2 * Copyright (C) 1991, 1992 Linus Torvalds
3 * Copyright (C) 1994, Karl Keyte: Added support for disk statistics
4 * Elevator latency, (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
5 * Queue request tables / lock, selectable elevator, Jens Axboe <axboe@suse.de>
6 * kernel-doc documentation started by NeilBrown <neilb@cse.unsw.edu.au> - July2000
7 * bio rewrite, highmem i/o, etc, Jens Axboe <axboe@suse.de> - may 2001
11 * This handles all read/write requests to block devices
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/backing-dev.h>
16 #include <linux/bio.h>
17 #include <linux/blkdev.h>
18 #include <linux/highmem.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/string.h>
22 #include <linux/init.h>
23 #include <linux/bootmem.h> /* for max_pfn/max_low_pfn */
24 #include <linux/completion.h>
25 #include <linux/slab.h>
26 #include <linux/swap.h>
27 #include <linux/writeback.h>
28 #include <linux/interrupt.h>
29 #include <linux/cpu.h>
30 #include <linux/blktrace_api.h>
35 #include <scsi/scsi_cmnd.h>
37 static void blk_unplug_work(void *data
);
38 static void blk_unplug_timeout(unsigned long data
);
39 static void drive_stat_acct(struct request
*rq
, int nr_sectors
, int new_io
);
40 static void init_request_from_bio(struct request
*req
, struct bio
*bio
);
41 static int __make_request(request_queue_t
*q
, struct bio
*bio
);
44 * For the allocated request tables
46 static kmem_cache_t
*request_cachep
;
49 * For queue allocation
51 static kmem_cache_t
*requestq_cachep
;
54 * For io context allocations
56 static kmem_cache_t
*iocontext_cachep
;
58 static wait_queue_head_t congestion_wqh
[2] = {
59 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh
[0]),
60 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh
[1])
64 * Controlling structure to kblockd
66 static struct workqueue_struct
*kblockd_workqueue
;
68 unsigned long blk_max_low_pfn
, blk_max_pfn
;
70 EXPORT_SYMBOL(blk_max_low_pfn
);
71 EXPORT_SYMBOL(blk_max_pfn
);
73 static DEFINE_PER_CPU(struct list_head
, blk_cpu_done
);
75 /* Amount of time in which a process may batch requests */
76 #define BLK_BATCH_TIME (HZ/50UL)
78 /* Number of requests a "batching" process may submit */
79 #define BLK_BATCH_REQ 32
82 * Return the threshold (number of used requests) at which the queue is
83 * considered to be congested. It include a little hysteresis to keep the
84 * context switch rate down.
86 static inline int queue_congestion_on_threshold(struct request_queue
*q
)
88 return q
->nr_congestion_on
;
92 * The threshold at which a queue is considered to be uncongested
94 static inline int queue_congestion_off_threshold(struct request_queue
*q
)
96 return q
->nr_congestion_off
;
99 static void blk_queue_congestion_threshold(struct request_queue
*q
)
103 nr
= q
->nr_requests
- (q
->nr_requests
/ 8) + 1;
104 if (nr
> q
->nr_requests
)
106 q
->nr_congestion_on
= nr
;
108 nr
= q
->nr_requests
- (q
->nr_requests
/ 8) - (q
->nr_requests
/ 16) - 1;
111 q
->nr_congestion_off
= nr
;
115 * A queue has just exitted congestion. Note this in the global counter of
116 * congested queues, and wake up anyone who was waiting for requests to be
119 static void clear_queue_congested(request_queue_t
*q
, int rw
)
122 wait_queue_head_t
*wqh
= &congestion_wqh
[rw
];
124 bit
= (rw
== WRITE
) ? BDI_write_congested
: BDI_read_congested
;
125 clear_bit(bit
, &q
->backing_dev_info
.state
);
126 smp_mb__after_clear_bit();
127 if (waitqueue_active(wqh
))
132 * A queue has just entered congestion. Flag that in the queue's VM-visible
133 * state flags and increment the global gounter of congested queues.
135 static void set_queue_congested(request_queue_t
*q
, int rw
)
139 bit
= (rw
== WRITE
) ? BDI_write_congested
: BDI_read_congested
;
140 set_bit(bit
, &q
->backing_dev_info
.state
);
144 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
147 * Locates the passed device's request queue and returns the address of its
150 * Will return NULL if the request queue cannot be located.
152 struct backing_dev_info
*blk_get_backing_dev_info(struct block_device
*bdev
)
154 struct backing_dev_info
*ret
= NULL
;
155 request_queue_t
*q
= bdev_get_queue(bdev
);
158 ret
= &q
->backing_dev_info
;
162 EXPORT_SYMBOL(blk_get_backing_dev_info
);
164 void blk_queue_activity_fn(request_queue_t
*q
, activity_fn
*fn
, void *data
)
167 q
->activity_data
= data
;
170 EXPORT_SYMBOL(blk_queue_activity_fn
);
173 * blk_queue_prep_rq - set a prepare_request function for queue
175 * @pfn: prepare_request function
177 * It's possible for a queue to register a prepare_request callback which
178 * is invoked before the request is handed to the request_fn. The goal of
179 * the function is to prepare a request for I/O, it can be used to build a
180 * cdb from the request data for instance.
183 void blk_queue_prep_rq(request_queue_t
*q
, prep_rq_fn
*pfn
)
188 EXPORT_SYMBOL(blk_queue_prep_rq
);
191 * blk_queue_merge_bvec - set a merge_bvec function for queue
193 * @mbfn: merge_bvec_fn
195 * Usually queues have static limitations on the max sectors or segments that
196 * we can put in a request. Stacking drivers may have some settings that
197 * are dynamic, and thus we have to query the queue whether it is ok to
198 * add a new bio_vec to a bio at a given offset or not. If the block device
199 * has such limitations, it needs to register a merge_bvec_fn to control
200 * the size of bio's sent to it. Note that a block device *must* allow a
201 * single page to be added to an empty bio. The block device driver may want
202 * to use the bio_split() function to deal with these bio's. By default
203 * no merge_bvec_fn is defined for a queue, and only the fixed limits are
206 void blk_queue_merge_bvec(request_queue_t
*q
, merge_bvec_fn
*mbfn
)
208 q
->merge_bvec_fn
= mbfn
;
211 EXPORT_SYMBOL(blk_queue_merge_bvec
);
213 void blk_queue_softirq_done(request_queue_t
*q
, softirq_done_fn
*fn
)
215 q
->softirq_done_fn
= fn
;
218 EXPORT_SYMBOL(blk_queue_softirq_done
);
221 * blk_queue_make_request - define an alternate make_request function for a device
222 * @q: the request queue for the device to be affected
223 * @mfn: the alternate make_request function
226 * The normal way for &struct bios to be passed to a device
227 * driver is for them to be collected into requests on a request
228 * queue, and then to allow the device driver to select requests
229 * off that queue when it is ready. This works well for many block
230 * devices. However some block devices (typically virtual devices
231 * such as md or lvm) do not benefit from the processing on the
232 * request queue, and are served best by having the requests passed
233 * directly to them. This can be achieved by providing a function
234 * to blk_queue_make_request().
237 * The driver that does this *must* be able to deal appropriately
238 * with buffers in "highmemory". This can be accomplished by either calling
239 * __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
240 * blk_queue_bounce() to create a buffer in normal memory.
242 void blk_queue_make_request(request_queue_t
* q
, make_request_fn
* mfn
)
247 q
->nr_requests
= BLKDEV_MAX_RQ
;
248 blk_queue_max_phys_segments(q
, MAX_PHYS_SEGMENTS
);
249 blk_queue_max_hw_segments(q
, MAX_HW_SEGMENTS
);
250 q
->make_request_fn
= mfn
;
251 q
->backing_dev_info
.ra_pages
= (VM_MAX_READAHEAD
* 1024) / PAGE_CACHE_SIZE
;
252 q
->backing_dev_info
.state
= 0;
253 q
->backing_dev_info
.capabilities
= BDI_CAP_MAP_COPY
;
254 blk_queue_max_sectors(q
, SAFE_MAX_SECTORS
);
255 blk_queue_hardsect_size(q
, 512);
256 blk_queue_dma_alignment(q
, 511);
257 blk_queue_congestion_threshold(q
);
258 q
->nr_batching
= BLK_BATCH_REQ
;
260 q
->unplug_thresh
= 4; /* hmm */
261 q
->unplug_delay
= (3 * HZ
) / 1000; /* 3 milliseconds */
262 if (q
->unplug_delay
== 0)
265 INIT_WORK(&q
->unplug_work
, blk_unplug_work
, q
);
267 q
->unplug_timer
.function
= blk_unplug_timeout
;
268 q
->unplug_timer
.data
= (unsigned long)q
;
271 * by default assume old behaviour and bounce for any highmem page
273 blk_queue_bounce_limit(q
, BLK_BOUNCE_HIGH
);
275 blk_queue_activity_fn(q
, NULL
, NULL
);
278 EXPORT_SYMBOL(blk_queue_make_request
);
280 static inline void rq_init(request_queue_t
*q
, struct request
*rq
)
282 INIT_LIST_HEAD(&rq
->queuelist
);
283 INIT_LIST_HEAD(&rq
->donelist
);
286 rq
->rq_status
= RQ_ACTIVE
;
287 rq
->bio
= rq
->biotail
= NULL
;
288 INIT_HLIST_NODE(&rq
->hash
);
289 RB_CLEAR_NODE(&rq
->rb_node
);
298 rq
->nr_phys_segments
= 0;
301 rq
->end_io_data
= NULL
;
302 rq
->completion_data
= NULL
;
306 * blk_queue_ordered - does this queue support ordered writes
307 * @q: the request queue
308 * @ordered: one of QUEUE_ORDERED_*
309 * @prepare_flush_fn: rq setup helper for cache flush ordered writes
312 * For journalled file systems, doing ordered writes on a commit
313 * block instead of explicitly doing wait_on_buffer (which is bad
314 * for performance) can be a big win. Block drivers supporting this
315 * feature should call this function and indicate so.
318 int blk_queue_ordered(request_queue_t
*q
, unsigned ordered
,
319 prepare_flush_fn
*prepare_flush_fn
)
321 if (ordered
& (QUEUE_ORDERED_PREFLUSH
| QUEUE_ORDERED_POSTFLUSH
) &&
322 prepare_flush_fn
== NULL
) {
323 printk(KERN_ERR
"blk_queue_ordered: prepare_flush_fn required\n");
327 if (ordered
!= QUEUE_ORDERED_NONE
&&
328 ordered
!= QUEUE_ORDERED_DRAIN
&&
329 ordered
!= QUEUE_ORDERED_DRAIN_FLUSH
&&
330 ordered
!= QUEUE_ORDERED_DRAIN_FUA
&&
331 ordered
!= QUEUE_ORDERED_TAG
&&
332 ordered
!= QUEUE_ORDERED_TAG_FLUSH
&&
333 ordered
!= QUEUE_ORDERED_TAG_FUA
) {
334 printk(KERN_ERR
"blk_queue_ordered: bad value %d\n", ordered
);
338 q
->ordered
= ordered
;
339 q
->next_ordered
= ordered
;
340 q
->prepare_flush_fn
= prepare_flush_fn
;
345 EXPORT_SYMBOL(blk_queue_ordered
);
348 * blk_queue_issue_flush_fn - set function for issuing a flush
349 * @q: the request queue
350 * @iff: the function to be called issuing the flush
353 * If a driver supports issuing a flush command, the support is notified
354 * to the block layer by defining it through this call.
357 void blk_queue_issue_flush_fn(request_queue_t
*q
, issue_flush_fn
*iff
)
359 q
->issue_flush_fn
= iff
;
362 EXPORT_SYMBOL(blk_queue_issue_flush_fn
);
365 * Cache flushing for ordered writes handling
367 inline unsigned blk_ordered_cur_seq(request_queue_t
*q
)
371 return 1 << ffz(q
->ordseq
);
374 unsigned blk_ordered_req_seq(struct request
*rq
)
376 request_queue_t
*q
= rq
->q
;
378 BUG_ON(q
->ordseq
== 0);
380 if (rq
== &q
->pre_flush_rq
)
381 return QUEUE_ORDSEQ_PREFLUSH
;
382 if (rq
== &q
->bar_rq
)
383 return QUEUE_ORDSEQ_BAR
;
384 if (rq
== &q
->post_flush_rq
)
385 return QUEUE_ORDSEQ_POSTFLUSH
;
387 if ((rq
->cmd_flags
& REQ_ORDERED_COLOR
) ==
388 (q
->orig_bar_rq
->cmd_flags
& REQ_ORDERED_COLOR
))
389 return QUEUE_ORDSEQ_DRAIN
;
391 return QUEUE_ORDSEQ_DONE
;
394 void blk_ordered_complete_seq(request_queue_t
*q
, unsigned seq
, int error
)
399 if (error
&& !q
->orderr
)
402 BUG_ON(q
->ordseq
& seq
);
405 if (blk_ordered_cur_seq(q
) != QUEUE_ORDSEQ_DONE
)
409 * Okay, sequence complete.
412 uptodate
= q
->orderr
? q
->orderr
: 1;
416 end_that_request_first(rq
, uptodate
, rq
->hard_nr_sectors
);
417 end_that_request_last(rq
, uptodate
);
420 static void pre_flush_end_io(struct request
*rq
, int error
)
422 elv_completed_request(rq
->q
, rq
);
423 blk_ordered_complete_seq(rq
->q
, QUEUE_ORDSEQ_PREFLUSH
, error
);
426 static void bar_end_io(struct request
*rq
, int error
)
428 elv_completed_request(rq
->q
, rq
);
429 blk_ordered_complete_seq(rq
->q
, QUEUE_ORDSEQ_BAR
, error
);
432 static void post_flush_end_io(struct request
*rq
, int error
)
434 elv_completed_request(rq
->q
, rq
);
435 blk_ordered_complete_seq(rq
->q
, QUEUE_ORDSEQ_POSTFLUSH
, error
);
438 static void queue_flush(request_queue_t
*q
, unsigned which
)
441 rq_end_io_fn
*end_io
;
443 if (which
== QUEUE_ORDERED_PREFLUSH
) {
444 rq
= &q
->pre_flush_rq
;
445 end_io
= pre_flush_end_io
;
447 rq
= &q
->post_flush_rq
;
448 end_io
= post_flush_end_io
;
451 rq
->cmd_flags
= REQ_HARDBARRIER
;
453 rq
->elevator_private
= NULL
;
454 rq
->rq_disk
= q
->bar_rq
.rq_disk
;
457 q
->prepare_flush_fn(q
, rq
);
459 elv_insert(q
, rq
, ELEVATOR_INSERT_FRONT
);
462 static inline struct request
*start_ordered(request_queue_t
*q
,
467 q
->ordered
= q
->next_ordered
;
468 q
->ordseq
|= QUEUE_ORDSEQ_STARTED
;
471 * Prep proxy barrier request.
473 blkdev_dequeue_request(rq
);
478 if (bio_data_dir(q
->orig_bar_rq
->bio
) == WRITE
)
479 rq
->cmd_flags
|= REQ_RW
;
480 rq
->cmd_flags
|= q
->ordered
& QUEUE_ORDERED_FUA
? REQ_FUA
: 0;
481 rq
->elevator_private
= NULL
;
483 init_request_from_bio(rq
, q
->orig_bar_rq
->bio
);
484 rq
->end_io
= bar_end_io
;
487 * Queue ordered sequence. As we stack them at the head, we
488 * need to queue in reverse order. Note that we rely on that
489 * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
490 * request gets inbetween ordered sequence.
492 if (q
->ordered
& QUEUE_ORDERED_POSTFLUSH
)
493 queue_flush(q
, QUEUE_ORDERED_POSTFLUSH
);
495 q
->ordseq
|= QUEUE_ORDSEQ_POSTFLUSH
;
497 elv_insert(q
, rq
, ELEVATOR_INSERT_FRONT
);
499 if (q
->ordered
& QUEUE_ORDERED_PREFLUSH
) {
500 queue_flush(q
, QUEUE_ORDERED_PREFLUSH
);
501 rq
= &q
->pre_flush_rq
;
503 q
->ordseq
|= QUEUE_ORDSEQ_PREFLUSH
;
505 if ((q
->ordered
& QUEUE_ORDERED_TAG
) || q
->in_flight
== 0)
506 q
->ordseq
|= QUEUE_ORDSEQ_DRAIN
;
513 int blk_do_ordered(request_queue_t
*q
, struct request
**rqp
)
515 struct request
*rq
= *rqp
;
516 int is_barrier
= blk_fs_request(rq
) && blk_barrier_rq(rq
);
522 if (q
->next_ordered
!= QUEUE_ORDERED_NONE
) {
523 *rqp
= start_ordered(q
, rq
);
527 * This can happen when the queue switches to
528 * ORDERED_NONE while this request is on it.
530 blkdev_dequeue_request(rq
);
531 end_that_request_first(rq
, -EOPNOTSUPP
,
532 rq
->hard_nr_sectors
);
533 end_that_request_last(rq
, -EOPNOTSUPP
);
540 * Ordered sequence in progress
543 /* Special requests are not subject to ordering rules. */
544 if (!blk_fs_request(rq
) &&
545 rq
!= &q
->pre_flush_rq
&& rq
!= &q
->post_flush_rq
)
548 if (q
->ordered
& QUEUE_ORDERED_TAG
) {
549 /* Ordered by tag. Blocking the next barrier is enough. */
550 if (is_barrier
&& rq
!= &q
->bar_rq
)
553 /* Ordered by draining. Wait for turn. */
554 WARN_ON(blk_ordered_req_seq(rq
) < blk_ordered_cur_seq(q
));
555 if (blk_ordered_req_seq(rq
) > blk_ordered_cur_seq(q
))
562 static int flush_dry_bio_endio(struct bio
*bio
, unsigned int bytes
, int error
)
564 request_queue_t
*q
= bio
->bi_private
;
565 struct bio_vec
*bvec
;
569 * This is dry run, restore bio_sector and size. We'll finish
570 * this request again with the original bi_end_io after an
571 * error occurs or post flush is complete.
580 bio_for_each_segment(bvec
, bio
, i
) {
581 bvec
->bv_len
+= bvec
->bv_offset
;
586 set_bit(BIO_UPTODATE
, &bio
->bi_flags
);
587 bio
->bi_size
= q
->bi_size
;
588 bio
->bi_sector
-= (q
->bi_size
>> 9);
594 static inline int ordered_bio_endio(struct request
*rq
, struct bio
*bio
,
595 unsigned int nbytes
, int error
)
597 request_queue_t
*q
= rq
->q
;
601 if (&q
->bar_rq
!= rq
)
605 * Okay, this is the barrier request in progress, dry finish it.
607 if (error
&& !q
->orderr
)
610 endio
= bio
->bi_end_io
;
611 private = bio
->bi_private
;
612 bio
->bi_end_io
= flush_dry_bio_endio
;
615 bio_endio(bio
, nbytes
, error
);
617 bio
->bi_end_io
= endio
;
618 bio
->bi_private
= private;
624 * blk_queue_bounce_limit - set bounce buffer limit for queue
625 * @q: the request queue for the device
626 * @dma_addr: bus address limit
629 * Different hardware can have different requirements as to what pages
630 * it can do I/O directly to. A low level driver can call
631 * blk_queue_bounce_limit to have lower memory pages allocated as bounce
632 * buffers for doing I/O to pages residing above @page.
634 void blk_queue_bounce_limit(request_queue_t
*q
, u64 dma_addr
)
636 unsigned long bounce_pfn
= dma_addr
>> PAGE_SHIFT
;
639 q
->bounce_gfp
= GFP_NOIO
;
640 #if BITS_PER_LONG == 64
641 /* Assume anything <= 4GB can be handled by IOMMU.
642 Actually some IOMMUs can handle everything, but I don't
643 know of a way to test this here. */
644 if (bounce_pfn
< (min_t(u64
,0xffffffff,BLK_BOUNCE_HIGH
) >> PAGE_SHIFT
))
646 q
->bounce_pfn
= max_low_pfn
;
648 if (bounce_pfn
< blk_max_low_pfn
)
650 q
->bounce_pfn
= bounce_pfn
;
653 init_emergency_isa_pool();
654 q
->bounce_gfp
= GFP_NOIO
| GFP_DMA
;
655 q
->bounce_pfn
= bounce_pfn
;
659 EXPORT_SYMBOL(blk_queue_bounce_limit
);
662 * blk_queue_max_sectors - set max sectors for a request for this queue
663 * @q: the request queue for the device
664 * @max_sectors: max sectors in the usual 512b unit
667 * Enables a low level driver to set an upper limit on the size of
670 void blk_queue_max_sectors(request_queue_t
*q
, unsigned int max_sectors
)
672 if ((max_sectors
<< 9) < PAGE_CACHE_SIZE
) {
673 max_sectors
= 1 << (PAGE_CACHE_SHIFT
- 9);
674 printk("%s: set to minimum %d\n", __FUNCTION__
, max_sectors
);
677 if (BLK_DEF_MAX_SECTORS
> max_sectors
)
678 q
->max_hw_sectors
= q
->max_sectors
= max_sectors
;
680 q
->max_sectors
= BLK_DEF_MAX_SECTORS
;
681 q
->max_hw_sectors
= max_sectors
;
685 EXPORT_SYMBOL(blk_queue_max_sectors
);
688 * blk_queue_max_phys_segments - set max phys segments for a request for this queue
689 * @q: the request queue for the device
690 * @max_segments: max number of segments
693 * Enables a low level driver to set an upper limit on the number of
694 * physical data segments in a request. This would be the largest sized
695 * scatter list the driver could handle.
697 void blk_queue_max_phys_segments(request_queue_t
*q
, unsigned short max_segments
)
701 printk("%s: set to minimum %d\n", __FUNCTION__
, max_segments
);
704 q
->max_phys_segments
= max_segments
;
707 EXPORT_SYMBOL(blk_queue_max_phys_segments
);
710 * blk_queue_max_hw_segments - set max hw segments for a request for this queue
711 * @q: the request queue for the device
712 * @max_segments: max number of segments
715 * Enables a low level driver to set an upper limit on the number of
716 * hw data segments in a request. This would be the largest number of
717 * address/length pairs the host adapter can actually give as once
720 void blk_queue_max_hw_segments(request_queue_t
*q
, unsigned short max_segments
)
724 printk("%s: set to minimum %d\n", __FUNCTION__
, max_segments
);
727 q
->max_hw_segments
= max_segments
;
730 EXPORT_SYMBOL(blk_queue_max_hw_segments
);
733 * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
734 * @q: the request queue for the device
735 * @max_size: max size of segment in bytes
738 * Enables a low level driver to set an upper limit on the size of a
741 void blk_queue_max_segment_size(request_queue_t
*q
, unsigned int max_size
)
743 if (max_size
< PAGE_CACHE_SIZE
) {
744 max_size
= PAGE_CACHE_SIZE
;
745 printk("%s: set to minimum %d\n", __FUNCTION__
, max_size
);
748 q
->max_segment_size
= max_size
;
751 EXPORT_SYMBOL(blk_queue_max_segment_size
);
754 * blk_queue_hardsect_size - set hardware sector size for the queue
755 * @q: the request queue for the device
756 * @size: the hardware sector size, in bytes
759 * This should typically be set to the lowest possible sector size
760 * that the hardware can operate on (possible without reverting to
761 * even internal read-modify-write operations). Usually the default
762 * of 512 covers most hardware.
764 void blk_queue_hardsect_size(request_queue_t
*q
, unsigned short size
)
766 q
->hardsect_size
= size
;
769 EXPORT_SYMBOL(blk_queue_hardsect_size
);
772 * Returns the minimum that is _not_ zero, unless both are zero.
774 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
777 * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
778 * @t: the stacking driver (top)
779 * @b: the underlying device (bottom)
781 void blk_queue_stack_limits(request_queue_t
*t
, request_queue_t
*b
)
783 /* zero is "infinity" */
784 t
->max_sectors
= min_not_zero(t
->max_sectors
,b
->max_sectors
);
785 t
->max_hw_sectors
= min_not_zero(t
->max_hw_sectors
,b
->max_hw_sectors
);
787 t
->max_phys_segments
= min(t
->max_phys_segments
,b
->max_phys_segments
);
788 t
->max_hw_segments
= min(t
->max_hw_segments
,b
->max_hw_segments
);
789 t
->max_segment_size
= min(t
->max_segment_size
,b
->max_segment_size
);
790 t
->hardsect_size
= max(t
->hardsect_size
,b
->hardsect_size
);
791 if (!test_bit(QUEUE_FLAG_CLUSTER
, &b
->queue_flags
))
792 clear_bit(QUEUE_FLAG_CLUSTER
, &t
->queue_flags
);
795 EXPORT_SYMBOL(blk_queue_stack_limits
);
798 * blk_queue_segment_boundary - set boundary rules for segment merging
799 * @q: the request queue for the device
800 * @mask: the memory boundary mask
802 void blk_queue_segment_boundary(request_queue_t
*q
, unsigned long mask
)
804 if (mask
< PAGE_CACHE_SIZE
- 1) {
805 mask
= PAGE_CACHE_SIZE
- 1;
806 printk("%s: set to minimum %lx\n", __FUNCTION__
, mask
);
809 q
->seg_boundary_mask
= mask
;
812 EXPORT_SYMBOL(blk_queue_segment_boundary
);
815 * blk_queue_dma_alignment - set dma length and memory alignment
816 * @q: the request queue for the device
817 * @mask: alignment mask
820 * set required memory and length aligment for direct dma transactions.
821 * this is used when buiding direct io requests for the queue.
824 void blk_queue_dma_alignment(request_queue_t
*q
, int mask
)
826 q
->dma_alignment
= mask
;
829 EXPORT_SYMBOL(blk_queue_dma_alignment
);
832 * blk_queue_find_tag - find a request by its tag and queue
833 * @q: The request queue for the device
834 * @tag: The tag of the request
837 * Should be used when a device returns a tag and you want to match
840 * no locks need be held.
842 struct request
*blk_queue_find_tag(request_queue_t
*q
, int tag
)
844 struct blk_queue_tag
*bqt
= q
->queue_tags
;
846 if (unlikely(bqt
== NULL
|| tag
>= bqt
->real_max_depth
))
849 return bqt
->tag_index
[tag
];
852 EXPORT_SYMBOL(blk_queue_find_tag
);
855 * __blk_free_tags - release a given set of tag maintenance info
856 * @bqt: the tag map to free
858 * Tries to free the specified @bqt@. Returns true if it was
859 * actually freed and false if there are still references using it
861 static int __blk_free_tags(struct blk_queue_tag
*bqt
)
865 retval
= atomic_dec_and_test(&bqt
->refcnt
);
868 BUG_ON(!list_empty(&bqt
->busy_list
));
870 kfree(bqt
->tag_index
);
871 bqt
->tag_index
= NULL
;
884 * __blk_queue_free_tags - release tag maintenance info
885 * @q: the request queue for the device
888 * blk_cleanup_queue() will take care of calling this function, if tagging
889 * has been used. So there's no need to call this directly.
891 static void __blk_queue_free_tags(request_queue_t
*q
)
893 struct blk_queue_tag
*bqt
= q
->queue_tags
;
898 __blk_free_tags(bqt
);
900 q
->queue_tags
= NULL
;
901 q
->queue_flags
&= ~(1 << QUEUE_FLAG_QUEUED
);
906 * blk_free_tags - release a given set of tag maintenance info
907 * @bqt: the tag map to free
909 * For externally managed @bqt@ frees the map. Callers of this
910 * function must guarantee to have released all the queues that
911 * might have been using this tag map.
913 void blk_free_tags(struct blk_queue_tag
*bqt
)
915 if (unlikely(!__blk_free_tags(bqt
)))
918 EXPORT_SYMBOL(blk_free_tags
);
921 * blk_queue_free_tags - release tag maintenance info
922 * @q: the request queue for the device
925 * This is used to disabled tagged queuing to a device, yet leave
928 void blk_queue_free_tags(request_queue_t
*q
)
930 clear_bit(QUEUE_FLAG_QUEUED
, &q
->queue_flags
);
933 EXPORT_SYMBOL(blk_queue_free_tags
);
936 init_tag_map(request_queue_t
*q
, struct blk_queue_tag
*tags
, int depth
)
938 struct request
**tag_index
;
939 unsigned long *tag_map
;
942 if (q
&& depth
> q
->nr_requests
* 2) {
943 depth
= q
->nr_requests
* 2;
944 printk(KERN_ERR
"%s: adjusted depth to %d\n",
945 __FUNCTION__
, depth
);
948 tag_index
= kzalloc(depth
* sizeof(struct request
*), GFP_ATOMIC
);
952 nr_ulongs
= ALIGN(depth
, BITS_PER_LONG
) / BITS_PER_LONG
;
953 tag_map
= kzalloc(nr_ulongs
* sizeof(unsigned long), GFP_ATOMIC
);
957 tags
->real_max_depth
= depth
;
958 tags
->max_depth
= depth
;
959 tags
->tag_index
= tag_index
;
960 tags
->tag_map
= tag_map
;
968 static struct blk_queue_tag
*__blk_queue_init_tags(struct request_queue
*q
,
971 struct blk_queue_tag
*tags
;
973 tags
= kmalloc(sizeof(struct blk_queue_tag
), GFP_ATOMIC
);
977 if (init_tag_map(q
, tags
, depth
))
980 INIT_LIST_HEAD(&tags
->busy_list
);
982 atomic_set(&tags
->refcnt
, 1);
990 * blk_init_tags - initialize the tag info for an external tag map
991 * @depth: the maximum queue depth supported
992 * @tags: the tag to use
994 struct blk_queue_tag
*blk_init_tags(int depth
)
996 return __blk_queue_init_tags(NULL
, depth
);
998 EXPORT_SYMBOL(blk_init_tags
);
1001 * blk_queue_init_tags - initialize the queue tag info
1002 * @q: the request queue for the device
1003 * @depth: the maximum queue depth supported
1004 * @tags: the tag to use
1006 int blk_queue_init_tags(request_queue_t
*q
, int depth
,
1007 struct blk_queue_tag
*tags
)
1011 BUG_ON(tags
&& q
->queue_tags
&& tags
!= q
->queue_tags
);
1013 if (!tags
&& !q
->queue_tags
) {
1014 tags
= __blk_queue_init_tags(q
, depth
);
1018 } else if (q
->queue_tags
) {
1019 if ((rc
= blk_queue_resize_tags(q
, depth
)))
1021 set_bit(QUEUE_FLAG_QUEUED
, &q
->queue_flags
);
1024 atomic_inc(&tags
->refcnt
);
1027 * assign it, all done
1029 q
->queue_tags
= tags
;
1030 q
->queue_flags
|= (1 << QUEUE_FLAG_QUEUED
);
1037 EXPORT_SYMBOL(blk_queue_init_tags
);
1040 * blk_queue_resize_tags - change the queueing depth
1041 * @q: the request queue for the device
1042 * @new_depth: the new max command queueing depth
1045 * Must be called with the queue lock held.
1047 int blk_queue_resize_tags(request_queue_t
*q
, int new_depth
)
1049 struct blk_queue_tag
*bqt
= q
->queue_tags
;
1050 struct request
**tag_index
;
1051 unsigned long *tag_map
;
1052 int max_depth
, nr_ulongs
;
1058 * if we already have large enough real_max_depth. just
1059 * adjust max_depth. *NOTE* as requests with tag value
1060 * between new_depth and real_max_depth can be in-flight, tag
1061 * map can not be shrunk blindly here.
1063 if (new_depth
<= bqt
->real_max_depth
) {
1064 bqt
->max_depth
= new_depth
;
1069 * Currently cannot replace a shared tag map with a new
1070 * one, so error out if this is the case
1072 if (atomic_read(&bqt
->refcnt
) != 1)
1076 * save the old state info, so we can copy it back
1078 tag_index
= bqt
->tag_index
;
1079 tag_map
= bqt
->tag_map
;
1080 max_depth
= bqt
->real_max_depth
;
1082 if (init_tag_map(q
, bqt
, new_depth
))
1085 memcpy(bqt
->tag_index
, tag_index
, max_depth
* sizeof(struct request
*));
1086 nr_ulongs
= ALIGN(max_depth
, BITS_PER_LONG
) / BITS_PER_LONG
;
1087 memcpy(bqt
->tag_map
, tag_map
, nr_ulongs
* sizeof(unsigned long));
1094 EXPORT_SYMBOL(blk_queue_resize_tags
);
1097 * blk_queue_end_tag - end tag operations for a request
1098 * @q: the request queue for the device
1099 * @rq: the request that has completed
1102 * Typically called when end_that_request_first() returns 0, meaning
1103 * all transfers have been done for a request. It's important to call
1104 * this function before end_that_request_last(), as that will put the
1105 * request back on the free list thus corrupting the internal tag list.
1108 * queue lock must be held.
1110 void blk_queue_end_tag(request_queue_t
*q
, struct request
*rq
)
1112 struct blk_queue_tag
*bqt
= q
->queue_tags
;
1117 if (unlikely(tag
>= bqt
->real_max_depth
))
1119 * This can happen after tag depth has been reduced.
1120 * FIXME: how about a warning or info message here?
1124 if (unlikely(!__test_and_clear_bit(tag
, bqt
->tag_map
))) {
1125 printk(KERN_ERR
"%s: attempt to clear non-busy tag (%d)\n",
1130 list_del_init(&rq
->queuelist
);
1131 rq
->cmd_flags
&= ~REQ_QUEUED
;
1134 if (unlikely(bqt
->tag_index
[tag
] == NULL
))
1135 printk(KERN_ERR
"%s: tag %d is missing\n",
1138 bqt
->tag_index
[tag
] = NULL
;
1142 EXPORT_SYMBOL(blk_queue_end_tag
);
1145 * blk_queue_start_tag - find a free tag and assign it
1146 * @q: the request queue for the device
1147 * @rq: the block request that needs tagging
1150 * This can either be used as a stand-alone helper, or possibly be
1151 * assigned as the queue &prep_rq_fn (in which case &struct request
1152 * automagically gets a tag assigned). Note that this function
1153 * assumes that any type of request can be queued! if this is not
1154 * true for your device, you must check the request type before
1155 * calling this function. The request will also be removed from
1156 * the request queue, so it's the drivers responsibility to readd
1157 * it if it should need to be restarted for some reason.
1160 * queue lock must be held.
1162 int blk_queue_start_tag(request_queue_t
*q
, struct request
*rq
)
1164 struct blk_queue_tag
*bqt
= q
->queue_tags
;
1167 if (unlikely((rq
->cmd_flags
& REQ_QUEUED
))) {
1169 "%s: request %p for device [%s] already tagged %d",
1171 rq
->rq_disk
? rq
->rq_disk
->disk_name
: "?", rq
->tag
);
1175 tag
= find_first_zero_bit(bqt
->tag_map
, bqt
->max_depth
);
1176 if (tag
>= bqt
->max_depth
)
1179 __set_bit(tag
, bqt
->tag_map
);
1181 rq
->cmd_flags
|= REQ_QUEUED
;
1183 bqt
->tag_index
[tag
] = rq
;
1184 blkdev_dequeue_request(rq
);
1185 list_add(&rq
->queuelist
, &bqt
->busy_list
);
1190 EXPORT_SYMBOL(blk_queue_start_tag
);
1193 * blk_queue_invalidate_tags - invalidate all pending tags
1194 * @q: the request queue for the device
1197 * Hardware conditions may dictate a need to stop all pending requests.
1198 * In this case, we will safely clear the block side of the tag queue and
1199 * readd all requests to the request queue in the right order.
1202 * queue lock must be held.
1204 void blk_queue_invalidate_tags(request_queue_t
*q
)
1206 struct blk_queue_tag
*bqt
= q
->queue_tags
;
1207 struct list_head
*tmp
, *n
;
1210 list_for_each_safe(tmp
, n
, &bqt
->busy_list
) {
1211 rq
= list_entry_rq(tmp
);
1213 if (rq
->tag
== -1) {
1215 "%s: bad tag found on list\n", __FUNCTION__
);
1216 list_del_init(&rq
->queuelist
);
1217 rq
->cmd_flags
&= ~REQ_QUEUED
;
1219 blk_queue_end_tag(q
, rq
);
1221 rq
->cmd_flags
&= ~REQ_STARTED
;
1222 __elv_add_request(q
, rq
, ELEVATOR_INSERT_BACK
, 0);
1226 EXPORT_SYMBOL(blk_queue_invalidate_tags
);
1228 void blk_dump_rq_flags(struct request
*rq
, char *msg
)
1232 printk("%s: dev %s: type=%x, flags=%x\n", msg
,
1233 rq
->rq_disk
? rq
->rq_disk
->disk_name
: "?", rq
->cmd_type
,
1236 printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq
->sector
,
1238 rq
->current_nr_sectors
);
1239 printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq
->bio
, rq
->biotail
, rq
->buffer
, rq
->data
, rq
->data_len
);
1241 if (blk_pc_request(rq
)) {
1243 for (bit
= 0; bit
< sizeof(rq
->cmd
); bit
++)
1244 printk("%02x ", rq
->cmd
[bit
]);
1249 EXPORT_SYMBOL(blk_dump_rq_flags
);
1251 void blk_recount_segments(request_queue_t
*q
, struct bio
*bio
)
1253 struct bio_vec
*bv
, *bvprv
= NULL
;
1254 int i
, nr_phys_segs
, nr_hw_segs
, seg_size
, hw_seg_size
, cluster
;
1255 int high
, highprv
= 1;
1257 if (unlikely(!bio
->bi_io_vec
))
1260 cluster
= q
->queue_flags
& (1 << QUEUE_FLAG_CLUSTER
);
1261 hw_seg_size
= seg_size
= nr_phys_segs
= nr_hw_segs
= 0;
1262 bio_for_each_segment(bv
, bio
, i
) {
1264 * the trick here is making sure that a high page is never
1265 * considered part of another segment, since that might
1266 * change with the bounce page.
1268 high
= page_to_pfn(bv
->bv_page
) >= q
->bounce_pfn
;
1269 if (high
|| highprv
)
1270 goto new_hw_segment
;
1272 if (seg_size
+ bv
->bv_len
> q
->max_segment_size
)
1274 if (!BIOVEC_PHYS_MERGEABLE(bvprv
, bv
))
1276 if (!BIOVEC_SEG_BOUNDARY(q
, bvprv
, bv
))
1278 if (BIOVEC_VIRT_OVERSIZE(hw_seg_size
+ bv
->bv_len
))
1279 goto new_hw_segment
;
1281 seg_size
+= bv
->bv_len
;
1282 hw_seg_size
+= bv
->bv_len
;
1287 if (BIOVEC_VIRT_MERGEABLE(bvprv
, bv
) &&
1288 !BIOVEC_VIRT_OVERSIZE(hw_seg_size
+ bv
->bv_len
)) {
1289 hw_seg_size
+= bv
->bv_len
;
1292 if (hw_seg_size
> bio
->bi_hw_front_size
)
1293 bio
->bi_hw_front_size
= hw_seg_size
;
1294 hw_seg_size
= BIOVEC_VIRT_START_SIZE(bv
) + bv
->bv_len
;
1300 seg_size
= bv
->bv_len
;
1303 if (hw_seg_size
> bio
->bi_hw_back_size
)
1304 bio
->bi_hw_back_size
= hw_seg_size
;
1305 if (nr_hw_segs
== 1 && hw_seg_size
> bio
->bi_hw_front_size
)
1306 bio
->bi_hw_front_size
= hw_seg_size
;
1307 bio
->bi_phys_segments
= nr_phys_segs
;
1308 bio
->bi_hw_segments
= nr_hw_segs
;
1309 bio
->bi_flags
|= (1 << BIO_SEG_VALID
);
1313 static int blk_phys_contig_segment(request_queue_t
*q
, struct bio
*bio
,
1316 if (!(q
->queue_flags
& (1 << QUEUE_FLAG_CLUSTER
)))
1319 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio
), __BVEC_START(nxt
)))
1321 if (bio
->bi_size
+ nxt
->bi_size
> q
->max_segment_size
)
1325 * bio and nxt are contigous in memory, check if the queue allows
1326 * these two to be merged into one
1328 if (BIO_SEG_BOUNDARY(q
, bio
, nxt
))
1334 static int blk_hw_contig_segment(request_queue_t
*q
, struct bio
*bio
,
1337 if (unlikely(!bio_flagged(bio
, BIO_SEG_VALID
)))
1338 blk_recount_segments(q
, bio
);
1339 if (unlikely(!bio_flagged(nxt
, BIO_SEG_VALID
)))
1340 blk_recount_segments(q
, nxt
);
1341 if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio
), __BVEC_START(nxt
)) ||
1342 BIOVEC_VIRT_OVERSIZE(bio
->bi_hw_front_size
+ bio
->bi_hw_back_size
))
1344 if (bio
->bi_size
+ nxt
->bi_size
> q
->max_segment_size
)
1351 * map a request to scatterlist, return number of sg entries setup. Caller
1352 * must make sure sg can hold rq->nr_phys_segments entries
1354 int blk_rq_map_sg(request_queue_t
*q
, struct request
*rq
, struct scatterlist
*sg
)
1356 struct bio_vec
*bvec
, *bvprv
;
1358 int nsegs
, i
, cluster
;
1361 cluster
= q
->queue_flags
& (1 << QUEUE_FLAG_CLUSTER
);
1364 * for each bio in rq
1367 rq_for_each_bio(bio
, rq
) {
1369 * for each segment in bio
1371 bio_for_each_segment(bvec
, bio
, i
) {
1372 int nbytes
= bvec
->bv_len
;
1374 if (bvprv
&& cluster
) {
1375 if (sg
[nsegs
- 1].length
+ nbytes
> q
->max_segment_size
)
1378 if (!BIOVEC_PHYS_MERGEABLE(bvprv
, bvec
))
1380 if (!BIOVEC_SEG_BOUNDARY(q
, bvprv
, bvec
))
1383 sg
[nsegs
- 1].length
+= nbytes
;
1386 memset(&sg
[nsegs
],0,sizeof(struct scatterlist
));
1387 sg
[nsegs
].page
= bvec
->bv_page
;
1388 sg
[nsegs
].length
= nbytes
;
1389 sg
[nsegs
].offset
= bvec
->bv_offset
;
1394 } /* segments in bio */
1400 EXPORT_SYMBOL(blk_rq_map_sg
);
1403 * the standard queue merge functions, can be overridden with device
1404 * specific ones if so desired
1407 static inline int ll_new_mergeable(request_queue_t
*q
,
1408 struct request
*req
,
1411 int nr_phys_segs
= bio_phys_segments(q
, bio
);
1413 if (req
->nr_phys_segments
+ nr_phys_segs
> q
->max_phys_segments
) {
1414 req
->cmd_flags
|= REQ_NOMERGE
;
1415 if (req
== q
->last_merge
)
1416 q
->last_merge
= NULL
;
1421 * A hw segment is just getting larger, bump just the phys
1424 req
->nr_phys_segments
+= nr_phys_segs
;
1428 static inline int ll_new_hw_segment(request_queue_t
*q
,
1429 struct request
*req
,
1432 int nr_hw_segs
= bio_hw_segments(q
, bio
);
1433 int nr_phys_segs
= bio_phys_segments(q
, bio
);
1435 if (req
->nr_hw_segments
+ nr_hw_segs
> q
->max_hw_segments
1436 || req
->nr_phys_segments
+ nr_phys_segs
> q
->max_phys_segments
) {
1437 req
->cmd_flags
|= REQ_NOMERGE
;
1438 if (req
== q
->last_merge
)
1439 q
->last_merge
= NULL
;
1444 * This will form the start of a new hw segment. Bump both
1447 req
->nr_hw_segments
+= nr_hw_segs
;
1448 req
->nr_phys_segments
+= nr_phys_segs
;
1452 static int ll_back_merge_fn(request_queue_t
*q
, struct request
*req
,
1455 unsigned short max_sectors
;
1458 if (unlikely(blk_pc_request(req
)))
1459 max_sectors
= q
->max_hw_sectors
;
1461 max_sectors
= q
->max_sectors
;
1463 if (req
->nr_sectors
+ bio_sectors(bio
) > max_sectors
) {
1464 req
->cmd_flags
|= REQ_NOMERGE
;
1465 if (req
== q
->last_merge
)
1466 q
->last_merge
= NULL
;
1469 if (unlikely(!bio_flagged(req
->biotail
, BIO_SEG_VALID
)))
1470 blk_recount_segments(q
, req
->biotail
);
1471 if (unlikely(!bio_flagged(bio
, BIO_SEG_VALID
)))
1472 blk_recount_segments(q
, bio
);
1473 len
= req
->biotail
->bi_hw_back_size
+ bio
->bi_hw_front_size
;
1474 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req
->biotail
), __BVEC_START(bio
)) &&
1475 !BIOVEC_VIRT_OVERSIZE(len
)) {
1476 int mergeable
= ll_new_mergeable(q
, req
, bio
);
1479 if (req
->nr_hw_segments
== 1)
1480 req
->bio
->bi_hw_front_size
= len
;
1481 if (bio
->bi_hw_segments
== 1)
1482 bio
->bi_hw_back_size
= len
;
1487 return ll_new_hw_segment(q
, req
, bio
);
1490 static int ll_front_merge_fn(request_queue_t
*q
, struct request
*req
,
1493 unsigned short max_sectors
;
1496 if (unlikely(blk_pc_request(req
)))
1497 max_sectors
= q
->max_hw_sectors
;
1499 max_sectors
= q
->max_sectors
;
1502 if (req
->nr_sectors
+ bio_sectors(bio
) > max_sectors
) {
1503 req
->cmd_flags
|= REQ_NOMERGE
;
1504 if (req
== q
->last_merge
)
1505 q
->last_merge
= NULL
;
1508 len
= bio
->bi_hw_back_size
+ req
->bio
->bi_hw_front_size
;
1509 if (unlikely(!bio_flagged(bio
, BIO_SEG_VALID
)))
1510 blk_recount_segments(q
, bio
);
1511 if (unlikely(!bio_flagged(req
->bio
, BIO_SEG_VALID
)))
1512 blk_recount_segments(q
, req
->bio
);
1513 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio
), __BVEC_START(req
->bio
)) &&
1514 !BIOVEC_VIRT_OVERSIZE(len
)) {
1515 int mergeable
= ll_new_mergeable(q
, req
, bio
);
1518 if (bio
->bi_hw_segments
== 1)
1519 bio
->bi_hw_front_size
= len
;
1520 if (req
->nr_hw_segments
== 1)
1521 req
->biotail
->bi_hw_back_size
= len
;
1526 return ll_new_hw_segment(q
, req
, bio
);
1529 static int ll_merge_requests_fn(request_queue_t
*q
, struct request
*req
,
1530 struct request
*next
)
1532 int total_phys_segments
;
1533 int total_hw_segments
;
1536 * First check if the either of the requests are re-queued
1537 * requests. Can't merge them if they are.
1539 if (req
->special
|| next
->special
)
1543 * Will it become too large?
1545 if ((req
->nr_sectors
+ next
->nr_sectors
) > q
->max_sectors
)
1548 total_phys_segments
= req
->nr_phys_segments
+ next
->nr_phys_segments
;
1549 if (blk_phys_contig_segment(q
, req
->biotail
, next
->bio
))
1550 total_phys_segments
--;
1552 if (total_phys_segments
> q
->max_phys_segments
)
1555 total_hw_segments
= req
->nr_hw_segments
+ next
->nr_hw_segments
;
1556 if (blk_hw_contig_segment(q
, req
->biotail
, next
->bio
)) {
1557 int len
= req
->biotail
->bi_hw_back_size
+ next
->bio
->bi_hw_front_size
;
1559 * propagate the combined length to the end of the requests
1561 if (req
->nr_hw_segments
== 1)
1562 req
->bio
->bi_hw_front_size
= len
;
1563 if (next
->nr_hw_segments
== 1)
1564 next
->biotail
->bi_hw_back_size
= len
;
1565 total_hw_segments
--;
1568 if (total_hw_segments
> q
->max_hw_segments
)
1571 /* Merge is OK... */
1572 req
->nr_phys_segments
= total_phys_segments
;
1573 req
->nr_hw_segments
= total_hw_segments
;
1578 * "plug" the device if there are no outstanding requests: this will
1579 * force the transfer to start only after we have put all the requests
1582 * This is called with interrupts off and no requests on the queue and
1583 * with the queue lock held.
1585 void blk_plug_device(request_queue_t
*q
)
1587 WARN_ON(!irqs_disabled());
1590 * don't plug a stopped queue, it must be paired with blk_start_queue()
1591 * which will restart the queueing
1593 if (blk_queue_stopped(q
))
1596 if (!test_and_set_bit(QUEUE_FLAG_PLUGGED
, &q
->queue_flags
)) {
1597 mod_timer(&q
->unplug_timer
, jiffies
+ q
->unplug_delay
);
1598 blk_add_trace_generic(q
, NULL
, 0, BLK_TA_PLUG
);
1602 EXPORT_SYMBOL(blk_plug_device
);
1605 * remove the queue from the plugged list, if present. called with
1606 * queue lock held and interrupts disabled.
1608 int blk_remove_plug(request_queue_t
*q
)
1610 WARN_ON(!irqs_disabled());
1612 if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED
, &q
->queue_flags
))
1615 del_timer(&q
->unplug_timer
);
1619 EXPORT_SYMBOL(blk_remove_plug
);
1622 * remove the plug and let it rip..
1624 void __generic_unplug_device(request_queue_t
*q
)
1626 if (unlikely(blk_queue_stopped(q
)))
1629 if (!blk_remove_plug(q
))
1634 EXPORT_SYMBOL(__generic_unplug_device
);
1637 * generic_unplug_device - fire a request queue
1638 * @q: The &request_queue_t in question
1641 * Linux uses plugging to build bigger requests queues before letting
1642 * the device have at them. If a queue is plugged, the I/O scheduler
1643 * is still adding and merging requests on the queue. Once the queue
1644 * gets unplugged, the request_fn defined for the queue is invoked and
1645 * transfers started.
1647 void generic_unplug_device(request_queue_t
*q
)
1649 spin_lock_irq(q
->queue_lock
);
1650 __generic_unplug_device(q
);
1651 spin_unlock_irq(q
->queue_lock
);
1653 EXPORT_SYMBOL(generic_unplug_device
);
1655 static void blk_backing_dev_unplug(struct backing_dev_info
*bdi
,
1658 request_queue_t
*q
= bdi
->unplug_io_data
;
1661 * devices don't necessarily have an ->unplug_fn defined
1664 blk_add_trace_pdu_int(q
, BLK_TA_UNPLUG_IO
, NULL
,
1665 q
->rq
.count
[READ
] + q
->rq
.count
[WRITE
]);
1671 static void blk_unplug_work(void *data
)
1673 request_queue_t
*q
= data
;
1675 blk_add_trace_pdu_int(q
, BLK_TA_UNPLUG_IO
, NULL
,
1676 q
->rq
.count
[READ
] + q
->rq
.count
[WRITE
]);
1681 static void blk_unplug_timeout(unsigned long data
)
1683 request_queue_t
*q
= (request_queue_t
*)data
;
1685 blk_add_trace_pdu_int(q
, BLK_TA_UNPLUG_TIMER
, NULL
,
1686 q
->rq
.count
[READ
] + q
->rq
.count
[WRITE
]);
1688 kblockd_schedule_work(&q
->unplug_work
);
1692 * blk_start_queue - restart a previously stopped queue
1693 * @q: The &request_queue_t in question
1696 * blk_start_queue() will clear the stop flag on the queue, and call
1697 * the request_fn for the queue if it was in a stopped state when
1698 * entered. Also see blk_stop_queue(). Queue lock must be held.
1700 void blk_start_queue(request_queue_t
*q
)
1702 WARN_ON(!irqs_disabled());
1704 clear_bit(QUEUE_FLAG_STOPPED
, &q
->queue_flags
);
1707 * one level of recursion is ok and is much faster than kicking
1708 * the unplug handling
1710 if (!test_and_set_bit(QUEUE_FLAG_REENTER
, &q
->queue_flags
)) {
1712 clear_bit(QUEUE_FLAG_REENTER
, &q
->queue_flags
);
1715 kblockd_schedule_work(&q
->unplug_work
);
1719 EXPORT_SYMBOL(blk_start_queue
);
1722 * blk_stop_queue - stop a queue
1723 * @q: The &request_queue_t in question
1726 * The Linux block layer assumes that a block driver will consume all
1727 * entries on the request queue when the request_fn strategy is called.
1728 * Often this will not happen, because of hardware limitations (queue
1729 * depth settings). If a device driver gets a 'queue full' response,
1730 * or if it simply chooses not to queue more I/O at one point, it can
1731 * call this function to prevent the request_fn from being called until
1732 * the driver has signalled it's ready to go again. This happens by calling
1733 * blk_start_queue() to restart queue operations. Queue lock must be held.
1735 void blk_stop_queue(request_queue_t
*q
)
1738 set_bit(QUEUE_FLAG_STOPPED
, &q
->queue_flags
);
1740 EXPORT_SYMBOL(blk_stop_queue
);
1743 * blk_sync_queue - cancel any pending callbacks on a queue
1747 * The block layer may perform asynchronous callback activity
1748 * on a queue, such as calling the unplug function after a timeout.
1749 * A block device may call blk_sync_queue to ensure that any
1750 * such activity is cancelled, thus allowing it to release resources
1751 * the the callbacks might use. The caller must already have made sure
1752 * that its ->make_request_fn will not re-add plugging prior to calling
1756 void blk_sync_queue(struct request_queue
*q
)
1758 del_timer_sync(&q
->unplug_timer
);
1761 EXPORT_SYMBOL(blk_sync_queue
);
1764 * blk_run_queue - run a single device queue
1765 * @q: The queue to run
1767 void blk_run_queue(struct request_queue
*q
)
1769 unsigned long flags
;
1771 spin_lock_irqsave(q
->queue_lock
, flags
);
1775 * Only recurse once to avoid overrunning the stack, let the unplug
1776 * handling reinvoke the handler shortly if we already got there.
1778 if (!elv_queue_empty(q
)) {
1779 if (!test_and_set_bit(QUEUE_FLAG_REENTER
, &q
->queue_flags
)) {
1781 clear_bit(QUEUE_FLAG_REENTER
, &q
->queue_flags
);
1784 kblockd_schedule_work(&q
->unplug_work
);
1788 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1790 EXPORT_SYMBOL(blk_run_queue
);
1793 * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1794 * @kobj: the kobj belonging of the request queue to be released
1797 * blk_cleanup_queue is the pair to blk_init_queue() or
1798 * blk_queue_make_request(). It should be called when a request queue is
1799 * being released; typically when a block device is being de-registered.
1800 * Currently, its primary task it to free all the &struct request
1801 * structures that were allocated to the queue and the queue itself.
1804 * Hopefully the low level driver will have finished any
1805 * outstanding requests first...
1807 static void blk_release_queue(struct kobject
*kobj
)
1809 request_queue_t
*q
= container_of(kobj
, struct request_queue
, kobj
);
1810 struct request_list
*rl
= &q
->rq
;
1815 mempool_destroy(rl
->rq_pool
);
1818 __blk_queue_free_tags(q
);
1820 blk_trace_shutdown(q
);
1822 kmem_cache_free(requestq_cachep
, q
);
1825 void blk_put_queue(request_queue_t
*q
)
1827 kobject_put(&q
->kobj
);
1829 EXPORT_SYMBOL(blk_put_queue
);
1831 void blk_cleanup_queue(request_queue_t
* q
)
1833 mutex_lock(&q
->sysfs_lock
);
1834 set_bit(QUEUE_FLAG_DEAD
, &q
->queue_flags
);
1835 mutex_unlock(&q
->sysfs_lock
);
1838 elevator_exit(q
->elevator
);
1843 EXPORT_SYMBOL(blk_cleanup_queue
);
1845 static int blk_init_free_list(request_queue_t
*q
)
1847 struct request_list
*rl
= &q
->rq
;
1849 rl
->count
[READ
] = rl
->count
[WRITE
] = 0;
1850 rl
->starved
[READ
] = rl
->starved
[WRITE
] = 0;
1852 init_waitqueue_head(&rl
->wait
[READ
]);
1853 init_waitqueue_head(&rl
->wait
[WRITE
]);
1855 rl
->rq_pool
= mempool_create_node(BLKDEV_MIN_RQ
, mempool_alloc_slab
,
1856 mempool_free_slab
, request_cachep
, q
->node
);
1864 request_queue_t
*blk_alloc_queue(gfp_t gfp_mask
)
1866 return blk_alloc_queue_node(gfp_mask
, -1);
1868 EXPORT_SYMBOL(blk_alloc_queue
);
1870 static struct kobj_type queue_ktype
;
1872 request_queue_t
*blk_alloc_queue_node(gfp_t gfp_mask
, int node_id
)
1876 q
= kmem_cache_alloc_node(requestq_cachep
, gfp_mask
, node_id
);
1880 memset(q
, 0, sizeof(*q
));
1881 init_timer(&q
->unplug_timer
);
1883 snprintf(q
->kobj
.name
, KOBJ_NAME_LEN
, "%s", "queue");
1884 q
->kobj
.ktype
= &queue_ktype
;
1885 kobject_init(&q
->kobj
);
1887 q
->backing_dev_info
.unplug_io_fn
= blk_backing_dev_unplug
;
1888 q
->backing_dev_info
.unplug_io_data
= q
;
1890 mutex_init(&q
->sysfs_lock
);
1894 EXPORT_SYMBOL(blk_alloc_queue_node
);
1897 * blk_init_queue - prepare a request queue for use with a block device
1898 * @rfn: The function to be called to process requests that have been
1899 * placed on the queue.
1900 * @lock: Request queue spin lock
1903 * If a block device wishes to use the standard request handling procedures,
1904 * which sorts requests and coalesces adjacent requests, then it must
1905 * call blk_init_queue(). The function @rfn will be called when there
1906 * are requests on the queue that need to be processed. If the device
1907 * supports plugging, then @rfn may not be called immediately when requests
1908 * are available on the queue, but may be called at some time later instead.
1909 * Plugged queues are generally unplugged when a buffer belonging to one
1910 * of the requests on the queue is needed, or due to memory pressure.
1912 * @rfn is not required, or even expected, to remove all requests off the
1913 * queue, but only as many as it can handle at a time. If it does leave
1914 * requests on the queue, it is responsible for arranging that the requests
1915 * get dealt with eventually.
1917 * The queue spin lock must be held while manipulating the requests on the
1918 * request queue; this lock will be taken also from interrupt context, so irq
1919 * disabling is needed for it.
1921 * Function returns a pointer to the initialized request queue, or NULL if
1922 * it didn't succeed.
1925 * blk_init_queue() must be paired with a blk_cleanup_queue() call
1926 * when the block device is deactivated (such as at module unload).
1929 request_queue_t
*blk_init_queue(request_fn_proc
*rfn
, spinlock_t
*lock
)
1931 return blk_init_queue_node(rfn
, lock
, -1);
1933 EXPORT_SYMBOL(blk_init_queue
);
1936 blk_init_queue_node(request_fn_proc
*rfn
, spinlock_t
*lock
, int node_id
)
1938 request_queue_t
*q
= blk_alloc_queue_node(GFP_KERNEL
, node_id
);
1944 if (blk_init_free_list(q
)) {
1945 kmem_cache_free(requestq_cachep
, q
);
1950 * if caller didn't supply a lock, they get per-queue locking with
1954 spin_lock_init(&q
->__queue_lock
);
1955 lock
= &q
->__queue_lock
;
1958 q
->request_fn
= rfn
;
1959 q
->back_merge_fn
= ll_back_merge_fn
;
1960 q
->front_merge_fn
= ll_front_merge_fn
;
1961 q
->merge_requests_fn
= ll_merge_requests_fn
;
1962 q
->prep_rq_fn
= NULL
;
1963 q
->unplug_fn
= generic_unplug_device
;
1964 q
->queue_flags
= (1 << QUEUE_FLAG_CLUSTER
);
1965 q
->queue_lock
= lock
;
1967 blk_queue_segment_boundary(q
, 0xffffffff);
1969 blk_queue_make_request(q
, __make_request
);
1970 blk_queue_max_segment_size(q
, MAX_SEGMENT_SIZE
);
1972 blk_queue_max_hw_segments(q
, MAX_HW_SEGMENTS
);
1973 blk_queue_max_phys_segments(q
, MAX_PHYS_SEGMENTS
);
1978 if (!elevator_init(q
, NULL
)) {
1979 blk_queue_congestion_threshold(q
);
1986 EXPORT_SYMBOL(blk_init_queue_node
);
1988 int blk_get_queue(request_queue_t
*q
)
1990 if (likely(!test_bit(QUEUE_FLAG_DEAD
, &q
->queue_flags
))) {
1991 kobject_get(&q
->kobj
);
1998 EXPORT_SYMBOL(blk_get_queue
);
2000 static inline void blk_free_request(request_queue_t
*q
, struct request
*rq
)
2002 if (rq
->cmd_flags
& REQ_ELVPRIV
)
2003 elv_put_request(q
, rq
);
2004 mempool_free(rq
, q
->rq
.rq_pool
);
2007 static inline struct request
*
2008 blk_alloc_request(request_queue_t
*q
, int rw
, struct bio
*bio
,
2009 int priv
, gfp_t gfp_mask
)
2011 struct request
*rq
= mempool_alloc(q
->rq
.rq_pool
, gfp_mask
);
2017 * first three bits are identical in rq->cmd_flags and bio->bi_rw,
2018 * see bio.h and blkdev.h
2023 if (unlikely(elv_set_request(q
, rq
, bio
, gfp_mask
))) {
2024 mempool_free(rq
, q
->rq
.rq_pool
);
2027 rq
->cmd_flags
|= REQ_ELVPRIV
;
2034 * ioc_batching returns true if the ioc is a valid batching request and
2035 * should be given priority access to a request.
2037 static inline int ioc_batching(request_queue_t
*q
, struct io_context
*ioc
)
2043 * Make sure the process is able to allocate at least 1 request
2044 * even if the batch times out, otherwise we could theoretically
2047 return ioc
->nr_batch_requests
== q
->nr_batching
||
2048 (ioc
->nr_batch_requests
> 0
2049 && time_before(jiffies
, ioc
->last_waited
+ BLK_BATCH_TIME
));
2053 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
2054 * will cause the process to be a "batcher" on all queues in the system. This
2055 * is the behaviour we want though - once it gets a wakeup it should be given
2058 static void ioc_set_batching(request_queue_t
*q
, struct io_context
*ioc
)
2060 if (!ioc
|| ioc_batching(q
, ioc
))
2063 ioc
->nr_batch_requests
= q
->nr_batching
;
2064 ioc
->last_waited
= jiffies
;
2067 static void __freed_request(request_queue_t
*q
, int rw
)
2069 struct request_list
*rl
= &q
->rq
;
2071 if (rl
->count
[rw
] < queue_congestion_off_threshold(q
))
2072 clear_queue_congested(q
, rw
);
2074 if (rl
->count
[rw
] + 1 <= q
->nr_requests
) {
2075 if (waitqueue_active(&rl
->wait
[rw
]))
2076 wake_up(&rl
->wait
[rw
]);
2078 blk_clear_queue_full(q
, rw
);
2083 * A request has just been released. Account for it, update the full and
2084 * congestion status, wake up any waiters. Called under q->queue_lock.
2086 static void freed_request(request_queue_t
*q
, int rw
, int priv
)
2088 struct request_list
*rl
= &q
->rq
;
2094 __freed_request(q
, rw
);
2096 if (unlikely(rl
->starved
[rw
^ 1]))
2097 __freed_request(q
, rw
^ 1);
2100 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2102 * Get a free request, queue_lock must be held.
2103 * Returns NULL on failure, with queue_lock held.
2104 * Returns !NULL on success, with queue_lock *not held*.
2106 static struct request
*get_request(request_queue_t
*q
, int rw
, struct bio
*bio
,
2109 struct request
*rq
= NULL
;
2110 struct request_list
*rl
= &q
->rq
;
2111 struct io_context
*ioc
= NULL
;
2112 int may_queue
, priv
;
2114 may_queue
= elv_may_queue(q
, rw
, bio
);
2115 if (may_queue
== ELV_MQUEUE_NO
)
2118 if (rl
->count
[rw
]+1 >= queue_congestion_on_threshold(q
)) {
2119 if (rl
->count
[rw
]+1 >= q
->nr_requests
) {
2120 ioc
= current_io_context(GFP_ATOMIC
);
2122 * The queue will fill after this allocation, so set
2123 * it as full, and mark this process as "batching".
2124 * This process will be allowed to complete a batch of
2125 * requests, others will be blocked.
2127 if (!blk_queue_full(q
, rw
)) {
2128 ioc_set_batching(q
, ioc
);
2129 blk_set_queue_full(q
, rw
);
2131 if (may_queue
!= ELV_MQUEUE_MUST
2132 && !ioc_batching(q
, ioc
)) {
2134 * The queue is full and the allocating
2135 * process is not a "batcher", and not
2136 * exempted by the IO scheduler
2142 set_queue_congested(q
, rw
);
2146 * Only allow batching queuers to allocate up to 50% over the defined
2147 * limit of requests, otherwise we could have thousands of requests
2148 * allocated with any setting of ->nr_requests
2150 if (rl
->count
[rw
] >= (3 * q
->nr_requests
/ 2))
2154 rl
->starved
[rw
] = 0;
2156 priv
= !test_bit(QUEUE_FLAG_ELVSWITCH
, &q
->queue_flags
);
2160 spin_unlock_irq(q
->queue_lock
);
2162 rq
= blk_alloc_request(q
, rw
, bio
, priv
, gfp_mask
);
2163 if (unlikely(!rq
)) {
2165 * Allocation failed presumably due to memory. Undo anything
2166 * we might have messed up.
2168 * Allocating task should really be put onto the front of the
2169 * wait queue, but this is pretty rare.
2171 spin_lock_irq(q
->queue_lock
);
2172 freed_request(q
, rw
, priv
);
2175 * in the very unlikely event that allocation failed and no
2176 * requests for this direction was pending, mark us starved
2177 * so that freeing of a request in the other direction will
2178 * notice us. another possible fix would be to split the
2179 * rq mempool into READ and WRITE
2182 if (unlikely(rl
->count
[rw
] == 0))
2183 rl
->starved
[rw
] = 1;
2189 * ioc may be NULL here, and ioc_batching will be false. That's
2190 * OK, if the queue is under the request limit then requests need
2191 * not count toward the nr_batch_requests limit. There will always
2192 * be some limit enforced by BLK_BATCH_TIME.
2194 if (ioc_batching(q
, ioc
))
2195 ioc
->nr_batch_requests
--;
2200 blk_add_trace_generic(q
, bio
, rw
, BLK_TA_GETRQ
);
2206 * No available requests for this queue, unplug the device and wait for some
2207 * requests to become available.
2209 * Called with q->queue_lock held, and returns with it unlocked.
2211 static struct request
*get_request_wait(request_queue_t
*q
, int rw
,
2216 rq
= get_request(q
, rw
, bio
, GFP_NOIO
);
2219 struct request_list
*rl
= &q
->rq
;
2221 prepare_to_wait_exclusive(&rl
->wait
[rw
], &wait
,
2222 TASK_UNINTERRUPTIBLE
);
2224 rq
= get_request(q
, rw
, bio
, GFP_NOIO
);
2227 struct io_context
*ioc
;
2229 blk_add_trace_generic(q
, bio
, rw
, BLK_TA_SLEEPRQ
);
2231 __generic_unplug_device(q
);
2232 spin_unlock_irq(q
->queue_lock
);
2236 * After sleeping, we become a "batching" process and
2237 * will be able to allocate at least one request, and
2238 * up to a big batch of them for a small period time.
2239 * See ioc_batching, ioc_set_batching
2241 ioc
= current_io_context(GFP_NOIO
);
2242 ioc_set_batching(q
, ioc
);
2244 spin_lock_irq(q
->queue_lock
);
2246 finish_wait(&rl
->wait
[rw
], &wait
);
2252 struct request
*blk_get_request(request_queue_t
*q
, int rw
, gfp_t gfp_mask
)
2256 BUG_ON(rw
!= READ
&& rw
!= WRITE
);
2258 spin_lock_irq(q
->queue_lock
);
2259 if (gfp_mask
& __GFP_WAIT
) {
2260 rq
= get_request_wait(q
, rw
, NULL
);
2262 rq
= get_request(q
, rw
, NULL
, gfp_mask
);
2264 spin_unlock_irq(q
->queue_lock
);
2266 /* q->queue_lock is unlocked at this point */
2270 EXPORT_SYMBOL(blk_get_request
);
2273 * blk_requeue_request - put a request back on queue
2274 * @q: request queue where request should be inserted
2275 * @rq: request to be inserted
2278 * Drivers often keep queueing requests until the hardware cannot accept
2279 * more, when that condition happens we need to put the request back
2280 * on the queue. Must be called with queue lock held.
2282 void blk_requeue_request(request_queue_t
*q
, struct request
*rq
)
2284 blk_add_trace_rq(q
, rq
, BLK_TA_REQUEUE
);
2286 if (blk_rq_tagged(rq
))
2287 blk_queue_end_tag(q
, rq
);
2289 elv_requeue_request(q
, rq
);
2292 EXPORT_SYMBOL(blk_requeue_request
);
2295 * blk_insert_request - insert a special request in to a request queue
2296 * @q: request queue where request should be inserted
2297 * @rq: request to be inserted
2298 * @at_head: insert request at head or tail of queue
2299 * @data: private data
2302 * Many block devices need to execute commands asynchronously, so they don't
2303 * block the whole kernel from preemption during request execution. This is
2304 * accomplished normally by inserting aritficial requests tagged as
2305 * REQ_SPECIAL in to the corresponding request queue, and letting them be
2306 * scheduled for actual execution by the request queue.
2308 * We have the option of inserting the head or the tail of the queue.
2309 * Typically we use the tail for new ioctls and so forth. We use the head
2310 * of the queue for things like a QUEUE_FULL message from a device, or a
2311 * host that is unable to accept a particular command.
2313 void blk_insert_request(request_queue_t
*q
, struct request
*rq
,
2314 int at_head
, void *data
)
2316 int where
= at_head
? ELEVATOR_INSERT_FRONT
: ELEVATOR_INSERT_BACK
;
2317 unsigned long flags
;
2320 * tell I/O scheduler that this isn't a regular read/write (ie it
2321 * must not attempt merges on this) and that it acts as a soft
2324 rq
->cmd_type
= REQ_TYPE_SPECIAL
;
2325 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
2329 spin_lock_irqsave(q
->queue_lock
, flags
);
2332 * If command is tagged, release the tag
2334 if (blk_rq_tagged(rq
))
2335 blk_queue_end_tag(q
, rq
);
2337 drive_stat_acct(rq
, rq
->nr_sectors
, 1);
2338 __elv_add_request(q
, rq
, where
, 0);
2340 if (blk_queue_plugged(q
))
2341 __generic_unplug_device(q
);
2344 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2347 EXPORT_SYMBOL(blk_insert_request
);
2350 * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2351 * @q: request queue where request should be inserted
2352 * @rq: request structure to fill
2353 * @ubuf: the user buffer
2354 * @len: length of user data
2357 * Data will be mapped directly for zero copy io, if possible. Otherwise
2358 * a kernel bounce buffer is used.
2360 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2361 * still in process context.
2363 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2364 * before being submitted to the device, as pages mapped may be out of
2365 * reach. It's the callers responsibility to make sure this happens. The
2366 * original bio must be passed back in to blk_rq_unmap_user() for proper
2369 int blk_rq_map_user(request_queue_t
*q
, struct request
*rq
, void __user
*ubuf
,
2372 unsigned long uaddr
;
2376 if (len
> (q
->max_hw_sectors
<< 9))
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 rq
->bio
= rq
->biotail
= bio
;
2395 blk_rq_bio_prep(q
, rq
, bio
);
2397 rq
->buffer
= rq
->data
= NULL
;
2403 * bio is the err-ptr
2405 return PTR_ERR(bio
);
2408 EXPORT_SYMBOL(blk_rq_map_user
);
2411 * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2412 * @q: request queue where request should be inserted
2413 * @rq: request to map data to
2414 * @iov: pointer to the iovec
2415 * @iov_count: number of elements in the iovec
2418 * Data will be mapped directly for zero copy io, if possible. Otherwise
2419 * a kernel bounce buffer is used.
2421 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2422 * still in process context.
2424 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2425 * before being submitted to the device, as pages mapped may be out of
2426 * reach. It's the callers responsibility to make sure this happens. The
2427 * original bio must be passed back in to blk_rq_unmap_user() for proper
2430 int blk_rq_map_user_iov(request_queue_t
*q
, struct request
*rq
,
2431 struct sg_iovec
*iov
, int iov_count
)
2435 if (!iov
|| iov_count
<= 0)
2438 /* we don't allow misaligned data like bio_map_user() does. If the
2439 * user is using sg, they're expected to know the alignment constraints
2440 * and respect them accordingly */
2441 bio
= bio_map_user_iov(q
, NULL
, iov
, iov_count
, rq_data_dir(rq
)== READ
);
2443 return PTR_ERR(bio
);
2445 rq
->bio
= rq
->biotail
= bio
;
2446 blk_rq_bio_prep(q
, rq
, bio
);
2447 rq
->buffer
= rq
->data
= NULL
;
2448 rq
->data_len
= bio
->bi_size
;
2452 EXPORT_SYMBOL(blk_rq_map_user_iov
);
2455 * blk_rq_unmap_user - unmap a request with user data
2456 * @bio: bio to be unmapped
2457 * @ulen: length of user buffer
2460 * Unmap a bio previously mapped by blk_rq_map_user().
2462 int blk_rq_unmap_user(struct bio
*bio
, unsigned int ulen
)
2467 if (bio_flagged(bio
, BIO_USER_MAPPED
))
2468 bio_unmap_user(bio
);
2470 ret
= bio_uncopy_user(bio
);
2476 EXPORT_SYMBOL(blk_rq_unmap_user
);
2479 * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2480 * @q: request queue where request should be inserted
2481 * @rq: request to fill
2482 * @kbuf: the kernel buffer
2483 * @len: length of user data
2484 * @gfp_mask: memory allocation flags
2486 int blk_rq_map_kern(request_queue_t
*q
, struct request
*rq
, void *kbuf
,
2487 unsigned int len
, gfp_t gfp_mask
)
2491 if (len
> (q
->max_hw_sectors
<< 9))
2496 bio
= bio_map_kern(q
, kbuf
, len
, gfp_mask
);
2498 return PTR_ERR(bio
);
2500 if (rq_data_dir(rq
) == WRITE
)
2501 bio
->bi_rw
|= (1 << BIO_RW
);
2503 rq
->bio
= rq
->biotail
= bio
;
2504 blk_rq_bio_prep(q
, rq
, bio
);
2506 rq
->buffer
= rq
->data
= NULL
;
2511 EXPORT_SYMBOL(blk_rq_map_kern
);
2514 * blk_execute_rq_nowait - insert a request into queue for execution
2515 * @q: queue to insert the request in
2516 * @bd_disk: matching gendisk
2517 * @rq: request to insert
2518 * @at_head: insert request at head or tail of queue
2519 * @done: I/O completion handler
2522 * Insert a fully prepared request at the back of the io scheduler queue
2523 * for execution. Don't wait for completion.
2525 void blk_execute_rq_nowait(request_queue_t
*q
, struct gendisk
*bd_disk
,
2526 struct request
*rq
, int at_head
,
2529 int where
= at_head
? ELEVATOR_INSERT_FRONT
: ELEVATOR_INSERT_BACK
;
2531 rq
->rq_disk
= bd_disk
;
2532 rq
->cmd_flags
|= REQ_NOMERGE
;
2534 WARN_ON(irqs_disabled());
2535 spin_lock_irq(q
->queue_lock
);
2536 __elv_add_request(q
, rq
, where
, 1);
2537 __generic_unplug_device(q
);
2538 spin_unlock_irq(q
->queue_lock
);
2540 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait
);
2543 * blk_execute_rq - insert a request into queue for execution
2544 * @q: queue to insert the request in
2545 * @bd_disk: matching gendisk
2546 * @rq: request to insert
2547 * @at_head: insert request at head or tail of queue
2550 * Insert a fully prepared request at the back of the io scheduler queue
2551 * for execution and wait for completion.
2553 int blk_execute_rq(request_queue_t
*q
, struct gendisk
*bd_disk
,
2554 struct request
*rq
, int at_head
)
2556 DECLARE_COMPLETION_ONSTACK(wait
);
2557 char sense
[SCSI_SENSE_BUFFERSIZE
];
2561 * we need an extra reference to the request, so we can look at
2562 * it after io completion
2567 memset(sense
, 0, sizeof(sense
));
2572 rq
->waiting
= &wait
;
2573 blk_execute_rq_nowait(q
, bd_disk
, rq
, at_head
, blk_end_sync_rq
);
2574 wait_for_completion(&wait
);
2583 EXPORT_SYMBOL(blk_execute_rq
);
2586 * blkdev_issue_flush - queue a flush
2587 * @bdev: blockdev to issue flush for
2588 * @error_sector: error sector
2591 * Issue a flush for the block device in question. Caller can supply
2592 * room for storing the error offset in case of a flush error, if they
2593 * wish to. Caller must run wait_for_completion() on its own.
2595 int blkdev_issue_flush(struct block_device
*bdev
, sector_t
*error_sector
)
2599 if (bdev
->bd_disk
== NULL
)
2602 q
= bdev_get_queue(bdev
);
2605 if (!q
->issue_flush_fn
)
2608 return q
->issue_flush_fn(q
, bdev
->bd_disk
, error_sector
);
2611 EXPORT_SYMBOL(blkdev_issue_flush
);
2613 static void drive_stat_acct(struct request
*rq
, int nr_sectors
, int new_io
)
2615 int rw
= rq_data_dir(rq
);
2617 if (!blk_fs_request(rq
) || !rq
->rq_disk
)
2621 __disk_stat_inc(rq
->rq_disk
, merges
[rw
]);
2623 disk_round_stats(rq
->rq_disk
);
2624 rq
->rq_disk
->in_flight
++;
2629 * add-request adds a request to the linked list.
2630 * queue lock is held and interrupts disabled, as we muck with the
2631 * request queue list.
2633 static inline void add_request(request_queue_t
* q
, struct request
* req
)
2635 drive_stat_acct(req
, req
->nr_sectors
, 1);
2638 q
->activity_fn(q
->activity_data
, rq_data_dir(req
));
2641 * elevator indicated where it wants this request to be
2642 * inserted at elevator_merge time
2644 __elv_add_request(q
, req
, ELEVATOR_INSERT_SORT
, 0);
2648 * disk_round_stats() - Round off the performance stats on a struct
2651 * The average IO queue length and utilisation statistics are maintained
2652 * by observing the current state of the queue length and the amount of
2653 * time it has been in this state for.
2655 * Normally, that accounting is done on IO completion, but that can result
2656 * in more than a second's worth of IO being accounted for within any one
2657 * second, leading to >100% utilisation. To deal with that, we call this
2658 * function to do a round-off before returning the results when reading
2659 * /proc/diskstats. This accounts immediately for all queue usage up to
2660 * the current jiffies and restarts the counters again.
2662 void disk_round_stats(struct gendisk
*disk
)
2664 unsigned long now
= jiffies
;
2666 if (now
== disk
->stamp
)
2669 if (disk
->in_flight
) {
2670 __disk_stat_add(disk
, time_in_queue
,
2671 disk
->in_flight
* (now
- disk
->stamp
));
2672 __disk_stat_add(disk
, io_ticks
, (now
- disk
->stamp
));
2677 EXPORT_SYMBOL_GPL(disk_round_stats
);
2680 * queue lock must be held
2682 void __blk_put_request(request_queue_t
*q
, struct request
*req
)
2684 struct request_list
*rl
= req
->rl
;
2688 if (unlikely(--req
->ref_count
))
2691 elv_completed_request(q
, req
);
2693 req
->rq_status
= RQ_INACTIVE
;
2697 * Request may not have originated from ll_rw_blk. if not,
2698 * it didn't come out of our reserved rq pools
2701 int rw
= rq_data_dir(req
);
2702 int priv
= req
->cmd_flags
& REQ_ELVPRIV
;
2704 BUG_ON(!list_empty(&req
->queuelist
));
2705 BUG_ON(!hlist_unhashed(&req
->hash
));
2707 blk_free_request(q
, req
);
2708 freed_request(q
, rw
, priv
);
2712 EXPORT_SYMBOL_GPL(__blk_put_request
);
2714 void blk_put_request(struct request
*req
)
2716 unsigned long flags
;
2717 request_queue_t
*q
= req
->q
;
2720 * Gee, IDE calls in w/ NULL q. Fix IDE and remove the
2721 * following if (q) test.
2724 spin_lock_irqsave(q
->queue_lock
, flags
);
2725 __blk_put_request(q
, req
);
2726 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2730 EXPORT_SYMBOL(blk_put_request
);
2733 * blk_end_sync_rq - executes a completion event on a request
2734 * @rq: request to complete
2735 * @error: end io status of the request
2737 void blk_end_sync_rq(struct request
*rq
, int error
)
2739 struct completion
*waiting
= rq
->waiting
;
2742 __blk_put_request(rq
->q
, rq
);
2745 * complete last, if this is a stack request the process (and thus
2746 * the rq pointer) could be invalid right after this complete()
2750 EXPORT_SYMBOL(blk_end_sync_rq
);
2753 * blk_congestion_wait - wait for a queue to become uncongested
2754 * @rw: READ or WRITE
2755 * @timeout: timeout in jiffies
2757 * Waits for up to @timeout jiffies for a queue (any queue) to exit congestion.
2758 * If no queues are congested then just wait for the next request to be
2761 long blk_congestion_wait(int rw
, long timeout
)
2765 wait_queue_head_t
*wqh
= &congestion_wqh
[rw
];
2767 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
2768 ret
= io_schedule_timeout(timeout
);
2769 finish_wait(wqh
, &wait
);
2773 EXPORT_SYMBOL(blk_congestion_wait
);
2776 * blk_congestion_end - wake up sleepers on a congestion queue
2777 * @rw: READ or WRITE
2779 void blk_congestion_end(int rw
)
2781 wait_queue_head_t
*wqh
= &congestion_wqh
[rw
];
2783 if (waitqueue_active(wqh
))
2788 * Has to be called with the request spinlock acquired
2790 static int attempt_merge(request_queue_t
*q
, struct request
*req
,
2791 struct request
*next
)
2793 if (!rq_mergeable(req
) || !rq_mergeable(next
))
2799 if (req
->sector
+ req
->nr_sectors
!= next
->sector
)
2802 if (rq_data_dir(req
) != rq_data_dir(next
)
2803 || req
->rq_disk
!= next
->rq_disk
2804 || next
->waiting
|| next
->special
)
2808 * If we are allowed to merge, then append bio list
2809 * from next to rq and release next. merge_requests_fn
2810 * will have updated segment counts, update sector
2813 if (!q
->merge_requests_fn(q
, req
, next
))
2817 * At this point we have either done a back merge
2818 * or front merge. We need the smaller start_time of
2819 * the merged requests to be the current request
2820 * for accounting purposes.
2822 if (time_after(req
->start_time
, next
->start_time
))
2823 req
->start_time
= next
->start_time
;
2825 req
->biotail
->bi_next
= next
->bio
;
2826 req
->biotail
= next
->biotail
;
2828 req
->nr_sectors
= req
->hard_nr_sectors
+= next
->hard_nr_sectors
;
2830 elv_merge_requests(q
, req
, next
);
2833 disk_round_stats(req
->rq_disk
);
2834 req
->rq_disk
->in_flight
--;
2837 req
->ioprio
= ioprio_best(req
->ioprio
, next
->ioprio
);
2839 __blk_put_request(q
, next
);
2843 static inline int attempt_back_merge(request_queue_t
*q
, struct request
*rq
)
2845 struct request
*next
= elv_latter_request(q
, rq
);
2848 return attempt_merge(q
, rq
, next
);
2853 static inline int attempt_front_merge(request_queue_t
*q
, struct request
*rq
)
2855 struct request
*prev
= elv_former_request(q
, rq
);
2858 return attempt_merge(q
, prev
, rq
);
2863 static void init_request_from_bio(struct request
*req
, struct bio
*bio
)
2865 req
->cmd_type
= REQ_TYPE_FS
;
2868 * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2870 if (bio_rw_ahead(bio
) || bio_failfast(bio
))
2871 req
->cmd_flags
|= REQ_FAILFAST
;
2874 * REQ_BARRIER implies no merging, but lets make it explicit
2876 if (unlikely(bio_barrier(bio
)))
2877 req
->cmd_flags
|= (REQ_HARDBARRIER
| REQ_NOMERGE
);
2880 req
->cmd_flags
|= REQ_RW_SYNC
;
2883 req
->hard_sector
= req
->sector
= bio
->bi_sector
;
2884 req
->hard_nr_sectors
= req
->nr_sectors
= bio_sectors(bio
);
2885 req
->current_nr_sectors
= req
->hard_cur_sectors
= bio_cur_sectors(bio
);
2886 req
->nr_phys_segments
= bio_phys_segments(req
->q
, bio
);
2887 req
->nr_hw_segments
= bio_hw_segments(req
->q
, bio
);
2888 req
->buffer
= bio_data(bio
); /* see ->buffer comment above */
2889 req
->waiting
= NULL
;
2890 req
->bio
= req
->biotail
= bio
;
2891 req
->ioprio
= bio_prio(bio
);
2892 req
->rq_disk
= bio
->bi_bdev
->bd_disk
;
2893 req
->start_time
= jiffies
;
2896 static int __make_request(request_queue_t
*q
, struct bio
*bio
)
2898 struct request
*req
;
2899 int el_ret
, rw
, nr_sectors
, cur_nr_sectors
, barrier
, err
, sync
;
2900 unsigned short prio
;
2903 sector
= bio
->bi_sector
;
2904 nr_sectors
= bio_sectors(bio
);
2905 cur_nr_sectors
= bio_cur_sectors(bio
);
2906 prio
= bio_prio(bio
);
2908 rw
= bio_data_dir(bio
);
2909 sync
= bio_sync(bio
);
2912 * low level driver can indicate that it wants pages above a
2913 * certain limit bounced to low memory (ie for highmem, or even
2914 * ISA dma in theory)
2916 blk_queue_bounce(q
, &bio
);
2918 spin_lock_prefetch(q
->queue_lock
);
2920 barrier
= bio_barrier(bio
);
2921 if (unlikely(barrier
) && (q
->next_ordered
== QUEUE_ORDERED_NONE
)) {
2926 spin_lock_irq(q
->queue_lock
);
2928 if (unlikely(barrier
) || elv_queue_empty(q
))
2931 el_ret
= elv_merge(q
, &req
, bio
);
2933 case ELEVATOR_BACK_MERGE
:
2934 BUG_ON(!rq_mergeable(req
));
2936 if (!q
->back_merge_fn(q
, req
, bio
))
2939 blk_add_trace_bio(q
, bio
, BLK_TA_BACKMERGE
);
2941 req
->biotail
->bi_next
= bio
;
2943 req
->nr_sectors
= req
->hard_nr_sectors
+= nr_sectors
;
2944 req
->ioprio
= ioprio_best(req
->ioprio
, prio
);
2945 drive_stat_acct(req
, nr_sectors
, 0);
2946 if (!attempt_back_merge(q
, req
))
2947 elv_merged_request(q
, req
, el_ret
);
2950 case ELEVATOR_FRONT_MERGE
:
2951 BUG_ON(!rq_mergeable(req
));
2953 if (!q
->front_merge_fn(q
, req
, bio
))
2956 blk_add_trace_bio(q
, bio
, BLK_TA_FRONTMERGE
);
2958 bio
->bi_next
= req
->bio
;
2962 * may not be valid. if the low level driver said
2963 * it didn't need a bounce buffer then it better
2964 * not touch req->buffer either...
2966 req
->buffer
= bio_data(bio
);
2967 req
->current_nr_sectors
= cur_nr_sectors
;
2968 req
->hard_cur_sectors
= cur_nr_sectors
;
2969 req
->sector
= req
->hard_sector
= sector
;
2970 req
->nr_sectors
= req
->hard_nr_sectors
+= nr_sectors
;
2971 req
->ioprio
= ioprio_best(req
->ioprio
, prio
);
2972 drive_stat_acct(req
, nr_sectors
, 0);
2973 if (!attempt_front_merge(q
, req
))
2974 elv_merged_request(q
, req
, el_ret
);
2977 /* ELV_NO_MERGE: elevator says don't/can't merge. */
2984 * Grab a free request. This is might sleep but can not fail.
2985 * Returns with the queue unlocked.
2987 req
= get_request_wait(q
, rw
, bio
);
2990 * After dropping the lock and possibly sleeping here, our request
2991 * may now be mergeable after it had proven unmergeable (above).
2992 * We don't worry about that case for efficiency. It won't happen
2993 * often, and the elevators are able to handle it.
2995 init_request_from_bio(req
, bio
);
2997 spin_lock_irq(q
->queue_lock
);
2998 if (elv_queue_empty(q
))
3000 add_request(q
, req
);
3003 __generic_unplug_device(q
);
3005 spin_unlock_irq(q
->queue_lock
);
3009 bio_endio(bio
, nr_sectors
<< 9, err
);
3014 * If bio->bi_dev is a partition, remap the location
3016 static inline void blk_partition_remap(struct bio
*bio
)
3018 struct block_device
*bdev
= bio
->bi_bdev
;
3020 if (bdev
!= bdev
->bd_contains
) {
3021 struct hd_struct
*p
= bdev
->bd_part
;
3022 const int rw
= bio_data_dir(bio
);
3024 p
->sectors
[rw
] += bio_sectors(bio
);
3027 bio
->bi_sector
+= p
->start_sect
;
3028 bio
->bi_bdev
= bdev
->bd_contains
;
3032 static void handle_bad_sector(struct bio
*bio
)
3034 char b
[BDEVNAME_SIZE
];
3036 printk(KERN_INFO
"attempt to access beyond end of device\n");
3037 printk(KERN_INFO
"%s: rw=%ld, want=%Lu, limit=%Lu\n",
3038 bdevname(bio
->bi_bdev
, b
),
3040 (unsigned long long)bio
->bi_sector
+ bio_sectors(bio
),
3041 (long long)(bio
->bi_bdev
->bd_inode
->i_size
>> 9));
3043 set_bit(BIO_EOF
, &bio
->bi_flags
);
3047 * generic_make_request: hand a buffer to its device driver for I/O
3048 * @bio: The bio describing the location in memory and on the device.
3050 * generic_make_request() is used to make I/O requests of block
3051 * devices. It is passed a &struct bio, which describes the I/O that needs
3054 * generic_make_request() does not return any status. The
3055 * success/failure status of the request, along with notification of
3056 * completion, is delivered asynchronously through the bio->bi_end_io
3057 * function described (one day) else where.
3059 * The caller of generic_make_request must make sure that bi_io_vec
3060 * are set to describe the memory buffer, and that bi_dev and bi_sector are
3061 * set to describe the device address, and the
3062 * bi_end_io and optionally bi_private are set to describe how
3063 * completion notification should be signaled.
3065 * generic_make_request and the drivers it calls may use bi_next if this
3066 * bio happens to be merged with someone else, and may change bi_dev and
3067 * bi_sector for remaps as it sees fit. So the values of these fields
3068 * should NOT be depended on after the call to generic_make_request.
3070 void generic_make_request(struct bio
*bio
)
3074 int ret
, nr_sectors
= bio_sectors(bio
);
3078 /* Test device or partition size, when known. */
3079 maxsector
= bio
->bi_bdev
->bd_inode
->i_size
>> 9;
3081 sector_t sector
= bio
->bi_sector
;
3083 if (maxsector
< nr_sectors
|| maxsector
- nr_sectors
< sector
) {
3085 * This may well happen - the kernel calls bread()
3086 * without checking the size of the device, e.g., when
3087 * mounting a device.
3089 handle_bad_sector(bio
);
3095 * Resolve the mapping until finished. (drivers are
3096 * still free to implement/resolve their own stacking
3097 * by explicitly returning 0)
3099 * NOTE: we don't repeat the blk_size check for each new device.
3100 * Stacking drivers are expected to know what they are doing.
3105 char b
[BDEVNAME_SIZE
];
3107 q
= bdev_get_queue(bio
->bi_bdev
);
3110 "generic_make_request: Trying to access "
3111 "nonexistent block-device %s (%Lu)\n",
3112 bdevname(bio
->bi_bdev
, b
),
3113 (long long) bio
->bi_sector
);
3115 bio_endio(bio
, bio
->bi_size
, -EIO
);
3119 if (unlikely(bio_sectors(bio
) > q
->max_hw_sectors
)) {
3120 printk("bio too big device %s (%u > %u)\n",
3121 bdevname(bio
->bi_bdev
, b
),
3127 if (unlikely(test_bit(QUEUE_FLAG_DEAD
, &q
->queue_flags
)))
3131 * If this device has partitions, remap block n
3132 * of partition p to block n+start(p) of the disk.
3134 blk_partition_remap(bio
);
3136 if (maxsector
!= -1)
3137 blk_add_trace_remap(q
, bio
, old_dev
, bio
->bi_sector
,
3140 blk_add_trace_bio(q
, bio
, BLK_TA_QUEUE
);
3142 maxsector
= bio
->bi_sector
;
3143 old_dev
= bio
->bi_bdev
->bd_dev
;
3145 ret
= q
->make_request_fn(q
, bio
);
3149 EXPORT_SYMBOL(generic_make_request
);
3152 * submit_bio: submit a bio to the block device layer for I/O
3153 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3154 * @bio: The &struct bio which describes the I/O
3156 * submit_bio() is very similar in purpose to generic_make_request(), and
3157 * uses that function to do most of the work. Both are fairly rough
3158 * interfaces, @bio must be presetup and ready for I/O.
3161 void submit_bio(int rw
, struct bio
*bio
)
3163 int count
= bio_sectors(bio
);
3165 BIO_BUG_ON(!bio
->bi_size
);
3166 BIO_BUG_ON(!bio
->bi_io_vec
);
3169 count_vm_events(PGPGOUT
, count
);
3171 count_vm_events(PGPGIN
, count
);
3173 if (unlikely(block_dump
)) {
3174 char b
[BDEVNAME_SIZE
];
3175 printk(KERN_DEBUG
"%s(%d): %s block %Lu on %s\n",
3176 current
->comm
, current
->pid
,
3177 (rw
& WRITE
) ? "WRITE" : "READ",
3178 (unsigned long long)bio
->bi_sector
,
3179 bdevname(bio
->bi_bdev
,b
));
3182 generic_make_request(bio
);
3185 EXPORT_SYMBOL(submit_bio
);
3187 static void blk_recalc_rq_segments(struct request
*rq
)
3189 struct bio
*bio
, *prevbio
= NULL
;
3190 int nr_phys_segs
, nr_hw_segs
;
3191 unsigned int phys_size
, hw_size
;
3192 request_queue_t
*q
= rq
->q
;
3197 phys_size
= hw_size
= nr_phys_segs
= nr_hw_segs
= 0;
3198 rq_for_each_bio(bio
, rq
) {
3199 /* Force bio hw/phys segs to be recalculated. */
3200 bio
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3202 nr_phys_segs
+= bio_phys_segments(q
, bio
);
3203 nr_hw_segs
+= bio_hw_segments(q
, bio
);
3205 int pseg
= phys_size
+ prevbio
->bi_size
+ bio
->bi_size
;
3206 int hseg
= hw_size
+ prevbio
->bi_size
+ bio
->bi_size
;
3208 if (blk_phys_contig_segment(q
, prevbio
, bio
) &&
3209 pseg
<= q
->max_segment_size
) {
3211 phys_size
+= prevbio
->bi_size
+ bio
->bi_size
;
3215 if (blk_hw_contig_segment(q
, prevbio
, bio
) &&
3216 hseg
<= q
->max_segment_size
) {
3218 hw_size
+= prevbio
->bi_size
+ bio
->bi_size
;
3225 rq
->nr_phys_segments
= nr_phys_segs
;
3226 rq
->nr_hw_segments
= nr_hw_segs
;
3229 static void blk_recalc_rq_sectors(struct request
*rq
, int nsect
)
3231 if (blk_fs_request(rq
)) {
3232 rq
->hard_sector
+= nsect
;
3233 rq
->hard_nr_sectors
-= nsect
;
3236 * Move the I/O submission pointers ahead if required.
3238 if ((rq
->nr_sectors
>= rq
->hard_nr_sectors
) &&
3239 (rq
->sector
<= rq
->hard_sector
)) {
3240 rq
->sector
= rq
->hard_sector
;
3241 rq
->nr_sectors
= rq
->hard_nr_sectors
;
3242 rq
->hard_cur_sectors
= bio_cur_sectors(rq
->bio
);
3243 rq
->current_nr_sectors
= rq
->hard_cur_sectors
;
3244 rq
->buffer
= bio_data(rq
->bio
);
3248 * if total number of sectors is less than the first segment
3249 * size, something has gone terribly wrong
3251 if (rq
->nr_sectors
< rq
->current_nr_sectors
) {
3252 printk("blk: request botched\n");
3253 rq
->nr_sectors
= rq
->current_nr_sectors
;
3258 static int __end_that_request_first(struct request
*req
, int uptodate
,
3261 int total_bytes
, bio_nbytes
, error
, next_idx
= 0;
3264 blk_add_trace_rq(req
->q
, req
, BLK_TA_COMPLETE
);
3267 * extend uptodate bool to allow < 0 value to be direct io error
3270 if (end_io_error(uptodate
))
3271 error
= !uptodate
? -EIO
: uptodate
;
3274 * for a REQ_BLOCK_PC request, we want to carry any eventual
3275 * sense key with us all the way through
3277 if (!blk_pc_request(req
))
3281 if (blk_fs_request(req
) && !(req
->cmd_flags
& REQ_QUIET
))
3282 printk("end_request: I/O error, dev %s, sector %llu\n",
3283 req
->rq_disk
? req
->rq_disk
->disk_name
: "?",
3284 (unsigned long long)req
->sector
);
3287 if (blk_fs_request(req
) && req
->rq_disk
) {
3288 const int rw
= rq_data_dir(req
);
3290 disk_stat_add(req
->rq_disk
, sectors
[rw
], nr_bytes
>> 9);
3293 total_bytes
= bio_nbytes
= 0;
3294 while ((bio
= req
->bio
) != NULL
) {
3297 if (nr_bytes
>= bio
->bi_size
) {
3298 req
->bio
= bio
->bi_next
;
3299 nbytes
= bio
->bi_size
;
3300 if (!ordered_bio_endio(req
, bio
, nbytes
, error
))
3301 bio_endio(bio
, nbytes
, error
);
3305 int idx
= bio
->bi_idx
+ next_idx
;
3307 if (unlikely(bio
->bi_idx
>= bio
->bi_vcnt
)) {
3308 blk_dump_rq_flags(req
, "__end_that");
3309 printk("%s: bio idx %d >= vcnt %d\n",
3311 bio
->bi_idx
, bio
->bi_vcnt
);
3315 nbytes
= bio_iovec_idx(bio
, idx
)->bv_len
;
3316 BIO_BUG_ON(nbytes
> bio
->bi_size
);
3319 * not a complete bvec done
3321 if (unlikely(nbytes
> nr_bytes
)) {
3322 bio_nbytes
+= nr_bytes
;
3323 total_bytes
+= nr_bytes
;
3328 * advance to the next vector
3331 bio_nbytes
+= nbytes
;
3334 total_bytes
+= nbytes
;
3337 if ((bio
= req
->bio
)) {
3339 * end more in this run, or just return 'not-done'
3341 if (unlikely(nr_bytes
<= 0))
3353 * if the request wasn't completed, update state
3356 if (!ordered_bio_endio(req
, bio
, bio_nbytes
, error
))
3357 bio_endio(bio
, bio_nbytes
, error
);
3358 bio
->bi_idx
+= next_idx
;
3359 bio_iovec(bio
)->bv_offset
+= nr_bytes
;
3360 bio_iovec(bio
)->bv_len
-= nr_bytes
;
3363 blk_recalc_rq_sectors(req
, total_bytes
>> 9);
3364 blk_recalc_rq_segments(req
);
3369 * end_that_request_first - end I/O on a request
3370 * @req: the request being processed
3371 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3372 * @nr_sectors: number of sectors to end I/O on
3375 * Ends I/O on a number of sectors attached to @req, and sets it up
3376 * for the next range of segments (if any) in the cluster.
3379 * 0 - we are done with this request, call end_that_request_last()
3380 * 1 - still buffers pending for this request
3382 int end_that_request_first(struct request
*req
, int uptodate
, int nr_sectors
)
3384 return __end_that_request_first(req
, uptodate
, nr_sectors
<< 9);
3387 EXPORT_SYMBOL(end_that_request_first
);
3390 * end_that_request_chunk - end I/O on a request
3391 * @req: the request being processed
3392 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3393 * @nr_bytes: number of bytes to complete
3396 * Ends I/O on a number of bytes attached to @req, and sets it up
3397 * for the next range of segments (if any). Like end_that_request_first(),
3398 * but deals with bytes instead of sectors.
3401 * 0 - we are done with this request, call end_that_request_last()
3402 * 1 - still buffers pending for this request
3404 int end_that_request_chunk(struct request
*req
, int uptodate
, int nr_bytes
)
3406 return __end_that_request_first(req
, uptodate
, nr_bytes
);
3409 EXPORT_SYMBOL(end_that_request_chunk
);
3412 * splice the completion data to a local structure and hand off to
3413 * process_completion_queue() to complete the requests
3415 static void blk_done_softirq(struct softirq_action
*h
)
3417 struct list_head
*cpu_list
, local_list
;
3419 local_irq_disable();
3420 cpu_list
= &__get_cpu_var(blk_cpu_done
);
3421 list_replace_init(cpu_list
, &local_list
);
3424 while (!list_empty(&local_list
)) {
3425 struct request
*rq
= list_entry(local_list
.next
, struct request
, donelist
);
3427 list_del_init(&rq
->donelist
);
3428 rq
->q
->softirq_done_fn(rq
);
3432 #ifdef CONFIG_HOTPLUG_CPU
3434 static int blk_cpu_notify(struct notifier_block
*self
, unsigned long action
,
3438 * If a CPU goes away, splice its entries to the current CPU
3439 * and trigger a run of the softirq
3441 if (action
== CPU_DEAD
) {
3442 int cpu
= (unsigned long) hcpu
;
3444 local_irq_disable();
3445 list_splice_init(&per_cpu(blk_cpu_done
, cpu
),
3446 &__get_cpu_var(blk_cpu_done
));
3447 raise_softirq_irqoff(BLOCK_SOFTIRQ
);
3455 static struct notifier_block __devinitdata blk_cpu_notifier
= {
3456 .notifier_call
= blk_cpu_notify
,
3459 #endif /* CONFIG_HOTPLUG_CPU */
3462 * blk_complete_request - end I/O on a request
3463 * @req: the request being processed
3466 * Ends all I/O on a request. It does not handle partial completions,
3467 * unless the driver actually implements this in its completion callback
3468 * through requeueing. Theh actual completion happens out-of-order,
3469 * through a softirq handler. The user must have registered a completion
3470 * callback through blk_queue_softirq_done().
3473 void blk_complete_request(struct request
*req
)
3475 struct list_head
*cpu_list
;
3476 unsigned long flags
;
3478 BUG_ON(!req
->q
->softirq_done_fn
);
3480 local_irq_save(flags
);
3482 cpu_list
= &__get_cpu_var(blk_cpu_done
);
3483 list_add_tail(&req
->donelist
, cpu_list
);
3484 raise_softirq_irqoff(BLOCK_SOFTIRQ
);
3486 local_irq_restore(flags
);
3489 EXPORT_SYMBOL(blk_complete_request
);
3492 * queue lock must be held
3494 void end_that_request_last(struct request
*req
, int uptodate
)
3496 struct gendisk
*disk
= req
->rq_disk
;
3500 * extend uptodate bool to allow < 0 value to be direct io error
3503 if (end_io_error(uptodate
))
3504 error
= !uptodate
? -EIO
: uptodate
;
3506 if (unlikely(laptop_mode
) && blk_fs_request(req
))
3507 laptop_io_completion();
3510 * Account IO completion. bar_rq isn't accounted as a normal
3511 * IO on queueing nor completion. Accounting the containing
3512 * request is enough.
3514 if (disk
&& blk_fs_request(req
) && req
!= &req
->q
->bar_rq
) {
3515 unsigned long duration
= jiffies
- req
->start_time
;
3516 const int rw
= rq_data_dir(req
);
3518 __disk_stat_inc(disk
, ios
[rw
]);
3519 __disk_stat_add(disk
, ticks
[rw
], duration
);
3520 disk_round_stats(disk
);
3524 req
->end_io(req
, error
);
3526 __blk_put_request(req
->q
, req
);
3529 EXPORT_SYMBOL(end_that_request_last
);
3531 void end_request(struct request
*req
, int uptodate
)
3533 if (!end_that_request_first(req
, uptodate
, req
->hard_cur_sectors
)) {
3534 add_disk_randomness(req
->rq_disk
);
3535 blkdev_dequeue_request(req
);
3536 end_that_request_last(req
, uptodate
);
3540 EXPORT_SYMBOL(end_request
);
3542 void blk_rq_bio_prep(request_queue_t
*q
, struct request
*rq
, struct bio
*bio
)
3544 /* first two bits are identical in rq->cmd_flags and bio->bi_rw */
3545 rq
->cmd_flags
|= (bio
->bi_rw
& 3);
3547 rq
->nr_phys_segments
= bio_phys_segments(q
, bio
);
3548 rq
->nr_hw_segments
= bio_hw_segments(q
, bio
);
3549 rq
->current_nr_sectors
= bio_cur_sectors(bio
);
3550 rq
->hard_cur_sectors
= rq
->current_nr_sectors
;
3551 rq
->hard_nr_sectors
= rq
->nr_sectors
= bio_sectors(bio
);
3552 rq
->buffer
= bio_data(bio
);
3554 rq
->bio
= rq
->biotail
= bio
;
3557 EXPORT_SYMBOL(blk_rq_bio_prep
);
3559 int kblockd_schedule_work(struct work_struct
*work
)
3561 return queue_work(kblockd_workqueue
, work
);
3564 EXPORT_SYMBOL(kblockd_schedule_work
);
3566 void kblockd_flush(void)
3568 flush_workqueue(kblockd_workqueue
);
3570 EXPORT_SYMBOL(kblockd_flush
);
3572 int __init
blk_dev_init(void)
3576 kblockd_workqueue
= create_workqueue("kblockd");
3577 if (!kblockd_workqueue
)
3578 panic("Failed to create kblockd\n");
3580 request_cachep
= kmem_cache_create("blkdev_requests",
3581 sizeof(struct request
), 0, SLAB_PANIC
, NULL
, NULL
);
3583 requestq_cachep
= kmem_cache_create("blkdev_queue",
3584 sizeof(request_queue_t
), 0, SLAB_PANIC
, NULL
, NULL
);
3586 iocontext_cachep
= kmem_cache_create("blkdev_ioc",
3587 sizeof(struct io_context
), 0, SLAB_PANIC
, NULL
, NULL
);
3589 for_each_possible_cpu(i
)
3590 INIT_LIST_HEAD(&per_cpu(blk_cpu_done
, i
));
3592 open_softirq(BLOCK_SOFTIRQ
, blk_done_softirq
, NULL
);
3593 register_hotcpu_notifier(&blk_cpu_notifier
);
3595 blk_max_low_pfn
= max_low_pfn
;
3596 blk_max_pfn
= max_pfn
;
3602 * IO Context helper functions
3604 void put_io_context(struct io_context
*ioc
)
3609 BUG_ON(atomic_read(&ioc
->refcount
) == 0);
3611 if (atomic_dec_and_test(&ioc
->refcount
)) {
3612 struct cfq_io_context
*cic
;
3615 if (ioc
->aic
&& ioc
->aic
->dtor
)
3616 ioc
->aic
->dtor(ioc
->aic
);
3617 if (ioc
->cic_root
.rb_node
!= NULL
) {
3618 struct rb_node
*n
= rb_first(&ioc
->cic_root
);
3620 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
3625 kmem_cache_free(iocontext_cachep
, ioc
);
3628 EXPORT_SYMBOL(put_io_context
);
3630 /* Called by the exitting task */
3631 void exit_io_context(void)
3633 unsigned long flags
;
3634 struct io_context
*ioc
;
3635 struct cfq_io_context
*cic
;
3637 local_irq_save(flags
);
3639 ioc
= current
->io_context
;
3640 current
->io_context
= NULL
;
3642 task_unlock(current
);
3643 local_irq_restore(flags
);
3645 if (ioc
->aic
&& ioc
->aic
->exit
)
3646 ioc
->aic
->exit(ioc
->aic
);
3647 if (ioc
->cic_root
.rb_node
!= NULL
) {
3648 cic
= rb_entry(rb_first(&ioc
->cic_root
), struct cfq_io_context
, rb_node
);
3652 put_io_context(ioc
);
3656 * If the current task has no IO context then create one and initialise it.
3657 * Otherwise, return its existing IO context.
3659 * This returned IO context doesn't have a specifically elevated refcount,
3660 * but since the current task itself holds a reference, the context can be
3661 * used in general code, so long as it stays within `current` context.
3663 struct io_context
*current_io_context(gfp_t gfp_flags
)
3665 struct task_struct
*tsk
= current
;
3666 struct io_context
*ret
;
3668 ret
= tsk
->io_context
;
3672 ret
= kmem_cache_alloc(iocontext_cachep
, gfp_flags
);
3674 atomic_set(&ret
->refcount
, 1);
3675 ret
->task
= current
;
3676 ret
->set_ioprio
= NULL
;
3677 ret
->last_waited
= jiffies
; /* doesn't matter... */
3678 ret
->nr_batch_requests
= 0; /* because this is 0 */
3680 ret
->cic_root
.rb_node
= NULL
;
3681 /* make sure set_task_ioprio() sees the settings above */
3683 tsk
->io_context
= ret
;
3688 EXPORT_SYMBOL(current_io_context
);
3691 * If the current task has no IO context then create one and initialise it.
3692 * If it does have a context, take a ref on it.
3694 * This is always called in the context of the task which submitted the I/O.
3696 struct io_context
*get_io_context(gfp_t gfp_flags
)
3698 struct io_context
*ret
;
3699 ret
= current_io_context(gfp_flags
);
3701 atomic_inc(&ret
->refcount
);
3704 EXPORT_SYMBOL(get_io_context
);
3706 void copy_io_context(struct io_context
**pdst
, struct io_context
**psrc
)
3708 struct io_context
*src
= *psrc
;
3709 struct io_context
*dst
= *pdst
;
3712 BUG_ON(atomic_read(&src
->refcount
) == 0);
3713 atomic_inc(&src
->refcount
);
3714 put_io_context(dst
);
3718 EXPORT_SYMBOL(copy_io_context
);
3720 void swap_io_context(struct io_context
**ioc1
, struct io_context
**ioc2
)
3722 struct io_context
*temp
;
3727 EXPORT_SYMBOL(swap_io_context
);
3732 struct queue_sysfs_entry
{
3733 struct attribute attr
;
3734 ssize_t (*show
)(struct request_queue
*, char *);
3735 ssize_t (*store
)(struct request_queue
*, const char *, size_t);
3739 queue_var_show(unsigned int var
, char *page
)
3741 return sprintf(page
, "%d\n", var
);
3745 queue_var_store(unsigned long *var
, const char *page
, size_t count
)
3747 char *p
= (char *) page
;
3749 *var
= simple_strtoul(p
, &p
, 10);
3753 static ssize_t
queue_requests_show(struct request_queue
*q
, char *page
)
3755 return queue_var_show(q
->nr_requests
, (page
));
3759 queue_requests_store(struct request_queue
*q
, const char *page
, size_t count
)
3761 struct request_list
*rl
= &q
->rq
;
3763 int ret
= queue_var_store(&nr
, page
, count
);
3764 if (nr
< BLKDEV_MIN_RQ
)
3767 spin_lock_irq(q
->queue_lock
);
3768 q
->nr_requests
= nr
;
3769 blk_queue_congestion_threshold(q
);
3771 if (rl
->count
[READ
] >= queue_congestion_on_threshold(q
))
3772 set_queue_congested(q
, READ
);
3773 else if (rl
->count
[READ
] < queue_congestion_off_threshold(q
))
3774 clear_queue_congested(q
, READ
);
3776 if (rl
->count
[WRITE
] >= queue_congestion_on_threshold(q
))
3777 set_queue_congested(q
, WRITE
);
3778 else if (rl
->count
[WRITE
] < queue_congestion_off_threshold(q
))
3779 clear_queue_congested(q
, WRITE
);
3781 if (rl
->count
[READ
] >= q
->nr_requests
) {
3782 blk_set_queue_full(q
, READ
);
3783 } else if (rl
->count
[READ
]+1 <= q
->nr_requests
) {
3784 blk_clear_queue_full(q
, READ
);
3785 wake_up(&rl
->wait
[READ
]);
3788 if (rl
->count
[WRITE
] >= q
->nr_requests
) {
3789 blk_set_queue_full(q
, WRITE
);
3790 } else if (rl
->count
[WRITE
]+1 <= q
->nr_requests
) {
3791 blk_clear_queue_full(q
, WRITE
);
3792 wake_up(&rl
->wait
[WRITE
]);
3794 spin_unlock_irq(q
->queue_lock
);
3798 static ssize_t
queue_ra_show(struct request_queue
*q
, char *page
)
3800 int ra_kb
= q
->backing_dev_info
.ra_pages
<< (PAGE_CACHE_SHIFT
- 10);
3802 return queue_var_show(ra_kb
, (page
));
3806 queue_ra_store(struct request_queue
*q
, const char *page
, size_t count
)
3808 unsigned long ra_kb
;
3809 ssize_t ret
= queue_var_store(&ra_kb
, page
, count
);
3811 spin_lock_irq(q
->queue_lock
);
3812 if (ra_kb
> (q
->max_sectors
>> 1))
3813 ra_kb
= (q
->max_sectors
>> 1);
3815 q
->backing_dev_info
.ra_pages
= ra_kb
>> (PAGE_CACHE_SHIFT
- 10);
3816 spin_unlock_irq(q
->queue_lock
);
3821 static ssize_t
queue_max_sectors_show(struct request_queue
*q
, char *page
)
3823 int max_sectors_kb
= q
->max_sectors
>> 1;
3825 return queue_var_show(max_sectors_kb
, (page
));
3829 queue_max_sectors_store(struct request_queue
*q
, const char *page
, size_t count
)
3831 unsigned long max_sectors_kb
,
3832 max_hw_sectors_kb
= q
->max_hw_sectors
>> 1,
3833 page_kb
= 1 << (PAGE_CACHE_SHIFT
- 10);
3834 ssize_t ret
= queue_var_store(&max_sectors_kb
, page
, count
);
3837 if (max_sectors_kb
> max_hw_sectors_kb
|| max_sectors_kb
< page_kb
)
3840 * Take the queue lock to update the readahead and max_sectors
3841 * values synchronously:
3843 spin_lock_irq(q
->queue_lock
);
3845 * Trim readahead window as well, if necessary:
3847 ra_kb
= q
->backing_dev_info
.ra_pages
<< (PAGE_CACHE_SHIFT
- 10);
3848 if (ra_kb
> max_sectors_kb
)
3849 q
->backing_dev_info
.ra_pages
=
3850 max_sectors_kb
>> (PAGE_CACHE_SHIFT
- 10);
3852 q
->max_sectors
= max_sectors_kb
<< 1;
3853 spin_unlock_irq(q
->queue_lock
);
3858 static ssize_t
queue_max_hw_sectors_show(struct request_queue
*q
, char *page
)
3860 int max_hw_sectors_kb
= q
->max_hw_sectors
>> 1;
3862 return queue_var_show(max_hw_sectors_kb
, (page
));
3866 static struct queue_sysfs_entry queue_requests_entry
= {
3867 .attr
= {.name
= "nr_requests", .mode
= S_IRUGO
| S_IWUSR
},
3868 .show
= queue_requests_show
,
3869 .store
= queue_requests_store
,
3872 static struct queue_sysfs_entry queue_ra_entry
= {
3873 .attr
= {.name
= "read_ahead_kb", .mode
= S_IRUGO
| S_IWUSR
},
3874 .show
= queue_ra_show
,
3875 .store
= queue_ra_store
,
3878 static struct queue_sysfs_entry queue_max_sectors_entry
= {
3879 .attr
= {.name
= "max_sectors_kb", .mode
= S_IRUGO
| S_IWUSR
},
3880 .show
= queue_max_sectors_show
,
3881 .store
= queue_max_sectors_store
,
3884 static struct queue_sysfs_entry queue_max_hw_sectors_entry
= {
3885 .attr
= {.name
= "max_hw_sectors_kb", .mode
= S_IRUGO
},
3886 .show
= queue_max_hw_sectors_show
,
3889 static struct queue_sysfs_entry queue_iosched_entry
= {
3890 .attr
= {.name
= "scheduler", .mode
= S_IRUGO
| S_IWUSR
},
3891 .show
= elv_iosched_show
,
3892 .store
= elv_iosched_store
,
3895 static struct attribute
*default_attrs
[] = {
3896 &queue_requests_entry
.attr
,
3897 &queue_ra_entry
.attr
,
3898 &queue_max_hw_sectors_entry
.attr
,
3899 &queue_max_sectors_entry
.attr
,
3900 &queue_iosched_entry
.attr
,
3904 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
3907 queue_attr_show(struct kobject
*kobj
, struct attribute
*attr
, char *page
)
3909 struct queue_sysfs_entry
*entry
= to_queue(attr
);
3910 request_queue_t
*q
= container_of(kobj
, struct request_queue
, kobj
);
3915 mutex_lock(&q
->sysfs_lock
);
3916 if (test_bit(QUEUE_FLAG_DEAD
, &q
->queue_flags
)) {
3917 mutex_unlock(&q
->sysfs_lock
);
3920 res
= entry
->show(q
, page
);
3921 mutex_unlock(&q
->sysfs_lock
);
3926 queue_attr_store(struct kobject
*kobj
, struct attribute
*attr
,
3927 const char *page
, size_t length
)
3929 struct queue_sysfs_entry
*entry
= to_queue(attr
);
3930 request_queue_t
*q
= container_of(kobj
, struct request_queue
, kobj
);
3936 mutex_lock(&q
->sysfs_lock
);
3937 if (test_bit(QUEUE_FLAG_DEAD
, &q
->queue_flags
)) {
3938 mutex_unlock(&q
->sysfs_lock
);
3941 res
= entry
->store(q
, page
, length
);
3942 mutex_unlock(&q
->sysfs_lock
);
3946 static struct sysfs_ops queue_sysfs_ops
= {
3947 .show
= queue_attr_show
,
3948 .store
= queue_attr_store
,
3951 static struct kobj_type queue_ktype
= {
3952 .sysfs_ops
= &queue_sysfs_ops
,
3953 .default_attrs
= default_attrs
,
3954 .release
= blk_release_queue
,
3957 int blk_register_queue(struct gendisk
*disk
)
3961 request_queue_t
*q
= disk
->queue
;
3963 if (!q
|| !q
->request_fn
)
3966 q
->kobj
.parent
= kobject_get(&disk
->kobj
);
3968 ret
= kobject_add(&q
->kobj
);
3972 kobject_uevent(&q
->kobj
, KOBJ_ADD
);
3974 ret
= elv_register_queue(q
);
3976 kobject_uevent(&q
->kobj
, KOBJ_REMOVE
);
3977 kobject_del(&q
->kobj
);
3984 void blk_unregister_queue(struct gendisk
*disk
)
3986 request_queue_t
*q
= disk
->queue
;
3988 if (q
&& q
->request_fn
) {
3989 elv_unregister_queue(q
);
3991 kobject_uevent(&q
->kobj
, KOBJ_REMOVE
);
3992 kobject_del(&q
->kobj
);
3993 kobject_put(&disk
->kobj
);