blob: 602e8441bc0fb6b094ec96dfb3273ff73b8b1506 [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;
Jens Axboe15c4f632015-10-27 14:09:51 +0900112 blk_qc_t bio_cookie;
113 struct block_device *bio_bdev;
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700114 struct inode *inode;
Andi Kleeneb28be22011-08-01 21:38:03 -0700115 loff_t i_size; /* i_size when submitted */
116 dio_iodone_t *end_io; /* IO completion function */
Andi Kleeneb28be22011-08-01 21:38:03 -0700117
Andi Kleen18772642011-08-01 21:38:07 -0700118 void *private; /* copy from map_bh.b_private */
Andi Kleeneb28be22011-08-01 21:38:03 -0700119
120 /* BIO completion state */
121 spinlock_t bio_lock; /* protects BIO fields below */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700122 int page_errors; /* errno from get_user_pages() */
123 int is_async; /* is IO async ? */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200124 bool defer_completion; /* defer AIO completion to workqueue? */
Ming Lei53cbf3b2015-08-17 10:31:46 +0800125 bool should_dirty; /* if pages should be dirtied */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700126 int io_error; /* IO error in completion path */
Andi Kleeneb28be22011-08-01 21:38:03 -0700127 unsigned long refcount; /* direct_io_worker() and bios */
128 struct bio *bio_list; /* singly linked via bi_private */
129 struct task_struct *waiter; /* waiting task (NULL if none) */
130
131 /* AIO related stuff */
132 struct kiocb *iocb; /* kiocb */
Andi Kleeneb28be22011-08-01 21:38:03 -0700133 ssize_t result; /* IO result */
134
Jeff Moyer23aee092009-12-15 16:47:49 -0800135 /*
136 * pages[] (and any fields placed after it) are not zeroed out at
137 * allocation time. Don't add new fields after pages[] unless you
138 * wish that they not be zeroed.
139 */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200140 union {
141 struct page *pages[DIO_PAGES]; /* page buffer */
142 struct work_struct complete_work;/* deferred AIO completion */
143 };
Andi Kleen6e8267f2011-08-01 21:38:06 -0700144} ____cacheline_aligned_in_smp;
145
146static struct kmem_cache *dio_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147
148/*
149 * How many pages are in the queue?
150 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700151static inline unsigned dio_pages_present(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152{
Andi Kleeneb28be22011-08-01 21:38:03 -0700153 return sdio->tail - sdio->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154}
155
156/*
157 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
158 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700159static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160{
Al Viro7b2c99d2014-03-15 04:05:57 -0400161 ssize_t ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162
Miklos Szeredi2c809292014-09-24 17:09:11 +0200163 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
Al Viro7b2c99d2014-03-15 04:05:57 -0400164 &sdio->from);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165
Andi Kleeneb28be22011-08-01 21:38:03 -0700166 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
Nick Piggin557ed1f2007-10-16 01:24:40 -0700167 struct page *page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168 /*
169 * A memory fault, but the filesystem has some outstanding
170 * mapped blocks. We need to use those blocks up to avoid
171 * leaking stale data in the file.
172 */
173 if (dio->page_errors == 0)
174 dio->page_errors = ret;
Nick Pigginb5810032005-10-29 18:16:12 -0700175 page_cache_get(page);
176 dio->pages[0] = page;
Andi Kleeneb28be22011-08-01 21:38:03 -0700177 sdio->head = 0;
178 sdio->tail = 1;
Al Viro7b2c99d2014-03-15 04:05:57 -0400179 sdio->from = 0;
180 sdio->to = PAGE_SIZE;
181 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182 }
183
184 if (ret >= 0) {
Al Viro7b2c99d2014-03-15 04:05:57 -0400185 iov_iter_advance(sdio->iter, ret);
186 ret += sdio->from;
Andi Kleeneb28be22011-08-01 21:38:03 -0700187 sdio->head = 0;
Al Viro7b2c99d2014-03-15 04:05:57 -0400188 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
189 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
190 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 return ret;
193}
194
195/*
196 * Get another userspace page. Returns an ERR_PTR on error. Pages are
197 * buffered inside the dio so that we can call get_user_pages() against a
198 * decent number of pages, less frequently. To provide nicer use of the
199 * L1 cache.
200 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700201static inline struct page *dio_get_page(struct dio *dio,
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300202 struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203{
Andi Kleeneb28be22011-08-01 21:38:03 -0700204 if (dio_pages_present(sdio) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205 int ret;
206
Andi Kleeneb28be22011-08-01 21:38:03 -0700207 ret = dio_refill_pages(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208 if (ret)
209 return ERR_PTR(ret);
Andi Kleeneb28be22011-08-01 21:38:03 -0700210 BUG_ON(dio_pages_present(sdio) == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211 }
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300212 return dio->pages[sdio->head];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213}
214
Zach Brown6d544bb2006-12-10 02:20:54 -0800215/**
216 * dio_complete() - called when all DIO BIO I/O has been completed
217 * @offset: the byte offset in the file of the completed operation
218 *
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200219 * This drops i_dio_count, lets interested parties know that a DIO operation
220 * has completed, and calculates the resulting return code for the operation.
Zach Brown6d544bb2006-12-10 02:20:54 -0800221 *
222 * It lets the filesystem know if it registered an interest earlier via
223 * get_block. Pass the private field of the map buffer_head so that
224 * filesystems can use it to hold additional state between get_block calls and
225 * dio_complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226 */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200227static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret,
228 bool is_async)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229{
Zach Brown6d544bb2006-12-10 02:20:54 -0800230 ssize_t transferred = 0;
231
Zach Brown8459d862006-12-10 02:21:05 -0800232 /*
233 * AIO submission can race with bio completion to get here while
234 * expecting to have the last io completed by bio completion.
235 * In that case -EIOCBQUEUED is in fact not an error we want
236 * to preserve through this call.
237 */
238 if (ret == -EIOCBQUEUED)
239 ret = 0;
240
Zach Brown6d544bb2006-12-10 02:20:54 -0800241 if (dio->result) {
242 transferred = dio->result;
243
244 /* Check for short read case */
245 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
246 transferred = dio->i_size - offset;
247 }
248
Zach Brown6d544bb2006-12-10 02:20:54 -0800249 if (ret == 0)
250 ret = dio->page_errors;
251 if (ret == 0)
252 ret = dio->io_error;
253 if (ret == 0)
254 ret = transferred;
255
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200256 if (dio->end_io && dio->result)
257 dio->end_io(dio->iocb, offset, transferred, dio->private);
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000258
Jens Axboefe0f07d2015-04-15 17:05:48 -0600259 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
260 inode_dio_end(dio->inode);
261
Christoph Hellwig02afc272013-09-04 15:04:40 +0200262 if (is_async) {
263 if (dio->rw & WRITE) {
264 int err;
265
266 err = generic_write_sync(dio->iocb->ki_filp, offset,
267 transferred);
268 if (err < 0 && ret > 0)
269 ret = err;
270 }
271
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100272 dio->iocb->ki_complete(dio->iocb, ret, 0);
Christoph Hellwig02afc272013-09-04 15:04:40 +0200273 }
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200274
275 kmem_cache_free(dio_cache, dio);
Zach Brown6d544bb2006-12-10 02:20:54 -0800276 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277}
278
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200279static void dio_aio_complete_work(struct work_struct *work)
280{
281 struct dio *dio = container_of(work, struct dio, complete_work);
282
283 dio_complete(dio, dio->iocb->ki_pos, 0, true);
284}
285
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286static int dio_bio_complete(struct dio *dio, struct bio *bio);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200287
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288/*
289 * Asynchronous IO callback.
290 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200291static void dio_bio_end_aio(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292{
293 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800294 unsigned long remaining;
295 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297 /* cleanup the bio */
298 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800299
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800300 spin_lock_irqsave(&dio->bio_lock, flags);
301 remaining = --dio->refcount;
302 if (remaining == 1 && dio->waiter)
Zach Brown20258b22006-12-10 02:21:01 -0800303 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800304 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b22006-12-10 02:21:01 -0800305
Zach Brown8459d862006-12-10 02:21:05 -0800306 if (remaining == 0) {
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200307 if (dio->result && dio->defer_completion) {
308 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
309 queue_work(dio->inode->i_sb->s_dio_done_wq,
310 &dio->complete_work);
311 } else {
312 dio_complete(dio, dio->iocb->ki_pos, 0, true);
313 }
Zach Brown8459d862006-12-10 02:21:05 -0800314 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315}
316
317/*
318 * The BIO completion handler simply queues the BIO up for the process-context
319 * handler.
320 *
321 * During I/O bi_private points at the dio. After I/O, bi_private is used to
322 * implement a singly-linked list of completed BIOs, at dio->bio_list.
323 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200324static void dio_bio_end_io(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325{
326 struct dio *dio = bio->bi_private;
327 unsigned long flags;
328
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329 spin_lock_irqsave(&dio->bio_lock, flags);
330 bio->bi_private = dio->bio_list;
331 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800332 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333 wake_up_process(dio->waiter);
334 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335}
336
Josef Bacikfacd07b2010-05-23 11:00:55 -0400337/**
338 * dio_end_io - handle the end io action for the given bio
339 * @bio: The direct io bio thats being completed
340 * @error: Error if there was one
341 *
342 * This is meant to be called by any filesystem that uses their own dio_submit_t
343 * so that the DIO specific endio actions are dealt with after the filesystem
344 * has done it's completion work.
345 */
346void dio_end_io(struct bio *bio, int error)
347{
348 struct dio *dio = bio->bi_private;
349
350 if (dio->is_async)
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200351 dio_bio_end_aio(bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400352 else
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200353 dio_bio_end_io(bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400354}
355EXPORT_SYMBOL_GPL(dio_end_io);
356
Andi Kleenba253fb2011-08-01 21:38:08 -0700357static inline void
Andi Kleeneb28be22011-08-01 21:38:03 -0700358dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
359 struct block_device *bdev,
360 sector_t first_sector, int nr_vecs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361{
362 struct bio *bio;
363
David Dillow20d96002011-01-20 14:44:22 -0800364 /*
365 * bio_alloc() is guaranteed to return a bio when called with
Mel Gorman71baba42015-11-06 16:28:28 -0800366 * __GFP_RECLAIM and we request a valid number of vectors.
David Dillow20d96002011-01-20 14:44:22 -0800367 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369
370 bio->bi_bdev = bdev;
Kent Overstreet4f024f32013-10-11 15:44:27 -0700371 bio->bi_iter.bi_sector = first_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372 if (dio->is_async)
373 bio->bi_end_io = dio_bio_end_aio;
374 else
375 bio->bi_end_io = dio_bio_end_io;
376
Andi Kleeneb28be22011-08-01 21:38:03 -0700377 sdio->bio = bio;
378 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379}
380
381/*
382 * In the AIO read case we speculatively dirty the pages before starting IO.
383 * During IO completion, any of these pages which happen to have been written
384 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800385 *
386 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700388static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389{
Andi Kleeneb28be22011-08-01 21:38:03 -0700390 struct bio *bio = sdio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800391 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392
393 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800394
395 spin_lock_irqsave(&dio->bio_lock, flags);
396 dio->refcount++;
397 spin_unlock_irqrestore(&dio->bio_lock, flags);
398
Ming Lei53cbf3b2015-08-17 10:31:46 +0800399 if (dio->is_async && dio->rw == READ && dio->should_dirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800401
Jens Axboec1c53462015-11-10 10:14:38 -0700402 dio->bio_bdev = bio->bi_bdev;
403
Jens Axboe15c4f632015-10-27 14:09:51 +0900404 if (sdio->submit_io) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700405 sdio->submit_io(dio->rw, bio, dio->inode,
406 sdio->logical_offset_in_bio);
Jens Axboe15c4f632015-10-27 14:09:51 +0900407 dio->bio_cookie = BLK_QC_T_NONE;
Jens Axboec1c53462015-11-10 10:14:38 -0700408 } else
Jens Axboe15c4f632015-10-27 14:09:51 +0900409 dio->bio_cookie = submit_bio(dio->rw, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410
Andi Kleeneb28be22011-08-01 21:38:03 -0700411 sdio->bio = NULL;
412 sdio->boundary = 0;
413 sdio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414}
415
416/*
417 * Release any resources in case of a failure
418 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700419static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420{
Al Viro7b2c99d2014-03-15 04:05:57 -0400421 while (sdio->head < sdio->tail)
422 page_cache_release(dio->pages[sdio->head++]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423}
424
425/*
Zach Brown02732012006-12-10 02:20:59 -0800426 * Wait for the next BIO to complete. Remove it and return it. NULL is
427 * returned once all BIOs have been completed. This must only be called once
428 * all bios have been issued so that dio->refcount can only decrease. This
429 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430 */
431static struct bio *dio_await_one(struct dio *dio)
432{
433 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800434 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435
436 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800437
438 /*
439 * Wait as long as the list is empty and there are bios in flight. bio
440 * completion drops the count, maybe adds to the list, and wakes while
441 * holding the bio_lock so we don't need set_current_state()'s barrier
442 * and can call it after testing our condition.
443 */
444 while (dio->refcount > 1 && dio->bio_list == NULL) {
445 __set_current_state(TASK_UNINTERRUPTIBLE);
446 dio->waiter = current;
447 spin_unlock_irqrestore(&dio->bio_lock, flags);
Jens Axboe15c4f632015-10-27 14:09:51 +0900448 if (!blk_poll(bdev_get_queue(dio->bio_bdev), dio->bio_cookie))
449 io_schedule();
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800450 /* wake up sets us TASK_RUNNING */
451 spin_lock_irqsave(&dio->bio_lock, flags);
452 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453 }
Zach Brown02732012006-12-10 02:20:59 -0800454 if (dio->bio_list) {
455 bio = dio->bio_list;
456 dio->bio_list = bio->bi_private;
457 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 spin_unlock_irqrestore(&dio->bio_lock, flags);
459 return bio;
460}
461
462/*
463 * Process one completed BIO. No locks are held.
464 */
465static int dio_bio_complete(struct dio *dio, struct bio *bio)
466{
Kent Overstreetcb34e052012-09-05 15:22:02 -0700467 struct bio_vec *bvec;
468 unsigned i;
Sasha Levin9b81c842015-08-10 19:05:18 -0400469 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200471 if (bio->bi_error)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800472 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473
Ming Lei53cbf3b2015-08-17 10:31:46 +0800474 if (dio->is_async && dio->rw == READ && dio->should_dirty) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475 bio_check_pages_dirty(bio); /* transfers ownership */
Sasha Levin9b81c842015-08-10 19:05:18 -0400476 err = bio->bi_error;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 } else {
Kent Overstreetcb34e052012-09-05 15:22:02 -0700478 bio_for_each_segment_all(bvec, bio, i) {
479 struct page *page = bvec->bv_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480
Ming Lei53cbf3b2015-08-17 10:31:46 +0800481 if (dio->rw == READ && !PageCompound(page) &&
482 dio->should_dirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483 set_page_dirty_lock(page);
484 page_cache_release(page);
485 }
Sasha Levin9b81c842015-08-10 19:05:18 -0400486 err = bio->bi_error;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 bio_put(bio);
488 }
Sasha Levin9b81c842015-08-10 19:05:18 -0400489 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490}
491
492/*
Zach Brown02732012006-12-10 02:20:59 -0800493 * Wait on and process all in-flight BIOs. This must only be called once
494 * all bios have been issued so that the refcount can only decrease.
495 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200496 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800497 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800499static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500{
Zach Brown02732012006-12-10 02:20:59 -0800501 struct bio *bio;
502 do {
503 bio = dio_await_one(dio);
504 if (bio)
505 dio_bio_complete(dio, bio);
506 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507}
508
509/*
510 * A really large O_DIRECT read or write can generate a lot of BIOs. So
511 * to keep the memory consumption sane we periodically reap any completed BIOs
512 * during the BIO generation phase.
513 *
514 * This also helps to limit the peak amount of pinned userspace memory.
515 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700516static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517{
518 int ret = 0;
519
Andi Kleeneb28be22011-08-01 21:38:03 -0700520 if (sdio->reap_counter++ >= 64) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 while (dio->bio_list) {
522 unsigned long flags;
523 struct bio *bio;
524 int ret2;
525
526 spin_lock_irqsave(&dio->bio_lock, flags);
527 bio = dio->bio_list;
528 dio->bio_list = bio->bi_private;
529 spin_unlock_irqrestore(&dio->bio_lock, flags);
530 ret2 = dio_bio_complete(dio, bio);
531 if (ret == 0)
532 ret = ret2;
533 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700534 sdio->reap_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 }
536 return ret;
537}
538
539/*
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200540 * Create workqueue for deferred direct IO completions. We allocate the
541 * workqueue when it's first needed. This avoids creating workqueue for
542 * filesystems that don't need it and also allows us to create the workqueue
543 * late enough so the we can include s_id in the name of the workqueue.
544 */
545static int sb_init_dio_done_wq(struct super_block *sb)
546{
Olof Johansson45150c42013-09-09 10:34:23 -0700547 struct workqueue_struct *old;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200548 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
549 WQ_MEM_RECLAIM, 0,
550 sb->s_id);
551 if (!wq)
552 return -ENOMEM;
553 /*
554 * This has to be atomic as more DIOs can race to create the workqueue
555 */
Olof Johansson45150c42013-09-09 10:34:23 -0700556 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200557 /* Someone created workqueue before us? Free ours... */
Olof Johansson45150c42013-09-09 10:34:23 -0700558 if (old)
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200559 destroy_workqueue(wq);
560 return 0;
561}
562
563static int dio_set_defer_completion(struct dio *dio)
564{
565 struct super_block *sb = dio->inode->i_sb;
566
567 if (dio->defer_completion)
568 return 0;
569 dio->defer_completion = true;
570 if (!sb->s_dio_done_wq)
571 return sb_init_dio_done_wq(sb);
572 return 0;
573}
574
575/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576 * Call into the fs to map some more disk blocks. We record the current number
Andi Kleeneb28be22011-08-01 21:38:03 -0700577 * of available blocks at sdio->blocks_available. These are in units of the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578 * fs blocksize, (1 << inode->i_blkbits).
579 *
580 * The fs is allowed to map lots of blocks at once. If it wants to do that,
581 * it uses the passed inode-relative block number as the file offset, as usual.
582 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800583 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 * has remaining to do. The fs should not map more than this number of blocks.
585 *
586 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
587 * indicate how much contiguous disk space has been made available at
588 * bh->b_blocknr.
589 *
590 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
591 * This isn't very efficient...
592 *
593 * In the case of filesystem holes: the fs may return an arbitrarily-large
594 * hole by returning an appropriate value in b_size and by clearing
595 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800596 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597 */
Andi Kleen18772642011-08-01 21:38:07 -0700598static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
599 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600{
601 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
Tao Maae55e1a2012-01-12 17:20:33 -0800603 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604 unsigned long fs_count; /* Number of filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605 int create;
Linus Torvaldsab738572012-11-29 12:27:00 -0800606 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607
608 /*
609 * If there was a memory error and we've overwritten all the
610 * mapped blocks then we can now return that memory error
611 */
612 ret = dio->page_errors;
613 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700614 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
615 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
Tao Maae55e1a2012-01-12 17:20:33 -0800616 fs_endblk = (sdio->final_block_in_request - 1) >>
617 sdio->blkfactor;
618 fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619
Nathan Scott3c674e72006-03-29 09:26:15 +1000620 map_bh->b_state = 0;
Linus Torvaldsab738572012-11-29 12:27:00 -0800621 map_bh->b_size = fs_count << i_blkbits;
Nathan Scott3c674e72006-03-29 09:26:15 +1000622
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800623 /*
624 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
625 * forbid block creations: only overwrites are permitted.
626 * We will return early to the caller once we see an
627 * unmapped buffer head returned, and the caller will fall
628 * back to buffered I/O.
629 *
630 * Otherwise the decision is left to the get_blocks method,
631 * which may decide to handle it or also return an unmapped
632 * buffer head.
633 */
Jens Axboeb31dc662006-06-13 08:26:10 +0200634 create = dio->rw & WRITE;
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800635 if (dio->flags & DIO_SKIP_HOLES) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700636 if (sdio->block_in_file < (i_size_read(dio->inode) >>
637 sdio->blkbits))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000640
Andi Kleeneb28be22011-08-01 21:38:03 -0700641 ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642 map_bh, create);
Andi Kleen18772642011-08-01 21:38:07 -0700643
644 /* Store for completion */
645 dio->private = map_bh->b_private;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200646
647 if (ret == 0 && buffer_defer_completion(map_bh))
648 ret = dio_set_defer_completion(dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 }
650 return ret;
651}
652
653/*
654 * There is no bio. Make one now.
655 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700656static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
657 sector_t start_sector, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658{
659 sector_t sector;
660 int ret, nr_pages;
661
Andi Kleeneb28be22011-08-01 21:38:03 -0700662 ret = dio_bio_reap(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663 if (ret)
664 goto out;
Andi Kleeneb28be22011-08-01 21:38:03 -0700665 sector = start_sector << (sdio->blkbits - 9);
Kent Overstreetb54ffb72015-05-19 14:31:01 +0200666 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 BUG_ON(nr_pages <= 0);
Andi Kleen18772642011-08-01 21:38:07 -0700668 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
Andi Kleeneb28be22011-08-01 21:38:03 -0700669 sdio->boundary = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670out:
671 return ret;
672}
673
674/*
675 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
676 * that was successful then update final_block_in_bio and take a ref against
677 * the just-added page.
678 *
679 * Return zero on success. Non-zero means the caller needs to start a new BIO.
680 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700681static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682{
683 int ret;
684
Andi Kleeneb28be22011-08-01 21:38:03 -0700685 ret = bio_add_page(sdio->bio, sdio->cur_page,
686 sdio->cur_page_len, sdio->cur_page_offset);
687 if (ret == sdio->cur_page_len) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 /*
689 * Decrement count only, if we are done with this page
690 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700691 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
692 sdio->pages_in_io--;
693 page_cache_get(sdio->cur_page);
694 sdio->final_block_in_bio = sdio->cur_page_block +
695 (sdio->cur_page_len >> sdio->blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696 ret = 0;
697 } else {
698 ret = 1;
699 }
700 return ret;
701}
702
703/*
704 * Put cur_page under IO. The section of cur_page which is described by
705 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
706 * starts on-disk at cur_page_block.
707 *
708 * We take a ref against the page here (on behalf of its presence in the bio).
709 *
710 * The caller of this function is responsible for removing cur_page from the
711 * dio, and for dropping the refcount which came from that presence.
712 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700713static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
714 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715{
716 int ret = 0;
717
Andi Kleeneb28be22011-08-01 21:38:03 -0700718 if (sdio->bio) {
719 loff_t cur_offset = sdio->cur_page_fs_offset;
720 loff_t bio_next_offset = sdio->logical_offset_in_bio +
Kent Overstreet4f024f32013-10-11 15:44:27 -0700721 sdio->bio->bi_iter.bi_size;
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400722
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400724 * See whether this new request is contiguous with the old.
725 *
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900726 * Btrfs cannot handle having logically non-contiguous requests
727 * submitted. For example if you have
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400728 *
729 * Logical: [0-4095][HOLE][8192-12287]
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900730 * Physical: [0-4095] [4096-8191]
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400731 *
732 * We cannot submit those pages together as one BIO. So if our
733 * current logical offset in the file does not equal what would
734 * be the next logical offset in the bio, submit the bio we
735 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700737 if (sdio->final_block_in_bio != sdio->cur_page_block ||
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400738 cur_offset != bio_next_offset)
Andi Kleeneb28be22011-08-01 21:38:03 -0700739 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 }
741
Andi Kleeneb28be22011-08-01 21:38:03 -0700742 if (sdio->bio == NULL) {
Andi Kleen18772642011-08-01 21:38:07 -0700743 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 if (ret)
745 goto out;
746 }
747
Andi Kleeneb28be22011-08-01 21:38:03 -0700748 if (dio_bio_add_page(sdio) != 0) {
749 dio_bio_submit(dio, sdio);
Andi Kleen18772642011-08-01 21:38:07 -0700750 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700752 ret = dio_bio_add_page(sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 BUG_ON(ret != 0);
754 }
755 }
756out:
757 return ret;
758}
759
760/*
761 * An autonomous function to put a chunk of a page under deferred IO.
762 *
763 * The caller doesn't actually know (or care) whether this piece of page is in
764 * a BIO, or is under IO or whatever. We just take care of all possible
765 * situations here. The separation between the logic of do_direct_IO() and
766 * that of submit_page_section() is important for clarity. Please don't break.
767 *
768 * The chunk of page starts on-disk at blocknr.
769 *
770 * We perform deferred IO, by recording the last-submitted page inside our
771 * private part of the dio structure. If possible, we just expand the IO
772 * across that page here.
773 *
774 * If that doesn't work out then we put the old page into the bio and add this
775 * page to the dio instead.
776 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700777static inline int
Andi Kleeneb28be22011-08-01 21:38:03 -0700778submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
Andi Kleen18772642011-08-01 21:38:07 -0700779 unsigned offset, unsigned len, sector_t blocknr,
780 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781{
782 int ret = 0;
783
Andrew Morton98c4d572006-12-10 02:19:47 -0800784 if (dio->rw & WRITE) {
785 /*
786 * Read accounting is performed in submit_bio()
787 */
788 task_io_account_write(len);
789 }
790
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 /*
792 * Can we just grow the current page's presence in the dio?
793 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700794 if (sdio->cur_page == page &&
795 sdio->cur_page_offset + sdio->cur_page_len == offset &&
796 sdio->cur_page_block +
797 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
798 sdio->cur_page_len += len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 goto out;
800 }
801
802 /*
803 * If there's a deferred page already there then send it.
804 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700805 if (sdio->cur_page) {
Andi Kleen18772642011-08-01 21:38:07 -0700806 ret = dio_send_cur_page(dio, sdio, map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -0700807 page_cache_release(sdio->cur_page);
808 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809 if (ret)
Jan Karab1058b92013-04-29 15:06:18 -0700810 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 }
812
813 page_cache_get(page); /* It is in dio */
Andi Kleeneb28be22011-08-01 21:38:03 -0700814 sdio->cur_page = page;
815 sdio->cur_page_offset = offset;
816 sdio->cur_page_len = len;
817 sdio->cur_page_block = blocknr;
818 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819out:
Jan Karab1058b92013-04-29 15:06:18 -0700820 /*
821 * If sdio->boundary then we want to schedule the IO now to
822 * avoid metadata seeks.
823 */
824 if (sdio->boundary) {
825 ret = dio_send_cur_page(dio, sdio, map_bh);
826 dio_bio_submit(dio, sdio);
827 page_cache_release(sdio->cur_page);
828 sdio->cur_page = NULL;
829 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 return ret;
831}
832
833/*
834 * Clean any dirty buffers in the blockdev mapping which alias newly-created
835 * file blocks. Only called for S_ISREG files - blockdevs do not set
836 * buffer_new
837 */
Andi Kleen18772642011-08-01 21:38:07 -0700838static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839{
840 unsigned i;
841 unsigned nblocks;
842
Andi Kleen18772642011-08-01 21:38:07 -0700843 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844
845 for (i = 0; i < nblocks; i++) {
Andi Kleen18772642011-08-01 21:38:07 -0700846 unmap_underlying_metadata(map_bh->b_bdev,
847 map_bh->b_blocknr + i);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848 }
849}
850
851/*
852 * If we are not writing the entire block and get_block() allocated
853 * the block for us, we need to fill-in the unused portion of the
854 * block with zeros. This happens only if user-buffer, fileoffset or
855 * io length is not filesystem block-size multiple.
856 *
857 * `end' is zero if we're doing the start of the IO, 1 at the end of the
858 * IO.
859 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700860static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
861 int end, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862{
863 unsigned dio_blocks_per_fs_block;
864 unsigned this_chunk_blocks; /* In dio_blocks */
865 unsigned this_chunk_bytes;
866 struct page *page;
867
Andi Kleeneb28be22011-08-01 21:38:03 -0700868 sdio->start_zero_done = 1;
Andi Kleen18772642011-08-01 21:38:07 -0700869 if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870 return;
871
Andi Kleeneb28be22011-08-01 21:38:03 -0700872 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
873 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874
875 if (!this_chunk_blocks)
876 return;
877
878 /*
879 * We need to zero out part of an fs block. It is either at the
880 * beginning or the end of the fs block.
881 */
882 if (end)
883 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
884
Andi Kleeneb28be22011-08-01 21:38:03 -0700885 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886
Nick Piggin557ed1f2007-10-16 01:24:40 -0700887 page = ZERO_PAGE(0);
Andi Kleeneb28be22011-08-01 21:38:03 -0700888 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700889 sdio->next_block_for_io, map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 return;
891
Andi Kleeneb28be22011-08-01 21:38:03 -0700892 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893}
894
895/*
896 * Walk the user pages, and the file, mapping blocks to disk and generating
897 * a sequence of (page,offset,len,block) mappings. These mappings are injected
898 * into submit_page_section(), which takes care of the next stage of submission
899 *
900 * Direct IO against a blockdev is different from a file. Because we can
901 * happily perform page-sized but 512-byte aligned IOs. It is important that
902 * blockdev IO be able to have fine alignment and large sizes.
903 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800904 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905 * with the size of IO which is permitted at this offset and this i_blkbits.
906 *
907 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800908 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 * fine alignment but still allows this function to work in PAGE_SIZE units.
910 */
Andi Kleen18772642011-08-01 21:38:07 -0700911static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
912 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913{
Andi Kleeneb28be22011-08-01 21:38:03 -0700914 const unsigned blkbits = sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915 int ret = 0;
916
Andi Kleeneb28be22011-08-01 21:38:03 -0700917 while (sdio->block_in_file < sdio->final_block_in_request) {
Al Viro7b2c99d2014-03-15 04:05:57 -0400918 struct page *page;
919 size_t from, to;
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300920
921 page = dio_get_page(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 if (IS_ERR(page)) {
923 ret = PTR_ERR(page);
924 goto out;
925 }
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300926 from = sdio->head ? 0 : sdio->from;
927 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
928 sdio->head++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929
Al Viro7b2c99d2014-03-15 04:05:57 -0400930 while (from < to) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 unsigned this_chunk_bytes; /* # of bytes mapped */
932 unsigned this_chunk_blocks; /* # of blocks */
933 unsigned u;
934
Andi Kleeneb28be22011-08-01 21:38:03 -0700935 if (sdio->blocks_available == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936 /*
937 * Need to go and map some more disk
938 */
939 unsigned long blkmask;
940 unsigned long dio_remainder;
941
Andi Kleen18772642011-08-01 21:38:07 -0700942 ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 if (ret) {
944 page_cache_release(page);
945 goto out;
946 }
947 if (!buffer_mapped(map_bh))
948 goto do_holes;
949
Andi Kleeneb28be22011-08-01 21:38:03 -0700950 sdio->blocks_available =
951 map_bh->b_size >> sdio->blkbits;
952 sdio->next_block_for_io =
953 map_bh->b_blocknr << sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 if (buffer_new(map_bh))
Andi Kleen18772642011-08-01 21:38:07 -0700955 clean_blockdev_aliases(dio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956
Andi Kleeneb28be22011-08-01 21:38:03 -0700957 if (!sdio->blkfactor)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 goto do_holes;
959
Andi Kleeneb28be22011-08-01 21:38:03 -0700960 blkmask = (1 << sdio->blkfactor) - 1;
961 dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962
963 /*
964 * If we are at the start of IO and that IO
965 * starts partway into a fs-block,
966 * dio_remainder will be non-zero. If the IO
967 * is a read then we can simply advance the IO
968 * cursor to the first block which is to be
969 * read. But if the IO is a write and the
970 * block was newly allocated we cannot do that;
971 * the start of the fs block must be zeroed out
972 * on-disk
973 */
974 if (!buffer_new(map_bh))
Andi Kleeneb28be22011-08-01 21:38:03 -0700975 sdio->next_block_for_io += dio_remainder;
976 sdio->blocks_available -= dio_remainder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 }
978do_holes:
979 /* Handle holes */
980 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800981 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982
983 /* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboeb31dc662006-06-13 08:26:10 +0200984 if (dio->rw & WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985 page_cache_release(page);
986 return -ENOTBLK;
987 }
988
Jeff Moyer35dc8162006-02-03 03:04:27 -0800989 /*
990 * Be sure to account for a partial block as the
991 * last block in the file
992 */
993 i_size_aligned = ALIGN(i_size_read(dio->inode),
994 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -0700995 if (sdio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -0800996 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997 /* We hit eof */
998 page_cache_release(page);
999 goto out;
1000 }
Al Viro7b2c99d2014-03-15 04:05:57 -04001001 zero_user(page, from, 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -07001002 sdio->block_in_file++;
Al Viro7b2c99d2014-03-15 04:05:57 -04001003 from += 1 << blkbits;
Al Viro3320c602014-03-10 02:30:55 -04001004 dio->result += 1 << blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005 goto next_block;
1006 }
1007
1008 /*
1009 * If we're performing IO which has an alignment which
1010 * is finer than the underlying fs, go check to see if
1011 * we must zero out the start of this block.
1012 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001013 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
Andi Kleen18772642011-08-01 21:38:07 -07001014 dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015
1016 /*
1017 * Work out, in this_chunk_blocks, how much disk we
1018 * can add to this page
1019 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001020 this_chunk_blocks = sdio->blocks_available;
Al Viro7b2c99d2014-03-15 04:05:57 -04001021 u = (to - from) >> blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022 if (this_chunk_blocks > u)
1023 this_chunk_blocks = u;
Andi Kleeneb28be22011-08-01 21:38:03 -07001024 u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025 if (this_chunk_blocks > u)
1026 this_chunk_blocks = u;
1027 this_chunk_bytes = this_chunk_blocks << blkbits;
1028 BUG_ON(this_chunk_bytes == 0);
1029
Jan Kara092c8d42013-04-29 15:06:17 -07001030 if (this_chunk_blocks == sdio->blocks_available)
1031 sdio->boundary = buffer_boundary(map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -07001032 ret = submit_page_section(dio, sdio, page,
Al Viro7b2c99d2014-03-15 04:05:57 -04001033 from,
Andi Kleeneb28be22011-08-01 21:38:03 -07001034 this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -07001035 sdio->next_block_for_io,
1036 map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037 if (ret) {
1038 page_cache_release(page);
1039 goto out;
1040 }
Andi Kleeneb28be22011-08-01 21:38:03 -07001041 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042
Andi Kleeneb28be22011-08-01 21:38:03 -07001043 sdio->block_in_file += this_chunk_blocks;
Al Viro7b2c99d2014-03-15 04:05:57 -04001044 from += this_chunk_bytes;
1045 dio->result += this_chunk_bytes;
Andi Kleeneb28be22011-08-01 21:38:03 -07001046 sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047next_block:
Andi Kleeneb28be22011-08-01 21:38:03 -07001048 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1049 if (sdio->block_in_file == sdio->final_block_in_request)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050 break;
1051 }
1052
1053 /* Drop the ref which was taken in get_user_pages() */
1054 page_cache_release(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055 }
1056out:
1057 return ret;
1058}
1059
Andi Kleen847cc632011-08-01 21:38:09 -07001060static inline int drop_refcount(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061{
Andi Kleen847cc632011-08-01 21:38:09 -07001062 int ret2;
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001063 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064
Zach Brown8459d862006-12-10 02:21:05 -08001065 /*
1066 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001067 * operation. AIO can if it was a broken operation described above or
1068 * in fact if all the bios race to complete before we get here. In
1069 * that case dio_complete() translates the EIOCBQUEUED into the proper
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001070 * return code that the caller will hand to ->complete().
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001071 *
1072 * This is managed by the bio_lock instead of being an atomic_t so that
1073 * completion paths can drop their ref and use the remaining count to
1074 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001075 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001076 spin_lock_irqsave(&dio->bio_lock, flags);
1077 ret2 = --dio->refcount;
1078 spin_unlock_irqrestore(&dio->bio_lock, flags);
Andi Kleen847cc632011-08-01 21:38:09 -07001079 return ret2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080}
1081
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001082/*
1083 * This is a library function for use by filesystem drivers.
1084 *
1085 * The locking rules are governed by the flags parameter:
1086 * - if the flags value contains DIO_LOCKING we use a fancy locking
1087 * scheme for dumb filesystems.
1088 * For writes this function is called under i_mutex and returns with
1089 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1090 * taken and dropped again before returning.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001091 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1092 * internal locking but rather rely on the filesystem to synchronize
1093 * direct I/O reads/writes versus each other and truncate.
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001094 *
1095 * To help with locking against truncate we incremented the i_dio_count
1096 * counter before starting direct I/O, and decrement it once we are done.
1097 * Truncate can wait for it to reach zero to provide exclusion. It is
1098 * expected that filesystem provide exclusion between new direct I/O
1099 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1100 * but other filesystems need to take care of this on their own.
Andi Kleenba253fb2011-08-01 21:38:08 -07001101 *
1102 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1103 * is always inlined. Otherwise gcc is unable to split the structure into
1104 * individual fields and will generate much worse code. This is important
1105 * for the whole file.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001106 */
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001107static inline ssize_t
Omar Sandoval17f8c842015-03-16 04:33:50 -07001108do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1109 struct block_device *bdev, struct iov_iter *iter,
1110 loff_t offset, get_block_t get_block, dio_iodone_t end_io,
1111 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112{
Linus Torvaldsab738572012-11-29 12:27:00 -08001113 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1114 unsigned blkbits = i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 unsigned blocksize_mask = (1 << blkbits) - 1;
1116 ssize_t retval = -EINVAL;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001117 size_t count = iov_iter_count(iter);
1118 loff_t end = offset + count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119 struct dio *dio;
Andi Kleeneb28be22011-08-01 21:38:03 -07001120 struct dio_submit sdio = { 0, };
Andi Kleen847cc632011-08-01 21:38:09 -07001121 struct buffer_head map_bh = { 0, };
Fengguang Wu647d1e42012-08-09 15:23:09 +02001122 struct blk_plug plug;
Al Viro886a3912014-03-05 13:50:45 -05001123 unsigned long align = offset | iov_iter_alignment(iter);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001125 /*
1126 * Avoid references to bdev if not absolutely needed to give
1127 * the early prefetch in the caller enough time.
1128 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129
Al Viro886a3912014-03-05 13:50:45 -05001130 if (align & blocksize_mask) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001132 blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 blocksize_mask = (1 << blkbits) - 1;
Al Viro886a3912014-03-05 13:50:45 -05001134 if (align & blocksize_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 goto out;
1136 }
1137
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001138 /* watch out for a 0 len io from a tricksy fs */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001139 if (iov_iter_rw(iter) == READ && !iov_iter_count(iter))
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001140 return 0;
1141
Andi Kleen6e8267f2011-08-01 21:38:06 -07001142 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 retval = -ENOMEM;
1144 if (!dio)
1145 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001146 /*
1147 * Believe it or not, zeroing out the page array caused a .5%
1148 * performance regression in a database benchmark. So, we take
1149 * care to only zero out what's needed.
1150 */
1151 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001153 dio->flags = flags;
1154 if (dio->flags & DIO_LOCKING) {
Omar Sandoval17f8c842015-03-16 04:33:50 -07001155 if (iov_iter_rw(iter) == READ) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001156 struct address_space *mapping =
1157 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001159 /* will be released by direct_io_worker */
1160 mutex_lock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161
1162 retval = filemap_write_and_wait_range(mapping, offset,
1163 end - 1);
1164 if (retval) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001165 mutex_unlock(&inode->i_mutex);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001166 kmem_cache_free(dio_cache, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167 goto out;
1168 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170 }
1171
Jan Kara74cedf92015-11-30 10:15:42 -07001172 /* Once we sampled i_size check for reads beyond EOF */
1173 dio->i_size = i_size_read(inode);
1174 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1175 if (dio->flags & DIO_LOCKING)
1176 mutex_unlock(&inode->i_mutex);
1177 kmem_cache_free(dio_cache, dio);
Al Viro2d4594a2015-12-08 12:22:47 -05001178 retval = 0;
Jan Kara74cedf92015-11-30 10:15:42 -07001179 goto out;
1180 }
1181
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 /*
Christoph Hellwig60392572014-02-10 10:27:11 +11001183 * For file extending writes updating i_size before data writeouts
1184 * complete can expose uninitialized blocks in dumb filesystems.
1185 * In that case we need to wait for I/O completion even if asked
1186 * for an asynchronous write.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187 */
Christoph Hellwig60392572014-02-10 10:27:11 +11001188 if (is_sync_kiocb(iocb))
1189 dio->is_async = false;
1190 else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
Omar Sandoval17f8c842015-03-16 04:33:50 -07001191 iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
Christoph Hellwig60392572014-02-10 10:27:11 +11001192 dio->is_async = false;
1193 else
1194 dio->is_async = true;
1195
Andi Kleen847cc632011-08-01 21:38:09 -07001196 dio->inode = inode;
Omar Sandoval17f8c842015-03-16 04:33:50 -07001197 dio->rw = iov_iter_rw(iter) == WRITE ? WRITE_ODIRECT : READ;
Christoph Hellwig02afc272013-09-04 15:04:40 +02001198
1199 /*
1200 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1201 * so that we can call ->fsync.
1202 */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001203 if (dio->is_async && iov_iter_rw(iter) == WRITE &&
Christoph Hellwig02afc272013-09-04 15:04:40 +02001204 ((iocb->ki_filp->f_flags & O_DSYNC) ||
1205 IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1206 retval = dio_set_defer_completion(dio);
1207 if (retval) {
1208 /*
1209 * We grab i_mutex only for reads so we don't have
1210 * to release it here
1211 */
1212 kmem_cache_free(dio_cache, dio);
1213 goto out;
1214 }
1215 }
1216
1217 /*
1218 * Will be decremented at I/O completion time.
1219 */
Jens Axboefe0f07d2015-04-15 17:05:48 -06001220 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
1221 inode_dio_begin(inode);
Christoph Hellwig02afc272013-09-04 15:04:40 +02001222
1223 retval = 0;
Andi Kleen847cc632011-08-01 21:38:09 -07001224 sdio.blkbits = blkbits;
Linus Torvaldsab738572012-11-29 12:27:00 -08001225 sdio.blkfactor = i_blkbits - blkbits;
Andi Kleen847cc632011-08-01 21:38:09 -07001226 sdio.block_in_file = offset >> blkbits;
1227
1228 sdio.get_block = get_block;
1229 dio->end_io = end_io;
1230 sdio.submit_io = submit_io;
1231 sdio.final_block_in_bio = -1;
1232 sdio.next_block_for_io = -1;
1233
1234 dio->iocb = iocb;
Andi Kleen847cc632011-08-01 21:38:09 -07001235
1236 spin_lock_init(&dio->bio_lock);
1237 dio->refcount = 1;
1238
Ming Lei53cbf3b2015-08-17 10:31:46 +08001239 dio->should_dirty = (iter->type == ITER_IOVEC);
Al Viro7b2c99d2014-03-15 04:05:57 -04001240 sdio.iter = iter;
1241 sdio.final_block_in_request =
1242 (offset + iov_iter_count(iter)) >> blkbits;
1243
Andi Kleen847cc632011-08-01 21:38:09 -07001244 /*
1245 * In case of non-aligned buffers, we may need 2 more
1246 * pages since we need to zero out first and last block.
1247 */
1248 if (unlikely(sdio.blkfactor))
1249 sdio.pages_in_io = 2;
1250
Al Virof67da302014-03-19 01:16:16 -04001251 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
Andi Kleen847cc632011-08-01 21:38:09 -07001252
Fengguang Wu647d1e42012-08-09 15:23:09 +02001253 blk_start_plug(&plug);
1254
Al Viro7b2c99d2014-03-15 04:05:57 -04001255 retval = do_direct_IO(dio, &sdio, &map_bh);
1256 if (retval)
1257 dio_cleanup(dio, &sdio);
Andi Kleen847cc632011-08-01 21:38:09 -07001258
1259 if (retval == -ENOTBLK) {
1260 /*
1261 * The remaining part of the request will be
1262 * be handled by buffered I/O when we return
1263 */
1264 retval = 0;
1265 }
1266 /*
1267 * There may be some unwritten disk at the end of a part-written
1268 * fs-block-sized block. Go zero that now.
1269 */
1270 dio_zero_block(dio, &sdio, 1, &map_bh);
1271
1272 if (sdio.cur_page) {
1273 ssize_t ret2;
1274
1275 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1276 if (retval == 0)
1277 retval = ret2;
1278 page_cache_release(sdio.cur_page);
1279 sdio.cur_page = NULL;
1280 }
1281 if (sdio.bio)
1282 dio_bio_submit(dio, &sdio);
1283
Fengguang Wu647d1e42012-08-09 15:23:09 +02001284 blk_finish_plug(&plug);
1285
Andi Kleen847cc632011-08-01 21:38:09 -07001286 /*
1287 * It is possible that, we return short IO due to end of file.
1288 * In that case, we need to release all the pages we got hold on.
1289 */
1290 dio_cleanup(dio, &sdio);
1291
1292 /*
1293 * All block lookups have been performed. For READ requests
1294 * we can let i_mutex go now that its achieved its purpose
1295 * of protecting us from looking up uninitialized blocks.
1296 */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001297 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
Andi Kleen847cc632011-08-01 21:38:09 -07001298 mutex_unlock(&dio->inode->i_mutex);
1299
1300 /*
1301 * The only time we want to leave bios in flight is when a successful
1302 * partial aio read or full aio write have been setup. In that case
1303 * bio completion will call aio_complete. The only time it's safe to
1304 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1305 * This had *better* be the only place that raises -EIOCBQUEUED.
1306 */
1307 BUG_ON(retval == -EIOCBQUEUED);
1308 if (dio->is_async && retval == 0 && dio->result &&
Omar Sandoval17f8c842015-03-16 04:33:50 -07001309 (iov_iter_rw(iter) == READ || dio->result == count))
Andi Kleen847cc632011-08-01 21:38:09 -07001310 retval = -EIOCBQUEUED;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001311 else
Andi Kleen847cc632011-08-01 21:38:09 -07001312 dio_await_completion(dio);
1313
1314 if (drop_refcount(dio) == 0) {
1315 retval = dio_complete(dio, offset, retval, false);
Andi Kleen847cc632011-08-01 21:38:09 -07001316 } else
1317 BUG_ON(retval != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001319out:
1320 return retval;
1321}
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001322
Omar Sandoval17f8c842015-03-16 04:33:50 -07001323ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1324 struct block_device *bdev, struct iov_iter *iter,
1325 loff_t offset, get_block_t get_block,
1326 dio_iodone_t end_io, dio_submit_t submit_io,
1327 int flags)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001328{
1329 /*
1330 * The block device state is needed in the end to finally
1331 * submit everything. Since it's likely to be cache cold
1332 * prefetch it here as first thing to hide some of the
1333 * latency.
1334 *
1335 * Attempt to prefetch the pieces we likely need later.
1336 */
1337 prefetch(&bdev->bd_disk->part_tbl);
1338 prefetch(bdev->bd_queue);
1339 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1340
Omar Sandoval17f8c842015-03-16 04:33:50 -07001341 return do_blockdev_direct_IO(iocb, inode, bdev, iter, offset, get_block,
1342 end_io, submit_io, flags);
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001343}
1344
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345EXPORT_SYMBOL(__blockdev_direct_IO);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001346
1347static __init int dio_init(void)
1348{
1349 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1350 return 0;
1351}
1352module_init(dio_init)