Auto merge with /home/aegl/GIT/linus
[linux-2.6] / 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->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->max_depth = depth;
802         tags->tag_index = tag_index;
803         tags->tag_map = tag_map;
804
805         return 0;
806 fail:
807         kfree(tag_index);
808         return -ENOMEM;
809 }
810
811 /**
812  * blk_queue_init_tags - initialize the queue tag info
813  * @q:  the request queue for the device
814  * @depth:  the maximum queue depth supported
815  * @tags: the tag to use
816  **/
817 int blk_queue_init_tags(request_queue_t *q, int depth,
818                         struct blk_queue_tag *tags)
819 {
820         int rc;
821
822         BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
823
824         if (!tags && !q->queue_tags) {
825                 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
826                 if (!tags)
827                         goto fail;
828
829                 if (init_tag_map(q, tags, depth))
830                         goto fail;
831
832                 INIT_LIST_HEAD(&tags->busy_list);
833                 tags->busy = 0;
834                 atomic_set(&tags->refcnt, 1);
835         } else if (q->queue_tags) {
836                 if ((rc = blk_queue_resize_tags(q, depth)))
837                         return rc;
838                 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
839                 return 0;
840         } else
841                 atomic_inc(&tags->refcnt);
842
843         /*
844          * assign it, all done
845          */
846         q->queue_tags = tags;
847         q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
848         return 0;
849 fail:
850         kfree(tags);
851         return -ENOMEM;
852 }
853
854 EXPORT_SYMBOL(blk_queue_init_tags);
855
856 /**
857  * blk_queue_resize_tags - change the queueing depth
858  * @q:  the request queue for the device
859  * @new_depth: the new max command queueing depth
860  *
861  *  Notes:
862  *    Must be called with the queue lock held.
863  **/
864 int blk_queue_resize_tags(request_queue_t *q, int new_depth)
865 {
866         struct blk_queue_tag *bqt = q->queue_tags;
867         struct request **tag_index;
868         unsigned long *tag_map;
869         int max_depth, nr_ulongs;
870
871         if (!bqt)
872                 return -ENXIO;
873
874         /*
875          * save the old state info, so we can copy it back
876          */
877         tag_index = bqt->tag_index;
878         tag_map = bqt->tag_map;
879         max_depth = bqt->max_depth;
880
881         if (init_tag_map(q, bqt, new_depth))
882                 return -ENOMEM;
883
884         memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
885         nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
886         memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
887
888         kfree(tag_index);
889         kfree(tag_map);
890         return 0;
891 }
892
893 EXPORT_SYMBOL(blk_queue_resize_tags);
894
895 /**
896  * blk_queue_end_tag - end tag operations for a request
897  * @q:  the request queue for the device
898  * @rq: the request that has completed
899  *
900  *  Description:
901  *    Typically called when end_that_request_first() returns 0, meaning
902  *    all transfers have been done for a request. It's important to call
903  *    this function before end_that_request_last(), as that will put the
904  *    request back on the free list thus corrupting the internal tag list.
905  *
906  *  Notes:
907  *   queue lock must be held.
908  **/
909 void blk_queue_end_tag(request_queue_t *q, struct request *rq)
910 {
911         struct blk_queue_tag *bqt = q->queue_tags;
912         int tag = rq->tag;
913
914         BUG_ON(tag == -1);
915
916         if (unlikely(tag >= bqt->max_depth))
917                 /*
918                  * This can happen after tag depth has been reduced.
919                  * FIXME: how about a warning or info message here?
920                  */
921                 return;
922
923         if (unlikely(!__test_and_clear_bit(tag, bqt->tag_map))) {
924                 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
925                        __FUNCTION__, tag);
926                 return;
927         }
928
929         list_del_init(&rq->queuelist);
930         rq->flags &= ~REQ_QUEUED;
931         rq->tag = -1;
932
933         if (unlikely(bqt->tag_index[tag] == NULL))
934                 printk(KERN_ERR "%s: tag %d is missing\n",
935                        __FUNCTION__, tag);
936
937         bqt->tag_index[tag] = NULL;
938         bqt->busy--;
939 }
940
941 EXPORT_SYMBOL(blk_queue_end_tag);
942
943 /**
944  * blk_queue_start_tag - find a free tag and assign it
945  * @q:  the request queue for the device
946  * @rq:  the block request that needs tagging
947  *
948  *  Description:
949  *    This can either be used as a stand-alone helper, or possibly be
950  *    assigned as the queue &prep_rq_fn (in which case &struct request
951  *    automagically gets a tag assigned). Note that this function
952  *    assumes that any type of request can be queued! if this is not
953  *    true for your device, you must check the request type before
954  *    calling this function.  The request will also be removed from
955  *    the request queue, so it's the drivers responsibility to readd
956  *    it if it should need to be restarted for some reason.
957  *
958  *  Notes:
959  *   queue lock must be held.
960  **/
961 int blk_queue_start_tag(request_queue_t *q, struct request *rq)
962 {
963         struct blk_queue_tag *bqt = q->queue_tags;
964         int tag;
965
966         if (unlikely((rq->flags & REQ_QUEUED))) {
967                 printk(KERN_ERR 
968                        "%s: request %p for device [%s] already tagged %d",
969                        __FUNCTION__, rq,
970                        rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
971                 BUG();
972         }
973
974         tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
975         if (tag >= bqt->max_depth)
976                 return 1;
977
978         __set_bit(tag, bqt->tag_map);
979
980         rq->flags |= REQ_QUEUED;
981         rq->tag = tag;
982         bqt->tag_index[tag] = rq;
983         blkdev_dequeue_request(rq);
984         list_add(&rq->queuelist, &bqt->busy_list);
985         bqt->busy++;
986         return 0;
987 }
988
989 EXPORT_SYMBOL(blk_queue_start_tag);
990
991 /**
992  * blk_queue_invalidate_tags - invalidate all pending tags
993  * @q:  the request queue for the device
994  *
995  *  Description:
996  *   Hardware conditions may dictate a need to stop all pending requests.
997  *   In this case, we will safely clear the block side of the tag queue and
998  *   readd all requests to the request queue in the right order.
999  *
1000  *  Notes:
1001  *   queue lock must be held.
1002  **/
1003 void blk_queue_invalidate_tags(request_queue_t *q)
1004 {
1005         struct blk_queue_tag *bqt = q->queue_tags;
1006         struct list_head *tmp, *n;
1007         struct request *rq;
1008
1009         list_for_each_safe(tmp, n, &bqt->busy_list) {
1010                 rq = list_entry_rq(tmp);
1011
1012                 if (rq->tag == -1) {
1013                         printk(KERN_ERR
1014                                "%s: bad tag found on list\n", __FUNCTION__);
1015                         list_del_init(&rq->queuelist);
1016                         rq->flags &= ~REQ_QUEUED;
1017                 } else
1018                         blk_queue_end_tag(q, rq);
1019
1020                 rq->flags &= ~REQ_STARTED;
1021                 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1022         }
1023 }
1024
1025 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1026
1027 static char *rq_flags[] = {
1028         "REQ_RW",
1029         "REQ_FAILFAST",
1030         "REQ_SOFTBARRIER",
1031         "REQ_HARDBARRIER",
1032         "REQ_CMD",
1033         "REQ_NOMERGE",
1034         "REQ_STARTED",
1035         "REQ_DONTPREP",
1036         "REQ_QUEUED",
1037         "REQ_PC",
1038         "REQ_BLOCK_PC",
1039         "REQ_SENSE",
1040         "REQ_FAILED",
1041         "REQ_QUIET",
1042         "REQ_SPECIAL",
1043         "REQ_DRIVE_CMD",
1044         "REQ_DRIVE_TASK",
1045         "REQ_DRIVE_TASKFILE",
1046         "REQ_PREEMPT",
1047         "REQ_PM_SUSPEND",
1048         "REQ_PM_RESUME",
1049         "REQ_PM_SHUTDOWN",
1050 };
1051
1052 void blk_dump_rq_flags(struct request *rq, char *msg)
1053 {
1054         int bit;
1055
1056         printk("%s: dev %s: flags = ", msg,
1057                 rq->rq_disk ? rq->rq_disk->disk_name : "?");
1058         bit = 0;
1059         do {
1060                 if (rq->flags & (1 << bit))
1061                         printk("%s ", rq_flags[bit]);
1062                 bit++;
1063         } while (bit < __REQ_NR_BITS);
1064
1065         printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1066                                                        rq->nr_sectors,
1067                                                        rq->current_nr_sectors);
1068         printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1069
1070         if (rq->flags & (REQ_BLOCK_PC | REQ_PC)) {
1071                 printk("cdb: ");
1072                 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1073                         printk("%02x ", rq->cmd[bit]);
1074                 printk("\n");
1075         }
1076 }
1077
1078 EXPORT_SYMBOL(blk_dump_rq_flags);
1079
1080 void blk_recount_segments(request_queue_t *q, struct bio *bio)
1081 {
1082         struct bio_vec *bv, *bvprv = NULL;
1083         int i, nr_phys_segs, nr_hw_segs, seg_size, hw_seg_size, cluster;
1084         int high, highprv = 1;
1085
1086         if (unlikely(!bio->bi_io_vec))
1087                 return;
1088
1089         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1090         hw_seg_size = seg_size = nr_phys_segs = nr_hw_segs = 0;
1091         bio_for_each_segment(bv, bio, i) {
1092                 /*
1093                  * the trick here is making sure that a high page is never
1094                  * considered part of another segment, since that might
1095                  * change with the bounce page.
1096                  */
1097                 high = page_to_pfn(bv->bv_page) >= q->bounce_pfn;
1098                 if (high || highprv)
1099                         goto new_hw_segment;
1100                 if (cluster) {
1101                         if (seg_size + bv->bv_len > q->max_segment_size)
1102                                 goto new_segment;
1103                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1104                                 goto new_segment;
1105                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1106                                 goto new_segment;
1107                         if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1108                                 goto new_hw_segment;
1109
1110                         seg_size += bv->bv_len;
1111                         hw_seg_size += bv->bv_len;
1112                         bvprv = bv;
1113                         continue;
1114                 }
1115 new_segment:
1116                 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1117                     !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len)) {
1118                         hw_seg_size += bv->bv_len;
1119                 } else {
1120 new_hw_segment:
1121                         if (hw_seg_size > bio->bi_hw_front_size)
1122                                 bio->bi_hw_front_size = hw_seg_size;
1123                         hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1124                         nr_hw_segs++;
1125                 }
1126
1127                 nr_phys_segs++;
1128                 bvprv = bv;
1129                 seg_size = bv->bv_len;
1130                 highprv = high;
1131         }
1132         if (hw_seg_size > bio->bi_hw_back_size)
1133                 bio->bi_hw_back_size = hw_seg_size;
1134         if (nr_hw_segs == 1 && hw_seg_size > bio->bi_hw_front_size)
1135                 bio->bi_hw_front_size = hw_seg_size;
1136         bio->bi_phys_segments = nr_phys_segs;
1137         bio->bi_hw_segments = nr_hw_segs;
1138         bio->bi_flags |= (1 << BIO_SEG_VALID);
1139 }
1140
1141
1142 static int blk_phys_contig_segment(request_queue_t *q, struct bio *bio,
1143                                    struct bio *nxt)
1144 {
1145         if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1146                 return 0;
1147
1148         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1149                 return 0;
1150         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1151                 return 0;
1152
1153         /*
1154          * bio and nxt are contigous in memory, check if the queue allows
1155          * these two to be merged into one
1156          */
1157         if (BIO_SEG_BOUNDARY(q, bio, nxt))
1158                 return 1;
1159
1160         return 0;
1161 }
1162
1163 static int blk_hw_contig_segment(request_queue_t *q, struct bio *bio,
1164                                  struct bio *nxt)
1165 {
1166         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1167                 blk_recount_segments(q, bio);
1168         if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1169                 blk_recount_segments(q, nxt);
1170         if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1171             BIOVEC_VIRT_OVERSIZE(bio->bi_hw_front_size + bio->bi_hw_back_size))
1172                 return 0;
1173         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1174                 return 0;
1175
1176         return 1;
1177 }
1178
1179 /*
1180  * map a request to scatterlist, return number of sg entries setup. Caller
1181  * must make sure sg can hold rq->nr_phys_segments entries
1182  */
1183 int blk_rq_map_sg(request_queue_t *q, struct request *rq, struct scatterlist *sg)
1184 {
1185         struct bio_vec *bvec, *bvprv;
1186         struct bio *bio;
1187         int nsegs, i, cluster;
1188
1189         nsegs = 0;
1190         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1191
1192         /*
1193          * for each bio in rq
1194          */
1195         bvprv = NULL;
1196         rq_for_each_bio(bio, rq) {
1197                 /*
1198                  * for each segment in bio
1199                  */
1200                 bio_for_each_segment(bvec, bio, i) {
1201                         int nbytes = bvec->bv_len;
1202
1203                         if (bvprv && cluster) {
1204                                 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1205                                         goto new_segment;
1206
1207                                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1208                                         goto new_segment;
1209                                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1210                                         goto new_segment;
1211
1212                                 sg[nsegs - 1].length += nbytes;
1213                         } else {
1214 new_segment:
1215                                 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1216                                 sg[nsegs].page = bvec->bv_page;
1217                                 sg[nsegs].length = nbytes;
1218                                 sg[nsegs].offset = bvec->bv_offset;
1219
1220                                 nsegs++;
1221                         }
1222                         bvprv = bvec;
1223                 } /* segments in bio */
1224         } /* bios in rq */
1225
1226         return nsegs;
1227 }
1228
1229 EXPORT_SYMBOL(blk_rq_map_sg);
1230
1231 /*
1232  * the standard queue merge functions, can be overridden with device
1233  * specific ones if so desired
1234  */
1235
1236 static inline int ll_new_mergeable(request_queue_t *q,
1237                                    struct request *req,
1238                                    struct bio *bio)
1239 {
1240         int nr_phys_segs = bio_phys_segments(q, bio);
1241
1242         if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1243                 req->flags |= REQ_NOMERGE;
1244                 if (req == q->last_merge)
1245                         q->last_merge = NULL;
1246                 return 0;
1247         }
1248
1249         /*
1250          * A hw segment is just getting larger, bump just the phys
1251          * counter.
1252          */
1253         req->nr_phys_segments += nr_phys_segs;
1254         return 1;
1255 }
1256
1257 static inline int ll_new_hw_segment(request_queue_t *q,
1258                                     struct request *req,
1259                                     struct bio *bio)
1260 {
1261         int nr_hw_segs = bio_hw_segments(q, bio);
1262         int nr_phys_segs = bio_phys_segments(q, bio);
1263
1264         if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1265             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1266                 req->flags |= REQ_NOMERGE;
1267                 if (req == q->last_merge)
1268                         q->last_merge = NULL;
1269                 return 0;
1270         }
1271
1272         /*
1273          * This will form the start of a new hw segment.  Bump both
1274          * counters.
1275          */
1276         req->nr_hw_segments += nr_hw_segs;
1277         req->nr_phys_segments += nr_phys_segs;
1278         return 1;
1279 }
1280
1281 static int ll_back_merge_fn(request_queue_t *q, struct request *req, 
1282                             struct bio *bio)
1283 {
1284         int len;
1285
1286         if (req->nr_sectors + bio_sectors(bio) > q->max_sectors) {
1287                 req->flags |= REQ_NOMERGE;
1288                 if (req == q->last_merge)
1289                         q->last_merge = NULL;
1290                 return 0;
1291         }
1292         if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1293                 blk_recount_segments(q, req->biotail);
1294         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1295                 blk_recount_segments(q, bio);
1296         len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1297         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1298             !BIOVEC_VIRT_OVERSIZE(len)) {
1299                 int mergeable =  ll_new_mergeable(q, req, bio);
1300
1301                 if (mergeable) {
1302                         if (req->nr_hw_segments == 1)
1303                                 req->bio->bi_hw_front_size = len;
1304                         if (bio->bi_hw_segments == 1)
1305                                 bio->bi_hw_back_size = len;
1306                 }
1307                 return mergeable;
1308         }
1309
1310         return ll_new_hw_segment(q, req, bio);
1311 }
1312
1313 static int ll_front_merge_fn(request_queue_t *q, struct request *req, 
1314                              struct bio *bio)
1315 {
1316         int len;
1317
1318         if (req->nr_sectors + bio_sectors(bio) > q->max_sectors) {
1319                 req->flags |= REQ_NOMERGE;
1320                 if (req == q->last_merge)
1321                         q->last_merge = NULL;
1322                 return 0;
1323         }
1324         len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1325         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1326                 blk_recount_segments(q, bio);
1327         if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1328                 blk_recount_segments(q, req->bio);
1329         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1330             !BIOVEC_VIRT_OVERSIZE(len)) {
1331                 int mergeable =  ll_new_mergeable(q, req, bio);
1332
1333                 if (mergeable) {
1334                         if (bio->bi_hw_segments == 1)
1335                                 bio->bi_hw_front_size = len;
1336                         if (req->nr_hw_segments == 1)
1337                                 req->biotail->bi_hw_back_size = len;
1338                 }
1339                 return mergeable;
1340         }
1341
1342         return ll_new_hw_segment(q, req, bio);
1343 }
1344
1345 static int ll_merge_requests_fn(request_queue_t *q, struct request *req,
1346                                 struct request *next)
1347 {
1348         int total_phys_segments;
1349         int total_hw_segments;
1350
1351         /*
1352          * First check if the either of the requests are re-queued
1353          * requests.  Can't merge them if they are.
1354          */
1355         if (req->special || next->special)
1356                 return 0;
1357
1358         /*
1359          * Will it become too large?
1360          */
1361         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1362                 return 0;
1363
1364         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1365         if (blk_phys_contig_segment(q, req->biotail, next->bio))
1366                 total_phys_segments--;
1367
1368         if (total_phys_segments > q->max_phys_segments)
1369                 return 0;
1370
1371         total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1372         if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1373                 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1374                 /*
1375                  * propagate the combined length to the end of the requests
1376                  */
1377                 if (req->nr_hw_segments == 1)
1378                         req->bio->bi_hw_front_size = len;
1379                 if (next->nr_hw_segments == 1)
1380                         next->biotail->bi_hw_back_size = len;
1381                 total_hw_segments--;
1382         }
1383
1384         if (total_hw_segments > q->max_hw_segments)
1385                 return 0;
1386
1387         /* Merge is OK... */
1388         req->nr_phys_segments = total_phys_segments;
1389         req->nr_hw_segments = total_hw_segments;
1390         return 1;
1391 }
1392
1393 /*
1394  * "plug" the device if there are no outstanding requests: this will
1395  * force the transfer to start only after we have put all the requests
1396  * on the list.
1397  *
1398  * This is called with interrupts off and no requests on the queue and
1399  * with the queue lock held.
1400  */
1401 void blk_plug_device(request_queue_t *q)
1402 {
1403         WARN_ON(!irqs_disabled());
1404
1405         /*
1406          * don't plug a stopped queue, it must be paired with blk_start_queue()
1407          * which will restart the queueing
1408          */
1409         if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
1410                 return;
1411
1412         if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1413                 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1414 }
1415
1416 EXPORT_SYMBOL(blk_plug_device);
1417
1418 /*
1419  * remove the queue from the plugged list, if present. called with
1420  * queue lock held and interrupts disabled.
1421  */
1422 int blk_remove_plug(request_queue_t *q)
1423 {
1424         WARN_ON(!irqs_disabled());
1425
1426         if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1427                 return 0;
1428
1429         del_timer(&q->unplug_timer);
1430         return 1;
1431 }
1432
1433 EXPORT_SYMBOL(blk_remove_plug);
1434
1435 /*
1436  * remove the plug and let it rip..
1437  */
1438 void __generic_unplug_device(request_queue_t *q)
1439 {
1440         if (unlikely(test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags)))
1441                 return;
1442
1443         if (!blk_remove_plug(q))
1444                 return;
1445
1446         q->request_fn(q);
1447 }
1448 EXPORT_SYMBOL(__generic_unplug_device);
1449
1450 /**
1451  * generic_unplug_device - fire a request queue
1452  * @q:    The &request_queue_t in question
1453  *
1454  * Description:
1455  *   Linux uses plugging to build bigger requests queues before letting
1456  *   the device have at them. If a queue is plugged, the I/O scheduler
1457  *   is still adding and merging requests on the queue. Once the queue
1458  *   gets unplugged, the request_fn defined for the queue is invoked and
1459  *   transfers started.
1460  **/
1461 void generic_unplug_device(request_queue_t *q)
1462 {
1463         spin_lock_irq(q->queue_lock);
1464         __generic_unplug_device(q);
1465         spin_unlock_irq(q->queue_lock);
1466 }
1467 EXPORT_SYMBOL(generic_unplug_device);
1468
1469 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1470                                    struct page *page)
1471 {
1472         request_queue_t *q = bdi->unplug_io_data;
1473
1474         /*
1475          * devices don't necessarily have an ->unplug_fn defined
1476          */
1477         if (q->unplug_fn)
1478                 q->unplug_fn(q);
1479 }
1480
1481 static void blk_unplug_work(void *data)
1482 {
1483         request_queue_t *q = data;
1484
1485         q->unplug_fn(q);
1486 }
1487
1488 static void blk_unplug_timeout(unsigned long data)
1489 {
1490         request_queue_t *q = (request_queue_t *)data;
1491
1492         kblockd_schedule_work(&q->unplug_work);
1493 }
1494
1495 /**
1496  * blk_start_queue - restart a previously stopped queue
1497  * @q:    The &request_queue_t in question
1498  *
1499  * Description:
1500  *   blk_start_queue() will clear the stop flag on the queue, and call
1501  *   the request_fn for the queue if it was in a stopped state when
1502  *   entered. Also see blk_stop_queue(). Queue lock must be held.
1503  **/
1504 void blk_start_queue(request_queue_t *q)
1505 {
1506         clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1507
1508         /*
1509          * one level of recursion is ok and is much faster than kicking
1510          * the unplug handling
1511          */
1512         if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1513                 q->request_fn(q);
1514                 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1515         } else {
1516                 blk_plug_device(q);
1517                 kblockd_schedule_work(&q->unplug_work);
1518         }
1519 }
1520
1521 EXPORT_SYMBOL(blk_start_queue);
1522
1523 /**
1524  * blk_stop_queue - stop a queue
1525  * @q:    The &request_queue_t in question
1526  *
1527  * Description:
1528  *   The Linux block layer assumes that a block driver will consume all
1529  *   entries on the request queue when the request_fn strategy is called.
1530  *   Often this will not happen, because of hardware limitations (queue
1531  *   depth settings). If a device driver gets a 'queue full' response,
1532  *   or if it simply chooses not to queue more I/O at one point, it can
1533  *   call this function to prevent the request_fn from being called until
1534  *   the driver has signalled it's ready to go again. This happens by calling
1535  *   blk_start_queue() to restart queue operations. Queue lock must be held.
1536  **/
1537 void blk_stop_queue(request_queue_t *q)
1538 {
1539         blk_remove_plug(q);
1540         set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1541 }
1542 EXPORT_SYMBOL(blk_stop_queue);
1543
1544 /**
1545  * blk_sync_queue - cancel any pending callbacks on a queue
1546  * @q: the queue
1547  *
1548  * Description:
1549  *     The block layer may perform asynchronous callback activity
1550  *     on a queue, such as calling the unplug function after a timeout.
1551  *     A block device may call blk_sync_queue to ensure that any
1552  *     such activity is cancelled, thus allowing it to release resources
1553  *     the the callbacks might use. The caller must already have made sure
1554  *     that its ->make_request_fn will not re-add plugging prior to calling
1555  *     this function.
1556  *
1557  */
1558 void blk_sync_queue(struct request_queue *q)
1559 {
1560         del_timer_sync(&q->unplug_timer);
1561         kblockd_flush();
1562 }
1563 EXPORT_SYMBOL(blk_sync_queue);
1564
1565 /**
1566  * blk_run_queue - run a single device queue
1567  * @q:  The queue to run
1568  */
1569 void blk_run_queue(struct request_queue *q)
1570 {
1571         unsigned long flags;
1572
1573         spin_lock_irqsave(q->queue_lock, flags);
1574         blk_remove_plug(q);
1575         if (!elv_queue_empty(q))
1576                 q->request_fn(q);
1577         spin_unlock_irqrestore(q->queue_lock, flags);
1578 }
1579 EXPORT_SYMBOL(blk_run_queue);
1580
1581 /**
1582  * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1583  * @q:    the request queue to be released
1584  *
1585  * Description:
1586  *     blk_cleanup_queue is the pair to blk_init_queue() or
1587  *     blk_queue_make_request().  It should be called when a request queue is
1588  *     being released; typically when a block device is being de-registered.
1589  *     Currently, its primary task it to free all the &struct request
1590  *     structures that were allocated to the queue and the queue itself.
1591  *
1592  * Caveat:
1593  *     Hopefully the low level driver will have finished any
1594  *     outstanding requests first...
1595  **/
1596 void blk_cleanup_queue(request_queue_t * q)
1597 {
1598         struct request_list *rl = &q->rq;
1599
1600         if (!atomic_dec_and_test(&q->refcnt))
1601                 return;
1602
1603         if (q->elevator)
1604                 elevator_exit(q->elevator);
1605
1606         blk_sync_queue(q);
1607
1608         if (rl->rq_pool)
1609                 mempool_destroy(rl->rq_pool);
1610
1611         if (q->queue_tags)
1612                 __blk_queue_free_tags(q);
1613
1614         blk_queue_ordered(q, QUEUE_ORDERED_NONE);
1615
1616         kmem_cache_free(requestq_cachep, q);
1617 }
1618
1619 EXPORT_SYMBOL(blk_cleanup_queue);
1620
1621 static int blk_init_free_list(request_queue_t *q)
1622 {
1623         struct request_list *rl = &q->rq;
1624
1625         rl->count[READ] = rl->count[WRITE] = 0;
1626         rl->starved[READ] = rl->starved[WRITE] = 0;
1627         init_waitqueue_head(&rl->wait[READ]);
1628         init_waitqueue_head(&rl->wait[WRITE]);
1629         init_waitqueue_head(&rl->drain);
1630
1631         rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1632                                 mempool_free_slab, request_cachep, q->node);
1633
1634         if (!rl->rq_pool)
1635                 return -ENOMEM;
1636
1637         return 0;
1638 }
1639
1640 static int __make_request(request_queue_t *, struct bio *);
1641
1642 request_queue_t *blk_alloc_queue(int gfp_mask)
1643 {
1644         return blk_alloc_queue_node(gfp_mask, -1);
1645 }
1646 EXPORT_SYMBOL(blk_alloc_queue);
1647
1648 request_queue_t *blk_alloc_queue_node(int gfp_mask, int node_id)
1649 {
1650         request_queue_t *q;
1651
1652         q = kmem_cache_alloc_node(requestq_cachep, gfp_mask, node_id);
1653         if (!q)
1654                 return NULL;
1655
1656         memset(q, 0, sizeof(*q));
1657         init_timer(&q->unplug_timer);
1658         atomic_set(&q->refcnt, 1);
1659
1660         q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1661         q->backing_dev_info.unplug_io_data = q;
1662
1663         return q;
1664 }
1665 EXPORT_SYMBOL(blk_alloc_queue_node);
1666
1667 /**
1668  * blk_init_queue  - prepare a request queue for use with a block device
1669  * @rfn:  The function to be called to process requests that have been
1670  *        placed on the queue.
1671  * @lock: Request queue spin lock
1672  *
1673  * Description:
1674  *    If a block device wishes to use the standard request handling procedures,
1675  *    which sorts requests and coalesces adjacent requests, then it must
1676  *    call blk_init_queue().  The function @rfn will be called when there
1677  *    are requests on the queue that need to be processed.  If the device
1678  *    supports plugging, then @rfn may not be called immediately when requests
1679  *    are available on the queue, but may be called at some time later instead.
1680  *    Plugged queues are generally unplugged when a buffer belonging to one
1681  *    of the requests on the queue is needed, or due to memory pressure.
1682  *
1683  *    @rfn is not required, or even expected, to remove all requests off the
1684  *    queue, but only as many as it can handle at a time.  If it does leave
1685  *    requests on the queue, it is responsible for arranging that the requests
1686  *    get dealt with eventually.
1687  *
1688  *    The queue spin lock must be held while manipulating the requests on the
1689  *    request queue.
1690  *
1691  *    Function returns a pointer to the initialized request queue, or NULL if
1692  *    it didn't succeed.
1693  *
1694  * Note:
1695  *    blk_init_queue() must be paired with a blk_cleanup_queue() call
1696  *    when the block device is deactivated (such as at module unload).
1697  **/
1698
1699 request_queue_t *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1700 {
1701         return blk_init_queue_node(rfn, lock, -1);
1702 }
1703 EXPORT_SYMBOL(blk_init_queue);
1704
1705 request_queue_t *
1706 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1707 {
1708         request_queue_t *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1709
1710         if (!q)
1711                 return NULL;
1712
1713         q->node = node_id;
1714         if (blk_init_free_list(q))
1715                 goto out_init;
1716
1717         /*
1718          * if caller didn't supply a lock, they get per-queue locking with
1719          * our embedded lock
1720          */
1721         if (!lock) {
1722                 spin_lock_init(&q->__queue_lock);
1723                 lock = &q->__queue_lock;
1724         }
1725
1726         q->request_fn           = rfn;
1727         q->back_merge_fn        = ll_back_merge_fn;
1728         q->front_merge_fn       = ll_front_merge_fn;
1729         q->merge_requests_fn    = ll_merge_requests_fn;
1730         q->prep_rq_fn           = NULL;
1731         q->unplug_fn            = generic_unplug_device;
1732         q->queue_flags          = (1 << QUEUE_FLAG_CLUSTER);
1733         q->queue_lock           = lock;
1734
1735         blk_queue_segment_boundary(q, 0xffffffff);
1736
1737         blk_queue_make_request(q, __make_request);
1738         blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1739
1740         blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1741         blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1742
1743         /*
1744          * all done
1745          */
1746         if (!elevator_init(q, NULL)) {
1747                 blk_queue_congestion_threshold(q);
1748                 return q;
1749         }
1750
1751         blk_cleanup_queue(q);
1752 out_init:
1753         kmem_cache_free(requestq_cachep, q);
1754         return NULL;
1755 }
1756 EXPORT_SYMBOL(blk_init_queue_node);
1757
1758 int blk_get_queue(request_queue_t *q)
1759 {
1760         if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1761                 atomic_inc(&q->refcnt);
1762                 return 0;
1763         }
1764
1765         return 1;
1766 }
1767
1768 EXPORT_SYMBOL(blk_get_queue);
1769
1770 static inline void blk_free_request(request_queue_t *q, struct request *rq)
1771 {
1772         elv_put_request(q, rq);
1773         mempool_free(rq, q->rq.rq_pool);
1774 }
1775
1776 static inline struct request *
1777 blk_alloc_request(request_queue_t *q, int rw, struct bio *bio, int gfp_mask)
1778 {
1779         struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1780
1781         if (!rq)
1782                 return NULL;
1783
1784         /*
1785          * first three bits are identical in rq->flags and bio->bi_rw,
1786          * see bio.h and blkdev.h
1787          */
1788         rq->flags = rw;
1789
1790         if (!elv_set_request(q, rq, bio, gfp_mask))
1791                 return rq;
1792
1793         mempool_free(rq, q->rq.rq_pool);
1794         return NULL;
1795 }
1796
1797 /*
1798  * ioc_batching returns true if the ioc is a valid batching request and
1799  * should be given priority access to a request.
1800  */
1801 static inline int ioc_batching(request_queue_t *q, struct io_context *ioc)
1802 {
1803         if (!ioc)
1804                 return 0;
1805
1806         /*
1807          * Make sure the process is able to allocate at least 1 request
1808          * even if the batch times out, otherwise we could theoretically
1809          * lose wakeups.
1810          */
1811         return ioc->nr_batch_requests == q->nr_batching ||
1812                 (ioc->nr_batch_requests > 0
1813                 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
1814 }
1815
1816 /*
1817  * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
1818  * will cause the process to be a "batcher" on all queues in the system. This
1819  * is the behaviour we want though - once it gets a wakeup it should be given
1820  * a nice run.
1821  */
1822 static void ioc_set_batching(request_queue_t *q, struct io_context *ioc)
1823 {
1824         if (!ioc || ioc_batching(q, ioc))
1825                 return;
1826
1827         ioc->nr_batch_requests = q->nr_batching;
1828         ioc->last_waited = jiffies;
1829 }
1830
1831 static void __freed_request(request_queue_t *q, int rw)
1832 {
1833         struct request_list *rl = &q->rq;
1834
1835         if (rl->count[rw] < queue_congestion_off_threshold(q))
1836                 clear_queue_congested(q, rw);
1837
1838         if (rl->count[rw] + 1 <= q->nr_requests) {
1839                 if (waitqueue_active(&rl->wait[rw]))
1840                         wake_up(&rl->wait[rw]);
1841
1842                 blk_clear_queue_full(q, rw);
1843         }
1844 }
1845
1846 /*
1847  * A request has just been released.  Account for it, update the full and
1848  * congestion status, wake up any waiters.   Called under q->queue_lock.
1849  */
1850 static void freed_request(request_queue_t *q, int rw)
1851 {
1852         struct request_list *rl = &q->rq;
1853
1854         rl->count[rw]--;
1855
1856         __freed_request(q, rw);
1857
1858         if (unlikely(rl->starved[rw ^ 1]))
1859                 __freed_request(q, rw ^ 1);
1860
1861         if (!rl->count[READ] && !rl->count[WRITE]) {
1862                 smp_mb();
1863                 if (unlikely(waitqueue_active(&rl->drain)))
1864                         wake_up(&rl->drain);
1865         }
1866 }
1867
1868 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
1869 /*
1870  * Get a free request, queue_lock must be held.
1871  * Returns NULL on failure, with queue_lock held.
1872  * Returns !NULL on success, with queue_lock *not held*.
1873  */
1874 static struct request *get_request(request_queue_t *q, int rw, struct bio *bio,
1875                                    int gfp_mask)
1876 {
1877         struct request *rq = NULL;
1878         struct request_list *rl = &q->rq;
1879         struct io_context *ioc = current_io_context(GFP_ATOMIC);
1880
1881         if (unlikely(test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags)))
1882                 goto out;
1883
1884         if (rl->count[rw]+1 >= q->nr_requests) {
1885                 /*
1886                  * The queue will fill after this allocation, so set it as
1887                  * full, and mark this process as "batching". This process
1888                  * will be allowed to complete a batch of requests, others
1889                  * will be blocked.
1890                  */
1891                 if (!blk_queue_full(q, rw)) {
1892                         ioc_set_batching(q, ioc);
1893                         blk_set_queue_full(q, rw);
1894                 }
1895         }
1896
1897         switch (elv_may_queue(q, rw, bio)) {
1898                 case ELV_MQUEUE_NO:
1899                         goto rq_starved;
1900                 case ELV_MQUEUE_MAY:
1901                         break;
1902                 case ELV_MQUEUE_MUST:
1903                         goto get_rq;
1904         }
1905
1906         if (blk_queue_full(q, rw) && !ioc_batching(q, ioc)) {
1907                 /*
1908                  * The queue is full and the allocating process is not a
1909                  * "batcher", and not exempted by the IO scheduler
1910                  */
1911                 goto out;
1912         }
1913
1914 get_rq:
1915         /*
1916          * Only allow batching queuers to allocate up to 50% over the defined
1917          * limit of requests, otherwise we could have thousands of requests
1918          * allocated with any setting of ->nr_requests
1919          */
1920         if (rl->count[rw] >= (3 * q->nr_requests / 2))
1921                 goto out;
1922
1923         rl->count[rw]++;
1924         rl->starved[rw] = 0;
1925         if (rl->count[rw] >= queue_congestion_on_threshold(q))
1926                 set_queue_congested(q, rw);
1927         spin_unlock_irq(q->queue_lock);
1928
1929         rq = blk_alloc_request(q, rw, bio, gfp_mask);
1930         if (!rq) {
1931                 /*
1932                  * Allocation failed presumably due to memory. Undo anything
1933                  * we might have messed up.
1934                  *
1935                  * Allocating task should really be put onto the front of the
1936                  * wait queue, but this is pretty rare.
1937                  */
1938                 spin_lock_irq(q->queue_lock);
1939                 freed_request(q, rw);
1940
1941                 /*
1942                  * in the very unlikely event that allocation failed and no
1943                  * requests for this direction was pending, mark us starved
1944                  * so that freeing of a request in the other direction will
1945                  * notice us. another possible fix would be to split the
1946                  * rq mempool into READ and WRITE
1947                  */
1948 rq_starved:
1949                 if (unlikely(rl->count[rw] == 0))
1950                         rl->starved[rw] = 1;
1951
1952                 goto out;
1953         }
1954
1955         if (ioc_batching(q, ioc))
1956                 ioc->nr_batch_requests--;
1957         
1958         rq_init(q, rq);
1959         rq->rl = rl;
1960 out:
1961         return rq;
1962 }
1963
1964 /*
1965  * No available requests for this queue, unplug the device and wait for some
1966  * requests to become available.
1967  *
1968  * Called with q->queue_lock held, and returns with it unlocked.
1969  */
1970 static struct request *get_request_wait(request_queue_t *q, int rw,
1971                                         struct bio *bio)
1972 {
1973         struct request *rq;
1974
1975         rq = get_request(q, rw, bio, GFP_NOIO);
1976         while (!rq) {
1977                 DEFINE_WAIT(wait);
1978                 struct request_list *rl = &q->rq;
1979
1980                 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
1981                                 TASK_UNINTERRUPTIBLE);
1982
1983                 rq = get_request(q, rw, bio, GFP_NOIO);
1984
1985                 if (!rq) {
1986                         struct io_context *ioc;
1987
1988                         __generic_unplug_device(q);
1989                         spin_unlock_irq(q->queue_lock);
1990                         io_schedule();
1991
1992                         /*
1993                          * After sleeping, we become a "batching" process and
1994                          * will be able to allocate at least one request, and
1995                          * up to a big batch of them for a small period time.
1996                          * See ioc_batching, ioc_set_batching
1997                          */
1998                         ioc = current_io_context(GFP_NOIO);
1999                         ioc_set_batching(q, ioc);
2000
2001                         spin_lock_irq(q->queue_lock);
2002                 }
2003                 finish_wait(&rl->wait[rw], &wait);
2004         }
2005
2006         return rq;
2007 }
2008
2009 struct request *blk_get_request(request_queue_t *q, int rw, int gfp_mask)
2010 {
2011         struct request *rq;
2012
2013         BUG_ON(rw != READ && rw != WRITE);
2014
2015         spin_lock_irq(q->queue_lock);
2016         if (gfp_mask & __GFP_WAIT) {
2017                 rq = get_request_wait(q, rw, NULL);
2018         } else {
2019                 rq = get_request(q, rw, NULL, gfp_mask);
2020                 if (!rq)
2021                         spin_unlock_irq(q->queue_lock);
2022         }
2023         /* q->queue_lock is unlocked at this point */
2024
2025         return rq;
2026 }
2027 EXPORT_SYMBOL(blk_get_request);
2028
2029 /**
2030  * blk_requeue_request - put a request back on queue
2031  * @q:          request queue where request should be inserted
2032  * @rq:         request to be inserted
2033  *
2034  * Description:
2035  *    Drivers often keep queueing requests until the hardware cannot accept
2036  *    more, when that condition happens we need to put the request back
2037  *    on the queue. Must be called with queue lock held.
2038  */
2039 void blk_requeue_request(request_queue_t *q, struct request *rq)
2040 {
2041         if (blk_rq_tagged(rq))
2042                 blk_queue_end_tag(q, rq);
2043
2044         elv_requeue_request(q, rq);
2045 }
2046
2047 EXPORT_SYMBOL(blk_requeue_request);
2048
2049 /**
2050  * blk_insert_request - insert a special request in to a request queue
2051  * @q:          request queue where request should be inserted
2052  * @rq:         request to be inserted
2053  * @at_head:    insert request at head or tail of queue
2054  * @data:       private data
2055  *
2056  * Description:
2057  *    Many block devices need to execute commands asynchronously, so they don't
2058  *    block the whole kernel from preemption during request execution.  This is
2059  *    accomplished normally by inserting aritficial requests tagged as
2060  *    REQ_SPECIAL in to the corresponding request queue, and letting them be
2061  *    scheduled for actual execution by the request queue.
2062  *
2063  *    We have the option of inserting the head or the tail of the queue.
2064  *    Typically we use the tail for new ioctls and so forth.  We use the head
2065  *    of the queue for things like a QUEUE_FULL message from a device, or a
2066  *    host that is unable to accept a particular command.
2067  */
2068 void blk_insert_request(request_queue_t *q, struct request *rq,
2069                         int at_head, void *data)
2070 {
2071         int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2072         unsigned long flags;
2073
2074         /*
2075          * tell I/O scheduler that this isn't a regular read/write (ie it
2076          * must not attempt merges on this) and that it acts as a soft
2077          * barrier
2078          */
2079         rq->flags |= REQ_SPECIAL | REQ_SOFTBARRIER;
2080
2081         rq->special = data;
2082
2083         spin_lock_irqsave(q->queue_lock, flags);
2084
2085         /*
2086          * If command is tagged, release the tag
2087          */
2088         if (blk_rq_tagged(rq))
2089                 blk_queue_end_tag(q, rq);
2090
2091         drive_stat_acct(rq, rq->nr_sectors, 1);
2092         __elv_add_request(q, rq, where, 0);
2093
2094         if (blk_queue_plugged(q))
2095                 __generic_unplug_device(q);
2096         else
2097                 q->request_fn(q);
2098         spin_unlock_irqrestore(q->queue_lock, flags);
2099 }
2100
2101 EXPORT_SYMBOL(blk_insert_request);
2102
2103 /**
2104  * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2105  * @q:          request queue where request should be inserted
2106  * @rw:         READ or WRITE data
2107  * @ubuf:       the user buffer
2108  * @len:        length of user data
2109  *
2110  * Description:
2111  *    Data will be mapped directly for zero copy io, if possible. Otherwise
2112  *    a kernel bounce buffer is used.
2113  *
2114  *    A matching blk_rq_unmap_user() must be issued at the end of io, while
2115  *    still in process context.
2116  *
2117  *    Note: The mapped bio may need to be bounced through blk_queue_bounce()
2118  *    before being submitted to the device, as pages mapped may be out of
2119  *    reach. It's the callers responsibility to make sure this happens. The
2120  *    original bio must be passed back in to blk_rq_unmap_user() for proper
2121  *    unmapping.
2122  */
2123 struct request *blk_rq_map_user(request_queue_t *q, int rw, void __user *ubuf,
2124                                 unsigned int len)
2125 {
2126         unsigned long uaddr;
2127         struct request *rq;
2128         struct bio *bio;
2129
2130         if (len > (q->max_sectors << 9))
2131                 return ERR_PTR(-EINVAL);
2132         if ((!len && ubuf) || (len && !ubuf))
2133                 return ERR_PTR(-EINVAL);
2134
2135         rq = blk_get_request(q, rw, __GFP_WAIT);
2136         if (!rq)
2137                 return ERR_PTR(-ENOMEM);
2138
2139         /*
2140          * if alignment requirement is satisfied, map in user pages for
2141          * direct dma. else, set up kernel bounce buffers
2142          */
2143         uaddr = (unsigned long) ubuf;
2144         if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2145                 bio = bio_map_user(q, NULL, uaddr, len, rw == READ);
2146         else
2147                 bio = bio_copy_user(q, uaddr, len, rw == READ);
2148
2149         if (!IS_ERR(bio)) {
2150                 rq->bio = rq->biotail = bio;
2151                 blk_rq_bio_prep(q, rq, bio);
2152
2153                 rq->buffer = rq->data = NULL;
2154                 rq->data_len = len;
2155                 return rq;
2156         }
2157
2158         /*
2159          * bio is the err-ptr
2160          */
2161         blk_put_request(rq);
2162         return (struct request *) bio;
2163 }
2164
2165 EXPORT_SYMBOL(blk_rq_map_user);
2166
2167 /**
2168  * blk_rq_unmap_user - unmap a request with user data
2169  * @rq:         request to be unmapped
2170  * @bio:        bio for the request
2171  * @ulen:       length of user buffer
2172  *
2173  * Description:
2174  *    Unmap a request previously mapped by blk_rq_map_user().
2175  */
2176 int blk_rq_unmap_user(struct request *rq, struct bio *bio, unsigned int ulen)
2177 {
2178         int ret = 0;
2179
2180         if (bio) {
2181                 if (bio_flagged(bio, BIO_USER_MAPPED))
2182                         bio_unmap_user(bio);
2183                 else
2184                         ret = bio_uncopy_user(bio);
2185         }
2186
2187         blk_put_request(rq);
2188         return ret;
2189 }
2190
2191 EXPORT_SYMBOL(blk_rq_unmap_user);
2192
2193 /**
2194  * blk_execute_rq - insert a request into queue for execution
2195  * @q:          queue to insert the request in
2196  * @bd_disk:    matching gendisk
2197  * @rq:         request to insert
2198  *
2199  * Description:
2200  *    Insert a fully prepared request at the back of the io scheduler queue
2201  *    for execution.
2202  */
2203 int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
2204                    struct request *rq)
2205 {
2206         DECLARE_COMPLETION(wait);
2207         char sense[SCSI_SENSE_BUFFERSIZE];
2208         int err = 0;
2209
2210         rq->rq_disk = bd_disk;
2211
2212         /*
2213          * we need an extra reference to the request, so we can look at
2214          * it after io completion
2215          */
2216         rq->ref_count++;
2217
2218         if (!rq->sense) {
2219                 memset(sense, 0, sizeof(sense));
2220                 rq->sense = sense;
2221                 rq->sense_len = 0;
2222         }
2223
2224         rq->flags |= REQ_NOMERGE;
2225         rq->waiting = &wait;
2226         rq->end_io = blk_end_sync_rq;
2227         elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 1);
2228         generic_unplug_device(q);
2229         wait_for_completion(&wait);
2230         rq->waiting = NULL;
2231
2232         if (rq->errors)
2233                 err = -EIO;
2234
2235         return err;
2236 }
2237
2238 EXPORT_SYMBOL(blk_execute_rq);
2239
2240 /**
2241  * blkdev_issue_flush - queue a flush
2242  * @bdev:       blockdev to issue flush for
2243  * @error_sector:       error sector
2244  *
2245  * Description:
2246  *    Issue a flush for the block device in question. Caller can supply
2247  *    room for storing the error offset in case of a flush error, if they
2248  *    wish to.  Caller must run wait_for_completion() on its own.
2249  */
2250 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2251 {
2252         request_queue_t *q;
2253
2254         if (bdev->bd_disk == NULL)
2255                 return -ENXIO;
2256
2257         q = bdev_get_queue(bdev);
2258         if (!q)
2259                 return -ENXIO;
2260         if (!q->issue_flush_fn)
2261                 return -EOPNOTSUPP;
2262
2263         return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2264 }
2265
2266 EXPORT_SYMBOL(blkdev_issue_flush);
2267
2268 static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2269 {
2270         int rw = rq_data_dir(rq);
2271
2272         if (!blk_fs_request(rq) || !rq->rq_disk)
2273                 return;
2274
2275         if (rw == READ) {
2276                 __disk_stat_add(rq->rq_disk, read_sectors, nr_sectors);
2277                 if (!new_io)
2278                         __disk_stat_inc(rq->rq_disk, read_merges);
2279         } else if (rw == WRITE) {
2280                 __disk_stat_add(rq->rq_disk, write_sectors, nr_sectors);
2281                 if (!new_io)
2282                         __disk_stat_inc(rq->rq_disk, write_merges);
2283         }
2284         if (new_io) {
2285                 disk_round_stats(rq->rq_disk);
2286                 rq->rq_disk->in_flight++;
2287         }
2288 }
2289
2290 /*
2291  * add-request adds a request to the linked list.
2292  * queue lock is held and interrupts disabled, as we muck with the
2293  * request queue list.
2294  */
2295 static inline void add_request(request_queue_t * q, struct request * req)
2296 {
2297         drive_stat_acct(req, req->nr_sectors, 1);
2298
2299         if (q->activity_fn)
2300                 q->activity_fn(q->activity_data, rq_data_dir(req));
2301
2302         /*
2303          * elevator indicated where it wants this request to be
2304          * inserted at elevator_merge time
2305          */
2306         __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2307 }
2308  
2309 /*
2310  * disk_round_stats()   - Round off the performance stats on a struct
2311  * disk_stats.
2312  *
2313  * The average IO queue length and utilisation statistics are maintained
2314  * by observing the current state of the queue length and the amount of
2315  * time it has been in this state for.
2316  *
2317  * Normally, that accounting is done on IO completion, but that can result
2318  * in more than a second's worth of IO being accounted for within any one
2319  * second, leading to >100% utilisation.  To deal with that, we call this
2320  * function to do a round-off before returning the results when reading
2321  * /proc/diskstats.  This accounts immediately for all queue usage up to
2322  * the current jiffies and restarts the counters again.
2323  */
2324 void disk_round_stats(struct gendisk *disk)
2325 {
2326         unsigned long now = jiffies;
2327
2328         __disk_stat_add(disk, time_in_queue,
2329                         disk->in_flight * (now - disk->stamp));
2330         disk->stamp = now;
2331
2332         if (disk->in_flight)
2333                 __disk_stat_add(disk, io_ticks, (now - disk->stamp_idle));
2334         disk->stamp_idle = now;
2335 }
2336
2337 /*
2338  * queue lock must be held
2339  */
2340 static void __blk_put_request(request_queue_t *q, struct request *req)
2341 {
2342         struct request_list *rl = req->rl;
2343
2344         if (unlikely(!q))
2345                 return;
2346         if (unlikely(--req->ref_count))
2347                 return;
2348
2349         req->rq_status = RQ_INACTIVE;
2350         req->rl = NULL;
2351
2352         /*
2353          * Request may not have originated from ll_rw_blk. if not,
2354          * it didn't come out of our reserved rq pools
2355          */
2356         if (rl) {
2357                 int rw = rq_data_dir(req);
2358
2359                 elv_completed_request(q, req);
2360
2361                 BUG_ON(!list_empty(&req->queuelist));
2362
2363                 blk_free_request(q, req);
2364                 freed_request(q, rw);
2365         }
2366 }
2367
2368 void blk_put_request(struct request *req)
2369 {
2370         /*
2371          * if req->rl isn't set, this request didnt originate from the
2372          * block layer, so it's safe to just disregard it
2373          */
2374         if (req->rl) {
2375                 unsigned long flags;
2376                 request_queue_t *q = req->q;
2377
2378                 spin_lock_irqsave(q->queue_lock, flags);
2379                 __blk_put_request(q, req);
2380                 spin_unlock_irqrestore(q->queue_lock, flags);
2381         }
2382 }
2383
2384 EXPORT_SYMBOL(blk_put_request);
2385
2386 /**
2387  * blk_end_sync_rq - executes a completion event on a request
2388  * @rq: request to complete
2389  */
2390 void blk_end_sync_rq(struct request *rq)
2391 {
2392         struct completion *waiting = rq->waiting;
2393
2394         rq->waiting = NULL;
2395         __blk_put_request(rq->q, rq);
2396
2397         /*
2398          * complete last, if this is a stack request the process (and thus
2399          * the rq pointer) could be invalid right after this complete()
2400          */
2401         complete(waiting);
2402 }
2403 EXPORT_SYMBOL(blk_end_sync_rq);
2404
2405 /**
2406  * blk_congestion_wait - wait for a queue to become uncongested
2407  * @rw: READ or WRITE
2408  * @timeout: timeout in jiffies
2409  *
2410  * Waits for up to @timeout jiffies for a queue (any queue) to exit congestion.
2411  * If no queues are congested then just wait for the next request to be
2412  * returned.
2413  */
2414 long blk_congestion_wait(int rw, long timeout)
2415 {
2416         long ret;
2417         DEFINE_WAIT(wait);
2418         wait_queue_head_t *wqh = &congestion_wqh[rw];
2419
2420         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
2421         ret = io_schedule_timeout(timeout);
2422         finish_wait(wqh, &wait);
2423         return ret;
2424 }
2425
2426 EXPORT_SYMBOL(blk_congestion_wait);
2427
2428 /*
2429  * Has to be called with the request spinlock acquired
2430  */
2431 static int attempt_merge(request_queue_t *q, struct request *req,
2432                           struct request *next)
2433 {
2434         if (!rq_mergeable(req) || !rq_mergeable(next))
2435                 return 0;
2436
2437         /*
2438          * not contigious
2439          */
2440         if (req->sector + req->nr_sectors != next->sector)
2441                 return 0;
2442
2443         if (rq_data_dir(req) != rq_data_dir(next)
2444             || req->rq_disk != next->rq_disk
2445             || next->waiting || next->special)
2446                 return 0;
2447
2448         /*
2449          * If we are allowed to merge, then append bio list
2450          * from next to rq and release next. merge_requests_fn
2451          * will have updated segment counts, update sector
2452          * counts here.
2453          */
2454         if (!q->merge_requests_fn(q, req, next))
2455                 return 0;
2456
2457         /*
2458          * At this point we have either done a back merge
2459          * or front merge. We need the smaller start_time of
2460          * the merged requests to be the current request
2461          * for accounting purposes.
2462          */
2463         if (time_after(req->start_time, next->start_time))
2464                 req->start_time = next->start_time;
2465
2466         req->biotail->bi_next = next->bio;
2467         req->biotail = next->biotail;
2468
2469         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2470
2471         elv_merge_requests(q, req, next);
2472
2473         if (req->rq_disk) {
2474                 disk_round_stats(req->rq_disk);
2475                 req->rq_disk->in_flight--;
2476         }
2477
2478         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2479
2480         __blk_put_request(q, next);
2481         return 1;
2482 }
2483
2484 static inline int attempt_back_merge(request_queue_t *q, struct request *rq)
2485 {
2486         struct request *next = elv_latter_request(q, rq);
2487
2488         if (next)
2489                 return attempt_merge(q, rq, next);
2490
2491         return 0;
2492 }
2493
2494 static inline int attempt_front_merge(request_queue_t *q, struct request *rq)
2495 {
2496         struct request *prev = elv_former_request(q, rq);
2497
2498         if (prev)
2499                 return attempt_merge(q, prev, rq);
2500
2501         return 0;
2502 }
2503
2504 /**
2505  * blk_attempt_remerge  - attempt to remerge active head with next request
2506  * @q:    The &request_queue_t belonging to the device
2507  * @rq:   The head request (usually)
2508  *
2509  * Description:
2510  *    For head-active devices, the queue can easily be unplugged so quickly
2511  *    that proper merging is not done on the front request. This may hurt
2512  *    performance greatly for some devices. The block layer cannot safely
2513  *    do merging on that first request for these queues, but the driver can
2514  *    call this function and make it happen any way. Only the driver knows
2515  *    when it is safe to do so.
2516  **/
2517 void blk_attempt_remerge(request_queue_t *q, struct request *rq)
2518 {
2519         unsigned long flags;
2520
2521         spin_lock_irqsave(q->queue_lock, flags);
2522         attempt_back_merge(q, rq);
2523         spin_unlock_irqrestore(q->queue_lock, flags);
2524 }
2525
2526 EXPORT_SYMBOL(blk_attempt_remerge);
2527
2528 static int __make_request(request_queue_t *q, struct bio *bio)
2529 {
2530         struct request *req;
2531         int el_ret, rw, nr_sectors, cur_nr_sectors, barrier, err, sync;
2532         unsigned short prio;
2533         sector_t sector;
2534
2535         sector = bio->bi_sector;
2536         nr_sectors = bio_sectors(bio);
2537         cur_nr_sectors = bio_cur_sectors(bio);
2538         prio = bio_prio(bio);
2539
2540         rw = bio_data_dir(bio);
2541         sync = bio_sync(bio);
2542
2543         /*
2544          * low level driver can indicate that it wants pages above a
2545          * certain limit bounced to low memory (ie for highmem, or even
2546          * ISA dma in theory)
2547          */
2548         blk_queue_bounce(q, &bio);
2549
2550         spin_lock_prefetch(q->queue_lock);
2551
2552         barrier = bio_barrier(bio);
2553         if (unlikely(barrier) && (q->ordered == QUEUE_ORDERED_NONE)) {
2554                 err = -EOPNOTSUPP;
2555                 goto end_io;
2556         }
2557
2558         spin_lock_irq(q->queue_lock);
2559
2560         if (unlikely(barrier) || elv_queue_empty(q))
2561                 goto get_rq;
2562
2563         el_ret = elv_merge(q, &req, bio);
2564         switch (el_ret) {
2565                 case ELEVATOR_BACK_MERGE:
2566                         BUG_ON(!rq_mergeable(req));
2567
2568                         if (!q->back_merge_fn(q, req, bio))
2569                                 break;
2570
2571                         req->biotail->bi_next = bio;
2572                         req->biotail = bio;
2573                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2574                         req->ioprio = ioprio_best(req->ioprio, prio);
2575                         drive_stat_acct(req, nr_sectors, 0);
2576                         if (!attempt_back_merge(q, req))
2577                                 elv_merged_request(q, req);
2578                         goto out;
2579
2580                 case ELEVATOR_FRONT_MERGE:
2581                         BUG_ON(!rq_mergeable(req));
2582
2583                         if (!q->front_merge_fn(q, req, bio))
2584                                 break;
2585
2586                         bio->bi_next = req->bio;
2587                         req->bio = bio;
2588
2589                         /*
2590                          * may not be valid. if the low level driver said
2591                          * it didn't need a bounce buffer then it better
2592                          * not touch req->buffer either...
2593                          */
2594                         req->buffer = bio_data(bio);
2595                         req->current_nr_sectors = cur_nr_sectors;
2596                         req->hard_cur_sectors = cur_nr_sectors;
2597                         req->sector = req->hard_sector = sector;
2598                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2599                         req->ioprio = ioprio_best(req->ioprio, prio);
2600                         drive_stat_acct(req, nr_sectors, 0);
2601                         if (!attempt_front_merge(q, req))
2602                                 elv_merged_request(q, req);
2603                         goto out;
2604
2605                 /* ELV_NO_MERGE: elevator says don't/can't merge. */
2606                 default:
2607                         ;
2608         }
2609
2610 get_rq:
2611         /*
2612          * Grab a free request. This is might sleep but can not fail.
2613          * Returns with the queue unlocked.
2614          */
2615         req = get_request_wait(q, rw, bio);
2616
2617         /*
2618          * After dropping the lock and possibly sleeping here, our request
2619          * may now be mergeable after it had proven unmergeable (above).
2620          * We don't worry about that case for efficiency. It won't happen
2621          * often, and the elevators are able to handle it.
2622          */
2623
2624         req->flags |= REQ_CMD;
2625
2626         /*
2627          * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2628          */
2629         if (bio_rw_ahead(bio) || bio_failfast(bio))
2630                 req->flags |= REQ_FAILFAST;
2631
2632         /*
2633          * REQ_BARRIER implies no merging, but lets make it explicit
2634          */
2635         if (unlikely(barrier))
2636                 req->flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2637
2638         req->errors = 0;
2639         req->hard_sector = req->sector = sector;
2640         req->hard_nr_sectors = req->nr_sectors = nr_sectors;
2641         req->current_nr_sectors = req->hard_cur_sectors = cur_nr_sectors;
2642         req->nr_phys_segments = bio_phys_segments(q, bio);
2643         req->nr_hw_segments = bio_hw_segments(q, bio);
2644         req->buffer = bio_data(bio);    /* see ->buffer comment above */
2645         req->waiting = NULL;
2646         req->bio = req->biotail = bio;
2647         req->ioprio = prio;
2648         req->rq_disk = bio->bi_bdev->bd_disk;
2649         req->start_time = jiffies;
2650
2651         spin_lock_irq(q->queue_lock);
2652         if (elv_queue_empty(q))
2653                 blk_plug_device(q);
2654         add_request(q, req);
2655 out:
2656         if (sync)
2657                 __generic_unplug_device(q);
2658
2659         spin_unlock_irq(q->queue_lock);
2660         return 0;
2661
2662 end_io:
2663         bio_endio(bio, nr_sectors << 9, err);
2664         return 0;
2665 }
2666
2667 /*
2668  * If bio->bi_dev is a partition, remap the location
2669  */
2670 static inline void blk_partition_remap(struct bio *bio)
2671 {
2672         struct block_device *bdev = bio->bi_bdev;
2673
2674         if (bdev != bdev->bd_contains) {
2675                 struct hd_struct *p = bdev->bd_part;
2676
2677                 switch (bio_data_dir(bio)) {
2678                 case READ:
2679                         p->read_sectors += bio_sectors(bio);
2680                         p->reads++;
2681                         break;
2682                 case WRITE:
2683                         p->write_sectors += bio_sectors(bio);
2684                         p->writes++;
2685                         break;
2686                 }
2687                 bio->bi_sector += p->start_sect;
2688                 bio->bi_bdev = bdev->bd_contains;
2689         }
2690 }
2691
2692 void blk_finish_queue_drain(request_queue_t *q)
2693 {
2694         struct request_list *rl = &q->rq;
2695         struct request *rq;
2696         int requeued = 0;
2697
2698         spin_lock_irq(q->queue_lock);
2699         clear_bit(QUEUE_FLAG_DRAIN, &q->queue_flags);
2700
2701         while (!list_empty(&q->drain_list)) {
2702                 rq = list_entry_rq(q->drain_list.next);
2703
2704                 list_del_init(&rq->queuelist);
2705                 elv_requeue_request(q, rq);
2706                 requeued++;
2707         }
2708
2709         if (requeued)
2710                 q->request_fn(q);
2711
2712         spin_unlock_irq(q->queue_lock);
2713
2714         wake_up(&rl->wait[0]);
2715         wake_up(&rl->wait[1]);
2716         wake_up(&rl->drain);
2717 }
2718
2719 static int wait_drain(request_queue_t *q, struct request_list *rl, int dispatch)
2720 {
2721         int wait = rl->count[READ] + rl->count[WRITE];
2722
2723         if (dispatch)
2724                 wait += !list_empty(&q->queue_head);
2725
2726         return wait;
2727 }
2728
2729 /*
2730  * We rely on the fact that only requests allocated through blk_alloc_request()
2731  * have io scheduler private data structures associated with them. Any other
2732  * type of request (allocated on stack or through kmalloc()) should not go
2733  * to the io scheduler core, but be attached to the queue head instead.
2734  */
2735 void blk_wait_queue_drained(request_queue_t *q, int wait_dispatch)
2736 {
2737         struct request_list *rl = &q->rq;
2738         DEFINE_WAIT(wait);
2739
2740         spin_lock_irq(q->queue_lock);
2741         set_bit(QUEUE_FLAG_DRAIN, &q->queue_flags);
2742
2743         while (wait_drain(q, rl, wait_dispatch)) {
2744                 prepare_to_wait(&rl->drain, &wait, TASK_UNINTERRUPTIBLE);
2745
2746                 if (wait_drain(q, rl, wait_dispatch)) {
2747                         __generic_unplug_device(q);
2748                         spin_unlock_irq(q->queue_lock);
2749                         io_schedule();
2750                         spin_lock_irq(q->queue_lock);
2751                 }
2752
2753                 finish_wait(&rl->drain, &wait);
2754         }
2755
2756         spin_unlock_irq(q->queue_lock);
2757 }
2758
2759 /*
2760  * block waiting for the io scheduler being started again.
2761  */
2762 static inline void block_wait_queue_running(request_queue_t *q)
2763 {
2764         DEFINE_WAIT(wait);
2765
2766         while (unlikely(test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags))) {
2767                 struct request_list *rl = &q->rq;
2768
2769                 prepare_to_wait_exclusive(&rl->drain, &wait,
2770                                 TASK_UNINTERRUPTIBLE);
2771
2772                 /*
2773                  * re-check the condition. avoids using prepare_to_wait()
2774                  * in the fast path (queue is running)
2775                  */
2776                 if (test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags))
2777                         io_schedule();
2778
2779                 finish_wait(&rl->drain, &wait);
2780         }
2781 }
2782
2783 static void handle_bad_sector(struct bio *bio)
2784 {
2785         char b[BDEVNAME_SIZE];
2786
2787         printk(KERN_INFO "attempt to access beyond end of device\n");
2788         printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
2789                         bdevname(bio->bi_bdev, b),
2790                         bio->bi_rw,
2791                         (unsigned long long)bio->bi_sector + bio_sectors(bio),
2792                         (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
2793
2794         set_bit(BIO_EOF, &bio->bi_flags);
2795 }
2796
2797 /**
2798  * generic_make_request: hand a buffer to its device driver for I/O
2799  * @bio:  The bio describing the location in memory and on the device.
2800  *
2801  * generic_make_request() is used to make I/O requests of block
2802  * devices. It is passed a &struct bio, which describes the I/O that needs
2803  * to be done.
2804  *
2805  * generic_make_request() does not return any status.  The
2806  * success/failure status of the request, along with notification of
2807  * completion, is delivered asynchronously through the bio->bi_end_io
2808  * function described (one day) else where.
2809  *
2810  * The caller of generic_make_request must make sure that bi_io_vec
2811  * are set to describe the memory buffer, and that bi_dev and bi_sector are
2812  * set to describe the device address, and the
2813  * bi_end_io and optionally bi_private are set to describe how
2814  * completion notification should be signaled.
2815  *
2816  * generic_make_request and the drivers it calls may use bi_next if this
2817  * bio happens to be merged with someone else, and may change bi_dev and
2818  * bi_sector for remaps as it sees fit.  So the values of these fields
2819  * should NOT be depended on after the call to generic_make_request.
2820  */
2821 void generic_make_request(struct bio *bio)
2822 {
2823         request_queue_t *q;
2824         sector_t maxsector;
2825         int ret, nr_sectors = bio_sectors(bio);
2826
2827         might_sleep();
2828         /* Test device or partition size, when known. */
2829         maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
2830         if (maxsector) {
2831                 sector_t sector = bio->bi_sector;
2832
2833                 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
2834                         /*
2835                          * This may well happen - the kernel calls bread()
2836                          * without checking the size of the device, e.g., when
2837                          * mounting a device.
2838                          */
2839                         handle_bad_sector(bio);
2840                         goto end_io;
2841                 }
2842         }
2843
2844         /*
2845          * Resolve the mapping until finished. (drivers are
2846          * still free to implement/resolve their own stacking
2847          * by explicitly returning 0)
2848          *
2849          * NOTE: we don't repeat the blk_size check for each new device.
2850          * Stacking drivers are expected to know what they are doing.
2851          */
2852         do {
2853                 char b[BDEVNAME_SIZE];
2854
2855                 q = bdev_get_queue(bio->bi_bdev);
2856                 if (!q) {
2857                         printk(KERN_ERR
2858                                "generic_make_request: Trying to access "
2859                                 "nonexistent block-device %s (%Lu)\n",
2860                                 bdevname(bio->bi_bdev, b),
2861                                 (long long) bio->bi_sector);
2862 end_io:
2863                         bio_endio(bio, bio->bi_size, -EIO);
2864                         break;
2865                 }
2866
2867                 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
2868                         printk("bio too big device %s (%u > %u)\n", 
2869                                 bdevname(bio->bi_bdev, b),
2870                                 bio_sectors(bio),
2871                                 q->max_hw_sectors);
2872                         goto end_io;
2873                 }
2874
2875                 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
2876                         goto end_io;
2877
2878                 block_wait_queue_running(q);
2879
2880                 /*
2881                  * If this device has partitions, remap block n
2882                  * of partition p to block n+start(p) of the disk.
2883                  */
2884                 blk_partition_remap(bio);
2885
2886                 ret = q->make_request_fn(q, bio);
2887         } while (ret);
2888 }
2889
2890 EXPORT_SYMBOL(generic_make_request);
2891
2892 /**
2893  * submit_bio: submit a bio to the block device layer for I/O
2894  * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
2895  * @bio: The &struct bio which describes the I/O
2896  *
2897  * submit_bio() is very similar in purpose to generic_make_request(), and
2898  * uses that function to do most of the work. Both are fairly rough
2899  * interfaces, @bio must be presetup and ready for I/O.
2900  *
2901  */
2902 void submit_bio(int rw, struct bio *bio)
2903 {
2904         int count = bio_sectors(bio);
2905
2906         BIO_BUG_ON(!bio->bi_size);
2907         BIO_BUG_ON(!bio->bi_io_vec);
2908         bio->bi_rw |= rw;
2909         if (rw & WRITE)
2910                 mod_page_state(pgpgout, count);
2911         else
2912                 mod_page_state(pgpgin, count);
2913
2914         if (unlikely(block_dump)) {
2915                 char b[BDEVNAME_SIZE];
2916                 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
2917                         current->comm, current->pid,
2918                         (rw & WRITE) ? "WRITE" : "READ",
2919                         (unsigned long long)bio->bi_sector,
2920                         bdevname(bio->bi_bdev,b));
2921         }
2922
2923         generic_make_request(bio);
2924 }
2925
2926 EXPORT_SYMBOL(submit_bio);
2927
2928 static void blk_recalc_rq_segments(struct request *rq)
2929 {
2930         struct bio *bio, *prevbio = NULL;
2931         int nr_phys_segs, nr_hw_segs;
2932         unsigned int phys_size, hw_size;
2933         request_queue_t *q = rq->q;
2934
2935         if (!rq->bio)
2936                 return;
2937
2938         phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
2939         rq_for_each_bio(bio, rq) {
2940                 /* Force bio hw/phys segs to be recalculated. */
2941                 bio->bi_flags &= ~(1 << BIO_SEG_VALID);
2942
2943                 nr_phys_segs += bio_phys_segments(q, bio);
2944                 nr_hw_segs += bio_hw_segments(q, bio);
2945                 if (prevbio) {
2946                         int pseg = phys_size + prevbio->bi_size + bio->bi_size;
2947                         int hseg = hw_size + prevbio->bi_size + bio->bi_size;
2948
2949                         if (blk_phys_contig_segment(q, prevbio, bio) &&
2950                             pseg <= q->max_segment_size) {
2951                                 nr_phys_segs--;
2952                                 phys_size += prevbio->bi_size + bio->bi_size;
2953                         } else
2954                                 phys_size = 0;
2955
2956                         if (blk_hw_contig_segment(q, prevbio, bio) &&
2957                             hseg <= q->max_segment_size) {
2958                                 nr_hw_segs--;
2959                                 hw_size += prevbio->bi_size + bio->bi_size;
2960                         } else
2961                                 hw_size = 0;
2962                 }
2963                 prevbio = bio;
2964         }
2965
2966         rq->nr_phys_segments = nr_phys_segs;
2967         rq->nr_hw_segments = nr_hw_segs;
2968 }
2969
2970 static void blk_recalc_rq_sectors(struct request *rq, int nsect)
2971 {
2972         if (blk_fs_request(rq)) {
2973                 rq->hard_sector += nsect;
2974                 rq->hard_nr_sectors -= nsect;
2975
2976                 /*
2977                  * Move the I/O submission pointers ahead if required.
2978                  */
2979                 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
2980                     (rq->sector <= rq->hard_sector)) {
2981                         rq->sector = rq->hard_sector;
2982                         rq->nr_sectors = rq->hard_nr_sectors;
2983                         rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
2984                         rq->current_nr_sectors = rq->hard_cur_sectors;
2985                         rq->buffer = bio_data(rq->bio);
2986                 }
2987
2988                 /*
2989                  * if total number of sectors is less than the first segment
2990                  * size, something has gone terribly wrong
2991                  */
2992                 if (rq->nr_sectors < rq->current_nr_sectors) {
2993                         printk("blk: request botched\n");
2994                         rq->nr_sectors = rq->current_nr_sectors;
2995                 }
2996         }
2997 }
2998
2999 static int __end_that_request_first(struct request *req, int uptodate,
3000                                     int nr_bytes)
3001 {
3002         int total_bytes, bio_nbytes, error, next_idx = 0;
3003         struct bio *bio;
3004
3005         /*
3006          * extend uptodate bool to allow < 0 value to be direct io error
3007          */
3008         error = 0;
3009         if (end_io_error(uptodate))
3010                 error = !uptodate ? -EIO : uptodate;
3011
3012         /*
3013          * for a REQ_BLOCK_PC request, we want to carry any eventual
3014          * sense key with us all the way through
3015          */
3016         if (!blk_pc_request(req))
3017                 req->errors = 0;
3018
3019         if (!uptodate) {
3020                 if (blk_fs_request(req) && !(req->flags & REQ_QUIET))
3021                         printk("end_request: I/O error, dev %s, sector %llu\n",
3022                                 req->rq_disk ? req->rq_disk->disk_name : "?",
3023                                 (unsigned long long)req->sector);
3024         }
3025
3026         total_bytes = bio_nbytes = 0;
3027         while ((bio = req->bio) != NULL) {
3028                 int nbytes;
3029
3030                 if (nr_bytes >= bio->bi_size) {
3031                         req->bio = bio->bi_next;
3032                         nbytes = bio->bi_size;
3033                         bio_endio(bio, nbytes, error);
3034                         next_idx = 0;
3035                         bio_nbytes = 0;
3036                 } else {
3037                         int idx = bio->bi_idx + next_idx;
3038
3039                         if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3040                                 blk_dump_rq_flags(req, "__end_that");
3041                                 printk("%s: bio idx %d >= vcnt %d\n",
3042                                                 __FUNCTION__,
3043                                                 bio->bi_idx, bio->bi_vcnt);
3044                                 break;
3045                         }
3046
3047                         nbytes = bio_iovec_idx(bio, idx)->bv_len;
3048                         BIO_BUG_ON(nbytes > bio->bi_size);
3049
3050                         /*
3051                          * not a complete bvec done
3052                          */
3053                         if (unlikely(nbytes > nr_bytes)) {
3054                                 bio_nbytes += nr_bytes;
3055                                 total_bytes += nr_bytes;
3056                                 break;
3057                         }
3058
3059                         /*
3060                          * advance to the next vector
3061                          */
3062                         next_idx++;
3063                         bio_nbytes += nbytes;
3064                 }
3065
3066                 total_bytes += nbytes;
3067                 nr_bytes -= nbytes;
3068
3069                 if ((bio = req->bio)) {
3070                         /*
3071                          * end more in this run, or just return 'not-done'
3072                          */
3073                         if (unlikely(nr_bytes <= 0))
3074                                 break;
3075                 }
3076         }
3077
3078         /*
3079          * completely done
3080          */
3081         if (!req->bio)
3082                 return 0;
3083
3084         /*
3085          * if the request wasn't completed, update state
3086          */
3087         if (bio_nbytes) {
3088                 bio_endio(bio, bio_nbytes, error);
3089                 bio->bi_idx += next_idx;
3090                 bio_iovec(bio)->bv_offset += nr_bytes;
3091                 bio_iovec(bio)->bv_len -= nr_bytes;
3092         }
3093
3094         blk_recalc_rq_sectors(req, total_bytes >> 9);
3095         blk_recalc_rq_segments(req);
3096         return 1;
3097 }
3098
3099 /**
3100  * end_that_request_first - end I/O on a request
3101  * @req:      the request being processed
3102  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3103  * @nr_sectors: number of sectors to end I/O on
3104  *
3105  * Description:
3106  *     Ends I/O on a number of sectors attached to @req, and sets it up
3107  *     for the next range of segments (if any) in the cluster.
3108  *
3109  * Return:
3110  *     0 - we are done with this request, call end_that_request_last()
3111  *     1 - still buffers pending for this request
3112  **/
3113 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3114 {
3115         return __end_that_request_first(req, uptodate, nr_sectors << 9);
3116 }
3117
3118 EXPORT_SYMBOL(end_that_request_first);
3119
3120 /**
3121  * end_that_request_chunk - end I/O on a request
3122  * @req:      the request being processed
3123  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3124  * @nr_bytes: number of bytes to complete
3125  *
3126  * Description:
3127  *     Ends I/O on a number of bytes attached to @req, and sets it up
3128  *     for the next range of segments (if any). Like end_that_request_first(),
3129  *     but deals with bytes instead of sectors.
3130  *
3131  * Return:
3132  *     0 - we are done with this request, call end_that_request_last()
3133  *     1 - still buffers pending for this request
3134  **/
3135 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3136 {
3137         return __end_that_request_first(req, uptodate, nr_bytes);
3138 }
3139
3140 EXPORT_SYMBOL(end_that_request_chunk);
3141
3142 /*
3143  * queue lock must be held
3144  */
3145 void end_that_request_last(struct request *req)
3146 {
3147         struct gendisk *disk = req->rq_disk;
3148
3149         if (unlikely(laptop_mode) && blk_fs_request(req))
3150                 laptop_io_completion();
3151
3152         if (disk && blk_fs_request(req)) {
3153                 unsigned long duration = jiffies - req->start_time;
3154                 switch (rq_data_dir(req)) {
3155                     case WRITE:
3156                         __disk_stat_inc(disk, writes);
3157                         __disk_stat_add(disk, write_ticks, duration);
3158                         break;
3159                     case READ:
3160                         __disk_stat_inc(disk, reads);
3161                         __disk_stat_add(disk, read_ticks, duration);
3162                         break;
3163                 }
3164                 disk_round_stats(disk);
3165                 disk->in_flight--;
3166         }
3167         if (req->end_io)
3168                 req->end_io(req);
3169         else
3170                 __blk_put_request(req->q, req);
3171 }
3172
3173 EXPORT_SYMBOL(end_that_request_last);
3174
3175 void end_request(struct request *req, int uptodate)
3176 {
3177         if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3178                 add_disk_randomness(req->rq_disk);
3179                 blkdev_dequeue_request(req);
3180                 end_that_request_last(req);
3181         }
3182 }
3183
3184 EXPORT_SYMBOL(end_request);
3185
3186 void blk_rq_bio_prep(request_queue_t *q, struct request *rq, struct bio *bio)
3187 {
3188         /* first three bits are identical in rq->flags and bio->bi_rw */
3189         rq->flags |= (bio->bi_rw & 7);
3190
3191         rq->nr_phys_segments = bio_phys_segments(q, bio);
3192         rq->nr_hw_segments = bio_hw_segments(q, bio);
3193         rq->current_nr_sectors = bio_cur_sectors(bio);
3194         rq->hard_cur_sectors = rq->current_nr_sectors;
3195         rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3196         rq->buffer = bio_data(bio);
3197
3198         rq->bio = rq->biotail = bio;
3199 }
3200
3201 EXPORT_SYMBOL(blk_rq_bio_prep);
3202
3203 int kblockd_schedule_work(struct work_struct *work)
3204 {
3205         return queue_work(kblockd_workqueue, work);
3206 }
3207
3208 EXPORT_SYMBOL(kblockd_schedule_work);
3209
3210 void kblockd_flush(void)
3211 {
3212         flush_workqueue(kblockd_workqueue);
3213 }
3214 EXPORT_SYMBOL(kblockd_flush);
3215
3216 int __init blk_dev_init(void)
3217 {
3218         kblockd_workqueue = create_workqueue("kblockd");
3219         if (!kblockd_workqueue)
3220                 panic("Failed to create kblockd\n");
3221
3222         request_cachep = kmem_cache_create("blkdev_requests",
3223                         sizeof(struct request), 0, SLAB_PANIC, NULL, NULL);
3224
3225         requestq_cachep = kmem_cache_create("blkdev_queue",
3226                         sizeof(request_queue_t), 0, SLAB_PANIC, NULL, NULL);
3227
3228         iocontext_cachep = kmem_cache_create("blkdev_ioc",
3229                         sizeof(struct io_context), 0, SLAB_PANIC, NULL, NULL);
3230
3231         blk_max_low_pfn = max_low_pfn;
3232         blk_max_pfn = max_pfn;
3233
3234         return 0;
3235 }
3236
3237 /*
3238  * IO Context helper functions
3239  */
3240 void put_io_context(struct io_context *ioc)
3241 {
3242         if (ioc == NULL)
3243                 return;
3244
3245         BUG_ON(atomic_read(&ioc->refcount) == 0);
3246
3247         if (atomic_dec_and_test(&ioc->refcount)) {
3248                 if (ioc->aic && ioc->aic->dtor)
3249                         ioc->aic->dtor(ioc->aic);
3250                 if (ioc->cic && ioc->cic->dtor)
3251                         ioc->cic->dtor(ioc->cic);
3252
3253                 kmem_cache_free(iocontext_cachep, ioc);
3254         }
3255 }
3256 EXPORT_SYMBOL(put_io_context);
3257
3258 /* Called by the exitting task */
3259 void exit_io_context(void)
3260 {
3261         unsigned long flags;
3262         struct io_context *ioc;
3263
3264         local_irq_save(flags);
3265         task_lock(current);
3266         ioc = current->io_context;
3267         current->io_context = NULL;
3268         ioc->task = NULL;
3269         task_unlock(current);
3270         local_irq_restore(flags);
3271
3272         if (ioc->aic && ioc->aic->exit)
3273                 ioc->aic->exit(ioc->aic);
3274         if (ioc->cic && ioc->cic->exit)
3275                 ioc->cic->exit(ioc->cic);
3276
3277         put_io_context(ioc);
3278 }
3279
3280 /*
3281  * If the current task has no IO context then create one and initialise it.
3282  * Otherwise, return its existing IO context.
3283  *
3284  * This returned IO context doesn't have a specifically elevated refcount,
3285  * but since the current task itself holds a reference, the context can be
3286  * used in general code, so long as it stays within `current` context.
3287  */
3288 struct io_context *current_io_context(int gfp_flags)
3289 {
3290         struct task_struct *tsk = current;
3291         struct io_context *ret;
3292
3293         ret = tsk->io_context;
3294         if (likely(ret))
3295                 return ret;
3296
3297         ret = kmem_cache_alloc(iocontext_cachep, gfp_flags);
3298         if (ret) {
3299                 atomic_set(&ret->refcount, 1);
3300                 ret->task = current;
3301                 ret->set_ioprio = NULL;
3302                 ret->last_waited = jiffies; /* doesn't matter... */
3303                 ret->nr_batch_requests = 0; /* because this is 0 */
3304                 ret->aic = NULL;
3305                 ret->cic = NULL;
3306                 tsk->io_context = ret;
3307         }
3308
3309         return ret;
3310 }
3311 EXPORT_SYMBOL(current_io_context);
3312
3313 /*
3314  * If the current task has no IO context then create one and initialise it.
3315  * If it does have a context, take a ref on it.
3316  *
3317  * This is always called in the context of the task which submitted the I/O.
3318  */
3319 struct io_context *get_io_context(int gfp_flags)
3320 {
3321         struct io_context *ret;
3322         ret = current_io_context(gfp_flags);
3323         if (likely(ret))
3324                 atomic_inc(&ret->refcount);
3325         return ret;
3326 }
3327 EXPORT_SYMBOL(get_io_context);
3328
3329 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3330 {
3331         struct io_context *src = *psrc;
3332         struct io_context *dst = *pdst;
3333
3334         if (src) {
3335                 BUG_ON(atomic_read(&src->refcount) == 0);
3336                 atomic_inc(&src->refcount);
3337                 put_io_context(dst);
3338                 *pdst = src;
3339         }
3340 }
3341 EXPORT_SYMBOL(copy_io_context);
3342
3343 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3344 {
3345         struct io_context *temp;
3346         temp = *ioc1;
3347         *ioc1 = *ioc2;
3348         *ioc2 = temp;
3349 }
3350 EXPORT_SYMBOL(swap_io_context);
3351
3352 /*
3353  * sysfs parts below
3354  */
3355 struct queue_sysfs_entry {
3356         struct attribute attr;
3357         ssize_t (*show)(struct request_queue *, char *);
3358         ssize_t (*store)(struct request_queue *, const char *, size_t);
3359 };
3360
3361 static ssize_t
3362 queue_var_show(unsigned int var, char *page)
3363 {
3364         return sprintf(page, "%d\n", var);
3365 }
3366
3367 static ssize_t
3368 queue_var_store(unsigned long *var, const char *page, size_t count)
3369 {
3370         char *p = (char *) page;
3371
3372         *var = simple_strtoul(p, &p, 10);
3373         return count;
3374 }
3375
3376 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3377 {
3378         return queue_var_show(q->nr_requests, (page));
3379 }
3380
3381 static ssize_t
3382 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3383 {
3384         struct request_list *rl = &q->rq;
3385
3386         int ret = queue_var_store(&q->nr_requests, page, count);
3387         if (q->nr_requests < BLKDEV_MIN_RQ)
3388                 q->nr_requests = BLKDEV_MIN_RQ;
3389         blk_queue_congestion_threshold(q);
3390
3391         if (rl->count[READ] >= queue_congestion_on_threshold(q))
3392                 set_queue_congested(q, READ);
3393         else if (rl->count[READ] < queue_congestion_off_threshold(q))
3394                 clear_queue_congested(q, READ);
3395
3396         if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3397                 set_queue_congested(q, WRITE);
3398         else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3399                 clear_queue_congested(q, WRITE);
3400
3401         if (rl->count[READ] >= q->nr_requests) {
3402                 blk_set_queue_full(q, READ);
3403         } else if (rl->count[READ]+1 <= q->nr_requests) {
3404                 blk_clear_queue_full(q, READ);
3405                 wake_up(&rl->wait[READ]);
3406         }
3407
3408         if (rl->count[WRITE] >= q->nr_requests) {
3409                 blk_set_queue_full(q, WRITE);
3410         } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3411                 blk_clear_queue_full(q, WRITE);
3412                 wake_up(&rl->wait[WRITE]);
3413         }
3414         return ret;
3415 }
3416
3417 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3418 {
3419         int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3420
3421         return queue_var_show(ra_kb, (page));
3422 }
3423
3424 static ssize_t
3425 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3426 {
3427         unsigned long ra_kb;
3428         ssize_t ret = queue_var_store(&ra_kb, page, count);
3429
3430         spin_lock_irq(q->queue_lock);
3431         if (ra_kb > (q->max_sectors >> 1))
3432                 ra_kb = (q->max_sectors >> 1);
3433
3434         q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3435         spin_unlock_irq(q->queue_lock);
3436
3437         return ret;
3438 }
3439
3440 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3441 {
3442         int max_sectors_kb = q->max_sectors >> 1;
3443
3444         return queue_var_show(max_sectors_kb, (page));
3445 }
3446
3447 static ssize_t
3448 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3449 {
3450         unsigned long max_sectors_kb,
3451                         max_hw_sectors_kb = q->max_hw_sectors >> 1,
3452                         page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3453         ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3454         int ra_kb;
3455
3456         if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3457                 return -EINVAL;
3458         /*
3459          * Take the queue lock to update the readahead and max_sectors
3460          * values synchronously:
3461          */
3462         spin_lock_irq(q->queue_lock);
3463         /*
3464          * Trim readahead window as well, if necessary:
3465          */
3466         ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3467         if (ra_kb > max_sectors_kb)
3468                 q->backing_dev_info.ra_pages =
3469                                 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3470
3471         q->max_sectors = max_sectors_kb << 1;
3472         spin_unlock_irq(q->queue_lock);
3473
3474         return ret;
3475 }
3476
3477 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3478 {
3479         int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3480
3481         return queue_var_show(max_hw_sectors_kb, (page));
3482 }
3483
3484
3485 static struct queue_sysfs_entry queue_requests_entry = {
3486         .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3487         .show = queue_requests_show,
3488         .store = queue_requests_store,
3489 };
3490
3491 static struct queue_sysfs_entry queue_ra_entry = {
3492         .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3493         .show = queue_ra_show,
3494         .store = queue_ra_store,
3495 };
3496
3497 static struct queue_sysfs_entry queue_max_sectors_entry = {
3498         .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3499         .show = queue_max_sectors_show,
3500         .store = queue_max_sectors_store,
3501 };
3502
3503 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3504         .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
3505         .show = queue_max_hw_sectors_show,
3506 };
3507
3508 static struct queue_sysfs_entry queue_iosched_entry = {
3509         .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
3510         .show = elv_iosched_show,
3511         .store = elv_iosched_store,
3512 };
3513
3514 static struct attribute *default_attrs[] = {
3515         &queue_requests_entry.attr,
3516         &queue_ra_entry.attr,
3517         &queue_max_hw_sectors_entry.attr,
3518         &queue_max_sectors_entry.attr,
3519         &queue_iosched_entry.attr,
3520         NULL,
3521 };
3522
3523 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
3524
3525 static ssize_t
3526 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
3527 {
3528         struct queue_sysfs_entry *entry = to_queue(attr);
3529         struct request_queue *q;
3530
3531         q = container_of(kobj, struct request_queue, kobj);
3532         if (!entry->show)
3533                 return -EIO;
3534
3535         return entry->show(q, page);
3536 }
3537
3538 static ssize_t
3539 queue_attr_store(struct kobject *kobj, struct attribute *attr,
3540                     const char *page, size_t length)
3541 {
3542         struct queue_sysfs_entry *entry = to_queue(attr);
3543         struct request_queue *q;
3544
3545         q = container_of(kobj, struct request_queue, kobj);
3546         if (!entry->store)
3547                 return -EIO;
3548
3549         return entry->store(q, page, length);
3550 }
3551
3552 static struct sysfs_ops queue_sysfs_ops = {
3553         .show   = queue_attr_show,
3554         .store  = queue_attr_store,
3555 };
3556
3557 static struct kobj_type queue_ktype = {
3558         .sysfs_ops      = &queue_sysfs_ops,
3559         .default_attrs  = default_attrs,
3560 };
3561
3562 int blk_register_queue(struct gendisk *disk)
3563 {
3564         int ret;
3565
3566         request_queue_t *q = disk->queue;
3567
3568         if (!q || !q->request_fn)
3569                 return -ENXIO;
3570
3571         q->kobj.parent = kobject_get(&disk->kobj);
3572         if (!q->kobj.parent)
3573                 return -EBUSY;
3574
3575         snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
3576         q->kobj.ktype = &queue_ktype;
3577
3578         ret = kobject_register(&q->kobj);
3579         if (ret < 0)
3580                 return ret;
3581
3582         ret = elv_register_queue(q);
3583         if (ret) {
3584                 kobject_unregister(&q->kobj);
3585                 return ret;
3586         }
3587
3588         return 0;
3589 }
3590
3591 void blk_unregister_queue(struct gendisk *disk)
3592 {
3593         request_queue_t *q = disk->queue;
3594
3595         if (q && q->request_fn) {
3596                 elv_unregister_queue(q);
3597
3598                 kobject_unregister(&q->kobj);
3599                 kobject_put(&disk->kobj);
3600         }
3601 }