2 * Copyright (C) 1991, 1992 Linus Torvalds
3 * Copyright (C) 1994, Karl Keyte: Added support for disk statistics
4 * Elevator latency, (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
5 * Queue request tables / lock, selectable elevator, Jens Axboe <axboe@suse.de>
6 * kernel-doc documentation started by NeilBrown <neilb@cse.unsw.edu.au> - July2000
7 * bio rewrite, highmem i/o, etc, Jens Axboe <axboe@suse.de> - may 2001
11 * This handles all read/write requests to block devices
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/backing-dev.h>
16 #include <linux/bio.h>
17 #include <linux/blkdev.h>
18 #include <linux/highmem.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/string.h>
22 #include <linux/init.h>
23 #include <linux/completion.h>
24 #include <linux/slab.h>
25 #include <linux/swap.h>
26 #include <linux/writeback.h>
27 #include <linux/task_io_accounting_ops.h>
28 #include <linux/interrupt.h>
29 #include <linux/cpu.h>
30 #include <linux/blktrace_api.h>
31 #include <linux/fault-inject.h>
32 #include <linux/scatterlist.h>
36 static void drive_stat_acct(struct request *rq, int new_io);
37 static int __make_request(struct request_queue *q, struct bio *bio);
38 static void blk_recalc_rq_segments(struct request *rq);
41 * For the allocated request tables
43 struct kmem_cache *request_cachep;
46 * For queue allocation
48 struct kmem_cache *blk_requestq_cachep = NULL;
51 * Controlling structure to kblockd
53 static struct workqueue_struct *kblockd_workqueue;
55 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
57 void blk_queue_congestion_threshold(struct request_queue *q)
61 nr = q->nr_requests - (q->nr_requests / 8) + 1;
62 if (nr > q->nr_requests)
64 q->nr_congestion_on = nr;
66 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
69 q->nr_congestion_off = nr;
73 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
76 * Locates the passed device's request queue and returns the address of its
79 * Will return NULL if the request queue cannot be located.
81 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
83 struct backing_dev_info *ret = NULL;
84 struct request_queue *q = bdev_get_queue(bdev);
87 ret = &q->backing_dev_info;
90 EXPORT_SYMBOL(blk_get_backing_dev_info);
92 void rq_init(struct request_queue *q, struct request *rq)
94 INIT_LIST_HEAD(&rq->queuelist);
95 INIT_LIST_HEAD(&rq->donelist);
98 rq->bio = rq->biotail = NULL;
99 INIT_HLIST_NODE(&rq->hash);
100 RB_CLEAR_NODE(&rq->rb_node);
108 rq->nr_phys_segments = 0;
111 rq->end_io_data = NULL;
112 rq->completion_data = NULL;
116 static void req_bio_endio(struct request *rq, struct bio *bio,
117 unsigned int nbytes, int error)
119 struct request_queue *q = rq->q;
121 if (&q->bar_rq != rq) {
123 clear_bit(BIO_UPTODATE, &bio->bi_flags);
124 else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
127 if (unlikely(nbytes > bio->bi_size)) {
128 printk("%s: want %u bytes done, only %u left\n",
129 __FUNCTION__, nbytes, bio->bi_size);
130 nbytes = bio->bi_size;
133 bio->bi_size -= nbytes;
134 bio->bi_sector += (nbytes >> 9);
135 if (bio->bi_size == 0)
136 bio_endio(bio, error);
140 * Okay, this is the barrier request in progress, just
143 if (error && !q->orderr)
148 void blk_dump_rq_flags(struct request *rq, char *msg)
152 printk("%s: dev %s: type=%x, flags=%x\n", msg,
153 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->cmd_type,
156 printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
158 rq->current_nr_sectors);
159 printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
161 if (blk_pc_request(rq)) {
163 for (bit = 0; bit < sizeof(rq->cmd); bit++)
164 printk("%02x ", rq->cmd[bit]);
169 EXPORT_SYMBOL(blk_dump_rq_flags);
171 void blk_recount_segments(struct request_queue *q, struct bio *bio)
174 struct bio *nxt = bio->bi_next;
176 rq.bio = rq.biotail = bio;
178 blk_recalc_rq_segments(&rq);
180 bio->bi_phys_segments = rq.nr_phys_segments;
181 bio->bi_hw_segments = rq.nr_hw_segments;
182 bio->bi_flags |= (1 << BIO_SEG_VALID);
184 EXPORT_SYMBOL(blk_recount_segments);
186 static void blk_recalc_rq_segments(struct request *rq)
190 unsigned int phys_size;
191 unsigned int hw_size;
192 struct bio_vec *bv, *bvprv = NULL;
196 struct req_iterator iter;
197 int high, highprv = 1;
198 struct request_queue *q = rq->q;
203 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
204 hw_seg_size = seg_size = 0;
205 phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
206 rq_for_each_segment(bv, rq, iter) {
208 * the trick here is making sure that a high page is never
209 * considered part of another segment, since that might
210 * change with the bounce page.
212 high = page_to_pfn(bv->bv_page) > q->bounce_pfn;
216 if (seg_size + bv->bv_len > q->max_segment_size)
218 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
220 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
222 if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
225 seg_size += bv->bv_len;
226 hw_seg_size += bv->bv_len;
231 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
232 !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
233 hw_seg_size += bv->bv_len;
236 if (nr_hw_segs == 1 &&
237 hw_seg_size > rq->bio->bi_hw_front_size)
238 rq->bio->bi_hw_front_size = hw_seg_size;
239 hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
245 seg_size = bv->bv_len;
249 if (nr_hw_segs == 1 &&
250 hw_seg_size > rq->bio->bi_hw_front_size)
251 rq->bio->bi_hw_front_size = hw_seg_size;
252 if (hw_seg_size > rq->biotail->bi_hw_back_size)
253 rq->biotail->bi_hw_back_size = hw_seg_size;
254 rq->nr_phys_segments = nr_phys_segs;
255 rq->nr_hw_segments = nr_hw_segs;
258 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
261 if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
264 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
266 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
270 * bio and nxt are contigous in memory, check if the queue allows
271 * these two to be merged into one
273 if (BIO_SEG_BOUNDARY(q, bio, nxt))
279 static int blk_hw_contig_segment(struct request_queue *q, struct bio *bio,
282 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
283 blk_recount_segments(q, bio);
284 if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
285 blk_recount_segments(q, nxt);
286 if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
287 BIOVEC_VIRT_OVERSIZE(bio->bi_hw_back_size + nxt->bi_hw_front_size))
289 if (bio->bi_hw_back_size + nxt->bi_hw_front_size > q->max_segment_size)
296 * map a request to scatterlist, return number of sg entries setup. Caller
297 * must make sure sg can hold rq->nr_phys_segments entries
299 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
300 struct scatterlist *sglist)
302 struct bio_vec *bvec, *bvprv;
303 struct req_iterator iter;
304 struct scatterlist *sg;
308 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
315 rq_for_each_segment(bvec, rq, iter) {
316 int nbytes = bvec->bv_len;
318 if (bvprv && cluster) {
319 if (sg->length + nbytes > q->max_segment_size)
322 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
324 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
327 sg->length += nbytes;
334 * If the driver previously mapped a shorter
335 * list, we could see a termination bit
336 * prematurely unless it fully inits the sg
337 * table on each mapping. We KNOW that there
338 * must be more entries here or the driver
339 * would be buggy, so force clear the
340 * termination bit to avoid doing a full
341 * sg_init_table() in drivers for each command.
343 sg->page_link &= ~0x02;
347 sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
351 } /* segments in rq */
353 if (q->dma_drain_size) {
354 sg->page_link &= ~0x02;
356 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
358 ((unsigned long)q->dma_drain_buffer) &
369 EXPORT_SYMBOL(blk_rq_map_sg);
372 * the standard queue merge functions, can be overridden with device
373 * specific ones if so desired
376 static inline int ll_new_mergeable(struct request_queue *q,
380 int nr_phys_segs = bio_phys_segments(q, bio);
382 if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
383 req->cmd_flags |= REQ_NOMERGE;
384 if (req == q->last_merge)
385 q->last_merge = NULL;
390 * A hw segment is just getting larger, bump just the phys
393 req->nr_phys_segments += nr_phys_segs;
397 static inline int ll_new_hw_segment(struct request_queue *q,
401 int nr_hw_segs = bio_hw_segments(q, bio);
402 int nr_phys_segs = bio_phys_segments(q, bio);
404 if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
405 || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
406 req->cmd_flags |= REQ_NOMERGE;
407 if (req == q->last_merge)
408 q->last_merge = NULL;
413 * This will form the start of a new hw segment. Bump both
416 req->nr_hw_segments += nr_hw_segs;
417 req->nr_phys_segments += nr_phys_segs;
421 int ll_back_merge_fn(struct request_queue *q, struct request *req,
424 unsigned short max_sectors;
427 if (unlikely(blk_pc_request(req)))
428 max_sectors = q->max_hw_sectors;
430 max_sectors = q->max_sectors;
432 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
433 req->cmd_flags |= REQ_NOMERGE;
434 if (req == q->last_merge)
435 q->last_merge = NULL;
438 if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
439 blk_recount_segments(q, req->biotail);
440 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
441 blk_recount_segments(q, bio);
442 len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
443 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
444 !BIOVEC_VIRT_OVERSIZE(len)) {
445 int mergeable = ll_new_mergeable(q, req, bio);
448 if (req->nr_hw_segments == 1)
449 req->bio->bi_hw_front_size = len;
450 if (bio->bi_hw_segments == 1)
451 bio->bi_hw_back_size = len;
456 return ll_new_hw_segment(q, req, bio);
459 static int ll_front_merge_fn(struct request_queue *q, struct request *req,
462 unsigned short max_sectors;
465 if (unlikely(blk_pc_request(req)))
466 max_sectors = q->max_hw_sectors;
468 max_sectors = q->max_sectors;
471 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
472 req->cmd_flags |= REQ_NOMERGE;
473 if (req == q->last_merge)
474 q->last_merge = NULL;
477 len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
478 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
479 blk_recount_segments(q, bio);
480 if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
481 blk_recount_segments(q, req->bio);
482 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
483 !BIOVEC_VIRT_OVERSIZE(len)) {
484 int mergeable = ll_new_mergeable(q, req, bio);
487 if (bio->bi_hw_segments == 1)
488 bio->bi_hw_front_size = len;
489 if (req->nr_hw_segments == 1)
490 req->biotail->bi_hw_back_size = len;
495 return ll_new_hw_segment(q, req, bio);
498 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
499 struct request *next)
501 int total_phys_segments;
502 int total_hw_segments;
505 * First check if the either of the requests are re-queued
506 * requests. Can't merge them if they are.
508 if (req->special || next->special)
512 * Will it become too large?
514 if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
517 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
518 if (blk_phys_contig_segment(q, req->biotail, next->bio))
519 total_phys_segments--;
521 if (total_phys_segments > q->max_phys_segments)
524 total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
525 if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
526 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
528 * propagate the combined length to the end of the requests
530 if (req->nr_hw_segments == 1)
531 req->bio->bi_hw_front_size = len;
532 if (next->nr_hw_segments == 1)
533 next->biotail->bi_hw_back_size = len;
537 if (total_hw_segments > q->max_hw_segments)
541 req->nr_phys_segments = total_phys_segments;
542 req->nr_hw_segments = total_hw_segments;
547 * "plug" the device if there are no outstanding requests: this will
548 * force the transfer to start only after we have put all the requests
551 * This is called with interrupts off and no requests on the queue and
552 * with the queue lock held.
554 void blk_plug_device(struct request_queue *q)
556 WARN_ON(!irqs_disabled());
559 * don't plug a stopped queue, it must be paired with blk_start_queue()
560 * which will restart the queueing
562 if (blk_queue_stopped(q))
565 if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
566 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
567 blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
571 EXPORT_SYMBOL(blk_plug_device);
574 * remove the queue from the plugged list, if present. called with
575 * queue lock held and interrupts disabled.
577 int blk_remove_plug(struct request_queue *q)
579 WARN_ON(!irqs_disabled());
581 if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
584 del_timer(&q->unplug_timer);
588 EXPORT_SYMBOL(blk_remove_plug);
591 * remove the plug and let it rip..
593 void __generic_unplug_device(struct request_queue *q)
595 if (unlikely(blk_queue_stopped(q)))
598 if (!blk_remove_plug(q))
603 EXPORT_SYMBOL(__generic_unplug_device);
606 * generic_unplug_device - fire a request queue
607 * @q: The &struct request_queue in question
610 * Linux uses plugging to build bigger requests queues before letting
611 * the device have at them. If a queue is plugged, the I/O scheduler
612 * is still adding and merging requests on the queue. Once the queue
613 * gets unplugged, the request_fn defined for the queue is invoked and
616 void generic_unplug_device(struct request_queue *q)
618 spin_lock_irq(q->queue_lock);
619 __generic_unplug_device(q);
620 spin_unlock_irq(q->queue_lock);
622 EXPORT_SYMBOL(generic_unplug_device);
624 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
627 struct request_queue *q = bdi->unplug_io_data;
632 void blk_unplug_work(struct work_struct *work)
634 struct request_queue *q =
635 container_of(work, struct request_queue, unplug_work);
637 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
638 q->rq.count[READ] + q->rq.count[WRITE]);
643 void blk_unplug_timeout(unsigned long data)
645 struct request_queue *q = (struct request_queue *)data;
647 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
648 q->rq.count[READ] + q->rq.count[WRITE]);
650 kblockd_schedule_work(&q->unplug_work);
653 void blk_unplug(struct request_queue *q)
656 * devices don't necessarily have an ->unplug_fn defined
659 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
660 q->rq.count[READ] + q->rq.count[WRITE]);
665 EXPORT_SYMBOL(blk_unplug);
668 * blk_start_queue - restart a previously stopped queue
669 * @q: The &struct request_queue in question
672 * blk_start_queue() will clear the stop flag on the queue, and call
673 * the request_fn for the queue if it was in a stopped state when
674 * entered. Also see blk_stop_queue(). Queue lock must be held.
676 void blk_start_queue(struct request_queue *q)
678 WARN_ON(!irqs_disabled());
680 clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
683 * one level of recursion is ok and is much faster than kicking
684 * the unplug handling
686 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
688 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
691 kblockd_schedule_work(&q->unplug_work);
695 EXPORT_SYMBOL(blk_start_queue);
698 * blk_stop_queue - stop a queue
699 * @q: The &struct request_queue in question
702 * The Linux block layer assumes that a block driver will consume all
703 * entries on the request queue when the request_fn strategy is called.
704 * Often this will not happen, because of hardware limitations (queue
705 * depth settings). If a device driver gets a 'queue full' response,
706 * or if it simply chooses not to queue more I/O at one point, it can
707 * call this function to prevent the request_fn from being called until
708 * the driver has signalled it's ready to go again. This happens by calling
709 * blk_start_queue() to restart queue operations. Queue lock must be held.
711 void blk_stop_queue(struct request_queue *q)
714 set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
716 EXPORT_SYMBOL(blk_stop_queue);
719 * blk_sync_queue - cancel any pending callbacks on a queue
723 * The block layer may perform asynchronous callback activity
724 * on a queue, such as calling the unplug function after a timeout.
725 * A block device may call blk_sync_queue to ensure that any
726 * such activity is cancelled, thus allowing it to release resources
727 * that the callbacks might use. The caller must already have made sure
728 * that its ->make_request_fn will not re-add plugging prior to calling
732 void blk_sync_queue(struct request_queue *q)
734 del_timer_sync(&q->unplug_timer);
735 kblockd_flush_work(&q->unplug_work);
737 EXPORT_SYMBOL(blk_sync_queue);
740 * blk_run_queue - run a single device queue
741 * @q: The queue to run
743 void blk_run_queue(struct request_queue *q)
747 spin_lock_irqsave(q->queue_lock, flags);
751 * Only recurse once to avoid overrunning the stack, let the unplug
752 * handling reinvoke the handler shortly if we already got there.
754 if (!elv_queue_empty(q)) {
755 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
757 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
760 kblockd_schedule_work(&q->unplug_work);
764 spin_unlock_irqrestore(q->queue_lock, flags);
766 EXPORT_SYMBOL(blk_run_queue);
768 void blk_put_queue(struct request_queue *q)
770 kobject_put(&q->kobj);
772 EXPORT_SYMBOL(blk_put_queue);
774 void blk_cleanup_queue(struct request_queue * q)
776 mutex_lock(&q->sysfs_lock);
777 set_bit(QUEUE_FLAG_DEAD, &q->queue_flags);
778 mutex_unlock(&q->sysfs_lock);
781 elevator_exit(q->elevator);
786 EXPORT_SYMBOL(blk_cleanup_queue);
788 static int blk_init_free_list(struct request_queue *q)
790 struct request_list *rl = &q->rq;
792 rl->count[READ] = rl->count[WRITE] = 0;
793 rl->starved[READ] = rl->starved[WRITE] = 0;
795 init_waitqueue_head(&rl->wait[READ]);
796 init_waitqueue_head(&rl->wait[WRITE]);
798 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
799 mempool_free_slab, request_cachep, q->node);
807 struct request_queue *blk_alloc_queue(gfp_t gfp_mask)
809 return blk_alloc_queue_node(gfp_mask, -1);
811 EXPORT_SYMBOL(blk_alloc_queue);
813 struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
815 struct request_queue *q;
818 q = kmem_cache_alloc_node(blk_requestq_cachep,
819 gfp_mask | __GFP_ZERO, node_id);
823 q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
824 q->backing_dev_info.unplug_io_data = q;
825 err = bdi_init(&q->backing_dev_info);
827 kmem_cache_free(blk_requestq_cachep, q);
831 init_timer(&q->unplug_timer);
833 kobject_init(&q->kobj, &blk_queue_ktype);
835 mutex_init(&q->sysfs_lock);
839 EXPORT_SYMBOL(blk_alloc_queue_node);
842 * blk_init_queue - prepare a request queue for use with a block device
843 * @rfn: The function to be called to process requests that have been
844 * placed on the queue.
845 * @lock: Request queue spin lock
848 * If a block device wishes to use the standard request handling procedures,
849 * which sorts requests and coalesces adjacent requests, then it must
850 * call blk_init_queue(). The function @rfn will be called when there
851 * are requests on the queue that need to be processed. If the device
852 * supports plugging, then @rfn may not be called immediately when requests
853 * are available on the queue, but may be called at some time later instead.
854 * Plugged queues are generally unplugged when a buffer belonging to one
855 * of the requests on the queue is needed, or due to memory pressure.
857 * @rfn is not required, or even expected, to remove all requests off the
858 * queue, but only as many as it can handle at a time. If it does leave
859 * requests on the queue, it is responsible for arranging that the requests
860 * get dealt with eventually.
862 * The queue spin lock must be held while manipulating the requests on the
863 * request queue; this lock will be taken also from interrupt context, so irq
864 * disabling is needed for it.
866 * Function returns a pointer to the initialized request queue, or NULL if
870 * blk_init_queue() must be paired with a blk_cleanup_queue() call
871 * when the block device is deactivated (such as at module unload).
874 struct request_queue *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
876 return blk_init_queue_node(rfn, lock, -1);
878 EXPORT_SYMBOL(blk_init_queue);
880 struct request_queue *
881 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
883 struct request_queue *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
889 if (blk_init_free_list(q)) {
890 kmem_cache_free(blk_requestq_cachep, q);
895 * if caller didn't supply a lock, they get per-queue locking with
899 spin_lock_init(&q->__queue_lock);
900 lock = &q->__queue_lock;
904 q->prep_rq_fn = NULL;
905 q->unplug_fn = generic_unplug_device;
906 q->queue_flags = (1 << QUEUE_FLAG_CLUSTER);
907 q->queue_lock = lock;
909 blk_queue_segment_boundary(q, 0xffffffff);
911 blk_queue_make_request(q, __make_request);
912 blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
914 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
915 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
917 q->sg_reserved_size = INT_MAX;
922 if (!elevator_init(q, NULL)) {
923 blk_queue_congestion_threshold(q);
930 EXPORT_SYMBOL(blk_init_queue_node);
932 int blk_get_queue(struct request_queue *q)
934 if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
935 kobject_get(&q->kobj);
942 EXPORT_SYMBOL(blk_get_queue);
944 static inline void blk_free_request(struct request_queue *q, struct request *rq)
946 if (rq->cmd_flags & REQ_ELVPRIV)
947 elv_put_request(q, rq);
948 mempool_free(rq, q->rq.rq_pool);
951 static struct request *
952 blk_alloc_request(struct request_queue *q, int rw, int priv, gfp_t gfp_mask)
954 struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
960 * first three bits are identical in rq->cmd_flags and bio->bi_rw,
961 * see bio.h and blkdev.h
963 rq->cmd_flags = rw | REQ_ALLOCED;
966 if (unlikely(elv_set_request(q, rq, gfp_mask))) {
967 mempool_free(rq, q->rq.rq_pool);
970 rq->cmd_flags |= REQ_ELVPRIV;
977 * ioc_batching returns true if the ioc is a valid batching request and
978 * should be given priority access to a request.
980 static inline int ioc_batching(struct request_queue *q, struct io_context *ioc)
986 * Make sure the process is able to allocate at least 1 request
987 * even if the batch times out, otherwise we could theoretically
990 return ioc->nr_batch_requests == q->nr_batching ||
991 (ioc->nr_batch_requests > 0
992 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
996 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
997 * will cause the process to be a "batcher" on all queues in the system. This
998 * is the behaviour we want though - once it gets a wakeup it should be given
1001 static void ioc_set_batching(struct request_queue *q, struct io_context *ioc)
1003 if (!ioc || ioc_batching(q, ioc))
1006 ioc->nr_batch_requests = q->nr_batching;
1007 ioc->last_waited = jiffies;
1010 static void __freed_request(struct request_queue *q, int rw)
1012 struct request_list *rl = &q->rq;
1014 if (rl->count[rw] < queue_congestion_off_threshold(q))
1015 blk_clear_queue_congested(q, rw);
1017 if (rl->count[rw] + 1 <= q->nr_requests) {
1018 if (waitqueue_active(&rl->wait[rw]))
1019 wake_up(&rl->wait[rw]);
1021 blk_clear_queue_full(q, rw);
1026 * A request has just been released. Account for it, update the full and
1027 * congestion status, wake up any waiters. Called under q->queue_lock.
1029 static void freed_request(struct request_queue *q, int rw, int priv)
1031 struct request_list *rl = &q->rq;
1037 __freed_request(q, rw);
1039 if (unlikely(rl->starved[rw ^ 1]))
1040 __freed_request(q, rw ^ 1);
1043 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
1045 * Get a free request, queue_lock must be held.
1046 * Returns NULL on failure, with queue_lock held.
1047 * Returns !NULL on success, with queue_lock *not held*.
1049 static struct request *get_request(struct request_queue *q, int rw_flags,
1050 struct bio *bio, gfp_t gfp_mask)
1052 struct request *rq = NULL;
1053 struct request_list *rl = &q->rq;
1054 struct io_context *ioc = NULL;
1055 const int rw = rw_flags & 0x01;
1056 int may_queue, priv;
1058 may_queue = elv_may_queue(q, rw_flags);
1059 if (may_queue == ELV_MQUEUE_NO)
1062 if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
1063 if (rl->count[rw]+1 >= q->nr_requests) {
1064 ioc = current_io_context(GFP_ATOMIC, q->node);
1066 * The queue will fill after this allocation, so set
1067 * it as full, and mark this process as "batching".
1068 * This process will be allowed to complete a batch of
1069 * requests, others will be blocked.
1071 if (!blk_queue_full(q, rw)) {
1072 ioc_set_batching(q, ioc);
1073 blk_set_queue_full(q, rw);
1075 if (may_queue != ELV_MQUEUE_MUST
1076 && !ioc_batching(q, ioc)) {
1078 * The queue is full and the allocating
1079 * process is not a "batcher", and not
1080 * exempted by the IO scheduler
1086 blk_set_queue_congested(q, rw);
1090 * Only allow batching queuers to allocate up to 50% over the defined
1091 * limit of requests, otherwise we could have thousands of requests
1092 * allocated with any setting of ->nr_requests
1094 if (rl->count[rw] >= (3 * q->nr_requests / 2))
1098 rl->starved[rw] = 0;
1100 priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1104 spin_unlock_irq(q->queue_lock);
1106 rq = blk_alloc_request(q, rw_flags, priv, gfp_mask);
1107 if (unlikely(!rq)) {
1109 * Allocation failed presumably due to memory. Undo anything
1110 * we might have messed up.
1112 * Allocating task should really be put onto the front of the
1113 * wait queue, but this is pretty rare.
1115 spin_lock_irq(q->queue_lock);
1116 freed_request(q, rw, priv);
1119 * in the very unlikely event that allocation failed and no
1120 * requests for this direction was pending, mark us starved
1121 * so that freeing of a request in the other direction will
1122 * notice us. another possible fix would be to split the
1123 * rq mempool into READ and WRITE
1126 if (unlikely(rl->count[rw] == 0))
1127 rl->starved[rw] = 1;
1133 * ioc may be NULL here, and ioc_batching will be false. That's
1134 * OK, if the queue is under the request limit then requests need
1135 * not count toward the nr_batch_requests limit. There will always
1136 * be some limit enforced by BLK_BATCH_TIME.
1138 if (ioc_batching(q, ioc))
1139 ioc->nr_batch_requests--;
1143 blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
1149 * No available requests for this queue, unplug the device and wait for some
1150 * requests to become available.
1152 * Called with q->queue_lock held, and returns with it unlocked.
1154 static struct request *get_request_wait(struct request_queue *q, int rw_flags,
1157 const int rw = rw_flags & 0x01;
1160 rq = get_request(q, rw_flags, bio, GFP_NOIO);
1163 struct request_list *rl = &q->rq;
1165 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
1166 TASK_UNINTERRUPTIBLE);
1168 rq = get_request(q, rw_flags, bio, GFP_NOIO);
1171 struct io_context *ioc;
1173 blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
1175 __generic_unplug_device(q);
1176 spin_unlock_irq(q->queue_lock);
1180 * After sleeping, we become a "batching" process and
1181 * will be able to allocate at least one request, and
1182 * up to a big batch of them for a small period time.
1183 * See ioc_batching, ioc_set_batching
1185 ioc = current_io_context(GFP_NOIO, q->node);
1186 ioc_set_batching(q, ioc);
1188 spin_lock_irq(q->queue_lock);
1190 finish_wait(&rl->wait[rw], &wait);
1196 struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
1200 BUG_ON(rw != READ && rw != WRITE);
1202 spin_lock_irq(q->queue_lock);
1203 if (gfp_mask & __GFP_WAIT) {
1204 rq = get_request_wait(q, rw, NULL);
1206 rq = get_request(q, rw, NULL, gfp_mask);
1208 spin_unlock_irq(q->queue_lock);
1210 /* q->queue_lock is unlocked at this point */
1214 EXPORT_SYMBOL(blk_get_request);
1217 * blk_start_queueing - initiate dispatch of requests to device
1218 * @q: request queue to kick into gear
1220 * This is basically a helper to remove the need to know whether a queue
1221 * is plugged or not if someone just wants to initiate dispatch of requests
1224 * The queue lock must be held with interrupts disabled.
1226 void blk_start_queueing(struct request_queue *q)
1228 if (!blk_queue_plugged(q))
1231 __generic_unplug_device(q);
1233 EXPORT_SYMBOL(blk_start_queueing);
1236 * blk_requeue_request - put a request back on queue
1237 * @q: request queue where request should be inserted
1238 * @rq: request to be inserted
1241 * Drivers often keep queueing requests until the hardware cannot accept
1242 * more, when that condition happens we need to put the request back
1243 * on the queue. Must be called with queue lock held.
1245 void blk_requeue_request(struct request_queue *q, struct request *rq)
1247 blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
1249 if (blk_rq_tagged(rq))
1250 blk_queue_end_tag(q, rq);
1252 elv_requeue_request(q, rq);
1255 EXPORT_SYMBOL(blk_requeue_request);
1258 * blk_insert_request - insert a special request in to a request queue
1259 * @q: request queue where request should be inserted
1260 * @rq: request to be inserted
1261 * @at_head: insert request at head or tail of queue
1262 * @data: private data
1265 * Many block devices need to execute commands asynchronously, so they don't
1266 * block the whole kernel from preemption during request execution. This is
1267 * accomplished normally by inserting aritficial requests tagged as
1268 * REQ_SPECIAL in to the corresponding request queue, and letting them be
1269 * scheduled for actual execution by the request queue.
1271 * We have the option of inserting the head or the tail of the queue.
1272 * Typically we use the tail for new ioctls and so forth. We use the head
1273 * of the queue for things like a QUEUE_FULL message from a device, or a
1274 * host that is unable to accept a particular command.
1276 void blk_insert_request(struct request_queue *q, struct request *rq,
1277 int at_head, void *data)
1279 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
1280 unsigned long flags;
1283 * tell I/O scheduler that this isn't a regular read/write (ie it
1284 * must not attempt merges on this) and that it acts as a soft
1287 rq->cmd_type = REQ_TYPE_SPECIAL;
1288 rq->cmd_flags |= REQ_SOFTBARRIER;
1292 spin_lock_irqsave(q->queue_lock, flags);
1295 * If command is tagged, release the tag
1297 if (blk_rq_tagged(rq))
1298 blk_queue_end_tag(q, rq);
1300 drive_stat_acct(rq, 1);
1301 __elv_add_request(q, rq, where, 0);
1302 blk_start_queueing(q);
1303 spin_unlock_irqrestore(q->queue_lock, flags);
1306 EXPORT_SYMBOL(blk_insert_request);
1308 static void drive_stat_acct(struct request *rq, int new_io)
1310 int rw = rq_data_dir(rq);
1312 if (!blk_fs_request(rq) || !rq->rq_disk)
1316 __disk_stat_inc(rq->rq_disk, merges[rw]);
1318 disk_round_stats(rq->rq_disk);
1319 rq->rq_disk->in_flight++;
1324 * add-request adds a request to the linked list.
1325 * queue lock is held and interrupts disabled, as we muck with the
1326 * request queue list.
1328 static inline void add_request(struct request_queue * q, struct request * req)
1330 drive_stat_acct(req, 1);
1333 * elevator indicated where it wants this request to be
1334 * inserted at elevator_merge time
1336 __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
1340 * disk_round_stats() - Round off the performance stats on a struct
1343 * The average IO queue length and utilisation statistics are maintained
1344 * by observing the current state of the queue length and the amount of
1345 * time it has been in this state for.
1347 * Normally, that accounting is done on IO completion, but that can result
1348 * in more than a second's worth of IO being accounted for within any one
1349 * second, leading to >100% utilisation. To deal with that, we call this
1350 * function to do a round-off before returning the results when reading
1351 * /proc/diskstats. This accounts immediately for all queue usage up to
1352 * the current jiffies and restarts the counters again.
1354 void disk_round_stats(struct gendisk *disk)
1356 unsigned long now = jiffies;
1358 if (now == disk->stamp)
1361 if (disk->in_flight) {
1362 __disk_stat_add(disk, time_in_queue,
1363 disk->in_flight * (now - disk->stamp));
1364 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
1369 EXPORT_SYMBOL_GPL(disk_round_stats);
1372 * queue lock must be held
1374 void __blk_put_request(struct request_queue *q, struct request *req)
1378 if (unlikely(--req->ref_count))
1381 elv_completed_request(q, req);
1384 * Request may not have originated from ll_rw_blk. if not,
1385 * it didn't come out of our reserved rq pools
1387 if (req->cmd_flags & REQ_ALLOCED) {
1388 int rw = rq_data_dir(req);
1389 int priv = req->cmd_flags & REQ_ELVPRIV;
1391 BUG_ON(!list_empty(&req->queuelist));
1392 BUG_ON(!hlist_unhashed(&req->hash));
1394 blk_free_request(q, req);
1395 freed_request(q, rw, priv);
1399 EXPORT_SYMBOL_GPL(__blk_put_request);
1401 void blk_put_request(struct request *req)
1403 unsigned long flags;
1404 struct request_queue *q = req->q;
1407 * Gee, IDE calls in w/ NULL q. Fix IDE and remove the
1408 * following if (q) test.
1411 spin_lock_irqsave(q->queue_lock, flags);
1412 __blk_put_request(q, req);
1413 spin_unlock_irqrestore(q->queue_lock, flags);
1417 EXPORT_SYMBOL(blk_put_request);
1420 * Has to be called with the request spinlock acquired
1422 static int attempt_merge(struct request_queue *q, struct request *req,
1423 struct request *next)
1425 if (!rq_mergeable(req) || !rq_mergeable(next))
1431 if (req->sector + req->nr_sectors != next->sector)
1434 if (rq_data_dir(req) != rq_data_dir(next)
1435 || req->rq_disk != next->rq_disk
1440 * If we are allowed to merge, then append bio list
1441 * from next to rq and release next. merge_requests_fn
1442 * will have updated segment counts, update sector
1445 if (!ll_merge_requests_fn(q, req, next))
1449 * At this point we have either done a back merge
1450 * or front merge. We need the smaller start_time of
1451 * the merged requests to be the current request
1452 * for accounting purposes.
1454 if (time_after(req->start_time, next->start_time))
1455 req->start_time = next->start_time;
1457 req->biotail->bi_next = next->bio;
1458 req->biotail = next->biotail;
1460 req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
1462 elv_merge_requests(q, req, next);
1465 disk_round_stats(req->rq_disk);
1466 req->rq_disk->in_flight--;
1469 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
1471 __blk_put_request(q, next);
1475 static inline int attempt_back_merge(struct request_queue *q,
1478 struct request *next = elv_latter_request(q, rq);
1481 return attempt_merge(q, rq, next);
1486 static inline int attempt_front_merge(struct request_queue *q,
1489 struct request *prev = elv_former_request(q, rq);
1492 return attempt_merge(q, prev, rq);
1497 void init_request_from_bio(struct request *req, struct bio *bio)
1499 req->cmd_type = REQ_TYPE_FS;
1502 * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
1504 if (bio_rw_ahead(bio) || bio_failfast(bio))
1505 req->cmd_flags |= REQ_FAILFAST;
1508 * REQ_BARRIER implies no merging, but lets make it explicit
1510 if (unlikely(bio_barrier(bio)))
1511 req->cmd_flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
1514 req->cmd_flags |= REQ_RW_SYNC;
1515 if (bio_rw_meta(bio))
1516 req->cmd_flags |= REQ_RW_META;
1519 req->hard_sector = req->sector = bio->bi_sector;
1520 req->ioprio = bio_prio(bio);
1521 req->start_time = jiffies;
1522 blk_rq_bio_prep(req->q, req, bio);
1525 static int __make_request(struct request_queue *q, struct bio *bio)
1527 struct request *req;
1528 int el_ret, nr_sectors, barrier, err;
1529 const unsigned short prio = bio_prio(bio);
1530 const int sync = bio_sync(bio);
1533 nr_sectors = bio_sectors(bio);
1536 * low level driver can indicate that it wants pages above a
1537 * certain limit bounced to low memory (ie for highmem, or even
1538 * ISA dma in theory)
1540 blk_queue_bounce(q, &bio);
1542 barrier = bio_barrier(bio);
1543 if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
1548 spin_lock_irq(q->queue_lock);
1550 if (unlikely(barrier) || elv_queue_empty(q))
1553 el_ret = elv_merge(q, &req, bio);
1555 case ELEVATOR_BACK_MERGE:
1556 BUG_ON(!rq_mergeable(req));
1558 if (!ll_back_merge_fn(q, req, bio))
1561 blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
1563 req->biotail->bi_next = bio;
1565 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
1566 req->ioprio = ioprio_best(req->ioprio, prio);
1567 drive_stat_acct(req, 0);
1568 if (!attempt_back_merge(q, req))
1569 elv_merged_request(q, req, el_ret);
1572 case ELEVATOR_FRONT_MERGE:
1573 BUG_ON(!rq_mergeable(req));
1575 if (!ll_front_merge_fn(q, req, bio))
1578 blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
1580 bio->bi_next = req->bio;
1584 * may not be valid. if the low level driver said
1585 * it didn't need a bounce buffer then it better
1586 * not touch req->buffer either...
1588 req->buffer = bio_data(bio);
1589 req->current_nr_sectors = bio_cur_sectors(bio);
1590 req->hard_cur_sectors = req->current_nr_sectors;
1591 req->sector = req->hard_sector = bio->bi_sector;
1592 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
1593 req->ioprio = ioprio_best(req->ioprio, prio);
1594 drive_stat_acct(req, 0);
1595 if (!attempt_front_merge(q, req))
1596 elv_merged_request(q, req, el_ret);
1599 /* ELV_NO_MERGE: elevator says don't/can't merge. */
1606 * This sync check and mask will be re-done in init_request_from_bio(),
1607 * but we need to set it earlier to expose the sync flag to the
1608 * rq allocator and io schedulers.
1610 rw_flags = bio_data_dir(bio);
1612 rw_flags |= REQ_RW_SYNC;
1615 * Grab a free request. This is might sleep but can not fail.
1616 * Returns with the queue unlocked.
1618 req = get_request_wait(q, rw_flags, bio);
1621 * After dropping the lock and possibly sleeping here, our request
1622 * may now be mergeable after it had proven unmergeable (above).
1623 * We don't worry about that case for efficiency. It won't happen
1624 * often, and the elevators are able to handle it.
1626 init_request_from_bio(req, bio);
1628 spin_lock_irq(q->queue_lock);
1629 if (elv_queue_empty(q))
1631 add_request(q, req);
1634 __generic_unplug_device(q);
1636 spin_unlock_irq(q->queue_lock);
1640 bio_endio(bio, err);
1645 * If bio->bi_dev is a partition, remap the location
1647 static inline void blk_partition_remap(struct bio *bio)
1649 struct block_device *bdev = bio->bi_bdev;
1651 if (bio_sectors(bio) && bdev != bdev->bd_contains) {
1652 struct hd_struct *p = bdev->bd_part;
1653 const int rw = bio_data_dir(bio);
1655 p->sectors[rw] += bio_sectors(bio);
1658 bio->bi_sector += p->start_sect;
1659 bio->bi_bdev = bdev->bd_contains;
1661 blk_add_trace_remap(bdev_get_queue(bio->bi_bdev), bio,
1662 bdev->bd_dev, bio->bi_sector,
1663 bio->bi_sector - p->start_sect);
1667 static void handle_bad_sector(struct bio *bio)
1669 char b[BDEVNAME_SIZE];
1671 printk(KERN_INFO "attempt to access beyond end of device\n");
1672 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
1673 bdevname(bio->bi_bdev, b),
1675 (unsigned long long)bio->bi_sector + bio_sectors(bio),
1676 (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
1678 set_bit(BIO_EOF, &bio->bi_flags);
1681 #ifdef CONFIG_FAIL_MAKE_REQUEST
1683 static DECLARE_FAULT_ATTR(fail_make_request);
1685 static int __init setup_fail_make_request(char *str)
1687 return setup_fault_attr(&fail_make_request, str);
1689 __setup("fail_make_request=", setup_fail_make_request);
1691 static int should_fail_request(struct bio *bio)
1693 if ((bio->bi_bdev->bd_disk->flags & GENHD_FL_FAIL) ||
1694 (bio->bi_bdev->bd_part && bio->bi_bdev->bd_part->make_it_fail))
1695 return should_fail(&fail_make_request, bio->bi_size);
1700 static int __init fail_make_request_debugfs(void)
1702 return init_fault_attr_dentries(&fail_make_request,
1703 "fail_make_request");
1706 late_initcall(fail_make_request_debugfs);
1708 #else /* CONFIG_FAIL_MAKE_REQUEST */
1710 static inline int should_fail_request(struct bio *bio)
1715 #endif /* CONFIG_FAIL_MAKE_REQUEST */
1718 * Check whether this bio extends beyond the end of the device.
1720 static inline int bio_check_eod(struct bio *bio, unsigned int nr_sectors)
1727 /* Test device or partition size, when known. */
1728 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
1730 sector_t sector = bio->bi_sector;
1732 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
1734 * This may well happen - the kernel calls bread()
1735 * without checking the size of the device, e.g., when
1736 * mounting a device.
1738 handle_bad_sector(bio);
1747 * generic_make_request: hand a buffer to its device driver for I/O
1748 * @bio: The bio describing the location in memory and on the device.
1750 * generic_make_request() is used to make I/O requests of block
1751 * devices. It is passed a &struct bio, which describes the I/O that needs
1754 * generic_make_request() does not return any status. The
1755 * success/failure status of the request, along with notification of
1756 * completion, is delivered asynchronously through the bio->bi_end_io
1757 * function described (one day) else where.
1759 * The caller of generic_make_request must make sure that bi_io_vec
1760 * are set to describe the memory buffer, and that bi_dev and bi_sector are
1761 * set to describe the device address, and the
1762 * bi_end_io and optionally bi_private are set to describe how
1763 * completion notification should be signaled.
1765 * generic_make_request and the drivers it calls may use bi_next if this
1766 * bio happens to be merged with someone else, and may change bi_dev and
1767 * bi_sector for remaps as it sees fit. So the values of these fields
1768 * should NOT be depended on after the call to generic_make_request.
1770 static inline void __generic_make_request(struct bio *bio)
1772 struct request_queue *q;
1773 sector_t old_sector;
1774 int ret, nr_sectors = bio_sectors(bio);
1780 if (bio_check_eod(bio, nr_sectors))
1784 * Resolve the mapping until finished. (drivers are
1785 * still free to implement/resolve their own stacking
1786 * by explicitly returning 0)
1788 * NOTE: we don't repeat the blk_size check for each new device.
1789 * Stacking drivers are expected to know what they are doing.
1794 char b[BDEVNAME_SIZE];
1796 q = bdev_get_queue(bio->bi_bdev);
1799 "generic_make_request: Trying to access "
1800 "nonexistent block-device %s (%Lu)\n",
1801 bdevname(bio->bi_bdev, b),
1802 (long long) bio->bi_sector);
1804 bio_endio(bio, err);
1808 if (unlikely(nr_sectors > q->max_hw_sectors)) {
1809 printk("bio too big device %s (%u > %u)\n",
1810 bdevname(bio->bi_bdev, b),
1816 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
1819 if (should_fail_request(bio))
1823 * If this device has partitions, remap block n
1824 * of partition p to block n+start(p) of the disk.
1826 blk_partition_remap(bio);
1828 if (old_sector != -1)
1829 blk_add_trace_remap(q, bio, old_dev, bio->bi_sector,
1832 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
1834 old_sector = bio->bi_sector;
1835 old_dev = bio->bi_bdev->bd_dev;
1837 if (bio_check_eod(bio, nr_sectors))
1839 if (bio_empty_barrier(bio) && !q->prepare_flush_fn) {
1844 ret = q->make_request_fn(q, bio);
1849 * We only want one ->make_request_fn to be active at a time,
1850 * else stack usage with stacked devices could be a problem.
1851 * So use current->bio_{list,tail} to keep a list of requests
1852 * submited by a make_request_fn function.
1853 * current->bio_tail is also used as a flag to say if
1854 * generic_make_request is currently active in this task or not.
1855 * If it is NULL, then no make_request is active. If it is non-NULL,
1856 * then a make_request is active, and new requests should be added
1859 void generic_make_request(struct bio *bio)
1861 if (current->bio_tail) {
1862 /* make_request is active */
1863 *(current->bio_tail) = bio;
1864 bio->bi_next = NULL;
1865 current->bio_tail = &bio->bi_next;
1868 /* following loop may be a bit non-obvious, and so deserves some
1870 * Before entering the loop, bio->bi_next is NULL (as all callers
1871 * ensure that) so we have a list with a single bio.
1872 * We pretend that we have just taken it off a longer list, so
1873 * we assign bio_list to the next (which is NULL) and bio_tail
1874 * to &bio_list, thus initialising the bio_list of new bios to be
1875 * added. __generic_make_request may indeed add some more bios
1876 * through a recursive call to generic_make_request. If it
1877 * did, we find a non-NULL value in bio_list and re-enter the loop
1878 * from the top. In this case we really did just take the bio
1879 * of the top of the list (no pretending) and so fixup bio_list and
1880 * bio_tail or bi_next, and call into __generic_make_request again.
1882 * The loop was structured like this to make only one call to
1883 * __generic_make_request (which is important as it is large and
1884 * inlined) and to keep the structure simple.
1886 BUG_ON(bio->bi_next);
1888 current->bio_list = bio->bi_next;
1889 if (bio->bi_next == NULL)
1890 current->bio_tail = ¤t->bio_list;
1892 bio->bi_next = NULL;
1893 __generic_make_request(bio);
1894 bio = current->bio_list;
1896 current->bio_tail = NULL; /* deactivate */
1899 EXPORT_SYMBOL(generic_make_request);
1902 * submit_bio: submit a bio to the block device layer for I/O
1903 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
1904 * @bio: The &struct bio which describes the I/O
1906 * submit_bio() is very similar in purpose to generic_make_request(), and
1907 * uses that function to do most of the work. Both are fairly rough
1908 * interfaces, @bio must be presetup and ready for I/O.
1911 void submit_bio(int rw, struct bio *bio)
1913 int count = bio_sectors(bio);
1918 * If it's a regular read/write or a barrier with data attached,
1919 * go through the normal accounting stuff before submission.
1921 if (!bio_empty_barrier(bio)) {
1923 BIO_BUG_ON(!bio->bi_size);
1924 BIO_BUG_ON(!bio->bi_io_vec);
1927 count_vm_events(PGPGOUT, count);
1929 task_io_account_read(bio->bi_size);
1930 count_vm_events(PGPGIN, count);
1933 if (unlikely(block_dump)) {
1934 char b[BDEVNAME_SIZE];
1935 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
1936 current->comm, task_pid_nr(current),
1937 (rw & WRITE) ? "WRITE" : "READ",
1938 (unsigned long long)bio->bi_sector,
1939 bdevname(bio->bi_bdev,b));
1943 generic_make_request(bio);
1946 EXPORT_SYMBOL(submit_bio);
1948 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
1950 if (blk_fs_request(rq)) {
1951 rq->hard_sector += nsect;
1952 rq->hard_nr_sectors -= nsect;
1955 * Move the I/O submission pointers ahead if required.
1957 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
1958 (rq->sector <= rq->hard_sector)) {
1959 rq->sector = rq->hard_sector;
1960 rq->nr_sectors = rq->hard_nr_sectors;
1961 rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
1962 rq->current_nr_sectors = rq->hard_cur_sectors;
1963 rq->buffer = bio_data(rq->bio);
1967 * if total number of sectors is less than the first segment
1968 * size, something has gone terribly wrong
1970 if (rq->nr_sectors < rq->current_nr_sectors) {
1971 printk("blk: request botched\n");
1972 rq->nr_sectors = rq->current_nr_sectors;
1978 * __end_that_request_first - end I/O on a request
1979 * @req: the request being processed
1980 * @error: 0 for success, < 0 for error
1981 * @nr_bytes: number of bytes to complete
1984 * Ends I/O on a number of bytes attached to @req, and sets it up
1985 * for the next range of segments (if any) in the cluster.
1988 * 0 - we are done with this request, call end_that_request_last()
1989 * 1 - still buffers pending for this request
1991 static int __end_that_request_first(struct request *req, int error,
1994 int total_bytes, bio_nbytes, next_idx = 0;
1997 blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
2000 * for a REQ_BLOCK_PC request, we want to carry any eventual
2001 * sense key with us all the way through
2003 if (!blk_pc_request(req))
2007 if (blk_fs_request(req) && !(req->cmd_flags & REQ_QUIET))
2008 printk("end_request: I/O error, dev %s, sector %llu\n",
2009 req->rq_disk ? req->rq_disk->disk_name : "?",
2010 (unsigned long long)req->sector);
2013 if (blk_fs_request(req) && req->rq_disk) {
2014 const int rw = rq_data_dir(req);
2016 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
2019 total_bytes = bio_nbytes = 0;
2020 while ((bio = req->bio) != NULL) {
2024 * For an empty barrier request, the low level driver must
2025 * store a potential error location in ->sector. We pass
2026 * that back up in ->bi_sector.
2028 if (blk_empty_barrier(req))
2029 bio->bi_sector = req->sector;
2031 if (nr_bytes >= bio->bi_size) {
2032 req->bio = bio->bi_next;
2033 nbytes = bio->bi_size;
2034 req_bio_endio(req, bio, nbytes, error);
2038 int idx = bio->bi_idx + next_idx;
2040 if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
2041 blk_dump_rq_flags(req, "__end_that");
2042 printk("%s: bio idx %d >= vcnt %d\n",
2044 bio->bi_idx, bio->bi_vcnt);
2048 nbytes = bio_iovec_idx(bio, idx)->bv_len;
2049 BIO_BUG_ON(nbytes > bio->bi_size);
2052 * not a complete bvec done
2054 if (unlikely(nbytes > nr_bytes)) {
2055 bio_nbytes += nr_bytes;
2056 total_bytes += nr_bytes;
2061 * advance to the next vector
2064 bio_nbytes += nbytes;
2067 total_bytes += nbytes;
2070 if ((bio = req->bio)) {
2072 * end more in this run, or just return 'not-done'
2074 if (unlikely(nr_bytes <= 0))
2086 * if the request wasn't completed, update state
2089 req_bio_endio(req, bio, bio_nbytes, error);
2090 bio->bi_idx += next_idx;
2091 bio_iovec(bio)->bv_offset += nr_bytes;
2092 bio_iovec(bio)->bv_len -= nr_bytes;
2095 blk_recalc_rq_sectors(req, total_bytes >> 9);
2096 blk_recalc_rq_segments(req);
2101 * splice the completion data to a local structure and hand off to
2102 * process_completion_queue() to complete the requests
2104 static void blk_done_softirq(struct softirq_action *h)
2106 struct list_head *cpu_list, local_list;
2108 local_irq_disable();
2109 cpu_list = &__get_cpu_var(blk_cpu_done);
2110 list_replace_init(cpu_list, &local_list);
2113 while (!list_empty(&local_list)) {
2114 struct request *rq = list_entry(local_list.next, struct request, donelist);
2116 list_del_init(&rq->donelist);
2117 rq->q->softirq_done_fn(rq);
2121 static int __cpuinit blk_cpu_notify(struct notifier_block *self, unsigned long action,
2125 * If a CPU goes away, splice its entries to the current CPU
2126 * and trigger a run of the softirq
2128 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
2129 int cpu = (unsigned long) hcpu;
2131 local_irq_disable();
2132 list_splice_init(&per_cpu(blk_cpu_done, cpu),
2133 &__get_cpu_var(blk_cpu_done));
2134 raise_softirq_irqoff(BLOCK_SOFTIRQ);
2142 static struct notifier_block blk_cpu_notifier __cpuinitdata = {
2143 .notifier_call = blk_cpu_notify,
2147 * blk_complete_request - end I/O on a request
2148 * @req: the request being processed
2151 * Ends all I/O on a request. It does not handle partial completions,
2152 * unless the driver actually implements this in its completion callback
2153 * through requeueing. The actual completion happens out-of-order,
2154 * through a softirq handler. The user must have registered a completion
2155 * callback through blk_queue_softirq_done().
2158 void blk_complete_request(struct request *req)
2160 struct list_head *cpu_list;
2161 unsigned long flags;
2163 BUG_ON(!req->q->softirq_done_fn);
2165 local_irq_save(flags);
2167 cpu_list = &__get_cpu_var(blk_cpu_done);
2168 list_add_tail(&req->donelist, cpu_list);
2169 raise_softirq_irqoff(BLOCK_SOFTIRQ);
2171 local_irq_restore(flags);
2174 EXPORT_SYMBOL(blk_complete_request);
2177 * queue lock must be held
2179 static void end_that_request_last(struct request *req, int error)
2181 struct gendisk *disk = req->rq_disk;
2183 if (blk_rq_tagged(req))
2184 blk_queue_end_tag(req->q, req);
2186 if (blk_queued_rq(req))
2187 blkdev_dequeue_request(req);
2189 if (unlikely(laptop_mode) && blk_fs_request(req))
2190 laptop_io_completion();
2193 * Account IO completion. bar_rq isn't accounted as a normal
2194 * IO on queueing nor completion. Accounting the containing
2195 * request is enough.
2197 if (disk && blk_fs_request(req) && req != &req->q->bar_rq) {
2198 unsigned long duration = jiffies - req->start_time;
2199 const int rw = rq_data_dir(req);
2201 __disk_stat_inc(disk, ios[rw]);
2202 __disk_stat_add(disk, ticks[rw], duration);
2203 disk_round_stats(disk);
2208 req->end_io(req, error);
2210 if (blk_bidi_rq(req))
2211 __blk_put_request(req->next_rq->q, req->next_rq);
2213 __blk_put_request(req->q, req);
2217 static inline void __end_request(struct request *rq, int uptodate,
2218 unsigned int nr_bytes)
2223 error = uptodate ? uptodate : -EIO;
2225 __blk_end_request(rq, error, nr_bytes);
2229 * blk_rq_bytes - Returns bytes left to complete in the entire request
2231 unsigned int blk_rq_bytes(struct request *rq)
2233 if (blk_fs_request(rq))
2234 return rq->hard_nr_sectors << 9;
2236 return rq->data_len;
2238 EXPORT_SYMBOL_GPL(blk_rq_bytes);
2241 * blk_rq_cur_bytes - Returns bytes left to complete in the current segment
2243 unsigned int blk_rq_cur_bytes(struct request *rq)
2245 if (blk_fs_request(rq))
2246 return rq->current_nr_sectors << 9;
2249 return rq->bio->bi_size;
2251 return rq->data_len;
2253 EXPORT_SYMBOL_GPL(blk_rq_cur_bytes);
2256 * end_queued_request - end all I/O on a queued request
2257 * @rq: the request being processed
2258 * @uptodate: error value or 0/1 uptodate flag
2261 * Ends all I/O on a request, and removes it from the block layer queues.
2262 * Not suitable for normal IO completion, unless the driver still has
2263 * the request attached to the block layer.
2266 void end_queued_request(struct request *rq, int uptodate)
2268 __end_request(rq, uptodate, blk_rq_bytes(rq));
2270 EXPORT_SYMBOL(end_queued_request);
2273 * end_dequeued_request - end all I/O on a dequeued request
2274 * @rq: the request being processed
2275 * @uptodate: error value or 0/1 uptodate flag
2278 * Ends all I/O on a request. The request must already have been
2279 * dequeued using blkdev_dequeue_request(), as is normally the case
2283 void end_dequeued_request(struct request *rq, int uptodate)
2285 __end_request(rq, uptodate, blk_rq_bytes(rq));
2287 EXPORT_SYMBOL(end_dequeued_request);
2291 * end_request - end I/O on the current segment of the request
2292 * @req: the request being processed
2293 * @uptodate: error value or 0/1 uptodate flag
2296 * Ends I/O on the current segment of a request. If that is the only
2297 * remaining segment, the request is also completed and freed.
2299 * This is a remnant of how older block drivers handled IO completions.
2300 * Modern drivers typically end IO on the full request in one go, unless
2301 * they have a residual value to account for. For that case this function
2302 * isn't really useful, unless the residual just happens to be the
2303 * full current segment. In other words, don't use this function in new
2304 * code. Either use end_request_completely(), or the
2305 * end_that_request_chunk() (along with end_that_request_last()) for
2306 * partial completions.
2309 void end_request(struct request *req, int uptodate)
2311 __end_request(req, uptodate, req->hard_cur_sectors << 9);
2313 EXPORT_SYMBOL(end_request);
2316 * blk_end_io - Generic end_io function to complete a request.
2317 * @rq: the request being processed
2318 * @error: 0 for success, < 0 for error
2319 * @nr_bytes: number of bytes to complete @rq
2320 * @bidi_bytes: number of bytes to complete @rq->next_rq
2321 * @drv_callback: function called between completion of bios in the request
2322 * and completion of the request.
2323 * If the callback returns non 0, this helper returns without
2324 * completion of the request.
2327 * Ends I/O on a number of bytes attached to @rq and @rq->next_rq.
2328 * If @rq has leftover, sets it up for the next range of segments.
2331 * 0 - we are done with this request
2332 * 1 - this request is not freed yet, it still has pending buffers.
2334 static int blk_end_io(struct request *rq, int error, int nr_bytes,
2335 int bidi_bytes, int (drv_callback)(struct request *))
2337 struct request_queue *q = rq->q;
2338 unsigned long flags = 0UL;
2340 if (blk_fs_request(rq) || blk_pc_request(rq)) {
2341 if (__end_that_request_first(rq, error, nr_bytes))
2344 /* Bidi request must be completed as a whole */
2345 if (blk_bidi_rq(rq) &&
2346 __end_that_request_first(rq->next_rq, error, bidi_bytes))
2350 /* Special feature for tricky drivers */
2351 if (drv_callback && drv_callback(rq))
2354 add_disk_randomness(rq->rq_disk);
2356 spin_lock_irqsave(q->queue_lock, flags);
2357 end_that_request_last(rq, error);
2358 spin_unlock_irqrestore(q->queue_lock, flags);
2364 * blk_end_request - Helper function for drivers to complete the request.
2365 * @rq: the request being processed
2366 * @error: 0 for success, < 0 for error
2367 * @nr_bytes: number of bytes to complete
2370 * Ends I/O on a number of bytes attached to @rq.
2371 * If @rq has leftover, sets it up for the next range of segments.
2374 * 0 - we are done with this request
2375 * 1 - still buffers pending for this request
2377 int blk_end_request(struct request *rq, int error, int nr_bytes)
2379 return blk_end_io(rq, error, nr_bytes, 0, NULL);
2381 EXPORT_SYMBOL_GPL(blk_end_request);
2384 * __blk_end_request - Helper function for drivers to complete the request.
2385 * @rq: the request being processed
2386 * @error: 0 for success, < 0 for error
2387 * @nr_bytes: number of bytes to complete
2390 * Must be called with queue lock held unlike blk_end_request().
2393 * 0 - we are done with this request
2394 * 1 - still buffers pending for this request
2396 int __blk_end_request(struct request *rq, int error, int nr_bytes)
2398 if (blk_fs_request(rq) || blk_pc_request(rq)) {
2399 if (__end_that_request_first(rq, error, nr_bytes))
2403 add_disk_randomness(rq->rq_disk);
2405 end_that_request_last(rq, error);
2409 EXPORT_SYMBOL_GPL(__blk_end_request);
2412 * blk_end_bidi_request - Helper function for drivers to complete bidi request.
2413 * @rq: the bidi request being processed
2414 * @error: 0 for success, < 0 for error
2415 * @nr_bytes: number of bytes to complete @rq
2416 * @bidi_bytes: number of bytes to complete @rq->next_rq
2419 * Ends I/O on a number of bytes attached to @rq and @rq->next_rq.
2422 * 0 - we are done with this request
2423 * 1 - still buffers pending for this request
2425 int blk_end_bidi_request(struct request *rq, int error, int nr_bytes,
2428 return blk_end_io(rq, error, nr_bytes, bidi_bytes, NULL);
2430 EXPORT_SYMBOL_GPL(blk_end_bidi_request);
2433 * blk_end_request_callback - Special helper function for tricky drivers
2434 * @rq: the request being processed
2435 * @error: 0 for success, < 0 for error
2436 * @nr_bytes: number of bytes to complete
2437 * @drv_callback: function called between completion of bios in the request
2438 * and completion of the request.
2439 * If the callback returns non 0, this helper returns without
2440 * completion of the request.
2443 * Ends I/O on a number of bytes attached to @rq.
2444 * If @rq has leftover, sets it up for the next range of segments.
2446 * This special helper function is used only for existing tricky drivers.
2447 * (e.g. cdrom_newpc_intr() of ide-cd)
2448 * This interface will be removed when such drivers are rewritten.
2449 * Don't use this interface in other places anymore.
2452 * 0 - we are done with this request
2453 * 1 - this request is not freed yet.
2454 * this request still has pending buffers or
2455 * the driver doesn't want to finish this request yet.
2457 int blk_end_request_callback(struct request *rq, int error, int nr_bytes,
2458 int (drv_callback)(struct request *))
2460 return blk_end_io(rq, error, nr_bytes, 0, drv_callback);
2462 EXPORT_SYMBOL_GPL(blk_end_request_callback);
2464 void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
2467 /* first two bits are identical in rq->cmd_flags and bio->bi_rw */
2468 rq->cmd_flags |= (bio->bi_rw & 3);
2470 rq->nr_phys_segments = bio_phys_segments(q, bio);
2471 rq->nr_hw_segments = bio_hw_segments(q, bio);
2472 rq->current_nr_sectors = bio_cur_sectors(bio);
2473 rq->hard_cur_sectors = rq->current_nr_sectors;
2474 rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
2475 rq->buffer = bio_data(bio);
2476 rq->data_len = bio->bi_size;
2478 rq->bio = rq->biotail = bio;
2481 rq->rq_disk = bio->bi_bdev->bd_disk;
2484 int kblockd_schedule_work(struct work_struct *work)
2486 return queue_work(kblockd_workqueue, work);
2489 EXPORT_SYMBOL(kblockd_schedule_work);
2491 void kblockd_flush_work(struct work_struct *work)
2493 cancel_work_sync(work);
2495 EXPORT_SYMBOL(kblockd_flush_work);
2497 int __init blk_dev_init(void)
2501 kblockd_workqueue = create_workqueue("kblockd");
2502 if (!kblockd_workqueue)
2503 panic("Failed to create kblockd\n");
2505 request_cachep = kmem_cache_create("blkdev_requests",
2506 sizeof(struct request), 0, SLAB_PANIC, NULL);
2508 blk_requestq_cachep = kmem_cache_create("blkdev_queue",
2509 sizeof(struct request_queue), 0, SLAB_PANIC, NULL);
2511 for_each_possible_cpu(i)
2512 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
2514 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
2515 register_hotcpu_notifier(&blk_cpu_notifier);