7fc903b5f3cd1c765acbc25e45886a80e1590e5b
[linux-2.6.git] / block / ll_rw_blk.c
1 /*
2  * Copyright (C) 1991, 1992 Linus Torvalds
3  * Copyright (C) 1994,      Karl Keyte: Added support for disk statistics
4  * Elevator latency, (C) 2000  Andrea Arcangeli <andrea@suse.de> SuSE
5  * Queue request tables / lock, selectable elevator, Jens Axboe <axboe@suse.de>
6  * kernel-doc documentation started by NeilBrown <neilb@cse.unsw.edu.au> -  July2000
7  * bio rewrite, highmem i/o, etc, Jens Axboe <axboe@suse.de> - may 2001
8  */
9
10 /*
11  * This handles all read/write requests to block devices
12  */
13 #include <linux/config.h>
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/backing-dev.h>
17 #include <linux/bio.h>
18 #include <linux/blkdev.h>
19 #include <linux/highmem.h>
20 #include <linux/mm.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/string.h>
23 #include <linux/init.h>
24 #include <linux/bootmem.h>      /* for max_pfn/max_low_pfn */
25 #include <linux/completion.h>
26 #include <linux/slab.h>
27 #include <linux/swap.h>
28 #include <linux/writeback.h>
29 #include <linux/interrupt.h>
30 #include <linux/cpu.h>
31 #include <linux/blktrace_api.h>
32
33 /*
34  * for max sense size
35  */
36 #include <scsi/scsi_cmnd.h>
37
38 static void blk_unplug_work(void *data);
39 static void blk_unplug_timeout(unsigned long data);
40 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io);
41 static void init_request_from_bio(struct request *req, struct bio *bio);
42 static int __make_request(request_queue_t *q, struct bio *bio);
43
44 /*
45  * For the allocated request tables
46  */
47 static kmem_cache_t *request_cachep;
48
49 /*
50  * For queue allocation
51  */
52 static kmem_cache_t *requestq_cachep;
53
54 /*
55  * For io context allocations
56  */
57 static kmem_cache_t *iocontext_cachep;
58
59 static wait_queue_head_t congestion_wqh[2] = {
60                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
61                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
62         };
63
64 /*
65  * Controlling structure to kblockd
66  */
67 static struct workqueue_struct *kblockd_workqueue;
68
69 unsigned long blk_max_low_pfn, blk_max_pfn;
70
71 EXPORT_SYMBOL(blk_max_low_pfn);
72 EXPORT_SYMBOL(blk_max_pfn);
73
74 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
75
76 /* Amount of time in which a process may batch requests */
77 #define BLK_BATCH_TIME  (HZ/50UL)
78
79 /* Number of requests a "batching" process may submit */
80 #define BLK_BATCH_REQ   32
81
82 /*
83  * Return the threshold (number of used requests) at which the queue is
84  * considered to be congested.  It include a little hysteresis to keep the
85  * context switch rate down.
86  */
87 static inline int queue_congestion_on_threshold(struct request_queue *q)
88 {
89         return q->nr_congestion_on;
90 }
91
92 /*
93  * The threshold at which a queue is considered to be uncongested
94  */
95 static inline int queue_congestion_off_threshold(struct request_queue *q)
96 {
97         return q->nr_congestion_off;
98 }
99
100 static void blk_queue_congestion_threshold(struct request_queue *q)
101 {
102         int nr;
103
104         nr = q->nr_requests - (q->nr_requests / 8) + 1;
105         if (nr > q->nr_requests)
106                 nr = q->nr_requests;
107         q->nr_congestion_on = nr;
108
109         nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
110         if (nr < 1)
111                 nr = 1;
112         q->nr_congestion_off = nr;
113 }
114
115 /*
116  * A queue has just exitted congestion.  Note this in the global counter of
117  * congested queues, and wake up anyone who was waiting for requests to be
118  * put back.
119  */
120 static void clear_queue_congested(request_queue_t *q, int rw)
121 {
122         enum bdi_state bit;
123         wait_queue_head_t *wqh = &congestion_wqh[rw];
124
125         bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
126         clear_bit(bit, &q->backing_dev_info.state);
127         smp_mb__after_clear_bit();
128         if (waitqueue_active(wqh))
129                 wake_up(wqh);
130 }
131
132 /*
133  * A queue has just entered congestion.  Flag that in the queue's VM-visible
134  * state flags and increment the global gounter of congested queues.
135  */
136 static void set_queue_congested(request_queue_t *q, int rw)
137 {
138         enum bdi_state bit;
139
140         bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
141         set_bit(bit, &q->backing_dev_info.state);
142 }
143
144 /**
145  * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
146  * @bdev:       device
147  *
148  * Locates the passed device's request queue and returns the address of its
149  * backing_dev_info
150  *
151  * Will return NULL if the request queue cannot be located.
152  */
153 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
154 {
155         struct backing_dev_info *ret = NULL;
156         request_queue_t *q = bdev_get_queue(bdev);
157
158         if (q)
159                 ret = &q->backing_dev_info;
160         return ret;
161 }
162
163 EXPORT_SYMBOL(blk_get_backing_dev_info);
164
165 void blk_queue_activity_fn(request_queue_t *q, activity_fn *fn, void *data)
166 {
167         q->activity_fn = fn;
168         q->activity_data = data;
169 }
170
171 EXPORT_SYMBOL(blk_queue_activity_fn);
172
173 /**
174  * blk_queue_prep_rq - set a prepare_request function for queue
175  * @q:          queue
176  * @pfn:        prepare_request function
177  *
178  * It's possible for a queue to register a prepare_request callback which
179  * is invoked before the request is handed to the request_fn. The goal of
180  * the function is to prepare a request for I/O, it can be used to build a
181  * cdb from the request data for instance.
182  *
183  */
184 void blk_queue_prep_rq(request_queue_t *q, prep_rq_fn *pfn)
185 {
186         q->prep_rq_fn = pfn;
187 }
188
189 EXPORT_SYMBOL(blk_queue_prep_rq);
190
191 /**
192  * blk_queue_merge_bvec - set a merge_bvec function for queue
193  * @q:          queue
194  * @mbfn:       merge_bvec_fn
195  *
196  * Usually queues have static limitations on the max sectors or segments that
197  * we can put in a request. Stacking drivers may have some settings that
198  * are dynamic, and thus we have to query the queue whether it is ok to
199  * add a new bio_vec to a bio at a given offset or not. If the block device
200  * has such limitations, it needs to register a merge_bvec_fn to control
201  * the size of bio's sent to it. Note that a block device *must* allow a
202  * single page to be added to an empty bio. The block device driver may want
203  * to use the bio_split() function to deal with these bio's. By default
204  * no merge_bvec_fn is defined for a queue, and only the fixed limits are
205  * honored.
206  */
207 void blk_queue_merge_bvec(request_queue_t *q, merge_bvec_fn *mbfn)
208 {
209         q->merge_bvec_fn = mbfn;
210 }
211
212 EXPORT_SYMBOL(blk_queue_merge_bvec);
213
214 void blk_queue_softirq_done(request_queue_t *q, softirq_done_fn *fn)
215 {
216         q->softirq_done_fn = fn;
217 }
218
219 EXPORT_SYMBOL(blk_queue_softirq_done);
220
221 /**
222  * blk_queue_make_request - define an alternate make_request function for a device
223  * @q:  the request queue for the device to be affected
224  * @mfn: the alternate make_request function
225  *
226  * Description:
227  *    The normal way for &struct bios to be passed to a device
228  *    driver is for them to be collected into requests on a request
229  *    queue, and then to allow the device driver to select requests
230  *    off that queue when it is ready.  This works well for many block
231  *    devices. However some block devices (typically virtual devices
232  *    such as md or lvm) do not benefit from the processing on the
233  *    request queue, and are served best by having the requests passed
234  *    directly to them.  This can be achieved by providing a function
235  *    to blk_queue_make_request().
236  *
237  * Caveat:
238  *    The driver that does this *must* be able to deal appropriately
239  *    with buffers in "highmemory". This can be accomplished by either calling
240  *    __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
241  *    blk_queue_bounce() to create a buffer in normal memory.
242  **/
243 void blk_queue_make_request(request_queue_t * q, make_request_fn * mfn)
244 {
245         /*
246          * set defaults
247          */
248         q->nr_requests = BLKDEV_MAX_RQ;
249         blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
250         blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
251         q->make_request_fn = mfn;
252         q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
253         q->backing_dev_info.state = 0;
254         q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
255         blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
256         blk_queue_hardsect_size(q, 512);
257         blk_queue_dma_alignment(q, 511);
258         blk_queue_congestion_threshold(q);
259         q->nr_batching = BLK_BATCH_REQ;
260
261         q->unplug_thresh = 4;           /* hmm */
262         q->unplug_delay = (3 * HZ) / 1000;      /* 3 milliseconds */
263         if (q->unplug_delay == 0)
264                 q->unplug_delay = 1;
265
266         INIT_WORK(&q->unplug_work, blk_unplug_work, q);
267
268         q->unplug_timer.function = blk_unplug_timeout;
269         q->unplug_timer.data = (unsigned long)q;
270
271         /*
272          * by default assume old behaviour and bounce for any highmem page
273          */
274         blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
275
276         blk_queue_activity_fn(q, NULL, NULL);
277 }
278
279 EXPORT_SYMBOL(blk_queue_make_request);
280
281 static inline void rq_init(request_queue_t *q, struct request *rq)
282 {
283         INIT_LIST_HEAD(&rq->queuelist);
284         INIT_LIST_HEAD(&rq->donelist);
285
286         rq->errors = 0;
287         rq->rq_status = RQ_ACTIVE;
288         rq->bio = rq->biotail = NULL;
289         rq->ioprio = 0;
290         rq->buffer = NULL;
291         rq->ref_count = 1;
292         rq->q = q;
293         rq->waiting = NULL;
294         rq->special = NULL;
295         rq->data_len = 0;
296         rq->data = NULL;
297         rq->nr_phys_segments = 0;
298         rq->sense = NULL;
299         rq->end_io = NULL;
300         rq->end_io_data = NULL;
301         rq->completion_data = NULL;
302 }
303
304 /**
305  * blk_queue_ordered - does this queue support ordered writes
306  * @q:        the request queue
307  * @ordered:  one of QUEUE_ORDERED_*
308  * @prepare_flush_fn: rq setup helper for cache flush ordered writes
309  *
310  * Description:
311  *   For journalled file systems, doing ordered writes on a commit
312  *   block instead of explicitly doing wait_on_buffer (which is bad
313  *   for performance) can be a big win. Block drivers supporting this
314  *   feature should call this function and indicate so.
315  *
316  **/
317 int blk_queue_ordered(request_queue_t *q, unsigned ordered,
318                       prepare_flush_fn *prepare_flush_fn)
319 {
320         if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
321             prepare_flush_fn == NULL) {
322                 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
323                 return -EINVAL;
324         }
325
326         if (ordered != QUEUE_ORDERED_NONE &&
327             ordered != QUEUE_ORDERED_DRAIN &&
328             ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
329             ordered != QUEUE_ORDERED_DRAIN_FUA &&
330             ordered != QUEUE_ORDERED_TAG &&
331             ordered != QUEUE_ORDERED_TAG_FLUSH &&
332             ordered != QUEUE_ORDERED_TAG_FUA) {
333                 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
334                 return -EINVAL;
335         }
336
337         q->ordered = ordered;
338         q->next_ordered = ordered;
339         q->prepare_flush_fn = prepare_flush_fn;
340
341         return 0;
342 }
343
344 EXPORT_SYMBOL(blk_queue_ordered);
345
346 /**
347  * blk_queue_issue_flush_fn - set function for issuing a flush
348  * @q:     the request queue
349  * @iff:   the function to be called issuing the flush
350  *
351  * Description:
352  *   If a driver supports issuing a flush command, the support is notified
353  *   to the block layer by defining it through this call.
354  *
355  **/
356 void blk_queue_issue_flush_fn(request_queue_t *q, issue_flush_fn *iff)
357 {
358         q->issue_flush_fn = iff;
359 }
360
361 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
362
363 /*
364  * Cache flushing for ordered writes handling
365  */
366 inline unsigned blk_ordered_cur_seq(request_queue_t *q)
367 {
368         if (!q->ordseq)
369                 return 0;
370         return 1 << ffz(q->ordseq);
371 }
372
373 unsigned blk_ordered_req_seq(struct request *rq)
374 {
375         request_queue_t *q = rq->q;
376
377         BUG_ON(q->ordseq == 0);
378
379         if (rq == &q->pre_flush_rq)
380                 return QUEUE_ORDSEQ_PREFLUSH;
381         if (rq == &q->bar_rq)
382                 return QUEUE_ORDSEQ_BAR;
383         if (rq == &q->post_flush_rq)
384                 return QUEUE_ORDSEQ_POSTFLUSH;
385
386         if ((rq->flags & REQ_ORDERED_COLOR) ==
387             (q->orig_bar_rq->flags & REQ_ORDERED_COLOR))
388                 return QUEUE_ORDSEQ_DRAIN;
389         else
390                 return QUEUE_ORDSEQ_DONE;
391 }
392
393 void blk_ordered_complete_seq(request_queue_t *q, unsigned seq, int error)
394 {
395         struct request *rq;
396         int uptodate;
397
398         if (error && !q->orderr)
399                 q->orderr = error;
400
401         BUG_ON(q->ordseq & seq);
402         q->ordseq |= seq;
403
404         if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
405                 return;
406
407         /*
408          * Okay, sequence complete.
409          */
410         rq = q->orig_bar_rq;
411         uptodate = q->orderr ? q->orderr : 1;
412
413         q->ordseq = 0;
414
415         end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
416         end_that_request_last(rq, uptodate);
417 }
418
419 static void pre_flush_end_io(struct request *rq, int error)
420 {
421         elv_completed_request(rq->q, rq);
422         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
423 }
424
425 static void bar_end_io(struct request *rq, int error)
426 {
427         elv_completed_request(rq->q, rq);
428         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
429 }
430
431 static void post_flush_end_io(struct request *rq, int error)
432 {
433         elv_completed_request(rq->q, rq);
434         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
435 }
436
437 static void queue_flush(request_queue_t *q, unsigned which)
438 {
439         struct request *rq;
440         rq_end_io_fn *end_io;
441
442         if (which == QUEUE_ORDERED_PREFLUSH) {
443                 rq = &q->pre_flush_rq;
444                 end_io = pre_flush_end_io;
445         } else {
446                 rq = &q->post_flush_rq;
447                 end_io = post_flush_end_io;
448         }
449
450         rq_init(q, rq);
451         rq->flags = REQ_HARDBARRIER;
452         rq->elevator_private = NULL;
453         rq->rq_disk = q->bar_rq.rq_disk;
454         rq->rl = NULL;
455         rq->end_io = end_io;
456         q->prepare_flush_fn(q, rq);
457
458         elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
459 }
460
461 static inline struct request *start_ordered(request_queue_t *q,
462                                             struct request *rq)
463 {
464         q->bi_size = 0;
465         q->orderr = 0;
466         q->ordered = q->next_ordered;
467         q->ordseq |= QUEUE_ORDSEQ_STARTED;
468
469         /*
470          * Prep proxy barrier request.
471          */
472         blkdev_dequeue_request(rq);
473         q->orig_bar_rq = rq;
474         rq = &q->bar_rq;
475         rq_init(q, rq);
476         rq->flags = bio_data_dir(q->orig_bar_rq->bio);
477         rq->flags |= q->ordered & QUEUE_ORDERED_FUA ? REQ_FUA : 0;
478         rq->elevator_private = NULL;
479         rq->rl = NULL;
480         init_request_from_bio(rq, q->orig_bar_rq->bio);
481         rq->end_io = bar_end_io;
482
483         /*
484          * Queue ordered sequence.  As we stack them at the head, we
485          * need to queue in reverse order.  Note that we rely on that
486          * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
487          * request gets inbetween ordered sequence.
488          */
489         if (q->ordered & QUEUE_ORDERED_POSTFLUSH)
490                 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
491         else
492                 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
493
494         elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
495
496         if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
497                 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
498                 rq = &q->pre_flush_rq;
499         } else
500                 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
501
502         if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
503                 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
504         else
505                 rq = NULL;
506
507         return rq;
508 }
509
510 int blk_do_ordered(request_queue_t *q, struct request **rqp)
511 {
512         struct request *rq = *rqp;
513         int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
514
515         if (!q->ordseq) {
516                 if (!is_barrier)
517                         return 1;
518
519                 if (q->next_ordered != QUEUE_ORDERED_NONE) {
520                         *rqp = start_ordered(q, rq);
521                         return 1;
522                 } else {
523                         /*
524                          * This can happen when the queue switches to
525                          * ORDERED_NONE while this request is on it.
526                          */
527                         blkdev_dequeue_request(rq);
528                         end_that_request_first(rq, -EOPNOTSUPP,
529                                                rq->hard_nr_sectors);
530                         end_that_request_last(rq, -EOPNOTSUPP);
531                         *rqp = NULL;
532                         return 0;
533                 }
534         }
535
536         /*
537          * Ordered sequence in progress
538          */
539
540         /* Special requests are not subject to ordering rules. */
541         if (!blk_fs_request(rq) &&
542             rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
543                 return 1;
544
545         if (q->ordered & QUEUE_ORDERED_TAG) {
546                 /* Ordered by tag.  Blocking the next barrier is enough. */
547                 if (is_barrier && rq != &q->bar_rq)
548                         *rqp = NULL;
549         } else {
550                 /* Ordered by draining.  Wait for turn. */
551                 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
552                 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
553                         *rqp = NULL;
554         }
555
556         return 1;
557 }
558
559 static int flush_dry_bio_endio(struct bio *bio, unsigned int bytes, int error)
560 {
561         request_queue_t *q = bio->bi_private;
562         struct bio_vec *bvec;
563         int i;
564
565         /*
566          * This is dry run, restore bio_sector and size.  We'll finish
567          * this request again with the original bi_end_io after an
568          * error occurs or post flush is complete.
569          */
570         q->bi_size += bytes;
571
572         if (bio->bi_size)
573                 return 1;
574
575         /* Rewind bvec's */
576         bio->bi_idx = 0;
577         bio_for_each_segment(bvec, bio, i) {
578                 bvec->bv_len += bvec->bv_offset;
579                 bvec->bv_offset = 0;
580         }
581
582         /* Reset bio */
583         set_bit(BIO_UPTODATE, &bio->bi_flags);
584         bio->bi_size = q->bi_size;
585         bio->bi_sector -= (q->bi_size >> 9);
586         q->bi_size = 0;
587
588         return 0;
589 }
590
591 static inline int ordered_bio_endio(struct request *rq, struct bio *bio,
592                                     unsigned int nbytes, int error)
593 {
594         request_queue_t *q = rq->q;
595         bio_end_io_t *endio;
596         void *private;
597
598         if (&q->bar_rq != rq)
599                 return 0;
600
601         /*
602          * Okay, this is the barrier request in progress, dry finish it.
603          */
604         if (error && !q->orderr)
605                 q->orderr = error;
606
607         endio = bio->bi_end_io;
608         private = bio->bi_private;
609         bio->bi_end_io = flush_dry_bio_endio;
610         bio->bi_private = q;
611
612         bio_endio(bio, nbytes, error);
613
614         bio->bi_end_io = endio;
615         bio->bi_private = private;
616
617         return 1;
618 }
619
620 /**
621  * blk_queue_bounce_limit - set bounce buffer limit for queue
622  * @q:  the request queue for the device
623  * @dma_addr:   bus address limit
624  *
625  * Description:
626  *    Different hardware can have different requirements as to what pages
627  *    it can do I/O directly to. A low level driver can call
628  *    blk_queue_bounce_limit to have lower memory pages allocated as bounce
629  *    buffers for doing I/O to pages residing above @page.
630  **/
631 void blk_queue_bounce_limit(request_queue_t *q, u64 dma_addr)
632 {
633         unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
634         int dma = 0;
635
636         q->bounce_gfp = GFP_NOIO;
637 #if BITS_PER_LONG == 64
638         /* Assume anything <= 4GB can be handled by IOMMU.
639            Actually some IOMMUs can handle everything, but I don't
640            know of a way to test this here. */
641         if (bounce_pfn < (0xffffffff>>PAGE_SHIFT))
642                 dma = 1;
643         q->bounce_pfn = max_low_pfn;
644 #else
645         if (bounce_pfn < blk_max_low_pfn)
646                 dma = 1;
647         q->bounce_pfn = bounce_pfn;
648 #endif
649         if (dma) {
650                 init_emergency_isa_pool();
651                 q->bounce_gfp = GFP_NOIO | GFP_DMA;
652                 q->bounce_pfn = bounce_pfn;
653         }
654 }
655
656 EXPORT_SYMBOL(blk_queue_bounce_limit);
657
658 /**
659  * blk_queue_max_sectors - set max sectors for a request for this queue
660  * @q:  the request queue for the device
661  * @max_sectors:  max sectors in the usual 512b unit
662  *
663  * Description:
664  *    Enables a low level driver to set an upper limit on the size of
665  *    received requests.
666  **/
667 void blk_queue_max_sectors(request_queue_t *q, unsigned int max_sectors)
668 {
669         if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
670                 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
671                 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
672         }
673
674         if (BLK_DEF_MAX_SECTORS > max_sectors)
675                 q->max_hw_sectors = q->max_sectors = max_sectors;
676         else {
677                 q->max_sectors = BLK_DEF_MAX_SECTORS;
678                 q->max_hw_sectors = max_sectors;
679         }
680 }
681
682 EXPORT_SYMBOL(blk_queue_max_sectors);
683
684 /**
685  * blk_queue_max_phys_segments - set max phys segments for a request for this queue
686  * @q:  the request queue for the device
687  * @max_segments:  max number of segments
688  *
689  * Description:
690  *    Enables a low level driver to set an upper limit on the number of
691  *    physical data segments in a request.  This would be the largest sized
692  *    scatter list the driver could handle.
693  **/
694 void blk_queue_max_phys_segments(request_queue_t *q, unsigned short max_segments)
695 {
696         if (!max_segments) {
697                 max_segments = 1;
698                 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
699         }
700
701         q->max_phys_segments = max_segments;
702 }
703
704 EXPORT_SYMBOL(blk_queue_max_phys_segments);
705
706 /**
707  * blk_queue_max_hw_segments - set max hw segments for a request for this queue
708  * @q:  the request queue for the device
709  * @max_segments:  max number of segments
710  *
711  * Description:
712  *    Enables a low level driver to set an upper limit on the number of
713  *    hw data segments in a request.  This would be the largest number of
714  *    address/length pairs the host adapter can actually give as once
715  *    to the device.
716  **/
717 void blk_queue_max_hw_segments(request_queue_t *q, unsigned short max_segments)
718 {
719         if (!max_segments) {
720                 max_segments = 1;
721                 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
722         }
723
724         q->max_hw_segments = max_segments;
725 }
726
727 EXPORT_SYMBOL(blk_queue_max_hw_segments);
728
729 /**
730  * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
731  * @q:  the request queue for the device
732  * @max_size:  max size of segment in bytes
733  *
734  * Description:
735  *    Enables a low level driver to set an upper limit on the size of a
736  *    coalesced segment
737  **/
738 void blk_queue_max_segment_size(request_queue_t *q, unsigned int max_size)
739 {
740         if (max_size < PAGE_CACHE_SIZE) {
741                 max_size = PAGE_CACHE_SIZE;
742                 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
743         }
744
745         q->max_segment_size = max_size;
746 }
747
748 EXPORT_SYMBOL(blk_queue_max_segment_size);
749
750 /**
751  * blk_queue_hardsect_size - set hardware sector size for the queue
752  * @q:  the request queue for the device
753  * @size:  the hardware sector size, in bytes
754  *
755  * Description:
756  *   This should typically be set to the lowest possible sector size
757  *   that the hardware can operate on (possible without reverting to
758  *   even internal read-modify-write operations). Usually the default
759  *   of 512 covers most hardware.
760  **/
761 void blk_queue_hardsect_size(request_queue_t *q, unsigned short size)
762 {
763         q->hardsect_size = size;
764 }
765
766 EXPORT_SYMBOL(blk_queue_hardsect_size);
767
768 /*
769  * Returns the minimum that is _not_ zero, unless both are zero.
770  */
771 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
772
773 /**
774  * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
775  * @t:  the stacking driver (top)
776  * @b:  the underlying device (bottom)
777  **/
778 void blk_queue_stack_limits(request_queue_t *t, request_queue_t *b)
779 {
780         /* zero is "infinity" */
781         t->max_sectors = min_not_zero(t->max_sectors,b->max_sectors);
782         t->max_hw_sectors = min_not_zero(t->max_hw_sectors,b->max_hw_sectors);
783
784         t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
785         t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
786         t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
787         t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
788 }
789
790 EXPORT_SYMBOL(blk_queue_stack_limits);
791
792 /**
793  * blk_queue_segment_boundary - set boundary rules for segment merging
794  * @q:  the request queue for the device
795  * @mask:  the memory boundary mask
796  **/
797 void blk_queue_segment_boundary(request_queue_t *q, unsigned long mask)
798 {
799         if (mask < PAGE_CACHE_SIZE - 1) {
800                 mask = PAGE_CACHE_SIZE - 1;
801                 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
802         }
803
804         q->seg_boundary_mask = mask;
805 }
806
807 EXPORT_SYMBOL(blk_queue_segment_boundary);
808
809 /**
810  * blk_queue_dma_alignment - set dma length and memory alignment
811  * @q:     the request queue for the device
812  * @mask:  alignment mask
813  *
814  * description:
815  *    set required memory and length aligment for direct dma transactions.
816  *    this is used when buiding direct io requests for the queue.
817  *
818  **/
819 void blk_queue_dma_alignment(request_queue_t *q, int mask)
820 {
821         q->dma_alignment = mask;
822 }
823
824 EXPORT_SYMBOL(blk_queue_dma_alignment);
825
826 /**
827  * blk_queue_find_tag - find a request by its tag and queue
828  * @q:   The request queue for the device
829  * @tag: The tag of the request
830  *
831  * Notes:
832  *    Should be used when a device returns a tag and you want to match
833  *    it with a request.
834  *
835  *    no locks need be held.
836  **/
837 struct request *blk_queue_find_tag(request_queue_t *q, int tag)
838 {
839         struct blk_queue_tag *bqt = q->queue_tags;
840
841         if (unlikely(bqt == NULL || tag >= bqt->real_max_depth))
842                 return NULL;
843
844         return bqt->tag_index[tag];
845 }
846
847 EXPORT_SYMBOL(blk_queue_find_tag);
848
849 /**
850  * __blk_queue_free_tags - release tag maintenance info
851  * @q:  the request queue for the device
852  *
853  *  Notes:
854  *    blk_cleanup_queue() will take care of calling this function, if tagging
855  *    has been used. So there's no need to call this directly.
856  **/
857 static void __blk_queue_free_tags(request_queue_t *q)
858 {
859         struct blk_queue_tag *bqt = q->queue_tags;
860
861         if (!bqt)
862                 return;
863
864         if (atomic_dec_and_test(&bqt->refcnt)) {
865                 BUG_ON(bqt->busy);
866                 BUG_ON(!list_empty(&bqt->busy_list));
867
868                 kfree(bqt->tag_index);
869                 bqt->tag_index = NULL;
870
871                 kfree(bqt->tag_map);
872                 bqt->tag_map = NULL;
873
874                 kfree(bqt);
875         }
876
877         q->queue_tags = NULL;
878         q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
879 }
880
881 /**
882  * blk_queue_free_tags - release tag maintenance info
883  * @q:  the request queue for the device
884  *
885  *  Notes:
886  *      This is used to disabled tagged queuing to a device, yet leave
887  *      queue in function.
888  **/
889 void blk_queue_free_tags(request_queue_t *q)
890 {
891         clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
892 }
893
894 EXPORT_SYMBOL(blk_queue_free_tags);
895
896 static int
897 init_tag_map(request_queue_t *q, struct blk_queue_tag *tags, int depth)
898 {
899         struct request **tag_index;
900         unsigned long *tag_map;
901         int nr_ulongs;
902
903         if (depth > q->nr_requests * 2) {
904                 depth = q->nr_requests * 2;
905                 printk(KERN_ERR "%s: adjusted depth to %d\n",
906                                 __FUNCTION__, depth);
907         }
908
909         tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC);
910         if (!tag_index)
911                 goto fail;
912
913         nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
914         tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
915         if (!tag_map)
916                 goto fail;
917
918         tags->real_max_depth = depth;
919         tags->max_depth = depth;
920         tags->tag_index = tag_index;
921         tags->tag_map = tag_map;
922
923         return 0;
924 fail:
925         kfree(tag_index);
926         return -ENOMEM;
927 }
928
929 /**
930  * blk_queue_init_tags - initialize the queue tag info
931  * @q:  the request queue for the device
932  * @depth:  the maximum queue depth supported
933  * @tags: the tag to use
934  **/
935 int blk_queue_init_tags(request_queue_t *q, int depth,
936                         struct blk_queue_tag *tags)
937 {
938         int rc;
939
940         BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
941
942         if (!tags && !q->queue_tags) {
943                 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
944                 if (!tags)
945                         goto fail;
946
947                 if (init_tag_map(q, tags, depth))
948                         goto fail;
949
950                 INIT_LIST_HEAD(&tags->busy_list);
951                 tags->busy = 0;
952                 atomic_set(&tags->refcnt, 1);
953         } else if (q->queue_tags) {
954                 if ((rc = blk_queue_resize_tags(q, depth)))
955                         return rc;
956                 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
957                 return 0;
958         } else
959                 atomic_inc(&tags->refcnt);
960
961         /*
962          * assign it, all done
963          */
964         q->queue_tags = tags;
965         q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
966         return 0;
967 fail:
968         kfree(tags);
969         return -ENOMEM;
970 }
971
972 EXPORT_SYMBOL(blk_queue_init_tags);
973
974 /**
975  * blk_queue_resize_tags - change the queueing depth
976  * @q:  the request queue for the device
977  * @new_depth: the new max command queueing depth
978  *
979  *  Notes:
980  *    Must be called with the queue lock held.
981  **/
982 int blk_queue_resize_tags(request_queue_t *q, int new_depth)
983 {
984         struct blk_queue_tag *bqt = q->queue_tags;
985         struct request **tag_index;
986         unsigned long *tag_map;
987         int max_depth, nr_ulongs;
988
989         if (!bqt)
990                 return -ENXIO;
991
992         /*
993          * if we already have large enough real_max_depth.  just
994          * adjust max_depth.  *NOTE* as requests with tag value
995          * between new_depth and real_max_depth can be in-flight, tag
996          * map can not be shrunk blindly here.
997          */
998         if (new_depth <= bqt->real_max_depth) {
999                 bqt->max_depth = new_depth;
1000                 return 0;
1001         }
1002
1003         /*
1004          * save the old state info, so we can copy it back
1005          */
1006         tag_index = bqt->tag_index;
1007         tag_map = bqt->tag_map;
1008         max_depth = bqt->real_max_depth;
1009
1010         if (init_tag_map(q, bqt, new_depth))
1011                 return -ENOMEM;
1012
1013         memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
1014         nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
1015         memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
1016
1017         kfree(tag_index);
1018         kfree(tag_map);
1019         return 0;
1020 }
1021
1022 EXPORT_SYMBOL(blk_queue_resize_tags);
1023
1024 /**
1025  * blk_queue_end_tag - end tag operations for a request
1026  * @q:  the request queue for the device
1027  * @rq: the request that has completed
1028  *
1029  *  Description:
1030  *    Typically called when end_that_request_first() returns 0, meaning
1031  *    all transfers have been done for a request. It's important to call
1032  *    this function before end_that_request_last(), as that will put the
1033  *    request back on the free list thus corrupting the internal tag list.
1034  *
1035  *  Notes:
1036  *   queue lock must be held.
1037  **/
1038 void blk_queue_end_tag(request_queue_t *q, struct request *rq)
1039 {
1040         struct blk_queue_tag *bqt = q->queue_tags;
1041         int tag = rq->tag;
1042
1043         BUG_ON(tag == -1);
1044
1045         if (unlikely(tag >= bqt->real_max_depth))
1046                 /*
1047                  * This can happen after tag depth has been reduced.
1048                  * FIXME: how about a warning or info message here?
1049                  */
1050                 return;
1051
1052         if (unlikely(!__test_and_clear_bit(tag, bqt->tag_map))) {
1053                 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
1054                        __FUNCTION__, tag);
1055                 return;
1056         }
1057
1058         list_del_init(&rq->queuelist);
1059         rq->flags &= ~REQ_QUEUED;
1060         rq->tag = -1;
1061
1062         if (unlikely(bqt->tag_index[tag] == NULL))
1063                 printk(KERN_ERR "%s: tag %d is missing\n",
1064                        __FUNCTION__, tag);
1065
1066         bqt->tag_index[tag] = NULL;
1067         bqt->busy--;
1068 }
1069
1070 EXPORT_SYMBOL(blk_queue_end_tag);
1071
1072 /**
1073  * blk_queue_start_tag - find a free tag and assign it
1074  * @q:  the request queue for the device
1075  * @rq:  the block request that needs tagging
1076  *
1077  *  Description:
1078  *    This can either be used as a stand-alone helper, or possibly be
1079  *    assigned as the queue &prep_rq_fn (in which case &struct request
1080  *    automagically gets a tag assigned). Note that this function
1081  *    assumes that any type of request can be queued! if this is not
1082  *    true for your device, you must check the request type before
1083  *    calling this function.  The request will also be removed from
1084  *    the request queue, so it's the drivers responsibility to readd
1085  *    it if it should need to be restarted for some reason.
1086  *
1087  *  Notes:
1088  *   queue lock must be held.
1089  **/
1090 int blk_queue_start_tag(request_queue_t *q, struct request *rq)
1091 {
1092         struct blk_queue_tag *bqt = q->queue_tags;
1093         int tag;
1094
1095         if (unlikely((rq->flags & REQ_QUEUED))) {
1096                 printk(KERN_ERR 
1097                        "%s: request %p for device [%s] already tagged %d",
1098                        __FUNCTION__, rq,
1099                        rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
1100                 BUG();
1101         }
1102
1103         tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
1104         if (tag >= bqt->max_depth)
1105                 return 1;
1106
1107         __set_bit(tag, bqt->tag_map);
1108
1109         rq->flags |= REQ_QUEUED;
1110         rq->tag = tag;
1111         bqt->tag_index[tag] = rq;
1112         blkdev_dequeue_request(rq);
1113         list_add(&rq->queuelist, &bqt->busy_list);
1114         bqt->busy++;
1115         return 0;
1116 }
1117
1118 EXPORT_SYMBOL(blk_queue_start_tag);
1119
1120 /**
1121  * blk_queue_invalidate_tags - invalidate all pending tags
1122  * @q:  the request queue for the device
1123  *
1124  *  Description:
1125  *   Hardware conditions may dictate a need to stop all pending requests.
1126  *   In this case, we will safely clear the block side of the tag queue and
1127  *   readd all requests to the request queue in the right order.
1128  *
1129  *  Notes:
1130  *   queue lock must be held.
1131  **/
1132 void blk_queue_invalidate_tags(request_queue_t *q)
1133 {
1134         struct blk_queue_tag *bqt = q->queue_tags;
1135         struct list_head *tmp, *n;
1136         struct request *rq;
1137
1138         list_for_each_safe(tmp, n, &bqt->busy_list) {
1139                 rq = list_entry_rq(tmp);
1140
1141                 if (rq->tag == -1) {
1142                         printk(KERN_ERR
1143                                "%s: bad tag found on list\n", __FUNCTION__);
1144                         list_del_init(&rq->queuelist);
1145                         rq->flags &= ~REQ_QUEUED;
1146                 } else
1147                         blk_queue_end_tag(q, rq);
1148
1149                 rq->flags &= ~REQ_STARTED;
1150                 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1151         }
1152 }
1153
1154 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1155
1156 static const char * const rq_flags[] = {
1157         "REQ_RW",
1158         "REQ_FAILFAST",
1159         "REQ_SORTED",
1160         "REQ_SOFTBARRIER",
1161         "REQ_HARDBARRIER",
1162         "REQ_FUA",
1163         "REQ_CMD",
1164         "REQ_NOMERGE",
1165         "REQ_STARTED",
1166         "REQ_DONTPREP",
1167         "REQ_QUEUED",
1168         "REQ_ELVPRIV",
1169         "REQ_PC",
1170         "REQ_BLOCK_PC",
1171         "REQ_SENSE",
1172         "REQ_FAILED",
1173         "REQ_QUIET",
1174         "REQ_SPECIAL",
1175         "REQ_DRIVE_CMD",
1176         "REQ_DRIVE_TASK",
1177         "REQ_DRIVE_TASKFILE",
1178         "REQ_PREEMPT",
1179         "REQ_PM_SUSPEND",
1180         "REQ_PM_RESUME",
1181         "REQ_PM_SHUTDOWN",
1182         "REQ_ORDERED_COLOR",
1183 };
1184
1185 void blk_dump_rq_flags(struct request *rq, char *msg)
1186 {
1187         int bit;
1188
1189         printk("%s: dev %s: flags = ", msg,
1190                 rq->rq_disk ? rq->rq_disk->disk_name : "?");
1191         bit = 0;
1192         do {
1193                 if (rq->flags & (1 << bit))
1194                         printk("%s ", rq_flags[bit]);
1195                 bit++;
1196         } while (bit < __REQ_NR_BITS);
1197
1198         printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1199                                                        rq->nr_sectors,
1200                                                        rq->current_nr_sectors);
1201         printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1202
1203         if (rq->flags & (REQ_BLOCK_PC | REQ_PC)) {
1204                 printk("cdb: ");
1205                 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1206                         printk("%02x ", rq->cmd[bit]);
1207                 printk("\n");
1208         }
1209 }
1210
1211 EXPORT_SYMBOL(blk_dump_rq_flags);
1212
1213 void blk_recount_segments(request_queue_t *q, struct bio *bio)
1214 {
1215         struct bio_vec *bv, *bvprv = NULL;
1216         int i, nr_phys_segs, nr_hw_segs, seg_size, hw_seg_size, cluster;
1217         int high, highprv = 1;
1218
1219         if (unlikely(!bio->bi_io_vec))
1220                 return;
1221
1222         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1223         hw_seg_size = seg_size = nr_phys_segs = nr_hw_segs = 0;
1224         bio_for_each_segment(bv, bio, i) {
1225                 /*
1226                  * the trick here is making sure that a high page is never
1227                  * considered part of another segment, since that might
1228                  * change with the bounce page.
1229                  */
1230                 high = page_to_pfn(bv->bv_page) >= q->bounce_pfn;
1231                 if (high || highprv)
1232                         goto new_hw_segment;
1233                 if (cluster) {
1234                         if (seg_size + bv->bv_len > q->max_segment_size)
1235                                 goto new_segment;
1236                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1237                                 goto new_segment;
1238                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1239                                 goto new_segment;
1240                         if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1241                                 goto new_hw_segment;
1242
1243                         seg_size += bv->bv_len;
1244                         hw_seg_size += bv->bv_len;
1245                         bvprv = bv;
1246                         continue;
1247                 }
1248 new_segment:
1249                 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1250                     !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len)) {
1251                         hw_seg_size += bv->bv_len;
1252                 } else {
1253 new_hw_segment:
1254                         if (hw_seg_size > bio->bi_hw_front_size)
1255                                 bio->bi_hw_front_size = hw_seg_size;
1256                         hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1257                         nr_hw_segs++;
1258                 }
1259
1260                 nr_phys_segs++;
1261                 bvprv = bv;
1262                 seg_size = bv->bv_len;
1263                 highprv = high;
1264         }
1265         if (hw_seg_size > bio->bi_hw_back_size)
1266                 bio->bi_hw_back_size = hw_seg_size;
1267         if (nr_hw_segs == 1 && hw_seg_size > bio->bi_hw_front_size)
1268                 bio->bi_hw_front_size = hw_seg_size;
1269         bio->bi_phys_segments = nr_phys_segs;
1270         bio->bi_hw_segments = nr_hw_segs;
1271         bio->bi_flags |= (1 << BIO_SEG_VALID);
1272 }
1273
1274
1275 static int blk_phys_contig_segment(request_queue_t *q, struct bio *bio,
1276                                    struct bio *nxt)
1277 {
1278         if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1279                 return 0;
1280
1281         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1282                 return 0;
1283         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1284                 return 0;
1285
1286         /*
1287          * bio and nxt are contigous in memory, check if the queue allows
1288          * these two to be merged into one
1289          */
1290         if (BIO_SEG_BOUNDARY(q, bio, nxt))
1291                 return 1;
1292
1293         return 0;
1294 }
1295
1296 static int blk_hw_contig_segment(request_queue_t *q, struct bio *bio,
1297                                  struct bio *nxt)
1298 {
1299         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1300                 blk_recount_segments(q, bio);
1301         if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1302                 blk_recount_segments(q, nxt);
1303         if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1304             BIOVEC_VIRT_OVERSIZE(bio->bi_hw_front_size + bio->bi_hw_back_size))
1305                 return 0;
1306         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1307                 return 0;
1308
1309         return 1;
1310 }
1311
1312 /*
1313  * map a request to scatterlist, return number of sg entries setup. Caller
1314  * must make sure sg can hold rq->nr_phys_segments entries
1315  */
1316 int blk_rq_map_sg(request_queue_t *q, struct request *rq, struct scatterlist *sg)
1317 {
1318         struct bio_vec *bvec, *bvprv;
1319         struct bio *bio;
1320         int nsegs, i, cluster;
1321
1322         nsegs = 0;
1323         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1324
1325         /*
1326          * for each bio in rq
1327          */
1328         bvprv = NULL;
1329         rq_for_each_bio(bio, rq) {
1330                 /*
1331                  * for each segment in bio
1332                  */
1333                 bio_for_each_segment(bvec, bio, i) {
1334                         int nbytes = bvec->bv_len;
1335
1336                         if (bvprv && cluster) {
1337                                 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1338                                         goto new_segment;
1339
1340                                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1341                                         goto new_segment;
1342                                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1343                                         goto new_segment;
1344
1345                                 sg[nsegs - 1].length += nbytes;
1346                         } else {
1347 new_segment:
1348                                 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1349                                 sg[nsegs].page = bvec->bv_page;
1350                                 sg[nsegs].length = nbytes;
1351                                 sg[nsegs].offset = bvec->bv_offset;
1352
1353                                 nsegs++;
1354                         }
1355                         bvprv = bvec;
1356                 } /* segments in bio */
1357         } /* bios in rq */
1358
1359         return nsegs;
1360 }
1361
1362 EXPORT_SYMBOL(blk_rq_map_sg);
1363
1364 /*
1365  * the standard queue merge functions, can be overridden with device
1366  * specific ones if so desired
1367  */
1368
1369 static inline int ll_new_mergeable(request_queue_t *q,
1370                                    struct request *req,
1371                                    struct bio *bio)
1372 {
1373         int nr_phys_segs = bio_phys_segments(q, bio);
1374
1375         if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1376                 req->flags |= REQ_NOMERGE;
1377                 if (req == q->last_merge)
1378                         q->last_merge = NULL;
1379                 return 0;
1380         }
1381
1382         /*
1383          * A hw segment is just getting larger, bump just the phys
1384          * counter.
1385          */
1386         req->nr_phys_segments += nr_phys_segs;
1387         return 1;
1388 }
1389
1390 static inline int ll_new_hw_segment(request_queue_t *q,
1391                                     struct request *req,
1392                                     struct bio *bio)
1393 {
1394         int nr_hw_segs = bio_hw_segments(q, bio);
1395         int nr_phys_segs = bio_phys_segments(q, bio);
1396
1397         if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1398             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1399                 req->flags |= REQ_NOMERGE;
1400                 if (req == q->last_merge)
1401                         q->last_merge = NULL;
1402                 return 0;
1403         }
1404
1405         /*
1406          * This will form the start of a new hw segment.  Bump both
1407          * counters.
1408          */
1409         req->nr_hw_segments += nr_hw_segs;
1410         req->nr_phys_segments += nr_phys_segs;
1411         return 1;
1412 }
1413
1414 static int ll_back_merge_fn(request_queue_t *q, struct request *req, 
1415                             struct bio *bio)
1416 {
1417         unsigned short max_sectors;
1418         int len;
1419
1420         if (unlikely(blk_pc_request(req)))
1421                 max_sectors = q->max_hw_sectors;
1422         else
1423                 max_sectors = q->max_sectors;
1424
1425         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1426                 req->flags |= REQ_NOMERGE;
1427                 if (req == q->last_merge)
1428                         q->last_merge = NULL;
1429                 return 0;
1430         }
1431         if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1432                 blk_recount_segments(q, req->biotail);
1433         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1434                 blk_recount_segments(q, bio);
1435         len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1436         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1437             !BIOVEC_VIRT_OVERSIZE(len)) {
1438                 int mergeable =  ll_new_mergeable(q, req, bio);
1439
1440                 if (mergeable) {
1441                         if (req->nr_hw_segments == 1)
1442                                 req->bio->bi_hw_front_size = len;
1443                         if (bio->bi_hw_segments == 1)
1444                                 bio->bi_hw_back_size = len;
1445                 }
1446                 return mergeable;
1447         }
1448
1449         return ll_new_hw_segment(q, req, bio);
1450 }
1451
1452 static int ll_front_merge_fn(request_queue_t *q, struct request *req, 
1453                              struct bio *bio)
1454 {
1455         unsigned short max_sectors;
1456         int len;
1457
1458         if (unlikely(blk_pc_request(req)))
1459                 max_sectors = q->max_hw_sectors;
1460         else
1461                 max_sectors = q->max_sectors;
1462
1463
1464         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1465                 req->flags |= REQ_NOMERGE;
1466                 if (req == q->last_merge)
1467                         q->last_merge = NULL;
1468                 return 0;
1469         }
1470         len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1471         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1472                 blk_recount_segments(q, bio);
1473         if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1474                 blk_recount_segments(q, req->bio);
1475         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1476             !BIOVEC_VIRT_OVERSIZE(len)) {
1477                 int mergeable =  ll_new_mergeable(q, req, bio);
1478
1479                 if (mergeable) {
1480                         if (bio->bi_hw_segments == 1)
1481                                 bio->bi_hw_front_size = len;
1482                         if (req->nr_hw_segments == 1)
1483                                 req->biotail->bi_hw_back_size = len;
1484                 }
1485                 return mergeable;
1486         }
1487
1488         return ll_new_hw_segment(q, req, bio);
1489 }
1490
1491 static int ll_merge_requests_fn(request_queue_t *q, struct request *req,
1492                                 struct request *next)
1493 {
1494         int total_phys_segments;
1495         int total_hw_segments;
1496
1497         /*
1498          * First check if the either of the requests are re-queued
1499          * requests.  Can't merge them if they are.
1500          */
1501         if (req->special || next->special)
1502                 return 0;
1503
1504         /*
1505          * Will it become too large?
1506          */
1507         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1508                 return 0;
1509
1510         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1511         if (blk_phys_contig_segment(q, req->biotail, next->bio))
1512                 total_phys_segments--;
1513
1514         if (total_phys_segments > q->max_phys_segments)
1515                 return 0;
1516
1517         total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1518         if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1519                 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1520                 /*
1521                  * propagate the combined length to the end of the requests
1522                  */
1523                 if (req->nr_hw_segments == 1)
1524                         req->bio->bi_hw_front_size = len;
1525                 if (next->nr_hw_segments == 1)
1526                         next->biotail->bi_hw_back_size = len;
1527                 total_hw_segments--;
1528         }
1529
1530         if (total_hw_segments > q->max_hw_segments)
1531                 return 0;
1532
1533         /* Merge is OK... */
1534         req->nr_phys_segments = total_phys_segments;
1535         req->nr_hw_segments = total_hw_segments;
1536         return 1;
1537 }
1538
1539 /*
1540  * "plug" the device if there are no outstanding requests: this will
1541  * force the transfer to start only after we have put all the requests
1542  * on the list.
1543  *
1544  * This is called with interrupts off and no requests on the queue and
1545  * with the queue lock held.
1546  */
1547 void blk_plug_device(request_queue_t *q)
1548 {
1549         WARN_ON(!irqs_disabled());
1550
1551         /*
1552          * don't plug a stopped queue, it must be paired with blk_start_queue()
1553          * which will restart the queueing
1554          */
1555         if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
1556                 return;
1557
1558         if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
1559                 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1560                 blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
1561         }
1562 }
1563
1564 EXPORT_SYMBOL(blk_plug_device);
1565
1566 /*
1567  * remove the queue from the plugged list, if present. called with
1568  * queue lock held and interrupts disabled.
1569  */
1570 int blk_remove_plug(request_queue_t *q)
1571 {
1572         WARN_ON(!irqs_disabled());
1573
1574         if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1575                 return 0;
1576
1577         del_timer(&q->unplug_timer);
1578         return 1;
1579 }
1580
1581 EXPORT_SYMBOL(blk_remove_plug);
1582
1583 /*
1584  * remove the plug and let it rip..
1585  */
1586 void __generic_unplug_device(request_queue_t *q)
1587 {
1588         if (unlikely(test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags)))
1589                 return;
1590
1591         if (!blk_remove_plug(q))
1592                 return;
1593
1594         q->request_fn(q);
1595 }
1596 EXPORT_SYMBOL(__generic_unplug_device);
1597
1598 /**
1599  * generic_unplug_device - fire a request queue
1600  * @q:    The &request_queue_t in question
1601  *
1602  * Description:
1603  *   Linux uses plugging to build bigger requests queues before letting
1604  *   the device have at them. If a queue is plugged, the I/O scheduler
1605  *   is still adding and merging requests on the queue. Once the queue
1606  *   gets unplugged, the request_fn defined for the queue is invoked and
1607  *   transfers started.
1608  **/
1609 void generic_unplug_device(request_queue_t *q)
1610 {
1611         spin_lock_irq(q->queue_lock);
1612         __generic_unplug_device(q);
1613         spin_unlock_irq(q->queue_lock);
1614 }
1615 EXPORT_SYMBOL(generic_unplug_device);
1616
1617 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1618                                    struct page *page)
1619 {
1620         request_queue_t *q = bdi->unplug_io_data;
1621
1622         /*
1623          * devices don't necessarily have an ->unplug_fn defined
1624          */
1625         if (q->unplug_fn) {
1626                 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1627                                         q->rq.count[READ] + q->rq.count[WRITE]);
1628
1629                 q->unplug_fn(q);
1630         }
1631 }
1632
1633 static void blk_unplug_work(void *data)
1634 {
1635         request_queue_t *q = data;
1636
1637         blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1638                                 q->rq.count[READ] + q->rq.count[WRITE]);
1639
1640         q->unplug_fn(q);
1641 }
1642
1643 static void blk_unplug_timeout(unsigned long data)
1644 {
1645         request_queue_t *q = (request_queue_t *)data;
1646
1647         blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
1648                                 q->rq.count[READ] + q->rq.count[WRITE]);
1649
1650         kblockd_schedule_work(&q->unplug_work);
1651 }
1652
1653 /**
1654  * blk_start_queue - restart a previously stopped queue
1655  * @q:    The &request_queue_t in question
1656  *
1657  * Description:
1658  *   blk_start_queue() will clear the stop flag on the queue, and call
1659  *   the request_fn for the queue if it was in a stopped state when
1660  *   entered. Also see blk_stop_queue(). Queue lock must be held.
1661  **/
1662 void blk_start_queue(request_queue_t *q)
1663 {
1664         clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1665
1666         /*
1667          * one level of recursion is ok and is much faster than kicking
1668          * the unplug handling
1669          */
1670         if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1671                 q->request_fn(q);
1672                 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1673         } else {
1674                 blk_plug_device(q);
1675                 kblockd_schedule_work(&q->unplug_work);
1676         }
1677 }
1678
1679 EXPORT_SYMBOL(blk_start_queue);
1680
1681 /**
1682  * blk_stop_queue - stop a queue
1683  * @q:    The &request_queue_t in question
1684  *
1685  * Description:
1686  *   The Linux block layer assumes that a block driver will consume all
1687  *   entries on the request queue when the request_fn strategy is called.
1688  *   Often this will not happen, because of hardware limitations (queue
1689  *   depth settings). If a device driver gets a 'queue full' response,
1690  *   or if it simply chooses not to queue more I/O at one point, it can
1691  *   call this function to prevent the request_fn from being called until
1692  *   the driver has signalled it's ready to go again. This happens by calling
1693  *   blk_start_queue() to restart queue operations. Queue lock must be held.
1694  **/
1695 void blk_stop_queue(request_queue_t *q)
1696 {
1697         blk_remove_plug(q);
1698         set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1699 }
1700 EXPORT_SYMBOL(blk_stop_queue);
1701
1702 /**
1703  * blk_sync_queue - cancel any pending callbacks on a queue
1704  * @q: the queue
1705  *
1706  * Description:
1707  *     The block layer may perform asynchronous callback activity
1708  *     on a queue, such as calling the unplug function after a timeout.
1709  *     A block device may call blk_sync_queue to ensure that any
1710  *     such activity is cancelled, thus allowing it to release resources
1711  *     the the callbacks might use. The caller must already have made sure
1712  *     that its ->make_request_fn will not re-add plugging prior to calling
1713  *     this function.
1714  *
1715  */
1716 void blk_sync_queue(struct request_queue *q)
1717 {
1718         del_timer_sync(&q->unplug_timer);
1719         kblockd_flush();
1720 }
1721 EXPORT_SYMBOL(blk_sync_queue);
1722
1723 /**
1724  * blk_run_queue - run a single device queue
1725  * @q:  The queue to run
1726  */
1727 void blk_run_queue(struct request_queue *q)
1728 {
1729         unsigned long flags;
1730
1731         spin_lock_irqsave(q->queue_lock, flags);
1732         blk_remove_plug(q);
1733         if (!elv_queue_empty(q))
1734                 q->request_fn(q);
1735         spin_unlock_irqrestore(q->queue_lock, flags);
1736 }
1737 EXPORT_SYMBOL(blk_run_queue);
1738
1739 /**
1740  * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1741  * @q:    the request queue to be released
1742  *
1743  * Description:
1744  *     blk_cleanup_queue is the pair to blk_init_queue() or
1745  *     blk_queue_make_request().  It should be called when a request queue is
1746  *     being released; typically when a block device is being de-registered.
1747  *     Currently, its primary task it to free all the &struct request
1748  *     structures that were allocated to the queue and the queue itself.
1749  *
1750  * Caveat:
1751  *     Hopefully the low level driver will have finished any
1752  *     outstanding requests first...
1753  **/
1754 static void blk_release_queue(struct kobject *kobj)
1755 {
1756         request_queue_t *q = container_of(kobj, struct request_queue, kobj);
1757         struct request_list *rl = &q->rq;
1758
1759         blk_sync_queue(q);
1760
1761         if (rl->rq_pool)
1762                 mempool_destroy(rl->rq_pool);
1763
1764         if (q->queue_tags)
1765                 __blk_queue_free_tags(q);
1766
1767         if (q->blk_trace)
1768                 blk_trace_shutdown(q);
1769
1770         kmem_cache_free(requestq_cachep, q);
1771 }
1772
1773 void blk_put_queue(request_queue_t *q)
1774 {
1775         kobject_put(&q->kobj);
1776 }
1777 EXPORT_SYMBOL(blk_put_queue);
1778
1779 void blk_cleanup_queue(request_queue_t * q)
1780 {
1781         mutex_lock(&q->sysfs_lock);
1782         set_bit(QUEUE_FLAG_DEAD, &q->queue_flags);
1783         mutex_unlock(&q->sysfs_lock);
1784
1785         if (q->elevator)
1786                 elevator_exit(q->elevator);
1787
1788         blk_put_queue(q);
1789 }
1790
1791 EXPORT_SYMBOL(blk_cleanup_queue);
1792
1793 static int blk_init_free_list(request_queue_t *q)
1794 {
1795         struct request_list *rl = &q->rq;
1796
1797         rl->count[READ] = rl->count[WRITE] = 0;
1798         rl->starved[READ] = rl->starved[WRITE] = 0;
1799         rl->elvpriv = 0;
1800         init_waitqueue_head(&rl->wait[READ]);
1801         init_waitqueue_head(&rl->wait[WRITE]);
1802
1803         rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1804                                 mempool_free_slab, request_cachep, q->node);
1805
1806         if (!rl->rq_pool)
1807                 return -ENOMEM;
1808
1809         return 0;
1810 }
1811
1812 request_queue_t *blk_alloc_queue(gfp_t gfp_mask)
1813 {
1814         return blk_alloc_queue_node(gfp_mask, -1);
1815 }
1816 EXPORT_SYMBOL(blk_alloc_queue);
1817
1818 static struct kobj_type queue_ktype;
1819
1820 request_queue_t *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
1821 {
1822         request_queue_t *q;
1823
1824         q = kmem_cache_alloc_node(requestq_cachep, gfp_mask, node_id);
1825         if (!q)
1826                 return NULL;
1827
1828         memset(q, 0, sizeof(*q));
1829         init_timer(&q->unplug_timer);
1830
1831         snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
1832         q->kobj.ktype = &queue_ktype;
1833         kobject_init(&q->kobj);
1834
1835         q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1836         q->backing_dev_info.unplug_io_data = q;
1837
1838         mutex_init(&q->sysfs_lock);
1839
1840         return q;
1841 }
1842 EXPORT_SYMBOL(blk_alloc_queue_node);
1843
1844 /**
1845  * blk_init_queue  - prepare a request queue for use with a block device
1846  * @rfn:  The function to be called to process requests that have been
1847  *        placed on the queue.
1848  * @lock: Request queue spin lock
1849  *
1850  * Description:
1851  *    If a block device wishes to use the standard request handling procedures,
1852  *    which sorts requests and coalesces adjacent requests, then it must
1853  *    call blk_init_queue().  The function @rfn will be called when there
1854  *    are requests on the queue that need to be processed.  If the device
1855  *    supports plugging, then @rfn may not be called immediately when requests
1856  *    are available on the queue, but may be called at some time later instead.
1857  *    Plugged queues are generally unplugged when a buffer belonging to one
1858  *    of the requests on the queue is needed, or due to memory pressure.
1859  *
1860  *    @rfn is not required, or even expected, to remove all requests off the
1861  *    queue, but only as many as it can handle at a time.  If it does leave
1862  *    requests on the queue, it is responsible for arranging that the requests
1863  *    get dealt with eventually.
1864  *
1865  *    The queue spin lock must be held while manipulating the requests on the
1866  *    request queue.
1867  *
1868  *    Function returns a pointer to the initialized request queue, or NULL if
1869  *    it didn't succeed.
1870  *
1871  * Note:
1872  *    blk_init_queue() must be paired with a blk_cleanup_queue() call
1873  *    when the block device is deactivated (such as at module unload).
1874  **/
1875
1876 request_queue_t *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1877 {
1878         return blk_init_queue_node(rfn, lock, -1);
1879 }
1880 EXPORT_SYMBOL(blk_init_queue);
1881
1882 request_queue_t *
1883 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1884 {
1885         request_queue_t *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1886
1887         if (!q)
1888                 return NULL;
1889
1890         q->node = node_id;
1891         if (blk_init_free_list(q)) {
1892                 kmem_cache_free(requestq_cachep, q);
1893                 return NULL;
1894         }
1895
1896         /*
1897          * if caller didn't supply a lock, they get per-queue locking with
1898          * our embedded lock
1899          */
1900         if (!lock) {
1901                 spin_lock_init(&q->__queue_lock);
1902                 lock = &q->__queue_lock;
1903         }
1904
1905         q->request_fn           = rfn;
1906         q->back_merge_fn        = ll_back_merge_fn;
1907         q->front_merge_fn       = ll_front_merge_fn;
1908         q->merge_requests_fn    = ll_merge_requests_fn;
1909         q->prep_rq_fn           = NULL;
1910         q->unplug_fn            = generic_unplug_device;
1911         q->queue_flags          = (1 << QUEUE_FLAG_CLUSTER);
1912         q->queue_lock           = lock;
1913
1914         blk_queue_segment_boundary(q, 0xffffffff);
1915
1916         blk_queue_make_request(q, __make_request);
1917         blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1918
1919         blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1920         blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1921
1922         /*
1923          * all done
1924          */
1925         if (!elevator_init(q, NULL)) {
1926                 blk_queue_congestion_threshold(q);
1927                 return q;
1928         }
1929
1930         blk_put_queue(q);
1931         return NULL;
1932 }
1933 EXPORT_SYMBOL(blk_init_queue_node);
1934
1935 int blk_get_queue(request_queue_t *q)
1936 {
1937         if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1938                 kobject_get(&q->kobj);
1939                 return 0;
1940         }
1941
1942         return 1;
1943 }
1944
1945 EXPORT_SYMBOL(blk_get_queue);
1946
1947 static inline void blk_free_request(request_queue_t *q, struct request *rq)
1948 {
1949         if (rq->flags & REQ_ELVPRIV)
1950                 elv_put_request(q, rq);
1951         mempool_free(rq, q->rq.rq_pool);
1952 }
1953
1954 static inline struct request *
1955 blk_alloc_request(request_queue_t *q, int rw, struct bio *bio,
1956                   int priv, gfp_t gfp_mask)
1957 {
1958         struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1959
1960         if (!rq)
1961                 return NULL;
1962
1963         /*
1964          * first three bits are identical in rq->flags and bio->bi_rw,
1965          * see bio.h and blkdev.h
1966          */
1967         rq->flags = rw;
1968
1969         if (priv) {
1970                 if (unlikely(elv_set_request(q, rq, bio, gfp_mask))) {
1971                         mempool_free(rq, q->rq.rq_pool);
1972                         return NULL;
1973                 }
1974                 rq->flags |= REQ_ELVPRIV;
1975         }
1976
1977         return rq;
1978 }
1979
1980 /*
1981  * ioc_batching returns true if the ioc is a valid batching request and
1982  * should be given priority access to a request.
1983  */
1984 static inline int ioc_batching(request_queue_t *q, struct io_context *ioc)
1985 {
1986         if (!ioc)
1987                 return 0;
1988
1989         /*
1990          * Make sure the process is able to allocate at least 1 request
1991          * even if the batch times out, otherwise we could theoretically
1992          * lose wakeups.
1993          */
1994         return ioc->nr_batch_requests == q->nr_batching ||
1995                 (ioc->nr_batch_requests > 0
1996                 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
1997 }
1998
1999 /*
2000  * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
2001  * will cause the process to be a "batcher" on all queues in the system. This
2002  * is the behaviour we want though - once it gets a wakeup it should be given
2003  * a nice run.
2004  */
2005 static void ioc_set_batching(request_queue_t *q, struct io_context *ioc)
2006 {
2007         if (!ioc || ioc_batching(q, ioc))
2008                 return;
2009
2010         ioc->nr_batch_requests = q->nr_batching;
2011         ioc->last_waited = jiffies;
2012 }
2013
2014 static void __freed_request(request_queue_t *q, int rw)
2015 {
2016         struct request_list *rl = &q->rq;
2017
2018         if (rl->count[rw] < queue_congestion_off_threshold(q))
2019                 clear_queue_congested(q, rw);
2020
2021         if (rl->count[rw] + 1 <= q->nr_requests) {
2022                 if (waitqueue_active(&rl->wait[rw]))
2023                         wake_up(&rl->wait[rw]);
2024
2025                 blk_clear_queue_full(q, rw);
2026         }
2027 }
2028
2029 /*
2030  * A request has just been released.  Account for it, update the full and
2031  * congestion status, wake up any waiters.   Called under q->queue_lock.
2032  */
2033 static void freed_request(request_queue_t *q, int rw, int priv)
2034 {
2035         struct request_list *rl = &q->rq;
2036
2037         rl->count[rw]--;
2038         if (priv)
2039                 rl->elvpriv--;
2040
2041         __freed_request(q, rw);
2042
2043         if (unlikely(rl->starved[rw ^ 1]))
2044                 __freed_request(q, rw ^ 1);
2045 }
2046
2047 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2048 /*
2049  * Get a free request, queue_lock must be held.
2050  * Returns NULL on failure, with queue_lock held.
2051  * Returns !NULL on success, with queue_lock *not held*.
2052  */
2053 static struct request *get_request(request_queue_t *q, int rw, struct bio *bio,
2054                                    gfp_t gfp_mask)
2055 {
2056         struct request *rq = NULL;
2057         struct request_list *rl = &q->rq;
2058         struct io_context *ioc = NULL;
2059         int may_queue, priv;
2060
2061         may_queue = elv_may_queue(q, rw, bio);
2062         if (may_queue == ELV_MQUEUE_NO)
2063                 goto rq_starved;
2064
2065         if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
2066                 if (rl->count[rw]+1 >= q->nr_requests) {
2067                         ioc = current_io_context(GFP_ATOMIC);
2068                         /*
2069                          * The queue will fill after this allocation, so set
2070                          * it as full, and mark this process as "batching".
2071                          * This process will be allowed to complete a batch of
2072                          * requests, others will be blocked.
2073                          */
2074                         if (!blk_queue_full(q, rw)) {
2075                                 ioc_set_batching(q, ioc);
2076                                 blk_set_queue_full(q, rw);
2077                         } else {
2078                                 if (may_queue != ELV_MQUEUE_MUST
2079                                                 && !ioc_batching(q, ioc)) {
2080                                         /*
2081                                          * The queue is full and the allocating
2082                                          * process is not a "batcher", and not
2083                                          * exempted by the IO scheduler
2084                                          */
2085                                         goto out;
2086                                 }
2087                         }
2088                 }
2089                 set_queue_congested(q, rw);
2090         }
2091
2092         /*
2093          * Only allow batching queuers to allocate up to 50% over the defined
2094          * limit of requests, otherwise we could have thousands of requests
2095          * allocated with any setting of ->nr_requests
2096          */
2097         if (rl->count[rw] >= (3 * q->nr_requests / 2))
2098                 goto out;
2099
2100         rl->count[rw]++;
2101         rl->starved[rw] = 0;
2102
2103         priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
2104         if (priv)
2105                 rl->elvpriv++;
2106
2107         spin_unlock_irq(q->queue_lock);
2108
2109         rq = blk_alloc_request(q, rw, bio, priv, gfp_mask);
2110         if (unlikely(!rq)) {
2111                 /*
2112                  * Allocation failed presumably due to memory. Undo anything
2113                  * we might have messed up.
2114                  *
2115                  * Allocating task should really be put onto the front of the
2116                  * wait queue, but this is pretty rare.
2117                  */
2118                 spin_lock_irq(q->queue_lock);
2119                 freed_request(q, rw, priv);
2120
2121                 /*
2122                  * in the very unlikely event that allocation failed and no
2123                  * requests for this direction was pending, mark us starved
2124                  * so that freeing of a request in the other direction will
2125                  * notice us. another possible fix would be to split the
2126                  * rq mempool into READ and WRITE
2127                  */
2128 rq_starved:
2129                 if (unlikely(rl->count[rw] == 0))
2130                         rl->starved[rw] = 1;
2131
2132                 goto out;
2133         }
2134
2135         /*
2136          * ioc may be NULL here, and ioc_batching will be false. That's
2137          * OK, if the queue is under the request limit then requests need
2138          * not count toward the nr_batch_requests limit. There will always
2139          * be some limit enforced by BLK_BATCH_TIME.
2140          */
2141         if (ioc_batching(q, ioc))
2142                 ioc->nr_batch_requests--;
2143         
2144         rq_init(q, rq);
2145         rq->rl = rl;
2146
2147         blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
2148 out:
2149         return rq;
2150 }
2151
2152 /*
2153  * No available requests for this queue, unplug the device and wait for some
2154  * requests to become available.
2155  *
2156  * Called with q->queue_lock held, and returns with it unlocked.
2157  */
2158 static struct request *get_request_wait(request_queue_t *q, int rw,
2159                                         struct bio *bio)
2160 {
2161         struct request *rq;
2162
2163         rq = get_request(q, rw, bio, GFP_NOIO);
2164         while (!rq) {
2165                 DEFINE_WAIT(wait);
2166                 struct request_list *rl = &q->rq;
2167
2168                 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2169                                 TASK_UNINTERRUPTIBLE);
2170
2171                 rq = get_request(q, rw, bio, GFP_NOIO);
2172
2173                 if (!rq) {
2174                         struct io_context *ioc;
2175
2176                         blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
2177
2178                         __generic_unplug_device(q);
2179                         spin_unlock_irq(q->queue_lock);
2180                         io_schedule();
2181
2182                         /*
2183                          * After sleeping, we become a "batching" process and
2184                          * will be able to allocate at least one request, and
2185                          * up to a big batch of them for a small period time.
2186                          * See ioc_batching, ioc_set_batching
2187                          */
2188                         ioc = current_io_context(GFP_NOIO);
2189                         ioc_set_batching(q, ioc);
2190
2191                         spin_lock_irq(q->queue_lock);
2192                 }
2193                 finish_wait(&rl->wait[rw], &wait);
2194         }
2195
2196         return rq;
2197 }
2198
2199 struct request *blk_get_request(request_queue_t *q, int rw, gfp_t gfp_mask)
2200 {
2201         struct request *rq;
2202
2203         BUG_ON(rw != READ && rw != WRITE);
2204
2205         spin_lock_irq(q->queue_lock);
2206         if (gfp_mask & __GFP_WAIT) {
2207                 rq = get_request_wait(q, rw, NULL);
2208         } else {
2209                 rq = get_request(q, rw, NULL, gfp_mask);
2210                 if (!rq)
2211                         spin_unlock_irq(q->queue_lock);
2212         }
2213         /* q->queue_lock is unlocked at this point */
2214
2215         return rq;
2216 }
2217 EXPORT_SYMBOL(blk_get_request);
2218
2219 /**
2220  * blk_requeue_request - put a request back on queue
2221  * @q:          request queue where request should be inserted
2222  * @rq:         request to be inserted
2223  *
2224  * Description:
2225  *    Drivers often keep queueing requests until the hardware cannot accept
2226  *    more, when that condition happens we need to put the request back
2227  *    on the queue. Must be called with queue lock held.
2228  */
2229 void blk_requeue_request(request_queue_t *q, struct request *rq)
2230 {
2231         blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
2232
2233         if (blk_rq_tagged(rq))
2234                 blk_queue_end_tag(q, rq);
2235
2236         elv_requeue_request(q, rq);
2237 }
2238
2239 EXPORT_SYMBOL(blk_requeue_request);
2240
2241 /**
2242  * blk_insert_request - insert a special request in to a request queue
2243  * @q:          request queue where request should be inserted
2244  * @rq:         request to be inserted
2245  * @at_head:    insert request at head or tail of queue
2246  * @data:       private data
2247  *
2248  * Description:
2249  *    Many block devices need to execute commands asynchronously, so they don't
2250  *    block the whole kernel from preemption during request execution.  This is
2251  *    accomplished normally by inserting aritficial requests tagged as
2252  *    REQ_SPECIAL in to the corresponding request queue, and letting them be
2253  *    scheduled for actual execution by the request queue.
2254  *
2255  *    We have the option of inserting the head or the tail of the queue.
2256  *    Typically we use the tail for new ioctls and so forth.  We use the head
2257  *    of the queue for things like a QUEUE_FULL message from a device, or a
2258  *    host that is unable to accept a particular command.
2259  */
2260 void blk_insert_request(request_queue_t *q, struct request *rq,
2261                         int at_head, void *data)
2262 {
2263         int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2264         unsigned long flags;
2265
2266         /*
2267          * tell I/O scheduler that this isn't a regular read/write (ie it
2268          * must not attempt merges on this) and that it acts as a soft
2269          * barrier
2270          */
2271         rq->flags |= REQ_SPECIAL | REQ_SOFTBARRIER;
2272
2273         rq->special = data;
2274
2275         spin_lock_irqsave(q->queue_lock, flags);
2276
2277         /*
2278          * If command is tagged, release the tag
2279          */
2280         if (blk_rq_tagged(rq))
2281                 blk_queue_end_tag(q, rq);
2282
2283         drive_stat_acct(rq, rq->nr_sectors, 1);
2284         __elv_add_request(q, rq, where, 0);
2285
2286         if (blk_queue_plugged(q))
2287                 __generic_unplug_device(q);
2288         else
2289                 q->request_fn(q);
2290         spin_unlock_irqrestore(q->queue_lock, flags);
2291 }
2292
2293 EXPORT_SYMBOL(blk_insert_request);
2294
2295 /**
2296  * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2297  * @q:          request queue where request should be inserted
2298  * @rq:         request structure to fill
2299  * @ubuf:       the user buffer
2300  * @len:        length of user data
2301  *
2302  * Description:
2303  *    Data will be mapped directly for zero copy io, if possible. Otherwise
2304  *    a kernel bounce buffer is used.
2305  *
2306  *    A matching blk_rq_unmap_user() must be issued at the end of io, while
2307  *    still in process context.
2308  *
2309  *    Note: The mapped bio may need to be bounced through blk_queue_bounce()
2310  *    before being submitted to the device, as pages mapped may be out of
2311  *    reach. It's the callers responsibility to make sure this happens. The
2312  *    original bio must be passed back in to blk_rq_unmap_user() for proper
2313  *    unmapping.
2314  */
2315 int blk_rq_map_user(request_queue_t *q, struct request *rq, void __user *ubuf,
2316                     unsigned int len)
2317 {
2318         unsigned long uaddr;
2319         struct bio *bio;
2320         int reading;
2321
2322         if (len > (q->max_hw_sectors << 9))
2323                 return -EINVAL;
2324         if (!len || !ubuf)
2325                 return -EINVAL;
2326
2327         reading = rq_data_dir(rq) == READ;
2328
2329         /*
2330          * if alignment requirement is satisfied, map in user pages for
2331          * direct dma. else, set up kernel bounce buffers
2332          */
2333         uaddr = (unsigned long) ubuf;
2334         if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2335                 bio = bio_map_user(q, NULL, uaddr, len, reading);
2336         else
2337                 bio = bio_copy_user(q, uaddr, len, reading);
2338
2339         if (!IS_ERR(bio)) {
2340                 rq->bio = rq->biotail = bio;
2341                 blk_rq_bio_prep(q, rq, bio);
2342
2343                 rq->buffer = rq->data = NULL;
2344                 rq->data_len = len;
2345                 return 0;
2346         }
2347
2348         /*
2349          * bio is the err-ptr
2350          */
2351         return PTR_ERR(bio);
2352 }
2353
2354 EXPORT_SYMBOL(blk_rq_map_user);
2355
2356 /**
2357  * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2358  * @q:          request queue where request should be inserted
2359  * @rq:         request to map data to
2360  * @iov:        pointer to the iovec
2361  * @iov_count:  number of elements in the iovec
2362  *
2363  * Description:
2364  *    Data will be mapped directly for zero copy io, if possible. Otherwise
2365  *    a kernel bounce buffer is used.
2366  *
2367  *    A matching blk_rq_unmap_user() must be issued at the end of io, while
2368  *    still in process context.
2369  *
2370  *    Note: The mapped bio may need to be bounced through blk_queue_bounce()
2371  *    before being submitted to the device, as pages mapped may be out of
2372  *    reach. It's the callers responsibility to make sure this happens. The
2373  *    original bio must be passed back in to blk_rq_unmap_user() for proper
2374  *    unmapping.
2375  */
2376 int blk_rq_map_user_iov(request_queue_t *q, struct request *rq,
2377                         struct sg_iovec *iov, int iov_count)
2378 {
2379         struct bio *bio;
2380
2381         if (!iov || iov_count <= 0)
2382                 return -EINVAL;
2383
2384         /* we don't allow misaligned data like bio_map_user() does.  If the
2385          * user is using sg, they're expected to know the alignment constraints
2386          * and respect them accordingly */
2387         bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2388         if (IS_ERR(bio))
2389                 return PTR_ERR(bio);
2390
2391         rq->bio = rq->biotail = bio;
2392         blk_rq_bio_prep(q, rq, bio);
2393         rq->buffer = rq->data = NULL;
2394         rq->data_len = bio->bi_size;
2395         return 0;
2396 }
2397
2398 EXPORT_SYMBOL(blk_rq_map_user_iov);
2399
2400 /**
2401  * blk_rq_unmap_user - unmap a request with user data
2402  * @bio:        bio to be unmapped
2403  * @ulen:       length of user buffer
2404  *
2405  * Description:
2406  *    Unmap a bio previously mapped by blk_rq_map_user().
2407  */
2408 int blk_rq_unmap_user(struct bio *bio, unsigned int ulen)
2409 {
2410         int ret = 0;
2411
2412         if (bio) {
2413                 if (bio_flagged(bio, BIO_USER_MAPPED))
2414                         bio_unmap_user(bio);
2415                 else
2416                         ret = bio_uncopy_user(bio);
2417         }
2418
2419         return 0;
2420 }
2421
2422 EXPORT_SYMBOL(blk_rq_unmap_user);
2423
2424 /**
2425  * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2426  * @q:          request queue where request should be inserted
2427  * @rq:         request to fill
2428  * @kbuf:       the kernel buffer
2429  * @len:        length of user data
2430  * @gfp_mask:   memory allocation flags
2431  */
2432 int blk_rq_map_kern(request_queue_t *q, struct request *rq, void *kbuf,
2433                     unsigned int len, gfp_t gfp_mask)
2434 {
2435         struct bio *bio;
2436
2437         if (len > (q->max_hw_sectors << 9))
2438                 return -EINVAL;
2439         if (!len || !kbuf)
2440                 return -EINVAL;
2441
2442         bio = bio_map_kern(q, kbuf, len, gfp_mask);
2443         if (IS_ERR(bio))
2444                 return PTR_ERR(bio);
2445
2446         if (rq_data_dir(rq) == WRITE)
2447                 bio->bi_rw |= (1 << BIO_RW);
2448
2449         rq->bio = rq->biotail = bio;
2450         blk_rq_bio_prep(q, rq, bio);
2451
2452         rq->buffer = rq->data = NULL;
2453         rq->data_len = len;
2454         return 0;
2455 }
2456
2457 EXPORT_SYMBOL(blk_rq_map_kern);
2458
2459 /**
2460  * blk_execute_rq_nowait - insert a request into queue for execution
2461  * @q:          queue to insert the request in
2462  * @bd_disk:    matching gendisk
2463  * @rq:         request to insert
2464  * @at_head:    insert request at head or tail of queue
2465  * @done:       I/O completion handler
2466  *
2467  * Description:
2468  *    Insert a fully prepared request at the back of the io scheduler queue
2469  *    for execution.  Don't wait for completion.
2470  */
2471 void blk_execute_rq_nowait(request_queue_t *q, struct gendisk *bd_disk,
2472                            struct request *rq, int at_head,
2473                            rq_end_io_fn *done)
2474 {
2475         int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2476
2477         rq->rq_disk = bd_disk;
2478         rq->flags |= REQ_NOMERGE;
2479         rq->end_io = done;
2480         WARN_ON(irqs_disabled());
2481         spin_lock_irq(q->queue_lock);
2482         __elv_add_request(q, rq, where, 1);
2483         __generic_unplug_device(q);
2484         spin_unlock_irq(q->queue_lock);
2485 }
2486 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
2487
2488 /**
2489  * blk_execute_rq - insert a request into queue for execution
2490  * @q:          queue to insert the request in
2491  * @bd_disk:    matching gendisk
2492  * @rq:         request to insert
2493  * @at_head:    insert request at head or tail of queue
2494  *
2495  * Description:
2496  *    Insert a fully prepared request at the back of the io scheduler queue
2497  *    for execution and wait for completion.
2498  */
2499 int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
2500                    struct request *rq, int at_head)
2501 {
2502         DECLARE_COMPLETION(wait);
2503         char sense[SCSI_SENSE_BUFFERSIZE];
2504         int err = 0;
2505
2506         /*
2507          * we need an extra reference to the request, so we can look at
2508          * it after io completion
2509          */
2510         rq->ref_count++;
2511
2512         if (!rq->sense) {
2513                 memset(sense, 0, sizeof(sense));
2514                 rq->sense = sense;
2515                 rq->sense_len = 0;
2516         }
2517
2518         rq->waiting = &wait;
2519         blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
2520         wait_for_completion(&wait);
2521         rq->waiting = NULL;
2522
2523         if (rq->errors)
2524                 err = -EIO;
2525
2526         return err;
2527 }
2528
2529 EXPORT_SYMBOL(blk_execute_rq);
2530
2531 /**
2532  * blkdev_issue_flush - queue a flush
2533  * @bdev:       blockdev to issue flush for
2534  * @error_sector:       error sector
2535  *
2536  * Description:
2537  *    Issue a flush for the block device in question. Caller can supply
2538  *    room for storing the error offset in case of a flush error, if they
2539  *    wish to.  Caller must run wait_for_completion() on its own.
2540  */
2541 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2542 {
2543         request_queue_t *q;
2544
2545         if (bdev->bd_disk == NULL)
2546                 return -ENXIO;
2547
2548         q = bdev_get_queue(bdev);
2549         if (!q)
2550                 return -ENXIO;
2551         if (!q->issue_flush_fn)
2552                 return -EOPNOTSUPP;
2553
2554         return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2555 }
2556
2557 EXPORT_SYMBOL(blkdev_issue_flush);
2558
2559 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2560 {
2561         int rw = rq_data_dir(rq);
2562
2563         if (!blk_fs_request(rq) || !rq->rq_disk)
2564                 return;
2565
2566         if (!new_io) {
2567                 __disk_stat_inc(rq->rq_disk, merges[rw]);
2568         } else {
2569                 disk_round_stats(rq->rq_disk);
2570                 rq->rq_disk->in_flight++;
2571         }
2572 }
2573
2574 /*
2575  * add-request adds a request to the linked list.
2576  * queue lock is held and interrupts disabled, as we muck with the
2577  * request queue list.
2578  */
2579 static inline void add_request(request_queue_t * q, struct request * req)
2580 {
2581         drive_stat_acct(req, req->nr_sectors, 1);
2582
2583         if (q->activity_fn)
2584                 q->activity_fn(q->activity_data, rq_data_dir(req));
2585
2586         /*
2587          * elevator indicated where it wants this request to be
2588          * inserted at elevator_merge time
2589          */
2590         __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2591 }
2592  
2593 /*
2594  * disk_round_stats()   - Round off the performance stats on a struct
2595  * disk_stats.
2596  *
2597  * The average IO queue length and utilisation statistics are maintained
2598  * by observing the current state of the queue length and the amount of
2599  * time it has been in this state for.
2600  *
2601  * Normally, that accounting is done on IO completion, but that can result
2602  * in more than a second's worth of IO being accounted for within any one
2603  * second, leading to >100% utilisation.  To deal with that, we call this
2604  * function to do a round-off before returning the results when reading
2605  * /proc/diskstats.  This accounts immediately for all queue usage up to
2606  * the current jiffies and restarts the counters again.
2607  */
2608 void disk_round_stats(struct gendisk *disk)
2609 {
2610         unsigned long now = jiffies;
2611
2612         if (now == disk->stamp)
2613                 return;
2614
2615         if (disk->in_flight) {
2616                 __disk_stat_add(disk, time_in_queue,
2617                                 disk->in_flight * (now - disk->stamp));
2618                 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2619         }
2620         disk->stamp = now;
2621 }
2622
2623 EXPORT_SYMBOL_GPL(disk_round_stats);
2624
2625 /*
2626  * queue lock must be held
2627  */
2628 void __blk_put_request(request_queue_t *q, struct request *req)
2629 {
2630         struct request_list *rl = req->rl;
2631
2632         if (unlikely(!q))
2633                 return;
2634         if (unlikely(--req->ref_count))
2635                 return;
2636
2637         elv_completed_request(q, req);
2638
2639         req->rq_status = RQ_INACTIVE;
2640         req->rl = NULL;
2641
2642         /*
2643          * Request may not have originated from ll_rw_blk. if not,
2644          * it didn't come out of our reserved rq pools
2645          */
2646         if (rl) {
2647                 int rw = rq_data_dir(req);
2648                 int priv = req->flags & REQ_ELVPRIV;
2649
2650                 BUG_ON(!list_empty(&req->queuelist));
2651
2652                 blk_free_request(q, req);
2653                 freed_request(q, rw, priv);
2654         }
2655 }
2656
2657 EXPORT_SYMBOL_GPL(__blk_put_request);
2658
2659 void blk_put_request(struct request *req)
2660 {
2661         unsigned long flags;
2662         request_queue_t *q = req->q;
2663
2664         /*
2665          * Gee, IDE calls in w/ NULL q.  Fix IDE and remove the
2666          * following if (q) test.
2667          */
2668         if (q) {
2669                 spin_lock_irqsave(q->queue_lock, flags);
2670                 __blk_put_request(q, req);
2671                 spin_unlock_irqrestore(q->queue_lock, flags);
2672         }
2673 }
2674
2675 EXPORT_SYMBOL(blk_put_request);
2676
2677 /**
2678  * blk_end_sync_rq - executes a completion event on a request
2679  * @rq: request to complete
2680  * @error: end io status of the request
2681  */
2682 void blk_end_sync_rq(struct request *rq, int error)
2683 {
2684         struct completion *waiting = rq->waiting;
2685
2686         rq->waiting = NULL;
2687         __blk_put_request(rq->q, rq);
2688
2689         /*
2690          * complete last, if this is a stack request the process (and thus
2691          * the rq pointer) could be invalid right after this complete()
2692          */
2693         complete(waiting);
2694 }
2695 EXPORT_SYMBOL(blk_end_sync_rq);
2696
2697 /**
2698  * blk_congestion_wait - wait for a queue to become uncongested
2699  * @rw: READ or WRITE
2700  * @timeout: timeout in jiffies
2701  *
2702  * Waits for up to @timeout jiffies for a queue (any queue) to exit congestion.
2703  * If no queues are congested then just wait for the next request to be
2704  * returned.
2705  */
2706 long blk_congestion_wait(int rw, long timeout)
2707 {
2708         long ret;
2709         DEFINE_WAIT(wait);
2710         wait_queue_head_t *wqh = &congestion_wqh[rw];
2711
2712         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
2713         ret = io_schedule_timeout(timeout);
2714         finish_wait(wqh, &wait);
2715         return ret;
2716 }
2717
2718 EXPORT_SYMBOL(blk_congestion_wait);
2719
2720 /*
2721  * Has to be called with the request spinlock acquired
2722  */
2723 static int attempt_merge(request_queue_t *q, struct request *req,
2724                           struct request *next)
2725 {
2726         if (!rq_mergeable(req) || !rq_mergeable(next))
2727                 return 0;
2728
2729         /*
2730          * not contigious
2731          */
2732         if (req->sector + req->nr_sectors != next->sector)
2733                 return 0;
2734
2735         if (rq_data_dir(req) != rq_data_dir(next)
2736             || req->rq_disk != next->rq_disk
2737             || next->waiting || next->special)
2738                 return 0;
2739
2740         /*
2741          * If we are allowed to merge, then append bio list
2742          * from next to rq and release next. merge_requests_fn
2743          * will have updated segment counts, update sector
2744          * counts here.
2745          */
2746         if (!q->merge_requests_fn(q, req, next))
2747                 return 0;
2748
2749         /*
2750          * At this point we have either done a back merge
2751          * or front merge. We need the smaller start_time of
2752          * the merged requests to be the current request
2753          * for accounting purposes.
2754          */
2755         if (time_after(req->start_time, next->start_time))
2756                 req->start_time = next->start_time;
2757
2758         req->biotail->bi_next = next->bio;
2759         req->biotail = next->biotail;
2760
2761         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2762
2763         elv_merge_requests(q, req, next);
2764
2765         if (req->rq_disk) {
2766                 disk_round_stats(req->rq_disk);
2767                 req->rq_disk->in_flight--;
2768         }
2769
2770         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2771
2772         __blk_put_request(q, next);
2773         return 1;
2774 }
2775
2776 static inline int attempt_back_merge(request_queue_t *q, struct request *rq)
2777 {
2778         struct request *next = elv_latter_request(q, rq);
2779
2780         if (next)
2781                 return attempt_merge(q, rq, next);
2782
2783         return 0;
2784 }
2785
2786 static inline int attempt_front_merge(request_queue_t *q, struct request *rq)
2787 {
2788         struct request *prev = elv_former_request(q, rq);
2789
2790         if (prev)
2791                 return attempt_merge(q, prev, rq);
2792
2793         return 0;
2794 }
2795
2796 static void init_request_from_bio(struct request *req, struct bio *bio)
2797 {
2798         req->flags |= REQ_CMD;
2799
2800         /*
2801          * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2802          */
2803         if (bio_rw_ahead(bio) || bio_failfast(bio))
2804                 req->flags |= REQ_FAILFAST;
2805
2806         /*
2807          * REQ_BARRIER implies no merging, but lets make it explicit
2808          */
2809         if (unlikely(bio_barrier(bio)))
2810                 req->flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2811
2812         req->errors = 0;
2813         req->hard_sector = req->sector = bio->bi_sector;
2814         req->hard_nr_sectors = req->nr_sectors = bio_sectors(bio);
2815         req->current_nr_sectors = req->hard_cur_sectors = bio_cur_sectors(bio);
2816         req->nr_phys_segments = bio_phys_segments(req->q, bio);
2817         req->nr_hw_segments = bio_hw_segments(req->q, bio);
2818         req->buffer = bio_data(bio);    /* see ->buffer comment above */
2819         req->waiting = NULL;
2820         req->bio = req->biotail = bio;
2821         req->ioprio = bio_prio(bio);
2822         req->rq_disk = bio->bi_bdev->bd_disk;
2823         req->start_time = jiffies;
2824 }
2825
2826 static int __make_request(request_queue_t *q, struct bio *bio)
2827 {
2828         struct request *req;
2829         int el_ret, rw, nr_sectors, cur_nr_sectors, barrier, err, sync;
2830         unsigned short prio;
2831         sector_t sector;
2832
2833         sector = bio->bi_sector;
2834         nr_sectors = bio_sectors(bio);
2835         cur_nr_sectors = bio_cur_sectors(bio);
2836         prio = bio_prio(bio);
2837
2838         rw = bio_data_dir(bio);
2839         sync = bio_sync(bio);
2840
2841         /*
2842          * low level driver can indicate that it wants pages above a
2843          * certain limit bounced to low memory (ie for highmem, or even
2844          * ISA dma in theory)
2845          */
2846         blk_queue_bounce(q, &bio);
2847
2848         spin_lock_prefetch(q->queue_lock);
2849
2850         barrier = bio_barrier(bio);
2851         if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
2852                 err = -EOPNOTSUPP;
2853                 goto end_io;
2854         }
2855
2856         spin_lock_irq(q->queue_lock);
2857
2858         if (unlikely(barrier) || elv_queue_empty(q))
2859                 goto get_rq;
2860
2861         el_ret = elv_merge(q, &req, bio);
2862         switch (el_ret) {
2863                 case ELEVATOR_BACK_MERGE:
2864                         BUG_ON(!rq_mergeable(req));
2865
2866                         if (!q->back_merge_fn(q, req, bio))
2867                                 break;
2868
2869                         blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
2870
2871                         req->biotail->bi_next = bio;
2872                         req->biotail = bio;
2873                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2874                         req->ioprio = ioprio_best(req->ioprio, prio);
2875                         drive_stat_acct(req, nr_sectors, 0);
2876                         if (!attempt_back_merge(q, req))
2877                                 elv_merged_request(q, req);
2878                         goto out;
2879
2880                 case ELEVATOR_FRONT_MERGE:
2881                         BUG_ON(!rq_mergeable(req));
2882
2883                         if (!q->front_merge_fn(q, req, bio))
2884                                 break;
2885
2886                         blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
2887
2888                         bio->bi_next = req->bio;
2889                         req->bio = bio;
2890
2891                         /*
2892                          * may not be valid. if the low level driver said
2893                          * it didn't need a bounce buffer then it better
2894                          * not touch req->buffer either...
2895                          */
2896                         req->buffer = bio_data(bio);
2897                         req->current_nr_sectors = cur_nr_sectors;
2898                         req->hard_cur_sectors = cur_nr_sectors;
2899                         req->sector = req->hard_sector = sector;
2900                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2901                         req->ioprio = ioprio_best(req->ioprio, prio);
2902                         drive_stat_acct(req, nr_sectors, 0);
2903                         if (!attempt_front_merge(q, req))
2904                                 elv_merged_request(q, req);
2905                         goto out;
2906
2907                 /* ELV_NO_MERGE: elevator says don't/can't merge. */
2908                 default:
2909                         ;
2910         }
2911
2912 get_rq:
2913         /*
2914          * Grab a free request. This is might sleep but can not fail.
2915          * Returns with the queue unlocked.
2916          */
2917         req = get_request_wait(q, rw, bio);
2918
2919         /*
2920          * After dropping the lock and possibly sleeping here, our request
2921          * may now be mergeable after it had proven unmergeable (above).
2922          * We don't worry about that case for efficiency. It won't happen
2923          * often, and the elevators are able to handle it.
2924          */
2925         init_request_from_bio(req, bio);
2926
2927         spin_lock_irq(q->queue_lock);
2928         if (elv_queue_empty(q))
2929                 blk_plug_device(q);
2930         add_request(q, req);
2931 out:
2932         if (sync)
2933                 __generic_unplug_device(q);
2934
2935         spin_unlock_irq(q->queue_lock);
2936         return 0;
2937
2938 end_io:
2939         bio_endio(bio, nr_sectors << 9, err);
2940         return 0;
2941 }
2942
2943 /*
2944  * If bio->bi_dev is a partition, remap the location
2945  */
2946 static inline void blk_partition_remap(struct bio *bio)
2947 {
2948         struct block_device *bdev = bio->bi_bdev;
2949
2950         if (bdev != bdev->bd_contains) {
2951                 struct hd_struct *p = bdev->bd_part;
2952                 const int rw = bio_data_dir(bio);
2953
2954                 p->sectors[rw] += bio_sectors(bio);
2955                 p->ios[rw]++;
2956
2957                 bio->bi_sector += p->start_sect;
2958                 bio->bi_bdev = bdev->bd_contains;
2959         }
2960 }
2961
2962 static void handle_bad_sector(struct bio *bio)
2963 {
2964         char b[BDEVNAME_SIZE];
2965
2966         printk(KERN_INFO "attempt to access beyond end of device\n");
2967         printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
2968                         bdevname(bio->bi_bdev, b),
2969                         bio->bi_rw,
2970                         (unsigned long long)bio->bi_sector + bio_sectors(bio),
2971                         (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
2972
2973         set_bit(BIO_EOF, &bio->bi_flags);
2974 }
2975
2976 /**
2977  * generic_make_request: hand a buffer to its device driver for I/O
2978  * @bio:  The bio describing the location in memory and on the device.
2979  *
2980  * generic_make_request() is used to make I/O requests of block
2981  * devices. It is passed a &struct bio, which describes the I/O that needs
2982  * to be done.
2983  *
2984  * generic_make_request() does not return any status.  The
2985  * success/failure status of the request, along with notification of
2986  * completion, is delivered asynchronously through the bio->bi_end_io
2987  * function described (one day) else where.
2988  *
2989  * The caller of generic_make_request must make sure that bi_io_vec
2990  * are set to describe the memory buffer, and that bi_dev and bi_sector are
2991  * set to describe the device address, and the
2992  * bi_end_io and optionally bi_private are set to describe how
2993  * completion notification should be signaled.
2994  *
2995  * generic_make_request and the drivers it calls may use bi_next if this
2996  * bio happens to be merged with someone else, and may change bi_dev and
2997  * bi_sector for remaps as it sees fit.  So the values of these fields
2998  * should NOT be depended on after the call to generic_make_request.
2999  */
3000 void generic_make_request(struct bio *bio)
3001 {
3002         request_queue_t *q;
3003         sector_t maxsector;
3004         int ret, nr_sectors = bio_sectors(bio);
3005         dev_t old_dev;
3006
3007         might_sleep();
3008         /* Test device or partition size, when known. */
3009         maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3010         if (maxsector) {
3011                 sector_t sector = bio->bi_sector;
3012
3013                 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
3014                         /*
3015                          * This may well happen - the kernel calls bread()
3016                          * without checking the size of the device, e.g., when
3017                          * mounting a device.
3018                          */
3019                         handle_bad_sector(bio);
3020                         goto end_io;
3021                 }
3022         }
3023
3024         /*
3025          * Resolve the mapping until finished. (drivers are
3026          * still free to implement/resolve their own stacking
3027          * by explicitly returning 0)
3028          *
3029          * NOTE: we don't repeat the blk_size check for each new device.
3030          * Stacking drivers are expected to know what they are doing.
3031          */
3032         maxsector = -1;
3033         old_dev = 0;
3034         do {
3035                 char b[BDEVNAME_SIZE];
3036
3037                 q = bdev_get_queue(bio->bi_bdev);
3038                 if (!q) {
3039                         printk(KERN_ERR
3040                                "generic_make_request: Trying to access "
3041                                 "nonexistent block-device %s (%Lu)\n",
3042                                 bdevname(bio->bi_bdev, b),
3043                                 (long long) bio->bi_sector);
3044 end_io:
3045                         bio_endio(bio, bio->bi_size, -EIO);
3046                         break;
3047                 }
3048
3049                 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
3050                         printk("bio too big device %s (%u > %u)\n", 
3051                                 bdevname(bio->bi_bdev, b),
3052                                 bio_sectors(bio),
3053                                 q->max_hw_sectors);
3054                         goto end_io;
3055                 }
3056
3057                 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
3058                         goto end_io;
3059
3060                 /*
3061                  * If this device has partitions, remap block n
3062                  * of partition p to block n+start(p) of the disk.
3063                  */
3064                 blk_partition_remap(bio);
3065
3066                 if (maxsector != -1)
3067                         blk_add_trace_remap(q, bio, old_dev, bio->bi_sector, 
3068                                             maxsector);
3069
3070                 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
3071
3072                 maxsector = bio->bi_sector;
3073                 old_dev = bio->bi_bdev->bd_dev;
3074
3075                 ret = q->make_request_fn(q, bio);
3076         } while (ret);
3077 }
3078
3079 EXPORT_SYMBOL(generic_make_request);
3080
3081 /**
3082  * submit_bio: submit a bio to the block device layer for I/O
3083  * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3084  * @bio: The &struct bio which describes the I/O
3085  *
3086  * submit_bio() is very similar in purpose to generic_make_request(), and
3087  * uses that function to do most of the work. Both are fairly rough
3088  * interfaces, @bio must be presetup and ready for I/O.
3089  *
3090  */
3091 void submit_bio(int rw, struct bio *bio)
3092 {
3093         int count = bio_sectors(bio);
3094
3095         BIO_BUG_ON(!bio->bi_size);
3096         BIO_BUG_ON(!bio->bi_io_vec);
3097         bio->bi_rw |= rw;
3098         if (rw & WRITE)
3099                 mod_page_state(pgpgout, count);
3100         else
3101                 mod_page_state(pgpgin, count);
3102
3103         if (unlikely(block_dump)) {
3104                 char b[BDEVNAME_SIZE];
3105                 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
3106                         current->comm, current->pid,
3107                         (rw & WRITE) ? "WRITE" : "READ",
3108                         (unsigned long long)bio->bi_sector,
3109                         bdevname(bio->bi_bdev,b));
3110         }
3111
3112         generic_make_request(bio);
3113 }
3114
3115 EXPORT_SYMBOL(submit_bio);
3116
3117 static void blk_recalc_rq_segments(struct request *rq)
3118 {
3119         struct bio *bio, *prevbio = NULL;
3120         int nr_phys_segs, nr_hw_segs;
3121         unsigned int phys_size, hw_size;
3122         request_queue_t *q = rq->q;
3123
3124         if (!rq->bio)
3125                 return;
3126
3127         phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
3128         rq_for_each_bio(bio, rq) {
3129                 /* Force bio hw/phys segs to be recalculated. */
3130                 bio->bi_flags &= ~(1 << BIO_SEG_VALID);
3131
3132                 nr_phys_segs += bio_phys_segments(q, bio);
3133                 nr_hw_segs += bio_hw_segments(q, bio);
3134                 if (prevbio) {
3135                         int pseg = phys_size + prevbio->bi_size + bio->bi_size;
3136                         int hseg = hw_size + prevbio->bi_size + bio->bi_size;
3137
3138                         if (blk_phys_contig_segment(q, prevbio, bio) &&
3139                             pseg <= q->max_segment_size) {
3140                                 nr_phys_segs--;
3141                                 phys_size += prevbio->bi_size + bio->bi_size;
3142                         } else
3143                                 phys_size = 0;
3144
3145                         if (blk_hw_contig_segment(q, prevbio, bio) &&
3146                             hseg <= q->max_segment_size) {
3147                                 nr_hw_segs--;
3148                                 hw_size += prevbio->bi_size + bio->bi_size;
3149                         } else
3150                                 hw_size = 0;
3151                 }
3152                 prevbio = bio;
3153         }
3154
3155         rq->nr_phys_segments = nr_phys_segs;
3156         rq->nr_hw_segments = nr_hw_segs;
3157 }
3158
3159 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
3160 {
3161         if (blk_fs_request(rq)) {
3162                 rq->hard_sector += nsect;
3163                 rq->hard_nr_sectors -= nsect;
3164
3165                 /*
3166                  * Move the I/O submission pointers ahead if required.
3167                  */
3168                 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3169                     (rq->sector <= rq->hard_sector)) {
3170                         rq->sector = rq->hard_sector;
3171                         rq->nr_sectors = rq->hard_nr_sectors;
3172                         rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3173                         rq->current_nr_sectors = rq->hard_cur_sectors;
3174                         rq->buffer = bio_data(rq->bio);
3175                 }
3176
3177                 /*
3178                  * if total number of sectors is less than the first segment
3179                  * size, something has gone terribly wrong
3180                  */
3181                 if (rq->nr_sectors < rq->current_nr_sectors) {
3182                         printk("blk: request botched\n");
3183                         rq->nr_sectors = rq->current_nr_sectors;
3184                 }
3185         }
3186 }
3187
3188 static int __end_that_request_first(struct request *req, int uptodate,
3189                                     int nr_bytes)
3190 {
3191         int total_bytes, bio_nbytes, error, next_idx = 0;
3192         struct bio *bio;
3193
3194         blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
3195
3196         /*
3197          * extend uptodate bool to allow < 0 value to be direct io error
3198          */
3199         error = 0;
3200         if (end_io_error(uptodate))
3201                 error = !uptodate ? -EIO : uptodate;
3202
3203         /*
3204          * for a REQ_BLOCK_PC request, we want to carry any eventual
3205          * sense key with us all the way through
3206          */
3207         if (!blk_pc_request(req))
3208                 req->errors = 0;
3209
3210         if (!uptodate) {
3211                 if (blk_fs_request(req) && !(req->flags & REQ_QUIET))
3212                         printk("end_request: I/O error, dev %s, sector %llu\n",
3213                                 req->rq_disk ? req->rq_disk->disk_name : "?",
3214                                 (unsigned long long)req->sector);
3215         }
3216
3217         if (blk_fs_request(req) && req->rq_disk) {
3218                 const int rw = rq_data_dir(req);
3219
3220                 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
3221         }
3222
3223         total_bytes = bio_nbytes = 0;
3224         while ((bio = req->bio) != NULL) {
3225                 int nbytes;
3226
3227                 if (nr_bytes >= bio->bi_size) {
3228                         req->bio = bio->bi_next;
3229                         nbytes = bio->bi_size;
3230                         if (!ordered_bio_endio(req, bio, nbytes, error))
3231                                 bio_endio(bio, nbytes, error);
3232                         next_idx = 0;
3233                         bio_nbytes = 0;
3234                 } else {
3235                         int idx = bio->bi_idx + next_idx;
3236
3237                         if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3238                                 blk_dump_rq_flags(req, "__end_that");
3239                                 printk("%s: bio idx %d >= vcnt %d\n",
3240                                                 __FUNCTION__,
3241                                                 bio->bi_idx, bio->bi_vcnt);
3242                                 break;
3243                         }
3244
3245                         nbytes = bio_iovec_idx(bio, idx)->bv_len;
3246                         BIO_BUG_ON(nbytes > bio->bi_size);
3247
3248                         /*
3249                          * not a complete bvec done
3250                          */
3251                         if (unlikely(nbytes > nr_bytes)) {
3252                                 bio_nbytes += nr_bytes;
3253                                 total_bytes += nr_bytes;
3254                                 break;
3255                         }
3256
3257                         /*
3258                          * advance to the next vector
3259                          */
3260                         next_idx++;
3261                         bio_nbytes += nbytes;
3262                 }
3263
3264                 total_bytes += nbytes;
3265                 nr_bytes -= nbytes;
3266
3267                 if ((bio = req->bio)) {
3268                         /*
3269                          * end more in this run, or just return 'not-done'
3270                          */
3271                         if (unlikely(nr_bytes <= 0))
3272                                 break;
3273                 }
3274         }
3275
3276         /*
3277          * completely done
3278          */
3279         if (!req->bio)
3280                 return 0;
3281
3282         /*
3283          * if the request wasn't completed, update state
3284          */
3285         if (bio_nbytes) {
3286                 if (!ordered_bio_endio(req, bio, bio_nbytes, error))
3287                         bio_endio(bio, bio_nbytes, error);
3288                 bio->bi_idx += next_idx;
3289                 bio_iovec(bio)->bv_offset += nr_bytes;
3290                 bio_iovec(bio)->bv_len -= nr_bytes;
3291         }
3292
3293         blk_recalc_rq_sectors(req, total_bytes >> 9);
3294         blk_recalc_rq_segments(req);
3295         return 1;
3296 }
3297
3298 /**
3299  * end_that_request_first - end I/O on a request
3300  * @req:      the request being processed
3301  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3302  * @nr_sectors: number of sectors to end I/O on
3303  *
3304  * Description:
3305  *     Ends I/O on a number of sectors attached to @req, and sets it up
3306  *     for the next range of segments (if any) in the cluster.
3307  *
3308  * Return:
3309  *     0 - we are done with this request, call end_that_request_last()
3310  *     1 - still buffers pending for this request
3311  **/
3312 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3313 {
3314         return __end_that_request_first(req, uptodate, nr_sectors << 9);
3315 }
3316
3317 EXPORT_SYMBOL(end_that_request_first);
3318
3319 /**
3320  * end_that_request_chunk - end I/O on a request
3321  * @req:      the request being processed
3322  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3323  * @nr_bytes: number of bytes to complete
3324  *
3325  * Description:
3326  *     Ends I/O on a number of bytes attached to @req, and sets it up
3327  *     for the next range of segments (if any). Like end_that_request_first(),
3328  *     but deals with bytes instead of sectors.
3329  *
3330  * Return:
3331  *     0 - we are done with this request, call end_that_request_last()
3332  *     1 - still buffers pending for this request
3333  **/
3334 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3335 {
3336         return __end_that_request_first(req, uptodate, nr_bytes);
3337 }
3338
3339 EXPORT_SYMBOL(end_that_request_chunk);
3340
3341 /*
3342  * splice the completion data to a local structure and hand off to
3343  * process_completion_queue() to complete the requests
3344  */
3345 static void blk_done_softirq(struct softirq_action *h)
3346 {
3347         struct list_head *cpu_list;
3348         LIST_HEAD(local_list);
3349
3350         local_irq_disable();
3351         cpu_list = &__get_cpu_var(blk_cpu_done);
3352         list_splice_init(cpu_list, &local_list);
3353         local_irq_enable();
3354
3355         while (!list_empty(&local_list)) {
3356                 struct request *rq = list_entry(local_list.next, struct request, donelist);
3357
3358                 list_del_init(&rq->donelist);
3359                 rq->q->softirq_done_fn(rq);
3360         }
3361 }
3362
3363 #ifdef CONFIG_HOTPLUG_CPU
3364
3365 static int blk_cpu_notify(struct notifier_block *self, unsigned long action,
3366                           void *hcpu)
3367 {
3368         /*
3369          * If a CPU goes away, splice its entries to the current CPU
3370          * and trigger a run of the softirq
3371          */
3372         if (action == CPU_DEAD) {
3373                 int cpu = (unsigned long) hcpu;
3374
3375                 local_irq_disable();
3376                 list_splice_init(&per_cpu(blk_cpu_done, cpu),
3377                                  &__get_cpu_var(blk_cpu_done));
3378                 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3379                 local_irq_enable();
3380         }
3381
3382         return NOTIFY_OK;
3383 }
3384
3385
3386 static struct notifier_block __devinitdata blk_cpu_notifier = {
3387         .notifier_call  = blk_cpu_notify,
3388 };
3389
3390 #endif /* CONFIG_HOTPLUG_CPU */
3391
3392 /**
3393  * blk_complete_request - end I/O on a request
3394  * @req:      the request being processed
3395  *
3396  * Description:
3397  *     Ends all I/O on a request. It does not handle partial completions,
3398  *     unless the driver actually implements this in its completionc callback
3399  *     through requeueing. Theh actual completion happens out-of-order,
3400  *     through a softirq handler. The user must have registered a completion
3401  *     callback through blk_queue_softirq_done().
3402  **/
3403
3404 void blk_complete_request(struct request *req)
3405 {
3406         struct list_head *cpu_list;
3407         unsigned long flags;
3408
3409         BUG_ON(!req->q->softirq_done_fn);
3410                 
3411         local_irq_save(flags);
3412
3413         cpu_list = &__get_cpu_var(blk_cpu_done);
3414         list_add_tail(&req->donelist, cpu_list);
3415         raise_softirq_irqoff(BLOCK_SOFTIRQ);
3416
3417         local_irq_restore(flags);
3418 }
3419
3420 EXPORT_SYMBOL(blk_complete_request);
3421         
3422 /*
3423  * queue lock must be held
3424  */
3425 void end_that_request_last(struct request *req, int uptodate)
3426 {
3427         struct gendisk *disk = req->rq_disk;
3428         int error;
3429
3430         /*
3431          * extend uptodate bool to allow < 0 value to be direct io error
3432          */
3433         error = 0;
3434         if (end_io_error(uptodate))
3435                 error = !uptodate ? -EIO : uptodate;
3436
3437         if (unlikely(laptop_mode) && blk_fs_request(req))
3438                 laptop_io_completion();
3439
3440         if (disk && blk_fs_request(req)) {
3441                 unsigned long duration = jiffies - req->start_time;
3442                 const int rw = rq_data_dir(req);
3443
3444                 __disk_stat_inc(disk, ios[rw]);
3445                 __disk_stat_add(disk, ticks[rw], duration);
3446                 disk_round_stats(disk);
3447                 disk->in_flight--;
3448         }
3449         if (req->end_io)
3450                 req->end_io(req, error);
3451         else
3452                 __blk_put_request(req->q, req);
3453 }
3454
3455 EXPORT_SYMBOL(end_that_request_last);
3456
3457 void end_request(struct request *req, int uptodate)
3458 {
3459         if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3460                 add_disk_randomness(req->rq_disk);
3461                 blkdev_dequeue_request(req);
3462                 end_that_request_last(req, uptodate);
3463         }
3464 }
3465
3466 EXPORT_SYMBOL(end_request);
3467
3468 void blk_rq_bio_prep(request_queue_t *q, struct request *rq, struct bio *bio)
3469 {
3470         /* first three bits are identical in rq->flags and bio->bi_rw */
3471         rq->flags |= (bio->bi_rw & 7);
3472
3473         rq->nr_phys_segments = bio_phys_segments(q, bio);
3474         rq->nr_hw_segments = bio_hw_segments(q, bio);
3475         rq->current_nr_sectors = bio_cur_sectors(bio);
3476         rq->hard_cur_sectors = rq->current_nr_sectors;
3477         rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3478         rq->buffer = bio_data(bio);
3479
3480         rq->bio = rq->biotail = bio;
3481 }
3482
3483 EXPORT_SYMBOL(blk_rq_bio_prep);
3484
3485 int kblockd_schedule_work(struct work_struct *work)
3486 {
3487         return queue_work(kblockd_workqueue, work);
3488 }
3489
3490 EXPORT_SYMBOL(kblockd_schedule_work);
3491
3492 void kblockd_flush(void)
3493 {
3494         flush_workqueue(kblockd_workqueue);
3495 }
3496 EXPORT_SYMBOL(kblockd_flush);
3497
3498 int __init blk_dev_init(void)
3499 {
3500         int i;
3501
3502         kblockd_workqueue = create_workqueue("kblockd");
3503         if (!kblockd_workqueue)
3504                 panic("Failed to create kblockd\n");
3505
3506         request_cachep = kmem_cache_create("blkdev_requests",
3507                         sizeof(struct request), 0, SLAB_PANIC, NULL, NULL);
3508
3509         requestq_cachep = kmem_cache_create("blkdev_queue",
3510                         sizeof(request_queue_t), 0, SLAB_PANIC, NULL, NULL);
3511
3512         iocontext_cachep = kmem_cache_create("blkdev_ioc",
3513                         sizeof(struct io_context), 0, SLAB_PANIC, NULL, NULL);
3514
3515         for_each_cpu(i)
3516                 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3517
3518         open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
3519 #ifdef CONFIG_HOTPLUG_CPU
3520         register_cpu_notifier(&blk_cpu_notifier);
3521 #endif
3522
3523         blk_max_low_pfn = max_low_pfn;
3524         blk_max_pfn = max_pfn;
3525
3526         return 0;
3527 }
3528
3529 /*
3530  * IO Context helper functions
3531  */
3532 void put_io_context(struct io_context *ioc)
3533 {
3534         if (ioc == NULL)
3535                 return;
3536
3537         BUG_ON(atomic_read(&ioc->refcount) == 0);
3538
3539         if (atomic_dec_and_test(&ioc->refcount)) {
3540                 rcu_read_lock();
3541                 if (ioc->aic && ioc->aic->dtor)
3542                         ioc->aic->dtor(ioc->aic);
3543                 if (ioc->cic && ioc->cic->dtor)
3544                         ioc->cic->dtor(ioc->cic);
3545                 rcu_read_unlock();
3546
3547                 kmem_cache_free(iocontext_cachep, ioc);
3548         }
3549 }
3550 EXPORT_SYMBOL(put_io_context);
3551
3552 /* Called by the exitting task */
3553 void exit_io_context(void)
3554 {
3555         unsigned long flags;
3556         struct io_context *ioc;
3557
3558         local_irq_save(flags);
3559         task_lock(current);
3560         ioc = current->io_context;
3561         current->io_context = NULL;
3562         ioc->task = NULL;
3563         task_unlock(current);
3564         local_irq_restore(flags);
3565
3566         if (ioc->aic && ioc->aic->exit)
3567                 ioc->aic->exit(ioc->aic);
3568         if (ioc->cic && ioc->cic->exit)
3569                 ioc->cic->exit(ioc->cic);
3570
3571         put_io_context(ioc);
3572 }
3573
3574 /*
3575  * If the current task has no IO context then create one and initialise it.
3576  * Otherwise, return its existing IO context.
3577  *
3578  * This returned IO context doesn't have a specifically elevated refcount,
3579  * but since the current task itself holds a reference, the context can be
3580  * used in general code, so long as it stays within `current` context.
3581  */
3582 struct io_context *current_io_context(gfp_t gfp_flags)
3583 {
3584         struct task_struct *tsk = current;
3585         struct io_context *ret;
3586
3587         ret = tsk->io_context;
3588         if (likely(ret))
3589                 return ret;
3590
3591         ret = kmem_cache_alloc(iocontext_cachep, gfp_flags);
3592         if (ret) {
3593                 atomic_set(&ret->refcount, 1);
3594                 ret->task = current;
3595                 ret->set_ioprio = NULL;
3596                 ret->last_waited = jiffies; /* doesn't matter... */
3597                 ret->nr_batch_requests = 0; /* because this is 0 */
3598                 ret->aic = NULL;
3599                 ret->cic = NULL;
3600                 tsk->io_context = ret;
3601         }
3602
3603         return ret;
3604 }
3605 EXPORT_SYMBOL(current_io_context);
3606
3607 /*
3608  * If the current task has no IO context then create one and initialise it.
3609  * If it does have a context, take a ref on it.
3610  *
3611  * This is always called in the context of the task which submitted the I/O.
3612  */
3613 struct io_context *get_io_context(gfp_t gfp_flags)
3614 {
3615         struct io_context *ret;
3616         ret = current_io_context(gfp_flags);
3617         if (likely(ret))
3618                 atomic_inc(&ret->refcount);
3619         return ret;
3620 }
3621 EXPORT_SYMBOL(get_io_context);
3622
3623 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3624 {
3625         struct io_context *src = *psrc;
3626         struct io_context *dst = *pdst;
3627
3628         if (src) {
3629                 BUG_ON(atomic_read(&src->refcount) == 0);
3630                 atomic_inc(&src->refcount);
3631                 put_io_context(dst);
3632                 *pdst = src;
3633         }
3634 }
3635 EXPORT_SYMBOL(copy_io_context);
3636
3637 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3638 {
3639         struct io_context *temp;
3640         temp = *ioc1;
3641         *ioc1 = *ioc2;
3642         *ioc2 = temp;
3643 }
3644 EXPORT_SYMBOL(swap_io_context);
3645
3646 /*
3647  * sysfs parts below
3648  */
3649 struct queue_sysfs_entry {
3650         struct attribute attr;
3651         ssize_t (*show)(struct request_queue *, char *);
3652         ssize_t (*store)(struct request_queue *, const char *, size_t);
3653 };
3654
3655 static ssize_t
3656 queue_var_show(unsigned int var, char *page)
3657 {
3658         return sprintf(page, "%d\n", var);
3659 }
3660
3661 static ssize_t
3662 queue_var_store(unsigned long *var, const char *page, size_t count)
3663 {
3664         char *p = (char *) page;
3665
3666         *var = simple_strtoul(p, &p, 10);
3667         return count;
3668 }
3669
3670 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3671 {
3672         return queue_var_show(q->nr_requests, (page));
3673 }
3674
3675 static ssize_t
3676 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3677 {
3678         struct request_list *rl = &q->rq;
3679         unsigned long nr;
3680         int ret = queue_var_store(&nr, page, count);
3681         if (nr < BLKDEV_MIN_RQ)
3682                 nr = BLKDEV_MIN_RQ;
3683
3684         spin_lock_irq(q->queue_lock);
3685         q->nr_requests = nr;
3686         blk_queue_congestion_threshold(q);
3687
3688         if (rl->count[READ] >= queue_congestion_on_threshold(q))
3689                 set_queue_congested(q, READ);
3690         else if (rl->count[READ] < queue_congestion_off_threshold(q))
3691                 clear_queue_congested(q, READ);
3692
3693         if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3694                 set_queue_congested(q, WRITE);
3695         else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3696                 clear_queue_congested(q, WRITE);
3697
3698         if (rl->count[READ] >= q->nr_requests) {
3699                 blk_set_queue_full(q, READ);
3700         } else if (rl->count[READ]+1 <= q->nr_requests) {
3701                 blk_clear_queue_full(q, READ);
3702                 wake_up(&rl->wait[READ]);
3703         }
3704
3705         if (rl->count[WRITE] >= q->nr_requests) {
3706                 blk_set_queue_full(q, WRITE);
3707         } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3708                 blk_clear_queue_full(q, WRITE);
3709                 wake_up(&rl->wait[WRITE]);
3710         }
3711         spin_unlock_irq(q->queue_lock);
3712         return ret;
3713 }
3714
3715 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3716 {
3717         int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3718
3719         return queue_var_show(ra_kb, (page));
3720 }
3721
3722 static ssize_t
3723 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3724 {
3725         unsigned long ra_kb;
3726         ssize_t ret = queue_var_store(&ra_kb, page, count);
3727
3728         spin_lock_irq(q->queue_lock);
3729         if (ra_kb > (q->max_sectors >> 1))
3730                 ra_kb = (q->max_sectors >> 1);
3731
3732         q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3733         spin_unlock_irq(q->queue_lock);
3734
3735         return ret;
3736 }
3737
3738 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3739 {
3740         int max_sectors_kb = q->max_sectors >> 1;
3741
3742         return queue_var_show(max_sectors_kb, (page));
3743 }
3744
3745 static ssize_t
3746 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3747 {
3748         unsigned long max_sectors_kb,
3749                         max_hw_sectors_kb = q->max_hw_sectors >> 1,
3750                         page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3751         ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3752         int ra_kb;
3753
3754         if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3755                 return -EINVAL;
3756         /*
3757          * Take the queue lock to update the readahead and max_sectors
3758          * values synchronously:
3759          */
3760         spin_lock_irq(q->queue_lock);
3761         /*
3762          * Trim readahead window as well, if necessary:
3763          */
3764         ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3765         if (ra_kb > max_sectors_kb)
3766                 q->backing_dev_info.ra_pages =
3767                                 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3768
3769         q->max_sectors = max_sectors_kb << 1;
3770         spin_unlock_irq(q->queue_lock);
3771
3772         return ret;
3773 }
3774
3775 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3776 {
3777         int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3778
3779         return queue_var_show(max_hw_sectors_kb, (page));
3780 }
3781
3782
3783 static struct queue_sysfs_entry queue_requests_entry = {
3784         .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3785         .show = queue_requests_show,
3786         .store = queue_requests_store,
3787 };
3788
3789 static struct queue_sysfs_entry queue_ra_entry = {
3790         .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3791         .show = queue_ra_show,
3792         .store = queue_ra_store,
3793 };
3794
3795 static struct queue_sysfs_entry queue_max_sectors_entry = {
3796         .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3797         .show = queue_max_sectors_show,
3798         .store = queue_max_sectors_store,
3799 };
3800
3801 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3802         .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
3803         .show = queue_max_hw_sectors_show,
3804 };
3805
3806 static struct queue_sysfs_entry queue_iosched_entry = {
3807         .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
3808         .show = elv_iosched_show,
3809         .store = elv_iosched_store,
3810 };
3811
3812 static struct attribute *default_attrs[] = {
3813         &queue_requests_entry.attr,
3814         &queue_ra_entry.attr,
3815         &queue_max_hw_sectors_entry.attr,
3816         &queue_max_sectors_entry.attr,
3817         &queue_iosched_entry.attr,
3818         NULL,
3819 };
3820
3821 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
3822
3823 static ssize_t
3824 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
3825 {
3826         struct queue_sysfs_entry *entry = to_queue(attr);
3827         request_queue_t *q = container_of(kobj, struct request_queue, kobj);
3828         ssize_t res;
3829
3830         if (!entry->show)
3831                 return -EIO;
3832         mutex_lock(&q->sysfs_lock);
3833         if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
3834                 mutex_unlock(&q->sysfs_lock);
3835                 return -ENOENT;
3836         }
3837         res = entry->show(q, page);
3838         mutex_unlock(&q->sysfs_lock);
3839         return res;
3840 }
3841
3842 static ssize_t
3843 queue_attr_store(struct kobject *kobj, struct attribute *attr,
3844                     const char *page, size_t length)
3845 {
3846         struct queue_sysfs_entry *entry = to_queue(attr);
3847         request_queue_t *q = container_of(kobj, struct request_queue, kobj);
3848
3849         ssize_t res;
3850
3851         if (!entry->store)
3852                 return -EIO;
3853         mutex_lock(&q->sysfs_lock);
3854         if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
3855                 mutex_unlock(&q->sysfs_lock);
3856                 return -ENOENT;
3857         }
3858         res = entry->store(q, page, length);
3859         mutex_unlock(&q->sysfs_lock);
3860         return res;
3861 }
3862
3863 static struct sysfs_ops queue_sysfs_ops = {
3864         .show   = queue_attr_show,
3865         .store  = queue_attr_store,
3866 };
3867
3868 static struct kobj_type queue_ktype = {
3869         .sysfs_ops      = &queue_sysfs_ops,
3870         .default_attrs  = default_attrs,
3871         .release        = blk_release_queue,
3872 };
3873
3874 int blk_register_queue(struct gendisk *disk)
3875 {
3876         int ret;
3877
3878         request_queue_t *q = disk->queue;
3879
3880         if (!q || !q->request_fn)
3881                 return -ENXIO;
3882
3883         q->kobj.parent = kobject_get(&disk->kobj);
3884
3885         ret = kobject_add(&q->kobj);
3886         if (ret < 0)
3887                 return ret;
3888
3889         kobject_uevent(&q->kobj, KOBJ_ADD);
3890
3891         ret = elv_register_queue(q);
3892         if (ret) {
3893                 kobject_uevent(&q->kobj, KOBJ_REMOVE);
3894                 kobject_del(&q->kobj);
3895                 return ret;
3896         }
3897
3898         return 0;
3899 }
3900
3901 void blk_unregister_queue(struct gendisk *disk)
3902 {
3903         request_queue_t *q = disk->queue;
3904
3905         if (q && q->request_fn) {
3906                 elv_unregister_queue(q);
3907
3908                 kobject_uevent(&q->kobj, KOBJ_REMOVE);
3909                 kobject_del(&q->kobj);
3910                 kobject_put(&disk->kobj);
3911         }
3912 }