Commit | Line | Data |
---|---|---|
e43473b7 VG |
1 | /* |
2 | * Interface for controlling IO bandwidth on a request queue | |
3 | * | |
4 | * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com> | |
5 | */ | |
6 | ||
7 | #include <linux/module.h> | |
8 | #include <linux/slab.h> | |
9 | #include <linux/blkdev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blktrace_api.h> | |
eea8f41c | 12 | #include <linux/blk-cgroup.h> |
bc9fcbf9 | 13 | #include "blk.h" |
e43473b7 VG |
14 | |
15 | /* Max dispatch from a group in 1 round */ | |
16 | static int throtl_grp_quantum = 8; | |
17 | ||
18 | /* Total max dispatch from all groups in one round */ | |
19 | static int throtl_quantum = 32; | |
20 | ||
d61fcfa4 SL |
21 | /* Throttling is performed over a slice and after that slice is renewed */ |
22 | #define DFL_THROTL_SLICE_HD (HZ / 10) | |
23 | #define DFL_THROTL_SLICE_SSD (HZ / 50) | |
297e3d85 | 24 | #define MAX_THROTL_SLICE (HZ) |
9e234eea | 25 | #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */ |
9bb67aeb SL |
26 | #define MIN_THROTL_BPS (320 * 1024) |
27 | #define MIN_THROTL_IOPS (10) | |
b4f428ef SL |
28 | #define DFL_LATENCY_TARGET (-1L) |
29 | #define DFL_IDLE_THRESHOLD (0) | |
6679a90c SL |
30 | #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */ |
31 | #define LATENCY_FILTERED_SSD (0) | |
32 | /* | |
33 | * For HD, very small latency comes from sequential IO. Such IO is helpless to | |
34 | * help determine if its IO is impacted by others, hence we ignore the IO | |
35 | */ | |
36 | #define LATENCY_FILTERED_HD (1000L) /* 1ms */ | |
e43473b7 | 37 | |
b9147dd1 SL |
38 | #define SKIP_LATENCY (((u64)1) << BLK_STAT_RES_SHIFT) |
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 | */ | |
c9e0332e TH |
88 | struct rb_root pending_tree; /* RB tree of active tgs */ |
89 | struct rb_node *first_pending; /* first node in the tree */ | |
90 | unsigned int nr_pending; /* # queued in the tree */ | |
91 | unsigned long first_pending_disptime; /* disptime of the first tg */ | |
69df0ab0 | 92 | struct timer_list pending_timer; /* fires on first_pending_disptime */ |
e43473b7 VG |
93 | }; |
94 | ||
5b2c16aa TH |
95 | enum tg_state_flags { |
96 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
0e9f4164 | 97 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ |
5b2c16aa TH |
98 | }; |
99 | ||
e43473b7 VG |
100 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
101 | ||
9f626e37 | 102 | enum { |
cd5ab1b0 | 103 | LIMIT_LOW, |
9f626e37 SL |
104 | LIMIT_MAX, |
105 | LIMIT_CNT, | |
106 | }; | |
107 | ||
e43473b7 | 108 | struct throtl_grp { |
f95a04af TH |
109 | /* must be the first member */ |
110 | struct blkg_policy_data pd; | |
111 | ||
c9e0332e | 112 | /* active throtl group service_queue member */ |
e43473b7 VG |
113 | struct rb_node rb_node; |
114 | ||
0f3457f6 TH |
115 | /* throtl_data this group belongs to */ |
116 | struct throtl_data *td; | |
117 | ||
49a2f1e3 TH |
118 | /* this group's service queue */ |
119 | struct throtl_service_queue service_queue; | |
120 | ||
c5cc2070 TH |
121 | /* |
122 | * qnode_on_self is used when bios are directly queued to this | |
123 | * throtl_grp so that local bios compete fairly with bios | |
124 | * dispatched from children. qnode_on_parent is used when bios are | |
125 | * dispatched from this throtl_grp into its parent and will compete | |
126 | * with the sibling qnode_on_parents and the parent's | |
127 | * qnode_on_self. | |
128 | */ | |
129 | struct throtl_qnode qnode_on_self[2]; | |
130 | struct throtl_qnode qnode_on_parent[2]; | |
131 | ||
e43473b7 VG |
132 | /* |
133 | * Dispatch time in jiffies. This is the estimated time when group | |
134 | * will unthrottle and is ready to dispatch more bio. It is used as | |
135 | * key to sort active groups in service tree. | |
136 | */ | |
137 | unsigned long disptime; | |
138 | ||
e43473b7 VG |
139 | unsigned int flags; |
140 | ||
693e751e TH |
141 | /* are there any throtl rules between this group and td? */ |
142 | bool has_rules[2]; | |
143 | ||
cd5ab1b0 | 144 | /* internally used bytes per second rate limits */ |
9f626e37 | 145 | uint64_t bps[2][LIMIT_CNT]; |
cd5ab1b0 SL |
146 | /* user configured bps limits */ |
147 | uint64_t bps_conf[2][LIMIT_CNT]; | |
e43473b7 | 148 | |
cd5ab1b0 | 149 | /* internally used IOPS limits */ |
9f626e37 | 150 | unsigned int iops[2][LIMIT_CNT]; |
cd5ab1b0 SL |
151 | /* user configured IOPS limits */ |
152 | unsigned int iops_conf[2][LIMIT_CNT]; | |
8e89d13f | 153 | |
e43473b7 VG |
154 | /* Number of bytes disptached in current slice */ |
155 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
156 | /* Number of bio's dispatched in current slice */ |
157 | unsigned int io_disp[2]; | |
e43473b7 | 158 | |
3f0abd80 SL |
159 | unsigned long last_low_overflow_time[2]; |
160 | ||
161 | uint64_t last_bytes_disp[2]; | |
162 | unsigned int last_io_disp[2]; | |
163 | ||
164 | unsigned long last_check_time; | |
165 | ||
ec80991d | 166 | unsigned long latency_target; /* us */ |
5b81fc3c | 167 | unsigned long latency_target_conf; /* us */ |
e43473b7 VG |
168 | /* When did we start a new slice */ |
169 | unsigned long slice_start[2]; | |
170 | unsigned long slice_end[2]; | |
9e234eea SL |
171 | |
172 | unsigned long last_finish_time; /* ns / 1024 */ | |
173 | unsigned long checked_last_finish_time; /* ns / 1024 */ | |
174 | unsigned long avg_idletime; /* ns / 1024 */ | |
175 | unsigned long idletime_threshold; /* us */ | |
5b81fc3c | 176 | unsigned long idletime_threshold_conf; /* us */ |
53696b8d SL |
177 | |
178 | unsigned int bio_cnt; /* total bios */ | |
179 | unsigned int bad_bio_cnt; /* bios exceeding latency threshold */ | |
180 | unsigned long bio_cnt_reset_time; | |
e43473b7 VG |
181 | }; |
182 | ||
b9147dd1 SL |
183 | /* We measure latency for request size from <= 4k to >= 1M */ |
184 | #define LATENCY_BUCKET_SIZE 9 | |
185 | ||
186 | struct latency_bucket { | |
187 | unsigned long total_latency; /* ns / 1024 */ | |
188 | int samples; | |
189 | }; | |
190 | ||
191 | struct avg_latency_bucket { | |
192 | unsigned long latency; /* ns / 1024 */ | |
193 | bool valid; | |
194 | }; | |
195 | ||
e43473b7 VG |
196 | struct throtl_data |
197 | { | |
e43473b7 | 198 | /* service tree for active throtl groups */ |
c9e0332e | 199 | struct throtl_service_queue service_queue; |
e43473b7 | 200 | |
e43473b7 VG |
201 | struct request_queue *queue; |
202 | ||
203 | /* Total Number of queued bios on READ and WRITE lists */ | |
204 | unsigned int nr_queued[2]; | |
205 | ||
297e3d85 SL |
206 | unsigned int throtl_slice; |
207 | ||
e43473b7 | 208 | /* Work for dispatching throttled bios */ |
69df0ab0 | 209 | struct work_struct dispatch_work; |
9f626e37 SL |
210 | unsigned int limit_index; |
211 | bool limit_valid[LIMIT_CNT]; | |
3f0abd80 SL |
212 | |
213 | unsigned long low_upgrade_time; | |
214 | unsigned long low_downgrade_time; | |
7394e31f SL |
215 | |
216 | unsigned int scale; | |
b9147dd1 SL |
217 | |
218 | struct latency_bucket tmp_buckets[LATENCY_BUCKET_SIZE]; | |
219 | struct avg_latency_bucket avg_buckets[LATENCY_BUCKET_SIZE]; | |
220 | struct latency_bucket __percpu *latency_buckets; | |
221 | unsigned long last_calculate_time; | |
6679a90c | 222 | unsigned long filtered_latency; |
b9147dd1 SL |
223 | |
224 | bool track_bio_latency; | |
e43473b7 VG |
225 | }; |
226 | ||
69df0ab0 TH |
227 | static void throtl_pending_timer_fn(unsigned long arg); |
228 | ||
f95a04af TH |
229 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
230 | { | |
231 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
232 | } | |
233 | ||
3c798398 | 234 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 235 | { |
f95a04af | 236 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
237 | } |
238 | ||
3c798398 | 239 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 240 | { |
f95a04af | 241 | return pd_to_blkg(&tg->pd); |
0381411e TH |
242 | } |
243 | ||
fda6f272 TH |
244 | /** |
245 | * sq_to_tg - return the throl_grp the specified service queue belongs to | |
246 | * @sq: the throtl_service_queue of interest | |
247 | * | |
248 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one | |
249 | * embedded in throtl_data, %NULL is returned. | |
250 | */ | |
251 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) | |
252 | { | |
253 | if (sq && sq->parent_sq) | |
254 | return container_of(sq, struct throtl_grp, service_queue); | |
255 | else | |
256 | return NULL; | |
257 | } | |
258 | ||
259 | /** | |
260 | * sq_to_td - return throtl_data the specified service queue belongs to | |
261 | * @sq: the throtl_service_queue of interest | |
262 | * | |
b43daedc | 263 | * A service_queue can be embedded in either a throtl_grp or throtl_data. |
fda6f272 TH |
264 | * Determine the associated throtl_data accordingly and return it. |
265 | */ | |
266 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) | |
267 | { | |
268 | struct throtl_grp *tg = sq_to_tg(sq); | |
269 | ||
270 | if (tg) | |
271 | return tg->td; | |
272 | else | |
273 | return container_of(sq, struct throtl_data, service_queue); | |
274 | } | |
275 | ||
7394e31f SL |
276 | /* |
277 | * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to | |
278 | * make the IO dispatch more smooth. | |
279 | * Scale up: linearly scale up according to lapsed time since upgrade. For | |
280 | * every throtl_slice, the limit scales up 1/2 .low limit till the | |
281 | * limit hits .max limit | |
282 | * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit | |
283 | */ | |
284 | static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td) | |
285 | { | |
286 | /* arbitrary value to avoid too big scale */ | |
287 | if (td->scale < 4096 && time_after_eq(jiffies, | |
288 | td->low_upgrade_time + td->scale * td->throtl_slice)) | |
289 | td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice; | |
290 | ||
291 | return low + (low >> 1) * td->scale; | |
292 | } | |
293 | ||
9f626e37 SL |
294 | static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw) |
295 | { | |
b22c417c | 296 | struct blkcg_gq *blkg = tg_to_blkg(tg); |
7394e31f | 297 | struct throtl_data *td; |
b22c417c SL |
298 | uint64_t ret; |
299 | ||
300 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
301 | return U64_MAX; | |
7394e31f SL |
302 | |
303 | td = tg->td; | |
304 | ret = tg->bps[rw][td->limit_index]; | |
9bb67aeb SL |
305 | if (ret == 0 && td->limit_index == LIMIT_LOW) { |
306 | /* intermediate node or iops isn't 0 */ | |
307 | if (!list_empty(&blkg->blkcg->css.children) || | |
308 | tg->iops[rw][td->limit_index]) | |
309 | return U64_MAX; | |
310 | else | |
311 | return MIN_THROTL_BPS; | |
312 | } | |
7394e31f SL |
313 | |
314 | if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] && | |
315 | tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) { | |
316 | uint64_t adjusted; | |
317 | ||
318 | adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td); | |
319 | ret = min(tg->bps[rw][LIMIT_MAX], adjusted); | |
320 | } | |
b22c417c | 321 | return ret; |
9f626e37 SL |
322 | } |
323 | ||
324 | static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw) | |
325 | { | |
b22c417c | 326 | struct blkcg_gq *blkg = tg_to_blkg(tg); |
7394e31f | 327 | struct throtl_data *td; |
b22c417c SL |
328 | unsigned int ret; |
329 | ||
330 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
331 | return UINT_MAX; | |
9bb67aeb | 332 | |
7394e31f SL |
333 | td = tg->td; |
334 | ret = tg->iops[rw][td->limit_index]; | |
9bb67aeb SL |
335 | if (ret == 0 && tg->td->limit_index == LIMIT_LOW) { |
336 | /* intermediate node or bps isn't 0 */ | |
337 | if (!list_empty(&blkg->blkcg->css.children) || | |
338 | tg->bps[rw][td->limit_index]) | |
339 | return UINT_MAX; | |
340 | else | |
341 | return MIN_THROTL_IOPS; | |
342 | } | |
7394e31f SL |
343 | |
344 | if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] && | |
345 | tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) { | |
346 | uint64_t adjusted; | |
347 | ||
348 | adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td); | |
349 | if (adjusted > UINT_MAX) | |
350 | adjusted = UINT_MAX; | |
351 | ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted); | |
352 | } | |
b22c417c | 353 | return ret; |
9f626e37 SL |
354 | } |
355 | ||
b9147dd1 SL |
356 | #define request_bucket_index(sectors) \ |
357 | clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1) | |
358 | ||
fda6f272 TH |
359 | /** |
360 | * throtl_log - log debug message via blktrace | |
361 | * @sq: the service_queue being reported | |
362 | * @fmt: printf format string | |
363 | * @args: printf args | |
364 | * | |
365 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a | |
366 | * throtl_grp; otherwise, just "throtl". | |
fda6f272 TH |
367 | */ |
368 | #define throtl_log(sq, fmt, args...) do { \ | |
369 | struct throtl_grp *__tg = sq_to_tg((sq)); \ | |
370 | struct throtl_data *__td = sq_to_td((sq)); \ | |
371 | \ | |
372 | (void)__td; \ | |
59fa0224 SL |
373 | if (likely(!blk_trace_note_message_enabled(__td->queue))) \ |
374 | break; \ | |
fda6f272 | 375 | if ((__tg)) { \ |
35fe6d76 SL |
376 | blk_add_cgroup_trace_msg(__td->queue, \ |
377 | tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\ | |
fda6f272 TH |
378 | } else { \ |
379 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
380 | } \ | |
54e7ed12 | 381 | } while (0) |
e43473b7 | 382 | |
ea0ea2bc SL |
383 | static inline unsigned int throtl_bio_data_size(struct bio *bio) |
384 | { | |
385 | /* assume it's one sector */ | |
386 | if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) | |
387 | return 512; | |
388 | return bio->bi_iter.bi_size; | |
389 | } | |
390 | ||
c5cc2070 TH |
391 | static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg) |
392 | { | |
393 | INIT_LIST_HEAD(&qn->node); | |
394 | bio_list_init(&qn->bios); | |
395 | qn->tg = tg; | |
396 | } | |
397 | ||
398 | /** | |
399 | * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it | |
400 | * @bio: bio being added | |
401 | * @qn: qnode to add bio to | |
402 | * @queued: the service_queue->queued[] list @qn belongs to | |
403 | * | |
404 | * Add @bio to @qn and put @qn on @queued if it's not already on. | |
405 | * @qn->tg's reference count is bumped when @qn is activated. See the | |
406 | * comment on top of throtl_qnode definition for details. | |
407 | */ | |
408 | static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn, | |
409 | struct list_head *queued) | |
410 | { | |
411 | bio_list_add(&qn->bios, bio); | |
412 | if (list_empty(&qn->node)) { | |
413 | list_add_tail(&qn->node, queued); | |
414 | blkg_get(tg_to_blkg(qn->tg)); | |
415 | } | |
416 | } | |
417 | ||
418 | /** | |
419 | * throtl_peek_queued - peek the first bio on a qnode list | |
420 | * @queued: the qnode list to peek | |
421 | */ | |
422 | static struct bio *throtl_peek_queued(struct list_head *queued) | |
423 | { | |
424 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
425 | struct bio *bio; | |
426 | ||
427 | if (list_empty(queued)) | |
428 | return NULL; | |
429 | ||
430 | bio = bio_list_peek(&qn->bios); | |
431 | WARN_ON_ONCE(!bio); | |
432 | return bio; | |
433 | } | |
434 | ||
435 | /** | |
436 | * throtl_pop_queued - pop the first bio form a qnode list | |
437 | * @queued: the qnode list to pop a bio from | |
438 | * @tg_to_put: optional out argument for throtl_grp to put | |
439 | * | |
440 | * Pop the first bio from the qnode list @queued. After popping, the first | |
441 | * qnode is removed from @queued if empty or moved to the end of @queued so | |
442 | * that the popping order is round-robin. | |
443 | * | |
444 | * When the first qnode is removed, its associated throtl_grp should be put | |
445 | * too. If @tg_to_put is NULL, this function automatically puts it; | |
446 | * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is | |
447 | * responsible for putting it. | |
448 | */ | |
449 | static struct bio *throtl_pop_queued(struct list_head *queued, | |
450 | struct throtl_grp **tg_to_put) | |
451 | { | |
452 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
453 | struct bio *bio; | |
454 | ||
455 | if (list_empty(queued)) | |
456 | return NULL; | |
457 | ||
458 | bio = bio_list_pop(&qn->bios); | |
459 | WARN_ON_ONCE(!bio); | |
460 | ||
461 | if (bio_list_empty(&qn->bios)) { | |
462 | list_del_init(&qn->node); | |
463 | if (tg_to_put) | |
464 | *tg_to_put = qn->tg; | |
465 | else | |
466 | blkg_put(tg_to_blkg(qn->tg)); | |
467 | } else { | |
468 | list_move_tail(&qn->node, queued); | |
469 | } | |
470 | ||
471 | return bio; | |
472 | } | |
473 | ||
49a2f1e3 | 474 | /* init a service_queue, assumes the caller zeroed it */ |
b2ce2643 | 475 | static void throtl_service_queue_init(struct throtl_service_queue *sq) |
49a2f1e3 | 476 | { |
c5cc2070 TH |
477 | INIT_LIST_HEAD(&sq->queued[0]); |
478 | INIT_LIST_HEAD(&sq->queued[1]); | |
49a2f1e3 | 479 | sq->pending_tree = RB_ROOT; |
69df0ab0 TH |
480 | setup_timer(&sq->pending_timer, throtl_pending_timer_fn, |
481 | (unsigned long)sq); | |
482 | } | |
483 | ||
001bea73 TH |
484 | static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node) |
485 | { | |
4fb72036 | 486 | struct throtl_grp *tg; |
24bdb8ef | 487 | int rw; |
4fb72036 TH |
488 | |
489 | tg = kzalloc_node(sizeof(*tg), gfp, node); | |
490 | if (!tg) | |
77ea7338 | 491 | return NULL; |
4fb72036 | 492 | |
b2ce2643 TH |
493 | throtl_service_queue_init(&tg->service_queue); |
494 | ||
495 | for (rw = READ; rw <= WRITE; rw++) { | |
496 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); | |
497 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); | |
498 | } | |
499 | ||
500 | RB_CLEAR_NODE(&tg->rb_node); | |
9f626e37 SL |
501 | tg->bps[READ][LIMIT_MAX] = U64_MAX; |
502 | tg->bps[WRITE][LIMIT_MAX] = U64_MAX; | |
503 | tg->iops[READ][LIMIT_MAX] = UINT_MAX; | |
504 | tg->iops[WRITE][LIMIT_MAX] = UINT_MAX; | |
cd5ab1b0 SL |
505 | tg->bps_conf[READ][LIMIT_MAX] = U64_MAX; |
506 | tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX; | |
507 | tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX; | |
508 | tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX; | |
509 | /* LIMIT_LOW will have default value 0 */ | |
b2ce2643 | 510 | |
ec80991d | 511 | tg->latency_target = DFL_LATENCY_TARGET; |
5b81fc3c | 512 | tg->latency_target_conf = DFL_LATENCY_TARGET; |
b4f428ef SL |
513 | tg->idletime_threshold = DFL_IDLE_THRESHOLD; |
514 | tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD; | |
ec80991d | 515 | |
4fb72036 | 516 | return &tg->pd; |
001bea73 TH |
517 | } |
518 | ||
a9520cd6 | 519 | static void throtl_pd_init(struct blkg_policy_data *pd) |
a29a171e | 520 | { |
a9520cd6 TH |
521 | struct throtl_grp *tg = pd_to_tg(pd); |
522 | struct blkcg_gq *blkg = tg_to_blkg(tg); | |
77216b04 | 523 | struct throtl_data *td = blkg->q->td; |
b2ce2643 | 524 | struct throtl_service_queue *sq = &tg->service_queue; |
cd1604fa | 525 | |
9138125b | 526 | /* |
aa6ec29b | 527 | * If on the default hierarchy, we switch to properly hierarchical |
9138125b TH |
528 | * behavior where limits on a given throtl_grp are applied to the |
529 | * whole subtree rather than just the group itself. e.g. If 16M | |
530 | * read_bps limit is set on the root group, the whole system can't | |
531 | * exceed 16M for the device. | |
532 | * | |
aa6ec29b | 533 | * If not on the default hierarchy, the broken flat hierarchy |
9138125b TH |
534 | * behavior is retained where all throtl_grps are treated as if |
535 | * they're all separate root groups right below throtl_data. | |
536 | * Limits of a group don't interact with limits of other groups | |
537 | * regardless of the position of the group in the hierarchy. | |
538 | */ | |
b2ce2643 | 539 | sq->parent_sq = &td->service_queue; |
9e10a130 | 540 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent) |
b2ce2643 | 541 | sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue; |
77216b04 | 542 | tg->td = td; |
8a3d2615 TH |
543 | } |
544 | ||
693e751e TH |
545 | /* |
546 | * Set has_rules[] if @tg or any of its parents have limits configured. | |
547 | * This doesn't require walking up to the top of the hierarchy as the | |
548 | * parent's has_rules[] is guaranteed to be correct. | |
549 | */ | |
550 | static void tg_update_has_rules(struct throtl_grp *tg) | |
551 | { | |
552 | struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); | |
9f626e37 | 553 | struct throtl_data *td = tg->td; |
693e751e TH |
554 | int rw; |
555 | ||
556 | for (rw = READ; rw <= WRITE; rw++) | |
557 | tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) || | |
9f626e37 SL |
558 | (td->limit_valid[td->limit_index] && |
559 | (tg_bps_limit(tg, rw) != U64_MAX || | |
560 | tg_iops_limit(tg, rw) != UINT_MAX)); | |
693e751e TH |
561 | } |
562 | ||
a9520cd6 | 563 | static void throtl_pd_online(struct blkg_policy_data *pd) |
693e751e | 564 | { |
aec24246 | 565 | struct throtl_grp *tg = pd_to_tg(pd); |
693e751e TH |
566 | /* |
567 | * We don't want new groups to escape the limits of its ancestors. | |
568 | * Update has_rules[] after a new group is brought online. | |
569 | */ | |
aec24246 | 570 | tg_update_has_rules(tg); |
693e751e TH |
571 | } |
572 | ||
cd5ab1b0 SL |
573 | static void blk_throtl_update_limit_valid(struct throtl_data *td) |
574 | { | |
575 | struct cgroup_subsys_state *pos_css; | |
576 | struct blkcg_gq *blkg; | |
577 | bool low_valid = false; | |
578 | ||
579 | rcu_read_lock(); | |
580 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
581 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
582 | ||
583 | if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] || | |
584 | tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) | |
585 | low_valid = true; | |
586 | } | |
587 | rcu_read_unlock(); | |
588 | ||
589 | td->limit_valid[LIMIT_LOW] = low_valid; | |
590 | } | |
591 | ||
c79892c5 | 592 | static void throtl_upgrade_state(struct throtl_data *td); |
cd5ab1b0 SL |
593 | static void throtl_pd_offline(struct blkg_policy_data *pd) |
594 | { | |
595 | struct throtl_grp *tg = pd_to_tg(pd); | |
596 | ||
597 | tg->bps[READ][LIMIT_LOW] = 0; | |
598 | tg->bps[WRITE][LIMIT_LOW] = 0; | |
599 | tg->iops[READ][LIMIT_LOW] = 0; | |
600 | tg->iops[WRITE][LIMIT_LOW] = 0; | |
601 | ||
602 | blk_throtl_update_limit_valid(tg->td); | |
603 | ||
c79892c5 SL |
604 | if (!tg->td->limit_valid[tg->td->limit_index]) |
605 | throtl_upgrade_state(tg->td); | |
cd5ab1b0 SL |
606 | } |
607 | ||
001bea73 TH |
608 | static void throtl_pd_free(struct blkg_policy_data *pd) |
609 | { | |
4fb72036 TH |
610 | struct throtl_grp *tg = pd_to_tg(pd); |
611 | ||
b2ce2643 | 612 | del_timer_sync(&tg->service_queue.pending_timer); |
4fb72036 | 613 | kfree(tg); |
001bea73 TH |
614 | } |
615 | ||
0049af73 TH |
616 | static struct throtl_grp * |
617 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
618 | { |
619 | /* Service tree is empty */ | |
0049af73 | 620 | if (!parent_sq->nr_pending) |
e43473b7 VG |
621 | return NULL; |
622 | ||
0049af73 TH |
623 | if (!parent_sq->first_pending) |
624 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
e43473b7 | 625 | |
0049af73 TH |
626 | if (parent_sq->first_pending) |
627 | return rb_entry_tg(parent_sq->first_pending); | |
e43473b7 VG |
628 | |
629 | return NULL; | |
630 | } | |
631 | ||
632 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
633 | { | |
634 | rb_erase(n, root); | |
635 | RB_CLEAR_NODE(n); | |
636 | } | |
637 | ||
0049af73 TH |
638 | static void throtl_rb_erase(struct rb_node *n, |
639 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 640 | { |
0049af73 TH |
641 | if (parent_sq->first_pending == n) |
642 | parent_sq->first_pending = NULL; | |
643 | rb_erase_init(n, &parent_sq->pending_tree); | |
644 | --parent_sq->nr_pending; | |
e43473b7 VG |
645 | } |
646 | ||
0049af73 | 647 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
648 | { |
649 | struct throtl_grp *tg; | |
650 | ||
0049af73 | 651 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
652 | if (!tg) |
653 | return; | |
654 | ||
0049af73 | 655 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
656 | } |
657 | ||
77216b04 | 658 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 659 | { |
77216b04 | 660 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
0049af73 | 661 | struct rb_node **node = &parent_sq->pending_tree.rb_node; |
e43473b7 VG |
662 | struct rb_node *parent = NULL; |
663 | struct throtl_grp *__tg; | |
664 | unsigned long key = tg->disptime; | |
665 | int left = 1; | |
666 | ||
667 | while (*node != NULL) { | |
668 | parent = *node; | |
669 | __tg = rb_entry_tg(parent); | |
670 | ||
671 | if (time_before(key, __tg->disptime)) | |
672 | node = &parent->rb_left; | |
673 | else { | |
674 | node = &parent->rb_right; | |
675 | left = 0; | |
676 | } | |
677 | } | |
678 | ||
679 | if (left) | |
0049af73 | 680 | parent_sq->first_pending = &tg->rb_node; |
e43473b7 VG |
681 | |
682 | rb_link_node(&tg->rb_node, parent, node); | |
0049af73 | 683 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); |
e43473b7 VG |
684 | } |
685 | ||
77216b04 | 686 | static void __throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 687 | { |
77216b04 | 688 | tg_service_queue_add(tg); |
5b2c16aa | 689 | tg->flags |= THROTL_TG_PENDING; |
77216b04 | 690 | tg->service_queue.parent_sq->nr_pending++; |
e43473b7 VG |
691 | } |
692 | ||
77216b04 | 693 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 694 | { |
5b2c16aa | 695 | if (!(tg->flags & THROTL_TG_PENDING)) |
77216b04 | 696 | __throtl_enqueue_tg(tg); |
e43473b7 VG |
697 | } |
698 | ||
77216b04 | 699 | static void __throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 700 | { |
77216b04 | 701 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
5b2c16aa | 702 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
703 | } |
704 | ||
77216b04 | 705 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 706 | { |
5b2c16aa | 707 | if (tg->flags & THROTL_TG_PENDING) |
77216b04 | 708 | __throtl_dequeue_tg(tg); |
e43473b7 VG |
709 | } |
710 | ||
a9131a27 | 711 | /* Call with queue lock held */ |
69df0ab0 TH |
712 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, |
713 | unsigned long expires) | |
a9131a27 | 714 | { |
a41b816c | 715 | unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice; |
06cceedc SL |
716 | |
717 | /* | |
718 | * Since we are adjusting the throttle limit dynamically, the sleep | |
719 | * time calculated according to previous limit might be invalid. It's | |
720 | * possible the cgroup sleep time is very long and no other cgroups | |
721 | * have IO running so notify the limit changes. Make sure the cgroup | |
722 | * doesn't sleep too long to avoid the missed notification. | |
723 | */ | |
724 | if (time_after(expires, max_expire)) | |
725 | expires = max_expire; | |
69df0ab0 TH |
726 | mod_timer(&sq->pending_timer, expires); |
727 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
728 | expires - jiffies, jiffies); | |
a9131a27 TH |
729 | } |
730 | ||
7f52f98c TH |
731 | /** |
732 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle | |
733 | * @sq: the service_queue to schedule dispatch for | |
734 | * @force: force scheduling | |
735 | * | |
736 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the | |
737 | * dispatch time of the first pending child. Returns %true if either timer | |
738 | * is armed or there's no pending child left. %false if the current | |
739 | * dispatch window is still open and the caller should continue | |
740 | * dispatching. | |
741 | * | |
742 | * If @force is %true, the dispatch timer is always scheduled and this | |
743 | * function is guaranteed to return %true. This is to be used when the | |
744 | * caller can't dispatch itself and needs to invoke pending_timer | |
745 | * unconditionally. Note that forced scheduling is likely to induce short | |
746 | * delay before dispatch starts even if @sq->first_pending_disptime is not | |
747 | * in the future and thus shouldn't be used in hot paths. | |
748 | */ | |
749 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
750 | bool force) | |
e43473b7 | 751 | { |
6a525600 | 752 | /* any pending children left? */ |
c9e0332e | 753 | if (!sq->nr_pending) |
7f52f98c | 754 | return true; |
e43473b7 | 755 | |
c9e0332e | 756 | update_min_dispatch_time(sq); |
e43473b7 | 757 | |
69df0ab0 | 758 | /* is the next dispatch time in the future? */ |
7f52f98c | 759 | if (force || time_after(sq->first_pending_disptime, jiffies)) { |
69df0ab0 | 760 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); |
7f52f98c | 761 | return true; |
69df0ab0 TH |
762 | } |
763 | ||
7f52f98c TH |
764 | /* tell the caller to continue dispatching */ |
765 | return false; | |
e43473b7 VG |
766 | } |
767 | ||
32ee5bc4 VG |
768 | static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, |
769 | bool rw, unsigned long start) | |
770 | { | |
771 | tg->bytes_disp[rw] = 0; | |
772 | tg->io_disp[rw] = 0; | |
773 | ||
774 | /* | |
775 | * Previous slice has expired. We must have trimmed it after last | |
776 | * bio dispatch. That means since start of last slice, we never used | |
777 | * that bandwidth. Do try to make use of that bandwidth while giving | |
778 | * credit. | |
779 | */ | |
780 | if (time_after_eq(start, tg->slice_start[rw])) | |
781 | tg->slice_start[rw] = start; | |
782 | ||
297e3d85 | 783 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; |
32ee5bc4 VG |
784 | throtl_log(&tg->service_queue, |
785 | "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", | |
786 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
787 | tg->slice_end[rw], jiffies); | |
788 | } | |
789 | ||
0f3457f6 | 790 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
791 | { |
792 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 793 | tg->io_disp[rw] = 0; |
e43473b7 | 794 | tg->slice_start[rw] = jiffies; |
297e3d85 | 795 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; |
fda6f272 TH |
796 | throtl_log(&tg->service_queue, |
797 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
798 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
799 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
800 | } |
801 | ||
0f3457f6 TH |
802 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
803 | unsigned long jiffy_end) | |
d1ae8ffd | 804 | { |
297e3d85 | 805 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); |
d1ae8ffd VG |
806 | } |
807 | ||
0f3457f6 TH |
808 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
809 | unsigned long jiffy_end) | |
e43473b7 | 810 | { |
297e3d85 | 811 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); |
fda6f272 TH |
812 | throtl_log(&tg->service_queue, |
813 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
814 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
815 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
816 | } |
817 | ||
818 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 819 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
820 | { |
821 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
5cf8c227 | 822 | return false; |
e43473b7 VG |
823 | |
824 | return 1; | |
825 | } | |
826 | ||
827 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 828 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 829 | { |
3aad5d3e VG |
830 | unsigned long nr_slices, time_elapsed, io_trim; |
831 | u64 bytes_trim, tmp; | |
e43473b7 VG |
832 | |
833 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
834 | ||
835 | /* | |
836 | * If bps are unlimited (-1), then time slice don't get | |
837 | * renewed. Don't try to trim the slice if slice is used. A new | |
838 | * slice will start when appropriate. | |
839 | */ | |
0f3457f6 | 840 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
841 | return; |
842 | ||
d1ae8ffd VG |
843 | /* |
844 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
845 | * that initially cgroup limit was very low resulting in high | |
846 | * slice_end, but later limit was bumped up and bio was dispached | |
847 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
848 | * is bad because it does not allow new slice to start. | |
849 | */ | |
850 | ||
297e3d85 | 851 | throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice); |
d1ae8ffd | 852 | |
e43473b7 VG |
853 | time_elapsed = jiffies - tg->slice_start[rw]; |
854 | ||
297e3d85 | 855 | nr_slices = time_elapsed / tg->td->throtl_slice; |
e43473b7 VG |
856 | |
857 | if (!nr_slices) | |
858 | return; | |
297e3d85 | 859 | tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices; |
3aad5d3e VG |
860 | do_div(tmp, HZ); |
861 | bytes_trim = tmp; | |
e43473b7 | 862 | |
297e3d85 SL |
863 | io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) / |
864 | HZ; | |
e43473b7 | 865 | |
8e89d13f | 866 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
867 | return; |
868 | ||
869 | if (tg->bytes_disp[rw] >= bytes_trim) | |
870 | tg->bytes_disp[rw] -= bytes_trim; | |
871 | else | |
872 | tg->bytes_disp[rw] = 0; | |
873 | ||
8e89d13f VG |
874 | if (tg->io_disp[rw] >= io_trim) |
875 | tg->io_disp[rw] -= io_trim; | |
876 | else | |
877 | tg->io_disp[rw] = 0; | |
878 | ||
297e3d85 | 879 | tg->slice_start[rw] += nr_slices * tg->td->throtl_slice; |
e43473b7 | 880 | |
fda6f272 TH |
881 | throtl_log(&tg->service_queue, |
882 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
883 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
884 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
e43473b7 VG |
885 | } |
886 | ||
0f3457f6 TH |
887 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
888 | unsigned long *wait) | |
e43473b7 VG |
889 | { |
890 | bool rw = bio_data_dir(bio); | |
8e89d13f | 891 | unsigned int io_allowed; |
e43473b7 | 892 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 893 | u64 tmp; |
e43473b7 | 894 | |
8e89d13f | 895 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 896 | |
8e89d13f VG |
897 | /* Slice has just started. Consider one slice interval */ |
898 | if (!jiffy_elapsed) | |
297e3d85 | 899 | jiffy_elapsed_rnd = tg->td->throtl_slice; |
8e89d13f | 900 | |
297e3d85 | 901 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); |
8e89d13f | 902 | |
c49c06e4 VG |
903 | /* |
904 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
905 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
906 | * will allow dispatch after 1 second and after that slice should | |
907 | * have been trimmed. | |
908 | */ | |
909 | ||
9f626e37 | 910 | tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd; |
c49c06e4 VG |
911 | do_div(tmp, HZ); |
912 | ||
913 | if (tmp > UINT_MAX) | |
914 | io_allowed = UINT_MAX; | |
915 | else | |
916 | io_allowed = tmp; | |
8e89d13f VG |
917 | |
918 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
919 | if (wait) |
920 | *wait = 0; | |
5cf8c227 | 921 | return true; |
e43473b7 VG |
922 | } |
923 | ||
8e89d13f | 924 | /* Calc approx time to dispatch */ |
9f626e37 | 925 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ) / tg_iops_limit(tg, rw) + 1; |
8e89d13f VG |
926 | |
927 | if (jiffy_wait > jiffy_elapsed) | |
928 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
929 | else | |
930 | jiffy_wait = 1; | |
931 | ||
932 | if (wait) | |
933 | *wait = jiffy_wait; | |
934 | return 0; | |
935 | } | |
936 | ||
0f3457f6 TH |
937 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
938 | unsigned long *wait) | |
8e89d13f VG |
939 | { |
940 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 941 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 942 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
ea0ea2bc | 943 | unsigned int bio_size = throtl_bio_data_size(bio); |
e43473b7 VG |
944 | |
945 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
946 | ||
947 | /* Slice has just started. Consider one slice interval */ | |
948 | if (!jiffy_elapsed) | |
297e3d85 | 949 | jiffy_elapsed_rnd = tg->td->throtl_slice; |
e43473b7 | 950 | |
297e3d85 | 951 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); |
e43473b7 | 952 | |
9f626e37 | 953 | tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd; |
5e901a2b | 954 | do_div(tmp, HZ); |
3aad5d3e | 955 | bytes_allowed = tmp; |
e43473b7 | 956 | |
ea0ea2bc | 957 | if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) { |
e43473b7 VG |
958 | if (wait) |
959 | *wait = 0; | |
5cf8c227 | 960 | return true; |
e43473b7 VG |
961 | } |
962 | ||
963 | /* Calc approx time to dispatch */ | |
ea0ea2bc | 964 | extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed; |
9f626e37 | 965 | jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw)); |
e43473b7 VG |
966 | |
967 | if (!jiffy_wait) | |
968 | jiffy_wait = 1; | |
969 | ||
970 | /* | |
971 | * This wait time is without taking into consideration the rounding | |
972 | * up we did. Add that time also. | |
973 | */ | |
974 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
975 | if (wait) |
976 | *wait = jiffy_wait; | |
8e89d13f VG |
977 | return 0; |
978 | } | |
979 | ||
980 | /* | |
981 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
982 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
983 | */ | |
0f3457f6 TH |
984 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
985 | unsigned long *wait) | |
8e89d13f VG |
986 | { |
987 | bool rw = bio_data_dir(bio); | |
988 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
989 | ||
990 | /* | |
991 | * Currently whole state machine of group depends on first bio | |
992 | * queued in the group bio list. So one should not be calling | |
993 | * this function with a different bio if there are other bios | |
994 | * queued. | |
995 | */ | |
73f0d49a | 996 | BUG_ON(tg->service_queue.nr_queued[rw] && |
c5cc2070 | 997 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); |
e43473b7 | 998 | |
8e89d13f | 999 | /* If tg->bps = -1, then BW is unlimited */ |
9f626e37 SL |
1000 | if (tg_bps_limit(tg, rw) == U64_MAX && |
1001 | tg_iops_limit(tg, rw) == UINT_MAX) { | |
8e89d13f VG |
1002 | if (wait) |
1003 | *wait = 0; | |
5cf8c227 | 1004 | return true; |
8e89d13f VG |
1005 | } |
1006 | ||
1007 | /* | |
1008 | * If previous slice expired, start a new one otherwise renew/extend | |
1009 | * existing slice to make sure it is at least throtl_slice interval | |
164c80ed VG |
1010 | * long since now. New slice is started only for empty throttle group. |
1011 | * If there is queued bio, that means there should be an active | |
1012 | * slice and it should be extended instead. | |
8e89d13f | 1013 | */ |
164c80ed | 1014 | if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw])) |
0f3457f6 | 1015 | throtl_start_new_slice(tg, rw); |
8e89d13f | 1016 | else { |
297e3d85 SL |
1017 | if (time_before(tg->slice_end[rw], |
1018 | jiffies + tg->td->throtl_slice)) | |
1019 | throtl_extend_slice(tg, rw, | |
1020 | jiffies + tg->td->throtl_slice); | |
8e89d13f VG |
1021 | } |
1022 | ||
0f3457f6 TH |
1023 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
1024 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
1025 | if (wait) |
1026 | *wait = 0; | |
1027 | return 1; | |
1028 | } | |
1029 | ||
1030 | max_wait = max(bps_wait, iops_wait); | |
1031 | ||
1032 | if (wait) | |
1033 | *wait = max_wait; | |
1034 | ||
1035 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 1036 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
1037 | |
1038 | return 0; | |
1039 | } | |
1040 | ||
1041 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
1042 | { | |
1043 | bool rw = bio_data_dir(bio); | |
ea0ea2bc | 1044 | unsigned int bio_size = throtl_bio_data_size(bio); |
e43473b7 VG |
1045 | |
1046 | /* Charge the bio to the group */ | |
ea0ea2bc | 1047 | tg->bytes_disp[rw] += bio_size; |
8e89d13f | 1048 | tg->io_disp[rw]++; |
ea0ea2bc | 1049 | tg->last_bytes_disp[rw] += bio_size; |
3f0abd80 | 1050 | tg->last_io_disp[rw]++; |
e43473b7 | 1051 | |
2a0f61e6 | 1052 | /* |
8d2bbd4c | 1053 | * BIO_THROTTLED is used to prevent the same bio to be throttled |
2a0f61e6 TH |
1054 | * more than once as a throttled bio will go through blk-throtl the |
1055 | * second time when it eventually gets issued. Set it when a bio | |
1056 | * is being charged to a tg. | |
2a0f61e6 | 1057 | */ |
8d2bbd4c CH |
1058 | if (!bio_flagged(bio, BIO_THROTTLED)) |
1059 | bio_set_flag(bio, BIO_THROTTLED); | |
e43473b7 VG |
1060 | } |
1061 | ||
c5cc2070 TH |
1062 | /** |
1063 | * throtl_add_bio_tg - add a bio to the specified throtl_grp | |
1064 | * @bio: bio to add | |
1065 | * @qn: qnode to use | |
1066 | * @tg: the target throtl_grp | |
1067 | * | |
1068 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, | |
1069 | * tg->qnode_on_self[] is used. | |
1070 | */ | |
1071 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, | |
1072 | struct throtl_grp *tg) | |
e43473b7 | 1073 | { |
73f0d49a | 1074 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1075 | bool rw = bio_data_dir(bio); |
1076 | ||
c5cc2070 TH |
1077 | if (!qn) |
1078 | qn = &tg->qnode_on_self[rw]; | |
1079 | ||
0e9f4164 TH |
1080 | /* |
1081 | * If @tg doesn't currently have any bios queued in the same | |
1082 | * direction, queueing @bio can change when @tg should be | |
1083 | * dispatched. Mark that @tg was empty. This is automatically | |
1084 | * cleaered on the next tg_update_disptime(). | |
1085 | */ | |
1086 | if (!sq->nr_queued[rw]) | |
1087 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
1088 | ||
c5cc2070 TH |
1089 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); |
1090 | ||
73f0d49a | 1091 | sq->nr_queued[rw]++; |
77216b04 | 1092 | throtl_enqueue_tg(tg); |
e43473b7 VG |
1093 | } |
1094 | ||
77216b04 | 1095 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 1096 | { |
73f0d49a | 1097 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1098 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
1099 | struct bio *bio; | |
1100 | ||
d609af3a ME |
1101 | bio = throtl_peek_queued(&sq->queued[READ]); |
1102 | if (bio) | |
0f3457f6 | 1103 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 1104 | |
d609af3a ME |
1105 | bio = throtl_peek_queued(&sq->queued[WRITE]); |
1106 | if (bio) | |
0f3457f6 | 1107 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
1108 | |
1109 | min_wait = min(read_wait, write_wait); | |
1110 | disptime = jiffies + min_wait; | |
1111 | ||
e43473b7 | 1112 | /* Update dispatch time */ |
77216b04 | 1113 | throtl_dequeue_tg(tg); |
e43473b7 | 1114 | tg->disptime = disptime; |
77216b04 | 1115 | throtl_enqueue_tg(tg); |
0e9f4164 TH |
1116 | |
1117 | /* see throtl_add_bio_tg() */ | |
1118 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
1119 | } |
1120 | ||
32ee5bc4 VG |
1121 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, |
1122 | struct throtl_grp *parent_tg, bool rw) | |
1123 | { | |
1124 | if (throtl_slice_used(parent_tg, rw)) { | |
1125 | throtl_start_new_slice_with_credit(parent_tg, rw, | |
1126 | child_tg->slice_start[rw]); | |
1127 | } | |
1128 | ||
1129 | } | |
1130 | ||
77216b04 | 1131 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 1132 | { |
73f0d49a | 1133 | struct throtl_service_queue *sq = &tg->service_queue; |
6bc9c2b4 TH |
1134 | struct throtl_service_queue *parent_sq = sq->parent_sq; |
1135 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
c5cc2070 | 1136 | struct throtl_grp *tg_to_put = NULL; |
e43473b7 VG |
1137 | struct bio *bio; |
1138 | ||
c5cc2070 TH |
1139 | /* |
1140 | * @bio is being transferred from @tg to @parent_sq. Popping a bio | |
1141 | * from @tg may put its reference and @parent_sq might end up | |
1142 | * getting released prematurely. Remember the tg to put and put it | |
1143 | * after @bio is transferred to @parent_sq. | |
1144 | */ | |
1145 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); | |
73f0d49a | 1146 | sq->nr_queued[rw]--; |
e43473b7 VG |
1147 | |
1148 | throtl_charge_bio(tg, bio); | |
6bc9c2b4 TH |
1149 | |
1150 | /* | |
1151 | * If our parent is another tg, we just need to transfer @bio to | |
1152 | * the parent using throtl_add_bio_tg(). If our parent is | |
1153 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
1154 | * bio_lists[] and decrease total number queued. The caller is | |
1155 | * responsible for issuing these bios. | |
1156 | */ | |
1157 | if (parent_tg) { | |
c5cc2070 | 1158 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); |
32ee5bc4 | 1159 | start_parent_slice_with_credit(tg, parent_tg, rw); |
6bc9c2b4 | 1160 | } else { |
c5cc2070 TH |
1161 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], |
1162 | &parent_sq->queued[rw]); | |
6bc9c2b4 TH |
1163 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
1164 | tg->td->nr_queued[rw]--; | |
1165 | } | |
e43473b7 | 1166 | |
0f3457f6 | 1167 | throtl_trim_slice(tg, rw); |
6bc9c2b4 | 1168 | |
c5cc2070 TH |
1169 | if (tg_to_put) |
1170 | blkg_put(tg_to_blkg(tg_to_put)); | |
e43473b7 VG |
1171 | } |
1172 | ||
77216b04 | 1173 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 1174 | { |
73f0d49a | 1175 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1176 | unsigned int nr_reads = 0, nr_writes = 0; |
1177 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 1178 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
1179 | struct bio *bio; |
1180 | ||
1181 | /* Try to dispatch 75% READS and 25% WRITES */ | |
1182 | ||
c5cc2070 | 1183 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && |
0f3457f6 | 1184 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1185 | |
77216b04 | 1186 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1187 | nr_reads++; |
1188 | ||
1189 | if (nr_reads >= max_nr_reads) | |
1190 | break; | |
1191 | } | |
1192 | ||
c5cc2070 | 1193 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && |
0f3457f6 | 1194 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1195 | |
77216b04 | 1196 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1197 | nr_writes++; |
1198 | ||
1199 | if (nr_writes >= max_nr_writes) | |
1200 | break; | |
1201 | } | |
1202 | ||
1203 | return nr_reads + nr_writes; | |
1204 | } | |
1205 | ||
651930bc | 1206 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
1207 | { |
1208 | unsigned int nr_disp = 0; | |
e43473b7 VG |
1209 | |
1210 | while (1) { | |
73f0d49a TH |
1211 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
1212 | struct throtl_service_queue *sq = &tg->service_queue; | |
e43473b7 VG |
1213 | |
1214 | if (!tg) | |
1215 | break; | |
1216 | ||
1217 | if (time_before(jiffies, tg->disptime)) | |
1218 | break; | |
1219 | ||
77216b04 | 1220 | throtl_dequeue_tg(tg); |
e43473b7 | 1221 | |
77216b04 | 1222 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 1223 | |
73f0d49a | 1224 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
77216b04 | 1225 | tg_update_disptime(tg); |
e43473b7 VG |
1226 | |
1227 | if (nr_disp >= throtl_quantum) | |
1228 | break; | |
1229 | } | |
1230 | ||
1231 | return nr_disp; | |
1232 | } | |
1233 | ||
c79892c5 SL |
1234 | static bool throtl_can_upgrade(struct throtl_data *td, |
1235 | struct throtl_grp *this_tg); | |
6e1a5704 TH |
1236 | /** |
1237 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
1238 | * @arg: the throtl_service_queue being serviced | |
1239 | * | |
1240 | * This timer is armed when a child throtl_grp with active bio's become | |
1241 | * pending and queued on the service_queue's pending_tree and expires when | |
1242 | * the first child throtl_grp should be dispatched. This function | |
2e48a530 TH |
1243 | * dispatches bio's from the children throtl_grps to the parent |
1244 | * service_queue. | |
1245 | * | |
1246 | * If the parent's parent is another throtl_grp, dispatching is propagated | |
1247 | * by either arming its pending_timer or repeating dispatch directly. If | |
1248 | * the top-level service_tree is reached, throtl_data->dispatch_work is | |
1249 | * kicked so that the ready bio's are issued. | |
6e1a5704 | 1250 | */ |
69df0ab0 TH |
1251 | static void throtl_pending_timer_fn(unsigned long arg) |
1252 | { | |
1253 | struct throtl_service_queue *sq = (void *)arg; | |
2e48a530 | 1254 | struct throtl_grp *tg = sq_to_tg(sq); |
69df0ab0 | 1255 | struct throtl_data *td = sq_to_td(sq); |
cb76199c | 1256 | struct request_queue *q = td->queue; |
2e48a530 TH |
1257 | struct throtl_service_queue *parent_sq; |
1258 | bool dispatched; | |
6e1a5704 | 1259 | int ret; |
e43473b7 VG |
1260 | |
1261 | spin_lock_irq(q->queue_lock); | |
c79892c5 SL |
1262 | if (throtl_can_upgrade(td, NULL)) |
1263 | throtl_upgrade_state(td); | |
1264 | ||
2e48a530 TH |
1265 | again: |
1266 | parent_sq = sq->parent_sq; | |
1267 | dispatched = false; | |
e43473b7 | 1268 | |
7f52f98c TH |
1269 | while (true) { |
1270 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
2e48a530 TH |
1271 | sq->nr_queued[READ] + sq->nr_queued[WRITE], |
1272 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
7f52f98c TH |
1273 | |
1274 | ret = throtl_select_dispatch(sq); | |
1275 | if (ret) { | |
7f52f98c TH |
1276 | throtl_log(sq, "bios disp=%u", ret); |
1277 | dispatched = true; | |
1278 | } | |
e43473b7 | 1279 | |
7f52f98c TH |
1280 | if (throtl_schedule_next_dispatch(sq, false)) |
1281 | break; | |
e43473b7 | 1282 | |
7f52f98c TH |
1283 | /* this dispatch windows is still open, relax and repeat */ |
1284 | spin_unlock_irq(q->queue_lock); | |
1285 | cpu_relax(); | |
1286 | spin_lock_irq(q->queue_lock); | |
651930bc | 1287 | } |
e43473b7 | 1288 | |
2e48a530 TH |
1289 | if (!dispatched) |
1290 | goto out_unlock; | |
6e1a5704 | 1291 | |
2e48a530 TH |
1292 | if (parent_sq) { |
1293 | /* @parent_sq is another throl_grp, propagate dispatch */ | |
1294 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1295 | tg_update_disptime(tg); | |
1296 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { | |
1297 | /* window is already open, repeat dispatching */ | |
1298 | sq = parent_sq; | |
1299 | tg = sq_to_tg(sq); | |
1300 | goto again; | |
1301 | } | |
1302 | } | |
1303 | } else { | |
1304 | /* reached the top-level, queue issueing */ | |
1305 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1306 | } | |
1307 | out_unlock: | |
e43473b7 | 1308 | spin_unlock_irq(q->queue_lock); |
6e1a5704 | 1309 | } |
e43473b7 | 1310 | |
6e1a5704 TH |
1311 | /** |
1312 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
1313 | * @work: work item being executed | |
1314 | * | |
1315 | * This function is queued for execution when bio's reach the bio_lists[] | |
1316 | * of throtl_data->service_queue. Those bio's are ready and issued by this | |
1317 | * function. | |
1318 | */ | |
8876e140 | 1319 | static void blk_throtl_dispatch_work_fn(struct work_struct *work) |
6e1a5704 TH |
1320 | { |
1321 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1322 | dispatch_work); | |
1323 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1324 | struct request_queue *q = td->queue; | |
1325 | struct bio_list bio_list_on_stack; | |
1326 | struct bio *bio; | |
1327 | struct blk_plug plug; | |
1328 | int rw; | |
1329 | ||
1330 | bio_list_init(&bio_list_on_stack); | |
1331 | ||
1332 | spin_lock_irq(q->queue_lock); | |
c5cc2070 TH |
1333 | for (rw = READ; rw <= WRITE; rw++) |
1334 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) | |
1335 | bio_list_add(&bio_list_on_stack, bio); | |
6e1a5704 TH |
1336 | spin_unlock_irq(q->queue_lock); |
1337 | ||
1338 | if (!bio_list_empty(&bio_list_on_stack)) { | |
69d60eb9 | 1339 | blk_start_plug(&plug); |
e43473b7 VG |
1340 | while((bio = bio_list_pop(&bio_list_on_stack))) |
1341 | generic_make_request(bio); | |
69d60eb9 | 1342 | blk_finish_plug(&plug); |
e43473b7 | 1343 | } |
e43473b7 VG |
1344 | } |
1345 | ||
f95a04af TH |
1346 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
1347 | int off) | |
60c2bc2d | 1348 | { |
f95a04af TH |
1349 | struct throtl_grp *tg = pd_to_tg(pd); |
1350 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 1351 | |
2ab5492d | 1352 | if (v == U64_MAX) |
60c2bc2d | 1353 | return 0; |
f95a04af | 1354 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
1355 | } |
1356 | ||
f95a04af TH |
1357 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
1358 | int off) | |
e43473b7 | 1359 | { |
f95a04af TH |
1360 | struct throtl_grp *tg = pd_to_tg(pd); |
1361 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 1362 | |
2ab5492d | 1363 | if (v == UINT_MAX) |
af133ceb | 1364 | return 0; |
f95a04af | 1365 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
1366 | } |
1367 | ||
2da8ca82 | 1368 | static int tg_print_conf_u64(struct seq_file *sf, void *v) |
8e89d13f | 1369 | { |
2da8ca82 TH |
1370 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64, |
1371 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1372 | return 0; |
8e89d13f VG |
1373 | } |
1374 | ||
2da8ca82 | 1375 | static int tg_print_conf_uint(struct seq_file *sf, void *v) |
8e89d13f | 1376 | { |
2da8ca82 TH |
1377 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint, |
1378 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1379 | return 0; |
60c2bc2d TH |
1380 | } |
1381 | ||
9bb67aeb | 1382 | static void tg_conf_updated(struct throtl_grp *tg, bool global) |
60c2bc2d | 1383 | { |
69948b07 | 1384 | struct throtl_service_queue *sq = &tg->service_queue; |
492eb21b | 1385 | struct cgroup_subsys_state *pos_css; |
69948b07 | 1386 | struct blkcg_gq *blkg; |
af133ceb | 1387 | |
fda6f272 TH |
1388 | throtl_log(&tg->service_queue, |
1389 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
9f626e37 SL |
1390 | tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE), |
1391 | tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE)); | |
632b4493 | 1392 | |
693e751e TH |
1393 | /* |
1394 | * Update has_rules[] flags for the updated tg's subtree. A tg is | |
1395 | * considered to have rules if either the tg itself or any of its | |
1396 | * ancestors has rules. This identifies groups without any | |
1397 | * restrictions in the whole hierarchy and allows them to bypass | |
1398 | * blk-throttle. | |
1399 | */ | |
9bb67aeb SL |
1400 | blkg_for_each_descendant_pre(blkg, pos_css, |
1401 | global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) { | |
5b81fc3c SL |
1402 | struct throtl_grp *this_tg = blkg_to_tg(blkg); |
1403 | struct throtl_grp *parent_tg; | |
1404 | ||
1405 | tg_update_has_rules(this_tg); | |
1406 | /* ignore root/second level */ | |
1407 | if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent || | |
1408 | !blkg->parent->parent) | |
1409 | continue; | |
1410 | parent_tg = blkg_to_tg(blkg->parent); | |
1411 | /* | |
1412 | * make sure all children has lower idle time threshold and | |
1413 | * higher latency target | |
1414 | */ | |
1415 | this_tg->idletime_threshold = min(this_tg->idletime_threshold, | |
1416 | parent_tg->idletime_threshold); | |
1417 | this_tg->latency_target = max(this_tg->latency_target, | |
1418 | parent_tg->latency_target); | |
1419 | } | |
693e751e | 1420 | |
632b4493 TH |
1421 | /* |
1422 | * We're already holding queue_lock and know @tg is valid. Let's | |
1423 | * apply the new config directly. | |
1424 | * | |
1425 | * Restart the slices for both READ and WRITES. It might happen | |
1426 | * that a group's limit are dropped suddenly and we don't want to | |
1427 | * account recently dispatched IO with new low rate. | |
1428 | */ | |
0f3457f6 TH |
1429 | throtl_start_new_slice(tg, 0); |
1430 | throtl_start_new_slice(tg, 1); | |
632b4493 | 1431 | |
5b2c16aa | 1432 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 1433 | tg_update_disptime(tg); |
7f52f98c | 1434 | throtl_schedule_next_dispatch(sq->parent_sq, true); |
632b4493 | 1435 | } |
69948b07 TH |
1436 | } |
1437 | ||
1438 | static ssize_t tg_set_conf(struct kernfs_open_file *of, | |
1439 | char *buf, size_t nbytes, loff_t off, bool is_u64) | |
1440 | { | |
1441 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1442 | struct blkg_conf_ctx ctx; | |
1443 | struct throtl_grp *tg; | |
1444 | int ret; | |
1445 | u64 v; | |
1446 | ||
1447 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1448 | if (ret) | |
1449 | return ret; | |
1450 | ||
1451 | ret = -EINVAL; | |
1452 | if (sscanf(ctx.body, "%llu", &v) != 1) | |
1453 | goto out_finish; | |
1454 | if (!v) | |
2ab5492d | 1455 | v = U64_MAX; |
69948b07 TH |
1456 | |
1457 | tg = blkg_to_tg(ctx.blkg); | |
1458 | ||
1459 | if (is_u64) | |
1460 | *(u64 *)((void *)tg + of_cft(of)->private) = v; | |
1461 | else | |
1462 | *(unsigned int *)((void *)tg + of_cft(of)->private) = v; | |
60c2bc2d | 1463 | |
9bb67aeb | 1464 | tg_conf_updated(tg, false); |
36aa9e5f TH |
1465 | ret = 0; |
1466 | out_finish: | |
60c2bc2d | 1467 | blkg_conf_finish(&ctx); |
36aa9e5f | 1468 | return ret ?: nbytes; |
8e89d13f VG |
1469 | } |
1470 | ||
451af504 TH |
1471 | static ssize_t tg_set_conf_u64(struct kernfs_open_file *of, |
1472 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1473 | { |
451af504 | 1474 | return tg_set_conf(of, buf, nbytes, off, true); |
60c2bc2d TH |
1475 | } |
1476 | ||
451af504 TH |
1477 | static ssize_t tg_set_conf_uint(struct kernfs_open_file *of, |
1478 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1479 | { |
451af504 | 1480 | return tg_set_conf(of, buf, nbytes, off, false); |
60c2bc2d TH |
1481 | } |
1482 | ||
880f50e2 | 1483 | static struct cftype throtl_legacy_files[] = { |
60c2bc2d TH |
1484 | { |
1485 | .name = "throttle.read_bps_device", | |
9f626e37 | 1486 | .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]), |
2da8ca82 | 1487 | .seq_show = tg_print_conf_u64, |
451af504 | 1488 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1489 | }, |
1490 | { | |
1491 | .name = "throttle.write_bps_device", | |
9f626e37 | 1492 | .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]), |
2da8ca82 | 1493 | .seq_show = tg_print_conf_u64, |
451af504 | 1494 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1495 | }, |
1496 | { | |
1497 | .name = "throttle.read_iops_device", | |
9f626e37 | 1498 | .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]), |
2da8ca82 | 1499 | .seq_show = tg_print_conf_uint, |
451af504 | 1500 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1501 | }, |
1502 | { | |
1503 | .name = "throttle.write_iops_device", | |
9f626e37 | 1504 | .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]), |
2da8ca82 | 1505 | .seq_show = tg_print_conf_uint, |
451af504 | 1506 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1507 | }, |
1508 | { | |
1509 | .name = "throttle.io_service_bytes", | |
77ea7338 TH |
1510 | .private = (unsigned long)&blkcg_policy_throtl, |
1511 | .seq_show = blkg_print_stat_bytes, | |
60c2bc2d TH |
1512 | }, |
1513 | { | |
1514 | .name = "throttle.io_serviced", | |
77ea7338 TH |
1515 | .private = (unsigned long)&blkcg_policy_throtl, |
1516 | .seq_show = blkg_print_stat_ios, | |
60c2bc2d TH |
1517 | }, |
1518 | { } /* terminate */ | |
1519 | }; | |
1520 | ||
cd5ab1b0 | 1521 | static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd, |
2ee867dc TH |
1522 | int off) |
1523 | { | |
1524 | struct throtl_grp *tg = pd_to_tg(pd); | |
1525 | const char *dname = blkg_dev_name(pd->blkg); | |
1526 | char bufs[4][21] = { "max", "max", "max", "max" }; | |
cd5ab1b0 SL |
1527 | u64 bps_dft; |
1528 | unsigned int iops_dft; | |
ada75b6e | 1529 | char idle_time[26] = ""; |
ec80991d | 1530 | char latency_time[26] = ""; |
2ee867dc TH |
1531 | |
1532 | if (!dname) | |
1533 | return 0; | |
9f626e37 | 1534 | |
cd5ab1b0 SL |
1535 | if (off == LIMIT_LOW) { |
1536 | bps_dft = 0; | |
1537 | iops_dft = 0; | |
1538 | } else { | |
1539 | bps_dft = U64_MAX; | |
1540 | iops_dft = UINT_MAX; | |
1541 | } | |
1542 | ||
1543 | if (tg->bps_conf[READ][off] == bps_dft && | |
1544 | tg->bps_conf[WRITE][off] == bps_dft && | |
1545 | tg->iops_conf[READ][off] == iops_dft && | |
ada75b6e | 1546 | tg->iops_conf[WRITE][off] == iops_dft && |
ec80991d | 1547 | (off != LIMIT_LOW || |
b4f428ef | 1548 | (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD && |
5b81fc3c | 1549 | tg->latency_target_conf == DFL_LATENCY_TARGET))) |
2ee867dc TH |
1550 | return 0; |
1551 | ||
9bb67aeb | 1552 | if (tg->bps_conf[READ][off] != U64_MAX) |
9f626e37 | 1553 | snprintf(bufs[0], sizeof(bufs[0]), "%llu", |
cd5ab1b0 | 1554 | tg->bps_conf[READ][off]); |
9bb67aeb | 1555 | if (tg->bps_conf[WRITE][off] != U64_MAX) |
9f626e37 | 1556 | snprintf(bufs[1], sizeof(bufs[1]), "%llu", |
cd5ab1b0 | 1557 | tg->bps_conf[WRITE][off]); |
9bb67aeb | 1558 | if (tg->iops_conf[READ][off] != UINT_MAX) |
9f626e37 | 1559 | snprintf(bufs[2], sizeof(bufs[2]), "%u", |
cd5ab1b0 | 1560 | tg->iops_conf[READ][off]); |
9bb67aeb | 1561 | if (tg->iops_conf[WRITE][off] != UINT_MAX) |
9f626e37 | 1562 | snprintf(bufs[3], sizeof(bufs[3]), "%u", |
cd5ab1b0 | 1563 | tg->iops_conf[WRITE][off]); |
ada75b6e | 1564 | if (off == LIMIT_LOW) { |
5b81fc3c | 1565 | if (tg->idletime_threshold_conf == ULONG_MAX) |
ada75b6e SL |
1566 | strcpy(idle_time, " idle=max"); |
1567 | else | |
1568 | snprintf(idle_time, sizeof(idle_time), " idle=%lu", | |
5b81fc3c | 1569 | tg->idletime_threshold_conf); |
ec80991d | 1570 | |
5b81fc3c | 1571 | if (tg->latency_target_conf == ULONG_MAX) |
ec80991d SL |
1572 | strcpy(latency_time, " latency=max"); |
1573 | else | |
1574 | snprintf(latency_time, sizeof(latency_time), | |
5b81fc3c | 1575 | " latency=%lu", tg->latency_target_conf); |
ada75b6e | 1576 | } |
2ee867dc | 1577 | |
ec80991d SL |
1578 | seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n", |
1579 | dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time, | |
1580 | latency_time); | |
2ee867dc TH |
1581 | return 0; |
1582 | } | |
1583 | ||
cd5ab1b0 | 1584 | static int tg_print_limit(struct seq_file *sf, void *v) |
2ee867dc | 1585 | { |
cd5ab1b0 | 1586 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit, |
2ee867dc TH |
1587 | &blkcg_policy_throtl, seq_cft(sf)->private, false); |
1588 | return 0; | |
1589 | } | |
1590 | ||
cd5ab1b0 | 1591 | static ssize_t tg_set_limit(struct kernfs_open_file *of, |
2ee867dc TH |
1592 | char *buf, size_t nbytes, loff_t off) |
1593 | { | |
1594 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1595 | struct blkg_conf_ctx ctx; | |
1596 | struct throtl_grp *tg; | |
1597 | u64 v[4]; | |
ada75b6e | 1598 | unsigned long idle_time; |
ec80991d | 1599 | unsigned long latency_time; |
2ee867dc | 1600 | int ret; |
cd5ab1b0 | 1601 | int index = of_cft(of)->private; |
2ee867dc TH |
1602 | |
1603 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1604 | if (ret) | |
1605 | return ret; | |
1606 | ||
1607 | tg = blkg_to_tg(ctx.blkg); | |
1608 | ||
cd5ab1b0 SL |
1609 | v[0] = tg->bps_conf[READ][index]; |
1610 | v[1] = tg->bps_conf[WRITE][index]; | |
1611 | v[2] = tg->iops_conf[READ][index]; | |
1612 | v[3] = tg->iops_conf[WRITE][index]; | |
2ee867dc | 1613 | |
5b81fc3c SL |
1614 | idle_time = tg->idletime_threshold_conf; |
1615 | latency_time = tg->latency_target_conf; | |
2ee867dc TH |
1616 | while (true) { |
1617 | char tok[27]; /* wiops=18446744073709551616 */ | |
1618 | char *p; | |
2ab5492d | 1619 | u64 val = U64_MAX; |
2ee867dc TH |
1620 | int len; |
1621 | ||
1622 | if (sscanf(ctx.body, "%26s%n", tok, &len) != 1) | |
1623 | break; | |
1624 | if (tok[0] == '\0') | |
1625 | break; | |
1626 | ctx.body += len; | |
1627 | ||
1628 | ret = -EINVAL; | |
1629 | p = tok; | |
1630 | strsep(&p, "="); | |
1631 | if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max"))) | |
1632 | goto out_finish; | |
1633 | ||
1634 | ret = -ERANGE; | |
1635 | if (!val) | |
1636 | goto out_finish; | |
1637 | ||
1638 | ret = -EINVAL; | |
1639 | if (!strcmp(tok, "rbps")) | |
1640 | v[0] = val; | |
1641 | else if (!strcmp(tok, "wbps")) | |
1642 | v[1] = val; | |
1643 | else if (!strcmp(tok, "riops")) | |
1644 | v[2] = min_t(u64, val, UINT_MAX); | |
1645 | else if (!strcmp(tok, "wiops")) | |
1646 | v[3] = min_t(u64, val, UINT_MAX); | |
ada75b6e SL |
1647 | else if (off == LIMIT_LOW && !strcmp(tok, "idle")) |
1648 | idle_time = val; | |
ec80991d SL |
1649 | else if (off == LIMIT_LOW && !strcmp(tok, "latency")) |
1650 | latency_time = val; | |
2ee867dc TH |
1651 | else |
1652 | goto out_finish; | |
1653 | } | |
1654 | ||
cd5ab1b0 SL |
1655 | tg->bps_conf[READ][index] = v[0]; |
1656 | tg->bps_conf[WRITE][index] = v[1]; | |
1657 | tg->iops_conf[READ][index] = v[2]; | |
1658 | tg->iops_conf[WRITE][index] = v[3]; | |
2ee867dc | 1659 | |
cd5ab1b0 SL |
1660 | if (index == LIMIT_MAX) { |
1661 | tg->bps[READ][index] = v[0]; | |
1662 | tg->bps[WRITE][index] = v[1]; | |
1663 | tg->iops[READ][index] = v[2]; | |
1664 | tg->iops[WRITE][index] = v[3]; | |
1665 | } | |
1666 | tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW], | |
1667 | tg->bps_conf[READ][LIMIT_MAX]); | |
1668 | tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW], | |
1669 | tg->bps_conf[WRITE][LIMIT_MAX]); | |
1670 | tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW], | |
1671 | tg->iops_conf[READ][LIMIT_MAX]); | |
1672 | tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW], | |
1673 | tg->iops_conf[WRITE][LIMIT_MAX]); | |
b4f428ef SL |
1674 | tg->idletime_threshold_conf = idle_time; |
1675 | tg->latency_target_conf = latency_time; | |
1676 | ||
1677 | /* force user to configure all settings for low limit */ | |
1678 | if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] || | |
1679 | tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) || | |
1680 | tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD || | |
1681 | tg->latency_target_conf == DFL_LATENCY_TARGET) { | |
1682 | tg->bps[READ][LIMIT_LOW] = 0; | |
1683 | tg->bps[WRITE][LIMIT_LOW] = 0; | |
1684 | tg->iops[READ][LIMIT_LOW] = 0; | |
1685 | tg->iops[WRITE][LIMIT_LOW] = 0; | |
1686 | tg->idletime_threshold = DFL_IDLE_THRESHOLD; | |
1687 | tg->latency_target = DFL_LATENCY_TARGET; | |
1688 | } else if (index == LIMIT_LOW) { | |
5b81fc3c | 1689 | tg->idletime_threshold = tg->idletime_threshold_conf; |
5b81fc3c | 1690 | tg->latency_target = tg->latency_target_conf; |
cd5ab1b0 | 1691 | } |
b4f428ef SL |
1692 | |
1693 | blk_throtl_update_limit_valid(tg->td); | |
1694 | if (tg->td->limit_valid[LIMIT_LOW]) { | |
1695 | if (index == LIMIT_LOW) | |
1696 | tg->td->limit_index = LIMIT_LOW; | |
1697 | } else | |
1698 | tg->td->limit_index = LIMIT_MAX; | |
9bb67aeb SL |
1699 | tg_conf_updated(tg, index == LIMIT_LOW && |
1700 | tg->td->limit_valid[LIMIT_LOW]); | |
2ee867dc TH |
1701 | ret = 0; |
1702 | out_finish: | |
1703 | blkg_conf_finish(&ctx); | |
1704 | return ret ?: nbytes; | |
1705 | } | |
1706 | ||
1707 | static struct cftype throtl_files[] = { | |
cd5ab1b0 SL |
1708 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
1709 | { | |
1710 | .name = "low", | |
1711 | .flags = CFTYPE_NOT_ON_ROOT, | |
1712 | .seq_show = tg_print_limit, | |
1713 | .write = tg_set_limit, | |
1714 | .private = LIMIT_LOW, | |
1715 | }, | |
1716 | #endif | |
2ee867dc TH |
1717 | { |
1718 | .name = "max", | |
1719 | .flags = CFTYPE_NOT_ON_ROOT, | |
cd5ab1b0 SL |
1720 | .seq_show = tg_print_limit, |
1721 | .write = tg_set_limit, | |
1722 | .private = LIMIT_MAX, | |
2ee867dc TH |
1723 | }, |
1724 | { } /* terminate */ | |
1725 | }; | |
1726 | ||
da527770 | 1727 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1728 | { |
1729 | struct throtl_data *td = q->td; | |
1730 | ||
69df0ab0 | 1731 | cancel_work_sync(&td->dispatch_work); |
e43473b7 VG |
1732 | } |
1733 | ||
3c798398 | 1734 | static struct blkcg_policy blkcg_policy_throtl = { |
2ee867dc | 1735 | .dfl_cftypes = throtl_files, |
880f50e2 | 1736 | .legacy_cftypes = throtl_legacy_files, |
f9fcc2d3 | 1737 | |
001bea73 | 1738 | .pd_alloc_fn = throtl_pd_alloc, |
f9fcc2d3 | 1739 | .pd_init_fn = throtl_pd_init, |
693e751e | 1740 | .pd_online_fn = throtl_pd_online, |
cd5ab1b0 | 1741 | .pd_offline_fn = throtl_pd_offline, |
001bea73 | 1742 | .pd_free_fn = throtl_pd_free, |
e43473b7 VG |
1743 | }; |
1744 | ||
3f0abd80 SL |
1745 | static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg) |
1746 | { | |
1747 | unsigned long rtime = jiffies, wtime = jiffies; | |
1748 | ||
1749 | if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]) | |
1750 | rtime = tg->last_low_overflow_time[READ]; | |
1751 | if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) | |
1752 | wtime = tg->last_low_overflow_time[WRITE]; | |
1753 | return min(rtime, wtime); | |
1754 | } | |
1755 | ||
1756 | /* tg should not be an intermediate node */ | |
1757 | static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg) | |
1758 | { | |
1759 | struct throtl_service_queue *parent_sq; | |
1760 | struct throtl_grp *parent = tg; | |
1761 | unsigned long ret = __tg_last_low_overflow_time(tg); | |
1762 | ||
1763 | while (true) { | |
1764 | parent_sq = parent->service_queue.parent_sq; | |
1765 | parent = sq_to_tg(parent_sq); | |
1766 | if (!parent) | |
1767 | break; | |
1768 | ||
1769 | /* | |
1770 | * The parent doesn't have low limit, it always reaches low | |
1771 | * limit. Its overflow time is useless for children | |
1772 | */ | |
1773 | if (!parent->bps[READ][LIMIT_LOW] && | |
1774 | !parent->iops[READ][LIMIT_LOW] && | |
1775 | !parent->bps[WRITE][LIMIT_LOW] && | |
1776 | !parent->iops[WRITE][LIMIT_LOW]) | |
1777 | continue; | |
1778 | if (time_after(__tg_last_low_overflow_time(parent), ret)) | |
1779 | ret = __tg_last_low_overflow_time(parent); | |
1780 | } | |
1781 | return ret; | |
1782 | } | |
1783 | ||
9e234eea SL |
1784 | static bool throtl_tg_is_idle(struct throtl_grp *tg) |
1785 | { | |
1786 | /* | |
1787 | * cgroup is idle if: | |
1788 | * - single idle is too long, longer than a fixed value (in case user | |
b4f428ef | 1789 | * configure a too big threshold) or 4 times of idletime threshold |
9e234eea | 1790 | * - average think time is more than threshold |
53696b8d | 1791 | * - IO latency is largely below threshold |
9e234eea | 1792 | */ |
b4f428ef | 1793 | unsigned long time; |
4cff729f | 1794 | bool ret; |
9e234eea | 1795 | |
b4f428ef SL |
1796 | time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold); |
1797 | ret = tg->latency_target == DFL_LATENCY_TARGET || | |
1798 | tg->idletime_threshold == DFL_IDLE_THRESHOLD || | |
1799 | (ktime_get_ns() >> 10) - tg->last_finish_time > time || | |
1800 | tg->avg_idletime > tg->idletime_threshold || | |
1801 | (tg->latency_target && tg->bio_cnt && | |
53696b8d | 1802 | tg->bad_bio_cnt * 5 < tg->bio_cnt); |
4cff729f SL |
1803 | throtl_log(&tg->service_queue, |
1804 | "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d", | |
1805 | tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt, | |
1806 | tg->bio_cnt, ret, tg->td->scale); | |
1807 | return ret; | |
9e234eea SL |
1808 | } |
1809 | ||
c79892c5 SL |
1810 | static bool throtl_tg_can_upgrade(struct throtl_grp *tg) |
1811 | { | |
1812 | struct throtl_service_queue *sq = &tg->service_queue; | |
1813 | bool read_limit, write_limit; | |
1814 | ||
1815 | /* | |
1816 | * if cgroup reaches low limit (if low limit is 0, the cgroup always | |
1817 | * reaches), it's ok to upgrade to next limit | |
1818 | */ | |
1819 | read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]; | |
1820 | write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]; | |
1821 | if (!read_limit && !write_limit) | |
1822 | return true; | |
1823 | if (read_limit && sq->nr_queued[READ] && | |
1824 | (!write_limit || sq->nr_queued[WRITE])) | |
1825 | return true; | |
1826 | if (write_limit && sq->nr_queued[WRITE] && | |
1827 | (!read_limit || sq->nr_queued[READ])) | |
1828 | return true; | |
aec24246 SL |
1829 | |
1830 | if (time_after_eq(jiffies, | |
fa6fb5aa SL |
1831 | tg_last_low_overflow_time(tg) + tg->td->throtl_slice) && |
1832 | throtl_tg_is_idle(tg)) | |
aec24246 | 1833 | return true; |
c79892c5 SL |
1834 | return false; |
1835 | } | |
1836 | ||
1837 | static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg) | |
1838 | { | |
1839 | while (true) { | |
1840 | if (throtl_tg_can_upgrade(tg)) | |
1841 | return true; | |
1842 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1843 | if (!tg || !tg_to_blkg(tg)->parent) | |
1844 | return false; | |
1845 | } | |
1846 | return false; | |
1847 | } | |
1848 | ||
1849 | static bool throtl_can_upgrade(struct throtl_data *td, | |
1850 | struct throtl_grp *this_tg) | |
1851 | { | |
1852 | struct cgroup_subsys_state *pos_css; | |
1853 | struct blkcg_gq *blkg; | |
1854 | ||
1855 | if (td->limit_index != LIMIT_LOW) | |
1856 | return false; | |
1857 | ||
297e3d85 | 1858 | if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice)) |
3f0abd80 SL |
1859 | return false; |
1860 | ||
c79892c5 SL |
1861 | rcu_read_lock(); |
1862 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1863 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1864 | ||
1865 | if (tg == this_tg) | |
1866 | continue; | |
1867 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1868 | continue; | |
1869 | if (!throtl_hierarchy_can_upgrade(tg)) { | |
1870 | rcu_read_unlock(); | |
1871 | return false; | |
1872 | } | |
1873 | } | |
1874 | rcu_read_unlock(); | |
1875 | return true; | |
1876 | } | |
1877 | ||
fa6fb5aa SL |
1878 | static void throtl_upgrade_check(struct throtl_grp *tg) |
1879 | { | |
1880 | unsigned long now = jiffies; | |
1881 | ||
1882 | if (tg->td->limit_index != LIMIT_LOW) | |
1883 | return; | |
1884 | ||
1885 | if (time_after(tg->last_check_time + tg->td->throtl_slice, now)) | |
1886 | return; | |
1887 | ||
1888 | tg->last_check_time = now; | |
1889 | ||
1890 | if (!time_after_eq(now, | |
1891 | __tg_last_low_overflow_time(tg) + tg->td->throtl_slice)) | |
1892 | return; | |
1893 | ||
1894 | if (throtl_can_upgrade(tg->td, NULL)) | |
1895 | throtl_upgrade_state(tg->td); | |
1896 | } | |
1897 | ||
c79892c5 SL |
1898 | static void throtl_upgrade_state(struct throtl_data *td) |
1899 | { | |
1900 | struct cgroup_subsys_state *pos_css; | |
1901 | struct blkcg_gq *blkg; | |
1902 | ||
4cff729f | 1903 | throtl_log(&td->service_queue, "upgrade to max"); |
c79892c5 | 1904 | td->limit_index = LIMIT_MAX; |
3f0abd80 | 1905 | td->low_upgrade_time = jiffies; |
7394e31f | 1906 | td->scale = 0; |
c79892c5 SL |
1907 | rcu_read_lock(); |
1908 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1909 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1910 | struct throtl_service_queue *sq = &tg->service_queue; | |
1911 | ||
1912 | tg->disptime = jiffies - 1; | |
1913 | throtl_select_dispatch(sq); | |
1914 | throtl_schedule_next_dispatch(sq, false); | |
1915 | } | |
1916 | rcu_read_unlock(); | |
1917 | throtl_select_dispatch(&td->service_queue); | |
1918 | throtl_schedule_next_dispatch(&td->service_queue, false); | |
1919 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1920 | } | |
1921 | ||
3f0abd80 SL |
1922 | static void throtl_downgrade_state(struct throtl_data *td, int new) |
1923 | { | |
7394e31f SL |
1924 | td->scale /= 2; |
1925 | ||
4cff729f | 1926 | throtl_log(&td->service_queue, "downgrade, scale %d", td->scale); |
7394e31f SL |
1927 | if (td->scale) { |
1928 | td->low_upgrade_time = jiffies - td->scale * td->throtl_slice; | |
1929 | return; | |
1930 | } | |
1931 | ||
3f0abd80 SL |
1932 | td->limit_index = new; |
1933 | td->low_downgrade_time = jiffies; | |
1934 | } | |
1935 | ||
1936 | static bool throtl_tg_can_downgrade(struct throtl_grp *tg) | |
1937 | { | |
1938 | struct throtl_data *td = tg->td; | |
1939 | unsigned long now = jiffies; | |
1940 | ||
1941 | /* | |
1942 | * If cgroup is below low limit, consider downgrade and throttle other | |
1943 | * cgroups | |
1944 | */ | |
297e3d85 SL |
1945 | if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) && |
1946 | time_after_eq(now, tg_last_low_overflow_time(tg) + | |
fa6fb5aa SL |
1947 | td->throtl_slice) && |
1948 | (!throtl_tg_is_idle(tg) || | |
1949 | !list_empty(&tg_to_blkg(tg)->blkcg->css.children))) | |
3f0abd80 SL |
1950 | return true; |
1951 | return false; | |
1952 | } | |
1953 | ||
1954 | static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg) | |
1955 | { | |
1956 | while (true) { | |
1957 | if (!throtl_tg_can_downgrade(tg)) | |
1958 | return false; | |
1959 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1960 | if (!tg || !tg_to_blkg(tg)->parent) | |
1961 | break; | |
1962 | } | |
1963 | return true; | |
1964 | } | |
1965 | ||
1966 | static void throtl_downgrade_check(struct throtl_grp *tg) | |
1967 | { | |
1968 | uint64_t bps; | |
1969 | unsigned int iops; | |
1970 | unsigned long elapsed_time; | |
1971 | unsigned long now = jiffies; | |
1972 | ||
1973 | if (tg->td->limit_index != LIMIT_MAX || | |
1974 | !tg->td->limit_valid[LIMIT_LOW]) | |
1975 | return; | |
1976 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1977 | return; | |
297e3d85 | 1978 | if (time_after(tg->last_check_time + tg->td->throtl_slice, now)) |
3f0abd80 SL |
1979 | return; |
1980 | ||
1981 | elapsed_time = now - tg->last_check_time; | |
1982 | tg->last_check_time = now; | |
1983 | ||
297e3d85 SL |
1984 | if (time_before(now, tg_last_low_overflow_time(tg) + |
1985 | tg->td->throtl_slice)) | |
3f0abd80 SL |
1986 | return; |
1987 | ||
1988 | if (tg->bps[READ][LIMIT_LOW]) { | |
1989 | bps = tg->last_bytes_disp[READ] * HZ; | |
1990 | do_div(bps, elapsed_time); | |
1991 | if (bps >= tg->bps[READ][LIMIT_LOW]) | |
1992 | tg->last_low_overflow_time[READ] = now; | |
1993 | } | |
1994 | ||
1995 | if (tg->bps[WRITE][LIMIT_LOW]) { | |
1996 | bps = tg->last_bytes_disp[WRITE] * HZ; | |
1997 | do_div(bps, elapsed_time); | |
1998 | if (bps >= tg->bps[WRITE][LIMIT_LOW]) | |
1999 | tg->last_low_overflow_time[WRITE] = now; | |
2000 | } | |
2001 | ||
2002 | if (tg->iops[READ][LIMIT_LOW]) { | |
2003 | iops = tg->last_io_disp[READ] * HZ / elapsed_time; | |
2004 | if (iops >= tg->iops[READ][LIMIT_LOW]) | |
2005 | tg->last_low_overflow_time[READ] = now; | |
2006 | } | |
2007 | ||
2008 | if (tg->iops[WRITE][LIMIT_LOW]) { | |
2009 | iops = tg->last_io_disp[WRITE] * HZ / elapsed_time; | |
2010 | if (iops >= tg->iops[WRITE][LIMIT_LOW]) | |
2011 | tg->last_low_overflow_time[WRITE] = now; | |
2012 | } | |
2013 | ||
2014 | /* | |
2015 | * If cgroup is below low limit, consider downgrade and throttle other | |
2016 | * cgroups | |
2017 | */ | |
2018 | if (throtl_hierarchy_can_downgrade(tg)) | |
2019 | throtl_downgrade_state(tg->td, LIMIT_LOW); | |
2020 | ||
2021 | tg->last_bytes_disp[READ] = 0; | |
2022 | tg->last_bytes_disp[WRITE] = 0; | |
2023 | tg->last_io_disp[READ] = 0; | |
2024 | tg->last_io_disp[WRITE] = 0; | |
2025 | } | |
2026 | ||
9e234eea SL |
2027 | static void blk_throtl_update_idletime(struct throtl_grp *tg) |
2028 | { | |
2029 | unsigned long now = ktime_get_ns() >> 10; | |
2030 | unsigned long last_finish_time = tg->last_finish_time; | |
2031 | ||
2032 | if (now <= last_finish_time || last_finish_time == 0 || | |
2033 | last_finish_time == tg->checked_last_finish_time) | |
2034 | return; | |
2035 | ||
2036 | tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3; | |
2037 | tg->checked_last_finish_time = last_finish_time; | |
2038 | } | |
2039 | ||
b9147dd1 SL |
2040 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
2041 | static void throtl_update_latency_buckets(struct throtl_data *td) | |
2042 | { | |
2043 | struct avg_latency_bucket avg_latency[LATENCY_BUCKET_SIZE]; | |
2044 | int i, cpu; | |
2045 | unsigned long last_latency = 0; | |
2046 | unsigned long latency; | |
2047 | ||
2048 | if (!blk_queue_nonrot(td->queue)) | |
2049 | return; | |
2050 | if (time_before(jiffies, td->last_calculate_time + HZ)) | |
2051 | return; | |
2052 | td->last_calculate_time = jiffies; | |
2053 | ||
2054 | memset(avg_latency, 0, sizeof(avg_latency)); | |
2055 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) { | |
2056 | struct latency_bucket *tmp = &td->tmp_buckets[i]; | |
2057 | ||
2058 | for_each_possible_cpu(cpu) { | |
2059 | struct latency_bucket *bucket; | |
2060 | ||
2061 | /* this isn't race free, but ok in practice */ | |
2062 | bucket = per_cpu_ptr(td->latency_buckets, cpu); | |
2063 | tmp->total_latency += bucket[i].total_latency; | |
2064 | tmp->samples += bucket[i].samples; | |
2065 | bucket[i].total_latency = 0; | |
2066 | bucket[i].samples = 0; | |
2067 | } | |
2068 | ||
2069 | if (tmp->samples >= 32) { | |
2070 | int samples = tmp->samples; | |
2071 | ||
2072 | latency = tmp->total_latency; | |
2073 | ||
2074 | tmp->total_latency = 0; | |
2075 | tmp->samples = 0; | |
2076 | latency /= samples; | |
2077 | if (latency == 0) | |
2078 | continue; | |
2079 | avg_latency[i].latency = latency; | |
2080 | } | |
2081 | } | |
2082 | ||
2083 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) { | |
2084 | if (!avg_latency[i].latency) { | |
2085 | if (td->avg_buckets[i].latency < last_latency) | |
2086 | td->avg_buckets[i].latency = last_latency; | |
2087 | continue; | |
2088 | } | |
2089 | ||
2090 | if (!td->avg_buckets[i].valid) | |
2091 | latency = avg_latency[i].latency; | |
2092 | else | |
2093 | latency = (td->avg_buckets[i].latency * 7 + | |
2094 | avg_latency[i].latency) >> 3; | |
2095 | ||
2096 | td->avg_buckets[i].latency = max(latency, last_latency); | |
2097 | td->avg_buckets[i].valid = true; | |
2098 | last_latency = td->avg_buckets[i].latency; | |
2099 | } | |
4cff729f SL |
2100 | |
2101 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) | |
2102 | throtl_log(&td->service_queue, | |
2103 | "Latency bucket %d: latency=%ld, valid=%d", i, | |
2104 | td->avg_buckets[i].latency, td->avg_buckets[i].valid); | |
b9147dd1 SL |
2105 | } |
2106 | #else | |
2107 | static inline void throtl_update_latency_buckets(struct throtl_data *td) | |
2108 | { | |
2109 | } | |
2110 | #endif | |
2111 | ||
2bc19cd5 JA |
2112 | static void blk_throtl_assoc_bio(struct throtl_grp *tg, struct bio *bio) |
2113 | { | |
2114 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
53cfdc10 JX |
2115 | if (bio->bi_css) { |
2116 | if (bio->bi_cg_private) | |
2117 | blkg_put(tg_to_blkg(bio->bi_cg_private)); | |
2bc19cd5 | 2118 | bio->bi_cg_private = tg; |
53cfdc10 JX |
2119 | blkg_get(tg_to_blkg(tg)); |
2120 | } | |
2bc19cd5 | 2121 | blk_stat_set_issue(&bio->bi_issue_stat, bio_sectors(bio)); |
2bc19cd5 JA |
2122 | #endif |
2123 | } | |
2124 | ||
ae118896 TH |
2125 | bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg, |
2126 | struct bio *bio) | |
e43473b7 | 2127 | { |
c5cc2070 | 2128 | struct throtl_qnode *qn = NULL; |
ae118896 | 2129 | struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg); |
73f0d49a | 2130 | struct throtl_service_queue *sq; |
0e9f4164 | 2131 | bool rw = bio_data_dir(bio); |
bc16a4f9 | 2132 | bool throttled = false; |
b9147dd1 | 2133 | struct throtl_data *td = tg->td; |
e43473b7 | 2134 | |
ae118896 TH |
2135 | WARN_ON_ONCE(!rcu_read_lock_held()); |
2136 | ||
2a0f61e6 | 2137 | /* see throtl_charge_bio() */ |
8d2bbd4c | 2138 | if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw]) |
bc16a4f9 | 2139 | goto out; |
e43473b7 VG |
2140 | |
2141 | spin_lock_irq(q->queue_lock); | |
c9589f03 | 2142 | |
b9147dd1 SL |
2143 | throtl_update_latency_buckets(td); |
2144 | ||
c9589f03 | 2145 | if (unlikely(blk_queue_bypass(q))) |
bc16a4f9 | 2146 | goto out_unlock; |
f469a7b4 | 2147 | |
2bc19cd5 | 2148 | blk_throtl_assoc_bio(tg, bio); |
9e234eea SL |
2149 | blk_throtl_update_idletime(tg); |
2150 | ||
73f0d49a TH |
2151 | sq = &tg->service_queue; |
2152 | ||
c79892c5 | 2153 | again: |
9e660acf | 2154 | while (true) { |
3f0abd80 SL |
2155 | if (tg->last_low_overflow_time[rw] == 0) |
2156 | tg->last_low_overflow_time[rw] = jiffies; | |
2157 | throtl_downgrade_check(tg); | |
fa6fb5aa | 2158 | throtl_upgrade_check(tg); |
9e660acf TH |
2159 | /* throtl is FIFO - if bios are already queued, should queue */ |
2160 | if (sq->nr_queued[rw]) | |
2161 | break; | |
de701c74 | 2162 | |
9e660acf | 2163 | /* if above limits, break to queue */ |
c79892c5 | 2164 | if (!tg_may_dispatch(tg, bio, NULL)) { |
3f0abd80 | 2165 | tg->last_low_overflow_time[rw] = jiffies; |
b9147dd1 SL |
2166 | if (throtl_can_upgrade(td, tg)) { |
2167 | throtl_upgrade_state(td); | |
c79892c5 SL |
2168 | goto again; |
2169 | } | |
9e660acf | 2170 | break; |
c79892c5 | 2171 | } |
9e660acf TH |
2172 | |
2173 | /* within limits, let's charge and dispatch directly */ | |
e43473b7 | 2174 | throtl_charge_bio(tg, bio); |
04521db0 VG |
2175 | |
2176 | /* | |
2177 | * We need to trim slice even when bios are not being queued | |
2178 | * otherwise it might happen that a bio is not queued for | |
2179 | * a long time and slice keeps on extending and trim is not | |
2180 | * called for a long time. Now if limits are reduced suddenly | |
2181 | * we take into account all the IO dispatched so far at new | |
2182 | * low rate and * newly queued IO gets a really long dispatch | |
2183 | * time. | |
2184 | * | |
2185 | * So keep on trimming slice even if bio is not queued. | |
2186 | */ | |
0f3457f6 | 2187 | throtl_trim_slice(tg, rw); |
9e660acf TH |
2188 | |
2189 | /* | |
2190 | * @bio passed through this layer without being throttled. | |
2191 | * Climb up the ladder. If we''re already at the top, it | |
2192 | * can be executed directly. | |
2193 | */ | |
c5cc2070 | 2194 | qn = &tg->qnode_on_parent[rw]; |
9e660acf TH |
2195 | sq = sq->parent_sq; |
2196 | tg = sq_to_tg(sq); | |
2197 | if (!tg) | |
2198 | goto out_unlock; | |
e43473b7 VG |
2199 | } |
2200 | ||
9e660acf | 2201 | /* out-of-limit, queue to @tg */ |
fda6f272 TH |
2202 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
2203 | rw == READ ? 'R' : 'W', | |
9f626e37 SL |
2204 | tg->bytes_disp[rw], bio->bi_iter.bi_size, |
2205 | tg_bps_limit(tg, rw), | |
2206 | tg->io_disp[rw], tg_iops_limit(tg, rw), | |
fda6f272 | 2207 | sq->nr_queued[READ], sq->nr_queued[WRITE]); |
e43473b7 | 2208 | |
3f0abd80 SL |
2209 | tg->last_low_overflow_time[rw] = jiffies; |
2210 | ||
b9147dd1 | 2211 | td->nr_queued[rw]++; |
c5cc2070 | 2212 | throtl_add_bio_tg(bio, qn, tg); |
bc16a4f9 | 2213 | throttled = true; |
e43473b7 | 2214 | |
7f52f98c TH |
2215 | /* |
2216 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
2217 | * was empty before @bio. The forced scheduling isn't likely to | |
2218 | * cause undue delay as @bio is likely to be dispatched directly if | |
2219 | * its @tg's disptime is not in the future. | |
2220 | */ | |
0e9f4164 | 2221 | if (tg->flags & THROTL_TG_WAS_EMPTY) { |
77216b04 | 2222 | tg_update_disptime(tg); |
7f52f98c | 2223 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); |
e43473b7 VG |
2224 | } |
2225 | ||
bc16a4f9 | 2226 | out_unlock: |
e43473b7 | 2227 | spin_unlock_irq(q->queue_lock); |
bc16a4f9 | 2228 | out: |
2a0f61e6 TH |
2229 | /* |
2230 | * As multiple blk-throtls may stack in the same issue path, we | |
2231 | * don't want bios to leave with the flag set. Clear the flag if | |
2232 | * being issued. | |
2233 | */ | |
2234 | if (!throttled) | |
8d2bbd4c | 2235 | bio_clear_flag(bio, BIO_THROTTLED); |
b9147dd1 SL |
2236 | |
2237 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
2238 | if (throttled || !td->track_bio_latency) | |
2239 | bio->bi_issue_stat.stat |= SKIP_LATENCY; | |
2240 | #endif | |
bc16a4f9 | 2241 | return throttled; |
e43473b7 VG |
2242 | } |
2243 | ||
9e234eea | 2244 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
b9147dd1 SL |
2245 | static void throtl_track_latency(struct throtl_data *td, sector_t size, |
2246 | int op, unsigned long time) | |
2247 | { | |
2248 | struct latency_bucket *latency; | |
2249 | int index; | |
2250 | ||
2251 | if (!td || td->limit_index != LIMIT_LOW || op != REQ_OP_READ || | |
2252 | !blk_queue_nonrot(td->queue)) | |
2253 | return; | |
2254 | ||
2255 | index = request_bucket_index(size); | |
2256 | ||
2257 | latency = get_cpu_ptr(td->latency_buckets); | |
2258 | latency[index].total_latency += time; | |
2259 | latency[index].samples++; | |
2260 | put_cpu_ptr(td->latency_buckets); | |
2261 | } | |
2262 | ||
2263 | void blk_throtl_stat_add(struct request *rq, u64 time_ns) | |
2264 | { | |
2265 | struct request_queue *q = rq->q; | |
2266 | struct throtl_data *td = q->td; | |
2267 | ||
2268 | throtl_track_latency(td, blk_stat_size(&rq->issue_stat), | |
2269 | req_op(rq), time_ns >> 10); | |
2270 | } | |
2271 | ||
9e234eea SL |
2272 | void blk_throtl_bio_endio(struct bio *bio) |
2273 | { | |
2274 | struct throtl_grp *tg; | |
b9147dd1 SL |
2275 | u64 finish_time_ns; |
2276 | unsigned long finish_time; | |
2277 | unsigned long start_time; | |
2278 | unsigned long lat; | |
9e234eea SL |
2279 | |
2280 | tg = bio->bi_cg_private; | |
2281 | if (!tg) | |
2282 | return; | |
2283 | bio->bi_cg_private = NULL; | |
2284 | ||
b9147dd1 SL |
2285 | finish_time_ns = ktime_get_ns(); |
2286 | tg->last_finish_time = finish_time_ns >> 10; | |
2287 | ||
2288 | start_time = blk_stat_time(&bio->bi_issue_stat) >> 10; | |
2289 | finish_time = __blk_stat_time(finish_time_ns) >> 10; | |
53cfdc10 JX |
2290 | if (!start_time || finish_time <= start_time) { |
2291 | blkg_put(tg_to_blkg(tg)); | |
53696b8d | 2292 | return; |
53cfdc10 | 2293 | } |
53696b8d SL |
2294 | |
2295 | lat = finish_time - start_time; | |
b9147dd1 | 2296 | /* this is only for bio based driver */ |
53696b8d | 2297 | if (!(bio->bi_issue_stat.stat & SKIP_LATENCY)) |
b9147dd1 SL |
2298 | throtl_track_latency(tg->td, blk_stat_size(&bio->bi_issue_stat), |
2299 | bio_op(bio), lat); | |
53696b8d | 2300 | |
6679a90c | 2301 | if (tg->latency_target && lat >= tg->td->filtered_latency) { |
53696b8d SL |
2302 | int bucket; |
2303 | unsigned int threshold; | |
2304 | ||
2305 | bucket = request_bucket_index( | |
2306 | blk_stat_size(&bio->bi_issue_stat)); | |
2307 | threshold = tg->td->avg_buckets[bucket].latency + | |
2308 | tg->latency_target; | |
2309 | if (lat > threshold) | |
2310 | tg->bad_bio_cnt++; | |
2311 | /* | |
2312 | * Not race free, could get wrong count, which means cgroups | |
2313 | * will be throttled | |
2314 | */ | |
2315 | tg->bio_cnt++; | |
2316 | } | |
2317 | ||
2318 | if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) { | |
2319 | tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies; | |
2320 | tg->bio_cnt /= 2; | |
2321 | tg->bad_bio_cnt /= 2; | |
b9147dd1 | 2322 | } |
53cfdc10 JX |
2323 | |
2324 | blkg_put(tg_to_blkg(tg)); | |
9e234eea SL |
2325 | } |
2326 | #endif | |
2327 | ||
2a12f0dc TH |
2328 | /* |
2329 | * Dispatch all bios from all children tg's queued on @parent_sq. On | |
2330 | * return, @parent_sq is guaranteed to not have any active children tg's | |
2331 | * and all bios from previously active tg's are on @parent_sq->bio_lists[]. | |
2332 | */ | |
2333 | static void tg_drain_bios(struct throtl_service_queue *parent_sq) | |
2334 | { | |
2335 | struct throtl_grp *tg; | |
2336 | ||
2337 | while ((tg = throtl_rb_first(parent_sq))) { | |
2338 | struct throtl_service_queue *sq = &tg->service_queue; | |
2339 | struct bio *bio; | |
2340 | ||
2341 | throtl_dequeue_tg(tg); | |
2342 | ||
c5cc2070 | 2343 | while ((bio = throtl_peek_queued(&sq->queued[READ]))) |
2a12f0dc | 2344 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
c5cc2070 | 2345 | while ((bio = throtl_peek_queued(&sq->queued[WRITE]))) |
2a12f0dc TH |
2346 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
2347 | } | |
2348 | } | |
2349 | ||
c9a929dd TH |
2350 | /** |
2351 | * blk_throtl_drain - drain throttled bios | |
2352 | * @q: request_queue to drain throttled bios for | |
2353 | * | |
2354 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
2355 | */ | |
2356 | void blk_throtl_drain(struct request_queue *q) | |
2357 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
2358 | { | |
2359 | struct throtl_data *td = q->td; | |
2a12f0dc | 2360 | struct blkcg_gq *blkg; |
492eb21b | 2361 | struct cgroup_subsys_state *pos_css; |
c9a929dd | 2362 | struct bio *bio; |
651930bc | 2363 | int rw; |
c9a929dd | 2364 | |
8bcb6c7d | 2365 | queue_lockdep_assert_held(q); |
2a12f0dc | 2366 | rcu_read_lock(); |
c9a929dd | 2367 | |
2a12f0dc TH |
2368 | /* |
2369 | * Drain each tg while doing post-order walk on the blkg tree, so | |
2370 | * that all bios are propagated to td->service_queue. It'd be | |
2371 | * better to walk service_queue tree directly but blkg walk is | |
2372 | * easier. | |
2373 | */ | |
492eb21b | 2374 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) |
2a12f0dc | 2375 | tg_drain_bios(&blkg_to_tg(blkg)->service_queue); |
73f0d49a | 2376 | |
2a12f0dc TH |
2377 | /* finally, transfer bios from top-level tg's into the td */ |
2378 | tg_drain_bios(&td->service_queue); | |
2379 | ||
2380 | rcu_read_unlock(); | |
c9a929dd TH |
2381 | spin_unlock_irq(q->queue_lock); |
2382 | ||
2a12f0dc | 2383 | /* all bios now should be in td->service_queue, issue them */ |
651930bc | 2384 | for (rw = READ; rw <= WRITE; rw++) |
c5cc2070 TH |
2385 | while ((bio = throtl_pop_queued(&td->service_queue.queued[rw], |
2386 | NULL))) | |
651930bc | 2387 | generic_make_request(bio); |
c9a929dd TH |
2388 | |
2389 | spin_lock_irq(q->queue_lock); | |
2390 | } | |
2391 | ||
e43473b7 VG |
2392 | int blk_throtl_init(struct request_queue *q) |
2393 | { | |
2394 | struct throtl_data *td; | |
a2b1693b | 2395 | int ret; |
e43473b7 VG |
2396 | |
2397 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
2398 | if (!td) | |
2399 | return -ENOMEM; | |
b9147dd1 SL |
2400 | td->latency_buckets = __alloc_percpu(sizeof(struct latency_bucket) * |
2401 | LATENCY_BUCKET_SIZE, __alignof__(u64)); | |
2402 | if (!td->latency_buckets) { | |
2403 | kfree(td); | |
2404 | return -ENOMEM; | |
2405 | } | |
e43473b7 | 2406 | |
69df0ab0 | 2407 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
b2ce2643 | 2408 | throtl_service_queue_init(&td->service_queue); |
e43473b7 | 2409 | |
cd1604fa | 2410 | q->td = td; |
29b12589 | 2411 | td->queue = q; |
02977e4a | 2412 | |
9f626e37 | 2413 | td->limit_valid[LIMIT_MAX] = true; |
cd5ab1b0 | 2414 | td->limit_index = LIMIT_MAX; |
3f0abd80 SL |
2415 | td->low_upgrade_time = jiffies; |
2416 | td->low_downgrade_time = jiffies; | |
9e234eea | 2417 | |
a2b1693b | 2418 | /* activate policy */ |
3c798398 | 2419 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
b9147dd1 SL |
2420 | if (ret) { |
2421 | free_percpu(td->latency_buckets); | |
f51b802c | 2422 | kfree(td); |
b9147dd1 | 2423 | } |
a2b1693b | 2424 | return ret; |
e43473b7 VG |
2425 | } |
2426 | ||
2427 | void blk_throtl_exit(struct request_queue *q) | |
2428 | { | |
c875f4d0 | 2429 | BUG_ON(!q->td); |
da527770 | 2430 | throtl_shutdown_wq(q); |
3c798398 | 2431 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
b9147dd1 | 2432 | free_percpu(q->td->latency_buckets); |
c9a929dd | 2433 | kfree(q->td); |
e43473b7 VG |
2434 | } |
2435 | ||
d61fcfa4 SL |
2436 | void blk_throtl_register_queue(struct request_queue *q) |
2437 | { | |
2438 | struct throtl_data *td; | |
6679a90c | 2439 | int i; |
d61fcfa4 SL |
2440 | |
2441 | td = q->td; | |
2442 | BUG_ON(!td); | |
2443 | ||
6679a90c | 2444 | if (blk_queue_nonrot(q)) { |
d61fcfa4 | 2445 | td->throtl_slice = DFL_THROTL_SLICE_SSD; |
6679a90c SL |
2446 | td->filtered_latency = LATENCY_FILTERED_SSD; |
2447 | } else { | |
d61fcfa4 | 2448 | td->throtl_slice = DFL_THROTL_SLICE_HD; |
6679a90c SL |
2449 | td->filtered_latency = LATENCY_FILTERED_HD; |
2450 | for (i = 0; i < LATENCY_BUCKET_SIZE; i++) | |
2451 | td->avg_buckets[i].latency = DFL_HD_BASELINE_LATENCY; | |
2452 | } | |
d61fcfa4 SL |
2453 | #ifndef CONFIG_BLK_DEV_THROTTLING_LOW |
2454 | /* if no low limit, use previous default */ | |
2455 | td->throtl_slice = DFL_THROTL_SLICE_HD; | |
2456 | #endif | |
9e234eea | 2457 | |
b9147dd1 SL |
2458 | td->track_bio_latency = !q->mq_ops && !q->request_fn; |
2459 | if (!td->track_bio_latency) | |
2460 | blk_stat_enable_accounting(q); | |
d61fcfa4 SL |
2461 | } |
2462 | ||
297e3d85 SL |
2463 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
2464 | ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page) | |
2465 | { | |
2466 | if (!q->td) | |
2467 | return -EINVAL; | |
2468 | return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice)); | |
2469 | } | |
2470 | ||
2471 | ssize_t blk_throtl_sample_time_store(struct request_queue *q, | |
2472 | const char *page, size_t count) | |
2473 | { | |
2474 | unsigned long v; | |
2475 | unsigned long t; | |
2476 | ||
2477 | if (!q->td) | |
2478 | return -EINVAL; | |
2479 | if (kstrtoul(page, 10, &v)) | |
2480 | return -EINVAL; | |
2481 | t = msecs_to_jiffies(v); | |
2482 | if (t == 0 || t > MAX_THROTL_SLICE) | |
2483 | return -EINVAL; | |
2484 | q->td->throtl_slice = t; | |
2485 | return count; | |
2486 | } | |
2487 | #endif | |
2488 | ||
e43473b7 VG |
2489 | static int __init throtl_init(void) |
2490 | { | |
450adcbe VG |
2491 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
2492 | if (!kthrotld_workqueue) | |
2493 | panic("Failed to create kthrotld\n"); | |
2494 | ||
3c798398 | 2495 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
2496 | } |
2497 | ||
2498 | module_init(throtl_init); |