blob: 31ba0935e32ed2f271253a1d828778a91193b211 [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 Sharma600634972011-07-26 16:09:06 -070038#include <linux/atomic.h>
Andi Kleen65dd2aa2012-01-12 17:20:35 -080039#include <linux/prefetch.h>
Kent Overstreeta27bb332013-05-07 16:19:08 -070040#include <linux/aio.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070041
42/*
43 * How many user pages to map in one call to get_user_pages(). This determines
Andi Kleencde1ecb2011-08-01 21:38:04 -070044 * the size of a structure in the slab cache
Linus Torvalds1da177e2005-04-16 15:20:36 -070045 */
46#define DIO_PAGES 64
47
48/*
49 * This code generally works in units of "dio_blocks". A dio_block is
50 * somewhere between the hard sector size and the filesystem block size. it
51 * is determined on a per-invocation basis. When talking to the filesystem
52 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
53 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
54 * to bio_block quantities by shifting left by blkfactor.
55 *
56 * If blkfactor is zero then the user's request was aligned to the filesystem's
57 * blocksize.
Linus Torvalds1da177e2005-04-16 15:20:36 -070058 */
59
Andi Kleeneb28be22011-08-01 21:38:03 -070060/* dio_state only used in the submission path */
61
62struct dio_submit {
Linus Torvalds1da177e2005-04-16 15:20:36 -070063 struct bio *bio; /* bio under assembly */
Linus Torvalds1da177e2005-04-16 15:20:36 -070064 unsigned blkbits; /* doesn't change */
65 unsigned blkfactor; /* When we're using an alignment which
66 is finer than the filesystem's soft
67 blocksize, this specifies how much
68 finer. blkfactor=2 means 1/4-block
69 alignment. Does not change */
70 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
71 been performed at the start of a
72 write */
73 int pages_in_io; /* approximate total IO pages */
74 size_t size; /* total request size (doesn't change)*/
75 sector_t block_in_file; /* Current offset into the underlying
76 file in dio_block units. */
77 unsigned blocks_available; /* At block_in_file. changes */
Andi Kleen0dc2bc42011-08-01 21:38:05 -070078 int reap_counter; /* rate limit reaping */
Linus Torvalds1da177e2005-04-16 15:20:36 -070079 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 */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080082 get_block_t *get_block; /* block mapping function */
Josef Bacikfacd07b2010-05-23 11:00:55 -040083 dio_submit_t *submit_io; /* IO submition function */
Andi Kleeneb28be22011-08-01 21:38:03 -070084
Josef Bacikfacd07b2010-05-23 11:00:55 -040085 loff_t logical_offset_in_bio; /* current first logical block in bio */
Linus Torvalds1da177e2005-04-16 15:20:36 -070086 sector_t final_block_in_bio; /* current final block in bio + 1 */
87 sector_t next_block_for_io; /* next block to be put under IO,
88 in dio_blocks units */
Linus Torvalds1da177e2005-04-16 15:20:36 -070089
90 /*
91 * Deferred addition of a page to the dio. These variables are
92 * private to dio_send_cur_page(), submit_page_section() and
93 * dio_bio_add_page().
94 */
95 struct page *cur_page; /* The page */
96 unsigned cur_page_offset; /* Offset into it, in bytes */
97 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
98 sector_t cur_page_block; /* Where it starts */
Josef Bacikfacd07b2010-05-23 11:00:55 -040099 loff_t cur_page_fs_offset; /* Offset in file */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100
Jeff Moyer23aee092009-12-15 16:47:49 -0800101 /*
102 * Page fetching state. These variables belong to dio_refill_pages().
103 */
104 int curr_page; /* changes */
105 int total_pages; /* doesn't change */
106 unsigned long curr_user_address;/* changes */
107
108 /*
109 * Page queue. These variables belong to dio_refill_pages() and
110 * dio_get_page().
111 */
112 unsigned head; /* next page to process */
113 unsigned tail; /* last valid page + 1 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700114};
115
116/* dio_state communicated between submission path and end_io */
117struct dio {
118 int flags; /* doesn't change */
Andi Kleeneb28be22011-08-01 21:38:03 -0700119 int rw;
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700120 struct inode *inode;
Andi Kleeneb28be22011-08-01 21:38:03 -0700121 loff_t i_size; /* i_size when submitted */
122 dio_iodone_t *end_io; /* IO completion function */
Andi Kleeneb28be22011-08-01 21:38:03 -0700123
Andi Kleen18772642011-08-01 21:38:07 -0700124 void *private; /* copy from map_bh.b_private */
Andi Kleeneb28be22011-08-01 21:38:03 -0700125
126 /* BIO completion state */
127 spinlock_t bio_lock; /* protects BIO fields below */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700128 int page_errors; /* errno from get_user_pages() */
129 int is_async; /* is IO async ? */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200130 bool defer_completion; /* defer AIO completion to workqueue? */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700131 int io_error; /* IO error in completion path */
Andi Kleeneb28be22011-08-01 21:38:03 -0700132 unsigned long refcount; /* direct_io_worker() and bios */
133 struct bio *bio_list; /* singly linked via bi_private */
134 struct task_struct *waiter; /* waiting task (NULL if none) */
135
136 /* AIO related stuff */
137 struct kiocb *iocb; /* kiocb */
Andi Kleeneb28be22011-08-01 21:38:03 -0700138 ssize_t result; /* IO result */
139
Jeff Moyer23aee092009-12-15 16:47:49 -0800140 /*
141 * pages[] (and any fields placed after it) are not zeroed out at
142 * allocation time. Don't add new fields after pages[] unless you
143 * wish that they not be zeroed.
144 */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200145 union {
146 struct page *pages[DIO_PAGES]; /* page buffer */
147 struct work_struct complete_work;/* deferred AIO completion */
148 };
Andi Kleen6e8267f2011-08-01 21:38:06 -0700149} ____cacheline_aligned_in_smp;
150
151static struct kmem_cache *dio_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152
153/*
154 * How many pages are in the queue?
155 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700156static inline unsigned dio_pages_present(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157{
Andi Kleeneb28be22011-08-01 21:38:03 -0700158 return sdio->tail - sdio->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159}
160
161/*
162 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
163 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700164static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165{
166 int ret;
167 int nr_pages;
168
Andi Kleeneb28be22011-08-01 21:38:03 -0700169 nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
Nick Pigginf5dd33c2008-07-25 19:45:25 -0700170 ret = get_user_pages_fast(
Andi Kleeneb28be22011-08-01 21:38:03 -0700171 sdio->curr_user_address, /* Where from? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 nr_pages, /* How many pages? */
173 dio->rw == READ, /* Write to memory? */
Nick Pigginf5dd33c2008-07-25 19:45:25 -0700174 &dio->pages[0]); /* Put results here */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175
Andi Kleeneb28be22011-08-01 21:38:03 -0700176 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
Nick Piggin557ed1f2007-10-16 01:24:40 -0700177 struct page *page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178 /*
179 * A memory fault, but the filesystem has some outstanding
180 * mapped blocks. We need to use those blocks up to avoid
181 * leaking stale data in the file.
182 */
183 if (dio->page_errors == 0)
184 dio->page_errors = ret;
Nick Pigginb5810032005-10-29 18:16:12 -0700185 page_cache_get(page);
186 dio->pages[0] = page;
Andi Kleeneb28be22011-08-01 21:38:03 -0700187 sdio->head = 0;
188 sdio->tail = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189 ret = 0;
190 goto out;
191 }
192
193 if (ret >= 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700194 sdio->curr_user_address += ret * PAGE_SIZE;
195 sdio->curr_page += ret;
196 sdio->head = 0;
197 sdio->tail = ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198 ret = 0;
199 }
200out:
201 return ret;
202}
203
204/*
205 * Get another userspace page. Returns an ERR_PTR on error. Pages are
206 * buffered inside the dio so that we can call get_user_pages() against a
207 * decent number of pages, less frequently. To provide nicer use of the
208 * L1 cache.
209 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700210static inline struct page *dio_get_page(struct dio *dio,
211 struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212{
Andi Kleeneb28be22011-08-01 21:38:03 -0700213 if (dio_pages_present(sdio) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214 int ret;
215
Andi Kleeneb28be22011-08-01 21:38:03 -0700216 ret = dio_refill_pages(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217 if (ret)
218 return ERR_PTR(ret);
Andi Kleeneb28be22011-08-01 21:38:03 -0700219 BUG_ON(dio_pages_present(sdio) == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700221 return dio->pages[sdio->head++];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222}
223
Zach Brown6d544bb2006-12-10 02:20:54 -0800224/**
225 * dio_complete() - called when all DIO BIO I/O has been completed
226 * @offset: the byte offset in the file of the completed operation
227 *
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200228 * This drops i_dio_count, lets interested parties know that a DIO operation
229 * has completed, and calculates the resulting return code for the operation.
Zach Brown6d544bb2006-12-10 02:20:54 -0800230 *
231 * It lets the filesystem know if it registered an interest earlier via
232 * get_block. Pass the private field of the map buffer_head so that
233 * filesystems can use it to hold additional state between get_block calls and
234 * dio_complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235 */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200236static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret,
237 bool is_async)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238{
Zach Brown6d544bb2006-12-10 02:20:54 -0800239 ssize_t transferred = 0;
240
Zach Brown8459d862006-12-10 02:21:05 -0800241 /*
242 * AIO submission can race with bio completion to get here while
243 * expecting to have the last io completed by bio completion.
244 * In that case -EIOCBQUEUED is in fact not an error we want
245 * to preserve through this call.
246 */
247 if (ret == -EIOCBQUEUED)
248 ret = 0;
249
Zach Brown6d544bb2006-12-10 02:20:54 -0800250 if (dio->result) {
251 transferred = dio->result;
252
253 /* Check for short read case */
254 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
255 transferred = dio->i_size - offset;
256 }
257
Zach Brown6d544bb2006-12-10 02:20:54 -0800258 if (ret == 0)
259 ret = dio->page_errors;
260 if (ret == 0)
261 ret = dio->io_error;
262 if (ret == 0)
263 ret = transferred;
264
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200265 if (dio->end_io && dio->result)
266 dio->end_io(dio->iocb, offset, transferred, dio->private);
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000267
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200268 inode_dio_done(dio->inode);
Christoph Hellwig02afc272013-09-04 15:04:40 +0200269 if (is_async) {
270 if (dio->rw & WRITE) {
271 int err;
272
273 err = generic_write_sync(dio->iocb->ki_filp, offset,
274 transferred);
275 if (err < 0 && ret > 0)
276 ret = err;
277 }
278
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200279 aio_complete(dio->iocb, ret, 0);
Christoph Hellwig02afc272013-09-04 15:04:40 +0200280 }
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200281
282 kmem_cache_free(dio_cache, dio);
Zach Brown6d544bb2006-12-10 02:20:54 -0800283 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284}
285
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200286static void dio_aio_complete_work(struct work_struct *work)
287{
288 struct dio *dio = container_of(work, struct dio, complete_work);
289
290 dio_complete(dio, dio->iocb->ki_pos, 0, true);
291}
292
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293static int dio_bio_complete(struct dio *dio, struct bio *bio);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200294
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295/*
296 * Asynchronous IO callback.
297 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200298static void dio_bio_end_aio(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299{
300 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800301 unsigned long remaining;
302 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304 /* cleanup the bio */
305 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800306
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800307 spin_lock_irqsave(&dio->bio_lock, flags);
308 remaining = --dio->refcount;
309 if (remaining == 1 && dio->waiter)
Zach Brown20258b22006-12-10 02:21:01 -0800310 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800311 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b22006-12-10 02:21:01 -0800312
Zach Brown8459d862006-12-10 02:21:05 -0800313 if (remaining == 0) {
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200314 if (dio->result && dio->defer_completion) {
315 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
316 queue_work(dio->inode->i_sb->s_dio_done_wq,
317 &dio->complete_work);
318 } else {
319 dio_complete(dio, dio->iocb->ki_pos, 0, true);
320 }
Zach Brown8459d862006-12-10 02:21:05 -0800321 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322}
323
324/*
325 * The BIO completion handler simply queues the BIO up for the process-context
326 * handler.
327 *
328 * During I/O bi_private points at the dio. After I/O, bi_private is used to
329 * implement a singly-linked list of completed BIOs, at dio->bio_list.
330 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200331static void dio_bio_end_io(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332{
333 struct dio *dio = bio->bi_private;
334 unsigned long flags;
335
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336 spin_lock_irqsave(&dio->bio_lock, flags);
337 bio->bi_private = dio->bio_list;
338 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800339 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 wake_up_process(dio->waiter);
341 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342}
343
Josef Bacikfacd07b2010-05-23 11:00:55 -0400344/**
345 * dio_end_io - handle the end io action for the given bio
346 * @bio: The direct io bio thats being completed
347 * @error: Error if there was one
348 *
349 * This is meant to be called by any filesystem that uses their own dio_submit_t
350 * so that the DIO specific endio actions are dealt with after the filesystem
351 * has done it's completion work.
352 */
353void dio_end_io(struct bio *bio, int error)
354{
355 struct dio *dio = bio->bi_private;
356
357 if (dio->is_async)
358 dio_bio_end_aio(bio, error);
359 else
360 dio_bio_end_io(bio, error);
361}
362EXPORT_SYMBOL_GPL(dio_end_io);
363
Andi Kleenba253fb2011-08-01 21:38:08 -0700364static inline void
Andi Kleeneb28be22011-08-01 21:38:03 -0700365dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
366 struct block_device *bdev,
367 sector_t first_sector, int nr_vecs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368{
369 struct bio *bio;
370
David Dillow20d96002011-01-20 14:44:22 -0800371 /*
372 * bio_alloc() is guaranteed to return a bio when called with
373 * __GFP_WAIT and we request a valid number of vectors.
374 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376
377 bio->bi_bdev = bdev;
Kent Overstreet4f024f32013-10-11 15:44:27 -0700378 bio->bi_iter.bi_sector = first_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 if (dio->is_async)
380 bio->bi_end_io = dio_bio_end_aio;
381 else
382 bio->bi_end_io = dio_bio_end_io;
383
Andi Kleeneb28be22011-08-01 21:38:03 -0700384 sdio->bio = bio;
385 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386}
387
388/*
389 * In the AIO read case we speculatively dirty the pages before starting IO.
390 * During IO completion, any of these pages which happen to have been written
391 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800392 *
393 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700395static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396{
Andi Kleeneb28be22011-08-01 21:38:03 -0700397 struct bio *bio = sdio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800398 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399
400 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800401
402 spin_lock_irqsave(&dio->bio_lock, flags);
403 dio->refcount++;
404 spin_unlock_irqrestore(&dio->bio_lock, flags);
405
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 if (dio->is_async && dio->rw == READ)
407 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800408
Andi Kleeneb28be22011-08-01 21:38:03 -0700409 if (sdio->submit_io)
410 sdio->submit_io(dio->rw, bio, dio->inode,
411 sdio->logical_offset_in_bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400412 else
413 submit_bio(dio->rw, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414
Andi Kleeneb28be22011-08-01 21:38:03 -0700415 sdio->bio = NULL;
416 sdio->boundary = 0;
417 sdio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418}
419
420/*
421 * Release any resources in case of a failure
422 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700423static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424{
Andi Kleeneb28be22011-08-01 21:38:03 -0700425 while (dio_pages_present(sdio))
426 page_cache_release(dio_get_page(dio, sdio));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427}
428
429/*
Zach Brown02732012006-12-10 02:20:59 -0800430 * Wait for the next BIO to complete. Remove it and return it. NULL is
431 * returned once all BIOs have been completed. This must only be called once
432 * all bios have been issued so that dio->refcount can only decrease. This
433 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 */
435static struct bio *dio_await_one(struct dio *dio)
436{
437 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800438 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439
440 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800441
442 /*
443 * Wait as long as the list is empty and there are bios in flight. bio
444 * completion drops the count, maybe adds to the list, and wakes while
445 * holding the bio_lock so we don't need set_current_state()'s barrier
446 * and can call it after testing our condition.
447 */
448 while (dio->refcount > 1 && dio->bio_list == NULL) {
449 __set_current_state(TASK_UNINTERRUPTIBLE);
450 dio->waiter = current;
451 spin_unlock_irqrestore(&dio->bio_lock, flags);
452 io_schedule();
453 /* wake up sets us TASK_RUNNING */
454 spin_lock_irqsave(&dio->bio_lock, flags);
455 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 }
Zach Brown02732012006-12-10 02:20:59 -0800457 if (dio->bio_list) {
458 bio = dio->bio_list;
459 dio->bio_list = bio->bi_private;
460 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 spin_unlock_irqrestore(&dio->bio_lock, flags);
462 return bio;
463}
464
465/*
466 * Process one completed BIO. No locks are held.
467 */
468static int dio_bio_complete(struct dio *dio, struct bio *bio)
469{
470 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
Kent Overstreetcb34e052012-09-05 15:22:02 -0700471 struct bio_vec *bvec;
472 unsigned i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473
474 if (!uptodate)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800475 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476
477 if (dio->is_async && dio->rw == READ) {
478 bio_check_pages_dirty(bio); /* transfers ownership */
479 } else {
Kent Overstreetcb34e052012-09-05 15:22:02 -0700480 bio_for_each_segment_all(bvec, bio, i) {
481 struct page *page = bvec->bv_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482
483 if (dio->rw == READ && !PageCompound(page))
484 set_page_dirty_lock(page);
485 page_cache_release(page);
486 }
487 bio_put(bio);
488 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 return uptodate ? 0 : -EIO;
490}
491
492/*
Zach Brown02732012006-12-10 02:20:59 -0800493 * Wait on and process all in-flight BIOs. This must only be called once
494 * all bios have been issued so that the refcount can only decrease.
495 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200496 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800497 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800499static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500{
Zach Brown02732012006-12-10 02:20:59 -0800501 struct bio *bio;
502 do {
503 bio = dio_await_one(dio);
504 if (bio)
505 dio_bio_complete(dio, bio);
506 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507}
508
509/*
510 * A really large O_DIRECT read or write can generate a lot of BIOs. So
511 * to keep the memory consumption sane we periodically reap any completed BIOs
512 * during the BIO generation phase.
513 *
514 * This also helps to limit the peak amount of pinned userspace memory.
515 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700516static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517{
518 int ret = 0;
519
Andi Kleeneb28be22011-08-01 21:38:03 -0700520 if (sdio->reap_counter++ >= 64) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 while (dio->bio_list) {
522 unsigned long flags;
523 struct bio *bio;
524 int ret2;
525
526 spin_lock_irqsave(&dio->bio_lock, flags);
527 bio = dio->bio_list;
528 dio->bio_list = bio->bi_private;
529 spin_unlock_irqrestore(&dio->bio_lock, flags);
530 ret2 = dio_bio_complete(dio, bio);
531 if (ret == 0)
532 ret = ret2;
533 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700534 sdio->reap_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 }
536 return ret;
537}
538
539/*
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200540 * Create workqueue for deferred direct IO completions. We allocate the
541 * workqueue when it's first needed. This avoids creating workqueue for
542 * filesystems that don't need it and also allows us to create the workqueue
543 * late enough so the we can include s_id in the name of the workqueue.
544 */
545static int sb_init_dio_done_wq(struct super_block *sb)
546{
Olof Johansson45150c42013-09-09 10:34:23 -0700547 struct workqueue_struct *old;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200548 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
549 WQ_MEM_RECLAIM, 0,
550 sb->s_id);
551 if (!wq)
552 return -ENOMEM;
553 /*
554 * This has to be atomic as more DIOs can race to create the workqueue
555 */
Olof Johansson45150c42013-09-09 10:34:23 -0700556 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200557 /* Someone created workqueue before us? Free ours... */
Olof Johansson45150c42013-09-09 10:34:23 -0700558 if (old)
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200559 destroy_workqueue(wq);
560 return 0;
561}
562
563static int dio_set_defer_completion(struct dio *dio)
564{
565 struct super_block *sb = dio->inode->i_sb;
566
567 if (dio->defer_completion)
568 return 0;
569 dio->defer_completion = true;
570 if (!sb->s_dio_done_wq)
571 return sb_init_dio_done_wq(sb);
572 return 0;
573}
574
575/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576 * Call into the fs to map some more disk blocks. We record the current number
Andi Kleeneb28be22011-08-01 21:38:03 -0700577 * of available blocks at sdio->blocks_available. These are in units of the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578 * fs blocksize, (1 << inode->i_blkbits).
579 *
580 * The fs is allowed to map lots of blocks at once. If it wants to do that,
581 * it uses the passed inode-relative block number as the file offset, as usual.
582 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800583 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 * has remaining to do. The fs should not map more than this number of blocks.
585 *
586 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
587 * indicate how much contiguous disk space has been made available at
588 * bh->b_blocknr.
589 *
590 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
591 * This isn't very efficient...
592 *
593 * In the case of filesystem holes: the fs may return an arbitrarily-large
594 * hole by returning an appropriate value in b_size and by clearing
595 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800596 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597 */
Andi Kleen18772642011-08-01 21:38:07 -0700598static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
599 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600{
601 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
Tao Maae55e1a2012-01-12 17:20:33 -0800603 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604 unsigned long fs_count; /* Number of filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605 int create;
Linus Torvaldsab738572012-11-29 12:27:00 -0800606 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607
608 /*
609 * If there was a memory error and we've overwritten all the
610 * mapped blocks then we can now return that memory error
611 */
612 ret = dio->page_errors;
613 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700614 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
615 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
Tao Maae55e1a2012-01-12 17:20:33 -0800616 fs_endblk = (sdio->final_block_in_request - 1) >>
617 sdio->blkfactor;
618 fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619
Nathan Scott3c674e72006-03-29 09:26:15 +1000620 map_bh->b_state = 0;
Linus Torvaldsab738572012-11-29 12:27:00 -0800621 map_bh->b_size = fs_count << i_blkbits;
Nathan Scott3c674e72006-03-29 09:26:15 +1000622
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800623 /*
624 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
625 * forbid block creations: only overwrites are permitted.
626 * We will return early to the caller once we see an
627 * unmapped buffer head returned, and the caller will fall
628 * back to buffered I/O.
629 *
630 * Otherwise the decision is left to the get_blocks method,
631 * which may decide to handle it or also return an unmapped
632 * buffer head.
633 */
Jens Axboeb31dc662006-06-13 08:26:10 +0200634 create = dio->rw & WRITE;
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800635 if (dio->flags & DIO_SKIP_HOLES) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700636 if (sdio->block_in_file < (i_size_read(dio->inode) >>
637 sdio->blkbits))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000640
Andi Kleeneb28be22011-08-01 21:38:03 -0700641 ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642 map_bh, create);
Andi Kleen18772642011-08-01 21:38:07 -0700643
644 /* Store for completion */
645 dio->private = map_bh->b_private;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200646
647 if (ret == 0 && buffer_defer_completion(map_bh))
648 ret = dio_set_defer_completion(dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 }
650 return ret;
651}
652
653/*
654 * There is no bio. Make one now.
655 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700656static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
657 sector_t start_sector, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658{
659 sector_t sector;
660 int ret, nr_pages;
661
Andi Kleeneb28be22011-08-01 21:38:03 -0700662 ret = dio_bio_reap(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663 if (ret)
664 goto out;
Andi Kleeneb28be22011-08-01 21:38:03 -0700665 sector = start_sector << (sdio->blkbits - 9);
Andi Kleen18772642011-08-01 21:38:07 -0700666 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 BUG_ON(nr_pages <= 0);
Andi Kleen18772642011-08-01 21:38:07 -0700668 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
Andi Kleeneb28be22011-08-01 21:38:03 -0700669 sdio->boundary = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670out:
671 return ret;
672}
673
674/*
675 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
676 * that was successful then update final_block_in_bio and take a ref against
677 * the just-added page.
678 *
679 * Return zero on success. Non-zero means the caller needs to start a new BIO.
680 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700681static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682{
683 int ret;
684
Andi Kleeneb28be22011-08-01 21:38:03 -0700685 ret = bio_add_page(sdio->bio, sdio->cur_page,
686 sdio->cur_page_len, sdio->cur_page_offset);
687 if (ret == sdio->cur_page_len) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 /*
689 * Decrement count only, if we are done with this page
690 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700691 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
692 sdio->pages_in_io--;
693 page_cache_get(sdio->cur_page);
694 sdio->final_block_in_bio = sdio->cur_page_block +
695 (sdio->cur_page_len >> sdio->blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696 ret = 0;
697 } else {
698 ret = 1;
699 }
700 return ret;
701}
702
703/*
704 * Put cur_page under IO. The section of cur_page which is described by
705 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
706 * starts on-disk at cur_page_block.
707 *
708 * We take a ref against the page here (on behalf of its presence in the bio).
709 *
710 * The caller of this function is responsible for removing cur_page from the
711 * dio, and for dropping the refcount which came from that presence.
712 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700713static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
714 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715{
716 int ret = 0;
717
Andi Kleeneb28be22011-08-01 21:38:03 -0700718 if (sdio->bio) {
719 loff_t cur_offset = sdio->cur_page_fs_offset;
720 loff_t bio_next_offset = sdio->logical_offset_in_bio +
Kent Overstreet4f024f32013-10-11 15:44:27 -0700721 sdio->bio->bi_iter.bi_size;
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400722
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400724 * See whether this new request is contiguous with the old.
725 *
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900726 * Btrfs cannot handle having logically non-contiguous requests
727 * submitted. For example if you have
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400728 *
729 * Logical: [0-4095][HOLE][8192-12287]
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900730 * Physical: [0-4095] [4096-8191]
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400731 *
732 * We cannot submit those pages together as one BIO. So if our
733 * current logical offset in the file does not equal what would
734 * be the next logical offset in the bio, submit the bio we
735 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700737 if (sdio->final_block_in_bio != sdio->cur_page_block ||
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400738 cur_offset != bio_next_offset)
Andi Kleeneb28be22011-08-01 21:38:03 -0700739 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 }
741
Andi Kleeneb28be22011-08-01 21:38:03 -0700742 if (sdio->bio == NULL) {
Andi Kleen18772642011-08-01 21:38:07 -0700743 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 if (ret)
745 goto out;
746 }
747
Andi Kleeneb28be22011-08-01 21:38:03 -0700748 if (dio_bio_add_page(sdio) != 0) {
749 dio_bio_submit(dio, sdio);
Andi Kleen18772642011-08-01 21:38:07 -0700750 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700752 ret = dio_bio_add_page(sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 BUG_ON(ret != 0);
754 }
755 }
756out:
757 return ret;
758}
759
760/*
761 * An autonomous function to put a chunk of a page under deferred IO.
762 *
763 * The caller doesn't actually know (or care) whether this piece of page is in
764 * a BIO, or is under IO or whatever. We just take care of all possible
765 * situations here. The separation between the logic of do_direct_IO() and
766 * that of submit_page_section() is important for clarity. Please don't break.
767 *
768 * The chunk of page starts on-disk at blocknr.
769 *
770 * We perform deferred IO, by recording the last-submitted page inside our
771 * private part of the dio structure. If possible, we just expand the IO
772 * across that page here.
773 *
774 * If that doesn't work out then we put the old page into the bio and add this
775 * page to the dio instead.
776 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700777static inline int
Andi Kleeneb28be22011-08-01 21:38:03 -0700778submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
Andi Kleen18772642011-08-01 21:38:07 -0700779 unsigned offset, unsigned len, sector_t blocknr,
780 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781{
782 int ret = 0;
783
Andrew Morton98c4d572006-12-10 02:19:47 -0800784 if (dio->rw & WRITE) {
785 /*
786 * Read accounting is performed in submit_bio()
787 */
788 task_io_account_write(len);
789 }
790
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 /*
792 * Can we just grow the current page's presence in the dio?
793 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700794 if (sdio->cur_page == page &&
795 sdio->cur_page_offset + sdio->cur_page_len == offset &&
796 sdio->cur_page_block +
797 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
798 sdio->cur_page_len += len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 goto out;
800 }
801
802 /*
803 * If there's a deferred page already there then send it.
804 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700805 if (sdio->cur_page) {
Andi Kleen18772642011-08-01 21:38:07 -0700806 ret = dio_send_cur_page(dio, sdio, map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -0700807 page_cache_release(sdio->cur_page);
808 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809 if (ret)
Jan Karab1058b92013-04-29 15:06:18 -0700810 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 }
812
813 page_cache_get(page); /* It is in dio */
Andi Kleeneb28be22011-08-01 21:38:03 -0700814 sdio->cur_page = page;
815 sdio->cur_page_offset = offset;
816 sdio->cur_page_len = len;
817 sdio->cur_page_block = blocknr;
818 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819out:
Jan Karab1058b92013-04-29 15:06:18 -0700820 /*
821 * If sdio->boundary then we want to schedule the IO now to
822 * avoid metadata seeks.
823 */
824 if (sdio->boundary) {
825 ret = dio_send_cur_page(dio, sdio, map_bh);
826 dio_bio_submit(dio, sdio);
827 page_cache_release(sdio->cur_page);
828 sdio->cur_page = NULL;
829 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 return ret;
831}
832
833/*
834 * Clean any dirty buffers in the blockdev mapping which alias newly-created
835 * file blocks. Only called for S_ISREG files - blockdevs do not set
836 * buffer_new
837 */
Andi Kleen18772642011-08-01 21:38:07 -0700838static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839{
840 unsigned i;
841 unsigned nblocks;
842
Andi Kleen18772642011-08-01 21:38:07 -0700843 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844
845 for (i = 0; i < nblocks; i++) {
Andi Kleen18772642011-08-01 21:38:07 -0700846 unmap_underlying_metadata(map_bh->b_bdev,
847 map_bh->b_blocknr + i);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848 }
849}
850
851/*
852 * If we are not writing the entire block and get_block() allocated
853 * the block for us, we need to fill-in the unused portion of the
854 * block with zeros. This happens only if user-buffer, fileoffset or
855 * io length is not filesystem block-size multiple.
856 *
857 * `end' is zero if we're doing the start of the IO, 1 at the end of the
858 * IO.
859 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700860static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
861 int end, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862{
863 unsigned dio_blocks_per_fs_block;
864 unsigned this_chunk_blocks; /* In dio_blocks */
865 unsigned this_chunk_bytes;
866 struct page *page;
867
Andi Kleeneb28be22011-08-01 21:38:03 -0700868 sdio->start_zero_done = 1;
Andi Kleen18772642011-08-01 21:38:07 -0700869 if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870 return;
871
Andi Kleeneb28be22011-08-01 21:38:03 -0700872 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
873 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874
875 if (!this_chunk_blocks)
876 return;
877
878 /*
879 * We need to zero out part of an fs block. It is either at the
880 * beginning or the end of the fs block.
881 */
882 if (end)
883 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
884
Andi Kleeneb28be22011-08-01 21:38:03 -0700885 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886
Nick Piggin557ed1f2007-10-16 01:24:40 -0700887 page = ZERO_PAGE(0);
Andi Kleeneb28be22011-08-01 21:38:03 -0700888 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700889 sdio->next_block_for_io, map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 return;
891
Andi Kleeneb28be22011-08-01 21:38:03 -0700892 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893}
894
895/*
896 * Walk the user pages, and the file, mapping blocks to disk and generating
897 * a sequence of (page,offset,len,block) mappings. These mappings are injected
898 * into submit_page_section(), which takes care of the next stage of submission
899 *
900 * Direct IO against a blockdev is different from a file. Because we can
901 * happily perform page-sized but 512-byte aligned IOs. It is important that
902 * blockdev IO be able to have fine alignment and large sizes.
903 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800904 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905 * with the size of IO which is permitted at this offset and this i_blkbits.
906 *
907 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800908 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 * fine alignment but still allows this function to work in PAGE_SIZE units.
910 */
Andi Kleen18772642011-08-01 21:38:07 -0700911static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
912 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913{
Andi Kleeneb28be22011-08-01 21:38:03 -0700914 const unsigned blkbits = sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
916 struct page *page;
917 unsigned block_in_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 int ret = 0;
919
920 /* The I/O can start at any block offset within the first page */
Andi Kleeneb28be22011-08-01 21:38:03 -0700921 block_in_page = sdio->first_block_in_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922
Andi Kleeneb28be22011-08-01 21:38:03 -0700923 while (sdio->block_in_file < sdio->final_block_in_request) {
924 page = dio_get_page(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925 if (IS_ERR(page)) {
926 ret = PTR_ERR(page);
927 goto out;
928 }
929
930 while (block_in_page < blocks_per_page) {
931 unsigned offset_in_page = block_in_page << blkbits;
932 unsigned this_chunk_bytes; /* # of bytes mapped */
933 unsigned this_chunk_blocks; /* # of blocks */
934 unsigned u;
935
Andi Kleeneb28be22011-08-01 21:38:03 -0700936 if (sdio->blocks_available == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937 /*
938 * Need to go and map some more disk
939 */
940 unsigned long blkmask;
941 unsigned long dio_remainder;
942
Andi Kleen18772642011-08-01 21:38:07 -0700943 ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944 if (ret) {
945 page_cache_release(page);
946 goto out;
947 }
948 if (!buffer_mapped(map_bh))
949 goto do_holes;
950
Andi Kleeneb28be22011-08-01 21:38:03 -0700951 sdio->blocks_available =
952 map_bh->b_size >> sdio->blkbits;
953 sdio->next_block_for_io =
954 map_bh->b_blocknr << sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955 if (buffer_new(map_bh))
Andi Kleen18772642011-08-01 21:38:07 -0700956 clean_blockdev_aliases(dio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957
Andi Kleeneb28be22011-08-01 21:38:03 -0700958 if (!sdio->blkfactor)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 goto do_holes;
960
Andi Kleeneb28be22011-08-01 21:38:03 -0700961 blkmask = (1 << sdio->blkfactor) - 1;
962 dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963
964 /*
965 * If we are at the start of IO and that IO
966 * starts partway into a fs-block,
967 * dio_remainder will be non-zero. If the IO
968 * is a read then we can simply advance the IO
969 * cursor to the first block which is to be
970 * read. But if the IO is a write and the
971 * block was newly allocated we cannot do that;
972 * the start of the fs block must be zeroed out
973 * on-disk
974 */
975 if (!buffer_new(map_bh))
Andi Kleeneb28be22011-08-01 21:38:03 -0700976 sdio->next_block_for_io += dio_remainder;
977 sdio->blocks_available -= dio_remainder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978 }
979do_holes:
980 /* Handle holes */
981 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800982 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983
984 /* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboeb31dc662006-06-13 08:26:10 +0200985 if (dio->rw & WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 page_cache_release(page);
987 return -ENOTBLK;
988 }
989
Jeff Moyer35dc8162006-02-03 03:04:27 -0800990 /*
991 * Be sure to account for a partial block as the
992 * last block in the file
993 */
994 i_size_aligned = ALIGN(i_size_read(dio->inode),
995 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -0700996 if (sdio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -0800997 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998 /* We hit eof */
999 page_cache_release(page);
1000 goto out;
1001 }
Christoph Lametereebd2aa2008-02-04 22:28:29 -08001002 zero_user(page, block_in_page << blkbits,
1003 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -07001004 sdio->block_in_file++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005 block_in_page++;
1006 goto next_block;
1007 }
1008
1009 /*
1010 * If we're performing IO which has an alignment which
1011 * is finer than the underlying fs, go check to see if
1012 * we must zero out the start of this block.
1013 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001014 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
Andi Kleen18772642011-08-01 21:38:07 -07001015 dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001016
1017 /*
1018 * Work out, in this_chunk_blocks, how much disk we
1019 * can add to this page
1020 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001021 this_chunk_blocks = sdio->blocks_available;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022 u = (PAGE_SIZE - offset_in_page) >> blkbits;
1023 if (this_chunk_blocks > u)
1024 this_chunk_blocks = u;
Andi Kleeneb28be22011-08-01 21:38:03 -07001025 u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026 if (this_chunk_blocks > u)
1027 this_chunk_blocks = u;
1028 this_chunk_bytes = this_chunk_blocks << blkbits;
1029 BUG_ON(this_chunk_bytes == 0);
1030
Jan Kara092c8d42013-04-29 15:06:17 -07001031 if (this_chunk_blocks == sdio->blocks_available)
1032 sdio->boundary = buffer_boundary(map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -07001033 ret = submit_page_section(dio, sdio, page,
1034 offset_in_page,
1035 this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -07001036 sdio->next_block_for_io,
1037 map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 if (ret) {
1039 page_cache_release(page);
1040 goto out;
1041 }
Andi Kleeneb28be22011-08-01 21:38:03 -07001042 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043
Andi Kleeneb28be22011-08-01 21:38:03 -07001044 sdio->block_in_file += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045 block_in_page += this_chunk_blocks;
Andi Kleeneb28be22011-08-01 21:38:03 -07001046 sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047next_block:
Andi Kleeneb28be22011-08-01 21:38:03 -07001048 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1049 if (sdio->block_in_file == sdio->final_block_in_request)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050 break;
1051 }
1052
1053 /* Drop the ref which was taken in get_user_pages() */
1054 page_cache_release(page);
1055 block_in_page = 0;
1056 }
1057out:
1058 return ret;
1059}
1060
Andi Kleen847cc632011-08-01 21:38:09 -07001061static inline int drop_refcount(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062{
Andi Kleen847cc632011-08-01 21:38:09 -07001063 int ret2;
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001064 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065
Zach Brown8459d862006-12-10 02:21:05 -08001066 /*
1067 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001068 * operation. AIO can if it was a broken operation described above or
1069 * in fact if all the bios race to complete before we get here. In
1070 * that case dio_complete() translates the EIOCBQUEUED into the proper
1071 * return code that the caller will hand to aio_complete().
1072 *
1073 * This is managed by the bio_lock instead of being an atomic_t so that
1074 * completion paths can drop their ref and use the remaining count to
1075 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001076 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001077 spin_lock_irqsave(&dio->bio_lock, flags);
1078 ret2 = --dio->refcount;
1079 spin_unlock_irqrestore(&dio->bio_lock, flags);
Andi Kleen847cc632011-08-01 21:38:09 -07001080 return ret2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081}
1082
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001083/*
1084 * This is a library function for use by filesystem drivers.
1085 *
1086 * The locking rules are governed by the flags parameter:
1087 * - if the flags value contains DIO_LOCKING we use a fancy locking
1088 * scheme for dumb filesystems.
1089 * For writes this function is called under i_mutex and returns with
1090 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1091 * taken and dropped again before returning.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001092 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1093 * internal locking but rather rely on the filesystem to synchronize
1094 * direct I/O reads/writes versus each other and truncate.
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001095 *
1096 * To help with locking against truncate we incremented the i_dio_count
1097 * counter before starting direct I/O, and decrement it once we are done.
1098 * Truncate can wait for it to reach zero to provide exclusion. It is
1099 * expected that filesystem provide exclusion between new direct I/O
1100 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1101 * but other filesystems need to take care of this on their own.
Andi Kleenba253fb2011-08-01 21:38:08 -07001102 *
1103 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1104 * is always inlined. Otherwise gcc is unable to split the structure into
1105 * individual fields and will generate much worse code. This is important
1106 * for the whole file.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001107 */
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001108static inline ssize_t
1109do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 struct block_device *bdev, const struct iovec *iov, loff_t offset,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -08001111 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
Josef Bacikfacd07b2010-05-23 11:00:55 -04001112 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113{
1114 int seg;
1115 size_t size;
1116 unsigned long addr;
Linus Torvaldsab738572012-11-29 12:27:00 -08001117 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1118 unsigned blkbits = i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119 unsigned blocksize_mask = (1 << blkbits) - 1;
1120 ssize_t retval = -EINVAL;
1121 loff_t end = offset;
1122 struct dio *dio;
Andi Kleeneb28be22011-08-01 21:38:03 -07001123 struct dio_submit sdio = { 0, };
Andi Kleen847cc632011-08-01 21:38:09 -07001124 unsigned long user_addr;
1125 size_t bytes;
1126 struct buffer_head map_bh = { 0, };
Fengguang Wu647d1e42012-08-09 15:23:09 +02001127 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128
1129 if (rw & WRITE)
Jens Axboe721a9602011-03-09 11:56:30 +01001130 rw = WRITE_ODIRECT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001132 /*
1133 * Avoid references to bdev if not absolutely needed to give
1134 * the early prefetch in the caller enough time.
1135 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136
1137 if (offset & blocksize_mask) {
1138 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001139 blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 blocksize_mask = (1 << blkbits) - 1;
1141 if (offset & blocksize_mask)
1142 goto out;
1143 }
1144
1145 /* Check the memory alignment. Blocks cannot straddle pages */
1146 for (seg = 0; seg < nr_segs; seg++) {
1147 addr = (unsigned long)iov[seg].iov_base;
1148 size = iov[seg].iov_len;
1149 end += size;
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001150 if (unlikely((addr & blocksize_mask) ||
1151 (size & blocksize_mask))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001153 blkbits = blksize_bits(
1154 bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155 blocksize_mask = (1 << blkbits) - 1;
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001156 if ((addr & blocksize_mask) || (size & blocksize_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 goto out;
1158 }
1159 }
1160
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001161 /* watch out for a 0 len io from a tricksy fs */
1162 if (rw == READ && end == offset)
1163 return 0;
1164
Andi Kleen6e8267f2011-08-01 21:38:06 -07001165 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166 retval = -ENOMEM;
1167 if (!dio)
1168 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001169 /*
1170 * Believe it or not, zeroing out the page array caused a .5%
1171 * performance regression in a database benchmark. So, we take
1172 * care to only zero out what's needed.
1173 */
1174 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001176 dio->flags = flags;
1177 if (dio->flags & DIO_LOCKING) {
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001178 if (rw == READ) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001179 struct address_space *mapping =
1180 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001182 /* will be released by direct_io_worker */
1183 mutex_lock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184
1185 retval = filemap_write_and_wait_range(mapping, offset,
1186 end - 1);
1187 if (retval) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001188 mutex_unlock(&inode->i_mutex);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001189 kmem_cache_free(dio_cache, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190 goto out;
1191 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 }
1194
1195 /*
Christoph Hellwig60392572014-02-10 10:27:11 +11001196 * For file extending writes updating i_size before data writeouts
1197 * complete can expose uninitialized blocks in dumb filesystems.
1198 * In that case we need to wait for I/O completion even if asked
1199 * for an asynchronous write.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200 */
Christoph Hellwig60392572014-02-10 10:27:11 +11001201 if (is_sync_kiocb(iocb))
1202 dio->is_async = false;
1203 else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
1204 (rw & WRITE) && end > i_size_read(inode))
1205 dio->is_async = false;
1206 else
1207 dio->is_async = true;
1208
Andi Kleen847cc632011-08-01 21:38:09 -07001209 dio->inode = inode;
1210 dio->rw = rw;
Christoph Hellwig02afc272013-09-04 15:04:40 +02001211
1212 /*
1213 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1214 * so that we can call ->fsync.
1215 */
1216 if (dio->is_async && (rw & WRITE) &&
1217 ((iocb->ki_filp->f_flags & O_DSYNC) ||
1218 IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1219 retval = dio_set_defer_completion(dio);
1220 if (retval) {
1221 /*
1222 * We grab i_mutex only for reads so we don't have
1223 * to release it here
1224 */
1225 kmem_cache_free(dio_cache, dio);
1226 goto out;
1227 }
1228 }
1229
1230 /*
1231 * Will be decremented at I/O completion time.
1232 */
1233 atomic_inc(&inode->i_dio_count);
1234
1235 retval = 0;
Andi Kleen847cc632011-08-01 21:38:09 -07001236 sdio.blkbits = blkbits;
Linus Torvaldsab738572012-11-29 12:27:00 -08001237 sdio.blkfactor = i_blkbits - blkbits;
Andi Kleen847cc632011-08-01 21:38:09 -07001238 sdio.block_in_file = offset >> blkbits;
1239
1240 sdio.get_block = get_block;
1241 dio->end_io = end_io;
1242 sdio.submit_io = submit_io;
1243 sdio.final_block_in_bio = -1;
1244 sdio.next_block_for_io = -1;
1245
1246 dio->iocb = iocb;
1247 dio->i_size = i_size_read(inode);
1248
1249 spin_lock_init(&dio->bio_lock);
1250 dio->refcount = 1;
1251
1252 /*
1253 * In case of non-aligned buffers, we may need 2 more
1254 * pages since we need to zero out first and last block.
1255 */
1256 if (unlikely(sdio.blkfactor))
1257 sdio.pages_in_io = 2;
1258
1259 for (seg = 0; seg < nr_segs; seg++) {
1260 user_addr = (unsigned long)iov[seg].iov_base;
1261 sdio.pages_in_io +=
1262 ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) /
1263 PAGE_SIZE - user_addr / PAGE_SIZE);
1264 }
1265
Fengguang Wu647d1e42012-08-09 15:23:09 +02001266 blk_start_plug(&plug);
1267
Andi Kleen847cc632011-08-01 21:38:09 -07001268 for (seg = 0; seg < nr_segs; seg++) {
1269 user_addr = (unsigned long)iov[seg].iov_base;
1270 sdio.size += bytes = iov[seg].iov_len;
1271
1272 /* Index into the first page of the first block */
1273 sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1274 sdio.final_block_in_request = sdio.block_in_file +
1275 (bytes >> blkbits);
1276 /* Page fetching state */
1277 sdio.head = 0;
1278 sdio.tail = 0;
1279 sdio.curr_page = 0;
1280
1281 sdio.total_pages = 0;
1282 if (user_addr & (PAGE_SIZE-1)) {
1283 sdio.total_pages++;
1284 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1285 }
1286 sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1287 sdio.curr_user_address = user_addr;
1288
1289 retval = do_direct_IO(dio, &sdio, &map_bh);
1290
1291 dio->result += iov[seg].iov_len -
1292 ((sdio.final_block_in_request - sdio.block_in_file) <<
1293 blkbits);
1294
1295 if (retval) {
1296 dio_cleanup(dio, &sdio);
1297 break;
1298 }
1299 } /* end iovec loop */
1300
1301 if (retval == -ENOTBLK) {
1302 /*
1303 * The remaining part of the request will be
1304 * be handled by buffered I/O when we return
1305 */
1306 retval = 0;
1307 }
1308 /*
1309 * There may be some unwritten disk at the end of a part-written
1310 * fs-block-sized block. Go zero that now.
1311 */
1312 dio_zero_block(dio, &sdio, 1, &map_bh);
1313
1314 if (sdio.cur_page) {
1315 ssize_t ret2;
1316
1317 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1318 if (retval == 0)
1319 retval = ret2;
1320 page_cache_release(sdio.cur_page);
1321 sdio.cur_page = NULL;
1322 }
1323 if (sdio.bio)
1324 dio_bio_submit(dio, &sdio);
1325
Fengguang Wu647d1e42012-08-09 15:23:09 +02001326 blk_finish_plug(&plug);
1327
Andi Kleen847cc632011-08-01 21:38:09 -07001328 /*
1329 * It is possible that, we return short IO due to end of file.
1330 * In that case, we need to release all the pages we got hold on.
1331 */
1332 dio_cleanup(dio, &sdio);
1333
1334 /*
1335 * All block lookups have been performed. For READ requests
1336 * we can let i_mutex go now that its achieved its purpose
1337 * of protecting us from looking up uninitialized blocks.
1338 */
1339 if (rw == READ && (dio->flags & DIO_LOCKING))
1340 mutex_unlock(&dio->inode->i_mutex);
1341
1342 /*
1343 * The only time we want to leave bios in flight is when a successful
1344 * partial aio read or full aio write have been setup. In that case
1345 * bio completion will call aio_complete. The only time it's safe to
1346 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1347 * This had *better* be the only place that raises -EIOCBQUEUED.
1348 */
1349 BUG_ON(retval == -EIOCBQUEUED);
1350 if (dio->is_async && retval == 0 && dio->result &&
Julia Lawalld1876632012-06-07 15:45:00 -07001351 ((rw == READ) || (dio->result == sdio.size)))
Andi Kleen847cc632011-08-01 21:38:09 -07001352 retval = -EIOCBQUEUED;
1353
1354 if (retval != -EIOCBQUEUED)
1355 dio_await_completion(dio);
1356
1357 if (drop_refcount(dio) == 0) {
1358 retval = dio_complete(dio, offset, retval, false);
Andi Kleen847cc632011-08-01 21:38:09 -07001359 } else
1360 BUG_ON(retval != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001362out:
1363 return retval;
1364}
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001365
1366ssize_t
1367__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1368 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1369 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1370 dio_submit_t submit_io, int flags)
1371{
1372 /*
1373 * The block device state is needed in the end to finally
1374 * submit everything. Since it's likely to be cache cold
1375 * prefetch it here as first thing to hide some of the
1376 * latency.
1377 *
1378 * Attempt to prefetch the pieces we likely need later.
1379 */
1380 prefetch(&bdev->bd_disk->part_tbl);
1381 prefetch(bdev->bd_queue);
1382 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1383
1384 return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset,
1385 nr_segs, get_block, end_io,
1386 submit_io, flags);
1387}
1388
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389EXPORT_SYMBOL(__blockdev_direct_IO);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001390
1391static __init int dio_init(void)
1392{
1393 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1394 return 0;
1395}
1396module_init(dio_init)