1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * cgroups support for the BFQ I/O scheduler.
4  */
5 #include <linux/module.h>
6 #include <linux/slab.h>
7 #include <linux/blkdev.h>
8 #include <linux/cgroup.h>
9 #include <linux/elevator.h>
10 #include <linux/ktime.h>
11 #include <linux/rbtree.h>
12 #include <linux/ioprio.h>
13 #include <linux/sbitmap.h>
14 #include <linux/delay.h>
15 
16 #include "bfq-iosched.h"
17 
18 #ifdef CONFIG_BFQ_CGROUP_DEBUG
bfq_stat_init(struct bfq_stat * stat,gfp_t gfp)19 static 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 
bfq_stat_exit(struct bfq_stat * stat)31 static 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  */
bfq_stat_add(struct bfq_stat * stat,uint64_t val)44 static 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  */
bfq_stat_read(struct bfq_stat * stat)53 static 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  */
bfq_stat_reset(struct bfq_stat * stat)62 static 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  */
bfq_stat_add_aux(struct bfq_stat * to,struct bfq_stat * from)75 static 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 /**
83  * 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  */
blkg_prfill_stat(struct seq_file * sf,struct blkg_policy_data * pd,int off)90 static 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 
96 /* bfqg stats flags */
97 enum bfqg_stats_flags {
98 	BFQG_stats_waiting = 0,
99 	BFQG_stats_idling,
100 	BFQG_stats_empty,
101 };
102 
103 #define BFQG_FLAG_FNS(name)						\
104 static void bfqg_stats_mark_##name(struct bfqg_stats *stats)	\
105 {									\
106 	stats->flags |= (1 << BFQG_stats_##name);			\
107 }									\
108 static void bfqg_stats_clear_##name(struct bfqg_stats *stats)	\
109 {									\
110 	stats->flags &= ~(1 << BFQG_stats_##name);			\
111 }									\
112 static int bfqg_stats_##name(struct bfqg_stats *stats)		\
113 {									\
114 	return (stats->flags & (1 << BFQG_stats_##name)) != 0;		\
115 }									\
116 
117 BFQG_FLAG_FNS(waiting)
BFQG_FLAG_FNS(idling)118 BFQG_FLAG_FNS(idling)
119 BFQG_FLAG_FNS(empty)
120 #undef BFQG_FLAG_FNS
121 
122 /* This should be called with the scheduler lock held. */
123 static void bfqg_stats_update_group_wait_time(struct bfqg_stats *stats)
124 {
125 	u64 now;
126 
127 	if (!bfqg_stats_waiting(stats))
128 		return;
129 
130 	now = ktime_get_ns();
131 	if (now > stats->start_group_wait_time)
132 		bfq_stat_add(&stats->group_wait_time,
133 			      now - stats->start_group_wait_time);
134 	bfqg_stats_clear_waiting(stats);
135 }
136 
137 /* This should be called with the scheduler lock held. */
bfqg_stats_set_start_group_wait_time(struct bfq_group * bfqg,struct bfq_group * curr_bfqg)138 static 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;
147 	stats->start_group_wait_time = ktime_get_ns();
148 	bfqg_stats_mark_waiting(stats);
149 }
150 
151 /* This should be called with the scheduler lock held. */
bfqg_stats_end_empty_time(struct bfqg_stats * stats)152 static void bfqg_stats_end_empty_time(struct bfqg_stats *stats)
153 {
154 	u64 now;
155 
156 	if (!bfqg_stats_empty(stats))
157 		return;
158 
159 	now = ktime_get_ns();
160 	if (now > stats->start_empty_time)
161 		bfq_stat_add(&stats->empty_time,
162 			      now - stats->start_empty_time);
163 	bfqg_stats_clear_empty(stats);
164 }
165 
bfqg_stats_update_dequeue(struct bfq_group * bfqg)166 void bfqg_stats_update_dequeue(struct bfq_group *bfqg)
167 {
168 	bfq_stat_add(&bfqg->stats.dequeue, 1);
169 }
170 
bfqg_stats_set_start_empty_time(struct bfq_group * bfqg)171 void 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 
186 	stats->start_empty_time = ktime_get_ns();
187 	bfqg_stats_mark_empty(stats);
188 }
189 
bfqg_stats_update_idle_time(struct bfq_group * bfqg)190 void bfqg_stats_update_idle_time(struct bfq_group *bfqg)
191 {
192 	struct bfqg_stats *stats = &bfqg->stats;
193 
194 	if (bfqg_stats_idling(stats)) {
195 		u64 now = ktime_get_ns();
196 
197 		if (now > stats->start_idle_time)
198 			bfq_stat_add(&stats->idle_time,
199 				      now - stats->start_idle_time);
200 		bfqg_stats_clear_idling(stats);
201 	}
202 }
203 
bfqg_stats_set_start_idle_time(struct bfq_group * bfqg)204 void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg)
205 {
206 	struct bfqg_stats *stats = &bfqg->stats;
207 
208 	stats->start_idle_time = ktime_get_ns();
209 	bfqg_stats_mark_idling(stats);
210 }
211 
bfqg_stats_update_avg_queue_size(struct bfq_group * bfqg)212 void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg)
213 {
214 	struct bfqg_stats *stats = &bfqg->stats;
215 
216 	bfq_stat_add(&stats->avg_queue_size_sum,
217 		      blkg_rwstat_total(&stats->queued));
218 	bfq_stat_add(&stats->avg_queue_size_samples, 1);
219 	bfqg_stats_update_group_wait_time(stats);
220 }
221 
bfqg_stats_update_io_add(struct bfq_group * bfqg,struct bfq_queue * bfqq,unsigned int op)222 void 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 
bfqg_stats_update_io_remove(struct bfq_group * bfqg,unsigned int op)231 void bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op)
232 {
233 	blkg_rwstat_add(&bfqg->stats.queued, op, -1);
234 }
235 
bfqg_stats_update_io_merged(struct bfq_group * bfqg,unsigned int op)236 void bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op)
237 {
238 	blkg_rwstat_add(&bfqg->stats.merged, op, 1);
239 }
240 
bfqg_stats_update_completion(struct bfq_group * bfqg,u64 start_time_ns,u64 io_start_time_ns,unsigned int op)241 void bfqg_stats_update_completion(struct bfq_group *bfqg, u64 start_time_ns,
242 				  u64 io_start_time_ns, unsigned int op)
243 {
244 	struct bfqg_stats *stats = &bfqg->stats;
245 	u64 now = ktime_get_ns();
246 
247 	if (now > io_start_time_ns)
248 		blkg_rwstat_add(&stats->service_time, op,
249 				now - io_start_time_ns);
250 	if (io_start_time_ns > start_time_ns)
251 		blkg_rwstat_add(&stats->wait_time, op,
252 				io_start_time_ns - start_time_ns);
253 }
254 
255 #else /* CONFIG_BFQ_CGROUP_DEBUG */
256 
bfqg_stats_update_io_add(struct bfq_group * bfqg,struct bfq_queue * bfqq,unsigned int op)257 void bfqg_stats_update_io_add(struct bfq_group *bfqg, struct bfq_queue *bfqq,
258 			      unsigned int op) { }
bfqg_stats_update_io_remove(struct bfq_group * bfqg,unsigned int op)259 void bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op) { }
bfqg_stats_update_io_merged(struct bfq_group * bfqg,unsigned int op)260 void bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op) { }
bfqg_stats_update_completion(struct bfq_group * bfqg,u64 start_time_ns,u64 io_start_time_ns,unsigned int op)261 void bfqg_stats_update_completion(struct bfq_group *bfqg, u64 start_time_ns,
262 				  u64 io_start_time_ns, unsigned int op) { }
bfqg_stats_update_dequeue(struct bfq_group * bfqg)263 void bfqg_stats_update_dequeue(struct bfq_group *bfqg) { }
bfqg_stats_set_start_empty_time(struct bfq_group * bfqg)264 void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg) { }
bfqg_stats_update_idle_time(struct bfq_group * bfqg)265 void bfqg_stats_update_idle_time(struct bfq_group *bfqg) { }
bfqg_stats_set_start_idle_time(struct bfq_group * bfqg)266 void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg) { }
bfqg_stats_update_avg_queue_size(struct bfq_group * bfqg)267 void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg) { }
268 
269 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
270 
271 #ifdef CONFIG_BFQ_GROUP_IOSCHED
272 
273 /*
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 
pd_to_bfqg(struct blkg_policy_data * pd)279 static 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 
bfqg_to_blkg(struct bfq_group * bfqg)284 struct blkcg_gq *bfqg_to_blkg(struct bfq_group *bfqg)
285 {
286 	return pd_to_blkg(&bfqg->pd);
287 }
288 
blkg_to_bfqg(struct blkcg_gq * blkg)289 static 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 
bfqg_parent(struct bfq_group * bfqg)301 static 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 
bfqq_group(struct bfq_queue * bfqq)308 struct 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 
bfqg_get(struct bfq_group * bfqg)322 static void bfqg_get(struct bfq_group *bfqg)
323 {
324 	bfqg->ref++;
325 }
326 
bfqg_put(struct bfq_group * bfqg)327 static void bfqg_put(struct bfq_group *bfqg)
328 {
329 	bfqg->ref--;
330 
331 	if (bfqg->ref == 0)
332 		kfree(bfqg);
333 }
334 
bfqg_and_blkg_get(struct bfq_group * bfqg)335 static void bfqg_and_blkg_get(struct bfq_group *bfqg)
336 {
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 
bfqg_and_blkg_put(struct bfq_group * bfqg)343 void bfqg_and_blkg_put(struct bfq_group *bfqg)
344 {
345 	blkg_put(bfqg_to_blkg(bfqg));
346 
347 	bfqg_put(bfqg);
348 }
349 
bfqg_stats_update_legacy_io(struct request_queue * q,struct request * rq)350 void 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 
354 	if (!bfqg)
355 		return;
356 
357 	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 
361 /* @stats = 0 */
bfqg_stats_reset(struct bfqg_stats * stats)362 static void bfqg_stats_reset(struct bfqg_stats *stats)
363 {
364 #ifdef CONFIG_BFQ_CGROUP_DEBUG
365 	/* 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);
369 	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);
376 #endif
377 }
378 
379 /* @to += @from */
bfqg_stats_add_aux(struct bfqg_stats * to,struct bfqg_stats * from)380 static void bfqg_stats_add_aux(struct bfqg_stats *to, struct bfqg_stats *from)
381 {
382 	if (!to || !from)
383 		return;
384 
385 #ifdef CONFIG_BFQ_CGROUP_DEBUG
386 	/* 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);
390 	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,
393 			  &from->avg_queue_size_samples);
394 	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);
398 #endif
399 }
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  */
bfqg_stats_xfer_dead(struct bfq_group * bfqg)406 static 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 
415 	lockdep_assert_held(&bfqg_to_blkg(bfqg)->q->queue_lock);
416 
417 	if (unlikely(!parent))
418 		return;
419 
420 	bfqg_stats_add_aux(&parent->stats, &bfqg->stats);
421 	bfqg_stats_reset(&bfqg->stats);
422 }
423 
bfq_init_entity(struct bfq_entity * entity,struct bfq_group * bfqg)424 void 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;
433 		/*
434 		 * Make sure that bfqg and its associated blkg do not
435 		 * disappear before entity.
436 		 */
437 		bfqg_and_blkg_get(bfqg);
438 	}
439 	entity->parent = bfqg->my_entity; /* NULL for root group */
440 	entity->sched_data = &bfqg->sched_data;
441 }
442 
bfqg_stats_exit(struct bfqg_stats * stats)443 static void bfqg_stats_exit(struct bfqg_stats *stats)
444 {
445 	blkg_rwstat_exit(&stats->bytes);
446 	blkg_rwstat_exit(&stats->ios);
447 #ifdef CONFIG_BFQ_CGROUP_DEBUG
448 	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);
452 	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);
459 #endif
460 }
461 
bfqg_stats_init(struct bfqg_stats * stats,gfp_t gfp)462 static int bfqg_stats_init(struct bfqg_stats *stats, gfp_t gfp)
463 {
464 	if (blkg_rwstat_init(&stats->bytes, gfp) ||
465 	    blkg_rwstat_init(&stats->ios, gfp))
466 		return -ENOMEM;
467 
468 #ifdef CONFIG_BFQ_CGROUP_DEBUG
469 	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) ||
473 	    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) ||
479 	    bfq_stat_init(&stats->empty_time, gfp)) {
480 		bfqg_stats_exit(stats);
481 		return -ENOMEM;
482 	}
483 #endif
484 
485 	return 0;
486 }
487 
cpd_to_bfqgd(struct blkcg_policy_data * cpd)488 static struct bfq_group_data *cpd_to_bfqgd(struct blkcg_policy_data *cpd)
489 {
490 	return cpd ? container_of(cpd, struct bfq_group_data, pd) : NULL;
491 }
492 
blkcg_to_bfqgd(struct blkcg * blkcg)493 static struct bfq_group_data *blkcg_to_bfqgd(struct blkcg *blkcg)
494 {
495 	return cpd_to_bfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_bfq));
496 }
497 
bfq_cpd_alloc(gfp_t gfp)498 static struct blkcg_policy_data *bfq_cpd_alloc(gfp_t gfp)
499 {
500 	struct bfq_group_data *bgd;
501 
502 	bgd = kzalloc(sizeof(*bgd), gfp);
503 	if (!bgd)
504 		return NULL;
505 	return &bgd->pd;
506 }
507 
bfq_cpd_init(struct blkcg_policy_data * cpd)508 static void bfq_cpd_init(struct blkcg_policy_data *cpd)
509 {
510 	struct bfq_group_data *d = cpd_to_bfqgd(cpd);
511 
512 	d->weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
513 		CGROUP_WEIGHT_DFL : BFQ_WEIGHT_LEGACY_DFL;
514 }
515 
bfq_cpd_free(struct blkcg_policy_data * cpd)516 static void bfq_cpd_free(struct blkcg_policy_data *cpd)
517 {
518 	kfree(cpd_to_bfqgd(cpd));
519 }
520 
bfq_pd_alloc(gfp_t gfp,struct request_queue * q,struct blkcg * blkcg)521 static struct blkg_policy_data *bfq_pd_alloc(gfp_t gfp, struct request_queue *q,
522 					     struct blkcg *blkcg)
523 {
524 	struct bfq_group *bfqg;
525 
526 	bfqg = kzalloc_node(sizeof(*bfqg), gfp, q->node);
527 	if (!bfqg)
528 		return NULL;
529 
530 	if (bfqg_stats_init(&bfqg->stats, gfp)) {
531 		kfree(bfqg);
532 		return NULL;
533 	}
534 
535 	/* see comments in bfq_bic_update_cgroup for why refcounting */
536 	bfqg_get(bfqg);
537 	return &bfqg->pd;
538 }
539 
bfq_pd_init(struct blkg_policy_data * pd)540 static void bfq_pd_init(struct blkg_policy_data *pd)
541 {
542 	struct blkcg_gq *blkg = pd_to_blkg(pd);
543 	struct bfq_group *bfqg = blkg_to_bfqg(blkg);
544 	struct bfq_data *bfqd = blkg->q->elevator->elevator_data;
545 	struct bfq_entity *entity = &bfqg->entity;
546 	struct bfq_group_data *d = blkcg_to_bfqgd(blkg->blkcg);
547 
548 	entity->orig_weight = entity->weight = entity->new_weight = d->weight;
549 	entity->my_sched_data = &bfqg->sched_data;
550 	bfqg->my_entity = entity; /*
551 				   * the root_group's will be set to NULL
552 				   * in bfq_init_queue()
553 				   */
554 	bfqg->bfqd = bfqd;
555 	bfqg->active_entities = 0;
556 	bfqg->rq_pos_tree = RB_ROOT;
557 }
558 
bfq_pd_free(struct blkg_policy_data * pd)559 static void bfq_pd_free(struct blkg_policy_data *pd)
560 {
561 	struct bfq_group *bfqg = pd_to_bfqg(pd);
562 
563 	bfqg_stats_exit(&bfqg->stats);
564 	bfqg_put(bfqg);
565 }
566 
bfq_pd_reset_stats(struct blkg_policy_data * pd)567 static void bfq_pd_reset_stats(struct blkg_policy_data *pd)
568 {
569 	struct bfq_group *bfqg = pd_to_bfqg(pd);
570 
571 	bfqg_stats_reset(&bfqg->stats);
572 }
573 
bfq_group_set_parent(struct bfq_group * bfqg,struct bfq_group * parent)574 static void bfq_group_set_parent(struct bfq_group *bfqg,
575 					struct bfq_group *parent)
576 {
577 	struct bfq_entity *entity;
578 
579 	entity = &bfqg->entity;
580 	entity->parent = parent->my_entity;
581 	entity->sched_data = &parent->sched_data;
582 }
583 
bfq_lookup_bfqg(struct bfq_data * bfqd,struct blkcg * blkcg)584 static struct bfq_group *bfq_lookup_bfqg(struct bfq_data *bfqd,
585 					 struct blkcg *blkcg)
586 {
587 	struct blkcg_gq *blkg;
588 
589 	blkg = blkg_lookup(blkcg, bfqd->queue);
590 	if (likely(blkg))
591 		return blkg_to_bfqg(blkg);
592 	return NULL;
593 }
594 
bfq_find_set_group(struct bfq_data * bfqd,struct blkcg * blkcg)595 struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd,
596 				     struct blkcg *blkcg)
597 {
598 	struct bfq_group *bfqg, *parent;
599 	struct bfq_entity *entity;
600 
601 	bfqg = bfq_lookup_bfqg(bfqd, blkcg);
602 
603 	if (unlikely(!bfqg))
604 		return NULL;
605 
606 	/*
607 	 * Update chain of bfq_groups as we might be handling a leaf group
608 	 * which, along with some of its relatives, has not been hooked yet
609 	 * to the private hierarchy of BFQ.
610 	 */
611 	entity = &bfqg->entity;
612 	for_each_entity(entity) {
613 		struct bfq_group *curr_bfqg = container_of(entity,
614 						struct bfq_group, entity);
615 		if (curr_bfqg != bfqd->root_group) {
616 			parent = bfqg_parent(curr_bfqg);
617 			if (!parent)
618 				parent = bfqd->root_group;
619 			bfq_group_set_parent(curr_bfqg, parent);
620 		}
621 	}
622 
623 	return bfqg;
624 }
625 
626 /**
627  * bfq_bfqq_move - migrate @bfqq to @bfqg.
628  * @bfqd: queue descriptor.
629  * @bfqq: the queue to move.
630  * @bfqg: the group to move to.
631  *
632  * Move @bfqq to @bfqg, deactivating it from its old group and reactivating
633  * it on the new one.  Avoid putting the entity on the old group idle tree.
634  *
635  * Must be called under the scheduler lock, to make sure that the blkg
636  * owning @bfqg does not disappear (see comments in
637  * bfq_bic_update_cgroup on guaranteeing the consistency of blkg
638  * objects).
639  */
bfq_bfqq_move(struct bfq_data * bfqd,struct bfq_queue * bfqq,struct bfq_group * bfqg)640 void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
641 		   struct bfq_group *bfqg)
642 {
643 	struct bfq_entity *entity = &bfqq->entity;
644 
645 	/*
646 	 * Get extra reference to prevent bfqq from being freed in
647 	 * next possible expire or deactivate.
648 	 */
649 	bfqq->ref++;
650 
651 	/* If bfqq is empty, then bfq_bfqq_expire also invokes
652 	 * bfq_del_bfqq_busy, thereby removing bfqq and its entity
653 	 * from data structures related to current group. Otherwise we
654 	 * need to remove bfqq explicitly with bfq_deactivate_bfqq, as
655 	 * we do below.
656 	 */
657 	if (bfqq == bfqd->in_service_queue)
658 		bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
659 				false, BFQQE_PREEMPTED);
660 
661 	if (bfq_bfqq_busy(bfqq))
662 		bfq_deactivate_bfqq(bfqd, bfqq, false, false);
663 	else if (entity->on_st_or_in_serv)
664 		bfq_put_idle_entity(bfq_entity_service_tree(entity), entity);
665 	bfqg_and_blkg_put(bfqq_group(bfqq));
666 
667 	entity->parent = bfqg->my_entity;
668 	entity->sched_data = &bfqg->sched_data;
669 	/* pin down bfqg and its associated blkg  */
670 	bfqg_and_blkg_get(bfqg);
671 
672 	if (bfq_bfqq_busy(bfqq)) {
673 		if (unlikely(!bfqd->nonrot_with_queueing))
674 			bfq_pos_tree_add_move(bfqd, bfqq);
675 		bfq_activate_bfqq(bfqd, bfqq);
676 	}
677 
678 	if (!bfqd->in_service_queue && !bfqd->rq_in_driver)
679 		bfq_schedule_dispatch(bfqd);
680 	/* release extra ref taken above, bfqq may happen to be freed now */
681 	bfq_put_queue(bfqq);
682 }
683 
684 /**
685  * __bfq_bic_change_cgroup - move @bic to @cgroup.
686  * @bfqd: the queue descriptor.
687  * @bic: the bic to move.
688  * @blkcg: the blk-cgroup to move to.
689  *
690  * Move bic to blkcg, assuming that bfqd->lock is held; which makes
691  * sure that the reference to cgroup is valid across the call (see
692  * comments in bfq_bic_update_cgroup on this issue)
693  *
694  * NOTE: an alternative approach might have been to store the current
695  * cgroup in bfqq and getting a reference to it, reducing the lookup
696  * time here, at the price of slightly more complex code.
697  */
__bfq_bic_change_cgroup(struct bfq_data * bfqd,struct bfq_io_cq * bic,struct blkcg * blkcg)698 static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd,
699 						struct bfq_io_cq *bic,
700 						struct blkcg *blkcg)
701 {
702 	struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0);
703 	struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1);
704 	struct bfq_group *bfqg;
705 	struct bfq_entity *entity;
706 
707 	bfqg = bfq_find_set_group(bfqd, blkcg);
708 
709 	if (unlikely(!bfqg))
710 		bfqg = bfqd->root_group;
711 
712 	if (async_bfqq) {
713 		entity = &async_bfqq->entity;
714 
715 		if (entity->sched_data != &bfqg->sched_data) {
716 			bic_set_bfqq(bic, NULL, 0);
717 			bfq_release_process_ref(bfqd, async_bfqq);
718 		}
719 	}
720 
721 	if (sync_bfqq) {
722 		entity = &sync_bfqq->entity;
723 		if (entity->sched_data != &bfqg->sched_data)
724 			bfq_bfqq_move(bfqd, sync_bfqq, bfqg);
725 	}
726 
727 	return bfqg;
728 }
729 
bfq_bic_update_cgroup(struct bfq_io_cq * bic,struct bio * bio)730 void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio)
731 {
732 	struct bfq_data *bfqd = bic_to_bfqd(bic);
733 	struct bfq_group *bfqg = NULL;
734 	uint64_t serial_nr;
735 
736 	rcu_read_lock();
737 	serial_nr = __bio_blkcg(bio)->css.serial_nr;
738 
739 	/*
740 	 * Check whether blkcg has changed.  The condition may trigger
741 	 * spuriously on a newly created cic but there's no harm.
742 	 */
743 	if (unlikely(!bfqd) || likely(bic->blkcg_serial_nr == serial_nr))
744 		goto out;
745 
746 	bfqg = __bfq_bic_change_cgroup(bfqd, bic, __bio_blkcg(bio));
747 	/*
748 	 * Update blkg_path for bfq_log_* functions. We cache this
749 	 * path, and update it here, for the following
750 	 * reasons. Operations on blkg objects in blk-cgroup are
751 	 * protected with the request_queue lock, and not with the
752 	 * lock that protects the instances of this scheduler
753 	 * (bfqd->lock). This exposes BFQ to the following sort of
754 	 * race.
755 	 *
756 	 * The blkg_lookup performed in bfq_get_queue, protected
757 	 * through rcu, may happen to return the address of a copy of
758 	 * the original blkg. If this is the case, then the
759 	 * bfqg_and_blkg_get performed in bfq_get_queue, to pin down
760 	 * the blkg, is useless: it does not prevent blk-cgroup code
761 	 * from destroying both the original blkg and all objects
762 	 * directly or indirectly referred by the copy of the
763 	 * blkg.
764 	 *
765 	 * On the bright side, destroy operations on a blkg invoke, as
766 	 * a first step, hooks of the scheduler associated with the
767 	 * blkg. And these hooks are executed with bfqd->lock held for
768 	 * BFQ. As a consequence, for any blkg associated with the
769 	 * request queue this instance of the scheduler is attached
770 	 * to, we are guaranteed that such a blkg is not destroyed, and
771 	 * that all the pointers it contains are consistent, while we
772 	 * are holding bfqd->lock. A blkg_lookup performed with
773 	 * bfqd->lock held then returns a fully consistent blkg, which
774 	 * remains consistent until this lock is held.
775 	 *
776 	 * Thanks to the last fact, and to the fact that: (1) bfqg has
777 	 * been obtained through a blkg_lookup in the above
778 	 * assignment, and (2) bfqd->lock is being held, here we can
779 	 * safely use the policy data for the involved blkg (i.e., the
780 	 * field bfqg->pd) to get to the blkg associated with bfqg,
781 	 * and then we can safely use any field of blkg. After we
782 	 * release bfqd->lock, even just getting blkg through this
783 	 * bfqg may cause dangling references to be traversed, as
784 	 * bfqg->pd may not exist any more.
785 	 *
786 	 * In view of the above facts, here we cache, in the bfqg, any
787 	 * blkg data we may need for this bic, and for its associated
788 	 * bfq_queue. As of now, we need to cache only the path of the
789 	 * blkg, which is used in the bfq_log_* functions.
790 	 *
791 	 * Finally, note that bfqg itself needs to be protected from
792 	 * destruction on the blkg_free of the original blkg (which
793 	 * invokes bfq_pd_free). We use an additional private
794 	 * refcounter for bfqg, to let it disappear only after no
795 	 * bfq_queue refers to it any longer.
796 	 */
797 	blkg_path(bfqg_to_blkg(bfqg), bfqg->blkg_path, sizeof(bfqg->blkg_path));
798 	bic->blkcg_serial_nr = serial_nr;
799 out:
800 	rcu_read_unlock();
801 }
802 
803 /**
804  * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st.
805  * @st: the service tree being flushed.
806  */
bfq_flush_idle_tree(struct bfq_service_tree * st)807 static void bfq_flush_idle_tree(struct bfq_service_tree *st)
808 {
809 	struct bfq_entity *entity = st->first_idle;
810 
811 	for (; entity ; entity = st->first_idle)
812 		__bfq_deactivate_entity(entity, false);
813 }
814 
815 /**
816  * bfq_reparent_leaf_entity - move leaf entity to the root_group.
817  * @bfqd: the device data structure with the root group.
818  * @entity: the entity to move, if entity is a leaf; or the parent entity
819  *	    of an active leaf entity to move, if entity is not a leaf.
820  */
bfq_reparent_leaf_entity(struct bfq_data * bfqd,struct bfq_entity * entity,int ioprio_class)821 static void bfq_reparent_leaf_entity(struct bfq_data *bfqd,
822 				     struct bfq_entity *entity,
823 				     int ioprio_class)
824 {
825 	struct bfq_queue *bfqq;
826 	struct bfq_entity *child_entity = entity;
827 
828 	while (child_entity->my_sched_data) { /* leaf not reached yet */
829 		struct bfq_sched_data *child_sd = child_entity->my_sched_data;
830 		struct bfq_service_tree *child_st = child_sd->service_tree +
831 			ioprio_class;
832 		struct rb_root *child_active = &child_st->active;
833 
834 		child_entity = bfq_entity_of(rb_first(child_active));
835 
836 		if (!child_entity)
837 			child_entity = child_sd->in_service_entity;
838 	}
839 
840 	bfqq = bfq_entity_to_bfqq(child_entity);
841 	bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
842 }
843 
844 /**
845  * bfq_reparent_active_queues - move to the root group all active queues.
846  * @bfqd: the device data structure with the root group.
847  * @bfqg: the group to move from.
848  * @st: the service tree to start the search from.
849  */
bfq_reparent_active_queues(struct bfq_data * bfqd,struct bfq_group * bfqg,struct bfq_service_tree * st,int ioprio_class)850 static void bfq_reparent_active_queues(struct bfq_data *bfqd,
851 				       struct bfq_group *bfqg,
852 				       struct bfq_service_tree *st,
853 				       int ioprio_class)
854 {
855 	struct rb_root *active = &st->active;
856 	struct bfq_entity *entity;
857 
858 	while ((entity = bfq_entity_of(rb_first(active))))
859 		bfq_reparent_leaf_entity(bfqd, entity, ioprio_class);
860 
861 	if (bfqg->sched_data.in_service_entity)
862 		bfq_reparent_leaf_entity(bfqd,
863 					 bfqg->sched_data.in_service_entity,
864 					 ioprio_class);
865 }
866 
867 /**
868  * bfq_pd_offline - deactivate the entity associated with @pd,
869  *		    and reparent its children entities.
870  * @pd: descriptor of the policy going offline.
871  *
872  * blkio already grabs the queue_lock for us, so no need to use
873  * RCU-based magic
874  */
bfq_pd_offline(struct blkg_policy_data * pd)875 static void bfq_pd_offline(struct blkg_policy_data *pd)
876 {
877 	struct bfq_service_tree *st;
878 	struct bfq_group *bfqg = pd_to_bfqg(pd);
879 	struct bfq_data *bfqd = bfqg->bfqd;
880 	struct bfq_entity *entity = bfqg->my_entity;
881 	unsigned long flags;
882 	int i;
883 
884 	spin_lock_irqsave(&bfqd->lock, flags);
885 
886 	if (!entity) /* root group */
887 		goto put_async_queues;
888 
889 	/*
890 	 * Empty all service_trees belonging to this group before
891 	 * deactivating the group itself.
892 	 */
893 	for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) {
894 		st = bfqg->sched_data.service_tree + i;
895 
896 		/*
897 		 * It may happen that some queues are still active
898 		 * (busy) upon group destruction (if the corresponding
899 		 * processes have been forced to terminate). We move
900 		 * all the leaf entities corresponding to these queues
901 		 * to the root_group.
902 		 * Also, it may happen that the group has an entity
903 		 * in service, which is disconnected from the active
904 		 * tree: it must be moved, too.
905 		 * There is no need to put the sync queues, as the
906 		 * scheduler has taken no reference.
907 		 */
908 		bfq_reparent_active_queues(bfqd, bfqg, st, i);
909 
910 		/*
911 		 * The idle tree may still contain bfq_queues
912 		 * belonging to exited task because they never
913 		 * migrated to a different cgroup from the one being
914 		 * destroyed now. In addition, even
915 		 * bfq_reparent_active_queues() may happen to add some
916 		 * entities to the idle tree. It happens if, in some
917 		 * of the calls to bfq_bfqq_move() performed by
918 		 * bfq_reparent_active_queues(), the queue to move is
919 		 * empty and gets expired.
920 		 */
921 		bfq_flush_idle_tree(st);
922 	}
923 
924 	__bfq_deactivate_entity(entity, false);
925 
926 put_async_queues:
927 	bfq_put_async_queues(bfqd, bfqg);
928 
929 	spin_unlock_irqrestore(&bfqd->lock, flags);
930 	/*
931 	 * @blkg is going offline and will be ignored by
932 	 * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
933 	 * that they don't get lost.  If IOs complete after this point, the
934 	 * stats for them will be lost.  Oh well...
935 	 */
936 	bfqg_stats_xfer_dead(bfqg);
937 }
938 
bfq_end_wr_async(struct bfq_data * bfqd)939 void bfq_end_wr_async(struct bfq_data *bfqd)
940 {
941 	struct blkcg_gq *blkg;
942 
943 	list_for_each_entry(blkg, &bfqd->queue->blkg_list, q_node) {
944 		struct bfq_group *bfqg = blkg_to_bfqg(blkg);
945 
946 		bfq_end_wr_async_queues(bfqd, bfqg);
947 	}
948 	bfq_end_wr_async_queues(bfqd, bfqd->root_group);
949 }
950 
bfq_io_show_weight_legacy(struct seq_file * sf,void * v)951 static int bfq_io_show_weight_legacy(struct seq_file *sf, void *v)
952 {
953 	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
954 	struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
955 	unsigned int val = 0;
956 
957 	if (bfqgd)
958 		val = bfqgd->weight;
959 
960 	seq_printf(sf, "%u\n", val);
961 
962 	return 0;
963 }
964 
bfqg_prfill_weight_device(struct seq_file * sf,struct blkg_policy_data * pd,int off)965 static u64 bfqg_prfill_weight_device(struct seq_file *sf,
966 				     struct blkg_policy_data *pd, int off)
967 {
968 	struct bfq_group *bfqg = pd_to_bfqg(pd);
969 
970 	if (!bfqg->entity.dev_weight)
971 		return 0;
972 	return __blkg_prfill_u64(sf, pd, bfqg->entity.dev_weight);
973 }
974 
bfq_io_show_weight(struct seq_file * sf,void * v)975 static int bfq_io_show_weight(struct seq_file *sf, void *v)
976 {
977 	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
978 	struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
979 
980 	seq_printf(sf, "default %u\n", bfqgd->weight);
981 	blkcg_print_blkgs(sf, blkcg, bfqg_prfill_weight_device,
982 			  &blkcg_policy_bfq, 0, false);
983 	return 0;
984 }
985 
bfq_group_set_weight(struct bfq_group * bfqg,u64 weight,u64 dev_weight)986 static void bfq_group_set_weight(struct bfq_group *bfqg, u64 weight, u64 dev_weight)
987 {
988 	weight = dev_weight ?: weight;
989 
990 	bfqg->entity.dev_weight = dev_weight;
991 	/*
992 	 * Setting the prio_changed flag of the entity
993 	 * to 1 with new_weight == weight would re-set
994 	 * the value of the weight to its ioprio mapping.
995 	 * Set the flag only if necessary.
996 	 */
997 	if ((unsigned short)weight != bfqg->entity.new_weight) {
998 		bfqg->entity.new_weight = (unsigned short)weight;
999 		/*
1000 		 * Make sure that the above new value has been
1001 		 * stored in bfqg->entity.new_weight before
1002 		 * setting the prio_changed flag. In fact,
1003 		 * this flag may be read asynchronously (in
1004 		 * critical sections protected by a different
1005 		 * lock than that held here), and finding this
1006 		 * flag set may cause the execution of the code
1007 		 * for updating parameters whose value may
1008 		 * depend also on bfqg->entity.new_weight (in
1009 		 * __bfq_entity_update_weight_prio).
1010 		 * This barrier makes sure that the new value
1011 		 * of bfqg->entity.new_weight is correctly
1012 		 * seen in that code.
1013 		 */
1014 		smp_wmb();
1015 		bfqg->entity.prio_changed = 1;
1016 	}
1017 }
1018 
bfq_io_set_weight_legacy(struct cgroup_subsys_state * css,struct cftype * cftype,u64 val)1019 static int bfq_io_set_weight_legacy(struct cgroup_subsys_state *css,
1020 				    struct cftype *cftype,
1021 				    u64 val)
1022 {
1023 	struct blkcg *blkcg = css_to_blkcg(css);
1024 	struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
1025 	struct blkcg_gq *blkg;
1026 	int ret = -ERANGE;
1027 
1028 	if (val < BFQ_MIN_WEIGHT || val > BFQ_MAX_WEIGHT)
1029 		return ret;
1030 
1031 	ret = 0;
1032 	spin_lock_irq(&blkcg->lock);
1033 	bfqgd->weight = (unsigned short)val;
1034 	hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1035 		struct bfq_group *bfqg = blkg_to_bfqg(blkg);
1036 
1037 		if (bfqg)
1038 			bfq_group_set_weight(bfqg, val, 0);
1039 	}
1040 	spin_unlock_irq(&blkcg->lock);
1041 
1042 	return ret;
1043 }
1044 
bfq_io_set_device_weight(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)1045 static ssize_t bfq_io_set_device_weight(struct kernfs_open_file *of,
1046 					char *buf, size_t nbytes,
1047 					loff_t off)
1048 {
1049 	int ret;
1050 	struct blkg_conf_ctx ctx;
1051 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1052 	struct bfq_group *bfqg;
1053 	u64 v;
1054 
1055 	ret = blkg_conf_prep(blkcg, &blkcg_policy_bfq, buf, &ctx);
1056 	if (ret)
1057 		return ret;
1058 
1059 	if (sscanf(ctx.body, "%llu", &v) == 1) {
1060 		/* require "default" on dfl */
1061 		ret = -ERANGE;
1062 		if (!v)
1063 			goto out;
1064 	} else if (!strcmp(strim(ctx.body), "default")) {
1065 		v = 0;
1066 	} else {
1067 		ret = -EINVAL;
1068 		goto out;
1069 	}
1070 
1071 	bfqg = blkg_to_bfqg(ctx.blkg);
1072 
1073 	ret = -ERANGE;
1074 	if (!v || (v >= BFQ_MIN_WEIGHT && v <= BFQ_MAX_WEIGHT)) {
1075 		bfq_group_set_weight(bfqg, bfqg->entity.weight, v);
1076 		ret = 0;
1077 	}
1078 out:
1079 	blkg_conf_finish(&ctx);
1080 	return ret ?: nbytes;
1081 }
1082 
bfq_io_set_weight(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)1083 static ssize_t bfq_io_set_weight(struct kernfs_open_file *of,
1084 				 char *buf, size_t nbytes,
1085 				 loff_t off)
1086 {
1087 	char *endp;
1088 	int ret;
1089 	u64 v;
1090 
1091 	buf = strim(buf);
1092 
1093 	/* "WEIGHT" or "default WEIGHT" sets the default weight */
1094 	v = simple_strtoull(buf, &endp, 0);
1095 	if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
1096 		ret = bfq_io_set_weight_legacy(of_css(of), NULL, v);
1097 		return ret ?: nbytes;
1098 	}
1099 
1100 	return bfq_io_set_device_weight(of, buf, nbytes, off);
1101 }
1102 
bfqg_print_rwstat(struct seq_file * sf,void * v)1103 static int bfqg_print_rwstat(struct seq_file *sf, void *v)
1104 {
1105 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1106 			  &blkcg_policy_bfq, seq_cft(sf)->private, true);
1107 	return 0;
1108 }
1109 
bfqg_prfill_rwstat_recursive(struct seq_file * sf,struct blkg_policy_data * pd,int off)1110 static u64 bfqg_prfill_rwstat_recursive(struct seq_file *sf,
1111 					struct blkg_policy_data *pd, int off)
1112 {
1113 	struct blkg_rwstat_sample sum;
1114 
1115 	blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_bfq, off, &sum);
1116 	return __blkg_prfill_rwstat(sf, pd, &sum);
1117 }
1118 
bfqg_print_rwstat_recursive(struct seq_file * sf,void * v)1119 static int bfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1120 {
1121 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1122 			  bfqg_prfill_rwstat_recursive, &blkcg_policy_bfq,
1123 			  seq_cft(sf)->private, true);
1124 	return 0;
1125 }
1126 
1127 #ifdef CONFIG_BFQ_CGROUP_DEBUG
bfqg_print_stat(struct seq_file * sf,void * v)1128 static int bfqg_print_stat(struct seq_file *sf, void *v)
1129 {
1130 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1131 			  &blkcg_policy_bfq, seq_cft(sf)->private, false);
1132 	return 0;
1133 }
1134 
bfqg_prfill_stat_recursive(struct seq_file * sf,struct blkg_policy_data * pd,int off)1135 static u64 bfqg_prfill_stat_recursive(struct seq_file *sf,
1136 				      struct blkg_policy_data *pd, int off)
1137 {
1138 	struct blkcg_gq *blkg = pd_to_blkg(pd);
1139 	struct blkcg_gq *pos_blkg;
1140 	struct cgroup_subsys_state *pos_css;
1141 	u64 sum = 0;
1142 
1143 	lockdep_assert_held(&blkg->q->queue_lock);
1144 
1145 	rcu_read_lock();
1146 	blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
1147 		struct bfq_stat *stat;
1148 
1149 		if (!pos_blkg->online)
1150 			continue;
1151 
1152 		stat = (void *)blkg_to_pd(pos_blkg, &blkcg_policy_bfq) + off;
1153 		sum += bfq_stat_read(stat) + atomic64_read(&stat->aux_cnt);
1154 	}
1155 	rcu_read_unlock();
1156 
1157 	return __blkg_prfill_u64(sf, pd, sum);
1158 }
1159 
bfqg_print_stat_recursive(struct seq_file * sf,void * v)1160 static int bfqg_print_stat_recursive(struct seq_file *sf, void *v)
1161 {
1162 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1163 			  bfqg_prfill_stat_recursive, &blkcg_policy_bfq,
1164 			  seq_cft(sf)->private, false);
1165 	return 0;
1166 }
1167 
bfqg_prfill_sectors(struct seq_file * sf,struct blkg_policy_data * pd,int off)1168 static u64 bfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1169 			       int off)
1170 {
1171 	struct bfq_group *bfqg = blkg_to_bfqg(pd->blkg);
1172 	u64 sum = blkg_rwstat_total(&bfqg->stats.bytes);
1173 
1174 	return __blkg_prfill_u64(sf, pd, sum >> 9);
1175 }
1176 
bfqg_print_stat_sectors(struct seq_file * sf,void * v)1177 static int bfqg_print_stat_sectors(struct seq_file *sf, void *v)
1178 {
1179 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1180 			  bfqg_prfill_sectors, &blkcg_policy_bfq, 0, false);
1181 	return 0;
1182 }
1183 
bfqg_prfill_sectors_recursive(struct seq_file * sf,struct blkg_policy_data * pd,int off)1184 static u64 bfqg_prfill_sectors_recursive(struct seq_file *sf,
1185 					 struct blkg_policy_data *pd, int off)
1186 {
1187 	struct blkg_rwstat_sample tmp;
1188 
1189 	blkg_rwstat_recursive_sum(pd->blkg, &blkcg_policy_bfq,
1190 			offsetof(struct bfq_group, stats.bytes), &tmp);
1191 
1192 	return __blkg_prfill_u64(sf, pd,
1193 		(tmp.cnt[BLKG_RWSTAT_READ] + tmp.cnt[BLKG_RWSTAT_WRITE]) >> 9);
1194 }
1195 
bfqg_print_stat_sectors_recursive(struct seq_file * sf,void * v)1196 static int bfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1197 {
1198 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1199 			  bfqg_prfill_sectors_recursive, &blkcg_policy_bfq, 0,
1200 			  false);
1201 	return 0;
1202 }
1203 
bfqg_prfill_avg_queue_size(struct seq_file * sf,struct blkg_policy_data * pd,int off)1204 static u64 bfqg_prfill_avg_queue_size(struct seq_file *sf,
1205 				      struct blkg_policy_data *pd, int off)
1206 {
1207 	struct bfq_group *bfqg = pd_to_bfqg(pd);
1208 	u64 samples = bfq_stat_read(&bfqg->stats.avg_queue_size_samples);
1209 	u64 v = 0;
1210 
1211 	if (samples) {
1212 		v = bfq_stat_read(&bfqg->stats.avg_queue_size_sum);
1213 		v = div64_u64(v, samples);
1214 	}
1215 	__blkg_prfill_u64(sf, pd, v);
1216 	return 0;
1217 }
1218 
1219 /* print avg_queue_size */
bfqg_print_avg_queue_size(struct seq_file * sf,void * v)1220 static int bfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1221 {
1222 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1223 			  bfqg_prfill_avg_queue_size, &blkcg_policy_bfq,
1224 			  0, false);
1225 	return 0;
1226 }
1227 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1228 
bfq_create_group_hierarchy(struct bfq_data * bfqd,int node)1229 struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1230 {
1231 	int ret;
1232 
1233 	ret = blkcg_activate_policy(bfqd->queue, &blkcg_policy_bfq);
1234 	if (ret)
1235 		return NULL;
1236 
1237 	return blkg_to_bfqg(bfqd->queue->root_blkg);
1238 }
1239 
1240 struct blkcg_policy blkcg_policy_bfq = {
1241 	.dfl_cftypes		= bfq_blkg_files,
1242 	.legacy_cftypes		= bfq_blkcg_legacy_files,
1243 
1244 	.cpd_alloc_fn		= bfq_cpd_alloc,
1245 	.cpd_init_fn		= bfq_cpd_init,
1246 	.cpd_bind_fn	        = bfq_cpd_init,
1247 	.cpd_free_fn		= bfq_cpd_free,
1248 
1249 	.pd_alloc_fn		= bfq_pd_alloc,
1250 	.pd_init_fn		= bfq_pd_init,
1251 	.pd_offline_fn		= bfq_pd_offline,
1252 	.pd_free_fn		= bfq_pd_free,
1253 	.pd_reset_stats_fn	= bfq_pd_reset_stats,
1254 };
1255 
1256 struct cftype bfq_blkcg_legacy_files[] = {
1257 	{
1258 		.name = "bfq.weight",
1259 		.flags = CFTYPE_NOT_ON_ROOT,
1260 		.seq_show = bfq_io_show_weight_legacy,
1261 		.write_u64 = bfq_io_set_weight_legacy,
1262 	},
1263 	{
1264 		.name = "bfq.weight_device",
1265 		.flags = CFTYPE_NOT_ON_ROOT,
1266 		.seq_show = bfq_io_show_weight,
1267 		.write = bfq_io_set_weight,
1268 	},
1269 
1270 	/* statistics, covers only the tasks in the bfqg */
1271 	{
1272 		.name = "bfq.io_service_bytes",
1273 		.private = offsetof(struct bfq_group, stats.bytes),
1274 		.seq_show = bfqg_print_rwstat,
1275 	},
1276 	{
1277 		.name = "bfq.io_serviced",
1278 		.private = offsetof(struct bfq_group, stats.ios),
1279 		.seq_show = bfqg_print_rwstat,
1280 	},
1281 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1282 	{
1283 		.name = "bfq.time",
1284 		.private = offsetof(struct bfq_group, stats.time),
1285 		.seq_show = bfqg_print_stat,
1286 	},
1287 	{
1288 		.name = "bfq.sectors",
1289 		.seq_show = bfqg_print_stat_sectors,
1290 	},
1291 	{
1292 		.name = "bfq.io_service_time",
1293 		.private = offsetof(struct bfq_group, stats.service_time),
1294 		.seq_show = bfqg_print_rwstat,
1295 	},
1296 	{
1297 		.name = "bfq.io_wait_time",
1298 		.private = offsetof(struct bfq_group, stats.wait_time),
1299 		.seq_show = bfqg_print_rwstat,
1300 	},
1301 	{
1302 		.name = "bfq.io_merged",
1303 		.private = offsetof(struct bfq_group, stats.merged),
1304 		.seq_show = bfqg_print_rwstat,
1305 	},
1306 	{
1307 		.name = "bfq.io_queued",
1308 		.private = offsetof(struct bfq_group, stats.queued),
1309 		.seq_show = bfqg_print_rwstat,
1310 	},
1311 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1312 
1313 	/* the same statistics which cover the bfqg and its descendants */
1314 	{
1315 		.name = "bfq.io_service_bytes_recursive",
1316 		.private = offsetof(struct bfq_group, stats.bytes),
1317 		.seq_show = bfqg_print_rwstat_recursive,
1318 	},
1319 	{
1320 		.name = "bfq.io_serviced_recursive",
1321 		.private = offsetof(struct bfq_group, stats.ios),
1322 		.seq_show = bfqg_print_rwstat_recursive,
1323 	},
1324 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1325 	{
1326 		.name = "bfq.time_recursive",
1327 		.private = offsetof(struct bfq_group, stats.time),
1328 		.seq_show = bfqg_print_stat_recursive,
1329 	},
1330 	{
1331 		.name = "bfq.sectors_recursive",
1332 		.seq_show = bfqg_print_stat_sectors_recursive,
1333 	},
1334 	{
1335 		.name = "bfq.io_service_time_recursive",
1336 		.private = offsetof(struct bfq_group, stats.service_time),
1337 		.seq_show = bfqg_print_rwstat_recursive,
1338 	},
1339 	{
1340 		.name = "bfq.io_wait_time_recursive",
1341 		.private = offsetof(struct bfq_group, stats.wait_time),
1342 		.seq_show = bfqg_print_rwstat_recursive,
1343 	},
1344 	{
1345 		.name = "bfq.io_merged_recursive",
1346 		.private = offsetof(struct bfq_group, stats.merged),
1347 		.seq_show = bfqg_print_rwstat_recursive,
1348 	},
1349 	{
1350 		.name = "bfq.io_queued_recursive",
1351 		.private = offsetof(struct bfq_group, stats.queued),
1352 		.seq_show = bfqg_print_rwstat_recursive,
1353 	},
1354 	{
1355 		.name = "bfq.avg_queue_size",
1356 		.seq_show = bfqg_print_avg_queue_size,
1357 	},
1358 	{
1359 		.name = "bfq.group_wait_time",
1360 		.private = offsetof(struct bfq_group, stats.group_wait_time),
1361 		.seq_show = bfqg_print_stat,
1362 	},
1363 	{
1364 		.name = "bfq.idle_time",
1365 		.private = offsetof(struct bfq_group, stats.idle_time),
1366 		.seq_show = bfqg_print_stat,
1367 	},
1368 	{
1369 		.name = "bfq.empty_time",
1370 		.private = offsetof(struct bfq_group, stats.empty_time),
1371 		.seq_show = bfqg_print_stat,
1372 	},
1373 	{
1374 		.name = "bfq.dequeue",
1375 		.private = offsetof(struct bfq_group, stats.dequeue),
1376 		.seq_show = bfqg_print_stat,
1377 	},
1378 #endif	/* CONFIG_BFQ_CGROUP_DEBUG */
1379 	{ }	/* terminate */
1380 };
1381 
1382 struct cftype bfq_blkg_files[] = {
1383 	{
1384 		.name = "bfq.weight",
1385 		.flags = CFTYPE_NOT_ON_ROOT,
1386 		.seq_show = bfq_io_show_weight,
1387 		.write = bfq_io_set_weight,
1388 	},
1389 	{} /* terminate */
1390 };
1391 
1392 #else	/* CONFIG_BFQ_GROUP_IOSCHED */
1393 
bfq_bfqq_move(struct bfq_data * bfqd,struct bfq_queue * bfqq,struct bfq_group * bfqg)1394 void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1395 		   struct bfq_group *bfqg) {}
1396 
bfq_init_entity(struct bfq_entity * entity,struct bfq_group * bfqg)1397 void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
1398 {
1399 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
1400 
1401 	entity->weight = entity->new_weight;
1402 	entity->orig_weight = entity->new_weight;
1403 	if (bfqq) {
1404 		bfqq->ioprio = bfqq->new_ioprio;
1405 		bfqq->ioprio_class = bfqq->new_ioprio_class;
1406 	}
1407 	entity->sched_data = &bfqg->sched_data;
1408 }
1409 
bfq_bic_update_cgroup(struct bfq_io_cq * bic,struct bio * bio)1410 void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio) {}
1411 
bfq_end_wr_async(struct bfq_data * bfqd)1412 void bfq_end_wr_async(struct bfq_data *bfqd)
1413 {
1414 	bfq_end_wr_async_queues(bfqd, bfqd->root_group);
1415 }
1416 
bfq_find_set_group(struct bfq_data * bfqd,struct blkcg * blkcg)1417 struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd, struct blkcg *blkcg)
1418 {
1419 	return bfqd->root_group;
1420 }
1421 
bfqq_group(struct bfq_queue * bfqq)1422 struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
1423 {
1424 	return bfqq->bfqd->root_group;
1425 }
1426 
bfqg_and_blkg_get(struct bfq_group * bfqg)1427 void bfqg_and_blkg_get(struct bfq_group *bfqg) {}
1428 
bfqg_and_blkg_put(struct bfq_group * bfqg)1429 void bfqg_and_blkg_put(struct bfq_group *bfqg) {}
1430 
bfq_create_group_hierarchy(struct bfq_data * bfqd,int node)1431 struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1432 {
1433 	struct bfq_group *bfqg;
1434 	int i;
1435 
1436 	bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node);
1437 	if (!bfqg)
1438 		return NULL;
1439 
1440 	for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
1441 		bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
1442 
1443 	return bfqg;
1444 }
1445 #endif	/* CONFIG_BFQ_GROUP_IOSCHED */
1446