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 | ||
693e751e TH |
127 | /* are there any throtl rules between this group and td? */ |
128 | bool has_rules[2]; | |
129 | ||
e43473b7 VG |
130 | /* bytes per second rate limits */ |
131 | uint64_t bps[2]; | |
132 | ||
8e89d13f VG |
133 | /* IOPS limits */ |
134 | unsigned int iops[2]; | |
135 | ||
e43473b7 VG |
136 | /* Number of bytes disptached in current slice */ |
137 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
138 | /* Number of bio's dispatched in current slice */ |
139 | unsigned int io_disp[2]; | |
e43473b7 VG |
140 | |
141 | /* When did we start a new slice */ | |
142 | unsigned long slice_start[2]; | |
143 | unsigned long slice_end[2]; | |
fe071437 | 144 | |
8a3d2615 TH |
145 | /* Per cpu stats pointer */ |
146 | struct tg_stats_cpu __percpu *stats_cpu; | |
147 | ||
148 | /* List of tgs waiting for per cpu stats memory to be allocated */ | |
149 | struct list_head stats_alloc_node; | |
e43473b7 VG |
150 | }; |
151 | ||
152 | struct throtl_data | |
153 | { | |
e43473b7 | 154 | /* service tree for active throtl groups */ |
c9e0332e | 155 | struct throtl_service_queue service_queue; |
e43473b7 | 156 | |
e43473b7 VG |
157 | struct request_queue *queue; |
158 | ||
159 | /* Total Number of queued bios on READ and WRITE lists */ | |
160 | unsigned int nr_queued[2]; | |
161 | ||
162 | /* | |
02977e4a | 163 | * number of total undestroyed groups |
e43473b7 VG |
164 | */ |
165 | unsigned int nr_undestroyed_grps; | |
166 | ||
167 | /* Work for dispatching throttled bios */ | |
69df0ab0 | 168 | struct work_struct dispatch_work; |
e43473b7 VG |
169 | }; |
170 | ||
8a3d2615 TH |
171 | /* list and work item to allocate percpu group stats */ |
172 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
173 | static LIST_HEAD(tg_stats_alloc_list); | |
174 | ||
175 | static void tg_stats_alloc_fn(struct work_struct *); | |
176 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
177 | ||
69df0ab0 TH |
178 | static void throtl_pending_timer_fn(unsigned long arg); |
179 | ||
f95a04af TH |
180 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
181 | { | |
182 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
183 | } | |
184 | ||
3c798398 | 185 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 186 | { |
f95a04af | 187 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
188 | } |
189 | ||
3c798398 | 190 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 191 | { |
f95a04af | 192 | return pd_to_blkg(&tg->pd); |
0381411e TH |
193 | } |
194 | ||
03d8e111 TH |
195 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) |
196 | { | |
197 | return blkg_to_tg(td->queue->root_blkg); | |
198 | } | |
199 | ||
fda6f272 TH |
200 | /** |
201 | * sq_to_tg - return the throl_grp the specified service queue belongs to | |
202 | * @sq: the throtl_service_queue of interest | |
203 | * | |
204 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one | |
205 | * embedded in throtl_data, %NULL is returned. | |
206 | */ | |
207 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) | |
208 | { | |
209 | if (sq && sq->parent_sq) | |
210 | return container_of(sq, struct throtl_grp, service_queue); | |
211 | else | |
212 | return NULL; | |
213 | } | |
214 | ||
215 | /** | |
216 | * sq_to_td - return throtl_data the specified service queue belongs to | |
217 | * @sq: the throtl_service_queue of interest | |
218 | * | |
219 | * A service_queue can be embeded in either a throtl_grp or throtl_data. | |
220 | * Determine the associated throtl_data accordingly and return it. | |
221 | */ | |
222 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) | |
223 | { | |
224 | struct throtl_grp *tg = sq_to_tg(sq); | |
225 | ||
226 | if (tg) | |
227 | return tg->td; | |
228 | else | |
229 | return container_of(sq, struct throtl_data, service_queue); | |
230 | } | |
231 | ||
232 | /** | |
233 | * throtl_log - log debug message via blktrace | |
234 | * @sq: the service_queue being reported | |
235 | * @fmt: printf format string | |
236 | * @args: printf args | |
237 | * | |
238 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a | |
239 | * throtl_grp; otherwise, just "throtl". | |
240 | * | |
241 | * TODO: this should be made a function and name formatting should happen | |
242 | * after testing whether blktrace is enabled. | |
243 | */ | |
244 | #define throtl_log(sq, fmt, args...) do { \ | |
245 | struct throtl_grp *__tg = sq_to_tg((sq)); \ | |
246 | struct throtl_data *__td = sq_to_td((sq)); \ | |
247 | \ | |
248 | (void)__td; \ | |
249 | if ((__tg)) { \ | |
250 | char __pbuf[128]; \ | |
54e7ed12 | 251 | \ |
fda6f272 TH |
252 | blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \ |
253 | blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
254 | } else { \ | |
255 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
256 | } \ | |
54e7ed12 | 257 | } while (0) |
e43473b7 | 258 | |
8a3d2615 TH |
259 | /* |
260 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
3b07e9ca | 261 | * system_wq once there are some groups on the alloc_list waiting for |
8a3d2615 TH |
262 | * allocation. |
263 | */ | |
264 | static void tg_stats_alloc_fn(struct work_struct *work) | |
265 | { | |
266 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
267 | struct delayed_work *dwork = to_delayed_work(work); | |
268 | bool empty = false; | |
269 | ||
270 | alloc_stats: | |
271 | if (!stats_cpu) { | |
272 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
273 | if (!stats_cpu) { | |
274 | /* allocation failed, try again after some time */ | |
3b07e9ca | 275 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); |
8a3d2615 TH |
276 | return; |
277 | } | |
278 | } | |
279 | ||
280 | spin_lock_irq(&tg_stats_alloc_lock); | |
281 | ||
282 | if (!list_empty(&tg_stats_alloc_list)) { | |
283 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
284 | struct throtl_grp, | |
285 | stats_alloc_node); | |
286 | swap(tg->stats_cpu, stats_cpu); | |
287 | list_del_init(&tg->stats_alloc_node); | |
288 | } | |
289 | ||
290 | empty = list_empty(&tg_stats_alloc_list); | |
291 | spin_unlock_irq(&tg_stats_alloc_lock); | |
292 | if (!empty) | |
293 | goto alloc_stats; | |
294 | } | |
295 | ||
c5cc2070 TH |
296 | static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg) |
297 | { | |
298 | INIT_LIST_HEAD(&qn->node); | |
299 | bio_list_init(&qn->bios); | |
300 | qn->tg = tg; | |
301 | } | |
302 | ||
303 | /** | |
304 | * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it | |
305 | * @bio: bio being added | |
306 | * @qn: qnode to add bio to | |
307 | * @queued: the service_queue->queued[] list @qn belongs to | |
308 | * | |
309 | * Add @bio to @qn and put @qn on @queued if it's not already on. | |
310 | * @qn->tg's reference count is bumped when @qn is activated. See the | |
311 | * comment on top of throtl_qnode definition for details. | |
312 | */ | |
313 | static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn, | |
314 | struct list_head *queued) | |
315 | { | |
316 | bio_list_add(&qn->bios, bio); | |
317 | if (list_empty(&qn->node)) { | |
318 | list_add_tail(&qn->node, queued); | |
319 | blkg_get(tg_to_blkg(qn->tg)); | |
320 | } | |
321 | } | |
322 | ||
323 | /** | |
324 | * throtl_peek_queued - peek the first bio on a qnode list | |
325 | * @queued: the qnode list to peek | |
326 | */ | |
327 | static struct bio *throtl_peek_queued(struct list_head *queued) | |
328 | { | |
329 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
330 | struct bio *bio; | |
331 | ||
332 | if (list_empty(queued)) | |
333 | return NULL; | |
334 | ||
335 | bio = bio_list_peek(&qn->bios); | |
336 | WARN_ON_ONCE(!bio); | |
337 | return bio; | |
338 | } | |
339 | ||
340 | /** | |
341 | * throtl_pop_queued - pop the first bio form a qnode list | |
342 | * @queued: the qnode list to pop a bio from | |
343 | * @tg_to_put: optional out argument for throtl_grp to put | |
344 | * | |
345 | * Pop the first bio from the qnode list @queued. After popping, the first | |
346 | * qnode is removed from @queued if empty or moved to the end of @queued so | |
347 | * that the popping order is round-robin. | |
348 | * | |
349 | * When the first qnode is removed, its associated throtl_grp should be put | |
350 | * too. If @tg_to_put is NULL, this function automatically puts it; | |
351 | * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is | |
352 | * responsible for putting it. | |
353 | */ | |
354 | static struct bio *throtl_pop_queued(struct list_head *queued, | |
355 | struct throtl_grp **tg_to_put) | |
356 | { | |
357 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
358 | struct bio *bio; | |
359 | ||
360 | if (list_empty(queued)) | |
361 | return NULL; | |
362 | ||
363 | bio = bio_list_pop(&qn->bios); | |
364 | WARN_ON_ONCE(!bio); | |
365 | ||
366 | if (bio_list_empty(&qn->bios)) { | |
367 | list_del_init(&qn->node); | |
368 | if (tg_to_put) | |
369 | *tg_to_put = qn->tg; | |
370 | else | |
371 | blkg_put(tg_to_blkg(qn->tg)); | |
372 | } else { | |
373 | list_move_tail(&qn->node, queued); | |
374 | } | |
375 | ||
376 | return bio; | |
377 | } | |
378 | ||
49a2f1e3 | 379 | /* init a service_queue, assumes the caller zeroed it */ |
77216b04 TH |
380 | static void throtl_service_queue_init(struct throtl_service_queue *sq, |
381 | struct throtl_service_queue *parent_sq) | |
49a2f1e3 | 382 | { |
c5cc2070 TH |
383 | INIT_LIST_HEAD(&sq->queued[0]); |
384 | INIT_LIST_HEAD(&sq->queued[1]); | |
49a2f1e3 | 385 | sq->pending_tree = RB_ROOT; |
77216b04 | 386 | sq->parent_sq = parent_sq; |
69df0ab0 TH |
387 | setup_timer(&sq->pending_timer, throtl_pending_timer_fn, |
388 | (unsigned long)sq); | |
389 | } | |
390 | ||
391 | static void throtl_service_queue_exit(struct throtl_service_queue *sq) | |
392 | { | |
393 | del_timer_sync(&sq->pending_timer); | |
49a2f1e3 TH |
394 | } |
395 | ||
3c798398 | 396 | static void throtl_pd_init(struct blkcg_gq *blkg) |
a29a171e | 397 | { |
0381411e | 398 | struct throtl_grp *tg = blkg_to_tg(blkg); |
77216b04 | 399 | struct throtl_data *td = blkg->q->td; |
9138125b | 400 | struct throtl_service_queue *parent_sq; |
ff26eaad | 401 | unsigned long flags; |
c5cc2070 | 402 | int rw; |
cd1604fa | 403 | |
9138125b TH |
404 | /* |
405 | * If sane_hierarchy is enabled, we switch to properly hierarchical | |
406 | * behavior where limits on a given throtl_grp are applied to the | |
407 | * whole subtree rather than just the group itself. e.g. If 16M | |
408 | * read_bps limit is set on the root group, the whole system can't | |
409 | * exceed 16M for the device. | |
410 | * | |
411 | * If sane_hierarchy is not enabled, the broken flat hierarchy | |
412 | * behavior is retained where all throtl_grps are treated as if | |
413 | * they're all separate root groups right below throtl_data. | |
414 | * Limits of a group don't interact with limits of other groups | |
415 | * regardless of the position of the group in the hierarchy. | |
416 | */ | |
417 | parent_sq = &td->service_queue; | |
418 | ||
419 | if (cgroup_sane_behavior(blkg->blkcg->css.cgroup) && blkg->parent) | |
420 | parent_sq = &blkg_to_tg(blkg->parent)->service_queue; | |
421 | ||
422 | throtl_service_queue_init(&tg->service_queue, parent_sq); | |
423 | ||
c5cc2070 TH |
424 | for (rw = READ; rw <= WRITE; rw++) { |
425 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); | |
426 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); | |
427 | } | |
428 | ||
a29a171e | 429 | RB_CLEAR_NODE(&tg->rb_node); |
77216b04 | 430 | tg->td = td; |
a29a171e | 431 | |
e56da7e2 TH |
432 | tg->bps[READ] = -1; |
433 | tg->bps[WRITE] = -1; | |
434 | tg->iops[READ] = -1; | |
435 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
436 | |
437 | /* | |
438 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
439 | * but percpu allocator can't be called from IO path. Queue tg on | |
440 | * tg_stats_alloc_list and allocate from work item. | |
441 | */ | |
ff26eaad | 442 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 443 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); |
3b07e9ca | 444 | schedule_delayed_work(&tg_stats_alloc_work, 0); |
ff26eaad | 445 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
446 | } |
447 | ||
693e751e TH |
448 | /* |
449 | * Set has_rules[] if @tg or any of its parents have limits configured. | |
450 | * This doesn't require walking up to the top of the hierarchy as the | |
451 | * parent's has_rules[] is guaranteed to be correct. | |
452 | */ | |
453 | static void tg_update_has_rules(struct throtl_grp *tg) | |
454 | { | |
455 | struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); | |
456 | int rw; | |
457 | ||
458 | for (rw = READ; rw <= WRITE; rw++) | |
459 | tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) || | |
460 | (tg->bps[rw] != -1 || tg->iops[rw] != -1); | |
461 | } | |
462 | ||
463 | static void throtl_pd_online(struct blkcg_gq *blkg) | |
464 | { | |
465 | /* | |
466 | * We don't want new groups to escape the limits of its ancestors. | |
467 | * Update has_rules[] after a new group is brought online. | |
468 | */ | |
469 | tg_update_has_rules(blkg_to_tg(blkg)); | |
470 | } | |
471 | ||
3c798398 | 472 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
8a3d2615 TH |
473 | { |
474 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
ff26eaad | 475 | unsigned long flags; |
8a3d2615 | 476 | |
ff26eaad | 477 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 478 | list_del_init(&tg->stats_alloc_node); |
ff26eaad | 479 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
480 | |
481 | free_percpu(tg->stats_cpu); | |
69df0ab0 TH |
482 | |
483 | throtl_service_queue_exit(&tg->service_queue); | |
8a3d2615 TH |
484 | } |
485 | ||
3c798398 | 486 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
8a3d2615 TH |
487 | { |
488 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
489 | int cpu; | |
490 | ||
491 | if (tg->stats_cpu == NULL) | |
492 | return; | |
493 | ||
494 | for_each_possible_cpu(cpu) { | |
495 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
496 | ||
497 | blkg_rwstat_reset(&sc->service_bytes); | |
498 | blkg_rwstat_reset(&sc->serviced); | |
499 | } | |
a29a171e VG |
500 | } |
501 | ||
3c798398 TH |
502 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
503 | struct blkcg *blkcg) | |
e43473b7 | 504 | { |
be2c6b19 | 505 | /* |
3c798398 TH |
506 | * This is the common case when there are no blkcgs. Avoid lookup |
507 | * in this case | |
cd1604fa | 508 | */ |
3c798398 | 509 | if (blkcg == &blkcg_root) |
03d8e111 | 510 | return td_root_tg(td); |
e43473b7 | 511 | |
e8989fae | 512 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
513 | } |
514 | ||
cd1604fa | 515 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
3c798398 | 516 | struct blkcg *blkcg) |
e43473b7 | 517 | { |
f469a7b4 | 518 | struct request_queue *q = td->queue; |
cd1604fa | 519 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 520 | |
f469a7b4 | 521 | /* |
3c798398 TH |
522 | * This is the common case when there are no blkcgs. Avoid lookup |
523 | * in this case | |
f469a7b4 | 524 | */ |
3c798398 | 525 | if (blkcg == &blkcg_root) { |
03d8e111 | 526 | tg = td_root_tg(td); |
cd1604fa | 527 | } else { |
3c798398 | 528 | struct blkcg_gq *blkg; |
f469a7b4 | 529 | |
3c96cb32 | 530 | blkg = blkg_lookup_create(blkcg, q); |
f469a7b4 | 531 | |
cd1604fa TH |
532 | /* if %NULL and @q is alive, fall back to root_tg */ |
533 | if (!IS_ERR(blkg)) | |
0381411e | 534 | tg = blkg_to_tg(blkg); |
3f3299d5 | 535 | else if (!blk_queue_dying(q)) |
03d8e111 | 536 | tg = td_root_tg(td); |
f469a7b4 VG |
537 | } |
538 | ||
e43473b7 VG |
539 | return tg; |
540 | } | |
541 | ||
0049af73 TH |
542 | static struct throtl_grp * |
543 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
544 | { |
545 | /* Service tree is empty */ | |
0049af73 | 546 | if (!parent_sq->nr_pending) |
e43473b7 VG |
547 | return NULL; |
548 | ||
0049af73 TH |
549 | if (!parent_sq->first_pending) |
550 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
e43473b7 | 551 | |
0049af73 TH |
552 | if (parent_sq->first_pending) |
553 | return rb_entry_tg(parent_sq->first_pending); | |
e43473b7 VG |
554 | |
555 | return NULL; | |
556 | } | |
557 | ||
558 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
559 | { | |
560 | rb_erase(n, root); | |
561 | RB_CLEAR_NODE(n); | |
562 | } | |
563 | ||
0049af73 TH |
564 | static void throtl_rb_erase(struct rb_node *n, |
565 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 566 | { |
0049af73 TH |
567 | if (parent_sq->first_pending == n) |
568 | parent_sq->first_pending = NULL; | |
569 | rb_erase_init(n, &parent_sq->pending_tree); | |
570 | --parent_sq->nr_pending; | |
e43473b7 VG |
571 | } |
572 | ||
0049af73 | 573 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
574 | { |
575 | struct throtl_grp *tg; | |
576 | ||
0049af73 | 577 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
578 | if (!tg) |
579 | return; | |
580 | ||
0049af73 | 581 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
582 | } |
583 | ||
77216b04 | 584 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 585 | { |
77216b04 | 586 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
0049af73 | 587 | struct rb_node **node = &parent_sq->pending_tree.rb_node; |
e43473b7 VG |
588 | struct rb_node *parent = NULL; |
589 | struct throtl_grp *__tg; | |
590 | unsigned long key = tg->disptime; | |
591 | int left = 1; | |
592 | ||
593 | while (*node != NULL) { | |
594 | parent = *node; | |
595 | __tg = rb_entry_tg(parent); | |
596 | ||
597 | if (time_before(key, __tg->disptime)) | |
598 | node = &parent->rb_left; | |
599 | else { | |
600 | node = &parent->rb_right; | |
601 | left = 0; | |
602 | } | |
603 | } | |
604 | ||
605 | if (left) | |
0049af73 | 606 | parent_sq->first_pending = &tg->rb_node; |
e43473b7 VG |
607 | |
608 | rb_link_node(&tg->rb_node, parent, node); | |
0049af73 | 609 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); |
e43473b7 VG |
610 | } |
611 | ||
77216b04 | 612 | static void __throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 613 | { |
77216b04 | 614 | tg_service_queue_add(tg); |
5b2c16aa | 615 | tg->flags |= THROTL_TG_PENDING; |
77216b04 | 616 | tg->service_queue.parent_sq->nr_pending++; |
e43473b7 VG |
617 | } |
618 | ||
77216b04 | 619 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 620 | { |
5b2c16aa | 621 | if (!(tg->flags & THROTL_TG_PENDING)) |
77216b04 | 622 | __throtl_enqueue_tg(tg); |
e43473b7 VG |
623 | } |
624 | ||
77216b04 | 625 | static void __throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 626 | { |
77216b04 | 627 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
5b2c16aa | 628 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
629 | } |
630 | ||
77216b04 | 631 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 632 | { |
5b2c16aa | 633 | if (tg->flags & THROTL_TG_PENDING) |
77216b04 | 634 | __throtl_dequeue_tg(tg); |
e43473b7 VG |
635 | } |
636 | ||
a9131a27 | 637 | /* Call with queue lock held */ |
69df0ab0 TH |
638 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, |
639 | unsigned long expires) | |
a9131a27 | 640 | { |
69df0ab0 TH |
641 | mod_timer(&sq->pending_timer, expires); |
642 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
643 | expires - jiffies, jiffies); | |
a9131a27 TH |
644 | } |
645 | ||
7f52f98c TH |
646 | /** |
647 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle | |
648 | * @sq: the service_queue to schedule dispatch for | |
649 | * @force: force scheduling | |
650 | * | |
651 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the | |
652 | * dispatch time of the first pending child. Returns %true if either timer | |
653 | * is armed or there's no pending child left. %false if the current | |
654 | * dispatch window is still open and the caller should continue | |
655 | * dispatching. | |
656 | * | |
657 | * If @force is %true, the dispatch timer is always scheduled and this | |
658 | * function is guaranteed to return %true. This is to be used when the | |
659 | * caller can't dispatch itself and needs to invoke pending_timer | |
660 | * unconditionally. Note that forced scheduling is likely to induce short | |
661 | * delay before dispatch starts even if @sq->first_pending_disptime is not | |
662 | * in the future and thus shouldn't be used in hot paths. | |
663 | */ | |
664 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
665 | bool force) | |
e43473b7 | 666 | { |
6a525600 | 667 | /* any pending children left? */ |
c9e0332e | 668 | if (!sq->nr_pending) |
7f52f98c | 669 | return true; |
e43473b7 | 670 | |
c9e0332e | 671 | update_min_dispatch_time(sq); |
e43473b7 | 672 | |
69df0ab0 | 673 | /* is the next dispatch time in the future? */ |
7f52f98c | 674 | if (force || time_after(sq->first_pending_disptime, jiffies)) { |
69df0ab0 | 675 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); |
7f52f98c | 676 | return true; |
69df0ab0 TH |
677 | } |
678 | ||
7f52f98c TH |
679 | /* tell the caller to continue dispatching */ |
680 | return false; | |
e43473b7 VG |
681 | } |
682 | ||
32ee5bc4 VG |
683 | static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, |
684 | bool rw, unsigned long start) | |
685 | { | |
686 | tg->bytes_disp[rw] = 0; | |
687 | tg->io_disp[rw] = 0; | |
688 | ||
689 | /* | |
690 | * Previous slice has expired. We must have trimmed it after last | |
691 | * bio dispatch. That means since start of last slice, we never used | |
692 | * that bandwidth. Do try to make use of that bandwidth while giving | |
693 | * credit. | |
694 | */ | |
695 | if (time_after_eq(start, tg->slice_start[rw])) | |
696 | tg->slice_start[rw] = start; | |
697 | ||
698 | tg->slice_end[rw] = jiffies + throtl_slice; | |
699 | throtl_log(&tg->service_queue, | |
700 | "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", | |
701 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
702 | tg->slice_end[rw], jiffies); | |
703 | } | |
704 | ||
0f3457f6 | 705 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
706 | { |
707 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 708 | tg->io_disp[rw] = 0; |
e43473b7 VG |
709 | tg->slice_start[rw] = jiffies; |
710 | tg->slice_end[rw] = jiffies + throtl_slice; | |
fda6f272 TH |
711 | throtl_log(&tg->service_queue, |
712 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
713 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
714 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
715 | } |
716 | ||
0f3457f6 TH |
717 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
718 | unsigned long jiffy_end) | |
d1ae8ffd VG |
719 | { |
720 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
721 | } | |
722 | ||
0f3457f6 TH |
723 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
724 | unsigned long jiffy_end) | |
e43473b7 VG |
725 | { |
726 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
fda6f272 TH |
727 | throtl_log(&tg->service_queue, |
728 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
729 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
730 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
731 | } |
732 | ||
733 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 734 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
735 | { |
736 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
737 | return 0; | |
738 | ||
739 | return 1; | |
740 | } | |
741 | ||
742 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 743 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 744 | { |
3aad5d3e VG |
745 | unsigned long nr_slices, time_elapsed, io_trim; |
746 | u64 bytes_trim, tmp; | |
e43473b7 VG |
747 | |
748 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
749 | ||
750 | /* | |
751 | * If bps are unlimited (-1), then time slice don't get | |
752 | * renewed. Don't try to trim the slice if slice is used. A new | |
753 | * slice will start when appropriate. | |
754 | */ | |
0f3457f6 | 755 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
756 | return; |
757 | ||
d1ae8ffd VG |
758 | /* |
759 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
760 | * that initially cgroup limit was very low resulting in high | |
761 | * slice_end, but later limit was bumped up and bio was dispached | |
762 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
763 | * is bad because it does not allow new slice to start. | |
764 | */ | |
765 | ||
0f3457f6 | 766 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); |
d1ae8ffd | 767 | |
e43473b7 VG |
768 | time_elapsed = jiffies - tg->slice_start[rw]; |
769 | ||
770 | nr_slices = time_elapsed / throtl_slice; | |
771 | ||
772 | if (!nr_slices) | |
773 | return; | |
3aad5d3e VG |
774 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
775 | do_div(tmp, HZ); | |
776 | bytes_trim = tmp; | |
e43473b7 | 777 | |
8e89d13f | 778 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 779 | |
8e89d13f | 780 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
781 | return; |
782 | ||
783 | if (tg->bytes_disp[rw] >= bytes_trim) | |
784 | tg->bytes_disp[rw] -= bytes_trim; | |
785 | else | |
786 | tg->bytes_disp[rw] = 0; | |
787 | ||
8e89d13f VG |
788 | if (tg->io_disp[rw] >= io_trim) |
789 | tg->io_disp[rw] -= io_trim; | |
790 | else | |
791 | tg->io_disp[rw] = 0; | |
792 | ||
e43473b7 VG |
793 | tg->slice_start[rw] += nr_slices * throtl_slice; |
794 | ||
fda6f272 TH |
795 | throtl_log(&tg->service_queue, |
796 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
797 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
798 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
e43473b7 VG |
799 | } |
800 | ||
0f3457f6 TH |
801 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
802 | unsigned long *wait) | |
e43473b7 VG |
803 | { |
804 | bool rw = bio_data_dir(bio); | |
8e89d13f | 805 | unsigned int io_allowed; |
e43473b7 | 806 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 807 | u64 tmp; |
e43473b7 | 808 | |
8e89d13f | 809 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 810 | |
8e89d13f VG |
811 | /* Slice has just started. Consider one slice interval */ |
812 | if (!jiffy_elapsed) | |
813 | jiffy_elapsed_rnd = throtl_slice; | |
814 | ||
815 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
816 | ||
c49c06e4 VG |
817 | /* |
818 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
819 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
820 | * will allow dispatch after 1 second and after that slice should | |
821 | * have been trimmed. | |
822 | */ | |
823 | ||
824 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
825 | do_div(tmp, HZ); | |
826 | ||
827 | if (tmp > UINT_MAX) | |
828 | io_allowed = UINT_MAX; | |
829 | else | |
830 | io_allowed = tmp; | |
8e89d13f VG |
831 | |
832 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
833 | if (wait) |
834 | *wait = 0; | |
835 | return 1; | |
836 | } | |
837 | ||
8e89d13f VG |
838 | /* Calc approx time to dispatch */ |
839 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
840 | ||
841 | if (jiffy_wait > jiffy_elapsed) | |
842 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
843 | else | |
844 | jiffy_wait = 1; | |
845 | ||
846 | if (wait) | |
847 | *wait = jiffy_wait; | |
848 | return 0; | |
849 | } | |
850 | ||
0f3457f6 TH |
851 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
852 | unsigned long *wait) | |
8e89d13f VG |
853 | { |
854 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 855 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 856 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
857 | |
858 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
859 | ||
860 | /* Slice has just started. Consider one slice interval */ | |
861 | if (!jiffy_elapsed) | |
862 | jiffy_elapsed_rnd = throtl_slice; | |
863 | ||
864 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
865 | ||
5e901a2b VG |
866 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
867 | do_div(tmp, HZ); | |
3aad5d3e | 868 | bytes_allowed = tmp; |
e43473b7 VG |
869 | |
870 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
871 | if (wait) | |
872 | *wait = 0; | |
873 | return 1; | |
874 | } | |
875 | ||
876 | /* Calc approx time to dispatch */ | |
877 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
878 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
879 | ||
880 | if (!jiffy_wait) | |
881 | jiffy_wait = 1; | |
882 | ||
883 | /* | |
884 | * This wait time is without taking into consideration the rounding | |
885 | * up we did. Add that time also. | |
886 | */ | |
887 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
888 | if (wait) |
889 | *wait = jiffy_wait; | |
8e89d13f VG |
890 | return 0; |
891 | } | |
892 | ||
893 | /* | |
894 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
895 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
896 | */ | |
0f3457f6 TH |
897 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
898 | unsigned long *wait) | |
8e89d13f VG |
899 | { |
900 | bool rw = bio_data_dir(bio); | |
901 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
902 | ||
903 | /* | |
904 | * Currently whole state machine of group depends on first bio | |
905 | * queued in the group bio list. So one should not be calling | |
906 | * this function with a different bio if there are other bios | |
907 | * queued. | |
908 | */ | |
73f0d49a | 909 | BUG_ON(tg->service_queue.nr_queued[rw] && |
c5cc2070 | 910 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); |
e43473b7 | 911 | |
8e89d13f VG |
912 | /* If tg->bps = -1, then BW is unlimited */ |
913 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
914 | if (wait) | |
915 | *wait = 0; | |
916 | return 1; | |
917 | } | |
918 | ||
919 | /* | |
920 | * If previous slice expired, start a new one otherwise renew/extend | |
921 | * existing slice to make sure it is at least throtl_slice interval | |
922 | * long since now. | |
923 | */ | |
0f3457f6 TH |
924 | if (throtl_slice_used(tg, rw)) |
925 | throtl_start_new_slice(tg, rw); | |
8e89d13f VG |
926 | else { |
927 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
0f3457f6 | 928 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); |
8e89d13f VG |
929 | } |
930 | ||
0f3457f6 TH |
931 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
932 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
933 | if (wait) |
934 | *wait = 0; | |
935 | return 1; | |
936 | } | |
937 | ||
938 | max_wait = max(bps_wait, iops_wait); | |
939 | ||
940 | if (wait) | |
941 | *wait = max_wait; | |
942 | ||
943 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 944 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
945 | |
946 | return 0; | |
947 | } | |
948 | ||
3c798398 | 949 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
950 | int rw) |
951 | { | |
8a3d2615 TH |
952 | struct throtl_grp *tg = blkg_to_tg(blkg); |
953 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
954 | unsigned long flags; |
955 | ||
956 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 957 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
958 | return; |
959 | ||
960 | /* | |
961 | * Disabling interrupts to provide mutual exclusion between two | |
962 | * writes on same cpu. It probably is not needed for 64bit. Not | |
963 | * optimizing that case yet. | |
964 | */ | |
965 | local_irq_save(flags); | |
966 | ||
8a3d2615 | 967 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 968 | |
629ed0b1 TH |
969 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
970 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
971 | ||
972 | local_irq_restore(flags); | |
973 | } | |
974 | ||
e43473b7 VG |
975 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
976 | { | |
977 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
978 | |
979 | /* Charge the bio to the group */ | |
980 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 981 | tg->io_disp[rw]++; |
e43473b7 | 982 | |
2a0f61e6 TH |
983 | /* |
984 | * REQ_THROTTLED is used to prevent the same bio to be throttled | |
985 | * more than once as a throttled bio will go through blk-throtl the | |
986 | * second time when it eventually gets issued. Set it when a bio | |
987 | * is being charged to a tg. | |
988 | * | |
989 | * Dispatch stats aren't recursive and each @bio should only be | |
990 | * accounted by the @tg it was originally associated with. Let's | |
991 | * update the stats when setting REQ_THROTTLED for the first time | |
992 | * which is guaranteed to be for the @bio's original tg. | |
993 | */ | |
994 | if (!(bio->bi_rw & REQ_THROTTLED)) { | |
995 | bio->bi_rw |= REQ_THROTTLED; | |
996 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, | |
997 | bio->bi_rw); | |
998 | } | |
e43473b7 VG |
999 | } |
1000 | ||
c5cc2070 TH |
1001 | /** |
1002 | * throtl_add_bio_tg - add a bio to the specified throtl_grp | |
1003 | * @bio: bio to add | |
1004 | * @qn: qnode to use | |
1005 | * @tg: the target throtl_grp | |
1006 | * | |
1007 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, | |
1008 | * tg->qnode_on_self[] is used. | |
1009 | */ | |
1010 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, | |
1011 | struct throtl_grp *tg) | |
e43473b7 | 1012 | { |
73f0d49a | 1013 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1014 | bool rw = bio_data_dir(bio); |
1015 | ||
c5cc2070 TH |
1016 | if (!qn) |
1017 | qn = &tg->qnode_on_self[rw]; | |
1018 | ||
0e9f4164 TH |
1019 | /* |
1020 | * If @tg doesn't currently have any bios queued in the same | |
1021 | * direction, queueing @bio can change when @tg should be | |
1022 | * dispatched. Mark that @tg was empty. This is automatically | |
1023 | * cleaered on the next tg_update_disptime(). | |
1024 | */ | |
1025 | if (!sq->nr_queued[rw]) | |
1026 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
1027 | ||
c5cc2070 TH |
1028 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); |
1029 | ||
73f0d49a | 1030 | sq->nr_queued[rw]++; |
77216b04 | 1031 | throtl_enqueue_tg(tg); |
e43473b7 VG |
1032 | } |
1033 | ||
77216b04 | 1034 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 1035 | { |
73f0d49a | 1036 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1037 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
1038 | struct bio *bio; | |
1039 | ||
c5cc2070 | 1040 | if ((bio = throtl_peek_queued(&sq->queued[READ]))) |
0f3457f6 | 1041 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 1042 | |
c5cc2070 | 1043 | if ((bio = throtl_peek_queued(&sq->queued[WRITE]))) |
0f3457f6 | 1044 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
1045 | |
1046 | min_wait = min(read_wait, write_wait); | |
1047 | disptime = jiffies + min_wait; | |
1048 | ||
e43473b7 | 1049 | /* Update dispatch time */ |
77216b04 | 1050 | throtl_dequeue_tg(tg); |
e43473b7 | 1051 | tg->disptime = disptime; |
77216b04 | 1052 | throtl_enqueue_tg(tg); |
0e9f4164 TH |
1053 | |
1054 | /* see throtl_add_bio_tg() */ | |
1055 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
1056 | } |
1057 | ||
32ee5bc4 VG |
1058 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, |
1059 | struct throtl_grp *parent_tg, bool rw) | |
1060 | { | |
1061 | if (throtl_slice_used(parent_tg, rw)) { | |
1062 | throtl_start_new_slice_with_credit(parent_tg, rw, | |
1063 | child_tg->slice_start[rw]); | |
1064 | } | |
1065 | ||
1066 | } | |
1067 | ||
77216b04 | 1068 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 1069 | { |
73f0d49a | 1070 | struct throtl_service_queue *sq = &tg->service_queue; |
6bc9c2b4 TH |
1071 | struct throtl_service_queue *parent_sq = sq->parent_sq; |
1072 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
c5cc2070 | 1073 | struct throtl_grp *tg_to_put = NULL; |
e43473b7 VG |
1074 | struct bio *bio; |
1075 | ||
c5cc2070 TH |
1076 | /* |
1077 | * @bio is being transferred from @tg to @parent_sq. Popping a bio | |
1078 | * from @tg may put its reference and @parent_sq might end up | |
1079 | * getting released prematurely. Remember the tg to put and put it | |
1080 | * after @bio is transferred to @parent_sq. | |
1081 | */ | |
1082 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); | |
73f0d49a | 1083 | sq->nr_queued[rw]--; |
e43473b7 VG |
1084 | |
1085 | throtl_charge_bio(tg, bio); | |
6bc9c2b4 TH |
1086 | |
1087 | /* | |
1088 | * If our parent is another tg, we just need to transfer @bio to | |
1089 | * the parent using throtl_add_bio_tg(). If our parent is | |
1090 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
1091 | * bio_lists[] and decrease total number queued. The caller is | |
1092 | * responsible for issuing these bios. | |
1093 | */ | |
1094 | if (parent_tg) { | |
c5cc2070 | 1095 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); |
32ee5bc4 | 1096 | start_parent_slice_with_credit(tg, parent_tg, rw); |
6bc9c2b4 | 1097 | } else { |
c5cc2070 TH |
1098 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], |
1099 | &parent_sq->queued[rw]); | |
6bc9c2b4 TH |
1100 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
1101 | tg->td->nr_queued[rw]--; | |
1102 | } | |
e43473b7 | 1103 | |
0f3457f6 | 1104 | throtl_trim_slice(tg, rw); |
6bc9c2b4 | 1105 | |
c5cc2070 TH |
1106 | if (tg_to_put) |
1107 | blkg_put(tg_to_blkg(tg_to_put)); | |
e43473b7 VG |
1108 | } |
1109 | ||
77216b04 | 1110 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 1111 | { |
73f0d49a | 1112 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1113 | unsigned int nr_reads = 0, nr_writes = 0; |
1114 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 1115 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
1116 | struct bio *bio; |
1117 | ||
1118 | /* Try to dispatch 75% READS and 25% WRITES */ | |
1119 | ||
c5cc2070 | 1120 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && |
0f3457f6 | 1121 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1122 | |
77216b04 | 1123 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1124 | nr_reads++; |
1125 | ||
1126 | if (nr_reads >= max_nr_reads) | |
1127 | break; | |
1128 | } | |
1129 | ||
c5cc2070 | 1130 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && |
0f3457f6 | 1131 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1132 | |
77216b04 | 1133 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1134 | nr_writes++; |
1135 | ||
1136 | if (nr_writes >= max_nr_writes) | |
1137 | break; | |
1138 | } | |
1139 | ||
1140 | return nr_reads + nr_writes; | |
1141 | } | |
1142 | ||
651930bc | 1143 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
1144 | { |
1145 | unsigned int nr_disp = 0; | |
e43473b7 VG |
1146 | |
1147 | while (1) { | |
73f0d49a TH |
1148 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
1149 | struct throtl_service_queue *sq = &tg->service_queue; | |
e43473b7 VG |
1150 | |
1151 | if (!tg) | |
1152 | break; | |
1153 | ||
1154 | if (time_before(jiffies, tg->disptime)) | |
1155 | break; | |
1156 | ||
77216b04 | 1157 | throtl_dequeue_tg(tg); |
e43473b7 | 1158 | |
77216b04 | 1159 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 1160 | |
73f0d49a | 1161 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
77216b04 | 1162 | tg_update_disptime(tg); |
e43473b7 VG |
1163 | |
1164 | if (nr_disp >= throtl_quantum) | |
1165 | break; | |
1166 | } | |
1167 | ||
1168 | return nr_disp; | |
1169 | } | |
1170 | ||
6e1a5704 TH |
1171 | /** |
1172 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
1173 | * @arg: the throtl_service_queue being serviced | |
1174 | * | |
1175 | * This timer is armed when a child throtl_grp with active bio's become | |
1176 | * pending and queued on the service_queue's pending_tree and expires when | |
1177 | * the first child throtl_grp should be dispatched. This function | |
2e48a530 TH |
1178 | * dispatches bio's from the children throtl_grps to the parent |
1179 | * service_queue. | |
1180 | * | |
1181 | * If the parent's parent is another throtl_grp, dispatching is propagated | |
1182 | * by either arming its pending_timer or repeating dispatch directly. If | |
1183 | * the top-level service_tree is reached, throtl_data->dispatch_work is | |
1184 | * kicked so that the ready bio's are issued. | |
6e1a5704 | 1185 | */ |
69df0ab0 TH |
1186 | static void throtl_pending_timer_fn(unsigned long arg) |
1187 | { | |
1188 | struct throtl_service_queue *sq = (void *)arg; | |
2e48a530 | 1189 | struct throtl_grp *tg = sq_to_tg(sq); |
69df0ab0 | 1190 | struct throtl_data *td = sq_to_td(sq); |
cb76199c | 1191 | struct request_queue *q = td->queue; |
2e48a530 TH |
1192 | struct throtl_service_queue *parent_sq; |
1193 | bool dispatched; | |
6e1a5704 | 1194 | int ret; |
e43473b7 VG |
1195 | |
1196 | spin_lock_irq(q->queue_lock); | |
2e48a530 TH |
1197 | again: |
1198 | parent_sq = sq->parent_sq; | |
1199 | dispatched = false; | |
e43473b7 | 1200 | |
7f52f98c TH |
1201 | while (true) { |
1202 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
2e48a530 TH |
1203 | sq->nr_queued[READ] + sq->nr_queued[WRITE], |
1204 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
7f52f98c TH |
1205 | |
1206 | ret = throtl_select_dispatch(sq); | |
1207 | if (ret) { | |
7f52f98c TH |
1208 | throtl_log(sq, "bios disp=%u", ret); |
1209 | dispatched = true; | |
1210 | } | |
e43473b7 | 1211 | |
7f52f98c TH |
1212 | if (throtl_schedule_next_dispatch(sq, false)) |
1213 | break; | |
e43473b7 | 1214 | |
7f52f98c TH |
1215 | /* this dispatch windows is still open, relax and repeat */ |
1216 | spin_unlock_irq(q->queue_lock); | |
1217 | cpu_relax(); | |
1218 | spin_lock_irq(q->queue_lock); | |
651930bc | 1219 | } |
e43473b7 | 1220 | |
2e48a530 TH |
1221 | if (!dispatched) |
1222 | goto out_unlock; | |
6e1a5704 | 1223 | |
2e48a530 TH |
1224 | if (parent_sq) { |
1225 | /* @parent_sq is another throl_grp, propagate dispatch */ | |
1226 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1227 | tg_update_disptime(tg); | |
1228 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { | |
1229 | /* window is already open, repeat dispatching */ | |
1230 | sq = parent_sq; | |
1231 | tg = sq_to_tg(sq); | |
1232 | goto again; | |
1233 | } | |
1234 | } | |
1235 | } else { | |
1236 | /* reached the top-level, queue issueing */ | |
1237 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1238 | } | |
1239 | out_unlock: | |
e43473b7 | 1240 | spin_unlock_irq(q->queue_lock); |
6e1a5704 | 1241 | } |
e43473b7 | 1242 | |
6e1a5704 TH |
1243 | /** |
1244 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
1245 | * @work: work item being executed | |
1246 | * | |
1247 | * This function is queued for execution when bio's reach the bio_lists[] | |
1248 | * of throtl_data->service_queue. Those bio's are ready and issued by this | |
1249 | * function. | |
1250 | */ | |
1251 | void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
1252 | { | |
1253 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1254 | dispatch_work); | |
1255 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1256 | struct request_queue *q = td->queue; | |
1257 | struct bio_list bio_list_on_stack; | |
1258 | struct bio *bio; | |
1259 | struct blk_plug plug; | |
1260 | int rw; | |
1261 | ||
1262 | bio_list_init(&bio_list_on_stack); | |
1263 | ||
1264 | spin_lock_irq(q->queue_lock); | |
c5cc2070 TH |
1265 | for (rw = READ; rw <= WRITE; rw++) |
1266 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) | |
1267 | bio_list_add(&bio_list_on_stack, bio); | |
6e1a5704 TH |
1268 | spin_unlock_irq(q->queue_lock); |
1269 | ||
1270 | if (!bio_list_empty(&bio_list_on_stack)) { | |
69d60eb9 | 1271 | blk_start_plug(&plug); |
e43473b7 VG |
1272 | while((bio = bio_list_pop(&bio_list_on_stack))) |
1273 | generic_make_request(bio); | |
69d60eb9 | 1274 | blk_finish_plug(&plug); |
e43473b7 | 1275 | } |
e43473b7 VG |
1276 | } |
1277 | ||
f95a04af TH |
1278 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
1279 | struct blkg_policy_data *pd, int off) | |
41b38b6d | 1280 | { |
f95a04af | 1281 | struct throtl_grp *tg = pd_to_tg(pd); |
41b38b6d TH |
1282 | struct blkg_rwstat rwstat = { }, tmp; |
1283 | int i, cpu; | |
1284 | ||
1285 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 1286 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
1287 | |
1288 | tmp = blkg_rwstat_read((void *)sc + off); | |
1289 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
1290 | rwstat.cnt[i] += tmp.cnt[i]; | |
1291 | } | |
1292 | ||
f95a04af | 1293 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
41b38b6d TH |
1294 | } |
1295 | ||
182446d0 TH |
1296 | static int tg_print_cpu_rwstat(struct cgroup_subsys_state *css, |
1297 | struct cftype *cft, struct seq_file *sf) | |
41b38b6d | 1298 | { |
182446d0 | 1299 | struct blkcg *blkcg = css_to_blkcg(css); |
41b38b6d | 1300 | |
3c798398 | 1301 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 1302 | cft->private, true); |
41b38b6d TH |
1303 | return 0; |
1304 | } | |
1305 | ||
f95a04af TH |
1306 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
1307 | int off) | |
60c2bc2d | 1308 | { |
f95a04af TH |
1309 | struct throtl_grp *tg = pd_to_tg(pd); |
1310 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 1311 | |
af133ceb | 1312 | if (v == -1) |
60c2bc2d | 1313 | return 0; |
f95a04af | 1314 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
1315 | } |
1316 | ||
f95a04af TH |
1317 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
1318 | int off) | |
e43473b7 | 1319 | { |
f95a04af TH |
1320 | struct throtl_grp *tg = pd_to_tg(pd); |
1321 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 1322 | |
af133ceb TH |
1323 | if (v == -1) |
1324 | return 0; | |
f95a04af | 1325 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
1326 | } |
1327 | ||
182446d0 TH |
1328 | static int tg_print_conf_u64(struct cgroup_subsys_state *css, |
1329 | struct cftype *cft, struct seq_file *sf) | |
8e89d13f | 1330 | { |
182446d0 | 1331 | blkcg_print_blkgs(sf, css_to_blkcg(css), tg_prfill_conf_u64, |
3c798398 | 1332 | &blkcg_policy_throtl, cft->private, false); |
af133ceb | 1333 | return 0; |
8e89d13f VG |
1334 | } |
1335 | ||
182446d0 TH |
1336 | static int tg_print_conf_uint(struct cgroup_subsys_state *css, |
1337 | struct cftype *cft, struct seq_file *sf) | |
8e89d13f | 1338 | { |
182446d0 | 1339 | blkcg_print_blkgs(sf, css_to_blkcg(css), tg_prfill_conf_uint, |
3c798398 | 1340 | &blkcg_policy_throtl, cft->private, false); |
af133ceb | 1341 | return 0; |
60c2bc2d TH |
1342 | } |
1343 | ||
182446d0 TH |
1344 | static int tg_set_conf(struct cgroup_subsys_state *css, struct cftype *cft, |
1345 | const char *buf, bool is_u64) | |
60c2bc2d | 1346 | { |
182446d0 | 1347 | struct blkcg *blkcg = css_to_blkcg(css); |
60c2bc2d | 1348 | struct blkg_conf_ctx ctx; |
af133ceb | 1349 | struct throtl_grp *tg; |
69df0ab0 | 1350 | struct throtl_service_queue *sq; |
693e751e | 1351 | struct blkcg_gq *blkg; |
492eb21b | 1352 | struct cgroup_subsys_state *pos_css; |
60c2bc2d TH |
1353 | int ret; |
1354 | ||
3c798398 | 1355 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
1356 | if (ret) |
1357 | return ret; | |
1358 | ||
af133ceb | 1359 | tg = blkg_to_tg(ctx.blkg); |
69df0ab0 | 1360 | sq = &tg->service_queue; |
af133ceb | 1361 | |
a2b1693b TH |
1362 | if (!ctx.v) |
1363 | ctx.v = -1; | |
af133ceb | 1364 | |
a2b1693b TH |
1365 | if (is_u64) |
1366 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
1367 | else | |
1368 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 1369 | |
fda6f272 TH |
1370 | throtl_log(&tg->service_queue, |
1371 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
1372 | tg->bps[READ], tg->bps[WRITE], | |
1373 | tg->iops[READ], tg->iops[WRITE]); | |
632b4493 | 1374 | |
693e751e TH |
1375 | /* |
1376 | * Update has_rules[] flags for the updated tg's subtree. A tg is | |
1377 | * considered to have rules if either the tg itself or any of its | |
1378 | * ancestors has rules. This identifies groups without any | |
1379 | * restrictions in the whole hierarchy and allows them to bypass | |
1380 | * blk-throttle. | |
1381 | */ | |
492eb21b | 1382 | blkg_for_each_descendant_pre(blkg, pos_css, ctx.blkg) |
693e751e TH |
1383 | tg_update_has_rules(blkg_to_tg(blkg)); |
1384 | ||
632b4493 TH |
1385 | /* |
1386 | * We're already holding queue_lock and know @tg is valid. Let's | |
1387 | * apply the new config directly. | |
1388 | * | |
1389 | * Restart the slices for both READ and WRITES. It might happen | |
1390 | * that a group's limit are dropped suddenly and we don't want to | |
1391 | * account recently dispatched IO with new low rate. | |
1392 | */ | |
0f3457f6 TH |
1393 | throtl_start_new_slice(tg, 0); |
1394 | throtl_start_new_slice(tg, 1); | |
632b4493 | 1395 | |
5b2c16aa | 1396 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 1397 | tg_update_disptime(tg); |
7f52f98c | 1398 | throtl_schedule_next_dispatch(sq->parent_sq, true); |
632b4493 | 1399 | } |
60c2bc2d TH |
1400 | |
1401 | blkg_conf_finish(&ctx); | |
a2b1693b | 1402 | return 0; |
8e89d13f VG |
1403 | } |
1404 | ||
182446d0 | 1405 | static int tg_set_conf_u64(struct cgroup_subsys_state *css, struct cftype *cft, |
af133ceb | 1406 | const char *buf) |
60c2bc2d | 1407 | { |
182446d0 | 1408 | return tg_set_conf(css, cft, buf, true); |
60c2bc2d TH |
1409 | } |
1410 | ||
182446d0 | 1411 | static int tg_set_conf_uint(struct cgroup_subsys_state *css, struct cftype *cft, |
af133ceb | 1412 | const char *buf) |
60c2bc2d | 1413 | { |
182446d0 | 1414 | return tg_set_conf(css, cft, buf, false); |
60c2bc2d TH |
1415 | } |
1416 | ||
1417 | static struct cftype throtl_files[] = { | |
1418 | { | |
1419 | .name = "throttle.read_bps_device", | |
af133ceb TH |
1420 | .private = offsetof(struct throtl_grp, bps[READ]), |
1421 | .read_seq_string = tg_print_conf_u64, | |
1422 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1423 | .max_write_len = 256, |
1424 | }, | |
1425 | { | |
1426 | .name = "throttle.write_bps_device", | |
af133ceb TH |
1427 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
1428 | .read_seq_string = tg_print_conf_u64, | |
1429 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
1430 | .max_write_len = 256, |
1431 | }, | |
1432 | { | |
1433 | .name = "throttle.read_iops_device", | |
af133ceb TH |
1434 | .private = offsetof(struct throtl_grp, iops[READ]), |
1435 | .read_seq_string = tg_print_conf_uint, | |
1436 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1437 | .max_write_len = 256, |
1438 | }, | |
1439 | { | |
1440 | .name = "throttle.write_iops_device", | |
af133ceb TH |
1441 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
1442 | .read_seq_string = tg_print_conf_uint, | |
1443 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
1444 | .max_write_len = 256, |
1445 | }, | |
1446 | { | |
1447 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1448 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1449 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1450 | }, |
1451 | { | |
1452 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1453 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1454 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1455 | }, |
1456 | { } /* terminate */ | |
1457 | }; | |
1458 | ||
da527770 | 1459 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1460 | { |
1461 | struct throtl_data *td = q->td; | |
1462 | ||
69df0ab0 | 1463 | cancel_work_sync(&td->dispatch_work); |
e43473b7 VG |
1464 | } |
1465 | ||
3c798398 | 1466 | static struct blkcg_policy blkcg_policy_throtl = { |
f9fcc2d3 TH |
1467 | .pd_size = sizeof(struct throtl_grp), |
1468 | .cftypes = throtl_files, | |
1469 | ||
1470 | .pd_init_fn = throtl_pd_init, | |
693e751e | 1471 | .pd_online_fn = throtl_pd_online, |
f9fcc2d3 TH |
1472 | .pd_exit_fn = throtl_pd_exit, |
1473 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 VG |
1474 | }; |
1475 | ||
bc16a4f9 | 1476 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1477 | { |
1478 | struct throtl_data *td = q->td; | |
c5cc2070 | 1479 | struct throtl_qnode *qn = NULL; |
e43473b7 | 1480 | struct throtl_grp *tg; |
73f0d49a | 1481 | struct throtl_service_queue *sq; |
0e9f4164 | 1482 | bool rw = bio_data_dir(bio); |
3c798398 | 1483 | struct blkcg *blkcg; |
bc16a4f9 | 1484 | bool throttled = false; |
e43473b7 | 1485 | |
2a0f61e6 TH |
1486 | /* see throtl_charge_bio() */ |
1487 | if (bio->bi_rw & REQ_THROTTLED) | |
bc16a4f9 | 1488 | goto out; |
e43473b7 | 1489 | |
af75cd3c VG |
1490 | /* |
1491 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1492 | * basic fields like stats and io rates. If a group has no rules, | |
1493 | * just update the dispatch stats in lockless manner and return. | |
1494 | */ | |
af75cd3c | 1495 | rcu_read_lock(); |
3c798398 | 1496 | blkcg = bio_blkcg(bio); |
cd1604fa | 1497 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1498 | if (tg) { |
693e751e | 1499 | if (!tg->has_rules[rw]) { |
629ed0b1 TH |
1500 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1501 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1502 | goto out_unlock_rcu; |
af75cd3c VG |
1503 | } |
1504 | } | |
af75cd3c VG |
1505 | |
1506 | /* | |
1507 | * Either group has not been allocated yet or it is not an unlimited | |
1508 | * IO group | |
1509 | */ | |
e43473b7 | 1510 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1511 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1512 | if (unlikely(!tg)) |
1513 | goto out_unlock; | |
f469a7b4 | 1514 | |
73f0d49a TH |
1515 | sq = &tg->service_queue; |
1516 | ||
9e660acf TH |
1517 | while (true) { |
1518 | /* throtl is FIFO - if bios are already queued, should queue */ | |
1519 | if (sq->nr_queued[rw]) | |
1520 | break; | |
de701c74 | 1521 | |
9e660acf TH |
1522 | /* if above limits, break to queue */ |
1523 | if (!tg_may_dispatch(tg, bio, NULL)) | |
1524 | break; | |
1525 | ||
1526 | /* within limits, let's charge and dispatch directly */ | |
e43473b7 | 1527 | throtl_charge_bio(tg, bio); |
04521db0 VG |
1528 | |
1529 | /* | |
1530 | * We need to trim slice even when bios are not being queued | |
1531 | * otherwise it might happen that a bio is not queued for | |
1532 | * a long time and slice keeps on extending and trim is not | |
1533 | * called for a long time. Now if limits are reduced suddenly | |
1534 | * we take into account all the IO dispatched so far at new | |
1535 | * low rate and * newly queued IO gets a really long dispatch | |
1536 | * time. | |
1537 | * | |
1538 | * So keep on trimming slice even if bio is not queued. | |
1539 | */ | |
0f3457f6 | 1540 | throtl_trim_slice(tg, rw); |
9e660acf TH |
1541 | |
1542 | /* | |
1543 | * @bio passed through this layer without being throttled. | |
1544 | * Climb up the ladder. If we''re already at the top, it | |
1545 | * can be executed directly. | |
1546 | */ | |
c5cc2070 | 1547 | qn = &tg->qnode_on_parent[rw]; |
9e660acf TH |
1548 | sq = sq->parent_sq; |
1549 | tg = sq_to_tg(sq); | |
1550 | if (!tg) | |
1551 | goto out_unlock; | |
e43473b7 VG |
1552 | } |
1553 | ||
9e660acf | 1554 | /* out-of-limit, queue to @tg */ |
fda6f272 TH |
1555 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
1556 | rw == READ ? 'R' : 'W', | |
1557 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], | |
1558 | tg->io_disp[rw], tg->iops[rw], | |
1559 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
e43473b7 | 1560 | |
671058fb | 1561 | bio_associate_current(bio); |
6bc9c2b4 | 1562 | tg->td->nr_queued[rw]++; |
c5cc2070 | 1563 | throtl_add_bio_tg(bio, qn, tg); |
bc16a4f9 | 1564 | throttled = true; |
e43473b7 | 1565 | |
7f52f98c TH |
1566 | /* |
1567 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
1568 | * was empty before @bio. The forced scheduling isn't likely to | |
1569 | * cause undue delay as @bio is likely to be dispatched directly if | |
1570 | * its @tg's disptime is not in the future. | |
1571 | */ | |
0e9f4164 | 1572 | if (tg->flags & THROTL_TG_WAS_EMPTY) { |
77216b04 | 1573 | tg_update_disptime(tg); |
7f52f98c | 1574 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); |
e43473b7 VG |
1575 | } |
1576 | ||
bc16a4f9 | 1577 | out_unlock: |
e43473b7 | 1578 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1579 | out_unlock_rcu: |
1580 | rcu_read_unlock(); | |
bc16a4f9 | 1581 | out: |
2a0f61e6 TH |
1582 | /* |
1583 | * As multiple blk-throtls may stack in the same issue path, we | |
1584 | * don't want bios to leave with the flag set. Clear the flag if | |
1585 | * being issued. | |
1586 | */ | |
1587 | if (!throttled) | |
1588 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1589 | return throttled; |
e43473b7 VG |
1590 | } |
1591 | ||
2a12f0dc TH |
1592 | /* |
1593 | * Dispatch all bios from all children tg's queued on @parent_sq. On | |
1594 | * return, @parent_sq is guaranteed to not have any active children tg's | |
1595 | * and all bios from previously active tg's are on @parent_sq->bio_lists[]. | |
1596 | */ | |
1597 | static void tg_drain_bios(struct throtl_service_queue *parent_sq) | |
1598 | { | |
1599 | struct throtl_grp *tg; | |
1600 | ||
1601 | while ((tg = throtl_rb_first(parent_sq))) { | |
1602 | struct throtl_service_queue *sq = &tg->service_queue; | |
1603 | struct bio *bio; | |
1604 | ||
1605 | throtl_dequeue_tg(tg); | |
1606 | ||
c5cc2070 | 1607 | while ((bio = throtl_peek_queued(&sq->queued[READ]))) |
2a12f0dc | 1608 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
c5cc2070 | 1609 | while ((bio = throtl_peek_queued(&sq->queued[WRITE]))) |
2a12f0dc TH |
1610 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
1611 | } | |
1612 | } | |
1613 | ||
c9a929dd TH |
1614 | /** |
1615 | * blk_throtl_drain - drain throttled bios | |
1616 | * @q: request_queue to drain throttled bios for | |
1617 | * | |
1618 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1619 | */ | |
1620 | void blk_throtl_drain(struct request_queue *q) | |
1621 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1622 | { | |
1623 | struct throtl_data *td = q->td; | |
2a12f0dc | 1624 | struct blkcg_gq *blkg; |
492eb21b | 1625 | struct cgroup_subsys_state *pos_css; |
c9a929dd | 1626 | struct bio *bio; |
651930bc | 1627 | int rw; |
c9a929dd | 1628 | |
8bcb6c7d | 1629 | queue_lockdep_assert_held(q); |
2a12f0dc | 1630 | rcu_read_lock(); |
c9a929dd | 1631 | |
2a12f0dc TH |
1632 | /* |
1633 | * Drain each tg while doing post-order walk on the blkg tree, so | |
1634 | * that all bios are propagated to td->service_queue. It'd be | |
1635 | * better to walk service_queue tree directly but blkg walk is | |
1636 | * easier. | |
1637 | */ | |
492eb21b | 1638 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) |
2a12f0dc | 1639 | tg_drain_bios(&blkg_to_tg(blkg)->service_queue); |
73f0d49a | 1640 | |
2a12f0dc TH |
1641 | /* finally, transfer bios from top-level tg's into the td */ |
1642 | tg_drain_bios(&td->service_queue); | |
1643 | ||
1644 | rcu_read_unlock(); | |
c9a929dd TH |
1645 | spin_unlock_irq(q->queue_lock); |
1646 | ||
2a12f0dc | 1647 | /* all bios now should be in td->service_queue, issue them */ |
651930bc | 1648 | for (rw = READ; rw <= WRITE; rw++) |
c5cc2070 TH |
1649 | while ((bio = throtl_pop_queued(&td->service_queue.queued[rw], |
1650 | NULL))) | |
651930bc | 1651 | generic_make_request(bio); |
c9a929dd TH |
1652 | |
1653 | spin_lock_irq(q->queue_lock); | |
1654 | } | |
1655 | ||
e43473b7 VG |
1656 | int blk_throtl_init(struct request_queue *q) |
1657 | { | |
1658 | struct throtl_data *td; | |
a2b1693b | 1659 | int ret; |
e43473b7 VG |
1660 | |
1661 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1662 | if (!td) | |
1663 | return -ENOMEM; | |
1664 | ||
69df0ab0 | 1665 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
77216b04 | 1666 | throtl_service_queue_init(&td->service_queue, NULL); |
e43473b7 | 1667 | |
cd1604fa | 1668 | q->td = td; |
29b12589 | 1669 | td->queue = q; |
02977e4a | 1670 | |
a2b1693b | 1671 | /* activate policy */ |
3c798398 | 1672 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1673 | if (ret) |
f51b802c | 1674 | kfree(td); |
a2b1693b | 1675 | return ret; |
e43473b7 VG |
1676 | } |
1677 | ||
1678 | void blk_throtl_exit(struct request_queue *q) | |
1679 | { | |
c875f4d0 | 1680 | BUG_ON(!q->td); |
da527770 | 1681 | throtl_shutdown_wq(q); |
3c798398 | 1682 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1683 | kfree(q->td); |
e43473b7 VG |
1684 | } |
1685 | ||
1686 | static int __init throtl_init(void) | |
1687 | { | |
450adcbe VG |
1688 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1689 | if (!kthrotld_workqueue) | |
1690 | panic("Failed to create kthrotld\n"); | |
1691 | ||
3c798398 | 1692 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1693 | } |
1694 | ||
1695 | module_init(throtl_init); |