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/interrupt.h>
29 #include <linux/cpu.h>
30 #include <linux/blktrace_api.h>
35 #include <scsi/scsi_cmnd.h>
37 static void blk_unplug_work(void *data);
38 static void blk_unplug_timeout(unsigned long data);
39 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io);
40 static void init_request_from_bio(struct request *req, struct bio *bio);
41 static int __make_request(request_queue_t *q, struct bio *bio);
44 * For the allocated request tables
46 static kmem_cache_t *request_cachep;
49 * For queue allocation
51 static kmem_cache_t *requestq_cachep;
54 * For io context allocations
56 static kmem_cache_t *iocontext_cachep;
58 static wait_queue_head_t congestion_wqh[2] = {
59 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
60 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
64 * Controlling structure to kblockd
66 static struct workqueue_struct *kblockd_workqueue;
68 unsigned long blk_max_low_pfn, blk_max_pfn;
70 EXPORT_SYMBOL(blk_max_low_pfn);
71 EXPORT_SYMBOL(blk_max_pfn);
73 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
75 /* Amount of time in which a process may batch requests */
76 #define BLK_BATCH_TIME (HZ/50UL)
78 /* Number of requests a "batching" process may submit */
79 #define BLK_BATCH_REQ 32
82 * Return the threshold (number of used requests) at which the queue is
83 * considered to be congested. It include a little hysteresis to keep the
84 * context switch rate down.
86 static inline int queue_congestion_on_threshold(struct request_queue *q)
88 return q->nr_congestion_on;
92 * The threshold at which a queue is considered to be uncongested
94 static inline int queue_congestion_off_threshold(struct request_queue *q)
96 return q->nr_congestion_off;
99 static void blk_queue_congestion_threshold(struct request_queue *q)
103 nr = q->nr_requests - (q->nr_requests / 8) + 1;
104 if (nr > q->nr_requests)
106 q->nr_congestion_on = nr;
108 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
111 q->nr_congestion_off = nr;
115 * A queue has just exitted congestion. Note this in the global counter of
116 * congested queues, and wake up anyone who was waiting for requests to be
119 static void clear_queue_congested(request_queue_t *q, int rw)
122 wait_queue_head_t *wqh = &congestion_wqh[rw];
124 bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
125 clear_bit(bit, &q->backing_dev_info.state);
126 smp_mb__after_clear_bit();
127 if (waitqueue_active(wqh))
132 * A queue has just entered congestion. Flag that in the queue's VM-visible
133 * state flags and increment the global gounter of congested queues.
135 static void set_queue_congested(request_queue_t *q, int rw)
139 bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
140 set_bit(bit, &q->backing_dev_info.state);
144 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
147 * Locates the passed device's request queue and returns the address of its
150 * Will return NULL if the request queue cannot be located.
152 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
154 struct backing_dev_info *ret = NULL;
155 request_queue_t *q = bdev_get_queue(bdev);
158 ret = &q->backing_dev_info;
162 EXPORT_SYMBOL(blk_get_backing_dev_info);
164 void blk_queue_activity_fn(request_queue_t *q, activity_fn *fn, void *data)
167 q->activity_data = data;
170 EXPORT_SYMBOL(blk_queue_activity_fn);
173 * blk_queue_prep_rq - set a prepare_request function for queue
175 * @pfn: prepare_request function
177 * It's possible for a queue to register a prepare_request callback which
178 * is invoked before the request is handed to the request_fn. The goal of
179 * the function is to prepare a request for I/O, it can be used to build a
180 * cdb from the request data for instance.
183 void blk_queue_prep_rq(request_queue_t *q, prep_rq_fn *pfn)
188 EXPORT_SYMBOL(blk_queue_prep_rq);
191 * blk_queue_merge_bvec - set a merge_bvec function for queue
193 * @mbfn: merge_bvec_fn
195 * Usually queues have static limitations on the max sectors or segments that
196 * we can put in a request. Stacking drivers may have some settings that
197 * are dynamic, and thus we have to query the queue whether it is ok to
198 * add a new bio_vec to a bio at a given offset or not. If the block device
199 * has such limitations, it needs to register a merge_bvec_fn to control
200 * the size of bio's sent to it. Note that a block device *must* allow a
201 * single page to be added to an empty bio. The block device driver may want
202 * to use the bio_split() function to deal with these bio's. By default
203 * no merge_bvec_fn is defined for a queue, and only the fixed limits are
206 void blk_queue_merge_bvec(request_queue_t *q, merge_bvec_fn *mbfn)
208 q->merge_bvec_fn = mbfn;
211 EXPORT_SYMBOL(blk_queue_merge_bvec);
213 void blk_queue_softirq_done(request_queue_t *q, softirq_done_fn *fn)
215 q->softirq_done_fn = fn;
218 EXPORT_SYMBOL(blk_queue_softirq_done);
221 * blk_queue_make_request - define an alternate make_request function for a device
222 * @q: the request queue for the device to be affected
223 * @mfn: the alternate make_request function
226 * The normal way for &struct bios to be passed to a device
227 * driver is for them to be collected into requests on a request
228 * queue, and then to allow the device driver to select requests
229 * off that queue when it is ready. This works well for many block
230 * devices. However some block devices (typically virtual devices
231 * such as md or lvm) do not benefit from the processing on the
232 * request queue, and are served best by having the requests passed
233 * directly to them. This can be achieved by providing a function
234 * to blk_queue_make_request().
237 * The driver that does this *must* be able to deal appropriately
238 * with buffers in "highmemory". This can be accomplished by either calling
239 * __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
240 * blk_queue_bounce() to create a buffer in normal memory.
242 void blk_queue_make_request(request_queue_t * q, make_request_fn * mfn)
247 q->nr_requests = BLKDEV_MAX_RQ;
248 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
249 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
250 q->make_request_fn = mfn;
251 q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
252 q->backing_dev_info.state = 0;
253 q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
254 blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
255 blk_queue_hardsect_size(q, 512);
256 blk_queue_dma_alignment(q, 511);
257 blk_queue_congestion_threshold(q);
258 q->nr_batching = BLK_BATCH_REQ;
260 q->unplug_thresh = 4; /* hmm */
261 q->unplug_delay = (3 * HZ) / 1000; /* 3 milliseconds */
262 if (q->unplug_delay == 0)
265 INIT_WORK(&q->unplug_work, blk_unplug_work, q);
267 q->unplug_timer.function = blk_unplug_timeout;
268 q->unplug_timer.data = (unsigned long)q;
271 * by default assume old behaviour and bounce for any highmem page
273 blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
275 blk_queue_activity_fn(q, NULL, NULL);
278 EXPORT_SYMBOL(blk_queue_make_request);
280 static inline void rq_init(request_queue_t *q, struct request *rq)
282 INIT_LIST_HEAD(&rq->queuelist);
283 INIT_LIST_HEAD(&rq->donelist);
286 rq->rq_status = RQ_ACTIVE;
287 rq->bio = rq->biotail = NULL;
296 rq->nr_phys_segments = 0;
299 rq->end_io_data = NULL;
300 rq->completion_data = NULL;
304 * blk_queue_ordered - does this queue support ordered writes
305 * @q: the request queue
306 * @ordered: one of QUEUE_ORDERED_*
307 * @prepare_flush_fn: rq setup helper for cache flush ordered writes
310 * For journalled file systems, doing ordered writes on a commit
311 * block instead of explicitly doing wait_on_buffer (which is bad
312 * for performance) can be a big win. Block drivers supporting this
313 * feature should call this function and indicate so.
316 int blk_queue_ordered(request_queue_t *q, unsigned ordered,
317 prepare_flush_fn *prepare_flush_fn)
319 if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
320 prepare_flush_fn == NULL) {
321 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
325 if (ordered != QUEUE_ORDERED_NONE &&
326 ordered != QUEUE_ORDERED_DRAIN &&
327 ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
328 ordered != QUEUE_ORDERED_DRAIN_FUA &&
329 ordered != QUEUE_ORDERED_TAG &&
330 ordered != QUEUE_ORDERED_TAG_FLUSH &&
331 ordered != QUEUE_ORDERED_TAG_FUA) {
332 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
336 q->ordered = ordered;
337 q->next_ordered = ordered;
338 q->prepare_flush_fn = prepare_flush_fn;
343 EXPORT_SYMBOL(blk_queue_ordered);
346 * blk_queue_issue_flush_fn - set function for issuing a flush
347 * @q: the request queue
348 * @iff: the function to be called issuing the flush
351 * If a driver supports issuing a flush command, the support is notified
352 * to the block layer by defining it through this call.
355 void blk_queue_issue_flush_fn(request_queue_t *q, issue_flush_fn *iff)
357 q->issue_flush_fn = iff;
360 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
363 * Cache flushing for ordered writes handling
365 inline unsigned blk_ordered_cur_seq(request_queue_t *q)
369 return 1 << ffz(q->ordseq);
372 unsigned blk_ordered_req_seq(struct request *rq)
374 request_queue_t *q = rq->q;
376 BUG_ON(q->ordseq == 0);
378 if (rq == &q->pre_flush_rq)
379 return QUEUE_ORDSEQ_PREFLUSH;
380 if (rq == &q->bar_rq)
381 return QUEUE_ORDSEQ_BAR;
382 if (rq == &q->post_flush_rq)
383 return QUEUE_ORDSEQ_POSTFLUSH;
385 if ((rq->cmd_flags & REQ_ORDERED_COLOR) ==
386 (q->orig_bar_rq->cmd_flags & REQ_ORDERED_COLOR))
387 return QUEUE_ORDSEQ_DRAIN;
389 return QUEUE_ORDSEQ_DONE;
392 void blk_ordered_complete_seq(request_queue_t *q, unsigned seq, int error)
397 if (error && !q->orderr)
400 BUG_ON(q->ordseq & seq);
403 if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
407 * Okay, sequence complete.
410 uptodate = q->orderr ? q->orderr : 1;
414 end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
415 end_that_request_last(rq, uptodate);
418 static void pre_flush_end_io(struct request *rq, int error)
420 elv_completed_request(rq->q, rq);
421 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
424 static void bar_end_io(struct request *rq, int error)
426 elv_completed_request(rq->q, rq);
427 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
430 static void post_flush_end_io(struct request *rq, int error)
432 elv_completed_request(rq->q, rq);
433 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
436 static void queue_flush(request_queue_t *q, unsigned which)
439 rq_end_io_fn *end_io;
441 if (which == QUEUE_ORDERED_PREFLUSH) {
442 rq = &q->pre_flush_rq;
443 end_io = pre_flush_end_io;
445 rq = &q->post_flush_rq;
446 end_io = post_flush_end_io;
449 rq->cmd_flags = REQ_HARDBARRIER;
451 rq->elevator_private = NULL;
452 rq->rq_disk = q->bar_rq.rq_disk;
455 q->prepare_flush_fn(q, rq);
457 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
460 static inline struct request *start_ordered(request_queue_t *q,
465 q->ordered = q->next_ordered;
466 q->ordseq |= QUEUE_ORDSEQ_STARTED;
469 * Prep proxy barrier request.
471 blkdev_dequeue_request(rq);
476 if (bio_data_dir(q->orig_bar_rq->bio) == WRITE)
477 rq->cmd_flags |= REQ_RW;
478 rq->cmd_flags |= q->ordered & QUEUE_ORDERED_FUA ? REQ_FUA : 0;
479 rq->elevator_private = NULL;
481 init_request_from_bio(rq, q->orig_bar_rq->bio);
482 rq->end_io = bar_end_io;
485 * Queue ordered sequence. As we stack them at the head, we
486 * need to queue in reverse order. Note that we rely on that
487 * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
488 * request gets inbetween ordered sequence.
490 if (q->ordered & QUEUE_ORDERED_POSTFLUSH)
491 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
493 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
495 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
497 if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
498 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
499 rq = &q->pre_flush_rq;
501 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
503 if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
504 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
511 int blk_do_ordered(request_queue_t *q, struct request **rqp)
513 struct request *rq = *rqp;
514 int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
520 if (q->next_ordered != QUEUE_ORDERED_NONE) {
521 *rqp = start_ordered(q, rq);
525 * This can happen when the queue switches to
526 * ORDERED_NONE while this request is on it.
528 blkdev_dequeue_request(rq);
529 end_that_request_first(rq, -EOPNOTSUPP,
530 rq->hard_nr_sectors);
531 end_that_request_last(rq, -EOPNOTSUPP);
538 * Ordered sequence in progress
541 /* Special requests are not subject to ordering rules. */
542 if (!blk_fs_request(rq) &&
543 rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
546 if (q->ordered & QUEUE_ORDERED_TAG) {
547 /* Ordered by tag. Blocking the next barrier is enough. */
548 if (is_barrier && rq != &q->bar_rq)
551 /* Ordered by draining. Wait for turn. */
552 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
553 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
560 static int flush_dry_bio_endio(struct bio *bio, unsigned int bytes, int error)
562 request_queue_t *q = bio->bi_private;
563 struct bio_vec *bvec;
567 * This is dry run, restore bio_sector and size. We'll finish
568 * this request again with the original bi_end_io after an
569 * error occurs or post flush is complete.
578 bio_for_each_segment(bvec, bio, i) {
579 bvec->bv_len += bvec->bv_offset;
584 set_bit(BIO_UPTODATE, &bio->bi_flags);
585 bio->bi_size = q->bi_size;
586 bio->bi_sector -= (q->bi_size >> 9);
592 static inline int ordered_bio_endio(struct request *rq, struct bio *bio,
593 unsigned int nbytes, int error)
595 request_queue_t *q = rq->q;
599 if (&q->bar_rq != rq)
603 * Okay, this is the barrier request in progress, dry finish it.
605 if (error && !q->orderr)
608 endio = bio->bi_end_io;
609 private = bio->bi_private;
610 bio->bi_end_io = flush_dry_bio_endio;
613 bio_endio(bio, nbytes, error);
615 bio->bi_end_io = endio;
616 bio->bi_private = private;
622 * blk_queue_bounce_limit - set bounce buffer limit for queue
623 * @q: the request queue for the device
624 * @dma_addr: bus address limit
627 * Different hardware can have different requirements as to what pages
628 * it can do I/O directly to. A low level driver can call
629 * blk_queue_bounce_limit to have lower memory pages allocated as bounce
630 * buffers for doing I/O to pages residing above @page.
632 void blk_queue_bounce_limit(request_queue_t *q, u64 dma_addr)
634 unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
637 q->bounce_gfp = GFP_NOIO;
638 #if BITS_PER_LONG == 64
639 /* Assume anything <= 4GB can be handled by IOMMU.
640 Actually some IOMMUs can handle everything, but I don't
641 know of a way to test this here. */
642 if (bounce_pfn < (min_t(u64,0xffffffff,BLK_BOUNCE_HIGH) >> PAGE_SHIFT))
644 q->bounce_pfn = max_low_pfn;
646 if (bounce_pfn < blk_max_low_pfn)
648 q->bounce_pfn = bounce_pfn;
651 init_emergency_isa_pool();
652 q->bounce_gfp = GFP_NOIO | GFP_DMA;
653 q->bounce_pfn = bounce_pfn;
657 EXPORT_SYMBOL(blk_queue_bounce_limit);
660 * blk_queue_max_sectors - set max sectors for a request for this queue
661 * @q: the request queue for the device
662 * @max_sectors: max sectors in the usual 512b unit
665 * Enables a low level driver to set an upper limit on the size of
668 void blk_queue_max_sectors(request_queue_t *q, unsigned int max_sectors)
670 if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
671 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
672 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
675 if (BLK_DEF_MAX_SECTORS > max_sectors)
676 q->max_hw_sectors = q->max_sectors = max_sectors;
678 q->max_sectors = BLK_DEF_MAX_SECTORS;
679 q->max_hw_sectors = max_sectors;
683 EXPORT_SYMBOL(blk_queue_max_sectors);
686 * blk_queue_max_phys_segments - set max phys segments for a request for this queue
687 * @q: the request queue for the device
688 * @max_segments: max number of segments
691 * Enables a low level driver to set an upper limit on the number of
692 * physical data segments in a request. This would be the largest sized
693 * scatter list the driver could handle.
695 void blk_queue_max_phys_segments(request_queue_t *q, unsigned short max_segments)
699 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
702 q->max_phys_segments = max_segments;
705 EXPORT_SYMBOL(blk_queue_max_phys_segments);
708 * blk_queue_max_hw_segments - set max hw segments for a request for this queue
709 * @q: the request queue for the device
710 * @max_segments: max number of segments
713 * Enables a low level driver to set an upper limit on the number of
714 * hw data segments in a request. This would be the largest number of
715 * address/length pairs the host adapter can actually give as once
718 void blk_queue_max_hw_segments(request_queue_t *q, unsigned short max_segments)
722 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
725 q->max_hw_segments = max_segments;
728 EXPORT_SYMBOL(blk_queue_max_hw_segments);
731 * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
732 * @q: the request queue for the device
733 * @max_size: max size of segment in bytes
736 * Enables a low level driver to set an upper limit on the size of a
739 void blk_queue_max_segment_size(request_queue_t *q, unsigned int max_size)
741 if (max_size < PAGE_CACHE_SIZE) {
742 max_size = PAGE_CACHE_SIZE;
743 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
746 q->max_segment_size = max_size;
749 EXPORT_SYMBOL(blk_queue_max_segment_size);
752 * blk_queue_hardsect_size - set hardware sector size for the queue
753 * @q: the request queue for the device
754 * @size: the hardware sector size, in bytes
757 * This should typically be set to the lowest possible sector size
758 * that the hardware can operate on (possible without reverting to
759 * even internal read-modify-write operations). Usually the default
760 * of 512 covers most hardware.
762 void blk_queue_hardsect_size(request_queue_t *q, unsigned short size)
764 q->hardsect_size = size;
767 EXPORT_SYMBOL(blk_queue_hardsect_size);
770 * Returns the minimum that is _not_ zero, unless both are zero.
772 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
775 * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
776 * @t: the stacking driver (top)
777 * @b: the underlying device (bottom)
779 void blk_queue_stack_limits(request_queue_t *t, request_queue_t *b)
781 /* zero is "infinity" */
782 t->max_sectors = min_not_zero(t->max_sectors,b->max_sectors);
783 t->max_hw_sectors = min_not_zero(t->max_hw_sectors,b->max_hw_sectors);
785 t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
786 t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
787 t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
788 t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
789 if (!test_bit(QUEUE_FLAG_CLUSTER, &b->queue_flags))
790 clear_bit(QUEUE_FLAG_CLUSTER, &t->queue_flags);
793 EXPORT_SYMBOL(blk_queue_stack_limits);
796 * blk_queue_segment_boundary - set boundary rules for segment merging
797 * @q: the request queue for the device
798 * @mask: the memory boundary mask
800 void blk_queue_segment_boundary(request_queue_t *q, unsigned long mask)
802 if (mask < PAGE_CACHE_SIZE - 1) {
803 mask = PAGE_CACHE_SIZE - 1;
804 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
807 q->seg_boundary_mask = mask;
810 EXPORT_SYMBOL(blk_queue_segment_boundary);
813 * blk_queue_dma_alignment - set dma length and memory alignment
814 * @q: the request queue for the device
815 * @mask: alignment mask
818 * set required memory and length aligment for direct dma transactions.
819 * this is used when buiding direct io requests for the queue.
822 void blk_queue_dma_alignment(request_queue_t *q, int mask)
824 q->dma_alignment = mask;
827 EXPORT_SYMBOL(blk_queue_dma_alignment);
830 * blk_queue_find_tag - find a request by its tag and queue
831 * @q: The request queue for the device
832 * @tag: The tag of the request
835 * Should be used when a device returns a tag and you want to match
838 * no locks need be held.
840 struct request *blk_queue_find_tag(request_queue_t *q, int tag)
842 struct blk_queue_tag *bqt = q->queue_tags;
844 if (unlikely(bqt == NULL || tag >= bqt->real_max_depth))
847 return bqt->tag_index[tag];
850 EXPORT_SYMBOL(blk_queue_find_tag);
853 * __blk_free_tags - release a given set of tag maintenance info
854 * @bqt: the tag map to free
856 * Tries to free the specified @bqt@. Returns true if it was
857 * actually freed and false if there are still references using it
859 static int __blk_free_tags(struct blk_queue_tag *bqt)
863 retval = atomic_dec_and_test(&bqt->refcnt);
866 BUG_ON(!list_empty(&bqt->busy_list));
868 kfree(bqt->tag_index);
869 bqt->tag_index = NULL;
882 * __blk_queue_free_tags - release tag maintenance info
883 * @q: the request queue for the device
886 * blk_cleanup_queue() will take care of calling this function, if tagging
887 * has been used. So there's no need to call this directly.
889 static void __blk_queue_free_tags(request_queue_t *q)
891 struct blk_queue_tag *bqt = q->queue_tags;
896 __blk_free_tags(bqt);
898 q->queue_tags = NULL;
899 q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
904 * blk_free_tags - release a given set of tag maintenance info
905 * @bqt: the tag map to free
907 * For externally managed @bqt@ frees the map. Callers of this
908 * function must guarantee to have released all the queues that
909 * might have been using this tag map.
911 void blk_free_tags(struct blk_queue_tag *bqt)
913 if (unlikely(!__blk_free_tags(bqt)))
916 EXPORT_SYMBOL(blk_free_tags);
919 * blk_queue_free_tags - release tag maintenance info
920 * @q: the request queue for the device
923 * This is used to disabled tagged queuing to a device, yet leave
926 void blk_queue_free_tags(request_queue_t *q)
928 clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
931 EXPORT_SYMBOL(blk_queue_free_tags);
934 init_tag_map(request_queue_t *q, struct blk_queue_tag *tags, int depth)
936 struct request **tag_index;
937 unsigned long *tag_map;
940 if (q && depth > q->nr_requests * 2) {
941 depth = q->nr_requests * 2;
942 printk(KERN_ERR "%s: adjusted depth to %d\n",
943 __FUNCTION__, depth);
946 tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC);
950 nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
951 tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
955 tags->real_max_depth = depth;
956 tags->max_depth = depth;
957 tags->tag_index = tag_index;
958 tags->tag_map = tag_map;
966 static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q,
969 struct blk_queue_tag *tags;
971 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
975 if (init_tag_map(q, tags, depth))
978 INIT_LIST_HEAD(&tags->busy_list);
980 atomic_set(&tags->refcnt, 1);
988 * blk_init_tags - initialize the tag info for an external tag map
989 * @depth: the maximum queue depth supported
990 * @tags: the tag to use
992 struct blk_queue_tag *blk_init_tags(int depth)
994 return __blk_queue_init_tags(NULL, depth);
996 EXPORT_SYMBOL(blk_init_tags);
999 * blk_queue_init_tags - initialize the queue tag info
1000 * @q: the request queue for the device
1001 * @depth: the maximum queue depth supported
1002 * @tags: the tag to use
1004 int blk_queue_init_tags(request_queue_t *q, int depth,
1005 struct blk_queue_tag *tags)
1009 BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
1011 if (!tags && !q->queue_tags) {
1012 tags = __blk_queue_init_tags(q, depth);
1016 } else if (q->queue_tags) {
1017 if ((rc = blk_queue_resize_tags(q, depth)))
1019 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
1022 atomic_inc(&tags->refcnt);
1025 * assign it, all done
1027 q->queue_tags = tags;
1028 q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
1035 EXPORT_SYMBOL(blk_queue_init_tags);
1038 * blk_queue_resize_tags - change the queueing depth
1039 * @q: the request queue for the device
1040 * @new_depth: the new max command queueing depth
1043 * Must be called with the queue lock held.
1045 int blk_queue_resize_tags(request_queue_t *q, int new_depth)
1047 struct blk_queue_tag *bqt = q->queue_tags;
1048 struct request **tag_index;
1049 unsigned long *tag_map;
1050 int max_depth, nr_ulongs;
1056 * if we already have large enough real_max_depth. just
1057 * adjust max_depth. *NOTE* as requests with tag value
1058 * between new_depth and real_max_depth can be in-flight, tag
1059 * map can not be shrunk blindly here.
1061 if (new_depth <= bqt->real_max_depth) {
1062 bqt->max_depth = new_depth;
1067 * Currently cannot replace a shared tag map with a new
1068 * one, so error out if this is the case
1070 if (atomic_read(&bqt->refcnt) != 1)
1074 * save the old state info, so we can copy it back
1076 tag_index = bqt->tag_index;
1077 tag_map = bqt->tag_map;
1078 max_depth = bqt->real_max_depth;
1080 if (init_tag_map(q, bqt, new_depth))
1083 memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
1084 nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
1085 memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
1092 EXPORT_SYMBOL(blk_queue_resize_tags);
1095 * blk_queue_end_tag - end tag operations for a request
1096 * @q: the request queue for the device
1097 * @rq: the request that has completed
1100 * Typically called when end_that_request_first() returns 0, meaning
1101 * all transfers have been done for a request. It's important to call
1102 * this function before end_that_request_last(), as that will put the
1103 * request back on the free list thus corrupting the internal tag list.
1106 * queue lock must be held.
1108 void blk_queue_end_tag(request_queue_t *q, struct request *rq)
1110 struct blk_queue_tag *bqt = q->queue_tags;
1115 if (unlikely(tag >= bqt->real_max_depth))
1117 * This can happen after tag depth has been reduced.
1118 * FIXME: how about a warning or info message here?
1122 if (unlikely(!__test_and_clear_bit(tag, bqt->tag_map))) {
1123 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
1128 list_del_init(&rq->queuelist);
1129 rq->cmd_flags &= ~REQ_QUEUED;
1132 if (unlikely(bqt->tag_index[tag] == NULL))
1133 printk(KERN_ERR "%s: tag %d is missing\n",
1136 bqt->tag_index[tag] = NULL;
1140 EXPORT_SYMBOL(blk_queue_end_tag);
1143 * blk_queue_start_tag - find a free tag and assign it
1144 * @q: the request queue for the device
1145 * @rq: the block request that needs tagging
1148 * This can either be used as a stand-alone helper, or possibly be
1149 * assigned as the queue &prep_rq_fn (in which case &struct request
1150 * automagically gets a tag assigned). Note that this function
1151 * assumes that any type of request can be queued! if this is not
1152 * true for your device, you must check the request type before
1153 * calling this function. The request will also be removed from
1154 * the request queue, so it's the drivers responsibility to readd
1155 * it if it should need to be restarted for some reason.
1158 * queue lock must be held.
1160 int blk_queue_start_tag(request_queue_t *q, struct request *rq)
1162 struct blk_queue_tag *bqt = q->queue_tags;
1165 if (unlikely((rq->cmd_flags & REQ_QUEUED))) {
1167 "%s: request %p for device [%s] already tagged %d",
1169 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
1173 tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
1174 if (tag >= bqt->max_depth)
1177 __set_bit(tag, bqt->tag_map);
1179 rq->cmd_flags |= REQ_QUEUED;
1181 bqt->tag_index[tag] = rq;
1182 blkdev_dequeue_request(rq);
1183 list_add(&rq->queuelist, &bqt->busy_list);
1188 EXPORT_SYMBOL(blk_queue_start_tag);
1191 * blk_queue_invalidate_tags - invalidate all pending tags
1192 * @q: the request queue for the device
1195 * Hardware conditions may dictate a need to stop all pending requests.
1196 * In this case, we will safely clear the block side of the tag queue and
1197 * readd all requests to the request queue in the right order.
1200 * queue lock must be held.
1202 void blk_queue_invalidate_tags(request_queue_t *q)
1204 struct blk_queue_tag *bqt = q->queue_tags;
1205 struct list_head *tmp, *n;
1208 list_for_each_safe(tmp, n, &bqt->busy_list) {
1209 rq = list_entry_rq(tmp);
1211 if (rq->tag == -1) {
1213 "%s: bad tag found on list\n", __FUNCTION__);
1214 list_del_init(&rq->queuelist);
1215 rq->cmd_flags &= ~REQ_QUEUED;
1217 blk_queue_end_tag(q, rq);
1219 rq->cmd_flags &= ~REQ_STARTED;
1220 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1224 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1226 void blk_dump_rq_flags(struct request *rq, char *msg)
1230 printk("%s: dev %s: type=%x, flags=%x\n", msg,
1231 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->cmd_type,
1234 printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1236 rq->current_nr_sectors);
1237 printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1239 if (blk_pc_request(rq)) {
1241 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1242 printk("%02x ", rq->cmd[bit]);
1247 EXPORT_SYMBOL(blk_dump_rq_flags);
1249 void blk_recount_segments(request_queue_t *q, struct bio *bio)
1251 struct bio_vec *bv, *bvprv = NULL;
1252 int i, nr_phys_segs, nr_hw_segs, seg_size, hw_seg_size, cluster;
1253 int high, highprv = 1;
1255 if (unlikely(!bio->bi_io_vec))
1258 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1259 hw_seg_size = seg_size = nr_phys_segs = nr_hw_segs = 0;
1260 bio_for_each_segment(bv, bio, i) {
1262 * the trick here is making sure that a high page is never
1263 * considered part of another segment, since that might
1264 * change with the bounce page.
1266 high = page_to_pfn(bv->bv_page) >= q->bounce_pfn;
1267 if (high || highprv)
1268 goto new_hw_segment;
1270 if (seg_size + bv->bv_len > q->max_segment_size)
1272 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1274 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1276 if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1277 goto new_hw_segment;
1279 seg_size += bv->bv_len;
1280 hw_seg_size += bv->bv_len;
1285 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1286 !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len)) {
1287 hw_seg_size += bv->bv_len;
1290 if (hw_seg_size > bio->bi_hw_front_size)
1291 bio->bi_hw_front_size = hw_seg_size;
1292 hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1298 seg_size = bv->bv_len;
1301 if (hw_seg_size > bio->bi_hw_back_size)
1302 bio->bi_hw_back_size = hw_seg_size;
1303 if (nr_hw_segs == 1 && hw_seg_size > bio->bi_hw_front_size)
1304 bio->bi_hw_front_size = hw_seg_size;
1305 bio->bi_phys_segments = nr_phys_segs;
1306 bio->bi_hw_segments = nr_hw_segs;
1307 bio->bi_flags |= (1 << BIO_SEG_VALID);
1311 static int blk_phys_contig_segment(request_queue_t *q, struct bio *bio,
1314 if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1317 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1319 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1323 * bio and nxt are contigous in memory, check if the queue allows
1324 * these two to be merged into one
1326 if (BIO_SEG_BOUNDARY(q, bio, nxt))
1332 static int blk_hw_contig_segment(request_queue_t *q, struct bio *bio,
1335 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1336 blk_recount_segments(q, bio);
1337 if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1338 blk_recount_segments(q, nxt);
1339 if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1340 BIOVEC_VIRT_OVERSIZE(bio->bi_hw_front_size + bio->bi_hw_back_size))
1342 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1349 * map a request to scatterlist, return number of sg entries setup. Caller
1350 * must make sure sg can hold rq->nr_phys_segments entries
1352 int blk_rq_map_sg(request_queue_t *q, struct request *rq, struct scatterlist *sg)
1354 struct bio_vec *bvec, *bvprv;
1356 int nsegs, i, cluster;
1359 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1362 * for each bio in rq
1365 rq_for_each_bio(bio, rq) {
1367 * for each segment in bio
1369 bio_for_each_segment(bvec, bio, i) {
1370 int nbytes = bvec->bv_len;
1372 if (bvprv && cluster) {
1373 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1376 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1378 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1381 sg[nsegs - 1].length += nbytes;
1384 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1385 sg[nsegs].page = bvec->bv_page;
1386 sg[nsegs].length = nbytes;
1387 sg[nsegs].offset = bvec->bv_offset;
1392 } /* segments in bio */
1398 EXPORT_SYMBOL(blk_rq_map_sg);
1401 * the standard queue merge functions, can be overridden with device
1402 * specific ones if so desired
1405 static inline int ll_new_mergeable(request_queue_t *q,
1406 struct request *req,
1409 int nr_phys_segs = bio_phys_segments(q, bio);
1411 if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1412 req->cmd_flags |= REQ_NOMERGE;
1413 if (req == q->last_merge)
1414 q->last_merge = NULL;
1419 * A hw segment is just getting larger, bump just the phys
1422 req->nr_phys_segments += nr_phys_segs;
1426 static inline int ll_new_hw_segment(request_queue_t *q,
1427 struct request *req,
1430 int nr_hw_segs = bio_hw_segments(q, bio);
1431 int nr_phys_segs = bio_phys_segments(q, bio);
1433 if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1434 || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1435 req->cmd_flags |= REQ_NOMERGE;
1436 if (req == q->last_merge)
1437 q->last_merge = NULL;
1442 * This will form the start of a new hw segment. Bump both
1445 req->nr_hw_segments += nr_hw_segs;
1446 req->nr_phys_segments += nr_phys_segs;
1450 static int ll_back_merge_fn(request_queue_t *q, struct request *req,
1453 unsigned short max_sectors;
1456 if (unlikely(blk_pc_request(req)))
1457 max_sectors = q->max_hw_sectors;
1459 max_sectors = q->max_sectors;
1461 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1462 req->cmd_flags |= REQ_NOMERGE;
1463 if (req == q->last_merge)
1464 q->last_merge = NULL;
1467 if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1468 blk_recount_segments(q, req->biotail);
1469 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1470 blk_recount_segments(q, bio);
1471 len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1472 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1473 !BIOVEC_VIRT_OVERSIZE(len)) {
1474 int mergeable = ll_new_mergeable(q, req, bio);
1477 if (req->nr_hw_segments == 1)
1478 req->bio->bi_hw_front_size = len;
1479 if (bio->bi_hw_segments == 1)
1480 bio->bi_hw_back_size = len;
1485 return ll_new_hw_segment(q, req, bio);
1488 static int ll_front_merge_fn(request_queue_t *q, struct request *req,
1491 unsigned short max_sectors;
1494 if (unlikely(blk_pc_request(req)))
1495 max_sectors = q->max_hw_sectors;
1497 max_sectors = q->max_sectors;
1500 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1501 req->cmd_flags |= REQ_NOMERGE;
1502 if (req == q->last_merge)
1503 q->last_merge = NULL;
1506 len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1507 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1508 blk_recount_segments(q, bio);
1509 if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1510 blk_recount_segments(q, req->bio);
1511 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1512 !BIOVEC_VIRT_OVERSIZE(len)) {
1513 int mergeable = ll_new_mergeable(q, req, bio);
1516 if (bio->bi_hw_segments == 1)
1517 bio->bi_hw_front_size = len;
1518 if (req->nr_hw_segments == 1)
1519 req->biotail->bi_hw_back_size = len;
1524 return ll_new_hw_segment(q, req, bio);
1527 static int ll_merge_requests_fn(request_queue_t *q, struct request *req,
1528 struct request *next)
1530 int total_phys_segments;
1531 int total_hw_segments;
1534 * First check if the either of the requests are re-queued
1535 * requests. Can't merge them if they are.
1537 if (req->special || next->special)
1541 * Will it become too large?
1543 if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1546 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1547 if (blk_phys_contig_segment(q, req->biotail, next->bio))
1548 total_phys_segments--;
1550 if (total_phys_segments > q->max_phys_segments)
1553 total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1554 if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1555 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1557 * propagate the combined length to the end of the requests
1559 if (req->nr_hw_segments == 1)
1560 req->bio->bi_hw_front_size = len;
1561 if (next->nr_hw_segments == 1)
1562 next->biotail->bi_hw_back_size = len;
1563 total_hw_segments--;
1566 if (total_hw_segments > q->max_hw_segments)
1569 /* Merge is OK... */
1570 req->nr_phys_segments = total_phys_segments;
1571 req->nr_hw_segments = total_hw_segments;
1576 * "plug" the device if there are no outstanding requests: this will
1577 * force the transfer to start only after we have put all the requests
1580 * This is called with interrupts off and no requests on the queue and
1581 * with the queue lock held.
1583 void blk_plug_device(request_queue_t *q)
1585 WARN_ON(!irqs_disabled());
1588 * don't plug a stopped queue, it must be paired with blk_start_queue()
1589 * which will restart the queueing
1591 if (blk_queue_stopped(q))
1594 if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
1595 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1596 blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
1600 EXPORT_SYMBOL(blk_plug_device);
1603 * remove the queue from the plugged list, if present. called with
1604 * queue lock held and interrupts disabled.
1606 int blk_remove_plug(request_queue_t *q)
1608 WARN_ON(!irqs_disabled());
1610 if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1613 del_timer(&q->unplug_timer);
1617 EXPORT_SYMBOL(blk_remove_plug);
1620 * remove the plug and let it rip..
1622 void __generic_unplug_device(request_queue_t *q)
1624 if (unlikely(blk_queue_stopped(q)))
1627 if (!blk_remove_plug(q))
1632 EXPORT_SYMBOL(__generic_unplug_device);
1635 * generic_unplug_device - fire a request queue
1636 * @q: The &request_queue_t in question
1639 * Linux uses plugging to build bigger requests queues before letting
1640 * the device have at them. If a queue is plugged, the I/O scheduler
1641 * is still adding and merging requests on the queue. Once the queue
1642 * gets unplugged, the request_fn defined for the queue is invoked and
1643 * transfers started.
1645 void generic_unplug_device(request_queue_t *q)
1647 spin_lock_irq(q->queue_lock);
1648 __generic_unplug_device(q);
1649 spin_unlock_irq(q->queue_lock);
1651 EXPORT_SYMBOL(generic_unplug_device);
1653 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1656 request_queue_t *q = bdi->unplug_io_data;
1659 * devices don't necessarily have an ->unplug_fn defined
1662 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1663 q->rq.count[READ] + q->rq.count[WRITE]);
1669 static void blk_unplug_work(void *data)
1671 request_queue_t *q = data;
1673 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1674 q->rq.count[READ] + q->rq.count[WRITE]);
1679 static void blk_unplug_timeout(unsigned long data)
1681 request_queue_t *q = (request_queue_t *)data;
1683 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
1684 q->rq.count[READ] + q->rq.count[WRITE]);
1686 kblockd_schedule_work(&q->unplug_work);
1690 * blk_start_queue - restart a previously stopped queue
1691 * @q: The &request_queue_t in question
1694 * blk_start_queue() will clear the stop flag on the queue, and call
1695 * the request_fn for the queue if it was in a stopped state when
1696 * entered. Also see blk_stop_queue(). Queue lock must be held.
1698 void blk_start_queue(request_queue_t *q)
1700 WARN_ON(!irqs_disabled());
1702 clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1705 * one level of recursion is ok and is much faster than kicking
1706 * the unplug handling
1708 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1710 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1713 kblockd_schedule_work(&q->unplug_work);
1717 EXPORT_SYMBOL(blk_start_queue);
1720 * blk_stop_queue - stop a queue
1721 * @q: The &request_queue_t in question
1724 * The Linux block layer assumes that a block driver will consume all
1725 * entries on the request queue when the request_fn strategy is called.
1726 * Often this will not happen, because of hardware limitations (queue
1727 * depth settings). If a device driver gets a 'queue full' response,
1728 * or if it simply chooses not to queue more I/O at one point, it can
1729 * call this function to prevent the request_fn from being called until
1730 * the driver has signalled it's ready to go again. This happens by calling
1731 * blk_start_queue() to restart queue operations. Queue lock must be held.
1733 void blk_stop_queue(request_queue_t *q)
1736 set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1738 EXPORT_SYMBOL(blk_stop_queue);
1741 * blk_sync_queue - cancel any pending callbacks on a queue
1745 * The block layer may perform asynchronous callback activity
1746 * on a queue, such as calling the unplug function after a timeout.
1747 * A block device may call blk_sync_queue to ensure that any
1748 * such activity is cancelled, thus allowing it to release resources
1749 * the the callbacks might use. The caller must already have made sure
1750 * that its ->make_request_fn will not re-add plugging prior to calling
1754 void blk_sync_queue(struct request_queue *q)
1756 del_timer_sync(&q->unplug_timer);
1759 EXPORT_SYMBOL(blk_sync_queue);
1762 * blk_run_queue - run a single device queue
1763 * @q: The queue to run
1765 void blk_run_queue(struct request_queue *q)
1767 unsigned long flags;
1769 spin_lock_irqsave(q->queue_lock, flags);
1773 * Only recurse once to avoid overrunning the stack, let the unplug
1774 * handling reinvoke the handler shortly if we already got there.
1776 if (!elv_queue_empty(q)) {
1777 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1779 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1782 kblockd_schedule_work(&q->unplug_work);
1786 spin_unlock_irqrestore(q->queue_lock, flags);
1788 EXPORT_SYMBOL(blk_run_queue);
1791 * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1792 * @kobj: the kobj belonging of the request queue to be released
1795 * blk_cleanup_queue is the pair to blk_init_queue() or
1796 * blk_queue_make_request(). It should be called when a request queue is
1797 * being released; typically when a block device is being de-registered.
1798 * Currently, its primary task it to free all the &struct request
1799 * structures that were allocated to the queue and the queue itself.
1802 * Hopefully the low level driver will have finished any
1803 * outstanding requests first...
1805 static void blk_release_queue(struct kobject *kobj)
1807 request_queue_t *q = container_of(kobj, struct request_queue, kobj);
1808 struct request_list *rl = &q->rq;
1813 mempool_destroy(rl->rq_pool);
1816 __blk_queue_free_tags(q);
1818 blk_trace_shutdown(q);
1820 kmem_cache_free(requestq_cachep, q);
1823 void blk_put_queue(request_queue_t *q)
1825 kobject_put(&q->kobj);
1827 EXPORT_SYMBOL(blk_put_queue);
1829 void blk_cleanup_queue(request_queue_t * q)
1831 mutex_lock(&q->sysfs_lock);
1832 set_bit(QUEUE_FLAG_DEAD, &q->queue_flags);
1833 mutex_unlock(&q->sysfs_lock);
1836 elevator_exit(q->elevator);
1841 EXPORT_SYMBOL(blk_cleanup_queue);
1843 static int blk_init_free_list(request_queue_t *q)
1845 struct request_list *rl = &q->rq;
1847 rl->count[READ] = rl->count[WRITE] = 0;
1848 rl->starved[READ] = rl->starved[WRITE] = 0;
1850 init_waitqueue_head(&rl->wait[READ]);
1851 init_waitqueue_head(&rl->wait[WRITE]);
1853 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1854 mempool_free_slab, request_cachep, q->node);
1862 request_queue_t *blk_alloc_queue(gfp_t gfp_mask)
1864 return blk_alloc_queue_node(gfp_mask, -1);
1866 EXPORT_SYMBOL(blk_alloc_queue);
1868 static struct kobj_type queue_ktype;
1870 request_queue_t *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
1874 q = kmem_cache_alloc_node(requestq_cachep, gfp_mask, node_id);
1878 memset(q, 0, sizeof(*q));
1879 init_timer(&q->unplug_timer);
1881 snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
1882 q->kobj.ktype = &queue_ktype;
1883 kobject_init(&q->kobj);
1885 q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1886 q->backing_dev_info.unplug_io_data = q;
1888 mutex_init(&q->sysfs_lock);
1892 EXPORT_SYMBOL(blk_alloc_queue_node);
1895 * blk_init_queue - prepare a request queue for use with a block device
1896 * @rfn: The function to be called to process requests that have been
1897 * placed on the queue.
1898 * @lock: Request queue spin lock
1901 * If a block device wishes to use the standard request handling procedures,
1902 * which sorts requests and coalesces adjacent requests, then it must
1903 * call blk_init_queue(). The function @rfn will be called when there
1904 * are requests on the queue that need to be processed. If the device
1905 * supports plugging, then @rfn may not be called immediately when requests
1906 * are available on the queue, but may be called at some time later instead.
1907 * Plugged queues are generally unplugged when a buffer belonging to one
1908 * of the requests on the queue is needed, or due to memory pressure.
1910 * @rfn is not required, or even expected, to remove all requests off the
1911 * queue, but only as many as it can handle at a time. If it does leave
1912 * requests on the queue, it is responsible for arranging that the requests
1913 * get dealt with eventually.
1915 * The queue spin lock must be held while manipulating the requests on the
1916 * request queue; this lock will be taken also from interrupt context, so irq
1917 * disabling is needed for it.
1919 * Function returns a pointer to the initialized request queue, or NULL if
1920 * it didn't succeed.
1923 * blk_init_queue() must be paired with a blk_cleanup_queue() call
1924 * when the block device is deactivated (such as at module unload).
1927 request_queue_t *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1929 return blk_init_queue_node(rfn, lock, -1);
1931 EXPORT_SYMBOL(blk_init_queue);
1934 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1936 request_queue_t *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1942 if (blk_init_free_list(q)) {
1943 kmem_cache_free(requestq_cachep, q);
1948 * if caller didn't supply a lock, they get per-queue locking with
1952 spin_lock_init(&q->__queue_lock);
1953 lock = &q->__queue_lock;
1956 q->request_fn = rfn;
1957 q->back_merge_fn = ll_back_merge_fn;
1958 q->front_merge_fn = ll_front_merge_fn;
1959 q->merge_requests_fn = ll_merge_requests_fn;
1960 q->prep_rq_fn = NULL;
1961 q->unplug_fn = generic_unplug_device;
1962 q->queue_flags = (1 << QUEUE_FLAG_CLUSTER);
1963 q->queue_lock = lock;
1965 blk_queue_segment_boundary(q, 0xffffffff);
1967 blk_queue_make_request(q, __make_request);
1968 blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1970 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1971 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1976 if (!elevator_init(q, NULL)) {
1977 blk_queue_congestion_threshold(q);
1984 EXPORT_SYMBOL(blk_init_queue_node);
1986 int blk_get_queue(request_queue_t *q)
1988 if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1989 kobject_get(&q->kobj);
1996 EXPORT_SYMBOL(blk_get_queue);
1998 static inline void blk_free_request(request_queue_t *q, struct request *rq)
2000 if (rq->cmd_flags & REQ_ELVPRIV)
2001 elv_put_request(q, rq);
2002 mempool_free(rq, q->rq.rq_pool);
2005 static inline struct request *
2006 blk_alloc_request(request_queue_t *q, int rw, struct bio *bio,
2007 int priv, gfp_t gfp_mask)
2009 struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
2015 * first three bits are identical in rq->cmd_flags and bio->bi_rw,
2016 * see bio.h and blkdev.h
2021 if (unlikely(elv_set_request(q, rq, bio, gfp_mask))) {
2022 mempool_free(rq, q->rq.rq_pool);
2025 rq->cmd_flags |= REQ_ELVPRIV;
2032 * ioc_batching returns true if the ioc is a valid batching request and
2033 * should be given priority access to a request.
2035 static inline int ioc_batching(request_queue_t *q, struct io_context *ioc)
2041 * Make sure the process is able to allocate at least 1 request
2042 * even if the batch times out, otherwise we could theoretically
2045 return ioc->nr_batch_requests == q->nr_batching ||
2046 (ioc->nr_batch_requests > 0
2047 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
2051 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
2052 * will cause the process to be a "batcher" on all queues in the system. This
2053 * is the behaviour we want though - once it gets a wakeup it should be given
2056 static void ioc_set_batching(request_queue_t *q, struct io_context *ioc)
2058 if (!ioc || ioc_batching(q, ioc))
2061 ioc->nr_batch_requests = q->nr_batching;
2062 ioc->last_waited = jiffies;
2065 static void __freed_request(request_queue_t *q, int rw)
2067 struct request_list *rl = &q->rq;
2069 if (rl->count[rw] < queue_congestion_off_threshold(q))
2070 clear_queue_congested(q, rw);
2072 if (rl->count[rw] + 1 <= q->nr_requests) {
2073 if (waitqueue_active(&rl->wait[rw]))
2074 wake_up(&rl->wait[rw]);
2076 blk_clear_queue_full(q, rw);
2081 * A request has just been released. Account for it, update the full and
2082 * congestion status, wake up any waiters. Called under q->queue_lock.
2084 static void freed_request(request_queue_t *q, int rw, int priv)
2086 struct request_list *rl = &q->rq;
2092 __freed_request(q, rw);
2094 if (unlikely(rl->starved[rw ^ 1]))
2095 __freed_request(q, rw ^ 1);
2098 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2100 * Get a free request, queue_lock must be held.
2101 * Returns NULL on failure, with queue_lock held.
2102 * Returns !NULL on success, with queue_lock *not held*.
2104 static struct request *get_request(request_queue_t *q, int rw, struct bio *bio,
2107 struct request *rq = NULL;
2108 struct request_list *rl = &q->rq;
2109 struct io_context *ioc = NULL;
2110 int may_queue, priv;
2112 may_queue = elv_may_queue(q, rw, bio);
2113 if (may_queue == ELV_MQUEUE_NO)
2116 if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
2117 if (rl->count[rw]+1 >= q->nr_requests) {
2118 ioc = current_io_context(GFP_ATOMIC);
2120 * The queue will fill after this allocation, so set
2121 * it as full, and mark this process as "batching".
2122 * This process will be allowed to complete a batch of
2123 * requests, others will be blocked.
2125 if (!blk_queue_full(q, rw)) {
2126 ioc_set_batching(q, ioc);
2127 blk_set_queue_full(q, rw);
2129 if (may_queue != ELV_MQUEUE_MUST
2130 && !ioc_batching(q, ioc)) {
2132 * The queue is full and the allocating
2133 * process is not a "batcher", and not
2134 * exempted by the IO scheduler
2140 set_queue_congested(q, rw);
2144 * Only allow batching queuers to allocate up to 50% over the defined
2145 * limit of requests, otherwise we could have thousands of requests
2146 * allocated with any setting of ->nr_requests
2148 if (rl->count[rw] >= (3 * q->nr_requests / 2))
2152 rl->starved[rw] = 0;
2154 priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
2158 spin_unlock_irq(q->queue_lock);
2160 rq = blk_alloc_request(q, rw, bio, priv, gfp_mask);
2161 if (unlikely(!rq)) {
2163 * Allocation failed presumably due to memory. Undo anything
2164 * we might have messed up.
2166 * Allocating task should really be put onto the front of the
2167 * wait queue, but this is pretty rare.
2169 spin_lock_irq(q->queue_lock);
2170 freed_request(q, rw, priv);
2173 * in the very unlikely event that allocation failed and no
2174 * requests for this direction was pending, mark us starved
2175 * so that freeing of a request in the other direction will
2176 * notice us. another possible fix would be to split the
2177 * rq mempool into READ and WRITE
2180 if (unlikely(rl->count[rw] == 0))
2181 rl->starved[rw] = 1;
2187 * ioc may be NULL here, and ioc_batching will be false. That's
2188 * OK, if the queue is under the request limit then requests need
2189 * not count toward the nr_batch_requests limit. There will always
2190 * be some limit enforced by BLK_BATCH_TIME.
2192 if (ioc_batching(q, ioc))
2193 ioc->nr_batch_requests--;
2198 blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
2204 * No available requests for this queue, unplug the device and wait for some
2205 * requests to become available.
2207 * Called with q->queue_lock held, and returns with it unlocked.
2209 static struct request *get_request_wait(request_queue_t *q, int rw,
2214 rq = get_request(q, rw, bio, GFP_NOIO);
2217 struct request_list *rl = &q->rq;
2219 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2220 TASK_UNINTERRUPTIBLE);
2222 rq = get_request(q, rw, bio, GFP_NOIO);
2225 struct io_context *ioc;
2227 blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
2229 __generic_unplug_device(q);
2230 spin_unlock_irq(q->queue_lock);
2234 * After sleeping, we become a "batching" process and
2235 * will be able to allocate at least one request, and
2236 * up to a big batch of them for a small period time.
2237 * See ioc_batching, ioc_set_batching
2239 ioc = current_io_context(GFP_NOIO);
2240 ioc_set_batching(q, ioc);
2242 spin_lock_irq(q->queue_lock);
2244 finish_wait(&rl->wait[rw], &wait);
2250 struct request *blk_get_request(request_queue_t *q, int rw, gfp_t gfp_mask)
2254 BUG_ON(rw != READ && rw != WRITE);
2256 spin_lock_irq(q->queue_lock);
2257 if (gfp_mask & __GFP_WAIT) {
2258 rq = get_request_wait(q, rw, NULL);
2260 rq = get_request(q, rw, NULL, gfp_mask);
2262 spin_unlock_irq(q->queue_lock);
2264 /* q->queue_lock is unlocked at this point */
2268 EXPORT_SYMBOL(blk_get_request);
2271 * blk_requeue_request - put a request back on queue
2272 * @q: request queue where request should be inserted
2273 * @rq: request to be inserted
2276 * Drivers often keep queueing requests until the hardware cannot accept
2277 * more, when that condition happens we need to put the request back
2278 * on the queue. Must be called with queue lock held.
2280 void blk_requeue_request(request_queue_t *q, struct request *rq)
2282 blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
2284 if (blk_rq_tagged(rq))
2285 blk_queue_end_tag(q, rq);
2287 elv_requeue_request(q, rq);
2290 EXPORT_SYMBOL(blk_requeue_request);
2293 * blk_insert_request - insert a special request in to a request queue
2294 * @q: request queue where request should be inserted
2295 * @rq: request to be inserted
2296 * @at_head: insert request at head or tail of queue
2297 * @data: private data
2300 * Many block devices need to execute commands asynchronously, so they don't
2301 * block the whole kernel from preemption during request execution. This is
2302 * accomplished normally by inserting aritficial requests tagged as
2303 * REQ_SPECIAL in to the corresponding request queue, and letting them be
2304 * scheduled for actual execution by the request queue.
2306 * We have the option of inserting the head or the tail of the queue.
2307 * Typically we use the tail for new ioctls and so forth. We use the head
2308 * of the queue for things like a QUEUE_FULL message from a device, or a
2309 * host that is unable to accept a particular command.
2311 void blk_insert_request(request_queue_t *q, struct request *rq,
2312 int at_head, void *data)
2314 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2315 unsigned long flags;
2318 * tell I/O scheduler that this isn't a regular read/write (ie it
2319 * must not attempt merges on this) and that it acts as a soft
2322 rq->cmd_type = REQ_TYPE_SPECIAL;
2323 rq->cmd_flags |= REQ_SOFTBARRIER;
2327 spin_lock_irqsave(q->queue_lock, flags);
2330 * If command is tagged, release the tag
2332 if (blk_rq_tagged(rq))
2333 blk_queue_end_tag(q, rq);
2335 drive_stat_acct(rq, rq->nr_sectors, 1);
2336 __elv_add_request(q, rq, where, 0);
2338 if (blk_queue_plugged(q))
2339 __generic_unplug_device(q);
2342 spin_unlock_irqrestore(q->queue_lock, flags);
2345 EXPORT_SYMBOL(blk_insert_request);
2348 * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2349 * @q: request queue where request should be inserted
2350 * @rq: request structure to fill
2351 * @ubuf: the user buffer
2352 * @len: length of user data
2355 * Data will be mapped directly for zero copy io, if possible. Otherwise
2356 * a kernel bounce buffer is used.
2358 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2359 * still in process context.
2361 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2362 * before being submitted to the device, as pages mapped may be out of
2363 * reach. It's the callers responsibility to make sure this happens. The
2364 * original bio must be passed back in to blk_rq_unmap_user() for proper
2367 int blk_rq_map_user(request_queue_t *q, struct request *rq, void __user *ubuf,
2370 unsigned long uaddr;
2374 if (len > (q->max_hw_sectors << 9))
2379 reading = rq_data_dir(rq) == READ;
2382 * if alignment requirement is satisfied, map in user pages for
2383 * direct dma. else, set up kernel bounce buffers
2385 uaddr = (unsigned long) ubuf;
2386 if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2387 bio = bio_map_user(q, NULL, uaddr, len, reading);
2389 bio = bio_copy_user(q, uaddr, len, reading);
2392 rq->bio = rq->biotail = bio;
2393 blk_rq_bio_prep(q, rq, bio);
2395 rq->buffer = rq->data = NULL;
2401 * bio is the err-ptr
2403 return PTR_ERR(bio);
2406 EXPORT_SYMBOL(blk_rq_map_user);
2409 * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2410 * @q: request queue where request should be inserted
2411 * @rq: request to map data to
2412 * @iov: pointer to the iovec
2413 * @iov_count: number of elements in the iovec
2416 * Data will be mapped directly for zero copy io, if possible. Otherwise
2417 * a kernel bounce buffer is used.
2419 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2420 * still in process context.
2422 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2423 * before being submitted to the device, as pages mapped may be out of
2424 * reach. It's the callers responsibility to make sure this happens. The
2425 * original bio must be passed back in to blk_rq_unmap_user() for proper
2428 int blk_rq_map_user_iov(request_queue_t *q, struct request *rq,
2429 struct sg_iovec *iov, int iov_count)
2433 if (!iov || iov_count <= 0)
2436 /* we don't allow misaligned data like bio_map_user() does. If the
2437 * user is using sg, they're expected to know the alignment constraints
2438 * and respect them accordingly */
2439 bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2441 return PTR_ERR(bio);
2443 rq->bio = rq->biotail = bio;
2444 blk_rq_bio_prep(q, rq, bio);
2445 rq->buffer = rq->data = NULL;
2446 rq->data_len = bio->bi_size;
2450 EXPORT_SYMBOL(blk_rq_map_user_iov);
2453 * blk_rq_unmap_user - unmap a request with user data
2454 * @bio: bio to be unmapped
2455 * @ulen: length of user buffer
2458 * Unmap a bio previously mapped by blk_rq_map_user().
2460 int blk_rq_unmap_user(struct bio *bio, unsigned int ulen)
2465 if (bio_flagged(bio, BIO_USER_MAPPED))
2466 bio_unmap_user(bio);
2468 ret = bio_uncopy_user(bio);
2474 EXPORT_SYMBOL(blk_rq_unmap_user);
2477 * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2478 * @q: request queue where request should be inserted
2479 * @rq: request to fill
2480 * @kbuf: the kernel buffer
2481 * @len: length of user data
2482 * @gfp_mask: memory allocation flags
2484 int blk_rq_map_kern(request_queue_t *q, struct request *rq, void *kbuf,
2485 unsigned int len, gfp_t gfp_mask)
2489 if (len > (q->max_hw_sectors << 9))
2494 bio = bio_map_kern(q, kbuf, len, gfp_mask);
2496 return PTR_ERR(bio);
2498 if (rq_data_dir(rq) == WRITE)
2499 bio->bi_rw |= (1 << BIO_RW);
2501 rq->bio = rq->biotail = bio;
2502 blk_rq_bio_prep(q, rq, bio);
2504 rq->buffer = rq->data = NULL;
2509 EXPORT_SYMBOL(blk_rq_map_kern);
2512 * blk_execute_rq_nowait - insert a request into queue for execution
2513 * @q: queue to insert the request in
2514 * @bd_disk: matching gendisk
2515 * @rq: request to insert
2516 * @at_head: insert request at head or tail of queue
2517 * @done: I/O completion handler
2520 * Insert a fully prepared request at the back of the io scheduler queue
2521 * for execution. Don't wait for completion.
2523 void blk_execute_rq_nowait(request_queue_t *q, struct gendisk *bd_disk,
2524 struct request *rq, int at_head,
2527 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2529 rq->rq_disk = bd_disk;
2530 rq->cmd_flags |= REQ_NOMERGE;
2532 WARN_ON(irqs_disabled());
2533 spin_lock_irq(q->queue_lock);
2534 __elv_add_request(q, rq, where, 1);
2535 __generic_unplug_device(q);
2536 spin_unlock_irq(q->queue_lock);
2538 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
2541 * blk_execute_rq - insert a request into queue for execution
2542 * @q: queue to insert the request in
2543 * @bd_disk: matching gendisk
2544 * @rq: request to insert
2545 * @at_head: insert request at head or tail of queue
2548 * Insert a fully prepared request at the back of the io scheduler queue
2549 * for execution and wait for completion.
2551 int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
2552 struct request *rq, int at_head)
2554 DECLARE_COMPLETION_ONSTACK(wait);
2555 char sense[SCSI_SENSE_BUFFERSIZE];
2559 * we need an extra reference to the request, so we can look at
2560 * it after io completion
2565 memset(sense, 0, sizeof(sense));
2570 rq->waiting = &wait;
2571 blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
2572 wait_for_completion(&wait);
2581 EXPORT_SYMBOL(blk_execute_rq);
2584 * blkdev_issue_flush - queue a flush
2585 * @bdev: blockdev to issue flush for
2586 * @error_sector: error sector
2589 * Issue a flush for the block device in question. Caller can supply
2590 * room for storing the error offset in case of a flush error, if they
2591 * wish to. Caller must run wait_for_completion() on its own.
2593 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2597 if (bdev->bd_disk == NULL)
2600 q = bdev_get_queue(bdev);
2603 if (!q->issue_flush_fn)
2606 return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2609 EXPORT_SYMBOL(blkdev_issue_flush);
2611 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2613 int rw = rq_data_dir(rq);
2615 if (!blk_fs_request(rq) || !rq->rq_disk)
2619 __disk_stat_inc(rq->rq_disk, merges[rw]);
2621 disk_round_stats(rq->rq_disk);
2622 rq->rq_disk->in_flight++;
2627 * add-request adds a request to the linked list.
2628 * queue lock is held and interrupts disabled, as we muck with the
2629 * request queue list.
2631 static inline void add_request(request_queue_t * q, struct request * req)
2633 drive_stat_acct(req, req->nr_sectors, 1);
2636 q->activity_fn(q->activity_data, rq_data_dir(req));
2639 * elevator indicated where it wants this request to be
2640 * inserted at elevator_merge time
2642 __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2646 * disk_round_stats() - Round off the performance stats on a struct
2649 * The average IO queue length and utilisation statistics are maintained
2650 * by observing the current state of the queue length and the amount of
2651 * time it has been in this state for.
2653 * Normally, that accounting is done on IO completion, but that can result
2654 * in more than a second's worth of IO being accounted for within any one
2655 * second, leading to >100% utilisation. To deal with that, we call this
2656 * function to do a round-off before returning the results when reading
2657 * /proc/diskstats. This accounts immediately for all queue usage up to
2658 * the current jiffies and restarts the counters again.
2660 void disk_round_stats(struct gendisk *disk)
2662 unsigned long now = jiffies;
2664 if (now == disk->stamp)
2667 if (disk->in_flight) {
2668 __disk_stat_add(disk, time_in_queue,
2669 disk->in_flight * (now - disk->stamp));
2670 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2675 EXPORT_SYMBOL_GPL(disk_round_stats);
2678 * queue lock must be held
2680 void __blk_put_request(request_queue_t *q, struct request *req)
2682 struct request_list *rl = req->rl;
2686 if (unlikely(--req->ref_count))
2689 elv_completed_request(q, req);
2691 req->rq_status = RQ_INACTIVE;
2695 * Request may not have originated from ll_rw_blk. if not,
2696 * it didn't come out of our reserved rq pools
2699 int rw = rq_data_dir(req);
2700 int priv = req->cmd_flags & REQ_ELVPRIV;
2702 BUG_ON(!list_empty(&req->queuelist));
2704 blk_free_request(q, req);
2705 freed_request(q, rw, priv);
2709 EXPORT_SYMBOL_GPL(__blk_put_request);
2711 void blk_put_request(struct request *req)
2713 unsigned long flags;
2714 request_queue_t *q = req->q;
2717 * Gee, IDE calls in w/ NULL q. Fix IDE and remove the
2718 * following if (q) test.
2721 spin_lock_irqsave(q->queue_lock, flags);
2722 __blk_put_request(q, req);
2723 spin_unlock_irqrestore(q->queue_lock, flags);
2727 EXPORT_SYMBOL(blk_put_request);
2730 * blk_end_sync_rq - executes a completion event on a request
2731 * @rq: request to complete
2732 * @error: end io status of the request
2734 void blk_end_sync_rq(struct request *rq, int error)
2736 struct completion *waiting = rq->waiting;
2739 __blk_put_request(rq->q, rq);
2742 * complete last, if this is a stack request the process (and thus
2743 * the rq pointer) could be invalid right after this complete()
2747 EXPORT_SYMBOL(blk_end_sync_rq);
2750 * blk_congestion_wait - wait for a queue to become uncongested
2751 * @rw: READ or WRITE
2752 * @timeout: timeout in jiffies
2754 * Waits for up to @timeout jiffies for a queue (any queue) to exit congestion.
2755 * If no queues are congested then just wait for the next request to be
2758 long blk_congestion_wait(int rw, long timeout)
2762 wait_queue_head_t *wqh = &congestion_wqh[rw];
2764 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
2765 ret = io_schedule_timeout(timeout);
2766 finish_wait(wqh, &wait);
2770 EXPORT_SYMBOL(blk_congestion_wait);
2773 * blk_congestion_end - wake up sleepers on a congestion queue
2774 * @rw: READ or WRITE
2776 void blk_congestion_end(int rw)
2778 wait_queue_head_t *wqh = &congestion_wqh[rw];
2780 if (waitqueue_active(wqh))
2785 * Has to be called with the request spinlock acquired
2787 static int attempt_merge(request_queue_t *q, struct request *req,
2788 struct request *next)
2790 if (!rq_mergeable(req) || !rq_mergeable(next))
2796 if (req->sector + req->nr_sectors != next->sector)
2799 if (rq_data_dir(req) != rq_data_dir(next)
2800 || req->rq_disk != next->rq_disk
2801 || next->waiting || next->special)
2805 * If we are allowed to merge, then append bio list
2806 * from next to rq and release next. merge_requests_fn
2807 * will have updated segment counts, update sector
2810 if (!q->merge_requests_fn(q, req, next))
2814 * At this point we have either done a back merge
2815 * or front merge. We need the smaller start_time of
2816 * the merged requests to be the current request
2817 * for accounting purposes.
2819 if (time_after(req->start_time, next->start_time))
2820 req->start_time = next->start_time;
2822 req->biotail->bi_next = next->bio;
2823 req->biotail = next->biotail;
2825 req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2827 elv_merge_requests(q, req, next);
2830 disk_round_stats(req->rq_disk);
2831 req->rq_disk->in_flight--;
2834 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2836 __blk_put_request(q, next);
2840 static inline int attempt_back_merge(request_queue_t *q, struct request *rq)
2842 struct request *next = elv_latter_request(q, rq);
2845 return attempt_merge(q, rq, next);
2850 static inline int attempt_front_merge(request_queue_t *q, struct request *rq)
2852 struct request *prev = elv_former_request(q, rq);
2855 return attempt_merge(q, prev, rq);
2860 static void init_request_from_bio(struct request *req, struct bio *bio)
2862 req->cmd_type = REQ_TYPE_FS;
2865 * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2867 if (bio_rw_ahead(bio) || bio_failfast(bio))
2868 req->cmd_flags |= REQ_FAILFAST;
2871 * REQ_BARRIER implies no merging, but lets make it explicit
2873 if (unlikely(bio_barrier(bio)))
2874 req->cmd_flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2877 req->cmd_flags |= REQ_RW_SYNC;
2880 req->hard_sector = req->sector = bio->bi_sector;
2881 req->hard_nr_sectors = req->nr_sectors = bio_sectors(bio);
2882 req->current_nr_sectors = req->hard_cur_sectors = bio_cur_sectors(bio);
2883 req->nr_phys_segments = bio_phys_segments(req->q, bio);
2884 req->nr_hw_segments = bio_hw_segments(req->q, bio);
2885 req->buffer = bio_data(bio); /* see ->buffer comment above */
2886 req->waiting = NULL;
2887 req->bio = req->biotail = bio;
2888 req->ioprio = bio_prio(bio);
2889 req->rq_disk = bio->bi_bdev->bd_disk;
2890 req->start_time = jiffies;
2893 static int __make_request(request_queue_t *q, struct bio *bio)
2895 struct request *req;
2896 int el_ret, rw, nr_sectors, cur_nr_sectors, barrier, err, sync;
2897 unsigned short prio;
2900 sector = bio->bi_sector;
2901 nr_sectors = bio_sectors(bio);
2902 cur_nr_sectors = bio_cur_sectors(bio);
2903 prio = bio_prio(bio);
2905 rw = bio_data_dir(bio);
2906 sync = bio_sync(bio);
2909 * low level driver can indicate that it wants pages above a
2910 * certain limit bounced to low memory (ie for highmem, or even
2911 * ISA dma in theory)
2913 blk_queue_bounce(q, &bio);
2915 spin_lock_prefetch(q->queue_lock);
2917 barrier = bio_barrier(bio);
2918 if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
2923 spin_lock_irq(q->queue_lock);
2925 if (unlikely(barrier) || elv_queue_empty(q))
2928 el_ret = elv_merge(q, &req, bio);
2930 case ELEVATOR_BACK_MERGE:
2931 BUG_ON(!rq_mergeable(req));
2933 if (!q->back_merge_fn(q, req, bio))
2936 blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
2938 req->biotail->bi_next = bio;
2940 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2941 req->ioprio = ioprio_best(req->ioprio, prio);
2942 drive_stat_acct(req, nr_sectors, 0);
2943 if (!attempt_back_merge(q, req))
2944 elv_merged_request(q, req);
2947 case ELEVATOR_FRONT_MERGE:
2948 BUG_ON(!rq_mergeable(req));
2950 if (!q->front_merge_fn(q, req, bio))
2953 blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
2955 bio->bi_next = req->bio;
2959 * may not be valid. if the low level driver said
2960 * it didn't need a bounce buffer then it better
2961 * not touch req->buffer either...
2963 req->buffer = bio_data(bio);
2964 req->current_nr_sectors = cur_nr_sectors;
2965 req->hard_cur_sectors = cur_nr_sectors;
2966 req->sector = req->hard_sector = sector;
2967 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2968 req->ioprio = ioprio_best(req->ioprio, prio);
2969 drive_stat_acct(req, nr_sectors, 0);
2970 if (!attempt_front_merge(q, req))
2971 elv_merged_request(q, req);
2974 /* ELV_NO_MERGE: elevator says don't/can't merge. */
2981 * Grab a free request. This is might sleep but can not fail.
2982 * Returns with the queue unlocked.
2984 req = get_request_wait(q, rw, bio);
2987 * After dropping the lock and possibly sleeping here, our request
2988 * may now be mergeable after it had proven unmergeable (above).
2989 * We don't worry about that case for efficiency. It won't happen
2990 * often, and the elevators are able to handle it.
2992 init_request_from_bio(req, bio);
2994 spin_lock_irq(q->queue_lock);
2995 if (elv_queue_empty(q))
2997 add_request(q, req);
3000 __generic_unplug_device(q);
3002 spin_unlock_irq(q->queue_lock);
3006 bio_endio(bio, nr_sectors << 9, err);
3011 * If bio->bi_dev is a partition, remap the location
3013 static inline void blk_partition_remap(struct bio *bio)
3015 struct block_device *bdev = bio->bi_bdev;
3017 if (bdev != bdev->bd_contains) {
3018 struct hd_struct *p = bdev->bd_part;
3019 const int rw = bio_data_dir(bio);
3021 p->sectors[rw] += bio_sectors(bio);
3024 bio->bi_sector += p->start_sect;
3025 bio->bi_bdev = bdev->bd_contains;
3029 static void handle_bad_sector(struct bio *bio)
3031 char b[BDEVNAME_SIZE];
3033 printk(KERN_INFO "attempt to access beyond end of device\n");
3034 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
3035 bdevname(bio->bi_bdev, b),
3037 (unsigned long long)bio->bi_sector + bio_sectors(bio),
3038 (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
3040 set_bit(BIO_EOF, &bio->bi_flags);
3044 * generic_make_request: hand a buffer to its device driver for I/O
3045 * @bio: The bio describing the location in memory and on the device.
3047 * generic_make_request() is used to make I/O requests of block
3048 * devices. It is passed a &struct bio, which describes the I/O that needs
3051 * generic_make_request() does not return any status. The
3052 * success/failure status of the request, along with notification of
3053 * completion, is delivered asynchronously through the bio->bi_end_io
3054 * function described (one day) else where.
3056 * The caller of generic_make_request must make sure that bi_io_vec
3057 * are set to describe the memory buffer, and that bi_dev and bi_sector are
3058 * set to describe the device address, and the
3059 * bi_end_io and optionally bi_private are set to describe how
3060 * completion notification should be signaled.
3062 * generic_make_request and the drivers it calls may use bi_next if this
3063 * bio happens to be merged with someone else, and may change bi_dev and
3064 * bi_sector for remaps as it sees fit. So the values of these fields
3065 * should NOT be depended on after the call to generic_make_request.
3067 void generic_make_request(struct bio *bio)
3071 int ret, nr_sectors = bio_sectors(bio);
3075 /* Test device or partition size, when known. */
3076 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3078 sector_t sector = bio->bi_sector;
3080 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
3082 * This may well happen - the kernel calls bread()
3083 * without checking the size of the device, e.g., when
3084 * mounting a device.
3086 handle_bad_sector(bio);
3092 * Resolve the mapping until finished. (drivers are
3093 * still free to implement/resolve their own stacking
3094 * by explicitly returning 0)
3096 * NOTE: we don't repeat the blk_size check for each new device.
3097 * Stacking drivers are expected to know what they are doing.
3102 char b[BDEVNAME_SIZE];
3104 q = bdev_get_queue(bio->bi_bdev);
3107 "generic_make_request: Trying to access "
3108 "nonexistent block-device %s (%Lu)\n",
3109 bdevname(bio->bi_bdev, b),
3110 (long long) bio->bi_sector);
3112 bio_endio(bio, bio->bi_size, -EIO);
3116 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
3117 printk("bio too big device %s (%u > %u)\n",
3118 bdevname(bio->bi_bdev, b),
3124 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
3128 * If this device has partitions, remap block n
3129 * of partition p to block n+start(p) of the disk.
3131 blk_partition_remap(bio);
3133 if (maxsector != -1)
3134 blk_add_trace_remap(q, bio, old_dev, bio->bi_sector,
3137 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
3139 maxsector = bio->bi_sector;
3140 old_dev = bio->bi_bdev->bd_dev;
3142 ret = q->make_request_fn(q, bio);
3146 EXPORT_SYMBOL(generic_make_request);
3149 * submit_bio: submit a bio to the block device layer for I/O
3150 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3151 * @bio: The &struct bio which describes the I/O
3153 * submit_bio() is very similar in purpose to generic_make_request(), and
3154 * uses that function to do most of the work. Both are fairly rough
3155 * interfaces, @bio must be presetup and ready for I/O.
3158 void submit_bio(int rw, struct bio *bio)
3160 int count = bio_sectors(bio);
3162 BIO_BUG_ON(!bio->bi_size);
3163 BIO_BUG_ON(!bio->bi_io_vec);
3166 count_vm_events(PGPGOUT, count);
3168 count_vm_events(PGPGIN, count);
3170 if (unlikely(block_dump)) {
3171 char b[BDEVNAME_SIZE];
3172 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
3173 current->comm, current->pid,
3174 (rw & WRITE) ? "WRITE" : "READ",
3175 (unsigned long long)bio->bi_sector,
3176 bdevname(bio->bi_bdev,b));
3179 generic_make_request(bio);
3182 EXPORT_SYMBOL(submit_bio);
3184 static void blk_recalc_rq_segments(struct request *rq)
3186 struct bio *bio, *prevbio = NULL;
3187 int nr_phys_segs, nr_hw_segs;
3188 unsigned int phys_size, hw_size;
3189 request_queue_t *q = rq->q;
3194 phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
3195 rq_for_each_bio(bio, rq) {
3196 /* Force bio hw/phys segs to be recalculated. */
3197 bio->bi_flags &= ~(1 << BIO_SEG_VALID);
3199 nr_phys_segs += bio_phys_segments(q, bio);
3200 nr_hw_segs += bio_hw_segments(q, bio);
3202 int pseg = phys_size + prevbio->bi_size + bio->bi_size;
3203 int hseg = hw_size + prevbio->bi_size + bio->bi_size;
3205 if (blk_phys_contig_segment(q, prevbio, bio) &&
3206 pseg <= q->max_segment_size) {
3208 phys_size += prevbio->bi_size + bio->bi_size;
3212 if (blk_hw_contig_segment(q, prevbio, bio) &&
3213 hseg <= q->max_segment_size) {
3215 hw_size += prevbio->bi_size + bio->bi_size;
3222 rq->nr_phys_segments = nr_phys_segs;
3223 rq->nr_hw_segments = nr_hw_segs;
3226 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
3228 if (blk_fs_request(rq)) {
3229 rq->hard_sector += nsect;
3230 rq->hard_nr_sectors -= nsect;
3233 * Move the I/O submission pointers ahead if required.
3235 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3236 (rq->sector <= rq->hard_sector)) {
3237 rq->sector = rq->hard_sector;
3238 rq->nr_sectors = rq->hard_nr_sectors;
3239 rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3240 rq->current_nr_sectors = rq->hard_cur_sectors;
3241 rq->buffer = bio_data(rq->bio);
3245 * if total number of sectors is less than the first segment
3246 * size, something has gone terribly wrong
3248 if (rq->nr_sectors < rq->current_nr_sectors) {
3249 printk("blk: request botched\n");
3250 rq->nr_sectors = rq->current_nr_sectors;
3255 static int __end_that_request_first(struct request *req, int uptodate,
3258 int total_bytes, bio_nbytes, error, next_idx = 0;
3261 blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
3264 * extend uptodate bool to allow < 0 value to be direct io error
3267 if (end_io_error(uptodate))
3268 error = !uptodate ? -EIO : uptodate;
3271 * for a REQ_BLOCK_PC request, we want to carry any eventual
3272 * sense key with us all the way through
3274 if (!blk_pc_request(req))
3278 if (blk_fs_request(req) && !(req->cmd_flags & REQ_QUIET))
3279 printk("end_request: I/O error, dev %s, sector %llu\n",
3280 req->rq_disk ? req->rq_disk->disk_name : "?",
3281 (unsigned long long)req->sector);
3284 if (blk_fs_request(req) && req->rq_disk) {
3285 const int rw = rq_data_dir(req);
3287 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
3290 total_bytes = bio_nbytes = 0;
3291 while ((bio = req->bio) != NULL) {
3294 if (nr_bytes >= bio->bi_size) {
3295 req->bio = bio->bi_next;
3296 nbytes = bio->bi_size;
3297 if (!ordered_bio_endio(req, bio, nbytes, error))
3298 bio_endio(bio, nbytes, error);
3302 int idx = bio->bi_idx + next_idx;
3304 if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3305 blk_dump_rq_flags(req, "__end_that");
3306 printk("%s: bio idx %d >= vcnt %d\n",
3308 bio->bi_idx, bio->bi_vcnt);
3312 nbytes = bio_iovec_idx(bio, idx)->bv_len;
3313 BIO_BUG_ON(nbytes > bio->bi_size);
3316 * not a complete bvec done
3318 if (unlikely(nbytes > nr_bytes)) {
3319 bio_nbytes += nr_bytes;
3320 total_bytes += nr_bytes;
3325 * advance to the next vector
3328 bio_nbytes += nbytes;
3331 total_bytes += nbytes;
3334 if ((bio = req->bio)) {
3336 * end more in this run, or just return 'not-done'
3338 if (unlikely(nr_bytes <= 0))
3350 * if the request wasn't completed, update state
3353 if (!ordered_bio_endio(req, bio, bio_nbytes, error))
3354 bio_endio(bio, bio_nbytes, error);
3355 bio->bi_idx += next_idx;
3356 bio_iovec(bio)->bv_offset += nr_bytes;
3357 bio_iovec(bio)->bv_len -= nr_bytes;
3360 blk_recalc_rq_sectors(req, total_bytes >> 9);
3361 blk_recalc_rq_segments(req);
3366 * end_that_request_first - end I/O on a request
3367 * @req: the request being processed
3368 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3369 * @nr_sectors: number of sectors to end I/O on
3372 * Ends I/O on a number of sectors attached to @req, and sets it up
3373 * for the next range of segments (if any) in the cluster.
3376 * 0 - we are done with this request, call end_that_request_last()
3377 * 1 - still buffers pending for this request
3379 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3381 return __end_that_request_first(req, uptodate, nr_sectors << 9);
3384 EXPORT_SYMBOL(end_that_request_first);
3387 * end_that_request_chunk - end I/O on a request
3388 * @req: the request being processed
3389 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3390 * @nr_bytes: number of bytes to complete
3393 * Ends I/O on a number of bytes attached to @req, and sets it up
3394 * for the next range of segments (if any). Like end_that_request_first(),
3395 * but deals with bytes instead of sectors.
3398 * 0 - we are done with this request, call end_that_request_last()
3399 * 1 - still buffers pending for this request
3401 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3403 return __end_that_request_first(req, uptodate, nr_bytes);
3406 EXPORT_SYMBOL(end_that_request_chunk);
3409 * splice the completion data to a local structure and hand off to
3410 * process_completion_queue() to complete the requests
3412 static void blk_done_softirq(struct softirq_action *h)
3414 struct list_head *cpu_list, local_list;
3416 local_irq_disable();
3417 cpu_list = &__get_cpu_var(blk_cpu_done);
3418 list_replace_init(cpu_list, &local_list);
3421 while (!list_empty(&local_list)) {
3422 struct request *rq = list_entry(local_list.next, struct request, donelist);
3424 list_del_init(&rq->donelist);
3425 rq->q->softirq_done_fn(rq);
3429 #ifdef CONFIG_HOTPLUG_CPU
3431 static int blk_cpu_notify(struct notifier_block *self, unsigned long action,
3435 * If a CPU goes away, splice its entries to the current CPU
3436 * and trigger a run of the softirq
3438 if (action == CPU_DEAD) {
3439 int cpu = (unsigned long) hcpu;
3441 local_irq_disable();
3442 list_splice_init(&per_cpu(blk_cpu_done, cpu),
3443 &__get_cpu_var(blk_cpu_done));
3444 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3452 static struct notifier_block __devinitdata blk_cpu_notifier = {
3453 .notifier_call = blk_cpu_notify,
3456 #endif /* CONFIG_HOTPLUG_CPU */
3459 * blk_complete_request - end I/O on a request
3460 * @req: the request being processed
3463 * Ends all I/O on a request. It does not handle partial completions,
3464 * unless the driver actually implements this in its completion callback
3465 * through requeueing. Theh actual completion happens out-of-order,
3466 * through a softirq handler. The user must have registered a completion
3467 * callback through blk_queue_softirq_done().
3470 void blk_complete_request(struct request *req)
3472 struct list_head *cpu_list;
3473 unsigned long flags;
3475 BUG_ON(!req->q->softirq_done_fn);
3477 local_irq_save(flags);
3479 cpu_list = &__get_cpu_var(blk_cpu_done);
3480 list_add_tail(&req->donelist, cpu_list);
3481 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3483 local_irq_restore(flags);
3486 EXPORT_SYMBOL(blk_complete_request);
3489 * queue lock must be held
3491 void end_that_request_last(struct request *req, int uptodate)
3493 struct gendisk *disk = req->rq_disk;
3497 * extend uptodate bool to allow < 0 value to be direct io error
3500 if (end_io_error(uptodate))
3501 error = !uptodate ? -EIO : uptodate;
3503 if (unlikely(laptop_mode) && blk_fs_request(req))
3504 laptop_io_completion();
3507 * Account IO completion. bar_rq isn't accounted as a normal
3508 * IO on queueing nor completion. Accounting the containing
3509 * request is enough.
3511 if (disk && blk_fs_request(req) && req != &req->q->bar_rq) {
3512 unsigned long duration = jiffies - req->start_time;
3513 const int rw = rq_data_dir(req);
3515 __disk_stat_inc(disk, ios[rw]);
3516 __disk_stat_add(disk, ticks[rw], duration);
3517 disk_round_stats(disk);
3521 req->end_io(req, error);
3523 __blk_put_request(req->q, req);
3526 EXPORT_SYMBOL(end_that_request_last);
3528 void end_request(struct request *req, int uptodate)
3530 if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3531 add_disk_randomness(req->rq_disk);
3532 blkdev_dequeue_request(req);
3533 end_that_request_last(req, uptodate);
3537 EXPORT_SYMBOL(end_request);
3539 void blk_rq_bio_prep(request_queue_t *q, struct request *rq, struct bio *bio)
3541 /* first two bits are identical in rq->cmd_flags and bio->bi_rw */
3542 rq->cmd_flags |= (bio->bi_rw & 3);
3544 rq->nr_phys_segments = bio_phys_segments(q, bio);
3545 rq->nr_hw_segments = bio_hw_segments(q, bio);
3546 rq->current_nr_sectors = bio_cur_sectors(bio);
3547 rq->hard_cur_sectors = rq->current_nr_sectors;
3548 rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3549 rq->buffer = bio_data(bio);
3551 rq->bio = rq->biotail = bio;
3554 EXPORT_SYMBOL(blk_rq_bio_prep);
3556 int kblockd_schedule_work(struct work_struct *work)
3558 return queue_work(kblockd_workqueue, work);
3561 EXPORT_SYMBOL(kblockd_schedule_work);
3563 void kblockd_flush(void)
3565 flush_workqueue(kblockd_workqueue);
3567 EXPORT_SYMBOL(kblockd_flush);
3569 int __init blk_dev_init(void)
3573 kblockd_workqueue = create_workqueue("kblockd");
3574 if (!kblockd_workqueue)
3575 panic("Failed to create kblockd\n");
3577 request_cachep = kmem_cache_create("blkdev_requests",
3578 sizeof(struct request), 0, SLAB_PANIC, NULL, NULL);
3580 requestq_cachep = kmem_cache_create("blkdev_queue",
3581 sizeof(request_queue_t), 0, SLAB_PANIC, NULL, NULL);
3583 iocontext_cachep = kmem_cache_create("blkdev_ioc",
3584 sizeof(struct io_context), 0, SLAB_PANIC, NULL, NULL);
3586 for_each_possible_cpu(i)
3587 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3589 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
3590 register_hotcpu_notifier(&blk_cpu_notifier);
3592 blk_max_low_pfn = max_low_pfn;
3593 blk_max_pfn = max_pfn;
3599 * IO Context helper functions
3601 void put_io_context(struct io_context *ioc)
3606 BUG_ON(atomic_read(&ioc->refcount) == 0);
3608 if (atomic_dec_and_test(&ioc->refcount)) {
3609 struct cfq_io_context *cic;
3612 if (ioc->aic && ioc->aic->dtor)
3613 ioc->aic->dtor(ioc->aic);
3614 if (ioc->cic_root.rb_node != NULL) {
3615 struct rb_node *n = rb_first(&ioc->cic_root);
3617 cic = rb_entry(n, struct cfq_io_context, rb_node);
3622 kmem_cache_free(iocontext_cachep, ioc);
3625 EXPORT_SYMBOL(put_io_context);
3627 /* Called by the exitting task */
3628 void exit_io_context(void)
3630 unsigned long flags;
3631 struct io_context *ioc;
3632 struct cfq_io_context *cic;
3634 local_irq_save(flags);
3636 ioc = current->io_context;
3637 current->io_context = NULL;
3639 task_unlock(current);
3640 local_irq_restore(flags);
3642 if (ioc->aic && ioc->aic->exit)
3643 ioc->aic->exit(ioc->aic);
3644 if (ioc->cic_root.rb_node != NULL) {
3645 cic = rb_entry(rb_first(&ioc->cic_root), struct cfq_io_context, rb_node);
3649 put_io_context(ioc);
3653 * If the current task has no IO context then create one and initialise it.
3654 * Otherwise, return its existing IO context.
3656 * This returned IO context doesn't have a specifically elevated refcount,
3657 * but since the current task itself holds a reference, the context can be
3658 * used in general code, so long as it stays within `current` context.
3660 struct io_context *current_io_context(gfp_t gfp_flags)
3662 struct task_struct *tsk = current;
3663 struct io_context *ret;
3665 ret = tsk->io_context;
3669 ret = kmem_cache_alloc(iocontext_cachep, gfp_flags);
3671 atomic_set(&ret->refcount, 1);
3672 ret->task = current;
3673 ret->set_ioprio = NULL;
3674 ret->last_waited = jiffies; /* doesn't matter... */
3675 ret->nr_batch_requests = 0; /* because this is 0 */
3677 ret->cic_root.rb_node = NULL;
3678 /* make sure set_task_ioprio() sees the settings above */
3680 tsk->io_context = ret;
3685 EXPORT_SYMBOL(current_io_context);
3688 * If the current task has no IO context then create one and initialise it.
3689 * If it does have a context, take a ref on it.
3691 * This is always called in the context of the task which submitted the I/O.
3693 struct io_context *get_io_context(gfp_t gfp_flags)
3695 struct io_context *ret;
3696 ret = current_io_context(gfp_flags);
3698 atomic_inc(&ret->refcount);
3701 EXPORT_SYMBOL(get_io_context);
3703 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3705 struct io_context *src = *psrc;
3706 struct io_context *dst = *pdst;
3709 BUG_ON(atomic_read(&src->refcount) == 0);
3710 atomic_inc(&src->refcount);
3711 put_io_context(dst);
3715 EXPORT_SYMBOL(copy_io_context);
3717 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3719 struct io_context *temp;
3724 EXPORT_SYMBOL(swap_io_context);
3729 struct queue_sysfs_entry {
3730 struct attribute attr;
3731 ssize_t (*show)(struct request_queue *, char *);
3732 ssize_t (*store)(struct request_queue *, const char *, size_t);
3736 queue_var_show(unsigned int var, char *page)
3738 return sprintf(page, "%d\n", var);
3742 queue_var_store(unsigned long *var, const char *page, size_t count)
3744 char *p = (char *) page;
3746 *var = simple_strtoul(p, &p, 10);
3750 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3752 return queue_var_show(q->nr_requests, (page));
3756 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3758 struct request_list *rl = &q->rq;
3760 int ret = queue_var_store(&nr, page, count);
3761 if (nr < BLKDEV_MIN_RQ)
3764 spin_lock_irq(q->queue_lock);
3765 q->nr_requests = nr;
3766 blk_queue_congestion_threshold(q);
3768 if (rl->count[READ] >= queue_congestion_on_threshold(q))
3769 set_queue_congested(q, READ);
3770 else if (rl->count[READ] < queue_congestion_off_threshold(q))
3771 clear_queue_congested(q, READ);
3773 if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3774 set_queue_congested(q, WRITE);
3775 else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3776 clear_queue_congested(q, WRITE);
3778 if (rl->count[READ] >= q->nr_requests) {
3779 blk_set_queue_full(q, READ);
3780 } else if (rl->count[READ]+1 <= q->nr_requests) {
3781 blk_clear_queue_full(q, READ);
3782 wake_up(&rl->wait[READ]);
3785 if (rl->count[WRITE] >= q->nr_requests) {
3786 blk_set_queue_full(q, WRITE);
3787 } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3788 blk_clear_queue_full(q, WRITE);
3789 wake_up(&rl->wait[WRITE]);
3791 spin_unlock_irq(q->queue_lock);
3795 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3797 int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3799 return queue_var_show(ra_kb, (page));
3803 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3805 unsigned long ra_kb;
3806 ssize_t ret = queue_var_store(&ra_kb, page, count);
3808 spin_lock_irq(q->queue_lock);
3809 if (ra_kb > (q->max_sectors >> 1))
3810 ra_kb = (q->max_sectors >> 1);
3812 q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3813 spin_unlock_irq(q->queue_lock);
3818 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3820 int max_sectors_kb = q->max_sectors >> 1;
3822 return queue_var_show(max_sectors_kb, (page));
3826 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3828 unsigned long max_sectors_kb,
3829 max_hw_sectors_kb = q->max_hw_sectors >> 1,
3830 page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3831 ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3834 if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3837 * Take the queue lock to update the readahead and max_sectors
3838 * values synchronously:
3840 spin_lock_irq(q->queue_lock);
3842 * Trim readahead window as well, if necessary:
3844 ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3845 if (ra_kb > max_sectors_kb)
3846 q->backing_dev_info.ra_pages =
3847 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3849 q->max_sectors = max_sectors_kb << 1;
3850 spin_unlock_irq(q->queue_lock);
3855 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3857 int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3859 return queue_var_show(max_hw_sectors_kb, (page));
3863 static struct queue_sysfs_entry queue_requests_entry = {
3864 .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3865 .show = queue_requests_show,
3866 .store = queue_requests_store,
3869 static struct queue_sysfs_entry queue_ra_entry = {
3870 .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3871 .show = queue_ra_show,
3872 .store = queue_ra_store,
3875 static struct queue_sysfs_entry queue_max_sectors_entry = {
3876 .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3877 .show = queue_max_sectors_show,
3878 .store = queue_max_sectors_store,
3881 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3882 .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
3883 .show = queue_max_hw_sectors_show,
3886 static struct queue_sysfs_entry queue_iosched_entry = {
3887 .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
3888 .show = elv_iosched_show,
3889 .store = elv_iosched_store,
3892 static struct attribute *default_attrs[] = {
3893 &queue_requests_entry.attr,
3894 &queue_ra_entry.attr,
3895 &queue_max_hw_sectors_entry.attr,
3896 &queue_max_sectors_entry.attr,
3897 &queue_iosched_entry.attr,
3901 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
3904 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
3906 struct queue_sysfs_entry *entry = to_queue(attr);
3907 request_queue_t *q = container_of(kobj, struct request_queue, kobj);
3912 mutex_lock(&q->sysfs_lock);
3913 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
3914 mutex_unlock(&q->sysfs_lock);
3917 res = entry->show(q, page);
3918 mutex_unlock(&q->sysfs_lock);
3923 queue_attr_store(struct kobject *kobj, struct attribute *attr,
3924 const char *page, size_t length)
3926 struct queue_sysfs_entry *entry = to_queue(attr);
3927 request_queue_t *q = container_of(kobj, struct request_queue, kobj);
3933 mutex_lock(&q->sysfs_lock);
3934 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
3935 mutex_unlock(&q->sysfs_lock);
3938 res = entry->store(q, page, length);
3939 mutex_unlock(&q->sysfs_lock);
3943 static struct sysfs_ops queue_sysfs_ops = {
3944 .show = queue_attr_show,
3945 .store = queue_attr_store,
3948 static struct kobj_type queue_ktype = {
3949 .sysfs_ops = &queue_sysfs_ops,
3950 .default_attrs = default_attrs,
3951 .release = blk_release_queue,
3954 int blk_register_queue(struct gendisk *disk)
3958 request_queue_t *q = disk->queue;
3960 if (!q || !q->request_fn)
3963 q->kobj.parent = kobject_get(&disk->kobj);
3965 ret = kobject_add(&q->kobj);
3969 kobject_uevent(&q->kobj, KOBJ_ADD);
3971 ret = elv_register_queue(q);
3973 kobject_uevent(&q->kobj, KOBJ_REMOVE);
3974 kobject_del(&q->kobj);
3981 void blk_unregister_queue(struct gendisk *disk)
3983 request_queue_t *q = disk->queue;
3985 if (q && q->request_fn) {
3986 elv_unregister_queue(q);
3988 kobject_uevent(&q->kobj, KOBJ_REMOVE);
3989 kobject_del(&q->kobj);
3990 kobject_put(&disk->kobj);