ocfs2: take meta data lock in ocfs2_file_aio_read()
[linux-2.6] / block / ll_rw_blk.c
1 /*
2  * Copyright (C) 1991, 1992 Linus Torvalds
3  * Copyright (C) 1994,      Karl Keyte: Added support for disk statistics
4  * Elevator latency, (C) 2000  Andrea Arcangeli <andrea@suse.de> SuSE
5  * Queue request tables / lock, selectable elevator, Jens Axboe <axboe@suse.de>
6  * kernel-doc documentation started by NeilBrown <neilb@cse.unsw.edu.au> -  July2000
7  * bio rewrite, highmem i/o, etc, Jens Axboe <axboe@suse.de> - may 2001
8  */
9
10 /*
11  * This handles all read/write requests to block devices
12  */
13 #include <linux/config.h>
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/backing-dev.h>
17 #include <linux/bio.h>
18 #include <linux/blkdev.h>
19 #include <linux/highmem.h>
20 #include <linux/mm.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/string.h>
23 #include <linux/init.h>
24 #include <linux/bootmem.h>      /* for max_pfn/max_low_pfn */
25 #include <linux/completion.h>
26 #include <linux/slab.h>
27 #include <linux/swap.h>
28 #include <linux/writeback.h>
29 #include <linux/interrupt.h>
30 #include <linux/cpu.h>
31 #include <linux/blktrace_api.h>
32
33 /*
34  * for max sense size
35  */
36 #include <scsi/scsi_cmnd.h>
37
38 static void blk_unplug_work(void *data);
39 static void blk_unplug_timeout(unsigned long data);
40 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io);
41 static void init_request_from_bio(struct request *req, struct bio *bio);
42 static int __make_request(request_queue_t *q, struct bio *bio);
43
44 /*
45  * For the allocated request tables
46  */
47 static kmem_cache_t *request_cachep;
48
49 /*
50  * For queue allocation
51  */
52 static kmem_cache_t *requestq_cachep;
53
54 /*
55  * For io context allocations
56  */
57 static kmem_cache_t *iocontext_cachep;
58
59 static wait_queue_head_t congestion_wqh[2] = {
60                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
61                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
62         };
63
64 /*
65  * Controlling structure to kblockd
66  */
67 static struct workqueue_struct *kblockd_workqueue;
68
69 unsigned long blk_max_low_pfn, blk_max_pfn;
70
71 EXPORT_SYMBOL(blk_max_low_pfn);
72 EXPORT_SYMBOL(blk_max_pfn);
73
74 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
75
76 /* Amount of time in which a process may batch requests */
77 #define BLK_BATCH_TIME  (HZ/50UL)
78
79 /* Number of requests a "batching" process may submit */
80 #define BLK_BATCH_REQ   32
81
82 /*
83  * Return the threshold (number of used requests) at which the queue is
84  * considered to be congested.  It include a little hysteresis to keep the
85  * context switch rate down.
86  */
87 static inline int queue_congestion_on_threshold(struct request_queue *q)
88 {
89         return q->nr_congestion_on;
90 }
91
92 /*
93  * The threshold at which a queue is considered to be uncongested
94  */
95 static inline int queue_congestion_off_threshold(struct request_queue *q)
96 {
97         return q->nr_congestion_off;
98 }
99
100 static void blk_queue_congestion_threshold(struct request_queue *q)
101 {
102         int nr;
103
104         nr = q->nr_requests - (q->nr_requests / 8) + 1;
105         if (nr > q->nr_requests)
106                 nr = q->nr_requests;
107         q->nr_congestion_on = nr;
108
109         nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
110         if (nr < 1)
111                 nr = 1;
112         q->nr_congestion_off = nr;
113 }
114
115 /*
116  * A queue has just exitted congestion.  Note this in the global counter of
117  * congested queues, and wake up anyone who was waiting for requests to be
118  * put back.
119  */
120 static void clear_queue_congested(request_queue_t *q, int rw)
121 {
122         enum bdi_state bit;
123         wait_queue_head_t *wqh = &congestion_wqh[rw];
124
125         bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
126         clear_bit(bit, &q->backing_dev_info.state);
127         smp_mb__after_clear_bit();
128         if (waitqueue_active(wqh))
129                 wake_up(wqh);
130 }
131
132 /*
133  * A queue has just entered congestion.  Flag that in the queue's VM-visible
134  * state flags and increment the global gounter of congested queues.
135  */
136 static void set_queue_congested(request_queue_t *q, int rw)
137 {
138         enum bdi_state bit;
139
140         bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
141         set_bit(bit, &q->backing_dev_info.state);
142 }
143
144 /**
145  * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
146  * @bdev:       device
147  *
148  * Locates the passed device's request queue and returns the address of its
149  * backing_dev_info
150  *
151  * Will return NULL if the request queue cannot be located.
152  */
153 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
154 {
155         struct backing_dev_info *ret = NULL;
156         request_queue_t *q = bdev_get_queue(bdev);
157
158         if (q)
159                 ret = &q->backing_dev_info;
160         return ret;
161 }
162
163 EXPORT_SYMBOL(blk_get_backing_dev_info);
164
165 void blk_queue_activity_fn(request_queue_t *q, activity_fn *fn, void *data)
166 {
167         q->activity_fn = fn;
168         q->activity_data = data;
169 }
170
171 EXPORT_SYMBOL(blk_queue_activity_fn);
172
173 /**
174  * blk_queue_prep_rq - set a prepare_request function for queue
175  * @q:          queue
176  * @pfn:        prepare_request function
177  *
178  * It's possible for a queue to register a prepare_request callback which
179  * is invoked before the request is handed to the request_fn. The goal of
180  * the function is to prepare a request for I/O, it can be used to build a
181  * cdb from the request data for instance.
182  *
183  */
184 void blk_queue_prep_rq(request_queue_t *q, prep_rq_fn *pfn)
185 {
186         q->prep_rq_fn = pfn;
187 }
188
189 EXPORT_SYMBOL(blk_queue_prep_rq);
190
191 /**
192  * blk_queue_merge_bvec - set a merge_bvec function for queue
193  * @q:          queue
194  * @mbfn:       merge_bvec_fn
195  *
196  * Usually queues have static limitations on the max sectors or segments that
197  * we can put in a request. Stacking drivers may have some settings that
198  * are dynamic, and thus we have to query the queue whether it is ok to
199  * add a new bio_vec to a bio at a given offset or not. If the block device
200  * has such limitations, it needs to register a merge_bvec_fn to control
201  * the size of bio's sent to it. Note that a block device *must* allow a
202  * single page to be added to an empty bio. The block device driver may want
203  * to use the bio_split() function to deal with these bio's. By default
204  * no merge_bvec_fn is defined for a queue, and only the fixed limits are
205  * honored.
206  */
207 void blk_queue_merge_bvec(request_queue_t *q, merge_bvec_fn *mbfn)
208 {
209         q->merge_bvec_fn = mbfn;
210 }
211
212 EXPORT_SYMBOL(blk_queue_merge_bvec);
213
214 void blk_queue_softirq_done(request_queue_t *q, softirq_done_fn *fn)
215 {
216         q->softirq_done_fn = fn;
217 }
218
219 EXPORT_SYMBOL(blk_queue_softirq_done);
220
221 /**
222  * blk_queue_make_request - define an alternate make_request function for a device
223  * @q:  the request queue for the device to be affected
224  * @mfn: the alternate make_request function
225  *
226  * Description:
227  *    The normal way for &struct bios to be passed to a device
228  *    driver is for them to be collected into requests on a request
229  *    queue, and then to allow the device driver to select requests
230  *    off that queue when it is ready.  This works well for many block
231  *    devices. However some block devices (typically virtual devices
232  *    such as md or lvm) do not benefit from the processing on the
233  *    request queue, and are served best by having the requests passed
234  *    directly to them.  This can be achieved by providing a function
235  *    to blk_queue_make_request().
236  *
237  * Caveat:
238  *    The driver that does this *must* be able to deal appropriately
239  *    with buffers in "highmemory". This can be accomplished by either calling
240  *    __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
241  *    blk_queue_bounce() to create a buffer in normal memory.
242  **/
243 void blk_queue_make_request(request_queue_t * q, make_request_fn * mfn)
244 {
245         /*
246          * set defaults
247          */
248         q->nr_requests = BLKDEV_MAX_RQ;
249         blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
250         blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
251         q->make_request_fn = mfn;
252         q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
253         q->backing_dev_info.state = 0;
254         q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
255         blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
256         blk_queue_hardsect_size(q, 512);
257         blk_queue_dma_alignment(q, 511);
258         blk_queue_congestion_threshold(q);
259         q->nr_batching = BLK_BATCH_REQ;
260
261         q->unplug_thresh = 4;           /* hmm */
262         q->unplug_delay = (3 * HZ) / 1000;      /* 3 milliseconds */
263         if (q->unplug_delay == 0)
264                 q->unplug_delay = 1;
265
266         INIT_WORK(&q->unplug_work, blk_unplug_work, q);
267
268         q->unplug_timer.function = blk_unplug_timeout;
269         q->unplug_timer.data = (unsigned long)q;
270
271         /*
272          * by default assume old behaviour and bounce for any highmem page
273          */
274         blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
275
276         blk_queue_activity_fn(q, NULL, NULL);
277 }
278
279 EXPORT_SYMBOL(blk_queue_make_request);
280
281 static inline void rq_init(request_queue_t *q, struct request *rq)
282 {
283         INIT_LIST_HEAD(&rq->queuelist);
284         INIT_LIST_HEAD(&rq->donelist);
285
286         rq->errors = 0;
287         rq->rq_status = RQ_ACTIVE;
288         rq->bio = rq->biotail = NULL;
289         rq->ioprio = 0;
290         rq->buffer = NULL;
291         rq->ref_count = 1;
292         rq->q = q;
293         rq->waiting = NULL;
294         rq->special = NULL;
295         rq->data_len = 0;
296         rq->data = NULL;
297         rq->nr_phys_segments = 0;
298         rq->sense = NULL;
299         rq->end_io = NULL;
300         rq->end_io_data = NULL;
301         rq->completion_data = NULL;
302 }
303
304 /**
305  * blk_queue_ordered - does this queue support ordered writes
306  * @q:        the request queue
307  * @ordered:  one of QUEUE_ORDERED_*
308  * @prepare_flush_fn: rq setup helper for cache flush ordered writes
309  *
310  * Description:
311  *   For journalled file systems, doing ordered writes on a commit
312  *   block instead of explicitly doing wait_on_buffer (which is bad
313  *   for performance) can be a big win. Block drivers supporting this
314  *   feature should call this function and indicate so.
315  *
316  **/
317 int blk_queue_ordered(request_queue_t *q, unsigned ordered,
318                       prepare_flush_fn *prepare_flush_fn)
319 {
320         if (ordered & (QUEUE_ORDERED_PREFLUSH | QUEUE_ORDERED_POSTFLUSH) &&
321             prepare_flush_fn == NULL) {
322                 printk(KERN_ERR "blk_queue_ordered: prepare_flush_fn required\n");
323                 return -EINVAL;
324         }
325
326         if (ordered != QUEUE_ORDERED_NONE &&
327             ordered != QUEUE_ORDERED_DRAIN &&
328             ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
329             ordered != QUEUE_ORDERED_DRAIN_FUA &&
330             ordered != QUEUE_ORDERED_TAG &&
331             ordered != QUEUE_ORDERED_TAG_FLUSH &&
332             ordered != QUEUE_ORDERED_TAG_FUA) {
333                 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
334                 return -EINVAL;
335         }
336
337         q->ordered = ordered;
338         q->next_ordered = ordered;
339         q->prepare_flush_fn = prepare_flush_fn;
340
341         return 0;
342 }
343
344 EXPORT_SYMBOL(blk_queue_ordered);
345
346 /**
347  * blk_queue_issue_flush_fn - set function for issuing a flush
348  * @q:     the request queue
349  * @iff:   the function to be called issuing the flush
350  *
351  * Description:
352  *   If a driver supports issuing a flush command, the support is notified
353  *   to the block layer by defining it through this call.
354  *
355  **/
356 void blk_queue_issue_flush_fn(request_queue_t *q, issue_flush_fn *iff)
357 {
358         q->issue_flush_fn = iff;
359 }
360
361 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
362
363 /*
364  * Cache flushing for ordered writes handling
365  */
366 inline unsigned blk_ordered_cur_seq(request_queue_t *q)
367 {
368         if (!q->ordseq)
369                 return 0;
370         return 1 << ffz(q->ordseq);
371 }
372
373 unsigned blk_ordered_req_seq(struct request *rq)
374 {
375         request_queue_t *q = rq->q;
376
377         BUG_ON(q->ordseq == 0);
378
379         if (rq == &q->pre_flush_rq)
380                 return QUEUE_ORDSEQ_PREFLUSH;
381         if (rq == &q->bar_rq)
382                 return QUEUE_ORDSEQ_BAR;
383         if (rq == &q->post_flush_rq)
384                 return QUEUE_ORDSEQ_POSTFLUSH;
385
386         if ((rq->flags & REQ_ORDERED_COLOR) ==
387             (q->orig_bar_rq->flags & REQ_ORDERED_COLOR))
388                 return QUEUE_ORDSEQ_DRAIN;
389         else
390                 return QUEUE_ORDSEQ_DONE;
391 }
392
393 void blk_ordered_complete_seq(request_queue_t *q, unsigned seq, int error)
394 {
395         struct request *rq;
396         int uptodate;
397
398         if (error && !q->orderr)
399                 q->orderr = error;
400
401         BUG_ON(q->ordseq & seq);
402         q->ordseq |= seq;
403
404         if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
405                 return;
406
407         /*
408          * Okay, sequence complete.
409          */
410         rq = q->orig_bar_rq;
411         uptodate = q->orderr ? q->orderr : 1;
412
413         q->ordseq = 0;
414
415         end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
416         end_that_request_last(rq, uptodate);
417 }
418
419 static void pre_flush_end_io(struct request *rq, int error)
420 {
421         elv_completed_request(rq->q, rq);
422         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
423 }
424
425 static void bar_end_io(struct request *rq, int error)
426 {
427         elv_completed_request(rq->q, rq);
428         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
429 }
430
431 static void post_flush_end_io(struct request *rq, int error)
432 {
433         elv_completed_request(rq->q, rq);
434         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
435 }
436
437 static void queue_flush(request_queue_t *q, unsigned which)
438 {
439         struct request *rq;
440         rq_end_io_fn *end_io;
441
442         if (which == QUEUE_ORDERED_PREFLUSH) {
443                 rq = &q->pre_flush_rq;
444                 end_io = pre_flush_end_io;
445         } else {
446                 rq = &q->post_flush_rq;
447                 end_io = post_flush_end_io;
448         }
449
450         rq_init(q, rq);
451         rq->flags = REQ_HARDBARRIER;
452         rq->elevator_private = NULL;
453         rq->rq_disk = q->bar_rq.rq_disk;
454         rq->rl = NULL;
455         rq->end_io = end_io;
456         q->prepare_flush_fn(q, rq);
457
458         elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
459 }
460
461 static inline struct request *start_ordered(request_queue_t *q,
462                                             struct request *rq)
463 {
464         q->bi_size = 0;
465         q->orderr = 0;
466         q->ordered = q->next_ordered;
467         q->ordseq |= QUEUE_ORDSEQ_STARTED;
468
469         /*
470          * Prep proxy barrier request.
471          */
472         blkdev_dequeue_request(rq);
473         q->orig_bar_rq = rq;
474         rq = &q->bar_rq;
475         rq_init(q, rq);
476         rq->flags = bio_data_dir(q->orig_bar_rq->bio);
477         rq->flags |= q->ordered & QUEUE_ORDERED_FUA ? REQ_FUA : 0;
478         rq->elevator_private = NULL;
479         rq->rl = NULL;
480         init_request_from_bio(rq, q->orig_bar_rq->bio);
481         rq->end_io = bar_end_io;
482
483         /*
484          * Queue ordered sequence.  As we stack them at the head, we
485          * need to queue in reverse order.  Note that we rely on that
486          * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
487          * request gets inbetween ordered sequence.
488          */
489         if (q->ordered & QUEUE_ORDERED_POSTFLUSH)
490                 queue_flush(q, QUEUE_ORDERED_POSTFLUSH);
491         else
492                 q->ordseq |= QUEUE_ORDSEQ_POSTFLUSH;
493
494         elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
495
496         if (q->ordered & QUEUE_ORDERED_PREFLUSH) {
497                 queue_flush(q, QUEUE_ORDERED_PREFLUSH);
498                 rq = &q->pre_flush_rq;
499         } else
500                 q->ordseq |= QUEUE_ORDSEQ_PREFLUSH;
501
502         if ((q->ordered & QUEUE_ORDERED_TAG) || q->in_flight == 0)
503                 q->ordseq |= QUEUE_ORDSEQ_DRAIN;
504         else
505                 rq = NULL;
506
507         return rq;
508 }
509
510 int blk_do_ordered(request_queue_t *q, struct request **rqp)
511 {
512         struct request *rq = *rqp;
513         int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq);
514
515         if (!q->ordseq) {
516                 if (!is_barrier)
517                         return 1;
518
519                 if (q->next_ordered != QUEUE_ORDERED_NONE) {
520                         *rqp = start_ordered(q, rq);
521                         return 1;
522                 } else {
523                         /*
524                          * This can happen when the queue switches to
525                          * ORDERED_NONE while this request is on it.
526                          */
527                         blkdev_dequeue_request(rq);
528                         end_that_request_first(rq, -EOPNOTSUPP,
529                                                rq->hard_nr_sectors);
530                         end_that_request_last(rq, -EOPNOTSUPP);
531                         *rqp = NULL;
532                         return 0;
533                 }
534         }
535
536         /*
537          * Ordered sequence in progress
538          */
539
540         /* Special requests are not subject to ordering rules. */
541         if (!blk_fs_request(rq) &&
542             rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
543                 return 1;
544
545         if (q->ordered & QUEUE_ORDERED_TAG) {
546                 /* Ordered by tag.  Blocking the next barrier is enough. */
547                 if (is_barrier && rq != &q->bar_rq)
548                         *rqp = NULL;
549         } else {
550                 /* Ordered by draining.  Wait for turn. */
551                 WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
552                 if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
553                         *rqp = NULL;
554         }
555
556         return 1;
557 }
558
559 static int flush_dry_bio_endio(struct bio *bio, unsigned int bytes, int error)
560 {
561         request_queue_t *q = bio->bi_private;
562         struct bio_vec *bvec;
563         int i;
564
565         /*
566          * This is dry run, restore bio_sector and size.  We'll finish
567          * this request again with the original bi_end_io after an
568          * error occurs or post flush is complete.
569          */
570         q->bi_size += bytes;
571
572         if (bio->bi_size)
573                 return 1;
574
575         /* Rewind bvec's */
576         bio->bi_idx = 0;
577         bio_for_each_segment(bvec, bio, i) {
578                 bvec->bv_len += bvec->bv_offset;
579                 bvec->bv_offset = 0;
580         }
581
582         /* Reset bio */
583         set_bit(BIO_UPTODATE, &bio->bi_flags);
584         bio->bi_size = q->bi_size;
585         bio->bi_sector -= (q->bi_size >> 9);
586         q->bi_size = 0;
587
588         return 0;
589 }
590
591 static inline int ordered_bio_endio(struct request *rq, struct bio *bio,
592                                     unsigned int nbytes, int error)
593 {
594         request_queue_t *q = rq->q;
595         bio_end_io_t *endio;
596         void *private;
597
598         if (&q->bar_rq != rq)
599                 return 0;
600
601         /*
602          * Okay, this is the barrier request in progress, dry finish it.
603          */
604         if (error && !q->orderr)
605                 q->orderr = error;
606
607         endio = bio->bi_end_io;
608         private = bio->bi_private;
609         bio->bi_end_io = flush_dry_bio_endio;
610         bio->bi_private = q;
611
612         bio_endio(bio, nbytes, error);
613
614         bio->bi_end_io = endio;
615         bio->bi_private = private;
616
617         return 1;
618 }
619
620 /**
621  * blk_queue_bounce_limit - set bounce buffer limit for queue
622  * @q:  the request queue for the device
623  * @dma_addr:   bus address limit
624  *
625  * Description:
626  *    Different hardware can have different requirements as to what pages
627  *    it can do I/O directly to. A low level driver can call
628  *    blk_queue_bounce_limit to have lower memory pages allocated as bounce
629  *    buffers for doing I/O to pages residing above @page.
630  **/
631 void blk_queue_bounce_limit(request_queue_t *q, u64 dma_addr)
632 {
633         unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
634         int dma = 0;
635
636         q->bounce_gfp = GFP_NOIO;
637 #if BITS_PER_LONG == 64
638         /* Assume anything <= 4GB can be handled by IOMMU.
639            Actually some IOMMUs can handle everything, but I don't
640            know of a way to test this here. */
641         if (bounce_pfn < (0xffffffff>>PAGE_SHIFT))
642                 dma = 1;
643         q->bounce_pfn = max_low_pfn;
644 #else
645         if (bounce_pfn < blk_max_low_pfn)
646                 dma = 1;
647         q->bounce_pfn = bounce_pfn;
648 #endif
649         if (dma) {
650                 init_emergency_isa_pool();
651                 q->bounce_gfp = GFP_NOIO | GFP_DMA;
652                 q->bounce_pfn = bounce_pfn;
653         }
654 }
655
656 EXPORT_SYMBOL(blk_queue_bounce_limit);
657
658 /**
659  * blk_queue_max_sectors - set max sectors for a request for this queue
660  * @q:  the request queue for the device
661  * @max_sectors:  max sectors in the usual 512b unit
662  *
663  * Description:
664  *    Enables a low level driver to set an upper limit on the size of
665  *    received requests.
666  **/
667 void blk_queue_max_sectors(request_queue_t *q, unsigned int max_sectors)
668 {
669         if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
670                 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
671                 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
672         }
673
674         if (BLK_DEF_MAX_SECTORS > max_sectors)
675                 q->max_hw_sectors = q->max_sectors = max_sectors;
676         else {
677                 q->max_sectors = BLK_DEF_MAX_SECTORS;
678                 q->max_hw_sectors = max_sectors;
679         }
680 }
681
682 EXPORT_SYMBOL(blk_queue_max_sectors);
683
684 /**
685  * blk_queue_max_phys_segments - set max phys segments for a request for this queue
686  * @q:  the request queue for the device
687  * @max_segments:  max number of segments
688  *
689  * Description:
690  *    Enables a low level driver to set an upper limit on the number of
691  *    physical data segments in a request.  This would be the largest sized
692  *    scatter list the driver could handle.
693  **/
694 void blk_queue_max_phys_segments(request_queue_t *q, unsigned short max_segments)
695 {
696         if (!max_segments) {
697                 max_segments = 1;
698                 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
699         }
700
701         q->max_phys_segments = max_segments;
702 }
703
704 EXPORT_SYMBOL(blk_queue_max_phys_segments);
705
706 /**
707  * blk_queue_max_hw_segments - set max hw segments for a request for this queue
708  * @q:  the request queue for the device
709  * @max_segments:  max number of segments
710  *
711  * Description:
712  *    Enables a low level driver to set an upper limit on the number of
713  *    hw data segments in a request.  This would be the largest number of
714  *    address/length pairs the host adapter can actually give as once
715  *    to the device.
716  **/
717 void blk_queue_max_hw_segments(request_queue_t *q, unsigned short max_segments)
718 {
719         if (!max_segments) {
720                 max_segments = 1;
721                 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
722         }
723
724         q->max_hw_segments = max_segments;
725 }
726
727 EXPORT_SYMBOL(blk_queue_max_hw_segments);
728
729 /**
730  * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
731  * @q:  the request queue for the device
732  * @max_size:  max size of segment in bytes
733  *
734  * Description:
735  *    Enables a low level driver to set an upper limit on the size of a
736  *    coalesced segment
737  **/
738 void blk_queue_max_segment_size(request_queue_t *q, unsigned int max_size)
739 {
740         if (max_size < PAGE_CACHE_SIZE) {
741                 max_size = PAGE_CACHE_SIZE;
742                 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
743         }
744
745         q->max_segment_size = max_size;
746 }
747
748 EXPORT_SYMBOL(blk_queue_max_segment_size);
749
750 /**
751  * blk_queue_hardsect_size - set hardware sector size for the queue
752  * @q:  the request queue for the device
753  * @size:  the hardware sector size, in bytes
754  *
755  * Description:
756  *   This should typically be set to the lowest possible sector size
757  *   that the hardware can operate on (possible without reverting to
758  *   even internal read-modify-write operations). Usually the default
759  *   of 512 covers most hardware.
760  **/
761 void blk_queue_hardsect_size(request_queue_t *q, unsigned short size)
762 {
763         q->hardsect_size = size;
764 }
765
766 EXPORT_SYMBOL(blk_queue_hardsect_size);
767
768 /*
769  * Returns the minimum that is _not_ zero, unless both are zero.
770  */
771 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
772
773 /**
774  * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
775  * @t:  the stacking driver (top)
776  * @b:  the underlying device (bottom)
777  **/
778 void blk_queue_stack_limits(request_queue_t *t, request_queue_t *b)
779 {
780         /* zero is "infinity" */
781         t->max_sectors = min_not_zero(t->max_sectors,b->max_sectors);
782         t->max_hw_sectors = min_not_zero(t->max_hw_sectors,b->max_hw_sectors);
783
784         t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
785         t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
786         t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
787         t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
788         if (!test_bit(QUEUE_FLAG_CLUSTER, &b->queue_flags))
789                 clear_bit(QUEUE_FLAG_CLUSTER, &t->queue_flags);
790 }
791
792 EXPORT_SYMBOL(blk_queue_stack_limits);
793
794 /**
795  * blk_queue_segment_boundary - set boundary rules for segment merging
796  * @q:  the request queue for the device
797  * @mask:  the memory boundary mask
798  **/
799 void blk_queue_segment_boundary(request_queue_t *q, unsigned long mask)
800 {
801         if (mask < PAGE_CACHE_SIZE - 1) {
802                 mask = PAGE_CACHE_SIZE - 1;
803                 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
804         }
805
806         q->seg_boundary_mask = mask;
807 }
808
809 EXPORT_SYMBOL(blk_queue_segment_boundary);
810
811 /**
812  * blk_queue_dma_alignment - set dma length and memory alignment
813  * @q:     the request queue for the device
814  * @mask:  alignment mask
815  *
816  * description:
817  *    set required memory and length aligment for direct dma transactions.
818  *    this is used when buiding direct io requests for the queue.
819  *
820  **/
821 void blk_queue_dma_alignment(request_queue_t *q, int mask)
822 {
823         q->dma_alignment = mask;
824 }
825
826 EXPORT_SYMBOL(blk_queue_dma_alignment);
827
828 /**
829  * blk_queue_find_tag - find a request by its tag and queue
830  * @q:   The request queue for the device
831  * @tag: The tag of the request
832  *
833  * Notes:
834  *    Should be used when a device returns a tag and you want to match
835  *    it with a request.
836  *
837  *    no locks need be held.
838  **/
839 struct request *blk_queue_find_tag(request_queue_t *q, int tag)
840 {
841         struct blk_queue_tag *bqt = q->queue_tags;
842
843         if (unlikely(bqt == NULL || tag >= bqt->real_max_depth))
844                 return NULL;
845
846         return bqt->tag_index[tag];
847 }
848
849 EXPORT_SYMBOL(blk_queue_find_tag);
850
851 /**
852  * __blk_queue_free_tags - release tag maintenance info
853  * @q:  the request queue for the device
854  *
855  *  Notes:
856  *    blk_cleanup_queue() will take care of calling this function, if tagging
857  *    has been used. So there's no need to call this directly.
858  **/
859 static void __blk_queue_free_tags(request_queue_t *q)
860 {
861         struct blk_queue_tag *bqt = q->queue_tags;
862
863         if (!bqt)
864                 return;
865
866         if (atomic_dec_and_test(&bqt->refcnt)) {
867                 BUG_ON(bqt->busy);
868                 BUG_ON(!list_empty(&bqt->busy_list));
869
870                 kfree(bqt->tag_index);
871                 bqt->tag_index = NULL;
872
873                 kfree(bqt->tag_map);
874                 bqt->tag_map = NULL;
875
876                 kfree(bqt);
877         }
878
879         q->queue_tags = NULL;
880         q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
881 }
882
883 /**
884  * blk_queue_free_tags - release tag maintenance info
885  * @q:  the request queue for the device
886  *
887  *  Notes:
888  *      This is used to disabled tagged queuing to a device, yet leave
889  *      queue in function.
890  **/
891 void blk_queue_free_tags(request_queue_t *q)
892 {
893         clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
894 }
895
896 EXPORT_SYMBOL(blk_queue_free_tags);
897
898 static int
899 init_tag_map(request_queue_t *q, struct blk_queue_tag *tags, int depth)
900 {
901         struct request **tag_index;
902         unsigned long *tag_map;
903         int nr_ulongs;
904
905         if (depth > q->nr_requests * 2) {
906                 depth = q->nr_requests * 2;
907                 printk(KERN_ERR "%s: adjusted depth to %d\n",
908                                 __FUNCTION__, depth);
909         }
910
911         tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC);
912         if (!tag_index)
913                 goto fail;
914
915         nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
916         tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
917         if (!tag_map)
918                 goto fail;
919
920         tags->real_max_depth = depth;
921         tags->max_depth = depth;
922         tags->tag_index = tag_index;
923         tags->tag_map = tag_map;
924
925         return 0;
926 fail:
927         kfree(tag_index);
928         return -ENOMEM;
929 }
930
931 /**
932  * blk_queue_init_tags - initialize the queue tag info
933  * @q:  the request queue for the device
934  * @depth:  the maximum queue depth supported
935  * @tags: the tag to use
936  **/
937 int blk_queue_init_tags(request_queue_t *q, int depth,
938                         struct blk_queue_tag *tags)
939 {
940         int rc;
941
942         BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
943
944         if (!tags && !q->queue_tags) {
945                 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
946                 if (!tags)
947                         goto fail;
948
949                 if (init_tag_map(q, tags, depth))
950                         goto fail;
951
952                 INIT_LIST_HEAD(&tags->busy_list);
953                 tags->busy = 0;
954                 atomic_set(&tags->refcnt, 1);
955         } else if (q->queue_tags) {
956                 if ((rc = blk_queue_resize_tags(q, depth)))
957                         return rc;
958                 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
959                 return 0;
960         } else
961                 atomic_inc(&tags->refcnt);
962
963         /*
964          * assign it, all done
965          */
966         q->queue_tags = tags;
967         q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
968         return 0;
969 fail:
970         kfree(tags);
971         return -ENOMEM;
972 }
973
974 EXPORT_SYMBOL(blk_queue_init_tags);
975
976 /**
977  * blk_queue_resize_tags - change the queueing depth
978  * @q:  the request queue for the device
979  * @new_depth: the new max command queueing depth
980  *
981  *  Notes:
982  *    Must be called with the queue lock held.
983  **/
984 int blk_queue_resize_tags(request_queue_t *q, int new_depth)
985 {
986         struct blk_queue_tag *bqt = q->queue_tags;
987         struct request **tag_index;
988         unsigned long *tag_map;
989         int max_depth, nr_ulongs;
990
991         if (!bqt)
992                 return -ENXIO;
993
994         /*
995          * if we already have large enough real_max_depth.  just
996          * adjust max_depth.  *NOTE* as requests with tag value
997          * between new_depth and real_max_depth can be in-flight, tag
998          * map can not be shrunk blindly here.
999          */
1000         if (new_depth <= bqt->real_max_depth) {
1001                 bqt->max_depth = new_depth;
1002                 return 0;
1003         }
1004
1005         /*
1006          * save the old state info, so we can copy it back
1007          */
1008         tag_index = bqt->tag_index;
1009         tag_map = bqt->tag_map;
1010         max_depth = bqt->real_max_depth;
1011
1012         if (init_tag_map(q, bqt, new_depth))
1013                 return -ENOMEM;
1014
1015         memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
1016         nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
1017         memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
1018
1019         kfree(tag_index);
1020         kfree(tag_map);
1021         return 0;
1022 }
1023
1024 EXPORT_SYMBOL(blk_queue_resize_tags);
1025
1026 /**
1027  * blk_queue_end_tag - end tag operations for a request
1028  * @q:  the request queue for the device
1029  * @rq: the request that has completed
1030  *
1031  *  Description:
1032  *    Typically called when end_that_request_first() returns 0, meaning
1033  *    all transfers have been done for a request. It's important to call
1034  *    this function before end_that_request_last(), as that will put the
1035  *    request back on the free list thus corrupting the internal tag list.
1036  *
1037  *  Notes:
1038  *   queue lock must be held.
1039  **/
1040 void blk_queue_end_tag(request_queue_t *q, struct request *rq)
1041 {
1042         struct blk_queue_tag *bqt = q->queue_tags;
1043         int tag = rq->tag;
1044
1045         BUG_ON(tag == -1);
1046
1047         if (unlikely(tag >= bqt->real_max_depth))
1048                 /*
1049                  * This can happen after tag depth has been reduced.
1050                  * FIXME: how about a warning or info message here?
1051                  */
1052                 return;
1053
1054         if (unlikely(!__test_and_clear_bit(tag, bqt->tag_map))) {
1055                 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
1056                        __FUNCTION__, tag);
1057                 return;
1058         }
1059
1060         list_del_init(&rq->queuelist);
1061         rq->flags &= ~REQ_QUEUED;
1062         rq->tag = -1;
1063
1064         if (unlikely(bqt->tag_index[tag] == NULL))
1065                 printk(KERN_ERR "%s: tag %d is missing\n",
1066                        __FUNCTION__, tag);
1067
1068         bqt->tag_index[tag] = NULL;
1069         bqt->busy--;
1070 }
1071
1072 EXPORT_SYMBOL(blk_queue_end_tag);
1073
1074 /**
1075  * blk_queue_start_tag - find a free tag and assign it
1076  * @q:  the request queue for the device
1077  * @rq:  the block request that needs tagging
1078  *
1079  *  Description:
1080  *    This can either be used as a stand-alone helper, or possibly be
1081  *    assigned as the queue &prep_rq_fn (in which case &struct request
1082  *    automagically gets a tag assigned). Note that this function
1083  *    assumes that any type of request can be queued! if this is not
1084  *    true for your device, you must check the request type before
1085  *    calling this function.  The request will also be removed from
1086  *    the request queue, so it's the drivers responsibility to readd
1087  *    it if it should need to be restarted for some reason.
1088  *
1089  *  Notes:
1090  *   queue lock must be held.
1091  **/
1092 int blk_queue_start_tag(request_queue_t *q, struct request *rq)
1093 {
1094         struct blk_queue_tag *bqt = q->queue_tags;
1095         int tag;
1096
1097         if (unlikely((rq->flags & REQ_QUEUED))) {
1098                 printk(KERN_ERR 
1099                        "%s: request %p for device [%s] already tagged %d",
1100                        __FUNCTION__, rq,
1101                        rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
1102                 BUG();
1103         }
1104
1105         tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
1106         if (tag >= bqt->max_depth)
1107                 return 1;
1108
1109         __set_bit(tag, bqt->tag_map);
1110
1111         rq->flags |= REQ_QUEUED;
1112         rq->tag = tag;
1113         bqt->tag_index[tag] = rq;
1114         blkdev_dequeue_request(rq);
1115         list_add(&rq->queuelist, &bqt->busy_list);
1116         bqt->busy++;
1117         return 0;
1118 }
1119
1120 EXPORT_SYMBOL(blk_queue_start_tag);
1121
1122 /**
1123  * blk_queue_invalidate_tags - invalidate all pending tags
1124  * @q:  the request queue for the device
1125  *
1126  *  Description:
1127  *   Hardware conditions may dictate a need to stop all pending requests.
1128  *   In this case, we will safely clear the block side of the tag queue and
1129  *   readd all requests to the request queue in the right order.
1130  *
1131  *  Notes:
1132  *   queue lock must be held.
1133  **/
1134 void blk_queue_invalidate_tags(request_queue_t *q)
1135 {
1136         struct blk_queue_tag *bqt = q->queue_tags;
1137         struct list_head *tmp, *n;
1138         struct request *rq;
1139
1140         list_for_each_safe(tmp, n, &bqt->busy_list) {
1141                 rq = list_entry_rq(tmp);
1142
1143                 if (rq->tag == -1) {
1144                         printk(KERN_ERR
1145                                "%s: bad tag found on list\n", __FUNCTION__);
1146                         list_del_init(&rq->queuelist);
1147                         rq->flags &= ~REQ_QUEUED;
1148                 } else
1149                         blk_queue_end_tag(q, rq);
1150
1151                 rq->flags &= ~REQ_STARTED;
1152                 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1153         }
1154 }
1155
1156 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1157
1158 static const char * const rq_flags[] = {
1159         "REQ_RW",
1160         "REQ_FAILFAST",
1161         "REQ_SORTED",
1162         "REQ_SOFTBARRIER",
1163         "REQ_HARDBARRIER",
1164         "REQ_FUA",
1165         "REQ_CMD",
1166         "REQ_NOMERGE",
1167         "REQ_STARTED",
1168         "REQ_DONTPREP",
1169         "REQ_QUEUED",
1170         "REQ_ELVPRIV",
1171         "REQ_PC",
1172         "REQ_BLOCK_PC",
1173         "REQ_SENSE",
1174         "REQ_FAILED",
1175         "REQ_QUIET",
1176         "REQ_SPECIAL",
1177         "REQ_DRIVE_CMD",
1178         "REQ_DRIVE_TASK",
1179         "REQ_DRIVE_TASKFILE",
1180         "REQ_PREEMPT",
1181         "REQ_PM_SUSPEND",
1182         "REQ_PM_RESUME",
1183         "REQ_PM_SHUTDOWN",
1184         "REQ_ORDERED_COLOR",
1185 };
1186
1187 void blk_dump_rq_flags(struct request *rq, char *msg)
1188 {
1189         int bit;
1190
1191         printk("%s: dev %s: flags = ", msg,
1192                 rq->rq_disk ? rq->rq_disk->disk_name : "?");
1193         bit = 0;
1194         do {
1195                 if (rq->flags & (1 << bit))
1196                         printk("%s ", rq_flags[bit]);
1197                 bit++;
1198         } while (bit < __REQ_NR_BITS);
1199
1200         printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1201                                                        rq->nr_sectors,
1202                                                        rq->current_nr_sectors);
1203         printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1204
1205         if (rq->flags & (REQ_BLOCK_PC | REQ_PC)) {
1206                 printk("cdb: ");
1207                 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1208                         printk("%02x ", rq->cmd[bit]);
1209                 printk("\n");
1210         }
1211 }
1212
1213 EXPORT_SYMBOL(blk_dump_rq_flags);
1214
1215 void blk_recount_segments(request_queue_t *q, struct bio *bio)
1216 {
1217         struct bio_vec *bv, *bvprv = NULL;
1218         int i, nr_phys_segs, nr_hw_segs, seg_size, hw_seg_size, cluster;
1219         int high, highprv = 1;
1220
1221         if (unlikely(!bio->bi_io_vec))
1222                 return;
1223
1224         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1225         hw_seg_size = seg_size = nr_phys_segs = nr_hw_segs = 0;
1226         bio_for_each_segment(bv, bio, i) {
1227                 /*
1228                  * the trick here is making sure that a high page is never
1229                  * considered part of another segment, since that might
1230                  * change with the bounce page.
1231                  */
1232                 high = page_to_pfn(bv->bv_page) >= q->bounce_pfn;
1233                 if (high || highprv)
1234                         goto new_hw_segment;
1235                 if (cluster) {
1236                         if (seg_size + bv->bv_len > q->max_segment_size)
1237                                 goto new_segment;
1238                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1239                                 goto new_segment;
1240                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1241                                 goto new_segment;
1242                         if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1243                                 goto new_hw_segment;
1244
1245                         seg_size += bv->bv_len;
1246                         hw_seg_size += bv->bv_len;
1247                         bvprv = bv;
1248                         continue;
1249                 }
1250 new_segment:
1251                 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1252                     !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len)) {
1253                         hw_seg_size += bv->bv_len;
1254                 } else {
1255 new_hw_segment:
1256                         if (hw_seg_size > bio->bi_hw_front_size)
1257                                 bio->bi_hw_front_size = hw_seg_size;
1258                         hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1259                         nr_hw_segs++;
1260                 }
1261
1262                 nr_phys_segs++;
1263                 bvprv = bv;
1264                 seg_size = bv->bv_len;
1265                 highprv = high;
1266         }
1267         if (hw_seg_size > bio->bi_hw_back_size)
1268                 bio->bi_hw_back_size = hw_seg_size;
1269         if (nr_hw_segs == 1 && hw_seg_size > bio->bi_hw_front_size)
1270                 bio->bi_hw_front_size = hw_seg_size;
1271         bio->bi_phys_segments = nr_phys_segs;
1272         bio->bi_hw_segments = nr_hw_segs;
1273         bio->bi_flags |= (1 << BIO_SEG_VALID);
1274 }
1275
1276
1277 static int blk_phys_contig_segment(request_queue_t *q, struct bio *bio,
1278                                    struct bio *nxt)
1279 {
1280         if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1281                 return 0;
1282
1283         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1284                 return 0;
1285         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1286                 return 0;
1287
1288         /*
1289          * bio and nxt are contigous in memory, check if the queue allows
1290          * these two to be merged into one
1291          */
1292         if (BIO_SEG_BOUNDARY(q, bio, nxt))
1293                 return 1;
1294
1295         return 0;
1296 }
1297
1298 static int blk_hw_contig_segment(request_queue_t *q, struct bio *bio,
1299                                  struct bio *nxt)
1300 {
1301         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1302                 blk_recount_segments(q, bio);
1303         if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1304                 blk_recount_segments(q, nxt);
1305         if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1306             BIOVEC_VIRT_OVERSIZE(bio->bi_hw_front_size + bio->bi_hw_back_size))
1307                 return 0;
1308         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1309                 return 0;
1310
1311         return 1;
1312 }
1313
1314 /*
1315  * map a request to scatterlist, return number of sg entries setup. Caller
1316  * must make sure sg can hold rq->nr_phys_segments entries
1317  */
1318 int blk_rq_map_sg(request_queue_t *q, struct request *rq, struct scatterlist *sg)
1319 {
1320         struct bio_vec *bvec, *bvprv;
1321         struct bio *bio;
1322         int nsegs, i, cluster;
1323
1324         nsegs = 0;
1325         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1326
1327         /*
1328          * for each bio in rq
1329          */
1330         bvprv = NULL;
1331         rq_for_each_bio(bio, rq) {
1332                 /*
1333                  * for each segment in bio
1334                  */
1335                 bio_for_each_segment(bvec, bio, i) {
1336                         int nbytes = bvec->bv_len;
1337
1338                         if (bvprv && cluster) {
1339                                 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1340                                         goto new_segment;
1341
1342                                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1343                                         goto new_segment;
1344                                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1345                                         goto new_segment;
1346
1347                                 sg[nsegs - 1].length += nbytes;
1348                         } else {
1349 new_segment:
1350                                 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1351                                 sg[nsegs].page = bvec->bv_page;
1352                                 sg[nsegs].length = nbytes;
1353                                 sg[nsegs].offset = bvec->bv_offset;
1354
1355                                 nsegs++;
1356                         }
1357                         bvprv = bvec;
1358                 } /* segments in bio */
1359         } /* bios in rq */
1360
1361         return nsegs;
1362 }
1363
1364 EXPORT_SYMBOL(blk_rq_map_sg);
1365
1366 /*
1367  * the standard queue merge functions, can be overridden with device
1368  * specific ones if so desired
1369  */
1370
1371 static inline int ll_new_mergeable(request_queue_t *q,
1372                                    struct request *req,
1373                                    struct bio *bio)
1374 {
1375         int nr_phys_segs = bio_phys_segments(q, bio);
1376
1377         if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1378                 req->flags |= REQ_NOMERGE;
1379                 if (req == q->last_merge)
1380                         q->last_merge = NULL;
1381                 return 0;
1382         }
1383
1384         /*
1385          * A hw segment is just getting larger, bump just the phys
1386          * counter.
1387          */
1388         req->nr_phys_segments += nr_phys_segs;
1389         return 1;
1390 }
1391
1392 static inline int ll_new_hw_segment(request_queue_t *q,
1393                                     struct request *req,
1394                                     struct bio *bio)
1395 {
1396         int nr_hw_segs = bio_hw_segments(q, bio);
1397         int nr_phys_segs = bio_phys_segments(q, bio);
1398
1399         if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1400             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1401                 req->flags |= REQ_NOMERGE;
1402                 if (req == q->last_merge)
1403                         q->last_merge = NULL;
1404                 return 0;
1405         }
1406
1407         /*
1408          * This will form the start of a new hw segment.  Bump both
1409          * counters.
1410          */
1411         req->nr_hw_segments += nr_hw_segs;
1412         req->nr_phys_segments += nr_phys_segs;
1413         return 1;
1414 }
1415
1416 static int ll_back_merge_fn(request_queue_t *q, struct request *req, 
1417                             struct bio *bio)
1418 {
1419         unsigned short max_sectors;
1420         int len;
1421
1422         if (unlikely(blk_pc_request(req)))
1423                 max_sectors = q->max_hw_sectors;
1424         else
1425                 max_sectors = q->max_sectors;
1426
1427         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1428                 req->flags |= REQ_NOMERGE;
1429                 if (req == q->last_merge)
1430                         q->last_merge = NULL;
1431                 return 0;
1432         }
1433         if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1434                 blk_recount_segments(q, req->biotail);
1435         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1436                 blk_recount_segments(q, bio);
1437         len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1438         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1439             !BIOVEC_VIRT_OVERSIZE(len)) {
1440                 int mergeable =  ll_new_mergeable(q, req, bio);
1441
1442                 if (mergeable) {
1443                         if (req->nr_hw_segments == 1)
1444                                 req->bio->bi_hw_front_size = len;
1445                         if (bio->bi_hw_segments == 1)
1446                                 bio->bi_hw_back_size = len;
1447                 }
1448                 return mergeable;
1449         }
1450
1451         return ll_new_hw_segment(q, req, bio);
1452 }
1453
1454 static int ll_front_merge_fn(request_queue_t *q, struct request *req, 
1455                              struct bio *bio)
1456 {
1457         unsigned short max_sectors;
1458         int len;
1459
1460         if (unlikely(blk_pc_request(req)))
1461                 max_sectors = q->max_hw_sectors;
1462         else
1463                 max_sectors = q->max_sectors;
1464
1465
1466         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
1467                 req->flags |= REQ_NOMERGE;
1468                 if (req == q->last_merge)
1469                         q->last_merge = NULL;
1470                 return 0;
1471         }
1472         len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1473         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1474                 blk_recount_segments(q, bio);
1475         if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1476                 blk_recount_segments(q, req->bio);
1477         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1478             !BIOVEC_VIRT_OVERSIZE(len)) {
1479                 int mergeable =  ll_new_mergeable(q, req, bio);
1480
1481                 if (mergeable) {
1482                         if (bio->bi_hw_segments == 1)
1483                                 bio->bi_hw_front_size = len;
1484                         if (req->nr_hw_segments == 1)
1485                                 req->biotail->bi_hw_back_size = len;
1486                 }
1487                 return mergeable;
1488         }
1489
1490         return ll_new_hw_segment(q, req, bio);
1491 }
1492
1493 static int ll_merge_requests_fn(request_queue_t *q, struct request *req,
1494                                 struct request *next)
1495 {
1496         int total_phys_segments;
1497         int total_hw_segments;
1498
1499         /*
1500          * First check if the either of the requests are re-queued
1501          * requests.  Can't merge them if they are.
1502          */
1503         if (req->special || next->special)
1504                 return 0;
1505
1506         /*
1507          * Will it become too large?
1508          */
1509         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1510                 return 0;
1511
1512         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1513         if (blk_phys_contig_segment(q, req->biotail, next->bio))
1514                 total_phys_segments--;
1515
1516         if (total_phys_segments > q->max_phys_segments)
1517                 return 0;
1518
1519         total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1520         if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1521                 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1522                 /*
1523                  * propagate the combined length to the end of the requests
1524                  */
1525                 if (req->nr_hw_segments == 1)
1526                         req->bio->bi_hw_front_size = len;
1527                 if (next->nr_hw_segments == 1)
1528                         next->biotail->bi_hw_back_size = len;
1529                 total_hw_segments--;
1530         }
1531
1532         if (total_hw_segments > q->max_hw_segments)
1533                 return 0;
1534
1535         /* Merge is OK... */
1536         req->nr_phys_segments = total_phys_segments;
1537         req->nr_hw_segments = total_hw_segments;
1538         return 1;
1539 }
1540
1541 /*
1542  * "plug" the device if there are no outstanding requests: this will
1543  * force the transfer to start only after we have put all the requests
1544  * on the list.
1545  *
1546  * This is called with interrupts off and no requests on the queue and
1547  * with the queue lock held.
1548  */
1549 void blk_plug_device(request_queue_t *q)
1550 {
1551         WARN_ON(!irqs_disabled());
1552
1553         /*
1554          * don't plug a stopped queue, it must be paired with blk_start_queue()
1555          * which will restart the queueing
1556          */
1557         if (blk_queue_stopped(q))
1558                 return;
1559
1560         if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
1561                 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1562                 blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
1563         }
1564 }
1565
1566 EXPORT_SYMBOL(blk_plug_device);
1567
1568 /*
1569  * remove the queue from the plugged list, if present. called with
1570  * queue lock held and interrupts disabled.
1571  */
1572 int blk_remove_plug(request_queue_t *q)
1573 {
1574         WARN_ON(!irqs_disabled());
1575
1576         if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1577                 return 0;
1578
1579         del_timer(&q->unplug_timer);
1580         return 1;
1581 }
1582
1583 EXPORT_SYMBOL(blk_remove_plug);
1584
1585 /*
1586  * remove the plug and let it rip..
1587  */
1588 void __generic_unplug_device(request_queue_t *q)
1589 {
1590         if (unlikely(blk_queue_stopped(q)))
1591                 return;
1592
1593         if (!blk_remove_plug(q))
1594                 return;
1595
1596         q->request_fn(q);
1597 }
1598 EXPORT_SYMBOL(__generic_unplug_device);
1599
1600 /**
1601  * generic_unplug_device - fire a request queue
1602  * @q:    The &request_queue_t in question
1603  *
1604  * Description:
1605  *   Linux uses plugging to build bigger requests queues before letting
1606  *   the device have at them. If a queue is plugged, the I/O scheduler
1607  *   is still adding and merging requests on the queue. Once the queue
1608  *   gets unplugged, the request_fn defined for the queue is invoked and
1609  *   transfers started.
1610  **/
1611 void generic_unplug_device(request_queue_t *q)
1612 {
1613         spin_lock_irq(q->queue_lock);
1614         __generic_unplug_device(q);
1615         spin_unlock_irq(q->queue_lock);
1616 }
1617 EXPORT_SYMBOL(generic_unplug_device);
1618
1619 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1620                                    struct page *page)
1621 {
1622         request_queue_t *q = bdi->unplug_io_data;
1623
1624         /*
1625          * devices don't necessarily have an ->unplug_fn defined
1626          */
1627         if (q->unplug_fn) {
1628                 blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1629                                         q->rq.count[READ] + q->rq.count[WRITE]);
1630
1631                 q->unplug_fn(q);
1632         }
1633 }
1634
1635 static void blk_unplug_work(void *data)
1636 {
1637         request_queue_t *q = data;
1638
1639         blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
1640                                 q->rq.count[READ] + q->rq.count[WRITE]);
1641
1642         q->unplug_fn(q);
1643 }
1644
1645 static void blk_unplug_timeout(unsigned long data)
1646 {
1647         request_queue_t *q = (request_queue_t *)data;
1648
1649         blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
1650                                 q->rq.count[READ] + q->rq.count[WRITE]);
1651
1652         kblockd_schedule_work(&q->unplug_work);
1653 }
1654
1655 /**
1656  * blk_start_queue - restart a previously stopped queue
1657  * @q:    The &request_queue_t in question
1658  *
1659  * Description:
1660  *   blk_start_queue() will clear the stop flag on the queue, and call
1661  *   the request_fn for the queue if it was in a stopped state when
1662  *   entered. Also see blk_stop_queue(). Queue lock must be held.
1663  **/
1664 void blk_start_queue(request_queue_t *q)
1665 {
1666         clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1667
1668         /*
1669          * one level of recursion is ok and is much faster than kicking
1670          * the unplug handling
1671          */
1672         if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1673                 q->request_fn(q);
1674                 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1675         } else {
1676                 blk_plug_device(q);
1677                 kblockd_schedule_work(&q->unplug_work);
1678         }
1679 }
1680
1681 EXPORT_SYMBOL(blk_start_queue);
1682
1683 /**
1684  * blk_stop_queue - stop a queue
1685  * @q:    The &request_queue_t in question
1686  *
1687  * Description:
1688  *   The Linux block layer assumes that a block driver will consume all
1689  *   entries on the request queue when the request_fn strategy is called.
1690  *   Often this will not happen, because of hardware limitations (queue
1691  *   depth settings). If a device driver gets a 'queue full' response,
1692  *   or if it simply chooses not to queue more I/O at one point, it can
1693  *   call this function to prevent the request_fn from being called until
1694  *   the driver has signalled it's ready to go again. This happens by calling
1695  *   blk_start_queue() to restart queue operations. Queue lock must be held.
1696  **/
1697 void blk_stop_queue(request_queue_t *q)
1698 {
1699         blk_remove_plug(q);
1700         set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1701 }
1702 EXPORT_SYMBOL(blk_stop_queue);
1703
1704 /**
1705  * blk_sync_queue - cancel any pending callbacks on a queue
1706  * @q: the queue
1707  *
1708  * Description:
1709  *     The block layer may perform asynchronous callback activity
1710  *     on a queue, such as calling the unplug function after a timeout.
1711  *     A block device may call blk_sync_queue to ensure that any
1712  *     such activity is cancelled, thus allowing it to release resources
1713  *     the the callbacks might use. The caller must already have made sure
1714  *     that its ->make_request_fn will not re-add plugging prior to calling
1715  *     this function.
1716  *
1717  */
1718 void blk_sync_queue(struct request_queue *q)
1719 {
1720         del_timer_sync(&q->unplug_timer);
1721         kblockd_flush();
1722 }
1723 EXPORT_SYMBOL(blk_sync_queue);
1724
1725 /**
1726  * blk_run_queue - run a single device queue
1727  * @q:  The queue to run
1728  */
1729 void blk_run_queue(struct request_queue *q)
1730 {
1731         unsigned long flags;
1732
1733         spin_lock_irqsave(q->queue_lock, flags);
1734         blk_remove_plug(q);
1735
1736         /*
1737          * Only recurse once to avoid overrunning the stack, let the unplug
1738          * handling reinvoke the handler shortly if we already got there.
1739          */
1740         if (!elv_queue_empty(q)) {
1741                 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1742                         q->request_fn(q);
1743                         clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1744                 } else {
1745                         blk_plug_device(q);
1746                         kblockd_schedule_work(&q->unplug_work);
1747                 }
1748         }
1749
1750         spin_unlock_irqrestore(q->queue_lock, flags);
1751 }
1752 EXPORT_SYMBOL(blk_run_queue);
1753
1754 /**
1755  * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1756  * @kobj:    the kobj belonging of the request queue to be released
1757  *
1758  * Description:
1759  *     blk_cleanup_queue is the pair to blk_init_queue() or
1760  *     blk_queue_make_request().  It should be called when a request queue is
1761  *     being released; typically when a block device is being de-registered.
1762  *     Currently, its primary task it to free all the &struct request
1763  *     structures that were allocated to the queue and the queue itself.
1764  *
1765  * Caveat:
1766  *     Hopefully the low level driver will have finished any
1767  *     outstanding requests first...
1768  **/
1769 static void blk_release_queue(struct kobject *kobj)
1770 {
1771         request_queue_t *q = container_of(kobj, struct request_queue, kobj);
1772         struct request_list *rl = &q->rq;
1773
1774         blk_sync_queue(q);
1775
1776         if (rl->rq_pool)
1777                 mempool_destroy(rl->rq_pool);
1778
1779         if (q->queue_tags)
1780                 __blk_queue_free_tags(q);
1781
1782         if (q->blk_trace)
1783                 blk_trace_shutdown(q);
1784
1785         kmem_cache_free(requestq_cachep, q);
1786 }
1787
1788 void blk_put_queue(request_queue_t *q)
1789 {
1790         kobject_put(&q->kobj);
1791 }
1792 EXPORT_SYMBOL(blk_put_queue);
1793
1794 void blk_cleanup_queue(request_queue_t * q)
1795 {
1796         mutex_lock(&q->sysfs_lock);
1797         set_bit(QUEUE_FLAG_DEAD, &q->queue_flags);
1798         mutex_unlock(&q->sysfs_lock);
1799
1800         if (q->elevator)
1801                 elevator_exit(q->elevator);
1802
1803         blk_put_queue(q);
1804 }
1805
1806 EXPORT_SYMBOL(blk_cleanup_queue);
1807
1808 static int blk_init_free_list(request_queue_t *q)
1809 {
1810         struct request_list *rl = &q->rq;
1811
1812         rl->count[READ] = rl->count[WRITE] = 0;
1813         rl->starved[READ] = rl->starved[WRITE] = 0;
1814         rl->elvpriv = 0;
1815         init_waitqueue_head(&rl->wait[READ]);
1816         init_waitqueue_head(&rl->wait[WRITE]);
1817
1818         rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1819                                 mempool_free_slab, request_cachep, q->node);
1820
1821         if (!rl->rq_pool)
1822                 return -ENOMEM;
1823
1824         return 0;
1825 }
1826
1827 request_queue_t *blk_alloc_queue(gfp_t gfp_mask)
1828 {
1829         return blk_alloc_queue_node(gfp_mask, -1);
1830 }
1831 EXPORT_SYMBOL(blk_alloc_queue);
1832
1833 static struct kobj_type queue_ktype;
1834
1835 request_queue_t *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
1836 {
1837         request_queue_t *q;
1838
1839         q = kmem_cache_alloc_node(requestq_cachep, gfp_mask, node_id);
1840         if (!q)
1841                 return NULL;
1842
1843         memset(q, 0, sizeof(*q));
1844         init_timer(&q->unplug_timer);
1845
1846         snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
1847         q->kobj.ktype = &queue_ktype;
1848         kobject_init(&q->kobj);
1849
1850         q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1851         q->backing_dev_info.unplug_io_data = q;
1852
1853         mutex_init(&q->sysfs_lock);
1854
1855         return q;
1856 }
1857 EXPORT_SYMBOL(blk_alloc_queue_node);
1858
1859 /**
1860  * blk_init_queue  - prepare a request queue for use with a block device
1861  * @rfn:  The function to be called to process requests that have been
1862  *        placed on the queue.
1863  * @lock: Request queue spin lock
1864  *
1865  * Description:
1866  *    If a block device wishes to use the standard request handling procedures,
1867  *    which sorts requests and coalesces adjacent requests, then it must
1868  *    call blk_init_queue().  The function @rfn will be called when there
1869  *    are requests on the queue that need to be processed.  If the device
1870  *    supports plugging, then @rfn may not be called immediately when requests
1871  *    are available on the queue, but may be called at some time later instead.
1872  *    Plugged queues are generally unplugged when a buffer belonging to one
1873  *    of the requests on the queue is needed, or due to memory pressure.
1874  *
1875  *    @rfn is not required, or even expected, to remove all requests off the
1876  *    queue, but only as many as it can handle at a time.  If it does leave
1877  *    requests on the queue, it is responsible for arranging that the requests
1878  *    get dealt with eventually.
1879  *
1880  *    The queue spin lock must be held while manipulating the requests on the
1881  *    request queue.
1882  *
1883  *    Function returns a pointer to the initialized request queue, or NULL if
1884  *    it didn't succeed.
1885  *
1886  * Note:
1887  *    blk_init_queue() must be paired with a blk_cleanup_queue() call
1888  *    when the block device is deactivated (such as at module unload).
1889  **/
1890
1891 request_queue_t *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1892 {
1893         return blk_init_queue_node(rfn, lock, -1);
1894 }
1895 EXPORT_SYMBOL(blk_init_queue);
1896
1897 request_queue_t *
1898 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1899 {
1900         request_queue_t *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1901
1902         if (!q)
1903                 return NULL;
1904
1905         q->node = node_id;
1906         if (blk_init_free_list(q)) {
1907                 kmem_cache_free(requestq_cachep, q);
1908                 return NULL;
1909         }
1910
1911         /*
1912          * if caller didn't supply a lock, they get per-queue locking with
1913          * our embedded lock
1914          */
1915         if (!lock) {
1916                 spin_lock_init(&q->__queue_lock);
1917                 lock = &q->__queue_lock;
1918         }
1919
1920         q->request_fn           = rfn;
1921         q->back_merge_fn        = ll_back_merge_fn;
1922         q->front_merge_fn       = ll_front_merge_fn;
1923         q->merge_requests_fn    = ll_merge_requests_fn;
1924         q->prep_rq_fn           = NULL;
1925         q->unplug_fn            = generic_unplug_device;
1926         q->queue_flags          = (1 << QUEUE_FLAG_CLUSTER);
1927         q->queue_lock           = lock;
1928
1929         blk_queue_segment_boundary(q, 0xffffffff);
1930
1931         blk_queue_make_request(q, __make_request);
1932         blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1933
1934         blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1935         blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1936
1937         /*
1938          * all done
1939          */
1940         if (!elevator_init(q, NULL)) {
1941                 blk_queue_congestion_threshold(q);
1942                 return q;
1943         }
1944
1945         blk_put_queue(q);
1946         return NULL;
1947 }
1948 EXPORT_SYMBOL(blk_init_queue_node);
1949
1950 int blk_get_queue(request_queue_t *q)
1951 {
1952         if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1953                 kobject_get(&q->kobj);
1954                 return 0;
1955         }
1956
1957         return 1;
1958 }
1959
1960 EXPORT_SYMBOL(blk_get_queue);
1961
1962 static inline void blk_free_request(request_queue_t *q, struct request *rq)
1963 {
1964         if (rq->flags & REQ_ELVPRIV)
1965                 elv_put_request(q, rq);
1966         mempool_free(rq, q->rq.rq_pool);
1967 }
1968
1969 static inline struct request *
1970 blk_alloc_request(request_queue_t *q, int rw, struct bio *bio,
1971                   int priv, gfp_t gfp_mask)
1972 {
1973         struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1974
1975         if (!rq)
1976                 return NULL;
1977
1978         /*
1979          * first three bits are identical in rq->flags and bio->bi_rw,
1980          * see bio.h and blkdev.h
1981          */
1982         rq->flags = rw;
1983
1984         if (priv) {
1985                 if (unlikely(elv_set_request(q, rq, bio, gfp_mask))) {
1986                         mempool_free(rq, q->rq.rq_pool);
1987                         return NULL;
1988                 }
1989                 rq->flags |= REQ_ELVPRIV;
1990         }
1991
1992         return rq;
1993 }
1994
1995 /*
1996  * ioc_batching returns true if the ioc is a valid batching request and
1997  * should be given priority access to a request.
1998  */
1999 static inline int ioc_batching(request_queue_t *q, struct io_context *ioc)
2000 {
2001         if (!ioc)
2002                 return 0;
2003
2004         /*
2005          * Make sure the process is able to allocate at least 1 request
2006          * even if the batch times out, otherwise we could theoretically
2007          * lose wakeups.
2008          */
2009         return ioc->nr_batch_requests == q->nr_batching ||
2010                 (ioc->nr_batch_requests > 0
2011                 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
2012 }
2013
2014 /*
2015  * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
2016  * will cause the process to be a "batcher" on all queues in the system. This
2017  * is the behaviour we want though - once it gets a wakeup it should be given
2018  * a nice run.
2019  */
2020 static void ioc_set_batching(request_queue_t *q, struct io_context *ioc)
2021 {
2022         if (!ioc || ioc_batching(q, ioc))
2023                 return;
2024
2025         ioc->nr_batch_requests = q->nr_batching;
2026         ioc->last_waited = jiffies;
2027 }
2028
2029 static void __freed_request(request_queue_t *q, int rw)
2030 {
2031         struct request_list *rl = &q->rq;
2032
2033         if (rl->count[rw] < queue_congestion_off_threshold(q))
2034                 clear_queue_congested(q, rw);
2035
2036         if (rl->count[rw] + 1 <= q->nr_requests) {
2037                 if (waitqueue_active(&rl->wait[rw]))
2038                         wake_up(&rl->wait[rw]);
2039
2040                 blk_clear_queue_full(q, rw);
2041         }
2042 }
2043
2044 /*
2045  * A request has just been released.  Account for it, update the full and
2046  * congestion status, wake up any waiters.   Called under q->queue_lock.
2047  */
2048 static void freed_request(request_queue_t *q, int rw, int priv)
2049 {
2050         struct request_list *rl = &q->rq;
2051
2052         rl->count[rw]--;
2053         if (priv)
2054                 rl->elvpriv--;
2055
2056         __freed_request(q, rw);
2057
2058         if (unlikely(rl->starved[rw ^ 1]))
2059                 __freed_request(q, rw ^ 1);
2060 }
2061
2062 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
2063 /*
2064  * Get a free request, queue_lock must be held.
2065  * Returns NULL on failure, with queue_lock held.
2066  * Returns !NULL on success, with queue_lock *not held*.
2067  */
2068 static struct request *get_request(request_queue_t *q, int rw, struct bio *bio,
2069                                    gfp_t gfp_mask)
2070 {
2071         struct request *rq = NULL;
2072         struct request_list *rl = &q->rq;
2073         struct io_context *ioc = NULL;
2074         int may_queue, priv;
2075
2076         may_queue = elv_may_queue(q, rw, bio);
2077         if (may_queue == ELV_MQUEUE_NO)
2078                 goto rq_starved;
2079
2080         if (rl->count[rw]+1 >= queue_congestion_on_threshold(q)) {
2081                 if (rl->count[rw]+1 >= q->nr_requests) {
2082                         ioc = current_io_context(GFP_ATOMIC);
2083                         /*
2084                          * The queue will fill after this allocation, so set
2085                          * it as full, and mark this process as "batching".
2086                          * This process will be allowed to complete a batch of
2087                          * requests, others will be blocked.
2088                          */
2089                         if (!blk_queue_full(q, rw)) {
2090                                 ioc_set_batching(q, ioc);
2091                                 blk_set_queue_full(q, rw);
2092                         } else {
2093                                 if (may_queue != ELV_MQUEUE_MUST
2094                                                 && !ioc_batching(q, ioc)) {
2095                                         /*
2096                                          * The queue is full and the allocating
2097                                          * process is not a "batcher", and not
2098                                          * exempted by the IO scheduler
2099                                          */
2100                                         goto out;
2101                                 }
2102                         }
2103                 }
2104                 set_queue_congested(q, rw);
2105         }
2106
2107         /*
2108          * Only allow batching queuers to allocate up to 50% over the defined
2109          * limit of requests, otherwise we could have thousands of requests
2110          * allocated with any setting of ->nr_requests
2111          */
2112         if (rl->count[rw] >= (3 * q->nr_requests / 2))
2113                 goto out;
2114
2115         rl->count[rw]++;
2116         rl->starved[rw] = 0;
2117
2118         priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
2119         if (priv)
2120                 rl->elvpriv++;
2121
2122         spin_unlock_irq(q->queue_lock);
2123
2124         rq = blk_alloc_request(q, rw, bio, priv, gfp_mask);
2125         if (unlikely(!rq)) {
2126                 /*
2127                  * Allocation failed presumably due to memory. Undo anything
2128                  * we might have messed up.
2129                  *
2130                  * Allocating task should really be put onto the front of the
2131                  * wait queue, but this is pretty rare.
2132                  */
2133                 spin_lock_irq(q->queue_lock);
2134                 freed_request(q, rw, priv);
2135
2136                 /*
2137                  * in the very unlikely event that allocation failed and no
2138                  * requests for this direction was pending, mark us starved
2139                  * so that freeing of a request in the other direction will
2140                  * notice us. another possible fix would be to split the
2141                  * rq mempool into READ and WRITE
2142                  */
2143 rq_starved:
2144                 if (unlikely(rl->count[rw] == 0))
2145                         rl->starved[rw] = 1;
2146
2147                 goto out;
2148         }
2149
2150         /*
2151          * ioc may be NULL here, and ioc_batching will be false. That's
2152          * OK, if the queue is under the request limit then requests need
2153          * not count toward the nr_batch_requests limit. There will always
2154          * be some limit enforced by BLK_BATCH_TIME.
2155          */
2156         if (ioc_batching(q, ioc))
2157                 ioc->nr_batch_requests--;
2158         
2159         rq_init(q, rq);
2160         rq->rl = rl;
2161
2162         blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
2163 out:
2164         return rq;
2165 }
2166
2167 /*
2168  * No available requests for this queue, unplug the device and wait for some
2169  * requests to become available.
2170  *
2171  * Called with q->queue_lock held, and returns with it unlocked.
2172  */
2173 static struct request *get_request_wait(request_queue_t *q, int rw,
2174                                         struct bio *bio)
2175 {
2176         struct request *rq;
2177
2178         rq = get_request(q, rw, bio, GFP_NOIO);
2179         while (!rq) {
2180                 DEFINE_WAIT(wait);
2181                 struct request_list *rl = &q->rq;
2182
2183                 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2184                                 TASK_UNINTERRUPTIBLE);
2185
2186                 rq = get_request(q, rw, bio, GFP_NOIO);
2187
2188                 if (!rq) {
2189                         struct io_context *ioc;
2190
2191                         blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
2192
2193                         __generic_unplug_device(q);
2194                         spin_unlock_irq(q->queue_lock);
2195                         io_schedule();
2196
2197                         /*
2198                          * After sleeping, we become a "batching" process and
2199                          * will be able to allocate at least one request, and
2200                          * up to a big batch of them for a small period time.
2201                          * See ioc_batching, ioc_set_batching
2202                          */
2203                         ioc = current_io_context(GFP_NOIO);
2204                         ioc_set_batching(q, ioc);
2205
2206                         spin_lock_irq(q->queue_lock);
2207                 }
2208                 finish_wait(&rl->wait[rw], &wait);
2209         }
2210
2211         return rq;
2212 }
2213
2214 struct request *blk_get_request(request_queue_t *q, int rw, gfp_t gfp_mask)
2215 {
2216         struct request *rq;
2217
2218         BUG_ON(rw != READ && rw != WRITE);
2219
2220         spin_lock_irq(q->queue_lock);
2221         if (gfp_mask & __GFP_WAIT) {
2222                 rq = get_request_wait(q, rw, NULL);
2223         } else {
2224                 rq = get_request(q, rw, NULL, gfp_mask);
2225                 if (!rq)
2226                         spin_unlock_irq(q->queue_lock);
2227         }
2228         /* q->queue_lock is unlocked at this point */
2229
2230         return rq;
2231 }
2232 EXPORT_SYMBOL(blk_get_request);
2233
2234 /**
2235  * blk_requeue_request - put a request back on queue
2236  * @q:          request queue where request should be inserted
2237  * @rq:         request to be inserted
2238  *
2239  * Description:
2240  *    Drivers often keep queueing requests until the hardware cannot accept
2241  *    more, when that condition happens we need to put the request back
2242  *    on the queue. Must be called with queue lock held.
2243  */
2244 void blk_requeue_request(request_queue_t *q, struct request *rq)
2245 {
2246         blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
2247
2248         if (blk_rq_tagged(rq))
2249                 blk_queue_end_tag(q, rq);
2250
2251         elv_requeue_request(q, rq);
2252 }
2253
2254 EXPORT_SYMBOL(blk_requeue_request);
2255
2256 /**
2257  * blk_insert_request - insert a special request in to a request queue
2258  * @q:          request queue where request should be inserted
2259  * @rq:         request to be inserted
2260  * @at_head:    insert request at head or tail of queue
2261  * @data:       private data
2262  *
2263  * Description:
2264  *    Many block devices need to execute commands asynchronously, so they don't
2265  *    block the whole kernel from preemption during request execution.  This is
2266  *    accomplished normally by inserting aritficial requests tagged as
2267  *    REQ_SPECIAL in to the corresponding request queue, and letting them be
2268  *    scheduled for actual execution by the request queue.
2269  *
2270  *    We have the option of inserting the head or the tail of the queue.
2271  *    Typically we use the tail for new ioctls and so forth.  We use the head
2272  *    of the queue for things like a QUEUE_FULL message from a device, or a
2273  *    host that is unable to accept a particular command.
2274  */
2275 void blk_insert_request(request_queue_t *q, struct request *rq,
2276                         int at_head, void *data)
2277 {
2278         int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2279         unsigned long flags;
2280
2281         /*
2282          * tell I/O scheduler that this isn't a regular read/write (ie it
2283          * must not attempt merges on this) and that it acts as a soft
2284          * barrier
2285          */
2286         rq->flags |= REQ_SPECIAL | REQ_SOFTBARRIER;
2287
2288         rq->special = data;
2289
2290         spin_lock_irqsave(q->queue_lock, flags);
2291
2292         /*
2293          * If command is tagged, release the tag
2294          */
2295         if (blk_rq_tagged(rq))
2296                 blk_queue_end_tag(q, rq);
2297
2298         drive_stat_acct(rq, rq->nr_sectors, 1);
2299         __elv_add_request(q, rq, where, 0);
2300
2301         if (blk_queue_plugged(q))
2302                 __generic_unplug_device(q);
2303         else
2304                 q->request_fn(q);
2305         spin_unlock_irqrestore(q->queue_lock, flags);
2306 }
2307
2308 EXPORT_SYMBOL(blk_insert_request);
2309
2310 /**
2311  * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2312  * @q:          request queue where request should be inserted
2313  * @rq:         request structure to fill
2314  * @ubuf:       the user buffer
2315  * @len:        length of user data
2316  *
2317  * Description:
2318  *    Data will be mapped directly for zero copy io, if possible. Otherwise
2319  *    a kernel bounce buffer is used.
2320  *
2321  *    A matching blk_rq_unmap_user() must be issued at the end of io, while
2322  *    still in process context.
2323  *
2324  *    Note: The mapped bio may need to be bounced through blk_queue_bounce()
2325  *    before being submitted to the device, as pages mapped may be out of
2326  *    reach. It's the callers responsibility to make sure this happens. The
2327  *    original bio must be passed back in to blk_rq_unmap_user() for proper
2328  *    unmapping.
2329  */
2330 int blk_rq_map_user(request_queue_t *q, struct request *rq, void __user *ubuf,
2331                     unsigned int len)
2332 {
2333         unsigned long uaddr;
2334         struct bio *bio;
2335         int reading;
2336
2337         if (len > (q->max_hw_sectors << 9))
2338                 return -EINVAL;
2339         if (!len || !ubuf)
2340                 return -EINVAL;
2341
2342         reading = rq_data_dir(rq) == READ;
2343
2344         /*
2345          * if alignment requirement is satisfied, map in user pages for
2346          * direct dma. else, set up kernel bounce buffers
2347          */
2348         uaddr = (unsigned long) ubuf;
2349         if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2350                 bio = bio_map_user(q, NULL, uaddr, len, reading);
2351         else
2352                 bio = bio_copy_user(q, uaddr, len, reading);
2353
2354         if (!IS_ERR(bio)) {
2355                 rq->bio = rq->biotail = bio;
2356                 blk_rq_bio_prep(q, rq, bio);
2357
2358                 rq->buffer = rq->data = NULL;
2359                 rq->data_len = len;
2360                 return 0;
2361         }
2362
2363         /*
2364          * bio is the err-ptr
2365          */
2366         return PTR_ERR(bio);
2367 }
2368
2369 EXPORT_SYMBOL(blk_rq_map_user);
2370
2371 /**
2372  * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2373  * @q:          request queue where request should be inserted
2374  * @rq:         request to map data to
2375  * @iov:        pointer to the iovec
2376  * @iov_count:  number of elements in the iovec
2377  *
2378  * Description:
2379  *    Data will be mapped directly for zero copy io, if possible. Otherwise
2380  *    a kernel bounce buffer is used.
2381  *
2382  *    A matching blk_rq_unmap_user() must be issued at the end of io, while
2383  *    still in process context.
2384  *
2385  *    Note: The mapped bio may need to be bounced through blk_queue_bounce()
2386  *    before being submitted to the device, as pages mapped may be out of
2387  *    reach. It's the callers responsibility to make sure this happens. The
2388  *    original bio must be passed back in to blk_rq_unmap_user() for proper
2389  *    unmapping.
2390  */
2391 int blk_rq_map_user_iov(request_queue_t *q, struct request *rq,
2392                         struct sg_iovec *iov, int iov_count)
2393 {
2394         struct bio *bio;
2395
2396         if (!iov || iov_count <= 0)
2397                 return -EINVAL;
2398
2399         /* we don't allow misaligned data like bio_map_user() does.  If the
2400          * user is using sg, they're expected to know the alignment constraints
2401          * and respect them accordingly */
2402         bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2403         if (IS_ERR(bio))
2404                 return PTR_ERR(bio);
2405
2406         rq->bio = rq->biotail = bio;
2407         blk_rq_bio_prep(q, rq, bio);
2408         rq->buffer = rq->data = NULL;
2409         rq->data_len = bio->bi_size;
2410         return 0;
2411 }
2412
2413 EXPORT_SYMBOL(blk_rq_map_user_iov);
2414
2415 /**
2416  * blk_rq_unmap_user - unmap a request with user data
2417  * @bio:        bio to be unmapped
2418  * @ulen:       length of user buffer
2419  *
2420  * Description:
2421  *    Unmap a bio previously mapped by blk_rq_map_user().
2422  */
2423 int blk_rq_unmap_user(struct bio *bio, unsigned int ulen)
2424 {
2425         int ret = 0;
2426
2427         if (bio) {
2428                 if (bio_flagged(bio, BIO_USER_MAPPED))
2429                         bio_unmap_user(bio);
2430                 else
2431                         ret = bio_uncopy_user(bio);
2432         }
2433
2434         return 0;
2435 }
2436
2437 EXPORT_SYMBOL(blk_rq_unmap_user);
2438
2439 /**
2440  * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2441  * @q:          request queue where request should be inserted
2442  * @rq:         request to fill
2443  * @kbuf:       the kernel buffer
2444  * @len:        length of user data
2445  * @gfp_mask:   memory allocation flags
2446  */
2447 int blk_rq_map_kern(request_queue_t *q, struct request *rq, void *kbuf,
2448                     unsigned int len, gfp_t gfp_mask)
2449 {
2450         struct bio *bio;
2451
2452         if (len > (q->max_hw_sectors << 9))
2453                 return -EINVAL;
2454         if (!len || !kbuf)
2455                 return -EINVAL;
2456
2457         bio = bio_map_kern(q, kbuf, len, gfp_mask);
2458         if (IS_ERR(bio))
2459                 return PTR_ERR(bio);
2460
2461         if (rq_data_dir(rq) == WRITE)
2462                 bio->bi_rw |= (1 << BIO_RW);
2463
2464         rq->bio = rq->biotail = bio;
2465         blk_rq_bio_prep(q, rq, bio);
2466
2467         rq->buffer = rq->data = NULL;
2468         rq->data_len = len;
2469         return 0;
2470 }
2471
2472 EXPORT_SYMBOL(blk_rq_map_kern);
2473
2474 /**
2475  * blk_execute_rq_nowait - insert a request into queue for execution
2476  * @q:          queue to insert the request in
2477  * @bd_disk:    matching gendisk
2478  * @rq:         request to insert
2479  * @at_head:    insert request at head or tail of queue
2480  * @done:       I/O completion handler
2481  *
2482  * Description:
2483  *    Insert a fully prepared request at the back of the io scheduler queue
2484  *    for execution.  Don't wait for completion.
2485  */
2486 void blk_execute_rq_nowait(request_queue_t *q, struct gendisk *bd_disk,
2487                            struct request *rq, int at_head,
2488                            rq_end_io_fn *done)
2489 {
2490         int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2491
2492         rq->rq_disk = bd_disk;
2493         rq->flags |= REQ_NOMERGE;
2494         rq->end_io = done;
2495         WARN_ON(irqs_disabled());
2496         spin_lock_irq(q->queue_lock);
2497         __elv_add_request(q, rq, where, 1);
2498         __generic_unplug_device(q);
2499         spin_unlock_irq(q->queue_lock);
2500 }
2501 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
2502
2503 /**
2504  * blk_execute_rq - insert a request into queue for execution
2505  * @q:          queue to insert the request in
2506  * @bd_disk:    matching gendisk
2507  * @rq:         request to insert
2508  * @at_head:    insert request at head or tail of queue
2509  *
2510  * Description:
2511  *    Insert a fully prepared request at the back of the io scheduler queue
2512  *    for execution and wait for completion.
2513  */
2514 int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
2515                    struct request *rq, int at_head)
2516 {
2517         DECLARE_COMPLETION(wait);
2518         char sense[SCSI_SENSE_BUFFERSIZE];
2519         int err = 0;
2520
2521         /*
2522          * we need an extra reference to the request, so we can look at
2523          * it after io completion
2524          */
2525         rq->ref_count++;
2526
2527         if (!rq->sense) {
2528                 memset(sense, 0, sizeof(sense));
2529                 rq->sense = sense;
2530                 rq->sense_len = 0;
2531         }
2532
2533         rq->waiting = &wait;
2534         blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
2535         wait_for_completion(&wait);
2536         rq->waiting = NULL;
2537
2538         if (rq->errors)
2539                 err = -EIO;
2540
2541         return err;
2542 }
2543
2544 EXPORT_SYMBOL(blk_execute_rq);
2545
2546 /**
2547  * blkdev_issue_flush - queue a flush
2548  * @bdev:       blockdev to issue flush for
2549  * @error_sector:       error sector
2550  *
2551  * Description:
2552  *    Issue a flush for the block device in question. Caller can supply
2553  *    room for storing the error offset in case of a flush error, if they
2554  *    wish to.  Caller must run wait_for_completion() on its own.
2555  */
2556 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2557 {
2558         request_queue_t *q;
2559
2560         if (bdev->bd_disk == NULL)
2561                 return -ENXIO;
2562
2563         q = bdev_get_queue(bdev);
2564         if (!q)
2565                 return -ENXIO;
2566         if (!q->issue_flush_fn)
2567                 return -EOPNOTSUPP;
2568
2569         return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2570 }
2571
2572 EXPORT_SYMBOL(blkdev_issue_flush);
2573
2574 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2575 {
2576         int rw = rq_data_dir(rq);
2577
2578         if (!blk_fs_request(rq) || !rq->rq_disk)
2579                 return;
2580
2581         if (!new_io) {
2582                 __disk_stat_inc(rq->rq_disk, merges[rw]);
2583         } else {
2584                 disk_round_stats(rq->rq_disk);
2585                 rq->rq_disk->in_flight++;
2586         }
2587 }
2588
2589 /*
2590  * add-request adds a request to the linked list.
2591  * queue lock is held and interrupts disabled, as we muck with the
2592  * request queue list.
2593  */
2594 static inline void add_request(request_queue_t * q, struct request * req)
2595 {
2596         drive_stat_acct(req, req->nr_sectors, 1);
2597
2598         if (q->activity_fn)
2599                 q->activity_fn(q->activity_data, rq_data_dir(req));
2600
2601         /*
2602          * elevator indicated where it wants this request to be
2603          * inserted at elevator_merge time
2604          */
2605         __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2606 }
2607  
2608 /*
2609  * disk_round_stats()   - Round off the performance stats on a struct
2610  * disk_stats.
2611  *
2612  * The average IO queue length and utilisation statistics are maintained
2613  * by observing the current state of the queue length and the amount of
2614  * time it has been in this state for.
2615  *
2616  * Normally, that accounting is done on IO completion, but that can result
2617  * in more than a second's worth of IO being accounted for within any one
2618  * second, leading to >100% utilisation.  To deal with that, we call this
2619  * function to do a round-off before returning the results when reading
2620  * /proc/diskstats.  This accounts immediately for all queue usage up to
2621  * the current jiffies and restarts the counters again.
2622  */
2623 void disk_round_stats(struct gendisk *disk)
2624 {
2625         unsigned long now = jiffies;
2626
2627         if (now == disk->stamp)
2628                 return;
2629
2630         if (disk->in_flight) {
2631                 __disk_stat_add(disk, time_in_queue,
2632                                 disk->in_flight * (now - disk->stamp));
2633                 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2634         }
2635         disk->stamp = now;
2636 }
2637
2638 EXPORT_SYMBOL_GPL(disk_round_stats);
2639
2640 /*
2641  * queue lock must be held
2642  */
2643 void __blk_put_request(request_queue_t *q, struct request *req)
2644 {
2645         struct request_list *rl = req->rl;
2646
2647         if (unlikely(!q))
2648                 return;
2649         if (unlikely(--req->ref_count))
2650                 return;
2651
2652         elv_completed_request(q, req);
2653
2654         req->rq_status = RQ_INACTIVE;
2655         req->rl = NULL;
2656
2657         /*
2658          * Request may not have originated from ll_rw_blk. if not,
2659          * it didn't come out of our reserved rq pools
2660          */
2661         if (rl) {
2662                 int rw = rq_data_dir(req);
2663                 int priv = req->flags & REQ_ELVPRIV;
2664
2665                 BUG_ON(!list_empty(&req->queuelist));
2666
2667                 blk_free_request(q, req);
2668                 freed_request(q, rw, priv);
2669         }
2670 }
2671
2672 EXPORT_SYMBOL_GPL(__blk_put_request);
2673
2674 void blk_put_request(struct request *req)
2675 {
2676         unsigned long flags;
2677         request_queue_t *q = req->q;
2678
2679         /*
2680          * Gee, IDE calls in w/ NULL q.  Fix IDE and remove the
2681          * following if (q) test.
2682          */
2683         if (q) {
2684                 spin_lock_irqsave(q->queue_lock, flags);
2685                 __blk_put_request(q, req);
2686                 spin_unlock_irqrestore(q->queue_lock, flags);
2687         }
2688 }
2689
2690 EXPORT_SYMBOL(blk_put_request);
2691
2692 /**
2693  * blk_end_sync_rq - executes a completion event on a request
2694  * @rq: request to complete
2695  * @error: end io status of the request
2696  */
2697 void blk_end_sync_rq(struct request *rq, int error)
2698 {
2699         struct completion *waiting = rq->waiting;
2700
2701         rq->waiting = NULL;
2702         __blk_put_request(rq->q, rq);
2703
2704         /*
2705          * complete last, if this is a stack request the process (and thus
2706          * the rq pointer) could be invalid right after this complete()
2707          */
2708         complete(waiting);
2709 }
2710 EXPORT_SYMBOL(blk_end_sync_rq);
2711
2712 /**
2713  * blk_congestion_wait - wait for a queue to become uncongested
2714  * @rw: READ or WRITE
2715  * @timeout: timeout in jiffies
2716  *
2717  * Waits for up to @timeout jiffies for a queue (any queue) to exit congestion.
2718  * If no queues are congested then just wait for the next request to be
2719  * returned.
2720  */
2721 long blk_congestion_wait(int rw, long timeout)
2722 {
2723         long ret;
2724         DEFINE_WAIT(wait);
2725         wait_queue_head_t *wqh = &congestion_wqh[rw];
2726
2727         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
2728         ret = io_schedule_timeout(timeout);
2729         finish_wait(wqh, &wait);
2730         return ret;
2731 }
2732
2733 EXPORT_SYMBOL(blk_congestion_wait);
2734
2735 /*
2736  * Has to be called with the request spinlock acquired
2737  */
2738 static int attempt_merge(request_queue_t *q, struct request *req,
2739                           struct request *next)
2740 {
2741         if (!rq_mergeable(req) || !rq_mergeable(next))
2742                 return 0;
2743
2744         /*
2745          * not contigious
2746          */
2747         if (req->sector + req->nr_sectors != next->sector)
2748                 return 0;
2749
2750         if (rq_data_dir(req) != rq_data_dir(next)
2751             || req->rq_disk != next->rq_disk
2752             || next->waiting || next->special)
2753                 return 0;
2754
2755         /*
2756          * If we are allowed to merge, then append bio list
2757          * from next to rq and release next. merge_requests_fn
2758          * will have updated segment counts, update sector
2759          * counts here.
2760          */
2761         if (!q->merge_requests_fn(q, req, next))
2762                 return 0;
2763
2764         /*
2765          * At this point we have either done a back merge
2766          * or front merge. We need the smaller start_time of
2767          * the merged requests to be the current request
2768          * for accounting purposes.
2769          */
2770         if (time_after(req->start_time, next->start_time))
2771                 req->start_time = next->start_time;
2772
2773         req->biotail->bi_next = next->bio;
2774         req->biotail = next->biotail;
2775
2776         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2777
2778         elv_merge_requests(q, req, next);
2779
2780         if (req->rq_disk) {
2781                 disk_round_stats(req->rq_disk);
2782                 req->rq_disk->in_flight--;
2783         }
2784
2785         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2786
2787         __blk_put_request(q, next);
2788         return 1;
2789 }
2790
2791 static inline int attempt_back_merge(request_queue_t *q, struct request *rq)
2792 {
2793         struct request *next = elv_latter_request(q, rq);
2794
2795         if (next)
2796                 return attempt_merge(q, rq, next);
2797
2798         return 0;
2799 }
2800
2801 static inline int attempt_front_merge(request_queue_t *q, struct request *rq)
2802 {
2803         struct request *prev = elv_former_request(q, rq);
2804
2805         if (prev)
2806                 return attempt_merge(q, prev, rq);
2807
2808         return 0;
2809 }
2810
2811 static void init_request_from_bio(struct request *req, struct bio *bio)
2812 {
2813         req->flags |= REQ_CMD;
2814
2815         /*
2816          * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2817          */
2818         if (bio_rw_ahead(bio) || bio_failfast(bio))
2819                 req->flags |= REQ_FAILFAST;
2820
2821         /*
2822          * REQ_BARRIER implies no merging, but lets make it explicit
2823          */
2824         if (unlikely(bio_barrier(bio)))
2825                 req->flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2826
2827         req->errors = 0;
2828         req->hard_sector = req->sector = bio->bi_sector;
2829         req->hard_nr_sectors = req->nr_sectors = bio_sectors(bio);
2830         req->current_nr_sectors = req->hard_cur_sectors = bio_cur_sectors(bio);
2831         req->nr_phys_segments = bio_phys_segments(req->q, bio);
2832         req->nr_hw_segments = bio_hw_segments(req->q, bio);
2833         req->buffer = bio_data(bio);    /* see ->buffer comment above */
2834         req->waiting = NULL;
2835         req->bio = req->biotail = bio;
2836         req->ioprio = bio_prio(bio);
2837         req->rq_disk = bio->bi_bdev->bd_disk;
2838         req->start_time = jiffies;
2839 }
2840
2841 static int __make_request(request_queue_t *q, struct bio *bio)
2842 {
2843         struct request *req;
2844         int el_ret, rw, nr_sectors, cur_nr_sectors, barrier, err, sync;
2845         unsigned short prio;
2846         sector_t sector;
2847
2848         sector = bio->bi_sector;
2849         nr_sectors = bio_sectors(bio);
2850         cur_nr_sectors = bio_cur_sectors(bio);
2851         prio = bio_prio(bio);
2852
2853         rw = bio_data_dir(bio);
2854         sync = bio_sync(bio);
2855
2856         /*
2857          * low level driver can indicate that it wants pages above a
2858          * certain limit bounced to low memory (ie for highmem, or even
2859          * ISA dma in theory)
2860          */
2861         blk_queue_bounce(q, &bio);
2862
2863         spin_lock_prefetch(q->queue_lock);
2864
2865         barrier = bio_barrier(bio);
2866         if (unlikely(barrier) && (q->next_ordered == QUEUE_ORDERED_NONE)) {
2867                 err = -EOPNOTSUPP;
2868                 goto end_io;
2869         }
2870
2871         spin_lock_irq(q->queue_lock);
2872
2873         if (unlikely(barrier) || elv_queue_empty(q))
2874                 goto get_rq;
2875
2876         el_ret = elv_merge(q, &req, bio);
2877         switch (el_ret) {
2878                 case ELEVATOR_BACK_MERGE:
2879                         BUG_ON(!rq_mergeable(req));
2880
2881                         if (!q->back_merge_fn(q, req, bio))
2882                                 break;
2883
2884                         blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
2885
2886                         req->biotail->bi_next = bio;
2887                         req->biotail = bio;
2888                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2889                         req->ioprio = ioprio_best(req->ioprio, prio);
2890                         drive_stat_acct(req, nr_sectors, 0);
2891                         if (!attempt_back_merge(q, req))
2892                                 elv_merged_request(q, req);
2893                         goto out;
2894
2895                 case ELEVATOR_FRONT_MERGE:
2896                         BUG_ON(!rq_mergeable(req));
2897
2898                         if (!q->front_merge_fn(q, req, bio))
2899                                 break;
2900
2901                         blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
2902
2903                         bio->bi_next = req->bio;
2904                         req->bio = bio;
2905
2906                         /*
2907                          * may not be valid. if the low level driver said
2908                          * it didn't need a bounce buffer then it better
2909                          * not touch req->buffer either...
2910                          */
2911                         req->buffer = bio_data(bio);
2912                         req->current_nr_sectors = cur_nr_sectors;
2913                         req->hard_cur_sectors = cur_nr_sectors;
2914                         req->sector = req->hard_sector = sector;
2915                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2916                         req->ioprio = ioprio_best(req->ioprio, prio);
2917                         drive_stat_acct(req, nr_sectors, 0);
2918                         if (!attempt_front_merge(q, req))
2919                                 elv_merged_request(q, req);
2920                         goto out;
2921
2922                 /* ELV_NO_MERGE: elevator says don't/can't merge. */
2923                 default:
2924                         ;
2925         }
2926
2927 get_rq:
2928         /*
2929          * Grab a free request. This is might sleep but can not fail.
2930          * Returns with the queue unlocked.
2931          */
2932         req = get_request_wait(q, rw, bio);
2933
2934         /*
2935          * After dropping the lock and possibly sleeping here, our request
2936          * may now be mergeable after it had proven unmergeable (above).
2937          * We don't worry about that case for efficiency. It won't happen
2938          * often, and the elevators are able to handle it.
2939          */
2940         init_request_from_bio(req, bio);
2941
2942         spin_lock_irq(q->queue_lock);
2943         if (elv_queue_empty(q))
2944                 blk_plug_device(q);
2945         add_request(q, req);
2946 out:
2947         if (sync)
2948                 __generic_unplug_device(q);
2949
2950         spin_unlock_irq(q->queue_lock);
2951         return 0;
2952
2953 end_io:
2954         bio_endio(bio, nr_sectors << 9, err);
2955         return 0;
2956 }
2957
2958 /*
2959  * If bio->bi_dev is a partition, remap the location
2960  */
2961 static inline void blk_partition_remap(struct bio *bio)
2962 {
2963         struct block_device *bdev = bio->bi_bdev;
2964
2965         if (bdev != bdev->bd_contains) {
2966                 struct hd_struct *p = bdev->bd_part;
2967                 const int rw = bio_data_dir(bio);
2968
2969                 p->sectors[rw] += bio_sectors(bio);
2970                 p->ios[rw]++;
2971
2972                 bio->bi_sector += p->start_sect;
2973                 bio->bi_bdev = bdev->bd_contains;
2974         }
2975 }
2976
2977 static void handle_bad_sector(struct bio *bio)
2978 {
2979         char b[BDEVNAME_SIZE];
2980
2981         printk(KERN_INFO "attempt to access beyond end of device\n");
2982         printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
2983                         bdevname(bio->bi_bdev, b),
2984                         bio->bi_rw,
2985                         (unsigned long long)bio->bi_sector + bio_sectors(bio),
2986                         (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
2987
2988         set_bit(BIO_EOF, &bio->bi_flags);
2989 }
2990
2991 /**
2992  * generic_make_request: hand a buffer to its device driver for I/O
2993  * @bio:  The bio describing the location in memory and on the device.
2994  *
2995  * generic_make_request() is used to make I/O requests of block
2996  * devices. It is passed a &struct bio, which describes the I/O that needs
2997  * to be done.
2998  *
2999  * generic_make_request() does not return any status.  The
3000  * success/failure status of the request, along with notification of
3001  * completion, is delivered asynchronously through the bio->bi_end_io
3002  * function described (one day) else where.
3003  *
3004  * The caller of generic_make_request must make sure that bi_io_vec
3005  * are set to describe the memory buffer, and that bi_dev and bi_sector are
3006  * set to describe the device address, and the
3007  * bi_end_io and optionally bi_private are set to describe how
3008  * completion notification should be signaled.
3009  *
3010  * generic_make_request and the drivers it calls may use bi_next if this
3011  * bio happens to be merged with someone else, and may change bi_dev and
3012  * bi_sector for remaps as it sees fit.  So the values of these fields
3013  * should NOT be depended on after the call to generic_make_request.
3014  */
3015 void generic_make_request(struct bio *bio)
3016 {
3017         request_queue_t *q;
3018         sector_t maxsector;
3019         int ret, nr_sectors = bio_sectors(bio);
3020         dev_t old_dev;
3021
3022         might_sleep();
3023         /* Test device or partition size, when known. */
3024         maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
3025         if (maxsector) {
3026                 sector_t sector = bio->bi_sector;
3027
3028                 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
3029                         /*
3030                          * This may well happen - the kernel calls bread()
3031                          * without checking the size of the device, e.g., when
3032                          * mounting a device.
3033                          */
3034                         handle_bad_sector(bio);
3035                         goto end_io;
3036                 }
3037         }
3038
3039         /*
3040          * Resolve the mapping until finished. (drivers are
3041          * still free to implement/resolve their own stacking
3042          * by explicitly returning 0)
3043          *
3044          * NOTE: we don't repeat the blk_size check for each new device.
3045          * Stacking drivers are expected to know what they are doing.
3046          */
3047         maxsector = -1;
3048         old_dev = 0;
3049         do {
3050                 char b[BDEVNAME_SIZE];
3051
3052                 q = bdev_get_queue(bio->bi_bdev);
3053                 if (!q) {
3054                         printk(KERN_ERR
3055                                "generic_make_request: Trying to access "
3056                                 "nonexistent block-device %s (%Lu)\n",
3057                                 bdevname(bio->bi_bdev, b),
3058                                 (long long) bio->bi_sector);
3059 end_io:
3060                         bio_endio(bio, bio->bi_size, -EIO);
3061                         break;
3062                 }
3063
3064                 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
3065                         printk("bio too big device %s (%u > %u)\n", 
3066                                 bdevname(bio->bi_bdev, b),
3067                                 bio_sectors(bio),
3068                                 q->max_hw_sectors);
3069                         goto end_io;
3070                 }
3071
3072                 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
3073                         goto end_io;
3074
3075                 /*
3076                  * If this device has partitions, remap block n
3077                  * of partition p to block n+start(p) of the disk.
3078                  */
3079                 blk_partition_remap(bio);
3080
3081                 if (maxsector != -1)
3082                         blk_add_trace_remap(q, bio, old_dev, bio->bi_sector, 
3083                                             maxsector);
3084
3085                 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
3086
3087                 maxsector = bio->bi_sector;
3088                 old_dev = bio->bi_bdev->bd_dev;
3089
3090                 ret = q->make_request_fn(q, bio);
3091         } while (ret);
3092 }
3093
3094 EXPORT_SYMBOL(generic_make_request);
3095
3096 /**
3097  * submit_bio: submit a bio to the block device layer for I/O
3098  * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
3099  * @bio: The &struct bio which describes the I/O
3100  *
3101  * submit_bio() is very similar in purpose to generic_make_request(), and
3102  * uses that function to do most of the work. Both are fairly rough
3103  * interfaces, @bio must be presetup and ready for I/O.
3104  *
3105  */
3106 void submit_bio(int rw, struct bio *bio)
3107 {
3108         int count = bio_sectors(bio);
3109
3110         BIO_BUG_ON(!bio->bi_size);
3111         BIO_BUG_ON(!bio->bi_io_vec);
3112         bio->bi_rw |= rw;
3113         if (rw & WRITE)
3114                 mod_page_state(pgpgout, count);
3115         else
3116                 mod_page_state(pgpgin, count);
3117
3118         if (unlikely(block_dump)) {
3119                 char b[BDEVNAME_SIZE];
3120                 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
3121                         current->comm, current->pid,
3122                         (rw & WRITE) ? "WRITE" : "READ",
3123                         (unsigned long long)bio->bi_sector,
3124                         bdevname(bio->bi_bdev,b));
3125         }
3126
3127         generic_make_request(bio);
3128 }
3129
3130 EXPORT_SYMBOL(submit_bio);
3131
3132 static void blk_recalc_rq_segments(struct request *rq)
3133 {
3134         struct bio *bio, *prevbio = NULL;
3135         int nr_phys_segs, nr_hw_segs;
3136         unsigned int phys_size, hw_size;
3137         request_queue_t *q = rq->q;
3138
3139         if (!rq->bio)
3140                 return;
3141
3142         phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
3143         rq_for_each_bio(bio, rq) {
3144                 /* Force bio hw/phys segs to be recalculated. */
3145                 bio->bi_flags &= ~(1 << BIO_SEG_VALID);
3146
3147                 nr_phys_segs += bio_phys_segments(q, bio);
3148                 nr_hw_segs += bio_hw_segments(q, bio);
3149                 if (prevbio) {
3150                         int pseg = phys_size + prevbio->bi_size + bio->bi_size;
3151                         int hseg = hw_size + prevbio->bi_size + bio->bi_size;
3152
3153                         if (blk_phys_contig_segment(q, prevbio, bio) &&
3154                             pseg <= q->max_segment_size) {
3155                                 nr_phys_segs--;
3156                                 phys_size += prevbio->bi_size + bio->bi_size;
3157                         } else
3158                                 phys_size = 0;
3159
3160                         if (blk_hw_contig_segment(q, prevbio, bio) &&
3161                             hseg <= q->max_segment_size) {
3162                                 nr_hw_segs--;
3163                                 hw_size += prevbio->bi_size + bio->bi_size;
3164                         } else
3165                                 hw_size = 0;
3166                 }
3167                 prevbio = bio;
3168         }
3169
3170         rq->nr_phys_segments = nr_phys_segs;
3171         rq->nr_hw_segments = nr_hw_segs;
3172 }
3173
3174 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
3175 {
3176         if (blk_fs_request(rq)) {
3177                 rq->hard_sector += nsect;
3178                 rq->hard_nr_sectors -= nsect;
3179
3180                 /*
3181                  * Move the I/O submission pointers ahead if required.
3182                  */
3183                 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3184                     (rq->sector <= rq->hard_sector)) {
3185                         rq->sector = rq->hard_sector;
3186                         rq->nr_sectors = rq->hard_nr_sectors;
3187                         rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3188                         rq->current_nr_sectors = rq->hard_cur_sectors;
3189                         rq->buffer = bio_data(rq->bio);
3190                 }
3191
3192                 /*
3193                  * if total number of sectors is less than the first segment
3194                  * size, something has gone terribly wrong
3195                  */
3196                 if (rq->nr_sectors < rq->current_nr_sectors) {
3197                         printk("blk: request botched\n");
3198                         rq->nr_sectors = rq->current_nr_sectors;
3199                 }
3200         }
3201 }
3202
3203 static int __end_that_request_first(struct request *req, int uptodate,
3204                                     int nr_bytes)
3205 {
3206         int total_bytes, bio_nbytes, error, next_idx = 0;
3207         struct bio *bio;
3208
3209         blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
3210
3211         /*
3212          * extend uptodate bool to allow < 0 value to be direct io error
3213          */
3214         error = 0;
3215         if (end_io_error(uptodate))
3216                 error = !uptodate ? -EIO : uptodate;
3217
3218         /*
3219          * for a REQ_BLOCK_PC request, we want to carry any eventual
3220          * sense key with us all the way through
3221          */
3222         if (!blk_pc_request(req))
3223                 req->errors = 0;
3224
3225         if (!uptodate) {
3226                 if (blk_fs_request(req) && !(req->flags & REQ_QUIET))
3227                         printk("end_request: I/O error, dev %s, sector %llu\n",
3228                                 req->rq_disk ? req->rq_disk->disk_name : "?",
3229                                 (unsigned long long)req->sector);
3230         }
3231
3232         if (blk_fs_request(req) && req->rq_disk) {
3233                 const int rw = rq_data_dir(req);
3234
3235                 disk_stat_add(req->rq_disk, sectors[rw], nr_bytes >> 9);
3236         }
3237
3238         total_bytes = bio_nbytes = 0;
3239         while ((bio = req->bio) != NULL) {
3240                 int nbytes;
3241
3242                 if (nr_bytes >= bio->bi_size) {
3243                         req->bio = bio->bi_next;
3244                         nbytes = bio->bi_size;
3245                         if (!ordered_bio_endio(req, bio, nbytes, error))
3246                                 bio_endio(bio, nbytes, error);
3247                         next_idx = 0;
3248                         bio_nbytes = 0;
3249                 } else {
3250                         int idx = bio->bi_idx + next_idx;
3251
3252                         if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3253                                 blk_dump_rq_flags(req, "__end_that");
3254                                 printk("%s: bio idx %d >= vcnt %d\n",
3255                                                 __FUNCTION__,
3256                                                 bio->bi_idx, bio->bi_vcnt);
3257                                 break;
3258                         }
3259
3260                         nbytes = bio_iovec_idx(bio, idx)->bv_len;
3261                         BIO_BUG_ON(nbytes > bio->bi_size);
3262
3263                         /*
3264                          * not a complete bvec done
3265                          */
3266                         if (unlikely(nbytes > nr_bytes)) {
3267                                 bio_nbytes += nr_bytes;
3268                                 total_bytes += nr_bytes;
3269                                 break;
3270                         }
3271
3272                         /*
3273                          * advance to the next vector
3274                          */
3275                         next_idx++;
3276                         bio_nbytes += nbytes;
3277                 }
3278
3279                 total_bytes += nbytes;
3280                 nr_bytes -= nbytes;
3281
3282                 if ((bio = req->bio)) {
3283                         /*
3284                          * end more in this run, or just return 'not-done'
3285                          */
3286                         if (unlikely(nr_bytes <= 0))
3287                                 break;
3288                 }
3289         }
3290
3291         /*
3292          * completely done
3293          */
3294         if (!req->bio)
3295                 return 0;
3296
3297         /*
3298          * if the request wasn't completed, update state
3299          */
3300         if (bio_nbytes) {
3301                 if (!ordered_bio_endio(req, bio, bio_nbytes, error))
3302                         bio_endio(bio, bio_nbytes, error);
3303                 bio->bi_idx += next_idx;
3304                 bio_iovec(bio)->bv_offset += nr_bytes;
3305                 bio_iovec(bio)->bv_len -= nr_bytes;
3306         }
3307
3308         blk_recalc_rq_sectors(req, total_bytes >> 9);
3309         blk_recalc_rq_segments(req);
3310         return 1;
3311 }
3312
3313 /**
3314  * end_that_request_first - end I/O on a request
3315  * @req:      the request being processed
3316  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3317  * @nr_sectors: number of sectors to end I/O on
3318  *
3319  * Description:
3320  *     Ends I/O on a number of sectors attached to @req, and sets it up
3321  *     for the next range of segments (if any) in the cluster.
3322  *
3323  * Return:
3324  *     0 - we are done with this request, call end_that_request_last()
3325  *     1 - still buffers pending for this request
3326  **/
3327 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3328 {
3329         return __end_that_request_first(req, uptodate, nr_sectors << 9);
3330 }
3331
3332 EXPORT_SYMBOL(end_that_request_first);
3333
3334 /**
3335  * end_that_request_chunk - end I/O on a request
3336  * @req:      the request being processed
3337  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3338  * @nr_bytes: number of bytes to complete
3339  *
3340  * Description:
3341  *     Ends I/O on a number of bytes attached to @req, and sets it up
3342  *     for the next range of segments (if any). Like end_that_request_first(),
3343  *     but deals with bytes instead of sectors.
3344  *
3345  * Return:
3346  *     0 - we are done with this request, call end_that_request_last()
3347  *     1 - still buffers pending for this request
3348  **/
3349 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3350 {
3351         return __end_that_request_first(req, uptodate, nr_bytes);
3352 }
3353
3354 EXPORT_SYMBOL(end_that_request_chunk);
3355
3356 /*
3357  * splice the completion data to a local structure and hand off to
3358  * process_completion_queue() to complete the requests
3359  */
3360 static void blk_done_softirq(struct softirq_action *h)
3361 {
3362         struct list_head *cpu_list;
3363         LIST_HEAD(local_list);
3364
3365         local_irq_disable();
3366         cpu_list = &__get_cpu_var(blk_cpu_done);
3367         list_splice_init(cpu_list, &local_list);
3368         local_irq_enable();
3369
3370         while (!list_empty(&local_list)) {
3371                 struct request *rq = list_entry(local_list.next, struct request, donelist);
3372
3373                 list_del_init(&rq->donelist);
3374                 rq->q->softirq_done_fn(rq);
3375         }
3376 }
3377
3378 #ifdef CONFIG_HOTPLUG_CPU
3379
3380 static int blk_cpu_notify(struct notifier_block *self, unsigned long action,
3381                           void *hcpu)
3382 {
3383         /*
3384          * If a CPU goes away, splice its entries to the current CPU
3385          * and trigger a run of the softirq
3386          */
3387         if (action == CPU_DEAD) {
3388                 int cpu = (unsigned long) hcpu;
3389
3390                 local_irq_disable();
3391                 list_splice_init(&per_cpu(blk_cpu_done, cpu),
3392                                  &__get_cpu_var(blk_cpu_done));
3393                 raise_softirq_irqoff(BLOCK_SOFTIRQ);
3394                 local_irq_enable();
3395         }
3396
3397         return NOTIFY_OK;
3398 }
3399
3400
3401 static struct notifier_block blk_cpu_notifier = {
3402         .notifier_call  = blk_cpu_notify,
3403 };
3404
3405 #endif /* CONFIG_HOTPLUG_CPU */
3406
3407 /**
3408  * blk_complete_request - end I/O on a request
3409  * @req:      the request being processed
3410  *
3411  * Description:
3412  *     Ends all I/O on a request. It does not handle partial completions,
3413  *     unless the driver actually implements this in its completionc callback
3414  *     through requeueing. Theh actual completion happens out-of-order,
3415  *     through a softirq handler. The user must have registered a completion
3416  *     callback through blk_queue_softirq_done().
3417  **/
3418
3419 void blk_complete_request(struct request *req)
3420 {
3421         struct list_head *cpu_list;
3422         unsigned long flags;
3423
3424         BUG_ON(!req->q->softirq_done_fn);
3425                 
3426         local_irq_save(flags);
3427
3428         cpu_list = &__get_cpu_var(blk_cpu_done);
3429         list_add_tail(&req->donelist, cpu_list);
3430         raise_softirq_irqoff(BLOCK_SOFTIRQ);
3431
3432         local_irq_restore(flags);
3433 }
3434
3435 EXPORT_SYMBOL(blk_complete_request);
3436         
3437 /*
3438  * queue lock must be held
3439  */
3440 void end_that_request_last(struct request *req, int uptodate)
3441 {
3442         struct gendisk *disk = req->rq_disk;
3443         int error;
3444
3445         /*
3446          * extend uptodate bool to allow < 0 value to be direct io error
3447          */
3448         error = 0;
3449         if (end_io_error(uptodate))
3450                 error = !uptodate ? -EIO : uptodate;
3451
3452         if (unlikely(laptop_mode) && blk_fs_request(req))
3453                 laptop_io_completion();
3454
3455         if (disk && blk_fs_request(req)) {
3456                 unsigned long duration = jiffies - req->start_time;
3457                 const int rw = rq_data_dir(req);
3458
3459                 __disk_stat_inc(disk, ios[rw]);
3460                 __disk_stat_add(disk, ticks[rw], duration);
3461                 disk_round_stats(disk);
3462                 disk->in_flight--;
3463         }
3464         if (req->end_io)
3465                 req->end_io(req, error);
3466         else
3467                 __blk_put_request(req->q, req);
3468 }
3469
3470 EXPORT_SYMBOL(end_that_request_last);
3471
3472 void end_request(struct request *req, int uptodate)
3473 {
3474         if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3475                 add_disk_randomness(req->rq_disk);
3476                 blkdev_dequeue_request(req);
3477                 end_that_request_last(req, uptodate);
3478         }
3479 }
3480
3481 EXPORT_SYMBOL(end_request);
3482
3483 void blk_rq_bio_prep(request_queue_t *q, struct request *rq, struct bio *bio)
3484 {
3485         /* first three bits are identical in rq->flags and bio->bi_rw */
3486         rq->flags |= (bio->bi_rw & 7);
3487
3488         rq->nr_phys_segments = bio_phys_segments(q, bio);
3489         rq->nr_hw_segments = bio_hw_segments(q, bio);
3490         rq->current_nr_sectors = bio_cur_sectors(bio);
3491         rq->hard_cur_sectors = rq->current_nr_sectors;
3492         rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3493         rq->buffer = bio_data(bio);
3494
3495         rq->bio = rq->biotail = bio;
3496 }
3497
3498 EXPORT_SYMBOL(blk_rq_bio_prep);
3499
3500 int kblockd_schedule_work(struct work_struct *work)
3501 {
3502         return queue_work(kblockd_workqueue, work);
3503 }
3504
3505 EXPORT_SYMBOL(kblockd_schedule_work);
3506
3507 void kblockd_flush(void)
3508 {
3509         flush_workqueue(kblockd_workqueue);
3510 }
3511 EXPORT_SYMBOL(kblockd_flush);
3512
3513 int __init blk_dev_init(void)
3514 {
3515         int i;
3516
3517         kblockd_workqueue = create_workqueue("kblockd");
3518         if (!kblockd_workqueue)
3519                 panic("Failed to create kblockd\n");
3520
3521         request_cachep = kmem_cache_create("blkdev_requests",
3522                         sizeof(struct request), 0, SLAB_PANIC, NULL, NULL);
3523
3524         requestq_cachep = kmem_cache_create("blkdev_queue",
3525                         sizeof(request_queue_t), 0, SLAB_PANIC, NULL, NULL);
3526
3527         iocontext_cachep = kmem_cache_create("blkdev_ioc",
3528                         sizeof(struct io_context), 0, SLAB_PANIC, NULL, NULL);
3529
3530         for_each_possible_cpu(i)
3531                 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3532
3533         open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
3534 #ifdef CONFIG_HOTPLUG_CPU
3535         register_cpu_notifier(&blk_cpu_notifier);
3536 #endif
3537
3538         blk_max_low_pfn = max_low_pfn;
3539         blk_max_pfn = max_pfn;
3540
3541         return 0;
3542 }
3543
3544 /*
3545  * IO Context helper functions
3546  */
3547 void put_io_context(struct io_context *ioc)
3548 {
3549         if (ioc == NULL)
3550                 return;
3551
3552         BUG_ON(atomic_read(&ioc->refcount) == 0);
3553
3554         if (atomic_dec_and_test(&ioc->refcount)) {
3555                 struct cfq_io_context *cic;
3556
3557                 rcu_read_lock();
3558                 if (ioc->aic && ioc->aic->dtor)
3559                         ioc->aic->dtor(ioc->aic);
3560                 if (ioc->cic_root.rb_node != NULL) {
3561                         struct rb_node *n = rb_first(&ioc->cic_root);
3562
3563                         cic = rb_entry(n, struct cfq_io_context, rb_node);
3564                         cic->dtor(ioc);
3565                 }
3566                 rcu_read_unlock();
3567
3568                 kmem_cache_free(iocontext_cachep, ioc);
3569         }
3570 }
3571 EXPORT_SYMBOL(put_io_context);
3572
3573 /* Called by the exitting task */
3574 void exit_io_context(void)
3575 {
3576         unsigned long flags;
3577         struct io_context *ioc;
3578         struct cfq_io_context *cic;
3579
3580         local_irq_save(flags);
3581         task_lock(current);
3582         ioc = current->io_context;
3583         current->io_context = NULL;
3584         ioc->task = NULL;
3585         task_unlock(current);
3586         local_irq_restore(flags);
3587
3588         if (ioc->aic && ioc->aic->exit)
3589                 ioc->aic->exit(ioc->aic);
3590         if (ioc->cic_root.rb_node != NULL) {
3591                 cic = rb_entry(rb_first(&ioc->cic_root), struct cfq_io_context, rb_node);
3592                 cic->exit(ioc);
3593         }
3594  
3595         put_io_context(ioc);
3596 }
3597
3598 /*
3599  * If the current task has no IO context then create one and initialise it.
3600  * Otherwise, return its existing IO context.
3601  *
3602  * This returned IO context doesn't have a specifically elevated refcount,
3603  * but since the current task itself holds a reference, the context can be
3604  * used in general code, so long as it stays within `current` context.
3605  */
3606 struct io_context *current_io_context(gfp_t gfp_flags)
3607 {
3608         struct task_struct *tsk = current;
3609         struct io_context *ret;
3610
3611         ret = tsk->io_context;
3612         if (likely(ret))
3613                 return ret;
3614
3615         ret = kmem_cache_alloc(iocontext_cachep, gfp_flags);
3616         if (ret) {
3617                 atomic_set(&ret->refcount, 1);
3618                 ret->task = current;
3619                 ret->set_ioprio = NULL;
3620                 ret->last_waited = jiffies; /* doesn't matter... */
3621                 ret->nr_batch_requests = 0; /* because this is 0 */
3622                 ret->aic = NULL;
3623                 ret->cic_root.rb_node = NULL;
3624                 tsk->io_context = ret;
3625         }
3626
3627         return ret;
3628 }
3629 EXPORT_SYMBOL(current_io_context);
3630
3631 /*
3632  * If the current task has no IO context then create one and initialise it.
3633  * If it does have a context, take a ref on it.
3634  *
3635  * This is always called in the context of the task which submitted the I/O.
3636  */
3637 struct io_context *get_io_context(gfp_t gfp_flags)
3638 {
3639         struct io_context *ret;
3640         ret = current_io_context(gfp_flags);
3641         if (likely(ret))
3642                 atomic_inc(&ret->refcount);
3643         return ret;
3644 }
3645 EXPORT_SYMBOL(get_io_context);
3646
3647 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3648 {
3649         struct io_context *src = *psrc;
3650         struct io_context *dst = *pdst;
3651
3652         if (src) {
3653                 BUG_ON(atomic_read(&src->refcount) == 0);
3654                 atomic_inc(&src->refcount);
3655                 put_io_context(dst);
3656                 *pdst = src;
3657         }
3658 }
3659 EXPORT_SYMBOL(copy_io_context);
3660
3661 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3662 {
3663         struct io_context *temp;
3664         temp = *ioc1;
3665         *ioc1 = *ioc2;
3666         *ioc2 = temp;
3667 }
3668 EXPORT_SYMBOL(swap_io_context);
3669
3670 /*
3671  * sysfs parts below
3672  */
3673 struct queue_sysfs_entry {
3674         struct attribute attr;
3675         ssize_t (*show)(struct request_queue *, char *);
3676         ssize_t (*store)(struct request_queue *, const char *, size_t);
3677 };
3678
3679 static ssize_t
3680 queue_var_show(unsigned int var, char *page)
3681 {
3682         return sprintf(page, "%d\n", var);
3683 }
3684
3685 static ssize_t
3686 queue_var_store(unsigned long *var, const char *page, size_t count)
3687 {
3688         char *p = (char *) page;
3689
3690         *var = simple_strtoul(p, &p, 10);
3691         return count;
3692 }
3693
3694 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3695 {
3696         return queue_var_show(q->nr_requests, (page));
3697 }
3698
3699 static ssize_t
3700 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3701 {
3702         struct request_list *rl = &q->rq;
3703         unsigned long nr;
3704         int ret = queue_var_store(&nr, page, count);
3705         if (nr < BLKDEV_MIN_RQ)
3706                 nr = BLKDEV_MIN_RQ;
3707
3708         spin_lock_irq(q->queue_lock);
3709         q->nr_requests = nr;
3710         blk_queue_congestion_threshold(q);
3711
3712         if (rl->count[READ] >= queue_congestion_on_threshold(q))
3713                 set_queue_congested(q, READ);
3714         else if (rl->count[READ] < queue_congestion_off_threshold(q))
3715                 clear_queue_congested(q, READ);
3716
3717         if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3718                 set_queue_congested(q, WRITE);
3719         else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3720                 clear_queue_congested(q, WRITE);
3721
3722         if (rl->count[READ] >= q->nr_requests) {
3723                 blk_set_queue_full(q, READ);
3724         } else if (rl->count[READ]+1 <= q->nr_requests) {
3725                 blk_clear_queue_full(q, READ);
3726                 wake_up(&rl->wait[READ]);
3727         }
3728
3729         if (rl->count[WRITE] >= q->nr_requests) {
3730                 blk_set_queue_full(q, WRITE);
3731         } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3732                 blk_clear_queue_full(q, WRITE);
3733                 wake_up(&rl->wait[WRITE]);
3734         }
3735         spin_unlock_irq(q->queue_lock);
3736         return ret;
3737 }
3738
3739 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3740 {
3741         int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3742
3743         return queue_var_show(ra_kb, (page));
3744 }
3745
3746 static ssize_t
3747 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3748 {
3749         unsigned long ra_kb;
3750         ssize_t ret = queue_var_store(&ra_kb, page, count);
3751
3752         spin_lock_irq(q->queue_lock);
3753         if (ra_kb > (q->max_sectors >> 1))
3754                 ra_kb = (q->max_sectors >> 1);
3755
3756         q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3757         spin_unlock_irq(q->queue_lock);
3758
3759         return ret;
3760 }
3761
3762 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3763 {
3764         int max_sectors_kb = q->max_sectors >> 1;
3765
3766         return queue_var_show(max_sectors_kb, (page));
3767 }
3768
3769 static ssize_t
3770 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3771 {
3772         unsigned long max_sectors_kb,
3773                         max_hw_sectors_kb = q->max_hw_sectors >> 1,
3774                         page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3775         ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3776         int ra_kb;
3777
3778         if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3779                 return -EINVAL;
3780         /*
3781          * Take the queue lock to update the readahead and max_sectors
3782          * values synchronously:
3783          */
3784         spin_lock_irq(q->queue_lock);
3785         /*
3786          * Trim readahead window as well, if necessary:
3787          */
3788         ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3789         if (ra_kb > max_sectors_kb)
3790                 q->backing_dev_info.ra_pages =
3791                                 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3792
3793         q->max_sectors = max_sectors_kb << 1;
3794         spin_unlock_irq(q->queue_lock);
3795
3796         return ret;
3797 }
3798
3799 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3800 {
3801         int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3802
3803         return queue_var_show(max_hw_sectors_kb, (page));
3804 }
3805
3806
3807 static struct queue_sysfs_entry queue_requests_entry = {
3808         .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3809         .show = queue_requests_show,
3810         .store = queue_requests_store,
3811 };
3812
3813 static struct queue_sysfs_entry queue_ra_entry = {
3814         .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3815         .show = queue_ra_show,
3816         .store = queue_ra_store,
3817 };
3818
3819 static struct queue_sysfs_entry queue_max_sectors_entry = {
3820         .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3821         .show = queue_max_sectors_show,
3822         .store = queue_max_sectors_store,
3823 };
3824
3825 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3826         .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
3827         .show = queue_max_hw_sectors_show,
3828 };
3829
3830 static struct queue_sysfs_entry queue_iosched_entry = {
3831         .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
3832         .show = elv_iosched_show,
3833         .store = elv_iosched_store,
3834 };
3835
3836 static struct attribute *default_attrs[] = {
3837         &queue_requests_entry.attr,
3838         &queue_ra_entry.attr,
3839         &queue_max_hw_sectors_entry.attr,
3840         &queue_max_sectors_entry.attr,
3841         &queue_iosched_entry.attr,
3842         NULL,
3843 };
3844
3845 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
3846
3847 static ssize_t
3848 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
3849 {
3850         struct queue_sysfs_entry *entry = to_queue(attr);
3851         request_queue_t *q = container_of(kobj, struct request_queue, kobj);
3852         ssize_t res;
3853
3854         if (!entry->show)
3855                 return -EIO;
3856         mutex_lock(&q->sysfs_lock);
3857         if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
3858                 mutex_unlock(&q->sysfs_lock);
3859                 return -ENOENT;
3860         }
3861         res = entry->show(q, page);
3862         mutex_unlock(&q->sysfs_lock);
3863         return res;
3864 }
3865
3866 static ssize_t
3867 queue_attr_store(struct kobject *kobj, struct attribute *attr,
3868                     const char *page, size_t length)
3869 {
3870         struct queue_sysfs_entry *entry = to_queue(attr);
3871         request_queue_t *q = container_of(kobj, struct request_queue, kobj);
3872
3873         ssize_t res;
3874
3875         if (!entry->store)
3876                 return -EIO;
3877         mutex_lock(&q->sysfs_lock);
3878         if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
3879                 mutex_unlock(&q->sysfs_lock);
3880                 return -ENOENT;
3881         }
3882         res = entry->store(q, page, length);
3883         mutex_unlock(&q->sysfs_lock);
3884         return res;
3885 }
3886
3887 static struct sysfs_ops queue_sysfs_ops = {
3888         .show   = queue_attr_show,
3889         .store  = queue_attr_store,
3890 };
3891
3892 static struct kobj_type queue_ktype = {
3893         .sysfs_ops      = &queue_sysfs_ops,
3894         .default_attrs  = default_attrs,
3895         .release        = blk_release_queue,
3896 };
3897
3898 int blk_register_queue(struct gendisk *disk)
3899 {
3900         int ret;
3901
3902         request_queue_t *q = disk->queue;
3903
3904         if (!q || !q->request_fn)
3905                 return -ENXIO;
3906
3907         q->kobj.parent = kobject_get(&disk->kobj);
3908
3909         ret = kobject_add(&q->kobj);
3910         if (ret < 0)
3911                 return ret;
3912
3913         kobject_uevent(&q->kobj, KOBJ_ADD);
3914
3915         ret = elv_register_queue(q);
3916         if (ret) {
3917                 kobject_uevent(&q->kobj, KOBJ_REMOVE);
3918                 kobject_del(&q->kobj);
3919                 return ret;
3920         }
3921
3922         return 0;
3923 }
3924
3925 void blk_unregister_queue(struct gendisk *disk)
3926 {
3927         request_queue_t *q = disk->queue;
3928
3929         if (q && q->request_fn) {
3930                 elv_unregister_queue(q);
3931
3932                 kobject_uevent(&q->kobj, KOBJ_REMOVE);
3933                 kobject_del(&q->kobj);
3934                 kobject_put(&disk->kobj);
3935         }
3936 }