]>
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 | #define DFL_IDLE_THRESHOLD_SSD (1000L) /* 1 ms */ | |
26 | #define DFL_IDLE_THRESHOLD_HD (100L * 1000) /* 100 ms */ | |
27 | #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */ | |
28 | /* default latency target is 0, eg, guarantee IO latency by default */ | |
29 | #define DFL_LATENCY_TARGET (0) | |
30 | ||
31 | #define SKIP_LATENCY (((u64)1) << BLK_STAT_RES_SHIFT) | |
32 | ||
33 | static struct blkcg_policy blkcg_policy_throtl; | |
34 | ||
35 | /* A workqueue to queue throttle related work */ | |
36 | static struct workqueue_struct *kthrotld_workqueue; | |
37 | ||
38 | /* | |
39 | * To implement hierarchical throttling, throtl_grps form a tree and bios | |
40 | * are dispatched upwards level by level until they reach the top and get | |
41 | * issued. When dispatching bios from the children and local group at each | |
42 | * level, if the bios are dispatched into a single bio_list, there's a risk | |
43 | * of a local or child group which can queue many bios at once filling up | |
44 | * the list starving others. | |
45 | * | |
46 | * To avoid such starvation, dispatched bios are queued separately | |
47 | * according to where they came from. When they are again dispatched to | |
48 | * the parent, they're popped in round-robin order so that no single source | |
49 | * hogs the dispatch window. | |
50 | * | |
51 | * throtl_qnode is used to keep the queued bios separated by their sources. | |
52 | * Bios are queued to throtl_qnode which in turn is queued to | |
53 | * throtl_service_queue and then dispatched in round-robin order. | |
54 | * | |
55 | * It's also used to track the reference counts on blkg's. A qnode always | |
56 | * belongs to a throtl_grp and gets queued on itself or the parent, so | |
57 | * incrementing the reference of the associated throtl_grp when a qnode is | |
58 | * queued and decrementing when dequeued is enough to keep the whole blkg | |
59 | * tree pinned while bios are in flight. | |
60 | */ | |
61 | struct throtl_qnode { | |
62 | struct list_head node; /* service_queue->queued[] */ | |
63 | struct bio_list bios; /* queued bios */ | |
64 | struct throtl_grp *tg; /* tg this qnode belongs to */ | |
65 | }; | |
66 | ||
67 | struct throtl_service_queue { | |
68 | struct throtl_service_queue *parent_sq; /* the parent service_queue */ | |
69 | ||
70 | /* | |
71 | * Bios queued directly to this service_queue or dispatched from | |
72 | * children throtl_grp's. | |
73 | */ | |
74 | struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */ | |
75 | unsigned int nr_queued[2]; /* number of queued bios */ | |
76 | ||
77 | /* | |
78 | * RB tree of active children throtl_grp's, which are sorted by | |
79 | * their ->disptime. | |
80 | */ | |
81 | struct rb_root pending_tree; /* RB tree of active tgs */ | |
82 | struct rb_node *first_pending; /* first node in the tree */ | |
83 | unsigned int nr_pending; /* # queued in the tree */ | |
84 | unsigned long first_pending_disptime; /* disptime of the first tg */ | |
85 | struct timer_list pending_timer; /* fires on first_pending_disptime */ | |
86 | }; | |
87 | ||
88 | enum tg_state_flags { | |
89 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
90 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ | |
91 | }; | |
92 | ||
93 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) | |
94 | ||
95 | enum { | |
96 | LIMIT_LOW, | |
97 | LIMIT_MAX, | |
98 | LIMIT_CNT, | |
99 | }; | |
100 | ||
101 | struct throtl_grp { | |
102 | /* must be the first member */ | |
103 | struct blkg_policy_data pd; | |
104 | ||
105 | /* active throtl group service_queue member */ | |
106 | struct rb_node rb_node; | |
107 | ||
108 | /* throtl_data this group belongs to */ | |
109 | struct throtl_data *td; | |
110 | ||
111 | /* this group's service queue */ | |
112 | struct throtl_service_queue service_queue; | |
113 | ||
114 | /* | |
115 | * qnode_on_self is used when bios are directly queued to this | |
116 | * throtl_grp so that local bios compete fairly with bios | |
117 | * dispatched from children. qnode_on_parent is used when bios are | |
118 | * dispatched from this throtl_grp into its parent and will compete | |
119 | * with the sibling qnode_on_parents and the parent's | |
120 | * qnode_on_self. | |
121 | */ | |
122 | struct throtl_qnode qnode_on_self[2]; | |
123 | struct throtl_qnode qnode_on_parent[2]; | |
124 | ||
125 | /* | |
126 | * Dispatch time in jiffies. This is the estimated time when group | |
127 | * will unthrottle and is ready to dispatch more bio. It is used as | |
128 | * key to sort active groups in service tree. | |
129 | */ | |
130 | unsigned long disptime; | |
131 | ||
132 | unsigned int flags; | |
133 | ||
134 | /* are there any throtl rules between this group and td? */ | |
135 | bool has_rules[2]; | |
136 | ||
137 | /* internally used bytes per second rate limits */ | |
138 | uint64_t bps[2][LIMIT_CNT]; | |
139 | /* user configured bps limits */ | |
140 | uint64_t bps_conf[2][LIMIT_CNT]; | |
141 | ||
142 | /* internally used IOPS limits */ | |
143 | unsigned int iops[2][LIMIT_CNT]; | |
144 | /* user configured IOPS limits */ | |
145 | unsigned int iops_conf[2][LIMIT_CNT]; | |
146 | ||
147 | /* Number of bytes disptached in current slice */ | |
148 | uint64_t bytes_disp[2]; | |
149 | /* Number of bio's dispatched in current slice */ | |
150 | unsigned int io_disp[2]; | |
151 | ||
152 | unsigned long last_low_overflow_time[2]; | |
153 | ||
154 | uint64_t last_bytes_disp[2]; | |
155 | unsigned int last_io_disp[2]; | |
156 | ||
157 | unsigned long last_check_time; | |
158 | ||
159 | unsigned long latency_target; /* us */ | |
160 | /* When did we start a new slice */ | |
161 | unsigned long slice_start[2]; | |
162 | unsigned long slice_end[2]; | |
163 | ||
164 | unsigned long last_finish_time; /* ns / 1024 */ | |
165 | unsigned long checked_last_finish_time; /* ns / 1024 */ | |
166 | unsigned long avg_idletime; /* ns / 1024 */ | |
167 | unsigned long idletime_threshold; /* us */ | |
168 | ||
169 | unsigned int bio_cnt; /* total bios */ | |
170 | unsigned int bad_bio_cnt; /* bios exceeding latency threshold */ | |
171 | unsigned long bio_cnt_reset_time; | |
172 | }; | |
173 | ||
174 | /* We measure latency for request size from <= 4k to >= 1M */ | |
175 | #define LATENCY_BUCKET_SIZE 9 | |
176 | ||
177 | struct latency_bucket { | |
178 | unsigned long total_latency; /* ns / 1024 */ | |
179 | int samples; | |
180 | }; | |
181 | ||
182 | struct avg_latency_bucket { | |
183 | unsigned long latency; /* ns / 1024 */ | |
184 | bool valid; | |
185 | }; | |
186 | ||
187 | struct throtl_data | |
188 | { | |
189 | /* service tree for active throtl groups */ | |
190 | struct throtl_service_queue service_queue; | |
191 | ||
192 | struct request_queue *queue; | |
193 | ||
194 | /* Total Number of queued bios on READ and WRITE lists */ | |
195 | unsigned int nr_queued[2]; | |
196 | ||
197 | unsigned int throtl_slice; | |
198 | ||
199 | /* Work for dispatching throttled bios */ | |
200 | struct work_struct dispatch_work; | |
201 | unsigned int limit_index; | |
202 | bool limit_valid[LIMIT_CNT]; | |
203 | ||
204 | unsigned long dft_idletime_threshold; /* us */ | |
205 | ||
206 | unsigned long low_upgrade_time; | |
207 | unsigned long low_downgrade_time; | |
208 | ||
209 | unsigned int scale; | |
210 | ||
211 | struct latency_bucket tmp_buckets[LATENCY_BUCKET_SIZE]; | |
212 | struct avg_latency_bucket avg_buckets[LATENCY_BUCKET_SIZE]; | |
213 | struct latency_bucket __percpu *latency_buckets; | |
214 | unsigned long last_calculate_time; | |
215 | ||
216 | bool track_bio_latency; | |
217 | }; | |
218 | ||
219 | static void throtl_pending_timer_fn(unsigned long arg); | |
220 | ||
221 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) | |
222 | { | |
223 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
224 | } | |
225 | ||
226 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) | |
227 | { | |
228 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); | |
229 | } | |
230 | ||
231 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) | |
232 | { | |
233 | return pd_to_blkg(&tg->pd); | |
234 | } | |
235 | ||
236 | /** | |
237 | * sq_to_tg - return the throl_grp the specified service queue belongs to | |
238 | * @sq: the throtl_service_queue of interest | |
239 | * | |
240 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one | |
241 | * embedded in throtl_data, %NULL is returned. | |
242 | */ | |
243 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) | |
244 | { | |
245 | if (sq && sq->parent_sq) | |
246 | return container_of(sq, struct throtl_grp, service_queue); | |
247 | else | |
248 | return NULL; | |
249 | } | |
250 | ||
251 | /** | |
252 | * sq_to_td - return throtl_data the specified service queue belongs to | |
253 | * @sq: the throtl_service_queue of interest | |
254 | * | |
255 | * A service_queue can be embedded in either a throtl_grp or throtl_data. | |
256 | * Determine the associated throtl_data accordingly and return it. | |
257 | */ | |
258 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) | |
259 | { | |
260 | struct throtl_grp *tg = sq_to_tg(sq); | |
261 | ||
262 | if (tg) | |
263 | return tg->td; | |
264 | else | |
265 | return container_of(sq, struct throtl_data, service_queue); | |
266 | } | |
267 | ||
268 | /* | |
269 | * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to | |
270 | * make the IO dispatch more smooth. | |
271 | * Scale up: linearly scale up according to lapsed time since upgrade. For | |
272 | * every throtl_slice, the limit scales up 1/2 .low limit till the | |
273 | * limit hits .max limit | |
274 | * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit | |
275 | */ | |
276 | static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td) | |
277 | { | |
278 | /* arbitrary value to avoid too big scale */ | |
279 | if (td->scale < 4096 && time_after_eq(jiffies, | |
280 | td->low_upgrade_time + td->scale * td->throtl_slice)) | |
281 | td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice; | |
282 | ||
283 | return low + (low >> 1) * td->scale; | |
284 | } | |
285 | ||
286 | static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw) | |
287 | { | |
288 | struct blkcg_gq *blkg = tg_to_blkg(tg); | |
289 | struct throtl_data *td; | |
290 | uint64_t ret; | |
291 | ||
292 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
293 | return U64_MAX; | |
294 | ||
295 | td = tg->td; | |
296 | ret = tg->bps[rw][td->limit_index]; | |
297 | if (ret == 0 && td->limit_index == LIMIT_LOW) | |
298 | return tg->bps[rw][LIMIT_MAX]; | |
299 | ||
300 | if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] && | |
301 | tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) { | |
302 | uint64_t adjusted; | |
303 | ||
304 | adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td); | |
305 | ret = min(tg->bps[rw][LIMIT_MAX], adjusted); | |
306 | } | |
307 | return ret; | |
308 | } | |
309 | ||
310 | static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw) | |
311 | { | |
312 | struct blkcg_gq *blkg = tg_to_blkg(tg); | |
313 | struct throtl_data *td; | |
314 | unsigned int ret; | |
315 | ||
316 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
317 | return UINT_MAX; | |
318 | td = tg->td; | |
319 | ret = tg->iops[rw][td->limit_index]; | |
320 | if (ret == 0 && tg->td->limit_index == LIMIT_LOW) | |
321 | return tg->iops[rw][LIMIT_MAX]; | |
322 | ||
323 | if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] && | |
324 | tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) { | |
325 | uint64_t adjusted; | |
326 | ||
327 | adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td); | |
328 | if (adjusted > UINT_MAX) | |
329 | adjusted = UINT_MAX; | |
330 | ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted); | |
331 | } | |
332 | return ret; | |
333 | } | |
334 | ||
335 | #define request_bucket_index(sectors) \ | |
336 | clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1) | |
337 | ||
338 | /** | |
339 | * throtl_log - log debug message via blktrace | |
340 | * @sq: the service_queue being reported | |
341 | * @fmt: printf format string | |
342 | * @args: printf args | |
343 | * | |
344 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a | |
345 | * throtl_grp; otherwise, just "throtl". | |
346 | */ | |
347 | #define throtl_log(sq, fmt, args...) do { \ | |
348 | struct throtl_grp *__tg = sq_to_tg((sq)); \ | |
349 | struct throtl_data *__td = sq_to_td((sq)); \ | |
350 | \ | |
351 | (void)__td; \ | |
352 | if (likely(!blk_trace_note_message_enabled(__td->queue))) \ | |
353 | break; \ | |
354 | if ((__tg)) { \ | |
355 | char __pbuf[128]; \ | |
356 | \ | |
357 | blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \ | |
358 | blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
359 | } else { \ | |
360 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
361 | } \ | |
362 | } while (0) | |
363 | ||
364 | static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg) | |
365 | { | |
366 | INIT_LIST_HEAD(&qn->node); | |
367 | bio_list_init(&qn->bios); | |
368 | qn->tg = tg; | |
369 | } | |
370 | ||
371 | /** | |
372 | * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it | |
373 | * @bio: bio being added | |
374 | * @qn: qnode to add bio to | |
375 | * @queued: the service_queue->queued[] list @qn belongs to | |
376 | * | |
377 | * Add @bio to @qn and put @qn on @queued if it's not already on. | |
378 | * @qn->tg's reference count is bumped when @qn is activated. See the | |
379 | * comment on top of throtl_qnode definition for details. | |
380 | */ | |
381 | static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn, | |
382 | struct list_head *queued) | |
383 | { | |
384 | bio_list_add(&qn->bios, bio); | |
385 | if (list_empty(&qn->node)) { | |
386 | list_add_tail(&qn->node, queued); | |
387 | blkg_get(tg_to_blkg(qn->tg)); | |
388 | } | |
389 | } | |
390 | ||
391 | /** | |
392 | * throtl_peek_queued - peek the first bio on a qnode list | |
393 | * @queued: the qnode list to peek | |
394 | */ | |
395 | static struct bio *throtl_peek_queued(struct list_head *queued) | |
396 | { | |
397 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
398 | struct bio *bio; | |
399 | ||
400 | if (list_empty(queued)) | |
401 | return NULL; | |
402 | ||
403 | bio = bio_list_peek(&qn->bios); | |
404 | WARN_ON_ONCE(!bio); | |
405 | return bio; | |
406 | } | |
407 | ||
408 | /** | |
409 | * throtl_pop_queued - pop the first bio form a qnode list | |
410 | * @queued: the qnode list to pop a bio from | |
411 | * @tg_to_put: optional out argument for throtl_grp to put | |
412 | * | |
413 | * Pop the first bio from the qnode list @queued. After popping, the first | |
414 | * qnode is removed from @queued if empty or moved to the end of @queued so | |
415 | * that the popping order is round-robin. | |
416 | * | |
417 | * When the first qnode is removed, its associated throtl_grp should be put | |
418 | * too. If @tg_to_put is NULL, this function automatically puts it; | |
419 | * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is | |
420 | * responsible for putting it. | |
421 | */ | |
422 | static struct bio *throtl_pop_queued(struct list_head *queued, | |
423 | struct throtl_grp **tg_to_put) | |
424 | { | |
425 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
426 | struct bio *bio; | |
427 | ||
428 | if (list_empty(queued)) | |
429 | return NULL; | |
430 | ||
431 | bio = bio_list_pop(&qn->bios); | |
432 | WARN_ON_ONCE(!bio); | |
433 | ||
434 | if (bio_list_empty(&qn->bios)) { | |
435 | list_del_init(&qn->node); | |
436 | if (tg_to_put) | |
437 | *tg_to_put = qn->tg; | |
438 | else | |
439 | blkg_put(tg_to_blkg(qn->tg)); | |
440 | } else { | |
441 | list_move_tail(&qn->node, queued); | |
442 | } | |
443 | ||
444 | return bio; | |
445 | } | |
446 | ||
447 | /* init a service_queue, assumes the caller zeroed it */ | |
448 | static void throtl_service_queue_init(struct throtl_service_queue *sq) | |
449 | { | |
450 | INIT_LIST_HEAD(&sq->queued[0]); | |
451 | INIT_LIST_HEAD(&sq->queued[1]); | |
452 | sq->pending_tree = RB_ROOT; | |
453 | setup_timer(&sq->pending_timer, throtl_pending_timer_fn, | |
454 | (unsigned long)sq); | |
455 | } | |
456 | ||
457 | static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node) | |
458 | { | |
459 | struct throtl_grp *tg; | |
460 | int rw; | |
461 | ||
462 | tg = kzalloc_node(sizeof(*tg), gfp, node); | |
463 | if (!tg) | |
464 | return NULL; | |
465 | ||
466 | throtl_service_queue_init(&tg->service_queue); | |
467 | ||
468 | for (rw = READ; rw <= WRITE; rw++) { | |
469 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); | |
470 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); | |
471 | } | |
472 | ||
473 | RB_CLEAR_NODE(&tg->rb_node); | |
474 | tg->bps[READ][LIMIT_MAX] = U64_MAX; | |
475 | tg->bps[WRITE][LIMIT_MAX] = U64_MAX; | |
476 | tg->iops[READ][LIMIT_MAX] = UINT_MAX; | |
477 | tg->iops[WRITE][LIMIT_MAX] = UINT_MAX; | |
478 | tg->bps_conf[READ][LIMIT_MAX] = U64_MAX; | |
479 | tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX; | |
480 | tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX; | |
481 | tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX; | |
482 | /* LIMIT_LOW will have default value 0 */ | |
483 | ||
484 | tg->latency_target = DFL_LATENCY_TARGET; | |
485 | ||
486 | return &tg->pd; | |
487 | } | |
488 | ||
489 | static void throtl_pd_init(struct blkg_policy_data *pd) | |
490 | { | |
491 | struct throtl_grp *tg = pd_to_tg(pd); | |
492 | struct blkcg_gq *blkg = tg_to_blkg(tg); | |
493 | struct throtl_data *td = blkg->q->td; | |
494 | struct throtl_service_queue *sq = &tg->service_queue; | |
495 | ||
496 | /* | |
497 | * If on the default hierarchy, we switch to properly hierarchical | |
498 | * behavior where limits on a given throtl_grp are applied to the | |
499 | * whole subtree rather than just the group itself. e.g. If 16M | |
500 | * read_bps limit is set on the root group, the whole system can't | |
501 | * exceed 16M for the device. | |
502 | * | |
503 | * If not on the default hierarchy, the broken flat hierarchy | |
504 | * behavior is retained where all throtl_grps are treated as if | |
505 | * they're all separate root groups right below throtl_data. | |
506 | * Limits of a group don't interact with limits of other groups | |
507 | * regardless of the position of the group in the hierarchy. | |
508 | */ | |
509 | sq->parent_sq = &td->service_queue; | |
510 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent) | |
511 | sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue; | |
512 | tg->td = td; | |
513 | ||
514 | tg->idletime_threshold = td->dft_idletime_threshold; | |
515 | } | |
516 | ||
517 | /* | |
518 | * Set has_rules[] if @tg or any of its parents have limits configured. | |
519 | * This doesn't require walking up to the top of the hierarchy as the | |
520 | * parent's has_rules[] is guaranteed to be correct. | |
521 | */ | |
522 | static void tg_update_has_rules(struct throtl_grp *tg) | |
523 | { | |
524 | struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); | |
525 | struct throtl_data *td = tg->td; | |
526 | int rw; | |
527 | ||
528 | for (rw = READ; rw <= WRITE; rw++) | |
529 | tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) || | |
530 | (td->limit_valid[td->limit_index] && | |
531 | (tg_bps_limit(tg, rw) != U64_MAX || | |
532 | tg_iops_limit(tg, rw) != UINT_MAX)); | |
533 | } | |
534 | ||
535 | static void throtl_pd_online(struct blkg_policy_data *pd) | |
536 | { | |
537 | struct throtl_grp *tg = pd_to_tg(pd); | |
538 | /* | |
539 | * We don't want new groups to escape the limits of its ancestors. | |
540 | * Update has_rules[] after a new group is brought online. | |
541 | */ | |
542 | tg_update_has_rules(tg); | |
543 | } | |
544 | ||
545 | static void blk_throtl_update_limit_valid(struct throtl_data *td) | |
546 | { | |
547 | struct cgroup_subsys_state *pos_css; | |
548 | struct blkcg_gq *blkg; | |
549 | bool low_valid = false; | |
550 | ||
551 | rcu_read_lock(); | |
552 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
553 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
554 | ||
555 | if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] || | |
556 | tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) | |
557 | low_valid = true; | |
558 | } | |
559 | rcu_read_unlock(); | |
560 | ||
561 | td->limit_valid[LIMIT_LOW] = low_valid; | |
562 | } | |
563 | ||
564 | static void throtl_upgrade_state(struct throtl_data *td); | |
565 | static void throtl_pd_offline(struct blkg_policy_data *pd) | |
566 | { | |
567 | struct throtl_grp *tg = pd_to_tg(pd); | |
568 | ||
569 | tg->bps[READ][LIMIT_LOW] = 0; | |
570 | tg->bps[WRITE][LIMIT_LOW] = 0; | |
571 | tg->iops[READ][LIMIT_LOW] = 0; | |
572 | tg->iops[WRITE][LIMIT_LOW] = 0; | |
573 | ||
574 | blk_throtl_update_limit_valid(tg->td); | |
575 | ||
576 | if (!tg->td->limit_valid[tg->td->limit_index]) | |
577 | throtl_upgrade_state(tg->td); | |
578 | } | |
579 | ||
580 | static void throtl_pd_free(struct blkg_policy_data *pd) | |
581 | { | |
582 | struct throtl_grp *tg = pd_to_tg(pd); | |
583 | ||
584 | del_timer_sync(&tg->service_queue.pending_timer); | |
585 | kfree(tg); | |
586 | } | |
587 | ||
588 | static struct throtl_grp * | |
589 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
590 | { | |
591 | /* Service tree is empty */ | |
592 | if (!parent_sq->nr_pending) | |
593 | return NULL; | |
594 | ||
595 | if (!parent_sq->first_pending) | |
596 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
597 | ||
598 | if (parent_sq->first_pending) | |
599 | return rb_entry_tg(parent_sq->first_pending); | |
600 | ||
601 | return NULL; | |
602 | } | |
603 | ||
604 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
605 | { | |
606 | rb_erase(n, root); | |
607 | RB_CLEAR_NODE(n); | |
608 | } | |
609 | ||
610 | static void throtl_rb_erase(struct rb_node *n, | |
611 | struct throtl_service_queue *parent_sq) | |
612 | { | |
613 | if (parent_sq->first_pending == n) | |
614 | parent_sq->first_pending = NULL; | |
615 | rb_erase_init(n, &parent_sq->pending_tree); | |
616 | --parent_sq->nr_pending; | |
617 | } | |
618 | ||
619 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) | |
620 | { | |
621 | struct throtl_grp *tg; | |
622 | ||
623 | tg = throtl_rb_first(parent_sq); | |
624 | if (!tg) | |
625 | return; | |
626 | ||
627 | parent_sq->first_pending_disptime = tg->disptime; | |
628 | } | |
629 | ||
630 | static void tg_service_queue_add(struct throtl_grp *tg) | |
631 | { | |
632 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; | |
633 | struct rb_node **node = &parent_sq->pending_tree.rb_node; | |
634 | struct rb_node *parent = NULL; | |
635 | struct throtl_grp *__tg; | |
636 | unsigned long key = tg->disptime; | |
637 | int left = 1; | |
638 | ||
639 | while (*node != NULL) { | |
640 | parent = *node; | |
641 | __tg = rb_entry_tg(parent); | |
642 | ||
643 | if (time_before(key, __tg->disptime)) | |
644 | node = &parent->rb_left; | |
645 | else { | |
646 | node = &parent->rb_right; | |
647 | left = 0; | |
648 | } | |
649 | } | |
650 | ||
651 | if (left) | |
652 | parent_sq->first_pending = &tg->rb_node; | |
653 | ||
654 | rb_link_node(&tg->rb_node, parent, node); | |
655 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); | |
656 | } | |
657 | ||
658 | static void __throtl_enqueue_tg(struct throtl_grp *tg) | |
659 | { | |
660 | tg_service_queue_add(tg); | |
661 | tg->flags |= THROTL_TG_PENDING; | |
662 | tg->service_queue.parent_sq->nr_pending++; | |
663 | } | |
664 | ||
665 | static void throtl_enqueue_tg(struct throtl_grp *tg) | |
666 | { | |
667 | if (!(tg->flags & THROTL_TG_PENDING)) | |
668 | __throtl_enqueue_tg(tg); | |
669 | } | |
670 | ||
671 | static void __throtl_dequeue_tg(struct throtl_grp *tg) | |
672 | { | |
673 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); | |
674 | tg->flags &= ~THROTL_TG_PENDING; | |
675 | } | |
676 | ||
677 | static void throtl_dequeue_tg(struct throtl_grp *tg) | |
678 | { | |
679 | if (tg->flags & THROTL_TG_PENDING) | |
680 | __throtl_dequeue_tg(tg); | |
681 | } | |
682 | ||
683 | /* Call with queue lock held */ | |
684 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, | |
685 | unsigned long expires) | |
686 | { | |
687 | unsigned long max_expire = jiffies + 8 * sq_to_tg(sq)->td->throtl_slice; | |
688 | ||
689 | /* | |
690 | * Since we are adjusting the throttle limit dynamically, the sleep | |
691 | * time calculated according to previous limit might be invalid. It's | |
692 | * possible the cgroup sleep time is very long and no other cgroups | |
693 | * have IO running so notify the limit changes. Make sure the cgroup | |
694 | * doesn't sleep too long to avoid the missed notification. | |
695 | */ | |
696 | if (time_after(expires, max_expire)) | |
697 | expires = max_expire; | |
698 | mod_timer(&sq->pending_timer, expires); | |
699 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
700 | expires - jiffies, jiffies); | |
701 | } | |
702 | ||
703 | /** | |
704 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle | |
705 | * @sq: the service_queue to schedule dispatch for | |
706 | * @force: force scheduling | |
707 | * | |
708 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the | |
709 | * dispatch time of the first pending child. Returns %true if either timer | |
710 | * is armed or there's no pending child left. %false if the current | |
711 | * dispatch window is still open and the caller should continue | |
712 | * dispatching. | |
713 | * | |
714 | * If @force is %true, the dispatch timer is always scheduled and this | |
715 | * function is guaranteed to return %true. This is to be used when the | |
716 | * caller can't dispatch itself and needs to invoke pending_timer | |
717 | * unconditionally. Note that forced scheduling is likely to induce short | |
718 | * delay before dispatch starts even if @sq->first_pending_disptime is not | |
719 | * in the future and thus shouldn't be used in hot paths. | |
720 | */ | |
721 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
722 | bool force) | |
723 | { | |
724 | /* any pending children left? */ | |
725 | if (!sq->nr_pending) | |
726 | return true; | |
727 | ||
728 | update_min_dispatch_time(sq); | |
729 | ||
730 | /* is the next dispatch time in the future? */ | |
731 | if (force || time_after(sq->first_pending_disptime, jiffies)) { | |
732 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); | |
733 | return true; | |
734 | } | |
735 | ||
736 | /* tell the caller to continue dispatching */ | |
737 | return false; | |
738 | } | |
739 | ||
740 | static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, | |
741 | bool rw, unsigned long start) | |
742 | { | |
743 | tg->bytes_disp[rw] = 0; | |
744 | tg->io_disp[rw] = 0; | |
745 | ||
746 | /* | |
747 | * Previous slice has expired. We must have trimmed it after last | |
748 | * bio dispatch. That means since start of last slice, we never used | |
749 | * that bandwidth. Do try to make use of that bandwidth while giving | |
750 | * credit. | |
751 | */ | |
752 | if (time_after_eq(start, tg->slice_start[rw])) | |
753 | tg->slice_start[rw] = start; | |
754 | ||
755 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; | |
756 | throtl_log(&tg->service_queue, | |
757 | "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", | |
758 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
759 | tg->slice_end[rw], jiffies); | |
760 | } | |
761 | ||
762 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) | |
763 | { | |
764 | tg->bytes_disp[rw] = 0; | |
765 | tg->io_disp[rw] = 0; | |
766 | tg->slice_start[rw] = jiffies; | |
767 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; | |
768 | throtl_log(&tg->service_queue, | |
769 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
770 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
771 | tg->slice_end[rw], jiffies); | |
772 | } | |
773 | ||
774 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, | |
775 | unsigned long jiffy_end) | |
776 | { | |
777 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); | |
778 | } | |
779 | ||
780 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, | |
781 | unsigned long jiffy_end) | |
782 | { | |
783 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); | |
784 | throtl_log(&tg->service_queue, | |
785 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
786 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
787 | tg->slice_end[rw], jiffies); | |
788 | } | |
789 | ||
790 | /* Determine if previously allocated or extended slice is complete or not */ | |
791 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) | |
792 | { | |
793 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
794 | return false; | |
795 | ||
796 | return 1; | |
797 | } | |
798 | ||
799 | /* Trim the used slices and adjust slice start accordingly */ | |
800 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) | |
801 | { | |
802 | unsigned long nr_slices, time_elapsed, io_trim; | |
803 | u64 bytes_trim, tmp; | |
804 | ||
805 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
806 | ||
807 | /* | |
808 | * If bps are unlimited (-1), then time slice don't get | |
809 | * renewed. Don't try to trim the slice if slice is used. A new | |
810 | * slice will start when appropriate. | |
811 | */ | |
812 | if (throtl_slice_used(tg, rw)) | |
813 | return; | |
814 | ||
815 | /* | |
816 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
817 | * that initially cgroup limit was very low resulting in high | |
818 | * slice_end, but later limit was bumped up and bio was dispached | |
819 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
820 | * is bad because it does not allow new slice to start. | |
821 | */ | |
822 | ||
823 | throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice); | |
824 | ||
825 | time_elapsed = jiffies - tg->slice_start[rw]; | |
826 | ||
827 | nr_slices = time_elapsed / tg->td->throtl_slice; | |
828 | ||
829 | if (!nr_slices) | |
830 | return; | |
831 | tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices; | |
832 | do_div(tmp, HZ); | |
833 | bytes_trim = tmp; | |
834 | ||
835 | io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) / | |
836 | HZ; | |
837 | ||
838 | if (!bytes_trim && !io_trim) | |
839 | return; | |
840 | ||
841 | if (tg->bytes_disp[rw] >= bytes_trim) | |
842 | tg->bytes_disp[rw] -= bytes_trim; | |
843 | else | |
844 | tg->bytes_disp[rw] = 0; | |
845 | ||
846 | if (tg->io_disp[rw] >= io_trim) | |
847 | tg->io_disp[rw] -= io_trim; | |
848 | else | |
849 | tg->io_disp[rw] = 0; | |
850 | ||
851 | tg->slice_start[rw] += nr_slices * tg->td->throtl_slice; | |
852 | ||
853 | throtl_log(&tg->service_queue, | |
854 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
855 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
856 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
857 | } | |
858 | ||
859 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, | |
860 | unsigned long *wait) | |
861 | { | |
862 | bool rw = bio_data_dir(bio); | |
863 | unsigned int io_allowed; | |
864 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; | |
865 | u64 tmp; | |
866 | ||
867 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
868 | ||
869 | /* Slice has just started. Consider one slice interval */ | |
870 | if (!jiffy_elapsed) | |
871 | jiffy_elapsed_rnd = tg->td->throtl_slice; | |
872 | ||
873 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); | |
874 | ||
875 | /* | |
876 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
877 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
878 | * will allow dispatch after 1 second and after that slice should | |
879 | * have been trimmed. | |
880 | */ | |
881 | ||
882 | tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd; | |
883 | do_div(tmp, HZ); | |
884 | ||
885 | if (tmp > UINT_MAX) | |
886 | io_allowed = UINT_MAX; | |
887 | else | |
888 | io_allowed = tmp; | |
889 | ||
890 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
891 | if (wait) | |
892 | *wait = 0; | |
893 | return true; | |
894 | } | |
895 | ||
896 | /* Calc approx time to dispatch */ | |
897 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ) / tg_iops_limit(tg, rw) + 1; | |
898 | ||
899 | if (jiffy_wait > jiffy_elapsed) | |
900 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
901 | else | |
902 | jiffy_wait = 1; | |
903 | ||
904 | if (wait) | |
905 | *wait = jiffy_wait; | |
906 | return 0; | |
907 | } | |
908 | ||
909 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, | |
910 | unsigned long *wait) | |
911 | { | |
912 | bool rw = bio_data_dir(bio); | |
913 | u64 bytes_allowed, extra_bytes, tmp; | |
914 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; | |
915 | ||
916 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
917 | ||
918 | /* Slice has just started. Consider one slice interval */ | |
919 | if (!jiffy_elapsed) | |
920 | jiffy_elapsed_rnd = tg->td->throtl_slice; | |
921 | ||
922 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); | |
923 | ||
924 | tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd; | |
925 | do_div(tmp, HZ); | |
926 | bytes_allowed = tmp; | |
927 | ||
928 | if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) { | |
929 | if (wait) | |
930 | *wait = 0; | |
931 | return true; | |
932 | } | |
933 | ||
934 | /* Calc approx time to dispatch */ | |
935 | extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed; | |
936 | jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw)); | |
937 | ||
938 | if (!jiffy_wait) | |
939 | jiffy_wait = 1; | |
940 | ||
941 | /* | |
942 | * This wait time is without taking into consideration the rounding | |
943 | * up we did. Add that time also. | |
944 | */ | |
945 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
946 | if (wait) | |
947 | *wait = jiffy_wait; | |
948 | return 0; | |
949 | } | |
950 | ||
951 | /* | |
952 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
953 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
954 | */ | |
955 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, | |
956 | unsigned long *wait) | |
957 | { | |
958 | bool rw = bio_data_dir(bio); | |
959 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
960 | ||
961 | /* | |
962 | * Currently whole state machine of group depends on first bio | |
963 | * queued in the group bio list. So one should not be calling | |
964 | * this function with a different bio if there are other bios | |
965 | * queued. | |
966 | */ | |
967 | BUG_ON(tg->service_queue.nr_queued[rw] && | |
968 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); | |
969 | ||
970 | /* If tg->bps = -1, then BW is unlimited */ | |
971 | if (tg_bps_limit(tg, rw) == U64_MAX && | |
972 | tg_iops_limit(tg, rw) == UINT_MAX) { | |
973 | if (wait) | |
974 | *wait = 0; | |
975 | return true; | |
976 | } | |
977 | ||
978 | /* | |
979 | * If previous slice expired, start a new one otherwise renew/extend | |
980 | * existing slice to make sure it is at least throtl_slice interval | |
981 | * long since now. New slice is started only for empty throttle group. | |
982 | * If there is queued bio, that means there should be an active | |
983 | * slice and it should be extended instead. | |
984 | */ | |
985 | if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw])) | |
986 | throtl_start_new_slice(tg, rw); | |
987 | else { | |
988 | if (time_before(tg->slice_end[rw], | |
989 | jiffies + tg->td->throtl_slice)) | |
990 | throtl_extend_slice(tg, rw, | |
991 | jiffies + tg->td->throtl_slice); | |
992 | } | |
993 | ||
994 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && | |
995 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
996 | if (wait) | |
997 | *wait = 0; | |
998 | return 1; | |
999 | } | |
1000 | ||
1001 | max_wait = max(bps_wait, iops_wait); | |
1002 | ||
1003 | if (wait) | |
1004 | *wait = max_wait; | |
1005 | ||
1006 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
1007 | throtl_extend_slice(tg, rw, jiffies + max_wait); | |
1008 | ||
1009 | return 0; | |
1010 | } | |
1011 | ||
1012 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
1013 | { | |
1014 | bool rw = bio_data_dir(bio); | |
1015 | ||
1016 | /* Charge the bio to the group */ | |
1017 | tg->bytes_disp[rw] += bio->bi_iter.bi_size; | |
1018 | tg->io_disp[rw]++; | |
1019 | tg->last_bytes_disp[rw] += bio->bi_iter.bi_size; | |
1020 | tg->last_io_disp[rw]++; | |
1021 | ||
1022 | /* | |
1023 | * BIO_THROTTLED is used to prevent the same bio to be throttled | |
1024 | * more than once as a throttled bio will go through blk-throtl the | |
1025 | * second time when it eventually gets issued. Set it when a bio | |
1026 | * is being charged to a tg. | |
1027 | */ | |
1028 | if (!bio_flagged(bio, BIO_THROTTLED)) | |
1029 | bio_set_flag(bio, BIO_THROTTLED); | |
1030 | } | |
1031 | ||
1032 | /** | |
1033 | * throtl_add_bio_tg - add a bio to the specified throtl_grp | |
1034 | * @bio: bio to add | |
1035 | * @qn: qnode to use | |
1036 | * @tg: the target throtl_grp | |
1037 | * | |
1038 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, | |
1039 | * tg->qnode_on_self[] is used. | |
1040 | */ | |
1041 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, | |
1042 | struct throtl_grp *tg) | |
1043 | { | |
1044 | struct throtl_service_queue *sq = &tg->service_queue; | |
1045 | bool rw = bio_data_dir(bio); | |
1046 | ||
1047 | if (!qn) | |
1048 | qn = &tg->qnode_on_self[rw]; | |
1049 | ||
1050 | /* | |
1051 | * If @tg doesn't currently have any bios queued in the same | |
1052 | * direction, queueing @bio can change when @tg should be | |
1053 | * dispatched. Mark that @tg was empty. This is automatically | |
1054 | * cleaered on the next tg_update_disptime(). | |
1055 | */ | |
1056 | if (!sq->nr_queued[rw]) | |
1057 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
1058 | ||
1059 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); | |
1060 | ||
1061 | sq->nr_queued[rw]++; | |
1062 | throtl_enqueue_tg(tg); | |
1063 | } | |
1064 | ||
1065 | static void tg_update_disptime(struct throtl_grp *tg) | |
1066 | { | |
1067 | struct throtl_service_queue *sq = &tg->service_queue; | |
1068 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
1069 | struct bio *bio; | |
1070 | ||
1071 | bio = throtl_peek_queued(&sq->queued[READ]); | |
1072 | if (bio) | |
1073 | tg_may_dispatch(tg, bio, &read_wait); | |
1074 | ||
1075 | bio = throtl_peek_queued(&sq->queued[WRITE]); | |
1076 | if (bio) | |
1077 | tg_may_dispatch(tg, bio, &write_wait); | |
1078 | ||
1079 | min_wait = min(read_wait, write_wait); | |
1080 | disptime = jiffies + min_wait; | |
1081 | ||
1082 | /* Update dispatch time */ | |
1083 | throtl_dequeue_tg(tg); | |
1084 | tg->disptime = disptime; | |
1085 | throtl_enqueue_tg(tg); | |
1086 | ||
1087 | /* see throtl_add_bio_tg() */ | |
1088 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
1089 | } | |
1090 | ||
1091 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, | |
1092 | struct throtl_grp *parent_tg, bool rw) | |
1093 | { | |
1094 | if (throtl_slice_used(parent_tg, rw)) { | |
1095 | throtl_start_new_slice_with_credit(parent_tg, rw, | |
1096 | child_tg->slice_start[rw]); | |
1097 | } | |
1098 | ||
1099 | } | |
1100 | ||
1101 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) | |
1102 | { | |
1103 | struct throtl_service_queue *sq = &tg->service_queue; | |
1104 | struct throtl_service_queue *parent_sq = sq->parent_sq; | |
1105 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
1106 | struct throtl_grp *tg_to_put = NULL; | |
1107 | struct bio *bio; | |
1108 | ||
1109 | /* | |
1110 | * @bio is being transferred from @tg to @parent_sq. Popping a bio | |
1111 | * from @tg may put its reference and @parent_sq might end up | |
1112 | * getting released prematurely. Remember the tg to put and put it | |
1113 | * after @bio is transferred to @parent_sq. | |
1114 | */ | |
1115 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); | |
1116 | sq->nr_queued[rw]--; | |
1117 | ||
1118 | throtl_charge_bio(tg, bio); | |
1119 | ||
1120 | /* | |
1121 | * If our parent is another tg, we just need to transfer @bio to | |
1122 | * the parent using throtl_add_bio_tg(). If our parent is | |
1123 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
1124 | * bio_lists[] and decrease total number queued. The caller is | |
1125 | * responsible for issuing these bios. | |
1126 | */ | |
1127 | if (parent_tg) { | |
1128 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); | |
1129 | start_parent_slice_with_credit(tg, parent_tg, rw); | |
1130 | } else { | |
1131 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], | |
1132 | &parent_sq->queued[rw]); | |
1133 | BUG_ON(tg->td->nr_queued[rw] <= 0); | |
1134 | tg->td->nr_queued[rw]--; | |
1135 | } | |
1136 | ||
1137 | throtl_trim_slice(tg, rw); | |
1138 | ||
1139 | if (tg_to_put) | |
1140 | blkg_put(tg_to_blkg(tg_to_put)); | |
1141 | } | |
1142 | ||
1143 | static int throtl_dispatch_tg(struct throtl_grp *tg) | |
1144 | { | |
1145 | struct throtl_service_queue *sq = &tg->service_queue; | |
1146 | unsigned int nr_reads = 0, nr_writes = 0; | |
1147 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
1148 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; | |
1149 | struct bio *bio; | |
1150 | ||
1151 | /* Try to dispatch 75% READS and 25% WRITES */ | |
1152 | ||
1153 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && | |
1154 | tg_may_dispatch(tg, bio, NULL)) { | |
1155 | ||
1156 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); | |
1157 | nr_reads++; | |
1158 | ||
1159 | if (nr_reads >= max_nr_reads) | |
1160 | break; | |
1161 | } | |
1162 | ||
1163 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && | |
1164 | tg_may_dispatch(tg, bio, NULL)) { | |
1165 | ||
1166 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); | |
1167 | nr_writes++; | |
1168 | ||
1169 | if (nr_writes >= max_nr_writes) | |
1170 | break; | |
1171 | } | |
1172 | ||
1173 | return nr_reads + nr_writes; | |
1174 | } | |
1175 | ||
1176 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) | |
1177 | { | |
1178 | unsigned int nr_disp = 0; | |
1179 | ||
1180 | while (1) { | |
1181 | struct throtl_grp *tg = throtl_rb_first(parent_sq); | |
1182 | struct throtl_service_queue *sq = &tg->service_queue; | |
1183 | ||
1184 | if (!tg) | |
1185 | break; | |
1186 | ||
1187 | if (time_before(jiffies, tg->disptime)) | |
1188 | break; | |
1189 | ||
1190 | throtl_dequeue_tg(tg); | |
1191 | ||
1192 | nr_disp += throtl_dispatch_tg(tg); | |
1193 | ||
1194 | if (sq->nr_queued[0] || sq->nr_queued[1]) | |
1195 | tg_update_disptime(tg); | |
1196 | ||
1197 | if (nr_disp >= throtl_quantum) | |
1198 | break; | |
1199 | } | |
1200 | ||
1201 | return nr_disp; | |
1202 | } | |
1203 | ||
1204 | static bool throtl_can_upgrade(struct throtl_data *td, | |
1205 | struct throtl_grp *this_tg); | |
1206 | /** | |
1207 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
1208 | * @arg: the throtl_service_queue being serviced | |
1209 | * | |
1210 | * This timer is armed when a child throtl_grp with active bio's become | |
1211 | * pending and queued on the service_queue's pending_tree and expires when | |
1212 | * the first child throtl_grp should be dispatched. This function | |
1213 | * dispatches bio's from the children throtl_grps to the parent | |
1214 | * service_queue. | |
1215 | * | |
1216 | * If the parent's parent is another throtl_grp, dispatching is propagated | |
1217 | * by either arming its pending_timer or repeating dispatch directly. If | |
1218 | * the top-level service_tree is reached, throtl_data->dispatch_work is | |
1219 | * kicked so that the ready bio's are issued. | |
1220 | */ | |
1221 | static void throtl_pending_timer_fn(unsigned long arg) | |
1222 | { | |
1223 | struct throtl_service_queue *sq = (void *)arg; | |
1224 | struct throtl_grp *tg = sq_to_tg(sq); | |
1225 | struct throtl_data *td = sq_to_td(sq); | |
1226 | struct request_queue *q = td->queue; | |
1227 | struct throtl_service_queue *parent_sq; | |
1228 | bool dispatched; | |
1229 | int ret; | |
1230 | ||
1231 | spin_lock_irq(q->queue_lock); | |
1232 | if (throtl_can_upgrade(td, NULL)) | |
1233 | throtl_upgrade_state(td); | |
1234 | ||
1235 | again: | |
1236 | parent_sq = sq->parent_sq; | |
1237 | dispatched = false; | |
1238 | ||
1239 | while (true) { | |
1240 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
1241 | sq->nr_queued[READ] + sq->nr_queued[WRITE], | |
1242 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
1243 | ||
1244 | ret = throtl_select_dispatch(sq); | |
1245 | if (ret) { | |
1246 | throtl_log(sq, "bios disp=%u", ret); | |
1247 | dispatched = true; | |
1248 | } | |
1249 | ||
1250 | if (throtl_schedule_next_dispatch(sq, false)) | |
1251 | break; | |
1252 | ||
1253 | /* this dispatch windows is still open, relax and repeat */ | |
1254 | spin_unlock_irq(q->queue_lock); | |
1255 | cpu_relax(); | |
1256 | spin_lock_irq(q->queue_lock); | |
1257 | } | |
1258 | ||
1259 | if (!dispatched) | |
1260 | goto out_unlock; | |
1261 | ||
1262 | if (parent_sq) { | |
1263 | /* @parent_sq is another throl_grp, propagate dispatch */ | |
1264 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1265 | tg_update_disptime(tg); | |
1266 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { | |
1267 | /* window is already open, repeat dispatching */ | |
1268 | sq = parent_sq; | |
1269 | tg = sq_to_tg(sq); | |
1270 | goto again; | |
1271 | } | |
1272 | } | |
1273 | } else { | |
1274 | /* reached the top-level, queue issueing */ | |
1275 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1276 | } | |
1277 | out_unlock: | |
1278 | spin_unlock_irq(q->queue_lock); | |
1279 | } | |
1280 | ||
1281 | /** | |
1282 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
1283 | * @work: work item being executed | |
1284 | * | |
1285 | * This function is queued for execution when bio's reach the bio_lists[] | |
1286 | * of throtl_data->service_queue. Those bio's are ready and issued by this | |
1287 | * function. | |
1288 | */ | |
1289 | static void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
1290 | { | |
1291 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1292 | dispatch_work); | |
1293 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1294 | struct request_queue *q = td->queue; | |
1295 | struct bio_list bio_list_on_stack; | |
1296 | struct bio *bio; | |
1297 | struct blk_plug plug; | |
1298 | int rw; | |
1299 | ||
1300 | bio_list_init(&bio_list_on_stack); | |
1301 | ||
1302 | spin_lock_irq(q->queue_lock); | |
1303 | for (rw = READ; rw <= WRITE; rw++) | |
1304 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) | |
1305 | bio_list_add(&bio_list_on_stack, bio); | |
1306 | spin_unlock_irq(q->queue_lock); | |
1307 | ||
1308 | if (!bio_list_empty(&bio_list_on_stack)) { | |
1309 | blk_start_plug(&plug); | |
1310 | while((bio = bio_list_pop(&bio_list_on_stack))) | |
1311 | generic_make_request(bio); | |
1312 | blk_finish_plug(&plug); | |
1313 | } | |
1314 | } | |
1315 | ||
1316 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, | |
1317 | int off) | |
1318 | { | |
1319 | struct throtl_grp *tg = pd_to_tg(pd); | |
1320 | u64 v = *(u64 *)((void *)tg + off); | |
1321 | ||
1322 | if (v == U64_MAX) | |
1323 | return 0; | |
1324 | return __blkg_prfill_u64(sf, pd, v); | |
1325 | } | |
1326 | ||
1327 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, | |
1328 | int off) | |
1329 | { | |
1330 | struct throtl_grp *tg = pd_to_tg(pd); | |
1331 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
1332 | ||
1333 | if (v == UINT_MAX) | |
1334 | return 0; | |
1335 | return __blkg_prfill_u64(sf, pd, v); | |
1336 | } | |
1337 | ||
1338 | static int tg_print_conf_u64(struct seq_file *sf, void *v) | |
1339 | { | |
1340 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64, | |
1341 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
1342 | return 0; | |
1343 | } | |
1344 | ||
1345 | static int tg_print_conf_uint(struct seq_file *sf, void *v) | |
1346 | { | |
1347 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint, | |
1348 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
1349 | return 0; | |
1350 | } | |
1351 | ||
1352 | static void tg_conf_updated(struct throtl_grp *tg) | |
1353 | { | |
1354 | struct throtl_service_queue *sq = &tg->service_queue; | |
1355 | struct cgroup_subsys_state *pos_css; | |
1356 | struct blkcg_gq *blkg; | |
1357 | ||
1358 | throtl_log(&tg->service_queue, | |
1359 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
1360 | tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE), | |
1361 | tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE)); | |
1362 | ||
1363 | /* | |
1364 | * Update has_rules[] flags for the updated tg's subtree. A tg is | |
1365 | * considered to have rules if either the tg itself or any of its | |
1366 | * ancestors has rules. This identifies groups without any | |
1367 | * restrictions in the whole hierarchy and allows them to bypass | |
1368 | * blk-throttle. | |
1369 | */ | |
1370 | blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg)) | |
1371 | tg_update_has_rules(blkg_to_tg(blkg)); | |
1372 | ||
1373 | /* | |
1374 | * We're already holding queue_lock and know @tg is valid. Let's | |
1375 | * apply the new config directly. | |
1376 | * | |
1377 | * Restart the slices for both READ and WRITES. It might happen | |
1378 | * that a group's limit are dropped suddenly and we don't want to | |
1379 | * account recently dispatched IO with new low rate. | |
1380 | */ | |
1381 | throtl_start_new_slice(tg, 0); | |
1382 | throtl_start_new_slice(tg, 1); | |
1383 | ||
1384 | if (tg->flags & THROTL_TG_PENDING) { | |
1385 | tg_update_disptime(tg); | |
1386 | throtl_schedule_next_dispatch(sq->parent_sq, true); | |
1387 | } | |
1388 | } | |
1389 | ||
1390 | static ssize_t tg_set_conf(struct kernfs_open_file *of, | |
1391 | char *buf, size_t nbytes, loff_t off, bool is_u64) | |
1392 | { | |
1393 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1394 | struct blkg_conf_ctx ctx; | |
1395 | struct throtl_grp *tg; | |
1396 | int ret; | |
1397 | u64 v; | |
1398 | ||
1399 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1400 | if (ret) | |
1401 | return ret; | |
1402 | ||
1403 | ret = -EINVAL; | |
1404 | if (sscanf(ctx.body, "%llu", &v) != 1) | |
1405 | goto out_finish; | |
1406 | if (!v) | |
1407 | v = U64_MAX; | |
1408 | ||
1409 | tg = blkg_to_tg(ctx.blkg); | |
1410 | ||
1411 | if (is_u64) | |
1412 | *(u64 *)((void *)tg + of_cft(of)->private) = v; | |
1413 | else | |
1414 | *(unsigned int *)((void *)tg + of_cft(of)->private) = v; | |
1415 | ||
1416 | tg_conf_updated(tg); | |
1417 | ret = 0; | |
1418 | out_finish: | |
1419 | blkg_conf_finish(&ctx); | |
1420 | return ret ?: 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_legacy_files[] = { | |
1436 | { | |
1437 | .name = "throttle.read_bps_device", | |
1438 | .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]), | |
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][LIMIT_MAX]), | |
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][LIMIT_MAX]), | |
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][LIMIT_MAX]), | |
1457 | .seq_show = tg_print_conf_uint, | |
1458 | .write = tg_set_conf_uint, | |
1459 | }, | |
1460 | { | |
1461 | .name = "throttle.io_service_bytes", | |
1462 | .private = (unsigned long)&blkcg_policy_throtl, | |
1463 | .seq_show = blkg_print_stat_bytes, | |
1464 | }, | |
1465 | { | |
1466 | .name = "throttle.io_serviced", | |
1467 | .private = (unsigned long)&blkcg_policy_throtl, | |
1468 | .seq_show = blkg_print_stat_ios, | |
1469 | }, | |
1470 | { } /* terminate */ | |
1471 | }; | |
1472 | ||
1473 | static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd, | |
1474 | int off) | |
1475 | { | |
1476 | struct throtl_grp *tg = pd_to_tg(pd); | |
1477 | const char *dname = blkg_dev_name(pd->blkg); | |
1478 | char bufs[4][21] = { "max", "max", "max", "max" }; | |
1479 | u64 bps_dft; | |
1480 | unsigned int iops_dft; | |
1481 | char idle_time[26] = ""; | |
1482 | char latency_time[26] = ""; | |
1483 | ||
1484 | if (!dname) | |
1485 | return 0; | |
1486 | ||
1487 | if (off == LIMIT_LOW) { | |
1488 | bps_dft = 0; | |
1489 | iops_dft = 0; | |
1490 | } else { | |
1491 | bps_dft = U64_MAX; | |
1492 | iops_dft = UINT_MAX; | |
1493 | } | |
1494 | ||
1495 | if (tg->bps_conf[READ][off] == bps_dft && | |
1496 | tg->bps_conf[WRITE][off] == bps_dft && | |
1497 | tg->iops_conf[READ][off] == iops_dft && | |
1498 | tg->iops_conf[WRITE][off] == iops_dft && | |
1499 | (off != LIMIT_LOW || | |
1500 | (tg->idletime_threshold == tg->td->dft_idletime_threshold && | |
1501 | tg->latency_target == DFL_LATENCY_TARGET))) | |
1502 | return 0; | |
1503 | ||
1504 | if (tg->bps_conf[READ][off] != bps_dft) | |
1505 | snprintf(bufs[0], sizeof(bufs[0]), "%llu", | |
1506 | tg->bps_conf[READ][off]); | |
1507 | if (tg->bps_conf[WRITE][off] != bps_dft) | |
1508 | snprintf(bufs[1], sizeof(bufs[1]), "%llu", | |
1509 | tg->bps_conf[WRITE][off]); | |
1510 | if (tg->iops_conf[READ][off] != iops_dft) | |
1511 | snprintf(bufs[2], sizeof(bufs[2]), "%u", | |
1512 | tg->iops_conf[READ][off]); | |
1513 | if (tg->iops_conf[WRITE][off] != iops_dft) | |
1514 | snprintf(bufs[3], sizeof(bufs[3]), "%u", | |
1515 | tg->iops_conf[WRITE][off]); | |
1516 | if (off == LIMIT_LOW) { | |
1517 | if (tg->idletime_threshold == ULONG_MAX) | |
1518 | strcpy(idle_time, " idle=max"); | |
1519 | else | |
1520 | snprintf(idle_time, sizeof(idle_time), " idle=%lu", | |
1521 | tg->idletime_threshold); | |
1522 | ||
1523 | if (tg->latency_target == ULONG_MAX) | |
1524 | strcpy(latency_time, " latency=max"); | |
1525 | else | |
1526 | snprintf(latency_time, sizeof(latency_time), | |
1527 | " latency=%lu", tg->latency_target); | |
1528 | } | |
1529 | ||
1530 | seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n", | |
1531 | dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time, | |
1532 | latency_time); | |
1533 | return 0; | |
1534 | } | |
1535 | ||
1536 | static int tg_print_limit(struct seq_file *sf, void *v) | |
1537 | { | |
1538 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit, | |
1539 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
1540 | return 0; | |
1541 | } | |
1542 | ||
1543 | static ssize_t tg_set_limit(struct kernfs_open_file *of, | |
1544 | char *buf, size_t nbytes, loff_t off) | |
1545 | { | |
1546 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1547 | struct blkg_conf_ctx ctx; | |
1548 | struct throtl_grp *tg; | |
1549 | u64 v[4]; | |
1550 | unsigned long idle_time; | |
1551 | unsigned long latency_time; | |
1552 | int ret; | |
1553 | int index = of_cft(of)->private; | |
1554 | ||
1555 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1556 | if (ret) | |
1557 | return ret; | |
1558 | ||
1559 | tg = blkg_to_tg(ctx.blkg); | |
1560 | ||
1561 | v[0] = tg->bps_conf[READ][index]; | |
1562 | v[1] = tg->bps_conf[WRITE][index]; | |
1563 | v[2] = tg->iops_conf[READ][index]; | |
1564 | v[3] = tg->iops_conf[WRITE][index]; | |
1565 | ||
1566 | idle_time = tg->idletime_threshold; | |
1567 | latency_time = tg->latency_target; | |
1568 | while (true) { | |
1569 | char tok[27]; /* wiops=18446744073709551616 */ | |
1570 | char *p; | |
1571 | u64 val = U64_MAX; | |
1572 | int len; | |
1573 | ||
1574 | if (sscanf(ctx.body, "%26s%n", tok, &len) != 1) | |
1575 | break; | |
1576 | if (tok[0] == '\0') | |
1577 | break; | |
1578 | ctx.body += len; | |
1579 | ||
1580 | ret = -EINVAL; | |
1581 | p = tok; | |
1582 | strsep(&p, "="); | |
1583 | if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max"))) | |
1584 | goto out_finish; | |
1585 | ||
1586 | ret = -ERANGE; | |
1587 | if (!val) | |
1588 | goto out_finish; | |
1589 | ||
1590 | ret = -EINVAL; | |
1591 | if (!strcmp(tok, "rbps")) | |
1592 | v[0] = val; | |
1593 | else if (!strcmp(tok, "wbps")) | |
1594 | v[1] = val; | |
1595 | else if (!strcmp(tok, "riops")) | |
1596 | v[2] = min_t(u64, val, UINT_MAX); | |
1597 | else if (!strcmp(tok, "wiops")) | |
1598 | v[3] = min_t(u64, val, UINT_MAX); | |
1599 | else if (off == LIMIT_LOW && !strcmp(tok, "idle")) | |
1600 | idle_time = val; | |
1601 | else if (off == LIMIT_LOW && !strcmp(tok, "latency")) | |
1602 | latency_time = val; | |
1603 | else | |
1604 | goto out_finish; | |
1605 | } | |
1606 | ||
1607 | tg->bps_conf[READ][index] = v[0]; | |
1608 | tg->bps_conf[WRITE][index] = v[1]; | |
1609 | tg->iops_conf[READ][index] = v[2]; | |
1610 | tg->iops_conf[WRITE][index] = v[3]; | |
1611 | ||
1612 | if (index == LIMIT_MAX) { | |
1613 | tg->bps[READ][index] = v[0]; | |
1614 | tg->bps[WRITE][index] = v[1]; | |
1615 | tg->iops[READ][index] = v[2]; | |
1616 | tg->iops[WRITE][index] = v[3]; | |
1617 | } | |
1618 | tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW], | |
1619 | tg->bps_conf[READ][LIMIT_MAX]); | |
1620 | tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW], | |
1621 | tg->bps_conf[WRITE][LIMIT_MAX]); | |
1622 | tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW], | |
1623 | tg->iops_conf[READ][LIMIT_MAX]); | |
1624 | tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW], | |
1625 | tg->iops_conf[WRITE][LIMIT_MAX]); | |
1626 | ||
1627 | if (index == LIMIT_LOW) { | |
1628 | blk_throtl_update_limit_valid(tg->td); | |
1629 | if (tg->td->limit_valid[LIMIT_LOW]) | |
1630 | tg->td->limit_index = LIMIT_LOW; | |
1631 | tg->idletime_threshold = (idle_time == ULONG_MAX) ? | |
1632 | ULONG_MAX : idle_time; | |
1633 | tg->latency_target = (latency_time == ULONG_MAX) ? | |
1634 | ULONG_MAX : latency_time; | |
1635 | } | |
1636 | tg_conf_updated(tg); | |
1637 | ret = 0; | |
1638 | out_finish: | |
1639 | blkg_conf_finish(&ctx); | |
1640 | return ret ?: nbytes; | |
1641 | } | |
1642 | ||
1643 | static struct cftype throtl_files[] = { | |
1644 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
1645 | { | |
1646 | .name = "low", | |
1647 | .flags = CFTYPE_NOT_ON_ROOT, | |
1648 | .seq_show = tg_print_limit, | |
1649 | .write = tg_set_limit, | |
1650 | .private = LIMIT_LOW, | |
1651 | }, | |
1652 | #endif | |
1653 | { | |
1654 | .name = "max", | |
1655 | .flags = CFTYPE_NOT_ON_ROOT, | |
1656 | .seq_show = tg_print_limit, | |
1657 | .write = tg_set_limit, | |
1658 | .private = LIMIT_MAX, | |
1659 | }, | |
1660 | { } /* terminate */ | |
1661 | }; | |
1662 | ||
1663 | static void throtl_shutdown_wq(struct request_queue *q) | |
1664 | { | |
1665 | struct throtl_data *td = q->td; | |
1666 | ||
1667 | cancel_work_sync(&td->dispatch_work); | |
1668 | } | |
1669 | ||
1670 | static struct blkcg_policy blkcg_policy_throtl = { | |
1671 | .dfl_cftypes = throtl_files, | |
1672 | .legacy_cftypes = throtl_legacy_files, | |
1673 | ||
1674 | .pd_alloc_fn = throtl_pd_alloc, | |
1675 | .pd_init_fn = throtl_pd_init, | |
1676 | .pd_online_fn = throtl_pd_online, | |
1677 | .pd_offline_fn = throtl_pd_offline, | |
1678 | .pd_free_fn = throtl_pd_free, | |
1679 | }; | |
1680 | ||
1681 | static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg) | |
1682 | { | |
1683 | unsigned long rtime = jiffies, wtime = jiffies; | |
1684 | ||
1685 | if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]) | |
1686 | rtime = tg->last_low_overflow_time[READ]; | |
1687 | if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) | |
1688 | wtime = tg->last_low_overflow_time[WRITE]; | |
1689 | return min(rtime, wtime); | |
1690 | } | |
1691 | ||
1692 | /* tg should not be an intermediate node */ | |
1693 | static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg) | |
1694 | { | |
1695 | struct throtl_service_queue *parent_sq; | |
1696 | struct throtl_grp *parent = tg; | |
1697 | unsigned long ret = __tg_last_low_overflow_time(tg); | |
1698 | ||
1699 | while (true) { | |
1700 | parent_sq = parent->service_queue.parent_sq; | |
1701 | parent = sq_to_tg(parent_sq); | |
1702 | if (!parent) | |
1703 | break; | |
1704 | ||
1705 | /* | |
1706 | * The parent doesn't have low limit, it always reaches low | |
1707 | * limit. Its overflow time is useless for children | |
1708 | */ | |
1709 | if (!parent->bps[READ][LIMIT_LOW] && | |
1710 | !parent->iops[READ][LIMIT_LOW] && | |
1711 | !parent->bps[WRITE][LIMIT_LOW] && | |
1712 | !parent->iops[WRITE][LIMIT_LOW]) | |
1713 | continue; | |
1714 | if (time_after(__tg_last_low_overflow_time(parent), ret)) | |
1715 | ret = __tg_last_low_overflow_time(parent); | |
1716 | } | |
1717 | return ret; | |
1718 | } | |
1719 | ||
1720 | static bool throtl_tg_is_idle(struct throtl_grp *tg) | |
1721 | { | |
1722 | /* | |
1723 | * cgroup is idle if: | |
1724 | * - single idle is too long, longer than a fixed value (in case user | |
1725 | * configure a too big threshold) or 4 times of slice | |
1726 | * - average think time is more than threshold | |
1727 | * - IO latency is largely below threshold | |
1728 | */ | |
1729 | unsigned long time = jiffies_to_usecs(4 * tg->td->throtl_slice); | |
1730 | ||
1731 | time = min_t(unsigned long, MAX_IDLE_TIME, time); | |
1732 | return (ktime_get_ns() >> 10) - tg->last_finish_time > time || | |
1733 | tg->avg_idletime > tg->idletime_threshold || | |
1734 | (tg->latency_target && tg->bio_cnt && | |
1735 | tg->bad_bio_cnt * 5 < tg->bio_cnt); | |
1736 | } | |
1737 | ||
1738 | static bool throtl_tg_can_upgrade(struct throtl_grp *tg) | |
1739 | { | |
1740 | struct throtl_service_queue *sq = &tg->service_queue; | |
1741 | bool read_limit, write_limit; | |
1742 | ||
1743 | /* | |
1744 | * if cgroup reaches low limit (if low limit is 0, the cgroup always | |
1745 | * reaches), it's ok to upgrade to next limit | |
1746 | */ | |
1747 | read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]; | |
1748 | write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]; | |
1749 | if (!read_limit && !write_limit) | |
1750 | return true; | |
1751 | if (read_limit && sq->nr_queued[READ] && | |
1752 | (!write_limit || sq->nr_queued[WRITE])) | |
1753 | return true; | |
1754 | if (write_limit && sq->nr_queued[WRITE] && | |
1755 | (!read_limit || sq->nr_queued[READ])) | |
1756 | return true; | |
1757 | ||
1758 | if (time_after_eq(jiffies, | |
1759 | tg_last_low_overflow_time(tg) + tg->td->throtl_slice) && | |
1760 | throtl_tg_is_idle(tg)) | |
1761 | return true; | |
1762 | return false; | |
1763 | } | |
1764 | ||
1765 | static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg) | |
1766 | { | |
1767 | while (true) { | |
1768 | if (throtl_tg_can_upgrade(tg)) | |
1769 | return true; | |
1770 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1771 | if (!tg || !tg_to_blkg(tg)->parent) | |
1772 | return false; | |
1773 | } | |
1774 | return false; | |
1775 | } | |
1776 | ||
1777 | static bool throtl_can_upgrade(struct throtl_data *td, | |
1778 | struct throtl_grp *this_tg) | |
1779 | { | |
1780 | struct cgroup_subsys_state *pos_css; | |
1781 | struct blkcg_gq *blkg; | |
1782 | ||
1783 | if (td->limit_index != LIMIT_LOW) | |
1784 | return false; | |
1785 | ||
1786 | if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice)) | |
1787 | return false; | |
1788 | ||
1789 | rcu_read_lock(); | |
1790 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1791 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1792 | ||
1793 | if (tg == this_tg) | |
1794 | continue; | |
1795 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1796 | continue; | |
1797 | if (!throtl_hierarchy_can_upgrade(tg)) { | |
1798 | rcu_read_unlock(); | |
1799 | return false; | |
1800 | } | |
1801 | } | |
1802 | rcu_read_unlock(); | |
1803 | return true; | |
1804 | } | |
1805 | ||
1806 | static void throtl_upgrade_check(struct throtl_grp *tg) | |
1807 | { | |
1808 | unsigned long now = jiffies; | |
1809 | ||
1810 | if (tg->td->limit_index != LIMIT_LOW) | |
1811 | return; | |
1812 | ||
1813 | if (time_after(tg->last_check_time + tg->td->throtl_slice, now)) | |
1814 | return; | |
1815 | ||
1816 | tg->last_check_time = now; | |
1817 | ||
1818 | if (!time_after_eq(now, | |
1819 | __tg_last_low_overflow_time(tg) + tg->td->throtl_slice)) | |
1820 | return; | |
1821 | ||
1822 | if (throtl_can_upgrade(tg->td, NULL)) | |
1823 | throtl_upgrade_state(tg->td); | |
1824 | } | |
1825 | ||
1826 | static void throtl_upgrade_state(struct throtl_data *td) | |
1827 | { | |
1828 | struct cgroup_subsys_state *pos_css; | |
1829 | struct blkcg_gq *blkg; | |
1830 | ||
1831 | td->limit_index = LIMIT_MAX; | |
1832 | td->low_upgrade_time = jiffies; | |
1833 | td->scale = 0; | |
1834 | rcu_read_lock(); | |
1835 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1836 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1837 | struct throtl_service_queue *sq = &tg->service_queue; | |
1838 | ||
1839 | tg->disptime = jiffies - 1; | |
1840 | throtl_select_dispatch(sq); | |
1841 | throtl_schedule_next_dispatch(sq, false); | |
1842 | } | |
1843 | rcu_read_unlock(); | |
1844 | throtl_select_dispatch(&td->service_queue); | |
1845 | throtl_schedule_next_dispatch(&td->service_queue, false); | |
1846 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1847 | } | |
1848 | ||
1849 | static void throtl_downgrade_state(struct throtl_data *td, int new) | |
1850 | { | |
1851 | td->scale /= 2; | |
1852 | ||
1853 | if (td->scale) { | |
1854 | td->low_upgrade_time = jiffies - td->scale * td->throtl_slice; | |
1855 | return; | |
1856 | } | |
1857 | ||
1858 | td->limit_index = new; | |
1859 | td->low_downgrade_time = jiffies; | |
1860 | } | |
1861 | ||
1862 | static bool throtl_tg_can_downgrade(struct throtl_grp *tg) | |
1863 | { | |
1864 | struct throtl_data *td = tg->td; | |
1865 | unsigned long now = jiffies; | |
1866 | ||
1867 | /* | |
1868 | * If cgroup is below low limit, consider downgrade and throttle other | |
1869 | * cgroups | |
1870 | */ | |
1871 | if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) && | |
1872 | time_after_eq(now, tg_last_low_overflow_time(tg) + | |
1873 | td->throtl_slice) && | |
1874 | (!throtl_tg_is_idle(tg) || | |
1875 | !list_empty(&tg_to_blkg(tg)->blkcg->css.children))) | |
1876 | return true; | |
1877 | return false; | |
1878 | } | |
1879 | ||
1880 | static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg) | |
1881 | { | |
1882 | while (true) { | |
1883 | if (!throtl_tg_can_downgrade(tg)) | |
1884 | return false; | |
1885 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1886 | if (!tg || !tg_to_blkg(tg)->parent) | |
1887 | break; | |
1888 | } | |
1889 | return true; | |
1890 | } | |
1891 | ||
1892 | static void throtl_downgrade_check(struct throtl_grp *tg) | |
1893 | { | |
1894 | uint64_t bps; | |
1895 | unsigned int iops; | |
1896 | unsigned long elapsed_time; | |
1897 | unsigned long now = jiffies; | |
1898 | ||
1899 | if (tg->td->limit_index != LIMIT_MAX || | |
1900 | !tg->td->limit_valid[LIMIT_LOW]) | |
1901 | return; | |
1902 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1903 | return; | |
1904 | if (time_after(tg->last_check_time + tg->td->throtl_slice, now)) | |
1905 | return; | |
1906 | ||
1907 | elapsed_time = now - tg->last_check_time; | |
1908 | tg->last_check_time = now; | |
1909 | ||
1910 | if (time_before(now, tg_last_low_overflow_time(tg) + | |
1911 | tg->td->throtl_slice)) | |
1912 | return; | |
1913 | ||
1914 | if (tg->bps[READ][LIMIT_LOW]) { | |
1915 | bps = tg->last_bytes_disp[READ] * HZ; | |
1916 | do_div(bps, elapsed_time); | |
1917 | if (bps >= tg->bps[READ][LIMIT_LOW]) | |
1918 | tg->last_low_overflow_time[READ] = now; | |
1919 | } | |
1920 | ||
1921 | if (tg->bps[WRITE][LIMIT_LOW]) { | |
1922 | bps = tg->last_bytes_disp[WRITE] * HZ; | |
1923 | do_div(bps, elapsed_time); | |
1924 | if (bps >= tg->bps[WRITE][LIMIT_LOW]) | |
1925 | tg->last_low_overflow_time[WRITE] = now; | |
1926 | } | |
1927 | ||
1928 | if (tg->iops[READ][LIMIT_LOW]) { | |
1929 | iops = tg->last_io_disp[READ] * HZ / elapsed_time; | |
1930 | if (iops >= tg->iops[READ][LIMIT_LOW]) | |
1931 | tg->last_low_overflow_time[READ] = now; | |
1932 | } | |
1933 | ||
1934 | if (tg->iops[WRITE][LIMIT_LOW]) { | |
1935 | iops = tg->last_io_disp[WRITE] * HZ / elapsed_time; | |
1936 | if (iops >= tg->iops[WRITE][LIMIT_LOW]) | |
1937 | tg->last_low_overflow_time[WRITE] = now; | |
1938 | } | |
1939 | ||
1940 | /* | |
1941 | * If cgroup is below low limit, consider downgrade and throttle other | |
1942 | * cgroups | |
1943 | */ | |
1944 | if (throtl_hierarchy_can_downgrade(tg)) | |
1945 | throtl_downgrade_state(tg->td, LIMIT_LOW); | |
1946 | ||
1947 | tg->last_bytes_disp[READ] = 0; | |
1948 | tg->last_bytes_disp[WRITE] = 0; | |
1949 | tg->last_io_disp[READ] = 0; | |
1950 | tg->last_io_disp[WRITE] = 0; | |
1951 | } | |
1952 | ||
1953 | static void blk_throtl_update_idletime(struct throtl_grp *tg) | |
1954 | { | |
1955 | unsigned long now = ktime_get_ns() >> 10; | |
1956 | unsigned long last_finish_time = tg->last_finish_time; | |
1957 | ||
1958 | if (now <= last_finish_time || last_finish_time == 0 || | |
1959 | last_finish_time == tg->checked_last_finish_time) | |
1960 | return; | |
1961 | ||
1962 | tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3; | |
1963 | tg->checked_last_finish_time = last_finish_time; | |
1964 | } | |
1965 | ||
1966 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
1967 | static void throtl_update_latency_buckets(struct throtl_data *td) | |
1968 | { | |
1969 | struct avg_latency_bucket avg_latency[LATENCY_BUCKET_SIZE]; | |
1970 | int i, cpu; | |
1971 | unsigned long last_latency = 0; | |
1972 | unsigned long latency; | |
1973 | ||
1974 | if (!blk_queue_nonrot(td->queue)) | |
1975 | return; | |
1976 | if (time_before(jiffies, td->last_calculate_time + HZ)) | |
1977 | return; | |
1978 | td->last_calculate_time = jiffies; | |
1979 | ||
1980 | memset(avg_latency, 0, sizeof(avg_latency)); | |
1981 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) { | |
1982 | struct latency_bucket *tmp = &td->tmp_buckets[i]; | |
1983 | ||
1984 | for_each_possible_cpu(cpu) { | |
1985 | struct latency_bucket *bucket; | |
1986 | ||
1987 | /* this isn't race free, but ok in practice */ | |
1988 | bucket = per_cpu_ptr(td->latency_buckets, cpu); | |
1989 | tmp->total_latency += bucket[i].total_latency; | |
1990 | tmp->samples += bucket[i].samples; | |
1991 | bucket[i].total_latency = 0; | |
1992 | bucket[i].samples = 0; | |
1993 | } | |
1994 | ||
1995 | if (tmp->samples >= 32) { | |
1996 | int samples = tmp->samples; | |
1997 | ||
1998 | latency = tmp->total_latency; | |
1999 | ||
2000 | tmp->total_latency = 0; | |
2001 | tmp->samples = 0; | |
2002 | latency /= samples; | |
2003 | if (latency == 0) | |
2004 | continue; | |
2005 | avg_latency[i].latency = latency; | |
2006 | } | |
2007 | } | |
2008 | ||
2009 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) { | |
2010 | if (!avg_latency[i].latency) { | |
2011 | if (td->avg_buckets[i].latency < last_latency) | |
2012 | td->avg_buckets[i].latency = last_latency; | |
2013 | continue; | |
2014 | } | |
2015 | ||
2016 | if (!td->avg_buckets[i].valid) | |
2017 | latency = avg_latency[i].latency; | |
2018 | else | |
2019 | latency = (td->avg_buckets[i].latency * 7 + | |
2020 | avg_latency[i].latency) >> 3; | |
2021 | ||
2022 | td->avg_buckets[i].latency = max(latency, last_latency); | |
2023 | td->avg_buckets[i].valid = true; | |
2024 | last_latency = td->avg_buckets[i].latency; | |
2025 | } | |
2026 | } | |
2027 | #else | |
2028 | static inline void throtl_update_latency_buckets(struct throtl_data *td) | |
2029 | { | |
2030 | } | |
2031 | #endif | |
2032 | ||
2033 | static void blk_throtl_assoc_bio(struct throtl_grp *tg, struct bio *bio) | |
2034 | { | |
2035 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
2036 | int ret; | |
2037 | ||
2038 | ret = bio_associate_current(bio); | |
2039 | if (ret == 0 || ret == -EBUSY) | |
2040 | bio->bi_cg_private = tg; | |
2041 | blk_stat_set_issue(&bio->bi_issue_stat, bio_sectors(bio)); | |
2042 | #else | |
2043 | bio_associate_current(bio); | |
2044 | #endif | |
2045 | } | |
2046 | ||
2047 | bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg, | |
2048 | struct bio *bio) | |
2049 | { | |
2050 | struct throtl_qnode *qn = NULL; | |
2051 | struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg); | |
2052 | struct throtl_service_queue *sq; | |
2053 | bool rw = bio_data_dir(bio); | |
2054 | bool throttled = false; | |
2055 | struct throtl_data *td = tg->td; | |
2056 | ||
2057 | WARN_ON_ONCE(!rcu_read_lock_held()); | |
2058 | ||
2059 | /* see throtl_charge_bio() */ | |
2060 | if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw]) | |
2061 | goto out; | |
2062 | ||
2063 | spin_lock_irq(q->queue_lock); | |
2064 | ||
2065 | throtl_update_latency_buckets(td); | |
2066 | ||
2067 | if (unlikely(blk_queue_bypass(q))) | |
2068 | goto out_unlock; | |
2069 | ||
2070 | blk_throtl_assoc_bio(tg, bio); | |
2071 | blk_throtl_update_idletime(tg); | |
2072 | ||
2073 | sq = &tg->service_queue; | |
2074 | ||
2075 | again: | |
2076 | while (true) { | |
2077 | if (tg->last_low_overflow_time[rw] == 0) | |
2078 | tg->last_low_overflow_time[rw] = jiffies; | |
2079 | throtl_downgrade_check(tg); | |
2080 | throtl_upgrade_check(tg); | |
2081 | /* throtl is FIFO - if bios are already queued, should queue */ | |
2082 | if (sq->nr_queued[rw]) | |
2083 | break; | |
2084 | ||
2085 | /* if above limits, break to queue */ | |
2086 | if (!tg_may_dispatch(tg, bio, NULL)) { | |
2087 | tg->last_low_overflow_time[rw] = jiffies; | |
2088 | if (throtl_can_upgrade(td, tg)) { | |
2089 | throtl_upgrade_state(td); | |
2090 | goto again; | |
2091 | } | |
2092 | break; | |
2093 | } | |
2094 | ||
2095 | /* within limits, let's charge and dispatch directly */ | |
2096 | throtl_charge_bio(tg, bio); | |
2097 | ||
2098 | /* | |
2099 | * We need to trim slice even when bios are not being queued | |
2100 | * otherwise it might happen that a bio is not queued for | |
2101 | * a long time and slice keeps on extending and trim is not | |
2102 | * called for a long time. Now if limits are reduced suddenly | |
2103 | * we take into account all the IO dispatched so far at new | |
2104 | * low rate and * newly queued IO gets a really long dispatch | |
2105 | * time. | |
2106 | * | |
2107 | * So keep on trimming slice even if bio is not queued. | |
2108 | */ | |
2109 | throtl_trim_slice(tg, rw); | |
2110 | ||
2111 | /* | |
2112 | * @bio passed through this layer without being throttled. | |
2113 | * Climb up the ladder. If we''re already at the top, it | |
2114 | * can be executed directly. | |
2115 | */ | |
2116 | qn = &tg->qnode_on_parent[rw]; | |
2117 | sq = sq->parent_sq; | |
2118 | tg = sq_to_tg(sq); | |
2119 | if (!tg) | |
2120 | goto out_unlock; | |
2121 | } | |
2122 | ||
2123 | /* out-of-limit, queue to @tg */ | |
2124 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", | |
2125 | rw == READ ? 'R' : 'W', | |
2126 | tg->bytes_disp[rw], bio->bi_iter.bi_size, | |
2127 | tg_bps_limit(tg, rw), | |
2128 | tg->io_disp[rw], tg_iops_limit(tg, rw), | |
2129 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
2130 | ||
2131 | tg->last_low_overflow_time[rw] = jiffies; | |
2132 | ||
2133 | td->nr_queued[rw]++; | |
2134 | throtl_add_bio_tg(bio, qn, tg); | |
2135 | throttled = true; | |
2136 | ||
2137 | /* | |
2138 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
2139 | * was empty before @bio. The forced scheduling isn't likely to | |
2140 | * cause undue delay as @bio is likely to be dispatched directly if | |
2141 | * its @tg's disptime is not in the future. | |
2142 | */ | |
2143 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
2144 | tg_update_disptime(tg); | |
2145 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); | |
2146 | } | |
2147 | ||
2148 | out_unlock: | |
2149 | spin_unlock_irq(q->queue_lock); | |
2150 | out: | |
2151 | /* | |
2152 | * As multiple blk-throtls may stack in the same issue path, we | |
2153 | * don't want bios to leave with the flag set. Clear the flag if | |
2154 | * being issued. | |
2155 | */ | |
2156 | if (!throttled) | |
2157 | bio_clear_flag(bio, BIO_THROTTLED); | |
2158 | ||
2159 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
2160 | if (throttled || !td->track_bio_latency) | |
2161 | bio->bi_issue_stat.stat |= SKIP_LATENCY; | |
2162 | #endif | |
2163 | return throttled; | |
2164 | } | |
2165 | ||
2166 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
2167 | static void throtl_track_latency(struct throtl_data *td, sector_t size, | |
2168 | int op, unsigned long time) | |
2169 | { | |
2170 | struct latency_bucket *latency; | |
2171 | int index; | |
2172 | ||
2173 | if (!td || td->limit_index != LIMIT_LOW || op != REQ_OP_READ || | |
2174 | !blk_queue_nonrot(td->queue)) | |
2175 | return; | |
2176 | ||
2177 | index = request_bucket_index(size); | |
2178 | ||
2179 | latency = get_cpu_ptr(td->latency_buckets); | |
2180 | latency[index].total_latency += time; | |
2181 | latency[index].samples++; | |
2182 | put_cpu_ptr(td->latency_buckets); | |
2183 | } | |
2184 | ||
2185 | void blk_throtl_stat_add(struct request *rq, u64 time_ns) | |
2186 | { | |
2187 | struct request_queue *q = rq->q; | |
2188 | struct throtl_data *td = q->td; | |
2189 | ||
2190 | throtl_track_latency(td, blk_stat_size(&rq->issue_stat), | |
2191 | req_op(rq), time_ns >> 10); | |
2192 | } | |
2193 | ||
2194 | void blk_throtl_bio_endio(struct bio *bio) | |
2195 | { | |
2196 | struct throtl_grp *tg; | |
2197 | u64 finish_time_ns; | |
2198 | unsigned long finish_time; | |
2199 | unsigned long start_time; | |
2200 | unsigned long lat; | |
2201 | ||
2202 | tg = bio->bi_cg_private; | |
2203 | if (!tg) | |
2204 | return; | |
2205 | bio->bi_cg_private = NULL; | |
2206 | ||
2207 | finish_time_ns = ktime_get_ns(); | |
2208 | tg->last_finish_time = finish_time_ns >> 10; | |
2209 | ||
2210 | start_time = blk_stat_time(&bio->bi_issue_stat) >> 10; | |
2211 | finish_time = __blk_stat_time(finish_time_ns) >> 10; | |
2212 | if (!start_time || finish_time <= start_time) | |
2213 | return; | |
2214 | ||
2215 | lat = finish_time - start_time; | |
2216 | /* this is only for bio based driver */ | |
2217 | if (!(bio->bi_issue_stat.stat & SKIP_LATENCY)) | |
2218 | throtl_track_latency(tg->td, blk_stat_size(&bio->bi_issue_stat), | |
2219 | bio_op(bio), lat); | |
2220 | ||
2221 | if (tg->latency_target) { | |
2222 | int bucket; | |
2223 | unsigned int threshold; | |
2224 | ||
2225 | bucket = request_bucket_index( | |
2226 | blk_stat_size(&bio->bi_issue_stat)); | |
2227 | threshold = tg->td->avg_buckets[bucket].latency + | |
2228 | tg->latency_target; | |
2229 | if (lat > threshold) | |
2230 | tg->bad_bio_cnt++; | |
2231 | /* | |
2232 | * Not race free, could get wrong count, which means cgroups | |
2233 | * will be throttled | |
2234 | */ | |
2235 | tg->bio_cnt++; | |
2236 | } | |
2237 | ||
2238 | if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) { | |
2239 | tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies; | |
2240 | tg->bio_cnt /= 2; | |
2241 | tg->bad_bio_cnt /= 2; | |
2242 | } | |
2243 | } | |
2244 | #endif | |
2245 | ||
2246 | /* | |
2247 | * Dispatch all bios from all children tg's queued on @parent_sq. On | |
2248 | * return, @parent_sq is guaranteed to not have any active children tg's | |
2249 | * and all bios from previously active tg's are on @parent_sq->bio_lists[]. | |
2250 | */ | |
2251 | static void tg_drain_bios(struct throtl_service_queue *parent_sq) | |
2252 | { | |
2253 | struct throtl_grp *tg; | |
2254 | ||
2255 | while ((tg = throtl_rb_first(parent_sq))) { | |
2256 | struct throtl_service_queue *sq = &tg->service_queue; | |
2257 | struct bio *bio; | |
2258 | ||
2259 | throtl_dequeue_tg(tg); | |
2260 | ||
2261 | while ((bio = throtl_peek_queued(&sq->queued[READ]))) | |
2262 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); | |
2263 | while ((bio = throtl_peek_queued(&sq->queued[WRITE]))) | |
2264 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); | |
2265 | } | |
2266 | } | |
2267 | ||
2268 | /** | |
2269 | * blk_throtl_drain - drain throttled bios | |
2270 | * @q: request_queue to drain throttled bios for | |
2271 | * | |
2272 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
2273 | */ | |
2274 | void blk_throtl_drain(struct request_queue *q) | |
2275 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
2276 | { | |
2277 | struct throtl_data *td = q->td; | |
2278 | struct blkcg_gq *blkg; | |
2279 | struct cgroup_subsys_state *pos_css; | |
2280 | struct bio *bio; | |
2281 | int rw; | |
2282 | ||
2283 | queue_lockdep_assert_held(q); | |
2284 | rcu_read_lock(); | |
2285 | ||
2286 | /* | |
2287 | * Drain each tg while doing post-order walk on the blkg tree, so | |
2288 | * that all bios are propagated to td->service_queue. It'd be | |
2289 | * better to walk service_queue tree directly but blkg walk is | |
2290 | * easier. | |
2291 | */ | |
2292 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) | |
2293 | tg_drain_bios(&blkg_to_tg(blkg)->service_queue); | |
2294 | ||
2295 | /* finally, transfer bios from top-level tg's into the td */ | |
2296 | tg_drain_bios(&td->service_queue); | |
2297 | ||
2298 | rcu_read_unlock(); | |
2299 | spin_unlock_irq(q->queue_lock); | |
2300 | ||
2301 | /* all bios now should be in td->service_queue, issue them */ | |
2302 | for (rw = READ; rw <= WRITE; rw++) | |
2303 | while ((bio = throtl_pop_queued(&td->service_queue.queued[rw], | |
2304 | NULL))) | |
2305 | generic_make_request(bio); | |
2306 | ||
2307 | spin_lock_irq(q->queue_lock); | |
2308 | } | |
2309 | ||
2310 | int blk_throtl_init(struct request_queue *q) | |
2311 | { | |
2312 | struct throtl_data *td; | |
2313 | int ret; | |
2314 | ||
2315 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
2316 | if (!td) | |
2317 | return -ENOMEM; | |
2318 | td->latency_buckets = __alloc_percpu(sizeof(struct latency_bucket) * | |
2319 | LATENCY_BUCKET_SIZE, __alignof__(u64)); | |
2320 | if (!td->latency_buckets) { | |
2321 | kfree(td); | |
2322 | return -ENOMEM; | |
2323 | } | |
2324 | ||
2325 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); | |
2326 | throtl_service_queue_init(&td->service_queue); | |
2327 | ||
2328 | q->td = td; | |
2329 | td->queue = q; | |
2330 | ||
2331 | td->limit_valid[LIMIT_MAX] = true; | |
2332 | td->limit_index = LIMIT_MAX; | |
2333 | td->low_upgrade_time = jiffies; | |
2334 | td->low_downgrade_time = jiffies; | |
2335 | ||
2336 | /* activate policy */ | |
2337 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); | |
2338 | if (ret) { | |
2339 | free_percpu(td->latency_buckets); | |
2340 | kfree(td); | |
2341 | } | |
2342 | return ret; | |
2343 | } | |
2344 | ||
2345 | void blk_throtl_exit(struct request_queue *q) | |
2346 | { | |
2347 | BUG_ON(!q->td); | |
2348 | throtl_shutdown_wq(q); | |
2349 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); | |
2350 | free_percpu(q->td->latency_buckets); | |
2351 | kfree(q->td); | |
2352 | } | |
2353 | ||
2354 | void blk_throtl_register_queue(struct request_queue *q) | |
2355 | { | |
2356 | struct throtl_data *td; | |
2357 | struct cgroup_subsys_state *pos_css; | |
2358 | struct blkcg_gq *blkg; | |
2359 | ||
2360 | td = q->td; | |
2361 | BUG_ON(!td); | |
2362 | ||
2363 | if (blk_queue_nonrot(q)) { | |
2364 | td->throtl_slice = DFL_THROTL_SLICE_SSD; | |
2365 | td->dft_idletime_threshold = DFL_IDLE_THRESHOLD_SSD; | |
2366 | } else { | |
2367 | td->throtl_slice = DFL_THROTL_SLICE_HD; | |
2368 | td->dft_idletime_threshold = DFL_IDLE_THRESHOLD_HD; | |
2369 | } | |
2370 | #ifndef CONFIG_BLK_DEV_THROTTLING_LOW | |
2371 | /* if no low limit, use previous default */ | |
2372 | td->throtl_slice = DFL_THROTL_SLICE_HD; | |
2373 | #endif | |
2374 | ||
2375 | td->track_bio_latency = !q->mq_ops && !q->request_fn; | |
2376 | if (!td->track_bio_latency) | |
2377 | blk_stat_enable_accounting(q); | |
2378 | ||
2379 | /* | |
2380 | * some tg are created before queue is fully initialized, eg, nonrot | |
2381 | * isn't initialized yet | |
2382 | */ | |
2383 | rcu_read_lock(); | |
2384 | blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) { | |
2385 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
2386 | ||
2387 | tg->idletime_threshold = td->dft_idletime_threshold; | |
2388 | } | |
2389 | rcu_read_unlock(); | |
2390 | } | |
2391 | ||
2392 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
2393 | ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page) | |
2394 | { | |
2395 | if (!q->td) | |
2396 | return -EINVAL; | |
2397 | return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice)); | |
2398 | } | |
2399 | ||
2400 | ssize_t blk_throtl_sample_time_store(struct request_queue *q, | |
2401 | const char *page, size_t count) | |
2402 | { | |
2403 | unsigned long v; | |
2404 | unsigned long t; | |
2405 | ||
2406 | if (!q->td) | |
2407 | return -EINVAL; | |
2408 | if (kstrtoul(page, 10, &v)) | |
2409 | return -EINVAL; | |
2410 | t = msecs_to_jiffies(v); | |
2411 | if (t == 0 || t > MAX_THROTL_SLICE) | |
2412 | return -EINVAL; | |
2413 | q->td->throtl_slice = t; | |
2414 | return count; | |
2415 | } | |
2416 | #endif | |
2417 | ||
2418 | static int __init throtl_init(void) | |
2419 | { | |
2420 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); | |
2421 | if (!kthrotld_workqueue) | |
2422 | panic("Failed to create kthrotld\n"); | |
2423 | ||
2424 | return blkcg_policy_register(&blkcg_policy_throtl); | |
2425 | } | |
2426 | ||
2427 | module_init(throtl_init); |