]>
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" |
1d156646 | 15 | #include "blk-cgroup-rwstat.h" |
e43473b7 VG |
16 | |
17 | /* Max dispatch from a group in 1 round */ | |
e675df2a | 18 | #define THROTL_GRP_QUANTUM 8 |
e43473b7 VG |
19 | |
20 | /* Total max dispatch from all groups in one round */ | |
e675df2a | 21 | #define THROTL_QUANTUM 32 |
e43473b7 | 22 | |
d61fcfa4 SL |
23 | /* Throttling is performed over a slice and after that slice is renewed */ |
24 | #define DFL_THROTL_SLICE_HD (HZ / 10) | |
25 | #define DFL_THROTL_SLICE_SSD (HZ / 50) | |
297e3d85 | 26 | #define MAX_THROTL_SLICE (HZ) |
9e234eea | 27 | #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */ |
9bb67aeb SL |
28 | #define MIN_THROTL_BPS (320 * 1024) |
29 | #define MIN_THROTL_IOPS (10) | |
b4f428ef SL |
30 | #define DFL_LATENCY_TARGET (-1L) |
31 | #define DFL_IDLE_THRESHOLD (0) | |
6679a90c SL |
32 | #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */ |
33 | #define LATENCY_FILTERED_SSD (0) | |
34 | /* | |
35 | * For HD, very small latency comes from sequential IO. Such IO is helpless to | |
36 | * help determine if its IO is impacted by others, hence we ignore the IO | |
37 | */ | |
38 | #define LATENCY_FILTERED_HD (1000L) /* 1ms */ | |
e43473b7 | 39 | |
3c798398 | 40 | static struct blkcg_policy blkcg_policy_throtl; |
0381411e | 41 | |
450adcbe VG |
42 | /* A workqueue to queue throttle related work */ |
43 | static struct workqueue_struct *kthrotld_workqueue; | |
450adcbe | 44 | |
c5cc2070 TH |
45 | /* |
46 | * To implement hierarchical throttling, throtl_grps form a tree and bios | |
47 | * are dispatched upwards level by level until they reach the top and get | |
48 | * issued. When dispatching bios from the children and local group at each | |
49 | * level, if the bios are dispatched into a single bio_list, there's a risk | |
50 | * of a local or child group which can queue many bios at once filling up | |
51 | * the list starving others. | |
52 | * | |
53 | * To avoid such starvation, dispatched bios are queued separately | |
54 | * according to where they came from. When they are again dispatched to | |
55 | * the parent, they're popped in round-robin order so that no single source | |
56 | * hogs the dispatch window. | |
57 | * | |
58 | * throtl_qnode is used to keep the queued bios separated by their sources. | |
59 | * Bios are queued to throtl_qnode which in turn is queued to | |
60 | * throtl_service_queue and then dispatched in round-robin order. | |
61 | * | |
62 | * It's also used to track the reference counts on blkg's. A qnode always | |
63 | * belongs to a throtl_grp and gets queued on itself or the parent, so | |
64 | * incrementing the reference of the associated throtl_grp when a qnode is | |
65 | * queued and decrementing when dequeued is enough to keep the whole blkg | |
66 | * tree pinned while bios are in flight. | |
67 | */ | |
68 | struct throtl_qnode { | |
69 | struct list_head node; /* service_queue->queued[] */ | |
70 | struct bio_list bios; /* queued bios */ | |
71 | struct throtl_grp *tg; /* tg this qnode belongs to */ | |
72 | }; | |
73 | ||
c9e0332e | 74 | struct throtl_service_queue { |
77216b04 TH |
75 | struct throtl_service_queue *parent_sq; /* the parent service_queue */ |
76 | ||
73f0d49a TH |
77 | /* |
78 | * Bios queued directly to this service_queue or dispatched from | |
79 | * children throtl_grp's. | |
80 | */ | |
c5cc2070 | 81 | struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */ |
73f0d49a TH |
82 | unsigned int nr_queued[2]; /* number of queued bios */ |
83 | ||
84 | /* | |
85 | * RB tree of active children throtl_grp's, which are sorted by | |
86 | * their ->disptime. | |
87 | */ | |
9ff01255 | 88 | struct rb_root_cached pending_tree; /* RB tree of active tgs */ |
c9e0332e TH |
89 | unsigned int nr_pending; /* # queued in the tree */ |
90 | unsigned long first_pending_disptime; /* disptime of the first tg */ | |
69df0ab0 | 91 | struct timer_list pending_timer; /* fires on first_pending_disptime */ |
e43473b7 VG |
92 | }; |
93 | ||
5b2c16aa TH |
94 | enum tg_state_flags { |
95 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
0e9f4164 | 96 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ |
5b2c16aa TH |
97 | }; |
98 | ||
e43473b7 VG |
99 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
100 | ||
9f626e37 | 101 | enum { |
cd5ab1b0 | 102 | LIMIT_LOW, |
9f626e37 SL |
103 | LIMIT_MAX, |
104 | LIMIT_CNT, | |
105 | }; | |
106 | ||
e43473b7 | 107 | struct throtl_grp { |
f95a04af TH |
108 | /* must be the first member */ |
109 | struct blkg_policy_data pd; | |
110 | ||
c9e0332e | 111 | /* active throtl group service_queue member */ |
e43473b7 VG |
112 | struct rb_node rb_node; |
113 | ||
0f3457f6 TH |
114 | /* throtl_data this group belongs to */ |
115 | struct throtl_data *td; | |
116 | ||
49a2f1e3 TH |
117 | /* this group's service queue */ |
118 | struct throtl_service_queue service_queue; | |
119 | ||
c5cc2070 TH |
120 | /* |
121 | * qnode_on_self is used when bios are directly queued to this | |
122 | * throtl_grp so that local bios compete fairly with bios | |
123 | * dispatched from children. qnode_on_parent is used when bios are | |
124 | * dispatched from this throtl_grp into its parent and will compete | |
125 | * with the sibling qnode_on_parents and the parent's | |
126 | * qnode_on_self. | |
127 | */ | |
128 | struct throtl_qnode qnode_on_self[2]; | |
129 | struct throtl_qnode qnode_on_parent[2]; | |
130 | ||
e43473b7 VG |
131 | /* |
132 | * Dispatch time in jiffies. This is the estimated time when group | |
133 | * will unthrottle and is ready to dispatch more bio. It is used as | |
134 | * key to sort active groups in service tree. | |
135 | */ | |
136 | unsigned long disptime; | |
137 | ||
e43473b7 VG |
138 | unsigned int flags; |
139 | ||
693e751e TH |
140 | /* are there any throtl rules between this group and td? */ |
141 | bool has_rules[2]; | |
142 | ||
cd5ab1b0 | 143 | /* internally used bytes per second rate limits */ |
9f626e37 | 144 | uint64_t bps[2][LIMIT_CNT]; |
cd5ab1b0 SL |
145 | /* user configured bps limits */ |
146 | uint64_t bps_conf[2][LIMIT_CNT]; | |
e43473b7 | 147 | |
cd5ab1b0 | 148 | /* internally used IOPS limits */ |
9f626e37 | 149 | unsigned int iops[2][LIMIT_CNT]; |
cd5ab1b0 SL |
150 | /* user configured IOPS limits */ |
151 | unsigned int iops_conf[2][LIMIT_CNT]; | |
8e89d13f | 152 | |
b53b072c | 153 | /* Number of bytes dispatched in current slice */ |
e43473b7 | 154 | uint64_t bytes_disp[2]; |
8e89d13f VG |
155 | /* Number of bio's dispatched in current slice */ |
156 | unsigned int io_disp[2]; | |
e43473b7 | 157 | |
3f0abd80 SL |
158 | unsigned long last_low_overflow_time[2]; |
159 | ||
160 | uint64_t last_bytes_disp[2]; | |
161 | unsigned int last_io_disp[2]; | |
162 | ||
163 | unsigned long last_check_time; | |
164 | ||
ec80991d | 165 | unsigned long latency_target; /* us */ |
5b81fc3c | 166 | unsigned long latency_target_conf; /* us */ |
e43473b7 VG |
167 | /* When did we start a new slice */ |
168 | unsigned long slice_start[2]; | |
169 | unsigned long slice_end[2]; | |
9e234eea SL |
170 | |
171 | unsigned long last_finish_time; /* ns / 1024 */ | |
172 | unsigned long checked_last_finish_time; /* ns / 1024 */ | |
173 | unsigned long avg_idletime; /* ns / 1024 */ | |
174 | unsigned long idletime_threshold; /* us */ | |
5b81fc3c | 175 | unsigned long idletime_threshold_conf; /* us */ |
53696b8d SL |
176 | |
177 | unsigned int bio_cnt; /* total bios */ | |
178 | unsigned int bad_bio_cnt; /* bios exceeding latency threshold */ | |
179 | unsigned long bio_cnt_reset_time; | |
7ca46438 TH |
180 | |
181 | struct blkg_rwstat stat_bytes; | |
182 | struct blkg_rwstat stat_ios; | |
e43473b7 VG |
183 | }; |
184 | ||
b9147dd1 SL |
185 | /* We measure latency for request size from <= 4k to >= 1M */ |
186 | #define LATENCY_BUCKET_SIZE 9 | |
187 | ||
188 | struct latency_bucket { | |
189 | unsigned long total_latency; /* ns / 1024 */ | |
190 | int samples; | |
191 | }; | |
192 | ||
193 | struct avg_latency_bucket { | |
194 | unsigned long latency; /* ns / 1024 */ | |
195 | bool valid; | |
196 | }; | |
197 | ||
e43473b7 VG |
198 | struct throtl_data |
199 | { | |
e43473b7 | 200 | /* service tree for active throtl groups */ |
c9e0332e | 201 | struct throtl_service_queue service_queue; |
e43473b7 | 202 | |
e43473b7 VG |
203 | struct request_queue *queue; |
204 | ||
205 | /* Total Number of queued bios on READ and WRITE lists */ | |
206 | unsigned int nr_queued[2]; | |
207 | ||
297e3d85 SL |
208 | unsigned int throtl_slice; |
209 | ||
e43473b7 | 210 | /* Work for dispatching throttled bios */ |
69df0ab0 | 211 | struct work_struct dispatch_work; |
9f626e37 SL |
212 | unsigned int limit_index; |
213 | bool limit_valid[LIMIT_CNT]; | |
3f0abd80 SL |
214 | |
215 | unsigned long low_upgrade_time; | |
216 | unsigned long low_downgrade_time; | |
7394e31f SL |
217 | |
218 | unsigned int scale; | |
b9147dd1 | 219 | |
b889bf66 JQ |
220 | struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE]; |
221 | struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE]; | |
222 | struct latency_bucket __percpu *latency_buckets[2]; | |
b9147dd1 | 223 | unsigned long last_calculate_time; |
6679a90c | 224 | unsigned long filtered_latency; |
b9147dd1 SL |
225 | |
226 | bool track_bio_latency; | |
e43473b7 VG |
227 | }; |
228 | ||
e99e88a9 | 229 | static void throtl_pending_timer_fn(struct timer_list *t); |
69df0ab0 | 230 | |
f95a04af TH |
231 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
232 | { | |
233 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
234 | } | |
235 | ||
3c798398 | 236 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 237 | { |
f95a04af | 238 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
239 | } |
240 | ||
3c798398 | 241 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 242 | { |
f95a04af | 243 | return pd_to_blkg(&tg->pd); |
0381411e TH |
244 | } |
245 | ||
fda6f272 TH |
246 | /** |
247 | * sq_to_tg - return the throl_grp the specified service queue belongs to | |
248 | * @sq: the throtl_service_queue of interest | |
249 | * | |
250 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one | |
251 | * embedded in throtl_data, %NULL is returned. | |
252 | */ | |
253 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) | |
254 | { | |
255 | if (sq && sq->parent_sq) | |
256 | return container_of(sq, struct throtl_grp, service_queue); | |
257 | else | |
258 | return NULL; | |
259 | } | |
260 | ||
261 | /** | |
262 | * sq_to_td - return throtl_data the specified service queue belongs to | |
263 | * @sq: the throtl_service_queue of interest | |
264 | * | |
b43daedc | 265 | * A service_queue can be embedded in either a throtl_grp or throtl_data. |
fda6f272 TH |
266 | * Determine the associated throtl_data accordingly and return it. |
267 | */ | |
268 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) | |
269 | { | |
270 | struct throtl_grp *tg = sq_to_tg(sq); | |
271 | ||
272 | if (tg) | |
273 | return tg->td; | |
274 | else | |
275 | return container_of(sq, struct throtl_data, service_queue); | |
276 | } | |
277 | ||
7394e31f SL |
278 | /* |
279 | * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to | |
280 | * make the IO dispatch more smooth. | |
281 | * Scale up: linearly scale up according to lapsed time since upgrade. For | |
282 | * every throtl_slice, the limit scales up 1/2 .low limit till the | |
283 | * limit hits .max limit | |
284 | * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit | |
285 | */ | |
286 | static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td) | |
287 | { | |
288 | /* arbitrary value to avoid too big scale */ | |
289 | if (td->scale < 4096 && time_after_eq(jiffies, | |
290 | td->low_upgrade_time + td->scale * td->throtl_slice)) | |
291 | td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice; | |
292 | ||
293 | return low + (low >> 1) * td->scale; | |
294 | } | |
295 | ||
9f626e37 SL |
296 | static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw) |
297 | { | |
b22c417c | 298 | struct blkcg_gq *blkg = tg_to_blkg(tg); |
7394e31f | 299 | struct throtl_data *td; |
b22c417c SL |
300 | uint64_t ret; |
301 | ||
302 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
303 | return U64_MAX; | |
7394e31f SL |
304 | |
305 | td = tg->td; | |
306 | ret = tg->bps[rw][td->limit_index]; | |
9bb67aeb SL |
307 | if (ret == 0 && td->limit_index == LIMIT_LOW) { |
308 | /* intermediate node or iops isn't 0 */ | |
309 | if (!list_empty(&blkg->blkcg->css.children) || | |
310 | tg->iops[rw][td->limit_index]) | |
311 | return U64_MAX; | |
312 | else | |
313 | return MIN_THROTL_BPS; | |
314 | } | |
7394e31f SL |
315 | |
316 | if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] && | |
317 | tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) { | |
318 | uint64_t adjusted; | |
319 | ||
320 | adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td); | |
321 | ret = min(tg->bps[rw][LIMIT_MAX], adjusted); | |
322 | } | |
b22c417c | 323 | return ret; |
9f626e37 SL |
324 | } |
325 | ||
326 | static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw) | |
327 | { | |
b22c417c | 328 | struct blkcg_gq *blkg = tg_to_blkg(tg); |
7394e31f | 329 | struct throtl_data *td; |
b22c417c SL |
330 | unsigned int ret; |
331 | ||
332 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
333 | return UINT_MAX; | |
9bb67aeb | 334 | |
7394e31f SL |
335 | td = tg->td; |
336 | ret = tg->iops[rw][td->limit_index]; | |
9bb67aeb SL |
337 | if (ret == 0 && tg->td->limit_index == LIMIT_LOW) { |
338 | /* intermediate node or bps isn't 0 */ | |
339 | if (!list_empty(&blkg->blkcg->css.children) || | |
340 | tg->bps[rw][td->limit_index]) | |
341 | return UINT_MAX; | |
342 | else | |
343 | return MIN_THROTL_IOPS; | |
344 | } | |
7394e31f SL |
345 | |
346 | if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] && | |
347 | tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) { | |
348 | uint64_t adjusted; | |
349 | ||
350 | adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td); | |
351 | if (adjusted > UINT_MAX) | |
352 | adjusted = UINT_MAX; | |
353 | ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted); | |
354 | } | |
b22c417c | 355 | return ret; |
9f626e37 SL |
356 | } |
357 | ||
b9147dd1 SL |
358 | #define request_bucket_index(sectors) \ |
359 | clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1) | |
360 | ||
fda6f272 TH |
361 | /** |
362 | * throtl_log - log debug message via blktrace | |
363 | * @sq: the service_queue being reported | |
364 | * @fmt: printf format string | |
365 | * @args: printf args | |
366 | * | |
367 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a | |
368 | * throtl_grp; otherwise, just "throtl". | |
fda6f272 TH |
369 | */ |
370 | #define throtl_log(sq, fmt, args...) do { \ | |
371 | struct throtl_grp *__tg = sq_to_tg((sq)); \ | |
372 | struct throtl_data *__td = sq_to_td((sq)); \ | |
373 | \ | |
374 | (void)__td; \ | |
59fa0224 SL |
375 | if (likely(!blk_trace_note_message_enabled(__td->queue))) \ |
376 | break; \ | |
fda6f272 | 377 | if ((__tg)) { \ |
35fe6d76 SL |
378 | blk_add_cgroup_trace_msg(__td->queue, \ |
379 | tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\ | |
fda6f272 TH |
380 | } else { \ |
381 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
382 | } \ | |
54e7ed12 | 383 | } while (0) |
e43473b7 | 384 | |
ea0ea2bc SL |
385 | static inline unsigned int throtl_bio_data_size(struct bio *bio) |
386 | { | |
387 | /* assume it's one sector */ | |
388 | if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) | |
389 | return 512; | |
390 | return bio->bi_iter.bi_size; | |
391 | } | |
392 | ||
c5cc2070 TH |
393 | static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg) |
394 | { | |
395 | INIT_LIST_HEAD(&qn->node); | |
396 | bio_list_init(&qn->bios); | |
397 | qn->tg = tg; | |
398 | } | |
399 | ||
400 | /** | |
401 | * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it | |
402 | * @bio: bio being added | |
403 | * @qn: qnode to add bio to | |
404 | * @queued: the service_queue->queued[] list @qn belongs to | |
405 | * | |
406 | * Add @bio to @qn and put @qn on @queued if it's not already on. | |
407 | * @qn->tg's reference count is bumped when @qn is activated. See the | |
408 | * comment on top of throtl_qnode definition for details. | |
409 | */ | |
410 | static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn, | |
411 | struct list_head *queued) | |
412 | { | |
413 | bio_list_add(&qn->bios, bio); | |
414 | if (list_empty(&qn->node)) { | |
415 | list_add_tail(&qn->node, queued); | |
416 | blkg_get(tg_to_blkg(qn->tg)); | |
417 | } | |
418 | } | |
419 | ||
420 | /** | |
421 | * throtl_peek_queued - peek the first bio on a qnode list | |
422 | * @queued: the qnode list to peek | |
423 | */ | |
424 | static struct bio *throtl_peek_queued(struct list_head *queued) | |
425 | { | |
426 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
427 | struct bio *bio; | |
428 | ||
429 | if (list_empty(queued)) | |
430 | return NULL; | |
431 | ||
432 | bio = bio_list_peek(&qn->bios); | |
433 | WARN_ON_ONCE(!bio); | |
434 | return bio; | |
435 | } | |
436 | ||
437 | /** | |
438 | * throtl_pop_queued - pop the first bio form a qnode list | |
439 | * @queued: the qnode list to pop a bio from | |
440 | * @tg_to_put: optional out argument for throtl_grp to put | |
441 | * | |
442 | * Pop the first bio from the qnode list @queued. After popping, the first | |
443 | * qnode is removed from @queued if empty or moved to the end of @queued so | |
444 | * that the popping order is round-robin. | |
445 | * | |
446 | * When the first qnode is removed, its associated throtl_grp should be put | |
447 | * too. If @tg_to_put is NULL, this function automatically puts it; | |
448 | * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is | |
449 | * responsible for putting it. | |
450 | */ | |
451 | static struct bio *throtl_pop_queued(struct list_head *queued, | |
452 | struct throtl_grp **tg_to_put) | |
453 | { | |
454 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
455 | struct bio *bio; | |
456 | ||
457 | if (list_empty(queued)) | |
458 | return NULL; | |
459 | ||
460 | bio = bio_list_pop(&qn->bios); | |
461 | WARN_ON_ONCE(!bio); | |
462 | ||
463 | if (bio_list_empty(&qn->bios)) { | |
464 | list_del_init(&qn->node); | |
465 | if (tg_to_put) | |
466 | *tg_to_put = qn->tg; | |
467 | else | |
468 | blkg_put(tg_to_blkg(qn->tg)); | |
469 | } else { | |
470 | list_move_tail(&qn->node, queued); | |
471 | } | |
472 | ||
473 | return bio; | |
474 | } | |
475 | ||
49a2f1e3 | 476 | /* init a service_queue, assumes the caller zeroed it */ |
b2ce2643 | 477 | static void throtl_service_queue_init(struct throtl_service_queue *sq) |
49a2f1e3 | 478 | { |
c5cc2070 TH |
479 | INIT_LIST_HEAD(&sq->queued[0]); |
480 | INIT_LIST_HEAD(&sq->queued[1]); | |
9ff01255 | 481 | sq->pending_tree = RB_ROOT_CACHED; |
e99e88a9 | 482 | timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0); |
69df0ab0 TH |
483 | } |
484 | ||
cf09a8ee TH |
485 | static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, |
486 | struct request_queue *q, | |
487 | struct blkcg *blkcg) | |
001bea73 | 488 | { |
4fb72036 | 489 | struct throtl_grp *tg; |
24bdb8ef | 490 | int rw; |
4fb72036 | 491 | |
cf09a8ee | 492 | tg = kzalloc_node(sizeof(*tg), gfp, q->node); |
4fb72036 | 493 | if (!tg) |
77ea7338 | 494 | return NULL; |
4fb72036 | 495 | |
7ca46438 TH |
496 | if (blkg_rwstat_init(&tg->stat_bytes, gfp)) |
497 | goto err_free_tg; | |
498 | ||
499 | if (blkg_rwstat_init(&tg->stat_ios, gfp)) | |
500 | goto err_exit_stat_bytes; | |
501 | ||
b2ce2643 TH |
502 | throtl_service_queue_init(&tg->service_queue); |
503 | ||
504 | for (rw = READ; rw <= WRITE; rw++) { | |
505 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); | |
506 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); | |
507 | } | |
508 | ||
509 | RB_CLEAR_NODE(&tg->rb_node); | |
9f626e37 SL |
510 | tg->bps[READ][LIMIT_MAX] = U64_MAX; |
511 | tg->bps[WRITE][LIMIT_MAX] = U64_MAX; | |
512 | tg->iops[READ][LIMIT_MAX] = UINT_MAX; | |
513 | tg->iops[WRITE][LIMIT_MAX] = UINT_MAX; | |
cd5ab1b0 SL |
514 | tg->bps_conf[READ][LIMIT_MAX] = U64_MAX; |
515 | tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX; | |
516 | tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX; | |
517 | tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX; | |
518 | /* LIMIT_LOW will have default value 0 */ | |
b2ce2643 | 519 | |
ec80991d | 520 | tg->latency_target = DFL_LATENCY_TARGET; |
5b81fc3c | 521 | tg->latency_target_conf = DFL_LATENCY_TARGET; |
b4f428ef SL |
522 | tg->idletime_threshold = DFL_IDLE_THRESHOLD; |
523 | tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD; | |
ec80991d | 524 | |
4fb72036 | 525 | return &tg->pd; |
7ca46438 TH |
526 | |
527 | err_exit_stat_bytes: | |
528 | blkg_rwstat_exit(&tg->stat_bytes); | |
529 | err_free_tg: | |
530 | kfree(tg); | |
531 | return NULL; | |
001bea73 TH |
532 | } |
533 | ||
a9520cd6 | 534 | static void throtl_pd_init(struct blkg_policy_data *pd) |
a29a171e | 535 | { |
a9520cd6 TH |
536 | struct throtl_grp *tg = pd_to_tg(pd); |
537 | struct blkcg_gq *blkg = tg_to_blkg(tg); | |
77216b04 | 538 | struct throtl_data *td = blkg->q->td; |
b2ce2643 | 539 | struct throtl_service_queue *sq = &tg->service_queue; |
cd1604fa | 540 | |
9138125b | 541 | /* |
aa6ec29b | 542 | * If on the default hierarchy, we switch to properly hierarchical |
9138125b TH |
543 | * behavior where limits on a given throtl_grp are applied to the |
544 | * whole subtree rather than just the group itself. e.g. If 16M | |
545 | * read_bps limit is set on the root group, the whole system can't | |
546 | * exceed 16M for the device. | |
547 | * | |
aa6ec29b | 548 | * If not on the default hierarchy, the broken flat hierarchy |
9138125b TH |
549 | * behavior is retained where all throtl_grps are treated as if |
550 | * they're all separate root groups right below throtl_data. | |
551 | * Limits of a group don't interact with limits of other groups | |
552 | * regardless of the position of the group in the hierarchy. | |
553 | */ | |
b2ce2643 | 554 | sq->parent_sq = &td->service_queue; |
9e10a130 | 555 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent) |
b2ce2643 | 556 | sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue; |
77216b04 | 557 | tg->td = td; |
8a3d2615 TH |
558 | } |
559 | ||
693e751e TH |
560 | /* |
561 | * Set has_rules[] if @tg or any of its parents have limits configured. | |
562 | * This doesn't require walking up to the top of the hierarchy as the | |
563 | * parent's has_rules[] is guaranteed to be correct. | |
564 | */ | |
565 | static void tg_update_has_rules(struct throtl_grp *tg) | |
566 | { | |
567 | struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); | |
9f626e37 | 568 | struct throtl_data *td = tg->td; |
693e751e TH |
569 | int rw; |
570 | ||
571 | for (rw = READ; rw <= WRITE; rw++) | |
572 | tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) || | |
9f626e37 SL |
573 | (td->limit_valid[td->limit_index] && |
574 | (tg_bps_limit(tg, rw) != U64_MAX || | |
575 | tg_iops_limit(tg, rw) != UINT_MAX)); | |
693e751e TH |
576 | } |
577 | ||
a9520cd6 | 578 | static void throtl_pd_online(struct blkg_policy_data *pd) |
693e751e | 579 | { |
aec24246 | 580 | struct throtl_grp *tg = pd_to_tg(pd); |
693e751e TH |
581 | /* |
582 | * We don't want new groups to escape the limits of its ancestors. | |
583 | * Update has_rules[] after a new group is brought online. | |
584 | */ | |
aec24246 | 585 | tg_update_has_rules(tg); |
693e751e TH |
586 | } |
587 | ||
cd5ab1b0 SL |
588 | static void blk_throtl_update_limit_valid(struct throtl_data *td) |
589 | { | |
590 | struct cgroup_subsys_state *pos_css; | |
591 | struct blkcg_gq *blkg; | |
592 | bool low_valid = false; | |
593 | ||
594 | rcu_read_lock(); | |
595 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
596 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
597 | ||
598 | if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] || | |
43ada787 | 599 | tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) { |
cd5ab1b0 | 600 | low_valid = true; |
43ada787 LB |
601 | break; |
602 | } | |
cd5ab1b0 SL |
603 | } |
604 | rcu_read_unlock(); | |
605 | ||
606 | td->limit_valid[LIMIT_LOW] = low_valid; | |
607 | } | |
608 | ||
c79892c5 | 609 | static void throtl_upgrade_state(struct throtl_data *td); |
cd5ab1b0 SL |
610 | static void throtl_pd_offline(struct blkg_policy_data *pd) |
611 | { | |
612 | struct throtl_grp *tg = pd_to_tg(pd); | |
613 | ||
614 | tg->bps[READ][LIMIT_LOW] = 0; | |
615 | tg->bps[WRITE][LIMIT_LOW] = 0; | |
616 | tg->iops[READ][LIMIT_LOW] = 0; | |
617 | tg->iops[WRITE][LIMIT_LOW] = 0; | |
618 | ||
619 | blk_throtl_update_limit_valid(tg->td); | |
620 | ||
c79892c5 SL |
621 | if (!tg->td->limit_valid[tg->td->limit_index]) |
622 | throtl_upgrade_state(tg->td); | |
cd5ab1b0 SL |
623 | } |
624 | ||
001bea73 TH |
625 | static void throtl_pd_free(struct blkg_policy_data *pd) |
626 | { | |
4fb72036 TH |
627 | struct throtl_grp *tg = pd_to_tg(pd); |
628 | ||
b2ce2643 | 629 | del_timer_sync(&tg->service_queue.pending_timer); |
7ca46438 TH |
630 | blkg_rwstat_exit(&tg->stat_bytes); |
631 | blkg_rwstat_exit(&tg->stat_ios); | |
4fb72036 | 632 | kfree(tg); |
001bea73 TH |
633 | } |
634 | ||
0049af73 TH |
635 | static struct throtl_grp * |
636 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 | 637 | { |
9ff01255 | 638 | struct rb_node *n; |
e43473b7 | 639 | /* Service tree is empty */ |
0049af73 | 640 | if (!parent_sq->nr_pending) |
e43473b7 VG |
641 | return NULL; |
642 | ||
9ff01255 LB |
643 | n = rb_first_cached(&parent_sq->pending_tree); |
644 | WARN_ON_ONCE(!n); | |
645 | if (!n) | |
646 | return NULL; | |
647 | return rb_entry_tg(n); | |
e43473b7 VG |
648 | } |
649 | ||
0049af73 TH |
650 | static void throtl_rb_erase(struct rb_node *n, |
651 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 652 | { |
9ff01255 LB |
653 | rb_erase_cached(n, &parent_sq->pending_tree); |
654 | RB_CLEAR_NODE(n); | |
0049af73 | 655 | --parent_sq->nr_pending; |
e43473b7 VG |
656 | } |
657 | ||
0049af73 | 658 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
659 | { |
660 | struct throtl_grp *tg; | |
661 | ||
0049af73 | 662 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
663 | if (!tg) |
664 | return; | |
665 | ||
0049af73 | 666 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
667 | } |
668 | ||
77216b04 | 669 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 670 | { |
77216b04 | 671 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
9ff01255 | 672 | struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node; |
e43473b7 VG |
673 | struct rb_node *parent = NULL; |
674 | struct throtl_grp *__tg; | |
675 | unsigned long key = tg->disptime; | |
9ff01255 | 676 | bool leftmost = true; |
e43473b7 VG |
677 | |
678 | while (*node != NULL) { | |
679 | parent = *node; | |
680 | __tg = rb_entry_tg(parent); | |
681 | ||
682 | if (time_before(key, __tg->disptime)) | |
683 | node = &parent->rb_left; | |
684 | else { | |
685 | node = &parent->rb_right; | |
9ff01255 | 686 | leftmost = false; |
e43473b7 VG |
687 | } |
688 | } | |
689 | ||
e43473b7 | 690 | rb_link_node(&tg->rb_node, parent, node); |
9ff01255 LB |
691 | rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree, |
692 | leftmost); | |
e43473b7 VG |
693 | } |
694 | ||
77216b04 | 695 | static void __throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 696 | { |
77216b04 | 697 | tg_service_queue_add(tg); |
5b2c16aa | 698 | tg->flags |= THROTL_TG_PENDING; |
77216b04 | 699 | tg->service_queue.parent_sq->nr_pending++; |
e43473b7 VG |
700 | } |
701 | ||
77216b04 | 702 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 703 | { |
5b2c16aa | 704 | if (!(tg->flags & THROTL_TG_PENDING)) |
77216b04 | 705 | __throtl_enqueue_tg(tg); |
e43473b7 VG |
706 | } |
707 | ||
77216b04 | 708 | static void __throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 709 | { |
77216b04 | 710 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
5b2c16aa | 711 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
712 | } |
713 | ||
77216b04 | 714 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 715 | { |
5b2c16aa | 716 | if (tg->flags & THROTL_TG_PENDING) |
77216b04 | 717 | __throtl_dequeue_tg(tg); |
e43473b7 VG |
718 | } |
719 | ||
a9131a27 | 720 | /* Call with queue lock held */ |
69df0ab0 TH |
721 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, |
722 | unsigned long expires) | |
a9131a27 | 723 | { |
a41b816c | 724 | unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice; |
06cceedc SL |
725 | |
726 | /* | |
727 | * Since we are adjusting the throttle limit dynamically, the sleep | |
728 | * time calculated according to previous limit might be invalid. It's | |
729 | * possible the cgroup sleep time is very long and no other cgroups | |
730 | * have IO running so notify the limit changes. Make sure the cgroup | |
731 | * doesn't sleep too long to avoid the missed notification. | |
732 | */ | |
733 | if (time_after(expires, max_expire)) | |
734 | expires = max_expire; | |
69df0ab0 TH |
735 | mod_timer(&sq->pending_timer, expires); |
736 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
737 | expires - jiffies, jiffies); | |
a9131a27 TH |
738 | } |
739 | ||
7f52f98c TH |
740 | /** |
741 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle | |
742 | * @sq: the service_queue to schedule dispatch for | |
743 | * @force: force scheduling | |
744 | * | |
745 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the | |
746 | * dispatch time of the first pending child. Returns %true if either timer | |
747 | * is armed or there's no pending child left. %false if the current | |
748 | * dispatch window is still open and the caller should continue | |
749 | * dispatching. | |
750 | * | |
751 | * If @force is %true, the dispatch timer is always scheduled and this | |
752 | * function is guaranteed to return %true. This is to be used when the | |
753 | * caller can't dispatch itself and needs to invoke pending_timer | |
754 | * unconditionally. Note that forced scheduling is likely to induce short | |
755 | * delay before dispatch starts even if @sq->first_pending_disptime is not | |
756 | * in the future and thus shouldn't be used in hot paths. | |
757 | */ | |
758 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
759 | bool force) | |
e43473b7 | 760 | { |
6a525600 | 761 | /* any pending children left? */ |
c9e0332e | 762 | if (!sq->nr_pending) |
7f52f98c | 763 | return true; |
e43473b7 | 764 | |
c9e0332e | 765 | update_min_dispatch_time(sq); |
e43473b7 | 766 | |
69df0ab0 | 767 | /* is the next dispatch time in the future? */ |
7f52f98c | 768 | if (force || time_after(sq->first_pending_disptime, jiffies)) { |
69df0ab0 | 769 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); |
7f52f98c | 770 | return true; |
69df0ab0 TH |
771 | } |
772 | ||
7f52f98c TH |
773 | /* tell the caller to continue dispatching */ |
774 | return false; | |
e43473b7 VG |
775 | } |
776 | ||
32ee5bc4 VG |
777 | static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, |
778 | bool rw, unsigned long start) | |
779 | { | |
780 | tg->bytes_disp[rw] = 0; | |
781 | tg->io_disp[rw] = 0; | |
782 | ||
783 | /* | |
784 | * Previous slice has expired. We must have trimmed it after last | |
785 | * bio dispatch. That means since start of last slice, we never used | |
786 | * that bandwidth. Do try to make use of that bandwidth while giving | |
787 | * credit. | |
788 | */ | |
789 | if (time_after_eq(start, tg->slice_start[rw])) | |
790 | tg->slice_start[rw] = start; | |
791 | ||
297e3d85 | 792 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; |
32ee5bc4 VG |
793 | throtl_log(&tg->service_queue, |
794 | "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", | |
795 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
796 | tg->slice_end[rw], jiffies); | |
797 | } | |
798 | ||
0f3457f6 | 799 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
800 | { |
801 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 802 | tg->io_disp[rw] = 0; |
e43473b7 | 803 | tg->slice_start[rw] = jiffies; |
297e3d85 | 804 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; |
fda6f272 TH |
805 | throtl_log(&tg->service_queue, |
806 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
807 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
808 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
809 | } |
810 | ||
0f3457f6 TH |
811 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
812 | unsigned long jiffy_end) | |
d1ae8ffd | 813 | { |
297e3d85 | 814 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); |
d1ae8ffd VG |
815 | } |
816 | ||
0f3457f6 TH |
817 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
818 | unsigned long jiffy_end) | |
e43473b7 | 819 | { |
297e3d85 | 820 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); |
fda6f272 TH |
821 | throtl_log(&tg->service_queue, |
822 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
823 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
824 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
825 | } |
826 | ||
827 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 828 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
829 | { |
830 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
5cf8c227 | 831 | return false; |
e43473b7 | 832 | |
0b6bad7d | 833 | return true; |
e43473b7 VG |
834 | } |
835 | ||
836 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 837 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 838 | { |
3aad5d3e VG |
839 | unsigned long nr_slices, time_elapsed, io_trim; |
840 | u64 bytes_trim, tmp; | |
e43473b7 VG |
841 | |
842 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
843 | ||
844 | /* | |
845 | * If bps are unlimited (-1), then time slice don't get | |
846 | * renewed. Don't try to trim the slice if slice is used. A new | |
847 | * slice will start when appropriate. | |
848 | */ | |
0f3457f6 | 849 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
850 | return; |
851 | ||
d1ae8ffd VG |
852 | /* |
853 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
854 | * that initially cgroup limit was very low resulting in high | |
b53b072c | 855 | * slice_end, but later limit was bumped up and bio was dispatched |
d1ae8ffd VG |
856 | * sooner, then we need to reduce slice_end. A high bogus slice_end |
857 | * is bad because it does not allow new slice to start. | |
858 | */ | |
859 | ||
297e3d85 | 860 | throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice); |
d1ae8ffd | 861 | |
e43473b7 VG |
862 | time_elapsed = jiffies - tg->slice_start[rw]; |
863 | ||
297e3d85 | 864 | nr_slices = time_elapsed / tg->td->throtl_slice; |
e43473b7 VG |
865 | |
866 | if (!nr_slices) | |
867 | return; | |
297e3d85 | 868 | tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices; |
3aad5d3e VG |
869 | do_div(tmp, HZ); |
870 | bytes_trim = tmp; | |
e43473b7 | 871 | |
297e3d85 SL |
872 | io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) / |
873 | HZ; | |
e43473b7 | 874 | |
8e89d13f | 875 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
876 | return; |
877 | ||
878 | if (tg->bytes_disp[rw] >= bytes_trim) | |
879 | tg->bytes_disp[rw] -= bytes_trim; | |
880 | else | |
881 | tg->bytes_disp[rw] = 0; | |
882 | ||
8e89d13f VG |
883 | if (tg->io_disp[rw] >= io_trim) |
884 | tg->io_disp[rw] -= io_trim; | |
885 | else | |
886 | tg->io_disp[rw] = 0; | |
887 | ||
297e3d85 | 888 | tg->slice_start[rw] += nr_slices * tg->td->throtl_slice; |
e43473b7 | 889 | |
fda6f272 TH |
890 | throtl_log(&tg->service_queue, |
891 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
892 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
893 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
e43473b7 VG |
894 | } |
895 | ||
0f3457f6 | 896 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
4599ea49 | 897 | u32 iops_limit, unsigned long *wait) |
e43473b7 VG |
898 | { |
899 | bool rw = bio_data_dir(bio); | |
8e89d13f | 900 | unsigned int io_allowed; |
e43473b7 | 901 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 902 | u64 tmp; |
e43473b7 | 903 | |
87fbeb88 BW |
904 | if (iops_limit == UINT_MAX) { |
905 | if (wait) | |
906 | *wait = 0; | |
907 | return true; | |
908 | } | |
909 | ||
3a10f999 | 910 | jiffy_elapsed = jiffies - tg->slice_start[rw]; |
e43473b7 | 911 | |
3a10f999 KK |
912 | /* Round up to the next throttle slice, wait time must be nonzero */ |
913 | jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice); | |
8e89d13f | 914 | |
c49c06e4 VG |
915 | /* |
916 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
917 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
918 | * will allow dispatch after 1 second and after that slice should | |
919 | * have been trimmed. | |
920 | */ | |
921 | ||
4599ea49 | 922 | tmp = (u64)iops_limit * jiffy_elapsed_rnd; |
c49c06e4 VG |
923 | do_div(tmp, HZ); |
924 | ||
925 | if (tmp > UINT_MAX) | |
926 | io_allowed = UINT_MAX; | |
927 | else | |
928 | io_allowed = tmp; | |
8e89d13f VG |
929 | |
930 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
931 | if (wait) |
932 | *wait = 0; | |
5cf8c227 | 933 | return true; |
e43473b7 VG |
934 | } |
935 | ||
8e89d13f | 936 | /* Calc approx time to dispatch */ |
991f61fe | 937 | jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed; |
8e89d13f VG |
938 | |
939 | if (wait) | |
940 | *wait = jiffy_wait; | |
0b6bad7d | 941 | return false; |
8e89d13f VG |
942 | } |
943 | ||
0f3457f6 | 944 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
4599ea49 | 945 | u64 bps_limit, unsigned long *wait) |
8e89d13f VG |
946 | { |
947 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 948 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 949 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
ea0ea2bc | 950 | unsigned int bio_size = throtl_bio_data_size(bio); |
e43473b7 | 951 | |
87fbeb88 BW |
952 | if (bps_limit == U64_MAX) { |
953 | if (wait) | |
954 | *wait = 0; | |
955 | return true; | |
956 | } | |
957 | ||
e43473b7 VG |
958 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
959 | ||
960 | /* Slice has just started. Consider one slice interval */ | |
961 | if (!jiffy_elapsed) | |
297e3d85 | 962 | jiffy_elapsed_rnd = tg->td->throtl_slice; |
e43473b7 | 963 | |
297e3d85 | 964 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); |
e43473b7 | 965 | |
4599ea49 | 966 | tmp = bps_limit * jiffy_elapsed_rnd; |
5e901a2b | 967 | do_div(tmp, HZ); |
3aad5d3e | 968 | bytes_allowed = tmp; |
e43473b7 | 969 | |
ea0ea2bc | 970 | if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) { |
e43473b7 VG |
971 | if (wait) |
972 | *wait = 0; | |
5cf8c227 | 973 | return true; |
e43473b7 VG |
974 | } |
975 | ||
976 | /* Calc approx time to dispatch */ | |
ea0ea2bc | 977 | extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed; |
4599ea49 | 978 | jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit); |
e43473b7 VG |
979 | |
980 | if (!jiffy_wait) | |
981 | jiffy_wait = 1; | |
982 | ||
983 | /* | |
984 | * This wait time is without taking into consideration the rounding | |
985 | * up we did. Add that time also. | |
986 | */ | |
987 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
988 | if (wait) |
989 | *wait = jiffy_wait; | |
0b6bad7d | 990 | return false; |
8e89d13f VG |
991 | } |
992 | ||
993 | /* | |
994 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
995 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
996 | */ | |
0f3457f6 TH |
997 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
998 | unsigned long *wait) | |
8e89d13f VG |
999 | { |
1000 | bool rw = bio_data_dir(bio); | |
1001 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
4599ea49 BW |
1002 | u64 bps_limit = tg_bps_limit(tg, rw); |
1003 | u32 iops_limit = tg_iops_limit(tg, rw); | |
8e89d13f VG |
1004 | |
1005 | /* | |
1006 | * Currently whole state machine of group depends on first bio | |
1007 | * queued in the group bio list. So one should not be calling | |
1008 | * this function with a different bio if there are other bios | |
1009 | * queued. | |
1010 | */ | |
73f0d49a | 1011 | BUG_ON(tg->service_queue.nr_queued[rw] && |
c5cc2070 | 1012 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); |
e43473b7 | 1013 | |
8e89d13f | 1014 | /* If tg->bps = -1, then BW is unlimited */ |
4599ea49 | 1015 | if (bps_limit == U64_MAX && iops_limit == UINT_MAX) { |
8e89d13f VG |
1016 | if (wait) |
1017 | *wait = 0; | |
5cf8c227 | 1018 | return true; |
8e89d13f VG |
1019 | } |
1020 | ||
1021 | /* | |
1022 | * If previous slice expired, start a new one otherwise renew/extend | |
1023 | * existing slice to make sure it is at least throtl_slice interval | |
164c80ed VG |
1024 | * long since now. New slice is started only for empty throttle group. |
1025 | * If there is queued bio, that means there should be an active | |
1026 | * slice and it should be extended instead. | |
8e89d13f | 1027 | */ |
164c80ed | 1028 | if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw])) |
0f3457f6 | 1029 | throtl_start_new_slice(tg, rw); |
8e89d13f | 1030 | else { |
297e3d85 SL |
1031 | if (time_before(tg->slice_end[rw], |
1032 | jiffies + tg->td->throtl_slice)) | |
1033 | throtl_extend_slice(tg, rw, | |
1034 | jiffies + tg->td->throtl_slice); | |
8e89d13f VG |
1035 | } |
1036 | ||
4599ea49 BW |
1037 | if (tg_with_in_bps_limit(tg, bio, bps_limit, &bps_wait) && |
1038 | tg_with_in_iops_limit(tg, bio, iops_limit, &iops_wait)) { | |
8e89d13f VG |
1039 | if (wait) |
1040 | *wait = 0; | |
0b6bad7d | 1041 | return true; |
8e89d13f VG |
1042 | } |
1043 | ||
1044 | max_wait = max(bps_wait, iops_wait); | |
1045 | ||
1046 | if (wait) | |
1047 | *wait = max_wait; | |
1048 | ||
1049 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 1050 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 | 1051 | |
0b6bad7d | 1052 | return false; |
e43473b7 VG |
1053 | } |
1054 | ||
1055 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
1056 | { | |
1057 | bool rw = bio_data_dir(bio); | |
ea0ea2bc | 1058 | unsigned int bio_size = throtl_bio_data_size(bio); |
e43473b7 VG |
1059 | |
1060 | /* Charge the bio to the group */ | |
ea0ea2bc | 1061 | tg->bytes_disp[rw] += bio_size; |
8e89d13f | 1062 | tg->io_disp[rw]++; |
ea0ea2bc | 1063 | tg->last_bytes_disp[rw] += bio_size; |
3f0abd80 | 1064 | tg->last_io_disp[rw]++; |
e43473b7 | 1065 | |
2a0f61e6 | 1066 | /* |
8d2bbd4c | 1067 | * BIO_THROTTLED is used to prevent the same bio to be throttled |
2a0f61e6 TH |
1068 | * more than once as a throttled bio will go through blk-throtl the |
1069 | * second time when it eventually gets issued. Set it when a bio | |
1070 | * is being charged to a tg. | |
2a0f61e6 | 1071 | */ |
8d2bbd4c CH |
1072 | if (!bio_flagged(bio, BIO_THROTTLED)) |
1073 | bio_set_flag(bio, BIO_THROTTLED); | |
e43473b7 VG |
1074 | } |
1075 | ||
c5cc2070 TH |
1076 | /** |
1077 | * throtl_add_bio_tg - add a bio to the specified throtl_grp | |
1078 | * @bio: bio to add | |
1079 | * @qn: qnode to use | |
1080 | * @tg: the target throtl_grp | |
1081 | * | |
1082 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, | |
1083 | * tg->qnode_on_self[] is used. | |
1084 | */ | |
1085 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, | |
1086 | struct throtl_grp *tg) | |
e43473b7 | 1087 | { |
73f0d49a | 1088 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1089 | bool rw = bio_data_dir(bio); |
1090 | ||
c5cc2070 TH |
1091 | if (!qn) |
1092 | qn = &tg->qnode_on_self[rw]; | |
1093 | ||
0e9f4164 TH |
1094 | /* |
1095 | * If @tg doesn't currently have any bios queued in the same | |
1096 | * direction, queueing @bio can change when @tg should be | |
1097 | * dispatched. Mark that @tg was empty. This is automatically | |
b53b072c | 1098 | * cleared on the next tg_update_disptime(). |
0e9f4164 TH |
1099 | */ |
1100 | if (!sq->nr_queued[rw]) | |
1101 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
1102 | ||
c5cc2070 TH |
1103 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); |
1104 | ||
73f0d49a | 1105 | sq->nr_queued[rw]++; |
77216b04 | 1106 | throtl_enqueue_tg(tg); |
e43473b7 VG |
1107 | } |
1108 | ||
77216b04 | 1109 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 1110 | { |
73f0d49a | 1111 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1112 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
1113 | struct bio *bio; | |
1114 | ||
d609af3a ME |
1115 | bio = throtl_peek_queued(&sq->queued[READ]); |
1116 | if (bio) | |
0f3457f6 | 1117 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 1118 | |
d609af3a ME |
1119 | bio = throtl_peek_queued(&sq->queued[WRITE]); |
1120 | if (bio) | |
0f3457f6 | 1121 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
1122 | |
1123 | min_wait = min(read_wait, write_wait); | |
1124 | disptime = jiffies + min_wait; | |
1125 | ||
e43473b7 | 1126 | /* Update dispatch time */ |
77216b04 | 1127 | throtl_dequeue_tg(tg); |
e43473b7 | 1128 | tg->disptime = disptime; |
77216b04 | 1129 | throtl_enqueue_tg(tg); |
0e9f4164 TH |
1130 | |
1131 | /* see throtl_add_bio_tg() */ | |
1132 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
1133 | } |
1134 | ||
32ee5bc4 VG |
1135 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, |
1136 | struct throtl_grp *parent_tg, bool rw) | |
1137 | { | |
1138 | if (throtl_slice_used(parent_tg, rw)) { | |
1139 | throtl_start_new_slice_with_credit(parent_tg, rw, | |
1140 | child_tg->slice_start[rw]); | |
1141 | } | |
1142 | ||
1143 | } | |
1144 | ||
77216b04 | 1145 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 1146 | { |
73f0d49a | 1147 | struct throtl_service_queue *sq = &tg->service_queue; |
6bc9c2b4 TH |
1148 | struct throtl_service_queue *parent_sq = sq->parent_sq; |
1149 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
c5cc2070 | 1150 | struct throtl_grp *tg_to_put = NULL; |
e43473b7 VG |
1151 | struct bio *bio; |
1152 | ||
c5cc2070 TH |
1153 | /* |
1154 | * @bio is being transferred from @tg to @parent_sq. Popping a bio | |
1155 | * from @tg may put its reference and @parent_sq might end up | |
1156 | * getting released prematurely. Remember the tg to put and put it | |
1157 | * after @bio is transferred to @parent_sq. | |
1158 | */ | |
1159 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); | |
73f0d49a | 1160 | sq->nr_queued[rw]--; |
e43473b7 VG |
1161 | |
1162 | throtl_charge_bio(tg, bio); | |
6bc9c2b4 TH |
1163 | |
1164 | /* | |
1165 | * If our parent is another tg, we just need to transfer @bio to | |
1166 | * the parent using throtl_add_bio_tg(). If our parent is | |
1167 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
1168 | * bio_lists[] and decrease total number queued. The caller is | |
1169 | * responsible for issuing these bios. | |
1170 | */ | |
1171 | if (parent_tg) { | |
c5cc2070 | 1172 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); |
32ee5bc4 | 1173 | start_parent_slice_with_credit(tg, parent_tg, rw); |
6bc9c2b4 | 1174 | } else { |
c5cc2070 TH |
1175 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], |
1176 | &parent_sq->queued[rw]); | |
6bc9c2b4 TH |
1177 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
1178 | tg->td->nr_queued[rw]--; | |
1179 | } | |
e43473b7 | 1180 | |
0f3457f6 | 1181 | throtl_trim_slice(tg, rw); |
6bc9c2b4 | 1182 | |
c5cc2070 TH |
1183 | if (tg_to_put) |
1184 | blkg_put(tg_to_blkg(tg_to_put)); | |
e43473b7 VG |
1185 | } |
1186 | ||
77216b04 | 1187 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 1188 | { |
73f0d49a | 1189 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 | 1190 | unsigned int nr_reads = 0, nr_writes = 0; |
e675df2a BW |
1191 | unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4; |
1192 | unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads; | |
e43473b7 VG |
1193 | struct bio *bio; |
1194 | ||
1195 | /* Try to dispatch 75% READS and 25% WRITES */ | |
1196 | ||
c5cc2070 | 1197 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && |
0f3457f6 | 1198 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1199 | |
77216b04 | 1200 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1201 | nr_reads++; |
1202 | ||
1203 | if (nr_reads >= max_nr_reads) | |
1204 | break; | |
1205 | } | |
1206 | ||
c5cc2070 | 1207 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && |
0f3457f6 | 1208 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1209 | |
77216b04 | 1210 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1211 | nr_writes++; |
1212 | ||
1213 | if (nr_writes >= max_nr_writes) | |
1214 | break; | |
1215 | } | |
1216 | ||
1217 | return nr_reads + nr_writes; | |
1218 | } | |
1219 | ||
651930bc | 1220 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
1221 | { |
1222 | unsigned int nr_disp = 0; | |
e43473b7 VG |
1223 | |
1224 | while (1) { | |
73f0d49a | 1225 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
2ab74cd2 | 1226 | struct throtl_service_queue *sq; |
e43473b7 VG |
1227 | |
1228 | if (!tg) | |
1229 | break; | |
1230 | ||
1231 | if (time_before(jiffies, tg->disptime)) | |
1232 | break; | |
1233 | ||
77216b04 | 1234 | throtl_dequeue_tg(tg); |
e43473b7 | 1235 | |
77216b04 | 1236 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 1237 | |
2ab74cd2 | 1238 | sq = &tg->service_queue; |
73f0d49a | 1239 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
77216b04 | 1240 | tg_update_disptime(tg); |
e43473b7 | 1241 | |
e675df2a | 1242 | if (nr_disp >= THROTL_QUANTUM) |
e43473b7 VG |
1243 | break; |
1244 | } | |
1245 | ||
1246 | return nr_disp; | |
1247 | } | |
1248 | ||
c79892c5 SL |
1249 | static bool throtl_can_upgrade(struct throtl_data *td, |
1250 | struct throtl_grp *this_tg); | |
6e1a5704 TH |
1251 | /** |
1252 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
216382dc | 1253 | * @t: the pending_timer member of the throtl_service_queue being serviced |
6e1a5704 TH |
1254 | * |
1255 | * This timer is armed when a child throtl_grp with active bio's become | |
1256 | * pending and queued on the service_queue's pending_tree and expires when | |
1257 | * the first child throtl_grp should be dispatched. This function | |
2e48a530 TH |
1258 | * dispatches bio's from the children throtl_grps to the parent |
1259 | * service_queue. | |
1260 | * | |
1261 | * If the parent's parent is another throtl_grp, dispatching is propagated | |
1262 | * by either arming its pending_timer or repeating dispatch directly. If | |
1263 | * the top-level service_tree is reached, throtl_data->dispatch_work is | |
1264 | * kicked so that the ready bio's are issued. | |
6e1a5704 | 1265 | */ |
e99e88a9 | 1266 | static void throtl_pending_timer_fn(struct timer_list *t) |
69df0ab0 | 1267 | { |
e99e88a9 | 1268 | struct throtl_service_queue *sq = from_timer(sq, t, pending_timer); |
2e48a530 | 1269 | struct throtl_grp *tg = sq_to_tg(sq); |
69df0ab0 | 1270 | struct throtl_data *td = sq_to_td(sq); |
cb76199c | 1271 | struct request_queue *q = td->queue; |
2e48a530 TH |
1272 | struct throtl_service_queue *parent_sq; |
1273 | bool dispatched; | |
6e1a5704 | 1274 | int ret; |
e43473b7 | 1275 | |
0d945c1f | 1276 | spin_lock_irq(&q->queue_lock); |
c79892c5 SL |
1277 | if (throtl_can_upgrade(td, NULL)) |
1278 | throtl_upgrade_state(td); | |
1279 | ||
2e48a530 TH |
1280 | again: |
1281 | parent_sq = sq->parent_sq; | |
1282 | dispatched = false; | |
e43473b7 | 1283 | |
7f52f98c TH |
1284 | while (true) { |
1285 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
2e48a530 TH |
1286 | sq->nr_queued[READ] + sq->nr_queued[WRITE], |
1287 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
7f52f98c TH |
1288 | |
1289 | ret = throtl_select_dispatch(sq); | |
1290 | if (ret) { | |
7f52f98c TH |
1291 | throtl_log(sq, "bios disp=%u", ret); |
1292 | dispatched = true; | |
1293 | } | |
e43473b7 | 1294 | |
7f52f98c TH |
1295 | if (throtl_schedule_next_dispatch(sq, false)) |
1296 | break; | |
e43473b7 | 1297 | |
7f52f98c | 1298 | /* this dispatch windows is still open, relax and repeat */ |
0d945c1f | 1299 | spin_unlock_irq(&q->queue_lock); |
7f52f98c | 1300 | cpu_relax(); |
0d945c1f | 1301 | spin_lock_irq(&q->queue_lock); |
651930bc | 1302 | } |
e43473b7 | 1303 | |
2e48a530 TH |
1304 | if (!dispatched) |
1305 | goto out_unlock; | |
6e1a5704 | 1306 | |
2e48a530 TH |
1307 | if (parent_sq) { |
1308 | /* @parent_sq is another throl_grp, propagate dispatch */ | |
1309 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1310 | tg_update_disptime(tg); | |
1311 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { | |
1312 | /* window is already open, repeat dispatching */ | |
1313 | sq = parent_sq; | |
1314 | tg = sq_to_tg(sq); | |
1315 | goto again; | |
1316 | } | |
1317 | } | |
1318 | } else { | |
b53b072c | 1319 | /* reached the top-level, queue issuing */ |
2e48a530 TH |
1320 | queue_work(kthrotld_workqueue, &td->dispatch_work); |
1321 | } | |
1322 | out_unlock: | |
0d945c1f | 1323 | spin_unlock_irq(&q->queue_lock); |
6e1a5704 | 1324 | } |
e43473b7 | 1325 | |
6e1a5704 TH |
1326 | /** |
1327 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
1328 | * @work: work item being executed | |
1329 | * | |
b53b072c BW |
1330 | * This function is queued for execution when bios reach the bio_lists[] |
1331 | * of throtl_data->service_queue. Those bios are ready and issued by this | |
6e1a5704 TH |
1332 | * function. |
1333 | */ | |
8876e140 | 1334 | static void blk_throtl_dispatch_work_fn(struct work_struct *work) |
6e1a5704 TH |
1335 | { |
1336 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1337 | dispatch_work); | |
1338 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1339 | struct request_queue *q = td->queue; | |
1340 | struct bio_list bio_list_on_stack; | |
1341 | struct bio *bio; | |
1342 | struct blk_plug plug; | |
1343 | int rw; | |
1344 | ||
1345 | bio_list_init(&bio_list_on_stack); | |
1346 | ||
0d945c1f | 1347 | spin_lock_irq(&q->queue_lock); |
c5cc2070 TH |
1348 | for (rw = READ; rw <= WRITE; rw++) |
1349 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) | |
1350 | bio_list_add(&bio_list_on_stack, bio); | |
0d945c1f | 1351 | spin_unlock_irq(&q->queue_lock); |
6e1a5704 TH |
1352 | |
1353 | if (!bio_list_empty(&bio_list_on_stack)) { | |
69d60eb9 | 1354 | blk_start_plug(&plug); |
ed00aabd CH |
1355 | while ((bio = bio_list_pop(&bio_list_on_stack))) |
1356 | submit_bio_noacct(bio); | |
69d60eb9 | 1357 | blk_finish_plug(&plug); |
e43473b7 | 1358 | } |
e43473b7 VG |
1359 | } |
1360 | ||
f95a04af TH |
1361 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
1362 | int off) | |
60c2bc2d | 1363 | { |
f95a04af TH |
1364 | struct throtl_grp *tg = pd_to_tg(pd); |
1365 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 1366 | |
2ab5492d | 1367 | if (v == U64_MAX) |
60c2bc2d | 1368 | return 0; |
f95a04af | 1369 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
1370 | } |
1371 | ||
f95a04af TH |
1372 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
1373 | int off) | |
e43473b7 | 1374 | { |
f95a04af TH |
1375 | struct throtl_grp *tg = pd_to_tg(pd); |
1376 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 1377 | |
2ab5492d | 1378 | if (v == UINT_MAX) |
af133ceb | 1379 | return 0; |
f95a04af | 1380 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
1381 | } |
1382 | ||
2da8ca82 | 1383 | static int tg_print_conf_u64(struct seq_file *sf, void *v) |
8e89d13f | 1384 | { |
2da8ca82 TH |
1385 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64, |
1386 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1387 | return 0; |
8e89d13f VG |
1388 | } |
1389 | ||
2da8ca82 | 1390 | static int tg_print_conf_uint(struct seq_file *sf, void *v) |
8e89d13f | 1391 | { |
2da8ca82 TH |
1392 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint, |
1393 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1394 | return 0; |
60c2bc2d TH |
1395 | } |
1396 | ||
9bb67aeb | 1397 | static void tg_conf_updated(struct throtl_grp *tg, bool global) |
60c2bc2d | 1398 | { |
69948b07 | 1399 | struct throtl_service_queue *sq = &tg->service_queue; |
492eb21b | 1400 | struct cgroup_subsys_state *pos_css; |
69948b07 | 1401 | struct blkcg_gq *blkg; |
af133ceb | 1402 | |
fda6f272 TH |
1403 | throtl_log(&tg->service_queue, |
1404 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
9f626e37 SL |
1405 | tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE), |
1406 | tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE)); | |
632b4493 | 1407 | |
693e751e TH |
1408 | /* |
1409 | * Update has_rules[] flags for the updated tg's subtree. A tg is | |
1410 | * considered to have rules if either the tg itself or any of its | |
1411 | * ancestors has rules. This identifies groups without any | |
1412 | * restrictions in the whole hierarchy and allows them to bypass | |
1413 | * blk-throttle. | |
1414 | */ | |
9bb67aeb SL |
1415 | blkg_for_each_descendant_pre(blkg, pos_css, |
1416 | global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) { | |
5b81fc3c SL |
1417 | struct throtl_grp *this_tg = blkg_to_tg(blkg); |
1418 | struct throtl_grp *parent_tg; | |
1419 | ||
1420 | tg_update_has_rules(this_tg); | |
1421 | /* ignore root/second level */ | |
1422 | if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent || | |
1423 | !blkg->parent->parent) | |
1424 | continue; | |
1425 | parent_tg = blkg_to_tg(blkg->parent); | |
1426 | /* | |
1427 | * make sure all children has lower idle time threshold and | |
1428 | * higher latency target | |
1429 | */ | |
1430 | this_tg->idletime_threshold = min(this_tg->idletime_threshold, | |
1431 | parent_tg->idletime_threshold); | |
1432 | this_tg->latency_target = max(this_tg->latency_target, | |
1433 | parent_tg->latency_target); | |
1434 | } | |
693e751e | 1435 | |
632b4493 TH |
1436 | /* |
1437 | * We're already holding queue_lock and know @tg is valid. Let's | |
1438 | * apply the new config directly. | |
1439 | * | |
1440 | * Restart the slices for both READ and WRITES. It might happen | |
1441 | * that a group's limit are dropped suddenly and we don't want to | |
1442 | * account recently dispatched IO with new low rate. | |
1443 | */ | |
ff8b22c0 BW |
1444 | throtl_start_new_slice(tg, READ); |
1445 | throtl_start_new_slice(tg, WRITE); | |
632b4493 | 1446 | |
5b2c16aa | 1447 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 1448 | tg_update_disptime(tg); |
7f52f98c | 1449 | throtl_schedule_next_dispatch(sq->parent_sq, true); |
632b4493 | 1450 | } |
69948b07 TH |
1451 | } |
1452 | ||
1453 | static ssize_t tg_set_conf(struct kernfs_open_file *of, | |
1454 | char *buf, size_t nbytes, loff_t off, bool is_u64) | |
1455 | { | |
1456 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1457 | struct blkg_conf_ctx ctx; | |
1458 | struct throtl_grp *tg; | |
1459 | int ret; | |
1460 | u64 v; | |
1461 | ||
1462 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1463 | if (ret) | |
1464 | return ret; | |
1465 | ||
1466 | ret = -EINVAL; | |
1467 | if (sscanf(ctx.body, "%llu", &v) != 1) | |
1468 | goto out_finish; | |
1469 | if (!v) | |
2ab5492d | 1470 | v = U64_MAX; |
69948b07 TH |
1471 | |
1472 | tg = blkg_to_tg(ctx.blkg); | |
1473 | ||
1474 | if (is_u64) | |
1475 | *(u64 *)((void *)tg + of_cft(of)->private) = v; | |
1476 | else | |
1477 | *(unsigned int *)((void *)tg + of_cft(of)->private) = v; | |
60c2bc2d | 1478 | |
9bb67aeb | 1479 | tg_conf_updated(tg, false); |
36aa9e5f TH |
1480 | ret = 0; |
1481 | out_finish: | |
60c2bc2d | 1482 | blkg_conf_finish(&ctx); |
36aa9e5f | 1483 | return ret ?: nbytes; |
8e89d13f VG |
1484 | } |
1485 | ||
451af504 TH |
1486 | static ssize_t tg_set_conf_u64(struct kernfs_open_file *of, |
1487 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1488 | { |
451af504 | 1489 | return tg_set_conf(of, buf, nbytes, off, true); |
60c2bc2d TH |
1490 | } |
1491 | ||
451af504 TH |
1492 | static ssize_t tg_set_conf_uint(struct kernfs_open_file *of, |
1493 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1494 | { |
451af504 | 1495 | return tg_set_conf(of, buf, nbytes, off, false); |
60c2bc2d TH |
1496 | } |
1497 | ||
7ca46438 TH |
1498 | static int tg_print_rwstat(struct seq_file *sf, void *v) |
1499 | { | |
1500 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
1501 | blkg_prfill_rwstat, &blkcg_policy_throtl, | |
1502 | seq_cft(sf)->private, true); | |
1503 | return 0; | |
1504 | } | |
1505 | ||
1506 | static u64 tg_prfill_rwstat_recursive(struct seq_file *sf, | |
1507 | struct blkg_policy_data *pd, int off) | |
1508 | { | |
1509 | struct blkg_rwstat_sample sum; | |
1510 | ||
1511 | blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off, | |
1512 | &sum); | |
1513 | return __blkg_prfill_rwstat(sf, pd, &sum); | |
1514 | } | |
1515 | ||
1516 | static int tg_print_rwstat_recursive(struct seq_file *sf, void *v) | |
1517 | { | |
1518 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
1519 | tg_prfill_rwstat_recursive, &blkcg_policy_throtl, | |
1520 | seq_cft(sf)->private, true); | |
1521 | return 0; | |
1522 | } | |
1523 | ||
880f50e2 | 1524 | static struct cftype throtl_legacy_files[] = { |
60c2bc2d TH |
1525 | { |
1526 | .name = "throttle.read_bps_device", | |
9f626e37 | 1527 | .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]), |
2da8ca82 | 1528 | .seq_show = tg_print_conf_u64, |
451af504 | 1529 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1530 | }, |
1531 | { | |
1532 | .name = "throttle.write_bps_device", | |
9f626e37 | 1533 | .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]), |
2da8ca82 | 1534 | .seq_show = tg_print_conf_u64, |
451af504 | 1535 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1536 | }, |
1537 | { | |
1538 | .name = "throttle.read_iops_device", | |
9f626e37 | 1539 | .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]), |
2da8ca82 | 1540 | .seq_show = tg_print_conf_uint, |
451af504 | 1541 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1542 | }, |
1543 | { | |
1544 | .name = "throttle.write_iops_device", | |
9f626e37 | 1545 | .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]), |
2da8ca82 | 1546 | .seq_show = tg_print_conf_uint, |
451af504 | 1547 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1548 | }, |
1549 | { | |
1550 | .name = "throttle.io_service_bytes", | |
7ca46438 TH |
1551 | .private = offsetof(struct throtl_grp, stat_bytes), |
1552 | .seq_show = tg_print_rwstat, | |
60c2bc2d | 1553 | }, |
17534c6f | 1554 | { |
1555 | .name = "throttle.io_service_bytes_recursive", | |
7ca46438 TH |
1556 | .private = offsetof(struct throtl_grp, stat_bytes), |
1557 | .seq_show = tg_print_rwstat_recursive, | |
17534c6f | 1558 | }, |
60c2bc2d TH |
1559 | { |
1560 | .name = "throttle.io_serviced", | |
7ca46438 TH |
1561 | .private = offsetof(struct throtl_grp, stat_ios), |
1562 | .seq_show = tg_print_rwstat, | |
60c2bc2d | 1563 | }, |
17534c6f | 1564 | { |
1565 | .name = "throttle.io_serviced_recursive", | |
7ca46438 TH |
1566 | .private = offsetof(struct throtl_grp, stat_ios), |
1567 | .seq_show = tg_print_rwstat_recursive, | |
17534c6f | 1568 | }, |
60c2bc2d TH |
1569 | { } /* terminate */ |
1570 | }; | |
1571 | ||
cd5ab1b0 | 1572 | static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd, |
2ee867dc TH |
1573 | int off) |
1574 | { | |
1575 | struct throtl_grp *tg = pd_to_tg(pd); | |
1576 | const char *dname = blkg_dev_name(pd->blkg); | |
1577 | char bufs[4][21] = { "max", "max", "max", "max" }; | |
cd5ab1b0 SL |
1578 | u64 bps_dft; |
1579 | unsigned int iops_dft; | |
ada75b6e | 1580 | char idle_time[26] = ""; |
ec80991d | 1581 | char latency_time[26] = ""; |
2ee867dc TH |
1582 | |
1583 | if (!dname) | |
1584 | return 0; | |
9f626e37 | 1585 | |
cd5ab1b0 SL |
1586 | if (off == LIMIT_LOW) { |
1587 | bps_dft = 0; | |
1588 | iops_dft = 0; | |
1589 | } else { | |
1590 | bps_dft = U64_MAX; | |
1591 | iops_dft = UINT_MAX; | |
1592 | } | |
1593 | ||
1594 | if (tg->bps_conf[READ][off] == bps_dft && | |
1595 | tg->bps_conf[WRITE][off] == bps_dft && | |
1596 | tg->iops_conf[READ][off] == iops_dft && | |
ada75b6e | 1597 | tg->iops_conf[WRITE][off] == iops_dft && |
ec80991d | 1598 | (off != LIMIT_LOW || |
b4f428ef | 1599 | (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD && |
5b81fc3c | 1600 | tg->latency_target_conf == DFL_LATENCY_TARGET))) |
2ee867dc TH |
1601 | return 0; |
1602 | ||
9bb67aeb | 1603 | if (tg->bps_conf[READ][off] != U64_MAX) |
9f626e37 | 1604 | snprintf(bufs[0], sizeof(bufs[0]), "%llu", |
cd5ab1b0 | 1605 | tg->bps_conf[READ][off]); |
9bb67aeb | 1606 | if (tg->bps_conf[WRITE][off] != U64_MAX) |
9f626e37 | 1607 | snprintf(bufs[1], sizeof(bufs[1]), "%llu", |
cd5ab1b0 | 1608 | tg->bps_conf[WRITE][off]); |
9bb67aeb | 1609 | if (tg->iops_conf[READ][off] != UINT_MAX) |
9f626e37 | 1610 | snprintf(bufs[2], sizeof(bufs[2]), "%u", |
cd5ab1b0 | 1611 | tg->iops_conf[READ][off]); |
9bb67aeb | 1612 | if (tg->iops_conf[WRITE][off] != UINT_MAX) |
9f626e37 | 1613 | snprintf(bufs[3], sizeof(bufs[3]), "%u", |
cd5ab1b0 | 1614 | tg->iops_conf[WRITE][off]); |
ada75b6e | 1615 | if (off == LIMIT_LOW) { |
5b81fc3c | 1616 | if (tg->idletime_threshold_conf == ULONG_MAX) |
ada75b6e SL |
1617 | strcpy(idle_time, " idle=max"); |
1618 | else | |
1619 | snprintf(idle_time, sizeof(idle_time), " idle=%lu", | |
5b81fc3c | 1620 | tg->idletime_threshold_conf); |
ec80991d | 1621 | |
5b81fc3c | 1622 | if (tg->latency_target_conf == ULONG_MAX) |
ec80991d SL |
1623 | strcpy(latency_time, " latency=max"); |
1624 | else | |
1625 | snprintf(latency_time, sizeof(latency_time), | |
5b81fc3c | 1626 | " latency=%lu", tg->latency_target_conf); |
ada75b6e | 1627 | } |
2ee867dc | 1628 | |
ec80991d SL |
1629 | seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n", |
1630 | dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time, | |
1631 | latency_time); | |
2ee867dc TH |
1632 | return 0; |
1633 | } | |
1634 | ||
cd5ab1b0 | 1635 | static int tg_print_limit(struct seq_file *sf, void *v) |
2ee867dc | 1636 | { |
cd5ab1b0 | 1637 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit, |
2ee867dc TH |
1638 | &blkcg_policy_throtl, seq_cft(sf)->private, false); |
1639 | return 0; | |
1640 | } | |
1641 | ||
cd5ab1b0 | 1642 | static ssize_t tg_set_limit(struct kernfs_open_file *of, |
2ee867dc TH |
1643 | char *buf, size_t nbytes, loff_t off) |
1644 | { | |
1645 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1646 | struct blkg_conf_ctx ctx; | |
1647 | struct throtl_grp *tg; | |
1648 | u64 v[4]; | |
ada75b6e | 1649 | unsigned long idle_time; |
ec80991d | 1650 | unsigned long latency_time; |
2ee867dc | 1651 | int ret; |
cd5ab1b0 | 1652 | int index = of_cft(of)->private; |
2ee867dc TH |
1653 | |
1654 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1655 | if (ret) | |
1656 | return ret; | |
1657 | ||
1658 | tg = blkg_to_tg(ctx.blkg); | |
1659 | ||
cd5ab1b0 SL |
1660 | v[0] = tg->bps_conf[READ][index]; |
1661 | v[1] = tg->bps_conf[WRITE][index]; | |
1662 | v[2] = tg->iops_conf[READ][index]; | |
1663 | v[3] = tg->iops_conf[WRITE][index]; | |
2ee867dc | 1664 | |
5b81fc3c SL |
1665 | idle_time = tg->idletime_threshold_conf; |
1666 | latency_time = tg->latency_target_conf; | |
2ee867dc TH |
1667 | while (true) { |
1668 | char tok[27]; /* wiops=18446744073709551616 */ | |
1669 | char *p; | |
2ab5492d | 1670 | u64 val = U64_MAX; |
2ee867dc TH |
1671 | int len; |
1672 | ||
1673 | if (sscanf(ctx.body, "%26s%n", tok, &len) != 1) | |
1674 | break; | |
1675 | if (tok[0] == '\0') | |
1676 | break; | |
1677 | ctx.body += len; | |
1678 | ||
1679 | ret = -EINVAL; | |
1680 | p = tok; | |
1681 | strsep(&p, "="); | |
1682 | if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max"))) | |
1683 | goto out_finish; | |
1684 | ||
1685 | ret = -ERANGE; | |
1686 | if (!val) | |
1687 | goto out_finish; | |
1688 | ||
1689 | ret = -EINVAL; | |
1690 | if (!strcmp(tok, "rbps")) | |
1691 | v[0] = val; | |
1692 | else if (!strcmp(tok, "wbps")) | |
1693 | v[1] = val; | |
1694 | else if (!strcmp(tok, "riops")) | |
1695 | v[2] = min_t(u64, val, UINT_MAX); | |
1696 | else if (!strcmp(tok, "wiops")) | |
1697 | v[3] = min_t(u64, val, UINT_MAX); | |
ada75b6e SL |
1698 | else if (off == LIMIT_LOW && !strcmp(tok, "idle")) |
1699 | idle_time = val; | |
ec80991d SL |
1700 | else if (off == LIMIT_LOW && !strcmp(tok, "latency")) |
1701 | latency_time = val; | |
2ee867dc TH |
1702 | else |
1703 | goto out_finish; | |
1704 | } | |
1705 | ||
cd5ab1b0 SL |
1706 | tg->bps_conf[READ][index] = v[0]; |
1707 | tg->bps_conf[WRITE][index] = v[1]; | |
1708 | tg->iops_conf[READ][index] = v[2]; | |
1709 | tg->iops_conf[WRITE][index] = v[3]; | |
2ee867dc | 1710 | |
cd5ab1b0 SL |
1711 | if (index == LIMIT_MAX) { |
1712 | tg->bps[READ][index] = v[0]; | |
1713 | tg->bps[WRITE][index] = v[1]; | |
1714 | tg->iops[READ][index] = v[2]; | |
1715 | tg->iops[WRITE][index] = v[3]; | |
1716 | } | |
1717 | tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW], | |
1718 | tg->bps_conf[READ][LIMIT_MAX]); | |
1719 | tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW], | |
1720 | tg->bps_conf[WRITE][LIMIT_MAX]); | |
1721 | tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW], | |
1722 | tg->iops_conf[READ][LIMIT_MAX]); | |
1723 | tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW], | |
1724 | tg->iops_conf[WRITE][LIMIT_MAX]); | |
b4f428ef SL |
1725 | tg->idletime_threshold_conf = idle_time; |
1726 | tg->latency_target_conf = latency_time; | |
1727 | ||
1728 | /* force user to configure all settings for low limit */ | |
1729 | if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] || | |
1730 | tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) || | |
1731 | tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD || | |
1732 | tg->latency_target_conf == DFL_LATENCY_TARGET) { | |
1733 | tg->bps[READ][LIMIT_LOW] = 0; | |
1734 | tg->bps[WRITE][LIMIT_LOW] = 0; | |
1735 | tg->iops[READ][LIMIT_LOW] = 0; | |
1736 | tg->iops[WRITE][LIMIT_LOW] = 0; | |
1737 | tg->idletime_threshold = DFL_IDLE_THRESHOLD; | |
1738 | tg->latency_target = DFL_LATENCY_TARGET; | |
1739 | } else if (index == LIMIT_LOW) { | |
5b81fc3c | 1740 | tg->idletime_threshold = tg->idletime_threshold_conf; |
5b81fc3c | 1741 | tg->latency_target = tg->latency_target_conf; |
cd5ab1b0 | 1742 | } |
b4f428ef SL |
1743 | |
1744 | blk_throtl_update_limit_valid(tg->td); | |
1745 | if (tg->td->limit_valid[LIMIT_LOW]) { | |
1746 | if (index == LIMIT_LOW) | |
1747 | tg->td->limit_index = LIMIT_LOW; | |
1748 | } else | |
1749 | tg->td->limit_index = LIMIT_MAX; | |
9bb67aeb SL |
1750 | tg_conf_updated(tg, index == LIMIT_LOW && |
1751 | tg->td->limit_valid[LIMIT_LOW]); | |
2ee867dc TH |
1752 | ret = 0; |
1753 | out_finish: | |
1754 | blkg_conf_finish(&ctx); | |
1755 | return ret ?: nbytes; | |
1756 | } | |
1757 | ||
1758 | static struct cftype throtl_files[] = { | |
cd5ab1b0 SL |
1759 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
1760 | { | |
1761 | .name = "low", | |
1762 | .flags = CFTYPE_NOT_ON_ROOT, | |
1763 | .seq_show = tg_print_limit, | |
1764 | .write = tg_set_limit, | |
1765 | .private = LIMIT_LOW, | |
1766 | }, | |
1767 | #endif | |
2ee867dc TH |
1768 | { |
1769 | .name = "max", | |
1770 | .flags = CFTYPE_NOT_ON_ROOT, | |
cd5ab1b0 SL |
1771 | .seq_show = tg_print_limit, |
1772 | .write = tg_set_limit, | |
1773 | .private = LIMIT_MAX, | |
2ee867dc TH |
1774 | }, |
1775 | { } /* terminate */ | |
1776 | }; | |
1777 | ||
da527770 | 1778 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1779 | { |
1780 | struct throtl_data *td = q->td; | |
1781 | ||
69df0ab0 | 1782 | cancel_work_sync(&td->dispatch_work); |
e43473b7 VG |
1783 | } |
1784 | ||
3c798398 | 1785 | static struct blkcg_policy blkcg_policy_throtl = { |
2ee867dc | 1786 | .dfl_cftypes = throtl_files, |
880f50e2 | 1787 | .legacy_cftypes = throtl_legacy_files, |
f9fcc2d3 | 1788 | |
001bea73 | 1789 | .pd_alloc_fn = throtl_pd_alloc, |
f9fcc2d3 | 1790 | .pd_init_fn = throtl_pd_init, |
693e751e | 1791 | .pd_online_fn = throtl_pd_online, |
cd5ab1b0 | 1792 | .pd_offline_fn = throtl_pd_offline, |
001bea73 | 1793 | .pd_free_fn = throtl_pd_free, |
e43473b7 VG |
1794 | }; |
1795 | ||
3f0abd80 SL |
1796 | static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg) |
1797 | { | |
1798 | unsigned long rtime = jiffies, wtime = jiffies; | |
1799 | ||
1800 | if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]) | |
1801 | rtime = tg->last_low_overflow_time[READ]; | |
1802 | if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) | |
1803 | wtime = tg->last_low_overflow_time[WRITE]; | |
1804 | return min(rtime, wtime); | |
1805 | } | |
1806 | ||
1807 | /* tg should not be an intermediate node */ | |
1808 | static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg) | |
1809 | { | |
1810 | struct throtl_service_queue *parent_sq; | |
1811 | struct throtl_grp *parent = tg; | |
1812 | unsigned long ret = __tg_last_low_overflow_time(tg); | |
1813 | ||
1814 | while (true) { | |
1815 | parent_sq = parent->service_queue.parent_sq; | |
1816 | parent = sq_to_tg(parent_sq); | |
1817 | if (!parent) | |
1818 | break; | |
1819 | ||
1820 | /* | |
1821 | * The parent doesn't have low limit, it always reaches low | |
1822 | * limit. Its overflow time is useless for children | |
1823 | */ | |
1824 | if (!parent->bps[READ][LIMIT_LOW] && | |
1825 | !parent->iops[READ][LIMIT_LOW] && | |
1826 | !parent->bps[WRITE][LIMIT_LOW] && | |
1827 | !parent->iops[WRITE][LIMIT_LOW]) | |
1828 | continue; | |
1829 | if (time_after(__tg_last_low_overflow_time(parent), ret)) | |
1830 | ret = __tg_last_low_overflow_time(parent); | |
1831 | } | |
1832 | return ret; | |
1833 | } | |
1834 | ||
9e234eea SL |
1835 | static bool throtl_tg_is_idle(struct throtl_grp *tg) |
1836 | { | |
1837 | /* | |
1838 | * cgroup is idle if: | |
1839 | * - single idle is too long, longer than a fixed value (in case user | |
b4f428ef | 1840 | * configure a too big threshold) or 4 times of idletime threshold |
9e234eea | 1841 | * - average think time is more than threshold |
53696b8d | 1842 | * - IO latency is largely below threshold |
9e234eea | 1843 | */ |
b4f428ef | 1844 | unsigned long time; |
4cff729f | 1845 | bool ret; |
9e234eea | 1846 | |
b4f428ef SL |
1847 | time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold); |
1848 | ret = tg->latency_target == DFL_LATENCY_TARGET || | |
1849 | tg->idletime_threshold == DFL_IDLE_THRESHOLD || | |
1850 | (ktime_get_ns() >> 10) - tg->last_finish_time > time || | |
1851 | tg->avg_idletime > tg->idletime_threshold || | |
1852 | (tg->latency_target && tg->bio_cnt && | |
53696b8d | 1853 | tg->bad_bio_cnt * 5 < tg->bio_cnt); |
4cff729f SL |
1854 | throtl_log(&tg->service_queue, |
1855 | "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d", | |
1856 | tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt, | |
1857 | tg->bio_cnt, ret, tg->td->scale); | |
1858 | return ret; | |
9e234eea SL |
1859 | } |
1860 | ||
c79892c5 SL |
1861 | static bool throtl_tg_can_upgrade(struct throtl_grp *tg) |
1862 | { | |
1863 | struct throtl_service_queue *sq = &tg->service_queue; | |
1864 | bool read_limit, write_limit; | |
1865 | ||
1866 | /* | |
1867 | * if cgroup reaches low limit (if low limit is 0, the cgroup always | |
1868 | * reaches), it's ok to upgrade to next limit | |
1869 | */ | |
1870 | read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]; | |
1871 | write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]; | |
1872 | if (!read_limit && !write_limit) | |
1873 | return true; | |
1874 | if (read_limit && sq->nr_queued[READ] && | |
1875 | (!write_limit || sq->nr_queued[WRITE])) | |
1876 | return true; | |
1877 | if (write_limit && sq->nr_queued[WRITE] && | |
1878 | (!read_limit || sq->nr_queued[READ])) | |
1879 | return true; | |
aec24246 SL |
1880 | |
1881 | if (time_after_eq(jiffies, | |
fa6fb5aa SL |
1882 | tg_last_low_overflow_time(tg) + tg->td->throtl_slice) && |
1883 | throtl_tg_is_idle(tg)) | |
aec24246 | 1884 | return true; |
c79892c5 SL |
1885 | return false; |
1886 | } | |
1887 | ||
1888 | static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg) | |
1889 | { | |
1890 | while (true) { | |
1891 | if (throtl_tg_can_upgrade(tg)) | |
1892 | return true; | |
1893 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1894 | if (!tg || !tg_to_blkg(tg)->parent) | |
1895 | return false; | |
1896 | } | |
1897 | return false; | |
1898 | } | |
1899 | ||
1900 | static bool throtl_can_upgrade(struct throtl_data *td, | |
1901 | struct throtl_grp *this_tg) | |
1902 | { | |
1903 | struct cgroup_subsys_state *pos_css; | |
1904 | struct blkcg_gq *blkg; | |
1905 | ||
1906 | if (td->limit_index != LIMIT_LOW) | |
1907 | return false; | |
1908 | ||
297e3d85 | 1909 | if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice)) |
3f0abd80 SL |
1910 | return false; |
1911 | ||
c79892c5 SL |
1912 | rcu_read_lock(); |
1913 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1914 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1915 | ||
1916 | if (tg == this_tg) | |
1917 | continue; | |
1918 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1919 | continue; | |
1920 | if (!throtl_hierarchy_can_upgrade(tg)) { | |
1921 | rcu_read_unlock(); | |
1922 | return false; | |
1923 | } | |
1924 | } | |
1925 | rcu_read_unlock(); | |
1926 | return true; | |
1927 | } | |
1928 | ||
fa6fb5aa SL |
1929 | static void throtl_upgrade_check(struct throtl_grp *tg) |
1930 | { | |
1931 | unsigned long now = jiffies; | |
1932 | ||
1933 | if (tg->td->limit_index != LIMIT_LOW) | |
1934 | return; | |
1935 | ||
1936 | if (time_after(tg->last_check_time + tg->td->throtl_slice, now)) | |
1937 | return; | |
1938 | ||
1939 | tg->last_check_time = now; | |
1940 | ||
1941 | if (!time_after_eq(now, | |
1942 | __tg_last_low_overflow_time(tg) + tg->td->throtl_slice)) | |
1943 | return; | |
1944 | ||
1945 | if (throtl_can_upgrade(tg->td, NULL)) | |
1946 | throtl_upgrade_state(tg->td); | |
1947 | } | |
1948 | ||
c79892c5 SL |
1949 | static void throtl_upgrade_state(struct throtl_data *td) |
1950 | { | |
1951 | struct cgroup_subsys_state *pos_css; | |
1952 | struct blkcg_gq *blkg; | |
1953 | ||
4cff729f | 1954 | throtl_log(&td->service_queue, "upgrade to max"); |
c79892c5 | 1955 | td->limit_index = LIMIT_MAX; |
3f0abd80 | 1956 | td->low_upgrade_time = jiffies; |
7394e31f | 1957 | td->scale = 0; |
c79892c5 SL |
1958 | rcu_read_lock(); |
1959 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1960 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1961 | struct throtl_service_queue *sq = &tg->service_queue; | |
1962 | ||
1963 | tg->disptime = jiffies - 1; | |
1964 | throtl_select_dispatch(sq); | |
4f02fb76 | 1965 | throtl_schedule_next_dispatch(sq, true); |
c79892c5 SL |
1966 | } |
1967 | rcu_read_unlock(); | |
1968 | throtl_select_dispatch(&td->service_queue); | |
4f02fb76 | 1969 | throtl_schedule_next_dispatch(&td->service_queue, true); |
c79892c5 SL |
1970 | queue_work(kthrotld_workqueue, &td->dispatch_work); |
1971 | } | |
1972 | ||
4247d9c8 | 1973 | static void throtl_downgrade_state(struct throtl_data *td) |
3f0abd80 | 1974 | { |
7394e31f SL |
1975 | td->scale /= 2; |
1976 | ||
4cff729f | 1977 | throtl_log(&td->service_queue, "downgrade, scale %d", td->scale); |
7394e31f SL |
1978 | if (td->scale) { |
1979 | td->low_upgrade_time = jiffies - td->scale * td->throtl_slice; | |
1980 | return; | |
1981 | } | |
1982 | ||
4247d9c8 | 1983 | td->limit_index = LIMIT_LOW; |
3f0abd80 SL |
1984 | td->low_downgrade_time = jiffies; |
1985 | } | |
1986 | ||
1987 | static bool throtl_tg_can_downgrade(struct throtl_grp *tg) | |
1988 | { | |
1989 | struct throtl_data *td = tg->td; | |
1990 | unsigned long now = jiffies; | |
1991 | ||
1992 | /* | |
1993 | * If cgroup is below low limit, consider downgrade and throttle other | |
1994 | * cgroups | |
1995 | */ | |
297e3d85 SL |
1996 | if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) && |
1997 | time_after_eq(now, tg_last_low_overflow_time(tg) + | |
fa6fb5aa SL |
1998 | td->throtl_slice) && |
1999 | (!throtl_tg_is_idle(tg) || | |
2000 | !list_empty(&tg_to_blkg(tg)->blkcg->css.children))) | |
3f0abd80 SL |
2001 | return true; |
2002 | return false; | |
2003 | } | |
2004 | ||
2005 | static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg) | |
2006 | { | |
2007 | while (true) { | |
2008 | if (!throtl_tg_can_downgrade(tg)) | |
2009 | return false; | |
2010 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
2011 | if (!tg || !tg_to_blkg(tg)->parent) | |
2012 | break; | |
2013 | } | |
2014 | return true; | |
2015 | } | |
2016 | ||
2017 | static void throtl_downgrade_check(struct throtl_grp *tg) | |
2018 | { | |
2019 | uint64_t bps; | |
2020 | unsigned int iops; | |
2021 | unsigned long elapsed_time; | |
2022 | unsigned long now = jiffies; | |
2023 | ||
2024 | if (tg->td->limit_index != LIMIT_MAX || | |
2025 | !tg->td->limit_valid[LIMIT_LOW]) | |
2026 | return; | |
2027 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
2028 | return; | |
297e3d85 | 2029 | if (time_after(tg->last_check_time + tg->td->throtl_slice, now)) |
3f0abd80 SL |
2030 | return; |
2031 | ||
2032 | elapsed_time = now - tg->last_check_time; | |
2033 | tg->last_check_time = now; | |
2034 | ||
297e3d85 SL |
2035 | if (time_before(now, tg_last_low_overflow_time(tg) + |
2036 | tg->td->throtl_slice)) | |
3f0abd80 SL |
2037 | return; |
2038 | ||
2039 | if (tg->bps[READ][LIMIT_LOW]) { | |
2040 | bps = tg->last_bytes_disp[READ] * HZ; | |
2041 | do_div(bps, elapsed_time); | |
2042 | if (bps >= tg->bps[READ][LIMIT_LOW]) | |
2043 | tg->last_low_overflow_time[READ] = now; | |
2044 | } | |
2045 | ||
2046 | if (tg->bps[WRITE][LIMIT_LOW]) { | |
2047 | bps = tg->last_bytes_disp[WRITE] * HZ; | |
2048 | do_div(bps, elapsed_time); | |
2049 | if (bps >= tg->bps[WRITE][LIMIT_LOW]) | |
2050 | tg->last_low_overflow_time[WRITE] = now; | |
2051 | } | |
2052 | ||
2053 | if (tg->iops[READ][LIMIT_LOW]) { | |
2054 | iops = tg->last_io_disp[READ] * HZ / elapsed_time; | |
2055 | if (iops >= tg->iops[READ][LIMIT_LOW]) | |
2056 | tg->last_low_overflow_time[READ] = now; | |
2057 | } | |
2058 | ||
2059 | if (tg->iops[WRITE][LIMIT_LOW]) { | |
2060 | iops = tg->last_io_disp[WRITE] * HZ / elapsed_time; | |
2061 | if (iops >= tg->iops[WRITE][LIMIT_LOW]) | |
2062 | tg->last_low_overflow_time[WRITE] = now; | |
2063 | } | |
2064 | ||
2065 | /* | |
2066 | * If cgroup is below low limit, consider downgrade and throttle other | |
2067 | * cgroups | |
2068 | */ | |
2069 | if (throtl_hierarchy_can_downgrade(tg)) | |
4247d9c8 | 2070 | throtl_downgrade_state(tg->td); |
3f0abd80 SL |
2071 | |
2072 | tg->last_bytes_disp[READ] = 0; | |
2073 | tg->last_bytes_disp[WRITE] = 0; | |
2074 | tg->last_io_disp[READ] = 0; | |
2075 | tg->last_io_disp[WRITE] = 0; | |
2076 | } | |
2077 | ||
9e234eea SL |
2078 | static void blk_throtl_update_idletime(struct throtl_grp *tg) |
2079 | { | |
2080 | unsigned long now = ktime_get_ns() >> 10; | |
2081 | unsigned long last_finish_time = tg->last_finish_time; | |
2082 | ||
2083 | if (now <= last_finish_time || last_finish_time == 0 || | |
2084 | last_finish_time == tg->checked_last_finish_time) | |
2085 | return; | |
2086 | ||
2087 | tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3; | |
2088 | tg->checked_last_finish_time = last_finish_time; | |
2089 | } | |
2090 | ||
b9147dd1 SL |
2091 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
2092 | static void throtl_update_latency_buckets(struct throtl_data *td) | |
2093 | { | |
b889bf66 JQ |
2094 | struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE]; |
2095 | int i, cpu, rw; | |
2096 | unsigned long last_latency[2] = { 0 }; | |
2097 | unsigned long latency[2]; | |
b9147dd1 SL |
2098 | |
2099 | if (!blk_queue_nonrot(td->queue)) | |
2100 | return; | |
2101 | if (time_before(jiffies, td->last_calculate_time + HZ)) | |
2102 | return; | |
2103 | td->last_calculate_time = jiffies; | |
2104 | ||
2105 | memset(avg_latency, 0, sizeof(avg_latency)); | |
b889bf66 JQ |
2106 | for (rw = READ; rw <= WRITE; rw++) { |
2107 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) { | |
2108 | struct latency_bucket *tmp = &td->tmp_buckets[rw][i]; | |
2109 | ||
2110 | for_each_possible_cpu(cpu) { | |
2111 | struct latency_bucket *bucket; | |
2112 | ||
2113 | /* this isn't race free, but ok in practice */ | |
2114 | bucket = per_cpu_ptr(td->latency_buckets[rw], | |
2115 | cpu); | |
2116 | tmp->total_latency += bucket[i].total_latency; | |
2117 | tmp->samples += bucket[i].samples; | |
2118 | bucket[i].total_latency = 0; | |
2119 | bucket[i].samples = 0; | |
2120 | } | |
b9147dd1 | 2121 | |
b889bf66 JQ |
2122 | if (tmp->samples >= 32) { |
2123 | int samples = tmp->samples; | |
b9147dd1 | 2124 | |
b889bf66 | 2125 | latency[rw] = tmp->total_latency; |
b9147dd1 | 2126 | |
b889bf66 JQ |
2127 | tmp->total_latency = 0; |
2128 | tmp->samples = 0; | |
2129 | latency[rw] /= samples; | |
2130 | if (latency[rw] == 0) | |
2131 | continue; | |
2132 | avg_latency[rw][i].latency = latency[rw]; | |
2133 | } | |
b9147dd1 SL |
2134 | } |
2135 | } | |
2136 | ||
b889bf66 JQ |
2137 | for (rw = READ; rw <= WRITE; rw++) { |
2138 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) { | |
2139 | if (!avg_latency[rw][i].latency) { | |
2140 | if (td->avg_buckets[rw][i].latency < last_latency[rw]) | |
2141 | td->avg_buckets[rw][i].latency = | |
2142 | last_latency[rw]; | |
2143 | continue; | |
2144 | } | |
b9147dd1 | 2145 | |
b889bf66 JQ |
2146 | if (!td->avg_buckets[rw][i].valid) |
2147 | latency[rw] = avg_latency[rw][i].latency; | |
2148 | else | |
2149 | latency[rw] = (td->avg_buckets[rw][i].latency * 7 + | |
2150 | avg_latency[rw][i].latency) >> 3; | |
b9147dd1 | 2151 | |
b889bf66 JQ |
2152 | td->avg_buckets[rw][i].latency = max(latency[rw], |
2153 | last_latency[rw]); | |
2154 | td->avg_buckets[rw][i].valid = true; | |
2155 | last_latency[rw] = td->avg_buckets[rw][i].latency; | |
2156 | } | |
b9147dd1 | 2157 | } |
4cff729f SL |
2158 | |
2159 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) | |
2160 | throtl_log(&td->service_queue, | |
b889bf66 JQ |
2161 | "Latency bucket %d: read latency=%ld, read valid=%d, " |
2162 | "write latency=%ld, write valid=%d", i, | |
2163 | td->avg_buckets[READ][i].latency, | |
2164 | td->avg_buckets[READ][i].valid, | |
2165 | td->avg_buckets[WRITE][i].latency, | |
2166 | td->avg_buckets[WRITE][i].valid); | |
b9147dd1 SL |
2167 | } |
2168 | #else | |
2169 | static inline void throtl_update_latency_buckets(struct throtl_data *td) | |
2170 | { | |
2171 | } | |
2172 | #endif | |
2173 | ||
db18a53e | 2174 | bool blk_throtl_bio(struct bio *bio) |
e43473b7 | 2175 | { |
db18a53e CH |
2176 | struct request_queue *q = bio->bi_disk->queue; |
2177 | struct blkcg_gq *blkg = bio->bi_blkg; | |
c5cc2070 | 2178 | struct throtl_qnode *qn = NULL; |
a2e83ef9 | 2179 | struct throtl_grp *tg = blkg_to_tg(blkg); |
73f0d49a | 2180 | struct throtl_service_queue *sq; |
0e9f4164 | 2181 | bool rw = bio_data_dir(bio); |
bc16a4f9 | 2182 | bool throttled = false; |
b9147dd1 | 2183 | struct throtl_data *td = tg->td; |
e43473b7 | 2184 | |
93b80638 | 2185 | rcu_read_lock(); |
ae118896 | 2186 | |
2a0f61e6 | 2187 | /* see throtl_charge_bio() */ |
7ca46438 TH |
2188 | if (bio_flagged(bio, BIO_THROTTLED)) |
2189 | goto out; | |
2190 | ||
2191 | if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) { | |
2192 | blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf, | |
2193 | bio->bi_iter.bi_size); | |
2194 | blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1); | |
2195 | } | |
2196 | ||
2197 | if (!tg->has_rules[rw]) | |
bc16a4f9 | 2198 | goto out; |
e43473b7 | 2199 | |
0d945c1f | 2200 | spin_lock_irq(&q->queue_lock); |
c9589f03 | 2201 | |
b9147dd1 SL |
2202 | throtl_update_latency_buckets(td); |
2203 | ||
9e234eea SL |
2204 | blk_throtl_update_idletime(tg); |
2205 | ||
73f0d49a TH |
2206 | sq = &tg->service_queue; |
2207 | ||
c79892c5 | 2208 | again: |
9e660acf | 2209 | while (true) { |
3f0abd80 SL |
2210 | if (tg->last_low_overflow_time[rw] == 0) |
2211 | tg->last_low_overflow_time[rw] = jiffies; | |
2212 | throtl_downgrade_check(tg); | |
fa6fb5aa | 2213 | throtl_upgrade_check(tg); |
9e660acf TH |
2214 | /* throtl is FIFO - if bios are already queued, should queue */ |
2215 | if (sq->nr_queued[rw]) | |
2216 | break; | |
de701c74 | 2217 | |
9e660acf | 2218 | /* if above limits, break to queue */ |
c79892c5 | 2219 | if (!tg_may_dispatch(tg, bio, NULL)) { |
3f0abd80 | 2220 | tg->last_low_overflow_time[rw] = jiffies; |
b9147dd1 SL |
2221 | if (throtl_can_upgrade(td, tg)) { |
2222 | throtl_upgrade_state(td); | |
c79892c5 SL |
2223 | goto again; |
2224 | } | |
9e660acf | 2225 | break; |
c79892c5 | 2226 | } |
9e660acf TH |
2227 | |
2228 | /* within limits, let's charge and dispatch directly */ | |
e43473b7 | 2229 | throtl_charge_bio(tg, bio); |
04521db0 VG |
2230 | |
2231 | /* | |
2232 | * We need to trim slice even when bios are not being queued | |
2233 | * otherwise it might happen that a bio is not queued for | |
2234 | * a long time and slice keeps on extending and trim is not | |
2235 | * called for a long time. Now if limits are reduced suddenly | |
2236 | * we take into account all the IO dispatched so far at new | |
2237 | * low rate and * newly queued IO gets a really long dispatch | |
2238 | * time. | |
2239 | * | |
2240 | * So keep on trimming slice even if bio is not queued. | |
2241 | */ | |
0f3457f6 | 2242 | throtl_trim_slice(tg, rw); |
9e660acf TH |
2243 | |
2244 | /* | |
2245 | * @bio passed through this layer without being throttled. | |
b53b072c | 2246 | * Climb up the ladder. If we're already at the top, it |
9e660acf TH |
2247 | * can be executed directly. |
2248 | */ | |
c5cc2070 | 2249 | qn = &tg->qnode_on_parent[rw]; |
9e660acf TH |
2250 | sq = sq->parent_sq; |
2251 | tg = sq_to_tg(sq); | |
2252 | if (!tg) | |
2253 | goto out_unlock; | |
e43473b7 VG |
2254 | } |
2255 | ||
9e660acf | 2256 | /* out-of-limit, queue to @tg */ |
fda6f272 TH |
2257 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
2258 | rw == READ ? 'R' : 'W', | |
9f626e37 SL |
2259 | tg->bytes_disp[rw], bio->bi_iter.bi_size, |
2260 | tg_bps_limit(tg, rw), | |
2261 | tg->io_disp[rw], tg_iops_limit(tg, rw), | |
fda6f272 | 2262 | sq->nr_queued[READ], sq->nr_queued[WRITE]); |
e43473b7 | 2263 | |
3f0abd80 SL |
2264 | tg->last_low_overflow_time[rw] = jiffies; |
2265 | ||
b9147dd1 | 2266 | td->nr_queued[rw]++; |
c5cc2070 | 2267 | throtl_add_bio_tg(bio, qn, tg); |
bc16a4f9 | 2268 | throttled = true; |
e43473b7 | 2269 | |
7f52f98c TH |
2270 | /* |
2271 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
2272 | * was empty before @bio. The forced scheduling isn't likely to | |
2273 | * cause undue delay as @bio is likely to be dispatched directly if | |
2274 | * its @tg's disptime is not in the future. | |
2275 | */ | |
0e9f4164 | 2276 | if (tg->flags & THROTL_TG_WAS_EMPTY) { |
77216b04 | 2277 | tg_update_disptime(tg); |
7f52f98c | 2278 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); |
e43473b7 VG |
2279 | } |
2280 | ||
bc16a4f9 | 2281 | out_unlock: |
0d945c1f | 2282 | spin_unlock_irq(&q->queue_lock); |
bc16a4f9 | 2283 | out: |
111be883 | 2284 | bio_set_flag(bio, BIO_THROTTLED); |
b9147dd1 SL |
2285 | |
2286 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
2287 | if (throttled || !td->track_bio_latency) | |
5238dcf4 | 2288 | bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY; |
b9147dd1 | 2289 | #endif |
93b80638 | 2290 | rcu_read_unlock(); |
bc16a4f9 | 2291 | return throttled; |
e43473b7 VG |
2292 | } |
2293 | ||
9e234eea | 2294 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
b9147dd1 SL |
2295 | static void throtl_track_latency(struct throtl_data *td, sector_t size, |
2296 | int op, unsigned long time) | |
2297 | { | |
2298 | struct latency_bucket *latency; | |
2299 | int index; | |
2300 | ||
b889bf66 JQ |
2301 | if (!td || td->limit_index != LIMIT_LOW || |
2302 | !(op == REQ_OP_READ || op == REQ_OP_WRITE) || | |
b9147dd1 SL |
2303 | !blk_queue_nonrot(td->queue)) |
2304 | return; | |
2305 | ||
2306 | index = request_bucket_index(size); | |
2307 | ||
b889bf66 | 2308 | latency = get_cpu_ptr(td->latency_buckets[op]); |
b9147dd1 SL |
2309 | latency[index].total_latency += time; |
2310 | latency[index].samples++; | |
b889bf66 | 2311 | put_cpu_ptr(td->latency_buckets[op]); |
b9147dd1 SL |
2312 | } |
2313 | ||
2314 | void blk_throtl_stat_add(struct request *rq, u64 time_ns) | |
2315 | { | |
2316 | struct request_queue *q = rq->q; | |
2317 | struct throtl_data *td = q->td; | |
2318 | ||
3d244306 HT |
2319 | throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq), |
2320 | time_ns >> 10); | |
b9147dd1 SL |
2321 | } |
2322 | ||
9e234eea SL |
2323 | void blk_throtl_bio_endio(struct bio *bio) |
2324 | { | |
08e18eab | 2325 | struct blkcg_gq *blkg; |
9e234eea | 2326 | struct throtl_grp *tg; |
b9147dd1 SL |
2327 | u64 finish_time_ns; |
2328 | unsigned long finish_time; | |
2329 | unsigned long start_time; | |
2330 | unsigned long lat; | |
b889bf66 | 2331 | int rw = bio_data_dir(bio); |
9e234eea | 2332 | |
08e18eab JB |
2333 | blkg = bio->bi_blkg; |
2334 | if (!blkg) | |
9e234eea | 2335 | return; |
08e18eab | 2336 | tg = blkg_to_tg(blkg); |
9e234eea | 2337 | |
b9147dd1 SL |
2338 | finish_time_ns = ktime_get_ns(); |
2339 | tg->last_finish_time = finish_time_ns >> 10; | |
2340 | ||
5238dcf4 OS |
2341 | start_time = bio_issue_time(&bio->bi_issue) >> 10; |
2342 | finish_time = __bio_issue_time(finish_time_ns) >> 10; | |
08e18eab | 2343 | if (!start_time || finish_time <= start_time) |
53696b8d SL |
2344 | return; |
2345 | ||
2346 | lat = finish_time - start_time; | |
b9147dd1 | 2347 | /* this is only for bio based driver */ |
5238dcf4 OS |
2348 | if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY)) |
2349 | throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue), | |
2350 | bio_op(bio), lat); | |
53696b8d | 2351 | |
6679a90c | 2352 | if (tg->latency_target && lat >= tg->td->filtered_latency) { |
53696b8d SL |
2353 | int bucket; |
2354 | unsigned int threshold; | |
2355 | ||
5238dcf4 | 2356 | bucket = request_bucket_index(bio_issue_size(&bio->bi_issue)); |
b889bf66 | 2357 | threshold = tg->td->avg_buckets[rw][bucket].latency + |
53696b8d SL |
2358 | tg->latency_target; |
2359 | if (lat > threshold) | |
2360 | tg->bad_bio_cnt++; | |
2361 | /* | |
2362 | * Not race free, could get wrong count, which means cgroups | |
2363 | * will be throttled | |
2364 | */ | |
2365 | tg->bio_cnt++; | |
2366 | } | |
2367 | ||
2368 | if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) { | |
2369 | tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies; | |
2370 | tg->bio_cnt /= 2; | |
2371 | tg->bad_bio_cnt /= 2; | |
b9147dd1 | 2372 | } |
9e234eea SL |
2373 | } |
2374 | #endif | |
2375 | ||
e43473b7 VG |
2376 | int blk_throtl_init(struct request_queue *q) |
2377 | { | |
2378 | struct throtl_data *td; | |
a2b1693b | 2379 | int ret; |
e43473b7 VG |
2380 | |
2381 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
2382 | if (!td) | |
2383 | return -ENOMEM; | |
b889bf66 | 2384 | td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) * |
b9147dd1 | 2385 | LATENCY_BUCKET_SIZE, __alignof__(u64)); |
b889bf66 JQ |
2386 | if (!td->latency_buckets[READ]) { |
2387 | kfree(td); | |
2388 | return -ENOMEM; | |
2389 | } | |
2390 | td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) * | |
b9147dd1 | 2391 | LATENCY_BUCKET_SIZE, __alignof__(u64)); |
b889bf66 JQ |
2392 | if (!td->latency_buckets[WRITE]) { |
2393 | free_percpu(td->latency_buckets[READ]); | |
b9147dd1 SL |
2394 | kfree(td); |
2395 | return -ENOMEM; | |
2396 | } | |
e43473b7 | 2397 | |
69df0ab0 | 2398 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
b2ce2643 | 2399 | throtl_service_queue_init(&td->service_queue); |
e43473b7 | 2400 | |
cd1604fa | 2401 | q->td = td; |
29b12589 | 2402 | td->queue = q; |
02977e4a | 2403 | |
9f626e37 | 2404 | td->limit_valid[LIMIT_MAX] = true; |
cd5ab1b0 | 2405 | td->limit_index = LIMIT_MAX; |
3f0abd80 SL |
2406 | td->low_upgrade_time = jiffies; |
2407 | td->low_downgrade_time = jiffies; | |
9e234eea | 2408 | |
a2b1693b | 2409 | /* activate policy */ |
3c798398 | 2410 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
b9147dd1 | 2411 | if (ret) { |
b889bf66 JQ |
2412 | free_percpu(td->latency_buckets[READ]); |
2413 | free_percpu(td->latency_buckets[WRITE]); | |
f51b802c | 2414 | kfree(td); |
b9147dd1 | 2415 | } |
a2b1693b | 2416 | return ret; |
e43473b7 VG |
2417 | } |
2418 | ||
2419 | void blk_throtl_exit(struct request_queue *q) | |
2420 | { | |
c875f4d0 | 2421 | BUG_ON(!q->td); |
da527770 | 2422 | throtl_shutdown_wq(q); |
3c798398 | 2423 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
b889bf66 JQ |
2424 | free_percpu(q->td->latency_buckets[READ]); |
2425 | free_percpu(q->td->latency_buckets[WRITE]); | |
c9a929dd | 2426 | kfree(q->td); |
e43473b7 VG |
2427 | } |
2428 | ||
d61fcfa4 SL |
2429 | void blk_throtl_register_queue(struct request_queue *q) |
2430 | { | |
2431 | struct throtl_data *td; | |
6679a90c | 2432 | int i; |
d61fcfa4 SL |
2433 | |
2434 | td = q->td; | |
2435 | BUG_ON(!td); | |
2436 | ||
6679a90c | 2437 | if (blk_queue_nonrot(q)) { |
d61fcfa4 | 2438 | td->throtl_slice = DFL_THROTL_SLICE_SSD; |
6679a90c SL |
2439 | td->filtered_latency = LATENCY_FILTERED_SSD; |
2440 | } else { | |
d61fcfa4 | 2441 | td->throtl_slice = DFL_THROTL_SLICE_HD; |
6679a90c | 2442 | td->filtered_latency = LATENCY_FILTERED_HD; |
b889bf66 JQ |
2443 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) { |
2444 | td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY; | |
2445 | td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY; | |
2446 | } | |
6679a90c | 2447 | } |
d61fcfa4 SL |
2448 | #ifndef CONFIG_BLK_DEV_THROTTLING_LOW |
2449 | /* if no low limit, use previous default */ | |
2450 | td->throtl_slice = DFL_THROTL_SLICE_HD; | |
2451 | #endif | |
9e234eea | 2452 | |
344e9ffc | 2453 | td->track_bio_latency = !queue_is_mq(q); |
b9147dd1 SL |
2454 | if (!td->track_bio_latency) |
2455 | blk_stat_enable_accounting(q); | |
d61fcfa4 SL |
2456 | } |
2457 | ||
297e3d85 SL |
2458 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
2459 | ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page) | |
2460 | { | |
2461 | if (!q->td) | |
2462 | return -EINVAL; | |
2463 | return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice)); | |
2464 | } | |
2465 | ||
2466 | ssize_t blk_throtl_sample_time_store(struct request_queue *q, | |
2467 | const char *page, size_t count) | |
2468 | { | |
2469 | unsigned long v; | |
2470 | unsigned long t; | |
2471 | ||
2472 | if (!q->td) | |
2473 | return -EINVAL; | |
2474 | if (kstrtoul(page, 10, &v)) | |
2475 | return -EINVAL; | |
2476 | t = msecs_to_jiffies(v); | |
2477 | if (t == 0 || t > MAX_THROTL_SLICE) | |
2478 | return -EINVAL; | |
2479 | q->td->throtl_slice = t; | |
2480 | return count; | |
2481 | } | |
2482 | #endif | |
2483 | ||
e43473b7 VG |
2484 | static int __init throtl_init(void) |
2485 | { | |
450adcbe VG |
2486 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
2487 | if (!kthrotld_workqueue) | |
2488 | panic("Failed to create kthrotld\n"); | |
2489 | ||
3c798398 | 2490 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
2491 | } |
2492 | ||
2493 | module_init(throtl_init); |