blob: c6220a2daefd96642180d642cd3a6bfde23dd262 [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 */
Mike Christie8a4c1e42016-06-05 14:31:50 -0500111 int op;
112 int op_flags;
Jens Axboe15c4f632015-10-27 14:09:51 +0900113 blk_qc_t bio_cookie;
114 struct block_device *bio_bdev;
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700115 struct inode *inode;
Andi Kleeneb28be22011-08-01 21:38:03 -0700116 loff_t i_size; /* i_size when submitted */
117 dio_iodone_t *end_io; /* IO completion function */
Andi Kleeneb28be22011-08-01 21:38:03 -0700118
Andi Kleen18772642011-08-01 21:38:07 -0700119 void *private; /* copy from map_bh.b_private */
Andi Kleeneb28be22011-08-01 21:38:03 -0700120
121 /* BIO completion state */
122 spinlock_t bio_lock; /* protects BIO fields below */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700123 int page_errors; /* errno from get_user_pages() */
124 int is_async; /* is IO async ? */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200125 bool defer_completion; /* defer AIO completion to workqueue? */
Ming Lei53cbf3b2015-08-17 10:31:46 +0800126 bool should_dirty; /* if pages should be dirtied */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700127 int io_error; /* IO error in completion path */
Andi Kleeneb28be22011-08-01 21:38:03 -0700128 unsigned long refcount; /* direct_io_worker() and bios */
129 struct bio *bio_list; /* singly linked via bi_private */
130 struct task_struct *waiter; /* waiting task (NULL if none) */
131
132 /* AIO related stuff */
133 struct kiocb *iocb; /* kiocb */
Andi Kleeneb28be22011-08-01 21:38:03 -0700134 ssize_t result; /* IO result */
135
Jeff Moyer23aee092009-12-15 16:47:49 -0800136 /*
137 * pages[] (and any fields placed after it) are not zeroed out at
138 * allocation time. Don't add new fields after pages[] unless you
139 * wish that they not be zeroed.
140 */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200141 union {
142 struct page *pages[DIO_PAGES]; /* page buffer */
143 struct work_struct complete_work;/* deferred AIO completion */
144 };
Andi Kleen6e8267f2011-08-01 21:38:06 -0700145} ____cacheline_aligned_in_smp;
146
147static struct kmem_cache *dio_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148
149/*
150 * How many pages are in the queue?
151 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700152static inline unsigned dio_pages_present(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153{
Andi Kleeneb28be22011-08-01 21:38:03 -0700154 return sdio->tail - sdio->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155}
156
157/*
158 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
159 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700160static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161{
Al Viro7b2c99d2014-03-15 04:05:57 -0400162 ssize_t ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163
Miklos Szeredi2c809292014-09-24 17:09:11 +0200164 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
Al Viro7b2c99d2014-03-15 04:05:57 -0400165 &sdio->from);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166
Mike Christie8a4c1e42016-06-05 14:31:50 -0500167 if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
Nick Piggin557ed1f2007-10-16 01:24:40 -0700168 struct page *page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169 /*
170 * A memory fault, but the filesystem has some outstanding
171 * mapped blocks. We need to use those blocks up to avoid
172 * leaking stale data in the file.
173 */
174 if (dio->page_errors == 0)
175 dio->page_errors = ret;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300176 get_page(page);
Nick Pigginb5810032005-10-29 18:16:12 -0700177 dio->pages[0] = page;
Andi Kleeneb28be22011-08-01 21:38:03 -0700178 sdio->head = 0;
179 sdio->tail = 1;
Al Viro7b2c99d2014-03-15 04:05:57 -0400180 sdio->from = 0;
181 sdio->to = PAGE_SIZE;
182 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183 }
184
185 if (ret >= 0) {
Al Viro7b2c99d2014-03-15 04:05:57 -0400186 iov_iter_advance(sdio->iter, ret);
187 ret += sdio->from;
Andi Kleeneb28be22011-08-01 21:38:03 -0700188 sdio->head = 0;
Al Viro7b2c99d2014-03-15 04:05:57 -0400189 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
190 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
191 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193 return ret;
194}
195
196/*
197 * Get another userspace page. Returns an ERR_PTR on error. Pages are
198 * buffered inside the dio so that we can call get_user_pages() against a
199 * decent number of pages, less frequently. To provide nicer use of the
200 * L1 cache.
201 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700202static inline struct page *dio_get_page(struct dio *dio,
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300203 struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204{
Andi Kleeneb28be22011-08-01 21:38:03 -0700205 if (dio_pages_present(sdio) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206 int ret;
207
Andi Kleeneb28be22011-08-01 21:38:03 -0700208 ret = dio_refill_pages(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209 if (ret)
210 return ERR_PTR(ret);
Andi Kleeneb28be22011-08-01 21:38:03 -0700211 BUG_ON(dio_pages_present(sdio) == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212 }
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300213 return dio->pages[sdio->head];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214}
215
Zach Brown6d544bb2006-12-10 02:20:54 -0800216/**
217 * dio_complete() - called when all DIO BIO I/O has been completed
218 * @offset: the byte offset in the file of the completed operation
219 *
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200220 * This drops i_dio_count, lets interested parties know that a DIO operation
221 * has completed, and calculates the resulting return code for the operation.
Zach Brown6d544bb2006-12-10 02:20:54 -0800222 *
223 * It lets the filesystem know if it registered an interest earlier via
224 * get_block. Pass the private field of the map buffer_head so that
225 * filesystems can use it to hold additional state between get_block calls and
226 * dio_complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227 */
Christoph Hellwig716b9bc2016-04-07 08:51:59 -0700228static ssize_t dio_complete(struct dio *dio, ssize_t ret, bool is_async)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229{
Christoph Hellwig716b9bc2016-04-07 08:51:59 -0700230 loff_t offset = dio->iocb->ki_pos;
Zach Brown6d544bb2006-12-10 02:20:54 -0800231 ssize_t transferred = 0;
232
Zach Brown8459d862006-12-10 02:21:05 -0800233 /*
234 * AIO submission can race with bio completion to get here while
235 * expecting to have the last io completed by bio completion.
236 * In that case -EIOCBQUEUED is in fact not an error we want
237 * to preserve through this call.
238 */
239 if (ret == -EIOCBQUEUED)
240 ret = 0;
241
Zach Brown6d544bb2006-12-10 02:20:54 -0800242 if (dio->result) {
243 transferred = dio->result;
244
245 /* Check for short read case */
Mike Christie8a4c1e42016-06-05 14:31:50 -0500246 if ((dio->op == REQ_OP_READ) &&
247 ((offset + transferred) > dio->i_size))
Zach Brown6d544bb2006-12-10 02:20:54 -0800248 transferred = dio->i_size - offset;
Al Viro4038acd2016-10-03 20:38:55 -0400249 /* ignore EFAULT if some IO has been done */
250 if (unlikely(ret == -EFAULT) && transferred)
251 ret = 0;
Zach Brown6d544bb2006-12-10 02:20:54 -0800252 }
253
Zach Brown6d544bb2006-12-10 02:20:54 -0800254 if (ret == 0)
255 ret = dio->page_errors;
256 if (ret == 0)
257 ret = dio->io_error;
258 if (ret == 0)
259 ret = transferred;
260
Christoph Hellwig187372a2016-02-08 14:40:51 +1100261 if (dio->end_io) {
262 int err;
263
Christoph Hellwige2592212016-04-07 08:52:01 -0700264 // XXX: ki_pos??
Christoph Hellwig187372a2016-02-08 14:40:51 +1100265 err = dio->end_io(dio->iocb, offset, ret, dio->private);
266 if (err)
267 ret = err;
268 }
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000269
Jens Axboefe0f07d2015-04-15 17:05:48 -0600270 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
271 inode_dio_end(dio->inode);
272
Christoph Hellwig02afc272013-09-04 15:04:40 +0200273 if (is_async) {
Christoph Hellwige2592212016-04-07 08:52:01 -0700274 /*
275 * generic_write_sync expects ki_pos to have been updated
276 * already, but the submission path only does this for
277 * synchronous I/O.
278 */
279 dio->iocb->ki_pos += transferred;
Christoph Hellwig02afc272013-09-04 15:04:40 +0200280
Mike Christie8a4c1e42016-06-05 14:31:50 -0500281 if (dio->op == REQ_OP_WRITE)
Christoph Hellwige2592212016-04-07 08:52:01 -0700282 ret = generic_write_sync(dio->iocb, transferred);
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100283 dio->iocb->ki_complete(dio->iocb, ret, 0);
Christoph Hellwig02afc272013-09-04 15:04:40 +0200284 }
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200285
286 kmem_cache_free(dio_cache, dio);
Zach Brown6d544bb2006-12-10 02:20:54 -0800287 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288}
289
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200290static void dio_aio_complete_work(struct work_struct *work)
291{
292 struct dio *dio = container_of(work, struct dio, complete_work);
293
Christoph Hellwig716b9bc2016-04-07 08:51:59 -0700294 dio_complete(dio, 0, true);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200295}
296
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297static int dio_bio_complete(struct dio *dio, struct bio *bio);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200298
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299/*
300 * Asynchronous IO callback.
301 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200302static void dio_bio_end_aio(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303{
304 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800305 unsigned long remaining;
306 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308 /* cleanup the bio */
309 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800310
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800311 spin_lock_irqsave(&dio->bio_lock, flags);
312 remaining = --dio->refcount;
313 if (remaining == 1 && dio->waiter)
Zach Brown20258b22006-12-10 02:21:01 -0800314 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800315 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b22006-12-10 02:21:01 -0800316
Zach Brown8459d862006-12-10 02:21:05 -0800317 if (remaining == 0) {
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200318 if (dio->result && dio->defer_completion) {
319 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
320 queue_work(dio->inode->i_sb->s_dio_done_wq,
321 &dio->complete_work);
322 } else {
Christoph Hellwig716b9bc2016-04-07 08:51:59 -0700323 dio_complete(dio, 0, true);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200324 }
Zach Brown8459d862006-12-10 02:21:05 -0800325 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326}
327
328/*
329 * The BIO completion handler simply queues the BIO up for the process-context
330 * handler.
331 *
332 * During I/O bi_private points at the dio. After I/O, bi_private is used to
333 * implement a singly-linked list of completed BIOs, at dio->bio_list.
334 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200335static void dio_bio_end_io(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336{
337 struct dio *dio = bio->bi_private;
338 unsigned long flags;
339
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 spin_lock_irqsave(&dio->bio_lock, flags);
341 bio->bi_private = dio->bio_list;
342 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800343 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344 wake_up_process(dio->waiter);
345 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346}
347
Josef Bacikfacd07b2010-05-23 11:00:55 -0400348/**
349 * dio_end_io - handle the end io action for the given bio
350 * @bio: The direct io bio thats being completed
351 * @error: Error if there was one
352 *
353 * This is meant to be called by any filesystem that uses their own dio_submit_t
354 * so that the DIO specific endio actions are dealt with after the filesystem
355 * has done it's completion work.
356 */
357void dio_end_io(struct bio *bio, int error)
358{
359 struct dio *dio = bio->bi_private;
360
361 if (dio->is_async)
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200362 dio_bio_end_aio(bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400363 else
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200364 dio_bio_end_io(bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400365}
366EXPORT_SYMBOL_GPL(dio_end_io);
367
Andi Kleenba253fb2011-08-01 21:38:08 -0700368static inline void
Andi Kleeneb28be22011-08-01 21:38:03 -0700369dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
370 struct block_device *bdev,
371 sector_t first_sector, int nr_vecs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372{
373 struct bio *bio;
374
David Dillow20d96002011-01-20 14:44:22 -0800375 /*
376 * bio_alloc() is guaranteed to return a bio when called with
Mel Gorman71baba42015-11-06 16:28:28 -0800377 * __GFP_RECLAIM and we request a valid number of vectors.
David Dillow20d96002011-01-20 14:44:22 -0800378 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380
381 bio->bi_bdev = bdev;
Kent Overstreet4f024f32013-10-11 15:44:27 -0700382 bio->bi_iter.bi_sector = first_sector;
Mike Christie8a4c1e42016-06-05 14:31:50 -0500383 bio_set_op_attrs(bio, dio->op, dio->op_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384 if (dio->is_async)
385 bio->bi_end_io = dio_bio_end_aio;
386 else
387 bio->bi_end_io = dio_bio_end_io;
388
Andi Kleeneb28be22011-08-01 21:38:03 -0700389 sdio->bio = bio;
390 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391}
392
393/*
394 * In the AIO read case we speculatively dirty the pages before starting IO.
395 * During IO completion, any of these pages which happen to have been written
396 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800397 *
398 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700400static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401{
Andi Kleeneb28be22011-08-01 21:38:03 -0700402 struct bio *bio = sdio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800403 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404
405 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800406
407 spin_lock_irqsave(&dio->bio_lock, flags);
408 dio->refcount++;
409 spin_unlock_irqrestore(&dio->bio_lock, flags);
410
Mike Christie8a4c1e42016-06-05 14:31:50 -0500411 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800413
Jens Axboec1c53462015-11-10 10:14:38 -0700414 dio->bio_bdev = bio->bi_bdev;
415
Jens Axboe15c4f632015-10-27 14:09:51 +0900416 if (sdio->submit_io) {
Mike Christie8a4c1e42016-06-05 14:31:50 -0500417 sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
Jens Axboe15c4f632015-10-27 14:09:51 +0900418 dio->bio_cookie = BLK_QC_T_NONE;
Jens Axboec1c53462015-11-10 10:14:38 -0700419 } else
Mike Christie4e49ea42016-06-05 14:31:41 -0500420 dio->bio_cookie = submit_bio(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421
Andi Kleeneb28be22011-08-01 21:38:03 -0700422 sdio->bio = NULL;
423 sdio->boundary = 0;
424 sdio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425}
426
427/*
428 * Release any resources in case of a failure
429 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700430static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431{
Al Viro7b2c99d2014-03-15 04:05:57 -0400432 while (sdio->head < sdio->tail)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300433 put_page(dio->pages[sdio->head++]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434}
435
436/*
Zach Brown02732012006-12-10 02:20:59 -0800437 * Wait for the next BIO to complete. Remove it and return it. NULL is
438 * returned once all BIOs have been completed. This must only be called once
439 * all bios have been issued so that dio->refcount can only decrease. This
440 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441 */
442static struct bio *dio_await_one(struct dio *dio)
443{
444 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800445 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446
447 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800448
449 /*
450 * Wait as long as the list is empty and there are bios in flight. bio
451 * completion drops the count, maybe adds to the list, and wakes while
452 * holding the bio_lock so we don't need set_current_state()'s barrier
453 * and can call it after testing our condition.
454 */
455 while (dio->refcount > 1 && dio->bio_list == NULL) {
456 __set_current_state(TASK_UNINTERRUPTIBLE);
457 dio->waiter = current;
458 spin_unlock_irqrestore(&dio->bio_lock, flags);
Christoph Hellwigc43c83a2016-03-03 16:04:02 +0100459 if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
460 !blk_poll(bdev_get_queue(dio->bio_bdev), dio->bio_cookie))
Jens Axboe15c4f632015-10-27 14:09:51 +0900461 io_schedule();
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800462 /* wake up sets us TASK_RUNNING */
463 spin_lock_irqsave(&dio->bio_lock, flags);
464 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 }
Zach Brown02732012006-12-10 02:20:59 -0800466 if (dio->bio_list) {
467 bio = dio->bio_list;
468 dio->bio_list = bio->bi_private;
469 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470 spin_unlock_irqrestore(&dio->bio_lock, flags);
471 return bio;
472}
473
474/*
475 * Process one completed BIO. No locks are held.
476 */
477static int dio_bio_complete(struct dio *dio, struct bio *bio)
478{
Kent Overstreetcb34e052012-09-05 15:22:02 -0700479 struct bio_vec *bvec;
480 unsigned i;
Sasha Levin9b81c842015-08-10 19:05:18 -0400481 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200483 if (bio->bi_error)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800484 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485
Mike Christie8a4c1e42016-06-05 14:31:50 -0500486 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) {
Sasha Levin9b81c842015-08-10 19:05:18 -0400487 err = bio->bi_error;
Mike Krinkin7ddc9712016-01-30 19:09:59 +0300488 bio_check_pages_dirty(bio); /* transfers ownership */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 } else {
Kent Overstreetcb34e052012-09-05 15:22:02 -0700490 bio_for_each_segment_all(bvec, bio, i) {
491 struct page *page = bvec->bv_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492
Mike Christie8a4c1e42016-06-05 14:31:50 -0500493 if (dio->op == REQ_OP_READ && !PageCompound(page) &&
Ming Lei53cbf3b2015-08-17 10:31:46 +0800494 dio->should_dirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495 set_page_dirty_lock(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300496 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497 }
Sasha Levin9b81c842015-08-10 19:05:18 -0400498 err = bio->bi_error;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499 bio_put(bio);
500 }
Sasha Levin9b81c842015-08-10 19:05:18 -0400501 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502}
503
504/*
Zach Brown02732012006-12-10 02:20:59 -0800505 * Wait on and process all in-flight BIOs. This must only be called once
506 * all bios have been issued so that the refcount can only decrease.
507 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200508 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800509 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800511static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512{
Zach Brown02732012006-12-10 02:20:59 -0800513 struct bio *bio;
514 do {
515 bio = dio_await_one(dio);
516 if (bio)
517 dio_bio_complete(dio, bio);
518 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519}
520
521/*
522 * A really large O_DIRECT read or write can generate a lot of BIOs. So
523 * to keep the memory consumption sane we periodically reap any completed BIOs
524 * during the BIO generation phase.
525 *
526 * This also helps to limit the peak amount of pinned userspace memory.
527 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700528static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529{
530 int ret = 0;
531
Andi Kleeneb28be22011-08-01 21:38:03 -0700532 if (sdio->reap_counter++ >= 64) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 while (dio->bio_list) {
534 unsigned long flags;
535 struct bio *bio;
536 int ret2;
537
538 spin_lock_irqsave(&dio->bio_lock, flags);
539 bio = dio->bio_list;
540 dio->bio_list = bio->bi_private;
541 spin_unlock_irqrestore(&dio->bio_lock, flags);
542 ret2 = dio_bio_complete(dio, bio);
543 if (ret == 0)
544 ret = ret2;
545 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700546 sdio->reap_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547 }
548 return ret;
549}
550
551/*
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200552 * Create workqueue for deferred direct IO completions. We allocate the
553 * workqueue when it's first needed. This avoids creating workqueue for
554 * filesystems that don't need it and also allows us to create the workqueue
555 * late enough so the we can include s_id in the name of the workqueue.
556 */
557static int sb_init_dio_done_wq(struct super_block *sb)
558{
Olof Johansson45150c42013-09-09 10:34:23 -0700559 struct workqueue_struct *old;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200560 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
561 WQ_MEM_RECLAIM, 0,
562 sb->s_id);
563 if (!wq)
564 return -ENOMEM;
565 /*
566 * This has to be atomic as more DIOs can race to create the workqueue
567 */
Olof Johansson45150c42013-09-09 10:34:23 -0700568 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200569 /* Someone created workqueue before us? Free ours... */
Olof Johansson45150c42013-09-09 10:34:23 -0700570 if (old)
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200571 destroy_workqueue(wq);
572 return 0;
573}
574
575static int dio_set_defer_completion(struct dio *dio)
576{
577 struct super_block *sb = dio->inode->i_sb;
578
579 if (dio->defer_completion)
580 return 0;
581 dio->defer_completion = true;
582 if (!sb->s_dio_done_wq)
583 return sb_init_dio_done_wq(sb);
584 return 0;
585}
586
587/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588 * Call into the fs to map some more disk blocks. We record the current number
Andi Kleeneb28be22011-08-01 21:38:03 -0700589 * of available blocks at sdio->blocks_available. These are in units of the
Fabian Frederick61604a22017-02-27 14:28:32 -0800590 * fs blocksize, i_blocksize(inode).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 *
592 * The fs is allowed to map lots of blocks at once. If it wants to do that,
593 * it uses the passed inode-relative block number as the file offset, as usual.
594 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800595 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596 * has remaining to do. The fs should not map more than this number of blocks.
597 *
598 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
599 * indicate how much contiguous disk space has been made available at
600 * bh->b_blocknr.
601 *
602 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
603 * This isn't very efficient...
604 *
605 * In the case of filesystem holes: the fs may return an arbitrarily-large
606 * hole by returning an appropriate value in b_size and by clearing
607 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800608 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 */
Andi Kleen18772642011-08-01 21:38:07 -0700610static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
611 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612{
613 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
Tao Maae55e1a2012-01-12 17:20:33 -0800615 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616 unsigned long fs_count; /* Number of filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 int create;
Linus Torvaldsab738572012-11-29 12:27:00 -0800618 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619
620 /*
621 * If there was a memory error and we've overwritten all the
622 * mapped blocks then we can now return that memory error
623 */
624 ret = dio->page_errors;
625 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700626 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
627 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
Tao Maae55e1a2012-01-12 17:20:33 -0800628 fs_endblk = (sdio->final_block_in_request - 1) >>
629 sdio->blkfactor;
630 fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631
Nathan Scott3c674e72006-03-29 09:26:15 +1000632 map_bh->b_state = 0;
Linus Torvaldsab738572012-11-29 12:27:00 -0800633 map_bh->b_size = fs_count << i_blkbits;
Nathan Scott3c674e72006-03-29 09:26:15 +1000634
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800635 /*
Eryu Guan9ecd10b2016-05-27 14:27:18 -0700636 * For writes that could fill holes inside i_size on a
637 * DIO_SKIP_HOLES filesystem we forbid block creations: only
638 * overwrites are permitted. We will return early to the caller
639 * once we see an unmapped buffer head returned, and the caller
640 * will fall back to buffered I/O.
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800641 *
642 * Otherwise the decision is left to the get_blocks method,
643 * which may decide to handle it or also return an unmapped
644 * buffer head.
645 */
Mike Christie8a4c1e42016-06-05 14:31:50 -0500646 create = dio->op == REQ_OP_WRITE;
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800647 if (dio->flags & DIO_SKIP_HOLES) {
Eryu Guan9ecd10b2016-05-27 14:27:18 -0700648 if (fs_startblk <= ((i_size_read(dio->inode) - 1) >>
649 i_blkbits))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000652
Andi Kleeneb28be22011-08-01 21:38:03 -0700653 ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 map_bh, create);
Andi Kleen18772642011-08-01 21:38:07 -0700655
656 /* Store for completion */
657 dio->private = map_bh->b_private;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200658
659 if (ret == 0 && buffer_defer_completion(map_bh))
660 ret = dio_set_defer_completion(dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661 }
662 return ret;
663}
664
665/*
666 * There is no bio. Make one now.
667 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700668static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
669 sector_t start_sector, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670{
671 sector_t sector;
672 int ret, nr_pages;
673
Andi Kleeneb28be22011-08-01 21:38:03 -0700674 ret = dio_bio_reap(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 if (ret)
676 goto out;
Andi Kleeneb28be22011-08-01 21:38:03 -0700677 sector = start_sector << (sdio->blkbits - 9);
Kent Overstreetb54ffb72015-05-19 14:31:01 +0200678 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 BUG_ON(nr_pages <= 0);
Andi Kleen18772642011-08-01 21:38:07 -0700680 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
Andi Kleeneb28be22011-08-01 21:38:03 -0700681 sdio->boundary = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682out:
683 return ret;
684}
685
686/*
687 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
688 * that was successful then update final_block_in_bio and take a ref against
689 * the just-added page.
690 *
691 * Return zero on success. Non-zero means the caller needs to start a new BIO.
692 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700693static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694{
695 int ret;
696
Andi Kleeneb28be22011-08-01 21:38:03 -0700697 ret = bio_add_page(sdio->bio, sdio->cur_page,
698 sdio->cur_page_len, sdio->cur_page_offset);
699 if (ret == sdio->cur_page_len) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 /*
701 * Decrement count only, if we are done with this page
702 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700703 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
704 sdio->pages_in_io--;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300705 get_page(sdio->cur_page);
Andi Kleeneb28be22011-08-01 21:38:03 -0700706 sdio->final_block_in_bio = sdio->cur_page_block +
707 (sdio->cur_page_len >> sdio->blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708 ret = 0;
709 } else {
710 ret = 1;
711 }
712 return ret;
713}
714
715/*
716 * Put cur_page under IO. The section of cur_page which is described by
717 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
718 * starts on-disk at cur_page_block.
719 *
720 * We take a ref against the page here (on behalf of its presence in the bio).
721 *
722 * The caller of this function is responsible for removing cur_page from the
723 * dio, and for dropping the refcount which came from that presence.
724 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700725static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
726 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727{
728 int ret = 0;
729
Andi Kleeneb28be22011-08-01 21:38:03 -0700730 if (sdio->bio) {
731 loff_t cur_offset = sdio->cur_page_fs_offset;
732 loff_t bio_next_offset = sdio->logical_offset_in_bio +
Kent Overstreet4f024f32013-10-11 15:44:27 -0700733 sdio->bio->bi_iter.bi_size;
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400734
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400736 * See whether this new request is contiguous with the old.
737 *
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900738 * Btrfs cannot handle having logically non-contiguous requests
739 * submitted. For example if you have
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400740 *
741 * Logical: [0-4095][HOLE][8192-12287]
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900742 * Physical: [0-4095] [4096-8191]
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400743 *
744 * We cannot submit those pages together as one BIO. So if our
745 * current logical offset in the file does not equal what would
746 * be the next logical offset in the bio, submit the bio we
747 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700749 if (sdio->final_block_in_bio != sdio->cur_page_block ||
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400750 cur_offset != bio_next_offset)
Andi Kleeneb28be22011-08-01 21:38:03 -0700751 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 }
753
Andi Kleeneb28be22011-08-01 21:38:03 -0700754 if (sdio->bio == NULL) {
Andi Kleen18772642011-08-01 21:38:07 -0700755 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756 if (ret)
757 goto out;
758 }
759
Andi Kleeneb28be22011-08-01 21:38:03 -0700760 if (dio_bio_add_page(sdio) != 0) {
761 dio_bio_submit(dio, sdio);
Andi Kleen18772642011-08-01 21:38:07 -0700762 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700764 ret = dio_bio_add_page(sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 BUG_ON(ret != 0);
766 }
767 }
768out:
769 return ret;
770}
771
772/*
773 * An autonomous function to put a chunk of a page under deferred IO.
774 *
775 * The caller doesn't actually know (or care) whether this piece of page is in
776 * a BIO, or is under IO or whatever. We just take care of all possible
777 * situations here. The separation between the logic of do_direct_IO() and
778 * that of submit_page_section() is important for clarity. Please don't break.
779 *
780 * The chunk of page starts on-disk at blocknr.
781 *
782 * We perform deferred IO, by recording the last-submitted page inside our
783 * private part of the dio structure. If possible, we just expand the IO
784 * across that page here.
785 *
786 * If that doesn't work out then we put the old page into the bio and add this
787 * page to the dio instead.
788 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700789static inline int
Andi Kleeneb28be22011-08-01 21:38:03 -0700790submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
Andi Kleen18772642011-08-01 21:38:07 -0700791 unsigned offset, unsigned len, sector_t blocknr,
792 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793{
794 int ret = 0;
795
Mike Christie8a4c1e42016-06-05 14:31:50 -0500796 if (dio->op == REQ_OP_WRITE) {
Andrew Morton98c4d572006-12-10 02:19:47 -0800797 /*
798 * Read accounting is performed in submit_bio()
799 */
800 task_io_account_write(len);
801 }
802
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 /*
804 * Can we just grow the current page's presence in the dio?
805 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700806 if (sdio->cur_page == page &&
807 sdio->cur_page_offset + sdio->cur_page_len == offset &&
808 sdio->cur_page_block +
809 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
810 sdio->cur_page_len += len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 goto out;
812 }
813
814 /*
815 * If there's a deferred page already there then send it.
816 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700817 if (sdio->cur_page) {
Andi Kleen18772642011-08-01 21:38:07 -0700818 ret = dio_send_cur_page(dio, sdio, map_bh);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300819 put_page(sdio->cur_page);
Andi Kleeneb28be22011-08-01 21:38:03 -0700820 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821 if (ret)
Jan Karab1058b92013-04-29 15:06:18 -0700822 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823 }
824
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300825 get_page(page); /* It is in dio */
Andi Kleeneb28be22011-08-01 21:38:03 -0700826 sdio->cur_page = page;
827 sdio->cur_page_offset = offset;
828 sdio->cur_page_len = len;
829 sdio->cur_page_block = blocknr;
830 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831out:
Jan Karab1058b92013-04-29 15:06:18 -0700832 /*
833 * If sdio->boundary then we want to schedule the IO now to
834 * avoid metadata seeks.
835 */
836 if (sdio->boundary) {
837 ret = dio_send_cur_page(dio, sdio, map_bh);
Andreas Gruenbacherf9139a12017-10-09 11:13:18 +0200838 if (sdio->bio)
839 dio_bio_submit(dio, sdio);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300840 put_page(sdio->cur_page);
Jan Karab1058b92013-04-29 15:06:18 -0700841 sdio->cur_page = NULL;
842 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 return ret;
844}
845
846/*
847 * Clean any dirty buffers in the blockdev mapping which alias newly-created
848 * file blocks. Only called for S_ISREG files - blockdevs do not set
849 * buffer_new
850 */
Andi Kleen18772642011-08-01 21:38:07 -0700851static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852{
853 unsigned i;
854 unsigned nblocks;
855
Andi Kleen18772642011-08-01 21:38:07 -0700856 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857
858 for (i = 0; i < nblocks; i++) {
Andi Kleen18772642011-08-01 21:38:07 -0700859 unmap_underlying_metadata(map_bh->b_bdev,
860 map_bh->b_blocknr + i);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861 }
862}
863
864/*
865 * If we are not writing the entire block and get_block() allocated
866 * the block for us, we need to fill-in the unused portion of the
867 * block with zeros. This happens only if user-buffer, fileoffset or
868 * io length is not filesystem block-size multiple.
869 *
870 * `end' is zero if we're doing the start of the IO, 1 at the end of the
871 * IO.
872 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700873static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
874 int end, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875{
876 unsigned dio_blocks_per_fs_block;
877 unsigned this_chunk_blocks; /* In dio_blocks */
878 unsigned this_chunk_bytes;
879 struct page *page;
880
Andi Kleeneb28be22011-08-01 21:38:03 -0700881 sdio->start_zero_done = 1;
Andi Kleen18772642011-08-01 21:38:07 -0700882 if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 return;
884
Andi Kleeneb28be22011-08-01 21:38:03 -0700885 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
886 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887
888 if (!this_chunk_blocks)
889 return;
890
891 /*
892 * We need to zero out part of an fs block. It is either at the
893 * beginning or the end of the fs block.
894 */
895 if (end)
896 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
897
Andi Kleeneb28be22011-08-01 21:38:03 -0700898 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899
Nick Piggin557ed1f2007-10-16 01:24:40 -0700900 page = ZERO_PAGE(0);
Andi Kleeneb28be22011-08-01 21:38:03 -0700901 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700902 sdio->next_block_for_io, map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903 return;
904
Andi Kleeneb28be22011-08-01 21:38:03 -0700905 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906}
907
908/*
909 * Walk the user pages, and the file, mapping blocks to disk and generating
910 * a sequence of (page,offset,len,block) mappings. These mappings are injected
911 * into submit_page_section(), which takes care of the next stage of submission
912 *
913 * Direct IO against a blockdev is different from a file. Because we can
914 * happily perform page-sized but 512-byte aligned IOs. It is important that
915 * blockdev IO be able to have fine alignment and large sizes.
916 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800917 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 * with the size of IO which is permitted at this offset and this i_blkbits.
919 *
920 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800921 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 * fine alignment but still allows this function to work in PAGE_SIZE units.
923 */
Andi Kleen18772642011-08-01 21:38:07 -0700924static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
925 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926{
Andi Kleeneb28be22011-08-01 21:38:03 -0700927 const unsigned blkbits = sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 int ret = 0;
929
Andi Kleeneb28be22011-08-01 21:38:03 -0700930 while (sdio->block_in_file < sdio->final_block_in_request) {
Al Viro7b2c99d2014-03-15 04:05:57 -0400931 struct page *page;
932 size_t from, to;
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300933
934 page = dio_get_page(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 if (IS_ERR(page)) {
936 ret = PTR_ERR(page);
937 goto out;
938 }
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300939 from = sdio->head ? 0 : sdio->from;
940 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
941 sdio->head++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942
Al Viro7b2c99d2014-03-15 04:05:57 -0400943 while (from < to) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944 unsigned this_chunk_bytes; /* # of bytes mapped */
945 unsigned this_chunk_blocks; /* # of blocks */
946 unsigned u;
947
Andi Kleeneb28be22011-08-01 21:38:03 -0700948 if (sdio->blocks_available == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 /*
950 * Need to go and map some more disk
951 */
952 unsigned long blkmask;
953 unsigned long dio_remainder;
954
Andi Kleen18772642011-08-01 21:38:07 -0700955 ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 if (ret) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300957 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 goto out;
959 }
960 if (!buffer_mapped(map_bh))
961 goto do_holes;
962
Andi Kleeneb28be22011-08-01 21:38:03 -0700963 sdio->blocks_available =
964 map_bh->b_size >> sdio->blkbits;
965 sdio->next_block_for_io =
966 map_bh->b_blocknr << sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967 if (buffer_new(map_bh))
Andi Kleen18772642011-08-01 21:38:07 -0700968 clean_blockdev_aliases(dio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969
Andi Kleeneb28be22011-08-01 21:38:03 -0700970 if (!sdio->blkfactor)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971 goto do_holes;
972
Andi Kleeneb28be22011-08-01 21:38:03 -0700973 blkmask = (1 << sdio->blkfactor) - 1;
974 dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975
976 /*
977 * If we are at the start of IO and that IO
978 * starts partway into a fs-block,
979 * dio_remainder will be non-zero. If the IO
980 * is a read then we can simply advance the IO
981 * cursor to the first block which is to be
982 * read. But if the IO is a write and the
983 * block was newly allocated we cannot do that;
984 * the start of the fs block must be zeroed out
985 * on-disk
986 */
987 if (!buffer_new(map_bh))
Andi Kleeneb28be22011-08-01 21:38:03 -0700988 sdio->next_block_for_io += dio_remainder;
989 sdio->blocks_available -= dio_remainder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990 }
991do_holes:
992 /* Handle holes */
993 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800994 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995
996 /* AKPM: eargh, -ENOTBLK is a hack */
Mike Christie8a4c1e42016-06-05 14:31:50 -0500997 if (dio->op == REQ_OP_WRITE) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300998 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999 return -ENOTBLK;
1000 }
1001
Jeff Moyer35dc8162006-02-03 03:04:27 -08001002 /*
1003 * Be sure to account for a partial block as the
1004 * last block in the file
1005 */
1006 i_size_aligned = ALIGN(i_size_read(dio->inode),
1007 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -07001008 if (sdio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -08001009 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010 /* We hit eof */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001011 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012 goto out;
1013 }
Al Viro7b2c99d2014-03-15 04:05:57 -04001014 zero_user(page, from, 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -07001015 sdio->block_in_file++;
Al Viro7b2c99d2014-03-15 04:05:57 -04001016 from += 1 << blkbits;
Al Viro3320c602014-03-10 02:30:55 -04001017 dio->result += 1 << blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001018 goto next_block;
1019 }
1020
1021 /*
1022 * If we're performing IO which has an alignment which
1023 * is finer than the underlying fs, go check to see if
1024 * we must zero out the start of this block.
1025 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001026 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
Andi Kleen18772642011-08-01 21:38:07 -07001027 dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028
1029 /*
1030 * Work out, in this_chunk_blocks, how much disk we
1031 * can add to this page
1032 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001033 this_chunk_blocks = sdio->blocks_available;
Al Viro7b2c99d2014-03-15 04:05:57 -04001034 u = (to - from) >> blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035 if (this_chunk_blocks > u)
1036 this_chunk_blocks = u;
Andi Kleeneb28be22011-08-01 21:38:03 -07001037 u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 if (this_chunk_blocks > u)
1039 this_chunk_blocks = u;
1040 this_chunk_bytes = this_chunk_blocks << blkbits;
1041 BUG_ON(this_chunk_bytes == 0);
1042
Jan Kara092c8d42013-04-29 15:06:17 -07001043 if (this_chunk_blocks == sdio->blocks_available)
1044 sdio->boundary = buffer_boundary(map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -07001045 ret = submit_page_section(dio, sdio, page,
Al Viro7b2c99d2014-03-15 04:05:57 -04001046 from,
Andi Kleeneb28be22011-08-01 21:38:03 -07001047 this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -07001048 sdio->next_block_for_io,
1049 map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050 if (ret) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001051 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 goto out;
1053 }
Andi Kleeneb28be22011-08-01 21:38:03 -07001054 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055
Andi Kleeneb28be22011-08-01 21:38:03 -07001056 sdio->block_in_file += this_chunk_blocks;
Al Viro7b2c99d2014-03-15 04:05:57 -04001057 from += this_chunk_bytes;
1058 dio->result += this_chunk_bytes;
Andi Kleeneb28be22011-08-01 21:38:03 -07001059 sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060next_block:
Andi Kleeneb28be22011-08-01 21:38:03 -07001061 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1062 if (sdio->block_in_file == sdio->final_block_in_request)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063 break;
1064 }
1065
1066 /* Drop the ref which was taken in get_user_pages() */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001067 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068 }
1069out:
1070 return ret;
1071}
1072
Andi Kleen847cc632011-08-01 21:38:09 -07001073static inline int drop_refcount(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001074{
Andi Kleen847cc632011-08-01 21:38:09 -07001075 int ret2;
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001076 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077
Zach Brown8459d862006-12-10 02:21:05 -08001078 /*
1079 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001080 * operation. AIO can if it was a broken operation described above or
1081 * in fact if all the bios race to complete before we get here. In
1082 * that case dio_complete() translates the EIOCBQUEUED into the proper
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001083 * return code that the caller will hand to ->complete().
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001084 *
1085 * This is managed by the bio_lock instead of being an atomic_t so that
1086 * completion paths can drop their ref and use the remaining count to
1087 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001088 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001089 spin_lock_irqsave(&dio->bio_lock, flags);
1090 ret2 = --dio->refcount;
1091 spin_unlock_irqrestore(&dio->bio_lock, flags);
Andi Kleen847cc632011-08-01 21:38:09 -07001092 return ret2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093}
1094
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001095/*
1096 * This is a library function for use by filesystem drivers.
1097 *
1098 * The locking rules are governed by the flags parameter:
1099 * - if the flags value contains DIO_LOCKING we use a fancy locking
1100 * scheme for dumb filesystems.
1101 * For writes this function is called under i_mutex and returns with
1102 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1103 * taken and dropped again before returning.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001104 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1105 * internal locking but rather rely on the filesystem to synchronize
1106 * direct I/O reads/writes versus each other and truncate.
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001107 *
1108 * To help with locking against truncate we incremented the i_dio_count
1109 * counter before starting direct I/O, and decrement it once we are done.
1110 * Truncate can wait for it to reach zero to provide exclusion. It is
1111 * expected that filesystem provide exclusion between new direct I/O
1112 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1113 * but other filesystems need to take care of this on their own.
Andi Kleenba253fb2011-08-01 21:38:08 -07001114 *
1115 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1116 * is always inlined. Otherwise gcc is unable to split the structure into
1117 * individual fields and will generate much worse code. This is important
1118 * for the whole file.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001119 */
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001120static inline ssize_t
Omar Sandoval17f8c842015-03-16 04:33:50 -07001121do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1122 struct block_device *bdev, struct iov_iter *iter,
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001123 get_block_t get_block, dio_iodone_t end_io,
Omar Sandoval17f8c842015-03-16 04:33:50 -07001124 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125{
Linus Torvaldsab738572012-11-29 12:27:00 -08001126 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1127 unsigned blkbits = i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 unsigned blocksize_mask = (1 << blkbits) - 1;
1129 ssize_t retval = -EINVAL;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001130 size_t count = iov_iter_count(iter);
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001131 loff_t offset = iocb->ki_pos;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001132 loff_t end = offset + count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 struct dio *dio;
Andi Kleeneb28be22011-08-01 21:38:03 -07001134 struct dio_submit sdio = { 0, };
Andi Kleen847cc632011-08-01 21:38:09 -07001135 struct buffer_head map_bh = { 0, };
Fengguang Wu647d1e42012-08-09 15:23:09 +02001136 struct blk_plug plug;
Al Viro886a3912014-03-05 13:50:45 -05001137 unsigned long align = offset | iov_iter_alignment(iter);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001139 /*
1140 * Avoid references to bdev if not absolutely needed to give
1141 * the early prefetch in the caller enough time.
1142 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143
Al Viro886a3912014-03-05 13:50:45 -05001144 if (align & blocksize_mask) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001146 blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 blocksize_mask = (1 << blkbits) - 1;
Al Viro886a3912014-03-05 13:50:45 -05001148 if (align & blocksize_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 goto out;
1150 }
1151
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001152 /* watch out for a 0 len io from a tricksy fs */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001153 if (iov_iter_rw(iter) == READ && !iov_iter_count(iter))
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001154 return 0;
1155
Andi Kleen6e8267f2011-08-01 21:38:06 -07001156 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 retval = -ENOMEM;
1158 if (!dio)
1159 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001160 /*
1161 * Believe it or not, zeroing out the page array caused a .5%
1162 * performance regression in a database benchmark. So, we take
1163 * care to only zero out what's needed.
1164 */
1165 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001167 dio->flags = flags;
1168 if (dio->flags & DIO_LOCKING) {
Omar Sandoval17f8c842015-03-16 04:33:50 -07001169 if (iov_iter_rw(iter) == READ) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001170 struct address_space *mapping =
1171 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001173 /* will be released by direct_io_worker */
Al Viro59551022016-01-22 15:40:57 -05001174 inode_lock(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175
1176 retval = filemap_write_and_wait_range(mapping, offset,
1177 end - 1);
1178 if (retval) {
Al Viro59551022016-01-22 15:40:57 -05001179 inode_unlock(inode);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001180 kmem_cache_free(dio_cache, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 goto out;
1182 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184 }
1185
Jan Kara74cedf92015-11-30 10:15:42 -07001186 /* Once we sampled i_size check for reads beyond EOF */
1187 dio->i_size = i_size_read(inode);
1188 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1189 if (dio->flags & DIO_LOCKING)
Al Viro59551022016-01-22 15:40:57 -05001190 inode_unlock(inode);
Jan Kara74cedf92015-11-30 10:15:42 -07001191 kmem_cache_free(dio_cache, dio);
Al Viro2d4594a2015-12-08 12:22:47 -05001192 retval = 0;
Jan Kara74cedf92015-11-30 10:15:42 -07001193 goto out;
1194 }
1195
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 /*
Christoph Hellwig60392572014-02-10 10:27:11 +11001197 * For file extending writes updating i_size before data writeouts
1198 * complete can expose uninitialized blocks in dumb filesystems.
1199 * In that case we need to wait for I/O completion even if asked
1200 * for an asynchronous write.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201 */
Christoph Hellwig60392572014-02-10 10:27:11 +11001202 if (is_sync_kiocb(iocb))
1203 dio->is_async = false;
1204 else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
Omar Sandoval17f8c842015-03-16 04:33:50 -07001205 iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
Christoph Hellwig60392572014-02-10 10:27:11 +11001206 dio->is_async = false;
1207 else
1208 dio->is_async = true;
1209
Andi Kleen847cc632011-08-01 21:38:09 -07001210 dio->inode = inode;
Mike Christie8a4c1e42016-06-05 14:31:50 -05001211 if (iov_iter_rw(iter) == WRITE) {
1212 dio->op = REQ_OP_WRITE;
1213 dio->op_flags = WRITE_ODIRECT;
1214 } else {
1215 dio->op = REQ_OP_READ;
1216 }
Christoph Hellwig02afc272013-09-04 15:04:40 +02001217
1218 /*
1219 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1220 * so that we can call ->fsync.
1221 */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001222 if (dio->is_async && iov_iter_rw(iter) == WRITE &&
Christoph Hellwig02afc272013-09-04 15:04:40 +02001223 ((iocb->ki_filp->f_flags & O_DSYNC) ||
1224 IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1225 retval = dio_set_defer_completion(dio);
1226 if (retval) {
1227 /*
1228 * We grab i_mutex only for reads so we don't have
1229 * to release it here
1230 */
1231 kmem_cache_free(dio_cache, dio);
1232 goto out;
1233 }
1234 }
1235
1236 /*
1237 * Will be decremented at I/O completion time.
1238 */
Jens Axboefe0f07d2015-04-15 17:05:48 -06001239 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
1240 inode_dio_begin(inode);
Christoph Hellwig02afc272013-09-04 15:04:40 +02001241
1242 retval = 0;
Andi Kleen847cc632011-08-01 21:38:09 -07001243 sdio.blkbits = blkbits;
Linus Torvaldsab738572012-11-29 12:27:00 -08001244 sdio.blkfactor = i_blkbits - blkbits;
Andi Kleen847cc632011-08-01 21:38:09 -07001245 sdio.block_in_file = offset >> blkbits;
1246
1247 sdio.get_block = get_block;
1248 dio->end_io = end_io;
1249 sdio.submit_io = submit_io;
1250 sdio.final_block_in_bio = -1;
1251 sdio.next_block_for_io = -1;
1252
1253 dio->iocb = iocb;
Andi Kleen847cc632011-08-01 21:38:09 -07001254
1255 spin_lock_init(&dio->bio_lock);
1256 dio->refcount = 1;
1257
Ming Lei53cbf3b2015-08-17 10:31:46 +08001258 dio->should_dirty = (iter->type == ITER_IOVEC);
Al Viro7b2c99d2014-03-15 04:05:57 -04001259 sdio.iter = iter;
1260 sdio.final_block_in_request =
1261 (offset + iov_iter_count(iter)) >> blkbits;
1262
Andi Kleen847cc632011-08-01 21:38:09 -07001263 /*
1264 * In case of non-aligned buffers, we may need 2 more
1265 * pages since we need to zero out first and last block.
1266 */
1267 if (unlikely(sdio.blkfactor))
1268 sdio.pages_in_io = 2;
1269
Al Virof67da302014-03-19 01:16:16 -04001270 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
Andi Kleen847cc632011-08-01 21:38:09 -07001271
Fengguang Wu647d1e42012-08-09 15:23:09 +02001272 blk_start_plug(&plug);
1273
Al Viro7b2c99d2014-03-15 04:05:57 -04001274 retval = do_direct_IO(dio, &sdio, &map_bh);
1275 if (retval)
1276 dio_cleanup(dio, &sdio);
Andi Kleen847cc632011-08-01 21:38:09 -07001277
1278 if (retval == -ENOTBLK) {
1279 /*
1280 * The remaining part of the request will be
1281 * be handled by buffered I/O when we return
1282 */
1283 retval = 0;
1284 }
1285 /*
1286 * There may be some unwritten disk at the end of a part-written
1287 * fs-block-sized block. Go zero that now.
1288 */
1289 dio_zero_block(dio, &sdio, 1, &map_bh);
1290
1291 if (sdio.cur_page) {
1292 ssize_t ret2;
1293
1294 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1295 if (retval == 0)
1296 retval = ret2;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001297 put_page(sdio.cur_page);
Andi Kleen847cc632011-08-01 21:38:09 -07001298 sdio.cur_page = NULL;
1299 }
1300 if (sdio.bio)
1301 dio_bio_submit(dio, &sdio);
1302
Fengguang Wu647d1e42012-08-09 15:23:09 +02001303 blk_finish_plug(&plug);
1304
Andi Kleen847cc632011-08-01 21:38:09 -07001305 /*
1306 * It is possible that, we return short IO due to end of file.
1307 * In that case, we need to release all the pages we got hold on.
1308 */
1309 dio_cleanup(dio, &sdio);
1310
1311 /*
1312 * All block lookups have been performed. For READ requests
1313 * we can let i_mutex go now that its achieved its purpose
1314 * of protecting us from looking up uninitialized blocks.
1315 */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001316 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
Al Viro59551022016-01-22 15:40:57 -05001317 inode_unlock(dio->inode);
Andi Kleen847cc632011-08-01 21:38:09 -07001318
1319 /*
1320 * The only time we want to leave bios in flight is when a successful
1321 * partial aio read or full aio write have been setup. In that case
1322 * bio completion will call aio_complete. The only time it's safe to
1323 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1324 * This had *better* be the only place that raises -EIOCBQUEUED.
1325 */
1326 BUG_ON(retval == -EIOCBQUEUED);
1327 if (dio->is_async && retval == 0 && dio->result &&
Omar Sandoval17f8c842015-03-16 04:33:50 -07001328 (iov_iter_rw(iter) == READ || dio->result == count))
Andi Kleen847cc632011-08-01 21:38:09 -07001329 retval = -EIOCBQUEUED;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001330 else
Andi Kleen847cc632011-08-01 21:38:09 -07001331 dio_await_completion(dio);
1332
1333 if (drop_refcount(dio) == 0) {
Christoph Hellwig716b9bc2016-04-07 08:51:59 -07001334 retval = dio_complete(dio, retval, false);
Andi Kleen847cc632011-08-01 21:38:09 -07001335 } else
1336 BUG_ON(retval != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001338out:
1339 return retval;
1340}
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001341
Omar Sandoval17f8c842015-03-16 04:33:50 -07001342ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1343 struct block_device *bdev, struct iov_iter *iter,
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001344 get_block_t get_block,
Omar Sandoval17f8c842015-03-16 04:33:50 -07001345 dio_iodone_t end_io, dio_submit_t submit_io,
1346 int flags)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001347{
1348 /*
1349 * The block device state is needed in the end to finally
1350 * submit everything. Since it's likely to be cache cold
1351 * prefetch it here as first thing to hide some of the
1352 * latency.
1353 *
1354 * Attempt to prefetch the pieces we likely need later.
1355 */
1356 prefetch(&bdev->bd_disk->part_tbl);
1357 prefetch(bdev->bd_queue);
1358 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1359
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001360 return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
Omar Sandoval17f8c842015-03-16 04:33:50 -07001361 end_io, submit_io, flags);
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001362}
1363
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364EXPORT_SYMBOL(__blockdev_direct_IO);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001365
1366static __init int dio_init(void)
1367{
1368 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1369 return 0;
1370}
1371module_init(dio_init)