/drivers/block/elevator.c

  1/*
  2 *  linux/drivers/block/elevator.c
  3 *
  4 *  Block device elevator/IO-scheduler.
  5 *
  6 *  Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
  7 *
  8 * 30042000 Jens Axboe <axboe@suse.de> :
  9 *
 10 * Split the elevator a bit so that it is possible to choose a different
 11 * one or even write a new "plug in". There are three pieces:
 12 * - elevator_fn, inserts a new request in the queue list
 13 * - elevator_merge_fn, decides whether a new buffer can be merged with
 14 *   an existing request
 15 * - elevator_dequeue_fn, called when a request is taken off the active list
 16 *
 17 * 20082000 Dave Jones <davej@suse.de> :
 18 * Removed tests for max-bomb-segments, which was breaking elvtune
 19 *  when run without -bN
 20 *
 21 * Jens:
 22 * - Rework again to work with bio instead of buffer_heads
 23 * - loose bi_dev comparisons, partition handling is right now
 24 * - completely modularize elevator setup and teardown
 25 *
 26 */
 27#include <linux/kernel.h>
 28#include <linux/fs.h>
 29#include <linux/blkdev.h>
 30#include <linux/elevator.h>
 31#include <linux/bio.h>
 32#include <linux/config.h>
 33#include <linux/module.h>
 34#include <linux/slab.h>
 35#include <linux/init.h>
 36#include <linux/compiler.h>
 37
 38#include <asm/uaccess.h>
 39
 40static DEFINE_SPINLOCK(elv_list_lock);
 41static LIST_HEAD(elv_list);
 42
 43/*
 44 * can we safely merge with this request?
 45 */
 46inline int elv_rq_merge_ok(struct request *rq, struct bio *bio)
 47{
 48	if (!rq_mergeable(rq))
 49		return 0;
 50
 51	/*
 52	 * different data direction or already started, don't merge
 53	 */
 54	if (bio_data_dir(bio) != rq_data_dir(rq))
 55		return 0;
 56
 57	/*
 58	 * same device and no special stuff set, merge is ok
 59	 */
 60	if (rq->rq_disk == bio->bi_bdev->bd_disk &&
 61	    !rq->waiting && !rq->special)
 62		return 1;
 63
 64	return 0;
 65}
 66EXPORT_SYMBOL(elv_rq_merge_ok);
 67
 68inline int elv_try_merge(struct request *__rq, struct bio *bio)
 69{
 70	int ret = ELEVATOR_NO_MERGE;
 71
 72	/*
 73	 * we can merge and sequence is ok, check if it's possible
 74	 */
 75	if (elv_rq_merge_ok(__rq, bio)) {
 76		if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
 77			ret = ELEVATOR_BACK_MERGE;
 78		else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
 79			ret = ELEVATOR_FRONT_MERGE;
 80	}
 81
 82	return ret;
 83}
 84EXPORT_SYMBOL(elv_try_merge);
 85
 86inline int elv_try_last_merge(request_queue_t *q, struct bio *bio)
 87{
 88	if (q->last_merge)
 89		return elv_try_merge(q->last_merge, bio);
 90
 91	return ELEVATOR_NO_MERGE;
 92}
 93EXPORT_SYMBOL(elv_try_last_merge);
 94
 95static struct elevator_type *elevator_find(const char *name)
 96{
 97	struct elevator_type *e = NULL;
 98	struct list_head *entry;
 99
100	spin_lock_irq(&elv_list_lock);
101	list_for_each(entry, &elv_list) {
102		struct elevator_type *__e;
103
104		__e = list_entry(entry, struct elevator_type, list);
105
106		if (!strcmp(__e->elevator_name, name)) {
107			e = __e;
108			break;
109		}
110	}
111	spin_unlock_irq(&elv_list_lock);
112
113	return e;
114}
115
116static void elevator_put(struct elevator_type *e)
117{
118	module_put(e->elevator_owner);
119}
120
121static struct elevator_type *elevator_get(const char *name)
122{
123	struct elevator_type *e = elevator_find(name);
124
125	if (!e)
126		return NULL;
127	if (!try_module_get(e->elevator_owner))
128		return NULL;
129
130	return e;
131}
132
133static int elevator_attach(request_queue_t *q, struct elevator_type *e,
134			   struct elevator_queue *eq)
135{
136	int ret = 0;
137
138	memset(eq, 0, sizeof(*eq));
139	eq->ops = &e->ops;
140	eq->elevator_type = e;
141
142	INIT_LIST_HEAD(&q->queue_head);
143	q->last_merge = NULL;
144	q->elevator = eq;
145
146	if (eq->ops->elevator_init_fn)
147		ret = eq->ops->elevator_init_fn(q, eq);
148
149	return ret;
150}
151
152static char chosen_elevator[16];
153
154static void elevator_setup_default(void)
155{
156	/*
157	 * check if default is set and exists
158	 */
159	if (chosen_elevator[0] && elevator_find(chosen_elevator))
160		return;
161
162#if defined(CONFIG_IOSCHED_AS)
163	strcpy(chosen_elevator, "anticipatory");
164#elif defined(CONFIG_IOSCHED_DEADLINE)
165	strcpy(chosen_elevator, "deadline");
166#elif defined(CONFIG_IOSCHED_CFQ)
167	strcpy(chosen_elevator, "cfq");
168#elif defined(CONFIG_IOSCHED_NOOP)
169	strcpy(chosen_elevator, "noop");
170#else
171#error "You must build at least 1 IO scheduler into the kernel"
172#endif
173}
174
175static int __init elevator_setup(char *str)
176{
177	strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
178	return 0;
179}
180
181__setup("elevator=", elevator_setup);
182
183int elevator_init(request_queue_t *q, char *name)
184{
185	struct elevator_type *e = NULL;
186	struct elevator_queue *eq;
187	int ret = 0;
188
189	elevator_setup_default();
190
191	if (!name)
192		name = chosen_elevator;
193
194	e = elevator_get(name);
195	if (!e)
196		return -EINVAL;
197
198	eq = kmalloc(sizeof(struct elevator_queue), GFP_KERNEL);
199	if (!eq) {
200		elevator_put(e->elevator_type);
201		return -ENOMEM;
202	}
203
204	ret = elevator_attach(q, e, eq);
205	if (ret) {
206		kfree(eq);
207		elevator_put(e->elevator_type);
208	}
209
210	return ret;
211}
212
213void elevator_exit(elevator_t *e)
214{
215	if (e->ops->elevator_exit_fn)
216		e->ops->elevator_exit_fn(e);
217
218	elevator_put(e->elevator_type);
219	e->elevator_type = NULL;
220	kfree(e);
221}
222
223static int elevator_global_init(void)
224{
225	return 0;
226}
227
228int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
229{
230	elevator_t *e = q->elevator;
231
232	if (e->ops->elevator_merge_fn)
233		return e->ops->elevator_merge_fn(q, req, bio);
234
235	return ELEVATOR_NO_MERGE;
236}
237
238void elv_merged_request(request_queue_t *q, struct request *rq)
239{
240	elevator_t *e = q->elevator;
241
242	if (e->ops->elevator_merged_fn)
243		e->ops->elevator_merged_fn(q, rq);
244}
245
246void elv_merge_requests(request_queue_t *q, struct request *rq,
247			     struct request *next)
248{
249	elevator_t *e = q->elevator;
250
251	if (q->last_merge == next)
252		q->last_merge = NULL;
253
254	if (e->ops->elevator_merge_req_fn)
255		e->ops->elevator_merge_req_fn(q, rq, next);
256}
257
258/*
259 * For careful internal use by the block layer. Essentially the same as
260 * a requeue in that it tells the io scheduler that this request is not
261 * active in the driver or hardware anymore, but we don't want the request
262 * added back to the scheduler. Function is not exported.
263 */
264void elv_deactivate_request(request_queue_t *q, struct request *rq)
265{
266	elevator_t *e = q->elevator;
267
268	/*
269	 * it already went through dequeue, we need to decrement the
270	 * in_flight count again
271	 */
272	if (blk_account_rq(rq))
273		q->in_flight--;
274
275	rq->flags &= ~REQ_STARTED;
276
277	if (e->ops->elevator_deactivate_req_fn)
278		e->ops->elevator_deactivate_req_fn(q, rq);
279}
280
281void elv_requeue_request(request_queue_t *q, struct request *rq)
282{
283	elv_deactivate_request(q, rq);
284
285	/*
286	 * if this is the flush, requeue the original instead and drop the flush
287	 */
288	if (rq->flags & REQ_BAR_FLUSH) {
289		clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
290		rq = rq->end_io_data;
291	}
292
293	/*
294	 * if iosched has an explicit requeue hook, then use that. otherwise
295	 * just put the request at the front of the queue
296	 */
297	if (q->elevator->ops->elevator_requeue_req_fn)
298		q->elevator->ops->elevator_requeue_req_fn(q, rq);
299	else
300		__elv_add_request(q, rq, ELEVATOR_INSERT_FRONT, 0);
301}
302
303void __elv_add_request(request_queue_t *q, struct request *rq, int where,
304		       int plug)
305{
306	/*
307	 * barriers implicitly indicate back insertion
308	 */
309	if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER) &&
310	    where == ELEVATOR_INSERT_SORT)
311		where = ELEVATOR_INSERT_BACK;
312
313	if (plug)
314		blk_plug_device(q);
315
316	rq->q = q;
317
318	if (!test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags)) {
319		q->elevator->ops->elevator_add_req_fn(q, rq, where);
320
321		if (blk_queue_plugged(q)) {
322			int nrq = q->rq.count[READ] + q->rq.count[WRITE]
323				  - q->in_flight;
324
325			if (nrq == q->unplug_thresh)
326				__generic_unplug_device(q);
327		}
328	} else
329		/*
330		 * if drain is set, store the request "locally". when the drain
331		 * is finished, the requests will be handed ordered to the io
332		 * scheduler
333		 */
334		list_add_tail(&rq->queuelist, &q->drain_list);
335}
336
337void elv_add_request(request_queue_t *q, struct request *rq, int where,
338		     int plug)
339{
340	unsigned long flags;
341
342	spin_lock_irqsave(q->queue_lock, flags);
343	__elv_add_request(q, rq, where, plug);
344	spin_unlock_irqrestore(q->queue_lock, flags);
345}
346
347static inline struct request *__elv_next_request(request_queue_t *q)
348{
349	struct request *rq = q->elevator->ops->elevator_next_req_fn(q);
350
351	/*
352	 * if this is a barrier write and the device has to issue a
353	 * flush sequence to support it, check how far we are
354	 */
355	if (rq && blk_fs_request(rq) && blk_barrier_rq(rq)) {
356		BUG_ON(q->ordered == QUEUE_ORDERED_NONE);
357
358		if (q->ordered == QUEUE_ORDERED_FLUSH &&
359		    !blk_barrier_preflush(rq))
360			rq = blk_start_pre_flush(q, rq);
361	}
362
363	return rq;
364}
365
366struct request *elv_next_request(request_queue_t *q)
367{
368	struct request *rq;
369	int ret;
370
371	while ((rq = __elv_next_request(q)) != NULL) {
372		/*
373		 * just mark as started even if we don't start it, a request
374		 * that has been delayed should not be passed by new incoming
375		 * requests
376		 */
377		rq->flags |= REQ_STARTED;
378
379		if (rq == q->last_merge)
380			q->last_merge = NULL;
381
382		if ((rq->flags & REQ_DONTPREP) || !q->prep_rq_fn)
383			break;
384
385		ret = q->prep_rq_fn(q, rq);
386		if (ret == BLKPREP_OK) {
387			break;
388		} else if (ret == BLKPREP_DEFER) {
389			rq = NULL;
390			break;
391		} else if (ret == BLKPREP_KILL) {
392			int nr_bytes = rq->hard_nr_sectors << 9;
393
394			if (!nr_bytes)
395				nr_bytes = rq->data_len;
396
397			blkdev_dequeue_request(rq);
398			rq->flags |= REQ_QUIET;
399			end_that_request_chunk(rq, 0, nr_bytes);
400			end_that_request_last(rq);
401		} else {
402			printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
403								ret);
404			break;
405		}
406	}
407
408	return rq;
409}
410
411void elv_remove_request(request_queue_t *q, struct request *rq)
412{
413	elevator_t *e = q->elevator;
414
415	/*
416	 * the time frame between a request being removed from the lists
417	 * and to it is freed is accounted as io that is in progress at
418	 * the driver side. note that we only account requests that the
419	 * driver has seen (REQ_STARTED set), to avoid false accounting
420	 * for request-request merges
421	 */
422	if (blk_account_rq(rq))
423		q->in_flight++;
424
425	/*
426	 * the main clearing point for q->last_merge is on retrieval of
427	 * request by driver (it calls elv_next_request()), but it _can_
428	 * also happen here if a request is added to the queue but later
429	 * deleted without ever being given to driver (merged with another
430	 * request).
431	 */
432	if (rq == q->last_merge)
433		q->last_merge = NULL;
434
435	if (e->ops->elevator_remove_req_fn)
436		e->ops->elevator_remove_req_fn(q, rq);
437}
438
439int elv_queue_empty(request_queue_t *q)
440{
441	elevator_t *e = q->elevator;
442
443	if (e->ops->elevator_queue_empty_fn)
444		return e->ops->elevator_queue_empty_fn(q);
445
446	return list_empty(&q->queue_head);
447}
448
449struct request *elv_latter_request(request_queue_t *q, struct request *rq)
450{
451	struct list_head *next;
452
453	elevator_t *e = q->elevator;
454
455	if (e->ops->elevator_latter_req_fn)
456		return e->ops->elevator_latter_req_fn(q, rq);
457
458	next = rq->queuelist.next;
459	if (next != &q->queue_head && next != &rq->queuelist)
460		return list_entry_rq(next);
461
462	return NULL;
463}
464
465struct request *elv_former_request(request_queue_t *q, struct request *rq)
466{
467	struct list_head *prev;
468
469	elevator_t *e = q->elevator;
470
471	if (e->ops->elevator_former_req_fn)
472		return e->ops->elevator_former_req_fn(q, rq);
473
474	prev = rq->queuelist.prev;
475	if (prev != &q->queue_head && prev != &rq->queuelist)
476		return list_entry_rq(prev);
477
478	return NULL;
479}
480
481int elv_set_request(request_queue_t *q, struct request *rq, int gfp_mask)
482{
483	elevator_t *e = q->elevator;
484
485	if (e->ops->elevator_set_req_fn)
486		return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
487
488	rq->elevator_private = NULL;
489	return 0;
490}
491
492void elv_put_request(request_queue_t *q, struct request *rq)
493{
494	elevator_t *e = q->elevator;
495
496	if (e->ops->elevator_put_req_fn)
497		e->ops->elevator_put_req_fn(q, rq);
498}
499
500int elv_may_queue(request_queue_t *q, int rw)
501{
502	elevator_t *e = q->elevator;
503
504	if (e->ops->elevator_may_queue_fn)
505		return e->ops->elevator_may_queue_fn(q, rw);
506
507	return ELV_MQUEUE_MAY;
508}
509
510void elv_completed_request(request_queue_t *q, struct request *rq)
511{
512	elevator_t *e = q->elevator;
513
514	/*
515	 * request is released from the driver, io must be done
516	 */
517	if (blk_account_rq(rq))
518		q->in_flight--;
519
520	if (e->ops->elevator_completed_req_fn)
521		e->ops->elevator_completed_req_fn(q, rq);
522}
523
524int elv_register_queue(struct request_queue *q)
525{
526	elevator_t *e = q->elevator;
527
528	e->kobj.parent = kobject_get(&q->kobj);
529	if (!e->kobj.parent)
530		return -EBUSY;
531
532	snprintf(e->kobj.name, KOBJ_NAME_LEN, "%s", "iosched");
533	e->kobj.ktype = e->elevator_type->elevator_ktype;
534
535	return kobject_register(&e->kobj);
536}
537
538void elv_unregister_queue(struct request_queue *q)
539{
540	if (q) {
541		elevator_t *e = q->elevator;
542		kobject_unregister(&e->kobj);
543		kobject_put(&q->kobj);
544	}
545}
546
547int elv_register(struct elevator_type *e)
548{
549	if (elevator_find(e->elevator_name))
550		BUG();
551
552	spin_lock_irq(&elv_list_lock);
553	list_add_tail(&e->list, &elv_list);
554	spin_unlock_irq(&elv_list_lock);
555
556	printk(KERN_INFO "io scheduler %s registered", e->elevator_name);
557	if (!strcmp(e->elevator_name, chosen_elevator))
558		printk(" (default)");
559	printk("\n");
560	return 0;
561}
562EXPORT_SYMBOL_GPL(elv_register);
563
564void elv_unregister(struct elevator_type *e)
565{
566	spin_lock_irq(&elv_list_lock);
567	list_del_init(&e->list);
568	spin_unlock_irq(&elv_list_lock);
569}
570EXPORT_SYMBOL_GPL(elv_unregister);
571
572/*
573 * switch to new_e io scheduler. be careful not to introduce deadlocks -
574 * we don't free the old io scheduler, before we have allocated what we
575 * need for the new one. this way we have a chance of going back to the old
576 * one, if the new one fails init for some reason. we also do an intermediate
577 * switch to noop to ensure safety with stack-allocated requests, since they
578 * don't originate from the block layer allocator. noop is safe here, because
579 * it never needs to touch the elevator itself for completion events. DRAIN
580 * flags will make sure we don't touch it for additions either.
581 */
582static void elevator_switch(request_queue_t *q, struct elevator_type *new_e)
583{
584	elevator_t *e = kmalloc(sizeof(elevator_t), GFP_KERNEL);
585	struct elevator_type *noop_elevator = NULL;
586	elevator_t *old_elevator;
587
588	if (!e)
589		goto error;
590
591	/*
592	 * first step, drain requests from the block freelist
593	 */
594	blk_wait_queue_drained(q, 0);
595
596	/*
597	 * unregister old elevator data
598	 */
599	elv_unregister_queue(q);
600	old_elevator = q->elevator;
601
602	/*
603 	 * next step, switch to noop since it uses no private rq structures
604	 * and doesn't allocate any memory for anything. then wait for any
605	 * non-fs requests in-flight
606 	 */
607	noop_elevator = elevator_get("noop");
608	spin_lock_irq(q->queue_lock);
609	elevator_attach(q, noop_elevator, e);
610	spin_unlock_irq(q->queue_lock);
611
612	blk_wait_queue_drained(q, 1);
613
614	/*
615	 * attach and start new elevator
616	 */
617	if (elevator_attach(q, new_e, e))
618		goto fail;
619
620	if (elv_register_queue(q))
621		goto fail_register;
622
623	/*
624	 * finally exit old elevator and start queue again
625	 */
626	elevator_exit(old_elevator);
627	blk_finish_queue_drain(q);
628	elevator_put(noop_elevator);
629	return;
630
631fail_register:
632	/*
633	 * switch failed, exit the new io scheduler and reattach the old
634	 * one again (along with re-adding the sysfs dir)
635	 */
636	elevator_exit(e);
637fail:
638	q->elevator = old_elevator;
639	elv_register_queue(q);
640	blk_finish_queue_drain(q);
641error:
642	if (noop_elevator)
643		elevator_put(noop_elevator);
644	elevator_put(new_e);
645	printk(KERN_ERR "elevator: switch to %s failed\n",new_e->elevator_name);
646}
647
648ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
649{
650	char elevator_name[ELV_NAME_MAX];
651	struct elevator_type *e;
652
653	memset(elevator_name, 0, sizeof(elevator_name));
654	strncpy(elevator_name, name, sizeof(elevator_name));
655
656	if (elevator_name[strlen(elevator_name) - 1] == '\n')
657		elevator_name[strlen(elevator_name) - 1] = '\0';
658
659	e = elevator_get(elevator_name);
660	if (!e) {
661		printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
662		return -EINVAL;
663	}
664
665	if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name))
666		return count;
667
668	elevator_switch(q, e);
669	return count;
670}
671
672ssize_t elv_iosched_show(request_queue_t *q, char *name)
673{
674	elevator_t *e = q->elevator;
675	struct elevator_type *elv = e->elevator_type;
676	struct list_head *entry;
677	int len = 0;
678
679	spin_lock_irq(q->queue_lock);
680	list_for_each(entry, &elv_list) {
681		struct elevator_type *__e;
682
683		__e = list_entry(entry, struct elevator_type, list);
684		if (!strcmp(elv->elevator_name, __e->elevator_name))
685			len += sprintf(name+len, "[%s] ", elv->elevator_name);
686		else
687			len += sprintf(name+len, "%s ", __e->elevator_name);
688	}
689	spin_unlock_irq(q->queue_lock);
690
691	len += sprintf(len+name, "\n");
692	return len;
693}
694
695module_init(elevator_global_init);
696
697EXPORT_SYMBOL(elv_add_request);
698EXPORT_SYMBOL(__elv_add_request);
699EXPORT_SYMBOL(elv_requeue_request);
700EXPORT_SYMBOL(elv_next_request);
701EXPORT_SYMBOL(elv_remove_request);
702EXPORT_SYMBOL(elv_queue_empty);
703EXPORT_SYMBOL(elv_completed_request);
704EXPORT_SYMBOL(elevator_exit);
705EXPORT_SYMBOL(elevator_init);