blob: f7c9931cb3803e64590d6998e9645ffff153b6e2 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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>
Christoph Lameter19460892005-06-23 00:08:19 -070031#include <linux/blkdev.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070032
33/*
34 * for max sense size
35 */
36#include <scsi/scsi_cmnd.h>
37
38static void blk_unplug_work(void *data);
39static void blk_unplug_timeout(unsigned long data);
Adrian Bunk93d17d32005-06-25 14:59:10 -070040static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -070041
42/*
43 * For the allocated request tables
44 */
45static kmem_cache_t *request_cachep;
46
47/*
48 * For queue allocation
49 */
50static kmem_cache_t *requestq_cachep;
51
52/*
53 * For io context allocations
54 */
55static kmem_cache_t *iocontext_cachep;
56
57static wait_queue_head_t congestion_wqh[2] = {
58 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
59 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
60 };
61
62/*
63 * Controlling structure to kblockd
64 */
65static struct workqueue_struct *kblockd_workqueue;
66
67unsigned long blk_max_low_pfn, blk_max_pfn;
68
69EXPORT_SYMBOL(blk_max_low_pfn);
70EXPORT_SYMBOL(blk_max_pfn);
71
72/* Amount of time in which a process may batch requests */
73#define BLK_BATCH_TIME (HZ/50UL)
74
75/* Number of requests a "batching" process may submit */
76#define BLK_BATCH_REQ 32
77
78/*
79 * Return the threshold (number of used requests) at which the queue is
80 * considered to be congested. It include a little hysteresis to keep the
81 * context switch rate down.
82 */
83static inline int queue_congestion_on_threshold(struct request_queue *q)
84{
85 return q->nr_congestion_on;
86}
87
88/*
89 * The threshold at which a queue is considered to be uncongested
90 */
91static inline int queue_congestion_off_threshold(struct request_queue *q)
92{
93 return q->nr_congestion_off;
94}
95
96static void blk_queue_congestion_threshold(struct request_queue *q)
97{
98 int nr;
99
100 nr = q->nr_requests - (q->nr_requests / 8) + 1;
101 if (nr > q->nr_requests)
102 nr = q->nr_requests;
103 q->nr_congestion_on = nr;
104
105 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
106 if (nr < 1)
107 nr = 1;
108 q->nr_congestion_off = nr;
109}
110
111/*
112 * A queue has just exitted congestion. Note this in the global counter of
113 * congested queues, and wake up anyone who was waiting for requests to be
114 * put back.
115 */
116static void clear_queue_congested(request_queue_t *q, int rw)
117{
118 enum bdi_state bit;
119 wait_queue_head_t *wqh = &congestion_wqh[rw];
120
121 bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
122 clear_bit(bit, &q->backing_dev_info.state);
123 smp_mb__after_clear_bit();
124 if (waitqueue_active(wqh))
125 wake_up(wqh);
126}
127
128/*
129 * A queue has just entered congestion. Flag that in the queue's VM-visible
130 * state flags and increment the global gounter of congested queues.
131 */
132static void set_queue_congested(request_queue_t *q, int rw)
133{
134 enum bdi_state bit;
135
136 bit = (rw == WRITE) ? BDI_write_congested : BDI_read_congested;
137 set_bit(bit, &q->backing_dev_info.state);
138}
139
140/**
141 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
142 * @bdev: device
143 *
144 * Locates the passed device's request queue and returns the address of its
145 * backing_dev_info
146 *
147 * Will return NULL if the request queue cannot be located.
148 */
149struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
150{
151 struct backing_dev_info *ret = NULL;
152 request_queue_t *q = bdev_get_queue(bdev);
153
154 if (q)
155 ret = &q->backing_dev_info;
156 return ret;
157}
158
159EXPORT_SYMBOL(blk_get_backing_dev_info);
160
161void blk_queue_activity_fn(request_queue_t *q, activity_fn *fn, void *data)
162{
163 q->activity_fn = fn;
164 q->activity_data = data;
165}
166
167EXPORT_SYMBOL(blk_queue_activity_fn);
168
169/**
170 * blk_queue_prep_rq - set a prepare_request function for queue
171 * @q: queue
172 * @pfn: prepare_request function
173 *
174 * It's possible for a queue to register a prepare_request callback which
175 * is invoked before the request is handed to the request_fn. The goal of
176 * the function is to prepare a request for I/O, it can be used to build a
177 * cdb from the request data for instance.
178 *
179 */
180void blk_queue_prep_rq(request_queue_t *q, prep_rq_fn *pfn)
181{
182 q->prep_rq_fn = pfn;
183}
184
185EXPORT_SYMBOL(blk_queue_prep_rq);
186
187/**
188 * blk_queue_merge_bvec - set a merge_bvec function for queue
189 * @q: queue
190 * @mbfn: merge_bvec_fn
191 *
192 * Usually queues have static limitations on the max sectors or segments that
193 * we can put in a request. Stacking drivers may have some settings that
194 * are dynamic, and thus we have to query the queue whether it is ok to
195 * add a new bio_vec to a bio at a given offset or not. If the block device
196 * has such limitations, it needs to register a merge_bvec_fn to control
197 * the size of bio's sent to it. Note that a block device *must* allow a
198 * single page to be added to an empty bio. The block device driver may want
199 * to use the bio_split() function to deal with these bio's. By default
200 * no merge_bvec_fn is defined for a queue, and only the fixed limits are
201 * honored.
202 */
203void blk_queue_merge_bvec(request_queue_t *q, merge_bvec_fn *mbfn)
204{
205 q->merge_bvec_fn = mbfn;
206}
207
208EXPORT_SYMBOL(blk_queue_merge_bvec);
209
210/**
211 * blk_queue_make_request - define an alternate make_request function for a device
212 * @q: the request queue for the device to be affected
213 * @mfn: the alternate make_request function
214 *
215 * Description:
216 * The normal way for &struct bios to be passed to a device
217 * driver is for them to be collected into requests on a request
218 * queue, and then to allow the device driver to select requests
219 * off that queue when it is ready. This works well for many block
220 * devices. However some block devices (typically virtual devices
221 * such as md or lvm) do not benefit from the processing on the
222 * request queue, and are served best by having the requests passed
223 * directly to them. This can be achieved by providing a function
224 * to blk_queue_make_request().
225 *
226 * Caveat:
227 * The driver that does this *must* be able to deal appropriately
228 * with buffers in "highmemory". This can be accomplished by either calling
229 * __bio_kmap_atomic() to get a temporary kernel mapping, or by calling
230 * blk_queue_bounce() to create a buffer in normal memory.
231 **/
232void blk_queue_make_request(request_queue_t * q, make_request_fn * mfn)
233{
234 /*
235 * set defaults
236 */
237 q->nr_requests = BLKDEV_MAX_RQ;
Stuart McLaren309c0a12005-09-06 15:17:47 -0700238 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
239 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240 q->make_request_fn = mfn;
241 q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
242 q->backing_dev_info.state = 0;
243 q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
244 blk_queue_max_sectors(q, MAX_SECTORS);
245 blk_queue_hardsect_size(q, 512);
246 blk_queue_dma_alignment(q, 511);
247 blk_queue_congestion_threshold(q);
248 q->nr_batching = BLK_BATCH_REQ;
249
250 q->unplug_thresh = 4; /* hmm */
251 q->unplug_delay = (3 * HZ) / 1000; /* 3 milliseconds */
252 if (q->unplug_delay == 0)
253 q->unplug_delay = 1;
254
255 INIT_WORK(&q->unplug_work, blk_unplug_work, q);
256
257 q->unplug_timer.function = blk_unplug_timeout;
258 q->unplug_timer.data = (unsigned long)q;
259
260 /*
261 * by default assume old behaviour and bounce for any highmem page
262 */
263 blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
264
265 blk_queue_activity_fn(q, NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266}
267
268EXPORT_SYMBOL(blk_queue_make_request);
269
270static 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;
Jens Axboe22e2c502005-06-27 10:55:12 +0200277 rq->ioprio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 rq->buffer = NULL;
279 rq->ref_count = 1;
280 rq->q = q;
281 rq->waiting = NULL;
282 rq->special = NULL;
283 rq->data_len = 0;
284 rq->data = NULL;
Mike Christie df46b9a2005-06-20 14:04:44 +0200285 rq->nr_phys_segments = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286 rq->sense = NULL;
287 rq->end_io = NULL;
288 rq->end_io_data = NULL;
289}
290
291/**
292 * blk_queue_ordered - does this queue support ordered writes
293 * @q: the request queue
294 * @flag: see below
295 *
296 * Description:
297 * For journalled file systems, doing ordered writes on a commit
298 * block instead of explicitly doing wait_on_buffer (which is bad
299 * for performance) can be a big win. Block drivers supporting this
300 * feature should call this function and indicate so.
301 *
302 **/
303void blk_queue_ordered(request_queue_t *q, int flag)
304{
305 switch (flag) {
306 case QUEUE_ORDERED_NONE:
307 if (q->flush_rq)
308 kmem_cache_free(request_cachep, q->flush_rq);
309 q->flush_rq = NULL;
310 q->ordered = flag;
311 break;
312 case QUEUE_ORDERED_TAG:
313 q->ordered = flag;
314 break;
315 case QUEUE_ORDERED_FLUSH:
316 q->ordered = flag;
317 if (!q->flush_rq)
318 q->flush_rq = kmem_cache_alloc(request_cachep,
319 GFP_KERNEL);
320 break;
321 default:
322 printk("blk_queue_ordered: bad value %d\n", flag);
323 break;
324 }
325}
326
327EXPORT_SYMBOL(blk_queue_ordered);
328
329/**
330 * blk_queue_issue_flush_fn - set function for issuing a flush
331 * @q: the request queue
332 * @iff: the function to be called issuing the flush
333 *
334 * Description:
335 * If a driver supports issuing a flush command, the support is notified
336 * to the block layer by defining it through this call.
337 *
338 **/
339void blk_queue_issue_flush_fn(request_queue_t *q, issue_flush_fn *iff)
340{
341 q->issue_flush_fn = iff;
342}
343
344EXPORT_SYMBOL(blk_queue_issue_flush_fn);
345
346/*
347 * Cache flushing for ordered writes handling
348 */
349static void blk_pre_flush_end_io(struct request *flush_rq)
350{
351 struct request *rq = flush_rq->end_io_data;
352 request_queue_t *q = rq->q;
353
Tejun Heo8922e162005-10-20 16:23:44 +0200354 elv_completed_request(q, flush_rq);
355
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356 rq->flags |= REQ_BAR_PREFLUSH;
357
358 if (!flush_rq->errors)
359 elv_requeue_request(q, rq);
360 else {
361 q->end_flush_fn(q, flush_rq);
362 clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
363 q->request_fn(q);
364 }
365}
366
367static void blk_post_flush_end_io(struct request *flush_rq)
368{
369 struct request *rq = flush_rq->end_io_data;
370 request_queue_t *q = rq->q;
371
Tejun Heo8922e162005-10-20 16:23:44 +0200372 elv_completed_request(q, flush_rq);
373
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 rq->flags |= REQ_BAR_POSTFLUSH;
375
376 q->end_flush_fn(q, flush_rq);
377 clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
378 q->request_fn(q);
379}
380
381struct request *blk_start_pre_flush(request_queue_t *q, struct request *rq)
382{
383 struct request *flush_rq = q->flush_rq;
384
385 BUG_ON(!blk_barrier_rq(rq));
386
387 if (test_and_set_bit(QUEUE_FLAG_FLUSH, &q->queue_flags))
388 return NULL;
389
390 rq_init(q, flush_rq);
391 flush_rq->elevator_private = NULL;
392 flush_rq->flags = REQ_BAR_FLUSH;
393 flush_rq->rq_disk = rq->rq_disk;
394 flush_rq->rl = NULL;
395
396 /*
397 * prepare_flush returns 0 if no flush is needed, just mark both
398 * pre and post flush as done in that case
399 */
400 if (!q->prepare_flush_fn(q, flush_rq)) {
401 rq->flags |= REQ_BAR_PREFLUSH | REQ_BAR_POSTFLUSH;
402 clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
403 return rq;
404 }
405
406 /*
407 * some drivers dequeue requests right away, some only after io
408 * completion. make sure the request is dequeued.
409 */
410 if (!list_empty(&rq->queuelist))
411 blkdev_dequeue_request(rq);
412
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413 flush_rq->end_io_data = rq;
414 flush_rq->end_io = blk_pre_flush_end_io;
415
416 __elv_add_request(q, flush_rq, ELEVATOR_INSERT_FRONT, 0);
417 return flush_rq;
418}
419
420static void blk_start_post_flush(request_queue_t *q, struct request *rq)
421{
422 struct request *flush_rq = q->flush_rq;
423
424 BUG_ON(!blk_barrier_rq(rq));
425
426 rq_init(q, flush_rq);
427 flush_rq->elevator_private = NULL;
428 flush_rq->flags = REQ_BAR_FLUSH;
429 flush_rq->rq_disk = rq->rq_disk;
430 flush_rq->rl = NULL;
431
432 if (q->prepare_flush_fn(q, flush_rq)) {
433 flush_rq->end_io_data = rq;
434 flush_rq->end_io = blk_post_flush_end_io;
435
436 __elv_add_request(q, flush_rq, ELEVATOR_INSERT_FRONT, 0);
437 q->request_fn(q);
438 }
439}
440
441static inline int blk_check_end_barrier(request_queue_t *q, struct request *rq,
442 int sectors)
443{
444 if (sectors > rq->nr_sectors)
445 sectors = rq->nr_sectors;
446
447 rq->nr_sectors -= sectors;
448 return rq->nr_sectors;
449}
450
451static int __blk_complete_barrier_rq(request_queue_t *q, struct request *rq,
452 int sectors, int queue_locked)
453{
454 if (q->ordered != QUEUE_ORDERED_FLUSH)
455 return 0;
456 if (!blk_fs_request(rq) || !blk_barrier_rq(rq))
457 return 0;
458 if (blk_barrier_postflush(rq))
459 return 0;
460
461 if (!blk_check_end_barrier(q, rq, sectors)) {
462 unsigned long flags = 0;
463
464 if (!queue_locked)
465 spin_lock_irqsave(q->queue_lock, flags);
466
467 blk_start_post_flush(q, rq);
468
469 if (!queue_locked)
470 spin_unlock_irqrestore(q->queue_lock, flags);
471 }
472
473 return 1;
474}
475
476/**
477 * blk_complete_barrier_rq - complete possible barrier request
478 * @q: the request queue for the device
479 * @rq: the request
480 * @sectors: number of sectors to complete
481 *
482 * Description:
483 * Used in driver end_io handling to determine whether to postpone
484 * completion of a barrier request until a post flush has been done. This
485 * is the unlocked variant, used if the caller doesn't already hold the
486 * queue lock.
487 **/
488int blk_complete_barrier_rq(request_queue_t *q, struct request *rq, int sectors)
489{
490 return __blk_complete_barrier_rq(q, rq, sectors, 0);
491}
492EXPORT_SYMBOL(blk_complete_barrier_rq);
493
494/**
495 * blk_complete_barrier_rq_locked - complete possible barrier request
496 * @q: the request queue for the device
497 * @rq: the request
498 * @sectors: number of sectors to complete
499 *
500 * Description:
501 * See blk_complete_barrier_rq(). This variant must be used if the caller
502 * holds the queue lock.
503 **/
504int blk_complete_barrier_rq_locked(request_queue_t *q, struct request *rq,
505 int sectors)
506{
507 return __blk_complete_barrier_rq(q, rq, sectors, 1);
508}
509EXPORT_SYMBOL(blk_complete_barrier_rq_locked);
510
511/**
512 * blk_queue_bounce_limit - set bounce buffer limit for queue
513 * @q: the request queue for the device
514 * @dma_addr: bus address limit
515 *
516 * Description:
517 * Different hardware can have different requirements as to what pages
518 * it can do I/O directly to. A low level driver can call
519 * blk_queue_bounce_limit to have lower memory pages allocated as bounce
520 * buffers for doing I/O to pages residing above @page. By default
521 * the block layer sets this to the highest numbered "low" memory page.
522 **/
523void blk_queue_bounce_limit(request_queue_t *q, u64 dma_addr)
524{
525 unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
526
527 /*
528 * set appropriate bounce gfp mask -- unfortunately we don't have a
529 * full 4GB zone, so we have to resort to low memory for any bounces.
530 * ISA has its own < 16MB zone.
531 */
532 if (bounce_pfn < blk_max_low_pfn) {
533 BUG_ON(dma_addr < BLK_BOUNCE_ISA);
534 init_emergency_isa_pool();
535 q->bounce_gfp = GFP_NOIO | GFP_DMA;
536 } else
537 q->bounce_gfp = GFP_NOIO;
538
539 q->bounce_pfn = bounce_pfn;
540}
541
542EXPORT_SYMBOL(blk_queue_bounce_limit);
543
544/**
545 * blk_queue_max_sectors - set max sectors for a request for this queue
546 * @q: the request queue for the device
547 * @max_sectors: max sectors in the usual 512b unit
548 *
549 * Description:
550 * Enables a low level driver to set an upper limit on the size of
551 * received requests.
552 **/
553void blk_queue_max_sectors(request_queue_t *q, unsigned short max_sectors)
554{
555 if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
556 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
557 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
558 }
559
560 q->max_sectors = q->max_hw_sectors = max_sectors;
561}
562
563EXPORT_SYMBOL(blk_queue_max_sectors);
564
565/**
566 * blk_queue_max_phys_segments - set max phys segments for a request for this queue
567 * @q: the request queue for the device
568 * @max_segments: max number of segments
569 *
570 * Description:
571 * Enables a low level driver to set an upper limit on the number of
572 * physical data segments in a request. This would be the largest sized
573 * scatter list the driver could handle.
574 **/
575void blk_queue_max_phys_segments(request_queue_t *q, unsigned short max_segments)
576{
577 if (!max_segments) {
578 max_segments = 1;
579 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
580 }
581
582 q->max_phys_segments = max_segments;
583}
584
585EXPORT_SYMBOL(blk_queue_max_phys_segments);
586
587/**
588 * blk_queue_max_hw_segments - set max hw segments for a request for this queue
589 * @q: the request queue for the device
590 * @max_segments: max number of segments
591 *
592 * Description:
593 * Enables a low level driver to set an upper limit on the number of
594 * hw data segments in a request. This would be the largest number of
595 * address/length pairs the host adapter can actually give as once
596 * to the device.
597 **/
598void blk_queue_max_hw_segments(request_queue_t *q, unsigned short max_segments)
599{
600 if (!max_segments) {
601 max_segments = 1;
602 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
603 }
604
605 q->max_hw_segments = max_segments;
606}
607
608EXPORT_SYMBOL(blk_queue_max_hw_segments);
609
610/**
611 * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
612 * @q: the request queue for the device
613 * @max_size: max size of segment in bytes
614 *
615 * Description:
616 * Enables a low level driver to set an upper limit on the size of a
617 * coalesced segment
618 **/
619void blk_queue_max_segment_size(request_queue_t *q, unsigned int max_size)
620{
621 if (max_size < PAGE_CACHE_SIZE) {
622 max_size = PAGE_CACHE_SIZE;
623 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
624 }
625
626 q->max_segment_size = max_size;
627}
628
629EXPORT_SYMBOL(blk_queue_max_segment_size);
630
631/**
632 * blk_queue_hardsect_size - set hardware sector size for the queue
633 * @q: the request queue for the device
634 * @size: the hardware sector size, in bytes
635 *
636 * Description:
637 * This should typically be set to the lowest possible sector size
638 * that the hardware can operate on (possible without reverting to
639 * even internal read-modify-write operations). Usually the default
640 * of 512 covers most hardware.
641 **/
642void blk_queue_hardsect_size(request_queue_t *q, unsigned short size)
643{
644 q->hardsect_size = size;
645}
646
647EXPORT_SYMBOL(blk_queue_hardsect_size);
648
649/*
650 * Returns the minimum that is _not_ zero, unless both are zero.
651 */
652#define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
653
654/**
655 * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
656 * @t: the stacking driver (top)
657 * @b: the underlying device (bottom)
658 **/
659void blk_queue_stack_limits(request_queue_t *t, request_queue_t *b)
660{
661 /* zero is "infinity" */
662 t->max_sectors = t->max_hw_sectors =
663 min_not_zero(t->max_sectors,b->max_sectors);
664
665 t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
666 t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
667 t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
668 t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
669}
670
671EXPORT_SYMBOL(blk_queue_stack_limits);
672
673/**
674 * blk_queue_segment_boundary - set boundary rules for segment merging
675 * @q: the request queue for the device
676 * @mask: the memory boundary mask
677 **/
678void blk_queue_segment_boundary(request_queue_t *q, unsigned long mask)
679{
680 if (mask < PAGE_CACHE_SIZE - 1) {
681 mask = PAGE_CACHE_SIZE - 1;
682 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
683 }
684
685 q->seg_boundary_mask = mask;
686}
687
688EXPORT_SYMBOL(blk_queue_segment_boundary);
689
690/**
691 * blk_queue_dma_alignment - set dma length and memory alignment
692 * @q: the request queue for the device
693 * @mask: alignment mask
694 *
695 * description:
696 * set required memory and length aligment for direct dma transactions.
697 * this is used when buiding direct io requests for the queue.
698 *
699 **/
700void blk_queue_dma_alignment(request_queue_t *q, int mask)
701{
702 q->dma_alignment = mask;
703}
704
705EXPORT_SYMBOL(blk_queue_dma_alignment);
706
707/**
708 * blk_queue_find_tag - find a request by its tag and queue
709 *
710 * @q: The request queue for the device
711 * @tag: The tag of the request
712 *
713 * Notes:
714 * Should be used when a device returns a tag and you want to match
715 * it with a request.
716 *
717 * no locks need be held.
718 **/
719struct request *blk_queue_find_tag(request_queue_t *q, int tag)
720{
721 struct blk_queue_tag *bqt = q->queue_tags;
722
Tejun Heoba025082005-08-05 13:28:11 -0700723 if (unlikely(bqt == NULL || tag >= bqt->real_max_depth))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724 return NULL;
725
726 return bqt->tag_index[tag];
727}
728
729EXPORT_SYMBOL(blk_queue_find_tag);
730
731/**
732 * __blk_queue_free_tags - release tag maintenance info
733 * @q: the request queue for the device
734 *
735 * Notes:
736 * blk_cleanup_queue() will take care of calling this function, if tagging
737 * has been used. So there's no need to call this directly.
738 **/
739static void __blk_queue_free_tags(request_queue_t *q)
740{
741 struct blk_queue_tag *bqt = q->queue_tags;
742
743 if (!bqt)
744 return;
745
746 if (atomic_dec_and_test(&bqt->refcnt)) {
747 BUG_ON(bqt->busy);
748 BUG_ON(!list_empty(&bqt->busy_list));
749
750 kfree(bqt->tag_index);
751 bqt->tag_index = NULL;
752
753 kfree(bqt->tag_map);
754 bqt->tag_map = NULL;
755
756 kfree(bqt);
757 }
758
759 q->queue_tags = NULL;
760 q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
761}
762
763/**
764 * blk_queue_free_tags - release tag maintenance info
765 * @q: the request queue for the device
766 *
767 * Notes:
768 * This is used to disabled tagged queuing to a device, yet leave
769 * queue in function.
770 **/
771void blk_queue_free_tags(request_queue_t *q)
772{
773 clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
774}
775
776EXPORT_SYMBOL(blk_queue_free_tags);
777
778static int
779init_tag_map(request_queue_t *q, struct blk_queue_tag *tags, int depth)
780{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781 struct request **tag_index;
782 unsigned long *tag_map;
Tejun Heofa72b902005-06-23 00:08:49 -0700783 int nr_ulongs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784
785 if (depth > q->nr_requests * 2) {
786 depth = q->nr_requests * 2;
787 printk(KERN_ERR "%s: adjusted depth to %d\n",
788 __FUNCTION__, depth);
789 }
790
791 tag_index = kmalloc(depth * sizeof(struct request *), GFP_ATOMIC);
792 if (!tag_index)
793 goto fail;
794
Tejun Heof7d37d02005-06-23 00:08:50 -0700795 nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
Tejun Heofa72b902005-06-23 00:08:49 -0700796 tag_map = kmalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797 if (!tag_map)
798 goto fail;
799
800 memset(tag_index, 0, depth * sizeof(struct request *));
Tejun Heofa72b902005-06-23 00:08:49 -0700801 memset(tag_map, 0, nr_ulongs * sizeof(unsigned long));
Tejun Heoba025082005-08-05 13:28:11 -0700802 tags->real_max_depth = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 tags->max_depth = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804 tags->tag_index = tag_index;
805 tags->tag_map = tag_map;
806
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 return 0;
808fail:
809 kfree(tag_index);
810 return -ENOMEM;
811}
812
813/**
814 * blk_queue_init_tags - initialize the queue tag info
815 * @q: the request queue for the device
816 * @depth: the maximum queue depth supported
817 * @tags: the tag to use
818 **/
819int blk_queue_init_tags(request_queue_t *q, int depth,
820 struct blk_queue_tag *tags)
821{
822 int rc;
823
824 BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
825
826 if (!tags && !q->queue_tags) {
827 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
828 if (!tags)
829 goto fail;
830
831 if (init_tag_map(q, tags, depth))
832 goto fail;
833
834 INIT_LIST_HEAD(&tags->busy_list);
835 tags->busy = 0;
836 atomic_set(&tags->refcnt, 1);
837 } else if (q->queue_tags) {
838 if ((rc = blk_queue_resize_tags(q, depth)))
839 return rc;
840 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
841 return 0;
842 } else
843 atomic_inc(&tags->refcnt);
844
845 /*
846 * assign it, all done
847 */
848 q->queue_tags = tags;
849 q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
850 return 0;
851fail:
852 kfree(tags);
853 return -ENOMEM;
854}
855
856EXPORT_SYMBOL(blk_queue_init_tags);
857
858/**
859 * blk_queue_resize_tags - change the queueing depth
860 * @q: the request queue for the device
861 * @new_depth: the new max command queueing depth
862 *
863 * Notes:
864 * Must be called with the queue lock held.
865 **/
866int blk_queue_resize_tags(request_queue_t *q, int new_depth)
867{
868 struct blk_queue_tag *bqt = q->queue_tags;
869 struct request **tag_index;
870 unsigned long *tag_map;
Tejun Heofa72b902005-06-23 00:08:49 -0700871 int max_depth, nr_ulongs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872
873 if (!bqt)
874 return -ENXIO;
875
876 /*
Tejun Heoba025082005-08-05 13:28:11 -0700877 * if we already have large enough real_max_depth. just
878 * adjust max_depth. *NOTE* as requests with tag value
879 * between new_depth and real_max_depth can be in-flight, tag
880 * map can not be shrunk blindly here.
881 */
882 if (new_depth <= bqt->real_max_depth) {
883 bqt->max_depth = new_depth;
884 return 0;
885 }
886
887 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 * save the old state info, so we can copy it back
889 */
890 tag_index = bqt->tag_index;
891 tag_map = bqt->tag_map;
Tejun Heoba025082005-08-05 13:28:11 -0700892 max_depth = bqt->real_max_depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893
894 if (init_tag_map(q, bqt, new_depth))
895 return -ENOMEM;
896
897 memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
Tejun Heof7d37d02005-06-23 00:08:50 -0700898 nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
Tejun Heofa72b902005-06-23 00:08:49 -0700899 memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900
901 kfree(tag_index);
902 kfree(tag_map);
903 return 0;
904}
905
906EXPORT_SYMBOL(blk_queue_resize_tags);
907
908/**
909 * blk_queue_end_tag - end tag operations for a request
910 * @q: the request queue for the device
911 * @rq: the request that has completed
912 *
913 * Description:
914 * Typically called when end_that_request_first() returns 0, meaning
915 * all transfers have been done for a request. It's important to call
916 * this function before end_that_request_last(), as that will put the
917 * request back on the free list thus corrupting the internal tag list.
918 *
919 * Notes:
920 * queue lock must be held.
921 **/
922void blk_queue_end_tag(request_queue_t *q, struct request *rq)
923{
924 struct blk_queue_tag *bqt = q->queue_tags;
925 int tag = rq->tag;
926
927 BUG_ON(tag == -1);
928
Tejun Heoba025082005-08-05 13:28:11 -0700929 if (unlikely(tag >= bqt->real_max_depth))
Tejun Heo040c9282005-06-23 00:08:51 -0700930 /*
931 * This can happen after tag depth has been reduced.
932 * FIXME: how about a warning or info message here?
933 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934 return;
935
936 if (unlikely(!__test_and_clear_bit(tag, bqt->tag_map))) {
Tejun Heo040c9282005-06-23 00:08:51 -0700937 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
938 __FUNCTION__, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 return;
940 }
941
942 list_del_init(&rq->queuelist);
943 rq->flags &= ~REQ_QUEUED;
944 rq->tag = -1;
945
946 if (unlikely(bqt->tag_index[tag] == NULL))
Tejun Heo040c9282005-06-23 00:08:51 -0700947 printk(KERN_ERR "%s: tag %d is missing\n",
948 __FUNCTION__, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949
950 bqt->tag_index[tag] = NULL;
951 bqt->busy--;
952}
953
954EXPORT_SYMBOL(blk_queue_end_tag);
955
956/**
957 * blk_queue_start_tag - find a free tag and assign it
958 * @q: the request queue for the device
959 * @rq: the block request that needs tagging
960 *
961 * Description:
962 * This can either be used as a stand-alone helper, or possibly be
963 * assigned as the queue &prep_rq_fn (in which case &struct request
964 * automagically gets a tag assigned). Note that this function
965 * assumes that any type of request can be queued! if this is not
966 * true for your device, you must check the request type before
967 * calling this function. The request will also be removed from
968 * the request queue, so it's the drivers responsibility to readd
969 * it if it should need to be restarted for some reason.
970 *
971 * Notes:
972 * queue lock must be held.
973 **/
974int blk_queue_start_tag(request_queue_t *q, struct request *rq)
975{
976 struct blk_queue_tag *bqt = q->queue_tags;
Tejun Heo2bf0fda2005-06-23 00:08:48 -0700977 int tag;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978
979 if (unlikely((rq->flags & REQ_QUEUED))) {
980 printk(KERN_ERR
Tejun Heo040c9282005-06-23 00:08:51 -0700981 "%s: request %p for device [%s] already tagged %d",
982 __FUNCTION__, rq,
983 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984 BUG();
985 }
986
Tejun Heo2bf0fda2005-06-23 00:08:48 -0700987 tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
988 if (tag >= bqt->max_depth)
989 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991 __set_bit(tag, bqt->tag_map);
992
993 rq->flags |= REQ_QUEUED;
994 rq->tag = tag;
995 bqt->tag_index[tag] = rq;
996 blkdev_dequeue_request(rq);
997 list_add(&rq->queuelist, &bqt->busy_list);
998 bqt->busy++;
999 return 0;
1000}
1001
1002EXPORT_SYMBOL(blk_queue_start_tag);
1003
1004/**
1005 * blk_queue_invalidate_tags - invalidate all pending tags
1006 * @q: the request queue for the device
1007 *
1008 * Description:
1009 * Hardware conditions may dictate a need to stop all pending requests.
1010 * In this case, we will safely clear the block side of the tag queue and
1011 * readd all requests to the request queue in the right order.
1012 *
1013 * Notes:
1014 * queue lock must be held.
1015 **/
1016void blk_queue_invalidate_tags(request_queue_t *q)
1017{
1018 struct blk_queue_tag *bqt = q->queue_tags;
1019 struct list_head *tmp, *n;
1020 struct request *rq;
1021
1022 list_for_each_safe(tmp, n, &bqt->busy_list) {
1023 rq = list_entry_rq(tmp);
1024
1025 if (rq->tag == -1) {
Tejun Heo040c9282005-06-23 00:08:51 -07001026 printk(KERN_ERR
1027 "%s: bad tag found on list\n", __FUNCTION__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028 list_del_init(&rq->queuelist);
1029 rq->flags &= ~REQ_QUEUED;
1030 } else
1031 blk_queue_end_tag(q, rq);
1032
1033 rq->flags &= ~REQ_STARTED;
1034 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1035 }
1036}
1037
1038EXPORT_SYMBOL(blk_queue_invalidate_tags);
1039
1040static char *rq_flags[] = {
1041 "REQ_RW",
1042 "REQ_FAILFAST",
Tejun Heo8922e162005-10-20 16:23:44 +02001043 "REQ_SORTED",
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044 "REQ_SOFTBARRIER",
1045 "REQ_HARDBARRIER",
1046 "REQ_CMD",
1047 "REQ_NOMERGE",
1048 "REQ_STARTED",
1049 "REQ_DONTPREP",
1050 "REQ_QUEUED",
Tejun Heocb98fc82005-10-28 08:29:39 +02001051 "REQ_ELVPRIV",
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 "REQ_PC",
1053 "REQ_BLOCK_PC",
1054 "REQ_SENSE",
1055 "REQ_FAILED",
1056 "REQ_QUIET",
1057 "REQ_SPECIAL",
1058 "REQ_DRIVE_CMD",
1059 "REQ_DRIVE_TASK",
1060 "REQ_DRIVE_TASKFILE",
1061 "REQ_PREEMPT",
1062 "REQ_PM_SUSPEND",
1063 "REQ_PM_RESUME",
1064 "REQ_PM_SHUTDOWN",
1065};
1066
1067void blk_dump_rq_flags(struct request *rq, char *msg)
1068{
1069 int bit;
1070
1071 printk("%s: dev %s: flags = ", msg,
1072 rq->rq_disk ? rq->rq_disk->disk_name : "?");
1073 bit = 0;
1074 do {
1075 if (rq->flags & (1 << bit))
1076 printk("%s ", rq_flags[bit]);
1077 bit++;
1078 } while (bit < __REQ_NR_BITS);
1079
1080 printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1081 rq->nr_sectors,
1082 rq->current_nr_sectors);
1083 printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1084
1085 if (rq->flags & (REQ_BLOCK_PC | REQ_PC)) {
1086 printk("cdb: ");
1087 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1088 printk("%02x ", rq->cmd[bit]);
1089 printk("\n");
1090 }
1091}
1092
1093EXPORT_SYMBOL(blk_dump_rq_flags);
1094
1095void blk_recount_segments(request_queue_t *q, struct bio *bio)
1096{
1097 struct bio_vec *bv, *bvprv = NULL;
1098 int i, nr_phys_segs, nr_hw_segs, seg_size, hw_seg_size, cluster;
1099 int high, highprv = 1;
1100
1101 if (unlikely(!bio->bi_io_vec))
1102 return;
1103
1104 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1105 hw_seg_size = seg_size = nr_phys_segs = nr_hw_segs = 0;
1106 bio_for_each_segment(bv, bio, i) {
1107 /*
1108 * the trick here is making sure that a high page is never
1109 * considered part of another segment, since that might
1110 * change with the bounce page.
1111 */
1112 high = page_to_pfn(bv->bv_page) >= q->bounce_pfn;
1113 if (high || highprv)
1114 goto new_hw_segment;
1115 if (cluster) {
1116 if (seg_size + bv->bv_len > q->max_segment_size)
1117 goto new_segment;
1118 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1119 goto new_segment;
1120 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1121 goto new_segment;
1122 if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1123 goto new_hw_segment;
1124
1125 seg_size += bv->bv_len;
1126 hw_seg_size += bv->bv_len;
1127 bvprv = bv;
1128 continue;
1129 }
1130new_segment:
1131 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1132 !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len)) {
1133 hw_seg_size += bv->bv_len;
1134 } else {
1135new_hw_segment:
1136 if (hw_seg_size > bio->bi_hw_front_size)
1137 bio->bi_hw_front_size = hw_seg_size;
1138 hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1139 nr_hw_segs++;
1140 }
1141
1142 nr_phys_segs++;
1143 bvprv = bv;
1144 seg_size = bv->bv_len;
1145 highprv = high;
1146 }
1147 if (hw_seg_size > bio->bi_hw_back_size)
1148 bio->bi_hw_back_size = hw_seg_size;
1149 if (nr_hw_segs == 1 && hw_seg_size > bio->bi_hw_front_size)
1150 bio->bi_hw_front_size = hw_seg_size;
1151 bio->bi_phys_segments = nr_phys_segs;
1152 bio->bi_hw_segments = nr_hw_segs;
1153 bio->bi_flags |= (1 << BIO_SEG_VALID);
1154}
1155
1156
Adrian Bunk93d17d32005-06-25 14:59:10 -07001157static int blk_phys_contig_segment(request_queue_t *q, struct bio *bio,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 struct bio *nxt)
1159{
1160 if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1161 return 0;
1162
1163 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1164 return 0;
1165 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1166 return 0;
1167
1168 /*
1169 * bio and nxt are contigous in memory, check if the queue allows
1170 * these two to be merged into one
1171 */
1172 if (BIO_SEG_BOUNDARY(q, bio, nxt))
1173 return 1;
1174
1175 return 0;
1176}
1177
Adrian Bunk93d17d32005-06-25 14:59:10 -07001178static int blk_hw_contig_segment(request_queue_t *q, struct bio *bio,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 struct bio *nxt)
1180{
1181 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1182 blk_recount_segments(q, bio);
1183 if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1184 blk_recount_segments(q, nxt);
1185 if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1186 BIOVEC_VIRT_OVERSIZE(bio->bi_hw_front_size + bio->bi_hw_back_size))
1187 return 0;
1188 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1189 return 0;
1190
1191 return 1;
1192}
1193
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194/*
1195 * map a request to scatterlist, return number of sg entries setup. Caller
1196 * must make sure sg can hold rq->nr_phys_segments entries
1197 */
1198int blk_rq_map_sg(request_queue_t *q, struct request *rq, struct scatterlist *sg)
1199{
1200 struct bio_vec *bvec, *bvprv;
1201 struct bio *bio;
1202 int nsegs, i, cluster;
1203
1204 nsegs = 0;
1205 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1206
1207 /*
1208 * for each bio in rq
1209 */
1210 bvprv = NULL;
1211 rq_for_each_bio(bio, rq) {
1212 /*
1213 * for each segment in bio
1214 */
1215 bio_for_each_segment(bvec, bio, i) {
1216 int nbytes = bvec->bv_len;
1217
1218 if (bvprv && cluster) {
1219 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1220 goto new_segment;
1221
1222 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1223 goto new_segment;
1224 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1225 goto new_segment;
1226
1227 sg[nsegs - 1].length += nbytes;
1228 } else {
1229new_segment:
1230 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1231 sg[nsegs].page = bvec->bv_page;
1232 sg[nsegs].length = nbytes;
1233 sg[nsegs].offset = bvec->bv_offset;
1234
1235 nsegs++;
1236 }
1237 bvprv = bvec;
1238 } /* segments in bio */
1239 } /* bios in rq */
1240
1241 return nsegs;
1242}
1243
1244EXPORT_SYMBOL(blk_rq_map_sg);
1245
1246/*
1247 * the standard queue merge functions, can be overridden with device
1248 * specific ones if so desired
1249 */
1250
1251static inline int ll_new_mergeable(request_queue_t *q,
1252 struct request *req,
1253 struct bio *bio)
1254{
1255 int nr_phys_segs = bio_phys_segments(q, bio);
1256
1257 if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1258 req->flags |= REQ_NOMERGE;
1259 if (req == q->last_merge)
1260 q->last_merge = NULL;
1261 return 0;
1262 }
1263
1264 /*
1265 * A hw segment is just getting larger, bump just the phys
1266 * counter.
1267 */
1268 req->nr_phys_segments += nr_phys_segs;
1269 return 1;
1270}
1271
1272static inline int ll_new_hw_segment(request_queue_t *q,
1273 struct request *req,
1274 struct bio *bio)
1275{
1276 int nr_hw_segs = bio_hw_segments(q, bio);
1277 int nr_phys_segs = bio_phys_segments(q, bio);
1278
1279 if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1280 || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1281 req->flags |= REQ_NOMERGE;
1282 if (req == q->last_merge)
1283 q->last_merge = NULL;
1284 return 0;
1285 }
1286
1287 /*
1288 * This will form the start of a new hw segment. Bump both
1289 * counters.
1290 */
1291 req->nr_hw_segments += nr_hw_segs;
1292 req->nr_phys_segments += nr_phys_segs;
1293 return 1;
1294}
1295
1296static int ll_back_merge_fn(request_queue_t *q, struct request *req,
1297 struct bio *bio)
1298{
1299 int len;
1300
1301 if (req->nr_sectors + bio_sectors(bio) > q->max_sectors) {
1302 req->flags |= REQ_NOMERGE;
1303 if (req == q->last_merge)
1304 q->last_merge = NULL;
1305 return 0;
1306 }
1307 if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1308 blk_recount_segments(q, req->biotail);
1309 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1310 blk_recount_segments(q, bio);
1311 len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1312 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1313 !BIOVEC_VIRT_OVERSIZE(len)) {
1314 int mergeable = ll_new_mergeable(q, req, bio);
1315
1316 if (mergeable) {
1317 if (req->nr_hw_segments == 1)
1318 req->bio->bi_hw_front_size = len;
1319 if (bio->bi_hw_segments == 1)
1320 bio->bi_hw_back_size = len;
1321 }
1322 return mergeable;
1323 }
1324
1325 return ll_new_hw_segment(q, req, bio);
1326}
1327
1328static int ll_front_merge_fn(request_queue_t *q, struct request *req,
1329 struct bio *bio)
1330{
1331 int len;
1332
1333 if (req->nr_sectors + bio_sectors(bio) > q->max_sectors) {
1334 req->flags |= REQ_NOMERGE;
1335 if (req == q->last_merge)
1336 q->last_merge = NULL;
1337 return 0;
1338 }
1339 len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1340 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1341 blk_recount_segments(q, bio);
1342 if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1343 blk_recount_segments(q, req->bio);
1344 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1345 !BIOVEC_VIRT_OVERSIZE(len)) {
1346 int mergeable = ll_new_mergeable(q, req, bio);
1347
1348 if (mergeable) {
1349 if (bio->bi_hw_segments == 1)
1350 bio->bi_hw_front_size = len;
1351 if (req->nr_hw_segments == 1)
1352 req->biotail->bi_hw_back_size = len;
1353 }
1354 return mergeable;
1355 }
1356
1357 return ll_new_hw_segment(q, req, bio);
1358}
1359
1360static int ll_merge_requests_fn(request_queue_t *q, struct request *req,
1361 struct request *next)
1362{
Nikita Danilovdfa1a552005-06-25 14:59:20 -07001363 int total_phys_segments;
1364 int total_hw_segments;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365
1366 /*
1367 * First check if the either of the requests are re-queued
1368 * requests. Can't merge them if they are.
1369 */
1370 if (req->special || next->special)
1371 return 0;
1372
1373 /*
Nikita Danilovdfa1a552005-06-25 14:59:20 -07001374 * Will it become too large?
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375 */
1376 if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1377 return 0;
1378
1379 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1380 if (blk_phys_contig_segment(q, req->biotail, next->bio))
1381 total_phys_segments--;
1382
1383 if (total_phys_segments > q->max_phys_segments)
1384 return 0;
1385
1386 total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1387 if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1388 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1389 /*
1390 * propagate the combined length to the end of the requests
1391 */
1392 if (req->nr_hw_segments == 1)
1393 req->bio->bi_hw_front_size = len;
1394 if (next->nr_hw_segments == 1)
1395 next->biotail->bi_hw_back_size = len;
1396 total_hw_segments--;
1397 }
1398
1399 if (total_hw_segments > q->max_hw_segments)
1400 return 0;
1401
1402 /* Merge is OK... */
1403 req->nr_phys_segments = total_phys_segments;
1404 req->nr_hw_segments = total_hw_segments;
1405 return 1;
1406}
1407
1408/*
1409 * "plug" the device if there are no outstanding requests: this will
1410 * force the transfer to start only after we have put all the requests
1411 * on the list.
1412 *
1413 * This is called with interrupts off and no requests on the queue and
1414 * with the queue lock held.
1415 */
1416void blk_plug_device(request_queue_t *q)
1417{
1418 WARN_ON(!irqs_disabled());
1419
1420 /*
1421 * don't plug a stopped queue, it must be paired with blk_start_queue()
1422 * which will restart the queueing
1423 */
1424 if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
1425 return;
1426
1427 if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1428 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1429}
1430
1431EXPORT_SYMBOL(blk_plug_device);
1432
1433/*
1434 * remove the queue from the plugged list, if present. called with
1435 * queue lock held and interrupts disabled.
1436 */
1437int blk_remove_plug(request_queue_t *q)
1438{
1439 WARN_ON(!irqs_disabled());
1440
1441 if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1442 return 0;
1443
1444 del_timer(&q->unplug_timer);
1445 return 1;
1446}
1447
1448EXPORT_SYMBOL(blk_remove_plug);
1449
1450/*
1451 * remove the plug and let it rip..
1452 */
1453void __generic_unplug_device(request_queue_t *q)
1454{
Nick Pigginfde6ad22005-06-23 00:08:53 -07001455 if (unlikely(test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456 return;
1457
1458 if (!blk_remove_plug(q))
1459 return;
1460
Jens Axboe22e2c502005-06-27 10:55:12 +02001461 q->request_fn(q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462}
1463EXPORT_SYMBOL(__generic_unplug_device);
1464
1465/**
1466 * generic_unplug_device - fire a request queue
1467 * @q: The &request_queue_t in question
1468 *
1469 * Description:
1470 * Linux uses plugging to build bigger requests queues before letting
1471 * the device have at them. If a queue is plugged, the I/O scheduler
1472 * is still adding and merging requests on the queue. Once the queue
1473 * gets unplugged, the request_fn defined for the queue is invoked and
1474 * transfers started.
1475 **/
1476void generic_unplug_device(request_queue_t *q)
1477{
1478 spin_lock_irq(q->queue_lock);
1479 __generic_unplug_device(q);
1480 spin_unlock_irq(q->queue_lock);
1481}
1482EXPORT_SYMBOL(generic_unplug_device);
1483
1484static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1485 struct page *page)
1486{
1487 request_queue_t *q = bdi->unplug_io_data;
1488
1489 /*
1490 * devices don't necessarily have an ->unplug_fn defined
1491 */
1492 if (q->unplug_fn)
1493 q->unplug_fn(q);
1494}
1495
1496static void blk_unplug_work(void *data)
1497{
1498 request_queue_t *q = data;
1499
1500 q->unplug_fn(q);
1501}
1502
1503static void blk_unplug_timeout(unsigned long data)
1504{
1505 request_queue_t *q = (request_queue_t *)data;
1506
1507 kblockd_schedule_work(&q->unplug_work);
1508}
1509
1510/**
1511 * blk_start_queue - restart a previously stopped queue
1512 * @q: The &request_queue_t in question
1513 *
1514 * Description:
1515 * blk_start_queue() will clear the stop flag on the queue, and call
1516 * the request_fn for the queue if it was in a stopped state when
1517 * entered. Also see blk_stop_queue(). Queue lock must be held.
1518 **/
1519void blk_start_queue(request_queue_t *q)
1520{
1521 clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1522
1523 /*
1524 * one level of recursion is ok and is much faster than kicking
1525 * the unplug handling
1526 */
1527 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1528 q->request_fn(q);
1529 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1530 } else {
1531 blk_plug_device(q);
1532 kblockd_schedule_work(&q->unplug_work);
1533 }
1534}
1535
1536EXPORT_SYMBOL(blk_start_queue);
1537
1538/**
1539 * blk_stop_queue - stop a queue
1540 * @q: The &request_queue_t in question
1541 *
1542 * Description:
1543 * The Linux block layer assumes that a block driver will consume all
1544 * entries on the request queue when the request_fn strategy is called.
1545 * Often this will not happen, because of hardware limitations (queue
1546 * depth settings). If a device driver gets a 'queue full' response,
1547 * or if it simply chooses not to queue more I/O at one point, it can
1548 * call this function to prevent the request_fn from being called until
1549 * the driver has signalled it's ready to go again. This happens by calling
1550 * blk_start_queue() to restart queue operations. Queue lock must be held.
1551 **/
1552void blk_stop_queue(request_queue_t *q)
1553{
1554 blk_remove_plug(q);
1555 set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1556}
1557EXPORT_SYMBOL(blk_stop_queue);
1558
1559/**
1560 * blk_sync_queue - cancel any pending callbacks on a queue
1561 * @q: the queue
1562 *
1563 * Description:
1564 * The block layer may perform asynchronous callback activity
1565 * on a queue, such as calling the unplug function after a timeout.
1566 * A block device may call blk_sync_queue to ensure that any
1567 * such activity is cancelled, thus allowing it to release resources
1568 * the the callbacks might use. The caller must already have made sure
1569 * that its ->make_request_fn will not re-add plugging prior to calling
1570 * this function.
1571 *
1572 */
1573void blk_sync_queue(struct request_queue *q)
1574{
1575 del_timer_sync(&q->unplug_timer);
1576 kblockd_flush();
1577}
1578EXPORT_SYMBOL(blk_sync_queue);
1579
1580/**
1581 * blk_run_queue - run a single device queue
1582 * @q: The queue to run
1583 */
1584void blk_run_queue(struct request_queue *q)
1585{
1586 unsigned long flags;
1587
1588 spin_lock_irqsave(q->queue_lock, flags);
1589 blk_remove_plug(q);
Ken Chena2997382005-04-16 15:25:43 -07001590 if (!elv_queue_empty(q))
1591 q->request_fn(q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592 spin_unlock_irqrestore(q->queue_lock, flags);
1593}
1594EXPORT_SYMBOL(blk_run_queue);
1595
1596/**
1597 * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1598 * @q: the request queue to be released
1599 *
1600 * Description:
1601 * blk_cleanup_queue is the pair to blk_init_queue() or
1602 * blk_queue_make_request(). It should be called when a request queue is
1603 * being released; typically when a block device is being de-registered.
1604 * Currently, its primary task it to free all the &struct request
1605 * structures that were allocated to the queue and the queue itself.
1606 *
1607 * Caveat:
1608 * Hopefully the low level driver will have finished any
1609 * outstanding requests first...
1610 **/
1611void blk_cleanup_queue(request_queue_t * q)
1612{
1613 struct request_list *rl = &q->rq;
1614
1615 if (!atomic_dec_and_test(&q->refcnt))
1616 return;
1617
1618 if (q->elevator)
1619 elevator_exit(q->elevator);
1620
1621 blk_sync_queue(q);
1622
1623 if (rl->rq_pool)
1624 mempool_destroy(rl->rq_pool);
1625
1626 if (q->queue_tags)
1627 __blk_queue_free_tags(q);
1628
1629 blk_queue_ordered(q, QUEUE_ORDERED_NONE);
1630
1631 kmem_cache_free(requestq_cachep, q);
1632}
1633
1634EXPORT_SYMBOL(blk_cleanup_queue);
1635
1636static int blk_init_free_list(request_queue_t *q)
1637{
1638 struct request_list *rl = &q->rq;
1639
1640 rl->count[READ] = rl->count[WRITE] = 0;
1641 rl->starved[READ] = rl->starved[WRITE] = 0;
Tejun Heocb98fc82005-10-28 08:29:39 +02001642 rl->elvpriv = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643 init_waitqueue_head(&rl->wait[READ]);
1644 init_waitqueue_head(&rl->wait[WRITE]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645
Christoph Lameter19460892005-06-23 00:08:19 -07001646 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1647 mempool_free_slab, request_cachep, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648
1649 if (!rl->rq_pool)
1650 return -ENOMEM;
1651
1652 return 0;
1653}
1654
1655static int __make_request(request_queue_t *, struct bio *);
1656
1657request_queue_t *blk_alloc_queue(int gfp_mask)
1658{
Christoph Lameter19460892005-06-23 00:08:19 -07001659 return blk_alloc_queue_node(gfp_mask, -1);
1660}
1661EXPORT_SYMBOL(blk_alloc_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662
Christoph Lameter19460892005-06-23 00:08:19 -07001663request_queue_t *blk_alloc_queue_node(int gfp_mask, int node_id)
1664{
1665 request_queue_t *q;
1666
1667 q = kmem_cache_alloc_node(requestq_cachep, gfp_mask, node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668 if (!q)
1669 return NULL;
1670
1671 memset(q, 0, sizeof(*q));
1672 init_timer(&q->unplug_timer);
1673 atomic_set(&q->refcnt, 1);
1674
1675 q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1676 q->backing_dev_info.unplug_io_data = q;
1677
1678 return q;
1679}
Christoph Lameter19460892005-06-23 00:08:19 -07001680EXPORT_SYMBOL(blk_alloc_queue_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681
1682/**
1683 * blk_init_queue - prepare a request queue for use with a block device
1684 * @rfn: The function to be called to process requests that have been
1685 * placed on the queue.
1686 * @lock: Request queue spin lock
1687 *
1688 * Description:
1689 * If a block device wishes to use the standard request handling procedures,
1690 * which sorts requests and coalesces adjacent requests, then it must
1691 * call blk_init_queue(). The function @rfn will be called when there
1692 * are requests on the queue that need to be processed. If the device
1693 * supports plugging, then @rfn may not be called immediately when requests
1694 * are available on the queue, but may be called at some time later instead.
1695 * Plugged queues are generally unplugged when a buffer belonging to one
1696 * of the requests on the queue is needed, or due to memory pressure.
1697 *
1698 * @rfn is not required, or even expected, to remove all requests off the
1699 * queue, but only as many as it can handle at a time. If it does leave
1700 * requests on the queue, it is responsible for arranging that the requests
1701 * get dealt with eventually.
1702 *
1703 * The queue spin lock must be held while manipulating the requests on the
1704 * request queue.
1705 *
1706 * Function returns a pointer to the initialized request queue, or NULL if
1707 * it didn't succeed.
1708 *
1709 * Note:
1710 * blk_init_queue() must be paired with a blk_cleanup_queue() call
1711 * when the block device is deactivated (such as at module unload).
1712 **/
Christoph Lameter19460892005-06-23 00:08:19 -07001713
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714request_queue_t *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1715{
Christoph Lameter19460892005-06-23 00:08:19 -07001716 return blk_init_queue_node(rfn, lock, -1);
1717}
1718EXPORT_SYMBOL(blk_init_queue);
1719
1720request_queue_t *
1721blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1722{
1723 request_queue_t *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724
1725 if (!q)
1726 return NULL;
1727
Christoph Lameter19460892005-06-23 00:08:19 -07001728 q->node = node_id;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 if (blk_init_free_list(q))
1730 goto out_init;
1731
152587d2005-04-12 16:22:06 -05001732 /*
1733 * if caller didn't supply a lock, they get per-queue locking with
1734 * our embedded lock
1735 */
1736 if (!lock) {
1737 spin_lock_init(&q->__queue_lock);
1738 lock = &q->__queue_lock;
1739 }
1740
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741 q->request_fn = rfn;
1742 q->back_merge_fn = ll_back_merge_fn;
1743 q->front_merge_fn = ll_front_merge_fn;
1744 q->merge_requests_fn = ll_merge_requests_fn;
1745 q->prep_rq_fn = NULL;
1746 q->unplug_fn = generic_unplug_device;
1747 q->queue_flags = (1 << QUEUE_FLAG_CLUSTER);
1748 q->queue_lock = lock;
1749
1750 blk_queue_segment_boundary(q, 0xffffffff);
1751
1752 blk_queue_make_request(q, __make_request);
1753 blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1754
1755 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1756 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1757
1758 /*
1759 * all done
1760 */
1761 if (!elevator_init(q, NULL)) {
1762 blk_queue_congestion_threshold(q);
1763 return q;
1764 }
1765
1766 blk_cleanup_queue(q);
1767out_init:
1768 kmem_cache_free(requestq_cachep, q);
1769 return NULL;
1770}
Christoph Lameter19460892005-06-23 00:08:19 -07001771EXPORT_SYMBOL(blk_init_queue_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772
1773int blk_get_queue(request_queue_t *q)
1774{
Nick Pigginfde6ad22005-06-23 00:08:53 -07001775 if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776 atomic_inc(&q->refcnt);
1777 return 0;
1778 }
1779
1780 return 1;
1781}
1782
1783EXPORT_SYMBOL(blk_get_queue);
1784
1785static inline void blk_free_request(request_queue_t *q, struct request *rq)
1786{
Tejun Heocb98fc82005-10-28 08:29:39 +02001787 if (rq->flags & REQ_ELVPRIV)
1788 elv_put_request(q, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789 mempool_free(rq, q->rq.rq_pool);
1790}
1791
Jens Axboe22e2c502005-06-27 10:55:12 +02001792static inline struct request *
Tejun Heocb98fc82005-10-28 08:29:39 +02001793blk_alloc_request(request_queue_t *q, int rw, struct bio *bio,
1794 int priv, int gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795{
1796 struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1797
1798 if (!rq)
1799 return NULL;
1800
1801 /*
1802 * first three bits are identical in rq->flags and bio->bi_rw,
1803 * see bio.h and blkdev.h
1804 */
1805 rq->flags = rw;
1806
Tejun Heocb98fc82005-10-28 08:29:39 +02001807 if (priv) {
1808 if (unlikely(elv_set_request(q, rq, bio, gfp_mask))) {
1809 mempool_free(rq, q->rq.rq_pool);
1810 return NULL;
1811 }
1812 rq->flags |= REQ_ELVPRIV;
1813 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814
Tejun Heocb98fc82005-10-28 08:29:39 +02001815 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816}
1817
1818/*
1819 * ioc_batching returns true if the ioc is a valid batching request and
1820 * should be given priority access to a request.
1821 */
1822static inline int ioc_batching(request_queue_t *q, struct io_context *ioc)
1823{
1824 if (!ioc)
1825 return 0;
1826
1827 /*
1828 * Make sure the process is able to allocate at least 1 request
1829 * even if the batch times out, otherwise we could theoretically
1830 * lose wakeups.
1831 */
1832 return ioc->nr_batch_requests == q->nr_batching ||
1833 (ioc->nr_batch_requests > 0
1834 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
1835}
1836
1837/*
1838 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
1839 * will cause the process to be a "batcher" on all queues in the system. This
1840 * is the behaviour we want though - once it gets a wakeup it should be given
1841 * a nice run.
1842 */
Adrian Bunk93d17d32005-06-25 14:59:10 -07001843static void ioc_set_batching(request_queue_t *q, struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844{
1845 if (!ioc || ioc_batching(q, ioc))
1846 return;
1847
1848 ioc->nr_batch_requests = q->nr_batching;
1849 ioc->last_waited = jiffies;
1850}
1851
1852static void __freed_request(request_queue_t *q, int rw)
1853{
1854 struct request_list *rl = &q->rq;
1855
1856 if (rl->count[rw] < queue_congestion_off_threshold(q))
1857 clear_queue_congested(q, rw);
1858
1859 if (rl->count[rw] + 1 <= q->nr_requests) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001860 if (waitqueue_active(&rl->wait[rw]))
1861 wake_up(&rl->wait[rw]);
1862
1863 blk_clear_queue_full(q, rw);
1864 }
1865}
1866
1867/*
1868 * A request has just been released. Account for it, update the full and
1869 * congestion status, wake up any waiters. Called under q->queue_lock.
1870 */
Tejun Heocb98fc82005-10-28 08:29:39 +02001871static void freed_request(request_queue_t *q, int rw, int priv)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872{
1873 struct request_list *rl = &q->rq;
1874
1875 rl->count[rw]--;
Tejun Heocb98fc82005-10-28 08:29:39 +02001876 if (priv)
1877 rl->elvpriv--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878
1879 __freed_request(q, rw);
1880
1881 if (unlikely(rl->starved[rw ^ 1]))
1882 __freed_request(q, rw ^ 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883}
1884
1885#define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
1886/*
Nick Piggind6344532005-06-28 20:45:14 -07001887 * Get a free request, queue_lock must be held.
1888 * Returns NULL on failure, with queue_lock held.
1889 * Returns !NULL on success, with queue_lock *not held*.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001891static struct request *get_request(request_queue_t *q, int rw, struct bio *bio,
1892 int gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893{
1894 struct request *rq = NULL;
1895 struct request_list *rl = &q->rq;
Nick Pigginfb3cc432005-06-28 20:45:15 -07001896 struct io_context *ioc = current_io_context(GFP_ATOMIC);
Tejun Heocb98fc82005-10-28 08:29:39 +02001897 int priv;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899 if (rl->count[rw]+1 >= q->nr_requests) {
1900 /*
1901 * The queue will fill after this allocation, so set it as
1902 * full, and mark this process as "batching". This process
1903 * will be allowed to complete a batch of requests, others
1904 * will be blocked.
1905 */
1906 if (!blk_queue_full(q, rw)) {
1907 ioc_set_batching(q, ioc);
1908 blk_set_queue_full(q, rw);
1909 }
1910 }
1911
Jens Axboe22e2c502005-06-27 10:55:12 +02001912 switch (elv_may_queue(q, rw, bio)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913 case ELV_MQUEUE_NO:
1914 goto rq_starved;
1915 case ELV_MQUEUE_MAY:
1916 break;
1917 case ELV_MQUEUE_MUST:
1918 goto get_rq;
1919 }
1920
1921 if (blk_queue_full(q, rw) && !ioc_batching(q, ioc)) {
1922 /*
1923 * The queue is full and the allocating process is not a
1924 * "batcher", and not exempted by the IO scheduler
1925 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 goto out;
1927 }
1928
1929get_rq:
Jens Axboe082cf692005-06-28 16:35:11 +02001930 /*
1931 * Only allow batching queuers to allocate up to 50% over the defined
1932 * limit of requests, otherwise we could have thousands of requests
1933 * allocated with any setting of ->nr_requests
1934 */
Hugh Dickinsfd782a42005-06-29 15:15:40 +01001935 if (rl->count[rw] >= (3 * q->nr_requests / 2))
Jens Axboe082cf692005-06-28 16:35:11 +02001936 goto out;
Hugh Dickinsfd782a42005-06-29 15:15:40 +01001937
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938 rl->count[rw]++;
1939 rl->starved[rw] = 0;
1940 if (rl->count[rw] >= queue_congestion_on_threshold(q))
1941 set_queue_congested(q, rw);
Tejun Heocb98fc82005-10-28 08:29:39 +02001942
1943 priv = !test_bit(QUEUE_FLAG_BYPASS, &q->queue_flags);
1944 if (priv)
1945 rl->elvpriv++;
1946
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947 spin_unlock_irq(q->queue_lock);
1948
Tejun Heocb98fc82005-10-28 08:29:39 +02001949 rq = blk_alloc_request(q, rw, bio, priv, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950 if (!rq) {
1951 /*
1952 * Allocation failed presumably due to memory. Undo anything
1953 * we might have messed up.
1954 *
1955 * Allocating task should really be put onto the front of the
1956 * wait queue, but this is pretty rare.
1957 */
1958 spin_lock_irq(q->queue_lock);
Tejun Heocb98fc82005-10-28 08:29:39 +02001959 freed_request(q, rw, priv);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960
1961 /*
1962 * in the very unlikely event that allocation failed and no
1963 * requests for this direction was pending, mark us starved
1964 * so that freeing of a request in the other direction will
1965 * notice us. another possible fix would be to split the
1966 * rq mempool into READ and WRITE
1967 */
1968rq_starved:
1969 if (unlikely(rl->count[rw] == 0))
1970 rl->starved[rw] = 1;
1971
Linus Torvalds1da177e2005-04-16 15:20:36 -07001972 goto out;
1973 }
1974
1975 if (ioc_batching(q, ioc))
1976 ioc->nr_batch_requests--;
1977
1978 rq_init(q, rq);
1979 rq->rl = rl;
1980out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981 return rq;
1982}
1983
1984/*
1985 * No available requests for this queue, unplug the device and wait for some
1986 * requests to become available.
Nick Piggind6344532005-06-28 20:45:14 -07001987 *
1988 * Called with q->queue_lock held, and returns with it unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001990static struct request *get_request_wait(request_queue_t *q, int rw,
1991 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001993 struct request *rq;
1994
Nick Piggin450991b2005-06-28 20:45:13 -07001995 rq = get_request(q, rw, bio, GFP_NOIO);
1996 while (!rq) {
1997 DEFINE_WAIT(wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998 struct request_list *rl = &q->rq;
1999
2000 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
2001 TASK_UNINTERRUPTIBLE);
2002
Jens Axboe22e2c502005-06-27 10:55:12 +02002003 rq = get_request(q, rw, bio, GFP_NOIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004
2005 if (!rq) {
2006 struct io_context *ioc;
2007
Nick Piggind6344532005-06-28 20:45:14 -07002008 __generic_unplug_device(q);
2009 spin_unlock_irq(q->queue_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 io_schedule();
2011
2012 /*
2013 * After sleeping, we become a "batching" process and
2014 * will be able to allocate at least one request, and
2015 * up to a big batch of them for a small period time.
2016 * See ioc_batching, ioc_set_batching
2017 */
Nick Pigginfb3cc432005-06-28 20:45:15 -07002018 ioc = current_io_context(GFP_NOIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019 ioc_set_batching(q, ioc);
Nick Piggind6344532005-06-28 20:45:14 -07002020
2021 spin_lock_irq(q->queue_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022 }
2023 finish_wait(&rl->wait[rw], &wait);
Nick Piggin450991b2005-06-28 20:45:13 -07002024 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025
2026 return rq;
2027}
2028
2029struct request *blk_get_request(request_queue_t *q, int rw, int gfp_mask)
2030{
2031 struct request *rq;
2032
2033 BUG_ON(rw != READ && rw != WRITE);
2034
Nick Piggind6344532005-06-28 20:45:14 -07002035 spin_lock_irq(q->queue_lock);
2036 if (gfp_mask & __GFP_WAIT) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002037 rq = get_request_wait(q, rw, NULL);
Nick Piggind6344532005-06-28 20:45:14 -07002038 } else {
Jens Axboe22e2c502005-06-27 10:55:12 +02002039 rq = get_request(q, rw, NULL, gfp_mask);
Nick Piggind6344532005-06-28 20:45:14 -07002040 if (!rq)
2041 spin_unlock_irq(q->queue_lock);
2042 }
2043 /* q->queue_lock is unlocked at this point */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044
2045 return rq;
2046}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047EXPORT_SYMBOL(blk_get_request);
2048
2049/**
2050 * blk_requeue_request - put a request back on queue
2051 * @q: request queue where request should be inserted
2052 * @rq: request to be inserted
2053 *
2054 * Description:
2055 * Drivers often keep queueing requests until the hardware cannot accept
2056 * more, when that condition happens we need to put the request back
2057 * on the queue. Must be called with queue lock held.
2058 */
2059void blk_requeue_request(request_queue_t *q, struct request *rq)
2060{
2061 if (blk_rq_tagged(rq))
2062 blk_queue_end_tag(q, rq);
2063
2064 elv_requeue_request(q, rq);
2065}
2066
2067EXPORT_SYMBOL(blk_requeue_request);
2068
2069/**
2070 * blk_insert_request - insert a special request in to a request queue
2071 * @q: request queue where request should be inserted
2072 * @rq: request to be inserted
2073 * @at_head: insert request at head or tail of queue
2074 * @data: private data
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075 *
2076 * Description:
2077 * Many block devices need to execute commands asynchronously, so they don't
2078 * block the whole kernel from preemption during request execution. This is
2079 * accomplished normally by inserting aritficial requests tagged as
2080 * REQ_SPECIAL in to the corresponding request queue, and letting them be
2081 * scheduled for actual execution by the request queue.
2082 *
2083 * We have the option of inserting the head or the tail of the queue.
2084 * Typically we use the tail for new ioctls and so forth. We use the head
2085 * of the queue for things like a QUEUE_FULL message from a device, or a
2086 * host that is unable to accept a particular command.
2087 */
2088void blk_insert_request(request_queue_t *q, struct request *rq,
Tejun Heo 867d1192005-04-24 02:06:05 -05002089 int at_head, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090{
Tejun Heo 867d1192005-04-24 02:06:05 -05002091 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092 unsigned long flags;
2093
2094 /*
2095 * tell I/O scheduler that this isn't a regular read/write (ie it
2096 * must not attempt merges on this) and that it acts as a soft
2097 * barrier
2098 */
2099 rq->flags |= REQ_SPECIAL | REQ_SOFTBARRIER;
2100
2101 rq->special = data;
2102
2103 spin_lock_irqsave(q->queue_lock, flags);
2104
2105 /*
2106 * If command is tagged, release the tag
2107 */
Tejun Heo 867d1192005-04-24 02:06:05 -05002108 if (blk_rq_tagged(rq))
2109 blk_queue_end_tag(q, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110
Tejun Heo 867d1192005-04-24 02:06:05 -05002111 drive_stat_acct(rq, rq->nr_sectors, 1);
2112 __elv_add_request(q, rq, where, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114 if (blk_queue_plugged(q))
2115 __generic_unplug_device(q);
2116 else
2117 q->request_fn(q);
2118 spin_unlock_irqrestore(q->queue_lock, flags);
2119}
2120
2121EXPORT_SYMBOL(blk_insert_request);
2122
2123/**
2124 * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2125 * @q: request queue where request should be inserted
Christoph Hellwig 73747ae2005-06-20 14:21:01 +02002126 * @rq: request structure to fill
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127 * @ubuf: the user buffer
2128 * @len: length of user data
2129 *
2130 * Description:
2131 * Data will be mapped directly for zero copy io, if possible. Otherwise
2132 * a kernel bounce buffer is used.
2133 *
2134 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2135 * still in process context.
2136 *
2137 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2138 * before being submitted to the device, as pages mapped may be out of
2139 * reach. It's the callers responsibility to make sure this happens. The
2140 * original bio must be passed back in to blk_rq_unmap_user() for proper
2141 * unmapping.
2142 */
Jens Axboedd1cab92005-06-20 14:06:01 +02002143int blk_rq_map_user(request_queue_t *q, struct request *rq, void __user *ubuf,
2144 unsigned int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145{
2146 unsigned long uaddr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 struct bio *bio;
Jens Axboedd1cab92005-06-20 14:06:01 +02002148 int reading;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149
2150 if (len > (q->max_sectors << 9))
Jens Axboedd1cab92005-06-20 14:06:01 +02002151 return -EINVAL;
2152 if (!len || !ubuf)
2153 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154
Jens Axboedd1cab92005-06-20 14:06:01 +02002155 reading = rq_data_dir(rq) == READ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156
2157 /*
2158 * if alignment requirement is satisfied, map in user pages for
2159 * direct dma. else, set up kernel bounce buffers
2160 */
2161 uaddr = (unsigned long) ubuf;
2162 if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
Jens Axboedd1cab92005-06-20 14:06:01 +02002163 bio = bio_map_user(q, NULL, uaddr, len, reading);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 else
Jens Axboedd1cab92005-06-20 14:06:01 +02002165 bio = bio_copy_user(q, uaddr, len, reading);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166
2167 if (!IS_ERR(bio)) {
2168 rq->bio = rq->biotail = bio;
2169 blk_rq_bio_prep(q, rq, bio);
2170
2171 rq->buffer = rq->data = NULL;
2172 rq->data_len = len;
Jens Axboedd1cab92005-06-20 14:06:01 +02002173 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174 }
2175
2176 /*
2177 * bio is the err-ptr
2178 */
Jens Axboedd1cab92005-06-20 14:06:01 +02002179 return PTR_ERR(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180}
2181
2182EXPORT_SYMBOL(blk_rq_map_user);
2183
2184/**
James Bottomley f1970ba2005-06-20 14:06:52 +02002185 * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
2186 * @q: request queue where request should be inserted
2187 * @rq: request to map data to
2188 * @iov: pointer to the iovec
2189 * @iov_count: number of elements in the iovec
2190 *
2191 * Description:
2192 * Data will be mapped directly for zero copy io, if possible. Otherwise
2193 * a kernel bounce buffer is used.
2194 *
2195 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2196 * still in process context.
2197 *
2198 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2199 * before being submitted to the device, as pages mapped may be out of
2200 * reach. It's the callers responsibility to make sure this happens. The
2201 * original bio must be passed back in to blk_rq_unmap_user() for proper
2202 * unmapping.
2203 */
2204int blk_rq_map_user_iov(request_queue_t *q, struct request *rq,
2205 struct sg_iovec *iov, int iov_count)
2206{
2207 struct bio *bio;
2208
2209 if (!iov || iov_count <= 0)
2210 return -EINVAL;
2211
2212 /* we don't allow misaligned data like bio_map_user() does. If the
2213 * user is using sg, they're expected to know the alignment constraints
2214 * and respect them accordingly */
2215 bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
2216 if (IS_ERR(bio))
2217 return PTR_ERR(bio);
2218
2219 rq->bio = rq->biotail = bio;
2220 blk_rq_bio_prep(q, rq, bio);
2221 rq->buffer = rq->data = NULL;
2222 rq->data_len = bio->bi_size;
2223 return 0;
2224}
2225
2226EXPORT_SYMBOL(blk_rq_map_user_iov);
2227
2228/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 * blk_rq_unmap_user - unmap a request with user data
Christoph Hellwig 73747ae2005-06-20 14:21:01 +02002230 * @bio: bio to be unmapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 * @ulen: length of user buffer
2232 *
2233 * Description:
Christoph Hellwig 73747ae2005-06-20 14:21:01 +02002234 * Unmap a bio previously mapped by blk_rq_map_user().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 */
Jens Axboedd1cab92005-06-20 14:06:01 +02002236int blk_rq_unmap_user(struct bio *bio, unsigned int ulen)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237{
2238 int ret = 0;
2239
2240 if (bio) {
2241 if (bio_flagged(bio, BIO_USER_MAPPED))
2242 bio_unmap_user(bio);
2243 else
2244 ret = bio_uncopy_user(bio);
2245 }
2246
Jens Axboedd1cab92005-06-20 14:06:01 +02002247 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248}
2249
2250EXPORT_SYMBOL(blk_rq_unmap_user);
2251
2252/**
Mike Christie df46b9a2005-06-20 14:04:44 +02002253 * blk_rq_map_kern - map kernel data to a request, for REQ_BLOCK_PC usage
2254 * @q: request queue where request should be inserted
Christoph Hellwig 73747ae2005-06-20 14:21:01 +02002255 * @rq: request to fill
Mike Christie df46b9a2005-06-20 14:04:44 +02002256 * @kbuf: the kernel buffer
2257 * @len: length of user data
Christoph Hellwig 73747ae2005-06-20 14:21:01 +02002258 * @gfp_mask: memory allocation flags
Mike Christie df46b9a2005-06-20 14:04:44 +02002259 */
Jens Axboedd1cab92005-06-20 14:06:01 +02002260int blk_rq_map_kern(request_queue_t *q, struct request *rq, void *kbuf,
2261 unsigned int len, unsigned int gfp_mask)
Mike Christie df46b9a2005-06-20 14:04:44 +02002262{
Mike Christie df46b9a2005-06-20 14:04:44 +02002263 struct bio *bio;
2264
2265 if (len > (q->max_sectors << 9))
Jens Axboedd1cab92005-06-20 14:06:01 +02002266 return -EINVAL;
2267 if (!len || !kbuf)
2268 return -EINVAL;
Mike Christie df46b9a2005-06-20 14:04:44 +02002269
2270 bio = bio_map_kern(q, kbuf, len, gfp_mask);
Jens Axboedd1cab92005-06-20 14:06:01 +02002271 if (IS_ERR(bio))
2272 return PTR_ERR(bio);
Mike Christie df46b9a2005-06-20 14:04:44 +02002273
Jens Axboedd1cab92005-06-20 14:06:01 +02002274 if (rq_data_dir(rq) == WRITE)
2275 bio->bi_rw |= (1 << BIO_RW);
Mike Christie df46b9a2005-06-20 14:04:44 +02002276
Jens Axboedd1cab92005-06-20 14:06:01 +02002277 rq->bio = rq->biotail = bio;
2278 blk_rq_bio_prep(q, rq, bio);
Mike Christie df46b9a2005-06-20 14:04:44 +02002279
Jens Axboedd1cab92005-06-20 14:06:01 +02002280 rq->buffer = rq->data = NULL;
2281 rq->data_len = len;
2282 return 0;
Mike Christie df46b9a2005-06-20 14:04:44 +02002283}
2284
2285EXPORT_SYMBOL(blk_rq_map_kern);
2286
Christoph Hellwig 73747ae2005-06-20 14:21:01 +02002287/**
2288 * blk_execute_rq_nowait - insert a request into queue for execution
2289 * @q: queue to insert the request in
2290 * @bd_disk: matching gendisk
2291 * @rq: request to insert
2292 * @at_head: insert request at head or tail of queue
2293 * @done: I/O completion handler
2294 *
2295 * Description:
2296 * Insert a fully prepared request at the back of the io scheduler queue
2297 * for execution. Don't wait for completion.
2298 */
James Bottomley f1970ba2005-06-20 14:06:52 +02002299void blk_execute_rq_nowait(request_queue_t *q, struct gendisk *bd_disk,
2300 struct request *rq, int at_head,
2301 void (*done)(struct request *))
2302{
2303 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
2304
2305 rq->rq_disk = bd_disk;
2306 rq->flags |= REQ_NOMERGE;
2307 rq->end_io = done;
2308 elv_add_request(q, rq, where, 1);
2309 generic_unplug_device(q);
2310}
2311
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312/**
2313 * blk_execute_rq - insert a request into queue for execution
2314 * @q: queue to insert the request in
2315 * @bd_disk: matching gendisk
2316 * @rq: request to insert
James Bottomley 994ca9a2005-06-20 14:11:09 +02002317 * @at_head: insert request at head or tail of queue
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318 *
2319 * Description:
2320 * Insert a fully prepared request at the back of the io scheduler queue
Christoph Hellwig 73747ae2005-06-20 14:21:01 +02002321 * for execution and wait for completion.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 */
2323int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
James Bottomley 994ca9a2005-06-20 14:11:09 +02002324 struct request *rq, int at_head)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325{
2326 DECLARE_COMPLETION(wait);
2327 char sense[SCSI_SENSE_BUFFERSIZE];
2328 int err = 0;
2329
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330 /*
2331 * we need an extra reference to the request, so we can look at
2332 * it after io completion
2333 */
2334 rq->ref_count++;
2335
2336 if (!rq->sense) {
2337 memset(sense, 0, sizeof(sense));
2338 rq->sense = sense;
2339 rq->sense_len = 0;
2340 }
2341
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342 rq->waiting = &wait;
James Bottomley 994ca9a2005-06-20 14:11:09 +02002343 blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344 wait_for_completion(&wait);
2345 rq->waiting = NULL;
2346
2347 if (rq->errors)
2348 err = -EIO;
2349
2350 return err;
2351}
2352
2353EXPORT_SYMBOL(blk_execute_rq);
2354
2355/**
2356 * blkdev_issue_flush - queue a flush
2357 * @bdev: blockdev to issue flush for
2358 * @error_sector: error sector
2359 *
2360 * Description:
2361 * Issue a flush for the block device in question. Caller can supply
2362 * room for storing the error offset in case of a flush error, if they
2363 * wish to. Caller must run wait_for_completion() on its own.
2364 */
2365int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2366{
2367 request_queue_t *q;
2368
2369 if (bdev->bd_disk == NULL)
2370 return -ENXIO;
2371
2372 q = bdev_get_queue(bdev);
2373 if (!q)
2374 return -ENXIO;
2375 if (!q->issue_flush_fn)
2376 return -EOPNOTSUPP;
2377
2378 return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2379}
2380
2381EXPORT_SYMBOL(blkdev_issue_flush);
2382
Adrian Bunk93d17d32005-06-25 14:59:10 -07002383static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384{
2385 int rw = rq_data_dir(rq);
2386
2387 if (!blk_fs_request(rq) || !rq->rq_disk)
2388 return;
2389
2390 if (rw == READ) {
2391 __disk_stat_add(rq->rq_disk, read_sectors, nr_sectors);
2392 if (!new_io)
2393 __disk_stat_inc(rq->rq_disk, read_merges);
2394 } else if (rw == WRITE) {
2395 __disk_stat_add(rq->rq_disk, write_sectors, nr_sectors);
2396 if (!new_io)
2397 __disk_stat_inc(rq->rq_disk, write_merges);
2398 }
2399 if (new_io) {
2400 disk_round_stats(rq->rq_disk);
2401 rq->rq_disk->in_flight++;
2402 }
2403}
2404
2405/*
2406 * add-request adds a request to the linked list.
2407 * queue lock is held and interrupts disabled, as we muck with the
2408 * request queue list.
2409 */
2410static inline void add_request(request_queue_t * q, struct request * req)
2411{
2412 drive_stat_acct(req, req->nr_sectors, 1);
2413
2414 if (q->activity_fn)
2415 q->activity_fn(q->activity_data, rq_data_dir(req));
2416
2417 /*
2418 * elevator indicated where it wants this request to be
2419 * inserted at elevator_merge time
2420 */
2421 __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2422}
2423
2424/*
2425 * disk_round_stats() - Round off the performance stats on a struct
2426 * disk_stats.
2427 *
2428 * The average IO queue length and utilisation statistics are maintained
2429 * by observing the current state of the queue length and the amount of
2430 * time it has been in this state for.
2431 *
2432 * Normally, that accounting is done on IO completion, but that can result
2433 * in more than a second's worth of IO being accounted for within any one
2434 * second, leading to >100% utilisation. To deal with that, we call this
2435 * function to do a round-off before returning the results when reading
2436 * /proc/diskstats. This accounts immediately for all queue usage up to
2437 * the current jiffies and restarts the counters again.
2438 */
2439void disk_round_stats(struct gendisk *disk)
2440{
2441 unsigned long now = jiffies;
2442
Chen, Kenneth Wb2982642005-10-13 21:49:29 +02002443 if (now == disk->stamp)
2444 return;
2445
Chen, Kenneth W20e5c812005-10-13 21:48:42 +02002446 if (disk->in_flight) {
2447 __disk_stat_add(disk, time_in_queue,
2448 disk->in_flight * (now - disk->stamp));
2449 __disk_stat_add(disk, io_ticks, (now - disk->stamp));
2450 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451 disk->stamp = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452}
2453
2454/*
2455 * queue lock must be held
2456 */
2457static void __blk_put_request(request_queue_t *q, struct request *req)
2458{
2459 struct request_list *rl = req->rl;
2460
2461 if (unlikely(!q))
2462 return;
2463 if (unlikely(--req->ref_count))
2464 return;
2465
Tejun Heo8922e162005-10-20 16:23:44 +02002466 elv_completed_request(q, req);
2467
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468 req->rq_status = RQ_INACTIVE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469 req->rl = NULL;
2470
2471 /*
2472 * Request may not have originated from ll_rw_blk. if not,
2473 * it didn't come out of our reserved rq pools
2474 */
2475 if (rl) {
2476 int rw = rq_data_dir(req);
Tejun Heocb98fc82005-10-28 08:29:39 +02002477 int priv = req->flags & REQ_ELVPRIV;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479 BUG_ON(!list_empty(&req->queuelist));
2480
2481 blk_free_request(q, req);
Tejun Heocb98fc82005-10-28 08:29:39 +02002482 freed_request(q, rw, priv);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483 }
2484}
2485
2486void blk_put_request(struct request *req)
2487{
Tejun Heo8922e162005-10-20 16:23:44 +02002488 unsigned long flags;
2489 request_queue_t *q = req->q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490
Tejun Heo8922e162005-10-20 16:23:44 +02002491 /*
2492 * Gee, IDE calls in w/ NULL q. Fix IDE and remove the
2493 * following if (q) test.
2494 */
2495 if (q) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496 spin_lock_irqsave(q->queue_lock, flags);
2497 __blk_put_request(q, req);
2498 spin_unlock_irqrestore(q->queue_lock, flags);
2499 }
2500}
2501
2502EXPORT_SYMBOL(blk_put_request);
2503
2504/**
2505 * blk_end_sync_rq - executes a completion event on a request
2506 * @rq: request to complete
2507 */
2508void blk_end_sync_rq(struct request *rq)
2509{
2510 struct completion *waiting = rq->waiting;
2511
2512 rq->waiting = NULL;
2513 __blk_put_request(rq->q, rq);
2514
2515 /*
2516 * complete last, if this is a stack request the process (and thus
2517 * the rq pointer) could be invalid right after this complete()
2518 */
2519 complete(waiting);
2520}
2521EXPORT_SYMBOL(blk_end_sync_rq);
2522
2523/**
2524 * blk_congestion_wait - wait for a queue to become uncongested
2525 * @rw: READ or WRITE
2526 * @timeout: timeout in jiffies
2527 *
2528 * Waits for up to @timeout jiffies for a queue (any queue) to exit congestion.
2529 * If no queues are congested then just wait for the next request to be
2530 * returned.
2531 */
2532long blk_congestion_wait(int rw, long timeout)
2533{
2534 long ret;
2535 DEFINE_WAIT(wait);
2536 wait_queue_head_t *wqh = &congestion_wqh[rw];
2537
2538 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
2539 ret = io_schedule_timeout(timeout);
2540 finish_wait(wqh, &wait);
2541 return ret;
2542}
2543
2544EXPORT_SYMBOL(blk_congestion_wait);
2545
2546/*
2547 * Has to be called with the request spinlock acquired
2548 */
2549static int attempt_merge(request_queue_t *q, struct request *req,
2550 struct request *next)
2551{
2552 if (!rq_mergeable(req) || !rq_mergeable(next))
2553 return 0;
2554
2555 /*
2556 * not contigious
2557 */
2558 if (req->sector + req->nr_sectors != next->sector)
2559 return 0;
2560
2561 if (rq_data_dir(req) != rq_data_dir(next)
2562 || req->rq_disk != next->rq_disk
2563 || next->waiting || next->special)
2564 return 0;
2565
2566 /*
2567 * If we are allowed to merge, then append bio list
2568 * from next to rq and release next. merge_requests_fn
2569 * will have updated segment counts, update sector
2570 * counts here.
2571 */
2572 if (!q->merge_requests_fn(q, req, next))
2573 return 0;
2574
2575 /*
2576 * At this point we have either done a back merge
2577 * or front merge. We need the smaller start_time of
2578 * the merged requests to be the current request
2579 * for accounting purposes.
2580 */
2581 if (time_after(req->start_time, next->start_time))
2582 req->start_time = next->start_time;
2583
2584 req->biotail->bi_next = next->bio;
2585 req->biotail = next->biotail;
2586
2587 req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2588
2589 elv_merge_requests(q, req, next);
2590
2591 if (req->rq_disk) {
2592 disk_round_stats(req->rq_disk);
2593 req->rq_disk->in_flight--;
2594 }
2595
Jens Axboe22e2c502005-06-27 10:55:12 +02002596 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2597
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598 __blk_put_request(q, next);
2599 return 1;
2600}
2601
2602static inline int attempt_back_merge(request_queue_t *q, struct request *rq)
2603{
2604 struct request *next = elv_latter_request(q, rq);
2605
2606 if (next)
2607 return attempt_merge(q, rq, next);
2608
2609 return 0;
2610}
2611
2612static inline int attempt_front_merge(request_queue_t *q, struct request *rq)
2613{
2614 struct request *prev = elv_former_request(q, rq);
2615
2616 if (prev)
2617 return attempt_merge(q, prev, rq);
2618
2619 return 0;
2620}
2621
2622/**
2623 * blk_attempt_remerge - attempt to remerge active head with next request
2624 * @q: The &request_queue_t belonging to the device
2625 * @rq: The head request (usually)
2626 *
2627 * Description:
2628 * For head-active devices, the queue can easily be unplugged so quickly
2629 * that proper merging is not done on the front request. This may hurt
2630 * performance greatly for some devices. The block layer cannot safely
2631 * do merging on that first request for these queues, but the driver can
2632 * call this function and make it happen any way. Only the driver knows
2633 * when it is safe to do so.
2634 **/
2635void blk_attempt_remerge(request_queue_t *q, struct request *rq)
2636{
2637 unsigned long flags;
2638
2639 spin_lock_irqsave(q->queue_lock, flags);
2640 attempt_back_merge(q, rq);
2641 spin_unlock_irqrestore(q->queue_lock, flags);
2642}
2643
2644EXPORT_SYMBOL(blk_attempt_remerge);
2645
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646static int __make_request(request_queue_t *q, struct bio *bio)
2647{
Nick Piggin450991b2005-06-28 20:45:13 -07002648 struct request *req;
Jens Axboe4a534f92005-04-16 15:25:40 -07002649 int el_ret, rw, nr_sectors, cur_nr_sectors, barrier, err, sync;
Jens Axboe22e2c502005-06-27 10:55:12 +02002650 unsigned short prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651 sector_t sector;
2652
2653 sector = bio->bi_sector;
2654 nr_sectors = bio_sectors(bio);
2655 cur_nr_sectors = bio_cur_sectors(bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002656 prio = bio_prio(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657
2658 rw = bio_data_dir(bio);
Jens Axboe4a534f92005-04-16 15:25:40 -07002659 sync = bio_sync(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660
2661 /*
2662 * low level driver can indicate that it wants pages above a
2663 * certain limit bounced to low memory (ie for highmem, or even
2664 * ISA dma in theory)
2665 */
2666 blk_queue_bounce(q, &bio);
2667
2668 spin_lock_prefetch(q->queue_lock);
2669
2670 barrier = bio_barrier(bio);
Nick Pigginfde6ad22005-06-23 00:08:53 -07002671 if (unlikely(barrier) && (q->ordered == QUEUE_ORDERED_NONE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672 err = -EOPNOTSUPP;
2673 goto end_io;
2674 }
2675
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676 spin_lock_irq(q->queue_lock);
2677
Nick Piggin450991b2005-06-28 20:45:13 -07002678 if (unlikely(barrier) || elv_queue_empty(q))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679 goto get_rq;
2680
2681 el_ret = elv_merge(q, &req, bio);
2682 switch (el_ret) {
2683 case ELEVATOR_BACK_MERGE:
2684 BUG_ON(!rq_mergeable(req));
2685
2686 if (!q->back_merge_fn(q, req, bio))
2687 break;
2688
2689 req->biotail->bi_next = bio;
2690 req->biotail = bio;
2691 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02002692 req->ioprio = ioprio_best(req->ioprio, prio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002693 drive_stat_acct(req, nr_sectors, 0);
2694 if (!attempt_back_merge(q, req))
2695 elv_merged_request(q, req);
2696 goto out;
2697
2698 case ELEVATOR_FRONT_MERGE:
2699 BUG_ON(!rq_mergeable(req));
2700
2701 if (!q->front_merge_fn(q, req, bio))
2702 break;
2703
2704 bio->bi_next = req->bio;
2705 req->bio = bio;
2706
2707 /*
2708 * may not be valid. if the low level driver said
2709 * it didn't need a bounce buffer then it better
2710 * not touch req->buffer either...
2711 */
2712 req->buffer = bio_data(bio);
2713 req->current_nr_sectors = cur_nr_sectors;
2714 req->hard_cur_sectors = cur_nr_sectors;
2715 req->sector = req->hard_sector = sector;
2716 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02002717 req->ioprio = ioprio_best(req->ioprio, prio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718 drive_stat_acct(req, nr_sectors, 0);
2719 if (!attempt_front_merge(q, req))
2720 elv_merged_request(q, req);
2721 goto out;
2722
Nick Piggin450991b2005-06-28 20:45:13 -07002723 /* ELV_NO_MERGE: elevator says don't/can't merge. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724 default:
Nick Piggin450991b2005-06-28 20:45:13 -07002725 ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 }
2727
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728get_rq:
Nick Piggin450991b2005-06-28 20:45:13 -07002729 /*
2730 * Grab a free request. This is might sleep but can not fail.
Nick Piggind6344532005-06-28 20:45:14 -07002731 * Returns with the queue unlocked.
Nick Piggin450991b2005-06-28 20:45:13 -07002732 */
Nick Piggin450991b2005-06-28 20:45:13 -07002733 req = get_request_wait(q, rw, bio);
Nick Piggind6344532005-06-28 20:45:14 -07002734
Nick Piggin450991b2005-06-28 20:45:13 -07002735 /*
2736 * After dropping the lock and possibly sleeping here, our request
2737 * may now be mergeable after it had proven unmergeable (above).
2738 * We don't worry about that case for efficiency. It won't happen
2739 * often, and the elevators are able to handle it.
2740 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741
2742 req->flags |= REQ_CMD;
2743
2744 /*
2745 * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2746 */
2747 if (bio_rw_ahead(bio) || bio_failfast(bio))
2748 req->flags |= REQ_FAILFAST;
2749
2750 /*
2751 * REQ_BARRIER implies no merging, but lets make it explicit
2752 */
Nick Pigginfde6ad22005-06-23 00:08:53 -07002753 if (unlikely(barrier))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754 req->flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2755
2756 req->errors = 0;
2757 req->hard_sector = req->sector = sector;
2758 req->hard_nr_sectors = req->nr_sectors = nr_sectors;
2759 req->current_nr_sectors = req->hard_cur_sectors = cur_nr_sectors;
2760 req->nr_phys_segments = bio_phys_segments(q, bio);
2761 req->nr_hw_segments = bio_hw_segments(q, bio);
2762 req->buffer = bio_data(bio); /* see ->buffer comment above */
2763 req->waiting = NULL;
2764 req->bio = req->biotail = bio;
Jens Axboe22e2c502005-06-27 10:55:12 +02002765 req->ioprio = prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 req->rq_disk = bio->bi_bdev->bd_disk;
2767 req->start_time = jiffies;
2768
Nick Piggin450991b2005-06-28 20:45:13 -07002769 spin_lock_irq(q->queue_lock);
2770 if (elv_queue_empty(q))
2771 blk_plug_device(q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772 add_request(q, req);
2773out:
Jens Axboe4a534f92005-04-16 15:25:40 -07002774 if (sync)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 __generic_unplug_device(q);
2776
2777 spin_unlock_irq(q->queue_lock);
2778 return 0;
2779
2780end_io:
2781 bio_endio(bio, nr_sectors << 9, err);
2782 return 0;
2783}
2784
2785/*
2786 * If bio->bi_dev is a partition, remap the location
2787 */
2788static inline void blk_partition_remap(struct bio *bio)
2789{
2790 struct block_device *bdev = bio->bi_bdev;
2791
2792 if (bdev != bdev->bd_contains) {
2793 struct hd_struct *p = bdev->bd_part;
2794
Jens Axboe22e2c502005-06-27 10:55:12 +02002795 switch (bio_data_dir(bio)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 case READ:
2797 p->read_sectors += bio_sectors(bio);
2798 p->reads++;
2799 break;
2800 case WRITE:
2801 p->write_sectors += bio_sectors(bio);
2802 p->writes++;
2803 break;
2804 }
2805 bio->bi_sector += p->start_sect;
2806 bio->bi_bdev = bdev->bd_contains;
2807 }
2808}
2809
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810static void handle_bad_sector(struct bio *bio)
2811{
2812 char b[BDEVNAME_SIZE];
2813
2814 printk(KERN_INFO "attempt to access beyond end of device\n");
2815 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
2816 bdevname(bio->bi_bdev, b),
2817 bio->bi_rw,
2818 (unsigned long long)bio->bi_sector + bio_sectors(bio),
2819 (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
2820
2821 set_bit(BIO_EOF, &bio->bi_flags);
2822}
2823
2824/**
2825 * generic_make_request: hand a buffer to its device driver for I/O
2826 * @bio: The bio describing the location in memory and on the device.
2827 *
2828 * generic_make_request() is used to make I/O requests of block
2829 * devices. It is passed a &struct bio, which describes the I/O that needs
2830 * to be done.
2831 *
2832 * generic_make_request() does not return any status. The
2833 * success/failure status of the request, along with notification of
2834 * completion, is delivered asynchronously through the bio->bi_end_io
2835 * function described (one day) else where.
2836 *
2837 * The caller of generic_make_request must make sure that bi_io_vec
2838 * are set to describe the memory buffer, and that bi_dev and bi_sector are
2839 * set to describe the device address, and the
2840 * bi_end_io and optionally bi_private are set to describe how
2841 * completion notification should be signaled.
2842 *
2843 * generic_make_request and the drivers it calls may use bi_next if this
2844 * bio happens to be merged with someone else, and may change bi_dev and
2845 * bi_sector for remaps as it sees fit. So the values of these fields
2846 * should NOT be depended on after the call to generic_make_request.
2847 */
2848void generic_make_request(struct bio *bio)
2849{
2850 request_queue_t *q;
2851 sector_t maxsector;
2852 int ret, nr_sectors = bio_sectors(bio);
2853
2854 might_sleep();
2855 /* Test device or partition size, when known. */
2856 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
2857 if (maxsector) {
2858 sector_t sector = bio->bi_sector;
2859
2860 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
2861 /*
2862 * This may well happen - the kernel calls bread()
2863 * without checking the size of the device, e.g., when
2864 * mounting a device.
2865 */
2866 handle_bad_sector(bio);
2867 goto end_io;
2868 }
2869 }
2870
2871 /*
2872 * Resolve the mapping until finished. (drivers are
2873 * still free to implement/resolve their own stacking
2874 * by explicitly returning 0)
2875 *
2876 * NOTE: we don't repeat the blk_size check for each new device.
2877 * Stacking drivers are expected to know what they are doing.
2878 */
2879 do {
2880 char b[BDEVNAME_SIZE];
2881
2882 q = bdev_get_queue(bio->bi_bdev);
2883 if (!q) {
2884 printk(KERN_ERR
2885 "generic_make_request: Trying to access "
2886 "nonexistent block-device %s (%Lu)\n",
2887 bdevname(bio->bi_bdev, b),
2888 (long long) bio->bi_sector);
2889end_io:
2890 bio_endio(bio, bio->bi_size, -EIO);
2891 break;
2892 }
2893
2894 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
2895 printk("bio too big device %s (%u > %u)\n",
2896 bdevname(bio->bi_bdev, b),
2897 bio_sectors(bio),
2898 q->max_hw_sectors);
2899 goto end_io;
2900 }
2901
Nick Pigginfde6ad22005-06-23 00:08:53 -07002902 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903 goto end_io;
2904
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 /*
2906 * If this device has partitions, remap block n
2907 * of partition p to block n+start(p) of the disk.
2908 */
2909 blk_partition_remap(bio);
2910
2911 ret = q->make_request_fn(q, bio);
2912 } while (ret);
2913}
2914
2915EXPORT_SYMBOL(generic_make_request);
2916
2917/**
2918 * submit_bio: submit a bio to the block device layer for I/O
2919 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
2920 * @bio: The &struct bio which describes the I/O
2921 *
2922 * submit_bio() is very similar in purpose to generic_make_request(), and
2923 * uses that function to do most of the work. Both are fairly rough
2924 * interfaces, @bio must be presetup and ready for I/O.
2925 *
2926 */
2927void submit_bio(int rw, struct bio *bio)
2928{
2929 int count = bio_sectors(bio);
2930
2931 BIO_BUG_ON(!bio->bi_size);
2932 BIO_BUG_ON(!bio->bi_io_vec);
Jens Axboe22e2c502005-06-27 10:55:12 +02002933 bio->bi_rw |= rw;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934 if (rw & WRITE)
2935 mod_page_state(pgpgout, count);
2936 else
2937 mod_page_state(pgpgin, count);
2938
2939 if (unlikely(block_dump)) {
2940 char b[BDEVNAME_SIZE];
2941 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
2942 current->comm, current->pid,
2943 (rw & WRITE) ? "WRITE" : "READ",
2944 (unsigned long long)bio->bi_sector,
2945 bdevname(bio->bi_bdev,b));
2946 }
2947
2948 generic_make_request(bio);
2949}
2950
2951EXPORT_SYMBOL(submit_bio);
2952
Adrian Bunk93d17d32005-06-25 14:59:10 -07002953static void blk_recalc_rq_segments(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954{
2955 struct bio *bio, *prevbio = NULL;
2956 int nr_phys_segs, nr_hw_segs;
2957 unsigned int phys_size, hw_size;
2958 request_queue_t *q = rq->q;
2959
2960 if (!rq->bio)
2961 return;
2962
2963 phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
2964 rq_for_each_bio(bio, rq) {
2965 /* Force bio hw/phys segs to be recalculated. */
2966 bio->bi_flags &= ~(1 << BIO_SEG_VALID);
2967
2968 nr_phys_segs += bio_phys_segments(q, bio);
2969 nr_hw_segs += bio_hw_segments(q, bio);
2970 if (prevbio) {
2971 int pseg = phys_size + prevbio->bi_size + bio->bi_size;
2972 int hseg = hw_size + prevbio->bi_size + bio->bi_size;
2973
2974 if (blk_phys_contig_segment(q, prevbio, bio) &&
2975 pseg <= q->max_segment_size) {
2976 nr_phys_segs--;
2977 phys_size += prevbio->bi_size + bio->bi_size;
2978 } else
2979 phys_size = 0;
2980
2981 if (blk_hw_contig_segment(q, prevbio, bio) &&
2982 hseg <= q->max_segment_size) {
2983 nr_hw_segs--;
2984 hw_size += prevbio->bi_size + bio->bi_size;
2985 } else
2986 hw_size = 0;
2987 }
2988 prevbio = bio;
2989 }
2990
2991 rq->nr_phys_segments = nr_phys_segs;
2992 rq->nr_hw_segments = nr_hw_segs;
2993}
2994
Adrian Bunk93d17d32005-06-25 14:59:10 -07002995static void blk_recalc_rq_sectors(struct request *rq, int nsect)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996{
2997 if (blk_fs_request(rq)) {
2998 rq->hard_sector += nsect;
2999 rq->hard_nr_sectors -= nsect;
3000
3001 /*
3002 * Move the I/O submission pointers ahead if required.
3003 */
3004 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
3005 (rq->sector <= rq->hard_sector)) {
3006 rq->sector = rq->hard_sector;
3007 rq->nr_sectors = rq->hard_nr_sectors;
3008 rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
3009 rq->current_nr_sectors = rq->hard_cur_sectors;
3010 rq->buffer = bio_data(rq->bio);
3011 }
3012
3013 /*
3014 * if total number of sectors is less than the first segment
3015 * size, something has gone terribly wrong
3016 */
3017 if (rq->nr_sectors < rq->current_nr_sectors) {
3018 printk("blk: request botched\n");
3019 rq->nr_sectors = rq->current_nr_sectors;
3020 }
3021 }
3022}
3023
3024static int __end_that_request_first(struct request *req, int uptodate,
3025 int nr_bytes)
3026{
3027 int total_bytes, bio_nbytes, error, next_idx = 0;
3028 struct bio *bio;
3029
3030 /*
3031 * extend uptodate bool to allow < 0 value to be direct io error
3032 */
3033 error = 0;
3034 if (end_io_error(uptodate))
3035 error = !uptodate ? -EIO : uptodate;
3036
3037 /*
3038 * for a REQ_BLOCK_PC request, we want to carry any eventual
3039 * sense key with us all the way through
3040 */
3041 if (!blk_pc_request(req))
3042 req->errors = 0;
3043
3044 if (!uptodate) {
3045 if (blk_fs_request(req) && !(req->flags & REQ_QUIET))
3046 printk("end_request: I/O error, dev %s, sector %llu\n",
3047 req->rq_disk ? req->rq_disk->disk_name : "?",
3048 (unsigned long long)req->sector);
3049 }
3050
3051 total_bytes = bio_nbytes = 0;
3052 while ((bio = req->bio) != NULL) {
3053 int nbytes;
3054
3055 if (nr_bytes >= bio->bi_size) {
3056 req->bio = bio->bi_next;
3057 nbytes = bio->bi_size;
3058 bio_endio(bio, nbytes, error);
3059 next_idx = 0;
3060 bio_nbytes = 0;
3061 } else {
3062 int idx = bio->bi_idx + next_idx;
3063
3064 if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3065 blk_dump_rq_flags(req, "__end_that");
3066 printk("%s: bio idx %d >= vcnt %d\n",
3067 __FUNCTION__,
3068 bio->bi_idx, bio->bi_vcnt);
3069 break;
3070 }
3071
3072 nbytes = bio_iovec_idx(bio, idx)->bv_len;
3073 BIO_BUG_ON(nbytes > bio->bi_size);
3074
3075 /*
3076 * not a complete bvec done
3077 */
3078 if (unlikely(nbytes > nr_bytes)) {
3079 bio_nbytes += nr_bytes;
3080 total_bytes += nr_bytes;
3081 break;
3082 }
3083
3084 /*
3085 * advance to the next vector
3086 */
3087 next_idx++;
3088 bio_nbytes += nbytes;
3089 }
3090
3091 total_bytes += nbytes;
3092 nr_bytes -= nbytes;
3093
3094 if ((bio = req->bio)) {
3095 /*
3096 * end more in this run, or just return 'not-done'
3097 */
3098 if (unlikely(nr_bytes <= 0))
3099 break;
3100 }
3101 }
3102
3103 /*
3104 * completely done
3105 */
3106 if (!req->bio)
3107 return 0;
3108
3109 /*
3110 * if the request wasn't completed, update state
3111 */
3112 if (bio_nbytes) {
3113 bio_endio(bio, bio_nbytes, error);
3114 bio->bi_idx += next_idx;
3115 bio_iovec(bio)->bv_offset += nr_bytes;
3116 bio_iovec(bio)->bv_len -= nr_bytes;
3117 }
3118
3119 blk_recalc_rq_sectors(req, total_bytes >> 9);
3120 blk_recalc_rq_segments(req);
3121 return 1;
3122}
3123
3124/**
3125 * end_that_request_first - end I/O on a request
3126 * @req: the request being processed
3127 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3128 * @nr_sectors: number of sectors to end I/O on
3129 *
3130 * Description:
3131 * Ends I/O on a number of sectors attached to @req, and sets it up
3132 * for the next range of segments (if any) in the cluster.
3133 *
3134 * Return:
3135 * 0 - we are done with this request, call end_that_request_last()
3136 * 1 - still buffers pending for this request
3137 **/
3138int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3139{
3140 return __end_that_request_first(req, uptodate, nr_sectors << 9);
3141}
3142
3143EXPORT_SYMBOL(end_that_request_first);
3144
3145/**
3146 * end_that_request_chunk - end I/O on a request
3147 * @req: the request being processed
3148 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3149 * @nr_bytes: number of bytes to complete
3150 *
3151 * Description:
3152 * Ends I/O on a number of bytes attached to @req, and sets it up
3153 * for the next range of segments (if any). Like end_that_request_first(),
3154 * but deals with bytes instead of sectors.
3155 *
3156 * Return:
3157 * 0 - we are done with this request, call end_that_request_last()
3158 * 1 - still buffers pending for this request
3159 **/
3160int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3161{
3162 return __end_that_request_first(req, uptodate, nr_bytes);
3163}
3164
3165EXPORT_SYMBOL(end_that_request_chunk);
3166
3167/*
3168 * queue lock must be held
3169 */
3170void end_that_request_last(struct request *req)
3171{
3172 struct gendisk *disk = req->rq_disk;
3173
3174 if (unlikely(laptop_mode) && blk_fs_request(req))
3175 laptop_io_completion();
3176
3177 if (disk && blk_fs_request(req)) {
3178 unsigned long duration = jiffies - req->start_time;
3179 switch (rq_data_dir(req)) {
3180 case WRITE:
3181 __disk_stat_inc(disk, writes);
3182 __disk_stat_add(disk, write_ticks, duration);
3183 break;
3184 case READ:
3185 __disk_stat_inc(disk, reads);
3186 __disk_stat_add(disk, read_ticks, duration);
3187 break;
3188 }
3189 disk_round_stats(disk);
3190 disk->in_flight--;
3191 }
3192 if (req->end_io)
3193 req->end_io(req);
3194 else
3195 __blk_put_request(req->q, req);
3196}
3197
3198EXPORT_SYMBOL(end_that_request_last);
3199
3200void end_request(struct request *req, int uptodate)
3201{
3202 if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3203 add_disk_randomness(req->rq_disk);
3204 blkdev_dequeue_request(req);
3205 end_that_request_last(req);
3206 }
3207}
3208
3209EXPORT_SYMBOL(end_request);
3210
3211void blk_rq_bio_prep(request_queue_t *q, struct request *rq, struct bio *bio)
3212{
3213 /* first three bits are identical in rq->flags and bio->bi_rw */
3214 rq->flags |= (bio->bi_rw & 7);
3215
3216 rq->nr_phys_segments = bio_phys_segments(q, bio);
3217 rq->nr_hw_segments = bio_hw_segments(q, bio);
3218 rq->current_nr_sectors = bio_cur_sectors(bio);
3219 rq->hard_cur_sectors = rq->current_nr_sectors;
3220 rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3221 rq->buffer = bio_data(bio);
3222
3223 rq->bio = rq->biotail = bio;
3224}
3225
3226EXPORT_SYMBOL(blk_rq_bio_prep);
3227
3228int kblockd_schedule_work(struct work_struct *work)
3229{
3230 return queue_work(kblockd_workqueue, work);
3231}
3232
3233EXPORT_SYMBOL(kblockd_schedule_work);
3234
3235void kblockd_flush(void)
3236{
3237 flush_workqueue(kblockd_workqueue);
3238}
3239EXPORT_SYMBOL(kblockd_flush);
3240
3241int __init blk_dev_init(void)
3242{
3243 kblockd_workqueue = create_workqueue("kblockd");
3244 if (!kblockd_workqueue)
3245 panic("Failed to create kblockd\n");
3246
3247 request_cachep = kmem_cache_create("blkdev_requests",
3248 sizeof(struct request), 0, SLAB_PANIC, NULL, NULL);
3249
3250 requestq_cachep = kmem_cache_create("blkdev_queue",
3251 sizeof(request_queue_t), 0, SLAB_PANIC, NULL, NULL);
3252
3253 iocontext_cachep = kmem_cache_create("blkdev_ioc",
3254 sizeof(struct io_context), 0, SLAB_PANIC, NULL, NULL);
3255
3256 blk_max_low_pfn = max_low_pfn;
3257 blk_max_pfn = max_pfn;
3258
3259 return 0;
3260}
3261
3262/*
3263 * IO Context helper functions
3264 */
3265void put_io_context(struct io_context *ioc)
3266{
3267 if (ioc == NULL)
3268 return;
3269
3270 BUG_ON(atomic_read(&ioc->refcount) == 0);
3271
3272 if (atomic_dec_and_test(&ioc->refcount)) {
3273 if (ioc->aic && ioc->aic->dtor)
3274 ioc->aic->dtor(ioc->aic);
3275 if (ioc->cic && ioc->cic->dtor)
3276 ioc->cic->dtor(ioc->cic);
3277
3278 kmem_cache_free(iocontext_cachep, ioc);
3279 }
3280}
3281EXPORT_SYMBOL(put_io_context);
3282
3283/* Called by the exitting task */
3284void exit_io_context(void)
3285{
3286 unsigned long flags;
3287 struct io_context *ioc;
3288
3289 local_irq_save(flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02003290 task_lock(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291 ioc = current->io_context;
3292 current->io_context = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02003293 ioc->task = NULL;
3294 task_unlock(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003295 local_irq_restore(flags);
3296
3297 if (ioc->aic && ioc->aic->exit)
3298 ioc->aic->exit(ioc->aic);
3299 if (ioc->cic && ioc->cic->exit)
3300 ioc->cic->exit(ioc->cic);
3301
3302 put_io_context(ioc);
3303}
3304
3305/*
3306 * If the current task has no IO context then create one and initialise it.
Nick Pigginfb3cc432005-06-28 20:45:15 -07003307 * Otherwise, return its existing IO context.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003308 *
Nick Pigginfb3cc432005-06-28 20:45:15 -07003309 * This returned IO context doesn't have a specifically elevated refcount,
3310 * but since the current task itself holds a reference, the context can be
3311 * used in general code, so long as it stays within `current` context.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312 */
Nick Pigginfb3cc432005-06-28 20:45:15 -07003313struct io_context *current_io_context(int gfp_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314{
3315 struct task_struct *tsk = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003316 struct io_context *ret;
3317
Linus Torvalds1da177e2005-04-16 15:20:36 -07003318 ret = tsk->io_context;
Nick Pigginfb3cc432005-06-28 20:45:15 -07003319 if (likely(ret))
3320 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003321
3322 ret = kmem_cache_alloc(iocontext_cachep, gfp_flags);
3323 if (ret) {
3324 atomic_set(&ret->refcount, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003325 ret->task = current;
3326 ret->set_ioprio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003327 ret->last_waited = jiffies; /* doesn't matter... */
3328 ret->nr_batch_requests = 0; /* because this is 0 */
3329 ret->aic = NULL;
3330 ret->cic = NULL;
Nick Pigginfb3cc432005-06-28 20:45:15 -07003331 tsk->io_context = ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003332 }
3333
3334 return ret;
3335}
Nick Pigginfb3cc432005-06-28 20:45:15 -07003336EXPORT_SYMBOL(current_io_context);
3337
3338/*
3339 * If the current task has no IO context then create one and initialise it.
3340 * If it does have a context, take a ref on it.
3341 *
3342 * This is always called in the context of the task which submitted the I/O.
3343 */
3344struct io_context *get_io_context(int gfp_flags)
3345{
3346 struct io_context *ret;
3347 ret = current_io_context(gfp_flags);
3348 if (likely(ret))
3349 atomic_inc(&ret->refcount);
3350 return ret;
3351}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003352EXPORT_SYMBOL(get_io_context);
3353
3354void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3355{
3356 struct io_context *src = *psrc;
3357 struct io_context *dst = *pdst;
3358
3359 if (src) {
3360 BUG_ON(atomic_read(&src->refcount) == 0);
3361 atomic_inc(&src->refcount);
3362 put_io_context(dst);
3363 *pdst = src;
3364 }
3365}
3366EXPORT_SYMBOL(copy_io_context);
3367
3368void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3369{
3370 struct io_context *temp;
3371 temp = *ioc1;
3372 *ioc1 = *ioc2;
3373 *ioc2 = temp;
3374}
3375EXPORT_SYMBOL(swap_io_context);
3376
3377/*
3378 * sysfs parts below
3379 */
3380struct queue_sysfs_entry {
3381 struct attribute attr;
3382 ssize_t (*show)(struct request_queue *, char *);
3383 ssize_t (*store)(struct request_queue *, const char *, size_t);
3384};
3385
3386static ssize_t
3387queue_var_show(unsigned int var, char *page)
3388{
3389 return sprintf(page, "%d\n", var);
3390}
3391
3392static ssize_t
3393queue_var_store(unsigned long *var, const char *page, size_t count)
3394{
3395 char *p = (char *) page;
3396
3397 *var = simple_strtoul(p, &p, 10);
3398 return count;
3399}
3400
3401static ssize_t queue_requests_show(struct request_queue *q, char *page)
3402{
3403 return queue_var_show(q->nr_requests, (page));
3404}
3405
3406static ssize_t
3407queue_requests_store(struct request_queue *q, const char *page, size_t count)
3408{
3409 struct request_list *rl = &q->rq;
3410
3411 int ret = queue_var_store(&q->nr_requests, page, count);
3412 if (q->nr_requests < BLKDEV_MIN_RQ)
3413 q->nr_requests = BLKDEV_MIN_RQ;
3414 blk_queue_congestion_threshold(q);
3415
3416 if (rl->count[READ] >= queue_congestion_on_threshold(q))
3417 set_queue_congested(q, READ);
3418 else if (rl->count[READ] < queue_congestion_off_threshold(q))
3419 clear_queue_congested(q, READ);
3420
3421 if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3422 set_queue_congested(q, WRITE);
3423 else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3424 clear_queue_congested(q, WRITE);
3425
3426 if (rl->count[READ] >= q->nr_requests) {
3427 blk_set_queue_full(q, READ);
3428 } else if (rl->count[READ]+1 <= q->nr_requests) {
3429 blk_clear_queue_full(q, READ);
3430 wake_up(&rl->wait[READ]);
3431 }
3432
3433 if (rl->count[WRITE] >= q->nr_requests) {
3434 blk_set_queue_full(q, WRITE);
3435 } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3436 blk_clear_queue_full(q, WRITE);
3437 wake_up(&rl->wait[WRITE]);
3438 }
3439 return ret;
3440}
3441
3442static ssize_t queue_ra_show(struct request_queue *q, char *page)
3443{
3444 int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3445
3446 return queue_var_show(ra_kb, (page));
3447}
3448
3449static ssize_t
3450queue_ra_store(struct request_queue *q, const char *page, size_t count)
3451{
3452 unsigned long ra_kb;
3453 ssize_t ret = queue_var_store(&ra_kb, page, count);
3454
3455 spin_lock_irq(q->queue_lock);
3456 if (ra_kb > (q->max_sectors >> 1))
3457 ra_kb = (q->max_sectors >> 1);
3458
3459 q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3460 spin_unlock_irq(q->queue_lock);
3461
3462 return ret;
3463}
3464
3465static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3466{
3467 int max_sectors_kb = q->max_sectors >> 1;
3468
3469 return queue_var_show(max_sectors_kb, (page));
3470}
3471
3472static ssize_t
3473queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3474{
3475 unsigned long max_sectors_kb,
3476 max_hw_sectors_kb = q->max_hw_sectors >> 1,
3477 page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3478 ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3479 int ra_kb;
3480
3481 if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3482 return -EINVAL;
3483 /*
3484 * Take the queue lock to update the readahead and max_sectors
3485 * values synchronously:
3486 */
3487 spin_lock_irq(q->queue_lock);
3488 /*
3489 * Trim readahead window as well, if necessary:
3490 */
3491 ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3492 if (ra_kb > max_sectors_kb)
3493 q->backing_dev_info.ra_pages =
3494 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3495
3496 q->max_sectors = max_sectors_kb << 1;
3497 spin_unlock_irq(q->queue_lock);
3498
3499 return ret;
3500}
3501
3502static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3503{
3504 int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3505
3506 return queue_var_show(max_hw_sectors_kb, (page));
3507}
3508
3509
3510static struct queue_sysfs_entry queue_requests_entry = {
3511 .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3512 .show = queue_requests_show,
3513 .store = queue_requests_store,
3514};
3515
3516static struct queue_sysfs_entry queue_ra_entry = {
3517 .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3518 .show = queue_ra_show,
3519 .store = queue_ra_store,
3520};
3521
3522static struct queue_sysfs_entry queue_max_sectors_entry = {
3523 .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3524 .show = queue_max_sectors_show,
3525 .store = queue_max_sectors_store,
3526};
3527
3528static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3529 .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
3530 .show = queue_max_hw_sectors_show,
3531};
3532
3533static struct queue_sysfs_entry queue_iosched_entry = {
3534 .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
3535 .show = elv_iosched_show,
3536 .store = elv_iosched_store,
3537};
3538
3539static struct attribute *default_attrs[] = {
3540 &queue_requests_entry.attr,
3541 &queue_ra_entry.attr,
3542 &queue_max_hw_sectors_entry.attr,
3543 &queue_max_sectors_entry.attr,
3544 &queue_iosched_entry.attr,
3545 NULL,
3546};
3547
3548#define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
3549
3550static ssize_t
3551queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
3552{
3553 struct queue_sysfs_entry *entry = to_queue(attr);
3554 struct request_queue *q;
3555
3556 q = container_of(kobj, struct request_queue, kobj);
3557 if (!entry->show)
Dmitry Torokhov6c1852a2005-04-29 01:26:06 -05003558 return -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003559
3560 return entry->show(q, page);
3561}
3562
3563static ssize_t
3564queue_attr_store(struct kobject *kobj, struct attribute *attr,
3565 const char *page, size_t length)
3566{
3567 struct queue_sysfs_entry *entry = to_queue(attr);
3568 struct request_queue *q;
3569
3570 q = container_of(kobj, struct request_queue, kobj);
3571 if (!entry->store)
Dmitry Torokhov6c1852a2005-04-29 01:26:06 -05003572 return -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573
3574 return entry->store(q, page, length);
3575}
3576
3577static struct sysfs_ops queue_sysfs_ops = {
3578 .show = queue_attr_show,
3579 .store = queue_attr_store,
3580};
3581
Adrian Bunk93d17d32005-06-25 14:59:10 -07003582static struct kobj_type queue_ktype = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583 .sysfs_ops = &queue_sysfs_ops,
3584 .default_attrs = default_attrs,
3585};
3586
3587int blk_register_queue(struct gendisk *disk)
3588{
3589 int ret;
3590
3591 request_queue_t *q = disk->queue;
3592
3593 if (!q || !q->request_fn)
3594 return -ENXIO;
3595
3596 q->kobj.parent = kobject_get(&disk->kobj);
3597 if (!q->kobj.parent)
3598 return -EBUSY;
3599
3600 snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
3601 q->kobj.ktype = &queue_ktype;
3602
3603 ret = kobject_register(&q->kobj);
3604 if (ret < 0)
3605 return ret;
3606
3607 ret = elv_register_queue(q);
3608 if (ret) {
3609 kobject_unregister(&q->kobj);
3610 return ret;
3611 }
3612
3613 return 0;
3614}
3615
3616void blk_unregister_queue(struct gendisk *disk)
3617{
3618 request_queue_t *q = disk->queue;
3619
3620 if (q && q->request_fn) {
3621 elv_unregister_queue(q);
3622
3623 kobject_unregister(&q->kobj);
3624 kobject_put(&disk->kobj);
3625 }
3626}