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 | ||
8a3d2615 TH |
43 | /* Per-cpu group stats */ |
44 | struct tg_stats_cpu { | |
45 | /* total bytes transferred */ | |
46 | struct blkg_rwstat service_bytes; | |
47 | /* total IOs serviced, post merge */ | |
48 | struct blkg_rwstat serviced; | |
49 | }; | |
50 | ||
e43473b7 | 51 | struct throtl_grp { |
e43473b7 VG |
52 | /* active throtl group service_tree member */ |
53 | struct rb_node rb_node; | |
54 | ||
55 | /* | |
56 | * Dispatch time in jiffies. This is the estimated time when group | |
57 | * will unthrottle and is ready to dispatch more bio. It is used as | |
58 | * key to sort active groups in service tree. | |
59 | */ | |
60 | unsigned long disptime; | |
61 | ||
e43473b7 VG |
62 | unsigned int flags; |
63 | ||
64 | /* Two lists for READ and WRITE */ | |
65 | struct bio_list bio_lists[2]; | |
66 | ||
67 | /* Number of queued bios on READ and WRITE lists */ | |
68 | unsigned int nr_queued[2]; | |
69 | ||
70 | /* bytes per second rate limits */ | |
71 | uint64_t bps[2]; | |
72 | ||
8e89d13f VG |
73 | /* IOPS limits */ |
74 | unsigned int iops[2]; | |
75 | ||
e43473b7 VG |
76 | /* Number of bytes disptached in current slice */ |
77 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
78 | /* Number of bio's dispatched in current slice */ |
79 | unsigned int io_disp[2]; | |
e43473b7 VG |
80 | |
81 | /* When did we start a new slice */ | |
82 | unsigned long slice_start[2]; | |
83 | unsigned long slice_end[2]; | |
fe071437 VG |
84 | |
85 | /* Some throttle limits got updated for the group */ | |
6f037937 | 86 | int limits_changed; |
8a3d2615 TH |
87 | |
88 | /* Per cpu stats pointer */ | |
89 | struct tg_stats_cpu __percpu *stats_cpu; | |
90 | ||
91 | /* List of tgs waiting for per cpu stats memory to be allocated */ | |
92 | struct list_head stats_alloc_node; | |
e43473b7 VG |
93 | }; |
94 | ||
95 | struct throtl_data | |
96 | { | |
e43473b7 VG |
97 | /* service tree for active throtl groups */ |
98 | struct throtl_rb_root tg_service_tree; | |
99 | ||
e43473b7 VG |
100 | struct request_queue *queue; |
101 | ||
102 | /* Total Number of queued bios on READ and WRITE lists */ | |
103 | unsigned int nr_queued[2]; | |
104 | ||
105 | /* | |
02977e4a | 106 | * number of total undestroyed groups |
e43473b7 VG |
107 | */ |
108 | unsigned int nr_undestroyed_grps; | |
109 | ||
110 | /* Work for dispatching throttled bios */ | |
111 | struct delayed_work throtl_work; | |
fe071437 | 112 | |
6f037937 | 113 | int limits_changed; |
e43473b7 VG |
114 | }; |
115 | ||
8a3d2615 TH |
116 | /* list and work item to allocate percpu group stats */ |
117 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
118 | static LIST_HEAD(tg_stats_alloc_list); | |
119 | ||
120 | static void tg_stats_alloc_fn(struct work_struct *); | |
121 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
122 | ||
0381411e TH |
123 | static inline struct throtl_grp *blkg_to_tg(struct blkio_group *blkg) |
124 | { | |
125 | return blkg_to_pdata(blkg, &blkio_policy_throtl); | |
126 | } | |
127 | ||
128 | static inline struct blkio_group *tg_to_blkg(struct throtl_grp *tg) | |
129 | { | |
aaec55a0 | 130 | return pdata_to_blkg(tg); |
0381411e TH |
131 | } |
132 | ||
03d8e111 TH |
133 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) |
134 | { | |
135 | return blkg_to_tg(td->queue->root_blkg); | |
136 | } | |
137 | ||
e43473b7 VG |
138 | enum tg_state_flags { |
139 | THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */ | |
140 | }; | |
141 | ||
142 | #define THROTL_TG_FNS(name) \ | |
143 | static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \ | |
144 | { \ | |
145 | (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \ | |
146 | } \ | |
147 | static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \ | |
148 | { \ | |
149 | (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \ | |
150 | } \ | |
151 | static inline int throtl_tg_##name(const struct throtl_grp *tg) \ | |
152 | { \ | |
153 | return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \ | |
154 | } | |
155 | ||
156 | THROTL_TG_FNS(on_rr); | |
157 | ||
158 | #define throtl_log_tg(td, tg, fmt, args...) \ | |
159 | blk_add_trace_msg((td)->queue, "throtl %s " fmt, \ | |
0381411e | 160 | blkg_path(tg_to_blkg(tg)), ##args); \ |
e43473b7 VG |
161 | |
162 | #define throtl_log(td, fmt, args...) \ | |
163 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) | |
164 | ||
d2f31a5f | 165 | static inline unsigned int total_nr_queued(struct throtl_data *td) |
e43473b7 | 166 | { |
d2f31a5f | 167 | return td->nr_queued[0] + td->nr_queued[1]; |
e43473b7 VG |
168 | } |
169 | ||
8a3d2615 TH |
170 | /* |
171 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
172 | * system_nrt_wq once there are some groups on the alloc_list waiting for | |
173 | * allocation. | |
174 | */ | |
175 | static void tg_stats_alloc_fn(struct work_struct *work) | |
176 | { | |
177 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
178 | struct delayed_work *dwork = to_delayed_work(work); | |
179 | bool empty = false; | |
180 | ||
181 | alloc_stats: | |
182 | if (!stats_cpu) { | |
183 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
184 | if (!stats_cpu) { | |
185 | /* allocation failed, try again after some time */ | |
186 | queue_delayed_work(system_nrt_wq, dwork, | |
187 | msecs_to_jiffies(10)); | |
188 | return; | |
189 | } | |
190 | } | |
191 | ||
192 | spin_lock_irq(&tg_stats_alloc_lock); | |
193 | ||
194 | if (!list_empty(&tg_stats_alloc_list)) { | |
195 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
196 | struct throtl_grp, | |
197 | stats_alloc_node); | |
198 | swap(tg->stats_cpu, stats_cpu); | |
199 | list_del_init(&tg->stats_alloc_node); | |
200 | } | |
201 | ||
202 | empty = list_empty(&tg_stats_alloc_list); | |
203 | spin_unlock_irq(&tg_stats_alloc_lock); | |
204 | if (!empty) | |
205 | goto alloc_stats; | |
206 | } | |
207 | ||
0381411e | 208 | static void throtl_init_blkio_group(struct blkio_group *blkg) |
a29a171e | 209 | { |
0381411e | 210 | struct throtl_grp *tg = blkg_to_tg(blkg); |
cd1604fa | 211 | |
a29a171e VG |
212 | RB_CLEAR_NODE(&tg->rb_node); |
213 | bio_list_init(&tg->bio_lists[0]); | |
214 | bio_list_init(&tg->bio_lists[1]); | |
215 | tg->limits_changed = false; | |
216 | ||
e56da7e2 TH |
217 | tg->bps[READ] = -1; |
218 | tg->bps[WRITE] = -1; | |
219 | tg->iops[READ] = -1; | |
220 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
221 | |
222 | /* | |
223 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
224 | * but percpu allocator can't be called from IO path. Queue tg on | |
225 | * tg_stats_alloc_list and allocate from work item. | |
226 | */ | |
227 | spin_lock(&tg_stats_alloc_lock); | |
228 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); | |
229 | queue_delayed_work(system_nrt_wq, &tg_stats_alloc_work, 0); | |
230 | spin_unlock(&tg_stats_alloc_lock); | |
231 | } | |
232 | ||
233 | static void throtl_exit_blkio_group(struct blkio_group *blkg) | |
234 | { | |
235 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
236 | ||
237 | spin_lock(&tg_stats_alloc_lock); | |
238 | list_del_init(&tg->stats_alloc_node); | |
239 | spin_unlock(&tg_stats_alloc_lock); | |
240 | ||
241 | free_percpu(tg->stats_cpu); | |
242 | } | |
243 | ||
244 | static void throtl_reset_group_stats(struct blkio_group *blkg) | |
245 | { | |
246 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
247 | int cpu; | |
248 | ||
249 | if (tg->stats_cpu == NULL) | |
250 | return; | |
251 | ||
252 | for_each_possible_cpu(cpu) { | |
253 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
254 | ||
255 | blkg_rwstat_reset(&sc->service_bytes); | |
256 | blkg_rwstat_reset(&sc->serviced); | |
257 | } | |
a29a171e VG |
258 | } |
259 | ||
f469a7b4 | 260 | static struct |
cd1604fa | 261 | throtl_grp *throtl_lookup_tg(struct throtl_data *td, struct blkio_cgroup *blkcg) |
e43473b7 | 262 | { |
be2c6b19 VG |
263 | /* |
264 | * This is the common case when there are no blkio cgroups. | |
cd1604fa TH |
265 | * Avoid lookup in this case |
266 | */ | |
be2c6b19 | 267 | if (blkcg == &blkio_root_cgroup) |
03d8e111 | 268 | return td_root_tg(td); |
e43473b7 | 269 | |
e8989fae | 270 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
271 | } |
272 | ||
cd1604fa TH |
273 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
274 | struct blkio_cgroup *blkcg) | |
e43473b7 | 275 | { |
f469a7b4 | 276 | struct request_queue *q = td->queue; |
cd1604fa | 277 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 278 | |
f469a7b4 | 279 | /* |
cd1604fa TH |
280 | * This is the common case when there are no blkio cgroups. |
281 | * Avoid lookup in this case | |
f469a7b4 | 282 | */ |
cd1604fa | 283 | if (blkcg == &blkio_root_cgroup) { |
03d8e111 | 284 | tg = td_root_tg(td); |
cd1604fa TH |
285 | } else { |
286 | struct blkio_group *blkg; | |
f469a7b4 | 287 | |
aaec55a0 | 288 | blkg = blkg_lookup_create(blkcg, q, false); |
f469a7b4 | 289 | |
cd1604fa TH |
290 | /* if %NULL and @q is alive, fall back to root_tg */ |
291 | if (!IS_ERR(blkg)) | |
0381411e | 292 | tg = blkg_to_tg(blkg); |
cd1604fa | 293 | else if (!blk_queue_dead(q)) |
03d8e111 | 294 | tg = td_root_tg(td); |
f469a7b4 VG |
295 | } |
296 | ||
e43473b7 VG |
297 | return tg; |
298 | } | |
299 | ||
300 | static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root) | |
301 | { | |
302 | /* Service tree is empty */ | |
303 | if (!root->count) | |
304 | return NULL; | |
305 | ||
306 | if (!root->left) | |
307 | root->left = rb_first(&root->rb); | |
308 | ||
309 | if (root->left) | |
310 | return rb_entry_tg(root->left); | |
311 | ||
312 | return NULL; | |
313 | } | |
314 | ||
315 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
316 | { | |
317 | rb_erase(n, root); | |
318 | RB_CLEAR_NODE(n); | |
319 | } | |
320 | ||
321 | static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root) | |
322 | { | |
323 | if (root->left == n) | |
324 | root->left = NULL; | |
325 | rb_erase_init(n, &root->rb); | |
326 | --root->count; | |
327 | } | |
328 | ||
329 | static void update_min_dispatch_time(struct throtl_rb_root *st) | |
330 | { | |
331 | struct throtl_grp *tg; | |
332 | ||
333 | tg = throtl_rb_first(st); | |
334 | if (!tg) | |
335 | return; | |
336 | ||
337 | st->min_disptime = tg->disptime; | |
338 | } | |
339 | ||
340 | static void | |
341 | tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg) | |
342 | { | |
343 | struct rb_node **node = &st->rb.rb_node; | |
344 | struct rb_node *parent = NULL; | |
345 | struct throtl_grp *__tg; | |
346 | unsigned long key = tg->disptime; | |
347 | int left = 1; | |
348 | ||
349 | while (*node != NULL) { | |
350 | parent = *node; | |
351 | __tg = rb_entry_tg(parent); | |
352 | ||
353 | if (time_before(key, __tg->disptime)) | |
354 | node = &parent->rb_left; | |
355 | else { | |
356 | node = &parent->rb_right; | |
357 | left = 0; | |
358 | } | |
359 | } | |
360 | ||
361 | if (left) | |
362 | st->left = &tg->rb_node; | |
363 | ||
364 | rb_link_node(&tg->rb_node, parent, node); | |
365 | rb_insert_color(&tg->rb_node, &st->rb); | |
366 | } | |
367 | ||
368 | static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
369 | { | |
370 | struct throtl_rb_root *st = &td->tg_service_tree; | |
371 | ||
372 | tg_service_tree_add(st, tg); | |
373 | throtl_mark_tg_on_rr(tg); | |
374 | st->count++; | |
375 | } | |
376 | ||
377 | static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
378 | { | |
379 | if (!throtl_tg_on_rr(tg)) | |
380 | __throtl_enqueue_tg(td, tg); | |
381 | } | |
382 | ||
383 | static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
384 | { | |
385 | throtl_rb_erase(&tg->rb_node, &td->tg_service_tree); | |
386 | throtl_clear_tg_on_rr(tg); | |
387 | } | |
388 | ||
389 | static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
390 | { | |
391 | if (throtl_tg_on_rr(tg)) | |
392 | __throtl_dequeue_tg(td, tg); | |
393 | } | |
394 | ||
395 | static void throtl_schedule_next_dispatch(struct throtl_data *td) | |
396 | { | |
397 | struct throtl_rb_root *st = &td->tg_service_tree; | |
398 | ||
399 | /* | |
400 | * If there are more bios pending, schedule more work. | |
401 | */ | |
402 | if (!total_nr_queued(td)) | |
403 | return; | |
404 | ||
405 | BUG_ON(!st->count); | |
406 | ||
407 | update_min_dispatch_time(st); | |
408 | ||
409 | if (time_before_eq(st->min_disptime, jiffies)) | |
450adcbe | 410 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 411 | else |
450adcbe | 412 | throtl_schedule_delayed_work(td, (st->min_disptime - jiffies)); |
e43473b7 VG |
413 | } |
414 | ||
415 | static inline void | |
416 | throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
417 | { | |
418 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 419 | tg->io_disp[rw] = 0; |
e43473b7 VG |
420 | tg->slice_start[rw] = jiffies; |
421 | tg->slice_end[rw] = jiffies + throtl_slice; | |
422 | throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
423 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
424 | tg->slice_end[rw], jiffies); | |
425 | } | |
426 | ||
d1ae8ffd VG |
427 | static inline void throtl_set_slice_end(struct throtl_data *td, |
428 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
429 | { | |
430 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
431 | } | |
432 | ||
e43473b7 VG |
433 | static inline void throtl_extend_slice(struct throtl_data *td, |
434 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
435 | { | |
436 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
437 | throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
438 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
439 | tg->slice_end[rw], jiffies); | |
440 | } | |
441 | ||
442 | /* Determine if previously allocated or extended slice is complete or not */ | |
443 | static bool | |
444 | throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
445 | { | |
446 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
447 | return 0; | |
448 | ||
449 | return 1; | |
450 | } | |
451 | ||
452 | /* Trim the used slices and adjust slice start accordingly */ | |
453 | static inline void | |
454 | throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
455 | { | |
3aad5d3e VG |
456 | unsigned long nr_slices, time_elapsed, io_trim; |
457 | u64 bytes_trim, tmp; | |
e43473b7 VG |
458 | |
459 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
460 | ||
461 | /* | |
462 | * If bps are unlimited (-1), then time slice don't get | |
463 | * renewed. Don't try to trim the slice if slice is used. A new | |
464 | * slice will start when appropriate. | |
465 | */ | |
466 | if (throtl_slice_used(td, tg, rw)) | |
467 | return; | |
468 | ||
d1ae8ffd VG |
469 | /* |
470 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
471 | * that initially cgroup limit was very low resulting in high | |
472 | * slice_end, but later limit was bumped up and bio was dispached | |
473 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
474 | * is bad because it does not allow new slice to start. | |
475 | */ | |
476 | ||
477 | throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice); | |
478 | ||
e43473b7 VG |
479 | time_elapsed = jiffies - tg->slice_start[rw]; |
480 | ||
481 | nr_slices = time_elapsed / throtl_slice; | |
482 | ||
483 | if (!nr_slices) | |
484 | return; | |
3aad5d3e VG |
485 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
486 | do_div(tmp, HZ); | |
487 | bytes_trim = tmp; | |
e43473b7 | 488 | |
8e89d13f | 489 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 490 | |
8e89d13f | 491 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
492 | return; |
493 | ||
494 | if (tg->bytes_disp[rw] >= bytes_trim) | |
495 | tg->bytes_disp[rw] -= bytes_trim; | |
496 | else | |
497 | tg->bytes_disp[rw] = 0; | |
498 | ||
8e89d13f VG |
499 | if (tg->io_disp[rw] >= io_trim) |
500 | tg->io_disp[rw] -= io_trim; | |
501 | else | |
502 | tg->io_disp[rw] = 0; | |
503 | ||
e43473b7 VG |
504 | tg->slice_start[rw] += nr_slices * throtl_slice; |
505 | ||
3aad5d3e | 506 | throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 507 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 508 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
509 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
510 | } | |
511 | ||
8e89d13f VG |
512 | static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg, |
513 | struct bio *bio, unsigned long *wait) | |
e43473b7 VG |
514 | { |
515 | bool rw = bio_data_dir(bio); | |
8e89d13f | 516 | unsigned int io_allowed; |
e43473b7 | 517 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 518 | u64 tmp; |
e43473b7 | 519 | |
8e89d13f | 520 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 521 | |
8e89d13f VG |
522 | /* Slice has just started. Consider one slice interval */ |
523 | if (!jiffy_elapsed) | |
524 | jiffy_elapsed_rnd = throtl_slice; | |
525 | ||
526 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
527 | ||
c49c06e4 VG |
528 | /* |
529 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
530 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
531 | * will allow dispatch after 1 second and after that slice should | |
532 | * have been trimmed. | |
533 | */ | |
534 | ||
535 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
536 | do_div(tmp, HZ); | |
537 | ||
538 | if (tmp > UINT_MAX) | |
539 | io_allowed = UINT_MAX; | |
540 | else | |
541 | io_allowed = tmp; | |
8e89d13f VG |
542 | |
543 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
544 | if (wait) |
545 | *wait = 0; | |
546 | return 1; | |
547 | } | |
548 | ||
8e89d13f VG |
549 | /* Calc approx time to dispatch */ |
550 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
551 | ||
552 | if (jiffy_wait > jiffy_elapsed) | |
553 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
554 | else | |
555 | jiffy_wait = 1; | |
556 | ||
557 | if (wait) | |
558 | *wait = jiffy_wait; | |
559 | return 0; | |
560 | } | |
561 | ||
562 | static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg, | |
563 | struct bio *bio, unsigned long *wait) | |
564 | { | |
565 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 566 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 567 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
568 | |
569 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
570 | ||
571 | /* Slice has just started. Consider one slice interval */ | |
572 | if (!jiffy_elapsed) | |
573 | jiffy_elapsed_rnd = throtl_slice; | |
574 | ||
575 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
576 | ||
5e901a2b VG |
577 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
578 | do_div(tmp, HZ); | |
3aad5d3e | 579 | bytes_allowed = tmp; |
e43473b7 VG |
580 | |
581 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
582 | if (wait) | |
583 | *wait = 0; | |
584 | return 1; | |
585 | } | |
586 | ||
587 | /* Calc approx time to dispatch */ | |
588 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
589 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
590 | ||
591 | if (!jiffy_wait) | |
592 | jiffy_wait = 1; | |
593 | ||
594 | /* | |
595 | * This wait time is without taking into consideration the rounding | |
596 | * up we did. Add that time also. | |
597 | */ | |
598 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
599 | if (wait) |
600 | *wait = jiffy_wait; | |
8e89d13f VG |
601 | return 0; |
602 | } | |
603 | ||
af75cd3c VG |
604 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
605 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
606 | return 1; | |
607 | return 0; | |
608 | } | |
609 | ||
8e89d13f VG |
610 | /* |
611 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
612 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
613 | */ | |
614 | static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg, | |
615 | struct bio *bio, unsigned long *wait) | |
616 | { | |
617 | bool rw = bio_data_dir(bio); | |
618 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
619 | ||
620 | /* | |
621 | * Currently whole state machine of group depends on first bio | |
622 | * queued in the group bio list. So one should not be calling | |
623 | * this function with a different bio if there are other bios | |
624 | * queued. | |
625 | */ | |
626 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); | |
e43473b7 | 627 | |
8e89d13f VG |
628 | /* If tg->bps = -1, then BW is unlimited */ |
629 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
630 | if (wait) | |
631 | *wait = 0; | |
632 | return 1; | |
633 | } | |
634 | ||
635 | /* | |
636 | * If previous slice expired, start a new one otherwise renew/extend | |
637 | * existing slice to make sure it is at least throtl_slice interval | |
638 | * long since now. | |
639 | */ | |
640 | if (throtl_slice_used(td, tg, rw)) | |
641 | throtl_start_new_slice(td, tg, rw); | |
642 | else { | |
643 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
644 | throtl_extend_slice(td, tg, rw, jiffies + throtl_slice); | |
645 | } | |
646 | ||
647 | if (tg_with_in_bps_limit(td, tg, bio, &bps_wait) | |
648 | && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) { | |
649 | if (wait) | |
650 | *wait = 0; | |
651 | return 1; | |
652 | } | |
653 | ||
654 | max_wait = max(bps_wait, iops_wait); | |
655 | ||
656 | if (wait) | |
657 | *wait = max_wait; | |
658 | ||
659 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
660 | throtl_extend_slice(td, tg, rw, jiffies + max_wait); | |
e43473b7 VG |
661 | |
662 | return 0; | |
663 | } | |
664 | ||
629ed0b1 TH |
665 | static void throtl_update_dispatch_stats(struct blkio_group *blkg, u64 bytes, |
666 | int rw) | |
667 | { | |
8a3d2615 TH |
668 | struct throtl_grp *tg = blkg_to_tg(blkg); |
669 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
670 | unsigned long flags; |
671 | ||
672 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 673 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
674 | return; |
675 | ||
676 | /* | |
677 | * Disabling interrupts to provide mutual exclusion between two | |
678 | * writes on same cpu. It probably is not needed for 64bit. Not | |
679 | * optimizing that case yet. | |
680 | */ | |
681 | local_irq_save(flags); | |
682 | ||
8a3d2615 | 683 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 684 | |
629ed0b1 TH |
685 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
686 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
687 | ||
688 | local_irq_restore(flags); | |
689 | } | |
690 | ||
e43473b7 VG |
691 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
692 | { | |
693 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
694 | |
695 | /* Charge the bio to the group */ | |
696 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 697 | tg->io_disp[rw]++; |
e43473b7 | 698 | |
629ed0b1 | 699 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw); |
e43473b7 VG |
700 | } |
701 | ||
702 | static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg, | |
703 | struct bio *bio) | |
704 | { | |
705 | bool rw = bio_data_dir(bio); | |
706 | ||
707 | bio_list_add(&tg->bio_lists[rw], bio); | |
708 | /* Take a bio reference on tg */ | |
1adaf3dd | 709 | blkg_get(tg_to_blkg(tg)); |
e43473b7 VG |
710 | tg->nr_queued[rw]++; |
711 | td->nr_queued[rw]++; | |
712 | throtl_enqueue_tg(td, tg); | |
713 | } | |
714 | ||
715 | static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg) | |
716 | { | |
717 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
718 | struct bio *bio; | |
719 | ||
720 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
721 | tg_may_dispatch(td, tg, bio, &read_wait); | |
722 | ||
723 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
724 | tg_may_dispatch(td, tg, bio, &write_wait); | |
725 | ||
726 | min_wait = min(read_wait, write_wait); | |
727 | disptime = jiffies + min_wait; | |
728 | ||
e43473b7 VG |
729 | /* Update dispatch time */ |
730 | throtl_dequeue_tg(td, tg); | |
731 | tg->disptime = disptime; | |
732 | throtl_enqueue_tg(td, tg); | |
733 | } | |
734 | ||
735 | static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg, | |
736 | bool rw, struct bio_list *bl) | |
737 | { | |
738 | struct bio *bio; | |
739 | ||
740 | bio = bio_list_pop(&tg->bio_lists[rw]); | |
741 | tg->nr_queued[rw]--; | |
1adaf3dd TH |
742 | /* Drop bio reference on blkg */ |
743 | blkg_put(tg_to_blkg(tg)); | |
e43473b7 VG |
744 | |
745 | BUG_ON(td->nr_queued[rw] <= 0); | |
746 | td->nr_queued[rw]--; | |
747 | ||
748 | throtl_charge_bio(tg, bio); | |
749 | bio_list_add(bl, bio); | |
750 | bio->bi_rw |= REQ_THROTTLED; | |
751 | ||
752 | throtl_trim_slice(td, tg, rw); | |
753 | } | |
754 | ||
755 | static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg, | |
756 | struct bio_list *bl) | |
757 | { | |
758 | unsigned int nr_reads = 0, nr_writes = 0; | |
759 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 760 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
761 | struct bio *bio; |
762 | ||
763 | /* Try to dispatch 75% READS and 25% WRITES */ | |
764 | ||
765 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) | |
766 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
767 | ||
768 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
769 | nr_reads++; | |
770 | ||
771 | if (nr_reads >= max_nr_reads) | |
772 | break; | |
773 | } | |
774 | ||
775 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) | |
776 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
777 | ||
778 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
779 | nr_writes++; | |
780 | ||
781 | if (nr_writes >= max_nr_writes) | |
782 | break; | |
783 | } | |
784 | ||
785 | return nr_reads + nr_writes; | |
786 | } | |
787 | ||
788 | static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl) | |
789 | { | |
790 | unsigned int nr_disp = 0; | |
791 | struct throtl_grp *tg; | |
792 | struct throtl_rb_root *st = &td->tg_service_tree; | |
793 | ||
794 | while (1) { | |
795 | tg = throtl_rb_first(st); | |
796 | ||
797 | if (!tg) | |
798 | break; | |
799 | ||
800 | if (time_before(jiffies, tg->disptime)) | |
801 | break; | |
802 | ||
803 | throtl_dequeue_tg(td, tg); | |
804 | ||
805 | nr_disp += throtl_dispatch_tg(td, tg, bl); | |
806 | ||
807 | if (tg->nr_queued[0] || tg->nr_queued[1]) { | |
808 | tg_update_disptime(td, tg); | |
809 | throtl_enqueue_tg(td, tg); | |
810 | } | |
811 | ||
812 | if (nr_disp >= throtl_quantum) | |
813 | break; | |
814 | } | |
815 | ||
816 | return nr_disp; | |
817 | } | |
818 | ||
fe071437 VG |
819 | static void throtl_process_limit_change(struct throtl_data *td) |
820 | { | |
4eef3049 TH |
821 | struct request_queue *q = td->queue; |
822 | struct blkio_group *blkg, *n; | |
fe071437 | 823 | |
de701c74 | 824 | if (!td->limits_changed) |
fe071437 VG |
825 | return; |
826 | ||
de701c74 | 827 | xchg(&td->limits_changed, false); |
fe071437 | 828 | |
de701c74 | 829 | throtl_log(td, "limits changed"); |
fe071437 | 830 | |
e8989fae | 831 | list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) { |
4eef3049 TH |
832 | struct throtl_grp *tg = blkg_to_tg(blkg); |
833 | ||
de701c74 VG |
834 | if (!tg->limits_changed) |
835 | continue; | |
836 | ||
837 | if (!xchg(&tg->limits_changed, false)) | |
838 | continue; | |
839 | ||
840 | throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu" | |
841 | " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE], | |
842 | tg->iops[READ], tg->iops[WRITE]); | |
843 | ||
04521db0 VG |
844 | /* |
845 | * Restart the slices for both READ and WRITES. It | |
846 | * might happen that a group's limit are dropped | |
847 | * suddenly and we don't want to account recently | |
848 | * dispatched IO with new low rate | |
849 | */ | |
850 | throtl_start_new_slice(td, tg, 0); | |
851 | throtl_start_new_slice(td, tg, 1); | |
852 | ||
de701c74 | 853 | if (throtl_tg_on_rr(tg)) |
fe071437 | 854 | tg_update_disptime(td, tg); |
fe071437 | 855 | } |
fe071437 VG |
856 | } |
857 | ||
e43473b7 VG |
858 | /* Dispatch throttled bios. Should be called without queue lock held. */ |
859 | static int throtl_dispatch(struct request_queue *q) | |
860 | { | |
861 | struct throtl_data *td = q->td; | |
862 | unsigned int nr_disp = 0; | |
863 | struct bio_list bio_list_on_stack; | |
864 | struct bio *bio; | |
69d60eb9 | 865 | struct blk_plug plug; |
e43473b7 VG |
866 | |
867 | spin_lock_irq(q->queue_lock); | |
868 | ||
fe071437 VG |
869 | throtl_process_limit_change(td); |
870 | ||
e43473b7 VG |
871 | if (!total_nr_queued(td)) |
872 | goto out; | |
873 | ||
874 | bio_list_init(&bio_list_on_stack); | |
875 | ||
d2f31a5f | 876 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
e43473b7 VG |
877 | total_nr_queued(td), td->nr_queued[READ], |
878 | td->nr_queued[WRITE]); | |
879 | ||
880 | nr_disp = throtl_select_dispatch(td, &bio_list_on_stack); | |
881 | ||
882 | if (nr_disp) | |
883 | throtl_log(td, "bios disp=%u", nr_disp); | |
884 | ||
885 | throtl_schedule_next_dispatch(td); | |
886 | out: | |
887 | spin_unlock_irq(q->queue_lock); | |
888 | ||
889 | /* | |
890 | * If we dispatched some requests, unplug the queue to make sure | |
891 | * immediate dispatch | |
892 | */ | |
893 | if (nr_disp) { | |
69d60eb9 | 894 | blk_start_plug(&plug); |
e43473b7 VG |
895 | while((bio = bio_list_pop(&bio_list_on_stack))) |
896 | generic_make_request(bio); | |
69d60eb9 | 897 | blk_finish_plug(&plug); |
e43473b7 VG |
898 | } |
899 | return nr_disp; | |
900 | } | |
901 | ||
902 | void blk_throtl_work(struct work_struct *work) | |
903 | { | |
904 | struct throtl_data *td = container_of(work, struct throtl_data, | |
905 | throtl_work.work); | |
906 | struct request_queue *q = td->queue; | |
907 | ||
908 | throtl_dispatch(q); | |
909 | } | |
910 | ||
911 | /* Call with queue lock held */ | |
450adcbe VG |
912 | static void |
913 | throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay) | |
e43473b7 VG |
914 | { |
915 | ||
e43473b7 VG |
916 | struct delayed_work *dwork = &td->throtl_work; |
917 | ||
04521db0 | 918 | /* schedule work if limits changed even if no bio is queued */ |
d2f31a5f | 919 | if (total_nr_queued(td) || td->limits_changed) { |
e43473b7 VG |
920 | /* |
921 | * We might have a work scheduled to be executed in future. | |
922 | * Cancel that and schedule a new one. | |
923 | */ | |
924 | __cancel_delayed_work(dwork); | |
450adcbe | 925 | queue_delayed_work(kthrotld_workqueue, dwork, delay); |
e43473b7 VG |
926 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", |
927 | delay, jiffies); | |
928 | } | |
929 | } | |
e43473b7 | 930 | |
d366e7ec | 931 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, void *pdata, int off) |
41b38b6d | 932 | { |
d366e7ec | 933 | struct throtl_grp *tg = pdata; |
41b38b6d TH |
934 | struct blkg_rwstat rwstat = { }, tmp; |
935 | int i, cpu; | |
936 | ||
937 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 938 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
939 | |
940 | tmp = blkg_rwstat_read((void *)sc + off); | |
941 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
942 | rwstat.cnt[i] += tmp.cnt[i]; | |
943 | } | |
944 | ||
d366e7ec | 945 | return __blkg_prfill_rwstat(sf, pdata, &rwstat); |
41b38b6d TH |
946 | } |
947 | ||
8a3d2615 TH |
948 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
949 | struct seq_file *sf) | |
41b38b6d TH |
950 | { |
951 | struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgrp); | |
952 | ||
ec399347 | 953 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkio_policy_throtl, |
5bc4afb1 | 954 | cft->private, true); |
41b38b6d TH |
955 | return 0; |
956 | } | |
957 | ||
d366e7ec | 958 | static u64 tg_prfill_conf_u64(struct seq_file *sf, void *pdata, int off) |
60c2bc2d | 959 | { |
d366e7ec | 960 | u64 v = *(u64 *)(pdata + off); |
60c2bc2d | 961 | |
af133ceb | 962 | if (v == -1) |
60c2bc2d | 963 | return 0; |
d366e7ec | 964 | return __blkg_prfill_u64(sf, pdata, v); |
60c2bc2d TH |
965 | } |
966 | ||
d366e7ec | 967 | static u64 tg_prfill_conf_uint(struct seq_file *sf, void *pdata, int off) |
e43473b7 | 968 | { |
d366e7ec | 969 | unsigned int v = *(unsigned int *)(pdata + off); |
fe071437 | 970 | |
af133ceb TH |
971 | if (v == -1) |
972 | return 0; | |
d366e7ec | 973 | return __blkg_prfill_u64(sf, pdata, v); |
e43473b7 VG |
974 | } |
975 | ||
af133ceb TH |
976 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
977 | struct seq_file *sf) | |
8e89d13f | 978 | { |
af133ceb | 979 | blkcg_print_blkgs(sf, cgroup_to_blkio_cgroup(cgrp), tg_prfill_conf_u64, |
ec399347 | 980 | &blkio_policy_throtl, cft->private, false); |
af133ceb | 981 | return 0; |
8e89d13f VG |
982 | } |
983 | ||
af133ceb TH |
984 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
985 | struct seq_file *sf) | |
8e89d13f | 986 | { |
af133ceb | 987 | blkcg_print_blkgs(sf, cgroup_to_blkio_cgroup(cgrp), tg_prfill_conf_uint, |
ec399347 | 988 | &blkio_policy_throtl, cft->private, false); |
af133ceb | 989 | return 0; |
60c2bc2d TH |
990 | } |
991 | ||
af133ceb TH |
992 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
993 | bool is_u64) | |
60c2bc2d TH |
994 | { |
995 | struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgrp); | |
60c2bc2d | 996 | struct blkg_conf_ctx ctx; |
af133ceb | 997 | struct throtl_grp *tg; |
a2b1693b | 998 | struct throtl_data *td; |
60c2bc2d TH |
999 | int ret; |
1000 | ||
da8b0662 | 1001 | ret = blkg_conf_prep(blkcg, &blkio_policy_throtl, buf, &ctx); |
60c2bc2d TH |
1002 | if (ret) |
1003 | return ret; | |
1004 | ||
af133ceb | 1005 | tg = blkg_to_tg(ctx.blkg); |
a2b1693b | 1006 | td = ctx.blkg->q->td; |
af133ceb | 1007 | |
a2b1693b TH |
1008 | if (!ctx.v) |
1009 | ctx.v = -1; | |
af133ceb | 1010 | |
a2b1693b TH |
1011 | if (is_u64) |
1012 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
1013 | else | |
1014 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 1015 | |
a2b1693b TH |
1016 | /* XXX: we don't need the following deferred processing */ |
1017 | xchg(&tg->limits_changed, true); | |
1018 | xchg(&td->limits_changed, true); | |
1019 | throtl_schedule_delayed_work(td, 0); | |
60c2bc2d TH |
1020 | |
1021 | blkg_conf_finish(&ctx); | |
a2b1693b | 1022 | return 0; |
8e89d13f VG |
1023 | } |
1024 | ||
af133ceb TH |
1025 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1026 | const char *buf) | |
60c2bc2d | 1027 | { |
af133ceb | 1028 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
1029 | } |
1030 | ||
af133ceb TH |
1031 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1032 | const char *buf) | |
60c2bc2d | 1033 | { |
af133ceb | 1034 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
1035 | } |
1036 | ||
1037 | static struct cftype throtl_files[] = { | |
1038 | { | |
1039 | .name = "throttle.read_bps_device", | |
af133ceb TH |
1040 | .private = offsetof(struct throtl_grp, bps[READ]), |
1041 | .read_seq_string = tg_print_conf_u64, | |
1042 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1043 | .max_write_len = 256, |
1044 | }, | |
1045 | { | |
1046 | .name = "throttle.write_bps_device", | |
af133ceb TH |
1047 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
1048 | .read_seq_string = tg_print_conf_u64, | |
1049 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1050 | .max_write_len = 256, |
1051 | }, | |
1052 | { | |
1053 | .name = "throttle.read_iops_device", | |
af133ceb TH |
1054 | .private = offsetof(struct throtl_grp, iops[READ]), |
1055 | .read_seq_string = tg_print_conf_uint, | |
1056 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1057 | .max_write_len = 256, |
1058 | }, | |
1059 | { | |
1060 | .name = "throttle.write_iops_device", | |
af133ceb TH |
1061 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
1062 | .read_seq_string = tg_print_conf_uint, | |
1063 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1064 | .max_write_len = 256, |
1065 | }, | |
1066 | { | |
1067 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1068 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1069 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1070 | }, |
1071 | { | |
1072 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1073 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1074 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1075 | }, |
1076 | { } /* terminate */ | |
1077 | }; | |
1078 | ||
da527770 | 1079 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1080 | { |
1081 | struct throtl_data *td = q->td; | |
1082 | ||
1083 | cancel_delayed_work_sync(&td->throtl_work); | |
1084 | } | |
1085 | ||
1086 | static struct blkio_policy_type blkio_policy_throtl = { | |
1087 | .ops = { | |
0381411e | 1088 | .blkio_init_group_fn = throtl_init_blkio_group, |
8a3d2615 TH |
1089 | .blkio_exit_group_fn = throtl_exit_blkio_group, |
1090 | .blkio_reset_group_stats_fn = throtl_reset_group_stats, | |
e43473b7 | 1091 | }, |
0381411e | 1092 | .pdata_size = sizeof(struct throtl_grp), |
60c2bc2d | 1093 | .cftypes = throtl_files, |
e43473b7 VG |
1094 | }; |
1095 | ||
bc16a4f9 | 1096 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1097 | { |
1098 | struct throtl_data *td = q->td; | |
1099 | struct throtl_grp *tg; | |
e43473b7 | 1100 | bool rw = bio_data_dir(bio), update_disptime = true; |
af75cd3c | 1101 | struct blkio_cgroup *blkcg; |
bc16a4f9 | 1102 | bool throttled = false; |
e43473b7 VG |
1103 | |
1104 | if (bio->bi_rw & REQ_THROTTLED) { | |
1105 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1106 | goto out; |
e43473b7 VG |
1107 | } |
1108 | ||
671058fb TH |
1109 | /* bio_associate_current() needs ioc, try creating */ |
1110 | create_io_context(GFP_ATOMIC, q->node); | |
1111 | ||
af75cd3c VG |
1112 | /* |
1113 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1114 | * basic fields like stats and io rates. If a group has no rules, | |
1115 | * just update the dispatch stats in lockless manner and return. | |
1116 | */ | |
af75cd3c | 1117 | rcu_read_lock(); |
4f85cb96 | 1118 | blkcg = bio_blkio_cgroup(bio); |
cd1604fa | 1119 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1120 | if (tg) { |
af75cd3c | 1121 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1122 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1123 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1124 | goto out_unlock_rcu; |
af75cd3c VG |
1125 | } |
1126 | } | |
af75cd3c VG |
1127 | |
1128 | /* | |
1129 | * Either group has not been allocated yet or it is not an unlimited | |
1130 | * IO group | |
1131 | */ | |
e43473b7 | 1132 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1133 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1134 | if (unlikely(!tg)) |
1135 | goto out_unlock; | |
f469a7b4 | 1136 | |
e43473b7 VG |
1137 | if (tg->nr_queued[rw]) { |
1138 | /* | |
1139 | * There is already another bio queued in same dir. No | |
1140 | * need to update dispatch time. | |
1141 | */ | |
231d704b | 1142 | update_disptime = false; |
e43473b7 | 1143 | goto queue_bio; |
de701c74 | 1144 | |
e43473b7 VG |
1145 | } |
1146 | ||
1147 | /* Bio is with-in rate limit of group */ | |
1148 | if (tg_may_dispatch(td, tg, bio, NULL)) { | |
1149 | throtl_charge_bio(tg, bio); | |
04521db0 VG |
1150 | |
1151 | /* | |
1152 | * We need to trim slice even when bios are not being queued | |
1153 | * otherwise it might happen that a bio is not queued for | |
1154 | * a long time and slice keeps on extending and trim is not | |
1155 | * called for a long time. Now if limits are reduced suddenly | |
1156 | * we take into account all the IO dispatched so far at new | |
1157 | * low rate and * newly queued IO gets a really long dispatch | |
1158 | * time. | |
1159 | * | |
1160 | * So keep on trimming slice even if bio is not queued. | |
1161 | */ | |
1162 | throtl_trim_slice(td, tg, rw); | |
bc16a4f9 | 1163 | goto out_unlock; |
e43473b7 VG |
1164 | } |
1165 | ||
1166 | queue_bio: | |
fd16d263 | 1167 | throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
8e89d13f VG |
1168 | " iodisp=%u iops=%u queued=%d/%d", |
1169 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1170 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1171 | tg->io_disp[rw], tg->iops[rw], |
e43473b7 VG |
1172 | tg->nr_queued[READ], tg->nr_queued[WRITE]); |
1173 | ||
671058fb | 1174 | bio_associate_current(bio); |
e43473b7 | 1175 | throtl_add_bio_tg(q->td, tg, bio); |
bc16a4f9 | 1176 | throttled = true; |
e43473b7 VG |
1177 | |
1178 | if (update_disptime) { | |
1179 | tg_update_disptime(td, tg); | |
1180 | throtl_schedule_next_dispatch(td); | |
1181 | } | |
1182 | ||
bc16a4f9 | 1183 | out_unlock: |
e43473b7 | 1184 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1185 | out_unlock_rcu: |
1186 | rcu_read_unlock(); | |
bc16a4f9 TH |
1187 | out: |
1188 | return throttled; | |
e43473b7 VG |
1189 | } |
1190 | ||
c9a929dd TH |
1191 | /** |
1192 | * blk_throtl_drain - drain throttled bios | |
1193 | * @q: request_queue to drain throttled bios for | |
1194 | * | |
1195 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1196 | */ | |
1197 | void blk_throtl_drain(struct request_queue *q) | |
1198 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1199 | { | |
1200 | struct throtl_data *td = q->td; | |
1201 | struct throtl_rb_root *st = &td->tg_service_tree; | |
1202 | struct throtl_grp *tg; | |
1203 | struct bio_list bl; | |
1204 | struct bio *bio; | |
1205 | ||
334c2b0b | 1206 | WARN_ON_ONCE(!queue_is_locked(q)); |
c9a929dd TH |
1207 | |
1208 | bio_list_init(&bl); | |
1209 | ||
1210 | while ((tg = throtl_rb_first(st))) { | |
1211 | throtl_dequeue_tg(td, tg); | |
1212 | ||
1213 | while ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
1214 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1215 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
1216 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1217 | } | |
1218 | spin_unlock_irq(q->queue_lock); | |
1219 | ||
1220 | while ((bio = bio_list_pop(&bl))) | |
1221 | generic_make_request(bio); | |
1222 | ||
1223 | spin_lock_irq(q->queue_lock); | |
1224 | } | |
1225 | ||
e43473b7 VG |
1226 | int blk_throtl_init(struct request_queue *q) |
1227 | { | |
1228 | struct throtl_data *td; | |
a2b1693b | 1229 | int ret; |
e43473b7 VG |
1230 | |
1231 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1232 | if (!td) | |
1233 | return -ENOMEM; | |
1234 | ||
e43473b7 | 1235 | td->tg_service_tree = THROTL_RB_ROOT; |
de701c74 | 1236 | td->limits_changed = false; |
a29a171e | 1237 | INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work); |
e43473b7 | 1238 | |
cd1604fa | 1239 | q->td = td; |
29b12589 | 1240 | td->queue = q; |
02977e4a | 1241 | |
a2b1693b TH |
1242 | /* activate policy */ |
1243 | ret = blkcg_activate_policy(q, &blkio_policy_throtl); | |
1244 | if (ret) | |
f51b802c | 1245 | kfree(td); |
a2b1693b | 1246 | return ret; |
e43473b7 VG |
1247 | } |
1248 | ||
1249 | void blk_throtl_exit(struct request_queue *q) | |
1250 | { | |
c875f4d0 | 1251 | BUG_ON(!q->td); |
da527770 | 1252 | throtl_shutdown_wq(q); |
a2b1693b | 1253 | blkcg_deactivate_policy(q, &blkio_policy_throtl); |
c9a929dd | 1254 | kfree(q->td); |
e43473b7 VG |
1255 | } |
1256 | ||
1257 | static int __init throtl_init(void) | |
1258 | { | |
450adcbe VG |
1259 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1260 | if (!kthrotld_workqueue) | |
1261 | panic("Failed to create kthrotld\n"); | |
1262 | ||
8bd435b3 | 1263 | return blkio_policy_register(&blkio_policy_throtl); |
e43473b7 VG |
1264 | } |
1265 | ||
1266 | module_init(throtl_init); |