[PATCH] lockdep: annotate on-stack completions
[linux-2.6.git] / block / ll_rw_blk.c
1 /*
2  * Copyright (C) 1991, 1992 Linus Torvalds
3  * Copyright (C) 1994,      Karl Keyte: Added support for disk statistics
4  * Elevator latency, (C) 2000  Andrea Arcangeli <andrea@suse.de> SuSE
5  * Queue request tables / lock, selectable elevator, Jens Axboe <axboe@suse.de>
6  * kernel-doc documentation started by NeilBrown <neilb@cse.unsw.edu.au> -  July2000
7  * bio rewrite, highmem i/o, etc, Jens Axboe <axboe@suse.de> - may 2001
8  */
9
10 /*
11  * This handles all read/write requests to block devices
12  */
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/backing-dev.h>
16 #include <linux/bio.h>
17 #include <linux/blkdev.h>
18 #include <linux/highmem.h>
19 #include <linux/mm.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/string.h>
22 #include <linux/init.h>
23 #include <linux/bootmem.h>      /* for max_pfn/max_low_pfn */
24 #include <linux/completion.h>
25 #include <linux/slab.h>
26 #include <linux/swap.h>
27 #include <linux/writeback.h>
28 #include <linux/interrupt.h>
29 #include <linux/cpu.h>
30 #include <linux/blktrace_api.h>
31
32 /*
33  * for max sense size
34  */
35 #include <scsi/scsi_cmnd.h>
36
37 static void blk_unplug_work(void *data);
38 static void blk_unplug_timeout(unsigned long data);
39 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io);
40 static void init_request_from_bio(struct request *req, struct bio *bio);
41 static int __make_request(request_queue_t *q, struct bio *bio);
42
43 /*
44  * For the allocated request tables
45  */
46 static kmem_cache_t *request_cachep;
47
48 /*
49  * For queue allocation
50  */
51 static kmem_cache_t *requestq_cachep;
52
53 /*
54  * For io context allocations
55  */
56 static kmem_cache_t *iocontext_cachep;
57
58 static wait_queue_head_t congestion_wqh[2] = {
59                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
60                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
61         };
62
63 /*
64  * Controlling structure to kblockd
65  */
66 static struct workqueue_struct *kblockd_workqueue;
67
68 unsigned long blk_max_low_pfn, blk_max_pfn;
69
70 EXPORT_SYMBOL(blk_max_low_pfn);
71 EXPORT_SYMBOL(blk_max_pfn);
72
73 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
74
75 /* Amount of time in which a process may batch requests */
76 #define BLK_BATCH_TIME  (HZ/50UL)
77
78 /* Number of requests a "batching" process may submit */
79 #define BLK_BATCH_REQ   32
80
81 /*
82  * Return the threshold (number of used requests) at which the queue is
83  * considered to be congested.  It include a little hysteresis to keep the
84  * context switch rate down.
85  */
86 static inline int queue_congestion_on_threshold(struct request_queue *q)
87 {
88         return q->nr_congestion_on;
89 }
90
91 /*
92  * The threshold at which a queue is considered to be uncongested
93  */
94 static inline int queue_congestion_off_threshold(struct request_queue *q)
95 {
96         return q->nr_congestion_off;
97 }
98
99 static void blk_queue_congestion_threshold(struct request_queue *q)
100 {
101         int nr;
102
103         nr = q->nr_requests - (q->nr_requests / 8) + 1;
104         if (nr > q->nr_requests)
105                 nr = q->nr_requests;
106         q->nr_congestion_on = nr;
107
108         nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
109         if (nr < 1)
110                 nr = 1;
111         q->nr_congestion_off = nr;
112 }
113
114 /*
115  * A queue has just exitted congestion.  Note this in the global counter of
116  * congested queues, and wake up anyone who was waiting for requests to be
117  * put back.
118  */
119 static void clear_queue_congested(request_queue_t *q, int rw)
120 {
121         enum bdi_state bit;
122         wait_queue_head_t *wqh = &congestion_wqh[rw];
123
124         bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
125         clear_bit(bit, &q->backing_dev_info.state);
126         smp_mb__after_clear_bit();
127         if (waitqueue_active(wqh))
128                 wake_up(wqh);
129 }
130
131 /*
132  * A queue has just entered congestion.  Flag that in the queue's VM-visible
133  * state flags and increment the global gounter of congested queues.
134  */
135 static void set_queue_congested(request_queue_t *q, int rw)
136 {
137         enum bdi_state bit;
138
139         bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
140         set_bit(bit, &q->backing_dev_info.state);
141 }
142
143 /**
144  * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
145  * @bdev:       device
146  *
147  * Locates the passed device's request queue and returns the address of its
148  * backing_dev_info
149  *
150  * Will return NULL if the request queue cannot be located.
151  */
152 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
153 {
154         struct backing_dev_info *ret = NULL;
155         request_queue_t *q = bdev_get_queue(bdev);
156
157         if (q)
158                 ret = &q->backing_dev_info;
159         return ret;
160 }
161
162 EXPORT_SYMBOL(blk_get_backing_dev_info);
163
164 void blk_queue_activity_fn(request_queue_t *q, activity_fn *fn, void *data)
165 {
166         q->activity_fn = fn;
167         q->activity_data = data;
168 }
169
170 EXPORT_SYMBOL(blk_queue_activity_fn);
171
172 /**
173  * blk_queue_prep_rq - set a prepare_request function for queue
174  * @q:          queue
175  * @pfn:        prepare_request function
176  *
177  * It's possible for a queue to register a prepare_request callback which
178  * is invoked before the request is handed to the request_fn. The goal of
179  * the function is to prepare a request for I/O, it can be used to build a
180  * cdb from the request data for instance.
181  *
182  */
183 void blk_queue_prep_rq(request_queue_t *q, prep_rq_fn *pfn)
184 {
185         q->prep_rq_fn = pfn;
186 }
187
188 EXPORT_SYMBOL(blk_queue_prep_rq);
189
190 /**
191  * blk_queue_merge_bvec - set a merge_bvec function for queue
192  * @q:          queue
193  * @mbfn:       merge_bvec_fn
194  *
195  * Usually queues have static limitations on the max sectors or segments that
196  * we can put in a request. Stacking drivers may have some settings that
197  * are dynamic, and thus we have to query the queue whether it is ok to
198  * add a new bio_vec to a bio at a given offset or not. If the block device
199  * has such limitations, it needs to register a merge_bvec_fn to control
200  * the size of bio's sent to it. Note that a block device *must* allow a
201  * single page to be added to an empty bio. The block device driver may want
202  * to use the bio_split() function to deal with these bio's. By default
203  * no merge_bvec_fn is defined for a queue, and only the fixed limits are
204  * honored.
205  */
206 void blk_queue_merge_bvec(request_queue_t *q, merge_bvec_fn *mbfn)
207 {
208         q->merge_bvec_fn = mbfn;
209 }
210
211 EXPORT_SYMBOL(blk_queue_merge_bvec);
212
213 void blk_queue_softirq_done(request_queue_t *q, softirq_done_fn *fn)
214 {
215         q->softirq_done_fn = fn;
216 }
217
218 EXPORT_SYMBOL(blk_queue_softirq_done);
219
220 /**
221  * blk_queue_make_request - define an alternate make_request function for a device
222  * @q:  the request queue for the device to be affected
223  * @mfn: the alternate make_request function
224  *
225  * Description:
226  *    The normal way for &struct bios to be passed to a device
227  *    driver is for them to be collected into requests on a request
228  *    queue, and then to allow the device driver to select requests
229  *    off that queue when it is ready.  This works well for many block
230  *    devices. However some block devices (typically virtual devices
231  *    such as md or lvm) do not benefit from the processing on the
232  *    request queue, and are served best by having the requests passed
233  *    directly to them.  This can be achieved by providing a function
234  *    to blk_queue_make_request().
235  *
236  * Caveat:
237  *    The driver that does this *must* be able to deal appropriately
238  *    with buffers in "highmemory". This can be accomplished by either calling
239  *    __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
240  *    blk_queue_bounce() to create a buffer in normal memory.
241  **/
242 void blk_queue_make_request(request_queue_t * q, make_request_fn * mfn)
243 {
244         /*
245          * set defaults
246          */
247         q->nr_requests = BLKDEV_MAX_RQ;
248         blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
249         blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
250         q->make_request_fn = mfn;
251         q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
252         q->backing_dev_info.state = 0;
253         q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
254         blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
255         blk_queue_hardsect_size(q, 512);
256         blk_queue_dma_alignment(q, 511);
257         blk_queue_congestion_threshold(q);
258         q->nr_batching = BLK_BATCH_REQ;
259
260         q->unplug_thresh = 4;           /* hmm */
261         q->unplug_delay = (3 * HZ) / 1000;      /* 3 milliseconds */
262         if (q->unplug_delay == 0)
263                 q->unplug_delay = 1;
264
265         INIT_WORK(&q->unplug_work, blk_unplug_work, q);
266
267         q->unplug_timer.function = blk_unplug_timeout;
268         q->unplug_timer.data = (unsigned long)q;
269
270         /*
271          * by default assume old behaviour and bounce for any highmem page
272          */
273         blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
274
275         blk_queue_activity_fn(q, NULL, NULL);
276 }
277
278 EXPORT_SYMBOL(blk_queue_make_request);
279
280 static inline void rq_init(request_queue_t *q, struct request *rq)
281 {
282         INIT_LIST_HEAD(&rq->queuelist);
283         INIT_LIST_HEAD(&rq->donelist);
284
285         rq->errors = 0;
286         rq->rq_status = RQ_ACTIVE;
287         rq->bio = rq->biotail = NULL;
288         rq->ioprio = 0;
289         rq->buffer = NULL;
290         rq->ref_count = 1;
291         rq->q = q;
292         rq->waiting = NULL;
293         rq->special = NULL;
294         rq->data_len = 0;
295         rq->data = NULL;
296         rq->nr_phys_segments = 0;
297         rq->sense = NULL;
298         rq->end_io = NULL;
299         rq->end_io_data = NULL;
300         rq->completion_data = NULL;
301 }
302
303 /**
304  * blk_queue_ordered - does this queue support ordered writes
305  * @q:        the request queue
306  * @ordered:  one of QUEUE_ORDERED_*
307  * @prepare_flush_fn: rq setup helper for cache flush ordered writes
308  *
309  * Description:
310  *   For journalled file systems, doing ordered writes on a commit
311  *   block instead of explicitly doing wait_on_buffer (which is bad
312  *   for performance) can be a big win. Block drivers supporting this
313  *   feature should call this function and indicate so.
314  *
315  **/
316 int blk_queue_ordered(request_queue_t *q, unsigned ordered,
317                       prepare_flush_fn *prepare_flush_fn)
318 {
319         if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
320             prepare_flush_fn == NULL) {
321                 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
322                 return -EINVAL;
323         }
324
325         if (ordered != QUEUE_ORDERED_NONE &&
326             ordered != QUEUE_ORDERED_DRAIN &&
327             ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
328             ordered != QUEUE_ORDERED_DRAIN_FUA &&
329             ordered != QUEUE_ORDERED_TAG &&
330             ordered != QUEUE_ORDERED_TAG_FLUSH &&
331             ordered != QUEUE_ORDERED_TAG_FUA) {
332                 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
333                 return -EINVAL;
334         }
335
336         q->ordered = ordered;
337         q->next_ordered = ordered;
338         q->prepare_flush_fn = prepare_flush_fn;
339
340         return 0;
341 }
342
343 EXPORT_SYMBOL(blk_queue_ordered);
344
345 /**
346  * blk_queue_issue_flush_fn - set function for issuing a flush
347  * @q:     the request queue
348  * @iff:   the function to be called issuing the flush
349  *
350  * Description:
351  *   If a driver supports issuing a flush command, the support is notified
352  *   to the block layer by defining it through this call.
353  *
354  **/
355 void blk_queue_issue_flush_fn(request_queue_t *q, issue_flush_fn *iff)
356 {
357         q->issue_flush_fn = iff;
358 }
359
360 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
361
362 /*
363  * Cache flushing for ordered writes handling
364  */
365 inline unsigned blk_ordered_cur_seq(request_queue_t *q)
366 {
367         if (!q->ordseq)
368                 return 0;
369         return 1 << ffz(q->ordseq);
370 }
371
372 unsigned blk_ordered_req_seq(struct request *rq)
373 {
374         request_queue_t *q = rq->q;
375
376         BUG_ON(q->ordseq == 0);
377
378         if (rq == &q->pre_flush_rq)
379                 return QUEUE_ORDSEQ_PREFLUSH;
380         if (rq == &q->bar_rq)
381                 return QUEUE_ORDSEQ_BAR;
382         if (rq == &q->post_flush_rq)
383                 return QUEUE_ORDSEQ_POSTFLUSH;
384
385         if ((rq->flags & REQ_ORDERED_COLOR) ==
386             (q->orig_bar_rq->flags & REQ_ORDERED_COLOR))
387                 return QUEUE_ORDSEQ_DRAIN;
388         else
389                 return QUEUE_ORDSEQ_DONE;
390 }
391
392 void blk_ordered_complete_seq(request_queue_t *q, unsigned seq, int error)
393 {
394         struct request *rq;
395         int uptodate;
396
397         if (error && !q->orderr)
398                 q->orderr = error;
399
400         BUG_ON(q->ordseq & seq);
401         q->ordseq |= seq;
402
403         if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
404                 return;
405
406         /*
407          * Okay, sequence complete.
408          */
409         rq = q->orig_bar_rq;
410         uptodate = q->orderr ? q->orderr : 1;
411
412         q->ordseq = 0;
413
414         end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
415         end_that_request_last(rq, uptodate);
416 }
417
418 static void pre_flush_end_io(struct request *rq, int error)
419 {
420         elv_completed_request(rq->q, rq);
421         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
422 }
423
424 static void bar_end_io(struct request *rq, int error)
425 {
426         elv_completed_request(rq->q, rq);
427         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
428 }
429
430 static void post_flush_end_io(struct request *rq, int error)
431 {
432         elv_completed_request(rq->q, rq);
433         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
434 }
435
436 static void queue_flush(request_queue_t *q, unsigned which)
437 {
438         struct request *rq;
439         rq_end_io_fn *end_io;
440
441         if (which == QUEUE_ORDERED_PREFLUSH) {
442                 rq = &q->pre_flush_rq;
443                 end_io = pre_flush_end_io;
444         } else {
445                 rq = &q->post_flush_rq;
446                 end_io = post_flush_end_io;
447         }
448
449         rq_init(q, rq);
450         rq->flags = REQ_HARDBARRIER;
451         rq->elevator_private = NULL;
452         rq->rq_disk = q->bar_rq.rq_disk;
453         rq->rl = NULL;
454         rq->end_io = end_io;
455         q->prepare_flush_fn(q, rq);
456
457         elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
458 }
459
460 static inline struct request *start_ordered(request_queue_t *q,
461                                             struct request *rq)
462 {
463         q->bi_size = 0;
464         q->orderr = 0;
465         q->ordered = q->next_ordered;
466         q->ordseq |= QUEUE_ORDSEQ_STARTED;
467
468         /*
469          * Prep proxy barrier request.
470          */
471         blkdev_dequeue_request(rq);
472         q->orig_bar_rq = rq;
473         rq = &q->bar_rq;
474         rq_init(q, rq);
475         rq->flags = bio_data_dir(q->orig_bar_rq->bio);
476         rq->flags |= q->ordered & QUEUE_ORDERED_FUA ? REQ_FUA : 0;
477         rq->elevator_private = NULL;
478         rq->rl = NULL;
479         init_request_from_bio(rq, q->orig_bar_rq->bio);
480         rq->end_io = bar_end_io;
481
482         /*
483          * Queue ordered sequence.  As we stack them at the head, we
484          * need to queue in reverse order.  Note that we rely on that
485          * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
486          * request gets inbetween ordered sequence.
487          */
488         if (q->ordered & QUEUE_ORDERED_POSTFLUSH)
489                 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
490         else
491                 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
492
493         elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
494
495         if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
496                 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
497                 rq = &q->pre_flush_rq;
498         } else
499                 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
500
501         if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
502                 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
503         else
504                 rq = NULL;
505
506         return rq;
507 }
508
509 int blk_do_ordered(request_queue_t *q, struct request **rqp)
510 {
511         struct request *rq = *rqp;
512         int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
513
514         if (!q->ordseq) {
515                 if (!is_barrier)
516                         return 1;
517
518                 if (q->next_ordered != QUEUE_ORDERED_NONE) {
519                         *rqp = start_ordered(q, rq);
520                         return 1;
521                 } else {
522                         /*
523                          * This can happen when the queue switches to
524                          * ORDERED_NONE while this request is on it.
525                          */
526                         blkdev_dequeue_request(rq);
527                         end_that_request_first(rq, -EOPNOTSUPP,
528                                                rq->hard_nr_sectors);
529                         end_that_request_last(rq, -EOPNOTSUPP);
530                         *rqp = NULL;
531                         return 0;
532                 }
533         }
534
535         /*
536          * Ordered sequence in progress
537          */
538
539         /* Special requests are not subject to ordering rules. */
540         if (!blk_fs_request(rq) &&
541             rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
542                 return 1;
543
544         if (q->ordered & QUEUE_ORDERED_TAG) {
545                 /* Ordered by tag.  Blocking the next barrier is enough. */
546                 if (is_barrier && rq != &q->bar_rq)
547                         *rqp = NULL;
548         } else {
549                 /* Ordered by draining.  Wait for turn. */
550                 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
551                 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
552                         *rqp = NULL;
553         }
554
555         return 1;
556 }
557
558 static int flush_dry_bio_endio(struct bio *bio, unsigned int bytes, int error)
559 {
560         request_queue_t *q = bio->bi_private;
561         struct bio_vec *bvec;
562         int i;
563
564         /*
565          * This is dry run, restore bio_sector and size.  We'll finish
566          * this request again with the original bi_end_io after an
567          * error occurs or post flush is complete.
568          */
569         q->bi_size += bytes;
570
571         if (bio->bi_size)
572                 return 1;
573
574         /* Rewind bvec's */
575         bio->bi_idx = 0;
576         bio_for_each_segment(bvec, bio, i) {
577                 bvec->bv_len += bvec->bv_offset;
578                 bvec->bv_offset = 0;
579         }
580
581         /* Reset bio */
582         set_bit(BIO_UPTODATE, &bio->bi_flags);
583         bio->bi_size = q->bi_size;
584         bio->bi_sector -= (q->bi_size >> 9);
585         q->bi_size = 0;
586
587         return 0;
588 }
589
590 static inline int ordered_bio_endio(struct request *rq, struct bio *bio,
591                                     unsigned int nbytes, int error)
592 {
593         request_queue_t *q = rq->q;
594         bio_end_io_t *endio;
595         void *private;
596
597         if (&q->bar_rq != rq)
598                 return 0;
599
600         /*
601          * Okay, this is the barrier request in progress, dry finish it.
602          */
603         if (error && !q->orderr)
604                 q->orderr = error;
605
606         endio = bio->bi_end_io;
607         private = bio->bi_private;
608         bio->bi_end_io = flush_dry_bio_endio;
609         bio->bi_private = q;
610
611         bio_endio(bio, nbytes, error);
612
613         bio->bi_end_io = endio;
614         bio->bi_private = private;
615
616         return 1;
617 }
618
619 /**
620  * blk_queue_bounce_limit - set bounce buffer limit for queue
621  * @q:  the request queue for the device
622  * @dma_addr:   bus address limit
623  *
624  * Description:
625  *    Different hardware can have different requirements as to what pages
626  *    it can do I/O directly to. A low level driver can call
627  *    blk_queue_bounce_limit to have lower memory pages allocated as bounce
628  *    buffers for doing I/O to pages residing above @page.
629  **/
630 void blk_queue_bounce_limit(request_queue_t *q, u64 dma_addr)
631 {
632         unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
633         int dma = 0;
634
635         q->bounce_gfp = GFP_NOIO;
636 #if BITS_PER_LONG == 64
637         /* Assume anything <= 4GB can be handled by IOMMU.
638            Actually some IOMMUs can handle everything, but I don't
639            know of a way to test this here. */
640         if (bounce_pfn < (min_t(u64,0xffffffff,BLK_BOUNCE_HIGH) >> PAGE_SHIFT))
641                 dma = 1;
642         q->bounce_pfn = max_low_pfn;
643 #else
644         if (bounce_pfn < blk_max_low_pfn)
645                 dma = 1;
646         q->bounce_pfn = bounce_pfn;
647 #endif
648         if (dma) {
649                 init_emergency_isa_pool();
650                 q->bounce_gfp = GFP_NOIO | GFP_DMA;
651                 q->bounce_pfn = bounce_pfn;
652         }
653 }
654
655 EXPORT_SYMBOL(blk_queue_bounce_limit);
656
657 /**
658  * blk_queue_max_sectors - set max sectors for a request for this queue
659  * @q:  the request queue for the device
660  * @max_sectors:  max sectors in the usual 512b unit
661  *
662  * Description:
663  *    Enables a low level driver to set an upper limit on the size of
664  *    received requests.
665  **/
666 void blk_queue_max_sectors(request_queue_t *q, unsigned int max_sectors)
667 {
668         if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
669                 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
670                 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
671         }
672
673         if (BLK_DEF_MAX_SECTORS > max_sectors)
674                 q->max_hw_sectors = q->max_sectors = max_sectors;
675         else {
676                 q->max_sectors = BLK_DEF_MAX_SECTORS;
677                 q->max_hw_sectors = max_sectors;
678         }
679 }
680
681 EXPORT_SYMBOL(blk_queue_max_sectors);
682
683 /**
684  * blk_queue_max_phys_segments - set max phys segments for a request for this queue
685  * @q:  the request queue for the device
686  * @max_segments:  max number of segments
687  *
688  * Description:
689  *    Enables a low level driver to set an upper limit on the number of
690  *    physical data segments in a request.  This would be the largest sized
691  *    scatter list the driver could handle.
692  **/
693 void blk_queue_max_phys_segments(request_queue_t *q, unsigned short max_segments)
694 {
695         if (!max_segments) {
696                 max_segments = 1;
697                 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
698         }
699
700         q->max_phys_segments = max_segments;
701 }
702
703 EXPORT_SYMBOL(blk_queue_max_phys_segments);
704
705 /**
706  * blk_queue_max_hw_segments - set max hw segments for a request for this queue
707  * @q:  the request queue for the device
708  * @max_segments:  max number of segments
709  *
710  * Description:
711  *    Enables a low level driver to set an upper limit on the number of
712  *    hw data segments in a request.  This would be the largest number of
713  *    address/length pairs the host adapter can actually give as once
714  *    to the device.
715  **/
716 void blk_queue_max_hw_segments(request_queue_t *q, unsigned short max_segments)
717 {
718         if (!max_segments) {
719                 max_segments = 1;
720                 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
721         }
722
723         q->max_hw_segments = max_segments;
724 }
725
726 EXPORT_SYMBOL(blk_queue_max_hw_segments);
727
728 /**
729  * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
730  * @q:  the request queue for the device
731  * @max_size:  max size of segment in bytes
732  *
733  * Description:
734  *    Enables a low level driver to set an upper limit on the size of a
735  *    coalesced segment
736  **/
737 void blk_queue_max_segment_size(request_queue_t *q, unsigned int max_size)
738 {
739         if (max_size < PAGE_CACHE_SIZE) {
740                 max_size = PAGE_CACHE_SIZE;
741                 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
742         }
743
744         q->max_segment_size = max_size;
745 }
746
747 EXPORT_SYMBOL(blk_queue_max_segment_size);
748
749 /**
750  * blk_queue_hardsect_size - set hardware sector size for the queue
751  * @q:  the request queue for the device
752  * @size:  the hardware sector size, in bytes
753  *
754  * Description:
755  *   This should typically be set to the lowest possible sector size
756  *   that the hardware can operate on (possible without reverting to
757  *   even internal read-modify-write operations). Usually the default
758  *   of 512 covers most hardware.
759  **/
760 void blk_queue_hardsect_size(request_queue_t *q, unsigned short size)
761 {
762         q->hardsect_size = size;
763 }
764
765 EXPORT_SYMBOL(blk_queue_hardsect_size);
766
767 /*
768  * Returns the minimum that is _not_ zero, unless both are zero.
769  */
770 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
771
772 /**
773  * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
774  * @t:  the stacking driver (top)
775  * @b:  the underlying device (bottom)
776  **/
777 void blk_queue_stack_limits(request_queue_t *t, request_queue_t *b)
778 {
779         /* zero is "infinity" */
780         t->max_sectors = min_not_zero(t->max_sectors,b->max_sectors);
781         t->max_hw_sectors = min_not_zero(t->max_hw_sectors,b->max_hw_sectors);
782
783         t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
784         t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
785         t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
786         t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
787         if (!test_bit(QUEUE_FLAG_CLUSTER, &b->queue_flags))
788                 clear_bit(QUEUE_FLAG_CLUSTER, &t->queue_flags);
789 }
790
791 EXPORT_SYMBOL(blk_queue_stack_limits);
792
793 /**
794  * blk_queue_segment_boundary - set boundary rules for segment merging
795  * @q:  the request queue for the device
796  * @mask:  the memory boundary mask
797  **/
798 void blk_queue_segment_boundary(request_queue_t *q, unsigned long mask)
799 {
800         if (mask < PAGE_CACHE_SIZE - 1) {
801                 mask = PAGE_CACHE_SIZE - 1;
802                 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
803         }
804
805         q->seg_boundary_mask = mask;
806 }
807
808 EXPORT_SYMBOL(blk_queue_segment_boundary);
809
810 /**
811  * blk_queue_dma_alignment - set dma length and memory alignment
812  * @q:     the request queue for the device
813  * @mask:  alignment mask
814  *
815  * description:
816  *    set required memory and length aligment for direct dma transactions.
817  *    this is used when buiding direct io requests for the queue.
818  *
819  **/
820 void blk_queue_dma_alignment(request_queue_t *q, int mask)
821 {
822         q->dma_alignment = mask;
823 }
824
825 EXPORT_SYMBOL(blk_queue_dma_alignment);
826
827 /**
828  * blk_queue_find_tag - find a request by its tag and queue
829  * @q:   The request queue for the device
830  * @tag: The tag of the request
831  *
832  * Notes:
833  *    Should be used when a device returns a tag and you want to match
834  *    it with a request.
835  *
836  *    no locks need be held.
837  **/
838 struct request *blk_queue_find_tag(request_queue_t *q, int tag)
839 {
840         struct blk_queue_tag *bqt = q->queue_tags;
841
842         if (unlikely(bqt == NULL || tag >= bqt->real_max_depth))
843                 return NULL;
844
845         return bqt->tag_index[tag];
846 }
847
848 EXPORT_SYMBOL(blk_queue_find_tag);
849
850 /**
851  * __blk_queue_free_tags - release tag maintenance info
852  * @q:  the request queue for the device
853  *
854  *  Notes:
855  *    blk_cleanup_queue() will take care of calling this function, if tagging
856  *    has been used. So there's no need to call this directly.
857  **/
858 static void __blk_queue_free_tags(request_queue_t *q)
859 {
860         struct blk_queue_tag *bqt = q->queue_tags;
861
862         if (!bqt)
863                 return;
864
865         if (atomic_dec_and_test(&bqt->refcnt)) {
866                 BUG_ON(bqt->busy);
867                 BUG_ON(!list_empty(&bqt->busy_list));
868
869                 kfree(bqt->tag_index);
870                 bqt->tag_index = NULL;
871
872                 kfree(bqt->tag_map);
873                 bqt->tag_map = NULL;
874
875                 kfree(bqt);
876         }
877
878         q->queue_tags = NULL;
879         q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
880 }
881
882 /**
883  * blk_queue_free_tags - release tag maintenance info
884  * @q:  the request queue for the device
885  *
886  *  Notes:
887  *      This is used to disabled tagged queuing to a device, yet leave
888  *      queue in function.
889  **/
890 void blk_queue_free_tags(request_queue_t *q)
891 {
892         clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
893 }
894
895 EXPORT_SYMBOL(blk_queue_free_tags);
896
897 static int
898 init_tag_map(request_queue_t *q, struct blk_queue_tag *tags, int depth)
899 {
900         struct request **tag_index;
901         unsigned long *tag_map;
902         int nr_ulongs;
903
904         if (depth > q->nr_requests * 2) {
905                 depth = q->nr_requests * 2;
906                 printk(KERN_ERR "%s: adjusted depth to %d\n",
907                                 __FUNCTION__, depth);
908         }
909
910         tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC);
911         if (!tag_index)
912                 goto fail;
913
914         nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
915         tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
916         if (!tag_map)
917                 goto fail;
918
919         tags->real_max_depth = depth;
920         tags->max_depth = depth;
921         tags->tag_index = tag_index;
922         tags->tag_map = tag_map;
923
924         return 0;
925 fail:
926         kfree(tag_index);
927         return -ENOMEM;
928 }
929
930 /**
931  * blk_queue_init_tags - initialize the queue tag info
932  * @q:  the request queue for the device
933  * @depth:  the maximum queue depth supported
934  * @tags: the tag to use
935  **/
936 int blk_queue_init_tags(request_queue_t *q, int depth,
937                         struct blk_queue_tag *tags)
938 {
939         int rc;
940
941         BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
942
943         if (!tags && !q->queue_tags) {
944                 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
945                 if (!tags)
946                         goto fail;
947
948                 if (init_tag_map(q, tags, depth))
949                         goto fail;
950
951                 INIT_LIST_HEAD(&tags->busy_list);
952                 tags->busy = 0;
953                 atomic_set(&tags->refcnt, 1);
954         } else if (q->queue_tags) {
955                 if ((rc = blk_queue_resize_tags(q, depth)))
956                         return rc;
957                 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
958                 return 0;
959         } else
960                 atomic_inc(&tags->refcnt);
961
962         /*
963          * assign it, all done
964          */
965         q->queue_tags = tags;
966         q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
967         return 0;
968 fail:
969         kfree(tags);
970         return -ENOMEM;
971 }
972
973 EXPORT_SYMBOL(blk_queue_init_tags);
974
975 /**
976  * blk_queue_resize_tags - change the queueing depth
977  * @q:  the request queue for the device
978  * @new_depth: the new max command queueing depth
979  *
980  *  Notes:
981  *    Must be called with the queue lock held.
982  **/
983 int blk_queue_resize_tags(request_queue_t *q, int new_depth)
984 {
985         struct blk_queue_tag *bqt = q->queue_tags;
986         struct request **tag_index;
987         unsigned long *tag_map;
988         int max_depth, nr_ulongs;
989
990         if (!bqt)
991                 return -ENXIO;
992
993         /*
994          * if we already have large enough real_max_depth.  just
995          * adjust max_depth.  *NOTE* as requests with tag value
996          * between new_depth and real_max_depth can be in-flight, tag
997          * map can not be shrunk blindly here.
998          */
999         if (new_depth <= bqt->real_max_depth) {
1000                 bqt->max_depth = new_depth;
1001                 return 0;
1002         }
1003
1004         /*
1005          * save the old state info, so we can copy it back
1006          */
1007         tag_index = bqt->tag_index;
1008         tag_map = bqt->tag_map;
1009         max_depth = bqt->real_max_depth;
1010
1011         if (init_tag_map(q, bqt, new_depth))
1012                 return -ENOMEM;
1013
1014         memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
1015         nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
1016         memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
1017
1018         kfree(tag_index);
1019         kfree(tag_map);
1020         return 0;
1021 }
1022
1023 EXPORT_SYMBOL(blk_queue_resize_tags);
1024
1025 /**
1026  * blk_queue_end_tag - end tag operations for a request
1027  * @q:  the request queue for the device
1028  * @rq: the request that has completed
1029  *
1030  *  Description:
1031  *    Typically called when end_that_request_first() returns 0, meaning
1032  *    all transfers have been done for a request. It's important to call
1033  *    this function before end_that_request_last(), as that will put the
1034  *    request back on the free list thus corrupting the internal tag list.
1035  *
1036  *  Notes:
1037  *   queue lock must be held.
1038  **/
1039 void blk_queue_end_tag(request_queue_t *q, struct request *rq)
1040 {
1041         struct blk_queue_tag *bqt = q->queue_tags;
1042         int tag = rq->tag;
1043
1044         BUG_ON(tag == -1);
1045
1046         if (unlikely(tag >= bqt->real_max_depth))
1047                 /*
1048                  * This can happen after tag depth has been reduced.
1049                  * FIXME: how about a warning or info message here?
1050                  */
1051                 return;
1052
1053         if (unlikely(!__test_and_clear_bit(tag, bqt->tag_map))) {
1054                 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
1055                        __FUNCTION__, tag);
1056                 return;
1057         }
1058
1059         list_del_init(&rq->queuelist);
1060         rq->flags &= ~REQ_QUEUED;
1061         rq->tag = -1;
1062
1063         if (unlikely(bqt->tag_index[tag] == NULL))
1064                 printk(KERN_ERR "%s: tag %d is missing\n",
1065                        __FUNCTION__, tag);
1066
1067         bqt->tag_index[tag] = NULL;
1068         bqt->busy--;
1069 }
1070
1071 EXPORT_SYMBOL(blk_queue_end_tag);
1072
1073 /**
1074  * blk_queue_start_tag - find a free tag and assign it
1075  * @q:  the request queue for the device
1076  * @rq:  the block request that needs tagging
1077  *
1078  *  Description:
1079  *    This can either be used as a stand-alone helper, or possibly be
1080  *    assigned as the queue &prep_rq_fn (in which case &struct request
1081  *    automagically gets a tag assigned). Note that this function
1082  *    assumes that any type of request can be queued! if this is not
1083  *    true for your device, you must check the request type before
1084  *    calling this function.  The request will also be removed from
1085  *    the request queue, so it's the drivers responsibility to readd
1086  *    it if it should need to be restarted for some reason.
1087  *
1088  *  Notes:
1089  *   queue lock must be held.
1090  **/
1091 int blk_queue_start_tag(request_queue_t *q, struct request *rq)
1092 {
1093         struct blk_queue_tag *bqt = q->queue_tags;
1094         int tag;
1095
1096         if (unlikely((rq->flags & REQ_QUEUED))) {
1097                 printk(KERN_ERR 
1098                        "%s: request %p for device [%s] already tagged %d",
1099                        __FUNCTION__, rq,
1100                        rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
1101                 BUG();
1102         }
1103
1104         tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
1105         if (tag >= bqt->max_depth)
1106                 return 1;
1107
1108         __set_bit(tag, bqt->tag_map);
1109
1110         rq->flags |= REQ_QUEUED;
1111         rq->tag = tag;
1112         bqt->tag_index[tag] = rq;
1113         blkdev_dequeue_request(rq);
1114         list_add(&rq->queuelist, &bqt->busy_list);
1115         bqt->busy++;
1116         return 0;
1117 }
1118
1119 EXPORT_SYMBOL(blk_queue_start_tag);
1120
1121 /**
1122  * blk_queue_invalidate_tags - invalidate all pending tags
1123  * @q:  the request queue for the device
1124  *
1125  *  Description:
1126  *   Hardware conditions may dictate a need to stop all pending requests.
1127  *   In this case, we will safely clear the block side of the tag queue and
1128  *   readd all requests to the request queue in the right order.
1129  *
1130  *  Notes:
1131  *   queue lock must be held.
1132  **/
1133 void blk_queue_invalidate_tags(request_queue_t *q)
1134 {
1135         struct blk_queue_tag *bqt = q->queue_tags;
1136         struct list_head *tmp, *n;
1137         struct request *rq;
1138
1139         list_for_each_safe(tmp, n, &bqt->busy_list) {
1140                 rq = list_entry_rq(tmp);
1141
1142                 if (rq->tag == -1) {
1143                         printk(KERN_ERR
1144                                "%s: bad tag found on list\n", __FUNCTION__);
1145                         list_del_init(&rq->queuelist);
1146                         rq->flags &= ~REQ_QUEUED;
1147                 } else
1148                         blk_queue_end_tag(q, rq);
1149
1150                 rq->flags &= ~REQ_STARTED;
1151                 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1152         }
1153 }
1154
1155 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1156
1157 static const char * const rq_flags[] = {
1158         "REQ_RW",
1159         "REQ_FAILFAST",
1160         "REQ_SORTED",
1161         "REQ_SOFTBARRIER",
1162         "REQ_HARDBARRIER",
1163         "REQ_FUA",
1164         "REQ_CMD",
1165         "REQ_NOMERGE",
1166         "REQ_STARTED",
1167         "REQ_DONTPREP",
1168         "REQ_QUEUED",
1169         "REQ_ELVPRIV",
1170         "REQ_PC",
1171         "REQ_BLOCK_PC",
1172         "REQ_SENSE",
1173         "REQ_FAILED",
1174         "REQ_QUIET",
1175         "REQ_SPECIAL",
1176         "REQ_DRIVE_CMD",
1177         "REQ_DRIVE_TASK",
1178         "REQ_DRIVE_TASKFILE",
1179         "REQ_PREEMPT",
1180         "REQ_PM_SUSPEND",
1181         "REQ_PM_RESUME",
1182         "REQ_PM_SHUTDOWN",
1183         "REQ_ORDERED_COLOR",
1184 };
1185
1186 void blk_dump_rq_flags(struct request *rq, char *msg)
1187 {
1188         int bit;
1189
1190         printk("%s: dev %s: flags = ", msg,
1191                 rq->rq_disk ? rq->rq_disk->disk_name : "?");
1192         bit = 0;
1193         do {
1194                 if (rq->flags & (1 << bit))
1195                         printk("%s ", rq_flags[bit]);
1196                 bit++;
1197         } while (bit < __REQ_NR_BITS);
1198
1199         printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1200                                                        rq->nr_sectors,
1201                                                        rq->current_nr_sectors);
1202         printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1203
1204         if (rq->flags & (REQ_BLOCK_PC | REQ_PC)) {
1205                 printk("cdb: ");
1206                 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1207                         printk("%02x ", rq->cmd[bit]);
1208                 printk("\n");
1209         }
1210 }
1211
1212 EXPORT_SYMBOL(blk_dump_rq_flags);
1213
1214 void blk_recount_segments(request_queue_t *q, struct bio *bio)
1215 {
1216         struct bio_vec *bv, *bvprv = NULL;
1217         int i, nr_phys_segs, nr_hw_segs, seg_size, hw_seg_size, cluster;
1218         int high, highprv = 1;
1219
1220         if (unlikely(!bio->bi_io_vec))
1221                 return;
1222
1223         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1224         hw_seg_size = seg_size = nr_phys_segs = nr_hw_segs = 0;
1225         bio_for_each_segment(bv, bio, i) {
1226                 /*
1227                  * the trick here is making sure that a high page is never
1228                  * considered part of another segment, since that might
1229                  * change with the bounce page.
1230                  */
1231                 high = page_to_pfn(bv->bv_page) >= q->bounce_pfn;
1232                 if (high || highprv)
1233                         goto new_hw_segment;
1234                 if (cluster) {
1235                         if (seg_size + bv->bv_len > q->max_segment_size)
1236                                 goto new_segment;
1237                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1238                                 goto new_segment;
1239                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1240                                 goto new_segment;
1241                         if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1242                                 goto new_hw_segment;
1243
1244                         seg_size += bv->bv_len;
1245                         hw_seg_size += bv->bv_len;
1246                         bvprv = bv;
1247                         continue;
1248                 }
1249 new_segment:
1250                 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1251                     !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len)) {
1252                         hw_seg_size += bv->bv_len;
1253                 } else {
1254 new_hw_segment:
1255                         if (hw_seg_size > bio->bi_hw_front_size)
1256                                 bio->bi_hw_front_size = hw_seg_size;
1257                         hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1258                         nr_hw_segs++;
1259                 }
1260
1261                 nr_phys_segs++;
1262                 bvprv = bv;
1263                 seg_size = bv->bv_len;
1264                 highprv = high;
1265         }
1266         if (hw_seg_size > bio->bi_hw_back_size)
1267                 bio->bi_hw_back_size = hw_seg_size;
1268         if (nr_hw_segs == 1 && hw_seg_size > bio->bi_hw_front_size)
1269                 bio->bi_hw_front_size = hw_seg_size;
1270         bio->bi_phys_segments = nr_phys_segs;
1271         bio->bi_hw_segments = nr_hw_segs;
1272         bio->bi_flags |= (1 << BIO_SEG_VALID);
1273 }
1274
1275
1276 static int blk_phys_contig_segment(request_queue_t *q, struct bio *bio,
1277                                    struct bio *nxt)
1278 {
1279         if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1280                 return 0;
1281
1282         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1283                 return 0;
1284         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1285                 return 0;
1286
1287         /*
1288          * bio and nxt are contigous in memory, check if the queue allows
1289          * these two to be merged into one
1290          */
1291         if (BIO_SEG_BOUNDARY(q, bio, nxt))
1292                 return 1;
1293
1294         return 0;
1295 }
1296
1297 static int blk_hw_contig_segment(request_queue_t *q, struct bio *bio,
1298                                  struct bio *nxt)
1299 {
1300         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1301                 blk_recount_segments(q, bio);
1302         if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1303                 blk_recount_segments(q, nxt);
1304         if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1305             BIOVEC_VIRT_OVERSIZE(bio->bi_hw_front_size + bio->bi_hw_back_size))
1306                 return 0;
1307         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1308                 return 0;
1309
1310         return 1;
1311 }
1312
1313 /*
1314  * map a request to scatterlist, return number of sg entries setup. Caller
1315  * must make sure sg can hold rq->nr_phys_segments entries
1316  */
1317 int blk_rq_map_sg(request_queue_t *q, struct request *rq, struct scatterlist *sg)
1318 {
1319         struct bio_vec *bvec, *bvprv;
1320         struct bio *bio;
1321         int nsegs, i, cluster;
1322
1323         nsegs = 0;
1324         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1325
1326         /*
1327          * for each bio in rq
1328          */
1329         bvprv = NULL;
1330         rq_for_each_bio(bio, rq) {
1331                 /*
1332                  * for each segment in bio
1333                  */
1334                 bio_for_each_segment(bvec, bio, i) {
1335                         int nbytes = bvec->bv_len;
1336
1337                         if (bvprv && cluster) {
1338                                 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1339                                         goto new_segment;
1340
1341                                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1342                                         goto new_segment;
1343                                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1344                                         goto new_segment;
1345
1346                                 sg[nsegs - 1].length += nbytes;
1347                         } else {
1348 new_segment:
1349                                 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1350                                 sg[nsegs].page = bvec->bv_page;
1351                                 sg[nsegs].length = nbytes;
1352                                 sg[nsegs].offset = bvec->bv_offset;
1353
1354                                 nsegs++;
1355                         }
1356                         bvprv = bvec;
1357                 } /* segments in bio */
1358         } /* bios in rq */
1359
1360         return nsegs;
1361 }
1362
1363 EXPORT_SYMBOL(blk_rq_map_sg);
1364
1365 /*
1366  * the standard queue merge functions, can be overridden with device
1367  * specific ones if so desired
1368  */
1369
1370 static inline int ll_new_mergeable(request_queue_t *q,
1371                                    struct request *req,
1372                                    struct bio *bio)
1373 {
1374         int nr_phys_segs = bio_phys_segments(q, bio);
1375
1376         if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1377                 req->flags |= REQ_NOMERGE;
1378                 if (req == q->last_merge)
1379                         q->last_merge = NULL;
1380                 return 0;
1381         }
1382
1383         /*
1384          * A hw segment is just getting larger, bump just the phys
1385          * counter.
1386          */
1387         req->nr_phys_segments += nr_phys_segs;
1388         return 1;
1389 }
1390
1391 static inline int ll_new_hw_segment(request_queue_t *q,
1392                                     struct request *req,
1393                                     struct bio *bio)
1394 {
1395         int nr_hw_segs = bio_hw_segments(q, bio);
1396         int nr_phys_segs = bio_phys_segments(q, bio);
1397
1398         if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1399             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1400                 req->flags |= REQ_NOMERGE;
1401                 if (req == q->last_merge)
1402                         q->last_merge = NULL;
1403                 return 0;
1404         }
1405
1406         /*
1407          * This will form the start of a new hw segment.  Bump both
1408          * counters.
1409          */
1410         req->nr_hw_segments += nr_hw_segs;
1411         req->nr_phys_segments += nr_phys_segs;
1412         return 1;
1413 }
1414
1415 static int ll_back_merge_fn(request_queue_t *q, struct request *req, 
1416                             struct bio *bio)
1417 {
1418         unsigned short max_sectors;
1419         int len;
1420
1421         if (unlikely(blk_pc_request(req)))
1422                 max_sectors = q->max_hw_sectors;
1423         else
1424                 max_sectors = q->max_sectors;
1425
1426         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1427                 req->flags |= REQ_NOMERGE;
1428                 if (req == q->last_merge)
1429                         q->last_merge = NULL;
1430                 return 0;
1431         }
1432         if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1433                 blk_recount_segments(q, req->biotail);
1434         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1435                 blk_recount_segments(q, bio);
1436         len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1437         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1438             !BIOVEC_VIRT_OVERSIZE(len)) {
1439                 int mergeable =  ll_new_mergeable(q, req, bio);
1440
1441                 if (mergeable) {
1442                         if (req->nr_hw_segments == 1)
1443                                 req->bio->bi_hw_front_size = len;
1444                         if (bio->bi_hw_segments == 1)
1445                                 bio->bi_hw_back_size = len;
1446                 }
1447                 return mergeable;
1448         }
1449
1450         return ll_new_hw_segment(q, req, bio);
1451 }
1452
1453 static int ll_front_merge_fn(request_queue_t *q, struct request *req, 
1454                              struct bio *bio)
1455 {
1456         unsigned short max_sectors;
1457         int len;
1458
1459         if (unlikely(blk_pc_request(req)))
1460                 max_sectors = q->max_hw_sectors;
1461         else
1462                 max_sectors = q->max_sectors;
1463
1464
1465         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1466                 req->flags |= REQ_NOMERGE;
1467                 if (req == q->last_merge)
1468                         q->last_merge = NULL;
1469                 return 0;
1470         }
1471         len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1472         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1473                 blk_recount_segments(q, bio);
1474         if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1475                 blk_recount_segments(q, req->bio);
1476         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1477             !BIOVEC_VIRT_OVERSIZE(len)) {
1478                 int mergeable =  ll_new_mergeable(q, req, bio);
1479
1480                 if (mergeable) {
1481                         if (bio->bi_hw_segments == 1)
1482                                 bio->bi_hw_front_size = len;
1483                         if (req->nr_hw_segments == 1)
1484                                 req->biotail->bi_hw_back_size = len;
1485                 }
1486                 return mergeable;
1487         }
1488
1489         return ll_new_hw_segment(q, req, bio);
1490 }
1491
1492 static int ll_merge_requests_fn(request_queue_t *q, struct request *req,
1493                                 struct request *next)
1494 {
1495         int total_phys_segments;
1496         int total_hw_segments;
1497
1498         /*
1499          * First check if the either of the requests are re-queued
1500          * requests.  Can't merge them if they are.
1501          */
1502         if (req->special || next->special)
1503                 return 0;
1504
1505         /*
1506          * Will it become too large?
1507          */
1508         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1509                 return 0;
1510
1511         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1512         if (blk_phys_contig_segment(q, req->biotail, next->bio))
1513                 total_phys_segments--;
1514
1515         if (total_phys_segments > q->max_phys_segments)
1516                 return 0;
1517
1518         total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1519         if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1520                 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1521                 /*
1522                  * propagate the combined length to the end of the requests
1523                  */
1524                 if (req->nr_hw_segments == 1)
1525                         req->bio->bi_hw_front_size = len;
1526                 if (next->nr_hw_segments == 1)
1527                         next->biotail->bi_hw_back_size = len;
1528                 total_hw_segments--;
1529         }
1530
1531         if (total_hw_segments > q->max_hw_segments)
1532                 return 0;
1533
1534         /* Merge is OK... */
1535         req->nr_phys_segments = total_phys_segments;
1536         req->nr_hw_segments = total_hw_segments;
1537         return 1;
1538 }
1539
1540 /*
1541  * "plug" the device if there are no outstanding requests: this will
1542  * force the transfer to start only after we have put all the requests
1543  * on the list.
1544  *
1545  * This is called with interrupts off and no requests on the queue and
1546  * with the queue lock held.
1547  */
1548 void blk_plug_device(request_queue_t *q)
1549 {
1550         WARN_ON(!irqs_disabled());
1551
1552         /*
1553          * don't plug a stopped queue, it must be paired with blk_start_queue()
1554          * which will restart the queueing
1555          */
1556         if (blk_queue_stopped(q))
1557                 return;
1558
1559         if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
1560                 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1561                 blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
1562         }
1563 }
1564
1565 EXPORT_SYMBOL(blk_plug_device);
1566
1567 /*
1568  * remove the queue from the plugged list, if present. called with
1569  * queue lock held and interrupts disabled.
1570  */
1571 int blk_remove_plug(request_queue_t *q)
1572 {
1573         WARN_ON(!irqs_disabled());
1574
1575         if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1576                 return 0;
1577
1578         del_timer(&q->unplug_timer);
1579         return 1;
1580 }
1581
1582 EXPORT_SYMBOL(blk_remove_plug);
1583
1584 /*
1585  * remove the plug and let it rip..
1586  */
1587 void __generic_unplug_device(request_queue_t *q)
1588 {
1589         if (unlikely(blk_queue_stopped(q)))
1590                 return;
1591
1592         if (!blk_remove_plug(q))
1593                 return;
1594
1595         q->request_fn(q);
1596 }
1597 EXPORT_SYMBOL(__generic_unplug_device);
1598
1599 /**
1600  * generic_unplug_device - fire a request queue
1601  * @q:    The &request_queue_t in question
1602  *
1603  * Description:
1604  *   Linux uses plugging to build bigger requests queues before letting
1605  *   the device have at them. If a queue is plugged, the I/O scheduler
1606  *   is still adding and merging requests on the queue. Once the queue
1607  *   gets unplugged, the request_fn defined for the queue is invoked and
1608  *   transfers started.
1609  **/
1610 void generic_unplug_device(request_queue_t *q)
1611 {
1612         spin_lock_irq(q->queue_lock);
1613         __generic_unplug_device(q);
1614         spin_unlock_irq(q->queue_lock);
1615 }
1616 EXPORT_SYMBOL(generic_unplug_device);
1617
1618 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1619                                    struct page *page)
1620 {
1621         request_queue_t *q = bdi->unplug_io_data;
1622
1623         /*
1624          * devices don't necessarily have an ->unplug_fn defined
1625          */
1626         if (q->unplug_fn) {
1627                 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1628                                         q->rq.count[READ] + q->rq.count[WRITE]);
1629
1630                 q->unplug_fn(q);
1631         }
1632 }
1633
1634 static void blk_unplug_work(void *data)
1635 {
1636         request_queue_t *q = data;
1637
1638         blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1639                                 q->rq.count[READ] + q->rq.count[WRITE]);
1640
1641         q->unplug_fn(q);
1642 }
1643
1644 static void blk_unplug_timeout(unsigned long data)
1645 {
1646         request_queue_t *q = (request_queue_t *)data;
1647
1648         blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
1649                                 q->rq.count[READ] + q->rq.count[WRITE]);
1650
1651         kblockd_schedule_work(&q->unplug_work);
1652 }
1653
1654 /**
1655  * blk_start_queue - restart a previously stopped queue
1656  * @q:    The &request_queue_t in question
1657  *
1658  * Description:
1659  *   blk_start_queue() will clear the stop flag on the queue, and call
1660  *   the request_fn for the queue if it was in a stopped state when
1661  *   entered. Also see blk_stop_queue(). Queue lock must be held.
1662  **/
1663 void blk_start_queue(request_queue_t *q)
1664 {
1665         WARN_ON(!irqs_disabled());
1666
1667         clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1668
1669         /*
1670          * one level of recursion is ok and is much faster than kicking
1671          * the unplug handling
1672          */
1673         if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1674                 q->request_fn(q);
1675                 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1676         } else {
1677                 blk_plug_device(q);
1678                 kblockd_schedule_work(&q->unplug_work);
1679         }
1680 }
1681
1682 EXPORT_SYMBOL(blk_start_queue);
1683
1684 /**
1685  * blk_stop_queue - stop a queue
1686  * @q:    The &request_queue_t in question
1687  *
1688  * Description:
1689  *   The Linux block layer assumes that a block driver will consume all
1690  *   entries on the request queue when the request_fn strategy is called.
1691  *   Often this will not happen, because of hardware limitations (queue
1692  *   depth settings). If a device driver gets a 'queue full' response,
1693  *   or if it simply chooses not to queue more I/O at one point, it can
1694  *   call this function to prevent the request_fn from being called until
1695  *   the driver has signalled it's ready to go again. This happens by calling
1696  *   blk_start_queue() to restart queue operations. Queue lock must be held.
1697  **/
1698 void blk_stop_queue(request_queue_t *q)
1699 {
1700         blk_remove_plug(q);
1701         set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1702 }
1703 EXPORT_SYMBOL(blk_stop_queue);
1704
1705 /**
1706  * blk_sync_queue - cancel any pending callbacks on a queue
1707  * @q: the queue
1708  *
1709  * Description:
1710  *     The block layer may perform asynchronous callback activity
1711  *     on a queue, such as calling the unplug function after a timeout.
1712  *     A block device may call blk_sync_queue to ensure that any
1713  *     such activity is cancelled, thus allowing it to release resources
1714  *     the the callbacks might use. The caller must already have made sure
1715  *     that its ->make_request_fn will not re-add plugging prior to calling
1716  *     this function.
1717  *
1718  */
1719 void blk_sync_queue(struct request_queue *q)
1720 {
1721         del_timer_sync(&q->unplug_timer);
1722         kblockd_flush();
1723 }
1724 EXPORT_SYMBOL(blk_sync_queue);
1725
1726 /**
1727  * blk_run_queue - run a single device queue
1728  * @q:  The queue to run
1729  */
1730 void blk_run_queue(struct request_queue *q)
1731 {
1732         unsigned long flags;
1733
1734         spin_lock_irqsave(q->queue_lock, flags);
1735         blk_remove_plug(q);
1736
1737         /*
1738          * Only recurse once to avoid overrunning the stack, let the unplug
1739          * handling reinvoke the handler shortly if we already got there.
1740          */
1741         if (!elv_queue_empty(q)) {
1742                 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1743                         q->request_fn(q);
1744                         clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1745                 } else {
1746                         blk_plug_device(q);
1747                         kblockd_schedule_work(&q->unplug_work);
1748                 }
1749         }
1750
1751         spin_unlock_irqrestore(q->queue_lock, flags);
1752 }
1753 EXPORT_SYMBOL(blk_run_queue);
1754
1755 /**
1756  * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1757  * @kobj:    the kobj belonging of the request queue to be released
1758  *
1759  * Description:
1760  *     blk_cleanup_queue is the pair to blk_init_queue() or
1761  *     blk_queue_make_request().  It should be called when a request queue is
1762  *     being released; typically when a block device is being de-registered.
1763  *     Currently, its primary task it to free all the &struct request
1764  *     structures that were allocated to the queue and the queue itself.
1765  *
1766  * Caveat:
1767  *     Hopefully the low level driver will have finished any
1768  *     outstanding requests first...
1769  **/
1770 static void blk_release_queue(struct kobject *kobj)
1771 {
1772         request_queue_t *q = container_of(kobj, struct request_queue, kobj);
1773         struct request_list *rl = &q->rq;
1774
1775         blk_sync_queue(q);
1776
1777         if (rl->rq_pool)
1778                 mempool_destroy(rl->rq_pool);
1779
1780         if (q->queue_tags)
1781                 __blk_queue_free_tags(q);
1782
1783         if (q->blk_trace)
1784                 blk_trace_shutdown(q);
1785
1786         kmem_cache_free(requestq_cachep, q);
1787 }
1788
1789 void blk_put_queue(request_queue_t *q)
1790 {
1791         kobject_put(&q->kobj);
1792 }
1793 EXPORT_SYMBOL(blk_put_queue);
1794
1795 void blk_cleanup_queue(request_queue_t * q)
1796 {
1797         mutex_lock(&q->sysfs_lock);
1798         set_bit(QUEUE_FLAG_DEAD, &q->queue_flags);
1799         mutex_unlock(&q->sysfs_lock);
1800
1801         if (q->elevator)
1802                 elevator_exit(q->elevator);
1803
1804         blk_put_queue(q);
1805 }
1806
1807 EXPORT_SYMBOL(blk_cleanup_queue);
1808
1809 static int blk_init_free_list(request_queue_t *q)
1810 {
1811         struct request_list *rl = &q->rq;
1812
1813         rl->count[READ] = rl->count[WRITE] = 0;
1814         rl->starved[READ] = rl->starved[WRITE] = 0;
1815         rl->elvpriv = 0;
1816         init_waitqueue_head(&rl->wait[READ]);
1817         init_waitqueue_head(&rl->wait[WRITE]);
1818
1819         rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1820                                 mempool_free_slab, request_cachep, q->node);
1821
1822         if (!rl->rq_pool)
1823                 return -ENOMEM;
1824
1825         return 0;
1826 }
1827
1828 request_queue_t *blk_alloc_queue(gfp_t gfp_mask)
1829 {
1830         return blk_alloc_queue_node(gfp_mask, -1);
1831 }
1832 EXPORT_SYMBOL(blk_alloc_queue);
1833
1834 static struct kobj_type queue_ktype;
1835
1836 request_queue_t *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
1837 {
1838         request_queue_t *q;
1839
1840         q = kmem_cache_alloc_node(requestq_cachep, gfp_mask, node_id);
1841         if (!q)
1842                 return NULL;
1843
1844         memset(q, 0, sizeof(*q));
1845         init_timer(&q->unplug_timer);
1846
1847         snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
1848         q->kobj.ktype = &queue_ktype;
1849         kobject_init(&q->kobj);
1850
1851         q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1852         q->backing_dev_info.unplug_io_data = q;
1853
1854         mutex_init(&q->sysfs_lock);
1855
1856         return q;
1857 }
1858 EXPORT_SYMBOL(blk_alloc_queue_node);
1859
1860 /**
1861  * blk_init_queue  - prepare a request queue for use with a block device
1862  * @rfn:  The function to be called to process requests that have been
1863  *        placed on the queue.
1864  * @lock: Request queue spin lock
1865  *
1866  * Description:
1867  *    If a block device wishes to use the standard request handling procedures,
1868  *    which sorts requests and coalesces adjacent requests, then it must
1869  *    call blk_init_queue().  The function @rfn will be called when there
1870  *    are requests on the queue that need to be processed.  If the device
1871  *    supports plugging, then @rfn may not be called immediately when requests
1872  *    are available on the queue, but may be called at some time later instead.
1873  *    Plugged queues are generally unplugged when a buffer belonging to one
1874  *    of the requests on the queue is needed, or due to memory pressure.
1875  *
1876  *    @rfn is not required, or even expected, to remove all requests off the
1877  *    queue, but only as many as it can handle at a time.  If it does leave
1878  *    requests on the queue, it is responsible for arranging that the requests
1879  *    get dealt with eventually.
1880  *
1881  *    The queue spin lock must be held while manipulating the requests on the
1882  *    request queue; this lock will be taken also from interrupt context, so irq
1883  *    disabling is needed for it.
1884  *
1885  *    Function returns a pointer to the initialized request queue, or NULL if
1886  *    it didn't succeed.
1887  *
1888  * Note:
1889  *    blk_init_queue() must be paired with a blk_cleanup_queue() call
1890  *    when the block device is deactivated (such as at module unload).
1891  **/
1892
1893 request_queue_t *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1894 {
1895         return blk_init_queue_node(rfn, lock, -1);
1896 }
1897 EXPORT_SYMBOL(blk_init_queue);
1898
1899 request_queue_t *
1900 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1901 {
1902         request_queue_t *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1903
1904         if (!q)
1905                 return NULL;
1906
1907         q->node = node_id;
1908         if (blk_init_free_list(q)) {
1909                 kmem_cache_free(requestq_cachep, q);
1910                 return NULL;
1911         }
1912
1913         /*
1914          * if caller didn't supply a lock, they get per-queue locking with
1915          * our embedded lock
1916          */
1917         if (!lock) {
1918                 spin_lock_init(&q->__queue_lock);
1919                 lock = &q->__queue_lock;
1920         }
1921
1922         q->request_fn           = rfn;
1923         q->back_merge_fn        = ll_back_merge_fn;
1924         q->front_merge_fn       = ll_front_merge_fn;
1925         q->merge_requests_fn    = ll_merge_requests_fn;
1926         q->prep_rq_fn           = NULL;
1927         q->unplug_fn            = generic_unplug_device;
1928         q->queue_flags          = (1 << QUEUE_FLAG_CLUSTER);
1929         q->queue_lock           = lock;
1930
1931         blk_queue_segment_boundary(q, 0xffffffff);
1932
1933         blk_queue_make_request(q, __make_request);
1934         blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1935
1936         blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1937         blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1938
1939         /*
1940          * all done
1941          */
1942         if (!elevator_init(q, NULL)) {
1943                 blk_queue_congestion_threshold(q);
1944                 return q;
1945         }
1946
1947         blk_put_queue(q);
1948         return NULL;
1949 }
1950 EXPORT_SYMBOL(blk_init_queue_node);
1951
1952 int blk_get_queue(request_queue_t *q)
1953 {
1954         if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1955                 kobject_get(&q->kobj);
1956                 return 0;
1957         }
1958
1959         return 1;
1960 }
1961
1962 EXPORT_SYMBOL(blk_get_queue);
1963
1964 static inline void blk_free_request(request_queue_t *q, struct request *rq)
1965 {
1966         if (rq->flags & REQ_ELVPRIV)
1967                 elv_put_request(q, rq);
1968         mempool_free(rq, q->rq.rq_pool);
1969 }
1970
1971 static inline struct request *
1972 blk_alloc_request(request_queue_t *q, int rw, struct bio *bio,
1973                   int priv, gfp_t gfp_mask)
1974 {
1975         struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1976
1977         if (!rq)
1978                 return NULL;
1979
1980         /*
1981          * first three bits are identical in rq->flags and bio->bi_rw,
1982          * see bio.h and blkdev.h
1983          */
1984         rq->flags = rw;
1985
1986         if (priv) {
1987                 if (unlikely(elv_set_request(q, rq, bio, gfp_mask))) {
1988                         mempool_free(rq, q->rq.rq_pool);
1989                         return NULL;
1990                 }
1991                 rq->flags |= REQ_ELVPRIV;
1992         }
1993
1994         return rq;
1995 }
1996
1997 /*
1998  * ioc_batching returns true if the ioc is a valid batching request and
1999  * should be given priority access to a request.
2000  */
2001 static inline int ioc_batching(request_queue_t *q, struct io_context *ioc)
2002 {
2003         if (!ioc)
2004                 return 0;
2005
2006         /*
2007          * Make sure the process is able to allocate at least 1 request
2008          * even if the batch times out, otherwise we could theoretically
2009          * lose wakeups.
2010          */
2011         return ioc->nr_batch_requests == q->nr_batching ||
2012                 (ioc->nr_batch_requests > 0
2013                 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
2014 }
2015
2016 /*
2017  * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
2018  * will cause the process to be a "batcher" on all queues in the system. This
2019  * is the behaviour we want though - once it gets a wakeup it should be given
2020  * a nice run.
2021  */
2022 static void ioc_set_batching(request_queue_t *q, struct io_context *ioc)
2023 {
2024         if (!ioc || ioc_batching(q, ioc))
2025                 return;
2026
2027         ioc->nr_batch_requests = q->nr_batching;
2028         ioc->last_waited = jiffies;
2029 }
2030
2031 static void __freed_request(request_queue_t *q, int rw)
2032 {
2033         struct request_list *rl = &q->rq;
2034
2035         if (rl->count[rw] < queue_congestion_off_threshold(q))
2036                 clear_queue_congested(q, rw);
2037
2038         if (rl->count[rw] + 1 <= q->nr_requests) {
2039                 if (waitqueue_active(&rl->wait[rw]))
2040                         wake_up(&rl->wait[rw]);
2041
2042                 blk_clear_queue_full(q, rw);
2043         }
2044 }
2045
2046 /*
2047  * A request has just been released.  Account for it, update the full and
2048  * congestion status, wake up any waiters.   Called under q->queue_lock.
2049  */
2050 static void freed_request(request_queue_t *q, int rw, int priv)
2051 {
2052         struct request_list *rl = &q->rq;
2053
2054         rl->count[rw]--;
2055         if (priv)
2056                 rl->elvpriv--;
2057
2058         __freed_request(q, rw);
2059
2060         if (unlikely(rl->starved[rw ^ 1]))
2061                 __freed_request(q, rw ^ 1);
2062 }
2063
2064 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2065 /*
2066  * Get a free request, queue_lock must be held.
2067  * Returns NULL on failure, with queue_lock held.
2068  * Returns !NULL on success, with queue_lock *not held*.
2069  */
2070 static struct request *get_request(request_queue_t *q, int rw, struct bio *bio,
2071                                    gfp_t gfp_mask)
2072 {
2073         struct request *rq = NULL;
2074         struct request_list *rl = &q->rq;
2075         struct io_context *ioc = NULL;
2076         int may_queue, priv;
2077
2078         may_queue = elv_may_queue(q, rw, bio);
2079         if (may_queue == ELV_MQUEUE_NO)
2080                 goto rq_starved;
2081
2082         if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
2083                 if (rl->count[rw]+1 >= q->nr_requests) {
2084                         ioc = current_io_context(GFP_ATOMIC);
2085                         /*
2086                          * The queue will fill after this allocation, so set
2087                          * it as full, and mark this process as "batching".
2088                          * This process will be allowed to complete a batch of
2089                          * requests, others will be blocked.
2090                          */
2091                         if (!blk_queue_full(q, rw)) {
2092                                 ioc_set_batching(q, ioc);
2093                                 blk_set_queue_full(q, rw);
2094                         } else {
2095                                 if (may_queue != ELV_MQUEUE_MUST
2096                                                 && !ioc_batching(q, ioc)) {
2097                                         /*
2098                                          * The queue is full and the allocating
2099                                          * process is not a "batcher", and not
2100                                          * exempted by the IO scheduler
2101                                          */
2102                                         goto out;
2103                                 }
2104                         }
2105                 }
2106                 set_queue_congested(q, rw);
2107         }
2108
2109         /*
2110          * Only allow batching queuers to allocate up to 50% over the defined
2111          * limit of requests, otherwise we could have thousands of requests
2112          * allocated with any setting of ->nr_requests
2113          */
2114         if (rl->count[rw] >= (3 * q->nr_requests / 2))
2115                 goto out;
2116
2117         rl->count[rw]++;
2118         rl->starved[rw] = 0;
2119
2120         priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
2121         if (priv)
2122                 rl->elvpriv++;
2123
2124         spin_unlock_irq(q->queue_lock);
2125
2126         rq = blk_alloc_request(q, rw, bio, priv, gfp_mask);
2127         if (unlikely(!rq)) {
2128                 /*
2129                  * Allocation failed presumably due to memory. Undo anything
2130                  * we might have messed up.
2131                  *
2132                  * Allocating task should really be put onto the front of the
2133                  * wait queue, but this is pretty rare.
2134                  */
2135                 spin_lock_irq(q->queue_lock);
2136                 freed_request(q, rw, priv);
2137
2138                 /*
2139                  * in the very unlikely event that allocation failed and no
2140                  * requests for this direction was pending, mark us starved
2141                  * so that freeing of a request in the other direction will
2142                  * notice us. another possible fix would be to split the
2143                  * rq mempool into READ and WRITE
2144                  */
2145 rq_starved:
2146                 if (unlikely(rl->count[rw] == 0))
2147                         rl->starved[rw] = 1;
2148
2149                 goto out;
2150         }
2151
2152         /*
2153          * ioc may be NULL here, and ioc_batching will be false. That's
2154          * OK, if the queue is under the request limit then requests need
2155          * not count toward the nr_batch_requests limit. There will always
2156          * be some limit enforced by BLK_BATCH_TIME.
2157          */
2158         if (ioc_batching(q, ioc))
2159                 ioc->nr_batch_requests--;
2160         
2161         rq_init(q, rq);
2162         rq->rl = rl;
2163
2164         blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
2165 out:
2166         return rq;
2167 }
2168
2169 /*
2170  * No available requests for this queue, unplug the device and wait for some
2171  * requests to become available.
2172  *
2173  * Called with q->queue_lock held, and returns with it unlocked.
2174  */
2175 static struct request *get_request_wait(request_queue_t *q, int rw,
2176                                         struct bio *bio)
2177 {
2178         struct request *rq;
2179
2180         rq = get_request(q, rw, bio, GFP_NOIO);
2181         while (!rq) {
2182                 DEFINE_WAIT(wait);
2183                 struct request_list *rl = &q->rq;
2184
2185                 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2186                                 TASK_UNINTERRUPTIBLE);
2187
2188                 rq = get_request(q, rw, bio, GFP_NOIO);
2189
2190                 if (!rq) {
2191                         struct io_context *ioc;
2192
2193                         blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
2194
2195                         __generic_unplug_device(q);
2196                         spin_unlock_irq(q->queue_lock);
2197                         io_schedule();
2198
2199                         /*
2200                          * After sleeping, we become a "batching" process and
2201                          * will be able to allocate at least one request, and
2202                          * up to a big batch of them for a small period time.
2203                          * See ioc_batching, ioc_set_batching
2204                          */
2205                         ioc = current_io_context(GFP_NOIO);
2206                         ioc_set_batching(q, ioc);
2207
2208                         spin_lock_irq(q->queue_lock);
2209                 }
2210                 finish_wait(&rl->wait[rw], &wait);
2211         }
2212
2213         return rq;
2214 }
2215
2216 struct request *blk_get_request(request_queue_t *q, int rw, gfp_t gfp_mask)
2217 {
2218         struct request *rq;
2219
2220         BUG_ON(rw != READ && rw != WRITE);
2221
2222         spin_lock_irq(q->queue_lock);
2223         if (gfp_mask & __GFP_WAIT) {
2224                 rq = get_request_wait(q, rw, NULL);
2225         } else {
2226                 rq = get_request(q, rw, NULL, gfp_mask);
2227                 if (!rq)
2228                         spin_unlock_irq(q->queue_lock);
2229         }
2230         /* q->queue_lock is unlocked at this point */
2231
2232         return rq;
2233 }
2234 EXPORT_SYMBOL(blk_get_request);
2235
2236 /**
2237  * blk_requeue_request - put a request back on queue
2238  * @q:          request queue where request should be inserted
2239  * @rq:         request to be inserted
2240  *
2241  * Description:
2242  *    Drivers often keep queueing requests until the hardware cannot accept
2243  *    more, when that condition happens we need to put the request back
2244  *    on the queue. Must be called with queue lock held.
2245  */
2246 void blk_requeue_request(request_queue_t *q, struct request *rq)
2247 {
2248         blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
2249
2250         if (blk_rq_tagged(rq))
2251                 blk_queue_end_tag(q, rq);
2252
2253         elv_requeue_request(q, rq);
2254 }
2255
2256 EXPORT_SYMBOL(blk_requeue_request);
2257
2258 /**
2259  * blk_insert_request - insert a special request in to a request queue
2260  * @q:          request queue where request should be inserted
2261  * @rq:         request to be inserted
2262  * @at_head:    insert request at head or tail of queue
2263  * @data:       private data
2264  *
2265  * Description:
2266  *    Many block devices need to execute commands asynchronously, so they don't
2267  *    block the whole kernel from preemption during request execution.  This is
2268  *    accomplished normally by inserting aritficial requests tagged as
2269  *    REQ_SPECIAL in to the corresponding request queue, and letting them be
2270  *    scheduled for actual execution by the request queue.
2271  *
2272  *    We have the option of inserting the head or the tail of the queue.
2273  *    Typically we use the tail for new ioctls and so forth.  We use the head
2274  *    of the queue for things like a QUEUE_FULL message from a device, or a
2275  *    host that is unable to accept a particular command.
2276  */
2277 void blk_insert_request(request_queue_t *q, struct request *rq,
2278                         int at_head, void *data)
2279 {
2280         int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2281         unsigned long flags;
2282
2283         /*
2284          * tell I/O scheduler that this isn't a regular read/write (ie it
2285          * must not attempt merges on this) and that it acts as a soft
2286          * barrier
2287          */
2288         rq->flags |= REQ_SPECIAL | REQ_SOFTBARRIER;
2289
2290         rq->special = data;
2291
2292         spin_lock_irqsave(q->queue_lock, flags);
2293
2294         /*
2295          * If command is tagged, release the tag
2296          */
2297         if (blk_rq_tagged(rq))
2298                 blk_queue_end_tag(q, rq);
2299
2300         drive_stat_acct(rq, rq->nr_sectors, 1);
2301         __elv_add_request(q, rq, where, 0);
2302
2303         if (blk_queue_plugged(q))
2304                 __generic_unplug_device(q);
2305         else
2306                 q->request_fn(q);
2307         spin_unlock_irqrestore(q->queue_lock, flags);
2308 }
2309
2310 EXPORT_SYMBOL(blk_insert_request);
2311
2312 /**
2313  * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2314  * @q:          request queue where request should be inserted
2315  * @rq:         request structure to fill
2316  * @ubuf:       the user buffer
2317  * @len:        length of user data
2318  *
2319  * Description:
2320  *    Data will be mapped directly for zero copy io, if possible. Otherwise
2321  *    a kernel bounce buffer is used.
2322  *
2323  *    A matching blk_rq_unmap_user() must be issued at the end of io, while
2324  *    still in process context.
2325  *
2326  *    Note: The mapped bio may need to be bounced through blk_queue_bounce()
2327  *    before being submitted to the device, as pages mapped may be out of
2328  *    reach. It's the callers responsibility to make sure this happens. The
2329  *    original bio must be passed back in to blk_rq_unmap_user() for proper
2330  *    unmapping.
2331  */
2332 int blk_rq_map_user(request_queue_t *q, struct request *rq, void __user *ubuf,
2333                     unsigned int len)
2334 {
2335         unsigned long uaddr;
2336         struct bio *bio;
2337         int reading;
2338
2339         if (len > (q->max_hw_sectors << 9))
2340                 return -EINVAL;
2341         if (!len || !ubuf)
2342                 return -EINVAL;
2343
2344         reading = rq_data_dir(rq) == READ;
2345
2346         /*
2347          * if alignment requirement is satisfied, map in user pages for
2348          * direct dma. else, set up kernel bounce buffers
2349          */
2350         uaddr = (unsigned long) ubuf;
2351         if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2352                 bio = bio_map_user(q, NULL, uaddr, len, reading);
2353         else
2354                 bio = bio_copy_user(q, uaddr, len, reading);
2355
2356         if (!IS_ERR(bio)) {
2357                 rq->bio = rq->biotail = bio;
2358                 blk_rq_bio_prep(q, rq, bio);
2359
2360                 rq->buffer = rq->data = NULL;
2361                 rq->data_len = len;
2362                 return 0;
2363         }
2364
2365         /*
2366          * bio is the err-ptr
2367          */
2368         return PTR_ERR(bio);
2369 }
2370
2371 EXPORT_SYMBOL(blk_rq_map_user);
2372
2373 /**
2374  * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2375  * @q:          request queue where request should be inserted
2376  * @rq:         request to map data to
2377  * @iov:        pointer to the iovec
2378  * @iov_count:  number of elements in the iovec
2379  *
2380  * Description:
2381  *    Data will be mapped directly for zero copy io, if possible. Otherwise
2382  *    a kernel bounce buffer is used.
2383  *
2384  *    A matching blk_rq_unmap_user() must be issued at the end of io, while
2385  *    still in process context.
2386  *
2387  *    Note: The mapped bio may need to be bounced through blk_queue_bounce()
2388  *    before being submitted to the device, as pages mapped may be out of
2389  *    reach. It's the callers responsibility to make sure this happens. The
2390  *    original bio must be passed back in to blk_rq_unmap_user() for proper
2391  *    unmapping.
2392  */
2393 int blk_rq_map_user_iov(request_queue_t *q, struct request *rq,
2394                         struct sg_iovec *iov, int iov_count)
2395 {
2396         struct bio *bio;
2397
2398         if (!iov || iov_count <= 0)
2399                 return -EINVAL;
2400
2401         /* we don't allow misaligned data like bio_map_user() does.  If the
2402          * user is using sg, they're expected to know the alignment constraints
2403          * and respect them accordingly */
2404         bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2405         if (IS_ERR(bio))
2406                 return PTR_ERR(bio);
2407
2408         rq->bio = rq->biotail = bio;
2409         blk_rq_bio_prep(q, rq, bio);
2410         rq->buffer = rq->data = NULL;
2411         rq->data_len = bio->bi_size;
2412         return 0;
2413 }
2414
2415 EXPORT_SYMBOL(blk_rq_map_user_iov);
2416
2417 /**
2418  * blk_rq_unmap_user - unmap a request with user data
2419  * @bio:        bio to be unmapped
2420  * @ulen:       length of user buffer
2421  *
2422  * Description:
2423  *    Unmap a bio previously mapped by blk_rq_map_user().
2424  */
2425 int blk_rq_unmap_user(struct bio *bio, unsigned int ulen)
2426 {
2427         int ret = 0;
2428
2429         if (bio) {
2430                 if (bio_flagged(bio, BIO_USER_MAPPED))
2431                         bio_unmap_user(bio);
2432                 else
2433                         ret = bio_uncopy_user(bio);
2434         }
2435
2436         return 0;
2437 }
2438
2439 EXPORT_SYMBOL(blk_rq_unmap_user);
2440
2441 /**
2442  * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2443  * @q:          request queue where request should be inserted
2444  * @rq:         request to fill
2445  * @kbuf:       the kernel buffer
2446  * @len:        length of user data
2447  * @gfp_mask:   memory allocation flags
2448  */
2449 int blk_rq_map_kern(request_queue_t *q, struct request *rq, void *kbuf,
2450                     unsigned int len, gfp_t gfp_mask)
2451 {
2452         struct bio *bio;
2453
2454         if (len > (q->max_hw_sectors << 9))
2455                 return -EINVAL;
2456         if (!len || !kbuf)
2457                 return -EINVAL;
2458
2459         bio = bio_map_kern(q, kbuf, len, gfp_mask);
2460         if (IS_ERR(bio))
2461                 return PTR_ERR(bio);
2462
2463         if (rq_data_dir(rq) == WRITE)
2464                 bio->bi_rw |= (1 << BIO_RW);
2465
2466         rq->bio = rq->biotail = bio;
2467         blk_rq_bio_prep(q, rq, bio);
2468
2469         rq->buffer = rq->data = NULL;
2470         rq->data_len = len;
2471         return 0;
2472 }
2473
2474 EXPORT_SYMBOL(blk_rq_map_kern);
2475
2476 /**
2477  * blk_execute_rq_nowait - insert a request into queue for execution
2478  * @q:          queue to insert the request in
2479  * @bd_disk:    matching gendisk
2480  * @rq:         request to insert
2481  * @at_head:    insert request at head or tail of queue
2482  * @done:       I/O completion handler
2483  *
2484  * Description:
2485  *    Insert a fully prepared request at the back of the io scheduler queue
2486  *    for execution.  Don't wait for completion.
2487  */
2488 void blk_execute_rq_nowait(request_queue_t *q, struct gendisk *bd_disk,
2489                            struct request *rq, int at_head,
2490                            rq_end_io_fn *done)
2491 {
2492         int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2493
2494         rq->rq_disk = bd_disk;
2495         rq->flags |= REQ_NOMERGE;
2496         rq->end_io = done;
2497         WARN_ON(irqs_disabled());
2498         spin_lock_irq(q->queue_lock);
2499         __elv_add_request(q, rq, where, 1);
2500         __generic_unplug_device(q);
2501         spin_unlock_irq(q->queue_lock);
2502 }
2503 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
2504
2505 /**
2506  * blk_execute_rq - insert a request into queue for execution
2507  * @q:          queue to insert the request in
2508  * @bd_disk:    matching gendisk
2509  * @rq:         request to insert
2510  * @at_head:    insert request at head or tail of queue
2511  *
2512  * Description:
2513  *    Insert a fully prepared request at the back of the io scheduler queue
2514  *    for execution and wait for completion.
2515  */
2516 int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
2517                    struct request *rq, int at_head)
2518 {
2519         DECLARE_COMPLETION_ONSTACK(wait);
2520         char sense[SCSI_SENSE_BUFFERSIZE];
2521         int err = 0;
2522
2523         /*
2524          * we need an extra reference to the request, so we can look at
2525          * it after io completion
2526          */
2527         rq->ref_count++;
2528
2529         if (!rq->sense) {
2530                 memset(sense, 0, sizeof(sense));
2531                 rq->sense = sense;
2532                 rq->sense_len = 0;
2533         }
2534
2535         rq->waiting = &wait;
2536         blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
2537         wait_for_completion(&wait);
2538         rq->waiting = NULL;
2539
2540         if (rq->errors)
2541                 err = -EIO;
2542
2543         return err;
2544 }
2545
2546 EXPORT_SYMBOL(blk_execute_rq);
2547
2548 /**
2549  * blkdev_issue_flush - queue a flush
2550  * @bdev:       blockdev to issue flush for
2551  * @error_sector:       error sector
2552  *
2553  * Description:
2554  *    Issue a flush for the block device in question. Caller can supply
2555  *    room for storing the error offset in case of a flush error, if they
2556  *    wish to.  Caller must run wait_for_completion() on its own.
2557  */
2558 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2559 {
2560         request_queue_t *q;
2561
2562         if (bdev->bd_disk == NULL)
2563                 return -ENXIO;
2564
2565         q = bdev_get_queue(bdev);
2566         if (!q)
2567                 return -ENXIO;
2568         if (!q->issue_flush_fn)
2569                 return -EOPNOTSUPP;
2570
2571         return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2572 }
2573
2574 EXPORT_SYMBOL(blkdev_issue_flush);
2575
2576 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2577 {
2578         int rw = rq_data_dir(rq);
2579
2580         if (!blk_fs_request(rq) || !rq->rq_disk)
2581                 return;
2582
2583         if (!new_io) {
2584                 __disk_stat_inc(rq->rq_disk, merges[rw]);
2585         } else {
2586                 disk_round_stats(rq->rq_disk);
2587                 rq->rq_disk->in_flight++;
2588         }
2589 }
2590
2591 /*
2592  * add-request adds a request to the linked list.
2593  * queue lock is held and interrupts disabled, as we muck with the
2594  * request queue list.
2595  */
2596 static inline void add_request(request_queue_t * q, struct request * req)
2597 {
2598         drive_stat_acct(req, req->nr_sectors, 1);
2599
2600         if (q->activity_fn)
2601                 q->activity_fn(q->activity_data, rq_data_dir(req));
2602
2603         /*
2604          * elevator indicated where it wants this request to be
2605          * inserted at elevator_merge time
2606          */
2607         __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2608 }
2609  
2610 /*
2611  * disk_round_stats()   - Round off the performance stats on a struct
2612  * disk_stats.
2613  *
2614  * The average IO queue length and utilisation statistics are maintained
2615  * by observing the current state of the queue length and the amount of
2616  * time it has been in this state for.
2617  *
2618  * Normally, that accounting is done on IO completion, but that can result
2619  * in more than a second's worth of IO being accounted for within any one
2620  * second, leading to >100% utilisation.  To deal with that, we call this
2621  * function to do a round-off before returning the results when reading
2622  * /proc/diskstats.  This accounts immediately for all queue usage up to
2623  * the current jiffies and restarts the counters again.
2624  */
2625 void disk_round_stats(struct gendisk *disk)
2626 {
2627         unsigned long now = jiffies;
2628
2629         if (now == disk->stamp)
2630                 return;
2631
2632         if (disk->in_flight) {
2633                 __disk_stat_add(disk, time_in_queue,
2634                                 disk->in_flight * (now - disk->stamp));
2635                 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2636         }
2637         disk->stamp = now;
2638 }
2639
2640 EXPORT_SYMBOL_GPL(disk_round_stats);
2641
2642 /*
2643  * queue lock must be held
2644  */
2645 void __blk_put_request(request_queue_t *q, struct request *req)
2646 {
2647         struct request_list *rl = req->rl;
2648
2649         if (unlikely(!q))
2650                 return;
2651         if (unlikely(--req->ref_count))
2652                 return;
2653
2654         elv_completed_request(q, req);
2655
2656         req->rq_status = RQ_INACTIVE;
2657         req->rl = NULL;
2658
2659         /*
2660          * Request may not have originated from ll_rw_blk. if not,
2661          * it didn't come out of our reserved rq pools
2662          */
2663         if (rl) {
2664                 int rw = rq_data_dir(req);
2665                 int priv = req->flags & REQ_ELVPRIV;
2666
2667                 BUG_ON(!list_empty(&req->queuelist));
2668
2669                 blk_free_request(q, req);
2670                 freed_request(q, rw, priv);
2671         }
2672 }
2673
2674 EXPORT_SYMBOL_GPL(__blk_put_request);
2675
2676 void blk_put_request(struct request *req)
2677 {
2678         unsigned long flags;
2679         request_queue_t *q = req->q;
2680
2681         /*
2682          * Gee, IDE calls in w/ NULL q.  Fix IDE and remove the
2683          * following if (q) test.
2684          */
2685         if (q) {
2686                 spin_lock_irqsave(q->queue_lock, flags);
2687                 __blk_put_request(q, req);
2688                 spin_unlock_irqrestore(q->queue_lock, flags);
2689         }
2690 }
2691
2692 EXPORT_SYMBOL(blk_put_request);
2693
2694 /**
2695  * blk_end_sync_rq - executes a completion event on a request
2696  * @rq: request to complete
2697  * @error: end io status of the request
2698  */
2699 void blk_end_sync_rq(struct request *rq, int error)
2700 {
2701         struct completion *waiting = rq->waiting;
2702
2703         rq->waiting = NULL;
2704         __blk_put_request(rq->q, rq);
2705
2706         /*
2707          * complete last, if this is a stack request the process (and thus
2708          * the rq pointer) could be invalid right after this complete()
2709          */
2710         complete(waiting);
2711 }
2712 EXPORT_SYMBOL(blk_end_sync_rq);
2713
2714 /**
2715  * blk_congestion_wait - wait for a queue to become uncongested
2716  * @rw: READ or WRITE
2717  * @timeout: timeout in jiffies
2718  *
2719  * Waits for up to @timeout jiffies for a queue (any queue) to exit congestion.
2720  * If no queues are congested then just wait for the next request to be
2721  * returned.
2722  */
2723 long blk_congestion_wait(int rw, long timeout)
2724 {
2725         long ret;
2726         DEFINE_WAIT(wait);
2727         wait_queue_head_t *wqh = &congestion_wqh[rw];
2728
2729         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
2730         ret = io_schedule_timeout(timeout);
2731         finish_wait(wqh, &wait);
2732         return ret;
2733 }
2734
2735 EXPORT_SYMBOL(blk_congestion_wait);
2736
2737 /*
2738  * Has to be called with the request spinlock acquired
2739  */
2740 static int attempt_merge(request_queue_t *q, struct request *req,
2741                           struct request *next)
2742 {
2743         if (!rq_mergeable(req) || !rq_mergeable(next))
2744                 return 0;
2745
2746         /*
2747          * not contiguous
2748          */
2749         if (req->sector + req->nr_sectors != next->sector)
2750                 return 0;
2751
2752         if (rq_data_dir(req) != rq_data_dir(next)
2753             || req->rq_disk != next->rq_disk
2754             || next->waiting || next->special)
2755                 return 0;
2756
2757         /*
2758          * If we are allowed to merge, then append bio list
2759          * from next to rq and release next. merge_requests_fn
2760          * will have updated segment counts, update sector
2761          * counts here.
2762          */
2763         if (!q->merge_requests_fn(q, req, next))
2764                 return 0;
2765
2766         /*
2767          * At this point we have either done a back merge
2768          * or front merge. We need the smaller start_time of
2769          * the merged requests to be the current request
2770          * for accounting purposes.
2771          */
2772         if (time_after(req->start_time, next->start_time))
2773                 req->start_time = next->start_time;
2774
2775         req->biotail->bi_next = next->bio;
2776         req->biotail = next->biotail;
2777
2778         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2779
2780         elv_merge_requests(q, req, next);
2781
2782         if (req->rq_disk) {
2783                 disk_round_stats(req->rq_disk);
2784                 req->rq_disk->in_flight--;
2785         }
2786
2787         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2788
2789         __blk_put_request(q, next);
2790         return 1;
2791 }
2792
2793 static inline int attempt_back_merge(request_queue_t *q, struct request *rq)
2794 {
2795         struct request *next = elv_latter_request(q, rq);
2796
2797         if (next)
2798                 return attempt_merge(q, rq, next);
2799
2800         return 0;
2801 }
2802
2803 static inline int attempt_front_merge(request_queue_t *q, struct request *rq)
2804 {
2805         struct request *prev = elv_former_request(q, rq);
2806
2807         if (prev)
2808                 return attempt_merge(q, prev, rq);
2809
2810         return 0;
2811 }
2812
2813 static void init_request_from_bio(struct request *req, struct bio *bio)
2814 {
2815         req->flags |= REQ_CMD;
2816
2817         /*
2818          * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2819          */
2820         if (bio_rw_ahead(bio) || bio_failfast(bio))
2821                 req->flags |= REQ_FAILFAST;
2822
2823         /*
2824          * REQ_BARRIER implies no merging, but lets make it explicit
2825          */
2826         if (unlikely(bio_barrier(bio)))
2827                 req->flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2828
2829         if (bio_sync(bio))
2830                 req->flags |= REQ_RW_SYNC;
2831
2832         req->errors = 0;
2833         req->hard_sector = req->sector = bio->bi_sector;
2834         req->hard_nr_sectors = req->nr_sectors = bio_sectors(bio);
2835         req->current_nr_sectors = req->hard_cur_sectors = bio_cur_sectors(bio);
2836         req->nr_phys_segments = bio_phys_segments(req->q, bio);
2837         req->nr_hw_segments = bio_hw_segments(req->q, bio);
2838         req->buffer = bio_data(bio);    /* see ->buffer comment above */
2839         req->waiting = NULL;
2840         req->bio = req->biotail = bio;
2841         req->ioprio = bio_prio(bio);
2842         req->rq_disk = bio->bi_bdev->bd_disk;
2843         req->start_time = jiffies;
2844 }
2845
2846 static int __make_request(request_queue_t *q, struct bio *bio)
2847 {
2848         struct request *req;
2849         int el_ret, rw, nr_sectors, cur_nr_sectors, barrier, err, sync;
2850         unsigned short prio;
2851         sector_t sector;
2852
2853         sector = bio->bi_sector;
2854         nr_sectors = bio_sectors(bio);
2855         cur_nr_sectors = bio_cur_sectors(bio);
2856         prio = bio_prio(bio);
2857
2858         rw = bio_data_dir(bio);
2859         sync = bio_sync(bio);
2860
2861         /*
2862          * low level driver can indicate that it wants pages above a
2863          * certain limit bounced to low memory (ie for highmem, or even
2864          * ISA dma in theory)
2865          */
2866         blk_queue_bounce(q, &bio);
2867
2868         spin_lock_prefetch(q->queue_lock);
2869
2870         barrier = bio_barrier(bio);
2871         if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
2872                 err = -EOPNOTSUPP;
2873                 goto end_io;
2874         }
2875
2876         spin_lock_irq(q->queue_lock);
2877
2878         if (unlikely(barrier) || elv_queue_empty(q))
2879                 goto get_rq;
2880
2881         el_ret = elv_merge(q, &req, bio);
2882         switch (el_ret) {
2883                 case ELEVATOR_BACK_MERGE:
2884                         BUG_ON(!rq_mergeable(req));
2885
2886                         if (!q->back_merge_fn(q, req, bio))
2887                                 break;
2888
2889                         blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
2890
2891                         req->biotail->bi_next = bio;
2892                         req->biotail = bio;
2893                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2894                         req->ioprio = ioprio_best(req->ioprio, prio);
2895                         drive_stat_acct(req, nr_sectors, 0);
2896                         if (!attempt_back_merge(q, req))
2897                                 elv_merged_request(q, req);
2898                         goto out;
2899
2900                 case ELEVATOR_FRONT_MERGE:
2901                         BUG_ON(!rq_mergeable(req));
2902
2903                         if (!q->front_merge_fn(q, req, bio))
2904                                 break;
2905
2906                         blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
2907
2908                         bio->bi_next = req->bio;
2909                         req->bio = bio;
2910
2911                         /*
2912                          * may not be valid. if the low level driver said
2913                          * it didn't need a bounce buffer then it better
2914                          * not touch req->buffer either...
2915                          */
2916                         req->buffer = bio_data(bio);
2917                         req->current_nr_sectors = cur_nr_sectors;
2918                         req->hard_cur_sectors = cur_nr_sectors;
2919                         req->sector = req->hard_sector = sector;
2920                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2921                         req->ioprio = ioprio_best(req->ioprio, prio);
2922                         drive_stat_acct(req, nr_sectors, 0);
2923                         if (!attempt_front_merge(q, req))
2924                                 elv_merged_request(q, req);
2925                         goto out;
2926
2927                 /* ELV_NO_MERGE: elevator says don't/can't merge. */
2928                 default:
2929                         ;
2930         }
2931
2932 get_rq:
2933         /*
2934          * Grab a free request. This is might sleep but can not fail.
2935          * Returns with the queue unlocked.
2936          */
2937         req = get_request_wait(q, rw, bio);
2938
2939         /*
2940          * After dropping the lock and possibly sleeping here, our request
2941          * may now be mergeable after it had proven unmergeable (above).
2942          * We don't worry about that case for efficiency. It won't happen
2943          * often, and the elevators are able to handle it.
2944          */
2945         init_request_from_bio(req, bio);
2946
2947         spin_lock_irq(q->queue_lock);
2948         if (elv_queue_empty(q))
2949                 blk_plug_device(q);
2950         add_request(q, req);
2951 out:
2952         if (sync)
2953                 __generic_unplug_device(q);
2954
2955         spin_unlock_irq(q->queue_lock);
2956         return 0;
2957
2958 end_io:
2959         bio_endio(bio, nr_sectors << 9, err);
2960         return 0;
2961 }
2962
2963 /*
2964  * If bio->bi_dev is a partition, remap the location
2965  */
2966 static inline void blk_partition_remap(struct bio *bio)
2967 {
2968         struct block_device *bdev = bio->bi_bdev;
2969
2970         if (bdev != bdev->bd_contains) {
2971                 struct hd_struct *p = bdev->bd_part;
2972                 const int rw = bio_data_dir(bio);
2973
2974                 p->sectors[rw] += bio_sectors(bio);
2975                 p->ios[rw]++;
2976
2977                 bio->bi_sector += p->start_sect;
2978                 bio->bi_bdev = bdev->bd_contains;
2979         }
2980 }
2981
2982 static void handle_bad_sector(struct bio *bio)
2983 {
2984         char b[BDEVNAME_SIZE];
2985
2986         printk(KERN_INFO "attempt to access beyond end of device\n");
2987         printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
2988                         bdevname(bio->bi_bdev, b),
2989                         bio->bi_rw,
2990                         (unsigned long long)bio->bi_sector + bio_sectors(bio),
2991                         (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
2992
2993         set_bit(BIO_EOF, &bio->bi_flags);
2994 }
2995
2996 /**
2997  * generic_make_request: hand a buffer to its device driver for I/O
2998  * @bio:  The bio describing the location in memory and on the device.
2999  *
3000  * generic_make_request() is used to make I/O requests of block
3001  * devices. It is passed a &struct bio, which describes the I/O that needs
3002  * to be done.
3003  *
3004  * generic_make_request() does not return any status.  The
3005  * success/failure status of the request, along with notification of
3006  * completion, is delivered asynchronously through the bio->bi_end_io
3007  * function described (one day) else where.
3008  *
3009  * The caller of generic_make_request must make sure that bi_io_vec
3010  * are set to describe the memory buffer, and that bi_dev and bi_sector are
3011  * set to describe the device address, and the
3012  * bi_end_io and optionally bi_private are set to describe how
3013  * completion notification should be signaled.
3014  *
3015  * generic_make_request and the drivers it calls may use bi_next if this
3016  * bio happens to be merged with someone else, and may change bi_dev and
3017  * bi_sector for remaps as it sees fit.  So the values of these fields
3018  * should NOT be depended on after the call to generic_make_request.
3019  */
3020 void generic_make_request(struct bio *bio)
3021 {
3022         request_queue_t *q;
3023         sector_t maxsector;
3024         int ret, nr_sectors = bio_sectors(bio);
3025         dev_t old_dev;
3026
3027         might_sleep();
3028         /* Test device or partition size, when known. */
3029         maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3030         if (maxsector) {
3031                 sector_t sector = bio->bi_sector;
3032
3033                 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
3034                         /*
3035                          * This may well happen - the kernel calls bread()
3036                          * without checking the size of the device, e.g., when
3037                          * mounting a device.
3038                          */
3039                         handle_bad_sector(bio);
3040                         goto end_io;
3041                 }
3042         }
3043
3044         /*
3045          * Resolve the mapping until finished. (drivers are
3046          * still free to implement/resolve their own stacking
3047          * by explicitly returning 0)
3048          *
3049          * NOTE: we don't repeat the blk_size check for each new device.
3050          * Stacking drivers are expected to know what they are doing.
3051          */
3052         maxsector = -1;
3053         old_dev = 0;
3054         do {
3055                 char b[BDEVNAME_SIZE];
3056
3057                 q = bdev_get_queue(bio->bi_bdev);
3058                 if (!q) {
3059                         printk(KERN_ERR
3060                                "generic_make_request: Trying to access "
3061                                 "nonexistent block-device %s (%Lu)\n",
3062                                 bdevname(bio->bi_bdev, b),
3063                                 (long long) bio->bi_sector);
3064 end_io:
3065                         bio_endio(bio, bio->bi_size, -EIO);
3066                         break;
3067                 }
3068
3069                 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
3070                         printk("bio too big device %s (%u > %u)\n", 
3071                                 bdevname(bio->bi_bdev, b),
3072                                 bio_sectors(bio),
3073                                 q->max_hw_sectors);
3074                         goto end_io;
3075                 }
3076
3077                 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
3078                         goto end_io;
3079
3080                 /*
3081                  * If this device has partitions, remap block n
3082                  * of partition p to block n+start(p) of the disk.
3083                  */
3084                 blk_partition_remap(bio);
3085
3086                 if (maxsector != -1)
3087                         blk_add_trace_remap(q, bio, old_dev, bio->bi_sector, 
3088                                             maxsector);
3089
3090                 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
3091
3092                 maxsector = bio->bi_sector;
3093                 old_dev = bio->bi_bdev->bd_dev;
3094
3095                 ret = q->make_request_fn(q, bio);
3096         } while (ret);
3097 }
3098
3099 EXPORT_SYMBOL(generic_make_request);
3100
3101 /**
3102  * submit_bio: submit a bio to the block device layer for I/O
3103  * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3104  * @bio: The &struct bio which describes the I/O
3105  *
3106  * submit_bio() is very similar in purpose to generic_make_request(), and
3107  * uses that function to do most of the work. Both are fairly rough
3108  * interfaces, @bio must be presetup and ready for I/O.
3109  *
3110  */
3111 void submit_bio(int rw, struct bio *bio)
3112 {
3113         int count = bio_sectors(bio);
3114
3115         BIO_BUG_ON(!bio->bi_size);
3116         BIO_BUG_ON(!bio->bi_io_vec);
3117         bio->bi_rw |= rw;
3118         if (rw & WRITE)
3119                 count_vm_events(PGPGOUT, count);
3120         else
3121                 count_vm_events(PGPGIN, count);
3122
3123         if (unlikely(block_dump)) {
3124                 char b[BDEVNAME_SIZE];
3125                 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
3126                         current->comm, current->pid,
3127                         (rw & WRITE) ? "WRITE" : "READ",
3128                         (unsigned long long)bio->bi_sector,
3129                         bdevname(bio->bi_bdev,b));
3130         }
3131
3132         generic_make_request(bio);
3133 }
3134
3135 EXPORT_SYMBOL(submit_bio);
3136
3137 static void blk_recalc_rq_segments(struct request *rq)
3138 {
3139         struct bio *bio, *prevbio = NULL;
3140         int nr_phys_segs, nr_hw_segs;
3141         unsigned int phys_size, hw_size;
3142         request_queue_t *q = rq->q;
3143
3144         if (!rq->bio)
3145                 return;
3146
3147         phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
3148         rq_for_each_bio(bio, rq) {
3149                 /* Force bio hw/phys segs to be recalculated. */
3150                 bio->bi_flags &= ~(1 << BIO_SEG_VALID);
3151
3152                 nr_phys_segs += bio_phys_segments(q, bio);
3153                 nr_hw_segs += bio_hw_segments(q, bio);
3154                 if (prevbio) {
3155                         int pseg = phys_size + prevbio->bi_size + bio->bi_size;
3156                         int hseg = hw_size + prevbio->bi_size + bio->bi_size;
3157
3158                         if (blk_phys_contig_segment(q, prevbio, bio) &&
3159                             pseg <= q->max_segment_size) {
3160                                 nr_phys_segs--;
3161                                 phys_size += prevbio->bi_size + bio->bi_size;
3162                         } else
3163                                 phys_size = 0;
3164
3165                         if (blk_hw_contig_segment(q, prevbio, bio) &&
3166                             hseg <= q->max_segment_size) {
3167                                 nr_hw_segs--;
3168                                 hw_size += prevbio->bi_size + bio->bi_size;
3169                         } else
3170                                 hw_size = 0;
3171                 }
3172                 prevbio = bio;
3173         }
3174
3175         rq->nr_phys_segments = nr_phys_segs;
3176         rq->nr_hw_segments = nr_hw_segs;
3177 }
3178
3179 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
3180 {
3181         if (blk_fs_request(rq)) {
3182                 rq->hard_sector += nsect;
3183                 rq->hard_nr_sectors -= nsect;
3184
3185                 /*
3186                  * Move the I/O submission pointers ahead if required.
3187                  */
3188                 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3189                     (rq->sector <= rq->hard_sector)) {
3190                         rq->sector = rq->hard_sector;
3191                         rq->nr_sectors = rq->hard_nr_sectors;
3192                         rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3193                         rq->current_nr_sectors = rq->hard_cur_sectors;
3194                         rq->buffer = bio_data(rq->bio);
3195                 }
3196
3197                 /*
3198                  * if total number of sectors is less than the first segment
3199                  * size, something has gone terribly wrong
3200                  */
3201                 if (rq->nr_sectors < rq->current_nr_sectors) {
3202                         printk("blk: request botched\n");
3203                         rq->nr_sectors = rq->current_nr_sectors;
3204                 }
3205         }
3206 }
3207
3208 static int __end_that_request_first(struct request *req, int uptodate,
3209                                     int nr_bytes)
3210 {
3211         int total_bytes, bio_nbytes, error, next_idx = 0;
3212         struct bio *bio;
3213
3214         blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
3215
3216         /*
3217          * extend uptodate bool to allow < 0 value to be direct io error
3218          */
3219         error = 0;
3220         if (end_io_error(uptodate))
3221                 error = !uptodate ? -EIO : uptodate;
3222
3223         /*
3224          * for a REQ_BLOCK_PC request, we want to carry any eventual
3225          * sense key with us all the way through
3226          */
3227         if (!blk_pc_request(req))
3228                 req->errors = 0;
3229
3230         if (!uptodate) {
3231                 if (blk_fs_request(req) && !(req->flags & REQ_QUIET))
3232                         printk("end_request: I/O error, dev %s, sector %llu\n",
3233                                 req->rq_disk ? req->rq_disk->disk_name : "?",
3234                                 (unsigned long long)req->sector);
3235         }
3236
3237         if (blk_fs_request(req) && req->rq_disk) {
3238                 const int rw = rq_data_dir(req);
3239
3240                 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
3241         }
3242
3243         total_bytes = bio_nbytes = 0;
3244         while ((bio = req->bio) != NULL) {
3245                 int nbytes;
3246
3247                 if (nr_bytes >= bio->bi_size) {
3248                         req->bio = bio->bi_next;
3249                         nbytes = bio->bi_size;
3250                         if (!ordered_bio_endio(req, bio, nbytes, error))
3251                                 bio_endio(bio, nbytes, error);
3252                         next_idx = 0;
3253                         bio_nbytes = 0;
3254                 } else {
3255                         int idx = bio->bi_idx + next_idx;
3256
3257                         if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3258                                 blk_dump_rq_flags(req, "__end_that");
3259                                 printk("%s: bio idx %d >= vcnt %d\n",
3260                                                 __FUNCTION__,
3261                                                 bio->bi_idx, bio->bi_vcnt);
3262                                 break;
3263                         }
3264
3265                         nbytes = bio_iovec_idx(bio, idx)->bv_len;
3266                         BIO_BUG_ON(nbytes > bio->bi_size);
3267
3268                         /*
3269                          * not a complete bvec done
3270                          */
3271                         if (unlikely(nbytes > nr_bytes)) {
3272                                 bio_nbytes += nr_bytes;
3273                                 total_bytes += nr_bytes;
3274                                 break;
3275                         }
3276
3277                         /*
3278                          * advance to the next vector
3279                          */
3280                         next_idx++;
3281                         bio_nbytes += nbytes;
3282                 }
3283
3284                 total_bytes += nbytes;
3285                 nr_bytes -= nbytes;
3286
3287                 if ((bio = req->bio)) {
3288                         /*
3289                          * end more in this run, or just return 'not-done'
3290                          */
3291                         if (unlikely(nr_bytes <= 0))
3292                                 break;
3293                 }
3294         }
3295
3296         /*
3297          * completely done
3298          */
3299         if (!req->bio)
3300                 return 0;
3301
3302         /*
3303          * if the request wasn't completed, update state
3304          */
3305         if (bio_nbytes) {
3306                 if (!ordered_bio_endio(req, bio, bio_nbytes, error))
3307                         bio_endio(bio, bio_nbytes, error);
3308                 bio->bi_idx += next_idx;
3309                 bio_iovec(bio)->bv_offset += nr_bytes;
3310                 bio_iovec(bio)->bv_len -= nr_bytes;
3311         }
3312
3313         blk_recalc_rq_sectors(req, total_bytes >> 9);
3314         blk_recalc_rq_segments(req);
3315         return 1;
3316 }
3317
3318 /**
3319  * end_that_request_first - end I/O on a request
3320  * @req:      the request being processed
3321  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3322  * @nr_sectors: number of sectors to end I/O on
3323  *
3324  * Description:
3325  *     Ends I/O on a number of sectors attached to @req, and sets it up
3326  *     for the next range of segments (if any) in the cluster.
3327  *
3328  * Return:
3329  *     0 - we are done with this request, call end_that_request_last()
3330  *     1 - still buffers pending for this request
3331  **/
3332 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3333 {
3334         return __end_that_request_first(req, uptodate, nr_sectors << 9);
3335 }
3336
3337 EXPORT_SYMBOL(end_that_request_first);
3338
3339 /**
3340  * end_that_request_chunk - end I/O on a request
3341  * @req:      the request being processed
3342  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3343  * @nr_bytes: number of bytes to complete
3344  *
3345  * Description:
3346  *     Ends I/O on a number of bytes attached to @req, and sets it up
3347  *     for the next range of segments (if any). Like end_that_request_first(),
3348  *     but deals with bytes instead of sectors.
3349  *
3350  * Return:
3351  *     0 - we are done with this request, call end_that_request_last()
3352  *     1 - still buffers pending for this request
3353  **/
3354 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3355 {
3356         return __end_that_request_first(req, uptodate, nr_bytes);
3357 }
3358
3359 EXPORT_SYMBOL(end_that_request_chunk);
3360
3361 /*
3362  * splice the completion data to a local structure and hand off to
3363  * process_completion_queue() to complete the requests
3364  */
3365 static void blk_done_softirq(struct softirq_action *h)
3366 {
3367         struct list_head *cpu_list, local_list;
3368
3369         local_irq_disable();
3370         cpu_list = &__get_cpu_var(blk_cpu_done);
3371         list_replace_init(cpu_list, &local_list);
3372         local_irq_enable();
3373
3374         while (!list_empty(&local_list)) {
3375                 struct request *rq = list_entry(local_list.next, struct request, donelist);
3376
3377                 list_del_init(&rq->donelist);
3378                 rq->q->softirq_done_fn(rq);
3379         }
3380 }
3381
3382 #ifdef CONFIG_HOTPLUG_CPU
3383
3384 static int blk_cpu_notify(struct notifier_block *self, unsigned long action,
3385                           void *hcpu)
3386 {
3387         /*
3388          * If a CPU goes away, splice its entries to the current CPU
3389          * and trigger a run of the softirq
3390          */
3391         if (action == CPU_DEAD) {
3392                 int cpu = (unsigned long) hcpu;
3393
3394                 local_irq_disable();
3395                 list_splice_init(&per_cpu(blk_cpu_done, cpu),
3396                                  &__get_cpu_var(blk_cpu_done));
3397                 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3398                 local_irq_enable();
3399         }
3400
3401         return NOTIFY_OK;
3402 }
3403
3404
3405 static struct notifier_block __devinitdata blk_cpu_notifier = {
3406         .notifier_call  = blk_cpu_notify,
3407 };
3408
3409 #endif /* CONFIG_HOTPLUG_CPU */
3410
3411 /**
3412  * blk_complete_request - end I/O on a request
3413  * @req:      the request being processed
3414  *
3415  * Description:
3416  *     Ends all I/O on a request. It does not handle partial completions,
3417  *     unless the driver actually implements this in its completion callback
3418  *     through requeueing. Theh actual completion happens out-of-order,
3419  *     through a softirq handler. The user must have registered a completion
3420  *     callback through blk_queue_softirq_done().
3421  **/
3422
3423 void blk_complete_request(struct request *req)
3424 {
3425         struct list_head *cpu_list;
3426         unsigned long flags;
3427
3428         BUG_ON(!req->q->softirq_done_fn);
3429                 
3430         local_irq_save(flags);
3431
3432         cpu_list = &__get_cpu_var(blk_cpu_done);
3433         list_add_tail(&req->donelist, cpu_list);
3434         raise_softirq_irqoff(BLOCK_SOFTIRQ);
3435
3436         local_irq_restore(flags);
3437 }
3438
3439 EXPORT_SYMBOL(blk_complete_request);
3440         
3441 /*
3442  * queue lock must be held
3443  */
3444 void end_that_request_last(struct request *req, int uptodate)
3445 {
3446         struct gendisk *disk = req->rq_disk;
3447         int error;
3448
3449         /*
3450          * extend uptodate bool to allow < 0 value to be direct io error
3451          */
3452         error = 0;
3453         if (end_io_error(uptodate))
3454                 error = !uptodate ? -EIO : uptodate;
3455
3456         if (unlikely(laptop_mode) && blk_fs_request(req))
3457                 laptop_io_completion();
3458
3459         /*
3460          * Account IO completion.  bar_rq isn't accounted as a normal
3461          * IO on queueing nor completion.  Accounting the containing
3462          * request is enough.
3463          */
3464         if (disk && blk_fs_request(req) && req != &req->q->bar_rq) {
3465                 unsigned long duration = jiffies - req->start_time;
3466                 const int rw = rq_data_dir(req);
3467
3468                 __disk_stat_inc(disk, ios[rw]);
3469                 __disk_stat_add(disk, ticks[rw], duration);
3470                 disk_round_stats(disk);
3471                 disk->in_flight--;
3472         }
3473         if (req->end_io)
3474                 req->end_io(req, error);
3475         else
3476                 __blk_put_request(req->q, req);
3477 }
3478
3479 EXPORT_SYMBOL(end_that_request_last);
3480
3481 void end_request(struct request *req, int uptodate)
3482 {
3483         if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3484                 add_disk_randomness(req->rq_disk);
3485                 blkdev_dequeue_request(req);
3486                 end_that_request_last(req, uptodate);
3487         }
3488 }
3489
3490 EXPORT_SYMBOL(end_request);
3491
3492 void blk_rq_bio_prep(request_queue_t *q, struct request *rq, struct bio *bio)
3493 {
3494         /* first three bits are identical in rq->flags and bio->bi_rw */
3495         rq->flags |= (bio->bi_rw & 7);
3496
3497         rq->nr_phys_segments = bio_phys_segments(q, bio);
3498         rq->nr_hw_segments = bio_hw_segments(q, bio);
3499         rq->current_nr_sectors = bio_cur_sectors(bio);
3500         rq->hard_cur_sectors = rq->current_nr_sectors;
3501         rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3502         rq->buffer = bio_data(bio);
3503
3504         rq->bio = rq->biotail = bio;
3505 }
3506
3507 EXPORT_SYMBOL(blk_rq_bio_prep);
3508
3509 int kblockd_schedule_work(struct work_struct *work)
3510 {
3511         return queue_work(kblockd_workqueue, work);
3512 }
3513
3514 EXPORT_SYMBOL(kblockd_schedule_work);
3515
3516 void kblockd_flush(void)
3517 {
3518         flush_workqueue(kblockd_workqueue);
3519 }
3520 EXPORT_SYMBOL(kblockd_flush);
3521
3522 int __init blk_dev_init(void)
3523 {
3524         int i;
3525
3526         kblockd_workqueue = create_workqueue("kblockd");
3527         if (!kblockd_workqueue)
3528                 panic("Failed to create kblockd\n");
3529
3530         request_cachep = kmem_cache_create("blkdev_requests",
3531                         sizeof(struct request), 0, SLAB_PANIC, NULL, NULL);
3532
3533         requestq_cachep = kmem_cache_create("blkdev_queue",
3534                         sizeof(request_queue_t), 0, SLAB_PANIC, NULL, NULL);
3535
3536         iocontext_cachep = kmem_cache_create("blkdev_ioc",
3537                         sizeof(struct io_context), 0, SLAB_PANIC, NULL, NULL);
3538
3539         for_each_possible_cpu(i)
3540                 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3541
3542         open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
3543         register_hotcpu_notifier(&blk_cpu_notifier);
3544
3545         blk_max_low_pfn = max_low_pfn;
3546         blk_max_pfn = max_pfn;
3547
3548         return 0;
3549 }
3550
3551 /*
3552  * IO Context helper functions
3553  */
3554 void put_io_context(struct io_context *ioc)
3555 {
3556         if (ioc == NULL)
3557                 return;
3558
3559         BUG_ON(atomic_read(&ioc->refcount) == 0);
3560
3561         if (atomic_dec_and_test(&ioc->refcount)) {
3562                 struct cfq_io_context *cic;
3563
3564                 rcu_read_lock();
3565                 if (ioc->aic && ioc->aic->dtor)
3566                         ioc->aic->dtor(ioc->aic);
3567                 if (ioc->cic_root.rb_node != NULL) {
3568                         struct rb_node *n = rb_first(&ioc->cic_root);
3569
3570                         cic = rb_entry(n, struct cfq_io_context, rb_node);
3571                         cic->dtor(ioc);
3572                 }
3573                 rcu_read_unlock();
3574
3575                 kmem_cache_free(iocontext_cachep, ioc);
3576         }
3577 }
3578 EXPORT_SYMBOL(put_io_context);
3579
3580 /* Called by the exitting task */
3581 void exit_io_context(void)
3582 {
3583         unsigned long flags;
3584         struct io_context *ioc;
3585         struct cfq_io_context *cic;
3586
3587         local_irq_save(flags);
3588         task_lock(current);
3589         ioc = current->io_context;
3590         current->io_context = NULL;
3591         ioc->task = NULL;
3592         task_unlock(current);
3593         local_irq_restore(flags);
3594
3595         if (ioc->aic && ioc->aic->exit)
3596                 ioc->aic->exit(ioc->aic);
3597         if (ioc->cic_root.rb_node != NULL) {
3598                 cic = rb_entry(rb_first(&ioc->cic_root), struct cfq_io_context, rb_node);
3599                 cic->exit(ioc);
3600         }
3601  
3602         put_io_context(ioc);
3603 }
3604
3605 /*
3606  * If the current task has no IO context then create one and initialise it.
3607  * Otherwise, return its existing IO context.
3608  *
3609  * This returned IO context doesn't have a specifically elevated refcount,
3610  * but since the current task itself holds a reference, the context can be
3611  * used in general code, so long as it stays within `current` context.
3612  */
3613 struct io_context *current_io_context(gfp_t gfp_flags)
3614 {
3615         struct task_struct *tsk = current;
3616         struct io_context *ret;
3617
3618         ret = tsk->io_context;
3619         if (likely(ret))
3620                 return ret;
3621
3622         ret = kmem_cache_alloc(iocontext_cachep, gfp_flags);
3623         if (ret) {
3624                 atomic_set(&ret->refcount, 1);
3625                 ret->task = current;
3626                 ret->set_ioprio = NULL;
3627                 ret->last_waited = jiffies; /* doesn't matter... */
3628                 ret->nr_batch_requests = 0; /* because this is 0 */
3629                 ret->aic = NULL;
3630                 ret->cic_root.rb_node = NULL;
3631                 tsk->io_context = ret;
3632         }
3633
3634         return ret;
3635 }
3636 EXPORT_SYMBOL(current_io_context);
3637
3638 /*
3639  * If the current task has no IO context then create one and initialise it.
3640  * If it does have a context, take a ref on it.
3641  *
3642  * This is always called in the context of the task which submitted the I/O.
3643  */
3644 struct io_context *get_io_context(gfp_t gfp_flags)
3645 {
3646         struct io_context *ret;
3647         ret = current_io_context(gfp_flags);
3648         if (likely(ret))
3649                 atomic_inc(&ret->refcount);
3650         return ret;
3651 }
3652 EXPORT_SYMBOL(get_io_context);
3653
3654 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3655 {
3656         struct io_context *src = *psrc;
3657         struct io_context *dst = *pdst;
3658
3659         if (src) {
3660                 BUG_ON(atomic_read(&src->refcount) == 0);
3661                 atomic_inc(&src->refcount);
3662                 put_io_context(dst);
3663                 *pdst = src;
3664         }
3665 }
3666 EXPORT_SYMBOL(copy_io_context);
3667
3668 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3669 {
3670         struct io_context *temp;
3671         temp = *ioc1;
3672         *ioc1 = *ioc2;
3673         *ioc2 = temp;
3674 }
3675 EXPORT_SYMBOL(swap_io_context);
3676
3677 /*
3678  * sysfs parts below
3679  */
3680 struct queue_sysfs_entry {
3681         struct attribute attr;
3682         ssize_t (*show)(struct request_queue *, char *);
3683         ssize_t (*store)(struct request_queue *, const char *, size_t);
3684 };
3685
3686 static ssize_t
3687 queue_var_show(unsigned int var, char *page)
3688 {
3689         return sprintf(page, "%d\n", var);
3690 }
3691
3692 static ssize_t
3693 queue_var_store(unsigned long *var, const char *page, size_t count)
3694 {
3695         char *p = (char *) page;
3696
3697         *var = simple_strtoul(p, &p, 10);
3698         return count;
3699 }
3700
3701 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3702 {
3703         return queue_var_show(q->nr_requests, (page));
3704 }
3705
3706 static ssize_t
3707 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3708 {
3709         struct request_list *rl = &q->rq;
3710         unsigned long nr;
3711         int ret = queue_var_store(&nr, page, count);
3712         if (nr < BLKDEV_MIN_RQ)
3713                 nr = BLKDEV_MIN_RQ;
3714
3715         spin_lock_irq(q->queue_lock);
3716         q->nr_requests = nr;
3717         blk_queue_congestion_threshold(q);
3718
3719         if (rl->count[READ] >= queue_congestion_on_threshold(q))
3720                 set_queue_congested(q, READ);
3721         else if (rl->count[READ] < queue_congestion_off_threshold(q))
3722                 clear_queue_congested(q, READ);
3723
3724         if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3725                 set_queue_congested(q, WRITE);
3726         else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3727                 clear_queue_congested(q, WRITE);
3728
3729         if (rl->count[READ] >= q->nr_requests) {
3730                 blk_set_queue_full(q, READ);
3731         } else if (rl->count[READ]+1 <= q->nr_requests) {
3732                 blk_clear_queue_full(q, READ);
3733                 wake_up(&rl->wait[READ]);
3734         }
3735
3736         if (rl->count[WRITE] >= q->nr_requests) {
3737                 blk_set_queue_full(q, WRITE);
3738         } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3739                 blk_clear_queue_full(q, WRITE);
3740                 wake_up(&rl->wait[WRITE]);
3741         }
3742         spin_unlock_irq(q->queue_lock);
3743         return ret;
3744 }
3745
3746 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3747 {
3748         int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3749
3750         return queue_var_show(ra_kb, (page));
3751 }
3752
3753 static ssize_t
3754 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3755 {
3756         unsigned long ra_kb;
3757         ssize_t ret = queue_var_store(&ra_kb, page, count);
3758
3759         spin_lock_irq(q->queue_lock);
3760         if (ra_kb > (q->max_sectors >> 1))
3761                 ra_kb = (q->max_sectors >> 1);
3762
3763         q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3764         spin_unlock_irq(q->queue_lock);
3765
3766         return ret;
3767 }
3768
3769 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3770 {
3771         int max_sectors_kb = q->max_sectors >> 1;
3772
3773         return queue_var_show(max_sectors_kb, (page));
3774 }
3775
3776 static ssize_t
3777 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3778 {
3779         unsigned long max_sectors_kb,
3780                         max_hw_sectors_kb = q->max_hw_sectors >> 1,
3781                         page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3782         ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3783         int ra_kb;
3784
3785         if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3786                 return -EINVAL;
3787         /*
3788          * Take the queue lock to update the readahead and max_sectors
3789          * values synchronously:
3790          */
3791         spin_lock_irq(q->queue_lock);
3792         /*
3793          * Trim readahead window as well, if necessary:
3794          */
3795         ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3796         if (ra_kb > max_sectors_kb)
3797                 q->backing_dev_info.ra_pages =
3798                                 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3799
3800         q->max_sectors = max_sectors_kb << 1;
3801         spin_unlock_irq(q->queue_lock);
3802
3803         return ret;
3804 }
3805
3806 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3807 {
3808         int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3809
3810         return queue_var_show(max_hw_sectors_kb, (page));
3811 }
3812
3813
3814 static struct queue_sysfs_entry queue_requests_entry = {
3815         .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3816         .show = queue_requests_show,
3817         .store = queue_requests_store,
3818 };
3819
3820 static struct queue_sysfs_entry queue_ra_entry = {
3821         .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3822         .show = queue_ra_show,
3823         .store = queue_ra_store,
3824 };
3825
3826 static struct queue_sysfs_entry queue_max_sectors_entry = {
3827         .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3828         .show = queue_max_sectors_show,
3829         .store = queue_max_sectors_store,
3830 };
3831
3832 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3833         .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
3834         .show = queue_max_hw_sectors_show,
3835 };
3836
3837 static struct queue_sysfs_entry queue_iosched_entry = {
3838         .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
3839         .show = elv_iosched_show,
3840         .store = elv_iosched_store,
3841 };
3842
3843 static struct attribute *default_attrs[] = {
3844         &queue_requests_entry.attr,
3845         &queue_ra_entry.attr,
3846         &queue_max_hw_sectors_entry.attr,
3847         &queue_max_sectors_entry.attr,
3848         &queue_iosched_entry.attr,
3849         NULL,
3850 };
3851
3852 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
3853
3854 static ssize_t
3855 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
3856 {
3857         struct queue_sysfs_entry *entry = to_queue(attr);
3858         request_queue_t *q = container_of(kobj, struct request_queue, kobj);
3859         ssize_t res;
3860
3861         if (!entry->show)
3862                 return -EIO;
3863         mutex_lock(&q->sysfs_lock);
3864         if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
3865                 mutex_unlock(&q->sysfs_lock);
3866                 return -ENOENT;
3867         }
3868         res = entry->show(q, page);
3869         mutex_unlock(&q->sysfs_lock);
3870         return res;
3871 }
3872
3873 static ssize_t
3874 queue_attr_store(struct kobject *kobj, struct attribute *attr,
3875                     const char *page, size_t length)
3876 {
3877         struct queue_sysfs_entry *entry = to_queue(attr);
3878         request_queue_t *q = container_of(kobj, struct request_queue, kobj);
3879
3880         ssize_t res;
3881
3882         if (!entry->store)
3883                 return -EIO;
3884         mutex_lock(&q->sysfs_lock);
3885         if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
3886                 mutex_unlock(&q->sysfs_lock);
3887                 return -ENOENT;
3888         }
3889         res = entry->store(q, page, length);
3890         mutex_unlock(&q->sysfs_lock);
3891         return res;
3892 }
3893
3894 static struct sysfs_ops queue_sysfs_ops = {
3895         .show   = queue_attr_show,
3896         .store  = queue_attr_store,
3897 };
3898
3899 static struct kobj_type queue_ktype = {
3900         .sysfs_ops      = &queue_sysfs_ops,
3901         .default_attrs  = default_attrs,
3902         .release        = blk_release_queue,
3903 };
3904
3905 int blk_register_queue(struct gendisk *disk)
3906 {
3907         int ret;
3908
3909         request_queue_t *q = disk->queue;
3910
3911         if (!q || !q->request_fn)
3912                 return -ENXIO;
3913
3914         q->kobj.parent = kobject_get(&disk->kobj);
3915
3916         ret = kobject_add(&q->kobj);
3917         if (ret < 0)
3918                 return ret;
3919
3920         kobject_uevent(&q->kobj, KOBJ_ADD);
3921
3922         ret = elv_register_queue(q);
3923         if (ret) {
3924                 kobject_uevent(&q->kobj, KOBJ_REMOVE);
3925                 kobject_del(&q->kobj);
3926                 return ret;
3927         }
3928
3929         return 0;
3930 }
3931
3932 void blk_unregister_queue(struct gendisk *disk)
3933 {
3934         request_queue_t *q = disk->queue;
3935
3936         if (q && q->request_fn) {
3937                 elv_unregister_queue(q);
3938
3939                 kobject_uevent(&q->kobj, KOBJ_REMOVE);
3940                 kobject_del(&q->kobj);
3941                 kobject_put(&disk->kobj);
3942         }
3943 }