blob: 7600aacf531dc8ed16ccfb727878d13834a16503 [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 *
Francois Camie1f8e872008-10-15 22:01:59 -07008 * 04Jul2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 * Initial version
10 * 11Sep2002 janetinc@us.ibm.com
11 * added readv/writev support.
Francois Camie1f8e872008-10-15 22:01:59 -070012 * 29Oct2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070013 * 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.
Linus Torvalds1da177e2005-04-16 15:20:36 -070056 */
57
58struct dio {
59 /* BIO submission state */
60 struct bio *bio; /* bio under assembly */
61 struct inode *inode;
62 int rw;
Daniel McNeil29504ff2005-04-16 15:25:50 -070063 loff_t i_size; /* i_size when submitted */
Christoph Hellwig5fe878a2009-12-15 16:47:50 -080064 int flags; /* doesn't change */
Linus Torvalds1da177e2005-04-16 15:20:36 -070065 unsigned blkbits; /* doesn't change */
66 unsigned blkfactor; /* When we're using an alignment which
67 is finer than the filesystem's soft
68 blocksize, this specifies how much
69 finer. blkfactor=2 means 1/4-block
70 alignment. Does not change */
71 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
72 been performed at the start of a
73 write */
74 int pages_in_io; /* approximate total IO pages */
75 size_t size; /* total request size (doesn't change)*/
76 sector_t block_in_file; /* Current offset into the underlying
77 file in dio_block units. */
78 unsigned blocks_available; /* At block_in_file. changes */
79 sector_t final_block_in_request;/* doesn't change */
80 unsigned first_block_in_page; /* doesn't change, Used only once */
81 int boundary; /* prev block is at a boundary */
82 int reap_counter; /* rate limit reaping */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080083 get_block_t *get_block; /* block mapping function */
Linus Torvalds1da177e2005-04-16 15:20:36 -070084 dio_iodone_t *end_io; /* IO completion function */
Josef Bacikfacd07b2010-05-23 11:00:55 -040085 dio_submit_t *submit_io; /* IO submition function */
86 loff_t logical_offset_in_bio; /* current first logical block in bio */
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 sector_t final_block_in_bio; /* current final block in bio + 1 */
88 sector_t next_block_for_io; /* next block to be put under IO,
89 in dio_blocks units */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080090 struct buffer_head map_bh; /* last get_block() result */
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
92 /*
93 * Deferred addition of a page to the dio. These variables are
94 * private to dio_send_cur_page(), submit_page_section() and
95 * dio_bio_add_page().
96 */
97 struct page *cur_page; /* The page */
98 unsigned cur_page_offset; /* Offset into it, in bytes */
99 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
100 sector_t cur_page_block; /* Where it starts */
Josef Bacikfacd07b2010-05-23 11:00:55 -0400101 loff_t cur_page_fs_offset; /* Offset in file */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103 /* BIO completion state */
104 spinlock_t bio_lock; /* protects BIO fields below */
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800105 unsigned long refcount; /* direct_io_worker() and bios */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106 struct bio *bio_list; /* singly linked via bi_private */
107 struct task_struct *waiter; /* waiting task (NULL if none) */
108
109 /* AIO related stuff */
110 struct kiocb *iocb; /* kiocb */
111 int is_async; /* is IO async ? */
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800112 int io_error; /* IO error in completion path */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 ssize_t result; /* IO result */
Jeff Moyer23aee092009-12-15 16:47:49 -0800114
115 /*
116 * Page fetching state. These variables belong to dio_refill_pages().
117 */
118 int curr_page; /* changes */
119 int total_pages; /* doesn't change */
120 unsigned long curr_user_address;/* changes */
121
122 /*
123 * Page queue. These variables belong to dio_refill_pages() and
124 * dio_get_page().
125 */
126 unsigned head; /* next page to process */
127 unsigned tail; /* last valid page + 1 */
128 int page_errors; /* errno from get_user_pages() */
129
130 /*
131 * pages[] (and any fields placed after it) are not zeroed out at
132 * allocation time. Don't add new fields after pages[] unless you
133 * wish that they not be zeroed.
134 */
135 struct page *pages[DIO_PAGES]; /* page buffer */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136};
137
138/*
139 * How many pages are in the queue?
140 */
141static inline unsigned dio_pages_present(struct dio *dio)
142{
143 return dio->tail - dio->head;
144}
145
146/*
147 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
148 */
149static int dio_refill_pages(struct dio *dio)
150{
151 int ret;
152 int nr_pages;
153
154 nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES);
Nick Pigginf5dd33c2008-07-25 19:45:25 -0700155 ret = get_user_pages_fast(
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156 dio->curr_user_address, /* Where from? */
157 nr_pages, /* How many pages? */
158 dio->rw == READ, /* Write to memory? */
Nick Pigginf5dd33c2008-07-25 19:45:25 -0700159 &dio->pages[0]); /* Put results here */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160
Jens Axboeb31dc662006-06-13 08:26:10 +0200161 if (ret < 0 && dio->blocks_available && (dio->rw & WRITE)) {
Nick Piggin557ed1f2007-10-16 01:24:40 -0700162 struct page *page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163 /*
164 * A memory fault, but the filesystem has some outstanding
165 * mapped blocks. We need to use those blocks up to avoid
166 * leaking stale data in the file.
167 */
168 if (dio->page_errors == 0)
169 dio->page_errors = ret;
Nick Pigginb5810032005-10-29 18:16:12 -0700170 page_cache_get(page);
171 dio->pages[0] = page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 dio->head = 0;
173 dio->tail = 1;
174 ret = 0;
175 goto out;
176 }
177
178 if (ret >= 0) {
179 dio->curr_user_address += ret * PAGE_SIZE;
180 dio->curr_page += ret;
181 dio->head = 0;
182 dio->tail = ret;
183 ret = 0;
184 }
185out:
186 return ret;
187}
188
189/*
190 * Get another userspace page. Returns an ERR_PTR on error. Pages are
191 * buffered inside the dio so that we can call get_user_pages() against a
192 * decent number of pages, less frequently. To provide nicer use of the
193 * L1 cache.
194 */
195static struct page *dio_get_page(struct dio *dio)
196{
197 if (dio_pages_present(dio) == 0) {
198 int ret;
199
200 ret = dio_refill_pages(dio);
201 if (ret)
202 return ERR_PTR(ret);
203 BUG_ON(dio_pages_present(dio) == 0);
204 }
205 return dio->pages[dio->head++];
206}
207
Zach Brown6d544bb2006-12-10 02:20:54 -0800208/**
209 * dio_complete() - called when all DIO BIO I/O has been completed
210 * @offset: the byte offset in the file of the completed operation
211 *
212 * This releases locks as dictated by the locking type, lets interested parties
213 * know that a DIO operation has completed, and calculates the resulting return
214 * code for the operation.
215 *
216 * It lets the filesystem know if it registered an interest earlier via
217 * get_block. Pass the private field of the map buffer_head so that
218 * filesystems can use it to hold additional state between get_block calls and
219 * dio_complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800221static int dio_complete(struct dio *dio, loff_t offset, int ret)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222{
Zach Brown6d544bb2006-12-10 02:20:54 -0800223 ssize_t transferred = 0;
224
Zach Brown8459d862006-12-10 02:21:05 -0800225 /*
226 * AIO submission can race with bio completion to get here while
227 * expecting to have the last io completed by bio completion.
228 * In that case -EIOCBQUEUED is in fact not an error we want
229 * to preserve through this call.
230 */
231 if (ret == -EIOCBQUEUED)
232 ret = 0;
233
Zach Brown6d544bb2006-12-10 02:20:54 -0800234 if (dio->result) {
235 transferred = dio->result;
236
237 /* Check for short read case */
238 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
239 transferred = dio->i_size - offset;
240 }
241
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242 if (dio->end_io && dio->result)
Zach Brown6d544bb2006-12-10 02:20:54 -0800243 dio->end_io(dio->iocb, offset, transferred,
244 dio->map_bh.b_private);
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800245
246 if (dio->flags & DIO_LOCKING)
Ingo Molnard8aa9052006-07-03 00:25:02 -0700247 /* lockdep: non-owner release */
248 up_read_non_owner(&dio->inode->i_alloc_sem);
Zach Brown6d544bb2006-12-10 02:20:54 -0800249
250 if (ret == 0)
251 ret = dio->page_errors;
252 if (ret == 0)
253 ret = dio->io_error;
254 if (ret == 0)
255 ret = transferred;
256
257 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258}
259
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260static int dio_bio_complete(struct dio *dio, struct bio *bio);
261/*
262 * Asynchronous IO callback.
263 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200264static void dio_bio_end_aio(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265{
266 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800267 unsigned long remaining;
268 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270 /* cleanup the bio */
271 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800272
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800273 spin_lock_irqsave(&dio->bio_lock, flags);
274 remaining = --dio->refcount;
275 if (remaining == 1 && dio->waiter)
Zach Brown20258b22006-12-10 02:21:01 -0800276 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800277 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b22006-12-10 02:21:01 -0800278
Zach Brown8459d862006-12-10 02:21:05 -0800279 if (remaining == 0) {
280 int ret = dio_complete(dio, dio->iocb->ki_pos, 0);
281 aio_complete(dio->iocb, ret, 0);
282 kfree(dio);
283 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284}
285
286/*
287 * The BIO completion handler simply queues the BIO up for the process-context
288 * handler.
289 *
290 * During I/O bi_private points at the dio. After I/O, bi_private is used to
291 * implement a singly-linked list of completed BIOs, at dio->bio_list.
292 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200293static void dio_bio_end_io(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294{
295 struct dio *dio = bio->bi_private;
296 unsigned long flags;
297
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298 spin_lock_irqsave(&dio->bio_lock, flags);
299 bio->bi_private = dio->bio_list;
300 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800301 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302 wake_up_process(dio->waiter);
303 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304}
305
Josef Bacikfacd07b2010-05-23 11:00:55 -0400306/**
307 * dio_end_io - handle the end io action for the given bio
308 * @bio: The direct io bio thats being completed
309 * @error: Error if there was one
310 *
311 * This is meant to be called by any filesystem that uses their own dio_submit_t
312 * so that the DIO specific endio actions are dealt with after the filesystem
313 * has done it's completion work.
314 */
315void dio_end_io(struct bio *bio, int error)
316{
317 struct dio *dio = bio->bi_private;
318
319 if (dio->is_async)
320 dio_bio_end_aio(bio, error);
321 else
322 dio_bio_end_io(bio, error);
323}
324EXPORT_SYMBOL_GPL(dio_end_io);
325
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326static int
327dio_bio_alloc(struct dio *dio, struct block_device *bdev,
328 sector_t first_sector, int nr_vecs)
329{
330 struct bio *bio;
331
332 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333
334 bio->bi_bdev = bdev;
335 bio->bi_sector = first_sector;
336 if (dio->is_async)
337 bio->bi_end_io = dio_bio_end_aio;
338 else
339 bio->bi_end_io = dio_bio_end_io;
340
341 dio->bio = bio;
Josef Bacikfacd07b2010-05-23 11:00:55 -0400342 dio->logical_offset_in_bio = dio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343 return 0;
344}
345
346/*
347 * In the AIO read case we speculatively dirty the pages before starting IO.
348 * During IO completion, any of these pages which happen to have been written
349 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800350 *
351 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352 */
353static void dio_bio_submit(struct dio *dio)
354{
355 struct bio *bio = dio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800356 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357
358 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800359
360 spin_lock_irqsave(&dio->bio_lock, flags);
361 dio->refcount++;
362 spin_unlock_irqrestore(&dio->bio_lock, flags);
363
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 if (dio->is_async && dio->rw == READ)
365 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800366
Josef Bacikfacd07b2010-05-23 11:00:55 -0400367 if (dio->submit_io)
368 dio->submit_io(dio->rw, bio, dio->inode,
369 dio->logical_offset_in_bio);
370 else
371 submit_bio(dio->rw, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372
373 dio->bio = NULL;
374 dio->boundary = 0;
Josef Bacikfacd07b2010-05-23 11:00:55 -0400375 dio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376}
377
378/*
379 * Release any resources in case of a failure
380 */
381static void dio_cleanup(struct dio *dio)
382{
383 while (dio_pages_present(dio))
384 page_cache_release(dio_get_page(dio));
385}
386
387/*
Zach Brown02732012006-12-10 02:20:59 -0800388 * Wait for the next BIO to complete. Remove it and return it. NULL is
389 * returned once all BIOs have been completed. This must only be called once
390 * all bios have been issued so that dio->refcount can only decrease. This
391 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 */
393static struct bio *dio_await_one(struct dio *dio)
394{
395 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800396 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397
398 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800399
400 /*
401 * Wait as long as the list is empty and there are bios in flight. bio
402 * completion drops the count, maybe adds to the list, and wakes while
403 * holding the bio_lock so we don't need set_current_state()'s barrier
404 * and can call it after testing our condition.
405 */
406 while (dio->refcount > 1 && dio->bio_list == NULL) {
407 __set_current_state(TASK_UNINTERRUPTIBLE);
408 dio->waiter = current;
409 spin_unlock_irqrestore(&dio->bio_lock, flags);
410 io_schedule();
411 /* wake up sets us TASK_RUNNING */
412 spin_lock_irqsave(&dio->bio_lock, flags);
413 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 }
Zach Brown02732012006-12-10 02:20:59 -0800415 if (dio->bio_list) {
416 bio = dio->bio_list;
417 dio->bio_list = bio->bi_private;
418 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 spin_unlock_irqrestore(&dio->bio_lock, flags);
420 return bio;
421}
422
423/*
424 * Process one completed BIO. No locks are held.
425 */
426static int dio_bio_complete(struct dio *dio, struct bio *bio)
427{
428 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
429 struct bio_vec *bvec = bio->bi_io_vec;
430 int page_no;
431
432 if (!uptodate)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800433 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434
435 if (dio->is_async && dio->rw == READ) {
436 bio_check_pages_dirty(bio); /* transfers ownership */
437 } else {
438 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
439 struct page *page = bvec[page_no].bv_page;
440
441 if (dio->rw == READ && !PageCompound(page))
442 set_page_dirty_lock(page);
443 page_cache_release(page);
444 }
445 bio_put(bio);
446 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 return uptodate ? 0 : -EIO;
448}
449
450/*
Zach Brown02732012006-12-10 02:20:59 -0800451 * Wait on and process all in-flight BIOs. This must only be called once
452 * all bios have been issued so that the refcount can only decrease.
453 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200454 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800455 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800457static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458{
Zach Brown02732012006-12-10 02:20:59 -0800459 struct bio *bio;
460 do {
461 bio = dio_await_one(dio);
462 if (bio)
463 dio_bio_complete(dio, bio);
464 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465}
466
467/*
468 * A really large O_DIRECT read or write can generate a lot of BIOs. So
469 * to keep the memory consumption sane we periodically reap any completed BIOs
470 * during the BIO generation phase.
471 *
472 * This also helps to limit the peak amount of pinned userspace memory.
473 */
474static int dio_bio_reap(struct dio *dio)
475{
476 int ret = 0;
477
478 if (dio->reap_counter++ >= 64) {
479 while (dio->bio_list) {
480 unsigned long flags;
481 struct bio *bio;
482 int ret2;
483
484 spin_lock_irqsave(&dio->bio_lock, flags);
485 bio = dio->bio_list;
486 dio->bio_list = bio->bi_private;
487 spin_unlock_irqrestore(&dio->bio_lock, flags);
488 ret2 = dio_bio_complete(dio, bio);
489 if (ret == 0)
490 ret = ret2;
491 }
492 dio->reap_counter = 0;
493 }
494 return ret;
495}
496
497/*
498 * Call into the fs to map some more disk blocks. We record the current number
499 * of available blocks at dio->blocks_available. These are in units of the
500 * fs blocksize, (1 << inode->i_blkbits).
501 *
502 * The fs is allowed to map lots of blocks at once. If it wants to do that,
503 * it uses the passed inode-relative block number as the file offset, as usual.
504 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800505 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 * has remaining to do. The fs should not map more than this number of blocks.
507 *
508 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
509 * indicate how much contiguous disk space has been made available at
510 * bh->b_blocknr.
511 *
512 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
513 * This isn't very efficient...
514 *
515 * In the case of filesystem holes: the fs may return an arbitrarily-large
516 * hole by returning an appropriate value in b_size and by clearing
517 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800518 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 */
520static int get_more_blocks(struct dio *dio)
521{
522 int ret;
523 struct buffer_head *map_bh = &dio->map_bh;
524 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
525 unsigned long fs_count; /* Number of filesystem-sized blocks */
526 unsigned long dio_count;/* Number of dio_block-sized blocks */
527 unsigned long blkmask;
528 int create;
529
530 /*
531 * If there was a memory error and we've overwritten all the
532 * mapped blocks then we can now return that memory error
533 */
534 ret = dio->page_errors;
535 if (ret == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 BUG_ON(dio->block_in_file >= dio->final_block_in_request);
537 fs_startblk = dio->block_in_file >> dio->blkfactor;
538 dio_count = dio->final_block_in_request - dio->block_in_file;
539 fs_count = dio_count >> dio->blkfactor;
540 blkmask = (1 << dio->blkfactor) - 1;
541 if (dio_count & blkmask)
542 fs_count++;
543
Nathan Scott3c674e72006-03-29 09:26:15 +1000544 map_bh->b_state = 0;
545 map_bh->b_size = fs_count << dio->inode->i_blkbits;
546
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800547 /*
548 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
549 * forbid block creations: only overwrites are permitted.
550 * We will return early to the caller once we see an
551 * unmapped buffer head returned, and the caller will fall
552 * back to buffered I/O.
553 *
554 * Otherwise the decision is left to the get_blocks method,
555 * which may decide to handle it or also return an unmapped
556 * buffer head.
557 */
Jens Axboeb31dc662006-06-13 08:26:10 +0200558 create = dio->rw & WRITE;
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800559 if (dio->flags & DIO_SKIP_HOLES) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560 if (dio->block_in_file < (i_size_read(dio->inode) >>
561 dio->blkbits))
562 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000564
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800565 ret = (*dio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566 map_bh, create);
567 }
568 return ret;
569}
570
571/*
572 * There is no bio. Make one now.
573 */
574static int dio_new_bio(struct dio *dio, sector_t start_sector)
575{
576 sector_t sector;
577 int ret, nr_pages;
578
579 ret = dio_bio_reap(dio);
580 if (ret)
581 goto out;
582 sector = start_sector << (dio->blkbits - 9);
583 nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
584 BUG_ON(nr_pages <= 0);
585 ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
586 dio->boundary = 0;
587out:
588 return ret;
589}
590
591/*
592 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
593 * that was successful then update final_block_in_bio and take a ref against
594 * the just-added page.
595 *
596 * Return zero on success. Non-zero means the caller needs to start a new BIO.
597 */
598static int dio_bio_add_page(struct dio *dio)
599{
600 int ret;
601
602 ret = bio_add_page(dio->bio, dio->cur_page,
603 dio->cur_page_len, dio->cur_page_offset);
604 if (ret == dio->cur_page_len) {
605 /*
606 * Decrement count only, if we are done with this page
607 */
608 if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE)
609 dio->pages_in_io--;
610 page_cache_get(dio->cur_page);
611 dio->final_block_in_bio = dio->cur_page_block +
612 (dio->cur_page_len >> dio->blkbits);
613 ret = 0;
614 } else {
615 ret = 1;
616 }
617 return ret;
618}
619
620/*
621 * Put cur_page under IO. The section of cur_page which is described by
622 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
623 * starts on-disk at cur_page_block.
624 *
625 * We take a ref against the page here (on behalf of its presence in the bio).
626 *
627 * The caller of this function is responsible for removing cur_page from the
628 * dio, and for dropping the refcount which came from that presence.
629 */
630static int dio_send_cur_page(struct dio *dio)
631{
632 int ret = 0;
633
634 if (dio->bio) {
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400635 loff_t cur_offset = dio->block_in_file << dio->blkbits;
636 loff_t bio_next_offset = dio->logical_offset_in_bio +
637 dio->bio->bi_size;
638
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400640 * See whether this new request is contiguous with the old.
641 *
642 * Btrfs cannot handl having logically non-contiguous requests
643 * submitted. For exmple if you have
644 *
645 * Logical: [0-4095][HOLE][8192-12287]
646 * Phyiscal: [0-4095] [4096-8181]
647 *
648 * We cannot submit those pages together as one BIO. So if our
649 * current logical offset in the file does not equal what would
650 * be the next logical offset in the bio, submit the bio we
651 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 */
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400653 if (dio->final_block_in_bio != dio->cur_page_block ||
654 cur_offset != bio_next_offset)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655 dio_bio_submit(dio);
656 /*
657 * Submit now if the underlying fs is about to perform a
658 * metadata read
659 */
660 if (dio->boundary)
661 dio_bio_submit(dio);
662 }
663
664 if (dio->bio == NULL) {
665 ret = dio_new_bio(dio, dio->cur_page_block);
666 if (ret)
667 goto out;
668 }
669
670 if (dio_bio_add_page(dio) != 0) {
671 dio_bio_submit(dio);
672 ret = dio_new_bio(dio, dio->cur_page_block);
673 if (ret == 0) {
674 ret = dio_bio_add_page(dio);
675 BUG_ON(ret != 0);
676 }
677 }
678out:
679 return ret;
680}
681
682/*
683 * An autonomous function to put a chunk of a page under deferred IO.
684 *
685 * The caller doesn't actually know (or care) whether this piece of page is in
686 * a BIO, or is under IO or whatever. We just take care of all possible
687 * situations here. The separation between the logic of do_direct_IO() and
688 * that of submit_page_section() is important for clarity. Please don't break.
689 *
690 * The chunk of page starts on-disk at blocknr.
691 *
692 * We perform deferred IO, by recording the last-submitted page inside our
693 * private part of the dio structure. If possible, we just expand the IO
694 * across that page here.
695 *
696 * If that doesn't work out then we put the old page into the bio and add this
697 * page to the dio instead.
698 */
699static int
700submit_page_section(struct dio *dio, struct page *page,
701 unsigned offset, unsigned len, sector_t blocknr)
702{
703 int ret = 0;
704
Andrew Morton98c4d572006-12-10 02:19:47 -0800705 if (dio->rw & WRITE) {
706 /*
707 * Read accounting is performed in submit_bio()
708 */
709 task_io_account_write(len);
710 }
711
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 /*
713 * Can we just grow the current page's presence in the dio?
714 */
715 if ( (dio->cur_page == page) &&
716 (dio->cur_page_offset + dio->cur_page_len == offset) &&
717 (dio->cur_page_block +
718 (dio->cur_page_len >> dio->blkbits) == blocknr)) {
719 dio->cur_page_len += len;
720
721 /*
722 * If dio->boundary then we want to schedule the IO now to
723 * avoid metadata seeks.
724 */
725 if (dio->boundary) {
726 ret = dio_send_cur_page(dio);
727 page_cache_release(dio->cur_page);
728 dio->cur_page = NULL;
729 }
730 goto out;
731 }
732
733 /*
734 * If there's a deferred page already there then send it.
735 */
736 if (dio->cur_page) {
737 ret = dio_send_cur_page(dio);
738 page_cache_release(dio->cur_page);
739 dio->cur_page = NULL;
740 if (ret)
741 goto out;
742 }
743
744 page_cache_get(page); /* It is in dio */
745 dio->cur_page = page;
746 dio->cur_page_offset = offset;
747 dio->cur_page_len = len;
748 dio->cur_page_block = blocknr;
Josef Bacikfacd07b2010-05-23 11:00:55 -0400749 dio->cur_page_fs_offset = dio->block_in_file << dio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750out:
751 return ret;
752}
753
754/*
755 * Clean any dirty buffers in the blockdev mapping which alias newly-created
756 * file blocks. Only called for S_ISREG files - blockdevs do not set
757 * buffer_new
758 */
759static void clean_blockdev_aliases(struct dio *dio)
760{
761 unsigned i;
762 unsigned nblocks;
763
764 nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
765
766 for (i = 0; i < nblocks; i++) {
767 unmap_underlying_metadata(dio->map_bh.b_bdev,
768 dio->map_bh.b_blocknr + i);
769 }
770}
771
772/*
773 * If we are not writing the entire block and get_block() allocated
774 * the block for us, we need to fill-in the unused portion of the
775 * block with zeros. This happens only if user-buffer, fileoffset or
776 * io length is not filesystem block-size multiple.
777 *
778 * `end' is zero if we're doing the start of the IO, 1 at the end of the
779 * IO.
780 */
781static void dio_zero_block(struct dio *dio, int end)
782{
783 unsigned dio_blocks_per_fs_block;
784 unsigned this_chunk_blocks; /* In dio_blocks */
785 unsigned this_chunk_bytes;
786 struct page *page;
787
788 dio->start_zero_done = 1;
789 if (!dio->blkfactor || !buffer_new(&dio->map_bh))
790 return;
791
792 dio_blocks_per_fs_block = 1 << dio->blkfactor;
793 this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
794
795 if (!this_chunk_blocks)
796 return;
797
798 /*
799 * We need to zero out part of an fs block. It is either at the
800 * beginning or the end of the fs block.
801 */
802 if (end)
803 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
804
805 this_chunk_bytes = this_chunk_blocks << dio->blkbits;
806
Nick Piggin557ed1f2007-10-16 01:24:40 -0700807 page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 if (submit_page_section(dio, page, 0, this_chunk_bytes,
809 dio->next_block_for_io))
810 return;
811
812 dio->next_block_for_io += this_chunk_blocks;
813}
814
815/*
816 * Walk the user pages, and the file, mapping blocks to disk and generating
817 * a sequence of (page,offset,len,block) mappings. These mappings are injected
818 * into submit_page_section(), which takes care of the next stage of submission
819 *
820 * Direct IO against a blockdev is different from a file. Because we can
821 * happily perform page-sized but 512-byte aligned IOs. It is important that
822 * blockdev IO be able to have fine alignment and large sizes.
823 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800824 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 * with the size of IO which is permitted at this offset and this i_blkbits.
826 *
827 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800828 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829 * fine alignment but still allows this function to work in PAGE_SIZE units.
830 */
831static int do_direct_IO(struct dio *dio)
832{
833 const unsigned blkbits = dio->blkbits;
834 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
835 struct page *page;
836 unsigned block_in_page;
837 struct buffer_head *map_bh = &dio->map_bh;
838 int ret = 0;
839
840 /* The I/O can start at any block offset within the first page */
841 block_in_page = dio->first_block_in_page;
842
843 while (dio->block_in_file < dio->final_block_in_request) {
844 page = dio_get_page(dio);
845 if (IS_ERR(page)) {
846 ret = PTR_ERR(page);
847 goto out;
848 }
849
850 while (block_in_page < blocks_per_page) {
851 unsigned offset_in_page = block_in_page << blkbits;
852 unsigned this_chunk_bytes; /* # of bytes mapped */
853 unsigned this_chunk_blocks; /* # of blocks */
854 unsigned u;
855
856 if (dio->blocks_available == 0) {
857 /*
858 * Need to go and map some more disk
859 */
860 unsigned long blkmask;
861 unsigned long dio_remainder;
862
863 ret = get_more_blocks(dio);
864 if (ret) {
865 page_cache_release(page);
866 goto out;
867 }
868 if (!buffer_mapped(map_bh))
869 goto do_holes;
870
871 dio->blocks_available =
872 map_bh->b_size >> dio->blkbits;
873 dio->next_block_for_io =
874 map_bh->b_blocknr << dio->blkfactor;
875 if (buffer_new(map_bh))
876 clean_blockdev_aliases(dio);
877
878 if (!dio->blkfactor)
879 goto do_holes;
880
881 blkmask = (1 << dio->blkfactor) - 1;
882 dio_remainder = (dio->block_in_file & blkmask);
883
884 /*
885 * If we are at the start of IO and that IO
886 * starts partway into a fs-block,
887 * dio_remainder will be non-zero. If the IO
888 * is a read then we can simply advance the IO
889 * cursor to the first block which is to be
890 * read. But if the IO is a write and the
891 * block was newly allocated we cannot do that;
892 * the start of the fs block must be zeroed out
893 * on-disk
894 */
895 if (!buffer_new(map_bh))
896 dio->next_block_for_io += dio_remainder;
897 dio->blocks_available -= dio_remainder;
898 }
899do_holes:
900 /* Handle holes */
901 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800902 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903
904 /* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboeb31dc662006-06-13 08:26:10 +0200905 if (dio->rw & WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 page_cache_release(page);
907 return -ENOTBLK;
908 }
909
Jeff Moyer35dc8162006-02-03 03:04:27 -0800910 /*
911 * Be sure to account for a partial block as the
912 * last block in the file
913 */
914 i_size_aligned = ALIGN(i_size_read(dio->inode),
915 1 << blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 if (dio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -0800917 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 /* We hit eof */
919 page_cache_release(page);
920 goto out;
921 }
Christoph Lametereebd2aa2008-02-04 22:28:29 -0800922 zero_user(page, block_in_page << blkbits,
923 1 << blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924 dio->block_in_file++;
925 block_in_page++;
926 goto next_block;
927 }
928
929 /*
930 * If we're performing IO which has an alignment which
931 * is finer than the underlying fs, go check to see if
932 * we must zero out the start of this block.
933 */
934 if (unlikely(dio->blkfactor && !dio->start_zero_done))
935 dio_zero_block(dio, 0);
936
937 /*
938 * Work out, in this_chunk_blocks, how much disk we
939 * can add to this page
940 */
941 this_chunk_blocks = dio->blocks_available;
942 u = (PAGE_SIZE - offset_in_page) >> blkbits;
943 if (this_chunk_blocks > u)
944 this_chunk_blocks = u;
945 u = dio->final_block_in_request - dio->block_in_file;
946 if (this_chunk_blocks > u)
947 this_chunk_blocks = u;
948 this_chunk_bytes = this_chunk_blocks << blkbits;
949 BUG_ON(this_chunk_bytes == 0);
950
951 dio->boundary = buffer_boundary(map_bh);
952 ret = submit_page_section(dio, page, offset_in_page,
953 this_chunk_bytes, dio->next_block_for_io);
954 if (ret) {
955 page_cache_release(page);
956 goto out;
957 }
958 dio->next_block_for_io += this_chunk_blocks;
959
960 dio->block_in_file += this_chunk_blocks;
961 block_in_page += this_chunk_blocks;
962 dio->blocks_available -= this_chunk_blocks;
963next_block:
Eric Sesterhennd4569d22006-04-01 01:10:13 +0200964 BUG_ON(dio->block_in_file > dio->final_block_in_request);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 if (dio->block_in_file == dio->final_block_in_request)
966 break;
967 }
968
969 /* Drop the ref which was taken in get_user_pages() */
970 page_cache_release(page);
971 block_in_page = 0;
972 }
973out:
974 return ret;
975}
976
977/*
Jes Sorensen1b1dcc12006-01-09 15:59:24 -0800978 * Releases both i_mutex and i_alloc_sem
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979 */
980static ssize_t
981direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
982 const struct iovec *iov, loff_t offset, unsigned long nr_segs,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800983 unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
Josef Bacikfacd07b2010-05-23 11:00:55 -0400984 dio_submit_t submit_io, struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985{
986 unsigned long user_addr;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800987 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988 int seg;
989 ssize_t ret = 0;
990 ssize_t ret2;
991 size_t bytes;
992
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 dio->inode = inode;
994 dio->rw = rw;
995 dio->blkbits = blkbits;
996 dio->blkfactor = inode->i_blkbits - blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997 dio->block_in_file = offset >> blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800999 dio->get_block = get_block;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 dio->end_io = end_io;
Josef Bacikfacd07b2010-05-23 11:00:55 -04001001 dio->submit_io = submit_io;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002 dio->final_block_in_bio = -1;
1003 dio->next_block_for_io = -1;
1004
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005 dio->iocb = iocb;
Daniel McNeil29504ff2005-04-16 15:25:50 -07001006 dio->i_size = i_size_read(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008 spin_lock_init(&dio->bio_lock);
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001009 dio->refcount = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010
1011 /*
1012 * In case of non-aligned buffers, we may need 2 more
1013 * pages since we need to zero out first and last block.
1014 */
1015 if (unlikely(dio->blkfactor))
1016 dio->pages_in_io = 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017
1018 for (seg = 0; seg < nr_segs; seg++) {
1019 user_addr = (unsigned long)iov[seg].iov_base;
1020 dio->pages_in_io +=
1021 ((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
1022 - user_addr/PAGE_SIZE);
1023 }
1024
1025 for (seg = 0; seg < nr_segs; seg++) {
1026 user_addr = (unsigned long)iov[seg].iov_base;
1027 dio->size += bytes = iov[seg].iov_len;
1028
1029 /* Index into the first page of the first block */
1030 dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1031 dio->final_block_in_request = dio->block_in_file +
1032 (bytes >> blkbits);
1033 /* Page fetching state */
1034 dio->head = 0;
1035 dio->tail = 0;
1036 dio->curr_page = 0;
1037
1038 dio->total_pages = 0;
1039 if (user_addr & (PAGE_SIZE-1)) {
1040 dio->total_pages++;
1041 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1042 }
1043 dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1044 dio->curr_user_address = user_addr;
1045
1046 ret = do_direct_IO(dio);
1047
1048 dio->result += iov[seg].iov_len -
1049 ((dio->final_block_in_request - dio->block_in_file) <<
1050 blkbits);
1051
1052 if (ret) {
1053 dio_cleanup(dio);
1054 break;
1055 }
1056 } /* end iovec loop */
1057
Josef Bacikfacd07b2010-05-23 11:00:55 -04001058 if (ret == -ENOTBLK) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059 /*
1060 * The remaining part of the request will be
1061 * be handled by buffered I/O when we return
1062 */
1063 ret = 0;
1064 }
1065 /*
1066 * There may be some unwritten disk at the end of a part-written
1067 * fs-block-sized block. Go zero that now.
1068 */
1069 dio_zero_block(dio, 1);
1070
1071 if (dio->cur_page) {
1072 ret2 = dio_send_cur_page(dio);
1073 if (ret == 0)
1074 ret = ret2;
1075 page_cache_release(dio->cur_page);
1076 dio->cur_page = NULL;
1077 }
1078 if (dio->bio)
1079 dio_bio_submit(dio);
1080
1081 /*
1082 * It is possible that, we return short IO due to end of file.
1083 * In that case, we need to release all the pages we got hold on.
1084 */
1085 dio_cleanup(dio);
1086
1087 /*
1088 * All block lookups have been performed. For READ requests
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001089 * we can let i_mutex go now that its achieved its purpose
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 * of protecting us from looking up uninitialized blocks.
1091 */
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001092 if (rw == READ && (dio->flags & DIO_LOCKING))
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001093 mutex_unlock(&dio->inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094
1095 /*
Zach Brown8459d862006-12-10 02:21:05 -08001096 * The only time we want to leave bios in flight is when a successful
1097 * partial aio read or full aio write have been setup. In that case
1098 * bio completion will call aio_complete. The only time it's safe to
1099 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1100 * This had *better* be the only place that raises -EIOCBQUEUED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 */
Zach Brown8459d862006-12-10 02:21:05 -08001102 BUG_ON(ret == -EIOCBQUEUED);
1103 if (dio->is_async && ret == 0 && dio->result &&
1104 ((rw & READ) || (dio->result == dio->size)))
1105 ret = -EIOCBQUEUED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106
Jeff Moyercfb1e332009-10-02 18:57:36 -04001107 if (ret != -EIOCBQUEUED) {
1108 /* All IO is now issued, send it on its way */
1109 blk_run_address_space(inode->i_mapping);
Zach Brown6d544bb2006-12-10 02:20:54 -08001110 dio_await_completion(dio);
Jeff Moyercfb1e332009-10-02 18:57:36 -04001111 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112
Zach Brown8459d862006-12-10 02:21:05 -08001113 /*
1114 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001115 * operation. AIO can if it was a broken operation described above or
1116 * in fact if all the bios race to complete before we get here. In
1117 * that case dio_complete() translates the EIOCBQUEUED into the proper
1118 * return code that the caller will hand to aio_complete().
1119 *
1120 * This is managed by the bio_lock instead of being an atomic_t so that
1121 * completion paths can drop their ref and use the remaining count to
1122 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001123 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001124 spin_lock_irqsave(&dio->bio_lock, flags);
1125 ret2 = --dio->refcount;
1126 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brownfcb82f82007-07-03 15:28:55 -07001127
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001128 if (ret2 == 0) {
Zach Brown6d544bb2006-12-10 02:20:54 -08001129 ret = dio_complete(dio, offset, ret);
Zach Brown8459d862006-12-10 02:21:05 -08001130 kfree(dio);
1131 } else
1132 BUG_ON(ret != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134 return ret;
1135}
1136
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137ssize_t
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001138__blockdev_direct_IO_newtrunc(int rw, struct kiocb *iocb, struct inode *inode,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 struct block_device *bdev, const struct iovec *iov, loff_t offset,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -08001140 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
Josef Bacikfacd07b2010-05-23 11:00:55 -04001141 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142{
1143 int seg;
1144 size_t size;
1145 unsigned long addr;
1146 unsigned blkbits = inode->i_blkbits;
1147 unsigned bdev_blkbits = 0;
1148 unsigned blocksize_mask = (1 << blkbits) - 1;
1149 ssize_t retval = -EINVAL;
1150 loff_t end = offset;
1151 struct dio *dio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152
1153 if (rw & WRITE)
Vivek Goyald9449ce2009-11-26 09:45:40 +01001154 rw = WRITE_ODIRECT_PLUG;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155
1156 if (bdev)
Martin K. Petersene1defc42009-05-22 17:17:49 -04001157 bdev_blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158
1159 if (offset & blocksize_mask) {
1160 if (bdev)
1161 blkbits = bdev_blkbits;
1162 blocksize_mask = (1 << blkbits) - 1;
1163 if (offset & blocksize_mask)
1164 goto out;
1165 }
1166
1167 /* Check the memory alignment. Blocks cannot straddle pages */
1168 for (seg = 0; seg < nr_segs; seg++) {
1169 addr = (unsigned long)iov[seg].iov_base;
1170 size = iov[seg].iov_len;
1171 end += size;
1172 if ((addr & blocksize_mask) || (size & blocksize_mask)) {
1173 if (bdev)
1174 blkbits = bdev_blkbits;
1175 blocksize_mask = (1 << blkbits) - 1;
1176 if ((addr & blocksize_mask) || (size & blocksize_mask))
1177 goto out;
1178 }
1179 }
1180
Jeff Moyer23aee092009-12-15 16:47:49 -08001181 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 retval = -ENOMEM;
1183 if (!dio)
1184 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001185 /*
1186 * Believe it or not, zeroing out the page array caused a .5%
1187 * performance regression in a database benchmark. So, we take
1188 * care to only zero out what's needed.
1189 */
1190 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001192 dio->flags = flags;
1193 if (dio->flags & DIO_LOCKING) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 /* watch out for a 0 len io from a tricksy fs */
1195 if (rw == READ && end > offset) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001196 struct address_space *mapping =
1197 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001199 /* will be released by direct_io_worker */
1200 mutex_lock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201
1202 retval = filemap_write_and_wait_range(mapping, offset,
1203 end - 1);
1204 if (retval) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001205 mutex_unlock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206 kfree(dio);
1207 goto out;
1208 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209 }
1210
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001211 /*
1212 * Will be released at I/O completion, possibly in a
1213 * different thread.
1214 */
1215 down_read_non_owner(&inode->i_alloc_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216 }
1217
1218 /*
1219 * For file extending writes updating i_size before data
1220 * writeouts complete can expose uninitialized blocks. So
1221 * even for AIO, we need to wait for i/o to complete before
1222 * returning in this case.
1223 */
Jens Axboeb31dc662006-06-13 08:26:10 +02001224 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225 (end > i_size_read(inode)));
1226
1227 retval = direct_io_worker(rw, iocb, inode, iov, offset,
Josef Bacikfacd07b2010-05-23 11:00:55 -04001228 nr_segs, blkbits, get_block, end_io,
1229 submit_io, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001231out:
1232 return retval;
1233}
1234EXPORT_SYMBOL(__blockdev_direct_IO_newtrunc);
1235
1236/*
1237 * This is a library function for use by filesystem drivers.
1238 *
1239 * The locking rules are governed by the flags parameter:
1240 * - if the flags value contains DIO_LOCKING we use a fancy locking
1241 * scheme for dumb filesystems.
1242 * For writes this function is called under i_mutex and returns with
1243 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1244 * taken and dropped again before returning.
1245 * For reads and writes i_alloc_sem is taken in shared mode and released
1246 * on I/O completion (which may happen asynchronously after returning to
1247 * the caller).
1248 *
1249 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1250 * internal locking but rather rely on the filesystem to synchronize
1251 * direct I/O reads/writes versus each other and truncate.
1252 * For reads and writes both i_mutex and i_alloc_sem are not held on
1253 * entry and are never taken.
1254 */
1255ssize_t
1256__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1257 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1258 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1259 dio_submit_t submit_io, int flags)
1260{
1261 ssize_t retval;
1262
1263 retval = __blockdev_direct_IO_newtrunc(rw, iocb, inode, bdev, iov,
1264 offset, nr_segs, get_block, end_io, submit_io, flags);
Dmitri Monakhov0f644152009-01-06 14:40:04 -08001265 /*
1266 * In case of error extending write may have instantiated a few
1267 * blocks outside i_size. Trim these off again for DIO_LOCKING.
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001268 * NOTE: DIO_NO_LOCK/DIO_OWN_LOCK callers have to handle this in
1269 * their own manner. This is a further example of where the old
1270 * truncate sequence is inadequate.
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001271 *
1272 * NOTE: filesystems with their own locking have to handle this
1273 * on their own.
Dmitri Monakhov0f644152009-01-06 14:40:04 -08001274 */
Al Viro06777d32009-12-17 04:52:13 -05001275 if (flags & DIO_LOCKING) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001276 if (unlikely((rw & WRITE) && retval < 0)) {
1277 loff_t isize = i_size_read(inode);
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001278 loff_t end = offset + iov_length(iov, nr_segs);
1279
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001280 if (end > isize)
1281 vmtruncate(inode, isize);
1282 }
Dmitri Monakhov0f644152009-01-06 14:40:04 -08001283 }
1284
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 return retval;
1286}
1287EXPORT_SYMBOL(__blockdev_direct_IO);