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 | ||
3c798398 | 24 | static struct blkcg_policy blkcg_policy_throtl; |
0381411e | 25 | |
450adcbe VG |
26 | /* A workqueue to queue throttle related work */ |
27 | static struct workqueue_struct *kthrotld_workqueue; | |
450adcbe | 28 | |
c5cc2070 TH |
29 | /* |
30 | * To implement hierarchical throttling, throtl_grps form a tree and bios | |
31 | * are dispatched upwards level by level until they reach the top and get | |
32 | * issued. When dispatching bios from the children and local group at each | |
33 | * level, if the bios are dispatched into a single bio_list, there's a risk | |
34 | * of a local or child group which can queue many bios at once filling up | |
35 | * the list starving others. | |
36 | * | |
37 | * To avoid such starvation, dispatched bios are queued separately | |
38 | * according to where they came from. When they are again dispatched to | |
39 | * the parent, they're popped in round-robin order so that no single source | |
40 | * hogs the dispatch window. | |
41 | * | |
42 | * throtl_qnode is used to keep the queued bios separated by their sources. | |
43 | * Bios are queued to throtl_qnode which in turn is queued to | |
44 | * throtl_service_queue and then dispatched in round-robin order. | |
45 | * | |
46 | * It's also used to track the reference counts on blkg's. A qnode always | |
47 | * belongs to a throtl_grp and gets queued on itself or the parent, so | |
48 | * incrementing the reference of the associated throtl_grp when a qnode is | |
49 | * queued and decrementing when dequeued is enough to keep the whole blkg | |
50 | * tree pinned while bios are in flight. | |
51 | */ | |
52 | struct throtl_qnode { | |
53 | struct list_head node; /* service_queue->queued[] */ | |
54 | struct bio_list bios; /* queued bios */ | |
55 | struct throtl_grp *tg; /* tg this qnode belongs to */ | |
56 | }; | |
57 | ||
c9e0332e | 58 | struct throtl_service_queue { |
77216b04 TH |
59 | struct throtl_service_queue *parent_sq; /* the parent service_queue */ |
60 | ||
73f0d49a TH |
61 | /* |
62 | * Bios queued directly to this service_queue or dispatched from | |
63 | * children throtl_grp's. | |
64 | */ | |
c5cc2070 | 65 | struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */ |
73f0d49a TH |
66 | unsigned int nr_queued[2]; /* number of queued bios */ |
67 | ||
68 | /* | |
69 | * RB tree of active children throtl_grp's, which are sorted by | |
70 | * their ->disptime. | |
71 | */ | |
c9e0332e TH |
72 | struct rb_root pending_tree; /* RB tree of active tgs */ |
73 | struct rb_node *first_pending; /* first node in the tree */ | |
74 | unsigned int nr_pending; /* # queued in the tree */ | |
75 | unsigned long first_pending_disptime; /* disptime of the first tg */ | |
69df0ab0 | 76 | struct timer_list pending_timer; /* fires on first_pending_disptime */ |
e43473b7 VG |
77 | }; |
78 | ||
5b2c16aa TH |
79 | enum tg_state_flags { |
80 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
0e9f4164 | 81 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ |
5b2c16aa TH |
82 | }; |
83 | ||
e43473b7 VG |
84 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
85 | ||
8a3d2615 TH |
86 | /* Per-cpu group stats */ |
87 | struct tg_stats_cpu { | |
88 | /* total bytes transferred */ | |
89 | struct blkg_rwstat service_bytes; | |
90 | /* total IOs serviced, post merge */ | |
91 | struct blkg_rwstat serviced; | |
92 | }; | |
93 | ||
e43473b7 | 94 | struct throtl_grp { |
f95a04af TH |
95 | /* must be the first member */ |
96 | struct blkg_policy_data pd; | |
97 | ||
c9e0332e | 98 | /* active throtl group service_queue member */ |
e43473b7 VG |
99 | struct rb_node rb_node; |
100 | ||
0f3457f6 TH |
101 | /* throtl_data this group belongs to */ |
102 | struct throtl_data *td; | |
103 | ||
49a2f1e3 TH |
104 | /* this group's service queue */ |
105 | struct throtl_service_queue service_queue; | |
106 | ||
c5cc2070 TH |
107 | /* |
108 | * qnode_on_self is used when bios are directly queued to this | |
109 | * throtl_grp so that local bios compete fairly with bios | |
110 | * dispatched from children. qnode_on_parent is used when bios are | |
111 | * dispatched from this throtl_grp into its parent and will compete | |
112 | * with the sibling qnode_on_parents and the parent's | |
113 | * qnode_on_self. | |
114 | */ | |
115 | struct throtl_qnode qnode_on_self[2]; | |
116 | struct throtl_qnode qnode_on_parent[2]; | |
117 | ||
e43473b7 VG |
118 | /* |
119 | * Dispatch time in jiffies. This is the estimated time when group | |
120 | * will unthrottle and is ready to dispatch more bio. It is used as | |
121 | * key to sort active groups in service tree. | |
122 | */ | |
123 | unsigned long disptime; | |
124 | ||
e43473b7 VG |
125 | unsigned int flags; |
126 | ||
e43473b7 VG |
127 | /* bytes per second rate limits */ |
128 | uint64_t bps[2]; | |
129 | ||
8e89d13f VG |
130 | /* IOPS limits */ |
131 | unsigned int iops[2]; | |
132 | ||
e43473b7 VG |
133 | /* Number of bytes disptached in current slice */ |
134 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
135 | /* Number of bio's dispatched in current slice */ |
136 | unsigned int io_disp[2]; | |
e43473b7 VG |
137 | |
138 | /* When did we start a new slice */ | |
139 | unsigned long slice_start[2]; | |
140 | unsigned long slice_end[2]; | |
fe071437 | 141 | |
8a3d2615 TH |
142 | /* Per cpu stats pointer */ |
143 | struct tg_stats_cpu __percpu *stats_cpu; | |
144 | ||
145 | /* List of tgs waiting for per cpu stats memory to be allocated */ | |
146 | struct list_head stats_alloc_node; | |
e43473b7 VG |
147 | }; |
148 | ||
149 | struct throtl_data | |
150 | { | |
e43473b7 | 151 | /* service tree for active throtl groups */ |
c9e0332e | 152 | struct throtl_service_queue service_queue; |
e43473b7 | 153 | |
e43473b7 VG |
154 | struct request_queue *queue; |
155 | ||
156 | /* Total Number of queued bios on READ and WRITE lists */ | |
157 | unsigned int nr_queued[2]; | |
158 | ||
159 | /* | |
02977e4a | 160 | * number of total undestroyed groups |
e43473b7 VG |
161 | */ |
162 | unsigned int nr_undestroyed_grps; | |
163 | ||
164 | /* Work for dispatching throttled bios */ | |
69df0ab0 | 165 | struct work_struct dispatch_work; |
e43473b7 VG |
166 | }; |
167 | ||
8a3d2615 TH |
168 | /* list and work item to allocate percpu group stats */ |
169 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
170 | static LIST_HEAD(tg_stats_alloc_list); | |
171 | ||
172 | static void tg_stats_alloc_fn(struct work_struct *); | |
173 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
174 | ||
69df0ab0 TH |
175 | static void throtl_pending_timer_fn(unsigned long arg); |
176 | ||
f95a04af TH |
177 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
178 | { | |
179 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
180 | } | |
181 | ||
3c798398 | 182 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 183 | { |
f95a04af | 184 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
185 | } |
186 | ||
3c798398 | 187 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 188 | { |
f95a04af | 189 | return pd_to_blkg(&tg->pd); |
0381411e TH |
190 | } |
191 | ||
03d8e111 TH |
192 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) |
193 | { | |
194 | return blkg_to_tg(td->queue->root_blkg); | |
195 | } | |
196 | ||
fda6f272 TH |
197 | /** |
198 | * sq_to_tg - return the throl_grp the specified service queue belongs to | |
199 | * @sq: the throtl_service_queue of interest | |
200 | * | |
201 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one | |
202 | * embedded in throtl_data, %NULL is returned. | |
203 | */ | |
204 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) | |
205 | { | |
206 | if (sq && sq->parent_sq) | |
207 | return container_of(sq, struct throtl_grp, service_queue); | |
208 | else | |
209 | return NULL; | |
210 | } | |
211 | ||
212 | /** | |
213 | * sq_to_td - return throtl_data the specified service queue belongs to | |
214 | * @sq: the throtl_service_queue of interest | |
215 | * | |
216 | * A service_queue can be embeded in either a throtl_grp or throtl_data. | |
217 | * Determine the associated throtl_data accordingly and return it. | |
218 | */ | |
219 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) | |
220 | { | |
221 | struct throtl_grp *tg = sq_to_tg(sq); | |
222 | ||
223 | if (tg) | |
224 | return tg->td; | |
225 | else | |
226 | return container_of(sq, struct throtl_data, service_queue); | |
227 | } | |
228 | ||
229 | /** | |
230 | * throtl_log - log debug message via blktrace | |
231 | * @sq: the service_queue being reported | |
232 | * @fmt: printf format string | |
233 | * @args: printf args | |
234 | * | |
235 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a | |
236 | * throtl_grp; otherwise, just "throtl". | |
237 | * | |
238 | * TODO: this should be made a function and name formatting should happen | |
239 | * after testing whether blktrace is enabled. | |
240 | */ | |
241 | #define throtl_log(sq, fmt, args...) do { \ | |
242 | struct throtl_grp *__tg = sq_to_tg((sq)); \ | |
243 | struct throtl_data *__td = sq_to_td((sq)); \ | |
244 | \ | |
245 | (void)__td; \ | |
246 | if ((__tg)) { \ | |
247 | char __pbuf[128]; \ | |
54e7ed12 | 248 | \ |
fda6f272 TH |
249 | blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \ |
250 | blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
251 | } else { \ | |
252 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
253 | } \ | |
54e7ed12 | 254 | } while (0) |
e43473b7 | 255 | |
8a3d2615 TH |
256 | /* |
257 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
3b07e9ca | 258 | * system_wq once there are some groups on the alloc_list waiting for |
8a3d2615 TH |
259 | * allocation. |
260 | */ | |
261 | static void tg_stats_alloc_fn(struct work_struct *work) | |
262 | { | |
263 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
264 | struct delayed_work *dwork = to_delayed_work(work); | |
265 | bool empty = false; | |
266 | ||
267 | alloc_stats: | |
268 | if (!stats_cpu) { | |
269 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
270 | if (!stats_cpu) { | |
271 | /* allocation failed, try again after some time */ | |
3b07e9ca | 272 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); |
8a3d2615 TH |
273 | return; |
274 | } | |
275 | } | |
276 | ||
277 | spin_lock_irq(&tg_stats_alloc_lock); | |
278 | ||
279 | if (!list_empty(&tg_stats_alloc_list)) { | |
280 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
281 | struct throtl_grp, | |
282 | stats_alloc_node); | |
283 | swap(tg->stats_cpu, stats_cpu); | |
284 | list_del_init(&tg->stats_alloc_node); | |
285 | } | |
286 | ||
287 | empty = list_empty(&tg_stats_alloc_list); | |
288 | spin_unlock_irq(&tg_stats_alloc_lock); | |
289 | if (!empty) | |
290 | goto alloc_stats; | |
291 | } | |
292 | ||
c5cc2070 TH |
293 | static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg) |
294 | { | |
295 | INIT_LIST_HEAD(&qn->node); | |
296 | bio_list_init(&qn->bios); | |
297 | qn->tg = tg; | |
298 | } | |
299 | ||
300 | /** | |
301 | * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it | |
302 | * @bio: bio being added | |
303 | * @qn: qnode to add bio to | |
304 | * @queued: the service_queue->queued[] list @qn belongs to | |
305 | * | |
306 | * Add @bio to @qn and put @qn on @queued if it's not already on. | |
307 | * @qn->tg's reference count is bumped when @qn is activated. See the | |
308 | * comment on top of throtl_qnode definition for details. | |
309 | */ | |
310 | static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn, | |
311 | struct list_head *queued) | |
312 | { | |
313 | bio_list_add(&qn->bios, bio); | |
314 | if (list_empty(&qn->node)) { | |
315 | list_add_tail(&qn->node, queued); | |
316 | blkg_get(tg_to_blkg(qn->tg)); | |
317 | } | |
318 | } | |
319 | ||
320 | /** | |
321 | * throtl_peek_queued - peek the first bio on a qnode list | |
322 | * @queued: the qnode list to peek | |
323 | */ | |
324 | static struct bio *throtl_peek_queued(struct list_head *queued) | |
325 | { | |
326 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
327 | struct bio *bio; | |
328 | ||
329 | if (list_empty(queued)) | |
330 | return NULL; | |
331 | ||
332 | bio = bio_list_peek(&qn->bios); | |
333 | WARN_ON_ONCE(!bio); | |
334 | return bio; | |
335 | } | |
336 | ||
337 | /** | |
338 | * throtl_pop_queued - pop the first bio form a qnode list | |
339 | * @queued: the qnode list to pop a bio from | |
340 | * @tg_to_put: optional out argument for throtl_grp to put | |
341 | * | |
342 | * Pop the first bio from the qnode list @queued. After popping, the first | |
343 | * qnode is removed from @queued if empty or moved to the end of @queued so | |
344 | * that the popping order is round-robin. | |
345 | * | |
346 | * When the first qnode is removed, its associated throtl_grp should be put | |
347 | * too. If @tg_to_put is NULL, this function automatically puts it; | |
348 | * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is | |
349 | * responsible for putting it. | |
350 | */ | |
351 | static struct bio *throtl_pop_queued(struct list_head *queued, | |
352 | struct throtl_grp **tg_to_put) | |
353 | { | |
354 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
355 | struct bio *bio; | |
356 | ||
357 | if (list_empty(queued)) | |
358 | return NULL; | |
359 | ||
360 | bio = bio_list_pop(&qn->bios); | |
361 | WARN_ON_ONCE(!bio); | |
362 | ||
363 | if (bio_list_empty(&qn->bios)) { | |
364 | list_del_init(&qn->node); | |
365 | if (tg_to_put) | |
366 | *tg_to_put = qn->tg; | |
367 | else | |
368 | blkg_put(tg_to_blkg(qn->tg)); | |
369 | } else { | |
370 | list_move_tail(&qn->node, queued); | |
371 | } | |
372 | ||
373 | return bio; | |
374 | } | |
375 | ||
49a2f1e3 | 376 | /* init a service_queue, assumes the caller zeroed it */ |
77216b04 TH |
377 | static void throtl_service_queue_init(struct throtl_service_queue *sq, |
378 | struct throtl_service_queue *parent_sq) | |
49a2f1e3 | 379 | { |
c5cc2070 TH |
380 | INIT_LIST_HEAD(&sq->queued[0]); |
381 | INIT_LIST_HEAD(&sq->queued[1]); | |
49a2f1e3 | 382 | sq->pending_tree = RB_ROOT; |
77216b04 | 383 | sq->parent_sq = parent_sq; |
69df0ab0 TH |
384 | setup_timer(&sq->pending_timer, throtl_pending_timer_fn, |
385 | (unsigned long)sq); | |
386 | } | |
387 | ||
388 | static void throtl_service_queue_exit(struct throtl_service_queue *sq) | |
389 | { | |
390 | del_timer_sync(&sq->pending_timer); | |
49a2f1e3 TH |
391 | } |
392 | ||
3c798398 | 393 | static void throtl_pd_init(struct blkcg_gq *blkg) |
a29a171e | 394 | { |
0381411e | 395 | struct throtl_grp *tg = blkg_to_tg(blkg); |
77216b04 | 396 | struct throtl_data *td = blkg->q->td; |
ff26eaad | 397 | unsigned long flags; |
c5cc2070 | 398 | int rw; |
cd1604fa | 399 | |
77216b04 | 400 | throtl_service_queue_init(&tg->service_queue, &td->service_queue); |
c5cc2070 TH |
401 | for (rw = READ; rw <= WRITE; rw++) { |
402 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); | |
403 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); | |
404 | } | |
405 | ||
a29a171e | 406 | RB_CLEAR_NODE(&tg->rb_node); |
77216b04 | 407 | tg->td = td; |
a29a171e | 408 | |
e56da7e2 TH |
409 | tg->bps[READ] = -1; |
410 | tg->bps[WRITE] = -1; | |
411 | tg->iops[READ] = -1; | |
412 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
413 | |
414 | /* | |
415 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
416 | * but percpu allocator can't be called from IO path. Queue tg on | |
417 | * tg_stats_alloc_list and allocate from work item. | |
418 | */ | |
ff26eaad | 419 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 420 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); |
3b07e9ca | 421 | schedule_delayed_work(&tg_stats_alloc_work, 0); |
ff26eaad | 422 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
423 | } |
424 | ||
3c798398 | 425 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
8a3d2615 TH |
426 | { |
427 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
ff26eaad | 428 | unsigned long flags; |
8a3d2615 | 429 | |
ff26eaad | 430 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 431 | list_del_init(&tg->stats_alloc_node); |
ff26eaad | 432 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
433 | |
434 | free_percpu(tg->stats_cpu); | |
69df0ab0 TH |
435 | |
436 | throtl_service_queue_exit(&tg->service_queue); | |
8a3d2615 TH |
437 | } |
438 | ||
3c798398 | 439 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
8a3d2615 TH |
440 | { |
441 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
442 | int cpu; | |
443 | ||
444 | if (tg->stats_cpu == NULL) | |
445 | return; | |
446 | ||
447 | for_each_possible_cpu(cpu) { | |
448 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
449 | ||
450 | blkg_rwstat_reset(&sc->service_bytes); | |
451 | blkg_rwstat_reset(&sc->serviced); | |
452 | } | |
a29a171e VG |
453 | } |
454 | ||
3c798398 TH |
455 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
456 | struct blkcg *blkcg) | |
e43473b7 | 457 | { |
be2c6b19 | 458 | /* |
3c798398 TH |
459 | * This is the common case when there are no blkcgs. Avoid lookup |
460 | * in this case | |
cd1604fa | 461 | */ |
3c798398 | 462 | if (blkcg == &blkcg_root) |
03d8e111 | 463 | return td_root_tg(td); |
e43473b7 | 464 | |
e8989fae | 465 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
466 | } |
467 | ||
cd1604fa | 468 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
3c798398 | 469 | struct blkcg *blkcg) |
e43473b7 | 470 | { |
f469a7b4 | 471 | struct request_queue *q = td->queue; |
cd1604fa | 472 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 473 | |
f469a7b4 | 474 | /* |
3c798398 TH |
475 | * This is the common case when there are no blkcgs. Avoid lookup |
476 | * in this case | |
f469a7b4 | 477 | */ |
3c798398 | 478 | if (blkcg == &blkcg_root) { |
03d8e111 | 479 | tg = td_root_tg(td); |
cd1604fa | 480 | } else { |
3c798398 | 481 | struct blkcg_gq *blkg; |
f469a7b4 | 482 | |
3c96cb32 | 483 | blkg = blkg_lookup_create(blkcg, q); |
f469a7b4 | 484 | |
cd1604fa TH |
485 | /* if %NULL and @q is alive, fall back to root_tg */ |
486 | if (!IS_ERR(blkg)) | |
0381411e | 487 | tg = blkg_to_tg(blkg); |
3f3299d5 | 488 | else if (!blk_queue_dying(q)) |
03d8e111 | 489 | tg = td_root_tg(td); |
f469a7b4 VG |
490 | } |
491 | ||
e43473b7 VG |
492 | return tg; |
493 | } | |
494 | ||
0049af73 TH |
495 | static struct throtl_grp * |
496 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
497 | { |
498 | /* Service tree is empty */ | |
0049af73 | 499 | if (!parent_sq->nr_pending) |
e43473b7 VG |
500 | return NULL; |
501 | ||
0049af73 TH |
502 | if (!parent_sq->first_pending) |
503 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
e43473b7 | 504 | |
0049af73 TH |
505 | if (parent_sq->first_pending) |
506 | return rb_entry_tg(parent_sq->first_pending); | |
e43473b7 VG |
507 | |
508 | return NULL; | |
509 | } | |
510 | ||
511 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
512 | { | |
513 | rb_erase(n, root); | |
514 | RB_CLEAR_NODE(n); | |
515 | } | |
516 | ||
0049af73 TH |
517 | static void throtl_rb_erase(struct rb_node *n, |
518 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 519 | { |
0049af73 TH |
520 | if (parent_sq->first_pending == n) |
521 | parent_sq->first_pending = NULL; | |
522 | rb_erase_init(n, &parent_sq->pending_tree); | |
523 | --parent_sq->nr_pending; | |
e43473b7 VG |
524 | } |
525 | ||
0049af73 | 526 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
527 | { |
528 | struct throtl_grp *tg; | |
529 | ||
0049af73 | 530 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
531 | if (!tg) |
532 | return; | |
533 | ||
0049af73 | 534 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
535 | } |
536 | ||
77216b04 | 537 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 538 | { |
77216b04 | 539 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
0049af73 | 540 | struct rb_node **node = &parent_sq->pending_tree.rb_node; |
e43473b7 VG |
541 | struct rb_node *parent = NULL; |
542 | struct throtl_grp *__tg; | |
543 | unsigned long key = tg->disptime; | |
544 | int left = 1; | |
545 | ||
546 | while (*node != NULL) { | |
547 | parent = *node; | |
548 | __tg = rb_entry_tg(parent); | |
549 | ||
550 | if (time_before(key, __tg->disptime)) | |
551 | node = &parent->rb_left; | |
552 | else { | |
553 | node = &parent->rb_right; | |
554 | left = 0; | |
555 | } | |
556 | } | |
557 | ||
558 | if (left) | |
0049af73 | 559 | parent_sq->first_pending = &tg->rb_node; |
e43473b7 VG |
560 | |
561 | rb_link_node(&tg->rb_node, parent, node); | |
0049af73 | 562 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); |
e43473b7 VG |
563 | } |
564 | ||
77216b04 | 565 | static void __throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 566 | { |
77216b04 | 567 | tg_service_queue_add(tg); |
5b2c16aa | 568 | tg->flags |= THROTL_TG_PENDING; |
77216b04 | 569 | tg->service_queue.parent_sq->nr_pending++; |
e43473b7 VG |
570 | } |
571 | ||
77216b04 | 572 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 573 | { |
5b2c16aa | 574 | if (!(tg->flags & THROTL_TG_PENDING)) |
77216b04 | 575 | __throtl_enqueue_tg(tg); |
e43473b7 VG |
576 | } |
577 | ||
77216b04 | 578 | static void __throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 579 | { |
77216b04 | 580 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
5b2c16aa | 581 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
582 | } |
583 | ||
77216b04 | 584 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 585 | { |
5b2c16aa | 586 | if (tg->flags & THROTL_TG_PENDING) |
77216b04 | 587 | __throtl_dequeue_tg(tg); |
e43473b7 VG |
588 | } |
589 | ||
a9131a27 | 590 | /* Call with queue lock held */ |
69df0ab0 TH |
591 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, |
592 | unsigned long expires) | |
a9131a27 | 593 | { |
69df0ab0 TH |
594 | mod_timer(&sq->pending_timer, expires); |
595 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
596 | expires - jiffies, jiffies); | |
a9131a27 TH |
597 | } |
598 | ||
7f52f98c TH |
599 | /** |
600 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle | |
601 | * @sq: the service_queue to schedule dispatch for | |
602 | * @force: force scheduling | |
603 | * | |
604 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the | |
605 | * dispatch time of the first pending child. Returns %true if either timer | |
606 | * is armed or there's no pending child left. %false if the current | |
607 | * dispatch window is still open and the caller should continue | |
608 | * dispatching. | |
609 | * | |
610 | * If @force is %true, the dispatch timer is always scheduled and this | |
611 | * function is guaranteed to return %true. This is to be used when the | |
612 | * caller can't dispatch itself and needs to invoke pending_timer | |
613 | * unconditionally. Note that forced scheduling is likely to induce short | |
614 | * delay before dispatch starts even if @sq->first_pending_disptime is not | |
615 | * in the future and thus shouldn't be used in hot paths. | |
616 | */ | |
617 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
618 | bool force) | |
e43473b7 | 619 | { |
6a525600 | 620 | /* any pending children left? */ |
c9e0332e | 621 | if (!sq->nr_pending) |
7f52f98c | 622 | return true; |
e43473b7 | 623 | |
c9e0332e | 624 | update_min_dispatch_time(sq); |
e43473b7 | 625 | |
69df0ab0 | 626 | /* is the next dispatch time in the future? */ |
7f52f98c | 627 | if (force || time_after(sq->first_pending_disptime, jiffies)) { |
69df0ab0 | 628 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); |
7f52f98c | 629 | return true; |
69df0ab0 TH |
630 | } |
631 | ||
7f52f98c TH |
632 | /* tell the caller to continue dispatching */ |
633 | return false; | |
e43473b7 VG |
634 | } |
635 | ||
32ee5bc4 VG |
636 | static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, |
637 | bool rw, unsigned long start) | |
638 | { | |
639 | tg->bytes_disp[rw] = 0; | |
640 | tg->io_disp[rw] = 0; | |
641 | ||
642 | /* | |
643 | * Previous slice has expired. We must have trimmed it after last | |
644 | * bio dispatch. That means since start of last slice, we never used | |
645 | * that bandwidth. Do try to make use of that bandwidth while giving | |
646 | * credit. | |
647 | */ | |
648 | if (time_after_eq(start, tg->slice_start[rw])) | |
649 | tg->slice_start[rw] = start; | |
650 | ||
651 | tg->slice_end[rw] = jiffies + throtl_slice; | |
652 | throtl_log(&tg->service_queue, | |
653 | "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", | |
654 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
655 | tg->slice_end[rw], jiffies); | |
656 | } | |
657 | ||
0f3457f6 | 658 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
659 | { |
660 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 661 | tg->io_disp[rw] = 0; |
e43473b7 VG |
662 | tg->slice_start[rw] = jiffies; |
663 | tg->slice_end[rw] = jiffies + throtl_slice; | |
fda6f272 TH |
664 | throtl_log(&tg->service_queue, |
665 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
666 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
667 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
668 | } |
669 | ||
0f3457f6 TH |
670 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
671 | unsigned long jiffy_end) | |
d1ae8ffd VG |
672 | { |
673 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
674 | } | |
675 | ||
0f3457f6 TH |
676 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
677 | unsigned long jiffy_end) | |
e43473b7 VG |
678 | { |
679 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
fda6f272 TH |
680 | throtl_log(&tg->service_queue, |
681 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
682 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
683 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
684 | } |
685 | ||
686 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 687 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
688 | { |
689 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
690 | return 0; | |
691 | ||
692 | return 1; | |
693 | } | |
694 | ||
695 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 696 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 697 | { |
3aad5d3e VG |
698 | unsigned long nr_slices, time_elapsed, io_trim; |
699 | u64 bytes_trim, tmp; | |
e43473b7 VG |
700 | |
701 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
702 | ||
703 | /* | |
704 | * If bps are unlimited (-1), then time slice don't get | |
705 | * renewed. Don't try to trim the slice if slice is used. A new | |
706 | * slice will start when appropriate. | |
707 | */ | |
0f3457f6 | 708 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
709 | return; |
710 | ||
d1ae8ffd VG |
711 | /* |
712 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
713 | * that initially cgroup limit was very low resulting in high | |
714 | * slice_end, but later limit was bumped up and bio was dispached | |
715 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
716 | * is bad because it does not allow new slice to start. | |
717 | */ | |
718 | ||
0f3457f6 | 719 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); |
d1ae8ffd | 720 | |
e43473b7 VG |
721 | time_elapsed = jiffies - tg->slice_start[rw]; |
722 | ||
723 | nr_slices = time_elapsed / throtl_slice; | |
724 | ||
725 | if (!nr_slices) | |
726 | return; | |
3aad5d3e VG |
727 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
728 | do_div(tmp, HZ); | |
729 | bytes_trim = tmp; | |
e43473b7 | 730 | |
8e89d13f | 731 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 732 | |
8e89d13f | 733 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
734 | return; |
735 | ||
736 | if (tg->bytes_disp[rw] >= bytes_trim) | |
737 | tg->bytes_disp[rw] -= bytes_trim; | |
738 | else | |
739 | tg->bytes_disp[rw] = 0; | |
740 | ||
8e89d13f VG |
741 | if (tg->io_disp[rw] >= io_trim) |
742 | tg->io_disp[rw] -= io_trim; | |
743 | else | |
744 | tg->io_disp[rw] = 0; | |
745 | ||
e43473b7 VG |
746 | tg->slice_start[rw] += nr_slices * throtl_slice; |
747 | ||
fda6f272 TH |
748 | throtl_log(&tg->service_queue, |
749 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
750 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
751 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
e43473b7 VG |
752 | } |
753 | ||
0f3457f6 TH |
754 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
755 | unsigned long *wait) | |
e43473b7 VG |
756 | { |
757 | bool rw = bio_data_dir(bio); | |
8e89d13f | 758 | unsigned int io_allowed; |
e43473b7 | 759 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 760 | u64 tmp; |
e43473b7 | 761 | |
8e89d13f | 762 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 763 | |
8e89d13f VG |
764 | /* Slice has just started. Consider one slice interval */ |
765 | if (!jiffy_elapsed) | |
766 | jiffy_elapsed_rnd = throtl_slice; | |
767 | ||
768 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
769 | ||
c49c06e4 VG |
770 | /* |
771 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
772 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
773 | * will allow dispatch after 1 second and after that slice should | |
774 | * have been trimmed. | |
775 | */ | |
776 | ||
777 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
778 | do_div(tmp, HZ); | |
779 | ||
780 | if (tmp > UINT_MAX) | |
781 | io_allowed = UINT_MAX; | |
782 | else | |
783 | io_allowed = tmp; | |
8e89d13f VG |
784 | |
785 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
786 | if (wait) |
787 | *wait = 0; | |
788 | return 1; | |
789 | } | |
790 | ||
8e89d13f VG |
791 | /* Calc approx time to dispatch */ |
792 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
793 | ||
794 | if (jiffy_wait > jiffy_elapsed) | |
795 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
796 | else | |
797 | jiffy_wait = 1; | |
798 | ||
799 | if (wait) | |
800 | *wait = jiffy_wait; | |
801 | return 0; | |
802 | } | |
803 | ||
0f3457f6 TH |
804 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
805 | unsigned long *wait) | |
8e89d13f VG |
806 | { |
807 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 808 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 809 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
810 | |
811 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
812 | ||
813 | /* Slice has just started. Consider one slice interval */ | |
814 | if (!jiffy_elapsed) | |
815 | jiffy_elapsed_rnd = throtl_slice; | |
816 | ||
817 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
818 | ||
5e901a2b VG |
819 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
820 | do_div(tmp, HZ); | |
3aad5d3e | 821 | bytes_allowed = tmp; |
e43473b7 VG |
822 | |
823 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
824 | if (wait) | |
825 | *wait = 0; | |
826 | return 1; | |
827 | } | |
828 | ||
829 | /* Calc approx time to dispatch */ | |
830 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
831 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
832 | ||
833 | if (!jiffy_wait) | |
834 | jiffy_wait = 1; | |
835 | ||
836 | /* | |
837 | * This wait time is without taking into consideration the rounding | |
838 | * up we did. Add that time also. | |
839 | */ | |
840 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
841 | if (wait) |
842 | *wait = jiffy_wait; | |
8e89d13f VG |
843 | return 0; |
844 | } | |
845 | ||
af75cd3c VG |
846 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
847 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
848 | return 1; | |
849 | return 0; | |
850 | } | |
851 | ||
8e89d13f VG |
852 | /* |
853 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
854 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
855 | */ | |
0f3457f6 TH |
856 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
857 | unsigned long *wait) | |
8e89d13f VG |
858 | { |
859 | bool rw = bio_data_dir(bio); | |
860 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
861 | ||
862 | /* | |
863 | * Currently whole state machine of group depends on first bio | |
864 | * queued in the group bio list. So one should not be calling | |
865 | * this function with a different bio if there are other bios | |
866 | * queued. | |
867 | */ | |
73f0d49a | 868 | BUG_ON(tg->service_queue.nr_queued[rw] && |
c5cc2070 | 869 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); |
e43473b7 | 870 | |
8e89d13f VG |
871 | /* If tg->bps = -1, then BW is unlimited */ |
872 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
873 | if (wait) | |
874 | *wait = 0; | |
875 | return 1; | |
876 | } | |
877 | ||
878 | /* | |
879 | * If previous slice expired, start a new one otherwise renew/extend | |
880 | * existing slice to make sure it is at least throtl_slice interval | |
881 | * long since now. | |
882 | */ | |
0f3457f6 TH |
883 | if (throtl_slice_used(tg, rw)) |
884 | throtl_start_new_slice(tg, rw); | |
8e89d13f VG |
885 | else { |
886 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
0f3457f6 | 887 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); |
8e89d13f VG |
888 | } |
889 | ||
0f3457f6 TH |
890 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
891 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
892 | if (wait) |
893 | *wait = 0; | |
894 | return 1; | |
895 | } | |
896 | ||
897 | max_wait = max(bps_wait, iops_wait); | |
898 | ||
899 | if (wait) | |
900 | *wait = max_wait; | |
901 | ||
902 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 903 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
904 | |
905 | return 0; | |
906 | } | |
907 | ||
3c798398 | 908 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
909 | int rw) |
910 | { | |
8a3d2615 TH |
911 | struct throtl_grp *tg = blkg_to_tg(blkg); |
912 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
913 | unsigned long flags; |
914 | ||
915 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 916 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
917 | return; |
918 | ||
919 | /* | |
920 | * Disabling interrupts to provide mutual exclusion between two | |
921 | * writes on same cpu. It probably is not needed for 64bit. Not | |
922 | * optimizing that case yet. | |
923 | */ | |
924 | local_irq_save(flags); | |
925 | ||
8a3d2615 | 926 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 927 | |
629ed0b1 TH |
928 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
929 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
930 | ||
931 | local_irq_restore(flags); | |
932 | } | |
933 | ||
e43473b7 VG |
934 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
935 | { | |
936 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
937 | |
938 | /* Charge the bio to the group */ | |
939 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 940 | tg->io_disp[rw]++; |
e43473b7 | 941 | |
2a0f61e6 TH |
942 | /* |
943 | * REQ_THROTTLED is used to prevent the same bio to be throttled | |
944 | * more than once as a throttled bio will go through blk-throtl the | |
945 | * second time when it eventually gets issued. Set it when a bio | |
946 | * is being charged to a tg. | |
947 | * | |
948 | * Dispatch stats aren't recursive and each @bio should only be | |
949 | * accounted by the @tg it was originally associated with. Let's | |
950 | * update the stats when setting REQ_THROTTLED for the first time | |
951 | * which is guaranteed to be for the @bio's original tg. | |
952 | */ | |
953 | if (!(bio->bi_rw & REQ_THROTTLED)) { | |
954 | bio->bi_rw |= REQ_THROTTLED; | |
955 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, | |
956 | bio->bi_rw); | |
957 | } | |
e43473b7 VG |
958 | } |
959 | ||
c5cc2070 TH |
960 | /** |
961 | * throtl_add_bio_tg - add a bio to the specified throtl_grp | |
962 | * @bio: bio to add | |
963 | * @qn: qnode to use | |
964 | * @tg: the target throtl_grp | |
965 | * | |
966 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, | |
967 | * tg->qnode_on_self[] is used. | |
968 | */ | |
969 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, | |
970 | struct throtl_grp *tg) | |
e43473b7 | 971 | { |
73f0d49a | 972 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
973 | bool rw = bio_data_dir(bio); |
974 | ||
c5cc2070 TH |
975 | if (!qn) |
976 | qn = &tg->qnode_on_self[rw]; | |
977 | ||
0e9f4164 TH |
978 | /* |
979 | * If @tg doesn't currently have any bios queued in the same | |
980 | * direction, queueing @bio can change when @tg should be | |
981 | * dispatched. Mark that @tg was empty. This is automatically | |
982 | * cleaered on the next tg_update_disptime(). | |
983 | */ | |
984 | if (!sq->nr_queued[rw]) | |
985 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
986 | ||
c5cc2070 TH |
987 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); |
988 | ||
73f0d49a | 989 | sq->nr_queued[rw]++; |
77216b04 | 990 | throtl_enqueue_tg(tg); |
e43473b7 VG |
991 | } |
992 | ||
77216b04 | 993 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 994 | { |
73f0d49a | 995 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
996 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
997 | struct bio *bio; | |
998 | ||
c5cc2070 | 999 | if ((bio = throtl_peek_queued(&sq->queued[READ]))) |
0f3457f6 | 1000 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 1001 | |
c5cc2070 | 1002 | if ((bio = throtl_peek_queued(&sq->queued[WRITE]))) |
0f3457f6 | 1003 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
1004 | |
1005 | min_wait = min(read_wait, write_wait); | |
1006 | disptime = jiffies + min_wait; | |
1007 | ||
e43473b7 | 1008 | /* Update dispatch time */ |
77216b04 | 1009 | throtl_dequeue_tg(tg); |
e43473b7 | 1010 | tg->disptime = disptime; |
77216b04 | 1011 | throtl_enqueue_tg(tg); |
0e9f4164 TH |
1012 | |
1013 | /* see throtl_add_bio_tg() */ | |
1014 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
1015 | } |
1016 | ||
32ee5bc4 VG |
1017 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, |
1018 | struct throtl_grp *parent_tg, bool rw) | |
1019 | { | |
1020 | if (throtl_slice_used(parent_tg, rw)) { | |
1021 | throtl_start_new_slice_with_credit(parent_tg, rw, | |
1022 | child_tg->slice_start[rw]); | |
1023 | } | |
1024 | ||
1025 | } | |
1026 | ||
77216b04 | 1027 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 1028 | { |
73f0d49a | 1029 | struct throtl_service_queue *sq = &tg->service_queue; |
6bc9c2b4 TH |
1030 | struct throtl_service_queue *parent_sq = sq->parent_sq; |
1031 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
c5cc2070 | 1032 | struct throtl_grp *tg_to_put = NULL; |
e43473b7 VG |
1033 | struct bio *bio; |
1034 | ||
c5cc2070 TH |
1035 | /* |
1036 | * @bio is being transferred from @tg to @parent_sq. Popping a bio | |
1037 | * from @tg may put its reference and @parent_sq might end up | |
1038 | * getting released prematurely. Remember the tg to put and put it | |
1039 | * after @bio is transferred to @parent_sq. | |
1040 | */ | |
1041 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); | |
73f0d49a | 1042 | sq->nr_queued[rw]--; |
e43473b7 VG |
1043 | |
1044 | throtl_charge_bio(tg, bio); | |
6bc9c2b4 TH |
1045 | |
1046 | /* | |
1047 | * If our parent is another tg, we just need to transfer @bio to | |
1048 | * the parent using throtl_add_bio_tg(). If our parent is | |
1049 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
1050 | * bio_lists[] and decrease total number queued. The caller is | |
1051 | * responsible for issuing these bios. | |
1052 | */ | |
1053 | if (parent_tg) { | |
c5cc2070 | 1054 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); |
32ee5bc4 | 1055 | start_parent_slice_with_credit(tg, parent_tg, rw); |
6bc9c2b4 | 1056 | } else { |
c5cc2070 TH |
1057 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], |
1058 | &parent_sq->queued[rw]); | |
6bc9c2b4 TH |
1059 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
1060 | tg->td->nr_queued[rw]--; | |
1061 | } | |
e43473b7 | 1062 | |
0f3457f6 | 1063 | throtl_trim_slice(tg, rw); |
6bc9c2b4 | 1064 | |
c5cc2070 TH |
1065 | if (tg_to_put) |
1066 | blkg_put(tg_to_blkg(tg_to_put)); | |
e43473b7 VG |
1067 | } |
1068 | ||
77216b04 | 1069 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 1070 | { |
73f0d49a | 1071 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1072 | unsigned int nr_reads = 0, nr_writes = 0; |
1073 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 1074 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
1075 | struct bio *bio; |
1076 | ||
1077 | /* Try to dispatch 75% READS and 25% WRITES */ | |
1078 | ||
c5cc2070 | 1079 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && |
0f3457f6 | 1080 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1081 | |
77216b04 | 1082 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1083 | nr_reads++; |
1084 | ||
1085 | if (nr_reads >= max_nr_reads) | |
1086 | break; | |
1087 | } | |
1088 | ||
c5cc2070 | 1089 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && |
0f3457f6 | 1090 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1091 | |
77216b04 | 1092 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1093 | nr_writes++; |
1094 | ||
1095 | if (nr_writes >= max_nr_writes) | |
1096 | break; | |
1097 | } | |
1098 | ||
1099 | return nr_reads + nr_writes; | |
1100 | } | |
1101 | ||
651930bc | 1102 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
1103 | { |
1104 | unsigned int nr_disp = 0; | |
e43473b7 VG |
1105 | |
1106 | while (1) { | |
73f0d49a TH |
1107 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
1108 | struct throtl_service_queue *sq = &tg->service_queue; | |
e43473b7 VG |
1109 | |
1110 | if (!tg) | |
1111 | break; | |
1112 | ||
1113 | if (time_before(jiffies, tg->disptime)) | |
1114 | break; | |
1115 | ||
77216b04 | 1116 | throtl_dequeue_tg(tg); |
e43473b7 | 1117 | |
77216b04 | 1118 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 1119 | |
73f0d49a | 1120 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
77216b04 | 1121 | tg_update_disptime(tg); |
e43473b7 VG |
1122 | |
1123 | if (nr_disp >= throtl_quantum) | |
1124 | break; | |
1125 | } | |
1126 | ||
1127 | return nr_disp; | |
1128 | } | |
1129 | ||
6e1a5704 TH |
1130 | /** |
1131 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
1132 | * @arg: the throtl_service_queue being serviced | |
1133 | * | |
1134 | * This timer is armed when a child throtl_grp with active bio's become | |
1135 | * pending and queued on the service_queue's pending_tree and expires when | |
1136 | * the first child throtl_grp should be dispatched. This function | |
2e48a530 TH |
1137 | * dispatches bio's from the children throtl_grps to the parent |
1138 | * service_queue. | |
1139 | * | |
1140 | * If the parent's parent is another throtl_grp, dispatching is propagated | |
1141 | * by either arming its pending_timer or repeating dispatch directly. If | |
1142 | * the top-level service_tree is reached, throtl_data->dispatch_work is | |
1143 | * kicked so that the ready bio's are issued. | |
6e1a5704 | 1144 | */ |
69df0ab0 TH |
1145 | static void throtl_pending_timer_fn(unsigned long arg) |
1146 | { | |
1147 | struct throtl_service_queue *sq = (void *)arg; | |
2e48a530 | 1148 | struct throtl_grp *tg = sq_to_tg(sq); |
69df0ab0 | 1149 | struct throtl_data *td = sq_to_td(sq); |
cb76199c | 1150 | struct request_queue *q = td->queue; |
2e48a530 TH |
1151 | struct throtl_service_queue *parent_sq; |
1152 | bool dispatched; | |
6e1a5704 | 1153 | int ret; |
e43473b7 VG |
1154 | |
1155 | spin_lock_irq(q->queue_lock); | |
2e48a530 TH |
1156 | again: |
1157 | parent_sq = sq->parent_sq; | |
1158 | dispatched = false; | |
e43473b7 | 1159 | |
7f52f98c TH |
1160 | while (true) { |
1161 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
2e48a530 TH |
1162 | sq->nr_queued[READ] + sq->nr_queued[WRITE], |
1163 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
7f52f98c TH |
1164 | |
1165 | ret = throtl_select_dispatch(sq); | |
1166 | if (ret) { | |
7f52f98c TH |
1167 | throtl_log(sq, "bios disp=%u", ret); |
1168 | dispatched = true; | |
1169 | } | |
e43473b7 | 1170 | |
7f52f98c TH |
1171 | if (throtl_schedule_next_dispatch(sq, false)) |
1172 | break; | |
e43473b7 | 1173 | |
7f52f98c TH |
1174 | /* this dispatch windows is still open, relax and repeat */ |
1175 | spin_unlock_irq(q->queue_lock); | |
1176 | cpu_relax(); | |
1177 | spin_lock_irq(q->queue_lock); | |
651930bc | 1178 | } |
e43473b7 | 1179 | |
2e48a530 TH |
1180 | if (!dispatched) |
1181 | goto out_unlock; | |
6e1a5704 | 1182 | |
2e48a530 TH |
1183 | if (parent_sq) { |
1184 | /* @parent_sq is another throl_grp, propagate dispatch */ | |
1185 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1186 | tg_update_disptime(tg); | |
1187 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { | |
1188 | /* window is already open, repeat dispatching */ | |
1189 | sq = parent_sq; | |
1190 | tg = sq_to_tg(sq); | |
1191 | goto again; | |
1192 | } | |
1193 | } | |
1194 | } else { | |
1195 | /* reached the top-level, queue issueing */ | |
1196 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1197 | } | |
1198 | out_unlock: | |
e43473b7 | 1199 | spin_unlock_irq(q->queue_lock); |
6e1a5704 | 1200 | } |
e43473b7 | 1201 | |
6e1a5704 TH |
1202 | /** |
1203 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
1204 | * @work: work item being executed | |
1205 | * | |
1206 | * This function is queued for execution when bio's reach the bio_lists[] | |
1207 | * of throtl_data->service_queue. Those bio's are ready and issued by this | |
1208 | * function. | |
1209 | */ | |
1210 | void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
1211 | { | |
1212 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1213 | dispatch_work); | |
1214 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1215 | struct request_queue *q = td->queue; | |
1216 | struct bio_list bio_list_on_stack; | |
1217 | struct bio *bio; | |
1218 | struct blk_plug plug; | |
1219 | int rw; | |
1220 | ||
1221 | bio_list_init(&bio_list_on_stack); | |
1222 | ||
1223 | spin_lock_irq(q->queue_lock); | |
c5cc2070 TH |
1224 | for (rw = READ; rw <= WRITE; rw++) |
1225 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) | |
1226 | bio_list_add(&bio_list_on_stack, bio); | |
6e1a5704 TH |
1227 | spin_unlock_irq(q->queue_lock); |
1228 | ||
1229 | if (!bio_list_empty(&bio_list_on_stack)) { | |
69d60eb9 | 1230 | blk_start_plug(&plug); |
e43473b7 VG |
1231 | while((bio = bio_list_pop(&bio_list_on_stack))) |
1232 | generic_make_request(bio); | |
69d60eb9 | 1233 | blk_finish_plug(&plug); |
e43473b7 | 1234 | } |
e43473b7 VG |
1235 | } |
1236 | ||
f95a04af TH |
1237 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
1238 | struct blkg_policy_data *pd, int off) | |
41b38b6d | 1239 | { |
f95a04af | 1240 | struct throtl_grp *tg = pd_to_tg(pd); |
41b38b6d TH |
1241 | struct blkg_rwstat rwstat = { }, tmp; |
1242 | int i, cpu; | |
1243 | ||
1244 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 1245 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
1246 | |
1247 | tmp = blkg_rwstat_read((void *)sc + off); | |
1248 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
1249 | rwstat.cnt[i] += tmp.cnt[i]; | |
1250 | } | |
1251 | ||
f95a04af | 1252 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
41b38b6d TH |
1253 | } |
1254 | ||
8a3d2615 TH |
1255 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
1256 | struct seq_file *sf) | |
41b38b6d | 1257 | { |
3c798398 | 1258 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
41b38b6d | 1259 | |
3c798398 | 1260 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 1261 | cft->private, true); |
41b38b6d TH |
1262 | return 0; |
1263 | } | |
1264 | ||
f95a04af TH |
1265 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
1266 | int off) | |
60c2bc2d | 1267 | { |
f95a04af TH |
1268 | struct throtl_grp *tg = pd_to_tg(pd); |
1269 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 1270 | |
af133ceb | 1271 | if (v == -1) |
60c2bc2d | 1272 | return 0; |
f95a04af | 1273 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
1274 | } |
1275 | ||
f95a04af TH |
1276 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
1277 | int off) | |
e43473b7 | 1278 | { |
f95a04af TH |
1279 | struct throtl_grp *tg = pd_to_tg(pd); |
1280 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 1281 | |
af133ceb TH |
1282 | if (v == -1) |
1283 | return 0; | |
f95a04af | 1284 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
1285 | } |
1286 | ||
af133ceb TH |
1287 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1288 | struct seq_file *sf) | |
8e89d13f | 1289 | { |
3c798398 TH |
1290 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, |
1291 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 1292 | return 0; |
8e89d13f VG |
1293 | } |
1294 | ||
af133ceb TH |
1295 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1296 | struct seq_file *sf) | |
8e89d13f | 1297 | { |
3c798398 TH |
1298 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, |
1299 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 1300 | return 0; |
60c2bc2d TH |
1301 | } |
1302 | ||
af133ceb TH |
1303 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
1304 | bool is_u64) | |
60c2bc2d | 1305 | { |
3c798398 | 1306 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
60c2bc2d | 1307 | struct blkg_conf_ctx ctx; |
af133ceb | 1308 | struct throtl_grp *tg; |
69df0ab0 | 1309 | struct throtl_service_queue *sq; |
60c2bc2d TH |
1310 | int ret; |
1311 | ||
3c798398 | 1312 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
1313 | if (ret) |
1314 | return ret; | |
1315 | ||
af133ceb | 1316 | tg = blkg_to_tg(ctx.blkg); |
69df0ab0 | 1317 | sq = &tg->service_queue; |
af133ceb | 1318 | |
a2b1693b TH |
1319 | if (!ctx.v) |
1320 | ctx.v = -1; | |
af133ceb | 1321 | |
a2b1693b TH |
1322 | if (is_u64) |
1323 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
1324 | else | |
1325 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 1326 | |
fda6f272 TH |
1327 | throtl_log(&tg->service_queue, |
1328 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
1329 | tg->bps[READ], tg->bps[WRITE], | |
1330 | tg->iops[READ], tg->iops[WRITE]); | |
632b4493 TH |
1331 | |
1332 | /* | |
1333 | * We're already holding queue_lock and know @tg is valid. Let's | |
1334 | * apply the new config directly. | |
1335 | * | |
1336 | * Restart the slices for both READ and WRITES. It might happen | |
1337 | * that a group's limit are dropped suddenly and we don't want to | |
1338 | * account recently dispatched IO with new low rate. | |
1339 | */ | |
0f3457f6 TH |
1340 | throtl_start_new_slice(tg, 0); |
1341 | throtl_start_new_slice(tg, 1); | |
632b4493 | 1342 | |
5b2c16aa | 1343 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 1344 | tg_update_disptime(tg); |
7f52f98c | 1345 | throtl_schedule_next_dispatch(sq->parent_sq, true); |
632b4493 | 1346 | } |
60c2bc2d TH |
1347 | |
1348 | blkg_conf_finish(&ctx); | |
a2b1693b | 1349 | return 0; |
8e89d13f VG |
1350 | } |
1351 | ||
af133ceb TH |
1352 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
1353 | const char *buf) | |
60c2bc2d | 1354 | { |
af133ceb | 1355 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
1356 | } |
1357 | ||
af133ceb TH |
1358 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
1359 | const char *buf) | |
60c2bc2d | 1360 | { |
af133ceb | 1361 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
1362 | } |
1363 | ||
1364 | static struct cftype throtl_files[] = { | |
1365 | { | |
1366 | .name = "throttle.read_bps_device", | |
af133ceb TH |
1367 | .private = offsetof(struct throtl_grp, bps[READ]), |
1368 | .read_seq_string = tg_print_conf_u64, | |
1369 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1370 | .max_write_len = 256, |
1371 | }, | |
1372 | { | |
1373 | .name = "throttle.write_bps_device", | |
af133ceb TH |
1374 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
1375 | .read_seq_string = tg_print_conf_u64, | |
1376 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1377 | .max_write_len = 256, |
1378 | }, | |
1379 | { | |
1380 | .name = "throttle.read_iops_device", | |
af133ceb TH |
1381 | .private = offsetof(struct throtl_grp, iops[READ]), |
1382 | .read_seq_string = tg_print_conf_uint, | |
1383 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1384 | .max_write_len = 256, |
1385 | }, | |
1386 | { | |
1387 | .name = "throttle.write_iops_device", | |
af133ceb TH |
1388 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
1389 | .read_seq_string = tg_print_conf_uint, | |
1390 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1391 | .max_write_len = 256, |
1392 | }, | |
1393 | { | |
1394 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1395 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1396 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1397 | }, |
1398 | { | |
1399 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1400 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1401 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1402 | }, |
1403 | { } /* terminate */ | |
1404 | }; | |
1405 | ||
da527770 | 1406 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1407 | { |
1408 | struct throtl_data *td = q->td; | |
1409 | ||
69df0ab0 | 1410 | cancel_work_sync(&td->dispatch_work); |
e43473b7 VG |
1411 | } |
1412 | ||
3c798398 | 1413 | static struct blkcg_policy blkcg_policy_throtl = { |
f9fcc2d3 TH |
1414 | .pd_size = sizeof(struct throtl_grp), |
1415 | .cftypes = throtl_files, | |
1416 | ||
1417 | .pd_init_fn = throtl_pd_init, | |
1418 | .pd_exit_fn = throtl_pd_exit, | |
1419 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 VG |
1420 | }; |
1421 | ||
bc16a4f9 | 1422 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1423 | { |
1424 | struct throtl_data *td = q->td; | |
c5cc2070 | 1425 | struct throtl_qnode *qn = NULL; |
e43473b7 | 1426 | struct throtl_grp *tg; |
73f0d49a | 1427 | struct throtl_service_queue *sq; |
0e9f4164 | 1428 | bool rw = bio_data_dir(bio); |
3c798398 | 1429 | struct blkcg *blkcg; |
bc16a4f9 | 1430 | bool throttled = false; |
e43473b7 | 1431 | |
2a0f61e6 TH |
1432 | /* see throtl_charge_bio() */ |
1433 | if (bio->bi_rw & REQ_THROTTLED) | |
bc16a4f9 | 1434 | goto out; |
e43473b7 | 1435 | |
af75cd3c VG |
1436 | /* |
1437 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1438 | * basic fields like stats and io rates. If a group has no rules, | |
1439 | * just update the dispatch stats in lockless manner and return. | |
1440 | */ | |
af75cd3c | 1441 | rcu_read_lock(); |
3c798398 | 1442 | blkcg = bio_blkcg(bio); |
cd1604fa | 1443 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1444 | if (tg) { |
af75cd3c | 1445 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1446 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1447 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1448 | goto out_unlock_rcu; |
af75cd3c VG |
1449 | } |
1450 | } | |
af75cd3c VG |
1451 | |
1452 | /* | |
1453 | * Either group has not been allocated yet or it is not an unlimited | |
1454 | * IO group | |
1455 | */ | |
e43473b7 | 1456 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1457 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1458 | if (unlikely(!tg)) |
1459 | goto out_unlock; | |
f469a7b4 | 1460 | |
73f0d49a TH |
1461 | sq = &tg->service_queue; |
1462 | ||
9e660acf TH |
1463 | while (true) { |
1464 | /* throtl is FIFO - if bios are already queued, should queue */ | |
1465 | if (sq->nr_queued[rw]) | |
1466 | break; | |
de701c74 | 1467 | |
9e660acf TH |
1468 | /* if above limits, break to queue */ |
1469 | if (!tg_may_dispatch(tg, bio, NULL)) | |
1470 | break; | |
1471 | ||
1472 | /* within limits, let's charge and dispatch directly */ | |
e43473b7 | 1473 | throtl_charge_bio(tg, bio); |
04521db0 VG |
1474 | |
1475 | /* | |
1476 | * We need to trim slice even when bios are not being queued | |
1477 | * otherwise it might happen that a bio is not queued for | |
1478 | * a long time and slice keeps on extending and trim is not | |
1479 | * called for a long time. Now if limits are reduced suddenly | |
1480 | * we take into account all the IO dispatched so far at new | |
1481 | * low rate and * newly queued IO gets a really long dispatch | |
1482 | * time. | |
1483 | * | |
1484 | * So keep on trimming slice even if bio is not queued. | |
1485 | */ | |
0f3457f6 | 1486 | throtl_trim_slice(tg, rw); |
9e660acf TH |
1487 | |
1488 | /* | |
1489 | * @bio passed through this layer without being throttled. | |
1490 | * Climb up the ladder. If we''re already at the top, it | |
1491 | * can be executed directly. | |
1492 | */ | |
c5cc2070 | 1493 | qn = &tg->qnode_on_parent[rw]; |
9e660acf TH |
1494 | sq = sq->parent_sq; |
1495 | tg = sq_to_tg(sq); | |
1496 | if (!tg) | |
1497 | goto out_unlock; | |
e43473b7 VG |
1498 | } |
1499 | ||
9e660acf | 1500 | /* out-of-limit, queue to @tg */ |
fda6f272 TH |
1501 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
1502 | rw == READ ? 'R' : 'W', | |
1503 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], | |
1504 | tg->io_disp[rw], tg->iops[rw], | |
1505 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
e43473b7 | 1506 | |
671058fb | 1507 | bio_associate_current(bio); |
6bc9c2b4 | 1508 | tg->td->nr_queued[rw]++; |
c5cc2070 | 1509 | throtl_add_bio_tg(bio, qn, tg); |
bc16a4f9 | 1510 | throttled = true; |
e43473b7 | 1511 | |
7f52f98c TH |
1512 | /* |
1513 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
1514 | * was empty before @bio. The forced scheduling isn't likely to | |
1515 | * cause undue delay as @bio is likely to be dispatched directly if | |
1516 | * its @tg's disptime is not in the future. | |
1517 | */ | |
0e9f4164 | 1518 | if (tg->flags & THROTL_TG_WAS_EMPTY) { |
77216b04 | 1519 | tg_update_disptime(tg); |
7f52f98c | 1520 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); |
e43473b7 VG |
1521 | } |
1522 | ||
bc16a4f9 | 1523 | out_unlock: |
e43473b7 | 1524 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1525 | out_unlock_rcu: |
1526 | rcu_read_unlock(); | |
bc16a4f9 | 1527 | out: |
2a0f61e6 TH |
1528 | /* |
1529 | * As multiple blk-throtls may stack in the same issue path, we | |
1530 | * don't want bios to leave with the flag set. Clear the flag if | |
1531 | * being issued. | |
1532 | */ | |
1533 | if (!throttled) | |
1534 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1535 | return throttled; |
e43473b7 VG |
1536 | } |
1537 | ||
2a12f0dc TH |
1538 | /* |
1539 | * Dispatch all bios from all children tg's queued on @parent_sq. On | |
1540 | * return, @parent_sq is guaranteed to not have any active children tg's | |
1541 | * and all bios from previously active tg's are on @parent_sq->bio_lists[]. | |
1542 | */ | |
1543 | static void tg_drain_bios(struct throtl_service_queue *parent_sq) | |
1544 | { | |
1545 | struct throtl_grp *tg; | |
1546 | ||
1547 | while ((tg = throtl_rb_first(parent_sq))) { | |
1548 | struct throtl_service_queue *sq = &tg->service_queue; | |
1549 | struct bio *bio; | |
1550 | ||
1551 | throtl_dequeue_tg(tg); | |
1552 | ||
c5cc2070 | 1553 | while ((bio = throtl_peek_queued(&sq->queued[READ]))) |
2a12f0dc | 1554 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
c5cc2070 | 1555 | while ((bio = throtl_peek_queued(&sq->queued[WRITE]))) |
2a12f0dc TH |
1556 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
1557 | } | |
1558 | } | |
1559 | ||
c9a929dd TH |
1560 | /** |
1561 | * blk_throtl_drain - drain throttled bios | |
1562 | * @q: request_queue to drain throttled bios for | |
1563 | * | |
1564 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1565 | */ | |
1566 | void blk_throtl_drain(struct request_queue *q) | |
1567 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1568 | { | |
1569 | struct throtl_data *td = q->td; | |
2a12f0dc TH |
1570 | struct blkcg_gq *blkg; |
1571 | struct cgroup *pos_cgrp; | |
c9a929dd | 1572 | struct bio *bio; |
651930bc | 1573 | int rw; |
c9a929dd | 1574 | |
8bcb6c7d | 1575 | queue_lockdep_assert_held(q); |
2a12f0dc | 1576 | rcu_read_lock(); |
c9a929dd | 1577 | |
2a12f0dc TH |
1578 | /* |
1579 | * Drain each tg while doing post-order walk on the blkg tree, so | |
1580 | * that all bios are propagated to td->service_queue. It'd be | |
1581 | * better to walk service_queue tree directly but blkg walk is | |
1582 | * easier. | |
1583 | */ | |
1584 | blkg_for_each_descendant_post(blkg, pos_cgrp, td->queue->root_blkg) | |
1585 | tg_drain_bios(&blkg_to_tg(blkg)->service_queue); | |
73f0d49a | 1586 | |
2a12f0dc | 1587 | tg_drain_bios(&td_root_tg(td)->service_queue); |
c9a929dd | 1588 | |
2a12f0dc TH |
1589 | /* finally, transfer bios from top-level tg's into the td */ |
1590 | tg_drain_bios(&td->service_queue); | |
1591 | ||
1592 | rcu_read_unlock(); | |
c9a929dd TH |
1593 | spin_unlock_irq(q->queue_lock); |
1594 | ||
2a12f0dc | 1595 | /* all bios now should be in td->service_queue, issue them */ |
651930bc | 1596 | for (rw = READ; rw <= WRITE; rw++) |
c5cc2070 TH |
1597 | while ((bio = throtl_pop_queued(&td->service_queue.queued[rw], |
1598 | NULL))) | |
651930bc | 1599 | generic_make_request(bio); |
c9a929dd TH |
1600 | |
1601 | spin_lock_irq(q->queue_lock); | |
1602 | } | |
1603 | ||
e43473b7 VG |
1604 | int blk_throtl_init(struct request_queue *q) |
1605 | { | |
1606 | struct throtl_data *td; | |
a2b1693b | 1607 | int ret; |
e43473b7 VG |
1608 | |
1609 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1610 | if (!td) | |
1611 | return -ENOMEM; | |
1612 | ||
69df0ab0 | 1613 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
77216b04 | 1614 | throtl_service_queue_init(&td->service_queue, NULL); |
e43473b7 | 1615 | |
cd1604fa | 1616 | q->td = td; |
29b12589 | 1617 | td->queue = q; |
02977e4a | 1618 | |
a2b1693b | 1619 | /* activate policy */ |
3c798398 | 1620 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1621 | if (ret) |
f51b802c | 1622 | kfree(td); |
a2b1693b | 1623 | return ret; |
e43473b7 VG |
1624 | } |
1625 | ||
1626 | void blk_throtl_exit(struct request_queue *q) | |
1627 | { | |
c875f4d0 | 1628 | BUG_ON(!q->td); |
da527770 | 1629 | throtl_shutdown_wq(q); |
3c798398 | 1630 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1631 | kfree(q->td); |
e43473b7 VG |
1632 | } |
1633 | ||
1634 | static int __init throtl_init(void) | |
1635 | { | |
450adcbe VG |
1636 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1637 | if (!kthrotld_workqueue) | |
1638 | panic("Failed to create kthrotld\n"); | |
1639 | ||
3c798398 | 1640 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1641 | } |
1642 | ||
1643 | module_init(throtl_init); |