blob: 24a5c5329bcd0de38f274f0166c90d32f01bb5ac [file] [log] [blame]
Christoph Hellwiga497ee32019-04-30 14:42:40 -04001// SPDX-License-Identifier: GPL-2.0-or-later
Paolo Valenteea25da42017-04-19 08:48:24 -06002/*
3 * cgroups support for the BFQ I/O scheduler.
Paolo Valenteea25da42017-04-19 08:48:24 -06004 */
5#include <linux/module.h>
6#include <linux/slab.h>
7#include <linux/blkdev.h>
8#include <linux/cgroup.h>
Paolo Valenteea25da42017-04-19 08:48:24 -06009#include <linux/ktime.h>
10#include <linux/rbtree.h>
11#include <linux/ioprio.h>
12#include <linux/sbitmap.h>
13#include <linux/delay.h>
14
Christoph Hellwig2e9bc342021-09-20 14:33:23 +020015#include "elevator.h"
Paolo Valenteea25da42017-04-19 08:48:24 -060016#include "bfq-iosched.h"
17
Christoph Hellwig8060c472019-06-06 12:26:24 +020018#ifdef CONFIG_BFQ_CGROUP_DEBUG
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +020019static int bfq_stat_init(struct bfq_stat *stat, gfp_t gfp)
20{
21 int ret;
22
23 ret = percpu_counter_init(&stat->cpu_cnt, 0, gfp);
24 if (ret)
25 return ret;
26
27 atomic64_set(&stat->aux_cnt, 0);
28 return 0;
29}
30
31static void bfq_stat_exit(struct bfq_stat *stat)
32{
33 percpu_counter_destroy(&stat->cpu_cnt);
34}
35
36/**
37 * bfq_stat_add - add a value to a bfq_stat
38 * @stat: target bfq_stat
39 * @val: value to add
40 *
41 * Add @val to @stat. The caller must ensure that IRQ on the same CPU
42 * don't re-enter this function for the same counter.
43 */
44static inline void bfq_stat_add(struct bfq_stat *stat, uint64_t val)
45{
46 percpu_counter_add_batch(&stat->cpu_cnt, val, BLKG_STAT_CPU_BATCH);
47}
48
49/**
50 * bfq_stat_read - read the current value of a bfq_stat
51 * @stat: bfq_stat to read
52 */
53static inline uint64_t bfq_stat_read(struct bfq_stat *stat)
54{
55 return percpu_counter_sum_positive(&stat->cpu_cnt);
56}
57
58/**
59 * bfq_stat_reset - reset a bfq_stat
60 * @stat: bfq_stat to reset
61 */
62static inline void bfq_stat_reset(struct bfq_stat *stat)
63{
64 percpu_counter_set(&stat->cpu_cnt, 0);
65 atomic64_set(&stat->aux_cnt, 0);
66}
67
68/**
69 * bfq_stat_add_aux - add a bfq_stat into another's aux count
70 * @to: the destination bfq_stat
71 * @from: the source
72 *
73 * Add @from's count including the aux one to @to's aux count.
74 */
75static inline void bfq_stat_add_aux(struct bfq_stat *to,
76 struct bfq_stat *from)
77{
78 atomic64_add(bfq_stat_read(from) + atomic64_read(&from->aux_cnt),
79 &to->aux_cnt);
80}
81
82/**
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +020083 * blkg_prfill_stat - prfill callback for bfq_stat
84 * @sf: seq_file to print to
85 * @pd: policy private data of interest
86 * @off: offset to the bfq_stat in @pd
87 *
88 * prfill callback for printing a bfq_stat.
89 */
90static u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd,
91 int off)
92{
93 return __blkg_prfill_u64(sf, pd, bfq_stat_read((void *)pd + off));
94}
95
Paolo Valenteea25da42017-04-19 08:48:24 -060096/* bfqg stats flags */
97enum bfqg_stats_flags {
98 BFQG_stats_waiting = 0,
99 BFQG_stats_idling,
100 BFQG_stats_empty,
101};
102
103#define BFQG_FLAG_FNS(name) \
104static void bfqg_stats_mark_##name(struct bfqg_stats *stats) \
105{ \
106 stats->flags |= (1 << BFQG_stats_##name); \
107} \
108static void bfqg_stats_clear_##name(struct bfqg_stats *stats) \
109{ \
110 stats->flags &= ~(1 << BFQG_stats_##name); \
111} \
112static int bfqg_stats_##name(struct bfqg_stats *stats) \
113{ \
114 return (stats->flags & (1 << BFQG_stats_##name)) != 0; \
115} \
116
117BFQG_FLAG_FNS(waiting)
118BFQG_FLAG_FNS(idling)
119BFQG_FLAG_FNS(empty)
120#undef BFQG_FLAG_FNS
121
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200122/* This should be called with the scheduler lock held. */
Paolo Valenteea25da42017-04-19 08:48:24 -0600123static void bfqg_stats_update_group_wait_time(struct bfqg_stats *stats)
124{
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700125 u64 now;
Paolo Valenteea25da42017-04-19 08:48:24 -0600126
127 if (!bfqg_stats_waiting(stats))
128 return;
129
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700130 now = ktime_get_ns();
131 if (now > stats->start_group_wait_time)
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +0200132 bfq_stat_add(&stats->group_wait_time,
Paolo Valenteea25da42017-04-19 08:48:24 -0600133 now - stats->start_group_wait_time);
134 bfqg_stats_clear_waiting(stats);
135}
136
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200137/* This should be called with the scheduler lock held. */
Paolo Valenteea25da42017-04-19 08:48:24 -0600138static void bfqg_stats_set_start_group_wait_time(struct bfq_group *bfqg,
139 struct bfq_group *curr_bfqg)
140{
141 struct bfqg_stats *stats = &bfqg->stats;
142
143 if (bfqg_stats_waiting(stats))
144 return;
145 if (bfqg == curr_bfqg)
146 return;
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700147 stats->start_group_wait_time = ktime_get_ns();
Paolo Valenteea25da42017-04-19 08:48:24 -0600148 bfqg_stats_mark_waiting(stats);
149}
150
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200151/* This should be called with the scheduler lock held. */
Paolo Valenteea25da42017-04-19 08:48:24 -0600152static void bfqg_stats_end_empty_time(struct bfqg_stats *stats)
153{
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700154 u64 now;
Paolo Valenteea25da42017-04-19 08:48:24 -0600155
156 if (!bfqg_stats_empty(stats))
157 return;
158
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700159 now = ktime_get_ns();
160 if (now > stats->start_empty_time)
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +0200161 bfq_stat_add(&stats->empty_time,
Paolo Valenteea25da42017-04-19 08:48:24 -0600162 now - stats->start_empty_time);
163 bfqg_stats_clear_empty(stats);
164}
165
166void bfqg_stats_update_dequeue(struct bfq_group *bfqg)
167{
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +0200168 bfq_stat_add(&bfqg->stats.dequeue, 1);
Paolo Valenteea25da42017-04-19 08:48:24 -0600169}
170
171void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg)
172{
173 struct bfqg_stats *stats = &bfqg->stats;
174
175 if (blkg_rwstat_total(&stats->queued))
176 return;
177
178 /*
179 * group is already marked empty. This can happen if bfqq got new
180 * request in parent group and moved to this group while being added
181 * to service tree. Just ignore the event and move on.
182 */
183 if (bfqg_stats_empty(stats))
184 return;
185
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700186 stats->start_empty_time = ktime_get_ns();
Paolo Valenteea25da42017-04-19 08:48:24 -0600187 bfqg_stats_mark_empty(stats);
188}
189
190void bfqg_stats_update_idle_time(struct bfq_group *bfqg)
191{
192 struct bfqg_stats *stats = &bfqg->stats;
193
194 if (bfqg_stats_idling(stats)) {
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700195 u64 now = ktime_get_ns();
Paolo Valenteea25da42017-04-19 08:48:24 -0600196
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700197 if (now > stats->start_idle_time)
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +0200198 bfq_stat_add(&stats->idle_time,
Paolo Valenteea25da42017-04-19 08:48:24 -0600199 now - stats->start_idle_time);
200 bfqg_stats_clear_idling(stats);
201 }
202}
203
204void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg)
205{
206 struct bfqg_stats *stats = &bfqg->stats;
207
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700208 stats->start_idle_time = ktime_get_ns();
Paolo Valenteea25da42017-04-19 08:48:24 -0600209 bfqg_stats_mark_idling(stats);
210}
211
212void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg)
213{
214 struct bfqg_stats *stats = &bfqg->stats;
215
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +0200216 bfq_stat_add(&stats->avg_queue_size_sum,
Paolo Valenteea25da42017-04-19 08:48:24 -0600217 blkg_rwstat_total(&stats->queued));
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +0200218 bfq_stat_add(&stats->avg_queue_size_samples, 1);
Paolo Valenteea25da42017-04-19 08:48:24 -0600219 bfqg_stats_update_group_wait_time(stats);
220}
221
Luca Miccioa33801e2017-11-13 07:34:10 +0100222void bfqg_stats_update_io_add(struct bfq_group *bfqg, struct bfq_queue *bfqq,
223 unsigned int op)
224{
225 blkg_rwstat_add(&bfqg->stats.queued, op, 1);
226 bfqg_stats_end_empty_time(&bfqg->stats);
227 if (!(bfqq == ((struct bfq_data *)bfqg->bfqd)->in_service_queue))
228 bfqg_stats_set_start_group_wait_time(bfqg, bfqq_group(bfqq));
229}
230
231void bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op)
232{
233 blkg_rwstat_add(&bfqg->stats.queued, op, -1);
234}
235
236void bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op)
237{
238 blkg_rwstat_add(&bfqg->stats.merged, op, 1);
239}
240
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700241void bfqg_stats_update_completion(struct bfq_group *bfqg, u64 start_time_ns,
242 u64 io_start_time_ns, unsigned int op)
Luca Miccioa33801e2017-11-13 07:34:10 +0100243{
244 struct bfqg_stats *stats = &bfqg->stats;
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700245 u64 now = ktime_get_ns();
Luca Miccioa33801e2017-11-13 07:34:10 +0100246
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700247 if (now > io_start_time_ns)
Luca Miccioa33801e2017-11-13 07:34:10 +0100248 blkg_rwstat_add(&stats->service_time, op,
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700249 now - io_start_time_ns);
250 if (io_start_time_ns > start_time_ns)
Luca Miccioa33801e2017-11-13 07:34:10 +0100251 blkg_rwstat_add(&stats->wait_time, op,
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700252 io_start_time_ns - start_time_ns);
Luca Miccioa33801e2017-11-13 07:34:10 +0100253}
254
Christoph Hellwig8060c472019-06-06 12:26:24 +0200255#else /* CONFIG_BFQ_CGROUP_DEBUG */
Luca Miccioa33801e2017-11-13 07:34:10 +0100256
257void bfqg_stats_update_io_add(struct bfq_group *bfqg, struct bfq_queue *bfqq,
258 unsigned int op) { }
259void bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op) { }
260void bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op) { }
Omar Sandoval84c7afc2018-05-09 02:08:51 -0700261void bfqg_stats_update_completion(struct bfq_group *bfqg, u64 start_time_ns,
262 u64 io_start_time_ns, unsigned int op) { }
Luca Miccioa33801e2017-11-13 07:34:10 +0100263void bfqg_stats_update_dequeue(struct bfq_group *bfqg) { }
264void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg) { }
265void bfqg_stats_update_idle_time(struct bfq_group *bfqg) { }
266void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg) { }
267void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg) { }
268
Christoph Hellwig8060c472019-06-06 12:26:24 +0200269#endif /* CONFIG_BFQ_CGROUP_DEBUG */
Luca Miccioa33801e2017-11-13 07:34:10 +0100270
271#ifdef CONFIG_BFQ_GROUP_IOSCHED
272
Paolo Valenteea25da42017-04-19 08:48:24 -0600273/*
274 * blk-cgroup policy-related handlers
275 * The following functions help in converting between blk-cgroup
276 * internal structures and BFQ-specific structures.
277 */
278
279static struct bfq_group *pd_to_bfqg(struct blkg_policy_data *pd)
280{
281 return pd ? container_of(pd, struct bfq_group, pd) : NULL;
282}
283
284struct blkcg_gq *bfqg_to_blkg(struct bfq_group *bfqg)
285{
286 return pd_to_blkg(&bfqg->pd);
287}
288
289static struct bfq_group *blkg_to_bfqg(struct blkcg_gq *blkg)
290{
291 return pd_to_bfqg(blkg_to_pd(blkg, &blkcg_policy_bfq));
292}
293
294/*
295 * bfq_group handlers
296 * The following functions help in navigating the bfq_group hierarchy
297 * by allowing to find the parent of a bfq_group or the bfq_group
298 * associated to a bfq_queue.
299 */
300
301static struct bfq_group *bfqg_parent(struct bfq_group *bfqg)
302{
303 struct blkcg_gq *pblkg = bfqg_to_blkg(bfqg)->parent;
304
305 return pblkg ? blkg_to_bfqg(pblkg) : NULL;
306}
307
308struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
309{
310 struct bfq_entity *group_entity = bfqq->entity.parent;
311
312 return group_entity ? container_of(group_entity, struct bfq_group,
313 entity) :
314 bfqq->bfqd->root_group;
315}
316
317/*
318 * The following two functions handle get and put of a bfq_group by
319 * wrapping the related blk-cgroup hooks.
320 */
321
322static void bfqg_get(struct bfq_group *bfqg)
323{
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200324 bfqg->ref++;
Paolo Valenteea25da42017-04-19 08:48:24 -0600325}
326
Bart Van Asschedfb79af2017-08-30 11:42:08 -0700327static void bfqg_put(struct bfq_group *bfqg)
Paolo Valenteea25da42017-04-19 08:48:24 -0600328{
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200329 bfqg->ref--;
330
331 if (bfqg->ref == 0)
332 kfree(bfqg);
333}
334
Dmitry Monakhov2de791a2020-08-11 06:43:40 +0000335static void bfqg_and_blkg_get(struct bfq_group *bfqg)
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200336{
337 /* see comments in bfq_bic_update_cgroup for why refcounting bfqg */
338 bfqg_get(bfqg);
339
340 blkg_get(bfqg_to_blkg(bfqg));
341}
342
343void bfqg_and_blkg_put(struct bfq_group *bfqg)
344{
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200345 blkg_put(bfqg_to_blkg(bfqg));
Konstantin Khlebnikovd5274b32018-09-06 11:05:44 +0300346
347 bfqg_put(bfqg);
Paolo Valenteea25da42017-04-19 08:48:24 -0600348}
349
Tejun Heofd41e602019-11-07 11:18:00 -0800350void bfqg_stats_update_legacy_io(struct request_queue *q, struct request *rq)
351{
352 struct bfq_group *bfqg = blkg_to_bfqg(rq->bio->bi_blkg);
353
Hou Tao08802ed2019-12-05 20:53:11 +0800354 if (!bfqg)
355 return;
356
Tejun Heofd41e602019-11-07 11:18:00 -0800357 blkg_rwstat_add(&bfqg->stats.bytes, rq->cmd_flags, blk_rq_bytes(rq));
358 blkg_rwstat_add(&bfqg->stats.ios, rq->cmd_flags, 1);
359}
360
Paolo Valenteea25da42017-04-19 08:48:24 -0600361/* @stats = 0 */
362static void bfqg_stats_reset(struct bfqg_stats *stats)
363{
Christoph Hellwig8060c472019-06-06 12:26:24 +0200364#ifdef CONFIG_BFQ_CGROUP_DEBUG
Paolo Valenteea25da42017-04-19 08:48:24 -0600365 /* queued stats shouldn't be cleared */
366 blkg_rwstat_reset(&stats->merged);
367 blkg_rwstat_reset(&stats->service_time);
368 blkg_rwstat_reset(&stats->wait_time);
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +0200369 bfq_stat_reset(&stats->time);
370 bfq_stat_reset(&stats->avg_queue_size_sum);
371 bfq_stat_reset(&stats->avg_queue_size_samples);
372 bfq_stat_reset(&stats->dequeue);
373 bfq_stat_reset(&stats->group_wait_time);
374 bfq_stat_reset(&stats->idle_time);
375 bfq_stat_reset(&stats->empty_time);
Luca Miccioa33801e2017-11-13 07:34:10 +0100376#endif
Paolo Valenteea25da42017-04-19 08:48:24 -0600377}
378
379/* @to += @from */
380static void bfqg_stats_add_aux(struct bfqg_stats *to, struct bfqg_stats *from)
381{
382 if (!to || !from)
383 return;
384
Christoph Hellwig8060c472019-06-06 12:26:24 +0200385#ifdef CONFIG_BFQ_CGROUP_DEBUG
Paolo Valenteea25da42017-04-19 08:48:24 -0600386 /* queued stats shouldn't be cleared */
387 blkg_rwstat_add_aux(&to->merged, &from->merged);
388 blkg_rwstat_add_aux(&to->service_time, &from->service_time);
389 blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +0200390 bfq_stat_add_aux(&from->time, &from->time);
391 bfq_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
392 bfq_stat_add_aux(&to->avg_queue_size_samples,
Paolo Valenteea25da42017-04-19 08:48:24 -0600393 &from->avg_queue_size_samples);
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +0200394 bfq_stat_add_aux(&to->dequeue, &from->dequeue);
395 bfq_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
396 bfq_stat_add_aux(&to->idle_time, &from->idle_time);
397 bfq_stat_add_aux(&to->empty_time, &from->empty_time);
Luca Miccioa33801e2017-11-13 07:34:10 +0100398#endif
Paolo Valenteea25da42017-04-19 08:48:24 -0600399}
400
401/*
402 * Transfer @bfqg's stats to its parent's aux counts so that the ancestors'
403 * recursive stats can still account for the amount used by this bfqg after
404 * it's gone.
405 */
406static void bfqg_stats_xfer_dead(struct bfq_group *bfqg)
407{
408 struct bfq_group *parent;
409
410 if (!bfqg) /* root_group */
411 return;
412
413 parent = bfqg_parent(bfqg);
414
Christoph Hellwig0d945c12018-11-15 12:17:28 -0700415 lockdep_assert_held(&bfqg_to_blkg(bfqg)->q->queue_lock);
Paolo Valenteea25da42017-04-19 08:48:24 -0600416
417 if (unlikely(!parent))
418 return;
419
420 bfqg_stats_add_aux(&parent->stats, &bfqg->stats);
421 bfqg_stats_reset(&bfqg->stats);
422}
423
424void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
425{
426 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
427
428 entity->weight = entity->new_weight;
429 entity->orig_weight = entity->new_weight;
430 if (bfqq) {
431 bfqq->ioprio = bfqq->new_ioprio;
432 bfqq->ioprio_class = bfqq->new_ioprio_class;
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200433 /*
434 * Make sure that bfqg and its associated blkg do not
435 * disappear before entity.
436 */
437 bfqg_and_blkg_get(bfqg);
Paolo Valenteea25da42017-04-19 08:48:24 -0600438 }
439 entity->parent = bfqg->my_entity; /* NULL for root group */
440 entity->sched_data = &bfqg->sched_data;
441}
442
443static void bfqg_stats_exit(struct bfqg_stats *stats)
444{
Tejun Heofd41e602019-11-07 11:18:00 -0800445 blkg_rwstat_exit(&stats->bytes);
446 blkg_rwstat_exit(&stats->ios);
Christoph Hellwig8060c472019-06-06 12:26:24 +0200447#ifdef CONFIG_BFQ_CGROUP_DEBUG
Paolo Valenteea25da42017-04-19 08:48:24 -0600448 blkg_rwstat_exit(&stats->merged);
449 blkg_rwstat_exit(&stats->service_time);
450 blkg_rwstat_exit(&stats->wait_time);
451 blkg_rwstat_exit(&stats->queued);
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +0200452 bfq_stat_exit(&stats->time);
453 bfq_stat_exit(&stats->avg_queue_size_sum);
454 bfq_stat_exit(&stats->avg_queue_size_samples);
455 bfq_stat_exit(&stats->dequeue);
456 bfq_stat_exit(&stats->group_wait_time);
457 bfq_stat_exit(&stats->idle_time);
458 bfq_stat_exit(&stats->empty_time);
Luca Miccioa33801e2017-11-13 07:34:10 +0100459#endif
Paolo Valenteea25da42017-04-19 08:48:24 -0600460}
461
462static int bfqg_stats_init(struct bfqg_stats *stats, gfp_t gfp)
463{
Tejun Heofd41e602019-11-07 11:18:00 -0800464 if (blkg_rwstat_init(&stats->bytes, gfp) ||
465 blkg_rwstat_init(&stats->ios, gfp))
Zheng Liang2fc428f2021-10-18 10:42:25 +0800466 goto error;
Tejun Heofd41e602019-11-07 11:18:00 -0800467
Christoph Hellwig8060c472019-06-06 12:26:24 +0200468#ifdef CONFIG_BFQ_CGROUP_DEBUG
Paolo Valenteea25da42017-04-19 08:48:24 -0600469 if (blkg_rwstat_init(&stats->merged, gfp) ||
470 blkg_rwstat_init(&stats->service_time, gfp) ||
471 blkg_rwstat_init(&stats->wait_time, gfp) ||
472 blkg_rwstat_init(&stats->queued, gfp) ||
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +0200473 bfq_stat_init(&stats->time, gfp) ||
474 bfq_stat_init(&stats->avg_queue_size_sum, gfp) ||
475 bfq_stat_init(&stats->avg_queue_size_samples, gfp) ||
476 bfq_stat_init(&stats->dequeue, gfp) ||
477 bfq_stat_init(&stats->group_wait_time, gfp) ||
478 bfq_stat_init(&stats->idle_time, gfp) ||
Zheng Liang2fc428f2021-10-18 10:42:25 +0800479 bfq_stat_init(&stats->empty_time, gfp))
480 goto error;
Luca Miccioa33801e2017-11-13 07:34:10 +0100481#endif
Paolo Valenteea25da42017-04-19 08:48:24 -0600482
483 return 0;
Zheng Liang2fc428f2021-10-18 10:42:25 +0800484
485error:
486 bfqg_stats_exit(stats);
487 return -ENOMEM;
Paolo Valenteea25da42017-04-19 08:48:24 -0600488}
489
490static struct bfq_group_data *cpd_to_bfqgd(struct blkcg_policy_data *cpd)
491{
492 return cpd ? container_of(cpd, struct bfq_group_data, pd) : NULL;
493}
494
495static struct bfq_group_data *blkcg_to_bfqgd(struct blkcg *blkcg)
496{
497 return cpd_to_bfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_bfq));
498}
499
Bart Van Asschedfb79af2017-08-30 11:42:08 -0700500static struct blkcg_policy_data *bfq_cpd_alloc(gfp_t gfp)
Paolo Valenteea25da42017-04-19 08:48:24 -0600501{
502 struct bfq_group_data *bgd;
503
504 bgd = kzalloc(sizeof(*bgd), gfp);
505 if (!bgd)
506 return NULL;
507 return &bgd->pd;
508}
509
Bart Van Asschedfb79af2017-08-30 11:42:08 -0700510static void bfq_cpd_init(struct blkcg_policy_data *cpd)
Paolo Valenteea25da42017-04-19 08:48:24 -0600511{
512 struct bfq_group_data *d = cpd_to_bfqgd(cpd);
513
514 d->weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
515 CGROUP_WEIGHT_DFL : BFQ_WEIGHT_LEGACY_DFL;
516}
517
Bart Van Asschedfb79af2017-08-30 11:42:08 -0700518static void bfq_cpd_free(struct blkcg_policy_data *cpd)
Paolo Valenteea25da42017-04-19 08:48:24 -0600519{
520 kfree(cpd_to_bfqgd(cpd));
521}
522
Tejun Heocf09a8e2019-08-28 15:05:51 -0700523static struct blkg_policy_data *bfq_pd_alloc(gfp_t gfp, struct request_queue *q,
524 struct blkcg *blkcg)
Paolo Valenteea25da42017-04-19 08:48:24 -0600525{
526 struct bfq_group *bfqg;
527
Tejun Heocf09a8e2019-08-28 15:05:51 -0700528 bfqg = kzalloc_node(sizeof(*bfqg), gfp, q->node);
Paolo Valenteea25da42017-04-19 08:48:24 -0600529 if (!bfqg)
530 return NULL;
531
532 if (bfqg_stats_init(&bfqg->stats, gfp)) {
533 kfree(bfqg);
534 return NULL;
535 }
536
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200537 /* see comments in bfq_bic_update_cgroup for why refcounting */
538 bfqg_get(bfqg);
Paolo Valenteea25da42017-04-19 08:48:24 -0600539 return &bfqg->pd;
540}
541
Bart Van Asschedfb79af2017-08-30 11:42:08 -0700542static void bfq_pd_init(struct blkg_policy_data *pd)
Paolo Valenteea25da42017-04-19 08:48:24 -0600543{
544 struct blkcg_gq *blkg = pd_to_blkg(pd);
545 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
546 struct bfq_data *bfqd = blkg->q->elevator->elevator_data;
547 struct bfq_entity *entity = &bfqg->entity;
548 struct bfq_group_data *d = blkcg_to_bfqgd(blkg->blkcg);
549
550 entity->orig_weight = entity->weight = entity->new_weight = d->weight;
551 entity->my_sched_data = &bfqg->sched_data;
Paolo Valente430a67f2021-03-04 18:46:27 +0100552 entity->last_bfqq_created = NULL;
553
Paolo Valenteea25da42017-04-19 08:48:24 -0600554 bfqg->my_entity = entity; /*
555 * the root_group's will be set to NULL
556 * in bfq_init_queue()
557 */
558 bfqg->bfqd = bfqd;
559 bfqg->active_entities = 0;
560 bfqg->rq_pos_tree = RB_ROOT;
561}
562
Bart Van Asschedfb79af2017-08-30 11:42:08 -0700563static void bfq_pd_free(struct blkg_policy_data *pd)
Paolo Valenteea25da42017-04-19 08:48:24 -0600564{
565 struct bfq_group *bfqg = pd_to_bfqg(pd);
566
567 bfqg_stats_exit(&bfqg->stats);
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200568 bfqg_put(bfqg);
Paolo Valenteea25da42017-04-19 08:48:24 -0600569}
570
Bart Van Asschedfb79af2017-08-30 11:42:08 -0700571static void bfq_pd_reset_stats(struct blkg_policy_data *pd)
Paolo Valenteea25da42017-04-19 08:48:24 -0600572{
573 struct bfq_group *bfqg = pd_to_bfqg(pd);
574
575 bfqg_stats_reset(&bfqg->stats);
576}
577
578static void bfq_group_set_parent(struct bfq_group *bfqg,
579 struct bfq_group *parent)
580{
581 struct bfq_entity *entity;
582
583 entity = &bfqg->entity;
584 entity->parent = parent->my_entity;
585 entity->sched_data = &parent->sched_data;
586}
587
588static struct bfq_group *bfq_lookup_bfqg(struct bfq_data *bfqd,
589 struct blkcg *blkcg)
590{
591 struct blkcg_gq *blkg;
592
593 blkg = blkg_lookup(blkcg, bfqd->queue);
594 if (likely(blkg))
595 return blkg_to_bfqg(blkg);
596 return NULL;
597}
598
599struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd,
600 struct blkcg *blkcg)
601{
602 struct bfq_group *bfqg, *parent;
603 struct bfq_entity *entity;
604
605 bfqg = bfq_lookup_bfqg(bfqd, blkcg);
606
607 if (unlikely(!bfqg))
608 return NULL;
609
610 /*
611 * Update chain of bfq_groups as we might be handling a leaf group
612 * which, along with some of its relatives, has not been hooked yet
613 * to the private hierarchy of BFQ.
614 */
615 entity = &bfqg->entity;
616 for_each_entity(entity) {
Carlo Nonato14afc592020-03-06 13:27:31 +0100617 struct bfq_group *curr_bfqg = container_of(entity,
618 struct bfq_group, entity);
619 if (curr_bfqg != bfqd->root_group) {
620 parent = bfqg_parent(curr_bfqg);
Paolo Valenteea25da42017-04-19 08:48:24 -0600621 if (!parent)
622 parent = bfqd->root_group;
Carlo Nonato14afc592020-03-06 13:27:31 +0100623 bfq_group_set_parent(curr_bfqg, parent);
Paolo Valenteea25da42017-04-19 08:48:24 -0600624 }
625 }
626
627 return bfqg;
628}
629
630/**
631 * bfq_bfqq_move - migrate @bfqq to @bfqg.
632 * @bfqd: queue descriptor.
633 * @bfqq: the queue to move.
634 * @bfqg: the group to move to.
635 *
636 * Move @bfqq to @bfqg, deactivating it from its old group and reactivating
637 * it on the new one. Avoid putting the entity on the old group idle tree.
638 *
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200639 * Must be called under the scheduler lock, to make sure that the blkg
640 * owning @bfqg does not disappear (see comments in
641 * bfq_bic_update_cgroup on guaranteeing the consistency of blkg
642 * objects).
Paolo Valenteea25da42017-04-19 08:48:24 -0600643 */
644void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
645 struct bfq_group *bfqg)
646{
647 struct bfq_entity *entity = &bfqq->entity;
648
Paolo Valentefd1bb3a2020-03-21 10:45:18 +0100649 /*
650 * Get extra reference to prevent bfqq from being freed in
651 * next possible expire or deactivate.
652 */
653 bfqq->ref++;
654
Paolo Valenteea25da42017-04-19 08:48:24 -0600655 /* If bfqq is empty, then bfq_bfqq_expire also invokes
656 * bfq_del_bfqq_busy, thereby removing bfqq and its entity
657 * from data structures related to current group. Otherwise we
658 * need to remove bfqq explicitly with bfq_deactivate_bfqq, as
659 * we do below.
660 */
661 if (bfqq == bfqd->in_service_queue)
662 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
663 false, BFQQE_PREEMPTED);
664
665 if (bfq_bfqq_busy(bfqq))
666 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
Paolo Valente33a16a92020-02-03 11:40:57 +0100667 else if (entity->on_st_or_in_serv)
Paolo Valenteea25da42017-04-19 08:48:24 -0600668 bfq_put_idle_entity(bfq_entity_service_tree(entity), entity);
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200669 bfqg_and_blkg_put(bfqq_group(bfqq));
Paolo Valenteea25da42017-04-19 08:48:24 -0600670
Paolo Valented29bd412021-10-15 16:43:36 +0200671 if (entity->parent &&
672 entity->parent->last_bfqq_created == bfqq)
673 entity->parent->last_bfqq_created = NULL;
674 else if (bfqd->last_bfqq_created == bfqq)
675 bfqd->last_bfqq_created = NULL;
676
Paolo Valenteea25da42017-04-19 08:48:24 -0600677 entity->parent = bfqg->my_entity;
678 entity->sched_data = &bfqg->sched_data;
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200679 /* pin down bfqg and its associated blkg */
680 bfqg_and_blkg_get(bfqg);
Paolo Valenteea25da42017-04-19 08:48:24 -0600681
682 if (bfq_bfqq_busy(bfqq)) {
Paolo Valente8cacc5a2019-03-12 09:59:30 +0100683 if (unlikely(!bfqd->nonrot_with_queueing))
684 bfq_pos_tree_add_move(bfqd, bfqq);
Paolo Valenteea25da42017-04-19 08:48:24 -0600685 bfq_activate_bfqq(bfqd, bfqq);
686 }
687
688 if (!bfqd->in_service_queue && !bfqd->rq_in_driver)
689 bfq_schedule_dispatch(bfqd);
Paolo Valentefd1bb3a2020-03-21 10:45:18 +0100690 /* release extra ref taken above, bfqq may happen to be freed now */
Paolo Valenteecedd3d2020-02-03 11:40:56 +0100691 bfq_put_queue(bfqq);
Paolo Valenteea25da42017-04-19 08:48:24 -0600692}
693
694/**
695 * __bfq_bic_change_cgroup - move @bic to @cgroup.
696 * @bfqd: the queue descriptor.
697 * @bic: the bic to move.
698 * @blkcg: the blk-cgroup to move to.
699 *
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200700 * Move bic to blkcg, assuming that bfqd->lock is held; which makes
701 * sure that the reference to cgroup is valid across the call (see
702 * comments in bfq_bic_update_cgroup on this issue)
Paolo Valenteea25da42017-04-19 08:48:24 -0600703 *
704 * NOTE: an alternative approach might have been to store the current
705 * cgroup in bfqq and getting a reference to it, reducing the lookup
706 * time here, at the price of slightly more complex code.
707 */
708static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd,
709 struct bfq_io_cq *bic,
710 struct blkcg *blkcg)
711{
712 struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0);
713 struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1);
714 struct bfq_group *bfqg;
715 struct bfq_entity *entity;
716
717 bfqg = bfq_find_set_group(bfqd, blkcg);
718
719 if (unlikely(!bfqg))
720 bfqg = bfqd->root_group;
721
722 if (async_bfqq) {
723 entity = &async_bfqq->entity;
724
725 if (entity->sched_data != &bfqg->sched_data) {
726 bic_set_bfqq(bic, NULL, 0);
Paolo Valentec8997732020-03-21 10:45:19 +0100727 bfq_release_process_ref(bfqd, async_bfqq);
Paolo Valenteea25da42017-04-19 08:48:24 -0600728 }
729 }
730
731 if (sync_bfqq) {
732 entity = &sync_bfqq->entity;
733 if (entity->sched_data != &bfqg->sched_data)
734 bfq_bfqq_move(bfqd, sync_bfqq, bfqg);
735 }
736
737 return bfqg;
738}
739
740void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio)
741{
742 struct bfq_data *bfqd = bic_to_bfqd(bic);
743 struct bfq_group *bfqg = NULL;
744 uint64_t serial_nr;
745
746 rcu_read_lock();
Dennis Zhou0fe061b2018-12-05 12:10:26 -0500747 serial_nr = __bio_blkcg(bio)->css.serial_nr;
Paolo Valenteea25da42017-04-19 08:48:24 -0600748
749 /*
750 * Check whether blkcg has changed. The condition may trigger
751 * spuriously on a newly created cic but there's no harm.
752 */
753 if (unlikely(!bfqd) || likely(bic->blkcg_serial_nr == serial_nr))
754 goto out;
755
Dennis Zhou0fe061b2018-12-05 12:10:26 -0500756 bfqg = __bfq_bic_change_cgroup(bfqd, bic, __bio_blkcg(bio));
Paolo Valente8f9bebc2017-06-05 10:11:15 +0200757 /*
758 * Update blkg_path for bfq_log_* functions. We cache this
759 * path, and update it here, for the following
760 * reasons. Operations on blkg objects in blk-cgroup are
761 * protected with the request_queue lock, and not with the
762 * lock that protects the instances of this scheduler
763 * (bfqd->lock). This exposes BFQ to the following sort of
764 * race.
765 *
766 * The blkg_lookup performed in bfq_get_queue, protected
767 * through rcu, may happen to return the address of a copy of
768 * the original blkg. If this is the case, then the
769 * bfqg_and_blkg_get performed in bfq_get_queue, to pin down
770 * the blkg, is useless: it does not prevent blk-cgroup code
771 * from destroying both the original blkg and all objects
772 * directly or indirectly referred by the copy of the
773 * blkg.
774 *
775 * On the bright side, destroy operations on a blkg invoke, as
776 * a first step, hooks of the scheduler associated with the
777 * blkg. And these hooks are executed with bfqd->lock held for
778 * BFQ. As a consequence, for any blkg associated with the
779 * request queue this instance of the scheduler is attached
780 * to, we are guaranteed that such a blkg is not destroyed, and
781 * that all the pointers it contains are consistent, while we
782 * are holding bfqd->lock. A blkg_lookup performed with
783 * bfqd->lock held then returns a fully consistent blkg, which
784 * remains consistent until this lock is held.
785 *
786 * Thanks to the last fact, and to the fact that: (1) bfqg has
787 * been obtained through a blkg_lookup in the above
788 * assignment, and (2) bfqd->lock is being held, here we can
789 * safely use the policy data for the involved blkg (i.e., the
790 * field bfqg->pd) to get to the blkg associated with bfqg,
791 * and then we can safely use any field of blkg. After we
792 * release bfqd->lock, even just getting blkg through this
793 * bfqg may cause dangling references to be traversed, as
794 * bfqg->pd may not exist any more.
795 *
796 * In view of the above facts, here we cache, in the bfqg, any
797 * blkg data we may need for this bic, and for its associated
798 * bfq_queue. As of now, we need to cache only the path of the
799 * blkg, which is used in the bfq_log_* functions.
800 *
801 * Finally, note that bfqg itself needs to be protected from
802 * destruction on the blkg_free of the original blkg (which
803 * invokes bfq_pd_free). We use an additional private
804 * refcounter for bfqg, to let it disappear only after no
805 * bfq_queue refers to it any longer.
806 */
807 blkg_path(bfqg_to_blkg(bfqg), bfqg->blkg_path, sizeof(bfqg->blkg_path));
Paolo Valenteea25da42017-04-19 08:48:24 -0600808 bic->blkcg_serial_nr = serial_nr;
809out:
810 rcu_read_unlock();
811}
812
813/**
814 * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st.
815 * @st: the service tree being flushed.
816 */
817static void bfq_flush_idle_tree(struct bfq_service_tree *st)
818{
819 struct bfq_entity *entity = st->first_idle;
820
821 for (; entity ; entity = st->first_idle)
822 __bfq_deactivate_entity(entity, false);
823}
824
825/**
826 * bfq_reparent_leaf_entity - move leaf entity to the root_group.
827 * @bfqd: the device data structure with the root group.
Paolo Valente576682f2020-03-21 10:45:20 +0100828 * @entity: the entity to move, if entity is a leaf; or the parent entity
829 * of an active leaf entity to move, if entity is not a leaf.
Paolo Valenteea25da42017-04-19 08:48:24 -0600830 */
831static void bfq_reparent_leaf_entity(struct bfq_data *bfqd,
Paolo Valente576682f2020-03-21 10:45:20 +0100832 struct bfq_entity *entity,
833 int ioprio_class)
Paolo Valenteea25da42017-04-19 08:48:24 -0600834{
Paolo Valente576682f2020-03-21 10:45:20 +0100835 struct bfq_queue *bfqq;
836 struct bfq_entity *child_entity = entity;
Paolo Valenteea25da42017-04-19 08:48:24 -0600837
Paolo Valente576682f2020-03-21 10:45:20 +0100838 while (child_entity->my_sched_data) { /* leaf not reached yet */
839 struct bfq_sched_data *child_sd = child_entity->my_sched_data;
840 struct bfq_service_tree *child_st = child_sd->service_tree +
841 ioprio_class;
842 struct rb_root *child_active = &child_st->active;
843
844 child_entity = bfq_entity_of(rb_first(child_active));
845
846 if (!child_entity)
847 child_entity = child_sd->in_service_entity;
848 }
849
850 bfqq = bfq_entity_to_bfqq(child_entity);
Paolo Valenteea25da42017-04-19 08:48:24 -0600851 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
852}
853
854/**
Paolo Valente576682f2020-03-21 10:45:20 +0100855 * bfq_reparent_active_queues - move to the root group all active queues.
Paolo Valenteea25da42017-04-19 08:48:24 -0600856 * @bfqd: the device data structure with the root group.
857 * @bfqg: the group to move from.
Paolo Valente576682f2020-03-21 10:45:20 +0100858 * @st: the service tree to start the search from.
Paolo Valenteea25da42017-04-19 08:48:24 -0600859 */
Paolo Valente576682f2020-03-21 10:45:20 +0100860static void bfq_reparent_active_queues(struct bfq_data *bfqd,
861 struct bfq_group *bfqg,
862 struct bfq_service_tree *st,
863 int ioprio_class)
Paolo Valenteea25da42017-04-19 08:48:24 -0600864{
865 struct rb_root *active = &st->active;
Paolo Valente576682f2020-03-21 10:45:20 +0100866 struct bfq_entity *entity;
Paolo Valenteea25da42017-04-19 08:48:24 -0600867
Paolo Valente576682f2020-03-21 10:45:20 +0100868 while ((entity = bfq_entity_of(rb_first(active))))
869 bfq_reparent_leaf_entity(bfqd, entity, ioprio_class);
Paolo Valenteea25da42017-04-19 08:48:24 -0600870
871 if (bfqg->sched_data.in_service_entity)
872 bfq_reparent_leaf_entity(bfqd,
Paolo Valente576682f2020-03-21 10:45:20 +0100873 bfqg->sched_data.in_service_entity,
874 ioprio_class);
Paolo Valenteea25da42017-04-19 08:48:24 -0600875}
876
877/**
878 * bfq_pd_offline - deactivate the entity associated with @pd,
879 * and reparent its children entities.
880 * @pd: descriptor of the policy going offline.
881 *
882 * blkio already grabs the queue_lock for us, so no need to use
883 * RCU-based magic
884 */
Bart Van Asschedfb79af2017-08-30 11:42:08 -0700885static void bfq_pd_offline(struct blkg_policy_data *pd)
Paolo Valenteea25da42017-04-19 08:48:24 -0600886{
887 struct bfq_service_tree *st;
888 struct bfq_group *bfqg = pd_to_bfqg(pd);
889 struct bfq_data *bfqd = bfqg->bfqd;
890 struct bfq_entity *entity = bfqg->my_entity;
891 unsigned long flags;
892 int i;
893
Paolo Valenteea25da42017-04-19 08:48:24 -0600894 spin_lock_irqsave(&bfqd->lock, flags);
Paolo Valente52257ff2018-01-09 10:27:58 +0100895
896 if (!entity) /* root group */
897 goto put_async_queues;
898
Paolo Valenteea25da42017-04-19 08:48:24 -0600899 /*
900 * Empty all service_trees belonging to this group before
901 * deactivating the group itself.
902 */
903 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) {
904 st = bfqg->sched_data.service_tree + i;
905
906 /*
Paolo Valenteea25da42017-04-19 08:48:24 -0600907 * It may happen that some queues are still active
908 * (busy) upon group destruction (if the corresponding
909 * processes have been forced to terminate). We move
910 * all the leaf entities corresponding to these queues
911 * to the root_group.
912 * Also, it may happen that the group has an entity
913 * in service, which is disconnected from the active
914 * tree: it must be moved, too.
915 * There is no need to put the sync queues, as the
916 * scheduler has taken no reference.
917 */
Paolo Valente576682f2020-03-21 10:45:20 +0100918 bfq_reparent_active_queues(bfqd, bfqg, st, i);
Paolo Valente4d38a872020-03-21 10:45:21 +0100919
920 /*
921 * The idle tree may still contain bfq_queues
922 * belonging to exited task because they never
923 * migrated to a different cgroup from the one being
924 * destroyed now. In addition, even
925 * bfq_reparent_active_queues() may happen to add some
926 * entities to the idle tree. It happens if, in some
927 * of the calls to bfq_bfqq_move() performed by
928 * bfq_reparent_active_queues(), the queue to move is
929 * empty and gets expired.
930 */
931 bfq_flush_idle_tree(st);
Paolo Valenteea25da42017-04-19 08:48:24 -0600932 }
933
934 __bfq_deactivate_entity(entity, false);
Paolo Valente52257ff2018-01-09 10:27:58 +0100935
936put_async_queues:
Paolo Valenteea25da42017-04-19 08:48:24 -0600937 bfq_put_async_queues(bfqd, bfqg);
938
939 spin_unlock_irqrestore(&bfqd->lock, flags);
940 /*
941 * @blkg is going offline and will be ignored by
942 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
943 * that they don't get lost. If IOs complete after this point, the
944 * stats for them will be lost. Oh well...
945 */
946 bfqg_stats_xfer_dead(bfqg);
947}
948
949void bfq_end_wr_async(struct bfq_data *bfqd)
950{
951 struct blkcg_gq *blkg;
952
953 list_for_each_entry(blkg, &bfqd->queue->blkg_list, q_node) {
954 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
955
956 bfq_end_wr_async_queues(bfqd, bfqg);
957 }
958 bfq_end_wr_async_queues(bfqd, bfqd->root_group);
959}
960
Fam Zheng795fe542019-08-28 11:54:53 +0800961static int bfq_io_show_weight_legacy(struct seq_file *sf, void *v)
Paolo Valenteea25da42017-04-19 08:48:24 -0600962{
963 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
964 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
965 unsigned int val = 0;
966
967 if (bfqgd)
968 val = bfqgd->weight;
969
970 seq_printf(sf, "%u\n", val);
971
972 return 0;
973}
974
Fam Zheng795fe542019-08-28 11:54:53 +0800975static u64 bfqg_prfill_weight_device(struct seq_file *sf,
976 struct blkg_policy_data *pd, int off)
Fam Zheng5ff047e2019-08-28 11:54:52 +0800977{
Fam Zheng795fe542019-08-28 11:54:53 +0800978 struct bfq_group *bfqg = pd_to_bfqg(pd);
979
980 if (!bfqg->entity.dev_weight)
981 return 0;
982 return __blkg_prfill_u64(sf, pd, bfqg->entity.dev_weight);
983}
984
985static int bfq_io_show_weight(struct seq_file *sf, void *v)
986{
987 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
988 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
989
990 seq_printf(sf, "default %u\n", bfqgd->weight);
991 blkcg_print_blkgs(sf, blkcg, bfqg_prfill_weight_device,
992 &blkcg_policy_bfq, 0, false);
993 return 0;
994}
995
996static void bfq_group_set_weight(struct bfq_group *bfqg, u64 weight, u64 dev_weight)
997{
998 weight = dev_weight ?: weight;
999
1000 bfqg->entity.dev_weight = dev_weight;
Fam Zheng5ff047e2019-08-28 11:54:52 +08001001 /*
1002 * Setting the prio_changed flag of the entity
1003 * to 1 with new_weight == weight would re-set
1004 * the value of the weight to its ioprio mapping.
1005 * Set the flag only if necessary.
1006 */
1007 if ((unsigned short)weight != bfqg->entity.new_weight) {
1008 bfqg->entity.new_weight = (unsigned short)weight;
1009 /*
1010 * Make sure that the above new value has been
1011 * stored in bfqg->entity.new_weight before
1012 * setting the prio_changed flag. In fact,
1013 * this flag may be read asynchronously (in
1014 * critical sections protected by a different
1015 * lock than that held here), and finding this
1016 * flag set may cause the execution of the code
1017 * for updating parameters whose value may
1018 * depend also on bfqg->entity.new_weight (in
1019 * __bfq_entity_update_weight_prio).
1020 * This barrier makes sure that the new value
1021 * of bfqg->entity.new_weight is correctly
1022 * seen in that code.
1023 */
1024 smp_wmb();
1025 bfqg->entity.prio_changed = 1;
1026 }
1027}
1028
Paolo Valenteea25da42017-04-19 08:48:24 -06001029static int bfq_io_set_weight_legacy(struct cgroup_subsys_state *css,
1030 struct cftype *cftype,
1031 u64 val)
1032{
1033 struct blkcg *blkcg = css_to_blkcg(css);
1034 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
1035 struct blkcg_gq *blkg;
1036 int ret = -ERANGE;
1037
1038 if (val < BFQ_MIN_WEIGHT || val > BFQ_MAX_WEIGHT)
1039 return ret;
1040
1041 ret = 0;
1042 spin_lock_irq(&blkcg->lock);
1043 bfqgd->weight = (unsigned short)val;
1044 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1045 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
1046
Fam Zheng5ff047e2019-08-28 11:54:52 +08001047 if (bfqg)
Fam Zheng795fe542019-08-28 11:54:53 +08001048 bfq_group_set_weight(bfqg, val, 0);
Paolo Valenteea25da42017-04-19 08:48:24 -06001049 }
1050 spin_unlock_irq(&blkcg->lock);
1051
1052 return ret;
1053}
1054
Fam Zheng795fe542019-08-28 11:54:53 +08001055static ssize_t bfq_io_set_device_weight(struct kernfs_open_file *of,
1056 char *buf, size_t nbytes,
1057 loff_t off)
1058{
1059 int ret;
1060 struct blkg_conf_ctx ctx;
1061 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1062 struct bfq_group *bfqg;
1063 u64 v;
1064
1065 ret = blkg_conf_prep(blkcg, &blkcg_policy_bfq, buf, &ctx);
1066 if (ret)
1067 return ret;
1068
1069 if (sscanf(ctx.body, "%llu", &v) == 1) {
1070 /* require "default" on dfl */
1071 ret = -ERANGE;
1072 if (!v)
1073 goto out;
1074 } else if (!strcmp(strim(ctx.body), "default")) {
1075 v = 0;
1076 } else {
1077 ret = -EINVAL;
1078 goto out;
1079 }
1080
1081 bfqg = blkg_to_bfqg(ctx.blkg);
1082
1083 ret = -ERANGE;
1084 if (!v || (v >= BFQ_MIN_WEIGHT && v <= BFQ_MAX_WEIGHT)) {
1085 bfq_group_set_weight(bfqg, bfqg->entity.weight, v);
1086 ret = 0;
1087 }
1088out:
1089 blkg_conf_finish(&ctx);
1090 return ret ?: nbytes;
1091}
1092
Paolo Valenteea25da42017-04-19 08:48:24 -06001093static ssize_t bfq_io_set_weight(struct kernfs_open_file *of,
1094 char *buf, size_t nbytes,
1095 loff_t off)
1096{
Fam Zheng795fe542019-08-28 11:54:53 +08001097 char *endp;
1098 int ret;
1099 u64 v;
Paolo Valenteea25da42017-04-19 08:48:24 -06001100
Fam Zheng795fe542019-08-28 11:54:53 +08001101 buf = strim(buf);
Paolo Valenteea25da42017-04-19 08:48:24 -06001102
Fam Zheng795fe542019-08-28 11:54:53 +08001103 /* "WEIGHT" or "default WEIGHT" sets the default weight */
1104 v = simple_strtoull(buf, &endp, 0);
1105 if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
1106 ret = bfq_io_set_weight_legacy(of_css(of), NULL, v);
1107 return ret ?: nbytes;
1108 }
1109
1110 return bfq_io_set_device_weight(of, buf, nbytes, off);
Paolo Valenteea25da42017-04-19 08:48:24 -06001111}
1112
Paolo Valenteea25da42017-04-19 08:48:24 -06001113static int bfqg_print_rwstat(struct seq_file *sf, void *v)
1114{
1115 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1116 &blkcg_policy_bfq, seq_cft(sf)->private, true);
1117 return 0;
1118}
1119
Tejun Heoa557f1c2019-11-07 11:17:59 -08001120static u64 bfqg_prfill_rwstat_recursive(struct seq_file *sf,
1121 struct blkg_policy_data *pd, int off)
1122{
1123 struct blkg_rwstat_sample sum;
1124
1125 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_bfq, off, &sum);
1126 return __blkg_prfill_rwstat(sf, pd, &sum);
1127}
1128
1129static int bfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1130{
1131 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1132 bfqg_prfill_rwstat_recursive, &blkcg_policy_bfq,
1133 seq_cft(sf)->private, true);
1134 return 0;
1135}
1136
Tejun Heofd41e602019-11-07 11:18:00 -08001137#ifdef CONFIG_BFQ_CGROUP_DEBUG
Tejun Heoa557f1c2019-11-07 11:17:59 -08001138static int bfqg_print_stat(struct seq_file *sf, void *v)
1139{
1140 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1141 &blkcg_policy_bfq, seq_cft(sf)->private, false);
1142 return 0;
1143}
1144
Paolo Valenteea25da42017-04-19 08:48:24 -06001145static u64 bfqg_prfill_stat_recursive(struct seq_file *sf,
1146 struct blkg_policy_data *pd, int off)
1147{
Christoph Hellwigd6258982019-06-06 12:26:23 +02001148 struct blkcg_gq *blkg = pd_to_blkg(pd);
1149 struct blkcg_gq *pos_blkg;
1150 struct cgroup_subsys_state *pos_css;
1151 u64 sum = 0;
1152
1153 lockdep_assert_held(&blkg->q->queue_lock);
1154
1155 rcu_read_lock();
1156 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
1157 struct bfq_stat *stat;
1158
1159 if (!pos_blkg->online)
1160 continue;
1161
1162 stat = (void *)blkg_to_pd(pos_blkg, &blkcg_policy_bfq) + off;
1163 sum += bfq_stat_read(stat) + atomic64_read(&stat->aux_cnt);
1164 }
1165 rcu_read_unlock();
1166
Paolo Valenteea25da42017-04-19 08:48:24 -06001167 return __blkg_prfill_u64(sf, pd, sum);
1168}
1169
Paolo Valenteea25da42017-04-19 08:48:24 -06001170static int bfqg_print_stat_recursive(struct seq_file *sf, void *v)
1171{
1172 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1173 bfqg_prfill_stat_recursive, &blkcg_policy_bfq,
1174 seq_cft(sf)->private, false);
1175 return 0;
1176}
1177
Paolo Valenteea25da42017-04-19 08:48:24 -06001178static u64 bfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1179 int off)
1180{
Tejun Heofd41e602019-11-07 11:18:00 -08001181 struct bfq_group *bfqg = blkg_to_bfqg(pd->blkg);
1182 u64 sum = blkg_rwstat_total(&bfqg->stats.bytes);
Paolo Valenteea25da42017-04-19 08:48:24 -06001183
1184 return __blkg_prfill_u64(sf, pd, sum >> 9);
1185}
1186
1187static int bfqg_print_stat_sectors(struct seq_file *sf, void *v)
1188{
1189 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1190 bfqg_prfill_sectors, &blkcg_policy_bfq, 0, false);
1191 return 0;
1192}
1193
1194static u64 bfqg_prfill_sectors_recursive(struct seq_file *sf,
1195 struct blkg_policy_data *pd, int off)
1196{
Christoph Hellwig7af6fd92019-06-06 12:26:21 +02001197 struct blkg_rwstat_sample tmp;
Christoph Hellwig5d0b6e42019-06-06 12:26:20 +02001198
Tejun Heofd41e602019-11-07 11:18:00 -08001199 blkg_rwstat_recursive_sum(pd->blkg, &blkcg_policy_bfq,
1200 offsetof(struct bfq_group, stats.bytes), &tmp);
Paolo Valenteea25da42017-04-19 08:48:24 -06001201
Christoph Hellwig7af6fd92019-06-06 12:26:21 +02001202 return __blkg_prfill_u64(sf, pd,
1203 (tmp.cnt[BLKG_RWSTAT_READ] + tmp.cnt[BLKG_RWSTAT_WRITE]) >> 9);
Paolo Valenteea25da42017-04-19 08:48:24 -06001204}
1205
1206static int bfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1207{
1208 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1209 bfqg_prfill_sectors_recursive, &blkcg_policy_bfq, 0,
1210 false);
1211 return 0;
1212}
1213
1214static u64 bfqg_prfill_avg_queue_size(struct seq_file *sf,
1215 struct blkg_policy_data *pd, int off)
1216{
1217 struct bfq_group *bfqg = pd_to_bfqg(pd);
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +02001218 u64 samples = bfq_stat_read(&bfqg->stats.avg_queue_size_samples);
Paolo Valenteea25da42017-04-19 08:48:24 -06001219 u64 v = 0;
1220
1221 if (samples) {
Christoph Hellwigc0ce79dc2019-06-06 12:26:22 +02001222 v = bfq_stat_read(&bfqg->stats.avg_queue_size_sum);
Paolo Valenteea25da42017-04-19 08:48:24 -06001223 v = div64_u64(v, samples);
1224 }
1225 __blkg_prfill_u64(sf, pd, v);
1226 return 0;
1227}
1228
1229/* print avg_queue_size */
1230static int bfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1231{
1232 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1233 bfqg_prfill_avg_queue_size, &blkcg_policy_bfq,
1234 0, false);
1235 return 0;
1236}
Christoph Hellwig8060c472019-06-06 12:26:24 +02001237#endif /* CONFIG_BFQ_CGROUP_DEBUG */
Paolo Valenteea25da42017-04-19 08:48:24 -06001238
1239struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1240{
1241 int ret;
1242
1243 ret = blkcg_activate_policy(bfqd->queue, &blkcg_policy_bfq);
1244 if (ret)
1245 return NULL;
1246
1247 return blkg_to_bfqg(bfqd->queue->root_blkg);
1248}
1249
1250struct blkcg_policy blkcg_policy_bfq = {
1251 .dfl_cftypes = bfq_blkg_files,
1252 .legacy_cftypes = bfq_blkcg_legacy_files,
1253
1254 .cpd_alloc_fn = bfq_cpd_alloc,
1255 .cpd_init_fn = bfq_cpd_init,
1256 .cpd_bind_fn = bfq_cpd_init,
1257 .cpd_free_fn = bfq_cpd_free,
1258
1259 .pd_alloc_fn = bfq_pd_alloc,
1260 .pd_init_fn = bfq_pd_init,
1261 .pd_offline_fn = bfq_pd_offline,
1262 .pd_free_fn = bfq_pd_free,
1263 .pd_reset_stats_fn = bfq_pd_reset_stats,
1264};
1265
1266struct cftype bfq_blkcg_legacy_files[] = {
1267 {
1268 .name = "bfq.weight",
Jens Axboecf892982019-06-10 03:35:41 -06001269 .flags = CFTYPE_NOT_ON_ROOT,
Fam Zheng795fe542019-08-28 11:54:53 +08001270 .seq_show = bfq_io_show_weight_legacy,
Paolo Valenteea25da42017-04-19 08:48:24 -06001271 .write_u64 = bfq_io_set_weight_legacy,
1272 },
Fam Zheng795fe542019-08-28 11:54:53 +08001273 {
1274 .name = "bfq.weight_device",
1275 .flags = CFTYPE_NOT_ON_ROOT,
1276 .seq_show = bfq_io_show_weight,
1277 .write = bfq_io_set_weight,
1278 },
Paolo Valenteea25da42017-04-19 08:48:24 -06001279
1280 /* statistics, covers only the tasks in the bfqg */
1281 {
Paolo Valenteea25da42017-04-19 08:48:24 -06001282 .name = "bfq.io_service_bytes",
Tejun Heofd41e602019-11-07 11:18:00 -08001283 .private = offsetof(struct bfq_group, stats.bytes),
1284 .seq_show = bfqg_print_rwstat,
Paolo Valenteea25da42017-04-19 08:48:24 -06001285 },
1286 {
1287 .name = "bfq.io_serviced",
Tejun Heofd41e602019-11-07 11:18:00 -08001288 .private = offsetof(struct bfq_group, stats.ios),
1289 .seq_show = bfqg_print_rwstat,
Paolo Valenteea25da42017-04-19 08:48:24 -06001290 },
Christoph Hellwig8060c472019-06-06 12:26:24 +02001291#ifdef CONFIG_BFQ_CGROUP_DEBUG
Luca Miccioa33801e2017-11-13 07:34:10 +01001292 {
1293 .name = "bfq.time",
1294 .private = offsetof(struct bfq_group, stats.time),
1295 .seq_show = bfqg_print_stat,
1296 },
1297 {
1298 .name = "bfq.sectors",
1299 .seq_show = bfqg_print_stat_sectors,
1300 },
Paolo Valenteea25da42017-04-19 08:48:24 -06001301 {
1302 .name = "bfq.io_service_time",
1303 .private = offsetof(struct bfq_group, stats.service_time),
1304 .seq_show = bfqg_print_rwstat,
1305 },
1306 {
1307 .name = "bfq.io_wait_time",
1308 .private = offsetof(struct bfq_group, stats.wait_time),
1309 .seq_show = bfqg_print_rwstat,
1310 },
1311 {
1312 .name = "bfq.io_merged",
1313 .private = offsetof(struct bfq_group, stats.merged),
1314 .seq_show = bfqg_print_rwstat,
1315 },
1316 {
1317 .name = "bfq.io_queued",
1318 .private = offsetof(struct bfq_group, stats.queued),
1319 .seq_show = bfqg_print_rwstat,
1320 },
Christoph Hellwig8060c472019-06-06 12:26:24 +02001321#endif /* CONFIG_BFQ_CGROUP_DEBUG */
Paolo Valenteea25da42017-04-19 08:48:24 -06001322
Angelo Ruocco636b8fe2019-04-08 17:35:34 +02001323 /* the same statistics which cover the bfqg and its descendants */
Paolo Valenteea25da42017-04-19 08:48:24 -06001324 {
Paolo Valenteea25da42017-04-19 08:48:24 -06001325 .name = "bfq.io_service_bytes_recursive",
Tejun Heofd41e602019-11-07 11:18:00 -08001326 .private = offsetof(struct bfq_group, stats.bytes),
1327 .seq_show = bfqg_print_rwstat_recursive,
Paolo Valenteea25da42017-04-19 08:48:24 -06001328 },
1329 {
1330 .name = "bfq.io_serviced_recursive",
Tejun Heofd41e602019-11-07 11:18:00 -08001331 .private = offsetof(struct bfq_group, stats.ios),
1332 .seq_show = bfqg_print_rwstat_recursive,
Paolo Valenteea25da42017-04-19 08:48:24 -06001333 },
Christoph Hellwig8060c472019-06-06 12:26:24 +02001334#ifdef CONFIG_BFQ_CGROUP_DEBUG
Luca Miccioa33801e2017-11-13 07:34:10 +01001335 {
1336 .name = "bfq.time_recursive",
1337 .private = offsetof(struct bfq_group, stats.time),
1338 .seq_show = bfqg_print_stat_recursive,
1339 },
1340 {
1341 .name = "bfq.sectors_recursive",
1342 .seq_show = bfqg_print_stat_sectors_recursive,
1343 },
Paolo Valenteea25da42017-04-19 08:48:24 -06001344 {
1345 .name = "bfq.io_service_time_recursive",
1346 .private = offsetof(struct bfq_group, stats.service_time),
1347 .seq_show = bfqg_print_rwstat_recursive,
1348 },
1349 {
1350 .name = "bfq.io_wait_time_recursive",
1351 .private = offsetof(struct bfq_group, stats.wait_time),
1352 .seq_show = bfqg_print_rwstat_recursive,
1353 },
1354 {
1355 .name = "bfq.io_merged_recursive",
1356 .private = offsetof(struct bfq_group, stats.merged),
1357 .seq_show = bfqg_print_rwstat_recursive,
1358 },
1359 {
1360 .name = "bfq.io_queued_recursive",
1361 .private = offsetof(struct bfq_group, stats.queued),
1362 .seq_show = bfqg_print_rwstat_recursive,
1363 },
1364 {
1365 .name = "bfq.avg_queue_size",
1366 .seq_show = bfqg_print_avg_queue_size,
1367 },
1368 {
1369 .name = "bfq.group_wait_time",
1370 .private = offsetof(struct bfq_group, stats.group_wait_time),
1371 .seq_show = bfqg_print_stat,
1372 },
1373 {
1374 .name = "bfq.idle_time",
1375 .private = offsetof(struct bfq_group, stats.idle_time),
1376 .seq_show = bfqg_print_stat,
1377 },
1378 {
1379 .name = "bfq.empty_time",
1380 .private = offsetof(struct bfq_group, stats.empty_time),
1381 .seq_show = bfqg_print_stat,
1382 },
1383 {
1384 .name = "bfq.dequeue",
1385 .private = offsetof(struct bfq_group, stats.dequeue),
1386 .seq_show = bfqg_print_stat,
1387 },
Christoph Hellwig8060c472019-06-06 12:26:24 +02001388#endif /* CONFIG_BFQ_CGROUP_DEBUG */
Paolo Valenteea25da42017-04-19 08:48:24 -06001389 { } /* terminate */
1390};
1391
1392struct cftype bfq_blkg_files[] = {
1393 {
1394 .name = "bfq.weight",
Jens Axboecf892982019-06-10 03:35:41 -06001395 .flags = CFTYPE_NOT_ON_ROOT,
Paolo Valenteea25da42017-04-19 08:48:24 -06001396 .seq_show = bfq_io_show_weight,
1397 .write = bfq_io_set_weight,
1398 },
1399 {} /* terminate */
1400};
1401
1402#else /* CONFIG_BFQ_GROUP_IOSCHED */
1403
Paolo Valenteea25da42017-04-19 08:48:24 -06001404void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1405 struct bfq_group *bfqg) {}
1406
1407void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
1408{
1409 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
1410
1411 entity->weight = entity->new_weight;
1412 entity->orig_weight = entity->new_weight;
1413 if (bfqq) {
1414 bfqq->ioprio = bfqq->new_ioprio;
1415 bfqq->ioprio_class = bfqq->new_ioprio_class;
1416 }
1417 entity->sched_data = &bfqg->sched_data;
1418}
1419
1420void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio) {}
1421
1422void bfq_end_wr_async(struct bfq_data *bfqd)
1423{
1424 bfq_end_wr_async_queues(bfqd, bfqd->root_group);
1425}
1426
1427struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd, struct blkcg *blkcg)
1428{
1429 return bfqd->root_group;
1430}
1431
1432struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
1433{
1434 return bfqq->bfqd->root_group;
1435}
1436
Paolo Valente4d8340d2020-02-03 11:40:58 +01001437void bfqg_and_blkg_get(struct bfq_group *bfqg) {}
1438
1439void bfqg_and_blkg_put(struct bfq_group *bfqg) {}
1440
Paolo Valenteea25da42017-04-19 08:48:24 -06001441struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1442{
1443 struct bfq_group *bfqg;
1444 int i;
1445
1446 bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node);
1447 if (!bfqg)
1448 return NULL;
1449
1450 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
1451 bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
1452
1453 return bfqg;
1454}
1455#endif /* CONFIG_BFQ_GROUP_IOSCHED */