blob: 6fb00e3f1059791d21b4ffc80671f3d051ecbc8e [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>
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 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070073 sector_t block_in_file; /* Current offset into the underlying
74 file in dio_block units. */
75 unsigned blocks_available; /* At block_in_file. changes */
Andi Kleen0dc2bc42011-08-01 21:38:05 -070076 int reap_counter; /* rate limit reaping */
Linus Torvalds1da177e2005-04-16 15:20:36 -070077 sector_t final_block_in_request;/* doesn't change */
Linus Torvalds1da177e2005-04-16 15:20:36 -070078 int boundary; /* prev block is at a boundary */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080079 get_block_t *get_block; /* block mapping function */
Josef Bacikfacd07b2010-05-23 11:00:55 -040080 dio_submit_t *submit_io; /* IO submition function */
Andi Kleeneb28be22011-08-01 21:38:03 -070081
Josef Bacikfacd07b2010-05-23 11:00:55 -040082 loff_t logical_offset_in_bio; /* current first logical block in bio */
Linus Torvalds1da177e2005-04-16 15:20:36 -070083 sector_t final_block_in_bio; /* current final block in bio + 1 */
84 sector_t next_block_for_io; /* next block to be put under IO,
85 in dio_blocks units */
Linus Torvalds1da177e2005-04-16 15:20:36 -070086
87 /*
88 * Deferred addition of a page to the dio. These variables are
89 * private to dio_send_cur_page(), submit_page_section() and
90 * dio_bio_add_page().
91 */
92 struct page *cur_page; /* The page */
93 unsigned cur_page_offset; /* Offset into it, in bytes */
94 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
95 sector_t cur_page_block; /* Where it starts */
Josef Bacikfacd07b2010-05-23 11:00:55 -040096 loff_t cur_page_fs_offset; /* Offset in file */
Linus Torvalds1da177e2005-04-16 15:20:36 -070097
Al Viro7b2c99d2014-03-15 04:05:57 -040098 struct iov_iter *iter;
Jeff Moyer23aee092009-12-15 16:47:49 -080099 /*
100 * Page queue. These variables belong to dio_refill_pages() and
101 * dio_get_page().
102 */
103 unsigned head; /* next page to process */
104 unsigned tail; /* last valid page + 1 */
Al Viro7b2c99d2014-03-15 04:05:57 -0400105 size_t from, to;
Andi Kleeneb28be22011-08-01 21:38:03 -0700106};
107
108/* dio_state communicated between submission path and end_io */
109struct dio {
110 int flags; /* doesn't change */
Andi Kleeneb28be22011-08-01 21:38:03 -0700111 int rw;
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700112 struct inode *inode;
Andi Kleeneb28be22011-08-01 21:38:03 -0700113 loff_t i_size; /* i_size when submitted */
114 dio_iodone_t *end_io; /* IO completion function */
Andi Kleeneb28be22011-08-01 21:38:03 -0700115
Andi Kleen18772642011-08-01 21:38:07 -0700116 void *private; /* copy from map_bh.b_private */
Andi Kleeneb28be22011-08-01 21:38:03 -0700117
118 /* BIO completion state */
119 spinlock_t bio_lock; /* protects BIO fields below */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700120 int page_errors; /* errno from get_user_pages() */
121 int is_async; /* is IO async ? */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200122 bool defer_completion; /* defer AIO completion to workqueue? */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700123 int io_error; /* IO error in completion path */
Andi Kleeneb28be22011-08-01 21:38:03 -0700124 unsigned long refcount; /* direct_io_worker() and bios */
125 struct bio *bio_list; /* singly linked via bi_private */
126 struct task_struct *waiter; /* waiting task (NULL if none) */
127
128 /* AIO related stuff */
129 struct kiocb *iocb; /* kiocb */
Andi Kleeneb28be22011-08-01 21:38:03 -0700130 ssize_t result; /* IO result */
131
Jeff Moyer23aee092009-12-15 16:47:49 -0800132 /*
133 * pages[] (and any fields placed after it) are not zeroed out at
134 * allocation time. Don't add new fields after pages[] unless you
135 * wish that they not be zeroed.
136 */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200137 union {
138 struct page *pages[DIO_PAGES]; /* page buffer */
139 struct work_struct complete_work;/* deferred AIO completion */
140 };
Andi Kleen6e8267f2011-08-01 21:38:06 -0700141} ____cacheline_aligned_in_smp;
142
143static struct kmem_cache *dio_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144
145/*
146 * How many pages are in the queue?
147 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700148static inline unsigned dio_pages_present(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149{
Andi Kleeneb28be22011-08-01 21:38:03 -0700150 return sdio->tail - sdio->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151}
152
153/*
154 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
155 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700156static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157{
Al Viro7b2c99d2014-03-15 04:05:57 -0400158 ssize_t ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159
Miklos Szeredi2c809292014-09-24 17:09:11 +0200160 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
Al Viro7b2c99d2014-03-15 04:05:57 -0400161 &sdio->from);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162
Andi Kleeneb28be22011-08-01 21:38:03 -0700163 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
Nick Piggin557ed1f2007-10-16 01:24:40 -0700164 struct page *page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165 /*
166 * A memory fault, but the filesystem has some outstanding
167 * mapped blocks. We need to use those blocks up to avoid
168 * leaking stale data in the file.
169 */
170 if (dio->page_errors == 0)
171 dio->page_errors = ret;
Nick Pigginb5810032005-10-29 18:16:12 -0700172 page_cache_get(page);
173 dio->pages[0] = page;
Andi Kleeneb28be22011-08-01 21:38:03 -0700174 sdio->head = 0;
175 sdio->tail = 1;
Al Viro7b2c99d2014-03-15 04:05:57 -0400176 sdio->from = 0;
177 sdio->to = PAGE_SIZE;
178 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179 }
180
181 if (ret >= 0) {
Al Viro7b2c99d2014-03-15 04:05:57 -0400182 iov_iter_advance(sdio->iter, ret);
183 ret += sdio->from;
Andi Kleeneb28be22011-08-01 21:38:03 -0700184 sdio->head = 0;
Al Viro7b2c99d2014-03-15 04:05:57 -0400185 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
186 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
187 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189 return ret;
190}
191
192/*
193 * Get another userspace page. Returns an ERR_PTR on error. Pages are
194 * buffered inside the dio so that we can call get_user_pages() against a
195 * decent number of pages, less frequently. To provide nicer use of the
196 * L1 cache.
197 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700198static inline struct page *dio_get_page(struct dio *dio,
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300199 struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200{
Andi Kleeneb28be22011-08-01 21:38:03 -0700201 if (dio_pages_present(sdio) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202 int ret;
203
Andi Kleeneb28be22011-08-01 21:38:03 -0700204 ret = dio_refill_pages(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205 if (ret)
206 return ERR_PTR(ret);
Andi Kleeneb28be22011-08-01 21:38:03 -0700207 BUG_ON(dio_pages_present(sdio) == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208 }
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300209 return dio->pages[sdio->head];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210}
211
Zach Brown6d544bb2006-12-10 02:20:54 -0800212/**
213 * dio_complete() - called when all DIO BIO I/O has been completed
214 * @offset: the byte offset in the file of the completed operation
215 *
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200216 * This drops i_dio_count, lets interested parties know that a DIO operation
217 * has completed, and calculates the resulting return code for the operation.
Zach Brown6d544bb2006-12-10 02:20:54 -0800218 *
219 * It lets the filesystem know if it registered an interest earlier via
220 * get_block. Pass the private field of the map buffer_head so that
221 * filesystems can use it to hold additional state between get_block calls and
222 * dio_complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223 */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200224static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret,
225 bool is_async)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226{
Zach Brown6d544bb2006-12-10 02:20:54 -0800227 ssize_t transferred = 0;
228
Zach Brown8459d862006-12-10 02:21:05 -0800229 /*
230 * AIO submission can race with bio completion to get here while
231 * expecting to have the last io completed by bio completion.
232 * In that case -EIOCBQUEUED is in fact not an error we want
233 * to preserve through this call.
234 */
235 if (ret == -EIOCBQUEUED)
236 ret = 0;
237
Zach Brown6d544bb2006-12-10 02:20:54 -0800238 if (dio->result) {
239 transferred = dio->result;
240
241 /* Check for short read case */
242 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
243 transferred = dio->i_size - offset;
244 }
245
Zach Brown6d544bb2006-12-10 02:20:54 -0800246 if (ret == 0)
247 ret = dio->page_errors;
248 if (ret == 0)
249 ret = dio->io_error;
250 if (ret == 0)
251 ret = transferred;
252
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200253 if (dio->end_io && dio->result)
254 dio->end_io(dio->iocb, offset, transferred, dio->private);
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000255
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200256 inode_dio_done(dio->inode);
Christoph Hellwig02afc272013-09-04 15:04:40 +0200257 if (is_async) {
258 if (dio->rw & WRITE) {
259 int err;
260
261 err = generic_write_sync(dio->iocb->ki_filp, offset,
262 transferred);
263 if (err < 0 && ret > 0)
264 ret = err;
265 }
266
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100267 dio->iocb->ki_complete(dio->iocb, ret, 0);
Christoph Hellwig02afc272013-09-04 15:04:40 +0200268 }
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200269
270 kmem_cache_free(dio_cache, dio);
Zach Brown6d544bb2006-12-10 02:20:54 -0800271 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272}
273
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200274static void dio_aio_complete_work(struct work_struct *work)
275{
276 struct dio *dio = container_of(work, struct dio, complete_work);
277
278 dio_complete(dio, dio->iocb->ki_pos, 0, true);
279}
280
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281static int dio_bio_complete(struct dio *dio, struct bio *bio);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200282
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283/*
284 * Asynchronous IO callback.
285 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200286static void dio_bio_end_aio(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287{
288 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800289 unsigned long remaining;
290 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 /* cleanup the bio */
293 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800294
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800295 spin_lock_irqsave(&dio->bio_lock, flags);
296 remaining = --dio->refcount;
297 if (remaining == 1 && dio->waiter)
Zach Brown20258b22006-12-10 02:21:01 -0800298 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800299 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b22006-12-10 02:21:01 -0800300
Zach Brown8459d862006-12-10 02:21:05 -0800301 if (remaining == 0) {
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200302 if (dio->result && dio->defer_completion) {
303 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
304 queue_work(dio->inode->i_sb->s_dio_done_wq,
305 &dio->complete_work);
306 } else {
307 dio_complete(dio, dio->iocb->ki_pos, 0, true);
308 }
Zach Brown8459d862006-12-10 02:21:05 -0800309 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310}
311
312/*
313 * The BIO completion handler simply queues the BIO up for the process-context
314 * handler.
315 *
316 * During I/O bi_private points at the dio. After I/O, bi_private is used to
317 * implement a singly-linked list of completed BIOs, at dio->bio_list.
318 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200319static void dio_bio_end_io(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320{
321 struct dio *dio = bio->bi_private;
322 unsigned long flags;
323
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324 spin_lock_irqsave(&dio->bio_lock, flags);
325 bio->bi_private = dio->bio_list;
326 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800327 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328 wake_up_process(dio->waiter);
329 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330}
331
Josef Bacikfacd07b2010-05-23 11:00:55 -0400332/**
333 * dio_end_io - handle the end io action for the given bio
334 * @bio: The direct io bio thats being completed
335 * @error: Error if there was one
336 *
337 * This is meant to be called by any filesystem that uses their own dio_submit_t
338 * so that the DIO specific endio actions are dealt with after the filesystem
339 * has done it's completion work.
340 */
341void dio_end_io(struct bio *bio, int error)
342{
343 struct dio *dio = bio->bi_private;
344
345 if (dio->is_async)
346 dio_bio_end_aio(bio, error);
347 else
348 dio_bio_end_io(bio, error);
349}
350EXPORT_SYMBOL_GPL(dio_end_io);
351
Andi Kleenba253fb2011-08-01 21:38:08 -0700352static inline void
Andi Kleeneb28be22011-08-01 21:38:03 -0700353dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
354 struct block_device *bdev,
355 sector_t first_sector, int nr_vecs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356{
357 struct bio *bio;
358
David Dillow20d96002011-01-20 14:44:22 -0800359 /*
360 * bio_alloc() is guaranteed to return a bio when called with
361 * __GFP_WAIT and we request a valid number of vectors.
362 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364
365 bio->bi_bdev = bdev;
Kent Overstreet4f024f32013-10-11 15:44:27 -0700366 bio->bi_iter.bi_sector = first_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 if (dio->is_async)
368 bio->bi_end_io = dio_bio_end_aio;
369 else
370 bio->bi_end_io = dio_bio_end_io;
371
Andi Kleeneb28be22011-08-01 21:38:03 -0700372 sdio->bio = bio;
373 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374}
375
376/*
377 * In the AIO read case we speculatively dirty the pages before starting IO.
378 * During IO completion, any of these pages which happen to have been written
379 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800380 *
381 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700383static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384{
Andi Kleeneb28be22011-08-01 21:38:03 -0700385 struct bio *bio = sdio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800386 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387
388 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800389
390 spin_lock_irqsave(&dio->bio_lock, flags);
391 dio->refcount++;
392 spin_unlock_irqrestore(&dio->bio_lock, flags);
393
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394 if (dio->is_async && dio->rw == READ)
395 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800396
Andi Kleeneb28be22011-08-01 21:38:03 -0700397 if (sdio->submit_io)
398 sdio->submit_io(dio->rw, bio, dio->inode,
399 sdio->logical_offset_in_bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400400 else
401 submit_bio(dio->rw, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402
Andi Kleeneb28be22011-08-01 21:38:03 -0700403 sdio->bio = NULL;
404 sdio->boundary = 0;
405 sdio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406}
407
408/*
409 * Release any resources in case of a failure
410 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700411static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412{
Al Viro7b2c99d2014-03-15 04:05:57 -0400413 while (sdio->head < sdio->tail)
414 page_cache_release(dio->pages[sdio->head++]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415}
416
417/*
Zach Brown02732012006-12-10 02:20:59 -0800418 * Wait for the next BIO to complete. Remove it and return it. NULL is
419 * returned once all BIOs have been completed. This must only be called once
420 * all bios have been issued so that dio->refcount can only decrease. This
421 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 */
423static struct bio *dio_await_one(struct dio *dio)
424{
425 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800426 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427
428 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800429
430 /*
431 * Wait as long as the list is empty and there are bios in flight. bio
432 * completion drops the count, maybe adds to the list, and wakes while
433 * holding the bio_lock so we don't need set_current_state()'s barrier
434 * and can call it after testing our condition.
435 */
436 while (dio->refcount > 1 && dio->bio_list == NULL) {
437 __set_current_state(TASK_UNINTERRUPTIBLE);
438 dio->waiter = current;
439 spin_unlock_irqrestore(&dio->bio_lock, flags);
440 io_schedule();
441 /* wake up sets us TASK_RUNNING */
442 spin_lock_irqsave(&dio->bio_lock, flags);
443 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 }
Zach Brown02732012006-12-10 02:20:59 -0800445 if (dio->bio_list) {
446 bio = dio->bio_list;
447 dio->bio_list = bio->bi_private;
448 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449 spin_unlock_irqrestore(&dio->bio_lock, flags);
450 return bio;
451}
452
453/*
454 * Process one completed BIO. No locks are held.
455 */
456static int dio_bio_complete(struct dio *dio, struct bio *bio)
457{
458 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
Kent Overstreetcb34e052012-09-05 15:22:02 -0700459 struct bio_vec *bvec;
460 unsigned i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461
462 if (!uptodate)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800463 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464
465 if (dio->is_async && dio->rw == READ) {
466 bio_check_pages_dirty(bio); /* transfers ownership */
467 } else {
Kent Overstreetcb34e052012-09-05 15:22:02 -0700468 bio_for_each_segment_all(bvec, bio, i) {
469 struct page *page = bvec->bv_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470
471 if (dio->rw == READ && !PageCompound(page))
472 set_page_dirty_lock(page);
473 page_cache_release(page);
474 }
475 bio_put(bio);
476 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 return uptodate ? 0 : -EIO;
478}
479
480/*
Zach Brown02732012006-12-10 02:20:59 -0800481 * Wait on and process all in-flight BIOs. This must only be called once
482 * all bios have been issued so that the refcount can only decrease.
483 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200484 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800485 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800487static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488{
Zach Brown02732012006-12-10 02:20:59 -0800489 struct bio *bio;
490 do {
491 bio = dio_await_one(dio);
492 if (bio)
493 dio_bio_complete(dio, bio);
494 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495}
496
497/*
498 * A really large O_DIRECT read or write can generate a lot of BIOs. So
499 * to keep the memory consumption sane we periodically reap any completed BIOs
500 * during the BIO generation phase.
501 *
502 * This also helps to limit the peak amount of pinned userspace memory.
503 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700504static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505{
506 int ret = 0;
507
Andi Kleeneb28be22011-08-01 21:38:03 -0700508 if (sdio->reap_counter++ >= 64) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 while (dio->bio_list) {
510 unsigned long flags;
511 struct bio *bio;
512 int ret2;
513
514 spin_lock_irqsave(&dio->bio_lock, flags);
515 bio = dio->bio_list;
516 dio->bio_list = bio->bi_private;
517 spin_unlock_irqrestore(&dio->bio_lock, flags);
518 ret2 = dio_bio_complete(dio, bio);
519 if (ret == 0)
520 ret = ret2;
521 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700522 sdio->reap_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 }
524 return ret;
525}
526
527/*
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200528 * Create workqueue for deferred direct IO completions. We allocate the
529 * workqueue when it's first needed. This avoids creating workqueue for
530 * filesystems that don't need it and also allows us to create the workqueue
531 * late enough so the we can include s_id in the name of the workqueue.
532 */
533static int sb_init_dio_done_wq(struct super_block *sb)
534{
Olof Johansson45150c42013-09-09 10:34:23 -0700535 struct workqueue_struct *old;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200536 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
537 WQ_MEM_RECLAIM, 0,
538 sb->s_id);
539 if (!wq)
540 return -ENOMEM;
541 /*
542 * This has to be atomic as more DIOs can race to create the workqueue
543 */
Olof Johansson45150c42013-09-09 10:34:23 -0700544 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200545 /* Someone created workqueue before us? Free ours... */
Olof Johansson45150c42013-09-09 10:34:23 -0700546 if (old)
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200547 destroy_workqueue(wq);
548 return 0;
549}
550
551static int dio_set_defer_completion(struct dio *dio)
552{
553 struct super_block *sb = dio->inode->i_sb;
554
555 if (dio->defer_completion)
556 return 0;
557 dio->defer_completion = true;
558 if (!sb->s_dio_done_wq)
559 return sb_init_dio_done_wq(sb);
560 return 0;
561}
562
563/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 * Call into the fs to map some more disk blocks. We record the current number
Andi Kleeneb28be22011-08-01 21:38:03 -0700565 * of available blocks at sdio->blocks_available. These are in units of the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566 * fs blocksize, (1 << inode->i_blkbits).
567 *
568 * The fs is allowed to map lots of blocks at once. If it wants to do that,
569 * it uses the passed inode-relative block number as the file offset, as usual.
570 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800571 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 * has remaining to do. The fs should not map more than this number of blocks.
573 *
574 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
575 * indicate how much contiguous disk space has been made available at
576 * bh->b_blocknr.
577 *
578 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
579 * This isn't very efficient...
580 *
581 * In the case of filesystem holes: the fs may return an arbitrarily-large
582 * hole by returning an appropriate value in b_size and by clearing
583 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800584 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 */
Andi Kleen18772642011-08-01 21:38:07 -0700586static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
587 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588{
589 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
Tao Maae55e1a2012-01-12 17:20:33 -0800591 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592 unsigned long fs_count; /* Number of filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593 int create;
Linus Torvaldsab738572012-11-29 12:27:00 -0800594 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595
596 /*
597 * If there was a memory error and we've overwritten all the
598 * mapped blocks then we can now return that memory error
599 */
600 ret = dio->page_errors;
601 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700602 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
603 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
Tao Maae55e1a2012-01-12 17:20:33 -0800604 fs_endblk = (sdio->final_block_in_request - 1) >>
605 sdio->blkfactor;
606 fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607
Nathan Scott3c674e72006-03-29 09:26:15 +1000608 map_bh->b_state = 0;
Linus Torvaldsab738572012-11-29 12:27:00 -0800609 map_bh->b_size = fs_count << i_blkbits;
Nathan Scott3c674e72006-03-29 09:26:15 +1000610
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800611 /*
612 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
613 * forbid block creations: only overwrites are permitted.
614 * We will return early to the caller once we see an
615 * unmapped buffer head returned, and the caller will fall
616 * back to buffered I/O.
617 *
618 * Otherwise the decision is left to the get_blocks method,
619 * which may decide to handle it or also return an unmapped
620 * buffer head.
621 */
Jens Axboeb31dc662006-06-13 08:26:10 +0200622 create = dio->rw & WRITE;
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800623 if (dio->flags & DIO_SKIP_HOLES) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700624 if (sdio->block_in_file < (i_size_read(dio->inode) >>
625 sdio->blkbits))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000628
Andi Kleeneb28be22011-08-01 21:38:03 -0700629 ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 map_bh, create);
Andi Kleen18772642011-08-01 21:38:07 -0700631
632 /* Store for completion */
633 dio->private = map_bh->b_private;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200634
635 if (ret == 0 && buffer_defer_completion(map_bh))
636 ret = dio_set_defer_completion(dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637 }
638 return ret;
639}
640
641/*
642 * There is no bio. Make one now.
643 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700644static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
645 sector_t start_sector, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646{
647 sector_t sector;
648 int ret, nr_pages;
649
Andi Kleeneb28be22011-08-01 21:38:03 -0700650 ret = dio_bio_reap(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 if (ret)
652 goto out;
Andi Kleeneb28be22011-08-01 21:38:03 -0700653 sector = start_sector << (sdio->blkbits - 9);
Andi Kleen18772642011-08-01 21:38:07 -0700654 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655 BUG_ON(nr_pages <= 0);
Andi Kleen18772642011-08-01 21:38:07 -0700656 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
Andi Kleeneb28be22011-08-01 21:38:03 -0700657 sdio->boundary = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658out:
659 return ret;
660}
661
662/*
663 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
664 * that was successful then update final_block_in_bio and take a ref against
665 * the just-added page.
666 *
667 * Return zero on success. Non-zero means the caller needs to start a new BIO.
668 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700669static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670{
671 int ret;
672
Andi Kleeneb28be22011-08-01 21:38:03 -0700673 ret = bio_add_page(sdio->bio, sdio->cur_page,
674 sdio->cur_page_len, sdio->cur_page_offset);
675 if (ret == sdio->cur_page_len) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 /*
677 * Decrement count only, if we are done with this page
678 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700679 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
680 sdio->pages_in_io--;
681 page_cache_get(sdio->cur_page);
682 sdio->final_block_in_bio = sdio->cur_page_block +
683 (sdio->cur_page_len >> sdio->blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 ret = 0;
685 } else {
686 ret = 1;
687 }
688 return ret;
689}
690
691/*
692 * Put cur_page under IO. The section of cur_page which is described by
693 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
694 * starts on-disk at cur_page_block.
695 *
696 * We take a ref against the page here (on behalf of its presence in the bio).
697 *
698 * The caller of this function is responsible for removing cur_page from the
699 * dio, and for dropping the refcount which came from that presence.
700 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700701static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
702 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703{
704 int ret = 0;
705
Andi Kleeneb28be22011-08-01 21:38:03 -0700706 if (sdio->bio) {
707 loff_t cur_offset = sdio->cur_page_fs_offset;
708 loff_t bio_next_offset = sdio->logical_offset_in_bio +
Kent Overstreet4f024f32013-10-11 15:44:27 -0700709 sdio->bio->bi_iter.bi_size;
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400710
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400712 * See whether this new request is contiguous with the old.
713 *
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900714 * Btrfs cannot handle having logically non-contiguous requests
715 * submitted. For example if you have
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400716 *
717 * Logical: [0-4095][HOLE][8192-12287]
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900718 * Physical: [0-4095] [4096-8191]
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400719 *
720 * We cannot submit those pages together as one BIO. So if our
721 * current logical offset in the file does not equal what would
722 * be the next logical offset in the bio, submit the bio we
723 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700725 if (sdio->final_block_in_bio != sdio->cur_page_block ||
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400726 cur_offset != bio_next_offset)
Andi Kleeneb28be22011-08-01 21:38:03 -0700727 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728 }
729
Andi Kleeneb28be22011-08-01 21:38:03 -0700730 if (sdio->bio == NULL) {
Andi Kleen18772642011-08-01 21:38:07 -0700731 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732 if (ret)
733 goto out;
734 }
735
Andi Kleeneb28be22011-08-01 21:38:03 -0700736 if (dio_bio_add_page(sdio) != 0) {
737 dio_bio_submit(dio, sdio);
Andi Kleen18772642011-08-01 21:38:07 -0700738 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700740 ret = dio_bio_add_page(sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 BUG_ON(ret != 0);
742 }
743 }
744out:
745 return ret;
746}
747
748/*
749 * An autonomous function to put a chunk of a page under deferred IO.
750 *
751 * The caller doesn't actually know (or care) whether this piece of page is in
752 * a BIO, or is under IO or whatever. We just take care of all possible
753 * situations here. The separation between the logic of do_direct_IO() and
754 * that of submit_page_section() is important for clarity. Please don't break.
755 *
756 * The chunk of page starts on-disk at blocknr.
757 *
758 * We perform deferred IO, by recording the last-submitted page inside our
759 * private part of the dio structure. If possible, we just expand the IO
760 * across that page here.
761 *
762 * If that doesn't work out then we put the old page into the bio and add this
763 * page to the dio instead.
764 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700765static inline int
Andi Kleeneb28be22011-08-01 21:38:03 -0700766submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
Andi Kleen18772642011-08-01 21:38:07 -0700767 unsigned offset, unsigned len, sector_t blocknr,
768 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769{
770 int ret = 0;
771
Andrew Morton98c4d572006-12-10 02:19:47 -0800772 if (dio->rw & WRITE) {
773 /*
774 * Read accounting is performed in submit_bio()
775 */
776 task_io_account_write(len);
777 }
778
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 /*
780 * Can we just grow the current page's presence in the dio?
781 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700782 if (sdio->cur_page == page &&
783 sdio->cur_page_offset + sdio->cur_page_len == offset &&
784 sdio->cur_page_block +
785 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
786 sdio->cur_page_len += len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787 goto out;
788 }
789
790 /*
791 * If there's a deferred page already there then send it.
792 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700793 if (sdio->cur_page) {
Andi Kleen18772642011-08-01 21:38:07 -0700794 ret = dio_send_cur_page(dio, sdio, map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -0700795 page_cache_release(sdio->cur_page);
796 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797 if (ret)
Jan Karab1058b92013-04-29 15:06:18 -0700798 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 }
800
801 page_cache_get(page); /* It is in dio */
Andi Kleeneb28be22011-08-01 21:38:03 -0700802 sdio->cur_page = page;
803 sdio->cur_page_offset = offset;
804 sdio->cur_page_len = len;
805 sdio->cur_page_block = blocknr;
806 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807out:
Jan Karab1058b92013-04-29 15:06:18 -0700808 /*
809 * If sdio->boundary then we want to schedule the IO now to
810 * avoid metadata seeks.
811 */
812 if (sdio->boundary) {
813 ret = dio_send_cur_page(dio, sdio, map_bh);
814 dio_bio_submit(dio, sdio);
815 page_cache_release(sdio->cur_page);
816 sdio->cur_page = NULL;
817 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818 return ret;
819}
820
821/*
822 * Clean any dirty buffers in the blockdev mapping which alias newly-created
823 * file blocks. Only called for S_ISREG files - blockdevs do not set
824 * buffer_new
825 */
Andi Kleen18772642011-08-01 21:38:07 -0700826static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827{
828 unsigned i;
829 unsigned nblocks;
830
Andi Kleen18772642011-08-01 21:38:07 -0700831 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832
833 for (i = 0; i < nblocks; i++) {
Andi Kleen18772642011-08-01 21:38:07 -0700834 unmap_underlying_metadata(map_bh->b_bdev,
835 map_bh->b_blocknr + i);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836 }
837}
838
839/*
840 * If we are not writing the entire block and get_block() allocated
841 * the block for us, we need to fill-in the unused portion of the
842 * block with zeros. This happens only if user-buffer, fileoffset or
843 * io length is not filesystem block-size multiple.
844 *
845 * `end' is zero if we're doing the start of the IO, 1 at the end of the
846 * IO.
847 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700848static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
849 int end, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850{
851 unsigned dio_blocks_per_fs_block;
852 unsigned this_chunk_blocks; /* In dio_blocks */
853 unsigned this_chunk_bytes;
854 struct page *page;
855
Andi Kleeneb28be22011-08-01 21:38:03 -0700856 sdio->start_zero_done = 1;
Andi Kleen18772642011-08-01 21:38:07 -0700857 if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858 return;
859
Andi Kleeneb28be22011-08-01 21:38:03 -0700860 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
861 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862
863 if (!this_chunk_blocks)
864 return;
865
866 /*
867 * We need to zero out part of an fs block. It is either at the
868 * beginning or the end of the fs block.
869 */
870 if (end)
871 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
872
Andi Kleeneb28be22011-08-01 21:38:03 -0700873 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874
Nick Piggin557ed1f2007-10-16 01:24:40 -0700875 page = ZERO_PAGE(0);
Andi Kleeneb28be22011-08-01 21:38:03 -0700876 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700877 sdio->next_block_for_io, map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878 return;
879
Andi Kleeneb28be22011-08-01 21:38:03 -0700880 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881}
882
883/*
884 * Walk the user pages, and the file, mapping blocks to disk and generating
885 * a sequence of (page,offset,len,block) mappings. These mappings are injected
886 * into submit_page_section(), which takes care of the next stage of submission
887 *
888 * Direct IO against a blockdev is different from a file. Because we can
889 * happily perform page-sized but 512-byte aligned IOs. It is important that
890 * blockdev IO be able to have fine alignment and large sizes.
891 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800892 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 * with the size of IO which is permitted at this offset and this i_blkbits.
894 *
895 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800896 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 * fine alignment but still allows this function to work in PAGE_SIZE units.
898 */
Andi Kleen18772642011-08-01 21:38:07 -0700899static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
900 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901{
Andi Kleeneb28be22011-08-01 21:38:03 -0700902 const unsigned blkbits = sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903 int ret = 0;
904
Andi Kleeneb28be22011-08-01 21:38:03 -0700905 while (sdio->block_in_file < sdio->final_block_in_request) {
Al Viro7b2c99d2014-03-15 04:05:57 -0400906 struct page *page;
907 size_t from, to;
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300908
909 page = dio_get_page(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910 if (IS_ERR(page)) {
911 ret = PTR_ERR(page);
912 goto out;
913 }
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300914 from = sdio->head ? 0 : sdio->from;
915 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
916 sdio->head++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917
Al Viro7b2c99d2014-03-15 04:05:57 -0400918 while (from < to) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919 unsigned this_chunk_bytes; /* # of bytes mapped */
920 unsigned this_chunk_blocks; /* # of blocks */
921 unsigned u;
922
Andi Kleeneb28be22011-08-01 21:38:03 -0700923 if (sdio->blocks_available == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924 /*
925 * Need to go and map some more disk
926 */
927 unsigned long blkmask;
928 unsigned long dio_remainder;
929
Andi Kleen18772642011-08-01 21:38:07 -0700930 ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 if (ret) {
932 page_cache_release(page);
933 goto out;
934 }
935 if (!buffer_mapped(map_bh))
936 goto do_holes;
937
Andi Kleeneb28be22011-08-01 21:38:03 -0700938 sdio->blocks_available =
939 map_bh->b_size >> sdio->blkbits;
940 sdio->next_block_for_io =
941 map_bh->b_blocknr << sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942 if (buffer_new(map_bh))
Andi Kleen18772642011-08-01 21:38:07 -0700943 clean_blockdev_aliases(dio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944
Andi Kleeneb28be22011-08-01 21:38:03 -0700945 if (!sdio->blkfactor)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 goto do_holes;
947
Andi Kleeneb28be22011-08-01 21:38:03 -0700948 blkmask = (1 << sdio->blkfactor) - 1;
949 dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950
951 /*
952 * If we are at the start of IO and that IO
953 * starts partway into a fs-block,
954 * dio_remainder will be non-zero. If the IO
955 * is a read then we can simply advance the IO
956 * cursor to the first block which is to be
957 * read. But if the IO is a write and the
958 * block was newly allocated we cannot do that;
959 * the start of the fs block must be zeroed out
960 * on-disk
961 */
962 if (!buffer_new(map_bh))
Andi Kleeneb28be22011-08-01 21:38:03 -0700963 sdio->next_block_for_io += dio_remainder;
964 sdio->blocks_available -= dio_remainder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 }
966do_holes:
967 /* Handle holes */
968 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800969 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970
971 /* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboeb31dc662006-06-13 08:26:10 +0200972 if (dio->rw & WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 page_cache_release(page);
974 return -ENOTBLK;
975 }
976
Jeff Moyer35dc8162006-02-03 03:04:27 -0800977 /*
978 * Be sure to account for a partial block as the
979 * last block in the file
980 */
981 i_size_aligned = ALIGN(i_size_read(dio->inode),
982 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -0700983 if (sdio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -0800984 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985 /* We hit eof */
986 page_cache_release(page);
987 goto out;
988 }
Al Viro7b2c99d2014-03-15 04:05:57 -0400989 zero_user(page, from, 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -0700990 sdio->block_in_file++;
Al Viro7b2c99d2014-03-15 04:05:57 -0400991 from += 1 << blkbits;
Al Viro3320c602014-03-10 02:30:55 -0400992 dio->result += 1 << blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 goto next_block;
994 }
995
996 /*
997 * If we're performing IO which has an alignment which
998 * is finer than the underlying fs, go check to see if
999 * we must zero out the start of this block.
1000 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001001 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
Andi Kleen18772642011-08-01 21:38:07 -07001002 dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003
1004 /*
1005 * Work out, in this_chunk_blocks, how much disk we
1006 * can add to this page
1007 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001008 this_chunk_blocks = sdio->blocks_available;
Al Viro7b2c99d2014-03-15 04:05:57 -04001009 u = (to - from) >> blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010 if (this_chunk_blocks > u)
1011 this_chunk_blocks = u;
Andi Kleeneb28be22011-08-01 21:38:03 -07001012 u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013 if (this_chunk_blocks > u)
1014 this_chunk_blocks = u;
1015 this_chunk_bytes = this_chunk_blocks << blkbits;
1016 BUG_ON(this_chunk_bytes == 0);
1017
Jan Kara092c8d42013-04-29 15:06:17 -07001018 if (this_chunk_blocks == sdio->blocks_available)
1019 sdio->boundary = buffer_boundary(map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -07001020 ret = submit_page_section(dio, sdio, page,
Al Viro7b2c99d2014-03-15 04:05:57 -04001021 from,
Andi Kleeneb28be22011-08-01 21:38:03 -07001022 this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -07001023 sdio->next_block_for_io,
1024 map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025 if (ret) {
1026 page_cache_release(page);
1027 goto out;
1028 }
Andi Kleeneb28be22011-08-01 21:38:03 -07001029 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030
Andi Kleeneb28be22011-08-01 21:38:03 -07001031 sdio->block_in_file += this_chunk_blocks;
Al Viro7b2c99d2014-03-15 04:05:57 -04001032 from += this_chunk_bytes;
1033 dio->result += this_chunk_bytes;
Andi Kleeneb28be22011-08-01 21:38:03 -07001034 sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035next_block:
Andi Kleeneb28be22011-08-01 21:38:03 -07001036 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1037 if (sdio->block_in_file == sdio->final_block_in_request)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 break;
1039 }
1040
1041 /* Drop the ref which was taken in get_user_pages() */
1042 page_cache_release(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 }
1044out:
1045 return ret;
1046}
1047
Andi Kleen847cc632011-08-01 21:38:09 -07001048static inline int drop_refcount(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001049{
Andi Kleen847cc632011-08-01 21:38:09 -07001050 int ret2;
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001051 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052
Zach Brown8459d862006-12-10 02:21:05 -08001053 /*
1054 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001055 * operation. AIO can if it was a broken operation described above or
1056 * in fact if all the bios race to complete before we get here. In
1057 * that case dio_complete() translates the EIOCBQUEUED into the proper
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001058 * return code that the caller will hand to ->complete().
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001059 *
1060 * This is managed by the bio_lock instead of being an atomic_t so that
1061 * completion paths can drop their ref and use the remaining count to
1062 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001063 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001064 spin_lock_irqsave(&dio->bio_lock, flags);
1065 ret2 = --dio->refcount;
1066 spin_unlock_irqrestore(&dio->bio_lock, flags);
Andi Kleen847cc632011-08-01 21:38:09 -07001067 return ret2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068}
1069
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001070/*
1071 * This is a library function for use by filesystem drivers.
1072 *
1073 * The locking rules are governed by the flags parameter:
1074 * - if the flags value contains DIO_LOCKING we use a fancy locking
1075 * scheme for dumb filesystems.
1076 * For writes this function is called under i_mutex and returns with
1077 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1078 * taken and dropped again before returning.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001079 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1080 * internal locking but rather rely on the filesystem to synchronize
1081 * direct I/O reads/writes versus each other and truncate.
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001082 *
1083 * To help with locking against truncate we incremented the i_dio_count
1084 * counter before starting direct I/O, and decrement it once we are done.
1085 * Truncate can wait for it to reach zero to provide exclusion. It is
1086 * expected that filesystem provide exclusion between new direct I/O
1087 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1088 * but other filesystems need to take care of this on their own.
Andi Kleenba253fb2011-08-01 21:38:08 -07001089 *
1090 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1091 * is always inlined. Otherwise gcc is unable to split the structure into
1092 * individual fields and will generate much worse code. This is important
1093 * for the whole file.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001094 */
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001095static inline ssize_t
1096do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
Al Viro31b14032014-03-05 01:33:16 -05001097 struct block_device *bdev, struct iov_iter *iter, loff_t offset,
1098 get_block_t get_block, dio_iodone_t end_io,
Josef Bacikfacd07b2010-05-23 11:00:55 -04001099 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100{
Linus Torvaldsab738572012-11-29 12:27:00 -08001101 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1102 unsigned blkbits = i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103 unsigned blocksize_mask = (1 << blkbits) - 1;
1104 ssize_t retval = -EINVAL;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001105 size_t count = iov_iter_count(iter);
1106 loff_t end = offset + count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107 struct dio *dio;
Andi Kleeneb28be22011-08-01 21:38:03 -07001108 struct dio_submit sdio = { 0, };
Andi Kleen847cc632011-08-01 21:38:09 -07001109 struct buffer_head map_bh = { 0, };
Fengguang Wu647d1e42012-08-09 15:23:09 +02001110 struct blk_plug plug;
Al Viro886a3912014-03-05 13:50:45 -05001111 unsigned long align = offset | iov_iter_alignment(iter);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112
1113 if (rw & WRITE)
Jens Axboe721a9602011-03-09 11:56:30 +01001114 rw = WRITE_ODIRECT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001116 /*
1117 * Avoid references to bdev if not absolutely needed to give
1118 * the early prefetch in the caller enough time.
1119 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120
Al Viro886a3912014-03-05 13:50:45 -05001121 if (align & blocksize_mask) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001123 blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 blocksize_mask = (1 << blkbits) - 1;
Al Viro886a3912014-03-05 13:50:45 -05001125 if (align & blocksize_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 goto out;
1127 }
1128
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001129 /* watch out for a 0 len io from a tricksy fs */
Al Viro886a3912014-03-05 13:50:45 -05001130 if (rw == READ && !iov_iter_count(iter))
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001131 return 0;
1132
Andi Kleen6e8267f2011-08-01 21:38:06 -07001133 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134 retval = -ENOMEM;
1135 if (!dio)
1136 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001137 /*
1138 * Believe it or not, zeroing out the page array caused a .5%
1139 * performance regression in a database benchmark. So, we take
1140 * care to only zero out what's needed.
1141 */
1142 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001144 dio->flags = flags;
1145 if (dio->flags & DIO_LOCKING) {
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001146 if (rw == READ) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001147 struct address_space *mapping =
1148 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001150 /* will be released by direct_io_worker */
1151 mutex_lock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152
1153 retval = filemap_write_and_wait_range(mapping, offset,
1154 end - 1);
1155 if (retval) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001156 mutex_unlock(&inode->i_mutex);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001157 kmem_cache_free(dio_cache, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 goto out;
1159 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161 }
1162
1163 /*
Christoph Hellwig60392572014-02-10 10:27:11 +11001164 * For file extending writes updating i_size before data writeouts
1165 * complete can expose uninitialized blocks in dumb filesystems.
1166 * In that case we need to wait for I/O completion even if asked
1167 * for an asynchronous write.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 */
Christoph Hellwig60392572014-02-10 10:27:11 +11001169 if (is_sync_kiocb(iocb))
1170 dio->is_async = false;
1171 else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
1172 (rw & WRITE) && end > i_size_read(inode))
1173 dio->is_async = false;
1174 else
1175 dio->is_async = true;
1176
Andi Kleen847cc632011-08-01 21:38:09 -07001177 dio->inode = inode;
1178 dio->rw = rw;
Christoph Hellwig02afc272013-09-04 15:04:40 +02001179
1180 /*
1181 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1182 * so that we can call ->fsync.
1183 */
1184 if (dio->is_async && (rw & WRITE) &&
1185 ((iocb->ki_filp->f_flags & O_DSYNC) ||
1186 IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1187 retval = dio_set_defer_completion(dio);
1188 if (retval) {
1189 /*
1190 * We grab i_mutex only for reads so we don't have
1191 * to release it here
1192 */
1193 kmem_cache_free(dio_cache, dio);
1194 goto out;
1195 }
1196 }
1197
1198 /*
1199 * Will be decremented at I/O completion time.
1200 */
1201 atomic_inc(&inode->i_dio_count);
1202
1203 retval = 0;
Andi Kleen847cc632011-08-01 21:38:09 -07001204 sdio.blkbits = blkbits;
Linus Torvaldsab738572012-11-29 12:27:00 -08001205 sdio.blkfactor = i_blkbits - blkbits;
Andi Kleen847cc632011-08-01 21:38:09 -07001206 sdio.block_in_file = offset >> blkbits;
1207
1208 sdio.get_block = get_block;
1209 dio->end_io = end_io;
1210 sdio.submit_io = submit_io;
1211 sdio.final_block_in_bio = -1;
1212 sdio.next_block_for_io = -1;
1213
1214 dio->iocb = iocb;
1215 dio->i_size = i_size_read(inode);
1216
1217 spin_lock_init(&dio->bio_lock);
1218 dio->refcount = 1;
1219
Al Viro7b2c99d2014-03-15 04:05:57 -04001220 sdio.iter = iter;
1221 sdio.final_block_in_request =
1222 (offset + iov_iter_count(iter)) >> blkbits;
1223
Andi Kleen847cc632011-08-01 21:38:09 -07001224 /*
1225 * In case of non-aligned buffers, we may need 2 more
1226 * pages since we need to zero out first and last block.
1227 */
1228 if (unlikely(sdio.blkfactor))
1229 sdio.pages_in_io = 2;
1230
Al Virof67da302014-03-19 01:16:16 -04001231 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
Andi Kleen847cc632011-08-01 21:38:09 -07001232
Fengguang Wu647d1e42012-08-09 15:23:09 +02001233 blk_start_plug(&plug);
1234
Al Viro7b2c99d2014-03-15 04:05:57 -04001235 retval = do_direct_IO(dio, &sdio, &map_bh);
1236 if (retval)
1237 dio_cleanup(dio, &sdio);
Andi Kleen847cc632011-08-01 21:38:09 -07001238
1239 if (retval == -ENOTBLK) {
1240 /*
1241 * The remaining part of the request will be
1242 * be handled by buffered I/O when we return
1243 */
1244 retval = 0;
1245 }
1246 /*
1247 * There may be some unwritten disk at the end of a part-written
1248 * fs-block-sized block. Go zero that now.
1249 */
1250 dio_zero_block(dio, &sdio, 1, &map_bh);
1251
1252 if (sdio.cur_page) {
1253 ssize_t ret2;
1254
1255 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1256 if (retval == 0)
1257 retval = ret2;
1258 page_cache_release(sdio.cur_page);
1259 sdio.cur_page = NULL;
1260 }
1261 if (sdio.bio)
1262 dio_bio_submit(dio, &sdio);
1263
Fengguang Wu647d1e42012-08-09 15:23:09 +02001264 blk_finish_plug(&plug);
1265
Andi Kleen847cc632011-08-01 21:38:09 -07001266 /*
1267 * It is possible that, we return short IO due to end of file.
1268 * In that case, we need to release all the pages we got hold on.
1269 */
1270 dio_cleanup(dio, &sdio);
1271
1272 /*
1273 * All block lookups have been performed. For READ requests
1274 * we can let i_mutex go now that its achieved its purpose
1275 * of protecting us from looking up uninitialized blocks.
1276 */
1277 if (rw == READ && (dio->flags & DIO_LOCKING))
1278 mutex_unlock(&dio->inode->i_mutex);
1279
1280 /*
1281 * The only time we want to leave bios in flight is when a successful
1282 * partial aio read or full aio write have been setup. In that case
1283 * bio completion will call aio_complete. The only time it's safe to
1284 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1285 * This had *better* be the only place that raises -EIOCBQUEUED.
1286 */
1287 BUG_ON(retval == -EIOCBQUEUED);
1288 if (dio->is_async && retval == 0 && dio->result &&
Christoph Hellwigaf436472014-07-30 07:18:48 -04001289 (rw == READ || dio->result == count))
Andi Kleen847cc632011-08-01 21:38:09 -07001290 retval = -EIOCBQUEUED;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001291 else
Andi Kleen847cc632011-08-01 21:38:09 -07001292 dio_await_completion(dio);
1293
1294 if (drop_refcount(dio) == 0) {
1295 retval = dio_complete(dio, offset, retval, false);
Andi Kleen847cc632011-08-01 21:38:09 -07001296 } else
1297 BUG_ON(retval != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001299out:
1300 return retval;
1301}
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001302
1303ssize_t
1304__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
Al Viro31b14032014-03-05 01:33:16 -05001305 struct block_device *bdev, struct iov_iter *iter, loff_t offset,
1306 get_block_t get_block, dio_iodone_t end_io,
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001307 dio_submit_t submit_io, int flags)
1308{
1309 /*
1310 * The block device state is needed in the end to finally
1311 * submit everything. Since it's likely to be cache cold
1312 * prefetch it here as first thing to hide some of the
1313 * latency.
1314 *
1315 * Attempt to prefetch the pieces we likely need later.
1316 */
1317 prefetch(&bdev->bd_disk->part_tbl);
1318 prefetch(bdev->bd_queue);
1319 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1320
Al Viro31b14032014-03-05 01:33:16 -05001321 return do_blockdev_direct_IO(rw, iocb, inode, bdev, iter, offset,
1322 get_block, end_io, submit_io, flags);
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001323}
1324
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325EXPORT_SYMBOL(__blockdev_direct_IO);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001326
1327static __init int dio_init(void)
1328{
1329 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1330 return 0;
1331}
1332module_init(dio_init)