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(request_queue_t *q, struct bio *bio);
44 static struct io_context *current_io_context(gfp_t gfp_flags, int node);
47 * For the allocated request tables
49 static struct kmem_cache *request_cachep;
52 * For queue allocation
54 static struct kmem_cache *requestq_cachep;
57 * For io context allocations
59 static struct kmem_cache *iocontext_cachep;
62 * Controlling structure to kblockd
64 static struct workqueue_struct *kblockd_workqueue;
66 unsigned long blk_max_low_pfn, blk_max_pfn;
68 EXPORT_SYMBOL(blk_max_low_pfn);
69 EXPORT_SYMBOL(blk_max_pfn);
71 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
73 /* Amount of time in which a process may batch requests */
74 #define BLK_BATCH_TIME (HZ/50UL)
76 /* Number of requests a "batching" process may submit */
77 #define BLK_BATCH_REQ 32
80 * Return the threshold (number of used requests) at which the queue is
81 * considered to be congested. It include a little hysteresis to keep the
82 * context switch rate down.
84 static inline int queue_congestion_on_threshold(struct request_queue *q)
86 return q->nr_congestion_on;
90 * The threshold at which a queue is considered to be uncongested
92 static inline int queue_congestion_off_threshold(struct request_queue *q)
94 return q->nr_congestion_off;
97 static void blk_queue_congestion_threshold(struct request_queue *q)
101 nr = q->nr_requests - (q->nr_requests / 8) + 1;
102 if (nr > q->nr_requests)
104 q->nr_congestion_on = nr;
106 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
109 q->nr_congestion_off = nr;
113 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
116 * Locates the passed device's request queue and returns the address of its
119 * Will return NULL if the request queue cannot be located.
121 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
123 struct backing_dev_info *ret = NULL;
124 request_queue_t *q = bdev_get_queue(bdev);
127 ret = &q->backing_dev_info;
130 EXPORT_SYMBOL(blk_get_backing_dev_info);
133 * blk_queue_prep_rq - set a prepare_request function for queue
135 * @pfn: prepare_request function
137 * It's possible for a queue to register a prepare_request callback which
138 * is invoked before the request is handed to the request_fn. The goal of
139 * the function is to prepare a request for I/O, it can be used to build a
140 * cdb from the request data for instance.
143 void blk_queue_prep_rq(request_queue_t *q, prep_rq_fn *pfn)
148 EXPORT_SYMBOL(blk_queue_prep_rq);
151 * blk_queue_merge_bvec - set a merge_bvec function for queue
153 * @mbfn: merge_bvec_fn
155 * Usually queues have static limitations on the max sectors or segments that
156 * we can put in a request. Stacking drivers may have some settings that
157 * are dynamic, and thus we have to query the queue whether it is ok to
158 * add a new bio_vec to a bio at a given offset or not. If the block device
159 * has such limitations, it needs to register a merge_bvec_fn to control
160 * the size of bio's sent to it. Note that a block device *must* allow a
161 * single page to be added to an empty bio. The block device driver may want
162 * to use the bio_split() function to deal with these bio's. By default
163 * no merge_bvec_fn is defined for a queue, and only the fixed limits are
166 void blk_queue_merge_bvec(request_queue_t *q, merge_bvec_fn *mbfn)
168 q->merge_bvec_fn = mbfn;
171 EXPORT_SYMBOL(blk_queue_merge_bvec);
173 void blk_queue_softirq_done(request_queue_t *q, softirq_done_fn *fn)
175 q->softirq_done_fn = fn;
178 EXPORT_SYMBOL(blk_queue_softirq_done);
181 * blk_queue_make_request - define an alternate make_request function for a device
182 * @q: the request queue for the device to be affected
183 * @mfn: the alternate make_request function
186 * The normal way for &struct bios to be passed to a device
187 * driver is for them to be collected into requests on a request
188 * queue, and then to allow the device driver to select requests
189 * off that queue when it is ready. This works well for many block
190 * devices. However some block devices (typically virtual devices
191 * such as md or lvm) do not benefit from the processing on the
192 * request queue, and are served best by having the requests passed
193 * directly to them. This can be achieved by providing a function
194 * to blk_queue_make_request().
197 * The driver that does this *must* be able to deal appropriately
198 * with buffers in "highmemory". This can be accomplished by either calling
199 * __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
200 * blk_queue_bounce() to create a buffer in normal memory.
202 void blk_queue_make_request(request_queue_t * q, make_request_fn * mfn)
207 q->nr_requests = BLKDEV_MAX_RQ;
208 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
209 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
210 q->make_request_fn = mfn;
211 q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
212 q->backing_dev_info.state = 0;
213 q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
214 blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
215 blk_queue_hardsect_size(q, 512);
216 blk_queue_dma_alignment(q, 511);
217 blk_queue_congestion_threshold(q);
218 q->nr_batching = BLK_BATCH_REQ;
220 q->unplug_thresh = 4; /* hmm */
221 q->unplug_delay = (3 * HZ) / 1000; /* 3 milliseconds */
222 if (q->unplug_delay == 0)
225 INIT_WORK(&q->unplug_work, blk_unplug_work);
227 q->unplug_timer.function = blk_unplug_timeout;
228 q->unplug_timer.data = (unsigned long)q;
231 * by default assume old behaviour and bounce for any highmem page
233 blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
236 EXPORT_SYMBOL(blk_queue_make_request);
238 static void rq_init(request_queue_t *q, struct request *rq)
240 INIT_LIST_HEAD(&rq->queuelist);
241 INIT_LIST_HEAD(&rq->donelist);
244 rq->bio = rq->biotail = NULL;
245 INIT_HLIST_NODE(&rq->hash);
246 RB_CLEAR_NODE(&rq->rb_node);
254 rq->nr_phys_segments = 0;
257 rq->end_io_data = NULL;
258 rq->completion_data = NULL;
262 * blk_queue_ordered - does this queue support ordered writes
263 * @q: the request queue
264 * @ordered: one of QUEUE_ORDERED_*
265 * @prepare_flush_fn: rq setup helper for cache flush ordered writes
268 * For journalled file systems, doing ordered writes on a commit
269 * block instead of explicitly doing wait_on_buffer (which is bad
270 * for performance) can be a big win. Block drivers supporting this
271 * feature should call this function and indicate so.
274 int blk_queue_ordered(request_queue_t *q, unsigned ordered,
275 prepare_flush_fn *prepare_flush_fn)
277 if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
278 prepare_flush_fn == NULL) {
279 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
283 if (ordered != QUEUE_ORDERED_NONE &&
284 ordered != QUEUE_ORDERED_DRAIN &&
285 ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
286 ordered != QUEUE_ORDERED_DRAIN_FUA &&
287 ordered != QUEUE_ORDERED_TAG &&
288 ordered != QUEUE_ORDERED_TAG_FLUSH &&
289 ordered != QUEUE_ORDERED_TAG_FUA) {
290 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
294 q->ordered = ordered;
295 q->next_ordered = ordered;
296 q->prepare_flush_fn = prepare_flush_fn;
301 EXPORT_SYMBOL(blk_queue_ordered);
304 * blk_queue_issue_flush_fn - set function for issuing a flush
305 * @q: the request queue
306 * @iff: the function to be called issuing the flush
309 * If a driver supports issuing a flush command, the support is notified
310 * to the block layer by defining it through this call.
313 void blk_queue_issue_flush_fn(request_queue_t *q, issue_flush_fn *iff)
315 q->issue_flush_fn = iff;
318 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
321 * Cache flushing for ordered writes handling
323 inline unsigned blk_ordered_cur_seq(request_queue_t *q)
327 return 1 << ffz(q->ordseq);
330 unsigned blk_ordered_req_seq(struct request *rq)
332 request_queue_t *q = rq->q;
334 BUG_ON(q->ordseq == 0);
336 if (rq == &q->pre_flush_rq)
337 return QUEUE_ORDSEQ_PREFLUSH;
338 if (rq == &q->bar_rq)
339 return QUEUE_ORDSEQ_BAR;
340 if (rq == &q->post_flush_rq)
341 return QUEUE_ORDSEQ_POSTFLUSH;
343 if ((rq->cmd_flags & REQ_ORDERED_COLOR) ==
344 (q->orig_bar_rq->cmd_flags & REQ_ORDERED_COLOR))
345 return QUEUE_ORDSEQ_DRAIN;
347 return QUEUE_ORDSEQ_DONE;
350 void blk_ordered_complete_seq(request_queue_t *q, unsigned seq, int error)
355 if (error && !q->orderr)
358 BUG_ON(q->ordseq & seq);
361 if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
365 * Okay, sequence complete.
368 uptodate = q->orderr ? q->orderr : 1;
372 end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
373 end_that_request_last(rq, uptodate);
376 static void pre_flush_end_io(struct request *rq, int error)
378 elv_completed_request(rq->q, rq);
379 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
382 static void bar_end_io(struct request *rq, int error)
384 elv_completed_request(rq->q, rq);
385 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
388 static void post_flush_end_io(struct request *rq, int error)
390 elv_completed_request(rq->q, rq);
391 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
394 static void queue_flush(request_queue_t *q, unsigned which)
397 rq_end_io_fn *end_io;
399 if (which == QUEUE_ORDERED_PREFLUSH) {
400 rq = &q->pre_flush_rq;
401 end_io = pre_flush_end_io;
403 rq = &q->post_flush_rq;
404 end_io = post_flush_end_io;
407 rq->cmd_flags = REQ_HARDBARRIER;
409 rq->elevator_private = NULL;
410 rq->elevator_private2 = NULL;
411 rq->rq_disk = q->bar_rq.rq_disk;
413 q->prepare_flush_fn(q, rq);
415 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
418 static inline struct request *start_ordered(request_queue_t *q,
423 q->ordered = q->next_ordered;
424 q->ordseq |= QUEUE_ORDSEQ_STARTED;
427 * Prep proxy barrier request.
429 blkdev_dequeue_request(rq);
434 if (bio_data_dir(q->orig_bar_rq->bio) == WRITE)
435 rq->cmd_flags |= REQ_RW;
436 rq->cmd_flags |= q->ordered & QUEUE_ORDERED_FUA ? REQ_FUA : 0;
437 rq->elevator_private = NULL;
438 rq->elevator_private2 = NULL;
439 init_request_from_bio(rq, q->orig_bar_rq->bio);
440 rq->end_io = bar_end_io;
443 * Queue ordered sequence. As we stack them at the head, we
444 * need to queue in reverse order. Note that we rely on that
445 * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
446 * request gets inbetween ordered sequence.
448 if (q->ordered & QUEUE_ORDERED_POSTFLUSH)
449 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
451 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
453 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
455 if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
456 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
457 rq = &q->pre_flush_rq;
459 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
461 if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
462 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
469 int blk_do_ordered(request_queue_t *q, struct request **rqp)
471 struct request *rq = *rqp;
472 int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
478 if (q->next_ordered != QUEUE_ORDERED_NONE) {
479 *rqp = start_ordered(q, rq);
483 * This can happen when the queue switches to
484 * ORDERED_NONE while this request is on it.
486 blkdev_dequeue_request(rq);
487 end_that_request_first(rq, -EOPNOTSUPP,
488 rq->hard_nr_sectors);
489 end_that_request_last(rq, -EOPNOTSUPP);
496 * Ordered sequence in progress
499 /* Special requests are not subject to ordering rules. */
500 if (!blk_fs_request(rq) &&
501 rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
504 if (q->ordered & QUEUE_ORDERED_TAG) {
505 /* Ordered by tag. Blocking the next barrier is enough. */
506 if (is_barrier && rq != &q->bar_rq)
509 /* Ordered by draining. Wait for turn. */
510 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
511 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
518 static int flush_dry_bio_endio(struct bio *bio, unsigned int bytes, int error)
520 request_queue_t *q = bio->bi_private;
521 struct bio_vec *bvec;
525 * This is dry run, restore bio_sector and size. We'll finish
526 * this request again with the original bi_end_io after an
527 * error occurs or post flush is complete.
536 bio_for_each_segment(bvec, bio, i) {
537 bvec->bv_len += bvec->bv_offset;
542 set_bit(BIO_UPTODATE, &bio->bi_flags);
543 bio->bi_size = q->bi_size;
544 bio->bi_sector -= (q->bi_size >> 9);
550 static int ordered_bio_endio(struct request *rq, struct bio *bio,
551 unsigned int nbytes, int error)
553 request_queue_t *q = rq->q;
557 if (&q->bar_rq != rq)
561 * Okay, this is the barrier request in progress, dry finish it.
563 if (error && !q->orderr)
566 endio = bio->bi_end_io;
567 private = bio->bi_private;
568 bio->bi_end_io = flush_dry_bio_endio;
571 bio_endio(bio, nbytes, error);
573 bio->bi_end_io = endio;
574 bio->bi_private = private;
580 * blk_queue_bounce_limit - set bounce buffer limit for queue
581 * @q: the request queue for the device
582 * @dma_addr: bus address limit
585 * Different hardware can have different requirements as to what pages
586 * it can do I/O directly to. A low level driver can call
587 * blk_queue_bounce_limit to have lower memory pages allocated as bounce
588 * buffers for doing I/O to pages residing above @page.
590 void blk_queue_bounce_limit(request_queue_t *q, u64 dma_addr)
592 unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
595 q->bounce_gfp = GFP_NOIO;
596 #if BITS_PER_LONG == 64
597 /* Assume anything <= 4GB can be handled by IOMMU.
598 Actually some IOMMUs can handle everything, but I don't
599 know of a way to test this here. */
600 if (bounce_pfn < (min_t(u64,0xffffffff,BLK_BOUNCE_HIGH) >> PAGE_SHIFT))
602 q->bounce_pfn = max_low_pfn;
604 if (bounce_pfn < blk_max_low_pfn)
606 q->bounce_pfn = bounce_pfn;
609 init_emergency_isa_pool();
610 q->bounce_gfp = GFP_NOIO | GFP_DMA;
611 q->bounce_pfn = bounce_pfn;
615 EXPORT_SYMBOL(blk_queue_bounce_limit);
618 * blk_queue_max_sectors - set max sectors for a request for this queue
619 * @q: the request queue for the device
620 * @max_sectors: max sectors in the usual 512b unit
623 * Enables a low level driver to set an upper limit on the size of
626 void blk_queue_max_sectors(request_queue_t *q, unsigned int max_sectors)
628 if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
629 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
630 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
633 if (BLK_DEF_MAX_SECTORS > max_sectors)
634 q->max_hw_sectors = q->max_sectors = max_sectors;
636 q->max_sectors = BLK_DEF_MAX_SECTORS;
637 q->max_hw_sectors = max_sectors;
641 EXPORT_SYMBOL(blk_queue_max_sectors);
644 * blk_queue_max_phys_segments - set max phys segments for a request for this queue
645 * @q: the request queue for the device
646 * @max_segments: max number of segments
649 * Enables a low level driver to set an upper limit on the number of
650 * physical data segments in a request. This would be the largest sized
651 * scatter list the driver could handle.
653 void blk_queue_max_phys_segments(request_queue_t *q, unsigned short max_segments)
657 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
660 q->max_phys_segments = max_segments;
663 EXPORT_SYMBOL(blk_queue_max_phys_segments);
666 * blk_queue_max_hw_segments - set max hw segments for a request for this queue
667 * @q: the request queue for the device
668 * @max_segments: max number of segments
671 * Enables a low level driver to set an upper limit on the number of
672 * hw data segments in a request. This would be the largest number of
673 * address/length pairs the host adapter can actually give as once
676 void blk_queue_max_hw_segments(request_queue_t *q, unsigned short max_segments)
680 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
683 q->max_hw_segments = max_segments;
686 EXPORT_SYMBOL(blk_queue_max_hw_segments);
689 * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
690 * @q: the request queue for the device
691 * @max_size: max size of segment in bytes
694 * Enables a low level driver to set an upper limit on the size of a
697 void blk_queue_max_segment_size(request_queue_t *q, unsigned int max_size)
699 if (max_size < PAGE_CACHE_SIZE) {
700 max_size = PAGE_CACHE_SIZE;
701 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
704 q->max_segment_size = max_size;
707 EXPORT_SYMBOL(blk_queue_max_segment_size);
710 * blk_queue_hardsect_size - set hardware sector size for the queue
711 * @q: the request queue for the device
712 * @size: the hardware sector size, in bytes
715 * This should typically be set to the lowest possible sector size
716 * that the hardware can operate on (possible without reverting to
717 * even internal read-modify-write operations). Usually the default
718 * of 512 covers most hardware.
720 void blk_queue_hardsect_size(request_queue_t *q, unsigned short size)
722 q->hardsect_size = size;
725 EXPORT_SYMBOL(blk_queue_hardsect_size);
728 * Returns the minimum that is _not_ zero, unless both are zero.
730 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
733 * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
734 * @t: the stacking driver (top)
735 * @b: the underlying device (bottom)
737 void blk_queue_stack_limits(request_queue_t *t, request_queue_t *b)
739 /* zero is "infinity" */
740 t->max_sectors = min_not_zero(t->max_sectors,b->max_sectors);
741 t->max_hw_sectors = min_not_zero(t->max_hw_sectors,b->max_hw_sectors);
743 t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
744 t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
745 t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
746 t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
747 if (!test_bit(QUEUE_FLAG_CLUSTER, &b->queue_flags))
748 clear_bit(QUEUE_FLAG_CLUSTER, &t->queue_flags);
751 EXPORT_SYMBOL(blk_queue_stack_limits);
754 * blk_queue_segment_boundary - set boundary rules for segment merging
755 * @q: the request queue for the device
756 * @mask: the memory boundary mask
758 void blk_queue_segment_boundary(request_queue_t *q, unsigned long mask)
760 if (mask < PAGE_CACHE_SIZE - 1) {
761 mask = PAGE_CACHE_SIZE - 1;
762 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
765 q->seg_boundary_mask = mask;
768 EXPORT_SYMBOL(blk_queue_segment_boundary);
771 * blk_queue_dma_alignment - set dma length and memory alignment
772 * @q: the request queue for the device
773 * @mask: alignment mask
776 * set required memory and length aligment for direct dma transactions.
777 * this is used when buiding direct io requests for the queue.
780 void blk_queue_dma_alignment(request_queue_t *q, int mask)
782 q->dma_alignment = mask;
785 EXPORT_SYMBOL(blk_queue_dma_alignment);
788 * blk_queue_find_tag - find a request by its tag and queue
789 * @q: The request queue for the device
790 * @tag: The tag of the request
793 * Should be used when a device returns a tag and you want to match
796 * no locks need be held.
798 struct request *blk_queue_find_tag(request_queue_t *q, int tag)
800 return blk_map_queue_find_tag(q->queue_tags, tag);
803 EXPORT_SYMBOL(blk_queue_find_tag);
806 * __blk_free_tags - release a given set of tag maintenance info
807 * @bqt: the tag map to free
809 * Tries to free the specified @bqt@. Returns true if it was
810 * actually freed and false if there are still references using it
812 static int __blk_free_tags(struct blk_queue_tag *bqt)
816 retval = atomic_dec_and_test(&bqt->refcnt);
819 BUG_ON(!list_empty(&bqt->busy_list));
821 kfree(bqt->tag_index);
822 bqt->tag_index = NULL;
835 * __blk_queue_free_tags - release tag maintenance info
836 * @q: the request queue for the device
839 * blk_cleanup_queue() will take care of calling this function, if tagging
840 * has been used. So there's no need to call this directly.
842 static void __blk_queue_free_tags(request_queue_t *q)
844 struct blk_queue_tag *bqt = q->queue_tags;
849 __blk_free_tags(bqt);
851 q->queue_tags = NULL;
852 q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
857 * blk_free_tags - release a given set of tag maintenance info
858 * @bqt: the tag map to free
860 * For externally managed @bqt@ frees the map. Callers of this
861 * function must guarantee to have released all the queues that
862 * might have been using this tag map.
864 void blk_free_tags(struct blk_queue_tag *bqt)
866 if (unlikely(!__blk_free_tags(bqt)))
869 EXPORT_SYMBOL(blk_free_tags);
872 * blk_queue_free_tags - release tag maintenance info
873 * @q: the request queue for the device
876 * This is used to disabled tagged queuing to a device, yet leave
879 void blk_queue_free_tags(request_queue_t *q)
881 clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
884 EXPORT_SYMBOL(blk_queue_free_tags);
887 init_tag_map(request_queue_t *q, struct blk_queue_tag *tags, int depth)
889 struct request **tag_index;
890 unsigned long *tag_map;
893 if (q && depth > q->nr_requests * 2) {
894 depth = q->nr_requests * 2;
895 printk(KERN_ERR "%s: adjusted depth to %d\n",
896 __FUNCTION__, depth);
899 tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC);
903 nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
904 tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
908 tags->real_max_depth = depth;
909 tags->max_depth = depth;
910 tags->tag_index = tag_index;
911 tags->tag_map = tag_map;
919 static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q,
922 struct blk_queue_tag *tags;
924 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
928 if (init_tag_map(q, tags, depth))
931 INIT_LIST_HEAD(&tags->busy_list);
933 atomic_set(&tags->refcnt, 1);
941 * blk_init_tags - initialize the tag info for an external tag map
942 * @depth: the maximum queue depth supported
943 * @tags: the tag to use
945 struct blk_queue_tag *blk_init_tags(int depth)
947 return __blk_queue_init_tags(NULL, depth);
949 EXPORT_SYMBOL(blk_init_tags);
952 * blk_queue_init_tags - initialize the queue tag info
953 * @q: the request queue for the device
954 * @depth: the maximum queue depth supported
955 * @tags: the tag to use
957 int blk_queue_init_tags(request_queue_t *q, int depth,
958 struct blk_queue_tag *tags)
962 BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
964 if (!tags && !q->queue_tags) {
965 tags = __blk_queue_init_tags(q, depth);
969 } else if (q->queue_tags) {
970 if ((rc = blk_queue_resize_tags(q, depth)))
972 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
975 atomic_inc(&tags->refcnt);
978 * assign it, all done
980 q->queue_tags = tags;
981 q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
988 EXPORT_SYMBOL(blk_queue_init_tags);
991 * blk_queue_resize_tags - change the queueing depth
992 * @q: the request queue for the device
993 * @new_depth: the new max command queueing depth
996 * Must be called with the queue lock held.
998 int blk_queue_resize_tags(request_queue_t *q, int new_depth)
1000 struct blk_queue_tag *bqt = q->queue_tags;
1001 struct request **tag_index;
1002 unsigned long *tag_map;
1003 int max_depth, nr_ulongs;
1009 * if we already have large enough real_max_depth. just
1010 * adjust max_depth. *NOTE* as requests with tag value
1011 * between new_depth and real_max_depth can be in-flight, tag
1012 * map can not be shrunk blindly here.
1014 if (new_depth <= bqt->real_max_depth) {
1015 bqt->max_depth = new_depth;
1020 * Currently cannot replace a shared tag map with a new
1021 * one, so error out if this is the case
1023 if (atomic_read(&bqt->refcnt) != 1)
1027 * save the old state info, so we can copy it back
1029 tag_index = bqt->tag_index;
1030 tag_map = bqt->tag_map;
1031 max_depth = bqt->real_max_depth;
1033 if (init_tag_map(q, bqt, new_depth))
1036 memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
1037 nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
1038 memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
1045 EXPORT_SYMBOL(blk_queue_resize_tags);
1048 * blk_queue_end_tag - end tag operations for a request
1049 * @q: the request queue for the device
1050 * @rq: the request that has completed
1053 * Typically called when end_that_request_first() returns 0, meaning
1054 * all transfers have been done for a request. It's important to call
1055 * this function before end_that_request_last(), as that will put the
1056 * request back on the free list thus corrupting the internal tag list.
1059 * queue lock must be held.
1061 void blk_queue_end_tag(request_queue_t *q, struct request *rq)
1063 struct blk_queue_tag *bqt = q->queue_tags;
1068 if (unlikely(tag >= bqt->real_max_depth))
1070 * This can happen after tag depth has been reduced.
1071 * FIXME: how about a warning or info message here?
1075 if (unlikely(!__test_and_clear_bit(tag, bqt->tag_map))) {
1076 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
1081 list_del_init(&rq->queuelist);
1082 rq->cmd_flags &= ~REQ_QUEUED;
1085 if (unlikely(bqt->tag_index[tag] == NULL))
1086 printk(KERN_ERR "%s: tag %d is missing\n",
1089 bqt->tag_index[tag] = NULL;
1093 EXPORT_SYMBOL(blk_queue_end_tag);
1096 * blk_queue_start_tag - find a free tag and assign it
1097 * @q: the request queue for the device
1098 * @rq: the block request that needs tagging
1101 * This can either be used as a stand-alone helper, or possibly be
1102 * assigned as the queue &prep_rq_fn (in which case &struct request
1103 * automagically gets a tag assigned). Note that this function
1104 * assumes that any type of request can be queued! if this is not
1105 * true for your device, you must check the request type before
1106 * calling this function. The request will also be removed from
1107 * the request queue, so it's the drivers responsibility to readd
1108 * it if it should need to be restarted for some reason.
1111 * queue lock must be held.
1113 int blk_queue_start_tag(request_queue_t *q, struct request *rq)
1115 struct blk_queue_tag *bqt = q->queue_tags;
1118 if (unlikely((rq->cmd_flags & REQ_QUEUED))) {
1120 "%s: request %p for device [%s] already tagged %d",
1122 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
1127 * Protect against shared tag maps, as we may not have exclusive
1128 * access to the tag map.
1131 tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
1132 if (tag >= bqt->max_depth)
1135 } while (test_and_set_bit(tag, bqt->tag_map));
1137 rq->cmd_flags |= REQ_QUEUED;
1139 bqt->tag_index[tag] = rq;
1140 blkdev_dequeue_request(rq);
1141 list_add(&rq->queuelist, &bqt->busy_list);
1146 EXPORT_SYMBOL(blk_queue_start_tag);
1149 * blk_queue_invalidate_tags - invalidate all pending tags
1150 * @q: the request queue for the device
1153 * Hardware conditions may dictate a need to stop all pending requests.
1154 * In this case, we will safely clear the block side of the tag queue and
1155 * readd all requests to the request queue in the right order.
1158 * queue lock must be held.
1160 void blk_queue_invalidate_tags(request_queue_t *q)
1162 struct blk_queue_tag *bqt = q->queue_tags;
1163 struct list_head *tmp, *n;
1166 list_for_each_safe(tmp, n, &bqt->busy_list) {
1167 rq = list_entry_rq(tmp);
1169 if (rq->tag == -1) {
1171 "%s: bad tag found on list\n", __FUNCTION__);
1172 list_del_init(&rq->queuelist);
1173 rq->cmd_flags &= ~REQ_QUEUED;
1175 blk_queue_end_tag(q, rq);
1177 rq->cmd_flags &= ~REQ_STARTED;
1178 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1182 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1184 void blk_dump_rq_flags(struct request *rq, char *msg)
1188 printk("%s: dev %s: type=%x, flags=%x\n", msg,
1189 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->cmd_type,
1192 printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1194 rq->current_nr_sectors);
1195 printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1197 if (blk_pc_request(rq)) {
1199 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1200 printk("%02x ", rq->cmd[bit]);
1205 EXPORT_SYMBOL(blk_dump_rq_flags);
1207 void blk_recount_segments(request_queue_t *q, struct bio *bio)
1209 struct bio_vec *bv, *bvprv = NULL;
1210 int i, nr_phys_segs, nr_hw_segs, seg_size, hw_seg_size, cluster;
1211 int high, highprv = 1;
1213 if (unlikely(!bio->bi_io_vec))
1216 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1217 hw_seg_size = seg_size = nr_phys_segs = nr_hw_segs = 0;
1218 bio_for_each_segment(bv, bio, i) {
1220 * the trick here is making sure that a high page is never
1221 * considered part of another segment, since that might
1222 * change with the bounce page.
1224 high = page_to_pfn(bv->bv_page) > q->bounce_pfn;
1225 if (high || highprv)
1226 goto new_hw_segment;
1228 if (seg_size + bv->bv_len > q->max_segment_size)
1230 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1232 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1234 if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1235 goto new_hw_segment;
1237 seg_size += bv->bv_len;
1238 hw_seg_size += bv->bv_len;
1243 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1244 !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len)) {
1245 hw_seg_size += bv->bv_len;
1248 if (hw_seg_size > bio->bi_hw_front_size)
1249 bio->bi_hw_front_size = hw_seg_size;
1250 hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1256 seg_size = bv->bv_len;
1259 if (hw_seg_size > bio->bi_hw_back_size)
1260 bio->bi_hw_back_size = hw_seg_size;
1261 if (nr_hw_segs == 1 && hw_seg_size > bio->bi_hw_front_size)
1262 bio->bi_hw_front_size = hw_seg_size;
1263 bio->bi_phys_segments = nr_phys_segs;
1264 bio->bi_hw_segments = nr_hw_segs;
1265 bio->bi_flags |= (1 << BIO_SEG_VALID);
1267 EXPORT_SYMBOL(blk_recount_segments);
1269 static int blk_phys_contig_segment(request_queue_t *q, struct bio *bio,
1272 if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1275 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1277 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1281 * bio and nxt are contigous in memory, check if the queue allows
1282 * these two to be merged into one
1284 if (BIO_SEG_BOUNDARY(q, bio, nxt))
1290 static int blk_hw_contig_segment(request_queue_t *q, struct bio *bio,
1293 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1294 blk_recount_segments(q, bio);
1295 if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1296 blk_recount_segments(q, nxt);
1297 if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1298 BIOVEC_VIRT_OVERSIZE(bio->bi_hw_front_size + bio->bi_hw_back_size))
1300 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1307 * map a request to scatterlist, return number of sg entries setup. Caller
1308 * must make sure sg can hold rq->nr_phys_segments entries
1310 int blk_rq_map_sg(request_queue_t *q, struct request *rq, struct scatterlist *sg)
1312 struct bio_vec *bvec, *bvprv;
1314 int nsegs, i, cluster;
1317 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1320 * for each bio in rq
1323 rq_for_each_bio(bio, rq) {
1325 * for each segment in bio
1327 bio_for_each_segment(bvec, bio, i) {
1328 int nbytes = bvec->bv_len;
1330 if (bvprv && cluster) {
1331 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1334 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1336 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1339 sg[nsegs - 1].length += nbytes;
1342 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1343 sg[nsegs].page = bvec->bv_page;
1344 sg[nsegs].length = nbytes;
1345 sg[nsegs].offset = bvec->bv_offset;
1350 } /* segments in bio */
1356 EXPORT_SYMBOL(blk_rq_map_sg);
1359 * the standard queue merge functions, can be overridden with device
1360 * specific ones if so desired
1363 static inline int ll_new_mergeable(request_queue_t *q,
1364 struct request *req,
1367 int nr_phys_segs = bio_phys_segments(q, bio);
1369 if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1370 req->cmd_flags |= REQ_NOMERGE;
1371 if (req == q->last_merge)
1372 q->last_merge = NULL;
1377 * A hw segment is just getting larger, bump just the phys
1380 req->nr_phys_segments += nr_phys_segs;
1384 static inline int ll_new_hw_segment(request_queue_t *q,
1385 struct request *req,
1388 int nr_hw_segs = bio_hw_segments(q, bio);
1389 int nr_phys_segs = bio_phys_segments(q, bio);
1391 if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1392 || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1393 req->cmd_flags |= REQ_NOMERGE;
1394 if (req == q->last_merge)
1395 q->last_merge = NULL;
1400 * This will form the start of a new hw segment. Bump both
1403 req->nr_hw_segments += nr_hw_segs;
1404 req->nr_phys_segments += nr_phys_segs;
1408 int ll_back_merge_fn(request_queue_t *q, struct request *req, struct bio *bio)
1410 unsigned short max_sectors;
1413 if (unlikely(blk_pc_request(req)))
1414 max_sectors = q->max_hw_sectors;
1416 max_sectors = q->max_sectors;
1418 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1419 req->cmd_flags |= REQ_NOMERGE;
1420 if (req == q->last_merge)
1421 q->last_merge = NULL;
1424 if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1425 blk_recount_segments(q, req->biotail);
1426 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1427 blk_recount_segments(q, bio);
1428 len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1429 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1430 !BIOVEC_VIRT_OVERSIZE(len)) {
1431 int mergeable = ll_new_mergeable(q, req, bio);
1434 if (req->nr_hw_segments == 1)
1435 req->bio->bi_hw_front_size = len;
1436 if (bio->bi_hw_segments == 1)
1437 bio->bi_hw_back_size = len;
1442 return ll_new_hw_segment(q, req, bio);
1444 EXPORT_SYMBOL(ll_back_merge_fn);
1446 static int ll_front_merge_fn(request_queue_t *q, struct request *req,
1449 unsigned short max_sectors;
1452 if (unlikely(blk_pc_request(req)))
1453 max_sectors = q->max_hw_sectors;
1455 max_sectors = q->max_sectors;
1458 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1459 req->cmd_flags |= REQ_NOMERGE;
1460 if (req == q->last_merge)
1461 q->last_merge = NULL;
1464 len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1465 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1466 blk_recount_segments(q, bio);
1467 if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1468 blk_recount_segments(q, req->bio);
1469 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1470 !BIOVEC_VIRT_OVERSIZE(len)) {
1471 int mergeable = ll_new_mergeable(q, req, bio);
1474 if (bio->bi_hw_segments == 1)
1475 bio->bi_hw_front_size = len;
1476 if (req->nr_hw_segments == 1)
1477 req->biotail->bi_hw_back_size = len;
1482 return ll_new_hw_segment(q, req, bio);
1485 static int ll_merge_requests_fn(request_queue_t *q, struct request *req,
1486 struct request *next)
1488 int total_phys_segments;
1489 int total_hw_segments;
1492 * First check if the either of the requests are re-queued
1493 * requests. Can't merge them if they are.
1495 if (req->special || next->special)
1499 * Will it become too large?
1501 if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1504 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1505 if (blk_phys_contig_segment(q, req->biotail, next->bio))
1506 total_phys_segments--;
1508 if (total_phys_segments > q->max_phys_segments)
1511 total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1512 if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1513 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1515 * propagate the combined length to the end of the requests
1517 if (req->nr_hw_segments == 1)
1518 req->bio->bi_hw_front_size = len;
1519 if (next->nr_hw_segments == 1)
1520 next->biotail->bi_hw_back_size = len;
1521 total_hw_segments--;
1524 if (total_hw_segments > q->max_hw_segments)
1527 /* Merge is OK... */
1528 req->nr_phys_segments = total_phys_segments;
1529 req->nr_hw_segments = total_hw_segments;
1534 * "plug" the device if there are no outstanding requests: this will
1535 * force the transfer to start only after we have put all the requests
1538 * This is called with interrupts off and no requests on the queue and
1539 * with the queue lock held.
1541 void blk_plug_device(request_queue_t *q)
1543 WARN_ON(!irqs_disabled());
1546 * don't plug a stopped queue, it must be paired with blk_start_queue()
1547 * which will restart the queueing
1549 if (blk_queue_stopped(q))
1552 if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
1553 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1554 blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
1558 EXPORT_SYMBOL(blk_plug_device);
1561 * remove the queue from the plugged list, if present. called with
1562 * queue lock held and interrupts disabled.
1564 int blk_remove_plug(request_queue_t *q)
1566 WARN_ON(!irqs_disabled());
1568 if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1571 del_timer(&q->unplug_timer);
1575 EXPORT_SYMBOL(blk_remove_plug);
1578 * remove the plug and let it rip..
1580 void __generic_unplug_device(request_queue_t *q)
1582 if (unlikely(blk_queue_stopped(q)))
1585 if (!blk_remove_plug(q))
1590 EXPORT_SYMBOL(__generic_unplug_device);
1593 * generic_unplug_device - fire a request queue
1594 * @q: The &request_queue_t in question
1597 * Linux uses plugging to build bigger requests queues before letting
1598 * the device have at them. If a queue is plugged, the I/O scheduler
1599 * is still adding and merging requests on the queue. Once the queue
1600 * gets unplugged, the request_fn defined for the queue is invoked and
1601 * transfers started.
1603 void generic_unplug_device(request_queue_t *q)
1605 spin_lock_irq(q->queue_lock);
1606 __generic_unplug_device(q);
1607 spin_unlock_irq(q->queue_lock);
1609 EXPORT_SYMBOL(generic_unplug_device);
1611 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1614 request_queue_t *q = bdi->unplug_io_data;
1617 * devices don't necessarily have an ->unplug_fn defined
1620 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1621 q->rq.count[READ] + q->rq.count[WRITE]);
1627 static void blk_unplug_work(struct work_struct *work)
1629 request_queue_t *q = container_of(work, request_queue_t, unplug_work);
1631 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1632 q->rq.count[READ] + q->rq.count[WRITE]);
1637 static void blk_unplug_timeout(unsigned long data)
1639 request_queue_t *q = (request_queue_t *)data;
1641 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
1642 q->rq.count[READ] + q->rq.count[WRITE]);
1644 kblockd_schedule_work(&q->unplug_work);
1648 * blk_start_queue - restart a previously stopped queue
1649 * @q: The &request_queue_t in question
1652 * blk_start_queue() will clear the stop flag on the queue, and call
1653 * the request_fn for the queue if it was in a stopped state when
1654 * entered. Also see blk_stop_queue(). Queue lock must be held.
1656 void blk_start_queue(request_queue_t *q)
1658 WARN_ON(!irqs_disabled());
1660 clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1663 * one level of recursion is ok and is much faster than kicking
1664 * the unplug handling
1666 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1668 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1671 kblockd_schedule_work(&q->unplug_work);
1675 EXPORT_SYMBOL(blk_start_queue);
1678 * blk_stop_queue - stop a queue
1679 * @q: The &request_queue_t in question
1682 * The Linux block layer assumes that a block driver will consume all
1683 * entries on the request queue when the request_fn strategy is called.
1684 * Often this will not happen, because of hardware limitations (queue
1685 * depth settings). If a device driver gets a 'queue full' response,
1686 * or if it simply chooses not to queue more I/O at one point, it can
1687 * call this function to prevent the request_fn from being called until
1688 * the driver has signalled it's ready to go again. This happens by calling
1689 * blk_start_queue() to restart queue operations. Queue lock must be held.
1691 void blk_stop_queue(request_queue_t *q)
1694 set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1696 EXPORT_SYMBOL(blk_stop_queue);
1699 * blk_sync_queue - cancel any pending callbacks on a queue
1703 * The block layer may perform asynchronous callback activity
1704 * on a queue, such as calling the unplug function after a timeout.
1705 * A block device may call blk_sync_queue to ensure that any
1706 * such activity is cancelled, thus allowing it to release resources
1707 * that the callbacks might use. The caller must already have made sure
1708 * that its ->make_request_fn will not re-add plugging prior to calling
1712 void blk_sync_queue(struct request_queue *q)
1714 del_timer_sync(&q->unplug_timer);
1716 EXPORT_SYMBOL(blk_sync_queue);
1719 * blk_run_queue - run a single device queue
1720 * @q: The queue to run
1722 void blk_run_queue(struct request_queue *q)
1724 unsigned long flags;
1726 spin_lock_irqsave(q->queue_lock, flags);
1730 * Only recurse once to avoid overrunning the stack, let the unplug
1731 * handling reinvoke the handler shortly if we already got there.
1733 if (!elv_queue_empty(q)) {
1734 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1736 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1739 kblockd_schedule_work(&q->unplug_work);
1743 spin_unlock_irqrestore(q->queue_lock, flags);
1745 EXPORT_SYMBOL(blk_run_queue);
1748 * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1749 * @kobj: the kobj belonging of the request queue to be released
1752 * blk_cleanup_queue is the pair to blk_init_queue() or
1753 * blk_queue_make_request(). It should be called when a request queue is
1754 * being released; typically when a block device is being de-registered.
1755 * Currently, its primary task it to free all the &struct request
1756 * structures that were allocated to the queue and the queue itself.
1759 * Hopefully the low level driver will have finished any
1760 * outstanding requests first...
1762 static void blk_release_queue(struct kobject *kobj)
1764 request_queue_t *q = container_of(kobj, struct request_queue, kobj);
1765 struct request_list *rl = &q->rq;
1770 mempool_destroy(rl->rq_pool);
1773 __blk_queue_free_tags(q);
1775 blk_trace_shutdown(q);
1777 kmem_cache_free(requestq_cachep, q);
1780 void blk_put_queue(request_queue_t *q)
1782 kobject_put(&q->kobj);
1784 EXPORT_SYMBOL(blk_put_queue);
1786 void blk_cleanup_queue(request_queue_t * q)
1788 mutex_lock(&q->sysfs_lock);
1789 set_bit(QUEUE_FLAG_DEAD, &q->queue_flags);
1790 mutex_unlock(&q->sysfs_lock);
1793 elevator_exit(q->elevator);
1798 EXPORT_SYMBOL(blk_cleanup_queue);
1800 static int blk_init_free_list(request_queue_t *q)
1802 struct request_list *rl = &q->rq;
1804 rl->count[READ] = rl->count[WRITE] = 0;
1805 rl->starved[READ] = rl->starved[WRITE] = 0;
1807 init_waitqueue_head(&rl->wait[READ]);
1808 init_waitqueue_head(&rl->wait[WRITE]);
1810 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1811 mempool_free_slab, request_cachep, q->node);
1819 request_queue_t *blk_alloc_queue(gfp_t gfp_mask)
1821 return blk_alloc_queue_node(gfp_mask, -1);
1823 EXPORT_SYMBOL(blk_alloc_queue);
1825 static struct kobj_type queue_ktype;
1827 request_queue_t *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
1831 q = kmem_cache_alloc_node(requestq_cachep, gfp_mask, node_id);
1835 memset(q, 0, sizeof(*q));
1836 init_timer(&q->unplug_timer);
1838 snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
1839 q->kobj.ktype = &queue_ktype;
1840 kobject_init(&q->kobj);
1842 q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1843 q->backing_dev_info.unplug_io_data = q;
1845 mutex_init(&q->sysfs_lock);
1849 EXPORT_SYMBOL(blk_alloc_queue_node);
1852 * blk_init_queue - prepare a request queue for use with a block device
1853 * @rfn: The function to be called to process requests that have been
1854 * placed on the queue.
1855 * @lock: Request queue spin lock
1858 * If a block device wishes to use the standard request handling procedures,
1859 * which sorts requests and coalesces adjacent requests, then it must
1860 * call blk_init_queue(). The function @rfn will be called when there
1861 * are requests on the queue that need to be processed. If the device
1862 * supports plugging, then @rfn may not be called immediately when requests
1863 * are available on the queue, but may be called at some time later instead.
1864 * Plugged queues are generally unplugged when a buffer belonging to one
1865 * of the requests on the queue is needed, or due to memory pressure.
1867 * @rfn is not required, or even expected, to remove all requests off the
1868 * queue, but only as many as it can handle at a time. If it does leave
1869 * requests on the queue, it is responsible for arranging that the requests
1870 * get dealt with eventually.
1872 * The queue spin lock must be held while manipulating the requests on the
1873 * request queue; this lock will be taken also from interrupt context, so irq
1874 * disabling is needed for it.
1876 * Function returns a pointer to the initialized request queue, or NULL if
1877 * it didn't succeed.
1880 * blk_init_queue() must be paired with a blk_cleanup_queue() call
1881 * when the block device is deactivated (such as at module unload).
1884 request_queue_t *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1886 return blk_init_queue_node(rfn, lock, -1);
1888 EXPORT_SYMBOL(blk_init_queue);
1891 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1893 request_queue_t *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1899 if (blk_init_free_list(q)) {
1900 kmem_cache_free(requestq_cachep, q);
1905 * if caller didn't supply a lock, they get per-queue locking with
1909 spin_lock_init(&q->__queue_lock);
1910 lock = &q->__queue_lock;
1913 q->request_fn = rfn;
1914 q->prep_rq_fn = NULL;
1915 q->unplug_fn = generic_unplug_device;
1916 q->queue_flags = (1 << QUEUE_FLAG_CLUSTER);
1917 q->queue_lock = lock;
1919 blk_queue_segment_boundary(q, 0xffffffff);
1921 blk_queue_make_request(q, __make_request);
1922 blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1924 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1925 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1927 q->sg_reserved_size = INT_MAX;
1932 if (!elevator_init(q, NULL)) {
1933 blk_queue_congestion_threshold(q);
1940 EXPORT_SYMBOL(blk_init_queue_node);
1942 int blk_get_queue(request_queue_t *q)
1944 if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1945 kobject_get(&q->kobj);
1952 EXPORT_SYMBOL(blk_get_queue);
1954 static inline void blk_free_request(request_queue_t *q, struct request *rq)
1956 if (rq->cmd_flags & REQ_ELVPRIV)
1957 elv_put_request(q, rq);
1958 mempool_free(rq, q->rq.rq_pool);
1961 static struct request *
1962 blk_alloc_request(request_queue_t *q, int rw, int priv, gfp_t gfp_mask)
1964 struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1970 * first three bits are identical in rq->cmd_flags and bio->bi_rw,
1971 * see bio.h and blkdev.h
1973 rq->cmd_flags = rw | REQ_ALLOCED;
1976 if (unlikely(elv_set_request(q, rq, gfp_mask))) {
1977 mempool_free(rq, q->rq.rq_pool);
1980 rq->cmd_flags |= REQ_ELVPRIV;
1987 * ioc_batching returns true if the ioc is a valid batching request and
1988 * should be given priority access to a request.
1990 static inline int ioc_batching(request_queue_t *q, struct io_context *ioc)
1996 * Make sure the process is able to allocate at least 1 request
1997 * even if the batch times out, otherwise we could theoretically
2000 return ioc->nr_batch_requests == q->nr_batching ||
2001 (ioc->nr_batch_requests > 0
2002 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
2006 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
2007 * will cause the process to be a "batcher" on all queues in the system. This
2008 * is the behaviour we want though - once it gets a wakeup it should be given
2011 static void ioc_set_batching(request_queue_t *q, struct io_context *ioc)
2013 if (!ioc || ioc_batching(q, ioc))
2016 ioc->nr_batch_requests = q->nr_batching;
2017 ioc->last_waited = jiffies;
2020 static void __freed_request(request_queue_t *q, int rw)
2022 struct request_list *rl = &q->rq;
2024 if (rl->count[rw] < queue_congestion_off_threshold(q))
2025 blk_clear_queue_congested(q, rw);
2027 if (rl->count[rw] + 1 <= q->nr_requests) {
2028 if (waitqueue_active(&rl->wait[rw]))
2029 wake_up(&rl->wait[rw]);
2031 blk_clear_queue_full(q, rw);
2036 * A request has just been released. Account for it, update the full and
2037 * congestion status, wake up any waiters. Called under q->queue_lock.
2039 static void freed_request(request_queue_t *q, int rw, int priv)
2041 struct request_list *rl = &q->rq;
2047 __freed_request(q, rw);
2049 if (unlikely(rl->starved[rw ^ 1]))
2050 __freed_request(q, rw ^ 1);
2053 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2055 * Get a free request, queue_lock must be held.
2056 * Returns NULL on failure, with queue_lock held.
2057 * Returns !NULL on success, with queue_lock *not held*.
2059 static struct request *get_request(request_queue_t *q, int rw_flags,
2060 struct bio *bio, gfp_t gfp_mask)
2062 struct request *rq = NULL;
2063 struct request_list *rl = &q->rq;
2064 struct io_context *ioc = NULL;
2065 const int rw = rw_flags & 0x01;
2066 int may_queue, priv;
2068 may_queue = elv_may_queue(q, rw_flags);
2069 if (may_queue == ELV_MQUEUE_NO)
2072 if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
2073 if (rl->count[rw]+1 >= q->nr_requests) {
2074 ioc = current_io_context(GFP_ATOMIC, q->node);
2076 * The queue will fill after this allocation, so set
2077 * it as full, and mark this process as "batching".
2078 * This process will be allowed to complete a batch of
2079 * requests, others will be blocked.
2081 if (!blk_queue_full(q, rw)) {
2082 ioc_set_batching(q, ioc);
2083 blk_set_queue_full(q, rw);
2085 if (may_queue != ELV_MQUEUE_MUST
2086 && !ioc_batching(q, ioc)) {
2088 * The queue is full and the allocating
2089 * process is not a "batcher", and not
2090 * exempted by the IO scheduler
2096 blk_set_queue_congested(q, rw);
2100 * Only allow batching queuers to allocate up to 50% over the defined
2101 * limit of requests, otherwise we could have thousands of requests
2102 * allocated with any setting of ->nr_requests
2104 if (rl->count[rw] >= (3 * q->nr_requests / 2))
2108 rl->starved[rw] = 0;
2110 priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
2114 spin_unlock_irq(q->queue_lock);
2116 rq = blk_alloc_request(q, rw_flags, priv, gfp_mask);
2117 if (unlikely(!rq)) {
2119 * Allocation failed presumably due to memory. Undo anything
2120 * we might have messed up.
2122 * Allocating task should really be put onto the front of the
2123 * wait queue, but this is pretty rare.
2125 spin_lock_irq(q->queue_lock);
2126 freed_request(q, rw, priv);
2129 * in the very unlikely event that allocation failed and no
2130 * requests for this direction was pending, mark us starved
2131 * so that freeing of a request in the other direction will
2132 * notice us. another possible fix would be to split the
2133 * rq mempool into READ and WRITE
2136 if (unlikely(rl->count[rw] == 0))
2137 rl->starved[rw] = 1;
2143 * ioc may be NULL here, and ioc_batching will be false. That's
2144 * OK, if the queue is under the request limit then requests need
2145 * not count toward the nr_batch_requests limit. There will always
2146 * be some limit enforced by BLK_BATCH_TIME.
2148 if (ioc_batching(q, ioc))
2149 ioc->nr_batch_requests--;
2153 blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
2159 * No available requests for this queue, unplug the device and wait for some
2160 * requests to become available.
2162 * Called with q->queue_lock held, and returns with it unlocked.
2164 static struct request *get_request_wait(request_queue_t *q, int rw_flags,
2167 const int rw = rw_flags & 0x01;
2170 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2173 struct request_list *rl = &q->rq;
2175 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2176 TASK_UNINTERRUPTIBLE);
2178 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2181 struct io_context *ioc;
2183 blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
2185 __generic_unplug_device(q);
2186 spin_unlock_irq(q->queue_lock);
2190 * After sleeping, we become a "batching" process and
2191 * will be able to allocate at least one request, and
2192 * up to a big batch of them for a small period time.
2193 * See ioc_batching, ioc_set_batching
2195 ioc = current_io_context(GFP_NOIO, q->node);
2196 ioc_set_batching(q, ioc);
2198 spin_lock_irq(q->queue_lock);
2200 finish_wait(&rl->wait[rw], &wait);
2206 struct request *blk_get_request(request_queue_t *q, int rw, gfp_t gfp_mask)
2210 BUG_ON(rw != READ && rw != WRITE);
2212 spin_lock_irq(q->queue_lock);
2213 if (gfp_mask & __GFP_WAIT) {
2214 rq = get_request_wait(q, rw, NULL);
2216 rq = get_request(q, rw, NULL, gfp_mask);
2218 spin_unlock_irq(q->queue_lock);
2220 /* q->queue_lock is unlocked at this point */
2224 EXPORT_SYMBOL(blk_get_request);
2227 * blk_start_queueing - initiate dispatch of requests to device
2228 * @q: request queue to kick into gear
2230 * This is basically a helper to remove the need to know whether a queue
2231 * is plugged or not if someone just wants to initiate dispatch of requests
2234 * The queue lock must be held with interrupts disabled.
2236 void blk_start_queueing(request_queue_t *q)
2238 if (!blk_queue_plugged(q))
2241 __generic_unplug_device(q);
2243 EXPORT_SYMBOL(blk_start_queueing);
2246 * blk_requeue_request - put a request back on queue
2247 * @q: request queue where request should be inserted
2248 * @rq: request to be inserted
2251 * Drivers often keep queueing requests until the hardware cannot accept
2252 * more, when that condition happens we need to put the request back
2253 * on the queue. Must be called with queue lock held.
2255 void blk_requeue_request(request_queue_t *q, struct request *rq)
2257 blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
2259 if (blk_rq_tagged(rq))
2260 blk_queue_end_tag(q, rq);
2262 elv_requeue_request(q, rq);
2265 EXPORT_SYMBOL(blk_requeue_request);
2268 * blk_insert_request - insert a special request in to a request queue
2269 * @q: request queue where request should be inserted
2270 * @rq: request to be inserted
2271 * @at_head: insert request at head or tail of queue
2272 * @data: private data
2275 * Many block devices need to execute commands asynchronously, so they don't
2276 * block the whole kernel from preemption during request execution. This is
2277 * accomplished normally by inserting aritficial requests tagged as
2278 * REQ_SPECIAL in to the corresponding request queue, and letting them be
2279 * scheduled for actual execution by the request queue.
2281 * We have the option of inserting the head or the tail of the queue.
2282 * Typically we use the tail for new ioctls and so forth. We use the head
2283 * of the queue for things like a QUEUE_FULL message from a device, or a
2284 * host that is unable to accept a particular command.
2286 void blk_insert_request(request_queue_t *q, struct request *rq,
2287 int at_head, void *data)
2289 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2290 unsigned long flags;
2293 * tell I/O scheduler that this isn't a regular read/write (ie it
2294 * must not attempt merges on this) and that it acts as a soft
2297 rq->cmd_type = REQ_TYPE_SPECIAL;
2298 rq->cmd_flags |= REQ_SOFTBARRIER;
2302 spin_lock_irqsave(q->queue_lock, flags);
2305 * If command is tagged, release the tag
2307 if (blk_rq_tagged(rq))
2308 blk_queue_end_tag(q, rq);
2310 drive_stat_acct(rq, rq->nr_sectors, 1);
2311 __elv_add_request(q, rq, where, 0);
2312 blk_start_queueing(q);
2313 spin_unlock_irqrestore(q->queue_lock, flags);
2316 EXPORT_SYMBOL(blk_insert_request);
2318 static int __blk_rq_unmap_user(struct bio *bio)
2323 if (bio_flagged(bio, BIO_USER_MAPPED))
2324 bio_unmap_user(bio);
2326 ret = bio_uncopy_user(bio);
2332 static int __blk_rq_map_user(request_queue_t *q, struct request *rq,
2333 void __user *ubuf, unsigned int len)
2335 unsigned long uaddr;
2336 struct bio *bio, *orig_bio;
2339 reading = rq_data_dir(rq) == READ;
2342 * if alignment requirement is satisfied, map in user pages for
2343 * direct dma. else, set up kernel bounce buffers
2345 uaddr = (unsigned long) ubuf;
2346 if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2347 bio = bio_map_user(q, NULL, uaddr, len, reading);
2349 bio = bio_copy_user(q, uaddr, len, reading);
2352 return PTR_ERR(bio);
2355 blk_queue_bounce(q, &bio);
2358 * We link the bounce buffer in and could have to traverse it
2359 * later so we have to get a ref to prevent it from being freed
2364 blk_rq_bio_prep(q, rq, bio);
2365 else if (!ll_back_merge_fn(q, rq, bio)) {
2369 rq->biotail->bi_next = bio;
2372 rq->data_len += bio->bi_size;
2375 return bio->bi_size;
2378 /* if it was boucned we must call the end io function */
2379 bio_endio(bio, bio->bi_size, 0);
2380 __blk_rq_unmap_user(orig_bio);
2386 * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2387 * @q: request queue where request should be inserted
2388 * @rq: request structure to fill
2389 * @ubuf: the user buffer
2390 * @len: length of user data
2393 * Data will be mapped directly for zero copy io, if possible. Otherwise
2394 * a kernel bounce buffer is used.
2396 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2397 * still in process context.
2399 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2400 * before being submitted to the device, as pages mapped may be out of
2401 * reach. It's the callers responsibility to make sure this happens. The
2402 * original bio must be passed back in to blk_rq_unmap_user() for proper
2405 int blk_rq_map_user(request_queue_t *q, struct request *rq, void __user *ubuf,
2408 unsigned long bytes_read = 0;
2409 struct bio *bio = NULL;
2412 if (len > (q->max_hw_sectors << 9))
2417 while (bytes_read != len) {
2418 unsigned long map_len, end, start;
2420 map_len = min_t(unsigned long, len - bytes_read, BIO_MAX_SIZE);
2421 end = ((unsigned long)ubuf + map_len + PAGE_SIZE - 1)
2423 start = (unsigned long)ubuf >> PAGE_SHIFT;
2426 * A bad offset could cause us to require BIO_MAX_PAGES + 1
2427 * pages. If this happens we just lower the requested
2428 * mapping len by a page so that we can fit
2430 if (end - start > BIO_MAX_PAGES)
2431 map_len -= PAGE_SIZE;
2433 ret = __blk_rq_map_user(q, rq, ubuf, map_len);
2442 rq->buffer = rq->data = NULL;
2445 blk_rq_unmap_user(bio);
2449 EXPORT_SYMBOL(blk_rq_map_user);
2452 * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2453 * @q: request queue where request should be inserted
2454 * @rq: request to map data to
2455 * @iov: pointer to the iovec
2456 * @iov_count: number of elements in the iovec
2457 * @len: I/O byte count
2460 * Data will be mapped directly for zero copy io, if possible. Otherwise
2461 * a kernel bounce buffer is used.
2463 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2464 * still in process context.
2466 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2467 * before being submitted to the device, as pages mapped may be out of
2468 * reach. It's the callers responsibility to make sure this happens. The
2469 * original bio must be passed back in to blk_rq_unmap_user() for proper
2472 int blk_rq_map_user_iov(request_queue_t *q, struct request *rq,
2473 struct sg_iovec *iov, int iov_count, unsigned int len)
2477 if (!iov || iov_count <= 0)
2480 /* we don't allow misaligned data like bio_map_user() does. If the
2481 * user is using sg, they're expected to know the alignment constraints
2482 * and respect them accordingly */
2483 bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2485 return PTR_ERR(bio);
2487 if (bio->bi_size != len) {
2488 bio_endio(bio, bio->bi_size, 0);
2489 bio_unmap_user(bio);
2494 blk_rq_bio_prep(q, rq, bio);
2495 rq->buffer = rq->data = NULL;
2499 EXPORT_SYMBOL(blk_rq_map_user_iov);
2502 * blk_rq_unmap_user - unmap a request with user data
2503 * @bio: start of bio list
2506 * Unmap a rq previously mapped by blk_rq_map_user(). The caller must
2507 * supply the original rq->bio from the blk_rq_map_user() return, since
2508 * the io completion may have changed rq->bio.
2510 int blk_rq_unmap_user(struct bio *bio)
2512 struct bio *mapped_bio;
2517 if (unlikely(bio_flagged(bio, BIO_BOUNCED)))
2518 mapped_bio = bio->bi_private;
2520 ret2 = __blk_rq_unmap_user(mapped_bio);
2526 bio_put(mapped_bio);
2532 EXPORT_SYMBOL(blk_rq_unmap_user);
2535 * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2536 * @q: request queue where request should be inserted
2537 * @rq: request to fill
2538 * @kbuf: the kernel buffer
2539 * @len: length of user data
2540 * @gfp_mask: memory allocation flags
2542 int blk_rq_map_kern(request_queue_t *q, struct request *rq, void *kbuf,
2543 unsigned int len, gfp_t gfp_mask)
2547 if (len > (q->max_hw_sectors << 9))
2552 bio = bio_map_kern(q, kbuf, len, gfp_mask);
2554 return PTR_ERR(bio);
2556 if (rq_data_dir(rq) == WRITE)
2557 bio->bi_rw |= (1 << BIO_RW);
2559 blk_rq_bio_prep(q, rq, bio);
2560 blk_queue_bounce(q, &rq->bio);
2561 rq->buffer = rq->data = NULL;
2565 EXPORT_SYMBOL(blk_rq_map_kern);
2568 * blk_execute_rq_nowait - insert a request into queue for execution
2569 * @q: queue to insert the request in
2570 * @bd_disk: matching gendisk
2571 * @rq: request to insert
2572 * @at_head: insert request at head or tail of queue
2573 * @done: I/O completion handler
2576 * Insert a fully prepared request at the back of the io scheduler queue
2577 * for execution. Don't wait for completion.
2579 void blk_execute_rq_nowait(request_queue_t *q, struct gendisk *bd_disk,
2580 struct request *rq, int at_head,
2583 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2585 rq->rq_disk = bd_disk;
2586 rq->cmd_flags |= REQ_NOMERGE;
2588 WARN_ON(irqs_disabled());
2589 spin_lock_irq(q->queue_lock);
2590 __elv_add_request(q, rq, where, 1);
2591 __generic_unplug_device(q);
2592 spin_unlock_irq(q->queue_lock);
2594 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
2597 * blk_execute_rq - insert a request into queue for execution
2598 * @q: queue to insert the request in
2599 * @bd_disk: matching gendisk
2600 * @rq: request to insert
2601 * @at_head: insert request at head or tail of queue
2604 * Insert a fully prepared request at the back of the io scheduler queue
2605 * for execution and wait for completion.
2607 int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
2608 struct request *rq, int at_head)
2610 DECLARE_COMPLETION_ONSTACK(wait);
2611 char sense[SCSI_SENSE_BUFFERSIZE];
2615 * we need an extra reference to the request, so we can look at
2616 * it after io completion
2621 memset(sense, 0, sizeof(sense));
2626 rq->end_io_data = &wait;
2627 blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
2628 wait_for_completion(&wait);
2636 EXPORT_SYMBOL(blk_execute_rq);
2639 * blkdev_issue_flush - queue a flush
2640 * @bdev: blockdev to issue flush for
2641 * @error_sector: error sector
2644 * Issue a flush for the block device in question. Caller can supply
2645 * room for storing the error offset in case of a flush error, if they
2646 * wish to. Caller must run wait_for_completion() on its own.
2648 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2652 if (bdev->bd_disk == NULL)
2655 q = bdev_get_queue(bdev);
2658 if (!q->issue_flush_fn)
2661 return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2664 EXPORT_SYMBOL(blkdev_issue_flush);
2666 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2668 int rw = rq_data_dir(rq);
2670 if (!blk_fs_request(rq) || !rq->rq_disk)
2674 __disk_stat_inc(rq->rq_disk, merges[rw]);
2676 disk_round_stats(rq->rq_disk);
2677 rq->rq_disk->in_flight++;
2682 * add-request adds a request to the linked list.
2683 * queue lock is held and interrupts disabled, as we muck with the
2684 * request queue list.
2686 static inline void add_request(request_queue_t * q, struct request * req)
2688 drive_stat_acct(req, req->nr_sectors, 1);
2691 * elevator indicated where it wants this request to be
2692 * inserted at elevator_merge time
2694 __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2698 * disk_round_stats() - Round off the performance stats on a struct
2701 * The average IO queue length and utilisation statistics are maintained
2702 * by observing the current state of the queue length and the amount of
2703 * time it has been in this state for.
2705 * Normally, that accounting is done on IO completion, but that can result
2706 * in more than a second's worth of IO being accounted for within any one
2707 * second, leading to >100% utilisation. To deal with that, we call this
2708 * function to do a round-off before returning the results when reading
2709 * /proc/diskstats. This accounts immediately for all queue usage up to
2710 * the current jiffies and restarts the counters again.
2712 void disk_round_stats(struct gendisk *disk)
2714 unsigned long now = jiffies;
2716 if (now == disk->stamp)
2719 if (disk->in_flight) {
2720 __disk_stat_add(disk, time_in_queue,
2721 disk->in_flight * (now - disk->stamp));
2722 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2727 EXPORT_SYMBOL_GPL(disk_round_stats);
2730 * queue lock must be held
2732 void __blk_put_request(request_queue_t *q, struct request *req)
2736 if (unlikely(--req->ref_count))
2739 elv_completed_request(q, req);
2742 * Request may not have originated from ll_rw_blk. if not,
2743 * it didn't come out of our reserved rq pools
2745 if (req->cmd_flags & REQ_ALLOCED) {
2746 int rw = rq_data_dir(req);
2747 int priv = req->cmd_flags & REQ_ELVPRIV;
2749 BUG_ON(!list_empty(&req->queuelist));
2750 BUG_ON(!hlist_unhashed(&req->hash));
2752 blk_free_request(q, req);
2753 freed_request(q, rw, priv);
2757 EXPORT_SYMBOL_GPL(__blk_put_request);
2759 void blk_put_request(struct request *req)
2761 unsigned long flags;
2762 request_queue_t *q = req->q;
2765 * Gee, IDE calls in w/ NULL q. Fix IDE and remove the
2766 * following if (q) test.
2769 spin_lock_irqsave(q->queue_lock, flags);
2770 __blk_put_request(q, req);
2771 spin_unlock_irqrestore(q->queue_lock, flags);
2775 EXPORT_SYMBOL(blk_put_request);
2778 * blk_end_sync_rq - executes a completion event on a request
2779 * @rq: request to complete
2780 * @error: end io status of the request
2782 void blk_end_sync_rq(struct request *rq, int error)
2784 struct completion *waiting = rq->end_io_data;
2786 rq->end_io_data = NULL;
2787 __blk_put_request(rq->q, rq);
2790 * complete last, if this is a stack request the process (and thus
2791 * the rq pointer) could be invalid right after this complete()
2795 EXPORT_SYMBOL(blk_end_sync_rq);
2798 * Has to be called with the request spinlock acquired
2800 static int attempt_merge(request_queue_t *q, struct request *req,
2801 struct request *next)
2803 if (!rq_mergeable(req) || !rq_mergeable(next))
2809 if (req->sector + req->nr_sectors != next->sector)
2812 if (rq_data_dir(req) != rq_data_dir(next)
2813 || req->rq_disk != next->rq_disk
2818 * If we are allowed to merge, then append bio list
2819 * from next to rq and release next. merge_requests_fn
2820 * will have updated segment counts, update sector
2823 if (!ll_merge_requests_fn(q, req, next))
2827 * At this point we have either done a back merge
2828 * or front merge. We need the smaller start_time of
2829 * the merged requests to be the current request
2830 * for accounting purposes.
2832 if (time_after(req->start_time, next->start_time))
2833 req->start_time = next->start_time;
2835 req->biotail->bi_next = next->bio;
2836 req->biotail = next->biotail;
2838 req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2840 elv_merge_requests(q, req, next);
2843 disk_round_stats(req->rq_disk);
2844 req->rq_disk->in_flight--;
2847 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2849 __blk_put_request(q, next);
2853 static inline int attempt_back_merge(request_queue_t *q, struct request *rq)
2855 struct request *next = elv_latter_request(q, rq);
2858 return attempt_merge(q, rq, next);
2863 static inline int attempt_front_merge(request_queue_t *q, struct request *rq)
2865 struct request *prev = elv_former_request(q, rq);
2868 return attempt_merge(q, prev, rq);
2873 static void init_request_from_bio(struct request *req, struct bio *bio)
2875 req->cmd_type = REQ_TYPE_FS;
2878 * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2880 if (bio_rw_ahead(bio) || bio_failfast(bio))
2881 req->cmd_flags |= REQ_FAILFAST;
2884 * REQ_BARRIER implies no merging, but lets make it explicit
2886 if (unlikely(bio_barrier(bio)))
2887 req->cmd_flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2890 req->cmd_flags |= REQ_RW_SYNC;
2891 if (bio_rw_meta(bio))
2892 req->cmd_flags |= REQ_RW_META;
2895 req->hard_sector = req->sector = bio->bi_sector;
2896 req->hard_nr_sectors = req->nr_sectors = bio_sectors(bio);
2897 req->current_nr_sectors = req->hard_cur_sectors = bio_cur_sectors(bio);
2898 req->nr_phys_segments = bio_phys_segments(req->q, bio);
2899 req->nr_hw_segments = bio_hw_segments(req->q, bio);
2900 req->buffer = bio_data(bio); /* see ->buffer comment above */
2901 req->bio = req->biotail = bio;
2902 req->ioprio = bio_prio(bio);
2903 req->rq_disk = bio->bi_bdev->bd_disk;
2904 req->start_time = jiffies;
2907 static int __make_request(request_queue_t *q, struct bio *bio)
2909 struct request *req;
2910 int el_ret, nr_sectors, barrier, err;
2911 const unsigned short prio = bio_prio(bio);
2912 const int sync = bio_sync(bio);
2915 nr_sectors = bio_sectors(bio);
2918 * low level driver can indicate that it wants pages above a
2919 * certain limit bounced to low memory (ie for highmem, or even
2920 * ISA dma in theory)
2922 blk_queue_bounce(q, &bio);
2924 barrier = bio_barrier(bio);
2925 if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
2930 spin_lock_irq(q->queue_lock);
2932 if (unlikely(barrier) || elv_queue_empty(q))
2935 el_ret = elv_merge(q, &req, bio);
2937 case ELEVATOR_BACK_MERGE:
2938 BUG_ON(!rq_mergeable(req));
2940 if (!ll_back_merge_fn(q, req, bio))
2943 blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
2945 req->biotail->bi_next = bio;
2947 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2948 req->ioprio = ioprio_best(req->ioprio, prio);
2949 drive_stat_acct(req, nr_sectors, 0);
2950 if (!attempt_back_merge(q, req))
2951 elv_merged_request(q, req, el_ret);
2954 case ELEVATOR_FRONT_MERGE:
2955 BUG_ON(!rq_mergeable(req));
2957 if (!ll_front_merge_fn(q, req, bio))
2960 blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
2962 bio->bi_next = req->bio;
2966 * may not be valid. if the low level driver said
2967 * it didn't need a bounce buffer then it better
2968 * not touch req->buffer either...
2970 req->buffer = bio_data(bio);
2971 req->current_nr_sectors = bio_cur_sectors(bio);
2972 req->hard_cur_sectors = req->current_nr_sectors;
2973 req->sector = req->hard_sector = bio->bi_sector;
2974 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2975 req->ioprio = ioprio_best(req->ioprio, prio);
2976 drive_stat_acct(req, nr_sectors, 0);
2977 if (!attempt_front_merge(q, req))
2978 elv_merged_request(q, req, el_ret);
2981 /* ELV_NO_MERGE: elevator says don't/can't merge. */
2988 * This sync check and mask will be re-done in init_request_from_bio(),
2989 * but we need to set it earlier to expose the sync flag to the
2990 * rq allocator and io schedulers.
2992 rw_flags = bio_data_dir(bio);
2994 rw_flags |= REQ_RW_SYNC;
2997 * Grab a free request. This is might sleep but can not fail.
2998 * Returns with the queue unlocked.
3000 req = get_request_wait(q, rw_flags, bio);
3003 * After dropping the lock and possibly sleeping here, our request
3004 * may now be mergeable after it had proven unmergeable (above).
3005 * We don't worry about that case for efficiency. It won't happen
3006 * often, and the elevators are able to handle it.
3008 init_request_from_bio(req, bio);
3010 spin_lock_irq(q->queue_lock);
3011 if (elv_queue_empty(q))
3013 add_request(q, req);
3016 __generic_unplug_device(q);
3018 spin_unlock_irq(q->queue_lock);
3022 bio_endio(bio, nr_sectors << 9, err);
3027 * If bio->bi_dev is a partition, remap the location
3029 static inline void blk_partition_remap(struct bio *bio)
3031 struct block_device *bdev = bio->bi_bdev;
3033 if (bdev != bdev->bd_contains) {
3034 struct hd_struct *p = bdev->bd_part;
3035 const int rw = bio_data_dir(bio);
3037 p->sectors[rw] += bio_sectors(bio);
3040 bio->bi_sector += p->start_sect;
3041 bio->bi_bdev = bdev->bd_contains;
3045 static void handle_bad_sector(struct bio *bio)
3047 char b[BDEVNAME_SIZE];
3049 printk(KERN_INFO "attempt to access beyond end of device\n");
3050 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
3051 bdevname(bio->bi_bdev, b),
3053 (unsigned long long)bio->bi_sector + bio_sectors(bio),
3054 (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
3056 set_bit(BIO_EOF, &bio->bi_flags);
3059 #ifdef CONFIG_FAIL_MAKE_REQUEST
3061 static DECLARE_FAULT_ATTR(fail_make_request);
3063 static int __init setup_fail_make_request(char *str)
3065 return setup_fault_attr(&fail_make_request, str);
3067 __setup("fail_make_request=", setup_fail_make_request);
3069 static int should_fail_request(struct bio *bio)
3071 if ((bio->bi_bdev->bd_disk->flags & GENHD_FL_FAIL) ||
3072 (bio->bi_bdev->bd_part && bio->bi_bdev->bd_part->make_it_fail))
3073 return should_fail(&fail_make_request, bio->bi_size);
3078 static int __init fail_make_request_debugfs(void)
3080 return init_fault_attr_dentries(&fail_make_request,
3081 "fail_make_request");
3084 late_initcall(fail_make_request_debugfs);
3086 #else /* CONFIG_FAIL_MAKE_REQUEST */
3088 static inline int should_fail_request(struct bio *bio)
3093 #endif /* CONFIG_FAIL_MAKE_REQUEST */
3096 * generic_make_request: hand a buffer to its device driver for I/O
3097 * @bio: The bio describing the location in memory and on the device.
3099 * generic_make_request() is used to make I/O requests of block
3100 * devices. It is passed a &struct bio, which describes the I/O that needs
3103 * generic_make_request() does not return any status. The
3104 * success/failure status of the request, along with notification of
3105 * completion, is delivered asynchronously through the bio->bi_end_io
3106 * function described (one day) else where.
3108 * The caller of generic_make_request must make sure that bi_io_vec
3109 * are set to describe the memory buffer, and that bi_dev and bi_sector are
3110 * set to describe the device address, and the
3111 * bi_end_io and optionally bi_private are set to describe how
3112 * completion notification should be signaled.
3114 * generic_make_request and the drivers it calls may use bi_next if this
3115 * bio happens to be merged with someone else, and may change bi_dev and
3116 * bi_sector for remaps as it sees fit. So the values of these fields
3117 * should NOT be depended on after the call to generic_make_request.
3119 static inline void __generic_make_request(struct bio *bio)
3123 sector_t old_sector;
3124 int ret, nr_sectors = bio_sectors(bio);
3128 /* Test device or partition size, when known. */
3129 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3131 sector_t sector = bio->bi_sector;
3133 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
3135 * This may well happen - the kernel calls bread()
3136 * without checking the size of the device, e.g., when
3137 * mounting a device.
3139 handle_bad_sector(bio);
3145 * Resolve the mapping until finished. (drivers are
3146 * still free to implement/resolve their own stacking
3147 * by explicitly returning 0)
3149 * NOTE: we don't repeat the blk_size check for each new device.
3150 * Stacking drivers are expected to know what they are doing.
3155 char b[BDEVNAME_SIZE];
3157 q = bdev_get_queue(bio->bi_bdev);
3160 "generic_make_request: Trying to access "
3161 "nonexistent block-device %s (%Lu)\n",
3162 bdevname(bio->bi_bdev, b),
3163 (long long) bio->bi_sector);
3165 bio_endio(bio, bio->bi_size, -EIO);
3169 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
3170 printk("bio too big device %s (%u > %u)\n",
3171 bdevname(bio->bi_bdev, b),
3177 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
3180 if (should_fail_request(bio))
3184 * If this device has partitions, remap block n
3185 * of partition p to block n+start(p) of the disk.
3187 blk_partition_remap(bio);
3189 if (old_sector != -1)
3190 blk_add_trace_remap(q, bio, old_dev, bio->bi_sector,
3193 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
3195 old_sector = bio->bi_sector;
3196 old_dev = bio->bi_bdev->bd_dev;
3198 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3200 sector_t sector = bio->bi_sector;
3202 if (maxsector < nr_sectors ||
3203 maxsector - nr_sectors < sector) {
3205 * This may well happen - partitions are not
3206 * checked to make sure they are within the size
3207 * of the whole device.
3209 handle_bad_sector(bio);
3214 ret = q->make_request_fn(q, bio);
3219 * We only want one ->make_request_fn to be active at a time,
3220 * else stack usage with stacked devices could be a problem.
3221 * So use current->bio_{list,tail} to keep a list of requests
3222 * submited by a make_request_fn function.
3223 * current->bio_tail is also used as a flag to say if
3224 * generic_make_request is currently active in this task or not.
3225 * If it is NULL, then no make_request is active. If it is non-NULL,
3226 * then a make_request is active, and new requests should be added
3229 void generic_make_request(struct bio *bio)
3231 if (current->bio_tail) {
3232 /* make_request is active */
3233 *(current->bio_tail) = bio;
3234 bio->bi_next = NULL;
3235 current->bio_tail = &bio->bi_next;
3238 /* following loop may be a bit non-obvious, and so deserves some
3240 * Before entering the loop, bio->bi_next is NULL (as all callers
3241 * ensure that) so we have a list with a single bio.
3242 * We pretend that we have just taken it off a longer list, so
3243 * we assign bio_list to the next (which is NULL) and bio_tail
3244 * to &bio_list, thus initialising the bio_list of new bios to be
3245 * added. __generic_make_request may indeed add some more bios
3246 * through a recursive call to generic_make_request. If it
3247 * did, we find a non-NULL value in bio_list and re-enter the loop
3248 * from the top. In this case we really did just take the bio
3249 * of the top of the list (no pretending) and so fixup bio_list and
3250 * bio_tail or bi_next, and call into __generic_make_request again.
3252 * The loop was structured like this to make only one call to
3253 * __generic_make_request (which is important as it is large and
3254 * inlined) and to keep the structure simple.
3256 BUG_ON(bio->bi_next);
3258 current->bio_list = bio->bi_next;
3259 if (bio->bi_next == NULL)
3260 current->bio_tail = ¤t->bio_list;
3262 bio->bi_next = NULL;
3263 __generic_make_request(bio);
3264 bio = current->bio_list;
3266 current->bio_tail = NULL; /* deactivate */
3269 EXPORT_SYMBOL(generic_make_request);
3272 * submit_bio: submit a bio to the block device layer for I/O
3273 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3274 * @bio: The &struct bio which describes the I/O
3276 * submit_bio() is very similar in purpose to generic_make_request(), and
3277 * uses that function to do most of the work. Both are fairly rough
3278 * interfaces, @bio must be presetup and ready for I/O.
3281 void submit_bio(int rw, struct bio *bio)
3283 int count = bio_sectors(bio);
3285 BIO_BUG_ON(!bio->bi_size);
3286 BIO_BUG_ON(!bio->bi_io_vec);
3289 count_vm_events(PGPGOUT, count);
3291 task_io_account_read(bio->bi_size);
3292 count_vm_events(PGPGIN, count);
3295 if (unlikely(block_dump)) {
3296 char b[BDEVNAME_SIZE];
3297 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
3298 current->comm, current->pid,
3299 (rw & WRITE) ? "WRITE" : "READ",
3300 (unsigned long long)bio->bi_sector,
3301 bdevname(bio->bi_bdev,b));
3304 generic_make_request(bio);
3307 EXPORT_SYMBOL(submit_bio);
3309 static void blk_recalc_rq_segments(struct request *rq)
3311 struct bio *bio, *prevbio = NULL;
3312 int nr_phys_segs, nr_hw_segs;
3313 unsigned int phys_size, hw_size;
3314 request_queue_t *q = rq->q;
3319 phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
3320 rq_for_each_bio(bio, rq) {
3321 /* Force bio hw/phys segs to be recalculated. */
3322 bio->bi_flags &= ~(1 << BIO_SEG_VALID);
3324 nr_phys_segs += bio_phys_segments(q, bio);
3325 nr_hw_segs += bio_hw_segments(q, bio);
3327 int pseg = phys_size + prevbio->bi_size + bio->bi_size;
3328 int hseg = hw_size + prevbio->bi_size + bio->bi_size;
3330 if (blk_phys_contig_segment(q, prevbio, bio) &&
3331 pseg <= q->max_segment_size) {
3333 phys_size += prevbio->bi_size + bio->bi_size;
3337 if (blk_hw_contig_segment(q, prevbio, bio) &&
3338 hseg <= q->max_segment_size) {
3340 hw_size += prevbio->bi_size + bio->bi_size;
3347 rq->nr_phys_segments = nr_phys_segs;
3348 rq->nr_hw_segments = nr_hw_segs;
3351 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
3353 if (blk_fs_request(rq)) {
3354 rq->hard_sector += nsect;
3355 rq->hard_nr_sectors -= nsect;
3358 * Move the I/O submission pointers ahead if required.
3360 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3361 (rq->sector <= rq->hard_sector)) {
3362 rq->sector = rq->hard_sector;
3363 rq->nr_sectors = rq->hard_nr_sectors;
3364 rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3365 rq->current_nr_sectors = rq->hard_cur_sectors;
3366 rq->buffer = bio_data(rq->bio);
3370 * if total number of sectors is less than the first segment
3371 * size, something has gone terribly wrong
3373 if (rq->nr_sectors < rq->current_nr_sectors) {
3374 printk("blk: request botched\n");
3375 rq->nr_sectors = rq->current_nr_sectors;
3380 static int __end_that_request_first(struct request *req, int uptodate,
3383 int total_bytes, bio_nbytes, error, next_idx = 0;
3386 blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
3389 * extend uptodate bool to allow < 0 value to be direct io error
3392 if (end_io_error(uptodate))
3393 error = !uptodate ? -EIO : uptodate;
3396 * for a REQ_BLOCK_PC request, we want to carry any eventual
3397 * sense key with us all the way through
3399 if (!blk_pc_request(req))
3403 if (blk_fs_request(req) && !(req->cmd_flags & REQ_QUIET))
3404 printk("end_request: I/O error, dev %s, sector %llu\n",
3405 req->rq_disk ? req->rq_disk->disk_name : "?",
3406 (unsigned long long)req->sector);
3409 if (blk_fs_request(req) && req->rq_disk) {
3410 const int rw = rq_data_dir(req);
3412 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
3415 total_bytes = bio_nbytes = 0;
3416 while ((bio = req->bio) != NULL) {
3419 if (nr_bytes >= bio->bi_size) {
3420 req->bio = bio->bi_next;
3421 nbytes = bio->bi_size;
3422 if (!ordered_bio_endio(req, bio, nbytes, error))
3423 bio_endio(bio, nbytes, error);
3427 int idx = bio->bi_idx + next_idx;
3429 if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3430 blk_dump_rq_flags(req, "__end_that");
3431 printk("%s: bio idx %d >= vcnt %d\n",
3433 bio->bi_idx, bio->bi_vcnt);
3437 nbytes = bio_iovec_idx(bio, idx)->bv_len;
3438 BIO_BUG_ON(nbytes > bio->bi_size);
3441 * not a complete bvec done
3443 if (unlikely(nbytes > nr_bytes)) {
3444 bio_nbytes += nr_bytes;
3445 total_bytes += nr_bytes;
3450 * advance to the next vector
3453 bio_nbytes += nbytes;
3456 total_bytes += nbytes;
3459 if ((bio = req->bio)) {
3461 * end more in this run, or just return 'not-done'
3463 if (unlikely(nr_bytes <= 0))
3475 * if the request wasn't completed, update state
3478 if (!ordered_bio_endio(req, bio, bio_nbytes, error))
3479 bio_endio(bio, bio_nbytes, error);
3480 bio->bi_idx += next_idx;
3481 bio_iovec(bio)->bv_offset += nr_bytes;
3482 bio_iovec(bio)->bv_len -= nr_bytes;
3485 blk_recalc_rq_sectors(req, total_bytes >> 9);
3486 blk_recalc_rq_segments(req);
3491 * end_that_request_first - end I/O on a request
3492 * @req: the request being processed
3493 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3494 * @nr_sectors: number of sectors to end I/O on
3497 * Ends I/O on a number of sectors attached to @req, and sets it up
3498 * for the next range of segments (if any) in the cluster.
3501 * 0 - we are done with this request, call end_that_request_last()
3502 * 1 - still buffers pending for this request
3504 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3506 return __end_that_request_first(req, uptodate, nr_sectors << 9);
3509 EXPORT_SYMBOL(end_that_request_first);
3512 * end_that_request_chunk - end I/O on a request
3513 * @req: the request being processed
3514 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3515 * @nr_bytes: number of bytes to complete
3518 * Ends I/O on a number of bytes attached to @req, and sets it up
3519 * for the next range of segments (if any). Like end_that_request_first(),
3520 * but deals with bytes instead of sectors.
3523 * 0 - we are done with this request, call end_that_request_last()
3524 * 1 - still buffers pending for this request
3526 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3528 return __end_that_request_first(req, uptodate, nr_bytes);
3531 EXPORT_SYMBOL(end_that_request_chunk);
3534 * splice the completion data to a local structure and hand off to
3535 * process_completion_queue() to complete the requests
3537 static void blk_done_softirq(struct softirq_action *h)
3539 struct list_head *cpu_list, local_list;
3541 local_irq_disable();
3542 cpu_list = &__get_cpu_var(blk_cpu_done);
3543 list_replace_init(cpu_list, &local_list);
3546 while (!list_empty(&local_list)) {
3547 struct request *rq = list_entry(local_list.next, struct request, donelist);
3549 list_del_init(&rq->donelist);
3550 rq->q->softirq_done_fn(rq);
3554 static int blk_cpu_notify(struct notifier_block *self, unsigned long action,
3558 * If a CPU goes away, splice its entries to the current CPU
3559 * and trigger a run of the softirq
3561 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
3562 int cpu = (unsigned long) hcpu;
3564 local_irq_disable();
3565 list_splice_init(&per_cpu(blk_cpu_done, cpu),
3566 &__get_cpu_var(blk_cpu_done));
3567 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3575 static struct notifier_block __devinitdata blk_cpu_notifier = {
3576 .notifier_call = blk_cpu_notify,
3580 * blk_complete_request - end I/O on a request
3581 * @req: the request being processed
3584 * Ends all I/O on a request. It does not handle partial completions,
3585 * unless the driver actually implements this in its completion callback
3586 * through requeueing. Theh actual completion happens out-of-order,
3587 * through a softirq handler. The user must have registered a completion
3588 * callback through blk_queue_softirq_done().
3591 void blk_complete_request(struct request *req)
3593 struct list_head *cpu_list;
3594 unsigned long flags;
3596 BUG_ON(!req->q->softirq_done_fn);
3598 local_irq_save(flags);
3600 cpu_list = &__get_cpu_var(blk_cpu_done);
3601 list_add_tail(&req->donelist, cpu_list);
3602 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3604 local_irq_restore(flags);
3607 EXPORT_SYMBOL(blk_complete_request);
3610 * queue lock must be held
3612 void end_that_request_last(struct request *req, int uptodate)
3614 struct gendisk *disk = req->rq_disk;
3618 * extend uptodate bool to allow < 0 value to be direct io error
3621 if (end_io_error(uptodate))
3622 error = !uptodate ? -EIO : uptodate;
3624 if (unlikely(laptop_mode) && blk_fs_request(req))
3625 laptop_io_completion();
3628 * Account IO completion. bar_rq isn't accounted as a normal
3629 * IO on queueing nor completion. Accounting the containing
3630 * request is enough.
3632 if (disk && blk_fs_request(req) && req != &req->q->bar_rq) {
3633 unsigned long duration = jiffies - req->start_time;
3634 const int rw = rq_data_dir(req);
3636 __disk_stat_inc(disk, ios[rw]);
3637 __disk_stat_add(disk, ticks[rw], duration);
3638 disk_round_stats(disk);
3642 req->end_io(req, error);
3644 __blk_put_request(req->q, req);
3647 EXPORT_SYMBOL(end_that_request_last);
3649 void end_request(struct request *req, int uptodate)
3651 if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3652 add_disk_randomness(req->rq_disk);
3653 blkdev_dequeue_request(req);
3654 end_that_request_last(req, uptodate);
3658 EXPORT_SYMBOL(end_request);
3660 void blk_rq_bio_prep(request_queue_t *q, struct request *rq, struct bio *bio)
3662 /* first two bits are identical in rq->cmd_flags and bio->bi_rw */
3663 rq->cmd_flags |= (bio->bi_rw & 3);
3665 rq->nr_phys_segments = bio_phys_segments(q, bio);
3666 rq->nr_hw_segments = bio_hw_segments(q, bio);
3667 rq->current_nr_sectors = bio_cur_sectors(bio);
3668 rq->hard_cur_sectors = rq->current_nr_sectors;
3669 rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3670 rq->buffer = bio_data(bio);
3671 rq->data_len = bio->bi_size;
3673 rq->bio = rq->biotail = bio;
3676 EXPORT_SYMBOL(blk_rq_bio_prep);
3678 int kblockd_schedule_work(struct work_struct *work)
3680 return queue_work(kblockd_workqueue, work);
3683 EXPORT_SYMBOL(kblockd_schedule_work);
3685 void kblockd_flush_work(struct work_struct *work)
3687 cancel_work_sync(work);
3689 EXPORT_SYMBOL(kblockd_flush_work);
3691 int __init blk_dev_init(void)
3695 kblockd_workqueue = create_workqueue("kblockd");
3696 if (!kblockd_workqueue)
3697 panic("Failed to create kblockd\n");
3699 request_cachep = kmem_cache_create("blkdev_requests",
3700 sizeof(struct request), 0, SLAB_PANIC, NULL, NULL);
3702 requestq_cachep = kmem_cache_create("blkdev_queue",
3703 sizeof(request_queue_t), 0, SLAB_PANIC, NULL, NULL);
3705 iocontext_cachep = kmem_cache_create("blkdev_ioc",
3706 sizeof(struct io_context), 0, SLAB_PANIC, NULL, NULL);
3708 for_each_possible_cpu(i)
3709 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3711 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
3712 register_hotcpu_notifier(&blk_cpu_notifier);
3714 blk_max_low_pfn = max_low_pfn - 1;
3715 blk_max_pfn = max_pfn - 1;
3721 * IO Context helper functions
3723 void put_io_context(struct io_context *ioc)
3728 BUG_ON(atomic_read(&ioc->refcount) == 0);
3730 if (atomic_dec_and_test(&ioc->refcount)) {
3731 struct cfq_io_context *cic;
3734 if (ioc->aic && ioc->aic->dtor)
3735 ioc->aic->dtor(ioc->aic);
3736 if (ioc->cic_root.rb_node != NULL) {
3737 struct rb_node *n = rb_first(&ioc->cic_root);
3739 cic = rb_entry(n, struct cfq_io_context, rb_node);
3744 kmem_cache_free(iocontext_cachep, ioc);
3747 EXPORT_SYMBOL(put_io_context);
3749 /* Called by the exitting task */
3750 void exit_io_context(void)
3752 struct io_context *ioc;
3753 struct cfq_io_context *cic;
3756 ioc = current->io_context;
3757 current->io_context = NULL;
3758 task_unlock(current);
3761 if (ioc->aic && ioc->aic->exit)
3762 ioc->aic->exit(ioc->aic);
3763 if (ioc->cic_root.rb_node != NULL) {
3764 cic = rb_entry(rb_first(&ioc->cic_root), struct cfq_io_context, rb_node);
3768 put_io_context(ioc);
3772 * If the current task has no IO context then create one and initialise it.
3773 * Otherwise, return its existing IO context.
3775 * This returned IO context doesn't have a specifically elevated refcount,
3776 * but since the current task itself holds a reference, the context can be
3777 * used in general code, so long as it stays within `current` context.
3779 static struct io_context *current_io_context(gfp_t gfp_flags, int node)
3781 struct task_struct *tsk = current;
3782 struct io_context *ret;
3784 ret = tsk->io_context;
3788 ret = kmem_cache_alloc_node(iocontext_cachep, gfp_flags, node);
3790 atomic_set(&ret->refcount, 1);
3791 ret->task = current;
3792 ret->ioprio_changed = 0;
3793 ret->last_waited = jiffies; /* doesn't matter... */
3794 ret->nr_batch_requests = 0; /* because this is 0 */
3796 ret->cic_root.rb_node = NULL;
3797 ret->ioc_data = NULL;
3798 /* make sure set_task_ioprio() sees the settings above */
3800 tsk->io_context = ret;
3805 EXPORT_SYMBOL(current_io_context);
3808 * If the current task has no IO context then create one and initialise it.
3809 * If it does have a context, take a ref on it.
3811 * This is always called in the context of the task which submitted the I/O.
3813 struct io_context *get_io_context(gfp_t gfp_flags, int node)
3815 struct io_context *ret;
3816 ret = current_io_context(gfp_flags, node);
3818 atomic_inc(&ret->refcount);
3821 EXPORT_SYMBOL(get_io_context);
3823 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3825 struct io_context *src = *psrc;
3826 struct io_context *dst = *pdst;
3829 BUG_ON(atomic_read(&src->refcount) == 0);
3830 atomic_inc(&src->refcount);
3831 put_io_context(dst);
3835 EXPORT_SYMBOL(copy_io_context);
3837 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3839 struct io_context *temp;
3844 EXPORT_SYMBOL(swap_io_context);
3849 struct queue_sysfs_entry {
3850 struct attribute attr;
3851 ssize_t (*show)(struct request_queue *, char *);
3852 ssize_t (*store)(struct request_queue *, const char *, size_t);
3856 queue_var_show(unsigned int var, char *page)
3858 return sprintf(page, "%d\n", var);
3862 queue_var_store(unsigned long *var, const char *page, size_t count)
3864 char *p = (char *) page;
3866 *var = simple_strtoul(p, &p, 10);
3870 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3872 return queue_var_show(q->nr_requests, (page));
3876 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3878 struct request_list *rl = &q->rq;
3880 int ret = queue_var_store(&nr, page, count);
3881 if (nr < BLKDEV_MIN_RQ)
3884 spin_lock_irq(q->queue_lock);
3885 q->nr_requests = nr;
3886 blk_queue_congestion_threshold(q);
3888 if (rl->count[READ] >= queue_congestion_on_threshold(q))
3889 blk_set_queue_congested(q, READ);
3890 else if (rl->count[READ] < queue_congestion_off_threshold(q))
3891 blk_clear_queue_congested(q, READ);
3893 if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3894 blk_set_queue_congested(q, WRITE);
3895 else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3896 blk_clear_queue_congested(q, WRITE);
3898 if (rl->count[READ] >= q->nr_requests) {
3899 blk_set_queue_full(q, READ);
3900 } else if (rl->count[READ]+1 <= q->nr_requests) {
3901 blk_clear_queue_full(q, READ);
3902 wake_up(&rl->wait[READ]);
3905 if (rl->count[WRITE] >= q->nr_requests) {
3906 blk_set_queue_full(q, WRITE);
3907 } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3908 blk_clear_queue_full(q, WRITE);
3909 wake_up(&rl->wait[WRITE]);
3911 spin_unlock_irq(q->queue_lock);
3915 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3917 int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3919 return queue_var_show(ra_kb, (page));
3923 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3925 unsigned long ra_kb;
3926 ssize_t ret = queue_var_store(&ra_kb, page, count);
3928 spin_lock_irq(q->queue_lock);
3929 q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3930 spin_unlock_irq(q->queue_lock);
3935 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3937 int max_sectors_kb = q->max_sectors >> 1;
3939 return queue_var_show(max_sectors_kb, (page));
3943 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3945 unsigned long max_sectors_kb,
3946 max_hw_sectors_kb = q->max_hw_sectors >> 1,
3947 page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3948 ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3951 if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3954 * Take the queue lock to update the readahead and max_sectors
3955 * values synchronously:
3957 spin_lock_irq(q->queue_lock);
3959 * Trim readahead window as well, if necessary:
3961 ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3962 if (ra_kb > max_sectors_kb)
3963 q->backing_dev_info.ra_pages =
3964 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3966 q->max_sectors = max_sectors_kb << 1;
3967 spin_unlock_irq(q->queue_lock);
3972 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3974 int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3976 return queue_var_show(max_hw_sectors_kb, (page));
3980 static struct queue_sysfs_entry queue_requests_entry = {
3981 .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3982 .show = queue_requests_show,
3983 .store = queue_requests_store,
3986 static struct queue_sysfs_entry queue_ra_entry = {
3987 .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3988 .show = queue_ra_show,
3989 .store = queue_ra_store,
3992 static struct queue_sysfs_entry queue_max_sectors_entry = {
3993 .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3994 .show = queue_max_sectors_show,
3995 .store = queue_max_sectors_store,
3998 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3999 .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
4000 .show = queue_max_hw_sectors_show,
4003 static struct queue_sysfs_entry queue_iosched_entry = {
4004 .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
4005 .show = elv_iosched_show,
4006 .store = elv_iosched_store,
4009 static struct attribute *default_attrs[] = {
4010 &queue_requests_entry.attr,
4011 &queue_ra_entry.attr,
4012 &queue_max_hw_sectors_entry.attr,
4013 &queue_max_sectors_entry.attr,
4014 &queue_iosched_entry.attr,
4018 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
4021 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
4023 struct queue_sysfs_entry *entry = to_queue(attr);
4024 request_queue_t *q = container_of(kobj, struct request_queue, kobj);
4029 mutex_lock(&q->sysfs_lock);
4030 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
4031 mutex_unlock(&q->sysfs_lock);
4034 res = entry->show(q, page);
4035 mutex_unlock(&q->sysfs_lock);
4040 queue_attr_store(struct kobject *kobj, struct attribute *attr,
4041 const char *page, size_t length)
4043 struct queue_sysfs_entry *entry = to_queue(attr);
4044 request_queue_t *q = container_of(kobj, struct request_queue, kobj);
4050 mutex_lock(&q->sysfs_lock);
4051 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
4052 mutex_unlock(&q->sysfs_lock);
4055 res = entry->store(q, page, length);
4056 mutex_unlock(&q->sysfs_lock);
4060 static struct sysfs_ops queue_sysfs_ops = {
4061 .show = queue_attr_show,
4062 .store = queue_attr_store,
4065 static struct kobj_type queue_ktype = {
4066 .sysfs_ops = &queue_sysfs_ops,
4067 .default_attrs = default_attrs,
4068 .release = blk_release_queue,
4071 int blk_register_queue(struct gendisk *disk)
4075 request_queue_t *q = disk->queue;
4077 if (!q || !q->request_fn)
4080 q->kobj.parent = kobject_get(&disk->kobj);
4082 ret = kobject_add(&q->kobj);
4086 kobject_uevent(&q->kobj, KOBJ_ADD);
4088 ret = elv_register_queue(q);
4090 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4091 kobject_del(&q->kobj);
4098 void blk_unregister_queue(struct gendisk *disk)
4100 request_queue_t *q = disk->queue;
4102 if (q && q->request_fn) {
4103 elv_unregister_queue(q);
4105 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4106 kobject_del(&q->kobj);
4107 kobject_put(&disk->kobj);