blob: 234fdcfbdf01105e9337e164f34a07dd7dfede5b [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;
238 q->max_phys_segments = MAX_PHYS_SEGMENTS;
239 q->max_hw_segments = MAX_HW_SEGMENTS;
240 q->make_request_fn = mfn;
241 q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
242 q->backing_dev_info.state = 0;
243 q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
244 blk_queue_max_sectors(q, MAX_SECTORS);
245 blk_queue_hardsect_size(q, 512);
246 blk_queue_dma_alignment(q, 511);
247 blk_queue_congestion_threshold(q);
248 q->nr_batching = BLK_BATCH_REQ;
249
250 q->unplug_thresh = 4; /* hmm */
251 q->unplug_delay = (3 * HZ) / 1000; /* 3 milliseconds */
252 if (q->unplug_delay == 0)
253 q->unplug_delay = 1;
254
255 INIT_WORK(&q->unplug_work, blk_unplug_work, q);
256
257 q->unplug_timer.function = blk_unplug_timeout;
258 q->unplug_timer.data = (unsigned long)q;
259
260 /*
261 * by default assume old behaviour and bounce for any highmem page
262 */
263 blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
264
265 blk_queue_activity_fn(q, NULL, NULL);
266
267 INIT_LIST_HEAD(&q->drain_list);
268}
269
270EXPORT_SYMBOL(blk_queue_make_request);
271
272static inline void rq_init(request_queue_t *q, struct request *rq)
273{
274 INIT_LIST_HEAD(&rq->queuelist);
275
276 rq->errors = 0;
277 rq->rq_status = RQ_ACTIVE;
278 rq->bio = rq->biotail = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +0200279 rq->ioprio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 rq->buffer = NULL;
281 rq->ref_count = 1;
282 rq->q = q;
283 rq->waiting = NULL;
284 rq->special = NULL;
285 rq->data_len = 0;
286 rq->data = NULL;
287 rq->sense = NULL;
288 rq->end_io = NULL;
289 rq->end_io_data = NULL;
290}
291
292/**
293 * blk_queue_ordered - does this queue support ordered writes
294 * @q: the request queue
295 * @flag: see below
296 *
297 * Description:
298 * For journalled file systems, doing ordered writes on a commit
299 * block instead of explicitly doing wait_on_buffer (which is bad
300 * for performance) can be a big win. Block drivers supporting this
301 * feature should call this function and indicate so.
302 *
303 **/
304void blk_queue_ordered(request_queue_t *q, int flag)
305{
306 switch (flag) {
307 case QUEUE_ORDERED_NONE:
308 if (q->flush_rq)
309 kmem_cache_free(request_cachep, q->flush_rq);
310 q->flush_rq = NULL;
311 q->ordered = flag;
312 break;
313 case QUEUE_ORDERED_TAG:
314 q->ordered = flag;
315 break;
316 case QUEUE_ORDERED_FLUSH:
317 q->ordered = flag;
318 if (!q->flush_rq)
319 q->flush_rq = kmem_cache_alloc(request_cachep,
320 GFP_KERNEL);
321 break;
322 default:
323 printk("blk_queue_ordered: bad value %d\n", flag);
324 break;
325 }
326}
327
328EXPORT_SYMBOL(blk_queue_ordered);
329
330/**
331 * blk_queue_issue_flush_fn - set function for issuing a flush
332 * @q: the request queue
333 * @iff: the function to be called issuing the flush
334 *
335 * Description:
336 * If a driver supports issuing a flush command, the support is notified
337 * to the block layer by defining it through this call.
338 *
339 **/
340void blk_queue_issue_flush_fn(request_queue_t *q, issue_flush_fn *iff)
341{
342 q->issue_flush_fn = iff;
343}
344
345EXPORT_SYMBOL(blk_queue_issue_flush_fn);
346
347/*
348 * Cache flushing for ordered writes handling
349 */
350static void blk_pre_flush_end_io(struct request *flush_rq)
351{
352 struct request *rq = flush_rq->end_io_data;
353 request_queue_t *q = rq->q;
354
355 rq->flags |= REQ_BAR_PREFLUSH;
356
357 if (!flush_rq->errors)
358 elv_requeue_request(q, rq);
359 else {
360 q->end_flush_fn(q, flush_rq);
361 clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
362 q->request_fn(q);
363 }
364}
365
366static void blk_post_flush_end_io(struct request *flush_rq)
367{
368 struct request *rq = flush_rq->end_io_data;
369 request_queue_t *q = rq->q;
370
371 rq->flags |= REQ_BAR_POSTFLUSH;
372
373 q->end_flush_fn(q, flush_rq);
374 clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
375 q->request_fn(q);
376}
377
378struct request *blk_start_pre_flush(request_queue_t *q, struct request *rq)
379{
380 struct request *flush_rq = q->flush_rq;
381
382 BUG_ON(!blk_barrier_rq(rq));
383
384 if (test_and_set_bit(QUEUE_FLAG_FLUSH, &q->queue_flags))
385 return NULL;
386
387 rq_init(q, flush_rq);
388 flush_rq->elevator_private = NULL;
389 flush_rq->flags = REQ_BAR_FLUSH;
390 flush_rq->rq_disk = rq->rq_disk;
391 flush_rq->rl = NULL;
392
393 /*
394 * prepare_flush returns 0 if no flush is needed, just mark both
395 * pre and post flush as done in that case
396 */
397 if (!q->prepare_flush_fn(q, flush_rq)) {
398 rq->flags |= REQ_BAR_PREFLUSH | REQ_BAR_POSTFLUSH;
399 clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
400 return rq;
401 }
402
403 /*
404 * some drivers dequeue requests right away, some only after io
405 * completion. make sure the request is dequeued.
406 */
407 if (!list_empty(&rq->queuelist))
408 blkdev_dequeue_request(rq);
409
410 elv_deactivate_request(q, rq);
411
412 flush_rq->end_io_data = rq;
413 flush_rq->end_io = blk_pre_flush_end_io;
414
415 __elv_add_request(q, flush_rq, ELEVATOR_INSERT_FRONT, 0);
416 return flush_rq;
417}
418
419static void blk_start_post_flush(request_queue_t *q, struct request *rq)
420{
421 struct request *flush_rq = q->flush_rq;
422
423 BUG_ON(!blk_barrier_rq(rq));
424
425 rq_init(q, flush_rq);
426 flush_rq->elevator_private = NULL;
427 flush_rq->flags = REQ_BAR_FLUSH;
428 flush_rq->rq_disk = rq->rq_disk;
429 flush_rq->rl = NULL;
430
431 if (q->prepare_flush_fn(q, flush_rq)) {
432 flush_rq->end_io_data = rq;
433 flush_rq->end_io = blk_post_flush_end_io;
434
435 __elv_add_request(q, flush_rq, ELEVATOR_INSERT_FRONT, 0);
436 q->request_fn(q);
437 }
438}
439
440static inline int blk_check_end_barrier(request_queue_t *q, struct request *rq,
441 int sectors)
442{
443 if (sectors > rq->nr_sectors)
444 sectors = rq->nr_sectors;
445
446 rq->nr_sectors -= sectors;
447 return rq->nr_sectors;
448}
449
450static int __blk_complete_barrier_rq(request_queue_t *q, struct request *rq,
451 int sectors, int queue_locked)
452{
453 if (q->ordered != QUEUE_ORDERED_FLUSH)
454 return 0;
455 if (!blk_fs_request(rq) || !blk_barrier_rq(rq))
456 return 0;
457 if (blk_barrier_postflush(rq))
458 return 0;
459
460 if (!blk_check_end_barrier(q, rq, sectors)) {
461 unsigned long flags = 0;
462
463 if (!queue_locked)
464 spin_lock_irqsave(q->queue_lock, flags);
465
466 blk_start_post_flush(q, rq);
467
468 if (!queue_locked)
469 spin_unlock_irqrestore(q->queue_lock, flags);
470 }
471
472 return 1;
473}
474
475/**
476 * blk_complete_barrier_rq - complete possible barrier request
477 * @q: the request queue for the device
478 * @rq: the request
479 * @sectors: number of sectors to complete
480 *
481 * Description:
482 * Used in driver end_io handling to determine whether to postpone
483 * completion of a barrier request until a post flush has been done. This
484 * is the unlocked variant, used if the caller doesn't already hold the
485 * queue lock.
486 **/
487int blk_complete_barrier_rq(request_queue_t *q, struct request *rq, int sectors)
488{
489 return __blk_complete_barrier_rq(q, rq, sectors, 0);
490}
491EXPORT_SYMBOL(blk_complete_barrier_rq);
492
493/**
494 * blk_complete_barrier_rq_locked - complete possible barrier request
495 * @q: the request queue for the device
496 * @rq: the request
497 * @sectors: number of sectors to complete
498 *
499 * Description:
500 * See blk_complete_barrier_rq(). This variant must be used if the caller
501 * holds the queue lock.
502 **/
503int blk_complete_barrier_rq_locked(request_queue_t *q, struct request *rq,
504 int sectors)
505{
506 return __blk_complete_barrier_rq(q, rq, sectors, 1);
507}
508EXPORT_SYMBOL(blk_complete_barrier_rq_locked);
509
510/**
511 * blk_queue_bounce_limit - set bounce buffer limit for queue
512 * @q: the request queue for the device
513 * @dma_addr: bus address limit
514 *
515 * Description:
516 * Different hardware can have different requirements as to what pages
517 * it can do I/O directly to. A low level driver can call
518 * blk_queue_bounce_limit to have lower memory pages allocated as bounce
519 * buffers for doing I/O to pages residing above @page. By default
520 * the block layer sets this to the highest numbered "low" memory page.
521 **/
522void blk_queue_bounce_limit(request_queue_t *q, u64 dma_addr)
523{
524 unsigned long bounce_pfn = dma_addr >> PAGE_SHIFT;
525
526 /*
527 * set appropriate bounce gfp mask -- unfortunately we don't have a
528 * full 4GB zone, so we have to resort to low memory for any bounces.
529 * ISA has its own < 16MB zone.
530 */
531 if (bounce_pfn < blk_max_low_pfn) {
532 BUG_ON(dma_addr < BLK_BOUNCE_ISA);
533 init_emergency_isa_pool();
534 q->bounce_gfp = GFP_NOIO | GFP_DMA;
535 } else
536 q->bounce_gfp = GFP_NOIO;
537
538 q->bounce_pfn = bounce_pfn;
539}
540
541EXPORT_SYMBOL(blk_queue_bounce_limit);
542
543/**
544 * blk_queue_max_sectors - set max sectors for a request for this queue
545 * @q: the request queue for the device
546 * @max_sectors: max sectors in the usual 512b unit
547 *
548 * Description:
549 * Enables a low level driver to set an upper limit on the size of
550 * received requests.
551 **/
552void blk_queue_max_sectors(request_queue_t *q, unsigned short max_sectors)
553{
554 if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
555 max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
556 printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
557 }
558
559 q->max_sectors = q->max_hw_sectors = max_sectors;
560}
561
562EXPORT_SYMBOL(blk_queue_max_sectors);
563
564/**
565 * blk_queue_max_phys_segments - set max phys segments for a request for this queue
566 * @q: the request queue for the device
567 * @max_segments: max number of segments
568 *
569 * Description:
570 * Enables a low level driver to set an upper limit on the number of
571 * physical data segments in a request. This would be the largest sized
572 * scatter list the driver could handle.
573 **/
574void blk_queue_max_phys_segments(request_queue_t *q, unsigned short max_segments)
575{
576 if (!max_segments) {
577 max_segments = 1;
578 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
579 }
580
581 q->max_phys_segments = max_segments;
582}
583
584EXPORT_SYMBOL(blk_queue_max_phys_segments);
585
586/**
587 * blk_queue_max_hw_segments - set max hw segments for a request for this queue
588 * @q: the request queue for the device
589 * @max_segments: max number of segments
590 *
591 * Description:
592 * Enables a low level driver to set an upper limit on the number of
593 * hw data segments in a request. This would be the largest number of
594 * address/length pairs the host adapter can actually give as once
595 * to the device.
596 **/
597void blk_queue_max_hw_segments(request_queue_t *q, unsigned short max_segments)
598{
599 if (!max_segments) {
600 max_segments = 1;
601 printk("%s: set to minimum %d\n", __FUNCTION__, max_segments);
602 }
603
604 q->max_hw_segments = max_segments;
605}
606
607EXPORT_SYMBOL(blk_queue_max_hw_segments);
608
609/**
610 * blk_queue_max_segment_size - set max segment size for blk_rq_map_sg
611 * @q: the request queue for the device
612 * @max_size: max size of segment in bytes
613 *
614 * Description:
615 * Enables a low level driver to set an upper limit on the size of a
616 * coalesced segment
617 **/
618void blk_queue_max_segment_size(request_queue_t *q, unsigned int max_size)
619{
620 if (max_size < PAGE_CACHE_SIZE) {
621 max_size = PAGE_CACHE_SIZE;
622 printk("%s: set to minimum %d\n", __FUNCTION__, max_size);
623 }
624
625 q->max_segment_size = max_size;
626}
627
628EXPORT_SYMBOL(blk_queue_max_segment_size);
629
630/**
631 * blk_queue_hardsect_size - set hardware sector size for the queue
632 * @q: the request queue for the device
633 * @size: the hardware sector size, in bytes
634 *
635 * Description:
636 * This should typically be set to the lowest possible sector size
637 * that the hardware can operate on (possible without reverting to
638 * even internal read-modify-write operations). Usually the default
639 * of 512 covers most hardware.
640 **/
641void blk_queue_hardsect_size(request_queue_t *q, unsigned short size)
642{
643 q->hardsect_size = size;
644}
645
646EXPORT_SYMBOL(blk_queue_hardsect_size);
647
648/*
649 * Returns the minimum that is _not_ zero, unless both are zero.
650 */
651#define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
652
653/**
654 * blk_queue_stack_limits - inherit underlying queue limits for stacked drivers
655 * @t: the stacking driver (top)
656 * @b: the underlying device (bottom)
657 **/
658void blk_queue_stack_limits(request_queue_t *t, request_queue_t *b)
659{
660 /* zero is "infinity" */
661 t->max_sectors = t->max_hw_sectors =
662 min_not_zero(t->max_sectors,b->max_sectors);
663
664 t->max_phys_segments = min(t->max_phys_segments,b->max_phys_segments);
665 t->max_hw_segments = min(t->max_hw_segments,b->max_hw_segments);
666 t->max_segment_size = min(t->max_segment_size,b->max_segment_size);
667 t->hardsect_size = max(t->hardsect_size,b->hardsect_size);
668}
669
670EXPORT_SYMBOL(blk_queue_stack_limits);
671
672/**
673 * blk_queue_segment_boundary - set boundary rules for segment merging
674 * @q: the request queue for the device
675 * @mask: the memory boundary mask
676 **/
677void blk_queue_segment_boundary(request_queue_t *q, unsigned long mask)
678{
679 if (mask < PAGE_CACHE_SIZE - 1) {
680 mask = PAGE_CACHE_SIZE - 1;
681 printk("%s: set to minimum %lx\n", __FUNCTION__, mask);
682 }
683
684 q->seg_boundary_mask = mask;
685}
686
687EXPORT_SYMBOL(blk_queue_segment_boundary);
688
689/**
690 * blk_queue_dma_alignment - set dma length and memory alignment
691 * @q: the request queue for the device
692 * @mask: alignment mask
693 *
694 * description:
695 * set required memory and length aligment for direct dma transactions.
696 * this is used when buiding direct io requests for the queue.
697 *
698 **/
699void blk_queue_dma_alignment(request_queue_t *q, int mask)
700{
701 q->dma_alignment = mask;
702}
703
704EXPORT_SYMBOL(blk_queue_dma_alignment);
705
706/**
707 * blk_queue_find_tag - find a request by its tag and queue
708 *
709 * @q: The request queue for the device
710 * @tag: The tag of the request
711 *
712 * Notes:
713 * Should be used when a device returns a tag and you want to match
714 * it with a request.
715 *
716 * no locks need be held.
717 **/
718struct request *blk_queue_find_tag(request_queue_t *q, int tag)
719{
720 struct blk_queue_tag *bqt = q->queue_tags;
721
Tejun Heofa72b902005-06-23 00:08:49 -0700722 if (unlikely(bqt == NULL || tag >= bqt->max_depth))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 return NULL;
724
725 return bqt->tag_index[tag];
726}
727
728EXPORT_SYMBOL(blk_queue_find_tag);
729
730/**
731 * __blk_queue_free_tags - release tag maintenance info
732 * @q: the request queue for the device
733 *
734 * Notes:
735 * blk_cleanup_queue() will take care of calling this function, if tagging
736 * has been used. So there's no need to call this directly.
737 **/
738static void __blk_queue_free_tags(request_queue_t *q)
739{
740 struct blk_queue_tag *bqt = q->queue_tags;
741
742 if (!bqt)
743 return;
744
745 if (atomic_dec_and_test(&bqt->refcnt)) {
746 BUG_ON(bqt->busy);
747 BUG_ON(!list_empty(&bqt->busy_list));
748
749 kfree(bqt->tag_index);
750 bqt->tag_index = NULL;
751
752 kfree(bqt->tag_map);
753 bqt->tag_map = NULL;
754
755 kfree(bqt);
756 }
757
758 q->queue_tags = NULL;
759 q->queue_flags &= ~(1 << QUEUE_FLAG_QUEUED);
760}
761
762/**
763 * blk_queue_free_tags - release tag maintenance info
764 * @q: the request queue for the device
765 *
766 * Notes:
767 * This is used to disabled tagged queuing to a device, yet leave
768 * queue in function.
769 **/
770void blk_queue_free_tags(request_queue_t *q)
771{
772 clear_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
773}
774
775EXPORT_SYMBOL(blk_queue_free_tags);
776
777static int
778init_tag_map(request_queue_t *q, struct blk_queue_tag *tags, int depth)
779{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 struct request **tag_index;
781 unsigned long *tag_map;
Tejun Heofa72b902005-06-23 00:08:49 -0700782 int nr_ulongs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783
784 if (depth > q->nr_requests * 2) {
785 depth = q->nr_requests * 2;
786 printk(KERN_ERR "%s: adjusted depth to %d\n",
787 __FUNCTION__, depth);
788 }
789
790 tag_index = kmalloc(depth * sizeof(struct request *), GFP_ATOMIC);
791 if (!tag_index)
792 goto fail;
793
Tejun Heof7d37d02005-06-23 00:08:50 -0700794 nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
Tejun Heofa72b902005-06-23 00:08:49 -0700795 tag_map = kmalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796 if (!tag_map)
797 goto fail;
798
799 memset(tag_index, 0, depth * sizeof(struct request *));
Tejun Heofa72b902005-06-23 00:08:49 -0700800 memset(tag_map, 0, nr_ulongs * sizeof(unsigned long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 tags->max_depth = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 tags->tag_index = tag_index;
803 tags->tag_map = tag_map;
804
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 return 0;
806fail:
807 kfree(tag_index);
808 return -ENOMEM;
809}
810
811/**
812 * blk_queue_init_tags - initialize the queue tag info
813 * @q: the request queue for the device
814 * @depth: the maximum queue depth supported
815 * @tags: the tag to use
816 **/
817int blk_queue_init_tags(request_queue_t *q, int depth,
818 struct blk_queue_tag *tags)
819{
820 int rc;
821
822 BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
823
824 if (!tags && !q->queue_tags) {
825 tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
826 if (!tags)
827 goto fail;
828
829 if (init_tag_map(q, tags, depth))
830 goto fail;
831
832 INIT_LIST_HEAD(&tags->busy_list);
833 tags->busy = 0;
834 atomic_set(&tags->refcnt, 1);
835 } else if (q->queue_tags) {
836 if ((rc = blk_queue_resize_tags(q, depth)))
837 return rc;
838 set_bit(QUEUE_FLAG_QUEUED, &q->queue_flags);
839 return 0;
840 } else
841 atomic_inc(&tags->refcnt);
842
843 /*
844 * assign it, all done
845 */
846 q->queue_tags = tags;
847 q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
848 return 0;
849fail:
850 kfree(tags);
851 return -ENOMEM;
852}
853
854EXPORT_SYMBOL(blk_queue_init_tags);
855
856/**
857 * blk_queue_resize_tags - change the queueing depth
858 * @q: the request queue for the device
859 * @new_depth: the new max command queueing depth
860 *
861 * Notes:
862 * Must be called with the queue lock held.
863 **/
864int blk_queue_resize_tags(request_queue_t *q, int new_depth)
865{
866 struct blk_queue_tag *bqt = q->queue_tags;
867 struct request **tag_index;
868 unsigned long *tag_map;
Tejun Heofa72b902005-06-23 00:08:49 -0700869 int max_depth, nr_ulongs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870
871 if (!bqt)
872 return -ENXIO;
873
874 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875 * save the old state info, so we can copy it back
876 */
877 tag_index = bqt->tag_index;
878 tag_map = bqt->tag_map;
Tejun Heofa72b902005-06-23 00:08:49 -0700879 max_depth = bqt->max_depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880
881 if (init_tag_map(q, bqt, new_depth))
882 return -ENOMEM;
883
884 memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
Tejun Heof7d37d02005-06-23 00:08:50 -0700885 nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
Tejun Heofa72b902005-06-23 00:08:49 -0700886 memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887
888 kfree(tag_index);
889 kfree(tag_map);
890 return 0;
891}
892
893EXPORT_SYMBOL(blk_queue_resize_tags);
894
895/**
896 * blk_queue_end_tag - end tag operations for a request
897 * @q: the request queue for the device
898 * @rq: the request that has completed
899 *
900 * Description:
901 * Typically called when end_that_request_first() returns 0, meaning
902 * all transfers have been done for a request. It's important to call
903 * this function before end_that_request_last(), as that will put the
904 * request back on the free list thus corrupting the internal tag list.
905 *
906 * Notes:
907 * queue lock must be held.
908 **/
909void blk_queue_end_tag(request_queue_t *q, struct request *rq)
910{
911 struct blk_queue_tag *bqt = q->queue_tags;
912 int tag = rq->tag;
913
914 BUG_ON(tag == -1);
915
Tejun Heofa72b902005-06-23 00:08:49 -0700916 if (unlikely(tag >= bqt->max_depth))
Tejun Heo040c9282005-06-23 00:08:51 -0700917 /*
918 * This can happen after tag depth has been reduced.
919 * FIXME: how about a warning or info message here?
920 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 return;
922
923 if (unlikely(!__test_and_clear_bit(tag, bqt->tag_map))) {
Tejun Heo040c9282005-06-23 00:08:51 -0700924 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
925 __FUNCTION__, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926 return;
927 }
928
929 list_del_init(&rq->queuelist);
930 rq->flags &= ~REQ_QUEUED;
931 rq->tag = -1;
932
933 if (unlikely(bqt->tag_index[tag] == NULL))
Tejun Heo040c9282005-06-23 00:08:51 -0700934 printk(KERN_ERR "%s: tag %d is missing\n",
935 __FUNCTION__, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936
937 bqt->tag_index[tag] = NULL;
938 bqt->busy--;
939}
940
941EXPORT_SYMBOL(blk_queue_end_tag);
942
943/**
944 * blk_queue_start_tag - find a free tag and assign it
945 * @q: the request queue for the device
946 * @rq: the block request that needs tagging
947 *
948 * Description:
949 * This can either be used as a stand-alone helper, or possibly be
950 * assigned as the queue &prep_rq_fn (in which case &struct request
951 * automagically gets a tag assigned). Note that this function
952 * assumes that any type of request can be queued! if this is not
953 * true for your device, you must check the request type before
954 * calling this function. The request will also be removed from
955 * the request queue, so it's the drivers responsibility to readd
956 * it if it should need to be restarted for some reason.
957 *
958 * Notes:
959 * queue lock must be held.
960 **/
961int blk_queue_start_tag(request_queue_t *q, struct request *rq)
962{
963 struct blk_queue_tag *bqt = q->queue_tags;
Tejun Heo2bf0fda2005-06-23 00:08:48 -0700964 int tag;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965
966 if (unlikely((rq->flags & REQ_QUEUED))) {
967 printk(KERN_ERR
Tejun Heo040c9282005-06-23 00:08:51 -0700968 "%s: request %p for device [%s] already tagged %d",
969 __FUNCTION__, rq,
970 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971 BUG();
972 }
973
Tejun Heo2bf0fda2005-06-23 00:08:48 -0700974 tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
975 if (tag >= bqt->max_depth)
976 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978 __set_bit(tag, bqt->tag_map);
979
980 rq->flags |= REQ_QUEUED;
981 rq->tag = tag;
982 bqt->tag_index[tag] = rq;
983 blkdev_dequeue_request(rq);
984 list_add(&rq->queuelist, &bqt->busy_list);
985 bqt->busy++;
986 return 0;
987}
988
989EXPORT_SYMBOL(blk_queue_start_tag);
990
991/**
992 * blk_queue_invalidate_tags - invalidate all pending tags
993 * @q: the request queue for the device
994 *
995 * Description:
996 * Hardware conditions may dictate a need to stop all pending requests.
997 * In this case, we will safely clear the block side of the tag queue and
998 * readd all requests to the request queue in the right order.
999 *
1000 * Notes:
1001 * queue lock must be held.
1002 **/
1003void blk_queue_invalidate_tags(request_queue_t *q)
1004{
1005 struct blk_queue_tag *bqt = q->queue_tags;
1006 struct list_head *tmp, *n;
1007 struct request *rq;
1008
1009 list_for_each_safe(tmp, n, &bqt->busy_list) {
1010 rq = list_entry_rq(tmp);
1011
1012 if (rq->tag == -1) {
Tejun Heo040c9282005-06-23 00:08:51 -07001013 printk(KERN_ERR
1014 "%s: bad tag found on list\n", __FUNCTION__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015 list_del_init(&rq->queuelist);
1016 rq->flags &= ~REQ_QUEUED;
1017 } else
1018 blk_queue_end_tag(q, rq);
1019
1020 rq->flags &= ~REQ_STARTED;
1021 __elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 0);
1022 }
1023}
1024
1025EXPORT_SYMBOL(blk_queue_invalidate_tags);
1026
1027static char *rq_flags[] = {
1028 "REQ_RW",
1029 "REQ_FAILFAST",
1030 "REQ_SOFTBARRIER",
1031 "REQ_HARDBARRIER",
1032 "REQ_CMD",
1033 "REQ_NOMERGE",
1034 "REQ_STARTED",
1035 "REQ_DONTPREP",
1036 "REQ_QUEUED",
1037 "REQ_PC",
1038 "REQ_BLOCK_PC",
1039 "REQ_SENSE",
1040 "REQ_FAILED",
1041 "REQ_QUIET",
1042 "REQ_SPECIAL",
1043 "REQ_DRIVE_CMD",
1044 "REQ_DRIVE_TASK",
1045 "REQ_DRIVE_TASKFILE",
1046 "REQ_PREEMPT",
1047 "REQ_PM_SUSPEND",
1048 "REQ_PM_RESUME",
1049 "REQ_PM_SHUTDOWN",
1050};
1051
1052void blk_dump_rq_flags(struct request *rq, char *msg)
1053{
1054 int bit;
1055
1056 printk("%s: dev %s: flags = ", msg,
1057 rq->rq_disk ? rq->rq_disk->disk_name : "?");
1058 bit = 0;
1059 do {
1060 if (rq->flags & (1 << bit))
1061 printk("%s ", rq_flags[bit]);
1062 bit++;
1063 } while (bit < __REQ_NR_BITS);
1064
1065 printk("\nsector %llu, nr/cnr %lu/%u\n", (unsigned long long)rq->sector,
1066 rq->nr_sectors,
1067 rq->current_nr_sectors);
1068 printk("bio %p, biotail %p, buffer %p, data %p, len %u\n", rq->bio, rq->biotail, rq->buffer, rq->data, rq->data_len);
1069
1070 if (rq->flags & (REQ_BLOCK_PC | REQ_PC)) {
1071 printk("cdb: ");
1072 for (bit = 0; bit < sizeof(rq->cmd); bit++)
1073 printk("%02x ", rq->cmd[bit]);
1074 printk("\n");
1075 }
1076}
1077
1078EXPORT_SYMBOL(blk_dump_rq_flags);
1079
1080void blk_recount_segments(request_queue_t *q, struct bio *bio)
1081{
1082 struct bio_vec *bv, *bvprv = NULL;
1083 int i, nr_phys_segs, nr_hw_segs, seg_size, hw_seg_size, cluster;
1084 int high, highprv = 1;
1085
1086 if (unlikely(!bio->bi_io_vec))
1087 return;
1088
1089 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1090 hw_seg_size = seg_size = nr_phys_segs = nr_hw_segs = 0;
1091 bio_for_each_segment(bv, bio, i) {
1092 /*
1093 * the trick here is making sure that a high page is never
1094 * considered part of another segment, since that might
1095 * change with the bounce page.
1096 */
1097 high = page_to_pfn(bv->bv_page) >= q->bounce_pfn;
1098 if (high || highprv)
1099 goto new_hw_segment;
1100 if (cluster) {
1101 if (seg_size + bv->bv_len > q->max_segment_size)
1102 goto new_segment;
1103 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
1104 goto new_segment;
1105 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
1106 goto new_segment;
1107 if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
1108 goto new_hw_segment;
1109
1110 seg_size += bv->bv_len;
1111 hw_seg_size += bv->bv_len;
1112 bvprv = bv;
1113 continue;
1114 }
1115new_segment:
1116 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
1117 !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len)) {
1118 hw_seg_size += bv->bv_len;
1119 } else {
1120new_hw_segment:
1121 if (hw_seg_size > bio->bi_hw_front_size)
1122 bio->bi_hw_front_size = hw_seg_size;
1123 hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
1124 nr_hw_segs++;
1125 }
1126
1127 nr_phys_segs++;
1128 bvprv = bv;
1129 seg_size = bv->bv_len;
1130 highprv = high;
1131 }
1132 if (hw_seg_size > bio->bi_hw_back_size)
1133 bio->bi_hw_back_size = hw_seg_size;
1134 if (nr_hw_segs == 1 && hw_seg_size > bio->bi_hw_front_size)
1135 bio->bi_hw_front_size = hw_seg_size;
1136 bio->bi_phys_segments = nr_phys_segs;
1137 bio->bi_hw_segments = nr_hw_segs;
1138 bio->bi_flags |= (1 << BIO_SEG_VALID);
1139}
1140
1141
Adrian Bunk93d17d32005-06-25 14:59:10 -07001142static int blk_phys_contig_segment(request_queue_t *q, struct bio *bio,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 struct bio *nxt)
1144{
1145 if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
1146 return 0;
1147
1148 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
1149 return 0;
1150 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1151 return 0;
1152
1153 /*
1154 * bio and nxt are contigous in memory, check if the queue allows
1155 * these two to be merged into one
1156 */
1157 if (BIO_SEG_BOUNDARY(q, bio, nxt))
1158 return 1;
1159
1160 return 0;
1161}
1162
Adrian Bunk93d17d32005-06-25 14:59:10 -07001163static int blk_hw_contig_segment(request_queue_t *q, struct bio *bio,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164 struct bio *nxt)
1165{
1166 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1167 blk_recount_segments(q, bio);
1168 if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
1169 blk_recount_segments(q, nxt);
1170 if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
1171 BIOVEC_VIRT_OVERSIZE(bio->bi_hw_front_size + bio->bi_hw_back_size))
1172 return 0;
1173 if (bio->bi_size + nxt->bi_size > q->max_segment_size)
1174 return 0;
1175
1176 return 1;
1177}
1178
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179/*
1180 * map a request to scatterlist, return number of sg entries setup. Caller
1181 * must make sure sg can hold rq->nr_phys_segments entries
1182 */
1183int blk_rq_map_sg(request_queue_t *q, struct request *rq, struct scatterlist *sg)
1184{
1185 struct bio_vec *bvec, *bvprv;
1186 struct bio *bio;
1187 int nsegs, i, cluster;
1188
1189 nsegs = 0;
1190 cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
1191
1192 /*
1193 * for each bio in rq
1194 */
1195 bvprv = NULL;
1196 rq_for_each_bio(bio, rq) {
1197 /*
1198 * for each segment in bio
1199 */
1200 bio_for_each_segment(bvec, bio, i) {
1201 int nbytes = bvec->bv_len;
1202
1203 if (bvprv && cluster) {
1204 if (sg[nsegs - 1].length + nbytes > q->max_segment_size)
1205 goto new_segment;
1206
1207 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
1208 goto new_segment;
1209 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
1210 goto new_segment;
1211
1212 sg[nsegs - 1].length += nbytes;
1213 } else {
1214new_segment:
1215 memset(&sg[nsegs],0,sizeof(struct scatterlist));
1216 sg[nsegs].page = bvec->bv_page;
1217 sg[nsegs].length = nbytes;
1218 sg[nsegs].offset = bvec->bv_offset;
1219
1220 nsegs++;
1221 }
1222 bvprv = bvec;
1223 } /* segments in bio */
1224 } /* bios in rq */
1225
1226 return nsegs;
1227}
1228
1229EXPORT_SYMBOL(blk_rq_map_sg);
1230
1231/*
1232 * the standard queue merge functions, can be overridden with device
1233 * specific ones if so desired
1234 */
1235
1236static inline int ll_new_mergeable(request_queue_t *q,
1237 struct request *req,
1238 struct bio *bio)
1239{
1240 int nr_phys_segs = bio_phys_segments(q, bio);
1241
1242 if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1243 req->flags |= REQ_NOMERGE;
1244 if (req == q->last_merge)
1245 q->last_merge = NULL;
1246 return 0;
1247 }
1248
1249 /*
1250 * A hw segment is just getting larger, bump just the phys
1251 * counter.
1252 */
1253 req->nr_phys_segments += nr_phys_segs;
1254 return 1;
1255}
1256
1257static inline int ll_new_hw_segment(request_queue_t *q,
1258 struct request *req,
1259 struct bio *bio)
1260{
1261 int nr_hw_segs = bio_hw_segments(q, bio);
1262 int nr_phys_segs = bio_phys_segments(q, bio);
1263
1264 if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
1265 || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
1266 req->flags |= REQ_NOMERGE;
1267 if (req == q->last_merge)
1268 q->last_merge = NULL;
1269 return 0;
1270 }
1271
1272 /*
1273 * This will form the start of a new hw segment. Bump both
1274 * counters.
1275 */
1276 req->nr_hw_segments += nr_hw_segs;
1277 req->nr_phys_segments += nr_phys_segs;
1278 return 1;
1279}
1280
1281static int ll_back_merge_fn(request_queue_t *q, struct request *req,
1282 struct bio *bio)
1283{
1284 int len;
1285
1286 if (req->nr_sectors + bio_sectors(bio) > q->max_sectors) {
1287 req->flags |= REQ_NOMERGE;
1288 if (req == q->last_merge)
1289 q->last_merge = NULL;
1290 return 0;
1291 }
1292 if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
1293 blk_recount_segments(q, req->biotail);
1294 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1295 blk_recount_segments(q, bio);
1296 len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
1297 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio)) &&
1298 !BIOVEC_VIRT_OVERSIZE(len)) {
1299 int mergeable = ll_new_mergeable(q, req, bio);
1300
1301 if (mergeable) {
1302 if (req->nr_hw_segments == 1)
1303 req->bio->bi_hw_front_size = len;
1304 if (bio->bi_hw_segments == 1)
1305 bio->bi_hw_back_size = len;
1306 }
1307 return mergeable;
1308 }
1309
1310 return ll_new_hw_segment(q, req, bio);
1311}
1312
1313static int ll_front_merge_fn(request_queue_t *q, struct request *req,
1314 struct bio *bio)
1315{
1316 int len;
1317
1318 if (req->nr_sectors + bio_sectors(bio) > q->max_sectors) {
1319 req->flags |= REQ_NOMERGE;
1320 if (req == q->last_merge)
1321 q->last_merge = NULL;
1322 return 0;
1323 }
1324 len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
1325 if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
1326 blk_recount_segments(q, bio);
1327 if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
1328 blk_recount_segments(q, req->bio);
1329 if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
1330 !BIOVEC_VIRT_OVERSIZE(len)) {
1331 int mergeable = ll_new_mergeable(q, req, bio);
1332
1333 if (mergeable) {
1334 if (bio->bi_hw_segments == 1)
1335 bio->bi_hw_front_size = len;
1336 if (req->nr_hw_segments == 1)
1337 req->biotail->bi_hw_back_size = len;
1338 }
1339 return mergeable;
1340 }
1341
1342 return ll_new_hw_segment(q, req, bio);
1343}
1344
1345static int ll_merge_requests_fn(request_queue_t *q, struct request *req,
1346 struct request *next)
1347{
Nikita Danilovdfa1a552005-06-25 14:59:20 -07001348 int total_phys_segments;
1349 int total_hw_segments;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350
1351 /*
1352 * First check if the either of the requests are re-queued
1353 * requests. Can't merge them if they are.
1354 */
1355 if (req->special || next->special)
1356 return 0;
1357
1358 /*
Nikita Danilovdfa1a552005-06-25 14:59:20 -07001359 * Will it become too large?
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360 */
1361 if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
1362 return 0;
1363
1364 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
1365 if (blk_phys_contig_segment(q, req->biotail, next->bio))
1366 total_phys_segments--;
1367
1368 if (total_phys_segments > q->max_phys_segments)
1369 return 0;
1370
1371 total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
1372 if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
1373 int len = req->biotail->bi_hw_back_size + next->bio->bi_hw_front_size;
1374 /*
1375 * propagate the combined length to the end of the requests
1376 */
1377 if (req->nr_hw_segments == 1)
1378 req->bio->bi_hw_front_size = len;
1379 if (next->nr_hw_segments == 1)
1380 next->biotail->bi_hw_back_size = len;
1381 total_hw_segments--;
1382 }
1383
1384 if (total_hw_segments > q->max_hw_segments)
1385 return 0;
1386
1387 /* Merge is OK... */
1388 req->nr_phys_segments = total_phys_segments;
1389 req->nr_hw_segments = total_hw_segments;
1390 return 1;
1391}
1392
1393/*
1394 * "plug" the device if there are no outstanding requests: this will
1395 * force the transfer to start only after we have put all the requests
1396 * on the list.
1397 *
1398 * This is called with interrupts off and no requests on the queue and
1399 * with the queue lock held.
1400 */
1401void blk_plug_device(request_queue_t *q)
1402{
1403 WARN_ON(!irqs_disabled());
1404
1405 /*
1406 * don't plug a stopped queue, it must be paired with blk_start_queue()
1407 * which will restart the queueing
1408 */
1409 if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
1410 return;
1411
1412 if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1413 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
1414}
1415
1416EXPORT_SYMBOL(blk_plug_device);
1417
1418/*
1419 * remove the queue from the plugged list, if present. called with
1420 * queue lock held and interrupts disabled.
1421 */
1422int blk_remove_plug(request_queue_t *q)
1423{
1424 WARN_ON(!irqs_disabled());
1425
1426 if (!test_and_clear_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
1427 return 0;
1428
1429 del_timer(&q->unplug_timer);
1430 return 1;
1431}
1432
1433EXPORT_SYMBOL(blk_remove_plug);
1434
1435/*
1436 * remove the plug and let it rip..
1437 */
1438void __generic_unplug_device(request_queue_t *q)
1439{
Nick Pigginfde6ad22005-06-23 00:08:53 -07001440 if (unlikely(test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441 return;
1442
1443 if (!blk_remove_plug(q))
1444 return;
1445
Jens Axboe22e2c502005-06-27 10:55:12 +02001446 q->request_fn(q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447}
1448EXPORT_SYMBOL(__generic_unplug_device);
1449
1450/**
1451 * generic_unplug_device - fire a request queue
1452 * @q: The &request_queue_t in question
1453 *
1454 * Description:
1455 * Linux uses plugging to build bigger requests queues before letting
1456 * the device have at them. If a queue is plugged, the I/O scheduler
1457 * is still adding and merging requests on the queue. Once the queue
1458 * gets unplugged, the request_fn defined for the queue is invoked and
1459 * transfers started.
1460 **/
1461void generic_unplug_device(request_queue_t *q)
1462{
1463 spin_lock_irq(q->queue_lock);
1464 __generic_unplug_device(q);
1465 spin_unlock_irq(q->queue_lock);
1466}
1467EXPORT_SYMBOL(generic_unplug_device);
1468
1469static void blk_backing_dev_unplug(struct backing_dev_info *bdi,
1470 struct page *page)
1471{
1472 request_queue_t *q = bdi->unplug_io_data;
1473
1474 /*
1475 * devices don't necessarily have an ->unplug_fn defined
1476 */
1477 if (q->unplug_fn)
1478 q->unplug_fn(q);
1479}
1480
1481static void blk_unplug_work(void *data)
1482{
1483 request_queue_t *q = data;
1484
1485 q->unplug_fn(q);
1486}
1487
1488static void blk_unplug_timeout(unsigned long data)
1489{
1490 request_queue_t *q = (request_queue_t *)data;
1491
1492 kblockd_schedule_work(&q->unplug_work);
1493}
1494
1495/**
1496 * blk_start_queue - restart a previously stopped queue
1497 * @q: The &request_queue_t in question
1498 *
1499 * Description:
1500 * blk_start_queue() will clear the stop flag on the queue, and call
1501 * the request_fn for the queue if it was in a stopped state when
1502 * entered. Also see blk_stop_queue(). Queue lock must be held.
1503 **/
1504void blk_start_queue(request_queue_t *q)
1505{
1506 clear_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1507
1508 /*
1509 * one level of recursion is ok and is much faster than kicking
1510 * the unplug handling
1511 */
1512 if (!test_and_set_bit(QUEUE_FLAG_REENTER, &q->queue_flags)) {
1513 q->request_fn(q);
1514 clear_bit(QUEUE_FLAG_REENTER, &q->queue_flags);
1515 } else {
1516 blk_plug_device(q);
1517 kblockd_schedule_work(&q->unplug_work);
1518 }
1519}
1520
1521EXPORT_SYMBOL(blk_start_queue);
1522
1523/**
1524 * blk_stop_queue - stop a queue
1525 * @q: The &request_queue_t in question
1526 *
1527 * Description:
1528 * The Linux block layer assumes that a block driver will consume all
1529 * entries on the request queue when the request_fn strategy is called.
1530 * Often this will not happen, because of hardware limitations (queue
1531 * depth settings). If a device driver gets a 'queue full' response,
1532 * or if it simply chooses not to queue more I/O at one point, it can
1533 * call this function to prevent the request_fn from being called until
1534 * the driver has signalled it's ready to go again. This happens by calling
1535 * blk_start_queue() to restart queue operations. Queue lock must be held.
1536 **/
1537void blk_stop_queue(request_queue_t *q)
1538{
1539 blk_remove_plug(q);
1540 set_bit(QUEUE_FLAG_STOPPED, &q->queue_flags);
1541}
1542EXPORT_SYMBOL(blk_stop_queue);
1543
1544/**
1545 * blk_sync_queue - cancel any pending callbacks on a queue
1546 * @q: the queue
1547 *
1548 * Description:
1549 * The block layer may perform asynchronous callback activity
1550 * on a queue, such as calling the unplug function after a timeout.
1551 * A block device may call blk_sync_queue to ensure that any
1552 * such activity is cancelled, thus allowing it to release resources
1553 * the the callbacks might use. The caller must already have made sure
1554 * that its ->make_request_fn will not re-add plugging prior to calling
1555 * this function.
1556 *
1557 */
1558void blk_sync_queue(struct request_queue *q)
1559{
1560 del_timer_sync(&q->unplug_timer);
1561 kblockd_flush();
1562}
1563EXPORT_SYMBOL(blk_sync_queue);
1564
1565/**
1566 * blk_run_queue - run a single device queue
1567 * @q: The queue to run
1568 */
1569void blk_run_queue(struct request_queue *q)
1570{
1571 unsigned long flags;
1572
1573 spin_lock_irqsave(q->queue_lock, flags);
1574 blk_remove_plug(q);
Ken Chena2997382005-04-16 15:25:43 -07001575 if (!elv_queue_empty(q))
1576 q->request_fn(q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577 spin_unlock_irqrestore(q->queue_lock, flags);
1578}
1579EXPORT_SYMBOL(blk_run_queue);
1580
1581/**
1582 * blk_cleanup_queue: - release a &request_queue_t when it is no longer needed
1583 * @q: the request queue to be released
1584 *
1585 * Description:
1586 * blk_cleanup_queue is the pair to blk_init_queue() or
1587 * blk_queue_make_request(). It should be called when a request queue is
1588 * being released; typically when a block device is being de-registered.
1589 * Currently, its primary task it to free all the &struct request
1590 * structures that were allocated to the queue and the queue itself.
1591 *
1592 * Caveat:
1593 * Hopefully the low level driver will have finished any
1594 * outstanding requests first...
1595 **/
1596void blk_cleanup_queue(request_queue_t * q)
1597{
1598 struct request_list *rl = &q->rq;
1599
1600 if (!atomic_dec_and_test(&q->refcnt))
1601 return;
1602
1603 if (q->elevator)
1604 elevator_exit(q->elevator);
1605
1606 blk_sync_queue(q);
1607
1608 if (rl->rq_pool)
1609 mempool_destroy(rl->rq_pool);
1610
1611 if (q->queue_tags)
1612 __blk_queue_free_tags(q);
1613
1614 blk_queue_ordered(q, QUEUE_ORDERED_NONE);
1615
1616 kmem_cache_free(requestq_cachep, q);
1617}
1618
1619EXPORT_SYMBOL(blk_cleanup_queue);
1620
1621static int blk_init_free_list(request_queue_t *q)
1622{
1623 struct request_list *rl = &q->rq;
1624
1625 rl->count[READ] = rl->count[WRITE] = 0;
1626 rl->starved[READ] = rl->starved[WRITE] = 0;
1627 init_waitqueue_head(&rl->wait[READ]);
1628 init_waitqueue_head(&rl->wait[WRITE]);
1629 init_waitqueue_head(&rl->drain);
1630
Christoph Lameter19460892005-06-23 00:08:19 -07001631 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1632 mempool_free_slab, request_cachep, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633
1634 if (!rl->rq_pool)
1635 return -ENOMEM;
1636
1637 return 0;
1638}
1639
1640static int __make_request(request_queue_t *, struct bio *);
1641
1642request_queue_t *blk_alloc_queue(int gfp_mask)
1643{
Christoph Lameter19460892005-06-23 00:08:19 -07001644 return blk_alloc_queue_node(gfp_mask, -1);
1645}
1646EXPORT_SYMBOL(blk_alloc_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647
Christoph Lameter19460892005-06-23 00:08:19 -07001648request_queue_t *blk_alloc_queue_node(int gfp_mask, int node_id)
1649{
1650 request_queue_t *q;
1651
1652 q = kmem_cache_alloc_node(requestq_cachep, gfp_mask, node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 if (!q)
1654 return NULL;
1655
1656 memset(q, 0, sizeof(*q));
1657 init_timer(&q->unplug_timer);
1658 atomic_set(&q->refcnt, 1);
1659
1660 q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
1661 q->backing_dev_info.unplug_io_data = q;
1662
1663 return q;
1664}
Christoph Lameter19460892005-06-23 00:08:19 -07001665EXPORT_SYMBOL(blk_alloc_queue_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666
1667/**
1668 * blk_init_queue - prepare a request queue for use with a block device
1669 * @rfn: The function to be called to process requests that have been
1670 * placed on the queue.
1671 * @lock: Request queue spin lock
1672 *
1673 * Description:
1674 * If a block device wishes to use the standard request handling procedures,
1675 * which sorts requests and coalesces adjacent requests, then it must
1676 * call blk_init_queue(). The function @rfn will be called when there
1677 * are requests on the queue that need to be processed. If the device
1678 * supports plugging, then @rfn may not be called immediately when requests
1679 * are available on the queue, but may be called at some time later instead.
1680 * Plugged queues are generally unplugged when a buffer belonging to one
1681 * of the requests on the queue is needed, or due to memory pressure.
1682 *
1683 * @rfn is not required, or even expected, to remove all requests off the
1684 * queue, but only as many as it can handle at a time. If it does leave
1685 * requests on the queue, it is responsible for arranging that the requests
1686 * get dealt with eventually.
1687 *
1688 * The queue spin lock must be held while manipulating the requests on the
1689 * request queue.
1690 *
1691 * Function returns a pointer to the initialized request queue, or NULL if
1692 * it didn't succeed.
1693 *
1694 * Note:
1695 * blk_init_queue() must be paired with a blk_cleanup_queue() call
1696 * when the block device is deactivated (such as at module unload).
1697 **/
Christoph Lameter19460892005-06-23 00:08:19 -07001698
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699request_queue_t *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
1700{
Christoph Lameter19460892005-06-23 00:08:19 -07001701 return blk_init_queue_node(rfn, lock, -1);
1702}
1703EXPORT_SYMBOL(blk_init_queue);
1704
1705request_queue_t *
1706blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
1707{
1708 request_queue_t *q = blk_alloc_queue_node(GFP_KERNEL, node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709
1710 if (!q)
1711 return NULL;
1712
Christoph Lameter19460892005-06-23 00:08:19 -07001713 q->node = node_id;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 if (blk_init_free_list(q))
1715 goto out_init;
1716
152587d2005-04-12 16:22:06 -05001717 /*
1718 * if caller didn't supply a lock, they get per-queue locking with
1719 * our embedded lock
1720 */
1721 if (!lock) {
1722 spin_lock_init(&q->__queue_lock);
1723 lock = &q->__queue_lock;
1724 }
1725
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 q->request_fn = rfn;
1727 q->back_merge_fn = ll_back_merge_fn;
1728 q->front_merge_fn = ll_front_merge_fn;
1729 q->merge_requests_fn = ll_merge_requests_fn;
1730 q->prep_rq_fn = NULL;
1731 q->unplug_fn = generic_unplug_device;
1732 q->queue_flags = (1 << QUEUE_FLAG_CLUSTER);
1733 q->queue_lock = lock;
1734
1735 blk_queue_segment_boundary(q, 0xffffffff);
1736
1737 blk_queue_make_request(q, __make_request);
1738 blk_queue_max_segment_size(q, MAX_SEGMENT_SIZE);
1739
1740 blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
1741 blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
1742
1743 /*
1744 * all done
1745 */
1746 if (!elevator_init(q, NULL)) {
1747 blk_queue_congestion_threshold(q);
1748 return q;
1749 }
1750
1751 blk_cleanup_queue(q);
1752out_init:
1753 kmem_cache_free(requestq_cachep, q);
1754 return NULL;
1755}
Christoph Lameter19460892005-06-23 00:08:19 -07001756EXPORT_SYMBOL(blk_init_queue_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757
1758int blk_get_queue(request_queue_t *q)
1759{
Nick Pigginfde6ad22005-06-23 00:08:53 -07001760 if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 atomic_inc(&q->refcnt);
1762 return 0;
1763 }
1764
1765 return 1;
1766}
1767
1768EXPORT_SYMBOL(blk_get_queue);
1769
1770static inline void blk_free_request(request_queue_t *q, struct request *rq)
1771{
1772 elv_put_request(q, rq);
1773 mempool_free(rq, q->rq.rq_pool);
1774}
1775
Jens Axboe22e2c502005-06-27 10:55:12 +02001776static inline struct request *
1777blk_alloc_request(request_queue_t *q, int rw, struct bio *bio, int gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778{
1779 struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
1780
1781 if (!rq)
1782 return NULL;
1783
1784 /*
1785 * first three bits are identical in rq->flags and bio->bi_rw,
1786 * see bio.h and blkdev.h
1787 */
1788 rq->flags = rw;
1789
Jens Axboe22e2c502005-06-27 10:55:12 +02001790 if (!elv_set_request(q, rq, bio, gfp_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791 return rq;
1792
1793 mempool_free(rq, q->rq.rq_pool);
1794 return NULL;
1795}
1796
1797/*
1798 * ioc_batching returns true if the ioc is a valid batching request and
1799 * should be given priority access to a request.
1800 */
1801static inline int ioc_batching(request_queue_t *q, struct io_context *ioc)
1802{
1803 if (!ioc)
1804 return 0;
1805
1806 /*
1807 * Make sure the process is able to allocate at least 1 request
1808 * even if the batch times out, otherwise we could theoretically
1809 * lose wakeups.
1810 */
1811 return ioc->nr_batch_requests == q->nr_batching ||
1812 (ioc->nr_batch_requests > 0
1813 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
1814}
1815
1816/*
1817 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
1818 * will cause the process to be a "batcher" on all queues in the system. This
1819 * is the behaviour we want though - once it gets a wakeup it should be given
1820 * a nice run.
1821 */
Adrian Bunk93d17d32005-06-25 14:59:10 -07001822static void ioc_set_batching(request_queue_t *q, struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823{
1824 if (!ioc || ioc_batching(q, ioc))
1825 return;
1826
1827 ioc->nr_batch_requests = q->nr_batching;
1828 ioc->last_waited = jiffies;
1829}
1830
1831static void __freed_request(request_queue_t *q, int rw)
1832{
1833 struct request_list *rl = &q->rq;
1834
1835 if (rl->count[rw] < queue_congestion_off_threshold(q))
1836 clear_queue_congested(q, rw);
1837
1838 if (rl->count[rw] + 1 <= q->nr_requests) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839 if (waitqueue_active(&rl->wait[rw]))
1840 wake_up(&rl->wait[rw]);
1841
1842 blk_clear_queue_full(q, rw);
1843 }
1844}
1845
1846/*
1847 * A request has just been released. Account for it, update the full and
1848 * congestion status, wake up any waiters. Called under q->queue_lock.
1849 */
1850static void freed_request(request_queue_t *q, int rw)
1851{
1852 struct request_list *rl = &q->rq;
1853
1854 rl->count[rw]--;
1855
1856 __freed_request(q, rw);
1857
1858 if (unlikely(rl->starved[rw ^ 1]))
1859 __freed_request(q, rw ^ 1);
1860
1861 if (!rl->count[READ] && !rl->count[WRITE]) {
1862 smp_mb();
1863 if (unlikely(waitqueue_active(&rl->drain)))
1864 wake_up(&rl->drain);
1865 }
1866}
1867
1868#define blkdev_free_rq(list) list_entry((list)->next, struct request, queuelist)
1869/*
1870 * Get a free request, queue_lock must not be held
1871 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001872static struct request *get_request(request_queue_t *q, int rw, struct bio *bio,
1873 int gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874{
1875 struct request *rq = NULL;
1876 struct request_list *rl = &q->rq;
1877 struct io_context *ioc = get_io_context(gfp_mask);
1878
1879 if (unlikely(test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags)))
1880 goto out;
1881
1882 spin_lock_irq(q->queue_lock);
1883 if (rl->count[rw]+1 >= q->nr_requests) {
1884 /*
1885 * The queue will fill after this allocation, so set it as
1886 * full, and mark this process as "batching". This process
1887 * will be allowed to complete a batch of requests, others
1888 * will be blocked.
1889 */
1890 if (!blk_queue_full(q, rw)) {
1891 ioc_set_batching(q, ioc);
1892 blk_set_queue_full(q, rw);
1893 }
1894 }
1895
Jens Axboe22e2c502005-06-27 10:55:12 +02001896 switch (elv_may_queue(q, rw, bio)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897 case ELV_MQUEUE_NO:
1898 goto rq_starved;
1899 case ELV_MQUEUE_MAY:
1900 break;
1901 case ELV_MQUEUE_MUST:
1902 goto get_rq;
1903 }
1904
1905 if (blk_queue_full(q, rw) && !ioc_batching(q, ioc)) {
1906 /*
1907 * The queue is full and the allocating process is not a
1908 * "batcher", and not exempted by the IO scheduler
1909 */
1910 spin_unlock_irq(q->queue_lock);
1911 goto out;
1912 }
1913
1914get_rq:
1915 rl->count[rw]++;
1916 rl->starved[rw] = 0;
1917 if (rl->count[rw] >= queue_congestion_on_threshold(q))
1918 set_queue_congested(q, rw);
1919 spin_unlock_irq(q->queue_lock);
1920
Jens Axboe22e2c502005-06-27 10:55:12 +02001921 rq = blk_alloc_request(q, rw, bio, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922 if (!rq) {
1923 /*
1924 * Allocation failed presumably due to memory. Undo anything
1925 * we might have messed up.
1926 *
1927 * Allocating task should really be put onto the front of the
1928 * wait queue, but this is pretty rare.
1929 */
1930 spin_lock_irq(q->queue_lock);
1931 freed_request(q, rw);
1932
1933 /*
1934 * in the very unlikely event that allocation failed and no
1935 * requests for this direction was pending, mark us starved
1936 * so that freeing of a request in the other direction will
1937 * notice us. another possible fix would be to split the
1938 * rq mempool into READ and WRITE
1939 */
1940rq_starved:
1941 if (unlikely(rl->count[rw] == 0))
1942 rl->starved[rw] = 1;
1943
1944 spin_unlock_irq(q->queue_lock);
1945 goto out;
1946 }
1947
1948 if (ioc_batching(q, ioc))
1949 ioc->nr_batch_requests--;
1950
1951 rq_init(q, rq);
1952 rq->rl = rl;
1953out:
1954 put_io_context(ioc);
1955 return rq;
1956}
1957
1958/*
1959 * No available requests for this queue, unplug the device and wait for some
1960 * requests to become available.
1961 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001962static struct request *get_request_wait(request_queue_t *q, int rw,
1963 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964{
1965 DEFINE_WAIT(wait);
1966 struct request *rq;
1967
Linus Torvalds1da177e2005-04-16 15:20:36 -07001968 do {
1969 struct request_list *rl = &q->rq;
1970
1971 prepare_to_wait_exclusive(&rl->wait[rw], &wait,
1972 TASK_UNINTERRUPTIBLE);
1973
Jens Axboe22e2c502005-06-27 10:55:12 +02001974 rq = get_request(q, rw, bio, GFP_NOIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975
1976 if (!rq) {
1977 struct io_context *ioc;
1978
Nick Pigginbdd646a2005-06-23 00:08:54 -07001979 generic_unplug_device(q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980 io_schedule();
1981
1982 /*
1983 * After sleeping, we become a "batching" process and
1984 * will be able to allocate at least one request, and
1985 * up to a big batch of them for a small period time.
1986 * See ioc_batching, ioc_set_batching
1987 */
1988 ioc = get_io_context(GFP_NOIO);
1989 ioc_set_batching(q, ioc);
1990 put_io_context(ioc);
1991 }
1992 finish_wait(&rl->wait[rw], &wait);
1993 } while (!rq);
1994
1995 return rq;
1996}
1997
1998struct request *blk_get_request(request_queue_t *q, int rw, int gfp_mask)
1999{
2000 struct request *rq;
2001
2002 BUG_ON(rw != READ && rw != WRITE);
2003
2004 if (gfp_mask & __GFP_WAIT)
Jens Axboe22e2c502005-06-27 10:55:12 +02002005 rq = get_request_wait(q, rw, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 else
Jens Axboe22e2c502005-06-27 10:55:12 +02002007 rq = get_request(q, rw, NULL, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008
2009 return rq;
2010}
2011
2012EXPORT_SYMBOL(blk_get_request);
2013
2014/**
2015 * blk_requeue_request - put a request back on queue
2016 * @q: request queue where request should be inserted
2017 * @rq: request to be inserted
2018 *
2019 * Description:
2020 * Drivers often keep queueing requests until the hardware cannot accept
2021 * more, when that condition happens we need to put the request back
2022 * on the queue. Must be called with queue lock held.
2023 */
2024void blk_requeue_request(request_queue_t *q, struct request *rq)
2025{
2026 if (blk_rq_tagged(rq))
2027 blk_queue_end_tag(q, rq);
2028
2029 elv_requeue_request(q, rq);
2030}
2031
2032EXPORT_SYMBOL(blk_requeue_request);
2033
2034/**
2035 * blk_insert_request - insert a special request in to a request queue
2036 * @q: request queue where request should be inserted
2037 * @rq: request to be inserted
2038 * @at_head: insert request at head or tail of queue
2039 * @data: private data
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040 *
2041 * Description:
2042 * Many block devices need to execute commands asynchronously, so they don't
2043 * block the whole kernel from preemption during request execution. This is
2044 * accomplished normally by inserting aritficial requests tagged as
2045 * REQ_SPECIAL in to the corresponding request queue, and letting them be
2046 * scheduled for actual execution by the request queue.
2047 *
2048 * We have the option of inserting the head or the tail of the queue.
2049 * Typically we use the tail for new ioctls and so forth. We use the head
2050 * of the queue for things like a QUEUE_FULL message from a device, or a
2051 * host that is unable to accept a particular command.
2052 */
2053void blk_insert_request(request_queue_t *q, struct request *rq,
Tejun Heo 867d1192005-04-24 02:06:05 -05002054 int at_head, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055{
Tejun Heo 867d1192005-04-24 02:06:05 -05002056 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057 unsigned long flags;
2058
2059 /*
2060 * tell I/O scheduler that this isn't a regular read/write (ie it
2061 * must not attempt merges on this) and that it acts as a soft
2062 * barrier
2063 */
2064 rq->flags |= REQ_SPECIAL | REQ_SOFTBARRIER;
2065
2066 rq->special = data;
2067
2068 spin_lock_irqsave(q->queue_lock, flags);
2069
2070 /*
2071 * If command is tagged, release the tag
2072 */
Tejun Heo 867d1192005-04-24 02:06:05 -05002073 if (blk_rq_tagged(rq))
2074 blk_queue_end_tag(q, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075
Tejun Heo 867d1192005-04-24 02:06:05 -05002076 drive_stat_acct(rq, rq->nr_sectors, 1);
2077 __elv_add_request(q, rq, where, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 if (blk_queue_plugged(q))
2080 __generic_unplug_device(q);
2081 else
2082 q->request_fn(q);
2083 spin_unlock_irqrestore(q->queue_lock, flags);
2084}
2085
2086EXPORT_SYMBOL(blk_insert_request);
2087
2088/**
2089 * blk_rq_map_user - map user data to a request, for REQ_BLOCK_PC usage
2090 * @q: request queue where request should be inserted
2091 * @rw: READ or WRITE data
2092 * @ubuf: the user buffer
2093 * @len: length of user data
2094 *
2095 * Description:
2096 * Data will be mapped directly for zero copy io, if possible. Otherwise
2097 * a kernel bounce buffer is used.
2098 *
2099 * A matching blk_rq_unmap_user() must be issued at the end of io, while
2100 * still in process context.
2101 *
2102 * Note: The mapped bio may need to be bounced through blk_queue_bounce()
2103 * before being submitted to the device, as pages mapped may be out of
2104 * reach. It's the callers responsibility to make sure this happens. The
2105 * original bio must be passed back in to blk_rq_unmap_user() for proper
2106 * unmapping.
2107 */
2108struct request *blk_rq_map_user(request_queue_t *q, int rw, void __user *ubuf,
2109 unsigned int len)
2110{
2111 unsigned long uaddr;
2112 struct request *rq;
2113 struct bio *bio;
2114
2115 if (len > (q->max_sectors << 9))
2116 return ERR_PTR(-EINVAL);
2117 if ((!len && ubuf) || (len && !ubuf))
2118 return ERR_PTR(-EINVAL);
2119
2120 rq = blk_get_request(q, rw, __GFP_WAIT);
2121 if (!rq)
2122 return ERR_PTR(-ENOMEM);
2123
2124 /*
2125 * if alignment requirement is satisfied, map in user pages for
2126 * direct dma. else, set up kernel bounce buffers
2127 */
2128 uaddr = (unsigned long) ubuf;
2129 if (!(uaddr & queue_dma_alignment(q)) && !(len & queue_dma_alignment(q)))
2130 bio = bio_map_user(q, NULL, uaddr, len, rw == READ);
2131 else
2132 bio = bio_copy_user(q, uaddr, len, rw == READ);
2133
2134 if (!IS_ERR(bio)) {
2135 rq->bio = rq->biotail = bio;
2136 blk_rq_bio_prep(q, rq, bio);
2137
2138 rq->buffer = rq->data = NULL;
2139 rq->data_len = len;
2140 return rq;
2141 }
2142
2143 /*
2144 * bio is the err-ptr
2145 */
2146 blk_put_request(rq);
2147 return (struct request *) bio;
2148}
2149
2150EXPORT_SYMBOL(blk_rq_map_user);
2151
2152/**
2153 * blk_rq_unmap_user - unmap a request with user data
2154 * @rq: request to be unmapped
2155 * @bio: bio for the request
2156 * @ulen: length of user buffer
2157 *
2158 * Description:
2159 * Unmap a request previously mapped by blk_rq_map_user().
2160 */
2161int blk_rq_unmap_user(struct request *rq, struct bio *bio, unsigned int ulen)
2162{
2163 int ret = 0;
2164
2165 if (bio) {
2166 if (bio_flagged(bio, BIO_USER_MAPPED))
2167 bio_unmap_user(bio);
2168 else
2169 ret = bio_uncopy_user(bio);
2170 }
2171
2172 blk_put_request(rq);
2173 return ret;
2174}
2175
2176EXPORT_SYMBOL(blk_rq_unmap_user);
2177
2178/**
2179 * blk_execute_rq - insert a request into queue for execution
2180 * @q: queue to insert the request in
2181 * @bd_disk: matching gendisk
2182 * @rq: request to insert
2183 *
2184 * Description:
2185 * Insert a fully prepared request at the back of the io scheduler queue
2186 * for execution.
2187 */
2188int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
2189 struct request *rq)
2190{
2191 DECLARE_COMPLETION(wait);
2192 char sense[SCSI_SENSE_BUFFERSIZE];
2193 int err = 0;
2194
2195 rq->rq_disk = bd_disk;
2196
2197 /*
2198 * we need an extra reference to the request, so we can look at
2199 * it after io completion
2200 */
2201 rq->ref_count++;
2202
2203 if (!rq->sense) {
2204 memset(sense, 0, sizeof(sense));
2205 rq->sense = sense;
2206 rq->sense_len = 0;
2207 }
2208
2209 rq->flags |= REQ_NOMERGE;
2210 rq->waiting = &wait;
2211 rq->end_io = blk_end_sync_rq;
2212 elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 1);
2213 generic_unplug_device(q);
2214 wait_for_completion(&wait);
2215 rq->waiting = NULL;
2216
2217 if (rq->errors)
2218 err = -EIO;
2219
2220 return err;
2221}
2222
2223EXPORT_SYMBOL(blk_execute_rq);
2224
2225/**
2226 * blkdev_issue_flush - queue a flush
2227 * @bdev: blockdev to issue flush for
2228 * @error_sector: error sector
2229 *
2230 * Description:
2231 * Issue a flush for the block device in question. Caller can supply
2232 * room for storing the error offset in case of a flush error, if they
2233 * wish to. Caller must run wait_for_completion() on its own.
2234 */
2235int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
2236{
2237 request_queue_t *q;
2238
2239 if (bdev->bd_disk == NULL)
2240 return -ENXIO;
2241
2242 q = bdev_get_queue(bdev);
2243 if (!q)
2244 return -ENXIO;
2245 if (!q->issue_flush_fn)
2246 return -EOPNOTSUPP;
2247
2248 return q->issue_flush_fn(q, bdev->bd_disk, error_sector);
2249}
2250
2251EXPORT_SYMBOL(blkdev_issue_flush);
2252
Adrian Bunk93d17d32005-06-25 14:59:10 -07002253static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254{
2255 int rw = rq_data_dir(rq);
2256
2257 if (!blk_fs_request(rq) || !rq->rq_disk)
2258 return;
2259
2260 if (rw == READ) {
2261 __disk_stat_add(rq->rq_disk, read_sectors, nr_sectors);
2262 if (!new_io)
2263 __disk_stat_inc(rq->rq_disk, read_merges);
2264 } else if (rw == WRITE) {
2265 __disk_stat_add(rq->rq_disk, write_sectors, nr_sectors);
2266 if (!new_io)
2267 __disk_stat_inc(rq->rq_disk, write_merges);
2268 }
2269 if (new_io) {
2270 disk_round_stats(rq->rq_disk);
2271 rq->rq_disk->in_flight++;
2272 }
2273}
2274
2275/*
2276 * add-request adds a request to the linked list.
2277 * queue lock is held and interrupts disabled, as we muck with the
2278 * request queue list.
2279 */
2280static inline void add_request(request_queue_t * q, struct request * req)
2281{
2282 drive_stat_acct(req, req->nr_sectors, 1);
2283
2284 if (q->activity_fn)
2285 q->activity_fn(q->activity_data, rq_data_dir(req));
2286
2287 /*
2288 * elevator indicated where it wants this request to be
2289 * inserted at elevator_merge time
2290 */
2291 __elv_add_request(q, req, ELEVATOR_INSERT_SORT, 0);
2292}
2293
2294/*
2295 * disk_round_stats() - Round off the performance stats on a struct
2296 * disk_stats.
2297 *
2298 * The average IO queue length and utilisation statistics are maintained
2299 * by observing the current state of the queue length and the amount of
2300 * time it has been in this state for.
2301 *
2302 * Normally, that accounting is done on IO completion, but that can result
2303 * in more than a second's worth of IO being accounted for within any one
2304 * second, leading to >100% utilisation. To deal with that, we call this
2305 * function to do a round-off before returning the results when reading
2306 * /proc/diskstats. This accounts immediately for all queue usage up to
2307 * the current jiffies and restarts the counters again.
2308 */
2309void disk_round_stats(struct gendisk *disk)
2310{
2311 unsigned long now = jiffies;
2312
2313 __disk_stat_add(disk, time_in_queue,
2314 disk->in_flight * (now - disk->stamp));
2315 disk->stamp = now;
2316
2317 if (disk->in_flight)
2318 __disk_stat_add(disk, io_ticks, (now - disk->stamp_idle));
2319 disk->stamp_idle = now;
2320}
2321
2322/*
2323 * queue lock must be held
2324 */
2325static void __blk_put_request(request_queue_t *q, struct request *req)
2326{
2327 struct request_list *rl = req->rl;
2328
2329 if (unlikely(!q))
2330 return;
2331 if (unlikely(--req->ref_count))
2332 return;
2333
2334 req->rq_status = RQ_INACTIVE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335 req->rl = NULL;
2336
2337 /*
2338 * Request may not have originated from ll_rw_blk. if not,
2339 * it didn't come out of our reserved rq pools
2340 */
2341 if (rl) {
2342 int rw = rq_data_dir(req);
2343
2344 elv_completed_request(q, req);
2345
2346 BUG_ON(!list_empty(&req->queuelist));
2347
2348 blk_free_request(q, req);
2349 freed_request(q, rw);
2350 }
2351}
2352
2353void blk_put_request(struct request *req)
2354{
2355 /*
2356 * if req->rl isn't set, this request didnt originate from the
2357 * block layer, so it's safe to just disregard it
2358 */
2359 if (req->rl) {
2360 unsigned long flags;
2361 request_queue_t *q = req->q;
2362
2363 spin_lock_irqsave(q->queue_lock, flags);
2364 __blk_put_request(q, req);
2365 spin_unlock_irqrestore(q->queue_lock, flags);
2366 }
2367}
2368
2369EXPORT_SYMBOL(blk_put_request);
2370
2371/**
2372 * blk_end_sync_rq - executes a completion event on a request
2373 * @rq: request to complete
2374 */
2375void blk_end_sync_rq(struct request *rq)
2376{
2377 struct completion *waiting = rq->waiting;
2378
2379 rq->waiting = NULL;
2380 __blk_put_request(rq->q, rq);
2381
2382 /*
2383 * complete last, if this is a stack request the process (and thus
2384 * the rq pointer) could be invalid right after this complete()
2385 */
2386 complete(waiting);
2387}
2388EXPORT_SYMBOL(blk_end_sync_rq);
2389
2390/**
2391 * blk_congestion_wait - wait for a queue to become uncongested
2392 * @rw: READ or WRITE
2393 * @timeout: timeout in jiffies
2394 *
2395 * Waits for up to @timeout jiffies for a queue (any queue) to exit congestion.
2396 * If no queues are congested then just wait for the next request to be
2397 * returned.
2398 */
2399long blk_congestion_wait(int rw, long timeout)
2400{
2401 long ret;
2402 DEFINE_WAIT(wait);
2403 wait_queue_head_t *wqh = &congestion_wqh[rw];
2404
2405 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
2406 ret = io_schedule_timeout(timeout);
2407 finish_wait(wqh, &wait);
2408 return ret;
2409}
2410
2411EXPORT_SYMBOL(blk_congestion_wait);
2412
2413/*
2414 * Has to be called with the request spinlock acquired
2415 */
2416static int attempt_merge(request_queue_t *q, struct request *req,
2417 struct request *next)
2418{
2419 if (!rq_mergeable(req) || !rq_mergeable(next))
2420 return 0;
2421
2422 /*
2423 * not contigious
2424 */
2425 if (req->sector + req->nr_sectors != next->sector)
2426 return 0;
2427
2428 if (rq_data_dir(req) != rq_data_dir(next)
2429 || req->rq_disk != next->rq_disk
2430 || next->waiting || next->special)
2431 return 0;
2432
2433 /*
2434 * If we are allowed to merge, then append bio list
2435 * from next to rq and release next. merge_requests_fn
2436 * will have updated segment counts, update sector
2437 * counts here.
2438 */
2439 if (!q->merge_requests_fn(q, req, next))
2440 return 0;
2441
2442 /*
2443 * At this point we have either done a back merge
2444 * or front merge. We need the smaller start_time of
2445 * the merged requests to be the current request
2446 * for accounting purposes.
2447 */
2448 if (time_after(req->start_time, next->start_time))
2449 req->start_time = next->start_time;
2450
2451 req->biotail->bi_next = next->bio;
2452 req->biotail = next->biotail;
2453
2454 req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
2455
2456 elv_merge_requests(q, req, next);
2457
2458 if (req->rq_disk) {
2459 disk_round_stats(req->rq_disk);
2460 req->rq_disk->in_flight--;
2461 }
2462
Jens Axboe22e2c502005-06-27 10:55:12 +02002463 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
2464
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465 __blk_put_request(q, next);
2466 return 1;
2467}
2468
2469static inline int attempt_back_merge(request_queue_t *q, struct request *rq)
2470{
2471 struct request *next = elv_latter_request(q, rq);
2472
2473 if (next)
2474 return attempt_merge(q, rq, next);
2475
2476 return 0;
2477}
2478
2479static inline int attempt_front_merge(request_queue_t *q, struct request *rq)
2480{
2481 struct request *prev = elv_former_request(q, rq);
2482
2483 if (prev)
2484 return attempt_merge(q, prev, rq);
2485
2486 return 0;
2487}
2488
2489/**
2490 * blk_attempt_remerge - attempt to remerge active head with next request
2491 * @q: The &request_queue_t belonging to the device
2492 * @rq: The head request (usually)
2493 *
2494 * Description:
2495 * For head-active devices, the queue can easily be unplugged so quickly
2496 * that proper merging is not done on the front request. This may hurt
2497 * performance greatly for some devices. The block layer cannot safely
2498 * do merging on that first request for these queues, but the driver can
2499 * call this function and make it happen any way. Only the driver knows
2500 * when it is safe to do so.
2501 **/
2502void blk_attempt_remerge(request_queue_t *q, struct request *rq)
2503{
2504 unsigned long flags;
2505
2506 spin_lock_irqsave(q->queue_lock, flags);
2507 attempt_back_merge(q, rq);
2508 spin_unlock_irqrestore(q->queue_lock, flags);
2509}
2510
2511EXPORT_SYMBOL(blk_attempt_remerge);
2512
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513static int __make_request(request_queue_t *q, struct bio *bio)
2514{
2515 struct request *req, *freereq = NULL;
Jens Axboe4a534f92005-04-16 15:25:40 -07002516 int el_ret, rw, nr_sectors, cur_nr_sectors, barrier, err, sync;
Jens Axboe22e2c502005-06-27 10:55:12 +02002517 unsigned short prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518 sector_t sector;
2519
2520 sector = bio->bi_sector;
2521 nr_sectors = bio_sectors(bio);
2522 cur_nr_sectors = bio_cur_sectors(bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002523 prio = bio_prio(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524
2525 rw = bio_data_dir(bio);
Jens Axboe4a534f92005-04-16 15:25:40 -07002526 sync = bio_sync(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527
2528 /*
2529 * low level driver can indicate that it wants pages above a
2530 * certain limit bounced to low memory (ie for highmem, or even
2531 * ISA dma in theory)
2532 */
2533 blk_queue_bounce(q, &bio);
2534
2535 spin_lock_prefetch(q->queue_lock);
2536
2537 barrier = bio_barrier(bio);
Nick Pigginfde6ad22005-06-23 00:08:53 -07002538 if (unlikely(barrier) && (q->ordered == QUEUE_ORDERED_NONE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 err = -EOPNOTSUPP;
2540 goto end_io;
2541 }
2542
2543again:
2544 spin_lock_irq(q->queue_lock);
2545
2546 if (elv_queue_empty(q)) {
2547 blk_plug_device(q);
2548 goto get_rq;
2549 }
2550 if (barrier)
2551 goto get_rq;
2552
2553 el_ret = elv_merge(q, &req, bio);
2554 switch (el_ret) {
2555 case ELEVATOR_BACK_MERGE:
2556 BUG_ON(!rq_mergeable(req));
2557
2558 if (!q->back_merge_fn(q, req, bio))
2559 break;
2560
2561 req->biotail->bi_next = bio;
2562 req->biotail = bio;
2563 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02002564 req->ioprio = ioprio_best(req->ioprio, prio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 drive_stat_acct(req, nr_sectors, 0);
2566 if (!attempt_back_merge(q, req))
2567 elv_merged_request(q, req);
2568 goto out;
2569
2570 case ELEVATOR_FRONT_MERGE:
2571 BUG_ON(!rq_mergeable(req));
2572
2573 if (!q->front_merge_fn(q, req, bio))
2574 break;
2575
2576 bio->bi_next = req->bio;
2577 req->bio = bio;
2578
2579 /*
2580 * may not be valid. if the low level driver said
2581 * it didn't need a bounce buffer then it better
2582 * not touch req->buffer either...
2583 */
2584 req->buffer = bio_data(bio);
2585 req->current_nr_sectors = cur_nr_sectors;
2586 req->hard_cur_sectors = cur_nr_sectors;
2587 req->sector = req->hard_sector = sector;
2588 req->nr_sectors = req->hard_nr_sectors += nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02002589 req->ioprio = ioprio_best(req->ioprio, prio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590 drive_stat_acct(req, nr_sectors, 0);
2591 if (!attempt_front_merge(q, req))
2592 elv_merged_request(q, req);
2593 goto out;
2594
2595 /*
2596 * elevator says don't/can't merge. get new request
2597 */
2598 case ELEVATOR_NO_MERGE:
2599 break;
2600
2601 default:
2602 printk("elevator returned crap (%d)\n", el_ret);
2603 BUG();
2604 }
2605
2606 /*
2607 * Grab a free request from the freelist - if that is empty, check
2608 * if we are doing read ahead and abort instead of blocking for
2609 * a free slot.
2610 */
2611get_rq:
2612 if (freereq) {
2613 req = freereq;
2614 freereq = NULL;
2615 } else {
2616 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02002617 if ((freereq = get_request(q, rw, bio, GFP_ATOMIC)) == NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002618 /*
2619 * READA bit set
2620 */
2621 err = -EWOULDBLOCK;
2622 if (bio_rw_ahead(bio))
2623 goto end_io;
2624
Jens Axboe22e2c502005-06-27 10:55:12 +02002625 freereq = get_request_wait(q, rw, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626 }
2627 goto again;
2628 }
2629
2630 req->flags |= REQ_CMD;
2631
2632 /*
2633 * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST)
2634 */
2635 if (bio_rw_ahead(bio) || bio_failfast(bio))
2636 req->flags |= REQ_FAILFAST;
2637
2638 /*
2639 * REQ_BARRIER implies no merging, but lets make it explicit
2640 */
Nick Pigginfde6ad22005-06-23 00:08:53 -07002641 if (unlikely(barrier))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642 req->flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
2643
2644 req->errors = 0;
2645 req->hard_sector = req->sector = sector;
2646 req->hard_nr_sectors = req->nr_sectors = nr_sectors;
2647 req->current_nr_sectors = req->hard_cur_sectors = cur_nr_sectors;
2648 req->nr_phys_segments = bio_phys_segments(q, bio);
2649 req->nr_hw_segments = bio_hw_segments(q, bio);
2650 req->buffer = bio_data(bio); /* see ->buffer comment above */
2651 req->waiting = NULL;
2652 req->bio = req->biotail = bio;
Jens Axboe22e2c502005-06-27 10:55:12 +02002653 req->ioprio = prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654 req->rq_disk = bio->bi_bdev->bd_disk;
2655 req->start_time = jiffies;
2656
2657 add_request(q, req);
2658out:
2659 if (freereq)
2660 __blk_put_request(q, freereq);
Jens Axboe4a534f92005-04-16 15:25:40 -07002661 if (sync)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662 __generic_unplug_device(q);
2663
2664 spin_unlock_irq(q->queue_lock);
2665 return 0;
2666
2667end_io:
2668 bio_endio(bio, nr_sectors << 9, err);
2669 return 0;
2670}
2671
2672/*
2673 * If bio->bi_dev is a partition, remap the location
2674 */
2675static inline void blk_partition_remap(struct bio *bio)
2676{
2677 struct block_device *bdev = bio->bi_bdev;
2678
2679 if (bdev != bdev->bd_contains) {
2680 struct hd_struct *p = bdev->bd_part;
2681
Jens Axboe22e2c502005-06-27 10:55:12 +02002682 switch (bio_data_dir(bio)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683 case READ:
2684 p->read_sectors += bio_sectors(bio);
2685 p->reads++;
2686 break;
2687 case WRITE:
2688 p->write_sectors += bio_sectors(bio);
2689 p->writes++;
2690 break;
2691 }
2692 bio->bi_sector += p->start_sect;
2693 bio->bi_bdev = bdev->bd_contains;
2694 }
2695}
2696
2697void blk_finish_queue_drain(request_queue_t *q)
2698{
2699 struct request_list *rl = &q->rq;
2700 struct request *rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002701 int requeued = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002702
2703 spin_lock_irq(q->queue_lock);
2704 clear_bit(QUEUE_FLAG_DRAIN, &q->queue_flags);
2705
2706 while (!list_empty(&q->drain_list)) {
2707 rq = list_entry_rq(q->drain_list.next);
2708
2709 list_del_init(&rq->queuelist);
Jens Axboe22e2c502005-06-27 10:55:12 +02002710 elv_requeue_request(q, rq);
2711 requeued++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712 }
2713
Jens Axboe22e2c502005-06-27 10:55:12 +02002714 if (requeued)
2715 q->request_fn(q);
2716
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717 spin_unlock_irq(q->queue_lock);
2718
2719 wake_up(&rl->wait[0]);
2720 wake_up(&rl->wait[1]);
2721 wake_up(&rl->drain);
2722}
2723
2724static int wait_drain(request_queue_t *q, struct request_list *rl, int dispatch)
2725{
2726 int wait = rl->count[READ] + rl->count[WRITE];
2727
2728 if (dispatch)
2729 wait += !list_empty(&q->queue_head);
2730
2731 return wait;
2732}
2733
2734/*
2735 * We rely on the fact that only requests allocated through blk_alloc_request()
2736 * have io scheduler private data structures associated with them. Any other
2737 * type of request (allocated on stack or through kmalloc()) should not go
2738 * to the io scheduler core, but be attached to the queue head instead.
2739 */
2740void blk_wait_queue_drained(request_queue_t *q, int wait_dispatch)
2741{
2742 struct request_list *rl = &q->rq;
2743 DEFINE_WAIT(wait);
2744
2745 spin_lock_irq(q->queue_lock);
2746 set_bit(QUEUE_FLAG_DRAIN, &q->queue_flags);
2747
2748 while (wait_drain(q, rl, wait_dispatch)) {
2749 prepare_to_wait(&rl->drain, &wait, TASK_UNINTERRUPTIBLE);
2750
2751 if (wait_drain(q, rl, wait_dispatch)) {
2752 __generic_unplug_device(q);
2753 spin_unlock_irq(q->queue_lock);
2754 io_schedule();
2755 spin_lock_irq(q->queue_lock);
2756 }
2757
2758 finish_wait(&rl->drain, &wait);
2759 }
2760
2761 spin_unlock_irq(q->queue_lock);
2762}
2763
2764/*
2765 * block waiting for the io scheduler being started again.
2766 */
2767static inline void block_wait_queue_running(request_queue_t *q)
2768{
2769 DEFINE_WAIT(wait);
2770
Nick Pigginfde6ad22005-06-23 00:08:53 -07002771 while (unlikely(test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772 struct request_list *rl = &q->rq;
2773
2774 prepare_to_wait_exclusive(&rl->drain, &wait,
2775 TASK_UNINTERRUPTIBLE);
2776
2777 /*
2778 * re-check the condition. avoids using prepare_to_wait()
2779 * in the fast path (queue is running)
2780 */
2781 if (test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags))
2782 io_schedule();
2783
2784 finish_wait(&rl->drain, &wait);
2785 }
2786}
2787
2788static void handle_bad_sector(struct bio *bio)
2789{
2790 char b[BDEVNAME_SIZE];
2791
2792 printk(KERN_INFO "attempt to access beyond end of device\n");
2793 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
2794 bdevname(bio->bi_bdev, b),
2795 bio->bi_rw,
2796 (unsigned long long)bio->bi_sector + bio_sectors(bio),
2797 (long long)(bio->bi_bdev->bd_inode->i_size >> 9));
2798
2799 set_bit(BIO_EOF, &bio->bi_flags);
2800}
2801
2802/**
2803 * generic_make_request: hand a buffer to its device driver for I/O
2804 * @bio: The bio describing the location in memory and on the device.
2805 *
2806 * generic_make_request() is used to make I/O requests of block
2807 * devices. It is passed a &struct bio, which describes the I/O that needs
2808 * to be done.
2809 *
2810 * generic_make_request() does not return any status. The
2811 * success/failure status of the request, along with notification of
2812 * completion, is delivered asynchronously through the bio->bi_end_io
2813 * function described (one day) else where.
2814 *
2815 * The caller of generic_make_request must make sure that bi_io_vec
2816 * are set to describe the memory buffer, and that bi_dev and bi_sector are
2817 * set to describe the device address, and the
2818 * bi_end_io and optionally bi_private are set to describe how
2819 * completion notification should be signaled.
2820 *
2821 * generic_make_request and the drivers it calls may use bi_next if this
2822 * bio happens to be merged with someone else, and may change bi_dev and
2823 * bi_sector for remaps as it sees fit. So the values of these fields
2824 * should NOT be depended on after the call to generic_make_request.
2825 */
2826void generic_make_request(struct bio *bio)
2827{
2828 request_queue_t *q;
2829 sector_t maxsector;
2830 int ret, nr_sectors = bio_sectors(bio);
2831
2832 might_sleep();
2833 /* Test device or partition size, when known. */
2834 maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
2835 if (maxsector) {
2836 sector_t sector = bio->bi_sector;
2837
2838 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
2839 /*
2840 * This may well happen - the kernel calls bread()
2841 * without checking the size of the device, e.g., when
2842 * mounting a device.
2843 */
2844 handle_bad_sector(bio);
2845 goto end_io;
2846 }
2847 }
2848
2849 /*
2850 * Resolve the mapping until finished. (drivers are
2851 * still free to implement/resolve their own stacking
2852 * by explicitly returning 0)
2853 *
2854 * NOTE: we don't repeat the blk_size check for each new device.
2855 * Stacking drivers are expected to know what they are doing.
2856 */
2857 do {
2858 char b[BDEVNAME_SIZE];
2859
2860 q = bdev_get_queue(bio->bi_bdev);
2861 if (!q) {
2862 printk(KERN_ERR
2863 "generic_make_request: Trying to access "
2864 "nonexistent block-device %s (%Lu)\n",
2865 bdevname(bio->bi_bdev, b),
2866 (long long) bio->bi_sector);
2867end_io:
2868 bio_endio(bio, bio->bi_size, -EIO);
2869 break;
2870 }
2871
2872 if (unlikely(bio_sectors(bio) > q->max_hw_sectors)) {
2873 printk("bio too big device %s (%u > %u)\n",
2874 bdevname(bio->bi_bdev, b),
2875 bio_sectors(bio),
2876 q->max_hw_sectors);
2877 goto end_io;
2878 }
2879
Nick Pigginfde6ad22005-06-23 00:08:53 -07002880 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881 goto end_io;
2882
2883 block_wait_queue_running(q);
2884
2885 /*
2886 * If this device has partitions, remap block n
2887 * of partition p to block n+start(p) of the disk.
2888 */
2889 blk_partition_remap(bio);
2890
2891 ret = q->make_request_fn(q, bio);
2892 } while (ret);
2893}
2894
2895EXPORT_SYMBOL(generic_make_request);
2896
2897/**
2898 * submit_bio: submit a bio to the block device layer for I/O
2899 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
2900 * @bio: The &struct bio which describes the I/O
2901 *
2902 * submit_bio() is very similar in purpose to generic_make_request(), and
2903 * uses that function to do most of the work. Both are fairly rough
2904 * interfaces, @bio must be presetup and ready for I/O.
2905 *
2906 */
2907void submit_bio(int rw, struct bio *bio)
2908{
2909 int count = bio_sectors(bio);
2910
2911 BIO_BUG_ON(!bio->bi_size);
2912 BIO_BUG_ON(!bio->bi_io_vec);
Jens Axboe22e2c502005-06-27 10:55:12 +02002913 bio->bi_rw |= rw;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002914 if (rw & WRITE)
2915 mod_page_state(pgpgout, count);
2916 else
2917 mod_page_state(pgpgin, count);
2918
2919 if (unlikely(block_dump)) {
2920 char b[BDEVNAME_SIZE];
2921 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
2922 current->comm, current->pid,
2923 (rw & WRITE) ? "WRITE" : "READ",
2924 (unsigned long long)bio->bi_sector,
2925 bdevname(bio->bi_bdev,b));
2926 }
2927
2928 generic_make_request(bio);
2929}
2930
2931EXPORT_SYMBOL(submit_bio);
2932
Adrian Bunk93d17d32005-06-25 14:59:10 -07002933static void blk_recalc_rq_segments(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934{
2935 struct bio *bio, *prevbio = NULL;
2936 int nr_phys_segs, nr_hw_segs;
2937 unsigned int phys_size, hw_size;
2938 request_queue_t *q = rq->q;
2939
2940 if (!rq->bio)
2941 return;
2942
2943 phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
2944 rq_for_each_bio(bio, rq) {
2945 /* Force bio hw/phys segs to be recalculated. */
2946 bio->bi_flags &= ~(1 << BIO_SEG_VALID);
2947
2948 nr_phys_segs += bio_phys_segments(q, bio);
2949 nr_hw_segs += bio_hw_segments(q, bio);
2950 if (prevbio) {
2951 int pseg = phys_size + prevbio->bi_size + bio->bi_size;
2952 int hseg = hw_size + prevbio->bi_size + bio->bi_size;
2953
2954 if (blk_phys_contig_segment(q, prevbio, bio) &&
2955 pseg <= q->max_segment_size) {
2956 nr_phys_segs--;
2957 phys_size += prevbio->bi_size + bio->bi_size;
2958 } else
2959 phys_size = 0;
2960
2961 if (blk_hw_contig_segment(q, prevbio, bio) &&
2962 hseg <= q->max_segment_size) {
2963 nr_hw_segs--;
2964 hw_size += prevbio->bi_size + bio->bi_size;
2965 } else
2966 hw_size = 0;
2967 }
2968 prevbio = bio;
2969 }
2970
2971 rq->nr_phys_segments = nr_phys_segs;
2972 rq->nr_hw_segments = nr_hw_segs;
2973}
2974
Adrian Bunk93d17d32005-06-25 14:59:10 -07002975static void blk_recalc_rq_sectors(struct request *rq, int nsect)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976{
2977 if (blk_fs_request(rq)) {
2978 rq->hard_sector += nsect;
2979 rq->hard_nr_sectors -= nsect;
2980
2981 /*
2982 * Move the I/O submission pointers ahead if required.
2983 */
2984 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
2985 (rq->sector <= rq->hard_sector)) {
2986 rq->sector = rq->hard_sector;
2987 rq->nr_sectors = rq->hard_nr_sectors;
2988 rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
2989 rq->current_nr_sectors = rq->hard_cur_sectors;
2990 rq->buffer = bio_data(rq->bio);
2991 }
2992
2993 /*
2994 * if total number of sectors is less than the first segment
2995 * size, something has gone terribly wrong
2996 */
2997 if (rq->nr_sectors < rq->current_nr_sectors) {
2998 printk("blk: request botched\n");
2999 rq->nr_sectors = rq->current_nr_sectors;
3000 }
3001 }
3002}
3003
3004static int __end_that_request_first(struct request *req, int uptodate,
3005 int nr_bytes)
3006{
3007 int total_bytes, bio_nbytes, error, next_idx = 0;
3008 struct bio *bio;
3009
3010 /*
3011 * extend uptodate bool to allow < 0 value to be direct io error
3012 */
3013 error = 0;
3014 if (end_io_error(uptodate))
3015 error = !uptodate ? -EIO : uptodate;
3016
3017 /*
3018 * for a REQ_BLOCK_PC request, we want to carry any eventual
3019 * sense key with us all the way through
3020 */
3021 if (!blk_pc_request(req))
3022 req->errors = 0;
3023
3024 if (!uptodate) {
3025 if (blk_fs_request(req) && !(req->flags & REQ_QUIET))
3026 printk("end_request: I/O error, dev %s, sector %llu\n",
3027 req->rq_disk ? req->rq_disk->disk_name : "?",
3028 (unsigned long long)req->sector);
3029 }
3030
3031 total_bytes = bio_nbytes = 0;
3032 while ((bio = req->bio) != NULL) {
3033 int nbytes;
3034
3035 if (nr_bytes >= bio->bi_size) {
3036 req->bio = bio->bi_next;
3037 nbytes = bio->bi_size;
3038 bio_endio(bio, nbytes, error);
3039 next_idx = 0;
3040 bio_nbytes = 0;
3041 } else {
3042 int idx = bio->bi_idx + next_idx;
3043
3044 if (unlikely(bio->bi_idx >= bio->bi_vcnt)) {
3045 blk_dump_rq_flags(req, "__end_that");
3046 printk("%s: bio idx %d >= vcnt %d\n",
3047 __FUNCTION__,
3048 bio->bi_idx, bio->bi_vcnt);
3049 break;
3050 }
3051
3052 nbytes = bio_iovec_idx(bio, idx)->bv_len;
3053 BIO_BUG_ON(nbytes > bio->bi_size);
3054
3055 /*
3056 * not a complete bvec done
3057 */
3058 if (unlikely(nbytes > nr_bytes)) {
3059 bio_nbytes += nr_bytes;
3060 total_bytes += nr_bytes;
3061 break;
3062 }
3063
3064 /*
3065 * advance to the next vector
3066 */
3067 next_idx++;
3068 bio_nbytes += nbytes;
3069 }
3070
3071 total_bytes += nbytes;
3072 nr_bytes -= nbytes;
3073
3074 if ((bio = req->bio)) {
3075 /*
3076 * end more in this run, or just return 'not-done'
3077 */
3078 if (unlikely(nr_bytes <= 0))
3079 break;
3080 }
3081 }
3082
3083 /*
3084 * completely done
3085 */
3086 if (!req->bio)
3087 return 0;
3088
3089 /*
3090 * if the request wasn't completed, update state
3091 */
3092 if (bio_nbytes) {
3093 bio_endio(bio, bio_nbytes, error);
3094 bio->bi_idx += next_idx;
3095 bio_iovec(bio)->bv_offset += nr_bytes;
3096 bio_iovec(bio)->bv_len -= nr_bytes;
3097 }
3098
3099 blk_recalc_rq_sectors(req, total_bytes >> 9);
3100 blk_recalc_rq_segments(req);
3101 return 1;
3102}
3103
3104/**
3105 * end_that_request_first - end I/O on a request
3106 * @req: the request being processed
3107 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3108 * @nr_sectors: number of sectors to end I/O on
3109 *
3110 * Description:
3111 * Ends I/O on a number of sectors attached to @req, and sets it up
3112 * for the next range of segments (if any) in the cluster.
3113 *
3114 * Return:
3115 * 0 - we are done with this request, call end_that_request_last()
3116 * 1 - still buffers pending for this request
3117 **/
3118int end_that_request_first(struct request *req, int uptodate, int nr_sectors)
3119{
3120 return __end_that_request_first(req, uptodate, nr_sectors << 9);
3121}
3122
3123EXPORT_SYMBOL(end_that_request_first);
3124
3125/**
3126 * end_that_request_chunk - end I/O on a request
3127 * @req: the request being processed
3128 * @uptodate: 1 for success, 0 for I/O error, < 0 for specific error
3129 * @nr_bytes: number of bytes to complete
3130 *
3131 * Description:
3132 * Ends I/O on a number of bytes attached to @req, and sets it up
3133 * for the next range of segments (if any). Like end_that_request_first(),
3134 * but deals with bytes instead of sectors.
3135 *
3136 * Return:
3137 * 0 - we are done with this request, call end_that_request_last()
3138 * 1 - still buffers pending for this request
3139 **/
3140int end_that_request_chunk(struct request *req, int uptodate, int nr_bytes)
3141{
3142 return __end_that_request_first(req, uptodate, nr_bytes);
3143}
3144
3145EXPORT_SYMBOL(end_that_request_chunk);
3146
3147/*
3148 * queue lock must be held
3149 */
3150void end_that_request_last(struct request *req)
3151{
3152 struct gendisk *disk = req->rq_disk;
3153
3154 if (unlikely(laptop_mode) && blk_fs_request(req))
3155 laptop_io_completion();
3156
3157 if (disk && blk_fs_request(req)) {
3158 unsigned long duration = jiffies - req->start_time;
3159 switch (rq_data_dir(req)) {
3160 case WRITE:
3161 __disk_stat_inc(disk, writes);
3162 __disk_stat_add(disk, write_ticks, duration);
3163 break;
3164 case READ:
3165 __disk_stat_inc(disk, reads);
3166 __disk_stat_add(disk, read_ticks, duration);
3167 break;
3168 }
3169 disk_round_stats(disk);
3170 disk->in_flight--;
3171 }
3172 if (req->end_io)
3173 req->end_io(req);
3174 else
3175 __blk_put_request(req->q, req);
3176}
3177
3178EXPORT_SYMBOL(end_that_request_last);
3179
3180void end_request(struct request *req, int uptodate)
3181{
3182 if (!end_that_request_first(req, uptodate, req->hard_cur_sectors)) {
3183 add_disk_randomness(req->rq_disk);
3184 blkdev_dequeue_request(req);
3185 end_that_request_last(req);
3186 }
3187}
3188
3189EXPORT_SYMBOL(end_request);
3190
3191void blk_rq_bio_prep(request_queue_t *q, struct request *rq, struct bio *bio)
3192{
3193 /* first three bits are identical in rq->flags and bio->bi_rw */
3194 rq->flags |= (bio->bi_rw & 7);
3195
3196 rq->nr_phys_segments = bio_phys_segments(q, bio);
3197 rq->nr_hw_segments = bio_hw_segments(q, bio);
3198 rq->current_nr_sectors = bio_cur_sectors(bio);
3199 rq->hard_cur_sectors = rq->current_nr_sectors;
3200 rq->hard_nr_sectors = rq->nr_sectors = bio_sectors(bio);
3201 rq->buffer = bio_data(bio);
3202
3203 rq->bio = rq->biotail = bio;
3204}
3205
3206EXPORT_SYMBOL(blk_rq_bio_prep);
3207
3208int kblockd_schedule_work(struct work_struct *work)
3209{
3210 return queue_work(kblockd_workqueue, work);
3211}
3212
3213EXPORT_SYMBOL(kblockd_schedule_work);
3214
3215void kblockd_flush(void)
3216{
3217 flush_workqueue(kblockd_workqueue);
3218}
3219EXPORT_SYMBOL(kblockd_flush);
3220
3221int __init blk_dev_init(void)
3222{
3223 kblockd_workqueue = create_workqueue("kblockd");
3224 if (!kblockd_workqueue)
3225 panic("Failed to create kblockd\n");
3226
3227 request_cachep = kmem_cache_create("blkdev_requests",
3228 sizeof(struct request), 0, SLAB_PANIC, NULL, NULL);
3229
3230 requestq_cachep = kmem_cache_create("blkdev_queue",
3231 sizeof(request_queue_t), 0, SLAB_PANIC, NULL, NULL);
3232
3233 iocontext_cachep = kmem_cache_create("blkdev_ioc",
3234 sizeof(struct io_context), 0, SLAB_PANIC, NULL, NULL);
3235
3236 blk_max_low_pfn = max_low_pfn;
3237 blk_max_pfn = max_pfn;
3238
3239 return 0;
3240}
3241
3242/*
3243 * IO Context helper functions
3244 */
3245void put_io_context(struct io_context *ioc)
3246{
3247 if (ioc == NULL)
3248 return;
3249
3250 BUG_ON(atomic_read(&ioc->refcount) == 0);
3251
3252 if (atomic_dec_and_test(&ioc->refcount)) {
3253 if (ioc->aic && ioc->aic->dtor)
3254 ioc->aic->dtor(ioc->aic);
3255 if (ioc->cic && ioc->cic->dtor)
3256 ioc->cic->dtor(ioc->cic);
3257
3258 kmem_cache_free(iocontext_cachep, ioc);
3259 }
3260}
3261EXPORT_SYMBOL(put_io_context);
3262
3263/* Called by the exitting task */
3264void exit_io_context(void)
3265{
3266 unsigned long flags;
3267 struct io_context *ioc;
3268
3269 local_irq_save(flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02003270 task_lock(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003271 ioc = current->io_context;
3272 current->io_context = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02003273 ioc->task = NULL;
3274 task_unlock(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003275 local_irq_restore(flags);
3276
3277 if (ioc->aic && ioc->aic->exit)
3278 ioc->aic->exit(ioc->aic);
3279 if (ioc->cic && ioc->cic->exit)
3280 ioc->cic->exit(ioc->cic);
3281
3282 put_io_context(ioc);
3283}
3284
3285/*
3286 * If the current task has no IO context then create one and initialise it.
3287 * If it does have a context, take a ref on it.
3288 *
3289 * This is always called in the context of the task which submitted the I/O.
3290 * But weird things happen, so we disable local interrupts to ensure exclusive
3291 * access to *current.
3292 */
3293struct io_context *get_io_context(int gfp_flags)
3294{
3295 struct task_struct *tsk = current;
3296 unsigned long flags;
3297 struct io_context *ret;
3298
3299 local_irq_save(flags);
3300 ret = tsk->io_context;
3301 if (ret)
3302 goto out;
3303
3304 local_irq_restore(flags);
3305
3306 ret = kmem_cache_alloc(iocontext_cachep, gfp_flags);
3307 if (ret) {
3308 atomic_set(&ret->refcount, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003309 ret->task = current;
3310 ret->set_ioprio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003311 ret->last_waited = jiffies; /* doesn't matter... */
3312 ret->nr_batch_requests = 0; /* because this is 0 */
3313 ret->aic = NULL;
3314 ret->cic = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003315
3316 local_irq_save(flags);
3317
3318 /*
3319 * very unlikely, someone raced with us in setting up the task
3320 * io context. free new context and just grab a reference.
3321 */
3322 if (!tsk->io_context)
3323 tsk->io_context = ret;
3324 else {
3325 kmem_cache_free(iocontext_cachep, ret);
3326 ret = tsk->io_context;
3327 }
3328
3329out:
3330 atomic_inc(&ret->refcount);
3331 local_irq_restore(flags);
3332 }
3333
3334 return ret;
3335}
3336EXPORT_SYMBOL(get_io_context);
3337
3338void copy_io_context(struct io_context **pdst, struct io_context **psrc)
3339{
3340 struct io_context *src = *psrc;
3341 struct io_context *dst = *pdst;
3342
3343 if (src) {
3344 BUG_ON(atomic_read(&src->refcount) == 0);
3345 atomic_inc(&src->refcount);
3346 put_io_context(dst);
3347 *pdst = src;
3348 }
3349}
3350EXPORT_SYMBOL(copy_io_context);
3351
3352void swap_io_context(struct io_context **ioc1, struct io_context **ioc2)
3353{
3354 struct io_context *temp;
3355 temp = *ioc1;
3356 *ioc1 = *ioc2;
3357 *ioc2 = temp;
3358}
3359EXPORT_SYMBOL(swap_io_context);
3360
3361/*
3362 * sysfs parts below
3363 */
3364struct queue_sysfs_entry {
3365 struct attribute attr;
3366 ssize_t (*show)(struct request_queue *, char *);
3367 ssize_t (*store)(struct request_queue *, const char *, size_t);
3368};
3369
3370static ssize_t
3371queue_var_show(unsigned int var, char *page)
3372{
3373 return sprintf(page, "%d\n", var);
3374}
3375
3376static ssize_t
3377queue_var_store(unsigned long *var, const char *page, size_t count)
3378{
3379 char *p = (char *) page;
3380
3381 *var = simple_strtoul(p, &p, 10);
3382 return count;
3383}
3384
3385static ssize_t queue_requests_show(struct request_queue *q, char *page)
3386{
3387 return queue_var_show(q->nr_requests, (page));
3388}
3389
3390static ssize_t
3391queue_requests_store(struct request_queue *q, const char *page, size_t count)
3392{
3393 struct request_list *rl = &q->rq;
3394
3395 int ret = queue_var_store(&q->nr_requests, page, count);
3396 if (q->nr_requests < BLKDEV_MIN_RQ)
3397 q->nr_requests = BLKDEV_MIN_RQ;
3398 blk_queue_congestion_threshold(q);
3399
3400 if (rl->count[READ] >= queue_congestion_on_threshold(q))
3401 set_queue_congested(q, READ);
3402 else if (rl->count[READ] < queue_congestion_off_threshold(q))
3403 clear_queue_congested(q, READ);
3404
3405 if (rl->count[WRITE] >= queue_congestion_on_threshold(q))
3406 set_queue_congested(q, WRITE);
3407 else if (rl->count[WRITE] < queue_congestion_off_threshold(q))
3408 clear_queue_congested(q, WRITE);
3409
3410 if (rl->count[READ] >= q->nr_requests) {
3411 blk_set_queue_full(q, READ);
3412 } else if (rl->count[READ]+1 <= q->nr_requests) {
3413 blk_clear_queue_full(q, READ);
3414 wake_up(&rl->wait[READ]);
3415 }
3416
3417 if (rl->count[WRITE] >= q->nr_requests) {
3418 blk_set_queue_full(q, WRITE);
3419 } else if (rl->count[WRITE]+1 <= q->nr_requests) {
3420 blk_clear_queue_full(q, WRITE);
3421 wake_up(&rl->wait[WRITE]);
3422 }
3423 return ret;
3424}
3425
3426static ssize_t queue_ra_show(struct request_queue *q, char *page)
3427{
3428 int ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3429
3430 return queue_var_show(ra_kb, (page));
3431}
3432
3433static ssize_t
3434queue_ra_store(struct request_queue *q, const char *page, size_t count)
3435{
3436 unsigned long ra_kb;
3437 ssize_t ret = queue_var_store(&ra_kb, page, count);
3438
3439 spin_lock_irq(q->queue_lock);
3440 if (ra_kb > (q->max_sectors >> 1))
3441 ra_kb = (q->max_sectors >> 1);
3442
3443 q->backing_dev_info.ra_pages = ra_kb >> (PAGE_CACHE_SHIFT - 10);
3444 spin_unlock_irq(q->queue_lock);
3445
3446 return ret;
3447}
3448
3449static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
3450{
3451 int max_sectors_kb = q->max_sectors >> 1;
3452
3453 return queue_var_show(max_sectors_kb, (page));
3454}
3455
3456static ssize_t
3457queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
3458{
3459 unsigned long max_sectors_kb,
3460 max_hw_sectors_kb = q->max_hw_sectors >> 1,
3461 page_kb = 1 << (PAGE_CACHE_SHIFT - 10);
3462 ssize_t ret = queue_var_store(&max_sectors_kb, page, count);
3463 int ra_kb;
3464
3465 if (max_sectors_kb > max_hw_sectors_kb || max_sectors_kb < page_kb)
3466 return -EINVAL;
3467 /*
3468 * Take the queue lock to update the readahead and max_sectors
3469 * values synchronously:
3470 */
3471 spin_lock_irq(q->queue_lock);
3472 /*
3473 * Trim readahead window as well, if necessary:
3474 */
3475 ra_kb = q->backing_dev_info.ra_pages << (PAGE_CACHE_SHIFT - 10);
3476 if (ra_kb > max_sectors_kb)
3477 q->backing_dev_info.ra_pages =
3478 max_sectors_kb >> (PAGE_CACHE_SHIFT - 10);
3479
3480 q->max_sectors = max_sectors_kb << 1;
3481 spin_unlock_irq(q->queue_lock);
3482
3483 return ret;
3484}
3485
3486static ssize_t queue_max_hw_sectors_show(struct request_queue *q, char *page)
3487{
3488 int max_hw_sectors_kb = q->max_hw_sectors >> 1;
3489
3490 return queue_var_show(max_hw_sectors_kb, (page));
3491}
3492
3493
3494static struct queue_sysfs_entry queue_requests_entry = {
3495 .attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
3496 .show = queue_requests_show,
3497 .store = queue_requests_store,
3498};
3499
3500static struct queue_sysfs_entry queue_ra_entry = {
3501 .attr = {.name = "read_ahead_kb", .mode = S_IRUGO | S_IWUSR },
3502 .show = queue_ra_show,
3503 .store = queue_ra_store,
3504};
3505
3506static struct queue_sysfs_entry queue_max_sectors_entry = {
3507 .attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
3508 .show = queue_max_sectors_show,
3509 .store = queue_max_sectors_store,
3510};
3511
3512static struct queue_sysfs_entry queue_max_hw_sectors_entry = {
3513 .attr = {.name = "max_hw_sectors_kb", .mode = S_IRUGO },
3514 .show = queue_max_hw_sectors_show,
3515};
3516
3517static struct queue_sysfs_entry queue_iosched_entry = {
3518 .attr = {.name = "scheduler", .mode = S_IRUGO | S_IWUSR },
3519 .show = elv_iosched_show,
3520 .store = elv_iosched_store,
3521};
3522
3523static struct attribute *default_attrs[] = {
3524 &queue_requests_entry.attr,
3525 &queue_ra_entry.attr,
3526 &queue_max_hw_sectors_entry.attr,
3527 &queue_max_sectors_entry.attr,
3528 &queue_iosched_entry.attr,
3529 NULL,
3530};
3531
3532#define to_queue(atr) container_of((atr), struct queue_sysfs_entry, attr)
3533
3534static ssize_t
3535queue_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
3536{
3537 struct queue_sysfs_entry *entry = to_queue(attr);
3538 struct request_queue *q;
3539
3540 q = container_of(kobj, struct request_queue, kobj);
3541 if (!entry->show)
Dmitry Torokhov6c1852a2005-04-29 01:26:06 -05003542 return -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003543
3544 return entry->show(q, page);
3545}
3546
3547static ssize_t
3548queue_attr_store(struct kobject *kobj, struct attribute *attr,
3549 const char *page, size_t length)
3550{
3551 struct queue_sysfs_entry *entry = to_queue(attr);
3552 struct request_queue *q;
3553
3554 q = container_of(kobj, struct request_queue, kobj);
3555 if (!entry->store)
Dmitry Torokhov6c1852a2005-04-29 01:26:06 -05003556 return -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557
3558 return entry->store(q, page, length);
3559}
3560
3561static struct sysfs_ops queue_sysfs_ops = {
3562 .show = queue_attr_show,
3563 .store = queue_attr_store,
3564};
3565
Adrian Bunk93d17d32005-06-25 14:59:10 -07003566static struct kobj_type queue_ktype = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567 .sysfs_ops = &queue_sysfs_ops,
3568 .default_attrs = default_attrs,
3569};
3570
3571int blk_register_queue(struct gendisk *disk)
3572{
3573 int ret;
3574
3575 request_queue_t *q = disk->queue;
3576
3577 if (!q || !q->request_fn)
3578 return -ENXIO;
3579
3580 q->kobj.parent = kobject_get(&disk->kobj);
3581 if (!q->kobj.parent)
3582 return -EBUSY;
3583
3584 snprintf(q->kobj.name, KOBJ_NAME_LEN, "%s", "queue");
3585 q->kobj.ktype = &queue_ktype;
3586
3587 ret = kobject_register(&q->kobj);
3588 if (ret < 0)
3589 return ret;
3590
3591 ret = elv_register_queue(q);
3592 if (ret) {
3593 kobject_unregister(&q->kobj);
3594 return ret;
3595 }
3596
3597 return 0;
3598}
3599
3600void blk_unregister_queue(struct gendisk *disk)
3601{
3602 request_queue_t *q = disk->queue;
3603
3604 if (q && q->request_fn) {
3605 elv_unregister_queue(q);
3606
3607 kobject_unregister(&q->kobj);
3608 kobject_put(&disk->kobj);
3609 }
3610}