | |
862 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
863 | else | |
864 | jiffy_wait = 1; | |
865 | ||
866 | if (wait) | |
867 | *wait = jiffy_wait; | |
868 | return 0; | |
869 | } | |
870 | ||
871 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, | |
872 | unsigned long *wait) | |
873 | { | |
874 | bool rw = bio_data_dir(bio); | |
875 | u64 bytes_allowed, extra_bytes, tmp; | |
876 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; | |
877 | ||
878 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
879 | ||
880 | /* Slice has just started. Consider one slice interval */ | |
881 | if (!jiffy_elapsed) | |
882 | jiffy_elapsed_rnd = throtl_slice; | |
883 | ||
884 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
885 | ||
886 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; | |
887 | do_div(tmp, HZ); | |
888 | bytes_allowed = tmp; | |
889 | ||
890 | if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) { | |
891 | if (wait) | |
892 | *wait = 0; | |
893 | return true; | |
894 | } | |
895 | ||
896 | /* Calc approx time to dispatch */ | |
897 | extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed; | |
898 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
899 | ||
900 | if (!jiffy_wait) | |
901 | jiffy_wait = 1; | |
902 | ||
903 | /* | |
904 | * This wait time is without taking into consideration the rounding | |
905 | * up we did. Add that time also. | |
906 | */ | |
907 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
908 | if (wait) | |
909 | *wait = jiffy_wait; | |
910 | return 0; | |
911 | } | |
912 | ||
913 | /* | |
914 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
915 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
916 | */ | |
917 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, | |
918 | unsigned long *wait) | |
919 | { | |
920 | bool rw = bio_data_dir(bio); | |
921 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
922 | ||
923 | /* | |
924 | * Currently whole state machine of group depends on first bio | |
925 | * queued in the group bio list. So one should not be calling | |
926 | * this function with a different bio if there are other bios | |
927 | * queued. | |
928 | */ | |
929 | BUG_ON(tg->service_queue.nr_queued[rw] && | |
930 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); | |
931 | ||
932 | /* If tg->bps = -1, then BW is unlimited */ | |
933 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
934 | if (wait) | |
935 | *wait = 0; | |
936 | return true; | |
937 | } | |
938 | ||
939 | /* | |
940 | * If previous slice expired, start a new one otherwise renew/extend | |
941 | * existing slice to make sure it is at least throtl_slice interval | |
942 | * long since now. | |
943 | */ | |
944 | if (throtl_slice_used(tg, rw)) | |
945 | throtl_start_new_slice(tg, rw); | |
946 | else { | |
947 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
948 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); | |
949 | } | |
950 | ||
951 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && | |
952 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
953 | if (wait) | |
954 | *wait = 0; | |
955 | return 1; | |
956 | } | |
957 | ||
958 | max_wait = max(bps_wait, iops_wait); | |
959 | ||
960 | if (wait) | |
961 | *wait = max_wait; | |
962 | ||
963 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
964 | throtl_extend_slice(tg, rw, jiffies + max_wait); | |
965 | ||
966 | return 0; | |
967 | } | |
968 | ||
969 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, | |
970 | int rw) | |
971 | { | |
972 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
973 | struct tg_stats_cpu *stats_cpu; | |
974 | unsigned long flags; | |
975 | ||
976 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
977 | if (tg->stats_cpu == NULL) | |
978 | return; | |
979 | ||
980 | /* | |
981 | * Disabling interrupts to provide mutual exclusion between two | |
982 | * writes on same cpu. It probably is not needed for 64bit. Not | |
983 | * optimizing that case yet. | |
984 | */ | |
985 | local_irq_save(flags); | |
986 | ||
987 | stats_cpu = this_cpu_ptr(tg->stats_cpu); | |
988 | ||
989 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); | |
990 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
991 | ||
992 | local_irq_restore(flags); | |
993 | } | |
994 | ||
995 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
996 | { | |
997 | bool rw = bio_data_dir(bio); | |
998 | ||
999 | /* Charge the bio to the group */ | |
1000 | tg->bytes_disp[rw] += bio->bi_iter.bi_size; | |
1001 | tg->io_disp[rw]++; | |
1002 | ||
1003 | /* | |
1004 | * REQ_THROTTLED is used to prevent the same bio to be throttled | |
1005 | * more than once as a throttled bio will go through blk-throtl the | |
1006 | * second time when it eventually gets issued. Set it when a bio | |
1007 | * is being charged to a tg. | |
1008 | * | |
1009 | * Dispatch stats aren't recursive and each @bio should only be | |
1010 | * accounted by the @tg it was originally associated with. Let's | |
1011 | * update the stats when setting REQ_THROTTLED for the first time | |
1012 | * which is guaranteed to be for the @bio's original tg. | |
1013 | */ | |
1014 | if (!(bio->bi_rw & REQ_THROTTLED)) { | |
1015 | bio->bi_rw |= REQ_THROTTLED; | |
1016 | throtl_update_dispatch_stats(tg_to_blkg(tg), | |
1017 | bio->bi_iter.bi_size, bio->bi_rw); | |
1018 | } | |
1019 | } | |
1020 | ||
1021 | /** | |
1022 | * throtl_add_bio_tg - add a bio to the specified throtl_grp | |
1023 | * @bio: bio to add | |
1024 | * @qn: qnode to use | |
1025 | * @tg: the target throtl_grp | |
1026 | * | |
1027 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, | |
1028 | * tg->qnode_on_self[] is used. | |
1029 | */ | |
1030 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, | |
1031 | struct throtl_grp *tg) | |
1032 | { | |
1033 | struct throtl_service_queue *sq = &tg->service_queue; | |
1034 | bool rw = bio_data_dir(bio); | |
1035 | ||
1036 | if (!qn) | |
1037 | qn = &tg->qnode_on_self[rw]; | |
1038 | ||
1039 | /* | |
1040 | * If @tg doesn't currently have any bios queued in the same | |
1041 | * direction, queueing @bio can change when @tg should be | |
1042 | * dispatched. Mark that @tg was empty. This is automatically | |
1043 | * cleaered on the next tg_update_disptime(). | |
1044 | */ | |
1045 | if (!sq->nr_queued[rw]) | |
1046 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
1047 | ||
1048 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); | |
1049 | ||
1050 | sq->nr_queued[rw]++; | |
1051 | throtl_enqueue_tg(tg); | |
1052 | } | |
1053 | ||
1054 | static void tg_update_disptime(struct throtl_grp *tg) | |
1055 | { | |
1056 | struct throtl_service_queue *sq = &tg->service_queue; | |
1057 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
1058 | struct bio *bio; | |
1059 | ||
1060 | if ((bio = throtl_peek_queued(&sq->queued[READ]))) | |
1061 | tg_may_dispatch(tg, bio, &read_wait); | |
1062 | ||
1063 | if ((bio = throtl_peek_queued(&sq->queued[WRITE]))) | |
1064 | tg_may_dispatch(tg, bio, &write_wait); | |
1065 | ||
1066 | min_wait = min(read_wait, write_wait); | |
1067 | disptime = jiffies + min_wait; | |
1068 | ||
1069 | /* Update dispatch time */ | |
1070 | throtl_dequeue_tg(tg); | |
1071 | tg->disptime = disptime; | |
1072 | throtl_enqueue_tg(tg); | |
1073 | ||
1074 | /* see throtl_add_bio_tg() */ | |
1075 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
1076 | } | |
1077 | ||
1078 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, | |
1079 | struct throtl_grp *parent_tg, bool rw) | |
1080 | { | |
1081 | if (throtl_slice_used(parent_tg, rw)) { | |
1082 | throtl_start_new_slice_with_credit(parent_tg, rw, | |
1083 | child_tg->slice_start[rw]); | |
1084 | } | |
1085 | ||
1086 | } | |
1087 | ||
1088 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) | |
1089 | { | |
1090 | struct throtl_service_queue *sq = &tg->service_queue; | |
1091 | struct throtl_service_queue *parent_sq = sq->parent_sq; | |
1092 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
1093 | struct throtl_grp *tg_to_put = NULL; | |
1094 | struct bio *bio; | |
1095 | ||
1096 | /* | |
1097 | * @bio is being transferred from @tg to @parent_sq. Popping a bio | |
1098 | * from @tg may put its reference and @parent_sq might end up | |
1099 | * getting released prematurely. Remember the tg to put and put it | |
1100 | * after @bio is transferred to @parent_sq. | |
1101 | */ | |
1102 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); | |
1103 | sq->nr_queued[rw]--; | |
1104 | ||
1105 | throtl_charge_bio(tg, bio); | |
1106 | ||
1107 | /* | |
1108 | * If our parent is another tg, we just need to transfer @bio to | |
1109 | * the parent using throtl_add_bio_tg(). If our parent is | |
1110 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
1111 | * bio_lists[] and decrease total number queued. The caller is | |
1112 | * responsible for issuing these bios. | |
1113 | */ | |
1114 | if (parent_tg) { | |
1115 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); | |
1116 | start_parent_slice_with_credit(tg, parent_tg, rw); | |
1117 | } else { | |
1118 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], | |
1119 | &parent_sq->queued[rw]); | |
1120 | BUG_ON(tg->td->nr_queued[rw] <= 0); | |
1121 | tg->td->nr_queued[rw]--; | |
1122 | } | |
1123 | ||
1124 | throtl_trim_slice(tg, rw); | |
1125 | ||
1126 | if (tg_to_put) | |
1127 | blkg_put(tg_to_blkg(tg_to_put)); | |
1128 | } | |
1129 | ||
1130 | static int throtl_dispatch_tg(struct throtl_grp *tg) | |
1131 | { | |
1132 | struct throtl_service_queue *sq = &tg->service_queue; | |
1133 | unsigned int nr_reads = 0, nr_writes = 0; | |
1134 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
1135 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; | |
1136 | struct bio *bio; | |
1137 | ||
1138 | /* Try to dispatch 75% READS and 25% WRITES */ | |
1139 | ||
1140 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && | |
1141 | tg_may_dispatch(tg, bio, NULL)) { | |
1142 | ||
1143 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); | |
1144 | nr_reads++; | |
1145 | ||
1146 | if (nr_reads >= max_nr_reads) | |
1147 | break; | |
1148 | } | |
1149 | ||
1150 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && | |
1151 | tg_may_dispatch(tg, bio, NULL)) { | |
1152 | ||
1153 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); | |
1154 | nr_writes++; | |
1155 | ||
1156 | if (nr_writes >= max_nr_writes) | |
1157 | break; | |
1158 | } | |
1159 | ||
1160 | return nr_reads + nr_writes; | |
1161 | } | |
1162 | ||
1163 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) | |
1164 | { | |
1165 | unsigned int nr_disp = 0; | |
1166 | ||
1167 | while (1) { | |
1168 | struct throtl_grp *tg = throtl_rb_first(parent_sq); | |
1169 | struct throtl_service_queue *sq = &tg->service_queue; | |
1170 | ||
1171 | if (!tg) | |
1172 | break; | |
1173 | ||
1174 | if (time_before(jiffies, tg->disptime)) | |
1175 | break; | |
1176 | ||
1177 | throtl_dequeue_tg(tg); | |
1178 | ||
1179 | nr_disp += throtl_dispatch_tg(tg); | |
1180 | ||
1181 | if (sq->nr_queued[0] || sq->nr_queued[1]) | |
1182 | tg_update_disptime(tg); | |
1183 | ||
1184 | if (nr_disp >= throtl_quantum) | |
1185 | break; | |
1186 | } | |
1187 | ||
1188 | return nr_disp; | |
1189 | } | |
1190 | ||
1191 | /** | |
1192 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
1193 | * @arg: the throtl_service_queue being serviced | |
1194 | * | |
1195 | * This timer is armed when a child throtl_grp with active bio's become | |
1196 | * pending and queued on the service_queue's pending_tree and expires when | |
1197 | * the first child throtl_grp should be dispatched. This function | |
1198 | * dispatches bio's from the children throtl_grps to the parent | |
1199 | * service_queue. | |
1200 | * | |
1201 | * If the parent's parent is another throtl_grp, dispatching is propagated | |
1202 | * by either arming its pending_timer or repeating dispatch directly. If | |
1203 | * the top-level service_tree is reached, throtl_data->dispatch_work is | |
1204 | * kicked so that the ready bio's are issued. | |
1205 | */ | |
1206 | static void throtl_pending_timer_fn(unsigned long arg) | |
1207 | { | |
1208 | struct throtl_service_queue *sq = (void *)arg; | |
1209 | struct throtl_grp *tg = sq_to_tg(sq); | |
1210 | struct throtl_data *td = sq_to_td(sq); | |
1211 | struct request_queue *q = td->queue; | |
1212 | struct throtl_service_queue *parent_sq; | |
1213 | bool dispatched; | |
1214 | int ret; | |
1215 | ||
1216 | spin_lock_irq(q->queue_lock); | |
1217 | again: | |
1218 | parent_sq = sq->parent_sq; | |
1219 | dispatched = false; | |
1220 | ||
1221 | while (true) { | |
1222 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
1223 | sq->nr_queued[READ] + sq->nr_queued[WRITE], | |
1224 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
1225 | ||
1226 | ret = throtl_select_dispatch(sq); | |
1227 | if (ret) { | |
1228 | throtl_log(sq, "bios disp=%u", ret); | |
1229 | dispatched = true; | |
1230 | } | |
1231 | ||
1232 | if (throtl_schedule_next_dispatch(sq, false)) | |
1233 | break; | |
1234 | ||
1235 | /* this dispatch windows is still open, relax and repeat */ | |
1236 | spin_unlock_irq(q->queue_lock); | |
1237 | cpu_relax(); | |
1238 | spin_lock_irq(q->queue_lock); | |
1239 | } | |
1240 | ||
1241 | if (!dispatched) | |
1242 | goto out_unlock; | |
1243 | ||
1244 | if (parent_sq) { | |
1245 | /* @parent_sq is another throl_grp, propagate dispatch */ | |
1246 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1247 | tg_update_disptime(tg); | |
1248 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { | |
1249 | /* window is already open, repeat dispatching */ | |
1250 | sq = parent_sq; | |
1251 | tg = sq_to_tg(sq); | |
1252 | goto again; | |
1253 | } | |
1254 | } | |
1255 | } else { | |
1256 | /* reached the top-level, queue issueing */ | |
1257 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1258 | } | |
1259 | out_unlock: | |
1260 | spin_unlock_irq(q->queue_lock); | |
1261 | } | |
1262 | ||
1263 | /** | |
1264 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
1265 | * @work: work item being executed | |
1266 | * | |
1267 | * This function is queued for execution when bio's reach the bio_lists[] | |
1268 | * of throtl_data->service_queue. Those bio's are ready and issued by this | |
1269 | * function. | |
1270 | */ | |
1271 | static void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
1272 | { | |
1273 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1274 | dispatch_work); | |
1275 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1276 | struct request_queue *q = td->queue; | |
1277 | struct bio_list bio_list_on_stack; | |
1278 | struct bio *bio; | |
1279 | struct blk_plug plug; | |
1280 | int rw; | |
1281 | ||
1282 | bio_list_init(&bio_list_on_stack); | |
1283 | ||
1284 | spin_lock_irq(q->queue_lock); | |
1285 | for (rw = READ; rw <= WRITE; rw++) | |
1286 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) | |
1287 | bio_list_add(&bio_list_on_stack, bio); | |
1288 | spin_unlock_irq(q->queue_lock); | |
1289 | ||
1290 | if (!bio_list_empty(&bio_list_on_stack)) { | |
1291 | blk_start_plug(&plug); | |
1292 | while((bio = bio_list_pop(&bio_list_on_stack))) | |
1293 | generic_make_request(bio); | |
1294 | blk_finish_plug(&plug); | |
1295 | } | |
1296 | } | |
1297 | ||
1298 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, | |
1299 | struct blkg_policy_data *pd, int off) | |
1300 | { | |
1301 | struct throtl_grp *tg = pd_to_tg(pd); | |
1302 | struct blkg_rwstat rwstat = { }, tmp; | |
1303 | int i, cpu; | |
1304 | ||
1305 | if (tg->stats_cpu == NULL) | |
1306 | return 0; | |
1307 | ||
1308 | for_each_possible_cpu(cpu) { | |
1309 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
1310 | ||
1311 | tmp = blkg_rwstat_read((void *)sc + off); | |
1312 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
1313 | rwstat.cnt[i] += tmp.cnt[i]; | |
1314 | } | |
1315 | ||
1316 | return __blkg_prfill_rwstat(sf, pd, &rwstat); | |
1317 | } | |
1318 | ||
1319 | static int tg_print_cpu_rwstat(struct seq_file *sf, void *v) | |
1320 | { | |
1321 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_cpu_rwstat, | |
1322 | &blkcg_policy_throtl, seq_cft(sf)->private, true); | |
1323 | return 0; | |
1324 | } | |
1325 | ||
1326 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, | |
1327 | int off) | |
1328 | { | |
1329 | struct throtl_grp *tg = pd_to_tg(pd); | |
1330 | u64 v = *(u64 *)((void *)tg + off); | |
1331 | ||
1332 | if (v == -1) | |
1333 | return 0; | |
1334 | return __blkg_prfill_u64(sf, pd, v); | |
1335 | } | |
1336 | ||
1337 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, | |
1338 | int off) | |
1339 | { | |
1340 | struct throtl_grp *tg = pd_to_tg(pd); | |
1341 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
1342 | ||
1343 | if (v == -1) | |
1344 | return 0; | |
1345 | return __blkg_prfill_u64(sf, pd, v); | |
1346 | } | |
1347 | ||
1348 | static int tg_print_conf_u64(struct seq_file *sf, void *v) | |
1349 | { | |
1350 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64, | |
1351 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
1352 | return 0; | |
1353 | } | |
1354 | ||
1355 | static int tg_print_conf_uint(struct seq_file *sf, void *v) | |
1356 | { | |
1357 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint, | |
1358 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
1359 | return 0; | |
1360 | } | |
1361 | ||
1362 | static ssize_t tg_set_conf(struct kernfs_open_file *of, | |
1363 | char *buf, size_t nbytes, loff_t off, bool is_u64) | |
1364 | { | |
1365 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1366 | struct blkg_conf_ctx ctx; | |
1367 | struct throtl_grp *tg; | |
1368 | struct throtl_service_queue *sq; | |
1369 | struct blkcg_gq *blkg; | |
1370 | struct cgroup_subsys_state *pos_css; | |
1371 | int ret; | |
1372 | ||
1373 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1374 | if (ret) | |
1375 | return ret; | |
1376 | ||
1377 | tg = blkg_to_tg(ctx.blkg); | |
1378 | sq = &tg->service_queue; | |
1379 | ||
1380 | if (!ctx.v) | |
1381 | ctx.v = -1; | |
1382 | ||
1383 | if (is_u64) | |
1384 | *(u64 *)((void *)tg + of_cft(of)->private) = ctx.v; | |
1385 | else | |
1386 | *(unsigned int *)((void *)tg + of_cft(of)->private) = ctx.v; | |
1387 | ||
1388 | throtl_log(&tg->service_queue, | |
1389 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
1390 | tg->bps[READ], tg->bps[WRITE], | |
1391 | tg->iops[READ], tg->iops[WRITE]); | |
1392 | ||
1393 | /* | |
1394 | * Update has_rules[] flags for the updated tg's subtree. A tg is | |
1395 | * considered to have rules if either the tg itself or any of its | |
1396 | * ancestors has rules. This identifies groups without any | |
1397 | * restrictions in the whole hierarchy and allows them to bypass | |
1398 | * blk-throttle. | |
1399 | */ | |
1400 | blkg_for_each_descendant_pre(blkg, pos_css, ctx.blkg) | |
1401 | tg_update_has_rules(blkg_to_tg(blkg)); | |
1402 | ||
1403 | /* | |
1404 | * We're already holding queue_lock and know @tg is valid. Let's | |
1405 | * apply the new config directly. | |
1406 | * | |
1407 | * Restart the slices for both READ and WRITES. It might happen | |
1408 | * that a group's limit are dropped suddenly and we don't want to | |
1409 | * account recently dispatched IO with new low rate. | |
1410 | */ | |
1411 | throtl_start_new_slice(tg, 0); | |
1412 | throtl_start_new_slice(tg, 1); | |
1413 | ||
1414 | if (tg->flags & THROTL_TG_PENDING) { | |
1415 | tg_update_disptime(tg); | |
1416 | throtl_schedule_next_dispatch(sq->parent_sq, true); | |
1417 | } | |
1418 | ||
1419 | blkg_conf_finish(&ctx); | |
1420 | return nbytes; | |
1421 | } | |
1422 | ||
1423 | static ssize_t tg_set_conf_u64(struct kernfs_open_file *of, | |
1424 | char *buf, size_t nbytes, loff_t off) | |
1425 | { | |
1426 | return tg_set_conf(of, buf, nbytes, off, true); | |
1427 | } | |
1428 | ||
1429 | static ssize_t tg_set_conf_uint(struct kernfs_open_file *of, | |
1430 | char *buf, size_t nbytes, loff_t off) | |
1431 | { | |
1432 | return tg_set_conf(of, buf, nbytes, off, false); | |
1433 | } | |
1434 | ||
1435 | static struct cftype throtl_files[] = { | |
1436 | { | |
1437 | .name = "throttle.read_bps_device", | |
1438 | .private = offsetof(struct throtl_grp, bps[READ]), | |
1439 | .seq_show = tg_print_conf_u64, | |
1440 | .write = tg_set_conf_u64, | |
1441 | }, | |
1442 | { | |
1443 | .name = "throttle.write_bps_device", | |
1444 | .private = offsetof(struct throtl_grp, bps[WRITE]), | |
1445 | .seq_show = tg_print_conf_u64, | |
1446 | .write = tg_set_conf_u64, | |
1447 | }, | |
1448 | { | |
1449 | .name = "throttle.read_iops_device", | |
1450 | .private = offsetof(struct throtl_grp, iops[READ]), | |
1451 | .seq_show = tg_print_conf_uint, | |
1452 | .write = tg_set_conf_uint, | |
1453 | }, | |
1454 | { | |
1455 | .name = "throttle.write_iops_device", | |
1456 | .private = offsetof(struct throtl_grp, iops[WRITE]), | |
1457 | .seq_show = tg_print_conf_uint, | |
1458 | .write = tg_set_conf_uint, | |
1459 | }, | |
1460 | { | |
1461 | .name = "throttle.io_service_bytes", | |
1462 | .private = offsetof(struct tg_stats_cpu, service_bytes), | |
1463 | .seq_show = tg_print_cpu_rwstat, | |
1464 | }, | |
1465 | { | |
1466 | .name = "throttle.io_serviced", | |
1467 | .private = offsetof(struct tg_stats_cpu, serviced), | |
1468 | .seq_show = tg_print_cpu_rwstat, | |
1469 | }, | |
1470 | { } /* terminate */ | |
1471 | }; | |
1472 | ||
1473 | static void throtl_shutdown_wq(struct request_queue *q) | |
1474 | { | |
1475 | struct throtl_data *td = q->td; | |
1476 | ||
1477 | cancel_work_sync(&td->dispatch_work); | |
1478 | } | |
1479 | ||
1480 | static struct blkcg_policy blkcg_policy_throtl = { | |
1481 | .cftypes = throtl_files, | |
1482 | ||
1483 | .pd_alloc_fn = throtl_pd_alloc, | |
1484 | .pd_init_fn = throtl_pd_init, | |
1485 | .pd_online_fn = throtl_pd_online, | |
1486 | .pd_exit_fn = throtl_pd_exit, | |
1487 | .pd_free_fn = throtl_pd_free, | |
1488 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
1489 | }; | |
1490 | ||
1491 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) | |
1492 | { | |
1493 | struct throtl_data *td = q->td; | |
1494 | struct throtl_qnode *qn = NULL; | |
1495 | struct throtl_grp *tg; | |
1496 | struct throtl_service_queue *sq; | |
1497 | bool rw = bio_data_dir(bio); | |
1498 | struct blkcg *blkcg; | |
1499 | bool throttled = false; | |
1500 | ||
1501 | /* see throtl_charge_bio() */ | |
1502 | if (bio->bi_rw & REQ_THROTTLED) | |
1503 | goto out; | |
1504 | ||
1505 | /* | |
1506 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1507 | * basic fields like stats and io rates. If a group has no rules, | |
1508 | * just update the dispatch stats in lockless manner and return. | |
1509 | */ | |
1510 | rcu_read_lock(); | |
1511 | blkcg = bio_blkcg(bio); | |
1512 | tg = throtl_lookup_tg(td, blkcg); | |
1513 | if (tg) { | |
1514 | if (!tg->has_rules[rw]) { | |
1515 | throtl_update_dispatch_stats(tg_to_blkg(tg), | |
1516 | bio->bi_iter.bi_size, bio->bi_rw); | |
1517 | goto out_unlock_rcu; | |
1518 | } | |
1519 | } | |
1520 | ||
1521 | /* | |
1522 | * Either group has not been allocated yet or it is not an unlimited | |
1523 | * IO group | |
1524 | */ | |
1525 | spin_lock_irq(q->queue_lock); | |
1526 | tg = throtl_lookup_create_tg(td, blkcg); | |
1527 | if (unlikely(!tg)) | |
1528 | goto out_unlock; | |
1529 | ||
1530 | sq = &tg->service_queue; | |
1531 | ||
1532 | while (true) { | |
1533 | /* throtl is FIFO - if bios are already queued, should queue */ | |
1534 | if (sq->nr_queued[rw]) | |
1535 | break; | |
1536 | ||
1537 | /* if above limits, break to queue */ | |
1538 | if (!tg_may_dispatch(tg, bio, NULL)) | |
1539 | break; | |
1540 | ||
1541 | /* within limits, let's charge and dispatch directly */ | |
1542 | throtl_charge_bio(tg, bio); | |
1543 | ||
1544 | /* | |
1545 | * We need to trim slice even when bios are not being queued | |
1546 | * otherwise it might happen that a bio is not queued for | |
1547 | * a long time and slice keeps on extending and trim is not | |
1548 | * called for a long time. Now if limits are reduced suddenly | |
1549 | * we take into account all the IO dispatched so far at new | |
1550 | * low rate and * newly queued IO gets a really long dispatch | |
1551 | * time. | |
1552 | * | |
1553 | * So keep on trimming slice even if bio is not queued. | |
1554 | */ | |
1555 | throtl_trim_slice(tg, rw); | |
1556 | ||
1557 | /* | |
1558 | * @bio passed through this layer without being throttled. | |
1559 | * Climb up the ladder. If we''re already at the top, it | |
1560 | * can be executed directly. | |
1561 | */ | |
1562 | qn = &tg->qnode_on_parent[rw]; | |
1563 | sq = sq->parent_sq; | |
1564 | tg = sq_to_tg(sq); | |
1565 | if (!tg) | |
1566 | goto out_unlock; | |
1567 | } | |
1568 | ||
1569 | /* out-of-limit, queue to @tg */ | |
1570 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", | |
1571 | rw == READ ? 'R' : 'W', | |
1572 | tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw], | |
1573 | tg->io_disp[rw], tg->iops[rw], | |
1574 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
1575 | ||
1576 | bio_associate_current(bio); | |
1577 | tg->td->nr_queued[rw]++; | |
1578 | throtl_add_bio_tg(bio, qn, tg); | |
1579 | throttled = true; | |
1580 | ||
1581 | /* | |
1582 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
1583 | * was empty before @bio. The forced scheduling isn't likely to | |
1584 | * cause undue delay as @bio is likely to be dispatched directly if | |
1585 | * its @tg's disptime is not in the future. | |
1586 | */ | |
1587 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1588 | tg_update_disptime(tg); | |
1589 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); | |
1590 | } | |
1591 | ||
1592 | out_unlock: | |
1593 | spin_unlock_irq(q->queue_lock); | |
1594 | out_unlock_rcu: | |
1595 | rcu_read_unlock(); | |
1596 | out: | |
1597 | /* | |
1598 | * As multiple blk-throtls may stack in the same issue path, we | |
1599 | * don't want bios to leave with the flag set. Clear the flag if | |
1600 | * being issued. | |
1601 | */ | |
1602 | if (!throttled) | |
1603 | bio->bi_rw &= ~REQ_THROTTLED; | |
1604 | return throttled; | |
1605 | } | |
1606 | ||
1607 | /* | |
1608 | * Dispatch all bios from all children tg's queued on @parent_sq. On | |
1609 | * return, @parent_sq is guaranteed to not have any active children tg's | |
1610 | * and all bios from previously active tg's are on @parent_sq->bio_lists[]. | |
1611 | */ | |
1612 | static void tg_drain_bios(struct throtl_service_queue *parent_sq) | |
1613 | { | |
1614 | struct throtl_grp *tg; | |
1615 | ||
1616 | while ((tg = throtl_rb_first(parent_sq))) { | |
1617 | struct throtl_service_queue *sq = &tg->service_queue; | |
1618 | struct bio *bio; | |
1619 | ||
1620 | throtl_dequeue_tg(tg); | |
1621 | ||
1622 | while ((bio = throtl_peek_queued(&sq->queued[READ]))) | |
1623 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); | |
1624 | while ((bio = throtl_peek_queued(&sq->queued[WRITE]))) | |
1625 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); | |
1626 | } | |
1627 | } | |
1628 | ||
1629 | /** | |
1630 | * blk_throtl_drain - drain throttled bios | |
1631 | * @q: request_queue to drain throttled bios for | |
1632 | * | |
1633 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1634 | */ | |
1635 | void blk_throtl_drain(struct request_queue *q) | |
1636 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1637 | { | |
1638 | struct throtl_data *td = q->td; | |
1639 | struct blkcg_gq *blkg; | |
1640 | struct cgroup_subsys_state *pos_css; | |
1641 | struct bio *bio; | |
1642 | int rw; | |
1643 | ||
1644 | queue_lockdep_assert_held(q); | |
1645 | rcu_read_lock(); | |
1646 | ||
1647 | /* | |
1648 | * Drain each tg while doing post-order walk on the blkg tree, so | |
1649 | * that all bios are propagated to td->service_queue. It'd be | |
1650 | * better to walk service_queue tree directly but blkg walk is | |
1651 | * easier. | |
1652 | */ | |
1653 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) | |
1654 | tg_drain_bios(&blkg_to_tg(blkg)->service_queue); | |
1655 | ||
1656 | /* finally, transfer bios from top-level tg's into the td */ | |
1657 | tg_drain_bios(&td->service_queue); | |
1658 | ||
1659 | rcu_read_unlock(); | |
1660 | spin_unlock_irq(q->queue_lock); | |
1661 | ||
1662 | /* all bios now should be in td->service_queue, issue them */ | |
1663 | for (rw = READ; rw <= WRITE; rw++) | |
1664 | while ((bio = throtl_pop_queued(&td->service_queue.queued[rw], | |
1665 | NULL))) | |
1666 | generic_make_request(bio); | |
1667 | ||
1668 | spin_lock_irq(q->queue_lock); | |
1669 | } | |
1670 | ||
1671 | int blk_throtl_init(struct request_queue *q) | |
1672 | { | |
1673 | struct throtl_data *td; | |
1674 | int ret; | |
1675 | ||
1676 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1677 | if (!td) | |
1678 | return -ENOMEM; | |
1679 | ||
1680 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); | |
1681 | throtl_service_queue_init(&td->service_queue, NULL); | |
1682 | ||
1683 | q->td = td; | |
1684 | td->queue = q; | |
1685 | ||
1686 | /* activate policy */ | |
1687 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); | |
1688 | if (ret) | |
1689 | kfree(td); | |
1690 | return ret; | |
1691 | } | |
1692 | ||
1693 | void blk_throtl_exit(struct request_queue *q) | |
1694 | { | |
1695 | BUG_ON(!q->td); | |
1696 | throtl_shutdown_wq(q); | |
1697 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); | |
1698 | kfree(q->td); | |
1699 | } | |
1700 | ||
1701 | static int __init throtl_init(void) | |
1702 | { | |
1703 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); | |
1704 | if (!kthrotld_workqueue) | |
1705 | panic("Failed to create kthrotld\n"); | |
1706 | ||
1707 | return blkcg_policy_register(&blkcg_policy_throtl); | |
1708 | } | |
1709 | ||
1710 | module_init(throtl_init); |