aboutsummaryrefslogtreecommitdiffstats
path: root/block/blk-core.c
diff options
context:
space:
mode:
authorJens Axboe <jens.axboe@oracle.com>2008-01-29 08:49:21 -0500
committerJens Axboe <jens.axboe@oracle.com>2008-01-29 15:55:05 -0500
commita168ee84c90b39ece357da127ab388f2f64db19c (patch)
treeacb5527d9835c06e4eb2c308850e74db857f7d64 /block/blk-core.c
parent9bf722598fcd51073974850ae026b44389430ecc (diff)
block: first step of splitting ll_rw_blk, rename it
Then we retain history in blk-core.c Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Diffstat (limited to 'block/blk-core.c')
-rw-r--r--block/blk-core.c4457
1 files changed, 4457 insertions, 0 deletions
diff --git a/block/blk-core.c b/block/blk-core.c
new file mode 100644
index 000000000000..1932a56f5e4b
--- /dev/null
+++ b/block/blk-core.c
@@ -0,0 +1,4457 @@
1/*
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
8 */
9
10/*
11 * This handles all read/write requests to block devices
12 */
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>
19#include <linux/mm.h>
20#include <linux/kernel_stat.h>
21#include <linux/string.h>
22#include <linux/init.h>
23#include <linux/bootmem.h> /* for max_pfn/max_low_pfn */
24#include <linux/completion.h>
25#include <linux/slab.h>
26#include <linux/swap.h>
27#include <linux/writeback.h>
28#include <linux/task_io_accounting_ops.h>
29#include <linux/interrupt.h>
30#include <linux/cpu.h>
31#include <linux/blktrace_api.h>
32#include <linux/fault-inject.h>
33#include <linux/scatterlist.h>
34
35/*
36 * for max sense size
37 */
38#include <scsi/scsi_cmnd.h>
39
40static void blk_unplug_work(struct work_struct *work);
41static void blk_unplug_timeout(unsigned long data);
42static void drive_stat_acct(struct request *rq, int new_io);
43static void init_request_from_bio(struct request *req, struct bio *bio);
44static int __make_request(struct request_queue *q, struct bio *bio);
45static struct io_context *current_io_context(gfp_t gfp_flags, int node);
46static void blk_recalc_rq_segments(struct request *rq);
47static void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
48 struct bio *bio);
49
50/*
51 * For the allocated request tables
52 */
53static struct kmem_cache *request_cachep;
54
55/*
56 * For queue allocation
57 */
58static struct kmem_cache *requestq_cachep;
59
60/*
61 * For io context allocations
62 */
63static struct kmem_cache *iocontext_cachep;
64
65/*
66 * Controlling structure to kblockd
67 */
68static struct workqueue_struct *kblockd_workqueue;
69
70unsigned long blk_max_low_pfn, blk_max_pfn;
71
72EXPORT_SYMBOL(blk_max_low_pfn);
73EXPORT_SYMBOL(blk_max_pfn);
74
75static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
76
77/* Amount of time in which a process may batch requests */
78#define BLK_BATCH_TIME (HZ/50UL)
79
80/* Number of requests a "batching" process may submit */
81#define BLK_BATCH_REQ 32
82
83/*
84 * Return the threshold (number of used requests) at which the queue is
85 * considered to be congested. It include a little hysteresis to keep the
86 * context switch rate down.
87 */
88static inline int queue_congestion_on_threshold(struct request_queue *q)
89{
90 return q->nr_congestion_on;
91}
92
93/*
94 * The threshold at which a queue is considered to be uncongested
95 */
96static inline int queue_congestion_off_threshold(struct request_queue *q)
97{
98 return q->nr_congestion_off;
99}
100
101static void blk_queue_congestion_threshold(struct request_queue *q)
102{
103 int nr;
104
105 nr = q->nr_requests - (q->nr_requests / 8) + 1;
106 if (nr > q->nr_requests)
107 nr = q->nr_requests;
108 q->nr_congestion_on = nr;
109
110 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
111 if (nr < 1)
112 nr = 1;
113 q->nr_congestion_off = nr;
114}
115
116/**
117 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
118 * @bdev: device
119 *
120 * Locates the passed device's request queue and returns the address of its
121 * backing_dev_info
122 *
123 * Will return NULL if the request queue cannot be located.
124 */
125struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
126{
127 struct backing_dev_info *ret = NULL;
128 struct request_queue *q = bdev_get_queue(bdev);
129
130 if (q)
131 ret = &q->backing_dev_info;
132 return ret;
133}
134EXPORT_SYMBOL(blk_get_backing_dev_info);
135
136/**
137 * blk_queue_prep_rq - set a prepare_request function for queue
138 * @q: queue
139 * @pfn: prepare_request function
140 *
141 * It's possible for a queue to register a prepare_request callback which
142 * is invoked before the request is handed to the request_fn. The goal of
143 * the function is to prepare a request for I/O, it can be used to build a
144 * cdb from the request data for instance.
145 *
146 */
147void blk_queue_prep_rq(struct request_queue *q, prep_rq_fn *pfn)
148{
149 q->prep_rq_fn = pfn;
150}
151
152EXPORT_SYMBOL(blk_queue_prep_rq);
153
154/**
155 * blk_queue_merge_bvec - set a merge_bvec function for queue
156 * @q: queue
157 * @mbfn: merge_bvec_fn
158 *
159 * Usually queues have static limitations on the max sectors or segments that
160 * we can put in a request. Stacking drivers may have some settings that
161 * are dynamic, and thus we have to query the queue whether it is ok to
162 * add a new bio_vec to a bio at a given offset or not. If the block device
163 * has such limitations, it needs to register a merge_bvec_fn to control
164 * the size of bio's sent to it. Note that a block device *must* allow a
165 * single page to be added to an empty bio. The block device driver may want
166 * to use the bio_split() function to deal with these bio's. By default
167 * no merge_bvec_fn is defined for a queue, and only the fixed limits are
168 * honored.
169 */
170void blk_queue_merge_bvec(struct request_queue *q, merge_bvec_fn *mbfn)
171{
172 q->merge_bvec_fn = mbfn;
173}
174
175EXPORT_SYMBOL(blk_queue_merge_bvec);
176
177void blk_queue_softirq_done(struct request_queue *q, softirq_done_fn *fn)
178{
179 q->softirq_done_fn = fn;
180}
181
182EXPORT_SYMBOL(blk_queue_softirq_done);
183
184/**
185 * blk_queue_make_request - define an alternate make_request function for a device
186 * @q: the request queue for the device to be affected
187 * @mfn: the alternate make_request function
188 *
189 * Description:
190 * The normal way for &struct bios to be passed to a device
191 * driver is for them to be collected into requests on a request
192 * queue, and then to allow the device driver to select requests
193 * off that queue when it is ready. This works well for many block
194 * devices. However some block devices (typically virtual devices
195 * such as md or lvm) do not benefit from the processing on the
196 * request queue, and are served best by having the requests passed
197 * directly to them. This can be achieved by providing a function
198 * to blk_queue_make_request().
199 *
200 * Caveat:
201 * The driver that does this *must* be able to deal appropriately
202 * with buffers in "highmemory". This can be accomplished by either calling
203 * __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
204 * blk_queue_bounce() to create a buffer in normal memory.
205 **/
206void blk_queue_make_request(struct request_queue * q, make_request_fn * mfn)
207{
208 /*
209 * set defaults
210 */
211 q->nr_requests = BLKDEV_MAX_RQ;
212 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
213 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
214 q->make_request_fn = mfn;
215 q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
216 q->backing_dev_info.state = 0;
217 q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
218 blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
219 blk_queue_hardsect_size(q, 512);
220 blk_queue_dma_alignment(q, 511);
221 blk_queue_congestion_threshold(q);
222 q->nr_batching = BLK_BATCH_REQ;
223
224 q->unplug_thresh = 4; /* hmm */
225 q->unplug_delay = (3 * HZ) / 1000; /* 3 milliseconds */
226 if (q->unplug_delay == 0)
227 q->unplug_delay = 1;
228
229 INIT_WORK(&q->unplug_work, blk_unplug_work);
230
231 q->unplug_timer.function = blk_unplug_timeout;
232 q->unplug_timer.data = (unsigned long)q;
233
234 /*
235 * by default assume old behaviour and bounce for any highmem page
236 */
237 blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
238}
239
240EXPORT_SYMBOL(blk_queue_make_request);
241
242static void rq_init(struct request_queue *q, struct request *rq)
243{
244 INIT_LIST_HEAD(&rq->queuelist);
245 INIT_LIST_HEAD(&rq->donelist);
246
247 rq->errors = 0;
248 rq->bio = rq->biotail = NULL;
249 INIT_HLIST_NODE(&rq->hash);
250 RB_CLEAR_NODE(&rq->rb_node);
251 rq->ioprio = 0;
252 rq->buffer = NULL;
253 rq->ref_count = 1;
254 rq->q = q;
255 rq->special = NULL;
256 rq->data_len = 0;
257 rq->data = NULL;
258 rq->nr_phys_segments = 0;
259 rq->sense = NULL;
260 rq->end_io = NULL;
261 rq->end_io_data = NULL;
262 rq->completion_data = NULL;
263 rq->next_rq = NULL;
264}
265
266/**
267 * blk_queue_ordered - does this queue support ordered writes
268 * @q: the request queue
269 * @ordered: one of QUEUE_ORDERED_*
270 * @prepare_flush_fn: rq setup helper for cache flush ordered writes
271 *
272 * Description:
273 * For journalled file systems, doing ordered writes on a commit
274 * block instead of explicitly doing wait_on_buffer (which is bad
275 * for performance) can be a big win. Block drivers supporting this
276 * feature should call this function and indicate so.
277 *
278 **/
279int blk_queue_ordered(struct request_queue *q, unsigned ordered,
280 prepare_flush_fn *prepare_flush_fn)
281{
282 if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
283 prepare_flush_fn == NULL) {
284 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
285 return -EINVAL;
286 }
287
288 if (ordered != QUEUE_ORDERED_NONE &&
289 ordered != QUEUE_ORDERED_DRAIN &&
290 ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
291 ordered != QUEUE_ORDERED_DRAIN_FUA &&
292 ordered != QUEUE_ORDERED_TAG &&
293 ordered != QUEUE_ORDERED_TAG_FLUSH &&
294 ordered != QUEUE_ORDERED_TAG_FUA) {
295 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
296 return -EINVAL;
297 }
298
299 q->ordered = ordered;
300 q->next_ordered = ordered;
301 q->prepare_flush_fn = prepare_flush_fn;
302
303 return 0;
304}
305
306EXPORT_SYMBOL(blk_queue_ordered);
307
308/*
309 * Cache flushing for ordered writes handling
310 */
311inline unsigned blk_ordered_cur_seq(struct request_queue *q)
312{
313 if (!q->ordseq)
314 return 0;
315 return 1 << ffz(q->ordseq);
316}
317
318unsigned blk_ordered_req_seq(struct request *rq)
319{
320 struct request_queue *q = rq->q;
321
322 BUG_ON(q->ordseq == 0);
323
324 if (rq == &q->pre_flush_rq)
325 return QUEUE_ORDSEQ_PREFLUSH;
326 if (rq == &q->bar_rq)
327 return QUEUE_ORDSEQ_BAR;
328 if (rq == &q->post_flush_rq)
329 return QUEUE_ORDSEQ_POSTFLUSH;
330
331 /*
332 * !fs requests don't need to follow barrier ordering. Always
333 * put them at the front. This fixes the following deadlock.
334 *
335 * http://thread.gmane.org/gmane.linux.kernel/537473
336 */
337 if (!blk_fs_request(rq))
338 return QUEUE_ORDSEQ_DRAIN;
339
340 if ((rq->cmd_flags & REQ_ORDERED_COLOR) ==
341 (q->orig_bar_rq->cmd_flags & REQ_ORDERED_COLOR))
342 return QUEUE_ORDSEQ_DRAIN;
343 else
344 return QUEUE_ORDSEQ_DONE;
345}
346
347void blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
348{
349 struct request *rq;
350
351 if (error && !q->orderr)
352 q->orderr = error;
353
354 BUG_ON(q->ordseq & seq);
355 q->ordseq |= seq;
356
357 if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
358 return;
359
360 /*
361 * Okay, sequence complete.
362 */
363 q->ordseq = 0;
364 rq = q->orig_bar_rq;
365
366 if (__blk_end_request(rq, q->orderr, blk_rq_bytes(rq)))
367 BUG();
368}
369
370static void pre_flush_end_io(struct request *rq, int error)
371{
372 elv_completed_request(rq->q, rq);
373 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
374}
375
376static void bar_end_io(struct request *rq, int error)
377{
378 elv_completed_request(rq->q, rq);
379 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
380}
381
382static void post_flush_end_io(struct request *rq, int error)
383{
384 elv_completed_request(rq->q, rq);
385 blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
386}
387
388static void queue_flush(struct request_queue *q, unsigned which)
389{
390 struct request *rq;
391 rq_end_io_fn *end_io;
392
393 if (which == QUEUE_ORDERED_PREFLUSH) {
394 rq = &q->pre_flush_rq;
395 end_io = pre_flush_end_io;
396 } else {
397 rq = &q->post_flush_rq;
398 end_io = post_flush_end_io;
399 }
400
401 rq->cmd_flags = REQ_HARDBARRIER;
402 rq_init(q, rq);
403 rq->elevator_private = NULL;
404 rq->elevator_private2 = NULL;
405 rq->rq_disk = q->bar_rq.rq_disk;
406 rq->end_io = end_io;
407 q->prepare_flush_fn(q, rq);
408
409 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
410}
411
412static inline struct request *start_ordered(struct request_queue *q,
413 struct request *rq)
414{
415 q->orderr = 0;
416 q->ordered = q->next_ordered;
417 q->ordseq |= QUEUE_ORDSEQ_STARTED;
418
419 /*
420 * Prep proxy barrier request.
421 */
422 blkdev_dequeue_request(rq);
423 q->orig_bar_rq = rq;
424 rq = &q->bar_rq;
425 rq->cmd_flags = 0;
426 rq_init(q, rq);
427 if (bio_data_dir(q->orig_bar_rq->bio) == WRITE)
428 rq->cmd_flags |= REQ_RW;
429 if (q->ordered & QUEUE_ORDERED_FUA)
430 rq->cmd_flags |= REQ_FUA;
431 rq->elevator_private = NULL;
432 rq->elevator_private2 = NULL;
433 init_request_from_bio(rq, q->orig_bar_rq->bio);
434 rq->end_io = bar_end_io;
435
436 /*
437 * Queue ordered sequence. As we stack them at the head, we
438 * need to queue in reverse order. Note that we rely on that
439 * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
440 * request gets inbetween ordered sequence. If this request is
441 * an empty barrier, we don't need to do a postflush ever since
442 * there will be no data written between the pre and post flush.
443 * Hence a single flush will suffice.
444 */
445 if ((q->ordered & QUEUE_ORDERED_POSTFLUSH) && !blk_empty_barrier(rq))
446 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
447 else
448 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
449
450 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
451
452 if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
453 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
454 rq = &q->pre_flush_rq;
455 } else
456 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
457
458 if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
459 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
460 else
461 rq = NULL;
462
463 return rq;
464}
465
466int blk_do_ordered(struct request_queue *q, struct request **rqp)
467{
468 struct request *rq = *rqp;
469 const int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
470
471 if (!q->ordseq) {
472 if (!is_barrier)
473 return 1;
474
475 if (q->next_ordered != QUEUE_ORDERED_NONE) {
476 *rqp = start_ordered(q, rq);
477 return 1;
478 } else {
479 /*
480 * This can happen when the queue switches to
481 * ORDERED_NONE while this request is on it.
482 */
483 blkdev_dequeue_request(rq);
484 if (__blk_end_request(rq, -EOPNOTSUPP,
485 blk_rq_bytes(rq)))
486 BUG();
487 *rqp = NULL;
488 return 0;
489 }
490 }
491
492 /*
493 * Ordered sequence in progress
494 */
495
496 /* Special requests are not subject to ordering rules. */
497 if (!blk_fs_request(rq) &&
498 rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
499 return 1;
500
501 if (q->ordered & QUEUE_ORDERED_TAG) {
502 /* Ordered by tag. Blocking the next barrier is enough. */
503 if (is_barrier && rq != &q->bar_rq)
504 *rqp = NULL;
505 } else {
506 /* Ordered by draining. Wait for turn. */
507 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
508 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
509 *rqp = NULL;
510 }
511
512 return 1;
513}
514
515static void req_bio_endio(struct request *rq, struct bio *bio,
516 unsigned int nbytes, int error)
517{
518 struct request_queue *q = rq->q;
519
520 if (&q->bar_rq != rq) {
521 if (error)
522 clear_bit(BIO_UPTODATE, &bio->bi_flags);
523 else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
524 error = -EIO;
525
526 if (unlikely(nbytes > bio->bi_size)) {
527 printk("%s: want %u bytes done, only %u left\n",
528 __FUNCTION__, nbytes, bio->bi_size);
529 nbytes = bio->bi_size;
530 }
531
532 bio->bi_size -= nbytes;
533 bio->bi_sector += (nbytes >> 9);
534 if (bio->bi_size == 0)
535 bio_endio(bio, error);
536 } else {
537
538 /*
539 * Okay, this is the barrier request in progress, just
540 * record the error;
541 */
542 if (error && !q->orderr)
543 q->orderr = error;
544 }
545}
546
547/**
548 * blk_queue_bounce_limit - set bounce buffer limit for queue
549 * @q: the request queue for the device
550 * @dma_addr: bus address limit
551 *
552 * Description:
553 * Different hardware can have different requirements as to what pages
554 * it can do I/O directly to. A low level driver can call
555 * blk_queue_bounce_limit to have lower memory pages allocated as bounce
556 * buffers for doing I/O to pages residing above @page.
557 **/
558void blk_queue_bounce_limit(struct request_queue *q, u64 dma_addr)
559{
560 unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
561 int dma = 0;
562
563 q->bounce_gfp = GFP_NOIO;
564#if BITS_PER_LONG == 64
565 /* Assume anything <= 4GB can be handled by IOMMU.
566 Actually some IOMMUs can handle everything, but I don't
567 know of a way to test this here. */
568 if (bounce_pfn < (min_t(u64,0xffffffff,BLK_BOUNCE_HIGH) >> PAGE_SHIFT))
569 dma = 1;
570 q->bounce_pfn = max_low_pfn;
571#else
572 if (bounce_pfn < blk_max_low_pfn)
573 dma = 1;
574 q->bounce_pfn = bounce_pfn;
575#endif
576 if (dma) {
577 init_emergency_isa_pool();
578 q->bounce_gfp = GFP_NOIO | GFP_DMA;
579 q->bounce_pfn = bounce_pfn;
580 }
581}
582
583EXPORT_SYMBOL(blk_queue_bounce_limit);
584
585/**
586 * blk_queue_max_sectors - set max sectors for a request for this queue
587 * @q: the request queue for the device
588 * @max_sectors: max sectors in the usual 512b unit
589 *
590 * Description:
591 * Enables a low level driver to set an upper limit on the size of
592 * received requests.
593 **/
594void blk_queue_max_sectors(struct request_queue *q, unsigned int max_sectors)
595{
596 if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
597 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
598 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
599 }
600
601 if (BLK_DEF_MAX_SECTORS > max_sectors)
602 q->max_hw_sectors = q->max_sectors = max_sectors;
603 else {
604 q->max_sectors = BLK_DEF_MAX_SECTORS;
605 q->max_hw_sectors = max_sectors;
606 }
607}
608
609EXPORT_SYMBOL(blk_queue_max_sectors);
610
611/**
612 * blk_queue_max_phys_segments - set max phys segments for a request for this queue
613 * @q: the request queue for the device
614 * @max_segments: max number of segments
615 *
616 * Description:
617 * Enables a low level driver to set an upper limit on the number of
618 * physical data segments in a request. This would be the largest sized
619 * scatter list the driver could handle.
620 **/
621void blk_queue_max_phys_segments(struct request_queue *q,
622 unsigned short max_segments)
623{
624 if (!max_segments) {
625 max_segments = 1;
626 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
627 }
628
629 q->max_phys_segments = max_segments;
630}
631
632EXPORT_SYMBOL(blk_queue_max_phys_segments);
633
634/**
635 * blk_queue_max_hw_segments - set max hw segments for a request for this queue
636 * @q: the request queue for the device
637 * @max_segments: max number of segments
638 *
639 * Description:
640 * Enables a low level driver to set an upper limit on the number of
641 * hw data segments in a request. This would be the largest number of
642 * address/length pairs the host adapter can actually give as once
643 * to the device.
644 **/
645void blk_queue_max_hw_segments(struct request_queue *q,
646 unsigned short max_segments)
647{
648 if (!max_segments) {
649 max_segments = 1;
650 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
651 }
652
653 q->max_hw_segments = max_segments;
654}
655
656EXPORT_SYMBOL(blk_queue_max_hw_segments);
657
658/**
659 * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
660 * @q: the request queue for the device
661 * @max_size: max size of segment in bytes
662 *
663 * Description:
664 * Enables a low level driver to set an upper limit on the size of a
665 * coalesced segment
666 **/
667void blk_queue_max_segment_size(struct request_queue *q, unsigned int max_size)
668{
669 if (max_size < PAGE_CACHE_SIZE) {
670 max_size = PAGE_CACHE_SIZE;
671 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
672 }
673
674 q->max_segment_size = max_size;
675}
676
677EXPORT_SYMBOL(blk_queue_max_segment_size);
678
679/**
680 * blk_queue_hardsect_size - set hardware sector size for the queue
681 * @q: the request queue for the device
682 * @size: the hardware sector size, in bytes
683 *
684 * Description:
685 * This should typically be set to the lowest possible sector size
686 * that the hardware can operate on (possible without reverting to
687 * even internal read-modify-write operations). Usually the default
688 * of 512 covers most hardware.
689 **/
690void blk_queue_hardsect_size(struct request_queue *q, unsigned short size)
691{
692 q->hardsect_size = size;
693}
694
695EXPORT_SYMBOL(blk_queue_hardsect_size);
696
697/*
698 * Returns the minimum that is _not_ zero, unless both are zero.
699 */
700#define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
701
702/**
703 * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
704 * @t: the stacking driver (top)
705 * @b: the underlying device (bottom)
706 **/
707void blk_queue_stack_limits(struct request_queue *t, struct request_queue *b)
708{
709 /* zero is "infinity" */
710 t->max_sectors = min_not_zero(t->max_sectors,b->max_sectors);
711 t->max_hw_sectors = min_not_zero(t->max_hw_sectors,b->max_hw_sectors);
712
713 t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
714 t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
715 t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
716 t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
717 if (!test_bit(QUEUE_FLAG_CLUSTER, &b->queue_flags))
718 clear_bit(QUEUE_FLAG_CLUSTER, &t->queue_flags);
719}
720
721EXPORT_SYMBOL(blk_queue_stack_limits);
722
723/**
724 * blk_queue_dma_drain - Set up a drain buffer for excess dma.
725 *
726 * @q: the request queue for the device
727 * @buf: physically contiguous buffer
728 * @size: size of the buffer in bytes
729 *
730 * Some devices have excess DMA problems and can't simply discard (or
731 * zero fill) the unwanted piece of the transfer. They have to have a
732 * real area of memory to transfer it into. The use case for this is
733 * ATAPI devices in DMA mode. If the packet command causes a transfer
734 * bigger than the transfer size some HBAs will lock up if there
735 * aren't DMA elements to contain the excess transfer. What this API
736 * does is adjust the queue so that the buf is always appended
737 * silently to the scatterlist.
738 *
739 * Note: This routine adjusts max_hw_segments to make room for
740 * appending the drain buffer. If you call
741 * blk_queue_max_hw_segments() or blk_queue_max_phys_segments() after
742 * calling this routine, you must set the limit to one fewer than your
743 * device can support otherwise there won't be room for the drain
744 * buffer.
745 */
746int blk_queue_dma_drain(struct request_queue *q, void *buf,
747 unsigned int size)
748{
749 if (q->max_hw_segments < 2 || q->max_phys_segments < 2)
750 return -EINVAL;
751 /* make room for appending the drain */
752 --q->max_hw_segments;
753 --q->max_phys_segments;
754 q->dma_drain_buffer = buf;
755 q->dma_drain_size = size;
756
757 return 0;
758}
759
760EXPORT_SYMBOL_GPL(blk_queue_dma_drain);
761
762/**
763 * blk_queue_segment_boundary - set boundary rules for segment merging
764 * @q: the request queue for the device
765 * @mask: the memory boundary mask
766 **/
767void blk_queue_segment_boundary(struct request_queue *q, unsigned long mask)
768{
769 if (mask < PAGE_CACHE_SIZE - 1) {
770 mask = PAGE_CACHE_SIZE - 1;
771 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
772 }
773
774 q->seg_boundary_mask = mask;
775}
776
777EXPORT_SYMBOL(blk_queue_segment_boundary);
778
779/**
780 * blk_queue_dma_alignment - set dma length and memory alignment
781 * @q: the request queue for the device
782 * @mask: alignment mask
783 *
784 * description:
785 * set required memory and length aligment for direct dma transactions.
786 * this is used when buiding direct io requests for the queue.
787 *
788 **/
789void blk_queue_dma_alignment(struct request_queue *q, int mask)
790{
791 q->dma_alignment = mask;
792}
793
794EXPORT_SYMBOL(blk_queue_dma_alignment);
795
796/**
797 * blk_queue_update_dma_alignment - update dma length and memory alignment
798 * @q: the request queue for the device
799 * @mask: alignment mask
800 *
801 * description:
802 * update required memory and length aligment for direct dma transactions.
803 * If the requested alignment is larger than the current alignment, then
804 * the current queue alignment is updated to the new value, otherwise it
805 * is left alone. The design of this is to allow multiple objects
806 * (driver, device, transport etc) to set their respective
807 * alignments without having them interfere.
808 *
809 **/
810void blk_queue_update_dma_alignment(struct request_queue *q, int mask)
811{
812 BUG_ON(mask > PAGE_SIZE);
813
814 if (mask > q->dma_alignment)
815 q->dma_alignment = mask;
816}
817
818EXPORT_SYMBOL(blk_queue_update_dma_alignment);
819
820/**
821 * blk_queue_find_tag - find a request by its tag and queue
822 * @q: The request queue for the device
823 * @tag: The tag of the request
824 *
825 * Notes:
826 * Should be used when a device returns a tag and you want to match
827 * it with a request.
828 *
829 * no locks need be held.
830 **/
831struct request *blk_queue_find_tag(struct request_queue *q, int tag)
832{
833 return blk_map_queue_find_tag(q->queue_tags, tag);
834}
835
836EXPORT_SYMBOL(blk_queue_find_tag);
837
838/**
839 * __blk_free_tags - release a given set of tag maintenance info
840 * @bqt: the tag map to free
841 *
842 * Tries to free the specified @bqt@. Returns true if it was
843 * actually freed and false if there are still references using it
844 */
845static int __blk_free_tags(struct blk_queue_tag *bqt)
846{
847 int retval;
848
849 retval = atomic_dec_and_test(&bqt->refcnt);
850 if (retval) {
851 BUG_ON(bqt->busy);
852
853 kfree(bqt->tag_index);
854 bqt->tag_index = NULL;
855
856 kfree(bqt->tag_map);
857 bqt->tag_map = NULL;
858
859 kfree(bqt);
860
861 }
862
863 return retval;
864}
865
866/**
867 * __blk_queue_free_tags - release tag maintenance info
868 * @q: the request queue for the device
869 *
870 * Notes:
871 * blk_cleanup_queue() will take care of calling this function, if tagging
872 * has been used. So there's no need to call this directly.
873 **/
874static void __blk_queue_free_tags(struct request_queue *q)
875{
876 struct blk_queue_tag *bqt = q->queue_tags;
877
878 if (!bqt)
879 return;
880
881 __blk_free_tags(bqt);
882
883 q->queue_tags = NULL;
884 q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
885}
886
887
888/**
889 * blk_free_tags - release a given set of tag maintenance info
890 * @bqt: the tag map to free
891 *
892 * For externally managed @bqt@ frees the map. Callers of this
893 * function must guarantee to have released all the queues that
894 * might have been using this tag map.
895 */
896void blk_free_tags(struct blk_queue_tag *bqt)
897{
898 if (unlikely(!__blk_free_tags(bqt)))
899 BUG();
900}
901EXPORT_SYMBOL(blk_free_tags);
902
903/**
904 * blk_queue_free_tags - release tag maintenance info
905 * @q: the request queue for the device
906 *
907 * Notes:
908 * This is used to disabled tagged queuing to a device, yet leave
909 * queue in function.
910 **/
911void blk_queue_free_tags(struct request_queue *q)
912{
913 clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
914}
915
916EXPORT_SYMBOL(blk_queue_free_tags);
917
918static int
919init_tag_map(struct request_queue *q, struct blk_queue_tag *tags, int depth)
920{
921 struct request **tag_index;
922 unsigned long *tag_map;
923 int nr_ulongs;
924
925 if (q && depth > q->nr_requests * 2) {
926 depth = q->nr_requests * 2;
927 printk(KERN_ERR "%s: adjusted depth to %d\n",
928 __FUNCTION__, depth);
929 }
930
931 tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC);
932 if (!tag_index)
933 goto fail;
934
935 nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
936 tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
937 if (!tag_map)
938 goto fail;
939
940 tags->real_max_depth = depth;
941 tags->max_depth = depth;
942 tags->tag_index = tag_index;
943 tags->tag_map = tag_map;
944
945 return 0;
946fail:
947 kfree(tag_index);
948 return -ENOMEM;
949}
950
951static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q,
952 int depth)
953{
954 struct blk_queue_tag *tags;
955
956 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
957 if (!tags)
958 goto fail;
959
960 if (init_tag_map(q, tags, depth))
961 goto fail;
962
963 tags->busy = 0;
964 atomic_set(&tags->refcnt, 1);
965 return tags;
966fail:
967 kfree(tags);
968 return NULL;
969}
970
971/**
972 * blk_init_tags - initialize the tag info for an external tag map
973 * @depth: the maximum queue depth supported
974 * @tags: the tag to use
975 **/
976struct blk_queue_tag *blk_init_tags(int depth)
977{
978 return __blk_queue_init_tags(NULL, depth);
979}
980EXPORT_SYMBOL(blk_init_tags);
981
982/**
983 * blk_queue_init_tags - initialize the queue tag info
984 * @q: the request queue for the device
985 * @depth: the maximum queue depth supported
986 * @tags: the tag to use
987 **/
988int blk_queue_init_tags(struct request_queue *q, int depth,
989 struct blk_queue_tag *tags)
990{
991 int rc;
992
993 BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
994
995 if (!tags && !q->queue_tags) {
996 tags = __blk_queue_init_tags(q, depth);
997
998 if (!tags)
999 goto fail;
1000 } else if (q->queue_tags) {
1001 if ((rc = blk_queue_resize_tags(q, depth)))
1002 return rc;
1003 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
1004 return 0;
1005 } else
1006 atomic_inc(&tags->refcnt);
1007
1008 /*
1009 * assign it, all done
1010 */
1011 q->queue_tags = tags;
1012 q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
1013 INIT_LIST_HEAD(&q->tag_busy_list);
1014 return 0;
1015fail:
1016 kfree(tags);
1017 return -ENOMEM;
1018}
1019
1020EXPORT_SYMBOL(blk_queue_init_tags);
1021
1022/**
1023 * blk_queue_resize_tags - change the queueing depth
1024 * @q: the request queue for the device
1025 * @new_depth: the new max command queueing depth
1026 *
1027 * Notes:
1028 * Must be called with the queue lock held.
1029 **/
1030int blk_queue_resize_tags(struct request_queue *q, int new_depth)
1031{
1032 struct blk_queue_tag *bqt = q->queue_tags;
1033 struct request **tag_index;
1034 unsigned long *tag_map;
1035 int max_depth, nr_ulongs;
1036
1037 if (!bqt)
1038 return -ENXIO;
1039
1040 /*
1041 * if we already have large enough real_max_depth. just
1042 * adjust max_depth. *NOTE* as requests with tag value
1043 * between new_depth and real_max_depth can be in-flight, tag
1044 * map can not be shrunk blindly here.
1045 */
1046 if (new_depth <= bqt->real_max_depth) {
1047 bqt->max_depth = new_depth;
1048 return 0;
1049 }
1050
1051 /*
1052 * Currently cannot replace a shared tag map with a new
1053 * one, so error out if this is the case
1054 */
1055 if (atomic_read(&bqt->refcnt) != 1)
1056 return -EBUSY;
1057
1058 /*
1059 * save the old state info, so we can copy it back
1060 */
1061 tag_index = bqt->tag_index;
1062 tag_map = bqt->tag_map;
1063 max_depth = bqt->real_max_depth;
1064
1065 if (init_tag_map(q, bqt, new_depth))
1066 return -ENOMEM;
1067
1068 memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
1069 nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
1070 memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
1071
1072 kfree(tag_index);
1073 kfree(tag_map);
1074 return 0;
1075}
1076
1077EXPORT_SYMBOL(blk_queue_resize_tags);
1078
1079/**
1080 * blk_queue_end_tag - end tag operations for a request
1081 * @q: the request queue for the device
1082 * @rq: the request that has completed
1083 *
1084 * Description:
1085 * Typically called when end_that_request_first() returns 0, meaning
1086 * all transfers have been done for a request. It's important to call
1087 * this function before end_that_request_last(), as that will put the
1088 * request back on the free list thus corrupting the internal tag list.
1089 *
1090 * Notes:
1091 * queue lock must be held.
1092 **/
1093void blk_queue_end_tag(struct request_queue *q, struct request *rq)
1094{
1095 struct blk_queue_tag *bqt = q->queue_tags;
1096 int tag = rq->tag;
1097
1098 BUG_ON(tag == -1);
1099
1100 if (unlikely(tag >= bqt->real_max_depth))
1101 /*
1102 * This can happen after tag depth has been reduced.
1103 * FIXME: how about a warning or info message here?
1104 */
1105 return;
1106
1107 list_del_init(&rq->queuelist);
1108 rq->cmd_flags &= ~REQ_QUEUED;
1109 rq->tag = -1;
1110
1111 if (unlikely(bqt->tag_index[tag] == NULL))
1112 printk(KERN_ERR "%s: tag %d is missing\n",
1113 __FUNCTION__, tag);
1114
1115 bqt->tag_index[tag] = NULL;
1116
1117 if (unlikely(!test_bit(tag, bqt->tag_map))) {
1118 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
1119 __FUNCTION__, tag);
1120 return;
1121 }
1122 /*
1123 * The tag_map bit acts as a lock for tag_index[bit], so we need
1124 * unlock memory barrier semantics.
1125 */
1126 clear_bit_unlock(tag, bqt->tag_map);
1127 bqt->busy--;
1128}
1129
1130EXPORT_SYMBOL(blk_queue_end_tag);
1131
1132/**
1133 * blk_queue_start_tag - find a free tag and assign it
1134 * @q: the request queue for the device
1135 * @rq: the block request that needs tagging
1136 *
1137 * Description:
1138 * This can either be used as a stand-alone helper, or possibly be
1139 * assigned as the queue &prep_rq_fn (in which case &struct request
1140 * automagically gets a tag assigned). Note that this function
1141 * assumes that any type of request can be queued! if this is not
1142 * true for your device, you must check the request type before
1143 * calling this function. The request will also be removed from
1144 * the request queue, so it's the drivers responsibility to readd
1145 * it if it should need to be restarted for some reason.
1146 *
1147 * Notes:
1148 * queue lock must be held.
1149 **/
1150int blk_queue_start_tag(struct request_queue *q, struct request *rq)
1151{
1152 struct blk_queue_tag *bqt = q->queue_tags;
1153 int tag;
1154
1155 if (unlikely((rq->cmd_flags & REQ_QUEUED))) {
1156 printk(KERN_ERR
1157 "%s: request %p for device [%s] already tagged %d",
1158 __FUNCTION__, rq,
1159 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
1160 BUG();
1161 }
1162
1163 /*
1164 * Protect against shared tag maps, as we may not have exclusive
1165 * access to the tag map.
1166 */
1167 do {
1168 tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
1169 if (tag >= bqt->max_depth)
1170 return 1;
1171
1172 } while (test_and_set_bit_lock(tag, bqt->tag_map));
1173 /*
1174 * We need lock ordering semantics given by test_and_set_bit_lock.
1175 * See blk_queue_end_tag for details.
1176 */
1177
1178 rq->cmd_flags |= REQ_QUEUED;
1179 rq->tag = tag;
1180 bqt->tag_index[tag] = rq;
1181 blkdev_dequeue_request(rq);
1182 list_add(&rq->queuelist, &q->tag_busy_list);
1183 bqt->busy++;
1184 return 0;
1185}
1186
1187EXPORT_SYMBOL(blk_queue_start_tag);
1188
1189/**
1190 * blk_queue_invalidate_tags - invalidate all pending tags
1191 * @q: the request queue for the device
1192 *
1193 * Description:
1194 * Hardware conditions may dictate a need to stop all pending requests.
1195 * In this case, we will safely clear the block side of the tag queue and
1196 * readd all requests to the request queue in the right order.
1197 *
1198 * Notes:
1199 * queue lock must be held.
1200 **/
1201void blk_queue_invalidate_tags(struct request_queue *q)
1202{
1203 struct list_head *tmp, *n;
1204
1205 list_for_each_safe(tmp, n, &q->tag_busy_list)
1206 blk_requeue_request(q, list_entry_rq(tmp));
1207}
1208
1209EXPORT_SYMBOL(blk_queue_invalidate_tags);
1210
1211void blk_dump_rq_flags(struct request *rq, char *msg)
1212{
1213 int bit;
1214
1215 printk("%s: dev %s: type=%x, flags=%x\n", msg,
1216 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->cmd_type,
1217 rq->cmd_flags);
1218
1219 printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1220 rq->nr_sectors,
1221 rq->current_nr_sectors);
1222 printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1223
1224 if (blk_pc_request(rq)) {
1225 printk("cdb: ");
1226 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1227 printk("%02x ", rq->cmd[bit]);
1228 printk("\n");
1229 }
1230}
1231
1232EXPORT_SYMBOL(blk_dump_rq_flags);
1233
1234void blk_recount_segments(struct request_queue *q, struct bio *bio)
1235{
1236 struct request rq;
1237 struct bio *nxt = bio->bi_next;
1238 rq.q = q;
1239 rq.bio = rq.biotail = bio;
1240 bio->bi_next = NULL;
1241 blk_recalc_rq_segments(&rq);
1242 bio->bi_next = nxt;
1243 bio->bi_phys_segments = rq.nr_phys_segments;
1244 bio->bi_hw_segments = rq.nr_hw_segments;
1245 bio->bi_flags |= (1 << BIO_SEG_VALID);
1246}
1247EXPORT_SYMBOL(blk_recount_segments);
1248
1249static void blk_recalc_rq_segments(struct request *rq)
1250{
1251 int nr_phys_segs;
1252 int nr_hw_segs;
1253 unsigned int phys_size;
1254 unsigned int hw_size;
1255 struct bio_vec *bv, *bvprv = NULL;
1256 int seg_size;
1257 int hw_seg_size;
1258 int cluster;
1259 struct req_iterator iter;
1260 int high, highprv = 1;
1261 struct request_queue *q = rq->q;
1262
1263 if (!rq->bio)
1264 return;
1265
1266 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1267 hw_seg_size = seg_size = 0;
1268 phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
1269 rq_for_each_segment(bv, rq, iter) {
1270 /*
1271 * the trick here is making sure that a high page is never
1272 * considered part of another segment, since that might
1273 * change with the bounce page.
1274 */
1275 high = page_to_pfn(bv->bv_page) > q->bounce_pfn;
1276 if (high || highprv)
1277 goto new_hw_segment;
1278 if (cluster) {
1279 if (seg_size + bv->bv_len > q->max_segment_size)
1280 goto new_segment;
1281 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1282 goto new_segment;
1283 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1284 goto new_segment;
1285 if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1286 goto new_hw_segment;
1287
1288 seg_size += bv->bv_len;
1289 hw_seg_size += bv->bv_len;
1290 bvprv = bv;
1291 continue;
1292 }
1293new_segment:
1294 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1295 !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1296 hw_seg_size += bv->bv_len;
1297 else {
1298new_hw_segment:
1299 if (nr_hw_segs == 1 &&
1300 hw_seg_size > rq->bio->bi_hw_front_size)
1301 rq->bio->bi_hw_front_size = hw_seg_size;
1302 hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1303 nr_hw_segs++;
1304 }
1305
1306 nr_phys_segs++;
1307 bvprv = bv;
1308 seg_size = bv->bv_len;
1309 highprv = high;
1310 }
1311
1312 if (nr_hw_segs == 1 &&
1313 hw_seg_size > rq->bio->bi_hw_front_size)
1314 rq->bio->bi_hw_front_size = hw_seg_size;
1315 if (hw_seg_size > rq->biotail->bi_hw_back_size)
1316 rq->biotail->bi_hw_back_size = hw_seg_size;
1317 rq->nr_phys_segments = nr_phys_segs;
1318 rq->nr_hw_segments = nr_hw_segs;
1319}
1320
1321static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
1322 struct bio *nxt)
1323{
1324 if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1325 return 0;
1326
1327 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1328 return 0;
1329 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1330 return 0;
1331
1332 /*
1333 * bio and nxt are contigous in memory, check if the queue allows
1334 * these two to be merged into one
1335 */
1336 if (BIO_SEG_BOUNDARY(q, bio, nxt))
1337 return 1;
1338
1339 return 0;
1340}
1341
1342static int blk_hw_contig_segment(struct request_queue *q, struct bio *bio,
1343 struct bio *nxt)
1344{
1345 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1346 blk_recount_segments(q, bio);
1347 if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1348 blk_recount_segments(q, nxt);
1349 if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1350 BIOVEC_VIRT_OVERSIZE(bio->bi_hw_back_size + nxt->bi_hw_front_size))
1351 return 0;
1352 if (bio->bi_hw_back_size + nxt->bi_hw_front_size > q->max_segment_size)
1353 return 0;
1354
1355 return 1;
1356}
1357
1358/*
1359 * map a request to scatterlist, return number of sg entries setup. Caller
1360 * must make sure sg can hold rq->nr_phys_segments entries
1361 */
1362int blk_rq_map_sg(struct request_queue *q, struct request *rq,
1363 struct scatterlist *sglist)
1364{
1365 struct bio_vec *bvec, *bvprv;
1366 struct req_iterator iter;
1367 struct scatterlist *sg;
1368 int nsegs, cluster;
1369
1370 nsegs = 0;
1371 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1372
1373 /*
1374 * for each bio in rq
1375 */
1376 bvprv = NULL;
1377 sg = NULL;
1378 rq_for_each_segment(bvec, rq, iter) {
1379 int nbytes = bvec->bv_len;
1380
1381 if (bvprv && cluster) {
1382 if (sg->length + nbytes > q->max_segment_size)
1383 goto new_segment;
1384
1385 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1386 goto new_segment;
1387 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1388 goto new_segment;
1389
1390 sg->length += nbytes;
1391 } else {
1392new_segment:
1393 if (!sg)
1394 sg = sglist;
1395 else {
1396 /*
1397 * If the driver previously mapped a shorter
1398 * list, we could see a termination bit
1399 * prematurely unless it fully inits the sg
1400 * table on each mapping. We KNOW that there
1401 * must be more entries here or the driver
1402 * would be buggy, so force clear the
1403 * termination bit to avoid doing a full
1404 * sg_init_table() in drivers for each command.
1405 */
1406 sg->page_link &= ~0x02;
1407 sg = sg_next(sg);
1408 }
1409
1410 sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
1411 nsegs++;
1412 }
1413 bvprv = bvec;
1414 } /* segments in rq */
1415
1416 if (q->dma_drain_size) {
1417 sg->page_link &= ~0x02;
1418 sg = sg_next(sg);
1419 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
1420 q->dma_drain_size,
1421 ((unsigned long)q->dma_drain_buffer) &
1422 (PAGE_SIZE - 1));
1423 nsegs++;
1424 }
1425
1426 if (sg)
1427 sg_mark_end(sg);
1428
1429 return nsegs;
1430}
1431
1432EXPORT_SYMBOL(blk_rq_map_sg);
1433
1434/*
1435 * the standard queue merge functions, can be overridden with device
1436 * specific ones if so desired
1437 */
1438
1439static inline int ll_new_mergeable(struct request_queue *q,
1440 struct request *req,
1441 struct bio *bio)
1442{
1443 int nr_phys_segs = bio_phys_segments(q, bio);
1444
1445 if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1446 req->cmd_flags |= REQ_NOMERGE;
1447 if (req == q->last_merge)
1448 q->last_merge = NULL;
1449 return 0;
1450 }
1451
1452 /*
1453 * A hw segment is just getting larger, bump just the phys
1454 * counter.
1455 */
1456 req->nr_phys_segments += nr_phys_segs;
1457 return 1;
1458}
1459
1460static inline int ll_new_hw_segment(struct request_queue *q,
1461 struct request *req,
1462 struct bio *bio)
1463{
1464 int nr_hw_segs = bio_hw_segments(q, bio);
1465 int nr_phys_segs = bio_phys_segments(q, bio);
1466
1467 if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1468 || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1469 req->cmd_flags |= REQ_NOMERGE;
1470 if (req == q->last_merge)
1471 q->last_merge = NULL;
1472 return 0;
1473 }
1474
1475 /*
1476 * This will form the start of a new hw segment. Bump both
1477 * counters.
1478 */
1479 req->nr_hw_segments += nr_hw_segs;
1480 req->nr_phys_segments += nr_phys_segs;
1481 return 1;
1482}
1483
1484static int ll_back_merge_fn(struct request_queue *q, struct request *req,
1485 struct bio *bio)
1486{
1487 unsigned short max_sectors;
1488 int len;
1489
1490 if (unlikely(blk_pc_request(req)))
1491 max_sectors = q->max_hw_sectors;
1492 else
1493 max_sectors = q->max_sectors;
1494
1495 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1496 req->cmd_flags |= REQ_NOMERGE;
1497 if (req == q->last_merge)
1498 q->last_merge = NULL;
1499 return 0;
1500 }
1501 if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1502 blk_recount_segments(q, req->biotail);
1503 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1504 blk_recount_segments(q, bio);
1505 len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1506 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1507 !BIOVEC_VIRT_OVERSIZE(len)) {
1508 int mergeable = ll_new_mergeable(q, req, bio);
1509
1510 if (mergeable) {
1511 if (req->nr_hw_segments == 1)
1512 req->bio->bi_hw_front_size = len;
1513 if (bio->bi_hw_segments == 1)
1514 bio->bi_hw_back_size = len;
1515 }
1516 return mergeable;
1517 }
1518
1519 return ll_new_hw_segment(q, req, bio);
1520}
1521
1522static int ll_front_merge_fn(struct request_queue *q, struct request *req,
1523 struct bio *bio)
1524{
1525 unsigned short max_sectors;
1526 int len;
1527
1528 if (unlikely(blk_pc_request(req)))
1529 max_sectors = q->max_hw_sectors;
1530 else
1531 max_sectors = q->max_sectors;
1532
1533
1534 if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1535 req->cmd_flags |= REQ_NOMERGE;
1536 if (req == q->last_merge)
1537 q->last_merge = NULL;
1538 return 0;
1539 }
1540 len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1541 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1542 blk_recount_segments(q, bio);
1543 if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1544 blk_recount_segments(q, req->bio);
1545 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1546 !BIOVEC_VIRT_OVERSIZE(len)) {
1547 int mergeable = ll_new_mergeable(q, req, bio);
1548
1549 if (mergeable) {
1550 if (bio->bi_hw_segments == 1)
1551 bio->bi_hw_front_size = len;
1552 if (req->nr_hw_segments == 1)
1553 req->biotail->bi_hw_back_size = len;
1554 }
1555 return mergeable;
1556 }
1557
1558 return ll_new_hw_segment(q, req, bio);
1559}
1560
1561static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
1562 struct request *next)
1563{
1564 int total_phys_segments;
1565 int total_hw_segments;
1566
1567 /*
1568 * First check if the either of the requests are re-queued
1569 * requests. Can't merge them if they are.
1570 */
1571 if (req->special || next->special)
1572 return 0;
1573
1574 /*
1575 * Will it become too large?
1576 */
1577 if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1578 return 0;
1579
1580 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1581 if (blk_phys_contig_segment(q, req->biotail, next->bio))
1582 total_phys_segments--;
1583
1584 if (total_phys_segments > q->max_phys_segments)
1585 return 0;
1586
1587 total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1588 if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1589 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1590 /*
1591 * propagate the combined length to the end of the requests
1592 */
1593 if (req->nr_hw_segments == 1)
1594 req->bio->bi_hw_front_size = len;
1595 if (next->nr_hw_segments == 1)
1596 next->biotail->bi_hw_back_size = len;
1597 total_hw_segments--;
1598 }
1599
1600 if (total_hw_segments > q->max_hw_segments)
1601 return 0;
1602
1603 /* Merge is OK... */
1604 req->nr_phys_segments = total_phys_segments;
1605 req->nr_hw_segments = total_hw_segments;
1606 return 1;
1607}
1608
1609/*
1610 * "plug" the device if there are no outstanding requests: this will
1611 * force the transfer to start only after we have put all the requests
1612 * on the list.
1613 *
1614 * This is called with interrupts off and no requests on the queue and
1615 * with the queue lock held.
1616 */
1617void blk_plug_device(struct request_queue *q)
1618{
1619 WARN_ON(!irqs_disabled());
1620
1621 /*
1622 * don't plug a stopped queue, it must be paired with blk_start_queue()
1623 * which will restart the queueing
1624 */
1625 if (blk_queue_stopped(q))
1626 return;
1627
1628 if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
1629 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1630 blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
1631 }
1632}
1633
1634EXPORT_SYMBOL(blk_plug_device);
1635
1636/*
1637 * remove the queue from the plugged list, if present. called with
1638 * queue lock held and interrupts disabled.
1639 */
1640int blk_remove_plug(struct request_queue *q)
1641{
1642 WARN_ON(!irqs_disabled());
1643
1644 if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1645 return 0;
1646
1647 del_timer(&q->unplug_timer);
1648 return 1;
1649}
1650
1651EXPORT_SYMBOL(blk_remove_plug);
1652
1653/*
1654 * remove the plug and let it rip..
1655 */
1656void __generic_unplug_device(struct request_queue *q)
1657{
1658 if (unlikely(blk_queue_stopped(q)))
1659 return;
1660
1661 if (!blk_remove_plug(q))
1662 return;
1663
1664 q->request_fn(q);
1665}
1666EXPORT_SYMBOL(__generic_unplug_device);
1667
1668/**
1669 * generic_unplug_device - fire a request queue
1670 * @q: The &struct request_queue in question
1671 *
1672 * Description:
1673 * Linux uses plugging to build bigger requests queues before letting
1674 * the device have at them. If a queue is plugged, the I/O scheduler
1675 * is still adding and merging requests on the queue. Once the queue
1676 * gets unplugged, the request_fn defined for the queue is invoked and
1677 * transfers started.
1678 **/
1679void generic_unplug_device(struct request_queue *q)
1680{
1681 spin_lock_irq(q->queue_lock);
1682 __generic_unplug_device(q);
1683 spin_unlock_irq(q->queue_lock);
1684}
1685EXPORT_SYMBOL(generic_unplug_device);
1686
1687static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1688 struct page *page)
1689{
1690 struct request_queue *q = bdi->unplug_io_data;
1691
1692 blk_unplug(q);
1693}
1694
1695static void blk_unplug_work(struct work_struct *work)
1696{
1697 struct request_queue *q =
1698 container_of(work, struct request_queue, unplug_work);
1699
1700 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1701 q->rq.count[READ] + q->rq.count[WRITE]);
1702
1703 q->unplug_fn(q);
1704}
1705
1706static void blk_unplug_timeout(unsigned long data)
1707{
1708 struct request_queue *q = (struct request_queue *)data;
1709
1710 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
1711 q->rq.count[READ] + q->rq.count[WRITE]);
1712
1713 kblockd_schedule_work(&q->unplug_work);
1714}
1715
1716void blk_unplug(struct request_queue *q)
1717{
1718 /*
1719 * devices don't necessarily have an ->unplug_fn defined
1720 */
1721 if (q->unplug_fn) {
1722 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1723 q->rq.count[READ] + q->rq.count[WRITE]);
1724
1725 q->unplug_fn(q);
1726 }
1727}
1728EXPORT_SYMBOL(blk_unplug);
1729
1730/**
1731 * blk_start_queue - restart a previously stopped queue
1732 * @q: The &struct request_queue in question
1733 *
1734 * Description:
1735 * blk_start_queue() will clear the stop flag on the queue, and call
1736 * the request_fn for the queue if it was in a stopped state when
1737 * entered. Also see blk_stop_queue(). Queue lock must be held.
1738 **/
1739void blk_start_queue(struct request_queue *q)
1740{
1741 WARN_ON(!irqs_disabled());
1742
1743 clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1744
1745 /*
1746 * one level of recursion is ok and is much faster than kicking
1747 * the unplug handling
1748 */
1749 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1750 q->request_fn(q);
1751 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1752 } else {
1753 blk_plug_device(q);
1754 kblockd_schedule_work(&q->unplug_work);
1755 }
1756}
1757
1758EXPORT_SYMBOL(blk_start_queue);
1759
1760/**
1761 * blk_stop_queue - stop a queue
1762 * @q: The &struct request_queue in question
1763 *
1764 * Description:
1765 * The Linux block layer assumes that a block driver will consume all
1766 * entries on the request queue when the request_fn strategy is called.
1767 * Often this will not happen, because of hardware limitations (queue
1768 * depth settings). If a device driver gets a 'queue full' response,
1769 * or if it simply chooses not to queue more I/O at one point, it can
1770 * call this function to prevent the request_fn from being called until
1771 * the driver has signalled it's ready to go again. This happens by calling
1772 * blk_start_queue() to restart queue operations. Queue lock must be held.
1773 **/
1774void blk_stop_queue(struct request_queue *q)
1775{
1776 blk_remove_plug(q);
1777 set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1778}
1779EXPORT_SYMBOL(blk_stop_queue);
1780
1781/**
1782 * blk_sync_queue - cancel any pending callbacks on a queue
1783 * @q: the queue
1784 *
1785 * Description:
1786 * The block layer may perform asynchronous callback activity
1787 * on a queue, such as calling the unplug function after a timeout.
1788 * A block device may call blk_sync_queue to ensure that any
1789 * such activity is cancelled, thus allowing it to release resources
1790 * that the callbacks might use. The caller must already have made sure
1791 * that its ->make_request_fn will not re-add plugging prior to calling
1792 * this function.
1793 *
1794 */
1795void blk_sync_queue(struct request_queue *q)
1796{
1797 del_timer_sync(&q->unplug_timer);
1798 kblockd_flush_work(&q->unplug_work);
1799}
1800EXPORT_SYMBOL(blk_sync_queue);
1801
1802/**
1803 * blk_run_queue - run a single device queue
1804 * @q: The queue to run
1805 */
1806void blk_run_queue(struct request_queue *q)
1807{
1808 unsigned long flags;
1809
1810 spin_lock_irqsave(q->queue_lock, flags);
1811 blk_remove_plug(q);
1812
1813 /*
1814 * Only recurse once to avoid overrunning the stack, let the unplug
1815 * handling reinvoke the handler shortly if we already got there.
1816 */
1817 if (!elv_queue_empty(q)) {
1818 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1819 q->request_fn(q);
1820 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1821 } else {
1822 blk_plug_device(q);
1823 kblockd_schedule_work(&q->unplug_work);
1824 }
1825 }
1826
1827 spin_unlock_irqrestore(q->queue_lock, flags);
1828}
1829EXPORT_SYMBOL(blk_run_queue);
1830
1831/**
1832 * blk_cleanup_queue: - release a &struct request_queue when it is no longer needed
1833 * @kobj: the kobj belonging of the request queue to be released
1834 *
1835 * Description:
1836 * blk_cleanup_queue is the pair to blk_init_queue() or
1837 * blk_queue_make_request(). It should be called when a request queue is
1838 * being released; typically when a block device is being de-registered.
1839 * Currently, its primary task it to free all the &struct request
1840 * structures that were allocated to the queue and the queue itself.
1841 *
1842 * Caveat:
1843 * Hopefully the low level driver will have finished any
1844 * outstanding requests first...
1845 **/
1846static void blk_release_queue(struct kobject *kobj)
1847{
1848 struct request_queue *q =
1849 container_of(kobj, struct request_queue, kobj);
1850 struct request_list *rl = &q->rq;
1851
1852 blk_sync_queue(q);
1853
1854 if (rl->rq_pool)
1855 mempool_destroy(rl->rq_pool);
1856
1857 if (q->queue_tags)
1858 __blk_queue_free_tags(q);
1859
1860 blk_trace_shutdown(q);
1861
1862 bdi_destroy(&q->backing_dev_info);
1863 kmem_cache_free(requestq_cachep, q);
1864}
1865
1866void blk_put_queue(struct request_queue *q)
1867{
1868 kobject_put(&q->kobj);
1869}
1870EXPORT_SYMBOL(blk_put_queue);
1871
1872void blk_cleanup_queue(struct request_queue * q)
1873{
1874 mutex_lock(&q->sysfs_lock);
1875 set_bit(QUEUE_FLAG_DEAD, &q->queue_flags);
1876 mutex_unlock(&q->sysfs_lock);
1877
1878 if (q->elevator)
1879 elevator_exit(q->elevator);
1880
1881 blk_put_queue(q);
1882}
1883
1884EXPORT_SYMBOL(blk_cleanup_queue);
1885
1886static int blk_init_free_list(struct request_queue *q)
1887{
1888 struct request_list *rl = &q->rq;
1889
1890 rl->count[READ] = rl->count[WRITE] = 0;
1891 rl->starved[READ] = rl->starved[WRITE] = 0;
1892 rl->elvpriv = 0;
1893 init_waitqueue_head(&rl->wait[READ]);
1894 init_waitqueue_head(&rl->wait[WRITE]);
1895
1896 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1897 mempool_free_slab, request_cachep, q->node);
1898
1899 if (!rl->rq_pool)
1900 return -ENOMEM;
1901
1902 return 0;
1903}
1904
1905struct request_queue *blk_alloc_queue(gfp_t gfp_mask)
1906{
1907 return blk_alloc_queue_node(gfp_mask, -1);
1908}
1909EXPORT_SYMBOL(blk_alloc_queue);
1910
1911static struct kobj_type queue_ktype;
1912
1913struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
1914{
1915 struct request_queue *q;
1916 int err;
1917
1918 q = kmem_cache_alloc_node(requestq_cachep,
1919 gfp_mask | __GFP_ZERO, node_id);
1920 if (!q)
1921 return NULL;
1922
1923 q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1924 q->backing_dev_info.unplug_io_data = q;
1925 err = bdi_init(&q->backing_dev_info);
1926 if (err) {
1927 kmem_cache_free(requestq_cachep, q);
1928 return NULL;
1929 }
1930
1931 init_timer(&q->unplug_timer);
1932
1933 kobject_init(&q->kobj, &queue_ktype);
1934
1935 mutex_init(&q->sysfs_lock);
1936
1937 return q;
1938}
1939EXPORT_SYMBOL(blk_alloc_queue_node);
1940
1941/**
1942 * blk_init_queue - prepare a request queue for use with a block device
1943 * @rfn: The function to be called to process requests that have been
1944 * placed on the queue.
1945 * @lock: Request queue spin lock
1946 *
1947 * Description:
1948 * If a block device wishes to use the standard request handling procedures,
1949 * which sorts requests and coalesces adjacent requests, then it must
1950 * call blk_init_queue(). The function @rfn will be called when there
1951 * are requests on the queue that need to be processed. If the device
1952 * supports plugging, then @rfn may not be called immediately when requests
1953 * are available on the queue, but may be called at some time later instead.
1954 * Plugged queues are generally unplugged when a buffer belonging to one
1955 * of the requests on the queue is needed, or due to memory pressure.
1956 *
1957 * @rfn is not required, or even expected, to remove all requests off the
1958 * queue, but only as many as it can handle at a time. If it does leave
1959 * requests on the queue, it is responsible for arranging that the requests
1960 * get dealt with eventually.
1961 *
1962 * The queue spin lock must be held while manipulating the requests on the
1963 * request queue; this lock will be taken also from interrupt context, so irq
1964 * disabling is needed for it.
1965 *
1966 * Function returns a pointer to the initialized request queue, or NULL if
1967 * it didn't succeed.
1968 *
1969 * Note:
1970 * blk_init_queue() must be paired with a blk_cleanup_queue() call
1971 * when the block device is deactivated (such as at module unload).
1972 **/
1973
1974struct request_queue *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1975{
1976 return blk_init_queue_node(rfn, lock, -1);
1977}
1978EXPORT_SYMBOL(blk_init_queue);
1979
1980struct request_queue *
1981blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1982{
1983 struct request_queue *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1984
1985 if (!q)
1986 return NULL;
1987
1988 q->node = node_id;
1989 if (blk_init_free_list(q)) {
1990 kmem_cache_free(requestq_cachep, q);
1991 return NULL;
1992 }
1993
1994 /*
1995 * if caller didn't supply a lock, they get per-queue locking with
1996 * our embedded lock
1997 */
1998 if (!lock) {
1999 spin_lock_init(&q->__queue_lock);
2000 lock = &q->__queue_lock;
2001 }
2002
2003 q->request_fn = rfn;
2004 q->prep_rq_fn = NULL;
2005 q->unplug_fn = generic_unplug_device;
2006 q->queue_flags = (1 << QUEUE_FLAG_CLUSTER);
2007 q->queue_lock = lock;
2008
2009 blk_queue_segment_boundary(q, 0xffffffff);
2010
2011 blk_queue_make_request(q, __make_request);
2012 blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
2013
2014 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
2015 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
2016
2017 q->sg_reserved_size = INT_MAX;
2018
2019 /*
2020 * all done
2021 */
2022 if (!elevator_init(q, NULL)) {
2023 blk_queue_congestion_threshold(q);
2024 return q;
2025 }
2026
2027 blk_put_queue(q);
2028 return NULL;
2029}
2030EXPORT_SYMBOL(blk_init_queue_node);
2031
2032int blk_get_queue(struct request_queue *q)
2033{
2034 if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
2035 kobject_get(&q->kobj);
2036 return 0;
2037 }
2038
2039 return 1;
2040}
2041
2042EXPORT_SYMBOL(blk_get_queue);
2043
2044static inline void blk_free_request(struct request_queue *q, struct request *rq)
2045{
2046 if (rq->cmd_flags & REQ_ELVPRIV)
2047 elv_put_request(q, rq);
2048 mempool_free(rq, q->rq.rq_pool);
2049}
2050
2051static struct request *
2052blk_alloc_request(struct request_queue *q, int rw, int priv, gfp_t gfp_mask)
2053{
2054 struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
2055
2056 if (!rq)
2057 return NULL;
2058
2059 /*
2060 * first three bits are identical in rq->cmd_flags and bio->bi_rw,
2061 * see bio.h and blkdev.h
2062 */
2063 rq->cmd_flags = rw | REQ_ALLOCED;
2064
2065 if (priv) {
2066 if (unlikely(elv_set_request(q, rq, gfp_mask))) {
2067 mempool_free(rq, q->rq.rq_pool);
2068 return NULL;
2069 }
2070 rq->cmd_flags |= REQ_ELVPRIV;
2071 }
2072
2073 return rq;
2074}
2075
2076/*
2077 * ioc_batching returns true if the ioc is a valid batching request and
2078 * should be given priority access to a request.
2079 */
2080static inline int ioc_batching(struct request_queue *q, struct io_context *ioc)
2081{
2082 if (!ioc)
2083 return 0;
2084
2085 /*
2086 * Make sure the process is able to allocate at least 1 request
2087 * even if the batch times out, otherwise we could theoretically
2088 * lose wakeups.
2089 */
2090 return ioc->nr_batch_requests == q->nr_batching ||
2091 (ioc->nr_batch_requests > 0
2092 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
2093}
2094
2095/*
2096 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
2097 * will cause the process to be a "batcher" on all queues in the system. This
2098 * is the behaviour we want though - once it gets a wakeup it should be given
2099 * a nice run.
2100 */
2101static void ioc_set_batching(struct request_queue *q, struct io_context *ioc)
2102{
2103 if (!ioc || ioc_batching(q, ioc))
2104 return;
2105
2106 ioc->nr_batch_requests = q->nr_batching;
2107 ioc->last_waited = jiffies;
2108}
2109
2110static void __freed_request(struct request_queue *q, int rw)
2111{
2112 struct request_list *rl = &q->rq;
2113
2114 if (rl->count[rw] < queue_congestion_off_threshold(q))
2115 blk_clear_queue_congested(q, rw);
2116
2117 if (rl->count[rw] + 1 <= q->nr_requests) {
2118 if (waitqueue_active(&rl->wait[rw]))
2119 wake_up(&rl->wait[rw]);
2120
2121 blk_clear_queue_full(q, rw);
2122 }
2123}
2124
2125/*
2126 * A request has just been released. Account for it, update the full and
2127 * congestion status, wake up any waiters. Called under q->queue_lock.
2128 */
2129static void freed_request(struct request_queue *q, int rw, int priv)
2130{
2131 struct request_list *rl = &q->rq;
2132
2133 rl->count[rw]--;
2134 if (priv)
2135 rl->elvpriv--;
2136
2137 __freed_request(q, rw);
2138
2139 if (unlikely(rl->starved[rw ^ 1]))
2140 __freed_request(q, rw ^ 1);
2141}
2142
2143#define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2144/*
2145 * Get a free request, queue_lock must be held.
2146 * Returns NULL on failure, with queue_lock held.
2147 * Returns !NULL on success, with queue_lock *not held*.
2148 */
2149static struct request *get_request(struct request_queue *q, int rw_flags,
2150 struct bio *bio, gfp_t gfp_mask)
2151{
2152 struct request *rq = NULL;
2153 struct request_list *rl = &q->rq;
2154 struct io_context *ioc = NULL;
2155 const int rw = rw_flags & 0x01;
2156 int may_queue, priv;
2157
2158 may_queue = elv_may_queue(q, rw_flags);
2159 if (may_queue == ELV_MQUEUE_NO)
2160 goto rq_starved;
2161
2162 if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
2163 if (rl->count[rw]+1 >= q->nr_requests) {
2164 ioc = current_io_context(GFP_ATOMIC, q->node);
2165 /*
2166 * The queue will fill after this allocation, so set
2167 * it as full, and mark this process as "batching".
2168 * This process will be allowed to complete a batch of
2169 * requests, others will be blocked.
2170 */
2171 if (!blk_queue_full(q, rw)) {
2172 ioc_set_batching(q, ioc);
2173 blk_set_queue_full(q, rw);
2174 } else {
2175 if (may_queue != ELV_MQUEUE_MUST
2176 && !ioc_batching(q, ioc)) {
2177 /*
2178 * The queue is full and the allocating
2179 * process is not a "batcher", and not
2180 * exempted by the IO scheduler
2181 */
2182 goto out;
2183 }
2184 }
2185 }
2186 blk_set_queue_congested(q, rw);
2187 }
2188
2189 /*
2190 * Only allow batching queuers to allocate up to 50% over the defined
2191 * limit of requests, otherwise we could have thousands of requests
2192 * allocated with any setting of ->nr_requests
2193 */
2194 if (rl->count[rw] >= (3 * q->nr_requests / 2))
2195 goto out;
2196
2197 rl->count[rw]++;
2198 rl->starved[rw] = 0;
2199
2200 priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
2201 if (priv)
2202 rl->elvpriv++;
2203
2204 spin_unlock_irq(q->queue_lock);
2205
2206 rq = blk_alloc_request(q, rw_flags, priv, gfp_mask);
2207 if (unlikely(!rq)) {
2208 /*
2209 * Allocation failed presumably due to memory. Undo anything
2210 * we might have messed up.
2211 *
2212 * Allocating task should really be put onto the front of the
2213 * wait queue, but this is pretty rare.
2214 */
2215 spin_lock_irq(q->queue_lock);
2216 freed_request(q, rw, priv);
2217
2218 /*
2219 * in the very unlikely event that allocation failed and no
2220 * requests for this direction was pending, mark us starved
2221 * so that freeing of a request in the other direction will
2222 * notice us. another possible fix would be to split the
2223 * rq mempool into READ and WRITE
2224 */
2225rq_starved:
2226 if (unlikely(rl->count[rw] == 0))
2227 rl->starved[rw] = 1;
2228
2229 goto out;
2230 }
2231
2232 /*
2233 * ioc may be NULL here, and ioc_batching will be false. That's
2234 * OK, if the queue is under the request limit then requests need
2235 * not count toward the nr_batch_requests limit. There will always
2236 * be some limit enforced by BLK_BATCH_TIME.
2237 */
2238 if (ioc_batching(q, ioc))
2239 ioc->nr_batch_requests--;
2240
2241 rq_init(q, rq);
2242
2243 blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
2244out:
2245 return rq;
2246}
2247
2248/*
2249 * No available requests for this queue, unplug the device and wait for some
2250 * requests to become available.
2251 *
2252 * Called with q->queue_lock held, and returns with it unlocked.
2253 */
2254static struct request *get_request_wait(struct request_queue *q, int rw_flags,
2255 struct bio *bio)
2256{
2257 const int rw = rw_flags & 0x01;
2258 struct request *rq;
2259
2260 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2261 while (!rq) {
2262 DEFINE_WAIT(wait);
2263 struct request_list *rl = &q->rq;
2264
2265 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2266 TASK_UNINTERRUPTIBLE);
2267
2268 rq = get_request(q, rw_flags, bio, GFP_NOIO);
2269
2270 if (!rq) {
2271 struct io_context *ioc;
2272
2273 blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
2274
2275 __generic_unplug_device(q);
2276 spin_unlock_irq(q->queue_lock);
2277 io_schedule();
2278
2279 /*
2280 * After sleeping, we become a "batching" process and
2281 * will be able to allocate at least one request, and
2282 * up to a big batch of them for a small period time.
2283 * See ioc_batching, ioc_set_batching
2284 */
2285 ioc = current_io_context(GFP_NOIO, q->node);
2286 ioc_set_batching(q, ioc);
2287
2288 spin_lock_irq(q->queue_lock);
2289 }
2290 finish_wait(&rl->wait[rw], &wait);
2291 }
2292
2293 return rq;
2294}
2295
2296struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
2297{
2298 struct request *rq;
2299
2300 BUG_ON(rw != READ && rw != WRITE);
2301
2302 spin_lock_irq(q->queue_lock);
2303 if (gfp_mask & __GFP_WAIT) {
2304 rq = get_request_wait(q, rw, NULL);
2305 } else {
2306 rq = get_request(q, rw, NULL, gfp_mask);
2307 if (!rq)
2308 spin_unlock_irq(q->queue_lock);
2309 }
2310 /* q->queue_lock is unlocked at this point */
2311
2312 return rq;
2313}
2314EXPORT_SYMBOL(blk_get_request);
2315
2316/**
2317 * blk_start_queueing - initiate dispatch of requests to device
2318 * @q: request queue to kick into gear
2319 *
2320 * This is basically a helper to remove the need to know whether a queue
2321 * is plugged or not if someone just wants to initiate dispatch of requests
2322 * for this queue.
2323 *
2324 * The queue lock must be held with interrupts disabled.
2325 */
2326void blk_start_queueing(struct request_queue *q)
2327{
2328 if (!blk_queue_plugged(q))
2329 q->request_fn(q);
2330 else
2331 __generic_unplug_device(q);
2332}
2333EXPORT_SYMBOL(blk_start_queueing);
2334
2335/**
2336 * blk_requeue_request - put a request back on queue
2337 * @q: request queue where request should be inserted
2338 * @rq: request to be inserted
2339 *
2340 * Description:
2341 * Drivers often keep queueing requests until the hardware cannot accept
2342 * more, when that condition happens we need to put the request back
2343 * on the queue. Must be called with queue lock held.
2344 */
2345void blk_requeue_request(struct request_queue *q, struct request *rq)
2346{
2347 blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
2348
2349 if (blk_rq_tagged(rq))
2350 blk_queue_end_tag(q, rq);
2351
2352 elv_requeue_request(q, rq);
2353}
2354
2355EXPORT_SYMBOL(blk_requeue_request);
2356
2357/**
2358 * blk_insert_request - insert a special request in to a request queue
2359 * @q: request queue where request should be inserted
2360 * @rq: request to be inserted
2361 * @at_head: insert request at head or tail of queue
2362 * @data: private data
2363 *
2364 * Description:
2365 * Many block devices need to execute commands asynchronously, so they don't
2366 * block the whole kernel from preemption during request execution. This is
2367 * accomplished normally by inserting aritficial requests tagged as
2368 * REQ_SPECIAL in to the corresponding request queue, and letting them be
2369 * scheduled for actual execution by the request queue.
2370 *
2371 * We have the option of inserting the head or the tail of the queue.
2372 * Typically we use the tail for new ioctls and so forth. We use the head
2373 * of the queue for things like a QUEUE_FULL message from a device, or a
2374 * host that is unable to accept a particular command.
2375 */
2376void blk_insert_request(struct request_queue *q, struct request *rq,
2377 int at_head, void *data)
2378{
2379 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2380 unsigned long flags;
2381
2382 /*
2383 * tell I/O scheduler that this isn't a regular read/write (ie it
2384 * must not attempt merges on this) and that it acts as a soft
2385 * barrier
2386 */
2387 rq->cmd_type = REQ_TYPE_SPECIAL;
2388 rq->cmd_flags |= REQ_SOFTBARRIER;
2389
2390 rq->special = data;
2391
2392 spin_lock_irqsave(q->queue_lock, flags);
2393
2394 /*
2395 * If command is tagged, release the tag
2396 */
2397 if (blk_rq_tagged(rq))
2398 blk_queue_end_tag(q, rq);
2399
2400 drive_stat_acct(rq, 1);
2401 __elv_add_request(q, rq, where, 0);
2402 blk_start_queueing(q);
2403 spin_unlock_irqrestore(q->queue_lock, flags);
2404}
2405
2406EXPORT_SYMBOL(blk_insert_request);
2407
2408static int __blk_rq_unmap_user(struct bio *bio)
2409{
2410 int ret = 0;
2411
2412 if (bio) {
2413 if (bio_flagged(bio, BIO_USER_MAPPED))
2414 bio_unmap_user(bio);
2415 else
2416 ret = bio_uncopy_user(bio);
2417 }
2418
2419 return ret;
2420}
2421
2422int blk_rq_append_bio(struct request_queue *q, struct request *rq,
2423 struct bio *bio)
2424{
2425 if (!rq->bio)
2426 blk_rq_bio_prep(q, rq, bio);
2427 else if (!ll_back_merge_fn(q, rq, bio))
2428 return -EINVAL;
2429 else {
2430 rq->biotail->bi_next = bio;
2431 rq->biotail = bio;
2432
2433 rq->data_len += bio->bi_size;
2434 }
2435 return 0;
2436}
2437EXPORT_SYMBOL(blk_rq_append_bio);
2438
2439static int __blk_rq_map_user(struct request_queue *q, struct request *rq,
2440 void __user *ubuf, unsigned int len)
2441{
2442 unsigned long uaddr;
2443 struct bio *bio, *orig_bio;
2444 int reading, ret;
2445
2446 reading = rq_data_dir(rq) == READ;
2447
2448 /*
2449 * if alignment requirement is satisfied, map in user pages for
2450 * direct dma. else, set up kernel bounce buffers
2451 */
2452 uaddr = (unsigned long) ubuf;
2453 if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2454 bio = bio_map_user(q, NULL, uaddr, len, reading);
2455 else
2456 bio = bio_copy_user(q, uaddr, len, reading);
2457
2458 if (IS_ERR(bio))
2459 return PTR_ERR(bio);
2460
2461 orig_bio = bio;
2462 blk_queue_bounce(q, &bio);
2463
2464 /*
2465 * We link the bounce buffer in and could have to traverse it
2466 * later so we have to get a ref to prevent it from being freed
2467 */
2468 bio_get(bio);
2469
2470 ret = blk_rq_append_bio(q, rq, bio);
2471 if (!ret)
2472 return bio->bi_size;
2473
2474 /* if it was boucned we must call the end io function */
2475 bio_endio(bio, 0);
2476 __blk_rq_unmap_user(orig_bio);
2477 bio_put(bio);
2478 return ret;
2479}
2480
2481/**
2482 * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2483 * @q: request queue where request should be inserted
2484 * @rq: request structure to fill
2485 * @ubuf: the user buffer
2486 * @len: length of user data
2487 *
2488 * Description:
2489 * Data will be mapped directly for zero copy io, if possible. Otherwise
2490 * a kernel bounce buffer is used.
2491 *
2492 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2493 * still in process context.
2494 *
2495 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2496 * before being submitted to the device, as pages mapped may be out of
2497 * reach. It's the callers responsibility to make sure this happens. The
2498 * original bio must be passed back in to blk_rq_unmap_user() for proper
2499 * unmapping.
2500 */
2501int blk_rq_map_user(struct request_queue *q, struct request *rq,
2502 void __user *ubuf, unsigned long len)
2503{
2504 unsigned long bytes_read = 0;
2505 struct bio *bio = NULL;
2506 int ret;
2507
2508 if (len > (q->max_hw_sectors << 9))
2509 return -EINVAL;
2510 if (!len || !ubuf)
2511 return -EINVAL;
2512
2513 while (bytes_read != len) {
2514 unsigned long map_len, end, start;
2515
2516 map_len = min_t(unsigned long, len - bytes_read, BIO_MAX_SIZE);
2517 end = ((unsigned long)ubuf + map_len + PAGE_SIZE - 1)
2518 >> PAGE_SHIFT;
2519 start = (unsigned long)ubuf >> PAGE_SHIFT;
2520
2521 /*
2522 * A bad offset could cause us to require BIO_MAX_PAGES + 1
2523 * pages. If this happens we just lower the requested
2524 * mapping len by a page so that we can fit
2525 */
2526 if (end - start > BIO_MAX_PAGES)
2527 map_len -= PAGE_SIZE;
2528
2529 ret = __blk_rq_map_user(q, rq, ubuf, map_len);
2530 if (ret < 0)
2531 goto unmap_rq;
2532 if (!bio)
2533 bio = rq->bio;
2534 bytes_read += ret;
2535 ubuf += ret;
2536 }
2537
2538 rq->buffer = rq->data = NULL;
2539 return 0;
2540unmap_rq:
2541 blk_rq_unmap_user(bio);
2542 return ret;
2543}
2544
2545EXPORT_SYMBOL(blk_rq_map_user);
2546
2547/**
2548 * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2549 * @q: request queue where request should be inserted
2550 * @rq: request to map data to
2551 * @iov: pointer to the iovec
2552 * @iov_count: number of elements in the iovec
2553 * @len: I/O byte count
2554 *
2555 * Description:
2556 * Data will be mapped directly for zero copy io, if possible. Otherwise
2557 * a kernel bounce buffer is used.
2558 *
2559 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2560 * still in process context.
2561 *
2562 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2563 * before being submitted to the device, as pages mapped may be out of
2564 * reach. It's the callers responsibility to make sure this happens. The
2565 * original bio must be passed back in to blk_rq_unmap_user() for proper
2566 * unmapping.
2567 */
2568int blk_rq_map_user_iov(struct request_queue *q, struct request *rq,
2569 struct sg_iovec *iov, int iov_count, unsigned int len)
2570{
2571 struct bio *bio;
2572
2573 if (!iov || iov_count <= 0)
2574 return -EINVAL;
2575
2576 /* we don't allow misaligned data like bio_map_user() does. If the
2577 * user is using sg, they're expected to know the alignment constraints
2578 * and respect them accordingly */
2579 bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2580 if (IS_ERR(bio))
2581 return PTR_ERR(bio);
2582
2583 if (bio->bi_size != len) {
2584 bio_endio(bio, 0);
2585 bio_unmap_user(bio);
2586 return -EINVAL;
2587 }
2588
2589 bio_get(bio);
2590 blk_rq_bio_prep(q, rq, bio);
2591 rq->buffer = rq->data = NULL;
2592 return 0;
2593}
2594
2595EXPORT_SYMBOL(blk_rq_map_user_iov);
2596
2597/**
2598 * blk_rq_unmap_user - unmap a request with user data
2599 * @bio: start of bio list
2600 *
2601 * Description:
2602 * Unmap a rq previously mapped by blk_rq_map_user(). The caller must
2603 * supply the original rq->bio from the blk_rq_map_user() return, since
2604 * the io completion may have changed rq->bio.
2605 */
2606int blk_rq_unmap_user(struct bio *bio)
2607{
2608 struct bio *mapped_bio;
2609 int ret = 0, ret2;
2610
2611 while (bio) {
2612 mapped_bio = bio;
2613 if (unlikely(bio_flagged(bio, BIO_BOUNCED)))
2614 mapped_bio = bio->bi_private;
2615
2616 ret2 = __blk_rq_unmap_user(mapped_bio);
2617 if (ret2 && !ret)
2618 ret = ret2;
2619
2620 mapped_bio = bio;
2621 bio = bio->bi_next;
2622 bio_put(mapped_bio);
2623 }
2624
2625 return ret;
2626}
2627
2628EXPORT_SYMBOL(blk_rq_unmap_user);
2629
2630/**
2631 * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2632 * @q: request queue where request should be inserted
2633 * @rq: request to fill
2634 * @kbuf: the kernel buffer
2635 * @len: length of user data
2636 * @gfp_mask: memory allocation flags
2637 */
2638int blk_rq_map_kern(struct request_queue *q, struct request *rq, void *kbuf,
2639 unsigned int len, gfp_t gfp_mask)
2640{
2641 struct bio *bio;
2642
2643 if (len > (q->max_hw_sectors << 9))
2644 return -EINVAL;
2645 if (!len || !kbuf)
2646 return -EINVAL;
2647
2648 bio = bio_map_kern(q, kbuf, len, gfp_mask);
2649 if (IS_ERR(bio))
2650 return PTR_ERR(bio);
2651
2652 if (rq_data_dir(rq) == WRITE)
2653 bio->bi_rw |= (1 << BIO_RW);
2654
2655 blk_rq_bio_prep(q, rq, bio);
2656 blk_queue_bounce(q, &rq->bio);
2657 rq->buffer = rq->data = NULL;
2658 return 0;
2659}
2660
2661EXPORT_SYMBOL(blk_rq_map_kern);
2662
2663/**
2664 * blk_execute_rq_nowait - insert a request into queue for execution
2665 * @q: queue to insert the request in
2666 * @bd_disk: matching gendisk
2667 * @rq: request to insert
2668 * @at_head: insert request at head or tail of queue
2669 * @done: I/O completion handler
2670 *
2671 * Description:
2672 * Insert a fully prepared request at the back of the io scheduler queue
2673 * for execution. Don't wait for completion.
2674 */
2675void blk_execute_rq_nowait(struct request_queue *q, struct gendisk *bd_disk,
2676 struct request *rq, int at_head,
2677 rq_end_io_fn *done)
2678{
2679 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2680
2681 rq->rq_disk = bd_disk;
2682 rq->cmd_flags |= REQ_NOMERGE;
2683 rq->end_io = done;
2684 WARN_ON(irqs_disabled());
2685 spin_lock_irq(q->queue_lock);
2686 __elv_add_request(q, rq, where, 1);
2687 __generic_unplug_device(q);
2688 spin_unlock_irq(q->queue_lock);
2689}
2690EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
2691
2692/**
2693 * blk_execute_rq - insert a request into queue for execution
2694 * @q: queue to insert the request in
2695 * @bd_disk: matching gendisk
2696 * @rq: request to insert
2697 * @at_head: insert request at head or tail of queue
2698 *
2699 * Description:
2700 * Insert a fully prepared request at the back of the io scheduler queue
2701 * for execution and wait for completion.
2702 */
2703int blk_execute_rq(struct request_queue *q, struct gendisk *bd_disk,
2704 struct request *rq, int at_head)
2705{
2706 DECLARE_COMPLETION_ONSTACK(wait);
2707 char sense[SCSI_SENSE_BUFFERSIZE];
2708 int err = 0;
2709
2710 /*
2711 * we need an extra reference to the request, so we can look at
2712 * it after io completion
2713 */
2714 rq->ref_count++;
2715
2716 if (!rq->sense) {
2717 memset(sense, 0, sizeof(sense));
2718 rq->sense = sense;
2719 rq->sense_len = 0;
2720 }
2721
2722 rq->end_io_data = &wait;
2723 blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
2724 wait_for_completion(&wait);
2725
2726 if (rq->errors)
2727 err = -EIO;
2728
2729 return err;
2730}
2731
2732EXPORT_SYMBOL(blk_execute_rq);
2733
2734static void bio_end_empty_barrier(struct bio *bio, int err)
2735{
2736 if (err)
2737 clear_bit(BIO_UPTODATE, &bio->bi_flags);
2738
2739 complete(bio->bi_private);
2740}
2741
2742/**
2743 * blkdev_issue_flush - queue a flush
2744 * @bdev: blockdev to issue flush for
2745 * @error_sector: error sector
2746 *
2747 * Description:
2748 * Issue a flush for the block device in question. Caller can supply
2749 * room for storing the error offset in case of a flush error, if they
2750 * wish to. Caller must run wait_for_completion() on its own.
2751 */
2752int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2753{
2754 DECLARE_COMPLETION_ONSTACK(wait);
2755 struct request_queue *q;
2756 struct bio *bio;
2757 int ret;
2758
2759 if (bdev->bd_disk == NULL)
2760 return -ENXIO;
2761
2762 q = bdev_get_queue(bdev);
2763 if (!q)
2764 return -ENXIO;
2765
2766 bio = bio_alloc(GFP_KERNEL, 0);
2767 if (!bio)
2768 return -ENOMEM;
2769
2770 bio->bi_end_io = bio_end_empty_barrier;
2771 bio->bi_private = &wait;
2772 bio->bi_bdev = bdev;
2773 submit_bio(1 << BIO_RW_BARRIER, bio);
2774
2775 wait_for_completion(&wait);
2776
2777 /*
2778 * The driver must store the error location in ->bi_sector, if
2779 * it supports it. For non-stacked drivers, this should be copied
2780 * from rq->sector.
2781 */
2782 if (error_sector)
2783 *error_sector = bio->bi_sector;
2784
2785 ret = 0;
2786 if (!bio_flagged(bio, BIO_UPTODATE))
2787 ret = -EIO;
2788
2789 bio_put(bio);
2790 return ret;
2791}
2792
2793EXPORT_SYMBOL(blkdev_issue_flush);
2794
2795static void drive_stat_acct(struct request *rq, int new_io)
2796{
2797 int rw = rq_data_dir(rq);
2798
2799 if (!blk_fs_request(rq) || !rq->rq_disk)
2800 return;
2801
2802 if (!new_io) {
2803 __disk_stat_inc(rq->rq_disk, merges[rw]);
2804 } else {
2805 disk_round_stats(rq->rq_disk);
2806 rq->rq_disk->in_flight++;
2807 }
2808}
2809
2810/*
2811 * add-request adds a request to the linked list.
2812 * queue lock is held and interrupts disabled, as we muck with the
2813 * request queue list.
2814 */
2815static inline void add_request(struct request_queue * q, struct request * req)
2816{
2817 drive_stat_acct(req, 1);
2818
2819 /*
2820 * elevator indicated where it wants this request to be
2821 * inserted at elevator_merge time
2822 */
2823 __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2824}
2825
2826/*
2827 * disk_round_stats() - Round off the performance stats on a struct
2828 * disk_stats.
2829 *
2830 * The average IO queue length and utilisation statistics are maintained
2831 * by observing the current state of the queue length and the amount of
2832 * time it has been in this state for.
2833 *
2834 * Normally, that accounting is done on IO completion, but that can result
2835 * in more than a second's worth of IO being accounted for within any one
2836 * second, leading to >100% utilisation. To deal with that, we call this
2837 * function to do a round-off before returning the results when reading
2838 * /proc/diskstats. This accounts immediately for all queue usage up to
2839 * the current jiffies and restarts the counters again.
2840 */
2841void disk_round_stats(struct gendisk *disk)
2842{
2843 unsigned long now = jiffies;
2844
2845 if (now == disk->stamp)
2846 return;
2847
2848 if (disk->in_flight) {
2849 __disk_stat_add(disk, time_in_queue,
2850 disk->in_flight * (now - disk->stamp));
2851 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2852 }
2853 disk->stamp = now;
2854}
2855
2856EXPORT_SYMBOL_GPL(disk_round_stats);
2857
2858/*
2859 * queue lock must be held
2860 */
2861void __blk_put_request(struct request_queue *q, struct request *req)
2862{
2863 if (unlikely(!q))
2864 return;
2865 if (unlikely(--req->ref_count))
2866 return;
2867
2868 elv_completed_request(q, req);
2869
2870 /*
2871 * Request may not have originated from ll_rw_blk. if not,
2872 * it didn't come out of our reserved rq pools
2873 */
2874 if (req->cmd_flags & REQ_ALLOCED) {
2875 int rw = rq_data_dir(req);
2876 int priv = req->cmd_flags & REQ_ELVPRIV;
2877
2878 BUG_ON(!list_empty(&req->queuelist));
2879 BUG_ON(!hlist_unhashed(&req->hash));
2880
2881 blk_free_request(q, req);
2882 freed_request(q, rw, priv);
2883 }
2884}
2885
2886EXPORT_SYMBOL_GPL(__blk_put_request);
2887
2888void blk_put_request(struct request *req)
2889{
2890 unsigned long flags;
2891 struct request_queue *q = req->q;
2892
2893 /*
2894 * Gee, IDE calls in w/ NULL q. Fix IDE and remove the
2895 * following if (q) test.
2896 */
2897 if (q) {
2898 spin_lock_irqsave(q->queue_lock, flags);
2899 __blk_put_request(q, req);
2900 spin_unlock_irqrestore(q->queue_lock, flags);
2901 }
2902}
2903
2904EXPORT_SYMBOL(blk_put_request);
2905
2906/**
2907 * blk_end_sync_rq - executes a completion event on a request
2908 * @rq: request to complete
2909 * @error: end io status of the request
2910 */
2911void blk_end_sync_rq(struct request *rq, int error)
2912{
2913 struct completion *waiting = rq->end_io_data;
2914
2915 rq->end_io_data = NULL;
2916 __blk_put_request(rq->q, rq);
2917
2918 /*
2919 * complete last, if this is a stack request the process (and thus
2920 * the rq pointer) could be invalid right after this complete()
2921 */
2922 complete(waiting);
2923}
2924EXPORT_SYMBOL(blk_end_sync_rq);
2925
2926/*
2927 * Has to be called with the request spinlock acquired
2928 */
2929static int attempt_merge(struct request_queue *q, struct request *req,
2930 struct request *next)
2931{
2932 if (!rq_mergeable(req) || !rq_mergeable(next))
2933 return 0;
2934
2935 /*
2936 * not contiguous
2937 */
2938 if (req->sector + req->nr_sectors != next->sector)
2939 return 0;
2940
2941 if (rq_data_dir(req) != rq_data_dir(next)
2942 || req->rq_disk != next->rq_disk
2943 || next->special)
2944 return 0;
2945
2946 /*
2947 * If we are allowed to merge, then append bio list
2948 * from next to rq and release next. merge_requests_fn
2949 * will have updated segment counts, update sector
2950 * counts here.
2951 */
2952 if (!ll_merge_requests_fn(q, req, next))
2953 return 0;
2954
2955 /*
2956 * At this point we have either done a back merge
2957 * or front merge. We need the smaller start_time of
2958 * the merged requests to be the current request
2959 * for accounting purposes.
2960 */
2961 if (time_after(req->start_time, next->start_time))
2962 req->start_time = next->start_time;
2963
2964 req->biotail->bi_next = next->bio;
2965 req->biotail = next->biotail;
2966
2967 req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2968
2969 elv_merge_requests(q, req, next);
2970
2971 if (req->rq_disk) {
2972 disk_round_stats(req->rq_disk);
2973 req->rq_disk->in_flight--;
2974 }
2975
2976 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2977
2978 __blk_put_request(q, next);
2979 return 1;
2980}
2981
2982static inline int attempt_back_merge(struct request_queue *q,
2983 struct request *rq)
2984{
2985 struct request *next = elv_latter_request(q, rq);
2986
2987 if (next)
2988 return attempt_merge(q, rq, next);
2989
2990 return 0;
2991}
2992
2993static inline int attempt_front_merge(struct request_queue *q,
2994 struct request *rq)
2995{
2996 struct request *prev = elv_former_request(q, rq);
2997
2998 if (prev)
2999 return attempt_merge(q, prev, rq);
3000
3001 return 0;
3002}
3003
3004static void init_request_from_bio(struct request *req, struct bio *bio)
3005{
3006 req->cmd_type = REQ_TYPE_FS;
3007
3008 /*
3009 * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
3010 */
3011 if (bio_rw_ahead(bio) || bio_failfast(bio))
3012 req->cmd_flags |= REQ_FAILFAST;
3013
3014 /*
3015 * REQ_BARRIER implies no merging, but lets make it explicit
3016 */
3017 if (unlikely(bio_barrier(bio)))
3018 req->cmd_flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
3019
3020 if (bio_sync(bio))
3021 req->cmd_flags |= REQ_RW_SYNC;
3022 if (bio_rw_meta(bio))
3023 req->cmd_flags |= REQ_RW_META;
3024
3025 req->errors = 0;
3026 req->hard_sector = req->sector = bio->bi_sector;
3027 req->ioprio = bio_prio(bio);
3028 req->start_time = jiffies;
3029 blk_rq_bio_prep(req->q, req, bio);
3030}
3031
3032static int __make_request(struct request_queue *q, struct bio *bio)
3033{
3034 struct request *req;
3035 int el_ret, nr_sectors, barrier, err;
3036 const unsigned short prio = bio_prio(bio);
3037 const int sync = bio_sync(bio);
3038 int rw_flags;
3039
3040 nr_sectors = bio_sectors(bio);
3041
3042 /*
3043 * low level driver can indicate that it wants pages above a
3044 * certain limit bounced to low memory (ie for highmem, or even
3045 * ISA dma in theory)
3046 */
3047 blk_queue_bounce(q, &bio);
3048
3049 barrier = bio_barrier(bio);
3050 if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
3051 err = -EOPNOTSUPP;
3052 goto end_io;
3053 }
3054
3055 spin_lock_irq(q->queue_lock);
3056
3057 if (unlikely(barrier) || elv_queue_empty(q))
3058 goto get_rq;
3059
3060 el_ret = elv_merge(q, &req, bio);
3061 switch (el_ret) {
3062 case ELEVATOR_BACK_MERGE:
3063 BUG_ON(!rq_mergeable(req));
3064
3065 if (!ll_back_merge_fn(q, req, bio))
3066 break;
3067
3068 blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
3069
3070 req->biotail->bi_next = bio;
3071 req->biotail = bio;
3072 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
3073 req->ioprio = ioprio_best(req->ioprio, prio);
3074 drive_stat_acct(req, 0);
3075 if (!attempt_back_merge(q, req))
3076 elv_merged_request(q, req, el_ret);
3077 goto out;
3078
3079 case ELEVATOR_FRONT_MERGE:
3080 BUG_ON(!rq_mergeable(req));
3081
3082 if (!ll_front_merge_fn(q, req, bio))
3083 break;
3084
3085 blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
3086
3087 bio->bi_next = req->bio;
3088 req->bio = bio;
3089
3090 /*
3091 * may not be valid. if the low level driver said
3092 * it didn't need a bounce buffer then it better
3093 * not touch req->buffer either...
3094 */
3095 req->buffer = bio_data(bio);
3096 req->current_nr_sectors = bio_cur_sectors(bio);
3097 req->hard_cur_sectors = req->current_nr_sectors;
3098 req->sector = req->hard_sector = bio->bi_sector;
3099 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
3100 req->ioprio = ioprio_best(req->ioprio, prio);
3101 drive_stat_acct(req, 0);
3102 if (!attempt_front_merge(q, req))
3103 elv_merged_request(q, req, el_ret);
3104 goto out;
3105
3106 /* ELV_NO_MERGE: elevator says don't/can't merge. */
3107 default:
3108 ;
3109 }
3110
3111get_rq:
3112 /*
3113 * This sync check and mask will be re-done in init_request_from_bio(),
3114 * but we need to set it earlier to expose the sync flag to the
3115 * rq allocator and io schedulers.
3116 */
3117 rw_flags = bio_data_dir(bio);
3118 if (sync)
3119 rw_flags |= REQ_RW_SYNC;
3120
3121 /*
3122 * Grab a free request. This is might sleep but can not fail.
3123 * Returns with the queue unlocked.
3124 */
3125 req = get_request_wait(q, rw_flags, bio);
3126
3127 /*
3128 * After dropping the lock and possibly sleeping here, our request
3129 * may now be mergeable after it had proven unmergeable (above).
3130 * We don't worry about that case for efficiency. It won't happen
3131 * often, and the elevators are able to handle it.
3132 */
3133 init_request_from_bio(req, bio);
3134
3135 spin_lock_irq(q->queue_lock);
3136 if (elv_queue_empty(q))
3137 blk_plug_device(q);
3138 add_request(q, req);
3139out:
3140 if (sync)
3141 __generic_unplug_device(q);
3142
3143 spin_unlock_irq(q->queue_lock);
3144 return 0;
3145
3146end_io:
3147 bio_endio(bio, err);
3148 return 0;
3149}
3150
3151/*
3152 * If bio->bi_dev is a partition, remap the location
3153 */
3154static inline void blk_partition_remap(struct bio *bio)
3155{
3156 struct block_device *bdev = bio->bi_bdev;
3157
3158 if (bio_sectors(bio) && bdev != bdev->bd_contains) {
3159 struct hd_struct *p = bdev->bd_part;
3160 const int rw = bio_data_dir(bio);
3161
3162 p->sectors[rw] += bio_sectors(bio);
3163 p->ios[rw]++;
3164
3165 bio->bi_sector += p->start_sect;
3166 bio->bi_bdev = bdev->bd_contains;
3167
3168 blk_add_trace_remap(bdev_get_queue(bio->bi_bdev), bio,
3169 bdev->bd_dev, bio->bi_sector,
3170 bio->bi_sector - p->start_sect);
3171 }
3172}
3173
3174static void handle_bad_sector(struct bio *bio)
3175{
3176 char b[BDEVNAME_SIZE];
3177
3178 printk(KERN_INFO "attempt to access beyond end of device\n");
3179 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
3180 bdevname(bio->bi_bdev, b),
3181 bio->bi_rw,
3182 (unsigned long long)bio->bi_sector + bio_sectors(bio),
3183 (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
3184
3185 set_bit(BIO_EOF, &bio->bi_flags);
3186}
3187
3188#ifdef CONFIG_FAIL_MAKE_REQUEST
3189
3190static DECLARE_FAULT_ATTR(fail_make_request);
3191
3192static int __init setup_fail_make_request(char *str)
3193{
3194 return setup_fault_attr(&fail_make_request, str);
3195}
3196__setup("fail_make_request=", setup_fail_make_request);
3197
3198static int should_fail_request(struct bio *bio)
3199{
3200 if ((bio->bi_bdev->bd_disk->flags & GENHD_FL_FAIL) ||
3201 (bio->bi_bdev->bd_part && bio->bi_bdev->bd_part->make_it_fail))
3202 return should_fail(&fail_make_request, bio->bi_size);
3203
3204 return 0;
3205}
3206
3207static int __init fail_make_request_debugfs(void)
3208{
3209 return init_fault_attr_dentries(&fail_make_request,
3210 "fail_make_request");
3211}
3212
3213late_initcall(fail_make_request_debugfs);
3214
3215#else /* CONFIG_FAIL_MAKE_REQUEST */
3216
3217static inline int should_fail_request(struct bio *bio)
3218{
3219 return 0;
3220}
3221
3222#endif /* CONFIG_FAIL_MAKE_REQUEST */
3223
3224/*
3225 * Check whether this bio extends beyond the end of the device.
3226 */
3227static inline int bio_check_eod(struct bio *bio, unsigned int nr_sectors)
3228{
3229 sector_t maxsector;
3230
3231 if (!nr_sectors)
3232 return 0;
3233
3234 /* Test device or partition size, when known. */
3235 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3236 if (maxsector) {
3237 sector_t sector = bio->bi_sector;
3238
3239 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
3240 /*
3241 * This may well happen - the kernel calls bread()
3242 * without checking the size of the device, e.g., when
3243 * mounting a device.
3244 */
3245 handle_bad_sector(bio);
3246 return 1;
3247 }
3248 }
3249
3250 return 0;
3251}
3252
3253/**
3254 * generic_make_request: hand a buffer to its device driver for I/O
3255 * @bio: The bio describing the location in memory and on the device.
3256 *
3257 * generic_make_request() is used to make I/O requests of block
3258 * devices. It is passed a &struct bio, which describes the I/O that needs
3259 * to be done.
3260 *
3261 * generic_make_request() does not return any status. The
3262 * success/failure status of the request, along with notification of
3263 * completion, is delivered asynchronously through the bio->bi_end_io
3264 * function described (one day) else where.
3265 *
3266 * The caller of generic_make_request must make sure that bi_io_vec
3267 * are set to describe the memory buffer, and that bi_dev and bi_sector are
3268 * set to describe the device address, and the
3269 * bi_end_io and optionally bi_private are set to describe how
3270 * completion notification should be signaled.
3271 *
3272 * generic_make_request and the drivers it calls may use bi_next if this
3273 * bio happens to be merged with someone else, and may change bi_dev and
3274 * bi_sector for remaps as it sees fit. So the values of these fields
3275 * should NOT be depended on after the call to generic_make_request.
3276 */
3277static inline void __generic_make_request(struct bio *bio)
3278{
3279 struct request_queue *q;
3280 sector_t old_sector;
3281 int ret, nr_sectors = bio_sectors(bio);
3282 dev_t old_dev;
3283 int err = -EIO;
3284
3285 might_sleep();
3286
3287 if (bio_check_eod(bio, nr_sectors))
3288 goto end_io;
3289
3290 /*
3291 * Resolve the mapping until finished. (drivers are
3292 * still free to implement/resolve their own stacking
3293 * by explicitly returning 0)
3294 *
3295 * NOTE: we don't repeat the blk_size check for each new device.
3296 * Stacking drivers are expected to know what they are doing.
3297 */
3298 old_sector = -1;
3299 old_dev = 0;
3300 do {
3301 char b[BDEVNAME_SIZE];
3302
3303 q = bdev_get_queue(bio->bi_bdev);
3304 if (!q) {
3305 printk(KERN_ERR
3306 "generic_make_request: Trying to access "
3307 "nonexistent block-device %s (%Lu)\n",
3308 bdevname(bio->bi_bdev, b),
3309 (long long) bio->bi_sector);
3310end_io:
3311 bio_endio(bio, err);
3312 break;
3313 }
3314
3315 if (unlikely(nr_sectors > q->max_hw_sectors)) {
3316 printk("bio too big device %s (%u > %u)\n",
3317 bdevname(bio->bi_bdev, b),
3318 bio_sectors(bio),
3319 q->max_hw_sectors);
3320 goto end_io;
3321 }
3322
3323 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
3324 goto end_io;
3325
3326 if (should_fail_request(bio))
3327 goto end_io;
3328
3329 /*
3330 * If this device has partitions, remap block n
3331 * of partition p to block n+start(p) of the disk.
3332 */
3333 blk_partition_remap(bio);
3334
3335 if (old_sector != -1)
3336 blk_add_trace_remap(q, bio, old_dev, bio->bi_sector,
3337 old_sector);
3338
3339 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
3340
3341 old_sector = bio->bi_sector;
3342 old_dev = bio->bi_bdev->bd_dev;
3343
3344 if (bio_check_eod(bio, nr_sectors))
3345 goto end_io;
3346 if (bio_empty_barrier(bio) && !q->prepare_flush_fn) {
3347 err = -EOPNOTSUPP;
3348 goto end_io;
3349 }
3350
3351 ret = q->make_request_fn(q, bio);
3352 } while (ret);
3353}
3354
3355/*
3356 * We only want one ->make_request_fn to be active at a time,
3357 * else stack usage with stacked devices could be a problem.
3358 * So use current->bio_{list,tail} to keep a list of requests
3359 * submited by a make_request_fn function.
3360 * current->bio_tail is also used as a flag to say if
3361 * generic_make_request is currently active in this task or not.
3362 * If it is NULL, then no make_request is active. If it is non-NULL,
3363 * then a make_request is active, and new requests should be added
3364 * at the tail
3365 */
3366void generic_make_request(struct bio *bio)
3367{
3368 if (current->bio_tail) {
3369 /* make_request is active */
3370 *(current->bio_tail) = bio;
3371 bio->bi_next = NULL;
3372 current->bio_tail = &bio->bi_next;
3373 return;
3374 }
3375 /* following loop may be a bit non-obvious, and so deserves some
3376 * explanation.
3377 * Before entering the loop, bio->bi_next is NULL (as all callers
3378 * ensure that) so we have a list with a single bio.
3379 * We pretend that we have just taken it off a longer list, so
3380 * we assign bio_list to the next (which is NULL) and bio_tail
3381 * to &bio_list, thus initialising the bio_list of new bios to be
3382 * added. __generic_make_request may indeed add some more bios
3383 * through a recursive call to generic_make_request. If it
3384 * did, we find a non-NULL value in bio_list and re-enter the loop
3385 * from the top. In this case we really did just take the bio
3386 * of the top of the list (no pretending) and so fixup bio_list and
3387 * bio_tail or bi_next, and call into __generic_make_request again.
3388 *
3389 * The loop was structured like this to make only one call to
3390 * __generic_make_request (which is important as it is large and
3391 * inlined) and to keep the structure simple.
3392 */
3393 BUG_ON(bio->bi_next);
3394 do {
3395 current->bio_list = bio->bi_next;
3396 if (bio->bi_next == NULL)
3397 current->bio_tail = &current->bio_list;
3398 else
3399 bio->bi_next = NULL;
3400 __generic_make_request(bio);
3401 bio = current->bio_list;
3402 } while (bio);
3403 current->bio_tail = NULL; /* deactivate */
3404}
3405
3406EXPORT_SYMBOL(generic_make_request);
3407
3408/**
3409 * submit_bio: submit a bio to the block device layer for I/O
3410 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3411 * @bio: The &struct bio which describes the I/O
3412 *
3413 * submit_bio() is very similar in purpose to generic_make_request(), and
3414 * uses that function to do most of the work. Both are fairly rough
3415 * interfaces, @bio must be presetup and ready for I/O.
3416 *
3417 */
3418void submit_bio(int rw, struct bio *bio)
3419{
3420 int count = bio_sectors(bio);
3421
3422 bio->bi_rw |= rw;
3423
3424 /*
3425 * If it's a regular read/write or a barrier with data attached,
3426 * go through the normal accounting stuff before submission.
3427 */
3428 if (!bio_empty_barrier(bio)) {
3429
3430 BIO_BUG_ON(!bio->bi_size);
3431 BIO_BUG_ON(!bio->bi_io_vec);
3432
3433 if (rw & WRITE) {
3434 count_vm_events(PGPGOUT, count);
3435 } else {
3436 task_io_account_read(bio->bi_size);
3437 count_vm_events(PGPGIN, count);
3438 }
3439
3440 if (unlikely(block_dump)) {
3441 char b[BDEVNAME_SIZE];
3442 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
3443 current->comm, task_pid_nr(current),
3444 (rw & WRITE) ? "WRITE" : "READ",
3445 (unsigned long long)bio->bi_sector,
3446 bdevname(bio->bi_bdev,b));
3447 }
3448 }
3449
3450 generic_make_request(bio);
3451}
3452
3453EXPORT_SYMBOL(submit_bio);
3454
3455static void blk_recalc_rq_sectors(struct request *rq, int nsect)
3456{
3457 if (blk_fs_request(rq)) {
3458 rq->hard_sector += nsect;
3459 rq->hard_nr_sectors -= nsect;
3460
3461 /*
3462 * Move the I/O submission pointers ahead if required.
3463 */
3464 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3465 (rq->sector <= rq->hard_sector)) {
3466 rq->sector = rq->hard_sector;
3467 rq->nr_sectors = rq->hard_nr_sectors;
3468 rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3469 rq->current_nr_sectors = rq->hard_cur_sectors;
3470 rq->buffer = bio_data(rq->bio);
3471 }
3472
3473 /*
3474 * if total number of sectors is less than the first segment
3475 * size, something has gone terribly wrong
3476 */
3477 if (rq->nr_sectors < rq->current_nr_sectors) {
3478 printk("blk: request botched\n");
3479 rq->nr_sectors = rq->current_nr_sectors;
3480 }
3481 }
3482}
3483
3484/**
3485 * __end_that_request_first - end I/O on a request
3486 * @req: the request being processed
3487 * @error: 0 for success, < 0 for error
3488 * @nr_bytes: number of bytes to complete
3489 *
3490 * Description:
3491 * Ends I/O on a number of bytes attached to @req, and sets it up
3492 * for the next range of segments (if any) in the cluster.
3493 *
3494 * Return:
3495 * 0 - we are done with this request, call end_that_request_last()
3496 * 1 - still buffers pending for this request
3497 **/
3498static int __end_that_request_first(struct request *req, int error,
3499 int nr_bytes)
3500{
3501 int total_bytes, bio_nbytes, next_idx = 0;
3502 struct bio *bio;
3503
3504 blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
3505
3506 /*
3507 * for a REQ_BLOCK_PC request, we want to carry any eventual
3508 * sense key with us all the way through
3509 */
3510 if (!blk_pc_request(req))
3511 req->errors = 0;
3512
3513 if (error) {
3514 if (blk_fs_request(req) && !(req->cmd_flags & REQ_QUIET))
3515 printk("end_request: I/O error, dev %s, sector %llu\n",
3516 req->rq_disk ? req->rq_disk->disk_name : "?",
3517 (unsigned long long)req->sector);
3518 }
3519
3520 if (blk_fs_request(req) && req->rq_disk) {
3521 const int rw = rq_data_dir(req);
3522
3523 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
3524 }
3525
3526 total_bytes = bio_nbytes = 0;
3527 while ((bio = req->bio) != NULL) {
3528 int nbytes;
3529
3530 /*
3531 * For an empty barrier request, the low level driver must
3532 * store a potential error location in ->sector. We pass
3533 * that back up in ->bi_sector.
3534 */
3535 if (blk_empty_barrier(req))
3536 bio->bi_sector = req->sector;
3537
3538 if (nr_bytes >= bio->bi_size) {
3539 req->bio = bio->bi_next;
3540 nbytes = bio->bi_size;
3541 req_bio_endio(req, bio, nbytes, error);
3542 next_idx = 0;
3543 bio_nbytes = 0;
3544 } else {
3545 int idx = bio->bi_idx + next_idx;
3546
3547 if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3548 blk_dump_rq_flags(req, "__end_that");
3549 printk("%s: bio idx %d >= vcnt %d\n",
3550 __FUNCTION__,
3551 bio->bi_idx, bio->bi_vcnt);
3552 break;
3553 }
3554
3555 nbytes = bio_iovec_idx(bio, idx)->bv_len;
3556 BIO_BUG_ON(nbytes > bio->bi_size);
3557
3558 /*
3559 * not a complete bvec done
3560 */
3561 if (unlikely(nbytes > nr_bytes)) {
3562 bio_nbytes += nr_bytes;
3563 total_bytes += nr_bytes;
3564 break;
3565 }
3566
3567 /*
3568 * advance to the next vector
3569 */
3570 next_idx++;
3571 bio_nbytes += nbytes;
3572 }
3573
3574 total_bytes += nbytes;
3575 nr_bytes -= nbytes;
3576
3577 if ((bio = req->bio)) {
3578 /*
3579 * end more in this run, or just return 'not-done'
3580 */
3581 if (unlikely(nr_bytes <= 0))
3582 break;
3583 }
3584 }
3585
3586 /*
3587 * completely done
3588 */
3589 if (!req->bio)
3590 return 0;
3591
3592 /*
3593 * if the request wasn't completed, update state
3594 */
3595 if (bio_nbytes) {
3596 req_bio_endio(req, bio, bio_nbytes, error);
3597 bio->bi_idx += next_idx;
3598 bio_iovec(bio)->bv_offset += nr_bytes;
3599 bio_iovec(bio)->bv_len -= nr_bytes;
3600 }
3601
3602 blk_recalc_rq_sectors(req, total_bytes >> 9);
3603 blk_recalc_rq_segments(req);
3604 return 1;
3605}
3606
3607/*
3608 * splice the completion data to a local structure and hand off to
3609 * process_completion_queue() to complete the requests
3610 */
3611static void blk_done_softirq(struct softirq_action *h)
3612{
3613 struct list_head *cpu_list, local_list;
3614
3615 local_irq_disable();
3616 cpu_list = &__get_cpu_var(blk_cpu_done);
3617 list_replace_init(cpu_list, &local_list);
3618 local_irq_enable();
3619
3620 while (!list_empty(&local_list)) {
3621 struct request *rq = list_entry(local_list.next, struct request, donelist);
3622
3623 list_del_init(&rq->donelist);
3624 rq->q->softirq_done_fn(rq);
3625 }
3626}
3627
3628static int __cpuinit blk_cpu_notify(struct notifier_block *self, unsigned long action,
3629 void *hcpu)
3630{
3631 /*
3632 * If a CPU goes away, splice its entries to the current CPU
3633 * and trigger a run of the softirq
3634 */
3635 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
3636 int cpu = (unsigned long) hcpu;
3637
3638 local_irq_disable();
3639 list_splice_init(&per_cpu(blk_cpu_done, cpu),
3640 &__get_cpu_var(blk_cpu_done));
3641 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3642 local_irq_enable();
3643 }
3644
3645 return NOTIFY_OK;
3646}
3647
3648
3649static struct notifier_block blk_cpu_notifier __cpuinitdata = {
3650 .notifier_call = blk_cpu_notify,
3651};
3652
3653/**
3654 * blk_complete_request - end I/O on a request
3655 * @req: the request being processed
3656 *
3657 * Description:
3658 * Ends all I/O on a request. It does not handle partial completions,
3659 * unless the driver actually implements this in its completion callback
3660 * through requeueing. The actual completion happens out-of-order,
3661 * through a softirq handler. The user must have registered a completion
3662 * callback through blk_queue_softirq_done().
3663 **/
3664
3665void blk_complete_request(struct request *req)
3666{
3667 struct list_head *cpu_list;
3668 unsigned long flags;
3669
3670 BUG_ON(!req->q->softirq_done_fn);
3671
3672 local_irq_save(flags);
3673
3674 cpu_list = &__get_cpu_var(blk_cpu_done);
3675 list_add_tail(&req->donelist, cpu_list);
3676 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3677
3678 local_irq_restore(flags);
3679}
3680
3681EXPORT_SYMBOL(blk_complete_request);
3682
3683/*
3684 * queue lock must be held
3685 */
3686static void end_that_request_last(struct request *req, int error)
3687{
3688 struct gendisk *disk = req->rq_disk;
3689
3690 if (blk_rq_tagged(req))
3691 blk_queue_end_tag(req->q, req);
3692
3693 if (blk_queued_rq(req))
3694 blkdev_dequeue_request(req);
3695
3696 if (unlikely(laptop_mode) && blk_fs_request(req))
3697 laptop_io_completion();
3698
3699 /*
3700 * Account IO completion. bar_rq isn't accounted as a normal
3701 * IO on queueing nor completion. Accounting the containing
3702 * request is enough.
3703 */
3704 if (disk && blk_fs_request(req) && req != &req->q->bar_rq) {
3705 unsigned long duration = jiffies - req->start_time;
3706 const int rw = rq_data_dir(req);
3707
3708 __disk_stat_inc(disk, ios[rw]);
3709 __disk_stat_add(disk, ticks[rw], duration);
3710 disk_round_stats(disk);
3711 disk->in_flight--;
3712 }
3713
3714 if (req->end_io)
3715 req->end_io(req, error);
3716 else {
3717 if (blk_bidi_rq(req))
3718 __blk_put_request(req->next_rq->q, req->next_rq);
3719
3720 __blk_put_request(req->q, req);
3721 }
3722}
3723
3724static inline void __end_request(struct request *rq, int uptodate,
3725 unsigned int nr_bytes)
3726{
3727 int error = 0;
3728
3729 if (uptodate <= 0)
3730 error = uptodate ? uptodate : -EIO;
3731
3732 __blk_end_request(rq, error, nr_bytes);
3733}
3734
3735/**
3736 * blk_rq_bytes - Returns bytes left to complete in the entire request
3737 **/
3738unsigned int blk_rq_bytes(struct request *rq)
3739{
3740 if (blk_fs_request(rq))
3741 return rq->hard_nr_sectors << 9;
3742
3743 return rq->data_len;
3744}
3745EXPORT_SYMBOL_GPL(blk_rq_bytes);
3746
3747/**
3748 * blk_rq_cur_bytes - Returns bytes left to complete in the current segment
3749 **/
3750unsigned int blk_rq_cur_bytes(struct request *rq)
3751{
3752 if (blk_fs_request(rq))
3753 return rq->current_nr_sectors << 9;
3754
3755 if (rq->bio)
3756 return rq->bio->bi_size;
3757
3758 return rq->data_len;
3759}
3760EXPORT_SYMBOL_GPL(blk_rq_cur_bytes);
3761
3762/**
3763 * end_queued_request - end all I/O on a queued request
3764 * @rq: the request being processed
3765 * @uptodate: error value or 0/1 uptodate flag
3766 *
3767 * Description:
3768 * Ends all I/O on a request, and removes it from the block layer queues.
3769 * Not suitable for normal IO completion, unless the driver still has
3770 * the request attached to the block layer.
3771 *
3772 **/
3773void end_queued_request(struct request *rq, int uptodate)
3774{
3775 __end_request(rq, uptodate, blk_rq_bytes(rq));
3776}
3777EXPORT_SYMBOL(end_queued_request);
3778
3779/**
3780 * end_dequeued_request - end all I/O on a dequeued request
3781 * @rq: the request being processed
3782 * @uptodate: error value or 0/1 uptodate flag
3783 *
3784 * Description:
3785 * Ends all I/O on a request. The request must already have been
3786 * dequeued using blkdev_dequeue_request(), as is normally the case
3787 * for most drivers.
3788 *
3789 **/
3790void end_dequeued_request(struct request *rq, int uptodate)
3791{
3792 __end_request(rq, uptodate, blk_rq_bytes(rq));
3793}
3794EXPORT_SYMBOL(end_dequeued_request);
3795
3796
3797/**
3798 * end_request - end I/O on the current segment of the request
3799 * @req: the request being processed
3800 * @uptodate: error value or 0/1 uptodate flag
3801 *
3802 * Description:
3803 * Ends I/O on the current segment of a request. If that is the only
3804 * remaining segment, the request is also completed and freed.
3805 *
3806 * This is a remnant of how older block drivers handled IO completions.
3807 * Modern drivers typically end IO on the full request in one go, unless
3808 * they have a residual value to account for. For that case this function
3809 * isn't really useful, unless the residual just happens to be the
3810 * full current segment. In other words, don't use this function in new
3811 * code. Either use end_request_completely(), or the
3812 * end_that_request_chunk() (along with end_that_request_last()) for
3813 * partial completions.
3814 *
3815 **/
3816void end_request(struct request *req, int uptodate)
3817{
3818 __end_request(req, uptodate, req->hard_cur_sectors << 9);
3819}
3820EXPORT_SYMBOL(end_request);
3821
3822/**
3823 * blk_end_io - Generic end_io function to complete a request.
3824 * @rq: the request being processed
3825 * @error: 0 for success, < 0 for error
3826 * @nr_bytes: number of bytes to complete @rq
3827 * @bidi_bytes: number of bytes to complete @rq->next_rq
3828 * @drv_callback: function called between completion of bios in the request
3829 * and completion of the request.
3830 * If the callback returns non 0, this helper returns without
3831 * completion of the request.
3832 *
3833 * Description:
3834 * Ends I/O on a number of bytes attached to @rq and @rq->next_rq.
3835 * If @rq has leftover, sets it up for the next range of segments.
3836 *
3837 * Return:
3838 * 0 - we are done with this request
3839 * 1 - this request is not freed yet, it still has pending buffers.
3840 **/
3841static int blk_end_io(struct request *rq, int error, int nr_bytes,
3842 int bidi_bytes, int (drv_callback)(struct request *))
3843{
3844 struct request_queue *q = rq->q;
3845 unsigned long flags = 0UL;
3846
3847 if (blk_fs_request(rq) || blk_pc_request(rq)) {
3848 if (__end_that_request_first(rq, error, nr_bytes))
3849 return 1;
3850
3851 /* Bidi request must be completed as a whole */
3852 if (blk_bidi_rq(rq) &&
3853 __end_that_request_first(rq->next_rq, error, bidi_bytes))
3854 return 1;
3855 }
3856
3857 /* Special feature for tricky drivers */
3858 if (drv_callback && drv_callback(rq))
3859 return 1;
3860
3861 add_disk_randomness(rq->rq_disk);
3862
3863 spin_lock_irqsave(q->queue_lock, flags);
3864 end_that_request_last(rq, error);
3865 spin_unlock_irqrestore(q->queue_lock, flags);
3866
3867 return 0;
3868}
3869
3870/**
3871 * blk_end_request - Helper function for drivers to complete the request.
3872 * @rq: the request being processed
3873 * @error: 0 for success, < 0 for error
3874 * @nr_bytes: number of bytes to complete
3875 *
3876 * Description:
3877 * Ends I/O on a number of bytes attached to @rq.
3878 * If @rq has leftover, sets it up for the next range of segments.
3879 *
3880 * Return:
3881 * 0 - we are done with this request
3882 * 1 - still buffers pending for this request
3883 **/
3884int blk_end_request(struct request *rq, int error, int nr_bytes)
3885{
3886 return blk_end_io(rq, error, nr_bytes, 0, NULL);
3887}
3888EXPORT_SYMBOL_GPL(blk_end_request);
3889
3890/**
3891 * __blk_end_request - Helper function for drivers to complete the request.
3892 * @rq: the request being processed
3893 * @error: 0 for success, < 0 for error
3894 * @nr_bytes: number of bytes to complete
3895 *
3896 * Description:
3897 * Must be called with queue lock held unlike blk_end_request().
3898 *
3899 * Return:
3900 * 0 - we are done with this request
3901 * 1 - still buffers pending for this request
3902 **/
3903int __blk_end_request(struct request *rq, int error, int nr_bytes)
3904{
3905 if (blk_fs_request(rq) || blk_pc_request(rq)) {
3906 if (__end_that_request_first(rq, error, nr_bytes))
3907 return 1;
3908 }
3909
3910 add_disk_randomness(rq->rq_disk);
3911
3912 end_that_request_last(rq, error);
3913
3914 return 0;
3915}
3916EXPORT_SYMBOL_GPL(__blk_end_request);
3917
3918/**
3919 * blk_end_bidi_request - Helper function for drivers to complete bidi request.
3920 * @rq: the bidi request being processed
3921 * @error: 0 for success, < 0 for error
3922 * @nr_bytes: number of bytes to complete @rq
3923 * @bidi_bytes: number of bytes to complete @rq->next_rq
3924 *
3925 * Description:
3926 * Ends I/O on a number of bytes attached to @rq and @rq->next_rq.
3927 *
3928 * Return:
3929 * 0 - we are done with this request
3930 * 1 - still buffers pending for this request
3931 **/
3932int blk_end_bidi_request(struct request *rq, int error, int nr_bytes,
3933 int bidi_bytes)
3934{
3935 return blk_end_io(rq, error, nr_bytes, bidi_bytes, NULL);
3936}
3937EXPORT_SYMBOL_GPL(blk_end_bidi_request);
3938
3939/**
3940 * blk_end_request_callback - Special helper function for tricky drivers
3941 * @rq: the request being processed
3942 * @error: 0 for success, < 0 for error
3943 * @nr_bytes: number of bytes to complete
3944 * @drv_callback: function called between completion of bios in the request
3945 * and completion of the request.
3946 * If the callback returns non 0, this helper returns without
3947 * completion of the request.
3948 *
3949 * Description:
3950 * Ends I/O on a number of bytes attached to @rq.
3951 * If @rq has leftover, sets it up for the next range of segments.
3952 *
3953 * This special helper function is used only for existing tricky drivers.
3954 * (e.g. cdrom_newpc_intr() of ide-cd)
3955 * This interface will be removed when such drivers are rewritten.
3956 * Don't use this interface in other places anymore.
3957 *
3958 * Return:
3959 * 0 - we are done with this request
3960 * 1 - this request is not freed yet.
3961 * this request still has pending buffers or
3962 * the driver doesn't want to finish this request yet.
3963 **/
3964int blk_end_request_callback(struct request *rq, int error, int nr_bytes,
3965 int (drv_callback)(struct request *))
3966{
3967 return blk_end_io(rq, error, nr_bytes, 0, drv_callback);
3968}
3969EXPORT_SYMBOL_GPL(blk_end_request_callback);
3970
3971static void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
3972 struct bio *bio)
3973{
3974 /* first two bits are identical in rq->cmd_flags and bio->bi_rw */
3975 rq->cmd_flags |= (bio->bi_rw & 3);
3976
3977 rq->nr_phys_segments = bio_phys_segments(q, bio);
3978 rq->nr_hw_segments = bio_hw_segments(q, bio);
3979 rq->current_nr_sectors = bio_cur_sectors(bio);
3980 rq->hard_cur_sectors = rq->current_nr_sectors;
3981 rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3982 rq->buffer = bio_data(bio);
3983 rq->data_len = bio->bi_size;
3984
3985 rq->bio = rq->biotail = bio;
3986
3987 if (bio->bi_bdev)
3988 rq->rq_disk = bio->bi_bdev->bd_disk;
3989}
3990
3991int kblockd_schedule_work(struct work_struct *work)
3992{
3993 return queue_work(kblockd_workqueue, work);
3994}
3995
3996EXPORT_SYMBOL(kblockd_schedule_work);
3997
3998void kblockd_flush_work(struct work_struct *work)
3999{
4000 cancel_work_sync(work);
4001}
4002EXPORT_SYMBOL(kblockd_flush_work);
4003
4004int __init blk_dev_init(void)
4005{
4006 int i;
4007
4008 kblockd_workqueue = create_workqueue("kblockd");
4009 if (!kblockd_workqueue)
4010 panic("Failed to create kblockd\n");
4011
4012 request_cachep = kmem_cache_create("blkdev_requests",
4013 sizeof(struct request), 0, SLAB_PANIC, NULL);
4014
4015 requestq_cachep = kmem_cache_create("blkdev_queue",
4016 sizeof(struct request_queue), 0, SLAB_PANIC, NULL);
4017
4018 iocontext_cachep = kmem_cache_create("blkdev_ioc",
4019 sizeof(struct io_context), 0, SLAB_PANIC, NULL);
4020
4021 for_each_possible_cpu(i)
4022 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
4023
4024 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
4025 register_hotcpu_notifier(&blk_cpu_notifier);
4026
4027 blk_max_low_pfn = max_low_pfn - 1;
4028 blk_max_pfn = max_pfn - 1;
4029
4030 return 0;
4031}
4032
4033static void cfq_dtor(struct io_context *ioc)
4034{
4035 struct cfq_io_context *cic[1];
4036 int r;
4037
4038 /*
4039 * We don't have a specific key to lookup with, so use the gang
4040 * lookup to just retrieve the first item stored. The cfq exit
4041 * function will iterate the full tree, so any member will do.
4042 */
4043 r = radix_tree_gang_lookup(&ioc->radix_root, (void **) cic, 0, 1);
4044 if (r > 0)
4045 cic[0]->dtor(ioc);
4046}
4047
4048/*
4049 * IO Context helper functions. put_io_context() returns 1 if there are no
4050 * more users of this io context, 0 otherwise.
4051 */
4052int put_io_context(struct io_context *ioc)
4053{
4054 if (ioc == NULL)
4055 return 1;
4056
4057 BUG_ON(atomic_read(&ioc->refcount) == 0);
4058
4059 if (atomic_dec_and_test(&ioc->refcount)) {
4060 rcu_read_lock();
4061 if (ioc->aic && ioc->aic->dtor)
4062 ioc->aic->dtor(ioc->aic);
4063 rcu_read_unlock();
4064 cfq_dtor(ioc);
4065
4066 kmem_cache_free(iocontext_cachep, ioc);
4067 return 1;
4068 }
4069 return 0;
4070}
4071EXPORT_SYMBOL(put_io_context);
4072
4073static void cfq_exit(struct io_context *ioc)
4074{
4075 struct cfq_io_context *cic[1];
4076 int r;
4077
4078 rcu_read_lock();
4079 /*
4080 * See comment for cfq_dtor()
4081 */
4082 r = radix_tree_gang_lookup(&ioc->radix_root, (void **) cic, 0, 1);
4083 rcu_read_unlock();
4084
4085 if (r > 0)
4086 cic[0]->exit(ioc);
4087}
4088
4089/* Called by the exitting task */
4090void exit_io_context(void)
4091{
4092 struct io_context *ioc;
4093
4094 task_lock(current);
4095 ioc = current->io_context;
4096 current->io_context = NULL;
4097 task_unlock(current);
4098
4099 if (atomic_dec_and_test(&ioc->nr_tasks)) {
4100 if (ioc->aic && ioc->aic->exit)
4101 ioc->aic->exit(ioc->aic);
4102 cfq_exit(ioc);
4103
4104 put_io_context(ioc);
4105 }
4106}
4107
4108struct io_context *alloc_io_context(gfp_t gfp_flags, int node)
4109{
4110 struct io_context *ret;
4111
4112 ret = kmem_cache_alloc_node(iocontext_cachep, gfp_flags, node);
4113 if (ret) {
4114 atomic_set(&ret->refcount, 1);
4115 atomic_set(&ret->nr_tasks, 1);
4116 spin_lock_init(&ret->lock);
4117 ret->ioprio_changed = 0;
4118 ret->ioprio = 0;
4119 ret->last_waited = jiffies; /* doesn't matter... */
4120 ret->nr_batch_requests = 0; /* because this is 0 */
4121 ret->aic = NULL;
4122 INIT_RADIX_TREE(&ret->radix_root, GFP_ATOMIC | __GFP_HIGH);
4123 ret->ioc_data = NULL;
4124 }
4125
4126 return ret;
4127}
4128
4129/*
4130 * If the current task has no IO context then create one and initialise it.
4131 * Otherwise, return its existing IO context.
4132 *
4133 * This returned IO context doesn't have a specifically elevated refcount,
4134 * but since the current task itself holds a reference, the context can be
4135 * used in general code, so long as it stays within `current` context.
4136 */
4137static struct io_context *current_io_context(gfp_t gfp_flags, int node)
4138{
4139 struct task_struct *tsk = current;
4140 struct io_context *ret;
4141
4142 ret = tsk->io_context;
4143 if (likely(ret))
4144 return ret;
4145
4146 ret = alloc_io_context(gfp_flags, node);
4147 if (ret) {
4148 /* make sure set_task_ioprio() sees the settings above */
4149 smp_wmb();
4150 tsk->io_context = ret;
4151 }
4152
4153 return ret;
4154}
4155
4156/*
4157 * If the current task has no IO context then create one and initialise it.
4158 * If it does have a context, take a ref on it.
4159 *
4160 * This is always called in the context of the task which submitted the I/O.
4161 */
4162struct io_context *get_io_context(gfp_t gfp_flags, int node)
4163{
4164 struct io_context *ret = NULL;
4165
4166 /*
4167 * Check for unlikely race with exiting task. ioc ref count is
4168 * zero when ioc is being detached.
4169 */
4170 do {
4171 ret = current_io_context(gfp_flags, node);
4172 if (unlikely(!ret))
4173 break;
4174 } while (!atomic_inc_not_zero(&ret->refcount));
4175
4176 return ret;
4177}
4178EXPORT_SYMBOL(get_io_context);
4179
4180void copy_io_context(struct io_context **pdst, struct io_context **psrc)
4181{
4182 struct io_context *src = *psrc;
4183 struct io_context *dst = *pdst;
4184
4185 if (src) {
4186 BUG_ON(atomic_read(&src->refcount) == 0);
4187 atomic_inc(&src->refcount);
4188 put_io_context(dst);
4189 *pdst = src;
4190 }
4191}
4192EXPORT_SYMBOL(copy_io_context);
4193
4194void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
4195{
4196 struct io_context *temp;
4197 temp = *ioc1;
4198 *ioc1 = *ioc2;
4199 *ioc2 = temp;
4200}
4201EXPORT_SYMBOL(swap_io_context);
4202
4203/*
4204 * sysfs parts below
4205 */
4206struct queue_sysfs_entry {
4207 struct attribute attr;
4208 ssize_t (*show)(struct request_queue *, char *);
4209 ssize_t (*store)(struct request_queue *, const char *, size_t);
4210};
4211
4212static ssize_t
4213queue_var_show(unsigned int var, char *page)
4214{
4215 return sprintf(page, "%d\n", var);
4216}
4217
4218static ssize_t
4219queue_var_store(unsigned long *var, const char *page, size_t count)
4220{
4221 char *p = (char *) page;
4222
4223 *var = simple_strtoul(p, &p, 10);
4224 return count;
4225}
4226
4227static ssize_t queue_requests_show(struct request_queue *q, char *page)
4228{
4229 return queue_var_show(q->nr_requests, (page));
4230}
4231
4232static ssize_t
4233queue_requests_store(struct request_queue *q, const char *page, size_t count)
4234{
4235 struct request_list *rl = &q->rq;
4236 unsigned long nr;
4237 int ret = queue_var_store(&nr, page, count);
4238 if (nr < BLKDEV_MIN_RQ)
4239 nr = BLKDEV_MIN_RQ;
4240
4241 spin_lock_irq(q->queue_lock);
4242 q->nr_requests = nr;
4243 blk_queue_congestion_threshold(q);
4244
4245 if (rl->count[READ] >= queue_congestion_on_threshold(q))
4246 blk_set_queue_congested(q, READ);
4247 else if (rl->count[READ] < queue_congestion_off_threshold(q))
4248 blk_clear_queue_congested(q, READ);
4249
4250 if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
4251 blk_set_queue_congested(q, WRITE);
4252 else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
4253 blk_clear_queue_congested(q, WRITE);
4254
4255 if (rl->count[READ] >= q->nr_requests) {
4256 blk_set_queue_full(q, READ);
4257 } else if (rl->count[READ]+1 <= q->nr_requests) {
4258 blk_clear_queue_full(q, READ);
4259 wake_up(&rl->wait[READ]);
4260 }
4261
4262 if (rl->count[WRITE] >= q->nr_requests) {
4263 blk_set_queue_full(q, WRITE);
4264 } else if (rl->count[WRITE]+1 <= q->nr_requests) {
4265 blk_clear_queue_full(q, WRITE);
4266 wake_up(&rl->wait[WRITE]);
4267 }
4268 spin_unlock_irq(q->queue_lock);
4269 return ret;
4270}
4271
4272static ssize_t queue_ra_show(struct request_queue *q, char *page)
4273{
4274 int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
4275
4276 return queue_var_show(ra_kb, (page));
4277}
4278
4279static ssize_t
4280queue_ra_store(struct request_queue *q, const char *page, size_t count)
4281{
4282 unsigned long ra_kb;
4283 ssize_t ret = queue_var_store(&ra_kb, page, count);
4284
4285 spin_lock_irq(q->queue_lock);
4286 q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
4287 spin_unlock_irq(q->queue_lock);
4288
4289 return ret;
4290}
4291
4292static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
4293{
4294 int max_sectors_kb = q->max_sectors >> 1;
4295
4296 return queue_var_show(max_sectors_kb, (page));
4297}
4298
4299static ssize_t
4300queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
4301{
4302 unsigned long max_sectors_kb,
4303 max_hw_sectors_kb = q->max_hw_sectors >> 1,
4304 page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
4305 ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
4306
4307 if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
4308 return -EINVAL;
4309 /*
4310 * Take the queue lock to update the readahead and max_sectors
4311 * values synchronously:
4312 */
4313 spin_lock_irq(q->queue_lock);
4314 q->max_sectors = max_sectors_kb << 1;
4315 spin_unlock_irq(q->queue_lock);
4316
4317 return ret;
4318}
4319
4320static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
4321{
4322 int max_hw_sectors_kb = q->max_hw_sectors >> 1;
4323
4324 return queue_var_show(max_hw_sectors_kb, (page));
4325}
4326
4327
4328static struct queue_sysfs_entry queue_requests_entry = {
4329 .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
4330 .show = queue_requests_show,
4331 .store = queue_requests_store,
4332};
4333
4334static struct queue_sysfs_entry queue_ra_entry = {
4335 .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
4336 .show = queue_ra_show,
4337 .store = queue_ra_store,
4338};
4339
4340static struct queue_sysfs_entry queue_max_sectors_entry = {
4341 .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
4342 .show = queue_max_sectors_show,
4343 .store = queue_max_sectors_store,
4344};
4345
4346static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
4347 .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
4348 .show = queue_max_hw_sectors_show,
4349};
4350
4351static struct queue_sysfs_entry queue_iosched_entry = {
4352 .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
4353 .show = elv_iosched_show,
4354 .store = elv_iosched_store,
4355};
4356
4357static struct attribute *default_attrs[] = {
4358 &queue_requests_entry.attr,
4359 &queue_ra_entry.attr,
4360 &queue_max_hw_sectors_entry.attr,
4361 &queue_max_sectors_entry.attr,
4362 &queue_iosched_entry.attr,
4363 NULL,
4364};
4365
4366#define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
4367
4368static ssize_t
4369queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
4370{
4371 struct queue_sysfs_entry *entry = to_queue(attr);
4372 struct request_queue *q =
4373 container_of(kobj, struct request_queue, kobj);
4374 ssize_t res;
4375
4376 if (!entry->show)
4377 return -EIO;
4378 mutex_lock(&q->sysfs_lock);
4379 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
4380 mutex_unlock(&q->sysfs_lock);
4381 return -ENOENT;
4382 }
4383 res = entry->show(q, page);
4384 mutex_unlock(&q->sysfs_lock);
4385 return res;
4386}
4387
4388static ssize_t
4389queue_attr_store(struct kobject *kobj, struct attribute *attr,
4390 const char *page, size_t length)
4391{
4392 struct queue_sysfs_entry *entry = to_queue(attr);
4393 struct request_queue *q = container_of(kobj, struct request_queue, kobj);
4394
4395 ssize_t res;
4396
4397 if (!entry->store)
4398 return -EIO;
4399 mutex_lock(&q->sysfs_lock);
4400 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
4401 mutex_unlock(&q->sysfs_lock);
4402 return -ENOENT;
4403 }
4404 res = entry->store(q, page, length);
4405 mutex_unlock(&q->sysfs_lock);
4406 return res;
4407}
4408
4409static struct sysfs_ops queue_sysfs_ops = {
4410 .show = queue_attr_show,
4411 .store = queue_attr_store,
4412};
4413
4414static struct kobj_type queue_ktype = {
4415 .sysfs_ops = &queue_sysfs_ops,
4416 .default_attrs = default_attrs,
4417 .release = blk_release_queue,
4418};
4419
4420int blk_register_queue(struct gendisk *disk)
4421{
4422 int ret;
4423
4424 struct request_queue *q = disk->queue;
4425
4426 if (!q || !q->request_fn)
4427 return -ENXIO;
4428
4429 ret = kobject_add(&q->kobj, kobject_get(&disk->dev.kobj),
4430 "%s", "queue");
4431 if (ret < 0)
4432 return ret;
4433
4434 kobject_uevent(&q->kobj, KOBJ_ADD);
4435
4436 ret = elv_register_queue(q);
4437 if (ret) {
4438 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4439 kobject_del(&q->kobj);
4440 return ret;
4441 }
4442
4443 return 0;
4444}
4445
4446void blk_unregister_queue(struct gendisk *disk)
4447{
4448 struct request_queue *q = disk->queue;
4449
4450 if (q && q->request_fn) {
4451 elv_unregister_queue(q);
4452
4453 kobject_uevent(&q->kobj, KOBJ_REMOVE);
4454 kobject_del(&q->kobj);
4455 kobject_put(&disk->dev.kobj);
4456 }
4457}