2 * Copyright (C) 1991, 1992 Linus Torvalds
3 * Copyright (C) 1994, Karl Keyte: Added support for disk statistics
4 * Elevator latency, (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
5 * Queue request tables / lock, selectable elevator, Jens Axboe <axboe@suse.de>
6 * kernel-doc documentation started by NeilBrown <neilb@cse.unsw.edu.au> - July2000
7 * bio rewrite, highmem i/o, etc, Jens Axboe <axboe@suse.de> - may 2001
11 * This handles all read/write requests to block devices
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/backing-dev.h>
16 #include <linux/bio.h>
17 #include <linux/blkdev.h>
18 #include <linux/highmem.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/string.h>
22 #include <linux/init.h>
23 #include <linux/bootmem.h> /* for max_pfn/max_low_pfn */
24 #include <linux/completion.h>
25 #include <linux/slab.h>
26 #include <linux/swap.h>
27 #include <linux/writeback.h>
28 #include <linux/task_io_accounting_ops.h>
29 #include <linux/interrupt.h>
30 #include <linux/cpu.h>
31 #include <linux/blktrace_api.h>
32 #include <linux/fault-inject.h>
37 #include <scsi/scsi_cmnd.h>
39 static void blk_unplug_work(struct work_struct *work);
40 static void blk_unplug_timeout(unsigned long data);
41 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io);
42 static void init_request_from_bio(struct request *req, struct bio *bio);
43 static int __make_request(struct request_queue *q, struct bio *bio);
44 static struct io_context *current_io_context(gfp_t gfp_flags, int node);
45 static void blk_recalc_rq_segments(struct request *rq);
46 static void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
50 * For the allocated request tables
52 static struct kmem_cache *request_cachep;
55 * For queue allocation
57 static struct kmem_cache *requestq_cachep;
60 * For io context allocations
62 static struct kmem_cache *iocontext_cachep;
65 * Controlling structure to kblockd
67 static struct workqueue_struct *kblockd_workqueue;
69 unsigned long blk_max_low_pfn, blk_max_pfn;
71 EXPORT_SYMBOL(blk_max_low_pfn);
72 EXPORT_SYMBOL(blk_max_pfn);
74 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
76 /* Amount of time in which a process may batch requests */
77 #define BLK_BATCH_TIME (HZ/50UL)
79 /* Number of requests a "batching" process may submit */
80 #define BLK_BATCH_REQ 32
83 * Return the threshold (number of used requests) at which the queue is
84 * considered to be congested. It include a little hysteresis to keep the
85 * context switch rate down.
87 static inline int queue_congestion_on_threshold(struct request_queue *q)
89 return q->nr_congestion_on;
93 * The threshold at which a queue is considered to be uncongested
95 static inline int queue_congestion_off_threshold(struct request_queue *q)
97 return q->nr_congestion_off;
100 static void blk_queue_congestion_threshold(struct request_queue *q)
104 nr = q->nr_requests - (q->nr_requests / 8) + 1;
105 if (nr > q->nr_requests)
107 q->nr_congestion_on = nr;
109 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
112 q->nr_congestion_off = nr;
116 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
119 * Locates the passed device's request queue and returns the address of its
122 * Will return NULL if the request queue cannot be located.
124 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
126 struct backing_dev_info *ret = NULL;
127 struct request_queue *q = bdev_get_queue(bdev);
130 ret = &q->backing_dev_info;
133 EXPORT_SYMBOL(blk_get_backing_dev_info);
136 * blk_queue_prep_rq - set a prepare_request function for queue
138 * @pfn: prepare_request function
140 * It's possible for a queue to register a prepare_request callback which
141 * is invoked before the request is handed to the request_fn. The goal of
142 * the function is to prepare a request for I/O, it can be used to build a
143 * cdb from the request data for instance.
146 void blk_queue_prep_rq(struct request_queue *q, prep_rq_fn *pfn)
151 EXPORT_SYMBOL(blk_queue_prep_rq);
154 * blk_queue_merge_bvec - set a merge_bvec function for queue
156 * @mbfn: merge_bvec_fn
158 * Usually queues have static limitations on the max sectors or segments that
159 * we can put in a request. Stacking drivers may have some settings that
160 * are dynamic, and thus we have to query the queue whether it is ok to
161 * add a new bio_vec to a bio at a given offset or not. If the block device
162 * has such limitations, it needs to register a merge_bvec_fn to control
163 * the size of bio's sent to it. Note that a block device *must* allow a
164 * single page to be added to an empty bio. The block device driver may want
165 * to use the bio_split() function to deal with these bio's. By default
166 * no merge_bvec_fn is defined for a queue, and only the fixed limits are
169 void blk_queue_merge_bvec(struct request_queue *q, merge_bvec_fn *mbfn)
171 q->merge_bvec_fn = mbfn;
174 EXPORT_SYMBOL(blk_queue_merge_bvec);
176 void blk_queue_softirq_done(struct request_queue *q, softirq_done_fn *fn)
178 q->softirq_done_fn = fn;
181 EXPORT_SYMBOL(blk_queue_softirq_done);
184 * blk_queue_make_request - define an alternate make_request function for a device
185 * @q: the request queue for the device to be affected
186 * @mfn: the alternate make_request function
189 * The normal way for &struct bios to be passed to a device
190 * driver is for them to be collected into requests on a request
191 * queue, and then to allow the device driver to select requests
192 * off that queue when it is ready. This works well for many block
193 * devices. However some block devices (typically virtual devices
194 * such as md or lvm) do not benefit from the processing on the
195 * request queue, and are served best by having the requests passed
196 * directly to them. This can be achieved by providing a function
197 * to blk_queue_make_request().
200 * The driver that does this *must* be able to deal appropriately
201 * with buffers in "highmemory". This can be accomplished by either calling
202 * __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
203 * blk_queue_bounce() to create a buffer in normal memory.
205 void blk_queue_make_request(struct request_queue * q, make_request_fn * mfn)
210 q->nr_requests = BLKDEV_MAX_RQ;
211 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
212 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
213 q->make_request_fn = mfn;
214 q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
215 q->backing_dev_info.state = 0;
216 q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
217 blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
218 blk_queue_hardsect_size(q, 512);
219 blk_queue_dma_alignment(q, 511);
220 blk_queue_congestion_threshold(q);
221 q->nr_batching = BLK_BATCH_REQ;
223 q->unplug_thresh = 4; /* hmm */
224 q->unplug_delay = (3 * HZ) / 1000; /* 3 milliseconds */
225 if (q->unplug_delay == 0)
228 INIT_WORK(&q->unplug_work, blk_unplug_work);
230 q->unplug_timer.function = blk_unplug_timeout;
231 q->unplug_timer.data = (unsigned long)q;
234 * by default assume old behaviour and bounce for any highmem page
236 blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
239 EXPORT_SYMBOL(blk_queue_make_request);
241 static void rq_init(struct request_queue *q, struct request *rq)
243 INIT_LIST_HEAD(&rq->queuelist);
244 INIT_LIST_HEAD(&rq->donelist);
247 rq->bio = rq->biotail = NULL;
248 INIT_HLIST_NODE(&rq->hash);
249 RB_CLEAR_NODE(&rq->rb_node);
257 rq->nr_phys_segments = 0;
260 rq->end_io_data = NULL;
261 rq->completion_data = NULL;
266 * blk_queue_ordered - does this queue support ordered writes
267 * @q: the request queue
268 * @ordered: one of QUEUE_ORDERED_*
269 * @prepare_flush_fn: rq setup helper for cache flush ordered writes
272 * For journalled file systems, doing ordered writes on a commit
273 * block instead of explicitly doing wait_on_buffer (which is bad
274 * for performance) can be a big win. Block drivers supporting this
275 * feature should call this function and indicate so.
278 int blk_queue_ordered(struct request_queue *q, unsigned ordered,
279 prepare_flush_fn *prepare_flush_fn)
281 if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
282 prepare_flush_fn == NULL) {
283 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
287 if (ordered != QUEUE_ORDERED_NONE &&
288 ordered != QUEUE_ORDERED_DRAIN &&
289 ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
290 ordered != QUEUE_ORDERED_DRAIN_FUA &&
291 ordered != QUEUE_ORDERED_TAG &&
292 ordered != QUEUE_ORDERED_TAG_FLUSH &&
293 ordered != QUEUE_ORDERED_TAG_FUA) {
294 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
298 q->ordered = ordered;
299 q->next_ordered = ordered;
300 q->prepare_flush_fn = prepare_flush_fn;
305 EXPORT_SYMBOL(blk_queue_ordered);
308 * blk_queue_issue_flush_fn - set function for issuing a flush
309 * @q: the request queue
310 * @iff: the function to be called issuing the flush
313 * If a driver supports issuing a flush command, the support is notified
314 * to the block layer by defining it through this call.
317 void blk_queue_issue_flush_fn(struct request_queue *q, issue_flush_fn *iff)
319 q->issue_flush_fn = iff;
322 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
325 * Cache flushing for ordered writes handling
327 inline unsigned blk_ordered_cur_seq(struct request_queue *q)
331 return 1 << ffz(q->ordseq);
334 unsigned blk_ordered_req_seq(struct request *rq)
336 struct request_queue *q = rq->q;
338 BUG_ON(q->ordseq == 0);
340 if (rq == &q->pre_flush_rq)
341 return QUEUE_ORDSEQ_PREFLUSH;
342 if (rq == &q->bar_rq)
343 return QUEUE_ORDSEQ_BAR;
344 if (rq == &q->post_flush_rq)
345 return QUEUE_ORDSEQ_POSTFLUSH;
348 * !fs requests don't need to follow barrier ordering. Always
349 * put them at the front. This fixes the following deadlock.
351 * http://thread.gmane.org/gmane.linux.kernel/537473
353 if (!blk_fs_request(rq))
354 return QUEUE_ORDSEQ_DRAIN;
356 if ((rq->cmd_flags & REQ_ORDERED_COLOR) ==
357 (q->orig_bar_rq->cmd_flags & REQ_ORDERED_COLOR))
358 return QUEUE_ORDSEQ_DRAIN;
360 return QUEUE_ORDSEQ_DONE;
363 void blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
368 if (error && !q->orderr)
371 BUG_ON(q->ordseq & seq);
374 if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
378 * Okay, sequence complete.
382 uptodate = q->orderr;
387 end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
388 end_that_request_last(rq, uptodate);
391 static void pre_flush_end_io(struct request *rq, int error)
393 elv_completed_request(rq->q, rq);
394 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
397 static void bar_end_io(struct request *rq, int error)
399 elv_completed_request(rq->q, rq);
400 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
403 static void post_flush_end_io(struct request *rq, int error)
405 elv_completed_request(rq->q, rq);
406 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
409 static void queue_flush(struct request_queue *q, unsigned which)
412 rq_end_io_fn *end_io;
414 if (which == QUEUE_ORDERED_PREFLUSH) {
415 rq = &q->pre_flush_rq;
416 end_io = pre_flush_end_io;
418 rq = &q->post_flush_rq;
419 end_io = post_flush_end_io;
422 rq->cmd_flags = REQ_HARDBARRIER;
424 rq->elevator_private = NULL;
425 rq->elevator_private2 = NULL;
426 rq->rq_disk = q->bar_rq.rq_disk;
428 q->prepare_flush_fn(q, rq);
430 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
433 static inline struct request *start_ordered(struct request_queue *q,
437 q->ordered = q->next_ordered;
438 q->ordseq |= QUEUE_ORDSEQ_STARTED;
441 * Prep proxy barrier request.
443 blkdev_dequeue_request(rq);
448 if (bio_data_dir(q->orig_bar_rq->bio) == WRITE)
449 rq->cmd_flags |= REQ_RW;
450 if (q->ordered & QUEUE_ORDERED_FUA)
451 rq->cmd_flags |= REQ_FUA;
452 rq->elevator_private = NULL;
453 rq->elevator_private2 = NULL;
454 init_request_from_bio(rq, q->orig_bar_rq->bio);
455 rq->end_io = bar_end_io;
458 * Queue ordered sequence. As we stack them at the head, we
459 * need to queue in reverse order. Note that we rely on that
460 * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
461 * request gets inbetween ordered sequence. If this request is
462 * an empty barrier, we don't need to do a postflush ever since
463 * there will be no data written between the pre and post flush.
464 * Hence a single flush will suffice.
466 if ((q->ordered & QUEUE_ORDERED_POSTFLUSH) && !blk_empty_barrier(rq))
467 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
469 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
471 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
473 if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
474 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
475 rq = &q->pre_flush_rq;
477 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
479 if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
480 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
487 int blk_do_ordered(struct request_queue *q, struct request **rqp)
489 struct request *rq = *rqp;
490 const int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
496 if (q->next_ordered != QUEUE_ORDERED_NONE) {
497 *rqp = start_ordered(q, rq);
501 * This can happen when the queue switches to
502 * ORDERED_NONE while this request is on it.
504 blkdev_dequeue_request(rq);
505 end_that_request_first(rq, -EOPNOTSUPP,
506 rq->hard_nr_sectors);
507 end_that_request_last(rq, -EOPNOTSUPP);
514 * Ordered sequence in progress
517 /* Special requests are not subject to ordering rules. */
518 if (!blk_fs_request(rq) &&
519 rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
522 if (q->ordered & QUEUE_ORDERED_TAG) {
523 /* Ordered by tag. Blocking the next barrier is enough. */
524 if (is_barrier && rq != &q->bar_rq)
527 /* Ordered by draining. Wait for turn. */
528 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
529 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
536 static void req_bio_endio(struct request *rq, struct bio *bio,
537 unsigned int nbytes, int error)
539 struct request_queue *q = rq->q;
541 if (&q->bar_rq != rq) {
543 clear_bit(BIO_UPTODATE, &bio->bi_flags);
544 else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
547 if (unlikely(nbytes > bio->bi_size)) {
548 printk("%s: want %u bytes done, only %u left\n",
549 __FUNCTION__, nbytes, bio->bi_size);
550 nbytes = bio->bi_size;
553 bio->bi_size -= nbytes;
554 bio->bi_sector += (nbytes >> 9);
555 if (bio->bi_size == 0)
556 bio_endio(bio, error);
560 * Okay, this is the barrier request in progress, just
563 if (error && !q->orderr)
569 * blk_queue_bounce_limit - set bounce buffer limit for queue
570 * @q: the request queue for the device
571 * @dma_addr: bus address limit
574 * Different hardware can have different requirements as to what pages
575 * it can do I/O directly to. A low level driver can call
576 * blk_queue_bounce_limit to have lower memory pages allocated as bounce
577 * buffers for doing I/O to pages residing above @page.
579 void blk_queue_bounce_limit(struct request_queue *q, u64 dma_addr)
581 unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
584 q->bounce_gfp = GFP_NOIO;
585 #if BITS_PER_LONG == 64
586 /* Assume anything <= 4GB can be handled by IOMMU.
587 Actually some IOMMUs can handle everything, but I don't
588 know of a way to test this here. */
589 if (bounce_pfn < (min_t(u64,0xffffffff,BLK_BOUNCE_HIGH) >> PAGE_SHIFT))
591 q->bounce_pfn = max_low_pfn;
593 if (bounce_pfn < blk_max_low_pfn)
595 q->bounce_pfn = bounce_pfn;
598 init_emergency_isa_pool();
599 q->bounce_gfp = GFP_NOIO | GFP_DMA;
600 q->bounce_pfn = bounce_pfn;
604 EXPORT_SYMBOL(blk_queue_bounce_limit);
607 * blk_queue_max_sectors - set max sectors for a request for this queue
608 * @q: the request queue for the device
609 * @max_sectors: max sectors in the usual 512b unit
612 * Enables a low level driver to set an upper limit on the size of
615 void blk_queue_max_sectors(struct request_queue *q, unsigned int max_sectors)
617 if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
618 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
619 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
622 if (BLK_DEF_MAX_SECTORS > max_sectors)
623 q->max_hw_sectors = q->max_sectors = max_sectors;
625 q->max_sectors = BLK_DEF_MAX_SECTORS;
626 q->max_hw_sectors = max_sectors;
630 EXPORT_SYMBOL(blk_queue_max_sectors);
633 * blk_queue_max_phys_segments - set max phys segments for a request for this queue
634 * @q: the request queue for the device
635 * @max_segments: max number of segments
638 * Enables a low level driver to set an upper limit on the number of
639 * physical data segments in a request. This would be the largest sized
640 * scatter list the driver could handle.
642 void blk_queue_max_phys_segments(struct request_queue *q,
643 unsigned short max_segments)
647 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
650 q->max_phys_segments = max_segments;
653 EXPORT_SYMBOL(blk_queue_max_phys_segments);
656 * blk_queue_max_hw_segments - set max hw segments for a request for this queue
657 * @q: the request queue for the device
658 * @max_segments: max number of segments
661 * Enables a low level driver to set an upper limit on the number of
662 * hw data segments in a request. This would be the largest number of
663 * address/length pairs the host adapter can actually give as once
666 void blk_queue_max_hw_segments(struct request_queue *q,
667 unsigned short max_segments)
671 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
674 q->max_hw_segments = max_segments;
677 EXPORT_SYMBOL(blk_queue_max_hw_segments);
680 * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
681 * @q: the request queue for the device
682 * @max_size: max size of segment in bytes
685 * Enables a low level driver to set an upper limit on the size of a
688 void blk_queue_max_segment_size(struct request_queue *q, unsigned int max_size)
690 if (max_size < PAGE_CACHE_SIZE) {
691 max_size = PAGE_CACHE_SIZE;
692 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
695 q->max_segment_size = max_size;
698 EXPORT_SYMBOL(blk_queue_max_segment_size);
701 * blk_queue_hardsect_size - set hardware sector size for the queue
702 * @q: the request queue for the device
703 * @size: the hardware sector size, in bytes
706 * This should typically be set to the lowest possible sector size
707 * that the hardware can operate on (possible without reverting to
708 * even internal read-modify-write operations). Usually the default
709 * of 512 covers most hardware.
711 void blk_queue_hardsect_size(struct request_queue *q, unsigned short size)
713 q->hardsect_size = size;
716 EXPORT_SYMBOL(blk_queue_hardsect_size);
719 * Returns the minimum that is _not_ zero, unless both are zero.
721 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
724 * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
725 * @t: the stacking driver (top)
726 * @b: the underlying device (bottom)
728 void blk_queue_stack_limits(struct request_queue *t, struct request_queue *b)
730 /* zero is "infinity" */
731 t->max_sectors = min_not_zero(t->max_sectors,b->max_sectors);
732 t->max_hw_sectors = min_not_zero(t->max_hw_sectors,b->max_hw_sectors);
734 t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
735 t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
736 t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
737 t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
738 if (!test_bit(QUEUE_FLAG_CLUSTER, &b->queue_flags))
739 clear_bit(QUEUE_FLAG_CLUSTER, &t->queue_flags);
742 EXPORT_SYMBOL(blk_queue_stack_limits);
745 * blk_queue_segment_boundary - set boundary rules for segment merging
746 * @q: the request queue for the device
747 * @mask: the memory boundary mask
749 void blk_queue_segment_boundary(struct request_queue *q, unsigned long mask)
751 if (mask < PAGE_CACHE_SIZE - 1) {
752 mask = PAGE_CACHE_SIZE - 1;
753 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
756 q->seg_boundary_mask = mask;
759 EXPORT_SYMBOL(blk_queue_segment_boundary);
762 * blk_queue_dma_alignment - set dma length and memory alignment
763 * @q: the request queue for the device
764 * @mask: alignment mask
767 * set required memory and length aligment for direct dma transactions.
768 * this is used when buiding direct io requests for the queue.
771 void blk_queue_dma_alignment(struct request_queue *q, int mask)
773 q->dma_alignment = mask;
776 EXPORT_SYMBOL(blk_queue_dma_alignment);
779 * blk_queue_find_tag - find a request by its tag and queue
780 * @q: The request queue for the device
781 * @tag: The tag of the request
784 * Should be used when a device returns a tag and you want to match
787 * no locks need be held.
789 struct request *blk_queue_find_tag(struct request_queue *q, int tag)
791 return blk_map_queue_find_tag(q->queue_tags, tag);
794 EXPORT_SYMBOL(blk_queue_find_tag);
797 * __blk_free_tags - release a given set of tag maintenance info
798 * @bqt: the tag map to free
800 * Tries to free the specified @bqt@. Returns true if it was
801 * actually freed and false if there are still references using it
803 static int __blk_free_tags(struct blk_queue_tag *bqt)
807 retval = atomic_dec_and_test(&bqt->refcnt);
810 BUG_ON(!list_empty(&bqt->busy_list));
812 kfree(bqt->tag_index);
813 bqt->tag_index = NULL;
826 * __blk_queue_free_tags - release tag maintenance info
827 * @q: the request queue for the device
830 * blk_cleanup_queue() will take care of calling this function, if tagging
831 * has been used. So there's no need to call this directly.
833 static void __blk_queue_free_tags(struct request_queue *q)
835 struct blk_queue_tag *bqt = q->queue_tags;
840 __blk_free_tags(bqt);
842 q->queue_tags = NULL;
843 q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
848 * blk_free_tags - release a given set of tag maintenance info
849 * @bqt: the tag map to free
851 * For externally managed @bqt@ frees the map. Callers of this
852 * function must guarantee to have released all the queues that
853 * might have been using this tag map.
855 void blk_free_tags(struct blk_queue_tag *bqt)
857 if (unlikely(!__blk_free_tags(bqt)))
860 EXPORT_SYMBOL(blk_free_tags);
863 * blk_queue_free_tags - release tag maintenance info
864 * @q: the request queue for the device
867 * This is used to disabled tagged queuing to a device, yet leave
870 void blk_queue_free_tags(struct request_queue *q)
872 clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
875 EXPORT_SYMBOL(blk_queue_free_tags);
878 init_tag_map(struct request_queue *q, struct blk_queue_tag *tags, int depth)
880 struct request **tag_index;
881 unsigned long *tag_map;
884 if (q && depth > q->nr_requests * 2) {
885 depth = q->nr_requests * 2;
886 printk(KERN_ERR "%s: adjusted depth to %d\n",
887 __FUNCTION__, depth);
890 tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC);
894 nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
895 tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
899 tags->real_max_depth = depth;
900 tags->max_depth = depth;
901 tags->tag_index = tag_index;
902 tags->tag_map = tag_map;
910 static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q,
913 struct blk_queue_tag *tags;
915 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
919 if (init_tag_map(q, tags, depth))
922 INIT_LIST_HEAD(&tags->busy_list);
924 atomic_set(&tags->refcnt, 1);
932 * blk_init_tags - initialize the tag info for an external tag map
933 * @depth: the maximum queue depth supported
934 * @tags: the tag to use
936 struct blk_queue_tag *blk_init_tags(int depth)
938 return __blk_queue_init_tags(NULL, depth);
940 EXPORT_SYMBOL(blk_init_tags);
943 * blk_queue_init_tags - initialize the queue tag info
944 * @q: the request queue for the device
945 * @depth: the maximum queue depth supported
946 * @tags: the tag to use
948 int blk_queue_init_tags(struct request_queue *q, int depth,
949 struct blk_queue_tag *tags)
953 BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
955 if (!tags && !q->queue_tags) {
956 tags = __blk_queue_init_tags(q, depth);
960 } else if (q->queue_tags) {
961 if ((rc = blk_queue_resize_tags(q, depth)))
963 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
966 atomic_inc(&tags->refcnt);
969 * assign it, all done
971 q->queue_tags = tags;
972 q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
979 EXPORT_SYMBOL(blk_queue_init_tags);
982 * blk_queue_resize_tags - change the queueing depth
983 * @q: the request queue for the device
984 * @new_depth: the new max command queueing depth
987 * Must be called with the queue lock held.
989 int blk_queue_resize_tags(struct request_queue *q, int new_depth)
991 struct blk_queue_tag *bqt = q->queue_tags;
992 struct request **tag_index;
993 unsigned long *tag_map;
994 int max_depth, nr_ulongs;
1000 * if we already have large enough real_max_depth. just
1001 * adjust max_depth. *NOTE* as requests with tag value
1002 * between new_depth and real_max_depth can be in-flight, tag
1003 * map can not be shrunk blindly here.
1005 if (new_depth <= bqt->real_max_depth) {
1006 bqt->max_depth = new_depth;
1011 * Currently cannot replace a shared tag map with a new
1012 * one, so error out if this is the case
1014 if (atomic_read(&bqt->refcnt) != 1)
1018 * save the old state info, so we can copy it back
1020 tag_index = bqt->tag_index;
1021 tag_map = bqt->tag_map;
1022 max_depth = bqt->real_max_depth;
1024 if (init_tag_map(q, bqt, new_depth))
1027 memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
1028 nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
1029 memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
1036 EXPORT_SYMBOL(blk_queue_resize_tags);
1039 * blk_queue_end_tag - end tag operations for a request
1040 * @q: the request queue for the device
1041 * @rq: the request that has completed
1044 * Typically called when end_that_request_first() returns 0, meaning
1045 * all transfers have been done for a request. It's important to call
1046 * this function before end_that_request_last(), as that will put the
1047 * request back on the free list thus corrupting the internal tag list.
1050 * queue lock must be held.
1052 void blk_queue_end_tag(struct request_queue *q, struct request *rq)
1054 struct blk_queue_tag *bqt = q->queue_tags;
1059 if (unlikely(tag >= bqt->real_max_depth))
1061 * This can happen after tag depth has been reduced.
1062 * FIXME: how about a warning or info message here?
1066 list_del_init(&rq->queuelist);
1067 rq->cmd_flags &= ~REQ_QUEUED;
1070 if (unlikely(bqt->tag_index[tag] == NULL))
1071 printk(KERN_ERR "%s: tag %d is missing\n",
1074 bqt->tag_index[tag] = NULL;
1077 * We use test_and_clear_bit's memory ordering properties here.
1078 * The tag_map bit acts as a lock for tag_index[bit], so we need
1079 * a barrer before clearing the bit (precisely: release semantics).
1080 * Could use clear_bit_unlock when it is merged.
1082 if (unlikely(!test_and_clear_bit(tag, bqt->tag_map))) {
1083 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
1091 EXPORT_SYMBOL(blk_queue_end_tag);
1094 * blk_queue_start_tag - find a free tag and assign it
1095 * @q: the request queue for the device
1096 * @rq: the block request that needs tagging
1099 * This can either be used as a stand-alone helper, or possibly be
1100 * assigned as the queue &prep_rq_fn (in which case &struct request
1101 * automagically gets a tag assigned). Note that this function
1102 * assumes that any type of request can be queued! if this is not
1103 * true for your device, you must check the request type before
1104 * calling this function. The request will also be removed from
1105 * the request queue, so it's the drivers responsibility to readd
1106 * it if it should need to be restarted for some reason.
1109 * queue lock must be held.
1111 int blk_queue_start_tag(struct request_queue *q, struct request *rq)
1113 struct blk_queue_tag *bqt = q->queue_tags;
1116 if (unlikely((rq->cmd_flags & REQ_QUEUED))) {
1118 "%s: request %p for device [%s] already tagged %d",
1120 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
1125 * Protect against shared tag maps, as we may not have exclusive
1126 * access to the tag map.
1129 tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
1130 if (tag >= bqt->max_depth)
1133 } while (test_and_set_bit(tag, bqt->tag_map));
1135 * We rely on test_and_set_bit providing lock memory ordering semantics
1136 * (could use test_and_set_bit_lock when it is merged).
1139 rq->cmd_flags |= REQ_QUEUED;
1141 bqt->tag_index[tag] = rq;
1142 blkdev_dequeue_request(rq);
1143 list_add(&rq->queuelist, &bqt->busy_list);
1148 EXPORT_SYMBOL(blk_queue_start_tag);
1151 * blk_queue_invalidate_tags - invalidate all pending tags
1152 * @q: the request queue for the device
1155 * Hardware conditions may dictate a need to stop all pending requests.
1156 * In this case, we will safely clear the block side of the tag queue and
1157 * readd all requests to the request queue in the right order.
1160 * queue lock must be held.
1162 void blk_queue_invalidate_tags(struct request_queue *q)
1164 struct blk_queue_tag *bqt = q->queue_tags;
1165 struct list_head *tmp, *n;
1168 list_for_each_safe(tmp, n, &bqt->busy_list) {
1169 rq = list_entry_rq(tmp);
1171 if (rq->tag == -1) {
1173 "%s: bad tag found on list\n", __FUNCTION__);
1174 list_del_init(&rq->queuelist);
1175 rq->cmd_flags &= ~REQ_QUEUED;
1177 blk_queue_end_tag(q, rq);
1179 rq->cmd_flags &= ~REQ_STARTED;
1180 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1184 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1186 void blk_dump_rq_flags(struct request *rq, char *msg)
1190 printk("%s: dev %s: type=%x, flags=%x\n", msg,
1191 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->cmd_type,
1194 printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1196 rq->current_nr_sectors);
1197 printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1199 if (blk_pc_request(rq)) {
1201 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1202 printk("%02x ", rq->cmd[bit]);
1207 EXPORT_SYMBOL(blk_dump_rq_flags);
1209 void blk_recount_segments(struct request_queue *q, struct bio *bio)
1212 struct bio *nxt = bio->bi_next;
1214 rq.bio = rq.biotail = bio;
1215 bio->bi_next = NULL;
1216 blk_recalc_rq_segments(&rq);
1218 bio->bi_phys_segments = rq.nr_phys_segments;
1219 bio->bi_hw_segments = rq.nr_hw_segments;
1220 bio->bi_flags |= (1 << BIO_SEG_VALID);
1222 EXPORT_SYMBOL(blk_recount_segments);
1224 static void blk_recalc_rq_segments(struct request *rq)
1228 unsigned int phys_size;
1229 unsigned int hw_size;
1230 struct bio_vec *bv, *bvprv = NULL;
1234 struct req_iterator iter;
1235 int high, highprv = 1;
1236 struct request_queue *q = rq->q;
1241 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1242 hw_seg_size = seg_size = 0;
1243 phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
1244 rq_for_each_segment(bv, rq, iter) {
1246 * the trick here is making sure that a high page is never
1247 * considered part of another segment, since that might
1248 * change with the bounce page.
1250 high = page_to_pfn(bv->bv_page) > q->bounce_pfn;
1251 if (high || highprv)
1252 goto new_hw_segment;
1254 if (seg_size + bv->bv_len > q->max_segment_size)
1256 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1258 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1260 if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1261 goto new_hw_segment;
1263 seg_size += bv->bv_len;
1264 hw_seg_size += bv->bv_len;
1269 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1270 !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1271 hw_seg_size += bv->bv_len;
1274 if (nr_hw_segs == 1 &&
1275 hw_seg_size > rq->bio->bi_hw_front_size)
1276 rq->bio->bi_hw_front_size = hw_seg_size;
1277 hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1283 seg_size = bv->bv_len;
1287 if (nr_hw_segs == 1 &&
1288 hw_seg_size > rq->bio->bi_hw_front_size)
1289 rq->bio->bi_hw_front_size = hw_seg_size;
1290 if (hw_seg_size > rq->biotail->bi_hw_back_size)
1291 rq->biotail->bi_hw_back_size = hw_seg_size;
1292 rq->nr_phys_segments = nr_phys_segs;
1293 rq->nr_hw_segments = nr_hw_segs;
1296 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
1299 if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1302 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1304 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1308 * bio and nxt are contigous in memory, check if the queue allows
1309 * these two to be merged into one
1311 if (BIO_SEG_BOUNDARY(q, bio, nxt))
1317 static int blk_hw_contig_segment(struct request_queue *q, struct bio *bio,
1320 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1321 blk_recount_segments(q, bio);
1322 if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1323 blk_recount_segments(q, nxt);
1324 if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1325 BIOVEC_VIRT_OVERSIZE(bio->bi_hw_back_size + nxt->bi_hw_front_size))
1327 if (bio->bi_hw_back_size + nxt->bi_hw_front_size > q->max_segment_size)
1334 * map a request to scatterlist, return number of sg entries setup. Caller
1335 * must make sure sg can hold rq->nr_phys_segments entries
1337 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
1338 struct scatterlist *sg)
1340 struct bio_vec *bvec, *bvprv;
1341 struct req_iterator iter;
1345 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1348 * for each bio in rq
1351 rq_for_each_segment(bvec, rq, iter) {
1352 int nbytes = bvec->bv_len;
1354 if (bvprv && cluster) {
1355 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1358 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1360 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1363 sg[nsegs - 1].length += nbytes;
1366 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1367 sg[nsegs].page = bvec->bv_page;
1368 sg[nsegs].length = nbytes;
1369 sg[nsegs].offset = bvec->bv_offset;
1374 } /* segments in rq */
1379 EXPORT_SYMBOL(blk_rq_map_sg);
1382 * the standard queue merge functions, can be overridden with device
1383 * specific ones if so desired
1386 static inline int ll_new_mergeable(struct request_queue *q,
1387 struct request *req,
1390 int nr_phys_segs = bio_phys_segments(q, bio);
1392 if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1393 req->cmd_flags |= REQ_NOMERGE;
1394 if (req == q->last_merge)
1395 q->last_merge = NULL;
1400 * A hw segment is just getting larger, bump just the phys
1403 req->nr_phys_segments += nr_phys_segs;
1407 static inline int ll_new_hw_segment(struct request_queue *q,
1408 struct request *req,
1411 int nr_hw_segs = bio_hw_segments(q, bio);
1412 int nr_phys_segs = bio_phys_segments(q, bio);
1414 if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1415 || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1416 req->cmd_flags |= REQ_NOMERGE;
1417 if (req == q->last_merge)
1418 q->last_merge = NULL;
1423 * This will form the start of a new hw segment. Bump both
1426 req->nr_hw_segments += nr_hw_segs;
1427 req->nr_phys_segments += nr_phys_segs;
1431 static int ll_back_merge_fn(struct request_queue *q, struct request *req,
1434 unsigned short max_sectors;
1437 if (unlikely(blk_pc_request(req)))
1438 max_sectors = q->max_hw_sectors;
1440 max_sectors = q->max_sectors;
1442 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1443 req->cmd_flags |= REQ_NOMERGE;
1444 if (req == q->last_merge)
1445 q->last_merge = NULL;
1448 if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1449 blk_recount_segments(q, req->biotail);
1450 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1451 blk_recount_segments(q, bio);
1452 len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1453 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1454 !BIOVEC_VIRT_OVERSIZE(len)) {
1455 int mergeable = ll_new_mergeable(q, req, bio);
1458 if (req->nr_hw_segments == 1)
1459 req->bio->bi_hw_front_size = len;
1460 if (bio->bi_hw_segments == 1)
1461 bio->bi_hw_back_size = len;
1466 return ll_new_hw_segment(q, req, bio);
1469 static int ll_front_merge_fn(struct request_queue *q, struct request *req,
1472 unsigned short max_sectors;
1475 if (unlikely(blk_pc_request(req)))
1476 max_sectors = q->max_hw_sectors;
1478 max_sectors = q->max_sectors;
1481 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1482 req->cmd_flags |= REQ_NOMERGE;
1483 if (req == q->last_merge)
1484 q->last_merge = NULL;
1487 len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1488 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1489 blk_recount_segments(q, bio);
1490 if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1491 blk_recount_segments(q, req->bio);
1492 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1493 !BIOVEC_VIRT_OVERSIZE(len)) {
1494 int mergeable = ll_new_mergeable(q, req, bio);
1497 if (bio->bi_hw_segments == 1)
1498 bio->bi_hw_front_size = len;
1499 if (req->nr_hw_segments == 1)
1500 req->biotail->bi_hw_back_size = len;
1505 return ll_new_hw_segment(q, req, bio);
1508 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
1509 struct request *next)
1511 int total_phys_segments;
1512 int total_hw_segments;
1515 * First check if the either of the requests are re-queued
1516 * requests. Can't merge them if they are.
1518 if (req->special || next->special)
1522 * Will it become too large?
1524 if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1527 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1528 if (blk_phys_contig_segment(q, req->biotail, next->bio))
1529 total_phys_segments--;
1531 if (total_phys_segments > q->max_phys_segments)
1534 total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1535 if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1536 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1538 * propagate the combined length to the end of the requests
1540 if (req->nr_hw_segments == 1)
1541 req->bio->bi_hw_front_size = len;
1542 if (next->nr_hw_segments == 1)
1543 next->biotail->bi_hw_back_size = len;
1544 total_hw_segments--;
1547 if (total_hw_segments > q->max_hw_segments)
1550 /* Merge is OK... */
1551 req->nr_phys_segments = total_phys_segments;
1552 req->nr_hw_segments = total_hw_segments;
1557 * "plug" the device if there are no outstanding requests: this will
1558 * force the transfer to start only after we have put all the requests
1561 * This is called with interrupts off and no requests on the queue and
1562 * with the queue lock held.
1564 void blk_plug_device(struct request_queue *q)
1566 WARN_ON(!irqs_disabled());
1569 * don't plug a stopped queue, it must be paired with blk_start_queue()
1570 * which will restart the queueing
1572 if (blk_queue_stopped(q))
1575 if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
1576 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1577 blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
1581 EXPORT_SYMBOL(blk_plug_device);
1584 * remove the queue from the plugged list, if present. called with
1585 * queue lock held and interrupts disabled.
1587 int blk_remove_plug(struct request_queue *q)
1589 WARN_ON(!irqs_disabled());
1591 if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1594 del_timer(&q->unplug_timer);
1598 EXPORT_SYMBOL(blk_remove_plug);
1601 * remove the plug and let it rip..
1603 void __generic_unplug_device(struct request_queue *q)
1605 if (unlikely(blk_queue_stopped(q)))
1608 if (!blk_remove_plug(q))
1613 EXPORT_SYMBOL(__generic_unplug_device);
1616 * generic_unplug_device - fire a request queue
1617 * @q: The &struct request_queue in question
1620 * Linux uses plugging to build bigger requests queues before letting
1621 * the device have at them. If a queue is plugged, the I/O scheduler
1622 * is still adding and merging requests on the queue. Once the queue
1623 * gets unplugged, the request_fn defined for the queue is invoked and
1624 * transfers started.
1626 void generic_unplug_device(struct request_queue *q)
1628 spin_lock_irq(q->queue_lock);
1629 __generic_unplug_device(q);
1630 spin_unlock_irq(q->queue_lock);
1632 EXPORT_SYMBOL(generic_unplug_device);
1634 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1637 struct request_queue *q = bdi->unplug_io_data;
1640 * devices don't necessarily have an ->unplug_fn defined
1643 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1644 q->rq.count[READ] + q->rq.count[WRITE]);
1650 static void blk_unplug_work(struct work_struct *work)
1652 struct request_queue *q =
1653 container_of(work, struct request_queue, unplug_work);
1655 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1656 q->rq.count[READ] + q->rq.count[WRITE]);
1661 static void blk_unplug_timeout(unsigned long data)
1663 struct request_queue *q = (struct request_queue *)data;
1665 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
1666 q->rq.count[READ] + q->rq.count[WRITE]);
1668 kblockd_schedule_work(&q->unplug_work);
1672 * blk_start_queue - restart a previously stopped queue
1673 * @q: The &struct request_queue in question
1676 * blk_start_queue() will clear the stop flag on the queue, and call
1677 * the request_fn for the queue if it was in a stopped state when
1678 * entered. Also see blk_stop_queue(). Queue lock must be held.
1680 void blk_start_queue(struct request_queue *q)
1682 WARN_ON(!irqs_disabled());
1684 clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1687 * one level of recursion is ok and is much faster than kicking
1688 * the unplug handling
1690 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1692 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1695 kblockd_schedule_work(&q->unplug_work);
1699 EXPORT_SYMBOL(blk_start_queue);
1702 * blk_stop_queue - stop a queue
1703 * @q: The &struct request_queue in question
1706 * The Linux block layer assumes that a block driver will consume all
1707 * entries on the request queue when the request_fn strategy is called.
1708 * Often this will not happen, because of hardware limitations (queue
1709 * depth settings). If a device driver gets a 'queue full' response,
1710 * or if it simply chooses not to queue more I/O at one point, it can
1711 * call this function to prevent the request_fn from being called until
1712 * the driver has signalled it's ready to go again. This happens by calling
1713 * blk_start_queue() to restart queue operations. Queue lock must be held.
1715 void blk_stop_queue(struct request_queue *q)
1718 set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1720 EXPORT_SYMBOL(blk_stop_queue);
1723 * blk_sync_queue - cancel any pending callbacks on a queue
1727 * The block layer may perform asynchronous callback activity
1728 * on a queue, such as calling the unplug function after a timeout.
1729 * A block device may call blk_sync_queue to ensure that any
1730 * such activity is cancelled, thus allowing it to release resources
1731 * that the callbacks might use. The caller must already have made sure
1732 * that its ->make_request_fn will not re-add plugging prior to calling
1736 void blk_sync_queue(struct request_queue *q)
1738 del_timer_sync(&q->unplug_timer);
1740 EXPORT_SYMBOL(blk_sync_queue);
1743 * blk_run_queue - run a single device queue
1744 * @q: The queue to run
1746 void blk_run_queue(struct request_queue *q)
1748 unsigned long flags;
1750 spin_lock_irqsave(q->queue_lock, flags);
1754 * Only recurse once to avoid overrunning the stack, let the unplug
1755 * handling reinvoke the handler shortly if we already got there.
1757 if (!elv_queue_empty(q)) {
1758 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1760 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1763 kblockd_schedule_work(&q->unplug_work);
1767 spin_unlock_irqrestore(q->queue_lock, flags);
1769 EXPORT_SYMBOL(blk_run_queue);
1772 * blk_cleanup_queue: - release a &struct request_queue when it is no longer needed
1773 * @kobj: the kobj belonging of the request queue to be released
1776 * blk_cleanup_queue is the pair to blk_init_queue() or
1777 * blk_queue_make_request(). It should be called when a request queue is
1778 * being released; typically when a block device is being de-registered.
1779 * Currently, its primary task it to free all the &struct request
1780 * structures that were allocated to the queue and the queue itself.
1783 * Hopefully the low level driver will have finished any
1784 * outstanding requests first...
1786 static void blk_release_queue(struct kobject *kobj)
1788 struct request_queue *q =
1789 container_of(kobj, struct request_queue, kobj);
1790 struct request_list *rl = &q->rq;
1795 mempool_destroy(rl->rq_pool);
1798 __blk_queue_free_tags(q);
1800 blk_trace_shutdown(q);
1802 kmem_cache_free(requestq_cachep, q);
1805 void blk_put_queue(struct request_queue *q)
1807 kobject_put(&q->kobj);
1809 EXPORT_SYMBOL(blk_put_queue);
1811 void blk_cleanup_queue(struct request_queue * q)
1813 mutex_lock(&q->sysfs_lock);
1814 set_bit(QUEUE_FLAG_DEAD, &q->queue_flags);
1815 mutex_unlock(&q->sysfs_lock);
1818 elevator_exit(q->elevator);
1823 EXPORT_SYMBOL(blk_cleanup_queue);
1825 static int blk_init_free_list(struct request_queue *q)
1827 struct request_list *rl = &q->rq;
1829 rl->count[READ] = rl->count[WRITE] = 0;
1830 rl->starved[READ] = rl->starved[WRITE] = 0;
1832 init_waitqueue_head(&rl->wait[READ]);
1833 init_waitqueue_head(&rl->wait[WRITE]);
1835 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1836 mempool_free_slab, request_cachep, q->node);
1844 struct request_queue *blk_alloc_queue(gfp_t gfp_mask)
1846 return blk_alloc_queue_node(gfp_mask, -1);
1848 EXPORT_SYMBOL(blk_alloc_queue);
1850 static struct kobj_type queue_ktype;
1852 struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
1854 struct request_queue *q;
1856 q = kmem_cache_alloc_node(requestq_cachep,
1857 gfp_mask | __GFP_ZERO, node_id);
1861 init_timer(&q->unplug_timer);
1863 kobject_set_name(&q->kobj, "%s", "queue");
1864 q->kobj.ktype = &queue_ktype;
1865 kobject_init(&q->kobj);
1867 q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1868 q->backing_dev_info.unplug_io_data = q;
1870 mutex_init(&q->sysfs_lock);
1874 EXPORT_SYMBOL(blk_alloc_queue_node);
1877 * blk_init_queue - prepare a request queue for use with a block device
1878 * @rfn: The function to be called to process requests that have been
1879 * placed on the queue.
1880 * @lock: Request queue spin lock
1883 * If a block device wishes to use the standard request handling procedures,
1884 * which sorts requests and coalesces adjacent requests, then it must
1885 * call blk_init_queue(). The function @rfn will be called when there
1886 * are requests on the queue that need to be processed. If the device
1887 * supports plugging, then @rfn may not be called immediately when requests
1888 * are available on the queue, but may be called at some time later instead.
1889 * Plugged queues are generally unplugged when a buffer belonging to one
1890 * of the requests on the queue is needed, or due to memory pressure.
1892 * @rfn is not required, or even expected, to remove all requests off the
1893 * queue, but only as many as it can handle at a time. If it does leave
1894 * requests on the queue, it is responsible for arranging that the requests
1895 * get dealt with eventually.
1897 * The queue spin lock must be held while manipulating the requests on the
1898 * request queue; this lock will be taken also from interrupt context, so irq
1899 * disabling is needed for it.
1901 * Function returns a pointer to the initialized request queue, or NULL if
1902 * it didn't succeed.
1905 * blk_init_queue() must be paired with a blk_cleanup_queue() call
1906 * when the block device is deactivated (such as at module unload).
1909 struct request_queue *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1911 return blk_init_queue_node(rfn, lock, -1);
1913 EXPORT_SYMBOL(blk_init_queue);
1915 struct request_queue *
1916 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1918 struct request_queue *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1924 if (blk_init_free_list(q)) {
1925 kmem_cache_free(requestq_cachep, q);
1930 * if caller didn't supply a lock, they get per-queue locking with
1934 spin_lock_init(&q->__queue_lock);
1935 lock = &q->__queue_lock;
1938 q->request_fn = rfn;
1939 q->prep_rq_fn = NULL;
1940 q->unplug_fn = generic_unplug_device;
1941 q->queue_flags = (1 << QUEUE_FLAG_CLUSTER);
1942 q->queue_lock = lock;
1944 blk_queue_segment_boundary(q, 0xffffffff);
1946 blk_queue_make_request(q, __make_request);
1947 blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1949 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1950 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1952 q->sg_reserved_size = INT_MAX;
1957 if (!elevator_init(q, NULL)) {
1958 blk_queue_congestion_threshold(q);
1965 EXPORT_SYMBOL(blk_init_queue_node);
1967 int blk_get_queue(struct request_queue *q)
1969 if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1970 kobject_get(&q->kobj);
1977 EXPORT_SYMBOL(blk_get_queue);
1979 static inline void blk_free_request(struct request_queue *q, struct request *rq)
1981 if (rq->cmd_flags & REQ_ELVPRIV)
1982 elv_put_request(q, rq);
1983 mempool_free(rq, q->rq.rq_pool);
1986 static struct request *
1987 blk_alloc_request(struct request_queue *q, int rw, int priv, gfp_t gfp_mask)
1989 struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1995 * first three bits are identical in rq->cmd_flags and bio->bi_rw,
1996 * see bio.h and blkdev.h
1998 rq->cmd_flags = rw | REQ_ALLOCED;
2001 if (unlikely(elv_set_request(q, rq, gfp_mask))) {
2002 mempool_free(rq, q->rq.rq_pool);
2005 rq->cmd_flags |= REQ_ELVPRIV;
2012 * ioc_batching returns true if the ioc is a valid batching request and
2013 * should be given priority access to a request.
2015 static inline int ioc_batching(struct request_queue *q, struct io_context *ioc)
2021 * Make sure the process is able to allocate at least 1 request
2022 * even if the batch times out, otherwise we could theoretically
2025 return ioc->nr_batch_requests == q->nr_batching ||
2026 (ioc->nr_batch_requests > 0
2027 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
2031 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
2032 * will cause the process to be a "batcher" on all queues in the system. This
2033 * is the behaviour we want though - once it gets a wakeup it should be given
2036 static void ioc_set_batching(struct request_queue *q, struct io_context *ioc)
2038 if (!ioc || ioc_batching(q, ioc))
2041 ioc->nr_batch_requests = q->nr_batching;
2042 ioc->last_waited = jiffies;
2045 static void __freed_request(struct request_queue *q, int rw)
2047 struct request_list *rl = &q->rq;
2049 if (rl->count[rw] < queue_congestion_off_threshold(q))
2050 blk_clear_queue_congested(q, rw);
2052 if (rl->count[rw] + 1 <= q->nr_requests) {
2053 if (waitqueue_active(&rl->wait[rw]))
2054 wake_up(&rl->wait[rw]);
2056 blk_clear_queue_full(q, rw);
2061 * A request has just been released. Account for it, update the full and
2062 * congestion status, wake up any waiters. Called under q->queue_lock.
2064 static void freed_request(struct request_queue *q, int rw, int priv)
2066 struct request_list *rl = &q->rq;
2072 __freed_request(q, rw);
2074 if (unlikely(rl->starved[rw ^ 1]))
2075 __freed_request(q, rw ^ 1);
2078 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2080 * Get a free request, queue_lock must be held.
2081 * Returns NULL on failure, with queue_lock held.
2082 * Returns !NULL on success, with queue_lock *not held*.
2084 static struct request *get_request(struct request_queue *q, int rw_flags,
2085 struct bio *bio, gfp_t gfp_mask)
2087 struct request *rq = NULL;
2088 struct request_list *rl = &q->rq;
2089 struct io_context *ioc = NULL;
2090 const int rw = rw_flags & 0x01;
2091 int may_queue, priv;
2093 may_queue = elv_may_queue(q, rw_flags);
2094 if (may_queue == ELV_MQUEUE_NO)
2097 if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
2098 if (rl->count[rw]+1 >= q->nr_requests) {
2099 ioc = current_io_context(GFP_ATOMIC, q->node);
2101 * The queue will fill after this allocation, so set
2102 * it as full, and mark this process as "batching".
2103 * This process will be allowed to complete a batch of
2104 * requests, others will be blocked.
2106 if (!blk_queue_full(q, rw)) {
2107 ioc_set_batching(q, ioc);
2108 blk_set_queue_full(q, rw);
2110 if (may_queue != ELV_MQUEUE_MUST
2111 && !ioc_batching(q, ioc)) {
2113 * The queue is full and the allocating
2114 * process is not a "batcher", and not
2115 * exempted by the IO scheduler
2121 blk_set_queue_congested(q, rw);
2125 * Only allow batching queuers to allocate up to 50% over the defined
2126 * limit of requests, otherwise we could have thousands of requests
2127 * allocated with any setting of ->nr_requests
2129 if (rl->count[rw] >= (3 * q->nr_requests / 2))
2133 rl->starved[rw] = 0;
2135 priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
2139 spin_unlock_irq(q->queue_lock);
2141 rq = blk_alloc_request(q, rw_flags, priv, gfp_mask);
2142 if (unlikely(!rq)) {
2144 * Allocation failed presumably due to memory. Undo anything
2145 * we might have messed up.
2147 * Allocating task should really be put onto the front of the
2148 * wait queue, but this is pretty rare.
2150 spin_lock_irq(q->queue_lock);
2151 freed_request(q, rw, priv);
2154 * in the very unlikely event that allocation failed and no
2155 * requests for this direction was pending, mark us starved
2156 * so that freeing of a request in the other direction will
2157 * notice us. another possible fix would be to split the
2158 * rq mempool into READ and WRITE
2161 if (unlikely(rl->count[rw] == 0))
2162 rl->starved[rw] = 1;
2168 * ioc may be NULL here, and ioc_batching will be false. That's
2169 * OK, if the queue is under the request limit then requests need
2170 * not count toward the nr_batch_requests limit. There will always
2171 * be some limit enforced by BLK_BATCH_TIME.
2173 if (ioc_batching(q, ioc))
2174 ioc->nr_batch_requests--;
2178 blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
2184 * No available requests for this queue, unplug the device and wait for some
2185 * requests to become available.
2187 * Called with q->queue_lock held, and returns with it unlocked.
2189 static struct request *get_request_wait(struct request_queue *q, int rw_flags,
2192 const int rw = rw_flags & 0x01;
2195 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2198 struct request_list *rl = &q->rq;
2200 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2201 TASK_UNINTERRUPTIBLE);
2203 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2206 struct io_context *ioc;
2208 blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
2210 __generic_unplug_device(q);
2211 spin_unlock_irq(q->queue_lock);
2215 * After sleeping, we become a "batching" process and
2216 * will be able to allocate at least one request, and
2217 * up to a big batch of them for a small period time.
2218 * See ioc_batching, ioc_set_batching
2220 ioc = current_io_context(GFP_NOIO, q->node);
2221 ioc_set_batching(q, ioc);
2223 spin_lock_irq(q->queue_lock);
2225 finish_wait(&rl->wait[rw], &wait);
2231 struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
2235 BUG_ON(rw != READ && rw != WRITE);
2237 spin_lock_irq(q->queue_lock);
2238 if (gfp_mask & __GFP_WAIT) {
2239 rq = get_request_wait(q, rw, NULL);
2241 rq = get_request(q, rw, NULL, gfp_mask);
2243 spin_unlock_irq(q->queue_lock);
2245 /* q->queue_lock is unlocked at this point */
2249 EXPORT_SYMBOL(blk_get_request);
2252 * blk_start_queueing - initiate dispatch of requests to device
2253 * @q: request queue to kick into gear
2255 * This is basically a helper to remove the need to know whether a queue
2256 * is plugged or not if someone just wants to initiate dispatch of requests
2259 * The queue lock must be held with interrupts disabled.
2261 void blk_start_queueing(struct request_queue *q)
2263 if (!blk_queue_plugged(q))
2266 __generic_unplug_device(q);
2268 EXPORT_SYMBOL(blk_start_queueing);
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(struct request_queue *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(struct request_queue *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);
2337 blk_start_queueing(q);
2338 spin_unlock_irqrestore(q->queue_lock, flags);
2341 EXPORT_SYMBOL(blk_insert_request);
2343 static int __blk_rq_unmap_user(struct bio *bio)
2348 if (bio_flagged(bio, BIO_USER_MAPPED))
2349 bio_unmap_user(bio);
2351 ret = bio_uncopy_user(bio);
2357 int blk_rq_append_bio(struct request_queue *q, struct request *rq,
2361 blk_rq_bio_prep(q, rq, bio);
2362 else if (!ll_back_merge_fn(q, rq, bio))
2365 rq->biotail->bi_next = bio;
2368 rq->data_len += bio->bi_size;
2372 EXPORT_SYMBOL(blk_rq_append_bio);
2374 static int __blk_rq_map_user(struct request_queue *q, struct request *rq,
2375 void __user *ubuf, unsigned int len)
2377 unsigned long uaddr;
2378 struct bio *bio, *orig_bio;
2381 reading = rq_data_dir(rq) == READ;
2384 * if alignment requirement is satisfied, map in user pages for
2385 * direct dma. else, set up kernel bounce buffers
2387 uaddr = (unsigned long) ubuf;
2388 if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2389 bio = bio_map_user(q, NULL, uaddr, len, reading);
2391 bio = bio_copy_user(q, uaddr, len, reading);
2394 return PTR_ERR(bio);
2397 blk_queue_bounce(q, &bio);
2400 * We link the bounce buffer in and could have to traverse it
2401 * later so we have to get a ref to prevent it from being freed
2405 ret = blk_rq_append_bio(q, rq, bio);
2407 return bio->bi_size;
2409 /* if it was boucned we must call the end io function */
2411 __blk_rq_unmap_user(orig_bio);
2417 * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2418 * @q: request queue where request should be inserted
2419 * @rq: request structure to fill
2420 * @ubuf: the user buffer
2421 * @len: length of user data
2424 * Data will be mapped directly for zero copy io, if possible. Otherwise
2425 * a kernel bounce buffer is used.
2427 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2428 * still in process context.
2430 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2431 * before being submitted to the device, as pages mapped may be out of
2432 * reach. It's the callers responsibility to make sure this happens. The
2433 * original bio must be passed back in to blk_rq_unmap_user() for proper
2436 int blk_rq_map_user(struct request_queue *q, struct request *rq,
2437 void __user *ubuf, unsigned long len)
2439 unsigned long bytes_read = 0;
2440 struct bio *bio = NULL;
2443 if (len > (q->max_hw_sectors << 9))
2448 while (bytes_read != len) {
2449 unsigned long map_len, end, start;
2451 map_len = min_t(unsigned long, len - bytes_read, BIO_MAX_SIZE);
2452 end = ((unsigned long)ubuf + map_len + PAGE_SIZE - 1)
2454 start = (unsigned long)ubuf >> PAGE_SHIFT;
2457 * A bad offset could cause us to require BIO_MAX_PAGES + 1
2458 * pages. If this happens we just lower the requested
2459 * mapping len by a page so that we can fit
2461 if (end - start > BIO_MAX_PAGES)
2462 map_len -= PAGE_SIZE;
2464 ret = __blk_rq_map_user(q, rq, ubuf, map_len);
2473 rq->buffer = rq->data = NULL;
2476 blk_rq_unmap_user(bio);
2480 EXPORT_SYMBOL(blk_rq_map_user);
2483 * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2484 * @q: request queue where request should be inserted
2485 * @rq: request to map data to
2486 * @iov: pointer to the iovec
2487 * @iov_count: number of elements in the iovec
2488 * @len: I/O byte count
2491 * Data will be mapped directly for zero copy io, if possible. Otherwise
2492 * a kernel bounce buffer is used.
2494 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2495 * still in process context.
2497 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2498 * before being submitted to the device, as pages mapped may be out of
2499 * reach. It's the callers responsibility to make sure this happens. The
2500 * original bio must be passed back in to blk_rq_unmap_user() for proper
2503 int blk_rq_map_user_iov(struct request_queue *q, struct request *rq,
2504 struct sg_iovec *iov, int iov_count, unsigned int len)
2508 if (!iov || iov_count <= 0)
2511 /* we don't allow misaligned data like bio_map_user() does. If the
2512 * user is using sg, they're expected to know the alignment constraints
2513 * and respect them accordingly */
2514 bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2516 return PTR_ERR(bio);
2518 if (bio->bi_size != len) {
2520 bio_unmap_user(bio);
2525 blk_rq_bio_prep(q, rq, bio);
2526 rq->buffer = rq->data = NULL;
2530 EXPORT_SYMBOL(blk_rq_map_user_iov);
2533 * blk_rq_unmap_user - unmap a request with user data
2534 * @bio: start of bio list
2537 * Unmap a rq previously mapped by blk_rq_map_user(). The caller must
2538 * supply the original rq->bio from the blk_rq_map_user() return, since
2539 * the io completion may have changed rq->bio.
2541 int blk_rq_unmap_user(struct bio *bio)
2543 struct bio *mapped_bio;
2548 if (unlikely(bio_flagged(bio, BIO_BOUNCED)))
2549 mapped_bio = bio->bi_private;
2551 ret2 = __blk_rq_unmap_user(mapped_bio);
2557 bio_put(mapped_bio);
2563 EXPORT_SYMBOL(blk_rq_unmap_user);
2566 * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2567 * @q: request queue where request should be inserted
2568 * @rq: request to fill
2569 * @kbuf: the kernel buffer
2570 * @len: length of user data
2571 * @gfp_mask: memory allocation flags
2573 int blk_rq_map_kern(struct request_queue *q, struct request *rq, void *kbuf,
2574 unsigned int len, gfp_t gfp_mask)
2578 if (len > (q->max_hw_sectors << 9))
2583 bio = bio_map_kern(q, kbuf, len, gfp_mask);
2585 return PTR_ERR(bio);
2587 if (rq_data_dir(rq) == WRITE)
2588 bio->bi_rw |= (1 << BIO_RW);
2590 blk_rq_bio_prep(q, rq, bio);
2591 blk_queue_bounce(q, &rq->bio);
2592 rq->buffer = rq->data = NULL;
2596 EXPORT_SYMBOL(blk_rq_map_kern);
2599 * blk_execute_rq_nowait - insert a request into queue for execution
2600 * @q: queue to insert the request in
2601 * @bd_disk: matching gendisk
2602 * @rq: request to insert
2603 * @at_head: insert request at head or tail of queue
2604 * @done: I/O completion handler
2607 * Insert a fully prepared request at the back of the io scheduler queue
2608 * for execution. Don't wait for completion.
2610 void blk_execute_rq_nowait(struct request_queue *q, struct gendisk *bd_disk,
2611 struct request *rq, int at_head,
2614 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2616 rq->rq_disk = bd_disk;
2617 rq->cmd_flags |= REQ_NOMERGE;
2619 WARN_ON(irqs_disabled());
2620 spin_lock_irq(q->queue_lock);
2621 __elv_add_request(q, rq, where, 1);
2622 __generic_unplug_device(q);
2623 spin_unlock_irq(q->queue_lock);
2625 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
2628 * blk_execute_rq - insert a request into queue for execution
2629 * @q: queue to insert the request in
2630 * @bd_disk: matching gendisk
2631 * @rq: request to insert
2632 * @at_head: insert request at head or tail of queue
2635 * Insert a fully prepared request at the back of the io scheduler queue
2636 * for execution and wait for completion.
2638 int blk_execute_rq(struct request_queue *q, struct gendisk *bd_disk,
2639 struct request *rq, int at_head)
2641 DECLARE_COMPLETION_ONSTACK(wait);
2642 char sense[SCSI_SENSE_BUFFERSIZE];
2646 * we need an extra reference to the request, so we can look at
2647 * it after io completion
2652 memset(sense, 0, sizeof(sense));
2657 rq->end_io_data = &wait;
2658 blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
2659 wait_for_completion(&wait);
2667 EXPORT_SYMBOL(blk_execute_rq);
2670 * blkdev_issue_flush - queue a flush
2671 * @bdev: blockdev to issue flush for
2672 * @error_sector: error sector
2675 * Issue a flush for the block device in question. Caller can supply
2676 * room for storing the error offset in case of a flush error, if they
2677 * wish to. Caller must run wait_for_completion() on its own.
2679 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2681 struct request_queue *q;
2683 if (bdev->bd_disk == NULL)
2686 q = bdev_get_queue(bdev);
2689 if (!q->issue_flush_fn)
2692 return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2695 EXPORT_SYMBOL(blkdev_issue_flush);
2697 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2699 int rw = rq_data_dir(rq);
2701 if (!blk_fs_request(rq) || !rq->rq_disk)
2705 __disk_stat_inc(rq->rq_disk, merges[rw]);
2707 disk_round_stats(rq->rq_disk);
2708 rq->rq_disk->in_flight++;
2713 * add-request adds a request to the linked list.
2714 * queue lock is held and interrupts disabled, as we muck with the
2715 * request queue list.
2717 static inline void add_request(struct request_queue * q, struct request * req)
2719 drive_stat_acct(req, req->nr_sectors, 1);
2722 * elevator indicated where it wants this request to be
2723 * inserted at elevator_merge time
2725 __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2729 * disk_round_stats() - Round off the performance stats on a struct
2732 * The average IO queue length and utilisation statistics are maintained
2733 * by observing the current state of the queue length and the amount of
2734 * time it has been in this state for.
2736 * Normally, that accounting is done on IO completion, but that can result
2737 * in more than a second's worth of IO being accounted for within any one
2738 * second, leading to >100% utilisation. To deal with that, we call this
2739 * function to do a round-off before returning the results when reading
2740 * /proc/diskstats. This accounts immediately for all queue usage up to
2741 * the current jiffies and restarts the counters again.
2743 void disk_round_stats(struct gendisk *disk)
2745 unsigned long now = jiffies;
2747 if (now == disk->stamp)
2750 if (disk->in_flight) {
2751 __disk_stat_add(disk, time_in_queue,
2752 disk->in_flight * (now - disk->stamp));
2753 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2758 EXPORT_SYMBOL_GPL(disk_round_stats);
2761 * queue lock must be held
2763 void __blk_put_request(struct request_queue *q, struct request *req)
2767 if (unlikely(--req->ref_count))
2770 elv_completed_request(q, req);
2773 * Request may not have originated from ll_rw_blk. if not,
2774 * it didn't come out of our reserved rq pools
2776 if (req->cmd_flags & REQ_ALLOCED) {
2777 int rw = rq_data_dir(req);
2778 int priv = req->cmd_flags & REQ_ELVPRIV;
2780 BUG_ON(!list_empty(&req->queuelist));
2781 BUG_ON(!hlist_unhashed(&req->hash));
2783 blk_free_request(q, req);
2784 freed_request(q, rw, priv);
2788 EXPORT_SYMBOL_GPL(__blk_put_request);
2790 void blk_put_request(struct request *req)
2792 unsigned long flags;
2793 struct request_queue *q = req->q;
2796 * Gee, IDE calls in w/ NULL q. Fix IDE and remove the
2797 * following if (q) test.
2800 spin_lock_irqsave(q->queue_lock, flags);
2801 __blk_put_request(q, req);
2802 spin_unlock_irqrestore(q->queue_lock, flags);
2806 EXPORT_SYMBOL(blk_put_request);
2809 * blk_end_sync_rq - executes a completion event on a request
2810 * @rq: request to complete
2811 * @error: end io status of the request
2813 void blk_end_sync_rq(struct request *rq, int error)
2815 struct completion *waiting = rq->end_io_data;
2817 rq->end_io_data = NULL;
2818 __blk_put_request(rq->q, rq);
2821 * complete last, if this is a stack request the process (and thus
2822 * the rq pointer) could be invalid right after this complete()
2826 EXPORT_SYMBOL(blk_end_sync_rq);
2829 * Has to be called with the request spinlock acquired
2831 static int attempt_merge(struct request_queue *q, struct request *req,
2832 struct request *next)
2834 if (!rq_mergeable(req) || !rq_mergeable(next))
2840 if (req->sector + req->nr_sectors != next->sector)
2843 if (rq_data_dir(req) != rq_data_dir(next)
2844 || req->rq_disk != next->rq_disk
2849 * If we are allowed to merge, then append bio list
2850 * from next to rq and release next. merge_requests_fn
2851 * will have updated segment counts, update sector
2854 if (!ll_merge_requests_fn(q, req, next))
2858 * At this point we have either done a back merge
2859 * or front merge. We need the smaller start_time of
2860 * the merged requests to be the current request
2861 * for accounting purposes.
2863 if (time_after(req->start_time, next->start_time))
2864 req->start_time = next->start_time;
2866 req->biotail->bi_next = next->bio;
2867 req->biotail = next->biotail;
2869 req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2871 elv_merge_requests(q, req, next);
2874 disk_round_stats(req->rq_disk);
2875 req->rq_disk->in_flight--;
2878 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2880 __blk_put_request(q, next);
2884 static inline int attempt_back_merge(struct request_queue *q,
2887 struct request *next = elv_latter_request(q, rq);
2890 return attempt_merge(q, rq, next);
2895 static inline int attempt_front_merge(struct request_queue *q,
2898 struct request *prev = elv_former_request(q, rq);
2901 return attempt_merge(q, prev, rq);
2906 static void init_request_from_bio(struct request *req, struct bio *bio)
2908 req->cmd_type = REQ_TYPE_FS;
2911 * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2913 if (bio_rw_ahead(bio) || bio_failfast(bio))
2914 req->cmd_flags |= REQ_FAILFAST;
2917 * REQ_BARRIER implies no merging, but lets make it explicit
2919 if (unlikely(bio_barrier(bio)))
2920 req->cmd_flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2923 req->cmd_flags |= REQ_RW_SYNC;
2924 if (bio_rw_meta(bio))
2925 req->cmd_flags |= REQ_RW_META;
2928 req->hard_sector = req->sector = bio->bi_sector;
2929 req->ioprio = bio_prio(bio);
2930 req->start_time = jiffies;
2931 blk_rq_bio_prep(req->q, req, bio);
2934 static int __make_request(struct request_queue *q, struct bio *bio)
2936 struct request *req;
2937 int el_ret, nr_sectors, barrier, err;
2938 const unsigned short prio = bio_prio(bio);
2939 const int sync = bio_sync(bio);
2942 nr_sectors = bio_sectors(bio);
2945 * low level driver can indicate that it wants pages above a
2946 * certain limit bounced to low memory (ie for highmem, or even
2947 * ISA dma in theory)
2949 blk_queue_bounce(q, &bio);
2951 barrier = bio_barrier(bio);
2952 if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
2957 spin_lock_irq(q->queue_lock);
2959 if (unlikely(barrier) || elv_queue_empty(q))
2962 el_ret = elv_merge(q, &req, bio);
2964 case ELEVATOR_BACK_MERGE:
2965 BUG_ON(!rq_mergeable(req));
2967 if (!ll_back_merge_fn(q, req, bio))
2970 blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
2972 req->biotail->bi_next = bio;
2974 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2975 req->ioprio = ioprio_best(req->ioprio, prio);
2976 drive_stat_acct(req, nr_sectors, 0);
2977 if (!attempt_back_merge(q, req))
2978 elv_merged_request(q, req, el_ret);
2981 case ELEVATOR_FRONT_MERGE:
2982 BUG_ON(!rq_mergeable(req));
2984 if (!ll_front_merge_fn(q, req, bio))
2987 blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
2989 bio->bi_next = req->bio;
2993 * may not be valid. if the low level driver said
2994 * it didn't need a bounce buffer then it better
2995 * not touch req->buffer either...
2997 req->buffer = bio_data(bio);
2998 req->current_nr_sectors = bio_cur_sectors(bio);
2999 req->hard_cur_sectors = req->current_nr_sectors;
3000 req->sector = req->hard_sector = bio->bi_sector;
3001 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
3002 req->ioprio = ioprio_best(req->ioprio, prio);
3003 drive_stat_acct(req, nr_sectors, 0);
3004 if (!attempt_front_merge(q, req))
3005 elv_merged_request(q, req, el_ret);
3008 /* ELV_NO_MERGE: elevator says don't/can't merge. */
3015 * This sync check and mask will be re-done in init_request_from_bio(),
3016 * but we need to set it earlier to expose the sync flag to the
3017 * rq allocator and io schedulers.
3019 rw_flags = bio_data_dir(bio);
3021 rw_flags |= REQ_RW_SYNC;
3024 * Grab a free request. This is might sleep but can not fail.
3025 * Returns with the queue unlocked.
3027 req = get_request_wait(q, rw_flags, bio);
3030 * After dropping the lock and possibly sleeping here, our request
3031 * may now be mergeable after it had proven unmergeable (above).
3032 * We don't worry about that case for efficiency. It won't happen
3033 * often, and the elevators are able to handle it.
3035 init_request_from_bio(req, bio);
3037 spin_lock_irq(q->queue_lock);
3038 if (elv_queue_empty(q))
3040 add_request(q, req);
3043 __generic_unplug_device(q);
3045 spin_unlock_irq(q->queue_lock);
3049 bio_endio(bio, err);
3054 * If bio->bi_dev is a partition, remap the location
3056 static inline void blk_partition_remap(struct bio *bio)
3058 struct block_device *bdev = bio->bi_bdev;
3060 if (bio_sectors(bio) && bdev != bdev->bd_contains) {
3061 struct hd_struct *p = bdev->bd_part;
3062 const int rw = bio_data_dir(bio);
3064 p->sectors[rw] += bio_sectors(bio);
3067 bio->bi_sector += p->start_sect;
3068 bio->bi_bdev = bdev->bd_contains;
3070 blk_add_trace_remap(bdev_get_queue(bio->bi_bdev), bio,
3071 bdev->bd_dev, bio->bi_sector,
3072 bio->bi_sector - p->start_sect);
3076 static void handle_bad_sector(struct bio *bio)
3078 char b[BDEVNAME_SIZE];
3080 printk(KERN_INFO "attempt to access beyond end of device\n");
3081 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
3082 bdevname(bio->bi_bdev, b),
3084 (unsigned long long)bio->bi_sector + bio_sectors(bio),
3085 (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
3087 set_bit(BIO_EOF, &bio->bi_flags);
3090 #ifdef CONFIG_FAIL_MAKE_REQUEST
3092 static DECLARE_FAULT_ATTR(fail_make_request);
3094 static int __init setup_fail_make_request(char *str)
3096 return setup_fault_attr(&fail_make_request, str);
3098 __setup("fail_make_request=", setup_fail_make_request);
3100 static int should_fail_request(struct bio *bio)
3102 if ((bio->bi_bdev->bd_disk->flags & GENHD_FL_FAIL) ||
3103 (bio->bi_bdev->bd_part && bio->bi_bdev->bd_part->make_it_fail))
3104 return should_fail(&fail_make_request, bio->bi_size);
3109 static int __init fail_make_request_debugfs(void)
3111 return init_fault_attr_dentries(&fail_make_request,
3112 "fail_make_request");
3115 late_initcall(fail_make_request_debugfs);
3117 #else /* CONFIG_FAIL_MAKE_REQUEST */
3119 static inline int should_fail_request(struct bio *bio)
3124 #endif /* CONFIG_FAIL_MAKE_REQUEST */
3127 * Check whether this bio extends beyond the end of the device.
3129 static inline int bio_check_eod(struct bio *bio, unsigned int nr_sectors)
3136 /* Test device or partition size, when known. */
3137 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3139 sector_t sector = bio->bi_sector;
3141 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
3143 * This may well happen - the kernel calls bread()
3144 * without checking the size of the device, e.g., when
3145 * mounting a device.
3147 handle_bad_sector(bio);
3156 * generic_make_request: hand a buffer to its device driver for I/O
3157 * @bio: The bio describing the location in memory and on the device.
3159 * generic_make_request() is used to make I/O requests of block
3160 * devices. It is passed a &struct bio, which describes the I/O that needs
3163 * generic_make_request() does not return any status. The
3164 * success/failure status of the request, along with notification of
3165 * completion, is delivered asynchronously through the bio->bi_end_io
3166 * function described (one day) else where.
3168 * The caller of generic_make_request must make sure that bi_io_vec
3169 * are set to describe the memory buffer, and that bi_dev and bi_sector are
3170 * set to describe the device address, and the
3171 * bi_end_io and optionally bi_private are set to describe how
3172 * completion notification should be signaled.
3174 * generic_make_request and the drivers it calls may use bi_next if this
3175 * bio happens to be merged with someone else, and may change bi_dev and
3176 * bi_sector for remaps as it sees fit. So the values of these fields
3177 * should NOT be depended on after the call to generic_make_request.
3179 static inline void __generic_make_request(struct bio *bio)
3181 struct request_queue *q;
3182 sector_t old_sector;
3183 int ret, nr_sectors = bio_sectors(bio);
3188 if (bio_check_eod(bio, nr_sectors))
3192 * Resolve the mapping until finished. (drivers are
3193 * still free to implement/resolve their own stacking
3194 * by explicitly returning 0)
3196 * NOTE: we don't repeat the blk_size check for each new device.
3197 * Stacking drivers are expected to know what they are doing.
3202 char b[BDEVNAME_SIZE];
3204 q = bdev_get_queue(bio->bi_bdev);
3207 "generic_make_request: Trying to access "
3208 "nonexistent block-device %s (%Lu)\n",
3209 bdevname(bio->bi_bdev, b),
3210 (long long) bio->bi_sector);
3212 bio_endio(bio, -EIO);
3216 if (unlikely(nr_sectors > q->max_hw_sectors)) {
3217 printk("bio too big device %s (%u > %u)\n",
3218 bdevname(bio->bi_bdev, b),
3224 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
3227 if (should_fail_request(bio))
3231 * If this device has partitions, remap block n
3232 * of partition p to block n+start(p) of the disk.
3234 blk_partition_remap(bio);
3236 if (old_sector != -1)
3237 blk_add_trace_remap(q, bio, old_dev, bio->bi_sector,
3240 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
3242 old_sector = bio->bi_sector;
3243 old_dev = bio->bi_bdev->bd_dev;
3245 if (bio_check_eod(bio, nr_sectors))
3248 ret = q->make_request_fn(q, bio);
3253 * We only want one ->make_request_fn to be active at a time,
3254 * else stack usage with stacked devices could be a problem.
3255 * So use current->bio_{list,tail} to keep a list of requests
3256 * submited by a make_request_fn function.
3257 * current->bio_tail is also used as a flag to say if
3258 * generic_make_request is currently active in this task or not.
3259 * If it is NULL, then no make_request is active. If it is non-NULL,
3260 * then a make_request is active, and new requests should be added
3263 void generic_make_request(struct bio *bio)
3265 if (current->bio_tail) {
3266 /* make_request is active */
3267 *(current->bio_tail) = bio;
3268 bio->bi_next = NULL;
3269 current->bio_tail = &bio->bi_next;
3272 /* following loop may be a bit non-obvious, and so deserves some
3274 * Before entering the loop, bio->bi_next is NULL (as all callers
3275 * ensure that) so we have a list with a single bio.
3276 * We pretend that we have just taken it off a longer list, so
3277 * we assign bio_list to the next (which is NULL) and bio_tail
3278 * to &bio_list, thus initialising the bio_list of new bios to be
3279 * added. __generic_make_request may indeed add some more bios
3280 * through a recursive call to generic_make_request. If it
3281 * did, we find a non-NULL value in bio_list and re-enter the loop
3282 * from the top. In this case we really did just take the bio
3283 * of the top of the list (no pretending) and so fixup bio_list and
3284 * bio_tail or bi_next, and call into __generic_make_request again.
3286 * The loop was structured like this to make only one call to
3287 * __generic_make_request (which is important as it is large and
3288 * inlined) and to keep the structure simple.
3290 BUG_ON(bio->bi_next);
3292 current->bio_list = bio->bi_next;
3293 if (bio->bi_next == NULL)
3294 current->bio_tail = ¤t->bio_list;
3296 bio->bi_next = NULL;
3297 __generic_make_request(bio);
3298 bio = current->bio_list;
3300 current->bio_tail = NULL; /* deactivate */
3303 EXPORT_SYMBOL(generic_make_request);
3306 * submit_bio: submit a bio to the block device layer for I/O
3307 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3308 * @bio: The &struct bio which describes the I/O
3310 * submit_bio() is very similar in purpose to generic_make_request(), and
3311 * uses that function to do most of the work. Both are fairly rough
3312 * interfaces, @bio must be presetup and ready for I/O.
3315 void submit_bio(int rw, struct bio *bio)
3317 int count = bio_sectors(bio);
3322 * If it's a regular read/write or a barrier with data attached,
3323 * go through the normal accounting stuff before submission.
3325 if (!bio_empty_barrier(bio)) {
3327 BIO_BUG_ON(!bio->bi_size);
3328 BIO_BUG_ON(!bio->bi_io_vec);
3331 count_vm_events(PGPGOUT, count);
3333 task_io_account_read(bio->bi_size);
3334 count_vm_events(PGPGIN, count);
3337 if (unlikely(block_dump)) {
3338 char b[BDEVNAME_SIZE];
3339 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
3340 current->comm, current->pid,
3341 (rw & WRITE) ? "WRITE" : "READ",
3342 (unsigned long long)bio->bi_sector,
3343 bdevname(bio->bi_bdev,b));
3347 generic_make_request(bio);
3350 EXPORT_SYMBOL(submit_bio);
3352 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
3354 if (blk_fs_request(rq)) {
3355 rq->hard_sector += nsect;
3356 rq->hard_nr_sectors -= nsect;
3359 * Move the I/O submission pointers ahead if required.
3361 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3362 (rq->sector <= rq->hard_sector)) {
3363 rq->sector = rq->hard_sector;
3364 rq->nr_sectors = rq->hard_nr_sectors;
3365 rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3366 rq->current_nr_sectors = rq->hard_cur_sectors;
3367 rq->buffer = bio_data(rq->bio);
3371 * if total number of sectors is less than the first segment
3372 * size, something has gone terribly wrong
3374 if (rq->nr_sectors < rq->current_nr_sectors) {
3375 printk("blk: request botched\n");
3376 rq->nr_sectors = rq->current_nr_sectors;
3381 static int __end_that_request_first(struct request *req, int uptodate,
3384 int total_bytes, bio_nbytes, error, next_idx = 0;
3387 blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
3390 * extend uptodate bool to allow < 0 value to be direct io error
3393 if (end_io_error(uptodate))
3394 error = !uptodate ? -EIO : uptodate;
3397 * for a REQ_BLOCK_PC request, we want to carry any eventual
3398 * sense key with us all the way through
3400 if (!blk_pc_request(req))
3404 if (blk_fs_request(req) && !(req->cmd_flags & REQ_QUIET))
3405 printk("end_request: I/O error, dev %s, sector %llu\n",
3406 req->rq_disk ? req->rq_disk->disk_name : "?",
3407 (unsigned long long)req->sector);
3410 if (blk_fs_request(req) && req->rq_disk) {
3411 const int rw = rq_data_dir(req);
3413 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
3416 total_bytes = bio_nbytes = 0;
3417 while ((bio = req->bio) != NULL) {
3421 * For an empty barrier request, the low level driver must
3422 * store a potential error location in ->sector. We pass
3423 * that back up in ->bi_sector.
3425 if (blk_empty_barrier(req))
3426 bio->bi_sector = req->sector;
3428 if (nr_bytes >= bio->bi_size) {
3429 req->bio = bio->bi_next;
3430 nbytes = bio->bi_size;
3431 req_bio_endio(req, bio, nbytes, error);
3435 int idx = bio->bi_idx + next_idx;
3437 if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3438 blk_dump_rq_flags(req, "__end_that");
3439 printk("%s: bio idx %d >= vcnt %d\n",
3441 bio->bi_idx, bio->bi_vcnt);
3445 nbytes = bio_iovec_idx(bio, idx)->bv_len;
3446 BIO_BUG_ON(nbytes > bio->bi_size);
3449 * not a complete bvec done
3451 if (unlikely(nbytes > nr_bytes)) {
3452 bio_nbytes += nr_bytes;
3453 total_bytes += nr_bytes;
3458 * advance to the next vector
3461 bio_nbytes += nbytes;
3464 total_bytes += nbytes;
3467 if ((bio = req->bio)) {
3469 * end more in this run, or just return 'not-done'
3471 if (unlikely(nr_bytes <= 0))
3483 * if the request wasn't completed, update state
3486 req_bio_endio(req, bio, bio_nbytes, error);
3487 bio->bi_idx += next_idx;
3488 bio_iovec(bio)->bv_offset += nr_bytes;
3489 bio_iovec(bio)->bv_len -= nr_bytes;
3492 blk_recalc_rq_sectors(req, total_bytes >> 9);
3493 blk_recalc_rq_segments(req);
3498 * end_that_request_first - end I/O on a request
3499 * @req: the request being processed
3500 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3501 * @nr_sectors: number of sectors to end I/O on
3504 * Ends I/O on a number of sectors attached to @req, and sets it up
3505 * for the next range of segments (if any) in the cluster.
3508 * 0 - we are done with this request, call end_that_request_last()
3509 * 1 - still buffers pending for this request
3511 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3513 return __end_that_request_first(req, uptodate, nr_sectors << 9);
3516 EXPORT_SYMBOL(end_that_request_first);
3519 * end_that_request_chunk - end I/O on a request
3520 * @req: the request being processed
3521 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3522 * @nr_bytes: number of bytes to complete
3525 * Ends I/O on a number of bytes attached to @req, and sets it up
3526 * for the next range of segments (if any). Like end_that_request_first(),
3527 * but deals with bytes instead of sectors.
3530 * 0 - we are done with this request, call end_that_request_last()
3531 * 1 - still buffers pending for this request
3533 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3535 return __end_that_request_first(req, uptodate, nr_bytes);
3538 EXPORT_SYMBOL(end_that_request_chunk);
3541 * splice the completion data to a local structure and hand off to
3542 * process_completion_queue() to complete the requests
3544 static void blk_done_softirq(struct softirq_action *h)
3546 struct list_head *cpu_list, local_list;
3548 local_irq_disable();
3549 cpu_list = &__get_cpu_var(blk_cpu_done);
3550 list_replace_init(cpu_list, &local_list);
3553 while (!list_empty(&local_list)) {
3554 struct request *rq = list_entry(local_list.next, struct request, donelist);
3556 list_del_init(&rq->donelist);
3557 rq->q->softirq_done_fn(rq);
3561 static int __cpuinit blk_cpu_notify(struct notifier_block *self, unsigned long action,
3565 * If a CPU goes away, splice its entries to the current CPU
3566 * and trigger a run of the softirq
3568 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
3569 int cpu = (unsigned long) hcpu;
3571 local_irq_disable();
3572 list_splice_init(&per_cpu(blk_cpu_done, cpu),
3573 &__get_cpu_var(blk_cpu_done));
3574 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3582 static struct notifier_block blk_cpu_notifier __cpuinitdata = {
3583 .notifier_call = blk_cpu_notify,
3587 * blk_complete_request - end I/O on a request
3588 * @req: the request being processed
3591 * Ends all I/O on a request. It does not handle partial completions,
3592 * unless the driver actually implements this in its completion callback
3593 * through requeueing. The actual completion happens out-of-order,
3594 * through a softirq handler. The user must have registered a completion
3595 * callback through blk_queue_softirq_done().
3598 void blk_complete_request(struct request *req)
3600 struct list_head *cpu_list;
3601 unsigned long flags;
3603 BUG_ON(!req->q->softirq_done_fn);
3605 local_irq_save(flags);
3607 cpu_list = &__get_cpu_var(blk_cpu_done);
3608 list_add_tail(&req->donelist, cpu_list);
3609 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3611 local_irq_restore(flags);
3614 EXPORT_SYMBOL(blk_complete_request);
3617 * queue lock must be held
3619 void end_that_request_last(struct request *req, int uptodate)
3621 struct gendisk *disk = req->rq_disk;
3625 * extend uptodate bool to allow < 0 value to be direct io error
3628 if (end_io_error(uptodate))
3629 error = !uptodate ? -EIO : uptodate;
3631 if (unlikely(laptop_mode) && blk_fs_request(req))
3632 laptop_io_completion();
3635 * Account IO completion. bar_rq isn't accounted as a normal
3636 * IO on queueing nor completion. Accounting the containing
3637 * request is enough.
3639 if (disk && blk_fs_request(req) && req != &req->q->bar_rq) {
3640 unsigned long duration = jiffies - req->start_time;
3641 const int rw = rq_data_dir(req);
3643 __disk_stat_inc(disk, ios[rw]);
3644 __disk_stat_add(disk, ticks[rw], duration);
3645 disk_round_stats(disk);
3649 req->end_io(req, error);
3651 __blk_put_request(req->q, req);
3654 EXPORT_SYMBOL(end_that_request_last);
3656 static inline void __end_request(struct request *rq, int uptodate,
3657 unsigned int nr_bytes, int dequeue)
3659 if (!end_that_request_chunk(rq, uptodate, nr_bytes)) {
3661 blkdev_dequeue_request(rq);
3662 add_disk_randomness(rq->rq_disk);
3663 end_that_request_last(rq, uptodate);
3667 static unsigned int rq_byte_size(struct request *rq)
3669 if (blk_fs_request(rq))
3670 return rq->hard_nr_sectors << 9;
3672 return rq->data_len;
3676 * end_queued_request - end all I/O on a queued request
3677 * @rq: the request being processed
3678 * @uptodate: error value or 0/1 uptodate flag
3681 * Ends all I/O on a request, and removes it from the block layer queues.
3682 * Not suitable for normal IO completion, unless the driver still has
3683 * the request attached to the block layer.
3686 void end_queued_request(struct request *rq, int uptodate)
3688 __end_request(rq, uptodate, rq_byte_size(rq), 1);
3690 EXPORT_SYMBOL(end_queued_request);
3693 * end_dequeued_request - end all I/O on a dequeued request
3694 * @rq: the request being processed
3695 * @uptodate: error value or 0/1 uptodate flag
3698 * Ends all I/O on a request. The request must already have been
3699 * dequeued using blkdev_dequeue_request(), as is normally the case
3703 void end_dequeued_request(struct request *rq, int uptodate)
3705 __end_request(rq, uptodate, rq_byte_size(rq), 0);
3707 EXPORT_SYMBOL(end_dequeued_request);
3711 * end_request - end I/O on the current segment of the request
3712 * @rq: the request being processed
3713 * @uptodate: error value or 0/1 uptodate flag
3716 * Ends I/O on the current segment of a request. If that is the only
3717 * remaining segment, the request is also completed and freed.
3719 * This is a remnant of how older block drivers handled IO completions.
3720 * Modern drivers typically end IO on the full request in one go, unless
3721 * they have a residual value to account for. For that case this function
3722 * isn't really useful, unless the residual just happens to be the
3723 * full current segment. In other words, don't use this function in new
3724 * code. Either use end_request_completely(), or the
3725 * end_that_request_chunk() (along with end_that_request_last()) for
3726 * partial completions.
3729 void end_request(struct request *req, int uptodate)
3731 __end_request(req, uptodate, req->hard_cur_sectors << 9, 1);
3733 EXPORT_SYMBOL(end_request);
3735 static void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
3738 /* first two bits are identical in rq->cmd_flags and bio->bi_rw */
3739 rq->cmd_flags |= (bio->bi_rw & 3);
3741 rq->nr_phys_segments = bio_phys_segments(q, bio);
3742 rq->nr_hw_segments = bio_hw_segments(q, bio);
3743 rq->current_nr_sectors = bio_cur_sectors(bio);
3744 rq->hard_cur_sectors = rq->current_nr_sectors;
3745 rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3746 rq->buffer = bio_data(bio);
3747 rq->data_len = bio->bi_size;
3749 rq->bio = rq->biotail = bio;
3752 rq->rq_disk = bio->bi_bdev->bd_disk;
3755 int kblockd_schedule_work(struct work_struct *work)
3757 return queue_work(kblockd_workqueue, work);
3760 EXPORT_SYMBOL(kblockd_schedule_work);
3762 void kblockd_flush_work(struct work_struct *work)
3764 cancel_work_sync(work);
3766 EXPORT_SYMBOL(kblockd_flush_work);
3768 int __init blk_dev_init(void)
3772 kblockd_workqueue = create_workqueue("kblockd");
3773 if (!kblockd_workqueue)
3774 panic("Failed to create kblockd\n");
3776 request_cachep = kmem_cache_create("blkdev_requests",
3777 sizeof(struct request), 0, SLAB_PANIC, NULL);
3779 requestq_cachep = kmem_cache_create("blkdev_queue",
3780 sizeof(struct request_queue), 0, SLAB_PANIC, NULL);
3782 iocontext_cachep = kmem_cache_create("blkdev_ioc",
3783 sizeof(struct io_context), 0, SLAB_PANIC, NULL);
3785 for_each_possible_cpu(i)
3786 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3788 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
3789 register_hotcpu_notifier(&blk_cpu_notifier);
3791 blk_max_low_pfn = max_low_pfn - 1;
3792 blk_max_pfn = max_pfn - 1;
3798 * IO Context helper functions
3800 void put_io_context(struct io_context *ioc)
3805 BUG_ON(atomic_read(&ioc->refcount) == 0);
3807 if (atomic_dec_and_test(&ioc->refcount)) {
3808 struct cfq_io_context *cic;
3811 if (ioc->aic && ioc->aic->dtor)
3812 ioc->aic->dtor(ioc->aic);
3813 if (ioc->cic_root.rb_node != NULL) {
3814 struct rb_node *n = rb_first(&ioc->cic_root);
3816 cic = rb_entry(n, struct cfq_io_context, rb_node);
3821 kmem_cache_free(iocontext_cachep, ioc);
3824 EXPORT_SYMBOL(put_io_context);
3826 /* Called by the exitting task */
3827 void exit_io_context(void)
3829 struct io_context *ioc;
3830 struct cfq_io_context *cic;
3833 ioc = current->io_context;
3834 current->io_context = NULL;
3835 task_unlock(current);
3838 if (ioc->aic && ioc->aic->exit)
3839 ioc->aic->exit(ioc->aic);
3840 if (ioc->cic_root.rb_node != NULL) {
3841 cic = rb_entry(rb_first(&ioc->cic_root), struct cfq_io_context, rb_node);
3845 put_io_context(ioc);
3849 * If the current task has no IO context then create one and initialise it.
3850 * Otherwise, return its existing IO context.
3852 * This returned IO context doesn't have a specifically elevated refcount,
3853 * but since the current task itself holds a reference, the context can be
3854 * used in general code, so long as it stays within `current` context.
3856 static struct io_context *current_io_context(gfp_t gfp_flags, int node)
3858 struct task_struct *tsk = current;
3859 struct io_context *ret;
3861 ret = tsk->io_context;
3865 ret = kmem_cache_alloc_node(iocontext_cachep, gfp_flags, node);
3867 atomic_set(&ret->refcount, 1);
3868 ret->task = current;
3869 ret->ioprio_changed = 0;
3870 ret->last_waited = jiffies; /* doesn't matter... */
3871 ret->nr_batch_requests = 0; /* because this is 0 */
3873 ret->cic_root.rb_node = NULL;
3874 ret->ioc_data = NULL;
3875 /* make sure set_task_ioprio() sees the settings above */
3877 tsk->io_context = ret;
3884 * If the current task has no IO context then create one and initialise it.
3885 * If it does have a context, take a ref on it.
3887 * This is always called in the context of the task which submitted the I/O.
3889 struct io_context *get_io_context(gfp_t gfp_flags, int node)
3891 struct io_context *ret;
3892 ret = current_io_context(gfp_flags, node);
3894 atomic_inc(&ret->refcount);
3897 EXPORT_SYMBOL(get_io_context);
3899 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3901 struct io_context *src = *psrc;
3902 struct io_context *dst = *pdst;
3905 BUG_ON(atomic_read(&src->refcount) == 0);
3906 atomic_inc(&src->refcount);
3907 put_io_context(dst);
3911 EXPORT_SYMBOL(copy_io_context);
3913 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3915 struct io_context *temp;
3920 EXPORT_SYMBOL(swap_io_context);
3925 struct queue_sysfs_entry {
3926 struct attribute attr;
3927 ssize_t (*show)(struct request_queue *, char *);
3928 ssize_t (*store)(struct request_queue *, const char *, size_t);
3932 queue_var_show(unsigned int var, char *page)
3934 return sprintf(page, "%d\n", var);
3938 queue_var_store(unsigned long *var, const char *page, size_t count)
3940 char *p = (char *) page;
3942 *var = simple_strtoul(p, &p, 10);
3946 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3948 return queue_var_show(q->nr_requests, (page));
3952 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3954 struct request_list *rl = &q->rq;
3956 int ret = queue_var_store(&nr, page, count);
3957 if (nr < BLKDEV_MIN_RQ)
3960 spin_lock_irq(q->queue_lock);
3961 q->nr_requests = nr;
3962 blk_queue_congestion_threshold(q);
3964 if (rl->count[READ] >= queue_congestion_on_threshold(q))
3965 blk_set_queue_congested(q, READ);
3966 else if (rl->count[READ] < queue_congestion_off_threshold(q))
3967 blk_clear_queue_congested(q, READ);
3969 if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3970 blk_set_queue_congested(q, WRITE);
3971 else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3972 blk_clear_queue_congested(q, WRITE);
3974 if (rl->count[READ] >= q->nr_requests) {
3975 blk_set_queue_full(q, READ);
3976 } else if (rl->count[READ]+1 <= q->nr_requests) {
3977 blk_clear_queue_full(q, READ);
3978 wake_up(&rl->wait[READ]);
3981 if (rl->count[WRITE] >= q->nr_requests) {
3982 blk_set_queue_full(q, WRITE);
3983 } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3984 blk_clear_queue_full(q, WRITE);
3985 wake_up(&rl->wait[WRITE]);
3987 spin_unlock_irq(q->queue_lock);
3991 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3993 int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3995 return queue_var_show(ra_kb, (page));
3999 queue_ra_store(struct request_queue *q, const char *page, size_t count)
4001 unsigned long ra_kb;
4002 ssize_t ret = queue_var_store(&ra_kb, page, count);
4004 spin_lock_irq(q->queue_lock);
4005 q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
4006 spin_unlock_irq(q->queue_lock);
4011 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
4013 int max_sectors_kb = q->max_sectors >> 1;
4015 return queue_var_show(max_sectors_kb, (page));
4019 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
4021 unsigned long max_sectors_kb,
4022 max_hw_sectors_kb = q->max_hw_sectors >> 1,
4023 page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
4024 ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
4027 if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
4030 * Take the queue lock to update the readahead and max_sectors
4031 * values synchronously:
4033 spin_lock_irq(q->queue_lock);
4035 * Trim readahead window as well, if necessary:
4037 ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
4038 if (ra_kb > max_sectors_kb)
4039 q->backing_dev_info.ra_pages =
4040 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
4042 q->max_sectors = max_sectors_kb << 1;
4043 spin_unlock_irq(q->queue_lock);
4048 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
4050 int max_hw_sectors_kb = q->max_hw_sectors >> 1;
4052 return queue_var_show(max_hw_sectors_kb, (page));
4056 static struct queue_sysfs_entry queue_requests_entry = {
4057 .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
4058 .show = queue_requests_show,
4059 .store = queue_requests_store,
4062 static struct queue_sysfs_entry queue_ra_entry = {
4063 .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
4064 .show = queue_ra_show,
4065 .store = queue_ra_store,
4068 static struct queue_sysfs_entry queue_max_sectors_entry = {
4069 .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
4070 .show = queue_max_sectors_show,
4071 .store = queue_max_sectors_store,
4074 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
4075 .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
4076 .show = queue_max_hw_sectors_show,
4079 static struct queue_sysfs_entry queue_iosched_entry = {
4080 .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
4081 .show = elv_iosched_show,
4082 .store = elv_iosched_store,
4085 static struct attribute *default_attrs[] = {
4086 &queue_requests_entry.attr,
4087 &queue_ra_entry.attr,
4088 &queue_max_hw_sectors_entry.attr,
4089 &queue_max_sectors_entry.attr,
4090 &queue_iosched_entry.attr,
4094 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
4097 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
4099 struct queue_sysfs_entry *entry = to_queue(attr);
4100 struct request_queue *q =
4101 container_of(kobj, struct request_queue, kobj);
4106 mutex_lock(&q->sysfs_lock);
4107 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
4108 mutex_unlock(&q->sysfs_lock);
4111 res = entry->show(q, page);
4112 mutex_unlock(&q->sysfs_lock);
4117 queue_attr_store(struct kobject *kobj, struct attribute *attr,
4118 const char *page, size_t length)
4120 struct queue_sysfs_entry *entry = to_queue(attr);
4121 struct request_queue *q = container_of(kobj, struct request_queue, kobj);
4127 mutex_lock(&q->sysfs_lock);
4128 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
4129 mutex_unlock(&q->sysfs_lock);
4132 res = entry->store(q, page, length);
4133 mutex_unlock(&q->sysfs_lock);
4137 static struct sysfs_ops queue_sysfs_ops = {
4138 .show = queue_attr_show,
4139 .store = queue_attr_store,
4142 static struct kobj_type queue_ktype = {
4143 .sysfs_ops = &queue_sysfs_ops,
4144 .default_attrs = default_attrs,
4145 .release = blk_release_queue,
4148 int blk_register_queue(struct gendisk *disk)
4152 struct request_queue *q = disk->queue;
4154 if (!q || !q->request_fn)
4157 q->kobj.parent = kobject_get(&disk->kobj);
4159 ret = kobject_add(&q->kobj);
4163 kobject_uevent(&q->kobj, KOBJ_ADD);
4165 ret = elv_register_queue(q);
4167 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4168 kobject_del(&q->kobj);
4175 void blk_unregister_queue(struct gendisk *disk)
4177 struct request_queue *q = disk->queue;
4179 if (q && q->request_fn) {
4180 elv_unregister_queue(q);
4182 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4183 kobject_del(&q->kobj);
4184 kobject_put(&disk->kobj);