blob: e181b6b2e297fb5d3bd03a07efe382f0dd204972 [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 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070074 sector_t block_in_file; /* Current offset into the underlying
75 file in dio_block units. */
76 unsigned blocks_available; /* At block_in_file. changes */
Andi Kleen0dc2bc42011-08-01 21:38:05 -070077 int reap_counter; /* rate limit reaping */
Linus Torvalds1da177e2005-04-16 15:20:36 -070078 sector_t final_block_in_request;/* doesn't change */
Linus Torvalds1da177e2005-04-16 15:20:36 -070079 int boundary; /* prev block is at a boundary */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080080 get_block_t *get_block; /* block mapping function */
Josef Bacikfacd07b2010-05-23 11:00:55 -040081 dio_submit_t *submit_io; /* IO submition function */
Andi Kleeneb28be22011-08-01 21:38:03 -070082
Josef Bacikfacd07b2010-05-23 11:00:55 -040083 loff_t logical_offset_in_bio; /* current first logical block in bio */
Linus Torvalds1da177e2005-04-16 15:20:36 -070084 sector_t final_block_in_bio; /* current final block in bio + 1 */
85 sector_t next_block_for_io; /* next block to be put under IO,
86 in dio_blocks units */
Linus Torvalds1da177e2005-04-16 15:20:36 -070087
88 /*
89 * Deferred addition of a page to the dio. These variables are
90 * private to dio_send_cur_page(), submit_page_section() and
91 * dio_bio_add_page().
92 */
93 struct page *cur_page; /* The page */
94 unsigned cur_page_offset; /* Offset into it, in bytes */
95 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
96 sector_t cur_page_block; /* Where it starts */
Josef Bacikfacd07b2010-05-23 11:00:55 -040097 loff_t cur_page_fs_offset; /* Offset in file */
Linus Torvalds1da177e2005-04-16 15:20:36 -070098
Al Viro7b2c99d2014-03-15 04:05:57 -040099 struct iov_iter *iter;
Jeff Moyer23aee092009-12-15 16:47:49 -0800100 /*
101 * Page queue. These variables belong to dio_refill_pages() and
102 * dio_get_page().
103 */
104 unsigned head; /* next page to process */
105 unsigned tail; /* last valid page + 1 */
Al Viro7b2c99d2014-03-15 04:05:57 -0400106 size_t from, to;
Andi Kleeneb28be22011-08-01 21:38:03 -0700107};
108
109/* dio_state communicated between submission path and end_io */
110struct dio {
111 int flags; /* doesn't change */
Andi Kleeneb28be22011-08-01 21:38:03 -0700112 int rw;
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700113 struct inode *inode;
Andi Kleeneb28be22011-08-01 21:38:03 -0700114 loff_t i_size; /* i_size when submitted */
115 dio_iodone_t *end_io; /* IO completion function */
Andi Kleeneb28be22011-08-01 21:38:03 -0700116
Andi Kleen18772642011-08-01 21:38:07 -0700117 void *private; /* copy from map_bh.b_private */
Andi Kleeneb28be22011-08-01 21:38:03 -0700118
119 /* BIO completion state */
120 spinlock_t bio_lock; /* protects BIO fields below */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700121 int page_errors; /* errno from get_user_pages() */
122 int is_async; /* is IO async ? */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200123 bool defer_completion; /* defer AIO completion to workqueue? */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700124 int io_error; /* IO error in completion path */
Andi Kleeneb28be22011-08-01 21:38:03 -0700125 unsigned long refcount; /* direct_io_worker() and bios */
126 struct bio *bio_list; /* singly linked via bi_private */
127 struct task_struct *waiter; /* waiting task (NULL if none) */
128
129 /* AIO related stuff */
130 struct kiocb *iocb; /* kiocb */
Andi Kleeneb28be22011-08-01 21:38:03 -0700131 ssize_t result; /* IO result */
132
Jeff Moyer23aee092009-12-15 16:47:49 -0800133 /*
134 * pages[] (and any fields placed after it) are not zeroed out at
135 * allocation time. Don't add new fields after pages[] unless you
136 * wish that they not be zeroed.
137 */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200138 union {
139 struct page *pages[DIO_PAGES]; /* page buffer */
140 struct work_struct complete_work;/* deferred AIO completion */
141 };
Andi Kleen6e8267f2011-08-01 21:38:06 -0700142} ____cacheline_aligned_in_smp;
143
144static struct kmem_cache *dio_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145
146/*
147 * How many pages are in the queue?
148 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700149static inline unsigned dio_pages_present(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150{
Andi Kleeneb28be22011-08-01 21:38:03 -0700151 return sdio->tail - sdio->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152}
153
154/*
155 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
156 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700157static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158{
Al Viro7b2c99d2014-03-15 04:05:57 -0400159 ssize_t ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160
Miklos Szeredi2c809292014-09-24 17:09:11 +0200161 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
Al Viro7b2c99d2014-03-15 04:05:57 -0400162 &sdio->from);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163
Andi Kleeneb28be22011-08-01 21:38:03 -0700164 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
Nick Piggin557ed1f2007-10-16 01:24:40 -0700165 struct page *page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 /*
167 * A memory fault, but the filesystem has some outstanding
168 * mapped blocks. We need to use those blocks up to avoid
169 * leaking stale data in the file.
170 */
171 if (dio->page_errors == 0)
172 dio->page_errors = ret;
Nick Pigginb5810032005-10-29 18:16:12 -0700173 page_cache_get(page);
174 dio->pages[0] = page;
Andi Kleeneb28be22011-08-01 21:38:03 -0700175 sdio->head = 0;
176 sdio->tail = 1;
Al Viro7b2c99d2014-03-15 04:05:57 -0400177 sdio->from = 0;
178 sdio->to = PAGE_SIZE;
179 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 }
181
182 if (ret >= 0) {
Al Viro7b2c99d2014-03-15 04:05:57 -0400183 iov_iter_advance(sdio->iter, ret);
184 ret += sdio->from;
Andi Kleeneb28be22011-08-01 21:38:03 -0700185 sdio->head = 0;
Al Viro7b2c99d2014-03-15 04:05:57 -0400186 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
187 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
188 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190 return ret;
191}
192
193/*
194 * Get another userspace page. Returns an ERR_PTR on error. Pages are
195 * buffered inside the dio so that we can call get_user_pages() against a
196 * decent number of pages, less frequently. To provide nicer use of the
197 * L1 cache.
198 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700199static inline struct page *dio_get_page(struct dio *dio,
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300200 struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201{
Andi Kleeneb28be22011-08-01 21:38:03 -0700202 if (dio_pages_present(sdio) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203 int ret;
204
Andi Kleeneb28be22011-08-01 21:38:03 -0700205 ret = dio_refill_pages(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206 if (ret)
207 return ERR_PTR(ret);
Andi Kleeneb28be22011-08-01 21:38:03 -0700208 BUG_ON(dio_pages_present(sdio) == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209 }
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300210 return dio->pages[sdio->head];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211}
212
Zach Brown6d544bb2006-12-10 02:20:54 -0800213/**
214 * dio_complete() - called when all DIO BIO I/O has been completed
215 * @offset: the byte offset in the file of the completed operation
216 *
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200217 * This drops i_dio_count, lets interested parties know that a DIO operation
218 * has completed, and calculates the resulting return code for the operation.
Zach Brown6d544bb2006-12-10 02:20:54 -0800219 *
220 * It lets the filesystem know if it registered an interest earlier via
221 * get_block. Pass the private field of the map buffer_head so that
222 * filesystems can use it to hold additional state between get_block calls and
223 * dio_complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224 */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200225static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret,
226 bool is_async)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227{
Zach Brown6d544bb2006-12-10 02:20:54 -0800228 ssize_t transferred = 0;
229
Zach Brown8459d862006-12-10 02:21:05 -0800230 /*
231 * AIO submission can race with bio completion to get here while
232 * expecting to have the last io completed by bio completion.
233 * In that case -EIOCBQUEUED is in fact not an error we want
234 * to preserve through this call.
235 */
236 if (ret == -EIOCBQUEUED)
237 ret = 0;
238
Zach Brown6d544bb2006-12-10 02:20:54 -0800239 if (dio->result) {
240 transferred = dio->result;
241
242 /* Check for short read case */
243 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
244 transferred = dio->i_size - offset;
245 }
246
Zach Brown6d544bb2006-12-10 02:20:54 -0800247 if (ret == 0)
248 ret = dio->page_errors;
249 if (ret == 0)
250 ret = dio->io_error;
251 if (ret == 0)
252 ret = transferred;
253
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200254 if (dio->end_io && dio->result)
255 dio->end_io(dio->iocb, offset, transferred, dio->private);
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000256
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200257 inode_dio_done(dio->inode);
Christoph Hellwig02afc272013-09-04 15:04:40 +0200258 if (is_async) {
259 if (dio->rw & WRITE) {
260 int err;
261
262 err = generic_write_sync(dio->iocb->ki_filp, offset,
263 transferred);
264 if (err < 0 && ret > 0)
265 ret = err;
266 }
267
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200268 aio_complete(dio->iocb, ret, 0);
Christoph Hellwig02afc272013-09-04 15:04:40 +0200269 }
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200270
271 kmem_cache_free(dio_cache, dio);
Zach Brown6d544bb2006-12-10 02:20:54 -0800272 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273}
274
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200275static void dio_aio_complete_work(struct work_struct *work)
276{
277 struct dio *dio = container_of(work, struct dio, complete_work);
278
279 dio_complete(dio, dio->iocb->ki_pos, 0, true);
280}
281
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282static int dio_bio_complete(struct dio *dio, struct bio *bio);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200283
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284/*
285 * Asynchronous IO callback.
286 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200287static void dio_bio_end_aio(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288{
289 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800290 unsigned long remaining;
291 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293 /* cleanup the bio */
294 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800295
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800296 spin_lock_irqsave(&dio->bio_lock, flags);
297 remaining = --dio->refcount;
298 if (remaining == 1 && dio->waiter)
Zach Brown20258b2b2006-12-10 02:21:01 -0800299 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800300 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b2b2006-12-10 02:21:01 -0800301
Zach Brown8459d862006-12-10 02:21:05 -0800302 if (remaining == 0) {
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200303 if (dio->result && dio->defer_completion) {
304 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
305 queue_work(dio->inode->i_sb->s_dio_done_wq,
306 &dio->complete_work);
307 } else {
308 dio_complete(dio, dio->iocb->ki_pos, 0, true);
309 }
Zach Brown8459d862006-12-10 02:21:05 -0800310 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311}
312
313/*
314 * The BIO completion handler simply queues the BIO up for the process-context
315 * handler.
316 *
317 * During I/O bi_private points at the dio. After I/O, bi_private is used to
318 * implement a singly-linked list of completed BIOs, at dio->bio_list.
319 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200320static void dio_bio_end_io(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321{
322 struct dio *dio = bio->bi_private;
323 unsigned long flags;
324
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 spin_lock_irqsave(&dio->bio_lock, flags);
326 bio->bi_private = dio->bio_list;
327 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800328 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329 wake_up_process(dio->waiter);
330 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331}
332
Josef Bacikfacd07b2010-05-23 11:00:55 -0400333/**
334 * dio_end_io - handle the end io action for the given bio
335 * @bio: The direct io bio thats being completed
336 * @error: Error if there was one
337 *
338 * This is meant to be called by any filesystem that uses their own dio_submit_t
339 * so that the DIO specific endio actions are dealt with after the filesystem
340 * has done it's completion work.
341 */
342void dio_end_io(struct bio *bio, int error)
343{
344 struct dio *dio = bio->bi_private;
345
346 if (dio->is_async)
347 dio_bio_end_aio(bio, error);
348 else
349 dio_bio_end_io(bio, error);
350}
351EXPORT_SYMBOL_GPL(dio_end_io);
352
Andi Kleenba253fb2011-08-01 21:38:08 -0700353static inline void
Andi Kleeneb28be22011-08-01 21:38:03 -0700354dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
355 struct block_device *bdev,
356 sector_t first_sector, int nr_vecs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357{
358 struct bio *bio;
359
David Dillow20d96002011-01-20 14:44:22 -0800360 /*
361 * bio_alloc() is guaranteed to return a bio when called with
362 * __GFP_WAIT and we request a valid number of vectors.
363 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365
366 bio->bi_bdev = bdev;
Kent Overstreet4f024f32013-10-11 15:44:27 -0700367 bio->bi_iter.bi_sector = first_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368 if (dio->is_async)
369 bio->bi_end_io = dio_bio_end_aio;
370 else
371 bio->bi_end_io = dio_bio_end_io;
372
Andi Kleeneb28be22011-08-01 21:38:03 -0700373 sdio->bio = bio;
374 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375}
376
377/*
378 * In the AIO read case we speculatively dirty the pages before starting IO.
379 * During IO completion, any of these pages which happen to have been written
380 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800381 *
382 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700384static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385{
Andi Kleeneb28be22011-08-01 21:38:03 -0700386 struct bio *bio = sdio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800387 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388
389 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800390
391 spin_lock_irqsave(&dio->bio_lock, flags);
392 dio->refcount++;
393 spin_unlock_irqrestore(&dio->bio_lock, flags);
394
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 if (dio->is_async && dio->rw == READ)
396 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800397
Andi Kleeneb28be22011-08-01 21:38:03 -0700398 if (sdio->submit_io)
399 sdio->submit_io(dio->rw, bio, dio->inode,
400 sdio->logical_offset_in_bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400401 else
402 submit_bio(dio->rw, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403
Andi Kleeneb28be22011-08-01 21:38:03 -0700404 sdio->bio = NULL;
405 sdio->boundary = 0;
406 sdio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407}
408
409/*
410 * Release any resources in case of a failure
411 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700412static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413{
Al Viro7b2c99d2014-03-15 04:05:57 -0400414 while (sdio->head < sdio->tail)
415 page_cache_release(dio->pages[sdio->head++]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416}
417
418/*
Zach Brown02732012006-12-10 02:20:59 -0800419 * Wait for the next BIO to complete. Remove it and return it. NULL is
420 * returned once all BIOs have been completed. This must only be called once
421 * all bios have been issued so that dio->refcount can only decrease. This
422 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 */
424static struct bio *dio_await_one(struct dio *dio)
425{
426 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800427 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428
429 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800430
431 /*
432 * Wait as long as the list is empty and there are bios in flight. bio
433 * completion drops the count, maybe adds to the list, and wakes while
434 * holding the bio_lock so we don't need set_current_state()'s barrier
435 * and can call it after testing our condition.
436 */
437 while (dio->refcount > 1 && dio->bio_list == NULL) {
438 __set_current_state(TASK_UNINTERRUPTIBLE);
439 dio->waiter = current;
440 spin_unlock_irqrestore(&dio->bio_lock, flags);
441 io_schedule();
442 /* wake up sets us TASK_RUNNING */
443 spin_lock_irqsave(&dio->bio_lock, flags);
444 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 }
Zach Brown02732012006-12-10 02:20:59 -0800446 if (dio->bio_list) {
447 bio = dio->bio_list;
448 dio->bio_list = bio->bi_private;
449 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 spin_unlock_irqrestore(&dio->bio_lock, flags);
451 return bio;
452}
453
454/*
455 * Process one completed BIO. No locks are held.
456 */
457static int dio_bio_complete(struct dio *dio, struct bio *bio)
458{
459 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
Kent Overstreetcb34e052012-09-05 15:22:02 -0700460 struct bio_vec *bvec;
461 unsigned i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462
463 if (!uptodate)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800464 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465
466 if (dio->is_async && dio->rw == READ) {
467 bio_check_pages_dirty(bio); /* transfers ownership */
468 } else {
Kent Overstreetcb34e052012-09-05 15:22:02 -0700469 bio_for_each_segment_all(bvec, bio, i) {
470 struct page *page = bvec->bv_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471
472 if (dio->rw == READ && !PageCompound(page))
473 set_page_dirty_lock(page);
474 page_cache_release(page);
475 }
476 bio_put(bio);
477 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 return uptodate ? 0 : -EIO;
479}
480
481/*
Zach Brown02732012006-12-10 02:20:59 -0800482 * Wait on and process all in-flight BIOs. This must only be called once
483 * all bios have been issued so that the refcount can only decrease.
484 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200485 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800486 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800488static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489{
Zach Brown02732012006-12-10 02:20:59 -0800490 struct bio *bio;
491 do {
492 bio = dio_await_one(dio);
493 if (bio)
494 dio_bio_complete(dio, bio);
495 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496}
497
498/*
499 * A really large O_DIRECT read or write can generate a lot of BIOs. So
500 * to keep the memory consumption sane we periodically reap any completed BIOs
501 * during the BIO generation phase.
502 *
503 * This also helps to limit the peak amount of pinned userspace memory.
504 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700505static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506{
507 int ret = 0;
508
Andi Kleeneb28be22011-08-01 21:38:03 -0700509 if (sdio->reap_counter++ >= 64) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 while (dio->bio_list) {
511 unsigned long flags;
512 struct bio *bio;
513 int ret2;
514
515 spin_lock_irqsave(&dio->bio_lock, flags);
516 bio = dio->bio_list;
517 dio->bio_list = bio->bi_private;
518 spin_unlock_irqrestore(&dio->bio_lock, flags);
519 ret2 = dio_bio_complete(dio, bio);
520 if (ret == 0)
521 ret = ret2;
522 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700523 sdio->reap_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524 }
525 return ret;
526}
527
528/*
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200529 * Create workqueue for deferred direct IO completions. We allocate the
530 * workqueue when it's first needed. This avoids creating workqueue for
531 * filesystems that don't need it and also allows us to create the workqueue
532 * late enough so the we can include s_id in the name of the workqueue.
533 */
534static int sb_init_dio_done_wq(struct super_block *sb)
535{
Olof Johansson45150c42013-09-09 10:34:23 -0700536 struct workqueue_struct *old;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200537 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
538 WQ_MEM_RECLAIM, 0,
539 sb->s_id);
540 if (!wq)
541 return -ENOMEM;
542 /*
543 * This has to be atomic as more DIOs can race to create the workqueue
544 */
Olof Johansson45150c42013-09-09 10:34:23 -0700545 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200546 /* Someone created workqueue before us? Free ours... */
Olof Johansson45150c42013-09-09 10:34:23 -0700547 if (old)
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200548 destroy_workqueue(wq);
549 return 0;
550}
551
552static int dio_set_defer_completion(struct dio *dio)
553{
554 struct super_block *sb = dio->inode->i_sb;
555
556 if (dio->defer_completion)
557 return 0;
558 dio->defer_completion = true;
559 if (!sb->s_dio_done_wq)
560 return sb_init_dio_done_wq(sb);
561 return 0;
562}
563
564/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565 * Call into the fs to map some more disk blocks. We record the current number
Andi Kleeneb28be22011-08-01 21:38:03 -0700566 * of available blocks at sdio->blocks_available. These are in units of the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 * fs blocksize, (1 << inode->i_blkbits).
568 *
569 * The fs is allowed to map lots of blocks at once. If it wants to do that,
570 * it uses the passed inode-relative block number as the file offset, as usual.
571 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800572 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573 * has remaining to do. The fs should not map more than this number of blocks.
574 *
575 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
576 * indicate how much contiguous disk space has been made available at
577 * bh->b_blocknr.
578 *
579 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
580 * This isn't very efficient...
581 *
582 * In the case of filesystem holes: the fs may return an arbitrarily-large
583 * hole by returning an appropriate value in b_size and by clearing
584 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800585 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586 */
Andi Kleen18772642011-08-01 21:38:07 -0700587static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
588 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589{
590 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
Tao Maae55e1a2012-01-12 17:20:33 -0800592 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593 unsigned long fs_count; /* Number of filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594 int create;
Linus Torvaldsab738572012-11-29 12:27:00 -0800595 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596
597 /*
598 * If there was a memory error and we've overwritten all the
599 * mapped blocks then we can now return that memory error
600 */
601 ret = dio->page_errors;
602 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700603 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
604 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
Tao Maae55e1a2012-01-12 17:20:33 -0800605 fs_endblk = (sdio->final_block_in_request - 1) >>
606 sdio->blkfactor;
607 fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608
Nathan Scott3c674e72006-03-29 09:26:15 +1000609 map_bh->b_state = 0;
Linus Torvaldsab738572012-11-29 12:27:00 -0800610 map_bh->b_size = fs_count << i_blkbits;
Nathan Scott3c674e72006-03-29 09:26:15 +1000611
Christoph Hellwig5fe878ae2009-12-15 16:47:50 -0800612 /*
613 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
614 * forbid block creations: only overwrites are permitted.
615 * We will return early to the caller once we see an
616 * unmapped buffer head returned, and the caller will fall
617 * back to buffered I/O.
618 *
619 * Otherwise the decision is left to the get_blocks method,
620 * which may decide to handle it or also return an unmapped
621 * buffer head.
622 */
Jens Axboeb31dc662006-06-13 08:26:10 +0200623 create = dio->rw & WRITE;
Christoph Hellwig5fe878ae2009-12-15 16:47:50 -0800624 if (dio->flags & DIO_SKIP_HOLES) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700625 if (sdio->block_in_file < (i_size_read(dio->inode) >>
626 sdio->blkbits))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000629
Andi Kleeneb28be22011-08-01 21:38:03 -0700630 ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631 map_bh, create);
Andi Kleen18772642011-08-01 21:38:07 -0700632
633 /* Store for completion */
634 dio->private = map_bh->b_private;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200635
636 if (ret == 0 && buffer_defer_completion(map_bh))
637 ret = dio_set_defer_completion(dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638 }
639 return ret;
640}
641
642/*
643 * There is no bio. Make one now.
644 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700645static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
646 sector_t start_sector, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647{
648 sector_t sector;
649 int ret, nr_pages;
650
Andi Kleeneb28be22011-08-01 21:38:03 -0700651 ret = dio_bio_reap(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 if (ret)
653 goto out;
Andi Kleeneb28be22011-08-01 21:38:03 -0700654 sector = start_sector << (sdio->blkbits - 9);
Andi Kleen18772642011-08-01 21:38:07 -0700655 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 BUG_ON(nr_pages <= 0);
Andi Kleen18772642011-08-01 21:38:07 -0700657 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
Andi Kleeneb28be22011-08-01 21:38:03 -0700658 sdio->boundary = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659out:
660 return ret;
661}
662
663/*
664 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
665 * that was successful then update final_block_in_bio and take a ref against
666 * the just-added page.
667 *
668 * Return zero on success. Non-zero means the caller needs to start a new BIO.
669 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700670static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671{
672 int ret;
673
Andi Kleeneb28be22011-08-01 21:38:03 -0700674 ret = bio_add_page(sdio->bio, sdio->cur_page,
675 sdio->cur_page_len, sdio->cur_page_offset);
676 if (ret == sdio->cur_page_len) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 /*
678 * Decrement count only, if we are done with this page
679 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700680 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
681 sdio->pages_in_io--;
682 page_cache_get(sdio->cur_page);
683 sdio->final_block_in_bio = sdio->cur_page_block +
684 (sdio->cur_page_len >> sdio->blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685 ret = 0;
686 } else {
687 ret = 1;
688 }
689 return ret;
690}
691
692/*
693 * Put cur_page under IO. The section of cur_page which is described by
694 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
695 * starts on-disk at cur_page_block.
696 *
697 * We take a ref against the page here (on behalf of its presence in the bio).
698 *
699 * The caller of this function is responsible for removing cur_page from the
700 * dio, and for dropping the refcount which came from that presence.
701 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700702static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
703 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704{
705 int ret = 0;
706
Andi Kleeneb28be22011-08-01 21:38:03 -0700707 if (sdio->bio) {
708 loff_t cur_offset = sdio->cur_page_fs_offset;
709 loff_t bio_next_offset = sdio->logical_offset_in_bio +
Kent Overstreet4f024f32013-10-11 15:44:27 -0700710 sdio->bio->bi_iter.bi_size;
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400711
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400713 * See whether this new request is contiguous with the old.
714 *
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900715 * Btrfs cannot handle having logically non-contiguous requests
716 * submitted. For example if you have
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400717 *
718 * Logical: [0-4095][HOLE][8192-12287]
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900719 * Physical: [0-4095] [4096-8191]
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400720 *
721 * We cannot submit those pages together as one BIO. So if our
722 * current logical offset in the file does not equal what would
723 * be the next logical offset in the bio, submit the bio we
724 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700726 if (sdio->final_block_in_bio != sdio->cur_page_block ||
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400727 cur_offset != bio_next_offset)
Andi Kleeneb28be22011-08-01 21:38:03 -0700728 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 }
730
Andi Kleeneb28be22011-08-01 21:38:03 -0700731 if (sdio->bio == NULL) {
Andi Kleen18772642011-08-01 21:38:07 -0700732 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 if (ret)
734 goto out;
735 }
736
Andi Kleeneb28be22011-08-01 21:38:03 -0700737 if (dio_bio_add_page(sdio) != 0) {
738 dio_bio_submit(dio, sdio);
Andi Kleen18772642011-08-01 21:38:07 -0700739 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700741 ret = dio_bio_add_page(sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742 BUG_ON(ret != 0);
743 }
744 }
745out:
746 return ret;
747}
748
749/*
750 * An autonomous function to put a chunk of a page under deferred IO.
751 *
752 * The caller doesn't actually know (or care) whether this piece of page is in
753 * a BIO, or is under IO or whatever. We just take care of all possible
754 * situations here. The separation between the logic of do_direct_IO() and
755 * that of submit_page_section() is important for clarity. Please don't break.
756 *
757 * The chunk of page starts on-disk at blocknr.
758 *
759 * We perform deferred IO, by recording the last-submitted page inside our
760 * private part of the dio structure. If possible, we just expand the IO
761 * across that page here.
762 *
763 * If that doesn't work out then we put the old page into the bio and add this
764 * page to the dio instead.
765 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700766static inline int
Andi Kleeneb28be22011-08-01 21:38:03 -0700767submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
Andi Kleen18772642011-08-01 21:38:07 -0700768 unsigned offset, unsigned len, sector_t blocknr,
769 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770{
771 int ret = 0;
772
Andrew Morton98c4d572006-12-10 02:19:47 -0800773 if (dio->rw & WRITE) {
774 /*
775 * Read accounting is performed in submit_bio()
776 */
777 task_io_account_write(len);
778 }
779
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 /*
781 * Can we just grow the current page's presence in the dio?
782 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700783 if (sdio->cur_page == page &&
784 sdio->cur_page_offset + sdio->cur_page_len == offset &&
785 sdio->cur_page_block +
786 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
787 sdio->cur_page_len += len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 goto out;
789 }
790
791 /*
792 * If there's a deferred page already there then send it.
793 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700794 if (sdio->cur_page) {
Andi Kleen18772642011-08-01 21:38:07 -0700795 ret = dio_send_cur_page(dio, sdio, map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -0700796 page_cache_release(sdio->cur_page);
797 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 if (ret)
Jan Karab1058b92013-04-29 15:06:18 -0700799 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 }
801
802 page_cache_get(page); /* It is in dio */
Andi Kleeneb28be22011-08-01 21:38:03 -0700803 sdio->cur_page = page;
804 sdio->cur_page_offset = offset;
805 sdio->cur_page_len = len;
806 sdio->cur_page_block = blocknr;
807 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808out:
Jan Karab1058b92013-04-29 15:06:18 -0700809 /*
810 * If sdio->boundary then we want to schedule the IO now to
811 * avoid metadata seeks.
812 */
813 if (sdio->boundary) {
814 ret = dio_send_cur_page(dio, sdio, map_bh);
815 dio_bio_submit(dio, sdio);
816 page_cache_release(sdio->cur_page);
817 sdio->cur_page = NULL;
818 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819 return ret;
820}
821
822/*
823 * Clean any dirty buffers in the blockdev mapping which alias newly-created
824 * file blocks. Only called for S_ISREG files - blockdevs do not set
825 * buffer_new
826 */
Andi Kleen18772642011-08-01 21:38:07 -0700827static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828{
829 unsigned i;
830 unsigned nblocks;
831
Andi Kleen18772642011-08-01 21:38:07 -0700832 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833
834 for (i = 0; i < nblocks; i++) {
Andi Kleen18772642011-08-01 21:38:07 -0700835 unmap_underlying_metadata(map_bh->b_bdev,
836 map_bh->b_blocknr + i);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 }
838}
839
840/*
841 * If we are not writing the entire block and get_block() allocated
842 * the block for us, we need to fill-in the unused portion of the
843 * block with zeros. This happens only if user-buffer, fileoffset or
844 * io length is not filesystem block-size multiple.
845 *
846 * `end' is zero if we're doing the start of the IO, 1 at the end of the
847 * IO.
848 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700849static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
850 int end, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851{
852 unsigned dio_blocks_per_fs_block;
853 unsigned this_chunk_blocks; /* In dio_blocks */
854 unsigned this_chunk_bytes;
855 struct page *page;
856
Andi Kleeneb28be22011-08-01 21:38:03 -0700857 sdio->start_zero_done = 1;
Andi Kleen18772642011-08-01 21:38:07 -0700858 if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 return;
860
Andi Kleeneb28be22011-08-01 21:38:03 -0700861 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
862 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863
864 if (!this_chunk_blocks)
865 return;
866
867 /*
868 * We need to zero out part of an fs block. It is either at the
869 * beginning or the end of the fs block.
870 */
871 if (end)
872 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
873
Andi Kleeneb28be22011-08-01 21:38:03 -0700874 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875
Nick Piggin557ed1f2007-10-16 01:24:40 -0700876 page = ZERO_PAGE(0);
Andi Kleeneb28be22011-08-01 21:38:03 -0700877 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700878 sdio->next_block_for_io, map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 return;
880
Andi Kleeneb28be22011-08-01 21:38:03 -0700881 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882}
883
884/*
885 * Walk the user pages, and the file, mapping blocks to disk and generating
886 * a sequence of (page,offset,len,block) mappings. These mappings are injected
887 * into submit_page_section(), which takes care of the next stage of submission
888 *
889 * Direct IO against a blockdev is different from a file. Because we can
890 * happily perform page-sized but 512-byte aligned IOs. It is important that
891 * blockdev IO be able to have fine alignment and large sizes.
892 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800893 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 * with the size of IO which is permitted at this offset and this i_blkbits.
895 *
896 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800897 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898 * fine alignment but still allows this function to work in PAGE_SIZE units.
899 */
Andi Kleen18772642011-08-01 21:38:07 -0700900static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
901 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902{
Andi Kleeneb28be22011-08-01 21:38:03 -0700903 const unsigned blkbits = sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904 int ret = 0;
905
Andi Kleeneb28be22011-08-01 21:38:03 -0700906 while (sdio->block_in_file < sdio->final_block_in_request) {
Al Viro7b2c99d2014-03-15 04:05:57 -0400907 struct page *page;
908 size_t from, to;
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300909
910 page = dio_get_page(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 if (IS_ERR(page)) {
912 ret = PTR_ERR(page);
913 goto out;
914 }
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300915 from = sdio->head ? 0 : sdio->from;
916 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
917 sdio->head++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918
Al Viro7b2c99d2014-03-15 04:05:57 -0400919 while (from < to) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 unsigned this_chunk_bytes; /* # of bytes mapped */
921 unsigned this_chunk_blocks; /* # of blocks */
922 unsigned u;
923
Andi Kleeneb28be22011-08-01 21:38:03 -0700924 if (sdio->blocks_available == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925 /*
926 * Need to go and map some more disk
927 */
928 unsigned long blkmask;
929 unsigned long dio_remainder;
930
Andi Kleen18772642011-08-01 21:38:07 -0700931 ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932 if (ret) {
933 page_cache_release(page);
934 goto out;
935 }
936 if (!buffer_mapped(map_bh))
937 goto do_holes;
938
Andi Kleeneb28be22011-08-01 21:38:03 -0700939 sdio->blocks_available =
940 map_bh->b_size >> sdio->blkbits;
941 sdio->next_block_for_io =
942 map_bh->b_blocknr << sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 if (buffer_new(map_bh))
Andi Kleen18772642011-08-01 21:38:07 -0700944 clean_blockdev_aliases(dio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945
Andi Kleeneb28be22011-08-01 21:38:03 -0700946 if (!sdio->blkfactor)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 goto do_holes;
948
Andi Kleeneb28be22011-08-01 21:38:03 -0700949 blkmask = (1 << sdio->blkfactor) - 1;
950 dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951
952 /*
953 * If we are at the start of IO and that IO
954 * starts partway into a fs-block,
955 * dio_remainder will be non-zero. If the IO
956 * is a read then we can simply advance the IO
957 * cursor to the first block which is to be
958 * read. But if the IO is a write and the
959 * block was newly allocated we cannot do that;
960 * the start of the fs block must be zeroed out
961 * on-disk
962 */
963 if (!buffer_new(map_bh))
Andi Kleeneb28be22011-08-01 21:38:03 -0700964 sdio->next_block_for_io += dio_remainder;
965 sdio->blocks_available -= dio_remainder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966 }
967do_holes:
968 /* Handle holes */
969 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800970 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971
972 /* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboeb31dc662006-06-13 08:26:10 +0200973 if (dio->rw & WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 page_cache_release(page);
975 return -ENOTBLK;
976 }
977
Jeff Moyer35dc8162006-02-03 03:04:27 -0800978 /*
979 * Be sure to account for a partial block as the
980 * last block in the file
981 */
982 i_size_aligned = ALIGN(i_size_read(dio->inode),
983 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -0700984 if (sdio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -0800985 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 /* We hit eof */
987 page_cache_release(page);
988 goto out;
989 }
Al Viro7b2c99d2014-03-15 04:05:57 -0400990 zero_user(page, from, 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -0700991 sdio->block_in_file++;
Al Viro7b2c99d2014-03-15 04:05:57 -0400992 from += 1 << blkbits;
Al Viro3320c602014-03-10 02:30:55 -0400993 dio->result += 1 << blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994 goto next_block;
995 }
996
997 /*
998 * If we're performing IO which has an alignment which
999 * is finer than the underlying fs, go check to see if
1000 * we must zero out the start of this block.
1001 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001002 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
Andi Kleen18772642011-08-01 21:38:07 -07001003 dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004
1005 /*
1006 * Work out, in this_chunk_blocks, how much disk we
1007 * can add to this page
1008 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001009 this_chunk_blocks = sdio->blocks_available;
Al Viro7b2c99d2014-03-15 04:05:57 -04001010 u = (to - from) >> blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011 if (this_chunk_blocks > u)
1012 this_chunk_blocks = u;
Andi Kleeneb28be22011-08-01 21:38:03 -07001013 u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014 if (this_chunk_blocks > u)
1015 this_chunk_blocks = u;
1016 this_chunk_bytes = this_chunk_blocks << blkbits;
1017 BUG_ON(this_chunk_bytes == 0);
1018
Jan Kara092c8d42013-04-29 15:06:17 -07001019 if (this_chunk_blocks == sdio->blocks_available)
1020 sdio->boundary = buffer_boundary(map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -07001021 ret = submit_page_section(dio, sdio, page,
Al Viro7b2c99d2014-03-15 04:05:57 -04001022 from,
Andi Kleeneb28be22011-08-01 21:38:03 -07001023 this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -07001024 sdio->next_block_for_io,
1025 map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026 if (ret) {
1027 page_cache_release(page);
1028 goto out;
1029 }
Andi Kleeneb28be22011-08-01 21:38:03 -07001030 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031
Andi Kleeneb28be22011-08-01 21:38:03 -07001032 sdio->block_in_file += this_chunk_blocks;
Al Viro7b2c99d2014-03-15 04:05:57 -04001033 from += this_chunk_bytes;
1034 dio->result += this_chunk_bytes;
Andi Kleeneb28be22011-08-01 21:38:03 -07001035 sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036next_block:
Andi Kleeneb28be22011-08-01 21:38:03 -07001037 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1038 if (sdio->block_in_file == sdio->final_block_in_request)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039 break;
1040 }
1041
1042 /* Drop the ref which was taken in get_user_pages() */
1043 page_cache_release(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044 }
1045out:
1046 return ret;
1047}
1048
Andi Kleen847cc632011-08-01 21:38:09 -07001049static inline int drop_refcount(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050{
Andi Kleen847cc632011-08-01 21:38:09 -07001051 int ret2;
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001052 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053
Zach Brown8459d862006-12-10 02:21:05 -08001054 /*
1055 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001056 * operation. AIO can if it was a broken operation described above or
1057 * in fact if all the bios race to complete before we get here. In
1058 * that case dio_complete() translates the EIOCBQUEUED into the proper
1059 * return code that the caller will hand to aio_complete().
1060 *
1061 * This is managed by the bio_lock instead of being an atomic_t so that
1062 * completion paths can drop their ref and use the remaining count to
1063 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001064 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001065 spin_lock_irqsave(&dio->bio_lock, flags);
1066 ret2 = --dio->refcount;
1067 spin_unlock_irqrestore(&dio->bio_lock, flags);
Andi Kleen847cc632011-08-01 21:38:09 -07001068 return ret2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069}
1070
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001071/*
1072 * This is a library function for use by filesystem drivers.
1073 *
1074 * The locking rules are governed by the flags parameter:
1075 * - if the flags value contains DIO_LOCKING we use a fancy locking
1076 * scheme for dumb filesystems.
1077 * For writes this function is called under i_mutex and returns with
1078 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1079 * taken and dropped again before returning.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001080 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1081 * internal locking but rather rely on the filesystem to synchronize
1082 * direct I/O reads/writes versus each other and truncate.
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001083 *
1084 * To help with locking against truncate we incremented the i_dio_count
1085 * counter before starting direct I/O, and decrement it once we are done.
1086 * Truncate can wait for it to reach zero to provide exclusion. It is
1087 * expected that filesystem provide exclusion between new direct I/O
1088 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1089 * but other filesystems need to take care of this on their own.
Andi Kleenba253fb2011-08-01 21:38:08 -07001090 *
1091 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1092 * is always inlined. Otherwise gcc is unable to split the structure into
1093 * individual fields and will generate much worse code. This is important
1094 * for the whole file.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001095 */
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001096static inline ssize_t
1097do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
Al Viro31b14032014-03-05 01:33:16 -05001098 struct block_device *bdev, struct iov_iter *iter, loff_t offset,
1099 get_block_t get_block, dio_iodone_t end_io,
Josef Bacikfacd07b2010-05-23 11:00:55 -04001100 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101{
Linus Torvaldsab738572012-11-29 12:27:00 -08001102 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1103 unsigned blkbits = i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104 unsigned blocksize_mask = (1 << blkbits) - 1;
1105 ssize_t retval = -EINVAL;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001106 size_t count = iov_iter_count(iter);
1107 loff_t end = offset + count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108 struct dio *dio;
Andi Kleeneb28be22011-08-01 21:38:03 -07001109 struct dio_submit sdio = { 0, };
Andi Kleen847cc632011-08-01 21:38:09 -07001110 struct buffer_head map_bh = { 0, };
Fengguang Wu647d1e42012-08-09 15:23:09 +02001111 struct blk_plug plug;
Al Viro886a3912014-03-05 13:50:45 -05001112 unsigned long align = offset | iov_iter_alignment(iter);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113
1114 if (rw & WRITE)
Jens Axboe721a9602011-03-09 11:56:30 +01001115 rw = WRITE_ODIRECT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001117 /*
1118 * Avoid references to bdev if not absolutely needed to give
1119 * the early prefetch in the caller enough time.
1120 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121
Al Viro886a3912014-03-05 13:50:45 -05001122 if (align & blocksize_mask) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001124 blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 blocksize_mask = (1 << blkbits) - 1;
Al Viro886a3912014-03-05 13:50:45 -05001126 if (align & blocksize_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 goto out;
1128 }
1129
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001130 /* watch out for a 0 len io from a tricksy fs */
Al Viro886a3912014-03-05 13:50:45 -05001131 if (rw == READ && !iov_iter_count(iter))
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001132 return 0;
1133
Andi Kleen6e8267f2011-08-01 21:38:06 -07001134 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 retval = -ENOMEM;
1136 if (!dio)
1137 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001138 /*
1139 * Believe it or not, zeroing out the page array caused a .5%
1140 * performance regression in a database benchmark. So, we take
1141 * care to only zero out what's needed.
1142 */
1143 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144
Christoph Hellwig5fe878ae2009-12-15 16:47:50 -08001145 dio->flags = flags;
1146 if (dio->flags & DIO_LOCKING) {
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001147 if (rw == READ) {
Christoph Hellwig5fe878ae2009-12-15 16:47:50 -08001148 struct address_space *mapping =
1149 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150
Christoph Hellwig5fe878ae2009-12-15 16:47:50 -08001151 /* will be released by direct_io_worker */
1152 mutex_lock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153
1154 retval = filemap_write_and_wait_range(mapping, offset,
1155 end - 1);
1156 if (retval) {
Christoph Hellwig5fe878ae2009-12-15 16:47:50 -08001157 mutex_unlock(&inode->i_mutex);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001158 kmem_cache_free(dio_cache, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159 goto out;
1160 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162 }
1163
1164 /*
Christoph Hellwig60392572014-02-10 10:27:11 +11001165 * For file extending writes updating i_size before data writeouts
1166 * complete can expose uninitialized blocks in dumb filesystems.
1167 * In that case we need to wait for I/O completion even if asked
1168 * for an asynchronous write.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169 */
Christoph Hellwig60392572014-02-10 10:27:11 +11001170 if (is_sync_kiocb(iocb))
1171 dio->is_async = false;
1172 else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
1173 (rw & WRITE) && end > i_size_read(inode))
1174 dio->is_async = false;
1175 else
1176 dio->is_async = true;
1177
Andi Kleen847cc632011-08-01 21:38:09 -07001178 dio->inode = inode;
1179 dio->rw = rw;
Christoph Hellwig02afc272013-09-04 15:04:40 +02001180
1181 /*
1182 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1183 * so that we can call ->fsync.
1184 */
1185 if (dio->is_async && (rw & WRITE) &&
1186 ((iocb->ki_filp->f_flags & O_DSYNC) ||
1187 IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1188 retval = dio_set_defer_completion(dio);
1189 if (retval) {
1190 /*
1191 * We grab i_mutex only for reads so we don't have
1192 * to release it here
1193 */
1194 kmem_cache_free(dio_cache, dio);
1195 goto out;
1196 }
1197 }
1198
1199 /*
1200 * Will be decremented at I/O completion time.
1201 */
1202 atomic_inc(&inode->i_dio_count);
1203
1204 retval = 0;
Andi Kleen847cc632011-08-01 21:38:09 -07001205 sdio.blkbits = blkbits;
Linus Torvaldsab738572012-11-29 12:27:00 -08001206 sdio.blkfactor = i_blkbits - blkbits;
Andi Kleen847cc632011-08-01 21:38:09 -07001207 sdio.block_in_file = offset >> blkbits;
1208
1209 sdio.get_block = get_block;
1210 dio->end_io = end_io;
1211 sdio.submit_io = submit_io;
1212 sdio.final_block_in_bio = -1;
1213 sdio.next_block_for_io = -1;
1214
1215 dio->iocb = iocb;
1216 dio->i_size = i_size_read(inode);
1217
1218 spin_lock_init(&dio->bio_lock);
1219 dio->refcount = 1;
1220
Al Viro7b2c99d2014-03-15 04:05:57 -04001221 sdio.iter = iter;
1222 sdio.final_block_in_request =
1223 (offset + iov_iter_count(iter)) >> blkbits;
1224
Andi Kleen847cc632011-08-01 21:38:09 -07001225 /*
1226 * In case of non-aligned buffers, we may need 2 more
1227 * pages since we need to zero out first and last block.
1228 */
1229 if (unlikely(sdio.blkfactor))
1230 sdio.pages_in_io = 2;
1231
Al Virof67da302014-03-19 01:16:16 -04001232 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
Andi Kleen847cc632011-08-01 21:38:09 -07001233
Fengguang Wu647d1e42012-08-09 15:23:09 +02001234 blk_start_plug(&plug);
1235
Al Viro7b2c99d2014-03-15 04:05:57 -04001236 retval = do_direct_IO(dio, &sdio, &map_bh);
1237 if (retval)
1238 dio_cleanup(dio, &sdio);
Andi Kleen847cc632011-08-01 21:38:09 -07001239
1240 if (retval == -ENOTBLK) {
1241 /*
1242 * The remaining part of the request will be
1243 * be handled by buffered I/O when we return
1244 */
1245 retval = 0;
1246 }
1247 /*
1248 * There may be some unwritten disk at the end of a part-written
1249 * fs-block-sized block. Go zero that now.
1250 */
1251 dio_zero_block(dio, &sdio, 1, &map_bh);
1252
1253 if (sdio.cur_page) {
1254 ssize_t ret2;
1255
1256 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1257 if (retval == 0)
1258 retval = ret2;
1259 page_cache_release(sdio.cur_page);
1260 sdio.cur_page = NULL;
1261 }
1262 if (sdio.bio)
1263 dio_bio_submit(dio, &sdio);
1264
Fengguang Wu647d1e42012-08-09 15:23:09 +02001265 blk_finish_plug(&plug);
1266
Andi Kleen847cc632011-08-01 21:38:09 -07001267 /*
1268 * It is possible that, we return short IO due to end of file.
1269 * In that case, we need to release all the pages we got hold on.
1270 */
1271 dio_cleanup(dio, &sdio);
1272
1273 /*
1274 * All block lookups have been performed. For READ requests
1275 * we can let i_mutex go now that its achieved its purpose
1276 * of protecting us from looking up uninitialized blocks.
1277 */
1278 if (rw == READ && (dio->flags & DIO_LOCKING))
1279 mutex_unlock(&dio->inode->i_mutex);
1280
1281 /*
1282 * The only time we want to leave bios in flight is when a successful
1283 * partial aio read or full aio write have been setup. In that case
1284 * bio completion will call aio_complete. The only time it's safe to
1285 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1286 * This had *better* be the only place that raises -EIOCBQUEUED.
1287 */
1288 BUG_ON(retval == -EIOCBQUEUED);
1289 if (dio->is_async && retval == 0 && dio->result &&
Christoph Hellwigaf436472014-07-30 07:18:48 -04001290 (rw == READ || dio->result == count))
Andi Kleen847cc632011-08-01 21:38:09 -07001291 retval = -EIOCBQUEUED;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001292 else
Andi Kleen847cc632011-08-01 21:38:09 -07001293 dio_await_completion(dio);
1294
1295 if (drop_refcount(dio) == 0) {
1296 retval = dio_complete(dio, offset, retval, false);
Andi Kleen847cc632011-08-01 21:38:09 -07001297 } else
1298 BUG_ON(retval != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001300out:
1301 return retval;
1302}
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001303
1304ssize_t
1305__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
Al Viro31b14032014-03-05 01:33:16 -05001306 struct block_device *bdev, struct iov_iter *iter, loff_t offset,
1307 get_block_t get_block, dio_iodone_t end_io,
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001308 dio_submit_t submit_io, int flags)
1309{
1310 /*
1311 * The block device state is needed in the end to finally
1312 * submit everything. Since it's likely to be cache cold
1313 * prefetch it here as first thing to hide some of the
1314 * latency.
1315 *
1316 * Attempt to prefetch the pieces we likely need later.
1317 */
1318 prefetch(&bdev->bd_disk->part_tbl);
1319 prefetch(bdev->bd_queue);
1320 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1321
Al Viro31b14032014-03-05 01:33:16 -05001322 return do_blockdev_direct_IO(rw, iocb, inode, bdev, iter, offset,
1323 get_block, end_io, submit_io, flags);
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001324}
1325
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326EXPORT_SYMBOL(__blockdev_direct_IO);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001327
1328static __init int dio_init(void)
1329{
1330 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1331 return 0;
1332}
1333module_init(dio_init)