Merge master.kernel.org:/home/rmk/linux-2.6-serial
[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
41 /*
42  * For the allocated request tables
43  */
44 static kmem_cache_t *request_cachep;
45
46 /*
47  * For queue allocation
48  */
49 static kmem_cache_t *requestq_cachep;
50
51 /*
52  * For io context allocations
53  */
54 static kmem_cache_t *iocontext_cachep;
55
56 static wait_queue_head_t congestion_wqh[2] = {
57                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
58                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
59         };
60
61 /*
62  * Controlling structure to kblockd
63  */
64 static struct workqueue_struct *kblockd_workqueue; 
65
66 unsigned long blk_max_low_pfn, blk_max_pfn;
67
68 EXPORT_SYMBOL(blk_max_low_pfn);
69 EXPORT_SYMBOL(blk_max_pfn);
70
71 /* Amount of time in which a process may batch requests */
72 #define BLK_BATCH_TIME  (HZ/50UL)
73
74 /* Number of requests a "batching" process may submit */
75 #define BLK_BATCH_REQ   32
76
77 /*
78  * Return the threshold (number of used requests) at which the queue is
79  * considered to be congested.  It include a little hysteresis to keep the
80  * context switch rate down.
81  */
82 static inline int queue_congestion_on_threshold(struct request_queue *q)
83 {
84         return q->nr_congestion_on;
85 }
86
87 /*
88  * The threshold at which a queue is considered to be uncongested
89  */
90 static inline int queue_congestion_off_threshold(struct request_queue *q)
91 {
92         return q->nr_congestion_off;
93 }
94
95 static void blk_queue_congestion_threshold(struct request_queue *q)
96 {
97         int nr;
98
99         nr = q->nr_requests - (q->nr_requests / 8) + 1;
100         if (nr > q->nr_requests)
101                 nr = q->nr_requests;
102         q->nr_congestion_on = nr;
103
104         nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
105         if (nr < 1)
106                 nr = 1;
107         q->nr_congestion_off = nr;
108 }
109
110 /*
111  * A queue has just exitted congestion.  Note this in the global counter of
112  * congested queues, and wake up anyone who was waiting for requests to be
113  * put back.
114  */
115 static void clear_queue_congested(request_queue_t *q, int rw)
116 {
117         enum bdi_state bit;
118         wait_queue_head_t *wqh = &congestion_wqh[rw];
119
120         bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
121         clear_bit(bit, &q->backing_dev_info.state);
122         smp_mb__after_clear_bit();
123         if (waitqueue_active(wqh))
124                 wake_up(wqh);
125 }
126
127 /*
128  * A queue has just entered congestion.  Flag that in the queue's VM-visible
129  * state flags and increment the global gounter of congested queues.
130  */
131 static void set_queue_congested(request_queue_t *q, int rw)
132 {
133         enum bdi_state bit;
134
135         bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
136         set_bit(bit, &q->backing_dev_info.state);
137 }
138
139 /**
140  * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
141  * @bdev:       device
142  *
143  * Locates the passed device's request queue and returns the address of its
144  * backing_dev_info
145  *
146  * Will return NULL if the request queue cannot be located.
147  */
148 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
149 {
150         struct backing_dev_info *ret = NULL;
151         request_queue_t *q = bdev_get_queue(bdev);
152
153         if (q)
154                 ret = &q->backing_dev_info;
155         return ret;
156 }
157
158 EXPORT_SYMBOL(blk_get_backing_dev_info);
159
160 void blk_queue_activity_fn(request_queue_t *q, activity_fn *fn, void *data)
161 {
162         q->activity_fn = fn;
163         q->activity_data = data;
164 }
165
166 EXPORT_SYMBOL(blk_queue_activity_fn);
167
168 /**
169  * blk_queue_prep_rq - set a prepare_request function for queue
170  * @q:          queue
171  * @pfn:        prepare_request function
172  *
173  * It's possible for a queue to register a prepare_request callback which
174  * is invoked before the request is handed to the request_fn. The goal of
175  * the function is to prepare a request for I/O, it can be used to build a
176  * cdb from the request data for instance.
177  *
178  */
179 void blk_queue_prep_rq(request_queue_t *q, prep_rq_fn *pfn)
180 {
181         q->prep_rq_fn = pfn;
182 }
183
184 EXPORT_SYMBOL(blk_queue_prep_rq);
185
186 /**
187  * blk_queue_merge_bvec - set a merge_bvec function for queue
188  * @q:          queue
189  * @mbfn:       merge_bvec_fn
190  *
191  * Usually queues have static limitations on the max sectors or segments that
192  * we can put in a request. Stacking drivers may have some settings that
193  * are dynamic, and thus we have to query the queue whether it is ok to
194  * add a new bio_vec to a bio at a given offset or not. If the block device
195  * has such limitations, it needs to register a merge_bvec_fn to control
196  * the size of bio's sent to it. Note that a block device *must* allow a
197  * single page to be added to an empty bio. The block device driver may want
198  * to use the bio_split() function to deal with these bio's. By default
199  * no merge_bvec_fn is defined for a queue, and only the fixed limits are
200  * honored.
201  */
202 void blk_queue_merge_bvec(request_queue_t *q, merge_bvec_fn *mbfn)
203 {
204         q->merge_bvec_fn = mbfn;
205 }
206
207 EXPORT_SYMBOL(blk_queue_merge_bvec);
208
209 /**
210  * blk_queue_make_request - define an alternate make_request function for a device
211  * @q:  the request queue for the device to be affected
212  * @mfn: the alternate make_request function
213  *
214  * Description:
215  *    The normal way for &struct bios to be passed to a device
216  *    driver is for them to be collected into requests on a request
217  *    queue, and then to allow the device driver to select requests
218  *    off that queue when it is ready.  This works well for many block
219  *    devices. However some block devices (typically virtual devices
220  *    such as md or lvm) do not benefit from the processing on the
221  *    request queue, and are served best by having the requests passed
222  *    directly to them.  This can be achieved by providing a function
223  *    to blk_queue_make_request().
224  *
225  * Caveat:
226  *    The driver that does this *must* be able to deal appropriately
227  *    with buffers in "highmemory". This can be accomplished by either calling
228  *    __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
229  *    blk_queue_bounce() to create a buffer in normal memory.
230  **/
231 void blk_queue_make_request(request_queue_t * q, make_request_fn * mfn)
232 {
233         /*
234          * set defaults
235          */
236         q->nr_requests = BLKDEV_MAX_RQ;
237         q->max_phys_segments = MAX_PHYS_SEGMENTS;
238         q->max_hw_segments = MAX_HW_SEGMENTS;
239         q->make_request_fn = mfn;
240         q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
241         q->backing_dev_info.state = 0;
242         q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
243         blk_queue_max_sectors(q, MAX_SECTORS);
244         blk_queue_hardsect_size(q, 512);
245         blk_queue_dma_alignment(q, 511);
246         blk_queue_congestion_threshold(q);
247         q->nr_batching = BLK_BATCH_REQ;
248
249         q->unplug_thresh = 4;           /* hmm */
250         q->unplug_delay = (3 * HZ) / 1000;      /* 3 milliseconds */
251         if (q->unplug_delay == 0)
252                 q->unplug_delay = 1;
253
254         INIT_WORK(&q->unplug_work, blk_unplug_work, q);
255
256         q->unplug_timer.function = blk_unplug_timeout;
257         q->unplug_timer.data = (unsigned long)q;
258
259         /*
260          * by default assume old behaviour and bounce for any highmem page
261          */
262         blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
263
264         blk_queue_activity_fn(q, NULL, NULL);
265
266         INIT_LIST_HEAD(&q->drain_list);
267 }
268
269 EXPORT_SYMBOL(blk_queue_make_request);
270
271 static inline void rq_init(request_queue_t *q, struct request *rq)
272 {
273         INIT_LIST_HEAD(&rq->queuelist);
274
275         rq->errors = 0;
276         rq->rq_status = RQ_ACTIVE;
277         rq->bio = rq->biotail = NULL;
278         rq->buffer = NULL;
279         rq->ref_count = 1;
280         rq->q = q;
281         rq->waiting = NULL;
282         rq->special = NULL;
283         rq->data_len = 0;
284         rq->data = NULL;
285         rq->sense = NULL;
286         rq->end_io = NULL;
287         rq->end_io_data = NULL;
288 }
289
290 /**
291  * blk_queue_ordered - does this queue support ordered writes
292  * @q:     the request queue
293  * @flag:  see below
294  *
295  * Description:
296  *   For journalled file systems, doing ordered writes on a commit
297  *   block instead of explicitly doing wait_on_buffer (which is bad
298  *   for performance) can be a big win. Block drivers supporting this
299  *   feature should call this function and indicate so.
300  *
301  **/
302 void blk_queue_ordered(request_queue_t *q, int flag)
303 {
304         switch (flag) {
305                 case QUEUE_ORDERED_NONE:
306                         if (q->flush_rq)
307                                 kmem_cache_free(request_cachep, q->flush_rq);
308                         q->flush_rq = NULL;
309                         q->ordered = flag;
310                         break;
311                 case QUEUE_ORDERED_TAG:
312                         q->ordered = flag;
313                         break;
314                 case QUEUE_ORDERED_FLUSH:
315                         q->ordered = flag;
316                         if (!q->flush_rq)
317                                 q->flush_rq = kmem_cache_alloc(request_cachep,
318                                                                 GFP_KERNEL);
319                         break;
320                 default:
321                         printk("blk_queue_ordered: bad value %d\n", flag);
322                         break;
323         }
324 }
325
326 EXPORT_SYMBOL(blk_queue_ordered);
327
328 /**
329  * blk_queue_issue_flush_fn - set function for issuing a flush
330  * @q:     the request queue
331  * @iff:   the function to be called issuing the flush
332  *
333  * Description:
334  *   If a driver supports issuing a flush command, the support is notified
335  *   to the block layer by defining it through this call.
336  *
337  **/
338 void blk_queue_issue_flush_fn(request_queue_t *q, issue_flush_fn *iff)
339 {
340         q->issue_flush_fn = iff;
341 }
342
343 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
344
345 /*
346  * Cache flushing for ordered writes handling
347  */
348 static void blk_pre_flush_end_io(struct request *flush_rq)
349 {
350         struct request *rq = flush_rq->end_io_data;
351         request_queue_t *q = rq->q;
352
353         rq->flags |= REQ_BAR_PREFLUSH;
354
355         if (!flush_rq->errors)
356                 elv_requeue_request(q, rq);
357         else {
358                 q->end_flush_fn(q, flush_rq);
359                 clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
360                 q->request_fn(q);
361         }
362 }
363
364 static void blk_post_flush_end_io(struct request *flush_rq)
365 {
366         struct request *rq = flush_rq->end_io_data;
367         request_queue_t *q = rq->q;
368
369         rq->flags |= REQ_BAR_POSTFLUSH;
370
371         q->end_flush_fn(q, flush_rq);
372         clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
373         q->request_fn(q);
374 }
375
376 struct request *blk_start_pre_flush(request_queue_t *q, struct request *rq)
377 {
378         struct request *flush_rq = q->flush_rq;
379
380         BUG_ON(!blk_barrier_rq(rq));
381
382         if (test_and_set_bit(QUEUE_FLAG_FLUSH, &q->queue_flags))
383                 return NULL;
384
385         rq_init(q, flush_rq);
386         flush_rq->elevator_private = NULL;
387         flush_rq->flags = REQ_BAR_FLUSH;
388         flush_rq->rq_disk = rq->rq_disk;
389         flush_rq->rl = NULL;
390
391         /*
392          * prepare_flush returns 0 if no flush is needed, just mark both
393          * pre and post flush as done in that case
394          */
395         if (!q->prepare_flush_fn(q, flush_rq)) {
396                 rq->flags |= REQ_BAR_PREFLUSH | REQ_BAR_POSTFLUSH;
397                 clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
398                 return rq;
399         }
400
401         /*
402          * some drivers dequeue requests right away, some only after io
403          * completion. make sure the request is dequeued.
404          */
405         if (!list_empty(&rq->queuelist))
406                 blkdev_dequeue_request(rq);
407
408         elv_deactivate_request(q, rq);
409
410         flush_rq->end_io_data = rq;
411         flush_rq->end_io = blk_pre_flush_end_io;
412
413         __elv_add_request(q, flush_rq, ELEVATOR_INSERT_FRONT, 0);
414         return flush_rq;
415 }
416
417 static void blk_start_post_flush(request_queue_t *q, struct request *rq)
418 {
419         struct request *flush_rq = q->flush_rq;
420
421         BUG_ON(!blk_barrier_rq(rq));
422
423         rq_init(q, flush_rq);
424         flush_rq->elevator_private = NULL;
425         flush_rq->flags = REQ_BAR_FLUSH;
426         flush_rq->rq_disk = rq->rq_disk;
427         flush_rq->rl = NULL;
428
429         if (q->prepare_flush_fn(q, flush_rq)) {
430                 flush_rq->end_io_data = rq;
431                 flush_rq->end_io = blk_post_flush_end_io;
432
433                 __elv_add_request(q, flush_rq, ELEVATOR_INSERT_FRONT, 0);
434                 q->request_fn(q);
435         }
436 }
437
438 static inline int blk_check_end_barrier(request_queue_t *q, struct request *rq,
439                                         int sectors)
440 {
441         if (sectors > rq->nr_sectors)
442                 sectors = rq->nr_sectors;
443
444         rq->nr_sectors -= sectors;
445         return rq->nr_sectors;
446 }
447
448 static int __blk_complete_barrier_rq(request_queue_t *q, struct request *rq,
449                                      int sectors, int queue_locked)
450 {
451         if (q->ordered != QUEUE_ORDERED_FLUSH)
452                 return 0;
453         if (!blk_fs_request(rq) || !blk_barrier_rq(rq))
454                 return 0;
455         if (blk_barrier_postflush(rq))
456                 return 0;
457
458         if (!blk_check_end_barrier(q, rq, sectors)) {
459                 unsigned long flags = 0;
460
461                 if (!queue_locked)
462                         spin_lock_irqsave(q->queue_lock, flags);
463
464                 blk_start_post_flush(q, rq);
465
466                 if (!queue_locked)
467                         spin_unlock_irqrestore(q->queue_lock, flags);
468         }
469
470         return 1;
471 }
472
473 /**
474  * blk_complete_barrier_rq - complete possible barrier request
475  * @q:  the request queue for the device
476  * @rq:  the request
477  * @sectors:  number of sectors to complete
478  *
479  * Description:
480  *   Used in driver end_io handling to determine whether to postpone
481  *   completion of a barrier request until a post flush has been done. This
482  *   is the unlocked variant, used if the caller doesn't already hold the
483  *   queue lock.
484  **/
485 int blk_complete_barrier_rq(request_queue_t *q, struct request *rq, int sectors)
486 {
487         return __blk_complete_barrier_rq(q, rq, sectors, 0);
488 }
489 EXPORT_SYMBOL(blk_complete_barrier_rq);
490
491 /**
492  * blk_complete_barrier_rq_locked - complete possible barrier request
493  * @q:  the request queue for the device
494  * @rq:  the request
495  * @sectors:  number of sectors to complete
496  *
497  * Description:
498  *   See blk_complete_barrier_rq(). This variant must be used if the caller
499  *   holds the queue lock.
500  **/
501 int blk_complete_barrier_rq_locked(request_queue_t *q, struct request *rq,
502                                    int sectors)
503 {
504         return __blk_complete_barrier_rq(q, rq, sectors, 1);
505 }
506 EXPORT_SYMBOL(blk_complete_barrier_rq_locked);
507
508 /**
509  * blk_queue_bounce_limit - set bounce buffer limit for queue
510  * @q:  the request queue for the device
511  * @dma_addr:   bus address limit
512  *
513  * Description:
514  *    Different hardware can have different requirements as to what pages
515  *    it can do I/O directly to. A low level driver can call
516  *    blk_queue_bounce_limit to have lower memory pages allocated as bounce
517  *    buffers for doing I/O to pages residing above @page. By default
518  *    the block layer sets this to the highest numbered "low" memory page.
519  **/
520 void blk_queue_bounce_limit(request_queue_t *q, u64 dma_addr)
521 {
522         unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
523
524         /*
525          * set appropriate bounce gfp mask -- unfortunately we don't have a
526          * full 4GB zone, so we have to resort to low memory for any bounces.
527          * ISA has its own < 16MB zone.
528          */
529         if (bounce_pfn < blk_max_low_pfn) {
530                 BUG_ON(dma_addr < BLK_BOUNCE_ISA);
531                 init_emergency_isa_pool();
532                 q->bounce_gfp = GFP_NOIO | GFP_DMA;
533         } else
534                 q->bounce_gfp = GFP_NOIO;
535
536         q->bounce_pfn = bounce_pfn;
537 }
538
539 EXPORT_SYMBOL(blk_queue_bounce_limit);
540
541 /**
542  * blk_queue_max_sectors - set max sectors for a request for this queue
543  * @q:  the request queue for the device
544  * @max_sectors:  max sectors in the usual 512b unit
545  *
546  * Description:
547  *    Enables a low level driver to set an upper limit on the size of
548  *    received requests.
549  **/
550 void blk_queue_max_sectors(request_queue_t *q, unsigned short max_sectors)
551 {
552         if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
553                 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
554                 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
555         }
556
557         q->max_sectors = q->max_hw_sectors = max_sectors;
558 }
559
560 EXPORT_SYMBOL(blk_queue_max_sectors);
561
562 /**
563  * blk_queue_max_phys_segments - set max phys segments for a request for this queue
564  * @q:  the request queue for the device
565  * @max_segments:  max number of segments
566  *
567  * Description:
568  *    Enables a low level driver to set an upper limit on the number of
569  *    physical data segments in a request.  This would be the largest sized
570  *    scatter list the driver could handle.
571  **/
572 void blk_queue_max_phys_segments(request_queue_t *q, unsigned short max_segments)
573 {
574         if (!max_segments) {
575                 max_segments = 1;
576                 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
577         }
578
579         q->max_phys_segments = max_segments;
580 }
581
582 EXPORT_SYMBOL(blk_queue_max_phys_segments);
583
584 /**
585  * blk_queue_max_hw_segments - set max hw segments for a request for this queue
586  * @q:  the request queue for the device
587  * @max_segments:  max number of segments
588  *
589  * Description:
590  *    Enables a low level driver to set an upper limit on the number of
591  *    hw data segments in a request.  This would be the largest number of
592  *    address/length pairs the host adapter can actually give as once
593  *    to the device.
594  **/
595 void blk_queue_max_hw_segments(request_queue_t *q, unsigned short max_segments)
596 {
597         if (!max_segments) {
598                 max_segments = 1;
599                 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
600         }
601
602         q->max_hw_segments = max_segments;
603 }
604
605 EXPORT_SYMBOL(blk_queue_max_hw_segments);
606
607 /**
608  * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
609  * @q:  the request queue for the device
610  * @max_size:  max size of segment in bytes
611  *
612  * Description:
613  *    Enables a low level driver to set an upper limit on the size of a
614  *    coalesced segment
615  **/
616 void blk_queue_max_segment_size(request_queue_t *q, unsigned int max_size)
617 {
618         if (max_size < PAGE_CACHE_SIZE) {
619                 max_size = PAGE_CACHE_SIZE;
620                 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
621         }
622
623         q->max_segment_size = max_size;
624 }
625
626 EXPORT_SYMBOL(blk_queue_max_segment_size);
627
628 /**
629  * blk_queue_hardsect_size - set hardware sector size for the queue
630  * @q:  the request queue for the device
631  * @size:  the hardware sector size, in bytes
632  *
633  * Description:
634  *   This should typically be set to the lowest possible sector size
635  *   that the hardware can operate on (possible without reverting to
636  *   even internal read-modify-write operations). Usually the default
637  *   of 512 covers most hardware.
638  **/
639 void blk_queue_hardsect_size(request_queue_t *q, unsigned short size)
640 {
641         q->hardsect_size = size;
642 }
643
644 EXPORT_SYMBOL(blk_queue_hardsect_size);
645
646 /*
647  * Returns the minimum that is _not_ zero, unless both are zero.
648  */
649 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
650
651 /**
652  * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
653  * @t:  the stacking driver (top)
654  * @b:  the underlying device (bottom)
655  **/
656 void blk_queue_stack_limits(request_queue_t *t, request_queue_t *b)
657 {
658         /* zero is "infinity" */
659         t->max_sectors = t->max_hw_sectors =
660                 min_not_zero(t->max_sectors,b->max_sectors);
661
662         t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
663         t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
664         t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
665         t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
666 }
667
668 EXPORT_SYMBOL(blk_queue_stack_limits);
669
670 /**
671  * blk_queue_segment_boundary - set boundary rules for segment merging
672  * @q:  the request queue for the device
673  * @mask:  the memory boundary mask
674  **/
675 void blk_queue_segment_boundary(request_queue_t *q, unsigned long mask)
676 {
677         if (mask < PAGE_CACHE_SIZE - 1) {
678                 mask = PAGE_CACHE_SIZE - 1;
679                 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
680         }
681
682         q->seg_boundary_mask = mask;
683 }
684
685 EXPORT_SYMBOL(blk_queue_segment_boundary);
686
687 /**
688  * blk_queue_dma_alignment - set dma length and memory alignment
689  * @q:     the request queue for the device
690  * @mask:  alignment mask
691  *
692  * description:
693  *    set required memory and length aligment for direct dma transactions.
694  *    this is used when buiding direct io requests for the queue.
695  *
696  **/
697 void blk_queue_dma_alignment(request_queue_t *q, int mask)
698 {
699         q->dma_alignment = mask;
700 }
701
702 EXPORT_SYMBOL(blk_queue_dma_alignment);
703
704 /**
705  * blk_queue_find_tag - find a request by its tag and queue
706  *
707  * @q:   The request queue for the device
708  * @tag: The tag of the request
709  *
710  * Notes:
711  *    Should be used when a device returns a tag and you want to match
712  *    it with a request.
713  *
714  *    no locks need be held.
715  **/
716 struct request *blk_queue_find_tag(request_queue_t *q, int tag)
717 {
718         struct blk_queue_tag *bqt = q->queue_tags;
719
720         if (unlikely(bqt == NULL || tag >= bqt->max_depth))
721                 return NULL;
722
723         return bqt->tag_index[tag];
724 }
725
726 EXPORT_SYMBOL(blk_queue_find_tag);
727
728 /**
729  * __blk_queue_free_tags - release tag maintenance info
730  * @q:  the request queue for the device
731  *
732  *  Notes:
733  *    blk_cleanup_queue() will take care of calling this function, if tagging
734  *    has been used. So there's no need to call this directly.
735  **/
736 static void __blk_queue_free_tags(request_queue_t *q)
737 {
738         struct blk_queue_tag *bqt = q->queue_tags;
739
740         if (!bqt)
741                 return;
742
743         if (atomic_dec_and_test(&bqt->refcnt)) {
744                 BUG_ON(bqt->busy);
745                 BUG_ON(!list_empty(&bqt->busy_list));
746
747                 kfree(bqt->tag_index);
748                 bqt->tag_index = NULL;
749
750                 kfree(bqt->tag_map);
751                 bqt->tag_map = NULL;
752
753                 kfree(bqt);
754         }
755
756         q->queue_tags = NULL;
757         q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
758 }
759
760 /**
761  * blk_queue_free_tags - release tag maintenance info
762  * @q:  the request queue for the device
763  *
764  *  Notes:
765  *      This is used to disabled tagged queuing to a device, yet leave
766  *      queue in function.
767  **/
768 void blk_queue_free_tags(request_queue_t *q)
769 {
770         clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
771 }
772
773 EXPORT_SYMBOL(blk_queue_free_tags);
774
775 static int
776 init_tag_map(request_queue_t *q, struct blk_queue_tag *tags, int depth)
777 {
778         struct request **tag_index;
779         unsigned long *tag_map;
780         int nr_ulongs;
781
782         if (depth > q->nr_requests * 2) {
783                 depth = q->nr_requests * 2;
784                 printk(KERN_ERR "%s: adjusted depth to %d\n",
785                                 __FUNCTION__, depth);
786         }
787
788         tag_index = kmalloc(depth * sizeof(struct request *), GFP_ATOMIC);
789         if (!tag_index)
790                 goto fail;
791
792         nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
793         tag_map = kmalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
794         if (!tag_map)
795                 goto fail;
796
797         memset(tag_index, 0, depth * sizeof(struct request *));
798         memset(tag_map, 0, nr_ulongs * sizeof(unsigned long));
799         tags->max_depth = depth;
800         tags->tag_index = tag_index;
801         tags->tag_map = tag_map;
802
803         return 0;
804 fail:
805         kfree(tag_index);
806         return -ENOMEM;
807 }
808
809 /**
810  * blk_queue_init_tags - initialize the queue tag info
811  * @q:  the request queue for the device
812  * @depth:  the maximum queue depth supported
813  * @tags: the tag to use
814  **/
815 int blk_queue_init_tags(request_queue_t *q, int depth,
816                         struct blk_queue_tag *tags)
817 {
818         int rc;
819
820         BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
821
822         if (!tags && !q->queue_tags) {
823                 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
824                 if (!tags)
825                         goto fail;
826
827                 if (init_tag_map(q, tags, depth))
828                         goto fail;
829
830                 INIT_LIST_HEAD(&tags->busy_list);
831                 tags->busy = 0;
832                 atomic_set(&tags->refcnt, 1);
833         } else if (q->queue_tags) {
834                 if ((rc = blk_queue_resize_tags(q, depth)))
835                         return rc;
836                 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
837                 return 0;
838         } else
839                 atomic_inc(&tags->refcnt);
840
841         /*
842          * assign it, all done
843          */
844         q->queue_tags = tags;
845         q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
846         return 0;
847 fail:
848         kfree(tags);
849         return -ENOMEM;
850 }
851
852 EXPORT_SYMBOL(blk_queue_init_tags);
853
854 /**
855  * blk_queue_resize_tags - change the queueing depth
856  * @q:  the request queue for the device
857  * @new_depth: the new max command queueing depth
858  *
859  *  Notes:
860  *    Must be called with the queue lock held.
861  **/
862 int blk_queue_resize_tags(request_queue_t *q, int new_depth)
863 {
864         struct blk_queue_tag *bqt = q->queue_tags;
865         struct request **tag_index;
866         unsigned long *tag_map;
867         int max_depth, nr_ulongs;
868
869         if (!bqt)
870                 return -ENXIO;
871
872         /*
873          * save the old state info, so we can copy it back
874          */
875         tag_index = bqt->tag_index;
876         tag_map = bqt->tag_map;
877         max_depth = bqt->max_depth;
878
879         if (init_tag_map(q, bqt, new_depth))
880                 return -ENOMEM;
881
882         memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
883         nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
884         memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
885
886         kfree(tag_index);
887         kfree(tag_map);
888         return 0;
889 }
890
891 EXPORT_SYMBOL(blk_queue_resize_tags);
892
893 /**
894  * blk_queue_end_tag - end tag operations for a request
895  * @q:  the request queue for the device
896  * @rq: the request that has completed
897  *
898  *  Description:
899  *    Typically called when end_that_request_first() returns 0, meaning
900  *    all transfers have been done for a request. It's important to call
901  *    this function before end_that_request_last(), as that will put the
902  *    request back on the free list thus corrupting the internal tag list.
903  *
904  *  Notes:
905  *   queue lock must be held.
906  **/
907 void blk_queue_end_tag(request_queue_t *q, struct request *rq)
908 {
909         struct blk_queue_tag *bqt = q->queue_tags;
910         int tag = rq->tag;
911
912         BUG_ON(tag == -1);
913
914         if (unlikely(tag >= bqt->max_depth))
915                 /*
916                  * This can happen after tag depth has been reduced.
917                  * FIXME: how about a warning or info message here?
918                  */
919                 return;
920
921         if (unlikely(!__test_and_clear_bit(tag, bqt->tag_map))) {
922                 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
923                        __FUNCTION__, tag);
924                 return;
925         }
926
927         list_del_init(&rq->queuelist);
928         rq->flags &= ~REQ_QUEUED;
929         rq->tag = -1;
930
931         if (unlikely(bqt->tag_index[tag] == NULL))
932                 printk(KERN_ERR "%s: tag %d is missing\n",
933                        __FUNCTION__, tag);
934
935         bqt->tag_index[tag] = NULL;
936         bqt->busy--;
937 }
938
939 EXPORT_SYMBOL(blk_queue_end_tag);
940
941 /**
942  * blk_queue_start_tag - find a free tag and assign it
943  * @q:  the request queue for the device
944  * @rq:  the block request that needs tagging
945  *
946  *  Description:
947  *    This can either be used as a stand-alone helper, or possibly be
948  *    assigned as the queue &prep_rq_fn (in which case &struct request
949  *    automagically gets a tag assigned). Note that this function
950  *    assumes that any type of request can be queued! if this is not
951  *    true for your device, you must check the request type before
952  *    calling this function.  The request will also be removed from
953  *    the request queue, so it's the drivers responsibility to readd
954  *    it if it should need to be restarted for some reason.
955  *
956  *  Notes:
957  *   queue lock must be held.
958  **/
959 int blk_queue_start_tag(request_queue_t *q, struct request *rq)
960 {
961         struct blk_queue_tag *bqt = q->queue_tags;
962         int tag;
963
964         if (unlikely((rq->flags & REQ_QUEUED))) {
965                 printk(KERN_ERR 
966                        "%s: request %p for device [%s] already tagged %d",
967                        __FUNCTION__, rq,
968                        rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
969                 BUG();
970         }
971
972         tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
973         if (tag >= bqt->max_depth)
974                 return 1;
975
976         __set_bit(tag, bqt->tag_map);
977
978         rq->flags |= REQ_QUEUED;
979         rq->tag = tag;
980         bqt->tag_index[tag] = rq;
981         blkdev_dequeue_request(rq);
982         list_add(&rq->queuelist, &bqt->busy_list);
983         bqt->busy++;
984         return 0;
985 }
986
987 EXPORT_SYMBOL(blk_queue_start_tag);
988
989 /**
990  * blk_queue_invalidate_tags - invalidate all pending tags
991  * @q:  the request queue for the device
992  *
993  *  Description:
994  *   Hardware conditions may dictate a need to stop all pending requests.
995  *   In this case, we will safely clear the block side of the tag queue and
996  *   readd all requests to the request queue in the right order.
997  *
998  *  Notes:
999  *   queue lock must be held.
1000  **/
1001 void blk_queue_invalidate_tags(request_queue_t *q)
1002 {
1003         struct blk_queue_tag *bqt = q->queue_tags;
1004         struct list_head *tmp, *n;
1005         struct request *rq;
1006
1007         list_for_each_safe(tmp, n, &bqt->busy_list) {
1008                 rq = list_entry_rq(tmp);
1009
1010                 if (rq->tag == -1) {
1011                         printk(KERN_ERR
1012                                "%s: bad tag found on list\n", __FUNCTION__);
1013                         list_del_init(&rq->queuelist);
1014                         rq->flags &= ~REQ_QUEUED;
1015                 } else
1016                         blk_queue_end_tag(q, rq);
1017
1018                 rq->flags &= ~REQ_STARTED;
1019                 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1020         }
1021 }
1022
1023 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1024
1025 static char *rq_flags[] = {
1026         "REQ_RW",
1027         "REQ_FAILFAST",
1028         "REQ_SOFTBARRIER",
1029         "REQ_HARDBARRIER",
1030         "REQ_CMD",
1031         "REQ_NOMERGE",
1032         "REQ_STARTED",
1033         "REQ_DONTPREP",
1034         "REQ_QUEUED",
1035         "REQ_PC",
1036         "REQ_BLOCK_PC",
1037         "REQ_SENSE",
1038         "REQ_FAILED",
1039         "REQ_QUIET",
1040         "REQ_SPECIAL",
1041         "REQ_DRIVE_CMD",
1042         "REQ_DRIVE_TASK",
1043         "REQ_DRIVE_TASKFILE",
1044         "REQ_PREEMPT",
1045         "REQ_PM_SUSPEND",
1046         "REQ_PM_RESUME",
1047         "REQ_PM_SHUTDOWN",
1048 };
1049
1050 void blk_dump_rq_flags(struct request *rq, char *msg)
1051 {
1052         int bit;
1053
1054         printk("%s: dev %s: flags = ", msg,
1055                 rq->rq_disk ? rq->rq_disk->disk_name : "?");
1056         bit = 0;
1057         do {
1058                 if (rq->flags & (1 << bit))
1059                         printk("%s ", rq_flags[bit]);
1060                 bit++;
1061         } while (bit < __REQ_NR_BITS);
1062
1063         printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1064                                                        rq->nr_sectors,
1065                                                        rq->current_nr_sectors);
1066         printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1067
1068         if (rq->flags & (REQ_BLOCK_PC | REQ_PC)) {
1069                 printk("cdb: ");
1070                 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1071                         printk("%02x ", rq->cmd[bit]);
1072                 printk("\n");
1073         }
1074 }
1075
1076 EXPORT_SYMBOL(blk_dump_rq_flags);
1077
1078 void blk_recount_segments(request_queue_t *q, struct bio *bio)
1079 {
1080         struct bio_vec *bv, *bvprv = NULL;
1081         int i, nr_phys_segs, nr_hw_segs, seg_size, hw_seg_size, cluster;
1082         int high, highprv = 1;
1083
1084         if (unlikely(!bio->bi_io_vec))
1085                 return;
1086
1087         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1088         hw_seg_size = seg_size = nr_phys_segs = nr_hw_segs = 0;
1089         bio_for_each_segment(bv, bio, i) {
1090                 /*
1091                  * the trick here is making sure that a high page is never
1092                  * considered part of another segment, since that might
1093                  * change with the bounce page.
1094                  */
1095                 high = page_to_pfn(bv->bv_page) >= q->bounce_pfn;
1096                 if (high || highprv)
1097                         goto new_hw_segment;
1098                 if (cluster) {
1099                         if (seg_size + bv->bv_len > q->max_segment_size)
1100                                 goto new_segment;
1101                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1102                                 goto new_segment;
1103                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1104                                 goto new_segment;
1105                         if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1106                                 goto new_hw_segment;
1107
1108                         seg_size += bv->bv_len;
1109                         hw_seg_size += bv->bv_len;
1110                         bvprv = bv;
1111                         continue;
1112                 }
1113 new_segment:
1114                 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1115                     !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len)) {
1116                         hw_seg_size += bv->bv_len;
1117                 } else {
1118 new_hw_segment:
1119                         if (hw_seg_size > bio->bi_hw_front_size)
1120                                 bio->bi_hw_front_size = hw_seg_size;
1121                         hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1122                         nr_hw_segs++;
1123                 }
1124
1125                 nr_phys_segs++;
1126                 bvprv = bv;
1127                 seg_size = bv->bv_len;
1128                 highprv = high;
1129         }
1130         if (hw_seg_size > bio->bi_hw_back_size)
1131                 bio->bi_hw_back_size = hw_seg_size;
1132         if (nr_hw_segs == 1 && hw_seg_size > bio->bi_hw_front_size)
1133                 bio->bi_hw_front_size = hw_seg_size;
1134         bio->bi_phys_segments = nr_phys_segs;
1135         bio->bi_hw_segments = nr_hw_segs;
1136         bio->bi_flags |= (1 << BIO_SEG_VALID);
1137 }
1138
1139
1140 int blk_phys_contig_segment(request_queue_t *q, struct bio *bio,
1141                                    struct bio *nxt)
1142 {
1143         if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1144                 return 0;
1145
1146         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1147                 return 0;
1148         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1149                 return 0;
1150
1151         /*
1152          * bio and nxt are contigous in memory, check if the queue allows
1153          * these two to be merged into one
1154          */
1155         if (BIO_SEG_BOUNDARY(q, bio, nxt))
1156                 return 1;
1157
1158         return 0;
1159 }
1160
1161 EXPORT_SYMBOL(blk_phys_contig_segment);
1162
1163 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 EXPORT_SYMBOL(blk_hw_contig_segment);
1180
1181 /*
1182  * map a request to scatterlist, return number of sg entries setup. Caller
1183  * must make sure sg can hold rq->nr_phys_segments entries
1184  */
1185 int blk_rq_map_sg(request_queue_t *q, struct request *rq, struct scatterlist *sg)
1186 {
1187         struct bio_vec *bvec, *bvprv;
1188         struct bio *bio;
1189         int nsegs, i, cluster;
1190
1191         nsegs = 0;
1192         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1193
1194         /*
1195          * for each bio in rq
1196          */
1197         bvprv = NULL;
1198         rq_for_each_bio(bio, rq) {
1199                 /*
1200                  * for each segment in bio
1201                  */
1202                 bio_for_each_segment(bvec, bio, i) {
1203                         int nbytes = bvec->bv_len;
1204
1205                         if (bvprv && cluster) {
1206                                 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1207                                         goto new_segment;
1208
1209                                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1210                                         goto new_segment;
1211                                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1212                                         goto new_segment;
1213
1214                                 sg[nsegs - 1].length += nbytes;
1215                         } else {
1216 new_segment:
1217                                 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1218                                 sg[nsegs].page = bvec->bv_page;
1219                                 sg[nsegs].length = nbytes;
1220                                 sg[nsegs].offset = bvec->bv_offset;
1221
1222                                 nsegs++;
1223                         }
1224                         bvprv = bvec;
1225                 } /* segments in bio */
1226         } /* bios in rq */
1227
1228         return nsegs;
1229 }
1230
1231 EXPORT_SYMBOL(blk_rq_map_sg);
1232
1233 /*
1234  * the standard queue merge functions, can be overridden with device
1235  * specific ones if so desired
1236  */
1237
1238 static inline int ll_new_mergeable(request_queue_t *q,
1239                                    struct request *req,
1240                                    struct bio *bio)
1241 {
1242         int nr_phys_segs = bio_phys_segments(q, bio);
1243
1244         if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1245                 req->flags |= REQ_NOMERGE;
1246                 if (req == q->last_merge)
1247                         q->last_merge = NULL;
1248                 return 0;
1249         }
1250
1251         /*
1252          * A hw segment is just getting larger, bump just the phys
1253          * counter.
1254          */
1255         req->nr_phys_segments += nr_phys_segs;
1256         return 1;
1257 }
1258
1259 static inline int ll_new_hw_segment(request_queue_t *q,
1260                                     struct request *req,
1261                                     struct bio *bio)
1262 {
1263         int nr_hw_segs = bio_hw_segments(q, bio);
1264         int nr_phys_segs = bio_phys_segments(q, bio);
1265
1266         if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1267             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1268                 req->flags |= REQ_NOMERGE;
1269                 if (req == q->last_merge)
1270                         q->last_merge = NULL;
1271                 return 0;
1272         }
1273
1274         /*
1275          * This will form the start of a new hw segment.  Bump both
1276          * counters.
1277          */
1278         req->nr_hw_segments += nr_hw_segs;
1279         req->nr_phys_segments += nr_phys_segs;
1280         return 1;
1281 }
1282
1283 static int ll_back_merge_fn(request_queue_t *q, struct request *req, 
1284                             struct bio *bio)
1285 {
1286         int len;
1287
1288         if (req->nr_sectors + bio_sectors(bio) > q->max_sectors) {
1289                 req->flags |= REQ_NOMERGE;
1290                 if (req == q->last_merge)
1291                         q->last_merge = NULL;
1292                 return 0;
1293         }
1294         if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1295                 blk_recount_segments(q, req->biotail);
1296         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1297                 blk_recount_segments(q, bio);
1298         len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1299         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1300             !BIOVEC_VIRT_OVERSIZE(len)) {
1301                 int mergeable =  ll_new_mergeable(q, req, bio);
1302
1303                 if (mergeable) {
1304                         if (req->nr_hw_segments == 1)
1305                                 req->bio->bi_hw_front_size = len;
1306                         if (bio->bi_hw_segments == 1)
1307                                 bio->bi_hw_back_size = len;
1308                 }
1309                 return mergeable;
1310         }
1311
1312         return ll_new_hw_segment(q, req, bio);
1313 }
1314
1315 static int ll_front_merge_fn(request_queue_t *q, struct request *req, 
1316                              struct bio *bio)
1317 {
1318         int len;
1319
1320         if (req->nr_sectors + bio_sectors(bio) > q->max_sectors) {
1321                 req->flags |= REQ_NOMERGE;
1322                 if (req == q->last_merge)
1323                         q->last_merge = NULL;
1324                 return 0;
1325         }
1326         len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1327         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1328                 blk_recount_segments(q, bio);
1329         if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1330                 blk_recount_segments(q, req->bio);
1331         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1332             !BIOVEC_VIRT_OVERSIZE(len)) {
1333                 int mergeable =  ll_new_mergeable(q, req, bio);
1334
1335                 if (mergeable) {
1336                         if (bio->bi_hw_segments == 1)
1337                                 bio->bi_hw_front_size = len;
1338                         if (req->nr_hw_segments == 1)
1339                                 req->biotail->bi_hw_back_size = len;
1340                 }
1341                 return mergeable;
1342         }
1343
1344         return ll_new_hw_segment(q, req, bio);
1345 }
1346
1347 static int ll_merge_requests_fn(request_queue_t *q, struct request *req,
1348                                 struct request *next)
1349 {
1350         int total_phys_segments = req->nr_phys_segments +next->nr_phys_segments;
1351         int total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1352
1353         /*
1354          * First check if the either of the requests are re-queued
1355          * requests.  Can't merge them if they are.
1356          */
1357         if (req->special || next->special)
1358                 return 0;
1359
1360         /*
1361          * Will it become to large?
1362          */
1363         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1364                 return 0;
1365
1366         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1367         if (blk_phys_contig_segment(q, req->biotail, next->bio))
1368                 total_phys_segments--;
1369
1370         if (total_phys_segments > q->max_phys_segments)
1371                 return 0;
1372
1373         total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1374         if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1375                 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1376                 /*
1377                  * propagate the combined length to the end of the requests
1378                  */
1379                 if (req->nr_hw_segments == 1)
1380                         req->bio->bi_hw_front_size = len;
1381                 if (next->nr_hw_segments == 1)
1382                         next->biotail->bi_hw_back_size = len;
1383                 total_hw_segments--;
1384         }
1385
1386         if (total_hw_segments > q->max_hw_segments)
1387                 return 0;
1388
1389         /* Merge is OK... */
1390         req->nr_phys_segments = total_phys_segments;
1391         req->nr_hw_segments = total_hw_segments;
1392         return 1;
1393 }
1394
1395 /*
1396  * "plug" the device if there are no outstanding requests: this will
1397  * force the transfer to start only after we have put all the requests
1398  * on the list.
1399  *
1400  * This is called with interrupts off and no requests on the queue and
1401  * with the queue lock held.
1402  */
1403 void blk_plug_device(request_queue_t *q)
1404 {
1405         WARN_ON(!irqs_disabled());
1406
1407         /*
1408          * don't plug a stopped queue, it must be paired with blk_start_queue()
1409          * which will restart the queueing
1410          */
1411         if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
1412                 return;
1413
1414         if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1415                 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1416 }
1417
1418 EXPORT_SYMBOL(blk_plug_device);
1419
1420 /*
1421  * remove the queue from the plugged list, if present. called with
1422  * queue lock held and interrupts disabled.
1423  */
1424 int blk_remove_plug(request_queue_t *q)
1425 {
1426         WARN_ON(!irqs_disabled());
1427
1428         if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1429                 return 0;
1430
1431         del_timer(&q->unplug_timer);
1432         return 1;
1433 }
1434
1435 EXPORT_SYMBOL(blk_remove_plug);
1436
1437 /*
1438  * remove the plug and let it rip..
1439  */
1440 void __generic_unplug_device(request_queue_t *q)
1441 {
1442         if (unlikely(test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags)))
1443                 return;
1444
1445         if (!blk_remove_plug(q))
1446                 return;
1447
1448         /*
1449          * was plugged, fire request_fn if queue has stuff to do
1450          */
1451         if (elv_next_request(q))
1452                 q->request_fn(q);
1453 }
1454 EXPORT_SYMBOL(__generic_unplug_device);
1455
1456 /**
1457  * generic_unplug_device - fire a request queue
1458  * @q:    The &request_queue_t in question
1459  *
1460  * Description:
1461  *   Linux uses plugging to build bigger requests queues before letting
1462  *   the device have at them. If a queue is plugged, the I/O scheduler
1463  *   is still adding and merging requests on the queue. Once the queue
1464  *   gets unplugged, the request_fn defined for the queue is invoked and
1465  *   transfers started.
1466  **/
1467 void generic_unplug_device(request_queue_t *q)
1468 {
1469         spin_lock_irq(q->queue_lock);
1470         __generic_unplug_device(q);
1471         spin_unlock_irq(q->queue_lock);
1472 }
1473 EXPORT_SYMBOL(generic_unplug_device);
1474
1475 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1476                                    struct page *page)
1477 {
1478         request_queue_t *q = bdi->unplug_io_data;
1479
1480         /*
1481          * devices don't necessarily have an ->unplug_fn defined
1482          */
1483         if (q->unplug_fn)
1484                 q->unplug_fn(q);
1485 }
1486
1487 static void blk_unplug_work(void *data)
1488 {
1489         request_queue_t *q = data;
1490
1491         q->unplug_fn(q);
1492 }
1493
1494 static void blk_unplug_timeout(unsigned long data)
1495 {
1496         request_queue_t *q = (request_queue_t *)data;
1497
1498         kblockd_schedule_work(&q->unplug_work);
1499 }
1500
1501 /**
1502  * blk_start_queue - restart a previously stopped queue
1503  * @q:    The &request_queue_t in question
1504  *
1505  * Description:
1506  *   blk_start_queue() will clear the stop flag on the queue, and call
1507  *   the request_fn for the queue if it was in a stopped state when
1508  *   entered. Also see blk_stop_queue(). Queue lock must be held.
1509  **/
1510 void blk_start_queue(request_queue_t *q)
1511 {
1512         clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1513
1514         /*
1515          * one level of recursion is ok and is much faster than kicking
1516          * the unplug handling
1517          */
1518         if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1519                 q->request_fn(q);
1520                 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1521         } else {
1522                 blk_plug_device(q);
1523                 kblockd_schedule_work(&q->unplug_work);
1524         }
1525 }
1526
1527 EXPORT_SYMBOL(blk_start_queue);
1528
1529 /**
1530  * blk_stop_queue - stop a queue
1531  * @q:    The &request_queue_t in question
1532  *
1533  * Description:
1534  *   The Linux block layer assumes that a block driver will consume all
1535  *   entries on the request queue when the request_fn strategy is called.
1536  *   Often this will not happen, because of hardware limitations (queue
1537  *   depth settings). If a device driver gets a 'queue full' response,
1538  *   or if it simply chooses not to queue more I/O at one point, it can
1539  *   call this function to prevent the request_fn from being called until
1540  *   the driver has signalled it's ready to go again. This happens by calling
1541  *   blk_start_queue() to restart queue operations. Queue lock must be held.
1542  **/
1543 void blk_stop_queue(request_queue_t *q)
1544 {
1545         blk_remove_plug(q);
1546         set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1547 }
1548 EXPORT_SYMBOL(blk_stop_queue);
1549
1550 /**
1551  * blk_sync_queue - cancel any pending callbacks on a queue
1552  * @q: the queue
1553  *
1554  * Description:
1555  *     The block layer may perform asynchronous callback activity
1556  *     on a queue, such as calling the unplug function after a timeout.
1557  *     A block device may call blk_sync_queue to ensure that any
1558  *     such activity is cancelled, thus allowing it to release resources
1559  *     the the callbacks might use. The caller must already have made sure
1560  *     that its ->make_request_fn will not re-add plugging prior to calling
1561  *     this function.
1562  *
1563  */
1564 void blk_sync_queue(struct request_queue *q)
1565 {
1566         del_timer_sync(&q->unplug_timer);
1567         kblockd_flush();
1568 }
1569 EXPORT_SYMBOL(blk_sync_queue);
1570
1571 /**
1572  * blk_run_queue - run a single device queue
1573  * @q:  The queue to run
1574  */
1575 void blk_run_queue(struct request_queue *q)
1576 {
1577         unsigned long flags;
1578
1579         spin_lock_irqsave(q->queue_lock, flags);
1580         blk_remove_plug(q);
1581         if (!elv_queue_empty(q))
1582                 q->request_fn(q);
1583         spin_unlock_irqrestore(q->queue_lock, flags);
1584 }
1585 EXPORT_SYMBOL(blk_run_queue);
1586
1587 /**
1588  * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1589  * @q:    the request queue to be released
1590  *
1591  * Description:
1592  *     blk_cleanup_queue is the pair to blk_init_queue() or
1593  *     blk_queue_make_request().  It should be called when a request queue is
1594  *     being released; typically when a block device is being de-registered.
1595  *     Currently, its primary task it to free all the &struct request
1596  *     structures that were allocated to the queue and the queue itself.
1597  *
1598  * Caveat:
1599  *     Hopefully the low level driver will have finished any
1600  *     outstanding requests first...
1601  **/
1602 void blk_cleanup_queue(request_queue_t * q)
1603 {
1604         struct request_list *rl = &q->rq;
1605
1606         if (!atomic_dec_and_test(&q->refcnt))
1607                 return;
1608
1609         if (q->elevator)
1610                 elevator_exit(q->elevator);
1611
1612         blk_sync_queue(q);
1613
1614         if (rl->rq_pool)
1615                 mempool_destroy(rl->rq_pool);
1616
1617         if (q->queue_tags)
1618                 __blk_queue_free_tags(q);
1619
1620         blk_queue_ordered(q, QUEUE_ORDERED_NONE);
1621
1622         kmem_cache_free(requestq_cachep, q);
1623 }
1624
1625 EXPORT_SYMBOL(blk_cleanup_queue);
1626
1627 static int blk_init_free_list(request_queue_t *q)
1628 {
1629         struct request_list *rl = &q->rq;
1630
1631         rl->count[READ] = rl->count[WRITE] = 0;
1632         rl->starved[READ] = rl->starved[WRITE] = 0;
1633         init_waitqueue_head(&rl->wait[READ]);
1634         init_waitqueue_head(&rl->wait[WRITE]);
1635         init_waitqueue_head(&rl->drain);
1636
1637         rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1638                                 mempool_free_slab, request_cachep, q->node);
1639
1640         if (!rl->rq_pool)
1641                 return -ENOMEM;
1642
1643         return 0;
1644 }
1645
1646 static int __make_request(request_queue_t *, struct bio *);
1647
1648 request_queue_t *blk_alloc_queue(int gfp_mask)
1649 {
1650         return blk_alloc_queue_node(gfp_mask, -1);
1651 }
1652 EXPORT_SYMBOL(blk_alloc_queue);
1653
1654 request_queue_t *blk_alloc_queue_node(int gfp_mask, int node_id)
1655 {
1656         request_queue_t *q;
1657
1658         q = kmem_cache_alloc_node(requestq_cachep, gfp_mask, node_id);
1659         if (!q)
1660                 return NULL;
1661
1662         memset(q, 0, sizeof(*q));
1663         init_timer(&q->unplug_timer);
1664         atomic_set(&q->refcnt, 1);
1665
1666         q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1667         q->backing_dev_info.unplug_io_data = q;
1668
1669         return q;
1670 }
1671 EXPORT_SYMBOL(blk_alloc_queue_node);
1672
1673 /**
1674  * blk_init_queue  - prepare a request queue for use with a block device
1675  * @rfn:  The function to be called to process requests that have been
1676  *        placed on the queue.
1677  * @lock: Request queue spin lock
1678  *
1679  * Description:
1680  *    If a block device wishes to use the standard request handling procedures,
1681  *    which sorts requests and coalesces adjacent requests, then it must
1682  *    call blk_init_queue().  The function @rfn will be called when there
1683  *    are requests on the queue that need to be processed.  If the device
1684  *    supports plugging, then @rfn may not be called immediately when requests
1685  *    are available on the queue, but may be called at some time later instead.
1686  *    Plugged queues are generally unplugged when a buffer belonging to one
1687  *    of the requests on the queue is needed, or due to memory pressure.
1688  *
1689  *    @rfn is not required, or even expected, to remove all requests off the
1690  *    queue, but only as many as it can handle at a time.  If it does leave
1691  *    requests on the queue, it is responsible for arranging that the requests
1692  *    get dealt with eventually.
1693  *
1694  *    The queue spin lock must be held while manipulating the requests on the
1695  *    request queue.
1696  *
1697  *    Function returns a pointer to the initialized request queue, or NULL if
1698  *    it didn't succeed.
1699  *
1700  * Note:
1701  *    blk_init_queue() must be paired with a blk_cleanup_queue() call
1702  *    when the block device is deactivated (such as at module unload).
1703  **/
1704
1705 request_queue_t *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1706 {
1707         return blk_init_queue_node(rfn, lock, -1);
1708 }
1709 EXPORT_SYMBOL(blk_init_queue);
1710
1711 request_queue_t *
1712 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1713 {
1714         request_queue_t *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
1715
1716         if (!q)
1717                 return NULL;
1718
1719         q->node = node_id;
1720         if (blk_init_free_list(q))
1721                 goto out_init;
1722
1723         /*
1724          * if caller didn't supply a lock, they get per-queue locking with
1725          * our embedded lock
1726          */
1727         if (!lock) {
1728                 spin_lock_init(&q->__queue_lock);
1729                 lock = &q->__queue_lock;
1730         }
1731
1732         q->request_fn           = rfn;
1733         q->back_merge_fn        = ll_back_merge_fn;
1734         q->front_merge_fn       = ll_front_merge_fn;
1735         q->merge_requests_fn    = ll_merge_requests_fn;
1736         q->prep_rq_fn           = NULL;
1737         q->unplug_fn            = generic_unplug_device;
1738         q->queue_flags          = (1 << QUEUE_FLAG_CLUSTER);
1739         q->queue_lock           = lock;
1740
1741         blk_queue_segment_boundary(q, 0xffffffff);
1742
1743         blk_queue_make_request(q, __make_request);
1744         blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1745
1746         blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1747         blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1748
1749         /*
1750          * all done
1751          */
1752         if (!elevator_init(q, NULL)) {
1753                 blk_queue_congestion_threshold(q);
1754                 return q;
1755         }
1756
1757         blk_cleanup_queue(q);
1758 out_init:
1759         kmem_cache_free(requestq_cachep, q);
1760         return NULL;
1761 }
1762 EXPORT_SYMBOL(blk_init_queue_node);
1763
1764 int blk_get_queue(request_queue_t *q)
1765 {
1766         if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
1767                 atomic_inc(&q->refcnt);
1768                 return 0;
1769         }
1770
1771         return 1;
1772 }
1773
1774 EXPORT_SYMBOL(blk_get_queue);
1775
1776 static inline void blk_free_request(request_queue_t *q, struct request *rq)
1777 {
1778         elv_put_request(q, rq);
1779         mempool_free(rq, q->rq.rq_pool);
1780 }
1781
1782 static inline struct request *blk_alloc_request(request_queue_t *q, int rw,
1783                                                 int gfp_mask)
1784 {
1785         struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1786
1787         if (!rq)
1788                 return NULL;
1789
1790         /*
1791          * first three bits are identical in rq->flags and bio->bi_rw,
1792          * see bio.h and blkdev.h
1793          */
1794         rq->flags = rw;
1795
1796         if (!elv_set_request(q, rq, gfp_mask))
1797                 return rq;
1798
1799         mempool_free(rq, q->rq.rq_pool);
1800         return NULL;
1801 }
1802
1803 /*
1804  * ioc_batching returns true if the ioc is a valid batching request and
1805  * should be given priority access to a request.
1806  */
1807 static inline int ioc_batching(request_queue_t *q, struct io_context *ioc)
1808 {
1809         if (!ioc)
1810                 return 0;
1811
1812         /*
1813          * Make sure the process is able to allocate at least 1 request
1814          * even if the batch times out, otherwise we could theoretically
1815          * lose wakeups.
1816          */
1817         return ioc->nr_batch_requests == q->nr_batching ||
1818                 (ioc->nr_batch_requests > 0
1819                 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
1820 }
1821
1822 /*
1823  * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
1824  * will cause the process to be a "batcher" on all queues in the system. This
1825  * is the behaviour we want though - once it gets a wakeup it should be given
1826  * a nice run.
1827  */
1828 void ioc_set_batching(request_queue_t *q, struct io_context *ioc)
1829 {
1830         if (!ioc || ioc_batching(q, ioc))
1831                 return;
1832
1833         ioc->nr_batch_requests = q->nr_batching;
1834         ioc->last_waited = jiffies;
1835 }
1836
1837 static void __freed_request(request_queue_t *q, int rw)
1838 {
1839         struct request_list *rl = &q->rq;
1840
1841         if (rl->count[rw] < queue_congestion_off_threshold(q))
1842                 clear_queue_congested(q, rw);
1843
1844         if (rl->count[rw] + 1 <= q->nr_requests) {
1845                 if (waitqueue_active(&rl->wait[rw]))
1846                         wake_up(&rl->wait[rw]);
1847
1848                 blk_clear_queue_full(q, rw);
1849         }
1850 }
1851
1852 /*
1853  * A request has just been released.  Account for it, update the full and
1854  * congestion status, wake up any waiters.   Called under q->queue_lock.
1855  */
1856 static void freed_request(request_queue_t *q, int rw)
1857 {
1858         struct request_list *rl = &q->rq;
1859
1860         rl->count[rw]--;
1861
1862         __freed_request(q, rw);
1863
1864         if (unlikely(rl->starved[rw ^ 1]))
1865                 __freed_request(q, rw ^ 1);
1866
1867         if (!rl->count[READ] && !rl->count[WRITE]) {
1868                 smp_mb();
1869                 if (unlikely(waitqueue_active(&rl->drain)))
1870                         wake_up(&rl->drain);
1871         }
1872 }
1873
1874 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
1875 /*
1876  * Get a free request, queue_lock must not be held
1877  */
1878 static struct request *get_request(request_queue_t *q, int rw, int gfp_mask)
1879 {
1880         struct request *rq = NULL;
1881         struct request_list *rl = &q->rq;
1882         struct io_context *ioc = get_io_context(gfp_mask);
1883
1884         if (unlikely(test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags)))
1885                 goto out;
1886
1887         spin_lock_irq(q->queue_lock);
1888         if (rl->count[rw]+1 >= q->nr_requests) {
1889                 /*
1890                  * The queue will fill after this allocation, so set it as
1891                  * full, and mark this process as "batching". This process
1892                  * will be allowed to complete a batch of requests, others
1893                  * will be blocked.
1894                  */
1895                 if (!blk_queue_full(q, rw)) {
1896                         ioc_set_batching(q, ioc);
1897                         blk_set_queue_full(q, rw);
1898                 }
1899         }
1900
1901         switch (elv_may_queue(q, rw)) {
1902                 case ELV_MQUEUE_NO:
1903                         goto rq_starved;
1904                 case ELV_MQUEUE_MAY:
1905                         break;
1906                 case ELV_MQUEUE_MUST:
1907                         goto get_rq;
1908         }
1909
1910         if (blk_queue_full(q, rw) && !ioc_batching(q, ioc)) {
1911                 /*
1912                  * The queue is full and the allocating process is not a
1913                  * "batcher", and not exempted by the IO scheduler
1914                  */
1915                 spin_unlock_irq(q->queue_lock);
1916                 goto out;
1917         }
1918
1919 get_rq:
1920         rl->count[rw]++;
1921         rl->starved[rw] = 0;
1922         if (rl->count[rw] >= queue_congestion_on_threshold(q))
1923                 set_queue_congested(q, rw);
1924         spin_unlock_irq(q->queue_lock);
1925
1926         rq = blk_alloc_request(q, rw, gfp_mask);
1927         if (!rq) {
1928                 /*
1929                  * Allocation failed presumably due to memory. Undo anything
1930                  * we might have messed up.
1931                  *
1932                  * Allocating task should really be put onto the front of the
1933                  * wait queue, but this is pretty rare.
1934                  */
1935                 spin_lock_irq(q->queue_lock);
1936                 freed_request(q, rw);
1937
1938                 /*
1939                  * in the very unlikely event that allocation failed and no
1940                  * requests for this direction was pending, mark us starved
1941                  * so that freeing of a request in the other direction will
1942                  * notice us. another possible fix would be to split the
1943                  * rq mempool into READ and WRITE
1944                  */
1945 rq_starved:
1946                 if (unlikely(rl->count[rw] == 0))
1947                         rl->starved[rw] = 1;
1948
1949                 spin_unlock_irq(q->queue_lock);
1950                 goto out;
1951         }
1952
1953         if (ioc_batching(q, ioc))
1954                 ioc->nr_batch_requests--;
1955         
1956         rq_init(q, rq);
1957         rq->rl = rl;
1958 out:
1959         put_io_context(ioc);
1960         return rq;
1961 }
1962
1963 /*
1964  * No available requests for this queue, unplug the device and wait for some
1965  * requests to become available.
1966  */
1967 static struct request *get_request_wait(request_queue_t *q, int rw)
1968 {
1969         DEFINE_WAIT(wait);
1970         struct request *rq;
1971
1972         do {
1973                 struct request_list *rl = &q->rq;
1974
1975                 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
1976                                 TASK_UNINTERRUPTIBLE);
1977
1978                 rq = get_request(q, rw, GFP_NOIO);
1979
1980                 if (!rq) {
1981                         struct io_context *ioc;
1982
1983                         generic_unplug_device(q);
1984                         io_schedule();
1985
1986                         /*
1987                          * After sleeping, we become a "batching" process and
1988                          * will be able to allocate at least one request, and
1989                          * up to a big batch of them for a small period time.
1990                          * See ioc_batching, ioc_set_batching
1991                          */
1992                         ioc = get_io_context(GFP_NOIO);
1993                         ioc_set_batching(q, ioc);
1994                         put_io_context(ioc);
1995                 }
1996                 finish_wait(&rl->wait[rw], &wait);
1997         } while (!rq);
1998
1999         return rq;
2000 }
2001
2002 struct request *blk_get_request(request_queue_t *q, int rw, int gfp_mask)
2003 {
2004         struct request *rq;
2005
2006         BUG_ON(rw != READ && rw != WRITE);
2007
2008         if (gfp_mask & __GFP_WAIT)
2009                 rq = get_request_wait(q, rw);
2010         else
2011                 rq = get_request(q, rw, gfp_mask);
2012
2013         return rq;
2014 }
2015
2016 EXPORT_SYMBOL(blk_get_request);
2017
2018 /**
2019  * blk_requeue_request - put a request back on queue
2020  * @q:          request queue where request should be inserted
2021  * @rq:         request to be inserted
2022  *
2023  * Description:
2024  *    Drivers often keep queueing requests until the hardware cannot accept
2025  *    more, when that condition happens we need to put the request back
2026  *    on the queue. Must be called with queue lock held.
2027  */
2028 void blk_requeue_request(request_queue_t *q, struct request *rq)
2029 {
2030         if (blk_rq_tagged(rq))
2031                 blk_queue_end_tag(q, rq);
2032
2033         elv_requeue_request(q, rq);
2034 }
2035
2036 EXPORT_SYMBOL(blk_requeue_request);
2037
2038 /**
2039  * blk_insert_request - insert a special request in to a request queue
2040  * @q:          request queue where request should be inserted
2041  * @rq:         request to be inserted
2042  * @at_head:    insert request at head or tail of queue
2043  * @data:       private data
2044  *
2045  * Description:
2046  *    Many block devices need to execute commands asynchronously, so they don't
2047  *    block the whole kernel from preemption during request execution.  This is
2048  *    accomplished normally by inserting aritficial requests tagged as
2049  *    REQ_SPECIAL in to the corresponding request queue, and letting them be
2050  *    scheduled for actual execution by the request queue.
2051  *
2052  *    We have the option of inserting the head or the tail of the queue.
2053  *    Typically we use the tail for new ioctls and so forth.  We use the head
2054  *    of the queue for things like a QUEUE_FULL message from a device, or a
2055  *    host that is unable to accept a particular command.
2056  */
2057 void blk_insert_request(request_queue_t *q, struct request *rq,
2058                         int at_head, void *data)
2059 {
2060         int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2061         unsigned long flags;
2062
2063         /*
2064          * tell I/O scheduler that this isn't a regular read/write (ie it
2065          * must not attempt merges on this) and that it acts as a soft
2066          * barrier
2067          */
2068         rq->flags |= REQ_SPECIAL | REQ_SOFTBARRIER;
2069
2070         rq->special = data;
2071
2072         spin_lock_irqsave(q->queue_lock, flags);
2073
2074         /*
2075          * If command is tagged, release the tag
2076          */
2077         if (blk_rq_tagged(rq))
2078                 blk_queue_end_tag(q, rq);
2079
2080         drive_stat_acct(rq, rq->nr_sectors, 1);
2081         __elv_add_request(q, rq, where, 0);
2082
2083         if (blk_queue_plugged(q))
2084                 __generic_unplug_device(q);
2085         else
2086                 q->request_fn(q);
2087         spin_unlock_irqrestore(q->queue_lock, flags);
2088 }
2089
2090 EXPORT_SYMBOL(blk_insert_request);
2091
2092 /**
2093  * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2094  * @q:          request queue where request should be inserted
2095  * @rw:         READ or WRITE data
2096  * @ubuf:       the user buffer
2097  * @len:        length of user data
2098  *
2099  * Description:
2100  *    Data will be mapped directly for zero copy io, if possible. Otherwise
2101  *    a kernel bounce buffer is used.
2102  *
2103  *    A matching blk_rq_unmap_user() must be issued at the end of io, while
2104  *    still in process context.
2105  *
2106  *    Note: The mapped bio may need to be bounced through blk_queue_bounce()
2107  *    before being submitted to the device, as pages mapped may be out of
2108  *    reach. It's the callers responsibility to make sure this happens. The
2109  *    original bio must be passed back in to blk_rq_unmap_user() for proper
2110  *    unmapping.
2111  */
2112 struct request *blk_rq_map_user(request_queue_t *q, int rw, void __user *ubuf,
2113                                 unsigned int len)
2114 {
2115         unsigned long uaddr;
2116         struct request *rq;
2117         struct bio *bio;
2118
2119         if (len > (q->max_sectors << 9))
2120                 return ERR_PTR(-EINVAL);
2121         if ((!len && ubuf) || (len && !ubuf))
2122                 return ERR_PTR(-EINVAL);
2123
2124         rq = blk_get_request(q, rw, __GFP_WAIT);
2125         if (!rq)
2126                 return ERR_PTR(-ENOMEM);
2127
2128         /*
2129          * if alignment requirement is satisfied, map in user pages for
2130          * direct dma. else, set up kernel bounce buffers
2131          */
2132         uaddr = (unsigned long) ubuf;
2133         if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2134                 bio = bio_map_user(q, NULL, uaddr, len, rw == READ);
2135         else
2136                 bio = bio_copy_user(q, uaddr, len, rw == READ);
2137
2138         if (!IS_ERR(bio)) {
2139                 rq->bio = rq->biotail = bio;
2140                 blk_rq_bio_prep(q, rq, bio);
2141
2142                 rq->buffer = rq->data = NULL;
2143                 rq->data_len = len;
2144                 return rq;
2145         }
2146
2147         /*
2148          * bio is the err-ptr
2149          */
2150         blk_put_request(rq);
2151         return (struct request *) bio;
2152 }
2153
2154 EXPORT_SYMBOL(blk_rq_map_user);
2155
2156 /**
2157  * blk_rq_unmap_user - unmap a request with user data
2158  * @rq:         request to be unmapped
2159  * @bio:        bio for the request
2160  * @ulen:       length of user buffer
2161  *
2162  * Description:
2163  *    Unmap a request previously mapped by blk_rq_map_user().
2164  */
2165 int blk_rq_unmap_user(struct request *rq, struct bio *bio, unsigned int ulen)
2166 {
2167         int ret = 0;
2168
2169         if (bio) {
2170                 if (bio_flagged(bio, BIO_USER_MAPPED))
2171                         bio_unmap_user(bio);
2172                 else
2173                         ret = bio_uncopy_user(bio);
2174         }
2175
2176         blk_put_request(rq);
2177         return ret;
2178 }
2179
2180 EXPORT_SYMBOL(blk_rq_unmap_user);
2181
2182 /**
2183  * blk_execute_rq - insert a request into queue for execution
2184  * @q:          queue to insert the request in
2185  * @bd_disk:    matching gendisk
2186  * @rq:         request to insert
2187  *
2188  * Description:
2189  *    Insert a fully prepared request at the back of the io scheduler queue
2190  *    for execution.
2191  */
2192 int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
2193                    struct request *rq)
2194 {
2195         DECLARE_COMPLETION(wait);
2196         char sense[SCSI_SENSE_BUFFERSIZE];
2197         int err = 0;
2198
2199         rq->rq_disk = bd_disk;
2200
2201         /*
2202          * we need an extra reference to the request, so we can look at
2203          * it after io completion
2204          */
2205         rq->ref_count++;
2206
2207         if (!rq->sense) {
2208                 memset(sense, 0, sizeof(sense));
2209                 rq->sense = sense;
2210                 rq->sense_len = 0;
2211         }
2212
2213         rq->flags |= REQ_NOMERGE;
2214         rq->waiting = &wait;
2215         rq->end_io = blk_end_sync_rq;
2216         elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 1);
2217         generic_unplug_device(q);
2218         wait_for_completion(&wait);
2219         rq->waiting = NULL;
2220
2221         if (rq->errors)
2222                 err = -EIO;
2223
2224         return err;
2225 }
2226
2227 EXPORT_SYMBOL(blk_execute_rq);
2228
2229 /**
2230  * blkdev_issue_flush - queue a flush
2231  * @bdev:       blockdev to issue flush for
2232  * @error_sector:       error sector
2233  *
2234  * Description:
2235  *    Issue a flush for the block device in question. Caller can supply
2236  *    room for storing the error offset in case of a flush error, if they
2237  *    wish to.  Caller must run wait_for_completion() on its own.
2238  */
2239 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2240 {
2241         request_queue_t *q;
2242
2243         if (bdev->bd_disk == NULL)
2244                 return -ENXIO;
2245
2246         q = bdev_get_queue(bdev);
2247         if (!q)
2248                 return -ENXIO;
2249         if (!q->issue_flush_fn)
2250                 return -EOPNOTSUPP;
2251
2252         return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2253 }
2254
2255 EXPORT_SYMBOL(blkdev_issue_flush);
2256
2257 /**
2258  * blkdev_scsi_issue_flush_fn - issue flush for SCSI devices
2259  * @q:          device queue
2260  * @disk:       gendisk
2261  * @error_sector:       error offset
2262  *
2263  * Description:
2264  *    Devices understanding the SCSI command set, can use this function as
2265  *    a helper for issuing a cache flush. Note: driver is required to store
2266  *    the error offset (in case of error flushing) in ->sector of struct
2267  *    request.
2268  */
2269 int blkdev_scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
2270                                sector_t *error_sector)
2271 {
2272         struct request *rq = blk_get_request(q, WRITE, __GFP_WAIT);
2273         int ret;
2274
2275         rq->flags |= REQ_BLOCK_PC | REQ_SOFTBARRIER;
2276         rq->sector = 0;
2277         memset(rq->cmd, 0, sizeof(rq->cmd));
2278         rq->cmd[0] = 0x35;
2279         rq->cmd_len = 12;
2280         rq->data = NULL;
2281         rq->data_len = 0;
2282         rq->timeout = 60 * HZ;
2283
2284         ret = blk_execute_rq(q, disk, rq);
2285
2286         if (ret && error_sector)
2287                 *error_sector = rq->sector;
2288
2289         blk_put_request(rq);
2290         return ret;
2291 }
2292
2293 EXPORT_SYMBOL(blkdev_scsi_issue_flush_fn);
2294
2295 void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2296 {
2297         int rw = rq_data_dir(rq);
2298
2299         if (!blk_fs_request(rq) || !rq->rq_disk)
2300                 return;
2301
2302         if (rw == READ) {
2303                 __disk_stat_add(rq->rq_disk, read_sectors, nr_sectors);
2304                 if (!new_io)
2305                         __disk_stat_inc(rq->rq_disk, read_merges);
2306         } else if (rw == WRITE) {
2307                 __disk_stat_add(rq->rq_disk, write_sectors, nr_sectors);
2308                 if (!new_io)
2309                         __disk_stat_inc(rq->rq_disk, write_merges);
2310         }
2311         if (new_io) {
2312                 disk_round_stats(rq->rq_disk);
2313                 rq->rq_disk->in_flight++;
2314         }
2315 }
2316
2317 /*
2318  * add-request adds a request to the linked list.
2319  * queue lock is held and interrupts disabled, as we muck with the
2320  * request queue list.
2321  */
2322 static inline void add_request(request_queue_t * q, struct request * req)
2323 {
2324         drive_stat_acct(req, req->nr_sectors, 1);
2325
2326         if (q->activity_fn)
2327                 q->activity_fn(q->activity_data, rq_data_dir(req));
2328
2329         /*
2330          * elevator indicated where it wants this request to be
2331          * inserted at elevator_merge time
2332          */
2333         __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2334 }
2335  
2336 /*
2337  * disk_round_stats()   - Round off the performance stats on a struct
2338  * disk_stats.
2339  *
2340  * The average IO queue length and utilisation statistics are maintained
2341  * by observing the current state of the queue length and the amount of
2342  * time it has been in this state for.
2343  *
2344  * Normally, that accounting is done on IO completion, but that can result
2345  * in more than a second's worth of IO being accounted for within any one
2346  * second, leading to >100% utilisation.  To deal with that, we call this
2347  * function to do a round-off before returning the results when reading
2348  * /proc/diskstats.  This accounts immediately for all queue usage up to
2349  * the current jiffies and restarts the counters again.
2350  */
2351 void disk_round_stats(struct gendisk *disk)
2352 {
2353         unsigned long now = jiffies;
2354
2355         __disk_stat_add(disk, time_in_queue,
2356                         disk->in_flight * (now - disk->stamp));
2357         disk->stamp = now;
2358
2359         if (disk->in_flight)
2360                 __disk_stat_add(disk, io_ticks, (now - disk->stamp_idle));
2361         disk->stamp_idle = now;
2362 }
2363
2364 /*
2365  * queue lock must be held
2366  */
2367 static void __blk_put_request(request_queue_t *q, struct request *req)
2368 {
2369         struct request_list *rl = req->rl;
2370
2371         if (unlikely(!q))
2372                 return;
2373         if (unlikely(--req->ref_count))
2374                 return;
2375
2376         req->rq_status = RQ_INACTIVE;
2377         req->q = NULL;
2378         req->rl = NULL;
2379
2380         /*
2381          * Request may not have originated from ll_rw_blk. if not,
2382          * it didn't come out of our reserved rq pools
2383          */
2384         if (rl) {
2385                 int rw = rq_data_dir(req);
2386
2387                 elv_completed_request(q, req);
2388
2389                 BUG_ON(!list_empty(&req->queuelist));
2390
2391                 blk_free_request(q, req);
2392                 freed_request(q, rw);
2393         }
2394 }
2395
2396 void blk_put_request(struct request *req)
2397 {
2398         /*
2399          * if req->rl isn't set, this request didnt originate from the
2400          * block layer, so it's safe to just disregard it
2401          */
2402         if (req->rl) {
2403                 unsigned long flags;
2404                 request_queue_t *q = req->q;
2405
2406                 spin_lock_irqsave(q->queue_lock, flags);
2407                 __blk_put_request(q, req);
2408                 spin_unlock_irqrestore(q->queue_lock, flags);
2409         }
2410 }
2411
2412 EXPORT_SYMBOL(blk_put_request);
2413
2414 /**
2415  * blk_end_sync_rq - executes a completion event on a request
2416  * @rq: request to complete
2417  */
2418 void blk_end_sync_rq(struct request *rq)
2419 {
2420         struct completion *waiting = rq->waiting;
2421
2422         rq->waiting = NULL;
2423         __blk_put_request(rq->q, rq);
2424
2425         /*
2426          * complete last, if this is a stack request the process (and thus
2427          * the rq pointer) could be invalid right after this complete()
2428          */
2429         complete(waiting);
2430 }
2431 EXPORT_SYMBOL(blk_end_sync_rq);
2432
2433 /**
2434  * blk_congestion_wait - wait for a queue to become uncongested
2435  * @rw: READ or WRITE
2436  * @timeout: timeout in jiffies
2437  *
2438  * Waits for up to @timeout jiffies for a queue (any queue) to exit congestion.
2439  * If no queues are congested then just wait for the next request to be
2440  * returned.
2441  */
2442 long blk_congestion_wait(int rw, long timeout)
2443 {
2444         long ret;
2445         DEFINE_WAIT(wait);
2446         wait_queue_head_t *wqh = &congestion_wqh[rw];
2447
2448         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
2449         ret = io_schedule_timeout(timeout);
2450         finish_wait(wqh, &wait);
2451         return ret;
2452 }
2453
2454 EXPORT_SYMBOL(blk_congestion_wait);
2455
2456 /*
2457  * Has to be called with the request spinlock acquired
2458  */
2459 static int attempt_merge(request_queue_t *q, struct request *req,
2460                           struct request *next)
2461 {
2462         if (!rq_mergeable(req) || !rq_mergeable(next))
2463                 return 0;
2464
2465         /*
2466          * not contigious
2467          */
2468         if (req->sector + req->nr_sectors != next->sector)
2469                 return 0;
2470
2471         if (rq_data_dir(req) != rq_data_dir(next)
2472             || req->rq_disk != next->rq_disk
2473             || next->waiting || next->special)
2474                 return 0;
2475
2476         /*
2477          * If we are allowed to merge, then append bio list
2478          * from next to rq and release next. merge_requests_fn
2479          * will have updated segment counts, update sector
2480          * counts here.
2481          */
2482         if (!q->merge_requests_fn(q, req, next))
2483                 return 0;
2484
2485         /*
2486          * At this point we have either done a back merge
2487          * or front merge. We need the smaller start_time of
2488          * the merged requests to be the current request
2489          * for accounting purposes.
2490          */
2491         if (time_after(req->start_time, next->start_time))
2492                 req->start_time = next->start_time;
2493
2494         req->biotail->bi_next = next->bio;
2495         req->biotail = next->biotail;
2496
2497         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2498
2499         elv_merge_requests(q, req, next);
2500
2501         if (req->rq_disk) {
2502                 disk_round_stats(req->rq_disk);
2503                 req->rq_disk->in_flight--;
2504         }
2505
2506         __blk_put_request(q, next);
2507         return 1;
2508 }
2509
2510 static inline int attempt_back_merge(request_queue_t *q, struct request *rq)
2511 {
2512         struct request *next = elv_latter_request(q, rq);
2513
2514         if (next)
2515                 return attempt_merge(q, rq, next);
2516
2517         return 0;
2518 }
2519
2520 static inline int attempt_front_merge(request_queue_t *q, struct request *rq)
2521 {
2522         struct request *prev = elv_former_request(q, rq);
2523
2524         if (prev)
2525                 return attempt_merge(q, prev, rq);
2526
2527         return 0;
2528 }
2529
2530 /**
2531  * blk_attempt_remerge  - attempt to remerge active head with next request
2532  * @q:    The &request_queue_t belonging to the device
2533  * @rq:   The head request (usually)
2534  *
2535  * Description:
2536  *    For head-active devices, the queue can easily be unplugged so quickly
2537  *    that proper merging is not done on the front request. This may hurt
2538  *    performance greatly for some devices. The block layer cannot safely
2539  *    do merging on that first request for these queues, but the driver can
2540  *    call this function and make it happen any way. Only the driver knows
2541  *    when it is safe to do so.
2542  **/
2543 void blk_attempt_remerge(request_queue_t *q, struct request *rq)
2544 {
2545         unsigned long flags;
2546
2547         spin_lock_irqsave(q->queue_lock, flags);
2548         attempt_back_merge(q, rq);
2549         spin_unlock_irqrestore(q->queue_lock, flags);
2550 }
2551
2552 EXPORT_SYMBOL(blk_attempt_remerge);
2553
2554 /*
2555  * Non-locking blk_attempt_remerge variant.
2556  */
2557 void __blk_attempt_remerge(request_queue_t *q, struct request *rq)
2558 {
2559         attempt_back_merge(q, rq);
2560 }
2561
2562 EXPORT_SYMBOL(__blk_attempt_remerge);
2563
2564 static int __make_request(request_queue_t *q, struct bio *bio)
2565 {
2566         struct request *req, *freereq = NULL;
2567         int el_ret, rw, nr_sectors, cur_nr_sectors, barrier, err, sync;
2568         sector_t sector;
2569
2570         sector = bio->bi_sector;
2571         nr_sectors = bio_sectors(bio);
2572         cur_nr_sectors = bio_cur_sectors(bio);
2573
2574         rw = bio_data_dir(bio);
2575         sync = bio_sync(bio);
2576
2577         /*
2578          * low level driver can indicate that it wants pages above a
2579          * certain limit bounced to low memory (ie for highmem, or even
2580          * ISA dma in theory)
2581          */
2582         blk_queue_bounce(q, &bio);
2583
2584         spin_lock_prefetch(q->queue_lock);
2585
2586         barrier = bio_barrier(bio);
2587         if (unlikely(barrier) && (q->ordered == QUEUE_ORDERED_NONE)) {
2588                 err = -EOPNOTSUPP;
2589                 goto end_io;
2590         }
2591
2592 again:
2593         spin_lock_irq(q->queue_lock);
2594
2595         if (elv_queue_empty(q)) {
2596                 blk_plug_device(q);
2597                 goto get_rq;
2598         }
2599         if (barrier)
2600                 goto get_rq;
2601
2602         el_ret = elv_merge(q, &req, bio);
2603         switch (el_ret) {
2604                 case ELEVATOR_BACK_MERGE:
2605                         BUG_ON(!rq_mergeable(req));
2606
2607                         if (!q->back_merge_fn(q, req, bio))
2608                                 break;
2609
2610                         req->biotail->bi_next = bio;
2611                         req->biotail = bio;
2612                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2613                         drive_stat_acct(req, nr_sectors, 0);
2614                         if (!attempt_back_merge(q, req))
2615                                 elv_merged_request(q, req);
2616                         goto out;
2617
2618                 case ELEVATOR_FRONT_MERGE:
2619                         BUG_ON(!rq_mergeable(req));
2620
2621                         if (!q->front_merge_fn(q, req, bio))
2622                                 break;
2623
2624                         bio->bi_next = req->bio;
2625                         req->bio = bio;
2626
2627                         /*
2628                          * may not be valid. if the low level driver said
2629                          * it didn't need a bounce buffer then it better
2630                          * not touch req->buffer either...
2631                          */
2632                         req->buffer = bio_data(bio);
2633                         req->current_nr_sectors = cur_nr_sectors;
2634                         req->hard_cur_sectors = cur_nr_sectors;
2635                         req->sector = req->hard_sector = sector;
2636                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2637                         drive_stat_acct(req, nr_sectors, 0);
2638                         if (!attempt_front_merge(q, req))
2639                                 elv_merged_request(q, req);
2640                         goto out;
2641
2642                 /*
2643                  * elevator says don't/can't merge. get new request
2644                  */
2645                 case ELEVATOR_NO_MERGE:
2646                         break;
2647
2648                 default:
2649                         printk("elevator returned crap (%d)\n", el_ret);
2650                         BUG();
2651         }
2652
2653         /*
2654          * Grab a free request from the freelist - if that is empty, check
2655          * if we are doing read ahead and abort instead of blocking for
2656          * a free slot.
2657          */
2658 get_rq:
2659         if (freereq) {
2660                 req = freereq;
2661                 freereq = NULL;
2662         } else {
2663                 spin_unlock_irq(q->queue_lock);
2664                 if ((freereq = get_request(q, rw, GFP_ATOMIC)) == NULL) {
2665                         /*
2666                          * READA bit set
2667                          */
2668                         err = -EWOULDBLOCK;
2669                         if (bio_rw_ahead(bio))
2670                                 goto end_io;
2671         
2672                         freereq = get_request_wait(q, rw);
2673                 }
2674                 goto again;
2675         }
2676
2677         req->flags |= REQ_CMD;
2678
2679         /*
2680          * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2681          */
2682         if (bio_rw_ahead(bio) || bio_failfast(bio))
2683                 req->flags |= REQ_FAILFAST;
2684
2685         /*
2686          * REQ_BARRIER implies no merging, but lets make it explicit
2687          */
2688         if (unlikely(barrier))
2689                 req->flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2690
2691         req->errors = 0;
2692         req->hard_sector = req->sector = sector;
2693         req->hard_nr_sectors = req->nr_sectors = nr_sectors;
2694         req->current_nr_sectors = req->hard_cur_sectors = cur_nr_sectors;
2695         req->nr_phys_segments = bio_phys_segments(q, bio);
2696         req->nr_hw_segments = bio_hw_segments(q, bio);
2697         req->buffer = bio_data(bio);    /* see ->buffer comment above */
2698         req->waiting = NULL;
2699         req->bio = req->biotail = bio;
2700         req->rq_disk = bio->bi_bdev->bd_disk;
2701         req->start_time = jiffies;
2702
2703         add_request(q, req);
2704 out:
2705         if (freereq)
2706                 __blk_put_request(q, freereq);
2707         if (sync)
2708                 __generic_unplug_device(q);
2709
2710         spin_unlock_irq(q->queue_lock);
2711         return 0;
2712
2713 end_io:
2714         bio_endio(bio, nr_sectors << 9, err);
2715         return 0;
2716 }
2717
2718 /*
2719  * If bio->bi_dev is a partition, remap the location
2720  */
2721 static inline void blk_partition_remap(struct bio *bio)
2722 {
2723         struct block_device *bdev = bio->bi_bdev;
2724
2725         if (bdev != bdev->bd_contains) {
2726                 struct hd_struct *p = bdev->bd_part;
2727
2728                 switch (bio->bi_rw) {
2729                 case READ:
2730                         p->read_sectors += bio_sectors(bio);
2731                         p->reads++;
2732                         break;
2733                 case WRITE:
2734                         p->write_sectors += bio_sectors(bio);
2735                         p->writes++;
2736                         break;
2737                 }
2738                 bio->bi_sector += p->start_sect;
2739                 bio->bi_bdev = bdev->bd_contains;
2740         }
2741 }
2742
2743 void blk_finish_queue_drain(request_queue_t *q)
2744 {
2745         struct request_list *rl = &q->rq;
2746         struct request *rq;
2747
2748         spin_lock_irq(q->queue_lock);
2749         clear_bit(QUEUE_FLAG_DRAIN, &q->queue_flags);
2750
2751         while (!list_empty(&q->drain_list)) {
2752                 rq = list_entry_rq(q->drain_list.next);
2753
2754                 list_del_init(&rq->queuelist);
2755                 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 1);
2756         }
2757
2758         spin_unlock_irq(q->queue_lock);
2759
2760         wake_up(&rl->wait[0]);
2761         wake_up(&rl->wait[1]);
2762         wake_up(&rl->drain);
2763 }
2764
2765 static int wait_drain(request_queue_t *q, struct request_list *rl, int dispatch)
2766 {
2767         int wait = rl->count[READ] + rl->count[WRITE];
2768
2769         if (dispatch)
2770                 wait += !list_empty(&q->queue_head);
2771
2772         return wait;
2773 }
2774
2775 /*
2776  * We rely on the fact that only requests allocated through blk_alloc_request()
2777  * have io scheduler private data structures associated with them. Any other
2778  * type of request (allocated on stack or through kmalloc()) should not go
2779  * to the io scheduler core, but be attached to the queue head instead.
2780  */
2781 void blk_wait_queue_drained(request_queue_t *q, int wait_dispatch)
2782 {
2783         struct request_list *rl = &q->rq;
2784         DEFINE_WAIT(wait);
2785
2786         spin_lock_irq(q->queue_lock);
2787         set_bit(QUEUE_FLAG_DRAIN, &q->queue_flags);
2788
2789         while (wait_drain(q, rl, wait_dispatch)) {
2790                 prepare_to_wait(&rl->drain, &wait, TASK_UNINTERRUPTIBLE);
2791
2792                 if (wait_drain(q, rl, wait_dispatch)) {
2793                         __generic_unplug_device(q);
2794                         spin_unlock_irq(q->queue_lock);
2795                         io_schedule();
2796                         spin_lock_irq(q->queue_lock);
2797                 }
2798
2799                 finish_wait(&rl->drain, &wait);
2800         }
2801
2802         spin_unlock_irq(q->queue_lock);
2803 }
2804
2805 /*
2806  * block waiting for the io scheduler being started again.
2807  */
2808 static inline void block_wait_queue_running(request_queue_t *q)
2809 {
2810         DEFINE_WAIT(wait);
2811
2812         while (unlikely(test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags))) {
2813                 struct request_list *rl = &q->rq;
2814
2815                 prepare_to_wait_exclusive(&rl->drain, &wait,
2816                                 TASK_UNINTERRUPTIBLE);
2817
2818                 /*
2819                  * re-check the condition. avoids using prepare_to_wait()
2820                  * in the fast path (queue is running)
2821                  */
2822                 if (test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags))
2823                         io_schedule();
2824
2825                 finish_wait(&rl->drain, &wait);
2826         }
2827 }
2828
2829 static void handle_bad_sector(struct bio *bio)
2830 {
2831         char b[BDEVNAME_SIZE];
2832
2833         printk(KERN_INFO "attempt to access beyond end of device\n");
2834         printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
2835                         bdevname(bio->bi_bdev, b),
2836                         bio->bi_rw,
2837                         (unsigned long long)bio->bi_sector + bio_sectors(bio),
2838                         (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
2839
2840         set_bit(BIO_EOF, &bio->bi_flags);
2841 }
2842
2843 /**
2844  * generic_make_request: hand a buffer to its device driver for I/O
2845  * @bio:  The bio describing the location in memory and on the device.
2846  *
2847  * generic_make_request() is used to make I/O requests of block
2848  * devices. It is passed a &struct bio, which describes the I/O that needs
2849  * to be done.
2850  *
2851  * generic_make_request() does not return any status.  The
2852  * success/failure status of the request, along with notification of
2853  * completion, is delivered asynchronously through the bio->bi_end_io
2854  * function described (one day) else where.
2855  *
2856  * The caller of generic_make_request must make sure that bi_io_vec
2857  * are set to describe the memory buffer, and that bi_dev and bi_sector are
2858  * set to describe the device address, and the
2859  * bi_end_io and optionally bi_private are set to describe how
2860  * completion notification should be signaled.
2861  *
2862  * generic_make_request and the drivers it calls may use bi_next if this
2863  * bio happens to be merged with someone else, and may change bi_dev and
2864  * bi_sector for remaps as it sees fit.  So the values of these fields
2865  * should NOT be depended on after the call to generic_make_request.
2866  */
2867 void generic_make_request(struct bio *bio)
2868 {
2869         request_queue_t *q;
2870         sector_t maxsector;
2871         int ret, nr_sectors = bio_sectors(bio);
2872
2873         might_sleep();
2874         /* Test device or partition size, when known. */
2875         maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
2876         if (maxsector) {
2877                 sector_t sector = bio->bi_sector;
2878
2879                 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
2880                         /*
2881                          * This may well happen - the kernel calls bread()
2882                          * without checking the size of the device, e.g., when
2883                          * mounting a device.
2884                          */
2885                         handle_bad_sector(bio);
2886                         goto end_io;
2887                 }
2888         }
2889
2890         /*
2891          * Resolve the mapping until finished. (drivers are
2892          * still free to implement/resolve their own stacking
2893          * by explicitly returning 0)
2894          *
2895          * NOTE: we don't repeat the blk_size check for each new device.
2896          * Stacking drivers are expected to know what they are doing.
2897          */
2898         do {
2899                 char b[BDEVNAME_SIZE];
2900
2901                 q = bdev_get_queue(bio->bi_bdev);
2902                 if (!q) {
2903                         printk(KERN_ERR
2904                                "generic_make_request: Trying to access "
2905                                 "nonexistent block-device %s (%Lu)\n",
2906                                 bdevname(bio->bi_bdev, b),
2907                                 (long long) bio->bi_sector);
2908 end_io:
2909                         bio_endio(bio, bio->bi_size, -EIO);
2910                         break;
2911                 }
2912
2913                 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
2914                         printk("bio too big device %s (%u > %u)\n", 
2915                                 bdevname(bio->bi_bdev, b),
2916                                 bio_sectors(bio),
2917                                 q->max_hw_sectors);
2918                         goto end_io;
2919                 }
2920
2921                 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
2922                         goto end_io;
2923
2924                 block_wait_queue_running(q);
2925
2926                 /*
2927                  * If this device has partitions, remap block n
2928                  * of partition p to block n+start(p) of the disk.
2929                  */
2930                 blk_partition_remap(bio);
2931
2932                 ret = q->make_request_fn(q, bio);
2933         } while (ret);
2934 }
2935
2936 EXPORT_SYMBOL(generic_make_request);
2937
2938 /**
2939  * submit_bio: submit a bio to the block device layer for I/O
2940  * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
2941  * @bio: The &struct bio which describes the I/O
2942  *
2943  * submit_bio() is very similar in purpose to generic_make_request(), and
2944  * uses that function to do most of the work. Both are fairly rough
2945  * interfaces, @bio must be presetup and ready for I/O.
2946  *
2947  */
2948 void submit_bio(int rw, struct bio *bio)
2949 {
2950         int count = bio_sectors(bio);
2951
2952         BIO_BUG_ON(!bio->bi_size);
2953         BIO_BUG_ON(!bio->bi_io_vec);
2954         bio->bi_rw = rw;
2955         if (rw & WRITE)
2956                 mod_page_state(pgpgout, count);
2957         else
2958                 mod_page_state(pgpgin, count);
2959
2960         if (unlikely(block_dump)) {
2961                 char b[BDEVNAME_SIZE];
2962                 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
2963                         current->comm, current->pid,
2964                         (rw & WRITE) ? "WRITE" : "READ",
2965                         (unsigned long long)bio->bi_sector,
2966                         bdevname(bio->bi_bdev,b));
2967         }
2968
2969         generic_make_request(bio);
2970 }
2971
2972 EXPORT_SYMBOL(submit_bio);
2973
2974 void blk_recalc_rq_segments(struct request *rq)
2975 {
2976         struct bio *bio, *prevbio = NULL;
2977         int nr_phys_segs, nr_hw_segs;
2978         unsigned int phys_size, hw_size;
2979         request_queue_t *q = rq->q;
2980
2981         if (!rq->bio)
2982                 return;
2983
2984         phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
2985         rq_for_each_bio(bio, rq) {
2986                 /* Force bio hw/phys segs to be recalculated. */
2987                 bio->bi_flags &= ~(1 << BIO_SEG_VALID);
2988
2989                 nr_phys_segs += bio_phys_segments(q, bio);
2990                 nr_hw_segs += bio_hw_segments(q, bio);
2991                 if (prevbio) {
2992                         int pseg = phys_size + prevbio->bi_size + bio->bi_size;
2993                         int hseg = hw_size + prevbio->bi_size + bio->bi_size;
2994
2995                         if (blk_phys_contig_segment(q, prevbio, bio) &&
2996                             pseg <= q->max_segment_size) {
2997                                 nr_phys_segs--;
2998                                 phys_size += prevbio->bi_size + bio->bi_size;
2999                         } else
3000                                 phys_size = 0;
3001
3002                         if (blk_hw_contig_segment(q, prevbio, bio) &&
3003                             hseg <= q->max_segment_size) {
3004                                 nr_hw_segs--;
3005                                 hw_size += prevbio->bi_size + bio->bi_size;
3006                         } else
3007                                 hw_size = 0;
3008                 }
3009                 prevbio = bio;
3010         }
3011
3012         rq->nr_phys_segments = nr_phys_segs;
3013         rq->nr_hw_segments = nr_hw_segs;
3014 }
3015
3016 void blk_recalc_rq_sectors(struct request *rq, int nsect)
3017 {
3018         if (blk_fs_request(rq)) {
3019                 rq->hard_sector += nsect;
3020                 rq->hard_nr_sectors -= nsect;
3021
3022                 /*
3023                  * Move the I/O submission pointers ahead if required.
3024                  */
3025                 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3026                     (rq->sector <= rq->hard_sector)) {
3027                         rq->sector = rq->hard_sector;
3028                         rq->nr_sectors = rq->hard_nr_sectors;
3029                         rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3030                         rq->current_nr_sectors = rq->hard_cur_sectors;
3031                         rq->buffer = bio_data(rq->bio);
3032                 }
3033
3034                 /*
3035                  * if total number of sectors is less than the first segment
3036                  * size, something has gone terribly wrong
3037                  */
3038                 if (rq->nr_sectors < rq->current_nr_sectors) {
3039                         printk("blk: request botched\n");
3040                         rq->nr_sectors = rq->current_nr_sectors;
3041                 }
3042         }
3043 }
3044
3045 static int __end_that_request_first(struct request *req, int uptodate,
3046                                     int nr_bytes)
3047 {
3048         int total_bytes, bio_nbytes, error, next_idx = 0;
3049         struct bio *bio;
3050
3051         /*
3052          * extend uptodate bool to allow < 0 value to be direct io error
3053          */
3054         error = 0;
3055         if (end_io_error(uptodate))
3056                 error = !uptodate ? -EIO : uptodate;
3057
3058         /*
3059          * for a REQ_BLOCK_PC request, we want to carry any eventual
3060          * sense key with us all the way through
3061          */
3062         if (!blk_pc_request(req))
3063                 req->errors = 0;
3064
3065         if (!uptodate) {
3066                 if (blk_fs_request(req) && !(req->flags & REQ_QUIET))
3067                         printk("end_request: I/O error, dev %s, sector %llu\n",
3068                                 req->rq_disk ? req->rq_disk->disk_name : "?",
3069                                 (unsigned long long)req->sector);
3070         }
3071
3072         total_bytes = bio_nbytes = 0;
3073         while ((bio = req->bio) != NULL) {
3074                 int nbytes;
3075
3076                 if (nr_bytes >= bio->bi_size) {
3077                         req->bio = bio->bi_next;
3078                         nbytes = bio->bi_size;
3079                         bio_endio(bio, nbytes, error);
3080                         next_idx = 0;
3081                         bio_nbytes = 0;
3082                 } else {
3083                         int idx = bio->bi_idx + next_idx;
3084
3085                         if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3086                                 blk_dump_rq_flags(req, "__end_that");
3087                                 printk("%s: bio idx %d >= vcnt %d\n",
3088                                                 __FUNCTION__,
3089                                                 bio->bi_idx, bio->bi_vcnt);
3090                                 break;
3091                         }
3092
3093                         nbytes = bio_iovec_idx(bio, idx)->bv_len;
3094                         BIO_BUG_ON(nbytes > bio->bi_size);
3095
3096                         /*
3097                          * not a complete bvec done
3098                          */
3099                         if (unlikely(nbytes > nr_bytes)) {
3100                                 bio_nbytes += nr_bytes;
3101                                 total_bytes += nr_bytes;
3102                                 break;
3103                         }
3104
3105                         /*
3106                          * advance to the next vector
3107                          */
3108                         next_idx++;
3109                         bio_nbytes += nbytes;
3110                 }
3111
3112                 total_bytes += nbytes;
3113                 nr_bytes -= nbytes;
3114
3115                 if ((bio = req->bio)) {
3116                         /*
3117                          * end more in this run, or just return 'not-done'
3118                          */
3119                         if (unlikely(nr_bytes <= 0))
3120                                 break;
3121                 }
3122         }
3123
3124         /*
3125          * completely done
3126          */
3127         if (!req->bio)
3128                 return 0;
3129
3130         /*
3131          * if the request wasn't completed, update state
3132          */
3133         if (bio_nbytes) {
3134                 bio_endio(bio, bio_nbytes, error);
3135                 bio->bi_idx += next_idx;
3136                 bio_iovec(bio)->bv_offset += nr_bytes;
3137                 bio_iovec(bio)->bv_len -= nr_bytes;
3138         }
3139
3140         blk_recalc_rq_sectors(req, total_bytes >> 9);
3141         blk_recalc_rq_segments(req);
3142         return 1;
3143 }
3144
3145 /**
3146  * end_that_request_first - end I/O on a request
3147  * @req:      the request being processed
3148  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3149  * @nr_sectors: number of sectors to end I/O on
3150  *
3151  * Description:
3152  *     Ends I/O on a number of sectors attached to @req, and sets it up
3153  *     for the next range of segments (if any) in the cluster.
3154  *
3155  * Return:
3156  *     0 - we are done with this request, call end_that_request_last()
3157  *     1 - still buffers pending for this request
3158  **/
3159 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3160 {
3161         return __end_that_request_first(req, uptodate, nr_sectors << 9);
3162 }
3163
3164 EXPORT_SYMBOL(end_that_request_first);
3165
3166 /**
3167  * end_that_request_chunk - end I/O on a request
3168  * @req:      the request being processed
3169  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3170  * @nr_bytes: number of bytes to complete
3171  *
3172  * Description:
3173  *     Ends I/O on a number of bytes attached to @req, and sets it up
3174  *     for the next range of segments (if any). Like end_that_request_first(),
3175  *     but deals with bytes instead of sectors.
3176  *
3177  * Return:
3178  *     0 - we are done with this request, call end_that_request_last()
3179  *     1 - still buffers pending for this request
3180  **/
3181 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3182 {
3183         return __end_that_request_first(req, uptodate, nr_bytes);
3184 }
3185
3186 EXPORT_SYMBOL(end_that_request_chunk);
3187
3188 /*
3189  * queue lock must be held
3190  */
3191 void end_that_request_last(struct request *req)
3192 {
3193         struct gendisk *disk = req->rq_disk;
3194
3195         if (unlikely(laptop_mode) && blk_fs_request(req))
3196                 laptop_io_completion();
3197
3198         if (disk && blk_fs_request(req)) {
3199                 unsigned long duration = jiffies - req->start_time;
3200                 switch (rq_data_dir(req)) {
3201                     case WRITE:
3202                         __disk_stat_inc(disk, writes);
3203                         __disk_stat_add(disk, write_ticks, duration);
3204                         break;
3205                     case READ:
3206                         __disk_stat_inc(disk, reads);
3207                         __disk_stat_add(disk, read_ticks, duration);
3208                         break;
3209                 }
3210                 disk_round_stats(disk);
3211                 disk->in_flight--;
3212         }
3213         if (req->end_io)
3214                 req->end_io(req);
3215         else
3216                 __blk_put_request(req->q, req);
3217 }
3218
3219 EXPORT_SYMBOL(end_that_request_last);
3220
3221 void end_request(struct request *req, int uptodate)
3222 {
3223         if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3224                 add_disk_randomness(req->rq_disk);
3225                 blkdev_dequeue_request(req);
3226                 end_that_request_last(req);
3227         }
3228 }
3229
3230 EXPORT_SYMBOL(end_request);
3231
3232 void blk_rq_bio_prep(request_queue_t *q, struct request *rq, struct bio *bio)
3233 {
3234         /* first three bits are identical in rq->flags and bio->bi_rw */
3235         rq->flags |= (bio->bi_rw & 7);
3236
3237         rq->nr_phys_segments = bio_phys_segments(q, bio);
3238         rq->nr_hw_segments = bio_hw_segments(q, bio);
3239         rq->current_nr_sectors = bio_cur_sectors(bio);
3240         rq->hard_cur_sectors = rq->current_nr_sectors;
3241         rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3242         rq->buffer = bio_data(bio);
3243
3244         rq->bio = rq->biotail = bio;
3245 }
3246
3247 EXPORT_SYMBOL(blk_rq_bio_prep);
3248
3249 int kblockd_schedule_work(struct work_struct *work)
3250 {
3251         return queue_work(kblockd_workqueue, work);
3252 }
3253
3254 EXPORT_SYMBOL(kblockd_schedule_work);
3255
3256 void kblockd_flush(void)
3257 {
3258         flush_workqueue(kblockd_workqueue);
3259 }
3260 EXPORT_SYMBOL(kblockd_flush);
3261
3262 int __init blk_dev_init(void)
3263 {
3264         kblockd_workqueue = create_workqueue("kblockd");
3265         if (!kblockd_workqueue)
3266                 panic("Failed to create kblockd\n");
3267
3268         request_cachep = kmem_cache_create("blkdev_requests",
3269                         sizeof(struct request), 0, SLAB_PANIC, NULL, NULL);
3270
3271         requestq_cachep = kmem_cache_create("blkdev_queue",
3272                         sizeof(request_queue_t), 0, SLAB_PANIC, NULL, NULL);
3273
3274         iocontext_cachep = kmem_cache_create("blkdev_ioc",
3275                         sizeof(struct io_context), 0, SLAB_PANIC, NULL, NULL);
3276
3277         blk_max_low_pfn = max_low_pfn;
3278         blk_max_pfn = max_pfn;
3279
3280         return 0;
3281 }
3282
3283 /*
3284  * IO Context helper functions
3285  */
3286 void put_io_context(struct io_context *ioc)
3287 {
3288         if (ioc == NULL)
3289                 return;
3290
3291         BUG_ON(atomic_read(&ioc->refcount) == 0);
3292
3293         if (atomic_dec_and_test(&ioc->refcount)) {
3294                 if (ioc->aic && ioc->aic->dtor)
3295                         ioc->aic->dtor(ioc->aic);
3296                 if (ioc->cic && ioc->cic->dtor)
3297                         ioc->cic->dtor(ioc->cic);
3298
3299                 kmem_cache_free(iocontext_cachep, ioc);
3300         }
3301 }
3302 EXPORT_SYMBOL(put_io_context);
3303
3304 /* Called by the exitting task */
3305 void exit_io_context(void)
3306 {
3307         unsigned long flags;
3308         struct io_context *ioc;
3309
3310         local_irq_save(flags);
3311         ioc = current->io_context;
3312         current->io_context = NULL;
3313         local_irq_restore(flags);
3314
3315         if (ioc->aic && ioc->aic->exit)
3316                 ioc->aic->exit(ioc->aic);
3317         if (ioc->cic && ioc->cic->exit)
3318                 ioc->cic->exit(ioc->cic);
3319
3320         put_io_context(ioc);
3321 }
3322
3323 /*
3324  * If the current task has no IO context then create one and initialise it.
3325  * If it does have a context, take a ref on it.
3326  *
3327  * This is always called in the context of the task which submitted the I/O.
3328  * But weird things happen, so we disable local interrupts to ensure exclusive
3329  * access to *current.
3330  */
3331 struct io_context *get_io_context(int gfp_flags)
3332 {
3333         struct task_struct *tsk = current;
3334         unsigned long flags;
3335         struct io_context *ret;
3336
3337         local_irq_save(flags);
3338         ret = tsk->io_context;
3339         if (ret)
3340                 goto out;
3341
3342         local_irq_restore(flags);
3343
3344         ret = kmem_cache_alloc(iocontext_cachep, gfp_flags);
3345         if (ret) {
3346                 atomic_set(&ret->refcount, 1);
3347                 ret->pid = tsk->pid;
3348                 ret->last_waited = jiffies; /* doesn't matter... */
3349                 ret->nr_batch_requests = 0; /* because this is 0 */
3350                 ret->aic = NULL;
3351                 ret->cic = NULL;
3352                 spin_lock_init(&ret->lock);
3353
3354                 local_irq_save(flags);
3355
3356                 /*
3357                  * very unlikely, someone raced with us in setting up the task
3358                  * io context. free new context and just grab a reference.
3359                  */
3360                 if (!tsk->io_context)
3361                         tsk->io_context = ret;
3362                 else {
3363                         kmem_cache_free(iocontext_cachep, ret);
3364                         ret = tsk->io_context;
3365                 }
3366
3367 out:
3368                 atomic_inc(&ret->refcount);
3369                 local_irq_restore(flags);
3370         }
3371
3372         return ret;
3373 }
3374 EXPORT_SYMBOL(get_io_context);
3375
3376 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3377 {
3378         struct io_context *src = *psrc;
3379         struct io_context *dst = *pdst;
3380
3381         if (src) {
3382                 BUG_ON(atomic_read(&src->refcount) == 0);
3383                 atomic_inc(&src->refcount);
3384                 put_io_context(dst);
3385                 *pdst = src;
3386         }
3387 }
3388 EXPORT_SYMBOL(copy_io_context);
3389
3390 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3391 {
3392         struct io_context *temp;
3393         temp = *ioc1;
3394         *ioc1 = *ioc2;
3395         *ioc2 = temp;
3396 }
3397 EXPORT_SYMBOL(swap_io_context);
3398
3399 /*
3400  * sysfs parts below
3401  */
3402 struct queue_sysfs_entry {
3403         struct attribute attr;
3404         ssize_t (*show)(struct request_queue *, char *);
3405         ssize_t (*store)(struct request_queue *, const char *, size_t);
3406 };
3407
3408 static ssize_t
3409 queue_var_show(unsigned int var, char *page)
3410 {
3411         return sprintf(page, "%d\n", var);
3412 }
3413
3414 static ssize_t
3415 queue_var_store(unsigned long *var, const char *page, size_t count)
3416 {
3417         char *p = (char *) page;
3418
3419         *var = simple_strtoul(p, &p, 10);
3420         return count;
3421 }
3422
3423 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3424 {
3425         return queue_var_show(q->nr_requests, (page));
3426 }
3427
3428 static ssize_t
3429 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3430 {
3431         struct request_list *rl = &q->rq;
3432
3433         int ret = queue_var_store(&q->nr_requests, page, count);
3434         if (q->nr_requests < BLKDEV_MIN_RQ)
3435                 q->nr_requests = BLKDEV_MIN_RQ;
3436         blk_queue_congestion_threshold(q);
3437
3438         if (rl->count[READ] >= queue_congestion_on_threshold(q))
3439                 set_queue_congested(q, READ);
3440         else if (rl->count[READ] < queue_congestion_off_threshold(q))
3441                 clear_queue_congested(q, READ);
3442
3443         if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3444                 set_queue_congested(q, WRITE);
3445         else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3446                 clear_queue_congested(q, WRITE);
3447
3448         if (rl->count[READ] >= q->nr_requests) {
3449                 blk_set_queue_full(q, READ);
3450         } else if (rl->count[READ]+1 <= q->nr_requests) {
3451                 blk_clear_queue_full(q, READ);
3452                 wake_up(&rl->wait[READ]);
3453         }
3454
3455         if (rl->count[WRITE] >= q->nr_requests) {
3456                 blk_set_queue_full(q, WRITE);
3457         } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3458                 blk_clear_queue_full(q, WRITE);
3459                 wake_up(&rl->wait[WRITE]);
3460         }
3461         return ret;
3462 }
3463
3464 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3465 {
3466         int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3467
3468         return queue_var_show(ra_kb, (page));
3469 }
3470
3471 static ssize_t
3472 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3473 {
3474         unsigned long ra_kb;
3475         ssize_t ret = queue_var_store(&ra_kb, page, count);
3476
3477         spin_lock_irq(q->queue_lock);
3478         if (ra_kb > (q->max_sectors >> 1))
3479                 ra_kb = (q->max_sectors >> 1);
3480
3481         q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3482         spin_unlock_irq(q->queue_lock);
3483
3484         return ret;
3485 }
3486
3487 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3488 {
3489         int max_sectors_kb = q->max_sectors >> 1;
3490
3491         return queue_var_show(max_sectors_kb, (page));
3492 }
3493
3494 static ssize_t
3495 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3496 {
3497         unsigned long max_sectors_kb,
3498                         max_hw_sectors_kb = q->max_hw_sectors >> 1,
3499                         page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3500         ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3501         int ra_kb;
3502
3503         if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3504                 return -EINVAL;
3505         /*
3506          * Take the queue lock to update the readahead and max_sectors
3507          * values synchronously:
3508          */
3509         spin_lock_irq(q->queue_lock);
3510         /*
3511          * Trim readahead window as well, if necessary:
3512          */
3513         ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3514         if (ra_kb > max_sectors_kb)
3515                 q->backing_dev_info.ra_pages =
3516                                 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3517
3518         q->max_sectors = max_sectors_kb << 1;
3519         spin_unlock_irq(q->queue_lock);
3520
3521         return ret;
3522 }
3523
3524 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3525 {
3526         int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3527
3528         return queue_var_show(max_hw_sectors_kb, (page));
3529 }
3530
3531
3532 static struct queue_sysfs_entry queue_requests_entry = {
3533         .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3534         .show = queue_requests_show,
3535         .store = queue_requests_store,
3536 };
3537
3538 static struct queue_sysfs_entry queue_ra_entry = {
3539         .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3540         .show = queue_ra_show,
3541         .store = queue_ra_store,
3542 };
3543
3544 static struct queue_sysfs_entry queue_max_sectors_entry = {
3545         .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3546         .show = queue_max_sectors_show,
3547         .store = queue_max_sectors_store,
3548 };
3549
3550 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3551         .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
3552         .show = queue_max_hw_sectors_show,
3553 };
3554
3555 static struct queue_sysfs_entry queue_iosched_entry = {
3556         .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
3557         .show = elv_iosched_show,
3558         .store = elv_iosched_store,
3559 };
3560
3561 static struct attribute *default_attrs[] = {
3562         &queue_requests_entry.attr,
3563         &queue_ra_entry.attr,
3564         &queue_max_hw_sectors_entry.attr,
3565         &queue_max_sectors_entry.attr,
3566         &queue_iosched_entry.attr,
3567         NULL,
3568 };
3569
3570 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
3571
3572 static ssize_t
3573 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
3574 {
3575         struct queue_sysfs_entry *entry = to_queue(attr);
3576         struct request_queue *q;
3577
3578         q = container_of(kobj, struct request_queue, kobj);
3579         if (!entry->show)
3580                 return -EIO;
3581
3582         return entry->show(q, page);
3583 }
3584
3585 static ssize_t
3586 queue_attr_store(struct kobject *kobj, struct attribute *attr,
3587                     const char *page, size_t length)
3588 {
3589         struct queue_sysfs_entry *entry = to_queue(attr);
3590         struct request_queue *q;
3591
3592         q = container_of(kobj, struct request_queue, kobj);
3593         if (!entry->store)
3594                 return -EIO;
3595
3596         return entry->store(q, page, length);
3597 }
3598
3599 static struct sysfs_ops queue_sysfs_ops = {
3600         .show   = queue_attr_show,
3601         .store  = queue_attr_store,
3602 };
3603
3604 struct kobj_type queue_ktype = {
3605         .sysfs_ops      = &queue_sysfs_ops,
3606         .default_attrs  = default_attrs,
3607 };
3608
3609 int blk_register_queue(struct gendisk *disk)
3610 {
3611         int ret;
3612
3613         request_queue_t *q = disk->queue;
3614
3615         if (!q || !q->request_fn)
3616                 return -ENXIO;
3617
3618         q->kobj.parent = kobject_get(&disk->kobj);
3619         if (!q->kobj.parent)
3620                 return -EBUSY;
3621
3622         snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
3623         q->kobj.ktype = &queue_ktype;
3624
3625         ret = kobject_register(&q->kobj);
3626         if (ret < 0)
3627                 return ret;
3628
3629         ret = elv_register_queue(q);
3630         if (ret) {
3631                 kobject_unregister(&q->kobj);
3632                 return ret;
3633         }
3634
3635         return 0;
3636 }
3637
3638 void blk_unregister_queue(struct gendisk *disk)
3639 {
3640         request_queue_t *q = disk->queue;
3641
3642         if (q && q->request_fn) {
3643                 elv_unregister_queue(q);
3644
3645                 kobject_unregister(&q->kobj);
3646                 kobject_put(&disk->kobj);
3647         }
3648 }