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