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