]>
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 a slice and after that slice is renewed */ | |
22 | #define DFL_THROTL_SLICE_HD (HZ / 10) | |
23 | #define DFL_THROTL_SLICE_SSD (HZ / 50) | |
24 | #define MAX_THROTL_SLICE (HZ) | |
25 | ||
26 | static struct blkcg_policy blkcg_policy_throtl; | |
27 | ||
28 | /* A workqueue to queue throttle related work */ | |
29 | static struct workqueue_struct *kthrotld_workqueue; | |
30 | ||
31 | /* | |
32 | * To implement hierarchical throttling, throtl_grps form a tree and bios | |
33 | * are dispatched upwards level by level until they reach the top and get | |
34 | * issued. When dispatching bios from the children and local group at each | |
35 | * level, if the bios are dispatched into a single bio_list, there's a risk | |
36 | * of a local or child group which can queue many bios at once filling up | |
37 | * the list starving others. | |
38 | * | |
39 | * To avoid such starvation, dispatched bios are queued separately | |
40 | * according to where they came from. When they are again dispatched to | |
41 | * the parent, they're popped in round-robin order so that no single source | |
42 | * hogs the dispatch window. | |
43 | * | |
44 | * throtl_qnode is used to keep the queued bios separated by their sources. | |
45 | * Bios are queued to throtl_qnode which in turn is queued to | |
46 | * throtl_service_queue and then dispatched in round-robin order. | |
47 | * | |
48 | * It's also used to track the reference counts on blkg's. A qnode always | |
49 | * belongs to a throtl_grp and gets queued on itself or the parent, so | |
50 | * incrementing the reference of the associated throtl_grp when a qnode is | |
51 | * queued and decrementing when dequeued is enough to keep the whole blkg | |
52 | * tree pinned while bios are in flight. | |
53 | */ | |
54 | struct throtl_qnode { | |
55 | struct list_head node; /* service_queue->queued[] */ | |
56 | struct bio_list bios; /* queued bios */ | |
57 | struct throtl_grp *tg; /* tg this qnode belongs to */ | |
58 | }; | |
59 | ||
60 | struct throtl_service_queue { | |
61 | struct throtl_service_queue *parent_sq; /* the parent service_queue */ | |
62 | ||
63 | /* | |
64 | * Bios queued directly to this service_queue or dispatched from | |
65 | * children throtl_grp's. | |
66 | */ | |
67 | struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */ | |
68 | unsigned int nr_queued[2]; /* number of queued bios */ | |
69 | ||
70 | /* | |
71 | * RB tree of active children throtl_grp's, which are sorted by | |
72 | * their ->disptime. | |
73 | */ | |
74 | struct rb_root pending_tree; /* RB tree of active tgs */ | |
75 | struct rb_node *first_pending; /* first node in the tree */ | |
76 | unsigned int nr_pending; /* # queued in the tree */ | |
77 | unsigned long first_pending_disptime; /* disptime of the first tg */ | |
78 | struct timer_list pending_timer; /* fires on first_pending_disptime */ | |
79 | }; | |
80 | ||
81 | enum tg_state_flags { | |
82 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
83 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ | |
84 | }; | |
85 | ||
86 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) | |
87 | ||
88 | enum { | |
89 | LIMIT_LOW, | |
90 | LIMIT_MAX, | |
91 | LIMIT_CNT, | |
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 | /* internally used bytes per second rate limits */ | |
131 | uint64_t bps[2][LIMIT_CNT]; | |
132 | /* user configured bps limits */ | |
133 | uint64_t bps_conf[2][LIMIT_CNT]; | |
134 | ||
135 | /* internally used IOPS limits */ | |
136 | unsigned int iops[2][LIMIT_CNT]; | |
137 | /* user configured IOPS limits */ | |
138 | unsigned int iops_conf[2][LIMIT_CNT]; | |
139 | ||
140 | /* Number of bytes disptached in current slice */ | |
141 | uint64_t bytes_disp[2]; | |
142 | /* Number of bio's dispatched in current slice */ | |
143 | unsigned int io_disp[2]; | |
144 | ||
145 | unsigned long last_low_overflow_time[2]; | |
146 | ||
147 | uint64_t last_bytes_disp[2]; | |
148 | unsigned int last_io_disp[2]; | |
149 | ||
150 | unsigned long last_check_time; | |
151 | ||
152 | unsigned long last_dispatch_time[2]; | |
153 | ||
154 | /* When did we start a new slice */ | |
155 | unsigned long slice_start[2]; | |
156 | unsigned long slice_end[2]; | |
157 | }; | |
158 | ||
159 | struct throtl_data | |
160 | { | |
161 | /* service tree for active throtl groups */ | |
162 | struct throtl_service_queue service_queue; | |
163 | ||
164 | struct request_queue *queue; | |
165 | ||
166 | /* Total Number of queued bios on READ and WRITE lists */ | |
167 | unsigned int nr_queued[2]; | |
168 | ||
169 | unsigned int throtl_slice; | |
170 | ||
171 | /* Work for dispatching throttled bios */ | |
172 | struct work_struct dispatch_work; | |
173 | unsigned int limit_index; | |
174 | bool limit_valid[LIMIT_CNT]; | |
175 | ||
176 | unsigned long low_upgrade_time; | |
177 | unsigned long low_downgrade_time; | |
178 | ||
179 | unsigned int scale; | |
180 | }; | |
181 | ||
182 | static void throtl_pending_timer_fn(unsigned long arg); | |
183 | ||
184 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) | |
185 | { | |
186 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
187 | } | |
188 | ||
189 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) | |
190 | { | |
191 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); | |
192 | } | |
193 | ||
194 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) | |
195 | { | |
196 | return pd_to_blkg(&tg->pd); | |
197 | } | |
198 | ||
199 | /** | |
200 | * sq_to_tg - return the throl_grp the specified service queue belongs to | |
201 | * @sq: the throtl_service_queue of interest | |
202 | * | |
203 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one | |
204 | * embedded in throtl_data, %NULL is returned. | |
205 | */ | |
206 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) | |
207 | { | |
208 | if (sq && sq->parent_sq) | |
209 | return container_of(sq, struct throtl_grp, service_queue); | |
210 | else | |
211 | return NULL; | |
212 | } | |
213 | ||
214 | /** | |
215 | * sq_to_td - return throtl_data the specified service queue belongs to | |
216 | * @sq: the throtl_service_queue of interest | |
217 | * | |
218 | * A service_queue can be embedded in either a throtl_grp or throtl_data. | |
219 | * Determine the associated throtl_data accordingly and return it. | |
220 | */ | |
221 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) | |
222 | { | |
223 | struct throtl_grp *tg = sq_to_tg(sq); | |
224 | ||
225 | if (tg) | |
226 | return tg->td; | |
227 | else | |
228 | return container_of(sq, struct throtl_data, service_queue); | |
229 | } | |
230 | ||
231 | /* | |
232 | * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to | |
233 | * make the IO dispatch more smooth. | |
234 | * Scale up: linearly scale up according to lapsed time since upgrade. For | |
235 | * every throtl_slice, the limit scales up 1/2 .low limit till the | |
236 | * limit hits .max limit | |
237 | * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit | |
238 | */ | |
239 | static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td) | |
240 | { | |
241 | /* arbitrary value to avoid too big scale */ | |
242 | if (td->scale < 4096 && time_after_eq(jiffies, | |
243 | td->low_upgrade_time + td->scale * td->throtl_slice)) | |
244 | td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice; | |
245 | ||
246 | return low + (low >> 1) * td->scale; | |
247 | } | |
248 | ||
249 | static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw) | |
250 | { | |
251 | struct blkcg_gq *blkg = tg_to_blkg(tg); | |
252 | struct throtl_data *td; | |
253 | uint64_t ret; | |
254 | ||
255 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
256 | return U64_MAX; | |
257 | ||
258 | td = tg->td; | |
259 | ret = tg->bps[rw][td->limit_index]; | |
260 | if (ret == 0 && td->limit_index == LIMIT_LOW) | |
261 | return tg->bps[rw][LIMIT_MAX]; | |
262 | ||
263 | if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] && | |
264 | tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) { | |
265 | uint64_t adjusted; | |
266 | ||
267 | adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td); | |
268 | ret = min(tg->bps[rw][LIMIT_MAX], adjusted); | |
269 | } | |
270 | return ret; | |
271 | } | |
272 | ||
273 | static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw) | |
274 | { | |
275 | struct blkcg_gq *blkg = tg_to_blkg(tg); | |
276 | struct throtl_data *td; | |
277 | unsigned int ret; | |
278 | ||
279 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
280 | return UINT_MAX; | |
281 | td = tg->td; | |
282 | ret = tg->iops[rw][td->limit_index]; | |
283 | if (ret == 0 && tg->td->limit_index == LIMIT_LOW) | |
284 | return tg->iops[rw][LIMIT_MAX]; | |
285 | ||
286 | if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] && | |
287 | tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) { | |
288 | uint64_t adjusted; | |
289 | ||
290 | adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td); | |
291 | if (adjusted > UINT_MAX) | |
292 | adjusted = UINT_MAX; | |
293 | ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted); | |
294 | } | |
295 | return ret; | |
296 | } | |
297 | ||
298 | /** | |
299 | * throtl_log - log debug message via blktrace | |
300 | * @sq: the service_queue being reported | |
301 | * @fmt: printf format string | |
302 | * @args: printf args | |
303 | * | |
304 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a | |
305 | * throtl_grp; otherwise, just "throtl". | |
306 | */ | |
307 | #define throtl_log(sq, fmt, args...) do { \ | |
308 | struct throtl_grp *__tg = sq_to_tg((sq)); \ | |
309 | struct throtl_data *__td = sq_to_td((sq)); \ | |
310 | \ | |
311 | (void)__td; \ | |
312 | if (likely(!blk_trace_note_message_enabled(__td->queue))) \ | |
313 | break; \ | |
314 | if ((__tg)) { \ | |
315 | char __pbuf[128]; \ | |
316 | \ | |
317 | blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \ | |
318 | blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
319 | } else { \ | |
320 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
321 | } \ | |
322 | } while (0) | |
323 | ||
324 | static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg) | |
325 | { | |
326 | INIT_LIST_HEAD(&qn->node); | |
327 | bio_list_init(&qn->bios); | |
328 | qn->tg = tg; | |
329 | } | |
330 | ||
331 | /** | |
332 | * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it | |
333 | * @bio: bio being added | |
334 | * @qn: qnode to add bio to | |
335 | * @queued: the service_queue->queued[] list @qn belongs to | |
336 | * | |
337 | * Add @bio to @qn and put @qn on @queued if it's not already on. | |
338 | * @qn->tg's reference count is bumped when @qn is activated. See the | |
339 | * comment on top of throtl_qnode definition for details. | |
340 | */ | |
341 | static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn, | |
342 | struct list_head *queued) | |
343 | { | |
344 | bio_list_add(&qn->bios, bio); | |
345 | if (list_empty(&qn->node)) { | |
346 | list_add_tail(&qn->node, queued); | |
347 | blkg_get(tg_to_blkg(qn->tg)); | |
348 | } | |
349 | } | |
350 | ||
351 | /** | |
352 | * throtl_peek_queued - peek the first bio on a qnode list | |
353 | * @queued: the qnode list to peek | |
354 | */ | |
355 | static struct bio *throtl_peek_queued(struct list_head *queued) | |
356 | { | |
357 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
358 | struct bio *bio; | |
359 | ||
360 | if (list_empty(queued)) | |
361 | return NULL; | |
362 | ||
363 | bio = bio_list_peek(&qn->bios); | |
364 | WARN_ON_ONCE(!bio); | |
365 | return bio; | |
366 | } | |
367 | ||
368 | /** | |
369 | * throtl_pop_queued - pop the first bio form a qnode list | |
370 | * @queued: the qnode list to pop a bio from | |
371 | * @tg_to_put: optional out argument for throtl_grp to put | |
372 | * | |
373 | * Pop the first bio from the qnode list @queued. After popping, the first | |
374 | * qnode is removed from @queued if empty or moved to the end of @queued so | |
375 | * that the popping order is round-robin. | |
376 | * | |
377 | * When the first qnode is removed, its associated throtl_grp should be put | |
378 | * too. If @tg_to_put is NULL, this function automatically puts it; | |
379 | * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is | |
380 | * responsible for putting it. | |
381 | */ | |
382 | static struct bio *throtl_pop_queued(struct list_head *queued, | |
383 | struct throtl_grp **tg_to_put) | |
384 | { | |
385 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
386 | struct bio *bio; | |
387 | ||
388 | if (list_empty(queued)) | |
389 | return NULL; | |
390 | ||
391 | bio = bio_list_pop(&qn->bios); | |
392 | WARN_ON_ONCE(!bio); | |
393 | ||
394 | if (bio_list_empty(&qn->bios)) { | |
395 | list_del_init(&qn->node); | |
396 | if (tg_to_put) | |
397 | *tg_to_put = qn->tg; | |
398 | else | |
399 | blkg_put(tg_to_blkg(qn->tg)); | |
400 | } else { | |
401 | list_move_tail(&qn->node, queued); | |
402 | } | |
403 | ||
404 | return bio; | |
405 | } | |
406 | ||
407 | /* init a service_queue, assumes the caller zeroed it */ | |
408 | static void throtl_service_queue_init(struct throtl_service_queue *sq) | |
409 | { | |
410 | INIT_LIST_HEAD(&sq->queued[0]); | |
411 | INIT_LIST_HEAD(&sq->queued[1]); | |
412 | sq->pending_tree = RB_ROOT; | |
413 | setup_timer(&sq->pending_timer, throtl_pending_timer_fn, | |
414 | (unsigned long)sq); | |
415 | } | |
416 | ||
417 | static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node) | |
418 | { | |
419 | struct throtl_grp *tg; | |
420 | int rw; | |
421 | ||
422 | tg = kzalloc_node(sizeof(*tg), gfp, node); | |
423 | if (!tg) | |
424 | return NULL; | |
425 | ||
426 | throtl_service_queue_init(&tg->service_queue); | |
427 | ||
428 | for (rw = READ; rw <= WRITE; rw++) { | |
429 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); | |
430 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); | |
431 | } | |
432 | ||
433 | RB_CLEAR_NODE(&tg->rb_node); | |
434 | tg->bps[READ][LIMIT_MAX] = U64_MAX; | |
435 | tg->bps[WRITE][LIMIT_MAX] = U64_MAX; | |
436 | tg->iops[READ][LIMIT_MAX] = UINT_MAX; | |
437 | tg->iops[WRITE][LIMIT_MAX] = UINT_MAX; | |
438 | tg->bps_conf[READ][LIMIT_MAX] = U64_MAX; | |
439 | tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX; | |
440 | tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX; | |
441 | tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX; | |
442 | /* LIMIT_LOW will have default value 0 */ | |
443 | ||
444 | return &tg->pd; | |
445 | } | |
446 | ||
447 | static void throtl_pd_init(struct blkg_policy_data *pd) | |
448 | { | |
449 | struct throtl_grp *tg = pd_to_tg(pd); | |
450 | struct blkcg_gq *blkg = tg_to_blkg(tg); | |
451 | struct throtl_data *td = blkg->q->td; | |
452 | struct throtl_service_queue *sq = &tg->service_queue; | |
453 | ||
454 | /* | |
455 | * If on the default hierarchy, we switch to properly hierarchical | |
456 | * behavior where limits on a given throtl_grp are applied to the | |
457 | * whole subtree rather than just the group itself. e.g. If 16M | |
458 | * read_bps limit is set on the root group, the whole system can't | |
459 | * exceed 16M for the device. | |
460 | * | |
461 | * If not on the default hierarchy, the broken flat hierarchy | |
462 | * behavior is retained where all throtl_grps are treated as if | |
463 | * they're all separate root groups right below throtl_data. | |
464 | * Limits of a group don't interact with limits of other groups | |
465 | * regardless of the position of the group in the hierarchy. | |
466 | */ | |
467 | sq->parent_sq = &td->service_queue; | |
468 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent) | |
469 | sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue; | |
470 | tg->td = td; | |
471 | } | |
472 | ||
473 | /* | |
474 | * Set has_rules[] if @tg or any of its parents have limits configured. | |
475 | * This doesn't require walking up to the top of the hierarchy as the | |
476 | * parent's has_rules[] is guaranteed to be correct. | |
477 | */ | |
478 | static void tg_update_has_rules(struct throtl_grp *tg) | |
479 | { | |
480 | struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); | |
481 | struct throtl_data *td = tg->td; | |
482 | int rw; | |
483 | ||
484 | for (rw = READ; rw <= WRITE; rw++) | |
485 | tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) || | |
486 | (td->limit_valid[td->limit_index] && | |
487 | (tg_bps_limit(tg, rw) != U64_MAX || | |
488 | tg_iops_limit(tg, rw) != UINT_MAX)); | |
489 | } | |
490 | ||
491 | static void throtl_pd_online(struct blkg_policy_data *pd) | |
492 | { | |
493 | struct throtl_grp *tg = pd_to_tg(pd); | |
494 | /* | |
495 | * We don't want new groups to escape the limits of its ancestors. | |
496 | * Update has_rules[] after a new group is brought online. | |
497 | */ | |
498 | tg_update_has_rules(tg); | |
499 | tg->last_dispatch_time[READ] = jiffies; | |
500 | tg->last_dispatch_time[WRITE] = jiffies; | |
501 | } | |
502 | ||
503 | static void blk_throtl_update_limit_valid(struct throtl_data *td) | |
504 | { | |
505 | struct cgroup_subsys_state *pos_css; | |
506 | struct blkcg_gq *blkg; | |
507 | bool low_valid = false; | |
508 | ||
509 | rcu_read_lock(); | |
510 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
511 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
512 | ||
513 | if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] || | |
514 | tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) | |
515 | low_valid = true; | |
516 | } | |
517 | rcu_read_unlock(); | |
518 | ||
519 | td->limit_valid[LIMIT_LOW] = low_valid; | |
520 | } | |
521 | ||
522 | static void throtl_upgrade_state(struct throtl_data *td); | |
523 | static void throtl_pd_offline(struct blkg_policy_data *pd) | |
524 | { | |
525 | struct throtl_grp *tg = pd_to_tg(pd); | |
526 | ||
527 | tg->bps[READ][LIMIT_LOW] = 0; | |
528 | tg->bps[WRITE][LIMIT_LOW] = 0; | |
529 | tg->iops[READ][LIMIT_LOW] = 0; | |
530 | tg->iops[WRITE][LIMIT_LOW] = 0; | |
531 | ||
532 | blk_throtl_update_limit_valid(tg->td); | |
533 | ||
534 | if (!tg->td->limit_valid[tg->td->limit_index]) | |
535 | throtl_upgrade_state(tg->td); | |
536 | } | |
537 | ||
538 | static void throtl_pd_free(struct blkg_policy_data *pd) | |
539 | { | |
540 | struct throtl_grp *tg = pd_to_tg(pd); | |
541 | ||
542 | del_timer_sync(&tg->service_queue.pending_timer); | |
543 | kfree(tg); | |
544 | } | |
545 | ||
546 | static struct throtl_grp * | |
547 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
548 | { | |
549 | /* Service tree is empty */ | |
550 | if (!parent_sq->nr_pending) | |
551 | return NULL; | |
552 | ||
553 | if (!parent_sq->first_pending) | |
554 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
555 | ||
556 | if (parent_sq->first_pending) | |
557 | return rb_entry_tg(parent_sq->first_pending); | |
558 | ||
559 | return NULL; | |
560 | } | |
561 | ||
562 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
563 | { | |
564 | rb_erase(n, root); | |
565 | RB_CLEAR_NODE(n); | |
566 | } | |
567 | ||
568 | static void throtl_rb_erase(struct rb_node *n, | |
569 | struct throtl_service_queue *parent_sq) | |
570 | { | |
571 | if (parent_sq->first_pending == n) | |
572 | parent_sq->first_pending = NULL; | |
573 | rb_erase_init(n, &parent_sq->pending_tree); | |
574 | --parent_sq->nr_pending; | |
575 | } | |
576 | ||
577 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) | |
578 | { | |
579 | struct throtl_grp *tg; | |
580 | ||
581 | tg = throtl_rb_first(parent_sq); | |
582 | if (!tg) | |
583 | return; | |
584 | ||
585 | parent_sq->first_pending_disptime = tg->disptime; | |
586 | } | |
587 | ||
588 | static void tg_service_queue_add(struct throtl_grp *tg) | |
589 | { | |
590 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; | |
591 | struct rb_node **node = &parent_sq->pending_tree.rb_node; | |
592 | struct rb_node *parent = NULL; | |
593 | struct throtl_grp *__tg; | |
594 | unsigned long key = tg->disptime; | |
595 | int left = 1; | |
596 | ||
597 | while (*node != NULL) { | |
598 | parent = *node; | |
599 | __tg = rb_entry_tg(parent); | |
600 | ||
601 | if (time_before(key, __tg->disptime)) | |
602 | node = &parent->rb_left; | |
603 | else { | |
604 | node = &parent->rb_right; | |
605 | left = 0; | |
606 | } | |
607 | } | |
608 | ||
609 | if (left) | |
610 | parent_sq->first_pending = &tg->rb_node; | |
611 | ||
612 | rb_link_node(&tg->rb_node, parent, node); | |
613 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); | |
614 | } | |
615 | ||
616 | static void __throtl_enqueue_tg(struct throtl_grp *tg) | |
617 | { | |
618 | tg_service_queue_add(tg); | |
619 | tg->flags |= THROTL_TG_PENDING; | |
620 | tg->service_queue.parent_sq->nr_pending++; | |
621 | } | |
622 | ||
623 | static void throtl_enqueue_tg(struct throtl_grp *tg) | |
624 | { | |
625 | if (!(tg->flags & THROTL_TG_PENDING)) | |
626 | __throtl_enqueue_tg(tg); | |
627 | } | |
628 | ||
629 | static void __throtl_dequeue_tg(struct throtl_grp *tg) | |
630 | { | |
631 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); | |
632 | tg->flags &= ~THROTL_TG_PENDING; | |
633 | } | |
634 | ||
635 | static void throtl_dequeue_tg(struct throtl_grp *tg) | |
636 | { | |
637 | if (tg->flags & THROTL_TG_PENDING) | |
638 | __throtl_dequeue_tg(tg); | |
639 | } | |
640 | ||
641 | /* Call with queue lock held */ | |
642 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, | |
643 | unsigned long expires) | |
644 | { | |
645 | unsigned long max_expire = jiffies + 8 * sq_to_tg(sq)->td->throtl_slice; | |
646 | ||
647 | /* | |
648 | * Since we are adjusting the throttle limit dynamically, the sleep | |
649 | * time calculated according to previous limit might be invalid. It's | |
650 | * possible the cgroup sleep time is very long and no other cgroups | |
651 | * have IO running so notify the limit changes. Make sure the cgroup | |
652 | * doesn't sleep too long to avoid the missed notification. | |
653 | */ | |
654 | if (time_after(expires, max_expire)) | |
655 | expires = max_expire; | |
656 | mod_timer(&sq->pending_timer, expires); | |
657 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
658 | expires - jiffies, jiffies); | |
659 | } | |
660 | ||
661 | /** | |
662 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle | |
663 | * @sq: the service_queue to schedule dispatch for | |
664 | * @force: force scheduling | |
665 | * | |
666 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the | |
667 | * dispatch time of the first pending child. Returns %true if either timer | |
668 | * is armed or there's no pending child left. %false if the current | |
669 | * dispatch window is still open and the caller should continue | |
670 | * dispatching. | |
671 | * | |
672 | * If @force is %true, the dispatch timer is always scheduled and this | |
673 | * function is guaranteed to return %true. This is to be used when the | |
674 | * caller can't dispatch itself and needs to invoke pending_timer | |
675 | * unconditionally. Note that forced scheduling is likely to induce short | |
676 | * delay before dispatch starts even if @sq->first_pending_disptime is not | |
677 | * in the future and thus shouldn't be used in hot paths. | |
678 | */ | |
679 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
680 | bool force) | |
681 | { | |
682 | /* any pending children left? */ | |
683 | if (!sq->nr_pending) | |
684 | return true; | |
685 | ||
686 | update_min_dispatch_time(sq); | |
687 | ||
688 | /* is the next dispatch time in the future? */ | |
689 | if (force || time_after(sq->first_pending_disptime, jiffies)) { | |
690 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); | |
691 | return true; | |
692 | } | |
693 | ||
694 | /* tell the caller to continue dispatching */ | |
695 | return false; | |
696 | } | |
697 | ||
698 | static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, | |
699 | bool rw, unsigned long start) | |
700 | { | |
701 | tg->bytes_disp[rw] = 0; | |
702 | tg->io_disp[rw] = 0; | |
703 | ||
704 | /* | |
705 | * Previous slice has expired. We must have trimmed it after last | |
706 | * bio dispatch. That means since start of last slice, we never used | |
707 | * that bandwidth. Do try to make use of that bandwidth while giving | |
708 | * credit. | |
709 | */ | |
710 | if (time_after_eq(start, tg->slice_start[rw])) | |
711 | tg->slice_start[rw] = start; | |
712 | ||
713 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; | |
714 | throtl_log(&tg->service_queue, | |
715 | "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", | |
716 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
717 | tg->slice_end[rw], jiffies); | |
718 | } | |
719 | ||
720 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) | |
721 | { | |
722 | tg->bytes_disp[rw] = 0; | |
723 | tg->io_disp[rw] = 0; | |
724 | tg->slice_start[rw] = jiffies; | |
725 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; | |
726 | throtl_log(&tg->service_queue, | |
727 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
728 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
729 | tg->slice_end[rw], jiffies); | |
730 | } | |
731 | ||
732 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, | |
733 | unsigned long jiffy_end) | |
734 | { | |
735 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); | |
736 | } | |
737 | ||
738 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, | |
739 | unsigned long jiffy_end) | |
740 | { | |
741 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); | |
742 | throtl_log(&tg->service_queue, | |
743 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
744 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
745 | tg->slice_end[rw], jiffies); | |
746 | } | |
747 | ||
748 | /* Determine if previously allocated or extended slice is complete or not */ | |
749 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) | |
750 | { | |
751 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
752 | return false; | |
753 | ||
754 | return 1; | |
755 | } | |
756 | ||
757 | /* Trim the used slices and adjust slice start accordingly */ | |
758 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) | |
759 | { | |
760 | unsigned long nr_slices, time_elapsed, io_trim; | |
761 | u64 bytes_trim, tmp; | |
762 | ||
763 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
764 | ||
765 | /* | |
766 | * If bps are unlimited (-1), then time slice don't get | |
767 | * renewed. Don't try to trim the slice if slice is used. A new | |
768 | * slice will start when appropriate. | |
769 | */ | |
770 | if (throtl_slice_used(tg, rw)) | |
771 | return; | |
772 | ||
773 | /* | |
774 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
775 | * that initially cgroup limit was very low resulting in high | |
776 | * slice_end, but later limit was bumped up and bio was dispached | |
777 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
778 | * is bad because it does not allow new slice to start. | |
779 | */ | |
780 | ||
781 | throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice); | |
782 | ||
783 | time_elapsed = jiffies - tg->slice_start[rw]; | |
784 | ||
785 | nr_slices = time_elapsed / tg->td->throtl_slice; | |
786 | ||
787 | if (!nr_slices) | |
788 | return; | |
789 | tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices; | |
790 | do_div(tmp, HZ); | |
791 | bytes_trim = tmp; | |
792 | ||
793 | io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) / | |
794 | HZ; | |
795 | ||
796 | if (!bytes_trim && !io_trim) | |
797 | return; | |
798 | ||
799 | if (tg->bytes_disp[rw] >= bytes_trim) | |
800 | tg->bytes_disp[rw] -= bytes_trim; | |
801 | else | |
802 | tg->bytes_disp[rw] = 0; | |
803 | ||
804 | if (tg->io_disp[rw] >= io_trim) | |
805 | tg->io_disp[rw] -= io_trim; | |
806 | else | |
807 | tg->io_disp[rw] = 0; | |
808 | ||
809 | tg->slice_start[rw] += nr_slices * tg->td->throtl_slice; | |
810 | ||
811 | throtl_log(&tg->service_queue, | |
812 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
813 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
814 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
815 | } | |
816 | ||
817 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, | |
818 | unsigned long *wait) | |
819 | { | |
820 | bool rw = bio_data_dir(bio); | |
821 | unsigned int io_allowed; | |
822 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; | |
823 | u64 tmp; | |
824 | ||
825 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
826 | ||
827 | /* Slice has just started. Consider one slice interval */ | |
828 | if (!jiffy_elapsed) | |
829 | jiffy_elapsed_rnd = tg->td->throtl_slice; | |
830 | ||
831 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); | |
832 | ||
833 | /* | |
834 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
835 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
836 | * will allow dispatch after 1 second and after that slice should | |
837 | * have been trimmed. | |
838 | */ | |
839 | ||
840 | tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd; | |
841 | do_div(tmp, HZ); | |
842 | ||
843 | if (tmp > UINT_MAX) | |
844 | io_allowed = UINT_MAX; | |
845 | else | |
846 | io_allowed = tmp; | |
847 | ||
848 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
849 | if (wait) | |
850 | *wait = 0; | |
851 | return true; | |
852 | } | |
853 | ||
854 | /* Calc approx time to dispatch */ | |
855 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ) / tg_iops_limit(tg, rw) + 1; | |
856 | ||
857 | if (jiffy_wait > jiffy_elapsed) | |
858 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
859 | else | |
860 | jiffy_wait = 1; | |
861 | ||
862 | if (wait) | |
863 | *wait = jiffy_wait; | |
864 | return 0; | |
865 | } | |
866 | ||
867 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, | |
868 | unsigned long *wait) | |
869 | { | |
870 | bool rw = bio_data_dir(bio); | |
871 | u64 bytes_allowed, extra_bytes, tmp; | |
872 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; | |
873 | ||
874 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
875 | ||
876 | /* Slice has just started. Consider one slice interval */ | |
877 | if (!jiffy_elapsed) | |
878 | jiffy_elapsed_rnd = tg->td->throtl_slice; | |
879 | ||
880 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); | |
881 | ||
882 | tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd; | |
883 | do_div(tmp, HZ); | |
884 | bytes_allowed = tmp; | |
885 | ||
886 | if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) { | |
887 | if (wait) | |
888 | *wait = 0; | |
889 | return true; | |
890 | } | |
891 | ||
892 | /* Calc approx time to dispatch */ | |
893 | extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed; | |
894 | jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw)); | |
895 | ||
896 | if (!jiffy_wait) | |
897 | jiffy_wait = 1; | |
898 | ||
899 | /* | |
900 | * This wait time is without taking into consideration the rounding | |
901 | * up we did. Add that time also. | |
902 | */ | |
903 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
904 | if (wait) | |
905 | *wait = jiffy_wait; | |
906 | return 0; | |
907 | } | |
908 | ||
909 | /* | |
910 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
911 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
912 | */ | |
913 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, | |
914 | unsigned long *wait) | |
915 | { | |
916 | bool rw = bio_data_dir(bio); | |
917 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
918 | ||
919 | /* | |
920 | * Currently whole state machine of group depends on first bio | |
921 | * queued in the group bio list. So one should not be calling | |
922 | * this function with a different bio if there are other bios | |
923 | * queued. | |
924 | */ | |
925 | BUG_ON(tg->service_queue.nr_queued[rw] && | |
926 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); | |
927 | ||
928 | /* If tg->bps = -1, then BW is unlimited */ | |
929 | if (tg_bps_limit(tg, rw) == U64_MAX && | |
930 | tg_iops_limit(tg, rw) == UINT_MAX) { | |
931 | if (wait) | |
932 | *wait = 0; | |
933 | return true; | |
934 | } | |
935 | ||
936 | /* | |
937 | * If previous slice expired, start a new one otherwise renew/extend | |
938 | * existing slice to make sure it is at least throtl_slice interval | |
939 | * long since now. New slice is started only for empty throttle group. | |
940 | * If there is queued bio, that means there should be an active | |
941 | * slice and it should be extended instead. | |
942 | */ | |
943 | if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw])) | |
944 | throtl_start_new_slice(tg, rw); | |
945 | else { | |
946 | if (time_before(tg->slice_end[rw], | |
947 | jiffies + tg->td->throtl_slice)) | |
948 | throtl_extend_slice(tg, rw, | |
949 | jiffies + tg->td->throtl_slice); | |
950 | } | |
951 | ||
952 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && | |
953 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
954 | if (wait) | |
955 | *wait = 0; | |
956 | return 1; | |
957 | } | |
958 | ||
959 | max_wait = max(bps_wait, iops_wait); | |
960 | ||
961 | if (wait) | |
962 | *wait = max_wait; | |
963 | ||
964 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
965 | throtl_extend_slice(tg, rw, jiffies + max_wait); | |
966 | ||
967 | return 0; | |
968 | } | |
969 | ||
970 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
971 | { | |
972 | bool rw = bio_data_dir(bio); | |
973 | ||
974 | /* Charge the bio to the group */ | |
975 | tg->bytes_disp[rw] += bio->bi_iter.bi_size; | |
976 | tg->io_disp[rw]++; | |
977 | tg->last_bytes_disp[rw] += bio->bi_iter.bi_size; | |
978 | tg->last_io_disp[rw]++; | |
979 | ||
980 | /* | |
981 | * BIO_THROTTLED is used to prevent the same bio to be throttled | |
982 | * more than once as a throttled bio will go through blk-throtl the | |
983 | * second time when it eventually gets issued. Set it when a bio | |
984 | * is being charged to a tg. | |
985 | */ | |
986 | if (!bio_flagged(bio, BIO_THROTTLED)) | |
987 | bio_set_flag(bio, BIO_THROTTLED); | |
988 | } | |
989 | ||
990 | /** | |
991 | * throtl_add_bio_tg - add a bio to the specified throtl_grp | |
992 | * @bio: bio to add | |
993 | * @qn: qnode to use | |
994 | * @tg: the target throtl_grp | |
995 | * | |
996 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, | |
997 | * tg->qnode_on_self[] is used. | |
998 | */ | |
999 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, | |
1000 | struct throtl_grp *tg) | |
1001 | { | |
1002 | struct throtl_service_queue *sq = &tg->service_queue; | |
1003 | bool rw = bio_data_dir(bio); | |
1004 | ||
1005 | if (!qn) | |
1006 | qn = &tg->qnode_on_self[rw]; | |
1007 | ||
1008 | /* | |
1009 | * If @tg doesn't currently have any bios queued in the same | |
1010 | * direction, queueing @bio can change when @tg should be | |
1011 | * dispatched. Mark that @tg was empty. This is automatically | |
1012 | * cleaered on the next tg_update_disptime(). | |
1013 | */ | |
1014 | if (!sq->nr_queued[rw]) | |
1015 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
1016 | ||
1017 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); | |
1018 | ||
1019 | sq->nr_queued[rw]++; | |
1020 | throtl_enqueue_tg(tg); | |
1021 | } | |
1022 | ||
1023 | static void tg_update_disptime(struct throtl_grp *tg) | |
1024 | { | |
1025 | struct throtl_service_queue *sq = &tg->service_queue; | |
1026 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
1027 | struct bio *bio; | |
1028 | ||
1029 | bio = throtl_peek_queued(&sq->queued[READ]); | |
1030 | if (bio) | |
1031 | tg_may_dispatch(tg, bio, &read_wait); | |
1032 | ||
1033 | bio = throtl_peek_queued(&sq->queued[WRITE]); | |
1034 | if (bio) | |
1035 | tg_may_dispatch(tg, bio, &write_wait); | |
1036 | ||
1037 | min_wait = min(read_wait, write_wait); | |
1038 | disptime = jiffies + min_wait; | |
1039 | ||
1040 | /* Update dispatch time */ | |
1041 | throtl_dequeue_tg(tg); | |
1042 | tg->disptime = disptime; | |
1043 | throtl_enqueue_tg(tg); | |
1044 | ||
1045 | /* see throtl_add_bio_tg() */ | |
1046 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
1047 | } | |
1048 | ||
1049 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, | |
1050 | struct throtl_grp *parent_tg, bool rw) | |
1051 | { | |
1052 | if (throtl_slice_used(parent_tg, rw)) { | |
1053 | throtl_start_new_slice_with_credit(parent_tg, rw, | |
1054 | child_tg->slice_start[rw]); | |
1055 | } | |
1056 | ||
1057 | } | |
1058 | ||
1059 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) | |
1060 | { | |
1061 | struct throtl_service_queue *sq = &tg->service_queue; | |
1062 | struct throtl_service_queue *parent_sq = sq->parent_sq; | |
1063 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
1064 | struct throtl_grp *tg_to_put = NULL; | |
1065 | struct bio *bio; | |
1066 | ||
1067 | /* | |
1068 | * @bio is being transferred from @tg to @parent_sq. Popping a bio | |
1069 | * from @tg may put its reference and @parent_sq might end up | |
1070 | * getting released prematurely. Remember the tg to put and put it | |
1071 | * after @bio is transferred to @parent_sq. | |
1072 | */ | |
1073 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); | |
1074 | sq->nr_queued[rw]--; | |
1075 | ||
1076 | throtl_charge_bio(tg, bio); | |
1077 | ||
1078 | /* | |
1079 | * If our parent is another tg, we just need to transfer @bio to | |
1080 | * the parent using throtl_add_bio_tg(). If our parent is | |
1081 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
1082 | * bio_lists[] and decrease total number queued. The caller is | |
1083 | * responsible for issuing these bios. | |
1084 | */ | |
1085 | if (parent_tg) { | |
1086 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); | |
1087 | start_parent_slice_with_credit(tg, parent_tg, rw); | |
1088 | } else { | |
1089 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], | |
1090 | &parent_sq->queued[rw]); | |
1091 | BUG_ON(tg->td->nr_queued[rw] <= 0); | |
1092 | tg->td->nr_queued[rw]--; | |
1093 | } | |
1094 | ||
1095 | throtl_trim_slice(tg, rw); | |
1096 | ||
1097 | if (tg_to_put) | |
1098 | blkg_put(tg_to_blkg(tg_to_put)); | |
1099 | } | |
1100 | ||
1101 | static int throtl_dispatch_tg(struct throtl_grp *tg) | |
1102 | { | |
1103 | struct throtl_service_queue *sq = &tg->service_queue; | |
1104 | unsigned int nr_reads = 0, nr_writes = 0; | |
1105 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
1106 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; | |
1107 | struct bio *bio; | |
1108 | ||
1109 | /* Try to dispatch 75% READS and 25% WRITES */ | |
1110 | ||
1111 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && | |
1112 | tg_may_dispatch(tg, bio, NULL)) { | |
1113 | ||
1114 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); | |
1115 | nr_reads++; | |
1116 | ||
1117 | if (nr_reads >= max_nr_reads) | |
1118 | break; | |
1119 | } | |
1120 | ||
1121 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && | |
1122 | tg_may_dispatch(tg, bio, NULL)) { | |
1123 | ||
1124 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); | |
1125 | nr_writes++; | |
1126 | ||
1127 | if (nr_writes >= max_nr_writes) | |
1128 | break; | |
1129 | } | |
1130 | ||
1131 | return nr_reads + nr_writes; | |
1132 | } | |
1133 | ||
1134 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) | |
1135 | { | |
1136 | unsigned int nr_disp = 0; | |
1137 | ||
1138 | while (1) { | |
1139 | struct throtl_grp *tg = throtl_rb_first(parent_sq); | |
1140 | struct throtl_service_queue *sq = &tg->service_queue; | |
1141 | ||
1142 | if (!tg) | |
1143 | break; | |
1144 | ||
1145 | if (time_before(jiffies, tg->disptime)) | |
1146 | break; | |
1147 | ||
1148 | throtl_dequeue_tg(tg); | |
1149 | ||
1150 | nr_disp += throtl_dispatch_tg(tg); | |
1151 | ||
1152 | if (sq->nr_queued[0] || sq->nr_queued[1]) | |
1153 | tg_update_disptime(tg); | |
1154 | ||
1155 | if (nr_disp >= throtl_quantum) | |
1156 | break; | |
1157 | } | |
1158 | ||
1159 | return nr_disp; | |
1160 | } | |
1161 | ||
1162 | static bool throtl_can_upgrade(struct throtl_data *td, | |
1163 | struct throtl_grp *this_tg); | |
1164 | /** | |
1165 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
1166 | * @arg: the throtl_service_queue being serviced | |
1167 | * | |
1168 | * This timer is armed when a child throtl_grp with active bio's become | |
1169 | * pending and queued on the service_queue's pending_tree and expires when | |
1170 | * the first child throtl_grp should be dispatched. This function | |
1171 | * dispatches bio's from the children throtl_grps to the parent | |
1172 | * service_queue. | |
1173 | * | |
1174 | * If the parent's parent is another throtl_grp, dispatching is propagated | |
1175 | * by either arming its pending_timer or repeating dispatch directly. If | |
1176 | * the top-level service_tree is reached, throtl_data->dispatch_work is | |
1177 | * kicked so that the ready bio's are issued. | |
1178 | */ | |
1179 | static void throtl_pending_timer_fn(unsigned long arg) | |
1180 | { | |
1181 | struct throtl_service_queue *sq = (void *)arg; | |
1182 | struct throtl_grp *tg = sq_to_tg(sq); | |
1183 | struct throtl_data *td = sq_to_td(sq); | |
1184 | struct request_queue *q = td->queue; | |
1185 | struct throtl_service_queue *parent_sq; | |
1186 | bool dispatched; | |
1187 | int ret; | |
1188 | ||
1189 | spin_lock_irq(q->queue_lock); | |
1190 | if (throtl_can_upgrade(td, NULL)) | |
1191 | throtl_upgrade_state(td); | |
1192 | ||
1193 | again: | |
1194 | parent_sq = sq->parent_sq; | |
1195 | dispatched = false; | |
1196 | ||
1197 | while (true) { | |
1198 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
1199 | sq->nr_queued[READ] + sq->nr_queued[WRITE], | |
1200 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
1201 | ||
1202 | ret = throtl_select_dispatch(sq); | |
1203 | if (ret) { | |
1204 | throtl_log(sq, "bios disp=%u", ret); | |
1205 | dispatched = true; | |
1206 | } | |
1207 | ||
1208 | if (throtl_schedule_next_dispatch(sq, false)) | |
1209 | break; | |
1210 | ||
1211 | /* this dispatch windows is still open, relax and repeat */ | |
1212 | spin_unlock_irq(q->queue_lock); | |
1213 | cpu_relax(); | |
1214 | spin_lock_irq(q->queue_lock); | |
1215 | } | |
1216 | ||
1217 | if (!dispatched) | |
1218 | goto out_unlock; | |
1219 | ||
1220 | if (parent_sq) { | |
1221 | /* @parent_sq is another throl_grp, propagate dispatch */ | |
1222 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1223 | tg_update_disptime(tg); | |
1224 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { | |
1225 | /* window is already open, repeat dispatching */ | |
1226 | sq = parent_sq; | |
1227 | tg = sq_to_tg(sq); | |
1228 | goto again; | |
1229 | } | |
1230 | } | |
1231 | } else { | |
1232 | /* reached the top-level, queue issueing */ | |
1233 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1234 | } | |
1235 | out_unlock: | |
1236 | spin_unlock_irq(q->queue_lock); | |
1237 | } | |
1238 | ||
1239 | /** | |
1240 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
1241 | * @work: work item being executed | |
1242 | * | |
1243 | * This function is queued for execution when bio's reach the bio_lists[] | |
1244 | * of throtl_data->service_queue. Those bio's are ready and issued by this | |
1245 | * function. | |
1246 | */ | |
1247 | static void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
1248 | { | |
1249 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1250 | dispatch_work); | |
1251 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1252 | struct request_queue *q = td->queue; | |
1253 | struct bio_list bio_list_on_stack; | |
1254 | struct bio *bio; | |
1255 | struct blk_plug plug; | |
1256 | int rw; | |
1257 | ||
1258 | bio_list_init(&bio_list_on_stack); | |
1259 | ||
1260 | spin_lock_irq(q->queue_lock); | |
1261 | for (rw = READ; rw <= WRITE; rw++) | |
1262 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) | |
1263 | bio_list_add(&bio_list_on_stack, bio); | |
1264 | spin_unlock_irq(q->queue_lock); | |
1265 | ||
1266 | if (!bio_list_empty(&bio_list_on_stack)) { | |
1267 | blk_start_plug(&plug); | |
1268 | while((bio = bio_list_pop(&bio_list_on_stack))) | |
1269 | generic_make_request(bio); | |
1270 | blk_finish_plug(&plug); | |
1271 | } | |
1272 | } | |
1273 | ||
1274 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, | |
1275 | int off) | |
1276 | { | |
1277 | struct throtl_grp *tg = pd_to_tg(pd); | |
1278 | u64 v = *(u64 *)((void *)tg + off); | |
1279 | ||
1280 | if (v == U64_MAX) | |
1281 | return 0; | |
1282 | return __blkg_prfill_u64(sf, pd, v); | |
1283 | } | |
1284 | ||
1285 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, | |
1286 | int off) | |
1287 | { | |
1288 | struct throtl_grp *tg = pd_to_tg(pd); | |
1289 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
1290 | ||
1291 | if (v == UINT_MAX) | |
1292 | return 0; | |
1293 | return __blkg_prfill_u64(sf, pd, v); | |
1294 | } | |
1295 | ||
1296 | static int tg_print_conf_u64(struct seq_file *sf, void *v) | |
1297 | { | |
1298 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64, | |
1299 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
1300 | return 0; | |
1301 | } | |
1302 | ||
1303 | static int tg_print_conf_uint(struct seq_file *sf, void *v) | |
1304 | { | |
1305 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint, | |
1306 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
1307 | return 0; | |
1308 | } | |
1309 | ||
1310 | static void tg_conf_updated(struct throtl_grp *tg) | |
1311 | { | |
1312 | struct throtl_service_queue *sq = &tg->service_queue; | |
1313 | struct cgroup_subsys_state *pos_css; | |
1314 | struct blkcg_gq *blkg; | |
1315 | ||
1316 | throtl_log(&tg->service_queue, | |
1317 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
1318 | tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE), | |
1319 | tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE)); | |
1320 | ||
1321 | /* | |
1322 | * Update has_rules[] flags for the updated tg's subtree. A tg is | |
1323 | * considered to have rules if either the tg itself or any of its | |
1324 | * ancestors has rules. This identifies groups without any | |
1325 | * restrictions in the whole hierarchy and allows them to bypass | |
1326 | * blk-throttle. | |
1327 | */ | |
1328 | blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg)) | |
1329 | tg_update_has_rules(blkg_to_tg(blkg)); | |
1330 | ||
1331 | /* | |
1332 | * We're already holding queue_lock and know @tg is valid. Let's | |
1333 | * apply the new config directly. | |
1334 | * | |
1335 | * Restart the slices for both READ and WRITES. It might happen | |
1336 | * that a group's limit are dropped suddenly and we don't want to | |
1337 | * account recently dispatched IO with new low rate. | |
1338 | */ | |
1339 | throtl_start_new_slice(tg, 0); | |
1340 | throtl_start_new_slice(tg, 1); | |
1341 | ||
1342 | if (tg->flags & THROTL_TG_PENDING) { | |
1343 | tg_update_disptime(tg); | |
1344 | throtl_schedule_next_dispatch(sq->parent_sq, true); | |
1345 | } | |
1346 | } | |
1347 | ||
1348 | static ssize_t tg_set_conf(struct kernfs_open_file *of, | |
1349 | char *buf, size_t nbytes, loff_t off, bool is_u64) | |
1350 | { | |
1351 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1352 | struct blkg_conf_ctx ctx; | |
1353 | struct throtl_grp *tg; | |
1354 | int ret; | |
1355 | u64 v; | |
1356 | ||
1357 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1358 | if (ret) | |
1359 | return ret; | |
1360 | ||
1361 | ret = -EINVAL; | |
1362 | if (sscanf(ctx.body, "%llu", &v) != 1) | |
1363 | goto out_finish; | |
1364 | if (!v) | |
1365 | v = U64_MAX; | |
1366 | ||
1367 | tg = blkg_to_tg(ctx.blkg); | |
1368 | ||
1369 | if (is_u64) | |
1370 | *(u64 *)((void *)tg + of_cft(of)->private) = v; | |
1371 | else | |
1372 | *(unsigned int *)((void *)tg + of_cft(of)->private) = v; | |
1373 | ||
1374 | tg_conf_updated(tg); | |
1375 | ret = 0; | |
1376 | out_finish: | |
1377 | blkg_conf_finish(&ctx); | |
1378 | return ret ?: nbytes; | |
1379 | } | |
1380 | ||
1381 | static ssize_t tg_set_conf_u64(struct kernfs_open_file *of, | |
1382 | char *buf, size_t nbytes, loff_t off) | |
1383 | { | |
1384 | return tg_set_conf(of, buf, nbytes, off, true); | |
1385 | } | |
1386 | ||
1387 | static ssize_t tg_set_conf_uint(struct kernfs_open_file *of, | |
1388 | char *buf, size_t nbytes, loff_t off) | |
1389 | { | |
1390 | return tg_set_conf(of, buf, nbytes, off, false); | |
1391 | } | |
1392 | ||
1393 | static struct cftype throtl_legacy_files[] = { | |
1394 | { | |
1395 | .name = "throttle.read_bps_device", | |
1396 | .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]), | |
1397 | .seq_show = tg_print_conf_u64, | |
1398 | .write = tg_set_conf_u64, | |
1399 | }, | |
1400 | { | |
1401 | .name = "throttle.write_bps_device", | |
1402 | .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]), | |
1403 | .seq_show = tg_print_conf_u64, | |
1404 | .write = tg_set_conf_u64, | |
1405 | }, | |
1406 | { | |
1407 | .name = "throttle.read_iops_device", | |
1408 | .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]), | |
1409 | .seq_show = tg_print_conf_uint, | |
1410 | .write = tg_set_conf_uint, | |
1411 | }, | |
1412 | { | |
1413 | .name = "throttle.write_iops_device", | |
1414 | .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]), | |
1415 | .seq_show = tg_print_conf_uint, | |
1416 | .write = tg_set_conf_uint, | |
1417 | }, | |
1418 | { | |
1419 | .name = "throttle.io_service_bytes", | |
1420 | .private = (unsigned long)&blkcg_policy_throtl, | |
1421 | .seq_show = blkg_print_stat_bytes, | |
1422 | }, | |
1423 | { | |
1424 | .name = "throttle.io_serviced", | |
1425 | .private = (unsigned long)&blkcg_policy_throtl, | |
1426 | .seq_show = blkg_print_stat_ios, | |
1427 | }, | |
1428 | { } /* terminate */ | |
1429 | }; | |
1430 | ||
1431 | static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd, | |
1432 | int off) | |
1433 | { | |
1434 | struct throtl_grp *tg = pd_to_tg(pd); | |
1435 | const char *dname = blkg_dev_name(pd->blkg); | |
1436 | char bufs[4][21] = { "max", "max", "max", "max" }; | |
1437 | u64 bps_dft; | |
1438 | unsigned int iops_dft; | |
1439 | ||
1440 | if (!dname) | |
1441 | return 0; | |
1442 | ||
1443 | if (off == LIMIT_LOW) { | |
1444 | bps_dft = 0; | |
1445 | iops_dft = 0; | |
1446 | } else { | |
1447 | bps_dft = U64_MAX; | |
1448 | iops_dft = UINT_MAX; | |
1449 | } | |
1450 | ||
1451 | if (tg->bps_conf[READ][off] == bps_dft && | |
1452 | tg->bps_conf[WRITE][off] == bps_dft && | |
1453 | tg->iops_conf[READ][off] == iops_dft && | |
1454 | tg->iops_conf[WRITE][off] == iops_dft) | |
1455 | return 0; | |
1456 | ||
1457 | if (tg->bps_conf[READ][off] != bps_dft) | |
1458 | snprintf(bufs[0], sizeof(bufs[0]), "%llu", | |
1459 | tg->bps_conf[READ][off]); | |
1460 | if (tg->bps_conf[WRITE][off] != bps_dft) | |
1461 | snprintf(bufs[1], sizeof(bufs[1]), "%llu", | |
1462 | tg->bps_conf[WRITE][off]); | |
1463 | if (tg->iops_conf[READ][off] != iops_dft) | |
1464 | snprintf(bufs[2], sizeof(bufs[2]), "%u", | |
1465 | tg->iops_conf[READ][off]); | |
1466 | if (tg->iops_conf[WRITE][off] != iops_dft) | |
1467 | snprintf(bufs[3], sizeof(bufs[3]), "%u", | |
1468 | tg->iops_conf[WRITE][off]); | |
1469 | ||
1470 | seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s\n", | |
1471 | dname, bufs[0], bufs[1], bufs[2], bufs[3]); | |
1472 | return 0; | |
1473 | } | |
1474 | ||
1475 | static int tg_print_limit(struct seq_file *sf, void *v) | |
1476 | { | |
1477 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit, | |
1478 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
1479 | return 0; | |
1480 | } | |
1481 | ||
1482 | static ssize_t tg_set_limit(struct kernfs_open_file *of, | |
1483 | char *buf, size_t nbytes, loff_t off) | |
1484 | { | |
1485 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1486 | struct blkg_conf_ctx ctx; | |
1487 | struct throtl_grp *tg; | |
1488 | u64 v[4]; | |
1489 | int ret; | |
1490 | int index = of_cft(of)->private; | |
1491 | ||
1492 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1493 | if (ret) | |
1494 | return ret; | |
1495 | ||
1496 | tg = blkg_to_tg(ctx.blkg); | |
1497 | ||
1498 | v[0] = tg->bps_conf[READ][index]; | |
1499 | v[1] = tg->bps_conf[WRITE][index]; | |
1500 | v[2] = tg->iops_conf[READ][index]; | |
1501 | v[3] = tg->iops_conf[WRITE][index]; | |
1502 | ||
1503 | while (true) { | |
1504 | char tok[27]; /* wiops=18446744073709551616 */ | |
1505 | char *p; | |
1506 | u64 val = U64_MAX; | |
1507 | int len; | |
1508 | ||
1509 | if (sscanf(ctx.body, "%26s%n", tok, &len) != 1) | |
1510 | break; | |
1511 | if (tok[0] == '\0') | |
1512 | break; | |
1513 | ctx.body += len; | |
1514 | ||
1515 | ret = -EINVAL; | |
1516 | p = tok; | |
1517 | strsep(&p, "="); | |
1518 | if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max"))) | |
1519 | goto out_finish; | |
1520 | ||
1521 | ret = -ERANGE; | |
1522 | if (!val) | |
1523 | goto out_finish; | |
1524 | ||
1525 | ret = -EINVAL; | |
1526 | if (!strcmp(tok, "rbps")) | |
1527 | v[0] = val; | |
1528 | else if (!strcmp(tok, "wbps")) | |
1529 | v[1] = val; | |
1530 | else if (!strcmp(tok, "riops")) | |
1531 | v[2] = min_t(u64, val, UINT_MAX); | |
1532 | else if (!strcmp(tok, "wiops")) | |
1533 | v[3] = min_t(u64, val, UINT_MAX); | |
1534 | else | |
1535 | goto out_finish; | |
1536 | } | |
1537 | ||
1538 | tg->bps_conf[READ][index] = v[0]; | |
1539 | tg->bps_conf[WRITE][index] = v[1]; | |
1540 | tg->iops_conf[READ][index] = v[2]; | |
1541 | tg->iops_conf[WRITE][index] = v[3]; | |
1542 | ||
1543 | if (index == LIMIT_MAX) { | |
1544 | tg->bps[READ][index] = v[0]; | |
1545 | tg->bps[WRITE][index] = v[1]; | |
1546 | tg->iops[READ][index] = v[2]; | |
1547 | tg->iops[WRITE][index] = v[3]; | |
1548 | } | |
1549 | tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW], | |
1550 | tg->bps_conf[READ][LIMIT_MAX]); | |
1551 | tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW], | |
1552 | tg->bps_conf[WRITE][LIMIT_MAX]); | |
1553 | tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW], | |
1554 | tg->iops_conf[READ][LIMIT_MAX]); | |
1555 | tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW], | |
1556 | tg->iops_conf[WRITE][LIMIT_MAX]); | |
1557 | ||
1558 | if (index == LIMIT_LOW) { | |
1559 | blk_throtl_update_limit_valid(tg->td); | |
1560 | if (tg->td->limit_valid[LIMIT_LOW]) | |
1561 | tg->td->limit_index = LIMIT_LOW; | |
1562 | } | |
1563 | tg_conf_updated(tg); | |
1564 | ret = 0; | |
1565 | out_finish: | |
1566 | blkg_conf_finish(&ctx); | |
1567 | return ret ?: nbytes; | |
1568 | } | |
1569 | ||
1570 | static struct cftype throtl_files[] = { | |
1571 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
1572 | { | |
1573 | .name = "low", | |
1574 | .flags = CFTYPE_NOT_ON_ROOT, | |
1575 | .seq_show = tg_print_limit, | |
1576 | .write = tg_set_limit, | |
1577 | .private = LIMIT_LOW, | |
1578 | }, | |
1579 | #endif | |
1580 | { | |
1581 | .name = "max", | |
1582 | .flags = CFTYPE_NOT_ON_ROOT, | |
1583 | .seq_show = tg_print_limit, | |
1584 | .write = tg_set_limit, | |
1585 | .private = LIMIT_MAX, | |
1586 | }, | |
1587 | { } /* terminate */ | |
1588 | }; | |
1589 | ||
1590 | static void throtl_shutdown_wq(struct request_queue *q) | |
1591 | { | |
1592 | struct throtl_data *td = q->td; | |
1593 | ||
1594 | cancel_work_sync(&td->dispatch_work); | |
1595 | } | |
1596 | ||
1597 | static struct blkcg_policy blkcg_policy_throtl = { | |
1598 | .dfl_cftypes = throtl_files, | |
1599 | .legacy_cftypes = throtl_legacy_files, | |
1600 | ||
1601 | .pd_alloc_fn = throtl_pd_alloc, | |
1602 | .pd_init_fn = throtl_pd_init, | |
1603 | .pd_online_fn = throtl_pd_online, | |
1604 | .pd_offline_fn = throtl_pd_offline, | |
1605 | .pd_free_fn = throtl_pd_free, | |
1606 | }; | |
1607 | ||
1608 | static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg) | |
1609 | { | |
1610 | unsigned long rtime = jiffies, wtime = jiffies; | |
1611 | ||
1612 | if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]) | |
1613 | rtime = tg->last_low_overflow_time[READ]; | |
1614 | if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) | |
1615 | wtime = tg->last_low_overflow_time[WRITE]; | |
1616 | return min(rtime, wtime); | |
1617 | } | |
1618 | ||
1619 | /* tg should not be an intermediate node */ | |
1620 | static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg) | |
1621 | { | |
1622 | struct throtl_service_queue *parent_sq; | |
1623 | struct throtl_grp *parent = tg; | |
1624 | unsigned long ret = __tg_last_low_overflow_time(tg); | |
1625 | ||
1626 | while (true) { | |
1627 | parent_sq = parent->service_queue.parent_sq; | |
1628 | parent = sq_to_tg(parent_sq); | |
1629 | if (!parent) | |
1630 | break; | |
1631 | ||
1632 | /* | |
1633 | * The parent doesn't have low limit, it always reaches low | |
1634 | * limit. Its overflow time is useless for children | |
1635 | */ | |
1636 | if (!parent->bps[READ][LIMIT_LOW] && | |
1637 | !parent->iops[READ][LIMIT_LOW] && | |
1638 | !parent->bps[WRITE][LIMIT_LOW] && | |
1639 | !parent->iops[WRITE][LIMIT_LOW]) | |
1640 | continue; | |
1641 | if (time_after(__tg_last_low_overflow_time(parent), ret)) | |
1642 | ret = __tg_last_low_overflow_time(parent); | |
1643 | } | |
1644 | return ret; | |
1645 | } | |
1646 | ||
1647 | static bool throtl_tg_can_upgrade(struct throtl_grp *tg) | |
1648 | { | |
1649 | struct throtl_service_queue *sq = &tg->service_queue; | |
1650 | bool read_limit, write_limit; | |
1651 | ||
1652 | /* | |
1653 | * if cgroup reaches low limit (if low limit is 0, the cgroup always | |
1654 | * reaches), it's ok to upgrade to next limit | |
1655 | */ | |
1656 | read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]; | |
1657 | write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]; | |
1658 | if (!read_limit && !write_limit) | |
1659 | return true; | |
1660 | if (read_limit && sq->nr_queued[READ] && | |
1661 | (!write_limit || sq->nr_queued[WRITE])) | |
1662 | return true; | |
1663 | if (write_limit && sq->nr_queued[WRITE] && | |
1664 | (!read_limit || sq->nr_queued[READ])) | |
1665 | return true; | |
1666 | ||
1667 | if (time_after_eq(jiffies, | |
1668 | tg->last_dispatch_time[READ] + tg->td->throtl_slice) && | |
1669 | time_after_eq(jiffies, | |
1670 | tg->last_dispatch_time[WRITE] + tg->td->throtl_slice)) | |
1671 | return true; | |
1672 | return false; | |
1673 | } | |
1674 | ||
1675 | static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg) | |
1676 | { | |
1677 | while (true) { | |
1678 | if (throtl_tg_can_upgrade(tg)) | |
1679 | return true; | |
1680 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1681 | if (!tg || !tg_to_blkg(tg)->parent) | |
1682 | return false; | |
1683 | } | |
1684 | return false; | |
1685 | } | |
1686 | ||
1687 | static bool throtl_can_upgrade(struct throtl_data *td, | |
1688 | struct throtl_grp *this_tg) | |
1689 | { | |
1690 | struct cgroup_subsys_state *pos_css; | |
1691 | struct blkcg_gq *blkg; | |
1692 | ||
1693 | if (td->limit_index != LIMIT_LOW) | |
1694 | return false; | |
1695 | ||
1696 | if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice)) | |
1697 | return false; | |
1698 | ||
1699 | rcu_read_lock(); | |
1700 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1701 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1702 | ||
1703 | if (tg == this_tg) | |
1704 | continue; | |
1705 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1706 | continue; | |
1707 | if (!throtl_hierarchy_can_upgrade(tg)) { | |
1708 | rcu_read_unlock(); | |
1709 | return false; | |
1710 | } | |
1711 | } | |
1712 | rcu_read_unlock(); | |
1713 | return true; | |
1714 | } | |
1715 | ||
1716 | static void throtl_upgrade_state(struct throtl_data *td) | |
1717 | { | |
1718 | struct cgroup_subsys_state *pos_css; | |
1719 | struct blkcg_gq *blkg; | |
1720 | ||
1721 | td->limit_index = LIMIT_MAX; | |
1722 | td->low_upgrade_time = jiffies; | |
1723 | td->scale = 0; | |
1724 | rcu_read_lock(); | |
1725 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1726 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1727 | struct throtl_service_queue *sq = &tg->service_queue; | |
1728 | ||
1729 | tg->disptime = jiffies - 1; | |
1730 | throtl_select_dispatch(sq); | |
1731 | throtl_schedule_next_dispatch(sq, false); | |
1732 | } | |
1733 | rcu_read_unlock(); | |
1734 | throtl_select_dispatch(&td->service_queue); | |
1735 | throtl_schedule_next_dispatch(&td->service_queue, false); | |
1736 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1737 | } | |
1738 | ||
1739 | static void throtl_downgrade_state(struct throtl_data *td, int new) | |
1740 | { | |
1741 | td->scale /= 2; | |
1742 | ||
1743 | if (td->scale) { | |
1744 | td->low_upgrade_time = jiffies - td->scale * td->throtl_slice; | |
1745 | return; | |
1746 | } | |
1747 | ||
1748 | td->limit_index = new; | |
1749 | td->low_downgrade_time = jiffies; | |
1750 | } | |
1751 | ||
1752 | static bool throtl_tg_can_downgrade(struct throtl_grp *tg) | |
1753 | { | |
1754 | struct throtl_data *td = tg->td; | |
1755 | unsigned long now = jiffies; | |
1756 | ||
1757 | if (time_after_eq(now, tg->last_dispatch_time[READ] + | |
1758 | td->throtl_slice) && | |
1759 | time_after_eq(now, tg->last_dispatch_time[WRITE] + | |
1760 | td->throtl_slice)) | |
1761 | return false; | |
1762 | /* | |
1763 | * If cgroup is below low limit, consider downgrade and throttle other | |
1764 | * cgroups | |
1765 | */ | |
1766 | if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) && | |
1767 | time_after_eq(now, tg_last_low_overflow_time(tg) + | |
1768 | td->throtl_slice)) | |
1769 | return true; | |
1770 | return false; | |
1771 | } | |
1772 | ||
1773 | static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg) | |
1774 | { | |
1775 | while (true) { | |
1776 | if (!throtl_tg_can_downgrade(tg)) | |
1777 | return false; | |
1778 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1779 | if (!tg || !tg_to_blkg(tg)->parent) | |
1780 | break; | |
1781 | } | |
1782 | return true; | |
1783 | } | |
1784 | ||
1785 | static void throtl_downgrade_check(struct throtl_grp *tg) | |
1786 | { | |
1787 | uint64_t bps; | |
1788 | unsigned int iops; | |
1789 | unsigned long elapsed_time; | |
1790 | unsigned long now = jiffies; | |
1791 | ||
1792 | if (tg->td->limit_index != LIMIT_MAX || | |
1793 | !tg->td->limit_valid[LIMIT_LOW]) | |
1794 | return; | |
1795 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1796 | return; | |
1797 | if (time_after(tg->last_check_time + tg->td->throtl_slice, now)) | |
1798 | return; | |
1799 | ||
1800 | elapsed_time = now - tg->last_check_time; | |
1801 | tg->last_check_time = now; | |
1802 | ||
1803 | if (time_before(now, tg_last_low_overflow_time(tg) + | |
1804 | tg->td->throtl_slice)) | |
1805 | return; | |
1806 | ||
1807 | if (tg->bps[READ][LIMIT_LOW]) { | |
1808 | bps = tg->last_bytes_disp[READ] * HZ; | |
1809 | do_div(bps, elapsed_time); | |
1810 | if (bps >= tg->bps[READ][LIMIT_LOW]) | |
1811 | tg->last_low_overflow_time[READ] = now; | |
1812 | } | |
1813 | ||
1814 | if (tg->bps[WRITE][LIMIT_LOW]) { | |
1815 | bps = tg->last_bytes_disp[WRITE] * HZ; | |
1816 | do_div(bps, elapsed_time); | |
1817 | if (bps >= tg->bps[WRITE][LIMIT_LOW]) | |
1818 | tg->last_low_overflow_time[WRITE] = now; | |
1819 | } | |
1820 | ||
1821 | if (tg->iops[READ][LIMIT_LOW]) { | |
1822 | iops = tg->last_io_disp[READ] * HZ / elapsed_time; | |
1823 | if (iops >= tg->iops[READ][LIMIT_LOW]) | |
1824 | tg->last_low_overflow_time[READ] = now; | |
1825 | } | |
1826 | ||
1827 | if (tg->iops[WRITE][LIMIT_LOW]) { | |
1828 | iops = tg->last_io_disp[WRITE] * HZ / elapsed_time; | |
1829 | if (iops >= tg->iops[WRITE][LIMIT_LOW]) | |
1830 | tg->last_low_overflow_time[WRITE] = now; | |
1831 | } | |
1832 | ||
1833 | /* | |
1834 | * If cgroup is below low limit, consider downgrade and throttle other | |
1835 | * cgroups | |
1836 | */ | |
1837 | if (throtl_hierarchy_can_downgrade(tg)) | |
1838 | throtl_downgrade_state(tg->td, LIMIT_LOW); | |
1839 | ||
1840 | tg->last_bytes_disp[READ] = 0; | |
1841 | tg->last_bytes_disp[WRITE] = 0; | |
1842 | tg->last_io_disp[READ] = 0; | |
1843 | tg->last_io_disp[WRITE] = 0; | |
1844 | } | |
1845 | ||
1846 | bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg, | |
1847 | struct bio *bio) | |
1848 | { | |
1849 | struct throtl_qnode *qn = NULL; | |
1850 | struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg); | |
1851 | struct throtl_service_queue *sq; | |
1852 | bool rw = bio_data_dir(bio); | |
1853 | bool throttled = false; | |
1854 | ||
1855 | WARN_ON_ONCE(!rcu_read_lock_held()); | |
1856 | ||
1857 | /* see throtl_charge_bio() */ | |
1858 | if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw]) | |
1859 | goto out; | |
1860 | ||
1861 | spin_lock_irq(q->queue_lock); | |
1862 | ||
1863 | if (unlikely(blk_queue_bypass(q))) | |
1864 | goto out_unlock; | |
1865 | ||
1866 | sq = &tg->service_queue; | |
1867 | ||
1868 | again: | |
1869 | while (true) { | |
1870 | tg->last_dispatch_time[rw] = jiffies; | |
1871 | if (tg->last_low_overflow_time[rw] == 0) | |
1872 | tg->last_low_overflow_time[rw] = jiffies; | |
1873 | throtl_downgrade_check(tg); | |
1874 | /* throtl is FIFO - if bios are already queued, should queue */ | |
1875 | if (sq->nr_queued[rw]) | |
1876 | break; | |
1877 | ||
1878 | /* if above limits, break to queue */ | |
1879 | if (!tg_may_dispatch(tg, bio, NULL)) { | |
1880 | tg->last_low_overflow_time[rw] = jiffies; | |
1881 | if (throtl_can_upgrade(tg->td, tg)) { | |
1882 | throtl_upgrade_state(tg->td); | |
1883 | goto again; | |
1884 | } | |
1885 | break; | |
1886 | } | |
1887 | ||
1888 | /* within limits, let's charge and dispatch directly */ | |
1889 | throtl_charge_bio(tg, bio); | |
1890 | ||
1891 | /* | |
1892 | * We need to trim slice even when bios are not being queued | |
1893 | * otherwise it might happen that a bio is not queued for | |
1894 | * a long time and slice keeps on extending and trim is not | |
1895 | * called for a long time. Now if limits are reduced suddenly | |
1896 | * we take into account all the IO dispatched so far at new | |
1897 | * low rate and * newly queued IO gets a really long dispatch | |
1898 | * time. | |
1899 | * | |
1900 | * So keep on trimming slice even if bio is not queued. | |
1901 | */ | |
1902 | throtl_trim_slice(tg, rw); | |
1903 | ||
1904 | /* | |
1905 | * @bio passed through this layer without being throttled. | |
1906 | * Climb up the ladder. If we''re already at the top, it | |
1907 | * can be executed directly. | |
1908 | */ | |
1909 | qn = &tg->qnode_on_parent[rw]; | |
1910 | sq = sq->parent_sq; | |
1911 | tg = sq_to_tg(sq); | |
1912 | if (!tg) | |
1913 | goto out_unlock; | |
1914 | } | |
1915 | ||
1916 | /* out-of-limit, queue to @tg */ | |
1917 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", | |
1918 | rw == READ ? 'R' : 'W', | |
1919 | tg->bytes_disp[rw], bio->bi_iter.bi_size, | |
1920 | tg_bps_limit(tg, rw), | |
1921 | tg->io_disp[rw], tg_iops_limit(tg, rw), | |
1922 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
1923 | ||
1924 | tg->last_low_overflow_time[rw] = jiffies; | |
1925 | ||
1926 | bio_associate_current(bio); | |
1927 | tg->td->nr_queued[rw]++; | |
1928 | throtl_add_bio_tg(bio, qn, tg); | |
1929 | throttled = true; | |
1930 | ||
1931 | /* | |
1932 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
1933 | * was empty before @bio. The forced scheduling isn't likely to | |
1934 | * cause undue delay as @bio is likely to be dispatched directly if | |
1935 | * its @tg's disptime is not in the future. | |
1936 | */ | |
1937 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1938 | tg_update_disptime(tg); | |
1939 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); | |
1940 | } | |
1941 | ||
1942 | out_unlock: | |
1943 | spin_unlock_irq(q->queue_lock); | |
1944 | out: | |
1945 | /* | |
1946 | * As multiple blk-throtls may stack in the same issue path, we | |
1947 | * don't want bios to leave with the flag set. Clear the flag if | |
1948 | * being issued. | |
1949 | */ | |
1950 | if (!throttled) | |
1951 | bio_clear_flag(bio, BIO_THROTTLED); | |
1952 | return throttled; | |
1953 | } | |
1954 | ||
1955 | /* | |
1956 | * Dispatch all bios from all children tg's queued on @parent_sq. On | |
1957 | * return, @parent_sq is guaranteed to not have any active children tg's | |
1958 | * and all bios from previously active tg's are on @parent_sq->bio_lists[]. | |
1959 | */ | |
1960 | static void tg_drain_bios(struct throtl_service_queue *parent_sq) | |
1961 | { | |
1962 | struct throtl_grp *tg; | |
1963 | ||
1964 | while ((tg = throtl_rb_first(parent_sq))) { | |
1965 | struct throtl_service_queue *sq = &tg->service_queue; | |
1966 | struct bio *bio; | |
1967 | ||
1968 | throtl_dequeue_tg(tg); | |
1969 | ||
1970 | while ((bio = throtl_peek_queued(&sq->queued[READ]))) | |
1971 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); | |
1972 | while ((bio = throtl_peek_queued(&sq->queued[WRITE]))) | |
1973 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); | |
1974 | } | |
1975 | } | |
1976 | ||
1977 | /** | |
1978 | * blk_throtl_drain - drain throttled bios | |
1979 | * @q: request_queue to drain throttled bios for | |
1980 | * | |
1981 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1982 | */ | |
1983 | void blk_throtl_drain(struct request_queue *q) | |
1984 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1985 | { | |
1986 | struct throtl_data *td = q->td; | |
1987 | struct blkcg_gq *blkg; | |
1988 | struct cgroup_subsys_state *pos_css; | |
1989 | struct bio *bio; | |
1990 | int rw; | |
1991 | ||
1992 | queue_lockdep_assert_held(q); | |
1993 | rcu_read_lock(); | |
1994 | ||
1995 | /* | |
1996 | * Drain each tg while doing post-order walk on the blkg tree, so | |
1997 | * that all bios are propagated to td->service_queue. It'd be | |
1998 | * better to walk service_queue tree directly but blkg walk is | |
1999 | * easier. | |
2000 | */ | |
2001 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) | |
2002 | tg_drain_bios(&blkg_to_tg(blkg)->service_queue); | |
2003 | ||
2004 | /* finally, transfer bios from top-level tg's into the td */ | |
2005 | tg_drain_bios(&td->service_queue); | |
2006 | ||
2007 | rcu_read_unlock(); | |
2008 | spin_unlock_irq(q->queue_lock); | |
2009 | ||
2010 | /* all bios now should be in td->service_queue, issue them */ | |
2011 | for (rw = READ; rw <= WRITE; rw++) | |
2012 | while ((bio = throtl_pop_queued(&td->service_queue.queued[rw], | |
2013 | NULL))) | |
2014 | generic_make_request(bio); | |
2015 | ||
2016 | spin_lock_irq(q->queue_lock); | |
2017 | } | |
2018 | ||
2019 | int blk_throtl_init(struct request_queue *q) | |
2020 | { | |
2021 | struct throtl_data *td; | |
2022 | int ret; | |
2023 | ||
2024 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
2025 | if (!td) | |
2026 | return -ENOMEM; | |
2027 | ||
2028 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); | |
2029 | throtl_service_queue_init(&td->service_queue); | |
2030 | ||
2031 | q->td = td; | |
2032 | td->queue = q; | |
2033 | ||
2034 | td->limit_valid[LIMIT_MAX] = true; | |
2035 | td->limit_index = LIMIT_MAX; | |
2036 | td->low_upgrade_time = jiffies; | |
2037 | td->low_downgrade_time = jiffies; | |
2038 | /* activate policy */ | |
2039 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); | |
2040 | if (ret) | |
2041 | kfree(td); | |
2042 | return ret; | |
2043 | } | |
2044 | ||
2045 | void blk_throtl_exit(struct request_queue *q) | |
2046 | { | |
2047 | BUG_ON(!q->td); | |
2048 | throtl_shutdown_wq(q); | |
2049 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); | |
2050 | kfree(q->td); | |
2051 | } | |
2052 | ||
2053 | void blk_throtl_register_queue(struct request_queue *q) | |
2054 | { | |
2055 | struct throtl_data *td; | |
2056 | ||
2057 | td = q->td; | |
2058 | BUG_ON(!td); | |
2059 | ||
2060 | if (blk_queue_nonrot(q)) | |
2061 | td->throtl_slice = DFL_THROTL_SLICE_SSD; | |
2062 | else | |
2063 | td->throtl_slice = DFL_THROTL_SLICE_HD; | |
2064 | #ifndef CONFIG_BLK_DEV_THROTTLING_LOW | |
2065 | /* if no low limit, use previous default */ | |
2066 | td->throtl_slice = DFL_THROTL_SLICE_HD; | |
2067 | #endif | |
2068 | } | |
2069 | ||
2070 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
2071 | ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page) | |
2072 | { | |
2073 | if (!q->td) | |
2074 | return -EINVAL; | |
2075 | return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice)); | |
2076 | } | |
2077 | ||
2078 | ssize_t blk_throtl_sample_time_store(struct request_queue *q, | |
2079 | const char *page, size_t count) | |
2080 | { | |
2081 | unsigned long v; | |
2082 | unsigned long t; | |
2083 | ||
2084 | if (!q->td) | |
2085 | return -EINVAL; | |
2086 | if (kstrtoul(page, 10, &v)) | |
2087 | return -EINVAL; | |
2088 | t = msecs_to_jiffies(v); | |
2089 | if (t == 0 || t > MAX_THROTL_SLICE) | |
2090 | return -EINVAL; | |
2091 | q->td->throtl_slice = t; | |
2092 | return count; | |
2093 | } | |
2094 | #endif | |
2095 | ||
2096 | static int __init throtl_init(void) | |
2097 | { | |
2098 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); | |
2099 | if (!kthrotld_workqueue) | |
2100 | panic("Failed to create kthrotld\n"); | |
2101 | ||
2102 | return blkcg_policy_register(&blkcg_policy_throtl); | |
2103 | } | |
2104 | ||
2105 | module_init(throtl_init); |