blob: be212458d1431fe7fb4f01262aca0bda0ed04f2d [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>
Arun Sharma60063492011-07-26 16:09:06 -070038#include <linux/atomic.h>
Andi Kleen65dd2aa2012-01-12 17:20:35 -080039#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070040
41/*
42 * How many user pages to map in one call to get_user_pages(). This determines
Andi Kleencde1ecb2011-08-01 21:38:04 -070043 * the size of a structure in the slab cache
Linus Torvalds1da177e2005-04-16 15:20:36 -070044 */
45#define DIO_PAGES 64
46
47/*
48 * This code generally works in units of "dio_blocks". A dio_block is
49 * somewhere between the hard sector size and the filesystem block size. it
50 * is determined on a per-invocation basis. When talking to the filesystem
51 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
52 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
53 * to bio_block quantities by shifting left by blkfactor.
54 *
55 * If blkfactor is zero then the user's request was aligned to the filesystem's
56 * blocksize.
Linus Torvalds1da177e2005-04-16 15:20:36 -070057 */
58
Andi Kleeneb28be22011-08-01 21:38:03 -070059/* dio_state only used in the submission path */
60
61struct dio_submit {
Linus Torvalds1da177e2005-04-16 15:20:36 -070062 struct bio *bio; /* bio under assembly */
Linus Torvalds1da177e2005-04-16 15:20:36 -070063 unsigned blkbits; /* doesn't change */
64 unsigned blkfactor; /* When we're using an alignment which
65 is finer than the filesystem's soft
66 blocksize, this specifies how much
67 finer. blkfactor=2 means 1/4-block
68 alignment. Does not change */
69 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
70 been performed at the start of a
71 write */
72 int pages_in_io; /* approximate total IO pages */
73 size_t size; /* total request size (doesn't change)*/
74 sector_t block_in_file; /* Current offset into the underlying
75 file in dio_block units. */
76 unsigned blocks_available; /* At block_in_file. changes */
Andi Kleen0dc2bc42011-08-01 21:38:05 -070077 int reap_counter; /* rate limit reaping */
Linus Torvalds1da177e2005-04-16 15:20:36 -070078 sector_t final_block_in_request;/* doesn't change */
79 unsigned first_block_in_page; /* doesn't change, Used only once */
80 int boundary; /* prev block is at a boundary */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080081 get_block_t *get_block; /* block mapping function */
Josef Bacikfacd07b2010-05-23 11:00:55 -040082 dio_submit_t *submit_io; /* IO submition function */
Andi Kleeneb28be22011-08-01 21:38:03 -070083
Josef Bacikfacd07b2010-05-23 11:00:55 -040084 loff_t logical_offset_in_bio; /* current first logical block in bio */
Linus Torvalds1da177e2005-04-16 15:20:36 -070085 sector_t final_block_in_bio; /* current final block in bio + 1 */
86 sector_t next_block_for_io; /* next block to be put under IO,
87 in dio_blocks units */
Linus Torvalds1da177e2005-04-16 15:20:36 -070088
89 /*
90 * Deferred addition of a page to the dio. These variables are
91 * private to dio_send_cur_page(), submit_page_section() and
92 * dio_bio_add_page().
93 */
94 struct page *cur_page; /* The page */
95 unsigned cur_page_offset; /* Offset into it, in bytes */
96 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
97 sector_t cur_page_block; /* Where it starts */
Josef Bacikfacd07b2010-05-23 11:00:55 -040098 loff_t cur_page_fs_offset; /* Offset in file */
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
Jeff Moyer23aee092009-12-15 16:47:49 -0800100 /*
101 * Page fetching state. These variables belong to dio_refill_pages().
102 */
103 int curr_page; /* changes */
104 int total_pages; /* doesn't change */
105 unsigned long curr_user_address;/* changes */
106
107 /*
108 * Page queue. These variables belong to dio_refill_pages() and
109 * dio_get_page().
110 */
111 unsigned head; /* next page to process */
112 unsigned tail; /* last valid page + 1 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700113};
114
115/* dio_state communicated between submission path and end_io */
116struct dio {
117 int flags; /* doesn't change */
Andi Kleeneb28be22011-08-01 21:38:03 -0700118 int rw;
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700119 struct inode *inode;
Andi Kleeneb28be22011-08-01 21:38:03 -0700120 loff_t i_size; /* i_size when submitted */
121 dio_iodone_t *end_io; /* IO completion function */
Andi Kleeneb28be22011-08-01 21:38:03 -0700122
Andi Kleen18772642011-08-01 21:38:07 -0700123 void *private; /* copy from map_bh.b_private */
Andi Kleeneb28be22011-08-01 21:38:03 -0700124
125 /* BIO completion state */
126 spinlock_t bio_lock; /* protects BIO fields below */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700127 int page_errors; /* errno from get_user_pages() */
128 int is_async; /* is IO async ? */
129 int io_error; /* IO error in completion path */
Andi Kleeneb28be22011-08-01 21:38:03 -0700130 unsigned long refcount; /* direct_io_worker() and bios */
131 struct bio *bio_list; /* singly linked via bi_private */
132 struct task_struct *waiter; /* waiting task (NULL if none) */
133
134 /* AIO related stuff */
135 struct kiocb *iocb; /* kiocb */
Andi Kleeneb28be22011-08-01 21:38:03 -0700136 ssize_t result; /* IO result */
137
Jeff Moyer23aee092009-12-15 16:47:49 -0800138 /*
139 * pages[] (and any fields placed after it) are not zeroed out at
140 * allocation time. Don't add new fields after pages[] unless you
141 * wish that they not be zeroed.
142 */
143 struct page *pages[DIO_PAGES]; /* page buffer */
Andi Kleen6e8267f2011-08-01 21:38:06 -0700144} ____cacheline_aligned_in_smp;
145
146static struct kmem_cache *dio_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147
Christoph Hellwigbd5fe6c2011-06-24 14:29:43 -0400148static void __inode_dio_wait(struct inode *inode)
149{
150 wait_queue_head_t *wq = bit_waitqueue(&inode->i_state, __I_DIO_WAKEUP);
151 DEFINE_WAIT_BIT(q, &inode->i_state, __I_DIO_WAKEUP);
152
153 do {
154 prepare_to_wait(wq, &q.wait, TASK_UNINTERRUPTIBLE);
155 if (atomic_read(&inode->i_dio_count))
156 schedule();
157 } while (atomic_read(&inode->i_dio_count));
158 finish_wait(wq, &q.wait);
159}
160
161/**
162 * inode_dio_wait - wait for outstanding DIO requests to finish
163 * @inode: inode to wait for
164 *
165 * Waits for all pending direct I/O requests to finish so that we can
166 * proceed with a truncate or equivalent operation.
167 *
168 * Must be called under a lock that serializes taking new references
169 * to i_dio_count, usually by inode->i_mutex.
170 */
171void inode_dio_wait(struct inode *inode)
172{
173 if (atomic_read(&inode->i_dio_count))
174 __inode_dio_wait(inode);
175}
Anton Altaparmakov37fbf4b2012-02-23 23:40:05 +0000176EXPORT_SYMBOL(inode_dio_wait);
Christoph Hellwigbd5fe6c2011-06-24 14:29:43 -0400177
178/*
179 * inode_dio_done - signal finish of a direct I/O requests
180 * @inode: inode the direct I/O happens on
181 *
182 * This is called once we've finished processing a direct I/O request,
183 * and is used to wake up callers waiting for direct I/O to be quiesced.
184 */
185void inode_dio_done(struct inode *inode)
186{
187 if (atomic_dec_and_test(&inode->i_dio_count))
188 wake_up_bit(&inode->i_state, __I_DIO_WAKEUP);
189}
Anton Altaparmakov37fbf4b2012-02-23 23:40:05 +0000190EXPORT_SYMBOL(inode_dio_done);
Christoph Hellwigbd5fe6c2011-06-24 14:29:43 -0400191
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192/*
193 * How many pages are in the queue?
194 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700195static inline unsigned dio_pages_present(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196{
Andi Kleeneb28be22011-08-01 21:38:03 -0700197 return sdio->tail - sdio->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198}
199
200/*
201 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
202 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700203static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204{
205 int ret;
206 int nr_pages;
207
Andi Kleeneb28be22011-08-01 21:38:03 -0700208 nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
Nick Pigginf5dd33c2008-07-25 19:45:25 -0700209 ret = get_user_pages_fast(
Andi Kleeneb28be22011-08-01 21:38:03 -0700210 sdio->curr_user_address, /* Where from? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211 nr_pages, /* How many pages? */
212 dio->rw == READ, /* Write to memory? */
Nick Pigginf5dd33c2008-07-25 19:45:25 -0700213 &dio->pages[0]); /* Put results here */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214
Andi Kleeneb28be22011-08-01 21:38:03 -0700215 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
Nick Piggin557ed1f2007-10-16 01:24:40 -0700216 struct page *page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217 /*
218 * A memory fault, but the filesystem has some outstanding
219 * mapped blocks. We need to use those blocks up to avoid
220 * leaking stale data in the file.
221 */
222 if (dio->page_errors == 0)
223 dio->page_errors = ret;
Nick Pigginb5810032005-10-29 18:16:12 -0700224 page_cache_get(page);
225 dio->pages[0] = page;
Andi Kleeneb28be22011-08-01 21:38:03 -0700226 sdio->head = 0;
227 sdio->tail = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228 ret = 0;
229 goto out;
230 }
231
232 if (ret >= 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700233 sdio->curr_user_address += ret * PAGE_SIZE;
234 sdio->curr_page += ret;
235 sdio->head = 0;
236 sdio->tail = ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 ret = 0;
238 }
239out:
240 return ret;
241}
242
243/*
244 * Get another userspace page. Returns an ERR_PTR on error. Pages are
245 * buffered inside the dio so that we can call get_user_pages() against a
246 * decent number of pages, less frequently. To provide nicer use of the
247 * L1 cache.
248 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700249static inline struct page *dio_get_page(struct dio *dio,
250 struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251{
Andi Kleeneb28be22011-08-01 21:38:03 -0700252 if (dio_pages_present(sdio) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253 int ret;
254
Andi Kleeneb28be22011-08-01 21:38:03 -0700255 ret = dio_refill_pages(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256 if (ret)
257 return ERR_PTR(ret);
Andi Kleeneb28be22011-08-01 21:38:03 -0700258 BUG_ON(dio_pages_present(sdio) == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700260 return dio->pages[sdio->head++];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261}
262
Zach Brown6d544bb2006-12-10 02:20:54 -0800263/**
264 * dio_complete() - called when all DIO BIO I/O has been completed
265 * @offset: the byte offset in the file of the completed operation
266 *
267 * This releases locks as dictated by the locking type, lets interested parties
268 * know that a DIO operation has completed, and calculates the resulting return
269 * code for the operation.
270 *
271 * It lets the filesystem know if it registered an interest earlier via
272 * get_block. Pass the private field of the map buffer_head so that
273 * filesystems can use it to hold additional state between get_block calls and
274 * dio_complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275 */
Edward Shishkincd1c5842010-10-26 14:22:28 -0700276static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277{
Zach Brown6d544bb2006-12-10 02:20:54 -0800278 ssize_t transferred = 0;
279
Zach Brown8459d862006-12-10 02:21:05 -0800280 /*
281 * AIO submission can race with bio completion to get here while
282 * expecting to have the last io completed by bio completion.
283 * In that case -EIOCBQUEUED is in fact not an error we want
284 * to preserve through this call.
285 */
286 if (ret == -EIOCBQUEUED)
287 ret = 0;
288
Zach Brown6d544bb2006-12-10 02:20:54 -0800289 if (dio->result) {
290 transferred = dio->result;
291
292 /* Check for short read case */
293 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
294 transferred = dio->i_size - offset;
295 }
296
Zach Brown6d544bb2006-12-10 02:20:54 -0800297 if (ret == 0)
298 ret = dio->page_errors;
299 if (ret == 0)
300 ret = dio->io_error;
301 if (ret == 0)
302 ret = transferred;
303
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000304 if (dio->end_io && dio->result) {
305 dio->end_io(dio->iocb, offset, transferred,
Andi Kleen18772642011-08-01 21:38:07 -0700306 dio->private, ret, is_async);
Christoph Hellwig72c50522011-06-24 14:29:48 -0400307 } else {
308 if (is_async)
309 aio_complete(dio->iocb, ret, 0);
310 inode_dio_done(dio->inode);
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000311 }
312
Zach Brown6d544bb2006-12-10 02:20:54 -0800313 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314}
315
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316static int dio_bio_complete(struct dio *dio, struct bio *bio);
317/*
318 * Asynchronous IO callback.
319 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200320static void dio_bio_end_aio(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321{
322 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800323 unsigned long remaining;
324 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326 /* cleanup the bio */
327 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800328
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800329 spin_lock_irqsave(&dio->bio_lock, flags);
330 remaining = --dio->refcount;
331 if (remaining == 1 && dio->waiter)
Zach Brown20258b22006-12-10 02:21:01 -0800332 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800333 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b22006-12-10 02:21:01 -0800334
Zach Brown8459d862006-12-10 02:21:05 -0800335 if (remaining == 0) {
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000336 dio_complete(dio, dio->iocb->ki_pos, 0, true);
Andi Kleen6e8267f2011-08-01 21:38:06 -0700337 kmem_cache_free(dio_cache, dio);
Zach Brown8459d862006-12-10 02:21:05 -0800338 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339}
340
341/*
342 * The BIO completion handler simply queues the BIO up for the process-context
343 * handler.
344 *
345 * During I/O bi_private points at the dio. After I/O, bi_private is used to
346 * implement a singly-linked list of completed BIOs, at dio->bio_list.
347 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200348static void dio_bio_end_io(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349{
350 struct dio *dio = bio->bi_private;
351 unsigned long flags;
352
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353 spin_lock_irqsave(&dio->bio_lock, flags);
354 bio->bi_private = dio->bio_list;
355 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800356 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357 wake_up_process(dio->waiter);
358 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359}
360
Josef Bacikfacd07b2010-05-23 11:00:55 -0400361/**
362 * dio_end_io - handle the end io action for the given bio
363 * @bio: The direct io bio thats being completed
364 * @error: Error if there was one
365 *
366 * This is meant to be called by any filesystem that uses their own dio_submit_t
367 * so that the DIO specific endio actions are dealt with after the filesystem
368 * has done it's completion work.
369 */
370void dio_end_io(struct bio *bio, int error)
371{
372 struct dio *dio = bio->bi_private;
373
374 if (dio->is_async)
375 dio_bio_end_aio(bio, error);
376 else
377 dio_bio_end_io(bio, error);
378}
379EXPORT_SYMBOL_GPL(dio_end_io);
380
Andi Kleenba253fb2011-08-01 21:38:08 -0700381static inline void
Andi Kleeneb28be22011-08-01 21:38:03 -0700382dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
383 struct block_device *bdev,
384 sector_t first_sector, int nr_vecs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385{
386 struct bio *bio;
387
David Dillow20d96002011-01-20 14:44:22 -0800388 /*
389 * bio_alloc() is guaranteed to return a bio when called with
390 * __GFP_WAIT and we request a valid number of vectors.
391 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393
394 bio->bi_bdev = bdev;
395 bio->bi_sector = first_sector;
396 if (dio->is_async)
397 bio->bi_end_io = dio_bio_end_aio;
398 else
399 bio->bi_end_io = dio_bio_end_io;
400
Andi Kleeneb28be22011-08-01 21:38:03 -0700401 sdio->bio = bio;
402 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403}
404
405/*
406 * In the AIO read case we speculatively dirty the pages before starting IO.
407 * During IO completion, any of these pages which happen to have been written
408 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800409 *
410 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700412static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413{
Andi Kleeneb28be22011-08-01 21:38:03 -0700414 struct bio *bio = sdio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800415 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416
417 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800418
419 spin_lock_irqsave(&dio->bio_lock, flags);
420 dio->refcount++;
421 spin_unlock_irqrestore(&dio->bio_lock, flags);
422
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 if (dio->is_async && dio->rw == READ)
424 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800425
Amir Samuelovacc2a3b2014-06-22 16:43:52 +0300426 bio->bi_dio_inode = dio->inode;
427
Andi Kleeneb28be22011-08-01 21:38:03 -0700428 if (sdio->submit_io)
429 sdio->submit_io(dio->rw, bio, dio->inode,
430 sdio->logical_offset_in_bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400431 else
432 submit_bio(dio->rw, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433
Andi Kleeneb28be22011-08-01 21:38:03 -0700434 sdio->bio = NULL;
435 sdio->boundary = 0;
436 sdio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437}
438
Amir Samuelovacc2a3b2014-06-22 16:43:52 +0300439struct inode *dio_bio_get_inode(struct bio *bio)
440{
441 struct inode *inode = NULL;
442
443 if (bio == NULL)
444 return NULL;
445
446 inode = bio->bi_dio_inode;
447
448 return inode;
449}
450EXPORT_SYMBOL(dio_bio_get_inode);
451
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452/*
453 * Release any resources in case of a failure
454 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700455static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456{
Andi Kleeneb28be22011-08-01 21:38:03 -0700457 while (dio_pages_present(sdio))
458 page_cache_release(dio_get_page(dio, sdio));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459}
460
461/*
Zach Brown02732012006-12-10 02:20:59 -0800462 * Wait for the next BIO to complete. Remove it and return it. NULL is
463 * returned once all BIOs have been completed. This must only be called once
464 * all bios have been issued so that dio->refcount can only decrease. This
465 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 */
467static struct bio *dio_await_one(struct dio *dio)
468{
469 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800470 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471
472 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800473
474 /*
475 * Wait as long as the list is empty and there are bios in flight. bio
476 * completion drops the count, maybe adds to the list, and wakes while
477 * holding the bio_lock so we don't need set_current_state()'s barrier
478 * and can call it after testing our condition.
479 */
480 while (dio->refcount > 1 && dio->bio_list == NULL) {
481 __set_current_state(TASK_UNINTERRUPTIBLE);
482 dio->waiter = current;
483 spin_unlock_irqrestore(&dio->bio_lock, flags);
484 io_schedule();
485 /* wake up sets us TASK_RUNNING */
486 spin_lock_irqsave(&dio->bio_lock, flags);
487 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488 }
Zach Brown02732012006-12-10 02:20:59 -0800489 if (dio->bio_list) {
490 bio = dio->bio_list;
491 dio->bio_list = bio->bi_private;
492 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 spin_unlock_irqrestore(&dio->bio_lock, flags);
494 return bio;
495}
496
497/*
498 * Process one completed BIO. No locks are held.
499 */
500static int dio_bio_complete(struct dio *dio, struct bio *bio)
501{
502 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
503 struct bio_vec *bvec = bio->bi_io_vec;
504 int page_no;
505
506 if (!uptodate)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800507 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508
509 if (dio->is_async && dio->rw == READ) {
510 bio_check_pages_dirty(bio); /* transfers ownership */
511 } else {
512 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
513 struct page *page = bvec[page_no].bv_page;
514
515 if (dio->rw == READ && !PageCompound(page))
516 set_page_dirty_lock(page);
517 page_cache_release(page);
518 }
519 bio_put(bio);
520 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 return uptodate ? 0 : -EIO;
522}
523
524/*
Zach Brown02732012006-12-10 02:20:59 -0800525 * Wait on and process all in-flight BIOs. This must only be called once
526 * all bios have been issued so that the refcount can only decrease.
527 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200528 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800529 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800531static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532{
Zach Brown02732012006-12-10 02:20:59 -0800533 struct bio *bio;
534 do {
535 bio = dio_await_one(dio);
536 if (bio)
537 dio_bio_complete(dio, bio);
538 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539}
540
541/*
542 * A really large O_DIRECT read or write can generate a lot of BIOs. So
543 * to keep the memory consumption sane we periodically reap any completed BIOs
544 * during the BIO generation phase.
545 *
546 * This also helps to limit the peak amount of pinned userspace memory.
547 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700548static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549{
550 int ret = 0;
551
Andi Kleeneb28be22011-08-01 21:38:03 -0700552 if (sdio->reap_counter++ >= 64) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553 while (dio->bio_list) {
554 unsigned long flags;
555 struct bio *bio;
556 int ret2;
557
558 spin_lock_irqsave(&dio->bio_lock, flags);
559 bio = dio->bio_list;
560 dio->bio_list = bio->bi_private;
561 spin_unlock_irqrestore(&dio->bio_lock, flags);
562 ret2 = dio_bio_complete(dio, bio);
563 if (ret == 0)
564 ret = ret2;
565 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700566 sdio->reap_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 }
568 return ret;
569}
570
571/*
572 * Call into the fs to map some more disk blocks. We record the current number
Andi Kleeneb28be22011-08-01 21:38:03 -0700573 * of available blocks at sdio->blocks_available. These are in units of the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 * fs blocksize, (1 << inode->i_blkbits).
575 *
576 * The fs is allowed to map lots of blocks at once. If it wants to do that,
577 * it uses the passed inode-relative block number as the file offset, as usual.
578 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800579 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 * has remaining to do. The fs should not map more than this number of blocks.
581 *
582 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
583 * indicate how much contiguous disk space has been made available at
584 * bh->b_blocknr.
585 *
586 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
587 * This isn't very efficient...
588 *
589 * In the case of filesystem holes: the fs may return an arbitrarily-large
590 * hole by returning an appropriate value in b_size and by clearing
591 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800592 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593 */
Andi Kleen18772642011-08-01 21:38:07 -0700594static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
595 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596{
597 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
Tao Maae55e1a2012-01-12 17:20:33 -0800599 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600 unsigned long fs_count; /* Number of filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601 int create;
602
603 /*
604 * If there was a memory error and we've overwritten all the
605 * mapped blocks then we can now return that memory error
606 */
607 ret = dio->page_errors;
608 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700609 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
610 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
Tao Maae55e1a2012-01-12 17:20:33 -0800611 fs_endblk = (sdio->final_block_in_request - 1) >>
612 sdio->blkfactor;
613 fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614
Nathan Scott3c674e72006-03-29 09:26:15 +1000615 map_bh->b_state = 0;
616 map_bh->b_size = fs_count << dio->inode->i_blkbits;
617
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800618 /*
619 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
620 * forbid block creations: only overwrites are permitted.
621 * We will return early to the caller once we see an
622 * unmapped buffer head returned, and the caller will fall
623 * back to buffered I/O.
624 *
625 * Otherwise the decision is left to the get_blocks method,
626 * which may decide to handle it or also return an unmapped
627 * buffer head.
628 */
Jens Axboeb31dc662006-06-13 08:26:10 +0200629 create = dio->rw & WRITE;
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800630 if (dio->flags & DIO_SKIP_HOLES) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700631 if (sdio->block_in_file < (i_size_read(dio->inode) >>
632 sdio->blkbits))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000635
Andi Kleeneb28be22011-08-01 21:38:03 -0700636 ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637 map_bh, create);
Andi Kleen18772642011-08-01 21:38:07 -0700638
639 /* Store for completion */
640 dio->private = map_bh->b_private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 }
642 return ret;
643}
644
645/*
646 * There is no bio. Make one now.
647 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700648static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
649 sector_t start_sector, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650{
651 sector_t sector;
652 int ret, nr_pages;
653
Andi Kleeneb28be22011-08-01 21:38:03 -0700654 ret = dio_bio_reap(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655 if (ret)
656 goto out;
Andi Kleeneb28be22011-08-01 21:38:03 -0700657 sector = start_sector << (sdio->blkbits - 9);
Andi Kleen18772642011-08-01 21:38:07 -0700658 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
David Dillow20d96002011-01-20 14:44:22 -0800659 nr_pages = min(nr_pages, BIO_MAX_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660 BUG_ON(nr_pages <= 0);
Andi Kleen18772642011-08-01 21:38:07 -0700661 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
Andi Kleeneb28be22011-08-01 21:38:03 -0700662 sdio->boundary = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663out:
664 return ret;
665}
666
667/*
668 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
669 * that was successful then update final_block_in_bio and take a ref against
670 * the just-added page.
671 *
672 * Return zero on success. Non-zero means the caller needs to start a new BIO.
673 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700674static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675{
676 int ret;
677
Andi Kleeneb28be22011-08-01 21:38:03 -0700678 ret = bio_add_page(sdio->bio, sdio->cur_page,
679 sdio->cur_page_len, sdio->cur_page_offset);
680 if (ret == sdio->cur_page_len) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681 /*
682 * Decrement count only, if we are done with this page
683 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700684 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
685 sdio->pages_in_io--;
686 page_cache_get(sdio->cur_page);
687 sdio->final_block_in_bio = sdio->cur_page_block +
688 (sdio->cur_page_len >> sdio->blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689 ret = 0;
690 } else {
691 ret = 1;
692 }
693 return ret;
694}
695
696/*
697 * Put cur_page under IO. The section of cur_page which is described by
698 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
699 * starts on-disk at cur_page_block.
700 *
701 * We take a ref against the page here (on behalf of its presence in the bio).
702 *
703 * The caller of this function is responsible for removing cur_page from the
704 * dio, and for dropping the refcount which came from that presence.
705 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700706static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
707 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708{
709 int ret = 0;
710
Andi Kleeneb28be22011-08-01 21:38:03 -0700711 if (sdio->bio) {
712 loff_t cur_offset = sdio->cur_page_fs_offset;
713 loff_t bio_next_offset = sdio->logical_offset_in_bio +
714 sdio->bio->bi_size;
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400715
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400717 * See whether this new request is contiguous with the old.
718 *
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900719 * Btrfs cannot handle having logically non-contiguous requests
720 * submitted. For example if you have
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400721 *
722 * Logical: [0-4095][HOLE][8192-12287]
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900723 * Physical: [0-4095] [4096-8191]
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400724 *
725 * We cannot submit those pages together as one BIO. So if our
726 * current logical offset in the file does not equal what would
727 * be the next logical offset in the bio, submit the bio we
728 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700730 if (sdio->final_block_in_bio != sdio->cur_page_block ||
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400731 cur_offset != bio_next_offset)
Andi Kleeneb28be22011-08-01 21:38:03 -0700732 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 /*
734 * Submit now if the underlying fs is about to perform a
735 * metadata read
736 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700737 else if (sdio->boundary)
738 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 }
740
Andi Kleeneb28be22011-08-01 21:38:03 -0700741 if (sdio->bio == NULL) {
Andi Kleen18772642011-08-01 21:38:07 -0700742 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743 if (ret)
744 goto out;
745 }
746
Andi Kleeneb28be22011-08-01 21:38:03 -0700747 if (dio_bio_add_page(sdio) != 0) {
748 dio_bio_submit(dio, sdio);
Andi Kleen18772642011-08-01 21:38:07 -0700749 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700751 ret = dio_bio_add_page(sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 BUG_ON(ret != 0);
753 }
754 }
755out:
756 return ret;
757}
758
759/*
760 * An autonomous function to put a chunk of a page under deferred IO.
761 *
762 * The caller doesn't actually know (or care) whether this piece of page is in
763 * a BIO, or is under IO or whatever. We just take care of all possible
764 * situations here. The separation between the logic of do_direct_IO() and
765 * that of submit_page_section() is important for clarity. Please don't break.
766 *
767 * The chunk of page starts on-disk at blocknr.
768 *
769 * We perform deferred IO, by recording the last-submitted page inside our
770 * private part of the dio structure. If possible, we just expand the IO
771 * across that page here.
772 *
773 * If that doesn't work out then we put the old page into the bio and add this
774 * page to the dio instead.
775 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700776static inline int
Andi Kleeneb28be22011-08-01 21:38:03 -0700777submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
Andi Kleen18772642011-08-01 21:38:07 -0700778 unsigned offset, unsigned len, sector_t blocknr,
779 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780{
781 int ret = 0;
782
Andrew Morton98c4d572006-12-10 02:19:47 -0800783 if (dio->rw & WRITE) {
784 /*
785 * Read accounting is performed in submit_bio()
786 */
787 task_io_account_write(len);
788 }
789
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790 /*
791 * Can we just grow the current page's presence in the dio?
792 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700793 if (sdio->cur_page == page &&
794 sdio->cur_page_offset + sdio->cur_page_len == offset &&
795 sdio->cur_page_block +
796 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
797 sdio->cur_page_len += len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798
799 /*
Andi Kleeneb28be22011-08-01 21:38:03 -0700800 * If sdio->boundary then we want to schedule the IO now to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 * avoid metadata seeks.
802 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700803 if (sdio->boundary) {
Andi Kleen18772642011-08-01 21:38:07 -0700804 ret = dio_send_cur_page(dio, sdio, map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -0700805 page_cache_release(sdio->cur_page);
806 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 }
808 goto out;
809 }
810
811 /*
812 * If there's a deferred page already there then send it.
813 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700814 if (sdio->cur_page) {
Andi Kleen18772642011-08-01 21:38:07 -0700815 ret = dio_send_cur_page(dio, sdio, map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -0700816 page_cache_release(sdio->cur_page);
817 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818 if (ret)
819 goto out;
820 }
821
822 page_cache_get(page); /* It is in dio */
Andi Kleeneb28be22011-08-01 21:38:03 -0700823 sdio->cur_page = page;
824 sdio->cur_page_offset = offset;
825 sdio->cur_page_len = len;
826 sdio->cur_page_block = blocknr;
827 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828out:
829 return ret;
830}
831
832/*
833 * Clean any dirty buffers in the blockdev mapping which alias newly-created
834 * file blocks. Only called for S_ISREG files - blockdevs do not set
835 * buffer_new
836 */
Andi Kleen18772642011-08-01 21:38:07 -0700837static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838{
839 unsigned i;
840 unsigned nblocks;
841
Andi Kleen18772642011-08-01 21:38:07 -0700842 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843
844 for (i = 0; i < nblocks; i++) {
Andi Kleen18772642011-08-01 21:38:07 -0700845 unmap_underlying_metadata(map_bh->b_bdev,
846 map_bh->b_blocknr + i);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847 }
848}
849
850/*
851 * If we are not writing the entire block and get_block() allocated
852 * the block for us, we need to fill-in the unused portion of the
853 * block with zeros. This happens only if user-buffer, fileoffset or
854 * io length is not filesystem block-size multiple.
855 *
856 * `end' is zero if we're doing the start of the IO, 1 at the end of the
857 * IO.
858 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700859static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
860 int end, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861{
862 unsigned dio_blocks_per_fs_block;
863 unsigned this_chunk_blocks; /* In dio_blocks */
864 unsigned this_chunk_bytes;
865 struct page *page;
866
Andi Kleeneb28be22011-08-01 21:38:03 -0700867 sdio->start_zero_done = 1;
Andi Kleen18772642011-08-01 21:38:07 -0700868 if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869 return;
870
Andi Kleeneb28be22011-08-01 21:38:03 -0700871 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
872 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873
874 if (!this_chunk_blocks)
875 return;
876
877 /*
878 * We need to zero out part of an fs block. It is either at the
879 * beginning or the end of the fs block.
880 */
881 if (end)
882 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
883
Andi Kleeneb28be22011-08-01 21:38:03 -0700884 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885
Nick Piggin557ed1f2007-10-16 01:24:40 -0700886 page = ZERO_PAGE(0);
Andi Kleeneb28be22011-08-01 21:38:03 -0700887 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700888 sdio->next_block_for_io, map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 return;
890
Andi Kleeneb28be22011-08-01 21:38:03 -0700891 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892}
893
894/*
895 * Walk the user pages, and the file, mapping blocks to disk and generating
896 * a sequence of (page,offset,len,block) mappings. These mappings are injected
897 * into submit_page_section(), which takes care of the next stage of submission
898 *
899 * Direct IO against a blockdev is different from a file. Because we can
900 * happily perform page-sized but 512-byte aligned IOs. It is important that
901 * blockdev IO be able to have fine alignment and large sizes.
902 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800903 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904 * with the size of IO which is permitted at this offset and this i_blkbits.
905 *
906 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800907 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908 * fine alignment but still allows this function to work in PAGE_SIZE units.
909 */
Andi Kleen18772642011-08-01 21:38:07 -0700910static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
911 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912{
Andi Kleeneb28be22011-08-01 21:38:03 -0700913 const unsigned blkbits = sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
915 struct page *page;
916 unsigned block_in_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917 int ret = 0;
918
919 /* The I/O can start at any block offset within the first page */
Andi Kleeneb28be22011-08-01 21:38:03 -0700920 block_in_page = sdio->first_block_in_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921
Andi Kleeneb28be22011-08-01 21:38:03 -0700922 while (sdio->block_in_file < sdio->final_block_in_request) {
923 page = dio_get_page(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924 if (IS_ERR(page)) {
925 ret = PTR_ERR(page);
926 goto out;
927 }
928
929 while (block_in_page < blocks_per_page) {
930 unsigned offset_in_page = block_in_page << blkbits;
931 unsigned this_chunk_bytes; /* # of bytes mapped */
932 unsigned this_chunk_blocks; /* # of blocks */
933 unsigned u;
934
Andi Kleeneb28be22011-08-01 21:38:03 -0700935 if (sdio->blocks_available == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936 /*
937 * Need to go and map some more disk
938 */
939 unsigned long blkmask;
940 unsigned long dio_remainder;
941
Andi Kleen18772642011-08-01 21:38:07 -0700942 ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 if (ret) {
944 page_cache_release(page);
945 goto out;
946 }
947 if (!buffer_mapped(map_bh))
948 goto do_holes;
949
Andi Kleeneb28be22011-08-01 21:38:03 -0700950 sdio->blocks_available =
951 map_bh->b_size >> sdio->blkbits;
952 sdio->next_block_for_io =
953 map_bh->b_blocknr << sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 if (buffer_new(map_bh))
Andi Kleen18772642011-08-01 21:38:07 -0700955 clean_blockdev_aliases(dio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956
Andi Kleeneb28be22011-08-01 21:38:03 -0700957 if (!sdio->blkfactor)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 goto do_holes;
959
Andi Kleeneb28be22011-08-01 21:38:03 -0700960 blkmask = (1 << sdio->blkfactor) - 1;
961 dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962
963 /*
964 * If we are at the start of IO and that IO
965 * starts partway into a fs-block,
966 * dio_remainder will be non-zero. If the IO
967 * is a read then we can simply advance the IO
968 * cursor to the first block which is to be
969 * read. But if the IO is a write and the
970 * block was newly allocated we cannot do that;
971 * the start of the fs block must be zeroed out
972 * on-disk
973 */
974 if (!buffer_new(map_bh))
Andi Kleeneb28be22011-08-01 21:38:03 -0700975 sdio->next_block_for_io += dio_remainder;
976 sdio->blocks_available -= dio_remainder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 }
978do_holes:
979 /* Handle holes */
980 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800981 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982
983 /* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboeb31dc662006-06-13 08:26:10 +0200984 if (dio->rw & WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985 page_cache_release(page);
986 return -ENOTBLK;
987 }
988
Jeff Moyer35dc8162006-02-03 03:04:27 -0800989 /*
990 * Be sure to account for a partial block as the
991 * last block in the file
992 */
993 i_size_aligned = ALIGN(i_size_read(dio->inode),
994 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -0700995 if (sdio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -0800996 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997 /* We hit eof */
998 page_cache_release(page);
999 goto out;
1000 }
Christoph Lametereebd2aa2008-02-04 22:28:29 -08001001 zero_user(page, block_in_page << blkbits,
1002 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -07001003 sdio->block_in_file++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004 block_in_page++;
1005 goto next_block;
1006 }
1007
1008 /*
1009 * If we're performing IO which has an alignment which
1010 * is finer than the underlying fs, go check to see if
1011 * we must zero out the start of this block.
1012 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001013 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
Andi Kleen18772642011-08-01 21:38:07 -07001014 dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015
1016 /*
1017 * Work out, in this_chunk_blocks, how much disk we
1018 * can add to this page
1019 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001020 this_chunk_blocks = sdio->blocks_available;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021 u = (PAGE_SIZE - offset_in_page) >> blkbits;
1022 if (this_chunk_blocks > u)
1023 this_chunk_blocks = u;
Andi Kleeneb28be22011-08-01 21:38:03 -07001024 u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025 if (this_chunk_blocks > u)
1026 this_chunk_blocks = u;
1027 this_chunk_bytes = this_chunk_blocks << blkbits;
1028 BUG_ON(this_chunk_bytes == 0);
1029
Andi Kleeneb28be22011-08-01 21:38:03 -07001030 sdio->boundary = buffer_boundary(map_bh);
1031 ret = submit_page_section(dio, sdio, page,
1032 offset_in_page,
1033 this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -07001034 sdio->next_block_for_io,
1035 map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036 if (ret) {
1037 page_cache_release(page);
1038 goto out;
1039 }
Andi Kleeneb28be22011-08-01 21:38:03 -07001040 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041
Andi Kleeneb28be22011-08-01 21:38:03 -07001042 sdio->block_in_file += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 block_in_page += this_chunk_blocks;
Andi Kleeneb28be22011-08-01 21:38:03 -07001044 sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045next_block:
Andi Kleeneb28be22011-08-01 21:38:03 -07001046 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1047 if (sdio->block_in_file == sdio->final_block_in_request)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 break;
1049 }
1050
1051 /* Drop the ref which was taken in get_user_pages() */
1052 page_cache_release(page);
1053 block_in_page = 0;
1054 }
1055out:
1056 return ret;
1057}
1058
Andi Kleen847cc632011-08-01 21:38:09 -07001059static inline int drop_refcount(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060{
Andi Kleen847cc632011-08-01 21:38:09 -07001061 int ret2;
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001062 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063
Zach Brown8459d862006-12-10 02:21:05 -08001064 /*
1065 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001066 * operation. AIO can if it was a broken operation described above or
1067 * in fact if all the bios race to complete before we get here. In
1068 * that case dio_complete() translates the EIOCBQUEUED into the proper
1069 * return code that the caller will hand to aio_complete().
1070 *
1071 * This is managed by the bio_lock instead of being an atomic_t so that
1072 * completion paths can drop their ref and use the remaining count to
1073 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001074 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001075 spin_lock_irqsave(&dio->bio_lock, flags);
1076 ret2 = --dio->refcount;
1077 spin_unlock_irqrestore(&dio->bio_lock, flags);
Andi Kleen847cc632011-08-01 21:38:09 -07001078 return ret2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079}
1080
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001081/*
1082 * This is a library function for use by filesystem drivers.
1083 *
1084 * The locking rules are governed by the flags parameter:
1085 * - if the flags value contains DIO_LOCKING we use a fancy locking
1086 * scheme for dumb filesystems.
1087 * For writes this function is called under i_mutex and returns with
1088 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1089 * taken and dropped again before returning.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001090 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1091 * internal locking but rather rely on the filesystem to synchronize
1092 * direct I/O reads/writes versus each other and truncate.
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001093 *
1094 * To help with locking against truncate we incremented the i_dio_count
1095 * counter before starting direct I/O, and decrement it once we are done.
1096 * Truncate can wait for it to reach zero to provide exclusion. It is
1097 * expected that filesystem provide exclusion between new direct I/O
1098 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1099 * but other filesystems need to take care of this on their own.
Andi Kleenba253fb2011-08-01 21:38:08 -07001100 *
1101 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1102 * is always inlined. Otherwise gcc is unable to split the structure into
1103 * individual fields and will generate much worse code. This is important
1104 * for the whole file.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001105 */
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001106static inline ssize_t
1107do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108 struct block_device *bdev, const struct iovec *iov, loff_t offset,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -08001109 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
Josef Bacikfacd07b2010-05-23 11:00:55 -04001110 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111{
1112 int seg;
1113 size_t size;
1114 unsigned long addr;
1115 unsigned blkbits = inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116 unsigned blocksize_mask = (1 << blkbits) - 1;
1117 ssize_t retval = -EINVAL;
1118 loff_t end = offset;
1119 struct dio *dio;
Andi Kleeneb28be22011-08-01 21:38:03 -07001120 struct dio_submit sdio = { 0, };
Andi Kleen847cc632011-08-01 21:38:09 -07001121 unsigned long user_addr;
1122 size_t bytes;
1123 struct buffer_head map_bh = { 0, };
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124
1125 if (rw & WRITE)
Jens Axboe721a9602011-03-09 11:56:30 +01001126 rw = WRITE_ODIRECT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001128 /*
1129 * Avoid references to bdev if not absolutely needed to give
1130 * the early prefetch in the caller enough time.
1131 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132
1133 if (offset & blocksize_mask) {
1134 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001135 blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136 blocksize_mask = (1 << blkbits) - 1;
1137 if (offset & blocksize_mask)
1138 goto out;
1139 }
1140
1141 /* Check the memory alignment. Blocks cannot straddle pages */
1142 for (seg = 0; seg < nr_segs; seg++) {
1143 addr = (unsigned long)iov[seg].iov_base;
1144 size = iov[seg].iov_len;
1145 end += size;
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001146 if (unlikely((addr & blocksize_mask) ||
1147 (size & blocksize_mask))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001149 blkbits = blksize_bits(
1150 bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 blocksize_mask = (1 << blkbits) - 1;
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001152 if ((addr & blocksize_mask) || (size & blocksize_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153 goto out;
1154 }
1155 }
1156
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001157 /* watch out for a 0 len io from a tricksy fs */
1158 if (rw == READ && end == offset)
1159 return 0;
1160
Andi Kleen6e8267f2011-08-01 21:38:06 -07001161 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162 retval = -ENOMEM;
1163 if (!dio)
1164 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001165 /*
1166 * Believe it or not, zeroing out the page array caused a .5%
1167 * performance regression in a database benchmark. So, we take
1168 * care to only zero out what's needed.
1169 */
1170 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001172 dio->flags = flags;
1173 if (dio->flags & DIO_LOCKING) {
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001174 if (rw == READ) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001175 struct address_space *mapping =
1176 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001178 /* will be released by direct_io_worker */
1179 mutex_lock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180
1181 retval = filemap_write_and_wait_range(mapping, offset,
1182 end - 1);
1183 if (retval) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001184 mutex_unlock(&inode->i_mutex);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001185 kmem_cache_free(dio_cache, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 goto out;
1187 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 }
1190
1191 /*
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001192 * Will be decremented at I/O completion time.
1193 */
1194 atomic_inc(&inode->i_dio_count);
1195
1196 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197 * For file extending writes updating i_size before data
1198 * writeouts complete can expose uninitialized blocks. So
1199 * even for AIO, we need to wait for i/o to complete before
1200 * returning in this case.
1201 */
Jens Axboeb31dc662006-06-13 08:26:10 +02001202 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203 (end > i_size_read(inode)));
1204
Andi Kleen847cc632011-08-01 21:38:09 -07001205 retval = 0;
1206
1207 dio->inode = inode;
1208 dio->rw = rw;
1209 sdio.blkbits = blkbits;
1210 sdio.blkfactor = inode->i_blkbits - blkbits;
1211 sdio.block_in_file = offset >> blkbits;
1212
1213 sdio.get_block = get_block;
1214 dio->end_io = end_io;
1215 sdio.submit_io = submit_io;
1216 sdio.final_block_in_bio = -1;
1217 sdio.next_block_for_io = -1;
1218
1219 dio->iocb = iocb;
1220 dio->i_size = i_size_read(inode);
1221
1222 spin_lock_init(&dio->bio_lock);
1223 dio->refcount = 1;
1224
1225 /*
1226 * In case of non-aligned buffers, we may need 2 more
1227 * pages since we need to zero out first and last block.
1228 */
1229 if (unlikely(sdio.blkfactor))
1230 sdio.pages_in_io = 2;
1231
1232 for (seg = 0; seg < nr_segs; seg++) {
1233 user_addr = (unsigned long)iov[seg].iov_base;
1234 sdio.pages_in_io +=
1235 ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) /
1236 PAGE_SIZE - user_addr / PAGE_SIZE);
1237 }
1238
1239 for (seg = 0; seg < nr_segs; seg++) {
1240 user_addr = (unsigned long)iov[seg].iov_base;
1241 sdio.size += bytes = iov[seg].iov_len;
1242
1243 /* Index into the first page of the first block */
1244 sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1245 sdio.final_block_in_request = sdio.block_in_file +
1246 (bytes >> blkbits);
1247 /* Page fetching state */
1248 sdio.head = 0;
1249 sdio.tail = 0;
1250 sdio.curr_page = 0;
1251
1252 sdio.total_pages = 0;
1253 if (user_addr & (PAGE_SIZE-1)) {
1254 sdio.total_pages++;
1255 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1256 }
1257 sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1258 sdio.curr_user_address = user_addr;
1259
1260 retval = do_direct_IO(dio, &sdio, &map_bh);
1261
1262 dio->result += iov[seg].iov_len -
1263 ((sdio.final_block_in_request - sdio.block_in_file) <<
1264 blkbits);
1265
1266 if (retval) {
1267 dio_cleanup(dio, &sdio);
1268 break;
1269 }
1270 } /* end iovec loop */
1271
1272 if (retval == -ENOTBLK) {
1273 /*
1274 * The remaining part of the request will be
1275 * be handled by buffered I/O when we return
1276 */
1277 retval = 0;
1278 }
1279 /*
1280 * There may be some unwritten disk at the end of a part-written
1281 * fs-block-sized block. Go zero that now.
1282 */
1283 dio_zero_block(dio, &sdio, 1, &map_bh);
1284
1285 if (sdio.cur_page) {
1286 ssize_t ret2;
1287
1288 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1289 if (retval == 0)
1290 retval = ret2;
1291 page_cache_release(sdio.cur_page);
1292 sdio.cur_page = NULL;
1293 }
1294 if (sdio.bio)
1295 dio_bio_submit(dio, &sdio);
1296
1297 /*
1298 * It is possible that, we return short IO due to end of file.
1299 * In that case, we need to release all the pages we got hold on.
1300 */
1301 dio_cleanup(dio, &sdio);
1302
1303 /*
1304 * All block lookups have been performed. For READ requests
1305 * we can let i_mutex go now that its achieved its purpose
1306 * of protecting us from looking up uninitialized blocks.
1307 */
1308 if (rw == READ && (dio->flags & DIO_LOCKING))
1309 mutex_unlock(&dio->inode->i_mutex);
1310
1311 /*
1312 * The only time we want to leave bios in flight is when a successful
1313 * partial aio read or full aio write have been setup. In that case
1314 * bio completion will call aio_complete. The only time it's safe to
1315 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1316 * This had *better* be the only place that raises -EIOCBQUEUED.
1317 */
1318 BUG_ON(retval == -EIOCBQUEUED);
1319 if (dio->is_async && retval == 0 && dio->result &&
1320 ((rw & READ) || (dio->result == sdio.size)))
1321 retval = -EIOCBQUEUED;
1322
1323 if (retval != -EIOCBQUEUED)
1324 dio_await_completion(dio);
1325
1326 if (drop_refcount(dio) == 0) {
1327 retval = dio_complete(dio, offset, retval, false);
1328 kmem_cache_free(dio_cache, dio);
1329 } else
1330 BUG_ON(retval != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001332out:
1333 return retval;
1334}
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001335
1336ssize_t
1337__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1338 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1339 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1340 dio_submit_t submit_io, int flags)
1341{
1342 /*
1343 * The block device state is needed in the end to finally
1344 * submit everything. Since it's likely to be cache cold
1345 * prefetch it here as first thing to hide some of the
1346 * latency.
1347 *
1348 * Attempt to prefetch the pieces we likely need later.
1349 */
1350 prefetch(&bdev->bd_disk->part_tbl);
1351 prefetch(bdev->bd_queue);
1352 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1353
1354 return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset,
1355 nr_segs, get_block, end_io,
1356 submit_io, flags);
1357}
1358
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359EXPORT_SYMBOL(__blockdev_direct_IO);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001360
1361static __init int dio_init(void)
1362{
1363 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1364 return 0;
1365}
1366module_init(dio_init)