[PATCH] percpu data: only iterate over possible CPUs
[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  * @prepare_flush_fn: rq setup helper for cache flush ordered writes
308  *
309  * Description:
310  *   For journalled file systems, doing ordered writes on a commit
311  *   block instead of explicitly doing wait_on_buffer (which is bad
312  *   for performance) can be a big win. Block drivers supporting this
313  *   feature should call this function and indicate so.
314  *
315  **/
316 int blk_queue_ordered(request_queue_t *q, unsigned ordered,
317                       prepare_flush_fn *prepare_flush_fn)
318 {
319         if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
320             prepare_flush_fn == NULL) {
321                 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
322                 return -EINVAL;
323         }
324
325         if (ordered != QUEUE_ORDERED_NONE &&
326             ordered != QUEUE_ORDERED_DRAIN &&
327             ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
328             ordered != QUEUE_ORDERED_DRAIN_FUA &&
329             ordered != QUEUE_ORDERED_TAG &&
330             ordered != QUEUE_ORDERED_TAG_FLUSH &&
331             ordered != QUEUE_ORDERED_TAG_FUA) {
332                 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
333                 return -EINVAL;
334         }
335
336         q->ordered = ordered;
337         q->next_ordered = ordered;
338         q->prepare_flush_fn = prepare_flush_fn;
339
340         return 0;
341 }
342
343 EXPORT_SYMBOL(blk_queue_ordered);
344
345 /**
346  * blk_queue_issue_flush_fn - set function for issuing a flush
347  * @q:     the request queue
348  * @iff:   the function to be called issuing the flush
349  *
350  * Description:
351  *   If a driver supports issuing a flush command, the support is notified
352  *   to the block layer by defining it through this call.
353  *
354  **/
355 void blk_queue_issue_flush_fn(request_queue_t *q, issue_flush_fn *iff)
356 {
357         q->issue_flush_fn = iff;
358 }
359
360 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
361
362 /*
363  * Cache flushing for ordered writes handling
364  */
365 inline unsigned blk_ordered_cur_seq(request_queue_t *q)
366 {
367         if (!q->ordseq)
368                 return 0;
369         return 1 << ffz(q->ordseq);
370 }
371
372 unsigned blk_ordered_req_seq(struct request *rq)
373 {
374         request_queue_t *q = rq->q;
375
376         BUG_ON(q->ordseq == 0);
377
378         if (rq == &q->pre_flush_rq)
379                 return QUEUE_ORDSEQ_PREFLUSH;
380         if (rq == &q->bar_rq)
381                 return QUEUE_ORDSEQ_BAR;
382         if (rq == &q->post_flush_rq)
383                 return QUEUE_ORDSEQ_POSTFLUSH;
384
385         if ((rq->flags & REQ_ORDERED_COLOR) ==
386             (q->orig_bar_rq->flags & REQ_ORDERED_COLOR))
387                 return QUEUE_ORDSEQ_DRAIN;
388         else
389                 return QUEUE_ORDSEQ_DONE;
390 }
391
392 void blk_ordered_complete_seq(request_queue_t *q, unsigned seq, int error)
393 {
394         struct request *rq;
395         int uptodate;
396
397         if (error && !q->orderr)
398                 q->orderr = error;
399
400         BUG_ON(q->ordseq & seq);
401         q->ordseq |= seq;
402
403         if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
404                 return;
405
406         /*
407          * Okay, sequence complete.
408          */
409         rq = q->orig_bar_rq;
410         uptodate = q->orderr ? q->orderr : 1;
411
412         q->ordseq = 0;
413
414         end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
415         end_that_request_last(rq, uptodate);
416 }
417
418 static void pre_flush_end_io(struct request *rq, int error)
419 {
420         elv_completed_request(rq->q, rq);
421         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
422 }
423
424 static void bar_end_io(struct request *rq, int error)
425 {
426         elv_completed_request(rq->q, rq);
427         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
428 }
429
430 static void post_flush_end_io(struct request *rq, int error)
431 {
432         elv_completed_request(rq->q, rq);
433         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
434 }
435
436 static void queue_flush(request_queue_t *q, unsigned which)
437 {
438         struct request *rq;
439         rq_end_io_fn *end_io;
440
441         if (which == QUEUE_ORDERED_PREFLUSH) {
442                 rq = &q->pre_flush_rq;
443                 end_io = pre_flush_end_io;
444         } else {
445                 rq = &q->post_flush_rq;
446                 end_io = post_flush_end_io;
447         }
448
449         rq_init(q, rq);
450         rq->flags = REQ_HARDBARRIER;
451         rq->elevator_private = NULL;
452         rq->rq_disk = q->bar_rq.rq_disk;
453         rq->rl = NULL;
454         rq->end_io = end_io;
455         q->prepare_flush_fn(q, rq);
456
457         __elv_add_request(q, rq, ELEVATOR_INSERT_FRONT, 0);
458 }
459
460 static inline struct request *start_ordered(request_queue_t *q,
461                                             struct request *rq)
462 {
463         q->bi_size = 0;
464         q->orderr = 0;
465         q->ordered = q->next_ordered;
466         q->ordseq |= QUEUE_ORDSEQ_STARTED;
467
468         /*
469          * Prep proxy barrier request.
470          */
471         blkdev_dequeue_request(rq);
472         q->orig_bar_rq = rq;
473         rq = &q->bar_rq;
474         rq_init(q, rq);
475         rq->flags = bio_data_dir(q->orig_bar_rq->bio);
476         rq->flags |= q->ordered & QUEUE_ORDERED_FUA ? REQ_FUA : 0;
477         rq->elevator_private = NULL;
478         rq->rl = NULL;
479         init_request_from_bio(rq, q->orig_bar_rq->bio);
480         rq->end_io = bar_end_io;
481
482         /*
483          * Queue ordered sequence.  As we stack them at the head, we
484          * need to queue in reverse order.  Note that we rely on that
485          * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
486          * request gets inbetween ordered sequence.
487          */
488         if (q->ordered & QUEUE_ORDERED_POSTFLUSH)
489                 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
490         else
491                 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
492
493         __elv_add_request(q, rq, ELEVATOR_INSERT_FRONT, 0);
494
495         if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
496                 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
497                 rq = &q->pre_flush_rq;
498         } else
499                 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
500
501         if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
502                 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
503         else
504                 rq = NULL;
505
506         return rq;
507 }
508
509 int blk_do_ordered(request_queue_t *q, struct request **rqp)
510 {
511         struct request *rq = *rqp, *allowed_rq;
512         int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
513
514         if (!q->ordseq) {
515                 if (!is_barrier)
516                         return 1;
517
518                 if (q->next_ordered != QUEUE_ORDERED_NONE) {
519                         *rqp = start_ordered(q, rq);
520                         return 1;
521                 } else {
522                         /*
523                          * This can happen when the queue switches to
524                          * ORDERED_NONE while this request is on it.
525                          */
526                         blkdev_dequeue_request(rq);
527                         end_that_request_first(rq, -EOPNOTSUPP,
528                                                rq->hard_nr_sectors);
529                         end_that_request_last(rq, -EOPNOTSUPP);
530                         *rqp = NULL;
531                         return 0;
532                 }
533         }
534
535         if (q->ordered & QUEUE_ORDERED_TAG) {
536                 if (is_barrier && rq != &q->bar_rq)
537                         *rqp = NULL;
538                 return 1;
539         }
540
541         switch (blk_ordered_cur_seq(q)) {
542         case QUEUE_ORDSEQ_PREFLUSH:
543                 allowed_rq = &q->pre_flush_rq;
544                 break;
545         case QUEUE_ORDSEQ_BAR:
546                 allowed_rq = &q->bar_rq;
547                 break;
548         case QUEUE_ORDSEQ_POSTFLUSH:
549                 allowed_rq = &q->post_flush_rq;
550                 break;
551         default:
552                 allowed_rq = NULL;
553                 break;
554         }
555
556         if (rq != allowed_rq &&
557             (blk_fs_request(rq) || rq == &q->pre_flush_rq ||
558              rq == &q->post_flush_rq))
559                 *rqp = NULL;
560
561         return 1;
562 }
563
564 static int flush_dry_bio_endio(struct bio *bio, unsigned int bytes, int error)
565 {
566         request_queue_t *q = bio->bi_private;
567         struct bio_vec *bvec;
568         int i;
569
570         /*
571          * This is dry run, restore bio_sector and size.  We'll finish
572          * this request again with the original bi_end_io after an
573          * error occurs or post flush is complete.
574          */
575         q->bi_size += bytes;
576
577         if (bio->bi_size)
578                 return 1;
579
580         /* Rewind bvec's */
581         bio->bi_idx = 0;
582         bio_for_each_segment(bvec, bio, i) {
583                 bvec->bv_len += bvec->bv_offset;
584                 bvec->bv_offset = 0;
585         }
586
587         /* Reset bio */
588         set_bit(BIO_UPTODATE, &bio->bi_flags);
589         bio->bi_size = q->bi_size;
590         bio->bi_sector -= (q->bi_size >> 9);
591         q->bi_size = 0;
592
593         return 0;
594 }
595
596 static inline int ordered_bio_endio(struct request *rq, struct bio *bio,
597                                     unsigned int nbytes, int error)
598 {
599         request_queue_t *q = rq->q;
600         bio_end_io_t *endio;
601         void *private;
602
603         if (&q->bar_rq != rq)
604                 return 0;
605
606         /*
607          * Okay, this is the barrier request in progress, dry finish it.
608          */
609         if (error && !q->orderr)
610                 q->orderr = error;
611
612         endio = bio->bi_end_io;
613         private = bio->bi_private;
614         bio->bi_end_io = flush_dry_bio_endio;
615         bio->bi_private = q;
616
617         bio_endio(bio, nbytes, error);
618
619         bio->bi_end_io = endio;
620         bio->bi_private = private;
621
622         return 1;
623 }
624
625 /**
626  * blk_queue_bounce_limit - set bounce buffer limit for queue
627  * @q:  the request queue for the device
628  * @dma_addr:   bus address limit
629  *
630  * Description:
631  *    Different hardware can have different requirements as to what pages
632  *    it can do I/O directly to. A low level driver can call
633  *    blk_queue_bounce_limit to have lower memory pages allocated as bounce
634  *    buffers for doing I/O to pages residing above @page. By default
635  *    the block layer sets this to the highest numbered "low" memory page.
636  **/
637 void blk_queue_bounce_limit(request_queue_t *q, u64 dma_addr)
638 {
639         unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
640
641         /*
642          * set appropriate bounce gfp mask -- unfortunately we don't have a
643          * full 4GB zone, so we have to resort to low memory for any bounces.
644          * ISA has its own < 16MB zone.
645          */
646         if (bounce_pfn < blk_max_low_pfn) {
647                 BUG_ON(dma_addr < BLK_BOUNCE_ISA);
648                 init_emergency_isa_pool();
649                 q->bounce_gfp = GFP_NOIO | GFP_DMA;
650         } else
651                 q->bounce_gfp = GFP_NOIO;
652
653         q->bounce_pfn = bounce_pfn;
654 }
655
656 EXPORT_SYMBOL(blk_queue_bounce_limit);
657
658 /**
659  * blk_queue_max_sectors - set max sectors for a request for this queue
660  * @q:  the request queue for the device
661  * @max_sectors:  max sectors in the usual 512b unit
662  *
663  * Description:
664  *    Enables a low level driver to set an upper limit on the size of
665  *    received requests.
666  **/
667 void blk_queue_max_sectors(request_queue_t *q, unsigned int max_sectors)
668 {
669         if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
670                 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
671                 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
672         }
673
674         if (BLK_DEF_MAX_SECTORS > max_sectors)
675                 q->max_hw_sectors = q->max_sectors = max_sectors;
676         else {
677                 q->max_sectors = BLK_DEF_MAX_SECTORS;
678                 q->max_hw_sectors = max_sectors;
679         }
680 }
681
682 EXPORT_SYMBOL(blk_queue_max_sectors);
683
684 /**
685  * blk_queue_max_phys_segments - set max phys segments for a request for this queue
686  * @q:  the request queue for the device
687  * @max_segments:  max number of segments
688  *
689  * Description:
690  *    Enables a low level driver to set an upper limit on the number of
691  *    physical data segments in a request.  This would be the largest sized
692  *    scatter list the driver could handle.
693  **/
694 void blk_queue_max_phys_segments(request_queue_t *q, unsigned short max_segments)
695 {
696         if (!max_segments) {
697                 max_segments = 1;
698                 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
699         }
700
701         q->max_phys_segments = max_segments;
702 }
703
704 EXPORT_SYMBOL(blk_queue_max_phys_segments);
705
706 /**
707  * blk_queue_max_hw_segments - set max hw segments for a request for this queue
708  * @q:  the request queue for the device
709  * @max_segments:  max number of segments
710  *
711  * Description:
712  *    Enables a low level driver to set an upper limit on the number of
713  *    hw data segments in a request.  This would be the largest number of
714  *    address/length pairs the host adapter can actually give as once
715  *    to the device.
716  **/
717 void blk_queue_max_hw_segments(request_queue_t *q, unsigned short max_segments)
718 {
719         if (!max_segments) {
720                 max_segments = 1;
721                 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
722         }
723
724         q->max_hw_segments = max_segments;
725 }
726
727 EXPORT_SYMBOL(blk_queue_max_hw_segments);
728
729 /**
730  * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
731  * @q:  the request queue for the device
732  * @max_size:  max size of segment in bytes
733  *
734  * Description:
735  *    Enables a low level driver to set an upper limit on the size of a
736  *    coalesced segment
737  **/
738 void blk_queue_max_segment_size(request_queue_t *q, unsigned int max_size)
739 {
740         if (max_size < PAGE_CACHE_SIZE) {
741                 max_size = PAGE_CACHE_SIZE;
742                 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
743         }
744
745         q->max_segment_size = max_size;
746 }
747
748 EXPORT_SYMBOL(blk_queue_max_segment_size);
749
750 /**
751  * blk_queue_hardsect_size - set hardware sector size for the queue
752  * @q:  the request queue for the device
753  * @size:  the hardware sector size, in bytes
754  *
755  * Description:
756  *   This should typically be set to the lowest possible sector size
757  *   that the hardware can operate on (possible without reverting to
758  *   even internal read-modify-write operations). Usually the default
759  *   of 512 covers most hardware.
760  **/
761 void blk_queue_hardsect_size(request_queue_t *q, unsigned short size)
762 {
763         q->hardsect_size = size;
764 }
765
766 EXPORT_SYMBOL(blk_queue_hardsect_size);
767
768 /*
769  * Returns the minimum that is _not_ zero, unless both are zero.
770  */
771 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
772
773 /**
774  * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
775  * @t:  the stacking driver (top)
776  * @b:  the underlying device (bottom)
777  **/
778 void blk_queue_stack_limits(request_queue_t *t, request_queue_t *b)
779 {
780         /* zero is "infinity" */
781         t->max_sectors = min_not_zero(t->max_sectors,b->max_sectors);
782         t->max_hw_sectors = min_not_zero(t->max_hw_sectors,b->max_hw_sectors);
783
784         t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
785         t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
786         t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
787         t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
788 }
789
790 EXPORT_SYMBOL(blk_queue_stack_limits);
791
792 /**
793  * blk_queue_segment_boundary - set boundary rules for segment merging
794  * @q:  the request queue for the device
795  * @mask:  the memory boundary mask
796  **/
797 void blk_queue_segment_boundary(request_queue_t *q, unsigned long mask)
798 {
799         if (mask < PAGE_CACHE_SIZE - 1) {
800                 mask = PAGE_CACHE_SIZE - 1;
801                 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
802         }
803
804         q->seg_boundary_mask = mask;
805 }
806
807 EXPORT_SYMBOL(blk_queue_segment_boundary);
808
809 /**
810  * blk_queue_dma_alignment - set dma length and memory alignment
811  * @q:     the request queue for the device
812  * @mask:  alignment mask
813  *
814  * description:
815  *    set required memory and length aligment for direct dma transactions.
816  *    this is used when buiding direct io requests for the queue.
817  *
818  **/
819 void blk_queue_dma_alignment(request_queue_t *q, int mask)
820 {
821         q->dma_alignment = mask;
822 }
823
824 EXPORT_SYMBOL(blk_queue_dma_alignment);
825
826 /**
827  * blk_queue_find_tag - find a request by its tag and queue
828  * @q:   The request queue for the device
829  * @tag: The tag of the request
830  *
831  * Notes:
832  *    Should be used when a device returns a tag and you want to match
833  *    it with a request.
834  *
835  *    no locks need be held.
836  **/
837 struct request *blk_queue_find_tag(request_queue_t *q, int tag)
838 {
839         struct blk_queue_tag *bqt = q->queue_tags;
840
841         if (unlikely(bqt == NULL || tag >= bqt->real_max_depth))
842                 return NULL;
843
844         return bqt->tag_index[tag];
845 }
846
847 EXPORT_SYMBOL(blk_queue_find_tag);
848
849 /**
850  * __blk_queue_free_tags - release tag maintenance info
851  * @q:  the request queue for the device
852  *
853  *  Notes:
854  *    blk_cleanup_queue() will take care of calling this function, if tagging
855  *    has been used. So there's no need to call this directly.
856  **/
857 static void __blk_queue_free_tags(request_queue_t *q)
858 {
859         struct blk_queue_tag *bqt = q->queue_tags;
860
861         if (!bqt)
862                 return;
863
864         if (atomic_dec_and_test(&bqt->refcnt)) {
865                 BUG_ON(bqt->busy);
866                 BUG_ON(!list_empty(&bqt->busy_list));
867
868                 kfree(bqt->tag_index);
869                 bqt->tag_index = NULL;
870
871                 kfree(bqt->tag_map);
872                 bqt->tag_map = NULL;
873
874                 kfree(bqt);
875         }
876
877         q->queue_tags = NULL;
878         q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
879 }
880
881 /**
882  * blk_queue_free_tags - release tag maintenance info
883  * @q:  the request queue for the device
884  *
885  *  Notes:
886  *      This is used to disabled tagged queuing to a device, yet leave
887  *      queue in function.
888  **/
889 void blk_queue_free_tags(request_queue_t *q)
890 {
891         clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
892 }
893
894 EXPORT_SYMBOL(blk_queue_free_tags);
895
896 static int
897 init_tag_map(request_queue_t *q, struct blk_queue_tag *tags, int depth)
898 {
899         struct request **tag_index;
900         unsigned long *tag_map;
901         int nr_ulongs;
902
903         if (depth > q->nr_requests * 2) {
904                 depth = q->nr_requests * 2;
905                 printk(KERN_ERR "%s: adjusted depth to %d\n",
906                                 __FUNCTION__, depth);
907         }
908
909         tag_index = kmalloc(depth * sizeof(struct request *), GFP_ATOMIC);
910         if (!tag_index)
911                 goto fail;
912
913         nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
914         tag_map = kmalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
915         if (!tag_map)
916                 goto fail;
917
918         memset(tag_index, 0, depth * sizeof(struct request *));
919         memset(tag_map, 0, nr_ulongs * sizeof(unsigned long));
920         tags->real_max_depth = depth;
921         tags->max_depth = depth;
922         tags->tag_index = tag_index;
923         tags->tag_map = tag_map;
924
925         return 0;
926 fail:
927         kfree(tag_index);
928         return -ENOMEM;
929 }
930
931 /**
932  * blk_queue_init_tags - initialize the queue tag info
933  * @q:  the request queue for the device
934  * @depth:  the maximum queue depth supported
935  * @tags: the tag to use
936  **/
937 int blk_queue_init_tags(request_queue_t *q, int depth,
938                         struct blk_queue_tag *tags)
939 {
940         int rc;
941
942         BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
943
944         if (!tags && !q->queue_tags) {
945                 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
946                 if (!tags)
947                         goto fail;
948
949                 if (init_tag_map(q, tags, depth))
950                         goto fail;
951
952                 INIT_LIST_HEAD(&tags->busy_list);
953                 tags->busy = 0;
954                 atomic_set(&tags->refcnt, 1);
955         } else if (q->queue_tags) {
956                 if ((rc = blk_queue_resize_tags(q, depth)))
957                         return rc;
958                 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
959                 return 0;
960         } else
961                 atomic_inc(&tags->refcnt);
962
963         /*
964          * assign it, all done
965          */
966         q->queue_tags = tags;
967         q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
968         return 0;
969 fail:
970         kfree(tags);
971         return -ENOMEM;
972 }
973
974 EXPORT_SYMBOL(blk_queue_init_tags);
975
976 /**
977  * blk_queue_resize_tags - change the queueing depth
978  * @q:  the request queue for the device
979  * @new_depth: the new max command queueing depth
980  *
981  *  Notes:
982  *    Must be called with the queue lock held.
983  **/
984 int blk_queue_resize_tags(request_queue_t *q, int new_depth)
985 {
986         struct blk_queue_tag *bqt = q->queue_tags;
987         struct request **tag_index;
988         unsigned long *tag_map;
989         int max_depth, nr_ulongs;
990
991         if (!bqt)
992                 return -ENXIO;
993
994         /*
995          * if we already have large enough real_max_depth.  just
996          * adjust max_depth.  *NOTE* as requests with tag value
997          * between new_depth and real_max_depth can be in-flight, tag
998          * map can not be shrunk blindly here.
999          */
1000         if (new_depth <= bqt->real_max_depth) {
1001                 bqt->max_depth = new_depth;
1002                 return 0;
1003         }
1004
1005         /*
1006          * save the old state info, so we can copy it back
1007          */
1008         tag_index = bqt->tag_index;
1009         tag_map = bqt->tag_map;
1010         max_depth = bqt->real_max_depth;
1011
1012         if (init_tag_map(q, bqt, new_depth))
1013                 return -ENOMEM;
1014
1015         memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
1016         nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
1017         memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
1018
1019         kfree(tag_index);
1020         kfree(tag_map);
1021         return 0;
1022 }
1023
1024 EXPORT_SYMBOL(blk_queue_resize_tags);
1025
1026 /**
1027  * blk_queue_end_tag - end tag operations for a request
1028  * @q:  the request queue for the device
1029  * @rq: the request that has completed
1030  *
1031  *  Description:
1032  *    Typically called when end_that_request_first() returns 0, meaning
1033  *    all transfers have been done for a request. It's important to call
1034  *    this function before end_that_request_last(), as that will put the
1035  *    request back on the free list thus corrupting the internal tag list.
1036  *
1037  *  Notes:
1038  *   queue lock must be held.
1039  **/
1040 void blk_queue_end_tag(request_queue_t *q, struct request *rq)
1041 {
1042         struct blk_queue_tag *bqt = q->queue_tags;
1043         int tag = rq->tag;
1044
1045         BUG_ON(tag == -1);
1046
1047         if (unlikely(tag >= bqt->real_max_depth))
1048                 /*
1049                  * This can happen after tag depth has been reduced.
1050                  * FIXME: how about a warning or info message here?
1051                  */
1052                 return;
1053
1054         if (unlikely(!__test_and_clear_bit(tag, bqt->tag_map))) {
1055                 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
1056                        __FUNCTION__, tag);
1057                 return;
1058         }
1059
1060         list_del_init(&rq->queuelist);
1061         rq->flags &= ~REQ_QUEUED;
1062         rq->tag = -1;
1063
1064         if (unlikely(bqt->tag_index[tag] == NULL))
1065                 printk(KERN_ERR "%s: tag %d is missing\n",
1066                        __FUNCTION__, tag);
1067
1068         bqt->tag_index[tag] = NULL;
1069         bqt->busy--;
1070 }
1071
1072 EXPORT_SYMBOL(blk_queue_end_tag);
1073
1074 /**
1075  * blk_queue_start_tag - find a free tag and assign it
1076  * @q:  the request queue for the device
1077  * @rq:  the block request that needs tagging
1078  *
1079  *  Description:
1080  *    This can either be used as a stand-alone helper, or possibly be
1081  *    assigned as the queue &prep_rq_fn (in which case &struct request
1082  *    automagically gets a tag assigned). Note that this function
1083  *    assumes that any type of request can be queued! if this is not
1084  *    true for your device, you must check the request type before
1085  *    calling this function.  The request will also be removed from
1086  *    the request queue, so it's the drivers responsibility to readd
1087  *    it if it should need to be restarted for some reason.
1088  *
1089  *  Notes:
1090  *   queue lock must be held.
1091  **/
1092 int blk_queue_start_tag(request_queue_t *q, struct request *rq)
1093 {
1094         struct blk_queue_tag *bqt = q->queue_tags;
1095         int tag;
1096
1097         if (unlikely((rq->flags & REQ_QUEUED))) {
1098                 printk(KERN_ERR 
1099                        "%s: request %p for device [%s] already tagged %d",
1100                        __FUNCTION__, rq,
1101                        rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
1102                 BUG();
1103         }
1104
1105         tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
1106         if (tag >= bqt->max_depth)
1107                 return 1;
1108
1109         __set_bit(tag, bqt->tag_map);
1110
1111         rq->flags |= REQ_QUEUED;
1112         rq->tag = tag;
1113         bqt->tag_index[tag] = rq;
1114         blkdev_dequeue_request(rq);
1115         list_add(&rq->queuelist, &bqt->busy_list);
1116         bqt->busy++;
1117         return 0;
1118 }
1119
1120 EXPORT_SYMBOL(blk_queue_start_tag);
1121
1122 /**
1123  * blk_queue_invalidate_tags - invalidate all pending tags
1124  * @q:  the request queue for the device
1125  *
1126  *  Description:
1127  *   Hardware conditions may dictate a need to stop all pending requests.
1128  *   In this case, we will safely clear the block side of the tag queue and
1129  *   readd all requests to the request queue in the right order.
1130  *
1131  *  Notes:
1132  *   queue lock must be held.
1133  **/
1134 void blk_queue_invalidate_tags(request_queue_t *q)
1135 {
1136         struct blk_queue_tag *bqt = q->queue_tags;
1137         struct list_head *tmp, *n;
1138         struct request *rq;
1139
1140         list_for_each_safe(tmp, n, &bqt->busy_list) {
1141                 rq = list_entry_rq(tmp);
1142
1143                 if (rq->tag == -1) {
1144                         printk(KERN_ERR
1145                                "%s: bad tag found on list\n", __FUNCTION__);
1146                         list_del_init(&rq->queuelist);
1147                         rq->flags &= ~REQ_QUEUED;
1148                 } else
1149                         blk_queue_end_tag(q, rq);
1150
1151                 rq->flags &= ~REQ_STARTED;
1152                 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1153         }
1154 }
1155
1156 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1157
1158 static const char * const rq_flags[] = {
1159         "REQ_RW",
1160         "REQ_FAILFAST",
1161         "REQ_SORTED",
1162         "REQ_SOFTBARRIER",
1163         "REQ_HARDBARRIER",
1164         "REQ_FUA",
1165         "REQ_CMD",
1166         "REQ_NOMERGE",
1167         "REQ_STARTED",
1168         "REQ_DONTPREP",
1169         "REQ_QUEUED",
1170         "REQ_ELVPRIV",
1171         "REQ_PC",
1172         "REQ_BLOCK_PC",
1173         "REQ_SENSE",
1174         "REQ_FAILED",
1175         "REQ_QUIET",
1176         "REQ_SPECIAL",
1177         "REQ_DRIVE_CMD",
1178         "REQ_DRIVE_TASK",
1179         "REQ_DRIVE_TASKFILE",
1180         "REQ_PREEMPT",
1181         "REQ_PM_SUSPEND",
1182         "REQ_PM_RESUME",
1183         "REQ_PM_SHUTDOWN",
1184         "REQ_ORDERED_COLOR",
1185 };
1186
1187 void blk_dump_rq_flags(struct request *rq, char *msg)
1188 {
1189         int bit;
1190
1191         printk("%s: dev %s: flags = ", msg,
1192                 rq->rq_disk ? rq->rq_disk->disk_name : "?");
1193         bit = 0;
1194         do {
1195                 if (rq->flags & (1 << bit))
1196                         printk("%s ", rq_flags[bit]);
1197                 bit++;
1198         } while (bit < __REQ_NR_BITS);
1199
1200         printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1201                                                        rq->nr_sectors,
1202                                                        rq->current_nr_sectors);
1203         printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1204
1205         if (rq->flags & (REQ_BLOCK_PC | REQ_PC)) {
1206                 printk("cdb: ");
1207                 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1208                         printk("%02x ", rq->cmd[bit]);
1209                 printk("\n");
1210         }
1211 }
1212
1213 EXPORT_SYMBOL(blk_dump_rq_flags);
1214
1215 void blk_recount_segments(request_queue_t *q, struct bio *bio)
1216 {
1217         struct bio_vec *bv, *bvprv = NULL;
1218         int i, nr_phys_segs, nr_hw_segs, seg_size, hw_seg_size, cluster;
1219         int high, highprv = 1;
1220
1221         if (unlikely(!bio->bi_io_vec))
1222                 return;
1223
1224         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1225         hw_seg_size = seg_size = nr_phys_segs = nr_hw_segs = 0;
1226         bio_for_each_segment(bv, bio, i) {
1227                 /*
1228                  * the trick here is making sure that a high page is never
1229                  * considered part of another segment, since that might
1230                  * change with the bounce page.
1231                  */
1232                 high = page_to_pfn(bv->bv_page) >= q->bounce_pfn;
1233                 if (high || highprv)
1234                         goto new_hw_segment;
1235                 if (cluster) {
1236                         if (seg_size + bv->bv_len > q->max_segment_size)
1237                                 goto new_segment;
1238                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1239                                 goto new_segment;
1240                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1241                                 goto new_segment;
1242                         if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1243                                 goto new_hw_segment;
1244
1245                         seg_size += bv->bv_len;
1246                         hw_seg_size += bv->bv_len;
1247                         bvprv = bv;
1248                         continue;
1249                 }
1250 new_segment:
1251                 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1252                     !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len)) {
1253                         hw_seg_size += bv->bv_len;
1254                 } else {
1255 new_hw_segment:
1256                         if (hw_seg_size > bio->bi_hw_front_size)
1257                                 bio->bi_hw_front_size = hw_seg_size;
1258                         hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1259                         nr_hw_segs++;
1260                 }
1261
1262                 nr_phys_segs++;
1263                 bvprv = bv;
1264                 seg_size = bv->bv_len;
1265                 highprv = high;
1266         }
1267         if (hw_seg_size > bio->bi_hw_back_size)
1268                 bio->bi_hw_back_size = hw_seg_size;
1269         if (nr_hw_segs == 1 && hw_seg_size > bio->bi_hw_front_size)
1270                 bio->bi_hw_front_size = hw_seg_size;
1271         bio->bi_phys_segments = nr_phys_segs;
1272         bio->bi_hw_segments = nr_hw_segs;
1273         bio->bi_flags |= (1 << BIO_SEG_VALID);
1274 }
1275
1276
1277 static int blk_phys_contig_segment(request_queue_t *q, struct bio *bio,
1278                                    struct bio *nxt)
1279 {
1280         if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1281                 return 0;
1282
1283         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1284                 return 0;
1285         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1286                 return 0;
1287
1288         /*
1289          * bio and nxt are contigous in memory, check if the queue allows
1290          * these two to be merged into one
1291          */
1292         if (BIO_SEG_BOUNDARY(q, bio, nxt))
1293                 return 1;
1294
1295         return 0;
1296 }
1297
1298 static int blk_hw_contig_segment(request_queue_t *q, struct bio *bio,
1299                                  struct bio *nxt)
1300 {
1301         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1302                 blk_recount_segments(q, bio);
1303         if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1304                 blk_recount_segments(q, nxt);
1305         if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1306             BIOVEC_VIRT_OVERSIZE(bio->bi_hw_front_size + bio->bi_hw_back_size))
1307                 return 0;
1308         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1309                 return 0;
1310
1311         return 1;
1312 }
1313
1314 /*
1315  * map a request to scatterlist, return number of sg entries setup. Caller
1316  * must make sure sg can hold rq->nr_phys_segments entries
1317  */
1318 int blk_rq_map_sg(request_queue_t *q, struct request *rq, struct scatterlist *sg)
1319 {
1320         struct bio_vec *bvec, *bvprv;
1321         struct bio *bio;
1322         int nsegs, i, cluster;
1323
1324         nsegs = 0;
1325         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1326
1327         /*
1328          * for each bio in rq
1329          */
1330         bvprv = NULL;
1331         rq_for_each_bio(bio, rq) {
1332                 /*
1333                  * for each segment in bio
1334                  */
1335                 bio_for_each_segment(bvec, bio, i) {
1336                         int nbytes = bvec->bv_len;
1337
1338                         if (bvprv && cluster) {
1339                                 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1340                                         goto new_segment;
1341
1342                                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1343                                         goto new_segment;
1344                                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1345                                         goto new_segment;
1346
1347                                 sg[nsegs - 1].length += nbytes;
1348                         } else {
1349 new_segment:
1350                                 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1351                                 sg[nsegs].page = bvec->bv_page;
1352                                 sg[nsegs].length = nbytes;
1353                                 sg[nsegs].offset = bvec->bv_offset;
1354
1355                                 nsegs++;
1356                         }
1357                         bvprv = bvec;
1358                 } /* segments in bio */
1359         } /* bios in rq */
1360
1361         return nsegs;
1362 }
1363
1364 EXPORT_SYMBOL(blk_rq_map_sg);
1365
1366 /*
1367  * the standard queue merge functions, can be overridden with device
1368  * specific ones if so desired
1369  */
1370
1371 static inline int ll_new_mergeable(request_queue_t *q,
1372                                    struct request *req,
1373                                    struct bio *bio)
1374 {
1375         int nr_phys_segs = bio_phys_segments(q, bio);
1376
1377         if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1378                 req->flags |= REQ_NOMERGE;
1379                 if (req == q->last_merge)
1380                         q->last_merge = NULL;
1381                 return 0;
1382         }
1383
1384         /*
1385          * A hw segment is just getting larger, bump just the phys
1386          * counter.
1387          */
1388         req->nr_phys_segments += nr_phys_segs;
1389         return 1;
1390 }
1391
1392 static inline int ll_new_hw_segment(request_queue_t *q,
1393                                     struct request *req,
1394                                     struct bio *bio)
1395 {
1396         int nr_hw_segs = bio_hw_segments(q, bio);
1397         int nr_phys_segs = bio_phys_segments(q, bio);
1398
1399         if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1400             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1401                 req->flags |= REQ_NOMERGE;
1402                 if (req == q->last_merge)
1403                         q->last_merge = NULL;
1404                 return 0;
1405         }
1406
1407         /*
1408          * This will form the start of a new hw segment.  Bump both
1409          * counters.
1410          */
1411         req->nr_hw_segments += nr_hw_segs;
1412         req->nr_phys_segments += nr_phys_segs;
1413         return 1;
1414 }
1415
1416 static int ll_back_merge_fn(request_queue_t *q, struct request *req, 
1417                             struct bio *bio)
1418 {
1419         unsigned short max_sectors;
1420         int len;
1421
1422         if (unlikely(blk_pc_request(req)))
1423                 max_sectors = q->max_hw_sectors;
1424         else
1425                 max_sectors = q->max_sectors;
1426
1427         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1428                 req->flags |= REQ_NOMERGE;
1429                 if (req == q->last_merge)
1430                         q->last_merge = NULL;
1431                 return 0;
1432         }
1433         if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1434                 blk_recount_segments(q, req->biotail);
1435         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1436                 blk_recount_segments(q, bio);
1437         len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1438         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1439             !BIOVEC_VIRT_OVERSIZE(len)) {
1440                 int mergeable =  ll_new_mergeable(q, req, bio);
1441
1442                 if (mergeable) {
1443                         if (req->nr_hw_segments == 1)
1444                                 req->bio->bi_hw_front_size = len;
1445                         if (bio->bi_hw_segments == 1)
1446                                 bio->bi_hw_back_size = len;
1447                 }
1448                 return mergeable;
1449         }
1450
1451         return ll_new_hw_segment(q, req, bio);
1452 }
1453
1454 static int ll_front_merge_fn(request_queue_t *q, struct request *req, 
1455                              struct bio *bio)
1456 {
1457         unsigned short max_sectors;
1458         int len;
1459
1460         if (unlikely(blk_pc_request(req)))
1461                 max_sectors = q->max_hw_sectors;
1462         else
1463                 max_sectors = q->max_sectors;
1464
1465
1466         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1467                 req->flags |= REQ_NOMERGE;
1468                 if (req == q->last_merge)
1469                         q->last_merge = NULL;
1470                 return 0;
1471         }
1472         len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1473         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1474                 blk_recount_segments(q, bio);
1475         if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1476                 blk_recount_segments(q, req->bio);
1477         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1478             !BIOVEC_VIRT_OVERSIZE(len)) {
1479                 int mergeable =  ll_new_mergeable(q, req, bio);
1480
1481                 if (mergeable) {
1482                         if (bio->bi_hw_segments == 1)
1483                                 bio->bi_hw_front_size = len;
1484                         if (req->nr_hw_segments == 1)
1485                                 req->biotail->bi_hw_back_size = len;
1486                 }
1487                 return mergeable;
1488         }
1489
1490         return ll_new_hw_segment(q, req, bio);
1491 }
1492
1493 static int ll_merge_requests_fn(request_queue_t *q, struct request *req,
1494                                 struct request *next)
1495 {
1496         int total_phys_segments;
1497         int total_hw_segments;
1498
1499         /*
1500          * First check if the either of the requests are re-queued
1501          * requests.  Can't merge them if they are.
1502          */
1503         if (req->special || next->special)
1504                 return 0;
1505
1506         /*
1507          * Will it become too large?
1508          */
1509         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1510                 return 0;
1511
1512         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1513         if (blk_phys_contig_segment(q, req->biotail, next->bio))
1514                 total_phys_segments--;
1515
1516         if (total_phys_segments > q->max_phys_segments)
1517                 return 0;
1518
1519         total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1520         if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1521                 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1522                 /*
1523                  * propagate the combined length to the end of the requests
1524                  */
1525                 if (req->nr_hw_segments == 1)
1526                         req->bio->bi_hw_front_size = len;
1527                 if (next->nr_hw_segments == 1)
1528                         next->biotail->bi_hw_back_size = len;
1529                 total_hw_segments--;
1530         }
1531
1532         if (total_hw_segments > q->max_hw_segments)
1533                 return 0;
1534
1535         /* Merge is OK... */
1536         req->nr_phys_segments = total_phys_segments;
1537         req->nr_hw_segments = total_hw_segments;
1538         return 1;
1539 }
1540
1541 /*
1542  * "plug" the device if there are no outstanding requests: this will
1543  * force the transfer to start only after we have put all the requests
1544  * on the list.
1545  *
1546  * This is called with interrupts off and no requests on the queue and
1547  * with the queue lock held.
1548  */
1549 void blk_plug_device(request_queue_t *q)
1550 {
1551         WARN_ON(!irqs_disabled());
1552
1553         /*
1554          * don't plug a stopped queue, it must be paired with blk_start_queue()
1555          * which will restart the queueing
1556          */
1557         if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
1558                 return;
1559
1560         if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1561                 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1562 }
1563
1564 EXPORT_SYMBOL(blk_plug_device);
1565
1566 /*
1567  * remove the queue from the plugged list, if present. called with
1568  * queue lock held and interrupts disabled.
1569  */
1570 int blk_remove_plug(request_queue_t *q)
1571 {
1572         WARN_ON(!irqs_disabled());
1573
1574         if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1575                 return 0;
1576
1577         del_timer(&q->unplug_timer);
1578         return 1;
1579 }
1580
1581 EXPORT_SYMBOL(blk_remove_plug);
1582
1583 /*
1584  * remove the plug and let it rip..
1585  */
1586 void __generic_unplug_device(request_queue_t *q)
1587 {
1588         if (unlikely(test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags)))
1589                 return;
1590
1591         if (!blk_remove_plug(q))
1592                 return;
1593
1594         q->request_fn(q);
1595 }
1596 EXPORT_SYMBOL(__generic_unplug_device);
1597
1598 /**
1599  * generic_unplug_device - fire a request queue
1600  * @q:    The &request_queue_t in question
1601  *
1602  * Description:
1603  *   Linux uses plugging to build bigger requests queues before letting
1604  *   the device have at them. If a queue is plugged, the I/O scheduler
1605  *   is still adding and merging requests on the queue. Once the queue
1606  *   gets unplugged, the request_fn defined for the queue is invoked and
1607  *   transfers started.
1608  **/
1609 void generic_unplug_device(request_queue_t *q)
1610 {
1611         spin_lock_irq(q->queue_lock);
1612         __generic_unplug_device(q);
1613         spin_unlock_irq(q->queue_lock);
1614 }
1615 EXPORT_SYMBOL(generic_unplug_device);
1616
1617 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1618                                    struct page *page)
1619 {
1620         request_queue_t *q = bdi->unplug_io_data;
1621
1622         /*
1623          * devices don't necessarily have an ->unplug_fn defined
1624          */
1625         if (q->unplug_fn)
1626                 q->unplug_fn(q);
1627 }
1628
1629 static void blk_unplug_work(void *data)
1630 {
1631         request_queue_t *q = data;
1632
1633         q->unplug_fn(q);
1634 }
1635
1636 static void blk_unplug_timeout(unsigned long data)
1637 {
1638         request_queue_t *q = (request_queue_t *)data;
1639
1640         kblockd_schedule_work(&q->unplug_work);
1641 }
1642
1643 /**
1644  * blk_start_queue - restart a previously stopped queue
1645  * @q:    The &request_queue_t in question
1646  *
1647  * Description:
1648  *   blk_start_queue() will clear the stop flag on the queue, and call
1649  *   the request_fn for the queue if it was in a stopped state when
1650  *   entered. Also see blk_stop_queue(). Queue lock must be held.
1651  **/
1652 void blk_start_queue(request_queue_t *q)
1653 {
1654         clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1655
1656         /*
1657          * one level of recursion is ok and is much faster than kicking
1658          * the unplug handling
1659          */
1660         if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1661                 q->request_fn(q);
1662                 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1663         } else {
1664                 blk_plug_device(q);
1665                 kblockd_schedule_work(&q->unplug_work);
1666         }
1667 }
1668
1669 EXPORT_SYMBOL(blk_start_queue);
1670
1671 /**
1672  * blk_stop_queue - stop a queue
1673  * @q:    The &request_queue_t in question
1674  *
1675  * Description:
1676  *   The Linux block layer assumes that a block driver will consume all
1677  *   entries on the request queue when the request_fn strategy is called.
1678  *   Often this will not happen, because of hardware limitations (queue
1679  *   depth settings). If a device driver gets a 'queue full' response,
1680  *   or if it simply chooses not to queue more I/O at one point, it can
1681  *   call this function to prevent the request_fn from being called until
1682  *   the driver has signalled it's ready to go again. This happens by calling
1683  *   blk_start_queue() to restart queue operations. Queue lock must be held.
1684  **/
1685 void blk_stop_queue(request_queue_t *q)
1686 {
1687         blk_remove_plug(q);
1688         set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1689 }
1690 EXPORT_SYMBOL(blk_stop_queue);
1691
1692 /**
1693  * blk_sync_queue - cancel any pending callbacks on a queue
1694  * @q: the queue
1695  *
1696  * Description:
1697  *     The block layer may perform asynchronous callback activity
1698  *     on a queue, such as calling the unplug function after a timeout.
1699  *     A block device may call blk_sync_queue to ensure that any
1700  *     such activity is cancelled, thus allowing it to release resources
1701  *     the the callbacks might use. The caller must already have made sure
1702  *     that its ->make_request_fn will not re-add plugging prior to calling
1703  *     this function.
1704  *
1705  */
1706 void blk_sync_queue(struct request_queue *q)
1707 {
1708         del_timer_sync(&q->unplug_timer);
1709         kblockd_flush();
1710 }
1711 EXPORT_SYMBOL(blk_sync_queue);
1712
1713 /**
1714  * blk_run_queue - run a single device queue
1715  * @q:  The queue to run
1716  */
1717 void blk_run_queue(struct request_queue *q)
1718 {
1719         unsigned long flags;
1720
1721         spin_lock_irqsave(q->queue_lock, flags);
1722         blk_remove_plug(q);
1723         if (!elv_queue_empty(q))
1724                 q->request_fn(q);
1725         spin_unlock_irqrestore(q->queue_lock, flags);
1726 }
1727 EXPORT_SYMBOL(blk_run_queue);
1728
1729 /**
1730  * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1731  * @q:    the request queue to be released
1732  *
1733  * Description:
1734  *     blk_cleanup_queue is the pair to blk_init_queue() or
1735  *     blk_queue_make_request().  It should be called when a request queue is
1736  *     being released; typically when a block device is being de-registered.
1737  *     Currently, its primary task it to free all the &struct request
1738  *     structures that were allocated to the queue and the queue itself.
1739  *
1740  * Caveat:
1741  *     Hopefully the low level driver will have finished any
1742  *     outstanding requests first...
1743  **/
1744 void blk_cleanup_queue(request_queue_t * q)
1745 {
1746         struct request_list *rl = &q->rq;
1747
1748         if (!atomic_dec_and_test(&q->refcnt))
1749                 return;
1750
1751         if (q->elevator)
1752                 elevator_exit(q->elevator);
1753
1754         blk_sync_queue(q);
1755
1756         if (rl->rq_pool)
1757                 mempool_destroy(rl->rq_pool);
1758
1759         if (q->queue_tags)
1760                 __blk_queue_free_tags(q);
1761
1762         kmem_cache_free(requestq_cachep, q);
1763 }
1764
1765 EXPORT_SYMBOL(blk_cleanup_queue);
1766
1767 static int blk_init_free_list(request_queue_t *q)
1768 {
1769         struct request_list *rl = &q->rq;
1770
1771         rl->count[READ] = rl->count[WRITE] = 0;
1772         rl->starved[READ] = rl->starved[WRITE] = 0;
1773         rl->elvpriv = 0;
1774         init_waitqueue_head(&rl->wait[READ]);
1775         init_waitqueue_head(&rl->wait[WRITE]);
1776
1777         rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1778                                 mempool_free_slab, request_cachep, q->node);
1779
1780         if (!rl->rq_pool)
1781                 return -ENOMEM;
1782
1783         return 0;
1784 }
1785
1786 request_queue_t *blk_alloc_queue(gfp_t gfp_mask)
1787 {
1788         return blk_alloc_queue_node(gfp_mask, -1);
1789 }
1790 EXPORT_SYMBOL(blk_alloc_queue);
1791
1792 request_queue_t *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
1793 {
1794         request_queue_t *q;
1795
1796         q = kmem_cache_alloc_node(requestq_cachep, gfp_mask, node_id);
1797         if (!q)
1798                 return NULL;
1799
1800         memset(q, 0, sizeof(*q));
1801         init_timer(&q->unplug_timer);
1802         atomic_set(&q->refcnt, 1);
1803
1804         q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1805         q->backing_dev_info.unplug_io_data = q;
1806
1807         return q;
1808 }
1809 EXPORT_SYMBOL(blk_alloc_queue_node);
1810
1811 /**
1812  * blk_init_queue  - prepare a request queue for use with a block device
1813  * @rfn:  The function to be called to process requests that have been
1814  *        placed on the queue.
1815  * @lock: Request queue spin lock
1816  *
1817  * Description:
1818  *    If a block device wishes to use the standard request handling procedures,
1819  *    which sorts requests and coalesces adjacent requests, then it must
1820  *    call blk_init_queue().  The function @rfn will be called when there
1821  *    are requests on the queue that need to be processed.  If the device
1822  *    supports plugging, then @rfn may not be called immediately when requests
1823  *    are available on the queue, but may be called at some time later instead.
1824  *    Plugged queues are generally unplugged when a buffer belonging to one
1825  *    of the requests on the queue is needed, or due to memory pressure.
1826  *
1827  *    @rfn is not required, or even expected, to remove all requests off the
1828  *    queue, but only as many as it can handle at a time.  If it does leave
1829  *    requests on the queue, it is responsible for arranging that the requests
1830  *    get dealt with eventually.
1831  *
1832  *    The queue spin lock must be held while manipulating the requests on the
1833  *    request queue.
1834  *
1835  *    Function returns a pointer to the initialized request queue, or NULL if
1836  *    it didn't succeed.
1837  *
1838  * Note:
1839  *    blk_init_queue() must be paired with a blk_cleanup_queue() call
1840  *    when the block device is deactivated (such as at module unload).
1841  **/
1842
1843 request_queue_t *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1844 {
1845         return blk_init_queue_node(rfn, lock, -1);
1846 }
1847 EXPORT_SYMBOL(blk_init_queue);
1848
1849 request_queue_t *
1850 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1851 {
1852         request_queue_t *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1853
1854         if (!q)
1855                 return NULL;
1856
1857         q->node = node_id;
1858         if (blk_init_free_list(q))
1859                 goto out_init;
1860
1861         /*
1862          * if caller didn't supply a lock, they get per-queue locking with
1863          * our embedded lock
1864          */
1865         if (!lock) {
1866                 spin_lock_init(&q->__queue_lock);
1867                 lock = &q->__queue_lock;
1868         }
1869
1870         q->request_fn           = rfn;
1871         q->back_merge_fn        = ll_back_merge_fn;
1872         q->front_merge_fn       = ll_front_merge_fn;
1873         q->merge_requests_fn    = ll_merge_requests_fn;
1874         q->prep_rq_fn           = NULL;
1875         q->unplug_fn            = generic_unplug_device;
1876         q->queue_flags          = (1 << QUEUE_FLAG_CLUSTER);
1877         q->queue_lock           = lock;
1878
1879         blk_queue_segment_boundary(q, 0xffffffff);
1880
1881         blk_queue_make_request(q, __make_request);
1882         blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1883
1884         blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1885         blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1886
1887         /*
1888          * all done
1889          */
1890         if (!elevator_init(q, NULL)) {
1891                 blk_queue_congestion_threshold(q);
1892                 return q;
1893         }
1894
1895         blk_cleanup_queue(q);
1896 out_init:
1897         kmem_cache_free(requestq_cachep, q);
1898         return NULL;
1899 }
1900 EXPORT_SYMBOL(blk_init_queue_node);
1901
1902 int blk_get_queue(request_queue_t *q)
1903 {
1904         if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1905                 atomic_inc(&q->refcnt);
1906                 return 0;
1907         }
1908
1909         return 1;
1910 }
1911
1912 EXPORT_SYMBOL(blk_get_queue);
1913
1914 static inline void blk_free_request(request_queue_t *q, struct request *rq)
1915 {
1916         if (rq->flags & REQ_ELVPRIV)
1917                 elv_put_request(q, rq);
1918         mempool_free(rq, q->rq.rq_pool);
1919 }
1920
1921 static inline struct request *
1922 blk_alloc_request(request_queue_t *q, int rw, struct bio *bio,
1923                   int priv, gfp_t gfp_mask)
1924 {
1925         struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1926
1927         if (!rq)
1928                 return NULL;
1929
1930         /*
1931          * first three bits are identical in rq->flags and bio->bi_rw,
1932          * see bio.h and blkdev.h
1933          */
1934         rq->flags = rw;
1935
1936         if (priv) {
1937                 if (unlikely(elv_set_request(q, rq, bio, gfp_mask))) {
1938                         mempool_free(rq, q->rq.rq_pool);
1939                         return NULL;
1940                 }
1941                 rq->flags |= REQ_ELVPRIV;
1942         }
1943
1944         return rq;
1945 }
1946
1947 /*
1948  * ioc_batching returns true if the ioc is a valid batching request and
1949  * should be given priority access to a request.
1950  */
1951 static inline int ioc_batching(request_queue_t *q, struct io_context *ioc)
1952 {
1953         if (!ioc)
1954                 return 0;
1955
1956         /*
1957          * Make sure the process is able to allocate at least 1 request
1958          * even if the batch times out, otherwise we could theoretically
1959          * lose wakeups.
1960          */
1961         return ioc->nr_batch_requests == q->nr_batching ||
1962                 (ioc->nr_batch_requests > 0
1963                 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
1964 }
1965
1966 /*
1967  * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
1968  * will cause the process to be a "batcher" on all queues in the system. This
1969  * is the behaviour we want though - once it gets a wakeup it should be given
1970  * a nice run.
1971  */
1972 static void ioc_set_batching(request_queue_t *q, struct io_context *ioc)
1973 {
1974         if (!ioc || ioc_batching(q, ioc))
1975                 return;
1976
1977         ioc->nr_batch_requests = q->nr_batching;
1978         ioc->last_waited = jiffies;
1979 }
1980
1981 static void __freed_request(request_queue_t *q, int rw)
1982 {
1983         struct request_list *rl = &q->rq;
1984
1985         if (rl->count[rw] < queue_congestion_off_threshold(q))
1986                 clear_queue_congested(q, rw);
1987
1988         if (rl->count[rw] + 1 <= q->nr_requests) {
1989                 if (waitqueue_active(&rl->wait[rw]))
1990                         wake_up(&rl->wait[rw]);
1991
1992                 blk_clear_queue_full(q, rw);
1993         }
1994 }
1995
1996 /*
1997  * A request has just been released.  Account for it, update the full and
1998  * congestion status, wake up any waiters.   Called under q->queue_lock.
1999  */
2000 static void freed_request(request_queue_t *q, int rw, int priv)
2001 {
2002         struct request_list *rl = &q->rq;
2003
2004         rl->count[rw]--;
2005         if (priv)
2006                 rl->elvpriv--;
2007
2008         __freed_request(q, rw);
2009
2010         if (unlikely(rl->starved[rw ^ 1]))
2011                 __freed_request(q, rw ^ 1);
2012 }
2013
2014 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2015 /*
2016  * Get a free request, queue_lock must be held.
2017  * Returns NULL on failure, with queue_lock held.
2018  * Returns !NULL on success, with queue_lock *not held*.
2019  */
2020 static struct request *get_request(request_queue_t *q, int rw, struct bio *bio,
2021                                    gfp_t gfp_mask)
2022 {
2023         struct request *rq = NULL;
2024         struct request_list *rl = &q->rq;
2025         struct io_context *ioc = NULL;
2026         int may_queue, priv;
2027
2028         may_queue = elv_may_queue(q, rw, bio);
2029         if (may_queue == ELV_MQUEUE_NO)
2030                 goto rq_starved;
2031
2032         if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
2033                 if (rl->count[rw]+1 >= q->nr_requests) {
2034                         ioc = current_io_context(GFP_ATOMIC);
2035                         /*
2036                          * The queue will fill after this allocation, so set
2037                          * it as full, and mark this process as "batching".
2038                          * This process will be allowed to complete a batch of
2039                          * requests, others will be blocked.
2040                          */
2041                         if (!blk_queue_full(q, rw)) {
2042                                 ioc_set_batching(q, ioc);
2043                                 blk_set_queue_full(q, rw);
2044                         } else {
2045                                 if (may_queue != ELV_MQUEUE_MUST
2046                                                 && !ioc_batching(q, ioc)) {
2047                                         /*
2048                                          * The queue is full and the allocating
2049                                          * process is not a "batcher", and not
2050                                          * exempted by the IO scheduler
2051                                          */
2052                                         goto out;
2053                                 }
2054                         }
2055                 }
2056                 set_queue_congested(q, rw);
2057         }
2058
2059         /*
2060          * Only allow batching queuers to allocate up to 50% over the defined
2061          * limit of requests, otherwise we could have thousands of requests
2062          * allocated with any setting of ->nr_requests
2063          */
2064         if (rl->count[rw] >= (3 * q->nr_requests / 2))
2065                 goto out;
2066
2067         rl->count[rw]++;
2068         rl->starved[rw] = 0;
2069
2070         priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
2071         if (priv)
2072                 rl->elvpriv++;
2073
2074         spin_unlock_irq(q->queue_lock);
2075
2076         rq = blk_alloc_request(q, rw, bio, priv, gfp_mask);
2077         if (unlikely(!rq)) {
2078                 /*
2079                  * Allocation failed presumably due to memory. Undo anything
2080                  * we might have messed up.
2081                  *
2082                  * Allocating task should really be put onto the front of the
2083                  * wait queue, but this is pretty rare.
2084                  */
2085                 spin_lock_irq(q->queue_lock);
2086                 freed_request(q, rw, priv);
2087
2088                 /*
2089                  * in the very unlikely event that allocation failed and no
2090                  * requests for this direction was pending, mark us starved
2091                  * so that freeing of a request in the other direction will
2092                  * notice us. another possible fix would be to split the
2093                  * rq mempool into READ and WRITE
2094                  */
2095 rq_starved:
2096                 if (unlikely(rl->count[rw] == 0))
2097                         rl->starved[rw] = 1;
2098
2099                 goto out;
2100         }
2101
2102         /*
2103          * ioc may be NULL here, and ioc_batching will be false. That's
2104          * OK, if the queue is under the request limit then requests need
2105          * not count toward the nr_batch_requests limit. There will always
2106          * be some limit enforced by BLK_BATCH_TIME.
2107          */
2108         if (ioc_batching(q, ioc))
2109                 ioc->nr_batch_requests--;
2110         
2111         rq_init(q, rq);
2112         rq->rl = rl;
2113 out:
2114         return rq;
2115 }
2116
2117 /*
2118  * No available requests for this queue, unplug the device and wait for some
2119  * requests to become available.
2120  *
2121  * Called with q->queue_lock held, and returns with it unlocked.
2122  */
2123 static struct request *get_request_wait(request_queue_t *q, int rw,
2124                                         struct bio *bio)
2125 {
2126         struct request *rq;
2127
2128         rq = get_request(q, rw, bio, GFP_NOIO);
2129         while (!rq) {
2130                 DEFINE_WAIT(wait);
2131                 struct request_list *rl = &q->rq;
2132
2133                 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2134                                 TASK_UNINTERRUPTIBLE);
2135
2136                 rq = get_request(q, rw, bio, GFP_NOIO);
2137
2138                 if (!rq) {
2139                         struct io_context *ioc;
2140
2141                         __generic_unplug_device(q);
2142                         spin_unlock_irq(q->queue_lock);
2143                         io_schedule();
2144
2145                         /*
2146                          * After sleeping, we become a "batching" process and
2147                          * will be able to allocate at least one request, and
2148                          * up to a big batch of them for a small period time.
2149                          * See ioc_batching, ioc_set_batching
2150                          */
2151                         ioc = current_io_context(GFP_NOIO);
2152                         ioc_set_batching(q, ioc);
2153
2154                         spin_lock_irq(q->queue_lock);
2155                 }
2156                 finish_wait(&rl->wait[rw], &wait);
2157         }
2158
2159         return rq;
2160 }
2161
2162 struct request *blk_get_request(request_queue_t *q, int rw, gfp_t gfp_mask)
2163 {
2164         struct request *rq;
2165
2166         BUG_ON(rw != READ && rw != WRITE);
2167
2168         spin_lock_irq(q->queue_lock);
2169         if (gfp_mask & __GFP_WAIT) {
2170                 rq = get_request_wait(q, rw, NULL);
2171         } else {
2172                 rq = get_request(q, rw, NULL, gfp_mask);
2173                 if (!rq)
2174                         spin_unlock_irq(q->queue_lock);
2175         }
2176         /* q->queue_lock is unlocked at this point */
2177
2178         return rq;
2179 }
2180 EXPORT_SYMBOL(blk_get_request);
2181
2182 /**
2183  * blk_requeue_request - put a request back on queue
2184  * @q:          request queue where request should be inserted
2185  * @rq:         request to be inserted
2186  *
2187  * Description:
2188  *    Drivers often keep queueing requests until the hardware cannot accept
2189  *    more, when that condition happens we need to put the request back
2190  *    on the queue. Must be called with queue lock held.
2191  */
2192 void blk_requeue_request(request_queue_t *q, struct request *rq)
2193 {
2194         if (blk_rq_tagged(rq))
2195                 blk_queue_end_tag(q, rq);
2196
2197         elv_requeue_request(q, rq);
2198 }
2199
2200 EXPORT_SYMBOL(blk_requeue_request);
2201
2202 /**
2203  * blk_insert_request - insert a special request in to a request queue
2204  * @q:          request queue where request should be inserted
2205  * @rq:         request to be inserted
2206  * @at_head:    insert request at head or tail of queue
2207  * @data:       private data
2208  *
2209  * Description:
2210  *    Many block devices need to execute commands asynchronously, so they don't
2211  *    block the whole kernel from preemption during request execution.  This is
2212  *    accomplished normally by inserting aritficial requests tagged as
2213  *    REQ_SPECIAL in to the corresponding request queue, and letting them be
2214  *    scheduled for actual execution by the request queue.
2215  *
2216  *    We have the option of inserting the head or the tail of the queue.
2217  *    Typically we use the tail for new ioctls and so forth.  We use the head
2218  *    of the queue for things like a QUEUE_FULL message from a device, or a
2219  *    host that is unable to accept a particular command.
2220  */
2221 void blk_insert_request(request_queue_t *q, struct request *rq,
2222                         int at_head, void *data)
2223 {
2224         int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2225         unsigned long flags;
2226
2227         /*
2228          * tell I/O scheduler that this isn't a regular read/write (ie it
2229          * must not attempt merges on this) and that it acts as a soft
2230          * barrier
2231          */
2232         rq->flags |= REQ_SPECIAL | REQ_SOFTBARRIER;
2233
2234         rq->special = data;
2235
2236         spin_lock_irqsave(q->queue_lock, flags);
2237
2238         /*
2239          * If command is tagged, release the tag
2240          */
2241         if (blk_rq_tagged(rq))
2242                 blk_queue_end_tag(q, rq);
2243
2244         drive_stat_acct(rq, rq->nr_sectors, 1);
2245         __elv_add_request(q, rq, where, 0);
2246
2247         if (blk_queue_plugged(q))
2248                 __generic_unplug_device(q);
2249         else
2250                 q->request_fn(q);
2251         spin_unlock_irqrestore(q->queue_lock, flags);
2252 }
2253
2254 EXPORT_SYMBOL(blk_insert_request);
2255
2256 /**
2257  * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2258  * @q:          request queue where request should be inserted
2259  * @rq:         request structure to fill
2260  * @ubuf:       the user buffer
2261  * @len:        length of user data
2262  *
2263  * Description:
2264  *    Data will be mapped directly for zero copy io, if possible. Otherwise
2265  *    a kernel bounce buffer is used.
2266  *
2267  *    A matching blk_rq_unmap_user() must be issued at the end of io, while
2268  *    still in process context.
2269  *
2270  *    Note: The mapped bio may need to be bounced through blk_queue_bounce()
2271  *    before being submitted to the device, as pages mapped may be out of
2272  *    reach. It's the callers responsibility to make sure this happens. The
2273  *    original bio must be passed back in to blk_rq_unmap_user() for proper
2274  *    unmapping.
2275  */
2276 int blk_rq_map_user(request_queue_t *q, struct request *rq, void __user *ubuf,
2277                     unsigned int len)
2278 {
2279         unsigned long uaddr;
2280         struct bio *bio;
2281         int reading;
2282
2283         if (len > (q->max_hw_sectors << 9))
2284                 return -EINVAL;
2285         if (!len || !ubuf)
2286                 return -EINVAL;
2287
2288         reading = rq_data_dir(rq) == READ;
2289
2290         /*
2291          * if alignment requirement is satisfied, map in user pages for
2292          * direct dma. else, set up kernel bounce buffers
2293          */
2294         uaddr = (unsigned long) ubuf;
2295         if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2296                 bio = bio_map_user(q, NULL, uaddr, len, reading);
2297         else
2298                 bio = bio_copy_user(q, uaddr, len, reading);
2299
2300         if (!IS_ERR(bio)) {
2301                 rq->bio = rq->biotail = bio;
2302                 blk_rq_bio_prep(q, rq, bio);
2303
2304                 rq->buffer = rq->data = NULL;
2305                 rq->data_len = len;
2306                 return 0;
2307         }
2308
2309         /*
2310          * bio is the err-ptr
2311          */
2312         return PTR_ERR(bio);
2313 }
2314
2315 EXPORT_SYMBOL(blk_rq_map_user);
2316
2317 /**
2318  * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2319  * @q:          request queue where request should be inserted
2320  * @rq:         request to map data to
2321  * @iov:        pointer to the iovec
2322  * @iov_count:  number of elements in the iovec
2323  *
2324  * Description:
2325  *    Data will be mapped directly for zero copy io, if possible. Otherwise
2326  *    a kernel bounce buffer is used.
2327  *
2328  *    A matching blk_rq_unmap_user() must be issued at the end of io, while
2329  *    still in process context.
2330  *
2331  *    Note: The mapped bio may need to be bounced through blk_queue_bounce()
2332  *    before being submitted to the device, as pages mapped may be out of
2333  *    reach. It's the callers responsibility to make sure this happens. The
2334  *    original bio must be passed back in to blk_rq_unmap_user() for proper
2335  *    unmapping.
2336  */
2337 int blk_rq_map_user_iov(request_queue_t *q, struct request *rq,
2338                         struct sg_iovec *iov, int iov_count)
2339 {
2340         struct bio *bio;
2341
2342         if (!iov || iov_count <= 0)
2343                 return -EINVAL;
2344
2345         /* we don't allow misaligned data like bio_map_user() does.  If the
2346          * user is using sg, they're expected to know the alignment constraints
2347          * and respect them accordingly */
2348         bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2349         if (IS_ERR(bio))
2350                 return PTR_ERR(bio);
2351
2352         rq->bio = rq->biotail = bio;
2353         blk_rq_bio_prep(q, rq, bio);
2354         rq->buffer = rq->data = NULL;
2355         rq->data_len = bio->bi_size;
2356         return 0;
2357 }
2358
2359 EXPORT_SYMBOL(blk_rq_map_user_iov);
2360
2361 /**
2362  * blk_rq_unmap_user - unmap a request with user data
2363  * @bio:        bio to be unmapped
2364  * @ulen:       length of user buffer
2365  *
2366  * Description:
2367  *    Unmap a bio previously mapped by blk_rq_map_user().
2368  */
2369 int blk_rq_unmap_user(struct bio *bio, unsigned int ulen)
2370 {
2371         int ret = 0;
2372
2373         if (bio) {
2374                 if (bio_flagged(bio, BIO_USER_MAPPED))
2375                         bio_unmap_user(bio);
2376                 else
2377                         ret = bio_uncopy_user(bio);
2378         }
2379
2380         return 0;
2381 }
2382
2383 EXPORT_SYMBOL(blk_rq_unmap_user);
2384
2385 /**
2386  * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2387  * @q:          request queue where request should be inserted
2388  * @rq:         request to fill
2389  * @kbuf:       the kernel buffer
2390  * @len:        length of user data
2391  * @gfp_mask:   memory allocation flags
2392  */
2393 int blk_rq_map_kern(request_queue_t *q, struct request *rq, void *kbuf,
2394                     unsigned int len, gfp_t gfp_mask)
2395 {
2396         struct bio *bio;
2397
2398         if (len > (q->max_hw_sectors << 9))
2399                 return -EINVAL;
2400         if (!len || !kbuf)
2401                 return -EINVAL;
2402
2403         bio = bio_map_kern(q, kbuf, len, gfp_mask);
2404         if (IS_ERR(bio))
2405                 return PTR_ERR(bio);
2406
2407         if (rq_data_dir(rq) == WRITE)
2408                 bio->bi_rw |= (1 << BIO_RW);
2409
2410         rq->bio = rq->biotail = bio;
2411         blk_rq_bio_prep(q, rq, bio);
2412
2413         rq->buffer = rq->data = NULL;
2414         rq->data_len = len;
2415         return 0;
2416 }
2417
2418 EXPORT_SYMBOL(blk_rq_map_kern);
2419
2420 /**
2421  * blk_execute_rq_nowait - insert a request into queue for execution
2422  * @q:          queue to insert the request in
2423  * @bd_disk:    matching gendisk
2424  * @rq:         request to insert
2425  * @at_head:    insert request at head or tail of queue
2426  * @done:       I/O completion handler
2427  *
2428  * Description:
2429  *    Insert a fully prepared request at the back of the io scheduler queue
2430  *    for execution.  Don't wait for completion.
2431  */
2432 void blk_execute_rq_nowait(request_queue_t *q, struct gendisk *bd_disk,
2433                            struct request *rq, int at_head,
2434                            rq_end_io_fn *done)
2435 {
2436         int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2437
2438         rq->rq_disk = bd_disk;
2439         rq->flags |= REQ_NOMERGE;
2440         rq->end_io = done;
2441         elv_add_request(q, rq, where, 1);
2442         generic_unplug_device(q);
2443 }
2444
2445 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
2446
2447 /**
2448  * blk_execute_rq - insert a request into queue for execution
2449  * @q:          queue to insert the request in
2450  * @bd_disk:    matching gendisk
2451  * @rq:         request to insert
2452  * @at_head:    insert request at head or tail of queue
2453  *
2454  * Description:
2455  *    Insert a fully prepared request at the back of the io scheduler queue
2456  *    for execution and wait for completion.
2457  */
2458 int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
2459                    struct request *rq, int at_head)
2460 {
2461         DECLARE_COMPLETION(wait);
2462         char sense[SCSI_SENSE_BUFFERSIZE];
2463         int err = 0;
2464
2465         /*
2466          * we need an extra reference to the request, so we can look at
2467          * it after io completion
2468          */
2469         rq->ref_count++;
2470
2471         if (!rq->sense) {
2472                 memset(sense, 0, sizeof(sense));
2473                 rq->sense = sense;
2474                 rq->sense_len = 0;
2475         }
2476
2477         rq->waiting = &wait;
2478         blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
2479         wait_for_completion(&wait);
2480         rq->waiting = NULL;
2481
2482         if (rq->errors)
2483                 err = -EIO;
2484
2485         return err;
2486 }
2487
2488 EXPORT_SYMBOL(blk_execute_rq);
2489
2490 /**
2491  * blkdev_issue_flush - queue a flush
2492  * @bdev:       blockdev to issue flush for
2493  * @error_sector:       error sector
2494  *
2495  * Description:
2496  *    Issue a flush for the block device in question. Caller can supply
2497  *    room for storing the error offset in case of a flush error, if they
2498  *    wish to.  Caller must run wait_for_completion() on its own.
2499  */
2500 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2501 {
2502         request_queue_t *q;
2503
2504         if (bdev->bd_disk == NULL)
2505                 return -ENXIO;
2506
2507         q = bdev_get_queue(bdev);
2508         if (!q)
2509                 return -ENXIO;
2510         if (!q->issue_flush_fn)
2511                 return -EOPNOTSUPP;
2512
2513         return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2514 }
2515
2516 EXPORT_SYMBOL(blkdev_issue_flush);
2517
2518 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2519 {
2520         int rw = rq_data_dir(rq);
2521
2522         if (!blk_fs_request(rq) || !rq->rq_disk)
2523                 return;
2524
2525         if (!new_io) {
2526                 __disk_stat_inc(rq->rq_disk, merges[rw]);
2527         } else {
2528                 disk_round_stats(rq->rq_disk);
2529                 rq->rq_disk->in_flight++;
2530         }
2531 }
2532
2533 /*
2534  * add-request adds a request to the linked list.
2535  * queue lock is held and interrupts disabled, as we muck with the
2536  * request queue list.
2537  */
2538 static inline void add_request(request_queue_t * q, struct request * req)
2539 {
2540         drive_stat_acct(req, req->nr_sectors, 1);
2541
2542         if (q->activity_fn)
2543                 q->activity_fn(q->activity_data, rq_data_dir(req));
2544
2545         /*
2546          * elevator indicated where it wants this request to be
2547          * inserted at elevator_merge time
2548          */
2549         __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2550 }
2551  
2552 /*
2553  * disk_round_stats()   - Round off the performance stats on a struct
2554  * disk_stats.
2555  *
2556  * The average IO queue length and utilisation statistics are maintained
2557  * by observing the current state of the queue length and the amount of
2558  * time it has been in this state for.
2559  *
2560  * Normally, that accounting is done on IO completion, but that can result
2561  * in more than a second's worth of IO being accounted for within any one
2562  * second, leading to >100% utilisation.  To deal with that, we call this
2563  * function to do a round-off before returning the results when reading
2564  * /proc/diskstats.  This accounts immediately for all queue usage up to
2565  * the current jiffies and restarts the counters again.
2566  */
2567 void disk_round_stats(struct gendisk *disk)
2568 {
2569         unsigned long now = jiffies;
2570
2571         if (now == disk->stamp)
2572                 return;
2573
2574         if (disk->in_flight) {
2575                 __disk_stat_add(disk, time_in_queue,
2576                                 disk->in_flight * (now - disk->stamp));
2577                 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2578         }
2579         disk->stamp = now;
2580 }
2581
2582 EXPORT_SYMBOL_GPL(disk_round_stats);
2583
2584 /*
2585  * queue lock must be held
2586  */
2587 void __blk_put_request(request_queue_t *q, struct request *req)
2588 {
2589         struct request_list *rl = req->rl;
2590
2591         if (unlikely(!q))
2592                 return;
2593         if (unlikely(--req->ref_count))
2594                 return;
2595
2596         elv_completed_request(q, req);
2597
2598         req->rq_status = RQ_INACTIVE;
2599         req->rl = NULL;
2600
2601         /*
2602          * Request may not have originated from ll_rw_blk. if not,
2603          * it didn't come out of our reserved rq pools
2604          */
2605         if (rl) {
2606                 int rw = rq_data_dir(req);
2607                 int priv = req->flags & REQ_ELVPRIV;
2608
2609                 BUG_ON(!list_empty(&req->queuelist));
2610
2611                 blk_free_request(q, req);
2612                 freed_request(q, rw, priv);
2613         }
2614 }
2615
2616 EXPORT_SYMBOL_GPL(__blk_put_request);
2617
2618 void blk_put_request(struct request *req)
2619 {
2620         unsigned long flags;
2621         request_queue_t *q = req->q;
2622
2623         /*
2624          * Gee, IDE calls in w/ NULL q.  Fix IDE and remove the
2625          * following if (q) test.
2626          */
2627         if (q) {
2628                 spin_lock_irqsave(q->queue_lock, flags);
2629                 __blk_put_request(q, req);
2630                 spin_unlock_irqrestore(q->queue_lock, flags);
2631         }
2632 }
2633
2634 EXPORT_SYMBOL(blk_put_request);
2635
2636 /**
2637  * blk_end_sync_rq - executes a completion event on a request
2638  * @rq: request to complete
2639  * @error: end io status of the request
2640  */
2641 void blk_end_sync_rq(struct request *rq, int error)
2642 {
2643         struct completion *waiting = rq->waiting;
2644
2645         rq->waiting = NULL;
2646         __blk_put_request(rq->q, rq);
2647
2648         /*
2649          * complete last, if this is a stack request the process (and thus
2650          * the rq pointer) could be invalid right after this complete()
2651          */
2652         complete(waiting);
2653 }
2654 EXPORT_SYMBOL(blk_end_sync_rq);
2655
2656 /**
2657  * blk_congestion_wait - wait for a queue to become uncongested
2658  * @rw: READ or WRITE
2659  * @timeout: timeout in jiffies
2660  *
2661  * Waits for up to @timeout jiffies for a queue (any queue) to exit congestion.
2662  * If no queues are congested then just wait for the next request to be
2663  * returned.
2664  */
2665 long blk_congestion_wait(int rw, long timeout)
2666 {
2667         long ret;
2668         DEFINE_WAIT(wait);
2669         wait_queue_head_t *wqh = &congestion_wqh[rw];
2670
2671         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
2672         ret = io_schedule_timeout(timeout);
2673         finish_wait(wqh, &wait);
2674         return ret;
2675 }
2676
2677 EXPORT_SYMBOL(blk_congestion_wait);
2678
2679 /*
2680  * Has to be called with the request spinlock acquired
2681  */
2682 static int attempt_merge(request_queue_t *q, struct request *req,
2683                           struct request *next)
2684 {
2685         if (!rq_mergeable(req) || !rq_mergeable(next))
2686                 return 0;
2687
2688         /*
2689          * not contigious
2690          */
2691         if (req->sector + req->nr_sectors != next->sector)
2692                 return 0;
2693
2694         if (rq_data_dir(req) != rq_data_dir(next)
2695             || req->rq_disk != next->rq_disk
2696             || next->waiting || next->special)
2697                 return 0;
2698
2699         /*
2700          * If we are allowed to merge, then append bio list
2701          * from next to rq and release next. merge_requests_fn
2702          * will have updated segment counts, update sector
2703          * counts here.
2704          */
2705         if (!q->merge_requests_fn(q, req, next))
2706                 return 0;
2707
2708         /*
2709          * At this point we have either done a back merge
2710          * or front merge. We need the smaller start_time of
2711          * the merged requests to be the current request
2712          * for accounting purposes.
2713          */
2714         if (time_after(req->start_time, next->start_time))
2715                 req->start_time = next->start_time;
2716
2717         req->biotail->bi_next = next->bio;
2718         req->biotail = next->biotail;
2719
2720         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2721
2722         elv_merge_requests(q, req, next);
2723
2724         if (req->rq_disk) {
2725                 disk_round_stats(req->rq_disk);
2726                 req->rq_disk->in_flight--;
2727         }
2728
2729         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2730
2731         __blk_put_request(q, next);
2732         return 1;
2733 }
2734
2735 static inline int attempt_back_merge(request_queue_t *q, struct request *rq)
2736 {
2737         struct request *next = elv_latter_request(q, rq);
2738
2739         if (next)
2740                 return attempt_merge(q, rq, next);
2741
2742         return 0;
2743 }
2744
2745 static inline int attempt_front_merge(request_queue_t *q, struct request *rq)
2746 {
2747         struct request *prev = elv_former_request(q, rq);
2748
2749         if (prev)
2750                 return attempt_merge(q, prev, rq);
2751
2752         return 0;
2753 }
2754
2755 static void init_request_from_bio(struct request *req, struct bio *bio)
2756 {
2757         req->flags |= REQ_CMD;
2758
2759         /*
2760          * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2761          */
2762         if (bio_rw_ahead(bio) || bio_failfast(bio))
2763                 req->flags |= REQ_FAILFAST;
2764
2765         /*
2766          * REQ_BARRIER implies no merging, but lets make it explicit
2767          */
2768         if (unlikely(bio_barrier(bio)))
2769                 req->flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2770
2771         req->errors = 0;
2772         req->hard_sector = req->sector = bio->bi_sector;
2773         req->hard_nr_sectors = req->nr_sectors = bio_sectors(bio);
2774         req->current_nr_sectors = req->hard_cur_sectors = bio_cur_sectors(bio);
2775         req->nr_phys_segments = bio_phys_segments(req->q, bio);
2776         req->nr_hw_segments = bio_hw_segments(req->q, bio);
2777         req->buffer = bio_data(bio);    /* see ->buffer comment above */
2778         req->waiting = NULL;
2779         req->bio = req->biotail = bio;
2780         req->ioprio = bio_prio(bio);
2781         req->rq_disk = bio->bi_bdev->bd_disk;
2782         req->start_time = jiffies;
2783 }
2784
2785 static int __make_request(request_queue_t *q, struct bio *bio)
2786 {
2787         struct request *req;
2788         int el_ret, rw, nr_sectors, cur_nr_sectors, barrier, err, sync;
2789         unsigned short prio;
2790         sector_t sector;
2791
2792         sector = bio->bi_sector;
2793         nr_sectors = bio_sectors(bio);
2794         cur_nr_sectors = bio_cur_sectors(bio);
2795         prio = bio_prio(bio);
2796
2797         rw = bio_data_dir(bio);
2798         sync = bio_sync(bio);
2799
2800         /*
2801          * low level driver can indicate that it wants pages above a
2802          * certain limit bounced to low memory (ie for highmem, or even
2803          * ISA dma in theory)
2804          */
2805         blk_queue_bounce(q, &bio);
2806
2807         spin_lock_prefetch(q->queue_lock);
2808
2809         barrier = bio_barrier(bio);
2810         if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
2811                 err = -EOPNOTSUPP;
2812                 goto end_io;
2813         }
2814
2815         spin_lock_irq(q->queue_lock);
2816
2817         if (unlikely(barrier) || elv_queue_empty(q))
2818                 goto get_rq;
2819
2820         el_ret = elv_merge(q, &req, bio);
2821         switch (el_ret) {
2822                 case ELEVATOR_BACK_MERGE:
2823                         BUG_ON(!rq_mergeable(req));
2824
2825                         if (!q->back_merge_fn(q, req, bio))
2826                                 break;
2827
2828                         req->biotail->bi_next = bio;
2829                         req->biotail = bio;
2830                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2831                         req->ioprio = ioprio_best(req->ioprio, prio);
2832                         drive_stat_acct(req, nr_sectors, 0);
2833                         if (!attempt_back_merge(q, req))
2834                                 elv_merged_request(q, req);
2835                         goto out;
2836
2837                 case ELEVATOR_FRONT_MERGE:
2838                         BUG_ON(!rq_mergeable(req));
2839
2840                         if (!q->front_merge_fn(q, req, bio))
2841                                 break;
2842
2843                         bio->bi_next = req->bio;
2844                         req->bio = bio;
2845
2846                         /*
2847                          * may not be valid. if the low level driver said
2848                          * it didn't need a bounce buffer then it better
2849                          * not touch req->buffer either...
2850                          */
2851                         req->buffer = bio_data(bio);
2852                         req->current_nr_sectors = cur_nr_sectors;
2853                         req->hard_cur_sectors = cur_nr_sectors;
2854                         req->sector = req->hard_sector = sector;
2855                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2856                         req->ioprio = ioprio_best(req->ioprio, prio);
2857                         drive_stat_acct(req, nr_sectors, 0);
2858                         if (!attempt_front_merge(q, req))
2859                                 elv_merged_request(q, req);
2860                         goto out;
2861
2862                 /* ELV_NO_MERGE: elevator says don't/can't merge. */
2863                 default:
2864                         ;
2865         }
2866
2867 get_rq:
2868         /*
2869          * Grab a free request. This is might sleep but can not fail.
2870          * Returns with the queue unlocked.
2871          */
2872         req = get_request_wait(q, rw, bio);
2873
2874         /*
2875          * After dropping the lock and possibly sleeping here, our request
2876          * may now be mergeable after it had proven unmergeable (above).
2877          * We don't worry about that case for efficiency. It won't happen
2878          * often, and the elevators are able to handle it.
2879          */
2880         init_request_from_bio(req, bio);
2881
2882         spin_lock_irq(q->queue_lock);
2883         if (elv_queue_empty(q))
2884                 blk_plug_device(q);
2885         add_request(q, req);
2886 out:
2887         if (sync)
2888                 __generic_unplug_device(q);
2889
2890         spin_unlock_irq(q->queue_lock);
2891         return 0;
2892
2893 end_io:
2894         bio_endio(bio, nr_sectors << 9, err);
2895         return 0;
2896 }
2897
2898 /*
2899  * If bio->bi_dev is a partition, remap the location
2900  */
2901 static inline void blk_partition_remap(struct bio *bio)
2902 {
2903         struct block_device *bdev = bio->bi_bdev;
2904
2905         if (bdev != bdev->bd_contains) {
2906                 struct hd_struct *p = bdev->bd_part;
2907                 const int rw = bio_data_dir(bio);
2908
2909                 p->sectors[rw] += bio_sectors(bio);
2910                 p->ios[rw]++;
2911
2912                 bio->bi_sector += p->start_sect;
2913                 bio->bi_bdev = bdev->bd_contains;
2914         }
2915 }
2916
2917 static void handle_bad_sector(struct bio *bio)
2918 {
2919         char b[BDEVNAME_SIZE];
2920
2921         printk(KERN_INFO "attempt to access beyond end of device\n");
2922         printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
2923                         bdevname(bio->bi_bdev, b),
2924                         bio->bi_rw,
2925                         (unsigned long long)bio->bi_sector + bio_sectors(bio),
2926                         (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
2927
2928         set_bit(BIO_EOF, &bio->bi_flags);
2929 }
2930
2931 /**
2932  * generic_make_request: hand a buffer to its device driver for I/O
2933  * @bio:  The bio describing the location in memory and on the device.
2934  *
2935  * generic_make_request() is used to make I/O requests of block
2936  * devices. It is passed a &struct bio, which describes the I/O that needs
2937  * to be done.
2938  *
2939  * generic_make_request() does not return any status.  The
2940  * success/failure status of the request, along with notification of
2941  * completion, is delivered asynchronously through the bio->bi_end_io
2942  * function described (one day) else where.
2943  *
2944  * The caller of generic_make_request must make sure that bi_io_vec
2945  * are set to describe the memory buffer, and that bi_dev and bi_sector are
2946  * set to describe the device address, and the
2947  * bi_end_io and optionally bi_private are set to describe how
2948  * completion notification should be signaled.
2949  *
2950  * generic_make_request and the drivers it calls may use bi_next if this
2951  * bio happens to be merged with someone else, and may change bi_dev and
2952  * bi_sector for remaps as it sees fit.  So the values of these fields
2953  * should NOT be depended on after the call to generic_make_request.
2954  */
2955 void generic_make_request(struct bio *bio)
2956 {
2957         request_queue_t *q;
2958         sector_t maxsector;
2959         int ret, nr_sectors = bio_sectors(bio);
2960
2961         might_sleep();
2962         /* Test device or partition size, when known. */
2963         maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
2964         if (maxsector) {
2965                 sector_t sector = bio->bi_sector;
2966
2967                 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
2968                         /*
2969                          * This may well happen - the kernel calls bread()
2970                          * without checking the size of the device, e.g., when
2971                          * mounting a device.
2972                          */
2973                         handle_bad_sector(bio);
2974                         goto end_io;
2975                 }
2976         }
2977
2978         /*
2979          * Resolve the mapping until finished. (drivers are
2980          * still free to implement/resolve their own stacking
2981          * by explicitly returning 0)
2982          *
2983          * NOTE: we don't repeat the blk_size check for each new device.
2984          * Stacking drivers are expected to know what they are doing.
2985          */
2986         do {
2987                 char b[BDEVNAME_SIZE];
2988
2989                 q = bdev_get_queue(bio->bi_bdev);
2990                 if (!q) {
2991                         printk(KERN_ERR
2992                                "generic_make_request: Trying to access "
2993                                 "nonexistent block-device %s (%Lu)\n",
2994                                 bdevname(bio->bi_bdev, b),
2995                                 (long long) bio->bi_sector);
2996 end_io:
2997                         bio_endio(bio, bio->bi_size, -EIO);
2998                         break;
2999                 }
3000
3001                 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
3002                         printk("bio too big device %s (%u > %u)\n", 
3003                                 bdevname(bio->bi_bdev, b),
3004                                 bio_sectors(bio),
3005                                 q->max_hw_sectors);
3006                         goto end_io;
3007                 }
3008
3009                 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
3010                         goto end_io;
3011
3012                 /*
3013                  * If this device has partitions, remap block n
3014                  * of partition p to block n+start(p) of the disk.
3015                  */
3016                 blk_partition_remap(bio);
3017
3018                 ret = q->make_request_fn(q, bio);
3019         } while (ret);
3020 }
3021
3022 EXPORT_SYMBOL(generic_make_request);
3023
3024 /**
3025  * submit_bio: submit a bio to the block device layer for I/O
3026  * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3027  * @bio: The &struct bio which describes the I/O
3028  *
3029  * submit_bio() is very similar in purpose to generic_make_request(), and
3030  * uses that function to do most of the work. Both are fairly rough
3031  * interfaces, @bio must be presetup and ready for I/O.
3032  *
3033  */
3034 void submit_bio(int rw, struct bio *bio)
3035 {
3036         int count = bio_sectors(bio);
3037
3038         BIO_BUG_ON(!bio->bi_size);
3039         BIO_BUG_ON(!bio->bi_io_vec);
3040         bio->bi_rw |= rw;
3041         if (rw & WRITE)
3042                 mod_page_state(pgpgout, count);
3043         else
3044                 mod_page_state(pgpgin, count);
3045
3046         if (unlikely(block_dump)) {
3047                 char b[BDEVNAME_SIZE];
3048                 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
3049                         current->comm, current->pid,
3050                         (rw & WRITE) ? "WRITE" : "READ",
3051                         (unsigned long long)bio->bi_sector,
3052                         bdevname(bio->bi_bdev,b));
3053         }
3054
3055         generic_make_request(bio);
3056 }
3057
3058 EXPORT_SYMBOL(submit_bio);
3059
3060 static void blk_recalc_rq_segments(struct request *rq)
3061 {
3062         struct bio *bio, *prevbio = NULL;
3063         int nr_phys_segs, nr_hw_segs;
3064         unsigned int phys_size, hw_size;
3065         request_queue_t *q = rq->q;
3066
3067         if (!rq->bio)
3068                 return;
3069
3070         phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
3071         rq_for_each_bio(bio, rq) {
3072                 /* Force bio hw/phys segs to be recalculated. */
3073                 bio->bi_flags &= ~(1 << BIO_SEG_VALID);
3074
3075                 nr_phys_segs += bio_phys_segments(q, bio);
3076                 nr_hw_segs += bio_hw_segments(q, bio);
3077                 if (prevbio) {
3078                         int pseg = phys_size + prevbio->bi_size + bio->bi_size;
3079                         int hseg = hw_size + prevbio->bi_size + bio->bi_size;
3080
3081                         if (blk_phys_contig_segment(q, prevbio, bio) &&
3082                             pseg <= q->max_segment_size) {
3083                                 nr_phys_segs--;
3084                                 phys_size += prevbio->bi_size + bio->bi_size;
3085                         } else
3086                                 phys_size = 0;
3087
3088                         if (blk_hw_contig_segment(q, prevbio, bio) &&
3089                             hseg <= q->max_segment_size) {
3090                                 nr_hw_segs--;
3091                                 hw_size += prevbio->bi_size + bio->bi_size;
3092                         } else
3093                                 hw_size = 0;
3094                 }
3095                 prevbio = bio;
3096         }
3097
3098         rq->nr_phys_segments = nr_phys_segs;
3099         rq->nr_hw_segments = nr_hw_segs;
3100 }
3101
3102 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
3103 {
3104         if (blk_fs_request(rq)) {
3105                 rq->hard_sector += nsect;
3106                 rq->hard_nr_sectors -= nsect;
3107
3108                 /*
3109                  * Move the I/O submission pointers ahead if required.
3110                  */
3111                 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3112                     (rq->sector <= rq->hard_sector)) {
3113                         rq->sector = rq->hard_sector;
3114                         rq->nr_sectors = rq->hard_nr_sectors;
3115                         rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3116                         rq->current_nr_sectors = rq->hard_cur_sectors;
3117                         rq->buffer = bio_data(rq->bio);
3118                 }
3119
3120                 /*
3121                  * if total number of sectors is less than the first segment
3122                  * size, something has gone terribly wrong
3123                  */
3124                 if (rq->nr_sectors < rq->current_nr_sectors) {
3125                         printk("blk: request botched\n");
3126                         rq->nr_sectors = rq->current_nr_sectors;
3127                 }
3128         }
3129 }
3130
3131 static int __end_that_request_first(struct request *req, int uptodate,
3132                                     int nr_bytes)
3133 {
3134         int total_bytes, bio_nbytes, error, next_idx = 0;
3135         struct bio *bio;
3136
3137         /*
3138          * extend uptodate bool to allow < 0 value to be direct io error
3139          */
3140         error = 0;
3141         if (end_io_error(uptodate))
3142                 error = !uptodate ? -EIO : uptodate;
3143
3144         /*
3145          * for a REQ_BLOCK_PC request, we want to carry any eventual
3146          * sense key with us all the way through
3147          */
3148         if (!blk_pc_request(req))
3149                 req->errors = 0;
3150
3151         if (!uptodate) {
3152                 if (blk_fs_request(req) && !(req->flags & REQ_QUIET))
3153                         printk("end_request: I/O error, dev %s, sector %llu\n",
3154                                 req->rq_disk ? req->rq_disk->disk_name : "?",
3155                                 (unsigned long long)req->sector);
3156         }
3157
3158         if (blk_fs_request(req) && req->rq_disk) {
3159                 const int rw = rq_data_dir(req);
3160
3161                 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
3162         }
3163
3164         total_bytes = bio_nbytes = 0;
3165         while ((bio = req->bio) != NULL) {
3166                 int nbytes;
3167
3168                 if (nr_bytes >= bio->bi_size) {
3169                         req->bio = bio->bi_next;
3170                         nbytes = bio->bi_size;
3171                         if (!ordered_bio_endio(req, bio, nbytes, error))
3172                                 bio_endio(bio, nbytes, error);
3173                         next_idx = 0;
3174                         bio_nbytes = 0;
3175                 } else {
3176                         int idx = bio->bi_idx + next_idx;
3177
3178                         if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3179                                 blk_dump_rq_flags(req, "__end_that");
3180                                 printk("%s: bio idx %d >= vcnt %d\n",
3181                                                 __FUNCTION__,
3182                                                 bio->bi_idx, bio->bi_vcnt);
3183                                 break;
3184                         }
3185
3186                         nbytes = bio_iovec_idx(bio, idx)->bv_len;
3187                         BIO_BUG_ON(nbytes > bio->bi_size);
3188
3189                         /*
3190                          * not a complete bvec done
3191                          */
3192                         if (unlikely(nbytes > nr_bytes)) {
3193                                 bio_nbytes += nr_bytes;
3194                                 total_bytes += nr_bytes;
3195                                 break;
3196                         }
3197
3198                         /*
3199                          * advance to the next vector
3200                          */
3201                         next_idx++;
3202                         bio_nbytes += nbytes;
3203                 }
3204
3205                 total_bytes += nbytes;
3206                 nr_bytes -= nbytes;
3207
3208                 if ((bio = req->bio)) {
3209                         /*
3210                          * end more in this run, or just return 'not-done'
3211                          */
3212                         if (unlikely(nr_bytes <= 0))
3213                                 break;
3214                 }
3215         }
3216
3217         /*
3218          * completely done
3219          */
3220         if (!req->bio)
3221                 return 0;
3222
3223         /*
3224          * if the request wasn't completed, update state
3225          */
3226         if (bio_nbytes) {
3227                 if (!ordered_bio_endio(req, bio, bio_nbytes, error))
3228                         bio_endio(bio, bio_nbytes, error);
3229                 bio->bi_idx += next_idx;
3230                 bio_iovec(bio)->bv_offset += nr_bytes;
3231                 bio_iovec(bio)->bv_len -= nr_bytes;
3232         }
3233
3234         blk_recalc_rq_sectors(req, total_bytes >> 9);
3235         blk_recalc_rq_segments(req);
3236         return 1;
3237 }
3238
3239 /**
3240  * end_that_request_first - end I/O on a request
3241  * @req:      the request being processed
3242  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3243  * @nr_sectors: number of sectors to end I/O on
3244  *
3245  * Description:
3246  *     Ends I/O on a number of sectors attached to @req, and sets it up
3247  *     for the next range of segments (if any) in the cluster.
3248  *
3249  * Return:
3250  *     0 - we are done with this request, call end_that_request_last()
3251  *     1 - still buffers pending for this request
3252  **/
3253 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3254 {
3255         return __end_that_request_first(req, uptodate, nr_sectors << 9);
3256 }
3257
3258 EXPORT_SYMBOL(end_that_request_first);
3259
3260 /**
3261  * end_that_request_chunk - end I/O on a request
3262  * @req:      the request being processed
3263  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3264  * @nr_bytes: number of bytes to complete
3265  *
3266  * Description:
3267  *     Ends I/O on a number of bytes attached to @req, and sets it up
3268  *     for the next range of segments (if any). Like end_that_request_first(),
3269  *     but deals with bytes instead of sectors.
3270  *
3271  * Return:
3272  *     0 - we are done with this request, call end_that_request_last()
3273  *     1 - still buffers pending for this request
3274  **/
3275 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3276 {
3277         return __end_that_request_first(req, uptodate, nr_bytes);
3278 }
3279
3280 EXPORT_SYMBOL(end_that_request_chunk);
3281
3282 /*
3283  * splice the completion data to a local structure and hand off to
3284  * process_completion_queue() to complete the requests
3285  */
3286 static void blk_done_softirq(struct softirq_action *h)
3287 {
3288         struct list_head *cpu_list;
3289         LIST_HEAD(local_list);
3290
3291         local_irq_disable();
3292         cpu_list = &__get_cpu_var(blk_cpu_done);
3293         list_splice_init(cpu_list, &local_list);
3294         local_irq_enable();
3295
3296         while (!list_empty(&local_list)) {
3297                 struct request *rq = list_entry(local_list.next, struct request, donelist);
3298
3299                 list_del_init(&rq->donelist);
3300                 rq->q->softirq_done_fn(rq);
3301         }
3302 }
3303
3304 #ifdef CONFIG_HOTPLUG_CPU
3305
3306 static int blk_cpu_notify(struct notifier_block *self, unsigned long action,
3307                           void *hcpu)
3308 {
3309         /*
3310          * If a CPU goes away, splice its entries to the current CPU
3311          * and trigger a run of the softirq
3312          */
3313         if (action == CPU_DEAD) {
3314                 int cpu = (unsigned long) hcpu;
3315
3316                 local_irq_disable();
3317                 list_splice_init(&per_cpu(blk_cpu_done, cpu),
3318                                  &__get_cpu_var(blk_cpu_done));
3319                 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3320                 local_irq_enable();
3321         }
3322
3323         return NOTIFY_OK;
3324 }
3325
3326
3327 static struct notifier_block __devinitdata blk_cpu_notifier = {
3328         .notifier_call  = blk_cpu_notify,
3329 };
3330
3331 #endif /* CONFIG_HOTPLUG_CPU */
3332
3333 /**
3334  * blk_complete_request - end I/O on a request
3335  * @req:      the request being processed
3336  *
3337  * Description:
3338  *     Ends all I/O on a request. It does not handle partial completions,
3339  *     unless the driver actually implements this in its completionc callback
3340  *     through requeueing. Theh actual completion happens out-of-order,
3341  *     through a softirq handler. The user must have registered a completion
3342  *     callback through blk_queue_softirq_done().
3343  **/
3344
3345 void blk_complete_request(struct request *req)
3346 {
3347         struct list_head *cpu_list;
3348         unsigned long flags;
3349
3350         BUG_ON(!req->q->softirq_done_fn);
3351                 
3352         local_irq_save(flags);
3353
3354         cpu_list = &__get_cpu_var(blk_cpu_done);
3355         list_add_tail(&req->donelist, cpu_list);
3356         raise_softirq_irqoff(BLOCK_SOFTIRQ);
3357
3358         local_irq_restore(flags);
3359 }
3360
3361 EXPORT_SYMBOL(blk_complete_request);
3362         
3363 /*
3364  * queue lock must be held
3365  */
3366 void end_that_request_last(struct request *req, int uptodate)
3367 {
3368         struct gendisk *disk = req->rq_disk;
3369         int error;
3370
3371         /*
3372          * extend uptodate bool to allow < 0 value to be direct io error
3373          */
3374         error = 0;
3375         if (end_io_error(uptodate))
3376                 error = !uptodate ? -EIO : uptodate;
3377
3378         if (unlikely(laptop_mode) && blk_fs_request(req))
3379                 laptop_io_completion();
3380
3381         if (disk && blk_fs_request(req)) {
3382                 unsigned long duration = jiffies - req->start_time;
3383                 const int rw = rq_data_dir(req);
3384
3385                 __disk_stat_inc(disk, ios[rw]);
3386                 __disk_stat_add(disk, ticks[rw], duration);
3387                 disk_round_stats(disk);
3388                 disk->in_flight--;
3389         }
3390         if (req->end_io)
3391                 req->end_io(req, error);
3392         else
3393                 __blk_put_request(req->q, req);
3394 }
3395
3396 EXPORT_SYMBOL(end_that_request_last);
3397
3398 void end_request(struct request *req, int uptodate)
3399 {
3400         if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3401                 add_disk_randomness(req->rq_disk);
3402                 blkdev_dequeue_request(req);
3403                 end_that_request_last(req, uptodate);
3404         }
3405 }
3406
3407 EXPORT_SYMBOL(end_request);
3408
3409 void blk_rq_bio_prep(request_queue_t *q, struct request *rq, struct bio *bio)
3410 {
3411         /* first three bits are identical in rq->flags and bio->bi_rw */
3412         rq->flags |= (bio->bi_rw & 7);
3413
3414         rq->nr_phys_segments = bio_phys_segments(q, bio);
3415         rq->nr_hw_segments = bio_hw_segments(q, bio);
3416         rq->current_nr_sectors = bio_cur_sectors(bio);
3417         rq->hard_cur_sectors = rq->current_nr_sectors;
3418         rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3419         rq->buffer = bio_data(bio);
3420
3421         rq->bio = rq->biotail = bio;
3422 }
3423
3424 EXPORT_SYMBOL(blk_rq_bio_prep);
3425
3426 int kblockd_schedule_work(struct work_struct *work)
3427 {
3428         return queue_work(kblockd_workqueue, work);
3429 }
3430
3431 EXPORT_SYMBOL(kblockd_schedule_work);
3432
3433 void kblockd_flush(void)
3434 {
3435         flush_workqueue(kblockd_workqueue);
3436 }
3437 EXPORT_SYMBOL(kblockd_flush);
3438
3439 int __init blk_dev_init(void)
3440 {
3441         int i;
3442
3443         kblockd_workqueue = create_workqueue("kblockd");
3444         if (!kblockd_workqueue)
3445                 panic("Failed to create kblockd\n");
3446
3447         request_cachep = kmem_cache_create("blkdev_requests",
3448                         sizeof(struct request), 0, SLAB_PANIC, NULL, NULL);
3449
3450         requestq_cachep = kmem_cache_create("blkdev_queue",
3451                         sizeof(request_queue_t), 0, SLAB_PANIC, NULL, NULL);
3452
3453         iocontext_cachep = kmem_cache_create("blkdev_ioc",
3454                         sizeof(struct io_context), 0, SLAB_PANIC, NULL, NULL);
3455
3456         for_each_cpu(i)
3457                 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3458
3459         open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
3460 #ifdef CONFIG_HOTPLUG_CPU
3461         register_cpu_notifier(&blk_cpu_notifier);
3462 #endif
3463
3464         blk_max_low_pfn = max_low_pfn;
3465         blk_max_pfn = max_pfn;
3466
3467         return 0;
3468 }
3469
3470 /*
3471  * IO Context helper functions
3472  */
3473 void put_io_context(struct io_context *ioc)
3474 {
3475         if (ioc == NULL)
3476                 return;
3477
3478         BUG_ON(atomic_read(&ioc->refcount) == 0);
3479
3480         if (atomic_dec_and_test(&ioc->refcount)) {
3481                 if (ioc->aic && ioc->aic->dtor)
3482                         ioc->aic->dtor(ioc->aic);
3483                 if (ioc->cic && ioc->cic->dtor)
3484                         ioc->cic->dtor(ioc->cic);
3485
3486                 kmem_cache_free(iocontext_cachep, ioc);
3487         }
3488 }
3489 EXPORT_SYMBOL(put_io_context);
3490
3491 /* Called by the exitting task */
3492 void exit_io_context(void)
3493 {
3494         unsigned long flags;
3495         struct io_context *ioc;
3496
3497         local_irq_save(flags);
3498         task_lock(current);
3499         ioc = current->io_context;
3500         current->io_context = NULL;
3501         ioc->task = NULL;
3502         task_unlock(current);
3503         local_irq_restore(flags);
3504
3505         if (ioc->aic && ioc->aic->exit)
3506                 ioc->aic->exit(ioc->aic);
3507         if (ioc->cic && ioc->cic->exit)
3508                 ioc->cic->exit(ioc->cic);
3509
3510         put_io_context(ioc);
3511 }
3512
3513 /*
3514  * If the current task has no IO context then create one and initialise it.
3515  * Otherwise, return its existing IO context.
3516  *
3517  * This returned IO context doesn't have a specifically elevated refcount,
3518  * but since the current task itself holds a reference, the context can be
3519  * used in general code, so long as it stays within `current` context.
3520  */
3521 struct io_context *current_io_context(gfp_t gfp_flags)
3522 {
3523         struct task_struct *tsk = current;
3524         struct io_context *ret;
3525
3526         ret = tsk->io_context;
3527         if (likely(ret))
3528                 return ret;
3529
3530         ret = kmem_cache_alloc(iocontext_cachep, gfp_flags);
3531         if (ret) {
3532                 atomic_set(&ret->refcount, 1);
3533                 ret->task = current;
3534                 ret->set_ioprio = NULL;
3535                 ret->last_waited = jiffies; /* doesn't matter... */
3536                 ret->nr_batch_requests = 0; /* because this is 0 */
3537                 ret->aic = NULL;
3538                 ret->cic = NULL;
3539                 tsk->io_context = ret;
3540         }
3541
3542         return ret;
3543 }
3544 EXPORT_SYMBOL(current_io_context);
3545
3546 /*
3547  * If the current task has no IO context then create one and initialise it.
3548  * If it does have a context, take a ref on it.
3549  *
3550  * This is always called in the context of the task which submitted the I/O.
3551  */
3552 struct io_context *get_io_context(gfp_t gfp_flags)
3553 {
3554         struct io_context *ret;
3555         ret = current_io_context(gfp_flags);
3556         if (likely(ret))
3557                 atomic_inc(&ret->refcount);
3558         return ret;
3559 }
3560 EXPORT_SYMBOL(get_io_context);
3561
3562 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3563 {
3564         struct io_context *src = *psrc;
3565         struct io_context *dst = *pdst;
3566
3567         if (src) {
3568                 BUG_ON(atomic_read(&src->refcount) == 0);
3569                 atomic_inc(&src->refcount);
3570                 put_io_context(dst);
3571                 *pdst = src;
3572         }
3573 }
3574 EXPORT_SYMBOL(copy_io_context);
3575
3576 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3577 {
3578         struct io_context *temp;
3579         temp = *ioc1;
3580         *ioc1 = *ioc2;
3581         *ioc2 = temp;
3582 }
3583 EXPORT_SYMBOL(swap_io_context);
3584
3585 /*
3586  * sysfs parts below
3587  */
3588 struct queue_sysfs_entry {
3589         struct attribute attr;
3590         ssize_t (*show)(struct request_queue *, char *);
3591         ssize_t (*store)(struct request_queue *, const char *, size_t);
3592 };
3593
3594 static ssize_t
3595 queue_var_show(unsigned int var, char *page)
3596 {
3597         return sprintf(page, "%d\n", var);
3598 }
3599
3600 static ssize_t
3601 queue_var_store(unsigned long *var, const char *page, size_t count)
3602 {
3603         char *p = (char *) page;
3604
3605         *var = simple_strtoul(p, &p, 10);
3606         return count;
3607 }
3608
3609 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3610 {
3611         return queue_var_show(q->nr_requests, (page));
3612 }
3613
3614 static ssize_t
3615 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3616 {
3617         struct request_list *rl = &q->rq;
3618
3619         int ret = queue_var_store(&q->nr_requests, page, count);
3620         if (q->nr_requests < BLKDEV_MIN_RQ)
3621                 q->nr_requests = BLKDEV_MIN_RQ;
3622         blk_queue_congestion_threshold(q);
3623
3624         if (rl->count[READ] >= queue_congestion_on_threshold(q))
3625                 set_queue_congested(q, READ);
3626         else if (rl->count[READ] < queue_congestion_off_threshold(q))
3627                 clear_queue_congested(q, READ);
3628
3629         if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3630                 set_queue_congested(q, WRITE);
3631         else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3632                 clear_queue_congested(q, WRITE);
3633
3634         if (rl->count[READ] >= q->nr_requests) {
3635                 blk_set_queue_full(q, READ);
3636         } else if (rl->count[READ]+1 <= q->nr_requests) {
3637                 blk_clear_queue_full(q, READ);
3638                 wake_up(&rl->wait[READ]);
3639         }
3640
3641         if (rl->count[WRITE] >= q->nr_requests) {
3642                 blk_set_queue_full(q, WRITE);
3643         } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3644                 blk_clear_queue_full(q, WRITE);
3645                 wake_up(&rl->wait[WRITE]);
3646         }
3647         return ret;
3648 }
3649
3650 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3651 {
3652         int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3653
3654         return queue_var_show(ra_kb, (page));
3655 }
3656
3657 static ssize_t
3658 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3659 {
3660         unsigned long ra_kb;
3661         ssize_t ret = queue_var_store(&ra_kb, page, count);
3662
3663         spin_lock_irq(q->queue_lock);
3664         if (ra_kb > (q->max_sectors >> 1))
3665                 ra_kb = (q->max_sectors >> 1);
3666
3667         q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3668         spin_unlock_irq(q->queue_lock);
3669
3670         return ret;
3671 }
3672
3673 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3674 {
3675         int max_sectors_kb = q->max_sectors >> 1;
3676
3677         return queue_var_show(max_sectors_kb, (page));
3678 }
3679
3680 static ssize_t
3681 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3682 {
3683         unsigned long max_sectors_kb,
3684                         max_hw_sectors_kb = q->max_hw_sectors >> 1,
3685                         page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3686         ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3687         int ra_kb;
3688
3689         if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3690                 return -EINVAL;
3691         /*
3692          * Take the queue lock to update the readahead and max_sectors
3693          * values synchronously:
3694          */
3695         spin_lock_irq(q->queue_lock);
3696         /*
3697          * Trim readahead window as well, if necessary:
3698          */
3699         ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3700         if (ra_kb > max_sectors_kb)
3701                 q->backing_dev_info.ra_pages =
3702                                 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3703
3704         q->max_sectors = max_sectors_kb << 1;
3705         spin_unlock_irq(q->queue_lock);
3706
3707         return ret;
3708 }
3709
3710 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3711 {
3712         int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3713
3714         return queue_var_show(max_hw_sectors_kb, (page));
3715 }
3716
3717
3718 static struct queue_sysfs_entry queue_requests_entry = {
3719         .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3720         .show = queue_requests_show,
3721         .store = queue_requests_store,
3722 };
3723
3724 static struct queue_sysfs_entry queue_ra_entry = {
3725         .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3726         .show = queue_ra_show,
3727         .store = queue_ra_store,
3728 };
3729
3730 static struct queue_sysfs_entry queue_max_sectors_entry = {
3731         .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3732         .show = queue_max_sectors_show,
3733         .store = queue_max_sectors_store,
3734 };
3735
3736 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3737         .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
3738         .show = queue_max_hw_sectors_show,
3739 };
3740
3741 static struct queue_sysfs_entry queue_iosched_entry = {
3742         .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
3743         .show = elv_iosched_show,
3744         .store = elv_iosched_store,
3745 };
3746
3747 static struct attribute *default_attrs[] = {
3748         &queue_requests_entry.attr,
3749         &queue_ra_entry.attr,
3750         &queue_max_hw_sectors_entry.attr,
3751         &queue_max_sectors_entry.attr,
3752         &queue_iosched_entry.attr,
3753         NULL,
3754 };
3755
3756 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
3757
3758 static ssize_t
3759 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
3760 {
3761         struct queue_sysfs_entry *entry = to_queue(attr);
3762         struct request_queue *q;
3763
3764         q = container_of(kobj, struct request_queue, kobj);
3765         if (!entry->show)
3766                 return -EIO;
3767
3768         return entry->show(q, page);
3769 }
3770
3771 static ssize_t
3772 queue_attr_store(struct kobject *kobj, struct attribute *attr,
3773                     const char *page, size_t length)
3774 {
3775         struct queue_sysfs_entry *entry = to_queue(attr);
3776         struct request_queue *q;
3777
3778         q = container_of(kobj, struct request_queue, kobj);
3779         if (!entry->store)
3780                 return -EIO;
3781
3782         return entry->store(q, page, length);
3783 }
3784
3785 static struct sysfs_ops queue_sysfs_ops = {
3786         .show   = queue_attr_show,
3787         .store  = queue_attr_store,
3788 };
3789
3790 static struct kobj_type queue_ktype = {
3791         .sysfs_ops      = &queue_sysfs_ops,
3792         .default_attrs  = default_attrs,
3793 };
3794
3795 int blk_register_queue(struct gendisk *disk)
3796 {
3797         int ret;
3798
3799         request_queue_t *q = disk->queue;
3800
3801         if (!q || !q->request_fn)
3802                 return -ENXIO;
3803
3804         q->kobj.parent = kobject_get(&disk->kobj);
3805         if (!q->kobj.parent)
3806                 return -EBUSY;
3807
3808         snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
3809         q->kobj.ktype = &queue_ktype;
3810
3811         ret = kobject_register(&q->kobj);
3812         if (ret < 0)
3813                 return ret;
3814
3815         ret = elv_register_queue(q);
3816         if (ret) {
3817                 kobject_unregister(&q->kobj);
3818                 return ret;
3819         }
3820
3821         return 0;
3822 }
3823
3824 void blk_unregister_queue(struct gendisk *disk)
3825 {
3826         request_queue_t *q = disk->queue;
3827
3828         if (q && q->request_fn) {
3829                 elv_unregister_queue(q);
3830
3831                 kobject_unregister(&q->kobj);
3832                 kobject_put(&disk->kobj);
3833         }
3834 }