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);
48 * For the allocated request tables
50 static struct kmem_cache *request_cachep;
53 * For queue allocation
55 static struct kmem_cache *requestq_cachep;
58 * For io context allocations
60 static struct kmem_cache *iocontext_cachep;
63 * Controlling structure to kblockd
65 static struct workqueue_struct *kblockd_workqueue;
67 unsigned long blk_max_low_pfn, blk_max_pfn;
69 EXPORT_SYMBOL(blk_max_low_pfn);
70 EXPORT_SYMBOL(blk_max_pfn);
72 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
74 /* Amount of time in which a process may batch requests */
75 #define BLK_BATCH_TIME (HZ/50UL)
77 /* Number of requests a "batching" process may submit */
78 #define BLK_BATCH_REQ 32
81 * Return the threshold (number of used requests) at which the queue is
82 * considered to be congested. It include a little hysteresis to keep the
83 * context switch rate down.
85 static inline int queue_congestion_on_threshold(struct request_queue *q)
87 return q->nr_congestion_on;
91 * The threshold at which a queue is considered to be uncongested
93 static inline int queue_congestion_off_threshold(struct request_queue *q)
95 return q->nr_congestion_off;
98 static void blk_queue_congestion_threshold(struct request_queue *q)
102 nr = q->nr_requests - (q->nr_requests / 8) + 1;
103 if (nr > q->nr_requests)
105 q->nr_congestion_on = nr;
107 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
110 q->nr_congestion_off = nr;
114 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
117 * Locates the passed device's request queue and returns the address of its
120 * Will return NULL if the request queue cannot be located.
122 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
124 struct backing_dev_info *ret = NULL;
125 struct request_queue *q = bdev_get_queue(bdev);
128 ret = &q->backing_dev_info;
131 EXPORT_SYMBOL(blk_get_backing_dev_info);
134 * blk_queue_prep_rq - set a prepare_request function for queue
136 * @pfn: prepare_request function
138 * It's possible for a queue to register a prepare_request callback which
139 * is invoked before the request is handed to the request_fn. The goal of
140 * the function is to prepare a request for I/O, it can be used to build a
141 * cdb from the request data for instance.
144 void blk_queue_prep_rq(struct request_queue *q, prep_rq_fn *pfn)
149 EXPORT_SYMBOL(blk_queue_prep_rq);
152 * blk_queue_merge_bvec - set a merge_bvec function for queue
154 * @mbfn: merge_bvec_fn
156 * Usually queues have static limitations on the max sectors or segments that
157 * we can put in a request. Stacking drivers may have some settings that
158 * are dynamic, and thus we have to query the queue whether it is ok to
159 * add a new bio_vec to a bio at a given offset or not. If the block device
160 * has such limitations, it needs to register a merge_bvec_fn to control
161 * the size of bio's sent to it. Note that a block device *must* allow a
162 * single page to be added to an empty bio. The block device driver may want
163 * to use the bio_split() function to deal with these bio's. By default
164 * no merge_bvec_fn is defined for a queue, and only the fixed limits are
167 void blk_queue_merge_bvec(struct request_queue *q, merge_bvec_fn *mbfn)
169 q->merge_bvec_fn = mbfn;
172 EXPORT_SYMBOL(blk_queue_merge_bvec);
174 void blk_queue_softirq_done(struct request_queue *q, softirq_done_fn *fn)
176 q->softirq_done_fn = fn;
179 EXPORT_SYMBOL(blk_queue_softirq_done);
182 * blk_queue_make_request - define an alternate make_request function for a device
183 * @q: the request queue for the device to be affected
184 * @mfn: the alternate make_request function
187 * The normal way for &struct bios to be passed to a device
188 * driver is for them to be collected into requests on a request
189 * queue, and then to allow the device driver to select requests
190 * off that queue when it is ready. This works well for many block
191 * devices. However some block devices (typically virtual devices
192 * such as md or lvm) do not benefit from the processing on the
193 * request queue, and are served best by having the requests passed
194 * directly to them. This can be achieved by providing a function
195 * to blk_queue_make_request().
198 * The driver that does this *must* be able to deal appropriately
199 * with buffers in "highmemory". This can be accomplished by either calling
200 * __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
201 * blk_queue_bounce() to create a buffer in normal memory.
203 void blk_queue_make_request(struct request_queue * q, make_request_fn * mfn)
208 q->nr_requests = BLKDEV_MAX_RQ;
209 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
210 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
211 q->make_request_fn = mfn;
212 q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
213 q->backing_dev_info.state = 0;
214 q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
215 blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
216 blk_queue_hardsect_size(q, 512);
217 blk_queue_dma_alignment(q, 511);
218 blk_queue_congestion_threshold(q);
219 q->nr_batching = BLK_BATCH_REQ;
221 q->unplug_thresh = 4; /* hmm */
222 q->unplug_delay = (3 * HZ) / 1000; /* 3 milliseconds */
223 if (q->unplug_delay == 0)
226 INIT_WORK(&q->unplug_work, blk_unplug_work);
228 q->unplug_timer.function = blk_unplug_timeout;
229 q->unplug_timer.data = (unsigned long)q;
232 * by default assume old behaviour and bounce for any highmem page
234 blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
237 EXPORT_SYMBOL(blk_queue_make_request);
239 static void rq_init(struct request_queue *q, struct request *rq)
241 INIT_LIST_HEAD(&rq->queuelist);
242 INIT_LIST_HEAD(&rq->donelist);
245 rq->bio = rq->biotail = NULL;
246 INIT_HLIST_NODE(&rq->hash);
247 RB_CLEAR_NODE(&rq->rb_node);
255 rq->nr_phys_segments = 0;
258 rq->end_io_data = NULL;
259 rq->completion_data = NULL;
264 * blk_queue_ordered - does this queue support ordered writes
265 * @q: the request queue
266 * @ordered: one of QUEUE_ORDERED_*
267 * @prepare_flush_fn: rq setup helper for cache flush ordered writes
270 * For journalled file systems, doing ordered writes on a commit
271 * block instead of explicitly doing wait_on_buffer (which is bad
272 * for performance) can be a big win. Block drivers supporting this
273 * feature should call this function and indicate so.
276 int blk_queue_ordered(struct request_queue *q, unsigned ordered,
277 prepare_flush_fn *prepare_flush_fn)
279 if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
280 prepare_flush_fn == NULL) {
281 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
285 if (ordered != QUEUE_ORDERED_NONE &&
286 ordered != QUEUE_ORDERED_DRAIN &&
287 ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
288 ordered != QUEUE_ORDERED_DRAIN_FUA &&
289 ordered != QUEUE_ORDERED_TAG &&
290 ordered != QUEUE_ORDERED_TAG_FLUSH &&
291 ordered != QUEUE_ORDERED_TAG_FUA) {
292 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
296 q->ordered = ordered;
297 q->next_ordered = ordered;
298 q->prepare_flush_fn = prepare_flush_fn;
303 EXPORT_SYMBOL(blk_queue_ordered);
306 * blk_queue_issue_flush_fn - set function for issuing a flush
307 * @q: the request queue
308 * @iff: the function to be called issuing the flush
311 * If a driver supports issuing a flush command, the support is notified
312 * to the block layer by defining it through this call.
315 void blk_queue_issue_flush_fn(struct request_queue *q, issue_flush_fn *iff)
317 q->issue_flush_fn = iff;
320 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
323 * Cache flushing for ordered writes handling
325 inline unsigned blk_ordered_cur_seq(struct request_queue *q)
329 return 1 << ffz(q->ordseq);
332 unsigned blk_ordered_req_seq(struct request *rq)
334 struct request_queue *q = rq->q;
336 BUG_ON(q->ordseq == 0);
338 if (rq == &q->pre_flush_rq)
339 return QUEUE_ORDSEQ_PREFLUSH;
340 if (rq == &q->bar_rq)
341 return QUEUE_ORDSEQ_BAR;
342 if (rq == &q->post_flush_rq)
343 return QUEUE_ORDSEQ_POSTFLUSH;
346 * !fs requests don't need to follow barrier ordering. Always
347 * put them at the front. This fixes the following deadlock.
349 * http://thread.gmane.org/gmane.linux.kernel/537473
351 if (!blk_fs_request(rq))
352 return QUEUE_ORDSEQ_DRAIN;
354 if ((rq->cmd_flags & REQ_ORDERED_COLOR) ==
355 (q->orig_bar_rq->cmd_flags & REQ_ORDERED_COLOR))
356 return QUEUE_ORDSEQ_DRAIN;
358 return QUEUE_ORDSEQ_DONE;
361 void blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
366 if (error && !q->orderr)
369 BUG_ON(q->ordseq & seq);
372 if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
376 * Okay, sequence complete.
379 uptodate = q->orderr ? q->orderr : 1;
383 end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
384 end_that_request_last(rq, uptodate);
387 static void pre_flush_end_io(struct request *rq, int error)
389 elv_completed_request(rq->q, rq);
390 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
393 static void bar_end_io(struct request *rq, int error)
395 elv_completed_request(rq->q, rq);
396 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
399 static void post_flush_end_io(struct request *rq, int error)
401 elv_completed_request(rq->q, rq);
402 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
405 static void queue_flush(struct request_queue *q, unsigned which)
408 rq_end_io_fn *end_io;
410 if (which == QUEUE_ORDERED_PREFLUSH) {
411 rq = &q->pre_flush_rq;
412 end_io = pre_flush_end_io;
414 rq = &q->post_flush_rq;
415 end_io = post_flush_end_io;
418 rq->cmd_flags = REQ_HARDBARRIER;
420 rq->elevator_private = NULL;
421 rq->elevator_private2 = NULL;
422 rq->rq_disk = q->bar_rq.rq_disk;
424 q->prepare_flush_fn(q, rq);
426 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
429 static inline struct request *start_ordered(struct request_queue *q,
434 q->ordered = q->next_ordered;
435 q->ordseq |= QUEUE_ORDSEQ_STARTED;
438 * Prep proxy barrier request.
440 blkdev_dequeue_request(rq);
445 if (bio_data_dir(q->orig_bar_rq->bio) == WRITE)
446 rq->cmd_flags |= REQ_RW;
447 rq->cmd_flags |= q->ordered & QUEUE_ORDERED_FUA ? REQ_FUA : 0;
448 rq->elevator_private = NULL;
449 rq->elevator_private2 = NULL;
450 init_request_from_bio(rq, q->orig_bar_rq->bio);
451 rq->end_io = bar_end_io;
454 * Queue ordered sequence. As we stack them at the head, we
455 * need to queue in reverse order. Note that we rely on that
456 * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
457 * request gets inbetween ordered sequence.
459 if (q->ordered & QUEUE_ORDERED_POSTFLUSH)
460 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
462 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
464 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
466 if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
467 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
468 rq = &q->pre_flush_rq;
470 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
472 if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
473 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
480 int blk_do_ordered(struct request_queue *q, struct request **rqp)
482 struct request *rq = *rqp;
483 int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
489 if (q->next_ordered != QUEUE_ORDERED_NONE) {
490 *rqp = start_ordered(q, rq);
494 * This can happen when the queue switches to
495 * ORDERED_NONE while this request is on it.
497 blkdev_dequeue_request(rq);
498 end_that_request_first(rq, -EOPNOTSUPP,
499 rq->hard_nr_sectors);
500 end_that_request_last(rq, -EOPNOTSUPP);
507 * Ordered sequence in progress
510 /* Special requests are not subject to ordering rules. */
511 if (!blk_fs_request(rq) &&
512 rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
515 if (q->ordered & QUEUE_ORDERED_TAG) {
516 /* Ordered by tag. Blocking the next barrier is enough. */
517 if (is_barrier && rq != &q->bar_rq)
520 /* Ordered by draining. Wait for turn. */
521 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
522 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
529 static int flush_dry_bio_endio(struct bio *bio, unsigned int bytes, int error)
531 struct request_queue *q = bio->bi_private;
534 * This is dry run, restore bio_sector and size. We'll finish
535 * this request again with the original bi_end_io after an
536 * error occurs or post flush is complete.
544 set_bit(BIO_UPTODATE, &bio->bi_flags);
545 bio->bi_size = q->bi_size;
546 bio->bi_sector -= (q->bi_size >> 9);
552 static int ordered_bio_endio(struct request *rq, struct bio *bio,
553 unsigned int nbytes, int error)
555 struct request_queue *q = rq->q;
559 if (&q->bar_rq != rq)
563 * Okay, this is the barrier request in progress, dry finish it.
565 if (error && !q->orderr)
568 endio = bio->bi_end_io;
569 private = bio->bi_private;
570 bio->bi_end_io = flush_dry_bio_endio;
573 bio_endio(bio, nbytes, error);
575 bio->bi_end_io = endio;
576 bio->bi_private = private;
582 * blk_queue_bounce_limit - set bounce buffer limit for queue
583 * @q: the request queue for the device
584 * @dma_addr: bus address limit
587 * Different hardware can have different requirements as to what pages
588 * it can do I/O directly to. A low level driver can call
589 * blk_queue_bounce_limit to have lower memory pages allocated as bounce
590 * buffers for doing I/O to pages residing above @page.
592 void blk_queue_bounce_limit(struct request_queue *q, u64 dma_addr)
594 unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
597 q->bounce_gfp = GFP_NOIO;
598 #if BITS_PER_LONG == 64
599 /* Assume anything <= 4GB can be handled by IOMMU.
600 Actually some IOMMUs can handle everything, but I don't
601 know of a way to test this here. */
602 if (bounce_pfn < (min_t(u64,0xffffffff,BLK_BOUNCE_HIGH) >> PAGE_SHIFT))
604 q->bounce_pfn = max_low_pfn;
606 if (bounce_pfn < blk_max_low_pfn)
608 q->bounce_pfn = bounce_pfn;
611 init_emergency_isa_pool();
612 q->bounce_gfp = GFP_NOIO | GFP_DMA;
613 q->bounce_pfn = bounce_pfn;
617 EXPORT_SYMBOL(blk_queue_bounce_limit);
620 * blk_queue_max_sectors - set max sectors for a request for this queue
621 * @q: the request queue for the device
622 * @max_sectors: max sectors in the usual 512b unit
625 * Enables a low level driver to set an upper limit on the size of
628 void blk_queue_max_sectors(struct request_queue *q, unsigned int max_sectors)
630 if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
631 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
632 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
635 if (BLK_DEF_MAX_SECTORS > max_sectors)
636 q->max_hw_sectors = q->max_sectors = max_sectors;
638 q->max_sectors = BLK_DEF_MAX_SECTORS;
639 q->max_hw_sectors = max_sectors;
643 EXPORT_SYMBOL(blk_queue_max_sectors);
646 * blk_queue_max_phys_segments - set max phys segments for a request for this queue
647 * @q: the request queue for the device
648 * @max_segments: max number of segments
651 * Enables a low level driver to set an upper limit on the number of
652 * physical data segments in a request. This would be the largest sized
653 * scatter list the driver could handle.
655 void blk_queue_max_phys_segments(struct request_queue *q,
656 unsigned short max_segments)
660 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
663 q->max_phys_segments = max_segments;
666 EXPORT_SYMBOL(blk_queue_max_phys_segments);
669 * blk_queue_max_hw_segments - set max hw segments for a request for this queue
670 * @q: the request queue for the device
671 * @max_segments: max number of segments
674 * Enables a low level driver to set an upper limit on the number of
675 * hw data segments in a request. This would be the largest number of
676 * address/length pairs the host adapter can actually give as once
679 void blk_queue_max_hw_segments(struct request_queue *q,
680 unsigned short max_segments)
684 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
687 q->max_hw_segments = max_segments;
690 EXPORT_SYMBOL(blk_queue_max_hw_segments);
693 * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
694 * @q: the request queue for the device
695 * @max_size: max size of segment in bytes
698 * Enables a low level driver to set an upper limit on the size of a
701 void blk_queue_max_segment_size(struct request_queue *q, unsigned int max_size)
703 if (max_size < PAGE_CACHE_SIZE) {
704 max_size = PAGE_CACHE_SIZE;
705 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
708 q->max_segment_size = max_size;
711 EXPORT_SYMBOL(blk_queue_max_segment_size);
714 * blk_queue_hardsect_size - set hardware sector size for the queue
715 * @q: the request queue for the device
716 * @size: the hardware sector size, in bytes
719 * This should typically be set to the lowest possible sector size
720 * that the hardware can operate on (possible without reverting to
721 * even internal read-modify-write operations). Usually the default
722 * of 512 covers most hardware.
724 void blk_queue_hardsect_size(struct request_queue *q, unsigned short size)
726 q->hardsect_size = size;
729 EXPORT_SYMBOL(blk_queue_hardsect_size);
732 * Returns the minimum that is _not_ zero, unless both are zero.
734 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
737 * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
738 * @t: the stacking driver (top)
739 * @b: the underlying device (bottom)
741 void blk_queue_stack_limits(struct request_queue *t, struct request_queue *b)
743 /* zero is "infinity" */
744 t->max_sectors = min_not_zero(t->max_sectors,b->max_sectors);
745 t->max_hw_sectors = min_not_zero(t->max_hw_sectors,b->max_hw_sectors);
747 t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
748 t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
749 t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
750 t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
751 if (!test_bit(QUEUE_FLAG_CLUSTER, &b->queue_flags))
752 clear_bit(QUEUE_FLAG_CLUSTER, &t->queue_flags);
755 EXPORT_SYMBOL(blk_queue_stack_limits);
758 * blk_queue_segment_boundary - set boundary rules for segment merging
759 * @q: the request queue for the device
760 * @mask: the memory boundary mask
762 void blk_queue_segment_boundary(struct request_queue *q, unsigned long mask)
764 if (mask < PAGE_CACHE_SIZE - 1) {
765 mask = PAGE_CACHE_SIZE - 1;
766 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
769 q->seg_boundary_mask = mask;
772 EXPORT_SYMBOL(blk_queue_segment_boundary);
775 * blk_queue_dma_alignment - set dma length and memory alignment
776 * @q: the request queue for the device
777 * @mask: alignment mask
780 * set required memory and length aligment for direct dma transactions.
781 * this is used when buiding direct io requests for the queue.
784 void blk_queue_dma_alignment(struct request_queue *q, int mask)
786 q->dma_alignment = mask;
789 EXPORT_SYMBOL(blk_queue_dma_alignment);
792 * blk_queue_find_tag - find a request by its tag and queue
793 * @q: The request queue for the device
794 * @tag: The tag of the request
797 * Should be used when a device returns a tag and you want to match
800 * no locks need be held.
802 struct request *blk_queue_find_tag(struct request_queue *q, int tag)
804 return blk_map_queue_find_tag(q->queue_tags, tag);
807 EXPORT_SYMBOL(blk_queue_find_tag);
810 * __blk_free_tags - release a given set of tag maintenance info
811 * @bqt: the tag map to free
813 * Tries to free the specified @bqt@. Returns true if it was
814 * actually freed and false if there are still references using it
816 static int __blk_free_tags(struct blk_queue_tag *bqt)
820 retval = atomic_dec_and_test(&bqt->refcnt);
823 BUG_ON(!list_empty(&bqt->busy_list));
825 kfree(bqt->tag_index);
826 bqt->tag_index = NULL;
839 * __blk_queue_free_tags - release tag maintenance info
840 * @q: the request queue for the device
843 * blk_cleanup_queue() will take care of calling this function, if tagging
844 * has been used. So there's no need to call this directly.
846 static void __blk_queue_free_tags(struct request_queue *q)
848 struct blk_queue_tag *bqt = q->queue_tags;
853 __blk_free_tags(bqt);
855 q->queue_tags = NULL;
856 q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
861 * blk_free_tags - release a given set of tag maintenance info
862 * @bqt: the tag map to free
864 * For externally managed @bqt@ frees the map. Callers of this
865 * function must guarantee to have released all the queues that
866 * might have been using this tag map.
868 void blk_free_tags(struct blk_queue_tag *bqt)
870 if (unlikely(!__blk_free_tags(bqt)))
873 EXPORT_SYMBOL(blk_free_tags);
876 * blk_queue_free_tags - release tag maintenance info
877 * @q: the request queue for the device
880 * This is used to disabled tagged queuing to a device, yet leave
883 void blk_queue_free_tags(struct request_queue *q)
885 clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
888 EXPORT_SYMBOL(blk_queue_free_tags);
891 init_tag_map(struct request_queue *q, struct blk_queue_tag *tags, int depth)
893 struct request **tag_index;
894 unsigned long *tag_map;
897 if (q && depth > q->nr_requests * 2) {
898 depth = q->nr_requests * 2;
899 printk(KERN_ERR "%s: adjusted depth to %d\n",
900 __FUNCTION__, depth);
903 tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC);
907 nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
908 tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
912 tags->real_max_depth = depth;
913 tags->max_depth = depth;
914 tags->tag_index = tag_index;
915 tags->tag_map = tag_map;
923 static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q,
926 struct blk_queue_tag *tags;
928 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
932 if (init_tag_map(q, tags, depth))
935 INIT_LIST_HEAD(&tags->busy_list);
937 atomic_set(&tags->refcnt, 1);
945 * blk_init_tags - initialize the tag info for an external tag map
946 * @depth: the maximum queue depth supported
947 * @tags: the tag to use
949 struct blk_queue_tag *blk_init_tags(int depth)
951 return __blk_queue_init_tags(NULL, depth);
953 EXPORT_SYMBOL(blk_init_tags);
956 * blk_queue_init_tags - initialize the queue tag info
957 * @q: the request queue for the device
958 * @depth: the maximum queue depth supported
959 * @tags: the tag to use
961 int blk_queue_init_tags(struct request_queue *q, int depth,
962 struct blk_queue_tag *tags)
966 BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
968 if (!tags && !q->queue_tags) {
969 tags = __blk_queue_init_tags(q, depth);
973 } else if (q->queue_tags) {
974 if ((rc = blk_queue_resize_tags(q, depth)))
976 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
979 atomic_inc(&tags->refcnt);
982 * assign it, all done
984 q->queue_tags = tags;
985 q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
992 EXPORT_SYMBOL(blk_queue_init_tags);
995 * blk_queue_resize_tags - change the queueing depth
996 * @q: the request queue for the device
997 * @new_depth: the new max command queueing depth
1000 * Must be called with the queue lock held.
1002 int blk_queue_resize_tags(struct request_queue *q, int new_depth)
1004 struct blk_queue_tag *bqt = q->queue_tags;
1005 struct request **tag_index;
1006 unsigned long *tag_map;
1007 int max_depth, nr_ulongs;
1013 * if we already have large enough real_max_depth. just
1014 * adjust max_depth. *NOTE* as requests with tag value
1015 * between new_depth and real_max_depth can be in-flight, tag
1016 * map can not be shrunk blindly here.
1018 if (new_depth <= bqt->real_max_depth) {
1019 bqt->max_depth = new_depth;
1024 * Currently cannot replace a shared tag map with a new
1025 * one, so error out if this is the case
1027 if (atomic_read(&bqt->refcnt) != 1)
1031 * save the old state info, so we can copy it back
1033 tag_index = bqt->tag_index;
1034 tag_map = bqt->tag_map;
1035 max_depth = bqt->real_max_depth;
1037 if (init_tag_map(q, bqt, new_depth))
1040 memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
1041 nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
1042 memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
1049 EXPORT_SYMBOL(blk_queue_resize_tags);
1052 * blk_queue_end_tag - end tag operations for a request
1053 * @q: the request queue for the device
1054 * @rq: the request that has completed
1057 * Typically called when end_that_request_first() returns 0, meaning
1058 * all transfers have been done for a request. It's important to call
1059 * this function before end_that_request_last(), as that will put the
1060 * request back on the free list thus corrupting the internal tag list.
1063 * queue lock must be held.
1065 void blk_queue_end_tag(struct request_queue *q, struct request *rq)
1067 struct blk_queue_tag *bqt = q->queue_tags;
1072 if (unlikely(tag >= bqt->real_max_depth))
1074 * This can happen after tag depth has been reduced.
1075 * FIXME: how about a warning or info message here?
1079 list_del_init(&rq->queuelist);
1080 rq->cmd_flags &= ~REQ_QUEUED;
1083 if (unlikely(bqt->tag_index[tag] == NULL))
1084 printk(KERN_ERR "%s: tag %d is missing\n",
1087 bqt->tag_index[tag] = NULL;
1090 * We use test_and_clear_bit's memory ordering properties here.
1091 * The tag_map bit acts as a lock for tag_index[bit], so we need
1092 * a barrer before clearing the bit (precisely: release semantics).
1093 * Could use clear_bit_unlock when it is merged.
1095 if (unlikely(!test_and_clear_bit(tag, bqt->tag_map))) {
1096 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
1104 EXPORT_SYMBOL(blk_queue_end_tag);
1107 * blk_queue_start_tag - find a free tag and assign it
1108 * @q: the request queue for the device
1109 * @rq: the block request that needs tagging
1112 * This can either be used as a stand-alone helper, or possibly be
1113 * assigned as the queue &prep_rq_fn (in which case &struct request
1114 * automagically gets a tag assigned). Note that this function
1115 * assumes that any type of request can be queued! if this is not
1116 * true for your device, you must check the request type before
1117 * calling this function. The request will also be removed from
1118 * the request queue, so it's the drivers responsibility to readd
1119 * it if it should need to be restarted for some reason.
1122 * queue lock must be held.
1124 int blk_queue_start_tag(struct request_queue *q, struct request *rq)
1126 struct blk_queue_tag *bqt = q->queue_tags;
1129 if (unlikely((rq->cmd_flags & REQ_QUEUED))) {
1131 "%s: request %p for device [%s] already tagged %d",
1133 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
1138 * Protect against shared tag maps, as we may not have exclusive
1139 * access to the tag map.
1142 tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
1143 if (tag >= bqt->max_depth)
1146 } while (test_and_set_bit(tag, bqt->tag_map));
1148 * We rely on test_and_set_bit providing lock memory ordering semantics
1149 * (could use test_and_set_bit_lock when it is merged).
1152 rq->cmd_flags |= REQ_QUEUED;
1154 bqt->tag_index[tag] = rq;
1155 blkdev_dequeue_request(rq);
1156 list_add(&rq->queuelist, &bqt->busy_list);
1161 EXPORT_SYMBOL(blk_queue_start_tag);
1164 * blk_queue_invalidate_tags - invalidate all pending tags
1165 * @q: the request queue for the device
1168 * Hardware conditions may dictate a need to stop all pending requests.
1169 * In this case, we will safely clear the block side of the tag queue and
1170 * readd all requests to the request queue in the right order.
1173 * queue lock must be held.
1175 void blk_queue_invalidate_tags(struct request_queue *q)
1177 struct blk_queue_tag *bqt = q->queue_tags;
1178 struct list_head *tmp, *n;
1181 list_for_each_safe(tmp, n, &bqt->busy_list) {
1182 rq = list_entry_rq(tmp);
1184 if (rq->tag == -1) {
1186 "%s: bad tag found on list\n", __FUNCTION__);
1187 list_del_init(&rq->queuelist);
1188 rq->cmd_flags &= ~REQ_QUEUED;
1190 blk_queue_end_tag(q, rq);
1192 rq->cmd_flags &= ~REQ_STARTED;
1193 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1197 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1199 void blk_dump_rq_flags(struct request *rq, char *msg)
1203 printk("%s: dev %s: type=%x, flags=%x\n", msg,
1204 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->cmd_type,
1207 printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1209 rq->current_nr_sectors);
1210 printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1212 if (blk_pc_request(rq)) {
1214 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1215 printk("%02x ", rq->cmd[bit]);
1220 EXPORT_SYMBOL(blk_dump_rq_flags);
1222 void blk_recount_segments(struct request_queue *q, struct bio *bio)
1225 struct bio *nxt = bio->bi_next;
1227 rq.bio = rq.biotail = bio;
1228 bio->bi_next = NULL;
1229 blk_recalc_rq_segments(&rq);
1231 bio->bi_phys_segments = rq.nr_phys_segments;
1232 bio->bi_hw_segments = rq.nr_hw_segments;
1233 bio->bi_flags |= (1 << BIO_SEG_VALID);
1235 EXPORT_SYMBOL(blk_recount_segments);
1237 static void blk_recalc_rq_segments(struct request *rq)
1241 unsigned int phys_size;
1242 unsigned int hw_size;
1243 struct bio_vec *bv, *bvprv = NULL;
1247 struct req_iterator iter;
1248 int high, highprv = 1;
1249 struct request_queue *q = rq->q;
1254 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1255 hw_seg_size = seg_size = 0;
1256 phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
1257 rq_for_each_segment(bv, rq, iter) {
1259 * the trick here is making sure that a high page is never
1260 * considered part of another segment, since that might
1261 * change with the bounce page.
1263 high = page_to_pfn(bv->bv_page) > q->bounce_pfn;
1264 if (high || highprv)
1265 goto new_hw_segment;
1267 if (seg_size + bv->bv_len > q->max_segment_size)
1269 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1271 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1273 if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1274 goto new_hw_segment;
1276 seg_size += bv->bv_len;
1277 hw_seg_size += bv->bv_len;
1282 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1283 !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1284 hw_seg_size += bv->bv_len;
1287 if (nr_hw_segs == 1 &&
1288 hw_seg_size > rq->bio->bi_hw_front_size)
1289 rq->bio->bi_hw_front_size = hw_seg_size;
1290 hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1296 seg_size = bv->bv_len;
1300 if (nr_hw_segs == 1 &&
1301 hw_seg_size > rq->bio->bi_hw_front_size)
1302 rq->bio->bi_hw_front_size = hw_seg_size;
1303 if (hw_seg_size > rq->biotail->bi_hw_back_size)
1304 rq->biotail->bi_hw_back_size = hw_seg_size;
1305 rq->nr_phys_segments = nr_phys_segs;
1306 rq->nr_hw_segments = nr_hw_segs;
1309 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
1312 if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1315 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1317 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1321 * bio and nxt are contigous in memory, check if the queue allows
1322 * these two to be merged into one
1324 if (BIO_SEG_BOUNDARY(q, bio, nxt))
1330 static int blk_hw_contig_segment(struct request_queue *q, struct bio *bio,
1333 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1334 blk_recount_segments(q, bio);
1335 if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1336 blk_recount_segments(q, nxt);
1337 if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1338 BIOVEC_VIRT_OVERSIZE(bio->bi_hw_back_size + nxt->bi_hw_front_size))
1340 if (bio->bi_hw_back_size + nxt->bi_hw_front_size > q->max_segment_size)
1347 * map a request to scatterlist, return number of sg entries setup. Caller
1348 * must make sure sg can hold rq->nr_phys_segments entries
1350 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
1351 struct scatterlist *sg)
1353 struct bio_vec *bvec, *bvprv;
1354 struct req_iterator iter;
1358 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1361 * for each bio in rq
1364 rq_for_each_segment(bvec, rq, iter) {
1365 int nbytes = bvec->bv_len;
1367 if (bvprv && cluster) {
1368 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1371 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1373 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1376 sg[nsegs - 1].length += nbytes;
1379 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1380 sg[nsegs].page = bvec->bv_page;
1381 sg[nsegs].length = nbytes;
1382 sg[nsegs].offset = bvec->bv_offset;
1387 } /* segments in rq */
1392 EXPORT_SYMBOL(blk_rq_map_sg);
1395 * the standard queue merge functions, can be overridden with device
1396 * specific ones if so desired
1399 static inline int ll_new_mergeable(struct request_queue *q,
1400 struct request *req,
1403 int nr_phys_segs = bio_phys_segments(q, bio);
1405 if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1406 req->cmd_flags |= REQ_NOMERGE;
1407 if (req == q->last_merge)
1408 q->last_merge = NULL;
1413 * A hw segment is just getting larger, bump just the phys
1416 req->nr_phys_segments += nr_phys_segs;
1420 static inline int ll_new_hw_segment(struct request_queue *q,
1421 struct request *req,
1424 int nr_hw_segs = bio_hw_segments(q, bio);
1425 int nr_phys_segs = bio_phys_segments(q, bio);
1427 if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1428 || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1429 req->cmd_flags |= REQ_NOMERGE;
1430 if (req == q->last_merge)
1431 q->last_merge = NULL;
1436 * This will form the start of a new hw segment. Bump both
1439 req->nr_hw_segments += nr_hw_segs;
1440 req->nr_phys_segments += nr_phys_segs;
1444 static int ll_back_merge_fn(struct request_queue *q, struct request *req,
1447 unsigned short max_sectors;
1450 if (unlikely(blk_pc_request(req)))
1451 max_sectors = q->max_hw_sectors;
1453 max_sectors = q->max_sectors;
1455 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1456 req->cmd_flags |= REQ_NOMERGE;
1457 if (req == q->last_merge)
1458 q->last_merge = NULL;
1461 if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1462 blk_recount_segments(q, req->biotail);
1463 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1464 blk_recount_segments(q, bio);
1465 len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1466 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1467 !BIOVEC_VIRT_OVERSIZE(len)) {
1468 int mergeable = ll_new_mergeable(q, req, bio);
1471 if (req->nr_hw_segments == 1)
1472 req->bio->bi_hw_front_size = len;
1473 if (bio->bi_hw_segments == 1)
1474 bio->bi_hw_back_size = len;
1479 return ll_new_hw_segment(q, req, bio);
1482 static int ll_front_merge_fn(struct request_queue *q, struct request *req,
1485 unsigned short max_sectors;
1488 if (unlikely(blk_pc_request(req)))
1489 max_sectors = q->max_hw_sectors;
1491 max_sectors = q->max_sectors;
1494 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1495 req->cmd_flags |= REQ_NOMERGE;
1496 if (req == q->last_merge)
1497 q->last_merge = NULL;
1500 len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1501 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1502 blk_recount_segments(q, bio);
1503 if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1504 blk_recount_segments(q, req->bio);
1505 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1506 !BIOVEC_VIRT_OVERSIZE(len)) {
1507 int mergeable = ll_new_mergeable(q, req, bio);
1510 if (bio->bi_hw_segments == 1)
1511 bio->bi_hw_front_size = len;
1512 if (req->nr_hw_segments == 1)
1513 req->biotail->bi_hw_back_size = len;
1518 return ll_new_hw_segment(q, req, bio);
1521 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
1522 struct request *next)
1524 int total_phys_segments;
1525 int total_hw_segments;
1528 * First check if the either of the requests are re-queued
1529 * requests. Can't merge them if they are.
1531 if (req->special || next->special)
1535 * Will it become too large?
1537 if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1540 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1541 if (blk_phys_contig_segment(q, req->biotail, next->bio))
1542 total_phys_segments--;
1544 if (total_phys_segments > q->max_phys_segments)
1547 total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1548 if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1549 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1551 * propagate the combined length to the end of the requests
1553 if (req->nr_hw_segments == 1)
1554 req->bio->bi_hw_front_size = len;
1555 if (next->nr_hw_segments == 1)
1556 next->biotail->bi_hw_back_size = len;
1557 total_hw_segments--;
1560 if (total_hw_segments > q->max_hw_segments)
1563 /* Merge is OK... */
1564 req->nr_phys_segments = total_phys_segments;
1565 req->nr_hw_segments = total_hw_segments;
1570 * "plug" the device if there are no outstanding requests: this will
1571 * force the transfer to start only after we have put all the requests
1574 * This is called with interrupts off and no requests on the queue and
1575 * with the queue lock held.
1577 void blk_plug_device(struct request_queue *q)
1579 WARN_ON(!irqs_disabled());
1582 * don't plug a stopped queue, it must be paired with blk_start_queue()
1583 * which will restart the queueing
1585 if (blk_queue_stopped(q))
1588 if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
1589 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1590 blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
1594 EXPORT_SYMBOL(blk_plug_device);
1597 * remove the queue from the plugged list, if present. called with
1598 * queue lock held and interrupts disabled.
1600 int blk_remove_plug(struct request_queue *q)
1602 WARN_ON(!irqs_disabled());
1604 if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1607 del_timer(&q->unplug_timer);
1611 EXPORT_SYMBOL(blk_remove_plug);
1614 * remove the plug and let it rip..
1616 void __generic_unplug_device(struct request_queue *q)
1618 if (unlikely(blk_queue_stopped(q)))
1621 if (!blk_remove_plug(q))
1626 EXPORT_SYMBOL(__generic_unplug_device);
1629 * generic_unplug_device - fire a request queue
1630 * @q: The &struct request_queue in question
1633 * Linux uses plugging to build bigger requests queues before letting
1634 * the device have at them. If a queue is plugged, the I/O scheduler
1635 * is still adding and merging requests on the queue. Once the queue
1636 * gets unplugged, the request_fn defined for the queue is invoked and
1637 * transfers started.
1639 void generic_unplug_device(struct request_queue *q)
1641 spin_lock_irq(q->queue_lock);
1642 __generic_unplug_device(q);
1643 spin_unlock_irq(q->queue_lock);
1645 EXPORT_SYMBOL(generic_unplug_device);
1647 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1650 struct request_queue *q = bdi->unplug_io_data;
1653 * devices don't necessarily have an ->unplug_fn defined
1656 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1657 q->rq.count[READ] + q->rq.count[WRITE]);
1663 static void blk_unplug_work(struct work_struct *work)
1665 struct request_queue *q =
1666 container_of(work, struct request_queue, unplug_work);
1668 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1669 q->rq.count[READ] + q->rq.count[WRITE]);
1674 static void blk_unplug_timeout(unsigned long data)
1676 struct request_queue *q = (struct request_queue *)data;
1678 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
1679 q->rq.count[READ] + q->rq.count[WRITE]);
1681 kblockd_schedule_work(&q->unplug_work);
1685 * blk_start_queue - restart a previously stopped queue
1686 * @q: The &struct request_queue in question
1689 * blk_start_queue() will clear the stop flag on the queue, and call
1690 * the request_fn for the queue if it was in a stopped state when
1691 * entered. Also see blk_stop_queue(). Queue lock must be held.
1693 void blk_start_queue(struct request_queue *q)
1695 WARN_ON(!irqs_disabled());
1697 clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1700 * one level of recursion is ok and is much faster than kicking
1701 * the unplug handling
1703 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1705 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1708 kblockd_schedule_work(&q->unplug_work);
1712 EXPORT_SYMBOL(blk_start_queue);
1715 * blk_stop_queue - stop a queue
1716 * @q: The &struct request_queue in question
1719 * The Linux block layer assumes that a block driver will consume all
1720 * entries on the request queue when the request_fn strategy is called.
1721 * Often this will not happen, because of hardware limitations (queue
1722 * depth settings). If a device driver gets a 'queue full' response,
1723 * or if it simply chooses not to queue more I/O at one point, it can
1724 * call this function to prevent the request_fn from being called until
1725 * the driver has signalled it's ready to go again. This happens by calling
1726 * blk_start_queue() to restart queue operations. Queue lock must be held.
1728 void blk_stop_queue(struct request_queue *q)
1731 set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1733 EXPORT_SYMBOL(blk_stop_queue);
1736 * blk_sync_queue - cancel any pending callbacks on a queue
1740 * The block layer may perform asynchronous callback activity
1741 * on a queue, such as calling the unplug function after a timeout.
1742 * A block device may call blk_sync_queue to ensure that any
1743 * such activity is cancelled, thus allowing it to release resources
1744 * that the callbacks might use. The caller must already have made sure
1745 * that its ->make_request_fn will not re-add plugging prior to calling
1749 void blk_sync_queue(struct request_queue *q)
1751 del_timer_sync(&q->unplug_timer);
1753 EXPORT_SYMBOL(blk_sync_queue);
1756 * blk_run_queue - run a single device queue
1757 * @q: The queue to run
1759 void blk_run_queue(struct request_queue *q)
1761 unsigned long flags;
1763 spin_lock_irqsave(q->queue_lock, flags);
1767 * Only recurse once to avoid overrunning the stack, let the unplug
1768 * handling reinvoke the handler shortly if we already got there.
1770 if (!elv_queue_empty(q)) {
1771 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1773 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1776 kblockd_schedule_work(&q->unplug_work);
1780 spin_unlock_irqrestore(q->queue_lock, flags);
1782 EXPORT_SYMBOL(blk_run_queue);
1785 * blk_cleanup_queue: - release a &struct request_queue when it is no longer needed
1786 * @kobj: the kobj belonging of the request queue to be released
1789 * blk_cleanup_queue is the pair to blk_init_queue() or
1790 * blk_queue_make_request(). It should be called when a request queue is
1791 * being released; typically when a block device is being de-registered.
1792 * Currently, its primary task it to free all the &struct request
1793 * structures that were allocated to the queue and the queue itself.
1796 * Hopefully the low level driver will have finished any
1797 * outstanding requests first...
1799 static void blk_release_queue(struct kobject *kobj)
1801 struct request_queue *q =
1802 container_of(kobj, struct request_queue, kobj);
1803 struct request_list *rl = &q->rq;
1808 mempool_destroy(rl->rq_pool);
1811 __blk_queue_free_tags(q);
1813 blk_trace_shutdown(q);
1815 kmem_cache_free(requestq_cachep, q);
1818 void blk_put_queue(struct request_queue *q)
1820 kobject_put(&q->kobj);
1822 EXPORT_SYMBOL(blk_put_queue);
1824 void blk_cleanup_queue(struct request_queue * q)
1826 mutex_lock(&q->sysfs_lock);
1827 set_bit(QUEUE_FLAG_DEAD, &q->queue_flags);
1828 mutex_unlock(&q->sysfs_lock);
1831 elevator_exit(q->elevator);
1836 EXPORT_SYMBOL(blk_cleanup_queue);
1838 static int blk_init_free_list(struct request_queue *q)
1840 struct request_list *rl = &q->rq;
1842 rl->count[READ] = rl->count[WRITE] = 0;
1843 rl->starved[READ] = rl->starved[WRITE] = 0;
1845 init_waitqueue_head(&rl->wait[READ]);
1846 init_waitqueue_head(&rl->wait[WRITE]);
1848 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1849 mempool_free_slab, request_cachep, q->node);
1857 struct request_queue *blk_alloc_queue(gfp_t gfp_mask)
1859 return blk_alloc_queue_node(gfp_mask, -1);
1861 EXPORT_SYMBOL(blk_alloc_queue);
1863 static struct kobj_type queue_ktype;
1865 struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
1867 struct request_queue *q;
1869 q = kmem_cache_alloc_node(requestq_cachep,
1870 gfp_mask | __GFP_ZERO, node_id);
1874 init_timer(&q->unplug_timer);
1876 snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
1877 q->kobj.ktype = &queue_ktype;
1878 kobject_init(&q->kobj);
1880 q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1881 q->backing_dev_info.unplug_io_data = q;
1883 mutex_init(&q->sysfs_lock);
1887 EXPORT_SYMBOL(blk_alloc_queue_node);
1890 * blk_init_queue - prepare a request queue for use with a block device
1891 * @rfn: The function to be called to process requests that have been
1892 * placed on the queue.
1893 * @lock: Request queue spin lock
1896 * If a block device wishes to use the standard request handling procedures,
1897 * which sorts requests and coalesces adjacent requests, then it must
1898 * call blk_init_queue(). The function @rfn will be called when there
1899 * are requests on the queue that need to be processed. If the device
1900 * supports plugging, then @rfn may not be called immediately when requests
1901 * are available on the queue, but may be called at some time later instead.
1902 * Plugged queues are generally unplugged when a buffer belonging to one
1903 * of the requests on the queue is needed, or due to memory pressure.
1905 * @rfn is not required, or even expected, to remove all requests off the
1906 * queue, but only as many as it can handle at a time. If it does leave
1907 * requests on the queue, it is responsible for arranging that the requests
1908 * get dealt with eventually.
1910 * The queue spin lock must be held while manipulating the requests on the
1911 * request queue; this lock will be taken also from interrupt context, so irq
1912 * disabling is needed for it.
1914 * Function returns a pointer to the initialized request queue, or NULL if
1915 * it didn't succeed.
1918 * blk_init_queue() must be paired with a blk_cleanup_queue() call
1919 * when the block device is deactivated (such as at module unload).
1922 struct request_queue *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1924 return blk_init_queue_node(rfn, lock, -1);
1926 EXPORT_SYMBOL(blk_init_queue);
1928 struct request_queue *
1929 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1931 struct request_queue *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1937 if (blk_init_free_list(q)) {
1938 kmem_cache_free(requestq_cachep, q);
1943 * if caller didn't supply a lock, they get per-queue locking with
1947 spin_lock_init(&q->__queue_lock);
1948 lock = &q->__queue_lock;
1951 q->request_fn = rfn;
1952 q->prep_rq_fn = NULL;
1953 q->unplug_fn = generic_unplug_device;
1954 q->queue_flags = (1 << QUEUE_FLAG_CLUSTER);
1955 q->queue_lock = lock;
1957 blk_queue_segment_boundary(q, 0xffffffff);
1959 blk_queue_make_request(q, __make_request);
1960 blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1962 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1963 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1965 q->sg_reserved_size = INT_MAX;
1970 if (!elevator_init(q, NULL)) {
1971 blk_queue_congestion_threshold(q);
1978 EXPORT_SYMBOL(blk_init_queue_node);
1980 int blk_get_queue(struct request_queue *q)
1982 if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1983 kobject_get(&q->kobj);
1990 EXPORT_SYMBOL(blk_get_queue);
1992 static inline void blk_free_request(struct request_queue *q, struct request *rq)
1994 if (rq->cmd_flags & REQ_ELVPRIV)
1995 elv_put_request(q, rq);
1996 mempool_free(rq, q->rq.rq_pool);
1999 static struct request *
2000 blk_alloc_request(struct request_queue *q, int rw, int priv, gfp_t gfp_mask)
2002 struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
2008 * first three bits are identical in rq->cmd_flags and bio->bi_rw,
2009 * see bio.h and blkdev.h
2011 rq->cmd_flags = rw | REQ_ALLOCED;
2014 if (unlikely(elv_set_request(q, rq, gfp_mask))) {
2015 mempool_free(rq, q->rq.rq_pool);
2018 rq->cmd_flags |= REQ_ELVPRIV;
2025 * ioc_batching returns true if the ioc is a valid batching request and
2026 * should be given priority access to a request.
2028 static inline int ioc_batching(struct request_queue *q, struct io_context *ioc)
2034 * Make sure the process is able to allocate at least 1 request
2035 * even if the batch times out, otherwise we could theoretically
2038 return ioc->nr_batch_requests == q->nr_batching ||
2039 (ioc->nr_batch_requests > 0
2040 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
2044 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
2045 * will cause the process to be a "batcher" on all queues in the system. This
2046 * is the behaviour we want though - once it gets a wakeup it should be given
2049 static void ioc_set_batching(struct request_queue *q, struct io_context *ioc)
2051 if (!ioc || ioc_batching(q, ioc))
2054 ioc->nr_batch_requests = q->nr_batching;
2055 ioc->last_waited = jiffies;
2058 static void __freed_request(struct request_queue *q, int rw)
2060 struct request_list *rl = &q->rq;
2062 if (rl->count[rw] < queue_congestion_off_threshold(q))
2063 blk_clear_queue_congested(q, rw);
2065 if (rl->count[rw] + 1 <= q->nr_requests) {
2066 if (waitqueue_active(&rl->wait[rw]))
2067 wake_up(&rl->wait[rw]);
2069 blk_clear_queue_full(q, rw);
2074 * A request has just been released. Account for it, update the full and
2075 * congestion status, wake up any waiters. Called under q->queue_lock.
2077 static void freed_request(struct request_queue *q, int rw, int priv)
2079 struct request_list *rl = &q->rq;
2085 __freed_request(q, rw);
2087 if (unlikely(rl->starved[rw ^ 1]))
2088 __freed_request(q, rw ^ 1);
2091 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2093 * Get a free request, queue_lock must be held.
2094 * Returns NULL on failure, with queue_lock held.
2095 * Returns !NULL on success, with queue_lock *not held*.
2097 static struct request *get_request(struct request_queue *q, int rw_flags,
2098 struct bio *bio, gfp_t gfp_mask)
2100 struct request *rq = NULL;
2101 struct request_list *rl = &q->rq;
2102 struct io_context *ioc = NULL;
2103 const int rw = rw_flags & 0x01;
2104 int may_queue, priv;
2106 may_queue = elv_may_queue(q, rw_flags);
2107 if (may_queue == ELV_MQUEUE_NO)
2110 if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
2111 if (rl->count[rw]+1 >= q->nr_requests) {
2112 ioc = current_io_context(GFP_ATOMIC, q->node);
2114 * The queue will fill after this allocation, so set
2115 * it as full, and mark this process as "batching".
2116 * This process will be allowed to complete a batch of
2117 * requests, others will be blocked.
2119 if (!blk_queue_full(q, rw)) {
2120 ioc_set_batching(q, ioc);
2121 blk_set_queue_full(q, rw);
2123 if (may_queue != ELV_MQUEUE_MUST
2124 && !ioc_batching(q, ioc)) {
2126 * The queue is full and the allocating
2127 * process is not a "batcher", and not
2128 * exempted by the IO scheduler
2134 blk_set_queue_congested(q, rw);
2138 * Only allow batching queuers to allocate up to 50% over the defined
2139 * limit of requests, otherwise we could have thousands of requests
2140 * allocated with any setting of ->nr_requests
2142 if (rl->count[rw] >= (3 * q->nr_requests / 2))
2146 rl->starved[rw] = 0;
2148 priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
2152 spin_unlock_irq(q->queue_lock);
2154 rq = blk_alloc_request(q, rw_flags, priv, gfp_mask);
2155 if (unlikely(!rq)) {
2157 * Allocation failed presumably due to memory. Undo anything
2158 * we might have messed up.
2160 * Allocating task should really be put onto the front of the
2161 * wait queue, but this is pretty rare.
2163 spin_lock_irq(q->queue_lock);
2164 freed_request(q, rw, priv);
2167 * in the very unlikely event that allocation failed and no
2168 * requests for this direction was pending, mark us starved
2169 * so that freeing of a request in the other direction will
2170 * notice us. another possible fix would be to split the
2171 * rq mempool into READ and WRITE
2174 if (unlikely(rl->count[rw] == 0))
2175 rl->starved[rw] = 1;
2181 * ioc may be NULL here, and ioc_batching will be false. That's
2182 * OK, if the queue is under the request limit then requests need
2183 * not count toward the nr_batch_requests limit. There will always
2184 * be some limit enforced by BLK_BATCH_TIME.
2186 if (ioc_batching(q, ioc))
2187 ioc->nr_batch_requests--;
2191 blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
2197 * No available requests for this queue, unplug the device and wait for some
2198 * requests to become available.
2200 * Called with q->queue_lock held, and returns with it unlocked.
2202 static struct request *get_request_wait(struct request_queue *q, int rw_flags,
2205 const int rw = rw_flags & 0x01;
2208 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2211 struct request_list *rl = &q->rq;
2213 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2214 TASK_UNINTERRUPTIBLE);
2216 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2219 struct io_context *ioc;
2221 blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
2223 __generic_unplug_device(q);
2224 spin_unlock_irq(q->queue_lock);
2228 * After sleeping, we become a "batching" process and
2229 * will be able to allocate at least one request, and
2230 * up to a big batch of them for a small period time.
2231 * See ioc_batching, ioc_set_batching
2233 ioc = current_io_context(GFP_NOIO, q->node);
2234 ioc_set_batching(q, ioc);
2236 spin_lock_irq(q->queue_lock);
2238 finish_wait(&rl->wait[rw], &wait);
2244 struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
2248 BUG_ON(rw != READ && rw != WRITE);
2250 spin_lock_irq(q->queue_lock);
2251 if (gfp_mask & __GFP_WAIT) {
2252 rq = get_request_wait(q, rw, NULL);
2254 rq = get_request(q, rw, NULL, gfp_mask);
2256 spin_unlock_irq(q->queue_lock);
2258 /* q->queue_lock is unlocked at this point */
2262 EXPORT_SYMBOL(blk_get_request);
2265 * blk_start_queueing - initiate dispatch of requests to device
2266 * @q: request queue to kick into gear
2268 * This is basically a helper to remove the need to know whether a queue
2269 * is plugged or not if someone just wants to initiate dispatch of requests
2272 * The queue lock must be held with interrupts disabled.
2274 void blk_start_queueing(struct request_queue *q)
2276 if (!blk_queue_plugged(q))
2279 __generic_unplug_device(q);
2281 EXPORT_SYMBOL(blk_start_queueing);
2284 * blk_requeue_request - put a request back on queue
2285 * @q: request queue where request should be inserted
2286 * @rq: request to be inserted
2289 * Drivers often keep queueing requests until the hardware cannot accept
2290 * more, when that condition happens we need to put the request back
2291 * on the queue. Must be called with queue lock held.
2293 void blk_requeue_request(struct request_queue *q, struct request *rq)
2295 blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
2297 if (blk_rq_tagged(rq))
2298 blk_queue_end_tag(q, rq);
2300 elv_requeue_request(q, rq);
2303 EXPORT_SYMBOL(blk_requeue_request);
2306 * blk_insert_request - insert a special request in to a request queue
2307 * @q: request queue where request should be inserted
2308 * @rq: request to be inserted
2309 * @at_head: insert request at head or tail of queue
2310 * @data: private data
2313 * Many block devices need to execute commands asynchronously, so they don't
2314 * block the whole kernel from preemption during request execution. This is
2315 * accomplished normally by inserting aritficial requests tagged as
2316 * REQ_SPECIAL in to the corresponding request queue, and letting them be
2317 * scheduled for actual execution by the request queue.
2319 * We have the option of inserting the head or the tail of the queue.
2320 * Typically we use the tail for new ioctls and so forth. We use the head
2321 * of the queue for things like a QUEUE_FULL message from a device, or a
2322 * host that is unable to accept a particular command.
2324 void blk_insert_request(struct request_queue *q, struct request *rq,
2325 int at_head, void *data)
2327 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2328 unsigned long flags;
2331 * tell I/O scheduler that this isn't a regular read/write (ie it
2332 * must not attempt merges on this) and that it acts as a soft
2335 rq->cmd_type = REQ_TYPE_SPECIAL;
2336 rq->cmd_flags |= REQ_SOFTBARRIER;
2340 spin_lock_irqsave(q->queue_lock, flags);
2343 * If command is tagged, release the tag
2345 if (blk_rq_tagged(rq))
2346 blk_queue_end_tag(q, rq);
2348 drive_stat_acct(rq, rq->nr_sectors, 1);
2349 __elv_add_request(q, rq, where, 0);
2350 blk_start_queueing(q);
2351 spin_unlock_irqrestore(q->queue_lock, flags);
2354 EXPORT_SYMBOL(blk_insert_request);
2356 static int __blk_rq_unmap_user(struct bio *bio)
2361 if (bio_flagged(bio, BIO_USER_MAPPED))
2362 bio_unmap_user(bio);
2364 ret = bio_uncopy_user(bio);
2370 int blk_rq_append_bio(struct request_queue *q, struct request *rq,
2374 blk_rq_bio_prep(q, rq, bio);
2375 else if (!ll_back_merge_fn(q, rq, bio))
2378 rq->biotail->bi_next = bio;
2381 rq->data_len += bio->bi_size;
2385 EXPORT_SYMBOL(blk_rq_append_bio);
2387 static int __blk_rq_map_user(struct request_queue *q, struct request *rq,
2388 void __user *ubuf, unsigned int len)
2390 unsigned long uaddr;
2391 struct bio *bio, *orig_bio;
2394 reading = rq_data_dir(rq) == READ;
2397 * if alignment requirement is satisfied, map in user pages for
2398 * direct dma. else, set up kernel bounce buffers
2400 uaddr = (unsigned long) ubuf;
2401 if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2402 bio = bio_map_user(q, NULL, uaddr, len, reading);
2404 bio = bio_copy_user(q, uaddr, len, reading);
2407 return PTR_ERR(bio);
2410 blk_queue_bounce(q, &bio);
2413 * We link the bounce buffer in and could have to traverse it
2414 * later so we have to get a ref to prevent it from being freed
2418 ret = blk_rq_append_bio(q, rq, bio);
2420 return bio->bi_size;
2422 /* if it was boucned we must call the end io function */
2423 bio_endio(bio, bio->bi_size, 0);
2424 __blk_rq_unmap_user(orig_bio);
2430 * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2431 * @q: request queue where request should be inserted
2432 * @rq: request structure to fill
2433 * @ubuf: the user buffer
2434 * @len: length of user data
2437 * Data will be mapped directly for zero copy io, if possible. Otherwise
2438 * a kernel bounce buffer is used.
2440 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2441 * still in process context.
2443 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2444 * before being submitted to the device, as pages mapped may be out of
2445 * reach. It's the callers responsibility to make sure this happens. The
2446 * original bio must be passed back in to blk_rq_unmap_user() for proper
2449 int blk_rq_map_user(struct request_queue *q, struct request *rq,
2450 void __user *ubuf, unsigned long len)
2452 unsigned long bytes_read = 0;
2453 struct bio *bio = NULL;
2456 if (len > (q->max_hw_sectors << 9))
2461 while (bytes_read != len) {
2462 unsigned long map_len, end, start;
2464 map_len = min_t(unsigned long, len - bytes_read, BIO_MAX_SIZE);
2465 end = ((unsigned long)ubuf + map_len + PAGE_SIZE - 1)
2467 start = (unsigned long)ubuf >> PAGE_SHIFT;
2470 * A bad offset could cause us to require BIO_MAX_PAGES + 1
2471 * pages. If this happens we just lower the requested
2472 * mapping len by a page so that we can fit
2474 if (end - start > BIO_MAX_PAGES)
2475 map_len -= PAGE_SIZE;
2477 ret = __blk_rq_map_user(q, rq, ubuf, map_len);
2486 rq->buffer = rq->data = NULL;
2489 blk_rq_unmap_user(bio);
2493 EXPORT_SYMBOL(blk_rq_map_user);
2496 * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2497 * @q: request queue where request should be inserted
2498 * @rq: request to map data to
2499 * @iov: pointer to the iovec
2500 * @iov_count: number of elements in the iovec
2501 * @len: I/O byte count
2504 * Data will be mapped directly for zero copy io, if possible. Otherwise
2505 * a kernel bounce buffer is used.
2507 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2508 * still in process context.
2510 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2511 * before being submitted to the device, as pages mapped may be out of
2512 * reach. It's the callers responsibility to make sure this happens. The
2513 * original bio must be passed back in to blk_rq_unmap_user() for proper
2516 int blk_rq_map_user_iov(struct request_queue *q, struct request *rq,
2517 struct sg_iovec *iov, int iov_count, unsigned int len)
2521 if (!iov || iov_count <= 0)
2524 /* we don't allow misaligned data like bio_map_user() does. If the
2525 * user is using sg, they're expected to know the alignment constraints
2526 * and respect them accordingly */
2527 bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2529 return PTR_ERR(bio);
2531 if (bio->bi_size != len) {
2532 bio_endio(bio, bio->bi_size, 0);
2533 bio_unmap_user(bio);
2538 blk_rq_bio_prep(q, rq, bio);
2539 rq->buffer = rq->data = NULL;
2543 EXPORT_SYMBOL(blk_rq_map_user_iov);
2546 * blk_rq_unmap_user - unmap a request with user data
2547 * @bio: start of bio list
2550 * Unmap a rq previously mapped by blk_rq_map_user(). The caller must
2551 * supply the original rq->bio from the blk_rq_map_user() return, since
2552 * the io completion may have changed rq->bio.
2554 int blk_rq_unmap_user(struct bio *bio)
2556 struct bio *mapped_bio;
2561 if (unlikely(bio_flagged(bio, BIO_BOUNCED)))
2562 mapped_bio = bio->bi_private;
2564 ret2 = __blk_rq_unmap_user(mapped_bio);
2570 bio_put(mapped_bio);
2576 EXPORT_SYMBOL(blk_rq_unmap_user);
2579 * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2580 * @q: request queue where request should be inserted
2581 * @rq: request to fill
2582 * @kbuf: the kernel buffer
2583 * @len: length of user data
2584 * @gfp_mask: memory allocation flags
2586 int blk_rq_map_kern(struct request_queue *q, struct request *rq, void *kbuf,
2587 unsigned int len, gfp_t gfp_mask)
2591 if (len > (q->max_hw_sectors << 9))
2596 bio = bio_map_kern(q, kbuf, len, gfp_mask);
2598 return PTR_ERR(bio);
2600 if (rq_data_dir(rq) == WRITE)
2601 bio->bi_rw |= (1 << BIO_RW);
2603 blk_rq_bio_prep(q, rq, bio);
2604 blk_queue_bounce(q, &rq->bio);
2605 rq->buffer = rq->data = NULL;
2609 EXPORT_SYMBOL(blk_rq_map_kern);
2612 * blk_execute_rq_nowait - insert a request into queue for execution
2613 * @q: queue to insert the request in
2614 * @bd_disk: matching gendisk
2615 * @rq: request to insert
2616 * @at_head: insert request at head or tail of queue
2617 * @done: I/O completion handler
2620 * Insert a fully prepared request at the back of the io scheduler queue
2621 * for execution. Don't wait for completion.
2623 void blk_execute_rq_nowait(struct request_queue *q, struct gendisk *bd_disk,
2624 struct request *rq, int at_head,
2627 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2629 rq->rq_disk = bd_disk;
2630 rq->cmd_flags |= REQ_NOMERGE;
2632 WARN_ON(irqs_disabled());
2633 spin_lock_irq(q->queue_lock);
2634 __elv_add_request(q, rq, where, 1);
2635 __generic_unplug_device(q);
2636 spin_unlock_irq(q->queue_lock);
2638 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
2641 * blk_execute_rq - insert a request into queue for execution
2642 * @q: queue to insert the request in
2643 * @bd_disk: matching gendisk
2644 * @rq: request to insert
2645 * @at_head: insert request at head or tail of queue
2648 * Insert a fully prepared request at the back of the io scheduler queue
2649 * for execution and wait for completion.
2651 int blk_execute_rq(struct request_queue *q, struct gendisk *bd_disk,
2652 struct request *rq, int at_head)
2654 DECLARE_COMPLETION_ONSTACK(wait);
2655 char sense[SCSI_SENSE_BUFFERSIZE];
2659 * we need an extra reference to the request, so we can look at
2660 * it after io completion
2665 memset(sense, 0, sizeof(sense));
2670 rq->end_io_data = &wait;
2671 blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
2672 wait_for_completion(&wait);
2680 EXPORT_SYMBOL(blk_execute_rq);
2683 * blkdev_issue_flush - queue a flush
2684 * @bdev: blockdev to issue flush for
2685 * @error_sector: error sector
2688 * Issue a flush for the block device in question. Caller can supply
2689 * room for storing the error offset in case of a flush error, if they
2690 * wish to. Caller must run wait_for_completion() on its own.
2692 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2694 struct request_queue *q;
2696 if (bdev->bd_disk == NULL)
2699 q = bdev_get_queue(bdev);
2702 if (!q->issue_flush_fn)
2705 return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2708 EXPORT_SYMBOL(blkdev_issue_flush);
2710 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2712 int rw = rq_data_dir(rq);
2714 if (!blk_fs_request(rq) || !rq->rq_disk)
2718 __disk_stat_inc(rq->rq_disk, merges[rw]);
2720 disk_round_stats(rq->rq_disk);
2721 rq->rq_disk->in_flight++;
2726 * add-request adds a request to the linked list.
2727 * queue lock is held and interrupts disabled, as we muck with the
2728 * request queue list.
2730 static inline void add_request(struct request_queue * q, struct request * req)
2732 drive_stat_acct(req, req->nr_sectors, 1);
2735 * elevator indicated where it wants this request to be
2736 * inserted at elevator_merge time
2738 __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2742 * disk_round_stats() - Round off the performance stats on a struct
2745 * The average IO queue length and utilisation statistics are maintained
2746 * by observing the current state of the queue length and the amount of
2747 * time it has been in this state for.
2749 * Normally, that accounting is done on IO completion, but that can result
2750 * in more than a second's worth of IO being accounted for within any one
2751 * second, leading to >100% utilisation. To deal with that, we call this
2752 * function to do a round-off before returning the results when reading
2753 * /proc/diskstats. This accounts immediately for all queue usage up to
2754 * the current jiffies and restarts the counters again.
2756 void disk_round_stats(struct gendisk *disk)
2758 unsigned long now = jiffies;
2760 if (now == disk->stamp)
2763 if (disk->in_flight) {
2764 __disk_stat_add(disk, time_in_queue,
2765 disk->in_flight * (now - disk->stamp));
2766 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2771 EXPORT_SYMBOL_GPL(disk_round_stats);
2774 * queue lock must be held
2776 void __blk_put_request(struct request_queue *q, struct request *req)
2780 if (unlikely(--req->ref_count))
2783 elv_completed_request(q, req);
2786 * Request may not have originated from ll_rw_blk. if not,
2787 * it didn't come out of our reserved rq pools
2789 if (req->cmd_flags & REQ_ALLOCED) {
2790 int rw = rq_data_dir(req);
2791 int priv = req->cmd_flags & REQ_ELVPRIV;
2793 BUG_ON(!list_empty(&req->queuelist));
2794 BUG_ON(!hlist_unhashed(&req->hash));
2796 blk_free_request(q, req);
2797 freed_request(q, rw, priv);
2801 EXPORT_SYMBOL_GPL(__blk_put_request);
2803 void blk_put_request(struct request *req)
2805 unsigned long flags;
2806 struct request_queue *q = req->q;
2809 * Gee, IDE calls in w/ NULL q. Fix IDE and remove the
2810 * following if (q) test.
2813 spin_lock_irqsave(q->queue_lock, flags);
2814 __blk_put_request(q, req);
2815 spin_unlock_irqrestore(q->queue_lock, flags);
2819 EXPORT_SYMBOL(blk_put_request);
2822 * blk_end_sync_rq - executes a completion event on a request
2823 * @rq: request to complete
2824 * @error: end io status of the request
2826 void blk_end_sync_rq(struct request *rq, int error)
2828 struct completion *waiting = rq->end_io_data;
2830 rq->end_io_data = NULL;
2831 __blk_put_request(rq->q, rq);
2834 * complete last, if this is a stack request the process (and thus
2835 * the rq pointer) could be invalid right after this complete()
2839 EXPORT_SYMBOL(blk_end_sync_rq);
2842 * Has to be called with the request spinlock acquired
2844 static int attempt_merge(struct request_queue *q, struct request *req,
2845 struct request *next)
2847 if (!rq_mergeable(req) || !rq_mergeable(next))
2853 if (req->sector + req->nr_sectors != next->sector)
2856 if (rq_data_dir(req) != rq_data_dir(next)
2857 || req->rq_disk != next->rq_disk
2862 * If we are allowed to merge, then append bio list
2863 * from next to rq and release next. merge_requests_fn
2864 * will have updated segment counts, update sector
2867 if (!ll_merge_requests_fn(q, req, next))
2871 * At this point we have either done a back merge
2872 * or front merge. We need the smaller start_time of
2873 * the merged requests to be the current request
2874 * for accounting purposes.
2876 if (time_after(req->start_time, next->start_time))
2877 req->start_time = next->start_time;
2879 req->biotail->bi_next = next->bio;
2880 req->biotail = next->biotail;
2882 req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2884 elv_merge_requests(q, req, next);
2887 disk_round_stats(req->rq_disk);
2888 req->rq_disk->in_flight--;
2891 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2893 __blk_put_request(q, next);
2897 static inline int attempt_back_merge(struct request_queue *q,
2900 struct request *next = elv_latter_request(q, rq);
2903 return attempt_merge(q, rq, next);
2908 static inline int attempt_front_merge(struct request_queue *q,
2911 struct request *prev = elv_former_request(q, rq);
2914 return attempt_merge(q, prev, rq);
2919 static void init_request_from_bio(struct request *req, struct bio *bio)
2921 req->cmd_type = REQ_TYPE_FS;
2924 * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2926 if (bio_rw_ahead(bio) || bio_failfast(bio))
2927 req->cmd_flags |= REQ_FAILFAST;
2930 * REQ_BARRIER implies no merging, but lets make it explicit
2932 if (unlikely(bio_barrier(bio)))
2933 req->cmd_flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2936 req->cmd_flags |= REQ_RW_SYNC;
2937 if (bio_rw_meta(bio))
2938 req->cmd_flags |= REQ_RW_META;
2941 req->hard_sector = req->sector = bio->bi_sector;
2942 req->hard_nr_sectors = req->nr_sectors = bio_sectors(bio);
2943 req->current_nr_sectors = req->hard_cur_sectors = bio_cur_sectors(bio);
2944 req->nr_phys_segments = bio_phys_segments(req->q, bio);
2945 req->nr_hw_segments = bio_hw_segments(req->q, bio);
2946 req->buffer = bio_data(bio); /* see ->buffer comment above */
2947 req->bio = req->biotail = bio;
2948 req->ioprio = bio_prio(bio);
2949 req->rq_disk = bio->bi_bdev->bd_disk;
2950 req->start_time = jiffies;
2953 static int __make_request(struct request_queue *q, struct bio *bio)
2955 struct request *req;
2956 int el_ret, nr_sectors, barrier, err;
2957 const unsigned short prio = bio_prio(bio);
2958 const int sync = bio_sync(bio);
2961 nr_sectors = bio_sectors(bio);
2964 * low level driver can indicate that it wants pages above a
2965 * certain limit bounced to low memory (ie for highmem, or even
2966 * ISA dma in theory)
2968 blk_queue_bounce(q, &bio);
2970 barrier = bio_barrier(bio);
2971 if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
2976 spin_lock_irq(q->queue_lock);
2978 if (unlikely(barrier) || elv_queue_empty(q))
2981 el_ret = elv_merge(q, &req, bio);
2983 case ELEVATOR_BACK_MERGE:
2984 BUG_ON(!rq_mergeable(req));
2986 if (!ll_back_merge_fn(q, req, bio))
2989 blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
2991 req->biotail->bi_next = bio;
2993 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2994 req->ioprio = ioprio_best(req->ioprio, prio);
2995 drive_stat_acct(req, nr_sectors, 0);
2996 if (!attempt_back_merge(q, req))
2997 elv_merged_request(q, req, el_ret);
3000 case ELEVATOR_FRONT_MERGE:
3001 BUG_ON(!rq_mergeable(req));
3003 if (!ll_front_merge_fn(q, req, bio))
3006 blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
3008 bio->bi_next = req->bio;
3012 * may not be valid. if the low level driver said
3013 * it didn't need a bounce buffer then it better
3014 * not touch req->buffer either...
3016 req->buffer = bio_data(bio);
3017 req->current_nr_sectors = bio_cur_sectors(bio);
3018 req->hard_cur_sectors = req->current_nr_sectors;
3019 req->sector = req->hard_sector = bio->bi_sector;
3020 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
3021 req->ioprio = ioprio_best(req->ioprio, prio);
3022 drive_stat_acct(req, nr_sectors, 0);
3023 if (!attempt_front_merge(q, req))
3024 elv_merged_request(q, req, el_ret);
3027 /* ELV_NO_MERGE: elevator says don't/can't merge. */
3034 * This sync check and mask will be re-done in init_request_from_bio(),
3035 * but we need to set it earlier to expose the sync flag to the
3036 * rq allocator and io schedulers.
3038 rw_flags = bio_data_dir(bio);
3040 rw_flags |= REQ_RW_SYNC;
3043 * Grab a free request. This is might sleep but can not fail.
3044 * Returns with the queue unlocked.
3046 req = get_request_wait(q, rw_flags, bio);
3049 * After dropping the lock and possibly sleeping here, our request
3050 * may now be mergeable after it had proven unmergeable (above).
3051 * We don't worry about that case for efficiency. It won't happen
3052 * often, and the elevators are able to handle it.
3054 init_request_from_bio(req, bio);
3056 spin_lock_irq(q->queue_lock);
3057 if (elv_queue_empty(q))
3059 add_request(q, req);
3062 __generic_unplug_device(q);
3064 spin_unlock_irq(q->queue_lock);
3068 bio_endio(bio, nr_sectors << 9, err);
3073 * If bio->bi_dev is a partition, remap the location
3075 static inline void blk_partition_remap(struct bio *bio)
3077 struct block_device *bdev = bio->bi_bdev;
3079 if (bdev != bdev->bd_contains) {
3080 struct hd_struct *p = bdev->bd_part;
3081 const int rw = bio_data_dir(bio);
3083 p->sectors[rw] += bio_sectors(bio);
3086 bio->bi_sector += p->start_sect;
3087 bio->bi_bdev = bdev->bd_contains;
3089 blk_add_trace_remap(bdev_get_queue(bio->bi_bdev), bio,
3090 bdev->bd_dev, bio->bi_sector,
3091 bio->bi_sector - p->start_sect);
3095 static void handle_bad_sector(struct bio *bio)
3097 char b[BDEVNAME_SIZE];
3099 printk(KERN_INFO "attempt to access beyond end of device\n");
3100 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
3101 bdevname(bio->bi_bdev, b),
3103 (unsigned long long)bio->bi_sector + bio_sectors(bio),
3104 (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
3106 set_bit(BIO_EOF, &bio->bi_flags);
3109 #ifdef CONFIG_FAIL_MAKE_REQUEST
3111 static DECLARE_FAULT_ATTR(fail_make_request);
3113 static int __init setup_fail_make_request(char *str)
3115 return setup_fault_attr(&fail_make_request, str);
3117 __setup("fail_make_request=", setup_fail_make_request);
3119 static int should_fail_request(struct bio *bio)
3121 if ((bio->bi_bdev->bd_disk->flags & GENHD_FL_FAIL) ||
3122 (bio->bi_bdev->bd_part && bio->bi_bdev->bd_part->make_it_fail))
3123 return should_fail(&fail_make_request, bio->bi_size);
3128 static int __init fail_make_request_debugfs(void)
3130 return init_fault_attr_dentries(&fail_make_request,
3131 "fail_make_request");
3134 late_initcall(fail_make_request_debugfs);
3136 #else /* CONFIG_FAIL_MAKE_REQUEST */
3138 static inline int should_fail_request(struct bio *bio)
3143 #endif /* CONFIG_FAIL_MAKE_REQUEST */
3146 * generic_make_request: hand a buffer to its device driver for I/O
3147 * @bio: The bio describing the location in memory and on the device.
3149 * generic_make_request() is used to make I/O requests of block
3150 * devices. It is passed a &struct bio, which describes the I/O that needs
3153 * generic_make_request() does not return any status. The
3154 * success/failure status of the request, along with notification of
3155 * completion, is delivered asynchronously through the bio->bi_end_io
3156 * function described (one day) else where.
3158 * The caller of generic_make_request must make sure that bi_io_vec
3159 * are set to describe the memory buffer, and that bi_dev and bi_sector are
3160 * set to describe the device address, and the
3161 * bi_end_io and optionally bi_private are set to describe how
3162 * completion notification should be signaled.
3164 * generic_make_request and the drivers it calls may use bi_next if this
3165 * bio happens to be merged with someone else, and may change bi_dev and
3166 * bi_sector for remaps as it sees fit. So the values of these fields
3167 * should NOT be depended on after the call to generic_make_request.
3169 static inline void __generic_make_request(struct bio *bio)
3171 struct request_queue *q;
3173 sector_t old_sector;
3174 int ret, nr_sectors = bio_sectors(bio);
3178 /* Test device or partition size, when known. */
3179 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3181 sector_t sector = bio->bi_sector;
3183 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
3185 * This may well happen - the kernel calls bread()
3186 * without checking the size of the device, e.g., when
3187 * mounting a device.
3189 handle_bad_sector(bio);
3195 * Resolve the mapping until finished. (drivers are
3196 * still free to implement/resolve their own stacking
3197 * by explicitly returning 0)
3199 * NOTE: we don't repeat the blk_size check for each new device.
3200 * Stacking drivers are expected to know what they are doing.
3205 char b[BDEVNAME_SIZE];
3207 q = bdev_get_queue(bio->bi_bdev);
3210 "generic_make_request: Trying to access "
3211 "nonexistent block-device %s (%Lu)\n",
3212 bdevname(bio->bi_bdev, b),
3213 (long long) bio->bi_sector);
3215 bio_endio(bio, bio->bi_size, -EIO);
3219 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
3220 printk("bio too big device %s (%u > %u)\n",
3221 bdevname(bio->bi_bdev, b),
3227 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
3230 if (should_fail_request(bio))
3234 * If this device has partitions, remap block n
3235 * of partition p to block n+start(p) of the disk.
3237 blk_partition_remap(bio);
3239 if (old_sector != -1)
3240 blk_add_trace_remap(q, bio, old_dev, bio->bi_sector,
3243 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
3245 old_sector = bio->bi_sector;
3246 old_dev = bio->bi_bdev->bd_dev;
3248 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3250 sector_t sector = bio->bi_sector;
3252 if (maxsector < nr_sectors ||
3253 maxsector - nr_sectors < sector) {
3255 * This may well happen - partitions are not
3256 * checked to make sure they are within the size
3257 * of the whole device.
3259 handle_bad_sector(bio);
3264 ret = q->make_request_fn(q, bio);
3269 * We only want one ->make_request_fn to be active at a time,
3270 * else stack usage with stacked devices could be a problem.
3271 * So use current->bio_{list,tail} to keep a list of requests
3272 * submited by a make_request_fn function.
3273 * current->bio_tail is also used as a flag to say if
3274 * generic_make_request is currently active in this task or not.
3275 * If it is NULL, then no make_request is active. If it is non-NULL,
3276 * then a make_request is active, and new requests should be added
3279 void generic_make_request(struct bio *bio)
3281 if (current->bio_tail) {
3282 /* make_request is active */
3283 *(current->bio_tail) = bio;
3284 bio->bi_next = NULL;
3285 current->bio_tail = &bio->bi_next;
3288 /* following loop may be a bit non-obvious, and so deserves some
3290 * Before entering the loop, bio->bi_next is NULL (as all callers
3291 * ensure that) so we have a list with a single bio.
3292 * We pretend that we have just taken it off a longer list, so
3293 * we assign bio_list to the next (which is NULL) and bio_tail
3294 * to &bio_list, thus initialising the bio_list of new bios to be
3295 * added. __generic_make_request may indeed add some more bios
3296 * through a recursive call to generic_make_request. If it
3297 * did, we find a non-NULL value in bio_list and re-enter the loop
3298 * from the top. In this case we really did just take the bio
3299 * of the top of the list (no pretending) and so fixup bio_list and
3300 * bio_tail or bi_next, and call into __generic_make_request again.
3302 * The loop was structured like this to make only one call to
3303 * __generic_make_request (which is important as it is large and
3304 * inlined) and to keep the structure simple.
3306 BUG_ON(bio->bi_next);
3308 current->bio_list = bio->bi_next;
3309 if (bio->bi_next == NULL)
3310 current->bio_tail = ¤t->bio_list;
3312 bio->bi_next = NULL;
3313 __generic_make_request(bio);
3314 bio = current->bio_list;
3316 current->bio_tail = NULL; /* deactivate */
3319 EXPORT_SYMBOL(generic_make_request);
3322 * submit_bio: submit a bio to the block device layer for I/O
3323 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3324 * @bio: The &struct bio which describes the I/O
3326 * submit_bio() is very similar in purpose to generic_make_request(), and
3327 * uses that function to do most of the work. Both are fairly rough
3328 * interfaces, @bio must be presetup and ready for I/O.
3331 void submit_bio(int rw, struct bio *bio)
3333 int count = bio_sectors(bio);
3335 BIO_BUG_ON(!bio->bi_size);
3336 BIO_BUG_ON(!bio->bi_io_vec);
3339 count_vm_events(PGPGOUT, count);
3341 task_io_account_read(bio->bi_size);
3342 count_vm_events(PGPGIN, count);
3345 if (unlikely(block_dump)) {
3346 char b[BDEVNAME_SIZE];
3347 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
3348 current->comm, current->pid,
3349 (rw & WRITE) ? "WRITE" : "READ",
3350 (unsigned long long)bio->bi_sector,
3351 bdevname(bio->bi_bdev,b));
3354 generic_make_request(bio);
3357 EXPORT_SYMBOL(submit_bio);
3359 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
3361 if (blk_fs_request(rq)) {
3362 rq->hard_sector += nsect;
3363 rq->hard_nr_sectors -= nsect;
3366 * Move the I/O submission pointers ahead if required.
3368 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3369 (rq->sector <= rq->hard_sector)) {
3370 rq->sector = rq->hard_sector;
3371 rq->nr_sectors = rq->hard_nr_sectors;
3372 rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3373 rq->current_nr_sectors = rq->hard_cur_sectors;
3374 rq->buffer = bio_data(rq->bio);
3378 * if total number of sectors is less than the first segment
3379 * size, something has gone terribly wrong
3381 if (rq->nr_sectors < rq->current_nr_sectors) {
3382 printk("blk: request botched\n");
3383 rq->nr_sectors = rq->current_nr_sectors;
3388 static int __end_that_request_first(struct request *req, int uptodate,
3391 int total_bytes, bio_nbytes, error, next_idx = 0;
3394 blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
3397 * extend uptodate bool to allow < 0 value to be direct io error
3400 if (end_io_error(uptodate))
3401 error = !uptodate ? -EIO : uptodate;
3404 * for a REQ_BLOCK_PC request, we want to carry any eventual
3405 * sense key with us all the way through
3407 if (!blk_pc_request(req))
3411 if (blk_fs_request(req) && !(req->cmd_flags & REQ_QUIET))
3412 printk("end_request: I/O error, dev %s, sector %llu\n",
3413 req->rq_disk ? req->rq_disk->disk_name : "?",
3414 (unsigned long long)req->sector);
3417 if (blk_fs_request(req) && req->rq_disk) {
3418 const int rw = rq_data_dir(req);
3420 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
3423 total_bytes = bio_nbytes = 0;
3424 while ((bio = req->bio) != NULL) {
3427 if (nr_bytes >= bio->bi_size) {
3428 req->bio = bio->bi_next;
3429 nbytes = bio->bi_size;
3430 if (!ordered_bio_endio(req, bio, nbytes, error))
3431 bio_endio(bio, nbytes, error);
3435 int idx = bio->bi_idx + next_idx;
3437 if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3438 blk_dump_rq_flags(req, "__end_that");
3439 printk("%s: bio idx %d >= vcnt %d\n",
3441 bio->bi_idx, bio->bi_vcnt);
3445 nbytes = bio_iovec_idx(bio, idx)->bv_len;
3446 BIO_BUG_ON(nbytes > bio->bi_size);
3449 * not a complete bvec done
3451 if (unlikely(nbytes > nr_bytes)) {
3452 bio_nbytes += nr_bytes;
3453 total_bytes += nr_bytes;
3458 * advance to the next vector
3461 bio_nbytes += nbytes;
3464 total_bytes += nbytes;
3467 if ((bio = req->bio)) {
3469 * end more in this run, or just return 'not-done'
3471 if (unlikely(nr_bytes <= 0))
3483 * if the request wasn't completed, update state
3486 if (!ordered_bio_endio(req, bio, bio_nbytes, error))
3487 bio_endio(bio, bio_nbytes, error);
3488 bio->bi_idx += next_idx;
3489 bio_iovec(bio)->bv_offset += nr_bytes;
3490 bio_iovec(bio)->bv_len -= nr_bytes;
3493 blk_recalc_rq_sectors(req, total_bytes >> 9);
3494 blk_recalc_rq_segments(req);
3499 * end_that_request_first - end I/O on a request
3500 * @req: the request being processed
3501 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3502 * @nr_sectors: number of sectors to end I/O on
3505 * Ends I/O on a number of sectors attached to @req, and sets it up
3506 * for the next range of segments (if any) in the cluster.
3509 * 0 - we are done with this request, call end_that_request_last()
3510 * 1 - still buffers pending for this request
3512 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3514 return __end_that_request_first(req, uptodate, nr_sectors << 9);
3517 EXPORT_SYMBOL(end_that_request_first);
3520 * end_that_request_chunk - end I/O on a request
3521 * @req: the request being processed
3522 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3523 * @nr_bytes: number of bytes to complete
3526 * Ends I/O on a number of bytes attached to @req, and sets it up
3527 * for the next range of segments (if any). Like end_that_request_first(),
3528 * but deals with bytes instead of sectors.
3531 * 0 - we are done with this request, call end_that_request_last()
3532 * 1 - still buffers pending for this request
3534 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3536 return __end_that_request_first(req, uptodate, nr_bytes);
3539 EXPORT_SYMBOL(end_that_request_chunk);
3542 * splice the completion data to a local structure and hand off to
3543 * process_completion_queue() to complete the requests
3545 static void blk_done_softirq(struct softirq_action *h)
3547 struct list_head *cpu_list, local_list;
3549 local_irq_disable();
3550 cpu_list = &__get_cpu_var(blk_cpu_done);
3551 list_replace_init(cpu_list, &local_list);
3554 while (!list_empty(&local_list)) {
3555 struct request *rq = list_entry(local_list.next, struct request, donelist);
3557 list_del_init(&rq->donelist);
3558 rq->q->softirq_done_fn(rq);
3562 static int blk_cpu_notify(struct notifier_block *self, unsigned long action,
3566 * If a CPU goes away, splice its entries to the current CPU
3567 * and trigger a run of the softirq
3569 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
3570 int cpu = (unsigned long) hcpu;
3572 local_irq_disable();
3573 list_splice_init(&per_cpu(blk_cpu_done, cpu),
3574 &__get_cpu_var(blk_cpu_done));
3575 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3583 static struct notifier_block __devinitdata blk_cpu_notifier = {
3584 .notifier_call = blk_cpu_notify,
3588 * blk_complete_request - end I/O on a request
3589 * @req: the request being processed
3592 * Ends all I/O on a request. It does not handle partial completions,
3593 * unless the driver actually implements this in its completion callback
3594 * through requeueing. Theh actual completion happens out-of-order,
3595 * through a softirq handler. The user must have registered a completion
3596 * callback through blk_queue_softirq_done().
3599 void blk_complete_request(struct request *req)
3601 struct list_head *cpu_list;
3602 unsigned long flags;
3604 BUG_ON(!req->q->softirq_done_fn);
3606 local_irq_save(flags);
3608 cpu_list = &__get_cpu_var(blk_cpu_done);
3609 list_add_tail(&req->donelist, cpu_list);
3610 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3612 local_irq_restore(flags);
3615 EXPORT_SYMBOL(blk_complete_request);
3618 * queue lock must be held
3620 void end_that_request_last(struct request *req, int uptodate)
3622 struct gendisk *disk = req->rq_disk;
3626 * extend uptodate bool to allow < 0 value to be direct io error
3629 if (end_io_error(uptodate))
3630 error = !uptodate ? -EIO : uptodate;
3632 if (unlikely(laptop_mode) && blk_fs_request(req))
3633 laptop_io_completion();
3636 * Account IO completion. bar_rq isn't accounted as a normal
3637 * IO on queueing nor completion. Accounting the containing
3638 * request is enough.
3640 if (disk && blk_fs_request(req) && req != &req->q->bar_rq) {
3641 unsigned long duration = jiffies - req->start_time;
3642 const int rw = rq_data_dir(req);
3644 __disk_stat_inc(disk, ios[rw]);
3645 __disk_stat_add(disk, ticks[rw], duration);
3646 disk_round_stats(disk);
3650 req->end_io(req, error);
3652 __blk_put_request(req->q, req);
3655 EXPORT_SYMBOL(end_that_request_last);
3657 void end_request(struct request *req, int uptodate)
3659 if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3660 add_disk_randomness(req->rq_disk);
3661 blkdev_dequeue_request(req);
3662 end_that_request_last(req, uptodate);
3666 EXPORT_SYMBOL(end_request);
3668 void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
3671 /* first two bits are identical in rq->cmd_flags and bio->bi_rw */
3672 rq->cmd_flags |= (bio->bi_rw & 3);
3674 rq->nr_phys_segments = bio_phys_segments(q, bio);
3675 rq->nr_hw_segments = bio_hw_segments(q, bio);
3676 rq->current_nr_sectors = bio_cur_sectors(bio);
3677 rq->hard_cur_sectors = rq->current_nr_sectors;
3678 rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3679 rq->buffer = bio_data(bio);
3680 rq->data_len = bio->bi_size;
3682 rq->bio = rq->biotail = bio;
3685 EXPORT_SYMBOL(blk_rq_bio_prep);
3687 int kblockd_schedule_work(struct work_struct *work)
3689 return queue_work(kblockd_workqueue, work);
3692 EXPORT_SYMBOL(kblockd_schedule_work);
3694 void kblockd_flush_work(struct work_struct *work)
3696 cancel_work_sync(work);
3698 EXPORT_SYMBOL(kblockd_flush_work);
3700 int __init blk_dev_init(void)
3704 kblockd_workqueue = create_workqueue("kblockd");
3705 if (!kblockd_workqueue)
3706 panic("Failed to create kblockd\n");
3708 request_cachep = kmem_cache_create("blkdev_requests",
3709 sizeof(struct request), 0, SLAB_PANIC, NULL);
3711 requestq_cachep = kmem_cache_create("blkdev_queue",
3712 sizeof(struct request_queue), 0, SLAB_PANIC, NULL);
3714 iocontext_cachep = kmem_cache_create("blkdev_ioc",
3715 sizeof(struct io_context), 0, SLAB_PANIC, NULL);
3717 for_each_possible_cpu(i)
3718 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3720 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
3721 register_hotcpu_notifier(&blk_cpu_notifier);
3723 blk_max_low_pfn = max_low_pfn - 1;
3724 blk_max_pfn = max_pfn - 1;
3730 * IO Context helper functions
3732 void put_io_context(struct io_context *ioc)
3737 BUG_ON(atomic_read(&ioc->refcount) == 0);
3739 if (atomic_dec_and_test(&ioc->refcount)) {
3740 struct cfq_io_context *cic;
3743 if (ioc->aic && ioc->aic->dtor)
3744 ioc->aic->dtor(ioc->aic);
3745 if (ioc->cic_root.rb_node != NULL) {
3746 struct rb_node *n = rb_first(&ioc->cic_root);
3748 cic = rb_entry(n, struct cfq_io_context, rb_node);
3753 kmem_cache_free(iocontext_cachep, ioc);
3756 EXPORT_SYMBOL(put_io_context);
3758 /* Called by the exitting task */
3759 void exit_io_context(void)
3761 struct io_context *ioc;
3762 struct cfq_io_context *cic;
3765 ioc = current->io_context;
3766 current->io_context = NULL;
3767 task_unlock(current);
3770 if (ioc->aic && ioc->aic->exit)
3771 ioc->aic->exit(ioc->aic);
3772 if (ioc->cic_root.rb_node != NULL) {
3773 cic = rb_entry(rb_first(&ioc->cic_root), struct cfq_io_context, rb_node);
3777 put_io_context(ioc);
3781 * If the current task has no IO context then create one and initialise it.
3782 * Otherwise, return its existing IO context.
3784 * This returned IO context doesn't have a specifically elevated refcount,
3785 * but since the current task itself holds a reference, the context can be
3786 * used in general code, so long as it stays within `current` context.
3788 static struct io_context *current_io_context(gfp_t gfp_flags, int node)
3790 struct task_struct *tsk = current;
3791 struct io_context *ret;
3793 ret = tsk->io_context;
3797 ret = kmem_cache_alloc_node(iocontext_cachep, gfp_flags, node);
3799 atomic_set(&ret->refcount, 1);
3800 ret->task = current;
3801 ret->ioprio_changed = 0;
3802 ret->last_waited = jiffies; /* doesn't matter... */
3803 ret->nr_batch_requests = 0; /* because this is 0 */
3805 ret->cic_root.rb_node = NULL;
3806 ret->ioc_data = NULL;
3807 /* make sure set_task_ioprio() sees the settings above */
3809 tsk->io_context = ret;
3816 * If the current task has no IO context then create one and initialise it.
3817 * If it does have a context, take a ref on it.
3819 * This is always called in the context of the task which submitted the I/O.
3821 struct io_context *get_io_context(gfp_t gfp_flags, int node)
3823 struct io_context *ret;
3824 ret = current_io_context(gfp_flags, node);
3826 atomic_inc(&ret->refcount);
3829 EXPORT_SYMBOL(get_io_context);
3831 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3833 struct io_context *src = *psrc;
3834 struct io_context *dst = *pdst;
3837 BUG_ON(atomic_read(&src->refcount) == 0);
3838 atomic_inc(&src->refcount);
3839 put_io_context(dst);
3843 EXPORT_SYMBOL(copy_io_context);
3845 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3847 struct io_context *temp;
3852 EXPORT_SYMBOL(swap_io_context);
3857 struct queue_sysfs_entry {
3858 struct attribute attr;
3859 ssize_t (*show)(struct request_queue *, char *);
3860 ssize_t (*store)(struct request_queue *, const char *, size_t);
3864 queue_var_show(unsigned int var, char *page)
3866 return sprintf(page, "%d\n", var);
3870 queue_var_store(unsigned long *var, const char *page, size_t count)
3872 char *p = (char *) page;
3874 *var = simple_strtoul(p, &p, 10);
3878 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3880 return queue_var_show(q->nr_requests, (page));
3884 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3886 struct request_list *rl = &q->rq;
3888 int ret = queue_var_store(&nr, page, count);
3889 if (nr < BLKDEV_MIN_RQ)
3892 spin_lock_irq(q->queue_lock);
3893 q->nr_requests = nr;
3894 blk_queue_congestion_threshold(q);
3896 if (rl->count[READ] >= queue_congestion_on_threshold(q))
3897 blk_set_queue_congested(q, READ);
3898 else if (rl->count[READ] < queue_congestion_off_threshold(q))
3899 blk_clear_queue_congested(q, READ);
3901 if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3902 blk_set_queue_congested(q, WRITE);
3903 else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3904 blk_clear_queue_congested(q, WRITE);
3906 if (rl->count[READ] >= q->nr_requests) {
3907 blk_set_queue_full(q, READ);
3908 } else if (rl->count[READ]+1 <= q->nr_requests) {
3909 blk_clear_queue_full(q, READ);
3910 wake_up(&rl->wait[READ]);
3913 if (rl->count[WRITE] >= q->nr_requests) {
3914 blk_set_queue_full(q, WRITE);
3915 } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3916 blk_clear_queue_full(q, WRITE);
3917 wake_up(&rl->wait[WRITE]);
3919 spin_unlock_irq(q->queue_lock);
3923 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3925 int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3927 return queue_var_show(ra_kb, (page));
3931 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3933 unsigned long ra_kb;
3934 ssize_t ret = queue_var_store(&ra_kb, page, count);
3936 spin_lock_irq(q->queue_lock);
3937 q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3938 spin_unlock_irq(q->queue_lock);
3943 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3945 int max_sectors_kb = q->max_sectors >> 1;
3947 return queue_var_show(max_sectors_kb, (page));
3951 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3953 unsigned long max_sectors_kb,
3954 max_hw_sectors_kb = q->max_hw_sectors >> 1,
3955 page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3956 ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3959 if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3962 * Take the queue lock to update the readahead and max_sectors
3963 * values synchronously:
3965 spin_lock_irq(q->queue_lock);
3967 * Trim readahead window as well, if necessary:
3969 ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3970 if (ra_kb > max_sectors_kb)
3971 q->backing_dev_info.ra_pages =
3972 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3974 q->max_sectors = max_sectors_kb << 1;
3975 spin_unlock_irq(q->queue_lock);
3980 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3982 int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3984 return queue_var_show(max_hw_sectors_kb, (page));
3988 static struct queue_sysfs_entry queue_requests_entry = {
3989 .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3990 .show = queue_requests_show,
3991 .store = queue_requests_store,
3994 static struct queue_sysfs_entry queue_ra_entry = {
3995 .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3996 .show = queue_ra_show,
3997 .store = queue_ra_store,
4000 static struct queue_sysfs_entry queue_max_sectors_entry = {
4001 .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
4002 .show = queue_max_sectors_show,
4003 .store = queue_max_sectors_store,
4006 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
4007 .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
4008 .show = queue_max_hw_sectors_show,
4011 static struct queue_sysfs_entry queue_iosched_entry = {
4012 .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
4013 .show = elv_iosched_show,
4014 .store = elv_iosched_store,
4017 static struct attribute *default_attrs[] = {
4018 &queue_requests_entry.attr,
4019 &queue_ra_entry.attr,
4020 &queue_max_hw_sectors_entry.attr,
4021 &queue_max_sectors_entry.attr,
4022 &queue_iosched_entry.attr,
4026 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
4029 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
4031 struct queue_sysfs_entry *entry = to_queue(attr);
4032 struct request_queue *q =
4033 container_of(kobj, struct request_queue, kobj);
4038 mutex_lock(&q->sysfs_lock);
4039 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
4040 mutex_unlock(&q->sysfs_lock);
4043 res = entry->show(q, page);
4044 mutex_unlock(&q->sysfs_lock);
4049 queue_attr_store(struct kobject *kobj, struct attribute *attr,
4050 const char *page, size_t length)
4052 struct queue_sysfs_entry *entry = to_queue(attr);
4053 struct request_queue *q = container_of(kobj, struct request_queue, kobj);
4059 mutex_lock(&q->sysfs_lock);
4060 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
4061 mutex_unlock(&q->sysfs_lock);
4064 res = entry->store(q, page, length);
4065 mutex_unlock(&q->sysfs_lock);
4069 static struct sysfs_ops queue_sysfs_ops = {
4070 .show = queue_attr_show,
4071 .store = queue_attr_store,
4074 static struct kobj_type queue_ktype = {
4075 .sysfs_ops = &queue_sysfs_ops,
4076 .default_attrs = default_attrs,
4077 .release = blk_release_queue,
4080 int blk_register_queue(struct gendisk *disk)
4084 struct request_queue *q = disk->queue;
4086 if (!q || !q->request_fn)
4089 q->kobj.parent = kobject_get(&disk->kobj);
4091 ret = kobject_add(&q->kobj);
4095 kobject_uevent(&q->kobj, KOBJ_ADD);
4097 ret = elv_register_queue(q);
4099 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4100 kobject_del(&q->kobj);
4107 void blk_unregister_queue(struct gendisk *disk)
4109 struct request_queue *q = disk->queue;
4111 if (q && q->request_fn) {
4112 elv_unregister_queue(q);
4114 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4115 kobject_del(&q->kobj);
4116 kobject_put(&disk->kobj);