blob: f3b4408be5904a996920397960b2df322fde07f3 [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;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300175 get_page(page);
Nick Pigginb5810032005-10-29 18:16:12 -0700176 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 Hellwig716b9bc2016-04-07 08:51:59 -0700227static ssize_t dio_complete(struct dio *dio, ssize_t ret, bool is_async)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228{
Christoph Hellwig716b9bc2016-04-07 08:51:59 -0700229 loff_t offset = dio->iocb->ki_pos;
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 Hellwig187372a2016-02-08 14:40:51 +1100256 if (dio->end_io) {
257 int err;
258
Christoph Hellwige2592212016-04-07 08:52:01 -0700259 // XXX: ki_pos??
Christoph Hellwig187372a2016-02-08 14:40:51 +1100260 err = dio->end_io(dio->iocb, offset, ret, dio->private);
261 if (err)
262 ret = err;
263 }
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000264
Jens Axboefe0f07d2015-04-15 17:05:48 -0600265 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
266 inode_dio_end(dio->inode);
267
Christoph Hellwig02afc272013-09-04 15:04:40 +0200268 if (is_async) {
Christoph Hellwige2592212016-04-07 08:52:01 -0700269 /*
270 * generic_write_sync expects ki_pos to have been updated
271 * already, but the submission path only does this for
272 * synchronous I/O.
273 */
274 dio->iocb->ki_pos += transferred;
Christoph Hellwig02afc272013-09-04 15:04:40 +0200275
Christoph Hellwige2592212016-04-07 08:52:01 -0700276 if (dio->rw & WRITE)
277 ret = generic_write_sync(dio->iocb, transferred);
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100278 dio->iocb->ki_complete(dio->iocb, ret, 0);
Christoph Hellwig02afc272013-09-04 15:04:40 +0200279 }
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200280
281 kmem_cache_free(dio_cache, dio);
Zach Brown6d544bb2006-12-10 02:20:54 -0800282 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283}
284
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200285static void dio_aio_complete_work(struct work_struct *work)
286{
287 struct dio *dio = container_of(work, struct dio, complete_work);
288
Christoph Hellwig716b9bc2016-04-07 08:51:59 -0700289 dio_complete(dio, 0, true);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200290}
291
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292static int dio_bio_complete(struct dio *dio, struct bio *bio);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200293
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294/*
295 * Asynchronous IO callback.
296 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200297static void dio_bio_end_aio(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298{
299 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800300 unsigned long remaining;
301 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303 /* cleanup the bio */
304 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800305
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800306 spin_lock_irqsave(&dio->bio_lock, flags);
307 remaining = --dio->refcount;
308 if (remaining == 1 && dio->waiter)
Zach Brown20258b22006-12-10 02:21:01 -0800309 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800310 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b22006-12-10 02:21:01 -0800311
Zach Brown8459d862006-12-10 02:21:05 -0800312 if (remaining == 0) {
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200313 if (dio->result && dio->defer_completion) {
314 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
315 queue_work(dio->inode->i_sb->s_dio_done_wq,
316 &dio->complete_work);
317 } else {
Christoph Hellwig716b9bc2016-04-07 08:51:59 -0700318 dio_complete(dio, 0, true);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200319 }
Zach Brown8459d862006-12-10 02:21:05 -0800320 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321}
322
323/*
324 * The BIO completion handler simply queues the BIO up for the process-context
325 * handler.
326 *
327 * During I/O bi_private points at the dio. After I/O, bi_private is used to
328 * implement a singly-linked list of completed BIOs, at dio->bio_list.
329 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200330static void dio_bio_end_io(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331{
332 struct dio *dio = bio->bi_private;
333 unsigned long flags;
334
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335 spin_lock_irqsave(&dio->bio_lock, flags);
336 bio->bi_private = dio->bio_list;
337 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800338 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339 wake_up_process(dio->waiter);
340 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341}
342
Josef Bacikfacd07b2010-05-23 11:00:55 -0400343/**
344 * dio_end_io - handle the end io action for the given bio
345 * @bio: The direct io bio thats being completed
346 * @error: Error if there was one
347 *
348 * This is meant to be called by any filesystem that uses their own dio_submit_t
349 * so that the DIO specific endio actions are dealt with after the filesystem
350 * has done it's completion work.
351 */
352void dio_end_io(struct bio *bio, int error)
353{
354 struct dio *dio = bio->bi_private;
355
356 if (dio->is_async)
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200357 dio_bio_end_aio(bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400358 else
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200359 dio_bio_end_io(bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400360}
361EXPORT_SYMBOL_GPL(dio_end_io);
362
Andi Kleenba253fb2011-08-01 21:38:08 -0700363static inline void
Andi Kleeneb28be22011-08-01 21:38:03 -0700364dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
365 struct block_device *bdev,
366 sector_t first_sector, int nr_vecs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367{
368 struct bio *bio;
369
David Dillow20d96002011-01-20 14:44:22 -0800370 /*
371 * bio_alloc() is guaranteed to return a bio when called with
Mel Gorman71baba42015-11-06 16:28:28 -0800372 * __GFP_RECLAIM and we request a valid number of vectors.
David Dillow20d96002011-01-20 14:44:22 -0800373 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375
376 bio->bi_bdev = bdev;
Kent Overstreet4f024f32013-10-11 15:44:27 -0700377 bio->bi_iter.bi_sector = first_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 if (dio->is_async)
379 bio->bi_end_io = dio_bio_end_aio;
380 else
381 bio->bi_end_io = dio_bio_end_io;
382
Andi Kleeneb28be22011-08-01 21:38:03 -0700383 sdio->bio = bio;
384 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385}
386
387/*
388 * In the AIO read case we speculatively dirty the pages before starting IO.
389 * During IO completion, any of these pages which happen to have been written
390 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800391 *
392 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700394static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395{
Andi Kleeneb28be22011-08-01 21:38:03 -0700396 struct bio *bio = sdio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800397 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398
399 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800400
401 spin_lock_irqsave(&dio->bio_lock, flags);
402 dio->refcount++;
403 spin_unlock_irqrestore(&dio->bio_lock, flags);
404
Ming Lei53cbf3b2015-08-17 10:31:46 +0800405 if (dio->is_async && dio->rw == READ && dio->should_dirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800407
Jens Axboec1c53462015-11-10 10:14:38 -0700408 dio->bio_bdev = bio->bi_bdev;
409
Jens Axboe15c4f632015-10-27 14:09:51 +0900410 if (sdio->submit_io) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700411 sdio->submit_io(dio->rw, bio, dio->inode,
412 sdio->logical_offset_in_bio);
Jens Axboe15c4f632015-10-27 14:09:51 +0900413 dio->bio_cookie = BLK_QC_T_NONE;
Jens Axboec1c53462015-11-10 10:14:38 -0700414 } else
Jens Axboe15c4f632015-10-27 14:09:51 +0900415 dio->bio_cookie = submit_bio(dio->rw, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416
Andi Kleeneb28be22011-08-01 21:38:03 -0700417 sdio->bio = NULL;
418 sdio->boundary = 0;
419 sdio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420}
421
422/*
423 * Release any resources in case of a failure
424 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700425static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426{
Al Viro7b2c99d2014-03-15 04:05:57 -0400427 while (sdio->head < sdio->tail)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300428 put_page(dio->pages[sdio->head++]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429}
430
431/*
Zach Brown02732012006-12-10 02:20:59 -0800432 * Wait for the next BIO to complete. Remove it and return it. NULL is
433 * returned once all BIOs have been completed. This must only be called once
434 * all bios have been issued so that dio->refcount can only decrease. This
435 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436 */
437static struct bio *dio_await_one(struct dio *dio)
438{
439 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800440 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441
442 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800443
444 /*
445 * Wait as long as the list is empty and there are bios in flight. bio
446 * completion drops the count, maybe adds to the list, and wakes while
447 * holding the bio_lock so we don't need set_current_state()'s barrier
448 * and can call it after testing our condition.
449 */
450 while (dio->refcount > 1 && dio->bio_list == NULL) {
451 __set_current_state(TASK_UNINTERRUPTIBLE);
452 dio->waiter = current;
453 spin_unlock_irqrestore(&dio->bio_lock, flags);
Christoph Hellwigc43c83a2016-03-03 16:04:02 +0100454 if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
455 !blk_poll(bdev_get_queue(dio->bio_bdev), dio->bio_cookie))
Jens Axboe15c4f632015-10-27 14:09:51 +0900456 io_schedule();
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800457 /* wake up sets us TASK_RUNNING */
458 spin_lock_irqsave(&dio->bio_lock, flags);
459 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 }
Zach Brown02732012006-12-10 02:20:59 -0800461 if (dio->bio_list) {
462 bio = dio->bio_list;
463 dio->bio_list = bio->bi_private;
464 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 spin_unlock_irqrestore(&dio->bio_lock, flags);
466 return bio;
467}
468
469/*
470 * Process one completed BIO. No locks are held.
471 */
472static int dio_bio_complete(struct dio *dio, struct bio *bio)
473{
Kent Overstreetcb34e052012-09-05 15:22:02 -0700474 struct bio_vec *bvec;
475 unsigned i;
Sasha Levin9b81c842015-08-10 19:05:18 -0400476 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200478 if (bio->bi_error)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800479 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480
Ming Lei53cbf3b2015-08-17 10:31:46 +0800481 if (dio->is_async && dio->rw == READ && dio->should_dirty) {
Sasha Levin9b81c842015-08-10 19:05:18 -0400482 err = bio->bi_error;
Mike Krinkin7ddc9712016-01-30 19:09:59 +0300483 bio_check_pages_dirty(bio); /* transfers ownership */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 } else {
Kent Overstreetcb34e052012-09-05 15:22:02 -0700485 bio_for_each_segment_all(bvec, bio, i) {
486 struct page *page = bvec->bv_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487
Ming Lei53cbf3b2015-08-17 10:31:46 +0800488 if (dio->rw == READ && !PageCompound(page) &&
489 dio->should_dirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 set_page_dirty_lock(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300491 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 }
Sasha Levin9b81c842015-08-10 19:05:18 -0400493 err = bio->bi_error;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 bio_put(bio);
495 }
Sasha Levin9b81c842015-08-10 19:05:18 -0400496 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497}
498
499/*
Zach Brown02732012006-12-10 02:20:59 -0800500 * Wait on and process all in-flight BIOs. This must only be called once
501 * all bios have been issued so that the refcount can only decrease.
502 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200503 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800504 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800506static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507{
Zach Brown02732012006-12-10 02:20:59 -0800508 struct bio *bio;
509 do {
510 bio = dio_await_one(dio);
511 if (bio)
512 dio_bio_complete(dio, bio);
513 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514}
515
516/*
517 * A really large O_DIRECT read or write can generate a lot of BIOs. So
518 * to keep the memory consumption sane we periodically reap any completed BIOs
519 * during the BIO generation phase.
520 *
521 * This also helps to limit the peak amount of pinned userspace memory.
522 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700523static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524{
525 int ret = 0;
526
Andi Kleeneb28be22011-08-01 21:38:03 -0700527 if (sdio->reap_counter++ >= 64) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 while (dio->bio_list) {
529 unsigned long flags;
530 struct bio *bio;
531 int ret2;
532
533 spin_lock_irqsave(&dio->bio_lock, flags);
534 bio = dio->bio_list;
535 dio->bio_list = bio->bi_private;
536 spin_unlock_irqrestore(&dio->bio_lock, flags);
537 ret2 = dio_bio_complete(dio, bio);
538 if (ret == 0)
539 ret = ret2;
540 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700541 sdio->reap_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542 }
543 return ret;
544}
545
546/*
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200547 * Create workqueue for deferred direct IO completions. We allocate the
548 * workqueue when it's first needed. This avoids creating workqueue for
549 * filesystems that don't need it and also allows us to create the workqueue
550 * late enough so the we can include s_id in the name of the workqueue.
551 */
552static int sb_init_dio_done_wq(struct super_block *sb)
553{
Olof Johansson45150c42013-09-09 10:34:23 -0700554 struct workqueue_struct *old;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200555 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
556 WQ_MEM_RECLAIM, 0,
557 sb->s_id);
558 if (!wq)
559 return -ENOMEM;
560 /*
561 * This has to be atomic as more DIOs can race to create the workqueue
562 */
Olof Johansson45150c42013-09-09 10:34:23 -0700563 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200564 /* Someone created workqueue before us? Free ours... */
Olof Johansson45150c42013-09-09 10:34:23 -0700565 if (old)
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200566 destroy_workqueue(wq);
567 return 0;
568}
569
570static int dio_set_defer_completion(struct dio *dio)
571{
572 struct super_block *sb = dio->inode->i_sb;
573
574 if (dio->defer_completion)
575 return 0;
576 dio->defer_completion = true;
577 if (!sb->s_dio_done_wq)
578 return sb_init_dio_done_wq(sb);
579 return 0;
580}
581
582/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 * Call into the fs to map some more disk blocks. We record the current number
Andi Kleeneb28be22011-08-01 21:38:03 -0700584 * of available blocks at sdio->blocks_available. These are in units of the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 * fs blocksize, (1 << inode->i_blkbits).
586 *
587 * The fs is allowed to map lots of blocks at once. If it wants to do that,
588 * it uses the passed inode-relative block number as the file offset, as usual.
589 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800590 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 * has remaining to do. The fs should not map more than this number of blocks.
592 *
593 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
594 * indicate how much contiguous disk space has been made available at
595 * bh->b_blocknr.
596 *
597 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
598 * This isn't very efficient...
599 *
600 * In the case of filesystem holes: the fs may return an arbitrarily-large
601 * hole by returning an appropriate value in b_size and by clearing
602 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800603 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604 */
Andi Kleen18772642011-08-01 21:38:07 -0700605static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
606 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607{
608 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
Tao Maae55e1a2012-01-12 17:20:33 -0800610 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611 unsigned long fs_count; /* Number of filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612 int create;
Linus Torvaldsab738572012-11-29 12:27:00 -0800613 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614
615 /*
616 * If there was a memory error and we've overwritten all the
617 * mapped blocks then we can now return that memory error
618 */
619 ret = dio->page_errors;
620 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700621 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
622 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
Tao Maae55e1a2012-01-12 17:20:33 -0800623 fs_endblk = (sdio->final_block_in_request - 1) >>
624 sdio->blkfactor;
625 fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626
Nathan Scott3c674e72006-03-29 09:26:15 +1000627 map_bh->b_state = 0;
Linus Torvaldsab738572012-11-29 12:27:00 -0800628 map_bh->b_size = fs_count << i_blkbits;
Nathan Scott3c674e72006-03-29 09:26:15 +1000629
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800630 /*
Eryu Guan9ecd10b2016-05-27 14:27:18 -0700631 * For writes that could fill holes inside i_size on a
632 * DIO_SKIP_HOLES filesystem we forbid block creations: only
633 * overwrites are permitted. We will return early to the caller
634 * once we see an unmapped buffer head returned, and the caller
635 * will fall back to buffered I/O.
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800636 *
637 * Otherwise the decision is left to the get_blocks method,
638 * which may decide to handle it or also return an unmapped
639 * buffer head.
640 */
Jens Axboeb31dc662006-06-13 08:26:10 +0200641 create = dio->rw & WRITE;
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800642 if (dio->flags & DIO_SKIP_HOLES) {
Eryu Guan9ecd10b2016-05-27 14:27:18 -0700643 if (fs_startblk <= ((i_size_read(dio->inode) - 1) >>
644 i_blkbits))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000647
Andi Kleeneb28be22011-08-01 21:38:03 -0700648 ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 map_bh, create);
Andi Kleen18772642011-08-01 21:38:07 -0700650
651 /* Store for completion */
652 dio->private = map_bh->b_private;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200653
654 if (ret == 0 && buffer_defer_completion(map_bh))
655 ret = dio_set_defer_completion(dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 }
657 return ret;
658}
659
660/*
661 * There is no bio. Make one now.
662 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700663static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
664 sector_t start_sector, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665{
666 sector_t sector;
667 int ret, nr_pages;
668
Andi Kleeneb28be22011-08-01 21:38:03 -0700669 ret = dio_bio_reap(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670 if (ret)
671 goto out;
Andi Kleeneb28be22011-08-01 21:38:03 -0700672 sector = start_sector << (sdio->blkbits - 9);
Kent Overstreetb54ffb72015-05-19 14:31:01 +0200673 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674 BUG_ON(nr_pages <= 0);
Andi Kleen18772642011-08-01 21:38:07 -0700675 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
Andi Kleeneb28be22011-08-01 21:38:03 -0700676 sdio->boundary = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677out:
678 return ret;
679}
680
681/*
682 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
683 * that was successful then update final_block_in_bio and take a ref against
684 * the just-added page.
685 *
686 * Return zero on success. Non-zero means the caller needs to start a new BIO.
687 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700688static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689{
690 int ret;
691
Andi Kleeneb28be22011-08-01 21:38:03 -0700692 ret = bio_add_page(sdio->bio, sdio->cur_page,
693 sdio->cur_page_len, sdio->cur_page_offset);
694 if (ret == sdio->cur_page_len) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 /*
696 * Decrement count only, if we are done with this page
697 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700698 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
699 sdio->pages_in_io--;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300700 get_page(sdio->cur_page);
Andi Kleeneb28be22011-08-01 21:38:03 -0700701 sdio->final_block_in_bio = sdio->cur_page_block +
702 (sdio->cur_page_len >> sdio->blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 ret = 0;
704 } else {
705 ret = 1;
706 }
707 return ret;
708}
709
710/*
711 * Put cur_page under IO. The section of cur_page which is described by
712 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
713 * starts on-disk at cur_page_block.
714 *
715 * We take a ref against the page here (on behalf of its presence in the bio).
716 *
717 * The caller of this function is responsible for removing cur_page from the
718 * dio, and for dropping the refcount which came from that presence.
719 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700720static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
721 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722{
723 int ret = 0;
724
Andi Kleeneb28be22011-08-01 21:38:03 -0700725 if (sdio->bio) {
726 loff_t cur_offset = sdio->cur_page_fs_offset;
727 loff_t bio_next_offset = sdio->logical_offset_in_bio +
Kent Overstreet4f024f32013-10-11 15:44:27 -0700728 sdio->bio->bi_iter.bi_size;
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400729
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400731 * See whether this new request is contiguous with the old.
732 *
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900733 * Btrfs cannot handle having logically non-contiguous requests
734 * submitted. For example if you have
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400735 *
736 * Logical: [0-4095][HOLE][8192-12287]
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900737 * Physical: [0-4095] [4096-8191]
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400738 *
739 * We cannot submit those pages together as one BIO. So if our
740 * current logical offset in the file does not equal what would
741 * be the next logical offset in the bio, submit the bio we
742 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700744 if (sdio->final_block_in_bio != sdio->cur_page_block ||
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400745 cur_offset != bio_next_offset)
Andi Kleeneb28be22011-08-01 21:38:03 -0700746 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 }
748
Andi Kleeneb28be22011-08-01 21:38:03 -0700749 if (sdio->bio == NULL) {
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)
752 goto out;
753 }
754
Andi Kleeneb28be22011-08-01 21:38:03 -0700755 if (dio_bio_add_page(sdio) != 0) {
756 dio_bio_submit(dio, sdio);
Andi Kleen18772642011-08-01 21:38:07 -0700757 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700759 ret = dio_bio_add_page(sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760 BUG_ON(ret != 0);
761 }
762 }
763out:
764 return ret;
765}
766
767/*
768 * An autonomous function to put a chunk of a page under deferred IO.
769 *
770 * The caller doesn't actually know (or care) whether this piece of page is in
771 * a BIO, or is under IO or whatever. We just take care of all possible
772 * situations here. The separation between the logic of do_direct_IO() and
773 * that of submit_page_section() is important for clarity. Please don't break.
774 *
775 * The chunk of page starts on-disk at blocknr.
776 *
777 * We perform deferred IO, by recording the last-submitted page inside our
778 * private part of the dio structure. If possible, we just expand the IO
779 * across that page here.
780 *
781 * If that doesn't work out then we put the old page into the bio and add this
782 * page to the dio instead.
783 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700784static inline int
Andi Kleeneb28be22011-08-01 21:38:03 -0700785submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
Andi Kleen18772642011-08-01 21:38:07 -0700786 unsigned offset, unsigned len, sector_t blocknr,
787 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788{
789 int ret = 0;
790
Andrew Morton98c4d572006-12-10 02:19:47 -0800791 if (dio->rw & WRITE) {
792 /*
793 * Read accounting is performed in submit_bio()
794 */
795 task_io_account_write(len);
796 }
797
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 /*
799 * Can we just grow the current page's presence in the dio?
800 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700801 if (sdio->cur_page == page &&
802 sdio->cur_page_offset + sdio->cur_page_len == offset &&
803 sdio->cur_page_block +
804 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
805 sdio->cur_page_len += len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806 goto out;
807 }
808
809 /*
810 * If there's a deferred page already there then send it.
811 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700812 if (sdio->cur_page) {
Andi Kleen18772642011-08-01 21:38:07 -0700813 ret = dio_send_cur_page(dio, sdio, map_bh);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300814 put_page(sdio->cur_page);
Andi Kleeneb28be22011-08-01 21:38:03 -0700815 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816 if (ret)
Jan Karab1058b92013-04-29 15:06:18 -0700817 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818 }
819
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300820 get_page(page); /* It is in dio */
Andi Kleeneb28be22011-08-01 21:38:03 -0700821 sdio->cur_page = page;
822 sdio->cur_page_offset = offset;
823 sdio->cur_page_len = len;
824 sdio->cur_page_block = blocknr;
825 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826out:
Jan Karab1058b92013-04-29 15:06:18 -0700827 /*
828 * If sdio->boundary then we want to schedule the IO now to
829 * avoid metadata seeks.
830 */
831 if (sdio->boundary) {
832 ret = dio_send_cur_page(dio, sdio, map_bh);
833 dio_bio_submit(dio, sdio);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300834 put_page(sdio->cur_page);
Jan Karab1058b92013-04-29 15:06:18 -0700835 sdio->cur_page = NULL;
836 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 return ret;
838}
839
840/*
841 * Clean any dirty buffers in the blockdev mapping which alias newly-created
842 * file blocks. Only called for S_ISREG files - blockdevs do not set
843 * buffer_new
844 */
Andi Kleen18772642011-08-01 21:38:07 -0700845static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846{
847 unsigned i;
848 unsigned nblocks;
849
Andi Kleen18772642011-08-01 21:38:07 -0700850 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851
852 for (i = 0; i < nblocks; i++) {
Andi Kleen18772642011-08-01 21:38:07 -0700853 unmap_underlying_metadata(map_bh->b_bdev,
854 map_bh->b_blocknr + i);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 }
856}
857
858/*
859 * If we are not writing the entire block and get_block() allocated
860 * the block for us, we need to fill-in the unused portion of the
861 * block with zeros. This happens only if user-buffer, fileoffset or
862 * io length is not filesystem block-size multiple.
863 *
864 * `end' is zero if we're doing the start of the IO, 1 at the end of the
865 * IO.
866 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700867static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
868 int end, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869{
870 unsigned dio_blocks_per_fs_block;
871 unsigned this_chunk_blocks; /* In dio_blocks */
872 unsigned this_chunk_bytes;
873 struct page *page;
874
Andi Kleeneb28be22011-08-01 21:38:03 -0700875 sdio->start_zero_done = 1;
Andi Kleen18772642011-08-01 21:38:07 -0700876 if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877 return;
878
Andi Kleeneb28be22011-08-01 21:38:03 -0700879 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
880 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881
882 if (!this_chunk_blocks)
883 return;
884
885 /*
886 * We need to zero out part of an fs block. It is either at the
887 * beginning or the end of the fs block.
888 */
889 if (end)
890 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
891
Andi Kleeneb28be22011-08-01 21:38:03 -0700892 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893
Nick Piggin557ed1f2007-10-16 01:24:40 -0700894 page = ZERO_PAGE(0);
Andi Kleeneb28be22011-08-01 21:38:03 -0700895 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700896 sdio->next_block_for_io, map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 return;
898
Andi Kleeneb28be22011-08-01 21:38:03 -0700899 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900}
901
902/*
903 * Walk the user pages, and the file, mapping blocks to disk and generating
904 * a sequence of (page,offset,len,block) mappings. These mappings are injected
905 * into submit_page_section(), which takes care of the next stage of submission
906 *
907 * Direct IO against a blockdev is different from a file. Because we can
908 * happily perform page-sized but 512-byte aligned IOs. It is important that
909 * blockdev IO be able to have fine alignment and large sizes.
910 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800911 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912 * with the size of IO which is permitted at this offset and this i_blkbits.
913 *
914 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800915 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 * fine alignment but still allows this function to work in PAGE_SIZE units.
917 */
Andi Kleen18772642011-08-01 21:38:07 -0700918static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
919 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920{
Andi Kleeneb28be22011-08-01 21:38:03 -0700921 const unsigned blkbits = sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 int ret = 0;
923
Andi Kleeneb28be22011-08-01 21:38:03 -0700924 while (sdio->block_in_file < sdio->final_block_in_request) {
Al Viro7b2c99d2014-03-15 04:05:57 -0400925 struct page *page;
926 size_t from, to;
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300927
928 page = dio_get_page(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 if (IS_ERR(page)) {
930 ret = PTR_ERR(page);
931 goto out;
932 }
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300933 from = sdio->head ? 0 : sdio->from;
934 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
935 sdio->head++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936
Al Viro7b2c99d2014-03-15 04:05:57 -0400937 while (from < to) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 unsigned this_chunk_bytes; /* # of bytes mapped */
939 unsigned this_chunk_blocks; /* # of blocks */
940 unsigned u;
941
Andi Kleeneb28be22011-08-01 21:38:03 -0700942 if (sdio->blocks_available == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 /*
944 * Need to go and map some more disk
945 */
946 unsigned long blkmask;
947 unsigned long dio_remainder;
948
Andi Kleen18772642011-08-01 21:38:07 -0700949 ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950 if (ret) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300951 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952 goto out;
953 }
954 if (!buffer_mapped(map_bh))
955 goto do_holes;
956
Andi Kleeneb28be22011-08-01 21:38:03 -0700957 sdio->blocks_available =
958 map_bh->b_size >> sdio->blkbits;
959 sdio->next_block_for_io =
960 map_bh->b_blocknr << sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 if (buffer_new(map_bh))
Andi Kleen18772642011-08-01 21:38:07 -0700962 clean_blockdev_aliases(dio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963
Andi Kleeneb28be22011-08-01 21:38:03 -0700964 if (!sdio->blkfactor)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 goto do_holes;
966
Andi Kleeneb28be22011-08-01 21:38:03 -0700967 blkmask = (1 << sdio->blkfactor) - 1;
968 dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969
970 /*
971 * If we are at the start of IO and that IO
972 * starts partway into a fs-block,
973 * dio_remainder will be non-zero. If the IO
974 * is a read then we can simply advance the IO
975 * cursor to the first block which is to be
976 * read. But if the IO is a write and the
977 * block was newly allocated we cannot do that;
978 * the start of the fs block must be zeroed out
979 * on-disk
980 */
981 if (!buffer_new(map_bh))
Andi Kleeneb28be22011-08-01 21:38:03 -0700982 sdio->next_block_for_io += dio_remainder;
983 sdio->blocks_available -= dio_remainder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984 }
985do_holes:
986 /* Handle holes */
987 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800988 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989
990 /* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboeb31dc662006-06-13 08:26:10 +0200991 if (dio->rw & WRITE) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300992 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 return -ENOTBLK;
994 }
995
Jeff Moyer35dc8162006-02-03 03:04:27 -0800996 /*
997 * Be sure to account for a partial block as the
998 * last block in the file
999 */
1000 i_size_aligned = ALIGN(i_size_read(dio->inode),
1001 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -07001002 if (sdio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -08001003 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004 /* We hit eof */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001005 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006 goto out;
1007 }
Al Viro7b2c99d2014-03-15 04:05:57 -04001008 zero_user(page, from, 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -07001009 sdio->block_in_file++;
Al Viro7b2c99d2014-03-15 04:05:57 -04001010 from += 1 << blkbits;
Al Viro3320c602014-03-10 02:30:55 -04001011 dio->result += 1 << blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012 goto next_block;
1013 }
1014
1015 /*
1016 * If we're performing IO which has an alignment which
1017 * is finer than the underlying fs, go check to see if
1018 * we must zero out the start of this block.
1019 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001020 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
Andi Kleen18772642011-08-01 21:38:07 -07001021 dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022
1023 /*
1024 * Work out, in this_chunk_blocks, how much disk we
1025 * can add to this page
1026 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001027 this_chunk_blocks = sdio->blocks_available;
Al Viro7b2c99d2014-03-15 04:05:57 -04001028 u = (to - from) >> blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029 if (this_chunk_blocks > u)
1030 this_chunk_blocks = u;
Andi Kleeneb28be22011-08-01 21:38:03 -07001031 u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032 if (this_chunk_blocks > u)
1033 this_chunk_blocks = u;
1034 this_chunk_bytes = this_chunk_blocks << blkbits;
1035 BUG_ON(this_chunk_bytes == 0);
1036
Jan Kara092c8d42013-04-29 15:06:17 -07001037 if (this_chunk_blocks == sdio->blocks_available)
1038 sdio->boundary = buffer_boundary(map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -07001039 ret = submit_page_section(dio, sdio, page,
Al Viro7b2c99d2014-03-15 04:05:57 -04001040 from,
Andi Kleeneb28be22011-08-01 21:38:03 -07001041 this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -07001042 sdio->next_block_for_io,
1043 map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044 if (ret) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001045 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046 goto out;
1047 }
Andi Kleeneb28be22011-08-01 21:38:03 -07001048 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001049
Andi Kleeneb28be22011-08-01 21:38:03 -07001050 sdio->block_in_file += this_chunk_blocks;
Al Viro7b2c99d2014-03-15 04:05:57 -04001051 from += this_chunk_bytes;
1052 dio->result += this_chunk_bytes;
Andi Kleeneb28be22011-08-01 21:38:03 -07001053 sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054next_block:
Andi Kleeneb28be22011-08-01 21:38:03 -07001055 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1056 if (sdio->block_in_file == sdio->final_block_in_request)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057 break;
1058 }
1059
1060 /* Drop the ref which was taken in get_user_pages() */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001061 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 }
1063out:
1064 return ret;
1065}
1066
Andi Kleen847cc632011-08-01 21:38:09 -07001067static inline int drop_refcount(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068{
Andi Kleen847cc632011-08-01 21:38:09 -07001069 int ret2;
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001070 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071
Zach Brown8459d862006-12-10 02:21:05 -08001072 /*
1073 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001074 * operation. AIO can if it was a broken operation described above or
1075 * in fact if all the bios race to complete before we get here. In
1076 * that case dio_complete() translates the EIOCBQUEUED into the proper
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001077 * return code that the caller will hand to ->complete().
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001078 *
1079 * This is managed by the bio_lock instead of being an atomic_t so that
1080 * completion paths can drop their ref and use the remaining count to
1081 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001082 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001083 spin_lock_irqsave(&dio->bio_lock, flags);
1084 ret2 = --dio->refcount;
1085 spin_unlock_irqrestore(&dio->bio_lock, flags);
Andi Kleen847cc632011-08-01 21:38:09 -07001086 return ret2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087}
1088
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001089/*
1090 * This is a library function for use by filesystem drivers.
1091 *
1092 * The locking rules are governed by the flags parameter:
1093 * - if the flags value contains DIO_LOCKING we use a fancy locking
1094 * scheme for dumb filesystems.
1095 * For writes this function is called under i_mutex and returns with
1096 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1097 * taken and dropped again before returning.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001098 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1099 * internal locking but rather rely on the filesystem to synchronize
1100 * direct I/O reads/writes versus each other and truncate.
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001101 *
1102 * To help with locking against truncate we incremented the i_dio_count
1103 * counter before starting direct I/O, and decrement it once we are done.
1104 * Truncate can wait for it to reach zero to provide exclusion. It is
1105 * expected that filesystem provide exclusion between new direct I/O
1106 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1107 * but other filesystems need to take care of this on their own.
Andi Kleenba253fb2011-08-01 21:38:08 -07001108 *
1109 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1110 * is always inlined. Otherwise gcc is unable to split the structure into
1111 * individual fields and will generate much worse code. This is important
1112 * for the whole file.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001113 */
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001114static inline ssize_t
Omar Sandoval17f8c842015-03-16 04:33:50 -07001115do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1116 struct block_device *bdev, struct iov_iter *iter,
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001117 get_block_t get_block, dio_iodone_t end_io,
Omar Sandoval17f8c842015-03-16 04:33:50 -07001118 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119{
Linus Torvaldsab738572012-11-29 12:27:00 -08001120 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1121 unsigned blkbits = i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 unsigned blocksize_mask = (1 << blkbits) - 1;
1123 ssize_t retval = -EINVAL;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001124 size_t count = iov_iter_count(iter);
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001125 loff_t offset = iocb->ki_pos;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001126 loff_t end = offset + count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 struct dio *dio;
Andi Kleeneb28be22011-08-01 21:38:03 -07001128 struct dio_submit sdio = { 0, };
Andi Kleen847cc632011-08-01 21:38:09 -07001129 struct buffer_head map_bh = { 0, };
Fengguang Wu647d1e42012-08-09 15:23:09 +02001130 struct blk_plug plug;
Al Viro886a3912014-03-05 13:50:45 -05001131 unsigned long align = offset | iov_iter_alignment(iter);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001133 /*
1134 * Avoid references to bdev if not absolutely needed to give
1135 * the early prefetch in the caller enough time.
1136 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137
Al Viro886a3912014-03-05 13:50:45 -05001138 if (align & blocksize_mask) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001140 blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 blocksize_mask = (1 << blkbits) - 1;
Al Viro886a3912014-03-05 13:50:45 -05001142 if (align & blocksize_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 goto out;
1144 }
1145
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001146 /* watch out for a 0 len io from a tricksy fs */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001147 if (iov_iter_rw(iter) == READ && !iov_iter_count(iter))
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001148 return 0;
1149
Andi Kleen6e8267f2011-08-01 21:38:06 -07001150 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 retval = -ENOMEM;
1152 if (!dio)
1153 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001154 /*
1155 * Believe it or not, zeroing out the page array caused a .5%
1156 * performance regression in a database benchmark. So, we take
1157 * care to only zero out what's needed.
1158 */
1159 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001161 dio->flags = flags;
1162 if (dio->flags & DIO_LOCKING) {
Omar Sandoval17f8c842015-03-16 04:33:50 -07001163 if (iov_iter_rw(iter) == READ) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001164 struct address_space *mapping =
1165 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001167 /* will be released by direct_io_worker */
Al Viro59551022016-01-22 15:40:57 -05001168 inode_lock(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169
1170 retval = filemap_write_and_wait_range(mapping, offset,
1171 end - 1);
1172 if (retval) {
Al Viro59551022016-01-22 15:40:57 -05001173 inode_unlock(inode);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001174 kmem_cache_free(dio_cache, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 goto out;
1176 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178 }
1179
Jan Kara74cedf92015-11-30 10:15:42 -07001180 /* Once we sampled i_size check for reads beyond EOF */
1181 dio->i_size = i_size_read(inode);
1182 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1183 if (dio->flags & DIO_LOCKING)
Al Viro59551022016-01-22 15:40:57 -05001184 inode_unlock(inode);
Jan Kara74cedf92015-11-30 10:15:42 -07001185 kmem_cache_free(dio_cache, dio);
Al Viro2d4594a2015-12-08 12:22:47 -05001186 retval = 0;
Jan Kara74cedf92015-11-30 10:15:42 -07001187 goto out;
1188 }
1189
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190 /*
Christoph Hellwig60392572014-02-10 10:27:11 +11001191 * For file extending writes updating i_size before data writeouts
1192 * complete can expose uninitialized blocks in dumb filesystems.
1193 * In that case we need to wait for I/O completion even if asked
1194 * for an asynchronous write.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195 */
Christoph Hellwig60392572014-02-10 10:27:11 +11001196 if (is_sync_kiocb(iocb))
1197 dio->is_async = false;
1198 else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
Omar Sandoval17f8c842015-03-16 04:33:50 -07001199 iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
Christoph Hellwig60392572014-02-10 10:27:11 +11001200 dio->is_async = false;
1201 else
1202 dio->is_async = true;
1203
Andi Kleen847cc632011-08-01 21:38:09 -07001204 dio->inode = inode;
Omar Sandoval17f8c842015-03-16 04:33:50 -07001205 dio->rw = iov_iter_rw(iter) == WRITE ? WRITE_ODIRECT : READ;
Christoph Hellwig02afc272013-09-04 15:04:40 +02001206
1207 /*
1208 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1209 * so that we can call ->fsync.
1210 */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001211 if (dio->is_async && iov_iter_rw(iter) == WRITE &&
Christoph Hellwig02afc272013-09-04 15:04:40 +02001212 ((iocb->ki_filp->f_flags & O_DSYNC) ||
1213 IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1214 retval = dio_set_defer_completion(dio);
1215 if (retval) {
1216 /*
1217 * We grab i_mutex only for reads so we don't have
1218 * to release it here
1219 */
1220 kmem_cache_free(dio_cache, dio);
1221 goto out;
1222 }
1223 }
1224
1225 /*
1226 * Will be decremented at I/O completion time.
1227 */
Jens Axboefe0f07d2015-04-15 17:05:48 -06001228 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
1229 inode_dio_begin(inode);
Christoph Hellwig02afc272013-09-04 15:04:40 +02001230
1231 retval = 0;
Andi Kleen847cc632011-08-01 21:38:09 -07001232 sdio.blkbits = blkbits;
Linus Torvaldsab738572012-11-29 12:27:00 -08001233 sdio.blkfactor = i_blkbits - blkbits;
Andi Kleen847cc632011-08-01 21:38:09 -07001234 sdio.block_in_file = offset >> blkbits;
1235
1236 sdio.get_block = get_block;
1237 dio->end_io = end_io;
1238 sdio.submit_io = submit_io;
1239 sdio.final_block_in_bio = -1;
1240 sdio.next_block_for_io = -1;
1241
1242 dio->iocb = iocb;
Andi Kleen847cc632011-08-01 21:38:09 -07001243
1244 spin_lock_init(&dio->bio_lock);
1245 dio->refcount = 1;
1246
Ming Lei53cbf3b2015-08-17 10:31:46 +08001247 dio->should_dirty = (iter->type == ITER_IOVEC);
Al Viro7b2c99d2014-03-15 04:05:57 -04001248 sdio.iter = iter;
1249 sdio.final_block_in_request =
1250 (offset + iov_iter_count(iter)) >> blkbits;
1251
Andi Kleen847cc632011-08-01 21:38:09 -07001252 /*
1253 * In case of non-aligned buffers, we may need 2 more
1254 * pages since we need to zero out first and last block.
1255 */
1256 if (unlikely(sdio.blkfactor))
1257 sdio.pages_in_io = 2;
1258
Al Virof67da302014-03-19 01:16:16 -04001259 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
Andi Kleen847cc632011-08-01 21:38:09 -07001260
Fengguang Wu647d1e42012-08-09 15:23:09 +02001261 blk_start_plug(&plug);
1262
Al Viro7b2c99d2014-03-15 04:05:57 -04001263 retval = do_direct_IO(dio, &sdio, &map_bh);
1264 if (retval)
1265 dio_cleanup(dio, &sdio);
Andi Kleen847cc632011-08-01 21:38:09 -07001266
1267 if (retval == -ENOTBLK) {
1268 /*
1269 * The remaining part of the request will be
1270 * be handled by buffered I/O when we return
1271 */
1272 retval = 0;
1273 }
1274 /*
1275 * There may be some unwritten disk at the end of a part-written
1276 * fs-block-sized block. Go zero that now.
1277 */
1278 dio_zero_block(dio, &sdio, 1, &map_bh);
1279
1280 if (sdio.cur_page) {
1281 ssize_t ret2;
1282
1283 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1284 if (retval == 0)
1285 retval = ret2;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001286 put_page(sdio.cur_page);
Andi Kleen847cc632011-08-01 21:38:09 -07001287 sdio.cur_page = NULL;
1288 }
1289 if (sdio.bio)
1290 dio_bio_submit(dio, &sdio);
1291
Fengguang Wu647d1e42012-08-09 15:23:09 +02001292 blk_finish_plug(&plug);
1293
Andi Kleen847cc632011-08-01 21:38:09 -07001294 /*
1295 * It is possible that, we return short IO due to end of file.
1296 * In that case, we need to release all the pages we got hold on.
1297 */
1298 dio_cleanup(dio, &sdio);
1299
1300 /*
1301 * All block lookups have been performed. For READ requests
1302 * we can let i_mutex go now that its achieved its purpose
1303 * of protecting us from looking up uninitialized blocks.
1304 */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001305 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
Al Viro59551022016-01-22 15:40:57 -05001306 inode_unlock(dio->inode);
Andi Kleen847cc632011-08-01 21:38:09 -07001307
1308 /*
1309 * The only time we want to leave bios in flight is when a successful
1310 * partial aio read or full aio write have been setup. In that case
1311 * bio completion will call aio_complete. The only time it's safe to
1312 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1313 * This had *better* be the only place that raises -EIOCBQUEUED.
1314 */
1315 BUG_ON(retval == -EIOCBQUEUED);
1316 if (dio->is_async && retval == 0 && dio->result &&
Omar Sandoval17f8c842015-03-16 04:33:50 -07001317 (iov_iter_rw(iter) == READ || dio->result == count))
Andi Kleen847cc632011-08-01 21:38:09 -07001318 retval = -EIOCBQUEUED;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001319 else
Andi Kleen847cc632011-08-01 21:38:09 -07001320 dio_await_completion(dio);
1321
1322 if (drop_refcount(dio) == 0) {
Christoph Hellwig716b9bc2016-04-07 08:51:59 -07001323 retval = dio_complete(dio, retval, false);
Andi Kleen847cc632011-08-01 21:38:09 -07001324 } else
1325 BUG_ON(retval != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001327out:
1328 return retval;
1329}
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001330
Omar Sandoval17f8c842015-03-16 04:33:50 -07001331ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1332 struct block_device *bdev, struct iov_iter *iter,
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001333 get_block_t get_block,
Omar Sandoval17f8c842015-03-16 04:33:50 -07001334 dio_iodone_t end_io, dio_submit_t submit_io,
1335 int flags)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001336{
1337 /*
1338 * The block device state is needed in the end to finally
1339 * submit everything. Since it's likely to be cache cold
1340 * prefetch it here as first thing to hide some of the
1341 * latency.
1342 *
1343 * Attempt to prefetch the pieces we likely need later.
1344 */
1345 prefetch(&bdev->bd_disk->part_tbl);
1346 prefetch(bdev->bd_queue);
1347 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1348
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001349 return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
Omar Sandoval17f8c842015-03-16 04:33:50 -07001350 end_io, submit_io, flags);
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001351}
1352
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353EXPORT_SYMBOL(__blockdev_direct_IO);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001354
1355static __init int dio_init(void)
1356{
1357 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1358 return 0;
1359}
1360module_init(dio_init)