]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
e43473b7 VG |
2 | /* |
3 | * Interface for controlling IO bandwidth on a request queue | |
4 | * | |
5 | * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com> | |
6 | */ | |
7 | ||
8 | #include <linux/module.h> | |
9 | #include <linux/slab.h> | |
10 | #include <linux/blkdev.h> | |
11 | #include <linux/bio.h> | |
12 | #include <linux/blktrace_api.h> | |
eea8f41c | 13 | #include <linux/blk-cgroup.h> |
bc9fcbf9 | 14 | #include "blk.h" |
e43473b7 VG |
15 | |
16 | /* Max dispatch from a group in 1 round */ | |
17 | static int throtl_grp_quantum = 8; | |
18 | ||
19 | /* Total max dispatch from all groups in one round */ | |
20 | static int throtl_quantum = 32; | |
21 | ||
d61fcfa4 SL |
22 | /* Throttling is performed over a slice and after that slice is renewed */ |
23 | #define DFL_THROTL_SLICE_HD (HZ / 10) | |
24 | #define DFL_THROTL_SLICE_SSD (HZ / 50) | |
297e3d85 | 25 | #define MAX_THROTL_SLICE (HZ) |
9e234eea | 26 | #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */ |
9bb67aeb SL |
27 | #define MIN_THROTL_BPS (320 * 1024) |
28 | #define MIN_THROTL_IOPS (10) | |
b4f428ef SL |
29 | #define DFL_LATENCY_TARGET (-1L) |
30 | #define DFL_IDLE_THRESHOLD (0) | |
6679a90c SL |
31 | #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */ |
32 | #define LATENCY_FILTERED_SSD (0) | |
33 | /* | |
34 | * For HD, very small latency comes from sequential IO. Such IO is helpless to | |
35 | * help determine if its IO is impacted by others, hence we ignore the IO | |
36 | */ | |
37 | #define LATENCY_FILTERED_HD (1000L) /* 1ms */ | |
e43473b7 | 38 | |
3c798398 | 39 | static struct blkcg_policy blkcg_policy_throtl; |
0381411e | 40 | |
450adcbe VG |
41 | /* A workqueue to queue throttle related work */ |
42 | static struct workqueue_struct *kthrotld_workqueue; | |
450adcbe | 43 | |
c5cc2070 TH |
44 | /* |
45 | * To implement hierarchical throttling, throtl_grps form a tree and bios | |
46 | * are dispatched upwards level by level until they reach the top and get | |
47 | * issued. When dispatching bios from the children and local group at each | |
48 | * level, if the bios are dispatched into a single bio_list, there's a risk | |
49 | * of a local or child group which can queue many bios at once filling up | |
50 | * the list starving others. | |
51 | * | |
52 | * To avoid such starvation, dispatched bios are queued separately | |
53 | * according to where they came from. When they are again dispatched to | |
54 | * the parent, they're popped in round-robin order so that no single source | |
55 | * hogs the dispatch window. | |
56 | * | |
57 | * throtl_qnode is used to keep the queued bios separated by their sources. | |
58 | * Bios are queued to throtl_qnode which in turn is queued to | |
59 | * throtl_service_queue and then dispatched in round-robin order. | |
60 | * | |
61 | * It's also used to track the reference counts on blkg's. A qnode always | |
62 | * belongs to a throtl_grp and gets queued on itself or the parent, so | |
63 | * incrementing the reference of the associated throtl_grp when a qnode is | |
64 | * queued and decrementing when dequeued is enough to keep the whole blkg | |
65 | * tree pinned while bios are in flight. | |
66 | */ | |
67 | struct throtl_qnode { | |
68 | struct list_head node; /* service_queue->queued[] */ | |
69 | struct bio_list bios; /* queued bios */ | |
70 | struct throtl_grp *tg; /* tg this qnode belongs to */ | |
71 | }; | |
72 | ||
c9e0332e | 73 | struct throtl_service_queue { |
77216b04 TH |
74 | struct throtl_service_queue *parent_sq; /* the parent service_queue */ |
75 | ||
73f0d49a TH |
76 | /* |
77 | * Bios queued directly to this service_queue or dispatched from | |
78 | * children throtl_grp's. | |
79 | */ | |
c5cc2070 | 80 | struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */ |
73f0d49a TH |
81 | unsigned int nr_queued[2]; /* number of queued bios */ |
82 | ||
83 | /* | |
84 | * RB tree of active children throtl_grp's, which are sorted by | |
85 | * their ->disptime. | |
86 | */ | |
9ff01255 | 87 | struct rb_root_cached pending_tree; /* RB tree of active tgs */ |
c9e0332e TH |
88 | unsigned int nr_pending; /* # queued in the tree */ |
89 | unsigned long first_pending_disptime; /* disptime of the first tg */ | |
69df0ab0 | 90 | struct timer_list pending_timer; /* fires on first_pending_disptime */ |
e43473b7 VG |
91 | }; |
92 | ||
5b2c16aa TH |
93 | enum tg_state_flags { |
94 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
0e9f4164 | 95 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ |
5b2c16aa TH |
96 | }; |
97 | ||
e43473b7 VG |
98 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
99 | ||
9f626e37 | 100 | enum { |
cd5ab1b0 | 101 | LIMIT_LOW, |
9f626e37 SL |
102 | LIMIT_MAX, |
103 | LIMIT_CNT, | |
104 | }; | |
105 | ||
e43473b7 | 106 | struct throtl_grp { |
f95a04af TH |
107 | /* must be the first member */ |
108 | struct blkg_policy_data pd; | |
109 | ||
c9e0332e | 110 | /* active throtl group service_queue member */ |
e43473b7 VG |
111 | struct rb_node rb_node; |
112 | ||
0f3457f6 TH |
113 | /* throtl_data this group belongs to */ |
114 | struct throtl_data *td; | |
115 | ||
49a2f1e3 TH |
116 | /* this group's service queue */ |
117 | struct throtl_service_queue service_queue; | |
118 | ||
c5cc2070 TH |
119 | /* |
120 | * qnode_on_self is used when bios are directly queued to this | |
121 | * throtl_grp so that local bios compete fairly with bios | |
122 | * dispatched from children. qnode_on_parent is used when bios are | |
123 | * dispatched from this throtl_grp into its parent and will compete | |
124 | * with the sibling qnode_on_parents and the parent's | |
125 | * qnode_on_self. | |
126 | */ | |
127 | struct throtl_qnode qnode_on_self[2]; | |
128 | struct throtl_qnode qnode_on_parent[2]; | |
129 | ||
e43473b7 VG |
130 | /* |
131 | * Dispatch time in jiffies. This is the estimated time when group | |
132 | * will unthrottle and is ready to dispatch more bio. It is used as | |
133 | * key to sort active groups in service tree. | |
134 | */ | |
135 | unsigned long disptime; | |
136 | ||
e43473b7 VG |
137 | unsigned int flags; |
138 | ||
693e751e TH |
139 | /* are there any throtl rules between this group and td? */ |
140 | bool has_rules[2]; | |
141 | ||
cd5ab1b0 | 142 | /* internally used bytes per second rate limits */ |
9f626e37 | 143 | uint64_t bps[2][LIMIT_CNT]; |
cd5ab1b0 SL |
144 | /* user configured bps limits */ |
145 | uint64_t bps_conf[2][LIMIT_CNT]; | |
e43473b7 | 146 | |
cd5ab1b0 | 147 | /* internally used IOPS limits */ |
9f626e37 | 148 | unsigned int iops[2][LIMIT_CNT]; |
cd5ab1b0 SL |
149 | /* user configured IOPS limits */ |
150 | unsigned int iops_conf[2][LIMIT_CNT]; | |
8e89d13f | 151 | |
e43473b7 VG |
152 | /* Number of bytes disptached in current slice */ |
153 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
154 | /* Number of bio's dispatched in current slice */ |
155 | unsigned int io_disp[2]; | |
e43473b7 | 156 | |
3f0abd80 SL |
157 | unsigned long last_low_overflow_time[2]; |
158 | ||
159 | uint64_t last_bytes_disp[2]; | |
160 | unsigned int last_io_disp[2]; | |
161 | ||
162 | unsigned long last_check_time; | |
163 | ||
ec80991d | 164 | unsigned long latency_target; /* us */ |
5b81fc3c | 165 | unsigned long latency_target_conf; /* us */ |
e43473b7 VG |
166 | /* When did we start a new slice */ |
167 | unsigned long slice_start[2]; | |
168 | unsigned long slice_end[2]; | |
9e234eea SL |
169 | |
170 | unsigned long last_finish_time; /* ns / 1024 */ | |
171 | unsigned long checked_last_finish_time; /* ns / 1024 */ | |
172 | unsigned long avg_idletime; /* ns / 1024 */ | |
173 | unsigned long idletime_threshold; /* us */ | |
5b81fc3c | 174 | unsigned long idletime_threshold_conf; /* us */ |
53696b8d SL |
175 | |
176 | unsigned int bio_cnt; /* total bios */ | |
177 | unsigned int bad_bio_cnt; /* bios exceeding latency threshold */ | |
178 | unsigned long bio_cnt_reset_time; | |
e43473b7 VG |
179 | }; |
180 | ||
b9147dd1 SL |
181 | /* We measure latency for request size from <= 4k to >= 1M */ |
182 | #define LATENCY_BUCKET_SIZE 9 | |
183 | ||
184 | struct latency_bucket { | |
185 | unsigned long total_latency; /* ns / 1024 */ | |
186 | int samples; | |
187 | }; | |
188 | ||
189 | struct avg_latency_bucket { | |
190 | unsigned long latency; /* ns / 1024 */ | |
191 | bool valid; | |
192 | }; | |
193 | ||
e43473b7 VG |
194 | struct throtl_data |
195 | { | |
e43473b7 | 196 | /* service tree for active throtl groups */ |
c9e0332e | 197 | struct throtl_service_queue service_queue; |
e43473b7 | 198 | |
e43473b7 VG |
199 | struct request_queue *queue; |
200 | ||
201 | /* Total Number of queued bios on READ and WRITE lists */ | |
202 | unsigned int nr_queued[2]; | |
203 | ||
297e3d85 SL |
204 | unsigned int throtl_slice; |
205 | ||
e43473b7 | 206 | /* Work for dispatching throttled bios */ |
69df0ab0 | 207 | struct work_struct dispatch_work; |
9f626e37 SL |
208 | unsigned int limit_index; |
209 | bool limit_valid[LIMIT_CNT]; | |
3f0abd80 SL |
210 | |
211 | unsigned long low_upgrade_time; | |
212 | unsigned long low_downgrade_time; | |
7394e31f SL |
213 | |
214 | unsigned int scale; | |
b9147dd1 | 215 | |
b889bf66 JQ |
216 | struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE]; |
217 | struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE]; | |
218 | struct latency_bucket __percpu *latency_buckets[2]; | |
b9147dd1 | 219 | unsigned long last_calculate_time; |
6679a90c | 220 | unsigned long filtered_latency; |
b9147dd1 SL |
221 | |
222 | bool track_bio_latency; | |
e43473b7 VG |
223 | }; |
224 | ||
e99e88a9 | 225 | static void throtl_pending_timer_fn(struct timer_list *t); |
69df0ab0 | 226 | |
f95a04af TH |
227 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
228 | { | |
229 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
230 | } | |
231 | ||
3c798398 | 232 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 233 | { |
f95a04af | 234 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
235 | } |
236 | ||
3c798398 | 237 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 238 | { |
f95a04af | 239 | return pd_to_blkg(&tg->pd); |
0381411e TH |
240 | } |
241 | ||
fda6f272 TH |
242 | /** |
243 | * sq_to_tg - return the throl_grp the specified service queue belongs to | |
244 | * @sq: the throtl_service_queue of interest | |
245 | * | |
246 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one | |
247 | * embedded in throtl_data, %NULL is returned. | |
248 | */ | |
249 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) | |
250 | { | |
251 | if (sq && sq->parent_sq) | |
252 | return container_of(sq, struct throtl_grp, service_queue); | |
253 | else | |
254 | return NULL; | |
255 | } | |
256 | ||
257 | /** | |
258 | * sq_to_td - return throtl_data the specified service queue belongs to | |
259 | * @sq: the throtl_service_queue of interest | |
260 | * | |
b43daedc | 261 | * A service_queue can be embedded in either a throtl_grp or throtl_data. |
fda6f272 TH |
262 | * Determine the associated throtl_data accordingly and return it. |
263 | */ | |
264 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) | |
265 | { | |
266 | struct throtl_grp *tg = sq_to_tg(sq); | |
267 | ||
268 | if (tg) | |
269 | return tg->td; | |
270 | else | |
271 | return container_of(sq, struct throtl_data, service_queue); | |
272 | } | |
273 | ||
7394e31f SL |
274 | /* |
275 | * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to | |
276 | * make the IO dispatch more smooth. | |
277 | * Scale up: linearly scale up according to lapsed time since upgrade. For | |
278 | * every throtl_slice, the limit scales up 1/2 .low limit till the | |
279 | * limit hits .max limit | |
280 | * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit | |
281 | */ | |
282 | static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td) | |
283 | { | |
284 | /* arbitrary value to avoid too big scale */ | |
285 | if (td->scale < 4096 && time_after_eq(jiffies, | |
286 | td->low_upgrade_time + td->scale * td->throtl_slice)) | |
287 | td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice; | |
288 | ||
289 | return low + (low >> 1) * td->scale; | |
290 | } | |
291 | ||
9f626e37 SL |
292 | static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw) |
293 | { | |
b22c417c | 294 | struct blkcg_gq *blkg = tg_to_blkg(tg); |
7394e31f | 295 | struct throtl_data *td; |
b22c417c SL |
296 | uint64_t ret; |
297 | ||
298 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
299 | return U64_MAX; | |
7394e31f SL |
300 | |
301 | td = tg->td; | |
302 | ret = tg->bps[rw][td->limit_index]; | |
9bb67aeb SL |
303 | if (ret == 0 && td->limit_index == LIMIT_LOW) { |
304 | /* intermediate node or iops isn't 0 */ | |
305 | if (!list_empty(&blkg->blkcg->css.children) || | |
306 | tg->iops[rw][td->limit_index]) | |
307 | return U64_MAX; | |
308 | else | |
309 | return MIN_THROTL_BPS; | |
310 | } | |
7394e31f SL |
311 | |
312 | if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] && | |
313 | tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) { | |
314 | uint64_t adjusted; | |
315 | ||
316 | adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td); | |
317 | ret = min(tg->bps[rw][LIMIT_MAX], adjusted); | |
318 | } | |
b22c417c | 319 | return ret; |
9f626e37 SL |
320 | } |
321 | ||
322 | static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw) | |
323 | { | |
b22c417c | 324 | struct blkcg_gq *blkg = tg_to_blkg(tg); |
7394e31f | 325 | struct throtl_data *td; |
b22c417c SL |
326 | unsigned int ret; |
327 | ||
328 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
329 | return UINT_MAX; | |
9bb67aeb | 330 | |
7394e31f SL |
331 | td = tg->td; |
332 | ret = tg->iops[rw][td->limit_index]; | |
9bb67aeb SL |
333 | if (ret == 0 && tg->td->limit_index == LIMIT_LOW) { |
334 | /* intermediate node or bps isn't 0 */ | |
335 | if (!list_empty(&blkg->blkcg->css.children) || | |
336 | tg->bps[rw][td->limit_index]) | |
337 | return UINT_MAX; | |
338 | else | |
339 | return MIN_THROTL_IOPS; | |
340 | } | |
7394e31f SL |
341 | |
342 | if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] && | |
343 | tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) { | |
344 | uint64_t adjusted; | |
345 | ||
346 | adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td); | |
347 | if (adjusted > UINT_MAX) | |
348 | adjusted = UINT_MAX; | |
349 | ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted); | |
350 | } | |
b22c417c | 351 | return ret; |
9f626e37 SL |
352 | } |
353 | ||
b9147dd1 SL |
354 | #define request_bucket_index(sectors) \ |
355 | clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1) | |
356 | ||
fda6f272 TH |
357 | /** |
358 | * throtl_log - log debug message via blktrace | |
359 | * @sq: the service_queue being reported | |
360 | * @fmt: printf format string | |
361 | * @args: printf args | |
362 | * | |
363 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a | |
364 | * throtl_grp; otherwise, just "throtl". | |
fda6f272 TH |
365 | */ |
366 | #define throtl_log(sq, fmt, args...) do { \ | |
367 | struct throtl_grp *__tg = sq_to_tg((sq)); \ | |
368 | struct throtl_data *__td = sq_to_td((sq)); \ | |
369 | \ | |
370 | (void)__td; \ | |
59fa0224 SL |
371 | if (likely(!blk_trace_note_message_enabled(__td->queue))) \ |
372 | break; \ | |
fda6f272 | 373 | if ((__tg)) { \ |
35fe6d76 SL |
374 | blk_add_cgroup_trace_msg(__td->queue, \ |
375 | tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\ | |
fda6f272 TH |
376 | } else { \ |
377 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
378 | } \ | |
54e7ed12 | 379 | } while (0) |
e43473b7 | 380 | |
ea0ea2bc SL |
381 | static inline unsigned int throtl_bio_data_size(struct bio *bio) |
382 | { | |
383 | /* assume it's one sector */ | |
384 | if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) | |
385 | return 512; | |
386 | return bio->bi_iter.bi_size; | |
387 | } | |
388 | ||
c5cc2070 TH |
389 | static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg) |
390 | { | |
391 | INIT_LIST_HEAD(&qn->node); | |
392 | bio_list_init(&qn->bios); | |
393 | qn->tg = tg; | |
394 | } | |
395 | ||
396 | /** | |
397 | * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it | |
398 | * @bio: bio being added | |
399 | * @qn: qnode to add bio to | |
400 | * @queued: the service_queue->queued[] list @qn belongs to | |
401 | * | |
402 | * Add @bio to @qn and put @qn on @queued if it's not already on. | |
403 | * @qn->tg's reference count is bumped when @qn is activated. See the | |
404 | * comment on top of throtl_qnode definition for details. | |
405 | */ | |
406 | static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn, | |
407 | struct list_head *queued) | |
408 | { | |
409 | bio_list_add(&qn->bios, bio); | |
410 | if (list_empty(&qn->node)) { | |
411 | list_add_tail(&qn->node, queued); | |
412 | blkg_get(tg_to_blkg(qn->tg)); | |
413 | } | |
414 | } | |
415 | ||
416 | /** | |
417 | * throtl_peek_queued - peek the first bio on a qnode list | |
418 | * @queued: the qnode list to peek | |
419 | */ | |
420 | static struct bio *throtl_peek_queued(struct list_head *queued) | |
421 | { | |
422 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
423 | struct bio *bio; | |
424 | ||
425 | if (list_empty(queued)) | |
426 | return NULL; | |
427 | ||
428 | bio = bio_list_peek(&qn->bios); | |
429 | WARN_ON_ONCE(!bio); | |
430 | return bio; | |
431 | } | |
432 | ||
433 | /** | |
434 | * throtl_pop_queued - pop the first bio form a qnode list | |
435 | * @queued: the qnode list to pop a bio from | |
436 | * @tg_to_put: optional out argument for throtl_grp to put | |
437 | * | |
438 | * Pop the first bio from the qnode list @queued. After popping, the first | |
439 | * qnode is removed from @queued if empty or moved to the end of @queued so | |
440 | * that the popping order is round-robin. | |
441 | * | |
442 | * When the first qnode is removed, its associated throtl_grp should be put | |
443 | * too. If @tg_to_put is NULL, this function automatically puts it; | |
444 | * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is | |
445 | * responsible for putting it. | |
446 | */ | |
447 | static struct bio *throtl_pop_queued(struct list_head *queued, | |
448 | struct throtl_grp **tg_to_put) | |
449 | { | |
450 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
451 | struct bio *bio; | |
452 | ||
453 | if (list_empty(queued)) | |
454 | return NULL; | |
455 | ||
456 | bio = bio_list_pop(&qn->bios); | |
457 | WARN_ON_ONCE(!bio); | |
458 | ||
459 | if (bio_list_empty(&qn->bios)) { | |
460 | list_del_init(&qn->node); | |
461 | if (tg_to_put) | |
462 | *tg_to_put = qn->tg; | |
463 | else | |
464 | blkg_put(tg_to_blkg(qn->tg)); | |
465 | } else { | |
466 | list_move_tail(&qn->node, queued); | |
467 | } | |
468 | ||
469 | return bio; | |
470 | } | |
471 | ||
49a2f1e3 | 472 | /* init a service_queue, assumes the caller zeroed it */ |
b2ce2643 | 473 | static void throtl_service_queue_init(struct throtl_service_queue *sq) |
49a2f1e3 | 474 | { |
c5cc2070 TH |
475 | INIT_LIST_HEAD(&sq->queued[0]); |
476 | INIT_LIST_HEAD(&sq->queued[1]); | |
9ff01255 | 477 | sq->pending_tree = RB_ROOT_CACHED; |
e99e88a9 | 478 | timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0); |
69df0ab0 TH |
479 | } |
480 | ||
001bea73 TH |
481 | static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node) |
482 | { | |
4fb72036 | 483 | struct throtl_grp *tg; |
24bdb8ef | 484 | int rw; |
4fb72036 TH |
485 | |
486 | tg = kzalloc_node(sizeof(*tg), gfp, node); | |
487 | if (!tg) | |
77ea7338 | 488 | return NULL; |
4fb72036 | 489 | |
b2ce2643 TH |
490 | throtl_service_queue_init(&tg->service_queue); |
491 | ||
492 | for (rw = READ; rw <= WRITE; rw++) { | |
493 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); | |
494 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); | |
495 | } | |
496 | ||
497 | RB_CLEAR_NODE(&tg->rb_node); | |
9f626e37 SL |
498 | tg->bps[READ][LIMIT_MAX] = U64_MAX; |
499 | tg->bps[WRITE][LIMIT_MAX] = U64_MAX; | |
500 | tg->iops[READ][LIMIT_MAX] = UINT_MAX; | |
501 | tg->iops[WRITE][LIMIT_MAX] = UINT_MAX; | |
cd5ab1b0 SL |
502 | tg->bps_conf[READ][LIMIT_MAX] = U64_MAX; |
503 | tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX; | |
504 | tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX; | |
505 | tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX; | |
506 | /* LIMIT_LOW will have default value 0 */ | |
b2ce2643 | 507 | |
ec80991d | 508 | tg->latency_target = DFL_LATENCY_TARGET; |
5b81fc3c | 509 | tg->latency_target_conf = DFL_LATENCY_TARGET; |
b4f428ef SL |
510 | tg->idletime_threshold = DFL_IDLE_THRESHOLD; |
511 | tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD; | |
ec80991d | 512 | |
4fb72036 | 513 | return &tg->pd; |
001bea73 TH |
514 | } |
515 | ||
a9520cd6 | 516 | static void throtl_pd_init(struct blkg_policy_data *pd) |
a29a171e | 517 | { |
a9520cd6 TH |
518 | struct throtl_grp *tg = pd_to_tg(pd); |
519 | struct blkcg_gq *blkg = tg_to_blkg(tg); | |
77216b04 | 520 | struct throtl_data *td = blkg->q->td; |
b2ce2643 | 521 | struct throtl_service_queue *sq = &tg->service_queue; |
cd1604fa | 522 | |
9138125b | 523 | /* |
aa6ec29b | 524 | * If on the default hierarchy, we switch to properly hierarchical |
9138125b TH |
525 | * behavior where limits on a given throtl_grp are applied to the |
526 | * whole subtree rather than just the group itself. e.g. If 16M | |
527 | * read_bps limit is set on the root group, the whole system can't | |
528 | * exceed 16M for the device. | |
529 | * | |
aa6ec29b | 530 | * If not on the default hierarchy, the broken flat hierarchy |
9138125b TH |
531 | * behavior is retained where all throtl_grps are treated as if |
532 | * they're all separate root groups right below throtl_data. | |
533 | * Limits of a group don't interact with limits of other groups | |
534 | * regardless of the position of the group in the hierarchy. | |
535 | */ | |
b2ce2643 | 536 | sq->parent_sq = &td->service_queue; |
9e10a130 | 537 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent) |
b2ce2643 | 538 | sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue; |
77216b04 | 539 | tg->td = td; |
8a3d2615 TH |
540 | } |
541 | ||
693e751e TH |
542 | /* |
543 | * Set has_rules[] if @tg or any of its parents have limits configured. | |
544 | * This doesn't require walking up to the top of the hierarchy as the | |
545 | * parent's has_rules[] is guaranteed to be correct. | |
546 | */ | |
547 | static void tg_update_has_rules(struct throtl_grp *tg) | |
548 | { | |
549 | struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); | |
9f626e37 | 550 | struct throtl_data *td = tg->td; |
693e751e TH |
551 | int rw; |
552 | ||
553 | for (rw = READ; rw <= WRITE; rw++) | |
554 | tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) || | |
9f626e37 SL |
555 | (td->limit_valid[td->limit_index] && |
556 | (tg_bps_limit(tg, rw) != U64_MAX || | |
557 | tg_iops_limit(tg, rw) != UINT_MAX)); | |
693e751e TH |
558 | } |
559 | ||
a9520cd6 | 560 | static void throtl_pd_online(struct blkg_policy_data *pd) |
693e751e | 561 | { |
aec24246 | 562 | struct throtl_grp *tg = pd_to_tg(pd); |
693e751e TH |
563 | /* |
564 | * We don't want new groups to escape the limits of its ancestors. | |
565 | * Update has_rules[] after a new group is brought online. | |
566 | */ | |
aec24246 | 567 | tg_update_has_rules(tg); |
693e751e TH |
568 | } |
569 | ||
cd5ab1b0 SL |
570 | static void blk_throtl_update_limit_valid(struct throtl_data *td) |
571 | { | |
572 | struct cgroup_subsys_state *pos_css; | |
573 | struct blkcg_gq *blkg; | |
574 | bool low_valid = false; | |
575 | ||
576 | rcu_read_lock(); | |
577 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
578 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
579 | ||
580 | if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] || | |
43ada787 | 581 | tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) { |
cd5ab1b0 | 582 | low_valid = true; |
43ada787 LB |
583 | break; |
584 | } | |
cd5ab1b0 SL |
585 | } |
586 | rcu_read_unlock(); | |
587 | ||
588 | td->limit_valid[LIMIT_LOW] = low_valid; | |
589 | } | |
590 | ||
c79892c5 | 591 | static void throtl_upgrade_state(struct throtl_data *td); |
cd5ab1b0 SL |
592 | static void throtl_pd_offline(struct blkg_policy_data *pd) |
593 | { | |
594 | struct throtl_grp *tg = pd_to_tg(pd); | |
595 | ||
596 | tg->bps[READ][LIMIT_LOW] = 0; | |
597 | tg->bps[WRITE][LIMIT_LOW] = 0; | |
598 | tg->iops[READ][LIMIT_LOW] = 0; | |
599 | tg->iops[WRITE][LIMIT_LOW] = 0; | |
600 | ||
601 | blk_throtl_update_limit_valid(tg->td); | |
602 | ||
c79892c5 SL |
603 | if (!tg->td->limit_valid[tg->td->limit_index]) |
604 | throtl_upgrade_state(tg->td); | |
cd5ab1b0 SL |
605 | } |
606 | ||
001bea73 TH |
607 | static void throtl_pd_free(struct blkg_policy_data *pd) |
608 | { | |
4fb72036 TH |
609 | struct throtl_grp *tg = pd_to_tg(pd); |
610 | ||
b2ce2643 | 611 | del_timer_sync(&tg->service_queue.pending_timer); |
4fb72036 | 612 | kfree(tg); |
001bea73 TH |
613 | } |
614 | ||
0049af73 TH |
615 | static struct throtl_grp * |
616 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 | 617 | { |
9ff01255 | 618 | struct rb_node *n; |
e43473b7 | 619 | /* Service tree is empty */ |
0049af73 | 620 | if (!parent_sq->nr_pending) |
e43473b7 VG |
621 | return NULL; |
622 | ||
9ff01255 LB |
623 | n = rb_first_cached(&parent_sq->pending_tree); |
624 | WARN_ON_ONCE(!n); | |
625 | if (!n) | |
626 | return NULL; | |
627 | return rb_entry_tg(n); | |
e43473b7 VG |
628 | } |
629 | ||
0049af73 TH |
630 | static void throtl_rb_erase(struct rb_node *n, |
631 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 632 | { |
9ff01255 LB |
633 | rb_erase_cached(n, &parent_sq->pending_tree); |
634 | RB_CLEAR_NODE(n); | |
0049af73 | 635 | --parent_sq->nr_pending; |
e43473b7 VG |
636 | } |
637 | ||
0049af73 | 638 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
639 | { |
640 | struct throtl_grp *tg; | |
641 | ||
0049af73 | 642 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
643 | if (!tg) |
644 | return; | |
645 | ||
0049af73 | 646 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
647 | } |
648 | ||
77216b04 | 649 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 650 | { |
77216b04 | 651 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
9ff01255 | 652 | struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node; |
e43473b7 VG |
653 | struct rb_node *parent = NULL; |
654 | struct throtl_grp *__tg; | |
655 | unsigned long key = tg->disptime; | |
9ff01255 | 656 | bool leftmost = true; |
e43473b7 VG |
657 | |
658 | while (*node != NULL) { | |
659 | parent = *node; | |
660 | __tg = rb_entry_tg(parent); | |
661 | ||
662 | if (time_before(key, __tg->disptime)) | |
663 | node = &parent->rb_left; | |
664 | else { | |
665 | node = &parent->rb_right; | |
9ff01255 | 666 | leftmost = false; |
e43473b7 VG |
667 | } |
668 | } | |
669 | ||
e43473b7 | 670 | rb_link_node(&tg->rb_node, parent, node); |
9ff01255 LB |
671 | rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree, |
672 | leftmost); | |
e43473b7 VG |
673 | } |
674 | ||
77216b04 | 675 | static void __throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 676 | { |
77216b04 | 677 | tg_service_queue_add(tg); |
5b2c16aa | 678 | tg->flags |= THROTL_TG_PENDING; |
77216b04 | 679 | tg->service_queue.parent_sq->nr_pending++; |
e43473b7 VG |
680 | } |
681 | ||
77216b04 | 682 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 683 | { |
5b2c16aa | 684 | if (!(tg->flags & THROTL_TG_PENDING)) |
77216b04 | 685 | __throtl_enqueue_tg(tg); |
e43473b7 VG |
686 | } |
687 | ||
77216b04 | 688 | static void __throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 689 | { |
77216b04 | 690 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
5b2c16aa | 691 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
692 | } |
693 | ||
77216b04 | 694 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 695 | { |
5b2c16aa | 696 | if (tg->flags & THROTL_TG_PENDING) |
77216b04 | 697 | __throtl_dequeue_tg(tg); |
e43473b7 VG |
698 | } |
699 | ||
a9131a27 | 700 | /* Call with queue lock held */ |
69df0ab0 TH |
701 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, |
702 | unsigned long expires) | |
a9131a27 | 703 | { |
a41b816c | 704 | unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice; |
06cceedc SL |
705 | |
706 | /* | |
707 | * Since we are adjusting the throttle limit dynamically, the sleep | |
708 | * time calculated according to previous limit might be invalid. It's | |
709 | * possible the cgroup sleep time is very long and no other cgroups | |
710 | * have IO running so notify the limit changes. Make sure the cgroup | |
711 | * doesn't sleep too long to avoid the missed notification. | |
712 | */ | |
713 | if (time_after(expires, max_expire)) | |
714 | expires = max_expire; | |
69df0ab0 TH |
715 | mod_timer(&sq->pending_timer, expires); |
716 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
717 | expires - jiffies, jiffies); | |
a9131a27 TH |
718 | } |
719 | ||
7f52f98c TH |
720 | /** |
721 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle | |
722 | * @sq: the service_queue to schedule dispatch for | |
723 | * @force: force scheduling | |
724 | * | |
725 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the | |
726 | * dispatch time of the first pending child. Returns %true if either timer | |
727 | * is armed or there's no pending child left. %false if the current | |
728 | * dispatch window is still open and the caller should continue | |
729 | * dispatching. | |
730 | * | |
731 | * If @force is %true, the dispatch timer is always scheduled and this | |
732 | * function is guaranteed to return %true. This is to be used when the | |
733 | * caller can't dispatch itself and needs to invoke pending_timer | |
734 | * unconditionally. Note that forced scheduling is likely to induce short | |
735 | * delay before dispatch starts even if @sq->first_pending_disptime is not | |
736 | * in the future and thus shouldn't be used in hot paths. | |
737 | */ | |
738 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
739 | bool force) | |
e43473b7 | 740 | { |
6a525600 | 741 | /* any pending children left? */ |
c9e0332e | 742 | if (!sq->nr_pending) |
7f52f98c | 743 | return true; |
e43473b7 | 744 | |
c9e0332e | 745 | update_min_dispatch_time(sq); |
e43473b7 | 746 | |
69df0ab0 | 747 | /* is the next dispatch time in the future? */ |
7f52f98c | 748 | if (force || time_after(sq->first_pending_disptime, jiffies)) { |
69df0ab0 | 749 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); |
7f52f98c | 750 | return true; |
69df0ab0 TH |
751 | } |
752 | ||
7f52f98c TH |
753 | /* tell the caller to continue dispatching */ |
754 | return false; | |
e43473b7 VG |
755 | } |
756 | ||
32ee5bc4 VG |
757 | static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, |
758 | bool rw, unsigned long start) | |
759 | { | |
760 | tg->bytes_disp[rw] = 0; | |
761 | tg->io_disp[rw] = 0; | |
762 | ||
763 | /* | |
764 | * Previous slice has expired. We must have trimmed it after last | |
765 | * bio dispatch. That means since start of last slice, we never used | |
766 | * that bandwidth. Do try to make use of that bandwidth while giving | |
767 | * credit. | |
768 | */ | |
769 | if (time_after_eq(start, tg->slice_start[rw])) | |
770 | tg->slice_start[rw] = start; | |
771 | ||
297e3d85 | 772 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; |
32ee5bc4 VG |
773 | throtl_log(&tg->service_queue, |
774 | "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", | |
775 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
776 | tg->slice_end[rw], jiffies); | |
777 | } | |
778 | ||
0f3457f6 | 779 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
780 | { |
781 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 782 | tg->io_disp[rw] = 0; |
e43473b7 | 783 | tg->slice_start[rw] = jiffies; |
297e3d85 | 784 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; |
fda6f272 TH |
785 | throtl_log(&tg->service_queue, |
786 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
787 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
788 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
789 | } |
790 | ||
0f3457f6 TH |
791 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
792 | unsigned long jiffy_end) | |
d1ae8ffd | 793 | { |
297e3d85 | 794 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); |
d1ae8ffd VG |
795 | } |
796 | ||
0f3457f6 TH |
797 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
798 | unsigned long jiffy_end) | |
e43473b7 | 799 | { |
297e3d85 | 800 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); |
fda6f272 TH |
801 | throtl_log(&tg->service_queue, |
802 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
803 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
804 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
805 | } |
806 | ||
807 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 808 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
809 | { |
810 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
5cf8c227 | 811 | return false; |
e43473b7 | 812 | |
0b6bad7d | 813 | return true; |
e43473b7 VG |
814 | } |
815 | ||
816 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 817 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 818 | { |
3aad5d3e VG |
819 | unsigned long nr_slices, time_elapsed, io_trim; |
820 | u64 bytes_trim, tmp; | |
e43473b7 VG |
821 | |
822 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
823 | ||
824 | /* | |
825 | * If bps are unlimited (-1), then time slice don't get | |
826 | * renewed. Don't try to trim the slice if slice is used. A new | |
827 | * slice will start when appropriate. | |
828 | */ | |
0f3457f6 | 829 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
830 | return; |
831 | ||
d1ae8ffd VG |
832 | /* |
833 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
834 | * that initially cgroup limit was very low resulting in high | |
835 | * slice_end, but later limit was bumped up and bio was dispached | |
836 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
837 | * is bad because it does not allow new slice to start. | |
838 | */ | |
839 | ||
297e3d85 | 840 | throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice); |
d1ae8ffd | 841 | |
e43473b7 VG |
842 | time_elapsed = jiffies - tg->slice_start[rw]; |
843 | ||
297e3d85 | 844 | nr_slices = time_elapsed / tg->td->throtl_slice; |
e43473b7 VG |
845 | |
846 | if (!nr_slices) | |
847 | return; | |
297e3d85 | 848 | tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices; |
3aad5d3e VG |
849 | do_div(tmp, HZ); |
850 | bytes_trim = tmp; | |
e43473b7 | 851 | |
297e3d85 SL |
852 | io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) / |
853 | HZ; | |
e43473b7 | 854 | |
8e89d13f | 855 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
856 | return; |
857 | ||
858 | if (tg->bytes_disp[rw] >= bytes_trim) | |
859 | tg->bytes_disp[rw] -= bytes_trim; | |
860 | else | |
861 | tg->bytes_disp[rw] = 0; | |
862 | ||
8e89d13f VG |
863 | if (tg->io_disp[rw] >= io_trim) |
864 | tg->io_disp[rw] -= io_trim; | |
865 | else | |
866 | tg->io_disp[rw] = 0; | |
867 | ||
297e3d85 | 868 | tg->slice_start[rw] += nr_slices * tg->td->throtl_slice; |
e43473b7 | 869 | |
fda6f272 TH |
870 | throtl_log(&tg->service_queue, |
871 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
872 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
873 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
e43473b7 VG |
874 | } |
875 | ||
0f3457f6 TH |
876 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
877 | unsigned long *wait) | |
e43473b7 VG |
878 | { |
879 | bool rw = bio_data_dir(bio); | |
8e89d13f | 880 | unsigned int io_allowed; |
e43473b7 | 881 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 882 | u64 tmp; |
e43473b7 | 883 | |
8e89d13f | 884 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 885 | |
8e89d13f VG |
886 | /* Slice has just started. Consider one slice interval */ |
887 | if (!jiffy_elapsed) | |
297e3d85 | 888 | jiffy_elapsed_rnd = tg->td->throtl_slice; |
8e89d13f | 889 | |
297e3d85 | 890 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); |
8e89d13f | 891 | |
c49c06e4 VG |
892 | /* |
893 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
894 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
895 | * will allow dispatch after 1 second and after that slice should | |
896 | * have been trimmed. | |
897 | */ | |
898 | ||
9f626e37 | 899 | tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd; |
c49c06e4 VG |
900 | do_div(tmp, HZ); |
901 | ||
902 | if (tmp > UINT_MAX) | |
903 | io_allowed = UINT_MAX; | |
904 | else | |
905 | io_allowed = tmp; | |
8e89d13f VG |
906 | |
907 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
908 | if (wait) |
909 | *wait = 0; | |
5cf8c227 | 910 | return true; |
e43473b7 VG |
911 | } |
912 | ||
8e89d13f | 913 | /* Calc approx time to dispatch */ |
991f61fe | 914 | jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed; |
8e89d13f VG |
915 | |
916 | if (wait) | |
917 | *wait = jiffy_wait; | |
0b6bad7d | 918 | return false; |
8e89d13f VG |
919 | } |
920 | ||
0f3457f6 TH |
921 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
922 | unsigned long *wait) | |
8e89d13f VG |
923 | { |
924 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 925 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 926 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
ea0ea2bc | 927 | unsigned int bio_size = throtl_bio_data_size(bio); |
e43473b7 VG |
928 | |
929 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
930 | ||
931 | /* Slice has just started. Consider one slice interval */ | |
932 | if (!jiffy_elapsed) | |
297e3d85 | 933 | jiffy_elapsed_rnd = tg->td->throtl_slice; |
e43473b7 | 934 | |
297e3d85 | 935 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); |
e43473b7 | 936 | |
9f626e37 | 937 | tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd; |
5e901a2b | 938 | do_div(tmp, HZ); |
3aad5d3e | 939 | bytes_allowed = tmp; |
e43473b7 | 940 | |
ea0ea2bc | 941 | if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) { |
e43473b7 VG |
942 | if (wait) |
943 | *wait = 0; | |
5cf8c227 | 944 | return true; |
e43473b7 VG |
945 | } |
946 | ||
947 | /* Calc approx time to dispatch */ | |
ea0ea2bc | 948 | extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed; |
9f626e37 | 949 | jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw)); |
e43473b7 VG |
950 | |
951 | if (!jiffy_wait) | |
952 | jiffy_wait = 1; | |
953 | ||
954 | /* | |
955 | * This wait time is without taking into consideration the rounding | |
956 | * up we did. Add that time also. | |
957 | */ | |
958 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
959 | if (wait) |
960 | *wait = jiffy_wait; | |
0b6bad7d | 961 | return false; |
8e89d13f VG |
962 | } |
963 | ||
964 | /* | |
965 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
966 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
967 | */ | |
0f3457f6 TH |
968 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
969 | unsigned long *wait) | |
8e89d13f VG |
970 | { |
971 | bool rw = bio_data_dir(bio); | |
972 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
973 | ||
974 | /* | |
975 | * Currently whole state machine of group depends on first bio | |
976 | * queued in the group bio list. So one should not be calling | |
977 | * this function with a different bio if there are other bios | |
978 | * queued. | |
979 | */ | |
73f0d49a | 980 | BUG_ON(tg->service_queue.nr_queued[rw] && |
c5cc2070 | 981 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); |
e43473b7 | 982 | |
8e89d13f | 983 | /* If tg->bps = -1, then BW is unlimited */ |
9f626e37 SL |
984 | if (tg_bps_limit(tg, rw) == U64_MAX && |
985 | tg_iops_limit(tg, rw) == UINT_MAX) { | |
8e89d13f VG |
986 | if (wait) |
987 | *wait = 0; | |
5cf8c227 | 988 | return true; |
8e89d13f VG |
989 | } |
990 | ||
991 | /* | |
992 | * If previous slice expired, start a new one otherwise renew/extend | |
993 | * existing slice to make sure it is at least throtl_slice interval | |
164c80ed VG |
994 | * long since now. New slice is started only for empty throttle group. |
995 | * If there is queued bio, that means there should be an active | |
996 | * slice and it should be extended instead. | |
8e89d13f | 997 | */ |
164c80ed | 998 | if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw])) |
0f3457f6 | 999 | throtl_start_new_slice(tg, rw); |
8e89d13f | 1000 | else { |
297e3d85 SL |
1001 | if (time_before(tg->slice_end[rw], |
1002 | jiffies + tg->td->throtl_slice)) | |
1003 | throtl_extend_slice(tg, rw, | |
1004 | jiffies + tg->td->throtl_slice); | |
8e89d13f VG |
1005 | } |
1006 | ||
0f3457f6 TH |
1007 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
1008 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
1009 | if (wait) |
1010 | *wait = 0; | |
0b6bad7d | 1011 | return true; |
8e89d13f VG |
1012 | } |
1013 | ||
1014 | max_wait = max(bps_wait, iops_wait); | |
1015 | ||
1016 | if (wait) | |
1017 | *wait = max_wait; | |
1018 | ||
1019 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 1020 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 | 1021 | |
0b6bad7d | 1022 | return false; |
e43473b7 VG |
1023 | } |
1024 | ||
1025 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
1026 | { | |
1027 | bool rw = bio_data_dir(bio); | |
ea0ea2bc | 1028 | unsigned int bio_size = throtl_bio_data_size(bio); |
e43473b7 VG |
1029 | |
1030 | /* Charge the bio to the group */ | |
ea0ea2bc | 1031 | tg->bytes_disp[rw] += bio_size; |
8e89d13f | 1032 | tg->io_disp[rw]++; |
ea0ea2bc | 1033 | tg->last_bytes_disp[rw] += bio_size; |
3f0abd80 | 1034 | tg->last_io_disp[rw]++; |
e43473b7 | 1035 | |
2a0f61e6 | 1036 | /* |
8d2bbd4c | 1037 | * BIO_THROTTLED is used to prevent the same bio to be throttled |
2a0f61e6 TH |
1038 | * more than once as a throttled bio will go through blk-throtl the |
1039 | * second time when it eventually gets issued. Set it when a bio | |
1040 | * is being charged to a tg. | |
2a0f61e6 | 1041 | */ |
8d2bbd4c CH |
1042 | if (!bio_flagged(bio, BIO_THROTTLED)) |
1043 | bio_set_flag(bio, BIO_THROTTLED); | |
e43473b7 VG |
1044 | } |
1045 | ||
c5cc2070 TH |
1046 | /** |
1047 | * throtl_add_bio_tg - add a bio to the specified throtl_grp | |
1048 | * @bio: bio to add | |
1049 | * @qn: qnode to use | |
1050 | * @tg: the target throtl_grp | |
1051 | * | |
1052 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, | |
1053 | * tg->qnode_on_self[] is used. | |
1054 | */ | |
1055 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, | |
1056 | struct throtl_grp *tg) | |
e43473b7 | 1057 | { |
73f0d49a | 1058 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1059 | bool rw = bio_data_dir(bio); |
1060 | ||
c5cc2070 TH |
1061 | if (!qn) |
1062 | qn = &tg->qnode_on_self[rw]; | |
1063 | ||
0e9f4164 TH |
1064 | /* |
1065 | * If @tg doesn't currently have any bios queued in the same | |
1066 | * direction, queueing @bio can change when @tg should be | |
1067 | * dispatched. Mark that @tg was empty. This is automatically | |
1068 | * cleaered on the next tg_update_disptime(). | |
1069 | */ | |
1070 | if (!sq->nr_queued[rw]) | |
1071 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
1072 | ||
c5cc2070 TH |
1073 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); |
1074 | ||
73f0d49a | 1075 | sq->nr_queued[rw]++; |
77216b04 | 1076 | throtl_enqueue_tg(tg); |
e43473b7 VG |
1077 | } |
1078 | ||
77216b04 | 1079 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 1080 | { |
73f0d49a | 1081 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1082 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
1083 | struct bio *bio; | |
1084 | ||
d609af3a ME |
1085 | bio = throtl_peek_queued(&sq->queued[READ]); |
1086 | if (bio) | |
0f3457f6 | 1087 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 1088 | |
d609af3a ME |
1089 | bio = throtl_peek_queued(&sq->queued[WRITE]); |
1090 | if (bio) | |
0f3457f6 | 1091 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
1092 | |
1093 | min_wait = min(read_wait, write_wait); | |
1094 | disptime = jiffies + min_wait; | |
1095 | ||
e43473b7 | 1096 | /* Update dispatch time */ |
77216b04 | 1097 | throtl_dequeue_tg(tg); |
e43473b7 | 1098 | tg->disptime = disptime; |
77216b04 | 1099 | throtl_enqueue_tg(tg); |
0e9f4164 TH |
1100 | |
1101 | /* see throtl_add_bio_tg() */ | |
1102 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
1103 | } |
1104 | ||
32ee5bc4 VG |
1105 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, |
1106 | struct throtl_grp *parent_tg, bool rw) | |
1107 | { | |
1108 | if (throtl_slice_used(parent_tg, rw)) { | |
1109 | throtl_start_new_slice_with_credit(parent_tg, rw, | |
1110 | child_tg->slice_start[rw]); | |
1111 | } | |
1112 | ||
1113 | } | |
1114 | ||
77216b04 | 1115 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 1116 | { |
73f0d49a | 1117 | struct throtl_service_queue *sq = &tg->service_queue; |
6bc9c2b4 TH |
1118 | struct throtl_service_queue *parent_sq = sq->parent_sq; |
1119 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
c5cc2070 | 1120 | struct throtl_grp *tg_to_put = NULL; |
e43473b7 VG |
1121 | struct bio *bio; |
1122 | ||
c5cc2070 TH |
1123 | /* |
1124 | * @bio is being transferred from @tg to @parent_sq. Popping a bio | |
1125 | * from @tg may put its reference and @parent_sq might end up | |
1126 | * getting released prematurely. Remember the tg to put and put it | |
1127 | * after @bio is transferred to @parent_sq. | |
1128 | */ | |
1129 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); | |
73f0d49a | 1130 | sq->nr_queued[rw]--; |
e43473b7 VG |
1131 | |
1132 | throtl_charge_bio(tg, bio); | |
6bc9c2b4 TH |
1133 | |
1134 | /* | |
1135 | * If our parent is another tg, we just need to transfer @bio to | |
1136 | * the parent using throtl_add_bio_tg(). If our parent is | |
1137 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
1138 | * bio_lists[] and decrease total number queued. The caller is | |
1139 | * responsible for issuing these bios. | |
1140 | */ | |
1141 | if (parent_tg) { | |
c5cc2070 | 1142 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); |
32ee5bc4 | 1143 | start_parent_slice_with_credit(tg, parent_tg, rw); |
6bc9c2b4 | 1144 | } else { |
c5cc2070 TH |
1145 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], |
1146 | &parent_sq->queued[rw]); | |
6bc9c2b4 TH |
1147 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
1148 | tg->td->nr_queued[rw]--; | |
1149 | } | |
e43473b7 | 1150 | |
0f3457f6 | 1151 | throtl_trim_slice(tg, rw); |
6bc9c2b4 | 1152 | |
c5cc2070 TH |
1153 | if (tg_to_put) |
1154 | blkg_put(tg_to_blkg(tg_to_put)); | |
e43473b7 VG |
1155 | } |
1156 | ||
77216b04 | 1157 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 1158 | { |
73f0d49a | 1159 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1160 | unsigned int nr_reads = 0, nr_writes = 0; |
1161 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 1162 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
1163 | struct bio *bio; |
1164 | ||
1165 | /* Try to dispatch 75% READS and 25% WRITES */ | |
1166 | ||
c5cc2070 | 1167 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && |
0f3457f6 | 1168 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1169 | |
77216b04 | 1170 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1171 | nr_reads++; |
1172 | ||
1173 | if (nr_reads >= max_nr_reads) | |
1174 | break; | |
1175 | } | |
1176 | ||
c5cc2070 | 1177 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && |
0f3457f6 | 1178 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1179 | |
77216b04 | 1180 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1181 | nr_writes++; |
1182 | ||
1183 | if (nr_writes >= max_nr_writes) | |
1184 | break; | |
1185 | } | |
1186 | ||
1187 | return nr_reads + nr_writes; | |
1188 | } | |
1189 | ||
651930bc | 1190 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
1191 | { |
1192 | unsigned int nr_disp = 0; | |
e43473b7 VG |
1193 | |
1194 | while (1) { | |
73f0d49a | 1195 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
2ab74cd2 | 1196 | struct throtl_service_queue *sq; |
e43473b7 VG |
1197 | |
1198 | if (!tg) | |
1199 | break; | |
1200 | ||
1201 | if (time_before(jiffies, tg->disptime)) | |
1202 | break; | |
1203 | ||
77216b04 | 1204 | throtl_dequeue_tg(tg); |
e43473b7 | 1205 | |
77216b04 | 1206 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 1207 | |
2ab74cd2 | 1208 | sq = &tg->service_queue; |
73f0d49a | 1209 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
77216b04 | 1210 | tg_update_disptime(tg); |
e43473b7 VG |
1211 | |
1212 | if (nr_disp >= throtl_quantum) | |
1213 | break; | |
1214 | } | |
1215 | ||
1216 | return nr_disp; | |
1217 | } | |
1218 | ||
c79892c5 SL |
1219 | static bool throtl_can_upgrade(struct throtl_data *td, |
1220 | struct throtl_grp *this_tg); | |
6e1a5704 TH |
1221 | /** |
1222 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
1223 | * @arg: the throtl_service_queue being serviced | |
1224 | * | |
1225 | * This timer is armed when a child throtl_grp with active bio's become | |
1226 | * pending and queued on the service_queue's pending_tree and expires when | |
1227 | * the first child throtl_grp should be dispatched. This function | |
2e48a530 TH |
1228 | * dispatches bio's from the children throtl_grps to the parent |
1229 | * service_queue. | |
1230 | * | |
1231 | * If the parent's parent is another throtl_grp, dispatching is propagated | |
1232 | * by either arming its pending_timer or repeating dispatch directly. If | |
1233 | * the top-level service_tree is reached, throtl_data->dispatch_work is | |
1234 | * kicked so that the ready bio's are issued. | |
6e1a5704 | 1235 | */ |
e99e88a9 | 1236 | static void throtl_pending_timer_fn(struct timer_list *t) |
69df0ab0 | 1237 | { |
e99e88a9 | 1238 | struct throtl_service_queue *sq = from_timer(sq, t, pending_timer); |
2e48a530 | 1239 | struct throtl_grp *tg = sq_to_tg(sq); |
69df0ab0 | 1240 | struct throtl_data *td = sq_to_td(sq); |
cb76199c | 1241 | struct request_queue *q = td->queue; |
2e48a530 TH |
1242 | struct throtl_service_queue *parent_sq; |
1243 | bool dispatched; | |
6e1a5704 | 1244 | int ret; |
e43473b7 VG |
1245 | |
1246 | spin_lock_irq(q->queue_lock); | |
c79892c5 SL |
1247 | if (throtl_can_upgrade(td, NULL)) |
1248 | throtl_upgrade_state(td); | |
1249 | ||
2e48a530 TH |
1250 | again: |
1251 | parent_sq = sq->parent_sq; | |
1252 | dispatched = false; | |
e43473b7 | 1253 | |
7f52f98c TH |
1254 | while (true) { |
1255 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
2e48a530 TH |
1256 | sq->nr_queued[READ] + sq->nr_queued[WRITE], |
1257 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
7f52f98c TH |
1258 | |
1259 | ret = throtl_select_dispatch(sq); | |
1260 | if (ret) { | |
7f52f98c TH |
1261 | throtl_log(sq, "bios disp=%u", ret); |
1262 | dispatched = true; | |
1263 | } | |
e43473b7 | 1264 | |
7f52f98c TH |
1265 | if (throtl_schedule_next_dispatch(sq, false)) |
1266 | break; | |
e43473b7 | 1267 | |
7f52f98c TH |
1268 | /* this dispatch windows is still open, relax and repeat */ |
1269 | spin_unlock_irq(q->queue_lock); | |
1270 | cpu_relax(); | |
1271 | spin_lock_irq(q->queue_lock); | |
651930bc | 1272 | } |
e43473b7 | 1273 | |
2e48a530 TH |
1274 | if (!dispatched) |
1275 | goto out_unlock; | |
6e1a5704 | 1276 | |
2e48a530 TH |
1277 | if (parent_sq) { |
1278 | /* @parent_sq is another throl_grp, propagate dispatch */ | |
1279 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1280 | tg_update_disptime(tg); | |
1281 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { | |
1282 | /* window is already open, repeat dispatching */ | |
1283 | sq = parent_sq; | |
1284 | tg = sq_to_tg(sq); | |
1285 | goto again; | |
1286 | } | |
1287 | } | |
1288 | } else { | |
1289 | /* reached the top-level, queue issueing */ | |
1290 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1291 | } | |
1292 | out_unlock: | |
e43473b7 | 1293 | spin_unlock_irq(q->queue_lock); |
6e1a5704 | 1294 | } |
e43473b7 | 1295 | |
6e1a5704 TH |
1296 | /** |
1297 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
1298 | * @work: work item being executed | |
1299 | * | |
1300 | * This function is queued for execution when bio's reach the bio_lists[] | |
1301 | * of throtl_data->service_queue. Those bio's are ready and issued by this | |
1302 | * function. | |
1303 | */ | |
8876e140 | 1304 | static void blk_throtl_dispatch_work_fn(struct work_struct *work) |
6e1a5704 TH |
1305 | { |
1306 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1307 | dispatch_work); | |
1308 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1309 | struct request_queue *q = td->queue; | |
1310 | struct bio_list bio_list_on_stack; | |
1311 | struct bio *bio; | |
1312 | struct blk_plug plug; | |
1313 | int rw; | |
1314 | ||
1315 | bio_list_init(&bio_list_on_stack); | |
1316 | ||
1317 | spin_lock_irq(q->queue_lock); | |
c5cc2070 TH |
1318 | for (rw = READ; rw <= WRITE; rw++) |
1319 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) | |
1320 | bio_list_add(&bio_list_on_stack, bio); | |
6e1a5704 TH |
1321 | spin_unlock_irq(q->queue_lock); |
1322 | ||
1323 | if (!bio_list_empty(&bio_list_on_stack)) { | |
69d60eb9 | 1324 | blk_start_plug(&plug); |
e43473b7 VG |
1325 | while((bio = bio_list_pop(&bio_list_on_stack))) |
1326 | generic_make_request(bio); | |
69d60eb9 | 1327 | blk_finish_plug(&plug); |
e43473b7 | 1328 | } |
e43473b7 VG |
1329 | } |
1330 | ||
f95a04af TH |
1331 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
1332 | int off) | |
60c2bc2d | 1333 | { |
f95a04af TH |
1334 | struct throtl_grp *tg = pd_to_tg(pd); |
1335 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 1336 | |
2ab5492d | 1337 | if (v == U64_MAX) |
60c2bc2d | 1338 | return 0; |
f95a04af | 1339 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
1340 | } |
1341 | ||
f95a04af TH |
1342 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
1343 | int off) | |
e43473b7 | 1344 | { |
f95a04af TH |
1345 | struct throtl_grp *tg = pd_to_tg(pd); |
1346 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 1347 | |
2ab5492d | 1348 | if (v == UINT_MAX) |
af133ceb | 1349 | return 0; |
f95a04af | 1350 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
1351 | } |
1352 | ||
2da8ca82 | 1353 | static int tg_print_conf_u64(struct seq_file *sf, void *v) |
8e89d13f | 1354 | { |
2da8ca82 TH |
1355 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64, |
1356 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1357 | return 0; |
8e89d13f VG |
1358 | } |
1359 | ||
2da8ca82 | 1360 | static int tg_print_conf_uint(struct seq_file *sf, void *v) |
8e89d13f | 1361 | { |
2da8ca82 TH |
1362 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint, |
1363 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1364 | return 0; |
60c2bc2d TH |
1365 | } |
1366 | ||
9bb67aeb | 1367 | static void tg_conf_updated(struct throtl_grp *tg, bool global) |
60c2bc2d | 1368 | { |
69948b07 | 1369 | struct throtl_service_queue *sq = &tg->service_queue; |
492eb21b | 1370 | struct cgroup_subsys_state *pos_css; |
69948b07 | 1371 | struct blkcg_gq *blkg; |
af133ceb | 1372 | |
fda6f272 TH |
1373 | throtl_log(&tg->service_queue, |
1374 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
9f626e37 SL |
1375 | tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE), |
1376 | tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE)); | |
632b4493 | 1377 | |
693e751e TH |
1378 | /* |
1379 | * Update has_rules[] flags for the updated tg's subtree. A tg is | |
1380 | * considered to have rules if either the tg itself or any of its | |
1381 | * ancestors has rules. This identifies groups without any | |
1382 | * restrictions in the whole hierarchy and allows them to bypass | |
1383 | * blk-throttle. | |
1384 | */ | |
9bb67aeb SL |
1385 | blkg_for_each_descendant_pre(blkg, pos_css, |
1386 | global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) { | |
5b81fc3c SL |
1387 | struct throtl_grp *this_tg = blkg_to_tg(blkg); |
1388 | struct throtl_grp *parent_tg; | |
1389 | ||
1390 | tg_update_has_rules(this_tg); | |
1391 | /* ignore root/second level */ | |
1392 | if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent || | |
1393 | !blkg->parent->parent) | |
1394 | continue; | |
1395 | parent_tg = blkg_to_tg(blkg->parent); | |
1396 | /* | |
1397 | * make sure all children has lower idle time threshold and | |
1398 | * higher latency target | |
1399 | */ | |
1400 | this_tg->idletime_threshold = min(this_tg->idletime_threshold, | |
1401 | parent_tg->idletime_threshold); | |
1402 | this_tg->latency_target = max(this_tg->latency_target, | |
1403 | parent_tg->latency_target); | |
1404 | } | |
693e751e | 1405 | |
632b4493 TH |
1406 | /* |
1407 | * We're already holding queue_lock and know @tg is valid. Let's | |
1408 | * apply the new config directly. | |
1409 | * | |
1410 | * Restart the slices for both READ and WRITES. It might happen | |
1411 | * that a group's limit are dropped suddenly and we don't want to | |
1412 | * account recently dispatched IO with new low rate. | |
1413 | */ | |
0f3457f6 TH |
1414 | throtl_start_new_slice(tg, 0); |
1415 | throtl_start_new_slice(tg, 1); | |
632b4493 | 1416 | |
5b2c16aa | 1417 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 1418 | tg_update_disptime(tg); |
7f52f98c | 1419 | throtl_schedule_next_dispatch(sq->parent_sq, true); |
632b4493 | 1420 | } |
69948b07 TH |
1421 | } |
1422 | ||
1423 | static ssize_t tg_set_conf(struct kernfs_open_file *of, | |
1424 | char *buf, size_t nbytes, loff_t off, bool is_u64) | |
1425 | { | |
1426 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1427 | struct blkg_conf_ctx ctx; | |
1428 | struct throtl_grp *tg; | |
1429 | int ret; | |
1430 | u64 v; | |
1431 | ||
1432 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1433 | if (ret) | |
1434 | return ret; | |
1435 | ||
1436 | ret = -EINVAL; | |
1437 | if (sscanf(ctx.body, "%llu", &v) != 1) | |
1438 | goto out_finish; | |
1439 | if (!v) | |
2ab5492d | 1440 | v = U64_MAX; |
69948b07 TH |
1441 | |
1442 | tg = blkg_to_tg(ctx.blkg); | |
1443 | ||
1444 | if (is_u64) | |
1445 | *(u64 *)((void *)tg + of_cft(of)->private) = v; | |
1446 | else | |
1447 | *(unsigned int *)((void *)tg + of_cft(of)->private) = v; | |
60c2bc2d | 1448 | |
9bb67aeb | 1449 | tg_conf_updated(tg, false); |
36aa9e5f TH |
1450 | ret = 0; |
1451 | out_finish: | |
60c2bc2d | 1452 | blkg_conf_finish(&ctx); |
36aa9e5f | 1453 | return ret ?: nbytes; |
8e89d13f VG |
1454 | } |
1455 | ||
451af504 TH |
1456 | static ssize_t tg_set_conf_u64(struct kernfs_open_file *of, |
1457 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1458 | { |
451af504 | 1459 | return tg_set_conf(of, buf, nbytes, off, true); |
60c2bc2d TH |
1460 | } |
1461 | ||
451af504 TH |
1462 | static ssize_t tg_set_conf_uint(struct kernfs_open_file *of, |
1463 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1464 | { |
451af504 | 1465 | return tg_set_conf(of, buf, nbytes, off, false); |
60c2bc2d TH |
1466 | } |
1467 | ||
880f50e2 | 1468 | static struct cftype throtl_legacy_files[] = { |
60c2bc2d TH |
1469 | { |
1470 | .name = "throttle.read_bps_device", | |
9f626e37 | 1471 | .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]), |
2da8ca82 | 1472 | .seq_show = tg_print_conf_u64, |
451af504 | 1473 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1474 | }, |
1475 | { | |
1476 | .name = "throttle.write_bps_device", | |
9f626e37 | 1477 | .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]), |
2da8ca82 | 1478 | .seq_show = tg_print_conf_u64, |
451af504 | 1479 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1480 | }, |
1481 | { | |
1482 | .name = "throttle.read_iops_device", | |
9f626e37 | 1483 | .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]), |
2da8ca82 | 1484 | .seq_show = tg_print_conf_uint, |
451af504 | 1485 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1486 | }, |
1487 | { | |
1488 | .name = "throttle.write_iops_device", | |
9f626e37 | 1489 | .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]), |
2da8ca82 | 1490 | .seq_show = tg_print_conf_uint, |
451af504 | 1491 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1492 | }, |
1493 | { | |
1494 | .name = "throttle.io_service_bytes", | |
77ea7338 TH |
1495 | .private = (unsigned long)&blkcg_policy_throtl, |
1496 | .seq_show = blkg_print_stat_bytes, | |
60c2bc2d | 1497 | }, |
17534c6f | 1498 | { |
1499 | .name = "throttle.io_service_bytes_recursive", | |
1500 | .private = (unsigned long)&blkcg_policy_throtl, | |
1501 | .seq_show = blkg_print_stat_bytes_recursive, | |
1502 | }, | |
60c2bc2d TH |
1503 | { |
1504 | .name = "throttle.io_serviced", | |
77ea7338 TH |
1505 | .private = (unsigned long)&blkcg_policy_throtl, |
1506 | .seq_show = blkg_print_stat_ios, | |
60c2bc2d | 1507 | }, |
17534c6f | 1508 | { |
1509 | .name = "throttle.io_serviced_recursive", | |
1510 | .private = (unsigned long)&blkcg_policy_throtl, | |
1511 | .seq_show = blkg_print_stat_ios_recursive, | |
1512 | }, | |
60c2bc2d TH |
1513 | { } /* terminate */ |
1514 | }; | |
1515 | ||
cd5ab1b0 | 1516 | static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd, |
2ee867dc TH |
1517 | int off) |
1518 | { | |
1519 | struct throtl_grp *tg = pd_to_tg(pd); | |
1520 | const char *dname = blkg_dev_name(pd->blkg); | |
1521 | char bufs[4][21] = { "max", "max", "max", "max" }; | |
cd5ab1b0 SL |
1522 | u64 bps_dft; |
1523 | unsigned int iops_dft; | |
ada75b6e | 1524 | char idle_time[26] = ""; |
ec80991d | 1525 | char latency_time[26] = ""; |
2ee867dc TH |
1526 | |
1527 | if (!dname) | |
1528 | return 0; | |
9f626e37 | 1529 | |
cd5ab1b0 SL |
1530 | if (off == LIMIT_LOW) { |
1531 | bps_dft = 0; | |
1532 | iops_dft = 0; | |
1533 | } else { | |
1534 | bps_dft = U64_MAX; | |
1535 | iops_dft = UINT_MAX; | |
1536 | } | |
1537 | ||
1538 | if (tg->bps_conf[READ][off] == bps_dft && | |
1539 | tg->bps_conf[WRITE][off] == bps_dft && | |
1540 | tg->iops_conf[READ][off] == iops_dft && | |
ada75b6e | 1541 | tg->iops_conf[WRITE][off] == iops_dft && |
ec80991d | 1542 | (off != LIMIT_LOW || |
b4f428ef | 1543 | (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD && |
5b81fc3c | 1544 | tg->latency_target_conf == DFL_LATENCY_TARGET))) |
2ee867dc TH |
1545 | return 0; |
1546 | ||
9bb67aeb | 1547 | if (tg->bps_conf[READ][off] != U64_MAX) |
9f626e37 | 1548 | snprintf(bufs[0], sizeof(bufs[0]), "%llu", |
cd5ab1b0 | 1549 | tg->bps_conf[READ][off]); |
9bb67aeb | 1550 | if (tg->bps_conf[WRITE][off] != U64_MAX) |
9f626e37 | 1551 | snprintf(bufs[1], sizeof(bufs[1]), "%llu", |
cd5ab1b0 | 1552 | tg->bps_conf[WRITE][off]); |
9bb67aeb | 1553 | if (tg->iops_conf[READ][off] != UINT_MAX) |
9f626e37 | 1554 | snprintf(bufs[2], sizeof(bufs[2]), "%u", |
cd5ab1b0 | 1555 | tg->iops_conf[READ][off]); |
9bb67aeb | 1556 | if (tg->iops_conf[WRITE][off] != UINT_MAX) |
9f626e37 | 1557 | snprintf(bufs[3], sizeof(bufs[3]), "%u", |
cd5ab1b0 | 1558 | tg->iops_conf[WRITE][off]); |
ada75b6e | 1559 | if (off == LIMIT_LOW) { |
5b81fc3c | 1560 | if (tg->idletime_threshold_conf == ULONG_MAX) |
ada75b6e SL |
1561 | strcpy(idle_time, " idle=max"); |
1562 | else | |
1563 | snprintf(idle_time, sizeof(idle_time), " idle=%lu", | |
5b81fc3c | 1564 | tg->idletime_threshold_conf); |
ec80991d | 1565 | |
5b81fc3c | 1566 | if (tg->latency_target_conf == ULONG_MAX) |
ec80991d SL |
1567 | strcpy(latency_time, " latency=max"); |
1568 | else | |
1569 | snprintf(latency_time, sizeof(latency_time), | |
5b81fc3c | 1570 | " latency=%lu", tg->latency_target_conf); |
ada75b6e | 1571 | } |
2ee867dc | 1572 | |
ec80991d SL |
1573 | seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n", |
1574 | dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time, | |
1575 | latency_time); | |
2ee867dc TH |
1576 | return 0; |
1577 | } | |
1578 | ||
cd5ab1b0 | 1579 | static int tg_print_limit(struct seq_file *sf, void *v) |
2ee867dc | 1580 | { |
cd5ab1b0 | 1581 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit, |
2ee867dc TH |
1582 | &blkcg_policy_throtl, seq_cft(sf)->private, false); |
1583 | return 0; | |
1584 | } | |
1585 | ||
cd5ab1b0 | 1586 | static ssize_t tg_set_limit(struct kernfs_open_file *of, |
2ee867dc TH |
1587 | char *buf, size_t nbytes, loff_t off) |
1588 | { | |
1589 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1590 | struct blkg_conf_ctx ctx; | |
1591 | struct throtl_grp *tg; | |
1592 | u64 v[4]; | |
ada75b6e | 1593 | unsigned long idle_time; |
ec80991d | 1594 | unsigned long latency_time; |
2ee867dc | 1595 | int ret; |
cd5ab1b0 | 1596 | int index = of_cft(of)->private; |
2ee867dc TH |
1597 | |
1598 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1599 | if (ret) | |
1600 | return ret; | |
1601 | ||
1602 | tg = blkg_to_tg(ctx.blkg); | |
1603 | ||
cd5ab1b0 SL |
1604 | v[0] = tg->bps_conf[READ][index]; |
1605 | v[1] = tg->bps_conf[WRITE][index]; | |
1606 | v[2] = tg->iops_conf[READ][index]; | |
1607 | v[3] = tg->iops_conf[WRITE][index]; | |
2ee867dc | 1608 | |
5b81fc3c SL |
1609 | idle_time = tg->idletime_threshold_conf; |
1610 | latency_time = tg->latency_target_conf; | |
2ee867dc TH |
1611 | while (true) { |
1612 | char tok[27]; /* wiops=18446744073709551616 */ | |
1613 | char *p; | |
2ab5492d | 1614 | u64 val = U64_MAX; |
2ee867dc TH |
1615 | int len; |
1616 | ||
1617 | if (sscanf(ctx.body, "%26s%n", tok, &len) != 1) | |
1618 | break; | |
1619 | if (tok[0] == '\0') | |
1620 | break; | |
1621 | ctx.body += len; | |
1622 | ||
1623 | ret = -EINVAL; | |
1624 | p = tok; | |
1625 | strsep(&p, "="); | |
1626 | if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max"))) | |
1627 | goto out_finish; | |
1628 | ||
1629 | ret = -ERANGE; | |
1630 | if (!val) | |
1631 | goto out_finish; | |
1632 | ||
1633 | ret = -EINVAL; | |
1634 | if (!strcmp(tok, "rbps")) | |
1635 | v[0] = val; | |
1636 | else if (!strcmp(tok, "wbps")) | |
1637 | v[1] = val; | |
1638 | else if (!strcmp(tok, "riops")) | |
1639 | v[2] = min_t(u64, val, UINT_MAX); | |
1640 | else if (!strcmp(tok, "wiops")) | |
1641 | v[3] = min_t(u64, val, UINT_MAX); | |
ada75b6e SL |
1642 | else if (off == LIMIT_LOW && !strcmp(tok, "idle")) |
1643 | idle_time = val; | |
ec80991d SL |
1644 | else if (off == LIMIT_LOW && !strcmp(tok, "latency")) |
1645 | latency_time = val; | |
2ee867dc TH |
1646 | else |
1647 | goto out_finish; | |
1648 | } | |
1649 | ||
cd5ab1b0 SL |
1650 | tg->bps_conf[READ][index] = v[0]; |
1651 | tg->bps_conf[WRITE][index] = v[1]; | |
1652 | tg->iops_conf[READ][index] = v[2]; | |
1653 | tg->iops_conf[WRITE][index] = v[3]; | |
2ee867dc | 1654 | |
cd5ab1b0 SL |
1655 | if (index == LIMIT_MAX) { |
1656 | tg->bps[READ][index] = v[0]; | |
1657 | tg->bps[WRITE][index] = v[1]; | |
1658 | tg->iops[READ][index] = v[2]; | |
1659 | tg->iops[WRITE][index] = v[3]; | |
1660 | } | |
1661 | tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW], | |
1662 | tg->bps_conf[READ][LIMIT_MAX]); | |
1663 | tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW], | |
1664 | tg->bps_conf[WRITE][LIMIT_MAX]); | |
1665 | tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW], | |
1666 | tg->iops_conf[READ][LIMIT_MAX]); | |
1667 | tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW], | |
1668 | tg->iops_conf[WRITE][LIMIT_MAX]); | |
b4f428ef SL |
1669 | tg->idletime_threshold_conf = idle_time; |
1670 | tg->latency_target_conf = latency_time; | |
1671 | ||
1672 | /* force user to configure all settings for low limit */ | |
1673 | if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] || | |
1674 | tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) || | |
1675 | tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD || | |
1676 | tg->latency_target_conf == DFL_LATENCY_TARGET) { | |
1677 | tg->bps[READ][LIMIT_LOW] = 0; | |
1678 | tg->bps[WRITE][LIMIT_LOW] = 0; | |
1679 | tg->iops[READ][LIMIT_LOW] = 0; | |
1680 | tg->iops[WRITE][LIMIT_LOW] = 0; | |
1681 | tg->idletime_threshold = DFL_IDLE_THRESHOLD; | |
1682 | tg->latency_target = DFL_LATENCY_TARGET; | |
1683 | } else if (index == LIMIT_LOW) { | |
5b81fc3c | 1684 | tg->idletime_threshold = tg->idletime_threshold_conf; |
5b81fc3c | 1685 | tg->latency_target = tg->latency_target_conf; |
cd5ab1b0 | 1686 | } |
b4f428ef SL |
1687 | |
1688 | blk_throtl_update_limit_valid(tg->td); | |
1689 | if (tg->td->limit_valid[LIMIT_LOW]) { | |
1690 | if (index == LIMIT_LOW) | |
1691 | tg->td->limit_index = LIMIT_LOW; | |
1692 | } else | |
1693 | tg->td->limit_index = LIMIT_MAX; | |
9bb67aeb SL |
1694 | tg_conf_updated(tg, index == LIMIT_LOW && |
1695 | tg->td->limit_valid[LIMIT_LOW]); | |
2ee867dc TH |
1696 | ret = 0; |
1697 | out_finish: | |
1698 | blkg_conf_finish(&ctx); | |
1699 | return ret ?: nbytes; | |
1700 | } | |
1701 | ||
1702 | static struct cftype throtl_files[] = { | |
cd5ab1b0 SL |
1703 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
1704 | { | |
1705 | .name = "low", | |
1706 | .flags = CFTYPE_NOT_ON_ROOT, | |
1707 | .seq_show = tg_print_limit, | |
1708 | .write = tg_set_limit, | |
1709 | .private = LIMIT_LOW, | |
1710 | }, | |
1711 | #endif | |
2ee867dc TH |
1712 | { |
1713 | .name = "max", | |
1714 | .flags = CFTYPE_NOT_ON_ROOT, | |
cd5ab1b0 SL |
1715 | .seq_show = tg_print_limit, |
1716 | .write = tg_set_limit, | |
1717 | .private = LIMIT_MAX, | |
2ee867dc TH |
1718 | }, |
1719 | { } /* terminate */ | |
1720 | }; | |
1721 | ||
da527770 | 1722 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1723 | { |
1724 | struct throtl_data *td = q->td; | |
1725 | ||
69df0ab0 | 1726 | cancel_work_sync(&td->dispatch_work); |
e43473b7 VG |
1727 | } |
1728 | ||
3c798398 | 1729 | static struct blkcg_policy blkcg_policy_throtl = { |
2ee867dc | 1730 | .dfl_cftypes = throtl_files, |
880f50e2 | 1731 | .legacy_cftypes = throtl_legacy_files, |
f9fcc2d3 | 1732 | |
001bea73 | 1733 | .pd_alloc_fn = throtl_pd_alloc, |
f9fcc2d3 | 1734 | .pd_init_fn = throtl_pd_init, |
693e751e | 1735 | .pd_online_fn = throtl_pd_online, |
cd5ab1b0 | 1736 | .pd_offline_fn = throtl_pd_offline, |
001bea73 | 1737 | .pd_free_fn = throtl_pd_free, |
e43473b7 VG |
1738 | }; |
1739 | ||
3f0abd80 SL |
1740 | static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg) |
1741 | { | |
1742 | unsigned long rtime = jiffies, wtime = jiffies; | |
1743 | ||
1744 | if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]) | |
1745 | rtime = tg->last_low_overflow_time[READ]; | |
1746 | if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) | |
1747 | wtime = tg->last_low_overflow_time[WRITE]; | |
1748 | return min(rtime, wtime); | |
1749 | } | |
1750 | ||
1751 | /* tg should not be an intermediate node */ | |
1752 | static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg) | |
1753 | { | |
1754 | struct throtl_service_queue *parent_sq; | |
1755 | struct throtl_grp *parent = tg; | |
1756 | unsigned long ret = __tg_last_low_overflow_time(tg); | |
1757 | ||
1758 | while (true) { | |
1759 | parent_sq = parent->service_queue.parent_sq; | |
1760 | parent = sq_to_tg(parent_sq); | |
1761 | if (!parent) | |
1762 | break; | |
1763 | ||
1764 | /* | |
1765 | * The parent doesn't have low limit, it always reaches low | |
1766 | * limit. Its overflow time is useless for children | |
1767 | */ | |
1768 | if (!parent->bps[READ][LIMIT_LOW] && | |
1769 | !parent->iops[READ][LIMIT_LOW] && | |
1770 | !parent->bps[WRITE][LIMIT_LOW] && | |
1771 | !parent->iops[WRITE][LIMIT_LOW]) | |
1772 | continue; | |
1773 | if (time_after(__tg_last_low_overflow_time(parent), ret)) | |
1774 | ret = __tg_last_low_overflow_time(parent); | |
1775 | } | |
1776 | return ret; | |
1777 | } | |
1778 | ||
9e234eea SL |
1779 | static bool throtl_tg_is_idle(struct throtl_grp *tg) |
1780 | { | |
1781 | /* | |
1782 | * cgroup is idle if: | |
1783 | * - single idle is too long, longer than a fixed value (in case user | |
b4f428ef | 1784 | * configure a too big threshold) or 4 times of idletime threshold |
9e234eea | 1785 | * - average think time is more than threshold |
53696b8d | 1786 | * - IO latency is largely below threshold |
9e234eea | 1787 | */ |
b4f428ef | 1788 | unsigned long time; |
4cff729f | 1789 | bool ret; |
9e234eea | 1790 | |
b4f428ef SL |
1791 | time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold); |
1792 | ret = tg->latency_target == DFL_LATENCY_TARGET || | |
1793 | tg->idletime_threshold == DFL_IDLE_THRESHOLD || | |
1794 | (ktime_get_ns() >> 10) - tg->last_finish_time > time || | |
1795 | tg->avg_idletime > tg->idletime_threshold || | |
1796 | (tg->latency_target && tg->bio_cnt && | |
53696b8d | 1797 | tg->bad_bio_cnt * 5 < tg->bio_cnt); |
4cff729f SL |
1798 | throtl_log(&tg->service_queue, |
1799 | "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d", | |
1800 | tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt, | |
1801 | tg->bio_cnt, ret, tg->td->scale); | |
1802 | return ret; | |
9e234eea SL |
1803 | } |
1804 | ||
c79892c5 SL |
1805 | static bool throtl_tg_can_upgrade(struct throtl_grp *tg) |
1806 | { | |
1807 | struct throtl_service_queue *sq = &tg->service_queue; | |
1808 | bool read_limit, write_limit; | |
1809 | ||
1810 | /* | |
1811 | * if cgroup reaches low limit (if low limit is 0, the cgroup always | |
1812 | * reaches), it's ok to upgrade to next limit | |
1813 | */ | |
1814 | read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]; | |
1815 | write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]; | |
1816 | if (!read_limit && !write_limit) | |
1817 | return true; | |
1818 | if (read_limit && sq->nr_queued[READ] && | |
1819 | (!write_limit || sq->nr_queued[WRITE])) | |
1820 | return true; | |
1821 | if (write_limit && sq->nr_queued[WRITE] && | |
1822 | (!read_limit || sq->nr_queued[READ])) | |
1823 | return true; | |
aec24246 SL |
1824 | |
1825 | if (time_after_eq(jiffies, | |
fa6fb5aa SL |
1826 | tg_last_low_overflow_time(tg) + tg->td->throtl_slice) && |
1827 | throtl_tg_is_idle(tg)) | |
aec24246 | 1828 | return true; |
c79892c5 SL |
1829 | return false; |
1830 | } | |
1831 | ||
1832 | static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg) | |
1833 | { | |
1834 | while (true) { | |
1835 | if (throtl_tg_can_upgrade(tg)) | |
1836 | return true; | |
1837 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1838 | if (!tg || !tg_to_blkg(tg)->parent) | |
1839 | return false; | |
1840 | } | |
1841 | return false; | |
1842 | } | |
1843 | ||
1844 | static bool throtl_can_upgrade(struct throtl_data *td, | |
1845 | struct throtl_grp *this_tg) | |
1846 | { | |
1847 | struct cgroup_subsys_state *pos_css; | |
1848 | struct blkcg_gq *blkg; | |
1849 | ||
1850 | if (td->limit_index != LIMIT_LOW) | |
1851 | return false; | |
1852 | ||
297e3d85 | 1853 | if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice)) |
3f0abd80 SL |
1854 | return false; |
1855 | ||
c79892c5 SL |
1856 | rcu_read_lock(); |
1857 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1858 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1859 | ||
1860 | if (tg == this_tg) | |
1861 | continue; | |
1862 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1863 | continue; | |
1864 | if (!throtl_hierarchy_can_upgrade(tg)) { | |
1865 | rcu_read_unlock(); | |
1866 | return false; | |
1867 | } | |
1868 | } | |
1869 | rcu_read_unlock(); | |
1870 | return true; | |
1871 | } | |
1872 | ||
fa6fb5aa SL |
1873 | static void throtl_upgrade_check(struct throtl_grp *tg) |
1874 | { | |
1875 | unsigned long now = jiffies; | |
1876 | ||
1877 | if (tg->td->limit_index != LIMIT_LOW) | |
1878 | return; | |
1879 | ||
1880 | if (time_after(tg->last_check_time + tg->td->throtl_slice, now)) | |
1881 | return; | |
1882 | ||
1883 | tg->last_check_time = now; | |
1884 | ||
1885 | if (!time_after_eq(now, | |
1886 | __tg_last_low_overflow_time(tg) + tg->td->throtl_slice)) | |
1887 | return; | |
1888 | ||
1889 | if (throtl_can_upgrade(tg->td, NULL)) | |
1890 | throtl_upgrade_state(tg->td); | |
1891 | } | |
1892 | ||
c79892c5 SL |
1893 | static void throtl_upgrade_state(struct throtl_data *td) |
1894 | { | |
1895 | struct cgroup_subsys_state *pos_css; | |
1896 | struct blkcg_gq *blkg; | |
1897 | ||
4cff729f | 1898 | throtl_log(&td->service_queue, "upgrade to max"); |
c79892c5 | 1899 | td->limit_index = LIMIT_MAX; |
3f0abd80 | 1900 | td->low_upgrade_time = jiffies; |
7394e31f | 1901 | td->scale = 0; |
c79892c5 SL |
1902 | rcu_read_lock(); |
1903 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1904 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1905 | struct throtl_service_queue *sq = &tg->service_queue; | |
1906 | ||
1907 | tg->disptime = jiffies - 1; | |
1908 | throtl_select_dispatch(sq); | |
4f02fb76 | 1909 | throtl_schedule_next_dispatch(sq, true); |
c79892c5 SL |
1910 | } |
1911 | rcu_read_unlock(); | |
1912 | throtl_select_dispatch(&td->service_queue); | |
4f02fb76 | 1913 | throtl_schedule_next_dispatch(&td->service_queue, true); |
c79892c5 SL |
1914 | queue_work(kthrotld_workqueue, &td->dispatch_work); |
1915 | } | |
1916 | ||
3f0abd80 SL |
1917 | static void throtl_downgrade_state(struct throtl_data *td, int new) |
1918 | { | |
7394e31f SL |
1919 | td->scale /= 2; |
1920 | ||
4cff729f | 1921 | throtl_log(&td->service_queue, "downgrade, scale %d", td->scale); |
7394e31f SL |
1922 | if (td->scale) { |
1923 | td->low_upgrade_time = jiffies - td->scale * td->throtl_slice; | |
1924 | return; | |
1925 | } | |
1926 | ||
3f0abd80 SL |
1927 | td->limit_index = new; |
1928 | td->low_downgrade_time = jiffies; | |
1929 | } | |
1930 | ||
1931 | static bool throtl_tg_can_downgrade(struct throtl_grp *tg) | |
1932 | { | |
1933 | struct throtl_data *td = tg->td; | |
1934 | unsigned long now = jiffies; | |
1935 | ||
1936 | /* | |
1937 | * If cgroup is below low limit, consider downgrade and throttle other | |
1938 | * cgroups | |
1939 | */ | |
297e3d85 SL |
1940 | if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) && |
1941 | time_after_eq(now, tg_last_low_overflow_time(tg) + | |
fa6fb5aa SL |
1942 | td->throtl_slice) && |
1943 | (!throtl_tg_is_idle(tg) || | |
1944 | !list_empty(&tg_to_blkg(tg)->blkcg->css.children))) | |
3f0abd80 SL |
1945 | return true; |
1946 | return false; | |
1947 | } | |
1948 | ||
1949 | static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg) | |
1950 | { | |
1951 | while (true) { | |
1952 | if (!throtl_tg_can_downgrade(tg)) | |
1953 | return false; | |
1954 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1955 | if (!tg || !tg_to_blkg(tg)->parent) | |
1956 | break; | |
1957 | } | |
1958 | return true; | |
1959 | } | |
1960 | ||
1961 | static void throtl_downgrade_check(struct throtl_grp *tg) | |
1962 | { | |
1963 | uint64_t bps; | |
1964 | unsigned int iops; | |
1965 | unsigned long elapsed_time; | |
1966 | unsigned long now = jiffies; | |
1967 | ||
1968 | if (tg->td->limit_index != LIMIT_MAX || | |
1969 | !tg->td->limit_valid[LIMIT_LOW]) | |
1970 | return; | |
1971 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1972 | return; | |
297e3d85 | 1973 | if (time_after(tg->last_check_time + tg->td->throtl_slice, now)) |
3f0abd80 SL |
1974 | return; |
1975 | ||
1976 | elapsed_time = now - tg->last_check_time; | |
1977 | tg->last_check_time = now; | |
1978 | ||
297e3d85 SL |
1979 | if (time_before(now, tg_last_low_overflow_time(tg) + |
1980 | tg->td->throtl_slice)) | |
3f0abd80 SL |
1981 | return; |
1982 | ||
1983 | if (tg->bps[READ][LIMIT_LOW]) { | |
1984 | bps = tg->last_bytes_disp[READ] * HZ; | |
1985 | do_div(bps, elapsed_time); | |
1986 | if (bps >= tg->bps[READ][LIMIT_LOW]) | |
1987 | tg->last_low_overflow_time[READ] = now; | |
1988 | } | |
1989 | ||
1990 | if (tg->bps[WRITE][LIMIT_LOW]) { | |
1991 | bps = tg->last_bytes_disp[WRITE] * HZ; | |
1992 | do_div(bps, elapsed_time); | |
1993 | if (bps >= tg->bps[WRITE][LIMIT_LOW]) | |
1994 | tg->last_low_overflow_time[WRITE] = now; | |
1995 | } | |
1996 | ||
1997 | if (tg->iops[READ][LIMIT_LOW]) { | |
1998 | iops = tg->last_io_disp[READ] * HZ / elapsed_time; | |
1999 | if (iops >= tg->iops[READ][LIMIT_LOW]) | |
2000 | tg->last_low_overflow_time[READ] = now; | |
2001 | } | |
2002 | ||
2003 | if (tg->iops[WRITE][LIMIT_LOW]) { | |
2004 | iops = tg->last_io_disp[WRITE] * HZ / elapsed_time; | |
2005 | if (iops >= tg->iops[WRITE][LIMIT_LOW]) | |
2006 | tg->last_low_overflow_time[WRITE] = now; | |
2007 | } | |
2008 | ||
2009 | /* | |
2010 | * If cgroup is below low limit, consider downgrade and throttle other | |
2011 | * cgroups | |
2012 | */ | |
2013 | if (throtl_hierarchy_can_downgrade(tg)) | |
2014 | throtl_downgrade_state(tg->td, LIMIT_LOW); | |
2015 | ||
2016 | tg->last_bytes_disp[READ] = 0; | |
2017 | tg->last_bytes_disp[WRITE] = 0; | |
2018 | tg->last_io_disp[READ] = 0; | |
2019 | tg->last_io_disp[WRITE] = 0; | |
2020 | } | |
2021 | ||
9e234eea SL |
2022 | static void blk_throtl_update_idletime(struct throtl_grp *tg) |
2023 | { | |
2024 | unsigned long now = ktime_get_ns() >> 10; | |
2025 | unsigned long last_finish_time = tg->last_finish_time; | |
2026 | ||
2027 | if (now <= last_finish_time || last_finish_time == 0 || | |
2028 | last_finish_time == tg->checked_last_finish_time) | |
2029 | return; | |
2030 | ||
2031 | tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3; | |
2032 | tg->checked_last_finish_time = last_finish_time; | |
2033 | } | |
2034 | ||
b9147dd1 SL |
2035 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
2036 | static void throtl_update_latency_buckets(struct throtl_data *td) | |
2037 | { | |
b889bf66 JQ |
2038 | struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE]; |
2039 | int i, cpu, rw; | |
2040 | unsigned long last_latency[2] = { 0 }; | |
2041 | unsigned long latency[2]; | |
b9147dd1 SL |
2042 | |
2043 | if (!blk_queue_nonrot(td->queue)) | |
2044 | return; | |
2045 | if (time_before(jiffies, td->last_calculate_time + HZ)) | |
2046 | return; | |
2047 | td->last_calculate_time = jiffies; | |
2048 | ||
2049 | memset(avg_latency, 0, sizeof(avg_latency)); | |
b889bf66 JQ |
2050 | for (rw = READ; rw <= WRITE; rw++) { |
2051 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) { | |
2052 | struct latency_bucket *tmp = &td->tmp_buckets[rw][i]; | |
2053 | ||
2054 | for_each_possible_cpu(cpu) { | |
2055 | struct latency_bucket *bucket; | |
2056 | ||
2057 | /* this isn't race free, but ok in practice */ | |
2058 | bucket = per_cpu_ptr(td->latency_buckets[rw], | |
2059 | cpu); | |
2060 | tmp->total_latency += bucket[i].total_latency; | |
2061 | tmp->samples += bucket[i].samples; | |
2062 | bucket[i].total_latency = 0; | |
2063 | bucket[i].samples = 0; | |
2064 | } | |
b9147dd1 | 2065 | |
b889bf66 JQ |
2066 | if (tmp->samples >= 32) { |
2067 | int samples = tmp->samples; | |
b9147dd1 | 2068 | |
b889bf66 | 2069 | latency[rw] = tmp->total_latency; |
b9147dd1 | 2070 | |
b889bf66 JQ |
2071 | tmp->total_latency = 0; |
2072 | tmp->samples = 0; | |
2073 | latency[rw] /= samples; | |
2074 | if (latency[rw] == 0) | |
2075 | continue; | |
2076 | avg_latency[rw][i].latency = latency[rw]; | |
2077 | } | |
b9147dd1 SL |
2078 | } |
2079 | } | |
2080 | ||
b889bf66 JQ |
2081 | for (rw = READ; rw <= WRITE; rw++) { |
2082 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) { | |
2083 | if (!avg_latency[rw][i].latency) { | |
2084 | if (td->avg_buckets[rw][i].latency < last_latency[rw]) | |
2085 | td->avg_buckets[rw][i].latency = | |
2086 | last_latency[rw]; | |
2087 | continue; | |
2088 | } | |
b9147dd1 | 2089 | |
b889bf66 JQ |
2090 | if (!td->avg_buckets[rw][i].valid) |
2091 | latency[rw] = avg_latency[rw][i].latency; | |
2092 | else | |
2093 | latency[rw] = (td->avg_buckets[rw][i].latency * 7 + | |
2094 | avg_latency[rw][i].latency) >> 3; | |
b9147dd1 | 2095 | |
b889bf66 JQ |
2096 | td->avg_buckets[rw][i].latency = max(latency[rw], |
2097 | last_latency[rw]); | |
2098 | td->avg_buckets[rw][i].valid = true; | |
2099 | last_latency[rw] = td->avg_buckets[rw][i].latency; | |
2100 | } | |
b9147dd1 | 2101 | } |
4cff729f SL |
2102 | |
2103 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) | |
2104 | throtl_log(&td->service_queue, | |
b889bf66 JQ |
2105 | "Latency bucket %d: read latency=%ld, read valid=%d, " |
2106 | "write latency=%ld, write valid=%d", i, | |
2107 | td->avg_buckets[READ][i].latency, | |
2108 | td->avg_buckets[READ][i].valid, | |
2109 | td->avg_buckets[WRITE][i].latency, | |
2110 | td->avg_buckets[WRITE][i].valid); | |
b9147dd1 SL |
2111 | } |
2112 | #else | |
2113 | static inline void throtl_update_latency_buckets(struct throtl_data *td) | |
2114 | { | |
2115 | } | |
2116 | #endif | |
2117 | ||
ae118896 TH |
2118 | bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg, |
2119 | struct bio *bio) | |
e43473b7 | 2120 | { |
c5cc2070 | 2121 | struct throtl_qnode *qn = NULL; |
a7b39b4e | 2122 | struct throtl_grp *tg = blkg_to_tg(blkg); |
73f0d49a | 2123 | struct throtl_service_queue *sq; |
0e9f4164 | 2124 | bool rw = bio_data_dir(bio); |
bc16a4f9 | 2125 | bool throttled = false; |
b9147dd1 | 2126 | struct throtl_data *td = tg->td; |
e43473b7 | 2127 | |
ae118896 TH |
2128 | WARN_ON_ONCE(!rcu_read_lock_held()); |
2129 | ||
2a0f61e6 | 2130 | /* see throtl_charge_bio() */ |
8d2bbd4c | 2131 | if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw]) |
bc16a4f9 | 2132 | goto out; |
e43473b7 VG |
2133 | |
2134 | spin_lock_irq(q->queue_lock); | |
c9589f03 | 2135 | |
b9147dd1 SL |
2136 | throtl_update_latency_buckets(td); |
2137 | ||
c9589f03 | 2138 | if (unlikely(blk_queue_bypass(q))) |
bc16a4f9 | 2139 | goto out_unlock; |
f469a7b4 | 2140 | |
9e234eea SL |
2141 | blk_throtl_update_idletime(tg); |
2142 | ||
73f0d49a TH |
2143 | sq = &tg->service_queue; |
2144 | ||
c79892c5 | 2145 | again: |
9e660acf | 2146 | while (true) { |
3f0abd80 SL |
2147 | if (tg->last_low_overflow_time[rw] == 0) |
2148 | tg->last_low_overflow_time[rw] = jiffies; | |
2149 | throtl_downgrade_check(tg); | |
fa6fb5aa | 2150 | throtl_upgrade_check(tg); |
9e660acf TH |
2151 | /* throtl is FIFO - if bios are already queued, should queue */ |
2152 | if (sq->nr_queued[rw]) | |
2153 | break; | |
de701c74 | 2154 | |
9e660acf | 2155 | /* if above limits, break to queue */ |
c79892c5 | 2156 | if (!tg_may_dispatch(tg, bio, NULL)) { |
3f0abd80 | 2157 | tg->last_low_overflow_time[rw] = jiffies; |
b9147dd1 SL |
2158 | if (throtl_can_upgrade(td, tg)) { |
2159 | throtl_upgrade_state(td); | |
c79892c5 SL |
2160 | goto again; |
2161 | } | |
9e660acf | 2162 | break; |
c79892c5 | 2163 | } |
9e660acf TH |
2164 | |
2165 | /* within limits, let's charge and dispatch directly */ | |
e43473b7 | 2166 | throtl_charge_bio(tg, bio); |
04521db0 VG |
2167 | |
2168 | /* | |
2169 | * We need to trim slice even when bios are not being queued | |
2170 | * otherwise it might happen that a bio is not queued for | |
2171 | * a long time and slice keeps on extending and trim is not | |
2172 | * called for a long time. Now if limits are reduced suddenly | |
2173 | * we take into account all the IO dispatched so far at new | |
2174 | * low rate and * newly queued IO gets a really long dispatch | |
2175 | * time. | |
2176 | * | |
2177 | * So keep on trimming slice even if bio is not queued. | |
2178 | */ | |
0f3457f6 | 2179 | throtl_trim_slice(tg, rw); |
9e660acf TH |
2180 | |
2181 | /* | |
2182 | * @bio passed through this layer without being throttled. | |
2183 | * Climb up the ladder. If we''re already at the top, it | |
2184 | * can be executed directly. | |
2185 | */ | |
c5cc2070 | 2186 | qn = &tg->qnode_on_parent[rw]; |
9e660acf TH |
2187 | sq = sq->parent_sq; |
2188 | tg = sq_to_tg(sq); | |
2189 | if (!tg) | |
2190 | goto out_unlock; | |
e43473b7 VG |
2191 | } |
2192 | ||
9e660acf | 2193 | /* out-of-limit, queue to @tg */ |
fda6f272 TH |
2194 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
2195 | rw == READ ? 'R' : 'W', | |
9f626e37 SL |
2196 | tg->bytes_disp[rw], bio->bi_iter.bi_size, |
2197 | tg_bps_limit(tg, rw), | |
2198 | tg->io_disp[rw], tg_iops_limit(tg, rw), | |
fda6f272 | 2199 | sq->nr_queued[READ], sq->nr_queued[WRITE]); |
e43473b7 | 2200 | |
3f0abd80 SL |
2201 | tg->last_low_overflow_time[rw] = jiffies; |
2202 | ||
b9147dd1 | 2203 | td->nr_queued[rw]++; |
c5cc2070 | 2204 | throtl_add_bio_tg(bio, qn, tg); |
bc16a4f9 | 2205 | throttled = true; |
e43473b7 | 2206 | |
7f52f98c TH |
2207 | /* |
2208 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
2209 | * was empty before @bio. The forced scheduling isn't likely to | |
2210 | * cause undue delay as @bio is likely to be dispatched directly if | |
2211 | * its @tg's disptime is not in the future. | |
2212 | */ | |
0e9f4164 | 2213 | if (tg->flags & THROTL_TG_WAS_EMPTY) { |
77216b04 | 2214 | tg_update_disptime(tg); |
7f52f98c | 2215 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); |
e43473b7 VG |
2216 | } |
2217 | ||
bc16a4f9 | 2218 | out_unlock: |
e43473b7 | 2219 | spin_unlock_irq(q->queue_lock); |
bc16a4f9 | 2220 | out: |
111be883 | 2221 | bio_set_flag(bio, BIO_THROTTLED); |
b9147dd1 SL |
2222 | |
2223 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
2224 | if (throttled || !td->track_bio_latency) | |
5238dcf4 | 2225 | bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY; |
b9147dd1 | 2226 | #endif |
bc16a4f9 | 2227 | return throttled; |
e43473b7 VG |
2228 | } |
2229 | ||
9e234eea | 2230 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
b9147dd1 SL |
2231 | static void throtl_track_latency(struct throtl_data *td, sector_t size, |
2232 | int op, unsigned long time) | |
2233 | { | |
2234 | struct latency_bucket *latency; | |
2235 | int index; | |
2236 | ||
b889bf66 JQ |
2237 | if (!td || td->limit_index != LIMIT_LOW || |
2238 | !(op == REQ_OP_READ || op == REQ_OP_WRITE) || | |
b9147dd1 SL |
2239 | !blk_queue_nonrot(td->queue)) |
2240 | return; | |
2241 | ||
2242 | index = request_bucket_index(size); | |
2243 | ||
b889bf66 | 2244 | latency = get_cpu_ptr(td->latency_buckets[op]); |
b9147dd1 SL |
2245 | latency[index].total_latency += time; |
2246 | latency[index].samples++; | |
b889bf66 | 2247 | put_cpu_ptr(td->latency_buckets[op]); |
b9147dd1 SL |
2248 | } |
2249 | ||
2250 | void blk_throtl_stat_add(struct request *rq, u64 time_ns) | |
2251 | { | |
2252 | struct request_queue *q = rq->q; | |
2253 | struct throtl_data *td = q->td; | |
2254 | ||
544ccc8d | 2255 | throtl_track_latency(td, rq->throtl_size, req_op(rq), time_ns >> 10); |
b9147dd1 SL |
2256 | } |
2257 | ||
9e234eea SL |
2258 | void blk_throtl_bio_endio(struct bio *bio) |
2259 | { | |
08e18eab | 2260 | struct blkcg_gq *blkg; |
9e234eea | 2261 | struct throtl_grp *tg; |
b9147dd1 SL |
2262 | u64 finish_time_ns; |
2263 | unsigned long finish_time; | |
2264 | unsigned long start_time; | |
2265 | unsigned long lat; | |
b889bf66 | 2266 | int rw = bio_data_dir(bio); |
9e234eea | 2267 | |
08e18eab JB |
2268 | blkg = bio->bi_blkg; |
2269 | if (!blkg) | |
9e234eea | 2270 | return; |
08e18eab | 2271 | tg = blkg_to_tg(blkg); |
9e234eea | 2272 | |
b9147dd1 SL |
2273 | finish_time_ns = ktime_get_ns(); |
2274 | tg->last_finish_time = finish_time_ns >> 10; | |
2275 | ||
5238dcf4 OS |
2276 | start_time = bio_issue_time(&bio->bi_issue) >> 10; |
2277 | finish_time = __bio_issue_time(finish_time_ns) >> 10; | |
08e18eab | 2278 | if (!start_time || finish_time <= start_time) |
53696b8d SL |
2279 | return; |
2280 | ||
2281 | lat = finish_time - start_time; | |
b9147dd1 | 2282 | /* this is only for bio based driver */ |
5238dcf4 OS |
2283 | if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY)) |
2284 | throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue), | |
2285 | bio_op(bio), lat); | |
53696b8d | 2286 | |
6679a90c | 2287 | if (tg->latency_target && lat >= tg->td->filtered_latency) { |
53696b8d SL |
2288 | int bucket; |
2289 | unsigned int threshold; | |
2290 | ||
5238dcf4 | 2291 | bucket = request_bucket_index(bio_issue_size(&bio->bi_issue)); |
b889bf66 | 2292 | threshold = tg->td->avg_buckets[rw][bucket].latency + |
53696b8d SL |
2293 | tg->latency_target; |
2294 | if (lat > threshold) | |
2295 | tg->bad_bio_cnt++; | |
2296 | /* | |
2297 | * Not race free, could get wrong count, which means cgroups | |
2298 | * will be throttled | |
2299 | */ | |
2300 | tg->bio_cnt++; | |
2301 | } | |
2302 | ||
2303 | if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) { | |
2304 | tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies; | |
2305 | tg->bio_cnt /= 2; | |
2306 | tg->bad_bio_cnt /= 2; | |
b9147dd1 | 2307 | } |
9e234eea SL |
2308 | } |
2309 | #endif | |
2310 | ||
2a12f0dc TH |
2311 | /* |
2312 | * Dispatch all bios from all children tg's queued on @parent_sq. On | |
2313 | * return, @parent_sq is guaranteed to not have any active children tg's | |
2314 | * and all bios from previously active tg's are on @parent_sq->bio_lists[]. | |
2315 | */ | |
2316 | static void tg_drain_bios(struct throtl_service_queue *parent_sq) | |
2317 | { | |
2318 | struct throtl_grp *tg; | |
2319 | ||
2320 | while ((tg = throtl_rb_first(parent_sq))) { | |
2321 | struct throtl_service_queue *sq = &tg->service_queue; | |
2322 | struct bio *bio; | |
2323 | ||
2324 | throtl_dequeue_tg(tg); | |
2325 | ||
c5cc2070 | 2326 | while ((bio = throtl_peek_queued(&sq->queued[READ]))) |
2a12f0dc | 2327 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
c5cc2070 | 2328 | while ((bio = throtl_peek_queued(&sq->queued[WRITE]))) |
2a12f0dc TH |
2329 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
2330 | } | |
2331 | } | |
2332 | ||
c9a929dd TH |
2333 | /** |
2334 | * blk_throtl_drain - drain throttled bios | |
2335 | * @q: request_queue to drain throttled bios for | |
2336 | * | |
2337 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
2338 | */ | |
2339 | void blk_throtl_drain(struct request_queue *q) | |
2340 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
2341 | { | |
2342 | struct throtl_data *td = q->td; | |
2a12f0dc | 2343 | struct blkcg_gq *blkg; |
492eb21b | 2344 | struct cgroup_subsys_state *pos_css; |
c9a929dd | 2345 | struct bio *bio; |
651930bc | 2346 | int rw; |
c9a929dd | 2347 | |
8bcb6c7d | 2348 | queue_lockdep_assert_held(q); |
2a12f0dc | 2349 | rcu_read_lock(); |
c9a929dd | 2350 | |
2a12f0dc TH |
2351 | /* |
2352 | * Drain each tg while doing post-order walk on the blkg tree, so | |
2353 | * that all bios are propagated to td->service_queue. It'd be | |
2354 | * better to walk service_queue tree directly but blkg walk is | |
2355 | * easier. | |
2356 | */ | |
492eb21b | 2357 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) |
2a12f0dc | 2358 | tg_drain_bios(&blkg_to_tg(blkg)->service_queue); |
73f0d49a | 2359 | |
2a12f0dc TH |
2360 | /* finally, transfer bios from top-level tg's into the td */ |
2361 | tg_drain_bios(&td->service_queue); | |
2362 | ||
2363 | rcu_read_unlock(); | |
c9a929dd TH |
2364 | spin_unlock_irq(q->queue_lock); |
2365 | ||
2a12f0dc | 2366 | /* all bios now should be in td->service_queue, issue them */ |
651930bc | 2367 | for (rw = READ; rw <= WRITE; rw++) |
c5cc2070 TH |
2368 | while ((bio = throtl_pop_queued(&td->service_queue.queued[rw], |
2369 | NULL))) | |
651930bc | 2370 | generic_make_request(bio); |
c9a929dd TH |
2371 | |
2372 | spin_lock_irq(q->queue_lock); | |
2373 | } | |
2374 | ||
e43473b7 VG |
2375 | int blk_throtl_init(struct request_queue *q) |
2376 | { | |
2377 | struct throtl_data *td; | |
a2b1693b | 2378 | int ret; |
e43473b7 VG |
2379 | |
2380 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
2381 | if (!td) | |
2382 | return -ENOMEM; | |
b889bf66 | 2383 | td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) * |
b9147dd1 | 2384 | LATENCY_BUCKET_SIZE, __alignof__(u64)); |
b889bf66 JQ |
2385 | if (!td->latency_buckets[READ]) { |
2386 | kfree(td); | |
2387 | return -ENOMEM; | |
2388 | } | |
2389 | td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) * | |
b9147dd1 | 2390 | LATENCY_BUCKET_SIZE, __alignof__(u64)); |
b889bf66 JQ |
2391 | if (!td->latency_buckets[WRITE]) { |
2392 | free_percpu(td->latency_buckets[READ]); | |
b9147dd1 SL |
2393 | kfree(td); |
2394 | return -ENOMEM; | |
2395 | } | |
e43473b7 | 2396 | |
69df0ab0 | 2397 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
b2ce2643 | 2398 | throtl_service_queue_init(&td->service_queue); |
e43473b7 | 2399 | |
cd1604fa | 2400 | q->td = td; |
29b12589 | 2401 | td->queue = q; |
02977e4a | 2402 | |
9f626e37 | 2403 | td->limit_valid[LIMIT_MAX] = true; |
cd5ab1b0 | 2404 | td->limit_index = LIMIT_MAX; |
3f0abd80 SL |
2405 | td->low_upgrade_time = jiffies; |
2406 | td->low_downgrade_time = jiffies; | |
9e234eea | 2407 | |
a2b1693b | 2408 | /* activate policy */ |
3c798398 | 2409 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
b9147dd1 | 2410 | if (ret) { |
b889bf66 JQ |
2411 | free_percpu(td->latency_buckets[READ]); |
2412 | free_percpu(td->latency_buckets[WRITE]); | |
f51b802c | 2413 | kfree(td); |
b9147dd1 | 2414 | } |
a2b1693b | 2415 | return ret; |
e43473b7 VG |
2416 | } |
2417 | ||
2418 | void blk_throtl_exit(struct request_queue *q) | |
2419 | { | |
c875f4d0 | 2420 | BUG_ON(!q->td); |
da527770 | 2421 | throtl_shutdown_wq(q); |
3c798398 | 2422 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
b889bf66 JQ |
2423 | free_percpu(q->td->latency_buckets[READ]); |
2424 | free_percpu(q->td->latency_buckets[WRITE]); | |
c9a929dd | 2425 | kfree(q->td); |
e43473b7 VG |
2426 | } |
2427 | ||
d61fcfa4 SL |
2428 | void blk_throtl_register_queue(struct request_queue *q) |
2429 | { | |
2430 | struct throtl_data *td; | |
6679a90c | 2431 | int i; |
d61fcfa4 SL |
2432 | |
2433 | td = q->td; | |
2434 | BUG_ON(!td); | |
2435 | ||
6679a90c | 2436 | if (blk_queue_nonrot(q)) { |
d61fcfa4 | 2437 | td->throtl_slice = DFL_THROTL_SLICE_SSD; |
6679a90c SL |
2438 | td->filtered_latency = LATENCY_FILTERED_SSD; |
2439 | } else { | |
d61fcfa4 | 2440 | td->throtl_slice = DFL_THROTL_SLICE_HD; |
6679a90c | 2441 | td->filtered_latency = LATENCY_FILTERED_HD; |
b889bf66 JQ |
2442 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) { |
2443 | td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY; | |
2444 | td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY; | |
2445 | } | |
6679a90c | 2446 | } |
d61fcfa4 SL |
2447 | #ifndef CONFIG_BLK_DEV_THROTTLING_LOW |
2448 | /* if no low limit, use previous default */ | |
2449 | td->throtl_slice = DFL_THROTL_SLICE_HD; | |
2450 | #endif | |
9e234eea | 2451 | |
475a055e | 2452 | td->track_bio_latency = !queue_is_rq_based(q); |
b9147dd1 SL |
2453 | if (!td->track_bio_latency) |
2454 | blk_stat_enable_accounting(q); | |
d61fcfa4 SL |
2455 | } |
2456 | ||
297e3d85 SL |
2457 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
2458 | ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page) | |
2459 | { | |
2460 | if (!q->td) | |
2461 | return -EINVAL; | |
2462 | return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice)); | |
2463 | } | |
2464 | ||
2465 | ssize_t blk_throtl_sample_time_store(struct request_queue *q, | |
2466 | const char *page, size_t count) | |
2467 | { | |
2468 | unsigned long v; | |
2469 | unsigned long t; | |
2470 | ||
2471 | if (!q->td) | |
2472 | return -EINVAL; | |
2473 | if (kstrtoul(page, 10, &v)) | |
2474 | return -EINVAL; | |
2475 | t = msecs_to_jiffies(v); | |
2476 | if (t == 0 || t > MAX_THROTL_SLICE) | |
2477 | return -EINVAL; | |
2478 | q->td->throtl_slice = t; | |
2479 | return count; | |
2480 | } | |
2481 | #endif | |
2482 | ||
e43473b7 VG |
2483 | static int __init throtl_init(void) |
2484 | { | |
450adcbe VG |
2485 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
2486 | if (!kthrotld_workqueue) | |
2487 | panic("Failed to create kthrotld\n"); | |
2488 | ||
3c798398 | 2489 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
2490 | } |
2491 | ||
2492 | module_init(throtl_init); |