Merge ../linux-2.6
[linux-2.6] / block / ll_rw_blk.c
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/config.h>
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/backing-dev.h>
17 #include <linux/bio.h>
18 #include <linux/blkdev.h>
19 #include <linux/highmem.h>
20 #include <linux/mm.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/string.h>
23 #include <linux/init.h>
24 #include <linux/bootmem.h>      /* for max_pfn/max_low_pfn */
25 #include <linux/completion.h>
26 #include <linux/slab.h>
27 #include <linux/swap.h>
28 #include <linux/writeback.h>
29 #include <linux/interrupt.h>
30 #include <linux/cpu.h>
31
32 /*
33  * for max sense size
34  */
35 #include <scsi/scsi_cmnd.h>
36
37 static void blk_unplug_work(void *data);
38 static void blk_unplug_timeout(unsigned long data);
39 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io);
40 static void init_request_from_bio(struct request *req, struct bio *bio);
41 static int __make_request(request_queue_t *q, struct bio *bio);
42
43 /*
44  * For the allocated request tables
45  */
46 static kmem_cache_t *request_cachep;
47
48 /*
49  * For queue allocation
50  */
51 static kmem_cache_t *requestq_cachep;
52
53 /*
54  * For io context allocations
55  */
56 static kmem_cache_t *iocontext_cachep;
57
58 static wait_queue_head_t congestion_wqh[2] = {
59                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
60                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
61         };
62
63 /*
64  * Controlling structure to kblockd
65  */
66 static struct workqueue_struct *kblockd_workqueue;
67
68 unsigned long blk_max_low_pfn, blk_max_pfn;
69
70 EXPORT_SYMBOL(blk_max_low_pfn);
71 EXPORT_SYMBOL(blk_max_pfn);
72
73 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
74
75 /* Amount of time in which a process may batch requests */
76 #define BLK_BATCH_TIME  (HZ/50UL)
77
78 /* Number of requests a "batching" process may submit */
79 #define BLK_BATCH_REQ   32
80
81 /*
82  * Return the threshold (number of used requests) at which the queue is
83  * considered to be congested.  It include a little hysteresis to keep the
84  * context switch rate down.
85  */
86 static inline int queue_congestion_on_threshold(struct request_queue *q)
87 {
88         return q->nr_congestion_on;
89 }
90
91 /*
92  * The threshold at which a queue is considered to be uncongested
93  */
94 static inline int queue_congestion_off_threshold(struct request_queue *q)
95 {
96         return q->nr_congestion_off;
97 }
98
99 static void blk_queue_congestion_threshold(struct request_queue *q)
100 {
101         int nr;
102
103         nr = q->nr_requests - (q->nr_requests / 8) + 1;
104         if (nr > q->nr_requests)
105                 nr = q->nr_requests;
106         q->nr_congestion_on = nr;
107
108         nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
109         if (nr < 1)
110                 nr = 1;
111         q->nr_congestion_off = nr;
112 }
113
114 /*
115  * A queue has just exitted congestion.  Note this in the global counter of
116  * congested queues, and wake up anyone who was waiting for requests to be
117  * put back.
118  */
119 static void clear_queue_congested(request_queue_t *q, int rw)
120 {
121         enum bdi_state bit;
122         wait_queue_head_t *wqh = &congestion_wqh[rw];
123
124         bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
125         clear_bit(bit, &q->backing_dev_info.state);
126         smp_mb__after_clear_bit();
127         if (waitqueue_active(wqh))
128                 wake_up(wqh);
129 }
130
131 /*
132  * A queue has just entered congestion.  Flag that in the queue's VM-visible
133  * state flags and increment the global gounter of congested queues.
134  */
135 static void set_queue_congested(request_queue_t *q, int rw)
136 {
137         enum bdi_state bit;
138
139         bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
140         set_bit(bit, &q->backing_dev_info.state);
141 }
142
143 /**
144  * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
145  * @bdev:       device
146  *
147  * Locates the passed device's request queue and returns the address of its
148  * backing_dev_info
149  *
150  * Will return NULL if the request queue cannot be located.
151  */
152 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
153 {
154         struct backing_dev_info *ret = NULL;
155         request_queue_t *q = bdev_get_queue(bdev);
156
157         if (q)
158                 ret = &q->backing_dev_info;
159         return ret;
160 }
161
162 EXPORT_SYMBOL(blk_get_backing_dev_info);
163
164 void blk_queue_activity_fn(request_queue_t *q, activity_fn *fn, void *data)
165 {
166         q->activity_fn = fn;
167         q->activity_data = data;
168 }
169
170 EXPORT_SYMBOL(blk_queue_activity_fn);
171
172 /**
173  * blk_queue_prep_rq - set a prepare_request function for queue
174  * @q:          queue
175  * @pfn:        prepare_request function
176  *
177  * It's possible for a queue to register a prepare_request callback which
178  * is invoked before the request is handed to the request_fn. The goal of
179  * the function is to prepare a request for I/O, it can be used to build a
180  * cdb from the request data for instance.
181  *
182  */
183 void blk_queue_prep_rq(request_queue_t *q, prep_rq_fn *pfn)
184 {
185         q->prep_rq_fn = pfn;
186 }
187
188 EXPORT_SYMBOL(blk_queue_prep_rq);
189
190 /**
191  * blk_queue_merge_bvec - set a merge_bvec function for queue
192  * @q:          queue
193  * @mbfn:       merge_bvec_fn
194  *
195  * Usually queues have static limitations on the max sectors or segments that
196  * we can put in a request. Stacking drivers may have some settings that
197  * are dynamic, and thus we have to query the queue whether it is ok to
198  * add a new bio_vec to a bio at a given offset or not. If the block device
199  * has such limitations, it needs to register a merge_bvec_fn to control
200  * the size of bio's sent to it. Note that a block device *must* allow a
201  * single page to be added to an empty bio. The block device driver may want
202  * to use the bio_split() function to deal with these bio's. By default
203  * no merge_bvec_fn is defined for a queue, and only the fixed limits are
204  * honored.
205  */
206 void blk_queue_merge_bvec(request_queue_t *q, merge_bvec_fn *mbfn)
207 {
208         q->merge_bvec_fn = mbfn;
209 }
210
211 EXPORT_SYMBOL(blk_queue_merge_bvec);
212
213 void blk_queue_softirq_done(request_queue_t *q, softirq_done_fn *fn)
214 {
215         q->softirq_done_fn = fn;
216 }
217
218 EXPORT_SYMBOL(blk_queue_softirq_done);
219
220 /**
221  * blk_queue_make_request - define an alternate make_request function for a device
222  * @q:  the request queue for the device to be affected
223  * @mfn: the alternate make_request function
224  *
225  * Description:
226  *    The normal way for &struct bios to be passed to a device
227  *    driver is for them to be collected into requests on a request
228  *    queue, and then to allow the device driver to select requests
229  *    off that queue when it is ready.  This works well for many block
230  *    devices. However some block devices (typically virtual devices
231  *    such as md or lvm) do not benefit from the processing on the
232  *    request queue, and are served best by having the requests passed
233  *    directly to them.  This can be achieved by providing a function
234  *    to blk_queue_make_request().
235  *
236  * Caveat:
237  *    The driver that does this *must* be able to deal appropriately
238  *    with buffers in "highmemory". This can be accomplished by either calling
239  *    __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
240  *    blk_queue_bounce() to create a buffer in normal memory.
241  **/
242 void blk_queue_make_request(request_queue_t * q, make_request_fn * mfn)
243 {
244         /*
245          * set defaults
246          */
247         q->nr_requests = BLKDEV_MAX_RQ;
248         blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
249         blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
250         q->make_request_fn = mfn;
251         q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
252         q->backing_dev_info.state = 0;
253         q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
254         blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
255         blk_queue_hardsect_size(q, 512);
256         blk_queue_dma_alignment(q, 511);
257         blk_queue_congestion_threshold(q);
258         q->nr_batching = BLK_BATCH_REQ;
259
260         q->unplug_thresh = 4;           /* hmm */
261         q->unplug_delay = (3 * HZ) / 1000;      /* 3 milliseconds */
262         if (q->unplug_delay == 0)
263                 q->unplug_delay = 1;
264
265         INIT_WORK(&q->unplug_work, blk_unplug_work, q);
266
267         q->unplug_timer.function = blk_unplug_timeout;
268         q->unplug_timer.data = (unsigned long)q;
269
270         /*
271          * by default assume old behaviour and bounce for any highmem page
272          */
273         blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
274
275         blk_queue_activity_fn(q, NULL, NULL);
276 }
277
278 EXPORT_SYMBOL(blk_queue_make_request);
279
280 static inline void rq_init(request_queue_t *q, struct request *rq)
281 {
282         INIT_LIST_HEAD(&rq->queuelist);
283         INIT_LIST_HEAD(&rq->donelist);
284
285         rq->errors = 0;
286         rq->rq_status = RQ_ACTIVE;
287         rq->bio = rq->biotail = NULL;
288         rq->ioprio = 0;
289         rq->buffer = NULL;
290         rq->ref_count = 1;
291         rq->q = q;
292         rq->waiting = NULL;
293         rq->special = NULL;
294         rq->data_len = 0;
295         rq->data = NULL;
296         rq->nr_phys_segments = 0;
297         rq->sense = NULL;
298         rq->end_io = NULL;
299         rq->end_io_data = NULL;
300         rq->completion_data = NULL;
301 }
302
303 /**
304  * blk_queue_ordered - does this queue support ordered writes
305  * @q:        the request queue
306  * @ordered:  one of QUEUE_ORDERED_*
307  * @prepare_flush_fn: rq setup helper for cache flush ordered writes
308  *
309  * Description:
310  *   For journalled file systems, doing ordered writes on a commit
311  *   block instead of explicitly doing wait_on_buffer (which is bad
312  *   for performance) can be a big win. Block drivers supporting this
313  *   feature should call this function and indicate so.
314  *
315  **/
316 int blk_queue_ordered(request_queue_t *q, unsigned ordered,
317                       prepare_flush_fn *prepare_flush_fn)
318 {
319         if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
320             prepare_flush_fn == NULL) {
321                 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
322                 return -EINVAL;
323         }
324
325         if (ordered != QUEUE_ORDERED_NONE &&
326             ordered != QUEUE_ORDERED_DRAIN &&
327             ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
328             ordered != QUEUE_ORDERED_DRAIN_FUA &&
329             ordered != QUEUE_ORDERED_TAG &&
330             ordered != QUEUE_ORDERED_TAG_FLUSH &&
331             ordered != QUEUE_ORDERED_TAG_FUA) {
332                 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
333                 return -EINVAL;
334         }
335
336         q->ordered = ordered;
337         q->next_ordered = ordered;
338         q->prepare_flush_fn = prepare_flush_fn;
339
340         return 0;
341 }
342
343 EXPORT_SYMBOL(blk_queue_ordered);
344
345 /**
346  * blk_queue_issue_flush_fn - set function for issuing a flush
347  * @q:     the request queue
348  * @iff:   the function to be called issuing the flush
349  *
350  * Description:
351  *   If a driver supports issuing a flush command, the support is notified
352  *   to the block layer by defining it through this call.
353  *
354  **/
355 void blk_queue_issue_flush_fn(request_queue_t *q, issue_flush_fn *iff)
356 {
357         q->issue_flush_fn = iff;
358 }
359
360 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
361
362 /*
363  * Cache flushing for ordered writes handling
364  */
365 inline unsigned blk_ordered_cur_seq(request_queue_t *q)
366 {
367         if (!q->ordseq)
368                 return 0;
369         return 1 << ffz(q->ordseq);
370 }
371
372 unsigned blk_ordered_req_seq(struct request *rq)
373 {
374         request_queue_t *q = rq->q;
375
376         BUG_ON(q->ordseq == 0);
377
378         if (rq == &q->pre_flush_rq)
379                 return QUEUE_ORDSEQ_PREFLUSH;
380         if (rq == &q->bar_rq)
381                 return QUEUE_ORDSEQ_BAR;
382         if (rq == &q->post_flush_rq)
383                 return QUEUE_ORDSEQ_POSTFLUSH;
384
385         if ((rq->flags & REQ_ORDERED_COLOR) ==
386             (q->orig_bar_rq->flags & REQ_ORDERED_COLOR))
387                 return QUEUE_ORDSEQ_DRAIN;
388         else
389                 return QUEUE_ORDSEQ_DONE;
390 }
391
392 void blk_ordered_complete_seq(request_queue_t *q, unsigned seq, int error)
393 {
394         struct request *rq;
395         int uptodate;
396
397         if (error && !q->orderr)
398                 q->orderr = error;
399
400         BUG_ON(q->ordseq & seq);
401         q->ordseq |= seq;
402
403         if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
404                 return;
405
406         /*
407          * Okay, sequence complete.
408          */
409         rq = q->orig_bar_rq;
410         uptodate = q->orderr ? q->orderr : 1;
411
412         q->ordseq = 0;
413
414         end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
415         end_that_request_last(rq, uptodate);
416 }
417
418 static void pre_flush_end_io(struct request *rq, int error)
419 {
420         elv_completed_request(rq->q, rq);
421         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
422 }
423
424 static void bar_end_io(struct request *rq, int error)
425 {
426         elv_completed_request(rq->q, rq);
427         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
428 }
429
430 static void post_flush_end_io(struct request *rq, int error)
431 {
432         elv_completed_request(rq->q, rq);
433         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
434 }
435
436 static void queue_flush(request_queue_t *q, unsigned which)
437 {
438         struct request *rq;
439         rq_end_io_fn *end_io;
440
441         if (which == QUEUE_ORDERED_PREFLUSH) {
442                 rq = &q->pre_flush_rq;
443                 end_io = pre_flush_end_io;
444         } else {
445                 rq = &q->post_flush_rq;
446                 end_io = post_flush_end_io;
447         }
448
449         rq_init(q, rq);
450         rq->flags = REQ_HARDBARRIER;
451         rq->elevator_private = NULL;
452         rq->rq_disk = q->bar_rq.rq_disk;
453         rq->rl = NULL;
454         rq->end_io = end_io;
455         q->prepare_flush_fn(q, rq);
456
457         elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
458 }
459
460 static inline struct request *start_ordered(request_queue_t *q,
461                                             struct request *rq)
462 {
463         q->bi_size = 0;
464         q->orderr = 0;
465         q->ordered = q->next_ordered;
466         q->ordseq |= QUEUE_ORDSEQ_STARTED;
467
468         /*
469          * Prep proxy barrier request.
470          */
471         blkdev_dequeue_request(rq);
472         q->orig_bar_rq = rq;
473         rq = &q->bar_rq;
474         rq_init(q, rq);
475         rq->flags = bio_data_dir(q->orig_bar_rq->bio);
476         rq->flags |= q->ordered & QUEUE_ORDERED_FUA ? REQ_FUA : 0;
477         rq->elevator_private = NULL;
478         rq->rl = NULL;
479         init_request_from_bio(rq, q->orig_bar_rq->bio);
480         rq->end_io = bar_end_io;
481
482         /*
483          * Queue ordered sequence.  As we stack them at the head, we
484          * need to queue in reverse order.  Note that we rely on that
485          * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
486          * request gets inbetween ordered sequence.
487          */
488         if (q->ordered & QUEUE_ORDERED_POSTFLUSH)
489                 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
490         else
491                 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
492
493         elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
494
495         if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
496                 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
497                 rq = &q->pre_flush_rq;
498         } else
499                 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
500
501         if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
502                 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
503         else
504                 rq = NULL;
505
506         return rq;
507 }
508
509 int blk_do_ordered(request_queue_t *q, struct request **rqp)
510 {
511         struct request *rq = *rqp;
512         int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
513
514         if (!q->ordseq) {
515                 if (!is_barrier)
516                         return 1;
517
518                 if (q->next_ordered != QUEUE_ORDERED_NONE) {
519                         *rqp = start_ordered(q, rq);
520                         return 1;
521                 } else {
522                         /*
523                          * This can happen when the queue switches to
524                          * ORDERED_NONE while this request is on it.
525                          */
526                         blkdev_dequeue_request(rq);
527                         end_that_request_first(rq, -EOPNOTSUPP,
528                                                rq->hard_nr_sectors);
529                         end_that_request_last(rq, -EOPNOTSUPP);
530                         *rqp = NULL;
531                         return 0;
532                 }
533         }
534
535         /*
536          * Ordered sequence in progress
537          */
538
539         /* Special requests are not subject to ordering rules. */
540         if (!blk_fs_request(rq) &&
541             rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
542                 return 1;
543
544         if (q->ordered & QUEUE_ORDERED_TAG) {
545                 /* Ordered by tag.  Blocking the next barrier is enough. */
546                 if (is_barrier && rq != &q->bar_rq)
547                         *rqp = NULL;
548         } else {
549                 /* Ordered by draining.  Wait for turn. */
550                 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
551                 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
552                         *rqp = NULL;
553         }
554
555         return 1;
556 }
557
558 static int flush_dry_bio_endio(struct bio *bio, unsigned int bytes, int error)
559 {
560         request_queue_t *q = bio->bi_private;
561         struct bio_vec *bvec;
562         int i;
563
564         /*
565          * This is dry run, restore bio_sector and size.  We'll finish
566          * this request again with the original bi_end_io after an
567          * error occurs or post flush is complete.
568          */
569         q->bi_size += bytes;
570
571         if (bio->bi_size)
572                 return 1;
573
574         /* Rewind bvec's */
575         bio->bi_idx = 0;
576         bio_for_each_segment(bvec, bio, i) {
577                 bvec->bv_len += bvec->bv_offset;
578                 bvec->bv_offset = 0;
579         }
580
581         /* Reset bio */
582         set_bit(BIO_UPTODATE, &bio->bi_flags);
583         bio->bi_size = q->bi_size;
584         bio->bi_sector -= (q->bi_size >> 9);
585         q->bi_size = 0;
586
587         return 0;
588 }
589
590 static inline int ordered_bio_endio(struct request *rq, struct bio *bio,
591                                     unsigned int nbytes, int error)
592 {
593         request_queue_t *q = rq->q;
594         bio_end_io_t *endio;
595         void *private;
596
597         if (&q->bar_rq != rq)
598                 return 0;
599
600         /*
601          * Okay, this is the barrier request in progress, dry finish it.
602          */
603         if (error && !q->orderr)
604                 q->orderr = error;
605
606         endio = bio->bi_end_io;
607         private = bio->bi_private;
608         bio->bi_end_io = flush_dry_bio_endio;
609         bio->bi_private = q;
610
611         bio_endio(bio, nbytes, error);
612
613         bio->bi_end_io = endio;
614         bio->bi_private = private;
615
616         return 1;
617 }
618
619 /**
620  * blk_queue_bounce_limit - set bounce buffer limit for queue
621  * @q:  the request queue for the device
622  * @dma_addr:   bus address limit
623  *
624  * Description:
625  *    Different hardware can have different requirements as to what pages
626  *    it can do I/O directly to. A low level driver can call
627  *    blk_queue_bounce_limit to have lower memory pages allocated as bounce
628  *    buffers for doing I/O to pages residing above @page.
629  **/
630 void blk_queue_bounce_limit(request_queue_t *q, u64 dma_addr)
631 {
632         unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
633         int dma = 0;
634
635         q->bounce_gfp = GFP_NOIO;
636 #if BITS_PER_LONG == 64
637         /* Assume anything <= 4GB can be handled by IOMMU.
638            Actually some IOMMUs can handle everything, but I don't
639            know of a way to test this here. */
640         if (bounce_pfn < (0xffffffff>>PAGE_SHIFT))
641                 dma = 1;
642         q->bounce_pfn = max_low_pfn;
643 #else
644         if (bounce_pfn < blk_max_low_pfn)
645                 dma = 1;
646         q->bounce_pfn = bounce_pfn;
647 #endif
648         if (dma) {
649                 init_emergency_isa_pool();
650                 q->bounce_gfp = GFP_NOIO | GFP_DMA;
651                 q->bounce_pfn = bounce_pfn;
652         }
653 }
654
655 EXPORT_SYMBOL(blk_queue_bounce_limit);
656
657 /**
658  * blk_queue_max_sectors - set max sectors for a request for this queue
659  * @q:  the request queue for the device
660  * @max_sectors:  max sectors in the usual 512b unit
661  *
662  * Description:
663  *    Enables a low level driver to set an upper limit on the size of
664  *    received requests.
665  **/
666 void blk_queue_max_sectors(request_queue_t *q, unsigned int max_sectors)
667 {
668         if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
669                 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
670                 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
671         }
672
673         if (BLK_DEF_MAX_SECTORS > max_sectors)
674                 q->max_hw_sectors = q->max_sectors = max_sectors;
675         else {
676                 q->max_sectors = BLK_DEF_MAX_SECTORS;
677                 q->max_hw_sectors = max_sectors;
678         }
679 }
680
681 EXPORT_SYMBOL(blk_queue_max_sectors);
682
683 /**
684  * blk_queue_max_phys_segments - set max phys segments for a request for this queue
685  * @q:  the request queue for the device
686  * @max_segments:  max number of segments
687  *
688  * Description:
689  *    Enables a low level driver to set an upper limit on the number of
690  *    physical data segments in a request.  This would be the largest sized
691  *    scatter list the driver could handle.
692  **/
693 void blk_queue_max_phys_segments(request_queue_t *q, unsigned short max_segments)
694 {
695         if (!max_segments) {
696                 max_segments = 1;
697                 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
698         }
699
700         q->max_phys_segments = max_segments;
701 }
702
703 EXPORT_SYMBOL(blk_queue_max_phys_segments);
704
705 /**
706  * blk_queue_max_hw_segments - set max hw segments for a request for this queue
707  * @q:  the request queue for the device
708  * @max_segments:  max number of segments
709  *
710  * Description:
711  *    Enables a low level driver to set an upper limit on the number of
712  *    hw data segments in a request.  This would be the largest number of
713  *    address/length pairs the host adapter can actually give as once
714  *    to the device.
715  **/
716 void blk_queue_max_hw_segments(request_queue_t *q, unsigned short max_segments)
717 {
718         if (!max_segments) {
719                 max_segments = 1;
720                 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
721         }
722
723         q->max_hw_segments = max_segments;
724 }
725
726 EXPORT_SYMBOL(blk_queue_max_hw_segments);
727
728 /**
729  * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
730  * @q:  the request queue for the device
731  * @max_size:  max size of segment in bytes
732  *
733  * Description:
734  *    Enables a low level driver to set an upper limit on the size of a
735  *    coalesced segment
736  **/
737 void blk_queue_max_segment_size(request_queue_t *q, unsigned int max_size)
738 {
739         if (max_size < PAGE_CACHE_SIZE) {
740                 max_size = PAGE_CACHE_SIZE;
741                 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
742         }
743
744         q->max_segment_size = max_size;
745 }
746
747 EXPORT_SYMBOL(blk_queue_max_segment_size);
748
749 /**
750  * blk_queue_hardsect_size - set hardware sector size for the queue
751  * @q:  the request queue for the device
752  * @size:  the hardware sector size, in bytes
753  *
754  * Description:
755  *   This should typically be set to the lowest possible sector size
756  *   that the hardware can operate on (possible without reverting to
757  *   even internal read-modify-write operations). Usually the default
758  *   of 512 covers most hardware.
759  **/
760 void blk_queue_hardsect_size(request_queue_t *q, unsigned short size)
761 {
762         q->hardsect_size = size;
763 }
764
765 EXPORT_SYMBOL(blk_queue_hardsect_size);
766
767 /*
768  * Returns the minimum that is _not_ zero, unless both are zero.
769  */
770 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
771
772 /**
773  * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
774  * @t:  the stacking driver (top)
775  * @b:  the underlying device (bottom)
776  **/
777 void blk_queue_stack_limits(request_queue_t *t, request_queue_t *b)
778 {
779         /* zero is "infinity" */
780         t->max_sectors = min_not_zero(t->max_sectors,b->max_sectors);
781         t->max_hw_sectors = min_not_zero(t->max_hw_sectors,b->max_hw_sectors);
782
783         t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
784         t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
785         t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
786         t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
787 }
788
789 EXPORT_SYMBOL(blk_queue_stack_limits);
790
791 /**
792  * blk_queue_segment_boundary - set boundary rules for segment merging
793  * @q:  the request queue for the device
794  * @mask:  the memory boundary mask
795  **/
796 void blk_queue_segment_boundary(request_queue_t *q, unsigned long mask)
797 {
798         if (mask < PAGE_CACHE_SIZE - 1) {
799                 mask = PAGE_CACHE_SIZE - 1;
800                 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
801         }
802
803         q->seg_boundary_mask = mask;
804 }
805
806 EXPORT_SYMBOL(blk_queue_segment_boundary);
807
808 /**
809  * blk_queue_dma_alignment - set dma length and memory alignment
810  * @q:     the request queue for the device
811  * @mask:  alignment mask
812  *
813  * description:
814  *    set required memory and length aligment for direct dma transactions.
815  *    this is used when buiding direct io requests for the queue.
816  *
817  **/
818 void blk_queue_dma_alignment(request_queue_t *q, int mask)
819 {
820         q->dma_alignment = mask;
821 }
822
823 EXPORT_SYMBOL(blk_queue_dma_alignment);
824
825 /**
826  * blk_queue_find_tag - find a request by its tag and queue
827  * @q:   The request queue for the device
828  * @tag: The tag of the request
829  *
830  * Notes:
831  *    Should be used when a device returns a tag and you want to match
832  *    it with a request.
833  *
834  *    no locks need be held.
835  **/
836 struct request *blk_queue_find_tag(request_queue_t *q, int tag)
837 {
838         struct blk_queue_tag *bqt = q->queue_tags;
839
840         if (unlikely(bqt == NULL || tag >= bqt->real_max_depth))
841                 return NULL;
842
843         return bqt->tag_index[tag];
844 }
845
846 EXPORT_SYMBOL(blk_queue_find_tag);
847
848 /**
849  * __blk_queue_free_tags - release tag maintenance info
850  * @q:  the request queue for the device
851  *
852  *  Notes:
853  *    blk_cleanup_queue() will take care of calling this function, if tagging
854  *    has been used. So there's no need to call this directly.
855  **/
856 static void __blk_queue_free_tags(request_queue_t *q)
857 {
858         struct blk_queue_tag *bqt = q->queue_tags;
859
860         if (!bqt)
861                 return;
862
863         if (atomic_dec_and_test(&bqt->refcnt)) {
864                 BUG_ON(bqt->busy);
865                 BUG_ON(!list_empty(&bqt->busy_list));
866
867                 kfree(bqt->tag_index);
868                 bqt->tag_index = NULL;
869
870                 kfree(bqt->tag_map);
871                 bqt->tag_map = NULL;
872
873                 kfree(bqt);
874         }
875
876         q->queue_tags = NULL;
877         q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
878 }
879
880 /**
881  * blk_queue_free_tags - release tag maintenance info
882  * @q:  the request queue for the device
883  *
884  *  Notes:
885  *      This is used to disabled tagged queuing to a device, yet leave
886  *      queue in function.
887  **/
888 void blk_queue_free_tags(request_queue_t *q)
889 {
890         clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
891 }
892
893 EXPORT_SYMBOL(blk_queue_free_tags);
894
895 static int
896 init_tag_map(request_queue_t *q, struct blk_queue_tag *tags, int depth)
897 {
898         struct request **tag_index;
899         unsigned long *tag_map;
900         int nr_ulongs;
901
902         if (depth > q->nr_requests * 2) {
903                 depth = q->nr_requests * 2;
904                 printk(KERN_ERR "%s: adjusted depth to %d\n",
905                                 __FUNCTION__, depth);
906         }
907
908         tag_index = kmalloc(depth * sizeof(struct request *), GFP_ATOMIC);
909         if (!tag_index)
910                 goto fail;
911
912         nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
913         tag_map = kmalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
914         if (!tag_map)
915                 goto fail;
916
917         memset(tag_index, 0, depth * sizeof(struct request *));
918         memset(tag_map, 0, nr_ulongs * sizeof(unsigned long));
919         tags->real_max_depth = depth;
920         tags->max_depth = depth;
921         tags->tag_index = tag_index;
922         tags->tag_map = tag_map;
923
924         return 0;
925 fail:
926         kfree(tag_index);
927         return -ENOMEM;
928 }
929
930 /**
931  * blk_queue_init_tags - initialize the queue tag info
932  * @q:  the request queue for the device
933  * @depth:  the maximum queue depth supported
934  * @tags: the tag to use
935  **/
936 int blk_queue_init_tags(request_queue_t *q, int depth,
937                         struct blk_queue_tag *tags)
938 {
939         int rc;
940
941         BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
942
943         if (!tags && !q->queue_tags) {
944                 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
945                 if (!tags)
946                         goto fail;
947
948                 if (init_tag_map(q, tags, depth))
949                         goto fail;
950
951                 INIT_LIST_HEAD(&tags->busy_list);
952                 tags->busy = 0;
953                 atomic_set(&tags->refcnt, 1);
954         } else if (q->queue_tags) {
955                 if ((rc = blk_queue_resize_tags(q, depth)))
956                         return rc;
957                 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
958                 return 0;
959         } else
960                 atomic_inc(&tags->refcnt);
961
962         /*
963          * assign it, all done
964          */
965         q->queue_tags = tags;
966         q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
967         return 0;
968 fail:
969         kfree(tags);
970         return -ENOMEM;
971 }
972
973 EXPORT_SYMBOL(blk_queue_init_tags);
974
975 /**
976  * blk_queue_resize_tags - change the queueing depth
977  * @q:  the request queue for the device
978  * @new_depth: the new max command queueing depth
979  *
980  *  Notes:
981  *    Must be called with the queue lock held.
982  **/
983 int blk_queue_resize_tags(request_queue_t *q, int new_depth)
984 {
985         struct blk_queue_tag *bqt = q->queue_tags;
986         struct request **tag_index;
987         unsigned long *tag_map;
988         int max_depth, nr_ulongs;
989
990         if (!bqt)
991                 return -ENXIO;
992
993         /*
994          * if we already have large enough real_max_depth.  just
995          * adjust max_depth.  *NOTE* as requests with tag value
996          * between new_depth and real_max_depth can be in-flight, tag
997          * map can not be shrunk blindly here.
998          */
999         if (new_depth <= bqt->real_max_depth) {
1000                 bqt->max_depth = new_depth;
1001                 return 0;
1002         }
1003
1004         /*
1005          * save the old state info, so we can copy it back
1006          */
1007         tag_index = bqt->tag_index;
1008         tag_map = bqt->tag_map;
1009         max_depth = bqt->real_max_depth;
1010
1011         if (init_tag_map(q, bqt, new_depth))
1012                 return -ENOMEM;
1013
1014         memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
1015         nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
1016         memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
1017
1018         kfree(tag_index);
1019         kfree(tag_map);
1020         return 0;
1021 }
1022
1023 EXPORT_SYMBOL(blk_queue_resize_tags);
1024
1025 /**
1026  * blk_queue_end_tag - end tag operations for a request
1027  * @q:  the request queue for the device
1028  * @rq: the request that has completed
1029  *
1030  *  Description:
1031  *    Typically called when end_that_request_first() returns 0, meaning
1032  *    all transfers have been done for a request. It's important to call
1033  *    this function before end_that_request_last(), as that will put the
1034  *    request back on the free list thus corrupting the internal tag list.
1035  *
1036  *  Notes:
1037  *   queue lock must be held.
1038  **/
1039 void blk_queue_end_tag(request_queue_t *q, struct request *rq)
1040 {
1041         struct blk_queue_tag *bqt = q->queue_tags;
1042         int tag = rq->tag;
1043
1044         BUG_ON(tag == -1);
1045
1046         if (unlikely(tag >= bqt->real_max_depth))
1047                 /*
1048                  * This can happen after tag depth has been reduced.
1049                  * FIXME: how about a warning or info message here?
1050                  */
1051                 return;
1052
1053         if (unlikely(!__test_and_clear_bit(tag, bqt->tag_map))) {
1054                 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
1055                        __FUNCTION__, tag);
1056                 return;
1057         }
1058
1059         list_del_init(&rq->queuelist);
1060         rq->flags &= ~REQ_QUEUED;
1061         rq->tag = -1;
1062
1063         if (unlikely(bqt->tag_index[tag] == NULL))
1064                 printk(KERN_ERR "%s: tag %d is missing\n",
1065                        __FUNCTION__, tag);
1066
1067         bqt->tag_index[tag] = NULL;
1068         bqt->busy--;
1069 }
1070
1071 EXPORT_SYMBOL(blk_queue_end_tag);
1072
1073 /**
1074  * blk_queue_start_tag - find a free tag and assign it
1075  * @q:  the request queue for the device
1076  * @rq:  the block request that needs tagging
1077  *
1078  *  Description:
1079  *    This can either be used as a stand-alone helper, or possibly be
1080  *    assigned as the queue &prep_rq_fn (in which case &struct request
1081  *    automagically gets a tag assigned). Note that this function
1082  *    assumes that any type of request can be queued! if this is not
1083  *    true for your device, you must check the request type before
1084  *    calling this function.  The request will also be removed from
1085  *    the request queue, so it's the drivers responsibility to readd
1086  *    it if it should need to be restarted for some reason.
1087  *
1088  *  Notes:
1089  *   queue lock must be held.
1090  **/
1091 int blk_queue_start_tag(request_queue_t *q, struct request *rq)
1092 {
1093         struct blk_queue_tag *bqt = q->queue_tags;
1094         int tag;
1095
1096         if (unlikely((rq->flags & REQ_QUEUED))) {
1097                 printk(KERN_ERR 
1098                        "%s: request %p for device [%s] already tagged %d",
1099                        __FUNCTION__, rq,
1100                        rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
1101                 BUG();
1102         }
1103
1104         tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
1105         if (tag >= bqt->max_depth)
1106                 return 1;
1107
1108         __set_bit(tag, bqt->tag_map);
1109
1110         rq->flags |= REQ_QUEUED;
1111         rq->tag = tag;
1112         bqt->tag_index[tag] = rq;
1113         blkdev_dequeue_request(rq);
1114         list_add(&rq->queuelist, &bqt->busy_list);
1115         bqt->busy++;
1116         return 0;
1117 }
1118
1119 EXPORT_SYMBOL(blk_queue_start_tag);
1120
1121 /**
1122  * blk_queue_invalidate_tags - invalidate all pending tags
1123  * @q:  the request queue for the device
1124  *
1125  *  Description:
1126  *   Hardware conditions may dictate a need to stop all pending requests.
1127  *   In this case, we will safely clear the block side of the tag queue and
1128  *   readd all requests to the request queue in the right order.
1129  *
1130  *  Notes:
1131  *   queue lock must be held.
1132  **/
1133 void blk_queue_invalidate_tags(request_queue_t *q)
1134 {
1135         struct blk_queue_tag *bqt = q->queue_tags;
1136         struct list_head *tmp, *n;
1137         struct request *rq;
1138
1139         list_for_each_safe(tmp, n, &bqt->busy_list) {
1140                 rq = list_entry_rq(tmp);
1141
1142                 if (rq->tag == -1) {
1143                         printk(KERN_ERR
1144                                "%s: bad tag found on list\n", __FUNCTION__);
1145                         list_del_init(&rq->queuelist);
1146                         rq->flags &= ~REQ_QUEUED;
1147                 } else
1148                         blk_queue_end_tag(q, rq);
1149
1150                 rq->flags &= ~REQ_STARTED;
1151                 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1152         }
1153 }
1154
1155 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1156
1157 static const char * const rq_flags[] = {
1158         "REQ_RW",
1159         "REQ_FAILFAST",
1160         "REQ_SORTED",
1161         "REQ_SOFTBARRIER",
1162         "REQ_HARDBARRIER",
1163         "REQ_FUA",
1164         "REQ_CMD",
1165         "REQ_NOMERGE",
1166         "REQ_STARTED",
1167         "REQ_DONTPREP",
1168         "REQ_QUEUED",
1169         "REQ_ELVPRIV",
1170         "REQ_PC",
1171         "REQ_BLOCK_PC",
1172         "REQ_SENSE",
1173         "REQ_FAILED",
1174         "REQ_QUIET",
1175         "REQ_SPECIAL",
1176         "REQ_DRIVE_CMD",
1177         "REQ_DRIVE_TASK",
1178         "REQ_DRIVE_TASKFILE",
1179         "REQ_PREEMPT",
1180         "REQ_PM_SUSPEND",
1181         "REQ_PM_RESUME",
1182         "REQ_PM_SHUTDOWN",
1183         "REQ_ORDERED_COLOR",
1184 };
1185
1186 void blk_dump_rq_flags(struct request *rq, char *msg)
1187 {
1188         int bit;
1189
1190         printk("%s: dev %s: flags = ", msg,
1191                 rq->rq_disk ? rq->rq_disk->disk_name : "?");
1192         bit = 0;
1193         do {
1194                 if (rq->flags & (1 << bit))
1195                         printk("%s ", rq_flags[bit]);
1196                 bit++;
1197         } while (bit < __REQ_NR_BITS);
1198
1199         printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1200                                                        rq->nr_sectors,
1201                                                        rq->current_nr_sectors);
1202         printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1203
1204         if (rq->flags & (REQ_BLOCK_PC | REQ_PC)) {
1205                 printk("cdb: ");
1206                 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1207                         printk("%02x ", rq->cmd[bit]);
1208                 printk("\n");
1209         }
1210 }
1211
1212 EXPORT_SYMBOL(blk_dump_rq_flags);
1213
1214 void blk_recount_segments(request_queue_t *q, struct bio *bio)
1215 {
1216         struct bio_vec *bv, *bvprv = NULL;
1217         int i, nr_phys_segs, nr_hw_segs, seg_size, hw_seg_size, cluster;
1218         int high, highprv = 1;
1219
1220         if (unlikely(!bio->bi_io_vec))
1221                 return;
1222
1223         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1224         hw_seg_size = seg_size = nr_phys_segs = nr_hw_segs = 0;
1225         bio_for_each_segment(bv, bio, i) {
1226                 /*
1227                  * the trick here is making sure that a high page is never
1228                  * considered part of another segment, since that might
1229                  * change with the bounce page.
1230                  */
1231                 high = page_to_pfn(bv->bv_page) >= q->bounce_pfn;
1232                 if (high || highprv)
1233                         goto new_hw_segment;
1234                 if (cluster) {
1235                         if (seg_size + bv->bv_len > q->max_segment_size)
1236                                 goto new_segment;
1237                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1238                                 goto new_segment;
1239                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1240                                 goto new_segment;
1241                         if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1242                                 goto new_hw_segment;
1243
1244                         seg_size += bv->bv_len;
1245                         hw_seg_size += bv->bv_len;
1246                         bvprv = bv;
1247                         continue;
1248                 }
1249 new_segment:
1250                 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1251                     !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len)) {
1252                         hw_seg_size += bv->bv_len;
1253                 } else {
1254 new_hw_segment:
1255                         if (hw_seg_size > bio->bi_hw_front_size)
1256                                 bio->bi_hw_front_size = hw_seg_size;
1257                         hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1258                         nr_hw_segs++;
1259                 }
1260
1261                 nr_phys_segs++;
1262                 bvprv = bv;
1263                 seg_size = bv->bv_len;
1264                 highprv = high;
1265         }
1266         if (hw_seg_size > bio->bi_hw_back_size)
1267                 bio->bi_hw_back_size = hw_seg_size;
1268         if (nr_hw_segs == 1 && hw_seg_size > bio->bi_hw_front_size)
1269                 bio->bi_hw_front_size = hw_seg_size;
1270         bio->bi_phys_segments = nr_phys_segs;
1271         bio->bi_hw_segments = nr_hw_segs;
1272         bio->bi_flags |= (1 << BIO_SEG_VALID);
1273 }
1274
1275
1276 static int blk_phys_contig_segment(request_queue_t *q, struct bio *bio,
1277                                    struct bio *nxt)
1278 {
1279         if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1280                 return 0;
1281
1282         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1283                 return 0;
1284         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1285                 return 0;
1286
1287         /*
1288          * bio and nxt are contigous in memory, check if the queue allows
1289          * these two to be merged into one
1290          */
1291         if (BIO_SEG_BOUNDARY(q, bio, nxt))
1292                 return 1;
1293
1294         return 0;
1295 }
1296
1297 static int blk_hw_contig_segment(request_queue_t *q, struct bio *bio,
1298                                  struct bio *nxt)
1299 {
1300         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1301                 blk_recount_segments(q, bio);
1302         if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1303                 blk_recount_segments(q, nxt);
1304         if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1305             BIOVEC_VIRT_OVERSIZE(bio->bi_hw_front_size + bio->bi_hw_back_size))
1306                 return 0;
1307         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1308                 return 0;
1309
1310         return 1;
1311 }
1312
1313 /*
1314  * map a request to scatterlist, return number of sg entries setup. Caller
1315  * must make sure sg can hold rq->nr_phys_segments entries
1316  */
1317 int blk_rq_map_sg(request_queue_t *q, struct request *rq, struct scatterlist *sg)
1318 {
1319         struct bio_vec *bvec, *bvprv;
1320         struct bio *bio;
1321         int nsegs, i, cluster;
1322
1323         nsegs = 0;
1324         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1325
1326         /*
1327          * for each bio in rq
1328          */
1329         bvprv = NULL;
1330         rq_for_each_bio(bio, rq) {
1331                 /*
1332                  * for each segment in bio
1333                  */
1334                 bio_for_each_segment(bvec, bio, i) {
1335                         int nbytes = bvec->bv_len;
1336
1337                         if (bvprv && cluster) {
1338                                 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1339                                         goto new_segment;
1340
1341                                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1342                                         goto new_segment;
1343                                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1344                                         goto new_segment;
1345
1346                                 sg[nsegs - 1].length += nbytes;
1347                         } else {
1348 new_segment:
1349                                 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1350                                 sg[nsegs].page = bvec->bv_page;
1351                                 sg[nsegs].length = nbytes;
1352                                 sg[nsegs].offset = bvec->bv_offset;
1353
1354                                 nsegs++;
1355                         }
1356                         bvprv = bvec;
1357                 } /* segments in bio */
1358         } /* bios in rq */
1359
1360         return nsegs;
1361 }
1362
1363 EXPORT_SYMBOL(blk_rq_map_sg);
1364
1365 /*
1366  * the standard queue merge functions, can be overridden with device
1367  * specific ones if so desired
1368  */
1369
1370 static inline int ll_new_mergeable(request_queue_t *q,
1371                                    struct request *req,
1372                                    struct bio *bio)
1373 {
1374         int nr_phys_segs = bio_phys_segments(q, bio);
1375
1376         if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1377                 req->flags |= REQ_NOMERGE;
1378                 if (req == q->last_merge)
1379                         q->last_merge = NULL;
1380                 return 0;
1381         }
1382
1383         /*
1384          * A hw segment is just getting larger, bump just the phys
1385          * counter.
1386          */
1387         req->nr_phys_segments += nr_phys_segs;
1388         return 1;
1389 }
1390
1391 static inline int ll_new_hw_segment(request_queue_t *q,
1392                                     struct request *req,
1393                                     struct bio *bio)
1394 {
1395         int nr_hw_segs = bio_hw_segments(q, bio);
1396         int nr_phys_segs = bio_phys_segments(q, bio);
1397
1398         if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1399             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1400                 req->flags |= REQ_NOMERGE;
1401                 if (req == q->last_merge)
1402                         q->last_merge = NULL;
1403                 return 0;
1404         }
1405
1406         /*
1407          * This will form the start of a new hw segment.  Bump both
1408          * counters.
1409          */
1410         req->nr_hw_segments += nr_hw_segs;
1411         req->nr_phys_segments += nr_phys_segs;
1412         return 1;
1413 }
1414
1415 static int ll_back_merge_fn(request_queue_t *q, struct request *req, 
1416                             struct bio *bio)
1417 {
1418         unsigned short max_sectors;
1419         int len;
1420
1421         if (unlikely(blk_pc_request(req)))
1422                 max_sectors = q->max_hw_sectors;
1423         else
1424                 max_sectors = q->max_sectors;
1425
1426         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1427                 req->flags |= REQ_NOMERGE;
1428                 if (req == q->last_merge)
1429                         q->last_merge = NULL;
1430                 return 0;
1431         }
1432         if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1433                 blk_recount_segments(q, req->biotail);
1434         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1435                 blk_recount_segments(q, bio);
1436         len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1437         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1438             !BIOVEC_VIRT_OVERSIZE(len)) {
1439                 int mergeable =  ll_new_mergeable(q, req, bio);
1440
1441                 if (mergeable) {
1442                         if (req->nr_hw_segments == 1)
1443                                 req->bio->bi_hw_front_size = len;
1444                         if (bio->bi_hw_segments == 1)
1445                                 bio->bi_hw_back_size = len;
1446                 }
1447                 return mergeable;
1448         }
1449
1450         return ll_new_hw_segment(q, req, bio);
1451 }
1452
1453 static int ll_front_merge_fn(request_queue_t *q, struct request *req, 
1454                              struct bio *bio)
1455 {
1456         unsigned short max_sectors;
1457         int len;
1458
1459         if (unlikely(blk_pc_request(req)))
1460                 max_sectors = q->max_hw_sectors;
1461         else
1462                 max_sectors = q->max_sectors;
1463
1464
1465         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1466                 req->flags |= REQ_NOMERGE;
1467                 if (req == q->last_merge)
1468                         q->last_merge = NULL;
1469                 return 0;
1470         }
1471         len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1472         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1473                 blk_recount_segments(q, bio);
1474         if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1475                 blk_recount_segments(q, req->bio);
1476         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1477             !BIOVEC_VIRT_OVERSIZE(len)) {
1478                 int mergeable =  ll_new_mergeable(q, req, bio);
1479
1480                 if (mergeable) {
1481                         if (bio->bi_hw_segments == 1)
1482                                 bio->bi_hw_front_size = len;
1483                         if (req->nr_hw_segments == 1)
1484                                 req->biotail->bi_hw_back_size = len;
1485                 }
1486                 return mergeable;
1487         }
1488
1489         return ll_new_hw_segment(q, req, bio);
1490 }
1491
1492 static int ll_merge_requests_fn(request_queue_t *q, struct request *req,
1493                                 struct request *next)
1494 {
1495         int total_phys_segments;
1496         int total_hw_segments;
1497
1498         /*
1499          * First check if the either of the requests are re-queued
1500          * requests.  Can't merge them if they are.
1501          */
1502         if (req->special || next->special)
1503                 return 0;
1504
1505         /*
1506          * Will it become too large?
1507          */
1508         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1509                 return 0;
1510
1511         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1512         if (blk_phys_contig_segment(q, req->biotail, next->bio))
1513                 total_phys_segments--;
1514
1515         if (total_phys_segments > q->max_phys_segments)
1516                 return 0;
1517
1518         total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1519         if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1520                 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1521                 /*
1522                  * propagate the combined length to the end of the requests
1523                  */
1524                 if (req->nr_hw_segments == 1)
1525                         req->bio->bi_hw_front_size = len;
1526                 if (next->nr_hw_segments == 1)
1527                         next->biotail->bi_hw_back_size = len;
1528                 total_hw_segments--;
1529         }
1530
1531         if (total_hw_segments > q->max_hw_segments)
1532                 return 0;
1533
1534         /* Merge is OK... */
1535         req->nr_phys_segments = total_phys_segments;
1536         req->nr_hw_segments = total_hw_segments;
1537         return 1;
1538 }
1539
1540 /*
1541  * "plug" the device if there are no outstanding requests: this will
1542  * force the transfer to start only after we have put all the requests
1543  * on the list.
1544  *
1545  * This is called with interrupts off and no requests on the queue and
1546  * with the queue lock held.
1547  */
1548 void blk_plug_device(request_queue_t *q)
1549 {
1550         WARN_ON(!irqs_disabled());
1551
1552         /*
1553          * don't plug a stopped queue, it must be paired with blk_start_queue()
1554          * which will restart the queueing
1555          */
1556         if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
1557                 return;
1558
1559         if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1560                 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1561 }
1562
1563 EXPORT_SYMBOL(blk_plug_device);
1564
1565 /*
1566  * remove the queue from the plugged list, if present. called with
1567  * queue lock held and interrupts disabled.
1568  */
1569 int blk_remove_plug(request_queue_t *q)
1570 {
1571         WARN_ON(!irqs_disabled());
1572
1573         if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1574                 return 0;
1575
1576         del_timer(&q->unplug_timer);
1577         return 1;
1578 }
1579
1580 EXPORT_SYMBOL(blk_remove_plug);
1581
1582 /*
1583  * remove the plug and let it rip..
1584  */
1585 void __generic_unplug_device(request_queue_t *q)
1586 {
1587         if (unlikely(test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags)))
1588                 return;
1589
1590         if (!blk_remove_plug(q))
1591                 return;
1592
1593         q->request_fn(q);
1594 }
1595 EXPORT_SYMBOL(__generic_unplug_device);
1596
1597 /**
1598  * generic_unplug_device - fire a request queue
1599  * @q:    The &request_queue_t in question
1600  *
1601  * Description:
1602  *   Linux uses plugging to build bigger requests queues before letting
1603  *   the device have at them. If a queue is plugged, the I/O scheduler
1604  *   is still adding and merging requests on the queue. Once the queue
1605  *   gets unplugged, the request_fn defined for the queue is invoked and
1606  *   transfers started.
1607  **/
1608 void generic_unplug_device(request_queue_t *q)
1609 {
1610         spin_lock_irq(q->queue_lock);
1611         __generic_unplug_device(q);
1612         spin_unlock_irq(q->queue_lock);
1613 }
1614 EXPORT_SYMBOL(generic_unplug_device);
1615
1616 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1617                                    struct page *page)
1618 {
1619         request_queue_t *q = bdi->unplug_io_data;
1620
1621         /*
1622          * devices don't necessarily have an ->unplug_fn defined
1623          */
1624         if (q->unplug_fn)
1625                 q->unplug_fn(q);
1626 }
1627
1628 static void blk_unplug_work(void *data)
1629 {
1630         request_queue_t *q = data;
1631
1632         q->unplug_fn(q);
1633 }
1634
1635 static void blk_unplug_timeout(unsigned long data)
1636 {
1637         request_queue_t *q = (request_queue_t *)data;
1638
1639         kblockd_schedule_work(&q->unplug_work);
1640 }
1641
1642 /**
1643  * blk_start_queue - restart a previously stopped queue
1644  * @q:    The &request_queue_t in question
1645  *
1646  * Description:
1647  *   blk_start_queue() will clear the stop flag on the queue, and call
1648  *   the request_fn for the queue if it was in a stopped state when
1649  *   entered. Also see blk_stop_queue(). Queue lock must be held.
1650  **/
1651 void blk_start_queue(request_queue_t *q)
1652 {
1653         clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1654
1655         /*
1656          * one level of recursion is ok and is much faster than kicking
1657          * the unplug handling
1658          */
1659         if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1660                 q->request_fn(q);
1661                 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1662         } else {
1663                 blk_plug_device(q);
1664                 kblockd_schedule_work(&q->unplug_work);
1665         }
1666 }
1667
1668 EXPORT_SYMBOL(blk_start_queue);
1669
1670 /**
1671  * blk_stop_queue - stop a queue
1672  * @q:    The &request_queue_t in question
1673  *
1674  * Description:
1675  *   The Linux block layer assumes that a block driver will consume all
1676  *   entries on the request queue when the request_fn strategy is called.
1677  *   Often this will not happen, because of hardware limitations (queue
1678  *   depth settings). If a device driver gets a 'queue full' response,
1679  *   or if it simply chooses not to queue more I/O at one point, it can
1680  *   call this function to prevent the request_fn from being called until
1681  *   the driver has signalled it's ready to go again. This happens by calling
1682  *   blk_start_queue() to restart queue operations. Queue lock must be held.
1683  **/
1684 void blk_stop_queue(request_queue_t *q)
1685 {
1686         blk_remove_plug(q);
1687         set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1688 }
1689 EXPORT_SYMBOL(blk_stop_queue);
1690
1691 /**
1692  * blk_sync_queue - cancel any pending callbacks on a queue
1693  * @q: the queue
1694  *
1695  * Description:
1696  *     The block layer may perform asynchronous callback activity
1697  *     on a queue, such as calling the unplug function after a timeout.
1698  *     A block device may call blk_sync_queue to ensure that any
1699  *     such activity is cancelled, thus allowing it to release resources
1700  *     the the callbacks might use. The caller must already have made sure
1701  *     that its ->make_request_fn will not re-add plugging prior to calling
1702  *     this function.
1703  *
1704  */
1705 void blk_sync_queue(struct request_queue *q)
1706 {
1707         del_timer_sync(&q->unplug_timer);
1708         kblockd_flush();
1709 }
1710 EXPORT_SYMBOL(blk_sync_queue);
1711
1712 /**
1713  * blk_run_queue - run a single device queue
1714  * @q:  The queue to run
1715  */
1716 void blk_run_queue(struct request_queue *q)
1717 {
1718         unsigned long flags;
1719
1720         spin_lock_irqsave(q->queue_lock, flags);
1721         blk_remove_plug(q);
1722         if (!elv_queue_empty(q))
1723                 q->request_fn(q);
1724         spin_unlock_irqrestore(q->queue_lock, flags);
1725 }
1726 EXPORT_SYMBOL(blk_run_queue);
1727
1728 /**
1729  * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1730  * @q:    the request queue to be released
1731  *
1732  * Description:
1733  *     blk_cleanup_queue is the pair to blk_init_queue() or
1734  *     blk_queue_make_request().  It should be called when a request queue is
1735  *     being released; typically when a block device is being de-registered.
1736  *     Currently, its primary task it to free all the &struct request
1737  *     structures that were allocated to the queue and the queue itself.
1738  *
1739  * Caveat:
1740  *     Hopefully the low level driver will have finished any
1741  *     outstanding requests first...
1742  **/
1743 void blk_cleanup_queue(request_queue_t * q)
1744 {
1745         struct request_list *rl = &q->rq;
1746
1747         if (!atomic_dec_and_test(&q->refcnt))
1748                 return;
1749
1750         if (q->elevator)
1751                 elevator_exit(q->elevator);
1752
1753         blk_sync_queue(q);
1754
1755         if (rl->rq_pool)
1756                 mempool_destroy(rl->rq_pool);
1757
1758         if (q->queue_tags)
1759                 __blk_queue_free_tags(q);
1760
1761         kmem_cache_free(requestq_cachep, q);
1762 }
1763
1764 EXPORT_SYMBOL(blk_cleanup_queue);
1765
1766 static int blk_init_free_list(request_queue_t *q)
1767 {
1768         struct request_list *rl = &q->rq;
1769
1770         rl->count[READ] = rl->count[WRITE] = 0;
1771         rl->starved[READ] = rl->starved[WRITE] = 0;
1772         rl->elvpriv = 0;
1773         init_waitqueue_head(&rl->wait[READ]);
1774         init_waitqueue_head(&rl->wait[WRITE]);
1775
1776         rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1777                                 mempool_free_slab, request_cachep, q->node);
1778
1779         if (!rl->rq_pool)
1780                 return -ENOMEM;
1781
1782         return 0;
1783 }
1784
1785 request_queue_t *blk_alloc_queue(gfp_t gfp_mask)
1786 {
1787         return blk_alloc_queue_node(gfp_mask, -1);
1788 }
1789 EXPORT_SYMBOL(blk_alloc_queue);
1790
1791 request_queue_t *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
1792 {
1793         request_queue_t *q;
1794
1795         q = kmem_cache_alloc_node(requestq_cachep, gfp_mask, node_id);
1796         if (!q)
1797                 return NULL;
1798
1799         memset(q, 0, sizeof(*q));
1800         init_timer(&q->unplug_timer);
1801         atomic_set(&q->refcnt, 1);
1802
1803         q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1804         q->backing_dev_info.unplug_io_data = q;
1805
1806         return q;
1807 }
1808 EXPORT_SYMBOL(blk_alloc_queue_node);
1809
1810 /**
1811  * blk_init_queue  - prepare a request queue for use with a block device
1812  * @rfn:  The function to be called to process requests that have been
1813  *        placed on the queue.
1814  * @lock: Request queue spin lock
1815  *
1816  * Description:
1817  *    If a block device wishes to use the standard request handling procedures,
1818  *    which sorts requests and coalesces adjacent requests, then it must
1819  *    call blk_init_queue().  The function @rfn will be called when there
1820  *    are requests on the queue that need to be processed.  If the device
1821  *    supports plugging, then @rfn may not be called immediately when requests
1822  *    are available on the queue, but may be called at some time later instead.
1823  *    Plugged queues are generally unplugged when a buffer belonging to one
1824  *    of the requests on the queue is needed, or due to memory pressure.
1825  *
1826  *    @rfn is not required, or even expected, to remove all requests off the
1827  *    queue, but only as many as it can handle at a time.  If it does leave
1828  *    requests on the queue, it is responsible for arranging that the requests
1829  *    get dealt with eventually.
1830  *
1831  *    The queue spin lock must be held while manipulating the requests on the
1832  *    request queue.
1833  *
1834  *    Function returns a pointer to the initialized request queue, or NULL if
1835  *    it didn't succeed.
1836  *
1837  * Note:
1838  *    blk_init_queue() must be paired with a blk_cleanup_queue() call
1839  *    when the block device is deactivated (such as at module unload).
1840  **/
1841
1842 request_queue_t *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1843 {
1844         return blk_init_queue_node(rfn, lock, -1);
1845 }
1846 EXPORT_SYMBOL(blk_init_queue);
1847
1848 request_queue_t *
1849 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1850 {
1851         request_queue_t *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1852
1853         if (!q)
1854                 return NULL;
1855
1856         q->node = node_id;
1857         if (blk_init_free_list(q))
1858                 goto out_init;
1859
1860         /*
1861          * if caller didn't supply a lock, they get per-queue locking with
1862          * our embedded lock
1863          */
1864         if (!lock) {
1865                 spin_lock_init(&q->__queue_lock);
1866                 lock = &q->__queue_lock;
1867         }
1868
1869         q->request_fn           = rfn;
1870         q->back_merge_fn        = ll_back_merge_fn;
1871         q->front_merge_fn       = ll_front_merge_fn;
1872         q->merge_requests_fn    = ll_merge_requests_fn;
1873         q->prep_rq_fn           = NULL;
1874         q->unplug_fn            = generic_unplug_device;
1875         q->queue_flags          = (1 << QUEUE_FLAG_CLUSTER);
1876         q->queue_lock           = lock;
1877
1878         blk_queue_segment_boundary(q, 0xffffffff);
1879
1880         blk_queue_make_request(q, __make_request);
1881         blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1882
1883         blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1884         blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1885
1886         /*
1887          * all done
1888          */
1889         if (!elevator_init(q, NULL)) {
1890                 blk_queue_congestion_threshold(q);
1891                 return q;
1892         }
1893
1894         blk_cleanup_queue(q);
1895 out_init:
1896         kmem_cache_free(requestq_cachep, q);
1897         return NULL;
1898 }
1899 EXPORT_SYMBOL(blk_init_queue_node);
1900
1901 int blk_get_queue(request_queue_t *q)
1902 {
1903         if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1904                 atomic_inc(&q->refcnt);
1905                 return 0;
1906         }
1907
1908         return 1;
1909 }
1910
1911 EXPORT_SYMBOL(blk_get_queue);
1912
1913 static inline void blk_free_request(request_queue_t *q, struct request *rq)
1914 {
1915         if (rq->flags & REQ_ELVPRIV)
1916                 elv_put_request(q, rq);
1917         mempool_free(rq, q->rq.rq_pool);
1918 }
1919
1920 static inline struct request *
1921 blk_alloc_request(request_queue_t *q, int rw, struct bio *bio,
1922                   int priv, gfp_t gfp_mask)
1923 {
1924         struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1925
1926         if (!rq)
1927                 return NULL;
1928
1929         /*
1930          * first three bits are identical in rq->flags and bio->bi_rw,
1931          * see bio.h and blkdev.h
1932          */
1933         rq->flags = rw;
1934
1935         if (priv) {
1936                 if (unlikely(elv_set_request(q, rq, bio, gfp_mask))) {
1937                         mempool_free(rq, q->rq.rq_pool);
1938                         return NULL;
1939                 }
1940                 rq->flags |= REQ_ELVPRIV;
1941         }
1942
1943         return rq;
1944 }
1945
1946 /*
1947  * ioc_batching returns true if the ioc is a valid batching request and
1948  * should be given priority access to a request.
1949  */
1950 static inline int ioc_batching(request_queue_t *q, struct io_context *ioc)
1951 {
1952         if (!ioc)
1953                 return 0;
1954
1955         /*
1956          * Make sure the process is able to allocate at least 1 request
1957          * even if the batch times out, otherwise we could theoretically
1958          * lose wakeups.
1959          */
1960         return ioc->nr_batch_requests == q->nr_batching ||
1961                 (ioc->nr_batch_requests > 0
1962                 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
1963 }
1964
1965 /*
1966  * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
1967  * will cause the process to be a "batcher" on all queues in the system. This
1968  * is the behaviour we want though - once it gets a wakeup it should be given
1969  * a nice run.
1970  */
1971 static void ioc_set_batching(request_queue_t *q, struct io_context *ioc)
1972 {
1973         if (!ioc || ioc_batching(q, ioc))
1974                 return;
1975
1976         ioc->nr_batch_requests = q->nr_batching;
1977         ioc->last_waited = jiffies;
1978 }
1979
1980 static void __freed_request(request_queue_t *q, int rw)
1981 {
1982         struct request_list *rl = &q->rq;
1983
1984         if (rl->count[rw] < queue_congestion_off_threshold(q))
1985                 clear_queue_congested(q, rw);
1986
1987         if (rl->count[rw] + 1 <= q->nr_requests) {
1988                 if (waitqueue_active(&rl->wait[rw]))
1989                         wake_up(&rl->wait[rw]);
1990
1991                 blk_clear_queue_full(q, rw);
1992         }
1993 }
1994
1995 /*
1996  * A request has just been released.  Account for it, update the full and
1997  * congestion status, wake up any waiters.   Called under q->queue_lock.
1998  */
1999 static void freed_request(request_queue_t *q, int rw, int priv)
2000 {
2001         struct request_list *rl = &q->rq;
2002
2003         rl->count[rw]--;
2004         if (priv)
2005                 rl->elvpriv--;
2006
2007         __freed_request(q, rw);
2008
2009         if (unlikely(rl->starved[rw ^ 1]))
2010                 __freed_request(q, rw ^ 1);
2011 }
2012
2013 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2014 /*
2015  * Get a free request, queue_lock must be held.
2016  * Returns NULL on failure, with queue_lock held.
2017  * Returns !NULL on success, with queue_lock *not held*.
2018  */
2019 static struct request *get_request(request_queue_t *q, int rw, struct bio *bio,
2020                                    gfp_t gfp_mask)
2021 {
2022         struct request *rq = NULL;
2023         struct request_list *rl = &q->rq;
2024         struct io_context *ioc = NULL;
2025         int may_queue, priv;
2026
2027         may_queue = elv_may_queue(q, rw, bio);
2028         if (may_queue == ELV_MQUEUE_NO)
2029                 goto rq_starved;
2030
2031         if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
2032                 if (rl->count[rw]+1 >= q->nr_requests) {
2033                         ioc = current_io_context(GFP_ATOMIC);
2034                         /*
2035                          * The queue will fill after this allocation, so set
2036                          * it as full, and mark this process as "batching".
2037                          * This process will be allowed to complete a batch of
2038                          * requests, others will be blocked.
2039                          */
2040                         if (!blk_queue_full(q, rw)) {
2041                                 ioc_set_batching(q, ioc);
2042                                 blk_set_queue_full(q, rw);
2043                         } else {
2044                                 if (may_queue != ELV_MQUEUE_MUST
2045                                                 && !ioc_batching(q, ioc)) {
2046                                         /*
2047                                          * The queue is full and the allocating
2048                                          * process is not a "batcher", and not
2049                                          * exempted by the IO scheduler
2050                                          */
2051                                         goto out;
2052                                 }
2053                         }
2054                 }
2055                 set_queue_congested(q, rw);
2056         }
2057
2058         /*
2059          * Only allow batching queuers to allocate up to 50% over the defined
2060          * limit of requests, otherwise we could have thousands of requests
2061          * allocated with any setting of ->nr_requests
2062          */
2063         if (rl->count[rw] >= (3 * q->nr_requests / 2))
2064                 goto out;
2065
2066         rl->count[rw]++;
2067         rl->starved[rw] = 0;
2068
2069         priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
2070         if (priv)
2071                 rl->elvpriv++;
2072
2073         spin_unlock_irq(q->queue_lock);
2074
2075         rq = blk_alloc_request(q, rw, bio, priv, gfp_mask);
2076         if (unlikely(!rq)) {
2077                 /*
2078                  * Allocation failed presumably due to memory. Undo anything
2079                  * we might have messed up.
2080                  *
2081                  * Allocating task should really be put onto the front of the
2082                  * wait queue, but this is pretty rare.
2083                  */
2084                 spin_lock_irq(q->queue_lock);
2085                 freed_request(q, rw, priv);
2086
2087                 /*
2088                  * in the very unlikely event that allocation failed and no
2089                  * requests for this direction was pending, mark us starved
2090                  * so that freeing of a request in the other direction will
2091                  * notice us. another possible fix would be to split the
2092                  * rq mempool into READ and WRITE
2093                  */
2094 rq_starved:
2095                 if (unlikely(rl->count[rw] == 0))
2096                         rl->starved[rw] = 1;
2097
2098                 goto out;
2099         }
2100
2101         /*
2102          * ioc may be NULL here, and ioc_batching will be false. That's
2103          * OK, if the queue is under the request limit then requests need
2104          * not count toward the nr_batch_requests limit. There will always
2105          * be some limit enforced by BLK_BATCH_TIME.
2106          */
2107         if (ioc_batching(q, ioc))
2108                 ioc->nr_batch_requests--;
2109         
2110         rq_init(q, rq);
2111         rq->rl = rl;
2112 out:
2113         return rq;
2114 }
2115
2116 /*
2117  * No available requests for this queue, unplug the device and wait for some
2118  * requests to become available.
2119  *
2120  * Called with q->queue_lock held, and returns with it unlocked.
2121  */
2122 static struct request *get_request_wait(request_queue_t *q, int rw,
2123                                         struct bio *bio)
2124 {
2125         struct request *rq;
2126
2127         rq = get_request(q, rw, bio, GFP_NOIO);
2128         while (!rq) {
2129                 DEFINE_WAIT(wait);
2130                 struct request_list *rl = &q->rq;
2131
2132                 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2133                                 TASK_UNINTERRUPTIBLE);
2134
2135                 rq = get_request(q, rw, bio, GFP_NOIO);
2136
2137                 if (!rq) {
2138                         struct io_context *ioc;
2139
2140                         __generic_unplug_device(q);
2141                         spin_unlock_irq(q->queue_lock);
2142                         io_schedule();
2143
2144                         /*
2145                          * After sleeping, we become a "batching" process and
2146                          * will be able to allocate at least one request, and
2147                          * up to a big batch of them for a small period time.
2148                          * See ioc_batching, ioc_set_batching
2149                          */
2150                         ioc = current_io_context(GFP_NOIO);
2151                         ioc_set_batching(q, ioc);
2152
2153                         spin_lock_irq(q->queue_lock);
2154                 }
2155                 finish_wait(&rl->wait[rw], &wait);
2156         }
2157
2158         return rq;
2159 }
2160
2161 struct request *blk_get_request(request_queue_t *q, int rw, gfp_t gfp_mask)
2162 {
2163         struct request *rq;
2164
2165         BUG_ON(rw != READ && rw != WRITE);
2166
2167         spin_lock_irq(q->queue_lock);
2168         if (gfp_mask & __GFP_WAIT) {
2169                 rq = get_request_wait(q, rw, NULL);
2170         } else {
2171                 rq = get_request(q, rw, NULL, gfp_mask);
2172                 if (!rq)
2173                         spin_unlock_irq(q->queue_lock);
2174         }
2175         /* q->queue_lock is unlocked at this point */
2176
2177         return rq;
2178 }
2179 EXPORT_SYMBOL(blk_get_request);
2180
2181 /**
2182  * blk_requeue_request - put a request back on queue
2183  * @q:          request queue where request should be inserted
2184  * @rq:         request to be inserted
2185  *
2186  * Description:
2187  *    Drivers often keep queueing requests until the hardware cannot accept
2188  *    more, when that condition happens we need to put the request back
2189  *    on the queue. Must be called with queue lock held.
2190  */
2191 void blk_requeue_request(request_queue_t *q, struct request *rq)
2192 {
2193         if (blk_rq_tagged(rq))
2194                 blk_queue_end_tag(q, rq);
2195
2196         elv_requeue_request(q, rq);
2197 }
2198
2199 EXPORT_SYMBOL(blk_requeue_request);
2200
2201 /**
2202  * blk_insert_request - insert a special request in to a request queue
2203  * @q:          request queue where request should be inserted
2204  * @rq:         request to be inserted
2205  * @at_head:    insert request at head or tail of queue
2206  * @data:       private data
2207  *
2208  * Description:
2209  *    Many block devices need to execute commands asynchronously, so they don't
2210  *    block the whole kernel from preemption during request execution.  This is
2211  *    accomplished normally by inserting aritficial requests tagged as
2212  *    REQ_SPECIAL in to the corresponding request queue, and letting them be
2213  *    scheduled for actual execution by the request queue.
2214  *
2215  *    We have the option of inserting the head or the tail of the queue.
2216  *    Typically we use the tail for new ioctls and so forth.  We use the head
2217  *    of the queue for things like a QUEUE_FULL message from a device, or a
2218  *    host that is unable to accept a particular command.
2219  */
2220 void blk_insert_request(request_queue_t *q, struct request *rq,
2221                         int at_head, void *data)
2222 {
2223         int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2224         unsigned long flags;
2225
2226         /*
2227          * tell I/O scheduler that this isn't a regular read/write (ie it
2228          * must not attempt merges on this) and that it acts as a soft
2229          * barrier
2230          */
2231         rq->flags |= REQ_SPECIAL | REQ_SOFTBARRIER;
2232
2233         rq->special = data;
2234
2235         spin_lock_irqsave(q->queue_lock, flags);
2236
2237         /*
2238          * If command is tagged, release the tag
2239          */
2240         if (blk_rq_tagged(rq))
2241                 blk_queue_end_tag(q, rq);
2242
2243         drive_stat_acct(rq, rq->nr_sectors, 1);
2244         __elv_add_request(q, rq, where, 0);
2245
2246         if (blk_queue_plugged(q))
2247                 __generic_unplug_device(q);
2248         else
2249                 q->request_fn(q);
2250         spin_unlock_irqrestore(q->queue_lock, flags);
2251 }
2252
2253 EXPORT_SYMBOL(blk_insert_request);
2254
2255 /**
2256  * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2257  * @q:          request queue where request should be inserted
2258  * @rq:         request structure to fill
2259  * @ubuf:       the user buffer
2260  * @len:        length of user data
2261  *
2262  * Description:
2263  *    Data will be mapped directly for zero copy io, if possible. Otherwise
2264  *    a kernel bounce buffer is used.
2265  *
2266  *    A matching blk_rq_unmap_user() must be issued at the end of io, while
2267  *    still in process context.
2268  *
2269  *    Note: The mapped bio may need to be bounced through blk_queue_bounce()
2270  *    before being submitted to the device, as pages mapped may be out of
2271  *    reach. It's the callers responsibility to make sure this happens. The
2272  *    original bio must be passed back in to blk_rq_unmap_user() for proper
2273  *    unmapping.
2274  */
2275 int blk_rq_map_user(request_queue_t *q, struct request *rq, void __user *ubuf,
2276                     unsigned int len)
2277 {
2278         unsigned long uaddr;
2279         struct bio *bio;
2280         int reading;
2281
2282         if (len > (q->max_hw_sectors << 9))
2283                 return -EINVAL;
2284         if (!len || !ubuf)
2285                 return -EINVAL;
2286
2287         reading = rq_data_dir(rq) == READ;
2288
2289         /*
2290          * if alignment requirement is satisfied, map in user pages for
2291          * direct dma. else, set up kernel bounce buffers
2292          */
2293         uaddr = (unsigned long) ubuf;
2294         if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2295                 bio = bio_map_user(q, NULL, uaddr, len, reading);
2296         else
2297                 bio = bio_copy_user(q, uaddr, len, reading);
2298
2299         if (!IS_ERR(bio)) {
2300                 rq->bio = rq->biotail = bio;
2301                 blk_rq_bio_prep(q, rq, bio);
2302
2303                 rq->buffer = rq->data = NULL;
2304                 rq->data_len = len;
2305                 return 0;
2306         }
2307
2308         /*
2309          * bio is the err-ptr
2310          */
2311         return PTR_ERR(bio);
2312 }
2313
2314 EXPORT_SYMBOL(blk_rq_map_user);
2315
2316 /**
2317  * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2318  * @q:          request queue where request should be inserted
2319  * @rq:         request to map data to
2320  * @iov:        pointer to the iovec
2321  * @iov_count:  number of elements in the iovec
2322  *
2323  * Description:
2324  *    Data will be mapped directly for zero copy io, if possible. Otherwise
2325  *    a kernel bounce buffer is used.
2326  *
2327  *    A matching blk_rq_unmap_user() must be issued at the end of io, while
2328  *    still in process context.
2329  *
2330  *    Note: The mapped bio may need to be bounced through blk_queue_bounce()
2331  *    before being submitted to the device, as pages mapped may be out of
2332  *    reach. It's the callers responsibility to make sure this happens. The
2333  *    original bio must be passed back in to blk_rq_unmap_user() for proper
2334  *    unmapping.
2335  */
2336 int blk_rq_map_user_iov(request_queue_t *q, struct request *rq,
2337                         struct sg_iovec *iov, int iov_count)
2338 {
2339         struct bio *bio;
2340
2341         if (!iov || iov_count <= 0)
2342                 return -EINVAL;
2343
2344         /* we don't allow misaligned data like bio_map_user() does.  If the
2345          * user is using sg, they're expected to know the alignment constraints
2346          * and respect them accordingly */
2347         bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2348         if (IS_ERR(bio))
2349                 return PTR_ERR(bio);
2350
2351         rq->bio = rq->biotail = bio;
2352         blk_rq_bio_prep(q, rq, bio);
2353         rq->buffer = rq->data = NULL;
2354         rq->data_len = bio->bi_size;
2355         return 0;
2356 }
2357
2358 EXPORT_SYMBOL(blk_rq_map_user_iov);
2359
2360 /**
2361  * blk_rq_unmap_user - unmap a request with user data
2362  * @bio:        bio to be unmapped
2363  * @ulen:       length of user buffer
2364  *
2365  * Description:
2366  *    Unmap a bio previously mapped by blk_rq_map_user().
2367  */
2368 int blk_rq_unmap_user(struct bio *bio, unsigned int ulen)
2369 {
2370         int ret = 0;
2371
2372         if (bio) {
2373                 if (bio_flagged(bio, BIO_USER_MAPPED))
2374                         bio_unmap_user(bio);
2375                 else
2376                         ret = bio_uncopy_user(bio);
2377         }
2378
2379         return 0;
2380 }
2381
2382 EXPORT_SYMBOL(blk_rq_unmap_user);
2383
2384 /**
2385  * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2386  * @q:          request queue where request should be inserted
2387  * @rq:         request to fill
2388  * @kbuf:       the kernel buffer
2389  * @len:        length of user data
2390  * @gfp_mask:   memory allocation flags
2391  */
2392 int blk_rq_map_kern(request_queue_t *q, struct request *rq, void *kbuf,
2393                     unsigned int len, gfp_t gfp_mask)
2394 {
2395         struct bio *bio;
2396
2397         if (len > (q->max_hw_sectors << 9))
2398                 return -EINVAL;
2399         if (!len || !kbuf)
2400                 return -EINVAL;
2401
2402         bio = bio_map_kern(q, kbuf, len, gfp_mask);
2403         if (IS_ERR(bio))
2404                 return PTR_ERR(bio);
2405
2406         if (rq_data_dir(rq) == WRITE)
2407                 bio->bi_rw |= (1 << BIO_RW);
2408
2409         rq->bio = rq->biotail = bio;
2410         blk_rq_bio_prep(q, rq, bio);
2411
2412         rq->buffer = rq->data = NULL;
2413         rq->data_len = len;
2414         return 0;
2415 }
2416
2417 EXPORT_SYMBOL(blk_rq_map_kern);
2418
2419 /**
2420  * blk_execute_rq_nowait - insert a request into queue for execution
2421  * @q:          queue to insert the request in
2422  * @bd_disk:    matching gendisk
2423  * @rq:         request to insert
2424  * @at_head:    insert request at head or tail of queue
2425  * @done:       I/O completion handler
2426  *
2427  * Description:
2428  *    Insert a fully prepared request at the back of the io scheduler queue
2429  *    for execution.  Don't wait for completion.
2430  */
2431 void blk_execute_rq_nowait(request_queue_t *q, struct gendisk *bd_disk,
2432                            struct request *rq, int at_head,
2433                            rq_end_io_fn *done)
2434 {
2435         int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2436
2437         rq->rq_disk = bd_disk;
2438         rq->flags |= REQ_NOMERGE;
2439         rq->end_io = done;
2440         elv_add_request(q, rq, where, 1);
2441         generic_unplug_device(q);
2442 }
2443
2444 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
2445
2446 /**
2447  * blk_execute_rq - insert a request into queue for execution
2448  * @q:          queue to insert the request in
2449  * @bd_disk:    matching gendisk
2450  * @rq:         request to insert
2451  * @at_head:    insert request at head or tail of queue
2452  *
2453  * Description:
2454  *    Insert a fully prepared request at the back of the io scheduler queue
2455  *    for execution and wait for completion.
2456  */
2457 int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
2458                    struct request *rq, int at_head)
2459 {
2460         DECLARE_COMPLETION(wait);
2461         char sense[SCSI_SENSE_BUFFERSIZE];
2462         int err = 0;
2463
2464         /*
2465          * we need an extra reference to the request, so we can look at
2466          * it after io completion
2467          */
2468         rq->ref_count++;
2469
2470         if (!rq->sense) {
2471                 memset(sense, 0, sizeof(sense));
2472                 rq->sense = sense;
2473                 rq->sense_len = 0;
2474         }
2475
2476         rq->waiting = &wait;
2477         blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
2478         wait_for_completion(&wait);
2479         rq->waiting = NULL;
2480
2481         if (rq->errors)
2482                 err = -EIO;
2483
2484         return err;
2485 }
2486
2487 EXPORT_SYMBOL(blk_execute_rq);
2488
2489 /**
2490  * blkdev_issue_flush - queue a flush
2491  * @bdev:       blockdev to issue flush for
2492  * @error_sector:       error sector
2493  *
2494  * Description:
2495  *    Issue a flush for the block device in question. Caller can supply
2496  *    room for storing the error offset in case of a flush error, if they
2497  *    wish to.  Caller must run wait_for_completion() on its own.
2498  */
2499 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2500 {
2501         request_queue_t *q;
2502
2503         if (bdev->bd_disk == NULL)
2504                 return -ENXIO;
2505
2506         q = bdev_get_queue(bdev);
2507         if (!q)
2508                 return -ENXIO;
2509         if (!q->issue_flush_fn)
2510                 return -EOPNOTSUPP;
2511
2512         return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2513 }
2514
2515 EXPORT_SYMBOL(blkdev_issue_flush);
2516
2517 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2518 {
2519         int rw = rq_data_dir(rq);
2520
2521         if (!blk_fs_request(rq) || !rq->rq_disk)
2522                 return;
2523
2524         if (!new_io) {
2525                 __disk_stat_inc(rq->rq_disk, merges[rw]);
2526         } else {
2527                 disk_round_stats(rq->rq_disk);
2528                 rq->rq_disk->in_flight++;
2529         }
2530 }
2531
2532 /*
2533  * add-request adds a request to the linked list.
2534  * queue lock is held and interrupts disabled, as we muck with the
2535  * request queue list.
2536  */
2537 static inline void add_request(request_queue_t * q, struct request * req)
2538 {
2539         drive_stat_acct(req, req->nr_sectors, 1);
2540
2541         if (q->activity_fn)
2542                 q->activity_fn(q->activity_data, rq_data_dir(req));
2543
2544         /*
2545          * elevator indicated where it wants this request to be
2546          * inserted at elevator_merge time
2547          */
2548         __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2549 }
2550  
2551 /*
2552  * disk_round_stats()   - Round off the performance stats on a struct
2553  * disk_stats.
2554  *
2555  * The average IO queue length and utilisation statistics are maintained
2556  * by observing the current state of the queue length and the amount of
2557  * time it has been in this state for.
2558  *
2559  * Normally, that accounting is done on IO completion, but that can result
2560  * in more than a second's worth of IO being accounted for within any one
2561  * second, leading to >100% utilisation.  To deal with that, we call this
2562  * function to do a round-off before returning the results when reading
2563  * /proc/diskstats.  This accounts immediately for all queue usage up to
2564  * the current jiffies and restarts the counters again.
2565  */
2566 void disk_round_stats(struct gendisk *disk)
2567 {
2568         unsigned long now = jiffies;
2569
2570         if (now == disk->stamp)
2571                 return;
2572
2573         if (disk->in_flight) {
2574                 __disk_stat_add(disk, time_in_queue,
2575                                 disk->in_flight * (now - disk->stamp));
2576                 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2577         }
2578         disk->stamp = now;
2579 }
2580
2581 EXPORT_SYMBOL_GPL(disk_round_stats);
2582
2583 /*
2584  * queue lock must be held
2585  */
2586 void __blk_put_request(request_queue_t *q, struct request *req)
2587 {
2588         struct request_list *rl = req->rl;
2589
2590         if (unlikely(!q))
2591                 return;
2592         if (unlikely(--req->ref_count))
2593                 return;
2594
2595         elv_completed_request(q, req);
2596
2597         req->rq_status = RQ_INACTIVE;
2598         req->rl = NULL;
2599
2600         /*
2601          * Request may not have originated from ll_rw_blk. if not,
2602          * it didn't come out of our reserved rq pools
2603          */
2604         if (rl) {
2605                 int rw = rq_data_dir(req);
2606                 int priv = req->flags & REQ_ELVPRIV;
2607
2608                 BUG_ON(!list_empty(&req->queuelist));
2609
2610                 blk_free_request(q, req);
2611                 freed_request(q, rw, priv);
2612         }
2613 }
2614
2615 EXPORT_SYMBOL_GPL(__blk_put_request);
2616
2617 void blk_put_request(struct request *req)
2618 {
2619         unsigned long flags;
2620         request_queue_t *q = req->q;
2621
2622         /*
2623          * Gee, IDE calls in w/ NULL q.  Fix IDE and remove the
2624          * following if (q) test.
2625          */
2626         if (q) {
2627                 spin_lock_irqsave(q->queue_lock, flags);
2628                 __blk_put_request(q, req);
2629                 spin_unlock_irqrestore(q->queue_lock, flags);
2630         }
2631 }
2632
2633 EXPORT_SYMBOL(blk_put_request);
2634
2635 /**
2636  * blk_end_sync_rq - executes a completion event on a request
2637  * @rq: request to complete
2638  * @error: end io status of the request
2639  */
2640 void blk_end_sync_rq(struct request *rq, int error)
2641 {
2642         struct completion *waiting = rq->waiting;
2643
2644         rq->waiting = NULL;
2645         __blk_put_request(rq->q, rq);
2646
2647         /*
2648          * complete last, if this is a stack request the process (and thus
2649          * the rq pointer) could be invalid right after this complete()
2650          */
2651         complete(waiting);
2652 }
2653 EXPORT_SYMBOL(blk_end_sync_rq);
2654
2655 /**
2656  * blk_congestion_wait - wait for a queue to become uncongested
2657  * @rw: READ or WRITE
2658  * @timeout: timeout in jiffies
2659  *
2660  * Waits for up to @timeout jiffies for a queue (any queue) to exit congestion.
2661  * If no queues are congested then just wait for the next request to be
2662  * returned.
2663  */
2664 long blk_congestion_wait(int rw, long timeout)
2665 {
2666         long ret;
2667         DEFINE_WAIT(wait);
2668         wait_queue_head_t *wqh = &congestion_wqh[rw];
2669
2670         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
2671         ret = io_schedule_timeout(timeout);
2672         finish_wait(wqh, &wait);
2673         return ret;
2674 }
2675
2676 EXPORT_SYMBOL(blk_congestion_wait);
2677
2678 /*
2679  * Has to be called with the request spinlock acquired
2680  */
2681 static int attempt_merge(request_queue_t *q, struct request *req,
2682                           struct request *next)
2683 {
2684         if (!rq_mergeable(req) || !rq_mergeable(next))
2685                 return 0;
2686
2687         /*
2688          * not contigious
2689          */
2690         if (req->sector + req->nr_sectors != next->sector)
2691                 return 0;
2692
2693         if (rq_data_dir(req) != rq_data_dir(next)
2694             || req->rq_disk != next->rq_disk
2695             || next->waiting || next->special)
2696                 return 0;
2697
2698         /*
2699          * If we are allowed to merge, then append bio list
2700          * from next to rq and release next. merge_requests_fn
2701          * will have updated segment counts, update sector
2702          * counts here.
2703          */
2704         if (!q->merge_requests_fn(q, req, next))
2705                 return 0;
2706
2707         /*
2708          * At this point we have either done a back merge
2709          * or front merge. We need the smaller start_time of
2710          * the merged requests to be the current request
2711          * for accounting purposes.
2712          */
2713         if (time_after(req->start_time, next->start_time))
2714                 req->start_time = next->start_time;
2715
2716         req->biotail->bi_next = next->bio;
2717         req->biotail = next->biotail;
2718
2719         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2720
2721         elv_merge_requests(q, req, next);
2722
2723         if (req->rq_disk) {
2724                 disk_round_stats(req->rq_disk);
2725                 req->rq_disk->in_flight--;
2726         }
2727
2728         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2729
2730         __blk_put_request(q, next);
2731         return 1;
2732 }
2733
2734 static inline int attempt_back_merge(request_queue_t *q, struct request *rq)
2735 {
2736         struct request *next = elv_latter_request(q, rq);
2737
2738         if (next)
2739                 return attempt_merge(q, rq, next);
2740
2741         return 0;
2742 }
2743
2744 static inline int attempt_front_merge(request_queue_t *q, struct request *rq)
2745 {
2746         struct request *prev = elv_former_request(q, rq);
2747
2748         if (prev)
2749                 return attempt_merge(q, prev, rq);
2750
2751         return 0;
2752 }
2753
2754 static void init_request_from_bio(struct request *req, struct bio *bio)
2755 {
2756         req->flags |= REQ_CMD;
2757
2758         /*
2759          * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2760          */
2761         if (bio_rw_ahead(bio) || bio_failfast(bio))
2762                 req->flags |= REQ_FAILFAST;
2763
2764         /*
2765          * REQ_BARRIER implies no merging, but lets make it explicit
2766          */
2767         if (unlikely(bio_barrier(bio)))
2768                 req->flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2769
2770         req->errors = 0;
2771         req->hard_sector = req->sector = bio->bi_sector;
2772         req->hard_nr_sectors = req->nr_sectors = bio_sectors(bio);
2773         req->current_nr_sectors = req->hard_cur_sectors = bio_cur_sectors(bio);
2774         req->nr_phys_segments = bio_phys_segments(req->q, bio);
2775         req->nr_hw_segments = bio_hw_segments(req->q, bio);
2776         req->buffer = bio_data(bio);    /* see ->buffer comment above */
2777         req->waiting = NULL;
2778         req->bio = req->biotail = bio;
2779         req->ioprio = bio_prio(bio);
2780         req->rq_disk = bio->bi_bdev->bd_disk;
2781         req->start_time = jiffies;
2782 }
2783
2784 static int __make_request(request_queue_t *q, struct bio *bio)
2785 {
2786         struct request *req;
2787         int el_ret, rw, nr_sectors, cur_nr_sectors, barrier, err, sync;
2788         unsigned short prio;
2789         sector_t sector;
2790
2791         sector = bio->bi_sector;
2792         nr_sectors = bio_sectors(bio);
2793         cur_nr_sectors = bio_cur_sectors(bio);
2794         prio = bio_prio(bio);
2795
2796         rw = bio_data_dir(bio);
2797         sync = bio_sync(bio);
2798
2799         /*
2800          * low level driver can indicate that it wants pages above a
2801          * certain limit bounced to low memory (ie for highmem, or even
2802          * ISA dma in theory)
2803          */
2804         blk_queue_bounce(q, &bio);
2805
2806         spin_lock_prefetch(q->queue_lock);
2807
2808         barrier = bio_barrier(bio);
2809         if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
2810                 err = -EOPNOTSUPP;
2811                 goto end_io;
2812         }
2813
2814         spin_lock_irq(q->queue_lock);
2815
2816         if (unlikely(barrier) || elv_queue_empty(q))
2817                 goto get_rq;
2818
2819         el_ret = elv_merge(q, &req, bio);
2820         switch (el_ret) {
2821                 case ELEVATOR_BACK_MERGE:
2822                         BUG_ON(!rq_mergeable(req));
2823
2824                         if (!q->back_merge_fn(q, req, bio))
2825                                 break;
2826
2827                         req->biotail->bi_next = bio;
2828                         req->biotail = bio;
2829                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2830                         req->ioprio = ioprio_best(req->ioprio, prio);
2831                         drive_stat_acct(req, nr_sectors, 0);
2832                         if (!attempt_back_merge(q, req))
2833                                 elv_merged_request(q, req);
2834                         goto out;
2835
2836                 case ELEVATOR_FRONT_MERGE:
2837                         BUG_ON(!rq_mergeable(req));
2838
2839                         if (!q->front_merge_fn(q, req, bio))
2840                                 break;
2841
2842                         bio->bi_next = req->bio;
2843                         req->bio = bio;
2844
2845                         /*
2846                          * may not be valid. if the low level driver said
2847                          * it didn't need a bounce buffer then it better
2848                          * not touch req->buffer either...
2849                          */
2850                         req->buffer = bio_data(bio);
2851                         req->current_nr_sectors = cur_nr_sectors;
2852                         req->hard_cur_sectors = cur_nr_sectors;
2853                         req->sector = req->hard_sector = sector;
2854                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2855                         req->ioprio = ioprio_best(req->ioprio, prio);
2856                         drive_stat_acct(req, nr_sectors, 0);
2857                         if (!attempt_front_merge(q, req))
2858                                 elv_merged_request(q, req);
2859                         goto out;
2860
2861                 /* ELV_NO_MERGE: elevator says don't/can't merge. */
2862                 default:
2863                         ;
2864         }
2865
2866 get_rq:
2867         /*
2868          * Grab a free request. This is might sleep but can not fail.
2869          * Returns with the queue unlocked.
2870          */
2871         req = get_request_wait(q, rw, bio);
2872
2873         /*
2874          * After dropping the lock and possibly sleeping here, our request
2875          * may now be mergeable after it had proven unmergeable (above).
2876          * We don't worry about that case for efficiency. It won't happen
2877          * often, and the elevators are able to handle it.
2878          */
2879         init_request_from_bio(req, bio);
2880
2881         spin_lock_irq(q->queue_lock);
2882         if (elv_queue_empty(q))
2883                 blk_plug_device(q);
2884         add_request(q, req);
2885 out:
2886         if (sync)
2887                 __generic_unplug_device(q);
2888
2889         spin_unlock_irq(q->queue_lock);
2890         return 0;
2891
2892 end_io:
2893         bio_endio(bio, nr_sectors << 9, err);
2894         return 0;
2895 }
2896
2897 /*
2898  * If bio->bi_dev is a partition, remap the location
2899  */
2900 static inline void blk_partition_remap(struct bio *bio)
2901 {
2902         struct block_device *bdev = bio->bi_bdev;
2903
2904         if (bdev != bdev->bd_contains) {
2905                 struct hd_struct *p = bdev->bd_part;
2906                 const int rw = bio_data_dir(bio);
2907
2908                 p->sectors[rw] += bio_sectors(bio);
2909                 p->ios[rw]++;
2910
2911                 bio->bi_sector += p->start_sect;
2912                 bio->bi_bdev = bdev->bd_contains;
2913         }
2914 }
2915
2916 static void handle_bad_sector(struct bio *bio)
2917 {
2918         char b[BDEVNAME_SIZE];
2919
2920         printk(KERN_INFO "attempt to access beyond end of device\n");
2921         printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
2922                         bdevname(bio->bi_bdev, b),
2923                         bio->bi_rw,
2924                         (unsigned long long)bio->bi_sector + bio_sectors(bio),
2925                         (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
2926
2927         set_bit(BIO_EOF, &bio->bi_flags);
2928 }
2929
2930 /**
2931  * generic_make_request: hand a buffer to its device driver for I/O
2932  * @bio:  The bio describing the location in memory and on the device.
2933  *
2934  * generic_make_request() is used to make I/O requests of block
2935  * devices. It is passed a &struct bio, which describes the I/O that needs
2936  * to be done.
2937  *
2938  * generic_make_request() does not return any status.  The
2939  * success/failure status of the request, along with notification of
2940  * completion, is delivered asynchronously through the bio->bi_end_io
2941  * function described (one day) else where.
2942  *
2943  * The caller of generic_make_request must make sure that bi_io_vec
2944  * are set to describe the memory buffer, and that bi_dev and bi_sector are
2945  * set to describe the device address, and the
2946  * bi_end_io and optionally bi_private are set to describe how
2947  * completion notification should be signaled.
2948  *
2949  * generic_make_request and the drivers it calls may use bi_next if this
2950  * bio happens to be merged with someone else, and may change bi_dev and
2951  * bi_sector for remaps as it sees fit.  So the values of these fields
2952  * should NOT be depended on after the call to generic_make_request.
2953  */
2954 void generic_make_request(struct bio *bio)
2955 {
2956         request_queue_t *q;
2957         sector_t maxsector;
2958         int ret, nr_sectors = bio_sectors(bio);
2959
2960         might_sleep();
2961         /* Test device or partition size, when known. */
2962         maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
2963         if (maxsector) {
2964                 sector_t sector = bio->bi_sector;
2965
2966                 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
2967                         /*
2968                          * This may well happen - the kernel calls bread()
2969                          * without checking the size of the device, e.g., when
2970                          * mounting a device.
2971                          */
2972                         handle_bad_sector(bio);
2973                         goto end_io;
2974                 }
2975         }
2976
2977         /*
2978          * Resolve the mapping until finished. (drivers are
2979          * still free to implement/resolve their own stacking
2980          * by explicitly returning 0)
2981          *
2982          * NOTE: we don't repeat the blk_size check for each new device.
2983          * Stacking drivers are expected to know what they are doing.
2984          */
2985         do {
2986                 char b[BDEVNAME_SIZE];
2987
2988                 q = bdev_get_queue(bio->bi_bdev);
2989                 if (!q) {
2990                         printk(KERN_ERR
2991                                "generic_make_request: Trying to access "
2992                                 "nonexistent block-device %s (%Lu)\n",
2993                                 bdevname(bio->bi_bdev, b),
2994                                 (long long) bio->bi_sector);
2995 end_io:
2996                         bio_endio(bio, bio->bi_size, -EIO);
2997                         break;
2998                 }
2999
3000                 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
3001                         printk("bio too big device %s (%u > %u)\n", 
3002                                 bdevname(bio->bi_bdev, b),
3003                                 bio_sectors(bio),
3004                                 q->max_hw_sectors);
3005                         goto end_io;
3006                 }
3007
3008                 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
3009                         goto end_io;
3010
3011                 /*
3012                  * If this device has partitions, remap block n
3013                  * of partition p to block n+start(p) of the disk.
3014                  */
3015                 blk_partition_remap(bio);
3016
3017                 ret = q->make_request_fn(q, bio);
3018         } while (ret);
3019 }
3020
3021 EXPORT_SYMBOL(generic_make_request);
3022
3023 /**
3024  * submit_bio: submit a bio to the block device layer for I/O
3025  * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3026  * @bio: The &struct bio which describes the I/O
3027  *
3028  * submit_bio() is very similar in purpose to generic_make_request(), and
3029  * uses that function to do most of the work. Both are fairly rough
3030  * interfaces, @bio must be presetup and ready for I/O.
3031  *
3032  */
3033 void submit_bio(int rw, struct bio *bio)
3034 {
3035         int count = bio_sectors(bio);
3036
3037         BIO_BUG_ON(!bio->bi_size);
3038         BIO_BUG_ON(!bio->bi_io_vec);
3039         bio->bi_rw |= rw;
3040         if (rw & WRITE)
3041                 mod_page_state(pgpgout, count);
3042         else
3043                 mod_page_state(pgpgin, count);
3044
3045         if (unlikely(block_dump)) {
3046                 char b[BDEVNAME_SIZE];
3047                 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
3048                         current->comm, current->pid,
3049                         (rw & WRITE) ? "WRITE" : "READ",
3050                         (unsigned long long)bio->bi_sector,
3051                         bdevname(bio->bi_bdev,b));
3052         }
3053
3054         generic_make_request(bio);
3055 }
3056
3057 EXPORT_SYMBOL(submit_bio);
3058
3059 static void blk_recalc_rq_segments(struct request *rq)
3060 {
3061         struct bio *bio, *prevbio = NULL;
3062         int nr_phys_segs, nr_hw_segs;
3063         unsigned int phys_size, hw_size;
3064         request_queue_t *q = rq->q;
3065
3066         if (!rq->bio)
3067                 return;
3068
3069         phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
3070         rq_for_each_bio(bio, rq) {
3071                 /* Force bio hw/phys segs to be recalculated. */
3072                 bio->bi_flags &= ~(1 << BIO_SEG_VALID);
3073
3074                 nr_phys_segs += bio_phys_segments(q, bio);
3075                 nr_hw_segs += bio_hw_segments(q, bio);
3076                 if (prevbio) {
3077                         int pseg = phys_size + prevbio->bi_size + bio->bi_size;
3078                         int hseg = hw_size + prevbio->bi_size + bio->bi_size;
3079
3080                         if (blk_phys_contig_segment(q, prevbio, bio) &&
3081                             pseg <= q->max_segment_size) {
3082                                 nr_phys_segs--;
3083                                 phys_size += prevbio->bi_size + bio->bi_size;
3084                         } else
3085                                 phys_size = 0;
3086
3087                         if (blk_hw_contig_segment(q, prevbio, bio) &&
3088                             hseg <= q->max_segment_size) {
3089                                 nr_hw_segs--;
3090                                 hw_size += prevbio->bi_size + bio->bi_size;
3091                         } else
3092                                 hw_size = 0;
3093                 }
3094                 prevbio = bio;
3095         }
3096
3097         rq->nr_phys_segments = nr_phys_segs;
3098         rq->nr_hw_segments = nr_hw_segs;
3099 }
3100
3101 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
3102 {
3103         if (blk_fs_request(rq)) {
3104                 rq->hard_sector += nsect;
3105                 rq->hard_nr_sectors -= nsect;
3106
3107                 /*
3108                  * Move the I/O submission pointers ahead if required.
3109                  */
3110                 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3111                     (rq->sector <= rq->hard_sector)) {
3112                         rq->sector = rq->hard_sector;
3113                         rq->nr_sectors = rq->hard_nr_sectors;
3114                         rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3115                         rq->current_nr_sectors = rq->hard_cur_sectors;
3116                         rq->buffer = bio_data(rq->bio);
3117                 }
3118
3119                 /*
3120                  * if total number of sectors is less than the first segment
3121                  * size, something has gone terribly wrong
3122                  */
3123                 if (rq->nr_sectors < rq->current_nr_sectors) {
3124                         printk("blk: request botched\n");
3125                         rq->nr_sectors = rq->current_nr_sectors;
3126                 }
3127         }
3128 }
3129
3130 static int __end_that_request_first(struct request *req, int uptodate,
3131                                     int nr_bytes)
3132 {
3133         int total_bytes, bio_nbytes, error, next_idx = 0;
3134         struct bio *bio;
3135
3136         /*
3137          * extend uptodate bool to allow < 0 value to be direct io error
3138          */
3139         error = 0;
3140         if (end_io_error(uptodate))
3141                 error = !uptodate ? -EIO : uptodate;
3142
3143         /*
3144          * for a REQ_BLOCK_PC request, we want to carry any eventual
3145          * sense key with us all the way through
3146          */
3147         if (!blk_pc_request(req))
3148                 req->errors = 0;
3149
3150         if (!uptodate) {
3151                 if (blk_fs_request(req) && !(req->flags & REQ_QUIET))
3152                         printk("end_request: I/O error, dev %s, sector %llu\n",
3153                                 req->rq_disk ? req->rq_disk->disk_name : "?",
3154                                 (unsigned long long)req->sector);
3155         }
3156
3157         if (blk_fs_request(req) && req->rq_disk) {
3158                 const int rw = rq_data_dir(req);
3159
3160                 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
3161         }
3162
3163         total_bytes = bio_nbytes = 0;
3164         while ((bio = req->bio) != NULL) {
3165                 int nbytes;
3166
3167                 if (nr_bytes >= bio->bi_size) {
3168                         req->bio = bio->bi_next;
3169                         nbytes = bio->bi_size;
3170                         if (!ordered_bio_endio(req, bio, nbytes, error))
3171                                 bio_endio(bio, nbytes, error);
3172                         next_idx = 0;
3173                         bio_nbytes = 0;
3174                 } else {
3175                         int idx = bio->bi_idx + next_idx;
3176
3177                         if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3178                                 blk_dump_rq_flags(req, "__end_that");
3179                                 printk("%s: bio idx %d >= vcnt %d\n",
3180                                                 __FUNCTION__,
3181                                                 bio->bi_idx, bio->bi_vcnt);
3182                                 break;
3183                         }
3184
3185                         nbytes = bio_iovec_idx(bio, idx)->bv_len;
3186                         BIO_BUG_ON(nbytes > bio->bi_size);
3187
3188                         /*
3189                          * not a complete bvec done
3190                          */
3191                         if (unlikely(nbytes > nr_bytes)) {
3192                                 bio_nbytes += nr_bytes;
3193                                 total_bytes += nr_bytes;
3194                                 break;
3195                         }
3196
3197                         /*
3198                          * advance to the next vector
3199                          */
3200                         next_idx++;
3201                         bio_nbytes += nbytes;
3202                 }
3203
3204                 total_bytes += nbytes;
3205                 nr_bytes -= nbytes;
3206
3207                 if ((bio = req->bio)) {
3208                         /*
3209                          * end more in this run, or just return 'not-done'
3210                          */
3211                         if (unlikely(nr_bytes <= 0))
3212                                 break;
3213                 }
3214         }
3215
3216         /*
3217          * completely done
3218          */
3219         if (!req->bio)
3220                 return 0;
3221
3222         /*
3223          * if the request wasn't completed, update state
3224          */
3225         if (bio_nbytes) {
3226                 if (!ordered_bio_endio(req, bio, bio_nbytes, error))
3227                         bio_endio(bio, bio_nbytes, error);
3228                 bio->bi_idx += next_idx;
3229                 bio_iovec(bio)->bv_offset += nr_bytes;
3230                 bio_iovec(bio)->bv_len -= nr_bytes;
3231         }
3232
3233         blk_recalc_rq_sectors(req, total_bytes >> 9);
3234         blk_recalc_rq_segments(req);
3235         return 1;
3236 }
3237
3238 /**
3239  * end_that_request_first - end I/O on a request
3240  * @req:      the request being processed
3241  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3242  * @nr_sectors: number of sectors to end I/O on
3243  *
3244  * Description:
3245  *     Ends I/O on a number of sectors attached to @req, and sets it up
3246  *     for the next range of segments (if any) in the cluster.
3247  *
3248  * Return:
3249  *     0 - we are done with this request, call end_that_request_last()
3250  *     1 - still buffers pending for this request
3251  **/
3252 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3253 {
3254         return __end_that_request_first(req, uptodate, nr_sectors << 9);
3255 }
3256
3257 EXPORT_SYMBOL(end_that_request_first);
3258
3259 /**
3260  * end_that_request_chunk - end I/O on a request
3261  * @req:      the request being processed
3262  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3263  * @nr_bytes: number of bytes to complete
3264  *
3265  * Description:
3266  *     Ends I/O on a number of bytes attached to @req, and sets it up
3267  *     for the next range of segments (if any). Like end_that_request_first(),
3268  *     but deals with bytes instead of sectors.
3269  *
3270  * Return:
3271  *     0 - we are done with this request, call end_that_request_last()
3272  *     1 - still buffers pending for this request
3273  **/
3274 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3275 {
3276         return __end_that_request_first(req, uptodate, nr_bytes);
3277 }
3278
3279 EXPORT_SYMBOL(end_that_request_chunk);
3280
3281 /*
3282  * splice the completion data to a local structure and hand off to
3283  * process_completion_queue() to complete the requests
3284  */
3285 static void blk_done_softirq(struct softirq_action *h)
3286 {
3287         struct list_head *cpu_list;
3288         LIST_HEAD(local_list);
3289
3290         local_irq_disable();
3291         cpu_list = &__get_cpu_var(blk_cpu_done);
3292         list_splice_init(cpu_list, &local_list);
3293         local_irq_enable();
3294
3295         while (!list_empty(&local_list)) {
3296                 struct request *rq = list_entry(local_list.next, struct request, donelist);
3297
3298                 list_del_init(&rq->donelist);
3299                 rq->q->softirq_done_fn(rq);
3300         }
3301 }
3302
3303 #ifdef CONFIG_HOTPLUG_CPU
3304
3305 static int blk_cpu_notify(struct notifier_block *self, unsigned long action,
3306                           void *hcpu)
3307 {
3308         /*
3309          * If a CPU goes away, splice its entries to the current CPU
3310          * and trigger a run of the softirq
3311          */
3312         if (action == CPU_DEAD) {
3313                 int cpu = (unsigned long) hcpu;
3314
3315                 local_irq_disable();
3316                 list_splice_init(&per_cpu(blk_cpu_done, cpu),
3317                                  &__get_cpu_var(blk_cpu_done));
3318                 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3319                 local_irq_enable();
3320         }
3321
3322         return NOTIFY_OK;
3323 }
3324
3325
3326 static struct notifier_block __devinitdata blk_cpu_notifier = {
3327         .notifier_call  = blk_cpu_notify,
3328 };
3329
3330 #endif /* CONFIG_HOTPLUG_CPU */
3331
3332 /**
3333  * blk_complete_request - end I/O on a request
3334  * @req:      the request being processed
3335  *
3336  * Description:
3337  *     Ends all I/O on a request. It does not handle partial completions,
3338  *     unless the driver actually implements this in its completionc callback
3339  *     through requeueing. Theh actual completion happens out-of-order,
3340  *     through a softirq handler. The user must have registered a completion
3341  *     callback through blk_queue_softirq_done().
3342  **/
3343
3344 void blk_complete_request(struct request *req)
3345 {
3346         struct list_head *cpu_list;
3347         unsigned long flags;
3348
3349         BUG_ON(!req->q->softirq_done_fn);
3350                 
3351         local_irq_save(flags);
3352
3353         cpu_list = &__get_cpu_var(blk_cpu_done);
3354         list_add_tail(&req->donelist, cpu_list);
3355         raise_softirq_irqoff(BLOCK_SOFTIRQ);
3356
3357         local_irq_restore(flags);
3358 }
3359
3360 EXPORT_SYMBOL(blk_complete_request);
3361         
3362 /*
3363  * queue lock must be held
3364  */
3365 void end_that_request_last(struct request *req, int uptodate)
3366 {
3367         struct gendisk *disk = req->rq_disk;
3368         int error;
3369
3370         /*
3371          * extend uptodate bool to allow < 0 value to be direct io error
3372          */
3373         error = 0;
3374         if (end_io_error(uptodate))
3375                 error = !uptodate ? -EIO : uptodate;
3376
3377         if (unlikely(laptop_mode) && blk_fs_request(req))
3378                 laptop_io_completion();
3379
3380         if (disk && blk_fs_request(req)) {
3381                 unsigned long duration = jiffies - req->start_time;
3382                 const int rw = rq_data_dir(req);
3383
3384                 __disk_stat_inc(disk, ios[rw]);
3385                 __disk_stat_add(disk, ticks[rw], duration);
3386                 disk_round_stats(disk);
3387                 disk->in_flight--;
3388         }
3389         if (req->end_io)
3390                 req->end_io(req, error);
3391         else
3392                 __blk_put_request(req->q, req);
3393 }
3394
3395 EXPORT_SYMBOL(end_that_request_last);
3396
3397 void end_request(struct request *req, int uptodate)
3398 {
3399         if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3400                 add_disk_randomness(req->rq_disk);
3401                 blkdev_dequeue_request(req);
3402                 end_that_request_last(req, uptodate);
3403         }
3404 }
3405
3406 EXPORT_SYMBOL(end_request);
3407
3408 void blk_rq_bio_prep(request_queue_t *q, struct request *rq, struct bio *bio)
3409 {
3410         /* first three bits are identical in rq->flags and bio->bi_rw */
3411         rq->flags |= (bio->bi_rw & 7);
3412
3413         rq->nr_phys_segments = bio_phys_segments(q, bio);
3414         rq->nr_hw_segments = bio_hw_segments(q, bio);
3415         rq->current_nr_sectors = bio_cur_sectors(bio);
3416         rq->hard_cur_sectors = rq->current_nr_sectors;
3417         rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3418         rq->buffer = bio_data(bio);
3419
3420         rq->bio = rq->biotail = bio;
3421 }
3422
3423 EXPORT_SYMBOL(blk_rq_bio_prep);
3424
3425 int kblockd_schedule_work(struct work_struct *work)
3426 {
3427         return queue_work(kblockd_workqueue, work);
3428 }
3429
3430 EXPORT_SYMBOL(kblockd_schedule_work);
3431
3432 void kblockd_flush(void)
3433 {
3434         flush_workqueue(kblockd_workqueue);
3435 }
3436 EXPORT_SYMBOL(kblockd_flush);
3437
3438 int __init blk_dev_init(void)
3439 {
3440         int i;
3441
3442         kblockd_workqueue = create_workqueue("kblockd");
3443         if (!kblockd_workqueue)
3444                 panic("Failed to create kblockd\n");
3445
3446         request_cachep = kmem_cache_create("blkdev_requests",
3447                         sizeof(struct request), 0, SLAB_PANIC, NULL, NULL);
3448
3449         requestq_cachep = kmem_cache_create("blkdev_queue",
3450                         sizeof(request_queue_t), 0, SLAB_PANIC, NULL, NULL);
3451
3452         iocontext_cachep = kmem_cache_create("blkdev_ioc",
3453                         sizeof(struct io_context), 0, SLAB_PANIC, NULL, NULL);
3454
3455         for_each_cpu(i)
3456                 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3457
3458         open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
3459 #ifdef CONFIG_HOTPLUG_CPU
3460         register_cpu_notifier(&blk_cpu_notifier);
3461 #endif
3462
3463         blk_max_low_pfn = max_low_pfn;
3464         blk_max_pfn = max_pfn;
3465
3466         return 0;
3467 }
3468
3469 /*
3470  * IO Context helper functions
3471  */
3472 void put_io_context(struct io_context *ioc)
3473 {
3474         if (ioc == NULL)
3475                 return;
3476
3477         BUG_ON(atomic_read(&ioc->refcount) == 0);
3478
3479         if (atomic_dec_and_test(&ioc->refcount)) {
3480                 if (ioc->aic && ioc->aic->dtor)
3481                         ioc->aic->dtor(ioc->aic);
3482                 if (ioc->cic && ioc->cic->dtor)
3483                         ioc->cic->dtor(ioc->cic);
3484
3485                 kmem_cache_free(iocontext_cachep, ioc);
3486         }
3487 }
3488 EXPORT_SYMBOL(put_io_context);
3489
3490 /* Called by the exitting task */
3491 void exit_io_context(void)
3492 {
3493         unsigned long flags;
3494         struct io_context *ioc;
3495
3496         local_irq_save(flags);
3497         task_lock(current);
3498         ioc = current->io_context;
3499         current->io_context = NULL;
3500         ioc->task = NULL;
3501         task_unlock(current);
3502         local_irq_restore(flags);
3503
3504         if (ioc->aic && ioc->aic->exit)
3505                 ioc->aic->exit(ioc->aic);
3506         if (ioc->cic && ioc->cic->exit)
3507                 ioc->cic->exit(ioc->cic);
3508
3509         put_io_context(ioc);
3510 }
3511
3512 /*
3513  * If the current task has no IO context then create one and initialise it.
3514  * Otherwise, return its existing IO context.
3515  *
3516  * This returned IO context doesn't have a specifically elevated refcount,
3517  * but since the current task itself holds a reference, the context can be
3518  * used in general code, so long as it stays within `current` context.
3519  */
3520 struct io_context *current_io_context(gfp_t gfp_flags)
3521 {
3522         struct task_struct *tsk = current;
3523         struct io_context *ret;
3524
3525         ret = tsk->io_context;
3526         if (likely(ret))
3527                 return ret;
3528
3529         ret = kmem_cache_alloc(iocontext_cachep, gfp_flags);
3530         if (ret) {
3531                 atomic_set(&ret->refcount, 1);
3532                 ret->task = current;
3533                 ret->set_ioprio = NULL;
3534                 ret->last_waited = jiffies; /* doesn't matter... */
3535                 ret->nr_batch_requests = 0; /* because this is 0 */
3536                 ret->aic = NULL;
3537                 ret->cic = NULL;
3538                 tsk->io_context = ret;
3539         }
3540
3541         return ret;
3542 }
3543 EXPORT_SYMBOL(current_io_context);
3544
3545 /*
3546  * If the current task has no IO context then create one and initialise it.
3547  * If it does have a context, take a ref on it.
3548  *
3549  * This is always called in the context of the task which submitted the I/O.
3550  */
3551 struct io_context *get_io_context(gfp_t gfp_flags)
3552 {
3553         struct io_context *ret;
3554         ret = current_io_context(gfp_flags);
3555         if (likely(ret))
3556                 atomic_inc(&ret->refcount);
3557         return ret;
3558 }
3559 EXPORT_SYMBOL(get_io_context);
3560
3561 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3562 {
3563         struct io_context *src = *psrc;
3564         struct io_context *dst = *pdst;
3565
3566         if (src) {
3567                 BUG_ON(atomic_read(&src->refcount) == 0);
3568                 atomic_inc(&src->refcount);
3569                 put_io_context(dst);
3570                 *pdst = src;
3571         }
3572 }
3573 EXPORT_SYMBOL(copy_io_context);
3574
3575 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3576 {
3577         struct io_context *temp;
3578         temp = *ioc1;
3579         *ioc1 = *ioc2;
3580         *ioc2 = temp;
3581 }
3582 EXPORT_SYMBOL(swap_io_context);
3583
3584 /*
3585  * sysfs parts below
3586  */
3587 struct queue_sysfs_entry {
3588         struct attribute attr;
3589         ssize_t (*show)(struct request_queue *, char *);
3590         ssize_t (*store)(struct request_queue *, const char *, size_t);
3591 };
3592
3593 static ssize_t
3594 queue_var_show(unsigned int var, char *page)
3595 {
3596         return sprintf(page, "%d\n", var);
3597 }
3598
3599 static ssize_t
3600 queue_var_store(unsigned long *var, const char *page, size_t count)
3601 {
3602         char *p = (char *) page;
3603
3604         *var = simple_strtoul(p, &p, 10);
3605         return count;
3606 }
3607
3608 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3609 {
3610         return queue_var_show(q->nr_requests, (page));
3611 }
3612
3613 static ssize_t
3614 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3615 {
3616         struct request_list *rl = &q->rq;
3617
3618         int ret = queue_var_store(&q->nr_requests, page, count);
3619         if (q->nr_requests < BLKDEV_MIN_RQ)
3620                 q->nr_requests = BLKDEV_MIN_RQ;
3621         blk_queue_congestion_threshold(q);
3622
3623         if (rl->count[READ] >= queue_congestion_on_threshold(q))
3624                 set_queue_congested(q, READ);
3625         else if (rl->count[READ] < queue_congestion_off_threshold(q))
3626                 clear_queue_congested(q, READ);
3627
3628         if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3629                 set_queue_congested(q, WRITE);
3630         else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3631                 clear_queue_congested(q, WRITE);
3632
3633         if (rl->count[READ] >= q->nr_requests) {
3634                 blk_set_queue_full(q, READ);
3635         } else if (rl->count[READ]+1 <= q->nr_requests) {
3636                 blk_clear_queue_full(q, READ);
3637                 wake_up(&rl->wait[READ]);
3638         }
3639
3640         if (rl->count[WRITE] >= q->nr_requests) {
3641                 blk_set_queue_full(q, WRITE);
3642         } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3643                 blk_clear_queue_full(q, WRITE);
3644                 wake_up(&rl->wait[WRITE]);
3645         }
3646         return ret;
3647 }
3648
3649 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3650 {
3651         int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3652
3653         return queue_var_show(ra_kb, (page));
3654 }
3655
3656 static ssize_t
3657 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3658 {
3659         unsigned long ra_kb;
3660         ssize_t ret = queue_var_store(&ra_kb, page, count);
3661
3662         spin_lock_irq(q->queue_lock);
3663         if (ra_kb > (q->max_sectors >> 1))
3664                 ra_kb = (q->max_sectors >> 1);
3665
3666         q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3667         spin_unlock_irq(q->queue_lock);
3668
3669         return ret;
3670 }
3671
3672 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3673 {
3674         int max_sectors_kb = q->max_sectors >> 1;
3675
3676         return queue_var_show(max_sectors_kb, (page));
3677 }
3678
3679 static ssize_t
3680 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3681 {
3682         unsigned long max_sectors_kb,
3683                         max_hw_sectors_kb = q->max_hw_sectors >> 1,
3684                         page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3685         ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3686         int ra_kb;
3687
3688         if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3689                 return -EINVAL;
3690         /*
3691          * Take the queue lock to update the readahead and max_sectors
3692          * values synchronously:
3693          */
3694         spin_lock_irq(q->queue_lock);
3695         /*
3696          * Trim readahead window as well, if necessary:
3697          */
3698         ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3699         if (ra_kb > max_sectors_kb)
3700                 q->backing_dev_info.ra_pages =
3701                                 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3702
3703         q->max_sectors = max_sectors_kb << 1;
3704         spin_unlock_irq(q->queue_lock);
3705
3706         return ret;
3707 }
3708
3709 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3710 {
3711         int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3712
3713         return queue_var_show(max_hw_sectors_kb, (page));
3714 }
3715
3716
3717 static struct queue_sysfs_entry queue_requests_entry = {
3718         .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3719         .show = queue_requests_show,
3720         .store = queue_requests_store,
3721 };
3722
3723 static struct queue_sysfs_entry queue_ra_entry = {
3724         .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3725         .show = queue_ra_show,
3726         .store = queue_ra_store,
3727 };
3728
3729 static struct queue_sysfs_entry queue_max_sectors_entry = {
3730         .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3731         .show = queue_max_sectors_show,
3732         .store = queue_max_sectors_store,
3733 };
3734
3735 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3736         .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
3737         .show = queue_max_hw_sectors_show,
3738 };
3739
3740 static struct queue_sysfs_entry queue_iosched_entry = {
3741         .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
3742         .show = elv_iosched_show,
3743         .store = elv_iosched_store,
3744 };
3745
3746 static struct attribute *default_attrs[] = {
3747         &queue_requests_entry.attr,
3748         &queue_ra_entry.attr,
3749         &queue_max_hw_sectors_entry.attr,
3750         &queue_max_sectors_entry.attr,
3751         &queue_iosched_entry.attr,
3752         NULL,
3753 };
3754
3755 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
3756
3757 static ssize_t
3758 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
3759 {
3760         struct queue_sysfs_entry *entry = to_queue(attr);
3761         struct request_queue *q;
3762
3763         q = container_of(kobj, struct request_queue, kobj);
3764         if (!entry->show)
3765                 return -EIO;
3766
3767         return entry->show(q, page);
3768 }
3769
3770 static ssize_t
3771 queue_attr_store(struct kobject *kobj, struct attribute *attr,
3772                     const char *page, size_t length)
3773 {
3774         struct queue_sysfs_entry *entry = to_queue(attr);
3775         struct request_queue *q;
3776
3777         q = container_of(kobj, struct request_queue, kobj);
3778         if (!entry->store)
3779                 return -EIO;
3780
3781         return entry->store(q, page, length);
3782 }
3783
3784 static struct sysfs_ops queue_sysfs_ops = {
3785         .show   = queue_attr_show,
3786         .store  = queue_attr_store,
3787 };
3788
3789 static struct kobj_type queue_ktype = {
3790         .sysfs_ops      = &queue_sysfs_ops,
3791         .default_attrs  = default_attrs,
3792 };
3793
3794 int blk_register_queue(struct gendisk *disk)
3795 {
3796         int ret;
3797
3798         request_queue_t *q = disk->queue;
3799
3800         if (!q || !q->request_fn)
3801                 return -ENXIO;
3802
3803         q->kobj.parent = kobject_get(&disk->kobj);
3804         if (!q->kobj.parent)
3805                 return -EBUSY;
3806
3807         snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
3808         q->kobj.ktype = &queue_ktype;
3809
3810         ret = kobject_register(&q->kobj);
3811         if (ret < 0)
3812                 return ret;
3813
3814         ret = elv_register_queue(q);
3815         if (ret) {
3816                 kobject_unregister(&q->kobj);
3817                 return ret;
3818         }
3819
3820         return 0;
3821 }
3822
3823 void blk_unregister_queue(struct gendisk *disk)
3824 {
3825         request_queue_t *q = disk->queue;
3826
3827         if (q && q->request_fn) {
3828                 elv_unregister_queue(q);
3829
3830                 kobject_unregister(&q->kobj);
3831                 kobject_put(&disk->kobj);
3832         }
3833 }