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 SL |
25 | #define DFL_IDLE_THRESHOLD_SSD (1000L) /* 1 ms */ |
26 | #define DFL_IDLE_THRESHOLD_HD (100L * 1000) /* 100 ms */ | |
27 | #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */ | |
e43473b7 | 28 | |
3c798398 | 29 | static struct blkcg_policy blkcg_policy_throtl; |
0381411e | 30 | |
450adcbe VG |
31 | /* A workqueue to queue throttle related work */ |
32 | static struct workqueue_struct *kthrotld_workqueue; | |
450adcbe | 33 | |
c5cc2070 TH |
34 | /* |
35 | * To implement hierarchical throttling, throtl_grps form a tree and bios | |
36 | * are dispatched upwards level by level until they reach the top and get | |
37 | * issued. When dispatching bios from the children and local group at each | |
38 | * level, if the bios are dispatched into a single bio_list, there's a risk | |
39 | * of a local or child group which can queue many bios at once filling up | |
40 | * the list starving others. | |
41 | * | |
42 | * To avoid such starvation, dispatched bios are queued separately | |
43 | * according to where they came from. When they are again dispatched to | |
44 | * the parent, they're popped in round-robin order so that no single source | |
45 | * hogs the dispatch window. | |
46 | * | |
47 | * throtl_qnode is used to keep the queued bios separated by their sources. | |
48 | * Bios are queued to throtl_qnode which in turn is queued to | |
49 | * throtl_service_queue and then dispatched in round-robin order. | |
50 | * | |
51 | * It's also used to track the reference counts on blkg's. A qnode always | |
52 | * belongs to a throtl_grp and gets queued on itself or the parent, so | |
53 | * incrementing the reference of the associated throtl_grp when a qnode is | |
54 | * queued and decrementing when dequeued is enough to keep the whole blkg | |
55 | * tree pinned while bios are in flight. | |
56 | */ | |
57 | struct throtl_qnode { | |
58 | struct list_head node; /* service_queue->queued[] */ | |
59 | struct bio_list bios; /* queued bios */ | |
60 | struct throtl_grp *tg; /* tg this qnode belongs to */ | |
61 | }; | |
62 | ||
c9e0332e | 63 | struct throtl_service_queue { |
77216b04 TH |
64 | struct throtl_service_queue *parent_sq; /* the parent service_queue */ |
65 | ||
73f0d49a TH |
66 | /* |
67 | * Bios queued directly to this service_queue or dispatched from | |
68 | * children throtl_grp's. | |
69 | */ | |
c5cc2070 | 70 | struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */ |
73f0d49a TH |
71 | unsigned int nr_queued[2]; /* number of queued bios */ |
72 | ||
73 | /* | |
74 | * RB tree of active children throtl_grp's, which are sorted by | |
75 | * their ->disptime. | |
76 | */ | |
c9e0332e TH |
77 | struct rb_root pending_tree; /* RB tree of active tgs */ |
78 | struct rb_node *first_pending; /* first node in the tree */ | |
79 | unsigned int nr_pending; /* # queued in the tree */ | |
80 | unsigned long first_pending_disptime; /* disptime of the first tg */ | |
69df0ab0 | 81 | struct timer_list pending_timer; /* fires on first_pending_disptime */ |
e43473b7 VG |
82 | }; |
83 | ||
5b2c16aa TH |
84 | enum tg_state_flags { |
85 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
0e9f4164 | 86 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ |
5b2c16aa TH |
87 | }; |
88 | ||
e43473b7 VG |
89 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
90 | ||
9f626e37 | 91 | enum { |
cd5ab1b0 | 92 | LIMIT_LOW, |
9f626e37 SL |
93 | LIMIT_MAX, |
94 | LIMIT_CNT, | |
95 | }; | |
96 | ||
e43473b7 | 97 | struct throtl_grp { |
f95a04af TH |
98 | /* must be the first member */ |
99 | struct blkg_policy_data pd; | |
100 | ||
c9e0332e | 101 | /* active throtl group service_queue member */ |
e43473b7 VG |
102 | struct rb_node rb_node; |
103 | ||
0f3457f6 TH |
104 | /* throtl_data this group belongs to */ |
105 | struct throtl_data *td; | |
106 | ||
49a2f1e3 TH |
107 | /* this group's service queue */ |
108 | struct throtl_service_queue service_queue; | |
109 | ||
c5cc2070 TH |
110 | /* |
111 | * qnode_on_self is used when bios are directly queued to this | |
112 | * throtl_grp so that local bios compete fairly with bios | |
113 | * dispatched from children. qnode_on_parent is used when bios are | |
114 | * dispatched from this throtl_grp into its parent and will compete | |
115 | * with the sibling qnode_on_parents and the parent's | |
116 | * qnode_on_self. | |
117 | */ | |
118 | struct throtl_qnode qnode_on_self[2]; | |
119 | struct throtl_qnode qnode_on_parent[2]; | |
120 | ||
e43473b7 VG |
121 | /* |
122 | * Dispatch time in jiffies. This is the estimated time when group | |
123 | * will unthrottle and is ready to dispatch more bio. It is used as | |
124 | * key to sort active groups in service tree. | |
125 | */ | |
126 | unsigned long disptime; | |
127 | ||
e43473b7 VG |
128 | unsigned int flags; |
129 | ||
693e751e TH |
130 | /* are there any throtl rules between this group and td? */ |
131 | bool has_rules[2]; | |
132 | ||
cd5ab1b0 | 133 | /* internally used bytes per second rate limits */ |
9f626e37 | 134 | uint64_t bps[2][LIMIT_CNT]; |
cd5ab1b0 SL |
135 | /* user configured bps limits */ |
136 | uint64_t bps_conf[2][LIMIT_CNT]; | |
e43473b7 | 137 | |
cd5ab1b0 | 138 | /* internally used IOPS limits */ |
9f626e37 | 139 | unsigned int iops[2][LIMIT_CNT]; |
cd5ab1b0 SL |
140 | /* user configured IOPS limits */ |
141 | unsigned int iops_conf[2][LIMIT_CNT]; | |
8e89d13f | 142 | |
e43473b7 VG |
143 | /* Number of bytes disptached in current slice */ |
144 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
145 | /* Number of bio's dispatched in current slice */ |
146 | unsigned int io_disp[2]; | |
e43473b7 | 147 | |
3f0abd80 SL |
148 | unsigned long last_low_overflow_time[2]; |
149 | ||
150 | uint64_t last_bytes_disp[2]; | |
151 | unsigned int last_io_disp[2]; | |
152 | ||
153 | unsigned long last_check_time; | |
154 | ||
aec24246 SL |
155 | unsigned long last_dispatch_time[2]; |
156 | ||
e43473b7 VG |
157 | /* When did we start a new slice */ |
158 | unsigned long slice_start[2]; | |
159 | unsigned long slice_end[2]; | |
9e234eea SL |
160 | |
161 | unsigned long last_finish_time; /* ns / 1024 */ | |
162 | unsigned long checked_last_finish_time; /* ns / 1024 */ | |
163 | unsigned long avg_idletime; /* ns / 1024 */ | |
164 | unsigned long idletime_threshold; /* us */ | |
e43473b7 VG |
165 | }; |
166 | ||
167 | struct throtl_data | |
168 | { | |
e43473b7 | 169 | /* service tree for active throtl groups */ |
c9e0332e | 170 | struct throtl_service_queue service_queue; |
e43473b7 | 171 | |
e43473b7 VG |
172 | struct request_queue *queue; |
173 | ||
174 | /* Total Number of queued bios on READ and WRITE lists */ | |
175 | unsigned int nr_queued[2]; | |
176 | ||
297e3d85 SL |
177 | unsigned int throtl_slice; |
178 | ||
e43473b7 | 179 | /* Work for dispatching throttled bios */ |
69df0ab0 | 180 | struct work_struct dispatch_work; |
9f626e37 SL |
181 | unsigned int limit_index; |
182 | bool limit_valid[LIMIT_CNT]; | |
3f0abd80 SL |
183 | |
184 | unsigned long low_upgrade_time; | |
185 | unsigned long low_downgrade_time; | |
7394e31f SL |
186 | |
187 | unsigned int scale; | |
e43473b7 VG |
188 | }; |
189 | ||
69df0ab0 TH |
190 | static void throtl_pending_timer_fn(unsigned long arg); |
191 | ||
f95a04af TH |
192 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
193 | { | |
194 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
195 | } | |
196 | ||
3c798398 | 197 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 198 | { |
f95a04af | 199 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
200 | } |
201 | ||
3c798398 | 202 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 203 | { |
f95a04af | 204 | return pd_to_blkg(&tg->pd); |
0381411e TH |
205 | } |
206 | ||
fda6f272 TH |
207 | /** |
208 | * sq_to_tg - return the throl_grp the specified service queue belongs to | |
209 | * @sq: the throtl_service_queue of interest | |
210 | * | |
211 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one | |
212 | * embedded in throtl_data, %NULL is returned. | |
213 | */ | |
214 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) | |
215 | { | |
216 | if (sq && sq->parent_sq) | |
217 | return container_of(sq, struct throtl_grp, service_queue); | |
218 | else | |
219 | return NULL; | |
220 | } | |
221 | ||
222 | /** | |
223 | * sq_to_td - return throtl_data the specified service queue belongs to | |
224 | * @sq: the throtl_service_queue of interest | |
225 | * | |
b43daedc | 226 | * A service_queue can be embedded in either a throtl_grp or throtl_data. |
fda6f272 TH |
227 | * Determine the associated throtl_data accordingly and return it. |
228 | */ | |
229 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) | |
230 | { | |
231 | struct throtl_grp *tg = sq_to_tg(sq); | |
232 | ||
233 | if (tg) | |
234 | return tg->td; | |
235 | else | |
236 | return container_of(sq, struct throtl_data, service_queue); | |
237 | } | |
238 | ||
7394e31f SL |
239 | /* |
240 | * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to | |
241 | * make the IO dispatch more smooth. | |
242 | * Scale up: linearly scale up according to lapsed time since upgrade. For | |
243 | * every throtl_slice, the limit scales up 1/2 .low limit till the | |
244 | * limit hits .max limit | |
245 | * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit | |
246 | */ | |
247 | static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td) | |
248 | { | |
249 | /* arbitrary value to avoid too big scale */ | |
250 | if (td->scale < 4096 && time_after_eq(jiffies, | |
251 | td->low_upgrade_time + td->scale * td->throtl_slice)) | |
252 | td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice; | |
253 | ||
254 | return low + (low >> 1) * td->scale; | |
255 | } | |
256 | ||
9f626e37 SL |
257 | static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw) |
258 | { | |
b22c417c | 259 | struct blkcg_gq *blkg = tg_to_blkg(tg); |
7394e31f | 260 | struct throtl_data *td; |
b22c417c SL |
261 | uint64_t ret; |
262 | ||
263 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
264 | return U64_MAX; | |
7394e31f SL |
265 | |
266 | td = tg->td; | |
267 | ret = tg->bps[rw][td->limit_index]; | |
268 | if (ret == 0 && td->limit_index == LIMIT_LOW) | |
b22c417c | 269 | return tg->bps[rw][LIMIT_MAX]; |
7394e31f SL |
270 | |
271 | if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] && | |
272 | tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) { | |
273 | uint64_t adjusted; | |
274 | ||
275 | adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td); | |
276 | ret = min(tg->bps[rw][LIMIT_MAX], adjusted); | |
277 | } | |
b22c417c | 278 | return ret; |
9f626e37 SL |
279 | } |
280 | ||
281 | static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw) | |
282 | { | |
b22c417c | 283 | struct blkcg_gq *blkg = tg_to_blkg(tg); |
7394e31f | 284 | struct throtl_data *td; |
b22c417c SL |
285 | unsigned int ret; |
286 | ||
287 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
288 | return UINT_MAX; | |
7394e31f SL |
289 | td = tg->td; |
290 | ret = tg->iops[rw][td->limit_index]; | |
b22c417c SL |
291 | if (ret == 0 && tg->td->limit_index == LIMIT_LOW) |
292 | return tg->iops[rw][LIMIT_MAX]; | |
7394e31f SL |
293 | |
294 | if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] && | |
295 | tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) { | |
296 | uint64_t adjusted; | |
297 | ||
298 | adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td); | |
299 | if (adjusted > UINT_MAX) | |
300 | adjusted = UINT_MAX; | |
301 | ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted); | |
302 | } | |
b22c417c | 303 | return ret; |
9f626e37 SL |
304 | } |
305 | ||
fda6f272 TH |
306 | /** |
307 | * throtl_log - log debug message via blktrace | |
308 | * @sq: the service_queue being reported | |
309 | * @fmt: printf format string | |
310 | * @args: printf args | |
311 | * | |
312 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a | |
313 | * throtl_grp; otherwise, just "throtl". | |
fda6f272 TH |
314 | */ |
315 | #define throtl_log(sq, fmt, args...) do { \ | |
316 | struct throtl_grp *__tg = sq_to_tg((sq)); \ | |
317 | struct throtl_data *__td = sq_to_td((sq)); \ | |
318 | \ | |
319 | (void)__td; \ | |
59fa0224 SL |
320 | if (likely(!blk_trace_note_message_enabled(__td->queue))) \ |
321 | break; \ | |
fda6f272 TH |
322 | if ((__tg)) { \ |
323 | char __pbuf[128]; \ | |
54e7ed12 | 324 | \ |
fda6f272 TH |
325 | blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \ |
326 | blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
327 | } else { \ | |
328 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
329 | } \ | |
54e7ed12 | 330 | } while (0) |
e43473b7 | 331 | |
c5cc2070 TH |
332 | static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg) |
333 | { | |
334 | INIT_LIST_HEAD(&qn->node); | |
335 | bio_list_init(&qn->bios); | |
336 | qn->tg = tg; | |
337 | } | |
338 | ||
339 | /** | |
340 | * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it | |
341 | * @bio: bio being added | |
342 | * @qn: qnode to add bio to | |
343 | * @queued: the service_queue->queued[] list @qn belongs to | |
344 | * | |
345 | * Add @bio to @qn and put @qn on @queued if it's not already on. | |
346 | * @qn->tg's reference count is bumped when @qn is activated. See the | |
347 | * comment on top of throtl_qnode definition for details. | |
348 | */ | |
349 | static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn, | |
350 | struct list_head *queued) | |
351 | { | |
352 | bio_list_add(&qn->bios, bio); | |
353 | if (list_empty(&qn->node)) { | |
354 | list_add_tail(&qn->node, queued); | |
355 | blkg_get(tg_to_blkg(qn->tg)); | |
356 | } | |
357 | } | |
358 | ||
359 | /** | |
360 | * throtl_peek_queued - peek the first bio on a qnode list | |
361 | * @queued: the qnode list to peek | |
362 | */ | |
363 | static struct bio *throtl_peek_queued(struct list_head *queued) | |
364 | { | |
365 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
366 | struct bio *bio; | |
367 | ||
368 | if (list_empty(queued)) | |
369 | return NULL; | |
370 | ||
371 | bio = bio_list_peek(&qn->bios); | |
372 | WARN_ON_ONCE(!bio); | |
373 | return bio; | |
374 | } | |
375 | ||
376 | /** | |
377 | * throtl_pop_queued - pop the first bio form a qnode list | |
378 | * @queued: the qnode list to pop a bio from | |
379 | * @tg_to_put: optional out argument for throtl_grp to put | |
380 | * | |
381 | * Pop the first bio from the qnode list @queued. After popping, the first | |
382 | * qnode is removed from @queued if empty or moved to the end of @queued so | |
383 | * that the popping order is round-robin. | |
384 | * | |
385 | * When the first qnode is removed, its associated throtl_grp should be put | |
386 | * too. If @tg_to_put is NULL, this function automatically puts it; | |
387 | * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is | |
388 | * responsible for putting it. | |
389 | */ | |
390 | static struct bio *throtl_pop_queued(struct list_head *queued, | |
391 | struct throtl_grp **tg_to_put) | |
392 | { | |
393 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
394 | struct bio *bio; | |
395 | ||
396 | if (list_empty(queued)) | |
397 | return NULL; | |
398 | ||
399 | bio = bio_list_pop(&qn->bios); | |
400 | WARN_ON_ONCE(!bio); | |
401 | ||
402 | if (bio_list_empty(&qn->bios)) { | |
403 | list_del_init(&qn->node); | |
404 | if (tg_to_put) | |
405 | *tg_to_put = qn->tg; | |
406 | else | |
407 | blkg_put(tg_to_blkg(qn->tg)); | |
408 | } else { | |
409 | list_move_tail(&qn->node, queued); | |
410 | } | |
411 | ||
412 | return bio; | |
413 | } | |
414 | ||
49a2f1e3 | 415 | /* init a service_queue, assumes the caller zeroed it */ |
b2ce2643 | 416 | static void throtl_service_queue_init(struct throtl_service_queue *sq) |
49a2f1e3 | 417 | { |
c5cc2070 TH |
418 | INIT_LIST_HEAD(&sq->queued[0]); |
419 | INIT_LIST_HEAD(&sq->queued[1]); | |
49a2f1e3 | 420 | sq->pending_tree = RB_ROOT; |
69df0ab0 TH |
421 | setup_timer(&sq->pending_timer, throtl_pending_timer_fn, |
422 | (unsigned long)sq); | |
423 | } | |
424 | ||
001bea73 TH |
425 | static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node) |
426 | { | |
4fb72036 | 427 | struct throtl_grp *tg; |
24bdb8ef | 428 | int rw; |
4fb72036 TH |
429 | |
430 | tg = kzalloc_node(sizeof(*tg), gfp, node); | |
431 | if (!tg) | |
77ea7338 | 432 | return NULL; |
4fb72036 | 433 | |
b2ce2643 TH |
434 | throtl_service_queue_init(&tg->service_queue); |
435 | ||
436 | for (rw = READ; rw <= WRITE; rw++) { | |
437 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); | |
438 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); | |
439 | } | |
440 | ||
441 | RB_CLEAR_NODE(&tg->rb_node); | |
9f626e37 SL |
442 | tg->bps[READ][LIMIT_MAX] = U64_MAX; |
443 | tg->bps[WRITE][LIMIT_MAX] = U64_MAX; | |
444 | tg->iops[READ][LIMIT_MAX] = UINT_MAX; | |
445 | tg->iops[WRITE][LIMIT_MAX] = UINT_MAX; | |
cd5ab1b0 SL |
446 | tg->bps_conf[READ][LIMIT_MAX] = U64_MAX; |
447 | tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX; | |
448 | tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX; | |
449 | tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX; | |
450 | /* LIMIT_LOW will have default value 0 */ | |
b2ce2643 | 451 | |
4fb72036 | 452 | return &tg->pd; |
001bea73 TH |
453 | } |
454 | ||
a9520cd6 | 455 | static void throtl_pd_init(struct blkg_policy_data *pd) |
a29a171e | 456 | { |
a9520cd6 TH |
457 | struct throtl_grp *tg = pd_to_tg(pd); |
458 | struct blkcg_gq *blkg = tg_to_blkg(tg); | |
77216b04 | 459 | struct throtl_data *td = blkg->q->td; |
b2ce2643 | 460 | struct throtl_service_queue *sq = &tg->service_queue; |
cd1604fa | 461 | |
9138125b | 462 | /* |
aa6ec29b | 463 | * If on the default hierarchy, we switch to properly hierarchical |
9138125b TH |
464 | * behavior where limits on a given throtl_grp are applied to the |
465 | * whole subtree rather than just the group itself. e.g. If 16M | |
466 | * read_bps limit is set on the root group, the whole system can't | |
467 | * exceed 16M for the device. | |
468 | * | |
aa6ec29b | 469 | * If not on the default hierarchy, the broken flat hierarchy |
9138125b TH |
470 | * behavior is retained where all throtl_grps are treated as if |
471 | * they're all separate root groups right below throtl_data. | |
472 | * Limits of a group don't interact with limits of other groups | |
473 | * regardless of the position of the group in the hierarchy. | |
474 | */ | |
b2ce2643 | 475 | sq->parent_sq = &td->service_queue; |
9e10a130 | 476 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent) |
b2ce2643 | 477 | sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue; |
77216b04 | 478 | tg->td = td; |
9e234eea SL |
479 | |
480 | if (blk_queue_nonrot(td->queue)) | |
481 | tg->idletime_threshold = DFL_IDLE_THRESHOLD_SSD; | |
482 | else | |
483 | tg->idletime_threshold = DFL_IDLE_THRESHOLD_HD; | |
8a3d2615 TH |
484 | } |
485 | ||
693e751e TH |
486 | /* |
487 | * Set has_rules[] if @tg or any of its parents have limits configured. | |
488 | * This doesn't require walking up to the top of the hierarchy as the | |
489 | * parent's has_rules[] is guaranteed to be correct. | |
490 | */ | |
491 | static void tg_update_has_rules(struct throtl_grp *tg) | |
492 | { | |
493 | struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); | |
9f626e37 | 494 | struct throtl_data *td = tg->td; |
693e751e TH |
495 | int rw; |
496 | ||
497 | for (rw = READ; rw <= WRITE; rw++) | |
498 | tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) || | |
9f626e37 SL |
499 | (td->limit_valid[td->limit_index] && |
500 | (tg_bps_limit(tg, rw) != U64_MAX || | |
501 | tg_iops_limit(tg, rw) != UINT_MAX)); | |
693e751e TH |
502 | } |
503 | ||
a9520cd6 | 504 | static void throtl_pd_online(struct blkg_policy_data *pd) |
693e751e | 505 | { |
aec24246 | 506 | struct throtl_grp *tg = pd_to_tg(pd); |
693e751e TH |
507 | /* |
508 | * We don't want new groups to escape the limits of its ancestors. | |
509 | * Update has_rules[] after a new group is brought online. | |
510 | */ | |
aec24246 SL |
511 | tg_update_has_rules(tg); |
512 | tg->last_dispatch_time[READ] = jiffies; | |
513 | tg->last_dispatch_time[WRITE] = jiffies; | |
693e751e TH |
514 | } |
515 | ||
cd5ab1b0 SL |
516 | static void blk_throtl_update_limit_valid(struct throtl_data *td) |
517 | { | |
518 | struct cgroup_subsys_state *pos_css; | |
519 | struct blkcg_gq *blkg; | |
520 | bool low_valid = false; | |
521 | ||
522 | rcu_read_lock(); | |
523 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
524 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
525 | ||
526 | if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] || | |
527 | tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) | |
528 | low_valid = true; | |
529 | } | |
530 | rcu_read_unlock(); | |
531 | ||
532 | td->limit_valid[LIMIT_LOW] = low_valid; | |
533 | } | |
534 | ||
c79892c5 | 535 | static void throtl_upgrade_state(struct throtl_data *td); |
cd5ab1b0 SL |
536 | static void throtl_pd_offline(struct blkg_policy_data *pd) |
537 | { | |
538 | struct throtl_grp *tg = pd_to_tg(pd); | |
539 | ||
540 | tg->bps[READ][LIMIT_LOW] = 0; | |
541 | tg->bps[WRITE][LIMIT_LOW] = 0; | |
542 | tg->iops[READ][LIMIT_LOW] = 0; | |
543 | tg->iops[WRITE][LIMIT_LOW] = 0; | |
544 | ||
545 | blk_throtl_update_limit_valid(tg->td); | |
546 | ||
c79892c5 SL |
547 | if (!tg->td->limit_valid[tg->td->limit_index]) |
548 | throtl_upgrade_state(tg->td); | |
cd5ab1b0 SL |
549 | } |
550 | ||
001bea73 TH |
551 | static void throtl_pd_free(struct blkg_policy_data *pd) |
552 | { | |
4fb72036 TH |
553 | struct throtl_grp *tg = pd_to_tg(pd); |
554 | ||
b2ce2643 | 555 | del_timer_sync(&tg->service_queue.pending_timer); |
4fb72036 | 556 | kfree(tg); |
001bea73 TH |
557 | } |
558 | ||
0049af73 TH |
559 | static struct throtl_grp * |
560 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
561 | { |
562 | /* Service tree is empty */ | |
0049af73 | 563 | if (!parent_sq->nr_pending) |
e43473b7 VG |
564 | return NULL; |
565 | ||
0049af73 TH |
566 | if (!parent_sq->first_pending) |
567 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
e43473b7 | 568 | |
0049af73 TH |
569 | if (parent_sq->first_pending) |
570 | return rb_entry_tg(parent_sq->first_pending); | |
e43473b7 VG |
571 | |
572 | return NULL; | |
573 | } | |
574 | ||
575 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
576 | { | |
577 | rb_erase(n, root); | |
578 | RB_CLEAR_NODE(n); | |
579 | } | |
580 | ||
0049af73 TH |
581 | static void throtl_rb_erase(struct rb_node *n, |
582 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 583 | { |
0049af73 TH |
584 | if (parent_sq->first_pending == n) |
585 | parent_sq->first_pending = NULL; | |
586 | rb_erase_init(n, &parent_sq->pending_tree); | |
587 | --parent_sq->nr_pending; | |
e43473b7 VG |
588 | } |
589 | ||
0049af73 | 590 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
591 | { |
592 | struct throtl_grp *tg; | |
593 | ||
0049af73 | 594 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
595 | if (!tg) |
596 | return; | |
597 | ||
0049af73 | 598 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
599 | } |
600 | ||
77216b04 | 601 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 602 | { |
77216b04 | 603 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
0049af73 | 604 | struct rb_node **node = &parent_sq->pending_tree.rb_node; |
e43473b7 VG |
605 | struct rb_node *parent = NULL; |
606 | struct throtl_grp *__tg; | |
607 | unsigned long key = tg->disptime; | |
608 | int left = 1; | |
609 | ||
610 | while (*node != NULL) { | |
611 | parent = *node; | |
612 | __tg = rb_entry_tg(parent); | |
613 | ||
614 | if (time_before(key, __tg->disptime)) | |
615 | node = &parent->rb_left; | |
616 | else { | |
617 | node = &parent->rb_right; | |
618 | left = 0; | |
619 | } | |
620 | } | |
621 | ||
622 | if (left) | |
0049af73 | 623 | parent_sq->first_pending = &tg->rb_node; |
e43473b7 VG |
624 | |
625 | rb_link_node(&tg->rb_node, parent, node); | |
0049af73 | 626 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); |
e43473b7 VG |
627 | } |
628 | ||
77216b04 | 629 | static void __throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 630 | { |
77216b04 | 631 | tg_service_queue_add(tg); |
5b2c16aa | 632 | tg->flags |= THROTL_TG_PENDING; |
77216b04 | 633 | tg->service_queue.parent_sq->nr_pending++; |
e43473b7 VG |
634 | } |
635 | ||
77216b04 | 636 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 637 | { |
5b2c16aa | 638 | if (!(tg->flags & THROTL_TG_PENDING)) |
77216b04 | 639 | __throtl_enqueue_tg(tg); |
e43473b7 VG |
640 | } |
641 | ||
77216b04 | 642 | static void __throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 643 | { |
77216b04 | 644 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
5b2c16aa | 645 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
646 | } |
647 | ||
77216b04 | 648 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 649 | { |
5b2c16aa | 650 | if (tg->flags & THROTL_TG_PENDING) |
77216b04 | 651 | __throtl_dequeue_tg(tg); |
e43473b7 VG |
652 | } |
653 | ||
a9131a27 | 654 | /* Call with queue lock held */ |
69df0ab0 TH |
655 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, |
656 | unsigned long expires) | |
a9131a27 | 657 | { |
297e3d85 | 658 | unsigned long max_expire = jiffies + 8 * sq_to_tg(sq)->td->throtl_slice; |
06cceedc SL |
659 | |
660 | /* | |
661 | * Since we are adjusting the throttle limit dynamically, the sleep | |
662 | * time calculated according to previous limit might be invalid. It's | |
663 | * possible the cgroup sleep time is very long and no other cgroups | |
664 | * have IO running so notify the limit changes. Make sure the cgroup | |
665 | * doesn't sleep too long to avoid the missed notification. | |
666 | */ | |
667 | if (time_after(expires, max_expire)) | |
668 | expires = max_expire; | |
69df0ab0 TH |
669 | mod_timer(&sq->pending_timer, expires); |
670 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
671 | expires - jiffies, jiffies); | |
a9131a27 TH |
672 | } |
673 | ||
7f52f98c TH |
674 | /** |
675 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle | |
676 | * @sq: the service_queue to schedule dispatch for | |
677 | * @force: force scheduling | |
678 | * | |
679 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the | |
680 | * dispatch time of the first pending child. Returns %true if either timer | |
681 | * is armed or there's no pending child left. %false if the current | |
682 | * dispatch window is still open and the caller should continue | |
683 | * dispatching. | |
684 | * | |
685 | * If @force is %true, the dispatch timer is always scheduled and this | |
686 | * function is guaranteed to return %true. This is to be used when the | |
687 | * caller can't dispatch itself and needs to invoke pending_timer | |
688 | * unconditionally. Note that forced scheduling is likely to induce short | |
689 | * delay before dispatch starts even if @sq->first_pending_disptime is not | |
690 | * in the future and thus shouldn't be used in hot paths. | |
691 | */ | |
692 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
693 | bool force) | |
e43473b7 | 694 | { |
6a525600 | 695 | /* any pending children left? */ |
c9e0332e | 696 | if (!sq->nr_pending) |
7f52f98c | 697 | return true; |
e43473b7 | 698 | |
c9e0332e | 699 | update_min_dispatch_time(sq); |
e43473b7 | 700 | |
69df0ab0 | 701 | /* is the next dispatch time in the future? */ |
7f52f98c | 702 | if (force || time_after(sq->first_pending_disptime, jiffies)) { |
69df0ab0 | 703 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); |
7f52f98c | 704 | return true; |
69df0ab0 TH |
705 | } |
706 | ||
7f52f98c TH |
707 | /* tell the caller to continue dispatching */ |
708 | return false; | |
e43473b7 VG |
709 | } |
710 | ||
32ee5bc4 VG |
711 | static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, |
712 | bool rw, unsigned long start) | |
713 | { | |
714 | tg->bytes_disp[rw] = 0; | |
715 | tg->io_disp[rw] = 0; | |
716 | ||
717 | /* | |
718 | * Previous slice has expired. We must have trimmed it after last | |
719 | * bio dispatch. That means since start of last slice, we never used | |
720 | * that bandwidth. Do try to make use of that bandwidth while giving | |
721 | * credit. | |
722 | */ | |
723 | if (time_after_eq(start, tg->slice_start[rw])) | |
724 | tg->slice_start[rw] = start; | |
725 | ||
297e3d85 | 726 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; |
32ee5bc4 VG |
727 | throtl_log(&tg->service_queue, |
728 | "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", | |
729 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
730 | tg->slice_end[rw], jiffies); | |
731 | } | |
732 | ||
0f3457f6 | 733 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
734 | { |
735 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 736 | tg->io_disp[rw] = 0; |
e43473b7 | 737 | tg->slice_start[rw] = jiffies; |
297e3d85 | 738 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; |
fda6f272 TH |
739 | throtl_log(&tg->service_queue, |
740 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
741 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
742 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
743 | } |
744 | ||
0f3457f6 TH |
745 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
746 | unsigned long jiffy_end) | |
d1ae8ffd | 747 | { |
297e3d85 | 748 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); |
d1ae8ffd VG |
749 | } |
750 | ||
0f3457f6 TH |
751 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
752 | unsigned long jiffy_end) | |
e43473b7 | 753 | { |
297e3d85 | 754 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); |
fda6f272 TH |
755 | throtl_log(&tg->service_queue, |
756 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
757 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
758 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
759 | } |
760 | ||
761 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 762 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
763 | { |
764 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
5cf8c227 | 765 | return false; |
e43473b7 VG |
766 | |
767 | return 1; | |
768 | } | |
769 | ||
770 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 771 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 772 | { |
3aad5d3e VG |
773 | unsigned long nr_slices, time_elapsed, io_trim; |
774 | u64 bytes_trim, tmp; | |
e43473b7 VG |
775 | |
776 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
777 | ||
778 | /* | |
779 | * If bps are unlimited (-1), then time slice don't get | |
780 | * renewed. Don't try to trim the slice if slice is used. A new | |
781 | * slice will start when appropriate. | |
782 | */ | |
0f3457f6 | 783 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
784 | return; |
785 | ||
d1ae8ffd VG |
786 | /* |
787 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
788 | * that initially cgroup limit was very low resulting in high | |
789 | * slice_end, but later limit was bumped up and bio was dispached | |
790 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
791 | * is bad because it does not allow new slice to start. | |
792 | */ | |
793 | ||
297e3d85 | 794 | throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice); |
d1ae8ffd | 795 | |
e43473b7 VG |
796 | time_elapsed = jiffies - tg->slice_start[rw]; |
797 | ||
297e3d85 | 798 | nr_slices = time_elapsed / tg->td->throtl_slice; |
e43473b7 VG |
799 | |
800 | if (!nr_slices) | |
801 | return; | |
297e3d85 | 802 | tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices; |
3aad5d3e VG |
803 | do_div(tmp, HZ); |
804 | bytes_trim = tmp; | |
e43473b7 | 805 | |
297e3d85 SL |
806 | io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) / |
807 | HZ; | |
e43473b7 | 808 | |
8e89d13f | 809 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
810 | return; |
811 | ||
812 | if (tg->bytes_disp[rw] >= bytes_trim) | |
813 | tg->bytes_disp[rw] -= bytes_trim; | |
814 | else | |
815 | tg->bytes_disp[rw] = 0; | |
816 | ||
8e89d13f VG |
817 | if (tg->io_disp[rw] >= io_trim) |
818 | tg->io_disp[rw] -= io_trim; | |
819 | else | |
820 | tg->io_disp[rw] = 0; | |
821 | ||
297e3d85 | 822 | tg->slice_start[rw] += nr_slices * tg->td->throtl_slice; |
e43473b7 | 823 | |
fda6f272 TH |
824 | throtl_log(&tg->service_queue, |
825 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
826 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
827 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
e43473b7 VG |
828 | } |
829 | ||
0f3457f6 TH |
830 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
831 | unsigned long *wait) | |
e43473b7 VG |
832 | { |
833 | bool rw = bio_data_dir(bio); | |
8e89d13f | 834 | unsigned int io_allowed; |
e43473b7 | 835 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 836 | u64 tmp; |
e43473b7 | 837 | |
8e89d13f | 838 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 839 | |
8e89d13f VG |
840 | /* Slice has just started. Consider one slice interval */ |
841 | if (!jiffy_elapsed) | |
297e3d85 | 842 | jiffy_elapsed_rnd = tg->td->throtl_slice; |
8e89d13f | 843 | |
297e3d85 | 844 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); |
8e89d13f | 845 | |
c49c06e4 VG |
846 | /* |
847 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
848 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
849 | * will allow dispatch after 1 second and after that slice should | |
850 | * have been trimmed. | |
851 | */ | |
852 | ||
9f626e37 | 853 | tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd; |
c49c06e4 VG |
854 | do_div(tmp, HZ); |
855 | ||
856 | if (tmp > UINT_MAX) | |
857 | io_allowed = UINT_MAX; | |
858 | else | |
859 | io_allowed = tmp; | |
8e89d13f VG |
860 | |
861 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
862 | if (wait) |
863 | *wait = 0; | |
5cf8c227 | 864 | return true; |
e43473b7 VG |
865 | } |
866 | ||
8e89d13f | 867 | /* Calc approx time to dispatch */ |
9f626e37 | 868 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ) / tg_iops_limit(tg, rw) + 1; |
8e89d13f VG |
869 | |
870 | if (jiffy_wait > jiffy_elapsed) | |
871 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
872 | else | |
873 | jiffy_wait = 1; | |
874 | ||
875 | if (wait) | |
876 | *wait = jiffy_wait; | |
877 | return 0; | |
878 | } | |
879 | ||
0f3457f6 TH |
880 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
881 | unsigned long *wait) | |
8e89d13f VG |
882 | { |
883 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 884 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 885 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
886 | |
887 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
888 | ||
889 | /* Slice has just started. Consider one slice interval */ | |
890 | if (!jiffy_elapsed) | |
297e3d85 | 891 | jiffy_elapsed_rnd = tg->td->throtl_slice; |
e43473b7 | 892 | |
297e3d85 | 893 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); |
e43473b7 | 894 | |
9f626e37 | 895 | tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd; |
5e901a2b | 896 | do_div(tmp, HZ); |
3aad5d3e | 897 | bytes_allowed = tmp; |
e43473b7 | 898 | |
4f024f37 | 899 | if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) { |
e43473b7 VG |
900 | if (wait) |
901 | *wait = 0; | |
5cf8c227 | 902 | return true; |
e43473b7 VG |
903 | } |
904 | ||
905 | /* Calc approx time to dispatch */ | |
4f024f37 | 906 | extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed; |
9f626e37 | 907 | jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw)); |
e43473b7 VG |
908 | |
909 | if (!jiffy_wait) | |
910 | jiffy_wait = 1; | |
911 | ||
912 | /* | |
913 | * This wait time is without taking into consideration the rounding | |
914 | * up we did. Add that time also. | |
915 | */ | |
916 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
917 | if (wait) |
918 | *wait = jiffy_wait; | |
8e89d13f VG |
919 | return 0; |
920 | } | |
921 | ||
922 | /* | |
923 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
924 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
925 | */ | |
0f3457f6 TH |
926 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
927 | unsigned long *wait) | |
8e89d13f VG |
928 | { |
929 | bool rw = bio_data_dir(bio); | |
930 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
931 | ||
932 | /* | |
933 | * Currently whole state machine of group depends on first bio | |
934 | * queued in the group bio list. So one should not be calling | |
935 | * this function with a different bio if there are other bios | |
936 | * queued. | |
937 | */ | |
73f0d49a | 938 | BUG_ON(tg->service_queue.nr_queued[rw] && |
c5cc2070 | 939 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); |
e43473b7 | 940 | |
8e89d13f | 941 | /* If tg->bps = -1, then BW is unlimited */ |
9f626e37 SL |
942 | if (tg_bps_limit(tg, rw) == U64_MAX && |
943 | tg_iops_limit(tg, rw) == UINT_MAX) { | |
8e89d13f VG |
944 | if (wait) |
945 | *wait = 0; | |
5cf8c227 | 946 | return true; |
8e89d13f VG |
947 | } |
948 | ||
949 | /* | |
950 | * If previous slice expired, start a new one otherwise renew/extend | |
951 | * existing slice to make sure it is at least throtl_slice interval | |
164c80ed VG |
952 | * long since now. New slice is started only for empty throttle group. |
953 | * If there is queued bio, that means there should be an active | |
954 | * slice and it should be extended instead. | |
8e89d13f | 955 | */ |
164c80ed | 956 | if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw])) |
0f3457f6 | 957 | throtl_start_new_slice(tg, rw); |
8e89d13f | 958 | else { |
297e3d85 SL |
959 | if (time_before(tg->slice_end[rw], |
960 | jiffies + tg->td->throtl_slice)) | |
961 | throtl_extend_slice(tg, rw, | |
962 | jiffies + tg->td->throtl_slice); | |
8e89d13f VG |
963 | } |
964 | ||
0f3457f6 TH |
965 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
966 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
967 | if (wait) |
968 | *wait = 0; | |
969 | return 1; | |
970 | } | |
971 | ||
972 | max_wait = max(bps_wait, iops_wait); | |
973 | ||
974 | if (wait) | |
975 | *wait = max_wait; | |
976 | ||
977 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 978 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
979 | |
980 | return 0; | |
981 | } | |
982 | ||
983 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
984 | { | |
985 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
986 | |
987 | /* Charge the bio to the group */ | |
4f024f37 | 988 | tg->bytes_disp[rw] += bio->bi_iter.bi_size; |
8e89d13f | 989 | tg->io_disp[rw]++; |
3f0abd80 SL |
990 | tg->last_bytes_disp[rw] += bio->bi_iter.bi_size; |
991 | tg->last_io_disp[rw]++; | |
e43473b7 | 992 | |
2a0f61e6 | 993 | /* |
8d2bbd4c | 994 | * BIO_THROTTLED is used to prevent the same bio to be throttled |
2a0f61e6 TH |
995 | * more than once as a throttled bio will go through blk-throtl the |
996 | * second time when it eventually gets issued. Set it when a bio | |
997 | * is being charged to a tg. | |
2a0f61e6 | 998 | */ |
8d2bbd4c CH |
999 | if (!bio_flagged(bio, BIO_THROTTLED)) |
1000 | bio_set_flag(bio, BIO_THROTTLED); | |
e43473b7 VG |
1001 | } |
1002 | ||
c5cc2070 TH |
1003 | /** |
1004 | * throtl_add_bio_tg - add a bio to the specified throtl_grp | |
1005 | * @bio: bio to add | |
1006 | * @qn: qnode to use | |
1007 | * @tg: the target throtl_grp | |
1008 | * | |
1009 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, | |
1010 | * tg->qnode_on_self[] is used. | |
1011 | */ | |
1012 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, | |
1013 | struct throtl_grp *tg) | |
e43473b7 | 1014 | { |
73f0d49a | 1015 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1016 | bool rw = bio_data_dir(bio); |
1017 | ||
c5cc2070 TH |
1018 | if (!qn) |
1019 | qn = &tg->qnode_on_self[rw]; | |
1020 | ||
0e9f4164 TH |
1021 | /* |
1022 | * If @tg doesn't currently have any bios queued in the same | |
1023 | * direction, queueing @bio can change when @tg should be | |
1024 | * dispatched. Mark that @tg was empty. This is automatically | |
1025 | * cleaered on the next tg_update_disptime(). | |
1026 | */ | |
1027 | if (!sq->nr_queued[rw]) | |
1028 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
1029 | ||
c5cc2070 TH |
1030 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); |
1031 | ||
73f0d49a | 1032 | sq->nr_queued[rw]++; |
77216b04 | 1033 | throtl_enqueue_tg(tg); |
e43473b7 VG |
1034 | } |
1035 | ||
77216b04 | 1036 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 1037 | { |
73f0d49a | 1038 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1039 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
1040 | struct bio *bio; | |
1041 | ||
d609af3a ME |
1042 | bio = throtl_peek_queued(&sq->queued[READ]); |
1043 | if (bio) | |
0f3457f6 | 1044 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 1045 | |
d609af3a ME |
1046 | bio = throtl_peek_queued(&sq->queued[WRITE]); |
1047 | if (bio) | |
0f3457f6 | 1048 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
1049 | |
1050 | min_wait = min(read_wait, write_wait); | |
1051 | disptime = jiffies + min_wait; | |
1052 | ||
e43473b7 | 1053 | /* Update dispatch time */ |
77216b04 | 1054 | throtl_dequeue_tg(tg); |
e43473b7 | 1055 | tg->disptime = disptime; |
77216b04 | 1056 | throtl_enqueue_tg(tg); |
0e9f4164 TH |
1057 | |
1058 | /* see throtl_add_bio_tg() */ | |
1059 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
1060 | } |
1061 | ||
32ee5bc4 VG |
1062 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, |
1063 | struct throtl_grp *parent_tg, bool rw) | |
1064 | { | |
1065 | if (throtl_slice_used(parent_tg, rw)) { | |
1066 | throtl_start_new_slice_with_credit(parent_tg, rw, | |
1067 | child_tg->slice_start[rw]); | |
1068 | } | |
1069 | ||
1070 | } | |
1071 | ||
77216b04 | 1072 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 1073 | { |
73f0d49a | 1074 | struct throtl_service_queue *sq = &tg->service_queue; |
6bc9c2b4 TH |
1075 | struct throtl_service_queue *parent_sq = sq->parent_sq; |
1076 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
c5cc2070 | 1077 | struct throtl_grp *tg_to_put = NULL; |
e43473b7 VG |
1078 | struct bio *bio; |
1079 | ||
c5cc2070 TH |
1080 | /* |
1081 | * @bio is being transferred from @tg to @parent_sq. Popping a bio | |
1082 | * from @tg may put its reference and @parent_sq might end up | |
1083 | * getting released prematurely. Remember the tg to put and put it | |
1084 | * after @bio is transferred to @parent_sq. | |
1085 | */ | |
1086 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); | |
73f0d49a | 1087 | sq->nr_queued[rw]--; |
e43473b7 VG |
1088 | |
1089 | throtl_charge_bio(tg, bio); | |
6bc9c2b4 TH |
1090 | |
1091 | /* | |
1092 | * If our parent is another tg, we just need to transfer @bio to | |
1093 | * the parent using throtl_add_bio_tg(). If our parent is | |
1094 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
1095 | * bio_lists[] and decrease total number queued. The caller is | |
1096 | * responsible for issuing these bios. | |
1097 | */ | |
1098 | if (parent_tg) { | |
c5cc2070 | 1099 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); |
32ee5bc4 | 1100 | start_parent_slice_with_credit(tg, parent_tg, rw); |
6bc9c2b4 | 1101 | } else { |
c5cc2070 TH |
1102 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], |
1103 | &parent_sq->queued[rw]); | |
6bc9c2b4 TH |
1104 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
1105 | tg->td->nr_queued[rw]--; | |
1106 | } | |
e43473b7 | 1107 | |
0f3457f6 | 1108 | throtl_trim_slice(tg, rw); |
6bc9c2b4 | 1109 | |
c5cc2070 TH |
1110 | if (tg_to_put) |
1111 | blkg_put(tg_to_blkg(tg_to_put)); | |
e43473b7 VG |
1112 | } |
1113 | ||
77216b04 | 1114 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 1115 | { |
73f0d49a | 1116 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1117 | unsigned int nr_reads = 0, nr_writes = 0; |
1118 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 1119 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
1120 | struct bio *bio; |
1121 | ||
1122 | /* Try to dispatch 75% READS and 25% WRITES */ | |
1123 | ||
c5cc2070 | 1124 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && |
0f3457f6 | 1125 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1126 | |
77216b04 | 1127 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1128 | nr_reads++; |
1129 | ||
1130 | if (nr_reads >= max_nr_reads) | |
1131 | break; | |
1132 | } | |
1133 | ||
c5cc2070 | 1134 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && |
0f3457f6 | 1135 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1136 | |
77216b04 | 1137 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1138 | nr_writes++; |
1139 | ||
1140 | if (nr_writes >= max_nr_writes) | |
1141 | break; | |
1142 | } | |
1143 | ||
1144 | return nr_reads + nr_writes; | |
1145 | } | |
1146 | ||
651930bc | 1147 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
1148 | { |
1149 | unsigned int nr_disp = 0; | |
e43473b7 VG |
1150 | |
1151 | while (1) { | |
73f0d49a TH |
1152 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
1153 | struct throtl_service_queue *sq = &tg->service_queue; | |
e43473b7 VG |
1154 | |
1155 | if (!tg) | |
1156 | break; | |
1157 | ||
1158 | if (time_before(jiffies, tg->disptime)) | |
1159 | break; | |
1160 | ||
77216b04 | 1161 | throtl_dequeue_tg(tg); |
e43473b7 | 1162 | |
77216b04 | 1163 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 1164 | |
73f0d49a | 1165 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
77216b04 | 1166 | tg_update_disptime(tg); |
e43473b7 VG |
1167 | |
1168 | if (nr_disp >= throtl_quantum) | |
1169 | break; | |
1170 | } | |
1171 | ||
1172 | return nr_disp; | |
1173 | } | |
1174 | ||
c79892c5 SL |
1175 | static bool throtl_can_upgrade(struct throtl_data *td, |
1176 | struct throtl_grp *this_tg); | |
6e1a5704 TH |
1177 | /** |
1178 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
1179 | * @arg: the throtl_service_queue being serviced | |
1180 | * | |
1181 | * This timer is armed when a child throtl_grp with active bio's become | |
1182 | * pending and queued on the service_queue's pending_tree and expires when | |
1183 | * the first child throtl_grp should be dispatched. This function | |
2e48a530 TH |
1184 | * dispatches bio's from the children throtl_grps to the parent |
1185 | * service_queue. | |
1186 | * | |
1187 | * If the parent's parent is another throtl_grp, dispatching is propagated | |
1188 | * by either arming its pending_timer or repeating dispatch directly. If | |
1189 | * the top-level service_tree is reached, throtl_data->dispatch_work is | |
1190 | * kicked so that the ready bio's are issued. | |
6e1a5704 | 1191 | */ |
69df0ab0 TH |
1192 | static void throtl_pending_timer_fn(unsigned long arg) |
1193 | { | |
1194 | struct throtl_service_queue *sq = (void *)arg; | |
2e48a530 | 1195 | struct throtl_grp *tg = sq_to_tg(sq); |
69df0ab0 | 1196 | struct throtl_data *td = sq_to_td(sq); |
cb76199c | 1197 | struct request_queue *q = td->queue; |
2e48a530 TH |
1198 | struct throtl_service_queue *parent_sq; |
1199 | bool dispatched; | |
6e1a5704 | 1200 | int ret; |
e43473b7 VG |
1201 | |
1202 | spin_lock_irq(q->queue_lock); | |
c79892c5 SL |
1203 | if (throtl_can_upgrade(td, NULL)) |
1204 | throtl_upgrade_state(td); | |
1205 | ||
2e48a530 TH |
1206 | again: |
1207 | parent_sq = sq->parent_sq; | |
1208 | dispatched = false; | |
e43473b7 | 1209 | |
7f52f98c TH |
1210 | while (true) { |
1211 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
2e48a530 TH |
1212 | sq->nr_queued[READ] + sq->nr_queued[WRITE], |
1213 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
7f52f98c TH |
1214 | |
1215 | ret = throtl_select_dispatch(sq); | |
1216 | if (ret) { | |
7f52f98c TH |
1217 | throtl_log(sq, "bios disp=%u", ret); |
1218 | dispatched = true; | |
1219 | } | |
e43473b7 | 1220 | |
7f52f98c TH |
1221 | if (throtl_schedule_next_dispatch(sq, false)) |
1222 | break; | |
e43473b7 | 1223 | |
7f52f98c TH |
1224 | /* this dispatch windows is still open, relax and repeat */ |
1225 | spin_unlock_irq(q->queue_lock); | |
1226 | cpu_relax(); | |
1227 | spin_lock_irq(q->queue_lock); | |
651930bc | 1228 | } |
e43473b7 | 1229 | |
2e48a530 TH |
1230 | if (!dispatched) |
1231 | goto out_unlock; | |
6e1a5704 | 1232 | |
2e48a530 TH |
1233 | if (parent_sq) { |
1234 | /* @parent_sq is another throl_grp, propagate dispatch */ | |
1235 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1236 | tg_update_disptime(tg); | |
1237 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { | |
1238 | /* window is already open, repeat dispatching */ | |
1239 | sq = parent_sq; | |
1240 | tg = sq_to_tg(sq); | |
1241 | goto again; | |
1242 | } | |
1243 | } | |
1244 | } else { | |
1245 | /* reached the top-level, queue issueing */ | |
1246 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1247 | } | |
1248 | out_unlock: | |
e43473b7 | 1249 | spin_unlock_irq(q->queue_lock); |
6e1a5704 | 1250 | } |
e43473b7 | 1251 | |
6e1a5704 TH |
1252 | /** |
1253 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
1254 | * @work: work item being executed | |
1255 | * | |
1256 | * This function is queued for execution when bio's reach the bio_lists[] | |
1257 | * of throtl_data->service_queue. Those bio's are ready and issued by this | |
1258 | * function. | |
1259 | */ | |
8876e140 | 1260 | static void blk_throtl_dispatch_work_fn(struct work_struct *work) |
6e1a5704 TH |
1261 | { |
1262 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1263 | dispatch_work); | |
1264 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1265 | struct request_queue *q = td->queue; | |
1266 | struct bio_list bio_list_on_stack; | |
1267 | struct bio *bio; | |
1268 | struct blk_plug plug; | |
1269 | int rw; | |
1270 | ||
1271 | bio_list_init(&bio_list_on_stack); | |
1272 | ||
1273 | spin_lock_irq(q->queue_lock); | |
c5cc2070 TH |
1274 | for (rw = READ; rw <= WRITE; rw++) |
1275 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) | |
1276 | bio_list_add(&bio_list_on_stack, bio); | |
6e1a5704 TH |
1277 | spin_unlock_irq(q->queue_lock); |
1278 | ||
1279 | if (!bio_list_empty(&bio_list_on_stack)) { | |
69d60eb9 | 1280 | blk_start_plug(&plug); |
e43473b7 VG |
1281 | while((bio = bio_list_pop(&bio_list_on_stack))) |
1282 | generic_make_request(bio); | |
69d60eb9 | 1283 | blk_finish_plug(&plug); |
e43473b7 | 1284 | } |
e43473b7 VG |
1285 | } |
1286 | ||
f95a04af TH |
1287 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
1288 | int off) | |
60c2bc2d | 1289 | { |
f95a04af TH |
1290 | struct throtl_grp *tg = pd_to_tg(pd); |
1291 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 1292 | |
2ab5492d | 1293 | if (v == U64_MAX) |
60c2bc2d | 1294 | return 0; |
f95a04af | 1295 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
1296 | } |
1297 | ||
f95a04af TH |
1298 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
1299 | int off) | |
e43473b7 | 1300 | { |
f95a04af TH |
1301 | struct throtl_grp *tg = pd_to_tg(pd); |
1302 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 1303 | |
2ab5492d | 1304 | if (v == UINT_MAX) |
af133ceb | 1305 | return 0; |
f95a04af | 1306 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
1307 | } |
1308 | ||
2da8ca82 | 1309 | static int tg_print_conf_u64(struct seq_file *sf, void *v) |
8e89d13f | 1310 | { |
2da8ca82 TH |
1311 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64, |
1312 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1313 | return 0; |
8e89d13f VG |
1314 | } |
1315 | ||
2da8ca82 | 1316 | static int tg_print_conf_uint(struct seq_file *sf, void *v) |
8e89d13f | 1317 | { |
2da8ca82 TH |
1318 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint, |
1319 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1320 | return 0; |
60c2bc2d TH |
1321 | } |
1322 | ||
69948b07 | 1323 | static void tg_conf_updated(struct throtl_grp *tg) |
60c2bc2d | 1324 | { |
69948b07 | 1325 | struct throtl_service_queue *sq = &tg->service_queue; |
492eb21b | 1326 | struct cgroup_subsys_state *pos_css; |
69948b07 | 1327 | struct blkcg_gq *blkg; |
af133ceb | 1328 | |
fda6f272 TH |
1329 | throtl_log(&tg->service_queue, |
1330 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
9f626e37 SL |
1331 | tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE), |
1332 | tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE)); | |
632b4493 | 1333 | |
693e751e TH |
1334 | /* |
1335 | * Update has_rules[] flags for the updated tg's subtree. A tg is | |
1336 | * considered to have rules if either the tg itself or any of its | |
1337 | * ancestors has rules. This identifies groups without any | |
1338 | * restrictions in the whole hierarchy and allows them to bypass | |
1339 | * blk-throttle. | |
1340 | */ | |
69948b07 | 1341 | blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg)) |
693e751e TH |
1342 | tg_update_has_rules(blkg_to_tg(blkg)); |
1343 | ||
632b4493 TH |
1344 | /* |
1345 | * We're already holding queue_lock and know @tg is valid. Let's | |
1346 | * apply the new config directly. | |
1347 | * | |
1348 | * Restart the slices for both READ and WRITES. It might happen | |
1349 | * that a group's limit are dropped suddenly and we don't want to | |
1350 | * account recently dispatched IO with new low rate. | |
1351 | */ | |
0f3457f6 TH |
1352 | throtl_start_new_slice(tg, 0); |
1353 | throtl_start_new_slice(tg, 1); | |
632b4493 | 1354 | |
5b2c16aa | 1355 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 1356 | tg_update_disptime(tg); |
7f52f98c | 1357 | throtl_schedule_next_dispatch(sq->parent_sq, true); |
632b4493 | 1358 | } |
69948b07 TH |
1359 | } |
1360 | ||
1361 | static ssize_t tg_set_conf(struct kernfs_open_file *of, | |
1362 | char *buf, size_t nbytes, loff_t off, bool is_u64) | |
1363 | { | |
1364 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1365 | struct blkg_conf_ctx ctx; | |
1366 | struct throtl_grp *tg; | |
1367 | int ret; | |
1368 | u64 v; | |
1369 | ||
1370 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1371 | if (ret) | |
1372 | return ret; | |
1373 | ||
1374 | ret = -EINVAL; | |
1375 | if (sscanf(ctx.body, "%llu", &v) != 1) | |
1376 | goto out_finish; | |
1377 | if (!v) | |
2ab5492d | 1378 | v = U64_MAX; |
69948b07 TH |
1379 | |
1380 | tg = blkg_to_tg(ctx.blkg); | |
1381 | ||
1382 | if (is_u64) | |
1383 | *(u64 *)((void *)tg + of_cft(of)->private) = v; | |
1384 | else | |
1385 | *(unsigned int *)((void *)tg + of_cft(of)->private) = v; | |
60c2bc2d | 1386 | |
69948b07 | 1387 | tg_conf_updated(tg); |
36aa9e5f TH |
1388 | ret = 0; |
1389 | out_finish: | |
60c2bc2d | 1390 | blkg_conf_finish(&ctx); |
36aa9e5f | 1391 | return ret ?: nbytes; |
8e89d13f VG |
1392 | } |
1393 | ||
451af504 TH |
1394 | static ssize_t tg_set_conf_u64(struct kernfs_open_file *of, |
1395 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1396 | { |
451af504 | 1397 | return tg_set_conf(of, buf, nbytes, off, true); |
60c2bc2d TH |
1398 | } |
1399 | ||
451af504 TH |
1400 | static ssize_t tg_set_conf_uint(struct kernfs_open_file *of, |
1401 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1402 | { |
451af504 | 1403 | return tg_set_conf(of, buf, nbytes, off, false); |
60c2bc2d TH |
1404 | } |
1405 | ||
880f50e2 | 1406 | static struct cftype throtl_legacy_files[] = { |
60c2bc2d TH |
1407 | { |
1408 | .name = "throttle.read_bps_device", | |
9f626e37 | 1409 | .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]), |
2da8ca82 | 1410 | .seq_show = tg_print_conf_u64, |
451af504 | 1411 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1412 | }, |
1413 | { | |
1414 | .name = "throttle.write_bps_device", | |
9f626e37 | 1415 | .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]), |
2da8ca82 | 1416 | .seq_show = tg_print_conf_u64, |
451af504 | 1417 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1418 | }, |
1419 | { | |
1420 | .name = "throttle.read_iops_device", | |
9f626e37 | 1421 | .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]), |
2da8ca82 | 1422 | .seq_show = tg_print_conf_uint, |
451af504 | 1423 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1424 | }, |
1425 | { | |
1426 | .name = "throttle.write_iops_device", | |
9f626e37 | 1427 | .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]), |
2da8ca82 | 1428 | .seq_show = tg_print_conf_uint, |
451af504 | 1429 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1430 | }, |
1431 | { | |
1432 | .name = "throttle.io_service_bytes", | |
77ea7338 TH |
1433 | .private = (unsigned long)&blkcg_policy_throtl, |
1434 | .seq_show = blkg_print_stat_bytes, | |
60c2bc2d TH |
1435 | }, |
1436 | { | |
1437 | .name = "throttle.io_serviced", | |
77ea7338 TH |
1438 | .private = (unsigned long)&blkcg_policy_throtl, |
1439 | .seq_show = blkg_print_stat_ios, | |
60c2bc2d TH |
1440 | }, |
1441 | { } /* terminate */ | |
1442 | }; | |
1443 | ||
cd5ab1b0 | 1444 | static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd, |
2ee867dc TH |
1445 | int off) |
1446 | { | |
1447 | struct throtl_grp *tg = pd_to_tg(pd); | |
1448 | const char *dname = blkg_dev_name(pd->blkg); | |
1449 | char bufs[4][21] = { "max", "max", "max", "max" }; | |
cd5ab1b0 SL |
1450 | u64 bps_dft; |
1451 | unsigned int iops_dft; | |
2ee867dc TH |
1452 | |
1453 | if (!dname) | |
1454 | return 0; | |
9f626e37 | 1455 | |
cd5ab1b0 SL |
1456 | if (off == LIMIT_LOW) { |
1457 | bps_dft = 0; | |
1458 | iops_dft = 0; | |
1459 | } else { | |
1460 | bps_dft = U64_MAX; | |
1461 | iops_dft = UINT_MAX; | |
1462 | } | |
1463 | ||
1464 | if (tg->bps_conf[READ][off] == bps_dft && | |
1465 | tg->bps_conf[WRITE][off] == bps_dft && | |
1466 | tg->iops_conf[READ][off] == iops_dft && | |
1467 | tg->iops_conf[WRITE][off] == iops_dft) | |
2ee867dc TH |
1468 | return 0; |
1469 | ||
cd5ab1b0 | 1470 | if (tg->bps_conf[READ][off] != bps_dft) |
9f626e37 | 1471 | snprintf(bufs[0], sizeof(bufs[0]), "%llu", |
cd5ab1b0 SL |
1472 | tg->bps_conf[READ][off]); |
1473 | if (tg->bps_conf[WRITE][off] != bps_dft) | |
9f626e37 | 1474 | snprintf(bufs[1], sizeof(bufs[1]), "%llu", |
cd5ab1b0 SL |
1475 | tg->bps_conf[WRITE][off]); |
1476 | if (tg->iops_conf[READ][off] != iops_dft) | |
9f626e37 | 1477 | snprintf(bufs[2], sizeof(bufs[2]), "%u", |
cd5ab1b0 SL |
1478 | tg->iops_conf[READ][off]); |
1479 | if (tg->iops_conf[WRITE][off] != iops_dft) | |
9f626e37 | 1480 | snprintf(bufs[3], sizeof(bufs[3]), "%u", |
cd5ab1b0 | 1481 | tg->iops_conf[WRITE][off]); |
2ee867dc TH |
1482 | |
1483 | seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s\n", | |
1484 | dname, bufs[0], bufs[1], bufs[2], bufs[3]); | |
1485 | return 0; | |
1486 | } | |
1487 | ||
cd5ab1b0 | 1488 | static int tg_print_limit(struct seq_file *sf, void *v) |
2ee867dc | 1489 | { |
cd5ab1b0 | 1490 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit, |
2ee867dc TH |
1491 | &blkcg_policy_throtl, seq_cft(sf)->private, false); |
1492 | return 0; | |
1493 | } | |
1494 | ||
cd5ab1b0 | 1495 | static ssize_t tg_set_limit(struct kernfs_open_file *of, |
2ee867dc TH |
1496 | char *buf, size_t nbytes, loff_t off) |
1497 | { | |
1498 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1499 | struct blkg_conf_ctx ctx; | |
1500 | struct throtl_grp *tg; | |
1501 | u64 v[4]; | |
1502 | int ret; | |
cd5ab1b0 | 1503 | int index = of_cft(of)->private; |
2ee867dc TH |
1504 | |
1505 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1506 | if (ret) | |
1507 | return ret; | |
1508 | ||
1509 | tg = blkg_to_tg(ctx.blkg); | |
1510 | ||
cd5ab1b0 SL |
1511 | v[0] = tg->bps_conf[READ][index]; |
1512 | v[1] = tg->bps_conf[WRITE][index]; | |
1513 | v[2] = tg->iops_conf[READ][index]; | |
1514 | v[3] = tg->iops_conf[WRITE][index]; | |
2ee867dc TH |
1515 | |
1516 | while (true) { | |
1517 | char tok[27]; /* wiops=18446744073709551616 */ | |
1518 | char *p; | |
2ab5492d | 1519 | u64 val = U64_MAX; |
2ee867dc TH |
1520 | int len; |
1521 | ||
1522 | if (sscanf(ctx.body, "%26s%n", tok, &len) != 1) | |
1523 | break; | |
1524 | if (tok[0] == '\0') | |
1525 | break; | |
1526 | ctx.body += len; | |
1527 | ||
1528 | ret = -EINVAL; | |
1529 | p = tok; | |
1530 | strsep(&p, "="); | |
1531 | if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max"))) | |
1532 | goto out_finish; | |
1533 | ||
1534 | ret = -ERANGE; | |
1535 | if (!val) | |
1536 | goto out_finish; | |
1537 | ||
1538 | ret = -EINVAL; | |
1539 | if (!strcmp(tok, "rbps")) | |
1540 | v[0] = val; | |
1541 | else if (!strcmp(tok, "wbps")) | |
1542 | v[1] = val; | |
1543 | else if (!strcmp(tok, "riops")) | |
1544 | v[2] = min_t(u64, val, UINT_MAX); | |
1545 | else if (!strcmp(tok, "wiops")) | |
1546 | v[3] = min_t(u64, val, UINT_MAX); | |
1547 | else | |
1548 | goto out_finish; | |
1549 | } | |
1550 | ||
cd5ab1b0 SL |
1551 | tg->bps_conf[READ][index] = v[0]; |
1552 | tg->bps_conf[WRITE][index] = v[1]; | |
1553 | tg->iops_conf[READ][index] = v[2]; | |
1554 | tg->iops_conf[WRITE][index] = v[3]; | |
2ee867dc | 1555 | |
cd5ab1b0 SL |
1556 | if (index == LIMIT_MAX) { |
1557 | tg->bps[READ][index] = v[0]; | |
1558 | tg->bps[WRITE][index] = v[1]; | |
1559 | tg->iops[READ][index] = v[2]; | |
1560 | tg->iops[WRITE][index] = v[3]; | |
1561 | } | |
1562 | tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW], | |
1563 | tg->bps_conf[READ][LIMIT_MAX]); | |
1564 | tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW], | |
1565 | tg->bps_conf[WRITE][LIMIT_MAX]); | |
1566 | tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW], | |
1567 | tg->iops_conf[READ][LIMIT_MAX]); | |
1568 | tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW], | |
1569 | tg->iops_conf[WRITE][LIMIT_MAX]); | |
1570 | ||
1571 | if (index == LIMIT_LOW) { | |
1572 | blk_throtl_update_limit_valid(tg->td); | |
1573 | if (tg->td->limit_valid[LIMIT_LOW]) | |
1574 | tg->td->limit_index = LIMIT_LOW; | |
1575 | } | |
2ee867dc TH |
1576 | tg_conf_updated(tg); |
1577 | ret = 0; | |
1578 | out_finish: | |
1579 | blkg_conf_finish(&ctx); | |
1580 | return ret ?: nbytes; | |
1581 | } | |
1582 | ||
1583 | static struct cftype throtl_files[] = { | |
cd5ab1b0 SL |
1584 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
1585 | { | |
1586 | .name = "low", | |
1587 | .flags = CFTYPE_NOT_ON_ROOT, | |
1588 | .seq_show = tg_print_limit, | |
1589 | .write = tg_set_limit, | |
1590 | .private = LIMIT_LOW, | |
1591 | }, | |
1592 | #endif | |
2ee867dc TH |
1593 | { |
1594 | .name = "max", | |
1595 | .flags = CFTYPE_NOT_ON_ROOT, | |
cd5ab1b0 SL |
1596 | .seq_show = tg_print_limit, |
1597 | .write = tg_set_limit, | |
1598 | .private = LIMIT_MAX, | |
2ee867dc TH |
1599 | }, |
1600 | { } /* terminate */ | |
1601 | }; | |
1602 | ||
da527770 | 1603 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1604 | { |
1605 | struct throtl_data *td = q->td; | |
1606 | ||
69df0ab0 | 1607 | cancel_work_sync(&td->dispatch_work); |
e43473b7 VG |
1608 | } |
1609 | ||
3c798398 | 1610 | static struct blkcg_policy blkcg_policy_throtl = { |
2ee867dc | 1611 | .dfl_cftypes = throtl_files, |
880f50e2 | 1612 | .legacy_cftypes = throtl_legacy_files, |
f9fcc2d3 | 1613 | |
001bea73 | 1614 | .pd_alloc_fn = throtl_pd_alloc, |
f9fcc2d3 | 1615 | .pd_init_fn = throtl_pd_init, |
693e751e | 1616 | .pd_online_fn = throtl_pd_online, |
cd5ab1b0 | 1617 | .pd_offline_fn = throtl_pd_offline, |
001bea73 | 1618 | .pd_free_fn = throtl_pd_free, |
e43473b7 VG |
1619 | }; |
1620 | ||
3f0abd80 SL |
1621 | static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg) |
1622 | { | |
1623 | unsigned long rtime = jiffies, wtime = jiffies; | |
1624 | ||
1625 | if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]) | |
1626 | rtime = tg->last_low_overflow_time[READ]; | |
1627 | if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) | |
1628 | wtime = tg->last_low_overflow_time[WRITE]; | |
1629 | return min(rtime, wtime); | |
1630 | } | |
1631 | ||
1632 | /* tg should not be an intermediate node */ | |
1633 | static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg) | |
1634 | { | |
1635 | struct throtl_service_queue *parent_sq; | |
1636 | struct throtl_grp *parent = tg; | |
1637 | unsigned long ret = __tg_last_low_overflow_time(tg); | |
1638 | ||
1639 | while (true) { | |
1640 | parent_sq = parent->service_queue.parent_sq; | |
1641 | parent = sq_to_tg(parent_sq); | |
1642 | if (!parent) | |
1643 | break; | |
1644 | ||
1645 | /* | |
1646 | * The parent doesn't have low limit, it always reaches low | |
1647 | * limit. Its overflow time is useless for children | |
1648 | */ | |
1649 | if (!parent->bps[READ][LIMIT_LOW] && | |
1650 | !parent->iops[READ][LIMIT_LOW] && | |
1651 | !parent->bps[WRITE][LIMIT_LOW] && | |
1652 | !parent->iops[WRITE][LIMIT_LOW]) | |
1653 | continue; | |
1654 | if (time_after(__tg_last_low_overflow_time(parent), ret)) | |
1655 | ret = __tg_last_low_overflow_time(parent); | |
1656 | } | |
1657 | return ret; | |
1658 | } | |
1659 | ||
9e234eea SL |
1660 | static bool throtl_tg_is_idle(struct throtl_grp *tg) |
1661 | { | |
1662 | /* | |
1663 | * cgroup is idle if: | |
1664 | * - single idle is too long, longer than a fixed value (in case user | |
1665 | * configure a too big threshold) or 4 times of slice | |
1666 | * - average think time is more than threshold | |
1667 | */ | |
1668 | unsigned long time = jiffies_to_usecs(4 * tg->td->throtl_slice); | |
1669 | ||
1670 | time = min_t(unsigned long, MAX_IDLE_TIME, time); | |
1671 | return (ktime_get_ns() >> 10) - tg->last_finish_time > time || | |
1672 | tg->avg_idletime > tg->idletime_threshold; | |
1673 | } | |
1674 | ||
c79892c5 SL |
1675 | static bool throtl_tg_can_upgrade(struct throtl_grp *tg) |
1676 | { | |
1677 | struct throtl_service_queue *sq = &tg->service_queue; | |
1678 | bool read_limit, write_limit; | |
1679 | ||
1680 | /* | |
1681 | * if cgroup reaches low limit (if low limit is 0, the cgroup always | |
1682 | * reaches), it's ok to upgrade to next limit | |
1683 | */ | |
1684 | read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]; | |
1685 | write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]; | |
1686 | if (!read_limit && !write_limit) | |
1687 | return true; | |
1688 | if (read_limit && sq->nr_queued[READ] && | |
1689 | (!write_limit || sq->nr_queued[WRITE])) | |
1690 | return true; | |
1691 | if (write_limit && sq->nr_queued[WRITE] && | |
1692 | (!read_limit || sq->nr_queued[READ])) | |
1693 | return true; | |
aec24246 SL |
1694 | |
1695 | if (time_after_eq(jiffies, | |
1696 | tg->last_dispatch_time[READ] + tg->td->throtl_slice) && | |
1697 | time_after_eq(jiffies, | |
1698 | tg->last_dispatch_time[WRITE] + tg->td->throtl_slice)) | |
1699 | return true; | |
c79892c5 SL |
1700 | return false; |
1701 | } | |
1702 | ||
1703 | static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg) | |
1704 | { | |
1705 | while (true) { | |
1706 | if (throtl_tg_can_upgrade(tg)) | |
1707 | return true; | |
1708 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1709 | if (!tg || !tg_to_blkg(tg)->parent) | |
1710 | return false; | |
1711 | } | |
1712 | return false; | |
1713 | } | |
1714 | ||
1715 | static bool throtl_can_upgrade(struct throtl_data *td, | |
1716 | struct throtl_grp *this_tg) | |
1717 | { | |
1718 | struct cgroup_subsys_state *pos_css; | |
1719 | struct blkcg_gq *blkg; | |
1720 | ||
1721 | if (td->limit_index != LIMIT_LOW) | |
1722 | return false; | |
1723 | ||
297e3d85 | 1724 | if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice)) |
3f0abd80 SL |
1725 | return false; |
1726 | ||
c79892c5 SL |
1727 | rcu_read_lock(); |
1728 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1729 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1730 | ||
1731 | if (tg == this_tg) | |
1732 | continue; | |
1733 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1734 | continue; | |
1735 | if (!throtl_hierarchy_can_upgrade(tg)) { | |
1736 | rcu_read_unlock(); | |
1737 | return false; | |
1738 | } | |
1739 | } | |
1740 | rcu_read_unlock(); | |
1741 | return true; | |
1742 | } | |
1743 | ||
1744 | static void throtl_upgrade_state(struct throtl_data *td) | |
1745 | { | |
1746 | struct cgroup_subsys_state *pos_css; | |
1747 | struct blkcg_gq *blkg; | |
1748 | ||
1749 | td->limit_index = LIMIT_MAX; | |
3f0abd80 | 1750 | td->low_upgrade_time = jiffies; |
7394e31f | 1751 | td->scale = 0; |
c79892c5 SL |
1752 | rcu_read_lock(); |
1753 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1754 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1755 | struct throtl_service_queue *sq = &tg->service_queue; | |
1756 | ||
1757 | tg->disptime = jiffies - 1; | |
1758 | throtl_select_dispatch(sq); | |
1759 | throtl_schedule_next_dispatch(sq, false); | |
1760 | } | |
1761 | rcu_read_unlock(); | |
1762 | throtl_select_dispatch(&td->service_queue); | |
1763 | throtl_schedule_next_dispatch(&td->service_queue, false); | |
1764 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1765 | } | |
1766 | ||
3f0abd80 SL |
1767 | static void throtl_downgrade_state(struct throtl_data *td, int new) |
1768 | { | |
7394e31f SL |
1769 | td->scale /= 2; |
1770 | ||
1771 | if (td->scale) { | |
1772 | td->low_upgrade_time = jiffies - td->scale * td->throtl_slice; | |
1773 | return; | |
1774 | } | |
1775 | ||
3f0abd80 SL |
1776 | td->limit_index = new; |
1777 | td->low_downgrade_time = jiffies; | |
1778 | } | |
1779 | ||
1780 | static bool throtl_tg_can_downgrade(struct throtl_grp *tg) | |
1781 | { | |
1782 | struct throtl_data *td = tg->td; | |
1783 | unsigned long now = jiffies; | |
1784 | ||
aec24246 SL |
1785 | if (time_after_eq(now, tg->last_dispatch_time[READ] + |
1786 | td->throtl_slice) && | |
1787 | time_after_eq(now, tg->last_dispatch_time[WRITE] + | |
1788 | td->throtl_slice)) | |
1789 | return false; | |
3f0abd80 SL |
1790 | /* |
1791 | * If cgroup is below low limit, consider downgrade and throttle other | |
1792 | * cgroups | |
1793 | */ | |
297e3d85 SL |
1794 | if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) && |
1795 | time_after_eq(now, tg_last_low_overflow_time(tg) + | |
1796 | td->throtl_slice)) | |
3f0abd80 SL |
1797 | return true; |
1798 | return false; | |
1799 | } | |
1800 | ||
1801 | static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg) | |
1802 | { | |
1803 | while (true) { | |
1804 | if (!throtl_tg_can_downgrade(tg)) | |
1805 | return false; | |
1806 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1807 | if (!tg || !tg_to_blkg(tg)->parent) | |
1808 | break; | |
1809 | } | |
1810 | return true; | |
1811 | } | |
1812 | ||
1813 | static void throtl_downgrade_check(struct throtl_grp *tg) | |
1814 | { | |
1815 | uint64_t bps; | |
1816 | unsigned int iops; | |
1817 | unsigned long elapsed_time; | |
1818 | unsigned long now = jiffies; | |
1819 | ||
1820 | if (tg->td->limit_index != LIMIT_MAX || | |
1821 | !tg->td->limit_valid[LIMIT_LOW]) | |
1822 | return; | |
1823 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1824 | return; | |
297e3d85 | 1825 | if (time_after(tg->last_check_time + tg->td->throtl_slice, now)) |
3f0abd80 SL |
1826 | return; |
1827 | ||
1828 | elapsed_time = now - tg->last_check_time; | |
1829 | tg->last_check_time = now; | |
1830 | ||
297e3d85 SL |
1831 | if (time_before(now, tg_last_low_overflow_time(tg) + |
1832 | tg->td->throtl_slice)) | |
3f0abd80 SL |
1833 | return; |
1834 | ||
1835 | if (tg->bps[READ][LIMIT_LOW]) { | |
1836 | bps = tg->last_bytes_disp[READ] * HZ; | |
1837 | do_div(bps, elapsed_time); | |
1838 | if (bps >= tg->bps[READ][LIMIT_LOW]) | |
1839 | tg->last_low_overflow_time[READ] = now; | |
1840 | } | |
1841 | ||
1842 | if (tg->bps[WRITE][LIMIT_LOW]) { | |
1843 | bps = tg->last_bytes_disp[WRITE] * HZ; | |
1844 | do_div(bps, elapsed_time); | |
1845 | if (bps >= tg->bps[WRITE][LIMIT_LOW]) | |
1846 | tg->last_low_overflow_time[WRITE] = now; | |
1847 | } | |
1848 | ||
1849 | if (tg->iops[READ][LIMIT_LOW]) { | |
1850 | iops = tg->last_io_disp[READ] * HZ / elapsed_time; | |
1851 | if (iops >= tg->iops[READ][LIMIT_LOW]) | |
1852 | tg->last_low_overflow_time[READ] = now; | |
1853 | } | |
1854 | ||
1855 | if (tg->iops[WRITE][LIMIT_LOW]) { | |
1856 | iops = tg->last_io_disp[WRITE] * HZ / elapsed_time; | |
1857 | if (iops >= tg->iops[WRITE][LIMIT_LOW]) | |
1858 | tg->last_low_overflow_time[WRITE] = now; | |
1859 | } | |
1860 | ||
1861 | /* | |
1862 | * If cgroup is below low limit, consider downgrade and throttle other | |
1863 | * cgroups | |
1864 | */ | |
1865 | if (throtl_hierarchy_can_downgrade(tg)) | |
1866 | throtl_downgrade_state(tg->td, LIMIT_LOW); | |
1867 | ||
1868 | tg->last_bytes_disp[READ] = 0; | |
1869 | tg->last_bytes_disp[WRITE] = 0; | |
1870 | tg->last_io_disp[READ] = 0; | |
1871 | tg->last_io_disp[WRITE] = 0; | |
1872 | } | |
1873 | ||
9e234eea SL |
1874 | static void blk_throtl_update_idletime(struct throtl_grp *tg) |
1875 | { | |
1876 | unsigned long now = ktime_get_ns() >> 10; | |
1877 | unsigned long last_finish_time = tg->last_finish_time; | |
1878 | ||
1879 | if (now <= last_finish_time || last_finish_time == 0 || | |
1880 | last_finish_time == tg->checked_last_finish_time) | |
1881 | return; | |
1882 | ||
1883 | tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3; | |
1884 | tg->checked_last_finish_time = last_finish_time; | |
1885 | } | |
1886 | ||
ae118896 TH |
1887 | bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg, |
1888 | struct bio *bio) | |
e43473b7 | 1889 | { |
c5cc2070 | 1890 | struct throtl_qnode *qn = NULL; |
ae118896 | 1891 | struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg); |
73f0d49a | 1892 | struct throtl_service_queue *sq; |
0e9f4164 | 1893 | bool rw = bio_data_dir(bio); |
bc16a4f9 | 1894 | bool throttled = false; |
9e234eea | 1895 | int ret; |
e43473b7 | 1896 | |
ae118896 TH |
1897 | WARN_ON_ONCE(!rcu_read_lock_held()); |
1898 | ||
2a0f61e6 | 1899 | /* see throtl_charge_bio() */ |
8d2bbd4c | 1900 | if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw]) |
bc16a4f9 | 1901 | goto out; |
e43473b7 VG |
1902 | |
1903 | spin_lock_irq(q->queue_lock); | |
c9589f03 TH |
1904 | |
1905 | if (unlikely(blk_queue_bypass(q))) | |
bc16a4f9 | 1906 | goto out_unlock; |
f469a7b4 | 1907 | |
9e234eea SL |
1908 | ret = bio_associate_current(bio); |
1909 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
1910 | if (ret == 0 || ret == -EBUSY) | |
1911 | bio->bi_cg_private = tg; | |
1912 | #endif | |
1913 | blk_throtl_update_idletime(tg); | |
1914 | ||
73f0d49a TH |
1915 | sq = &tg->service_queue; |
1916 | ||
c79892c5 | 1917 | again: |
9e660acf | 1918 | while (true) { |
aec24246 | 1919 | tg->last_dispatch_time[rw] = jiffies; |
3f0abd80 SL |
1920 | if (tg->last_low_overflow_time[rw] == 0) |
1921 | tg->last_low_overflow_time[rw] = jiffies; | |
1922 | throtl_downgrade_check(tg); | |
9e660acf TH |
1923 | /* throtl is FIFO - if bios are already queued, should queue */ |
1924 | if (sq->nr_queued[rw]) | |
1925 | break; | |
de701c74 | 1926 | |
9e660acf | 1927 | /* if above limits, break to queue */ |
c79892c5 | 1928 | if (!tg_may_dispatch(tg, bio, NULL)) { |
3f0abd80 | 1929 | tg->last_low_overflow_time[rw] = jiffies; |
c79892c5 SL |
1930 | if (throtl_can_upgrade(tg->td, tg)) { |
1931 | throtl_upgrade_state(tg->td); | |
1932 | goto again; | |
1933 | } | |
9e660acf | 1934 | break; |
c79892c5 | 1935 | } |
9e660acf TH |
1936 | |
1937 | /* within limits, let's charge and dispatch directly */ | |
e43473b7 | 1938 | throtl_charge_bio(tg, bio); |
04521db0 VG |
1939 | |
1940 | /* | |
1941 | * We need to trim slice even when bios are not being queued | |
1942 | * otherwise it might happen that a bio is not queued for | |
1943 | * a long time and slice keeps on extending and trim is not | |
1944 | * called for a long time. Now if limits are reduced suddenly | |
1945 | * we take into account all the IO dispatched so far at new | |
1946 | * low rate and * newly queued IO gets a really long dispatch | |
1947 | * time. | |
1948 | * | |
1949 | * So keep on trimming slice even if bio is not queued. | |
1950 | */ | |
0f3457f6 | 1951 | throtl_trim_slice(tg, rw); |
9e660acf TH |
1952 | |
1953 | /* | |
1954 | * @bio passed through this layer without being throttled. | |
1955 | * Climb up the ladder. If we''re already at the top, it | |
1956 | * can be executed directly. | |
1957 | */ | |
c5cc2070 | 1958 | qn = &tg->qnode_on_parent[rw]; |
9e660acf TH |
1959 | sq = sq->parent_sq; |
1960 | tg = sq_to_tg(sq); | |
1961 | if (!tg) | |
1962 | goto out_unlock; | |
e43473b7 VG |
1963 | } |
1964 | ||
9e660acf | 1965 | /* out-of-limit, queue to @tg */ |
fda6f272 TH |
1966 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
1967 | rw == READ ? 'R' : 'W', | |
9f626e37 SL |
1968 | tg->bytes_disp[rw], bio->bi_iter.bi_size, |
1969 | tg_bps_limit(tg, rw), | |
1970 | tg->io_disp[rw], tg_iops_limit(tg, rw), | |
fda6f272 | 1971 | sq->nr_queued[READ], sq->nr_queued[WRITE]); |
e43473b7 | 1972 | |
3f0abd80 SL |
1973 | tg->last_low_overflow_time[rw] = jiffies; |
1974 | ||
6bc9c2b4 | 1975 | tg->td->nr_queued[rw]++; |
c5cc2070 | 1976 | throtl_add_bio_tg(bio, qn, tg); |
bc16a4f9 | 1977 | throttled = true; |
e43473b7 | 1978 | |
7f52f98c TH |
1979 | /* |
1980 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
1981 | * was empty before @bio. The forced scheduling isn't likely to | |
1982 | * cause undue delay as @bio is likely to be dispatched directly if | |
1983 | * its @tg's disptime is not in the future. | |
1984 | */ | |
0e9f4164 | 1985 | if (tg->flags & THROTL_TG_WAS_EMPTY) { |
77216b04 | 1986 | tg_update_disptime(tg); |
7f52f98c | 1987 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); |
e43473b7 VG |
1988 | } |
1989 | ||
bc16a4f9 | 1990 | out_unlock: |
e43473b7 | 1991 | spin_unlock_irq(q->queue_lock); |
bc16a4f9 | 1992 | out: |
2a0f61e6 TH |
1993 | /* |
1994 | * As multiple blk-throtls may stack in the same issue path, we | |
1995 | * don't want bios to leave with the flag set. Clear the flag if | |
1996 | * being issued. | |
1997 | */ | |
1998 | if (!throttled) | |
8d2bbd4c | 1999 | bio_clear_flag(bio, BIO_THROTTLED); |
bc16a4f9 | 2000 | return throttled; |
e43473b7 VG |
2001 | } |
2002 | ||
9e234eea SL |
2003 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
2004 | void blk_throtl_bio_endio(struct bio *bio) | |
2005 | { | |
2006 | struct throtl_grp *tg; | |
2007 | ||
2008 | tg = bio->bi_cg_private; | |
2009 | if (!tg) | |
2010 | return; | |
2011 | bio->bi_cg_private = NULL; | |
2012 | ||
2013 | tg->last_finish_time = ktime_get_ns() >> 10; | |
2014 | } | |
2015 | #endif | |
2016 | ||
2a12f0dc TH |
2017 | /* |
2018 | * Dispatch all bios from all children tg's queued on @parent_sq. On | |
2019 | * return, @parent_sq is guaranteed to not have any active children tg's | |
2020 | * and all bios from previously active tg's are on @parent_sq->bio_lists[]. | |
2021 | */ | |
2022 | static void tg_drain_bios(struct throtl_service_queue *parent_sq) | |
2023 | { | |
2024 | struct throtl_grp *tg; | |
2025 | ||
2026 | while ((tg = throtl_rb_first(parent_sq))) { | |
2027 | struct throtl_service_queue *sq = &tg->service_queue; | |
2028 | struct bio *bio; | |
2029 | ||
2030 | throtl_dequeue_tg(tg); | |
2031 | ||
c5cc2070 | 2032 | while ((bio = throtl_peek_queued(&sq->queued[READ]))) |
2a12f0dc | 2033 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
c5cc2070 | 2034 | while ((bio = throtl_peek_queued(&sq->queued[WRITE]))) |
2a12f0dc TH |
2035 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
2036 | } | |
2037 | } | |
2038 | ||
c9a929dd TH |
2039 | /** |
2040 | * blk_throtl_drain - drain throttled bios | |
2041 | * @q: request_queue to drain throttled bios for | |
2042 | * | |
2043 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
2044 | */ | |
2045 | void blk_throtl_drain(struct request_queue *q) | |
2046 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
2047 | { | |
2048 | struct throtl_data *td = q->td; | |
2a12f0dc | 2049 | struct blkcg_gq *blkg; |
492eb21b | 2050 | struct cgroup_subsys_state *pos_css; |
c9a929dd | 2051 | struct bio *bio; |
651930bc | 2052 | int rw; |
c9a929dd | 2053 | |
8bcb6c7d | 2054 | queue_lockdep_assert_held(q); |
2a12f0dc | 2055 | rcu_read_lock(); |
c9a929dd | 2056 | |
2a12f0dc TH |
2057 | /* |
2058 | * Drain each tg while doing post-order walk on the blkg tree, so | |
2059 | * that all bios are propagated to td->service_queue. It'd be | |
2060 | * better to walk service_queue tree directly but blkg walk is | |
2061 | * easier. | |
2062 | */ | |
492eb21b | 2063 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) |
2a12f0dc | 2064 | tg_drain_bios(&blkg_to_tg(blkg)->service_queue); |
73f0d49a | 2065 | |
2a12f0dc TH |
2066 | /* finally, transfer bios from top-level tg's into the td */ |
2067 | tg_drain_bios(&td->service_queue); | |
2068 | ||
2069 | rcu_read_unlock(); | |
c9a929dd TH |
2070 | spin_unlock_irq(q->queue_lock); |
2071 | ||
2a12f0dc | 2072 | /* all bios now should be in td->service_queue, issue them */ |
651930bc | 2073 | for (rw = READ; rw <= WRITE; rw++) |
c5cc2070 TH |
2074 | while ((bio = throtl_pop_queued(&td->service_queue.queued[rw], |
2075 | NULL))) | |
651930bc | 2076 | generic_make_request(bio); |
c9a929dd TH |
2077 | |
2078 | spin_lock_irq(q->queue_lock); | |
2079 | } | |
2080 | ||
e43473b7 VG |
2081 | int blk_throtl_init(struct request_queue *q) |
2082 | { | |
2083 | struct throtl_data *td; | |
a2b1693b | 2084 | int ret; |
e43473b7 VG |
2085 | |
2086 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
2087 | if (!td) | |
2088 | return -ENOMEM; | |
2089 | ||
69df0ab0 | 2090 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
b2ce2643 | 2091 | throtl_service_queue_init(&td->service_queue); |
e43473b7 | 2092 | |
cd1604fa | 2093 | q->td = td; |
29b12589 | 2094 | td->queue = q; |
02977e4a | 2095 | |
9f626e37 | 2096 | td->limit_valid[LIMIT_MAX] = true; |
cd5ab1b0 | 2097 | td->limit_index = LIMIT_MAX; |
3f0abd80 SL |
2098 | td->low_upgrade_time = jiffies; |
2099 | td->low_downgrade_time = jiffies; | |
9e234eea | 2100 | |
a2b1693b | 2101 | /* activate policy */ |
3c798398 | 2102 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 2103 | if (ret) |
f51b802c | 2104 | kfree(td); |
a2b1693b | 2105 | return ret; |
e43473b7 VG |
2106 | } |
2107 | ||
2108 | void blk_throtl_exit(struct request_queue *q) | |
2109 | { | |
c875f4d0 | 2110 | BUG_ON(!q->td); |
da527770 | 2111 | throtl_shutdown_wq(q); |
3c798398 | 2112 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 2113 | kfree(q->td); |
e43473b7 VG |
2114 | } |
2115 | ||
d61fcfa4 SL |
2116 | void blk_throtl_register_queue(struct request_queue *q) |
2117 | { | |
2118 | struct throtl_data *td; | |
9e234eea SL |
2119 | struct cgroup_subsys_state *pos_css; |
2120 | struct blkcg_gq *blkg; | |
d61fcfa4 SL |
2121 | |
2122 | td = q->td; | |
2123 | BUG_ON(!td); | |
2124 | ||
2125 | if (blk_queue_nonrot(q)) | |
2126 | td->throtl_slice = DFL_THROTL_SLICE_SSD; | |
2127 | else | |
2128 | td->throtl_slice = DFL_THROTL_SLICE_HD; | |
2129 | #ifndef CONFIG_BLK_DEV_THROTTLING_LOW | |
2130 | /* if no low limit, use previous default */ | |
2131 | td->throtl_slice = DFL_THROTL_SLICE_HD; | |
2132 | #endif | |
9e234eea SL |
2133 | |
2134 | /* | |
2135 | * some tg are created before queue is fully initialized, eg, nonrot | |
2136 | * isn't initialized yet | |
2137 | */ | |
2138 | rcu_read_lock(); | |
2139 | blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) { | |
2140 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
2141 | ||
2142 | if (blk_queue_nonrot(q)) | |
2143 | tg->idletime_threshold = DFL_IDLE_THRESHOLD_SSD; | |
2144 | else | |
2145 | tg->idletime_threshold = DFL_IDLE_THRESHOLD_HD; | |
2146 | } | |
2147 | rcu_read_unlock(); | |
d61fcfa4 SL |
2148 | } |
2149 | ||
297e3d85 SL |
2150 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
2151 | ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page) | |
2152 | { | |
2153 | if (!q->td) | |
2154 | return -EINVAL; | |
2155 | return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice)); | |
2156 | } | |
2157 | ||
2158 | ssize_t blk_throtl_sample_time_store(struct request_queue *q, | |
2159 | const char *page, size_t count) | |
2160 | { | |
2161 | unsigned long v; | |
2162 | unsigned long t; | |
2163 | ||
2164 | if (!q->td) | |
2165 | return -EINVAL; | |
2166 | if (kstrtoul(page, 10, &v)) | |
2167 | return -EINVAL; | |
2168 | t = msecs_to_jiffies(v); | |
2169 | if (t == 0 || t > MAX_THROTL_SLICE) | |
2170 | return -EINVAL; | |
2171 | q->td->throtl_slice = t; | |
2172 | return count; | |
2173 | } | |
2174 | #endif | |
2175 | ||
e43473b7 VG |
2176 | static int __init throtl_init(void) |
2177 | { | |
450adcbe VG |
2178 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
2179 | if (!kthrotld_workqueue) | |
2180 | panic("Failed to create kthrotld\n"); | |
2181 | ||
3c798398 | 2182 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
2183 | } |
2184 | ||
2185 | module_init(throtl_init); |