2 * Copyright (C) 1991, 1992 Linus Torvalds
3 * Copyright (C) 1994, Karl Keyte: Added support for disk statistics
4 * Elevator latency, (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
5 * Queue request tables / lock, selectable elevator, Jens Axboe <axboe@suse.de>
6 * kernel-doc documentation started by NeilBrown <neilb@cse.unsw.edu.au> - July2000
7 * bio rewrite, highmem i/o, etc, Jens Axboe <axboe@suse.de> - may 2001
11 * This handles all read/write requests to block devices
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/backing-dev.h>
16 #include <linux/bio.h>
17 #include <linux/blkdev.h>
18 #include <linux/highmem.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/string.h>
22 #include <linux/init.h>
23 #include <linux/bootmem.h> /* for max_pfn/max_low_pfn */
24 #include <linux/completion.h>
25 #include <linux/slab.h>
26 #include <linux/swap.h>
27 #include <linux/writeback.h>
28 #include <linux/task_io_accounting_ops.h>
29 #include <linux/interrupt.h>
30 #include <linux/cpu.h>
31 #include <linux/blktrace_api.h>
32 #include <linux/fault-inject.h>
37 #include <scsi/scsi_cmnd.h>
39 static void blk_unplug_work(struct work_struct *work);
40 static void blk_unplug_timeout(unsigned long data);
41 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io);
42 static void init_request_from_bio(struct request *req, struct bio *bio);
43 static int __make_request(struct request_queue *q, struct bio *bio);
44 static struct io_context *current_io_context(gfp_t gfp_flags, int node);
45 static void blk_recalc_rq_segments(struct request *rq);
46 static void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
50 * For the allocated request tables
52 static struct kmem_cache *request_cachep;
55 * For queue allocation
57 static struct kmem_cache *requestq_cachep;
60 * For io context allocations
62 static struct kmem_cache *iocontext_cachep;
65 * Controlling structure to kblockd
67 static struct workqueue_struct *kblockd_workqueue;
69 unsigned long blk_max_low_pfn, blk_max_pfn;
71 EXPORT_SYMBOL(blk_max_low_pfn);
72 EXPORT_SYMBOL(blk_max_pfn);
74 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
76 /* Amount of time in which a process may batch requests */
77 #define BLK_BATCH_TIME (HZ/50UL)
79 /* Number of requests a "batching" process may submit */
80 #define BLK_BATCH_REQ 32
83 * Return the threshold (number of used requests) at which the queue is
84 * considered to be congested. It include a little hysteresis to keep the
85 * context switch rate down.
87 static inline int queue_congestion_on_threshold(struct request_queue *q)
89 return q->nr_congestion_on;
93 * The threshold at which a queue is considered to be uncongested
95 static inline int queue_congestion_off_threshold(struct request_queue *q)
97 return q->nr_congestion_off;
100 static void blk_queue_congestion_threshold(struct request_queue *q)
104 nr = q->nr_requests - (q->nr_requests / 8) + 1;
105 if (nr > q->nr_requests)
107 q->nr_congestion_on = nr;
109 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
112 q->nr_congestion_off = nr;
116 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
119 * Locates the passed device's request queue and returns the address of its
122 * Will return NULL if the request queue cannot be located.
124 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
126 struct backing_dev_info *ret = NULL;
127 struct request_queue *q = bdev_get_queue(bdev);
130 ret = &q->backing_dev_info;
133 EXPORT_SYMBOL(blk_get_backing_dev_info);
136 * blk_queue_prep_rq - set a prepare_request function for queue
138 * @pfn: prepare_request function
140 * It's possible for a queue to register a prepare_request callback which
141 * is invoked before the request is handed to the request_fn. The goal of
142 * the function is to prepare a request for I/O, it can be used to build a
143 * cdb from the request data for instance.
146 void blk_queue_prep_rq(struct request_queue *q, prep_rq_fn *pfn)
151 EXPORT_SYMBOL(blk_queue_prep_rq);
154 * blk_queue_merge_bvec - set a merge_bvec function for queue
156 * @mbfn: merge_bvec_fn
158 * Usually queues have static limitations on the max sectors or segments that
159 * we can put in a request. Stacking drivers may have some settings that
160 * are dynamic, and thus we have to query the queue whether it is ok to
161 * add a new bio_vec to a bio at a given offset or not. If the block device
162 * has such limitations, it needs to register a merge_bvec_fn to control
163 * the size of bio's sent to it. Note that a block device *must* allow a
164 * single page to be added to an empty bio. The block device driver may want
165 * to use the bio_split() function to deal with these bio's. By default
166 * no merge_bvec_fn is defined for a queue, and only the fixed limits are
169 void blk_queue_merge_bvec(struct request_queue *q, merge_bvec_fn *mbfn)
171 q->merge_bvec_fn = mbfn;
174 EXPORT_SYMBOL(blk_queue_merge_bvec);
176 void blk_queue_softirq_done(struct request_queue *q, softirq_done_fn *fn)
178 q->softirq_done_fn = fn;
181 EXPORT_SYMBOL(blk_queue_softirq_done);
184 * blk_queue_make_request - define an alternate make_request function for a device
185 * @q: the request queue for the device to be affected
186 * @mfn: the alternate make_request function
189 * The normal way for &struct bios to be passed to a device
190 * driver is for them to be collected into requests on a request
191 * queue, and then to allow the device driver to select requests
192 * off that queue when it is ready. This works well for many block
193 * devices. However some block devices (typically virtual devices
194 * such as md or lvm) do not benefit from the processing on the
195 * request queue, and are served best by having the requests passed
196 * directly to them. This can be achieved by providing a function
197 * to blk_queue_make_request().
200 * The driver that does this *must* be able to deal appropriately
201 * with buffers in "highmemory". This can be accomplished by either calling
202 * __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
203 * blk_queue_bounce() to create a buffer in normal memory.
205 void blk_queue_make_request(struct request_queue * q, make_request_fn * mfn)
210 q->nr_requests = BLKDEV_MAX_RQ;
211 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
212 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
213 q->make_request_fn = mfn;
214 q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
215 q->backing_dev_info.state = 0;
216 q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
217 blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
218 blk_queue_hardsect_size(q, 512);
219 blk_queue_dma_alignment(q, 511);
220 blk_queue_congestion_threshold(q);
221 q->nr_batching = BLK_BATCH_REQ;
223 q->unplug_thresh = 4; /* hmm */
224 q->unplug_delay = (3 * HZ) / 1000; /* 3 milliseconds */
225 if (q->unplug_delay == 0)
228 INIT_WORK(&q->unplug_work, blk_unplug_work);
230 q->unplug_timer.function = blk_unplug_timeout;
231 q->unplug_timer.data = (unsigned long)q;
234 * by default assume old behaviour and bounce for any highmem page
236 blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
239 EXPORT_SYMBOL(blk_queue_make_request);
241 static void rq_init(struct request_queue *q, struct request *rq)
243 INIT_LIST_HEAD(&rq->queuelist);
244 INIT_LIST_HEAD(&rq->donelist);
247 rq->bio = rq->biotail = NULL;
248 INIT_HLIST_NODE(&rq->hash);
249 RB_CLEAR_NODE(&rq->rb_node);
257 rq->nr_phys_segments = 0;
260 rq->end_io_data = NULL;
261 rq->completion_data = NULL;
266 * blk_queue_ordered - does this queue support ordered writes
267 * @q: the request queue
268 * @ordered: one of QUEUE_ORDERED_*
269 * @prepare_flush_fn: rq setup helper for cache flush ordered writes
272 * For journalled file systems, doing ordered writes on a commit
273 * block instead of explicitly doing wait_on_buffer (which is bad
274 * for performance) can be a big win. Block drivers supporting this
275 * feature should call this function and indicate so.
278 int blk_queue_ordered(struct request_queue *q, unsigned ordered,
279 prepare_flush_fn *prepare_flush_fn)
281 if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
282 prepare_flush_fn == NULL) {
283 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
287 if (ordered != QUEUE_ORDERED_NONE &&
288 ordered != QUEUE_ORDERED_DRAIN &&
289 ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
290 ordered != QUEUE_ORDERED_DRAIN_FUA &&
291 ordered != QUEUE_ORDERED_TAG &&
292 ordered != QUEUE_ORDERED_TAG_FLUSH &&
293 ordered != QUEUE_ORDERED_TAG_FUA) {
294 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
298 q->ordered = ordered;
299 q->next_ordered = ordered;
300 q->prepare_flush_fn = prepare_flush_fn;
305 EXPORT_SYMBOL(blk_queue_ordered);
308 * blk_queue_issue_flush_fn - set function for issuing a flush
309 * @q: the request queue
310 * @iff: the function to be called issuing the flush
313 * If a driver supports issuing a flush command, the support is notified
314 * to the block layer by defining it through this call.
317 void blk_queue_issue_flush_fn(struct request_queue *q, issue_flush_fn *iff)
319 q->issue_flush_fn = iff;
322 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
325 * Cache flushing for ordered writes handling
327 inline unsigned blk_ordered_cur_seq(struct request_queue *q)
331 return 1 << ffz(q->ordseq);
334 unsigned blk_ordered_req_seq(struct request *rq)
336 struct request_queue *q = rq->q;
338 BUG_ON(q->ordseq == 0);
340 if (rq == &q->pre_flush_rq)
341 return QUEUE_ORDSEQ_PREFLUSH;
342 if (rq == &q->bar_rq)
343 return QUEUE_ORDSEQ_BAR;
344 if (rq == &q->post_flush_rq)
345 return QUEUE_ORDSEQ_POSTFLUSH;
348 * !fs requests don't need to follow barrier ordering. Always
349 * put them at the front. This fixes the following deadlock.
351 * http://thread.gmane.org/gmane.linux.kernel/537473
353 if (!blk_fs_request(rq))
354 return QUEUE_ORDSEQ_DRAIN;
356 if ((rq->cmd_flags & REQ_ORDERED_COLOR) ==
357 (q->orig_bar_rq->cmd_flags & REQ_ORDERED_COLOR))
358 return QUEUE_ORDSEQ_DRAIN;
360 return QUEUE_ORDSEQ_DONE;
363 void blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
368 if (error && !q->orderr)
371 BUG_ON(q->ordseq & seq);
374 if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
378 * Okay, sequence complete.
382 uptodate = q->orderr;
387 end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
388 end_that_request_last(rq, uptodate);
391 static void pre_flush_end_io(struct request *rq, int error)
393 elv_completed_request(rq->q, rq);
394 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
397 static void bar_end_io(struct request *rq, int error)
399 elv_completed_request(rq->q, rq);
400 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
403 static void post_flush_end_io(struct request *rq, int error)
405 elv_completed_request(rq->q, rq);
406 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
409 static void queue_flush(struct request_queue *q, unsigned which)
412 rq_end_io_fn *end_io;
414 if (which == QUEUE_ORDERED_PREFLUSH) {
415 rq = &q->pre_flush_rq;
416 end_io = pre_flush_end_io;
418 rq = &q->post_flush_rq;
419 end_io = post_flush_end_io;
422 rq->cmd_flags = REQ_HARDBARRIER;
424 rq->elevator_private = NULL;
425 rq->elevator_private2 = NULL;
426 rq->rq_disk = q->bar_rq.rq_disk;
428 q->prepare_flush_fn(q, rq);
430 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
433 static inline struct request *start_ordered(struct request_queue *q,
437 q->ordered = q->next_ordered;
438 q->ordseq |= QUEUE_ORDSEQ_STARTED;
441 * Prep proxy barrier request.
443 blkdev_dequeue_request(rq);
448 if (bio_data_dir(q->orig_bar_rq->bio) == WRITE)
449 rq->cmd_flags |= REQ_RW;
450 if (q->ordered & QUEUE_ORDERED_FUA)
451 rq->cmd_flags |= REQ_FUA;
452 rq->elevator_private = NULL;
453 rq->elevator_private2 = NULL;
454 init_request_from_bio(rq, q->orig_bar_rq->bio);
455 rq->end_io = bar_end_io;
458 * Queue ordered sequence. As we stack them at the head, we
459 * need to queue in reverse order. Note that we rely on that
460 * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
461 * request gets inbetween ordered sequence.
463 if (q->ordered & QUEUE_ORDERED_POSTFLUSH)
464 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
466 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
468 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
470 if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
471 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
472 rq = &q->pre_flush_rq;
474 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
476 if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
477 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
484 int blk_do_ordered(struct request_queue *q, struct request **rqp)
486 struct request *rq = *rqp;
487 int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
493 if (q->next_ordered != QUEUE_ORDERED_NONE) {
494 *rqp = start_ordered(q, rq);
498 * This can happen when the queue switches to
499 * ORDERED_NONE while this request is on it.
501 blkdev_dequeue_request(rq);
502 end_that_request_first(rq, -EOPNOTSUPP,
503 rq->hard_nr_sectors);
504 end_that_request_last(rq, -EOPNOTSUPP);
511 * Ordered sequence in progress
514 /* Special requests are not subject to ordering rules. */
515 if (!blk_fs_request(rq) &&
516 rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
519 if (q->ordered & QUEUE_ORDERED_TAG) {
520 /* Ordered by tag. Blocking the next barrier is enough. */
521 if (is_barrier && rq != &q->bar_rq)
524 /* Ordered by draining. Wait for turn. */
525 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
526 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
533 static void req_bio_endio(struct request *rq, struct bio *bio,
534 unsigned int nbytes, int error)
536 struct request_queue *q = rq->q;
538 if (&q->bar_rq != rq) {
540 clear_bit(BIO_UPTODATE, &bio->bi_flags);
541 else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
544 if (unlikely(nbytes > bio->bi_size)) {
545 printk("%s: want %u bytes done, only %u left\n",
546 __FUNCTION__, nbytes, bio->bi_size);
547 nbytes = bio->bi_size;
550 bio->bi_size -= nbytes;
551 bio->bi_sector += (nbytes >> 9);
552 if (bio->bi_size == 0)
553 bio_endio(bio, error);
557 * Okay, this is the barrier request in progress, just
560 if (error && !q->orderr)
566 * blk_queue_bounce_limit - set bounce buffer limit for queue
567 * @q: the request queue for the device
568 * @dma_addr: bus address limit
571 * Different hardware can have different requirements as to what pages
572 * it can do I/O directly to. A low level driver can call
573 * blk_queue_bounce_limit to have lower memory pages allocated as bounce
574 * buffers for doing I/O to pages residing above @page.
576 void blk_queue_bounce_limit(struct request_queue *q, u64 dma_addr)
578 unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
581 q->bounce_gfp = GFP_NOIO;
582 #if BITS_PER_LONG == 64
583 /* Assume anything <= 4GB can be handled by IOMMU.
584 Actually some IOMMUs can handle everything, but I don't
585 know of a way to test this here. */
586 if (bounce_pfn < (min_t(u64,0xffffffff,BLK_BOUNCE_HIGH) >> PAGE_SHIFT))
588 q->bounce_pfn = max_low_pfn;
590 if (bounce_pfn < blk_max_low_pfn)
592 q->bounce_pfn = bounce_pfn;
595 init_emergency_isa_pool();
596 q->bounce_gfp = GFP_NOIO | GFP_DMA;
597 q->bounce_pfn = bounce_pfn;
601 EXPORT_SYMBOL(blk_queue_bounce_limit);
604 * blk_queue_max_sectors - set max sectors for a request for this queue
605 * @q: the request queue for the device
606 * @max_sectors: max sectors in the usual 512b unit
609 * Enables a low level driver to set an upper limit on the size of
612 void blk_queue_max_sectors(struct request_queue *q, unsigned int max_sectors)
614 if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
615 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
616 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
619 if (BLK_DEF_MAX_SECTORS > max_sectors)
620 q->max_hw_sectors = q->max_sectors = max_sectors;
622 q->max_sectors = BLK_DEF_MAX_SECTORS;
623 q->max_hw_sectors = max_sectors;
627 EXPORT_SYMBOL(blk_queue_max_sectors);
630 * blk_queue_max_phys_segments - set max phys segments for a request for this queue
631 * @q: the request queue for the device
632 * @max_segments: max number of segments
635 * Enables a low level driver to set an upper limit on the number of
636 * physical data segments in a request. This would be the largest sized
637 * scatter list the driver could handle.
639 void blk_queue_max_phys_segments(struct request_queue *q,
640 unsigned short max_segments)
644 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
647 q->max_phys_segments = max_segments;
650 EXPORT_SYMBOL(blk_queue_max_phys_segments);
653 * blk_queue_max_hw_segments - set max hw segments for a request for this queue
654 * @q: the request queue for the device
655 * @max_segments: max number of segments
658 * Enables a low level driver to set an upper limit on the number of
659 * hw data segments in a request. This would be the largest number of
660 * address/length pairs the host adapter can actually give as once
663 void blk_queue_max_hw_segments(struct request_queue *q,
664 unsigned short max_segments)
668 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
671 q->max_hw_segments = max_segments;
674 EXPORT_SYMBOL(blk_queue_max_hw_segments);
677 * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
678 * @q: the request queue for the device
679 * @max_size: max size of segment in bytes
682 * Enables a low level driver to set an upper limit on the size of a
685 void blk_queue_max_segment_size(struct request_queue *q, unsigned int max_size)
687 if (max_size < PAGE_CACHE_SIZE) {
688 max_size = PAGE_CACHE_SIZE;
689 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
692 q->max_segment_size = max_size;
695 EXPORT_SYMBOL(blk_queue_max_segment_size);
698 * blk_queue_hardsect_size - set hardware sector size for the queue
699 * @q: the request queue for the device
700 * @size: the hardware sector size, in bytes
703 * This should typically be set to the lowest possible sector size
704 * that the hardware can operate on (possible without reverting to
705 * even internal read-modify-write operations). Usually the default
706 * of 512 covers most hardware.
708 void blk_queue_hardsect_size(struct request_queue *q, unsigned short size)
710 q->hardsect_size = size;
713 EXPORT_SYMBOL(blk_queue_hardsect_size);
716 * Returns the minimum that is _not_ zero, unless both are zero.
718 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
721 * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
722 * @t: the stacking driver (top)
723 * @b: the underlying device (bottom)
725 void blk_queue_stack_limits(struct request_queue *t, struct request_queue *b)
727 /* zero is "infinity" */
728 t->max_sectors = min_not_zero(t->max_sectors,b->max_sectors);
729 t->max_hw_sectors = min_not_zero(t->max_hw_sectors,b->max_hw_sectors);
731 t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
732 t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
733 t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
734 t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
735 if (!test_bit(QUEUE_FLAG_CLUSTER, &b->queue_flags))
736 clear_bit(QUEUE_FLAG_CLUSTER, &t->queue_flags);
739 EXPORT_SYMBOL(blk_queue_stack_limits);
742 * blk_queue_segment_boundary - set boundary rules for segment merging
743 * @q: the request queue for the device
744 * @mask: the memory boundary mask
746 void blk_queue_segment_boundary(struct request_queue *q, unsigned long mask)
748 if (mask < PAGE_CACHE_SIZE - 1) {
749 mask = PAGE_CACHE_SIZE - 1;
750 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
753 q->seg_boundary_mask = mask;
756 EXPORT_SYMBOL(blk_queue_segment_boundary);
759 * blk_queue_dma_alignment - set dma length and memory alignment
760 * @q: the request queue for the device
761 * @mask: alignment mask
764 * set required memory and length aligment for direct dma transactions.
765 * this is used when buiding direct io requests for the queue.
768 void blk_queue_dma_alignment(struct request_queue *q, int mask)
770 q->dma_alignment = mask;
773 EXPORT_SYMBOL(blk_queue_dma_alignment);
776 * blk_queue_find_tag - find a request by its tag and queue
777 * @q: The request queue for the device
778 * @tag: The tag of the request
781 * Should be used when a device returns a tag and you want to match
784 * no locks need be held.
786 struct request *blk_queue_find_tag(struct request_queue *q, int tag)
788 return blk_map_queue_find_tag(q->queue_tags, tag);
791 EXPORT_SYMBOL(blk_queue_find_tag);
794 * __blk_free_tags - release a given set of tag maintenance info
795 * @bqt: the tag map to free
797 * Tries to free the specified @bqt@. Returns true if it was
798 * actually freed and false if there are still references using it
800 static int __blk_free_tags(struct blk_queue_tag *bqt)
804 retval = atomic_dec_and_test(&bqt->refcnt);
807 BUG_ON(!list_empty(&bqt->busy_list));
809 kfree(bqt->tag_index);
810 bqt->tag_index = NULL;
823 * __blk_queue_free_tags - release tag maintenance info
824 * @q: the request queue for the device
827 * blk_cleanup_queue() will take care of calling this function, if tagging
828 * has been used. So there's no need to call this directly.
830 static void __blk_queue_free_tags(struct request_queue *q)
832 struct blk_queue_tag *bqt = q->queue_tags;
837 __blk_free_tags(bqt);
839 q->queue_tags = NULL;
840 q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
845 * blk_free_tags - release a given set of tag maintenance info
846 * @bqt: the tag map to free
848 * For externally managed @bqt@ frees the map. Callers of this
849 * function must guarantee to have released all the queues that
850 * might have been using this tag map.
852 void blk_free_tags(struct blk_queue_tag *bqt)
854 if (unlikely(!__blk_free_tags(bqt)))
857 EXPORT_SYMBOL(blk_free_tags);
860 * blk_queue_free_tags - release tag maintenance info
861 * @q: the request queue for the device
864 * This is used to disabled tagged queuing to a device, yet leave
867 void blk_queue_free_tags(struct request_queue *q)
869 clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
872 EXPORT_SYMBOL(blk_queue_free_tags);
875 init_tag_map(struct request_queue *q, struct blk_queue_tag *tags, int depth)
877 struct request **tag_index;
878 unsigned long *tag_map;
881 if (q && depth > q->nr_requests * 2) {
882 depth = q->nr_requests * 2;
883 printk(KERN_ERR "%s: adjusted depth to %d\n",
884 __FUNCTION__, depth);
887 tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC);
891 nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
892 tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
896 tags->real_max_depth = depth;
897 tags->max_depth = depth;
898 tags->tag_index = tag_index;
899 tags->tag_map = tag_map;
907 static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q,
910 struct blk_queue_tag *tags;
912 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
916 if (init_tag_map(q, tags, depth))
919 INIT_LIST_HEAD(&tags->busy_list);
921 atomic_set(&tags->refcnt, 1);
929 * blk_init_tags - initialize the tag info for an external tag map
930 * @depth: the maximum queue depth supported
931 * @tags: the tag to use
933 struct blk_queue_tag *blk_init_tags(int depth)
935 return __blk_queue_init_tags(NULL, depth);
937 EXPORT_SYMBOL(blk_init_tags);
940 * blk_queue_init_tags - initialize the queue tag info
941 * @q: the request queue for the device
942 * @depth: the maximum queue depth supported
943 * @tags: the tag to use
945 int blk_queue_init_tags(struct request_queue *q, int depth,
946 struct blk_queue_tag *tags)
950 BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
952 if (!tags && !q->queue_tags) {
953 tags = __blk_queue_init_tags(q, depth);
957 } else if (q->queue_tags) {
958 if ((rc = blk_queue_resize_tags(q, depth)))
960 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
963 atomic_inc(&tags->refcnt);
966 * assign it, all done
968 q->queue_tags = tags;
969 q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
976 EXPORT_SYMBOL(blk_queue_init_tags);
979 * blk_queue_resize_tags - change the queueing depth
980 * @q: the request queue for the device
981 * @new_depth: the new max command queueing depth
984 * Must be called with the queue lock held.
986 int blk_queue_resize_tags(struct request_queue *q, int new_depth)
988 struct blk_queue_tag *bqt = q->queue_tags;
989 struct request **tag_index;
990 unsigned long *tag_map;
991 int max_depth, nr_ulongs;
997 * if we already have large enough real_max_depth. just
998 * adjust max_depth. *NOTE* as requests with tag value
999 * between new_depth and real_max_depth can be in-flight, tag
1000 * map can not be shrunk blindly here.
1002 if (new_depth <= bqt->real_max_depth) {
1003 bqt->max_depth = new_depth;
1008 * Currently cannot replace a shared tag map with a new
1009 * one, so error out if this is the case
1011 if (atomic_read(&bqt->refcnt) != 1)
1015 * save the old state info, so we can copy it back
1017 tag_index = bqt->tag_index;
1018 tag_map = bqt->tag_map;
1019 max_depth = bqt->real_max_depth;
1021 if (init_tag_map(q, bqt, new_depth))
1024 memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
1025 nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
1026 memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
1033 EXPORT_SYMBOL(blk_queue_resize_tags);
1036 * blk_queue_end_tag - end tag operations for a request
1037 * @q: the request queue for the device
1038 * @rq: the request that has completed
1041 * Typically called when end_that_request_first() returns 0, meaning
1042 * all transfers have been done for a request. It's important to call
1043 * this function before end_that_request_last(), as that will put the
1044 * request back on the free list thus corrupting the internal tag list.
1047 * queue lock must be held.
1049 void blk_queue_end_tag(struct request_queue *q, struct request *rq)
1051 struct blk_queue_tag *bqt = q->queue_tags;
1056 if (unlikely(tag >= bqt->real_max_depth))
1058 * This can happen after tag depth has been reduced.
1059 * FIXME: how about a warning or info message here?
1063 list_del_init(&rq->queuelist);
1064 rq->cmd_flags &= ~REQ_QUEUED;
1067 if (unlikely(bqt->tag_index[tag] == NULL))
1068 printk(KERN_ERR "%s: tag %d is missing\n",
1071 bqt->tag_index[tag] = NULL;
1074 * We use test_and_clear_bit's memory ordering properties here.
1075 * The tag_map bit acts as a lock for tag_index[bit], so we need
1076 * a barrer before clearing the bit (precisely: release semantics).
1077 * Could use clear_bit_unlock when it is merged.
1079 if (unlikely(!test_and_clear_bit(tag, bqt->tag_map))) {
1080 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
1088 EXPORT_SYMBOL(blk_queue_end_tag);
1091 * blk_queue_start_tag - find a free tag and assign it
1092 * @q: the request queue for the device
1093 * @rq: the block request that needs tagging
1096 * This can either be used as a stand-alone helper, or possibly be
1097 * assigned as the queue &prep_rq_fn (in which case &struct request
1098 * automagically gets a tag assigned). Note that this function
1099 * assumes that any type of request can be queued! if this is not
1100 * true for your device, you must check the request type before
1101 * calling this function. The request will also be removed from
1102 * the request queue, so it's the drivers responsibility to readd
1103 * it if it should need to be restarted for some reason.
1106 * queue lock must be held.
1108 int blk_queue_start_tag(struct request_queue *q, struct request *rq)
1110 struct blk_queue_tag *bqt = q->queue_tags;
1113 if (unlikely((rq->cmd_flags & REQ_QUEUED))) {
1115 "%s: request %p for device [%s] already tagged %d",
1117 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
1122 * Protect against shared tag maps, as we may not have exclusive
1123 * access to the tag map.
1126 tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
1127 if (tag >= bqt->max_depth)
1130 } while (test_and_set_bit(tag, bqt->tag_map));
1132 * We rely on test_and_set_bit providing lock memory ordering semantics
1133 * (could use test_and_set_bit_lock when it is merged).
1136 rq->cmd_flags |= REQ_QUEUED;
1138 bqt->tag_index[tag] = rq;
1139 blkdev_dequeue_request(rq);
1140 list_add(&rq->queuelist, &bqt->busy_list);
1145 EXPORT_SYMBOL(blk_queue_start_tag);
1148 * blk_queue_invalidate_tags - invalidate all pending tags
1149 * @q: the request queue for the device
1152 * Hardware conditions may dictate a need to stop all pending requests.
1153 * In this case, we will safely clear the block side of the tag queue and
1154 * readd all requests to the request queue in the right order.
1157 * queue lock must be held.
1159 void blk_queue_invalidate_tags(struct request_queue *q)
1161 struct blk_queue_tag *bqt = q->queue_tags;
1162 struct list_head *tmp, *n;
1165 list_for_each_safe(tmp, n, &bqt->busy_list) {
1166 rq = list_entry_rq(tmp);
1168 if (rq->tag == -1) {
1170 "%s: bad tag found on list\n", __FUNCTION__);
1171 list_del_init(&rq->queuelist);
1172 rq->cmd_flags &= ~REQ_QUEUED;
1174 blk_queue_end_tag(q, rq);
1176 rq->cmd_flags &= ~REQ_STARTED;
1177 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1181 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1183 void blk_dump_rq_flags(struct request *rq, char *msg)
1187 printk("%s: dev %s: type=%x, flags=%x\n", msg,
1188 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->cmd_type,
1191 printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1193 rq->current_nr_sectors);
1194 printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1196 if (blk_pc_request(rq)) {
1198 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1199 printk("%02x ", rq->cmd[bit]);
1204 EXPORT_SYMBOL(blk_dump_rq_flags);
1206 void blk_recount_segments(struct request_queue *q, struct bio *bio)
1209 struct bio *nxt = bio->bi_next;
1211 rq.bio = rq.biotail = bio;
1212 bio->bi_next = NULL;
1213 blk_recalc_rq_segments(&rq);
1215 bio->bi_phys_segments = rq.nr_phys_segments;
1216 bio->bi_hw_segments = rq.nr_hw_segments;
1217 bio->bi_flags |= (1 << BIO_SEG_VALID);
1219 EXPORT_SYMBOL(blk_recount_segments);
1221 static void blk_recalc_rq_segments(struct request *rq)
1225 unsigned int phys_size;
1226 unsigned int hw_size;
1227 struct bio_vec *bv, *bvprv = NULL;
1231 struct req_iterator iter;
1232 int high, highprv = 1;
1233 struct request_queue *q = rq->q;
1238 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1239 hw_seg_size = seg_size = 0;
1240 phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
1241 rq_for_each_segment(bv, rq, iter) {
1243 * the trick here is making sure that a high page is never
1244 * considered part of another segment, since that might
1245 * change with the bounce page.
1247 high = page_to_pfn(bv->bv_page) > q->bounce_pfn;
1248 if (high || highprv)
1249 goto new_hw_segment;
1251 if (seg_size + bv->bv_len > q->max_segment_size)
1253 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1255 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1257 if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1258 goto new_hw_segment;
1260 seg_size += bv->bv_len;
1261 hw_seg_size += bv->bv_len;
1266 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1267 !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1268 hw_seg_size += bv->bv_len;
1271 if (nr_hw_segs == 1 &&
1272 hw_seg_size > rq->bio->bi_hw_front_size)
1273 rq->bio->bi_hw_front_size = hw_seg_size;
1274 hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1280 seg_size = bv->bv_len;
1284 if (nr_hw_segs == 1 &&
1285 hw_seg_size > rq->bio->bi_hw_front_size)
1286 rq->bio->bi_hw_front_size = hw_seg_size;
1287 if (hw_seg_size > rq->biotail->bi_hw_back_size)
1288 rq->biotail->bi_hw_back_size = hw_seg_size;
1289 rq->nr_phys_segments = nr_phys_segs;
1290 rq->nr_hw_segments = nr_hw_segs;
1293 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
1296 if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1299 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1301 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1305 * bio and nxt are contigous in memory, check if the queue allows
1306 * these two to be merged into one
1308 if (BIO_SEG_BOUNDARY(q, bio, nxt))
1314 static int blk_hw_contig_segment(struct request_queue *q, struct bio *bio,
1317 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1318 blk_recount_segments(q, bio);
1319 if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1320 blk_recount_segments(q, nxt);
1321 if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1322 BIOVEC_VIRT_OVERSIZE(bio->bi_hw_back_size + nxt->bi_hw_front_size))
1324 if (bio->bi_hw_back_size + nxt->bi_hw_front_size > q->max_segment_size)
1331 * map a request to scatterlist, return number of sg entries setup. Caller
1332 * must make sure sg can hold rq->nr_phys_segments entries
1334 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
1335 struct scatterlist *sg)
1337 struct bio_vec *bvec, *bvprv;
1338 struct req_iterator iter;
1342 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1345 * for each bio in rq
1348 rq_for_each_segment(bvec, rq, iter) {
1349 int nbytes = bvec->bv_len;
1351 if (bvprv && cluster) {
1352 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1355 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1357 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1360 sg[nsegs - 1].length += nbytes;
1363 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1364 sg[nsegs].page = bvec->bv_page;
1365 sg[nsegs].length = nbytes;
1366 sg[nsegs].offset = bvec->bv_offset;
1371 } /* segments in rq */
1376 EXPORT_SYMBOL(blk_rq_map_sg);
1379 * the standard queue merge functions, can be overridden with device
1380 * specific ones if so desired
1383 static inline int ll_new_mergeable(struct request_queue *q,
1384 struct request *req,
1387 int nr_phys_segs = bio_phys_segments(q, bio);
1389 if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1390 req->cmd_flags |= REQ_NOMERGE;
1391 if (req == q->last_merge)
1392 q->last_merge = NULL;
1397 * A hw segment is just getting larger, bump just the phys
1400 req->nr_phys_segments += nr_phys_segs;
1404 static inline int ll_new_hw_segment(struct request_queue *q,
1405 struct request *req,
1408 int nr_hw_segs = bio_hw_segments(q, bio);
1409 int nr_phys_segs = bio_phys_segments(q, bio);
1411 if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1412 || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1413 req->cmd_flags |= REQ_NOMERGE;
1414 if (req == q->last_merge)
1415 q->last_merge = NULL;
1420 * This will form the start of a new hw segment. Bump both
1423 req->nr_hw_segments += nr_hw_segs;
1424 req->nr_phys_segments += nr_phys_segs;
1428 static int ll_back_merge_fn(struct request_queue *q, struct request *req,
1431 unsigned short max_sectors;
1434 if (unlikely(blk_pc_request(req)))
1435 max_sectors = q->max_hw_sectors;
1437 max_sectors = q->max_sectors;
1439 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1440 req->cmd_flags |= REQ_NOMERGE;
1441 if (req == q->last_merge)
1442 q->last_merge = NULL;
1445 if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1446 blk_recount_segments(q, req->biotail);
1447 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1448 blk_recount_segments(q, bio);
1449 len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1450 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1451 !BIOVEC_VIRT_OVERSIZE(len)) {
1452 int mergeable = ll_new_mergeable(q, req, bio);
1455 if (req->nr_hw_segments == 1)
1456 req->bio->bi_hw_front_size = len;
1457 if (bio->bi_hw_segments == 1)
1458 bio->bi_hw_back_size = len;
1463 return ll_new_hw_segment(q, req, bio);
1466 static int ll_front_merge_fn(struct request_queue *q, struct request *req,
1469 unsigned short max_sectors;
1472 if (unlikely(blk_pc_request(req)))
1473 max_sectors = q->max_hw_sectors;
1475 max_sectors = q->max_sectors;
1478 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1479 req->cmd_flags |= REQ_NOMERGE;
1480 if (req == q->last_merge)
1481 q->last_merge = NULL;
1484 len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1485 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1486 blk_recount_segments(q, bio);
1487 if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1488 blk_recount_segments(q, req->bio);
1489 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1490 !BIOVEC_VIRT_OVERSIZE(len)) {
1491 int mergeable = ll_new_mergeable(q, req, bio);
1494 if (bio->bi_hw_segments == 1)
1495 bio->bi_hw_front_size = len;
1496 if (req->nr_hw_segments == 1)
1497 req->biotail->bi_hw_back_size = len;
1502 return ll_new_hw_segment(q, req, bio);
1505 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
1506 struct request *next)
1508 int total_phys_segments;
1509 int total_hw_segments;
1512 * First check if the either of the requests are re-queued
1513 * requests. Can't merge them if they are.
1515 if (req->special || next->special)
1519 * Will it become too large?
1521 if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1524 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1525 if (blk_phys_contig_segment(q, req->biotail, next->bio))
1526 total_phys_segments--;
1528 if (total_phys_segments > q->max_phys_segments)
1531 total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1532 if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1533 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1535 * propagate the combined length to the end of the requests
1537 if (req->nr_hw_segments == 1)
1538 req->bio->bi_hw_front_size = len;
1539 if (next->nr_hw_segments == 1)
1540 next->biotail->bi_hw_back_size = len;
1541 total_hw_segments--;
1544 if (total_hw_segments > q->max_hw_segments)
1547 /* Merge is OK... */
1548 req->nr_phys_segments = total_phys_segments;
1549 req->nr_hw_segments = total_hw_segments;
1554 * "plug" the device if there are no outstanding requests: this will
1555 * force the transfer to start only after we have put all the requests
1558 * This is called with interrupts off and no requests on the queue and
1559 * with the queue lock held.
1561 void blk_plug_device(struct request_queue *q)
1563 WARN_ON(!irqs_disabled());
1566 * don't plug a stopped queue, it must be paired with blk_start_queue()
1567 * which will restart the queueing
1569 if (blk_queue_stopped(q))
1572 if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
1573 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1574 blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
1578 EXPORT_SYMBOL(blk_plug_device);
1581 * remove the queue from the plugged list, if present. called with
1582 * queue lock held and interrupts disabled.
1584 int blk_remove_plug(struct request_queue *q)
1586 WARN_ON(!irqs_disabled());
1588 if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1591 del_timer(&q->unplug_timer);
1595 EXPORT_SYMBOL(blk_remove_plug);
1598 * remove the plug and let it rip..
1600 void __generic_unplug_device(struct request_queue *q)
1602 if (unlikely(blk_queue_stopped(q)))
1605 if (!blk_remove_plug(q))
1610 EXPORT_SYMBOL(__generic_unplug_device);
1613 * generic_unplug_device - fire a request queue
1614 * @q: The &struct request_queue in question
1617 * Linux uses plugging to build bigger requests queues before letting
1618 * the device have at them. If a queue is plugged, the I/O scheduler
1619 * is still adding and merging requests on the queue. Once the queue
1620 * gets unplugged, the request_fn defined for the queue is invoked and
1621 * transfers started.
1623 void generic_unplug_device(struct request_queue *q)
1625 spin_lock_irq(q->queue_lock);
1626 __generic_unplug_device(q);
1627 spin_unlock_irq(q->queue_lock);
1629 EXPORT_SYMBOL(generic_unplug_device);
1631 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1634 struct request_queue *q = bdi->unplug_io_data;
1637 * devices don't necessarily have an ->unplug_fn defined
1640 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1641 q->rq.count[READ] + q->rq.count[WRITE]);
1647 static void blk_unplug_work(struct work_struct *work)
1649 struct request_queue *q =
1650 container_of(work, struct request_queue, unplug_work);
1652 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1653 q->rq.count[READ] + q->rq.count[WRITE]);
1658 static void blk_unplug_timeout(unsigned long data)
1660 struct request_queue *q = (struct request_queue *)data;
1662 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
1663 q->rq.count[READ] + q->rq.count[WRITE]);
1665 kblockd_schedule_work(&q->unplug_work);
1669 * blk_start_queue - restart a previously stopped queue
1670 * @q: The &struct request_queue in question
1673 * blk_start_queue() will clear the stop flag on the queue, and call
1674 * the request_fn for the queue if it was in a stopped state when
1675 * entered. Also see blk_stop_queue(). Queue lock must be held.
1677 void blk_start_queue(struct request_queue *q)
1679 WARN_ON(!irqs_disabled());
1681 clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1684 * one level of recursion is ok and is much faster than kicking
1685 * the unplug handling
1687 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1689 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1692 kblockd_schedule_work(&q->unplug_work);
1696 EXPORT_SYMBOL(blk_start_queue);
1699 * blk_stop_queue - stop a queue
1700 * @q: The &struct request_queue in question
1703 * The Linux block layer assumes that a block driver will consume all
1704 * entries on the request queue when the request_fn strategy is called.
1705 * Often this will not happen, because of hardware limitations (queue
1706 * depth settings). If a device driver gets a 'queue full' response,
1707 * or if it simply chooses not to queue more I/O at one point, it can
1708 * call this function to prevent the request_fn from being called until
1709 * the driver has signalled it's ready to go again. This happens by calling
1710 * blk_start_queue() to restart queue operations. Queue lock must be held.
1712 void blk_stop_queue(struct request_queue *q)
1715 set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1717 EXPORT_SYMBOL(blk_stop_queue);
1720 * blk_sync_queue - cancel any pending callbacks on a queue
1724 * The block layer may perform asynchronous callback activity
1725 * on a queue, such as calling the unplug function after a timeout.
1726 * A block device may call blk_sync_queue to ensure that any
1727 * such activity is cancelled, thus allowing it to release resources
1728 * that the callbacks might use. The caller must already have made sure
1729 * that its ->make_request_fn will not re-add plugging prior to calling
1733 void blk_sync_queue(struct request_queue *q)
1735 del_timer_sync(&q->unplug_timer);
1737 EXPORT_SYMBOL(blk_sync_queue);
1740 * blk_run_queue - run a single device queue
1741 * @q: The queue to run
1743 void blk_run_queue(struct request_queue *q)
1745 unsigned long flags;
1747 spin_lock_irqsave(q->queue_lock, flags);
1751 * Only recurse once to avoid overrunning the stack, let the unplug
1752 * handling reinvoke the handler shortly if we already got there.
1754 if (!elv_queue_empty(q)) {
1755 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1757 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1760 kblockd_schedule_work(&q->unplug_work);
1764 spin_unlock_irqrestore(q->queue_lock, flags);
1766 EXPORT_SYMBOL(blk_run_queue);
1769 * blk_cleanup_queue: - release a &struct request_queue when it is no longer needed
1770 * @kobj: the kobj belonging of the request queue to be released
1773 * blk_cleanup_queue is the pair to blk_init_queue() or
1774 * blk_queue_make_request(). It should be called when a request queue is
1775 * being released; typically when a block device is being de-registered.
1776 * Currently, its primary task it to free all the &struct request
1777 * structures that were allocated to the queue and the queue itself.
1780 * Hopefully the low level driver will have finished any
1781 * outstanding requests first...
1783 static void blk_release_queue(struct kobject *kobj)
1785 struct request_queue *q =
1786 container_of(kobj, struct request_queue, kobj);
1787 struct request_list *rl = &q->rq;
1792 mempool_destroy(rl->rq_pool);
1795 __blk_queue_free_tags(q);
1797 blk_trace_shutdown(q);
1799 kmem_cache_free(requestq_cachep, q);
1802 void blk_put_queue(struct request_queue *q)
1804 kobject_put(&q->kobj);
1806 EXPORT_SYMBOL(blk_put_queue);
1808 void blk_cleanup_queue(struct request_queue * q)
1810 mutex_lock(&q->sysfs_lock);
1811 set_bit(QUEUE_FLAG_DEAD, &q->queue_flags);
1812 mutex_unlock(&q->sysfs_lock);
1815 elevator_exit(q->elevator);
1820 EXPORT_SYMBOL(blk_cleanup_queue);
1822 static int blk_init_free_list(struct request_queue *q)
1824 struct request_list *rl = &q->rq;
1826 rl->count[READ] = rl->count[WRITE] = 0;
1827 rl->starved[READ] = rl->starved[WRITE] = 0;
1829 init_waitqueue_head(&rl->wait[READ]);
1830 init_waitqueue_head(&rl->wait[WRITE]);
1832 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1833 mempool_free_slab, request_cachep, q->node);
1841 struct request_queue *blk_alloc_queue(gfp_t gfp_mask)
1843 return blk_alloc_queue_node(gfp_mask, -1);
1845 EXPORT_SYMBOL(blk_alloc_queue);
1847 static struct kobj_type queue_ktype;
1849 struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
1851 struct request_queue *q;
1853 q = kmem_cache_alloc_node(requestq_cachep,
1854 gfp_mask | __GFP_ZERO, node_id);
1858 init_timer(&q->unplug_timer);
1860 kobject_set_name(&q->kobj, "%s", "queue");
1861 q->kobj.ktype = &queue_ktype;
1862 kobject_init(&q->kobj);
1864 q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1865 q->backing_dev_info.unplug_io_data = q;
1867 mutex_init(&q->sysfs_lock);
1871 EXPORT_SYMBOL(blk_alloc_queue_node);
1874 * blk_init_queue - prepare a request queue for use with a block device
1875 * @rfn: The function to be called to process requests that have been
1876 * placed on the queue.
1877 * @lock: Request queue spin lock
1880 * If a block device wishes to use the standard request handling procedures,
1881 * which sorts requests and coalesces adjacent requests, then it must
1882 * call blk_init_queue(). The function @rfn will be called when there
1883 * are requests on the queue that need to be processed. If the device
1884 * supports plugging, then @rfn may not be called immediately when requests
1885 * are available on the queue, but may be called at some time later instead.
1886 * Plugged queues are generally unplugged when a buffer belonging to one
1887 * of the requests on the queue is needed, or due to memory pressure.
1889 * @rfn is not required, or even expected, to remove all requests off the
1890 * queue, but only as many as it can handle at a time. If it does leave
1891 * requests on the queue, it is responsible for arranging that the requests
1892 * get dealt with eventually.
1894 * The queue spin lock must be held while manipulating the requests on the
1895 * request queue; this lock will be taken also from interrupt context, so irq
1896 * disabling is needed for it.
1898 * Function returns a pointer to the initialized request queue, or NULL if
1899 * it didn't succeed.
1902 * blk_init_queue() must be paired with a blk_cleanup_queue() call
1903 * when the block device is deactivated (such as at module unload).
1906 struct request_queue *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1908 return blk_init_queue_node(rfn, lock, -1);
1910 EXPORT_SYMBOL(blk_init_queue);
1912 struct request_queue *
1913 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1915 struct request_queue *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1921 if (blk_init_free_list(q)) {
1922 kmem_cache_free(requestq_cachep, q);
1927 * if caller didn't supply a lock, they get per-queue locking with
1931 spin_lock_init(&q->__queue_lock);
1932 lock = &q->__queue_lock;
1935 q->request_fn = rfn;
1936 q->prep_rq_fn = NULL;
1937 q->unplug_fn = generic_unplug_device;
1938 q->queue_flags = (1 << QUEUE_FLAG_CLUSTER);
1939 q->queue_lock = lock;
1941 blk_queue_segment_boundary(q, 0xffffffff);
1943 blk_queue_make_request(q, __make_request);
1944 blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1946 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1947 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1949 q->sg_reserved_size = INT_MAX;
1954 if (!elevator_init(q, NULL)) {
1955 blk_queue_congestion_threshold(q);
1962 EXPORT_SYMBOL(blk_init_queue_node);
1964 int blk_get_queue(struct request_queue *q)
1966 if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1967 kobject_get(&q->kobj);
1974 EXPORT_SYMBOL(blk_get_queue);
1976 static inline void blk_free_request(struct request_queue *q, struct request *rq)
1978 if (rq->cmd_flags & REQ_ELVPRIV)
1979 elv_put_request(q, rq);
1980 mempool_free(rq, q->rq.rq_pool);
1983 static struct request *
1984 blk_alloc_request(struct request_queue *q, int rw, int priv, gfp_t gfp_mask)
1986 struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1992 * first three bits are identical in rq->cmd_flags and bio->bi_rw,
1993 * see bio.h and blkdev.h
1995 rq->cmd_flags = rw | REQ_ALLOCED;
1998 if (unlikely(elv_set_request(q, rq, gfp_mask))) {
1999 mempool_free(rq, q->rq.rq_pool);
2002 rq->cmd_flags |= REQ_ELVPRIV;
2009 * ioc_batching returns true if the ioc is a valid batching request and
2010 * should be given priority access to a request.
2012 static inline int ioc_batching(struct request_queue *q, struct io_context *ioc)
2018 * Make sure the process is able to allocate at least 1 request
2019 * even if the batch times out, otherwise we could theoretically
2022 return ioc->nr_batch_requests == q->nr_batching ||
2023 (ioc->nr_batch_requests > 0
2024 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
2028 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
2029 * will cause the process to be a "batcher" on all queues in the system. This
2030 * is the behaviour we want though - once it gets a wakeup it should be given
2033 static void ioc_set_batching(struct request_queue *q, struct io_context *ioc)
2035 if (!ioc || ioc_batching(q, ioc))
2038 ioc->nr_batch_requests = q->nr_batching;
2039 ioc->last_waited = jiffies;
2042 static void __freed_request(struct request_queue *q, int rw)
2044 struct request_list *rl = &q->rq;
2046 if (rl->count[rw] < queue_congestion_off_threshold(q))
2047 blk_clear_queue_congested(q, rw);
2049 if (rl->count[rw] + 1 <= q->nr_requests) {
2050 if (waitqueue_active(&rl->wait[rw]))
2051 wake_up(&rl->wait[rw]);
2053 blk_clear_queue_full(q, rw);
2058 * A request has just been released. Account for it, update the full and
2059 * congestion status, wake up any waiters. Called under q->queue_lock.
2061 static void freed_request(struct request_queue *q, int rw, int priv)
2063 struct request_list *rl = &q->rq;
2069 __freed_request(q, rw);
2071 if (unlikely(rl->starved[rw ^ 1]))
2072 __freed_request(q, rw ^ 1);
2075 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2077 * Get a free request, queue_lock must be held.
2078 * Returns NULL on failure, with queue_lock held.
2079 * Returns !NULL on success, with queue_lock *not held*.
2081 static struct request *get_request(struct request_queue *q, int rw_flags,
2082 struct bio *bio, gfp_t gfp_mask)
2084 struct request *rq = NULL;
2085 struct request_list *rl = &q->rq;
2086 struct io_context *ioc = NULL;
2087 const int rw = rw_flags & 0x01;
2088 int may_queue, priv;
2090 may_queue = elv_may_queue(q, rw_flags);
2091 if (may_queue == ELV_MQUEUE_NO)
2094 if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
2095 if (rl->count[rw]+1 >= q->nr_requests) {
2096 ioc = current_io_context(GFP_ATOMIC, q->node);
2098 * The queue will fill after this allocation, so set
2099 * it as full, and mark this process as "batching".
2100 * This process will be allowed to complete a batch of
2101 * requests, others will be blocked.
2103 if (!blk_queue_full(q, rw)) {
2104 ioc_set_batching(q, ioc);
2105 blk_set_queue_full(q, rw);
2107 if (may_queue != ELV_MQUEUE_MUST
2108 && !ioc_batching(q, ioc)) {
2110 * The queue is full and the allocating
2111 * process is not a "batcher", and not
2112 * exempted by the IO scheduler
2118 blk_set_queue_congested(q, rw);
2122 * Only allow batching queuers to allocate up to 50% over the defined
2123 * limit of requests, otherwise we could have thousands of requests
2124 * allocated with any setting of ->nr_requests
2126 if (rl->count[rw] >= (3 * q->nr_requests / 2))
2130 rl->starved[rw] = 0;
2132 priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
2136 spin_unlock_irq(q->queue_lock);
2138 rq = blk_alloc_request(q, rw_flags, priv, gfp_mask);
2139 if (unlikely(!rq)) {
2141 * Allocation failed presumably due to memory. Undo anything
2142 * we might have messed up.
2144 * Allocating task should really be put onto the front of the
2145 * wait queue, but this is pretty rare.
2147 spin_lock_irq(q->queue_lock);
2148 freed_request(q, rw, priv);
2151 * in the very unlikely event that allocation failed and no
2152 * requests for this direction was pending, mark us starved
2153 * so that freeing of a request in the other direction will
2154 * notice us. another possible fix would be to split the
2155 * rq mempool into READ and WRITE
2158 if (unlikely(rl->count[rw] == 0))
2159 rl->starved[rw] = 1;
2165 * ioc may be NULL here, and ioc_batching will be false. That's
2166 * OK, if the queue is under the request limit then requests need
2167 * not count toward the nr_batch_requests limit. There will always
2168 * be some limit enforced by BLK_BATCH_TIME.
2170 if (ioc_batching(q, ioc))
2171 ioc->nr_batch_requests--;
2175 blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
2181 * No available requests for this queue, unplug the device and wait for some
2182 * requests to become available.
2184 * Called with q->queue_lock held, and returns with it unlocked.
2186 static struct request *get_request_wait(struct request_queue *q, int rw_flags,
2189 const int rw = rw_flags & 0x01;
2192 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2195 struct request_list *rl = &q->rq;
2197 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2198 TASK_UNINTERRUPTIBLE);
2200 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2203 struct io_context *ioc;
2205 blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
2207 __generic_unplug_device(q);
2208 spin_unlock_irq(q->queue_lock);
2212 * After sleeping, we become a "batching" process and
2213 * will be able to allocate at least one request, and
2214 * up to a big batch of them for a small period time.
2215 * See ioc_batching, ioc_set_batching
2217 ioc = current_io_context(GFP_NOIO, q->node);
2218 ioc_set_batching(q, ioc);
2220 spin_lock_irq(q->queue_lock);
2222 finish_wait(&rl->wait[rw], &wait);
2228 struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
2232 BUG_ON(rw != READ && rw != WRITE);
2234 spin_lock_irq(q->queue_lock);
2235 if (gfp_mask & __GFP_WAIT) {
2236 rq = get_request_wait(q, rw, NULL);
2238 rq = get_request(q, rw, NULL, gfp_mask);
2240 spin_unlock_irq(q->queue_lock);
2242 /* q->queue_lock is unlocked at this point */
2246 EXPORT_SYMBOL(blk_get_request);
2249 * blk_start_queueing - initiate dispatch of requests to device
2250 * @q: request queue to kick into gear
2252 * This is basically a helper to remove the need to know whether a queue
2253 * is plugged or not if someone just wants to initiate dispatch of requests
2256 * The queue lock must be held with interrupts disabled.
2258 void blk_start_queueing(struct request_queue *q)
2260 if (!blk_queue_plugged(q))
2263 __generic_unplug_device(q);
2265 EXPORT_SYMBOL(blk_start_queueing);
2268 * blk_requeue_request - put a request back on queue
2269 * @q: request queue where request should be inserted
2270 * @rq: request to be inserted
2273 * Drivers often keep queueing requests until the hardware cannot accept
2274 * more, when that condition happens we need to put the request back
2275 * on the queue. Must be called with queue lock held.
2277 void blk_requeue_request(struct request_queue *q, struct request *rq)
2279 blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
2281 if (blk_rq_tagged(rq))
2282 blk_queue_end_tag(q, rq);
2284 elv_requeue_request(q, rq);
2287 EXPORT_SYMBOL(blk_requeue_request);
2290 * blk_insert_request - insert a special request in to a request queue
2291 * @q: request queue where request should be inserted
2292 * @rq: request to be inserted
2293 * @at_head: insert request at head or tail of queue
2294 * @data: private data
2297 * Many block devices need to execute commands asynchronously, so they don't
2298 * block the whole kernel from preemption during request execution. This is
2299 * accomplished normally by inserting aritficial requests tagged as
2300 * REQ_SPECIAL in to the corresponding request queue, and letting them be
2301 * scheduled for actual execution by the request queue.
2303 * We have the option of inserting the head or the tail of the queue.
2304 * Typically we use the tail for new ioctls and so forth. We use the head
2305 * of the queue for things like a QUEUE_FULL message from a device, or a
2306 * host that is unable to accept a particular command.
2308 void blk_insert_request(struct request_queue *q, struct request *rq,
2309 int at_head, void *data)
2311 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2312 unsigned long flags;
2315 * tell I/O scheduler that this isn't a regular read/write (ie it
2316 * must not attempt merges on this) and that it acts as a soft
2319 rq->cmd_type = REQ_TYPE_SPECIAL;
2320 rq->cmd_flags |= REQ_SOFTBARRIER;
2324 spin_lock_irqsave(q->queue_lock, flags);
2327 * If command is tagged, release the tag
2329 if (blk_rq_tagged(rq))
2330 blk_queue_end_tag(q, rq);
2332 drive_stat_acct(rq, rq->nr_sectors, 1);
2333 __elv_add_request(q, rq, where, 0);
2334 blk_start_queueing(q);
2335 spin_unlock_irqrestore(q->queue_lock, flags);
2338 EXPORT_SYMBOL(blk_insert_request);
2340 static int __blk_rq_unmap_user(struct bio *bio)
2345 if (bio_flagged(bio, BIO_USER_MAPPED))
2346 bio_unmap_user(bio);
2348 ret = bio_uncopy_user(bio);
2354 int blk_rq_append_bio(struct request_queue *q, struct request *rq,
2358 blk_rq_bio_prep(q, rq, bio);
2359 else if (!ll_back_merge_fn(q, rq, bio))
2362 rq->biotail->bi_next = bio;
2365 rq->data_len += bio->bi_size;
2369 EXPORT_SYMBOL(blk_rq_append_bio);
2371 static int __blk_rq_map_user(struct request_queue *q, struct request *rq,
2372 void __user *ubuf, unsigned int len)
2374 unsigned long uaddr;
2375 struct bio *bio, *orig_bio;
2378 reading = rq_data_dir(rq) == READ;
2381 * if alignment requirement is satisfied, map in user pages for
2382 * direct dma. else, set up kernel bounce buffers
2384 uaddr = (unsigned long) ubuf;
2385 if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2386 bio = bio_map_user(q, NULL, uaddr, len, reading);
2388 bio = bio_copy_user(q, uaddr, len, reading);
2391 return PTR_ERR(bio);
2394 blk_queue_bounce(q, &bio);
2397 * We link the bounce buffer in and could have to traverse it
2398 * later so we have to get a ref to prevent it from being freed
2402 ret = blk_rq_append_bio(q, rq, bio);
2404 return bio->bi_size;
2406 /* if it was boucned we must call the end io function */
2408 __blk_rq_unmap_user(orig_bio);
2414 * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2415 * @q: request queue where request should be inserted
2416 * @rq: request structure to fill
2417 * @ubuf: the user buffer
2418 * @len: length of user data
2421 * Data will be mapped directly for zero copy io, if possible. Otherwise
2422 * a kernel bounce buffer is used.
2424 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2425 * still in process context.
2427 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2428 * before being submitted to the device, as pages mapped may be out of
2429 * reach. It's the callers responsibility to make sure this happens. The
2430 * original bio must be passed back in to blk_rq_unmap_user() for proper
2433 int blk_rq_map_user(struct request_queue *q, struct request *rq,
2434 void __user *ubuf, unsigned long len)
2436 unsigned long bytes_read = 0;
2437 struct bio *bio = NULL;
2440 if (len > (q->max_hw_sectors << 9))
2445 while (bytes_read != len) {
2446 unsigned long map_len, end, start;
2448 map_len = min_t(unsigned long, len - bytes_read, BIO_MAX_SIZE);
2449 end = ((unsigned long)ubuf + map_len + PAGE_SIZE - 1)
2451 start = (unsigned long)ubuf >> PAGE_SHIFT;
2454 * A bad offset could cause us to require BIO_MAX_PAGES + 1
2455 * pages. If this happens we just lower the requested
2456 * mapping len by a page so that we can fit
2458 if (end - start > BIO_MAX_PAGES)
2459 map_len -= PAGE_SIZE;
2461 ret = __blk_rq_map_user(q, rq, ubuf, map_len);
2470 rq->buffer = rq->data = NULL;
2473 blk_rq_unmap_user(bio);
2477 EXPORT_SYMBOL(blk_rq_map_user);
2480 * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2481 * @q: request queue where request should be inserted
2482 * @rq: request to map data to
2483 * @iov: pointer to the iovec
2484 * @iov_count: number of elements in the iovec
2485 * @len: I/O byte count
2488 * Data will be mapped directly for zero copy io, if possible. Otherwise
2489 * a kernel bounce buffer is used.
2491 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2492 * still in process context.
2494 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2495 * before being submitted to the device, as pages mapped may be out of
2496 * reach. It's the callers responsibility to make sure this happens. The
2497 * original bio must be passed back in to blk_rq_unmap_user() for proper
2500 int blk_rq_map_user_iov(struct request_queue *q, struct request *rq,
2501 struct sg_iovec *iov, int iov_count, unsigned int len)
2505 if (!iov || iov_count <= 0)
2508 /* we don't allow misaligned data like bio_map_user() does. If the
2509 * user is using sg, they're expected to know the alignment constraints
2510 * and respect them accordingly */
2511 bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2513 return PTR_ERR(bio);
2515 if (bio->bi_size != len) {
2517 bio_unmap_user(bio);
2522 blk_rq_bio_prep(q, rq, bio);
2523 rq->buffer = rq->data = NULL;
2527 EXPORT_SYMBOL(blk_rq_map_user_iov);
2530 * blk_rq_unmap_user - unmap a request with user data
2531 * @bio: start of bio list
2534 * Unmap a rq previously mapped by blk_rq_map_user(). The caller must
2535 * supply the original rq->bio from the blk_rq_map_user() return, since
2536 * the io completion may have changed rq->bio.
2538 int blk_rq_unmap_user(struct bio *bio)
2540 struct bio *mapped_bio;
2545 if (unlikely(bio_flagged(bio, BIO_BOUNCED)))
2546 mapped_bio = bio->bi_private;
2548 ret2 = __blk_rq_unmap_user(mapped_bio);
2554 bio_put(mapped_bio);
2560 EXPORT_SYMBOL(blk_rq_unmap_user);
2563 * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2564 * @q: request queue where request should be inserted
2565 * @rq: request to fill
2566 * @kbuf: the kernel buffer
2567 * @len: length of user data
2568 * @gfp_mask: memory allocation flags
2570 int blk_rq_map_kern(struct request_queue *q, struct request *rq, void *kbuf,
2571 unsigned int len, gfp_t gfp_mask)
2575 if (len > (q->max_hw_sectors << 9))
2580 bio = bio_map_kern(q, kbuf, len, gfp_mask);
2582 return PTR_ERR(bio);
2584 if (rq_data_dir(rq) == WRITE)
2585 bio->bi_rw |= (1 << BIO_RW);
2587 blk_rq_bio_prep(q, rq, bio);
2588 blk_queue_bounce(q, &rq->bio);
2589 rq->buffer = rq->data = NULL;
2593 EXPORT_SYMBOL(blk_rq_map_kern);
2596 * blk_execute_rq_nowait - insert a request into queue for execution
2597 * @q: queue to insert the request in
2598 * @bd_disk: matching gendisk
2599 * @rq: request to insert
2600 * @at_head: insert request at head or tail of queue
2601 * @done: I/O completion handler
2604 * Insert a fully prepared request at the back of the io scheduler queue
2605 * for execution. Don't wait for completion.
2607 void blk_execute_rq_nowait(struct request_queue *q, struct gendisk *bd_disk,
2608 struct request *rq, int at_head,
2611 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2613 rq->rq_disk = bd_disk;
2614 rq->cmd_flags |= REQ_NOMERGE;
2616 WARN_ON(irqs_disabled());
2617 spin_lock_irq(q->queue_lock);
2618 __elv_add_request(q, rq, where, 1);
2619 __generic_unplug_device(q);
2620 spin_unlock_irq(q->queue_lock);
2622 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
2625 * blk_execute_rq - insert a request into queue for execution
2626 * @q: queue to insert the request in
2627 * @bd_disk: matching gendisk
2628 * @rq: request to insert
2629 * @at_head: insert request at head or tail of queue
2632 * Insert a fully prepared request at the back of the io scheduler queue
2633 * for execution and wait for completion.
2635 int blk_execute_rq(struct request_queue *q, struct gendisk *bd_disk,
2636 struct request *rq, int at_head)
2638 DECLARE_COMPLETION_ONSTACK(wait);
2639 char sense[SCSI_SENSE_BUFFERSIZE];
2643 * we need an extra reference to the request, so we can look at
2644 * it after io completion
2649 memset(sense, 0, sizeof(sense));
2654 rq->end_io_data = &wait;
2655 blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
2656 wait_for_completion(&wait);
2664 EXPORT_SYMBOL(blk_execute_rq);
2667 * blkdev_issue_flush - queue a flush
2668 * @bdev: blockdev to issue flush for
2669 * @error_sector: error sector
2672 * Issue a flush for the block device in question. Caller can supply
2673 * room for storing the error offset in case of a flush error, if they
2674 * wish to. Caller must run wait_for_completion() on its own.
2676 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2678 struct request_queue *q;
2680 if (bdev->bd_disk == NULL)
2683 q = bdev_get_queue(bdev);
2686 if (!q->issue_flush_fn)
2689 return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2692 EXPORT_SYMBOL(blkdev_issue_flush);
2694 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2696 int rw = rq_data_dir(rq);
2698 if (!blk_fs_request(rq) || !rq->rq_disk)
2702 __disk_stat_inc(rq->rq_disk, merges[rw]);
2704 disk_round_stats(rq->rq_disk);
2705 rq->rq_disk->in_flight++;
2710 * add-request adds a request to the linked list.
2711 * queue lock is held and interrupts disabled, as we muck with the
2712 * request queue list.
2714 static inline void add_request(struct request_queue * q, struct request * req)
2716 drive_stat_acct(req, req->nr_sectors, 1);
2719 * elevator indicated where it wants this request to be
2720 * inserted at elevator_merge time
2722 __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2726 * disk_round_stats() - Round off the performance stats on a struct
2729 * The average IO queue length and utilisation statistics are maintained
2730 * by observing the current state of the queue length and the amount of
2731 * time it has been in this state for.
2733 * Normally, that accounting is done on IO completion, but that can result
2734 * in more than a second's worth of IO being accounted for within any one
2735 * second, leading to >100% utilisation. To deal with that, we call this
2736 * function to do a round-off before returning the results when reading
2737 * /proc/diskstats. This accounts immediately for all queue usage up to
2738 * the current jiffies and restarts the counters again.
2740 void disk_round_stats(struct gendisk *disk)
2742 unsigned long now = jiffies;
2744 if (now == disk->stamp)
2747 if (disk->in_flight) {
2748 __disk_stat_add(disk, time_in_queue,
2749 disk->in_flight * (now - disk->stamp));
2750 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2755 EXPORT_SYMBOL_GPL(disk_round_stats);
2758 * queue lock must be held
2760 void __blk_put_request(struct request_queue *q, struct request *req)
2764 if (unlikely(--req->ref_count))
2767 elv_completed_request(q, req);
2770 * Request may not have originated from ll_rw_blk. if not,
2771 * it didn't come out of our reserved rq pools
2773 if (req->cmd_flags & REQ_ALLOCED) {
2774 int rw = rq_data_dir(req);
2775 int priv = req->cmd_flags & REQ_ELVPRIV;
2777 BUG_ON(!list_empty(&req->queuelist));
2778 BUG_ON(!hlist_unhashed(&req->hash));
2780 blk_free_request(q, req);
2781 freed_request(q, rw, priv);
2785 EXPORT_SYMBOL_GPL(__blk_put_request);
2787 void blk_put_request(struct request *req)
2789 unsigned long flags;
2790 struct request_queue *q = req->q;
2793 * Gee, IDE calls in w/ NULL q. Fix IDE and remove the
2794 * following if (q) test.
2797 spin_lock_irqsave(q->queue_lock, flags);
2798 __blk_put_request(q, req);
2799 spin_unlock_irqrestore(q->queue_lock, flags);
2803 EXPORT_SYMBOL(blk_put_request);
2806 * blk_end_sync_rq - executes a completion event on a request
2807 * @rq: request to complete
2808 * @error: end io status of the request
2810 void blk_end_sync_rq(struct request *rq, int error)
2812 struct completion *waiting = rq->end_io_data;
2814 rq->end_io_data = NULL;
2815 __blk_put_request(rq->q, rq);
2818 * complete last, if this is a stack request the process (and thus
2819 * the rq pointer) could be invalid right after this complete()
2823 EXPORT_SYMBOL(blk_end_sync_rq);
2826 * Has to be called with the request spinlock acquired
2828 static int attempt_merge(struct request_queue *q, struct request *req,
2829 struct request *next)
2831 if (!rq_mergeable(req) || !rq_mergeable(next))
2837 if (req->sector + req->nr_sectors != next->sector)
2840 if (rq_data_dir(req) != rq_data_dir(next)
2841 || req->rq_disk != next->rq_disk
2846 * If we are allowed to merge, then append bio list
2847 * from next to rq and release next. merge_requests_fn
2848 * will have updated segment counts, update sector
2851 if (!ll_merge_requests_fn(q, req, next))
2855 * At this point we have either done a back merge
2856 * or front merge. We need the smaller start_time of
2857 * the merged requests to be the current request
2858 * for accounting purposes.
2860 if (time_after(req->start_time, next->start_time))
2861 req->start_time = next->start_time;
2863 req->biotail->bi_next = next->bio;
2864 req->biotail = next->biotail;
2866 req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2868 elv_merge_requests(q, req, next);
2871 disk_round_stats(req->rq_disk);
2872 req->rq_disk->in_flight--;
2875 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2877 __blk_put_request(q, next);
2881 static inline int attempt_back_merge(struct request_queue *q,
2884 struct request *next = elv_latter_request(q, rq);
2887 return attempt_merge(q, rq, next);
2892 static inline int attempt_front_merge(struct request_queue *q,
2895 struct request *prev = elv_former_request(q, rq);
2898 return attempt_merge(q, prev, rq);
2903 static void init_request_from_bio(struct request *req, struct bio *bio)
2905 req->cmd_type = REQ_TYPE_FS;
2908 * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2910 if (bio_rw_ahead(bio) || bio_failfast(bio))
2911 req->cmd_flags |= REQ_FAILFAST;
2914 * REQ_BARRIER implies no merging, but lets make it explicit
2916 if (unlikely(bio_barrier(bio)))
2917 req->cmd_flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2920 req->cmd_flags |= REQ_RW_SYNC;
2921 if (bio_rw_meta(bio))
2922 req->cmd_flags |= REQ_RW_META;
2925 req->hard_sector = req->sector = bio->bi_sector;
2926 req->ioprio = bio_prio(bio);
2927 req->start_time = jiffies;
2928 blk_rq_bio_prep(req->q, req, bio);
2931 static int __make_request(struct request_queue *q, struct bio *bio)
2933 struct request *req;
2934 int el_ret, nr_sectors, barrier, err;
2935 const unsigned short prio = bio_prio(bio);
2936 const int sync = bio_sync(bio);
2939 nr_sectors = bio_sectors(bio);
2942 * low level driver can indicate that it wants pages above a
2943 * certain limit bounced to low memory (ie for highmem, or even
2944 * ISA dma in theory)
2946 blk_queue_bounce(q, &bio);
2948 barrier = bio_barrier(bio);
2949 if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
2954 spin_lock_irq(q->queue_lock);
2956 if (unlikely(barrier) || elv_queue_empty(q))
2959 el_ret = elv_merge(q, &req, bio);
2961 case ELEVATOR_BACK_MERGE:
2962 BUG_ON(!rq_mergeable(req));
2964 if (!ll_back_merge_fn(q, req, bio))
2967 blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
2969 req->biotail->bi_next = bio;
2971 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2972 req->ioprio = ioprio_best(req->ioprio, prio);
2973 drive_stat_acct(req, nr_sectors, 0);
2974 if (!attempt_back_merge(q, req))
2975 elv_merged_request(q, req, el_ret);
2978 case ELEVATOR_FRONT_MERGE:
2979 BUG_ON(!rq_mergeable(req));
2981 if (!ll_front_merge_fn(q, req, bio))
2984 blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
2986 bio->bi_next = req->bio;
2990 * may not be valid. if the low level driver said
2991 * it didn't need a bounce buffer then it better
2992 * not touch req->buffer either...
2994 req->buffer = bio_data(bio);
2995 req->current_nr_sectors = bio_cur_sectors(bio);
2996 req->hard_cur_sectors = req->current_nr_sectors;
2997 req->sector = req->hard_sector = bio->bi_sector;
2998 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2999 req->ioprio = ioprio_best(req->ioprio, prio);
3000 drive_stat_acct(req, nr_sectors, 0);
3001 if (!attempt_front_merge(q, req))
3002 elv_merged_request(q, req, el_ret);
3005 /* ELV_NO_MERGE: elevator says don't/can't merge. */
3012 * This sync check and mask will be re-done in init_request_from_bio(),
3013 * but we need to set it earlier to expose the sync flag to the
3014 * rq allocator and io schedulers.
3016 rw_flags = bio_data_dir(bio);
3018 rw_flags |= REQ_RW_SYNC;
3021 * Grab a free request. This is might sleep but can not fail.
3022 * Returns with the queue unlocked.
3024 req = get_request_wait(q, rw_flags, bio);
3027 * After dropping the lock and possibly sleeping here, our request
3028 * may now be mergeable after it had proven unmergeable (above).
3029 * We don't worry about that case for efficiency. It won't happen
3030 * often, and the elevators are able to handle it.
3032 init_request_from_bio(req, bio);
3034 spin_lock_irq(q->queue_lock);
3035 if (elv_queue_empty(q))
3037 add_request(q, req);
3040 __generic_unplug_device(q);
3042 spin_unlock_irq(q->queue_lock);
3046 bio_endio(bio, err);
3051 * If bio->bi_dev is a partition, remap the location
3053 static inline void blk_partition_remap(struct bio *bio)
3055 struct block_device *bdev = bio->bi_bdev;
3057 if (bdev != bdev->bd_contains) {
3058 struct hd_struct *p = bdev->bd_part;
3059 const int rw = bio_data_dir(bio);
3061 p->sectors[rw] += bio_sectors(bio);
3064 bio->bi_sector += p->start_sect;
3065 bio->bi_bdev = bdev->bd_contains;
3067 blk_add_trace_remap(bdev_get_queue(bio->bi_bdev), bio,
3068 bdev->bd_dev, bio->bi_sector,
3069 bio->bi_sector - p->start_sect);
3073 static void handle_bad_sector(struct bio *bio)
3075 char b[BDEVNAME_SIZE];
3077 printk(KERN_INFO "attempt to access beyond end of device\n");
3078 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
3079 bdevname(bio->bi_bdev, b),
3081 (unsigned long long)bio->bi_sector + bio_sectors(bio),
3082 (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
3084 set_bit(BIO_EOF, &bio->bi_flags);
3087 #ifdef CONFIG_FAIL_MAKE_REQUEST
3089 static DECLARE_FAULT_ATTR(fail_make_request);
3091 static int __init setup_fail_make_request(char *str)
3093 return setup_fault_attr(&fail_make_request, str);
3095 __setup("fail_make_request=", setup_fail_make_request);
3097 static int should_fail_request(struct bio *bio)
3099 if ((bio->bi_bdev->bd_disk->flags & GENHD_FL_FAIL) ||
3100 (bio->bi_bdev->bd_part && bio->bi_bdev->bd_part->make_it_fail))
3101 return should_fail(&fail_make_request, bio->bi_size);
3106 static int __init fail_make_request_debugfs(void)
3108 return init_fault_attr_dentries(&fail_make_request,
3109 "fail_make_request");
3112 late_initcall(fail_make_request_debugfs);
3114 #else /* CONFIG_FAIL_MAKE_REQUEST */
3116 static inline int should_fail_request(struct bio *bio)
3121 #endif /* CONFIG_FAIL_MAKE_REQUEST */
3124 * generic_make_request: hand a buffer to its device driver for I/O
3125 * @bio: The bio describing the location in memory and on the device.
3127 * generic_make_request() is used to make I/O requests of block
3128 * devices. It is passed a &struct bio, which describes the I/O that needs
3131 * generic_make_request() does not return any status. The
3132 * success/failure status of the request, along with notification of
3133 * completion, is delivered asynchronously through the bio->bi_end_io
3134 * function described (one day) else where.
3136 * The caller of generic_make_request must make sure that bi_io_vec
3137 * are set to describe the memory buffer, and that bi_dev and bi_sector are
3138 * set to describe the device address, and the
3139 * bi_end_io and optionally bi_private are set to describe how
3140 * completion notification should be signaled.
3142 * generic_make_request and the drivers it calls may use bi_next if this
3143 * bio happens to be merged with someone else, and may change bi_dev and
3144 * bi_sector for remaps as it sees fit. So the values of these fields
3145 * should NOT be depended on after the call to generic_make_request.
3147 static inline void __generic_make_request(struct bio *bio)
3149 struct request_queue *q;
3151 sector_t old_sector;
3152 int ret, nr_sectors = bio_sectors(bio);
3156 /* Test device or partition size, when known. */
3157 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3159 sector_t sector = bio->bi_sector;
3161 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
3163 * This may well happen - the kernel calls bread()
3164 * without checking the size of the device, e.g., when
3165 * mounting a device.
3167 handle_bad_sector(bio);
3173 * Resolve the mapping until finished. (drivers are
3174 * still free to implement/resolve their own stacking
3175 * by explicitly returning 0)
3177 * NOTE: we don't repeat the blk_size check for each new device.
3178 * Stacking drivers are expected to know what they are doing.
3183 char b[BDEVNAME_SIZE];
3185 q = bdev_get_queue(bio->bi_bdev);
3188 "generic_make_request: Trying to access "
3189 "nonexistent block-device %s (%Lu)\n",
3190 bdevname(bio->bi_bdev, b),
3191 (long long) bio->bi_sector);
3193 bio_endio(bio, -EIO);
3197 if (unlikely(nr_sectors > q->max_hw_sectors)) {
3198 printk("bio too big device %s (%u > %u)\n",
3199 bdevname(bio->bi_bdev, b),
3205 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
3208 if (should_fail_request(bio))
3212 * If this device has partitions, remap block n
3213 * of partition p to block n+start(p) of the disk.
3215 blk_partition_remap(bio);
3217 if (old_sector != -1)
3218 blk_add_trace_remap(q, bio, old_dev, bio->bi_sector,
3221 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
3223 old_sector = bio->bi_sector;
3224 old_dev = bio->bi_bdev->bd_dev;
3226 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3228 sector_t sector = bio->bi_sector;
3230 if (maxsector < nr_sectors ||
3231 maxsector - nr_sectors < sector) {
3233 * This may well happen - partitions are not
3234 * checked to make sure they are within the size
3235 * of the whole device.
3237 handle_bad_sector(bio);
3242 ret = q->make_request_fn(q, bio);
3247 * We only want one ->make_request_fn to be active at a time,
3248 * else stack usage with stacked devices could be a problem.
3249 * So use current->bio_{list,tail} to keep a list of requests
3250 * submited by a make_request_fn function.
3251 * current->bio_tail is also used as a flag to say if
3252 * generic_make_request is currently active in this task or not.
3253 * If it is NULL, then no make_request is active. If it is non-NULL,
3254 * then a make_request is active, and new requests should be added
3257 void generic_make_request(struct bio *bio)
3259 if (current->bio_tail) {
3260 /* make_request is active */
3261 *(current->bio_tail) = bio;
3262 bio->bi_next = NULL;
3263 current->bio_tail = &bio->bi_next;
3266 /* following loop may be a bit non-obvious, and so deserves some
3268 * Before entering the loop, bio->bi_next is NULL (as all callers
3269 * ensure that) so we have a list with a single bio.
3270 * We pretend that we have just taken it off a longer list, so
3271 * we assign bio_list to the next (which is NULL) and bio_tail
3272 * to &bio_list, thus initialising the bio_list of new bios to be
3273 * added. __generic_make_request may indeed add some more bios
3274 * through a recursive call to generic_make_request. If it
3275 * did, we find a non-NULL value in bio_list and re-enter the loop
3276 * from the top. In this case we really did just take the bio
3277 * of the top of the list (no pretending) and so fixup bio_list and
3278 * bio_tail or bi_next, and call into __generic_make_request again.
3280 * The loop was structured like this to make only one call to
3281 * __generic_make_request (which is important as it is large and
3282 * inlined) and to keep the structure simple.
3284 BUG_ON(bio->bi_next);
3286 current->bio_list = bio->bi_next;
3287 if (bio->bi_next == NULL)
3288 current->bio_tail = ¤t->bio_list;
3290 bio->bi_next = NULL;
3291 __generic_make_request(bio);
3292 bio = current->bio_list;
3294 current->bio_tail = NULL; /* deactivate */
3297 EXPORT_SYMBOL(generic_make_request);
3300 * submit_bio: submit a bio to the block device layer for I/O
3301 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3302 * @bio: The &struct bio which describes the I/O
3304 * submit_bio() is very similar in purpose to generic_make_request(), and
3305 * uses that function to do most of the work. Both are fairly rough
3306 * interfaces, @bio must be presetup and ready for I/O.
3309 void submit_bio(int rw, struct bio *bio)
3311 int count = bio_sectors(bio);
3313 BIO_BUG_ON(!bio->bi_size);
3314 BIO_BUG_ON(!bio->bi_io_vec);
3317 count_vm_events(PGPGOUT, count);
3319 task_io_account_read(bio->bi_size);
3320 count_vm_events(PGPGIN, count);
3323 if (unlikely(block_dump)) {
3324 char b[BDEVNAME_SIZE];
3325 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
3326 current->comm, current->pid,
3327 (rw & WRITE) ? "WRITE" : "READ",
3328 (unsigned long long)bio->bi_sector,
3329 bdevname(bio->bi_bdev,b));
3332 generic_make_request(bio);
3335 EXPORT_SYMBOL(submit_bio);
3337 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
3339 if (blk_fs_request(rq)) {
3340 rq->hard_sector += nsect;
3341 rq->hard_nr_sectors -= nsect;
3344 * Move the I/O submission pointers ahead if required.
3346 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3347 (rq->sector <= rq->hard_sector)) {
3348 rq->sector = rq->hard_sector;
3349 rq->nr_sectors = rq->hard_nr_sectors;
3350 rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3351 rq->current_nr_sectors = rq->hard_cur_sectors;
3352 rq->buffer = bio_data(rq->bio);
3356 * if total number of sectors is less than the first segment
3357 * size, something has gone terribly wrong
3359 if (rq->nr_sectors < rq->current_nr_sectors) {
3360 printk("blk: request botched\n");
3361 rq->nr_sectors = rq->current_nr_sectors;
3366 static int __end_that_request_first(struct request *req, int uptodate,
3369 int total_bytes, bio_nbytes, error, next_idx = 0;
3372 blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
3375 * extend uptodate bool to allow < 0 value to be direct io error
3378 if (end_io_error(uptodate))
3379 error = !uptodate ? -EIO : uptodate;
3382 * for a REQ_BLOCK_PC request, we want to carry any eventual
3383 * sense key with us all the way through
3385 if (!blk_pc_request(req))
3389 if (blk_fs_request(req) && !(req->cmd_flags & REQ_QUIET))
3390 printk("end_request: I/O error, dev %s, sector %llu\n",
3391 req->rq_disk ? req->rq_disk->disk_name : "?",
3392 (unsigned long long)req->sector);
3395 if (blk_fs_request(req) && req->rq_disk) {
3396 const int rw = rq_data_dir(req);
3398 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
3401 total_bytes = bio_nbytes = 0;
3402 while ((bio = req->bio) != NULL) {
3405 if (nr_bytes >= bio->bi_size) {
3406 req->bio = bio->bi_next;
3407 nbytes = bio->bi_size;
3408 req_bio_endio(req, bio, nbytes, error);
3412 int idx = bio->bi_idx + next_idx;
3414 if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3415 blk_dump_rq_flags(req, "__end_that");
3416 printk("%s: bio idx %d >= vcnt %d\n",
3418 bio->bi_idx, bio->bi_vcnt);
3422 nbytes = bio_iovec_idx(bio, idx)->bv_len;
3423 BIO_BUG_ON(nbytes > bio->bi_size);
3426 * not a complete bvec done
3428 if (unlikely(nbytes > nr_bytes)) {
3429 bio_nbytes += nr_bytes;
3430 total_bytes += nr_bytes;
3435 * advance to the next vector
3438 bio_nbytes += nbytes;
3441 total_bytes += nbytes;
3444 if ((bio = req->bio)) {
3446 * end more in this run, or just return 'not-done'
3448 if (unlikely(nr_bytes <= 0))
3460 * if the request wasn't completed, update state
3463 req_bio_endio(req, bio, bio_nbytes, error);
3464 bio->bi_idx += next_idx;
3465 bio_iovec(bio)->bv_offset += nr_bytes;
3466 bio_iovec(bio)->bv_len -= nr_bytes;
3469 blk_recalc_rq_sectors(req, total_bytes >> 9);
3470 blk_recalc_rq_segments(req);
3475 * end_that_request_first - end I/O on a request
3476 * @req: the request being processed
3477 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3478 * @nr_sectors: number of sectors to end I/O on
3481 * Ends I/O on a number of sectors attached to @req, and sets it up
3482 * for the next range of segments (if any) in the cluster.
3485 * 0 - we are done with this request, call end_that_request_last()
3486 * 1 - still buffers pending for this request
3488 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3490 return __end_that_request_first(req, uptodate, nr_sectors << 9);
3493 EXPORT_SYMBOL(end_that_request_first);
3496 * end_that_request_chunk - end I/O on a request
3497 * @req: the request being processed
3498 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3499 * @nr_bytes: number of bytes to complete
3502 * Ends I/O on a number of bytes attached to @req, and sets it up
3503 * for the next range of segments (if any). Like end_that_request_first(),
3504 * but deals with bytes instead of sectors.
3507 * 0 - we are done with this request, call end_that_request_last()
3508 * 1 - still buffers pending for this request
3510 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3512 return __end_that_request_first(req, uptodate, nr_bytes);
3515 EXPORT_SYMBOL(end_that_request_chunk);
3518 * splice the completion data to a local structure and hand off to
3519 * process_completion_queue() to complete the requests
3521 static void blk_done_softirq(struct softirq_action *h)
3523 struct list_head *cpu_list, local_list;
3525 local_irq_disable();
3526 cpu_list = &__get_cpu_var(blk_cpu_done);
3527 list_replace_init(cpu_list, &local_list);
3530 while (!list_empty(&local_list)) {
3531 struct request *rq = list_entry(local_list.next, struct request, donelist);
3533 list_del_init(&rq->donelist);
3534 rq->q->softirq_done_fn(rq);
3538 static int __cpuinit blk_cpu_notify(struct notifier_block *self, unsigned long action,
3542 * If a CPU goes away, splice its entries to the current CPU
3543 * and trigger a run of the softirq
3545 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
3546 int cpu = (unsigned long) hcpu;
3548 local_irq_disable();
3549 list_splice_init(&per_cpu(blk_cpu_done, cpu),
3550 &__get_cpu_var(blk_cpu_done));
3551 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3559 static struct notifier_block blk_cpu_notifier __cpuinitdata = {
3560 .notifier_call = blk_cpu_notify,
3564 * blk_complete_request - end I/O on a request
3565 * @req: the request being processed
3568 * Ends all I/O on a request. It does not handle partial completions,
3569 * unless the driver actually implements this in its completion callback
3570 * through requeueing. The actual completion happens out-of-order,
3571 * through a softirq handler. The user must have registered a completion
3572 * callback through blk_queue_softirq_done().
3575 void blk_complete_request(struct request *req)
3577 struct list_head *cpu_list;
3578 unsigned long flags;
3580 BUG_ON(!req->q->softirq_done_fn);
3582 local_irq_save(flags);
3584 cpu_list = &__get_cpu_var(blk_cpu_done);
3585 list_add_tail(&req->donelist, cpu_list);
3586 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3588 local_irq_restore(flags);
3591 EXPORT_SYMBOL(blk_complete_request);
3594 * queue lock must be held
3596 void end_that_request_last(struct request *req, int uptodate)
3598 struct gendisk *disk = req->rq_disk;
3602 * extend uptodate bool to allow < 0 value to be direct io error
3605 if (end_io_error(uptodate))
3606 error = !uptodate ? -EIO : uptodate;
3608 if (unlikely(laptop_mode) && blk_fs_request(req))
3609 laptop_io_completion();
3612 * Account IO completion. bar_rq isn't accounted as a normal
3613 * IO on queueing nor completion. Accounting the containing
3614 * request is enough.
3616 if (disk && blk_fs_request(req) && req != &req->q->bar_rq) {
3617 unsigned long duration = jiffies - req->start_time;
3618 const int rw = rq_data_dir(req);
3620 __disk_stat_inc(disk, ios[rw]);
3621 __disk_stat_add(disk, ticks[rw], duration);
3622 disk_round_stats(disk);
3626 req->end_io(req, error);
3628 __blk_put_request(req->q, req);
3631 EXPORT_SYMBOL(end_that_request_last);
3633 void end_request(struct request *req, int uptodate)
3635 if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3636 add_disk_randomness(req->rq_disk);
3637 blkdev_dequeue_request(req);
3638 end_that_request_last(req, uptodate);
3642 EXPORT_SYMBOL(end_request);
3644 static void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
3647 /* first two bits are identical in rq->cmd_flags and bio->bi_rw */
3648 rq->cmd_flags |= (bio->bi_rw & 3);
3650 rq->nr_phys_segments = bio_phys_segments(q, bio);
3651 rq->nr_hw_segments = bio_hw_segments(q, bio);
3652 rq->current_nr_sectors = bio_cur_sectors(bio);
3653 rq->hard_cur_sectors = rq->current_nr_sectors;
3654 rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3655 rq->buffer = bio_data(bio);
3656 rq->data_len = bio->bi_size;
3658 rq->bio = rq->biotail = bio;
3661 rq->rq_disk = bio->bi_bdev->bd_disk;
3664 int kblockd_schedule_work(struct work_struct *work)
3666 return queue_work(kblockd_workqueue, work);
3669 EXPORT_SYMBOL(kblockd_schedule_work);
3671 void kblockd_flush_work(struct work_struct *work)
3673 cancel_work_sync(work);
3675 EXPORT_SYMBOL(kblockd_flush_work);
3677 int __init blk_dev_init(void)
3681 kblockd_workqueue = create_workqueue("kblockd");
3682 if (!kblockd_workqueue)
3683 panic("Failed to create kblockd\n");
3685 request_cachep = kmem_cache_create("blkdev_requests",
3686 sizeof(struct request), 0, SLAB_PANIC, NULL);
3688 requestq_cachep = kmem_cache_create("blkdev_queue",
3689 sizeof(struct request_queue), 0, SLAB_PANIC, NULL);
3691 iocontext_cachep = kmem_cache_create("blkdev_ioc",
3692 sizeof(struct io_context), 0, SLAB_PANIC, NULL);
3694 for_each_possible_cpu(i)
3695 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3697 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
3698 register_hotcpu_notifier(&blk_cpu_notifier);
3700 blk_max_low_pfn = max_low_pfn - 1;
3701 blk_max_pfn = max_pfn - 1;
3707 * IO Context helper functions
3709 void put_io_context(struct io_context *ioc)
3714 BUG_ON(atomic_read(&ioc->refcount) == 0);
3716 if (atomic_dec_and_test(&ioc->refcount)) {
3717 struct cfq_io_context *cic;
3720 if (ioc->aic && ioc->aic->dtor)
3721 ioc->aic->dtor(ioc->aic);
3722 if (ioc->cic_root.rb_node != NULL) {
3723 struct rb_node *n = rb_first(&ioc->cic_root);
3725 cic = rb_entry(n, struct cfq_io_context, rb_node);
3730 kmem_cache_free(iocontext_cachep, ioc);
3733 EXPORT_SYMBOL(put_io_context);
3735 /* Called by the exitting task */
3736 void exit_io_context(void)
3738 struct io_context *ioc;
3739 struct cfq_io_context *cic;
3742 ioc = current->io_context;
3743 current->io_context = NULL;
3744 task_unlock(current);
3747 if (ioc->aic && ioc->aic->exit)
3748 ioc->aic->exit(ioc->aic);
3749 if (ioc->cic_root.rb_node != NULL) {
3750 cic = rb_entry(rb_first(&ioc->cic_root), struct cfq_io_context, rb_node);
3754 put_io_context(ioc);
3758 * If the current task has no IO context then create one and initialise it.
3759 * Otherwise, return its existing IO context.
3761 * This returned IO context doesn't have a specifically elevated refcount,
3762 * but since the current task itself holds a reference, the context can be
3763 * used in general code, so long as it stays within `current` context.
3765 static struct io_context *current_io_context(gfp_t gfp_flags, int node)
3767 struct task_struct *tsk = current;
3768 struct io_context *ret;
3770 ret = tsk->io_context;
3774 ret = kmem_cache_alloc_node(iocontext_cachep, gfp_flags, node);
3776 atomic_set(&ret->refcount, 1);
3777 ret->task = current;
3778 ret->ioprio_changed = 0;
3779 ret->last_waited = jiffies; /* doesn't matter... */
3780 ret->nr_batch_requests = 0; /* because this is 0 */
3782 ret->cic_root.rb_node = NULL;
3783 ret->ioc_data = NULL;
3784 /* make sure set_task_ioprio() sees the settings above */
3786 tsk->io_context = ret;
3793 * If the current task has no IO context then create one and initialise it.
3794 * If it does have a context, take a ref on it.
3796 * This is always called in the context of the task which submitted the I/O.
3798 struct io_context *get_io_context(gfp_t gfp_flags, int node)
3800 struct io_context *ret;
3801 ret = current_io_context(gfp_flags, node);
3803 atomic_inc(&ret->refcount);
3806 EXPORT_SYMBOL(get_io_context);
3808 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3810 struct io_context *src = *psrc;
3811 struct io_context *dst = *pdst;
3814 BUG_ON(atomic_read(&src->refcount) == 0);
3815 atomic_inc(&src->refcount);
3816 put_io_context(dst);
3820 EXPORT_SYMBOL(copy_io_context);
3822 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3824 struct io_context *temp;
3829 EXPORT_SYMBOL(swap_io_context);
3834 struct queue_sysfs_entry {
3835 struct attribute attr;
3836 ssize_t (*show)(struct request_queue *, char *);
3837 ssize_t (*store)(struct request_queue *, const char *, size_t);
3841 queue_var_show(unsigned int var, char *page)
3843 return sprintf(page, "%d\n", var);
3847 queue_var_store(unsigned long *var, const char *page, size_t count)
3849 char *p = (char *) page;
3851 *var = simple_strtoul(p, &p, 10);
3855 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3857 return queue_var_show(q->nr_requests, (page));
3861 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3863 struct request_list *rl = &q->rq;
3865 int ret = queue_var_store(&nr, page, count);
3866 if (nr < BLKDEV_MIN_RQ)
3869 spin_lock_irq(q->queue_lock);
3870 q->nr_requests = nr;
3871 blk_queue_congestion_threshold(q);
3873 if (rl->count[READ] >= queue_congestion_on_threshold(q))
3874 blk_set_queue_congested(q, READ);
3875 else if (rl->count[READ] < queue_congestion_off_threshold(q))
3876 blk_clear_queue_congested(q, READ);
3878 if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3879 blk_set_queue_congested(q, WRITE);
3880 else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3881 blk_clear_queue_congested(q, WRITE);
3883 if (rl->count[READ] >= q->nr_requests) {
3884 blk_set_queue_full(q, READ);
3885 } else if (rl->count[READ]+1 <= q->nr_requests) {
3886 blk_clear_queue_full(q, READ);
3887 wake_up(&rl->wait[READ]);
3890 if (rl->count[WRITE] >= q->nr_requests) {
3891 blk_set_queue_full(q, WRITE);
3892 } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3893 blk_clear_queue_full(q, WRITE);
3894 wake_up(&rl->wait[WRITE]);
3896 spin_unlock_irq(q->queue_lock);
3900 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3902 int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3904 return queue_var_show(ra_kb, (page));
3908 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3910 unsigned long ra_kb;
3911 ssize_t ret = queue_var_store(&ra_kb, page, count);
3913 spin_lock_irq(q->queue_lock);
3914 q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3915 spin_unlock_irq(q->queue_lock);
3920 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3922 int max_sectors_kb = q->max_sectors >> 1;
3924 return queue_var_show(max_sectors_kb, (page));
3928 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3930 unsigned long max_sectors_kb,
3931 max_hw_sectors_kb = q->max_hw_sectors >> 1,
3932 page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3933 ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3936 if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3939 * Take the queue lock to update the readahead and max_sectors
3940 * values synchronously:
3942 spin_lock_irq(q->queue_lock);
3944 * Trim readahead window as well, if necessary:
3946 ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3947 if (ra_kb > max_sectors_kb)
3948 q->backing_dev_info.ra_pages =
3949 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3951 q->max_sectors = max_sectors_kb << 1;
3952 spin_unlock_irq(q->queue_lock);
3957 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3959 int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3961 return queue_var_show(max_hw_sectors_kb, (page));
3965 static struct queue_sysfs_entry queue_requests_entry = {
3966 .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3967 .show = queue_requests_show,
3968 .store = queue_requests_store,
3971 static struct queue_sysfs_entry queue_ra_entry = {
3972 .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3973 .show = queue_ra_show,
3974 .store = queue_ra_store,
3977 static struct queue_sysfs_entry queue_max_sectors_entry = {
3978 .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3979 .show = queue_max_sectors_show,
3980 .store = queue_max_sectors_store,
3983 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3984 .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
3985 .show = queue_max_hw_sectors_show,
3988 static struct queue_sysfs_entry queue_iosched_entry = {
3989 .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
3990 .show = elv_iosched_show,
3991 .store = elv_iosched_store,
3994 static struct attribute *default_attrs[] = {
3995 &queue_requests_entry.attr,
3996 &queue_ra_entry.attr,
3997 &queue_max_hw_sectors_entry.attr,
3998 &queue_max_sectors_entry.attr,
3999 &queue_iosched_entry.attr,
4003 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
4006 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
4008 struct queue_sysfs_entry *entry = to_queue(attr);
4009 struct request_queue *q =
4010 container_of(kobj, struct request_queue, kobj);
4015 mutex_lock(&q->sysfs_lock);
4016 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
4017 mutex_unlock(&q->sysfs_lock);
4020 res = entry->show(q, page);
4021 mutex_unlock(&q->sysfs_lock);
4026 queue_attr_store(struct kobject *kobj, struct attribute *attr,
4027 const char *page, size_t length)
4029 struct queue_sysfs_entry *entry = to_queue(attr);
4030 struct request_queue *q = container_of(kobj, struct request_queue, kobj);
4036 mutex_lock(&q->sysfs_lock);
4037 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
4038 mutex_unlock(&q->sysfs_lock);
4041 res = entry->store(q, page, length);
4042 mutex_unlock(&q->sysfs_lock);
4046 static struct sysfs_ops queue_sysfs_ops = {
4047 .show = queue_attr_show,
4048 .store = queue_attr_store,
4051 static struct kobj_type queue_ktype = {
4052 .sysfs_ops = &queue_sysfs_ops,
4053 .default_attrs = default_attrs,
4054 .release = blk_release_queue,
4057 int blk_register_queue(struct gendisk *disk)
4061 struct request_queue *q = disk->queue;
4063 if (!q || !q->request_fn)
4066 q->kobj.parent = kobject_get(&disk->kobj);
4068 ret = kobject_add(&q->kobj);
4072 kobject_uevent(&q->kobj, KOBJ_ADD);
4074 ret = elv_register_queue(q);
4076 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4077 kobject_del(&q->kobj);
4084 void blk_unregister_queue(struct gendisk *disk)
4086 struct request_queue *q = disk->queue;
4088 if (q && q->request_fn) {
4089 elv_unregister_queue(q);
4091 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4092 kobject_del(&q->kobj);
4093 kobject_put(&disk->kobj);