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