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 | ||
d61fcfa4 SL |
21 | /* Throttling is performed over a slice and after that slice is renewed */ |
22 | #define DFL_THROTL_SLICE_HD (HZ / 10) | |
23 | #define DFL_THROTL_SLICE_SSD (HZ / 50) | |
297e3d85 | 24 | #define MAX_THROTL_SLICE (HZ) |
e43473b7 | 25 | |
3c798398 | 26 | static struct blkcg_policy blkcg_policy_throtl; |
0381411e | 27 | |
450adcbe VG |
28 | /* A workqueue to queue throttle related work */ |
29 | static struct workqueue_struct *kthrotld_workqueue; | |
450adcbe | 30 | |
c5cc2070 TH |
31 | /* |
32 | * To implement hierarchical throttling, throtl_grps form a tree and bios | |
33 | * are dispatched upwards level by level until they reach the top and get | |
34 | * issued. When dispatching bios from the children and local group at each | |
35 | * level, if the bios are dispatched into a single bio_list, there's a risk | |
36 | * of a local or child group which can queue many bios at once filling up | |
37 | * the list starving others. | |
38 | * | |
39 | * To avoid such starvation, dispatched bios are queued separately | |
40 | * according to where they came from. When they are again dispatched to | |
41 | * the parent, they're popped in round-robin order so that no single source | |
42 | * hogs the dispatch window. | |
43 | * | |
44 | * throtl_qnode is used to keep the queued bios separated by their sources. | |
45 | * Bios are queued to throtl_qnode which in turn is queued to | |
46 | * throtl_service_queue and then dispatched in round-robin order. | |
47 | * | |
48 | * It's also used to track the reference counts on blkg's. A qnode always | |
49 | * belongs to a throtl_grp and gets queued on itself or the parent, so | |
50 | * incrementing the reference of the associated throtl_grp when a qnode is | |
51 | * queued and decrementing when dequeued is enough to keep the whole blkg | |
52 | * tree pinned while bios are in flight. | |
53 | */ | |
54 | struct throtl_qnode { | |
55 | struct list_head node; /* service_queue->queued[] */ | |
56 | struct bio_list bios; /* queued bios */ | |
57 | struct throtl_grp *tg; /* tg this qnode belongs to */ | |
58 | }; | |
59 | ||
c9e0332e | 60 | struct throtl_service_queue { |
77216b04 TH |
61 | struct throtl_service_queue *parent_sq; /* the parent service_queue */ |
62 | ||
73f0d49a TH |
63 | /* |
64 | * Bios queued directly to this service_queue or dispatched from | |
65 | * children throtl_grp's. | |
66 | */ | |
c5cc2070 | 67 | struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */ |
73f0d49a TH |
68 | unsigned int nr_queued[2]; /* number of queued bios */ |
69 | ||
70 | /* | |
71 | * RB tree of active children throtl_grp's, which are sorted by | |
72 | * their ->disptime. | |
73 | */ | |
c9e0332e TH |
74 | struct rb_root pending_tree; /* RB tree of active tgs */ |
75 | struct rb_node *first_pending; /* first node in the tree */ | |
76 | unsigned int nr_pending; /* # queued in the tree */ | |
77 | unsigned long first_pending_disptime; /* disptime of the first tg */ | |
69df0ab0 | 78 | struct timer_list pending_timer; /* fires on first_pending_disptime */ |
e43473b7 VG |
79 | }; |
80 | ||
5b2c16aa TH |
81 | enum tg_state_flags { |
82 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
0e9f4164 | 83 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ |
5b2c16aa TH |
84 | }; |
85 | ||
e43473b7 VG |
86 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
87 | ||
9f626e37 | 88 | enum { |
cd5ab1b0 | 89 | LIMIT_LOW, |
9f626e37 SL |
90 | LIMIT_MAX, |
91 | LIMIT_CNT, | |
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 | ||
cd5ab1b0 | 130 | /* internally used bytes per second rate limits */ |
9f626e37 | 131 | uint64_t bps[2][LIMIT_CNT]; |
cd5ab1b0 SL |
132 | /* user configured bps limits */ |
133 | uint64_t bps_conf[2][LIMIT_CNT]; | |
e43473b7 | 134 | |
cd5ab1b0 | 135 | /* internally used IOPS limits */ |
9f626e37 | 136 | unsigned int iops[2][LIMIT_CNT]; |
cd5ab1b0 SL |
137 | /* user configured IOPS limits */ |
138 | unsigned int iops_conf[2][LIMIT_CNT]; | |
8e89d13f | 139 | |
e43473b7 VG |
140 | /* Number of bytes disptached in current slice */ |
141 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
142 | /* Number of bio's dispatched in current slice */ |
143 | unsigned int io_disp[2]; | |
e43473b7 | 144 | |
3f0abd80 SL |
145 | unsigned long last_low_overflow_time[2]; |
146 | ||
147 | uint64_t last_bytes_disp[2]; | |
148 | unsigned int last_io_disp[2]; | |
149 | ||
150 | unsigned long last_check_time; | |
151 | ||
aec24246 SL |
152 | unsigned long last_dispatch_time[2]; |
153 | ||
e43473b7 VG |
154 | /* When did we start a new slice */ |
155 | unsigned long slice_start[2]; | |
156 | unsigned long slice_end[2]; | |
157 | }; | |
158 | ||
159 | struct throtl_data | |
160 | { | |
e43473b7 | 161 | /* service tree for active throtl groups */ |
c9e0332e | 162 | struct throtl_service_queue service_queue; |
e43473b7 | 163 | |
e43473b7 VG |
164 | struct request_queue *queue; |
165 | ||
166 | /* Total Number of queued bios on READ and WRITE lists */ | |
167 | unsigned int nr_queued[2]; | |
168 | ||
297e3d85 SL |
169 | unsigned int throtl_slice; |
170 | ||
e43473b7 | 171 | /* Work for dispatching throttled bios */ |
69df0ab0 | 172 | struct work_struct dispatch_work; |
9f626e37 SL |
173 | unsigned int limit_index; |
174 | bool limit_valid[LIMIT_CNT]; | |
3f0abd80 SL |
175 | |
176 | unsigned long low_upgrade_time; | |
177 | unsigned long low_downgrade_time; | |
e43473b7 VG |
178 | }; |
179 | ||
69df0ab0 TH |
180 | static void throtl_pending_timer_fn(unsigned long arg); |
181 | ||
f95a04af TH |
182 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
183 | { | |
184 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
185 | } | |
186 | ||
3c798398 | 187 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 188 | { |
f95a04af | 189 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
190 | } |
191 | ||
3c798398 | 192 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 193 | { |
f95a04af | 194 | return pd_to_blkg(&tg->pd); |
0381411e TH |
195 | } |
196 | ||
fda6f272 TH |
197 | /** |
198 | * sq_to_tg - return the throl_grp the specified service queue belongs to | |
199 | * @sq: the throtl_service_queue of interest | |
200 | * | |
201 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one | |
202 | * embedded in throtl_data, %NULL is returned. | |
203 | */ | |
204 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) | |
205 | { | |
206 | if (sq && sq->parent_sq) | |
207 | return container_of(sq, struct throtl_grp, service_queue); | |
208 | else | |
209 | return NULL; | |
210 | } | |
211 | ||
212 | /** | |
213 | * sq_to_td - return throtl_data the specified service queue belongs to | |
214 | * @sq: the throtl_service_queue of interest | |
215 | * | |
b43daedc | 216 | * A service_queue can be embedded in either a throtl_grp or throtl_data. |
fda6f272 TH |
217 | * Determine the associated throtl_data accordingly and return it. |
218 | */ | |
219 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) | |
220 | { | |
221 | struct throtl_grp *tg = sq_to_tg(sq); | |
222 | ||
223 | if (tg) | |
224 | return tg->td; | |
225 | else | |
226 | return container_of(sq, struct throtl_data, service_queue); | |
227 | } | |
228 | ||
9f626e37 SL |
229 | static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw) |
230 | { | |
b22c417c SL |
231 | struct blkcg_gq *blkg = tg_to_blkg(tg); |
232 | uint64_t ret; | |
233 | ||
234 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
235 | return U64_MAX; | |
236 | ret = tg->bps[rw][tg->td->limit_index]; | |
237 | if (ret == 0 && tg->td->limit_index == LIMIT_LOW) | |
238 | return tg->bps[rw][LIMIT_MAX]; | |
239 | return ret; | |
9f626e37 SL |
240 | } |
241 | ||
242 | static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw) | |
243 | { | |
b22c417c SL |
244 | struct blkcg_gq *blkg = tg_to_blkg(tg); |
245 | unsigned int ret; | |
246 | ||
247 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) | |
248 | return UINT_MAX; | |
249 | ret = tg->iops[rw][tg->td->limit_index]; | |
250 | if (ret == 0 && tg->td->limit_index == LIMIT_LOW) | |
251 | return tg->iops[rw][LIMIT_MAX]; | |
252 | return ret; | |
9f626e37 SL |
253 | } |
254 | ||
fda6f272 TH |
255 | /** |
256 | * throtl_log - log debug message via blktrace | |
257 | * @sq: the service_queue being reported | |
258 | * @fmt: printf format string | |
259 | * @args: printf args | |
260 | * | |
261 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a | |
262 | * throtl_grp; otherwise, just "throtl". | |
fda6f272 TH |
263 | */ |
264 | #define throtl_log(sq, fmt, args...) do { \ | |
265 | struct throtl_grp *__tg = sq_to_tg((sq)); \ | |
266 | struct throtl_data *__td = sq_to_td((sq)); \ | |
267 | \ | |
268 | (void)__td; \ | |
59fa0224 SL |
269 | if (likely(!blk_trace_note_message_enabled(__td->queue))) \ |
270 | break; \ | |
fda6f272 TH |
271 | if ((__tg)) { \ |
272 | char __pbuf[128]; \ | |
54e7ed12 | 273 | \ |
fda6f272 TH |
274 | blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \ |
275 | blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
276 | } else { \ | |
277 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ | |
278 | } \ | |
54e7ed12 | 279 | } while (0) |
e43473b7 | 280 | |
c5cc2070 TH |
281 | static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg) |
282 | { | |
283 | INIT_LIST_HEAD(&qn->node); | |
284 | bio_list_init(&qn->bios); | |
285 | qn->tg = tg; | |
286 | } | |
287 | ||
288 | /** | |
289 | * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it | |
290 | * @bio: bio being added | |
291 | * @qn: qnode to add bio to | |
292 | * @queued: the service_queue->queued[] list @qn belongs to | |
293 | * | |
294 | * Add @bio to @qn and put @qn on @queued if it's not already on. | |
295 | * @qn->tg's reference count is bumped when @qn is activated. See the | |
296 | * comment on top of throtl_qnode definition for details. | |
297 | */ | |
298 | static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn, | |
299 | struct list_head *queued) | |
300 | { | |
301 | bio_list_add(&qn->bios, bio); | |
302 | if (list_empty(&qn->node)) { | |
303 | list_add_tail(&qn->node, queued); | |
304 | blkg_get(tg_to_blkg(qn->tg)); | |
305 | } | |
306 | } | |
307 | ||
308 | /** | |
309 | * throtl_peek_queued - peek the first bio on a qnode list | |
310 | * @queued: the qnode list to peek | |
311 | */ | |
312 | static struct bio *throtl_peek_queued(struct list_head *queued) | |
313 | { | |
314 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
315 | struct bio *bio; | |
316 | ||
317 | if (list_empty(queued)) | |
318 | return NULL; | |
319 | ||
320 | bio = bio_list_peek(&qn->bios); | |
321 | WARN_ON_ONCE(!bio); | |
322 | return bio; | |
323 | } | |
324 | ||
325 | /** | |
326 | * throtl_pop_queued - pop the first bio form a qnode list | |
327 | * @queued: the qnode list to pop a bio from | |
328 | * @tg_to_put: optional out argument for throtl_grp to put | |
329 | * | |
330 | * Pop the first bio from the qnode list @queued. After popping, the first | |
331 | * qnode is removed from @queued if empty or moved to the end of @queued so | |
332 | * that the popping order is round-robin. | |
333 | * | |
334 | * When the first qnode is removed, its associated throtl_grp should be put | |
335 | * too. If @tg_to_put is NULL, this function automatically puts it; | |
336 | * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is | |
337 | * responsible for putting it. | |
338 | */ | |
339 | static struct bio *throtl_pop_queued(struct list_head *queued, | |
340 | struct throtl_grp **tg_to_put) | |
341 | { | |
342 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); | |
343 | struct bio *bio; | |
344 | ||
345 | if (list_empty(queued)) | |
346 | return NULL; | |
347 | ||
348 | bio = bio_list_pop(&qn->bios); | |
349 | WARN_ON_ONCE(!bio); | |
350 | ||
351 | if (bio_list_empty(&qn->bios)) { | |
352 | list_del_init(&qn->node); | |
353 | if (tg_to_put) | |
354 | *tg_to_put = qn->tg; | |
355 | else | |
356 | blkg_put(tg_to_blkg(qn->tg)); | |
357 | } else { | |
358 | list_move_tail(&qn->node, queued); | |
359 | } | |
360 | ||
361 | return bio; | |
362 | } | |
363 | ||
49a2f1e3 | 364 | /* init a service_queue, assumes the caller zeroed it */ |
b2ce2643 | 365 | static void throtl_service_queue_init(struct throtl_service_queue *sq) |
49a2f1e3 | 366 | { |
c5cc2070 TH |
367 | INIT_LIST_HEAD(&sq->queued[0]); |
368 | INIT_LIST_HEAD(&sq->queued[1]); | |
49a2f1e3 | 369 | sq->pending_tree = RB_ROOT; |
69df0ab0 TH |
370 | setup_timer(&sq->pending_timer, throtl_pending_timer_fn, |
371 | (unsigned long)sq); | |
372 | } | |
373 | ||
001bea73 TH |
374 | static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node) |
375 | { | |
4fb72036 | 376 | struct throtl_grp *tg; |
24bdb8ef | 377 | int rw; |
4fb72036 TH |
378 | |
379 | tg = kzalloc_node(sizeof(*tg), gfp, node); | |
380 | if (!tg) | |
77ea7338 | 381 | return NULL; |
4fb72036 | 382 | |
b2ce2643 TH |
383 | throtl_service_queue_init(&tg->service_queue); |
384 | ||
385 | for (rw = READ; rw <= WRITE; rw++) { | |
386 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); | |
387 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); | |
388 | } | |
389 | ||
390 | RB_CLEAR_NODE(&tg->rb_node); | |
9f626e37 SL |
391 | tg->bps[READ][LIMIT_MAX] = U64_MAX; |
392 | tg->bps[WRITE][LIMIT_MAX] = U64_MAX; | |
393 | tg->iops[READ][LIMIT_MAX] = UINT_MAX; | |
394 | tg->iops[WRITE][LIMIT_MAX] = UINT_MAX; | |
cd5ab1b0 SL |
395 | tg->bps_conf[READ][LIMIT_MAX] = U64_MAX; |
396 | tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX; | |
397 | tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX; | |
398 | tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX; | |
399 | /* LIMIT_LOW will have default value 0 */ | |
b2ce2643 | 400 | |
4fb72036 | 401 | return &tg->pd; |
001bea73 TH |
402 | } |
403 | ||
a9520cd6 | 404 | static void throtl_pd_init(struct blkg_policy_data *pd) |
a29a171e | 405 | { |
a9520cd6 TH |
406 | struct throtl_grp *tg = pd_to_tg(pd); |
407 | struct blkcg_gq *blkg = tg_to_blkg(tg); | |
77216b04 | 408 | struct throtl_data *td = blkg->q->td; |
b2ce2643 | 409 | struct throtl_service_queue *sq = &tg->service_queue; |
cd1604fa | 410 | |
9138125b | 411 | /* |
aa6ec29b | 412 | * If on the default hierarchy, we switch to properly hierarchical |
9138125b TH |
413 | * behavior where limits on a given throtl_grp are applied to the |
414 | * whole subtree rather than just the group itself. e.g. If 16M | |
415 | * read_bps limit is set on the root group, the whole system can't | |
416 | * exceed 16M for the device. | |
417 | * | |
aa6ec29b | 418 | * If not on the default hierarchy, the broken flat hierarchy |
9138125b TH |
419 | * behavior is retained where all throtl_grps are treated as if |
420 | * they're all separate root groups right below throtl_data. | |
421 | * Limits of a group don't interact with limits of other groups | |
422 | * regardless of the position of the group in the hierarchy. | |
423 | */ | |
b2ce2643 | 424 | sq->parent_sq = &td->service_queue; |
9e10a130 | 425 | if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent) |
b2ce2643 | 426 | sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue; |
77216b04 | 427 | tg->td = td; |
8a3d2615 TH |
428 | } |
429 | ||
693e751e TH |
430 | /* |
431 | * Set has_rules[] if @tg or any of its parents have limits configured. | |
432 | * This doesn't require walking up to the top of the hierarchy as the | |
433 | * parent's has_rules[] is guaranteed to be correct. | |
434 | */ | |
435 | static void tg_update_has_rules(struct throtl_grp *tg) | |
436 | { | |
437 | struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); | |
9f626e37 | 438 | struct throtl_data *td = tg->td; |
693e751e TH |
439 | int rw; |
440 | ||
441 | for (rw = READ; rw <= WRITE; rw++) | |
442 | tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) || | |
9f626e37 SL |
443 | (td->limit_valid[td->limit_index] && |
444 | (tg_bps_limit(tg, rw) != U64_MAX || | |
445 | tg_iops_limit(tg, rw) != UINT_MAX)); | |
693e751e TH |
446 | } |
447 | ||
a9520cd6 | 448 | static void throtl_pd_online(struct blkg_policy_data *pd) |
693e751e | 449 | { |
aec24246 | 450 | struct throtl_grp *tg = pd_to_tg(pd); |
693e751e TH |
451 | /* |
452 | * We don't want new groups to escape the limits of its ancestors. | |
453 | * Update has_rules[] after a new group is brought online. | |
454 | */ | |
aec24246 SL |
455 | tg_update_has_rules(tg); |
456 | tg->last_dispatch_time[READ] = jiffies; | |
457 | tg->last_dispatch_time[WRITE] = jiffies; | |
693e751e TH |
458 | } |
459 | ||
cd5ab1b0 SL |
460 | static void blk_throtl_update_limit_valid(struct throtl_data *td) |
461 | { | |
462 | struct cgroup_subsys_state *pos_css; | |
463 | struct blkcg_gq *blkg; | |
464 | bool low_valid = false; | |
465 | ||
466 | rcu_read_lock(); | |
467 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
468 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
469 | ||
470 | if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] || | |
471 | tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) | |
472 | low_valid = true; | |
473 | } | |
474 | rcu_read_unlock(); | |
475 | ||
476 | td->limit_valid[LIMIT_LOW] = low_valid; | |
477 | } | |
478 | ||
c79892c5 | 479 | static void throtl_upgrade_state(struct throtl_data *td); |
cd5ab1b0 SL |
480 | static void throtl_pd_offline(struct blkg_policy_data *pd) |
481 | { | |
482 | struct throtl_grp *tg = pd_to_tg(pd); | |
483 | ||
484 | tg->bps[READ][LIMIT_LOW] = 0; | |
485 | tg->bps[WRITE][LIMIT_LOW] = 0; | |
486 | tg->iops[READ][LIMIT_LOW] = 0; | |
487 | tg->iops[WRITE][LIMIT_LOW] = 0; | |
488 | ||
489 | blk_throtl_update_limit_valid(tg->td); | |
490 | ||
c79892c5 SL |
491 | if (!tg->td->limit_valid[tg->td->limit_index]) |
492 | throtl_upgrade_state(tg->td); | |
cd5ab1b0 SL |
493 | } |
494 | ||
001bea73 TH |
495 | static void throtl_pd_free(struct blkg_policy_data *pd) |
496 | { | |
4fb72036 TH |
497 | struct throtl_grp *tg = pd_to_tg(pd); |
498 | ||
b2ce2643 | 499 | del_timer_sync(&tg->service_queue.pending_timer); |
4fb72036 | 500 | kfree(tg); |
001bea73 TH |
501 | } |
502 | ||
0049af73 TH |
503 | static struct throtl_grp * |
504 | throtl_rb_first(struct throtl_service_queue *parent_sq) | |
e43473b7 VG |
505 | { |
506 | /* Service tree is empty */ | |
0049af73 | 507 | if (!parent_sq->nr_pending) |
e43473b7 VG |
508 | return NULL; |
509 | ||
0049af73 TH |
510 | if (!parent_sq->first_pending) |
511 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); | |
e43473b7 | 512 | |
0049af73 TH |
513 | if (parent_sq->first_pending) |
514 | return rb_entry_tg(parent_sq->first_pending); | |
e43473b7 VG |
515 | |
516 | return NULL; | |
517 | } | |
518 | ||
519 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
520 | { | |
521 | rb_erase(n, root); | |
522 | RB_CLEAR_NODE(n); | |
523 | } | |
524 | ||
0049af73 TH |
525 | static void throtl_rb_erase(struct rb_node *n, |
526 | struct throtl_service_queue *parent_sq) | |
e43473b7 | 527 | { |
0049af73 TH |
528 | if (parent_sq->first_pending == n) |
529 | parent_sq->first_pending = NULL; | |
530 | rb_erase_init(n, &parent_sq->pending_tree); | |
531 | --parent_sq->nr_pending; | |
e43473b7 VG |
532 | } |
533 | ||
0049af73 | 534 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
535 | { |
536 | struct throtl_grp *tg; | |
537 | ||
0049af73 | 538 | tg = throtl_rb_first(parent_sq); |
e43473b7 VG |
539 | if (!tg) |
540 | return; | |
541 | ||
0049af73 | 542 | parent_sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
543 | } |
544 | ||
77216b04 | 545 | static void tg_service_queue_add(struct throtl_grp *tg) |
e43473b7 | 546 | { |
77216b04 | 547 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
0049af73 | 548 | struct rb_node **node = &parent_sq->pending_tree.rb_node; |
e43473b7 VG |
549 | struct rb_node *parent = NULL; |
550 | struct throtl_grp *__tg; | |
551 | unsigned long key = tg->disptime; | |
552 | int left = 1; | |
553 | ||
554 | while (*node != NULL) { | |
555 | parent = *node; | |
556 | __tg = rb_entry_tg(parent); | |
557 | ||
558 | if (time_before(key, __tg->disptime)) | |
559 | node = &parent->rb_left; | |
560 | else { | |
561 | node = &parent->rb_right; | |
562 | left = 0; | |
563 | } | |
564 | } | |
565 | ||
566 | if (left) | |
0049af73 | 567 | parent_sq->first_pending = &tg->rb_node; |
e43473b7 VG |
568 | |
569 | rb_link_node(&tg->rb_node, parent, node); | |
0049af73 | 570 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); |
e43473b7 VG |
571 | } |
572 | ||
77216b04 | 573 | static void __throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 574 | { |
77216b04 | 575 | tg_service_queue_add(tg); |
5b2c16aa | 576 | tg->flags |= THROTL_TG_PENDING; |
77216b04 | 577 | tg->service_queue.parent_sq->nr_pending++; |
e43473b7 VG |
578 | } |
579 | ||
77216b04 | 580 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
e43473b7 | 581 | { |
5b2c16aa | 582 | if (!(tg->flags & THROTL_TG_PENDING)) |
77216b04 | 583 | __throtl_enqueue_tg(tg); |
e43473b7 VG |
584 | } |
585 | ||
77216b04 | 586 | static void __throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 587 | { |
77216b04 | 588 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
5b2c16aa | 589 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
590 | } |
591 | ||
77216b04 | 592 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
e43473b7 | 593 | { |
5b2c16aa | 594 | if (tg->flags & THROTL_TG_PENDING) |
77216b04 | 595 | __throtl_dequeue_tg(tg); |
e43473b7 VG |
596 | } |
597 | ||
a9131a27 | 598 | /* Call with queue lock held */ |
69df0ab0 TH |
599 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, |
600 | unsigned long expires) | |
a9131a27 | 601 | { |
297e3d85 | 602 | unsigned long max_expire = jiffies + 8 * sq_to_tg(sq)->td->throtl_slice; |
06cceedc SL |
603 | |
604 | /* | |
605 | * Since we are adjusting the throttle limit dynamically, the sleep | |
606 | * time calculated according to previous limit might be invalid. It's | |
607 | * possible the cgroup sleep time is very long and no other cgroups | |
608 | * have IO running so notify the limit changes. Make sure the cgroup | |
609 | * doesn't sleep too long to avoid the missed notification. | |
610 | */ | |
611 | if (time_after(expires, max_expire)) | |
612 | expires = max_expire; | |
69df0ab0 TH |
613 | mod_timer(&sq->pending_timer, expires); |
614 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", | |
615 | expires - jiffies, jiffies); | |
a9131a27 TH |
616 | } |
617 | ||
7f52f98c TH |
618 | /** |
619 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle | |
620 | * @sq: the service_queue to schedule dispatch for | |
621 | * @force: force scheduling | |
622 | * | |
623 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the | |
624 | * dispatch time of the first pending child. Returns %true if either timer | |
625 | * is armed or there's no pending child left. %false if the current | |
626 | * dispatch window is still open and the caller should continue | |
627 | * dispatching. | |
628 | * | |
629 | * If @force is %true, the dispatch timer is always scheduled and this | |
630 | * function is guaranteed to return %true. This is to be used when the | |
631 | * caller can't dispatch itself and needs to invoke pending_timer | |
632 | * unconditionally. Note that forced scheduling is likely to induce short | |
633 | * delay before dispatch starts even if @sq->first_pending_disptime is not | |
634 | * in the future and thus shouldn't be used in hot paths. | |
635 | */ | |
636 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, | |
637 | bool force) | |
e43473b7 | 638 | { |
6a525600 | 639 | /* any pending children left? */ |
c9e0332e | 640 | if (!sq->nr_pending) |
7f52f98c | 641 | return true; |
e43473b7 | 642 | |
c9e0332e | 643 | update_min_dispatch_time(sq); |
e43473b7 | 644 | |
69df0ab0 | 645 | /* is the next dispatch time in the future? */ |
7f52f98c | 646 | if (force || time_after(sq->first_pending_disptime, jiffies)) { |
69df0ab0 | 647 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); |
7f52f98c | 648 | return true; |
69df0ab0 TH |
649 | } |
650 | ||
7f52f98c TH |
651 | /* tell the caller to continue dispatching */ |
652 | return false; | |
e43473b7 VG |
653 | } |
654 | ||
32ee5bc4 VG |
655 | static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, |
656 | bool rw, unsigned long start) | |
657 | { | |
658 | tg->bytes_disp[rw] = 0; | |
659 | tg->io_disp[rw] = 0; | |
660 | ||
661 | /* | |
662 | * Previous slice has expired. We must have trimmed it after last | |
663 | * bio dispatch. That means since start of last slice, we never used | |
664 | * that bandwidth. Do try to make use of that bandwidth while giving | |
665 | * credit. | |
666 | */ | |
667 | if (time_after_eq(start, tg->slice_start[rw])) | |
668 | tg->slice_start[rw] = start; | |
669 | ||
297e3d85 | 670 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; |
32ee5bc4 VG |
671 | throtl_log(&tg->service_queue, |
672 | "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", | |
673 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
674 | tg->slice_end[rw], jiffies); | |
675 | } | |
676 | ||
0f3457f6 | 677 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
678 | { |
679 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 680 | tg->io_disp[rw] = 0; |
e43473b7 | 681 | tg->slice_start[rw] = jiffies; |
297e3d85 | 682 | tg->slice_end[rw] = jiffies + tg->td->throtl_slice; |
fda6f272 TH |
683 | throtl_log(&tg->service_queue, |
684 | "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
685 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
686 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
687 | } |
688 | ||
0f3457f6 TH |
689 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
690 | unsigned long jiffy_end) | |
d1ae8ffd | 691 | { |
297e3d85 | 692 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); |
d1ae8ffd VG |
693 | } |
694 | ||
0f3457f6 TH |
695 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
696 | unsigned long jiffy_end) | |
e43473b7 | 697 | { |
297e3d85 | 698 | tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); |
fda6f272 TH |
699 | throtl_log(&tg->service_queue, |
700 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
701 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
702 | tg->slice_end[rw], jiffies); | |
e43473b7 VG |
703 | } |
704 | ||
705 | /* Determine if previously allocated or extended slice is complete or not */ | |
0f3457f6 | 706 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
e43473b7 VG |
707 | { |
708 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
5cf8c227 | 709 | return false; |
e43473b7 VG |
710 | |
711 | return 1; | |
712 | } | |
713 | ||
714 | /* Trim the used slices and adjust slice start accordingly */ | |
0f3457f6 | 715 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
e43473b7 | 716 | { |
3aad5d3e VG |
717 | unsigned long nr_slices, time_elapsed, io_trim; |
718 | u64 bytes_trim, tmp; | |
e43473b7 VG |
719 | |
720 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
721 | ||
722 | /* | |
723 | * If bps are unlimited (-1), then time slice don't get | |
724 | * renewed. Don't try to trim the slice if slice is used. A new | |
725 | * slice will start when appropriate. | |
726 | */ | |
0f3457f6 | 727 | if (throtl_slice_used(tg, rw)) |
e43473b7 VG |
728 | return; |
729 | ||
d1ae8ffd VG |
730 | /* |
731 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
732 | * that initially cgroup limit was very low resulting in high | |
733 | * slice_end, but later limit was bumped up and bio was dispached | |
734 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
735 | * is bad because it does not allow new slice to start. | |
736 | */ | |
737 | ||
297e3d85 | 738 | throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice); |
d1ae8ffd | 739 | |
e43473b7 VG |
740 | time_elapsed = jiffies - tg->slice_start[rw]; |
741 | ||
297e3d85 | 742 | nr_slices = time_elapsed / tg->td->throtl_slice; |
e43473b7 VG |
743 | |
744 | if (!nr_slices) | |
745 | return; | |
297e3d85 | 746 | tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices; |
3aad5d3e VG |
747 | do_div(tmp, HZ); |
748 | bytes_trim = tmp; | |
e43473b7 | 749 | |
297e3d85 SL |
750 | io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) / |
751 | HZ; | |
e43473b7 | 752 | |
8e89d13f | 753 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
754 | return; |
755 | ||
756 | if (tg->bytes_disp[rw] >= bytes_trim) | |
757 | tg->bytes_disp[rw] -= bytes_trim; | |
758 | else | |
759 | tg->bytes_disp[rw] = 0; | |
760 | ||
8e89d13f VG |
761 | if (tg->io_disp[rw] >= io_trim) |
762 | tg->io_disp[rw] -= io_trim; | |
763 | else | |
764 | tg->io_disp[rw] = 0; | |
765 | ||
297e3d85 | 766 | tg->slice_start[rw] += nr_slices * tg->td->throtl_slice; |
e43473b7 | 767 | |
fda6f272 TH |
768 | throtl_log(&tg->service_queue, |
769 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", | |
770 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, | |
771 | tg->slice_start[rw], tg->slice_end[rw], jiffies); | |
e43473b7 VG |
772 | } |
773 | ||
0f3457f6 TH |
774 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
775 | unsigned long *wait) | |
e43473b7 VG |
776 | { |
777 | bool rw = bio_data_dir(bio); | |
8e89d13f | 778 | unsigned int io_allowed; |
e43473b7 | 779 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 780 | u64 tmp; |
e43473b7 | 781 | |
8e89d13f | 782 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 783 | |
8e89d13f VG |
784 | /* Slice has just started. Consider one slice interval */ |
785 | if (!jiffy_elapsed) | |
297e3d85 | 786 | jiffy_elapsed_rnd = tg->td->throtl_slice; |
8e89d13f | 787 | |
297e3d85 | 788 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); |
8e89d13f | 789 | |
c49c06e4 VG |
790 | /* |
791 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
792 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
793 | * will allow dispatch after 1 second and after that slice should | |
794 | * have been trimmed. | |
795 | */ | |
796 | ||
9f626e37 | 797 | tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd; |
c49c06e4 VG |
798 | do_div(tmp, HZ); |
799 | ||
800 | if (tmp > UINT_MAX) | |
801 | io_allowed = UINT_MAX; | |
802 | else | |
803 | io_allowed = tmp; | |
8e89d13f VG |
804 | |
805 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
806 | if (wait) |
807 | *wait = 0; | |
5cf8c227 | 808 | return true; |
e43473b7 VG |
809 | } |
810 | ||
8e89d13f | 811 | /* Calc approx time to dispatch */ |
9f626e37 | 812 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ) / tg_iops_limit(tg, rw) + 1; |
8e89d13f VG |
813 | |
814 | if (jiffy_wait > jiffy_elapsed) | |
815 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
816 | else | |
817 | jiffy_wait = 1; | |
818 | ||
819 | if (wait) | |
820 | *wait = jiffy_wait; | |
821 | return 0; | |
822 | } | |
823 | ||
0f3457f6 TH |
824 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
825 | unsigned long *wait) | |
8e89d13f VG |
826 | { |
827 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 828 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 829 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
830 | |
831 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
832 | ||
833 | /* Slice has just started. Consider one slice interval */ | |
834 | if (!jiffy_elapsed) | |
297e3d85 | 835 | jiffy_elapsed_rnd = tg->td->throtl_slice; |
e43473b7 | 836 | |
297e3d85 | 837 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); |
e43473b7 | 838 | |
9f626e37 | 839 | tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd; |
5e901a2b | 840 | do_div(tmp, HZ); |
3aad5d3e | 841 | bytes_allowed = tmp; |
e43473b7 | 842 | |
4f024f37 | 843 | if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) { |
e43473b7 VG |
844 | if (wait) |
845 | *wait = 0; | |
5cf8c227 | 846 | return true; |
e43473b7 VG |
847 | } |
848 | ||
849 | /* Calc approx time to dispatch */ | |
4f024f37 | 850 | extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed; |
9f626e37 | 851 | jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw)); |
e43473b7 VG |
852 | |
853 | if (!jiffy_wait) | |
854 | jiffy_wait = 1; | |
855 | ||
856 | /* | |
857 | * This wait time is without taking into consideration the rounding | |
858 | * up we did. Add that time also. | |
859 | */ | |
860 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
861 | if (wait) |
862 | *wait = jiffy_wait; | |
8e89d13f VG |
863 | return 0; |
864 | } | |
865 | ||
866 | /* | |
867 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
868 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
869 | */ | |
0f3457f6 TH |
870 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
871 | unsigned long *wait) | |
8e89d13f VG |
872 | { |
873 | bool rw = bio_data_dir(bio); | |
874 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
875 | ||
876 | /* | |
877 | * Currently whole state machine of group depends on first bio | |
878 | * queued in the group bio list. So one should not be calling | |
879 | * this function with a different bio if there are other bios | |
880 | * queued. | |
881 | */ | |
73f0d49a | 882 | BUG_ON(tg->service_queue.nr_queued[rw] && |
c5cc2070 | 883 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); |
e43473b7 | 884 | |
8e89d13f | 885 | /* If tg->bps = -1, then BW is unlimited */ |
9f626e37 SL |
886 | if (tg_bps_limit(tg, rw) == U64_MAX && |
887 | tg_iops_limit(tg, rw) == UINT_MAX) { | |
8e89d13f VG |
888 | if (wait) |
889 | *wait = 0; | |
5cf8c227 | 890 | return true; |
8e89d13f VG |
891 | } |
892 | ||
893 | /* | |
894 | * If previous slice expired, start a new one otherwise renew/extend | |
895 | * existing slice to make sure it is at least throtl_slice interval | |
164c80ed VG |
896 | * long since now. New slice is started only for empty throttle group. |
897 | * If there is queued bio, that means there should be an active | |
898 | * slice and it should be extended instead. | |
8e89d13f | 899 | */ |
164c80ed | 900 | if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw])) |
0f3457f6 | 901 | throtl_start_new_slice(tg, rw); |
8e89d13f | 902 | else { |
297e3d85 SL |
903 | if (time_before(tg->slice_end[rw], |
904 | jiffies + tg->td->throtl_slice)) | |
905 | throtl_extend_slice(tg, rw, | |
906 | jiffies + tg->td->throtl_slice); | |
8e89d13f VG |
907 | } |
908 | ||
0f3457f6 TH |
909 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
910 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { | |
8e89d13f VG |
911 | if (wait) |
912 | *wait = 0; | |
913 | return 1; | |
914 | } | |
915 | ||
916 | max_wait = max(bps_wait, iops_wait); | |
917 | ||
918 | if (wait) | |
919 | *wait = max_wait; | |
920 | ||
921 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
0f3457f6 | 922 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
e43473b7 VG |
923 | |
924 | return 0; | |
925 | } | |
926 | ||
927 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) | |
928 | { | |
929 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
930 | |
931 | /* Charge the bio to the group */ | |
4f024f37 | 932 | tg->bytes_disp[rw] += bio->bi_iter.bi_size; |
8e89d13f | 933 | tg->io_disp[rw]++; |
3f0abd80 SL |
934 | tg->last_bytes_disp[rw] += bio->bi_iter.bi_size; |
935 | tg->last_io_disp[rw]++; | |
e43473b7 | 936 | |
2a0f61e6 | 937 | /* |
8d2bbd4c | 938 | * BIO_THROTTLED is used to prevent the same bio to be throttled |
2a0f61e6 TH |
939 | * more than once as a throttled bio will go through blk-throtl the |
940 | * second time when it eventually gets issued. Set it when a bio | |
941 | * is being charged to a tg. | |
2a0f61e6 | 942 | */ |
8d2bbd4c CH |
943 | if (!bio_flagged(bio, BIO_THROTTLED)) |
944 | bio_set_flag(bio, BIO_THROTTLED); | |
e43473b7 VG |
945 | } |
946 | ||
c5cc2070 TH |
947 | /** |
948 | * throtl_add_bio_tg - add a bio to the specified throtl_grp | |
949 | * @bio: bio to add | |
950 | * @qn: qnode to use | |
951 | * @tg: the target throtl_grp | |
952 | * | |
953 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, | |
954 | * tg->qnode_on_self[] is used. | |
955 | */ | |
956 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, | |
957 | struct throtl_grp *tg) | |
e43473b7 | 958 | { |
73f0d49a | 959 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
960 | bool rw = bio_data_dir(bio); |
961 | ||
c5cc2070 TH |
962 | if (!qn) |
963 | qn = &tg->qnode_on_self[rw]; | |
964 | ||
0e9f4164 TH |
965 | /* |
966 | * If @tg doesn't currently have any bios queued in the same | |
967 | * direction, queueing @bio can change when @tg should be | |
968 | * dispatched. Mark that @tg was empty. This is automatically | |
969 | * cleaered on the next tg_update_disptime(). | |
970 | */ | |
971 | if (!sq->nr_queued[rw]) | |
972 | tg->flags |= THROTL_TG_WAS_EMPTY; | |
973 | ||
c5cc2070 TH |
974 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); |
975 | ||
73f0d49a | 976 | sq->nr_queued[rw]++; |
77216b04 | 977 | throtl_enqueue_tg(tg); |
e43473b7 VG |
978 | } |
979 | ||
77216b04 | 980 | static void tg_update_disptime(struct throtl_grp *tg) |
e43473b7 | 981 | { |
73f0d49a | 982 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
983 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
984 | struct bio *bio; | |
985 | ||
d609af3a ME |
986 | bio = throtl_peek_queued(&sq->queued[READ]); |
987 | if (bio) | |
0f3457f6 | 988 | tg_may_dispatch(tg, bio, &read_wait); |
e43473b7 | 989 | |
d609af3a ME |
990 | bio = throtl_peek_queued(&sq->queued[WRITE]); |
991 | if (bio) | |
0f3457f6 | 992 | tg_may_dispatch(tg, bio, &write_wait); |
e43473b7 VG |
993 | |
994 | min_wait = min(read_wait, write_wait); | |
995 | disptime = jiffies + min_wait; | |
996 | ||
e43473b7 | 997 | /* Update dispatch time */ |
77216b04 | 998 | throtl_dequeue_tg(tg); |
e43473b7 | 999 | tg->disptime = disptime; |
77216b04 | 1000 | throtl_enqueue_tg(tg); |
0e9f4164 TH |
1001 | |
1002 | /* see throtl_add_bio_tg() */ | |
1003 | tg->flags &= ~THROTL_TG_WAS_EMPTY; | |
e43473b7 VG |
1004 | } |
1005 | ||
32ee5bc4 VG |
1006 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, |
1007 | struct throtl_grp *parent_tg, bool rw) | |
1008 | { | |
1009 | if (throtl_slice_used(parent_tg, rw)) { | |
1010 | throtl_start_new_slice_with_credit(parent_tg, rw, | |
1011 | child_tg->slice_start[rw]); | |
1012 | } | |
1013 | ||
1014 | } | |
1015 | ||
77216b04 | 1016 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
e43473b7 | 1017 | { |
73f0d49a | 1018 | struct throtl_service_queue *sq = &tg->service_queue; |
6bc9c2b4 TH |
1019 | struct throtl_service_queue *parent_sq = sq->parent_sq; |
1020 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); | |
c5cc2070 | 1021 | struct throtl_grp *tg_to_put = NULL; |
e43473b7 VG |
1022 | struct bio *bio; |
1023 | ||
c5cc2070 TH |
1024 | /* |
1025 | * @bio is being transferred from @tg to @parent_sq. Popping a bio | |
1026 | * from @tg may put its reference and @parent_sq might end up | |
1027 | * getting released prematurely. Remember the tg to put and put it | |
1028 | * after @bio is transferred to @parent_sq. | |
1029 | */ | |
1030 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); | |
73f0d49a | 1031 | sq->nr_queued[rw]--; |
e43473b7 VG |
1032 | |
1033 | throtl_charge_bio(tg, bio); | |
6bc9c2b4 TH |
1034 | |
1035 | /* | |
1036 | * If our parent is another tg, we just need to transfer @bio to | |
1037 | * the parent using throtl_add_bio_tg(). If our parent is | |
1038 | * @td->service_queue, @bio is ready to be issued. Put it on its | |
1039 | * bio_lists[] and decrease total number queued. The caller is | |
1040 | * responsible for issuing these bios. | |
1041 | */ | |
1042 | if (parent_tg) { | |
c5cc2070 | 1043 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); |
32ee5bc4 | 1044 | start_parent_slice_with_credit(tg, parent_tg, rw); |
6bc9c2b4 | 1045 | } else { |
c5cc2070 TH |
1046 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], |
1047 | &parent_sq->queued[rw]); | |
6bc9c2b4 TH |
1048 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
1049 | tg->td->nr_queued[rw]--; | |
1050 | } | |
e43473b7 | 1051 | |
0f3457f6 | 1052 | throtl_trim_slice(tg, rw); |
6bc9c2b4 | 1053 | |
c5cc2070 TH |
1054 | if (tg_to_put) |
1055 | blkg_put(tg_to_blkg(tg_to_put)); | |
e43473b7 VG |
1056 | } |
1057 | ||
77216b04 | 1058 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
e43473b7 | 1059 | { |
73f0d49a | 1060 | struct throtl_service_queue *sq = &tg->service_queue; |
e43473b7 VG |
1061 | unsigned int nr_reads = 0, nr_writes = 0; |
1062 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 1063 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
1064 | struct bio *bio; |
1065 | ||
1066 | /* Try to dispatch 75% READS and 25% WRITES */ | |
1067 | ||
c5cc2070 | 1068 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && |
0f3457f6 | 1069 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1070 | |
77216b04 | 1071 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1072 | nr_reads++; |
1073 | ||
1074 | if (nr_reads >= max_nr_reads) | |
1075 | break; | |
1076 | } | |
1077 | ||
c5cc2070 | 1078 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && |
0f3457f6 | 1079 | tg_may_dispatch(tg, bio, NULL)) { |
e43473b7 | 1080 | |
77216b04 | 1081 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
e43473b7 VG |
1082 | nr_writes++; |
1083 | ||
1084 | if (nr_writes >= max_nr_writes) | |
1085 | break; | |
1086 | } | |
1087 | ||
1088 | return nr_reads + nr_writes; | |
1089 | } | |
1090 | ||
651930bc | 1091 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
e43473b7 VG |
1092 | { |
1093 | unsigned int nr_disp = 0; | |
e43473b7 VG |
1094 | |
1095 | while (1) { | |
73f0d49a TH |
1096 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
1097 | struct throtl_service_queue *sq = &tg->service_queue; | |
e43473b7 VG |
1098 | |
1099 | if (!tg) | |
1100 | break; | |
1101 | ||
1102 | if (time_before(jiffies, tg->disptime)) | |
1103 | break; | |
1104 | ||
77216b04 | 1105 | throtl_dequeue_tg(tg); |
e43473b7 | 1106 | |
77216b04 | 1107 | nr_disp += throtl_dispatch_tg(tg); |
e43473b7 | 1108 | |
73f0d49a | 1109 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
77216b04 | 1110 | tg_update_disptime(tg); |
e43473b7 VG |
1111 | |
1112 | if (nr_disp >= throtl_quantum) | |
1113 | break; | |
1114 | } | |
1115 | ||
1116 | return nr_disp; | |
1117 | } | |
1118 | ||
c79892c5 SL |
1119 | static bool throtl_can_upgrade(struct throtl_data *td, |
1120 | struct throtl_grp *this_tg); | |
6e1a5704 TH |
1121 | /** |
1122 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer | |
1123 | * @arg: the throtl_service_queue being serviced | |
1124 | * | |
1125 | * This timer is armed when a child throtl_grp with active bio's become | |
1126 | * pending and queued on the service_queue's pending_tree and expires when | |
1127 | * the first child throtl_grp should be dispatched. This function | |
2e48a530 TH |
1128 | * dispatches bio's from the children throtl_grps to the parent |
1129 | * service_queue. | |
1130 | * | |
1131 | * If the parent's parent is another throtl_grp, dispatching is propagated | |
1132 | * by either arming its pending_timer or repeating dispatch directly. If | |
1133 | * the top-level service_tree is reached, throtl_data->dispatch_work is | |
1134 | * kicked so that the ready bio's are issued. | |
6e1a5704 | 1135 | */ |
69df0ab0 TH |
1136 | static void throtl_pending_timer_fn(unsigned long arg) |
1137 | { | |
1138 | struct throtl_service_queue *sq = (void *)arg; | |
2e48a530 | 1139 | struct throtl_grp *tg = sq_to_tg(sq); |
69df0ab0 | 1140 | struct throtl_data *td = sq_to_td(sq); |
cb76199c | 1141 | struct request_queue *q = td->queue; |
2e48a530 TH |
1142 | struct throtl_service_queue *parent_sq; |
1143 | bool dispatched; | |
6e1a5704 | 1144 | int ret; |
e43473b7 VG |
1145 | |
1146 | spin_lock_irq(q->queue_lock); | |
c79892c5 SL |
1147 | if (throtl_can_upgrade(td, NULL)) |
1148 | throtl_upgrade_state(td); | |
1149 | ||
2e48a530 TH |
1150 | again: |
1151 | parent_sq = sq->parent_sq; | |
1152 | dispatched = false; | |
e43473b7 | 1153 | |
7f52f98c TH |
1154 | while (true) { |
1155 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", | |
2e48a530 TH |
1156 | sq->nr_queued[READ] + sq->nr_queued[WRITE], |
1157 | sq->nr_queued[READ], sq->nr_queued[WRITE]); | |
7f52f98c TH |
1158 | |
1159 | ret = throtl_select_dispatch(sq); | |
1160 | if (ret) { | |
7f52f98c TH |
1161 | throtl_log(sq, "bios disp=%u", ret); |
1162 | dispatched = true; | |
1163 | } | |
e43473b7 | 1164 | |
7f52f98c TH |
1165 | if (throtl_schedule_next_dispatch(sq, false)) |
1166 | break; | |
e43473b7 | 1167 | |
7f52f98c TH |
1168 | /* this dispatch windows is still open, relax and repeat */ |
1169 | spin_unlock_irq(q->queue_lock); | |
1170 | cpu_relax(); | |
1171 | spin_lock_irq(q->queue_lock); | |
651930bc | 1172 | } |
e43473b7 | 1173 | |
2e48a530 TH |
1174 | if (!dispatched) |
1175 | goto out_unlock; | |
6e1a5704 | 1176 | |
2e48a530 TH |
1177 | if (parent_sq) { |
1178 | /* @parent_sq is another throl_grp, propagate dispatch */ | |
1179 | if (tg->flags & THROTL_TG_WAS_EMPTY) { | |
1180 | tg_update_disptime(tg); | |
1181 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { | |
1182 | /* window is already open, repeat dispatching */ | |
1183 | sq = parent_sq; | |
1184 | tg = sq_to_tg(sq); | |
1185 | goto again; | |
1186 | } | |
1187 | } | |
1188 | } else { | |
1189 | /* reached the top-level, queue issueing */ | |
1190 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1191 | } | |
1192 | out_unlock: | |
e43473b7 | 1193 | spin_unlock_irq(q->queue_lock); |
6e1a5704 | 1194 | } |
e43473b7 | 1195 | |
6e1a5704 TH |
1196 | /** |
1197 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work | |
1198 | * @work: work item being executed | |
1199 | * | |
1200 | * This function is queued for execution when bio's reach the bio_lists[] | |
1201 | * of throtl_data->service_queue. Those bio's are ready and issued by this | |
1202 | * function. | |
1203 | */ | |
8876e140 | 1204 | static void blk_throtl_dispatch_work_fn(struct work_struct *work) |
6e1a5704 TH |
1205 | { |
1206 | struct throtl_data *td = container_of(work, struct throtl_data, | |
1207 | dispatch_work); | |
1208 | struct throtl_service_queue *td_sq = &td->service_queue; | |
1209 | struct request_queue *q = td->queue; | |
1210 | struct bio_list bio_list_on_stack; | |
1211 | struct bio *bio; | |
1212 | struct blk_plug plug; | |
1213 | int rw; | |
1214 | ||
1215 | bio_list_init(&bio_list_on_stack); | |
1216 | ||
1217 | spin_lock_irq(q->queue_lock); | |
c5cc2070 TH |
1218 | for (rw = READ; rw <= WRITE; rw++) |
1219 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) | |
1220 | bio_list_add(&bio_list_on_stack, bio); | |
6e1a5704 TH |
1221 | spin_unlock_irq(q->queue_lock); |
1222 | ||
1223 | if (!bio_list_empty(&bio_list_on_stack)) { | |
69d60eb9 | 1224 | blk_start_plug(&plug); |
e43473b7 VG |
1225 | while((bio = bio_list_pop(&bio_list_on_stack))) |
1226 | generic_make_request(bio); | |
69d60eb9 | 1227 | blk_finish_plug(&plug); |
e43473b7 | 1228 | } |
e43473b7 VG |
1229 | } |
1230 | ||
f95a04af TH |
1231 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
1232 | int off) | |
60c2bc2d | 1233 | { |
f95a04af TH |
1234 | struct throtl_grp *tg = pd_to_tg(pd); |
1235 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 1236 | |
2ab5492d | 1237 | if (v == U64_MAX) |
60c2bc2d | 1238 | return 0; |
f95a04af | 1239 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
1240 | } |
1241 | ||
f95a04af TH |
1242 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
1243 | int off) | |
e43473b7 | 1244 | { |
f95a04af TH |
1245 | struct throtl_grp *tg = pd_to_tg(pd); |
1246 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 1247 | |
2ab5492d | 1248 | if (v == UINT_MAX) |
af133ceb | 1249 | return 0; |
f95a04af | 1250 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
1251 | } |
1252 | ||
2da8ca82 | 1253 | static int tg_print_conf_u64(struct seq_file *sf, void *v) |
8e89d13f | 1254 | { |
2da8ca82 TH |
1255 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64, |
1256 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1257 | return 0; |
8e89d13f VG |
1258 | } |
1259 | ||
2da8ca82 | 1260 | static int tg_print_conf_uint(struct seq_file *sf, void *v) |
8e89d13f | 1261 | { |
2da8ca82 TH |
1262 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint, |
1263 | &blkcg_policy_throtl, seq_cft(sf)->private, false); | |
af133ceb | 1264 | return 0; |
60c2bc2d TH |
1265 | } |
1266 | ||
69948b07 | 1267 | static void tg_conf_updated(struct throtl_grp *tg) |
60c2bc2d | 1268 | { |
69948b07 | 1269 | struct throtl_service_queue *sq = &tg->service_queue; |
492eb21b | 1270 | struct cgroup_subsys_state *pos_css; |
69948b07 | 1271 | struct blkcg_gq *blkg; |
af133ceb | 1272 | |
fda6f272 TH |
1273 | throtl_log(&tg->service_queue, |
1274 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", | |
9f626e37 SL |
1275 | tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE), |
1276 | tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE)); | |
632b4493 | 1277 | |
693e751e TH |
1278 | /* |
1279 | * Update has_rules[] flags for the updated tg's subtree. A tg is | |
1280 | * considered to have rules if either the tg itself or any of its | |
1281 | * ancestors has rules. This identifies groups without any | |
1282 | * restrictions in the whole hierarchy and allows them to bypass | |
1283 | * blk-throttle. | |
1284 | */ | |
69948b07 | 1285 | blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg)) |
693e751e TH |
1286 | tg_update_has_rules(blkg_to_tg(blkg)); |
1287 | ||
632b4493 TH |
1288 | /* |
1289 | * We're already holding queue_lock and know @tg is valid. Let's | |
1290 | * apply the new config directly. | |
1291 | * | |
1292 | * Restart the slices for both READ and WRITES. It might happen | |
1293 | * that a group's limit are dropped suddenly and we don't want to | |
1294 | * account recently dispatched IO with new low rate. | |
1295 | */ | |
0f3457f6 TH |
1296 | throtl_start_new_slice(tg, 0); |
1297 | throtl_start_new_slice(tg, 1); | |
632b4493 | 1298 | |
5b2c16aa | 1299 | if (tg->flags & THROTL_TG_PENDING) { |
77216b04 | 1300 | tg_update_disptime(tg); |
7f52f98c | 1301 | throtl_schedule_next_dispatch(sq->parent_sq, true); |
632b4493 | 1302 | } |
69948b07 TH |
1303 | } |
1304 | ||
1305 | static ssize_t tg_set_conf(struct kernfs_open_file *of, | |
1306 | char *buf, size_t nbytes, loff_t off, bool is_u64) | |
1307 | { | |
1308 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1309 | struct blkg_conf_ctx ctx; | |
1310 | struct throtl_grp *tg; | |
1311 | int ret; | |
1312 | u64 v; | |
1313 | ||
1314 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1315 | if (ret) | |
1316 | return ret; | |
1317 | ||
1318 | ret = -EINVAL; | |
1319 | if (sscanf(ctx.body, "%llu", &v) != 1) | |
1320 | goto out_finish; | |
1321 | if (!v) | |
2ab5492d | 1322 | v = U64_MAX; |
69948b07 TH |
1323 | |
1324 | tg = blkg_to_tg(ctx.blkg); | |
1325 | ||
1326 | if (is_u64) | |
1327 | *(u64 *)((void *)tg + of_cft(of)->private) = v; | |
1328 | else | |
1329 | *(unsigned int *)((void *)tg + of_cft(of)->private) = v; | |
60c2bc2d | 1330 | |
69948b07 | 1331 | tg_conf_updated(tg); |
36aa9e5f TH |
1332 | ret = 0; |
1333 | out_finish: | |
60c2bc2d | 1334 | blkg_conf_finish(&ctx); |
36aa9e5f | 1335 | return ret ?: nbytes; |
8e89d13f VG |
1336 | } |
1337 | ||
451af504 TH |
1338 | static ssize_t tg_set_conf_u64(struct kernfs_open_file *of, |
1339 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1340 | { |
451af504 | 1341 | return tg_set_conf(of, buf, nbytes, off, true); |
60c2bc2d TH |
1342 | } |
1343 | ||
451af504 TH |
1344 | static ssize_t tg_set_conf_uint(struct kernfs_open_file *of, |
1345 | char *buf, size_t nbytes, loff_t off) | |
60c2bc2d | 1346 | { |
451af504 | 1347 | return tg_set_conf(of, buf, nbytes, off, false); |
60c2bc2d TH |
1348 | } |
1349 | ||
880f50e2 | 1350 | static struct cftype throtl_legacy_files[] = { |
60c2bc2d TH |
1351 | { |
1352 | .name = "throttle.read_bps_device", | |
9f626e37 | 1353 | .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]), |
2da8ca82 | 1354 | .seq_show = tg_print_conf_u64, |
451af504 | 1355 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1356 | }, |
1357 | { | |
1358 | .name = "throttle.write_bps_device", | |
9f626e37 | 1359 | .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]), |
2da8ca82 | 1360 | .seq_show = tg_print_conf_u64, |
451af504 | 1361 | .write = tg_set_conf_u64, |
60c2bc2d TH |
1362 | }, |
1363 | { | |
1364 | .name = "throttle.read_iops_device", | |
9f626e37 | 1365 | .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]), |
2da8ca82 | 1366 | .seq_show = tg_print_conf_uint, |
451af504 | 1367 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1368 | }, |
1369 | { | |
1370 | .name = "throttle.write_iops_device", | |
9f626e37 | 1371 | .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]), |
2da8ca82 | 1372 | .seq_show = tg_print_conf_uint, |
451af504 | 1373 | .write = tg_set_conf_uint, |
60c2bc2d TH |
1374 | }, |
1375 | { | |
1376 | .name = "throttle.io_service_bytes", | |
77ea7338 TH |
1377 | .private = (unsigned long)&blkcg_policy_throtl, |
1378 | .seq_show = blkg_print_stat_bytes, | |
60c2bc2d TH |
1379 | }, |
1380 | { | |
1381 | .name = "throttle.io_serviced", | |
77ea7338 TH |
1382 | .private = (unsigned long)&blkcg_policy_throtl, |
1383 | .seq_show = blkg_print_stat_ios, | |
60c2bc2d TH |
1384 | }, |
1385 | { } /* terminate */ | |
1386 | }; | |
1387 | ||
cd5ab1b0 | 1388 | static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd, |
2ee867dc TH |
1389 | int off) |
1390 | { | |
1391 | struct throtl_grp *tg = pd_to_tg(pd); | |
1392 | const char *dname = blkg_dev_name(pd->blkg); | |
1393 | char bufs[4][21] = { "max", "max", "max", "max" }; | |
cd5ab1b0 SL |
1394 | u64 bps_dft; |
1395 | unsigned int iops_dft; | |
2ee867dc TH |
1396 | |
1397 | if (!dname) | |
1398 | return 0; | |
9f626e37 | 1399 | |
cd5ab1b0 SL |
1400 | if (off == LIMIT_LOW) { |
1401 | bps_dft = 0; | |
1402 | iops_dft = 0; | |
1403 | } else { | |
1404 | bps_dft = U64_MAX; | |
1405 | iops_dft = UINT_MAX; | |
1406 | } | |
1407 | ||
1408 | if (tg->bps_conf[READ][off] == bps_dft && | |
1409 | tg->bps_conf[WRITE][off] == bps_dft && | |
1410 | tg->iops_conf[READ][off] == iops_dft && | |
1411 | tg->iops_conf[WRITE][off] == iops_dft) | |
2ee867dc TH |
1412 | return 0; |
1413 | ||
cd5ab1b0 | 1414 | if (tg->bps_conf[READ][off] != bps_dft) |
9f626e37 | 1415 | snprintf(bufs[0], sizeof(bufs[0]), "%llu", |
cd5ab1b0 SL |
1416 | tg->bps_conf[READ][off]); |
1417 | if (tg->bps_conf[WRITE][off] != bps_dft) | |
9f626e37 | 1418 | snprintf(bufs[1], sizeof(bufs[1]), "%llu", |
cd5ab1b0 SL |
1419 | tg->bps_conf[WRITE][off]); |
1420 | if (tg->iops_conf[READ][off] != iops_dft) | |
9f626e37 | 1421 | snprintf(bufs[2], sizeof(bufs[2]), "%u", |
cd5ab1b0 SL |
1422 | tg->iops_conf[READ][off]); |
1423 | if (tg->iops_conf[WRITE][off] != iops_dft) | |
9f626e37 | 1424 | snprintf(bufs[3], sizeof(bufs[3]), "%u", |
cd5ab1b0 | 1425 | tg->iops_conf[WRITE][off]); |
2ee867dc TH |
1426 | |
1427 | seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s\n", | |
1428 | dname, bufs[0], bufs[1], bufs[2], bufs[3]); | |
1429 | return 0; | |
1430 | } | |
1431 | ||
cd5ab1b0 | 1432 | static int tg_print_limit(struct seq_file *sf, void *v) |
2ee867dc | 1433 | { |
cd5ab1b0 | 1434 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit, |
2ee867dc TH |
1435 | &blkcg_policy_throtl, seq_cft(sf)->private, false); |
1436 | return 0; | |
1437 | } | |
1438 | ||
cd5ab1b0 | 1439 | static ssize_t tg_set_limit(struct kernfs_open_file *of, |
2ee867dc TH |
1440 | char *buf, size_t nbytes, loff_t off) |
1441 | { | |
1442 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1443 | struct blkg_conf_ctx ctx; | |
1444 | struct throtl_grp *tg; | |
1445 | u64 v[4]; | |
1446 | int ret; | |
cd5ab1b0 | 1447 | int index = of_cft(of)->private; |
2ee867dc TH |
1448 | |
1449 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); | |
1450 | if (ret) | |
1451 | return ret; | |
1452 | ||
1453 | tg = blkg_to_tg(ctx.blkg); | |
1454 | ||
cd5ab1b0 SL |
1455 | v[0] = tg->bps_conf[READ][index]; |
1456 | v[1] = tg->bps_conf[WRITE][index]; | |
1457 | v[2] = tg->iops_conf[READ][index]; | |
1458 | v[3] = tg->iops_conf[WRITE][index]; | |
2ee867dc TH |
1459 | |
1460 | while (true) { | |
1461 | char tok[27]; /* wiops=18446744073709551616 */ | |
1462 | char *p; | |
2ab5492d | 1463 | u64 val = U64_MAX; |
2ee867dc TH |
1464 | int len; |
1465 | ||
1466 | if (sscanf(ctx.body, "%26s%n", tok, &len) != 1) | |
1467 | break; | |
1468 | if (tok[0] == '\0') | |
1469 | break; | |
1470 | ctx.body += len; | |
1471 | ||
1472 | ret = -EINVAL; | |
1473 | p = tok; | |
1474 | strsep(&p, "="); | |
1475 | if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max"))) | |
1476 | goto out_finish; | |
1477 | ||
1478 | ret = -ERANGE; | |
1479 | if (!val) | |
1480 | goto out_finish; | |
1481 | ||
1482 | ret = -EINVAL; | |
1483 | if (!strcmp(tok, "rbps")) | |
1484 | v[0] = val; | |
1485 | else if (!strcmp(tok, "wbps")) | |
1486 | v[1] = val; | |
1487 | else if (!strcmp(tok, "riops")) | |
1488 | v[2] = min_t(u64, val, UINT_MAX); | |
1489 | else if (!strcmp(tok, "wiops")) | |
1490 | v[3] = min_t(u64, val, UINT_MAX); | |
1491 | else | |
1492 | goto out_finish; | |
1493 | } | |
1494 | ||
cd5ab1b0 SL |
1495 | tg->bps_conf[READ][index] = v[0]; |
1496 | tg->bps_conf[WRITE][index] = v[1]; | |
1497 | tg->iops_conf[READ][index] = v[2]; | |
1498 | tg->iops_conf[WRITE][index] = v[3]; | |
2ee867dc | 1499 | |
cd5ab1b0 SL |
1500 | if (index == LIMIT_MAX) { |
1501 | tg->bps[READ][index] = v[0]; | |
1502 | tg->bps[WRITE][index] = v[1]; | |
1503 | tg->iops[READ][index] = v[2]; | |
1504 | tg->iops[WRITE][index] = v[3]; | |
1505 | } | |
1506 | tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW], | |
1507 | tg->bps_conf[READ][LIMIT_MAX]); | |
1508 | tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW], | |
1509 | tg->bps_conf[WRITE][LIMIT_MAX]); | |
1510 | tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW], | |
1511 | tg->iops_conf[READ][LIMIT_MAX]); | |
1512 | tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW], | |
1513 | tg->iops_conf[WRITE][LIMIT_MAX]); | |
1514 | ||
1515 | if (index == LIMIT_LOW) { | |
1516 | blk_throtl_update_limit_valid(tg->td); | |
1517 | if (tg->td->limit_valid[LIMIT_LOW]) | |
1518 | tg->td->limit_index = LIMIT_LOW; | |
1519 | } | |
2ee867dc TH |
1520 | tg_conf_updated(tg); |
1521 | ret = 0; | |
1522 | out_finish: | |
1523 | blkg_conf_finish(&ctx); | |
1524 | return ret ?: nbytes; | |
1525 | } | |
1526 | ||
1527 | static struct cftype throtl_files[] = { | |
cd5ab1b0 SL |
1528 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
1529 | { | |
1530 | .name = "low", | |
1531 | .flags = CFTYPE_NOT_ON_ROOT, | |
1532 | .seq_show = tg_print_limit, | |
1533 | .write = tg_set_limit, | |
1534 | .private = LIMIT_LOW, | |
1535 | }, | |
1536 | #endif | |
2ee867dc TH |
1537 | { |
1538 | .name = "max", | |
1539 | .flags = CFTYPE_NOT_ON_ROOT, | |
cd5ab1b0 SL |
1540 | .seq_show = tg_print_limit, |
1541 | .write = tg_set_limit, | |
1542 | .private = LIMIT_MAX, | |
2ee867dc TH |
1543 | }, |
1544 | { } /* terminate */ | |
1545 | }; | |
1546 | ||
da527770 | 1547 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1548 | { |
1549 | struct throtl_data *td = q->td; | |
1550 | ||
69df0ab0 | 1551 | cancel_work_sync(&td->dispatch_work); |
e43473b7 VG |
1552 | } |
1553 | ||
3c798398 | 1554 | static struct blkcg_policy blkcg_policy_throtl = { |
2ee867dc | 1555 | .dfl_cftypes = throtl_files, |
880f50e2 | 1556 | .legacy_cftypes = throtl_legacy_files, |
f9fcc2d3 | 1557 | |
001bea73 | 1558 | .pd_alloc_fn = throtl_pd_alloc, |
f9fcc2d3 | 1559 | .pd_init_fn = throtl_pd_init, |
693e751e | 1560 | .pd_online_fn = throtl_pd_online, |
cd5ab1b0 | 1561 | .pd_offline_fn = throtl_pd_offline, |
001bea73 | 1562 | .pd_free_fn = throtl_pd_free, |
e43473b7 VG |
1563 | }; |
1564 | ||
3f0abd80 SL |
1565 | static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg) |
1566 | { | |
1567 | unsigned long rtime = jiffies, wtime = jiffies; | |
1568 | ||
1569 | if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]) | |
1570 | rtime = tg->last_low_overflow_time[READ]; | |
1571 | if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) | |
1572 | wtime = tg->last_low_overflow_time[WRITE]; | |
1573 | return min(rtime, wtime); | |
1574 | } | |
1575 | ||
1576 | /* tg should not be an intermediate node */ | |
1577 | static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg) | |
1578 | { | |
1579 | struct throtl_service_queue *parent_sq; | |
1580 | struct throtl_grp *parent = tg; | |
1581 | unsigned long ret = __tg_last_low_overflow_time(tg); | |
1582 | ||
1583 | while (true) { | |
1584 | parent_sq = parent->service_queue.parent_sq; | |
1585 | parent = sq_to_tg(parent_sq); | |
1586 | if (!parent) | |
1587 | break; | |
1588 | ||
1589 | /* | |
1590 | * The parent doesn't have low limit, it always reaches low | |
1591 | * limit. Its overflow time is useless for children | |
1592 | */ | |
1593 | if (!parent->bps[READ][LIMIT_LOW] && | |
1594 | !parent->iops[READ][LIMIT_LOW] && | |
1595 | !parent->bps[WRITE][LIMIT_LOW] && | |
1596 | !parent->iops[WRITE][LIMIT_LOW]) | |
1597 | continue; | |
1598 | if (time_after(__tg_last_low_overflow_time(parent), ret)) | |
1599 | ret = __tg_last_low_overflow_time(parent); | |
1600 | } | |
1601 | return ret; | |
1602 | } | |
1603 | ||
c79892c5 SL |
1604 | static bool throtl_tg_can_upgrade(struct throtl_grp *tg) |
1605 | { | |
1606 | struct throtl_service_queue *sq = &tg->service_queue; | |
1607 | bool read_limit, write_limit; | |
1608 | ||
1609 | /* | |
1610 | * if cgroup reaches low limit (if low limit is 0, the cgroup always | |
1611 | * reaches), it's ok to upgrade to next limit | |
1612 | */ | |
1613 | read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW]; | |
1614 | write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]; | |
1615 | if (!read_limit && !write_limit) | |
1616 | return true; | |
1617 | if (read_limit && sq->nr_queued[READ] && | |
1618 | (!write_limit || sq->nr_queued[WRITE])) | |
1619 | return true; | |
1620 | if (write_limit && sq->nr_queued[WRITE] && | |
1621 | (!read_limit || sq->nr_queued[READ])) | |
1622 | return true; | |
aec24246 SL |
1623 | |
1624 | if (time_after_eq(jiffies, | |
1625 | tg->last_dispatch_time[READ] + tg->td->throtl_slice) && | |
1626 | time_after_eq(jiffies, | |
1627 | tg->last_dispatch_time[WRITE] + tg->td->throtl_slice)) | |
1628 | return true; | |
c79892c5 SL |
1629 | return false; |
1630 | } | |
1631 | ||
1632 | static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg) | |
1633 | { | |
1634 | while (true) { | |
1635 | if (throtl_tg_can_upgrade(tg)) | |
1636 | return true; | |
1637 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1638 | if (!tg || !tg_to_blkg(tg)->parent) | |
1639 | return false; | |
1640 | } | |
1641 | return false; | |
1642 | } | |
1643 | ||
1644 | static bool throtl_can_upgrade(struct throtl_data *td, | |
1645 | struct throtl_grp *this_tg) | |
1646 | { | |
1647 | struct cgroup_subsys_state *pos_css; | |
1648 | struct blkcg_gq *blkg; | |
1649 | ||
1650 | if (td->limit_index != LIMIT_LOW) | |
1651 | return false; | |
1652 | ||
297e3d85 | 1653 | if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice)) |
3f0abd80 SL |
1654 | return false; |
1655 | ||
c79892c5 SL |
1656 | rcu_read_lock(); |
1657 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1658 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1659 | ||
1660 | if (tg == this_tg) | |
1661 | continue; | |
1662 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1663 | continue; | |
1664 | if (!throtl_hierarchy_can_upgrade(tg)) { | |
1665 | rcu_read_unlock(); | |
1666 | return false; | |
1667 | } | |
1668 | } | |
1669 | rcu_read_unlock(); | |
1670 | return true; | |
1671 | } | |
1672 | ||
1673 | static void throtl_upgrade_state(struct throtl_data *td) | |
1674 | { | |
1675 | struct cgroup_subsys_state *pos_css; | |
1676 | struct blkcg_gq *blkg; | |
1677 | ||
1678 | td->limit_index = LIMIT_MAX; | |
3f0abd80 | 1679 | td->low_upgrade_time = jiffies; |
c79892c5 SL |
1680 | rcu_read_lock(); |
1681 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) { | |
1682 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
1683 | struct throtl_service_queue *sq = &tg->service_queue; | |
1684 | ||
1685 | tg->disptime = jiffies - 1; | |
1686 | throtl_select_dispatch(sq); | |
1687 | throtl_schedule_next_dispatch(sq, false); | |
1688 | } | |
1689 | rcu_read_unlock(); | |
1690 | throtl_select_dispatch(&td->service_queue); | |
1691 | throtl_schedule_next_dispatch(&td->service_queue, false); | |
1692 | queue_work(kthrotld_workqueue, &td->dispatch_work); | |
1693 | } | |
1694 | ||
3f0abd80 SL |
1695 | static void throtl_downgrade_state(struct throtl_data *td, int new) |
1696 | { | |
1697 | td->limit_index = new; | |
1698 | td->low_downgrade_time = jiffies; | |
1699 | } | |
1700 | ||
1701 | static bool throtl_tg_can_downgrade(struct throtl_grp *tg) | |
1702 | { | |
1703 | struct throtl_data *td = tg->td; | |
1704 | unsigned long now = jiffies; | |
1705 | ||
aec24246 SL |
1706 | if (time_after_eq(now, tg->last_dispatch_time[READ] + |
1707 | td->throtl_slice) && | |
1708 | time_after_eq(now, tg->last_dispatch_time[WRITE] + | |
1709 | td->throtl_slice)) | |
1710 | return false; | |
3f0abd80 SL |
1711 | /* |
1712 | * If cgroup is below low limit, consider downgrade and throttle other | |
1713 | * cgroups | |
1714 | */ | |
297e3d85 SL |
1715 | if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) && |
1716 | time_after_eq(now, tg_last_low_overflow_time(tg) + | |
1717 | td->throtl_slice)) | |
3f0abd80 SL |
1718 | return true; |
1719 | return false; | |
1720 | } | |
1721 | ||
1722 | static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg) | |
1723 | { | |
1724 | while (true) { | |
1725 | if (!throtl_tg_can_downgrade(tg)) | |
1726 | return false; | |
1727 | tg = sq_to_tg(tg->service_queue.parent_sq); | |
1728 | if (!tg || !tg_to_blkg(tg)->parent) | |
1729 | break; | |
1730 | } | |
1731 | return true; | |
1732 | } | |
1733 | ||
1734 | static void throtl_downgrade_check(struct throtl_grp *tg) | |
1735 | { | |
1736 | uint64_t bps; | |
1737 | unsigned int iops; | |
1738 | unsigned long elapsed_time; | |
1739 | unsigned long now = jiffies; | |
1740 | ||
1741 | if (tg->td->limit_index != LIMIT_MAX || | |
1742 | !tg->td->limit_valid[LIMIT_LOW]) | |
1743 | return; | |
1744 | if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children)) | |
1745 | return; | |
297e3d85 | 1746 | if (time_after(tg->last_check_time + tg->td->throtl_slice, now)) |
3f0abd80 SL |
1747 | return; |
1748 | ||
1749 | elapsed_time = now - tg->last_check_time; | |
1750 | tg->last_check_time = now; | |
1751 | ||
297e3d85 SL |
1752 | if (time_before(now, tg_last_low_overflow_time(tg) + |
1753 | tg->td->throtl_slice)) | |
3f0abd80 SL |
1754 | return; |
1755 | ||
1756 | if (tg->bps[READ][LIMIT_LOW]) { | |
1757 | bps = tg->last_bytes_disp[READ] * HZ; | |
1758 | do_div(bps, elapsed_time); | |
1759 | if (bps >= tg->bps[READ][LIMIT_LOW]) | |
1760 | tg->last_low_overflow_time[READ] = now; | |
1761 | } | |
1762 | ||
1763 | if (tg->bps[WRITE][LIMIT_LOW]) { | |
1764 | bps = tg->last_bytes_disp[WRITE] * HZ; | |
1765 | do_div(bps, elapsed_time); | |
1766 | if (bps >= tg->bps[WRITE][LIMIT_LOW]) | |
1767 | tg->last_low_overflow_time[WRITE] = now; | |
1768 | } | |
1769 | ||
1770 | if (tg->iops[READ][LIMIT_LOW]) { | |
1771 | iops = tg->last_io_disp[READ] * HZ / elapsed_time; | |
1772 | if (iops >= tg->iops[READ][LIMIT_LOW]) | |
1773 | tg->last_low_overflow_time[READ] = now; | |
1774 | } | |
1775 | ||
1776 | if (tg->iops[WRITE][LIMIT_LOW]) { | |
1777 | iops = tg->last_io_disp[WRITE] * HZ / elapsed_time; | |
1778 | if (iops >= tg->iops[WRITE][LIMIT_LOW]) | |
1779 | tg->last_low_overflow_time[WRITE] = now; | |
1780 | } | |
1781 | ||
1782 | /* | |
1783 | * If cgroup is below low limit, consider downgrade and throttle other | |
1784 | * cgroups | |
1785 | */ | |
1786 | if (throtl_hierarchy_can_downgrade(tg)) | |
1787 | throtl_downgrade_state(tg->td, LIMIT_LOW); | |
1788 | ||
1789 | tg->last_bytes_disp[READ] = 0; | |
1790 | tg->last_bytes_disp[WRITE] = 0; | |
1791 | tg->last_io_disp[READ] = 0; | |
1792 | tg->last_io_disp[WRITE] = 0; | |
1793 | } | |
1794 | ||
ae118896 TH |
1795 | bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg, |
1796 | struct bio *bio) | |
e43473b7 | 1797 | { |
c5cc2070 | 1798 | struct throtl_qnode *qn = NULL; |
ae118896 | 1799 | struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg); |
73f0d49a | 1800 | struct throtl_service_queue *sq; |
0e9f4164 | 1801 | bool rw = bio_data_dir(bio); |
bc16a4f9 | 1802 | bool throttled = false; |
e43473b7 | 1803 | |
ae118896 TH |
1804 | WARN_ON_ONCE(!rcu_read_lock_held()); |
1805 | ||
2a0f61e6 | 1806 | /* see throtl_charge_bio() */ |
8d2bbd4c | 1807 | if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw]) |
bc16a4f9 | 1808 | goto out; |
e43473b7 VG |
1809 | |
1810 | spin_lock_irq(q->queue_lock); | |
c9589f03 TH |
1811 | |
1812 | if (unlikely(blk_queue_bypass(q))) | |
bc16a4f9 | 1813 | goto out_unlock; |
f469a7b4 | 1814 | |
73f0d49a TH |
1815 | sq = &tg->service_queue; |
1816 | ||
c79892c5 | 1817 | again: |
9e660acf | 1818 | while (true) { |
aec24246 | 1819 | tg->last_dispatch_time[rw] = jiffies; |
3f0abd80 SL |
1820 | if (tg->last_low_overflow_time[rw] == 0) |
1821 | tg->last_low_overflow_time[rw] = jiffies; | |
1822 | throtl_downgrade_check(tg); | |
9e660acf TH |
1823 | /* throtl is FIFO - if bios are already queued, should queue */ |
1824 | if (sq->nr_queued[rw]) | |
1825 | break; | |
de701c74 | 1826 | |
9e660acf | 1827 | /* if above limits, break to queue */ |
c79892c5 | 1828 | if (!tg_may_dispatch(tg, bio, NULL)) { |
3f0abd80 | 1829 | tg->last_low_overflow_time[rw] = jiffies; |
c79892c5 SL |
1830 | if (throtl_can_upgrade(tg->td, tg)) { |
1831 | throtl_upgrade_state(tg->td); | |
1832 | goto again; | |
1833 | } | |
9e660acf | 1834 | break; |
c79892c5 | 1835 | } |
9e660acf TH |
1836 | |
1837 | /* within limits, let's charge and dispatch directly */ | |
e43473b7 | 1838 | throtl_charge_bio(tg, bio); |
04521db0 VG |
1839 | |
1840 | /* | |
1841 | * We need to trim slice even when bios are not being queued | |
1842 | * otherwise it might happen that a bio is not queued for | |
1843 | * a long time and slice keeps on extending and trim is not | |
1844 | * called for a long time. Now if limits are reduced suddenly | |
1845 | * we take into account all the IO dispatched so far at new | |
1846 | * low rate and * newly queued IO gets a really long dispatch | |
1847 | * time. | |
1848 | * | |
1849 | * So keep on trimming slice even if bio is not queued. | |
1850 | */ | |
0f3457f6 | 1851 | throtl_trim_slice(tg, rw); |
9e660acf TH |
1852 | |
1853 | /* | |
1854 | * @bio passed through this layer without being throttled. | |
1855 | * Climb up the ladder. If we''re already at the top, it | |
1856 | * can be executed directly. | |
1857 | */ | |
c5cc2070 | 1858 | qn = &tg->qnode_on_parent[rw]; |
9e660acf TH |
1859 | sq = sq->parent_sq; |
1860 | tg = sq_to_tg(sq); | |
1861 | if (!tg) | |
1862 | goto out_unlock; | |
e43473b7 VG |
1863 | } |
1864 | ||
9e660acf | 1865 | /* out-of-limit, queue to @tg */ |
fda6f272 TH |
1866 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
1867 | rw == READ ? 'R' : 'W', | |
9f626e37 SL |
1868 | tg->bytes_disp[rw], bio->bi_iter.bi_size, |
1869 | tg_bps_limit(tg, rw), | |
1870 | tg->io_disp[rw], tg_iops_limit(tg, rw), | |
fda6f272 | 1871 | sq->nr_queued[READ], sq->nr_queued[WRITE]); |
e43473b7 | 1872 | |
3f0abd80 SL |
1873 | tg->last_low_overflow_time[rw] = jiffies; |
1874 | ||
671058fb | 1875 | bio_associate_current(bio); |
6bc9c2b4 | 1876 | tg->td->nr_queued[rw]++; |
c5cc2070 | 1877 | throtl_add_bio_tg(bio, qn, tg); |
bc16a4f9 | 1878 | throttled = true; |
e43473b7 | 1879 | |
7f52f98c TH |
1880 | /* |
1881 | * Update @tg's dispatch time and force schedule dispatch if @tg | |
1882 | * was empty before @bio. The forced scheduling isn't likely to | |
1883 | * cause undue delay as @bio is likely to be dispatched directly if | |
1884 | * its @tg's disptime is not in the future. | |
1885 | */ | |
0e9f4164 | 1886 | if (tg->flags & THROTL_TG_WAS_EMPTY) { |
77216b04 | 1887 | tg_update_disptime(tg); |
7f52f98c | 1888 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); |
e43473b7 VG |
1889 | } |
1890 | ||
bc16a4f9 | 1891 | out_unlock: |
e43473b7 | 1892 | spin_unlock_irq(q->queue_lock); |
bc16a4f9 | 1893 | out: |
2a0f61e6 TH |
1894 | /* |
1895 | * As multiple blk-throtls may stack in the same issue path, we | |
1896 | * don't want bios to leave with the flag set. Clear the flag if | |
1897 | * being issued. | |
1898 | */ | |
1899 | if (!throttled) | |
8d2bbd4c | 1900 | bio_clear_flag(bio, BIO_THROTTLED); |
bc16a4f9 | 1901 | return throttled; |
e43473b7 VG |
1902 | } |
1903 | ||
2a12f0dc TH |
1904 | /* |
1905 | * Dispatch all bios from all children tg's queued on @parent_sq. On | |
1906 | * return, @parent_sq is guaranteed to not have any active children tg's | |
1907 | * and all bios from previously active tg's are on @parent_sq->bio_lists[]. | |
1908 | */ | |
1909 | static void tg_drain_bios(struct throtl_service_queue *parent_sq) | |
1910 | { | |
1911 | struct throtl_grp *tg; | |
1912 | ||
1913 | while ((tg = throtl_rb_first(parent_sq))) { | |
1914 | struct throtl_service_queue *sq = &tg->service_queue; | |
1915 | struct bio *bio; | |
1916 | ||
1917 | throtl_dequeue_tg(tg); | |
1918 | ||
c5cc2070 | 1919 | while ((bio = throtl_peek_queued(&sq->queued[READ]))) |
2a12f0dc | 1920 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
c5cc2070 | 1921 | while ((bio = throtl_peek_queued(&sq->queued[WRITE]))) |
2a12f0dc TH |
1922 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
1923 | } | |
1924 | } | |
1925 | ||
c9a929dd TH |
1926 | /** |
1927 | * blk_throtl_drain - drain throttled bios | |
1928 | * @q: request_queue to drain throttled bios for | |
1929 | * | |
1930 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1931 | */ | |
1932 | void blk_throtl_drain(struct request_queue *q) | |
1933 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1934 | { | |
1935 | struct throtl_data *td = q->td; | |
2a12f0dc | 1936 | struct blkcg_gq *blkg; |
492eb21b | 1937 | struct cgroup_subsys_state *pos_css; |
c9a929dd | 1938 | struct bio *bio; |
651930bc | 1939 | int rw; |
c9a929dd | 1940 | |
8bcb6c7d | 1941 | queue_lockdep_assert_held(q); |
2a12f0dc | 1942 | rcu_read_lock(); |
c9a929dd | 1943 | |
2a12f0dc TH |
1944 | /* |
1945 | * Drain each tg while doing post-order walk on the blkg tree, so | |
1946 | * that all bios are propagated to td->service_queue. It'd be | |
1947 | * better to walk service_queue tree directly but blkg walk is | |
1948 | * easier. | |
1949 | */ | |
492eb21b | 1950 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) |
2a12f0dc | 1951 | tg_drain_bios(&blkg_to_tg(blkg)->service_queue); |
73f0d49a | 1952 | |
2a12f0dc TH |
1953 | /* finally, transfer bios from top-level tg's into the td */ |
1954 | tg_drain_bios(&td->service_queue); | |
1955 | ||
1956 | rcu_read_unlock(); | |
c9a929dd TH |
1957 | spin_unlock_irq(q->queue_lock); |
1958 | ||
2a12f0dc | 1959 | /* all bios now should be in td->service_queue, issue them */ |
651930bc | 1960 | for (rw = READ; rw <= WRITE; rw++) |
c5cc2070 TH |
1961 | while ((bio = throtl_pop_queued(&td->service_queue.queued[rw], |
1962 | NULL))) | |
651930bc | 1963 | generic_make_request(bio); |
c9a929dd TH |
1964 | |
1965 | spin_lock_irq(q->queue_lock); | |
1966 | } | |
1967 | ||
e43473b7 VG |
1968 | int blk_throtl_init(struct request_queue *q) |
1969 | { | |
1970 | struct throtl_data *td; | |
a2b1693b | 1971 | int ret; |
e43473b7 VG |
1972 | |
1973 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1974 | if (!td) | |
1975 | return -ENOMEM; | |
1976 | ||
69df0ab0 | 1977 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
b2ce2643 | 1978 | throtl_service_queue_init(&td->service_queue); |
e43473b7 | 1979 | |
cd1604fa | 1980 | q->td = td; |
29b12589 | 1981 | td->queue = q; |
02977e4a | 1982 | |
9f626e37 | 1983 | td->limit_valid[LIMIT_MAX] = true; |
cd5ab1b0 | 1984 | td->limit_index = LIMIT_MAX; |
3f0abd80 SL |
1985 | td->low_upgrade_time = jiffies; |
1986 | td->low_downgrade_time = jiffies; | |
a2b1693b | 1987 | /* activate policy */ |
3c798398 | 1988 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1989 | if (ret) |
f51b802c | 1990 | kfree(td); |
a2b1693b | 1991 | return ret; |
e43473b7 VG |
1992 | } |
1993 | ||
1994 | void blk_throtl_exit(struct request_queue *q) | |
1995 | { | |
c875f4d0 | 1996 | BUG_ON(!q->td); |
da527770 | 1997 | throtl_shutdown_wq(q); |
3c798398 | 1998 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1999 | kfree(q->td); |
e43473b7 VG |
2000 | } |
2001 | ||
d61fcfa4 SL |
2002 | void blk_throtl_register_queue(struct request_queue *q) |
2003 | { | |
2004 | struct throtl_data *td; | |
2005 | ||
2006 | td = q->td; | |
2007 | BUG_ON(!td); | |
2008 | ||
2009 | if (blk_queue_nonrot(q)) | |
2010 | td->throtl_slice = DFL_THROTL_SLICE_SSD; | |
2011 | else | |
2012 | td->throtl_slice = DFL_THROTL_SLICE_HD; | |
2013 | #ifndef CONFIG_BLK_DEV_THROTTLING_LOW | |
2014 | /* if no low limit, use previous default */ | |
2015 | td->throtl_slice = DFL_THROTL_SLICE_HD; | |
2016 | #endif | |
2017 | } | |
2018 | ||
297e3d85 SL |
2019 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW |
2020 | ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page) | |
2021 | { | |
2022 | if (!q->td) | |
2023 | return -EINVAL; | |
2024 | return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice)); | |
2025 | } | |
2026 | ||
2027 | ssize_t blk_throtl_sample_time_store(struct request_queue *q, | |
2028 | const char *page, size_t count) | |
2029 | { | |
2030 | unsigned long v; | |
2031 | unsigned long t; | |
2032 | ||
2033 | if (!q->td) | |
2034 | return -EINVAL; | |
2035 | if (kstrtoul(page, 10, &v)) | |
2036 | return -EINVAL; | |
2037 | t = msecs_to_jiffies(v); | |
2038 | if (t == 0 || t > MAX_THROTL_SLICE) | |
2039 | return -EINVAL; | |
2040 | q->td->throtl_slice = t; | |
2041 | return count; | |
2042 | } | |
2043 | #endif | |
2044 | ||
e43473b7 VG |
2045 | static int __init throtl_init(void) |
2046 | { | |
450adcbe VG |
2047 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
2048 | if (!kthrotld_workqueue) | |
2049 | panic("Failed to create kthrotld\n"); | |
2050 | ||
3c798398 | 2051 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
2052 | } |
2053 | ||
2054 | module_init(throtl_init); |