• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/rbtree.h>
13 #include <linux/ioprio.h>
14 #include <linux/blktrace_api.h>
15 
16 /*
17  * tunables
18  */
19 /* max queue in one round of service */
20 static const int cfq_quantum = 4;
21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
22 /* maximum backwards seek, in KiB */
23 static const int cfq_back_max = 16 * 1024;
24 /* penalty of a backwards seek */
25 static const int cfq_back_penalty = 2;
26 static const int cfq_slice_sync = HZ / 10;
27 static int cfq_slice_async = HZ / 25;
28 static const int cfq_slice_async_rq = 2;
29 static int cfq_slice_idle = HZ / 125;
30 
31 /*
32  * offset from end of service tree
33  */
34 #define CFQ_IDLE_DELAY		(HZ / 5)
35 
36 /*
37  * below this threshold, we consider thinktime immediate
38  */
39 #define CFQ_MIN_TT		(2)
40 
41 #define CFQ_SLICE_SCALE		(5)
42 #define CFQ_HW_QUEUE_MIN	(5)
43 
44 #define RQ_CIC(rq)		\
45 	((struct cfq_io_context *) (rq)->elevator_private)
46 #define RQ_CFQQ(rq)		(struct cfq_queue *) ((rq)->elevator_private2)
47 
48 static struct kmem_cache *cfq_pool;
49 static struct kmem_cache *cfq_ioc_pool;
50 
51 static DEFINE_PER_CPU(unsigned long, ioc_count);
52 static struct completion *ioc_gone;
53 static DEFINE_SPINLOCK(ioc_gone_lock);
54 
55 #define CFQ_PRIO_LISTS		IOPRIO_BE_NR
56 #define cfq_class_idle(cfqq)	((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
57 #define cfq_class_rt(cfqq)	((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
58 
59 #define ASYNC			(0)
60 #define SYNC			(1)
61 
62 #define sample_valid(samples)	((samples) > 80)
63 
64 /*
65  * Most of our rbtree usage is for sorting with min extraction, so
66  * if we cache the leftmost node we don't have to walk down the tree
67  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
68  * move this into the elevator for the rq sorting as well.
69  */
70 struct cfq_rb_root {
71 	struct rb_root rb;
72 	struct rb_node *left;
73 };
74 #define CFQ_RB_ROOT	(struct cfq_rb_root) { RB_ROOT, NULL, }
75 
76 /*
77  * Per block device queue structure
78  */
79 struct cfq_data {
80 	struct request_queue *queue;
81 
82 	/*
83 	 * rr list of queues with requests and the count of them
84 	 */
85 	struct cfq_rb_root service_tree;
86 	unsigned int busy_queues;
87 	/*
88 	 * Used to track any pending rt requests so we can pre-empt current
89 	 * non-RT cfqq in service when this value is non-zero.
90 	 */
91 	unsigned int busy_rt_queues;
92 
93 	int rq_in_driver;
94 	int sync_flight;
95 
96 	/*
97 	 * queue-depth detection
98 	 */
99 	int rq_queued;
100 	int hw_tag;
101 	int hw_tag_samples;
102 	int rq_in_driver_peak;
103 
104 	/*
105 	 * idle window management
106 	 */
107 	struct timer_list idle_slice_timer;
108 	struct work_struct unplug_work;
109 
110 	struct cfq_queue *active_queue;
111 	struct cfq_io_context *active_cic;
112 
113 	/*
114 	 * async queue for each priority case
115 	 */
116 	struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
117 	struct cfq_queue *async_idle_cfqq;
118 
119 	sector_t last_position;
120 	unsigned long last_end_request;
121 
122 	/*
123 	 * tunables, see top of file
124 	 */
125 	unsigned int cfq_quantum;
126 	unsigned int cfq_fifo_expire[2];
127 	unsigned int cfq_back_penalty;
128 	unsigned int cfq_back_max;
129 	unsigned int cfq_slice[2];
130 	unsigned int cfq_slice_async_rq;
131 	unsigned int cfq_slice_idle;
132 
133 	struct list_head cic_list;
134 };
135 
136 /*
137  * Per process-grouping structure
138  */
139 struct cfq_queue {
140 	/* reference count */
141 	atomic_t ref;
142 	/* various state flags, see below */
143 	unsigned int flags;
144 	/* parent cfq_data */
145 	struct cfq_data *cfqd;
146 	/* service_tree member */
147 	struct rb_node rb_node;
148 	/* service_tree key */
149 	unsigned long rb_key;
150 	/* sorted list of pending requests */
151 	struct rb_root sort_list;
152 	/* if fifo isn't expired, next request to serve */
153 	struct request *next_rq;
154 	/* requests queued in sort_list */
155 	int queued[2];
156 	/* currently allocated requests */
157 	int allocated[2];
158 	/* fifo list of requests in sort_list */
159 	struct list_head fifo;
160 
161 	unsigned long slice_end;
162 	long slice_resid;
163 
164 	/* pending metadata requests */
165 	int meta_pending;
166 	/* number of requests that are on the dispatch list or inside driver */
167 	int dispatched;
168 
169 	/* io prio of this group */
170 	unsigned short ioprio, org_ioprio;
171 	unsigned short ioprio_class, org_ioprio_class;
172 
173 	pid_t pid;
174 };
175 
176 enum cfqq_state_flags {
177 	CFQ_CFQQ_FLAG_on_rr = 0,	/* on round-robin busy list */
178 	CFQ_CFQQ_FLAG_wait_request,	/* waiting for a request */
179 	CFQ_CFQQ_FLAG_must_alloc,	/* must be allowed rq alloc */
180 	CFQ_CFQQ_FLAG_must_alloc_slice,	/* per-slice must_alloc flag */
181 	CFQ_CFQQ_FLAG_must_dispatch,	/* must dispatch, even if expired */
182 	CFQ_CFQQ_FLAG_fifo_expire,	/* FIFO checked in this slice */
183 	CFQ_CFQQ_FLAG_idle_window,	/* slice idling enabled */
184 	CFQ_CFQQ_FLAG_prio_changed,	/* task priority has changed */
185 	CFQ_CFQQ_FLAG_queue_new,	/* queue never been serviced */
186 	CFQ_CFQQ_FLAG_slice_new,	/* no requests dispatched in slice */
187 	CFQ_CFQQ_FLAG_sync,		/* synchronous queue */
188 };
189 
190 #define CFQ_CFQQ_FNS(name)						\
191 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)		\
192 {									\
193 	(cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);			\
194 }									\
195 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)	\
196 {									\
197 	(cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);			\
198 }									\
199 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)		\
200 {									\
201 	return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;	\
202 }
203 
204 CFQ_CFQQ_FNS(on_rr);
205 CFQ_CFQQ_FNS(wait_request);
206 CFQ_CFQQ_FNS(must_alloc);
207 CFQ_CFQQ_FNS(must_alloc_slice);
208 CFQ_CFQQ_FNS(must_dispatch);
209 CFQ_CFQQ_FNS(fifo_expire);
210 CFQ_CFQQ_FNS(idle_window);
211 CFQ_CFQQ_FNS(prio_changed);
212 CFQ_CFQQ_FNS(queue_new);
213 CFQ_CFQQ_FNS(slice_new);
214 CFQ_CFQQ_FNS(sync);
215 #undef CFQ_CFQQ_FNS
216 
217 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)	\
218 	blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
219 #define cfq_log(cfqd, fmt, args...)	\
220 	blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
221 
222 static void cfq_dispatch_insert(struct request_queue *, struct request *);
223 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
224 				       struct io_context *, gfp_t);
225 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
226 						struct io_context *);
227 
cic_to_cfqq(struct cfq_io_context * cic,int is_sync)228 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
229 					    int is_sync)
230 {
231 	return cic->cfqq[!!is_sync];
232 }
233 
cic_set_cfqq(struct cfq_io_context * cic,struct cfq_queue * cfqq,int is_sync)234 static inline void cic_set_cfqq(struct cfq_io_context *cic,
235 				struct cfq_queue *cfqq, int is_sync)
236 {
237 	cic->cfqq[!!is_sync] = cfqq;
238 }
239 
240 /*
241  * We regard a request as SYNC, if it's either a read or has the SYNC bit
242  * set (in which case it could also be direct WRITE).
243  */
cfq_bio_sync(struct bio * bio)244 static inline int cfq_bio_sync(struct bio *bio)
245 {
246 	if (bio_data_dir(bio) == READ || bio_sync(bio))
247 		return 1;
248 
249 	return 0;
250 }
251 
252 /*
253  * scheduler run of queue, if there are requests pending and no one in the
254  * driver that will restart queueing
255  */
cfq_schedule_dispatch(struct cfq_data * cfqd)256 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
257 {
258 	if (cfqd->busy_queues) {
259 		cfq_log(cfqd, "schedule dispatch");
260 		kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
261 	}
262 }
263 
cfq_queue_empty(struct request_queue * q)264 static int cfq_queue_empty(struct request_queue *q)
265 {
266 	struct cfq_data *cfqd = q->elevator->elevator_data;
267 
268 	return !cfqd->busy_queues;
269 }
270 
271 /*
272  * Scale schedule slice based on io priority. Use the sync time slice only
273  * if a queue is marked sync and has sync io queued. A sync queue with async
274  * io only, should not get full sync slice length.
275  */
cfq_prio_slice(struct cfq_data * cfqd,int sync,unsigned short prio)276 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
277 				 unsigned short prio)
278 {
279 	const int base_slice = cfqd->cfq_slice[sync];
280 
281 	WARN_ON(prio >= IOPRIO_BE_NR);
282 
283 	return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
284 }
285 
286 static inline int
cfq_prio_to_slice(struct cfq_data * cfqd,struct cfq_queue * cfqq)287 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
288 {
289 	return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
290 }
291 
292 static inline void
cfq_set_prio_slice(struct cfq_data * cfqd,struct cfq_queue * cfqq)293 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
294 {
295 	cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
296 	cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
297 }
298 
299 /*
300  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
301  * isn't valid until the first request from the dispatch is activated
302  * and the slice time set.
303  */
cfq_slice_used(struct cfq_queue * cfqq)304 static inline int cfq_slice_used(struct cfq_queue *cfqq)
305 {
306 	if (cfq_cfqq_slice_new(cfqq))
307 		return 0;
308 	if (time_before(jiffies, cfqq->slice_end))
309 		return 0;
310 
311 	return 1;
312 }
313 
314 /*
315  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
316  * We choose the request that is closest to the head right now. Distance
317  * behind the head is penalized and only allowed to a certain extent.
318  */
319 static struct request *
cfq_choose_req(struct cfq_data * cfqd,struct request * rq1,struct request * rq2)320 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
321 {
322 	sector_t last, s1, s2, d1 = 0, d2 = 0;
323 	unsigned long back_max;
324 #define CFQ_RQ1_WRAP	0x01 /* request 1 wraps */
325 #define CFQ_RQ2_WRAP	0x02 /* request 2 wraps */
326 	unsigned wrap = 0; /* bit mask: requests behind the disk head? */
327 
328 	if (rq1 == NULL || rq1 == rq2)
329 		return rq2;
330 	if (rq2 == NULL)
331 		return rq1;
332 
333 	if (rq_is_sync(rq1) && !rq_is_sync(rq2))
334 		return rq1;
335 	else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
336 		return rq2;
337 	if (rq_is_meta(rq1) && !rq_is_meta(rq2))
338 		return rq1;
339 	else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
340 		return rq2;
341 
342 	s1 = rq1->sector;
343 	s2 = rq2->sector;
344 
345 	last = cfqd->last_position;
346 
347 	/*
348 	 * by definition, 1KiB is 2 sectors
349 	 */
350 	back_max = cfqd->cfq_back_max * 2;
351 
352 	/*
353 	 * Strict one way elevator _except_ in the case where we allow
354 	 * short backward seeks which are biased as twice the cost of a
355 	 * similar forward seek.
356 	 */
357 	if (s1 >= last)
358 		d1 = s1 - last;
359 	else if (s1 + back_max >= last)
360 		d1 = (last - s1) * cfqd->cfq_back_penalty;
361 	else
362 		wrap |= CFQ_RQ1_WRAP;
363 
364 	if (s2 >= last)
365 		d2 = s2 - last;
366 	else if (s2 + back_max >= last)
367 		d2 = (last - s2) * cfqd->cfq_back_penalty;
368 	else
369 		wrap |= CFQ_RQ2_WRAP;
370 
371 	/* Found required data */
372 
373 	/*
374 	 * By doing switch() on the bit mask "wrap" we avoid having to
375 	 * check two variables for all permutations: --> faster!
376 	 */
377 	switch (wrap) {
378 	case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
379 		if (d1 < d2)
380 			return rq1;
381 		else if (d2 < d1)
382 			return rq2;
383 		else {
384 			if (s1 >= s2)
385 				return rq1;
386 			else
387 				return rq2;
388 		}
389 
390 	case CFQ_RQ2_WRAP:
391 		return rq1;
392 	case CFQ_RQ1_WRAP:
393 		return rq2;
394 	case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
395 	default:
396 		/*
397 		 * Since both rqs are wrapped,
398 		 * start with the one that's further behind head
399 		 * (--> only *one* back seek required),
400 		 * since back seek takes more time than forward.
401 		 */
402 		if (s1 <= s2)
403 			return rq1;
404 		else
405 			return rq2;
406 	}
407 }
408 
409 /*
410  * The below is leftmost cache rbtree addon
411  */
cfq_rb_first(struct cfq_rb_root * root)412 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
413 {
414 	if (!root->left)
415 		root->left = rb_first(&root->rb);
416 
417 	if (root->left)
418 		return rb_entry(root->left, struct cfq_queue, rb_node);
419 
420 	return NULL;
421 }
422 
cfq_rb_erase(struct rb_node * n,struct cfq_rb_root * root)423 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
424 {
425 	if (root->left == n)
426 		root->left = NULL;
427 
428 	rb_erase(n, &root->rb);
429 	RB_CLEAR_NODE(n);
430 }
431 
432 /*
433  * would be nice to take fifo expire time into account as well
434  */
435 static struct request *
cfq_find_next_rq(struct cfq_data * cfqd,struct cfq_queue * cfqq,struct request * last)436 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
437 		  struct request *last)
438 {
439 	struct rb_node *rbnext = rb_next(&last->rb_node);
440 	struct rb_node *rbprev = rb_prev(&last->rb_node);
441 	struct request *next = NULL, *prev = NULL;
442 
443 	BUG_ON(RB_EMPTY_NODE(&last->rb_node));
444 
445 	if (rbprev)
446 		prev = rb_entry_rq(rbprev);
447 
448 	if (rbnext)
449 		next = rb_entry_rq(rbnext);
450 	else {
451 		rbnext = rb_first(&cfqq->sort_list);
452 		if (rbnext && rbnext != &last->rb_node)
453 			next = rb_entry_rq(rbnext);
454 	}
455 
456 	return cfq_choose_req(cfqd, next, prev);
457 }
458 
cfq_slice_offset(struct cfq_data * cfqd,struct cfq_queue * cfqq)459 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
460 				      struct cfq_queue *cfqq)
461 {
462 	/*
463 	 * just an approximation, should be ok.
464 	 */
465 	return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
466 		       cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
467 }
468 
469 /*
470  * The cfqd->service_tree holds all pending cfq_queue's that have
471  * requests waiting to be processed. It is sorted in the order that
472  * we will service the queues.
473  */
cfq_service_tree_add(struct cfq_data * cfqd,struct cfq_queue * cfqq,int add_front)474 static void cfq_service_tree_add(struct cfq_data *cfqd,
475 				    struct cfq_queue *cfqq, int add_front)
476 {
477 	struct rb_node **p, *parent;
478 	struct cfq_queue *__cfqq;
479 	unsigned long rb_key;
480 	int left;
481 
482 	if (cfq_class_idle(cfqq)) {
483 		rb_key = CFQ_IDLE_DELAY;
484 		parent = rb_last(&cfqd->service_tree.rb);
485 		if (parent && parent != &cfqq->rb_node) {
486 			__cfqq = rb_entry(parent, struct cfq_queue, rb_node);
487 			rb_key += __cfqq->rb_key;
488 		} else
489 			rb_key += jiffies;
490 	} else if (!add_front) {
491 		rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
492 		rb_key += cfqq->slice_resid;
493 		cfqq->slice_resid = 0;
494 	} else
495 		rb_key = 0;
496 
497 	if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
498 		/*
499 		 * same position, nothing more to do
500 		 */
501 		if (rb_key == cfqq->rb_key)
502 			return;
503 
504 		cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
505 	}
506 
507 	left = 1;
508 	parent = NULL;
509 	p = &cfqd->service_tree.rb.rb_node;
510 	while (*p) {
511 		struct rb_node **n;
512 
513 		parent = *p;
514 		__cfqq = rb_entry(parent, struct cfq_queue, rb_node);
515 
516 		/*
517 		 * sort RT queues first, we always want to give
518 		 * preference to them. IDLE queues goes to the back.
519 		 * after that, sort on the next service time.
520 		 */
521 		if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
522 			n = &(*p)->rb_left;
523 		else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
524 			n = &(*p)->rb_right;
525 		else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
526 			n = &(*p)->rb_left;
527 		else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
528 			n = &(*p)->rb_right;
529 		else if (rb_key < __cfqq->rb_key)
530 			n = &(*p)->rb_left;
531 		else
532 			n = &(*p)->rb_right;
533 
534 		if (n == &(*p)->rb_right)
535 			left = 0;
536 
537 		p = n;
538 	}
539 
540 	if (left)
541 		cfqd->service_tree.left = &cfqq->rb_node;
542 
543 	cfqq->rb_key = rb_key;
544 	rb_link_node(&cfqq->rb_node, parent, p);
545 	rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
546 }
547 
548 /*
549  * Update cfqq's position in the service tree.
550  */
cfq_resort_rr_list(struct cfq_data * cfqd,struct cfq_queue * cfqq)551 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
552 {
553 	/*
554 	 * Resorting requires the cfqq to be on the RR list already.
555 	 */
556 	if (cfq_cfqq_on_rr(cfqq))
557 		cfq_service_tree_add(cfqd, cfqq, 0);
558 }
559 
560 /*
561  * add to busy list of queues for service, trying to be fair in ordering
562  * the pending list according to last request service
563  */
cfq_add_cfqq_rr(struct cfq_data * cfqd,struct cfq_queue * cfqq)564 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
565 {
566 	cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
567 	BUG_ON(cfq_cfqq_on_rr(cfqq));
568 	cfq_mark_cfqq_on_rr(cfqq);
569 	cfqd->busy_queues++;
570 	if (cfq_class_rt(cfqq))
571 		cfqd->busy_rt_queues++;
572 
573 	cfq_resort_rr_list(cfqd, cfqq);
574 }
575 
576 /*
577  * Called when the cfqq no longer has requests pending, remove it from
578  * the service tree.
579  */
cfq_del_cfqq_rr(struct cfq_data * cfqd,struct cfq_queue * cfqq)580 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
581 {
582 	cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
583 	BUG_ON(!cfq_cfqq_on_rr(cfqq));
584 	cfq_clear_cfqq_on_rr(cfqq);
585 
586 	if (!RB_EMPTY_NODE(&cfqq->rb_node))
587 		cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
588 
589 	BUG_ON(!cfqd->busy_queues);
590 	cfqd->busy_queues--;
591 	if (cfq_class_rt(cfqq))
592 		cfqd->busy_rt_queues--;
593 }
594 
595 /*
596  * rb tree support functions
597  */
cfq_del_rq_rb(struct request * rq)598 static void cfq_del_rq_rb(struct request *rq)
599 {
600 	struct cfq_queue *cfqq = RQ_CFQQ(rq);
601 	struct cfq_data *cfqd = cfqq->cfqd;
602 	const int sync = rq_is_sync(rq);
603 
604 	BUG_ON(!cfqq->queued[sync]);
605 	cfqq->queued[sync]--;
606 
607 	elv_rb_del(&cfqq->sort_list, rq);
608 
609 	if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
610 		cfq_del_cfqq_rr(cfqd, cfqq);
611 }
612 
cfq_add_rq_rb(struct request * rq)613 static void cfq_add_rq_rb(struct request *rq)
614 {
615 	struct cfq_queue *cfqq = RQ_CFQQ(rq);
616 	struct cfq_data *cfqd = cfqq->cfqd;
617 	struct request *__alias;
618 
619 	cfqq->queued[rq_is_sync(rq)]++;
620 
621 	/*
622 	 * looks a little odd, but the first insert might return an alias.
623 	 * if that happens, put the alias on the dispatch list
624 	 */
625 	while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
626 		cfq_dispatch_insert(cfqd->queue, __alias);
627 
628 	if (!cfq_cfqq_on_rr(cfqq))
629 		cfq_add_cfqq_rr(cfqd, cfqq);
630 
631 	/*
632 	 * check if this request is a better next-serve candidate
633 	 */
634 	cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
635 	BUG_ON(!cfqq->next_rq);
636 }
637 
cfq_reposition_rq_rb(struct cfq_queue * cfqq,struct request * rq)638 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
639 {
640 	elv_rb_del(&cfqq->sort_list, rq);
641 	cfqq->queued[rq_is_sync(rq)]--;
642 	cfq_add_rq_rb(rq);
643 }
644 
645 static struct request *
cfq_find_rq_fmerge(struct cfq_data * cfqd,struct bio * bio)646 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
647 {
648 	struct task_struct *tsk = current;
649 	struct cfq_io_context *cic;
650 	struct cfq_queue *cfqq;
651 
652 	cic = cfq_cic_lookup(cfqd, tsk->io_context);
653 	if (!cic)
654 		return NULL;
655 
656 	cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
657 	if (cfqq) {
658 		sector_t sector = bio->bi_sector + bio_sectors(bio);
659 
660 		return elv_rb_find(&cfqq->sort_list, sector);
661 	}
662 
663 	return NULL;
664 }
665 
cfq_activate_request(struct request_queue * q,struct request * rq)666 static void cfq_activate_request(struct request_queue *q, struct request *rq)
667 {
668 	struct cfq_data *cfqd = q->elevator->elevator_data;
669 
670 	cfqd->rq_in_driver++;
671 	cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
672 						cfqd->rq_in_driver);
673 
674 	cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
675 }
676 
cfq_deactivate_request(struct request_queue * q,struct request * rq)677 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
678 {
679 	struct cfq_data *cfqd = q->elevator->elevator_data;
680 
681 	WARN_ON(!cfqd->rq_in_driver);
682 	cfqd->rq_in_driver--;
683 	cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
684 						cfqd->rq_in_driver);
685 }
686 
cfq_remove_request(struct request * rq)687 static void cfq_remove_request(struct request *rq)
688 {
689 	struct cfq_queue *cfqq = RQ_CFQQ(rq);
690 
691 	if (cfqq->next_rq == rq)
692 		cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
693 
694 	list_del_init(&rq->queuelist);
695 	cfq_del_rq_rb(rq);
696 
697 	cfqq->cfqd->rq_queued--;
698 	if (rq_is_meta(rq)) {
699 		WARN_ON(!cfqq->meta_pending);
700 		cfqq->meta_pending--;
701 	}
702 }
703 
cfq_merge(struct request_queue * q,struct request ** req,struct bio * bio)704 static int cfq_merge(struct request_queue *q, struct request **req,
705 		     struct bio *bio)
706 {
707 	struct cfq_data *cfqd = q->elevator->elevator_data;
708 	struct request *__rq;
709 
710 	__rq = cfq_find_rq_fmerge(cfqd, bio);
711 	if (__rq && elv_rq_merge_ok(__rq, bio)) {
712 		*req = __rq;
713 		return ELEVATOR_FRONT_MERGE;
714 	}
715 
716 	return ELEVATOR_NO_MERGE;
717 }
718 
cfq_merged_request(struct request_queue * q,struct request * req,int type)719 static void cfq_merged_request(struct request_queue *q, struct request *req,
720 			       int type)
721 {
722 	if (type == ELEVATOR_FRONT_MERGE) {
723 		struct cfq_queue *cfqq = RQ_CFQQ(req);
724 
725 		cfq_reposition_rq_rb(cfqq, req);
726 	}
727 }
728 
729 static void
cfq_merged_requests(struct request_queue * q,struct request * rq,struct request * next)730 cfq_merged_requests(struct request_queue *q, struct request *rq,
731 		    struct request *next)
732 {
733 	/*
734 	 * reposition in fifo if next is older than rq
735 	 */
736 	if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
737 	    time_before(next->start_time, rq->start_time))
738 		list_move(&rq->queuelist, &next->queuelist);
739 
740 	cfq_remove_request(next);
741 }
742 
cfq_allow_merge(struct request_queue * q,struct request * rq,struct bio * bio)743 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
744 			   struct bio *bio)
745 {
746 	struct cfq_data *cfqd = q->elevator->elevator_data;
747 	struct cfq_io_context *cic;
748 	struct cfq_queue *cfqq;
749 
750 	/*
751 	 * Disallow merge of a sync bio into an async request.
752 	 */
753 	if (cfq_bio_sync(bio) && !rq_is_sync(rq))
754 		return 0;
755 
756 	/*
757 	 * Lookup the cfqq that this bio will be queued with. Allow
758 	 * merge only if rq is queued there.
759 	 */
760 	cic = cfq_cic_lookup(cfqd, current->io_context);
761 	if (!cic)
762 		return 0;
763 
764 	cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
765 	if (cfqq == RQ_CFQQ(rq))
766 		return 1;
767 
768 	return 0;
769 }
770 
__cfq_set_active_queue(struct cfq_data * cfqd,struct cfq_queue * cfqq)771 static void __cfq_set_active_queue(struct cfq_data *cfqd,
772 				   struct cfq_queue *cfqq)
773 {
774 	if (cfqq) {
775 		cfq_log_cfqq(cfqd, cfqq, "set_active");
776 		cfqq->slice_end = 0;
777 		cfq_clear_cfqq_must_alloc_slice(cfqq);
778 		cfq_clear_cfqq_fifo_expire(cfqq);
779 		cfq_mark_cfqq_slice_new(cfqq);
780 		cfq_clear_cfqq_queue_new(cfqq);
781 	}
782 
783 	cfqd->active_queue = cfqq;
784 }
785 
786 /*
787  * current cfqq expired its slice (or was too idle), select new one
788  */
789 static void
__cfq_slice_expired(struct cfq_data * cfqd,struct cfq_queue * cfqq,int timed_out)790 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
791 		    int timed_out)
792 {
793 	cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
794 
795 	if (cfq_cfqq_wait_request(cfqq))
796 		del_timer(&cfqd->idle_slice_timer);
797 
798 	cfq_clear_cfqq_must_dispatch(cfqq);
799 	cfq_clear_cfqq_wait_request(cfqq);
800 
801 	/*
802 	 * store what was left of this slice, if the queue idled/timed out
803 	 */
804 	if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
805 		cfqq->slice_resid = cfqq->slice_end - jiffies;
806 		cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
807 	}
808 
809 	cfq_resort_rr_list(cfqd, cfqq);
810 
811 	if (cfqq == cfqd->active_queue)
812 		cfqd->active_queue = NULL;
813 
814 	if (cfqd->active_cic) {
815 		put_io_context(cfqd->active_cic->ioc);
816 		cfqd->active_cic = NULL;
817 	}
818 }
819 
cfq_slice_expired(struct cfq_data * cfqd,int timed_out)820 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
821 {
822 	struct cfq_queue *cfqq = cfqd->active_queue;
823 
824 	if (cfqq)
825 		__cfq_slice_expired(cfqd, cfqq, timed_out);
826 }
827 
828 /*
829  * Get next queue for service. Unless we have a queue preemption,
830  * we'll simply select the first cfqq in the service tree.
831  */
cfq_get_next_queue(struct cfq_data * cfqd)832 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
833 {
834 	if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
835 		return NULL;
836 
837 	return cfq_rb_first(&cfqd->service_tree);
838 }
839 
840 /*
841  * Get and set a new active queue for service.
842  */
cfq_set_active_queue(struct cfq_data * cfqd)843 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
844 {
845 	struct cfq_queue *cfqq;
846 
847 	cfqq = cfq_get_next_queue(cfqd);
848 	__cfq_set_active_queue(cfqd, cfqq);
849 	return cfqq;
850 }
851 
cfq_dist_from_last(struct cfq_data * cfqd,struct request * rq)852 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
853 					  struct request *rq)
854 {
855 	if (rq->sector >= cfqd->last_position)
856 		return rq->sector - cfqd->last_position;
857 	else
858 		return cfqd->last_position - rq->sector;
859 }
860 
cfq_rq_close(struct cfq_data * cfqd,struct request * rq)861 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
862 {
863 	struct cfq_io_context *cic = cfqd->active_cic;
864 
865 	if (!sample_valid(cic->seek_samples))
866 		return 0;
867 
868 	return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
869 }
870 
cfq_close_cooperator(struct cfq_data * cfq_data,struct cfq_queue * cfqq)871 static int cfq_close_cooperator(struct cfq_data *cfq_data,
872 				struct cfq_queue *cfqq)
873 {
874 	/*
875 	 * We should notice if some of the queues are cooperating, eg
876 	 * working closely on the same area of the disk. In that case,
877 	 * we can group them together and don't waste time idling.
878 	 */
879 	return 0;
880 }
881 
882 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
883 
cfq_arm_slice_timer(struct cfq_data * cfqd)884 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
885 {
886 	struct cfq_queue *cfqq = cfqd->active_queue;
887 	struct cfq_io_context *cic;
888 	unsigned long sl;
889 
890 	/*
891 	 * SSD device without seek penalty, disable idling. But only do so
892 	 * for devices that support queuing, otherwise we still have a problem
893 	 * with sync vs async workloads.
894 	 */
895 	if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
896 		return;
897 
898 	WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
899 	WARN_ON(cfq_cfqq_slice_new(cfqq));
900 
901 	/*
902 	 * idle is disabled, either manually or by past process history
903 	 */
904 	if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
905 		return;
906 
907 	/*
908 	 * still requests with the driver, don't idle
909 	 */
910 	if (cfqd->rq_in_driver)
911 		return;
912 
913 	/*
914 	 * task has exited, don't wait
915 	 */
916 	cic = cfqd->active_cic;
917 	if (!cic || !atomic_read(&cic->ioc->nr_tasks))
918 		return;
919 
920 	/*
921 	 * See if this prio level has a good candidate
922 	 */
923 	if (cfq_close_cooperator(cfqd, cfqq) &&
924 	    (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
925 		return;
926 
927 	cfq_mark_cfqq_must_dispatch(cfqq);
928 	cfq_mark_cfqq_wait_request(cfqq);
929 
930 	/*
931 	 * we don't want to idle for seeks, but we do want to allow
932 	 * fair distribution of slice time for a process doing back-to-back
933 	 * seeks. so allow a little bit of time for him to submit a new rq
934 	 */
935 	sl = cfqd->cfq_slice_idle;
936 	if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
937 		sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
938 
939 	mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
940 	cfq_log(cfqd, "arm_idle: %lu", sl);
941 }
942 
943 /*
944  * Move request from internal lists to the request queue dispatch list.
945  */
cfq_dispatch_insert(struct request_queue * q,struct request * rq)946 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
947 {
948 	struct cfq_data *cfqd = q->elevator->elevator_data;
949 	struct cfq_queue *cfqq = RQ_CFQQ(rq);
950 
951 	cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
952 
953 	cfq_remove_request(rq);
954 	cfqq->dispatched++;
955 	elv_dispatch_sort(q, rq);
956 
957 	if (cfq_cfqq_sync(cfqq))
958 		cfqd->sync_flight++;
959 }
960 
961 /*
962  * return expired entry, or NULL to just start from scratch in rbtree
963  */
cfq_check_fifo(struct cfq_queue * cfqq)964 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
965 {
966 	struct cfq_data *cfqd = cfqq->cfqd;
967 	struct request *rq;
968 	int fifo;
969 
970 	if (cfq_cfqq_fifo_expire(cfqq))
971 		return NULL;
972 
973 	cfq_mark_cfqq_fifo_expire(cfqq);
974 
975 	if (list_empty(&cfqq->fifo))
976 		return NULL;
977 
978 	fifo = cfq_cfqq_sync(cfqq);
979 	rq = rq_entry_fifo(cfqq->fifo.next);
980 
981 	if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
982 		rq = NULL;
983 
984 	cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
985 	return rq;
986 }
987 
988 static inline int
cfq_prio_to_maxrq(struct cfq_data * cfqd,struct cfq_queue * cfqq)989 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
990 {
991 	const int base_rq = cfqd->cfq_slice_async_rq;
992 
993 	WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
994 
995 	return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
996 }
997 
998 /*
999  * Select a queue for service. If we have a current active queue,
1000  * check whether to continue servicing it, or retrieve and set a new one.
1001  */
cfq_select_queue(struct cfq_data * cfqd)1002 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1003 {
1004 	struct cfq_queue *cfqq;
1005 
1006 	cfqq = cfqd->active_queue;
1007 	if (!cfqq)
1008 		goto new_queue;
1009 
1010 	/*
1011 	 * The active queue has run out of time, expire it and select new.
1012 	 */
1013 	if (cfq_slice_used(cfqq))
1014 		goto expire;
1015 
1016 	/*
1017 	 * If we have a RT cfqq waiting, then we pre-empt the current non-rt
1018 	 * cfqq.
1019 	 */
1020 	if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues) {
1021 		/*
1022 		 * We simulate this as cfqq timed out so that it gets to bank
1023 		 * the remaining of its time slice.
1024 		 */
1025 		cfq_log_cfqq(cfqd, cfqq, "preempt");
1026 		cfq_slice_expired(cfqd, 1);
1027 		goto new_queue;
1028 	}
1029 
1030 	/*
1031 	 * The active queue has requests and isn't expired, allow it to
1032 	 * dispatch.
1033 	 */
1034 	if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1035 		goto keep_queue;
1036 
1037 	/*
1038 	 * No requests pending. If the active queue still has requests in
1039 	 * flight or is idling for a new request, allow either of these
1040 	 * conditions to happen (or time out) before selecting a new queue.
1041 	 */
1042 	if (timer_pending(&cfqd->idle_slice_timer) ||
1043 	    (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1044 		cfqq = NULL;
1045 		goto keep_queue;
1046 	}
1047 
1048 expire:
1049 	cfq_slice_expired(cfqd, 0);
1050 new_queue:
1051 	cfqq = cfq_set_active_queue(cfqd);
1052 keep_queue:
1053 	return cfqq;
1054 }
1055 
1056 /*
1057  * Dispatch some requests from cfqq, moving them to the request queue
1058  * dispatch list.
1059  */
1060 static int
__cfq_dispatch_requests(struct cfq_data * cfqd,struct cfq_queue * cfqq,int max_dispatch)1061 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1062 			int max_dispatch)
1063 {
1064 	int dispatched = 0;
1065 
1066 	BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1067 
1068 	do {
1069 		struct request *rq;
1070 
1071 		/*
1072 		 * follow expired path, else get first next available
1073 		 */
1074 		rq = cfq_check_fifo(cfqq);
1075 		if (rq == NULL)
1076 			rq = cfqq->next_rq;
1077 
1078 		/*
1079 		 * finally, insert request into driver dispatch list
1080 		 */
1081 		cfq_dispatch_insert(cfqd->queue, rq);
1082 
1083 		dispatched++;
1084 
1085 		if (!cfqd->active_cic) {
1086 			atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1087 			cfqd->active_cic = RQ_CIC(rq);
1088 		}
1089 
1090 		if (RB_EMPTY_ROOT(&cfqq->sort_list))
1091 			break;
1092 
1093 		/*
1094 		 * If there is a non-empty RT cfqq waiting for current
1095 		 * cfqq's timeslice to complete, pre-empt this cfqq
1096 		 */
1097 		if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues)
1098 			break;
1099 
1100 	} while (dispatched < max_dispatch);
1101 
1102 	/*
1103 	 * expire an async queue immediately if it has used up its slice. idle
1104 	 * queue always expire after 1 dispatch round.
1105 	 */
1106 	if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1107 	    dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1108 	    cfq_class_idle(cfqq))) {
1109 		cfqq->slice_end = jiffies + 1;
1110 		cfq_slice_expired(cfqd, 0);
1111 	}
1112 
1113 	return dispatched;
1114 }
1115 
__cfq_forced_dispatch_cfqq(struct cfq_queue * cfqq)1116 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1117 {
1118 	int dispatched = 0;
1119 
1120 	while (cfqq->next_rq) {
1121 		cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1122 		dispatched++;
1123 	}
1124 
1125 	BUG_ON(!list_empty(&cfqq->fifo));
1126 	return dispatched;
1127 }
1128 
1129 /*
1130  * Drain our current requests. Used for barriers and when switching
1131  * io schedulers on-the-fly.
1132  */
cfq_forced_dispatch(struct cfq_data * cfqd)1133 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1134 {
1135 	struct cfq_queue *cfqq;
1136 	int dispatched = 0;
1137 
1138 	while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1139 		dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1140 
1141 	cfq_slice_expired(cfqd, 0);
1142 
1143 	BUG_ON(cfqd->busy_queues);
1144 
1145 	cfq_log(cfqd, "forced_dispatch=%d\n", dispatched);
1146 	return dispatched;
1147 }
1148 
cfq_dispatch_requests(struct request_queue * q,int force)1149 static int cfq_dispatch_requests(struct request_queue *q, int force)
1150 {
1151 	struct cfq_data *cfqd = q->elevator->elevator_data;
1152 	struct cfq_queue *cfqq;
1153 	int dispatched;
1154 
1155 	if (!cfqd->busy_queues)
1156 		return 0;
1157 
1158 	if (unlikely(force))
1159 		return cfq_forced_dispatch(cfqd);
1160 
1161 	dispatched = 0;
1162 	while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1163 		int max_dispatch;
1164 
1165 		max_dispatch = cfqd->cfq_quantum;
1166 		if (cfq_class_idle(cfqq))
1167 			max_dispatch = 1;
1168 
1169 		if (cfqq->dispatched >= max_dispatch && cfqd->busy_queues > 1)
1170 			break;
1171 
1172 		if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1173 			break;
1174 
1175 		cfq_clear_cfqq_must_dispatch(cfqq);
1176 		cfq_clear_cfqq_wait_request(cfqq);
1177 		del_timer(&cfqd->idle_slice_timer);
1178 
1179 		dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1180 	}
1181 
1182 	cfq_log(cfqd, "dispatched=%d", dispatched);
1183 	return dispatched;
1184 }
1185 
1186 /*
1187  * task holds one reference to the queue, dropped when task exits. each rq
1188  * in-flight on this queue also holds a reference, dropped when rq is freed.
1189  *
1190  * queue lock must be held here.
1191  */
cfq_put_queue(struct cfq_queue * cfqq)1192 static void cfq_put_queue(struct cfq_queue *cfqq)
1193 {
1194 	struct cfq_data *cfqd = cfqq->cfqd;
1195 
1196 	BUG_ON(atomic_read(&cfqq->ref) <= 0);
1197 
1198 	if (!atomic_dec_and_test(&cfqq->ref))
1199 		return;
1200 
1201 	cfq_log_cfqq(cfqd, cfqq, "put_queue");
1202 	BUG_ON(rb_first(&cfqq->sort_list));
1203 	BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1204 	BUG_ON(cfq_cfqq_on_rr(cfqq));
1205 
1206 	if (unlikely(cfqd->active_queue == cfqq)) {
1207 		__cfq_slice_expired(cfqd, cfqq, 0);
1208 		cfq_schedule_dispatch(cfqd);
1209 	}
1210 
1211 	kmem_cache_free(cfq_pool, cfqq);
1212 }
1213 
1214 /*
1215  * Must always be called with the rcu_read_lock() held
1216  */
1217 static void
__call_for_each_cic(struct io_context * ioc,void (* func)(struct io_context *,struct cfq_io_context *))1218 __call_for_each_cic(struct io_context *ioc,
1219 		    void (*func)(struct io_context *, struct cfq_io_context *))
1220 {
1221 	struct cfq_io_context *cic;
1222 	struct hlist_node *n;
1223 
1224 	hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1225 		func(ioc, cic);
1226 }
1227 
1228 /*
1229  * Call func for each cic attached to this ioc.
1230  */
1231 static void
call_for_each_cic(struct io_context * ioc,void (* func)(struct io_context *,struct cfq_io_context *))1232 call_for_each_cic(struct io_context *ioc,
1233 		  void (*func)(struct io_context *, struct cfq_io_context *))
1234 {
1235 	rcu_read_lock();
1236 	__call_for_each_cic(ioc, func);
1237 	rcu_read_unlock();
1238 }
1239 
cfq_cic_free_rcu(struct rcu_head * head)1240 static void cfq_cic_free_rcu(struct rcu_head *head)
1241 {
1242 	struct cfq_io_context *cic;
1243 
1244 	cic = container_of(head, struct cfq_io_context, rcu_head);
1245 
1246 	kmem_cache_free(cfq_ioc_pool, cic);
1247 	elv_ioc_count_dec(ioc_count);
1248 
1249 	if (ioc_gone) {
1250 		/*
1251 		 * CFQ scheduler is exiting, grab exit lock and check
1252 		 * the pending io context count. If it hits zero,
1253 		 * complete ioc_gone and set it back to NULL
1254 		 */
1255 		spin_lock(&ioc_gone_lock);
1256 		if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1257 			complete(ioc_gone);
1258 			ioc_gone = NULL;
1259 		}
1260 		spin_unlock(&ioc_gone_lock);
1261 	}
1262 }
1263 
cfq_cic_free(struct cfq_io_context * cic)1264 static void cfq_cic_free(struct cfq_io_context *cic)
1265 {
1266 	call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1267 }
1268 
cic_free_func(struct io_context * ioc,struct cfq_io_context * cic)1269 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1270 {
1271 	unsigned long flags;
1272 
1273 	BUG_ON(!cic->dead_key);
1274 
1275 	spin_lock_irqsave(&ioc->lock, flags);
1276 	radix_tree_delete(&ioc->radix_root, cic->dead_key);
1277 	hlist_del_rcu(&cic->cic_list);
1278 	spin_unlock_irqrestore(&ioc->lock, flags);
1279 
1280 	cfq_cic_free(cic);
1281 }
1282 
1283 /*
1284  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1285  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1286  * and ->trim() which is called with the task lock held
1287  */
cfq_free_io_context(struct io_context * ioc)1288 static void cfq_free_io_context(struct io_context *ioc)
1289 {
1290 	/*
1291 	 * ioc->refcount is zero here, or we are called from elv_unregister(),
1292 	 * so no more cic's are allowed to be linked into this ioc.  So it
1293 	 * should be ok to iterate over the known list, we will see all cic's
1294 	 * since no new ones are added.
1295 	 */
1296 	__call_for_each_cic(ioc, cic_free_func);
1297 }
1298 
cfq_exit_cfqq(struct cfq_data * cfqd,struct cfq_queue * cfqq)1299 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1300 {
1301 	if (unlikely(cfqq == cfqd->active_queue)) {
1302 		__cfq_slice_expired(cfqd, cfqq, 0);
1303 		cfq_schedule_dispatch(cfqd);
1304 	}
1305 
1306 	cfq_put_queue(cfqq);
1307 }
1308 
__cfq_exit_single_io_context(struct cfq_data * cfqd,struct cfq_io_context * cic)1309 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1310 					 struct cfq_io_context *cic)
1311 {
1312 	struct io_context *ioc = cic->ioc;
1313 
1314 	list_del_init(&cic->queue_list);
1315 
1316 	/*
1317 	 * Make sure key == NULL is seen for dead queues
1318 	 */
1319 	smp_wmb();
1320 	cic->dead_key = (unsigned long) cic->key;
1321 	cic->key = NULL;
1322 
1323 	if (ioc->ioc_data == cic)
1324 		rcu_assign_pointer(ioc->ioc_data, NULL);
1325 
1326 	if (cic->cfqq[ASYNC]) {
1327 		cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1328 		cic->cfqq[ASYNC] = NULL;
1329 	}
1330 
1331 	if (cic->cfqq[SYNC]) {
1332 		cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1333 		cic->cfqq[SYNC] = NULL;
1334 	}
1335 }
1336 
cfq_exit_single_io_context(struct io_context * ioc,struct cfq_io_context * cic)1337 static void cfq_exit_single_io_context(struct io_context *ioc,
1338 				       struct cfq_io_context *cic)
1339 {
1340 	struct cfq_data *cfqd = cic->key;
1341 
1342 	if (cfqd) {
1343 		struct request_queue *q = cfqd->queue;
1344 		unsigned long flags;
1345 
1346 		spin_lock_irqsave(q->queue_lock, flags);
1347 
1348 		/*
1349 		 * Ensure we get a fresh copy of the ->key to prevent
1350 		 * race between exiting task and queue
1351 		 */
1352 		smp_read_barrier_depends();
1353 		if (cic->key)
1354 			__cfq_exit_single_io_context(cfqd, cic);
1355 
1356 		spin_unlock_irqrestore(q->queue_lock, flags);
1357 	}
1358 }
1359 
1360 /*
1361  * The process that ioc belongs to has exited, we need to clean up
1362  * and put the internal structures we have that belongs to that process.
1363  */
cfq_exit_io_context(struct io_context * ioc)1364 static void cfq_exit_io_context(struct io_context *ioc)
1365 {
1366 	call_for_each_cic(ioc, cfq_exit_single_io_context);
1367 }
1368 
1369 static struct cfq_io_context *
cfq_alloc_io_context(struct cfq_data * cfqd,gfp_t gfp_mask)1370 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1371 {
1372 	struct cfq_io_context *cic;
1373 
1374 	cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1375 							cfqd->queue->node);
1376 	if (cic) {
1377 		cic->last_end_request = jiffies;
1378 		INIT_LIST_HEAD(&cic->queue_list);
1379 		INIT_HLIST_NODE(&cic->cic_list);
1380 		cic->dtor = cfq_free_io_context;
1381 		cic->exit = cfq_exit_io_context;
1382 		elv_ioc_count_inc(ioc_count);
1383 	}
1384 
1385 	return cic;
1386 }
1387 
cfq_init_prio_data(struct cfq_queue * cfqq,struct io_context * ioc)1388 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1389 {
1390 	struct task_struct *tsk = current;
1391 	int ioprio_class;
1392 
1393 	if (!cfq_cfqq_prio_changed(cfqq))
1394 		return;
1395 
1396 	ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1397 	switch (ioprio_class) {
1398 	default:
1399 		printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1400 	case IOPRIO_CLASS_NONE:
1401 		/*
1402 		 * no prio set, inherit CPU scheduling settings
1403 		 */
1404 		cfqq->ioprio = task_nice_ioprio(tsk);
1405 		cfqq->ioprio_class = task_nice_ioclass(tsk);
1406 		break;
1407 	case IOPRIO_CLASS_RT:
1408 		cfqq->ioprio = task_ioprio(ioc);
1409 		cfqq->ioprio_class = IOPRIO_CLASS_RT;
1410 		break;
1411 	case IOPRIO_CLASS_BE:
1412 		cfqq->ioprio = task_ioprio(ioc);
1413 		cfqq->ioprio_class = IOPRIO_CLASS_BE;
1414 		break;
1415 	case IOPRIO_CLASS_IDLE:
1416 		cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1417 		cfqq->ioprio = 7;
1418 		cfq_clear_cfqq_idle_window(cfqq);
1419 		break;
1420 	}
1421 
1422 	/*
1423 	 * keep track of original prio settings in case we have to temporarily
1424 	 * elevate the priority of this queue
1425 	 */
1426 	cfqq->org_ioprio = cfqq->ioprio;
1427 	cfqq->org_ioprio_class = cfqq->ioprio_class;
1428 	cfq_clear_cfqq_prio_changed(cfqq);
1429 }
1430 
changed_ioprio(struct io_context * ioc,struct cfq_io_context * cic)1431 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1432 {
1433 	struct cfq_data *cfqd = cic->key;
1434 	struct cfq_queue *cfqq;
1435 	unsigned long flags;
1436 
1437 	if (unlikely(!cfqd))
1438 		return;
1439 
1440 	spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1441 
1442 	cfqq = cic->cfqq[ASYNC];
1443 	if (cfqq) {
1444 		struct cfq_queue *new_cfqq;
1445 		new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
1446 		if (new_cfqq) {
1447 			cic->cfqq[ASYNC] = new_cfqq;
1448 			cfq_put_queue(cfqq);
1449 		}
1450 	}
1451 
1452 	cfqq = cic->cfqq[SYNC];
1453 	if (cfqq)
1454 		cfq_mark_cfqq_prio_changed(cfqq);
1455 
1456 	spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1457 }
1458 
cfq_ioc_set_ioprio(struct io_context * ioc)1459 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1460 {
1461 	call_for_each_cic(ioc, changed_ioprio);
1462 	ioc->ioprio_changed = 0;
1463 }
1464 
1465 static struct cfq_queue *
cfq_find_alloc_queue(struct cfq_data * cfqd,int is_sync,struct io_context * ioc,gfp_t gfp_mask)1466 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1467 		     struct io_context *ioc, gfp_t gfp_mask)
1468 {
1469 	struct cfq_queue *cfqq, *new_cfqq = NULL;
1470 	struct cfq_io_context *cic;
1471 
1472 retry:
1473 	cic = cfq_cic_lookup(cfqd, ioc);
1474 	/* cic always exists here */
1475 	cfqq = cic_to_cfqq(cic, is_sync);
1476 
1477 	if (!cfqq) {
1478 		if (new_cfqq) {
1479 			cfqq = new_cfqq;
1480 			new_cfqq = NULL;
1481 		} else if (gfp_mask & __GFP_WAIT) {
1482 			/*
1483 			 * Inform the allocator of the fact that we will
1484 			 * just repeat this allocation if it fails, to allow
1485 			 * the allocator to do whatever it needs to attempt to
1486 			 * free memory.
1487 			 */
1488 			spin_unlock_irq(cfqd->queue->queue_lock);
1489 			new_cfqq = kmem_cache_alloc_node(cfq_pool,
1490 					gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1491 					cfqd->queue->node);
1492 			spin_lock_irq(cfqd->queue->queue_lock);
1493 			goto retry;
1494 		} else {
1495 			cfqq = kmem_cache_alloc_node(cfq_pool,
1496 					gfp_mask | __GFP_ZERO,
1497 					cfqd->queue->node);
1498 			if (!cfqq)
1499 				goto out;
1500 		}
1501 
1502 		RB_CLEAR_NODE(&cfqq->rb_node);
1503 		INIT_LIST_HEAD(&cfqq->fifo);
1504 
1505 		atomic_set(&cfqq->ref, 0);
1506 		cfqq->cfqd = cfqd;
1507 
1508 		cfq_mark_cfqq_prio_changed(cfqq);
1509 		cfq_mark_cfqq_queue_new(cfqq);
1510 
1511 		cfq_init_prio_data(cfqq, ioc);
1512 
1513 		if (is_sync) {
1514 			if (!cfq_class_idle(cfqq))
1515 				cfq_mark_cfqq_idle_window(cfqq);
1516 			cfq_mark_cfqq_sync(cfqq);
1517 		}
1518 		cfqq->pid = current->pid;
1519 		cfq_log_cfqq(cfqd, cfqq, "alloced");
1520 	}
1521 
1522 	if (new_cfqq)
1523 		kmem_cache_free(cfq_pool, new_cfqq);
1524 
1525 out:
1526 	WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1527 	return cfqq;
1528 }
1529 
1530 static struct cfq_queue **
cfq_async_queue_prio(struct cfq_data * cfqd,int ioprio_class,int ioprio)1531 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1532 {
1533 	switch (ioprio_class) {
1534 	case IOPRIO_CLASS_RT:
1535 		return &cfqd->async_cfqq[0][ioprio];
1536 	case IOPRIO_CLASS_BE:
1537 		return &cfqd->async_cfqq[1][ioprio];
1538 	case IOPRIO_CLASS_IDLE:
1539 		return &cfqd->async_idle_cfqq;
1540 	default:
1541 		BUG();
1542 	}
1543 }
1544 
1545 static struct cfq_queue *
cfq_get_queue(struct cfq_data * cfqd,int is_sync,struct io_context * ioc,gfp_t gfp_mask)1546 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1547 	      gfp_t gfp_mask)
1548 {
1549 	const int ioprio = task_ioprio(ioc);
1550 	const int ioprio_class = task_ioprio_class(ioc);
1551 	struct cfq_queue **async_cfqq = NULL;
1552 	struct cfq_queue *cfqq = NULL;
1553 
1554 	if (!is_sync) {
1555 		async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1556 		cfqq = *async_cfqq;
1557 	}
1558 
1559 	if (!cfqq) {
1560 		cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1561 		if (!cfqq)
1562 			return NULL;
1563 	}
1564 
1565 	/*
1566 	 * pin the queue now that it's allocated, scheduler exit will prune it
1567 	 */
1568 	if (!is_sync && !(*async_cfqq)) {
1569 		atomic_inc(&cfqq->ref);
1570 		*async_cfqq = cfqq;
1571 	}
1572 
1573 	atomic_inc(&cfqq->ref);
1574 	return cfqq;
1575 }
1576 
1577 /*
1578  * We drop cfq io contexts lazily, so we may find a dead one.
1579  */
1580 static void
cfq_drop_dead_cic(struct cfq_data * cfqd,struct io_context * ioc,struct cfq_io_context * cic)1581 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1582 		  struct cfq_io_context *cic)
1583 {
1584 	unsigned long flags;
1585 
1586 	WARN_ON(!list_empty(&cic->queue_list));
1587 
1588 	spin_lock_irqsave(&ioc->lock, flags);
1589 
1590 	BUG_ON(ioc->ioc_data == cic);
1591 
1592 	radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1593 	hlist_del_rcu(&cic->cic_list);
1594 	spin_unlock_irqrestore(&ioc->lock, flags);
1595 
1596 	cfq_cic_free(cic);
1597 }
1598 
1599 static struct cfq_io_context *
cfq_cic_lookup(struct cfq_data * cfqd,struct io_context * ioc)1600 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1601 {
1602 	struct cfq_io_context *cic;
1603 	unsigned long flags;
1604 	void *k;
1605 
1606 	if (unlikely(!ioc))
1607 		return NULL;
1608 
1609 	rcu_read_lock();
1610 
1611 	/*
1612 	 * we maintain a last-hit cache, to avoid browsing over the tree
1613 	 */
1614 	cic = rcu_dereference(ioc->ioc_data);
1615 	if (cic && cic->key == cfqd) {
1616 		rcu_read_unlock();
1617 		return cic;
1618 	}
1619 
1620 	do {
1621 		cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1622 		rcu_read_unlock();
1623 		if (!cic)
1624 			break;
1625 		/* ->key must be copied to avoid race with cfq_exit_queue() */
1626 		k = cic->key;
1627 		if (unlikely(!k)) {
1628 			cfq_drop_dead_cic(cfqd, ioc, cic);
1629 			rcu_read_lock();
1630 			continue;
1631 		}
1632 
1633 		spin_lock_irqsave(&ioc->lock, flags);
1634 		rcu_assign_pointer(ioc->ioc_data, cic);
1635 		spin_unlock_irqrestore(&ioc->lock, flags);
1636 		break;
1637 	} while (1);
1638 
1639 	return cic;
1640 }
1641 
1642 /*
1643  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1644  * the process specific cfq io context when entered from the block layer.
1645  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1646  */
cfq_cic_link(struct cfq_data * cfqd,struct io_context * ioc,struct cfq_io_context * cic,gfp_t gfp_mask)1647 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1648 			struct cfq_io_context *cic, gfp_t gfp_mask)
1649 {
1650 	unsigned long flags;
1651 	int ret;
1652 
1653 	ret = radix_tree_preload(gfp_mask);
1654 	if (!ret) {
1655 		cic->ioc = ioc;
1656 		cic->key = cfqd;
1657 
1658 		spin_lock_irqsave(&ioc->lock, flags);
1659 		ret = radix_tree_insert(&ioc->radix_root,
1660 						(unsigned long) cfqd, cic);
1661 		if (!ret)
1662 			hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1663 		spin_unlock_irqrestore(&ioc->lock, flags);
1664 
1665 		radix_tree_preload_end();
1666 
1667 		if (!ret) {
1668 			spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1669 			list_add(&cic->queue_list, &cfqd->cic_list);
1670 			spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1671 		}
1672 	}
1673 
1674 	if (ret)
1675 		printk(KERN_ERR "cfq: cic link failed!\n");
1676 
1677 	return ret;
1678 }
1679 
1680 /*
1681  * Setup general io context and cfq io context. There can be several cfq
1682  * io contexts per general io context, if this process is doing io to more
1683  * than one device managed by cfq.
1684  */
1685 static struct cfq_io_context *
cfq_get_io_context(struct cfq_data * cfqd,gfp_t gfp_mask)1686 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1687 {
1688 	struct io_context *ioc = NULL;
1689 	struct cfq_io_context *cic;
1690 
1691 	might_sleep_if(gfp_mask & __GFP_WAIT);
1692 
1693 	ioc = get_io_context(gfp_mask, cfqd->queue->node);
1694 	if (!ioc)
1695 		return NULL;
1696 
1697 	cic = cfq_cic_lookup(cfqd, ioc);
1698 	if (cic)
1699 		goto out;
1700 
1701 	cic = cfq_alloc_io_context(cfqd, gfp_mask);
1702 	if (cic == NULL)
1703 		goto err;
1704 
1705 	if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1706 		goto err_free;
1707 
1708 out:
1709 	smp_read_barrier_depends();
1710 	if (unlikely(ioc->ioprio_changed))
1711 		cfq_ioc_set_ioprio(ioc);
1712 
1713 	return cic;
1714 err_free:
1715 	cfq_cic_free(cic);
1716 err:
1717 	put_io_context(ioc);
1718 	return NULL;
1719 }
1720 
1721 static void
cfq_update_io_thinktime(struct cfq_data * cfqd,struct cfq_io_context * cic)1722 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1723 {
1724 	unsigned long elapsed = jiffies - cic->last_end_request;
1725 	unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1726 
1727 	cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1728 	cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1729 	cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1730 }
1731 
1732 static void
cfq_update_io_seektime(struct cfq_data * cfqd,struct cfq_io_context * cic,struct request * rq)1733 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1734 		       struct request *rq)
1735 {
1736 	sector_t sdist;
1737 	u64 total;
1738 
1739 	if (cic->last_request_pos < rq->sector)
1740 		sdist = rq->sector - cic->last_request_pos;
1741 	else
1742 		sdist = cic->last_request_pos - rq->sector;
1743 
1744 	/*
1745 	 * Don't allow the seek distance to get too large from the
1746 	 * odd fragment, pagein, etc
1747 	 */
1748 	if (cic->seek_samples <= 60) /* second&third seek */
1749 		sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1750 	else
1751 		sdist = min(sdist, (cic->seek_mean * 4)	+ 2*1024*64);
1752 
1753 	cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1754 	cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1755 	total = cic->seek_total + (cic->seek_samples/2);
1756 	do_div(total, cic->seek_samples);
1757 	cic->seek_mean = (sector_t)total;
1758 }
1759 
1760 /*
1761  * Disable idle window if the process thinks too long or seeks so much that
1762  * it doesn't matter
1763  */
1764 static void
cfq_update_idle_window(struct cfq_data * cfqd,struct cfq_queue * cfqq,struct cfq_io_context * cic)1765 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1766 		       struct cfq_io_context *cic)
1767 {
1768 	int old_idle, enable_idle;
1769 
1770 	/*
1771 	 * Don't idle for async or idle io prio class
1772 	 */
1773 	if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1774 		return;
1775 
1776 	enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1777 
1778 	if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1779 	    (cfqd->hw_tag && CIC_SEEKY(cic)))
1780 		enable_idle = 0;
1781 	else if (sample_valid(cic->ttime_samples)) {
1782 		if (cic->ttime_mean > cfqd->cfq_slice_idle)
1783 			enable_idle = 0;
1784 		else
1785 			enable_idle = 1;
1786 	}
1787 
1788 	if (old_idle != enable_idle) {
1789 		cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1790 		if (enable_idle)
1791 			cfq_mark_cfqq_idle_window(cfqq);
1792 		else
1793 			cfq_clear_cfqq_idle_window(cfqq);
1794 	}
1795 }
1796 
1797 /*
1798  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1799  * no or if we aren't sure, a 1 will cause a preempt.
1800  */
1801 static int
cfq_should_preempt(struct cfq_data * cfqd,struct cfq_queue * new_cfqq,struct request * rq)1802 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1803 		   struct request *rq)
1804 {
1805 	struct cfq_queue *cfqq;
1806 
1807 	cfqq = cfqd->active_queue;
1808 	if (!cfqq)
1809 		return 0;
1810 
1811 	if (cfq_slice_used(cfqq))
1812 		return 1;
1813 
1814 	if (cfq_class_idle(new_cfqq))
1815 		return 0;
1816 
1817 	if (cfq_class_idle(cfqq))
1818 		return 1;
1819 
1820 	/*
1821 	 * if the new request is sync, but the currently running queue is
1822 	 * not, let the sync request have priority.
1823 	 */
1824 	if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1825 		return 1;
1826 
1827 	/*
1828 	 * So both queues are sync. Let the new request get disk time if
1829 	 * it's a metadata request and the current queue is doing regular IO.
1830 	 */
1831 	if (rq_is_meta(rq) && !cfqq->meta_pending)
1832 		return 1;
1833 
1834 	/*
1835 	 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
1836 	 */
1837 	if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
1838 		return 1;
1839 
1840 	if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1841 		return 0;
1842 
1843 	/*
1844 	 * if this request is as-good as one we would expect from the
1845 	 * current cfqq, let it preempt
1846 	 */
1847 	if (cfq_rq_close(cfqd, rq))
1848 		return 1;
1849 
1850 	return 0;
1851 }
1852 
1853 /*
1854  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1855  * let it have half of its nominal slice.
1856  */
cfq_preempt_queue(struct cfq_data * cfqd,struct cfq_queue * cfqq)1857 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1858 {
1859 	cfq_log_cfqq(cfqd, cfqq, "preempt");
1860 	cfq_slice_expired(cfqd, 1);
1861 
1862 	/*
1863 	 * Put the new queue at the front of the of the current list,
1864 	 * so we know that it will be selected next.
1865 	 */
1866 	BUG_ON(!cfq_cfqq_on_rr(cfqq));
1867 
1868 	cfq_service_tree_add(cfqd, cfqq, 1);
1869 
1870 	cfqq->slice_end = 0;
1871 	cfq_mark_cfqq_slice_new(cfqq);
1872 }
1873 
1874 /*
1875  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1876  * something we should do about it
1877  */
1878 static void
cfq_rq_enqueued(struct cfq_data * cfqd,struct cfq_queue * cfqq,struct request * rq)1879 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1880 		struct request *rq)
1881 {
1882 	struct cfq_io_context *cic = RQ_CIC(rq);
1883 
1884 	cfqd->rq_queued++;
1885 	if (rq_is_meta(rq))
1886 		cfqq->meta_pending++;
1887 
1888 	cfq_update_io_thinktime(cfqd, cic);
1889 	cfq_update_io_seektime(cfqd, cic, rq);
1890 	cfq_update_idle_window(cfqd, cfqq, cic);
1891 
1892 	cic->last_request_pos = rq->sector + rq->nr_sectors;
1893 
1894 	if (cfqq == cfqd->active_queue) {
1895 		/*
1896 		 * if we are waiting for a request for this queue, let it rip
1897 		 * immediately and flag that we must not expire this queue
1898 		 * just now
1899 		 */
1900 		if (cfq_cfqq_wait_request(cfqq)) {
1901 			cfq_mark_cfqq_must_dispatch(cfqq);
1902 			del_timer(&cfqd->idle_slice_timer);
1903 			blk_start_queueing(cfqd->queue);
1904 		}
1905 	} else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1906 		/*
1907 		 * not the active queue - expire current slice if it is
1908 		 * idle and has expired it's mean thinktime or this new queue
1909 		 * has some old slice time left and is of higher priority or
1910 		 * this new queue is RT and the current one is BE
1911 		 */
1912 		cfq_preempt_queue(cfqd, cfqq);
1913 		cfq_mark_cfqq_must_dispatch(cfqq);
1914 		blk_start_queueing(cfqd->queue);
1915 	}
1916 }
1917 
cfq_insert_request(struct request_queue * q,struct request * rq)1918 static void cfq_insert_request(struct request_queue *q, struct request *rq)
1919 {
1920 	struct cfq_data *cfqd = q->elevator->elevator_data;
1921 	struct cfq_queue *cfqq = RQ_CFQQ(rq);
1922 
1923 	cfq_log_cfqq(cfqd, cfqq, "insert_request");
1924 	cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
1925 
1926 	cfq_add_rq_rb(rq);
1927 
1928 	list_add_tail(&rq->queuelist, &cfqq->fifo);
1929 
1930 	cfq_rq_enqueued(cfqd, cfqq, rq);
1931 }
1932 
1933 /*
1934  * Update hw_tag based on peak queue depth over 50 samples under
1935  * sufficient load.
1936  */
cfq_update_hw_tag(struct cfq_data * cfqd)1937 static void cfq_update_hw_tag(struct cfq_data *cfqd)
1938 {
1939 	if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
1940 		cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
1941 
1942 	if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
1943 	    cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
1944 		return;
1945 
1946 	if (cfqd->hw_tag_samples++ < 50)
1947 		return;
1948 
1949 	if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
1950 		cfqd->hw_tag = 1;
1951 	else
1952 		cfqd->hw_tag = 0;
1953 
1954 	cfqd->hw_tag_samples = 0;
1955 	cfqd->rq_in_driver_peak = 0;
1956 }
1957 
cfq_completed_request(struct request_queue * q,struct request * rq)1958 static void cfq_completed_request(struct request_queue *q, struct request *rq)
1959 {
1960 	struct cfq_queue *cfqq = RQ_CFQQ(rq);
1961 	struct cfq_data *cfqd = cfqq->cfqd;
1962 	const int sync = rq_is_sync(rq);
1963 	unsigned long now;
1964 
1965 	now = jiffies;
1966 	cfq_log_cfqq(cfqd, cfqq, "complete");
1967 
1968 	cfq_update_hw_tag(cfqd);
1969 
1970 	WARN_ON(!cfqd->rq_in_driver);
1971 	WARN_ON(!cfqq->dispatched);
1972 	cfqd->rq_in_driver--;
1973 	cfqq->dispatched--;
1974 
1975 	if (cfq_cfqq_sync(cfqq))
1976 		cfqd->sync_flight--;
1977 
1978 	if (!cfq_class_idle(cfqq))
1979 		cfqd->last_end_request = now;
1980 
1981 	if (sync)
1982 		RQ_CIC(rq)->last_end_request = now;
1983 
1984 	/*
1985 	 * If this is the active queue, check if it needs to be expired,
1986 	 * or if we want to idle in case it has no pending requests.
1987 	 */
1988 	if (cfqd->active_queue == cfqq) {
1989 		if (cfq_cfqq_slice_new(cfqq)) {
1990 			cfq_set_prio_slice(cfqd, cfqq);
1991 			cfq_clear_cfqq_slice_new(cfqq);
1992 		}
1993 		if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
1994 			cfq_slice_expired(cfqd, 1);
1995 		else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1996 			cfq_arm_slice_timer(cfqd);
1997 	}
1998 
1999 	if (!cfqd->rq_in_driver)
2000 		cfq_schedule_dispatch(cfqd);
2001 }
2002 
2003 /*
2004  * we temporarily boost lower priority queues if they are holding fs exclusive
2005  * resources. they are boosted to normal prio (CLASS_BE/4)
2006  */
cfq_prio_boost(struct cfq_queue * cfqq)2007 static void cfq_prio_boost(struct cfq_queue *cfqq)
2008 {
2009 	if (has_fs_excl()) {
2010 		/*
2011 		 * boost idle prio on transactions that would lock out other
2012 		 * users of the filesystem
2013 		 */
2014 		if (cfq_class_idle(cfqq))
2015 			cfqq->ioprio_class = IOPRIO_CLASS_BE;
2016 		if (cfqq->ioprio > IOPRIO_NORM)
2017 			cfqq->ioprio = IOPRIO_NORM;
2018 	} else {
2019 		/*
2020 		 * check if we need to unboost the queue
2021 		 */
2022 		if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2023 			cfqq->ioprio_class = cfqq->org_ioprio_class;
2024 		if (cfqq->ioprio != cfqq->org_ioprio)
2025 			cfqq->ioprio = cfqq->org_ioprio;
2026 	}
2027 }
2028 
__cfq_may_queue(struct cfq_queue * cfqq)2029 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2030 {
2031 	if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
2032 	    !cfq_cfqq_must_alloc_slice(cfqq)) {
2033 		cfq_mark_cfqq_must_alloc_slice(cfqq);
2034 		return ELV_MQUEUE_MUST;
2035 	}
2036 
2037 	return ELV_MQUEUE_MAY;
2038 }
2039 
cfq_may_queue(struct request_queue * q,int rw)2040 static int cfq_may_queue(struct request_queue *q, int rw)
2041 {
2042 	struct cfq_data *cfqd = q->elevator->elevator_data;
2043 	struct task_struct *tsk = current;
2044 	struct cfq_io_context *cic;
2045 	struct cfq_queue *cfqq;
2046 
2047 	/*
2048 	 * don't force setup of a queue from here, as a call to may_queue
2049 	 * does not necessarily imply that a request actually will be queued.
2050 	 * so just lookup a possibly existing queue, or return 'may queue'
2051 	 * if that fails
2052 	 */
2053 	cic = cfq_cic_lookup(cfqd, tsk->io_context);
2054 	if (!cic)
2055 		return ELV_MQUEUE_MAY;
2056 
2057 	cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
2058 	if (cfqq) {
2059 		cfq_init_prio_data(cfqq, cic->ioc);
2060 		cfq_prio_boost(cfqq);
2061 
2062 		return __cfq_may_queue(cfqq);
2063 	}
2064 
2065 	return ELV_MQUEUE_MAY;
2066 }
2067 
2068 /*
2069  * queue lock held here
2070  */
cfq_put_request(struct request * rq)2071 static void cfq_put_request(struct request *rq)
2072 {
2073 	struct cfq_queue *cfqq = RQ_CFQQ(rq);
2074 
2075 	if (cfqq) {
2076 		const int rw = rq_data_dir(rq);
2077 
2078 		BUG_ON(!cfqq->allocated[rw]);
2079 		cfqq->allocated[rw]--;
2080 
2081 		put_io_context(RQ_CIC(rq)->ioc);
2082 
2083 		rq->elevator_private = NULL;
2084 		rq->elevator_private2 = NULL;
2085 
2086 		cfq_put_queue(cfqq);
2087 	}
2088 }
2089 
2090 /*
2091  * Allocate cfq data structures associated with this request.
2092  */
2093 static int
cfq_set_request(struct request_queue * q,struct request * rq,gfp_t gfp_mask)2094 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2095 {
2096 	struct cfq_data *cfqd = q->elevator->elevator_data;
2097 	struct cfq_io_context *cic;
2098 	const int rw = rq_data_dir(rq);
2099 	const int is_sync = rq_is_sync(rq);
2100 	struct cfq_queue *cfqq;
2101 	unsigned long flags;
2102 
2103 	might_sleep_if(gfp_mask & __GFP_WAIT);
2104 
2105 	cic = cfq_get_io_context(cfqd, gfp_mask);
2106 
2107 	spin_lock_irqsave(q->queue_lock, flags);
2108 
2109 	if (!cic)
2110 		goto queue_fail;
2111 
2112 	cfqq = cic_to_cfqq(cic, is_sync);
2113 	if (!cfqq) {
2114 		cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2115 
2116 		if (!cfqq)
2117 			goto queue_fail;
2118 
2119 		cic_set_cfqq(cic, cfqq, is_sync);
2120 	}
2121 
2122 	cfqq->allocated[rw]++;
2123 	cfq_clear_cfqq_must_alloc(cfqq);
2124 	atomic_inc(&cfqq->ref);
2125 
2126 	spin_unlock_irqrestore(q->queue_lock, flags);
2127 
2128 	rq->elevator_private = cic;
2129 	rq->elevator_private2 = cfqq;
2130 	return 0;
2131 
2132 queue_fail:
2133 	if (cic)
2134 		put_io_context(cic->ioc);
2135 
2136 	cfq_schedule_dispatch(cfqd);
2137 	spin_unlock_irqrestore(q->queue_lock, flags);
2138 	cfq_log(cfqd, "set_request fail");
2139 	return 1;
2140 }
2141 
cfq_kick_queue(struct work_struct * work)2142 static void cfq_kick_queue(struct work_struct *work)
2143 {
2144 	struct cfq_data *cfqd =
2145 		container_of(work, struct cfq_data, unplug_work);
2146 	struct request_queue *q = cfqd->queue;
2147 	unsigned long flags;
2148 
2149 	spin_lock_irqsave(q->queue_lock, flags);
2150 	blk_start_queueing(q);
2151 	spin_unlock_irqrestore(q->queue_lock, flags);
2152 }
2153 
2154 /*
2155  * Timer running if the active_queue is currently idling inside its time slice
2156  */
cfq_idle_slice_timer(unsigned long data)2157 static void cfq_idle_slice_timer(unsigned long data)
2158 {
2159 	struct cfq_data *cfqd = (struct cfq_data *) data;
2160 	struct cfq_queue *cfqq;
2161 	unsigned long flags;
2162 	int timed_out = 1;
2163 
2164 	cfq_log(cfqd, "idle timer fired");
2165 
2166 	spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2167 
2168 	cfqq = cfqd->active_queue;
2169 	if (cfqq) {
2170 		timed_out = 0;
2171 
2172 		/*
2173 		 * expired
2174 		 */
2175 		if (cfq_slice_used(cfqq))
2176 			goto expire;
2177 
2178 		/*
2179 		 * only expire and reinvoke request handler, if there are
2180 		 * other queues with pending requests
2181 		 */
2182 		if (!cfqd->busy_queues)
2183 			goto out_cont;
2184 
2185 		/*
2186 		 * not expired and it has a request pending, let it dispatch
2187 		 */
2188 		if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
2189 			cfq_mark_cfqq_must_dispatch(cfqq);
2190 			goto out_kick;
2191 		}
2192 	}
2193 expire:
2194 	cfq_slice_expired(cfqd, timed_out);
2195 out_kick:
2196 	cfq_schedule_dispatch(cfqd);
2197 out_cont:
2198 	spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2199 }
2200 
cfq_shutdown_timer_wq(struct cfq_data * cfqd)2201 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2202 {
2203 	del_timer_sync(&cfqd->idle_slice_timer);
2204 	cancel_work_sync(&cfqd->unplug_work);
2205 }
2206 
cfq_put_async_queues(struct cfq_data * cfqd)2207 static void cfq_put_async_queues(struct cfq_data *cfqd)
2208 {
2209 	int i;
2210 
2211 	for (i = 0; i < IOPRIO_BE_NR; i++) {
2212 		if (cfqd->async_cfqq[0][i])
2213 			cfq_put_queue(cfqd->async_cfqq[0][i]);
2214 		if (cfqd->async_cfqq[1][i])
2215 			cfq_put_queue(cfqd->async_cfqq[1][i]);
2216 	}
2217 
2218 	if (cfqd->async_idle_cfqq)
2219 		cfq_put_queue(cfqd->async_idle_cfqq);
2220 }
2221 
cfq_exit_queue(struct elevator_queue * e)2222 static void cfq_exit_queue(struct elevator_queue *e)
2223 {
2224 	struct cfq_data *cfqd = e->elevator_data;
2225 	struct request_queue *q = cfqd->queue;
2226 
2227 	cfq_shutdown_timer_wq(cfqd);
2228 
2229 	spin_lock_irq(q->queue_lock);
2230 
2231 	if (cfqd->active_queue)
2232 		__cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2233 
2234 	while (!list_empty(&cfqd->cic_list)) {
2235 		struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2236 							struct cfq_io_context,
2237 							queue_list);
2238 
2239 		__cfq_exit_single_io_context(cfqd, cic);
2240 	}
2241 
2242 	cfq_put_async_queues(cfqd);
2243 
2244 	spin_unlock_irq(q->queue_lock);
2245 
2246 	cfq_shutdown_timer_wq(cfqd);
2247 
2248 	kfree(cfqd);
2249 }
2250 
cfq_init_queue(struct request_queue * q)2251 static void *cfq_init_queue(struct request_queue *q)
2252 {
2253 	struct cfq_data *cfqd;
2254 
2255 	cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2256 	if (!cfqd)
2257 		return NULL;
2258 
2259 	cfqd->service_tree = CFQ_RB_ROOT;
2260 	INIT_LIST_HEAD(&cfqd->cic_list);
2261 
2262 	cfqd->queue = q;
2263 
2264 	init_timer(&cfqd->idle_slice_timer);
2265 	cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2266 	cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2267 
2268 	INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2269 
2270 	cfqd->last_end_request = jiffies;
2271 	cfqd->cfq_quantum = cfq_quantum;
2272 	cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2273 	cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2274 	cfqd->cfq_back_max = cfq_back_max;
2275 	cfqd->cfq_back_penalty = cfq_back_penalty;
2276 	cfqd->cfq_slice[0] = cfq_slice_async;
2277 	cfqd->cfq_slice[1] = cfq_slice_sync;
2278 	cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2279 	cfqd->cfq_slice_idle = cfq_slice_idle;
2280 	cfqd->hw_tag = 1;
2281 
2282 	return cfqd;
2283 }
2284 
cfq_slab_kill(void)2285 static void cfq_slab_kill(void)
2286 {
2287 	/*
2288 	 * Caller already ensured that pending RCU callbacks are completed,
2289 	 * so we should have no busy allocations at this point.
2290 	 */
2291 	if (cfq_pool)
2292 		kmem_cache_destroy(cfq_pool);
2293 	if (cfq_ioc_pool)
2294 		kmem_cache_destroy(cfq_ioc_pool);
2295 }
2296 
cfq_slab_setup(void)2297 static int __init cfq_slab_setup(void)
2298 {
2299 	cfq_pool = KMEM_CACHE(cfq_queue, 0);
2300 	if (!cfq_pool)
2301 		goto fail;
2302 
2303 	cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2304 	if (!cfq_ioc_pool)
2305 		goto fail;
2306 
2307 	return 0;
2308 fail:
2309 	cfq_slab_kill();
2310 	return -ENOMEM;
2311 }
2312 
2313 /*
2314  * sysfs parts below -->
2315  */
2316 static ssize_t
cfq_var_show(unsigned int var,char * page)2317 cfq_var_show(unsigned int var, char *page)
2318 {
2319 	return sprintf(page, "%d\n", var);
2320 }
2321 
2322 static ssize_t
cfq_var_store(unsigned int * var,const char * page,size_t count)2323 cfq_var_store(unsigned int *var, const char *page, size_t count)
2324 {
2325 	char *p = (char *) page;
2326 
2327 	*var = simple_strtoul(p, &p, 10);
2328 	return count;
2329 }
2330 
2331 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)				\
2332 static ssize_t __FUNC(struct elevator_queue *e, char *page)		\
2333 {									\
2334 	struct cfq_data *cfqd = e->elevator_data;			\
2335 	unsigned int __data = __VAR;					\
2336 	if (__CONV)							\
2337 		__data = jiffies_to_msecs(__data);			\
2338 	return cfq_var_show(__data, (page));				\
2339 }
2340 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2341 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2342 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2343 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2344 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2345 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2346 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2347 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2348 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2349 #undef SHOW_FUNCTION
2350 
2351 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)			\
2352 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)	\
2353 {									\
2354 	struct cfq_data *cfqd = e->elevator_data;			\
2355 	unsigned int __data;						\
2356 	int ret = cfq_var_store(&__data, (page), count);		\
2357 	if (__data < (MIN))						\
2358 		__data = (MIN);						\
2359 	else if (__data > (MAX))					\
2360 		__data = (MAX);						\
2361 	if (__CONV)							\
2362 		*(__PTR) = msecs_to_jiffies(__data);			\
2363 	else								\
2364 		*(__PTR) = __data;					\
2365 	return ret;							\
2366 }
2367 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2368 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2369 		UINT_MAX, 1);
2370 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2371 		UINT_MAX, 1);
2372 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2373 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2374 		UINT_MAX, 0);
2375 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2376 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2377 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2378 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2379 		UINT_MAX, 0);
2380 #undef STORE_FUNCTION
2381 
2382 #define CFQ_ATTR(name) \
2383 	__ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2384 
2385 static struct elv_fs_entry cfq_attrs[] = {
2386 	CFQ_ATTR(quantum),
2387 	CFQ_ATTR(fifo_expire_sync),
2388 	CFQ_ATTR(fifo_expire_async),
2389 	CFQ_ATTR(back_seek_max),
2390 	CFQ_ATTR(back_seek_penalty),
2391 	CFQ_ATTR(slice_sync),
2392 	CFQ_ATTR(slice_async),
2393 	CFQ_ATTR(slice_async_rq),
2394 	CFQ_ATTR(slice_idle),
2395 	__ATTR_NULL
2396 };
2397 
2398 static struct elevator_type iosched_cfq = {
2399 	.ops = {
2400 		.elevator_merge_fn = 		cfq_merge,
2401 		.elevator_merged_fn =		cfq_merged_request,
2402 		.elevator_merge_req_fn =	cfq_merged_requests,
2403 		.elevator_allow_merge_fn =	cfq_allow_merge,
2404 		.elevator_dispatch_fn =		cfq_dispatch_requests,
2405 		.elevator_add_req_fn =		cfq_insert_request,
2406 		.elevator_activate_req_fn =	cfq_activate_request,
2407 		.elevator_deactivate_req_fn =	cfq_deactivate_request,
2408 		.elevator_queue_empty_fn =	cfq_queue_empty,
2409 		.elevator_completed_req_fn =	cfq_completed_request,
2410 		.elevator_former_req_fn =	elv_rb_former_request,
2411 		.elevator_latter_req_fn =	elv_rb_latter_request,
2412 		.elevator_set_req_fn =		cfq_set_request,
2413 		.elevator_put_req_fn =		cfq_put_request,
2414 		.elevator_may_queue_fn =	cfq_may_queue,
2415 		.elevator_init_fn =		cfq_init_queue,
2416 		.elevator_exit_fn =		cfq_exit_queue,
2417 		.trim =				cfq_free_io_context,
2418 	},
2419 	.elevator_attrs =	cfq_attrs,
2420 	.elevator_name =	"cfq",
2421 	.elevator_owner =	THIS_MODULE,
2422 };
2423 
cfq_init(void)2424 static int __init cfq_init(void)
2425 {
2426 	/*
2427 	 * could be 0 on HZ < 1000 setups
2428 	 */
2429 	if (!cfq_slice_async)
2430 		cfq_slice_async = 1;
2431 	if (!cfq_slice_idle)
2432 		cfq_slice_idle = 1;
2433 
2434 	if (cfq_slab_setup())
2435 		return -ENOMEM;
2436 
2437 	elv_register(&iosched_cfq);
2438 
2439 	return 0;
2440 }
2441 
cfq_exit(void)2442 static void __exit cfq_exit(void)
2443 {
2444 	DECLARE_COMPLETION_ONSTACK(all_gone);
2445 	elv_unregister(&iosched_cfq);
2446 	ioc_gone = &all_gone;
2447 	/* ioc_gone's update must be visible before reading ioc_count */
2448 	smp_wmb();
2449 
2450 	/*
2451 	 * this also protects us from entering cfq_slab_kill() with
2452 	 * pending RCU callbacks
2453 	 */
2454 	if (elv_ioc_count_read(ioc_count))
2455 		wait_for_completion(&all_gone);
2456 	cfq_slab_kill();
2457 }
2458 
2459 module_init(cfq_init);
2460 module_exit(cfq_exit);
2461 
2462 MODULE_AUTHOR("Jens Axboe");
2463 MODULE_LICENSE("GPL");
2464 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");
2465