blob: 472037732daf00c093846a93967a4858f0e382c7 [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 Hellwig7b7a8662013-09-04 15:04:39 +0200227static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret,
228 bool is_async)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229{
Zach Brown6d544bb2006-12-10 02:20:54 -0800230 ssize_t transferred = 0;
231
Zach Brown8459d862006-12-10 02:21:05 -0800232 /*
233 * AIO submission can race with bio completion to get here while
234 * expecting to have the last io completed by bio completion.
235 * In that case -EIOCBQUEUED is in fact not an error we want
236 * to preserve through this call.
237 */
238 if (ret == -EIOCBQUEUED)
239 ret = 0;
240
Zach Brown6d544bb2006-12-10 02:20:54 -0800241 if (dio->result) {
242 transferred = dio->result;
243
244 /* Check for short read case */
245 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
246 transferred = dio->i_size - offset;
247 }
248
Zach Brown6d544bb2006-12-10 02:20:54 -0800249 if (ret == 0)
250 ret = dio->page_errors;
251 if (ret == 0)
252 ret = dio->io_error;
253 if (ret == 0)
254 ret = transferred;
255
Christoph Hellwig187372a2016-02-08 14:40:51 +1100256 if (dio->end_io) {
257 int err;
258
259 err = dio->end_io(dio->iocb, offset, ret, dio->private);
260 if (err)
261 ret = err;
262 }
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000263
Jens Axboefe0f07d2015-04-15 17:05:48 -0600264 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
265 inode_dio_end(dio->inode);
266
Christoph Hellwig02afc272013-09-04 15:04:40 +0200267 if (is_async) {
268 if (dio->rw & WRITE) {
269 int err;
270
271 err = generic_write_sync(dio->iocb->ki_filp, offset,
272 transferred);
273 if (err < 0 && ret > 0)
274 ret = err;
275 }
276
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100277 dio->iocb->ki_complete(dio->iocb, ret, 0);
Christoph Hellwig02afc272013-09-04 15:04:40 +0200278 }
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200279
280 kmem_cache_free(dio_cache, dio);
Zach Brown6d544bb2006-12-10 02:20:54 -0800281 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282}
283
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200284static void dio_aio_complete_work(struct work_struct *work)
285{
286 struct dio *dio = container_of(work, struct dio, complete_work);
287
288 dio_complete(dio, dio->iocb->ki_pos, 0, true);
289}
290
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291static int dio_bio_complete(struct dio *dio, struct bio *bio);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200292
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293/*
294 * Asynchronous IO callback.
295 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200296static void dio_bio_end_aio(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297{
298 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800299 unsigned long remaining;
300 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302 /* cleanup the bio */
303 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800304
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800305 spin_lock_irqsave(&dio->bio_lock, flags);
306 remaining = --dio->refcount;
307 if (remaining == 1 && dio->waiter)
Zach Brown20258b22006-12-10 02:21:01 -0800308 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800309 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b22006-12-10 02:21:01 -0800310
Zach Brown8459d862006-12-10 02:21:05 -0800311 if (remaining == 0) {
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200312 if (dio->result && dio->defer_completion) {
313 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
314 queue_work(dio->inode->i_sb->s_dio_done_wq,
315 &dio->complete_work);
316 } else {
317 dio_complete(dio, dio->iocb->ki_pos, 0, true);
318 }
Zach Brown8459d862006-12-10 02:21:05 -0800319 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320}
321
322/*
323 * The BIO completion handler simply queues the BIO up for the process-context
324 * handler.
325 *
326 * During I/O bi_private points at the dio. After I/O, bi_private is used to
327 * implement a singly-linked list of completed BIOs, at dio->bio_list.
328 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200329static void dio_bio_end_io(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330{
331 struct dio *dio = bio->bi_private;
332 unsigned long flags;
333
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334 spin_lock_irqsave(&dio->bio_lock, flags);
335 bio->bi_private = dio->bio_list;
336 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800337 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338 wake_up_process(dio->waiter);
339 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340}
341
Josef Bacikfacd07b2010-05-23 11:00:55 -0400342/**
343 * dio_end_io - handle the end io action for the given bio
344 * @bio: The direct io bio thats being completed
345 * @error: Error if there was one
346 *
347 * This is meant to be called by any filesystem that uses their own dio_submit_t
348 * so that the DIO specific endio actions are dealt with after the filesystem
349 * has done it's completion work.
350 */
351void dio_end_io(struct bio *bio, int error)
352{
353 struct dio *dio = bio->bi_private;
354
355 if (dio->is_async)
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200356 dio_bio_end_aio(bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400357 else
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200358 dio_bio_end_io(bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400359}
360EXPORT_SYMBOL_GPL(dio_end_io);
361
Andi Kleenba253fb2011-08-01 21:38:08 -0700362static inline void
Andi Kleeneb28be22011-08-01 21:38:03 -0700363dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
364 struct block_device *bdev,
365 sector_t first_sector, int nr_vecs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366{
367 struct bio *bio;
368
David Dillow20d96002011-01-20 14:44:22 -0800369 /*
370 * bio_alloc() is guaranteed to return a bio when called with
Mel Gorman71baba42015-11-06 16:28:28 -0800371 * __GFP_RECLAIM and we request a valid number of vectors.
David Dillow20d96002011-01-20 14:44:22 -0800372 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374
375 bio->bi_bdev = bdev;
Kent Overstreet4f024f32013-10-11 15:44:27 -0700376 bio->bi_iter.bi_sector = first_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 if (dio->is_async)
378 bio->bi_end_io = dio_bio_end_aio;
379 else
380 bio->bi_end_io = dio_bio_end_io;
381
Andi Kleeneb28be22011-08-01 21:38:03 -0700382 sdio->bio = bio;
383 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384}
385
386/*
387 * In the AIO read case we speculatively dirty the pages before starting IO.
388 * During IO completion, any of these pages which happen to have been written
389 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800390 *
391 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700393static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394{
Andi Kleeneb28be22011-08-01 21:38:03 -0700395 struct bio *bio = sdio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800396 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397
398 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800399
400 spin_lock_irqsave(&dio->bio_lock, flags);
401 dio->refcount++;
402 spin_unlock_irqrestore(&dio->bio_lock, flags);
403
Ming Lei53cbf3b2015-08-17 10:31:46 +0800404 if (dio->is_async && dio->rw == READ && dio->should_dirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800406
Jens Axboec1c53462015-11-10 10:14:38 -0700407 dio->bio_bdev = bio->bi_bdev;
408
Jens Axboe15c4f632015-10-27 14:09:51 +0900409 if (sdio->submit_io) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700410 sdio->submit_io(dio->rw, bio, dio->inode,
411 sdio->logical_offset_in_bio);
Jens Axboe15c4f632015-10-27 14:09:51 +0900412 dio->bio_cookie = BLK_QC_T_NONE;
Jens Axboec1c53462015-11-10 10:14:38 -0700413 } else
Jens Axboe15c4f632015-10-27 14:09:51 +0900414 dio->bio_cookie = submit_bio(dio->rw, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415
Andi Kleeneb28be22011-08-01 21:38:03 -0700416 sdio->bio = NULL;
417 sdio->boundary = 0;
418 sdio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419}
420
421/*
422 * Release any resources in case of a failure
423 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700424static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425{
Al Viro7b2c99d2014-03-15 04:05:57 -0400426 while (sdio->head < sdio->tail)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300427 put_page(dio->pages[sdio->head++]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428}
429
430/*
Zach Brown02732012006-12-10 02:20:59 -0800431 * Wait for the next BIO to complete. Remove it and return it. NULL is
432 * returned once all BIOs have been completed. This must only be called once
433 * all bios have been issued so that dio->refcount can only decrease. This
434 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435 */
436static struct bio *dio_await_one(struct dio *dio)
437{
438 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800439 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440
441 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800442
443 /*
444 * Wait as long as the list is empty and there are bios in flight. bio
445 * completion drops the count, maybe adds to the list, and wakes while
446 * holding the bio_lock so we don't need set_current_state()'s barrier
447 * and can call it after testing our condition.
448 */
449 while (dio->refcount > 1 && dio->bio_list == NULL) {
450 __set_current_state(TASK_UNINTERRUPTIBLE);
451 dio->waiter = current;
452 spin_unlock_irqrestore(&dio->bio_lock, flags);
Christoph Hellwigc43c83a2016-03-03 16:04:02 +0100453 if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
454 !blk_poll(bdev_get_queue(dio->bio_bdev), dio->bio_cookie))
Jens Axboe15c4f632015-10-27 14:09:51 +0900455 io_schedule();
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800456 /* wake up sets us TASK_RUNNING */
457 spin_lock_irqsave(&dio->bio_lock, flags);
458 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 }
Zach Brown02732012006-12-10 02:20:59 -0800460 if (dio->bio_list) {
461 bio = dio->bio_list;
462 dio->bio_list = bio->bi_private;
463 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 spin_unlock_irqrestore(&dio->bio_lock, flags);
465 return bio;
466}
467
468/*
469 * Process one completed BIO. No locks are held.
470 */
471static int dio_bio_complete(struct dio *dio, struct bio *bio)
472{
Kent Overstreetcb34e052012-09-05 15:22:02 -0700473 struct bio_vec *bvec;
474 unsigned i;
Sasha Levin9b81c842015-08-10 19:05:18 -0400475 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200477 if (bio->bi_error)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800478 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479
Ming Lei53cbf3b2015-08-17 10:31:46 +0800480 if (dio->is_async && dio->rw == READ && dio->should_dirty) {
Sasha Levin9b81c842015-08-10 19:05:18 -0400481 err = bio->bi_error;
Mike Krinkin7ddc9712016-01-30 19:09:59 +0300482 bio_check_pages_dirty(bio); /* transfers ownership */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483 } else {
Kent Overstreetcb34e052012-09-05 15:22:02 -0700484 bio_for_each_segment_all(bvec, bio, i) {
485 struct page *page = bvec->bv_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486
Ming Lei53cbf3b2015-08-17 10:31:46 +0800487 if (dio->rw == READ && !PageCompound(page) &&
488 dio->should_dirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 set_page_dirty_lock(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300490 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491 }
Sasha Levin9b81c842015-08-10 19:05:18 -0400492 err = bio->bi_error;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 bio_put(bio);
494 }
Sasha Levin9b81c842015-08-10 19:05:18 -0400495 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496}
497
498/*
Zach Brown02732012006-12-10 02:20:59 -0800499 * Wait on and process all in-flight BIOs. This must only be called once
500 * all bios have been issued so that the refcount can only decrease.
501 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200502 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800503 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800505static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506{
Zach Brown02732012006-12-10 02:20:59 -0800507 struct bio *bio;
508 do {
509 bio = dio_await_one(dio);
510 if (bio)
511 dio_bio_complete(dio, bio);
512 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513}
514
515/*
516 * A really large O_DIRECT read or write can generate a lot of BIOs. So
517 * to keep the memory consumption sane we periodically reap any completed BIOs
518 * during the BIO generation phase.
519 *
520 * This also helps to limit the peak amount of pinned userspace memory.
521 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700522static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523{
524 int ret = 0;
525
Andi Kleeneb28be22011-08-01 21:38:03 -0700526 if (sdio->reap_counter++ >= 64) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527 while (dio->bio_list) {
528 unsigned long flags;
529 struct bio *bio;
530 int ret2;
531
532 spin_lock_irqsave(&dio->bio_lock, flags);
533 bio = dio->bio_list;
534 dio->bio_list = bio->bi_private;
535 spin_unlock_irqrestore(&dio->bio_lock, flags);
536 ret2 = dio_bio_complete(dio, bio);
537 if (ret == 0)
538 ret = ret2;
539 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700540 sdio->reap_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541 }
542 return ret;
543}
544
545/*
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200546 * Create workqueue for deferred direct IO completions. We allocate the
547 * workqueue when it's first needed. This avoids creating workqueue for
548 * filesystems that don't need it and also allows us to create the workqueue
549 * late enough so the we can include s_id in the name of the workqueue.
550 */
551static int sb_init_dio_done_wq(struct super_block *sb)
552{
Olof Johansson45150c42013-09-09 10:34:23 -0700553 struct workqueue_struct *old;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200554 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
555 WQ_MEM_RECLAIM, 0,
556 sb->s_id);
557 if (!wq)
558 return -ENOMEM;
559 /*
560 * This has to be atomic as more DIOs can race to create the workqueue
561 */
Olof Johansson45150c42013-09-09 10:34:23 -0700562 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200563 /* Someone created workqueue before us? Free ours... */
Olof Johansson45150c42013-09-09 10:34:23 -0700564 if (old)
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200565 destroy_workqueue(wq);
566 return 0;
567}
568
569static int dio_set_defer_completion(struct dio *dio)
570{
571 struct super_block *sb = dio->inode->i_sb;
572
573 if (dio->defer_completion)
574 return 0;
575 dio->defer_completion = true;
576 if (!sb->s_dio_done_wq)
577 return sb_init_dio_done_wq(sb);
578 return 0;
579}
580
581/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 * Call into the fs to map some more disk blocks. We record the current number
Andi Kleeneb28be22011-08-01 21:38:03 -0700583 * of available blocks at sdio->blocks_available. These are in units of the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 * fs blocksize, (1 << inode->i_blkbits).
585 *
586 * The fs is allowed to map lots of blocks at once. If it wants to do that,
587 * it uses the passed inode-relative block number as the file offset, as usual.
588 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800589 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590 * has remaining to do. The fs should not map more than this number of blocks.
591 *
592 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
593 * indicate how much contiguous disk space has been made available at
594 * bh->b_blocknr.
595 *
596 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
597 * This isn't very efficient...
598 *
599 * In the case of filesystem holes: the fs may return an arbitrarily-large
600 * hole by returning an appropriate value in b_size and by clearing
601 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800602 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603 */
Andi Kleen18772642011-08-01 21:38:07 -0700604static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
605 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606{
607 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
Tao Maae55e1a2012-01-12 17:20:33 -0800609 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610 unsigned long fs_count; /* Number of filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611 int create;
Linus Torvaldsab738572012-11-29 12:27:00 -0800612 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613
614 /*
615 * If there was a memory error and we've overwritten all the
616 * mapped blocks then we can now return that memory error
617 */
618 ret = dio->page_errors;
619 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700620 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
621 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
Tao Maae55e1a2012-01-12 17:20:33 -0800622 fs_endblk = (sdio->final_block_in_request - 1) >>
623 sdio->blkfactor;
624 fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625
Nathan Scott3c674e72006-03-29 09:26:15 +1000626 map_bh->b_state = 0;
Linus Torvaldsab738572012-11-29 12:27:00 -0800627 map_bh->b_size = fs_count << i_blkbits;
Nathan Scott3c674e72006-03-29 09:26:15 +1000628
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800629 /*
630 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
631 * forbid block creations: only overwrites are permitted.
632 * We will return early to the caller once we see an
633 * unmapped buffer head returned, and the caller will fall
634 * back to buffered I/O.
635 *
636 * Otherwise the decision is left to the get_blocks method,
637 * which may decide to handle it or also return an unmapped
638 * buffer head.
639 */
Jens Axboeb31dc662006-06-13 08:26:10 +0200640 create = dio->rw & WRITE;
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800641 if (dio->flags & DIO_SKIP_HOLES) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700642 if (sdio->block_in_file < (i_size_read(dio->inode) >>
643 sdio->blkbits))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000646
Andi Kleeneb28be22011-08-01 21:38:03 -0700647 ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648 map_bh, create);
Andi Kleen18772642011-08-01 21:38:07 -0700649
650 /* Store for completion */
651 dio->private = map_bh->b_private;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200652
653 if (ret == 0 && buffer_defer_completion(map_bh))
654 ret = dio_set_defer_completion(dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655 }
656 return ret;
657}
658
659/*
660 * There is no bio. Make one now.
661 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700662static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
663 sector_t start_sector, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664{
665 sector_t sector;
666 int ret, nr_pages;
667
Andi Kleeneb28be22011-08-01 21:38:03 -0700668 ret = dio_bio_reap(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669 if (ret)
670 goto out;
Andi Kleeneb28be22011-08-01 21:38:03 -0700671 sector = start_sector << (sdio->blkbits - 9);
Kent Overstreetb54ffb72015-05-19 14:31:01 +0200672 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 BUG_ON(nr_pages <= 0);
Andi Kleen18772642011-08-01 21:38:07 -0700674 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
Andi Kleeneb28be22011-08-01 21:38:03 -0700675 sdio->boundary = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676out:
677 return ret;
678}
679
680/*
681 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
682 * that was successful then update final_block_in_bio and take a ref against
683 * the just-added page.
684 *
685 * Return zero on success. Non-zero means the caller needs to start a new BIO.
686 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700687static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688{
689 int ret;
690
Andi Kleeneb28be22011-08-01 21:38:03 -0700691 ret = bio_add_page(sdio->bio, sdio->cur_page,
692 sdio->cur_page_len, sdio->cur_page_offset);
693 if (ret == sdio->cur_page_len) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 /*
695 * Decrement count only, if we are done with this page
696 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700697 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
698 sdio->pages_in_io--;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300699 get_page(sdio->cur_page);
Andi Kleeneb28be22011-08-01 21:38:03 -0700700 sdio->final_block_in_bio = sdio->cur_page_block +
701 (sdio->cur_page_len >> sdio->blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702 ret = 0;
703 } else {
704 ret = 1;
705 }
706 return ret;
707}
708
709/*
710 * Put cur_page under IO. The section of cur_page which is described by
711 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
712 * starts on-disk at cur_page_block.
713 *
714 * We take a ref against the page here (on behalf of its presence in the bio).
715 *
716 * The caller of this function is responsible for removing cur_page from the
717 * dio, and for dropping the refcount which came from that presence.
718 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700719static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
720 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721{
722 int ret = 0;
723
Andi Kleeneb28be22011-08-01 21:38:03 -0700724 if (sdio->bio) {
725 loff_t cur_offset = sdio->cur_page_fs_offset;
726 loff_t bio_next_offset = sdio->logical_offset_in_bio +
Kent Overstreet4f024f32013-10-11 15:44:27 -0700727 sdio->bio->bi_iter.bi_size;
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400728
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400730 * See whether this new request is contiguous with the old.
731 *
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900732 * Btrfs cannot handle having logically non-contiguous requests
733 * submitted. For example if you have
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400734 *
735 * Logical: [0-4095][HOLE][8192-12287]
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900736 * Physical: [0-4095] [4096-8191]
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400737 *
738 * We cannot submit those pages together as one BIO. So if our
739 * current logical offset in the file does not equal what would
740 * be the next logical offset in the bio, submit the bio we
741 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700743 if (sdio->final_block_in_bio != sdio->cur_page_block ||
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400744 cur_offset != bio_next_offset)
Andi Kleeneb28be22011-08-01 21:38:03 -0700745 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 }
747
Andi Kleeneb28be22011-08-01 21:38:03 -0700748 if (sdio->bio == NULL) {
Andi Kleen18772642011-08-01 21:38:07 -0700749 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 if (ret)
751 goto out;
752 }
753
Andi Kleeneb28be22011-08-01 21:38:03 -0700754 if (dio_bio_add_page(sdio) != 0) {
755 dio_bio_submit(dio, sdio);
Andi Kleen18772642011-08-01 21:38:07 -0700756 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700758 ret = dio_bio_add_page(sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759 BUG_ON(ret != 0);
760 }
761 }
762out:
763 return ret;
764}
765
766/*
767 * An autonomous function to put a chunk of a page under deferred IO.
768 *
769 * The caller doesn't actually know (or care) whether this piece of page is in
770 * a BIO, or is under IO or whatever. We just take care of all possible
771 * situations here. The separation between the logic of do_direct_IO() and
772 * that of submit_page_section() is important for clarity. Please don't break.
773 *
774 * The chunk of page starts on-disk at blocknr.
775 *
776 * We perform deferred IO, by recording the last-submitted page inside our
777 * private part of the dio structure. If possible, we just expand the IO
778 * across that page here.
779 *
780 * If that doesn't work out then we put the old page into the bio and add this
781 * page to the dio instead.
782 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700783static inline int
Andi Kleeneb28be22011-08-01 21:38:03 -0700784submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
Andi Kleen18772642011-08-01 21:38:07 -0700785 unsigned offset, unsigned len, sector_t blocknr,
786 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787{
788 int ret = 0;
789
Andrew Morton98c4d572006-12-10 02:19:47 -0800790 if (dio->rw & WRITE) {
791 /*
792 * Read accounting is performed in submit_bio()
793 */
794 task_io_account_write(len);
795 }
796
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797 /*
798 * Can we just grow the current page's presence in the dio?
799 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700800 if (sdio->cur_page == page &&
801 sdio->cur_page_offset + sdio->cur_page_len == offset &&
802 sdio->cur_page_block +
803 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
804 sdio->cur_page_len += len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 goto out;
806 }
807
808 /*
809 * If there's a deferred page already there then send it.
810 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700811 if (sdio->cur_page) {
Andi Kleen18772642011-08-01 21:38:07 -0700812 ret = dio_send_cur_page(dio, sdio, map_bh);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300813 put_page(sdio->cur_page);
Andi Kleeneb28be22011-08-01 21:38:03 -0700814 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 if (ret)
Jan Karab1058b92013-04-29 15:06:18 -0700816 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817 }
818
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300819 get_page(page); /* It is in dio */
Andi Kleeneb28be22011-08-01 21:38:03 -0700820 sdio->cur_page = page;
821 sdio->cur_page_offset = offset;
822 sdio->cur_page_len = len;
823 sdio->cur_page_block = blocknr;
824 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825out:
Jan Karab1058b92013-04-29 15:06:18 -0700826 /*
827 * If sdio->boundary then we want to schedule the IO now to
828 * avoid metadata seeks.
829 */
830 if (sdio->boundary) {
831 ret = dio_send_cur_page(dio, sdio, map_bh);
832 dio_bio_submit(dio, sdio);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300833 put_page(sdio->cur_page);
Jan Karab1058b92013-04-29 15:06:18 -0700834 sdio->cur_page = NULL;
835 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836 return ret;
837}
838
839/*
840 * Clean any dirty buffers in the blockdev mapping which alias newly-created
841 * file blocks. Only called for S_ISREG files - blockdevs do not set
842 * buffer_new
843 */
Andi Kleen18772642011-08-01 21:38:07 -0700844static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845{
846 unsigned i;
847 unsigned nblocks;
848
Andi Kleen18772642011-08-01 21:38:07 -0700849 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850
851 for (i = 0; i < nblocks; i++) {
Andi Kleen18772642011-08-01 21:38:07 -0700852 unmap_underlying_metadata(map_bh->b_bdev,
853 map_bh->b_blocknr + i);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 }
855}
856
857/*
858 * If we are not writing the entire block and get_block() allocated
859 * the block for us, we need to fill-in the unused portion of the
860 * block with zeros. This happens only if user-buffer, fileoffset or
861 * io length is not filesystem block-size multiple.
862 *
863 * `end' is zero if we're doing the start of the IO, 1 at the end of the
864 * IO.
865 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700866static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
867 int end, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868{
869 unsigned dio_blocks_per_fs_block;
870 unsigned this_chunk_blocks; /* In dio_blocks */
871 unsigned this_chunk_bytes;
872 struct page *page;
873
Andi Kleeneb28be22011-08-01 21:38:03 -0700874 sdio->start_zero_done = 1;
Andi Kleen18772642011-08-01 21:38:07 -0700875 if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 return;
877
Andi Kleeneb28be22011-08-01 21:38:03 -0700878 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
879 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880
881 if (!this_chunk_blocks)
882 return;
883
884 /*
885 * We need to zero out part of an fs block. It is either at the
886 * beginning or the end of the fs block.
887 */
888 if (end)
889 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
890
Andi Kleeneb28be22011-08-01 21:38:03 -0700891 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892
Nick Piggin557ed1f2007-10-16 01:24:40 -0700893 page = ZERO_PAGE(0);
Andi Kleeneb28be22011-08-01 21:38:03 -0700894 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700895 sdio->next_block_for_io, map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 return;
897
Andi Kleeneb28be22011-08-01 21:38:03 -0700898 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899}
900
901/*
902 * Walk the user pages, and the file, mapping blocks to disk and generating
903 * a sequence of (page,offset,len,block) mappings. These mappings are injected
904 * into submit_page_section(), which takes care of the next stage of submission
905 *
906 * Direct IO against a blockdev is different from a file. Because we can
907 * happily perform page-sized but 512-byte aligned IOs. It is important that
908 * blockdev IO be able to have fine alignment and large sizes.
909 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800910 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 * with the size of IO which is permitted at this offset and this i_blkbits.
912 *
913 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800914 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915 * fine alignment but still allows this function to work in PAGE_SIZE units.
916 */
Andi Kleen18772642011-08-01 21:38:07 -0700917static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
918 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919{
Andi Kleeneb28be22011-08-01 21:38:03 -0700920 const unsigned blkbits = sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 int ret = 0;
922
Andi Kleeneb28be22011-08-01 21:38:03 -0700923 while (sdio->block_in_file < sdio->final_block_in_request) {
Al Viro7b2c99d2014-03-15 04:05:57 -0400924 struct page *page;
925 size_t from, to;
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300926
927 page = dio_get_page(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 if (IS_ERR(page)) {
929 ret = PTR_ERR(page);
930 goto out;
931 }
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300932 from = sdio->head ? 0 : sdio->from;
933 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
934 sdio->head++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935
Al Viro7b2c99d2014-03-15 04:05:57 -0400936 while (from < to) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937 unsigned this_chunk_bytes; /* # of bytes mapped */
938 unsigned this_chunk_blocks; /* # of blocks */
939 unsigned u;
940
Andi Kleeneb28be22011-08-01 21:38:03 -0700941 if (sdio->blocks_available == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942 /*
943 * Need to go and map some more disk
944 */
945 unsigned long blkmask;
946 unsigned long dio_remainder;
947
Andi Kleen18772642011-08-01 21:38:07 -0700948 ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 if (ret) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300950 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951 goto out;
952 }
953 if (!buffer_mapped(map_bh))
954 goto do_holes;
955
Andi Kleeneb28be22011-08-01 21:38:03 -0700956 sdio->blocks_available =
957 map_bh->b_size >> sdio->blkbits;
958 sdio->next_block_for_io =
959 map_bh->b_blocknr << sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960 if (buffer_new(map_bh))
Andi Kleen18772642011-08-01 21:38:07 -0700961 clean_blockdev_aliases(dio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962
Andi Kleeneb28be22011-08-01 21:38:03 -0700963 if (!sdio->blkfactor)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 goto do_holes;
965
Andi Kleeneb28be22011-08-01 21:38:03 -0700966 blkmask = (1 << sdio->blkfactor) - 1;
967 dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968
969 /*
970 * If we are at the start of IO and that IO
971 * starts partway into a fs-block,
972 * dio_remainder will be non-zero. If the IO
973 * is a read then we can simply advance the IO
974 * cursor to the first block which is to be
975 * read. But if the IO is a write and the
976 * block was newly allocated we cannot do that;
977 * the start of the fs block must be zeroed out
978 * on-disk
979 */
980 if (!buffer_new(map_bh))
Andi Kleeneb28be22011-08-01 21:38:03 -0700981 sdio->next_block_for_io += dio_remainder;
982 sdio->blocks_available -= dio_remainder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983 }
984do_holes:
985 /* Handle holes */
986 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800987 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988
989 /* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboeb31dc662006-06-13 08:26:10 +0200990 if (dio->rw & WRITE) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300991 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992 return -ENOTBLK;
993 }
994
Jeff Moyer35dc8162006-02-03 03:04:27 -0800995 /*
996 * Be sure to account for a partial block as the
997 * last block in the file
998 */
999 i_size_aligned = ALIGN(i_size_read(dio->inode),
1000 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -07001001 if (sdio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -08001002 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003 /* We hit eof */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001004 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005 goto out;
1006 }
Al Viro7b2c99d2014-03-15 04:05:57 -04001007 zero_user(page, from, 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -07001008 sdio->block_in_file++;
Al Viro7b2c99d2014-03-15 04:05:57 -04001009 from += 1 << blkbits;
Al Viro3320c602014-03-10 02:30:55 -04001010 dio->result += 1 << blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011 goto next_block;
1012 }
1013
1014 /*
1015 * If we're performing IO which has an alignment which
1016 * is finer than the underlying fs, go check to see if
1017 * we must zero out the start of this block.
1018 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001019 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
Andi Kleen18772642011-08-01 21:38:07 -07001020 dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021
1022 /*
1023 * Work out, in this_chunk_blocks, how much disk we
1024 * can add to this page
1025 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001026 this_chunk_blocks = sdio->blocks_available;
Al Viro7b2c99d2014-03-15 04:05:57 -04001027 u = (to - from) >> blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028 if (this_chunk_blocks > u)
1029 this_chunk_blocks = u;
Andi Kleeneb28be22011-08-01 21:38:03 -07001030 u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 if (this_chunk_blocks > u)
1032 this_chunk_blocks = u;
1033 this_chunk_bytes = this_chunk_blocks << blkbits;
1034 BUG_ON(this_chunk_bytes == 0);
1035
Jan Kara092c8d42013-04-29 15:06:17 -07001036 if (this_chunk_blocks == sdio->blocks_available)
1037 sdio->boundary = buffer_boundary(map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -07001038 ret = submit_page_section(dio, sdio, page,
Al Viro7b2c99d2014-03-15 04:05:57 -04001039 from,
Andi Kleeneb28be22011-08-01 21:38:03 -07001040 this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -07001041 sdio->next_block_for_io,
1042 map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 if (ret) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001044 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045 goto out;
1046 }
Andi Kleeneb28be22011-08-01 21:38:03 -07001047 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048
Andi Kleeneb28be22011-08-01 21:38:03 -07001049 sdio->block_in_file += this_chunk_blocks;
Al Viro7b2c99d2014-03-15 04:05:57 -04001050 from += this_chunk_bytes;
1051 dio->result += this_chunk_bytes;
Andi Kleeneb28be22011-08-01 21:38:03 -07001052 sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053next_block:
Andi Kleeneb28be22011-08-01 21:38:03 -07001054 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1055 if (sdio->block_in_file == sdio->final_block_in_request)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056 break;
1057 }
1058
1059 /* Drop the ref which was taken in get_user_pages() */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001060 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061 }
1062out:
1063 return ret;
1064}
1065
Andi Kleen847cc632011-08-01 21:38:09 -07001066static inline int drop_refcount(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001067{
Andi Kleen847cc632011-08-01 21:38:09 -07001068 int ret2;
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001069 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070
Zach Brown8459d862006-12-10 02:21:05 -08001071 /*
1072 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001073 * operation. AIO can if it was a broken operation described above or
1074 * in fact if all the bios race to complete before we get here. In
1075 * that case dio_complete() translates the EIOCBQUEUED into the proper
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001076 * return code that the caller will hand to ->complete().
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001077 *
1078 * This is managed by the bio_lock instead of being an atomic_t so that
1079 * completion paths can drop their ref and use the remaining count to
1080 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001081 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001082 spin_lock_irqsave(&dio->bio_lock, flags);
1083 ret2 = --dio->refcount;
1084 spin_unlock_irqrestore(&dio->bio_lock, flags);
Andi Kleen847cc632011-08-01 21:38:09 -07001085 return ret2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086}
1087
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001088/*
1089 * This is a library function for use by filesystem drivers.
1090 *
1091 * The locking rules are governed by the flags parameter:
1092 * - if the flags value contains DIO_LOCKING we use a fancy locking
1093 * scheme for dumb filesystems.
1094 * For writes this function is called under i_mutex and returns with
1095 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1096 * taken and dropped again before returning.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001097 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1098 * internal locking but rather rely on the filesystem to synchronize
1099 * direct I/O reads/writes versus each other and truncate.
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001100 *
1101 * To help with locking against truncate we incremented the i_dio_count
1102 * counter before starting direct I/O, and decrement it once we are done.
1103 * Truncate can wait for it to reach zero to provide exclusion. It is
1104 * expected that filesystem provide exclusion between new direct I/O
1105 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1106 * but other filesystems need to take care of this on their own.
Andi Kleenba253fb2011-08-01 21:38:08 -07001107 *
1108 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1109 * is always inlined. Otherwise gcc is unable to split the structure into
1110 * individual fields and will generate much worse code. This is important
1111 * for the whole file.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001112 */
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001113static inline ssize_t
Omar Sandoval17f8c842015-03-16 04:33:50 -07001114do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1115 struct block_device *bdev, struct iov_iter *iter,
1116 loff_t offset, get_block_t get_block, dio_iodone_t end_io,
1117 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118{
Linus Torvaldsab738572012-11-29 12:27:00 -08001119 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1120 unsigned blkbits = i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 unsigned blocksize_mask = (1 << blkbits) - 1;
1122 ssize_t retval = -EINVAL;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001123 size_t count = iov_iter_count(iter);
1124 loff_t end = offset + count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 struct dio *dio;
Andi Kleeneb28be22011-08-01 21:38:03 -07001126 struct dio_submit sdio = { 0, };
Andi Kleen847cc632011-08-01 21:38:09 -07001127 struct buffer_head map_bh = { 0, };
Fengguang Wu647d1e42012-08-09 15:23:09 +02001128 struct blk_plug plug;
Al Viro886a3912014-03-05 13:50:45 -05001129 unsigned long align = offset | iov_iter_alignment(iter);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001131 /*
1132 * Avoid references to bdev if not absolutely needed to give
1133 * the early prefetch in the caller enough time.
1134 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135
Al Viro886a3912014-03-05 13:50:45 -05001136 if (align & blocksize_mask) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001138 blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 blocksize_mask = (1 << blkbits) - 1;
Al Viro886a3912014-03-05 13:50:45 -05001140 if (align & blocksize_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 goto out;
1142 }
1143
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001144 /* watch out for a 0 len io from a tricksy fs */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001145 if (iov_iter_rw(iter) == READ && !iov_iter_count(iter))
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001146 return 0;
1147
Andi Kleen6e8267f2011-08-01 21:38:06 -07001148 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 retval = -ENOMEM;
1150 if (!dio)
1151 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001152 /*
1153 * Believe it or not, zeroing out the page array caused a .5%
1154 * performance regression in a database benchmark. So, we take
1155 * care to only zero out what's needed.
1156 */
1157 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001159 dio->flags = flags;
1160 if (dio->flags & DIO_LOCKING) {
Omar Sandoval17f8c842015-03-16 04:33:50 -07001161 if (iov_iter_rw(iter) == READ) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001162 struct address_space *mapping =
1163 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001165 /* will be released by direct_io_worker */
Al Viro59551022016-01-22 15:40:57 -05001166 inode_lock(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167
1168 retval = filemap_write_and_wait_range(mapping, offset,
1169 end - 1);
1170 if (retval) {
Al Viro59551022016-01-22 15:40:57 -05001171 inode_unlock(inode);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001172 kmem_cache_free(dio_cache, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173 goto out;
1174 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176 }
1177
Jan Kara74cedf92015-11-30 10:15:42 -07001178 /* Once we sampled i_size check for reads beyond EOF */
1179 dio->i_size = i_size_read(inode);
1180 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1181 if (dio->flags & DIO_LOCKING)
Al Viro59551022016-01-22 15:40:57 -05001182 inode_unlock(inode);
Jan Kara74cedf92015-11-30 10:15:42 -07001183 kmem_cache_free(dio_cache, dio);
Al Viro2d4594a2015-12-08 12:22:47 -05001184 retval = 0;
Jan Kara74cedf92015-11-30 10:15:42 -07001185 goto out;
1186 }
1187
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188 /*
Christoph Hellwig60392572014-02-10 10:27:11 +11001189 * For file extending writes updating i_size before data writeouts
1190 * complete can expose uninitialized blocks in dumb filesystems.
1191 * In that case we need to wait for I/O completion even if asked
1192 * for an asynchronous write.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 */
Christoph Hellwig60392572014-02-10 10:27:11 +11001194 if (is_sync_kiocb(iocb))
1195 dio->is_async = false;
1196 else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
Omar Sandoval17f8c842015-03-16 04:33:50 -07001197 iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
Christoph Hellwig60392572014-02-10 10:27:11 +11001198 dio->is_async = false;
1199 else
1200 dio->is_async = true;
1201
Andi Kleen847cc632011-08-01 21:38:09 -07001202 dio->inode = inode;
Omar Sandoval17f8c842015-03-16 04:33:50 -07001203 dio->rw = iov_iter_rw(iter) == WRITE ? WRITE_ODIRECT : READ;
Christoph Hellwig02afc272013-09-04 15:04:40 +02001204
1205 /*
1206 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1207 * so that we can call ->fsync.
1208 */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001209 if (dio->is_async && iov_iter_rw(iter) == WRITE &&
Christoph Hellwig02afc272013-09-04 15:04:40 +02001210 ((iocb->ki_filp->f_flags & O_DSYNC) ||
1211 IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1212 retval = dio_set_defer_completion(dio);
1213 if (retval) {
1214 /*
1215 * We grab i_mutex only for reads so we don't have
1216 * to release it here
1217 */
1218 kmem_cache_free(dio_cache, dio);
1219 goto out;
1220 }
1221 }
1222
1223 /*
1224 * Will be decremented at I/O completion time.
1225 */
Jens Axboefe0f07d2015-04-15 17:05:48 -06001226 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
1227 inode_dio_begin(inode);
Christoph Hellwig02afc272013-09-04 15:04:40 +02001228
1229 retval = 0;
Andi Kleen847cc632011-08-01 21:38:09 -07001230 sdio.blkbits = blkbits;
Linus Torvaldsab738572012-11-29 12:27:00 -08001231 sdio.blkfactor = i_blkbits - blkbits;
Andi Kleen847cc632011-08-01 21:38:09 -07001232 sdio.block_in_file = offset >> blkbits;
1233
1234 sdio.get_block = get_block;
1235 dio->end_io = end_io;
1236 sdio.submit_io = submit_io;
1237 sdio.final_block_in_bio = -1;
1238 sdio.next_block_for_io = -1;
1239
1240 dio->iocb = iocb;
Andi Kleen847cc632011-08-01 21:38:09 -07001241
1242 spin_lock_init(&dio->bio_lock);
1243 dio->refcount = 1;
1244
Ming Lei53cbf3b2015-08-17 10:31:46 +08001245 dio->should_dirty = (iter->type == ITER_IOVEC);
Al Viro7b2c99d2014-03-15 04:05:57 -04001246 sdio.iter = iter;
1247 sdio.final_block_in_request =
1248 (offset + iov_iter_count(iter)) >> blkbits;
1249
Andi Kleen847cc632011-08-01 21:38:09 -07001250 /*
1251 * In case of non-aligned buffers, we may need 2 more
1252 * pages since we need to zero out first and last block.
1253 */
1254 if (unlikely(sdio.blkfactor))
1255 sdio.pages_in_io = 2;
1256
Al Virof67da302014-03-19 01:16:16 -04001257 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
Andi Kleen847cc632011-08-01 21:38:09 -07001258
Fengguang Wu647d1e42012-08-09 15:23:09 +02001259 blk_start_plug(&plug);
1260
Al Viro7b2c99d2014-03-15 04:05:57 -04001261 retval = do_direct_IO(dio, &sdio, &map_bh);
1262 if (retval)
1263 dio_cleanup(dio, &sdio);
Andi Kleen847cc632011-08-01 21:38:09 -07001264
1265 if (retval == -ENOTBLK) {
1266 /*
1267 * The remaining part of the request will be
1268 * be handled by buffered I/O when we return
1269 */
1270 retval = 0;
1271 }
1272 /*
1273 * There may be some unwritten disk at the end of a part-written
1274 * fs-block-sized block. Go zero that now.
1275 */
1276 dio_zero_block(dio, &sdio, 1, &map_bh);
1277
1278 if (sdio.cur_page) {
1279 ssize_t ret2;
1280
1281 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1282 if (retval == 0)
1283 retval = ret2;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001284 put_page(sdio.cur_page);
Andi Kleen847cc632011-08-01 21:38:09 -07001285 sdio.cur_page = NULL;
1286 }
1287 if (sdio.bio)
1288 dio_bio_submit(dio, &sdio);
1289
Fengguang Wu647d1e42012-08-09 15:23:09 +02001290 blk_finish_plug(&plug);
1291
Andi Kleen847cc632011-08-01 21:38:09 -07001292 /*
1293 * It is possible that, we return short IO due to end of file.
1294 * In that case, we need to release all the pages we got hold on.
1295 */
1296 dio_cleanup(dio, &sdio);
1297
1298 /*
1299 * All block lookups have been performed. For READ requests
1300 * we can let i_mutex go now that its achieved its purpose
1301 * of protecting us from looking up uninitialized blocks.
1302 */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001303 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
Al Viro59551022016-01-22 15:40:57 -05001304 inode_unlock(dio->inode);
Andi Kleen847cc632011-08-01 21:38:09 -07001305
1306 /*
1307 * The only time we want to leave bios in flight is when a successful
1308 * partial aio read or full aio write have been setup. In that case
1309 * bio completion will call aio_complete. The only time it's safe to
1310 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1311 * This had *better* be the only place that raises -EIOCBQUEUED.
1312 */
1313 BUG_ON(retval == -EIOCBQUEUED);
1314 if (dio->is_async && retval == 0 && dio->result &&
Omar Sandoval17f8c842015-03-16 04:33:50 -07001315 (iov_iter_rw(iter) == READ || dio->result == count))
Andi Kleen847cc632011-08-01 21:38:09 -07001316 retval = -EIOCBQUEUED;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001317 else
Andi Kleen847cc632011-08-01 21:38:09 -07001318 dio_await_completion(dio);
1319
1320 if (drop_refcount(dio) == 0) {
1321 retval = dio_complete(dio, offset, retval, false);
Andi Kleen847cc632011-08-01 21:38:09 -07001322 } else
1323 BUG_ON(retval != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001325out:
1326 return retval;
1327}
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001328
Omar Sandoval17f8c842015-03-16 04:33:50 -07001329ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1330 struct block_device *bdev, struct iov_iter *iter,
1331 loff_t offset, get_block_t get_block,
1332 dio_iodone_t end_io, dio_submit_t submit_io,
1333 int flags)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001334{
1335 /*
1336 * The block device state is needed in the end to finally
1337 * submit everything. Since it's likely to be cache cold
1338 * prefetch it here as first thing to hide some of the
1339 * latency.
1340 *
1341 * Attempt to prefetch the pieces we likely need later.
1342 */
1343 prefetch(&bdev->bd_disk->part_tbl);
1344 prefetch(bdev->bd_queue);
1345 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1346
Omar Sandoval17f8c842015-03-16 04:33:50 -07001347 return do_blockdev_direct_IO(iocb, inode, bdev, iter, offset, get_block,
1348 end_io, submit_io, flags);
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001349}
1350
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351EXPORT_SYMBOL(__blockdev_direct_IO);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001352
1353static __init int dio_init(void)
1354{
1355 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1356 return 0;
1357}
1358module_init(dio_init)