[PATCH] fix NMI lockup with CFQ scheduler
[linux-2.6.git] / drivers / block / ll_rw_blk.c
1 /*
2  *  linux/drivers/block/ll_rw_blk.c
3  *
4  * Copyright (C) 1991, 1992 Linus Torvalds
5  * Copyright (C) 1994,      Karl Keyte: Added support for disk statistics
6  * Elevator latency, (C) 2000  Andrea Arcangeli <andrea@suse.de> SuSE
7  * Queue request tables / lock, selectable elevator, Jens Axboe <axboe@suse.de>
8  * kernel-doc documentation started by NeilBrown <neilb@cse.unsw.edu.au> -  July2000
9  * bio rewrite, highmem i/o, etc, Jens Axboe <axboe@suse.de> - may 2001
10  */
11
12 /*
13  * This handles all read/write requests to block devices
14  */
15 #include <linux/config.h>
16 #include <linux/kernel.h>
17 #include <linux/module.h>
18 #include <linux/backing-dev.h>
19 #include <linux/bio.h>
20 #include <linux/blkdev.h>
21 #include <linux/highmem.h>
22 #include <linux/mm.h>
23 #include <linux/kernel_stat.h>
24 #include <linux/string.h>
25 #include <linux/init.h>
26 #include <linux/bootmem.h>      /* for max_pfn/max_low_pfn */
27 #include <linux/completion.h>
28 #include <linux/slab.h>
29 #include <linux/swap.h>
30 #include <linux/writeback.h>
31
32 /*
33  * for max sense size
34  */
35 #include <scsi/scsi_cmnd.h>
36
37 static void blk_unplug_work(void *data);
38 static void blk_unplug_timeout(unsigned long data);
39
40 /*
41  * For the allocated request tables
42  */
43 static kmem_cache_t *request_cachep;
44
45 /*
46  * For queue allocation
47  */
48 static kmem_cache_t *requestq_cachep;
49
50 /*
51  * For io context allocations
52  */
53 static kmem_cache_t *iocontext_cachep;
54
55 static wait_queue_head_t congestion_wqh[2] = {
56                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
57                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
58         };
59
60 /*
61  * Controlling structure to kblockd
62  */
63 static struct workqueue_struct *kblockd_workqueue; 
64
65 unsigned long blk_max_low_pfn, blk_max_pfn;
66
67 EXPORT_SYMBOL(blk_max_low_pfn);
68 EXPORT_SYMBOL(blk_max_pfn);
69
70 /* Amount of time in which a process may batch requests */
71 #define BLK_BATCH_TIME  (HZ/50UL)
72
73 /* Number of requests a "batching" process may submit */
74 #define BLK_BATCH_REQ   32
75
76 /*
77  * Return the threshold (number of used requests) at which the queue is
78  * considered to be congested.  It include a little hysteresis to keep the
79  * context switch rate down.
80  */
81 static inline int queue_congestion_on_threshold(struct request_queue *q)
82 {
83         return q->nr_congestion_on;
84 }
85
86 /*
87  * The threshold at which a queue is considered to be uncongested
88  */
89 static inline int queue_congestion_off_threshold(struct request_queue *q)
90 {
91         return q->nr_congestion_off;
92 }
93
94 static void blk_queue_congestion_threshold(struct request_queue *q)
95 {
96         int nr;
97
98         nr = q->nr_requests - (q->nr_requests / 8) + 1;
99         if (nr > q->nr_requests)
100                 nr = q->nr_requests;
101         q->nr_congestion_on = nr;
102
103         nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
104         if (nr < 1)
105                 nr = 1;
106         q->nr_congestion_off = nr;
107 }
108
109 /*
110  * A queue has just exitted congestion.  Note this in the global counter of
111  * congested queues, and wake up anyone who was waiting for requests to be
112  * put back.
113  */
114 static void clear_queue_congested(request_queue_t *q, int rw)
115 {
116         enum bdi_state bit;
117         wait_queue_head_t *wqh = &congestion_wqh[rw];
118
119         bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
120         clear_bit(bit, &q->backing_dev_info.state);
121         smp_mb__after_clear_bit();
122         if (waitqueue_active(wqh))
123                 wake_up(wqh);
124 }
125
126 /*
127  * A queue has just entered congestion.  Flag that in the queue's VM-visible
128  * state flags and increment the global gounter of congested queues.
129  */
130 static void set_queue_congested(request_queue_t *q, int rw)
131 {
132         enum bdi_state bit;
133
134         bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
135         set_bit(bit, &q->backing_dev_info.state);
136 }
137
138 /**
139  * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
140  * @bdev:       device
141  *
142  * Locates the passed device's request queue and returns the address of its
143  * backing_dev_info
144  *
145  * Will return NULL if the request queue cannot be located.
146  */
147 struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
148 {
149         struct backing_dev_info *ret = NULL;
150         request_queue_t *q = bdev_get_queue(bdev);
151
152         if (q)
153                 ret = &q->backing_dev_info;
154         return ret;
155 }
156
157 EXPORT_SYMBOL(blk_get_backing_dev_info);
158
159 void blk_queue_activity_fn(request_queue_t *q, activity_fn *fn, void *data)
160 {
161         q->activity_fn = fn;
162         q->activity_data = data;
163 }
164
165 EXPORT_SYMBOL(blk_queue_activity_fn);
166
167 /**
168  * blk_queue_prep_rq - set a prepare_request function for queue
169  * @q:          queue
170  * @pfn:        prepare_request function
171  *
172  * It's possible for a queue to register a prepare_request callback which
173  * is invoked before the request is handed to the request_fn. The goal of
174  * the function is to prepare a request for I/O, it can be used to build a
175  * cdb from the request data for instance.
176  *
177  */
178 void blk_queue_prep_rq(request_queue_t *q, prep_rq_fn *pfn)
179 {
180         q->prep_rq_fn = pfn;
181 }
182
183 EXPORT_SYMBOL(blk_queue_prep_rq);
184
185 /**
186  * blk_queue_merge_bvec - set a merge_bvec function for queue
187  * @q:          queue
188  * @mbfn:       merge_bvec_fn
189  *
190  * Usually queues have static limitations on the max sectors or segments that
191  * we can put in a request. Stacking drivers may have some settings that
192  * are dynamic, and thus we have to query the queue whether it is ok to
193  * add a new bio_vec to a bio at a given offset or not. If the block device
194  * has such limitations, it needs to register a merge_bvec_fn to control
195  * the size of bio's sent to it. Note that a block device *must* allow a
196  * single page to be added to an empty bio. The block device driver may want
197  * to use the bio_split() function to deal with these bio's. By default
198  * no merge_bvec_fn is defined for a queue, and only the fixed limits are
199  * honored.
200  */
201 void blk_queue_merge_bvec(request_queue_t *q, merge_bvec_fn *mbfn)
202 {
203         q->merge_bvec_fn = mbfn;
204 }
205
206 EXPORT_SYMBOL(blk_queue_merge_bvec);
207
208 /**
209  * blk_queue_make_request - define an alternate make_request function for a device
210  * @q:  the request queue for the device to be affected
211  * @mfn: the alternate make_request function
212  *
213  * Description:
214  *    The normal way for &struct bios to be passed to a device
215  *    driver is for them to be collected into requests on a request
216  *    queue, and then to allow the device driver to select requests
217  *    off that queue when it is ready.  This works well for many block
218  *    devices. However some block devices (typically virtual devices
219  *    such as md or lvm) do not benefit from the processing on the
220  *    request queue, and are served best by having the requests passed
221  *    directly to them.  This can be achieved by providing a function
222  *    to blk_queue_make_request().
223  *
224  * Caveat:
225  *    The driver that does this *must* be able to deal appropriately
226  *    with buffers in "highmemory". This can be accomplished by either calling
227  *    __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
228  *    blk_queue_bounce() to create a buffer in normal memory.
229  **/
230 void blk_queue_make_request(request_queue_t * q, make_request_fn * mfn)
231 {
232         /*
233          * set defaults
234          */
235         q->nr_requests = BLKDEV_MAX_RQ;
236         q->max_phys_segments = MAX_PHYS_SEGMENTS;
237         q->max_hw_segments = MAX_HW_SEGMENTS;
238         q->make_request_fn = mfn;
239         q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
240         q->backing_dev_info.state = 0;
241         q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
242         blk_queue_max_sectors(q, MAX_SECTORS);
243         blk_queue_hardsect_size(q, 512);
244         blk_queue_dma_alignment(q, 511);
245         blk_queue_congestion_threshold(q);
246         q->nr_batching = BLK_BATCH_REQ;
247
248         q->unplug_thresh = 4;           /* hmm */
249         q->unplug_delay = (3 * HZ) / 1000;      /* 3 milliseconds */
250         if (q->unplug_delay == 0)
251                 q->unplug_delay = 1;
252
253         INIT_WORK(&q->unplug_work, blk_unplug_work, q);
254
255         q->unplug_timer.function = blk_unplug_timeout;
256         q->unplug_timer.data = (unsigned long)q;
257
258         /*
259          * by default assume old behaviour and bounce for any highmem page
260          */
261         blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
262
263         blk_queue_activity_fn(q, NULL, NULL);
264
265         INIT_LIST_HEAD(&q->drain_list);
266 }
267
268 EXPORT_SYMBOL(blk_queue_make_request);
269
270 static inline void rq_init(request_queue_t *q, struct request *rq)
271 {
272         INIT_LIST_HEAD(&rq->queuelist);
273
274         rq->errors = 0;
275         rq->rq_status = RQ_ACTIVE;
276         rq->bio = rq->biotail = NULL;
277         rq->buffer = NULL;
278         rq->ref_count = 1;
279         rq->q = q;
280         rq->waiting = NULL;
281         rq->special = NULL;
282         rq->data_len = 0;
283         rq->data = NULL;
284         rq->sense = NULL;
285         rq->end_io = NULL;
286         rq->end_io_data = NULL;
287 }
288
289 /**
290  * blk_queue_ordered - does this queue support ordered writes
291  * @q:     the request queue
292  * @flag:  see below
293  *
294  * Description:
295  *   For journalled file systems, doing ordered writes on a commit
296  *   block instead of explicitly doing wait_on_buffer (which is bad
297  *   for performance) can be a big win. Block drivers supporting this
298  *   feature should call this function and indicate so.
299  *
300  **/
301 void blk_queue_ordered(request_queue_t *q, int flag)
302 {
303         switch (flag) {
304                 case QUEUE_ORDERED_NONE:
305                         if (q->flush_rq)
306                                 kmem_cache_free(request_cachep, q->flush_rq);
307                         q->flush_rq = NULL;
308                         q->ordered = flag;
309                         break;
310                 case QUEUE_ORDERED_TAG:
311                         q->ordered = flag;
312                         break;
313                 case QUEUE_ORDERED_FLUSH:
314                         q->ordered = flag;
315                         if (!q->flush_rq)
316                                 q->flush_rq = kmem_cache_alloc(request_cachep,
317                                                                 GFP_KERNEL);
318                         break;
319                 default:
320                         printk("blk_queue_ordered: bad value %d\n", flag);
321                         break;
322         }
323 }
324
325 EXPORT_SYMBOL(blk_queue_ordered);
326
327 /**
328  * blk_queue_issue_flush_fn - set function for issuing a flush
329  * @q:     the request queue
330  * @iff:   the function to be called issuing the flush
331  *
332  * Description:
333  *   If a driver supports issuing a flush command, the support is notified
334  *   to the block layer by defining it through this call.
335  *
336  **/
337 void blk_queue_issue_flush_fn(request_queue_t *q, issue_flush_fn *iff)
338 {
339         q->issue_flush_fn = iff;
340 }
341
342 EXPORT_SYMBOL(blk_queue_issue_flush_fn);
343
344 /*
345  * Cache flushing for ordered writes handling
346  */
347 static void blk_pre_flush_end_io(struct request *flush_rq)
348 {
349         struct request *rq = flush_rq->end_io_data;
350         request_queue_t *q = rq->q;
351
352         rq->flags |= REQ_BAR_PREFLUSH;
353
354         if (!flush_rq->errors)
355                 elv_requeue_request(q, rq);
356         else {
357                 q->end_flush_fn(q, flush_rq);
358                 clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
359                 q->request_fn(q);
360         }
361 }
362
363 static void blk_post_flush_end_io(struct request *flush_rq)
364 {
365         struct request *rq = flush_rq->end_io_data;
366         request_queue_t *q = rq->q;
367
368         rq->flags |= REQ_BAR_POSTFLUSH;
369
370         q->end_flush_fn(q, flush_rq);
371         clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
372         q->request_fn(q);
373 }
374
375 struct request *blk_start_pre_flush(request_queue_t *q, struct request *rq)
376 {
377         struct request *flush_rq = q->flush_rq;
378
379         BUG_ON(!blk_barrier_rq(rq));
380
381         if (test_and_set_bit(QUEUE_FLAG_FLUSH, &q->queue_flags))
382                 return NULL;
383
384         rq_init(q, flush_rq);
385         flush_rq->elevator_private = NULL;
386         flush_rq->flags = REQ_BAR_FLUSH;
387         flush_rq->rq_disk = rq->rq_disk;
388         flush_rq->rl = NULL;
389
390         /*
391          * prepare_flush returns 0 if no flush is needed, just mark both
392          * pre and post flush as done in that case
393          */
394         if (!q->prepare_flush_fn(q, flush_rq)) {
395                 rq->flags |= REQ_BAR_PREFLUSH | REQ_BAR_POSTFLUSH;
396                 clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
397                 return rq;
398         }
399
400         /*
401          * some drivers dequeue requests right away, some only after io
402          * completion. make sure the request is dequeued.
403          */
404         if (!list_empty(&rq->queuelist))
405                 blkdev_dequeue_request(rq);
406
407         elv_deactivate_request(q, rq);
408
409         flush_rq->end_io_data = rq;
410         flush_rq->end_io = blk_pre_flush_end_io;
411
412         __elv_add_request(q, flush_rq, ELEVATOR_INSERT_FRONT, 0);
413         return flush_rq;
414 }
415
416 static void blk_start_post_flush(request_queue_t *q, struct request *rq)
417 {
418         struct request *flush_rq = q->flush_rq;
419
420         BUG_ON(!blk_barrier_rq(rq));
421
422         rq_init(q, flush_rq);
423         flush_rq->elevator_private = NULL;
424         flush_rq->flags = REQ_BAR_FLUSH;
425         flush_rq->rq_disk = rq->rq_disk;
426         flush_rq->rl = NULL;
427
428         if (q->prepare_flush_fn(q, flush_rq)) {
429                 flush_rq->end_io_data = rq;
430                 flush_rq->end_io = blk_post_flush_end_io;
431
432                 __elv_add_request(q, flush_rq, ELEVATOR_INSERT_FRONT, 0);
433                 q->request_fn(q);
434         }
435 }
436
437 static inline int blk_check_end_barrier(request_queue_t *q, struct request *rq,
438                                         int sectors)
439 {
440         if (sectors > rq->nr_sectors)
441                 sectors = rq->nr_sectors;
442
443         rq->nr_sectors -= sectors;
444         return rq->nr_sectors;
445 }
446
447 static int __blk_complete_barrier_rq(request_queue_t *q, struct request *rq,
448                                      int sectors, int queue_locked)
449 {
450         if (q->ordered != QUEUE_ORDERED_FLUSH)
451                 return 0;
452         if (!blk_fs_request(rq) || !blk_barrier_rq(rq))
453                 return 0;
454         if (blk_barrier_postflush(rq))
455                 return 0;
456
457         if (!blk_check_end_barrier(q, rq, sectors)) {
458                 unsigned long flags = 0;
459
460                 if (!queue_locked)
461                         spin_lock_irqsave(q->queue_lock, flags);
462
463                 blk_start_post_flush(q, rq);
464
465                 if (!queue_locked)
466                         spin_unlock_irqrestore(q->queue_lock, flags);
467         }
468
469         return 1;
470 }
471
472 /**
473  * blk_complete_barrier_rq - complete possible barrier request
474  * @q:  the request queue for the device
475  * @rq:  the request
476  * @sectors:  number of sectors to complete
477  *
478  * Description:
479  *   Used in driver end_io handling to determine whether to postpone
480  *   completion of a barrier request until a post flush has been done. This
481  *   is the unlocked variant, used if the caller doesn't already hold the
482  *   queue lock.
483  **/
484 int blk_complete_barrier_rq(request_queue_t *q, struct request *rq, int sectors)
485 {
486         return __blk_complete_barrier_rq(q, rq, sectors, 0);
487 }
488 EXPORT_SYMBOL(blk_complete_barrier_rq);
489
490 /**
491  * blk_complete_barrier_rq_locked - complete possible barrier request
492  * @q:  the request queue for the device
493  * @rq:  the request
494  * @sectors:  number of sectors to complete
495  *
496  * Description:
497  *   See blk_complete_barrier_rq(). This variant must be used if the caller
498  *   holds the queue lock.
499  **/
500 int blk_complete_barrier_rq_locked(request_queue_t *q, struct request *rq,
501                                    int sectors)
502 {
503         return __blk_complete_barrier_rq(q, rq, sectors, 1);
504 }
505 EXPORT_SYMBOL(blk_complete_barrier_rq_locked);
506
507 /**
508  * blk_queue_bounce_limit - set bounce buffer limit for queue
509  * @q:  the request queue for the device
510  * @dma_addr:   bus address limit
511  *
512  * Description:
513  *    Different hardware can have different requirements as to what pages
514  *    it can do I/O directly to. A low level driver can call
515  *    blk_queue_bounce_limit to have lower memory pages allocated as bounce
516  *    buffers for doing I/O to pages residing above @page. By default
517  *    the block layer sets this to the highest numbered "low" memory page.
518  **/
519 void blk_queue_bounce_limit(request_queue_t *q, u64 dma_addr)
520 {
521         unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
522
523         /*
524          * set appropriate bounce gfp mask -- unfortunately we don't have a
525          * full 4GB zone, so we have to resort to low memory for any bounces.
526          * ISA has its own < 16MB zone.
527          */
528         if (bounce_pfn < blk_max_low_pfn) {
529                 BUG_ON(dma_addr < BLK_BOUNCE_ISA);
530                 init_emergency_isa_pool();
531                 q->bounce_gfp = GFP_NOIO | GFP_DMA;
532         } else
533                 q->bounce_gfp = GFP_NOIO;
534
535         q->bounce_pfn = bounce_pfn;
536 }
537
538 EXPORT_SYMBOL(blk_queue_bounce_limit);
539
540 /**
541  * blk_queue_max_sectors - set max sectors for a request for this queue
542  * @q:  the request queue for the device
543  * @max_sectors:  max sectors in the usual 512b unit
544  *
545  * Description:
546  *    Enables a low level driver to set an upper limit on the size of
547  *    received requests.
548  **/
549 void blk_queue_max_sectors(request_queue_t *q, unsigned short max_sectors)
550 {
551         if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
552                 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
553                 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
554         }
555
556         q->max_sectors = q->max_hw_sectors = max_sectors;
557 }
558
559 EXPORT_SYMBOL(blk_queue_max_sectors);
560
561 /**
562  * blk_queue_max_phys_segments - set max phys segments for a request for this queue
563  * @q:  the request queue for the device
564  * @max_segments:  max number of segments
565  *
566  * Description:
567  *    Enables a low level driver to set an upper limit on the number of
568  *    physical data segments in a request.  This would be the largest sized
569  *    scatter list the driver could handle.
570  **/
571 void blk_queue_max_phys_segments(request_queue_t *q, unsigned short max_segments)
572 {
573         if (!max_segments) {
574                 max_segments = 1;
575                 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
576         }
577
578         q->max_phys_segments = max_segments;
579 }
580
581 EXPORT_SYMBOL(blk_queue_max_phys_segments);
582
583 /**
584  * blk_queue_max_hw_segments - set max hw segments for a request for this queue
585  * @q:  the request queue for the device
586  * @max_segments:  max number of segments
587  *
588  * Description:
589  *    Enables a low level driver to set an upper limit on the number of
590  *    hw data segments in a request.  This would be the largest number of
591  *    address/length pairs the host adapter can actually give as once
592  *    to the device.
593  **/
594 void blk_queue_max_hw_segments(request_queue_t *q, unsigned short max_segments)
595 {
596         if (!max_segments) {
597                 max_segments = 1;
598                 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
599         }
600
601         q->max_hw_segments = max_segments;
602 }
603
604 EXPORT_SYMBOL(blk_queue_max_hw_segments);
605
606 /**
607  * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
608  * @q:  the request queue for the device
609  * @max_size:  max size of segment in bytes
610  *
611  * Description:
612  *    Enables a low level driver to set an upper limit on the size of a
613  *    coalesced segment
614  **/
615 void blk_queue_max_segment_size(request_queue_t *q, unsigned int max_size)
616 {
617         if (max_size < PAGE_CACHE_SIZE) {
618                 max_size = PAGE_CACHE_SIZE;
619                 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
620         }
621
622         q->max_segment_size = max_size;
623 }
624
625 EXPORT_SYMBOL(blk_queue_max_segment_size);
626
627 /**
628  * blk_queue_hardsect_size - set hardware sector size for the queue
629  * @q:  the request queue for the device
630  * @size:  the hardware sector size, in bytes
631  *
632  * Description:
633  *   This should typically be set to the lowest possible sector size
634  *   that the hardware can operate on (possible without reverting to
635  *   even internal read-modify-write operations). Usually the default
636  *   of 512 covers most hardware.
637  **/
638 void blk_queue_hardsect_size(request_queue_t *q, unsigned short size)
639 {
640         q->hardsect_size = size;
641 }
642
643 EXPORT_SYMBOL(blk_queue_hardsect_size);
644
645 /*
646  * Returns the minimum that is _not_ zero, unless both are zero.
647  */
648 #define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
649
650 /**
651  * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
652  * @t:  the stacking driver (top)
653  * @b:  the underlying device (bottom)
654  **/
655 void blk_queue_stack_limits(request_queue_t *t, request_queue_t *b)
656 {
657         /* zero is "infinity" */
658         t->max_sectors = t->max_hw_sectors =
659                 min_not_zero(t->max_sectors,b->max_sectors);
660
661         t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
662         t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
663         t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
664         t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
665 }
666
667 EXPORT_SYMBOL(blk_queue_stack_limits);
668
669 /**
670  * blk_queue_segment_boundary - set boundary rules for segment merging
671  * @q:  the request queue for the device
672  * @mask:  the memory boundary mask
673  **/
674 void blk_queue_segment_boundary(request_queue_t *q, unsigned long mask)
675 {
676         if (mask < PAGE_CACHE_SIZE - 1) {
677                 mask = PAGE_CACHE_SIZE - 1;
678                 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
679         }
680
681         q->seg_boundary_mask = mask;
682 }
683
684 EXPORT_SYMBOL(blk_queue_segment_boundary);
685
686 /**
687  * blk_queue_dma_alignment - set dma length and memory alignment
688  * @q:     the request queue for the device
689  * @mask:  alignment mask
690  *
691  * description:
692  *    set required memory and length aligment for direct dma transactions.
693  *    this is used when buiding direct io requests for the queue.
694  *
695  **/
696 void blk_queue_dma_alignment(request_queue_t *q, int mask)
697 {
698         q->dma_alignment = mask;
699 }
700
701 EXPORT_SYMBOL(blk_queue_dma_alignment);
702
703 /**
704  * blk_queue_find_tag - find a request by its tag and queue
705  *
706  * @q:   The request queue for the device
707  * @tag: The tag of the request
708  *
709  * Notes:
710  *    Should be used when a device returns a tag and you want to match
711  *    it with a request.
712  *
713  *    no locks need be held.
714  **/
715 struct request *blk_queue_find_tag(request_queue_t *q, int tag)
716 {
717         struct blk_queue_tag *bqt = q->queue_tags;
718
719         if (unlikely(bqt == NULL || tag >= bqt->real_max_depth))
720                 return NULL;
721
722         return bqt->tag_index[tag];
723 }
724
725 EXPORT_SYMBOL(blk_queue_find_tag);
726
727 /**
728  * __blk_queue_free_tags - release tag maintenance info
729  * @q:  the request queue for the device
730  *
731  *  Notes:
732  *    blk_cleanup_queue() will take care of calling this function, if tagging
733  *    has been used. So there's no need to call this directly.
734  **/
735 static void __blk_queue_free_tags(request_queue_t *q)
736 {
737         struct blk_queue_tag *bqt = q->queue_tags;
738
739         if (!bqt)
740                 return;
741
742         if (atomic_dec_and_test(&bqt->refcnt)) {
743                 BUG_ON(bqt->busy);
744                 BUG_ON(!list_empty(&bqt->busy_list));
745
746                 kfree(bqt->tag_index);
747                 bqt->tag_index = NULL;
748
749                 kfree(bqt->tag_map);
750                 bqt->tag_map = NULL;
751
752                 kfree(bqt);
753         }
754
755         q->queue_tags = NULL;
756         q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
757 }
758
759 /**
760  * blk_queue_free_tags - release tag maintenance info
761  * @q:  the request queue for the device
762  *
763  *  Notes:
764  *      This is used to disabled tagged queuing to a device, yet leave
765  *      queue in function.
766  **/
767 void blk_queue_free_tags(request_queue_t *q)
768 {
769         clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
770 }
771
772 EXPORT_SYMBOL(blk_queue_free_tags);
773
774 static int
775 init_tag_map(request_queue_t *q, struct blk_queue_tag *tags, int depth)
776 {
777         int bits, i;
778         struct request **tag_index;
779         unsigned long *tag_map;
780
781         if (depth > q->nr_requests * 2) {
782                 depth = q->nr_requests * 2;
783                 printk(KERN_ERR "%s: adjusted depth to %d\n",
784                                 __FUNCTION__, depth);
785         }
786
787         tag_index = kmalloc(depth * sizeof(struct request *), GFP_ATOMIC);
788         if (!tag_index)
789                 goto fail;
790
791         bits = (depth / BLK_TAGS_PER_LONG) + 1;
792         tag_map = kmalloc(bits * sizeof(unsigned long), GFP_ATOMIC);
793         if (!tag_map)
794                 goto fail;
795
796         memset(tag_index, 0, depth * sizeof(struct request *));
797         memset(tag_map, 0, bits * sizeof(unsigned long));
798         tags->max_depth = depth;
799         tags->real_max_depth = bits * BITS_PER_LONG;
800         tags->tag_index = tag_index;
801         tags->tag_map = tag_map;
802
803         /*
804          * set the upper bits if the depth isn't a multiple of the word size
805          */
806         for (i = depth; i < bits * BLK_TAGS_PER_LONG; i++)
807                 __set_bit(i, tag_map);
808
809         return 0;
810 fail:
811         kfree(tag_index);
812         return -ENOMEM;
813 }
814
815 /**
816  * blk_queue_init_tags - initialize the queue tag info
817  * @q:  the request queue for the device
818  * @depth:  the maximum queue depth supported
819  * @tags: the tag to use
820  **/
821 int blk_queue_init_tags(request_queue_t *q, int depth,
822                         struct blk_queue_tag *tags)
823 {
824         int rc;
825
826         BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
827
828         if (!tags && !q->queue_tags) {
829                 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
830                 if (!tags)
831                         goto fail;
832
833                 if (init_tag_map(q, tags, depth))
834                         goto fail;
835
836                 INIT_LIST_HEAD(&tags->busy_list);
837                 tags->busy = 0;
838                 atomic_set(&tags->refcnt, 1);
839         } else if (q->queue_tags) {
840                 if ((rc = blk_queue_resize_tags(q, depth)))
841                         return rc;
842                 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
843                 return 0;
844         } else
845                 atomic_inc(&tags->refcnt);
846
847         /*
848          * assign it, all done
849          */
850         q->queue_tags = tags;
851         q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
852         return 0;
853 fail:
854         kfree(tags);
855         return -ENOMEM;
856 }
857
858 EXPORT_SYMBOL(blk_queue_init_tags);
859
860 /**
861  * blk_queue_resize_tags - change the queueing depth
862  * @q:  the request queue for the device
863  * @new_depth: the new max command queueing depth
864  *
865  *  Notes:
866  *    Must be called with the queue lock held.
867  **/
868 int blk_queue_resize_tags(request_queue_t *q, int new_depth)
869 {
870         struct blk_queue_tag *bqt = q->queue_tags;
871         struct request **tag_index;
872         unsigned long *tag_map;
873         int bits, max_depth;
874
875         if (!bqt)
876                 return -ENXIO;
877
878         /*
879          * don't bother sizing down
880          */
881         if (new_depth <= bqt->real_max_depth) {
882                 bqt->max_depth = new_depth;
883                 return 0;
884         }
885
886         /*
887          * save the old state info, so we can copy it back
888          */
889         tag_index = bqt->tag_index;
890         tag_map = bqt->tag_map;
891         max_depth = bqt->real_max_depth;
892
893         if (init_tag_map(q, bqt, new_depth))
894                 return -ENOMEM;
895
896         memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
897         bits = max_depth / BLK_TAGS_PER_LONG;
898         memcpy(bqt->tag_map, tag_map, bits * sizeof(unsigned long));
899
900         kfree(tag_index);
901         kfree(tag_map);
902         return 0;
903 }
904
905 EXPORT_SYMBOL(blk_queue_resize_tags);
906
907 /**
908  * blk_queue_end_tag - end tag operations for a request
909  * @q:  the request queue for the device
910  * @rq: the request that has completed
911  *
912  *  Description:
913  *    Typically called when end_that_request_first() returns 0, meaning
914  *    all transfers have been done for a request. It's important to call
915  *    this function before end_that_request_last(), as that will put the
916  *    request back on the free list thus corrupting the internal tag list.
917  *
918  *  Notes:
919  *   queue lock must be held.
920  **/
921 void blk_queue_end_tag(request_queue_t *q, struct request *rq)
922 {
923         struct blk_queue_tag *bqt = q->queue_tags;
924         int tag = rq->tag;
925
926         BUG_ON(tag == -1);
927
928         if (unlikely(tag >= bqt->real_max_depth))
929                 return;
930
931         if (unlikely(!__test_and_clear_bit(tag, bqt->tag_map))) {
932                 printk("attempt to clear non-busy tag (%d)\n", tag);
933                 return;
934         }
935
936         list_del_init(&rq->queuelist);
937         rq->flags &= ~REQ_QUEUED;
938         rq->tag = -1;
939
940         if (unlikely(bqt->tag_index[tag] == NULL))
941                 printk("tag %d is missing\n", tag);
942
943         bqt->tag_index[tag] = NULL;
944         bqt->busy--;
945 }
946
947 EXPORT_SYMBOL(blk_queue_end_tag);
948
949 /**
950  * blk_queue_start_tag - find a free tag and assign it
951  * @q:  the request queue for the device
952  * @rq:  the block request that needs tagging
953  *
954  *  Description:
955  *    This can either be used as a stand-alone helper, or possibly be
956  *    assigned as the queue &prep_rq_fn (in which case &struct request
957  *    automagically gets a tag assigned). Note that this function
958  *    assumes that any type of request can be queued! if this is not
959  *    true for your device, you must check the request type before
960  *    calling this function.  The request will also be removed from
961  *    the request queue, so it's the drivers responsibility to readd
962  *    it if it should need to be restarted for some reason.
963  *
964  *  Notes:
965  *   queue lock must be held.
966  **/
967 int blk_queue_start_tag(request_queue_t *q, struct request *rq)
968 {
969         struct blk_queue_tag *bqt = q->queue_tags;
970         unsigned long *map = bqt->tag_map;
971         int tag = 0;
972
973         if (unlikely((rq->flags & REQ_QUEUED))) {
974                 printk(KERN_ERR 
975                        "request %p for device [%s] already tagged %d",
976                        rq, rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
977                 BUG();
978         }
979
980         for (map = bqt->tag_map; *map == -1UL; map++) {
981                 tag += BLK_TAGS_PER_LONG;
982
983                 if (tag >= bqt->max_depth)
984                         return 1;
985         }
986
987         tag += ffz(*map);
988         __set_bit(tag, bqt->tag_map);
989
990         rq->flags |= REQ_QUEUED;
991         rq->tag = tag;
992         bqt->tag_index[tag] = rq;
993         blkdev_dequeue_request(rq);
994         list_add(&rq->queuelist, &bqt->busy_list);
995         bqt->busy++;
996         return 0;
997 }
998
999 EXPORT_SYMBOL(blk_queue_start_tag);
1000
1001 /**
1002  * blk_queue_invalidate_tags - invalidate all pending tags
1003  * @q:  the request queue for the device
1004  *
1005  *  Description:
1006  *   Hardware conditions may dictate a need to stop all pending requests.
1007  *   In this case, we will safely clear the block side of the tag queue and
1008  *   readd all requests to the request queue in the right order.
1009  *
1010  *  Notes:
1011  *   queue lock must be held.
1012  **/
1013 void blk_queue_invalidate_tags(request_queue_t *q)
1014 {
1015         struct blk_queue_tag *bqt = q->queue_tags;
1016         struct list_head *tmp, *n;
1017         struct request *rq;
1018
1019         list_for_each_safe(tmp, n, &bqt->busy_list) {
1020                 rq = list_entry_rq(tmp);
1021
1022                 if (rq->tag == -1) {
1023                         printk("bad tag found on list\n");
1024                         list_del_init(&rq->queuelist);
1025                         rq->flags &= ~REQ_QUEUED;
1026                 } else
1027                         blk_queue_end_tag(q, rq);
1028
1029                 rq->flags &= ~REQ_STARTED;
1030                 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1031         }
1032 }
1033
1034 EXPORT_SYMBOL(blk_queue_invalidate_tags);
1035
1036 static char *rq_flags[] = {
1037         "REQ_RW",
1038         "REQ_FAILFAST",
1039         "REQ_SOFTBARRIER",
1040         "REQ_HARDBARRIER",
1041         "REQ_CMD",
1042         "REQ_NOMERGE",
1043         "REQ_STARTED",
1044         "REQ_DONTPREP",
1045         "REQ_QUEUED",
1046         "REQ_PC",
1047         "REQ_BLOCK_PC",
1048         "REQ_SENSE",
1049         "REQ_FAILED",
1050         "REQ_QUIET",
1051         "REQ_SPECIAL",
1052         "REQ_DRIVE_CMD",
1053         "REQ_DRIVE_TASK",
1054         "REQ_DRIVE_TASKFILE",
1055         "REQ_PREEMPT",
1056         "REQ_PM_SUSPEND",
1057         "REQ_PM_RESUME",
1058         "REQ_PM_SHUTDOWN",
1059 };
1060
1061 void blk_dump_rq_flags(struct request *rq, char *msg)
1062 {
1063         int bit;
1064
1065         printk("%s: dev %s: flags = ", msg,
1066                 rq->rq_disk ? rq->rq_disk->disk_name : "?");
1067         bit = 0;
1068         do {
1069                 if (rq->flags & (1 << bit))
1070                         printk("%s ", rq_flags[bit]);
1071                 bit++;
1072         } while (bit < __REQ_NR_BITS);
1073
1074         printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1075                                                        rq->nr_sectors,
1076                                                        rq->current_nr_sectors);
1077         printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1078
1079         if (rq->flags & (REQ_BLOCK_PC | REQ_PC)) {
1080                 printk("cdb: ");
1081                 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1082                         printk("%02x ", rq->cmd[bit]);
1083                 printk("\n");
1084         }
1085 }
1086
1087 EXPORT_SYMBOL(blk_dump_rq_flags);
1088
1089 void blk_recount_segments(request_queue_t *q, struct bio *bio)
1090 {
1091         struct bio_vec *bv, *bvprv = NULL;
1092         int i, nr_phys_segs, nr_hw_segs, seg_size, hw_seg_size, cluster;
1093         int high, highprv = 1;
1094
1095         if (unlikely(!bio->bi_io_vec))
1096                 return;
1097
1098         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1099         hw_seg_size = seg_size = nr_phys_segs = nr_hw_segs = 0;
1100         bio_for_each_segment(bv, bio, i) {
1101                 /*
1102                  * the trick here is making sure that a high page is never
1103                  * considered part of another segment, since that might
1104                  * change with the bounce page.
1105                  */
1106                 high = page_to_pfn(bv->bv_page) >= q->bounce_pfn;
1107                 if (high || highprv)
1108                         goto new_hw_segment;
1109                 if (cluster) {
1110                         if (seg_size + bv->bv_len > q->max_segment_size)
1111                                 goto new_segment;
1112                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1113                                 goto new_segment;
1114                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1115                                 goto new_segment;
1116                         if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1117                                 goto new_hw_segment;
1118
1119                         seg_size += bv->bv_len;
1120                         hw_seg_size += bv->bv_len;
1121                         bvprv = bv;
1122                         continue;
1123                 }
1124 new_segment:
1125                 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1126                     !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len)) {
1127                         hw_seg_size += bv->bv_len;
1128                 } else {
1129 new_hw_segment:
1130                         if (hw_seg_size > bio->bi_hw_front_size)
1131                                 bio->bi_hw_front_size = hw_seg_size;
1132                         hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1133                         nr_hw_segs++;
1134                 }
1135
1136                 nr_phys_segs++;
1137                 bvprv = bv;
1138                 seg_size = bv->bv_len;
1139                 highprv = high;
1140         }
1141         if (hw_seg_size > bio->bi_hw_back_size)
1142                 bio->bi_hw_back_size = hw_seg_size;
1143         if (nr_hw_segs == 1 && hw_seg_size > bio->bi_hw_front_size)
1144                 bio->bi_hw_front_size = hw_seg_size;
1145         bio->bi_phys_segments = nr_phys_segs;
1146         bio->bi_hw_segments = nr_hw_segs;
1147         bio->bi_flags |= (1 << BIO_SEG_VALID);
1148 }
1149
1150
1151 int blk_phys_contig_segment(request_queue_t *q, struct bio *bio,
1152                                    struct bio *nxt)
1153 {
1154         if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1155                 return 0;
1156
1157         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1158                 return 0;
1159         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1160                 return 0;
1161
1162         /*
1163          * bio and nxt are contigous in memory, check if the queue allows
1164          * these two to be merged into one
1165          */
1166         if (BIO_SEG_BOUNDARY(q, bio, nxt))
1167                 return 1;
1168
1169         return 0;
1170 }
1171
1172 EXPORT_SYMBOL(blk_phys_contig_segment);
1173
1174 int blk_hw_contig_segment(request_queue_t *q, struct bio *bio,
1175                                  struct bio *nxt)
1176 {
1177         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1178                 blk_recount_segments(q, bio);
1179         if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1180                 blk_recount_segments(q, nxt);
1181         if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1182             BIOVEC_VIRT_OVERSIZE(bio->bi_hw_front_size + bio->bi_hw_back_size))
1183                 return 0;
1184         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1185                 return 0;
1186
1187         return 1;
1188 }
1189
1190 EXPORT_SYMBOL(blk_hw_contig_segment);
1191
1192 /*
1193  * map a request to scatterlist, return number of sg entries setup. Caller
1194  * must make sure sg can hold rq->nr_phys_segments entries
1195  */
1196 int blk_rq_map_sg(request_queue_t *q, struct request *rq, struct scatterlist *sg)
1197 {
1198         struct bio_vec *bvec, *bvprv;
1199         struct bio *bio;
1200         int nsegs, i, cluster;
1201
1202         nsegs = 0;
1203         cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1204
1205         /*
1206          * for each bio in rq
1207          */
1208         bvprv = NULL;
1209         rq_for_each_bio(bio, rq) {
1210                 /*
1211                  * for each segment in bio
1212                  */
1213                 bio_for_each_segment(bvec, bio, i) {
1214                         int nbytes = bvec->bv_len;
1215
1216                         if (bvprv && cluster) {
1217                                 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1218                                         goto new_segment;
1219
1220                                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1221                                         goto new_segment;
1222                                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1223                                         goto new_segment;
1224
1225                                 sg[nsegs - 1].length += nbytes;
1226                         } else {
1227 new_segment:
1228                                 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1229                                 sg[nsegs].page = bvec->bv_page;
1230                                 sg[nsegs].length = nbytes;
1231                                 sg[nsegs].offset = bvec->bv_offset;
1232
1233                                 nsegs++;
1234                         }
1235                         bvprv = bvec;
1236                 } /* segments in bio */
1237         } /* bios in rq */
1238
1239         return nsegs;
1240 }
1241
1242 EXPORT_SYMBOL(blk_rq_map_sg);
1243
1244 /*
1245  * the standard queue merge functions, can be overridden with device
1246  * specific ones if so desired
1247  */
1248
1249 static inline int ll_new_mergeable(request_queue_t *q,
1250                                    struct request *req,
1251                                    struct bio *bio)
1252 {
1253         int nr_phys_segs = bio_phys_segments(q, bio);
1254
1255         if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1256                 req->flags |= REQ_NOMERGE;
1257                 if (req == q->last_merge)
1258                         q->last_merge = NULL;
1259                 return 0;
1260         }
1261
1262         /*
1263          * A hw segment is just getting larger, bump just the phys
1264          * counter.
1265          */
1266         req->nr_phys_segments += nr_phys_segs;
1267         return 1;
1268 }
1269
1270 static inline int ll_new_hw_segment(request_queue_t *q,
1271                                     struct request *req,
1272                                     struct bio *bio)
1273 {
1274         int nr_hw_segs = bio_hw_segments(q, bio);
1275         int nr_phys_segs = bio_phys_segments(q, bio);
1276
1277         if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1278             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1279                 req->flags |= REQ_NOMERGE;
1280                 if (req == q->last_merge)
1281                         q->last_merge = NULL;
1282                 return 0;
1283         }
1284
1285         /*
1286          * This will form the start of a new hw segment.  Bump both
1287          * counters.
1288          */
1289         req->nr_hw_segments += nr_hw_segs;
1290         req->nr_phys_segments += nr_phys_segs;
1291         return 1;
1292 }
1293
1294 static int ll_back_merge_fn(request_queue_t *q, struct request *req, 
1295                             struct bio *bio)
1296 {
1297         int len;
1298
1299         if (req->nr_sectors + bio_sectors(bio) > q->max_sectors) {
1300                 req->flags |= REQ_NOMERGE;
1301                 if (req == q->last_merge)
1302                         q->last_merge = NULL;
1303                 return 0;
1304         }
1305         if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1306                 blk_recount_segments(q, req->biotail);
1307         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1308                 blk_recount_segments(q, bio);
1309         len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1310         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1311             !BIOVEC_VIRT_OVERSIZE(len)) {
1312                 int mergeable =  ll_new_mergeable(q, req, bio);
1313
1314                 if (mergeable) {
1315                         if (req->nr_hw_segments == 1)
1316                                 req->bio->bi_hw_front_size = len;
1317                         if (bio->bi_hw_segments == 1)
1318                                 bio->bi_hw_back_size = len;
1319                 }
1320                 return mergeable;
1321         }
1322
1323         return ll_new_hw_segment(q, req, bio);
1324 }
1325
1326 static int ll_front_merge_fn(request_queue_t *q, struct request *req, 
1327                              struct bio *bio)
1328 {
1329         int len;
1330
1331         if (req->nr_sectors + bio_sectors(bio) > q->max_sectors) {
1332                 req->flags |= REQ_NOMERGE;
1333                 if (req == q->last_merge)
1334                         q->last_merge = NULL;
1335                 return 0;
1336         }
1337         len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1338         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1339                 blk_recount_segments(q, bio);
1340         if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1341                 blk_recount_segments(q, req->bio);
1342         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1343             !BIOVEC_VIRT_OVERSIZE(len)) {
1344                 int mergeable =  ll_new_mergeable(q, req, bio);
1345
1346                 if (mergeable) {
1347                         if (bio->bi_hw_segments == 1)
1348                                 bio->bi_hw_front_size = len;
1349                         if (req->nr_hw_segments == 1)
1350                                 req->biotail->bi_hw_back_size = len;
1351                 }
1352                 return mergeable;
1353         }
1354
1355         return ll_new_hw_segment(q, req, bio);
1356 }
1357
1358 static int ll_merge_requests_fn(request_queue_t *q, struct request *req,
1359                                 struct request *next)
1360 {
1361         int total_phys_segments = req->nr_phys_segments +next->nr_phys_segments;
1362         int total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1363
1364         /*
1365          * First check if the either of the requests are re-queued
1366          * requests.  Can't merge them if they are.
1367          */
1368         if (req->special || next->special)
1369                 return 0;
1370
1371         /*
1372          * Will it become to large?
1373          */
1374         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1375                 return 0;
1376
1377         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1378         if (blk_phys_contig_segment(q, req->biotail, next->bio))
1379                 total_phys_segments--;
1380
1381         if (total_phys_segments > q->max_phys_segments)
1382                 return 0;
1383
1384         total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1385         if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1386                 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1387                 /*
1388                  * propagate the combined length to the end of the requests
1389                  */
1390                 if (req->nr_hw_segments == 1)
1391                         req->bio->bi_hw_front_size = len;
1392                 if (next->nr_hw_segments == 1)
1393                         next->biotail->bi_hw_back_size = len;
1394                 total_hw_segments--;
1395         }
1396
1397         if (total_hw_segments > q->max_hw_segments)
1398                 return 0;
1399
1400         /* Merge is OK... */
1401         req->nr_phys_segments = total_phys_segments;
1402         req->nr_hw_segments = total_hw_segments;
1403         return 1;
1404 }
1405
1406 /*
1407  * "plug" the device if there are no outstanding requests: this will
1408  * force the transfer to start only after we have put all the requests
1409  * on the list.
1410  *
1411  * This is called with interrupts off and no requests on the queue and
1412  * with the queue lock held.
1413  */
1414 void blk_plug_device(request_queue_t *q)
1415 {
1416         WARN_ON(!irqs_disabled());
1417
1418         /*
1419          * don't plug a stopped queue, it must be paired with blk_start_queue()
1420          * which will restart the queueing
1421          */
1422         if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
1423                 return;
1424
1425         if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1426                 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1427 }
1428
1429 EXPORT_SYMBOL(blk_plug_device);
1430
1431 /*
1432  * remove the queue from the plugged list, if present. called with
1433  * queue lock held and interrupts disabled.
1434  */
1435 int blk_remove_plug(request_queue_t *q)
1436 {
1437         WARN_ON(!irqs_disabled());
1438
1439         if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1440                 return 0;
1441
1442         del_timer(&q->unplug_timer);
1443         return 1;
1444 }
1445
1446 EXPORT_SYMBOL(blk_remove_plug);
1447
1448 /*
1449  * remove the plug and let it rip..
1450  */
1451 void __generic_unplug_device(request_queue_t *q)
1452 {
1453         if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
1454                 return;
1455
1456         if (!blk_remove_plug(q))
1457                 return;
1458
1459         /*
1460          * was plugged, fire request_fn if queue has stuff to do
1461          */
1462         if (elv_next_request(q))
1463                 q->request_fn(q);
1464 }
1465 EXPORT_SYMBOL(__generic_unplug_device);
1466
1467 /**
1468  * generic_unplug_device - fire a request queue
1469  * @q:    The &request_queue_t in question
1470  *
1471  * Description:
1472  *   Linux uses plugging to build bigger requests queues before letting
1473  *   the device have at them. If a queue is plugged, the I/O scheduler
1474  *   is still adding and merging requests on the queue. Once the queue
1475  *   gets unplugged, the request_fn defined for the queue is invoked and
1476  *   transfers started.
1477  **/
1478 void generic_unplug_device(request_queue_t *q)
1479 {
1480         spin_lock_irq(q->queue_lock);
1481         __generic_unplug_device(q);
1482         spin_unlock_irq(q->queue_lock);
1483 }
1484 EXPORT_SYMBOL(generic_unplug_device);
1485
1486 static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1487                                    struct page *page)
1488 {
1489         request_queue_t *q = bdi->unplug_io_data;
1490
1491         /*
1492          * devices don't necessarily have an ->unplug_fn defined
1493          */
1494         if (q->unplug_fn)
1495                 q->unplug_fn(q);
1496 }
1497
1498 static void blk_unplug_work(void *data)
1499 {
1500         request_queue_t *q = data;
1501
1502         q->unplug_fn(q);
1503 }
1504
1505 static void blk_unplug_timeout(unsigned long data)
1506 {
1507         request_queue_t *q = (request_queue_t *)data;
1508
1509         kblockd_schedule_work(&q->unplug_work);
1510 }
1511
1512 /**
1513  * blk_start_queue - restart a previously stopped queue
1514  * @q:    The &request_queue_t in question
1515  *
1516  * Description:
1517  *   blk_start_queue() will clear the stop flag on the queue, and call
1518  *   the request_fn for the queue if it was in a stopped state when
1519  *   entered. Also see blk_stop_queue(). Queue lock must be held.
1520  **/
1521 void blk_start_queue(request_queue_t *q)
1522 {
1523         clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1524
1525         /*
1526          * one level of recursion is ok and is much faster than kicking
1527          * the unplug handling
1528          */
1529         if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1530                 q->request_fn(q);
1531                 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1532         } else {
1533                 blk_plug_device(q);
1534                 kblockd_schedule_work(&q->unplug_work);
1535         }
1536 }
1537
1538 EXPORT_SYMBOL(blk_start_queue);
1539
1540 /**
1541  * blk_stop_queue - stop a queue
1542  * @q:    The &request_queue_t in question
1543  *
1544  * Description:
1545  *   The Linux block layer assumes that a block driver will consume all
1546  *   entries on the request queue when the request_fn strategy is called.
1547  *   Often this will not happen, because of hardware limitations (queue
1548  *   depth settings). If a device driver gets a 'queue full' response,
1549  *   or if it simply chooses not to queue more I/O at one point, it can
1550  *   call this function to prevent the request_fn from being called until
1551  *   the driver has signalled it's ready to go again. This happens by calling
1552  *   blk_start_queue() to restart queue operations. Queue lock must be held.
1553  **/
1554 void blk_stop_queue(request_queue_t *q)
1555 {
1556         blk_remove_plug(q);
1557         set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1558 }
1559 EXPORT_SYMBOL(blk_stop_queue);
1560
1561 /**
1562  * blk_sync_queue - cancel any pending callbacks on a queue
1563  * @q: the queue
1564  *
1565  * Description:
1566  *     The block layer may perform asynchronous callback activity
1567  *     on a queue, such as calling the unplug function after a timeout.
1568  *     A block device may call blk_sync_queue to ensure that any
1569  *     such activity is cancelled, thus allowing it to release resources
1570  *     the the callbacks might use. The caller must already have made sure
1571  *     that its ->make_request_fn will not re-add plugging prior to calling
1572  *     this function.
1573  *
1574  */
1575 void blk_sync_queue(struct request_queue *q)
1576 {
1577         del_timer_sync(&q->unplug_timer);
1578         kblockd_flush();
1579 }
1580 EXPORT_SYMBOL(blk_sync_queue);
1581
1582 /**
1583  * blk_run_queue - run a single device queue
1584  * @q:  The queue to run
1585  */
1586 void blk_run_queue(struct request_queue *q)
1587 {
1588         unsigned long flags;
1589
1590         spin_lock_irqsave(q->queue_lock, flags);
1591         blk_remove_plug(q);
1592         if (!elv_queue_empty(q))
1593                 q->request_fn(q);
1594         spin_unlock_irqrestore(q->queue_lock, flags);
1595 }
1596 EXPORT_SYMBOL(blk_run_queue);
1597
1598 /**
1599  * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1600  * @q:    the request queue to be released
1601  *
1602  * Description:
1603  *     blk_cleanup_queue is the pair to blk_init_queue() or
1604  *     blk_queue_make_request().  It should be called when a request queue is
1605  *     being released; typically when a block device is being de-registered.
1606  *     Currently, its primary task it to free all the &struct request
1607  *     structures that were allocated to the queue and the queue itself.
1608  *
1609  * Caveat:
1610  *     Hopefully the low level driver will have finished any
1611  *     outstanding requests first...
1612  **/
1613 void blk_cleanup_queue(request_queue_t * q)
1614 {
1615         struct request_list *rl = &q->rq;
1616
1617         if (!atomic_dec_and_test(&q->refcnt))
1618                 return;
1619
1620         if (q->elevator)
1621                 elevator_exit(q->elevator);
1622
1623         blk_sync_queue(q);
1624
1625         if (rl->rq_pool)
1626                 mempool_destroy(rl->rq_pool);
1627
1628         if (q->queue_tags)
1629                 __blk_queue_free_tags(q);
1630
1631         blk_queue_ordered(q, QUEUE_ORDERED_NONE);
1632
1633         kmem_cache_free(requestq_cachep, q);
1634 }
1635
1636 EXPORT_SYMBOL(blk_cleanup_queue);
1637
1638 static int blk_init_free_list(request_queue_t *q)
1639 {
1640         struct request_list *rl = &q->rq;
1641
1642         rl->count[READ] = rl->count[WRITE] = 0;
1643         rl->starved[READ] = rl->starved[WRITE] = 0;
1644         init_waitqueue_head(&rl->wait[READ]);
1645         init_waitqueue_head(&rl->wait[WRITE]);
1646         init_waitqueue_head(&rl->drain);
1647
1648         rl->rq_pool = mempool_create(BLKDEV_MIN_RQ, mempool_alloc_slab, mempool_free_slab, request_cachep);
1649
1650         if (!rl->rq_pool)
1651                 return -ENOMEM;
1652
1653         return 0;
1654 }
1655
1656 static int __make_request(request_queue_t *, struct bio *);
1657
1658 request_queue_t *blk_alloc_queue(int gfp_mask)
1659 {
1660         request_queue_t *q = kmem_cache_alloc(requestq_cachep, gfp_mask);
1661
1662         if (!q)
1663                 return NULL;
1664
1665         memset(q, 0, sizeof(*q));
1666         init_timer(&q->unplug_timer);
1667         atomic_set(&q->refcnt, 1);
1668
1669         q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1670         q->backing_dev_info.unplug_io_data = q;
1671
1672         return q;
1673 }
1674
1675 EXPORT_SYMBOL(blk_alloc_queue);
1676
1677 /**
1678  * blk_init_queue  - prepare a request queue for use with a block device
1679  * @rfn:  The function to be called to process requests that have been
1680  *        placed on the queue.
1681  * @lock: Request queue spin lock
1682  *
1683  * Description:
1684  *    If a block device wishes to use the standard request handling procedures,
1685  *    which sorts requests and coalesces adjacent requests, then it must
1686  *    call blk_init_queue().  The function @rfn will be called when there
1687  *    are requests on the queue that need to be processed.  If the device
1688  *    supports plugging, then @rfn may not be called immediately when requests
1689  *    are available on the queue, but may be called at some time later instead.
1690  *    Plugged queues are generally unplugged when a buffer belonging to one
1691  *    of the requests on the queue is needed, or due to memory pressure.
1692  *
1693  *    @rfn is not required, or even expected, to remove all requests off the
1694  *    queue, but only as many as it can handle at a time.  If it does leave
1695  *    requests on the queue, it is responsible for arranging that the requests
1696  *    get dealt with eventually.
1697  *
1698  *    The queue spin lock must be held while manipulating the requests on the
1699  *    request queue.
1700  *
1701  *    Function returns a pointer to the initialized request queue, or NULL if
1702  *    it didn't succeed.
1703  *
1704  * Note:
1705  *    blk_init_queue() must be paired with a blk_cleanup_queue() call
1706  *    when the block device is deactivated (such as at module unload).
1707  **/
1708 request_queue_t *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1709 {
1710         request_queue_t *q = blk_alloc_queue(GFP_KERNEL);
1711
1712         if (!q)
1713                 return NULL;
1714
1715         if (blk_init_free_list(q))
1716                 goto out_init;
1717
1718         /*
1719          * if caller didn't supply a lock, they get per-queue locking with
1720          * our embedded lock
1721          */
1722         if (!lock) {
1723                 spin_lock_init(&q->__queue_lock);
1724                 lock = &q->__queue_lock;
1725         }
1726
1727         q->request_fn           = rfn;
1728         q->back_merge_fn        = ll_back_merge_fn;
1729         q->front_merge_fn       = ll_front_merge_fn;
1730         q->merge_requests_fn    = ll_merge_requests_fn;
1731         q->prep_rq_fn           = NULL;
1732         q->unplug_fn            = generic_unplug_device;
1733         q->queue_flags          = (1 << QUEUE_FLAG_CLUSTER);
1734         q->queue_lock           = lock;
1735
1736         blk_queue_segment_boundary(q, 0xffffffff);
1737
1738         blk_queue_make_request(q, __make_request);
1739         blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1740
1741         blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1742         blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1743
1744         /*
1745          * all done
1746          */
1747         if (!elevator_init(q, NULL)) {
1748                 blk_queue_congestion_threshold(q);
1749                 return q;
1750         }
1751
1752         blk_cleanup_queue(q);
1753 out_init:
1754         kmem_cache_free(requestq_cachep, q);
1755         return NULL;
1756 }
1757
1758 EXPORT_SYMBOL(blk_init_queue);
1759
1760 int blk_get_queue(request_queue_t *q)
1761 {
1762         if (!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)) {
1763                 atomic_inc(&q->refcnt);
1764                 return 0;
1765         }
1766
1767         return 1;
1768 }
1769
1770 EXPORT_SYMBOL(blk_get_queue);
1771
1772 static inline void blk_free_request(request_queue_t *q, struct request *rq)
1773 {
1774         elv_put_request(q, rq);
1775         mempool_free(rq, q->rq.rq_pool);
1776 }
1777
1778 static inline struct request *blk_alloc_request(request_queue_t *q, int rw,
1779                                                 int gfp_mask)
1780 {
1781         struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1782
1783         if (!rq)
1784                 return NULL;
1785
1786         /*
1787          * first three bits are identical in rq->flags and bio->bi_rw,
1788          * see bio.h and blkdev.h
1789          */
1790         rq->flags = rw;
1791
1792         if (!elv_set_request(q, rq, gfp_mask))
1793                 return rq;
1794
1795         mempool_free(rq, q->rq.rq_pool);
1796         return NULL;
1797 }
1798
1799 /*
1800  * ioc_batching returns true if the ioc is a valid batching request and
1801  * should be given priority access to a request.
1802  */
1803 static inline int ioc_batching(request_queue_t *q, struct io_context *ioc)
1804 {
1805         if (!ioc)
1806                 return 0;
1807
1808         /*
1809          * Make sure the process is able to allocate at least 1 request
1810          * even if the batch times out, otherwise we could theoretically
1811          * lose wakeups.
1812          */
1813         return ioc->nr_batch_requests == q->nr_batching ||
1814                 (ioc->nr_batch_requests > 0
1815                 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
1816 }
1817
1818 /*
1819  * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
1820  * will cause the process to be a "batcher" on all queues in the system. This
1821  * is the behaviour we want though - once it gets a wakeup it should be given
1822  * a nice run.
1823  */
1824 void ioc_set_batching(request_queue_t *q, struct io_context *ioc)
1825 {
1826         if (!ioc || ioc_batching(q, ioc))
1827                 return;
1828
1829         ioc->nr_batch_requests = q->nr_batching;
1830         ioc->last_waited = jiffies;
1831 }
1832
1833 static void __freed_request(request_queue_t *q, int rw)
1834 {
1835         struct request_list *rl = &q->rq;
1836
1837         if (rl->count[rw] < queue_congestion_off_threshold(q))
1838                 clear_queue_congested(q, rw);
1839
1840         if (rl->count[rw] + 1 <= q->nr_requests) {
1841                 smp_mb();
1842                 if (waitqueue_active(&rl->wait[rw]))
1843                         wake_up(&rl->wait[rw]);
1844
1845                 blk_clear_queue_full(q, rw);
1846         }
1847 }
1848
1849 /*
1850  * A request has just been released.  Account for it, update the full and
1851  * congestion status, wake up any waiters.   Called under q->queue_lock.
1852  */
1853 static void freed_request(request_queue_t *q, int rw)
1854 {
1855         struct request_list *rl = &q->rq;
1856
1857         rl->count[rw]--;
1858
1859         __freed_request(q, rw);
1860
1861         if (unlikely(rl->starved[rw ^ 1]))
1862                 __freed_request(q, rw ^ 1);
1863
1864         if (!rl->count[READ] && !rl->count[WRITE]) {
1865                 smp_mb();
1866                 if (unlikely(waitqueue_active(&rl->drain)))
1867                         wake_up(&rl->drain);
1868         }
1869 }
1870
1871 #define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
1872 /*
1873  * Get a free request, queue_lock must not be held
1874  */
1875 static struct request *get_request(request_queue_t *q, int rw, int gfp_mask)
1876 {
1877         struct request *rq = NULL;
1878         struct request_list *rl = &q->rq;
1879         struct io_context *ioc = get_io_context(gfp_mask);
1880
1881         if (unlikely(test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags)))
1882                 goto out;
1883
1884         spin_lock_irq(q->queue_lock);
1885         if (rl->count[rw]+1 >= q->nr_requests) {
1886                 /*
1887                  * The queue will fill after this allocation, so set it as
1888                  * full, and mark this process as "batching". This process
1889                  * will be allowed to complete a batch of requests, others
1890                  * will be blocked.
1891                  */
1892                 if (!blk_queue_full(q, rw)) {
1893                         ioc_set_batching(q, ioc);
1894                         blk_set_queue_full(q, rw);
1895                 }
1896         }
1897
1898         switch (elv_may_queue(q, rw)) {
1899                 case ELV_MQUEUE_NO:
1900                         goto rq_starved;
1901                 case ELV_MQUEUE_MAY:
1902                         break;
1903                 case ELV_MQUEUE_MUST:
1904                         goto get_rq;
1905         }
1906
1907         if (blk_queue_full(q, rw) && !ioc_batching(q, ioc)) {
1908                 /*
1909                  * The queue is full and the allocating process is not a
1910                  * "batcher", and not exempted by the IO scheduler
1911                  */
1912                 spin_unlock_irq(q->queue_lock);
1913                 goto out;
1914         }
1915
1916 get_rq:
1917         rl->count[rw]++;
1918         rl->starved[rw] = 0;
1919         if (rl->count[rw] >= queue_congestion_on_threshold(q))
1920                 set_queue_congested(q, rw);
1921         spin_unlock_irq(q->queue_lock);
1922
1923         rq = blk_alloc_request(q, rw, gfp_mask);
1924         if (!rq) {
1925                 /*
1926                  * Allocation failed presumably due to memory. Undo anything
1927                  * we might have messed up.
1928                  *
1929                  * Allocating task should really be put onto the front of the
1930                  * wait queue, but this is pretty rare.
1931                  */
1932                 spin_lock_irq(q->queue_lock);
1933                 freed_request(q, rw);
1934
1935                 /*
1936                  * in the very unlikely event that allocation failed and no
1937                  * requests for this direction was pending, mark us starved
1938                  * so that freeing of a request in the other direction will
1939                  * notice us. another possible fix would be to split the
1940                  * rq mempool into READ and WRITE
1941                  */
1942 rq_starved:
1943                 if (unlikely(rl->count[rw] == 0))
1944                         rl->starved[rw] = 1;
1945
1946                 spin_unlock_irq(q->queue_lock);
1947                 goto out;
1948         }
1949
1950         if (ioc_batching(q, ioc))
1951                 ioc->nr_batch_requests--;
1952         
1953         rq_init(q, rq);
1954         rq->rl = rl;
1955 out:
1956         put_io_context(ioc);
1957         return rq;
1958 }
1959
1960 /*
1961  * No available requests for this queue, unplug the device and wait for some
1962  * requests to become available.
1963  */
1964 static struct request *get_request_wait(request_queue_t *q, int rw)
1965 {
1966         DEFINE_WAIT(wait);
1967         struct request *rq;
1968
1969         generic_unplug_device(q);
1970         do {
1971                 struct request_list *rl = &q->rq;
1972
1973                 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
1974                                 TASK_UNINTERRUPTIBLE);
1975
1976                 rq = get_request(q, rw, GFP_NOIO);
1977
1978                 if (!rq) {
1979                         struct io_context *ioc;
1980
1981                         io_schedule();
1982
1983                         /*
1984                          * After sleeping, we become a "batching" process and
1985                          * will be able to allocate at least one request, and
1986                          * up to a big batch of them for a small period time.
1987                          * See ioc_batching, ioc_set_batching
1988                          */
1989                         ioc = get_io_context(GFP_NOIO);
1990                         ioc_set_batching(q, ioc);
1991                         put_io_context(ioc);
1992                 }
1993                 finish_wait(&rl->wait[rw], &wait);
1994         } while (!rq);
1995
1996         return rq;
1997 }
1998
1999 struct request *blk_get_request(request_queue_t *q, int rw, int gfp_mask)
2000 {
2001         struct request *rq;
2002
2003         BUG_ON(rw != READ && rw != WRITE);
2004
2005         if (gfp_mask & __GFP_WAIT)
2006                 rq = get_request_wait(q, rw);
2007         else
2008                 rq = get_request(q, rw, gfp_mask);
2009
2010         return rq;
2011 }
2012
2013 EXPORT_SYMBOL(blk_get_request);
2014
2015 /**
2016  * blk_requeue_request - put a request back on queue
2017  * @q:          request queue where request should be inserted
2018  * @rq:         request to be inserted
2019  *
2020  * Description:
2021  *    Drivers often keep queueing requests until the hardware cannot accept
2022  *    more, when that condition happens we need to put the request back
2023  *    on the queue. Must be called with queue lock held.
2024  */
2025 void blk_requeue_request(request_queue_t *q, struct request *rq)
2026 {
2027         if (blk_rq_tagged(rq))
2028                 blk_queue_end_tag(q, rq);
2029
2030         elv_requeue_request(q, rq);
2031 }
2032
2033 EXPORT_SYMBOL(blk_requeue_request);
2034
2035 /**
2036  * blk_insert_request - insert a special request in to a request queue
2037  * @q:          request queue where request should be inserted
2038  * @rq:         request to be inserted
2039  * @at_head:    insert request at head or tail of queue
2040  * @data:       private data
2041  * @reinsert:   true if request it a reinsertion of previously processed one
2042  *
2043  * Description:
2044  *    Many block devices need to execute commands asynchronously, so they don't
2045  *    block the whole kernel from preemption during request execution.  This is
2046  *    accomplished normally by inserting aritficial requests tagged as
2047  *    REQ_SPECIAL in to the corresponding request queue, and letting them be
2048  *    scheduled for actual execution by the request queue.
2049  *
2050  *    We have the option of inserting the head or the tail of the queue.
2051  *    Typically we use the tail for new ioctls and so forth.  We use the head
2052  *    of the queue for things like a QUEUE_FULL message from a device, or a
2053  *    host that is unable to accept a particular command.
2054  */
2055 void blk_insert_request(request_queue_t *q, struct request *rq,
2056                         int at_head, void *data, int reinsert)
2057 {
2058         unsigned long flags;
2059
2060         /*
2061          * tell I/O scheduler that this isn't a regular read/write (ie it
2062          * must not attempt merges on this) and that it acts as a soft
2063          * barrier
2064          */
2065         rq->flags |= REQ_SPECIAL | REQ_SOFTBARRIER;
2066
2067         rq->special = data;
2068
2069         spin_lock_irqsave(q->queue_lock, flags);
2070
2071         /*
2072          * If command is tagged, release the tag
2073          */
2074         if (reinsert)
2075                 blk_requeue_request(q, rq);
2076         else {
2077                 int where = ELEVATOR_INSERT_BACK;
2078
2079                 if (at_head)
2080                         where = ELEVATOR_INSERT_FRONT;
2081
2082                 if (blk_rq_tagged(rq))
2083                         blk_queue_end_tag(q, rq);
2084
2085                 drive_stat_acct(rq, rq->nr_sectors, 1);
2086                 __elv_add_request(q, rq, where, 0);
2087         }
2088         if (blk_queue_plugged(q))
2089                 __generic_unplug_device(q);
2090         else
2091                 q->request_fn(q);
2092         spin_unlock_irqrestore(q->queue_lock, flags);
2093 }
2094
2095 EXPORT_SYMBOL(blk_insert_request);
2096
2097 /**
2098  * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2099  * @q:          request queue where request should be inserted
2100  * @rw:         READ or WRITE data
2101  * @ubuf:       the user buffer
2102  * @len:        length of user data
2103  *
2104  * Description:
2105  *    Data will be mapped directly for zero copy io, if possible. Otherwise
2106  *    a kernel bounce buffer is used.
2107  *
2108  *    A matching blk_rq_unmap_user() must be issued at the end of io, while
2109  *    still in process context.
2110  *
2111  *    Note: The mapped bio may need to be bounced through blk_queue_bounce()
2112  *    before being submitted to the device, as pages mapped may be out of
2113  *    reach. It's the callers responsibility to make sure this happens. The
2114  *    original bio must be passed back in to blk_rq_unmap_user() for proper
2115  *    unmapping.
2116  */
2117 struct request *blk_rq_map_user(request_queue_t *q, int rw, void __user *ubuf,
2118                                 unsigned int len)
2119 {
2120         unsigned long uaddr;
2121         struct request *rq;
2122         struct bio *bio;
2123
2124         if (len > (q->max_sectors << 9))
2125                 return ERR_PTR(-EINVAL);
2126         if ((!len && ubuf) || (len && !ubuf))
2127                 return ERR_PTR(-EINVAL);
2128
2129         rq = blk_get_request(q, rw, __GFP_WAIT);
2130         if (!rq)
2131                 return ERR_PTR(-ENOMEM);
2132
2133         /*
2134          * if alignment requirement is satisfied, map in user pages for
2135          * direct dma. else, set up kernel bounce buffers
2136          */
2137         uaddr = (unsigned long) ubuf;
2138         if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2139                 bio = bio_map_user(q, NULL, uaddr, len, rw == READ);
2140         else
2141                 bio = bio_copy_user(q, uaddr, len, rw == READ);
2142
2143         if (!IS_ERR(bio)) {
2144                 rq->bio = rq->biotail = bio;
2145                 blk_rq_bio_prep(q, rq, bio);
2146
2147                 rq->buffer = rq->data = NULL;
2148                 rq->data_len = len;
2149                 return rq;
2150         }
2151
2152         /*
2153          * bio is the err-ptr
2154          */
2155         blk_put_request(rq);
2156         return (struct request *) bio;
2157 }
2158
2159 EXPORT_SYMBOL(blk_rq_map_user);
2160
2161 /**
2162  * blk_rq_unmap_user - unmap a request with user data
2163  * @rq:         request to be unmapped
2164  * @bio:        bio for the request
2165  * @ulen:       length of user buffer
2166  *
2167  * Description:
2168  *    Unmap a request previously mapped by blk_rq_map_user().
2169  */
2170 int blk_rq_unmap_user(struct request *rq, struct bio *bio, unsigned int ulen)
2171 {
2172         int ret = 0;
2173
2174         if (bio) {
2175                 if (bio_flagged(bio, BIO_USER_MAPPED))
2176                         bio_unmap_user(bio);
2177                 else
2178                         ret = bio_uncopy_user(bio);
2179         }
2180
2181         blk_put_request(rq);
2182         return ret;
2183 }
2184
2185 EXPORT_SYMBOL(blk_rq_unmap_user);
2186
2187 /**
2188  * blk_execute_rq - insert a request into queue for execution
2189  * @q:          queue to insert the request in
2190  * @bd_disk:    matching gendisk
2191  * @rq:         request to insert
2192  *
2193  * Description:
2194  *    Insert a fully prepared request at the back of the io scheduler queue
2195  *    for execution.
2196  */
2197 int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
2198                    struct request *rq)
2199 {
2200         DECLARE_COMPLETION(wait);
2201         char sense[SCSI_SENSE_BUFFERSIZE];
2202         int err = 0;
2203
2204         rq->rq_disk = bd_disk;
2205
2206         /*
2207          * we need an extra reference to the request, so we can look at
2208          * it after io completion
2209          */
2210         rq->ref_count++;
2211
2212         if (!rq->sense) {
2213                 memset(sense, 0, sizeof(sense));
2214                 rq->sense = sense;
2215                 rq->sense_len = 0;
2216         }
2217
2218         rq->flags |= REQ_NOMERGE;
2219         rq->waiting = &wait;
2220         rq->end_io = blk_end_sync_rq;
2221         elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 1);
2222         generic_unplug_device(q);
2223         wait_for_completion(&wait);
2224         rq->waiting = NULL;
2225
2226         if (rq->errors)
2227                 err = -EIO;
2228
2229         return err;
2230 }
2231
2232 EXPORT_SYMBOL(blk_execute_rq);
2233
2234 /**
2235  * blkdev_issue_flush - queue a flush
2236  * @bdev:       blockdev to issue flush for
2237  * @error_sector:       error sector
2238  *
2239  * Description:
2240  *    Issue a flush for the block device in question. Caller can supply
2241  *    room for storing the error offset in case of a flush error, if they
2242  *    wish to.  Caller must run wait_for_completion() on its own.
2243  */
2244 int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2245 {
2246         request_queue_t *q;
2247
2248         if (bdev->bd_disk == NULL)
2249                 return -ENXIO;
2250
2251         q = bdev_get_queue(bdev);
2252         if (!q)
2253                 return -ENXIO;
2254         if (!q->issue_flush_fn)
2255                 return -EOPNOTSUPP;
2256
2257         return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2258 }
2259
2260 EXPORT_SYMBOL(blkdev_issue_flush);
2261
2262 /**
2263  * blkdev_scsi_issue_flush_fn - issue flush for SCSI devices
2264  * @q:          device queue
2265  * @disk:       gendisk
2266  * @error_sector:       error offset
2267  *
2268  * Description:
2269  *    Devices understanding the SCSI command set, can use this function as
2270  *    a helper for issuing a cache flush. Note: driver is required to store
2271  *    the error offset (in case of error flushing) in ->sector of struct
2272  *    request.
2273  */
2274 int blkdev_scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
2275                                sector_t *error_sector)
2276 {
2277         struct request *rq = blk_get_request(q, WRITE, __GFP_WAIT);
2278         int ret;
2279
2280         rq->flags |= REQ_BLOCK_PC | REQ_SOFTBARRIER;
2281         rq->sector = 0;
2282         memset(rq->cmd, 0, sizeof(rq->cmd));
2283         rq->cmd[0] = 0x35;
2284         rq->cmd_len = 12;
2285         rq->data = NULL;
2286         rq->data_len = 0;
2287         rq->timeout = 60 * HZ;
2288
2289         ret = blk_execute_rq(q, disk, rq);
2290
2291         if (ret && error_sector)
2292                 *error_sector = rq->sector;
2293
2294         blk_put_request(rq);
2295         return ret;
2296 }
2297
2298 EXPORT_SYMBOL(blkdev_scsi_issue_flush_fn);
2299
2300 void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
2301 {
2302         int rw = rq_data_dir(rq);
2303
2304         if (!blk_fs_request(rq) || !rq->rq_disk)
2305                 return;
2306
2307         if (rw == READ) {
2308                 __disk_stat_add(rq->rq_disk, read_sectors, nr_sectors);
2309                 if (!new_io)
2310                         __disk_stat_inc(rq->rq_disk, read_merges);
2311         } else if (rw == WRITE) {
2312                 __disk_stat_add(rq->rq_disk, write_sectors, nr_sectors);
2313                 if (!new_io)
2314                         __disk_stat_inc(rq->rq_disk, write_merges);
2315         }
2316         if (new_io) {
2317                 disk_round_stats(rq->rq_disk);
2318                 rq->rq_disk->in_flight++;
2319         }
2320 }
2321
2322 /*
2323  * add-request adds a request to the linked list.
2324  * queue lock is held and interrupts disabled, as we muck with the
2325  * request queue list.
2326  */
2327 static inline void add_request(request_queue_t * q, struct request * req)
2328 {
2329         drive_stat_acct(req, req->nr_sectors, 1);
2330
2331         if (q->activity_fn)
2332                 q->activity_fn(q->activity_data, rq_data_dir(req));
2333
2334         /*
2335          * elevator indicated where it wants this request to be
2336          * inserted at elevator_merge time
2337          */
2338         __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2339 }
2340  
2341 /*
2342  * disk_round_stats()   - Round off the performance stats on a struct
2343  * disk_stats.
2344  *
2345  * The average IO queue length and utilisation statistics are maintained
2346  * by observing the current state of the queue length and the amount of
2347  * time it has been in this state for.
2348  *
2349  * Normally, that accounting is done on IO completion, but that can result
2350  * in more than a second's worth of IO being accounted for within any one
2351  * second, leading to >100% utilisation.  To deal with that, we call this
2352  * function to do a round-off before returning the results when reading
2353  * /proc/diskstats.  This accounts immediately for all queue usage up to
2354  * the current jiffies and restarts the counters again.
2355  */
2356 void disk_round_stats(struct gendisk *disk)
2357 {
2358         unsigned long now = jiffies;
2359
2360         __disk_stat_add(disk, time_in_queue,
2361                         disk->in_flight * (now - disk->stamp));
2362         disk->stamp = now;
2363
2364         if (disk->in_flight)
2365                 __disk_stat_add(disk, io_ticks, (now - disk->stamp_idle));
2366         disk->stamp_idle = now;
2367 }
2368
2369 /*
2370  * queue lock must be held
2371  */
2372 static void __blk_put_request(request_queue_t *q, struct request *req)
2373 {
2374         struct request_list *rl = req->rl;
2375
2376         if (unlikely(!q))
2377                 return;
2378         if (unlikely(--req->ref_count))
2379                 return;
2380
2381         req->rq_status = RQ_INACTIVE;
2382         req->q = NULL;
2383         req->rl = NULL;
2384
2385         /*
2386          * Request may not have originated from ll_rw_blk. if not,
2387          * it didn't come out of our reserved rq pools
2388          */
2389         if (rl) {
2390                 int rw = rq_data_dir(req);
2391
2392                 elv_completed_request(q, req);
2393
2394                 BUG_ON(!list_empty(&req->queuelist));
2395
2396                 blk_free_request(q, req);
2397                 freed_request(q, rw);
2398         }
2399 }
2400
2401 void blk_put_request(struct request *req)
2402 {
2403         /*
2404          * if req->rl isn't set, this request didnt originate from the
2405          * block layer, so it's safe to just disregard it
2406          */
2407         if (req->rl) {
2408                 unsigned long flags;
2409                 request_queue_t *q = req->q;
2410
2411                 spin_lock_irqsave(q->queue_lock, flags);
2412                 __blk_put_request(q, req);
2413                 spin_unlock_irqrestore(q->queue_lock, flags);
2414         }
2415 }
2416
2417 EXPORT_SYMBOL(blk_put_request);
2418
2419 /**
2420  * blk_end_sync_rq - executes a completion event on a request
2421  * @rq: request to complete
2422  */
2423 void blk_end_sync_rq(struct request *rq)
2424 {
2425         struct completion *waiting = rq->waiting;
2426
2427         rq->waiting = NULL;
2428         __blk_put_request(rq->q, rq);
2429
2430         /*
2431          * complete last, if this is a stack request the process (and thus
2432          * the rq pointer) could be invalid right after this complete()
2433          */
2434         complete(waiting);
2435 }
2436 EXPORT_SYMBOL(blk_end_sync_rq);
2437
2438 /**
2439  * blk_congestion_wait - wait for a queue to become uncongested
2440  * @rw: READ or WRITE
2441  * @timeout: timeout in jiffies
2442  *
2443  * Waits for up to @timeout jiffies for a queue (any queue) to exit congestion.
2444  * If no queues are congested then just wait for the next request to be
2445  * returned.
2446  */
2447 long blk_congestion_wait(int rw, long timeout)
2448 {
2449         long ret;
2450         DEFINE_WAIT(wait);
2451         wait_queue_head_t *wqh = &congestion_wqh[rw];
2452
2453         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
2454         ret = io_schedule_timeout(timeout);
2455         finish_wait(wqh, &wait);
2456         return ret;
2457 }
2458
2459 EXPORT_SYMBOL(blk_congestion_wait);
2460
2461 /*
2462  * Has to be called with the request spinlock acquired
2463  */
2464 static int attempt_merge(request_queue_t *q, struct request *req,
2465                           struct request *next)
2466 {
2467         if (!rq_mergeable(req) || !rq_mergeable(next))
2468                 return 0;
2469
2470         /*
2471          * not contigious
2472          */
2473         if (req->sector + req->nr_sectors != next->sector)
2474                 return 0;
2475
2476         if (rq_data_dir(req) != rq_data_dir(next)
2477             || req->rq_disk != next->rq_disk
2478             || next->waiting || next->special)
2479                 return 0;
2480
2481         /*
2482          * If we are allowed to merge, then append bio list
2483          * from next to rq and release next. merge_requests_fn
2484          * will have updated segment counts, update sector
2485          * counts here.
2486          */
2487         if (!q->merge_requests_fn(q, req, next))
2488                 return 0;
2489
2490         /*
2491          * At this point we have either done a back merge
2492          * or front merge. We need the smaller start_time of
2493          * the merged requests to be the current request
2494          * for accounting purposes.
2495          */
2496         if (time_after(req->start_time, next->start_time))
2497                 req->start_time = next->start_time;
2498
2499         req->biotail->bi_next = next->bio;
2500         req->biotail = next->biotail;
2501
2502         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2503
2504         elv_merge_requests(q, req, next);
2505
2506         if (req->rq_disk) {
2507                 disk_round_stats(req->rq_disk);
2508                 req->rq_disk->in_flight--;
2509         }
2510
2511         __blk_put_request(q, next);
2512         return 1;
2513 }
2514
2515 static inline int attempt_back_merge(request_queue_t *q, struct request *rq)
2516 {
2517         struct request *next = elv_latter_request(q, rq);
2518
2519         if (next)
2520                 return attempt_merge(q, rq, next);
2521
2522         return 0;
2523 }
2524
2525 static inline int attempt_front_merge(request_queue_t *q, struct request *rq)
2526 {
2527         struct request *prev = elv_former_request(q, rq);
2528
2529         if (prev)
2530                 return attempt_merge(q, prev, rq);
2531
2532         return 0;
2533 }
2534
2535 /**
2536  * blk_attempt_remerge  - attempt to remerge active head with next request
2537  * @q:    The &request_queue_t belonging to the device
2538  * @rq:   The head request (usually)
2539  *
2540  * Description:
2541  *    For head-active devices, the queue can easily be unplugged so quickly
2542  *    that proper merging is not done on the front request. This may hurt
2543  *    performance greatly for some devices. The block layer cannot safely
2544  *    do merging on that first request for these queues, but the driver can
2545  *    call this function and make it happen any way. Only the driver knows
2546  *    when it is safe to do so.
2547  **/
2548 void blk_attempt_remerge(request_queue_t *q, struct request *rq)
2549 {
2550         unsigned long flags;
2551
2552         spin_lock_irqsave(q->queue_lock, flags);
2553         attempt_back_merge(q, rq);
2554         spin_unlock_irqrestore(q->queue_lock, flags);
2555 }
2556
2557 EXPORT_SYMBOL(blk_attempt_remerge);
2558
2559 /*
2560  * Non-locking blk_attempt_remerge variant.
2561  */
2562 void __blk_attempt_remerge(request_queue_t *q, struct request *rq)
2563 {
2564         attempt_back_merge(q, rq);
2565 }
2566
2567 EXPORT_SYMBOL(__blk_attempt_remerge);
2568
2569 static int __make_request(request_queue_t *q, struct bio *bio)
2570 {
2571         struct request *req, *freereq = NULL;
2572         int el_ret, rw, nr_sectors, cur_nr_sectors, barrier, err, sync;
2573         sector_t sector;
2574
2575         sector = bio->bi_sector;
2576         nr_sectors = bio_sectors(bio);
2577         cur_nr_sectors = bio_cur_sectors(bio);
2578
2579         rw = bio_data_dir(bio);
2580         sync = bio_sync(bio);
2581
2582         /*
2583          * low level driver can indicate that it wants pages above a
2584          * certain limit bounced to low memory (ie for highmem, or even
2585          * ISA dma in theory)
2586          */
2587         blk_queue_bounce(q, &bio);
2588
2589         spin_lock_prefetch(q->queue_lock);
2590
2591         barrier = bio_barrier(bio);
2592         if (barrier && (q->ordered == QUEUE_ORDERED_NONE)) {
2593                 err = -EOPNOTSUPP;
2594                 goto end_io;
2595         }
2596
2597 again:
2598         spin_lock_irq(q->queue_lock);
2599
2600         if (elv_queue_empty(q)) {
2601                 blk_plug_device(q);
2602                 goto get_rq;
2603         }
2604         if (barrier)
2605                 goto get_rq;
2606
2607         el_ret = elv_merge(q, &req, bio);
2608         switch (el_ret) {
2609                 case ELEVATOR_BACK_MERGE:
2610                         BUG_ON(!rq_mergeable(req));
2611
2612                         if (!q->back_merge_fn(q, req, bio))
2613                                 break;
2614
2615                         req->biotail->bi_next = bio;
2616                         req->biotail = bio;
2617                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2618                         drive_stat_acct(req, nr_sectors, 0);
2619                         if (!attempt_back_merge(q, req))
2620                                 elv_merged_request(q, req);
2621                         goto out;
2622
2623                 case ELEVATOR_FRONT_MERGE:
2624                         BUG_ON(!rq_mergeable(req));
2625
2626                         if (!q->front_merge_fn(q, req, bio))
2627                                 break;
2628
2629                         bio->bi_next = req->bio;
2630                         req->bio = bio;
2631
2632                         /*
2633                          * may not be valid. if the low level driver said
2634                          * it didn't need a bounce buffer then it better
2635                          * not touch req->buffer either...
2636                          */
2637                         req->buffer = bio_data(bio);
2638                         req->current_nr_sectors = cur_nr_sectors;
2639                         req->hard_cur_sectors = cur_nr_sectors;
2640                         req->sector = req->hard_sector = sector;
2641                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
2642                         drive_stat_acct(req, nr_sectors, 0);
2643                         if (!attempt_front_merge(q, req))
2644                                 elv_merged_request(q, req);
2645                         goto out;
2646
2647                 /*
2648                  * elevator says don't/can't merge. get new request
2649                  */
2650                 case ELEVATOR_NO_MERGE:
2651                         break;
2652
2653                 default:
2654                         printk("elevator returned crap (%d)\n", el_ret);
2655                         BUG();
2656         }
2657
2658         /*
2659          * Grab a free request from the freelist - if that is empty, check
2660          * if we are doing read ahead and abort instead of blocking for
2661          * a free slot.
2662          */
2663 get_rq:
2664         if (freereq) {
2665                 req = freereq;
2666                 freereq = NULL;
2667         } else {
2668                 spin_unlock_irq(q->queue_lock);
2669                 if ((freereq = get_request(q, rw, GFP_ATOMIC)) == NULL) {
2670                         /*
2671                          * READA bit set
2672                          */
2673                         err = -EWOULDBLOCK;
2674                         if (bio_rw_ahead(bio))
2675                                 goto end_io;
2676         
2677                         freereq = get_request_wait(q, rw);
2678                 }
2679                 goto again;
2680         }
2681
2682         req->flags |= REQ_CMD;
2683
2684         /*
2685          * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2686          */
2687         if (bio_rw_ahead(bio) || bio_failfast(bio))
2688                 req->flags |= REQ_FAILFAST;
2689
2690         /*
2691          * REQ_BARRIER implies no merging, but lets make it explicit
2692          */
2693         if (barrier)
2694                 req->flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2695
2696         req->errors = 0;
2697         req->hard_sector = req->sector = sector;
2698         req->hard_nr_sectors = req->nr_sectors = nr_sectors;
2699         req->current_nr_sectors = req->hard_cur_sectors = cur_nr_sectors;
2700         req->nr_phys_segments = bio_phys_segments(q, bio);
2701         req->nr_hw_segments = bio_hw_segments(q, bio);
2702         req->buffer = bio_data(bio);    /* see ->buffer comment above */
2703         req->waiting = NULL;
2704         req->bio = req->biotail = bio;
2705         req->rq_disk = bio->bi_bdev->bd_disk;
2706         req->start_time = jiffies;
2707
2708         add_request(q, req);
2709 out:
2710         if (freereq)
2711                 __blk_put_request(q, freereq);
2712         if (sync)
2713                 __generic_unplug_device(q);
2714
2715         spin_unlock_irq(q->queue_lock);
2716         return 0;
2717
2718 end_io:
2719         bio_endio(bio, nr_sectors << 9, err);
2720         return 0;
2721 }
2722
2723 /*
2724  * If bio->bi_dev is a partition, remap the location
2725  */
2726 static inline void blk_partition_remap(struct bio *bio)
2727 {
2728         struct block_device *bdev = bio->bi_bdev;
2729
2730         if (bdev != bdev->bd_contains) {
2731                 struct hd_struct *p = bdev->bd_part;
2732
2733                 switch (bio->bi_rw) {
2734                 case READ:
2735                         p->read_sectors += bio_sectors(bio);
2736                         p->reads++;
2737                         break;
2738                 case WRITE:
2739                         p->write_sectors += bio_sectors(bio);
2740                         p->writes++;
2741                         break;
2742                 }
2743                 bio->bi_sector += p->start_sect;
2744                 bio->bi_bdev = bdev->bd_contains;
2745         }
2746 }
2747
2748 void blk_finish_queue_drain(request_queue_t *q)
2749 {
2750         struct request_list *rl = &q->rq;
2751         struct request *rq;
2752
2753         spin_lock_irq(q->queue_lock);
2754         clear_bit(QUEUE_FLAG_DRAIN, &q->queue_flags);
2755
2756         while (!list_empty(&q->drain_list)) {
2757                 rq = list_entry_rq(q->drain_list.next);
2758
2759                 list_del_init(&rq->queuelist);
2760                 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 1);
2761         }
2762
2763         spin_unlock_irq(q->queue_lock);
2764
2765         wake_up(&rl->wait[0]);
2766         wake_up(&rl->wait[1]);
2767         wake_up(&rl->drain);
2768 }
2769
2770 static int wait_drain(request_queue_t *q, struct request_list *rl, int dispatch)
2771 {
2772         int wait = rl->count[READ] + rl->count[WRITE];
2773
2774         if (dispatch)
2775                 wait += !list_empty(&q->queue_head);
2776
2777         return wait;
2778 }
2779
2780 /*
2781  * We rely on the fact that only requests allocated through blk_alloc_request()
2782  * have io scheduler private data structures associated with them. Any other
2783  * type of request (allocated on stack or through kmalloc()) should not go
2784  * to the io scheduler core, but be attached to the queue head instead.
2785  */
2786 void blk_wait_queue_drained(request_queue_t *q, int wait_dispatch)
2787 {
2788         struct request_list *rl = &q->rq;
2789         DEFINE_WAIT(wait);
2790
2791         spin_lock_irq(q->queue_lock);
2792         set_bit(QUEUE_FLAG_DRAIN, &q->queue_flags);
2793
2794         while (wait_drain(q, rl, wait_dispatch)) {
2795                 prepare_to_wait(&rl->drain, &wait, TASK_UNINTERRUPTIBLE);
2796
2797                 if (wait_drain(q, rl, wait_dispatch)) {
2798                         __generic_unplug_device(q);
2799                         spin_unlock_irq(q->queue_lock);
2800                         io_schedule();
2801                         spin_lock_irq(q->queue_lock);
2802                 }
2803
2804                 finish_wait(&rl->drain, &wait);
2805         }
2806
2807         spin_unlock_irq(q->queue_lock);
2808 }
2809
2810 /*
2811  * block waiting for the io scheduler being started again.
2812  */
2813 static inline void block_wait_queue_running(request_queue_t *q)
2814 {
2815         DEFINE_WAIT(wait);
2816
2817         while (test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags)) {
2818                 struct request_list *rl = &q->rq;
2819
2820                 prepare_to_wait_exclusive(&rl->drain, &wait,
2821                                 TASK_UNINTERRUPTIBLE);
2822
2823                 /*
2824                  * re-check the condition. avoids using prepare_to_wait()
2825                  * in the fast path (queue is running)
2826                  */
2827                 if (test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags))
2828                         io_schedule();
2829
2830                 finish_wait(&rl->drain, &wait);
2831         }
2832 }
2833
2834 static void handle_bad_sector(struct bio *bio)
2835 {
2836         char b[BDEVNAME_SIZE];
2837
2838         printk(KERN_INFO "attempt to access beyond end of device\n");
2839         printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
2840                         bdevname(bio->bi_bdev, b),
2841                         bio->bi_rw,
2842                         (unsigned long long)bio->bi_sector + bio_sectors(bio),
2843                         (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
2844
2845         set_bit(BIO_EOF, &bio->bi_flags);
2846 }
2847
2848 /**
2849  * generic_make_request: hand a buffer to its device driver for I/O
2850  * @bio:  The bio describing the location in memory and on the device.
2851  *
2852  * generic_make_request() is used to make I/O requests of block
2853  * devices. It is passed a &struct bio, which describes the I/O that needs
2854  * to be done.
2855  *
2856  * generic_make_request() does not return any status.  The
2857  * success/failure status of the request, along with notification of
2858  * completion, is delivered asynchronously through the bio->bi_end_io
2859  * function described (one day) else where.
2860  *
2861  * The caller of generic_make_request must make sure that bi_io_vec
2862  * are set to describe the memory buffer, and that bi_dev and bi_sector are
2863  * set to describe the device address, and the
2864  * bi_end_io and optionally bi_private are set to describe how
2865  * completion notification should be signaled.
2866  *
2867  * generic_make_request and the drivers it calls may use bi_next if this
2868  * bio happens to be merged with someone else, and may change bi_dev and
2869  * bi_sector for remaps as it sees fit.  So the values of these fields
2870  * should NOT be depended on after the call to generic_make_request.
2871  */
2872 void generic_make_request(struct bio *bio)
2873 {
2874         request_queue_t *q;
2875         sector_t maxsector;
2876         int ret, nr_sectors = bio_sectors(bio);
2877
2878         might_sleep();
2879         /* Test device or partition size, when known. */
2880         maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
2881         if (maxsector) {
2882                 sector_t sector = bio->bi_sector;
2883
2884                 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
2885                         /*
2886                          * This may well happen - the kernel calls bread()
2887                          * without checking the size of the device, e.g., when
2888                          * mounting a device.
2889                          */
2890                         handle_bad_sector(bio);
2891                         goto end_io;
2892                 }
2893         }
2894
2895         /*
2896          * Resolve the mapping until finished. (drivers are
2897          * still free to implement/resolve their own stacking
2898          * by explicitly returning 0)
2899          *
2900          * NOTE: we don't repeat the blk_size check for each new device.
2901          * Stacking drivers are expected to know what they are doing.
2902          */
2903         do {
2904                 char b[BDEVNAME_SIZE];
2905
2906                 q = bdev_get_queue(bio->bi_bdev);
2907                 if (!q) {
2908                         printk(KERN_ERR
2909                                "generic_make_request: Trying to access "
2910                                 "nonexistent block-device %s (%Lu)\n",
2911                                 bdevname(bio->bi_bdev, b),
2912                                 (long long) bio->bi_sector);
2913 end_io:
2914                         bio_endio(bio, bio->bi_size, -EIO);
2915                         break;
2916                 }
2917
2918                 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
2919                         printk("bio too big device %s (%u > %u)\n", 
2920                                 bdevname(bio->bi_bdev, b),
2921                                 bio_sectors(bio),
2922                                 q->max_hw_sectors);
2923                         goto end_io;
2924                 }
2925
2926                 if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))
2927                         goto end_io;
2928
2929                 block_wait_queue_running(q);
2930
2931                 /*
2932                  * If this device has partitions, remap block n
2933                  * of partition p to block n+start(p) of the disk.
2934                  */
2935                 blk_partition_remap(bio);
2936
2937                 ret = q->make_request_fn(q, bio);
2938         } while (ret);
2939 }
2940
2941 EXPORT_SYMBOL(generic_make_request);
2942
2943 /**
2944  * submit_bio: submit a bio to the block device layer for I/O
2945  * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
2946  * @bio: The &struct bio which describes the I/O
2947  *
2948  * submit_bio() is very similar in purpose to generic_make_request(), and
2949  * uses that function to do most of the work. Both are fairly rough
2950  * interfaces, @bio must be presetup and ready for I/O.
2951  *
2952  */
2953 void submit_bio(int rw, struct bio *bio)
2954 {
2955         int count = bio_sectors(bio);
2956
2957         BIO_BUG_ON(!bio->bi_size);
2958         BIO_BUG_ON(!bio->bi_io_vec);
2959         bio->bi_rw = rw;
2960         if (rw & WRITE)
2961                 mod_page_state(pgpgout, count);
2962         else
2963                 mod_page_state(pgpgin, count);
2964
2965         if (unlikely(block_dump)) {
2966                 char b[BDEVNAME_SIZE];
2967                 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
2968                         current->comm, current->pid,
2969                         (rw & WRITE) ? "WRITE" : "READ",
2970                         (unsigned long long)bio->bi_sector,
2971                         bdevname(bio->bi_bdev,b));
2972         }
2973
2974         generic_make_request(bio);
2975 }
2976
2977 EXPORT_SYMBOL(submit_bio);
2978
2979 void blk_recalc_rq_segments(struct request *rq)
2980 {
2981         struct bio *bio, *prevbio = NULL;
2982         int nr_phys_segs, nr_hw_segs;
2983         unsigned int phys_size, hw_size;
2984         request_queue_t *q = rq->q;
2985
2986         if (!rq->bio)
2987                 return;
2988
2989         phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
2990         rq_for_each_bio(bio, rq) {
2991                 /* Force bio hw/phys segs to be recalculated. */
2992                 bio->bi_flags &= ~(1 << BIO_SEG_VALID);
2993
2994                 nr_phys_segs += bio_phys_segments(q, bio);
2995                 nr_hw_segs += bio_hw_segments(q, bio);
2996                 if (prevbio) {
2997                         int pseg = phys_size + prevbio->bi_size + bio->bi_size;
2998                         int hseg = hw_size + prevbio->bi_size + bio->bi_size;
2999
3000                         if (blk_phys_contig_segment(q, prevbio, bio) &&
3001                             pseg <= q->max_segment_size) {
3002                                 nr_phys_segs--;
3003                                 phys_size += prevbio->bi_size + bio->bi_size;
3004                         } else
3005                                 phys_size = 0;
3006
3007                         if (blk_hw_contig_segment(q, prevbio, bio) &&
3008                             hseg <= q->max_segment_size) {
3009                                 nr_hw_segs--;
3010                                 hw_size += prevbio->bi_size + bio->bi_size;
3011                         } else
3012                                 hw_size = 0;
3013                 }
3014                 prevbio = bio;
3015         }
3016
3017         rq->nr_phys_segments = nr_phys_segs;
3018         rq->nr_hw_segments = nr_hw_segs;
3019 }
3020
3021 void blk_recalc_rq_sectors(struct request *rq, int nsect)
3022 {
3023         if (blk_fs_request(rq)) {
3024                 rq->hard_sector += nsect;
3025                 rq->hard_nr_sectors -= nsect;
3026
3027                 /*
3028                  * Move the I/O submission pointers ahead if required.
3029                  */
3030                 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3031                     (rq->sector <= rq->hard_sector)) {
3032                         rq->sector = rq->hard_sector;
3033                         rq->nr_sectors = rq->hard_nr_sectors;
3034                         rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3035                         rq->current_nr_sectors = rq->hard_cur_sectors;
3036                         rq->buffer = bio_data(rq->bio);
3037                 }
3038
3039                 /*
3040                  * if total number of sectors is less than the first segment
3041                  * size, something has gone terribly wrong
3042                  */
3043                 if (rq->nr_sectors < rq->current_nr_sectors) {
3044                         printk("blk: request botched\n");
3045                         rq->nr_sectors = rq->current_nr_sectors;
3046                 }
3047         }
3048 }
3049
3050 static int __end_that_request_first(struct request *req, int uptodate,
3051                                     int nr_bytes)
3052 {
3053         int total_bytes, bio_nbytes, error, next_idx = 0;
3054         struct bio *bio;
3055
3056         /*
3057          * extend uptodate bool to allow < 0 value to be direct io error
3058          */
3059         error = 0;
3060         if (end_io_error(uptodate))
3061                 error = !uptodate ? -EIO : uptodate;
3062
3063         /*
3064          * for a REQ_BLOCK_PC request, we want to carry any eventual
3065          * sense key with us all the way through
3066          */
3067         if (!blk_pc_request(req))
3068                 req->errors = 0;
3069
3070         if (!uptodate) {
3071                 if (blk_fs_request(req) && !(req->flags & REQ_QUIET))
3072                         printk("end_request: I/O error, dev %s, sector %llu\n",
3073                                 req->rq_disk ? req->rq_disk->disk_name : "?",
3074                                 (unsigned long long)req->sector);
3075         }
3076
3077         total_bytes = bio_nbytes = 0;
3078         while ((bio = req->bio) != NULL) {
3079                 int nbytes;
3080
3081                 if (nr_bytes >= bio->bi_size) {
3082                         req->bio = bio->bi_next;
3083                         nbytes = bio->bi_size;
3084                         bio_endio(bio, nbytes, error);
3085                         next_idx = 0;
3086                         bio_nbytes = 0;
3087                 } else {
3088                         int idx = bio->bi_idx + next_idx;
3089
3090                         if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3091                                 blk_dump_rq_flags(req, "__end_that");
3092                                 printk("%s: bio idx %d >= vcnt %d\n",
3093                                                 __FUNCTION__,
3094                                                 bio->bi_idx, bio->bi_vcnt);
3095                                 break;
3096                         }
3097
3098                         nbytes = bio_iovec_idx(bio, idx)->bv_len;
3099                         BIO_BUG_ON(nbytes > bio->bi_size);
3100
3101                         /*
3102                          * not a complete bvec done
3103                          */
3104                         if (unlikely(nbytes > nr_bytes)) {
3105                                 bio_nbytes += nr_bytes;
3106                                 total_bytes += nr_bytes;
3107                                 break;
3108                         }
3109
3110                         /*
3111                          * advance to the next vector
3112                          */
3113                         next_idx++;
3114                         bio_nbytes += nbytes;
3115                 }
3116
3117                 total_bytes += nbytes;
3118                 nr_bytes -= nbytes;
3119
3120                 if ((bio = req->bio)) {
3121                         /*
3122                          * end more in this run, or just return 'not-done'
3123                          */
3124                         if (unlikely(nr_bytes <= 0))
3125                                 break;
3126                 }
3127         }
3128
3129         /*
3130          * completely done
3131          */
3132         if (!req->bio)
3133                 return 0;
3134
3135         /*
3136          * if the request wasn't completed, update state
3137          */
3138         if (bio_nbytes) {
3139                 bio_endio(bio, bio_nbytes, error);
3140                 bio->bi_idx += next_idx;
3141                 bio_iovec(bio)->bv_offset += nr_bytes;
3142                 bio_iovec(bio)->bv_len -= nr_bytes;
3143         }
3144
3145         blk_recalc_rq_sectors(req, total_bytes >> 9);
3146         blk_recalc_rq_segments(req);
3147         return 1;
3148 }
3149
3150 /**
3151  * end_that_request_first - end I/O on a request
3152  * @req:      the request being processed
3153  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3154  * @nr_sectors: number of sectors to end I/O on
3155  *
3156  * Description:
3157  *     Ends I/O on a number of sectors attached to @req, and sets it up
3158  *     for the next range of segments (if any) in the cluster.
3159  *
3160  * Return:
3161  *     0 - we are done with this request, call end_that_request_last()
3162  *     1 - still buffers pending for this request
3163  **/
3164 int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3165 {
3166         return __end_that_request_first(req, uptodate, nr_sectors << 9);
3167 }
3168
3169 EXPORT_SYMBOL(end_that_request_first);
3170
3171 /**
3172  * end_that_request_chunk - end I/O on a request
3173  * @req:      the request being processed
3174  * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3175  * @nr_bytes: number of bytes to complete
3176  *
3177  * Description:
3178  *     Ends I/O on a number of bytes attached to @req, and sets it up
3179  *     for the next range of segments (if any). Like end_that_request_first(),
3180  *     but deals with bytes instead of sectors.
3181  *
3182  * Return:
3183  *     0 - we are done with this request, call end_that_request_last()
3184  *     1 - still buffers pending for this request
3185  **/
3186 int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3187 {
3188         return __end_that_request_first(req, uptodate, nr_bytes);
3189 }
3190
3191 EXPORT_SYMBOL(end_that_request_chunk);
3192
3193 /*
3194  * queue lock must be held
3195  */
3196 void end_that_request_last(struct request *req)
3197 {
3198         struct gendisk *disk = req->rq_disk;
3199
3200         if (unlikely(laptop_mode) && blk_fs_request(req))
3201                 laptop_io_completion();
3202
3203         if (disk && blk_fs_request(req)) {
3204                 unsigned long duration = jiffies - req->start_time;
3205                 switch (rq_data_dir(req)) {
3206                     case WRITE:
3207                         __disk_stat_inc(disk, writes);
3208                         __disk_stat_add(disk, write_ticks, duration);
3209                         break;
3210                     case READ:
3211                         __disk_stat_inc(disk, reads);
3212                         __disk_stat_add(disk, read_ticks, duration);
3213                         break;
3214                 }
3215                 disk_round_stats(disk);
3216                 disk->in_flight--;
3217         }
3218         if (req->end_io)
3219                 req->end_io(req);
3220         else
3221                 __blk_put_request(req->q, req);
3222 }
3223
3224 EXPORT_SYMBOL(end_that_request_last);
3225
3226 void end_request(struct request *req, int uptodate)
3227 {
3228         if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3229                 add_disk_randomness(req->rq_disk);
3230                 blkdev_dequeue_request(req);
3231                 end_that_request_last(req);
3232         }
3233 }
3234
3235 EXPORT_SYMBOL(end_request);
3236
3237 void blk_rq_bio_prep(request_queue_t *q, struct request *rq, struct bio *bio)
3238 {
3239         /* first three bits are identical in rq->flags and bio->bi_rw */
3240         rq->flags |= (bio->bi_rw & 7);
3241
3242         rq->nr_phys_segments = bio_phys_segments(q, bio);
3243         rq->nr_hw_segments = bio_hw_segments(q, bio);
3244         rq->current_nr_sectors = bio_cur_sectors(bio);
3245         rq->hard_cur_sectors = rq->current_nr_sectors;
3246         rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3247         rq->buffer = bio_data(bio);
3248
3249         rq->bio = rq->biotail = bio;
3250 }
3251
3252 EXPORT_SYMBOL(blk_rq_bio_prep);
3253
3254 int kblockd_schedule_work(struct work_struct *work)
3255 {
3256         return queue_work(kblockd_workqueue, work);
3257 }
3258
3259 EXPORT_SYMBOL(kblockd_schedule_work);
3260
3261 void kblockd_flush(void)
3262 {
3263         flush_workqueue(kblockd_workqueue);
3264 }
3265 EXPORT_SYMBOL(kblockd_flush);
3266
3267 int __init blk_dev_init(void)
3268 {
3269         kblockd_workqueue = create_workqueue("kblockd");
3270         if (!kblockd_workqueue)
3271                 panic("Failed to create kblockd\n");
3272
3273         request_cachep = kmem_cache_create("blkdev_requests",
3274                         sizeof(struct request), 0, SLAB_PANIC, NULL, NULL);
3275
3276         requestq_cachep = kmem_cache_create("blkdev_queue",
3277                         sizeof(request_queue_t), 0, SLAB_PANIC, NULL, NULL);
3278
3279         iocontext_cachep = kmem_cache_create("blkdev_ioc",
3280                         sizeof(struct io_context), 0, SLAB_PANIC, NULL, NULL);
3281
3282         blk_max_low_pfn = max_low_pfn;
3283         blk_max_pfn = max_pfn;
3284
3285         return 0;
3286 }
3287
3288 /*
3289  * IO Context helper functions
3290  */
3291 void put_io_context(struct io_context *ioc)
3292 {
3293         if (ioc == NULL)
3294                 return;
3295
3296         BUG_ON(atomic_read(&ioc->refcount) == 0);
3297
3298         if (atomic_dec_and_test(&ioc->refcount)) {
3299                 if (ioc->aic && ioc->aic->dtor)
3300                         ioc->aic->dtor(ioc->aic);
3301                 if (ioc->cic && ioc->cic->dtor)
3302                         ioc->cic->dtor(ioc->cic);
3303
3304                 kmem_cache_free(iocontext_cachep, ioc);
3305         }
3306 }
3307 EXPORT_SYMBOL(put_io_context);
3308
3309 /* Called by the exitting task */
3310 void exit_io_context(void)
3311 {
3312         unsigned long flags;
3313         struct io_context *ioc;
3314
3315         local_irq_save(flags);
3316         ioc = current->io_context;
3317         current->io_context = NULL;
3318         local_irq_restore(flags);
3319
3320         if (ioc->aic && ioc->aic->exit)
3321                 ioc->aic->exit(ioc->aic);
3322         if (ioc->cic && ioc->cic->exit)
3323                 ioc->cic->exit(ioc->cic);
3324
3325         put_io_context(ioc);
3326 }
3327
3328 /*
3329  * If the current task has no IO context then create one and initialise it.
3330  * If it does have a context, take a ref on it.
3331  *
3332  * This is always called in the context of the task which submitted the I/O.
3333  * But weird things happen, so we disable local interrupts to ensure exclusive
3334  * access to *current.
3335  */
3336 struct io_context *get_io_context(int gfp_flags)
3337 {
3338         struct task_struct *tsk = current;
3339         unsigned long flags;
3340         struct io_context *ret;
3341
3342         local_irq_save(flags);
3343         ret = tsk->io_context;
3344         if (ret)
3345                 goto out;
3346
3347         local_irq_restore(flags);
3348
3349         ret = kmem_cache_alloc(iocontext_cachep, gfp_flags);
3350         if (ret) {
3351                 atomic_set(&ret->refcount, 1);
3352                 ret->pid = tsk->pid;
3353                 ret->last_waited = jiffies; /* doesn't matter... */
3354                 ret->nr_batch_requests = 0; /* because this is 0 */
3355                 ret->aic = NULL;
3356                 ret->cic = NULL;
3357                 spin_lock_init(&ret->lock);
3358
3359                 local_irq_save(flags);
3360
3361                 /*
3362                  * very unlikely, someone raced with us in setting up the task
3363                  * io context. free new context and just grab a reference.
3364                  */
3365                 if (!tsk->io_context)
3366                         tsk->io_context = ret;
3367                 else {
3368                         kmem_cache_free(iocontext_cachep, ret);
3369                         ret = tsk->io_context;
3370                 }
3371
3372 out:
3373                 atomic_inc(&ret->refcount);
3374                 local_irq_restore(flags);
3375         }
3376
3377         return ret;
3378 }
3379 EXPORT_SYMBOL(get_io_context);
3380
3381 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3382 {
3383         struct io_context *src = *psrc;
3384         struct io_context *dst = *pdst;
3385
3386         if (src) {
3387                 BUG_ON(atomic_read(&src->refcount) == 0);
3388                 atomic_inc(&src->refcount);
3389                 put_io_context(dst);
3390                 *pdst = src;
3391         }
3392 }
3393 EXPORT_SYMBOL(copy_io_context);
3394
3395 void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3396 {
3397         struct io_context *temp;
3398         temp = *ioc1;
3399         *ioc1 = *ioc2;
3400         *ioc2 = temp;
3401 }
3402 EXPORT_SYMBOL(swap_io_context);
3403
3404 /*
3405  * sysfs parts below
3406  */
3407 struct queue_sysfs_entry {
3408         struct attribute attr;
3409         ssize_t (*show)(struct request_queue *, char *);
3410         ssize_t (*store)(struct request_queue *, const char *, size_t);
3411 };
3412
3413 static ssize_t
3414 queue_var_show(unsigned int var, char *page)
3415 {
3416         return sprintf(page, "%d\n", var);
3417 }
3418
3419 static ssize_t
3420 queue_var_store(unsigned long *var, const char *page, size_t count)
3421 {
3422         char *p = (char *) page;
3423
3424         *var = simple_strtoul(p, &p, 10);
3425         return count;
3426 }
3427
3428 static ssize_t queue_requests_show(struct request_queue *q, char *page)
3429 {
3430         return queue_var_show(q->nr_requests, (page));
3431 }
3432
3433 static ssize_t
3434 queue_requests_store(struct request_queue *q, const char *page, size_t count)
3435 {
3436         struct request_list *rl = &q->rq;
3437
3438         int ret = queue_var_store(&q->nr_requests, page, count);
3439         if (q->nr_requests < BLKDEV_MIN_RQ)
3440                 q->nr_requests = BLKDEV_MIN_RQ;
3441         blk_queue_congestion_threshold(q);
3442
3443         if (rl->count[READ] >= queue_congestion_on_threshold(q))
3444                 set_queue_congested(q, READ);
3445         else if (rl->count[READ] < queue_congestion_off_threshold(q))
3446                 clear_queue_congested(q, READ);
3447
3448         if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3449                 set_queue_congested(q, WRITE);
3450         else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3451                 clear_queue_congested(q, WRITE);
3452
3453         if (rl->count[READ] >= q->nr_requests) {
3454                 blk_set_queue_full(q, READ);
3455         } else if (rl->count[READ]+1 <= q->nr_requests) {
3456                 blk_clear_queue_full(q, READ);
3457                 wake_up(&rl->wait[READ]);
3458         }
3459
3460         if (rl->count[WRITE] >= q->nr_requests) {
3461                 blk_set_queue_full(q, WRITE);
3462         } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3463                 blk_clear_queue_full(q, WRITE);
3464                 wake_up(&rl->wait[WRITE]);
3465         }
3466         return ret;
3467 }
3468
3469 static ssize_t queue_ra_show(struct request_queue *q, char *page)
3470 {
3471         int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3472
3473         return queue_var_show(ra_kb, (page));
3474 }
3475
3476 static ssize_t
3477 queue_ra_store(struct request_queue *q, const char *page, size_t count)
3478 {
3479         unsigned long ra_kb;
3480         ssize_t ret = queue_var_store(&ra_kb, page, count);
3481
3482         spin_lock_irq(q->queue_lock);
3483         if (ra_kb > (q->max_sectors >> 1))
3484                 ra_kb = (q->max_sectors >> 1);
3485
3486         q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3487         spin_unlock_irq(q->queue_lock);
3488
3489         return ret;
3490 }
3491
3492 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3493 {
3494         int max_sectors_kb = q->max_sectors >> 1;
3495
3496         return queue_var_show(max_sectors_kb, (page));
3497 }
3498
3499 static ssize_t
3500 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3501 {
3502         unsigned long max_sectors_kb,
3503                         max_hw_sectors_kb = q->max_hw_sectors >> 1,
3504                         page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3505         ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3506         int ra_kb;
3507
3508         if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3509                 return -EINVAL;
3510         /*
3511          * Take the queue lock to update the readahead and max_sectors
3512          * values synchronously:
3513          */
3514         spin_lock_irq(q->queue_lock);
3515         /*
3516          * Trim readahead window as well, if necessary:
3517          */
3518         ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3519         if (ra_kb > max_sectors_kb)
3520                 q->backing_dev_info.ra_pages =
3521                                 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3522
3523         q->max_sectors = max_sectors_kb << 1;
3524         spin_unlock_irq(q->queue_lock);
3525
3526         return ret;
3527 }
3528
3529 static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3530 {
3531         int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3532
3533         return queue_var_show(max_hw_sectors_kb, (page));
3534 }
3535
3536
3537 static struct queue_sysfs_entry queue_requests_entry = {
3538         .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3539         .show = queue_requests_show,
3540         .store = queue_requests_store,
3541 };
3542
3543 static struct queue_sysfs_entry queue_ra_entry = {
3544         .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3545         .show = queue_ra_show,
3546         .store = queue_ra_store,
3547 };
3548
3549 static struct queue_sysfs_entry queue_max_sectors_entry = {
3550         .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3551         .show = queue_max_sectors_show,
3552         .store = queue_max_sectors_store,
3553 };
3554
3555 static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3556         .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
3557         .show = queue_max_hw_sectors_show,
3558 };
3559
3560 static struct queue_sysfs_entry queue_iosched_entry = {
3561         .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
3562         .show = elv_iosched_show,
3563         .store = elv_iosched_store,
3564 };
3565
3566 static struct attribute *default_attrs[] = {
3567         &queue_requests_entry.attr,
3568         &queue_ra_entry.attr,
3569         &queue_max_hw_sectors_entry.attr,
3570         &queue_max_sectors_entry.attr,
3571         &queue_iosched_entry.attr,
3572         NULL,
3573 };
3574
3575 #define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
3576
3577 static ssize_t
3578 queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
3579 {
3580         struct queue_sysfs_entry *entry = to_queue(attr);
3581         struct request_queue *q;
3582
3583         q = container_of(kobj, struct request_queue, kobj);
3584         if (!entry->show)
3585                 return 0;
3586
3587         return entry->show(q, page);
3588 }
3589
3590 static ssize_t
3591 queue_attr_store(struct kobject *kobj, struct attribute *attr,
3592                     const char *page, size_t length)
3593 {
3594         struct queue_sysfs_entry *entry = to_queue(attr);
3595         struct request_queue *q;
3596
3597         q = container_of(kobj, struct request_queue, kobj);
3598         if (!entry->store)
3599                 return -EINVAL;
3600
3601         return entry->store(q, page, length);
3602 }
3603
3604 static struct sysfs_ops queue_sysfs_ops = {
3605         .show   = queue_attr_show,
3606         .store  = queue_attr_store,
3607 };
3608
3609 struct kobj_type queue_ktype = {
3610         .sysfs_ops      = &queue_sysfs_ops,
3611         .default_attrs  = default_attrs,
3612 };
3613
3614 int blk_register_queue(struct gendisk *disk)
3615 {
3616         int ret;
3617
3618         request_queue_t *q = disk->queue;
3619
3620         if (!q || !q->request_fn)
3621                 return -ENXIO;
3622
3623         q->kobj.parent = kobject_get(&disk->kobj);
3624         if (!q->kobj.parent)
3625                 return -EBUSY;
3626
3627         snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
3628         q->kobj.ktype = &queue_ktype;
3629
3630         ret = kobject_register(&q->kobj);
3631         if (ret < 0)
3632                 return ret;
3633
3634         ret = elv_register_queue(q);
3635         if (ret) {
3636                 kobject_unregister(&q->kobj);
3637                 return ret;
3638         }
3639
3640         return 0;
3641 }
3642
3643 void blk_unregister_queue(struct gendisk *disk)
3644 {
3645         request_queue_t *q = disk->queue;
3646
3647         if (q && q->request_fn) {
3648                 elv_unregister_queue(q);
3649
3650                 kobject_unregister(&q->kobj);
3651                 kobject_put(&disk->kobj);
3652         }
3653 }