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>
37 #include <scsi/scsi_cmnd.h>
39 static void blk_unplug_work(struct work_struct *work);
40 static void blk_unplug_timeout(unsigned long data);
41 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io);
42 static void init_request_from_bio(struct request *req, struct bio *bio);
43 static int __make_request(struct request_queue *q, struct bio *bio);
44 static struct io_context *current_io_context(gfp_t gfp_flags, int node);
45 static void blk_recalc_rq_segments(struct request *rq);
46 static void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
50 * For the allocated request tables
52 static struct kmem_cache *request_cachep;
55 * For queue allocation
57 static struct kmem_cache *requestq_cachep;
60 * For io context allocations
62 static struct kmem_cache *iocontext_cachep;
65 * Controlling structure to kblockd
67 static struct workqueue_struct *kblockd_workqueue;
69 unsigned long blk_max_low_pfn, blk_max_pfn;
71 EXPORT_SYMBOL(blk_max_low_pfn);
72 EXPORT_SYMBOL(blk_max_pfn);
74 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
76 /* Amount of time in which a process may batch requests */
77 #define BLK_BATCH_TIME (HZ/50UL)
79 /* Number of requests a "batching" process may submit */
80 #define BLK_BATCH_REQ 32
83 * Return the threshold (number of used requests) at which the queue is
84 * considered to be congested. It include a little hysteresis to keep the
85 * context switch rate down.
87 static inline int queue_congestion_on_threshold(struct request_queue *q)
89 return q->nr_congestion_on;
93 * The threshold at which a queue is considered to be uncongested
95 static inline int queue_congestion_off_threshold(struct request_queue *q)
97 return q->nr_congestion_off;
100 static void blk_queue_congestion_threshold(struct request_queue *q)
104 nr = q->nr_requests - (q->nr_requests / 8) + 1;
105 if (nr > q->nr_requests)
107 q->nr_congestion_on = nr;
109 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
112 q->nr_congestion_off = nr;
116 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
119 * Locates the passed device's request queue and returns the address of its
122 * Will return NULL if the request queue cannot be located.
124 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
126 struct backing_dev_info *ret = NULL;
127 struct request_queue *q = bdev_get_queue(bdev);
130 ret = &q->backing_dev_info;
133 EXPORT_SYMBOL(blk_get_backing_dev_info);
136 * blk_queue_prep_rq - set a prepare_request function for queue
138 * @pfn: prepare_request function
140 * It's possible for a queue to register a prepare_request callback which
141 * is invoked before the request is handed to the request_fn. The goal of
142 * the function is to prepare a request for I/O, it can be used to build a
143 * cdb from the request data for instance.
146 void blk_queue_prep_rq(struct request_queue *q, prep_rq_fn *pfn)
151 EXPORT_SYMBOL(blk_queue_prep_rq);
154 * blk_queue_merge_bvec - set a merge_bvec function for queue
156 * @mbfn: merge_bvec_fn
158 * Usually queues have static limitations on the max sectors or segments that
159 * we can put in a request. Stacking drivers may have some settings that
160 * are dynamic, and thus we have to query the queue whether it is ok to
161 * add a new bio_vec to a bio at a given offset or not. If the block device
162 * has such limitations, it needs to register a merge_bvec_fn to control
163 * the size of bio's sent to it. Note that a block device *must* allow a
164 * single page to be added to an empty bio. The block device driver may want
165 * to use the bio_split() function to deal with these bio's. By default
166 * no merge_bvec_fn is defined for a queue, and only the fixed limits are
169 void blk_queue_merge_bvec(struct request_queue *q, merge_bvec_fn *mbfn)
171 q->merge_bvec_fn = mbfn;
174 EXPORT_SYMBOL(blk_queue_merge_bvec);
176 void blk_queue_softirq_done(struct request_queue *q, softirq_done_fn *fn)
178 q->softirq_done_fn = fn;
181 EXPORT_SYMBOL(blk_queue_softirq_done);
184 * blk_queue_make_request - define an alternate make_request function for a device
185 * @q: the request queue for the device to be affected
186 * @mfn: the alternate make_request function
189 * The normal way for &struct bios to be passed to a device
190 * driver is for them to be collected into requests on a request
191 * queue, and then to allow the device driver to select requests
192 * off that queue when it is ready. This works well for many block
193 * devices. However some block devices (typically virtual devices
194 * such as md or lvm) do not benefit from the processing on the
195 * request queue, and are served best by having the requests passed
196 * directly to them. This can be achieved by providing a function
197 * to blk_queue_make_request().
200 * The driver that does this *must* be able to deal appropriately
201 * with buffers in "highmemory". This can be accomplished by either calling
202 * __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
203 * blk_queue_bounce() to create a buffer in normal memory.
205 void blk_queue_make_request(struct request_queue * q, make_request_fn * mfn)
210 q->nr_requests = BLKDEV_MAX_RQ;
211 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
212 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
213 q->make_request_fn = mfn;
214 q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
215 q->backing_dev_info.state = 0;
216 q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
217 blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
218 blk_queue_hardsect_size(q, 512);
219 blk_queue_dma_alignment(q, 511);
220 blk_queue_congestion_threshold(q);
221 q->nr_batching = BLK_BATCH_REQ;
223 q->unplug_thresh = 4; /* hmm */
224 q->unplug_delay = (3 * HZ) / 1000; /* 3 milliseconds */
225 if (q->unplug_delay == 0)
228 INIT_WORK(&q->unplug_work, blk_unplug_work);
230 q->unplug_timer.function = blk_unplug_timeout;
231 q->unplug_timer.data = (unsigned long)q;
234 * by default assume old behaviour and bounce for any highmem page
236 blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
239 EXPORT_SYMBOL(blk_queue_make_request);
241 static void rq_init(struct request_queue *q, struct request *rq)
243 INIT_LIST_HEAD(&rq->queuelist);
244 INIT_LIST_HEAD(&rq->donelist);
247 rq->bio = rq->biotail = NULL;
248 INIT_HLIST_NODE(&rq->hash);
249 RB_CLEAR_NODE(&rq->rb_node);
257 rq->nr_phys_segments = 0;
260 rq->end_io_data = NULL;
261 rq->completion_data = NULL;
266 * blk_queue_ordered - does this queue support ordered writes
267 * @q: the request queue
268 * @ordered: one of QUEUE_ORDERED_*
269 * @prepare_flush_fn: rq setup helper for cache flush ordered writes
272 * For journalled file systems, doing ordered writes on a commit
273 * block instead of explicitly doing wait_on_buffer (which is bad
274 * for performance) can be a big win. Block drivers supporting this
275 * feature should call this function and indicate so.
278 int blk_queue_ordered(struct request_queue *q, unsigned ordered,
279 prepare_flush_fn *prepare_flush_fn)
281 if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
282 prepare_flush_fn == NULL) {
283 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
287 if (ordered != QUEUE_ORDERED_NONE &&
288 ordered != QUEUE_ORDERED_DRAIN &&
289 ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
290 ordered != QUEUE_ORDERED_DRAIN_FUA &&
291 ordered != QUEUE_ORDERED_TAG &&
292 ordered != QUEUE_ORDERED_TAG_FLUSH &&
293 ordered != QUEUE_ORDERED_TAG_FUA) {
294 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
298 q->ordered = ordered;
299 q->next_ordered = ordered;
300 q->prepare_flush_fn = prepare_flush_fn;
305 EXPORT_SYMBOL(blk_queue_ordered);
308 * blk_queue_issue_flush_fn - set function for issuing a flush
309 * @q: the request queue
310 * @iff: the function to be called issuing the flush
313 * If a driver supports issuing a flush command, the support is notified
314 * to the block layer by defining it through this call.
317 void blk_queue_issue_flush_fn(struct request_queue *q, issue_flush_fn *iff)
319 q->issue_flush_fn = iff;
322 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
325 * Cache flushing for ordered writes handling
327 inline unsigned blk_ordered_cur_seq(struct request_queue *q)
331 return 1 << ffz(q->ordseq);
334 unsigned blk_ordered_req_seq(struct request *rq)
336 struct request_queue *q = rq->q;
338 BUG_ON(q->ordseq == 0);
340 if (rq == &q->pre_flush_rq)
341 return QUEUE_ORDSEQ_PREFLUSH;
342 if (rq == &q->bar_rq)
343 return QUEUE_ORDSEQ_BAR;
344 if (rq == &q->post_flush_rq)
345 return QUEUE_ORDSEQ_POSTFLUSH;
348 * !fs requests don't need to follow barrier ordering. Always
349 * put them at the front. This fixes the following deadlock.
351 * http://thread.gmane.org/gmane.linux.kernel/537473
353 if (!blk_fs_request(rq))
354 return QUEUE_ORDSEQ_DRAIN;
356 if ((rq->cmd_flags & REQ_ORDERED_COLOR) ==
357 (q->orig_bar_rq->cmd_flags & REQ_ORDERED_COLOR))
358 return QUEUE_ORDSEQ_DRAIN;
360 return QUEUE_ORDSEQ_DONE;
363 void blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
368 if (error && !q->orderr)
371 BUG_ON(q->ordseq & seq);
374 if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
378 * Okay, sequence complete.
381 uptodate = q->orderr ? q->orderr : 1;
385 end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
386 end_that_request_last(rq, uptodate);
389 static void pre_flush_end_io(struct request *rq, int error)
391 elv_completed_request(rq->q, rq);
392 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
395 static void bar_end_io(struct request *rq, int error)
397 elv_completed_request(rq->q, rq);
398 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
401 static void post_flush_end_io(struct request *rq, int error)
403 elv_completed_request(rq->q, rq);
404 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
407 static void queue_flush(struct request_queue *q, unsigned which)
410 rq_end_io_fn *end_io;
412 if (which == QUEUE_ORDERED_PREFLUSH) {
413 rq = &q->pre_flush_rq;
414 end_io = pre_flush_end_io;
416 rq = &q->post_flush_rq;
417 end_io = post_flush_end_io;
420 rq->cmd_flags = REQ_HARDBARRIER;
422 rq->elevator_private = NULL;
423 rq->elevator_private2 = NULL;
424 rq->rq_disk = q->bar_rq.rq_disk;
426 q->prepare_flush_fn(q, rq);
428 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
431 static inline struct request *start_ordered(struct request_queue *q,
436 q->ordered = q->next_ordered;
437 q->ordseq |= QUEUE_ORDSEQ_STARTED;
440 * Prep proxy barrier request.
442 blkdev_dequeue_request(rq);
447 if (bio_data_dir(q->orig_bar_rq->bio) == WRITE)
448 rq->cmd_flags |= REQ_RW;
449 rq->cmd_flags |= q->ordered & QUEUE_ORDERED_FUA ? REQ_FUA : 0;
450 rq->elevator_private = NULL;
451 rq->elevator_private2 = NULL;
452 init_request_from_bio(rq, q->orig_bar_rq->bio);
453 rq->end_io = bar_end_io;
456 * Queue ordered sequence. As we stack them at the head, we
457 * need to queue in reverse order. Note that we rely on that
458 * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
459 * request gets inbetween ordered sequence.
461 if (q->ordered & QUEUE_ORDERED_POSTFLUSH)
462 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
464 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
466 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
468 if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
469 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
470 rq = &q->pre_flush_rq;
472 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
474 if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
475 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
482 int blk_do_ordered(struct request_queue *q, struct request **rqp)
484 struct request *rq = *rqp;
485 int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
491 if (q->next_ordered != QUEUE_ORDERED_NONE) {
492 *rqp = start_ordered(q, rq);
496 * This can happen when the queue switches to
497 * ORDERED_NONE while this request is on it.
499 blkdev_dequeue_request(rq);
500 end_that_request_first(rq, -EOPNOTSUPP,
501 rq->hard_nr_sectors);
502 end_that_request_last(rq, -EOPNOTSUPP);
509 * Ordered sequence in progress
512 /* Special requests are not subject to ordering rules. */
513 if (!blk_fs_request(rq) &&
514 rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
517 if (q->ordered & QUEUE_ORDERED_TAG) {
518 /* Ordered by tag. Blocking the next barrier is enough. */
519 if (is_barrier && rq != &q->bar_rq)
522 /* Ordered by draining. Wait for turn. */
523 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
524 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
531 static int flush_dry_bio_endio(struct bio *bio, unsigned int bytes, int error)
533 struct request_queue *q = bio->bi_private;
536 * This is dry run, restore bio_sector and size. We'll finish
537 * this request again with the original bi_end_io after an
538 * error occurs or post flush is complete.
546 set_bit(BIO_UPTODATE, &bio->bi_flags);
547 bio->bi_size = q->bi_size;
548 bio->bi_sector -= (q->bi_size >> 9);
554 static int ordered_bio_endio(struct request *rq, struct bio *bio,
555 unsigned int nbytes, int error)
557 struct request_queue *q = rq->q;
561 if (&q->bar_rq != rq)
565 * Okay, this is the barrier request in progress, dry finish it.
567 if (error && !q->orderr)
570 endio = bio->bi_end_io;
571 private = bio->bi_private;
572 bio->bi_end_io = flush_dry_bio_endio;
575 bio_endio(bio, nbytes, error);
577 bio->bi_end_io = endio;
578 bio->bi_private = private;
584 * blk_queue_bounce_limit - set bounce buffer limit for queue
585 * @q: the request queue for the device
586 * @dma_addr: bus address limit
589 * Different hardware can have different requirements as to what pages
590 * it can do I/O directly to. A low level driver can call
591 * blk_queue_bounce_limit to have lower memory pages allocated as bounce
592 * buffers for doing I/O to pages residing above @page.
594 void blk_queue_bounce_limit(struct request_queue *q, u64 dma_addr)
596 unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
599 q->bounce_gfp = GFP_NOIO;
600 #if BITS_PER_LONG == 64
601 /* Assume anything <= 4GB can be handled by IOMMU.
602 Actually some IOMMUs can handle everything, but I don't
603 know of a way to test this here. */
604 if (bounce_pfn < (min_t(u64,0xffffffff,BLK_BOUNCE_HIGH) >> PAGE_SHIFT))
606 q->bounce_pfn = max_low_pfn;
608 if (bounce_pfn < blk_max_low_pfn)
610 q->bounce_pfn = bounce_pfn;
613 init_emergency_isa_pool();
614 q->bounce_gfp = GFP_NOIO | GFP_DMA;
615 q->bounce_pfn = bounce_pfn;
619 EXPORT_SYMBOL(blk_queue_bounce_limit);
622 * blk_queue_max_sectors - set max sectors for a request for this queue
623 * @q: the request queue for the device
624 * @max_sectors: max sectors in the usual 512b unit
627 * Enables a low level driver to set an upper limit on the size of
630 void blk_queue_max_sectors(struct request_queue *q, unsigned int max_sectors)
632 if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
633 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
634 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
637 if (BLK_DEF_MAX_SECTORS > max_sectors)
638 q->max_hw_sectors = q->max_sectors = max_sectors;
640 q->max_sectors = BLK_DEF_MAX_SECTORS;
641 q->max_hw_sectors = max_sectors;
645 EXPORT_SYMBOL(blk_queue_max_sectors);
648 * blk_queue_max_phys_segments - set max phys segments for a request for this queue
649 * @q: the request queue for the device
650 * @max_segments: max number of segments
653 * Enables a low level driver to set an upper limit on the number of
654 * physical data segments in a request. This would be the largest sized
655 * scatter list the driver could handle.
657 void blk_queue_max_phys_segments(struct request_queue *q,
658 unsigned short max_segments)
662 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
665 q->max_phys_segments = max_segments;
668 EXPORT_SYMBOL(blk_queue_max_phys_segments);
671 * blk_queue_max_hw_segments - set max hw segments for a request for this queue
672 * @q: the request queue for the device
673 * @max_segments: max number of segments
676 * Enables a low level driver to set an upper limit on the number of
677 * hw data segments in a request. This would be the largest number of
678 * address/length pairs the host adapter can actually give as once
681 void blk_queue_max_hw_segments(struct request_queue *q,
682 unsigned short max_segments)
686 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
689 q->max_hw_segments = max_segments;
692 EXPORT_SYMBOL(blk_queue_max_hw_segments);
695 * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
696 * @q: the request queue for the device
697 * @max_size: max size of segment in bytes
700 * Enables a low level driver to set an upper limit on the size of a
703 void blk_queue_max_segment_size(struct request_queue *q, unsigned int max_size)
705 if (max_size < PAGE_CACHE_SIZE) {
706 max_size = PAGE_CACHE_SIZE;
707 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
710 q->max_segment_size = max_size;
713 EXPORT_SYMBOL(blk_queue_max_segment_size);
716 * blk_queue_hardsect_size - set hardware sector size for the queue
717 * @q: the request queue for the device
718 * @size: the hardware sector size, in bytes
721 * This should typically be set to the lowest possible sector size
722 * that the hardware can operate on (possible without reverting to
723 * even internal read-modify-write operations). Usually the default
724 * of 512 covers most hardware.
726 void blk_queue_hardsect_size(struct request_queue *q, unsigned short size)
728 q->hardsect_size = size;
731 EXPORT_SYMBOL(blk_queue_hardsect_size);
734 * Returns the minimum that is _not_ zero, unless both are zero.
736 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
739 * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
740 * @t: the stacking driver (top)
741 * @b: the underlying device (bottom)
743 void blk_queue_stack_limits(struct request_queue *t, struct request_queue *b)
745 /* zero is "infinity" */
746 t->max_sectors = min_not_zero(t->max_sectors,b->max_sectors);
747 t->max_hw_sectors = min_not_zero(t->max_hw_sectors,b->max_hw_sectors);
749 t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
750 t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
751 t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
752 t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
753 if (!test_bit(QUEUE_FLAG_CLUSTER, &b->queue_flags))
754 clear_bit(QUEUE_FLAG_CLUSTER, &t->queue_flags);
757 EXPORT_SYMBOL(blk_queue_stack_limits);
760 * blk_queue_segment_boundary - set boundary rules for segment merging
761 * @q: the request queue for the device
762 * @mask: the memory boundary mask
764 void blk_queue_segment_boundary(struct request_queue *q, unsigned long mask)
766 if (mask < PAGE_CACHE_SIZE - 1) {
767 mask = PAGE_CACHE_SIZE - 1;
768 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
771 q->seg_boundary_mask = mask;
774 EXPORT_SYMBOL(blk_queue_segment_boundary);
777 * blk_queue_dma_alignment - set dma length and memory alignment
778 * @q: the request queue for the device
779 * @mask: alignment mask
782 * set required memory and length aligment for direct dma transactions.
783 * this is used when buiding direct io requests for the queue.
786 void blk_queue_dma_alignment(struct request_queue *q, int mask)
788 q->dma_alignment = mask;
791 EXPORT_SYMBOL(blk_queue_dma_alignment);
794 * blk_queue_find_tag - find a request by its tag and queue
795 * @q: The request queue for the device
796 * @tag: The tag of the request
799 * Should be used when a device returns a tag and you want to match
802 * no locks need be held.
804 struct request *blk_queue_find_tag(struct request_queue *q, int tag)
806 return blk_map_queue_find_tag(q->queue_tags, tag);
809 EXPORT_SYMBOL(blk_queue_find_tag);
812 * __blk_free_tags - release a given set of tag maintenance info
813 * @bqt: the tag map to free
815 * Tries to free the specified @bqt@. Returns true if it was
816 * actually freed and false if there are still references using it
818 static int __blk_free_tags(struct blk_queue_tag *bqt)
822 retval = atomic_dec_and_test(&bqt->refcnt);
825 BUG_ON(!list_empty(&bqt->busy_list));
827 kfree(bqt->tag_index);
828 bqt->tag_index = NULL;
841 * __blk_queue_free_tags - release tag maintenance info
842 * @q: the request queue for the device
845 * blk_cleanup_queue() will take care of calling this function, if tagging
846 * has been used. So there's no need to call this directly.
848 static void __blk_queue_free_tags(struct request_queue *q)
850 struct blk_queue_tag *bqt = q->queue_tags;
855 __blk_free_tags(bqt);
857 q->queue_tags = NULL;
858 q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
863 * blk_free_tags - release a given set of tag maintenance info
864 * @bqt: the tag map to free
866 * For externally managed @bqt@ frees the map. Callers of this
867 * function must guarantee to have released all the queues that
868 * might have been using this tag map.
870 void blk_free_tags(struct blk_queue_tag *bqt)
872 if (unlikely(!__blk_free_tags(bqt)))
875 EXPORT_SYMBOL(blk_free_tags);
878 * blk_queue_free_tags - release tag maintenance info
879 * @q: the request queue for the device
882 * This is used to disabled tagged queuing to a device, yet leave
885 void blk_queue_free_tags(struct request_queue *q)
887 clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
890 EXPORT_SYMBOL(blk_queue_free_tags);
893 init_tag_map(struct request_queue *q, struct blk_queue_tag *tags, int depth)
895 struct request **tag_index;
896 unsigned long *tag_map;
899 if (q && depth > q->nr_requests * 2) {
900 depth = q->nr_requests * 2;
901 printk(KERN_ERR "%s: adjusted depth to %d\n",
902 __FUNCTION__, depth);
905 tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC);
909 nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
910 tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
914 tags->real_max_depth = depth;
915 tags->max_depth = depth;
916 tags->tag_index = tag_index;
917 tags->tag_map = tag_map;
925 static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q,
928 struct blk_queue_tag *tags;
930 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
934 if (init_tag_map(q, tags, depth))
937 INIT_LIST_HEAD(&tags->busy_list);
939 atomic_set(&tags->refcnt, 1);
947 * blk_init_tags - initialize the tag info for an external tag map
948 * @depth: the maximum queue depth supported
949 * @tags: the tag to use
951 struct blk_queue_tag *blk_init_tags(int depth)
953 return __blk_queue_init_tags(NULL, depth);
955 EXPORT_SYMBOL(blk_init_tags);
958 * blk_queue_init_tags - initialize the queue tag info
959 * @q: the request queue for the device
960 * @depth: the maximum queue depth supported
961 * @tags: the tag to use
963 int blk_queue_init_tags(struct request_queue *q, int depth,
964 struct blk_queue_tag *tags)
968 BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
970 if (!tags && !q->queue_tags) {
971 tags = __blk_queue_init_tags(q, depth);
975 } else if (q->queue_tags) {
976 if ((rc = blk_queue_resize_tags(q, depth)))
978 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
981 atomic_inc(&tags->refcnt);
984 * assign it, all done
986 q->queue_tags = tags;
987 q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
994 EXPORT_SYMBOL(blk_queue_init_tags);
997 * blk_queue_resize_tags - change the queueing depth
998 * @q: the request queue for the device
999 * @new_depth: the new max command queueing depth
1002 * Must be called with the queue lock held.
1004 int blk_queue_resize_tags(struct request_queue *q, int new_depth)
1006 struct blk_queue_tag *bqt = q->queue_tags;
1007 struct request **tag_index;
1008 unsigned long *tag_map;
1009 int max_depth, nr_ulongs;
1015 * if we already have large enough real_max_depth. just
1016 * adjust max_depth. *NOTE* as requests with tag value
1017 * between new_depth and real_max_depth can be in-flight, tag
1018 * map can not be shrunk blindly here.
1020 if (new_depth <= bqt->real_max_depth) {
1021 bqt->max_depth = new_depth;
1026 * Currently cannot replace a shared tag map with a new
1027 * one, so error out if this is the case
1029 if (atomic_read(&bqt->refcnt) != 1)
1033 * save the old state info, so we can copy it back
1035 tag_index = bqt->tag_index;
1036 tag_map = bqt->tag_map;
1037 max_depth = bqt->real_max_depth;
1039 if (init_tag_map(q, bqt, new_depth))
1042 memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
1043 nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
1044 memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
1051 EXPORT_SYMBOL(blk_queue_resize_tags);
1054 * blk_queue_end_tag - end tag operations for a request
1055 * @q: the request queue for the device
1056 * @rq: the request that has completed
1059 * Typically called when end_that_request_first() returns 0, meaning
1060 * all transfers have been done for a request. It's important to call
1061 * this function before end_that_request_last(), as that will put the
1062 * request back on the free list thus corrupting the internal tag list.
1065 * queue lock must be held.
1067 void blk_queue_end_tag(struct request_queue *q, struct request *rq)
1069 struct blk_queue_tag *bqt = q->queue_tags;
1074 if (unlikely(tag >= bqt->real_max_depth))
1076 * This can happen after tag depth has been reduced.
1077 * FIXME: how about a warning or info message here?
1081 list_del_init(&rq->queuelist);
1082 rq->cmd_flags &= ~REQ_QUEUED;
1085 if (unlikely(bqt->tag_index[tag] == NULL))
1086 printk(KERN_ERR "%s: tag %d is missing\n",
1089 bqt->tag_index[tag] = NULL;
1092 * We use test_and_clear_bit's memory ordering properties here.
1093 * The tag_map bit acts as a lock for tag_index[bit], so we need
1094 * a barrer before clearing the bit (precisely: release semantics).
1095 * Could use clear_bit_unlock when it is merged.
1097 if (unlikely(!test_and_clear_bit(tag, bqt->tag_map))) {
1098 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
1106 EXPORT_SYMBOL(blk_queue_end_tag);
1109 * blk_queue_start_tag - find a free tag and assign it
1110 * @q: the request queue for the device
1111 * @rq: the block request that needs tagging
1114 * This can either be used as a stand-alone helper, or possibly be
1115 * assigned as the queue &prep_rq_fn (in which case &struct request
1116 * automagically gets a tag assigned). Note that this function
1117 * assumes that any type of request can be queued! if this is not
1118 * true for your device, you must check the request type before
1119 * calling this function. The request will also be removed from
1120 * the request queue, so it's the drivers responsibility to readd
1121 * it if it should need to be restarted for some reason.
1124 * queue lock must be held.
1126 int blk_queue_start_tag(struct request_queue *q, struct request *rq)
1128 struct blk_queue_tag *bqt = q->queue_tags;
1131 if (unlikely((rq->cmd_flags & REQ_QUEUED))) {
1133 "%s: request %p for device [%s] already tagged %d",
1135 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
1140 * Protect against shared tag maps, as we may not have exclusive
1141 * access to the tag map.
1144 tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
1145 if (tag >= bqt->max_depth)
1148 } while (test_and_set_bit(tag, bqt->tag_map));
1150 * We rely on test_and_set_bit providing lock memory ordering semantics
1151 * (could use test_and_set_bit_lock when it is merged).
1154 rq->cmd_flags |= REQ_QUEUED;
1156 bqt->tag_index[tag] = rq;
1157 blkdev_dequeue_request(rq);
1158 list_add(&rq->queuelist, &bqt->busy_list);
1163 EXPORT_SYMBOL(blk_queue_start_tag);
1166 * blk_queue_invalidate_tags - invalidate all pending tags
1167 * @q: the request queue for the device
1170 * Hardware conditions may dictate a need to stop all pending requests.
1171 * In this case, we will safely clear the block side of the tag queue and
1172 * readd all requests to the request queue in the right order.
1175 * queue lock must be held.
1177 void blk_queue_invalidate_tags(struct request_queue *q)
1179 struct blk_queue_tag *bqt = q->queue_tags;
1180 struct list_head *tmp, *n;
1183 list_for_each_safe(tmp, n, &bqt->busy_list) {
1184 rq = list_entry_rq(tmp);
1186 if (rq->tag == -1) {
1188 "%s: bad tag found on list\n", __FUNCTION__);
1189 list_del_init(&rq->queuelist);
1190 rq->cmd_flags &= ~REQ_QUEUED;
1192 blk_queue_end_tag(q, rq);
1194 rq->cmd_flags &= ~REQ_STARTED;
1195 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1199 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1201 void blk_dump_rq_flags(struct request *rq, char *msg)
1205 printk("%s: dev %s: type=%x, flags=%x\n", msg,
1206 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->cmd_type,
1209 printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1211 rq->current_nr_sectors);
1212 printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1214 if (blk_pc_request(rq)) {
1216 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1217 printk("%02x ", rq->cmd[bit]);
1222 EXPORT_SYMBOL(blk_dump_rq_flags);
1224 void blk_recount_segments(struct request_queue *q, struct bio *bio)
1227 struct bio *nxt = bio->bi_next;
1229 rq.bio = rq.biotail = bio;
1230 bio->bi_next = NULL;
1231 blk_recalc_rq_segments(&rq);
1233 bio->bi_phys_segments = rq.nr_phys_segments;
1234 bio->bi_hw_segments = rq.nr_hw_segments;
1235 bio->bi_flags |= (1 << BIO_SEG_VALID);
1237 EXPORT_SYMBOL(blk_recount_segments);
1239 static void blk_recalc_rq_segments(struct request *rq)
1243 unsigned int phys_size;
1244 unsigned int hw_size;
1245 struct bio_vec *bv, *bvprv = NULL;
1249 struct req_iterator iter;
1250 int high, highprv = 1;
1251 struct request_queue *q = rq->q;
1256 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1257 hw_seg_size = seg_size = 0;
1258 phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
1259 rq_for_each_segment(bv, rq, iter) {
1261 * the trick here is making sure that a high page is never
1262 * considered part of another segment, since that might
1263 * change with the bounce page.
1265 high = page_to_pfn(bv->bv_page) > q->bounce_pfn;
1266 if (high || highprv)
1267 goto new_hw_segment;
1269 if (seg_size + bv->bv_len > q->max_segment_size)
1271 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1273 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1275 if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1276 goto new_hw_segment;
1278 seg_size += bv->bv_len;
1279 hw_seg_size += bv->bv_len;
1284 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1285 !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1286 hw_seg_size += bv->bv_len;
1289 if (nr_hw_segs == 1 &&
1290 hw_seg_size > rq->bio->bi_hw_front_size)
1291 rq->bio->bi_hw_front_size = hw_seg_size;
1292 hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1298 seg_size = bv->bv_len;
1302 if (nr_hw_segs == 1 &&
1303 hw_seg_size > rq->bio->bi_hw_front_size)
1304 rq->bio->bi_hw_front_size = hw_seg_size;
1305 if (hw_seg_size > rq->biotail->bi_hw_back_size)
1306 rq->biotail->bi_hw_back_size = hw_seg_size;
1307 rq->nr_phys_segments = nr_phys_segs;
1308 rq->nr_hw_segments = nr_hw_segs;
1311 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
1314 if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1317 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1319 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1323 * bio and nxt are contigous in memory, check if the queue allows
1324 * these two to be merged into one
1326 if (BIO_SEG_BOUNDARY(q, bio, nxt))
1332 static int blk_hw_contig_segment(struct request_queue *q, struct bio *bio,
1335 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1336 blk_recount_segments(q, bio);
1337 if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1338 blk_recount_segments(q, nxt);
1339 if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1340 BIOVEC_VIRT_OVERSIZE(bio->bi_hw_back_size + nxt->bi_hw_front_size))
1342 if (bio->bi_hw_back_size + nxt->bi_hw_front_size > q->max_segment_size)
1349 * map a request to scatterlist, return number of sg entries setup. Caller
1350 * must make sure sg can hold rq->nr_phys_segments entries
1352 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
1353 struct scatterlist *sg)
1355 struct bio_vec *bvec, *bvprv;
1356 struct req_iterator iter;
1360 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1363 * for each bio in rq
1366 rq_for_each_segment(bvec, rq, iter) {
1367 int nbytes = bvec->bv_len;
1369 if (bvprv && cluster) {
1370 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1373 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1375 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1378 sg[nsegs - 1].length += nbytes;
1381 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1382 sg[nsegs].page = bvec->bv_page;
1383 sg[nsegs].length = nbytes;
1384 sg[nsegs].offset = bvec->bv_offset;
1389 } /* segments in rq */
1394 EXPORT_SYMBOL(blk_rq_map_sg);
1397 * the standard queue merge functions, can be overridden with device
1398 * specific ones if so desired
1401 static inline int ll_new_mergeable(struct request_queue *q,
1402 struct request *req,
1405 int nr_phys_segs = bio_phys_segments(q, bio);
1407 if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1408 req->cmd_flags |= REQ_NOMERGE;
1409 if (req == q->last_merge)
1410 q->last_merge = NULL;
1415 * A hw segment is just getting larger, bump just the phys
1418 req->nr_phys_segments += nr_phys_segs;
1422 static inline int ll_new_hw_segment(struct request_queue *q,
1423 struct request *req,
1426 int nr_hw_segs = bio_hw_segments(q, bio);
1427 int nr_phys_segs = bio_phys_segments(q, bio);
1429 if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1430 || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1431 req->cmd_flags |= REQ_NOMERGE;
1432 if (req == q->last_merge)
1433 q->last_merge = NULL;
1438 * This will form the start of a new hw segment. Bump both
1441 req->nr_hw_segments += nr_hw_segs;
1442 req->nr_phys_segments += nr_phys_segs;
1446 static int ll_back_merge_fn(struct request_queue *q, struct request *req,
1449 unsigned short max_sectors;
1452 if (unlikely(blk_pc_request(req)))
1453 max_sectors = q->max_hw_sectors;
1455 max_sectors = q->max_sectors;
1457 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1458 req->cmd_flags |= REQ_NOMERGE;
1459 if (req == q->last_merge)
1460 q->last_merge = NULL;
1463 if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1464 blk_recount_segments(q, req->biotail);
1465 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1466 blk_recount_segments(q, bio);
1467 len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1468 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1469 !BIOVEC_VIRT_OVERSIZE(len)) {
1470 int mergeable = ll_new_mergeable(q, req, bio);
1473 if (req->nr_hw_segments == 1)
1474 req->bio->bi_hw_front_size = len;
1475 if (bio->bi_hw_segments == 1)
1476 bio->bi_hw_back_size = len;
1481 return ll_new_hw_segment(q, req, bio);
1484 static int ll_front_merge_fn(struct request_queue *q, struct request *req,
1487 unsigned short max_sectors;
1490 if (unlikely(blk_pc_request(req)))
1491 max_sectors = q->max_hw_sectors;
1493 max_sectors = q->max_sectors;
1496 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1497 req->cmd_flags |= REQ_NOMERGE;
1498 if (req == q->last_merge)
1499 q->last_merge = NULL;
1502 len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1503 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1504 blk_recount_segments(q, bio);
1505 if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1506 blk_recount_segments(q, req->bio);
1507 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1508 !BIOVEC_VIRT_OVERSIZE(len)) {
1509 int mergeable = ll_new_mergeable(q, req, bio);
1512 if (bio->bi_hw_segments == 1)
1513 bio->bi_hw_front_size = len;
1514 if (req->nr_hw_segments == 1)
1515 req->biotail->bi_hw_back_size = len;
1520 return ll_new_hw_segment(q, req, bio);
1523 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
1524 struct request *next)
1526 int total_phys_segments;
1527 int total_hw_segments;
1530 * First check if the either of the requests are re-queued
1531 * requests. Can't merge them if they are.
1533 if (req->special || next->special)
1537 * Will it become too large?
1539 if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1542 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1543 if (blk_phys_contig_segment(q, req->biotail, next->bio))
1544 total_phys_segments--;
1546 if (total_phys_segments > q->max_phys_segments)
1549 total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1550 if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1551 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1553 * propagate the combined length to the end of the requests
1555 if (req->nr_hw_segments == 1)
1556 req->bio->bi_hw_front_size = len;
1557 if (next->nr_hw_segments == 1)
1558 next->biotail->bi_hw_back_size = len;
1559 total_hw_segments--;
1562 if (total_hw_segments > q->max_hw_segments)
1565 /* Merge is OK... */
1566 req->nr_phys_segments = total_phys_segments;
1567 req->nr_hw_segments = total_hw_segments;
1572 * "plug" the device if there are no outstanding requests: this will
1573 * force the transfer to start only after we have put all the requests
1576 * This is called with interrupts off and no requests on the queue and
1577 * with the queue lock held.
1579 void blk_plug_device(struct request_queue *q)
1581 WARN_ON(!irqs_disabled());
1584 * don't plug a stopped queue, it must be paired with blk_start_queue()
1585 * which will restart the queueing
1587 if (blk_queue_stopped(q))
1590 if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
1591 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1592 blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
1596 EXPORT_SYMBOL(blk_plug_device);
1599 * remove the queue from the plugged list, if present. called with
1600 * queue lock held and interrupts disabled.
1602 int blk_remove_plug(struct request_queue *q)
1604 WARN_ON(!irqs_disabled());
1606 if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1609 del_timer(&q->unplug_timer);
1613 EXPORT_SYMBOL(blk_remove_plug);
1616 * remove the plug and let it rip..
1618 void __generic_unplug_device(struct request_queue *q)
1620 if (unlikely(blk_queue_stopped(q)))
1623 if (!blk_remove_plug(q))
1628 EXPORT_SYMBOL(__generic_unplug_device);
1631 * generic_unplug_device - fire a request queue
1632 * @q: The &struct request_queue in question
1635 * Linux uses plugging to build bigger requests queues before letting
1636 * the device have at them. If a queue is plugged, the I/O scheduler
1637 * is still adding and merging requests on the queue. Once the queue
1638 * gets unplugged, the request_fn defined for the queue is invoked and
1639 * transfers started.
1641 void generic_unplug_device(struct request_queue *q)
1643 spin_lock_irq(q->queue_lock);
1644 __generic_unplug_device(q);
1645 spin_unlock_irq(q->queue_lock);
1647 EXPORT_SYMBOL(generic_unplug_device);
1649 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1652 struct request_queue *q = bdi->unplug_io_data;
1655 * devices don't necessarily have an ->unplug_fn defined
1658 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1659 q->rq.count[READ] + q->rq.count[WRITE]);
1665 static void blk_unplug_work(struct work_struct *work)
1667 struct request_queue *q =
1668 container_of(work, struct request_queue, unplug_work);
1670 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1671 q->rq.count[READ] + q->rq.count[WRITE]);
1676 static void blk_unplug_timeout(unsigned long data)
1678 struct request_queue *q = (struct request_queue *)data;
1680 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
1681 q->rq.count[READ] + q->rq.count[WRITE]);
1683 kblockd_schedule_work(&q->unplug_work);
1687 * blk_start_queue - restart a previously stopped queue
1688 * @q: The &struct request_queue in question
1691 * blk_start_queue() will clear the stop flag on the queue, and call
1692 * the request_fn for the queue if it was in a stopped state when
1693 * entered. Also see blk_stop_queue(). Queue lock must be held.
1695 void blk_start_queue(struct request_queue *q)
1697 WARN_ON(!irqs_disabled());
1699 clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1702 * one level of recursion is ok and is much faster than kicking
1703 * the unplug handling
1705 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1707 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1710 kblockd_schedule_work(&q->unplug_work);
1714 EXPORT_SYMBOL(blk_start_queue);
1717 * blk_stop_queue - stop a queue
1718 * @q: The &struct request_queue in question
1721 * The Linux block layer assumes that a block driver will consume all
1722 * entries on the request queue when the request_fn strategy is called.
1723 * Often this will not happen, because of hardware limitations (queue
1724 * depth settings). If a device driver gets a 'queue full' response,
1725 * or if it simply chooses not to queue more I/O at one point, it can
1726 * call this function to prevent the request_fn from being called until
1727 * the driver has signalled it's ready to go again. This happens by calling
1728 * blk_start_queue() to restart queue operations. Queue lock must be held.
1730 void blk_stop_queue(struct request_queue *q)
1733 set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1735 EXPORT_SYMBOL(blk_stop_queue);
1738 * blk_sync_queue - cancel any pending callbacks on a queue
1742 * The block layer may perform asynchronous callback activity
1743 * on a queue, such as calling the unplug function after a timeout.
1744 * A block device may call blk_sync_queue to ensure that any
1745 * such activity is cancelled, thus allowing it to release resources
1746 * that the callbacks might use. The caller must already have made sure
1747 * that its ->make_request_fn will not re-add plugging prior to calling
1751 void blk_sync_queue(struct request_queue *q)
1753 del_timer_sync(&q->unplug_timer);
1755 EXPORT_SYMBOL(blk_sync_queue);
1758 * blk_run_queue - run a single device queue
1759 * @q: The queue to run
1761 void blk_run_queue(struct request_queue *q)
1763 unsigned long flags;
1765 spin_lock_irqsave(q->queue_lock, flags);
1769 * Only recurse once to avoid overrunning the stack, let the unplug
1770 * handling reinvoke the handler shortly if we already got there.
1772 if (!elv_queue_empty(q)) {
1773 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1775 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1778 kblockd_schedule_work(&q->unplug_work);
1782 spin_unlock_irqrestore(q->queue_lock, flags);
1784 EXPORT_SYMBOL(blk_run_queue);
1787 * blk_cleanup_queue: - release a &struct request_queue when it is no longer needed
1788 * @kobj: the kobj belonging of the request queue to be released
1791 * blk_cleanup_queue is the pair to blk_init_queue() or
1792 * blk_queue_make_request(). It should be called when a request queue is
1793 * being released; typically when a block device is being de-registered.
1794 * Currently, its primary task it to free all the &struct request
1795 * structures that were allocated to the queue and the queue itself.
1798 * Hopefully the low level driver will have finished any
1799 * outstanding requests first...
1801 static void blk_release_queue(struct kobject *kobj)
1803 struct request_queue *q =
1804 container_of(kobj, struct request_queue, kobj);
1805 struct request_list *rl = &q->rq;
1810 mempool_destroy(rl->rq_pool);
1813 __blk_queue_free_tags(q);
1815 blk_trace_shutdown(q);
1817 kmem_cache_free(requestq_cachep, q);
1820 void blk_put_queue(struct request_queue *q)
1822 kobject_put(&q->kobj);
1824 EXPORT_SYMBOL(blk_put_queue);
1826 void blk_cleanup_queue(struct request_queue * q)
1828 mutex_lock(&q->sysfs_lock);
1829 set_bit(QUEUE_FLAG_DEAD, &q->queue_flags);
1830 mutex_unlock(&q->sysfs_lock);
1833 elevator_exit(q->elevator);
1838 EXPORT_SYMBOL(blk_cleanup_queue);
1840 static int blk_init_free_list(struct request_queue *q)
1842 struct request_list *rl = &q->rq;
1844 rl->count[READ] = rl->count[WRITE] = 0;
1845 rl->starved[READ] = rl->starved[WRITE] = 0;
1847 init_waitqueue_head(&rl->wait[READ]);
1848 init_waitqueue_head(&rl->wait[WRITE]);
1850 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1851 mempool_free_slab, request_cachep, q->node);
1859 struct request_queue *blk_alloc_queue(gfp_t gfp_mask)
1861 return blk_alloc_queue_node(gfp_mask, -1);
1863 EXPORT_SYMBOL(blk_alloc_queue);
1865 static struct kobj_type queue_ktype;
1867 struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
1869 struct request_queue *q;
1871 q = kmem_cache_alloc_node(requestq_cachep,
1872 gfp_mask | __GFP_ZERO, node_id);
1876 init_timer(&q->unplug_timer);
1878 snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
1879 q->kobj.ktype = &queue_ktype;
1880 kobject_init(&q->kobj);
1882 q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1883 q->backing_dev_info.unplug_io_data = q;
1885 mutex_init(&q->sysfs_lock);
1889 EXPORT_SYMBOL(blk_alloc_queue_node);
1892 * blk_init_queue - prepare a request queue for use with a block device
1893 * @rfn: The function to be called to process requests that have been
1894 * placed on the queue.
1895 * @lock: Request queue spin lock
1898 * If a block device wishes to use the standard request handling procedures,
1899 * which sorts requests and coalesces adjacent requests, then it must
1900 * call blk_init_queue(). The function @rfn will be called when there
1901 * are requests on the queue that need to be processed. If the device
1902 * supports plugging, then @rfn may not be called immediately when requests
1903 * are available on the queue, but may be called at some time later instead.
1904 * Plugged queues are generally unplugged when a buffer belonging to one
1905 * of the requests on the queue is needed, or due to memory pressure.
1907 * @rfn is not required, or even expected, to remove all requests off the
1908 * queue, but only as many as it can handle at a time. If it does leave
1909 * requests on the queue, it is responsible for arranging that the requests
1910 * get dealt with eventually.
1912 * The queue spin lock must be held while manipulating the requests on the
1913 * request queue; this lock will be taken also from interrupt context, so irq
1914 * disabling is needed for it.
1916 * Function returns a pointer to the initialized request queue, or NULL if
1917 * it didn't succeed.
1920 * blk_init_queue() must be paired with a blk_cleanup_queue() call
1921 * when the block device is deactivated (such as at module unload).
1924 struct request_queue *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1926 return blk_init_queue_node(rfn, lock, -1);
1928 EXPORT_SYMBOL(blk_init_queue);
1930 struct request_queue *
1931 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1933 struct request_queue *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1939 if (blk_init_free_list(q)) {
1940 kmem_cache_free(requestq_cachep, q);
1945 * if caller didn't supply a lock, they get per-queue locking with
1949 spin_lock_init(&q->__queue_lock);
1950 lock = &q->__queue_lock;
1953 q->request_fn = rfn;
1954 q->prep_rq_fn = NULL;
1955 q->unplug_fn = generic_unplug_device;
1956 q->queue_flags = (1 << QUEUE_FLAG_CLUSTER);
1957 q->queue_lock = lock;
1959 blk_queue_segment_boundary(q, 0xffffffff);
1961 blk_queue_make_request(q, __make_request);
1962 blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1964 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1965 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1967 q->sg_reserved_size = INT_MAX;
1972 if (!elevator_init(q, NULL)) {
1973 blk_queue_congestion_threshold(q);
1980 EXPORT_SYMBOL(blk_init_queue_node);
1982 int blk_get_queue(struct request_queue *q)
1984 if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1985 kobject_get(&q->kobj);
1992 EXPORT_SYMBOL(blk_get_queue);
1994 static inline void blk_free_request(struct request_queue *q, struct request *rq)
1996 if (rq->cmd_flags & REQ_ELVPRIV)
1997 elv_put_request(q, rq);
1998 mempool_free(rq, q->rq.rq_pool);
2001 static struct request *
2002 blk_alloc_request(struct request_queue *q, int rw, int priv, gfp_t gfp_mask)
2004 struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
2010 * first three bits are identical in rq->cmd_flags and bio->bi_rw,
2011 * see bio.h and blkdev.h
2013 rq->cmd_flags = rw | REQ_ALLOCED;
2016 if (unlikely(elv_set_request(q, rq, gfp_mask))) {
2017 mempool_free(rq, q->rq.rq_pool);
2020 rq->cmd_flags |= REQ_ELVPRIV;
2027 * ioc_batching returns true if the ioc is a valid batching request and
2028 * should be given priority access to a request.
2030 static inline int ioc_batching(struct request_queue *q, struct io_context *ioc)
2036 * Make sure the process is able to allocate at least 1 request
2037 * even if the batch times out, otherwise we could theoretically
2040 return ioc->nr_batch_requests == q->nr_batching ||
2041 (ioc->nr_batch_requests > 0
2042 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
2046 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
2047 * will cause the process to be a "batcher" on all queues in the system. This
2048 * is the behaviour we want though - once it gets a wakeup it should be given
2051 static void ioc_set_batching(struct request_queue *q, struct io_context *ioc)
2053 if (!ioc || ioc_batching(q, ioc))
2056 ioc->nr_batch_requests = q->nr_batching;
2057 ioc->last_waited = jiffies;
2060 static void __freed_request(struct request_queue *q, int rw)
2062 struct request_list *rl = &q->rq;
2064 if (rl->count[rw] < queue_congestion_off_threshold(q))
2065 blk_clear_queue_congested(q, rw);
2067 if (rl->count[rw] + 1 <= q->nr_requests) {
2068 if (waitqueue_active(&rl->wait[rw]))
2069 wake_up(&rl->wait[rw]);
2071 blk_clear_queue_full(q, rw);
2076 * A request has just been released. Account for it, update the full and
2077 * congestion status, wake up any waiters. Called under q->queue_lock.
2079 static void freed_request(struct request_queue *q, int rw, int priv)
2081 struct request_list *rl = &q->rq;
2087 __freed_request(q, rw);
2089 if (unlikely(rl->starved[rw ^ 1]))
2090 __freed_request(q, rw ^ 1);
2093 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2095 * Get a free request, queue_lock must be held.
2096 * Returns NULL on failure, with queue_lock held.
2097 * Returns !NULL on success, with queue_lock *not held*.
2099 static struct request *get_request(struct request_queue *q, int rw_flags,
2100 struct bio *bio, gfp_t gfp_mask)
2102 struct request *rq = NULL;
2103 struct request_list *rl = &q->rq;
2104 struct io_context *ioc = NULL;
2105 const int rw = rw_flags & 0x01;
2106 int may_queue, priv;
2108 may_queue = elv_may_queue(q, rw_flags);
2109 if (may_queue == ELV_MQUEUE_NO)
2112 if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
2113 if (rl->count[rw]+1 >= q->nr_requests) {
2114 ioc = current_io_context(GFP_ATOMIC, q->node);
2116 * The queue will fill after this allocation, so set
2117 * it as full, and mark this process as "batching".
2118 * This process will be allowed to complete a batch of
2119 * requests, others will be blocked.
2121 if (!blk_queue_full(q, rw)) {
2122 ioc_set_batching(q, ioc);
2123 blk_set_queue_full(q, rw);
2125 if (may_queue != ELV_MQUEUE_MUST
2126 && !ioc_batching(q, ioc)) {
2128 * The queue is full and the allocating
2129 * process is not a "batcher", and not
2130 * exempted by the IO scheduler
2136 blk_set_queue_congested(q, rw);
2140 * Only allow batching queuers to allocate up to 50% over the defined
2141 * limit of requests, otherwise we could have thousands of requests
2142 * allocated with any setting of ->nr_requests
2144 if (rl->count[rw] >= (3 * q->nr_requests / 2))
2148 rl->starved[rw] = 0;
2150 priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
2154 spin_unlock_irq(q->queue_lock);
2156 rq = blk_alloc_request(q, rw_flags, priv, gfp_mask);
2157 if (unlikely(!rq)) {
2159 * Allocation failed presumably due to memory. Undo anything
2160 * we might have messed up.
2162 * Allocating task should really be put onto the front of the
2163 * wait queue, but this is pretty rare.
2165 spin_lock_irq(q->queue_lock);
2166 freed_request(q, rw, priv);
2169 * in the very unlikely event that allocation failed and no
2170 * requests for this direction was pending, mark us starved
2171 * so that freeing of a request in the other direction will
2172 * notice us. another possible fix would be to split the
2173 * rq mempool into READ and WRITE
2176 if (unlikely(rl->count[rw] == 0))
2177 rl->starved[rw] = 1;
2183 * ioc may be NULL here, and ioc_batching will be false. That's
2184 * OK, if the queue is under the request limit then requests need
2185 * not count toward the nr_batch_requests limit. There will always
2186 * be some limit enforced by BLK_BATCH_TIME.
2188 if (ioc_batching(q, ioc))
2189 ioc->nr_batch_requests--;
2193 blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
2199 * No available requests for this queue, unplug the device and wait for some
2200 * requests to become available.
2202 * Called with q->queue_lock held, and returns with it unlocked.
2204 static struct request *get_request_wait(struct request_queue *q, int rw_flags,
2207 const int rw = rw_flags & 0x01;
2210 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2213 struct request_list *rl = &q->rq;
2215 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2216 TASK_UNINTERRUPTIBLE);
2218 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2221 struct io_context *ioc;
2223 blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
2225 __generic_unplug_device(q);
2226 spin_unlock_irq(q->queue_lock);
2230 * After sleeping, we become a "batching" process and
2231 * will be able to allocate at least one request, and
2232 * up to a big batch of them for a small period time.
2233 * See ioc_batching, ioc_set_batching
2235 ioc = current_io_context(GFP_NOIO, q->node);
2236 ioc_set_batching(q, ioc);
2238 spin_lock_irq(q->queue_lock);
2240 finish_wait(&rl->wait[rw], &wait);
2246 struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
2250 BUG_ON(rw != READ && rw != WRITE);
2252 spin_lock_irq(q->queue_lock);
2253 if (gfp_mask & __GFP_WAIT) {
2254 rq = get_request_wait(q, rw, NULL);
2256 rq = get_request(q, rw, NULL, gfp_mask);
2258 spin_unlock_irq(q->queue_lock);
2260 /* q->queue_lock is unlocked at this point */
2264 EXPORT_SYMBOL(blk_get_request);
2267 * blk_start_queueing - initiate dispatch of requests to device
2268 * @q: request queue to kick into gear
2270 * This is basically a helper to remove the need to know whether a queue
2271 * is plugged or not if someone just wants to initiate dispatch of requests
2274 * The queue lock must be held with interrupts disabled.
2276 void blk_start_queueing(struct request_queue *q)
2278 if (!blk_queue_plugged(q))
2281 __generic_unplug_device(q);
2283 EXPORT_SYMBOL(blk_start_queueing);
2286 * blk_requeue_request - put a request back on queue
2287 * @q: request queue where request should be inserted
2288 * @rq: request to be inserted
2291 * Drivers often keep queueing requests until the hardware cannot accept
2292 * more, when that condition happens we need to put the request back
2293 * on the queue. Must be called with queue lock held.
2295 void blk_requeue_request(struct request_queue *q, struct request *rq)
2297 blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
2299 if (blk_rq_tagged(rq))
2300 blk_queue_end_tag(q, rq);
2302 elv_requeue_request(q, rq);
2305 EXPORT_SYMBOL(blk_requeue_request);
2308 * blk_insert_request - insert a special request in to a request queue
2309 * @q: request queue where request should be inserted
2310 * @rq: request to be inserted
2311 * @at_head: insert request at head or tail of queue
2312 * @data: private data
2315 * Many block devices need to execute commands asynchronously, so they don't
2316 * block the whole kernel from preemption during request execution. This is
2317 * accomplished normally by inserting aritficial requests tagged as
2318 * REQ_SPECIAL in to the corresponding request queue, and letting them be
2319 * scheduled for actual execution by the request queue.
2321 * We have the option of inserting the head or the tail of the queue.
2322 * Typically we use the tail for new ioctls and so forth. We use the head
2323 * of the queue for things like a QUEUE_FULL message from a device, or a
2324 * host that is unable to accept a particular command.
2326 void blk_insert_request(struct request_queue *q, struct request *rq,
2327 int at_head, void *data)
2329 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2330 unsigned long flags;
2333 * tell I/O scheduler that this isn't a regular read/write (ie it
2334 * must not attempt merges on this) and that it acts as a soft
2337 rq->cmd_type = REQ_TYPE_SPECIAL;
2338 rq->cmd_flags |= REQ_SOFTBARRIER;
2342 spin_lock_irqsave(q->queue_lock, flags);
2345 * If command is tagged, release the tag
2347 if (blk_rq_tagged(rq))
2348 blk_queue_end_tag(q, rq);
2350 drive_stat_acct(rq, rq->nr_sectors, 1);
2351 __elv_add_request(q, rq, where, 0);
2352 blk_start_queueing(q);
2353 spin_unlock_irqrestore(q->queue_lock, flags);
2356 EXPORT_SYMBOL(blk_insert_request);
2358 static int __blk_rq_unmap_user(struct bio *bio)
2363 if (bio_flagged(bio, BIO_USER_MAPPED))
2364 bio_unmap_user(bio);
2366 ret = bio_uncopy_user(bio);
2372 int blk_rq_append_bio(struct request_queue *q, struct request *rq,
2376 blk_rq_bio_prep(q, rq, bio);
2377 else if (!ll_back_merge_fn(q, rq, bio))
2380 rq->biotail->bi_next = bio;
2383 rq->data_len += bio->bi_size;
2387 EXPORT_SYMBOL(blk_rq_append_bio);
2389 static int __blk_rq_map_user(struct request_queue *q, struct request *rq,
2390 void __user *ubuf, unsigned int len)
2392 unsigned long uaddr;
2393 struct bio *bio, *orig_bio;
2396 reading = rq_data_dir(rq) == READ;
2399 * if alignment requirement is satisfied, map in user pages for
2400 * direct dma. else, set up kernel bounce buffers
2402 uaddr = (unsigned long) ubuf;
2403 if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2404 bio = bio_map_user(q, NULL, uaddr, len, reading);
2406 bio = bio_copy_user(q, uaddr, len, reading);
2409 return PTR_ERR(bio);
2412 blk_queue_bounce(q, &bio);
2415 * We link the bounce buffer in and could have to traverse it
2416 * later so we have to get a ref to prevent it from being freed
2420 ret = blk_rq_append_bio(q, rq, bio);
2422 return bio->bi_size;
2424 /* if it was boucned we must call the end io function */
2425 bio_endio(bio, bio->bi_size, 0);
2426 __blk_rq_unmap_user(orig_bio);
2432 * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2433 * @q: request queue where request should be inserted
2434 * @rq: request structure to fill
2435 * @ubuf: the user buffer
2436 * @len: length of user data
2439 * Data will be mapped directly for zero copy io, if possible. Otherwise
2440 * a kernel bounce buffer is used.
2442 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2443 * still in process context.
2445 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2446 * before being submitted to the device, as pages mapped may be out of
2447 * reach. It's the callers responsibility to make sure this happens. The
2448 * original bio must be passed back in to blk_rq_unmap_user() for proper
2451 int blk_rq_map_user(struct request_queue *q, struct request *rq,
2452 void __user *ubuf, unsigned long len)
2454 unsigned long bytes_read = 0;
2455 struct bio *bio = NULL;
2458 if (len > (q->max_hw_sectors << 9))
2463 while (bytes_read != len) {
2464 unsigned long map_len, end, start;
2466 map_len = min_t(unsigned long, len - bytes_read, BIO_MAX_SIZE);
2467 end = ((unsigned long)ubuf + map_len + PAGE_SIZE - 1)
2469 start = (unsigned long)ubuf >> PAGE_SHIFT;
2472 * A bad offset could cause us to require BIO_MAX_PAGES + 1
2473 * pages. If this happens we just lower the requested
2474 * mapping len by a page so that we can fit
2476 if (end - start > BIO_MAX_PAGES)
2477 map_len -= PAGE_SIZE;
2479 ret = __blk_rq_map_user(q, rq, ubuf, map_len);
2488 rq->buffer = rq->data = NULL;
2491 blk_rq_unmap_user(bio);
2495 EXPORT_SYMBOL(blk_rq_map_user);
2498 * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2499 * @q: request queue where request should be inserted
2500 * @rq: request to map data to
2501 * @iov: pointer to the iovec
2502 * @iov_count: number of elements in the iovec
2503 * @len: I/O byte count
2506 * Data will be mapped directly for zero copy io, if possible. Otherwise
2507 * a kernel bounce buffer is used.
2509 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2510 * still in process context.
2512 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2513 * before being submitted to the device, as pages mapped may be out of
2514 * reach. It's the callers responsibility to make sure this happens. The
2515 * original bio must be passed back in to blk_rq_unmap_user() for proper
2518 int blk_rq_map_user_iov(struct request_queue *q, struct request *rq,
2519 struct sg_iovec *iov, int iov_count, unsigned int len)
2523 if (!iov || iov_count <= 0)
2526 /* we don't allow misaligned data like bio_map_user() does. If the
2527 * user is using sg, they're expected to know the alignment constraints
2528 * and respect them accordingly */
2529 bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2531 return PTR_ERR(bio);
2533 if (bio->bi_size != len) {
2534 bio_endio(bio, bio->bi_size, 0);
2535 bio_unmap_user(bio);
2540 blk_rq_bio_prep(q, rq, bio);
2541 rq->buffer = rq->data = NULL;
2545 EXPORT_SYMBOL(blk_rq_map_user_iov);
2548 * blk_rq_unmap_user - unmap a request with user data
2549 * @bio: start of bio list
2552 * Unmap a rq previously mapped by blk_rq_map_user(). The caller must
2553 * supply the original rq->bio from the blk_rq_map_user() return, since
2554 * the io completion may have changed rq->bio.
2556 int blk_rq_unmap_user(struct bio *bio)
2558 struct bio *mapped_bio;
2563 if (unlikely(bio_flagged(bio, BIO_BOUNCED)))
2564 mapped_bio = bio->bi_private;
2566 ret2 = __blk_rq_unmap_user(mapped_bio);
2572 bio_put(mapped_bio);
2578 EXPORT_SYMBOL(blk_rq_unmap_user);
2581 * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2582 * @q: request queue where request should be inserted
2583 * @rq: request to fill
2584 * @kbuf: the kernel buffer
2585 * @len: length of user data
2586 * @gfp_mask: memory allocation flags
2588 int blk_rq_map_kern(struct request_queue *q, struct request *rq, void *kbuf,
2589 unsigned int len, gfp_t gfp_mask)
2593 if (len > (q->max_hw_sectors << 9))
2598 bio = bio_map_kern(q, kbuf, len, gfp_mask);
2600 return PTR_ERR(bio);
2602 if (rq_data_dir(rq) == WRITE)
2603 bio->bi_rw |= (1 << BIO_RW);
2605 blk_rq_bio_prep(q, rq, bio);
2606 blk_queue_bounce(q, &rq->bio);
2607 rq->buffer = rq->data = NULL;
2611 EXPORT_SYMBOL(blk_rq_map_kern);
2614 * blk_execute_rq_nowait - insert a request into queue for execution
2615 * @q: queue to insert the request in
2616 * @bd_disk: matching gendisk
2617 * @rq: request to insert
2618 * @at_head: insert request at head or tail of queue
2619 * @done: I/O completion handler
2622 * Insert a fully prepared request at the back of the io scheduler queue
2623 * for execution. Don't wait for completion.
2625 void blk_execute_rq_nowait(struct request_queue *q, struct gendisk *bd_disk,
2626 struct request *rq, int at_head,
2629 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2631 rq->rq_disk = bd_disk;
2632 rq->cmd_flags |= REQ_NOMERGE;
2634 WARN_ON(irqs_disabled());
2635 spin_lock_irq(q->queue_lock);
2636 __elv_add_request(q, rq, where, 1);
2637 __generic_unplug_device(q);
2638 spin_unlock_irq(q->queue_lock);
2640 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
2643 * blk_execute_rq - insert a request into queue for execution
2644 * @q: queue to insert the request in
2645 * @bd_disk: matching gendisk
2646 * @rq: request to insert
2647 * @at_head: insert request at head or tail of queue
2650 * Insert a fully prepared request at the back of the io scheduler queue
2651 * for execution and wait for completion.
2653 int blk_execute_rq(struct request_queue *q, struct gendisk *bd_disk,
2654 struct request *rq, int at_head)
2656 DECLARE_COMPLETION_ONSTACK(wait);
2657 char sense[SCSI_SENSE_BUFFERSIZE];
2661 * we need an extra reference to the request, so we can look at
2662 * it after io completion
2667 memset(sense, 0, sizeof(sense));
2672 rq->end_io_data = &wait;
2673 blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
2674 wait_for_completion(&wait);
2682 EXPORT_SYMBOL(blk_execute_rq);
2685 * blkdev_issue_flush - queue a flush
2686 * @bdev: blockdev to issue flush for
2687 * @error_sector: error sector
2690 * Issue a flush for the block device in question. Caller can supply
2691 * room for storing the error offset in case of a flush error, if they
2692 * wish to. Caller must run wait_for_completion() on its own.
2694 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2696 struct request_queue *q;
2698 if (bdev->bd_disk == NULL)
2701 q = bdev_get_queue(bdev);
2704 if (!q->issue_flush_fn)
2707 return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2710 EXPORT_SYMBOL(blkdev_issue_flush);
2712 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2714 int rw = rq_data_dir(rq);
2716 if (!blk_fs_request(rq) || !rq->rq_disk)
2720 __disk_stat_inc(rq->rq_disk, merges[rw]);
2722 disk_round_stats(rq->rq_disk);
2723 rq->rq_disk->in_flight++;
2728 * add-request adds a request to the linked list.
2729 * queue lock is held and interrupts disabled, as we muck with the
2730 * request queue list.
2732 static inline void add_request(struct request_queue * q, struct request * req)
2734 drive_stat_acct(req, req->nr_sectors, 1);
2737 * elevator indicated where it wants this request to be
2738 * inserted at elevator_merge time
2740 __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2744 * disk_round_stats() - Round off the performance stats on a struct
2747 * The average IO queue length and utilisation statistics are maintained
2748 * by observing the current state of the queue length and the amount of
2749 * time it has been in this state for.
2751 * Normally, that accounting is done on IO completion, but that can result
2752 * in more than a second's worth of IO being accounted for within any one
2753 * second, leading to >100% utilisation. To deal with that, we call this
2754 * function to do a round-off before returning the results when reading
2755 * /proc/diskstats. This accounts immediately for all queue usage up to
2756 * the current jiffies and restarts the counters again.
2758 void disk_round_stats(struct gendisk *disk)
2760 unsigned long now = jiffies;
2762 if (now == disk->stamp)
2765 if (disk->in_flight) {
2766 __disk_stat_add(disk, time_in_queue,
2767 disk->in_flight * (now - disk->stamp));
2768 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2773 EXPORT_SYMBOL_GPL(disk_round_stats);
2776 * queue lock must be held
2778 void __blk_put_request(struct request_queue *q, struct request *req)
2782 if (unlikely(--req->ref_count))
2785 elv_completed_request(q, req);
2788 * Request may not have originated from ll_rw_blk. if not,
2789 * it didn't come out of our reserved rq pools
2791 if (req->cmd_flags & REQ_ALLOCED) {
2792 int rw = rq_data_dir(req);
2793 int priv = req->cmd_flags & REQ_ELVPRIV;
2795 BUG_ON(!list_empty(&req->queuelist));
2796 BUG_ON(!hlist_unhashed(&req->hash));
2798 blk_free_request(q, req);
2799 freed_request(q, rw, priv);
2803 EXPORT_SYMBOL_GPL(__blk_put_request);
2805 void blk_put_request(struct request *req)
2807 unsigned long flags;
2808 struct request_queue *q = req->q;
2811 * Gee, IDE calls in w/ NULL q. Fix IDE and remove the
2812 * following if (q) test.
2815 spin_lock_irqsave(q->queue_lock, flags);
2816 __blk_put_request(q, req);
2817 spin_unlock_irqrestore(q->queue_lock, flags);
2821 EXPORT_SYMBOL(blk_put_request);
2824 * blk_end_sync_rq - executes a completion event on a request
2825 * @rq: request to complete
2826 * @error: end io status of the request
2828 void blk_end_sync_rq(struct request *rq, int error)
2830 struct completion *waiting = rq->end_io_data;
2832 rq->end_io_data = NULL;
2833 __blk_put_request(rq->q, rq);
2836 * complete last, if this is a stack request the process (and thus
2837 * the rq pointer) could be invalid right after this complete()
2841 EXPORT_SYMBOL(blk_end_sync_rq);
2844 * Has to be called with the request spinlock acquired
2846 static int attempt_merge(struct request_queue *q, struct request *req,
2847 struct request *next)
2849 if (!rq_mergeable(req) || !rq_mergeable(next))
2855 if (req->sector + req->nr_sectors != next->sector)
2858 if (rq_data_dir(req) != rq_data_dir(next)
2859 || req->rq_disk != next->rq_disk
2864 * If we are allowed to merge, then append bio list
2865 * from next to rq and release next. merge_requests_fn
2866 * will have updated segment counts, update sector
2869 if (!ll_merge_requests_fn(q, req, next))
2873 * At this point we have either done a back merge
2874 * or front merge. We need the smaller start_time of
2875 * the merged requests to be the current request
2876 * for accounting purposes.
2878 if (time_after(req->start_time, next->start_time))
2879 req->start_time = next->start_time;
2881 req->biotail->bi_next = next->bio;
2882 req->biotail = next->biotail;
2884 req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2886 elv_merge_requests(q, req, next);
2889 disk_round_stats(req->rq_disk);
2890 req->rq_disk->in_flight--;
2893 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2895 __blk_put_request(q, next);
2899 static inline int attempt_back_merge(struct request_queue *q,
2902 struct request *next = elv_latter_request(q, rq);
2905 return attempt_merge(q, rq, next);
2910 static inline int attempt_front_merge(struct request_queue *q,
2913 struct request *prev = elv_former_request(q, rq);
2916 return attempt_merge(q, prev, rq);
2921 static void init_request_from_bio(struct request *req, struct bio *bio)
2923 req->cmd_type = REQ_TYPE_FS;
2926 * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2928 if (bio_rw_ahead(bio) || bio_failfast(bio))
2929 req->cmd_flags |= REQ_FAILFAST;
2932 * REQ_BARRIER implies no merging, but lets make it explicit
2934 if (unlikely(bio_barrier(bio)))
2935 req->cmd_flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2938 req->cmd_flags |= REQ_RW_SYNC;
2939 if (bio_rw_meta(bio))
2940 req->cmd_flags |= REQ_RW_META;
2943 req->hard_sector = req->sector = bio->bi_sector;
2944 req->hard_nr_sectors = req->nr_sectors = bio_sectors(bio);
2945 req->current_nr_sectors = req->hard_cur_sectors = bio_cur_sectors(bio);
2946 req->nr_phys_segments = bio_phys_segments(req->q, bio);
2947 req->nr_hw_segments = bio_hw_segments(req->q, bio);
2948 req->buffer = bio_data(bio); /* see ->buffer comment above */
2949 req->bio = req->biotail = bio;
2950 req->ioprio = bio_prio(bio);
2951 req->rq_disk = bio->bi_bdev->bd_disk;
2952 req->start_time = jiffies;
2955 static int __make_request(struct request_queue *q, struct bio *bio)
2957 struct request *req;
2958 int el_ret, nr_sectors, barrier, err;
2959 const unsigned short prio = bio_prio(bio);
2960 const int sync = bio_sync(bio);
2963 nr_sectors = bio_sectors(bio);
2966 * low level driver can indicate that it wants pages above a
2967 * certain limit bounced to low memory (ie for highmem, or even
2968 * ISA dma in theory)
2970 blk_queue_bounce(q, &bio);
2972 barrier = bio_barrier(bio);
2973 if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
2978 spin_lock_irq(q->queue_lock);
2980 if (unlikely(barrier) || elv_queue_empty(q))
2983 el_ret = elv_merge(q, &req, bio);
2985 case ELEVATOR_BACK_MERGE:
2986 BUG_ON(!rq_mergeable(req));
2988 if (!ll_back_merge_fn(q, req, bio))
2991 blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
2993 req->biotail->bi_next = bio;
2995 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2996 req->ioprio = ioprio_best(req->ioprio, prio);
2997 drive_stat_acct(req, nr_sectors, 0);
2998 if (!attempt_back_merge(q, req))
2999 elv_merged_request(q, req, el_ret);
3002 case ELEVATOR_FRONT_MERGE:
3003 BUG_ON(!rq_mergeable(req));
3005 if (!ll_front_merge_fn(q, req, bio))
3008 blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
3010 bio->bi_next = req->bio;
3014 * may not be valid. if the low level driver said
3015 * it didn't need a bounce buffer then it better
3016 * not touch req->buffer either...
3018 req->buffer = bio_data(bio);
3019 req->current_nr_sectors = bio_cur_sectors(bio);
3020 req->hard_cur_sectors = req->current_nr_sectors;
3021 req->sector = req->hard_sector = bio->bi_sector;
3022 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
3023 req->ioprio = ioprio_best(req->ioprio, prio);
3024 drive_stat_acct(req, nr_sectors, 0);
3025 if (!attempt_front_merge(q, req))
3026 elv_merged_request(q, req, el_ret);
3029 /* ELV_NO_MERGE: elevator says don't/can't merge. */
3036 * This sync check and mask will be re-done in init_request_from_bio(),
3037 * but we need to set it earlier to expose the sync flag to the
3038 * rq allocator and io schedulers.
3040 rw_flags = bio_data_dir(bio);
3042 rw_flags |= REQ_RW_SYNC;
3045 * Grab a free request. This is might sleep but can not fail.
3046 * Returns with the queue unlocked.
3048 req = get_request_wait(q, rw_flags, bio);
3051 * After dropping the lock and possibly sleeping here, our request
3052 * may now be mergeable after it had proven unmergeable (above).
3053 * We don't worry about that case for efficiency. It won't happen
3054 * often, and the elevators are able to handle it.
3056 init_request_from_bio(req, bio);
3058 spin_lock_irq(q->queue_lock);
3059 if (elv_queue_empty(q))
3061 add_request(q, req);
3064 __generic_unplug_device(q);
3066 spin_unlock_irq(q->queue_lock);
3070 bio_endio(bio, nr_sectors << 9, err);
3075 * If bio->bi_dev is a partition, remap the location
3077 static inline void blk_partition_remap(struct bio *bio)
3079 struct block_device *bdev = bio->bi_bdev;
3081 if (bdev != bdev->bd_contains) {
3082 struct hd_struct *p = bdev->bd_part;
3083 const int rw = bio_data_dir(bio);
3085 p->sectors[rw] += bio_sectors(bio);
3088 bio->bi_sector += p->start_sect;
3089 bio->bi_bdev = bdev->bd_contains;
3091 blk_add_trace_remap(bdev_get_queue(bio->bi_bdev), bio,
3092 bdev->bd_dev, bio->bi_sector,
3093 bio->bi_sector - p->start_sect);
3097 static void handle_bad_sector(struct bio *bio)
3099 char b[BDEVNAME_SIZE];
3101 printk(KERN_INFO "attempt to access beyond end of device\n");
3102 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
3103 bdevname(bio->bi_bdev, b),
3105 (unsigned long long)bio->bi_sector + bio_sectors(bio),
3106 (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
3108 set_bit(BIO_EOF, &bio->bi_flags);
3111 #ifdef CONFIG_FAIL_MAKE_REQUEST
3113 static DECLARE_FAULT_ATTR(fail_make_request);
3115 static int __init setup_fail_make_request(char *str)
3117 return setup_fault_attr(&fail_make_request, str);
3119 __setup("fail_make_request=", setup_fail_make_request);
3121 static int should_fail_request(struct bio *bio)
3123 if ((bio->bi_bdev->bd_disk->flags & GENHD_FL_FAIL) ||
3124 (bio->bi_bdev->bd_part && bio->bi_bdev->bd_part->make_it_fail))
3125 return should_fail(&fail_make_request, bio->bi_size);
3130 static int __init fail_make_request_debugfs(void)
3132 return init_fault_attr_dentries(&fail_make_request,
3133 "fail_make_request");
3136 late_initcall(fail_make_request_debugfs);
3138 #else /* CONFIG_FAIL_MAKE_REQUEST */
3140 static inline int should_fail_request(struct bio *bio)
3145 #endif /* CONFIG_FAIL_MAKE_REQUEST */
3148 * generic_make_request: hand a buffer to its device driver for I/O
3149 * @bio: The bio describing the location in memory and on the device.
3151 * generic_make_request() is used to make I/O requests of block
3152 * devices. It is passed a &struct bio, which describes the I/O that needs
3155 * generic_make_request() does not return any status. The
3156 * success/failure status of the request, along with notification of
3157 * completion, is delivered asynchronously through the bio->bi_end_io
3158 * function described (one day) else where.
3160 * The caller of generic_make_request must make sure that bi_io_vec
3161 * are set to describe the memory buffer, and that bi_dev and bi_sector are
3162 * set to describe the device address, and the
3163 * bi_end_io and optionally bi_private are set to describe how
3164 * completion notification should be signaled.
3166 * generic_make_request and the drivers it calls may use bi_next if this
3167 * bio happens to be merged with someone else, and may change bi_dev and
3168 * bi_sector for remaps as it sees fit. So the values of these fields
3169 * should NOT be depended on after the call to generic_make_request.
3171 static inline void __generic_make_request(struct bio *bio)
3173 struct request_queue *q;
3175 sector_t old_sector;
3176 int ret, nr_sectors = bio_sectors(bio);
3180 /* Test device or partition size, when known. */
3181 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3183 sector_t sector = bio->bi_sector;
3185 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
3187 * This may well happen - the kernel calls bread()
3188 * without checking the size of the device, e.g., when
3189 * mounting a device.
3191 handle_bad_sector(bio);
3197 * Resolve the mapping until finished. (drivers are
3198 * still free to implement/resolve their own stacking
3199 * by explicitly returning 0)
3201 * NOTE: we don't repeat the blk_size check for each new device.
3202 * Stacking drivers are expected to know what they are doing.
3207 char b[BDEVNAME_SIZE];
3209 q = bdev_get_queue(bio->bi_bdev);
3212 "generic_make_request: Trying to access "
3213 "nonexistent block-device %s (%Lu)\n",
3214 bdevname(bio->bi_bdev, b),
3215 (long long) bio->bi_sector);
3217 bio_endio(bio, bio->bi_size, -EIO);
3221 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
3222 printk("bio too big device %s (%u > %u)\n",
3223 bdevname(bio->bi_bdev, b),
3229 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
3232 if (should_fail_request(bio))
3236 * If this device has partitions, remap block n
3237 * of partition p to block n+start(p) of the disk.
3239 blk_partition_remap(bio);
3241 if (old_sector != -1)
3242 blk_add_trace_remap(q, bio, old_dev, bio->bi_sector,
3245 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
3247 old_sector = bio->bi_sector;
3248 old_dev = bio->bi_bdev->bd_dev;
3250 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3252 sector_t sector = bio->bi_sector;
3254 if (maxsector < nr_sectors ||
3255 maxsector - nr_sectors < sector) {
3257 * This may well happen - partitions are not
3258 * checked to make sure they are within the size
3259 * of the whole device.
3261 handle_bad_sector(bio);
3266 ret = q->make_request_fn(q, bio);
3271 * We only want one ->make_request_fn to be active at a time,
3272 * else stack usage with stacked devices could be a problem.
3273 * So use current->bio_{list,tail} to keep a list of requests
3274 * submited by a make_request_fn function.
3275 * current->bio_tail is also used as a flag to say if
3276 * generic_make_request is currently active in this task or not.
3277 * If it is NULL, then no make_request is active. If it is non-NULL,
3278 * then a make_request is active, and new requests should be added
3281 void generic_make_request(struct bio *bio)
3283 if (current->bio_tail) {
3284 /* make_request is active */
3285 *(current->bio_tail) = bio;
3286 bio->bi_next = NULL;
3287 current->bio_tail = &bio->bi_next;
3290 /* following loop may be a bit non-obvious, and so deserves some
3292 * Before entering the loop, bio->bi_next is NULL (as all callers
3293 * ensure that) so we have a list with a single bio.
3294 * We pretend that we have just taken it off a longer list, so
3295 * we assign bio_list to the next (which is NULL) and bio_tail
3296 * to &bio_list, thus initialising the bio_list of new bios to be
3297 * added. __generic_make_request may indeed add some more bios
3298 * through a recursive call to generic_make_request. If it
3299 * did, we find a non-NULL value in bio_list and re-enter the loop
3300 * from the top. In this case we really did just take the bio
3301 * of the top of the list (no pretending) and so fixup bio_list and
3302 * bio_tail or bi_next, and call into __generic_make_request again.
3304 * The loop was structured like this to make only one call to
3305 * __generic_make_request (which is important as it is large and
3306 * inlined) and to keep the structure simple.
3308 BUG_ON(bio->bi_next);
3310 current->bio_list = bio->bi_next;
3311 if (bio->bi_next == NULL)
3312 current->bio_tail = ¤t->bio_list;
3314 bio->bi_next = NULL;
3315 __generic_make_request(bio);
3316 bio = current->bio_list;
3318 current->bio_tail = NULL; /* deactivate */
3321 EXPORT_SYMBOL(generic_make_request);
3324 * submit_bio: submit a bio to the block device layer for I/O
3325 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3326 * @bio: The &struct bio which describes the I/O
3328 * submit_bio() is very similar in purpose to generic_make_request(), and
3329 * uses that function to do most of the work. Both are fairly rough
3330 * interfaces, @bio must be presetup and ready for I/O.
3333 void submit_bio(int rw, struct bio *bio)
3335 int count = bio_sectors(bio);
3337 BIO_BUG_ON(!bio->bi_size);
3338 BIO_BUG_ON(!bio->bi_io_vec);
3341 count_vm_events(PGPGOUT, count);
3343 task_io_account_read(bio->bi_size);
3344 count_vm_events(PGPGIN, count);
3347 if (unlikely(block_dump)) {
3348 char b[BDEVNAME_SIZE];
3349 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
3350 current->comm, current->pid,
3351 (rw & WRITE) ? "WRITE" : "READ",
3352 (unsigned long long)bio->bi_sector,
3353 bdevname(bio->bi_bdev,b));
3356 generic_make_request(bio);
3359 EXPORT_SYMBOL(submit_bio);
3361 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
3363 if (blk_fs_request(rq)) {
3364 rq->hard_sector += nsect;
3365 rq->hard_nr_sectors -= nsect;
3368 * Move the I/O submission pointers ahead if required.
3370 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3371 (rq->sector <= rq->hard_sector)) {
3372 rq->sector = rq->hard_sector;
3373 rq->nr_sectors = rq->hard_nr_sectors;
3374 rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3375 rq->current_nr_sectors = rq->hard_cur_sectors;
3376 rq->buffer = bio_data(rq->bio);
3380 * if total number of sectors is less than the first segment
3381 * size, something has gone terribly wrong
3383 if (rq->nr_sectors < rq->current_nr_sectors) {
3384 printk("blk: request botched\n");
3385 rq->nr_sectors = rq->current_nr_sectors;
3390 static int __end_that_request_first(struct request *req, int uptodate,
3393 int total_bytes, bio_nbytes, error, next_idx = 0;
3396 blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
3399 * extend uptodate bool to allow < 0 value to be direct io error
3402 if (end_io_error(uptodate))
3403 error = !uptodate ? -EIO : uptodate;
3406 * for a REQ_BLOCK_PC request, we want to carry any eventual
3407 * sense key with us all the way through
3409 if (!blk_pc_request(req))
3413 if (blk_fs_request(req) && !(req->cmd_flags & REQ_QUIET))
3414 printk("end_request: I/O error, dev %s, sector %llu\n",
3415 req->rq_disk ? req->rq_disk->disk_name : "?",
3416 (unsigned long long)req->sector);
3419 if (blk_fs_request(req) && req->rq_disk) {
3420 const int rw = rq_data_dir(req);
3422 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
3425 total_bytes = bio_nbytes = 0;
3426 while ((bio = req->bio) != NULL) {
3429 if (nr_bytes >= bio->bi_size) {
3430 req->bio = bio->bi_next;
3431 nbytes = bio->bi_size;
3432 if (!ordered_bio_endio(req, bio, nbytes, error))
3433 bio_endio(bio, nbytes, error);
3437 int idx = bio->bi_idx + next_idx;
3439 if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3440 blk_dump_rq_flags(req, "__end_that");
3441 printk("%s: bio idx %d >= vcnt %d\n",
3443 bio->bi_idx, bio->bi_vcnt);
3447 nbytes = bio_iovec_idx(bio, idx)->bv_len;
3448 BIO_BUG_ON(nbytes > bio->bi_size);
3451 * not a complete bvec done
3453 if (unlikely(nbytes > nr_bytes)) {
3454 bio_nbytes += nr_bytes;
3455 total_bytes += nr_bytes;
3460 * advance to the next vector
3463 bio_nbytes += nbytes;
3466 total_bytes += nbytes;
3469 if ((bio = req->bio)) {
3471 * end more in this run, or just return 'not-done'
3473 if (unlikely(nr_bytes <= 0))
3485 * if the request wasn't completed, update state
3488 if (!ordered_bio_endio(req, bio, bio_nbytes, error))
3489 bio_endio(bio, bio_nbytes, error);
3490 bio->bi_idx += next_idx;
3491 bio_iovec(bio)->bv_offset += nr_bytes;
3492 bio_iovec(bio)->bv_len -= nr_bytes;
3495 blk_recalc_rq_sectors(req, total_bytes >> 9);
3496 blk_recalc_rq_segments(req);
3501 * end_that_request_first - end I/O on a request
3502 * @req: the request being processed
3503 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3504 * @nr_sectors: number of sectors to end I/O on
3507 * Ends I/O on a number of sectors attached to @req, and sets it up
3508 * for the next range of segments (if any) in the cluster.
3511 * 0 - we are done with this request, call end_that_request_last()
3512 * 1 - still buffers pending for this request
3514 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3516 return __end_that_request_first(req, uptodate, nr_sectors << 9);
3519 EXPORT_SYMBOL(end_that_request_first);
3522 * end_that_request_chunk - end I/O on a request
3523 * @req: the request being processed
3524 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3525 * @nr_bytes: number of bytes to complete
3528 * Ends I/O on a number of bytes attached to @req, and sets it up
3529 * for the next range of segments (if any). Like end_that_request_first(),
3530 * but deals with bytes instead of sectors.
3533 * 0 - we are done with this request, call end_that_request_last()
3534 * 1 - still buffers pending for this request
3536 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3538 return __end_that_request_first(req, uptodate, nr_bytes);
3541 EXPORT_SYMBOL(end_that_request_chunk);
3544 * splice the completion data to a local structure and hand off to
3545 * process_completion_queue() to complete the requests
3547 static void blk_done_softirq(struct softirq_action *h)
3549 struct list_head *cpu_list, local_list;
3551 local_irq_disable();
3552 cpu_list = &__get_cpu_var(blk_cpu_done);
3553 list_replace_init(cpu_list, &local_list);
3556 while (!list_empty(&local_list)) {
3557 struct request *rq = list_entry(local_list.next, struct request, donelist);
3559 list_del_init(&rq->donelist);
3560 rq->q->softirq_done_fn(rq);
3564 static int blk_cpu_notify(struct notifier_block *self, unsigned long action,
3568 * If a CPU goes away, splice its entries to the current CPU
3569 * and trigger a run of the softirq
3571 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
3572 int cpu = (unsigned long) hcpu;
3574 local_irq_disable();
3575 list_splice_init(&per_cpu(blk_cpu_done, cpu),
3576 &__get_cpu_var(blk_cpu_done));
3577 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3585 static struct notifier_block __devinitdata blk_cpu_notifier = {
3586 .notifier_call = blk_cpu_notify,
3590 * blk_complete_request - end I/O on a request
3591 * @req: the request being processed
3594 * Ends all I/O on a request. It does not handle partial completions,
3595 * unless the driver actually implements this in its completion callback
3596 * through requeueing. Theh actual completion happens out-of-order,
3597 * through a softirq handler. The user must have registered a completion
3598 * callback through blk_queue_softirq_done().
3601 void blk_complete_request(struct request *req)
3603 struct list_head *cpu_list;
3604 unsigned long flags;
3606 BUG_ON(!req->q->softirq_done_fn);
3608 local_irq_save(flags);
3610 cpu_list = &__get_cpu_var(blk_cpu_done);
3611 list_add_tail(&req->donelist, cpu_list);
3612 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3614 local_irq_restore(flags);
3617 EXPORT_SYMBOL(blk_complete_request);
3620 * queue lock must be held
3622 void end_that_request_last(struct request *req, int uptodate)
3624 struct gendisk *disk = req->rq_disk;
3628 * extend uptodate bool to allow < 0 value to be direct io error
3631 if (end_io_error(uptodate))
3632 error = !uptodate ? -EIO : uptodate;
3634 if (unlikely(laptop_mode) && blk_fs_request(req))
3635 laptop_io_completion();
3638 * Account IO completion. bar_rq isn't accounted as a normal
3639 * IO on queueing nor completion. Accounting the containing
3640 * request is enough.
3642 if (disk && blk_fs_request(req) && req != &req->q->bar_rq) {
3643 unsigned long duration = jiffies - req->start_time;
3644 const int rw = rq_data_dir(req);
3646 __disk_stat_inc(disk, ios[rw]);
3647 __disk_stat_add(disk, ticks[rw], duration);
3648 disk_round_stats(disk);
3652 req->end_io(req, error);
3654 __blk_put_request(req->q, req);
3657 EXPORT_SYMBOL(end_that_request_last);
3659 void end_request(struct request *req, int uptodate)
3661 if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3662 add_disk_randomness(req->rq_disk);
3663 blkdev_dequeue_request(req);
3664 end_that_request_last(req, uptodate);
3668 EXPORT_SYMBOL(end_request);
3670 static void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
3673 /* first two bits are identical in rq->cmd_flags and bio->bi_rw */
3674 rq->cmd_flags |= (bio->bi_rw & 3);
3676 rq->nr_phys_segments = bio_phys_segments(q, bio);
3677 rq->nr_hw_segments = bio_hw_segments(q, bio);
3678 rq->current_nr_sectors = bio_cur_sectors(bio);
3679 rq->hard_cur_sectors = rq->current_nr_sectors;
3680 rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3681 rq->buffer = bio_data(bio);
3682 rq->data_len = bio->bi_size;
3684 rq->bio = rq->biotail = bio;
3687 rq->rq_disk = bio->bi_bdev->bd_disk;
3690 int kblockd_schedule_work(struct work_struct *work)
3692 return queue_work(kblockd_workqueue, work);
3695 EXPORT_SYMBOL(kblockd_schedule_work);
3697 void kblockd_flush_work(struct work_struct *work)
3699 cancel_work_sync(work);
3701 EXPORT_SYMBOL(kblockd_flush_work);
3703 int __init blk_dev_init(void)
3707 kblockd_workqueue = create_workqueue("kblockd");
3708 if (!kblockd_workqueue)
3709 panic("Failed to create kblockd\n");
3711 request_cachep = kmem_cache_create("blkdev_requests",
3712 sizeof(struct request), 0, SLAB_PANIC, NULL);
3714 requestq_cachep = kmem_cache_create("blkdev_queue",
3715 sizeof(struct request_queue), 0, SLAB_PANIC, NULL);
3717 iocontext_cachep = kmem_cache_create("blkdev_ioc",
3718 sizeof(struct io_context), 0, SLAB_PANIC, NULL);
3720 for_each_possible_cpu(i)
3721 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3723 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
3724 register_hotcpu_notifier(&blk_cpu_notifier);
3726 blk_max_low_pfn = max_low_pfn - 1;
3727 blk_max_pfn = max_pfn - 1;
3733 * IO Context helper functions
3735 void put_io_context(struct io_context *ioc)
3740 BUG_ON(atomic_read(&ioc->refcount) == 0);
3742 if (atomic_dec_and_test(&ioc->refcount)) {
3743 struct cfq_io_context *cic;
3746 if (ioc->aic && ioc->aic->dtor)
3747 ioc->aic->dtor(ioc->aic);
3748 if (ioc->cic_root.rb_node != NULL) {
3749 struct rb_node *n = rb_first(&ioc->cic_root);
3751 cic = rb_entry(n, struct cfq_io_context, rb_node);
3756 kmem_cache_free(iocontext_cachep, ioc);
3759 EXPORT_SYMBOL(put_io_context);
3761 /* Called by the exitting task */
3762 void exit_io_context(void)
3764 struct io_context *ioc;
3765 struct cfq_io_context *cic;
3768 ioc = current->io_context;
3769 current->io_context = NULL;
3770 task_unlock(current);
3773 if (ioc->aic && ioc->aic->exit)
3774 ioc->aic->exit(ioc->aic);
3775 if (ioc->cic_root.rb_node != NULL) {
3776 cic = rb_entry(rb_first(&ioc->cic_root), struct cfq_io_context, rb_node);
3780 put_io_context(ioc);
3784 * If the current task has no IO context then create one and initialise it.
3785 * Otherwise, return its existing IO context.
3787 * This returned IO context doesn't have a specifically elevated refcount,
3788 * but since the current task itself holds a reference, the context can be
3789 * used in general code, so long as it stays within `current` context.
3791 static struct io_context *current_io_context(gfp_t gfp_flags, int node)
3793 struct task_struct *tsk = current;
3794 struct io_context *ret;
3796 ret = tsk->io_context;
3800 ret = kmem_cache_alloc_node(iocontext_cachep, gfp_flags, node);
3802 atomic_set(&ret->refcount, 1);
3803 ret->task = current;
3804 ret->ioprio_changed = 0;
3805 ret->last_waited = jiffies; /* doesn't matter... */
3806 ret->nr_batch_requests = 0; /* because this is 0 */
3808 ret->cic_root.rb_node = NULL;
3809 ret->ioc_data = NULL;
3810 /* make sure set_task_ioprio() sees the settings above */
3812 tsk->io_context = ret;
3819 * If the current task has no IO context then create one and initialise it.
3820 * If it does have a context, take a ref on it.
3822 * This is always called in the context of the task which submitted the I/O.
3824 struct io_context *get_io_context(gfp_t gfp_flags, int node)
3826 struct io_context *ret;
3827 ret = current_io_context(gfp_flags, node);
3829 atomic_inc(&ret->refcount);
3832 EXPORT_SYMBOL(get_io_context);
3834 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3836 struct io_context *src = *psrc;
3837 struct io_context *dst = *pdst;
3840 BUG_ON(atomic_read(&src->refcount) == 0);
3841 atomic_inc(&src->refcount);
3842 put_io_context(dst);
3846 EXPORT_SYMBOL(copy_io_context);
3848 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3850 struct io_context *temp;
3855 EXPORT_SYMBOL(swap_io_context);
3860 struct queue_sysfs_entry {
3861 struct attribute attr;
3862 ssize_t (*show)(struct request_queue *, char *);
3863 ssize_t (*store)(struct request_queue *, const char *, size_t);
3867 queue_var_show(unsigned int var, char *page)
3869 return sprintf(page, "%d\n", var);
3873 queue_var_store(unsigned long *var, const char *page, size_t count)
3875 char *p = (char *) page;
3877 *var = simple_strtoul(p, &p, 10);
3881 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3883 return queue_var_show(q->nr_requests, (page));
3887 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3889 struct request_list *rl = &q->rq;
3891 int ret = queue_var_store(&nr, page, count);
3892 if (nr < BLKDEV_MIN_RQ)
3895 spin_lock_irq(q->queue_lock);
3896 q->nr_requests = nr;
3897 blk_queue_congestion_threshold(q);
3899 if (rl->count[READ] >= queue_congestion_on_threshold(q))
3900 blk_set_queue_congested(q, READ);
3901 else if (rl->count[READ] < queue_congestion_off_threshold(q))
3902 blk_clear_queue_congested(q, READ);
3904 if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3905 blk_set_queue_congested(q, WRITE);
3906 else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3907 blk_clear_queue_congested(q, WRITE);
3909 if (rl->count[READ] >= q->nr_requests) {
3910 blk_set_queue_full(q, READ);
3911 } else if (rl->count[READ]+1 <= q->nr_requests) {
3912 blk_clear_queue_full(q, READ);
3913 wake_up(&rl->wait[READ]);
3916 if (rl->count[WRITE] >= q->nr_requests) {
3917 blk_set_queue_full(q, WRITE);
3918 } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3919 blk_clear_queue_full(q, WRITE);
3920 wake_up(&rl->wait[WRITE]);
3922 spin_unlock_irq(q->queue_lock);
3926 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3928 int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3930 return queue_var_show(ra_kb, (page));
3934 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3936 unsigned long ra_kb;
3937 ssize_t ret = queue_var_store(&ra_kb, page, count);
3939 spin_lock_irq(q->queue_lock);
3940 q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3941 spin_unlock_irq(q->queue_lock);
3946 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3948 int max_sectors_kb = q->max_sectors >> 1;
3950 return queue_var_show(max_sectors_kb, (page));
3954 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3956 unsigned long max_sectors_kb,
3957 max_hw_sectors_kb = q->max_hw_sectors >> 1,
3958 page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3959 ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3962 if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3965 * Take the queue lock to update the readahead and max_sectors
3966 * values synchronously:
3968 spin_lock_irq(q->queue_lock);
3970 * Trim readahead window as well, if necessary:
3972 ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3973 if (ra_kb > max_sectors_kb)
3974 q->backing_dev_info.ra_pages =
3975 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3977 q->max_sectors = max_sectors_kb << 1;
3978 spin_unlock_irq(q->queue_lock);
3983 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3985 int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3987 return queue_var_show(max_hw_sectors_kb, (page));
3991 static struct queue_sysfs_entry queue_requests_entry = {
3992 .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3993 .show = queue_requests_show,
3994 .store = queue_requests_store,
3997 static struct queue_sysfs_entry queue_ra_entry = {
3998 .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3999 .show = queue_ra_show,
4000 .store = queue_ra_store,
4003 static struct queue_sysfs_entry queue_max_sectors_entry = {
4004 .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
4005 .show = queue_max_sectors_show,
4006 .store = queue_max_sectors_store,
4009 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
4010 .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
4011 .show = queue_max_hw_sectors_show,
4014 static struct queue_sysfs_entry queue_iosched_entry = {
4015 .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
4016 .show = elv_iosched_show,
4017 .store = elv_iosched_store,
4020 static struct attribute *default_attrs[] = {
4021 &queue_requests_entry.attr,
4022 &queue_ra_entry.attr,
4023 &queue_max_hw_sectors_entry.attr,
4024 &queue_max_sectors_entry.attr,
4025 &queue_iosched_entry.attr,
4029 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
4032 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
4034 struct queue_sysfs_entry *entry = to_queue(attr);
4035 struct request_queue *q =
4036 container_of(kobj, struct request_queue, kobj);
4041 mutex_lock(&q->sysfs_lock);
4042 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
4043 mutex_unlock(&q->sysfs_lock);
4046 res = entry->show(q, page);
4047 mutex_unlock(&q->sysfs_lock);
4052 queue_attr_store(struct kobject *kobj, struct attribute *attr,
4053 const char *page, size_t length)
4055 struct queue_sysfs_entry *entry = to_queue(attr);
4056 struct request_queue *q = container_of(kobj, struct request_queue, kobj);
4062 mutex_lock(&q->sysfs_lock);
4063 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
4064 mutex_unlock(&q->sysfs_lock);
4067 res = entry->store(q, page, length);
4068 mutex_unlock(&q->sysfs_lock);
4072 static struct sysfs_ops queue_sysfs_ops = {
4073 .show = queue_attr_show,
4074 .store = queue_attr_store,
4077 static struct kobj_type queue_ktype = {
4078 .sysfs_ops = &queue_sysfs_ops,
4079 .default_attrs = default_attrs,
4080 .release = blk_release_queue,
4083 int blk_register_queue(struct gendisk *disk)
4087 struct request_queue *q = disk->queue;
4089 if (!q || !q->request_fn)
4092 q->kobj.parent = kobject_get(&disk->kobj);
4094 ret = kobject_add(&q->kobj);
4098 kobject_uevent(&q->kobj, KOBJ_ADD);
4100 ret = elv_register_queue(q);
4102 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4103 kobject_del(&q->kobj);
4110 void blk_unregister_queue(struct gendisk *disk)
4112 struct request_queue *q = disk->queue;
4114 if (q && q->request_fn) {
4115 elv_unregister_queue(q);
4117 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4118 kobject_del(&q->kobj);
4119 kobject_put(&disk->kobj);