Commit | Line | Data |
---|---|---|
e43473b7 VG |
1 | /* |
2 | * Interface for controlling IO bandwidth on a request queue | |
3 | * | |
4 | * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com> | |
5 | */ | |
6 | ||
7 | #include <linux/module.h> | |
8 | #include <linux/slab.h> | |
9 | #include <linux/blkdev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blktrace_api.h> | |
eea8f41c | 12 | #include <linux/blk-cgroup.h> |
bc9fcbf9 | 13 | #include "blk.h" |
e43473b7 VG |
14 | |
15 | /* Max dispatch from a group in 1 round */ | |
16 | static int throtl_grp_quantum = 8; | |
17 | ||
18 | /* Total max dispatch from all groups in one round */ | |
19 | static int throtl_quantum = 32; | |
20 | ||
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 | ||
86 | struct throtl_grp { | |
f95a04af TH |
87 | /* must be the first member */ |
88 | struct blkg_policy_data pd; | |
89 | ||
c9e0332e | 90 | /* active throtl group service_queue member */ |
e43473b7 VG |
91 | struct rb_node rb_node; |
92 | ||
0f3457f6 TH |
93 | /* throtl_data this group belongs to */ |
94 | struct throtl_data *td; | |
95 | ||
49a2f1e3 TH |
96 | /* this group's service queue */ |
97 | struct throtl_service_queue service_queue; | |
98 | ||
c5cc2070 TH |
99 | /* |
100 | * qnode_on_self is used when bios are directly queued to this | |
101 | * throtl_grp so that local bios compete fairly with bios | |
102 | * dispatched from children. qnode_on_parent is used when bios are | |
103 | * dispatched from this throtl_grp into its parent and will compete | |
104 | * with the sibling qnode_on_parents and the parent's | |
105 | * qnode_on_self. | |
106 | */ | |
107 | struct throtl_qnode qnode_on_self[2]; | |
108 | struct throtl_qnode qnode_on_parent[2]; | |
109 | ||
e43473b7 VG |
110 | /* |
111 | * Dispatch time in jiffies. This is the estimated time when group | |
112 | * will unthrottle and is ready to dispatch more bio. It is used as | |
113 | * key to sort active groups in service tree. | |
114 | */ | |
115 | unsigned long disptime; | |
116 | ||
e43473b7 VG |
117 | unsigned int flags; |
118 | ||
693e751e TH |
119 | /* are there any throtl rules between this group and td? */ |
120 | bool has_rules[2]; | |
121 | ||
e43473b7 VG |
122 | /* bytes per second rate limits */ |
123 | uint64_t bps[2]; | |
124 | ||
8e89d13f VG |
125 | /* IOPS limits */ |
126 | unsigned int iops[2]; | |
127 | ||
e43473b7 VG |
128 | /* Number of bytes disptached in current slice */ |
129 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
130 | /* Number of bio's dispatched in current slice */ |
131 | unsigned int io_disp[2]; | |
e43473b7 VG |
132 | |
133 | /* When did we start a new slice */ | |
134 | unsigned long slice_start[2]; | |
135 | unsigned long slice_end[2]; | |
136 | }; | |
137 | ||
138 | struct throtl_data | |
139 | { | |
e43473b7 | 140 | /* service tree for active throtl groups */ |
c9e0332e | 141 | struct throtl_service_queue service_queue; |
e43473b7 | 142 | |
e43473b7 VG |
143 | struct request_queue *queue; |
144 | ||
145 | /* Total Number of queued bios on READ and WRITE lists */ | |
146 | unsigned int nr_queued[2]; | |
147 | ||
e43473b7 | 148 | /* Work for dispatching throttled bios */ |
69df0ab0 | 149 | struct work_struct dispatch_work; |
e43473b7 VG |
150 | }; |
151 | ||
69df0ab0 TH |
152 | static void throtl_pending_timer_fn(unsigned long arg); |
153 | ||
f95a04af TH |
154 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
155 | { | |
156 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
157 | } | |
158 | ||
3c798398 | 159 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 160 | { |
f95a04af | 161 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
162 | } |
163 | ||
3c798398 | 164 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 165 | { |
f95a04af | 166 | return pd_to_blkg(&tg->pd); |
0381411e TH |
167 | } |
168 | ||
fda6f272 TH |
169 | /** |
170 | * sq_to_tg - return the throl_grp the specified service queue belongs to | |
171 | * @sq: the throtl_service_queue of interest | |
172 | * | |
173 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one | |
174 | * embedded in throtl_data, %NULL is returned. | |
175 | */ | |
176 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) | |
177 | { | |
178 | if (sq && sq->parent_sq) | |
179 | return container_of(sq, struct throtl_grp, service_queue); | |
180 | else | |
181 | return NULL; | |
182 | } | |
183 | ||
184 | /** | |
185 | * sq_to_td - return throtl_data the specified service queue belongs to | |
186 | * @sq: the throtl_service_queue of interest | |
187 | * | |
188 | * A service_queue can be embeded in either a throtl_grp or throtl_data. | |
189 | * Determine the associated throtl_data accordingly and return it. | |
190 | */ | |
191 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) | |
192 | { | |
193 | struct throtl_grp *tg = sq_to_tg(sq); | |
194 | ||
195 | if (tg) | |
196 | return tg->td; | |
197 | else | |
198 | return container_of(sq, struct throtl_data, service_queue); | |
199 | } | |
200 | ||
201 | /** | |
202 | * throtl_log - log debug message via blktrace | |
203 | * @sq: the service_queue being reported | |
204 | * @fmt: printf format string | |
205 | * @args: printf args | |
206 | * | |
207 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a | |
208 | * throtl_grp; otherwise, just "throtl". | |
fda6f272 TH |
209 | */ |
210 | #define throtl_log(sq, fmt, args...) do { \ | |
211 | struct throtl_grp *__tg = sq_to_tg((sq)); \ | |
212 | struct throtl_data *__td = sq_to_td((sq)); \ | |
213 | \ | |
214 | (void)__td; \ | |
59fa0224 SL |
215 | if (likely(!blk_trace_note_message_enabled(__td->queue))) \ |
216 | break; \ | |
fda6f272 TH |
217 | if ((__tg)) { \ |
218 | char __pbuf[128]; \ | |
54e7ed12 | 219 | \ |
fda6f272 TH |
220 | blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \ |
221 | blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
222 | } else { \ | |
223 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
224 | } \ | |
54e7ed12 | 225 | } while (0) |
e43473b7 | 226 | |
c5cc2070 TH |
227 | static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg) |
228 | { | |
229 | INIT_LIST_HEAD(&qn->node); | |
230 | bio_list_init(&qn->bios); | |
231 | qn->tg = tg; | |
232 | } | |
233 | ||
234 | /** | |
235 | * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it | |
236 | * @bio: bio being added | |
237 | * @qn: qnode to add bio to | |
238 | * @queued: the service_queue->queued[] list @qn belongs to | |
239 | * | |
240 | * Add @bio to @qn and put @qn on @queued if it's not already on. | |
241 | * @qn->tg's reference count is bumped when @qn is activated. See the | |
242 | * comment on top of throtl_qnode definition for details. | |
243 | */ | |
244 | static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn, | |
245 | struct list_head *queued) | |
246 | { | |
247 | bio_list_add(&qn->bios, bio); | |
248 | if (list_empty(&qn->node)) { | |
249 | list_add_tail(&qn->node, queued); | |
250 | blkg_get(tg_to_blkg(qn->tg)); | |
251 | } | |
252 | } | |
253 | ||
254 | /** | |
255 | * throtl_peek_queued - peek the first bio on a qnode list | |
256 | * @queued: the qnode list to peek | |
257 | */ | |
258 | static struct bio *throtl_peek_queued(struct list_head *queued) | |
259 | { | |
260 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
261 | struct bio *bio; | |
262 | ||
263 | if (list_empty(queued)) | |
264 | return NULL; | |
265 | ||
266 | bio = bio_list_peek(&qn->bios); | |
267 | WARN_ON_ONCE(!bio); | |
268 | return bio; | |
269 | } | |
270 | ||
271 | /** | |
272 | * throtl_pop_queued - pop the first bio form a qnode list | |
273 | * @queued: the qnode list to pop a bio from | |
274 | * @tg_to_put: optional out argument for throtl_grp to put | |
275 | * | |
276 | * Pop the first bio from the qnode list @queued. After popping, the first | |
277 | * qnode is removed from @queued if empty or moved to the end of @queued so | |
278 | * that the popping order is round-robin. | |
279 | * | |
280 | * When the first qnode is removed, its associated throtl_grp should be put | |
281 | * too. If @tg_to_put is NULL, this function automatically puts it; | |
282 | * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is | |
283 | * responsible for putting it. | |
284 | */ | |
285 | static struct bio *throtl_pop_queued(struct list_head *queued, | |
286 | struct throtl_grp **tg_to_put) | |
287 | { | |
288 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
289 | struct bio *bio; | |
290 | ||
291 | if (list_empty(queued)) | |
292 | return NULL; | |
293 | ||
294 | bio = bio_list_pop(&qn->bios); | |
295 | WARN_ON_ONCE(!bio); | |
296 | ||
297 | if (bio_list_empty(&qn->bios)) { | |
298 | list_del_init(&qn->node); | |
299 | if (tg_to_put) | |
300 | *tg_to_put = qn->tg; | |
301 | else | |
302 | blkg_put(tg_to_blkg(qn->tg)); | |
303 | } else { | |
304 | list_move_tail(&qn->node, queued); | |
305 | } | |
306 | ||
307 | return bio; | |
308 | } | |
309 | ||
49a2f1e3 | 310 | /* init a service_queue, assumes the caller zeroed it */ |
b2ce2643 | 311 | static void throtl_service_queue_init(struct throtl_service_queue *sq) |
49a2f1e3 | 312 | { |
c5cc2070 TH |
313 | INIT_LIST_HEAD(&sq->queued[0]); |
314 | INIT_LIST_HEAD(&sq->queued[1]); | |
49a2f1e3 | 315 | sq->pending_tree = RB_ROOT; |
69df0ab0 TH |
316 | setup_timer(&sq->pending_timer, throtl_pending_timer_fn, |
317 | (unsigned long)sq); | |
318 | } | |
319 | ||
001bea73 TH |
320 | static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node) |
321 | { | |
4fb72036 | 322 | struct throtl_grp *tg; |
24bdb8ef | 323 | int rw; |
4fb72036 TH |
324 | |
325 | tg = kzalloc_node(sizeof(*tg), gfp, node); | |
326 | if (!tg) | |
77ea7338 | 327 | return NULL; |
4fb72036 | 328 | |
b2ce2643 TH |
329 | throtl_service_queue_init(&tg->service_queue); |
330 | ||
331 | for (rw = READ; rw <= WRITE; rw++) { | |
332 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); | |
333 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); | |
334 | } | |
335 | ||
336 | RB_CLEAR_NODE(&tg->rb_node); | |
337 | tg->bps[READ] = -1; | |
338 | tg->bps[WRITE] = -1; | |
339 | tg->iops[READ] = -1; | |
340 | tg->iops[WRITE] = -1; | |
341 | ||
4fb72036 | 342 | return &tg->pd; |
001bea73 TH |
343 | } |
344 | ||
a9520cd6 | 345 | static void throtl_pd_init(struct blkg_policy_data *pd) |
a29a171e | 346 | { |
a9520cd6 TH |
347 | struct throtl_grp *tg = pd_to_tg(pd); |
348 | struct blkcg_gq *blkg = tg_to_blkg(tg); | |
77216b04 | 349 | struct throtl_data *td = blkg->q->td; |
b2ce2643 | 350 | struct throtl_service_queue *sq = &tg->service_queue; |
cd1604fa | 351 | |
9138125b | 352 | /* |
aa6ec29b | 353 | * If on the default hierarchy, we switch to properly hierarchical |
9138125b TH |
354 | * behavior where limits on a given throtl_grp are applied to the |
355 | * whole subtree rather than just the group itself. e.g. If 16M | |
356 | * read_bps limit is set on the root group, the whole system can't | |
357 | * exceed 16M for the device. | |
358 | * | |
aa6ec29b | 359 | * If not on the default hierarchy, the broken flat hierarchy |
9138125b TH |
360 | * behavior is retained where all throtl_grps are treated as if |
361 | * they're all separate root groups right below throtl_data. | |
362 | * Limits of a group don't interact with limits of other groups | |
363 | * regardless of the position of the group in the hierarchy. | |
364 | */ | |
b2ce2643 | 365 | sq->parent_sq = &td->service_queue; |
9e10a130 | 366 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent) |
b2ce2643 | 367 | sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue; |
77216b04 | 368 | tg->td = td; |
8a3d2615 TH |
369 | } |
370 | ||
693e751e TH |
371 | /* |
372 | * Set has_rules[] if @tg or any of its parents have limits configured. | |
373 | * This doesn't require walking up to the top of the hierarchy as the | |
374 | * parent's has_rules[] is guaranteed to be correct. | |
375 | */ | |
376 | static void tg_update_has_rules(struct throtl_grp *tg) | |
377 | { | |
378 | struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); | |
379 | int rw; | |
380 | ||
381 | for (rw = READ; rw <= WRITE; rw++) | |
382 | tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) || | |
383 | (tg->bps[rw] != -1 || tg->iops[rw] != -1); | |
384 | } | |
385 | ||
a9520cd6 | 386 | static void throtl_pd_online(struct blkg_policy_data *pd) |
693e751e TH |
387 | { |
388 | /* | |
389 | * We don't want new groups to escape the limits of its ancestors. | |
390 | * Update has_rules[] after a new group is brought online. | |
391 | */ | |
a9520cd6 | 392 | tg_update_has_rules(pd_to_tg(pd)); |
693e751e TH |
393 | } |
394 | ||
001bea73 TH |
395 | static void throtl_pd_free(struct blkg_policy_data *pd) |
396 | { | |
4fb72036 TH |
397 | struct throtl_grp *tg = pd_to_tg(pd); |
398 | ||
b2ce2643 | 399 | del_timer_sync(&tg->service_queue.pending_timer); |
4fb72036 | 400 | kfree(tg); |
001bea73 TH |
401 | } |
402 | ||
0049af73 TH |
403 | static struct throtl_grp * |
404 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
405 | { |
406 | /* Service tree is empty */ | |
0049af73 | 407 | if (!parent_sq->nr_pending) |
e43473b7 VG |
408 | return NULL; |
409 | ||
0049af73 TH |
410 | if (!parent_sq->first_pending) |
411 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
e43473b7 | 412 | |
0049af73 TH |
413 | if (parent_sq->first_pending) |
414 | return rb_entry_tg(parent_sq->first_pending); | |
e43473b7 VG |
415 | |
416 | return NULL; | |
417 | } | |
418 | ||
419 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
420 | { | |
421 | rb_erase(n, root); | |
422 | RB_CLEAR_NODE(n); | |
423 | } | |
424 | ||
0049af73 TH |
425 | static void throtl_rb_erase(struct rb_node *n, |
426 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 427 | { |
0049af73 TH |
428 | if (parent_sq->first_pending == n) |
429 | parent_sq->first_pending = NULL; | |
430 | rb_erase_init(n, &parent_sq->pending_tree); | |
431 | --parent_sq->nr_pending; | |
e43473b7 VG |
432 | } |
433 | ||
0049af73 | 434 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
435 | { |
436 | struct throtl_grp *tg; | |
437 | ||
0049af73 | 438 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
439 | if (!tg) |
440 | return; | |
441 | ||
0049af73 | 442 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
443 | } |
444 | ||
77216b04 | 445 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 446 | { |
77216b04 | 447 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
0049af73 | 448 | struct rb_node **node = &parent_sq->pending_tree.rb_node; |
e43473b7 VG |
449 | struct rb_node *parent = NULL; |
450 | struct throtl_grp *__tg; | |
451 | unsigned long key = tg->disptime; | |
452 | int left = 1; | |
453 | ||
454 | while (*node != NULL) { | |
455 | parent = *node; | |
456 | __tg = rb_entry_tg(parent); | |
457 | ||
458 | if (time_before(key, __tg->disptime)) | |
459 | node = &parent->rb_left; | |
460 | else { | |
461 | node = &parent->rb_right; | |
462 | left = 0; | |
463 | } | |
464 | } | |
465 | ||
466 | if (left) | |
0049af73 | 467 | parent_sq->first_pending = &tg->rb_node; |
e43473b7 VG |
468 | |
469 | rb_link_node(&tg->rb_node, parent, node); | |
0049af73 | 470 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); |
e43473b7 VG |
471 | } |
472 | ||
77216b04 | 473 | static void __throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 474 | { |
77216b04 | 475 | tg_service_queue_add(tg); |
5b2c16aa | 476 | tg->flags |= THROTL_TG_PENDING; |
77216b04 | 477 | tg->service_queue.parent_sq->nr_pending++; |
e43473b7 VG |
478 | } |
479 | ||
77216b04 | 480 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 481 | { |
5b2c16aa | 482 | if (!(tg->flags & THROTL_TG_PENDING)) |
77216b04 | 483 | __throtl_enqueue_tg(tg); |
e43473b7 VG |
484 | } |
485 | ||
77216b04 | 486 | static void __throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 487 | { |
77216b04 | 488 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
5b2c16aa | 489 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
490 | } |
491 | ||
77216b04 | 492 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 493 | { |
5b2c16aa | 494 | if (tg->flags & THROTL_TG_PENDING) |
77216b04 | 495 | __throtl_dequeue_tg(tg); |
e43473b7 VG |
496 | } |
497 | ||
a9131a27 | 498 | /* Call with queue lock held */ |
69df0ab0 TH |
499 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, |
500 | unsigned long expires) | |
a9131a27 | 501 | { |
69df0ab0 TH |
502 | mod_timer(&sq->pending_timer, expires); |
503 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
504 | expires - jiffies, jiffies); | |
a9131a27 TH |
505 | } |
506 | ||
7f52f98c TH |
507 | /** |
508 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle | |
509 | * @sq: the service_queue to schedule dispatch for | |
510 | * @force: force scheduling | |
511 | * | |
512 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the | |
513 | * dispatch time of the first pending child. Returns %true if either timer | |
514 | * is armed or there's no pending child left. %false if the current | |
515 | * dispatch window is still open and the caller should continue | |
516 | * dispatching. | |
517 | * | |
518 | * If @force is %true, the dispatch timer is always scheduled and this | |
519 | * function is guaranteed to return %true. This is to be used when the | |
520 | * caller can't dispatch itself and needs to invoke pending_timer | |
521 | * unconditionally. Note that forced scheduling is likely to induce short | |
522 | * delay before dispatch starts even if @sq->first_pending_disptime is not | |
523 | * in the future and thus shouldn't be used in hot paths. | |
524 | */ | |
525 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
526 | bool force) | |
e43473b7 | 527 | { |
6a525600 | 528 | /* any pending children left? */ |
c9e0332e | 529 | if (!sq->nr_pending) |
7f52f98c | 530 | return true; |
e43473b7 | 531 | |
c9e0332e | 532 | update_min_dispatch_time(sq); |
e43473b7 | 533 | |
69df0ab0 | 534 | /* is the next dispatch time in the future? */ |
7f52f98c | 535 | if (force || time_after(sq->first_pending_disptime, jiffies)) { |
69df0ab0 | 536 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); |
7f52f98c | 537 | return true; |
69df0ab0 TH |
538 | } |
539 | ||
7f52f98c TH |
540 | /* tell the caller to continue dispatching */ |
541 | return false; | |
e43473b7 VG |
542 | } |
543 | ||
32ee5bc4 VG |
544 | static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, |
545 | bool rw, unsigned long start) | |
546 | { | |
547 | tg->bytes_disp[rw] = 0; | |
548 | tg->io_disp[rw] = 0; | |
549 | ||
550 | /* | |
551 | * Previous slice has expired. We must have trimmed it after last | |
552 | * bio dispatch. That means since start of last slice, we never used | |
553 | * that bandwidth. Do try to make use of that bandwidth while giving | |
554 | * credit. | |
555 | */ | |
556 | if (time_after_eq(start, tg->slice_start[rw])) | |
557 | tg->slice_start[rw] = start; | |
558 | ||
559 | tg->slice_end[rw] = jiffies + throtl_slice; | |
560 | throtl_log(&tg->service_queue, | |
561 | "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", | |
562 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
563 | tg->slice_end[rw], jiffies); | |
564 | } | |
565 | ||
0f3457f6 | 566 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
567 | { |
568 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 569 | tg->io_disp[rw] = 0; |
e43473b7 VG |
570 | tg->slice_start[rw] = jiffies; |
571 | tg->slice_end[rw] = jiffies + throtl_slice; | |
fda6f272 TH |
572 | throtl_log(&tg->service_queue, |
573 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
574 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
575 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
576 | } |
577 | ||
0f3457f6 TH |
578 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
579 | unsigned long jiffy_end) | |
d1ae8ffd VG |
580 | { |
581 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
582 | } | |
583 | ||
0f3457f6 TH |
584 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
585 | unsigned long jiffy_end) | |
e43473b7 VG |
586 | { |
587 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
fda6f272 TH |
588 | throtl_log(&tg->service_queue, |
589 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
590 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
591 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
592 | } |
593 | ||
594 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 595 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
596 | { |
597 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
5cf8c227 | 598 | return false; |
e43473b7 VG |
599 | |
600 | return 1; | |
601 | } | |
602 | ||
603 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 604 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 605 | { |
3aad5d3e VG |
606 | unsigned long nr_slices, time_elapsed, io_trim; |
607 | u64 bytes_trim, tmp; | |
e43473b7 VG |
608 | |
609 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
610 | ||
611 | /* | |
612 | * If bps are unlimited (-1), then time slice don't get | |
613 | * renewed. Don't try to trim the slice if slice is used. A new | |
614 | * slice will start when appropriate. | |
615 | */ | |
0f3457f6 | 616 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
617 | return; |
618 | ||
d1ae8ffd VG |
619 | /* |
620 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
621 | * that initially cgroup limit was very low resulting in high | |
622 | * slice_end, but later limit was bumped up and bio was dispached | |
623 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
624 | * is bad because it does not allow new slice to start. | |
625 | */ | |
626 | ||
0f3457f6 | 627 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); |
d1ae8ffd | 628 | |
e43473b7 VG |
629 | time_elapsed = jiffies - tg->slice_start[rw]; |
630 | ||
631 | nr_slices = time_elapsed / throtl_slice; | |
632 | ||
633 | if (!nr_slices) | |
634 | return; | |
3aad5d3e VG |
635 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
636 | do_div(tmp, HZ); | |
637 | bytes_trim = tmp; | |
e43473b7 | 638 | |
8e89d13f | 639 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 640 | |
8e89d13f | 641 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
642 | return; |
643 | ||
644 | if (tg->bytes_disp[rw] >= bytes_trim) | |
645 | tg->bytes_disp[rw] -= bytes_trim; | |
646 | else | |
647 | tg->bytes_disp[rw] = 0; | |
648 | ||
8e89d13f VG |
649 | if (tg->io_disp[rw] >= io_trim) |
650 | tg->io_disp[rw] -= io_trim; | |
651 | else | |
652 | tg->io_disp[rw] = 0; | |
653 | ||
e43473b7 VG |
654 | tg->slice_start[rw] += nr_slices * throtl_slice; |
655 | ||
fda6f272 TH |
656 | throtl_log(&tg->service_queue, |
657 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
658 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
659 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
e43473b7 VG |
660 | } |
661 | ||
0f3457f6 TH |
662 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
663 | unsigned long *wait) | |
e43473b7 VG |
664 | { |
665 | bool rw = bio_data_dir(bio); | |
8e89d13f | 666 | unsigned int io_allowed; |
e43473b7 | 667 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 668 | u64 tmp; |
e43473b7 | 669 | |
8e89d13f | 670 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 671 | |
8e89d13f VG |
672 | /* Slice has just started. Consider one slice interval */ |
673 | if (!jiffy_elapsed) | |
674 | jiffy_elapsed_rnd = throtl_slice; | |
675 | ||
676 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
677 | ||
c49c06e4 VG |
678 | /* |
679 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
680 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
681 | * will allow dispatch after 1 second and after that slice should | |
682 | * have been trimmed. | |
683 | */ | |
684 | ||
685 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
686 | do_div(tmp, HZ); | |
687 | ||
688 | if (tmp > UINT_MAX) | |
689 | io_allowed = UINT_MAX; | |
690 | else | |
691 | io_allowed = tmp; | |
8e89d13f VG |
692 | |
693 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
694 | if (wait) |
695 | *wait = 0; | |
5cf8c227 | 696 | return true; |
e43473b7 VG |
697 | } |
698 | ||
8e89d13f VG |
699 | /* Calc approx time to dispatch */ |
700 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
701 | ||
702 | if (jiffy_wait > jiffy_elapsed) | |
703 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
704 | else | |
705 | jiffy_wait = 1; | |
706 | ||
707 | if (wait) | |
708 | *wait = jiffy_wait; | |
709 | return 0; | |
710 | } | |
711 | ||
0f3457f6 TH |
712 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
713 | unsigned long *wait) | |
8e89d13f VG |
714 | { |
715 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 716 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 717 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
718 | |
719 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
720 | ||
721 | /* Slice has just started. Consider one slice interval */ | |
722 | if (!jiffy_elapsed) | |
723 | jiffy_elapsed_rnd = throtl_slice; | |
724 | ||
725 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
726 | ||
5e901a2b VG |
727 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
728 | do_div(tmp, HZ); | |
3aad5d3e | 729 | bytes_allowed = tmp; |
e43473b7 | 730 | |
4f024f37 | 731 | if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) { |
e43473b7 VG |
732 | if (wait) |
733 | *wait = 0; | |
5cf8c227 | 734 | return true; |
e43473b7 VG |
735 | } |
736 | ||
737 | /* Calc approx time to dispatch */ | |
4f024f37 | 738 | extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed; |
e43473b7 VG |
739 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); |
740 | ||
741 | if (!jiffy_wait) | |
742 | jiffy_wait = 1; | |
743 | ||
744 | /* | |
745 | * This wait time is without taking into consideration the rounding | |
746 | * up we did. Add that time also. | |
747 | */ | |
748 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
749 | if (wait) |
750 | *wait = jiffy_wait; | |
8e89d13f VG |
751 | return 0; |
752 | } | |
753 | ||
754 | /* | |
755 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
756 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
757 | */ | |
0f3457f6 TH |
758 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
759 | unsigned long *wait) | |
8e89d13f VG |
760 | { |
761 | bool rw = bio_data_dir(bio); | |
762 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
763 | ||
764 | /* | |
765 | * Currently whole state machine of group depends on first bio | |
766 | * queued in the group bio list. So one should not be calling | |
767 | * this function with a different bio if there are other bios | |
768 | * queued. | |
769 | */ | |
73f0d49a | 770 | BUG_ON(tg->service_queue.nr_queued[rw] && |
c5cc2070 | 771 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); |
e43473b7 | 772 | |
8e89d13f VG |
773 | /* If tg->bps = -1, then BW is unlimited */ |
774 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
775 | if (wait) | |
776 | *wait = 0; | |
5cf8c227 | 777 | return true; |
8e89d13f VG |
778 | } |
779 | ||
780 | /* | |
781 | * If previous slice expired, start a new one otherwise renew/extend | |
782 | * existing slice to make sure it is at least throtl_slice interval | |
164c80ed VG |
783 | * long since now. New slice is started only for empty throttle group. |
784 | * If there is queued bio, that means there should be an active | |
785 | * slice and it should be extended instead. | |
8e89d13f | 786 | */ |
164c80ed | 787 | if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw])) |
0f3457f6 | 788 | throtl_start_new_slice(tg, rw); |
8e89d13f VG |
789 | else { |
790 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
0f3457f6 | 791 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); |
8e89d13f VG |
792 | } |
793 | ||
0f3457f6 TH |
794 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
795 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
796 | if (wait) |
797 | *wait = 0; | |
798 | return 1; | |
799 | } | |
800 | ||
801 | max_wait = max(bps_wait, iops_wait); | |
802 | ||
803 | if (wait) | |
804 | *wait = max_wait; | |
805 | ||
806 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 807 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
808 | |
809 | return 0; | |
810 | } | |
811 | ||
812 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
813 | { | |
814 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
815 | |
816 | /* Charge the bio to the group */ | |
4f024f37 | 817 | tg->bytes_disp[rw] += bio->bi_iter.bi_size; |
8e89d13f | 818 | tg->io_disp[rw]++; |
e43473b7 | 819 | |
2a0f61e6 | 820 | /* |
8d2bbd4c | 821 | * BIO_THROTTLED is used to prevent the same bio to be throttled |
2a0f61e6 TH |
822 | * more than once as a throttled bio will go through blk-throtl the |
823 | * second time when it eventually gets issued. Set it when a bio | |
824 | * is being charged to a tg. | |
2a0f61e6 | 825 | */ |
8d2bbd4c CH |
826 | if (!bio_flagged(bio, BIO_THROTTLED)) |
827 | bio_set_flag(bio, BIO_THROTTLED); | |
e43473b7 VG |
828 | } |
829 | ||
c5cc2070 TH |
830 | /** |
831 | * throtl_add_bio_tg - add a bio to the specified throtl_grp | |
832 | * @bio: bio to add | |
833 | * @qn: qnode to use | |
834 | * @tg: the target throtl_grp | |
835 | * | |
836 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, | |
837 | * tg->qnode_on_self[] is used. | |
838 | */ | |
839 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, | |
840 | struct throtl_grp *tg) | |
e43473b7 | 841 | { |
73f0d49a | 842 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
843 | bool rw = bio_data_dir(bio); |
844 | ||
c5cc2070 TH |
845 | if (!qn) |
846 | qn = &tg->qnode_on_self[rw]; | |
847 | ||
0e9f4164 TH |
848 | /* |
849 | * If @tg doesn't currently have any bios queued in the same | |
850 | * direction, queueing @bio can change when @tg should be | |
851 | * dispatched. Mark that @tg was empty. This is automatically | |
852 | * cleaered on the next tg_update_disptime(). | |
853 | */ | |
854 | if (!sq->nr_queued[rw]) | |
855 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
856 | ||
c5cc2070 TH |
857 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); |
858 | ||
73f0d49a | 859 | sq->nr_queued[rw]++; |
77216b04 | 860 | throtl_enqueue_tg(tg); |
e43473b7 VG |
861 | } |
862 | ||
77216b04 | 863 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 864 | { |
73f0d49a | 865 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
866 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
867 | struct bio *bio; | |
868 | ||
c5cc2070 | 869 | if ((bio = throtl_peek_queued(&sq->queued[READ]))) |
0f3457f6 | 870 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 871 | |
c5cc2070 | 872 | if ((bio = throtl_peek_queued(&sq->queued[WRITE]))) |
0f3457f6 | 873 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
874 | |
875 | min_wait = min(read_wait, write_wait); | |
876 | disptime = jiffies + min_wait; | |
877 | ||
e43473b7 | 878 | /* Update dispatch time */ |
77216b04 | 879 | throtl_dequeue_tg(tg); |
e43473b7 | 880 | tg->disptime = disptime; |
77216b04 | 881 | throtl_enqueue_tg(tg); |
0e9f4164 TH |
882 | |
883 | /* see throtl_add_bio_tg() */ | |
884 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
885 | } |
886 | ||
32ee5bc4 VG |
887 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, |
888 | struct throtl_grp *parent_tg, bool rw) | |
889 | { | |
890 | if (throtl_slice_used(parent_tg, rw)) { | |
891 | throtl_start_new_slice_with_credit(parent_tg, rw, | |
892 | child_tg->slice_start[rw]); | |
893 | } | |
894 | ||
895 | } | |
896 | ||
77216b04 | 897 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 898 | { |
73f0d49a | 899 | struct throtl_service_queue *sq = &tg->service_queue; |
6bc9c2b4 TH |
900 | struct throtl_service_queue *parent_sq = sq->parent_sq; |
901 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
c5cc2070 | 902 | struct throtl_grp *tg_to_put = NULL; |
e43473b7 VG |
903 | struct bio *bio; |
904 | ||
c5cc2070 TH |
905 | /* |
906 | * @bio is being transferred from @tg to @parent_sq. Popping a bio | |
907 | * from @tg may put its reference and @parent_sq might end up | |
908 | * getting released prematurely. Remember the tg to put and put it | |
909 | * after @bio is transferred to @parent_sq. | |
910 | */ | |
911 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); | |
73f0d49a | 912 | sq->nr_queued[rw]--; |
e43473b7 VG |
913 | |
914 | throtl_charge_bio(tg, bio); | |
6bc9c2b4 TH |
915 | |
916 | /* | |
917 | * If our parent is another tg, we just need to transfer @bio to | |
918 | * the parent using throtl_add_bio_tg(). If our parent is | |
919 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
920 | * bio_lists[] and decrease total number queued. The caller is | |
921 | * responsible for issuing these bios. | |
922 | */ | |
923 | if (parent_tg) { | |
c5cc2070 | 924 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); |
32ee5bc4 | 925 | start_parent_slice_with_credit(tg, parent_tg, rw); |
6bc9c2b4 | 926 | } else { |
c5cc2070 TH |
927 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], |
928 | &parent_sq->queued[rw]); | |
6bc9c2b4 TH |
929 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
930 | tg->td->nr_queued[rw]--; | |
931 | } | |
e43473b7 | 932 | |
0f3457f6 | 933 | throtl_trim_slice(tg, rw); |
6bc9c2b4 | 934 | |
c5cc2070 TH |
935 | if (tg_to_put) |
936 | blkg_put(tg_to_blkg(tg_to_put)); | |
e43473b7 VG |
937 | } |
938 | ||
77216b04 | 939 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 940 | { |
73f0d49a | 941 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
942 | unsigned int nr_reads = 0, nr_writes = 0; |
943 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 944 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
945 | struct bio *bio; |
946 | ||
947 | /* Try to dispatch 75% READS and 25% WRITES */ | |
948 | ||
c5cc2070 | 949 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && |
0f3457f6 | 950 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 951 | |
77216b04 | 952 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
953 | nr_reads++; |
954 | ||
955 | if (nr_reads >= max_nr_reads) | |
956 | break; | |
957 | } | |
958 | ||
c5cc2070 | 959 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && |
0f3457f6 | 960 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 961 | |
77216b04 | 962 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
963 | nr_writes++; |
964 | ||
965 | if (nr_writes >= max_nr_writes) | |
966 | break; | |
967 | } | |
968 | ||
969 | return nr_reads + nr_writes; | |
970 | } | |
971 | ||
651930bc | 972 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
973 | { |
974 | unsigned int nr_disp = 0; | |
e43473b7 VG |
975 | |
976 | while (1) { | |
73f0d49a TH |
977 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
978 | struct throtl_service_queue *sq = &tg->service_queue; | |
e43473b7 VG |
979 | |
980 | if (!tg) | |
981 | break; | |
982 | ||
983 | if (time_before(jiffies, tg->disptime)) | |
984 | break; | |
985 | ||
77216b04 | 986 | throtl_dequeue_tg(tg); |
e43473b7 | 987 | |
77216b04 | 988 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 989 | |
73f0d49a | 990 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
77216b04 | 991 | tg_update_disptime(tg); |
e43473b7 VG |
992 | |
993 | if (nr_disp >= throtl_quantum) | |
994 | break; | |
995 | } | |
996 | ||
997 | return nr_disp; | |
998 | } | |
999 | ||
6e1a5704 TH |
1000 | /** |
1001 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
1002 | * @arg: the throtl_service_queue being serviced | |
1003 | * | |
1004 | * This timer is armed when a child throtl_grp with active bio's become | |
1005 | * pending and queued on the service_queue's pending_tree and expires when | |
1006 | * the first child throtl_grp should be dispatched. This function | |
2e48a530 TH |
1007 | * dispatches bio's from the children throtl_grps to the parent |
1008 | * service_queue. | |
1009 | * | |
1010 | * If the parent's parent is another throtl_grp, dispatching is propagated | |
1011 | * by either arming its pending_timer or repeating dispatch directly. If | |
1012 | * the top-level service_tree is reached, throtl_data->dispatch_work is | |
1013 | * kicked so that the ready bio's are issued. | |
6e1a5704 | 1014 | */ |
69df0ab0 TH |
1015 | static void throtl_pending_timer_fn(unsigned long arg) |
1016 | { | |
1017 | struct throtl_service_queue *sq = (void *)arg; | |
2e48a530 | 1018 | struct throtl_grp *tg = sq_to_tg(sq); |
69df0ab0 | 1019 | struct throtl_data *td = sq_to_td(sq); |
cb76199c | 1020 | struct request_queue *q = td->queue; |
2e48a530 TH |
1021 | struct throtl_service_queue *parent_sq; |
1022 | bool dispatched; | |
6e1a5704 | 1023 | int ret; |
e43473b7 VG |
1024 | |
1025 | spin_lock_irq(q->queue_lock); | |
2e48a530 TH |
1026 | again: |
1027 | parent_sq = sq->parent_sq; | |
1028 | dispatched = false; | |
e43473b7 | 1029 | |
7f52f98c TH |
1030 | while (true) { |
1031 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
2e48a530 TH |
1032 | sq->nr_queued[READ] + sq->nr_queued[WRITE], |
1033 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
7f52f98c TH |
1034 | |
1035 | ret = throtl_select_dispatch(sq); | |
1036 | if (ret) { | |
7f52f98c TH |
1037 | throtl_log(sq, "bios disp=%u", ret); |
1038 | dispatched = true; | |
1039 | } | |
e43473b7 | 1040 | |
7f52f98c TH |
1041 | if (throtl_schedule_next_dispatch(sq, false)) |
1042 | break; | |
e43473b7 | 1043 | |
7f52f98c TH |
1044 | /* this dispatch windows is still open, relax and repeat */ |
1045 | spin_unlock_irq(q->queue_lock); | |
1046 | cpu_relax(); | |
1047 | spin_lock_irq(q->queue_lock); | |
651930bc | 1048 | } |
e43473b7 | 1049 | |
2e48a530 TH |
1050 | if (!dispatched) |
1051 | goto out_unlock; | |
6e1a5704 | 1052 | |
2e48a530 TH |
1053 | if (parent_sq) { |
1054 | /* @parent_sq is another throl_grp, propagate dispatch */ | |
1055 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1056 | tg_update_disptime(tg); | |
1057 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { | |
1058 | /* window is already open, repeat dispatching */ | |
1059 | sq = parent_sq; | |
1060 | tg = sq_to_tg(sq); | |
1061 | goto again; | |
1062 | } | |
1063 | } | |
1064 | } else { | |
1065 | /* reached the top-level, queue issueing */ | |
1066 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1067 | } | |
1068 | out_unlock: | |
e43473b7 | 1069 | spin_unlock_irq(q->queue_lock); |
6e1a5704 | 1070 | } |
e43473b7 | 1071 | |
6e1a5704 TH |
1072 | /** |
1073 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
1074 | * @work: work item being executed | |
1075 | * | |
1076 | * This function is queued for execution when bio's reach the bio_lists[] | |
1077 | * of throtl_data->service_queue. Those bio's are ready and issued by this | |
1078 | * function. | |
1079 | */ | |
8876e140 | 1080 | static void blk_throtl_dispatch_work_fn(struct work_struct *work) |
6e1a5704 TH |
1081 | { |
1082 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1083 | dispatch_work); | |
1084 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1085 | struct request_queue *q = td->queue; | |
1086 | struct bio_list bio_list_on_stack; | |
1087 | struct bio *bio; | |
1088 | struct blk_plug plug; | |
1089 | int rw; | |
1090 | ||
1091 | bio_list_init(&bio_list_on_stack); | |
1092 | ||
1093 | spin_lock_irq(q->queue_lock); | |
c5cc2070 TH |
1094 | for (rw = READ; rw <= WRITE; rw++) |
1095 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) | |
1096 | bio_list_add(&bio_list_on_stack, bio); | |
6e1a5704 TH |
1097 | spin_unlock_irq(q->queue_lock); |
1098 | ||
1099 | if (!bio_list_empty(&bio_list_on_stack)) { | |
69d60eb9 | 1100 | blk_start_plug(&plug); |
e43473b7 VG |
1101 | while((bio = bio_list_pop(&bio_list_on_stack))) |
1102 | generic_make_request(bio); | |
69d60eb9 | 1103 | blk_finish_plug(&plug); |
e43473b7 | 1104 | } |
e43473b7 VG |
1105 | } |
1106 | ||
f95a04af TH |
1107 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
1108 | int off) | |
60c2bc2d | 1109 | { |
f95a04af TH |
1110 | struct throtl_grp *tg = pd_to_tg(pd); |
1111 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 1112 | |
af133ceb | 1113 | if (v == -1) |
60c2bc2d | 1114 | return 0; |
f95a04af | 1115 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
1116 | } |
1117 | ||
f95a04af TH |
1118 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
1119 | int off) | |
e43473b7 | 1120 | { |
f95a04af TH |
1121 | struct throtl_grp *tg = pd_to_tg(pd); |
1122 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 1123 | |
af133ceb TH |
1124 | if (v == -1) |
1125 | return 0; | |
f95a04af | 1126 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
1127 | } |
1128 | ||
2da8ca82 | 1129 | static int tg_print_conf_u64(struct seq_file *sf, void *v) |
8e89d13f | 1130 | { |
2da8ca82 TH |
1131 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64, |
1132 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1133 | return 0; |
8e89d13f VG |
1134 | } |
1135 | ||
2da8ca82 | 1136 | static int tg_print_conf_uint(struct seq_file *sf, void *v) |
8e89d13f | 1137 | { |
2da8ca82 TH |
1138 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint, |
1139 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1140 | return 0; |
60c2bc2d TH |
1141 | } |
1142 | ||
69948b07 | 1143 | static void tg_conf_updated(struct throtl_grp *tg) |
60c2bc2d | 1144 | { |
69948b07 | 1145 | struct throtl_service_queue *sq = &tg->service_queue; |
492eb21b | 1146 | struct cgroup_subsys_state *pos_css; |
69948b07 | 1147 | struct blkcg_gq *blkg; |
af133ceb | 1148 | |
fda6f272 TH |
1149 | throtl_log(&tg->service_queue, |
1150 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
1151 | tg->bps[READ], tg->bps[WRITE], | |
1152 | tg->iops[READ], tg->iops[WRITE]); | |
632b4493 | 1153 | |
693e751e TH |
1154 | /* |
1155 | * Update has_rules[] flags for the updated tg's subtree. A tg is | |
1156 | * considered to have rules if either the tg itself or any of its | |
1157 | * ancestors has rules. This identifies groups without any | |
1158 | * restrictions in the whole hierarchy and allows them to bypass | |
1159 | * blk-throttle. | |
1160 | */ | |
69948b07 | 1161 | blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg)) |
693e751e TH |
1162 | tg_update_has_rules(blkg_to_tg(blkg)); |
1163 | ||
632b4493 TH |
1164 | /* |
1165 | * We're already holding queue_lock and know @tg is valid. Let's | |
1166 | * apply the new config directly. | |
1167 | * | |
1168 | * Restart the slices for both READ and WRITES. It might happen | |
1169 | * that a group's limit are dropped suddenly and we don't want to | |
1170 | * account recently dispatched IO with new low rate. | |
1171 | */ | |
0f3457f6 TH |
1172 | throtl_start_new_slice(tg, 0); |
1173 | throtl_start_new_slice(tg, 1); | |
632b4493 | 1174 | |
5b2c16aa | 1175 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 1176 | tg_update_disptime(tg); |
7f52f98c | 1177 | throtl_schedule_next_dispatch(sq->parent_sq, true); |
632b4493 | 1178 | } |
69948b07 TH |
1179 | } |
1180 | ||
1181 | static ssize_t tg_set_conf(struct kernfs_open_file *of, | |
1182 | char *buf, size_t nbytes, loff_t off, bool is_u64) | |
1183 | { | |
1184 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1185 | struct blkg_conf_ctx ctx; | |
1186 | struct throtl_grp *tg; | |
1187 | int ret; | |
1188 | u64 v; | |
1189 | ||
1190 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1191 | if (ret) | |
1192 | return ret; | |
1193 | ||
1194 | ret = -EINVAL; | |
1195 | if (sscanf(ctx.body, "%llu", &v) != 1) | |
1196 | goto out_finish; | |
1197 | if (!v) | |
1198 | v = -1; | |
1199 | ||
1200 | tg = blkg_to_tg(ctx.blkg); | |
1201 | ||
1202 | if (is_u64) | |
1203 | *(u64 *)((void *)tg + of_cft(of)->private) = v; | |
1204 | else | |
1205 | *(unsigned int *)((void *)tg + of_cft(of)->private) = v; | |
60c2bc2d | 1206 | |
69948b07 | 1207 | tg_conf_updated(tg); |
36aa9e5f TH |
1208 | ret = 0; |
1209 | out_finish: | |
60c2bc2d | 1210 | blkg_conf_finish(&ctx); |
36aa9e5f | 1211 | return ret ?: nbytes; |
8e89d13f VG |
1212 | } |
1213 | ||
451af504 TH |
1214 | static ssize_t tg_set_conf_u64(struct kernfs_open_file *of, |
1215 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1216 | { |
451af504 | 1217 | return tg_set_conf(of, buf, nbytes, off, true); |
60c2bc2d TH |
1218 | } |
1219 | ||
451af504 TH |
1220 | static ssize_t tg_set_conf_uint(struct kernfs_open_file *of, |
1221 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1222 | { |
451af504 | 1223 | return tg_set_conf(of, buf, nbytes, off, false); |
60c2bc2d TH |
1224 | } |
1225 | ||
880f50e2 | 1226 | static struct cftype throtl_legacy_files[] = { |
60c2bc2d TH |
1227 | { |
1228 | .name = "throttle.read_bps_device", | |
af133ceb | 1229 | .private = offsetof(struct throtl_grp, bps[READ]), |
2da8ca82 | 1230 | .seq_show = tg_print_conf_u64, |
451af504 | 1231 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1232 | }, |
1233 | { | |
1234 | .name = "throttle.write_bps_device", | |
af133ceb | 1235 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
2da8ca82 | 1236 | .seq_show = tg_print_conf_u64, |
451af504 | 1237 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1238 | }, |
1239 | { | |
1240 | .name = "throttle.read_iops_device", | |
af133ceb | 1241 | .private = offsetof(struct throtl_grp, iops[READ]), |
2da8ca82 | 1242 | .seq_show = tg_print_conf_uint, |
451af504 | 1243 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1244 | }, |
1245 | { | |
1246 | .name = "throttle.write_iops_device", | |
af133ceb | 1247 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
2da8ca82 | 1248 | .seq_show = tg_print_conf_uint, |
451af504 | 1249 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1250 | }, |
1251 | { | |
1252 | .name = "throttle.io_service_bytes", | |
77ea7338 TH |
1253 | .private = (unsigned long)&blkcg_policy_throtl, |
1254 | .seq_show = blkg_print_stat_bytes, | |
60c2bc2d TH |
1255 | }, |
1256 | { | |
1257 | .name = "throttle.io_serviced", | |
77ea7338 TH |
1258 | .private = (unsigned long)&blkcg_policy_throtl, |
1259 | .seq_show = blkg_print_stat_ios, | |
60c2bc2d TH |
1260 | }, |
1261 | { } /* terminate */ | |
1262 | }; | |
1263 | ||
2ee867dc TH |
1264 | static u64 tg_prfill_max(struct seq_file *sf, struct blkg_policy_data *pd, |
1265 | int off) | |
1266 | { | |
1267 | struct throtl_grp *tg = pd_to_tg(pd); | |
1268 | const char *dname = blkg_dev_name(pd->blkg); | |
1269 | char bufs[4][21] = { "max", "max", "max", "max" }; | |
1270 | ||
1271 | if (!dname) | |
1272 | return 0; | |
1273 | if (tg->bps[READ] == -1 && tg->bps[WRITE] == -1 && | |
1274 | tg->iops[READ] == -1 && tg->iops[WRITE] == -1) | |
1275 | return 0; | |
1276 | ||
1277 | if (tg->bps[READ] != -1) | |
1278 | snprintf(bufs[0], sizeof(bufs[0]), "%llu", tg->bps[READ]); | |
1279 | if (tg->bps[WRITE] != -1) | |
1280 | snprintf(bufs[1], sizeof(bufs[1]), "%llu", tg->bps[WRITE]); | |
1281 | if (tg->iops[READ] != -1) | |
1282 | snprintf(bufs[2], sizeof(bufs[2]), "%u", tg->iops[READ]); | |
1283 | if (tg->iops[WRITE] != -1) | |
1284 | snprintf(bufs[3], sizeof(bufs[3]), "%u", tg->iops[WRITE]); | |
1285 | ||
1286 | seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s\n", | |
1287 | dname, bufs[0], bufs[1], bufs[2], bufs[3]); | |
1288 | return 0; | |
1289 | } | |
1290 | ||
1291 | static int tg_print_max(struct seq_file *sf, void *v) | |
1292 | { | |
1293 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_max, | |
1294 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
1295 | return 0; | |
1296 | } | |
1297 | ||
1298 | static ssize_t tg_set_max(struct kernfs_open_file *of, | |
1299 | char *buf, size_t nbytes, loff_t off) | |
1300 | { | |
1301 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1302 | struct blkg_conf_ctx ctx; | |
1303 | struct throtl_grp *tg; | |
1304 | u64 v[4]; | |
1305 | int ret; | |
1306 | ||
1307 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1308 | if (ret) | |
1309 | return ret; | |
1310 | ||
1311 | tg = blkg_to_tg(ctx.blkg); | |
1312 | ||
1313 | v[0] = tg->bps[READ]; | |
1314 | v[1] = tg->bps[WRITE]; | |
1315 | v[2] = tg->iops[READ]; | |
1316 | v[3] = tg->iops[WRITE]; | |
1317 | ||
1318 | while (true) { | |
1319 | char tok[27]; /* wiops=18446744073709551616 */ | |
1320 | char *p; | |
1321 | u64 val = -1; | |
1322 | int len; | |
1323 | ||
1324 | if (sscanf(ctx.body, "%26s%n", tok, &len) != 1) | |
1325 | break; | |
1326 | if (tok[0] == '\0') | |
1327 | break; | |
1328 | ctx.body += len; | |
1329 | ||
1330 | ret = -EINVAL; | |
1331 | p = tok; | |
1332 | strsep(&p, "="); | |
1333 | if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max"))) | |
1334 | goto out_finish; | |
1335 | ||
1336 | ret = -ERANGE; | |
1337 | if (!val) | |
1338 | goto out_finish; | |
1339 | ||
1340 | ret = -EINVAL; | |
1341 | if (!strcmp(tok, "rbps")) | |
1342 | v[0] = val; | |
1343 | else if (!strcmp(tok, "wbps")) | |
1344 | v[1] = val; | |
1345 | else if (!strcmp(tok, "riops")) | |
1346 | v[2] = min_t(u64, val, UINT_MAX); | |
1347 | else if (!strcmp(tok, "wiops")) | |
1348 | v[3] = min_t(u64, val, UINT_MAX); | |
1349 | else | |
1350 | goto out_finish; | |
1351 | } | |
1352 | ||
1353 | tg->bps[READ] = v[0]; | |
1354 | tg->bps[WRITE] = v[1]; | |
1355 | tg->iops[READ] = v[2]; | |
1356 | tg->iops[WRITE] = v[3]; | |
1357 | ||
1358 | tg_conf_updated(tg); | |
1359 | ret = 0; | |
1360 | out_finish: | |
1361 | blkg_conf_finish(&ctx); | |
1362 | return ret ?: nbytes; | |
1363 | } | |
1364 | ||
1365 | static struct cftype throtl_files[] = { | |
1366 | { | |
1367 | .name = "max", | |
1368 | .flags = CFTYPE_NOT_ON_ROOT, | |
1369 | .seq_show = tg_print_max, | |
1370 | .write = tg_set_max, | |
1371 | }, | |
1372 | { } /* terminate */ | |
1373 | }; | |
1374 | ||
da527770 | 1375 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1376 | { |
1377 | struct throtl_data *td = q->td; | |
1378 | ||
69df0ab0 | 1379 | cancel_work_sync(&td->dispatch_work); |
e43473b7 VG |
1380 | } |
1381 | ||
3c798398 | 1382 | static struct blkcg_policy blkcg_policy_throtl = { |
2ee867dc | 1383 | .dfl_cftypes = throtl_files, |
880f50e2 | 1384 | .legacy_cftypes = throtl_legacy_files, |
f9fcc2d3 | 1385 | |
001bea73 | 1386 | .pd_alloc_fn = throtl_pd_alloc, |
f9fcc2d3 | 1387 | .pd_init_fn = throtl_pd_init, |
693e751e | 1388 | .pd_online_fn = throtl_pd_online, |
001bea73 | 1389 | .pd_free_fn = throtl_pd_free, |
e43473b7 VG |
1390 | }; |
1391 | ||
ae118896 TH |
1392 | bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg, |
1393 | struct bio *bio) | |
e43473b7 | 1394 | { |
c5cc2070 | 1395 | struct throtl_qnode *qn = NULL; |
ae118896 | 1396 | struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg); |
73f0d49a | 1397 | struct throtl_service_queue *sq; |
0e9f4164 | 1398 | bool rw = bio_data_dir(bio); |
bc16a4f9 | 1399 | bool throttled = false; |
e43473b7 | 1400 | |
ae118896 TH |
1401 | WARN_ON_ONCE(!rcu_read_lock_held()); |
1402 | ||
2a0f61e6 | 1403 | /* see throtl_charge_bio() */ |
8d2bbd4c | 1404 | if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw]) |
bc16a4f9 | 1405 | goto out; |
e43473b7 VG |
1406 | |
1407 | spin_lock_irq(q->queue_lock); | |
c9589f03 TH |
1408 | |
1409 | if (unlikely(blk_queue_bypass(q))) | |
bc16a4f9 | 1410 | goto out_unlock; |
f469a7b4 | 1411 | |
73f0d49a TH |
1412 | sq = &tg->service_queue; |
1413 | ||
9e660acf TH |
1414 | while (true) { |
1415 | /* throtl is FIFO - if bios are already queued, should queue */ | |
1416 | if (sq->nr_queued[rw]) | |
1417 | break; | |
de701c74 | 1418 | |
9e660acf TH |
1419 | /* if above limits, break to queue */ |
1420 | if (!tg_may_dispatch(tg, bio, NULL)) | |
1421 | break; | |
1422 | ||
1423 | /* within limits, let's charge and dispatch directly */ | |
e43473b7 | 1424 | throtl_charge_bio(tg, bio); |
04521db0 VG |
1425 | |
1426 | /* | |
1427 | * We need to trim slice even when bios are not being queued | |
1428 | * otherwise it might happen that a bio is not queued for | |
1429 | * a long time and slice keeps on extending and trim is not | |
1430 | * called for a long time. Now if limits are reduced suddenly | |
1431 | * we take into account all the IO dispatched so far at new | |
1432 | * low rate and * newly queued IO gets a really long dispatch | |
1433 | * time. | |
1434 | * | |
1435 | * So keep on trimming slice even if bio is not queued. | |
1436 | */ | |
0f3457f6 | 1437 | throtl_trim_slice(tg, rw); |
9e660acf TH |
1438 | |
1439 | /* | |
1440 | * @bio passed through this layer without being throttled. | |
1441 | * Climb up the ladder. If we''re already at the top, it | |
1442 | * can be executed directly. | |
1443 | */ | |
c5cc2070 | 1444 | qn = &tg->qnode_on_parent[rw]; |
9e660acf TH |
1445 | sq = sq->parent_sq; |
1446 | tg = sq_to_tg(sq); | |
1447 | if (!tg) | |
1448 | goto out_unlock; | |
e43473b7 VG |
1449 | } |
1450 | ||
9e660acf | 1451 | /* out-of-limit, queue to @tg */ |
fda6f272 TH |
1452 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
1453 | rw == READ ? 'R' : 'W', | |
4f024f37 | 1454 | tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw], |
fda6f272 TH |
1455 | tg->io_disp[rw], tg->iops[rw], |
1456 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
e43473b7 | 1457 | |
671058fb | 1458 | bio_associate_current(bio); |
6bc9c2b4 | 1459 | tg->td->nr_queued[rw]++; |
c5cc2070 | 1460 | throtl_add_bio_tg(bio, qn, tg); |
bc16a4f9 | 1461 | throttled = true; |
e43473b7 | 1462 | |
7f52f98c TH |
1463 | /* |
1464 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
1465 | * was empty before @bio. The forced scheduling isn't likely to | |
1466 | * cause undue delay as @bio is likely to be dispatched directly if | |
1467 | * its @tg's disptime is not in the future. | |
1468 | */ | |
0e9f4164 | 1469 | if (tg->flags & THROTL_TG_WAS_EMPTY) { |
77216b04 | 1470 | tg_update_disptime(tg); |
7f52f98c | 1471 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); |
e43473b7 VG |
1472 | } |
1473 | ||
bc16a4f9 | 1474 | out_unlock: |
e43473b7 | 1475 | spin_unlock_irq(q->queue_lock); |
bc16a4f9 | 1476 | out: |
2a0f61e6 TH |
1477 | /* |
1478 | * As multiple blk-throtls may stack in the same issue path, we | |
1479 | * don't want bios to leave with the flag set. Clear the flag if | |
1480 | * being issued. | |
1481 | */ | |
1482 | if (!throttled) | |
8d2bbd4c | 1483 | bio_clear_flag(bio, BIO_THROTTLED); |
bc16a4f9 | 1484 | return throttled; |
e43473b7 VG |
1485 | } |
1486 | ||
2a12f0dc TH |
1487 | /* |
1488 | * Dispatch all bios from all children tg's queued on @parent_sq. On | |
1489 | * return, @parent_sq is guaranteed to not have any active children tg's | |
1490 | * and all bios from previously active tg's are on @parent_sq->bio_lists[]. | |
1491 | */ | |
1492 | static void tg_drain_bios(struct throtl_service_queue *parent_sq) | |
1493 | { | |
1494 | struct throtl_grp *tg; | |
1495 | ||
1496 | while ((tg = throtl_rb_first(parent_sq))) { | |
1497 | struct throtl_service_queue *sq = &tg->service_queue; | |
1498 | struct bio *bio; | |
1499 | ||
1500 | throtl_dequeue_tg(tg); | |
1501 | ||
c5cc2070 | 1502 | while ((bio = throtl_peek_queued(&sq->queued[READ]))) |
2a12f0dc | 1503 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
c5cc2070 | 1504 | while ((bio = throtl_peek_queued(&sq->queued[WRITE]))) |
2a12f0dc TH |
1505 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
1506 | } | |
1507 | } | |
1508 | ||
c9a929dd TH |
1509 | /** |
1510 | * blk_throtl_drain - drain throttled bios | |
1511 | * @q: request_queue to drain throttled bios for | |
1512 | * | |
1513 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1514 | */ | |
1515 | void blk_throtl_drain(struct request_queue *q) | |
1516 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1517 | { | |
1518 | struct throtl_data *td = q->td; | |
2a12f0dc | 1519 | struct blkcg_gq *blkg; |
492eb21b | 1520 | struct cgroup_subsys_state *pos_css; |
c9a929dd | 1521 | struct bio *bio; |
651930bc | 1522 | int rw; |
c9a929dd | 1523 | |
8bcb6c7d | 1524 | queue_lockdep_assert_held(q); |
2a12f0dc | 1525 | rcu_read_lock(); |
c9a929dd | 1526 | |
2a12f0dc TH |
1527 | /* |
1528 | * Drain each tg while doing post-order walk on the blkg tree, so | |
1529 | * that all bios are propagated to td->service_queue. It'd be | |
1530 | * better to walk service_queue tree directly but blkg walk is | |
1531 | * easier. | |
1532 | */ | |
492eb21b | 1533 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) |
2a12f0dc | 1534 | tg_drain_bios(&blkg_to_tg(blkg)->service_queue); |
73f0d49a | 1535 | |
2a12f0dc TH |
1536 | /* finally, transfer bios from top-level tg's into the td */ |
1537 | tg_drain_bios(&td->service_queue); | |
1538 | ||
1539 | rcu_read_unlock(); | |
c9a929dd TH |
1540 | spin_unlock_irq(q->queue_lock); |
1541 | ||
2a12f0dc | 1542 | /* all bios now should be in td->service_queue, issue them */ |
651930bc | 1543 | for (rw = READ; rw <= WRITE; rw++) |
c5cc2070 TH |
1544 | while ((bio = throtl_pop_queued(&td->service_queue.queued[rw], |
1545 | NULL))) | |
651930bc | 1546 | generic_make_request(bio); |
c9a929dd TH |
1547 | |
1548 | spin_lock_irq(q->queue_lock); | |
1549 | } | |
1550 | ||
e43473b7 VG |
1551 | int blk_throtl_init(struct request_queue *q) |
1552 | { | |
1553 | struct throtl_data *td; | |
a2b1693b | 1554 | int ret; |
e43473b7 VG |
1555 | |
1556 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1557 | if (!td) | |
1558 | return -ENOMEM; | |
1559 | ||
69df0ab0 | 1560 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
b2ce2643 | 1561 | throtl_service_queue_init(&td->service_queue); |
e43473b7 | 1562 | |
cd1604fa | 1563 | q->td = td; |
29b12589 | 1564 | td->queue = q; |
02977e4a | 1565 | |
a2b1693b | 1566 | /* activate policy */ |
3c798398 | 1567 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1568 | if (ret) |
f51b802c | 1569 | kfree(td); |
a2b1693b | 1570 | return ret; |
e43473b7 VG |
1571 | } |
1572 | ||
1573 | void blk_throtl_exit(struct request_queue *q) | |
1574 | { | |
c875f4d0 | 1575 | BUG_ON(!q->td); |
da527770 | 1576 | throtl_shutdown_wq(q); |
3c798398 | 1577 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1578 | kfree(q->td); |
e43473b7 VG |
1579 | } |
1580 | ||
1581 | static int __init throtl_init(void) | |
1582 | { | |
450adcbe VG |
1583 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1584 | if (!kthrotld_workqueue) | |
1585 | panic("Failed to create kthrotld\n"); | |
1586 | ||
3c798398 | 1587 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1588 | } |
1589 | ||
1590 | module_init(throtl_init); |