blob: 9e81addbd6ea7293ad3d2578d20bfeaf92cf8354 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * fs/direct-io.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * O_DIRECT
7 *
8 * 04Jul2002 akpm@zip.com.au
9 * Initial version
10 * 11Sep2002 janetinc@us.ibm.com
11 * added readv/writev support.
12 * 29Oct2002 akpm@zip.com.au
13 * rewrote bio_add_page() support.
14 * 30Oct2002 pbadari@us.ibm.com
15 * added support for non-aligned IO.
16 * 06Nov2002 pbadari@us.ibm.com
17 * added asynchronous IO support.
18 * 21Jul2003 nathans@sgi.com
19 * added IO completion notifier.
20 */
21
22#include <linux/kernel.h>
23#include <linux/module.h>
24#include <linux/types.h>
25#include <linux/fs.h>
26#include <linux/mm.h>
27#include <linux/slab.h>
28#include <linux/highmem.h>
29#include <linux/pagemap.h>
Andrew Morton98c4d572006-12-10 02:19:47 -080030#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070031#include <linux/bio.h>
32#include <linux/wait.h>
33#include <linux/err.h>
34#include <linux/blkdev.h>
35#include <linux/buffer_head.h>
36#include <linux/rwsem.h>
37#include <linux/uio.h>
38#include <asm/atomic.h>
39
40/*
41 * How many user pages to map in one call to get_user_pages(). This determines
42 * the size of a structure on the stack.
43 */
44#define DIO_PAGES 64
45
46/*
47 * This code generally works in units of "dio_blocks". A dio_block is
48 * somewhere between the hard sector size and the filesystem block size. it
49 * is determined on a per-invocation basis. When talking to the filesystem
50 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
51 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
52 * to bio_block quantities by shifting left by blkfactor.
53 *
54 * If blkfactor is zero then the user's request was aligned to the filesystem's
55 * blocksize.
56 *
57 * lock_type is DIO_LOCKING for regular files on direct-IO-naive filesystems.
58 * This determines whether we need to do the fancy locking which prevents
59 * direct-IO from being able to read uninitialised disk blocks. If its zero
Jes Sorensen1b1dcc12006-01-09 15:59:24 -080060 * (blockdev) this locking is not done, and if it is DIO_OWN_LOCKING i_mutex is
Linus Torvalds1da177e2005-04-16 15:20:36 -070061 * not held for the entire direct write (taken briefly, initially, during a
62 * direct read though, but its never held for the duration of a direct-IO).
63 */
64
65struct dio {
66 /* BIO submission state */
67 struct bio *bio; /* bio under assembly */
68 struct inode *inode;
69 int rw;
Daniel McNeil29504ff2005-04-16 15:25:50 -070070 loff_t i_size; /* i_size when submitted */
Linus Torvalds1da177e2005-04-16 15:20:36 -070071 int lock_type; /* doesn't change */
72 unsigned blkbits; /* doesn't change */
73 unsigned blkfactor; /* When we're using an alignment which
74 is finer than the filesystem's soft
75 blocksize, this specifies how much
76 finer. blkfactor=2 means 1/4-block
77 alignment. Does not change */
78 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
79 been performed at the start of a
80 write */
81 int pages_in_io; /* approximate total IO pages */
82 size_t size; /* total request size (doesn't change)*/
83 sector_t block_in_file; /* Current offset into the underlying
84 file in dio_block units. */
85 unsigned blocks_available; /* At block_in_file. changes */
86 sector_t final_block_in_request;/* doesn't change */
87 unsigned first_block_in_page; /* doesn't change, Used only once */
88 int boundary; /* prev block is at a boundary */
89 int reap_counter; /* rate limit reaping */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080090 get_block_t *get_block; /* block mapping function */
Linus Torvalds1da177e2005-04-16 15:20:36 -070091 dio_iodone_t *end_io; /* IO completion function */
92 sector_t final_block_in_bio; /* current final block in bio + 1 */
93 sector_t next_block_for_io; /* next block to be put under IO,
94 in dio_blocks units */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080095 struct buffer_head map_bh; /* last get_block() result */
Linus Torvalds1da177e2005-04-16 15:20:36 -070096
97 /*
98 * Deferred addition of a page to the dio. These variables are
99 * private to dio_send_cur_page(), submit_page_section() and
100 * dio_bio_add_page().
101 */
102 struct page *cur_page; /* The page */
103 unsigned cur_page_offset; /* Offset into it, in bytes */
104 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
105 sector_t cur_page_block; /* Where it starts */
106
107 /*
108 * Page fetching state. These variables belong to dio_refill_pages().
109 */
110 int curr_page; /* changes */
111 int total_pages; /* doesn't change */
112 unsigned long curr_user_address;/* changes */
113
114 /*
115 * Page queue. These variables belong to dio_refill_pages() and
116 * dio_get_page().
117 */
118 struct page *pages[DIO_PAGES]; /* page buffer */
119 unsigned head; /* next page to process */
120 unsigned tail; /* last valid page + 1 */
121 int page_errors; /* errno from get_user_pages() */
122
123 /* BIO completion state */
124 spinlock_t bio_lock; /* protects BIO fields below */
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800125 unsigned long refcount; /* direct_io_worker() and bios */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126 struct bio *bio_list; /* singly linked via bi_private */
127 struct task_struct *waiter; /* waiting task (NULL if none) */
128
129 /* AIO related stuff */
130 struct kiocb *iocb; /* kiocb */
131 int is_async; /* is IO async ? */
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800132 int io_error; /* IO error in completion path */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133 ssize_t result; /* IO result */
134};
135
136/*
137 * How many pages are in the queue?
138 */
139static inline unsigned dio_pages_present(struct dio *dio)
140{
141 return dio->tail - dio->head;
142}
143
144/*
145 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
146 */
147static int dio_refill_pages(struct dio *dio)
148{
149 int ret;
150 int nr_pages;
151
152 nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES);
153 down_read(&current->mm->mmap_sem);
154 ret = get_user_pages(
155 current, /* Task for fault acounting */
156 current->mm, /* whose pages? */
157 dio->curr_user_address, /* Where from? */
158 nr_pages, /* How many pages? */
159 dio->rw == READ, /* Write to memory? */
160 0, /* force (?) */
161 &dio->pages[0],
162 NULL); /* vmas */
163 up_read(&current->mm->mmap_sem);
164
Jens Axboeb31dc662006-06-13 08:26:10 +0200165 if (ret < 0 && dio->blocks_available && (dio->rw & WRITE)) {
Nick Piggin557ed1f2007-10-16 01:24:40 -0700166 struct page *page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 /*
168 * A memory fault, but the filesystem has some outstanding
169 * mapped blocks. We need to use those blocks up to avoid
170 * leaking stale data in the file.
171 */
172 if (dio->page_errors == 0)
173 dio->page_errors = ret;
Nick Pigginb5810032005-10-29 18:16:12 -0700174 page_cache_get(page);
175 dio->pages[0] = page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176 dio->head = 0;
177 dio->tail = 1;
178 ret = 0;
179 goto out;
180 }
181
182 if (ret >= 0) {
183 dio->curr_user_address += ret * PAGE_SIZE;
184 dio->curr_page += ret;
185 dio->head = 0;
186 dio->tail = ret;
187 ret = 0;
188 }
189out:
190 return ret;
191}
192
193/*
194 * Get another userspace page. Returns an ERR_PTR on error. Pages are
195 * buffered inside the dio so that we can call get_user_pages() against a
196 * decent number of pages, less frequently. To provide nicer use of the
197 * L1 cache.
198 */
199static struct page *dio_get_page(struct dio *dio)
200{
201 if (dio_pages_present(dio) == 0) {
202 int ret;
203
204 ret = dio_refill_pages(dio);
205 if (ret)
206 return ERR_PTR(ret);
207 BUG_ON(dio_pages_present(dio) == 0);
208 }
209 return dio->pages[dio->head++];
210}
211
Zach Brown6d544bb2006-12-10 02:20:54 -0800212/**
213 * dio_complete() - called when all DIO BIO I/O has been completed
214 * @offset: the byte offset in the file of the completed operation
215 *
216 * This releases locks as dictated by the locking type, lets interested parties
217 * know that a DIO operation has completed, and calculates the resulting return
218 * code for the operation.
219 *
220 * It lets the filesystem know if it registered an interest earlier via
221 * get_block. Pass the private field of the map buffer_head so that
222 * filesystems can use it to hold additional state between get_block calls and
223 * dio_complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800225static int dio_complete(struct dio *dio, loff_t offset, int ret)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226{
Zach Brown6d544bb2006-12-10 02:20:54 -0800227 ssize_t transferred = 0;
228
Zach Brown8459d862006-12-10 02:21:05 -0800229 /*
230 * AIO submission can race with bio completion to get here while
231 * expecting to have the last io completed by bio completion.
232 * In that case -EIOCBQUEUED is in fact not an error we want
233 * to preserve through this call.
234 */
235 if (ret == -EIOCBQUEUED)
236 ret = 0;
237
Zach Brown6d544bb2006-12-10 02:20:54 -0800238 if (dio->result) {
239 transferred = dio->result;
240
241 /* Check for short read case */
242 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
243 transferred = dio->i_size - offset;
244 }
245
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246 if (dio->end_io && dio->result)
Zach Brown6d544bb2006-12-10 02:20:54 -0800247 dio->end_io(dio->iocb, offset, transferred,
248 dio->map_bh.b_private);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249 if (dio->lock_type == DIO_LOCKING)
Ingo Molnard8aa9052006-07-03 00:25:02 -0700250 /* lockdep: non-owner release */
251 up_read_non_owner(&dio->inode->i_alloc_sem);
Zach Brown6d544bb2006-12-10 02:20:54 -0800252
253 if (ret == 0)
254 ret = dio->page_errors;
255 if (ret == 0)
256 ret = dio->io_error;
257 if (ret == 0)
258 ret = transferred;
259
260 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261}
262
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263static int dio_bio_complete(struct dio *dio, struct bio *bio);
264/*
265 * Asynchronous IO callback.
266 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200267static void dio_bio_end_aio(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268{
269 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800270 unsigned long remaining;
271 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273 /* cleanup the bio */
274 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800275
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800276 spin_lock_irqsave(&dio->bio_lock, flags);
277 remaining = --dio->refcount;
278 if (remaining == 1 && dio->waiter)
Zach Brown20258b22006-12-10 02:21:01 -0800279 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800280 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b22006-12-10 02:21:01 -0800281
Zach Brown8459d862006-12-10 02:21:05 -0800282 if (remaining == 0) {
283 int ret = dio_complete(dio, dio->iocb->ki_pos, 0);
284 aio_complete(dio->iocb, ret, 0);
285 kfree(dio);
286 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287}
288
289/*
290 * The BIO completion handler simply queues the BIO up for the process-context
291 * handler.
292 *
293 * During I/O bi_private points at the dio. After I/O, bi_private is used to
294 * implement a singly-linked list of completed BIOs, at dio->bio_list.
295 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200296static void dio_bio_end_io(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297{
298 struct dio *dio = bio->bi_private;
299 unsigned long flags;
300
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301 spin_lock_irqsave(&dio->bio_lock, flags);
302 bio->bi_private = dio->bio_list;
303 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800304 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305 wake_up_process(dio->waiter);
306 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307}
308
309static int
310dio_bio_alloc(struct dio *dio, struct block_device *bdev,
311 sector_t first_sector, int nr_vecs)
312{
313 struct bio *bio;
314
315 bio = bio_alloc(GFP_KERNEL, nr_vecs);
316 if (bio == NULL)
317 return -ENOMEM;
318
319 bio->bi_bdev = bdev;
320 bio->bi_sector = first_sector;
321 if (dio->is_async)
322 bio->bi_end_io = dio_bio_end_aio;
323 else
324 bio->bi_end_io = dio_bio_end_io;
325
326 dio->bio = bio;
327 return 0;
328}
329
330/*
331 * In the AIO read case we speculatively dirty the pages before starting IO.
332 * During IO completion, any of these pages which happen to have been written
333 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800334 *
335 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336 */
337static void dio_bio_submit(struct dio *dio)
338{
339 struct bio *bio = dio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800340 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341
342 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800343
344 spin_lock_irqsave(&dio->bio_lock, flags);
345 dio->refcount++;
346 spin_unlock_irqrestore(&dio->bio_lock, flags);
347
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 if (dio->is_async && dio->rw == READ)
349 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800350
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351 submit_bio(dio->rw, bio);
352
353 dio->bio = NULL;
354 dio->boundary = 0;
355}
356
357/*
358 * Release any resources in case of a failure
359 */
360static void dio_cleanup(struct dio *dio)
361{
362 while (dio_pages_present(dio))
363 page_cache_release(dio_get_page(dio));
364}
365
366/*
Zach Brown02732012006-12-10 02:20:59 -0800367 * Wait for the next BIO to complete. Remove it and return it. NULL is
368 * returned once all BIOs have been completed. This must only be called once
369 * all bios have been issued so that dio->refcount can only decrease. This
370 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371 */
372static struct bio *dio_await_one(struct dio *dio)
373{
374 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800375 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376
377 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800378
379 /*
380 * Wait as long as the list is empty and there are bios in flight. bio
381 * completion drops the count, maybe adds to the list, and wakes while
382 * holding the bio_lock so we don't need set_current_state()'s barrier
383 * and can call it after testing our condition.
384 */
385 while (dio->refcount > 1 && dio->bio_list == NULL) {
386 __set_current_state(TASK_UNINTERRUPTIBLE);
387 dio->waiter = current;
388 spin_unlock_irqrestore(&dio->bio_lock, flags);
389 io_schedule();
390 /* wake up sets us TASK_RUNNING */
391 spin_lock_irqsave(&dio->bio_lock, flags);
392 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393 }
Zach Brown02732012006-12-10 02:20:59 -0800394 if (dio->bio_list) {
395 bio = dio->bio_list;
396 dio->bio_list = bio->bi_private;
397 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 spin_unlock_irqrestore(&dio->bio_lock, flags);
399 return bio;
400}
401
402/*
403 * Process one completed BIO. No locks are held.
404 */
405static int dio_bio_complete(struct dio *dio, struct bio *bio)
406{
407 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
408 struct bio_vec *bvec = bio->bi_io_vec;
409 int page_no;
410
411 if (!uptodate)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800412 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413
414 if (dio->is_async && dio->rw == READ) {
415 bio_check_pages_dirty(bio); /* transfers ownership */
416 } else {
417 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
418 struct page *page = bvec[page_no].bv_page;
419
420 if (dio->rw == READ && !PageCompound(page))
421 set_page_dirty_lock(page);
422 page_cache_release(page);
423 }
424 bio_put(bio);
425 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426 return uptodate ? 0 : -EIO;
427}
428
429/*
Zach Brown02732012006-12-10 02:20:59 -0800430 * Wait on and process all in-flight BIOs. This must only be called once
431 * all bios have been issued so that the refcount can only decrease.
432 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200433 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800434 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800436static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437{
Zach Brown02732012006-12-10 02:20:59 -0800438 struct bio *bio;
439 do {
440 bio = dio_await_one(dio);
441 if (bio)
442 dio_bio_complete(dio, bio);
443 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444}
445
446/*
447 * A really large O_DIRECT read or write can generate a lot of BIOs. So
448 * to keep the memory consumption sane we periodically reap any completed BIOs
449 * during the BIO generation phase.
450 *
451 * This also helps to limit the peak amount of pinned userspace memory.
452 */
453static int dio_bio_reap(struct dio *dio)
454{
455 int ret = 0;
456
457 if (dio->reap_counter++ >= 64) {
458 while (dio->bio_list) {
459 unsigned long flags;
460 struct bio *bio;
461 int ret2;
462
463 spin_lock_irqsave(&dio->bio_lock, flags);
464 bio = dio->bio_list;
465 dio->bio_list = bio->bi_private;
466 spin_unlock_irqrestore(&dio->bio_lock, flags);
467 ret2 = dio_bio_complete(dio, bio);
468 if (ret == 0)
469 ret = ret2;
470 }
471 dio->reap_counter = 0;
472 }
473 return ret;
474}
475
476/*
477 * Call into the fs to map some more disk blocks. We record the current number
478 * of available blocks at dio->blocks_available. These are in units of the
479 * fs blocksize, (1 << inode->i_blkbits).
480 *
481 * The fs is allowed to map lots of blocks at once. If it wants to do that,
482 * it uses the passed inode-relative block number as the file offset, as usual.
483 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800484 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 * has remaining to do. The fs should not map more than this number of blocks.
486 *
487 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
488 * indicate how much contiguous disk space has been made available at
489 * bh->b_blocknr.
490 *
491 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
492 * This isn't very efficient...
493 *
494 * In the case of filesystem holes: the fs may return an arbitrarily-large
495 * hole by returning an appropriate value in b_size and by clearing
496 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800497 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 */
499static int get_more_blocks(struct dio *dio)
500{
501 int ret;
502 struct buffer_head *map_bh = &dio->map_bh;
503 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
504 unsigned long fs_count; /* Number of filesystem-sized blocks */
505 unsigned long dio_count;/* Number of dio_block-sized blocks */
506 unsigned long blkmask;
507 int create;
508
509 /*
510 * If there was a memory error and we've overwritten all the
511 * mapped blocks then we can now return that memory error
512 */
513 ret = dio->page_errors;
514 if (ret == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 BUG_ON(dio->block_in_file >= dio->final_block_in_request);
516 fs_startblk = dio->block_in_file >> dio->blkfactor;
517 dio_count = dio->final_block_in_request - dio->block_in_file;
518 fs_count = dio_count >> dio->blkfactor;
519 blkmask = (1 << dio->blkfactor) - 1;
520 if (dio_count & blkmask)
521 fs_count++;
522
Nathan Scott3c674e72006-03-29 09:26:15 +1000523 map_bh->b_state = 0;
524 map_bh->b_size = fs_count << dio->inode->i_blkbits;
525
Jens Axboeb31dc662006-06-13 08:26:10 +0200526 create = dio->rw & WRITE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527 if (dio->lock_type == DIO_LOCKING) {
528 if (dio->block_in_file < (i_size_read(dio->inode) >>
529 dio->blkbits))
530 create = 0;
531 } else if (dio->lock_type == DIO_NO_LOCKING) {
532 create = 0;
533 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000534
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 /*
536 * For writes inside i_size we forbid block creations: only
537 * overwrites are permitted. We fall back to buffered writes
538 * at a higher level for inside-i_size block-instantiating
539 * writes.
540 */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800541 ret = (*dio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542 map_bh, create);
543 }
544 return ret;
545}
546
547/*
548 * There is no bio. Make one now.
549 */
550static int dio_new_bio(struct dio *dio, sector_t start_sector)
551{
552 sector_t sector;
553 int ret, nr_pages;
554
555 ret = dio_bio_reap(dio);
556 if (ret)
557 goto out;
558 sector = start_sector << (dio->blkbits - 9);
559 nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
560 BUG_ON(nr_pages <= 0);
561 ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
562 dio->boundary = 0;
563out:
564 return ret;
565}
566
567/*
568 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
569 * that was successful then update final_block_in_bio and take a ref against
570 * the just-added page.
571 *
572 * Return zero on success. Non-zero means the caller needs to start a new BIO.
573 */
574static int dio_bio_add_page(struct dio *dio)
575{
576 int ret;
577
578 ret = bio_add_page(dio->bio, dio->cur_page,
579 dio->cur_page_len, dio->cur_page_offset);
580 if (ret == dio->cur_page_len) {
581 /*
582 * Decrement count only, if we are done with this page
583 */
584 if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE)
585 dio->pages_in_io--;
586 page_cache_get(dio->cur_page);
587 dio->final_block_in_bio = dio->cur_page_block +
588 (dio->cur_page_len >> dio->blkbits);
589 ret = 0;
590 } else {
591 ret = 1;
592 }
593 return ret;
594}
595
596/*
597 * Put cur_page under IO. The section of cur_page which is described by
598 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
599 * starts on-disk at cur_page_block.
600 *
601 * We take a ref against the page here (on behalf of its presence in the bio).
602 *
603 * The caller of this function is responsible for removing cur_page from the
604 * dio, and for dropping the refcount which came from that presence.
605 */
606static int dio_send_cur_page(struct dio *dio)
607{
608 int ret = 0;
609
610 if (dio->bio) {
611 /*
612 * See whether this new request is contiguous with the old
613 */
614 if (dio->final_block_in_bio != dio->cur_page_block)
615 dio_bio_submit(dio);
616 /*
617 * Submit now if the underlying fs is about to perform a
618 * metadata read
619 */
620 if (dio->boundary)
621 dio_bio_submit(dio);
622 }
623
624 if (dio->bio == NULL) {
625 ret = dio_new_bio(dio, dio->cur_page_block);
626 if (ret)
627 goto out;
628 }
629
630 if (dio_bio_add_page(dio) != 0) {
631 dio_bio_submit(dio);
632 ret = dio_new_bio(dio, dio->cur_page_block);
633 if (ret == 0) {
634 ret = dio_bio_add_page(dio);
635 BUG_ON(ret != 0);
636 }
637 }
638out:
639 return ret;
640}
641
642/*
643 * An autonomous function to put a chunk of a page under deferred IO.
644 *
645 * The caller doesn't actually know (or care) whether this piece of page is in
646 * a BIO, or is under IO or whatever. We just take care of all possible
647 * situations here. The separation between the logic of do_direct_IO() and
648 * that of submit_page_section() is important for clarity. Please don't break.
649 *
650 * The chunk of page starts on-disk at blocknr.
651 *
652 * We perform deferred IO, by recording the last-submitted page inside our
653 * private part of the dio structure. If possible, we just expand the IO
654 * across that page here.
655 *
656 * If that doesn't work out then we put the old page into the bio and add this
657 * page to the dio instead.
658 */
659static int
660submit_page_section(struct dio *dio, struct page *page,
661 unsigned offset, unsigned len, sector_t blocknr)
662{
663 int ret = 0;
664
Andrew Morton98c4d572006-12-10 02:19:47 -0800665 if (dio->rw & WRITE) {
666 /*
667 * Read accounting is performed in submit_bio()
668 */
669 task_io_account_write(len);
670 }
671
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 /*
673 * Can we just grow the current page's presence in the dio?
674 */
675 if ( (dio->cur_page == page) &&
676 (dio->cur_page_offset + dio->cur_page_len == offset) &&
677 (dio->cur_page_block +
678 (dio->cur_page_len >> dio->blkbits) == blocknr)) {
679 dio->cur_page_len += len;
680
681 /*
682 * If dio->boundary then we want to schedule the IO now to
683 * avoid metadata seeks.
684 */
685 if (dio->boundary) {
686 ret = dio_send_cur_page(dio);
687 page_cache_release(dio->cur_page);
688 dio->cur_page = NULL;
689 }
690 goto out;
691 }
692
693 /*
694 * If there's a deferred page already there then send it.
695 */
696 if (dio->cur_page) {
697 ret = dio_send_cur_page(dio);
698 page_cache_release(dio->cur_page);
699 dio->cur_page = NULL;
700 if (ret)
701 goto out;
702 }
703
704 page_cache_get(page); /* It is in dio */
705 dio->cur_page = page;
706 dio->cur_page_offset = offset;
707 dio->cur_page_len = len;
708 dio->cur_page_block = blocknr;
709out:
710 return ret;
711}
712
713/*
714 * Clean any dirty buffers in the blockdev mapping which alias newly-created
715 * file blocks. Only called for S_ISREG files - blockdevs do not set
716 * buffer_new
717 */
718static void clean_blockdev_aliases(struct dio *dio)
719{
720 unsigned i;
721 unsigned nblocks;
722
723 nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
724
725 for (i = 0; i < nblocks; i++) {
726 unmap_underlying_metadata(dio->map_bh.b_bdev,
727 dio->map_bh.b_blocknr + i);
728 }
729}
730
731/*
732 * If we are not writing the entire block and get_block() allocated
733 * the block for us, we need to fill-in the unused portion of the
734 * block with zeros. This happens only if user-buffer, fileoffset or
735 * io length is not filesystem block-size multiple.
736 *
737 * `end' is zero if we're doing the start of the IO, 1 at the end of the
738 * IO.
739 */
740static void dio_zero_block(struct dio *dio, int end)
741{
742 unsigned dio_blocks_per_fs_block;
743 unsigned this_chunk_blocks; /* In dio_blocks */
744 unsigned this_chunk_bytes;
745 struct page *page;
746
747 dio->start_zero_done = 1;
748 if (!dio->blkfactor || !buffer_new(&dio->map_bh))
749 return;
750
751 dio_blocks_per_fs_block = 1 << dio->blkfactor;
752 this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
753
754 if (!this_chunk_blocks)
755 return;
756
757 /*
758 * We need to zero out part of an fs block. It is either at the
759 * beginning or the end of the fs block.
760 */
761 if (end)
762 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
763
764 this_chunk_bytes = this_chunk_blocks << dio->blkbits;
765
Nick Piggin557ed1f2007-10-16 01:24:40 -0700766 page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767 if (submit_page_section(dio, page, 0, this_chunk_bytes,
768 dio->next_block_for_io))
769 return;
770
771 dio->next_block_for_io += this_chunk_blocks;
772}
773
774/*
775 * Walk the user pages, and the file, mapping blocks to disk and generating
776 * a sequence of (page,offset,len,block) mappings. These mappings are injected
777 * into submit_page_section(), which takes care of the next stage of submission
778 *
779 * Direct IO against a blockdev is different from a file. Because we can
780 * happily perform page-sized but 512-byte aligned IOs. It is important that
781 * blockdev IO be able to have fine alignment and large sizes.
782 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800783 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 * with the size of IO which is permitted at this offset and this i_blkbits.
785 *
786 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800787 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 * fine alignment but still allows this function to work in PAGE_SIZE units.
789 */
790static int do_direct_IO(struct dio *dio)
791{
792 const unsigned blkbits = dio->blkbits;
793 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
794 struct page *page;
795 unsigned block_in_page;
796 struct buffer_head *map_bh = &dio->map_bh;
797 int ret = 0;
798
799 /* The I/O can start at any block offset within the first page */
800 block_in_page = dio->first_block_in_page;
801
802 while (dio->block_in_file < dio->final_block_in_request) {
803 page = dio_get_page(dio);
804 if (IS_ERR(page)) {
805 ret = PTR_ERR(page);
806 goto out;
807 }
808
809 while (block_in_page < blocks_per_page) {
810 unsigned offset_in_page = block_in_page << blkbits;
811 unsigned this_chunk_bytes; /* # of bytes mapped */
812 unsigned this_chunk_blocks; /* # of blocks */
813 unsigned u;
814
815 if (dio->blocks_available == 0) {
816 /*
817 * Need to go and map some more disk
818 */
819 unsigned long blkmask;
820 unsigned long dio_remainder;
821
822 ret = get_more_blocks(dio);
823 if (ret) {
824 page_cache_release(page);
825 goto out;
826 }
827 if (!buffer_mapped(map_bh))
828 goto do_holes;
829
830 dio->blocks_available =
831 map_bh->b_size >> dio->blkbits;
832 dio->next_block_for_io =
833 map_bh->b_blocknr << dio->blkfactor;
834 if (buffer_new(map_bh))
835 clean_blockdev_aliases(dio);
836
837 if (!dio->blkfactor)
838 goto do_holes;
839
840 blkmask = (1 << dio->blkfactor) - 1;
841 dio_remainder = (dio->block_in_file & blkmask);
842
843 /*
844 * If we are at the start of IO and that IO
845 * starts partway into a fs-block,
846 * dio_remainder will be non-zero. If the IO
847 * is a read then we can simply advance the IO
848 * cursor to the first block which is to be
849 * read. But if the IO is a write and the
850 * block was newly allocated we cannot do that;
851 * the start of the fs block must be zeroed out
852 * on-disk
853 */
854 if (!buffer_new(map_bh))
855 dio->next_block_for_io += dio_remainder;
856 dio->blocks_available -= dio_remainder;
857 }
858do_holes:
859 /* Handle holes */
860 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800861 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862
863 /* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboeb31dc662006-06-13 08:26:10 +0200864 if (dio->rw & WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865 page_cache_release(page);
866 return -ENOTBLK;
867 }
868
Jeff Moyer35dc8162006-02-03 03:04:27 -0800869 /*
870 * Be sure to account for a partial block as the
871 * last block in the file
872 */
873 i_size_aligned = ALIGN(i_size_read(dio->inode),
874 1 << blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875 if (dio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -0800876 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877 /* We hit eof */
878 page_cache_release(page);
879 goto out;
880 }
Christoph Lametereebd2aa2008-02-04 22:28:29 -0800881 zero_user(page, block_in_page << blkbits,
882 1 << blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 dio->block_in_file++;
884 block_in_page++;
885 goto next_block;
886 }
887
888 /*
889 * If we're performing IO which has an alignment which
890 * is finer than the underlying fs, go check to see if
891 * we must zero out the start of this block.
892 */
893 if (unlikely(dio->blkfactor && !dio->start_zero_done))
894 dio_zero_block(dio, 0);
895
896 /*
897 * Work out, in this_chunk_blocks, how much disk we
898 * can add to this page
899 */
900 this_chunk_blocks = dio->blocks_available;
901 u = (PAGE_SIZE - offset_in_page) >> blkbits;
902 if (this_chunk_blocks > u)
903 this_chunk_blocks = u;
904 u = dio->final_block_in_request - dio->block_in_file;
905 if (this_chunk_blocks > u)
906 this_chunk_blocks = u;
907 this_chunk_bytes = this_chunk_blocks << blkbits;
908 BUG_ON(this_chunk_bytes == 0);
909
910 dio->boundary = buffer_boundary(map_bh);
911 ret = submit_page_section(dio, page, offset_in_page,
912 this_chunk_bytes, dio->next_block_for_io);
913 if (ret) {
914 page_cache_release(page);
915 goto out;
916 }
917 dio->next_block_for_io += this_chunk_blocks;
918
919 dio->block_in_file += this_chunk_blocks;
920 block_in_page += this_chunk_blocks;
921 dio->blocks_available -= this_chunk_blocks;
922next_block:
Eric Sesterhennd4569d22006-04-01 01:10:13 +0200923 BUG_ON(dio->block_in_file > dio->final_block_in_request);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924 if (dio->block_in_file == dio->final_block_in_request)
925 break;
926 }
927
928 /* Drop the ref which was taken in get_user_pages() */
929 page_cache_release(page);
930 block_in_page = 0;
931 }
932out:
933 return ret;
934}
935
936/*
Jes Sorensen1b1dcc12006-01-09 15:59:24 -0800937 * Releases both i_mutex and i_alloc_sem
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 */
939static ssize_t
940direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
941 const struct iovec *iov, loff_t offset, unsigned long nr_segs,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800942 unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 struct dio *dio)
944{
945 unsigned long user_addr;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800946 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 int seg;
948 ssize_t ret = 0;
949 ssize_t ret2;
950 size_t bytes;
951
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952 dio->inode = inode;
953 dio->rw = rw;
954 dio->blkbits = blkbits;
955 dio->blkfactor = inode->i_blkbits - blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 dio->block_in_file = offset >> blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800958 dio->get_block = get_block;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 dio->end_io = end_io;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960 dio->final_block_in_bio = -1;
961 dio->next_block_for_io = -1;
962
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 dio->iocb = iocb;
Daniel McNeil29504ff2005-04-16 15:25:50 -0700964 dio->i_size = i_size_read(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966 spin_lock_init(&dio->bio_lock);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800967 dio->refcount = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968
969 /*
970 * In case of non-aligned buffers, we may need 2 more
971 * pages since we need to zero out first and last block.
972 */
973 if (unlikely(dio->blkfactor))
974 dio->pages_in_io = 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975
976 for (seg = 0; seg < nr_segs; seg++) {
977 user_addr = (unsigned long)iov[seg].iov_base;
978 dio->pages_in_io +=
979 ((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
980 - user_addr/PAGE_SIZE);
981 }
982
983 for (seg = 0; seg < nr_segs; seg++) {
984 user_addr = (unsigned long)iov[seg].iov_base;
985 dio->size += bytes = iov[seg].iov_len;
986
987 /* Index into the first page of the first block */
988 dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
989 dio->final_block_in_request = dio->block_in_file +
990 (bytes >> blkbits);
991 /* Page fetching state */
992 dio->head = 0;
993 dio->tail = 0;
994 dio->curr_page = 0;
995
996 dio->total_pages = 0;
997 if (user_addr & (PAGE_SIZE-1)) {
998 dio->total_pages++;
999 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1000 }
1001 dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1002 dio->curr_user_address = user_addr;
1003
1004 ret = do_direct_IO(dio);
1005
1006 dio->result += iov[seg].iov_len -
1007 ((dio->final_block_in_request - dio->block_in_file) <<
1008 blkbits);
1009
1010 if (ret) {
1011 dio_cleanup(dio);
1012 break;
1013 }
1014 } /* end iovec loop */
1015
Jens Axboeb31dc662006-06-13 08:26:10 +02001016 if (ret == -ENOTBLK && (rw & WRITE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017 /*
1018 * The remaining part of the request will be
1019 * be handled by buffered I/O when we return
1020 */
1021 ret = 0;
1022 }
1023 /*
1024 * There may be some unwritten disk at the end of a part-written
1025 * fs-block-sized block. Go zero that now.
1026 */
1027 dio_zero_block(dio, 1);
1028
1029 if (dio->cur_page) {
1030 ret2 = dio_send_cur_page(dio);
1031 if (ret == 0)
1032 ret = ret2;
1033 page_cache_release(dio->cur_page);
1034 dio->cur_page = NULL;
1035 }
1036 if (dio->bio)
1037 dio_bio_submit(dio);
1038
Zach Brown17a7b1d2006-12-10 02:20:56 -08001039 /* All IO is now issued, send it on its way */
1040 blk_run_address_space(inode->i_mapping);
1041
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042 /*
1043 * It is possible that, we return short IO due to end of file.
1044 * In that case, we need to release all the pages we got hold on.
1045 */
1046 dio_cleanup(dio);
1047
1048 /*
1049 * All block lookups have been performed. For READ requests
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001050 * we can let i_mutex go now that its achieved its purpose
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051 * of protecting us from looking up uninitialized blocks.
1052 */
1053 if ((rw == READ) && (dio->lock_type == DIO_LOCKING))
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001054 mutex_unlock(&dio->inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055
1056 /*
Zach Brown8459d862006-12-10 02:21:05 -08001057 * The only time we want to leave bios in flight is when a successful
1058 * partial aio read or full aio write have been setup. In that case
1059 * bio completion will call aio_complete. The only time it's safe to
1060 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1061 * This had *better* be the only place that raises -EIOCBQUEUED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 */
Zach Brown8459d862006-12-10 02:21:05 -08001063 BUG_ON(ret == -EIOCBQUEUED);
1064 if (dio->is_async && ret == 0 && dio->result &&
1065 ((rw & READ) || (dio->result == dio->size)))
1066 ret = -EIOCBQUEUED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001067
Zach Brown8459d862006-12-10 02:21:05 -08001068 if (ret != -EIOCBQUEUED)
Zach Brown6d544bb2006-12-10 02:20:54 -08001069 dio_await_completion(dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070
Zach Brown8459d862006-12-10 02:21:05 -08001071 /*
1072 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001073 * operation. AIO can if it was a broken operation described above or
1074 * in fact if all the bios race to complete before we get here. In
1075 * that case dio_complete() translates the EIOCBQUEUED into the proper
1076 * return code that the caller will hand to aio_complete().
1077 *
1078 * This is managed by the bio_lock instead of being an atomic_t so that
1079 * completion paths can drop their ref and use the remaining count to
1080 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001081 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001082 spin_lock_irqsave(&dio->bio_lock, flags);
1083 ret2 = --dio->refcount;
1084 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brownfcb82f82007-07-03 15:28:55 -07001085
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001086 if (ret2 == 0) {
Zach Brown6d544bb2006-12-10 02:20:54 -08001087 ret = dio_complete(dio, offset, ret);
Zach Brown8459d862006-12-10 02:21:05 -08001088 kfree(dio);
1089 } else
1090 BUG_ON(ret != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 return ret;
1093}
1094
1095/*
1096 * This is a library function for use by filesystem drivers.
1097 * The locking rules are governed by the dio_lock_type parameter.
1098 *
1099 * DIO_NO_LOCKING (no locking, for raw block device access)
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001100 * For writes, i_mutex is not held on entry; it is never taken.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 *
1102 * DIO_LOCKING (simple locking for regular files)
Nathan Scott3fb962b2006-03-15 15:14:45 +11001103 * For writes we are called under i_mutex and return with i_mutex held, even
1104 * though it is internally dropped.
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001105 * For reads, i_mutex is not held on entry, but it is taken and dropped before
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106 * returning.
1107 *
1108 * DIO_OWN_LOCKING (filesystem provides synchronisation and handling of
1109 * uninitialised data, allowing parallel direct readers and writers)
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001110 * For writes we are called without i_mutex, return without it, never touch it.
Nathan Scott3fb962b2006-03-15 15:14:45 +11001111 * For reads we are called under i_mutex and return with i_mutex held, even
1112 * though it may be internally dropped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 *
1114 * Additional i_alloc_sem locking requirements described inline below.
1115 */
1116ssize_t
1117__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1118 struct block_device *bdev, const struct iovec *iov, loff_t offset,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -08001119 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 int dio_lock_type)
1121{
1122 int seg;
1123 size_t size;
1124 unsigned long addr;
1125 unsigned blkbits = inode->i_blkbits;
1126 unsigned bdev_blkbits = 0;
1127 unsigned blocksize_mask = (1 << blkbits) - 1;
1128 ssize_t retval = -EINVAL;
1129 loff_t end = offset;
1130 struct dio *dio;
Nathan Scott3fb962b2006-03-15 15:14:45 +11001131 int release_i_mutex = 0;
1132 int acquire_i_mutex = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133
1134 if (rw & WRITE)
Jens Axboeb31dc662006-06-13 08:26:10 +02001135 rw = WRITE_SYNC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136
1137 if (bdev)
1138 bdev_blkbits = blksize_bits(bdev_hardsect_size(bdev));
1139
1140 if (offset & blocksize_mask) {
1141 if (bdev)
1142 blkbits = bdev_blkbits;
1143 blocksize_mask = (1 << blkbits) - 1;
1144 if (offset & blocksize_mask)
1145 goto out;
1146 }
1147
1148 /* Check the memory alignment. Blocks cannot straddle pages */
1149 for (seg = 0; seg < nr_segs; seg++) {
1150 addr = (unsigned long)iov[seg].iov_base;
1151 size = iov[seg].iov_len;
1152 end += size;
1153 if ((addr & blocksize_mask) || (size & blocksize_mask)) {
1154 if (bdev)
1155 blkbits = bdev_blkbits;
1156 blocksize_mask = (1 << blkbits) - 1;
1157 if ((addr & blocksize_mask) || (size & blocksize_mask))
1158 goto out;
1159 }
1160 }
1161
Zach Brown848c4dd2007-08-20 17:12:01 -07001162 dio = kzalloc(sizeof(*dio), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163 retval = -ENOMEM;
1164 if (!dio)
1165 goto out;
1166
1167 /*
1168 * For block device access DIO_NO_LOCKING is used,
1169 * neither readers nor writers do any locking at all
1170 * For regular files using DIO_LOCKING,
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001171 * readers need to grab i_mutex and i_alloc_sem
1172 * writers need to grab i_alloc_sem only (i_mutex is already held)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173 * For regular files using DIO_OWN_LOCKING,
1174 * neither readers nor writers take any locks here
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 */
1176 dio->lock_type = dio_lock_type;
1177 if (dio_lock_type != DIO_NO_LOCKING) {
1178 /* watch out for a 0 len io from a tricksy fs */
1179 if (rw == READ && end > offset) {
1180 struct address_space *mapping;
1181
1182 mapping = iocb->ki_filp->f_mapping;
1183 if (dio_lock_type != DIO_OWN_LOCKING) {
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001184 mutex_lock(&inode->i_mutex);
Nathan Scott3fb962b2006-03-15 15:14:45 +11001185 release_i_mutex = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 }
1187
1188 retval = filemap_write_and_wait_range(mapping, offset,
1189 end - 1);
1190 if (retval) {
1191 kfree(dio);
1192 goto out;
1193 }
1194
1195 if (dio_lock_type == DIO_OWN_LOCKING) {
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001196 mutex_unlock(&inode->i_mutex);
Nathan Scott3fb962b2006-03-15 15:14:45 +11001197 acquire_i_mutex = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198 }
1199 }
1200
1201 if (dio_lock_type == DIO_LOCKING)
Ingo Molnard8aa9052006-07-03 00:25:02 -07001202 /* lockdep: not the owner will release it */
1203 down_read_non_owner(&inode->i_alloc_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204 }
1205
1206 /*
1207 * For file extending writes updating i_size before data
1208 * writeouts complete can expose uninitialized blocks. So
1209 * even for AIO, we need to wait for i/o to complete before
1210 * returning in this case.
1211 */
Jens Axboeb31dc662006-06-13 08:26:10 +02001212 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213 (end > i_size_read(inode)));
1214
1215 retval = direct_io_worker(rw, iocb, inode, iov, offset,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -08001216 nr_segs, blkbits, get_block, end_io, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217
1218 if (rw == READ && dio_lock_type == DIO_LOCKING)
Nathan Scott3fb962b2006-03-15 15:14:45 +11001219 release_i_mutex = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220
1221out:
Nathan Scott3fb962b2006-03-15 15:14:45 +11001222 if (release_i_mutex)
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001223 mutex_unlock(&inode->i_mutex);
Nathan Scott3fb962b2006-03-15 15:14:45 +11001224 else if (acquire_i_mutex)
1225 mutex_lock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 return retval;
1227}
1228EXPORT_SYMBOL(__blockdev_direct_IO);