]>
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> | |
12 | #include "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 | ||
21 | /* Throttling is performed over 100ms slice and after that slice is renewed */ | |
22 | static unsigned long throtl_slice = HZ/10; /* 100 ms */ | |
23 | ||
0381411e TH |
24 | static struct blkio_policy_type blkio_policy_throtl; |
25 | ||
450adcbe VG |
26 | /* A workqueue to queue throttle related work */ |
27 | static struct workqueue_struct *kthrotld_workqueue; | |
28 | static void throtl_schedule_delayed_work(struct throtl_data *td, | |
29 | unsigned long delay); | |
30 | ||
e43473b7 VG |
31 | struct throtl_rb_root { |
32 | struct rb_root rb; | |
33 | struct rb_node *left; | |
34 | unsigned int count; | |
35 | unsigned long min_disptime; | |
36 | }; | |
37 | ||
38 | #define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \ | |
39 | .count = 0, .min_disptime = 0} | |
40 | ||
41 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) | |
42 | ||
43 | struct throtl_grp { | |
e43473b7 VG |
44 | /* active throtl group service_tree member */ |
45 | struct rb_node rb_node; | |
46 | ||
47 | /* | |
48 | * Dispatch time in jiffies. This is the estimated time when group | |
49 | * will unthrottle and is ready to dispatch more bio. It is used as | |
50 | * key to sort active groups in service tree. | |
51 | */ | |
52 | unsigned long disptime; | |
53 | ||
e43473b7 VG |
54 | unsigned int flags; |
55 | ||
56 | /* Two lists for READ and WRITE */ | |
57 | struct bio_list bio_lists[2]; | |
58 | ||
59 | /* Number of queued bios on READ and WRITE lists */ | |
60 | unsigned int nr_queued[2]; | |
61 | ||
62 | /* bytes per second rate limits */ | |
63 | uint64_t bps[2]; | |
64 | ||
8e89d13f VG |
65 | /* IOPS limits */ |
66 | unsigned int iops[2]; | |
67 | ||
e43473b7 VG |
68 | /* Number of bytes disptached in current slice */ |
69 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
70 | /* Number of bio's dispatched in current slice */ |
71 | unsigned int io_disp[2]; | |
e43473b7 VG |
72 | |
73 | /* When did we start a new slice */ | |
74 | unsigned long slice_start[2]; | |
75 | unsigned long slice_end[2]; | |
fe071437 VG |
76 | |
77 | /* Some throttle limits got updated for the group */ | |
6f037937 | 78 | int limits_changed; |
e43473b7 VG |
79 | }; |
80 | ||
81 | struct throtl_data | |
82 | { | |
e43473b7 VG |
83 | /* service tree for active throtl groups */ |
84 | struct throtl_rb_root tg_service_tree; | |
85 | ||
29b12589 | 86 | struct throtl_grp *root_tg; |
e43473b7 VG |
87 | struct request_queue *queue; |
88 | ||
89 | /* Total Number of queued bios on READ and WRITE lists */ | |
90 | unsigned int nr_queued[2]; | |
91 | ||
92 | /* | |
02977e4a | 93 | * number of total undestroyed groups |
e43473b7 VG |
94 | */ |
95 | unsigned int nr_undestroyed_grps; | |
96 | ||
97 | /* Work for dispatching throttled bios */ | |
98 | struct delayed_work throtl_work; | |
fe071437 | 99 | |
6f037937 | 100 | int limits_changed; |
e43473b7 VG |
101 | }; |
102 | ||
0381411e TH |
103 | static inline struct throtl_grp *blkg_to_tg(struct blkio_group *blkg) |
104 | { | |
105 | return blkg_to_pdata(blkg, &blkio_policy_throtl); | |
106 | } | |
107 | ||
108 | static inline struct blkio_group *tg_to_blkg(struct throtl_grp *tg) | |
109 | { | |
aaec55a0 | 110 | return pdata_to_blkg(tg); |
0381411e TH |
111 | } |
112 | ||
e43473b7 VG |
113 | enum tg_state_flags { |
114 | THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */ | |
115 | }; | |
116 | ||
117 | #define THROTL_TG_FNS(name) \ | |
118 | static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \ | |
119 | { \ | |
120 | (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \ | |
121 | } \ | |
122 | static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \ | |
123 | { \ | |
124 | (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \ | |
125 | } \ | |
126 | static inline int throtl_tg_##name(const struct throtl_grp *tg) \ | |
127 | { \ | |
128 | return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \ | |
129 | } | |
130 | ||
131 | THROTL_TG_FNS(on_rr); | |
132 | ||
133 | #define throtl_log_tg(td, tg, fmt, args...) \ | |
134 | blk_add_trace_msg((td)->queue, "throtl %s " fmt, \ | |
0381411e | 135 | blkg_path(tg_to_blkg(tg)), ##args); \ |
e43473b7 VG |
136 | |
137 | #define throtl_log(td, fmt, args...) \ | |
138 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) | |
139 | ||
d2f31a5f | 140 | static inline unsigned int total_nr_queued(struct throtl_data *td) |
e43473b7 | 141 | { |
d2f31a5f | 142 | return td->nr_queued[0] + td->nr_queued[1]; |
e43473b7 VG |
143 | } |
144 | ||
0381411e | 145 | static void throtl_init_blkio_group(struct blkio_group *blkg) |
a29a171e | 146 | { |
0381411e | 147 | struct throtl_grp *tg = blkg_to_tg(blkg); |
cd1604fa | 148 | |
a29a171e VG |
149 | RB_CLEAR_NODE(&tg->rb_node); |
150 | bio_list_init(&tg->bio_lists[0]); | |
151 | bio_list_init(&tg->bio_lists[1]); | |
152 | tg->limits_changed = false; | |
153 | ||
e56da7e2 TH |
154 | tg->bps[READ] = -1; |
155 | tg->bps[WRITE] = -1; | |
156 | tg->iops[READ] = -1; | |
157 | tg->iops[WRITE] = -1; | |
a29a171e VG |
158 | } |
159 | ||
f469a7b4 | 160 | static struct |
cd1604fa | 161 | throtl_grp *throtl_lookup_tg(struct throtl_data *td, struct blkio_cgroup *blkcg) |
e43473b7 | 162 | { |
be2c6b19 VG |
163 | /* |
164 | * This is the common case when there are no blkio cgroups. | |
cd1604fa TH |
165 | * Avoid lookup in this case |
166 | */ | |
be2c6b19 | 167 | if (blkcg == &blkio_root_cgroup) |
7a4dd281 | 168 | return td->root_tg; |
e43473b7 | 169 | |
e8989fae | 170 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
171 | } |
172 | ||
cd1604fa TH |
173 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
174 | struct blkio_cgroup *blkcg) | |
e43473b7 | 175 | { |
f469a7b4 | 176 | struct request_queue *q = td->queue; |
cd1604fa | 177 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 178 | |
f469a7b4 | 179 | /* |
cd1604fa TH |
180 | * This is the common case when there are no blkio cgroups. |
181 | * Avoid lookup in this case | |
f469a7b4 | 182 | */ |
cd1604fa TH |
183 | if (blkcg == &blkio_root_cgroup) { |
184 | tg = td->root_tg; | |
185 | } else { | |
186 | struct blkio_group *blkg; | |
f469a7b4 | 187 | |
aaec55a0 | 188 | blkg = blkg_lookup_create(blkcg, q, false); |
f469a7b4 | 189 | |
cd1604fa TH |
190 | /* if %NULL and @q is alive, fall back to root_tg */ |
191 | if (!IS_ERR(blkg)) | |
0381411e | 192 | tg = blkg_to_tg(blkg); |
cd1604fa TH |
193 | else if (!blk_queue_dead(q)) |
194 | tg = td->root_tg; | |
f469a7b4 VG |
195 | } |
196 | ||
e43473b7 VG |
197 | return tg; |
198 | } | |
199 | ||
200 | static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root) | |
201 | { | |
202 | /* Service tree is empty */ | |
203 | if (!root->count) | |
204 | return NULL; | |
205 | ||
206 | if (!root->left) | |
207 | root->left = rb_first(&root->rb); | |
208 | ||
209 | if (root->left) | |
210 | return rb_entry_tg(root->left); | |
211 | ||
212 | return NULL; | |
213 | } | |
214 | ||
215 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
216 | { | |
217 | rb_erase(n, root); | |
218 | RB_CLEAR_NODE(n); | |
219 | } | |
220 | ||
221 | static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root) | |
222 | { | |
223 | if (root->left == n) | |
224 | root->left = NULL; | |
225 | rb_erase_init(n, &root->rb); | |
226 | --root->count; | |
227 | } | |
228 | ||
229 | static void update_min_dispatch_time(struct throtl_rb_root *st) | |
230 | { | |
231 | struct throtl_grp *tg; | |
232 | ||
233 | tg = throtl_rb_first(st); | |
234 | if (!tg) | |
235 | return; | |
236 | ||
237 | st->min_disptime = tg->disptime; | |
238 | } | |
239 | ||
240 | static void | |
241 | tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg) | |
242 | { | |
243 | struct rb_node **node = &st->rb.rb_node; | |
244 | struct rb_node *parent = NULL; | |
245 | struct throtl_grp *__tg; | |
246 | unsigned long key = tg->disptime; | |
247 | int left = 1; | |
248 | ||
249 | while (*node != NULL) { | |
250 | parent = *node; | |
251 | __tg = rb_entry_tg(parent); | |
252 | ||
253 | if (time_before(key, __tg->disptime)) | |
254 | node = &parent->rb_left; | |
255 | else { | |
256 | node = &parent->rb_right; | |
257 | left = 0; | |
258 | } | |
259 | } | |
260 | ||
261 | if (left) | |
262 | st->left = &tg->rb_node; | |
263 | ||
264 | rb_link_node(&tg->rb_node, parent, node); | |
265 | rb_insert_color(&tg->rb_node, &st->rb); | |
266 | } | |
267 | ||
268 | static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
269 | { | |
270 | struct throtl_rb_root *st = &td->tg_service_tree; | |
271 | ||
272 | tg_service_tree_add(st, tg); | |
273 | throtl_mark_tg_on_rr(tg); | |
274 | st->count++; | |
275 | } | |
276 | ||
277 | static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
278 | { | |
279 | if (!throtl_tg_on_rr(tg)) | |
280 | __throtl_enqueue_tg(td, tg); | |
281 | } | |
282 | ||
283 | static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
284 | { | |
285 | throtl_rb_erase(&tg->rb_node, &td->tg_service_tree); | |
286 | throtl_clear_tg_on_rr(tg); | |
287 | } | |
288 | ||
289 | static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
290 | { | |
291 | if (throtl_tg_on_rr(tg)) | |
292 | __throtl_dequeue_tg(td, tg); | |
293 | } | |
294 | ||
295 | static void throtl_schedule_next_dispatch(struct throtl_data *td) | |
296 | { | |
297 | struct throtl_rb_root *st = &td->tg_service_tree; | |
298 | ||
299 | /* | |
300 | * If there are more bios pending, schedule more work. | |
301 | */ | |
302 | if (!total_nr_queued(td)) | |
303 | return; | |
304 | ||
305 | BUG_ON(!st->count); | |
306 | ||
307 | update_min_dispatch_time(st); | |
308 | ||
309 | if (time_before_eq(st->min_disptime, jiffies)) | |
450adcbe | 310 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 311 | else |
450adcbe | 312 | throtl_schedule_delayed_work(td, (st->min_disptime - jiffies)); |
e43473b7 VG |
313 | } |
314 | ||
315 | static inline void | |
316 | throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
317 | { | |
318 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 319 | tg->io_disp[rw] = 0; |
e43473b7 VG |
320 | tg->slice_start[rw] = jiffies; |
321 | tg->slice_end[rw] = jiffies + throtl_slice; | |
322 | throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
323 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
324 | tg->slice_end[rw], jiffies); | |
325 | } | |
326 | ||
d1ae8ffd VG |
327 | static inline void throtl_set_slice_end(struct throtl_data *td, |
328 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
329 | { | |
330 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
331 | } | |
332 | ||
e43473b7 VG |
333 | static inline void throtl_extend_slice(struct throtl_data *td, |
334 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
335 | { | |
336 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
337 | throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
338 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
339 | tg->slice_end[rw], jiffies); | |
340 | } | |
341 | ||
342 | /* Determine if previously allocated or extended slice is complete or not */ | |
343 | static bool | |
344 | throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
345 | { | |
346 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
347 | return 0; | |
348 | ||
349 | return 1; | |
350 | } | |
351 | ||
352 | /* Trim the used slices and adjust slice start accordingly */ | |
353 | static inline void | |
354 | throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
355 | { | |
3aad5d3e VG |
356 | unsigned long nr_slices, time_elapsed, io_trim; |
357 | u64 bytes_trim, tmp; | |
e43473b7 VG |
358 | |
359 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
360 | ||
361 | /* | |
362 | * If bps are unlimited (-1), then time slice don't get | |
363 | * renewed. Don't try to trim the slice if slice is used. A new | |
364 | * slice will start when appropriate. | |
365 | */ | |
366 | if (throtl_slice_used(td, tg, rw)) | |
367 | return; | |
368 | ||
d1ae8ffd VG |
369 | /* |
370 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
371 | * that initially cgroup limit was very low resulting in high | |
372 | * slice_end, but later limit was bumped up and bio was dispached | |
373 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
374 | * is bad because it does not allow new slice to start. | |
375 | */ | |
376 | ||
377 | throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice); | |
378 | ||
e43473b7 VG |
379 | time_elapsed = jiffies - tg->slice_start[rw]; |
380 | ||
381 | nr_slices = time_elapsed / throtl_slice; | |
382 | ||
383 | if (!nr_slices) | |
384 | return; | |
3aad5d3e VG |
385 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
386 | do_div(tmp, HZ); | |
387 | bytes_trim = tmp; | |
e43473b7 | 388 | |
8e89d13f | 389 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 390 | |
8e89d13f | 391 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
392 | return; |
393 | ||
394 | if (tg->bytes_disp[rw] >= bytes_trim) | |
395 | tg->bytes_disp[rw] -= bytes_trim; | |
396 | else | |
397 | tg->bytes_disp[rw] = 0; | |
398 | ||
8e89d13f VG |
399 | if (tg->io_disp[rw] >= io_trim) |
400 | tg->io_disp[rw] -= io_trim; | |
401 | else | |
402 | tg->io_disp[rw] = 0; | |
403 | ||
e43473b7 VG |
404 | tg->slice_start[rw] += nr_slices * throtl_slice; |
405 | ||
3aad5d3e | 406 | throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 407 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 408 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
409 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
410 | } | |
411 | ||
8e89d13f VG |
412 | static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg, |
413 | struct bio *bio, unsigned long *wait) | |
e43473b7 VG |
414 | { |
415 | bool rw = bio_data_dir(bio); | |
8e89d13f | 416 | unsigned int io_allowed; |
e43473b7 | 417 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 418 | u64 tmp; |
e43473b7 | 419 | |
8e89d13f | 420 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 421 | |
8e89d13f VG |
422 | /* Slice has just started. Consider one slice interval */ |
423 | if (!jiffy_elapsed) | |
424 | jiffy_elapsed_rnd = throtl_slice; | |
425 | ||
426 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
427 | ||
c49c06e4 VG |
428 | /* |
429 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
430 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
431 | * will allow dispatch after 1 second and after that slice should | |
432 | * have been trimmed. | |
433 | */ | |
434 | ||
435 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
436 | do_div(tmp, HZ); | |
437 | ||
438 | if (tmp > UINT_MAX) | |
439 | io_allowed = UINT_MAX; | |
440 | else | |
441 | io_allowed = tmp; | |
8e89d13f VG |
442 | |
443 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
444 | if (wait) |
445 | *wait = 0; | |
446 | return 1; | |
447 | } | |
448 | ||
8e89d13f VG |
449 | /* Calc approx time to dispatch */ |
450 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
451 | ||
452 | if (jiffy_wait > jiffy_elapsed) | |
453 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
454 | else | |
455 | jiffy_wait = 1; | |
456 | ||
457 | if (wait) | |
458 | *wait = jiffy_wait; | |
459 | return 0; | |
460 | } | |
461 | ||
462 | static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg, | |
463 | struct bio *bio, unsigned long *wait) | |
464 | { | |
465 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 466 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 467 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
468 | |
469 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
470 | ||
471 | /* Slice has just started. Consider one slice interval */ | |
472 | if (!jiffy_elapsed) | |
473 | jiffy_elapsed_rnd = throtl_slice; | |
474 | ||
475 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
476 | ||
5e901a2b VG |
477 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
478 | do_div(tmp, HZ); | |
3aad5d3e | 479 | bytes_allowed = tmp; |
e43473b7 VG |
480 | |
481 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
482 | if (wait) | |
483 | *wait = 0; | |
484 | return 1; | |
485 | } | |
486 | ||
487 | /* Calc approx time to dispatch */ | |
488 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
489 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
490 | ||
491 | if (!jiffy_wait) | |
492 | jiffy_wait = 1; | |
493 | ||
494 | /* | |
495 | * This wait time is without taking into consideration the rounding | |
496 | * up we did. Add that time also. | |
497 | */ | |
498 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
499 | if (wait) |
500 | *wait = jiffy_wait; | |
8e89d13f VG |
501 | return 0; |
502 | } | |
503 | ||
af75cd3c VG |
504 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
505 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
506 | return 1; | |
507 | return 0; | |
508 | } | |
509 | ||
8e89d13f VG |
510 | /* |
511 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
512 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
513 | */ | |
514 | static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg, | |
515 | struct bio *bio, unsigned long *wait) | |
516 | { | |
517 | bool rw = bio_data_dir(bio); | |
518 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
519 | ||
520 | /* | |
521 | * Currently whole state machine of group depends on first bio | |
522 | * queued in the group bio list. So one should not be calling | |
523 | * this function with a different bio if there are other bios | |
524 | * queued. | |
525 | */ | |
526 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); | |
e43473b7 | 527 | |
8e89d13f VG |
528 | /* If tg->bps = -1, then BW is unlimited */ |
529 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
530 | if (wait) | |
531 | *wait = 0; | |
532 | return 1; | |
533 | } | |
534 | ||
535 | /* | |
536 | * If previous slice expired, start a new one otherwise renew/extend | |
537 | * existing slice to make sure it is at least throtl_slice interval | |
538 | * long since now. | |
539 | */ | |
540 | if (throtl_slice_used(td, tg, rw)) | |
541 | throtl_start_new_slice(td, tg, rw); | |
542 | else { | |
543 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
544 | throtl_extend_slice(td, tg, rw, jiffies + throtl_slice); | |
545 | } | |
546 | ||
547 | if (tg_with_in_bps_limit(td, tg, bio, &bps_wait) | |
548 | && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) { | |
549 | if (wait) | |
550 | *wait = 0; | |
551 | return 1; | |
552 | } | |
553 | ||
554 | max_wait = max(bps_wait, iops_wait); | |
555 | ||
556 | if (wait) | |
557 | *wait = max_wait; | |
558 | ||
559 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
560 | throtl_extend_slice(td, tg, rw, jiffies + max_wait); | |
e43473b7 VG |
561 | |
562 | return 0; | |
563 | } | |
564 | ||
629ed0b1 TH |
565 | static void throtl_update_dispatch_stats(struct blkio_group *blkg, u64 bytes, |
566 | int rw) | |
567 | { | |
568 | struct blkg_policy_data *pd = blkg->pd[BLKIO_POLICY_THROTL]; | |
569 | struct blkio_group_stats_cpu *stats_cpu; | |
570 | unsigned long flags; | |
571 | ||
572 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
573 | if (pd->stats_cpu == NULL) | |
574 | return; | |
575 | ||
576 | /* | |
577 | * Disabling interrupts to provide mutual exclusion between two | |
578 | * writes on same cpu. It probably is not needed for 64bit. Not | |
579 | * optimizing that case yet. | |
580 | */ | |
581 | local_irq_save(flags); | |
582 | ||
583 | stats_cpu = this_cpu_ptr(pd->stats_cpu); | |
584 | ||
629ed0b1 TH |
585 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
586 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
587 | ||
588 | local_irq_restore(flags); | |
589 | } | |
590 | ||
e43473b7 VG |
591 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
592 | { | |
593 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
594 | |
595 | /* Charge the bio to the group */ | |
596 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 597 | tg->io_disp[rw]++; |
e43473b7 | 598 | |
629ed0b1 | 599 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw); |
e43473b7 VG |
600 | } |
601 | ||
602 | static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg, | |
603 | struct bio *bio) | |
604 | { | |
605 | bool rw = bio_data_dir(bio); | |
606 | ||
607 | bio_list_add(&tg->bio_lists[rw], bio); | |
608 | /* Take a bio reference on tg */ | |
1adaf3dd | 609 | blkg_get(tg_to_blkg(tg)); |
e43473b7 VG |
610 | tg->nr_queued[rw]++; |
611 | td->nr_queued[rw]++; | |
612 | throtl_enqueue_tg(td, tg); | |
613 | } | |
614 | ||
615 | static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg) | |
616 | { | |
617 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
618 | struct bio *bio; | |
619 | ||
620 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
621 | tg_may_dispatch(td, tg, bio, &read_wait); | |
622 | ||
623 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
624 | tg_may_dispatch(td, tg, bio, &write_wait); | |
625 | ||
626 | min_wait = min(read_wait, write_wait); | |
627 | disptime = jiffies + min_wait; | |
628 | ||
e43473b7 VG |
629 | /* Update dispatch time */ |
630 | throtl_dequeue_tg(td, tg); | |
631 | tg->disptime = disptime; | |
632 | throtl_enqueue_tg(td, tg); | |
633 | } | |
634 | ||
635 | static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg, | |
636 | bool rw, struct bio_list *bl) | |
637 | { | |
638 | struct bio *bio; | |
639 | ||
640 | bio = bio_list_pop(&tg->bio_lists[rw]); | |
641 | tg->nr_queued[rw]--; | |
1adaf3dd TH |
642 | /* Drop bio reference on blkg */ |
643 | blkg_put(tg_to_blkg(tg)); | |
e43473b7 VG |
644 | |
645 | BUG_ON(td->nr_queued[rw] <= 0); | |
646 | td->nr_queued[rw]--; | |
647 | ||
648 | throtl_charge_bio(tg, bio); | |
649 | bio_list_add(bl, bio); | |
650 | bio->bi_rw |= REQ_THROTTLED; | |
651 | ||
652 | throtl_trim_slice(td, tg, rw); | |
653 | } | |
654 | ||
655 | static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg, | |
656 | struct bio_list *bl) | |
657 | { | |
658 | unsigned int nr_reads = 0, nr_writes = 0; | |
659 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 660 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
661 | struct bio *bio; |
662 | ||
663 | /* Try to dispatch 75% READS and 25% WRITES */ | |
664 | ||
665 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) | |
666 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
667 | ||
668 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
669 | nr_reads++; | |
670 | ||
671 | if (nr_reads >= max_nr_reads) | |
672 | break; | |
673 | } | |
674 | ||
675 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) | |
676 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
677 | ||
678 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
679 | nr_writes++; | |
680 | ||
681 | if (nr_writes >= max_nr_writes) | |
682 | break; | |
683 | } | |
684 | ||
685 | return nr_reads + nr_writes; | |
686 | } | |
687 | ||
688 | static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl) | |
689 | { | |
690 | unsigned int nr_disp = 0; | |
691 | struct throtl_grp *tg; | |
692 | struct throtl_rb_root *st = &td->tg_service_tree; | |
693 | ||
694 | while (1) { | |
695 | tg = throtl_rb_first(st); | |
696 | ||
697 | if (!tg) | |
698 | break; | |
699 | ||
700 | if (time_before(jiffies, tg->disptime)) | |
701 | break; | |
702 | ||
703 | throtl_dequeue_tg(td, tg); | |
704 | ||
705 | nr_disp += throtl_dispatch_tg(td, tg, bl); | |
706 | ||
707 | if (tg->nr_queued[0] || tg->nr_queued[1]) { | |
708 | tg_update_disptime(td, tg); | |
709 | throtl_enqueue_tg(td, tg); | |
710 | } | |
711 | ||
712 | if (nr_disp >= throtl_quantum) | |
713 | break; | |
714 | } | |
715 | ||
716 | return nr_disp; | |
717 | } | |
718 | ||
fe071437 VG |
719 | static void throtl_process_limit_change(struct throtl_data *td) |
720 | { | |
4eef3049 TH |
721 | struct request_queue *q = td->queue; |
722 | struct blkio_group *blkg, *n; | |
fe071437 | 723 | |
de701c74 | 724 | if (!td->limits_changed) |
fe071437 VG |
725 | return; |
726 | ||
de701c74 | 727 | xchg(&td->limits_changed, false); |
fe071437 | 728 | |
de701c74 | 729 | throtl_log(td, "limits changed"); |
fe071437 | 730 | |
e8989fae | 731 | list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) { |
4eef3049 TH |
732 | struct throtl_grp *tg = blkg_to_tg(blkg); |
733 | ||
de701c74 VG |
734 | if (!tg->limits_changed) |
735 | continue; | |
736 | ||
737 | if (!xchg(&tg->limits_changed, false)) | |
738 | continue; | |
739 | ||
740 | throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu" | |
741 | " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE], | |
742 | tg->iops[READ], tg->iops[WRITE]); | |
743 | ||
04521db0 VG |
744 | /* |
745 | * Restart the slices for both READ and WRITES. It | |
746 | * might happen that a group's limit are dropped | |
747 | * suddenly and we don't want to account recently | |
748 | * dispatched IO with new low rate | |
749 | */ | |
750 | throtl_start_new_slice(td, tg, 0); | |
751 | throtl_start_new_slice(td, tg, 1); | |
752 | ||
de701c74 | 753 | if (throtl_tg_on_rr(tg)) |
fe071437 | 754 | tg_update_disptime(td, tg); |
fe071437 | 755 | } |
fe071437 VG |
756 | } |
757 | ||
e43473b7 VG |
758 | /* Dispatch throttled bios. Should be called without queue lock held. */ |
759 | static int throtl_dispatch(struct request_queue *q) | |
760 | { | |
761 | struct throtl_data *td = q->td; | |
762 | unsigned int nr_disp = 0; | |
763 | struct bio_list bio_list_on_stack; | |
764 | struct bio *bio; | |
69d60eb9 | 765 | struct blk_plug plug; |
e43473b7 VG |
766 | |
767 | spin_lock_irq(q->queue_lock); | |
768 | ||
fe071437 VG |
769 | throtl_process_limit_change(td); |
770 | ||
e43473b7 VG |
771 | if (!total_nr_queued(td)) |
772 | goto out; | |
773 | ||
774 | bio_list_init(&bio_list_on_stack); | |
775 | ||
d2f31a5f | 776 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
e43473b7 VG |
777 | total_nr_queued(td), td->nr_queued[READ], |
778 | td->nr_queued[WRITE]); | |
779 | ||
780 | nr_disp = throtl_select_dispatch(td, &bio_list_on_stack); | |
781 | ||
782 | if (nr_disp) | |
783 | throtl_log(td, "bios disp=%u", nr_disp); | |
784 | ||
785 | throtl_schedule_next_dispatch(td); | |
786 | out: | |
787 | spin_unlock_irq(q->queue_lock); | |
788 | ||
789 | /* | |
790 | * If we dispatched some requests, unplug the queue to make sure | |
791 | * immediate dispatch | |
792 | */ | |
793 | if (nr_disp) { | |
69d60eb9 | 794 | blk_start_plug(&plug); |
e43473b7 VG |
795 | while((bio = bio_list_pop(&bio_list_on_stack))) |
796 | generic_make_request(bio); | |
69d60eb9 | 797 | blk_finish_plug(&plug); |
e43473b7 VG |
798 | } |
799 | return nr_disp; | |
800 | } | |
801 | ||
802 | void blk_throtl_work(struct work_struct *work) | |
803 | { | |
804 | struct throtl_data *td = container_of(work, struct throtl_data, | |
805 | throtl_work.work); | |
806 | struct request_queue *q = td->queue; | |
807 | ||
808 | throtl_dispatch(q); | |
809 | } | |
810 | ||
811 | /* Call with queue lock held */ | |
450adcbe VG |
812 | static void |
813 | throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay) | |
e43473b7 VG |
814 | { |
815 | ||
e43473b7 VG |
816 | struct delayed_work *dwork = &td->throtl_work; |
817 | ||
04521db0 | 818 | /* schedule work if limits changed even if no bio is queued */ |
d2f31a5f | 819 | if (total_nr_queued(td) || td->limits_changed) { |
e43473b7 VG |
820 | /* |
821 | * We might have a work scheduled to be executed in future. | |
822 | * Cancel that and schedule a new one. | |
823 | */ | |
824 | __cancel_delayed_work(dwork); | |
450adcbe | 825 | queue_delayed_work(kthrotld_workqueue, dwork, delay); |
e43473b7 VG |
826 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", |
827 | delay, jiffies); | |
828 | } | |
829 | } | |
e43473b7 | 830 | |
60c2bc2d TH |
831 | /* |
832 | * Can not take queue lock in update functions as queue lock under | |
833 | * blkcg_lock is not allowed. Under other paths we take blkcg_lock under | |
834 | * queue_lock. | |
835 | */ | |
de701c74 VG |
836 | static void throtl_update_blkio_group_common(struct throtl_data *td, |
837 | struct throtl_grp *tg) | |
838 | { | |
839 | xchg(&tg->limits_changed, true); | |
840 | xchg(&td->limits_changed, true); | |
841 | /* Schedule a work now to process the limit change */ | |
842 | throtl_schedule_delayed_work(td, 0); | |
843 | } | |
844 | ||
41b38b6d TH |
845 | static u64 blkg_prfill_cpu_rwstat(struct seq_file *sf, |
846 | struct blkg_policy_data *pd, int off) | |
847 | { | |
848 | struct blkg_rwstat rwstat = { }, tmp; | |
849 | int i, cpu; | |
850 | ||
851 | for_each_possible_cpu(cpu) { | |
852 | struct blkio_group_stats_cpu *sc = | |
853 | per_cpu_ptr(pd->stats_cpu, cpu); | |
854 | ||
855 | tmp = blkg_rwstat_read((void *)sc + off); | |
856 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
857 | rwstat.cnt[i] += tmp.cnt[i]; | |
858 | } | |
859 | ||
860 | return __blkg_prfill_rwstat(sf, pd, &rwstat); | |
861 | } | |
862 | ||
863 | /* print per-cpu blkg_rwstat specified by BLKCG_STAT_PRIV() */ | |
864 | static int blkcg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, | |
865 | struct seq_file *sf) | |
866 | { | |
867 | struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgrp); | |
868 | ||
869 | blkcg_print_blkgs(sf, blkcg, blkg_prfill_cpu_rwstat, | |
870 | BLKCG_STAT_POL(cft->private), | |
871 | BLKCG_STAT_OFF(cft->private), true); | |
872 | return 0; | |
873 | } | |
874 | ||
60c2bc2d TH |
875 | static u64 blkg_prfill_conf_u64(struct seq_file *sf, |
876 | struct blkg_policy_data *pd, int off) | |
877 | { | |
878 | u64 v = *(u64 *)((void *)&pd->conf + off); | |
879 | ||
880 | if (!v) | |
881 | return 0; | |
882 | return __blkg_prfill_u64(sf, pd, v); | |
883 | } | |
884 | ||
885 | static int blkcg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, | |
886 | struct seq_file *sf) | |
887 | { | |
888 | blkcg_print_blkgs(sf, cgroup_to_blkio_cgroup(cgrp), | |
889 | blkg_prfill_conf_u64, BLKIO_POLICY_THROTL, | |
890 | cft->private, false); | |
891 | return 0; | |
892 | } | |
893 | ||
894 | static void throtl_update_blkio_group_read_bps(struct blkio_group *blkg, | |
895 | u64 read_bps) | |
e43473b7 | 896 | { |
0381411e | 897 | struct throtl_grp *tg = blkg_to_tg(blkg); |
fe071437 | 898 | |
de701c74 | 899 | tg->bps[READ] = read_bps; |
60c2bc2d | 900 | throtl_update_blkio_group_common(blkg->q->td, tg); |
e43473b7 VG |
901 | } |
902 | ||
60c2bc2d TH |
903 | static void throtl_update_blkio_group_write_bps(struct blkio_group *blkg, |
904 | u64 write_bps) | |
e43473b7 | 905 | { |
0381411e | 906 | struct throtl_grp *tg = blkg_to_tg(blkg); |
fe071437 | 907 | |
de701c74 | 908 | tg->bps[WRITE] = write_bps; |
60c2bc2d | 909 | throtl_update_blkio_group_common(blkg->q->td, tg); |
e43473b7 VG |
910 | } |
911 | ||
60c2bc2d TH |
912 | static void throtl_update_blkio_group_read_iops(struct blkio_group *blkg, |
913 | u64 read_iops) | |
8e89d13f | 914 | { |
0381411e | 915 | struct throtl_grp *tg = blkg_to_tg(blkg); |
fe071437 | 916 | |
de701c74 | 917 | tg->iops[READ] = read_iops; |
60c2bc2d | 918 | throtl_update_blkio_group_common(blkg->q->td, tg); |
8e89d13f VG |
919 | } |
920 | ||
60c2bc2d TH |
921 | static void throtl_update_blkio_group_write_iops(struct blkio_group *blkg, |
922 | u64 write_iops) | |
8e89d13f | 923 | { |
0381411e | 924 | struct throtl_grp *tg = blkg_to_tg(blkg); |
fe071437 | 925 | |
de701c74 | 926 | tg->iops[WRITE] = write_iops; |
60c2bc2d TH |
927 | throtl_update_blkio_group_common(blkg->q->td, tg); |
928 | } | |
929 | ||
930 | static int blkcg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, | |
931 | const char *buf, | |
932 | void (*update)(struct blkio_group *, u64)) | |
933 | { | |
934 | struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgrp); | |
935 | struct blkg_policy_data *pd; | |
936 | struct blkg_conf_ctx ctx; | |
937 | int ret; | |
938 | ||
939 | ret = blkg_conf_prep(blkcg, buf, &ctx); | |
940 | if (ret) | |
941 | return ret; | |
942 | ||
943 | ret = -EINVAL; | |
944 | pd = ctx.blkg->pd[BLKIO_POLICY_THROTL]; | |
945 | if (pd) { | |
946 | *(u64 *)((void *)&pd->conf + cft->private) = ctx.v; | |
947 | update(ctx.blkg, ctx.v ?: -1); | |
948 | ret = 0; | |
949 | } | |
950 | ||
951 | blkg_conf_finish(&ctx); | |
952 | return ret; | |
8e89d13f VG |
953 | } |
954 | ||
60c2bc2d TH |
955 | static int blkcg_set_conf_bps_r(struct cgroup *cgrp, struct cftype *cft, |
956 | const char *buf) | |
957 | { | |
958 | return blkcg_set_conf_u64(cgrp, cft, buf, | |
959 | throtl_update_blkio_group_read_bps); | |
960 | } | |
961 | ||
962 | static int blkcg_set_conf_bps_w(struct cgroup *cgrp, struct cftype *cft, | |
963 | const char *buf) | |
964 | { | |
965 | return blkcg_set_conf_u64(cgrp, cft, buf, | |
966 | throtl_update_blkio_group_write_bps); | |
967 | } | |
968 | ||
969 | static int blkcg_set_conf_iops_r(struct cgroup *cgrp, struct cftype *cft, | |
970 | const char *buf) | |
971 | { | |
972 | return blkcg_set_conf_u64(cgrp, cft, buf, | |
973 | throtl_update_blkio_group_read_iops); | |
974 | } | |
975 | ||
976 | static int blkcg_set_conf_iops_w(struct cgroup *cgrp, struct cftype *cft, | |
977 | const char *buf) | |
978 | { | |
979 | return blkcg_set_conf_u64(cgrp, cft, buf, | |
980 | throtl_update_blkio_group_write_iops); | |
981 | } | |
982 | ||
983 | static struct cftype throtl_files[] = { | |
984 | { | |
985 | .name = "throttle.read_bps_device", | |
986 | .private = offsetof(struct blkio_group_conf, bps[READ]), | |
987 | .read_seq_string = blkcg_print_conf_u64, | |
988 | .write_string = blkcg_set_conf_bps_r, | |
989 | .max_write_len = 256, | |
990 | }, | |
991 | { | |
992 | .name = "throttle.write_bps_device", | |
993 | .private = offsetof(struct blkio_group_conf, bps[WRITE]), | |
994 | .read_seq_string = blkcg_print_conf_u64, | |
995 | .write_string = blkcg_set_conf_bps_w, | |
996 | .max_write_len = 256, | |
997 | }, | |
998 | { | |
999 | .name = "throttle.read_iops_device", | |
1000 | .private = offsetof(struct blkio_group_conf, iops[READ]), | |
1001 | .read_seq_string = blkcg_print_conf_u64, | |
1002 | .write_string = blkcg_set_conf_iops_r, | |
1003 | .max_write_len = 256, | |
1004 | }, | |
1005 | { | |
1006 | .name = "throttle.write_iops_device", | |
1007 | .private = offsetof(struct blkio_group_conf, iops[WRITE]), | |
1008 | .read_seq_string = blkcg_print_conf_u64, | |
1009 | .write_string = blkcg_set_conf_iops_w, | |
1010 | .max_write_len = 256, | |
1011 | }, | |
1012 | { | |
1013 | .name = "throttle.io_service_bytes", | |
1014 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_THROTL, | |
1015 | offsetof(struct blkio_group_stats_cpu, service_bytes)), | |
1016 | .read_seq_string = blkcg_print_cpu_rwstat, | |
1017 | }, | |
1018 | { | |
1019 | .name = "throttle.io_serviced", | |
1020 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_THROTL, | |
1021 | offsetof(struct blkio_group_stats_cpu, serviced)), | |
1022 | .read_seq_string = blkcg_print_cpu_rwstat, | |
1023 | }, | |
1024 | { } /* terminate */ | |
1025 | }; | |
1026 | ||
da527770 | 1027 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1028 | { |
1029 | struct throtl_data *td = q->td; | |
1030 | ||
1031 | cancel_delayed_work_sync(&td->throtl_work); | |
1032 | } | |
1033 | ||
1034 | static struct blkio_policy_type blkio_policy_throtl = { | |
1035 | .ops = { | |
0381411e | 1036 | .blkio_init_group_fn = throtl_init_blkio_group, |
e43473b7 | 1037 | }, |
8e89d13f | 1038 | .plid = BLKIO_POLICY_THROTL, |
0381411e | 1039 | .pdata_size = sizeof(struct throtl_grp), |
60c2bc2d | 1040 | .cftypes = throtl_files, |
e43473b7 VG |
1041 | }; |
1042 | ||
bc16a4f9 | 1043 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1044 | { |
1045 | struct throtl_data *td = q->td; | |
1046 | struct throtl_grp *tg; | |
e43473b7 | 1047 | bool rw = bio_data_dir(bio), update_disptime = true; |
af75cd3c | 1048 | struct blkio_cgroup *blkcg; |
bc16a4f9 | 1049 | bool throttled = false; |
e43473b7 VG |
1050 | |
1051 | if (bio->bi_rw & REQ_THROTTLED) { | |
1052 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1053 | goto out; |
e43473b7 VG |
1054 | } |
1055 | ||
671058fb TH |
1056 | /* bio_associate_current() needs ioc, try creating */ |
1057 | create_io_context(GFP_ATOMIC, q->node); | |
1058 | ||
af75cd3c VG |
1059 | /* |
1060 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1061 | * basic fields like stats and io rates. If a group has no rules, | |
1062 | * just update the dispatch stats in lockless manner and return. | |
1063 | */ | |
af75cd3c | 1064 | rcu_read_lock(); |
4f85cb96 | 1065 | blkcg = bio_blkio_cgroup(bio); |
cd1604fa | 1066 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1067 | if (tg) { |
af75cd3c | 1068 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1069 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1070 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1071 | goto out_unlock_rcu; |
af75cd3c VG |
1072 | } |
1073 | } | |
af75cd3c VG |
1074 | |
1075 | /* | |
1076 | * Either group has not been allocated yet or it is not an unlimited | |
1077 | * IO group | |
1078 | */ | |
e43473b7 | 1079 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1080 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1081 | if (unlikely(!tg)) |
1082 | goto out_unlock; | |
f469a7b4 | 1083 | |
e43473b7 VG |
1084 | if (tg->nr_queued[rw]) { |
1085 | /* | |
1086 | * There is already another bio queued in same dir. No | |
1087 | * need to update dispatch time. | |
1088 | */ | |
231d704b | 1089 | update_disptime = false; |
e43473b7 | 1090 | goto queue_bio; |
de701c74 | 1091 | |
e43473b7 VG |
1092 | } |
1093 | ||
1094 | /* Bio is with-in rate limit of group */ | |
1095 | if (tg_may_dispatch(td, tg, bio, NULL)) { | |
1096 | throtl_charge_bio(tg, bio); | |
04521db0 VG |
1097 | |
1098 | /* | |
1099 | * We need to trim slice even when bios are not being queued | |
1100 | * otherwise it might happen that a bio is not queued for | |
1101 | * a long time and slice keeps on extending and trim is not | |
1102 | * called for a long time. Now if limits are reduced suddenly | |
1103 | * we take into account all the IO dispatched so far at new | |
1104 | * low rate and * newly queued IO gets a really long dispatch | |
1105 | * time. | |
1106 | * | |
1107 | * So keep on trimming slice even if bio is not queued. | |
1108 | */ | |
1109 | throtl_trim_slice(td, tg, rw); | |
bc16a4f9 | 1110 | goto out_unlock; |
e43473b7 VG |
1111 | } |
1112 | ||
1113 | queue_bio: | |
fd16d263 | 1114 | throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
8e89d13f VG |
1115 | " iodisp=%u iops=%u queued=%d/%d", |
1116 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1117 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1118 | tg->io_disp[rw], tg->iops[rw], |
e43473b7 VG |
1119 | tg->nr_queued[READ], tg->nr_queued[WRITE]); |
1120 | ||
671058fb | 1121 | bio_associate_current(bio); |
e43473b7 | 1122 | throtl_add_bio_tg(q->td, tg, bio); |
bc16a4f9 | 1123 | throttled = true; |
e43473b7 VG |
1124 | |
1125 | if (update_disptime) { | |
1126 | tg_update_disptime(td, tg); | |
1127 | throtl_schedule_next_dispatch(td); | |
1128 | } | |
1129 | ||
bc16a4f9 | 1130 | out_unlock: |
e43473b7 | 1131 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1132 | out_unlock_rcu: |
1133 | rcu_read_unlock(); | |
bc16a4f9 TH |
1134 | out: |
1135 | return throttled; | |
e43473b7 VG |
1136 | } |
1137 | ||
c9a929dd TH |
1138 | /** |
1139 | * blk_throtl_drain - drain throttled bios | |
1140 | * @q: request_queue to drain throttled bios for | |
1141 | * | |
1142 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1143 | */ | |
1144 | void blk_throtl_drain(struct request_queue *q) | |
1145 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1146 | { | |
1147 | struct throtl_data *td = q->td; | |
1148 | struct throtl_rb_root *st = &td->tg_service_tree; | |
1149 | struct throtl_grp *tg; | |
1150 | struct bio_list bl; | |
1151 | struct bio *bio; | |
1152 | ||
334c2b0b | 1153 | WARN_ON_ONCE(!queue_is_locked(q)); |
c9a929dd TH |
1154 | |
1155 | bio_list_init(&bl); | |
1156 | ||
1157 | while ((tg = throtl_rb_first(st))) { | |
1158 | throtl_dequeue_tg(td, tg); | |
1159 | ||
1160 | while ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
1161 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1162 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
1163 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1164 | } | |
1165 | spin_unlock_irq(q->queue_lock); | |
1166 | ||
1167 | while ((bio = bio_list_pop(&bl))) | |
1168 | generic_make_request(bio); | |
1169 | ||
1170 | spin_lock_irq(q->queue_lock); | |
1171 | } | |
1172 | ||
e43473b7 VG |
1173 | int blk_throtl_init(struct request_queue *q) |
1174 | { | |
1175 | struct throtl_data *td; | |
cd1604fa | 1176 | struct blkio_group *blkg; |
e43473b7 VG |
1177 | |
1178 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1179 | if (!td) | |
1180 | return -ENOMEM; | |
1181 | ||
e43473b7 | 1182 | td->tg_service_tree = THROTL_RB_ROOT; |
de701c74 | 1183 | td->limits_changed = false; |
a29a171e | 1184 | INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work); |
e43473b7 | 1185 | |
cd1604fa | 1186 | q->td = td; |
29b12589 | 1187 | td->queue = q; |
02977e4a | 1188 | |
cd1604fa | 1189 | /* alloc and init root group. */ |
f51b802c TH |
1190 | rcu_read_lock(); |
1191 | spin_lock_irq(q->queue_lock); | |
29b12589 | 1192 | |
aaec55a0 | 1193 | blkg = blkg_lookup_create(&blkio_root_cgroup, q, true); |
cd1604fa | 1194 | if (!IS_ERR(blkg)) |
0381411e | 1195 | td->root_tg = blkg_to_tg(blkg); |
e43473b7 | 1196 | |
f51b802c | 1197 | spin_unlock_irq(q->queue_lock); |
e43473b7 VG |
1198 | rcu_read_unlock(); |
1199 | ||
f51b802c TH |
1200 | if (!td->root_tg) { |
1201 | kfree(td); | |
1202 | return -ENOMEM; | |
1203 | } | |
e43473b7 VG |
1204 | return 0; |
1205 | } | |
1206 | ||
1207 | void blk_throtl_exit(struct request_queue *q) | |
1208 | { | |
c875f4d0 | 1209 | BUG_ON(!q->td); |
da527770 | 1210 | throtl_shutdown_wq(q); |
c9a929dd | 1211 | kfree(q->td); |
e43473b7 VG |
1212 | } |
1213 | ||
1214 | static int __init throtl_init(void) | |
1215 | { | |
450adcbe VG |
1216 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1217 | if (!kthrotld_workqueue) | |
1218 | panic("Failed to create kthrotld\n"); | |
1219 | ||
e43473b7 VG |
1220 | blkio_policy_register(&blkio_policy_throtl); |
1221 | return 0; | |
1222 | } | |
1223 | ||
1224 | module_init(throtl_init); |