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