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/bootmem.h> /* for max_pfn/max_low_pfn */
24 #include <linux/completion.h>
25 #include <linux/slab.h>
26 #include <linux/swap.h>
27 #include <linux/writeback.h>
28 #include <linux/task_io_accounting_ops.h>
29 #include <linux/interrupt.h>
30 #include <linux/cpu.h>
31 #include <linux/blktrace_api.h>
32 #include <linux/fault-inject.h>
33 #include <linux/scatterlist.h>
38 #include <scsi/scsi_cmnd.h>
40 static void blk_unplug_work(struct work_struct *work);
41 static void blk_unplug_timeout(unsigned long data);
42 static void drive_stat_acct(struct request *rq, int new_io);
43 static void init_request_from_bio(struct request *req, struct bio *bio);
44 static int __make_request(struct request_queue *q, struct bio *bio);
45 static struct io_context *current_io_context(gfp_t gfp_flags, int node);
46 static void blk_recalc_rq_segments(struct request *rq);
47 static void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
51 * For the allocated request tables
53 static struct kmem_cache *request_cachep;
56 * For queue allocation
58 static struct kmem_cache *requestq_cachep;
61 * For io context allocations
63 static struct kmem_cache *iocontext_cachep;
66 * Controlling structure to kblockd
68 static struct workqueue_struct *kblockd_workqueue;
70 unsigned long blk_max_low_pfn, blk_max_pfn;
72 EXPORT_SYMBOL(blk_max_low_pfn);
73 EXPORT_SYMBOL(blk_max_pfn);
75 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
77 /* Amount of time in which a process may batch requests */
78 #define BLK_BATCH_TIME (HZ/50UL)
80 /* Number of requests a "batching" process may submit */
81 #define BLK_BATCH_REQ 32
84 * Return the threshold (number of used requests) at which the queue is
85 * considered to be congested. It include a little hysteresis to keep the
86 * context switch rate down.
88 static inline int queue_congestion_on_threshold(struct request_queue *q)
90 return q->nr_congestion_on;
94 * The threshold at which a queue is considered to be uncongested
96 static inline int queue_congestion_off_threshold(struct request_queue *q)
98 return q->nr_congestion_off;
101 static void blk_queue_congestion_threshold(struct request_queue *q)
105 nr = q->nr_requests - (q->nr_requests / 8) + 1;
106 if (nr > q->nr_requests)
108 q->nr_congestion_on = nr;
110 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
113 q->nr_congestion_off = nr;
117 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
120 * Locates the passed device's request queue and returns the address of its
123 * Will return NULL if the request queue cannot be located.
125 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
127 struct backing_dev_info *ret = NULL;
128 struct request_queue *q = bdev_get_queue(bdev);
131 ret = &q->backing_dev_info;
134 EXPORT_SYMBOL(blk_get_backing_dev_info);
137 * blk_queue_prep_rq - set a prepare_request function for queue
139 * @pfn: prepare_request function
141 * It's possible for a queue to register a prepare_request callback which
142 * is invoked before the request is handed to the request_fn. The goal of
143 * the function is to prepare a request for I/O, it can be used to build a
144 * cdb from the request data for instance.
147 void blk_queue_prep_rq(struct request_queue *q, prep_rq_fn *pfn)
152 EXPORT_SYMBOL(blk_queue_prep_rq);
155 * blk_queue_merge_bvec - set a merge_bvec function for queue
157 * @mbfn: merge_bvec_fn
159 * Usually queues have static limitations on the max sectors or segments that
160 * we can put in a request. Stacking drivers may have some settings that
161 * are dynamic, and thus we have to query the queue whether it is ok to
162 * add a new bio_vec to a bio at a given offset or not. If the block device
163 * has such limitations, it needs to register a merge_bvec_fn to control
164 * the size of bio's sent to it. Note that a block device *must* allow a
165 * single page to be added to an empty bio. The block device driver may want
166 * to use the bio_split() function to deal with these bio's. By default
167 * no merge_bvec_fn is defined for a queue, and only the fixed limits are
170 void blk_queue_merge_bvec(struct request_queue *q, merge_bvec_fn *mbfn)
172 q->merge_bvec_fn = mbfn;
175 EXPORT_SYMBOL(blk_queue_merge_bvec);
177 void blk_queue_softirq_done(struct request_queue *q, softirq_done_fn *fn)
179 q->softirq_done_fn = fn;
182 EXPORT_SYMBOL(blk_queue_softirq_done);
185 * blk_queue_make_request - define an alternate make_request function for a device
186 * @q: the request queue for the device to be affected
187 * @mfn: the alternate make_request function
190 * The normal way for &struct bios to be passed to a device
191 * driver is for them to be collected into requests on a request
192 * queue, and then to allow the device driver to select requests
193 * off that queue when it is ready. This works well for many block
194 * devices. However some block devices (typically virtual devices
195 * such as md or lvm) do not benefit from the processing on the
196 * request queue, and are served best by having the requests passed
197 * directly to them. This can be achieved by providing a function
198 * to blk_queue_make_request().
201 * The driver that does this *must* be able to deal appropriately
202 * with buffers in "highmemory". This can be accomplished by either calling
203 * __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
204 * blk_queue_bounce() to create a buffer in normal memory.
206 void blk_queue_make_request(struct request_queue * q, make_request_fn * mfn)
211 q->nr_requests = BLKDEV_MAX_RQ;
212 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
213 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
214 q->make_request_fn = mfn;
215 q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
216 q->backing_dev_info.state = 0;
217 q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
218 blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
219 blk_queue_hardsect_size(q, 512);
220 blk_queue_dma_alignment(q, 511);
221 blk_queue_congestion_threshold(q);
222 q->nr_batching = BLK_BATCH_REQ;
224 q->unplug_thresh = 4; /* hmm */
225 q->unplug_delay = (3 * HZ) / 1000; /* 3 milliseconds */
226 if (q->unplug_delay == 0)
229 INIT_WORK(&q->unplug_work, blk_unplug_work);
231 q->unplug_timer.function = blk_unplug_timeout;
232 q->unplug_timer.data = (unsigned long)q;
235 * by default assume old behaviour and bounce for any highmem page
237 blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
240 EXPORT_SYMBOL(blk_queue_make_request);
242 static void rq_init(struct request_queue *q, struct request *rq)
244 INIT_LIST_HEAD(&rq->queuelist);
245 INIT_LIST_HEAD(&rq->donelist);
248 rq->bio = rq->biotail = NULL;
249 INIT_HLIST_NODE(&rq->hash);
250 RB_CLEAR_NODE(&rq->rb_node);
258 rq->nr_phys_segments = 0;
261 rq->end_io_data = NULL;
262 rq->completion_data = NULL;
267 * blk_queue_ordered - does this queue support ordered writes
268 * @q: the request queue
269 * @ordered: one of QUEUE_ORDERED_*
270 * @prepare_flush_fn: rq setup helper for cache flush ordered writes
273 * For journalled file systems, doing ordered writes on a commit
274 * block instead of explicitly doing wait_on_buffer (which is bad
275 * for performance) can be a big win. Block drivers supporting this
276 * feature should call this function and indicate so.
279 int blk_queue_ordered(struct request_queue *q, unsigned ordered,
280 prepare_flush_fn *prepare_flush_fn)
282 if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
283 prepare_flush_fn == NULL) {
284 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
288 if (ordered != QUEUE_ORDERED_NONE &&
289 ordered != QUEUE_ORDERED_DRAIN &&
290 ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
291 ordered != QUEUE_ORDERED_DRAIN_FUA &&
292 ordered != QUEUE_ORDERED_TAG &&
293 ordered != QUEUE_ORDERED_TAG_FLUSH &&
294 ordered != QUEUE_ORDERED_TAG_FUA) {
295 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
299 q->ordered = ordered;
300 q->next_ordered = ordered;
301 q->prepare_flush_fn = prepare_flush_fn;
306 EXPORT_SYMBOL(blk_queue_ordered);
309 * Cache flushing for ordered writes handling
311 inline unsigned blk_ordered_cur_seq(struct request_queue *q)
315 return 1 << ffz(q->ordseq);
318 unsigned blk_ordered_req_seq(struct request *rq)
320 struct request_queue *q = rq->q;
322 BUG_ON(q->ordseq == 0);
324 if (rq == &q->pre_flush_rq)
325 return QUEUE_ORDSEQ_PREFLUSH;
326 if (rq == &q->bar_rq)
327 return QUEUE_ORDSEQ_BAR;
328 if (rq == &q->post_flush_rq)
329 return QUEUE_ORDSEQ_POSTFLUSH;
332 * !fs requests don't need to follow barrier ordering. Always
333 * put them at the front. This fixes the following deadlock.
335 * http://thread.gmane.org/gmane.linux.kernel/537473
337 if (!blk_fs_request(rq))
338 return QUEUE_ORDSEQ_DRAIN;
340 if ((rq->cmd_flags & REQ_ORDERED_COLOR) ==
341 (q->orig_bar_rq->cmd_flags & REQ_ORDERED_COLOR))
342 return QUEUE_ORDSEQ_DRAIN;
344 return QUEUE_ORDSEQ_DONE;
347 void blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
352 if (error && !q->orderr)
355 BUG_ON(q->ordseq & seq);
358 if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
362 * Okay, sequence complete.
366 uptodate = q->orderr;
371 end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
372 end_that_request_last(rq, uptodate);
375 static void pre_flush_end_io(struct request *rq, int error)
377 elv_completed_request(rq->q, rq);
378 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
381 static void bar_end_io(struct request *rq, int error)
383 elv_completed_request(rq->q, rq);
384 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
387 static void post_flush_end_io(struct request *rq, int error)
389 elv_completed_request(rq->q, rq);
390 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
393 static void queue_flush(struct request_queue *q, unsigned which)
396 rq_end_io_fn *end_io;
398 if (which == QUEUE_ORDERED_PREFLUSH) {
399 rq = &q->pre_flush_rq;
400 end_io = pre_flush_end_io;
402 rq = &q->post_flush_rq;
403 end_io = post_flush_end_io;
406 rq->cmd_flags = REQ_HARDBARRIER;
408 rq->elevator_private = NULL;
409 rq->elevator_private2 = NULL;
410 rq->rq_disk = q->bar_rq.rq_disk;
412 q->prepare_flush_fn(q, rq);
414 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
417 static inline struct request *start_ordered(struct request_queue *q,
421 q->ordered = q->next_ordered;
422 q->ordseq |= QUEUE_ORDSEQ_STARTED;
425 * Prep proxy barrier request.
427 blkdev_dequeue_request(rq);
432 if (bio_data_dir(q->orig_bar_rq->bio) == WRITE)
433 rq->cmd_flags |= REQ_RW;
434 if (q->ordered & QUEUE_ORDERED_FUA)
435 rq->cmd_flags |= REQ_FUA;
436 rq->elevator_private = NULL;
437 rq->elevator_private2 = NULL;
438 init_request_from_bio(rq, q->orig_bar_rq->bio);
439 rq->end_io = bar_end_io;
442 * Queue ordered sequence. As we stack them at the head, we
443 * need to queue in reverse order. Note that we rely on that
444 * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
445 * request gets inbetween ordered sequence. If this request is
446 * an empty barrier, we don't need to do a postflush ever since
447 * there will be no data written between the pre and post flush.
448 * Hence a single flush will suffice.
450 if ((q->ordered & QUEUE_ORDERED_POSTFLUSH) && !blk_empty_barrier(rq))
451 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
453 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
455 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
457 if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
458 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
459 rq = &q->pre_flush_rq;
461 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
463 if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
464 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
471 int blk_do_ordered(struct request_queue *q, struct request **rqp)
473 struct request *rq = *rqp;
474 const int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
480 if (q->next_ordered != QUEUE_ORDERED_NONE) {
481 *rqp = start_ordered(q, rq);
485 * This can happen when the queue switches to
486 * ORDERED_NONE while this request is on it.
488 blkdev_dequeue_request(rq);
489 end_that_request_first(rq, -EOPNOTSUPP,
490 rq->hard_nr_sectors);
491 end_that_request_last(rq, -EOPNOTSUPP);
498 * Ordered sequence in progress
501 /* Special requests are not subject to ordering rules. */
502 if (!blk_fs_request(rq) &&
503 rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
506 if (q->ordered & QUEUE_ORDERED_TAG) {
507 /* Ordered by tag. Blocking the next barrier is enough. */
508 if (is_barrier && rq != &q->bar_rq)
511 /* Ordered by draining. Wait for turn. */
512 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
513 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
520 static void req_bio_endio(struct request *rq, struct bio *bio,
521 unsigned int nbytes, int error)
523 struct request_queue *q = rq->q;
525 if (&q->bar_rq != rq) {
527 clear_bit(BIO_UPTODATE, &bio->bi_flags);
528 else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
531 if (unlikely(nbytes > bio->bi_size)) {
532 printk("%s: want %u bytes done, only %u left\n",
533 __FUNCTION__, nbytes, bio->bi_size);
534 nbytes = bio->bi_size;
537 bio->bi_size -= nbytes;
538 bio->bi_sector += (nbytes >> 9);
539 if (bio->bi_size == 0)
540 bio_endio(bio, error);
544 * Okay, this is the barrier request in progress, just
547 if (error && !q->orderr)
553 * blk_queue_bounce_limit - set bounce buffer limit for queue
554 * @q: the request queue for the device
555 * @dma_addr: bus address limit
558 * Different hardware can have different requirements as to what pages
559 * it can do I/O directly to. A low level driver can call
560 * blk_queue_bounce_limit to have lower memory pages allocated as bounce
561 * buffers for doing I/O to pages residing above @page.
563 void blk_queue_bounce_limit(struct request_queue *q, u64 dma_addr)
565 unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
568 q->bounce_gfp = GFP_NOIO;
569 #if BITS_PER_LONG == 64
570 /* Assume anything <= 4GB can be handled by IOMMU.
571 Actually some IOMMUs can handle everything, but I don't
572 know of a way to test this here. */
573 if (bounce_pfn < (min_t(u64,0xffffffff,BLK_BOUNCE_HIGH) >> PAGE_SHIFT))
575 q->bounce_pfn = max_low_pfn;
577 if (bounce_pfn < blk_max_low_pfn)
579 q->bounce_pfn = bounce_pfn;
582 init_emergency_isa_pool();
583 q->bounce_gfp = GFP_NOIO | GFP_DMA;
584 q->bounce_pfn = bounce_pfn;
588 EXPORT_SYMBOL(blk_queue_bounce_limit);
591 * blk_queue_max_sectors - set max sectors for a request for this queue
592 * @q: the request queue for the device
593 * @max_sectors: max sectors in the usual 512b unit
596 * Enables a low level driver to set an upper limit on the size of
599 void blk_queue_max_sectors(struct request_queue *q, unsigned int max_sectors)
601 if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
602 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
603 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
606 if (BLK_DEF_MAX_SECTORS > max_sectors)
607 q->max_hw_sectors = q->max_sectors = max_sectors;
609 q->max_sectors = BLK_DEF_MAX_SECTORS;
610 q->max_hw_sectors = max_sectors;
614 EXPORT_SYMBOL(blk_queue_max_sectors);
617 * blk_queue_max_phys_segments - set max phys segments for a request for this queue
618 * @q: the request queue for the device
619 * @max_segments: max number of segments
622 * Enables a low level driver to set an upper limit on the number of
623 * physical data segments in a request. This would be the largest sized
624 * scatter list the driver could handle.
626 void blk_queue_max_phys_segments(struct request_queue *q,
627 unsigned short max_segments)
631 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
634 q->max_phys_segments = max_segments;
637 EXPORT_SYMBOL(blk_queue_max_phys_segments);
640 * blk_queue_max_hw_segments - set max hw segments for a request for this queue
641 * @q: the request queue for the device
642 * @max_segments: max number of segments
645 * Enables a low level driver to set an upper limit on the number of
646 * hw data segments in a request. This would be the largest number of
647 * address/length pairs the host adapter can actually give as once
650 void blk_queue_max_hw_segments(struct request_queue *q,
651 unsigned short max_segments)
655 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
658 q->max_hw_segments = max_segments;
661 EXPORT_SYMBOL(blk_queue_max_hw_segments);
664 * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
665 * @q: the request queue for the device
666 * @max_size: max size of segment in bytes
669 * Enables a low level driver to set an upper limit on the size of a
672 void blk_queue_max_segment_size(struct request_queue *q, unsigned int max_size)
674 if (max_size < PAGE_CACHE_SIZE) {
675 max_size = PAGE_CACHE_SIZE;
676 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
679 q->max_segment_size = max_size;
682 EXPORT_SYMBOL(blk_queue_max_segment_size);
685 * blk_queue_hardsect_size - set hardware sector size for the queue
686 * @q: the request queue for the device
687 * @size: the hardware sector size, in bytes
690 * This should typically be set to the lowest possible sector size
691 * that the hardware can operate on (possible without reverting to
692 * even internal read-modify-write operations). Usually the default
693 * of 512 covers most hardware.
695 void blk_queue_hardsect_size(struct request_queue *q, unsigned short size)
697 q->hardsect_size = size;
700 EXPORT_SYMBOL(blk_queue_hardsect_size);
703 * Returns the minimum that is _not_ zero, unless both are zero.
705 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
708 * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
709 * @t: the stacking driver (top)
710 * @b: the underlying device (bottom)
712 void blk_queue_stack_limits(struct request_queue *t, struct request_queue *b)
714 /* zero is "infinity" */
715 t->max_sectors = min_not_zero(t->max_sectors,b->max_sectors);
716 t->max_hw_sectors = min_not_zero(t->max_hw_sectors,b->max_hw_sectors);
718 t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
719 t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
720 t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
721 t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
722 if (!test_bit(QUEUE_FLAG_CLUSTER, &b->queue_flags))
723 clear_bit(QUEUE_FLAG_CLUSTER, &t->queue_flags);
726 EXPORT_SYMBOL(blk_queue_stack_limits);
729 * blk_queue_segment_boundary - set boundary rules for segment merging
730 * @q: the request queue for the device
731 * @mask: the memory boundary mask
733 void blk_queue_segment_boundary(struct request_queue *q, unsigned long mask)
735 if (mask < PAGE_CACHE_SIZE - 1) {
736 mask = PAGE_CACHE_SIZE - 1;
737 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
740 q->seg_boundary_mask = mask;
743 EXPORT_SYMBOL(blk_queue_segment_boundary);
746 * blk_queue_dma_alignment - set dma length and memory alignment
747 * @q: the request queue for the device
748 * @mask: alignment mask
751 * set required memory and length aligment for direct dma transactions.
752 * this is used when buiding direct io requests for the queue.
755 void blk_queue_dma_alignment(struct request_queue *q, int mask)
757 q->dma_alignment = mask;
760 EXPORT_SYMBOL(blk_queue_dma_alignment);
763 * blk_queue_find_tag - find a request by its tag and queue
764 * @q: The request queue for the device
765 * @tag: The tag of the request
768 * Should be used when a device returns a tag and you want to match
771 * no locks need be held.
773 struct request *blk_queue_find_tag(struct request_queue *q, int tag)
775 return blk_map_queue_find_tag(q->queue_tags, tag);
778 EXPORT_SYMBOL(blk_queue_find_tag);
781 * __blk_free_tags - release a given set of tag maintenance info
782 * @bqt: the tag map to free
784 * Tries to free the specified @bqt@. Returns true if it was
785 * actually freed and false if there are still references using it
787 static int __blk_free_tags(struct blk_queue_tag *bqt)
791 retval = atomic_dec_and_test(&bqt->refcnt);
794 BUG_ON(!list_empty(&bqt->busy_list));
796 kfree(bqt->tag_index);
797 bqt->tag_index = NULL;
810 * __blk_queue_free_tags - release tag maintenance info
811 * @q: the request queue for the device
814 * blk_cleanup_queue() will take care of calling this function, if tagging
815 * has been used. So there's no need to call this directly.
817 static void __blk_queue_free_tags(struct request_queue *q)
819 struct blk_queue_tag *bqt = q->queue_tags;
824 __blk_free_tags(bqt);
826 q->queue_tags = NULL;
827 q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
832 * blk_free_tags - release a given set of tag maintenance info
833 * @bqt: the tag map to free
835 * For externally managed @bqt@ frees the map. Callers of this
836 * function must guarantee to have released all the queues that
837 * might have been using this tag map.
839 void blk_free_tags(struct blk_queue_tag *bqt)
841 if (unlikely(!__blk_free_tags(bqt)))
844 EXPORT_SYMBOL(blk_free_tags);
847 * blk_queue_free_tags - release tag maintenance info
848 * @q: the request queue for the device
851 * This is used to disabled tagged queuing to a device, yet leave
854 void blk_queue_free_tags(struct request_queue *q)
856 clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
859 EXPORT_SYMBOL(blk_queue_free_tags);
862 init_tag_map(struct request_queue *q, struct blk_queue_tag *tags, int depth)
864 struct request **tag_index;
865 unsigned long *tag_map;
868 if (q && depth > q->nr_requests * 2) {
869 depth = q->nr_requests * 2;
870 printk(KERN_ERR "%s: adjusted depth to %d\n",
871 __FUNCTION__, depth);
874 tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC);
878 nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
879 tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
883 tags->real_max_depth = depth;
884 tags->max_depth = depth;
885 tags->tag_index = tag_index;
886 tags->tag_map = tag_map;
894 static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q,
897 struct blk_queue_tag *tags;
899 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
903 if (init_tag_map(q, tags, depth))
906 INIT_LIST_HEAD(&tags->busy_list);
908 atomic_set(&tags->refcnt, 1);
916 * blk_init_tags - initialize the tag info for an external tag map
917 * @depth: the maximum queue depth supported
918 * @tags: the tag to use
920 struct blk_queue_tag *blk_init_tags(int depth)
922 return __blk_queue_init_tags(NULL, depth);
924 EXPORT_SYMBOL(blk_init_tags);
927 * blk_queue_init_tags - initialize the queue tag info
928 * @q: the request queue for the device
929 * @depth: the maximum queue depth supported
930 * @tags: the tag to use
932 int blk_queue_init_tags(struct request_queue *q, int depth,
933 struct blk_queue_tag *tags)
937 BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
939 if (!tags && !q->queue_tags) {
940 tags = __blk_queue_init_tags(q, depth);
944 } else if (q->queue_tags) {
945 if ((rc = blk_queue_resize_tags(q, depth)))
947 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
950 atomic_inc(&tags->refcnt);
953 * assign it, all done
955 q->queue_tags = tags;
956 q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
963 EXPORT_SYMBOL(blk_queue_init_tags);
966 * blk_queue_resize_tags - change the queueing depth
967 * @q: the request queue for the device
968 * @new_depth: the new max command queueing depth
971 * Must be called with the queue lock held.
973 int blk_queue_resize_tags(struct request_queue *q, int new_depth)
975 struct blk_queue_tag *bqt = q->queue_tags;
976 struct request **tag_index;
977 unsigned long *tag_map;
978 int max_depth, nr_ulongs;
984 * if we already have large enough real_max_depth. just
985 * adjust max_depth. *NOTE* as requests with tag value
986 * between new_depth and real_max_depth can be in-flight, tag
987 * map can not be shrunk blindly here.
989 if (new_depth <= bqt->real_max_depth) {
990 bqt->max_depth = new_depth;
995 * Currently cannot replace a shared tag map with a new
996 * one, so error out if this is the case
998 if (atomic_read(&bqt->refcnt) != 1)
1002 * save the old state info, so we can copy it back
1004 tag_index = bqt->tag_index;
1005 tag_map = bqt->tag_map;
1006 max_depth = bqt->real_max_depth;
1008 if (init_tag_map(q, bqt, new_depth))
1011 memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
1012 nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
1013 memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
1020 EXPORT_SYMBOL(blk_queue_resize_tags);
1023 * blk_queue_end_tag - end tag operations for a request
1024 * @q: the request queue for the device
1025 * @rq: the request that has completed
1028 * Typically called when end_that_request_first() returns 0, meaning
1029 * all transfers have been done for a request. It's important to call
1030 * this function before end_that_request_last(), as that will put the
1031 * request back on the free list thus corrupting the internal tag list.
1034 * queue lock must be held.
1036 void blk_queue_end_tag(struct request_queue *q, struct request *rq)
1038 struct blk_queue_tag *bqt = q->queue_tags;
1043 if (unlikely(tag >= bqt->real_max_depth))
1045 * This can happen after tag depth has been reduced.
1046 * FIXME: how about a warning or info message here?
1050 list_del_init(&rq->queuelist);
1051 rq->cmd_flags &= ~REQ_QUEUED;
1054 if (unlikely(bqt->tag_index[tag] == NULL))
1055 printk(KERN_ERR "%s: tag %d is missing\n",
1058 bqt->tag_index[tag] = NULL;
1061 * We use test_and_clear_bit's memory ordering properties here.
1062 * The tag_map bit acts as a lock for tag_index[bit], so we need
1063 * a barrer before clearing the bit (precisely: release semantics).
1064 * Could use clear_bit_unlock when it is merged.
1066 if (unlikely(!test_and_clear_bit(tag, bqt->tag_map))) {
1067 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
1075 EXPORT_SYMBOL(blk_queue_end_tag);
1078 * blk_queue_start_tag - find a free tag and assign it
1079 * @q: the request queue for the device
1080 * @rq: the block request that needs tagging
1083 * This can either be used as a stand-alone helper, or possibly be
1084 * assigned as the queue &prep_rq_fn (in which case &struct request
1085 * automagically gets a tag assigned). Note that this function
1086 * assumes that any type of request can be queued! if this is not
1087 * true for your device, you must check the request type before
1088 * calling this function. The request will also be removed from
1089 * the request queue, so it's the drivers responsibility to readd
1090 * it if it should need to be restarted for some reason.
1093 * queue lock must be held.
1095 int blk_queue_start_tag(struct request_queue *q, struct request *rq)
1097 struct blk_queue_tag *bqt = q->queue_tags;
1100 if (unlikely((rq->cmd_flags & REQ_QUEUED))) {
1102 "%s: request %p for device [%s] already tagged %d",
1104 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
1109 * Protect against shared tag maps, as we may not have exclusive
1110 * access to the tag map.
1113 tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
1114 if (tag >= bqt->max_depth)
1117 } while (test_and_set_bit(tag, bqt->tag_map));
1119 * We rely on test_and_set_bit providing lock memory ordering semantics
1120 * (could use test_and_set_bit_lock when it is merged).
1123 rq->cmd_flags |= REQ_QUEUED;
1125 bqt->tag_index[tag] = rq;
1126 blkdev_dequeue_request(rq);
1127 list_add(&rq->queuelist, &bqt->busy_list);
1132 EXPORT_SYMBOL(blk_queue_start_tag);
1135 * blk_queue_invalidate_tags - invalidate all pending tags
1136 * @q: the request queue for the device
1139 * Hardware conditions may dictate a need to stop all pending requests.
1140 * In this case, we will safely clear the block side of the tag queue and
1141 * readd all requests to the request queue in the right order.
1144 * queue lock must be held.
1146 void blk_queue_invalidate_tags(struct request_queue *q)
1148 struct blk_queue_tag *bqt = q->queue_tags;
1149 struct list_head *tmp, *n;
1152 list_for_each_safe(tmp, n, &bqt->busy_list) {
1153 rq = list_entry_rq(tmp);
1155 if (rq->tag == -1) {
1157 "%s: bad tag found on list\n", __FUNCTION__);
1158 list_del_init(&rq->queuelist);
1159 rq->cmd_flags &= ~REQ_QUEUED;
1161 blk_queue_end_tag(q, rq);
1163 rq->cmd_flags &= ~REQ_STARTED;
1164 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1168 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1170 void blk_dump_rq_flags(struct request *rq, char *msg)
1174 printk("%s: dev %s: type=%x, flags=%x\n", msg,
1175 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->cmd_type,
1178 printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1180 rq->current_nr_sectors);
1181 printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1183 if (blk_pc_request(rq)) {
1185 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1186 printk("%02x ", rq->cmd[bit]);
1191 EXPORT_SYMBOL(blk_dump_rq_flags);
1193 void blk_recount_segments(struct request_queue *q, struct bio *bio)
1196 struct bio *nxt = bio->bi_next;
1198 rq.bio = rq.biotail = bio;
1199 bio->bi_next = NULL;
1200 blk_recalc_rq_segments(&rq);
1202 bio->bi_phys_segments = rq.nr_phys_segments;
1203 bio->bi_hw_segments = rq.nr_hw_segments;
1204 bio->bi_flags |= (1 << BIO_SEG_VALID);
1206 EXPORT_SYMBOL(blk_recount_segments);
1208 static void blk_recalc_rq_segments(struct request *rq)
1212 unsigned int phys_size;
1213 unsigned int hw_size;
1214 struct bio_vec *bv, *bvprv = NULL;
1218 struct req_iterator iter;
1219 int high, highprv = 1;
1220 struct request_queue *q = rq->q;
1225 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1226 hw_seg_size = seg_size = 0;
1227 phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
1228 rq_for_each_segment(bv, rq, iter) {
1230 * the trick here is making sure that a high page is never
1231 * considered part of another segment, since that might
1232 * change with the bounce page.
1234 high = page_to_pfn(bv->bv_page) > q->bounce_pfn;
1235 if (high || highprv)
1236 goto new_hw_segment;
1238 if (seg_size + bv->bv_len > q->max_segment_size)
1240 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1242 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1244 if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1245 goto new_hw_segment;
1247 seg_size += bv->bv_len;
1248 hw_seg_size += bv->bv_len;
1253 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1254 !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1255 hw_seg_size += bv->bv_len;
1258 if (nr_hw_segs == 1 &&
1259 hw_seg_size > rq->bio->bi_hw_front_size)
1260 rq->bio->bi_hw_front_size = hw_seg_size;
1261 hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1267 seg_size = bv->bv_len;
1271 if (nr_hw_segs == 1 &&
1272 hw_seg_size > rq->bio->bi_hw_front_size)
1273 rq->bio->bi_hw_front_size = hw_seg_size;
1274 if (hw_seg_size > rq->biotail->bi_hw_back_size)
1275 rq->biotail->bi_hw_back_size = hw_seg_size;
1276 rq->nr_phys_segments = nr_phys_segs;
1277 rq->nr_hw_segments = nr_hw_segs;
1280 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
1283 if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1286 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1288 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1292 * bio and nxt are contigous in memory, check if the queue allows
1293 * these two to be merged into one
1295 if (BIO_SEG_BOUNDARY(q, bio, nxt))
1301 static int blk_hw_contig_segment(struct request_queue *q, struct bio *bio,
1304 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1305 blk_recount_segments(q, bio);
1306 if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1307 blk_recount_segments(q, nxt);
1308 if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1309 BIOVEC_VIRT_OVERSIZE(bio->bi_hw_back_size + nxt->bi_hw_front_size))
1311 if (bio->bi_hw_back_size + nxt->bi_hw_front_size > q->max_segment_size)
1318 * map a request to scatterlist, return number of sg entries setup. Caller
1319 * must make sure sg can hold rq->nr_phys_segments entries
1321 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
1322 struct scatterlist *sglist)
1324 struct bio_vec *bvec, *bvprv;
1325 struct req_iterator iter;
1326 struct scatterlist *sg;
1330 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1333 * for each bio in rq
1337 rq_for_each_segment(bvec, rq, iter) {
1338 int nbytes = bvec->bv_len;
1340 if (bvprv && cluster) {
1341 if (sg->length + nbytes > q->max_segment_size)
1344 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1346 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1349 sg->length += nbytes;
1356 * If the driver previously mapped a shorter
1357 * list, we could see a termination bit
1358 * prematurely unless it fully inits the sg
1359 * table on each mapping. We KNOW that there
1360 * must be more entries here or the driver
1361 * would be buggy, so force clear the
1362 * termination bit to avoid doing a full
1363 * sg_init_table() in drivers for each command.
1365 sg->page_link &= ~0x02;
1369 sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
1373 } /* segments in rq */
1381 EXPORT_SYMBOL(blk_rq_map_sg);
1384 * the standard queue merge functions, can be overridden with device
1385 * specific ones if so desired
1388 static inline int ll_new_mergeable(struct request_queue *q,
1389 struct request *req,
1392 int nr_phys_segs = bio_phys_segments(q, bio);
1394 if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1395 req->cmd_flags |= REQ_NOMERGE;
1396 if (req == q->last_merge)
1397 q->last_merge = NULL;
1402 * A hw segment is just getting larger, bump just the phys
1405 req->nr_phys_segments += nr_phys_segs;
1409 static inline int ll_new_hw_segment(struct request_queue *q,
1410 struct request *req,
1413 int nr_hw_segs = bio_hw_segments(q, bio);
1414 int nr_phys_segs = bio_phys_segments(q, bio);
1416 if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1417 || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1418 req->cmd_flags |= REQ_NOMERGE;
1419 if (req == q->last_merge)
1420 q->last_merge = NULL;
1425 * This will form the start of a new hw segment. Bump both
1428 req->nr_hw_segments += nr_hw_segs;
1429 req->nr_phys_segments += nr_phys_segs;
1433 static int ll_back_merge_fn(struct request_queue *q, struct request *req,
1436 unsigned short max_sectors;
1439 if (unlikely(blk_pc_request(req)))
1440 max_sectors = q->max_hw_sectors;
1442 max_sectors = q->max_sectors;
1444 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1445 req->cmd_flags |= REQ_NOMERGE;
1446 if (req == q->last_merge)
1447 q->last_merge = NULL;
1450 if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1451 blk_recount_segments(q, req->biotail);
1452 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1453 blk_recount_segments(q, bio);
1454 len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1455 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1456 !BIOVEC_VIRT_OVERSIZE(len)) {
1457 int mergeable = ll_new_mergeable(q, req, bio);
1460 if (req->nr_hw_segments == 1)
1461 req->bio->bi_hw_front_size = len;
1462 if (bio->bi_hw_segments == 1)
1463 bio->bi_hw_back_size = len;
1468 return ll_new_hw_segment(q, req, bio);
1471 static int ll_front_merge_fn(struct request_queue *q, struct request *req,
1474 unsigned short max_sectors;
1477 if (unlikely(blk_pc_request(req)))
1478 max_sectors = q->max_hw_sectors;
1480 max_sectors = q->max_sectors;
1483 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1484 req->cmd_flags |= REQ_NOMERGE;
1485 if (req == q->last_merge)
1486 q->last_merge = NULL;
1489 len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1490 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1491 blk_recount_segments(q, bio);
1492 if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1493 blk_recount_segments(q, req->bio);
1494 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1495 !BIOVEC_VIRT_OVERSIZE(len)) {
1496 int mergeable = ll_new_mergeable(q, req, bio);
1499 if (bio->bi_hw_segments == 1)
1500 bio->bi_hw_front_size = len;
1501 if (req->nr_hw_segments == 1)
1502 req->biotail->bi_hw_back_size = len;
1507 return ll_new_hw_segment(q, req, bio);
1510 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
1511 struct request *next)
1513 int total_phys_segments;
1514 int total_hw_segments;
1517 * First check if the either of the requests are re-queued
1518 * requests. Can't merge them if they are.
1520 if (req->special || next->special)
1524 * Will it become too large?
1526 if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1529 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1530 if (blk_phys_contig_segment(q, req->biotail, next->bio))
1531 total_phys_segments--;
1533 if (total_phys_segments > q->max_phys_segments)
1536 total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1537 if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1538 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1540 * propagate the combined length to the end of the requests
1542 if (req->nr_hw_segments == 1)
1543 req->bio->bi_hw_front_size = len;
1544 if (next->nr_hw_segments == 1)
1545 next->biotail->bi_hw_back_size = len;
1546 total_hw_segments--;
1549 if (total_hw_segments > q->max_hw_segments)
1552 /* Merge is OK... */
1553 req->nr_phys_segments = total_phys_segments;
1554 req->nr_hw_segments = total_hw_segments;
1559 * "plug" the device if there are no outstanding requests: this will
1560 * force the transfer to start only after we have put all the requests
1563 * This is called with interrupts off and no requests on the queue and
1564 * with the queue lock held.
1566 void blk_plug_device(struct request_queue *q)
1568 WARN_ON(!irqs_disabled());
1571 * don't plug a stopped queue, it must be paired with blk_start_queue()
1572 * which will restart the queueing
1574 if (blk_queue_stopped(q))
1577 if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
1578 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1579 blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
1583 EXPORT_SYMBOL(blk_plug_device);
1586 * remove the queue from the plugged list, if present. called with
1587 * queue lock held and interrupts disabled.
1589 int blk_remove_plug(struct request_queue *q)
1591 WARN_ON(!irqs_disabled());
1593 if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1596 del_timer(&q->unplug_timer);
1600 EXPORT_SYMBOL(blk_remove_plug);
1603 * remove the plug and let it rip..
1605 void __generic_unplug_device(struct request_queue *q)
1607 if (unlikely(blk_queue_stopped(q)))
1610 if (!blk_remove_plug(q))
1615 EXPORT_SYMBOL(__generic_unplug_device);
1618 * generic_unplug_device - fire a request queue
1619 * @q: The &struct request_queue in question
1622 * Linux uses plugging to build bigger requests queues before letting
1623 * the device have at them. If a queue is plugged, the I/O scheduler
1624 * is still adding and merging requests on the queue. Once the queue
1625 * gets unplugged, the request_fn defined for the queue is invoked and
1626 * transfers started.
1628 void generic_unplug_device(struct request_queue *q)
1630 spin_lock_irq(q->queue_lock);
1631 __generic_unplug_device(q);
1632 spin_unlock_irq(q->queue_lock);
1634 EXPORT_SYMBOL(generic_unplug_device);
1636 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1639 struct request_queue *q = bdi->unplug_io_data;
1642 * devices don't necessarily have an ->unplug_fn defined
1645 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1646 q->rq.count[READ] + q->rq.count[WRITE]);
1652 static void blk_unplug_work(struct work_struct *work)
1654 struct request_queue *q =
1655 container_of(work, struct request_queue, unplug_work);
1657 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1658 q->rq.count[READ] + q->rq.count[WRITE]);
1663 static void blk_unplug_timeout(unsigned long data)
1665 struct request_queue *q = (struct request_queue *)data;
1667 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
1668 q->rq.count[READ] + q->rq.count[WRITE]);
1670 kblockd_schedule_work(&q->unplug_work);
1674 * blk_start_queue - restart a previously stopped queue
1675 * @q: The &struct request_queue in question
1678 * blk_start_queue() will clear the stop flag on the queue, and call
1679 * the request_fn for the queue if it was in a stopped state when
1680 * entered. Also see blk_stop_queue(). Queue lock must be held.
1682 void blk_start_queue(struct request_queue *q)
1684 WARN_ON(!irqs_disabled());
1686 clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1689 * one level of recursion is ok and is much faster than kicking
1690 * the unplug handling
1692 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1694 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1697 kblockd_schedule_work(&q->unplug_work);
1701 EXPORT_SYMBOL(blk_start_queue);
1704 * blk_stop_queue - stop a queue
1705 * @q: The &struct request_queue in question
1708 * The Linux block layer assumes that a block driver will consume all
1709 * entries on the request queue when the request_fn strategy is called.
1710 * Often this will not happen, because of hardware limitations (queue
1711 * depth settings). If a device driver gets a 'queue full' response,
1712 * or if it simply chooses not to queue more I/O at one point, it can
1713 * call this function to prevent the request_fn from being called until
1714 * the driver has signalled it's ready to go again. This happens by calling
1715 * blk_start_queue() to restart queue operations. Queue lock must be held.
1717 void blk_stop_queue(struct request_queue *q)
1720 set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1722 EXPORT_SYMBOL(blk_stop_queue);
1725 * blk_sync_queue - cancel any pending callbacks on a queue
1729 * The block layer may perform asynchronous callback activity
1730 * on a queue, such as calling the unplug function after a timeout.
1731 * A block device may call blk_sync_queue to ensure that any
1732 * such activity is cancelled, thus allowing it to release resources
1733 * that the callbacks might use. The caller must already have made sure
1734 * that its ->make_request_fn will not re-add plugging prior to calling
1738 void blk_sync_queue(struct request_queue *q)
1740 del_timer_sync(&q->unplug_timer);
1741 kblockd_flush_work(&q->unplug_work);
1743 EXPORT_SYMBOL(blk_sync_queue);
1746 * blk_run_queue - run a single device queue
1747 * @q: The queue to run
1749 void blk_run_queue(struct request_queue *q)
1751 unsigned long flags;
1753 spin_lock_irqsave(q->queue_lock, flags);
1757 * Only recurse once to avoid overrunning the stack, let the unplug
1758 * handling reinvoke the handler shortly if we already got there.
1760 if (!elv_queue_empty(q)) {
1761 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1763 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1766 kblockd_schedule_work(&q->unplug_work);
1770 spin_unlock_irqrestore(q->queue_lock, flags);
1772 EXPORT_SYMBOL(blk_run_queue);
1775 * blk_cleanup_queue: - release a &struct request_queue when it is no longer needed
1776 * @kobj: the kobj belonging of the request queue to be released
1779 * blk_cleanup_queue is the pair to blk_init_queue() or
1780 * blk_queue_make_request(). It should be called when a request queue is
1781 * being released; typically when a block device is being de-registered.
1782 * Currently, its primary task it to free all the &struct request
1783 * structures that were allocated to the queue and the queue itself.
1786 * Hopefully the low level driver will have finished any
1787 * outstanding requests first...
1789 static void blk_release_queue(struct kobject *kobj)
1791 struct request_queue *q =
1792 container_of(kobj, struct request_queue, kobj);
1793 struct request_list *rl = &q->rq;
1798 mempool_destroy(rl->rq_pool);
1801 __blk_queue_free_tags(q);
1803 blk_trace_shutdown(q);
1805 bdi_destroy(&q->backing_dev_info);
1806 kmem_cache_free(requestq_cachep, q);
1809 void blk_put_queue(struct request_queue *q)
1811 kobject_put(&q->kobj);
1813 EXPORT_SYMBOL(blk_put_queue);
1815 void blk_cleanup_queue(struct request_queue * q)
1817 mutex_lock(&q->sysfs_lock);
1818 set_bit(QUEUE_FLAG_DEAD, &q->queue_flags);
1819 mutex_unlock(&q->sysfs_lock);
1822 elevator_exit(q->elevator);
1827 EXPORT_SYMBOL(blk_cleanup_queue);
1829 static int blk_init_free_list(struct request_queue *q)
1831 struct request_list *rl = &q->rq;
1833 rl->count[READ] = rl->count[WRITE] = 0;
1834 rl->starved[READ] = rl->starved[WRITE] = 0;
1836 init_waitqueue_head(&rl->wait[READ]);
1837 init_waitqueue_head(&rl->wait[WRITE]);
1839 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1840 mempool_free_slab, request_cachep, q->node);
1848 struct request_queue *blk_alloc_queue(gfp_t gfp_mask)
1850 return blk_alloc_queue_node(gfp_mask, -1);
1852 EXPORT_SYMBOL(blk_alloc_queue);
1854 static struct kobj_type queue_ktype;
1856 struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
1858 struct request_queue *q;
1861 q = kmem_cache_alloc_node(requestq_cachep,
1862 gfp_mask | __GFP_ZERO, node_id);
1866 q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1867 q->backing_dev_info.unplug_io_data = q;
1868 err = bdi_init(&q->backing_dev_info);
1870 kmem_cache_free(requestq_cachep, q);
1874 init_timer(&q->unplug_timer);
1876 kobject_set_name(&q->kobj, "%s", "queue");
1877 q->kobj.ktype = &queue_ktype;
1878 kobject_init(&q->kobj);
1880 mutex_init(&q->sysfs_lock);
1884 EXPORT_SYMBOL(blk_alloc_queue_node);
1887 * blk_init_queue - prepare a request queue for use with a block device
1888 * @rfn: The function to be called to process requests that have been
1889 * placed on the queue.
1890 * @lock: Request queue spin lock
1893 * If a block device wishes to use the standard request handling procedures,
1894 * which sorts requests and coalesces adjacent requests, then it must
1895 * call blk_init_queue(). The function @rfn will be called when there
1896 * are requests on the queue that need to be processed. If the device
1897 * supports plugging, then @rfn may not be called immediately when requests
1898 * are available on the queue, but may be called at some time later instead.
1899 * Plugged queues are generally unplugged when a buffer belonging to one
1900 * of the requests on the queue is needed, or due to memory pressure.
1902 * @rfn is not required, or even expected, to remove all requests off the
1903 * queue, but only as many as it can handle at a time. If it does leave
1904 * requests on the queue, it is responsible for arranging that the requests
1905 * get dealt with eventually.
1907 * The queue spin lock must be held while manipulating the requests on the
1908 * request queue; this lock will be taken also from interrupt context, so irq
1909 * disabling is needed for it.
1911 * Function returns a pointer to the initialized request queue, or NULL if
1912 * it didn't succeed.
1915 * blk_init_queue() must be paired with a blk_cleanup_queue() call
1916 * when the block device is deactivated (such as at module unload).
1919 struct request_queue *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1921 return blk_init_queue_node(rfn, lock, -1);
1923 EXPORT_SYMBOL(blk_init_queue);
1925 struct request_queue *
1926 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1928 struct request_queue *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1934 if (blk_init_free_list(q)) {
1935 kmem_cache_free(requestq_cachep, q);
1940 * if caller didn't supply a lock, they get per-queue locking with
1944 spin_lock_init(&q->__queue_lock);
1945 lock = &q->__queue_lock;
1948 q->request_fn = rfn;
1949 q->prep_rq_fn = NULL;
1950 q->unplug_fn = generic_unplug_device;
1951 q->queue_flags = (1 << QUEUE_FLAG_CLUSTER);
1952 q->queue_lock = lock;
1954 blk_queue_segment_boundary(q, 0xffffffff);
1956 blk_queue_make_request(q, __make_request);
1957 blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1959 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1960 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1962 q->sg_reserved_size = INT_MAX;
1967 if (!elevator_init(q, NULL)) {
1968 blk_queue_congestion_threshold(q);
1975 EXPORT_SYMBOL(blk_init_queue_node);
1977 int blk_get_queue(struct request_queue *q)
1979 if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1980 kobject_get(&q->kobj);
1987 EXPORT_SYMBOL(blk_get_queue);
1989 static inline void blk_free_request(struct request_queue *q, struct request *rq)
1991 if (rq->cmd_flags & REQ_ELVPRIV)
1992 elv_put_request(q, rq);
1993 mempool_free(rq, q->rq.rq_pool);
1996 static struct request *
1997 blk_alloc_request(struct request_queue *q, int rw, int priv, gfp_t gfp_mask)
1999 struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
2005 * first three bits are identical in rq->cmd_flags and bio->bi_rw,
2006 * see bio.h and blkdev.h
2008 rq->cmd_flags = rw | REQ_ALLOCED;
2011 if (unlikely(elv_set_request(q, rq, gfp_mask))) {
2012 mempool_free(rq, q->rq.rq_pool);
2015 rq->cmd_flags |= REQ_ELVPRIV;
2022 * ioc_batching returns true if the ioc is a valid batching request and
2023 * should be given priority access to a request.
2025 static inline int ioc_batching(struct request_queue *q, struct io_context *ioc)
2031 * Make sure the process is able to allocate at least 1 request
2032 * even if the batch times out, otherwise we could theoretically
2035 return ioc->nr_batch_requests == q->nr_batching ||
2036 (ioc->nr_batch_requests > 0
2037 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
2041 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
2042 * will cause the process to be a "batcher" on all queues in the system. This
2043 * is the behaviour we want though - once it gets a wakeup it should be given
2046 static void ioc_set_batching(struct request_queue *q, struct io_context *ioc)
2048 if (!ioc || ioc_batching(q, ioc))
2051 ioc->nr_batch_requests = q->nr_batching;
2052 ioc->last_waited = jiffies;
2055 static void __freed_request(struct request_queue *q, int rw)
2057 struct request_list *rl = &q->rq;
2059 if (rl->count[rw] < queue_congestion_off_threshold(q))
2060 blk_clear_queue_congested(q, rw);
2062 if (rl->count[rw] + 1 <= q->nr_requests) {
2063 if (waitqueue_active(&rl->wait[rw]))
2064 wake_up(&rl->wait[rw]);
2066 blk_clear_queue_full(q, rw);
2071 * A request has just been released. Account for it, update the full and
2072 * congestion status, wake up any waiters. Called under q->queue_lock.
2074 static void freed_request(struct request_queue *q, int rw, int priv)
2076 struct request_list *rl = &q->rq;
2082 __freed_request(q, rw);
2084 if (unlikely(rl->starved[rw ^ 1]))
2085 __freed_request(q, rw ^ 1);
2088 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2090 * Get a free request, queue_lock must be held.
2091 * Returns NULL on failure, with queue_lock held.
2092 * Returns !NULL on success, with queue_lock *not held*.
2094 static struct request *get_request(struct request_queue *q, int rw_flags,
2095 struct bio *bio, gfp_t gfp_mask)
2097 struct request *rq = NULL;
2098 struct request_list *rl = &q->rq;
2099 struct io_context *ioc = NULL;
2100 const int rw = rw_flags & 0x01;
2101 int may_queue, priv;
2103 may_queue = elv_may_queue(q, rw_flags);
2104 if (may_queue == ELV_MQUEUE_NO)
2107 if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
2108 if (rl->count[rw]+1 >= q->nr_requests) {
2109 ioc = current_io_context(GFP_ATOMIC, q->node);
2111 * The queue will fill after this allocation, so set
2112 * it as full, and mark this process as "batching".
2113 * This process will be allowed to complete a batch of
2114 * requests, others will be blocked.
2116 if (!blk_queue_full(q, rw)) {
2117 ioc_set_batching(q, ioc);
2118 blk_set_queue_full(q, rw);
2120 if (may_queue != ELV_MQUEUE_MUST
2121 && !ioc_batching(q, ioc)) {
2123 * The queue is full and the allocating
2124 * process is not a "batcher", and not
2125 * exempted by the IO scheduler
2131 blk_set_queue_congested(q, rw);
2135 * Only allow batching queuers to allocate up to 50% over the defined
2136 * limit of requests, otherwise we could have thousands of requests
2137 * allocated with any setting of ->nr_requests
2139 if (rl->count[rw] >= (3 * q->nr_requests / 2))
2143 rl->starved[rw] = 0;
2145 priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
2149 spin_unlock_irq(q->queue_lock);
2151 rq = blk_alloc_request(q, rw_flags, priv, gfp_mask);
2152 if (unlikely(!rq)) {
2154 * Allocation failed presumably due to memory. Undo anything
2155 * we might have messed up.
2157 * Allocating task should really be put onto the front of the
2158 * wait queue, but this is pretty rare.
2160 spin_lock_irq(q->queue_lock);
2161 freed_request(q, rw, priv);
2164 * in the very unlikely event that allocation failed and no
2165 * requests for this direction was pending, mark us starved
2166 * so that freeing of a request in the other direction will
2167 * notice us. another possible fix would be to split the
2168 * rq mempool into READ and WRITE
2171 if (unlikely(rl->count[rw] == 0))
2172 rl->starved[rw] = 1;
2178 * ioc may be NULL here, and ioc_batching will be false. That's
2179 * OK, if the queue is under the request limit then requests need
2180 * not count toward the nr_batch_requests limit. There will always
2181 * be some limit enforced by BLK_BATCH_TIME.
2183 if (ioc_batching(q, ioc))
2184 ioc->nr_batch_requests--;
2188 blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
2194 * No available requests for this queue, unplug the device and wait for some
2195 * requests to become available.
2197 * Called with q->queue_lock held, and returns with it unlocked.
2199 static struct request *get_request_wait(struct request_queue *q, int rw_flags,
2202 const int rw = rw_flags & 0x01;
2205 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2208 struct request_list *rl = &q->rq;
2210 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2211 TASK_UNINTERRUPTIBLE);
2213 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2216 struct io_context *ioc;
2218 blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
2220 __generic_unplug_device(q);
2221 spin_unlock_irq(q->queue_lock);
2225 * After sleeping, we become a "batching" process and
2226 * will be able to allocate at least one request, and
2227 * up to a big batch of them for a small period time.
2228 * See ioc_batching, ioc_set_batching
2230 ioc = current_io_context(GFP_NOIO, q->node);
2231 ioc_set_batching(q, ioc);
2233 spin_lock_irq(q->queue_lock);
2235 finish_wait(&rl->wait[rw], &wait);
2241 struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
2245 BUG_ON(rw != READ && rw != WRITE);
2247 spin_lock_irq(q->queue_lock);
2248 if (gfp_mask & __GFP_WAIT) {
2249 rq = get_request_wait(q, rw, NULL);
2251 rq = get_request(q, rw, NULL, gfp_mask);
2253 spin_unlock_irq(q->queue_lock);
2255 /* q->queue_lock is unlocked at this point */
2259 EXPORT_SYMBOL(blk_get_request);
2262 * blk_start_queueing - initiate dispatch of requests to device
2263 * @q: request queue to kick into gear
2265 * This is basically a helper to remove the need to know whether a queue
2266 * is plugged or not if someone just wants to initiate dispatch of requests
2269 * The queue lock must be held with interrupts disabled.
2271 void blk_start_queueing(struct request_queue *q)
2273 if (!blk_queue_plugged(q))
2276 __generic_unplug_device(q);
2278 EXPORT_SYMBOL(blk_start_queueing);
2281 * blk_requeue_request - put a request back on queue
2282 * @q: request queue where request should be inserted
2283 * @rq: request to be inserted
2286 * Drivers often keep queueing requests until the hardware cannot accept
2287 * more, when that condition happens we need to put the request back
2288 * on the queue. Must be called with queue lock held.
2290 void blk_requeue_request(struct request_queue *q, struct request *rq)
2292 blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
2294 if (blk_rq_tagged(rq))
2295 blk_queue_end_tag(q, rq);
2297 elv_requeue_request(q, rq);
2300 EXPORT_SYMBOL(blk_requeue_request);
2303 * blk_insert_request - insert a special request in to a request queue
2304 * @q: request queue where request should be inserted
2305 * @rq: request to be inserted
2306 * @at_head: insert request at head or tail of queue
2307 * @data: private data
2310 * Many block devices need to execute commands asynchronously, so they don't
2311 * block the whole kernel from preemption during request execution. This is
2312 * accomplished normally by inserting aritficial requests tagged as
2313 * REQ_SPECIAL in to the corresponding request queue, and letting them be
2314 * scheduled for actual execution by the request queue.
2316 * We have the option of inserting the head or the tail of the queue.
2317 * Typically we use the tail for new ioctls and so forth. We use the head
2318 * of the queue for things like a QUEUE_FULL message from a device, or a
2319 * host that is unable to accept a particular command.
2321 void blk_insert_request(struct request_queue *q, struct request *rq,
2322 int at_head, void *data)
2324 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2325 unsigned long flags;
2328 * tell I/O scheduler that this isn't a regular read/write (ie it
2329 * must not attempt merges on this) and that it acts as a soft
2332 rq->cmd_type = REQ_TYPE_SPECIAL;
2333 rq->cmd_flags |= REQ_SOFTBARRIER;
2337 spin_lock_irqsave(q->queue_lock, flags);
2340 * If command is tagged, release the tag
2342 if (blk_rq_tagged(rq))
2343 blk_queue_end_tag(q, rq);
2345 drive_stat_acct(rq, 1);
2346 __elv_add_request(q, rq, where, 0);
2347 blk_start_queueing(q);
2348 spin_unlock_irqrestore(q->queue_lock, flags);
2351 EXPORT_SYMBOL(blk_insert_request);
2353 static int __blk_rq_unmap_user(struct bio *bio)
2358 if (bio_flagged(bio, BIO_USER_MAPPED))
2359 bio_unmap_user(bio);
2361 ret = bio_uncopy_user(bio);
2367 int blk_rq_append_bio(struct request_queue *q, struct request *rq,
2371 blk_rq_bio_prep(q, rq, bio);
2372 else if (!ll_back_merge_fn(q, rq, bio))
2375 rq->biotail->bi_next = bio;
2378 rq->data_len += bio->bi_size;
2382 EXPORT_SYMBOL(blk_rq_append_bio);
2384 static int __blk_rq_map_user(struct request_queue *q, struct request *rq,
2385 void __user *ubuf, unsigned int len)
2387 unsigned long uaddr;
2388 struct bio *bio, *orig_bio;
2391 reading = rq_data_dir(rq) == READ;
2394 * if alignment requirement is satisfied, map in user pages for
2395 * direct dma. else, set up kernel bounce buffers
2397 uaddr = (unsigned long) ubuf;
2398 if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2399 bio = bio_map_user(q, NULL, uaddr, len, reading);
2401 bio = bio_copy_user(q, uaddr, len, reading);
2404 return PTR_ERR(bio);
2407 blk_queue_bounce(q, &bio);
2410 * We link the bounce buffer in and could have to traverse it
2411 * later so we have to get a ref to prevent it from being freed
2415 ret = blk_rq_append_bio(q, rq, bio);
2417 return bio->bi_size;
2419 /* if it was boucned we must call the end io function */
2421 __blk_rq_unmap_user(orig_bio);
2427 * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2428 * @q: request queue where request should be inserted
2429 * @rq: request structure to fill
2430 * @ubuf: the user buffer
2431 * @len: length of user data
2434 * Data will be mapped directly for zero copy io, if possible. Otherwise
2435 * a kernel bounce buffer is used.
2437 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2438 * still in process context.
2440 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2441 * before being submitted to the device, as pages mapped may be out of
2442 * reach. It's the callers responsibility to make sure this happens. The
2443 * original bio must be passed back in to blk_rq_unmap_user() for proper
2446 int blk_rq_map_user(struct request_queue *q, struct request *rq,
2447 void __user *ubuf, unsigned long len)
2449 unsigned long bytes_read = 0;
2450 struct bio *bio = NULL;
2453 if (len > (q->max_hw_sectors << 9))
2458 while (bytes_read != len) {
2459 unsigned long map_len, end, start;
2461 map_len = min_t(unsigned long, len - bytes_read, BIO_MAX_SIZE);
2462 end = ((unsigned long)ubuf + map_len + PAGE_SIZE - 1)
2464 start = (unsigned long)ubuf >> PAGE_SHIFT;
2467 * A bad offset could cause us to require BIO_MAX_PAGES + 1
2468 * pages. If this happens we just lower the requested
2469 * mapping len by a page so that we can fit
2471 if (end - start > BIO_MAX_PAGES)
2472 map_len -= PAGE_SIZE;
2474 ret = __blk_rq_map_user(q, rq, ubuf, map_len);
2483 rq->buffer = rq->data = NULL;
2486 blk_rq_unmap_user(bio);
2490 EXPORT_SYMBOL(blk_rq_map_user);
2493 * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2494 * @q: request queue where request should be inserted
2495 * @rq: request to map data to
2496 * @iov: pointer to the iovec
2497 * @iov_count: number of elements in the iovec
2498 * @len: I/O byte count
2501 * Data will be mapped directly for zero copy io, if possible. Otherwise
2502 * a kernel bounce buffer is used.
2504 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2505 * still in process context.
2507 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2508 * before being submitted to the device, as pages mapped may be out of
2509 * reach. It's the callers responsibility to make sure this happens. The
2510 * original bio must be passed back in to blk_rq_unmap_user() for proper
2513 int blk_rq_map_user_iov(struct request_queue *q, struct request *rq,
2514 struct sg_iovec *iov, int iov_count, unsigned int len)
2518 if (!iov || iov_count <= 0)
2521 /* we don't allow misaligned data like bio_map_user() does. If the
2522 * user is using sg, they're expected to know the alignment constraints
2523 * and respect them accordingly */
2524 bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2526 return PTR_ERR(bio);
2528 if (bio->bi_size != len) {
2530 bio_unmap_user(bio);
2535 blk_rq_bio_prep(q, rq, bio);
2536 rq->buffer = rq->data = NULL;
2540 EXPORT_SYMBOL(blk_rq_map_user_iov);
2543 * blk_rq_unmap_user - unmap a request with user data
2544 * @bio: start of bio list
2547 * Unmap a rq previously mapped by blk_rq_map_user(). The caller must
2548 * supply the original rq->bio from the blk_rq_map_user() return, since
2549 * the io completion may have changed rq->bio.
2551 int blk_rq_unmap_user(struct bio *bio)
2553 struct bio *mapped_bio;
2558 if (unlikely(bio_flagged(bio, BIO_BOUNCED)))
2559 mapped_bio = bio->bi_private;
2561 ret2 = __blk_rq_unmap_user(mapped_bio);
2567 bio_put(mapped_bio);
2573 EXPORT_SYMBOL(blk_rq_unmap_user);
2576 * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2577 * @q: request queue where request should be inserted
2578 * @rq: request to fill
2579 * @kbuf: the kernel buffer
2580 * @len: length of user data
2581 * @gfp_mask: memory allocation flags
2583 int blk_rq_map_kern(struct request_queue *q, struct request *rq, void *kbuf,
2584 unsigned int len, gfp_t gfp_mask)
2588 if (len > (q->max_hw_sectors << 9))
2593 bio = bio_map_kern(q, kbuf, len, gfp_mask);
2595 return PTR_ERR(bio);
2597 if (rq_data_dir(rq) == WRITE)
2598 bio->bi_rw |= (1 << BIO_RW);
2600 blk_rq_bio_prep(q, rq, bio);
2601 blk_queue_bounce(q, &rq->bio);
2602 rq->buffer = rq->data = NULL;
2606 EXPORT_SYMBOL(blk_rq_map_kern);
2609 * blk_execute_rq_nowait - insert a request into queue for execution
2610 * @q: queue to insert the request in
2611 * @bd_disk: matching gendisk
2612 * @rq: request to insert
2613 * @at_head: insert request at head or tail of queue
2614 * @done: I/O completion handler
2617 * Insert a fully prepared request at the back of the io scheduler queue
2618 * for execution. Don't wait for completion.
2620 void blk_execute_rq_nowait(struct request_queue *q, struct gendisk *bd_disk,
2621 struct request *rq, int at_head,
2624 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2626 rq->rq_disk = bd_disk;
2627 rq->cmd_flags |= REQ_NOMERGE;
2629 WARN_ON(irqs_disabled());
2630 spin_lock_irq(q->queue_lock);
2631 __elv_add_request(q, rq, where, 1);
2632 __generic_unplug_device(q);
2633 spin_unlock_irq(q->queue_lock);
2635 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
2638 * blk_execute_rq - insert a request into queue for execution
2639 * @q: queue to insert the request in
2640 * @bd_disk: matching gendisk
2641 * @rq: request to insert
2642 * @at_head: insert request at head or tail of queue
2645 * Insert a fully prepared request at the back of the io scheduler queue
2646 * for execution and wait for completion.
2648 int blk_execute_rq(struct request_queue *q, struct gendisk *bd_disk,
2649 struct request *rq, int at_head)
2651 DECLARE_COMPLETION_ONSTACK(wait);
2652 char sense[SCSI_SENSE_BUFFERSIZE];
2656 * we need an extra reference to the request, so we can look at
2657 * it after io completion
2662 memset(sense, 0, sizeof(sense));
2667 rq->end_io_data = &wait;
2668 blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
2669 wait_for_completion(&wait);
2677 EXPORT_SYMBOL(blk_execute_rq);
2679 static void bio_end_empty_barrier(struct bio *bio, int err)
2682 clear_bit(BIO_UPTODATE, &bio->bi_flags);
2684 complete(bio->bi_private);
2688 * blkdev_issue_flush - queue a flush
2689 * @bdev: blockdev to issue flush for
2690 * @error_sector: error sector
2693 * Issue a flush for the block device in question. Caller can supply
2694 * room for storing the error offset in case of a flush error, if they
2695 * wish to. Caller must run wait_for_completion() on its own.
2697 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2699 DECLARE_COMPLETION_ONSTACK(wait);
2700 struct request_queue *q;
2704 if (bdev->bd_disk == NULL)
2707 q = bdev_get_queue(bdev);
2711 bio = bio_alloc(GFP_KERNEL, 0);
2715 bio->bi_end_io = bio_end_empty_barrier;
2716 bio->bi_private = &wait;
2717 bio->bi_bdev = bdev;
2718 submit_bio(1 << BIO_RW_BARRIER, bio);
2720 wait_for_completion(&wait);
2723 * The driver must store the error location in ->bi_sector, if
2724 * it supports it. For non-stacked drivers, this should be copied
2728 *error_sector = bio->bi_sector;
2731 if (!bio_flagged(bio, BIO_UPTODATE))
2738 EXPORT_SYMBOL(blkdev_issue_flush);
2740 static void drive_stat_acct(struct request *rq, int new_io)
2742 int rw = rq_data_dir(rq);
2744 if (!blk_fs_request(rq) || !rq->rq_disk)
2748 __disk_stat_inc(rq->rq_disk, merges[rw]);
2750 disk_round_stats(rq->rq_disk);
2751 rq->rq_disk->in_flight++;
2756 * add-request adds a request to the linked list.
2757 * queue lock is held and interrupts disabled, as we muck with the
2758 * request queue list.
2760 static inline void add_request(struct request_queue * q, struct request * req)
2762 drive_stat_acct(req, 1);
2765 * elevator indicated where it wants this request to be
2766 * inserted at elevator_merge time
2768 __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2772 * disk_round_stats() - Round off the performance stats on a struct
2775 * The average IO queue length and utilisation statistics are maintained
2776 * by observing the current state of the queue length and the amount of
2777 * time it has been in this state for.
2779 * Normally, that accounting is done on IO completion, but that can result
2780 * in more than a second's worth of IO being accounted for within any one
2781 * second, leading to >100% utilisation. To deal with that, we call this
2782 * function to do a round-off before returning the results when reading
2783 * /proc/diskstats. This accounts immediately for all queue usage up to
2784 * the current jiffies and restarts the counters again.
2786 void disk_round_stats(struct gendisk *disk)
2788 unsigned long now = jiffies;
2790 if (now == disk->stamp)
2793 if (disk->in_flight) {
2794 __disk_stat_add(disk, time_in_queue,
2795 disk->in_flight * (now - disk->stamp));
2796 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2801 EXPORT_SYMBOL_GPL(disk_round_stats);
2804 * queue lock must be held
2806 void __blk_put_request(struct request_queue *q, struct request *req)
2810 if (unlikely(--req->ref_count))
2813 elv_completed_request(q, req);
2816 * Request may not have originated from ll_rw_blk. if not,
2817 * it didn't come out of our reserved rq pools
2819 if (req->cmd_flags & REQ_ALLOCED) {
2820 int rw = rq_data_dir(req);
2821 int priv = req->cmd_flags & REQ_ELVPRIV;
2823 BUG_ON(!list_empty(&req->queuelist));
2824 BUG_ON(!hlist_unhashed(&req->hash));
2826 blk_free_request(q, req);
2827 freed_request(q, rw, priv);
2831 EXPORT_SYMBOL_GPL(__blk_put_request);
2833 void blk_put_request(struct request *req)
2835 unsigned long flags;
2836 struct request_queue *q = req->q;
2839 * Gee, IDE calls in w/ NULL q. Fix IDE and remove the
2840 * following if (q) test.
2843 spin_lock_irqsave(q->queue_lock, flags);
2844 __blk_put_request(q, req);
2845 spin_unlock_irqrestore(q->queue_lock, flags);
2849 EXPORT_SYMBOL(blk_put_request);
2852 * blk_end_sync_rq - executes a completion event on a request
2853 * @rq: request to complete
2854 * @error: end io status of the request
2856 void blk_end_sync_rq(struct request *rq, int error)
2858 struct completion *waiting = rq->end_io_data;
2860 rq->end_io_data = NULL;
2861 __blk_put_request(rq->q, rq);
2864 * complete last, if this is a stack request the process (and thus
2865 * the rq pointer) could be invalid right after this complete()
2869 EXPORT_SYMBOL(blk_end_sync_rq);
2872 * Has to be called with the request spinlock acquired
2874 static int attempt_merge(struct request_queue *q, struct request *req,
2875 struct request *next)
2877 if (!rq_mergeable(req) || !rq_mergeable(next))
2883 if (req->sector + req->nr_sectors != next->sector)
2886 if (rq_data_dir(req) != rq_data_dir(next)
2887 || req->rq_disk != next->rq_disk
2892 * If we are allowed to merge, then append bio list
2893 * from next to rq and release next. merge_requests_fn
2894 * will have updated segment counts, update sector
2897 if (!ll_merge_requests_fn(q, req, next))
2901 * At this point we have either done a back merge
2902 * or front merge. We need the smaller start_time of
2903 * the merged requests to be the current request
2904 * for accounting purposes.
2906 if (time_after(req->start_time, next->start_time))
2907 req->start_time = next->start_time;
2909 req->biotail->bi_next = next->bio;
2910 req->biotail = next->biotail;
2912 req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2914 elv_merge_requests(q, req, next);
2917 disk_round_stats(req->rq_disk);
2918 req->rq_disk->in_flight--;
2921 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2923 __blk_put_request(q, next);
2927 static inline int attempt_back_merge(struct request_queue *q,
2930 struct request *next = elv_latter_request(q, rq);
2933 return attempt_merge(q, rq, next);
2938 static inline int attempt_front_merge(struct request_queue *q,
2941 struct request *prev = elv_former_request(q, rq);
2944 return attempt_merge(q, prev, rq);
2949 static void init_request_from_bio(struct request *req, struct bio *bio)
2951 req->cmd_type = REQ_TYPE_FS;
2954 * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2956 if (bio_rw_ahead(bio) || bio_failfast(bio))
2957 req->cmd_flags |= REQ_FAILFAST;
2960 * REQ_BARRIER implies no merging, but lets make it explicit
2962 if (unlikely(bio_barrier(bio)))
2963 req->cmd_flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2966 req->cmd_flags |= REQ_RW_SYNC;
2967 if (bio_rw_meta(bio))
2968 req->cmd_flags |= REQ_RW_META;
2971 req->hard_sector = req->sector = bio->bi_sector;
2972 req->ioprio = bio_prio(bio);
2973 req->start_time = jiffies;
2974 blk_rq_bio_prep(req->q, req, bio);
2977 static int __make_request(struct request_queue *q, struct bio *bio)
2979 struct request *req;
2980 int el_ret, nr_sectors, barrier, err;
2981 const unsigned short prio = bio_prio(bio);
2982 const int sync = bio_sync(bio);
2985 nr_sectors = bio_sectors(bio);
2988 * low level driver can indicate that it wants pages above a
2989 * certain limit bounced to low memory (ie for highmem, or even
2990 * ISA dma in theory)
2992 blk_queue_bounce(q, &bio);
2994 barrier = bio_barrier(bio);
2995 if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
3000 spin_lock_irq(q->queue_lock);
3002 if (unlikely(barrier) || elv_queue_empty(q))
3005 el_ret = elv_merge(q, &req, bio);
3007 case ELEVATOR_BACK_MERGE:
3008 BUG_ON(!rq_mergeable(req));
3010 if (!ll_back_merge_fn(q, req, bio))
3013 blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
3015 req->biotail->bi_next = bio;
3017 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
3018 req->ioprio = ioprio_best(req->ioprio, prio);
3019 drive_stat_acct(req, 0);
3020 if (!attempt_back_merge(q, req))
3021 elv_merged_request(q, req, el_ret);
3024 case ELEVATOR_FRONT_MERGE:
3025 BUG_ON(!rq_mergeable(req));
3027 if (!ll_front_merge_fn(q, req, bio))
3030 blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
3032 bio->bi_next = req->bio;
3036 * may not be valid. if the low level driver said
3037 * it didn't need a bounce buffer then it better
3038 * not touch req->buffer either...
3040 req->buffer = bio_data(bio);
3041 req->current_nr_sectors = bio_cur_sectors(bio);
3042 req->hard_cur_sectors = req->current_nr_sectors;
3043 req->sector = req->hard_sector = bio->bi_sector;
3044 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
3045 req->ioprio = ioprio_best(req->ioprio, prio);
3046 drive_stat_acct(req, 0);
3047 if (!attempt_front_merge(q, req))
3048 elv_merged_request(q, req, el_ret);
3051 /* ELV_NO_MERGE: elevator says don't/can't merge. */
3058 * This sync check and mask will be re-done in init_request_from_bio(),
3059 * but we need to set it earlier to expose the sync flag to the
3060 * rq allocator and io schedulers.
3062 rw_flags = bio_data_dir(bio);
3064 rw_flags |= REQ_RW_SYNC;
3067 * Grab a free request. This is might sleep but can not fail.
3068 * Returns with the queue unlocked.
3070 req = get_request_wait(q, rw_flags, bio);
3073 * After dropping the lock and possibly sleeping here, our request
3074 * may now be mergeable after it had proven unmergeable (above).
3075 * We don't worry about that case for efficiency. It won't happen
3076 * often, and the elevators are able to handle it.
3078 init_request_from_bio(req, bio);
3080 spin_lock_irq(q->queue_lock);
3081 if (elv_queue_empty(q))
3083 add_request(q, req);
3086 __generic_unplug_device(q);
3088 spin_unlock_irq(q->queue_lock);
3092 bio_endio(bio, err);
3097 * If bio->bi_dev is a partition, remap the location
3099 static inline void blk_partition_remap(struct bio *bio)
3101 struct block_device *bdev = bio->bi_bdev;
3103 if (bio_sectors(bio) && bdev != bdev->bd_contains) {
3104 struct hd_struct *p = bdev->bd_part;
3105 const int rw = bio_data_dir(bio);
3107 p->sectors[rw] += bio_sectors(bio);
3110 bio->bi_sector += p->start_sect;
3111 bio->bi_bdev = bdev->bd_contains;
3113 blk_add_trace_remap(bdev_get_queue(bio->bi_bdev), bio,
3114 bdev->bd_dev, bio->bi_sector,
3115 bio->bi_sector - p->start_sect);
3119 static void handle_bad_sector(struct bio *bio)
3121 char b[BDEVNAME_SIZE];
3123 printk(KERN_INFO "attempt to access beyond end of device\n");
3124 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
3125 bdevname(bio->bi_bdev, b),
3127 (unsigned long long)bio->bi_sector + bio_sectors(bio),
3128 (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
3130 set_bit(BIO_EOF, &bio->bi_flags);
3133 #ifdef CONFIG_FAIL_MAKE_REQUEST
3135 static DECLARE_FAULT_ATTR(fail_make_request);
3137 static int __init setup_fail_make_request(char *str)
3139 return setup_fault_attr(&fail_make_request, str);
3141 __setup("fail_make_request=", setup_fail_make_request);
3143 static int should_fail_request(struct bio *bio)
3145 if ((bio->bi_bdev->bd_disk->flags & GENHD_FL_FAIL) ||
3146 (bio->bi_bdev->bd_part && bio->bi_bdev->bd_part->make_it_fail))
3147 return should_fail(&fail_make_request, bio->bi_size);
3152 static int __init fail_make_request_debugfs(void)
3154 return init_fault_attr_dentries(&fail_make_request,
3155 "fail_make_request");
3158 late_initcall(fail_make_request_debugfs);
3160 #else /* CONFIG_FAIL_MAKE_REQUEST */
3162 static inline int should_fail_request(struct bio *bio)
3167 #endif /* CONFIG_FAIL_MAKE_REQUEST */
3170 * Check whether this bio extends beyond the end of the device.
3172 static inline int bio_check_eod(struct bio *bio, unsigned int nr_sectors)
3179 /* Test device or partition size, when known. */
3180 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3182 sector_t sector = bio->bi_sector;
3184 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
3186 * This may well happen - the kernel calls bread()
3187 * without checking the size of the device, e.g., when
3188 * mounting a device.
3190 handle_bad_sector(bio);
3199 * generic_make_request: hand a buffer to its device driver for I/O
3200 * @bio: The bio describing the location in memory and on the device.
3202 * generic_make_request() is used to make I/O requests of block
3203 * devices. It is passed a &struct bio, which describes the I/O that needs
3206 * generic_make_request() does not return any status. The
3207 * success/failure status of the request, along with notification of
3208 * completion, is delivered asynchronously through the bio->bi_end_io
3209 * function described (one day) else where.
3211 * The caller of generic_make_request must make sure that bi_io_vec
3212 * are set to describe the memory buffer, and that bi_dev and bi_sector are
3213 * set to describe the device address, and the
3214 * bi_end_io and optionally bi_private are set to describe how
3215 * completion notification should be signaled.
3217 * generic_make_request and the drivers it calls may use bi_next if this
3218 * bio happens to be merged with someone else, and may change bi_dev and
3219 * bi_sector for remaps as it sees fit. So the values of these fields
3220 * should NOT be depended on after the call to generic_make_request.
3222 static inline void __generic_make_request(struct bio *bio)
3224 struct request_queue *q;
3225 sector_t old_sector;
3226 int ret, nr_sectors = bio_sectors(bio);
3231 if (bio_check_eod(bio, nr_sectors))
3235 * Resolve the mapping until finished. (drivers are
3236 * still free to implement/resolve their own stacking
3237 * by explicitly returning 0)
3239 * NOTE: we don't repeat the blk_size check for each new device.
3240 * Stacking drivers are expected to know what they are doing.
3245 char b[BDEVNAME_SIZE];
3247 q = bdev_get_queue(bio->bi_bdev);
3250 "generic_make_request: Trying to access "
3251 "nonexistent block-device %s (%Lu)\n",
3252 bdevname(bio->bi_bdev, b),
3253 (long long) bio->bi_sector);
3255 bio_endio(bio, -EIO);
3259 if (unlikely(nr_sectors > q->max_hw_sectors)) {
3260 printk("bio too big device %s (%u > %u)\n",
3261 bdevname(bio->bi_bdev, b),
3267 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
3270 if (should_fail_request(bio))
3274 * If this device has partitions, remap block n
3275 * of partition p to block n+start(p) of the disk.
3277 blk_partition_remap(bio);
3279 if (old_sector != -1)
3280 blk_add_trace_remap(q, bio, old_dev, bio->bi_sector,
3283 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
3285 old_sector = bio->bi_sector;
3286 old_dev = bio->bi_bdev->bd_dev;
3288 if (bio_check_eod(bio, nr_sectors))
3291 ret = q->make_request_fn(q, bio);
3296 * We only want one ->make_request_fn to be active at a time,
3297 * else stack usage with stacked devices could be a problem.
3298 * So use current->bio_{list,tail} to keep a list of requests
3299 * submited by a make_request_fn function.
3300 * current->bio_tail is also used as a flag to say if
3301 * generic_make_request is currently active in this task or not.
3302 * If it is NULL, then no make_request is active. If it is non-NULL,
3303 * then a make_request is active, and new requests should be added
3306 void generic_make_request(struct bio *bio)
3308 if (current->bio_tail) {
3309 /* make_request is active */
3310 *(current->bio_tail) = bio;
3311 bio->bi_next = NULL;
3312 current->bio_tail = &bio->bi_next;
3315 /* following loop may be a bit non-obvious, and so deserves some
3317 * Before entering the loop, bio->bi_next is NULL (as all callers
3318 * ensure that) so we have a list with a single bio.
3319 * We pretend that we have just taken it off a longer list, so
3320 * we assign bio_list to the next (which is NULL) and bio_tail
3321 * to &bio_list, thus initialising the bio_list of new bios to be
3322 * added. __generic_make_request may indeed add some more bios
3323 * through a recursive call to generic_make_request. If it
3324 * did, we find a non-NULL value in bio_list and re-enter the loop
3325 * from the top. In this case we really did just take the bio
3326 * of the top of the list (no pretending) and so fixup bio_list and
3327 * bio_tail or bi_next, and call into __generic_make_request again.
3329 * The loop was structured like this to make only one call to
3330 * __generic_make_request (which is important as it is large and
3331 * inlined) and to keep the structure simple.
3333 BUG_ON(bio->bi_next);
3335 current->bio_list = bio->bi_next;
3336 if (bio->bi_next == NULL)
3337 current->bio_tail = ¤t->bio_list;
3339 bio->bi_next = NULL;
3340 __generic_make_request(bio);
3341 bio = current->bio_list;
3343 current->bio_tail = NULL; /* deactivate */
3346 EXPORT_SYMBOL(generic_make_request);
3349 * submit_bio: submit a bio to the block device layer for I/O
3350 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3351 * @bio: The &struct bio which describes the I/O
3353 * submit_bio() is very similar in purpose to generic_make_request(), and
3354 * uses that function to do most of the work. Both are fairly rough
3355 * interfaces, @bio must be presetup and ready for I/O.
3358 void submit_bio(int rw, struct bio *bio)
3360 int count = bio_sectors(bio);
3365 * If it's a regular read/write or a barrier with data attached,
3366 * go through the normal accounting stuff before submission.
3368 if (!bio_empty_barrier(bio)) {
3370 BIO_BUG_ON(!bio->bi_size);
3371 BIO_BUG_ON(!bio->bi_io_vec);
3374 count_vm_events(PGPGOUT, count);
3376 task_io_account_read(bio->bi_size);
3377 count_vm_events(PGPGIN, count);
3380 if (unlikely(block_dump)) {
3381 char b[BDEVNAME_SIZE];
3382 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
3383 current->comm, task_pid_nr(current),
3384 (rw & WRITE) ? "WRITE" : "READ",
3385 (unsigned long long)bio->bi_sector,
3386 bdevname(bio->bi_bdev,b));
3390 generic_make_request(bio);
3393 EXPORT_SYMBOL(submit_bio);
3395 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
3397 if (blk_fs_request(rq)) {
3398 rq->hard_sector += nsect;
3399 rq->hard_nr_sectors -= nsect;
3402 * Move the I/O submission pointers ahead if required.
3404 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3405 (rq->sector <= rq->hard_sector)) {
3406 rq->sector = rq->hard_sector;
3407 rq->nr_sectors = rq->hard_nr_sectors;
3408 rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3409 rq->current_nr_sectors = rq->hard_cur_sectors;
3410 rq->buffer = bio_data(rq->bio);
3414 * if total number of sectors is less than the first segment
3415 * size, something has gone terribly wrong
3417 if (rq->nr_sectors < rq->current_nr_sectors) {
3418 printk("blk: request botched\n");
3419 rq->nr_sectors = rq->current_nr_sectors;
3424 static int __end_that_request_first(struct request *req, int uptodate,
3427 int total_bytes, bio_nbytes, error, next_idx = 0;
3430 blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
3433 * extend uptodate bool to allow < 0 value to be direct io error
3436 if (end_io_error(uptodate))
3437 error = !uptodate ? -EIO : uptodate;
3440 * for a REQ_BLOCK_PC request, we want to carry any eventual
3441 * sense key with us all the way through
3443 if (!blk_pc_request(req))
3447 if (blk_fs_request(req) && !(req->cmd_flags & REQ_QUIET))
3448 printk("end_request: I/O error, dev %s, sector %llu\n",
3449 req->rq_disk ? req->rq_disk->disk_name : "?",
3450 (unsigned long long)req->sector);
3453 if (blk_fs_request(req) && req->rq_disk) {
3454 const int rw = rq_data_dir(req);
3456 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
3459 total_bytes = bio_nbytes = 0;
3460 while ((bio = req->bio) != NULL) {
3464 * For an empty barrier request, the low level driver must
3465 * store a potential error location in ->sector. We pass
3466 * that back up in ->bi_sector.
3468 if (blk_empty_barrier(req))
3469 bio->bi_sector = req->sector;
3471 if (nr_bytes >= bio->bi_size) {
3472 req->bio = bio->bi_next;
3473 nbytes = bio->bi_size;
3474 req_bio_endio(req, bio, nbytes, error);
3478 int idx = bio->bi_idx + next_idx;
3480 if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3481 blk_dump_rq_flags(req, "__end_that");
3482 printk("%s: bio idx %d >= vcnt %d\n",
3484 bio->bi_idx, bio->bi_vcnt);
3488 nbytes = bio_iovec_idx(bio, idx)->bv_len;
3489 BIO_BUG_ON(nbytes > bio->bi_size);
3492 * not a complete bvec done
3494 if (unlikely(nbytes > nr_bytes)) {
3495 bio_nbytes += nr_bytes;
3496 total_bytes += nr_bytes;
3501 * advance to the next vector
3504 bio_nbytes += nbytes;
3507 total_bytes += nbytes;
3510 if ((bio = req->bio)) {
3512 * end more in this run, or just return 'not-done'
3514 if (unlikely(nr_bytes <= 0))
3526 * if the request wasn't completed, update state
3529 req_bio_endio(req, bio, bio_nbytes, error);
3530 bio->bi_idx += next_idx;
3531 bio_iovec(bio)->bv_offset += nr_bytes;
3532 bio_iovec(bio)->bv_len -= nr_bytes;
3535 blk_recalc_rq_sectors(req, total_bytes >> 9);
3536 blk_recalc_rq_segments(req);
3541 * end_that_request_first - end I/O on a request
3542 * @req: the request being processed
3543 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3544 * @nr_sectors: number of sectors to end I/O on
3547 * Ends I/O on a number of sectors attached to @req, and sets it up
3548 * for the next range of segments (if any) in the cluster.
3551 * 0 - we are done with this request, call end_that_request_last()
3552 * 1 - still buffers pending for this request
3554 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3556 return __end_that_request_first(req, uptodate, nr_sectors << 9);
3559 EXPORT_SYMBOL(end_that_request_first);
3562 * end_that_request_chunk - end I/O on a request
3563 * @req: the request being processed
3564 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3565 * @nr_bytes: number of bytes to complete
3568 * Ends I/O on a number of bytes attached to @req, and sets it up
3569 * for the next range of segments (if any). Like end_that_request_first(),
3570 * but deals with bytes instead of sectors.
3573 * 0 - we are done with this request, call end_that_request_last()
3574 * 1 - still buffers pending for this request
3576 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3578 return __end_that_request_first(req, uptodate, nr_bytes);
3581 EXPORT_SYMBOL(end_that_request_chunk);
3584 * splice the completion data to a local structure and hand off to
3585 * process_completion_queue() to complete the requests
3587 static void blk_done_softirq(struct softirq_action *h)
3589 struct list_head *cpu_list, local_list;
3591 local_irq_disable();
3592 cpu_list = &__get_cpu_var(blk_cpu_done);
3593 list_replace_init(cpu_list, &local_list);
3596 while (!list_empty(&local_list)) {
3597 struct request *rq = list_entry(local_list.next, struct request, donelist);
3599 list_del_init(&rq->donelist);
3600 rq->q->softirq_done_fn(rq);
3604 static int __cpuinit blk_cpu_notify(struct notifier_block *self, unsigned long action,
3608 * If a CPU goes away, splice its entries to the current CPU
3609 * and trigger a run of the softirq
3611 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
3612 int cpu = (unsigned long) hcpu;
3614 local_irq_disable();
3615 list_splice_init(&per_cpu(blk_cpu_done, cpu),
3616 &__get_cpu_var(blk_cpu_done));
3617 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3625 static struct notifier_block blk_cpu_notifier __cpuinitdata = {
3626 .notifier_call = blk_cpu_notify,
3630 * blk_complete_request - end I/O on a request
3631 * @req: the request being processed
3634 * Ends all I/O on a request. It does not handle partial completions,
3635 * unless the driver actually implements this in its completion callback
3636 * through requeueing. The actual completion happens out-of-order,
3637 * through a softirq handler. The user must have registered a completion
3638 * callback through blk_queue_softirq_done().
3641 void blk_complete_request(struct request *req)
3643 struct list_head *cpu_list;
3644 unsigned long flags;
3646 BUG_ON(!req->q->softirq_done_fn);
3648 local_irq_save(flags);
3650 cpu_list = &__get_cpu_var(blk_cpu_done);
3651 list_add_tail(&req->donelist, cpu_list);
3652 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3654 local_irq_restore(flags);
3657 EXPORT_SYMBOL(blk_complete_request);
3660 * queue lock must be held
3662 void end_that_request_last(struct request *req, int uptodate)
3664 struct gendisk *disk = req->rq_disk;
3668 * extend uptodate bool to allow < 0 value to be direct io error
3671 if (end_io_error(uptodate))
3672 error = !uptodate ? -EIO : uptodate;
3674 if (unlikely(laptop_mode) && blk_fs_request(req))
3675 laptop_io_completion();
3678 * Account IO completion. bar_rq isn't accounted as a normal
3679 * IO on queueing nor completion. Accounting the containing
3680 * request is enough.
3682 if (disk && blk_fs_request(req) && req != &req->q->bar_rq) {
3683 unsigned long duration = jiffies - req->start_time;
3684 const int rw = rq_data_dir(req);
3686 __disk_stat_inc(disk, ios[rw]);
3687 __disk_stat_add(disk, ticks[rw], duration);
3688 disk_round_stats(disk);
3692 req->end_io(req, error);
3694 __blk_put_request(req->q, req);
3697 EXPORT_SYMBOL(end_that_request_last);
3699 static inline void __end_request(struct request *rq, int uptodate,
3700 unsigned int nr_bytes, int dequeue)
3702 if (!end_that_request_chunk(rq, uptodate, nr_bytes)) {
3704 blkdev_dequeue_request(rq);
3705 add_disk_randomness(rq->rq_disk);
3706 end_that_request_last(rq, uptodate);
3710 static unsigned int rq_byte_size(struct request *rq)
3712 if (blk_fs_request(rq))
3713 return rq->hard_nr_sectors << 9;
3715 return rq->data_len;
3719 * end_queued_request - end all I/O on a queued request
3720 * @rq: the request being processed
3721 * @uptodate: error value or 0/1 uptodate flag
3724 * Ends all I/O on a request, and removes it from the block layer queues.
3725 * Not suitable for normal IO completion, unless the driver still has
3726 * the request attached to the block layer.
3729 void end_queued_request(struct request *rq, int uptodate)
3731 __end_request(rq, uptodate, rq_byte_size(rq), 1);
3733 EXPORT_SYMBOL(end_queued_request);
3736 * end_dequeued_request - end all I/O on a dequeued request
3737 * @rq: the request being processed
3738 * @uptodate: error value or 0/1 uptodate flag
3741 * Ends all I/O on a request. The request must already have been
3742 * dequeued using blkdev_dequeue_request(), as is normally the case
3746 void end_dequeued_request(struct request *rq, int uptodate)
3748 __end_request(rq, uptodate, rq_byte_size(rq), 0);
3750 EXPORT_SYMBOL(end_dequeued_request);
3754 * end_request - end I/O on the current segment of the request
3755 * @req: the request being processed
3756 * @uptodate: error value or 0/1 uptodate flag
3759 * Ends I/O on the current segment of a request. If that is the only
3760 * remaining segment, the request is also completed and freed.
3762 * This is a remnant of how older block drivers handled IO completions.
3763 * Modern drivers typically end IO on the full request in one go, unless
3764 * they have a residual value to account for. For that case this function
3765 * isn't really useful, unless the residual just happens to be the
3766 * full current segment. In other words, don't use this function in new
3767 * code. Either use end_request_completely(), or the
3768 * end_that_request_chunk() (along with end_that_request_last()) for
3769 * partial completions.
3772 void end_request(struct request *req, int uptodate)
3774 __end_request(req, uptodate, req->hard_cur_sectors << 9, 1);
3776 EXPORT_SYMBOL(end_request);
3778 static void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
3781 /* first two bits are identical in rq->cmd_flags and bio->bi_rw */
3782 rq->cmd_flags |= (bio->bi_rw & 3);
3784 rq->nr_phys_segments = bio_phys_segments(q, bio);
3785 rq->nr_hw_segments = bio_hw_segments(q, bio);
3786 rq->current_nr_sectors = bio_cur_sectors(bio);
3787 rq->hard_cur_sectors = rq->current_nr_sectors;
3788 rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3789 rq->buffer = bio_data(bio);
3790 rq->data_len = bio->bi_size;
3792 rq->bio = rq->biotail = bio;
3795 rq->rq_disk = bio->bi_bdev->bd_disk;
3798 int kblockd_schedule_work(struct work_struct *work)
3800 return queue_work(kblockd_workqueue, work);
3803 EXPORT_SYMBOL(kblockd_schedule_work);
3805 void kblockd_flush_work(struct work_struct *work)
3807 cancel_work_sync(work);
3809 EXPORT_SYMBOL(kblockd_flush_work);
3811 int __init blk_dev_init(void)
3815 kblockd_workqueue = create_workqueue("kblockd");
3816 if (!kblockd_workqueue)
3817 panic("Failed to create kblockd\n");
3819 request_cachep = kmem_cache_create("blkdev_requests",
3820 sizeof(struct request), 0, SLAB_PANIC, NULL);
3822 requestq_cachep = kmem_cache_create("blkdev_queue",
3823 sizeof(struct request_queue), 0, SLAB_PANIC, NULL);
3825 iocontext_cachep = kmem_cache_create("blkdev_ioc",
3826 sizeof(struct io_context), 0, SLAB_PANIC, NULL);
3828 for_each_possible_cpu(i)
3829 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3831 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
3832 register_hotcpu_notifier(&blk_cpu_notifier);
3834 blk_max_low_pfn = max_low_pfn - 1;
3835 blk_max_pfn = max_pfn - 1;
3841 * IO Context helper functions
3843 void put_io_context(struct io_context *ioc)
3848 BUG_ON(atomic_read(&ioc->refcount) == 0);
3850 if (atomic_dec_and_test(&ioc->refcount)) {
3851 struct cfq_io_context *cic;
3854 if (ioc->aic && ioc->aic->dtor)
3855 ioc->aic->dtor(ioc->aic);
3856 if (ioc->cic_root.rb_node != NULL) {
3857 struct rb_node *n = rb_first(&ioc->cic_root);
3859 cic = rb_entry(n, struct cfq_io_context, rb_node);
3864 kmem_cache_free(iocontext_cachep, ioc);
3867 EXPORT_SYMBOL(put_io_context);
3869 /* Called by the exitting task */
3870 void exit_io_context(void)
3872 struct io_context *ioc;
3873 struct cfq_io_context *cic;
3876 ioc = current->io_context;
3877 current->io_context = NULL;
3878 task_unlock(current);
3881 if (ioc->aic && ioc->aic->exit)
3882 ioc->aic->exit(ioc->aic);
3883 if (ioc->cic_root.rb_node != NULL) {
3884 cic = rb_entry(rb_first(&ioc->cic_root), struct cfq_io_context, rb_node);
3888 put_io_context(ioc);
3892 * If the current task has no IO context then create one and initialise it.
3893 * Otherwise, return its existing IO context.
3895 * This returned IO context doesn't have a specifically elevated refcount,
3896 * but since the current task itself holds a reference, the context can be
3897 * used in general code, so long as it stays within `current` context.
3899 static struct io_context *current_io_context(gfp_t gfp_flags, int node)
3901 struct task_struct *tsk = current;
3902 struct io_context *ret;
3904 ret = tsk->io_context;
3908 ret = kmem_cache_alloc_node(iocontext_cachep, gfp_flags, node);
3910 atomic_set(&ret->refcount, 1);
3911 ret->task = current;
3912 ret->ioprio_changed = 0;
3913 ret->last_waited = jiffies; /* doesn't matter... */
3914 ret->nr_batch_requests = 0; /* because this is 0 */
3916 ret->cic_root.rb_node = NULL;
3917 ret->ioc_data = NULL;
3918 /* make sure set_task_ioprio() sees the settings above */
3920 tsk->io_context = ret;
3927 * If the current task has no IO context then create one and initialise it.
3928 * If it does have a context, take a ref on it.
3930 * This is always called in the context of the task which submitted the I/O.
3932 struct io_context *get_io_context(gfp_t gfp_flags, int node)
3934 struct io_context *ret;
3935 ret = current_io_context(gfp_flags, node);
3937 atomic_inc(&ret->refcount);
3940 EXPORT_SYMBOL(get_io_context);
3942 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3944 struct io_context *src = *psrc;
3945 struct io_context *dst = *pdst;
3948 BUG_ON(atomic_read(&src->refcount) == 0);
3949 atomic_inc(&src->refcount);
3950 put_io_context(dst);
3954 EXPORT_SYMBOL(copy_io_context);
3956 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3958 struct io_context *temp;
3963 EXPORT_SYMBOL(swap_io_context);
3968 struct queue_sysfs_entry {
3969 struct attribute attr;
3970 ssize_t (*show)(struct request_queue *, char *);
3971 ssize_t (*store)(struct request_queue *, const char *, size_t);
3975 queue_var_show(unsigned int var, char *page)
3977 return sprintf(page, "%d\n", var);
3981 queue_var_store(unsigned long *var, const char *page, size_t count)
3983 char *p = (char *) page;
3985 *var = simple_strtoul(p, &p, 10);
3989 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3991 return queue_var_show(q->nr_requests, (page));
3995 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3997 struct request_list *rl = &q->rq;
3999 int ret = queue_var_store(&nr, page, count);
4000 if (nr < BLKDEV_MIN_RQ)
4003 spin_lock_irq(q->queue_lock);
4004 q->nr_requests = nr;
4005 blk_queue_congestion_threshold(q);
4007 if (rl->count[READ] >= queue_congestion_on_threshold(q))
4008 blk_set_queue_congested(q, READ);
4009 else if (rl->count[READ] < queue_congestion_off_threshold(q))
4010 blk_clear_queue_congested(q, READ);
4012 if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
4013 blk_set_queue_congested(q, WRITE);
4014 else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
4015 blk_clear_queue_congested(q, WRITE);
4017 if (rl->count[READ] >= q->nr_requests) {
4018 blk_set_queue_full(q, READ);
4019 } else if (rl->count[READ]+1 <= q->nr_requests) {
4020 blk_clear_queue_full(q, READ);
4021 wake_up(&rl->wait[READ]);
4024 if (rl->count[WRITE] >= q->nr_requests) {
4025 blk_set_queue_full(q, WRITE);
4026 } else if (rl->count[WRITE]+1 <= q->nr_requests) {
4027 blk_clear_queue_full(q, WRITE);
4028 wake_up(&rl->wait[WRITE]);
4030 spin_unlock_irq(q->queue_lock);
4034 static ssize_t queue_ra_show(struct request_queue *q, char *page)
4036 int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
4038 return queue_var_show(ra_kb, (page));
4042 queue_ra_store(struct request_queue *q, const char *page, size_t count)
4044 unsigned long ra_kb;
4045 ssize_t ret = queue_var_store(&ra_kb, page, count);
4047 spin_lock_irq(q->queue_lock);
4048 q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
4049 spin_unlock_irq(q->queue_lock);
4054 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
4056 int max_sectors_kb = q->max_sectors >> 1;
4058 return queue_var_show(max_sectors_kb, (page));
4062 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
4064 unsigned long max_sectors_kb,
4065 max_hw_sectors_kb = q->max_hw_sectors >> 1,
4066 page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
4067 ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
4069 if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
4072 * Take the queue lock to update the readahead and max_sectors
4073 * values synchronously:
4075 spin_lock_irq(q->queue_lock);
4076 q->max_sectors = max_sectors_kb << 1;
4077 spin_unlock_irq(q->queue_lock);
4082 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
4084 int max_hw_sectors_kb = q->max_hw_sectors >> 1;
4086 return queue_var_show(max_hw_sectors_kb, (page));
4089 static ssize_t queue_max_segments_show(struct request_queue *q, char *page)
4091 return queue_var_show(q->max_phys_segments, page);
4094 static ssize_t queue_max_segments_store(struct request_queue *q,
4095 const char *page, size_t count)
4097 unsigned long segments;
4098 ssize_t ret = queue_var_store(&segments, page, count);
4100 spin_lock_irq(q->queue_lock);
4101 q->max_phys_segments = segments;
4102 spin_unlock_irq(q->queue_lock);
4106 static struct queue_sysfs_entry queue_requests_entry = {
4107 .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
4108 .show = queue_requests_show,
4109 .store = queue_requests_store,
4112 static struct queue_sysfs_entry queue_ra_entry = {
4113 .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
4114 .show = queue_ra_show,
4115 .store = queue_ra_store,
4118 static struct queue_sysfs_entry queue_max_sectors_entry = {
4119 .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
4120 .show = queue_max_sectors_show,
4121 .store = queue_max_sectors_store,
4124 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
4125 .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
4126 .show = queue_max_hw_sectors_show,
4129 static struct queue_sysfs_entry queue_max_segments_entry = {
4130 .attr = {.name = "max_segments", .mode = S_IRUGO | S_IWUSR },
4131 .show = queue_max_segments_show,
4132 .store = queue_max_segments_store,
4135 static struct queue_sysfs_entry queue_iosched_entry = {
4136 .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
4137 .show = elv_iosched_show,
4138 .store = elv_iosched_store,
4141 static struct attribute *default_attrs[] = {
4142 &queue_requests_entry.attr,
4143 &queue_ra_entry.attr,
4144 &queue_max_hw_sectors_entry.attr,
4145 &queue_max_sectors_entry.attr,
4146 &queue_max_segments_entry.attr,
4147 &queue_iosched_entry.attr,
4151 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
4154 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
4156 struct queue_sysfs_entry *entry = to_queue(attr);
4157 struct request_queue *q =
4158 container_of(kobj, struct request_queue, kobj);
4163 mutex_lock(&q->sysfs_lock);
4164 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
4165 mutex_unlock(&q->sysfs_lock);
4168 res = entry->show(q, page);
4169 mutex_unlock(&q->sysfs_lock);
4174 queue_attr_store(struct kobject *kobj, struct attribute *attr,
4175 const char *page, size_t length)
4177 struct queue_sysfs_entry *entry = to_queue(attr);
4178 struct request_queue *q = container_of(kobj, struct request_queue, kobj);
4184 mutex_lock(&q->sysfs_lock);
4185 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
4186 mutex_unlock(&q->sysfs_lock);
4189 res = entry->store(q, page, length);
4190 mutex_unlock(&q->sysfs_lock);
4194 static struct sysfs_ops queue_sysfs_ops = {
4195 .show = queue_attr_show,
4196 .store = queue_attr_store,
4199 static struct kobj_type queue_ktype = {
4200 .sysfs_ops = &queue_sysfs_ops,
4201 .default_attrs = default_attrs,
4202 .release = blk_release_queue,
4205 int blk_register_queue(struct gendisk *disk)
4209 struct request_queue *q = disk->queue;
4211 if (!q || !q->request_fn)
4214 q->kobj.parent = kobject_get(&disk->kobj);
4216 ret = kobject_add(&q->kobj);
4220 kobject_uevent(&q->kobj, KOBJ_ADD);
4222 ret = elv_register_queue(q);
4224 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4225 kobject_del(&q->kobj);
4232 void blk_unregister_queue(struct gendisk *disk)
4234 struct request_queue *q = disk->queue;
4236 if (q && q->request_fn) {
4237 elv_unregister_queue(q);
4239 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4240 kobject_del(&q->kobj);
4241 kobject_put(&disk->kobj);