blob: 7c3ce73cb6170ee820aec560ece2fafc8b48b83c [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;
249 }
250
Zach Brown6d544bb2006-12-10 02:20:54 -0800251 if (ret == 0)
252 ret = dio->page_errors;
253 if (ret == 0)
254 ret = dio->io_error;
255 if (ret == 0)
256 ret = transferred;
257
Christoph Hellwig187372a2016-02-08 14:40:51 +1100258 if (dio->end_io) {
259 int err;
260
Christoph Hellwige2592212016-04-07 08:52:01 -0700261 // XXX: ki_pos??
Christoph Hellwig187372a2016-02-08 14:40:51 +1100262 err = dio->end_io(dio->iocb, offset, ret, dio->private);
263 if (err)
264 ret = err;
265 }
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000266
Jens Axboefe0f07d2015-04-15 17:05:48 -0600267 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
268 inode_dio_end(dio->inode);
269
Christoph Hellwig02afc272013-09-04 15:04:40 +0200270 if (is_async) {
Christoph Hellwige2592212016-04-07 08:52:01 -0700271 /*
272 * generic_write_sync expects ki_pos to have been updated
273 * already, but the submission path only does this for
274 * synchronous I/O.
275 */
276 dio->iocb->ki_pos += transferred;
Christoph Hellwig02afc272013-09-04 15:04:40 +0200277
Mike Christie8a4c1e42016-06-05 14:31:50 -0500278 if (dio->op == REQ_OP_WRITE)
Christoph Hellwige2592212016-04-07 08:52:01 -0700279 ret = generic_write_sync(dio->iocb, transferred);
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100280 dio->iocb->ki_complete(dio->iocb, ret, 0);
Christoph Hellwig02afc272013-09-04 15:04:40 +0200281 }
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200282
283 kmem_cache_free(dio_cache, dio);
Zach Brown6d544bb2006-12-10 02:20:54 -0800284 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285}
286
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200287static void dio_aio_complete_work(struct work_struct *work)
288{
289 struct dio *dio = container_of(work, struct dio, complete_work);
290
Christoph Hellwig716b9bc2016-04-07 08:51:59 -0700291 dio_complete(dio, 0, true);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200292}
293
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294static int dio_bio_complete(struct dio *dio, struct bio *bio);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200295
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296/*
297 * Asynchronous IO callback.
298 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200299static void dio_bio_end_aio(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300{
301 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800302 unsigned long remaining;
303 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305 /* cleanup the bio */
306 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800307
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800308 spin_lock_irqsave(&dio->bio_lock, flags);
309 remaining = --dio->refcount;
310 if (remaining == 1 && dio->waiter)
Zach Brown20258b22006-12-10 02:21:01 -0800311 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800312 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b22006-12-10 02:21:01 -0800313
Zach Brown8459d862006-12-10 02:21:05 -0800314 if (remaining == 0) {
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200315 if (dio->result && dio->defer_completion) {
316 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
317 queue_work(dio->inode->i_sb->s_dio_done_wq,
318 &dio->complete_work);
319 } else {
Christoph Hellwig716b9bc2016-04-07 08:51:59 -0700320 dio_complete(dio, 0, true);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200321 }
Zach Brown8459d862006-12-10 02:21:05 -0800322 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323}
324
325/*
326 * The BIO completion handler simply queues the BIO up for the process-context
327 * handler.
328 *
329 * During I/O bi_private points at the dio. After I/O, bi_private is used to
330 * implement a singly-linked list of completed BIOs, at dio->bio_list.
331 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200332static void dio_bio_end_io(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333{
334 struct dio *dio = bio->bi_private;
335 unsigned long flags;
336
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337 spin_lock_irqsave(&dio->bio_lock, flags);
338 bio->bi_private = dio->bio_list;
339 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800340 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341 wake_up_process(dio->waiter);
342 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343}
344
Josef Bacikfacd07b2010-05-23 11:00:55 -0400345/**
346 * dio_end_io - handle the end io action for the given bio
347 * @bio: The direct io bio thats being completed
348 * @error: Error if there was one
349 *
350 * This is meant to be called by any filesystem that uses their own dio_submit_t
351 * so that the DIO specific endio actions are dealt with after the filesystem
352 * has done it's completion work.
353 */
354void dio_end_io(struct bio *bio, int error)
355{
356 struct dio *dio = bio->bi_private;
357
358 if (dio->is_async)
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200359 dio_bio_end_aio(bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400360 else
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200361 dio_bio_end_io(bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400362}
363EXPORT_SYMBOL_GPL(dio_end_io);
364
Andi Kleenba253fb2011-08-01 21:38:08 -0700365static inline void
Andi Kleeneb28be22011-08-01 21:38:03 -0700366dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
367 struct block_device *bdev,
368 sector_t first_sector, int nr_vecs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369{
370 struct bio *bio;
371
David Dillow20d96002011-01-20 14:44:22 -0800372 /*
373 * bio_alloc() is guaranteed to return a bio when called with
Mel Gorman71baba42015-11-06 16:28:28 -0800374 * __GFP_RECLAIM and we request a valid number of vectors.
David Dillow20d96002011-01-20 14:44:22 -0800375 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377
378 bio->bi_bdev = bdev;
Kent Overstreet4f024f32013-10-11 15:44:27 -0700379 bio->bi_iter.bi_sector = first_sector;
Mike Christie8a4c1e42016-06-05 14:31:50 -0500380 bio_set_op_attrs(bio, dio->op, dio->op_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381 if (dio->is_async)
382 bio->bi_end_io = dio_bio_end_aio;
383 else
384 bio->bi_end_io = dio_bio_end_io;
385
Andi Kleeneb28be22011-08-01 21:38:03 -0700386 sdio->bio = bio;
387 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388}
389
390/*
391 * In the AIO read case we speculatively dirty the pages before starting IO.
392 * During IO completion, any of these pages which happen to have been written
393 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800394 *
395 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700397static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398{
Andi Kleeneb28be22011-08-01 21:38:03 -0700399 struct bio *bio = sdio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800400 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401
402 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800403
404 spin_lock_irqsave(&dio->bio_lock, flags);
405 dio->refcount++;
406 spin_unlock_irqrestore(&dio->bio_lock, flags);
407
Mike Christie8a4c1e42016-06-05 14:31:50 -0500408 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800410
Jens Axboec1c53462015-11-10 10:14:38 -0700411 dio->bio_bdev = bio->bi_bdev;
412
Jens Axboe15c4f632015-10-27 14:09:51 +0900413 if (sdio->submit_io) {
Mike Christie8a4c1e42016-06-05 14:31:50 -0500414 sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
Jens Axboe15c4f632015-10-27 14:09:51 +0900415 dio->bio_cookie = BLK_QC_T_NONE;
Jens Axboec1c53462015-11-10 10:14:38 -0700416 } else
Mike Christie4e49ea42016-06-05 14:31:41 -0500417 dio->bio_cookie = submit_bio(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418
Andi Kleeneb28be22011-08-01 21:38:03 -0700419 sdio->bio = NULL;
420 sdio->boundary = 0;
421 sdio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422}
423
424/*
425 * Release any resources in case of a failure
426 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700427static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428{
Al Viro7b2c99d2014-03-15 04:05:57 -0400429 while (sdio->head < sdio->tail)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300430 put_page(dio->pages[sdio->head++]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431}
432
433/*
Zach Brown02732012006-12-10 02:20:59 -0800434 * Wait for the next BIO to complete. Remove it and return it. NULL is
435 * returned once all BIOs have been completed. This must only be called once
436 * all bios have been issued so that dio->refcount can only decrease. This
437 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438 */
439static struct bio *dio_await_one(struct dio *dio)
440{
441 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800442 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443
444 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800445
446 /*
447 * Wait as long as the list is empty and there are bios in flight. bio
448 * completion drops the count, maybe adds to the list, and wakes while
449 * holding the bio_lock so we don't need set_current_state()'s barrier
450 * and can call it after testing our condition.
451 */
452 while (dio->refcount > 1 && dio->bio_list == NULL) {
453 __set_current_state(TASK_UNINTERRUPTIBLE);
454 dio->waiter = current;
455 spin_unlock_irqrestore(&dio->bio_lock, flags);
Christoph Hellwigc43c83a2016-03-03 16:04:02 +0100456 if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
457 !blk_poll(bdev_get_queue(dio->bio_bdev), dio->bio_cookie))
Jens Axboe15c4f632015-10-27 14:09:51 +0900458 io_schedule();
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800459 /* wake up sets us TASK_RUNNING */
460 spin_lock_irqsave(&dio->bio_lock, flags);
461 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 }
Zach Brown02732012006-12-10 02:20:59 -0800463 if (dio->bio_list) {
464 bio = dio->bio_list;
465 dio->bio_list = bio->bi_private;
466 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 spin_unlock_irqrestore(&dio->bio_lock, flags);
468 return bio;
469}
470
471/*
472 * Process one completed BIO. No locks are held.
473 */
474static int dio_bio_complete(struct dio *dio, struct bio *bio)
475{
Kent Overstreetcb34e052012-09-05 15:22:02 -0700476 struct bio_vec *bvec;
477 unsigned i;
Sasha Levin9b81c842015-08-10 19:05:18 -0400478 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200480 if (bio->bi_error)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800481 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482
Mike Christie8a4c1e42016-06-05 14:31:50 -0500483 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) {
Sasha Levin9b81c842015-08-10 19:05:18 -0400484 err = bio->bi_error;
Mike Krinkin7ddc9712016-01-30 19:09:59 +0300485 bio_check_pages_dirty(bio); /* transfers ownership */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486 } else {
Kent Overstreetcb34e052012-09-05 15:22:02 -0700487 bio_for_each_segment_all(bvec, bio, i) {
488 struct page *page = bvec->bv_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489
Mike Christie8a4c1e42016-06-05 14:31:50 -0500490 if (dio->op == REQ_OP_READ && !PageCompound(page) &&
Ming Lei53cbf3b2015-08-17 10:31:46 +0800491 dio->should_dirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 set_page_dirty_lock(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300493 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 }
Sasha Levin9b81c842015-08-10 19:05:18 -0400495 err = bio->bi_error;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496 bio_put(bio);
497 }
Sasha Levin9b81c842015-08-10 19:05:18 -0400498 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499}
500
501/*
Zach Brown02732012006-12-10 02:20:59 -0800502 * Wait on and process all in-flight BIOs. This must only be called once
503 * all bios have been issued so that the refcount can only decrease.
504 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200505 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800506 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800508static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509{
Zach Brown02732012006-12-10 02:20:59 -0800510 struct bio *bio;
511 do {
512 bio = dio_await_one(dio);
513 if (bio)
514 dio_bio_complete(dio, bio);
515 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516}
517
518/*
519 * A really large O_DIRECT read or write can generate a lot of BIOs. So
520 * to keep the memory consumption sane we periodically reap any completed BIOs
521 * during the BIO generation phase.
522 *
523 * This also helps to limit the peak amount of pinned userspace memory.
524 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700525static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526{
527 int ret = 0;
528
Andi Kleeneb28be22011-08-01 21:38:03 -0700529 if (sdio->reap_counter++ >= 64) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530 while (dio->bio_list) {
531 unsigned long flags;
532 struct bio *bio;
533 int ret2;
534
535 spin_lock_irqsave(&dio->bio_lock, flags);
536 bio = dio->bio_list;
537 dio->bio_list = bio->bi_private;
538 spin_unlock_irqrestore(&dio->bio_lock, flags);
539 ret2 = dio_bio_complete(dio, bio);
540 if (ret == 0)
541 ret = ret2;
542 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700543 sdio->reap_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544 }
545 return ret;
546}
547
548/*
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200549 * Create workqueue for deferred direct IO completions. We allocate the
550 * workqueue when it's first needed. This avoids creating workqueue for
551 * filesystems that don't need it and also allows us to create the workqueue
552 * late enough so the we can include s_id in the name of the workqueue.
553 */
554static int sb_init_dio_done_wq(struct super_block *sb)
555{
Olof Johansson45150c42013-09-09 10:34:23 -0700556 struct workqueue_struct *old;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200557 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
558 WQ_MEM_RECLAIM, 0,
559 sb->s_id);
560 if (!wq)
561 return -ENOMEM;
562 /*
563 * This has to be atomic as more DIOs can race to create the workqueue
564 */
Olof Johansson45150c42013-09-09 10:34:23 -0700565 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200566 /* Someone created workqueue before us? Free ours... */
Olof Johansson45150c42013-09-09 10:34:23 -0700567 if (old)
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200568 destroy_workqueue(wq);
569 return 0;
570}
571
572static int dio_set_defer_completion(struct dio *dio)
573{
574 struct super_block *sb = dio->inode->i_sb;
575
576 if (dio->defer_completion)
577 return 0;
578 dio->defer_completion = true;
579 if (!sb->s_dio_done_wq)
580 return sb_init_dio_done_wq(sb);
581 return 0;
582}
583
584/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 * Call into the fs to map some more disk blocks. We record the current number
Andi Kleeneb28be22011-08-01 21:38:03 -0700586 * of available blocks at sdio->blocks_available. These are in units of the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 * fs blocksize, (1 << inode->i_blkbits).
588 *
589 * The fs is allowed to map lots of blocks at once. If it wants to do that,
590 * it uses the passed inode-relative block number as the file offset, as usual.
591 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800592 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593 * has remaining to do. The fs should not map more than this number of blocks.
594 *
595 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
596 * indicate how much contiguous disk space has been made available at
597 * bh->b_blocknr.
598 *
599 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
600 * This isn't very efficient...
601 *
602 * In the case of filesystem holes: the fs may return an arbitrarily-large
603 * hole by returning an appropriate value in b_size and by clearing
604 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800605 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 */
Andi Kleen18772642011-08-01 21:38:07 -0700607static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
608 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609{
610 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
Tao Maae55e1a2012-01-12 17:20:33 -0800612 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613 unsigned long fs_count; /* Number of filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 int create;
Linus Torvaldsab738572012-11-29 12:27:00 -0800615 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616
617 /*
618 * If there was a memory error and we've overwritten all the
619 * mapped blocks then we can now return that memory error
620 */
621 ret = dio->page_errors;
622 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700623 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
624 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
Tao Maae55e1a2012-01-12 17:20:33 -0800625 fs_endblk = (sdio->final_block_in_request - 1) >>
626 sdio->blkfactor;
627 fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628
Nathan Scott3c674e72006-03-29 09:26:15 +1000629 map_bh->b_state = 0;
Linus Torvaldsab738572012-11-29 12:27:00 -0800630 map_bh->b_size = fs_count << i_blkbits;
Nathan Scott3c674e72006-03-29 09:26:15 +1000631
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800632 /*
Eryu Guan9ecd10b2016-05-27 14:27:18 -0700633 * For writes that could fill holes inside i_size on a
634 * DIO_SKIP_HOLES filesystem we forbid block creations: only
635 * overwrites are permitted. We will return early to the caller
636 * once we see an unmapped buffer head returned, and the caller
637 * will fall back to buffered I/O.
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800638 *
639 * Otherwise the decision is left to the get_blocks method,
640 * which may decide to handle it or also return an unmapped
641 * buffer head.
642 */
Mike Christie8a4c1e42016-06-05 14:31:50 -0500643 create = dio->op == REQ_OP_WRITE;
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800644 if (dio->flags & DIO_SKIP_HOLES) {
Eryu Guan9ecd10b2016-05-27 14:27:18 -0700645 if (fs_startblk <= ((i_size_read(dio->inode) - 1) >>
646 i_blkbits))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000649
Andi Kleeneb28be22011-08-01 21:38:03 -0700650 ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 map_bh, create);
Andi Kleen18772642011-08-01 21:38:07 -0700652
653 /* Store for completion */
654 dio->private = map_bh->b_private;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200655
656 if (ret == 0 && buffer_defer_completion(map_bh))
657 ret = dio_set_defer_completion(dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 }
659 return ret;
660}
661
662/*
663 * There is no bio. Make one now.
664 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700665static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
666 sector_t start_sector, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667{
668 sector_t sector;
669 int ret, nr_pages;
670
Andi Kleeneb28be22011-08-01 21:38:03 -0700671 ret = dio_bio_reap(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 if (ret)
673 goto out;
Andi Kleeneb28be22011-08-01 21:38:03 -0700674 sector = start_sector << (sdio->blkbits - 9);
Kent Overstreetb54ffb72015-05-19 14:31:01 +0200675 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 BUG_ON(nr_pages <= 0);
Andi Kleen18772642011-08-01 21:38:07 -0700677 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
Andi Kleeneb28be22011-08-01 21:38:03 -0700678 sdio->boundary = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679out:
680 return ret;
681}
682
683/*
684 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
685 * that was successful then update final_block_in_bio and take a ref against
686 * the just-added page.
687 *
688 * Return zero on success. Non-zero means the caller needs to start a new BIO.
689 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700690static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691{
692 int ret;
693
Andi Kleeneb28be22011-08-01 21:38:03 -0700694 ret = bio_add_page(sdio->bio, sdio->cur_page,
695 sdio->cur_page_len, sdio->cur_page_offset);
696 if (ret == sdio->cur_page_len) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 /*
698 * Decrement count only, if we are done with this page
699 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700700 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
701 sdio->pages_in_io--;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300702 get_page(sdio->cur_page);
Andi Kleeneb28be22011-08-01 21:38:03 -0700703 sdio->final_block_in_bio = sdio->cur_page_block +
704 (sdio->cur_page_len >> sdio->blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 ret = 0;
706 } else {
707 ret = 1;
708 }
709 return ret;
710}
711
712/*
713 * Put cur_page under IO. The section of cur_page which is described by
714 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
715 * starts on-disk at cur_page_block.
716 *
717 * We take a ref against the page here (on behalf of its presence in the bio).
718 *
719 * The caller of this function is responsible for removing cur_page from the
720 * dio, and for dropping the refcount which came from that presence.
721 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700722static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
723 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724{
725 int ret = 0;
726
Andi Kleeneb28be22011-08-01 21:38:03 -0700727 if (sdio->bio) {
728 loff_t cur_offset = sdio->cur_page_fs_offset;
729 loff_t bio_next_offset = sdio->logical_offset_in_bio +
Kent Overstreet4f024f32013-10-11 15:44:27 -0700730 sdio->bio->bi_iter.bi_size;
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400731
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400733 * See whether this new request is contiguous with the old.
734 *
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900735 * Btrfs cannot handle having logically non-contiguous requests
736 * submitted. For example if you have
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400737 *
738 * Logical: [0-4095][HOLE][8192-12287]
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900739 * Physical: [0-4095] [4096-8191]
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400740 *
741 * We cannot submit those pages together as one BIO. So if our
742 * current logical offset in the file does not equal what would
743 * be the next logical offset in the bio, submit the bio we
744 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700746 if (sdio->final_block_in_bio != sdio->cur_page_block ||
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400747 cur_offset != bio_next_offset)
Andi Kleeneb28be22011-08-01 21:38:03 -0700748 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749 }
750
Andi Kleeneb28be22011-08-01 21:38:03 -0700751 if (sdio->bio == NULL) {
Andi Kleen18772642011-08-01 21:38:07 -0700752 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 if (ret)
754 goto out;
755 }
756
Andi Kleeneb28be22011-08-01 21:38:03 -0700757 if (dio_bio_add_page(sdio) != 0) {
758 dio_bio_submit(dio, sdio);
Andi Kleen18772642011-08-01 21:38:07 -0700759 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700761 ret = dio_bio_add_page(sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762 BUG_ON(ret != 0);
763 }
764 }
765out:
766 return ret;
767}
768
769/*
770 * An autonomous function to put a chunk of a page under deferred IO.
771 *
772 * The caller doesn't actually know (or care) whether this piece of page is in
773 * a BIO, or is under IO or whatever. We just take care of all possible
774 * situations here. The separation between the logic of do_direct_IO() and
775 * that of submit_page_section() is important for clarity. Please don't break.
776 *
777 * The chunk of page starts on-disk at blocknr.
778 *
779 * We perform deferred IO, by recording the last-submitted page inside our
780 * private part of the dio structure. If possible, we just expand the IO
781 * across that page here.
782 *
783 * If that doesn't work out then we put the old page into the bio and add this
784 * page to the dio instead.
785 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700786static inline int
Andi Kleeneb28be22011-08-01 21:38:03 -0700787submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
Andi Kleen18772642011-08-01 21:38:07 -0700788 unsigned offset, unsigned len, sector_t blocknr,
789 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790{
791 int ret = 0;
792
Mike Christie8a4c1e42016-06-05 14:31:50 -0500793 if (dio->op == REQ_OP_WRITE) {
Andrew Morton98c4d572006-12-10 02:19:47 -0800794 /*
795 * Read accounting is performed in submit_bio()
796 */
797 task_io_account_write(len);
798 }
799
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 /*
801 * Can we just grow the current page's presence in the dio?
802 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700803 if (sdio->cur_page == page &&
804 sdio->cur_page_offset + sdio->cur_page_len == offset &&
805 sdio->cur_page_block +
806 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
807 sdio->cur_page_len += len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 goto out;
809 }
810
811 /*
812 * If there's a deferred page already there then send it.
813 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700814 if (sdio->cur_page) {
Andi Kleen18772642011-08-01 21:38:07 -0700815 ret = dio_send_cur_page(dio, sdio, map_bh);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300816 put_page(sdio->cur_page);
Andi Kleeneb28be22011-08-01 21:38:03 -0700817 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818 if (ret)
Jan Karab1058b92013-04-29 15:06:18 -0700819 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 }
821
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300822 get_page(page); /* It is in dio */
Andi Kleeneb28be22011-08-01 21:38:03 -0700823 sdio->cur_page = page;
824 sdio->cur_page_offset = offset;
825 sdio->cur_page_len = len;
826 sdio->cur_page_block = blocknr;
827 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828out:
Jan Karab1058b92013-04-29 15:06:18 -0700829 /*
830 * If sdio->boundary then we want to schedule the IO now to
831 * avoid metadata seeks.
832 */
833 if (sdio->boundary) {
834 ret = dio_send_cur_page(dio, sdio, map_bh);
835 dio_bio_submit(dio, sdio);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300836 put_page(sdio->cur_page);
Jan Karab1058b92013-04-29 15:06:18 -0700837 sdio->cur_page = NULL;
838 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 return ret;
840}
841
842/*
843 * Clean any dirty buffers in the blockdev mapping which alias newly-created
844 * file blocks. Only called for S_ISREG files - blockdevs do not set
845 * buffer_new
846 */
Andi Kleen18772642011-08-01 21:38:07 -0700847static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848{
849 unsigned i;
850 unsigned nblocks;
851
Andi Kleen18772642011-08-01 21:38:07 -0700852 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853
854 for (i = 0; i < nblocks; i++) {
Andi Kleen18772642011-08-01 21:38:07 -0700855 unmap_underlying_metadata(map_bh->b_bdev,
856 map_bh->b_blocknr + i);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857 }
858}
859
860/*
861 * If we are not writing the entire block and get_block() allocated
862 * the block for us, we need to fill-in the unused portion of the
863 * block with zeros. This happens only if user-buffer, fileoffset or
864 * io length is not filesystem block-size multiple.
865 *
866 * `end' is zero if we're doing the start of the IO, 1 at the end of the
867 * IO.
868 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700869static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
870 int end, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871{
872 unsigned dio_blocks_per_fs_block;
873 unsigned this_chunk_blocks; /* In dio_blocks */
874 unsigned this_chunk_bytes;
875 struct page *page;
876
Andi Kleeneb28be22011-08-01 21:38:03 -0700877 sdio->start_zero_done = 1;
Andi Kleen18772642011-08-01 21:38:07 -0700878 if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 return;
880
Andi Kleeneb28be22011-08-01 21:38:03 -0700881 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
882 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883
884 if (!this_chunk_blocks)
885 return;
886
887 /*
888 * We need to zero out part of an fs block. It is either at the
889 * beginning or the end of the fs block.
890 */
891 if (end)
892 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
893
Andi Kleeneb28be22011-08-01 21:38:03 -0700894 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895
Nick Piggin557ed1f2007-10-16 01:24:40 -0700896 page = ZERO_PAGE(0);
Andi Kleeneb28be22011-08-01 21:38:03 -0700897 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700898 sdio->next_block_for_io, map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 return;
900
Andi Kleeneb28be22011-08-01 21:38:03 -0700901 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902}
903
904/*
905 * Walk the user pages, and the file, mapping blocks to disk and generating
906 * a sequence of (page,offset,len,block) mappings. These mappings are injected
907 * into submit_page_section(), which takes care of the next stage of submission
908 *
909 * Direct IO against a blockdev is different from a file. Because we can
910 * happily perform page-sized but 512-byte aligned IOs. It is important that
911 * blockdev IO be able to have fine alignment and large sizes.
912 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800913 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 * with the size of IO which is permitted at this offset and this i_blkbits.
915 *
916 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800917 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 * fine alignment but still allows this function to work in PAGE_SIZE units.
919 */
Andi Kleen18772642011-08-01 21:38:07 -0700920static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
921 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922{
Andi Kleeneb28be22011-08-01 21:38:03 -0700923 const unsigned blkbits = sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924 int ret = 0;
925
Andi Kleeneb28be22011-08-01 21:38:03 -0700926 while (sdio->block_in_file < sdio->final_block_in_request) {
Al Viro7b2c99d2014-03-15 04:05:57 -0400927 struct page *page;
928 size_t from, to;
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300929
930 page = dio_get_page(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 if (IS_ERR(page)) {
932 ret = PTR_ERR(page);
933 goto out;
934 }
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300935 from = sdio->head ? 0 : sdio->from;
936 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
937 sdio->head++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938
Al Viro7b2c99d2014-03-15 04:05:57 -0400939 while (from < to) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940 unsigned this_chunk_bytes; /* # of bytes mapped */
941 unsigned this_chunk_blocks; /* # of blocks */
942 unsigned u;
943
Andi Kleeneb28be22011-08-01 21:38:03 -0700944 if (sdio->blocks_available == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 /*
946 * Need to go and map some more disk
947 */
948 unsigned long blkmask;
949 unsigned long dio_remainder;
950
Andi Kleen18772642011-08-01 21:38:07 -0700951 ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952 if (ret) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300953 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 goto out;
955 }
956 if (!buffer_mapped(map_bh))
957 goto do_holes;
958
Andi Kleeneb28be22011-08-01 21:38:03 -0700959 sdio->blocks_available =
960 map_bh->b_size >> sdio->blkbits;
961 sdio->next_block_for_io =
962 map_bh->b_blocknr << sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 if (buffer_new(map_bh))
Andi Kleen18772642011-08-01 21:38:07 -0700964 clean_blockdev_aliases(dio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965
Andi Kleeneb28be22011-08-01 21:38:03 -0700966 if (!sdio->blkfactor)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967 goto do_holes;
968
Andi Kleeneb28be22011-08-01 21:38:03 -0700969 blkmask = (1 << sdio->blkfactor) - 1;
970 dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971
972 /*
973 * If we are at the start of IO and that IO
974 * starts partway into a fs-block,
975 * dio_remainder will be non-zero. If the IO
976 * is a read then we can simply advance the IO
977 * cursor to the first block which is to be
978 * read. But if the IO is a write and the
979 * block was newly allocated we cannot do that;
980 * the start of the fs block must be zeroed out
981 * on-disk
982 */
983 if (!buffer_new(map_bh))
Andi Kleeneb28be22011-08-01 21:38:03 -0700984 sdio->next_block_for_io += dio_remainder;
985 sdio->blocks_available -= dio_remainder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 }
987do_holes:
988 /* Handle holes */
989 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800990 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991
992 /* AKPM: eargh, -ENOTBLK is a hack */
Mike Christie8a4c1e42016-06-05 14:31:50 -0500993 if (dio->op == REQ_OP_WRITE) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300994 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995 return -ENOTBLK;
996 }
997
Jeff Moyer35dc8162006-02-03 03:04:27 -0800998 /*
999 * Be sure to account for a partial block as the
1000 * last block in the file
1001 */
1002 i_size_aligned = ALIGN(i_size_read(dio->inode),
1003 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -07001004 if (sdio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -08001005 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006 /* We hit eof */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001007 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008 goto out;
1009 }
Al Viro7b2c99d2014-03-15 04:05:57 -04001010 zero_user(page, from, 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -07001011 sdio->block_in_file++;
Al Viro7b2c99d2014-03-15 04:05:57 -04001012 from += 1 << blkbits;
Al Viro3320c602014-03-10 02:30:55 -04001013 dio->result += 1 << blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014 goto next_block;
1015 }
1016
1017 /*
1018 * If we're performing IO which has an alignment which
1019 * is finer than the underlying fs, go check to see if
1020 * we must zero out the start of this block.
1021 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001022 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
Andi Kleen18772642011-08-01 21:38:07 -07001023 dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001024
1025 /*
1026 * Work out, in this_chunk_blocks, how much disk we
1027 * can add to this page
1028 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001029 this_chunk_blocks = sdio->blocks_available;
Al Viro7b2c99d2014-03-15 04:05:57 -04001030 u = (to - from) >> blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 if (this_chunk_blocks > u)
1032 this_chunk_blocks = u;
Andi Kleeneb28be22011-08-01 21:38:03 -07001033 u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034 if (this_chunk_blocks > u)
1035 this_chunk_blocks = u;
1036 this_chunk_bytes = this_chunk_blocks << blkbits;
1037 BUG_ON(this_chunk_bytes == 0);
1038
Jan Kara092c8d42013-04-29 15:06:17 -07001039 if (this_chunk_blocks == sdio->blocks_available)
1040 sdio->boundary = buffer_boundary(map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -07001041 ret = submit_page_section(dio, sdio, page,
Al Viro7b2c99d2014-03-15 04:05:57 -04001042 from,
Andi Kleeneb28be22011-08-01 21:38:03 -07001043 this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -07001044 sdio->next_block_for_io,
1045 map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046 if (ret) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001047 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 goto out;
1049 }
Andi Kleeneb28be22011-08-01 21:38:03 -07001050 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051
Andi Kleeneb28be22011-08-01 21:38:03 -07001052 sdio->block_in_file += this_chunk_blocks;
Al Viro7b2c99d2014-03-15 04:05:57 -04001053 from += this_chunk_bytes;
1054 dio->result += this_chunk_bytes;
Andi Kleeneb28be22011-08-01 21:38:03 -07001055 sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056next_block:
Andi Kleeneb28be22011-08-01 21:38:03 -07001057 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1058 if (sdio->block_in_file == sdio->final_block_in_request)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059 break;
1060 }
1061
1062 /* Drop the ref which was taken in get_user_pages() */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001063 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064 }
1065out:
1066 return ret;
1067}
1068
Andi Kleen847cc632011-08-01 21:38:09 -07001069static inline int drop_refcount(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070{
Andi Kleen847cc632011-08-01 21:38:09 -07001071 int ret2;
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001072 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073
Zach Brown8459d862006-12-10 02:21:05 -08001074 /*
1075 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001076 * operation. AIO can if it was a broken operation described above or
1077 * in fact if all the bios race to complete before we get here. In
1078 * that case dio_complete() translates the EIOCBQUEUED into the proper
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001079 * return code that the caller will hand to ->complete().
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001080 *
1081 * This is managed by the bio_lock instead of being an atomic_t so that
1082 * completion paths can drop their ref and use the remaining count to
1083 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001084 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001085 spin_lock_irqsave(&dio->bio_lock, flags);
1086 ret2 = --dio->refcount;
1087 spin_unlock_irqrestore(&dio->bio_lock, flags);
Andi Kleen847cc632011-08-01 21:38:09 -07001088 return ret2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089}
1090
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001091/*
1092 * This is a library function for use by filesystem drivers.
1093 *
1094 * The locking rules are governed by the flags parameter:
1095 * - if the flags value contains DIO_LOCKING we use a fancy locking
1096 * scheme for dumb filesystems.
1097 * For writes this function is called under i_mutex and returns with
1098 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1099 * taken and dropped again before returning.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001100 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1101 * internal locking but rather rely on the filesystem to synchronize
1102 * direct I/O reads/writes versus each other and truncate.
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001103 *
1104 * To help with locking against truncate we incremented the i_dio_count
1105 * counter before starting direct I/O, and decrement it once we are done.
1106 * Truncate can wait for it to reach zero to provide exclusion. It is
1107 * expected that filesystem provide exclusion between new direct I/O
1108 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1109 * but other filesystems need to take care of this on their own.
Andi Kleenba253fb2011-08-01 21:38:08 -07001110 *
1111 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1112 * is always inlined. Otherwise gcc is unable to split the structure into
1113 * individual fields and will generate much worse code. This is important
1114 * for the whole file.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001115 */
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001116static inline ssize_t
Omar Sandoval17f8c842015-03-16 04:33:50 -07001117do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1118 struct block_device *bdev, struct iov_iter *iter,
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001119 get_block_t get_block, dio_iodone_t end_io,
Omar Sandoval17f8c842015-03-16 04:33:50 -07001120 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121{
Linus Torvaldsab738572012-11-29 12:27:00 -08001122 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1123 unsigned blkbits = i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 unsigned blocksize_mask = (1 << blkbits) - 1;
1125 ssize_t retval = -EINVAL;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001126 size_t count = iov_iter_count(iter);
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001127 loff_t offset = iocb->ki_pos;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001128 loff_t end = offset + count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 struct dio *dio;
Andi Kleeneb28be22011-08-01 21:38:03 -07001130 struct dio_submit sdio = { 0, };
Andi Kleen847cc632011-08-01 21:38:09 -07001131 struct buffer_head map_bh = { 0, };
Fengguang Wu647d1e42012-08-09 15:23:09 +02001132 struct blk_plug plug;
Al Viro886a3912014-03-05 13:50:45 -05001133 unsigned long align = offset | iov_iter_alignment(iter);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001135 /*
1136 * Avoid references to bdev if not absolutely needed to give
1137 * the early prefetch in the caller enough time.
1138 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139
Al Viro886a3912014-03-05 13:50:45 -05001140 if (align & blocksize_mask) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001142 blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 blocksize_mask = (1 << blkbits) - 1;
Al Viro886a3912014-03-05 13:50:45 -05001144 if (align & blocksize_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145 goto out;
1146 }
1147
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001148 /* watch out for a 0 len io from a tricksy fs */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001149 if (iov_iter_rw(iter) == READ && !iov_iter_count(iter))
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001150 return 0;
1151
Andi Kleen6e8267f2011-08-01 21:38:06 -07001152 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153 retval = -ENOMEM;
1154 if (!dio)
1155 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001156 /*
1157 * Believe it or not, zeroing out the page array caused a .5%
1158 * performance regression in a database benchmark. So, we take
1159 * care to only zero out what's needed.
1160 */
1161 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001163 dio->flags = flags;
1164 if (dio->flags & DIO_LOCKING) {
Omar Sandoval17f8c842015-03-16 04:33:50 -07001165 if (iov_iter_rw(iter) == READ) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001166 struct address_space *mapping =
1167 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001169 /* will be released by direct_io_worker */
Al Viro59551022016-01-22 15:40:57 -05001170 inode_lock(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171
1172 retval = filemap_write_and_wait_range(mapping, offset,
1173 end - 1);
1174 if (retval) {
Al Viro59551022016-01-22 15:40:57 -05001175 inode_unlock(inode);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001176 kmem_cache_free(dio_cache, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177 goto out;
1178 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 }
1181
Jan Kara74cedf92015-11-30 10:15:42 -07001182 /* Once we sampled i_size check for reads beyond EOF */
1183 dio->i_size = i_size_read(inode);
1184 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1185 if (dio->flags & DIO_LOCKING)
Al Viro59551022016-01-22 15:40:57 -05001186 inode_unlock(inode);
Jan Kara74cedf92015-11-30 10:15:42 -07001187 kmem_cache_free(dio_cache, dio);
Al Viro2d4594a2015-12-08 12:22:47 -05001188 retval = 0;
Jan Kara74cedf92015-11-30 10:15:42 -07001189 goto out;
1190 }
1191
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 /*
Christoph Hellwig60392572014-02-10 10:27:11 +11001193 * For file extending writes updating i_size before data writeouts
1194 * complete can expose uninitialized blocks in dumb filesystems.
1195 * In that case we need to wait for I/O completion even if asked
1196 * for an asynchronous write.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197 */
Christoph Hellwig60392572014-02-10 10:27:11 +11001198 if (is_sync_kiocb(iocb))
1199 dio->is_async = false;
1200 else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
Omar Sandoval17f8c842015-03-16 04:33:50 -07001201 iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
Christoph Hellwig60392572014-02-10 10:27:11 +11001202 dio->is_async = false;
1203 else
1204 dio->is_async = true;
1205
Andi Kleen847cc632011-08-01 21:38:09 -07001206 dio->inode = inode;
Mike Christie8a4c1e42016-06-05 14:31:50 -05001207 if (iov_iter_rw(iter) == WRITE) {
1208 dio->op = REQ_OP_WRITE;
1209 dio->op_flags = WRITE_ODIRECT;
1210 } else {
1211 dio->op = REQ_OP_READ;
1212 }
Christoph Hellwig02afc272013-09-04 15:04:40 +02001213
1214 /*
1215 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1216 * so that we can call ->fsync.
1217 */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001218 if (dio->is_async && iov_iter_rw(iter) == WRITE &&
Christoph Hellwig02afc272013-09-04 15:04:40 +02001219 ((iocb->ki_filp->f_flags & O_DSYNC) ||
1220 IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1221 retval = dio_set_defer_completion(dio);
1222 if (retval) {
1223 /*
1224 * We grab i_mutex only for reads so we don't have
1225 * to release it here
1226 */
1227 kmem_cache_free(dio_cache, dio);
1228 goto out;
1229 }
1230 }
1231
1232 /*
1233 * Will be decremented at I/O completion time.
1234 */
Jens Axboefe0f07d2015-04-15 17:05:48 -06001235 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
1236 inode_dio_begin(inode);
Christoph Hellwig02afc272013-09-04 15:04:40 +02001237
1238 retval = 0;
Andi Kleen847cc632011-08-01 21:38:09 -07001239 sdio.blkbits = blkbits;
Linus Torvaldsab738572012-11-29 12:27:00 -08001240 sdio.blkfactor = i_blkbits - blkbits;
Andi Kleen847cc632011-08-01 21:38:09 -07001241 sdio.block_in_file = offset >> blkbits;
1242
1243 sdio.get_block = get_block;
1244 dio->end_io = end_io;
1245 sdio.submit_io = submit_io;
1246 sdio.final_block_in_bio = -1;
1247 sdio.next_block_for_io = -1;
1248
1249 dio->iocb = iocb;
Andi Kleen847cc632011-08-01 21:38:09 -07001250
1251 spin_lock_init(&dio->bio_lock);
1252 dio->refcount = 1;
1253
Ming Lei53cbf3b2015-08-17 10:31:46 +08001254 dio->should_dirty = (iter->type == ITER_IOVEC);
Al Viro7b2c99d2014-03-15 04:05:57 -04001255 sdio.iter = iter;
1256 sdio.final_block_in_request =
1257 (offset + iov_iter_count(iter)) >> blkbits;
1258
Andi Kleen847cc632011-08-01 21:38:09 -07001259 /*
1260 * In case of non-aligned buffers, we may need 2 more
1261 * pages since we need to zero out first and last block.
1262 */
1263 if (unlikely(sdio.blkfactor))
1264 sdio.pages_in_io = 2;
1265
Al Virof67da302014-03-19 01:16:16 -04001266 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
Andi Kleen847cc632011-08-01 21:38:09 -07001267
Fengguang Wu647d1e42012-08-09 15:23:09 +02001268 blk_start_plug(&plug);
1269
Al Viro7b2c99d2014-03-15 04:05:57 -04001270 retval = do_direct_IO(dio, &sdio, &map_bh);
1271 if (retval)
1272 dio_cleanup(dio, &sdio);
Andi Kleen847cc632011-08-01 21:38:09 -07001273
1274 if (retval == -ENOTBLK) {
1275 /*
1276 * The remaining part of the request will be
1277 * be handled by buffered I/O when we return
1278 */
1279 retval = 0;
1280 }
1281 /*
1282 * There may be some unwritten disk at the end of a part-written
1283 * fs-block-sized block. Go zero that now.
1284 */
1285 dio_zero_block(dio, &sdio, 1, &map_bh);
1286
1287 if (sdio.cur_page) {
1288 ssize_t ret2;
1289
1290 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1291 if (retval == 0)
1292 retval = ret2;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001293 put_page(sdio.cur_page);
Andi Kleen847cc632011-08-01 21:38:09 -07001294 sdio.cur_page = NULL;
1295 }
1296 if (sdio.bio)
1297 dio_bio_submit(dio, &sdio);
1298
Fengguang Wu647d1e42012-08-09 15:23:09 +02001299 blk_finish_plug(&plug);
1300
Andi Kleen847cc632011-08-01 21:38:09 -07001301 /*
1302 * It is possible that, we return short IO due to end of file.
1303 * In that case, we need to release all the pages we got hold on.
1304 */
1305 dio_cleanup(dio, &sdio);
1306
1307 /*
1308 * All block lookups have been performed. For READ requests
1309 * we can let i_mutex go now that its achieved its purpose
1310 * of protecting us from looking up uninitialized blocks.
1311 */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001312 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
Al Viro59551022016-01-22 15:40:57 -05001313 inode_unlock(dio->inode);
Andi Kleen847cc632011-08-01 21:38:09 -07001314
1315 /*
1316 * The only time we want to leave bios in flight is when a successful
1317 * partial aio read or full aio write have been setup. In that case
1318 * bio completion will call aio_complete. The only time it's safe to
1319 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1320 * This had *better* be the only place that raises -EIOCBQUEUED.
1321 */
1322 BUG_ON(retval == -EIOCBQUEUED);
1323 if (dio->is_async && retval == 0 && dio->result &&
Omar Sandoval17f8c842015-03-16 04:33:50 -07001324 (iov_iter_rw(iter) == READ || dio->result == count))
Andi Kleen847cc632011-08-01 21:38:09 -07001325 retval = -EIOCBQUEUED;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001326 else
Andi Kleen847cc632011-08-01 21:38:09 -07001327 dio_await_completion(dio);
1328
1329 if (drop_refcount(dio) == 0) {
Christoph Hellwig716b9bc2016-04-07 08:51:59 -07001330 retval = dio_complete(dio, retval, false);
Andi Kleen847cc632011-08-01 21:38:09 -07001331 } else
1332 BUG_ON(retval != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001334out:
1335 return retval;
1336}
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001337
Omar Sandoval17f8c842015-03-16 04:33:50 -07001338ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1339 struct block_device *bdev, struct iov_iter *iter,
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001340 get_block_t get_block,
Omar Sandoval17f8c842015-03-16 04:33:50 -07001341 dio_iodone_t end_io, dio_submit_t submit_io,
1342 int flags)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001343{
1344 /*
1345 * The block device state is needed in the end to finally
1346 * submit everything. Since it's likely to be cache cold
1347 * prefetch it here as first thing to hide some of the
1348 * latency.
1349 *
1350 * Attempt to prefetch the pieces we likely need later.
1351 */
1352 prefetch(&bdev->bd_disk->part_tbl);
1353 prefetch(bdev->bd_queue);
1354 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1355
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001356 return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
Omar Sandoval17f8c842015-03-16 04:33:50 -07001357 end_io, submit_io, flags);
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001358}
1359
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360EXPORT_SYMBOL(__blockdev_direct_IO);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001361
1362static __init int dio_init(void)
1363{
1364 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1365 return 0;
1366}
1367module_init(dio_init)