blob: 7ab90f5081eebc4ab8b0de88bef8d0b6310ed113 [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>
Kent Overstreeta27bb332013-05-07 16:19:08 -070040#include <linux/aio.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070041
42/*
43 * How many user pages to map in one call to get_user_pages(). This determines
Andi Kleencde1ecb2011-08-01 21:38:04 -070044 * the size of a structure in the slab cache
Linus Torvalds1da177e2005-04-16 15:20:36 -070045 */
46#define DIO_PAGES 64
47
48/*
49 * This code generally works in units of "dio_blocks". A dio_block is
50 * somewhere between the hard sector size and the filesystem block size. it
51 * is determined on a per-invocation basis. When talking to the filesystem
52 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
53 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
54 * to bio_block quantities by shifting left by blkfactor.
55 *
56 * If blkfactor is zero then the user's request was aligned to the filesystem's
57 * blocksize.
Linus Torvalds1da177e2005-04-16 15:20:36 -070058 */
59
Andi Kleeneb28be22011-08-01 21:38:03 -070060/* dio_state only used in the submission path */
61
62struct dio_submit {
Linus Torvalds1da177e2005-04-16 15:20:36 -070063 struct bio *bio; /* bio under assembly */
Linus Torvalds1da177e2005-04-16 15:20:36 -070064 unsigned blkbits; /* doesn't change */
65 unsigned blkfactor; /* When we're using an alignment which
66 is finer than the filesystem's soft
67 blocksize, this specifies how much
68 finer. blkfactor=2 means 1/4-block
69 alignment. Does not change */
70 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
71 been performed at the start of a
72 write */
73 int pages_in_io; /* approximate total IO pages */
74 size_t size; /* total request size (doesn't change)*/
75 sector_t block_in_file; /* Current offset into the underlying
76 file in dio_block units. */
77 unsigned blocks_available; /* At block_in_file. changes */
Andi Kleen0dc2bc42011-08-01 21:38:05 -070078 int reap_counter; /* rate limit reaping */
Linus Torvalds1da177e2005-04-16 15:20:36 -070079 sector_t final_block_in_request;/* doesn't change */
80 unsigned first_block_in_page; /* doesn't change, Used only once */
81 int boundary; /* prev block is at a boundary */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080082 get_block_t *get_block; /* block mapping function */
Josef Bacikfacd07b2010-05-23 11:00:55 -040083 dio_submit_t *submit_io; /* IO submition function */
Andi Kleeneb28be22011-08-01 21:38:03 -070084
Josef Bacikfacd07b2010-05-23 11:00:55 -040085 loff_t logical_offset_in_bio; /* current first logical block in bio */
Linus Torvalds1da177e2005-04-16 15:20:36 -070086 sector_t final_block_in_bio; /* current final block in bio + 1 */
87 sector_t next_block_for_io; /* next block to be put under IO,
88 in dio_blocks units */
Linus Torvalds1da177e2005-04-16 15:20:36 -070089
90 /*
91 * Deferred addition of a page to the dio. These variables are
92 * private to dio_send_cur_page(), submit_page_section() and
93 * dio_bio_add_page().
94 */
95 struct page *cur_page; /* The page */
96 unsigned cur_page_offset; /* Offset into it, in bytes */
97 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
98 sector_t cur_page_block; /* Where it starts */
Josef Bacikfacd07b2010-05-23 11:00:55 -040099 loff_t cur_page_fs_offset; /* Offset in file */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100
Jeff Moyer23aee092009-12-15 16:47:49 -0800101 /*
102 * Page fetching state. These variables belong to dio_refill_pages().
103 */
104 int curr_page; /* changes */
105 int total_pages; /* doesn't change */
106 unsigned long curr_user_address;/* changes */
107
108 /*
109 * Page queue. These variables belong to dio_refill_pages() and
110 * dio_get_page().
111 */
112 unsigned head; /* next page to process */
113 unsigned tail; /* last valid page + 1 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700114};
115
116/* dio_state communicated between submission path and end_io */
117struct dio {
118 int flags; /* doesn't change */
Andi Kleeneb28be22011-08-01 21:38:03 -0700119 int rw;
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700120 struct inode *inode;
Andi Kleeneb28be22011-08-01 21:38:03 -0700121 loff_t i_size; /* i_size when submitted */
122 dio_iodone_t *end_io; /* IO completion function */
Andi Kleeneb28be22011-08-01 21:38:03 -0700123
Andi Kleen18772642011-08-01 21:38:07 -0700124 void *private; /* copy from map_bh.b_private */
Andi Kleeneb28be22011-08-01 21:38:03 -0700125
126 /* BIO completion state */
127 spinlock_t bio_lock; /* protects BIO fields below */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700128 int page_errors; /* errno from get_user_pages() */
129 int is_async; /* is IO async ? */
130 int io_error; /* IO error in completion path */
Andi Kleeneb28be22011-08-01 21:38:03 -0700131 unsigned long refcount; /* direct_io_worker() and bios */
132 struct bio *bio_list; /* singly linked via bi_private */
133 struct task_struct *waiter; /* waiting task (NULL if none) */
134
135 /* AIO related stuff */
136 struct kiocb *iocb; /* kiocb */
Andi Kleeneb28be22011-08-01 21:38:03 -0700137 ssize_t result; /* IO result */
138
Jeff Moyer23aee092009-12-15 16:47:49 -0800139 /*
140 * pages[] (and any fields placed after it) are not zeroed out at
141 * allocation time. Don't add new fields after pages[] unless you
142 * wish that they not be zeroed.
143 */
144 struct page *pages[DIO_PAGES]; /* page buffer */
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{
162 int ret;
163 int nr_pages;
164
Andi Kleeneb28be22011-08-01 21:38:03 -0700165 nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
Nick Pigginf5dd33c2008-07-25 19:45:25 -0700166 ret = get_user_pages_fast(
Andi Kleeneb28be22011-08-01 21:38:03 -0700167 sdio->curr_user_address, /* Where from? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168 nr_pages, /* How many pages? */
169 dio->rw == READ, /* Write to memory? */
Nick Pigginf5dd33c2008-07-25 19:45:25 -0700170 &dio->pages[0]); /* Put results here */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171
Andi Kleeneb28be22011-08-01 21:38:03 -0700172 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
Nick Piggin557ed1f2007-10-16 01:24:40 -0700173 struct page *page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174 /*
175 * A memory fault, but the filesystem has some outstanding
176 * mapped blocks. We need to use those blocks up to avoid
177 * leaking stale data in the file.
178 */
179 if (dio->page_errors == 0)
180 dio->page_errors = ret;
Nick Pigginb5810032005-10-29 18:16:12 -0700181 page_cache_get(page);
182 dio->pages[0] = page;
Andi Kleeneb28be22011-08-01 21:38:03 -0700183 sdio->head = 0;
184 sdio->tail = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185 ret = 0;
186 goto out;
187 }
188
189 if (ret >= 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700190 sdio->curr_user_address += ret * PAGE_SIZE;
191 sdio->curr_page += ret;
192 sdio->head = 0;
193 sdio->tail = ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194 ret = 0;
195 }
196out:
197 return ret;
198}
199
200/*
201 * Get another userspace page. Returns an ERR_PTR on error. Pages are
202 * buffered inside the dio so that we can call get_user_pages() against a
203 * decent number of pages, less frequently. To provide nicer use of the
204 * L1 cache.
205 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700206static inline struct page *dio_get_page(struct dio *dio,
207 struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208{
Andi Kleeneb28be22011-08-01 21:38:03 -0700209 if (dio_pages_present(sdio) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210 int ret;
211
Andi Kleeneb28be22011-08-01 21:38:03 -0700212 ret = dio_refill_pages(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213 if (ret)
214 return ERR_PTR(ret);
Andi Kleeneb28be22011-08-01 21:38:03 -0700215 BUG_ON(dio_pages_present(sdio) == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700217 return dio->pages[sdio->head++];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218}
219
Zach Brown6d544bb2006-12-10 02:20:54 -0800220/**
221 * dio_complete() - called when all DIO BIO I/O has been completed
222 * @offset: the byte offset in the file of the completed operation
223 *
224 * This releases locks as dictated by the locking type, lets interested parties
225 * know that a DIO operation has completed, and calculates the resulting return
226 * code for the operation.
227 *
228 * It lets the filesystem know if it registered an interest earlier via
229 * get_block. Pass the private field of the map buffer_head so that
230 * filesystems can use it to hold additional state between get_block calls and
231 * dio_complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232 */
Edward Shishkincd1c5842010-10-26 14:22:28 -0700233static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234{
Zach Brown6d544bb2006-12-10 02:20:54 -0800235 ssize_t transferred = 0;
236
Zach Brown8459d862006-12-10 02:21:05 -0800237 /*
238 * AIO submission can race with bio completion to get here while
239 * expecting to have the last io completed by bio completion.
240 * In that case -EIOCBQUEUED is in fact not an error we want
241 * to preserve through this call.
242 */
243 if (ret == -EIOCBQUEUED)
244 ret = 0;
245
Zach Brown6d544bb2006-12-10 02:20:54 -0800246 if (dio->result) {
247 transferred = dio->result;
248
249 /* Check for short read case */
250 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
251 transferred = dio->i_size - offset;
252 }
253
Zach Brown6d544bb2006-12-10 02:20:54 -0800254 if (ret == 0)
255 ret = dio->page_errors;
256 if (ret == 0)
257 ret = dio->io_error;
258 if (ret == 0)
259 ret = transferred;
260
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000261 if (dio->end_io && dio->result) {
262 dio->end_io(dio->iocb, offset, transferred,
Andi Kleen18772642011-08-01 21:38:07 -0700263 dio->private, ret, is_async);
Christoph Hellwig72c50522011-06-24 14:29:48 -0400264 } else {
Jan Kara54c807e2013-01-30 00:28:01 +0100265 inode_dio_done(dio->inode);
Christoph Hellwig72c50522011-06-24 14:29:48 -0400266 if (is_async)
267 aio_complete(dio->iocb, ret, 0);
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000268 }
269
Zach Brown6d544bb2006-12-10 02:20:54 -0800270 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271}
272
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273static int dio_bio_complete(struct dio *dio, struct bio *bio);
274/*
275 * Asynchronous IO callback.
276 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200277static void dio_bio_end_aio(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278{
279 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800280 unsigned long remaining;
281 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283 /* cleanup the bio */
284 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800285
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800286 spin_lock_irqsave(&dio->bio_lock, flags);
287 remaining = --dio->refcount;
288 if (remaining == 1 && dio->waiter)
Zach Brown20258b22006-12-10 02:21:01 -0800289 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800290 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b22006-12-10 02:21:01 -0800291
Zach Brown8459d862006-12-10 02:21:05 -0800292 if (remaining == 0) {
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000293 dio_complete(dio, dio->iocb->ki_pos, 0, true);
Andi Kleen6e8267f2011-08-01 21:38:06 -0700294 kmem_cache_free(dio_cache, dio);
Zach Brown8459d862006-12-10 02:21:05 -0800295 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296}
297
298/*
299 * The BIO completion handler simply queues the BIO up for the process-context
300 * handler.
301 *
302 * During I/O bi_private points at the dio. After I/O, bi_private is used to
303 * implement a singly-linked list of completed BIOs, at dio->bio_list.
304 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200305static void dio_bio_end_io(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306{
307 struct dio *dio = bio->bi_private;
308 unsigned long flags;
309
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310 spin_lock_irqsave(&dio->bio_lock, flags);
311 bio->bi_private = dio->bio_list;
312 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800313 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314 wake_up_process(dio->waiter);
315 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316}
317
Josef Bacikfacd07b2010-05-23 11:00:55 -0400318/**
319 * dio_end_io - handle the end io action for the given bio
320 * @bio: The direct io bio thats being completed
321 * @error: Error if there was one
322 *
323 * This is meant to be called by any filesystem that uses their own dio_submit_t
324 * so that the DIO specific endio actions are dealt with after the filesystem
325 * has done it's completion work.
326 */
327void dio_end_io(struct bio *bio, int error)
328{
329 struct dio *dio = bio->bi_private;
330
331 if (dio->is_async)
332 dio_bio_end_aio(bio, error);
333 else
334 dio_bio_end_io(bio, error);
335}
336EXPORT_SYMBOL_GPL(dio_end_io);
337
Andi Kleenba253fb2011-08-01 21:38:08 -0700338static inline void
Andi Kleeneb28be22011-08-01 21:38:03 -0700339dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
340 struct block_device *bdev,
341 sector_t first_sector, int nr_vecs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342{
343 struct bio *bio;
344
David Dillow20d96002011-01-20 14:44:22 -0800345 /*
346 * bio_alloc() is guaranteed to return a bio when called with
347 * __GFP_WAIT and we request a valid number of vectors.
348 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350
351 bio->bi_bdev = bdev;
352 bio->bi_sector = first_sector;
353 if (dio->is_async)
354 bio->bi_end_io = dio_bio_end_aio;
355 else
356 bio->bi_end_io = dio_bio_end_io;
357
Andi Kleeneb28be22011-08-01 21:38:03 -0700358 sdio->bio = bio;
359 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360}
361
362/*
363 * In the AIO read case we speculatively dirty the pages before starting IO.
364 * During IO completion, any of these pages which happen to have been written
365 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800366 *
367 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700369static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370{
Andi Kleeneb28be22011-08-01 21:38:03 -0700371 struct bio *bio = sdio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800372 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373
374 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800375
376 spin_lock_irqsave(&dio->bio_lock, flags);
377 dio->refcount++;
378 spin_unlock_irqrestore(&dio->bio_lock, flags);
379
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 if (dio->is_async && dio->rw == READ)
381 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800382
Andi Kleeneb28be22011-08-01 21:38:03 -0700383 if (sdio->submit_io)
384 sdio->submit_io(dio->rw, bio, dio->inode,
385 sdio->logical_offset_in_bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400386 else
387 submit_bio(dio->rw, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388
Andi Kleeneb28be22011-08-01 21:38:03 -0700389 sdio->bio = NULL;
390 sdio->boundary = 0;
391 sdio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392}
393
394/*
395 * Release any resources in case of a failure
396 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700397static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398{
Andi Kleeneb28be22011-08-01 21:38:03 -0700399 while (dio_pages_present(sdio))
400 page_cache_release(dio_get_page(dio, sdio));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401}
402
403/*
Zach Brown02732012006-12-10 02:20:59 -0800404 * Wait for the next BIO to complete. Remove it and return it. NULL is
405 * returned once all BIOs have been completed. This must only be called once
406 * all bios have been issued so that dio->refcount can only decrease. This
407 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408 */
409static struct bio *dio_await_one(struct dio *dio)
410{
411 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800412 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413
414 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800415
416 /*
417 * Wait as long as the list is empty and there are bios in flight. bio
418 * completion drops the count, maybe adds to the list, and wakes while
419 * holding the bio_lock so we don't need set_current_state()'s barrier
420 * and can call it after testing our condition.
421 */
422 while (dio->refcount > 1 && dio->bio_list == NULL) {
423 __set_current_state(TASK_UNINTERRUPTIBLE);
424 dio->waiter = current;
425 spin_unlock_irqrestore(&dio->bio_lock, flags);
426 io_schedule();
427 /* wake up sets us TASK_RUNNING */
428 spin_lock_irqsave(&dio->bio_lock, flags);
429 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430 }
Zach Brown02732012006-12-10 02:20:59 -0800431 if (dio->bio_list) {
432 bio = dio->bio_list;
433 dio->bio_list = bio->bi_private;
434 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435 spin_unlock_irqrestore(&dio->bio_lock, flags);
436 return bio;
437}
438
439/*
440 * Process one completed BIO. No locks are held.
441 */
442static int dio_bio_complete(struct dio *dio, struct bio *bio)
443{
444 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
Kent Overstreetcb34e052012-09-05 15:22:02 -0700445 struct bio_vec *bvec;
446 unsigned i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447
448 if (!uptodate)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800449 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450
451 if (dio->is_async && dio->rw == READ) {
452 bio_check_pages_dirty(bio); /* transfers ownership */
453 } else {
Kent Overstreetcb34e052012-09-05 15:22:02 -0700454 bio_for_each_segment_all(bvec, bio, i) {
455 struct page *page = bvec->bv_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456
457 if (dio->rw == READ && !PageCompound(page))
458 set_page_dirty_lock(page);
459 page_cache_release(page);
460 }
461 bio_put(bio);
462 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 return uptodate ? 0 : -EIO;
464}
465
466/*
Zach Brown02732012006-12-10 02:20:59 -0800467 * Wait on and process all in-flight BIOs. This must only be called once
468 * all bios have been issued so that the refcount can only decrease.
469 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200470 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800471 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800473static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474{
Zach Brown02732012006-12-10 02:20:59 -0800475 struct bio *bio;
476 do {
477 bio = dio_await_one(dio);
478 if (bio)
479 dio_bio_complete(dio, bio);
480 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481}
482
483/*
484 * A really large O_DIRECT read or write can generate a lot of BIOs. So
485 * to keep the memory consumption sane we periodically reap any completed BIOs
486 * during the BIO generation phase.
487 *
488 * This also helps to limit the peak amount of pinned userspace memory.
489 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700490static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491{
492 int ret = 0;
493
Andi Kleeneb28be22011-08-01 21:38:03 -0700494 if (sdio->reap_counter++ >= 64) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495 while (dio->bio_list) {
496 unsigned long flags;
497 struct bio *bio;
498 int ret2;
499
500 spin_lock_irqsave(&dio->bio_lock, flags);
501 bio = dio->bio_list;
502 dio->bio_list = bio->bi_private;
503 spin_unlock_irqrestore(&dio->bio_lock, flags);
504 ret2 = dio_bio_complete(dio, bio);
505 if (ret == 0)
506 ret = ret2;
507 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700508 sdio->reap_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 }
510 return ret;
511}
512
513/*
514 * Call into the fs to map some more disk blocks. We record the current number
Andi Kleeneb28be22011-08-01 21:38:03 -0700515 * of available blocks at sdio->blocks_available. These are in units of the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 * fs blocksize, (1 << inode->i_blkbits).
517 *
518 * The fs is allowed to map lots of blocks at once. If it wants to do that,
519 * it uses the passed inode-relative block number as the file offset, as usual.
520 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800521 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 * has remaining to do. The fs should not map more than this number of blocks.
523 *
524 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
525 * indicate how much contiguous disk space has been made available at
526 * bh->b_blocknr.
527 *
528 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
529 * This isn't very efficient...
530 *
531 * In the case of filesystem holes: the fs may return an arbitrarily-large
532 * hole by returning an appropriate value in b_size and by clearing
533 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800534 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 */
Andi Kleen18772642011-08-01 21:38:07 -0700536static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
537 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538{
539 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
Tao Maae55e1a2012-01-12 17:20:33 -0800541 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542 unsigned long fs_count; /* Number of filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543 int create;
Linus Torvaldsab738572012-11-29 12:27:00 -0800544 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545
546 /*
547 * If there was a memory error and we've overwritten all the
548 * mapped blocks then we can now return that memory error
549 */
550 ret = dio->page_errors;
551 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700552 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
553 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
Tao Maae55e1a2012-01-12 17:20:33 -0800554 fs_endblk = (sdio->final_block_in_request - 1) >>
555 sdio->blkfactor;
556 fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557
Nathan Scott3c674e72006-03-29 09:26:15 +1000558 map_bh->b_state = 0;
Linus Torvaldsab738572012-11-29 12:27:00 -0800559 map_bh->b_size = fs_count << i_blkbits;
Nathan Scott3c674e72006-03-29 09:26:15 +1000560
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800561 /*
562 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
563 * forbid block creations: only overwrites are permitted.
564 * We will return early to the caller once we see an
565 * unmapped buffer head returned, and the caller will fall
566 * back to buffered I/O.
567 *
568 * Otherwise the decision is left to the get_blocks method,
569 * which may decide to handle it or also return an unmapped
570 * buffer head.
571 */
Jens Axboeb31dc662006-06-13 08:26:10 +0200572 create = dio->rw & WRITE;
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800573 if (dio->flags & DIO_SKIP_HOLES) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700574 if (sdio->block_in_file < (i_size_read(dio->inode) >>
575 sdio->blkbits))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000578
Andi Kleeneb28be22011-08-01 21:38:03 -0700579 ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 map_bh, create);
Andi Kleen18772642011-08-01 21:38:07 -0700581
582 /* Store for completion */
583 dio->private = map_bh->b_private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 }
585 return ret;
586}
587
588/*
589 * There is no bio. Make one now.
590 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700591static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
592 sector_t start_sector, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593{
594 sector_t sector;
595 int ret, nr_pages;
596
Andi Kleeneb28be22011-08-01 21:38:03 -0700597 ret = dio_bio_reap(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 if (ret)
599 goto out;
Andi Kleeneb28be22011-08-01 21:38:03 -0700600 sector = start_sector << (sdio->blkbits - 9);
Andi Kleen18772642011-08-01 21:38:07 -0700601 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
David Dillow20d96002011-01-20 14:44:22 -0800602 nr_pages = min(nr_pages, BIO_MAX_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603 BUG_ON(nr_pages <= 0);
Andi Kleen18772642011-08-01 21:38:07 -0700604 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
Andi Kleeneb28be22011-08-01 21:38:03 -0700605 sdio->boundary = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606out:
607 return ret;
608}
609
610/*
611 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
612 * that was successful then update final_block_in_bio and take a ref against
613 * the just-added page.
614 *
615 * Return zero on success. Non-zero means the caller needs to start a new BIO.
616 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700617static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618{
619 int ret;
620
Andi Kleeneb28be22011-08-01 21:38:03 -0700621 ret = bio_add_page(sdio->bio, sdio->cur_page,
622 sdio->cur_page_len, sdio->cur_page_offset);
623 if (ret == sdio->cur_page_len) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 /*
625 * Decrement count only, if we are done with this page
626 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700627 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
628 sdio->pages_in_io--;
629 page_cache_get(sdio->cur_page);
630 sdio->final_block_in_bio = sdio->cur_page_block +
631 (sdio->cur_page_len >> sdio->blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632 ret = 0;
633 } else {
634 ret = 1;
635 }
636 return ret;
637}
638
639/*
640 * Put cur_page under IO. The section of cur_page which is described by
641 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
642 * starts on-disk at cur_page_block.
643 *
644 * We take a ref against the page here (on behalf of its presence in the bio).
645 *
646 * The caller of this function is responsible for removing cur_page from the
647 * dio, and for dropping the refcount which came from that presence.
648 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700649static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
650 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651{
652 int ret = 0;
653
Andi Kleeneb28be22011-08-01 21:38:03 -0700654 if (sdio->bio) {
655 loff_t cur_offset = sdio->cur_page_fs_offset;
656 loff_t bio_next_offset = sdio->logical_offset_in_bio +
657 sdio->bio->bi_size;
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400658
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400660 * See whether this new request is contiguous with the old.
661 *
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900662 * Btrfs cannot handle having logically non-contiguous requests
663 * submitted. For example if you have
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400664 *
665 * Logical: [0-4095][HOLE][8192-12287]
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900666 * Physical: [0-4095] [4096-8191]
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400667 *
668 * We cannot submit those pages together as one BIO. So if our
669 * current logical offset in the file does not equal what would
670 * be the next logical offset in the bio, submit the bio we
671 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700673 if (sdio->final_block_in_bio != sdio->cur_page_block ||
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400674 cur_offset != bio_next_offset)
Andi Kleeneb28be22011-08-01 21:38:03 -0700675 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 }
677
Andi Kleeneb28be22011-08-01 21:38:03 -0700678 if (sdio->bio == NULL) {
Andi Kleen18772642011-08-01 21:38:07 -0700679 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680 if (ret)
681 goto out;
682 }
683
Andi Kleeneb28be22011-08-01 21:38:03 -0700684 if (dio_bio_add_page(sdio) != 0) {
685 dio_bio_submit(dio, sdio);
Andi Kleen18772642011-08-01 21:38:07 -0700686 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700688 ret = dio_bio_add_page(sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689 BUG_ON(ret != 0);
690 }
691 }
692out:
693 return ret;
694}
695
696/*
697 * An autonomous function to put a chunk of a page under deferred IO.
698 *
699 * The caller doesn't actually know (or care) whether this piece of page is in
700 * a BIO, or is under IO or whatever. We just take care of all possible
701 * situations here. The separation between the logic of do_direct_IO() and
702 * that of submit_page_section() is important for clarity. Please don't break.
703 *
704 * The chunk of page starts on-disk at blocknr.
705 *
706 * We perform deferred IO, by recording the last-submitted page inside our
707 * private part of the dio structure. If possible, we just expand the IO
708 * across that page here.
709 *
710 * If that doesn't work out then we put the old page into the bio and add this
711 * page to the dio instead.
712 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700713static inline int
Andi Kleeneb28be22011-08-01 21:38:03 -0700714submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
Andi Kleen18772642011-08-01 21:38:07 -0700715 unsigned offset, unsigned len, sector_t blocknr,
716 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717{
718 int ret = 0;
719
Andrew Morton98c4d572006-12-10 02:19:47 -0800720 if (dio->rw & WRITE) {
721 /*
722 * Read accounting is performed in submit_bio()
723 */
724 task_io_account_write(len);
725 }
726
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 /*
728 * Can we just grow the current page's presence in the dio?
729 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700730 if (sdio->cur_page == page &&
731 sdio->cur_page_offset + sdio->cur_page_len == offset &&
732 sdio->cur_page_block +
733 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
734 sdio->cur_page_len += len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735 goto out;
736 }
737
738 /*
739 * If there's a deferred page already there then send it.
740 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700741 if (sdio->cur_page) {
Andi Kleen18772642011-08-01 21:38:07 -0700742 ret = dio_send_cur_page(dio, sdio, map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -0700743 page_cache_release(sdio->cur_page);
744 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 if (ret)
Jan Karab1058b92013-04-29 15:06:18 -0700746 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 }
748
749 page_cache_get(page); /* It is in dio */
Andi Kleeneb28be22011-08-01 21:38:03 -0700750 sdio->cur_page = page;
751 sdio->cur_page_offset = offset;
752 sdio->cur_page_len = len;
753 sdio->cur_page_block = blocknr;
754 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755out:
Jan Karab1058b92013-04-29 15:06:18 -0700756 /*
757 * If sdio->boundary then we want to schedule the IO now to
758 * avoid metadata seeks.
759 */
760 if (sdio->boundary) {
761 ret = dio_send_cur_page(dio, sdio, map_bh);
762 dio_bio_submit(dio, sdio);
763 page_cache_release(sdio->cur_page);
764 sdio->cur_page = NULL;
765 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766 return ret;
767}
768
769/*
770 * Clean any dirty buffers in the blockdev mapping which alias newly-created
771 * file blocks. Only called for S_ISREG files - blockdevs do not set
772 * buffer_new
773 */
Andi Kleen18772642011-08-01 21:38:07 -0700774static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775{
776 unsigned i;
777 unsigned nblocks;
778
Andi Kleen18772642011-08-01 21:38:07 -0700779 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780
781 for (i = 0; i < nblocks; i++) {
Andi Kleen18772642011-08-01 21:38:07 -0700782 unmap_underlying_metadata(map_bh->b_bdev,
783 map_bh->b_blocknr + i);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 }
785}
786
787/*
788 * If we are not writing the entire block and get_block() allocated
789 * the block for us, we need to fill-in the unused portion of the
790 * block with zeros. This happens only if user-buffer, fileoffset or
791 * io length is not filesystem block-size multiple.
792 *
793 * `end' is zero if we're doing the start of the IO, 1 at the end of the
794 * IO.
795 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700796static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
797 int end, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798{
799 unsigned dio_blocks_per_fs_block;
800 unsigned this_chunk_blocks; /* In dio_blocks */
801 unsigned this_chunk_bytes;
802 struct page *page;
803
Andi Kleeneb28be22011-08-01 21:38:03 -0700804 sdio->start_zero_done = 1;
Andi Kleen18772642011-08-01 21:38:07 -0700805 if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806 return;
807
Andi Kleeneb28be22011-08-01 21:38:03 -0700808 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
809 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810
811 if (!this_chunk_blocks)
812 return;
813
814 /*
815 * We need to zero out part of an fs block. It is either at the
816 * beginning or the end of the fs block.
817 */
818 if (end)
819 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
820
Andi Kleeneb28be22011-08-01 21:38:03 -0700821 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822
Nick Piggin557ed1f2007-10-16 01:24:40 -0700823 page = ZERO_PAGE(0);
Andi Kleeneb28be22011-08-01 21:38:03 -0700824 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700825 sdio->next_block_for_io, map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 return;
827
Andi Kleeneb28be22011-08-01 21:38:03 -0700828 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829}
830
831/*
832 * Walk the user pages, and the file, mapping blocks to disk and generating
833 * a sequence of (page,offset,len,block) mappings. These mappings are injected
834 * into submit_page_section(), which takes care of the next stage of submission
835 *
836 * Direct IO against a blockdev is different from a file. Because we can
837 * happily perform page-sized but 512-byte aligned IOs. It is important that
838 * blockdev IO be able to have fine alignment and large sizes.
839 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800840 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 * with the size of IO which is permitted at this offset and this i_blkbits.
842 *
843 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800844 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845 * fine alignment but still allows this function to work in PAGE_SIZE units.
846 */
Andi Kleen18772642011-08-01 21:38:07 -0700847static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
848 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849{
Andi Kleeneb28be22011-08-01 21:38:03 -0700850 const unsigned blkbits = sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
852 struct page *page;
853 unsigned block_in_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 int ret = 0;
855
856 /* The I/O can start at any block offset within the first page */
Andi Kleeneb28be22011-08-01 21:38:03 -0700857 block_in_page = sdio->first_block_in_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858
Andi Kleeneb28be22011-08-01 21:38:03 -0700859 while (sdio->block_in_file < sdio->final_block_in_request) {
860 page = dio_get_page(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861 if (IS_ERR(page)) {
862 ret = PTR_ERR(page);
863 goto out;
864 }
865
866 while (block_in_page < blocks_per_page) {
867 unsigned offset_in_page = block_in_page << blkbits;
868 unsigned this_chunk_bytes; /* # of bytes mapped */
869 unsigned this_chunk_blocks; /* # of blocks */
870 unsigned u;
871
Andi Kleeneb28be22011-08-01 21:38:03 -0700872 if (sdio->blocks_available == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873 /*
874 * Need to go and map some more disk
875 */
876 unsigned long blkmask;
877 unsigned long dio_remainder;
878
Andi Kleen18772642011-08-01 21:38:07 -0700879 ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880 if (ret) {
881 page_cache_release(page);
882 goto out;
883 }
884 if (!buffer_mapped(map_bh))
885 goto do_holes;
886
Andi Kleeneb28be22011-08-01 21:38:03 -0700887 sdio->blocks_available =
888 map_bh->b_size >> sdio->blkbits;
889 sdio->next_block_for_io =
890 map_bh->b_blocknr << sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 if (buffer_new(map_bh))
Andi Kleen18772642011-08-01 21:38:07 -0700892 clean_blockdev_aliases(dio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893
Andi Kleeneb28be22011-08-01 21:38:03 -0700894 if (!sdio->blkfactor)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 goto do_holes;
896
Andi Kleeneb28be22011-08-01 21:38:03 -0700897 blkmask = (1 << sdio->blkfactor) - 1;
898 dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899
900 /*
901 * If we are at the start of IO and that IO
902 * starts partway into a fs-block,
903 * dio_remainder will be non-zero. If the IO
904 * is a read then we can simply advance the IO
905 * cursor to the first block which is to be
906 * read. But if the IO is a write and the
907 * block was newly allocated we cannot do that;
908 * the start of the fs block must be zeroed out
909 * on-disk
910 */
911 if (!buffer_new(map_bh))
Andi Kleeneb28be22011-08-01 21:38:03 -0700912 sdio->next_block_for_io += dio_remainder;
913 sdio->blocks_available -= dio_remainder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 }
915do_holes:
916 /* Handle holes */
917 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800918 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919
920 /* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboeb31dc662006-06-13 08:26:10 +0200921 if (dio->rw & WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 page_cache_release(page);
923 return -ENOTBLK;
924 }
925
Jeff Moyer35dc8162006-02-03 03:04:27 -0800926 /*
927 * Be sure to account for a partial block as the
928 * last block in the file
929 */
930 i_size_aligned = ALIGN(i_size_read(dio->inode),
931 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -0700932 if (sdio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -0800933 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934 /* We hit eof */
935 page_cache_release(page);
936 goto out;
937 }
Christoph Lametereebd2aa2008-02-04 22:28:29 -0800938 zero_user(page, block_in_page << blkbits,
939 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -0700940 sdio->block_in_file++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 block_in_page++;
942 goto next_block;
943 }
944
945 /*
946 * If we're performing IO which has an alignment which
947 * is finer than the underlying fs, go check to see if
948 * we must zero out the start of this block.
949 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700950 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
Andi Kleen18772642011-08-01 21:38:07 -0700951 dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952
953 /*
954 * Work out, in this_chunk_blocks, how much disk we
955 * can add to this page
956 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700957 this_chunk_blocks = sdio->blocks_available;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 u = (PAGE_SIZE - offset_in_page) >> blkbits;
959 if (this_chunk_blocks > u)
960 this_chunk_blocks = u;
Andi Kleeneb28be22011-08-01 21:38:03 -0700961 u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 if (this_chunk_blocks > u)
963 this_chunk_blocks = u;
964 this_chunk_bytes = this_chunk_blocks << blkbits;
965 BUG_ON(this_chunk_bytes == 0);
966
Jan Kara092c8d42013-04-29 15:06:17 -0700967 if (this_chunk_blocks == sdio->blocks_available)
968 sdio->boundary = buffer_boundary(map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -0700969 ret = submit_page_section(dio, sdio, page,
970 offset_in_page,
971 this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700972 sdio->next_block_for_io,
973 map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 if (ret) {
975 page_cache_release(page);
976 goto out;
977 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700978 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979
Andi Kleeneb28be22011-08-01 21:38:03 -0700980 sdio->block_in_file += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981 block_in_page += this_chunk_blocks;
Andi Kleeneb28be22011-08-01 21:38:03 -0700982 sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983next_block:
Andi Kleeneb28be22011-08-01 21:38:03 -0700984 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
985 if (sdio->block_in_file == sdio->final_block_in_request)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 break;
987 }
988
989 /* Drop the ref which was taken in get_user_pages() */
990 page_cache_release(page);
991 block_in_page = 0;
992 }
993out:
994 return ret;
995}
996
Andi Kleen847cc632011-08-01 21:38:09 -0700997static inline int drop_refcount(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998{
Andi Kleen847cc632011-08-01 21:38:09 -0700999 int ret2;
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001000 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001
Zach Brown8459d862006-12-10 02:21:05 -08001002 /*
1003 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001004 * operation. AIO can if it was a broken operation described above or
1005 * in fact if all the bios race to complete before we get here. In
1006 * that case dio_complete() translates the EIOCBQUEUED into the proper
1007 * return code that the caller will hand to aio_complete().
1008 *
1009 * This is managed by the bio_lock instead of being an atomic_t so that
1010 * completion paths can drop their ref and use the remaining count to
1011 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001012 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001013 spin_lock_irqsave(&dio->bio_lock, flags);
1014 ret2 = --dio->refcount;
1015 spin_unlock_irqrestore(&dio->bio_lock, flags);
Andi Kleen847cc632011-08-01 21:38:09 -07001016 return ret2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017}
1018
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001019/*
1020 * This is a library function for use by filesystem drivers.
1021 *
1022 * The locking rules are governed by the flags parameter:
1023 * - if the flags value contains DIO_LOCKING we use a fancy locking
1024 * scheme for dumb filesystems.
1025 * For writes this function is called under i_mutex and returns with
1026 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1027 * taken and dropped again before returning.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001028 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1029 * internal locking but rather rely on the filesystem to synchronize
1030 * direct I/O reads/writes versus each other and truncate.
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001031 *
1032 * To help with locking against truncate we incremented the i_dio_count
1033 * counter before starting direct I/O, and decrement it once we are done.
1034 * Truncate can wait for it to reach zero to provide exclusion. It is
1035 * expected that filesystem provide exclusion between new direct I/O
1036 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1037 * but other filesystems need to take care of this on their own.
Andi Kleenba253fb2011-08-01 21:38:08 -07001038 *
1039 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1040 * is always inlined. Otherwise gcc is unable to split the structure into
1041 * individual fields and will generate much worse code. This is important
1042 * for the whole file.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001043 */
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001044static inline ssize_t
1045do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046 struct block_device *bdev, const struct iovec *iov, loff_t offset,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -08001047 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
Josef Bacikfacd07b2010-05-23 11:00:55 -04001048 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001049{
1050 int seg;
1051 size_t size;
1052 unsigned long addr;
Linus Torvaldsab738572012-11-29 12:27:00 -08001053 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1054 unsigned blkbits = i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055 unsigned blocksize_mask = (1 << blkbits) - 1;
1056 ssize_t retval = -EINVAL;
1057 loff_t end = offset;
1058 struct dio *dio;
Andi Kleeneb28be22011-08-01 21:38:03 -07001059 struct dio_submit sdio = { 0, };
Andi Kleen847cc632011-08-01 21:38:09 -07001060 unsigned long user_addr;
1061 size_t bytes;
1062 struct buffer_head map_bh = { 0, };
Fengguang Wu647d1e42012-08-09 15:23:09 +02001063 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064
1065 if (rw & WRITE)
Jens Axboe721a9602011-03-09 11:56:30 +01001066 rw = WRITE_ODIRECT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001067
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001068 /*
1069 * Avoid references to bdev if not absolutely needed to give
1070 * the early prefetch in the caller enough time.
1071 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072
1073 if (offset & blocksize_mask) {
1074 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001075 blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076 blocksize_mask = (1 << blkbits) - 1;
1077 if (offset & blocksize_mask)
1078 goto out;
1079 }
1080
1081 /* Check the memory alignment. Blocks cannot straddle pages */
1082 for (seg = 0; seg < nr_segs; seg++) {
1083 addr = (unsigned long)iov[seg].iov_base;
1084 size = iov[seg].iov_len;
1085 end += size;
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001086 if (unlikely((addr & blocksize_mask) ||
1087 (size & blocksize_mask))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001089 blkbits = blksize_bits(
1090 bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091 blocksize_mask = (1 << blkbits) - 1;
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001092 if ((addr & blocksize_mask) || (size & blocksize_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 goto out;
1094 }
1095 }
1096
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001097 /* watch out for a 0 len io from a tricksy fs */
1098 if (rw == READ && end == offset)
1099 return 0;
1100
Andi Kleen6e8267f2011-08-01 21:38:06 -07001101 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 retval = -ENOMEM;
1103 if (!dio)
1104 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001105 /*
1106 * Believe it or not, zeroing out the page array caused a .5%
1107 * performance regression in a database benchmark. So, we take
1108 * care to only zero out what's needed.
1109 */
1110 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001112 dio->flags = flags;
1113 if (dio->flags & DIO_LOCKING) {
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001114 if (rw == READ) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001115 struct address_space *mapping =
1116 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001118 /* will be released by direct_io_worker */
1119 mutex_lock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120
1121 retval = filemap_write_and_wait_range(mapping, offset,
1122 end - 1);
1123 if (retval) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001124 mutex_unlock(&inode->i_mutex);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001125 kmem_cache_free(dio_cache, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 goto out;
1127 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 }
1130
1131 /*
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001132 * Will be decremented at I/O completion time.
1133 */
1134 atomic_inc(&inode->i_dio_count);
1135
1136 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137 * For file extending writes updating i_size before data
1138 * writeouts complete can expose uninitialized blocks. So
1139 * even for AIO, we need to wait for i/o to complete before
1140 * returning in this case.
1141 */
Jens Axboeb31dc662006-06-13 08:26:10 +02001142 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 (end > i_size_read(inode)));
1144
Andi Kleen847cc632011-08-01 21:38:09 -07001145 retval = 0;
1146
1147 dio->inode = inode;
1148 dio->rw = rw;
1149 sdio.blkbits = blkbits;
Linus Torvaldsab738572012-11-29 12:27:00 -08001150 sdio.blkfactor = i_blkbits - blkbits;
Andi Kleen847cc632011-08-01 21:38:09 -07001151 sdio.block_in_file = offset >> blkbits;
1152
1153 sdio.get_block = get_block;
1154 dio->end_io = end_io;
1155 sdio.submit_io = submit_io;
1156 sdio.final_block_in_bio = -1;
1157 sdio.next_block_for_io = -1;
1158
1159 dio->iocb = iocb;
1160 dio->i_size = i_size_read(inode);
1161
1162 spin_lock_init(&dio->bio_lock);
1163 dio->refcount = 1;
1164
1165 /*
1166 * In case of non-aligned buffers, we may need 2 more
1167 * pages since we need to zero out first and last block.
1168 */
1169 if (unlikely(sdio.blkfactor))
1170 sdio.pages_in_io = 2;
1171
1172 for (seg = 0; seg < nr_segs; seg++) {
1173 user_addr = (unsigned long)iov[seg].iov_base;
1174 sdio.pages_in_io +=
1175 ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) /
1176 PAGE_SIZE - user_addr / PAGE_SIZE);
1177 }
1178
Fengguang Wu647d1e42012-08-09 15:23:09 +02001179 blk_start_plug(&plug);
1180
Andi Kleen847cc632011-08-01 21:38:09 -07001181 for (seg = 0; seg < nr_segs; seg++) {
1182 user_addr = (unsigned long)iov[seg].iov_base;
1183 sdio.size += bytes = iov[seg].iov_len;
1184
1185 /* Index into the first page of the first block */
1186 sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1187 sdio.final_block_in_request = sdio.block_in_file +
1188 (bytes >> blkbits);
1189 /* Page fetching state */
1190 sdio.head = 0;
1191 sdio.tail = 0;
1192 sdio.curr_page = 0;
1193
1194 sdio.total_pages = 0;
1195 if (user_addr & (PAGE_SIZE-1)) {
1196 sdio.total_pages++;
1197 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1198 }
1199 sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1200 sdio.curr_user_address = user_addr;
1201
1202 retval = do_direct_IO(dio, &sdio, &map_bh);
1203
1204 dio->result += iov[seg].iov_len -
1205 ((sdio.final_block_in_request - sdio.block_in_file) <<
1206 blkbits);
1207
1208 if (retval) {
1209 dio_cleanup(dio, &sdio);
1210 break;
1211 }
1212 } /* end iovec loop */
1213
1214 if (retval == -ENOTBLK) {
1215 /*
1216 * The remaining part of the request will be
1217 * be handled by buffered I/O when we return
1218 */
1219 retval = 0;
1220 }
1221 /*
1222 * There may be some unwritten disk at the end of a part-written
1223 * fs-block-sized block. Go zero that now.
1224 */
1225 dio_zero_block(dio, &sdio, 1, &map_bh);
1226
1227 if (sdio.cur_page) {
1228 ssize_t ret2;
1229
1230 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1231 if (retval == 0)
1232 retval = ret2;
1233 page_cache_release(sdio.cur_page);
1234 sdio.cur_page = NULL;
1235 }
1236 if (sdio.bio)
1237 dio_bio_submit(dio, &sdio);
1238
Fengguang Wu647d1e42012-08-09 15:23:09 +02001239 blk_finish_plug(&plug);
1240
Andi Kleen847cc632011-08-01 21:38:09 -07001241 /*
1242 * It is possible that, we return short IO due to end of file.
1243 * In that case, we need to release all the pages we got hold on.
1244 */
1245 dio_cleanup(dio, &sdio);
1246
1247 /*
1248 * All block lookups have been performed. For READ requests
1249 * we can let i_mutex go now that its achieved its purpose
1250 * of protecting us from looking up uninitialized blocks.
1251 */
1252 if (rw == READ && (dio->flags & DIO_LOCKING))
1253 mutex_unlock(&dio->inode->i_mutex);
1254
1255 /*
1256 * The only time we want to leave bios in flight is when a successful
1257 * partial aio read or full aio write have been setup. In that case
1258 * bio completion will call aio_complete. The only time it's safe to
1259 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1260 * This had *better* be the only place that raises -EIOCBQUEUED.
1261 */
1262 BUG_ON(retval == -EIOCBQUEUED);
1263 if (dio->is_async && retval == 0 && dio->result &&
Julia Lawalld1876632012-06-07 15:45:00 -07001264 ((rw == READ) || (dio->result == sdio.size)))
Andi Kleen847cc632011-08-01 21:38:09 -07001265 retval = -EIOCBQUEUED;
1266
1267 if (retval != -EIOCBQUEUED)
1268 dio_await_completion(dio);
1269
1270 if (drop_refcount(dio) == 0) {
1271 retval = dio_complete(dio, offset, retval, false);
1272 kmem_cache_free(dio_cache, dio);
1273 } else
1274 BUG_ON(retval != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001276out:
1277 return retval;
1278}
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001279
1280ssize_t
1281__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1282 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1283 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1284 dio_submit_t submit_io, int flags)
1285{
1286 /*
1287 * The block device state is needed in the end to finally
1288 * submit everything. Since it's likely to be cache cold
1289 * prefetch it here as first thing to hide some of the
1290 * latency.
1291 *
1292 * Attempt to prefetch the pieces we likely need later.
1293 */
1294 prefetch(&bdev->bd_disk->part_tbl);
1295 prefetch(bdev->bd_queue);
1296 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1297
1298 return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset,
1299 nr_segs, get_block, end_io,
1300 submit_io, flags);
1301}
1302
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303EXPORT_SYMBOL(__blockdev_direct_IO);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001304
1305static __init int dio_init(void)
1306{
1307 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1308 return 0;
1309}
1310module_init(dio_init)