2 * linux/drivers/block/ll_rw_blk.c
4 * Copyright (C) 1991, 1992 Linus Torvalds
5 * Copyright (C) 1994, Karl Keyte: Added support for disk statistics
6 * Elevator latency, (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
7 * Queue request tables / lock, selectable elevator, Jens Axboe <axboe@suse.de>
8 * kernel-doc documentation started by NeilBrown <neilb@cse.unsw.edu.au> - July2000
9 * bio rewrite, highmem i/o, etc, Jens Axboe <axboe@suse.de> - may 2001
13 * This handles all read/write requests to block devices
15 #include <linux/config.h>
16 #include <linux/kernel.h>
17 #include <linux/module.h>
18 #include <linux/backing-dev.h>
19 #include <linux/bio.h>
20 #include <linux/blkdev.h>
21 #include <linux/highmem.h>
23 #include <linux/kernel_stat.h>
24 #include <linux/string.h>
25 #include <linux/init.h>
26 #include <linux/bootmem.h> /* for max_pfn/max_low_pfn */
27 #include <linux/completion.h>
28 #include <linux/slab.h>
29 #include <linux/swap.h>
30 #include <linux/writeback.h>
35 #include <scsi/scsi_cmnd.h>
37 static void blk_unplug_work(void *data);
38 static void blk_unplug_timeout(unsigned long data);
41 * For the allocated request tables
43 static kmem_cache_t *request_cachep;
46 * For queue allocation
48 static kmem_cache_t *requestq_cachep;
51 * For io context allocations
53 static kmem_cache_t *iocontext_cachep;
55 static wait_queue_head_t congestion_wqh[2] = {
56 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
57 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
61 * Controlling structure to kblockd
63 static struct workqueue_struct *kblockd_workqueue;
65 unsigned long blk_max_low_pfn, blk_max_pfn;
67 EXPORT_SYMBOL(blk_max_low_pfn);
68 EXPORT_SYMBOL(blk_max_pfn);
70 /* Amount of time in which a process may batch requests */
71 #define BLK_BATCH_TIME (HZ/50UL)
73 /* Number of requests a "batching" process may submit */
74 #define BLK_BATCH_REQ 32
77 * Return the threshold (number of used requests) at which the queue is
78 * considered to be congested. It include a little hysteresis to keep the
79 * context switch rate down.
81 static inline int queue_congestion_on_threshold(struct request_queue *q)
83 return q->nr_congestion_on;
87 * The threshold at which a queue is considered to be uncongested
89 static inline int queue_congestion_off_threshold(struct request_queue *q)
91 return q->nr_congestion_off;
94 static void blk_queue_congestion_threshold(struct request_queue *q)
98 nr = q->nr_requests - (q->nr_requests / 8) + 1;
99 if (nr > q->nr_requests)
101 q->nr_congestion_on = nr;
103 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
106 q->nr_congestion_off = nr;
110 * A queue has just exitted congestion. Note this in the global counter of
111 * congested queues, and wake up anyone who was waiting for requests to be
114 static void clear_queue_congested(request_queue_t *q, int rw)
117 wait_queue_head_t *wqh = &congestion_wqh[rw];
119 bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
120 clear_bit(bit, &q->backing_dev_info.state);
121 smp_mb__after_clear_bit();
122 if (waitqueue_active(wqh))
127 * A queue has just entered congestion. Flag that in the queue's VM-visible
128 * state flags and increment the global gounter of congested queues.
130 static void set_queue_congested(request_queue_t *q, int rw)
134 bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
135 set_bit(bit, &q->backing_dev_info.state);
139 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
142 * Locates the passed device's request queue and returns the address of its
145 * Will return NULL if the request queue cannot be located.
147 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
149 struct backing_dev_info *ret = NULL;
150 request_queue_t *q = bdev_get_queue(bdev);
153 ret = &q->backing_dev_info;
157 EXPORT_SYMBOL(blk_get_backing_dev_info);
159 void blk_queue_activity_fn(request_queue_t *q, activity_fn *fn, void *data)
162 q->activity_data = data;
165 EXPORT_SYMBOL(blk_queue_activity_fn);
168 * blk_queue_prep_rq - set a prepare_request function for queue
170 * @pfn: prepare_request function
172 * It's possible for a queue to register a prepare_request callback which
173 * is invoked before the request is handed to the request_fn. The goal of
174 * the function is to prepare a request for I/O, it can be used to build a
175 * cdb from the request data for instance.
178 void blk_queue_prep_rq(request_queue_t *q, prep_rq_fn *pfn)
183 EXPORT_SYMBOL(blk_queue_prep_rq);
186 * blk_queue_merge_bvec - set a merge_bvec function for queue
188 * @mbfn: merge_bvec_fn
190 * Usually queues have static limitations on the max sectors or segments that
191 * we can put in a request. Stacking drivers may have some settings that
192 * are dynamic, and thus we have to query the queue whether it is ok to
193 * add a new bio_vec to a bio at a given offset or not. If the block device
194 * has such limitations, it needs to register a merge_bvec_fn to control
195 * the size of bio's sent to it. Note that a block device *must* allow a
196 * single page to be added to an empty bio. The block device driver may want
197 * to use the bio_split() function to deal with these bio's. By default
198 * no merge_bvec_fn is defined for a queue, and only the fixed limits are
201 void blk_queue_merge_bvec(request_queue_t *q, merge_bvec_fn *mbfn)
203 q->merge_bvec_fn = mbfn;
206 EXPORT_SYMBOL(blk_queue_merge_bvec);
209 * blk_queue_make_request - define an alternate make_request function for a device
210 * @q: the request queue for the device to be affected
211 * @mfn: the alternate make_request function
214 * The normal way for &struct bios to be passed to a device
215 * driver is for them to be collected into requests on a request
216 * queue, and then to allow the device driver to select requests
217 * off that queue when it is ready. This works well for many block
218 * devices. However some block devices (typically virtual devices
219 * such as md or lvm) do not benefit from the processing on the
220 * request queue, and are served best by having the requests passed
221 * directly to them. This can be achieved by providing a function
222 * to blk_queue_make_request().
225 * The driver that does this *must* be able to deal appropriately
226 * with buffers in "highmemory". This can be accomplished by either calling
227 * __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
228 * blk_queue_bounce() to create a buffer in normal memory.
230 void blk_queue_make_request(request_queue_t * q, make_request_fn * mfn)
235 q->nr_requests = BLKDEV_MAX_RQ;
236 q->max_phys_segments = MAX_PHYS_SEGMENTS;
237 q->max_hw_segments = MAX_HW_SEGMENTS;
238 q->make_request_fn = mfn;
239 q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
240 q->backing_dev_info.state = 0;
241 q->backing_dev_info.memory_backed = 0;
242 blk_queue_max_sectors(q, MAX_SECTORS);
243 blk_queue_hardsect_size(q, 512);
244 blk_queue_dma_alignment(q, 511);
245 blk_queue_congestion_threshold(q);
246 q->nr_batching = BLK_BATCH_REQ;
248 q->unplug_thresh = 4; /* hmm */
249 q->unplug_delay = (3 * HZ) / 1000; /* 3 milliseconds */
250 if (q->unplug_delay == 0)
253 INIT_WORK(&q->unplug_work, blk_unplug_work, q);
255 q->unplug_timer.function = blk_unplug_timeout;
256 q->unplug_timer.data = (unsigned long)q;
259 * by default assume old behaviour and bounce for any highmem page
261 blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
263 blk_queue_activity_fn(q, NULL, NULL);
265 INIT_LIST_HEAD(&q->drain_list);
268 EXPORT_SYMBOL(blk_queue_make_request);
271 * blk_queue_ordered - does this queue support ordered writes
272 * @q: the request queue
276 * For journalled file systems, doing ordered writes on a commit
277 * block instead of explicitly doing wait_on_buffer (which is bad
278 * for performance) can be a big win. Block drivers supporting this
279 * feature should call this function and indicate so.
282 void blk_queue_ordered(request_queue_t *q, int flag)
285 set_bit(QUEUE_FLAG_ORDERED, &q->queue_flags);
287 clear_bit(QUEUE_FLAG_ORDERED, &q->queue_flags);
290 EXPORT_SYMBOL(blk_queue_ordered);
293 * blk_queue_issue_flush_fn - set function for issuing a flush
294 * @q: the request queue
295 * @iff: the function to be called issuing the flush
298 * If a driver supports issuing a flush command, the support is notified
299 * to the block layer by defining it through this call.
302 void blk_queue_issue_flush_fn(request_queue_t *q, issue_flush_fn *iff)
304 q->issue_flush_fn = iff;
307 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
310 * blk_queue_bounce_limit - set bounce buffer limit for queue
311 * @q: the request queue for the device
312 * @dma_addr: bus address limit
315 * Different hardware can have different requirements as to what pages
316 * it can do I/O directly to. A low level driver can call
317 * blk_queue_bounce_limit to have lower memory pages allocated as bounce
318 * buffers for doing I/O to pages residing above @page. By default
319 * the block layer sets this to the highest numbered "low" memory page.
321 void blk_queue_bounce_limit(request_queue_t *q, u64 dma_addr)
323 unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
326 * set appropriate bounce gfp mask -- unfortunately we don't have a
327 * full 4GB zone, so we have to resort to low memory for any bounces.
328 * ISA has its own < 16MB zone.
330 if (bounce_pfn < blk_max_low_pfn) {
331 BUG_ON(dma_addr < BLK_BOUNCE_ISA);
332 init_emergency_isa_pool();
333 q->bounce_gfp = GFP_NOIO | GFP_DMA;
335 q->bounce_gfp = GFP_NOIO;
337 q->bounce_pfn = bounce_pfn;
340 EXPORT_SYMBOL(blk_queue_bounce_limit);
343 * blk_queue_max_sectors - set max sectors for a request for this queue
344 * @q: the request queue for the device
345 * @max_sectors: max sectors in the usual 512b unit
348 * Enables a low level driver to set an upper limit on the size of
351 void blk_queue_max_sectors(request_queue_t *q, unsigned short max_sectors)
353 if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
354 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
355 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
358 q->max_sectors = q->max_hw_sectors = max_sectors;
361 EXPORT_SYMBOL(blk_queue_max_sectors);
364 * blk_queue_max_phys_segments - set max phys segments for a request for this queue
365 * @q: the request queue for the device
366 * @max_segments: max number of segments
369 * Enables a low level driver to set an upper limit on the number of
370 * physical data segments in a request. This would be the largest sized
371 * scatter list the driver could handle.
373 void blk_queue_max_phys_segments(request_queue_t *q, unsigned short max_segments)
377 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
380 q->max_phys_segments = max_segments;
383 EXPORT_SYMBOL(blk_queue_max_phys_segments);
386 * blk_queue_max_hw_segments - set max hw segments for a request for this queue
387 * @q: the request queue for the device
388 * @max_segments: max number of segments
391 * Enables a low level driver to set an upper limit on the number of
392 * hw data segments in a request. This would be the largest number of
393 * address/length pairs the host adapter can actually give as once
396 void blk_queue_max_hw_segments(request_queue_t *q, unsigned short max_segments)
400 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
403 q->max_hw_segments = max_segments;
406 EXPORT_SYMBOL(blk_queue_max_hw_segments);
409 * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
410 * @q: the request queue for the device
411 * @max_size: max size of segment in bytes
414 * Enables a low level driver to set an upper limit on the size of a
417 void blk_queue_max_segment_size(request_queue_t *q, unsigned int max_size)
419 if (max_size < PAGE_CACHE_SIZE) {
420 max_size = PAGE_CACHE_SIZE;
421 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
424 q->max_segment_size = max_size;
427 EXPORT_SYMBOL(blk_queue_max_segment_size);
430 * blk_queue_hardsect_size - set hardware sector size for the queue
431 * @q: the request queue for the device
432 * @size: the hardware sector size, in bytes
435 * This should typically be set to the lowest possible sector size
436 * that the hardware can operate on (possible without reverting to
437 * even internal read-modify-write operations). Usually the default
438 * of 512 covers most hardware.
440 void blk_queue_hardsect_size(request_queue_t *q, unsigned short size)
442 q->hardsect_size = size;
445 EXPORT_SYMBOL(blk_queue_hardsect_size);
448 * Returns the minimum that is _not_ zero, unless both are zero.
450 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
453 * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
454 * @t: the stacking driver (top)
455 * @b: the underlying device (bottom)
457 void blk_queue_stack_limits(request_queue_t *t, request_queue_t *b)
459 /* zero is "infinity" */
460 t->max_sectors = t->max_hw_sectors =
461 min_not_zero(t->max_sectors,b->max_sectors);
463 t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
464 t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
465 t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
466 t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
469 EXPORT_SYMBOL(blk_queue_stack_limits);
472 * blk_queue_segment_boundary - set boundary rules for segment merging
473 * @q: the request queue for the device
474 * @mask: the memory boundary mask
476 void blk_queue_segment_boundary(request_queue_t *q, unsigned long mask)
478 if (mask < PAGE_CACHE_SIZE - 1) {
479 mask = PAGE_CACHE_SIZE - 1;
480 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
483 q->seg_boundary_mask = mask;
486 EXPORT_SYMBOL(blk_queue_segment_boundary);
489 * blk_queue_dma_alignment - set dma length and memory alignment
490 * @q: the request queue for the device
491 * @mask: alignment mask
494 * set required memory and length aligment for direct dma transactions.
495 * this is used when buiding direct io requests for the queue.
498 void blk_queue_dma_alignment(request_queue_t *q, int mask)
500 q->dma_alignment = mask;
503 EXPORT_SYMBOL(blk_queue_dma_alignment);
506 * blk_queue_find_tag - find a request by its tag and queue
508 * @q: The request queue for the device
509 * @tag: The tag of the request
512 * Should be used when a device returns a tag and you want to match
515 * no locks need be held.
517 struct request *blk_queue_find_tag(request_queue_t *q, int tag)
519 struct blk_queue_tag *bqt = q->queue_tags;
521 if (unlikely(bqt == NULL || tag >= bqt->real_max_depth))
524 return bqt->tag_index[tag];
527 EXPORT_SYMBOL(blk_queue_find_tag);
530 * __blk_queue_free_tags - release tag maintenance info
531 * @q: the request queue for the device
534 * blk_cleanup_queue() will take care of calling this function, if tagging
535 * has been used. So there's no need to call this directly.
537 static void __blk_queue_free_tags(request_queue_t *q)
539 struct blk_queue_tag *bqt = q->queue_tags;
544 if (atomic_dec_and_test(&bqt->refcnt)) {
546 BUG_ON(!list_empty(&bqt->busy_list));
548 kfree(bqt->tag_index);
549 bqt->tag_index = NULL;
557 q->queue_tags = NULL;
558 q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
562 * blk_queue_free_tags - release tag maintenance info
563 * @q: the request queue for the device
566 * This is used to disabled tagged queuing to a device, yet leave
569 void blk_queue_free_tags(request_queue_t *q)
571 clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
574 EXPORT_SYMBOL(blk_queue_free_tags);
577 init_tag_map(request_queue_t *q, struct blk_queue_tag *tags, int depth)
580 struct request **tag_index;
581 unsigned long *tag_map;
583 if (depth > q->nr_requests * 2) {
584 depth = q->nr_requests * 2;
585 printk(KERN_ERR "%s: adjusted depth to %d\n",
586 __FUNCTION__, depth);
589 tag_index = kmalloc(depth * sizeof(struct request *), GFP_ATOMIC);
593 bits = (depth / BLK_TAGS_PER_LONG) + 1;
594 tag_map = kmalloc(bits * sizeof(unsigned long), GFP_ATOMIC);
598 memset(tag_index, 0, depth * sizeof(struct request *));
599 memset(tag_map, 0, bits * sizeof(unsigned long));
600 tags->max_depth = depth;
601 tags->real_max_depth = bits * BITS_PER_LONG;
602 tags->tag_index = tag_index;
603 tags->tag_map = tag_map;
606 * set the upper bits if the depth isn't a multiple of the word size
608 for (i = depth; i < bits * BLK_TAGS_PER_LONG; i++)
609 __set_bit(i, tag_map);
618 * blk_queue_init_tags - initialize the queue tag info
619 * @q: the request queue for the device
620 * @depth: the maximum queue depth supported
622 int blk_queue_init_tags(request_queue_t *q, int depth,
623 struct blk_queue_tag *tags)
627 BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
629 if (!tags && !q->queue_tags) {
630 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
634 if (init_tag_map(q, tags, depth))
637 INIT_LIST_HEAD(&tags->busy_list);
639 atomic_set(&tags->refcnt, 1);
640 } else if (q->queue_tags) {
641 if ((rc = blk_queue_resize_tags(q, depth)))
643 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
646 atomic_inc(&tags->refcnt);
649 * assign it, all done
651 q->queue_tags = tags;
652 q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
659 EXPORT_SYMBOL(blk_queue_init_tags);
662 * blk_queue_resize_tags - change the queueing depth
663 * @q: the request queue for the device
664 * @new_depth: the new max command queueing depth
667 * Must be called with the queue lock held.
669 int blk_queue_resize_tags(request_queue_t *q, int new_depth)
671 struct blk_queue_tag *bqt = q->queue_tags;
672 struct request **tag_index;
673 unsigned long *tag_map;
680 * don't bother sizing down
682 if (new_depth <= bqt->real_max_depth) {
683 bqt->max_depth = new_depth;
688 * save the old state info, so we can copy it back
690 tag_index = bqt->tag_index;
691 tag_map = bqt->tag_map;
692 max_depth = bqt->real_max_depth;
694 if (init_tag_map(q, bqt, new_depth))
697 memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
698 bits = max_depth / BLK_TAGS_PER_LONG;
699 memcpy(bqt->tag_map, tag_map, bits * sizeof(unsigned long));
706 EXPORT_SYMBOL(blk_queue_resize_tags);
709 * blk_queue_end_tag - end tag operations for a request
710 * @q: the request queue for the device
711 * @rq: the request that has completed
714 * Typically called when end_that_request_first() returns 0, meaning
715 * all transfers have been done for a request. It's important to call
716 * this function before end_that_request_last(), as that will put the
717 * request back on the free list thus corrupting the internal tag list.
720 * queue lock must be held.
722 void blk_queue_end_tag(request_queue_t *q, struct request *rq)
724 struct blk_queue_tag *bqt = q->queue_tags;
729 if (unlikely(tag >= bqt->real_max_depth))
732 if (unlikely(!__test_and_clear_bit(tag, bqt->tag_map))) {
733 printk("attempt to clear non-busy tag (%d)\n", tag);
737 list_del_init(&rq->queuelist);
738 rq->flags &= ~REQ_QUEUED;
741 if (unlikely(bqt->tag_index[tag] == NULL))
742 printk("tag %d is missing\n", tag);
744 bqt->tag_index[tag] = NULL;
748 EXPORT_SYMBOL(blk_queue_end_tag);
751 * blk_queue_start_tag - find a free tag and assign it
752 * @q: the request queue for the device
753 * @rq: the block request that needs tagging
756 * This can either be used as a stand-alone helper, or possibly be
757 * assigned as the queue &prep_rq_fn (in which case &struct request
758 * automagically gets a tag assigned). Note that this function
759 * assumes that any type of request can be queued! if this is not
760 * true for your device, you must check the request type before
761 * calling this function. The request will also be removed from
762 * the request queue, so it's the drivers responsibility to readd
763 * it if it should need to be restarted for some reason.
766 * queue lock must be held.
768 int blk_queue_start_tag(request_queue_t *q, struct request *rq)
770 struct blk_queue_tag *bqt = q->queue_tags;
771 unsigned long *map = bqt->tag_map;
774 if (unlikely((rq->flags & REQ_QUEUED))) {
776 "request %p for device [%s] already tagged %d",
777 rq, rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
781 for (map = bqt->tag_map; *map == -1UL; map++) {
782 tag += BLK_TAGS_PER_LONG;
784 if (tag >= bqt->max_depth)
789 __set_bit(tag, bqt->tag_map);
791 rq->flags |= REQ_QUEUED;
793 bqt->tag_index[tag] = rq;
794 blkdev_dequeue_request(rq);
795 list_add(&rq->queuelist, &bqt->busy_list);
800 EXPORT_SYMBOL(blk_queue_start_tag);
803 * blk_queue_invalidate_tags - invalidate all pending tags
804 * @q: the request queue for the device
807 * Hardware conditions may dictate a need to stop all pending requests.
808 * In this case, we will safely clear the block side of the tag queue and
809 * readd all requests to the request queue in the right order.
812 * queue lock must be held.
814 void blk_queue_invalidate_tags(request_queue_t *q)
816 struct blk_queue_tag *bqt = q->queue_tags;
817 struct list_head *tmp, *n;
820 list_for_each_safe(tmp, n, &bqt->busy_list) {
821 rq = list_entry_rq(tmp);
824 printk("bad tag found on list\n");
825 list_del_init(&rq->queuelist);
826 rq->flags &= ~REQ_QUEUED;
828 blk_queue_end_tag(q, rq);
830 rq->flags &= ~REQ_STARTED;
831 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
835 EXPORT_SYMBOL(blk_queue_invalidate_tags);
837 static char *rq_flags[] = {
855 "REQ_DRIVE_TASKFILE",
862 void blk_dump_rq_flags(struct request *rq, char *msg)
866 printk("%s: dev %s: flags = ", msg,
867 rq->rq_disk ? rq->rq_disk->disk_name : "?");
870 if (rq->flags & (1 << bit))
871 printk("%s ", rq_flags[bit]);
873 } while (bit < __REQ_NR_BITS);
875 printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
877 rq->current_nr_sectors);
878 printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
880 if (rq->flags & (REQ_BLOCK_PC | REQ_PC)) {
882 for (bit = 0; bit < sizeof(rq->cmd); bit++)
883 printk("%02x ", rq->cmd[bit]);
888 EXPORT_SYMBOL(blk_dump_rq_flags);
890 void blk_recount_segments(request_queue_t *q, struct bio *bio)
892 struct bio_vec *bv, *bvprv = NULL;
893 int i, nr_phys_segs, nr_hw_segs, seg_size, hw_seg_size, cluster;
894 int high, highprv = 1;
896 if (unlikely(!bio->bi_io_vec))
899 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
900 hw_seg_size = seg_size = nr_phys_segs = nr_hw_segs = 0;
901 bio_for_each_segment(bv, bio, i) {
903 * the trick here is making sure that a high page is never
904 * considered part of another segment, since that might
905 * change with the bounce page.
907 high = page_to_pfn(bv->bv_page) >= q->bounce_pfn;
911 if (seg_size + bv->bv_len > q->max_segment_size)
913 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
915 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
917 if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
920 seg_size += bv->bv_len;
921 hw_seg_size += bv->bv_len;
926 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
927 !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len)) {
928 hw_seg_size += bv->bv_len;
931 if (hw_seg_size > bio->bi_hw_front_size)
932 bio->bi_hw_front_size = hw_seg_size;
933 hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
939 seg_size = bv->bv_len;
942 if (hw_seg_size > bio->bi_hw_back_size)
943 bio->bi_hw_back_size = hw_seg_size;
944 if (nr_hw_segs == 1 && hw_seg_size > bio->bi_hw_front_size)
945 bio->bi_hw_front_size = hw_seg_size;
946 bio->bi_phys_segments = nr_phys_segs;
947 bio->bi_hw_segments = nr_hw_segs;
948 bio->bi_flags |= (1 << BIO_SEG_VALID);
952 int blk_phys_contig_segment(request_queue_t *q, struct bio *bio,
955 if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
958 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
960 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
964 * bio and nxt are contigous in memory, check if the queue allows
965 * these two to be merged into one
967 if (BIO_SEG_BOUNDARY(q, bio, nxt))
973 EXPORT_SYMBOL(blk_phys_contig_segment);
975 int blk_hw_contig_segment(request_queue_t *q, struct bio *bio,
978 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
979 blk_recount_segments(q, bio);
980 if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
981 blk_recount_segments(q, nxt);
982 if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
983 BIOVEC_VIRT_OVERSIZE(bio->bi_hw_front_size + bio->bi_hw_back_size))
985 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
991 EXPORT_SYMBOL(blk_hw_contig_segment);
994 * map a request to scatterlist, return number of sg entries setup. Caller
995 * must make sure sg can hold rq->nr_phys_segments entries
997 int blk_rq_map_sg(request_queue_t *q, struct request *rq, struct scatterlist *sg)
999 struct bio_vec *bvec, *bvprv;
1001 int nsegs, i, cluster;
1004 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1007 * for each bio in rq
1010 rq_for_each_bio(bio, rq) {
1012 * for each segment in bio
1014 bio_for_each_segment(bvec, bio, i) {
1015 int nbytes = bvec->bv_len;
1017 if (bvprv && cluster) {
1018 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1021 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1023 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1026 sg[nsegs - 1].length += nbytes;
1029 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1030 sg[nsegs].page = bvec->bv_page;
1031 sg[nsegs].length = nbytes;
1032 sg[nsegs].offset = bvec->bv_offset;
1037 } /* segments in bio */
1043 EXPORT_SYMBOL(blk_rq_map_sg);
1046 * the standard queue merge functions, can be overridden with device
1047 * specific ones if so desired
1050 static inline int ll_new_mergeable(request_queue_t *q,
1051 struct request *req,
1054 int nr_phys_segs = bio_phys_segments(q, bio);
1056 if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1057 req->flags |= REQ_NOMERGE;
1058 if (req == q->last_merge)
1059 q->last_merge = NULL;
1064 * A hw segment is just getting larger, bump just the phys
1067 req->nr_phys_segments += nr_phys_segs;
1071 static inline int ll_new_hw_segment(request_queue_t *q,
1072 struct request *req,
1075 int nr_hw_segs = bio_hw_segments(q, bio);
1076 int nr_phys_segs = bio_phys_segments(q, bio);
1078 if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1079 || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1080 req->flags |= REQ_NOMERGE;
1081 if (req == q->last_merge)
1082 q->last_merge = NULL;
1087 * This will form the start of a new hw segment. Bump both
1090 req->nr_hw_segments += nr_hw_segs;
1091 req->nr_phys_segments += nr_phys_segs;
1095 static int ll_back_merge_fn(request_queue_t *q, struct request *req,
1100 if (req->nr_sectors + bio_sectors(bio) > q->max_sectors) {
1101 req->flags |= REQ_NOMERGE;
1102 if (req == q->last_merge)
1103 q->last_merge = NULL;
1106 if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1107 blk_recount_segments(q, req->biotail);
1108 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1109 blk_recount_segments(q, bio);
1110 len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1111 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1112 !BIOVEC_VIRT_OVERSIZE(len)) {
1113 int mergeable = ll_new_mergeable(q, req, bio);
1116 if (req->nr_hw_segments == 1)
1117 req->bio->bi_hw_front_size = len;
1118 if (bio->bi_hw_segments == 1)
1119 bio->bi_hw_back_size = len;
1124 return ll_new_hw_segment(q, req, bio);
1127 static int ll_front_merge_fn(request_queue_t *q, struct request *req,
1132 if (req->nr_sectors + bio_sectors(bio) > q->max_sectors) {
1133 req->flags |= REQ_NOMERGE;
1134 if (req == q->last_merge)
1135 q->last_merge = NULL;
1138 len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1139 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1140 blk_recount_segments(q, bio);
1141 if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1142 blk_recount_segments(q, req->bio);
1143 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1144 !BIOVEC_VIRT_OVERSIZE(len)) {
1145 int mergeable = ll_new_mergeable(q, req, bio);
1148 if (bio->bi_hw_segments == 1)
1149 bio->bi_hw_front_size = len;
1150 if (req->nr_hw_segments == 1)
1151 req->biotail->bi_hw_back_size = len;
1156 return ll_new_hw_segment(q, req, bio);
1159 static int ll_merge_requests_fn(request_queue_t *q, struct request *req,
1160 struct request *next)
1162 int total_phys_segments = req->nr_phys_segments +next->nr_phys_segments;
1163 int total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1166 * First check if the either of the requests are re-queued
1167 * requests. Can't merge them if they are.
1169 if (req->special || next->special)
1173 * Will it become to large?
1175 if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1178 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1179 if (blk_phys_contig_segment(q, req->biotail, next->bio))
1180 total_phys_segments--;
1182 if (total_phys_segments > q->max_phys_segments)
1185 total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1186 if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1187 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1189 * propagate the combined length to the end of the requests
1191 if (req->nr_hw_segments == 1)
1192 req->bio->bi_hw_front_size = len;
1193 if (next->nr_hw_segments == 1)
1194 next->biotail->bi_hw_back_size = len;
1195 total_hw_segments--;
1198 if (total_hw_segments > q->max_hw_segments)
1201 /* Merge is OK... */
1202 req->nr_phys_segments = total_phys_segments;
1203 req->nr_hw_segments = total_hw_segments;
1208 * "plug" the device if there are no outstanding requests: this will
1209 * force the transfer to start only after we have put all the requests
1212 * This is called with interrupts off and no requests on the queue and
1213 * with the queue lock held.
1215 void blk_plug_device(request_queue_t *q)
1217 WARN_ON(!irqs_disabled());
1220 * don't plug a stopped queue, it must be paired with blk_start_queue()
1221 * which will restart the queueing
1223 if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
1226 if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1227 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1230 EXPORT_SYMBOL(blk_plug_device);
1233 * remove the queue from the plugged list, if present. called with
1234 * queue lock held and interrupts disabled.
1236 int blk_remove_plug(request_queue_t *q)
1238 WARN_ON(!irqs_disabled());
1240 if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1243 del_timer(&q->unplug_timer);
1247 EXPORT_SYMBOL(blk_remove_plug);
1250 * remove the plug and let it rip..
1252 void __generic_unplug_device(request_queue_t *q)
1254 if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
1257 if (!blk_remove_plug(q))
1261 * was plugged, fire request_fn if queue has stuff to do
1263 if (elv_next_request(q))
1266 EXPORT_SYMBOL(__generic_unplug_device);
1269 * generic_unplug_device - fire a request queue
1270 * @q: The &request_queue_t in question
1273 * Linux uses plugging to build bigger requests queues before letting
1274 * the device have at them. If a queue is plugged, the I/O scheduler
1275 * is still adding and merging requests on the queue. Once the queue
1276 * gets unplugged, the request_fn defined for the queue is invoked and
1277 * transfers started.
1279 void generic_unplug_device(request_queue_t *q)
1281 spin_lock_irq(q->queue_lock);
1282 __generic_unplug_device(q);
1283 spin_unlock_irq(q->queue_lock);
1285 EXPORT_SYMBOL(generic_unplug_device);
1287 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1290 request_queue_t *q = bdi->unplug_io_data;
1293 * devices don't necessarily have an ->unplug_fn defined
1299 static void blk_unplug_work(void *data)
1301 request_queue_t *q = data;
1306 static void blk_unplug_timeout(unsigned long data)
1308 request_queue_t *q = (request_queue_t *)data;
1310 kblockd_schedule_work(&q->unplug_work);
1314 * blk_start_queue - restart a previously stopped queue
1315 * @q: The &request_queue_t in question
1318 * blk_start_queue() will clear the stop flag on the queue, and call
1319 * the request_fn for the queue if it was in a stopped state when
1320 * entered. Also see blk_stop_queue(). Queue lock must be held.
1322 void blk_start_queue(request_queue_t *q)
1324 clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1327 * one level of recursion is ok and is much faster than kicking
1328 * the unplug handling
1330 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1332 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1335 kblockd_schedule_work(&q->unplug_work);
1339 EXPORT_SYMBOL(blk_start_queue);
1342 * blk_stop_queue - stop a queue
1343 * @q: The &request_queue_t in question
1346 * The Linux block layer assumes that a block driver will consume all
1347 * entries on the request queue when the request_fn strategy is called.
1348 * Often this will not happen, because of hardware limitations (queue
1349 * depth settings). If a device driver gets a 'queue full' response,
1350 * or if it simply chooses not to queue more I/O at one point, it can
1351 * call this function to prevent the request_fn from being called until
1352 * the driver has signalled it's ready to go again. This happens by calling
1353 * blk_start_queue() to restart queue operations. Queue lock must be held.
1355 void blk_stop_queue(request_queue_t *q)
1358 set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1360 EXPORT_SYMBOL(blk_stop_queue);
1363 * blk_sync_queue - cancel any pending callbacks on a queue
1367 * The block layer may perform asynchronous callback activity
1368 * on a queue, such as calling the unplug function after a timeout.
1369 * A block device may call blk_sync_queue to ensure that any
1370 * such activity is cancelled, thus allowing it to release resources
1371 * the the callbacks might use. The caller must already have made sure
1372 * that its ->make_request_fn will not re-add plugging prior to calling
1376 void blk_sync_queue(struct request_queue *q)
1378 del_timer_sync(&q->unplug_timer);
1381 EXPORT_SYMBOL(blk_sync_queue);
1384 * blk_run_queue - run a single device queue
1385 * @q: The queue to run
1387 void blk_run_queue(struct request_queue *q)
1389 unsigned long flags;
1391 spin_lock_irqsave(q->queue_lock, flags);
1394 spin_unlock_irqrestore(q->queue_lock, flags);
1396 EXPORT_SYMBOL(blk_run_queue);
1399 * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1400 * @q: the request queue to be released
1403 * blk_cleanup_queue is the pair to blk_init_queue() or
1404 * blk_queue_make_request(). It should be called when a request queue is
1405 * being released; typically when a block device is being de-registered.
1406 * Currently, its primary task it to free all the &struct request
1407 * structures that were allocated to the queue and the queue itself.
1410 * Hopefully the low level driver will have finished any
1411 * outstanding requests first...
1413 void blk_cleanup_queue(request_queue_t * q)
1415 struct request_list *rl = &q->rq;
1417 if (!atomic_dec_and_test(&q->refcnt))
1421 elevator_exit(q->elevator);
1426 mempool_destroy(rl->rq_pool);
1429 __blk_queue_free_tags(q);
1431 kmem_cache_free(requestq_cachep, q);
1434 EXPORT_SYMBOL(blk_cleanup_queue);
1436 static int blk_init_free_list(request_queue_t *q)
1438 struct request_list *rl = &q->rq;
1440 rl->count[READ] = rl->count[WRITE] = 0;
1441 init_waitqueue_head(&rl->wait[READ]);
1442 init_waitqueue_head(&rl->wait[WRITE]);
1443 init_waitqueue_head(&rl->drain);
1445 rl->rq_pool = mempool_create(BLKDEV_MIN_RQ, mempool_alloc_slab, mempool_free_slab, request_cachep);
1453 static int __make_request(request_queue_t *, struct bio *);
1455 request_queue_t *blk_alloc_queue(int gfp_mask)
1457 request_queue_t *q = kmem_cache_alloc(requestq_cachep, gfp_mask);
1462 memset(q, 0, sizeof(*q));
1463 init_timer(&q->unplug_timer);
1464 atomic_set(&q->refcnt, 1);
1466 q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1467 q->backing_dev_info.unplug_io_data = q;
1472 EXPORT_SYMBOL(blk_alloc_queue);
1475 * blk_init_queue - prepare a request queue for use with a block device
1476 * @rfn: The function to be called to process requests that have been
1477 * placed on the queue.
1478 * @lock: Request queue spin lock
1481 * If a block device wishes to use the standard request handling procedures,
1482 * which sorts requests and coalesces adjacent requests, then it must
1483 * call blk_init_queue(). The function @rfn will be called when there
1484 * are requests on the queue that need to be processed. If the device
1485 * supports plugging, then @rfn may not be called immediately when requests
1486 * are available on the queue, but may be called at some time later instead.
1487 * Plugged queues are generally unplugged when a buffer belonging to one
1488 * of the requests on the queue is needed, or due to memory pressure.
1490 * @rfn is not required, or even expected, to remove all requests off the
1491 * queue, but only as many as it can handle at a time. If it does leave
1492 * requests on the queue, it is responsible for arranging that the requests
1493 * get dealt with eventually.
1495 * The queue spin lock must be held while manipulating the requests on the
1498 * Function returns a pointer to the initialized request queue, or NULL if
1499 * it didn't succeed.
1502 * blk_init_queue() must be paired with a blk_cleanup_queue() call
1503 * when the block device is deactivated (such as at module unload).
1505 request_queue_t *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1507 request_queue_t *q = blk_alloc_queue(GFP_KERNEL);
1512 if (blk_init_free_list(q))
1515 q->request_fn = rfn;
1516 q->back_merge_fn = ll_back_merge_fn;
1517 q->front_merge_fn = ll_front_merge_fn;
1518 q->merge_requests_fn = ll_merge_requests_fn;
1519 q->prep_rq_fn = NULL;
1520 q->unplug_fn = generic_unplug_device;
1521 q->queue_flags = (1 << QUEUE_FLAG_CLUSTER);
1522 q->queue_lock = lock;
1524 blk_queue_segment_boundary(q, 0xffffffff);
1526 blk_queue_make_request(q, __make_request);
1527 blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1529 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1530 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1535 if (!elevator_init(q, NULL)) {
1536 blk_queue_congestion_threshold(q);
1540 blk_cleanup_queue(q);
1542 kmem_cache_free(requestq_cachep, q);
1546 EXPORT_SYMBOL(blk_init_queue);
1548 int blk_get_queue(request_queue_t *q)
1550 if (!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
1551 atomic_inc(&q->refcnt);
1558 EXPORT_SYMBOL(blk_get_queue);
1560 static inline void blk_free_request(request_queue_t *q, struct request *rq)
1562 elv_put_request(q, rq);
1563 mempool_free(rq, q->rq.rq_pool);
1566 static inline struct request *blk_alloc_request(request_queue_t *q, int rw,
1569 struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1575 * first three bits are identical in rq->flags and bio->bi_rw,
1576 * see bio.h and blkdev.h
1580 if (!elv_set_request(q, rq, gfp_mask))
1583 mempool_free(rq, q->rq.rq_pool);
1588 * ioc_batching returns true if the ioc is a valid batching request and
1589 * should be given priority access to a request.
1591 static inline int ioc_batching(request_queue_t *q, struct io_context *ioc)
1597 * Make sure the process is able to allocate at least 1 request
1598 * even if the batch times out, otherwise we could theoretically
1601 return ioc->nr_batch_requests == q->nr_batching ||
1602 (ioc->nr_batch_requests > 0
1603 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
1607 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
1608 * will cause the process to be a "batcher" on all queues in the system. This
1609 * is the behaviour we want though - once it gets a wakeup it should be given
1612 void ioc_set_batching(request_queue_t *q, struct io_context *ioc)
1614 if (!ioc || ioc_batching(q, ioc))
1617 ioc->nr_batch_requests = q->nr_batching;
1618 ioc->last_waited = jiffies;
1622 * A request has just been released. Account for it, update the full and
1623 * congestion status, wake up any waiters. Called under q->queue_lock.
1625 static void freed_request(request_queue_t *q, int rw)
1627 struct request_list *rl = &q->rq;
1630 if (rl->count[rw] < queue_congestion_off_threshold(q))
1631 clear_queue_congested(q, rw);
1632 if (rl->count[rw]+1 <= q->nr_requests) {
1634 if (waitqueue_active(&rl->wait[rw]))
1635 wake_up(&rl->wait[rw]);
1636 blk_clear_queue_full(q, rw);
1638 if (unlikely(waitqueue_active(&rl->drain)) &&
1639 !rl->count[READ] && !rl->count[WRITE])
1640 wake_up(&rl->drain);
1643 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
1645 * Get a free request, queue_lock must not be held
1647 static struct request *get_request(request_queue_t *q, int rw, int gfp_mask)
1649 struct request *rq = NULL;
1650 struct request_list *rl = &q->rq;
1651 struct io_context *ioc = get_io_context(gfp_mask);
1653 if (unlikely(test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags)))
1656 spin_lock_irq(q->queue_lock);
1657 if (rl->count[rw]+1 >= q->nr_requests) {
1659 * The queue will fill after this allocation, so set it as
1660 * full, and mark this process as "batching". This process
1661 * will be allowed to complete a batch of requests, others
1664 if (!blk_queue_full(q, rw)) {
1665 ioc_set_batching(q, ioc);
1666 blk_set_queue_full(q, rw);
1670 switch (elv_may_queue(q, rw)) {
1672 spin_unlock_irq(q->queue_lock);
1674 case ELV_MQUEUE_MAY:
1676 case ELV_MQUEUE_MUST:
1680 if (blk_queue_full(q, rw) && !ioc_batching(q, ioc)) {
1682 * The queue is full and the allocating process is not a
1683 * "batcher", and not exempted by the IO scheduler
1685 spin_unlock_irq(q->queue_lock);
1691 if (rl->count[rw] >= queue_congestion_on_threshold(q))
1692 set_queue_congested(q, rw);
1693 spin_unlock_irq(q->queue_lock);
1695 rq = blk_alloc_request(q, rw, gfp_mask);
1698 * Allocation failed presumably due to memory. Undo anything
1699 * we might have messed up.
1701 * Allocating task should really be put onto the front of the
1702 * wait queue, but this is pretty rare.
1704 spin_lock_irq(q->queue_lock);
1705 freed_request(q, rw);
1706 spin_unlock_irq(q->queue_lock);
1710 if (ioc_batching(q, ioc))
1711 ioc->nr_batch_requests--;
1713 INIT_LIST_HEAD(&rq->queuelist);
1716 rq->rq_status = RQ_ACTIVE;
1717 rq->bio = rq->biotail = NULL;
1729 put_io_context(ioc);
1734 * No available requests for this queue, unplug the device and wait for some
1735 * requests to become available.
1737 static struct request *get_request_wait(request_queue_t *q, int rw)
1742 generic_unplug_device(q);
1744 struct request_list *rl = &q->rq;
1746 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
1747 TASK_UNINTERRUPTIBLE);
1749 rq = get_request(q, rw, GFP_NOIO);
1752 struct io_context *ioc;
1757 * After sleeping, we become a "batching" process and
1758 * will be able to allocate at least one request, and
1759 * up to a big batch of them for a small period time.
1760 * See ioc_batching, ioc_set_batching
1762 ioc = get_io_context(GFP_NOIO);
1763 ioc_set_batching(q, ioc);
1764 put_io_context(ioc);
1766 finish_wait(&rl->wait[rw], &wait);
1772 struct request *blk_get_request(request_queue_t *q, int rw, int gfp_mask)
1776 BUG_ON(rw != READ && rw != WRITE);
1778 if (gfp_mask & __GFP_WAIT)
1779 rq = get_request_wait(q, rw);
1781 rq = get_request(q, rw, gfp_mask);
1786 EXPORT_SYMBOL(blk_get_request);
1789 * blk_requeue_request - put a request back on queue
1790 * @q: request queue where request should be inserted
1791 * @rq: request to be inserted
1794 * Drivers often keep queueing requests until the hardware cannot accept
1795 * more, when that condition happens we need to put the request back
1796 * on the queue. Must be called with queue lock held.
1798 void blk_requeue_request(request_queue_t *q, struct request *rq)
1800 if (blk_rq_tagged(rq))
1801 blk_queue_end_tag(q, rq);
1803 elv_requeue_request(q, rq);
1806 EXPORT_SYMBOL(blk_requeue_request);
1809 * blk_insert_request - insert a special request in to a request queue
1810 * @q: request queue where request should be inserted
1811 * @rq: request to be inserted
1812 * @at_head: insert request at head or tail of queue
1813 * @data: private data
1814 * @reinsert: true if request it a reinsertion of previously processed one
1817 * Many block devices need to execute commands asynchronously, so they don't
1818 * block the whole kernel from preemption during request execution. This is
1819 * accomplished normally by inserting aritficial requests tagged as
1820 * REQ_SPECIAL in to the corresponding request queue, and letting them be
1821 * scheduled for actual execution by the request queue.
1823 * We have the option of inserting the head or the tail of the queue.
1824 * Typically we use the tail for new ioctls and so forth. We use the head
1825 * of the queue for things like a QUEUE_FULL message from a device, or a
1826 * host that is unable to accept a particular command.
1828 void blk_insert_request(request_queue_t *q, struct request *rq,
1829 int at_head, void *data, int reinsert)
1831 unsigned long flags;
1834 * tell I/O scheduler that this isn't a regular read/write (ie it
1835 * must not attempt merges on this) and that it acts as a soft
1838 rq->flags |= REQ_SPECIAL | REQ_SOFTBARRIER;
1842 spin_lock_irqsave(q->queue_lock, flags);
1845 * If command is tagged, release the tag
1848 blk_requeue_request(q, rq);
1850 int where = ELEVATOR_INSERT_BACK;
1853 where = ELEVATOR_INSERT_FRONT;
1855 if (blk_rq_tagged(rq))
1856 blk_queue_end_tag(q, rq);
1858 drive_stat_acct(rq, rq->nr_sectors, 1);
1859 __elv_add_request(q, rq, where, 0);
1861 if (blk_queue_plugged(q))
1862 __generic_unplug_device(q);
1865 spin_unlock_irqrestore(q->queue_lock, flags);
1868 EXPORT_SYMBOL(blk_insert_request);
1871 * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
1872 * @q: request queue where request should be inserted
1873 * @rw: READ or WRITE data
1874 * @ubuf: the user buffer
1875 * @len: length of user data
1878 * Data will be mapped directly for zero copy io, if possible. Otherwise
1879 * a kernel bounce buffer is used.
1881 * A matching blk_rq_unmap_user() must be issued at the end of io, while
1882 * still in process context.
1884 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
1885 * before being submitted to the device, as pages mapped may be out of
1886 * reach. It's the callers responsibility to make sure this happens. The
1887 * original bio must be passed back in to blk_rq_unmap_user() for proper
1890 struct request *blk_rq_map_user(request_queue_t *q, int rw, void __user *ubuf,
1893 unsigned long uaddr;
1897 if (len > (q->max_sectors << 9))
1898 return ERR_PTR(-EINVAL);
1899 if ((!len && ubuf) || (len && !ubuf))
1900 return ERR_PTR(-EINVAL);
1902 rq = blk_get_request(q, rw, __GFP_WAIT);
1904 return ERR_PTR(-ENOMEM);
1907 * if alignment requirement is satisfied, map in user pages for
1908 * direct dma. else, set up kernel bounce buffers
1910 uaddr = (unsigned long) ubuf;
1911 if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
1912 bio = bio_map_user(q, NULL, uaddr, len, rw == READ);
1914 bio = bio_copy_user(q, uaddr, len, rw == READ);
1917 rq->bio = rq->biotail = bio;
1918 blk_rq_bio_prep(q, rq, bio);
1920 rq->buffer = rq->data = NULL;
1926 * bio is the err-ptr
1928 blk_put_request(rq);
1929 return (struct request *) bio;
1932 EXPORT_SYMBOL(blk_rq_map_user);
1935 * blk_rq_unmap_user - unmap a request with user data
1936 * @rq: request to be unmapped
1937 * @ubuf: user buffer
1938 * @ulen: length of user buffer
1941 * Unmap a request previously mapped by blk_rq_map_user().
1943 int blk_rq_unmap_user(struct request *rq, struct bio *bio, unsigned int ulen)
1948 if (bio_flagged(bio, BIO_USER_MAPPED))
1949 bio_unmap_user(bio);
1951 ret = bio_uncopy_user(bio);
1954 blk_put_request(rq);
1958 EXPORT_SYMBOL(blk_rq_unmap_user);
1961 * blk_execute_rq - insert a request into queue for execution
1962 * @q: queue to insert the request in
1963 * @bd_disk: matching gendisk
1964 * @rq: request to insert
1967 * Insert a fully prepared request at the back of the io scheduler queue
1970 int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
1973 DECLARE_COMPLETION(wait);
1974 char sense[SCSI_SENSE_BUFFERSIZE];
1977 rq->rq_disk = bd_disk;
1980 * we need an extra reference to the request, so we can look at
1981 * it after io completion
1986 memset(sense, 0, sizeof(sense));
1991 rq->flags |= REQ_NOMERGE;
1993 rq->waiting = &wait;
1994 elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 1);
1995 generic_unplug_device(q);
1996 wait_for_completion(rq->waiting);
2005 EXPORT_SYMBOL(blk_execute_rq);
2008 * blkdev_issue_flush - queue a flush
2009 * @bdev: blockdev to issue flush for
2010 * @error_sector: error sector
2013 * Issue a flush for the block device in question. Caller can supply
2014 * room for storing the error offset in case of a flush error, if they
2015 * wish to. Caller must run wait_for_completion() on its own.
2017 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2021 if (bdev->bd_disk == NULL)
2024 q = bdev_get_queue(bdev);
2027 if (!q->issue_flush_fn)
2030 return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2033 EXPORT_SYMBOL(blkdev_issue_flush);
2036 * blkdev_scsi_issue_flush_fn - issue flush for SCSI devices
2039 * @error_sector: error offset
2042 * Devices understanding the SCSI command set, can use this function as
2043 * a helper for issuing a cache flush. Note: driver is required to store
2044 * the error offset (in case of error flushing) in ->sector of struct
2047 int blkdev_scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
2048 sector_t *error_sector)
2050 struct request *rq = blk_get_request(q, WRITE, __GFP_WAIT);
2053 rq->flags |= REQ_BLOCK_PC | REQ_SOFTBARRIER;
2055 memset(rq->cmd, 0, sizeof(rq->cmd));
2060 rq->timeout = 60 * HZ;
2062 ret = blk_execute_rq(q, disk, rq);
2064 if (ret && error_sector)
2065 *error_sector = rq->sector;
2067 blk_put_request(rq);
2071 EXPORT_SYMBOL(blkdev_scsi_issue_flush_fn);
2073 void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2075 int rw = rq_data_dir(rq);
2077 if (!blk_fs_request(rq) || !rq->rq_disk)
2081 __disk_stat_add(rq->rq_disk, read_sectors, nr_sectors);
2083 __disk_stat_inc(rq->rq_disk, read_merges);
2084 } else if (rw == WRITE) {
2085 __disk_stat_add(rq->rq_disk, write_sectors, nr_sectors);
2087 __disk_stat_inc(rq->rq_disk, write_merges);
2090 disk_round_stats(rq->rq_disk);
2091 rq->rq_disk->in_flight++;
2096 * add-request adds a request to the linked list.
2097 * queue lock is held and interrupts disabled, as we muck with the
2098 * request queue list.
2100 static inline void add_request(request_queue_t * q, struct request * req)
2102 drive_stat_acct(req, req->nr_sectors, 1);
2105 q->activity_fn(q->activity_data, rq_data_dir(req));
2108 * elevator indicated where it wants this request to be
2109 * inserted at elevator_merge time
2111 __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2115 * disk_round_stats() - Round off the performance stats on a struct
2118 * The average IO queue length and utilisation statistics are maintained
2119 * by observing the current state of the queue length and the amount of
2120 * time it has been in this state for.
2122 * Normally, that accounting is done on IO completion, but that can result
2123 * in more than a second's worth of IO being accounted for within any one
2124 * second, leading to >100% utilisation. To deal with that, we call this
2125 * function to do a round-off before returning the results when reading
2126 * /proc/diskstats. This accounts immediately for all queue usage up to
2127 * the current jiffies and restarts the counters again.
2129 void disk_round_stats(struct gendisk *disk)
2131 unsigned long now = jiffies;
2133 __disk_stat_add(disk, time_in_queue,
2134 disk->in_flight * (now - disk->stamp));
2137 if (disk->in_flight)
2138 __disk_stat_add(disk, io_ticks, (now - disk->stamp_idle));
2139 disk->stamp_idle = now;
2143 * queue lock must be held
2145 void __blk_put_request(request_queue_t *q, struct request *req)
2147 struct request_list *rl = req->rl;
2151 if (unlikely(--req->ref_count))
2154 req->rq_status = RQ_INACTIVE;
2159 * Request may not have originated from ll_rw_blk. if not,
2160 * it didn't come out of our reserved rq pools
2163 int rw = rq_data_dir(req);
2165 elv_completed_request(q, req);
2167 BUG_ON(!list_empty(&req->queuelist));
2169 blk_free_request(q, req);
2170 freed_request(q, rw);
2174 void blk_put_request(struct request *req)
2177 * if req->rl isn't set, this request didnt originate from the
2178 * block layer, so it's safe to just disregard it
2181 unsigned long flags;
2182 request_queue_t *q = req->q;
2184 spin_lock_irqsave(q->queue_lock, flags);
2185 __blk_put_request(q, req);
2186 spin_unlock_irqrestore(q->queue_lock, flags);
2190 EXPORT_SYMBOL(blk_put_request);
2193 * blk_congestion_wait - wait for a queue to become uncongested
2194 * @rw: READ or WRITE
2195 * @timeout: timeout in jiffies
2197 * Waits for up to @timeout jiffies for a queue (any queue) to exit congestion.
2198 * If no queues are congested then just wait for the next request to be
2201 long blk_congestion_wait(int rw, long timeout)
2205 wait_queue_head_t *wqh = &congestion_wqh[rw];
2207 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
2208 ret = io_schedule_timeout(timeout);
2209 finish_wait(wqh, &wait);
2213 EXPORT_SYMBOL(blk_congestion_wait);
2216 * Has to be called with the request spinlock acquired
2218 static int attempt_merge(request_queue_t *q, struct request *req,
2219 struct request *next)
2221 if (!rq_mergeable(req) || !rq_mergeable(next))
2227 if (req->sector + req->nr_sectors != next->sector)
2230 if (rq_data_dir(req) != rq_data_dir(next)
2231 || req->rq_disk != next->rq_disk
2232 || next->waiting || next->special)
2236 * If we are allowed to merge, then append bio list
2237 * from next to rq and release next. merge_requests_fn
2238 * will have updated segment counts, update sector
2241 if (!q->merge_requests_fn(q, req, next))
2245 * At this point we have either done a back merge
2246 * or front merge. We need the smaller start_time of
2247 * the merged requests to be the current request
2248 * for accounting purposes.
2250 if (time_after(req->start_time, next->start_time))
2251 req->start_time = next->start_time;
2253 req->biotail->bi_next = next->bio;
2254 req->biotail = next->biotail;
2256 req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2258 elv_merge_requests(q, req, next);
2261 disk_round_stats(req->rq_disk);
2262 req->rq_disk->in_flight--;
2265 __blk_put_request(q, next);
2269 static inline int attempt_back_merge(request_queue_t *q, struct request *rq)
2271 struct request *next = elv_latter_request(q, rq);
2274 return attempt_merge(q, rq, next);
2279 static inline int attempt_front_merge(request_queue_t *q, struct request *rq)
2281 struct request *prev = elv_former_request(q, rq);
2284 return attempt_merge(q, prev, rq);
2290 * blk_attempt_remerge - attempt to remerge active head with next request
2291 * @q: The &request_queue_t belonging to the device
2292 * @rq: The head request (usually)
2295 * For head-active devices, the queue can easily be unplugged so quickly
2296 * that proper merging is not done on the front request. This may hurt
2297 * performance greatly for some devices. The block layer cannot safely
2298 * do merging on that first request for these queues, but the driver can
2299 * call this function and make it happen any way. Only the driver knows
2300 * when it is safe to do so.
2302 void blk_attempt_remerge(request_queue_t *q, struct request *rq)
2304 unsigned long flags;
2306 spin_lock_irqsave(q->queue_lock, flags);
2307 attempt_back_merge(q, rq);
2308 spin_unlock_irqrestore(q->queue_lock, flags);
2311 EXPORT_SYMBOL(blk_attempt_remerge);
2314 * Non-locking blk_attempt_remerge variant.
2316 void __blk_attempt_remerge(request_queue_t *q, struct request *rq)
2318 attempt_back_merge(q, rq);
2321 EXPORT_SYMBOL(__blk_attempt_remerge);
2323 static int __make_request(request_queue_t *q, struct bio *bio)
2325 struct request *req, *freereq = NULL;
2326 int el_ret, rw, nr_sectors, cur_nr_sectors, barrier, err;
2329 sector = bio->bi_sector;
2330 nr_sectors = bio_sectors(bio);
2331 cur_nr_sectors = bio_cur_sectors(bio);
2333 rw = bio_data_dir(bio);
2336 * low level driver can indicate that it wants pages above a
2337 * certain limit bounced to low memory (ie for highmem, or even
2338 * ISA dma in theory)
2340 blk_queue_bounce(q, &bio);
2342 spin_lock_prefetch(q->queue_lock);
2344 barrier = bio_barrier(bio);
2345 if (barrier && !(q->queue_flags & (1 << QUEUE_FLAG_ORDERED))) {
2351 spin_lock_irq(q->queue_lock);
2353 if (elv_queue_empty(q)) {
2360 el_ret = elv_merge(q, &req, bio);
2362 case ELEVATOR_BACK_MERGE:
2363 BUG_ON(!rq_mergeable(req));
2365 if (!q->back_merge_fn(q, req, bio))
2368 req->biotail->bi_next = bio;
2370 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2371 drive_stat_acct(req, nr_sectors, 0);
2372 if (!attempt_back_merge(q, req))
2373 elv_merged_request(q, req);
2376 case ELEVATOR_FRONT_MERGE:
2377 BUG_ON(!rq_mergeable(req));
2379 if (!q->front_merge_fn(q, req, bio))
2382 bio->bi_next = req->bio;
2386 * may not be valid. if the low level driver said
2387 * it didn't need a bounce buffer then it better
2388 * not touch req->buffer either...
2390 req->buffer = bio_data(bio);
2391 req->current_nr_sectors = cur_nr_sectors;
2392 req->hard_cur_sectors = cur_nr_sectors;
2393 req->sector = req->hard_sector = sector;
2394 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2395 drive_stat_acct(req, nr_sectors, 0);
2396 if (!attempt_front_merge(q, req))
2397 elv_merged_request(q, req);
2401 * elevator says don't/can't merge. get new request
2403 case ELEVATOR_NO_MERGE:
2407 printk("elevator returned crap (%d)\n", el_ret);
2412 * Grab a free request from the freelist - if that is empty, check
2413 * if we are doing read ahead and abort instead of blocking for
2421 spin_unlock_irq(q->queue_lock);
2422 if ((freereq = get_request(q, rw, GFP_ATOMIC)) == NULL) {
2427 if (bio_rw_ahead(bio))
2430 freereq = get_request_wait(q, rw);
2435 req->flags |= REQ_CMD;
2438 * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2440 if (bio_rw_ahead(bio) || bio_failfast(bio))
2441 req->flags |= REQ_FAILFAST;
2444 * REQ_BARRIER implies no merging, but lets make it explicit
2447 req->flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2450 req->hard_sector = req->sector = sector;
2451 req->hard_nr_sectors = req->nr_sectors = nr_sectors;
2452 req->current_nr_sectors = req->hard_cur_sectors = cur_nr_sectors;
2453 req->nr_phys_segments = bio_phys_segments(q, bio);
2454 req->nr_hw_segments = bio_hw_segments(q, bio);
2455 req->buffer = bio_data(bio); /* see ->buffer comment above */
2456 req->waiting = NULL;
2457 req->bio = req->biotail = bio;
2458 req->rq_disk = bio->bi_bdev->bd_disk;
2459 req->start_time = jiffies;
2461 add_request(q, req);
2464 __blk_put_request(q, freereq);
2466 __generic_unplug_device(q);
2468 spin_unlock_irq(q->queue_lock);
2472 bio_endio(bio, nr_sectors << 9, err);
2477 * If bio->bi_dev is a partition, remap the location
2479 static inline void blk_partition_remap(struct bio *bio)
2481 struct block_device *bdev = bio->bi_bdev;
2483 if (bdev != bdev->bd_contains) {
2484 struct hd_struct *p = bdev->bd_part;
2486 switch (bio->bi_rw) {
2488 p->read_sectors += bio_sectors(bio);
2492 p->write_sectors += bio_sectors(bio);
2496 bio->bi_sector += p->start_sect;
2497 bio->bi_bdev = bdev->bd_contains;
2501 void blk_finish_queue_drain(request_queue_t *q)
2503 struct request_list *rl = &q->rq;
2506 spin_lock_irq(q->queue_lock);
2507 clear_bit(QUEUE_FLAG_DRAIN, &q->queue_flags);
2509 while (!list_empty(&q->drain_list)) {
2510 rq = list_entry_rq(q->drain_list.next);
2512 list_del_init(&rq->queuelist);
2513 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 1);
2516 spin_unlock_irq(q->queue_lock);
2518 wake_up(&rl->wait[0]);
2519 wake_up(&rl->wait[1]);
2520 wake_up(&rl->drain);
2523 static int wait_drain(request_queue_t *q, struct request_list *rl, int dispatch)
2525 int wait = rl->count[READ] + rl->count[WRITE];
2528 wait += !list_empty(&q->queue_head);
2534 * We rely on the fact that only requests allocated through blk_alloc_request()
2535 * have io scheduler private data structures associated with them. Any other
2536 * type of request (allocated on stack or through kmalloc()) should not go
2537 * to the io scheduler core, but be attached to the queue head instead.
2539 void blk_wait_queue_drained(request_queue_t *q, int wait_dispatch)
2541 struct request_list *rl = &q->rq;
2544 spin_lock_irq(q->queue_lock);
2545 set_bit(QUEUE_FLAG_DRAIN, &q->queue_flags);
2547 while (wait_drain(q, rl, wait_dispatch)) {
2548 prepare_to_wait(&rl->drain, &wait, TASK_UNINTERRUPTIBLE);
2550 if (wait_drain(q, rl, wait_dispatch)) {
2551 __generic_unplug_device(q);
2552 spin_unlock_irq(q->queue_lock);
2554 spin_lock_irq(q->queue_lock);
2557 finish_wait(&rl->drain, &wait);
2560 spin_unlock_irq(q->queue_lock);
2564 * block waiting for the io scheduler being started again.
2566 static inline void block_wait_queue_running(request_queue_t *q)
2570 while (test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags)) {
2571 struct request_list *rl = &q->rq;
2573 prepare_to_wait_exclusive(&rl->drain, &wait,
2574 TASK_UNINTERRUPTIBLE);
2577 * re-check the condition. avoids using prepare_to_wait()
2578 * in the fast path (queue is running)
2580 if (test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags))
2583 finish_wait(&rl->drain, &wait);
2587 static void handle_bad_sector(struct bio *bio)
2589 char b[BDEVNAME_SIZE];
2591 printk(KERN_INFO "attempt to access beyond end of device\n");
2592 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
2593 bdevname(bio->bi_bdev, b),
2595 (unsigned long long)bio->bi_sector + bio_sectors(bio),
2596 (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
2598 set_bit(BIO_EOF, &bio->bi_flags);
2602 * generic_make_request: hand a buffer to its device driver for I/O
2603 * @bio: The bio describing the location in memory and on the device.
2605 * generic_make_request() is used to make I/O requests of block
2606 * devices. It is passed a &struct bio, which describes the I/O that needs
2609 * generic_make_request() does not return any status. The
2610 * success/failure status of the request, along with notification of
2611 * completion, is delivered asynchronously through the bio->bi_end_io
2612 * function described (one day) else where.
2614 * The caller of generic_make_request must make sure that bi_io_vec
2615 * are set to describe the memory buffer, and that bi_dev and bi_sector are
2616 * set to describe the device address, and the
2617 * bi_end_io and optionally bi_private are set to describe how
2618 * completion notification should be signaled.
2620 * generic_make_request and the drivers it calls may use bi_next if this
2621 * bio happens to be merged with someone else, and may change bi_dev and
2622 * bi_sector for remaps as it sees fit. So the values of these fields
2623 * should NOT be depended on after the call to generic_make_request.
2625 void generic_make_request(struct bio *bio)
2629 int ret, nr_sectors = bio_sectors(bio);
2632 /* Test device or partition size, when known. */
2633 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
2635 sector_t sector = bio->bi_sector;
2637 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
2639 * This may well happen - the kernel calls bread()
2640 * without checking the size of the device, e.g., when
2641 * mounting a device.
2643 handle_bad_sector(bio);
2649 * Resolve the mapping until finished. (drivers are
2650 * still free to implement/resolve their own stacking
2651 * by explicitly returning 0)
2653 * NOTE: we don't repeat the blk_size check for each new device.
2654 * Stacking drivers are expected to know what they are doing.
2657 char b[BDEVNAME_SIZE];
2659 q = bdev_get_queue(bio->bi_bdev);
2662 "generic_make_request: Trying to access "
2663 "nonexistent block-device %s (%Lu)\n",
2664 bdevname(bio->bi_bdev, b),
2665 (long long) bio->bi_sector);
2667 bio_endio(bio, bio->bi_size, -EIO);
2671 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
2672 printk("bio too big device %s (%u > %u)\n",
2673 bdevname(bio->bi_bdev, b),
2679 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))
2682 block_wait_queue_running(q);
2685 * If this device has partitions, remap block n
2686 * of partition p to block n+start(p) of the disk.
2688 blk_partition_remap(bio);
2690 ret = q->make_request_fn(q, bio);
2694 EXPORT_SYMBOL(generic_make_request);
2697 * submit_bio: submit a bio to the block device layer for I/O
2698 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
2699 * @bio: The &struct bio which describes the I/O
2701 * submit_bio() is very similar in purpose to generic_make_request(), and
2702 * uses that function to do most of the work. Both are fairly rough
2703 * interfaces, @bio must be presetup and ready for I/O.
2706 void submit_bio(int rw, struct bio *bio)
2708 int count = bio_sectors(bio);
2710 BIO_BUG_ON(!bio->bi_size);
2711 BIO_BUG_ON(!bio->bi_io_vec);
2714 mod_page_state(pgpgout, count);
2716 mod_page_state(pgpgin, count);
2718 if (unlikely(block_dump)) {
2719 char b[BDEVNAME_SIZE];
2720 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
2721 current->comm, current->pid,
2722 (rw & WRITE) ? "WRITE" : "READ",
2723 (unsigned long long)bio->bi_sector,
2724 bdevname(bio->bi_bdev,b));
2727 generic_make_request(bio);
2730 EXPORT_SYMBOL(submit_bio);
2732 void blk_recalc_rq_segments(struct request *rq)
2734 struct bio *bio, *prevbio = NULL;
2735 int nr_phys_segs, nr_hw_segs;
2736 unsigned int phys_size, hw_size;
2737 request_queue_t *q = rq->q;
2742 phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
2743 rq_for_each_bio(bio, rq) {
2744 /* Force bio hw/phys segs to be recalculated. */
2745 bio->bi_flags &= ~(1 << BIO_SEG_VALID);
2747 nr_phys_segs += bio_phys_segments(q, bio);
2748 nr_hw_segs += bio_hw_segments(q, bio);
2750 int pseg = phys_size + prevbio->bi_size + bio->bi_size;
2751 int hseg = hw_size + prevbio->bi_size + bio->bi_size;
2753 if (blk_phys_contig_segment(q, prevbio, bio) &&
2754 pseg <= q->max_segment_size) {
2756 phys_size += prevbio->bi_size + bio->bi_size;
2760 if (blk_hw_contig_segment(q, prevbio, bio) &&
2761 hseg <= q->max_segment_size) {
2763 hw_size += prevbio->bi_size + bio->bi_size;
2770 rq->nr_phys_segments = nr_phys_segs;
2771 rq->nr_hw_segments = nr_hw_segs;
2774 void blk_recalc_rq_sectors(struct request *rq, int nsect)
2776 if (blk_fs_request(rq)) {
2777 rq->hard_sector += nsect;
2778 rq->hard_nr_sectors -= nsect;
2781 * Move the I/O submission pointers ahead if required.
2783 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
2784 (rq->sector <= rq->hard_sector)) {
2785 rq->sector = rq->hard_sector;
2786 rq->nr_sectors = rq->hard_nr_sectors;
2787 rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
2788 rq->current_nr_sectors = rq->hard_cur_sectors;
2789 rq->buffer = bio_data(rq->bio);
2793 * if total number of sectors is less than the first segment
2794 * size, something has gone terribly wrong
2796 if (rq->nr_sectors < rq->current_nr_sectors) {
2797 printk("blk: request botched\n");
2798 rq->nr_sectors = rq->current_nr_sectors;
2803 static int __end_that_request_first(struct request *req, int uptodate,
2806 int total_bytes, bio_nbytes, error, next_idx = 0;
2810 * extend uptodate bool to allow < 0 value to be direct io error
2813 if (end_io_error(uptodate))
2814 error = !uptodate ? -EIO : uptodate;
2817 * for a REQ_BLOCK_PC request, we want to carry any eventual
2818 * sense key with us all the way through
2820 if (!blk_pc_request(req))
2824 if (blk_fs_request(req) && !(req->flags & REQ_QUIET))
2825 printk("end_request: I/O error, dev %s, sector %llu\n",
2826 req->rq_disk ? req->rq_disk->disk_name : "?",
2827 (unsigned long long)req->sector);
2830 total_bytes = bio_nbytes = 0;
2831 while ((bio = req->bio) != NULL) {
2834 if (nr_bytes >= bio->bi_size) {
2835 req->bio = bio->bi_next;
2836 nbytes = bio->bi_size;
2837 bio_endio(bio, nbytes, error);
2841 int idx = bio->bi_idx + next_idx;
2843 if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
2844 blk_dump_rq_flags(req, "__end_that");
2845 printk("%s: bio idx %d >= vcnt %d\n",
2847 bio->bi_idx, bio->bi_vcnt);
2851 nbytes = bio_iovec_idx(bio, idx)->bv_len;
2852 BIO_BUG_ON(nbytes > bio->bi_size);
2855 * not a complete bvec done
2857 if (unlikely(nbytes > nr_bytes)) {
2858 bio_nbytes += nr_bytes;
2859 total_bytes += nr_bytes;
2864 * advance to the next vector
2867 bio_nbytes += nbytes;
2870 total_bytes += nbytes;
2873 if ((bio = req->bio)) {
2875 * end more in this run, or just return 'not-done'
2877 if (unlikely(nr_bytes <= 0))
2889 * if the request wasn't completed, update state
2892 bio_endio(bio, bio_nbytes, error);
2893 bio->bi_idx += next_idx;
2894 bio_iovec(bio)->bv_offset += nr_bytes;
2895 bio_iovec(bio)->bv_len -= nr_bytes;
2898 blk_recalc_rq_sectors(req, total_bytes >> 9);
2899 blk_recalc_rq_segments(req);
2904 * end_that_request_first - end I/O on a request
2905 * @req: the request being processed
2906 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
2907 * @nr_sectors: number of sectors to end I/O on
2910 * Ends I/O on a number of sectors attached to @req, and sets it up
2911 * for the next range of segments (if any) in the cluster.
2914 * 0 - we are done with this request, call end_that_request_last()
2915 * 1 - still buffers pending for this request
2917 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
2919 return __end_that_request_first(req, uptodate, nr_sectors << 9);
2922 EXPORT_SYMBOL(end_that_request_first);
2925 * end_that_request_chunk - end I/O on a request
2926 * @req: the request being processed
2927 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
2928 * @nr_bytes: number of bytes to complete
2931 * Ends I/O on a number of bytes attached to @req, and sets it up
2932 * for the next range of segments (if any). Like end_that_request_first(),
2933 * but deals with bytes instead of sectors.
2936 * 0 - we are done with this request, call end_that_request_last()
2937 * 1 - still buffers pending for this request
2939 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
2941 return __end_that_request_first(req, uptodate, nr_bytes);
2944 EXPORT_SYMBOL(end_that_request_chunk);
2947 * queue lock must be held
2949 void end_that_request_last(struct request *req)
2951 struct gendisk *disk = req->rq_disk;
2952 struct completion *waiting = req->waiting;
2954 if (unlikely(laptop_mode) && blk_fs_request(req))
2955 laptop_io_completion();
2957 if (disk && blk_fs_request(req)) {
2958 unsigned long duration = jiffies - req->start_time;
2959 switch (rq_data_dir(req)) {
2961 __disk_stat_inc(disk, writes);
2962 __disk_stat_add(disk, write_ticks, duration);
2965 __disk_stat_inc(disk, reads);
2966 __disk_stat_add(disk, read_ticks, duration);
2969 disk_round_stats(disk);
2972 __blk_put_request(req->q, req);
2973 /* Do this LAST! The structure may be freed immediately afterwards */
2978 EXPORT_SYMBOL(end_that_request_last);
2980 void end_request(struct request *req, int uptodate)
2982 if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
2983 add_disk_randomness(req->rq_disk);
2984 blkdev_dequeue_request(req);
2985 end_that_request_last(req);
2989 EXPORT_SYMBOL(end_request);
2991 void blk_rq_bio_prep(request_queue_t *q, struct request *rq, struct bio *bio)
2993 /* first three bits are identical in rq->flags and bio->bi_rw */
2994 rq->flags |= (bio->bi_rw & 7);
2996 rq->nr_phys_segments = bio_phys_segments(q, bio);
2997 rq->nr_hw_segments = bio_hw_segments(q, bio);
2998 rq->current_nr_sectors = bio_cur_sectors(bio);
2999 rq->hard_cur_sectors = rq->current_nr_sectors;
3000 rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3001 rq->buffer = bio_data(bio);
3003 rq->bio = rq->biotail = bio;
3006 EXPORT_SYMBOL(blk_rq_bio_prep);
3008 int kblockd_schedule_work(struct work_struct *work)
3010 return queue_work(kblockd_workqueue, work);
3013 EXPORT_SYMBOL(kblockd_schedule_work);
3015 void kblockd_flush(void)
3017 flush_workqueue(kblockd_workqueue);
3019 EXPORT_SYMBOL(kblockd_flush);
3021 int __init blk_dev_init(void)
3023 kblockd_workqueue = create_workqueue("kblockd");
3024 if (!kblockd_workqueue)
3025 panic("Failed to create kblockd\n");
3027 request_cachep = kmem_cache_create("blkdev_requests",
3028 sizeof(struct request), 0, SLAB_PANIC, NULL, NULL);
3030 requestq_cachep = kmem_cache_create("blkdev_queue",
3031 sizeof(request_queue_t), 0, SLAB_PANIC, NULL, NULL);
3033 iocontext_cachep = kmem_cache_create("blkdev_ioc",
3034 sizeof(struct io_context), 0, SLAB_PANIC, NULL, NULL);
3036 blk_max_low_pfn = max_low_pfn;
3037 blk_max_pfn = max_pfn;
3043 * IO Context helper functions
3045 void put_io_context(struct io_context *ioc)
3050 BUG_ON(atomic_read(&ioc->refcount) == 0);
3052 if (atomic_dec_and_test(&ioc->refcount)) {
3053 if (ioc->aic && ioc->aic->dtor)
3054 ioc->aic->dtor(ioc->aic);
3055 if (ioc->cic && ioc->cic->dtor)
3056 ioc->cic->dtor(ioc->cic);
3058 kmem_cache_free(iocontext_cachep, ioc);
3061 EXPORT_SYMBOL(put_io_context);
3063 /* Called by the exitting task */
3064 void exit_io_context(void)
3066 unsigned long flags;
3067 struct io_context *ioc;
3069 local_irq_save(flags);
3070 ioc = current->io_context;
3071 current->io_context = NULL;
3072 local_irq_restore(flags);
3074 if (ioc->aic && ioc->aic->exit)
3075 ioc->aic->exit(ioc->aic);
3076 if (ioc->cic && ioc->cic->exit)
3077 ioc->cic->exit(ioc->cic);
3079 put_io_context(ioc);
3083 * If the current task has no IO context then create one and initialise it.
3084 * If it does have a context, take a ref on it.
3086 * This is always called in the context of the task which submitted the I/O.
3087 * But weird things happen, so we disable local interrupts to ensure exclusive
3088 * access to *current.
3090 struct io_context *get_io_context(int gfp_flags)
3092 struct task_struct *tsk = current;
3093 unsigned long flags;
3094 struct io_context *ret;
3096 local_irq_save(flags);
3097 ret = tsk->io_context;
3101 local_irq_restore(flags);
3103 ret = kmem_cache_alloc(iocontext_cachep, gfp_flags);
3105 atomic_set(&ret->refcount, 1);
3106 ret->pid = tsk->pid;
3107 ret->last_waited = jiffies; /* doesn't matter... */
3108 ret->nr_batch_requests = 0; /* because this is 0 */
3111 spin_lock_init(&ret->lock);
3113 local_irq_save(flags);
3116 * very unlikely, someone raced with us in setting up the task
3117 * io context. free new context and just grab a reference.
3119 if (!tsk->io_context)
3120 tsk->io_context = ret;
3122 kmem_cache_free(iocontext_cachep, ret);
3123 ret = tsk->io_context;
3127 atomic_inc(&ret->refcount);
3128 local_irq_restore(flags);
3133 EXPORT_SYMBOL(get_io_context);
3135 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3137 struct io_context *src = *psrc;
3138 struct io_context *dst = *pdst;
3141 BUG_ON(atomic_read(&src->refcount) == 0);
3142 atomic_inc(&src->refcount);
3143 put_io_context(dst);
3147 EXPORT_SYMBOL(copy_io_context);
3149 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3151 struct io_context *temp;
3156 EXPORT_SYMBOL(swap_io_context);
3161 struct queue_sysfs_entry {
3162 struct attribute attr;
3163 ssize_t (*show)(struct request_queue *, char *);
3164 ssize_t (*store)(struct request_queue *, const char *, size_t);
3168 queue_var_show(unsigned int var, char *page)
3170 return sprintf(page, "%d\n", var);
3174 queue_var_store(unsigned long *var, const char *page, size_t count)
3176 char *p = (char *) page;
3178 *var = simple_strtoul(p, &p, 10);
3182 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3184 return queue_var_show(q->nr_requests, (page));
3188 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3190 struct request_list *rl = &q->rq;
3192 int ret = queue_var_store(&q->nr_requests, page, count);
3193 if (q->nr_requests < BLKDEV_MIN_RQ)
3194 q->nr_requests = BLKDEV_MIN_RQ;
3195 blk_queue_congestion_threshold(q);
3197 if (rl->count[READ] >= queue_congestion_on_threshold(q))
3198 set_queue_congested(q, READ);
3199 else if (rl->count[READ] < queue_congestion_off_threshold(q))
3200 clear_queue_congested(q, READ);
3202 if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3203 set_queue_congested(q, WRITE);
3204 else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3205 clear_queue_congested(q, WRITE);
3207 if (rl->count[READ] >= q->nr_requests) {
3208 blk_set_queue_full(q, READ);
3209 } else if (rl->count[READ]+1 <= q->nr_requests) {
3210 blk_clear_queue_full(q, READ);
3211 wake_up(&rl->wait[READ]);
3214 if (rl->count[WRITE] >= q->nr_requests) {
3215 blk_set_queue_full(q, WRITE);
3216 } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3217 blk_clear_queue_full(q, WRITE);
3218 wake_up(&rl->wait[WRITE]);
3223 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3225 int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3227 return queue_var_show(ra_kb, (page));
3231 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3233 unsigned long ra_kb;
3234 ssize_t ret = queue_var_store(&ra_kb, page, count);
3236 spin_lock_irq(q->queue_lock);
3237 if (ra_kb > (q->max_sectors >> 1))
3238 ra_kb = (q->max_sectors >> 1);
3240 q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3241 spin_unlock_irq(q->queue_lock);
3246 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3248 int max_sectors_kb = q->max_sectors >> 1;
3250 return queue_var_show(max_sectors_kb, (page));
3254 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3256 unsigned long max_sectors_kb,
3257 max_hw_sectors_kb = q->max_hw_sectors >> 1,
3258 page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3259 ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3262 if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3265 * Take the queue lock to update the readahead and max_sectors
3266 * values synchronously:
3268 spin_lock_irq(q->queue_lock);
3270 * Trim readahead window as well, if necessary:
3272 ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3273 if (ra_kb > max_sectors_kb)
3274 q->backing_dev_info.ra_pages =
3275 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3277 q->max_sectors = max_sectors_kb << 1;
3278 spin_unlock_irq(q->queue_lock);
3283 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3285 int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3287 return queue_var_show(max_hw_sectors_kb, (page));
3291 static struct queue_sysfs_entry queue_requests_entry = {
3292 .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3293 .show = queue_requests_show,
3294 .store = queue_requests_store,
3297 static struct queue_sysfs_entry queue_ra_entry = {
3298 .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3299 .show = queue_ra_show,
3300 .store = queue_ra_store,
3303 static struct queue_sysfs_entry queue_max_sectors_entry = {
3304 .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3305 .show = queue_max_sectors_show,
3306 .store = queue_max_sectors_store,
3309 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3310 .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
3311 .show = queue_max_hw_sectors_show,
3314 static struct queue_sysfs_entry queue_iosched_entry = {
3315 .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
3316 .show = elv_iosched_show,
3317 .store = elv_iosched_store,
3320 static struct attribute *default_attrs[] = {
3321 &queue_requests_entry.attr,
3322 &queue_ra_entry.attr,
3323 &queue_max_hw_sectors_entry.attr,
3324 &queue_max_sectors_entry.attr,
3325 &queue_iosched_entry.attr,
3329 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
3332 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
3334 struct queue_sysfs_entry *entry = to_queue(attr);
3335 struct request_queue *q;
3337 q = container_of(kobj, struct request_queue, kobj);
3341 return entry->show(q, page);
3345 queue_attr_store(struct kobject *kobj, struct attribute *attr,
3346 const char *page, size_t length)
3348 struct queue_sysfs_entry *entry = to_queue(attr);
3349 struct request_queue *q;
3351 q = container_of(kobj, struct request_queue, kobj);
3355 return entry->store(q, page, length);
3358 static struct sysfs_ops queue_sysfs_ops = {
3359 .show = queue_attr_show,
3360 .store = queue_attr_store,
3363 struct kobj_type queue_ktype = {
3364 .sysfs_ops = &queue_sysfs_ops,
3365 .default_attrs = default_attrs,
3368 int blk_register_queue(struct gendisk *disk)
3372 request_queue_t *q = disk->queue;
3374 if (!q || !q->request_fn)
3377 q->kobj.parent = kobject_get(&disk->kobj);
3378 if (!q->kobj.parent)
3381 snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
3382 q->kobj.ktype = &queue_ktype;
3384 ret = kobject_register(&q->kobj);
3388 ret = elv_register_queue(q);
3390 kobject_unregister(&q->kobj);
3397 void blk_unregister_queue(struct gendisk *disk)
3399 request_queue_t *q = disk->queue;
3401 if (q && q->request_fn) {
3402 elv_unregister_queue(q);
3404 kobject_unregister(&q->kobj);
3405 kobject_put(&disk->kobj);