blob: f853263cf74f8abf29016db5e44eaa9e0de44abc [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 */
73 size_t size; /* total request size (doesn't change)*/
74 sector_t block_in_file; /* Current offset into the underlying
75 file in dio_block units. */
76 unsigned blocks_available; /* At block_in_file. changes */
Andi Kleen0dc2bc42011-08-01 21:38:05 -070077 int reap_counter; /* rate limit reaping */
Linus Torvalds1da177e2005-04-16 15:20:36 -070078 sector_t final_block_in_request;/* doesn't change */
79 unsigned first_block_in_page; /* doesn't change, Used only once */
80 int boundary; /* prev block is at a boundary */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080081 get_block_t *get_block; /* block mapping function */
Josef Bacikfacd07b2010-05-23 11:00:55 -040082 dio_submit_t *submit_io; /* IO submition function */
Andi Kleeneb28be22011-08-01 21:38:03 -070083
Josef Bacikfacd07b2010-05-23 11:00:55 -040084 loff_t logical_offset_in_bio; /* current first logical block in bio */
Linus Torvalds1da177e2005-04-16 15:20:36 -070085 sector_t final_block_in_bio; /* current final block in bio + 1 */
86 sector_t next_block_for_io; /* next block to be put under IO,
87 in dio_blocks units */
Linus Torvalds1da177e2005-04-16 15:20:36 -070088
89 /*
90 * Deferred addition of a page to the dio. These variables are
91 * private to dio_send_cur_page(), submit_page_section() and
92 * dio_bio_add_page().
93 */
94 struct page *cur_page; /* The page */
95 unsigned cur_page_offset; /* Offset into it, in bytes */
96 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
97 sector_t cur_page_block; /* Where it starts */
Josef Bacikfacd07b2010-05-23 11:00:55 -040098 loff_t cur_page_fs_offset; /* Offset in file */
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
Jeff Moyer23aee092009-12-15 16:47:49 -0800100 /*
101 * Page fetching state. These variables belong to dio_refill_pages().
102 */
103 int curr_page; /* changes */
104 int total_pages; /* doesn't change */
105 unsigned long curr_user_address;/* changes */
106
107 /*
108 * Page queue. These variables belong to dio_refill_pages() and
109 * dio_get_page().
110 */
111 unsigned head; /* next page to process */
112 unsigned tail; /* last valid page + 1 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700113};
114
115/* dio_state communicated between submission path and end_io */
116struct dio {
117 int flags; /* doesn't change */
Andi Kleeneb28be22011-08-01 21:38:03 -0700118 int rw;
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700119 struct inode *inode;
Andi Kleeneb28be22011-08-01 21:38:03 -0700120 loff_t i_size; /* i_size when submitted */
121 dio_iodone_t *end_io; /* IO completion function */
Andi Kleeneb28be22011-08-01 21:38:03 -0700122
Andi Kleen18772642011-08-01 21:38:07 -0700123 void *private; /* copy from map_bh.b_private */
Andi Kleeneb28be22011-08-01 21:38:03 -0700124
125 /* BIO completion state */
126 spinlock_t bio_lock; /* protects BIO fields below */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700127 int page_errors; /* errno from get_user_pages() */
128 int is_async; /* is IO async ? */
129 int io_error; /* IO error in completion path */
Andi Kleeneb28be22011-08-01 21:38:03 -0700130 unsigned long refcount; /* direct_io_worker() and bios */
131 struct bio *bio_list; /* singly linked via bi_private */
132 struct task_struct *waiter; /* waiting task (NULL if none) */
133
134 /* AIO related stuff */
135 struct kiocb *iocb; /* kiocb */
Andi Kleeneb28be22011-08-01 21:38:03 -0700136 ssize_t result; /* IO result */
137
Jeff Moyer23aee092009-12-15 16:47:49 -0800138 /*
139 * pages[] (and any fields placed after it) are not zeroed out at
140 * allocation time. Don't add new fields after pages[] unless you
141 * wish that they not be zeroed.
142 */
143 struct page *pages[DIO_PAGES]; /* page buffer */
Andi Kleen6e8267f2011-08-01 21:38:06 -0700144} ____cacheline_aligned_in_smp;
145
146static struct kmem_cache *dio_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147
148/*
149 * How many pages are in the queue?
150 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700151static inline unsigned dio_pages_present(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152{
Andi Kleeneb28be22011-08-01 21:38:03 -0700153 return sdio->tail - sdio->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154}
155
156/*
157 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
158 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700159static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160{
161 int ret;
162 int nr_pages;
163
Andi Kleeneb28be22011-08-01 21:38:03 -0700164 nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
Nick Pigginf5dd33c2008-07-25 19:45:25 -0700165 ret = get_user_pages_fast(
Andi Kleeneb28be22011-08-01 21:38:03 -0700166 sdio->curr_user_address, /* Where from? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 nr_pages, /* How many pages? */
168 dio->rw == READ, /* Write to memory? */
Nick Pigginf5dd33c2008-07-25 19:45:25 -0700169 &dio->pages[0]); /* Put results here */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170
Andi Kleeneb28be22011-08-01 21:38:03 -0700171 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
Nick Piggin557ed1f2007-10-16 01:24:40 -0700172 struct page *page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173 /*
174 * A memory fault, but the filesystem has some outstanding
175 * mapped blocks. We need to use those blocks up to avoid
176 * leaking stale data in the file.
177 */
178 if (dio->page_errors == 0)
179 dio->page_errors = ret;
Nick Pigginb5810032005-10-29 18:16:12 -0700180 page_cache_get(page);
181 dio->pages[0] = page;
Andi Kleeneb28be22011-08-01 21:38:03 -0700182 sdio->head = 0;
183 sdio->tail = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 ret = 0;
185 goto out;
186 }
187
188 if (ret >= 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700189 sdio->curr_user_address += ret * PAGE_SIZE;
190 sdio->curr_page += ret;
191 sdio->head = 0;
192 sdio->tail = ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193 ret = 0;
194 }
195out:
196 return ret;
197}
198
199/*
200 * Get another userspace page. Returns an ERR_PTR on error. Pages are
201 * buffered inside the dio so that we can call get_user_pages() against a
202 * decent number of pages, less frequently. To provide nicer use of the
203 * L1 cache.
204 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700205static inline struct page *dio_get_page(struct dio *dio,
206 struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207{
Andi Kleeneb28be22011-08-01 21:38:03 -0700208 if (dio_pages_present(sdio) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209 int ret;
210
Andi Kleeneb28be22011-08-01 21:38:03 -0700211 ret = dio_refill_pages(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212 if (ret)
213 return ERR_PTR(ret);
Andi Kleeneb28be22011-08-01 21:38:03 -0700214 BUG_ON(dio_pages_present(sdio) == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700216 return dio->pages[sdio->head++];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217}
218
Zach Brown6d544bb2006-12-10 02:20:54 -0800219/**
220 * dio_complete() - called when all DIO BIO I/O has been completed
221 * @offset: the byte offset in the file of the completed operation
222 *
223 * This releases locks as dictated by the locking type, lets interested parties
224 * know that a DIO operation has completed, and calculates the resulting return
225 * code for the operation.
226 *
227 * It lets the filesystem know if it registered an interest earlier via
228 * get_block. Pass the private field of the map buffer_head so that
229 * filesystems can use it to hold additional state between get_block calls and
230 * dio_complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231 */
Edward Shishkincd1c5842010-10-26 14:22:28 -0700232static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233{
Zach Brown6d544bb2006-12-10 02:20:54 -0800234 ssize_t transferred = 0;
235
Zach Brown8459d862006-12-10 02:21:05 -0800236 /*
237 * AIO submission can race with bio completion to get here while
238 * expecting to have the last io completed by bio completion.
239 * In that case -EIOCBQUEUED is in fact not an error we want
240 * to preserve through this call.
241 */
242 if (ret == -EIOCBQUEUED)
243 ret = 0;
244
Zach Brown6d544bb2006-12-10 02:20:54 -0800245 if (dio->result) {
246 transferred = dio->result;
247
248 /* Check for short read case */
249 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
250 transferred = dio->i_size - offset;
251 }
252
Zach Brown6d544bb2006-12-10 02:20:54 -0800253 if (ret == 0)
254 ret = dio->page_errors;
255 if (ret == 0)
256 ret = dio->io_error;
257 if (ret == 0)
258 ret = transferred;
259
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000260 if (dio->end_io && dio->result) {
261 dio->end_io(dio->iocb, offset, transferred,
Andi Kleen18772642011-08-01 21:38:07 -0700262 dio->private, ret, is_async);
Christoph Hellwig72c50522011-06-24 14:29:48 -0400263 } else {
Jan Kara54c807e2013-01-30 00:28:01 +0100264 inode_dio_done(dio->inode);
Christoph Hellwig72c50522011-06-24 14:29:48 -0400265 if (is_async)
266 aio_complete(dio->iocb, ret, 0);
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000267 }
268
Zach Brown6d544bb2006-12-10 02:20:54 -0800269 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270}
271
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272static int dio_bio_complete(struct dio *dio, struct bio *bio);
273/*
274 * Asynchronous IO callback.
275 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200276static void dio_bio_end_aio(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277{
278 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800279 unsigned long remaining;
280 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282 /* cleanup the bio */
283 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800284
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800285 spin_lock_irqsave(&dio->bio_lock, flags);
286 remaining = --dio->refcount;
287 if (remaining == 1 && dio->waiter)
Zach Brown20258b22006-12-10 02:21:01 -0800288 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800289 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b22006-12-10 02:21:01 -0800290
Zach Brown8459d862006-12-10 02:21:05 -0800291 if (remaining == 0) {
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000292 dio_complete(dio, dio->iocb->ki_pos, 0, true);
Andi Kleen6e8267f2011-08-01 21:38:06 -0700293 kmem_cache_free(dio_cache, dio);
Zach Brown8459d862006-12-10 02:21:05 -0800294 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295}
296
297/*
298 * The BIO completion handler simply queues the BIO up for the process-context
299 * handler.
300 *
301 * During I/O bi_private points at the dio. After I/O, bi_private is used to
302 * implement a singly-linked list of completed BIOs, at dio->bio_list.
303 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200304static void dio_bio_end_io(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305{
306 struct dio *dio = bio->bi_private;
307 unsigned long flags;
308
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309 spin_lock_irqsave(&dio->bio_lock, flags);
310 bio->bi_private = dio->bio_list;
311 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800312 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313 wake_up_process(dio->waiter);
314 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315}
316
Josef Bacikfacd07b2010-05-23 11:00:55 -0400317/**
318 * dio_end_io - handle the end io action for the given bio
319 * @bio: The direct io bio thats being completed
320 * @error: Error if there was one
321 *
322 * This is meant to be called by any filesystem that uses their own dio_submit_t
323 * so that the DIO specific endio actions are dealt with after the filesystem
324 * has done it's completion work.
325 */
326void dio_end_io(struct bio *bio, int error)
327{
328 struct dio *dio = bio->bi_private;
329
330 if (dio->is_async)
331 dio_bio_end_aio(bio, error);
332 else
333 dio_bio_end_io(bio, error);
334}
335EXPORT_SYMBOL_GPL(dio_end_io);
336
Andi Kleenba253fb2011-08-01 21:38:08 -0700337static inline void
Andi Kleeneb28be22011-08-01 21:38:03 -0700338dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
339 struct block_device *bdev,
340 sector_t first_sector, int nr_vecs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341{
342 struct bio *bio;
343
David Dillow20d96002011-01-20 14:44:22 -0800344 /*
345 * bio_alloc() is guaranteed to return a bio when called with
346 * __GFP_WAIT and we request a valid number of vectors.
347 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349
350 bio->bi_bdev = bdev;
351 bio->bi_sector = first_sector;
352 if (dio->is_async)
353 bio->bi_end_io = dio_bio_end_aio;
354 else
355 bio->bi_end_io = dio_bio_end_io;
356
Andi Kleeneb28be22011-08-01 21:38:03 -0700357 sdio->bio = bio;
358 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359}
360
361/*
362 * In the AIO read case we speculatively dirty the pages before starting IO.
363 * During IO completion, any of these pages which happen to have been written
364 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800365 *
366 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700368static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369{
Andi Kleeneb28be22011-08-01 21:38:03 -0700370 struct bio *bio = sdio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800371 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372
373 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800374
375 spin_lock_irqsave(&dio->bio_lock, flags);
376 dio->refcount++;
377 spin_unlock_irqrestore(&dio->bio_lock, flags);
378
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 if (dio->is_async && dio->rw == READ)
380 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800381
Andi Kleeneb28be22011-08-01 21:38:03 -0700382 if (sdio->submit_io)
383 sdio->submit_io(dio->rw, bio, dio->inode,
384 sdio->logical_offset_in_bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400385 else
386 submit_bio(dio->rw, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387
Andi Kleeneb28be22011-08-01 21:38:03 -0700388 sdio->bio = NULL;
389 sdio->boundary = 0;
390 sdio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391}
392
393/*
394 * Release any resources in case of a failure
395 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700396static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397{
Andi Kleeneb28be22011-08-01 21:38:03 -0700398 while (dio_pages_present(sdio))
399 page_cache_release(dio_get_page(dio, sdio));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400}
401
402/*
Zach Brown02732012006-12-10 02:20:59 -0800403 * Wait for the next BIO to complete. Remove it and return it. NULL is
404 * returned once all BIOs have been completed. This must only be called once
405 * all bios have been issued so that dio->refcount can only decrease. This
406 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407 */
408static struct bio *dio_await_one(struct dio *dio)
409{
410 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800411 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412
413 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800414
415 /*
416 * Wait as long as the list is empty and there are bios in flight. bio
417 * completion drops the count, maybe adds to the list, and wakes while
418 * holding the bio_lock so we don't need set_current_state()'s barrier
419 * and can call it after testing our condition.
420 */
421 while (dio->refcount > 1 && dio->bio_list == NULL) {
422 __set_current_state(TASK_UNINTERRUPTIBLE);
423 dio->waiter = current;
424 spin_unlock_irqrestore(&dio->bio_lock, flags);
425 io_schedule();
426 /* wake up sets us TASK_RUNNING */
427 spin_lock_irqsave(&dio->bio_lock, flags);
428 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429 }
Zach Brown02732012006-12-10 02:20:59 -0800430 if (dio->bio_list) {
431 bio = dio->bio_list;
432 dio->bio_list = bio->bi_private;
433 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 spin_unlock_irqrestore(&dio->bio_lock, flags);
435 return bio;
436}
437
438/*
439 * Process one completed BIO. No locks are held.
440 */
441static int dio_bio_complete(struct dio *dio, struct bio *bio)
442{
443 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
444 struct bio_vec *bvec = bio->bi_io_vec;
445 int page_no;
446
447 if (!uptodate)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800448 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449
450 if (dio->is_async && dio->rw == READ) {
451 bio_check_pages_dirty(bio); /* transfers ownership */
452 } else {
453 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
454 struct page *page = bvec[page_no].bv_page;
455
456 if (dio->rw == READ && !PageCompound(page))
457 set_page_dirty_lock(page);
458 page_cache_release(page);
459 }
460 bio_put(bio);
461 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 return uptodate ? 0 : -EIO;
463}
464
465/*
Zach Brown02732012006-12-10 02:20:59 -0800466 * Wait on and process all in-flight BIOs. This must only be called once
467 * all bios have been issued so that the refcount can only decrease.
468 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200469 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800470 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800472static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473{
Zach Brown02732012006-12-10 02:20:59 -0800474 struct bio *bio;
475 do {
476 bio = dio_await_one(dio);
477 if (bio)
478 dio_bio_complete(dio, bio);
479 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480}
481
482/*
483 * A really large O_DIRECT read or write can generate a lot of BIOs. So
484 * to keep the memory consumption sane we periodically reap any completed BIOs
485 * during the BIO generation phase.
486 *
487 * This also helps to limit the peak amount of pinned userspace memory.
488 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700489static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490{
491 int ret = 0;
492
Andi Kleeneb28be22011-08-01 21:38:03 -0700493 if (sdio->reap_counter++ >= 64) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 while (dio->bio_list) {
495 unsigned long flags;
496 struct bio *bio;
497 int ret2;
498
499 spin_lock_irqsave(&dio->bio_lock, flags);
500 bio = dio->bio_list;
501 dio->bio_list = bio->bi_private;
502 spin_unlock_irqrestore(&dio->bio_lock, flags);
503 ret2 = dio_bio_complete(dio, bio);
504 if (ret == 0)
505 ret = ret2;
506 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700507 sdio->reap_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508 }
509 return ret;
510}
511
512/*
513 * Call into the fs to map some more disk blocks. We record the current number
Andi Kleeneb28be22011-08-01 21:38:03 -0700514 * of available blocks at sdio->blocks_available. These are in units of the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 * fs blocksize, (1 << inode->i_blkbits).
516 *
517 * The fs is allowed to map lots of blocks at once. If it wants to do that,
518 * it uses the passed inode-relative block number as the file offset, as usual.
519 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800520 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 * has remaining to do. The fs should not map more than this number of blocks.
522 *
523 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
524 * indicate how much contiguous disk space has been made available at
525 * bh->b_blocknr.
526 *
527 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
528 * This isn't very efficient...
529 *
530 * In the case of filesystem holes: the fs may return an arbitrarily-large
531 * hole by returning an appropriate value in b_size and by clearing
532 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800533 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 */
Andi Kleen18772642011-08-01 21:38:07 -0700535static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
536 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537{
538 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
Tao Maae55e1a2012-01-12 17:20:33 -0800540 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541 unsigned long fs_count; /* Number of filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542 int create;
Linus Torvaldsab738572012-11-29 12:27:00 -0800543 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544
545 /*
546 * If there was a memory error and we've overwritten all the
547 * mapped blocks then we can now return that memory error
548 */
549 ret = dio->page_errors;
550 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700551 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
552 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
Tao Maae55e1a2012-01-12 17:20:33 -0800553 fs_endblk = (sdio->final_block_in_request - 1) >>
554 sdio->blkfactor;
555 fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556
Nathan Scott3c674e72006-03-29 09:26:15 +1000557 map_bh->b_state = 0;
Linus Torvaldsab738572012-11-29 12:27:00 -0800558 map_bh->b_size = fs_count << i_blkbits;
Nathan Scott3c674e72006-03-29 09:26:15 +1000559
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800560 /*
561 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
562 * forbid block creations: only overwrites are permitted.
563 * We will return early to the caller once we see an
564 * unmapped buffer head returned, and the caller will fall
565 * back to buffered I/O.
566 *
567 * Otherwise the decision is left to the get_blocks method,
568 * which may decide to handle it or also return an unmapped
569 * buffer head.
570 */
Jens Axboeb31dc662006-06-13 08:26:10 +0200571 create = dio->rw & WRITE;
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800572 if (dio->flags & DIO_SKIP_HOLES) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700573 if (sdio->block_in_file < (i_size_read(dio->inode) >>
574 sdio->blkbits))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000577
Andi Kleeneb28be22011-08-01 21:38:03 -0700578 ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 map_bh, create);
Andi Kleen18772642011-08-01 21:38:07 -0700580
581 /* Store for completion */
582 dio->private = map_bh->b_private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 }
584 return ret;
585}
586
587/*
588 * There is no bio. Make one now.
589 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700590static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
591 sector_t start_sector, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592{
593 sector_t sector;
594 int ret, nr_pages;
595
Andi Kleeneb28be22011-08-01 21:38:03 -0700596 ret = dio_bio_reap(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597 if (ret)
598 goto out;
Andi Kleeneb28be22011-08-01 21:38:03 -0700599 sector = start_sector << (sdio->blkbits - 9);
Andi Kleen18772642011-08-01 21:38:07 -0700600 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
David Dillow20d96002011-01-20 14:44:22 -0800601 nr_pages = min(nr_pages, BIO_MAX_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602 BUG_ON(nr_pages <= 0);
Andi Kleen18772642011-08-01 21:38:07 -0700603 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
Andi Kleeneb28be22011-08-01 21:38:03 -0700604 sdio->boundary = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605out:
606 return ret;
607}
608
609/*
610 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
611 * that was successful then update final_block_in_bio and take a ref against
612 * the just-added page.
613 *
614 * Return zero on success. Non-zero means the caller needs to start a new BIO.
615 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700616static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617{
618 int ret;
619
Andi Kleeneb28be22011-08-01 21:38:03 -0700620 ret = bio_add_page(sdio->bio, sdio->cur_page,
621 sdio->cur_page_len, sdio->cur_page_offset);
622 if (ret == sdio->cur_page_len) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623 /*
624 * Decrement count only, if we are done with this page
625 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700626 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
627 sdio->pages_in_io--;
628 page_cache_get(sdio->cur_page);
629 sdio->final_block_in_bio = sdio->cur_page_block +
630 (sdio->cur_page_len >> sdio->blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631 ret = 0;
632 } else {
633 ret = 1;
634 }
635 return ret;
636}
637
638/*
639 * Put cur_page under IO. The section of cur_page which is described by
640 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
641 * starts on-disk at cur_page_block.
642 *
643 * We take a ref against the page here (on behalf of its presence in the bio).
644 *
645 * The caller of this function is responsible for removing cur_page from the
646 * dio, and for dropping the refcount which came from that presence.
647 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700648static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
649 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650{
651 int ret = 0;
652
Andi Kleeneb28be22011-08-01 21:38:03 -0700653 if (sdio->bio) {
654 loff_t cur_offset = sdio->cur_page_fs_offset;
655 loff_t bio_next_offset = sdio->logical_offset_in_bio +
656 sdio->bio->bi_size;
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400657
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400659 * See whether this new request is contiguous with the old.
660 *
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900661 * Btrfs cannot handle having logically non-contiguous requests
662 * submitted. For example if you have
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400663 *
664 * Logical: [0-4095][HOLE][8192-12287]
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900665 * Physical: [0-4095] [4096-8191]
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400666 *
667 * We cannot submit those pages together as one BIO. So if our
668 * current logical offset in the file does not equal what would
669 * be the next logical offset in the bio, submit the bio we
670 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700672 if (sdio->final_block_in_bio != sdio->cur_page_block ||
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400673 cur_offset != bio_next_offset)
Andi Kleeneb28be22011-08-01 21:38:03 -0700674 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 /*
676 * Submit now if the underlying fs is about to perform a
677 * metadata read
678 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700679 else if (sdio->boundary)
680 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681 }
682
Andi Kleeneb28be22011-08-01 21:38:03 -0700683 if (sdio->bio == NULL) {
Andi Kleen18772642011-08-01 21:38:07 -0700684 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685 if (ret)
686 goto out;
687 }
688
Andi Kleeneb28be22011-08-01 21:38:03 -0700689 if (dio_bio_add_page(sdio) != 0) {
690 dio_bio_submit(dio, sdio);
Andi Kleen18772642011-08-01 21:38:07 -0700691 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700693 ret = dio_bio_add_page(sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 BUG_ON(ret != 0);
695 }
696 }
697out:
698 return ret;
699}
700
701/*
702 * An autonomous function to put a chunk of a page under deferred IO.
703 *
704 * The caller doesn't actually know (or care) whether this piece of page is in
705 * a BIO, or is under IO or whatever. We just take care of all possible
706 * situations here. The separation between the logic of do_direct_IO() and
707 * that of submit_page_section() is important for clarity. Please don't break.
708 *
709 * The chunk of page starts on-disk at blocknr.
710 *
711 * We perform deferred IO, by recording the last-submitted page inside our
712 * private part of the dio structure. If possible, we just expand the IO
713 * across that page here.
714 *
715 * If that doesn't work out then we put the old page into the bio and add this
716 * page to the dio instead.
717 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700718static inline int
Andi Kleeneb28be22011-08-01 21:38:03 -0700719submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
Andi Kleen18772642011-08-01 21:38:07 -0700720 unsigned offset, unsigned len, sector_t blocknr,
721 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722{
723 int ret = 0;
724
Andrew Morton98c4d572006-12-10 02:19:47 -0800725 if (dio->rw & WRITE) {
726 /*
727 * Read accounting is performed in submit_bio()
728 */
729 task_io_account_write(len);
730 }
731
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732 /*
733 * Can we just grow the current page's presence in the dio?
734 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700735 if (sdio->cur_page == page &&
736 sdio->cur_page_offset + sdio->cur_page_len == offset &&
737 sdio->cur_page_block +
738 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
739 sdio->cur_page_len += len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740
741 /*
Andi Kleeneb28be22011-08-01 21:38:03 -0700742 * If sdio->boundary then we want to schedule the IO now to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743 * avoid metadata seeks.
744 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700745 if (sdio->boundary) {
Andi Kleen18772642011-08-01 21:38:07 -0700746 ret = dio_send_cur_page(dio, sdio, map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -0700747 page_cache_release(sdio->cur_page);
748 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749 }
750 goto out;
751 }
752
753 /*
754 * If there's a deferred page already there then send it.
755 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700756 if (sdio->cur_page) {
Andi Kleen18772642011-08-01 21:38:07 -0700757 ret = dio_send_cur_page(dio, sdio, map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -0700758 page_cache_release(sdio->cur_page);
759 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760 if (ret)
761 goto out;
762 }
763
764 page_cache_get(page); /* It is in dio */
Andi Kleeneb28be22011-08-01 21:38:03 -0700765 sdio->cur_page = page;
766 sdio->cur_page_offset = offset;
767 sdio->cur_page_len = len;
768 sdio->cur_page_block = blocknr;
769 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770out:
771 return ret;
772}
773
774/*
775 * Clean any dirty buffers in the blockdev mapping which alias newly-created
776 * file blocks. Only called for S_ISREG files - blockdevs do not set
777 * buffer_new
778 */
Andi Kleen18772642011-08-01 21:38:07 -0700779static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780{
781 unsigned i;
782 unsigned nblocks;
783
Andi Kleen18772642011-08-01 21:38:07 -0700784 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785
786 for (i = 0; i < nblocks; i++) {
Andi Kleen18772642011-08-01 21:38:07 -0700787 unmap_underlying_metadata(map_bh->b_bdev,
788 map_bh->b_blocknr + i);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 }
790}
791
792/*
793 * If we are not writing the entire block and get_block() allocated
794 * the block for us, we need to fill-in the unused portion of the
795 * block with zeros. This happens only if user-buffer, fileoffset or
796 * io length is not filesystem block-size multiple.
797 *
798 * `end' is zero if we're doing the start of the IO, 1 at the end of the
799 * IO.
800 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700801static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
802 int end, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803{
804 unsigned dio_blocks_per_fs_block;
805 unsigned this_chunk_blocks; /* In dio_blocks */
806 unsigned this_chunk_bytes;
807 struct page *page;
808
Andi Kleeneb28be22011-08-01 21:38:03 -0700809 sdio->start_zero_done = 1;
Andi Kleen18772642011-08-01 21:38:07 -0700810 if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 return;
812
Andi Kleeneb28be22011-08-01 21:38:03 -0700813 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
814 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815
816 if (!this_chunk_blocks)
817 return;
818
819 /*
820 * We need to zero out part of an fs block. It is either at the
821 * beginning or the end of the fs block.
822 */
823 if (end)
824 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
825
Andi Kleeneb28be22011-08-01 21:38:03 -0700826 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827
Nick Piggin557ed1f2007-10-16 01:24:40 -0700828 page = ZERO_PAGE(0);
Andi Kleeneb28be22011-08-01 21:38:03 -0700829 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700830 sdio->next_block_for_io, map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 return;
832
Andi Kleeneb28be22011-08-01 21:38:03 -0700833 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834}
835
836/*
837 * Walk the user pages, and the file, mapping blocks to disk and generating
838 * a sequence of (page,offset,len,block) mappings. These mappings are injected
839 * into submit_page_section(), which takes care of the next stage of submission
840 *
841 * Direct IO against a blockdev is different from a file. Because we can
842 * happily perform page-sized but 512-byte aligned IOs. It is important that
843 * blockdev IO be able to have fine alignment and large sizes.
844 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800845 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846 * with the size of IO which is permitted at this offset and this i_blkbits.
847 *
848 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800849 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850 * fine alignment but still allows this function to work in PAGE_SIZE units.
851 */
Andi Kleen18772642011-08-01 21:38:07 -0700852static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
853 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854{
Andi Kleeneb28be22011-08-01 21:38:03 -0700855 const unsigned blkbits = sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
857 struct page *page;
858 unsigned block_in_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 int ret = 0;
860
861 /* The I/O can start at any block offset within the first page */
Andi Kleeneb28be22011-08-01 21:38:03 -0700862 block_in_page = sdio->first_block_in_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863
Andi Kleeneb28be22011-08-01 21:38:03 -0700864 while (sdio->block_in_file < sdio->final_block_in_request) {
865 page = dio_get_page(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866 if (IS_ERR(page)) {
867 ret = PTR_ERR(page);
868 goto out;
869 }
870
871 while (block_in_page < blocks_per_page) {
872 unsigned offset_in_page = block_in_page << blkbits;
873 unsigned this_chunk_bytes; /* # of bytes mapped */
874 unsigned this_chunk_blocks; /* # of blocks */
875 unsigned u;
876
Andi Kleeneb28be22011-08-01 21:38:03 -0700877 if (sdio->blocks_available == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878 /*
879 * Need to go and map some more disk
880 */
881 unsigned long blkmask;
882 unsigned long dio_remainder;
883
Andi Kleen18772642011-08-01 21:38:07 -0700884 ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 if (ret) {
886 page_cache_release(page);
887 goto out;
888 }
889 if (!buffer_mapped(map_bh))
890 goto do_holes;
891
Andi Kleeneb28be22011-08-01 21:38:03 -0700892 sdio->blocks_available =
893 map_bh->b_size >> sdio->blkbits;
894 sdio->next_block_for_io =
895 map_bh->b_blocknr << sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 if (buffer_new(map_bh))
Andi Kleen18772642011-08-01 21:38:07 -0700897 clean_blockdev_aliases(dio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898
Andi Kleeneb28be22011-08-01 21:38:03 -0700899 if (!sdio->blkfactor)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 goto do_holes;
901
Andi Kleeneb28be22011-08-01 21:38:03 -0700902 blkmask = (1 << sdio->blkfactor) - 1;
903 dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904
905 /*
906 * If we are at the start of IO and that IO
907 * starts partway into a fs-block,
908 * dio_remainder will be non-zero. If the IO
909 * is a read then we can simply advance the IO
910 * cursor to the first block which is to be
911 * read. But if the IO is a write and the
912 * block was newly allocated we cannot do that;
913 * the start of the fs block must be zeroed out
914 * on-disk
915 */
916 if (!buffer_new(map_bh))
Andi Kleeneb28be22011-08-01 21:38:03 -0700917 sdio->next_block_for_io += dio_remainder;
918 sdio->blocks_available -= dio_remainder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919 }
920do_holes:
921 /* Handle holes */
922 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800923 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924
925 /* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboeb31dc662006-06-13 08:26:10 +0200926 if (dio->rw & WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 page_cache_release(page);
928 return -ENOTBLK;
929 }
930
Jeff Moyer35dc8162006-02-03 03:04:27 -0800931 /*
932 * Be sure to account for a partial block as the
933 * last block in the file
934 */
935 i_size_aligned = ALIGN(i_size_read(dio->inode),
936 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -0700937 if (sdio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -0800938 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 /* We hit eof */
940 page_cache_release(page);
941 goto out;
942 }
Christoph Lametereebd2aa2008-02-04 22:28:29 -0800943 zero_user(page, block_in_page << blkbits,
944 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -0700945 sdio->block_in_file++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 block_in_page++;
947 goto next_block;
948 }
949
950 /*
951 * If we're performing IO which has an alignment which
952 * is finer than the underlying fs, go check to see if
953 * we must zero out the start of this block.
954 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700955 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
Andi Kleen18772642011-08-01 21:38:07 -0700956 dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957
958 /*
959 * Work out, in this_chunk_blocks, how much disk we
960 * can add to this page
961 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700962 this_chunk_blocks = sdio->blocks_available;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 u = (PAGE_SIZE - offset_in_page) >> blkbits;
964 if (this_chunk_blocks > u)
965 this_chunk_blocks = u;
Andi Kleeneb28be22011-08-01 21:38:03 -0700966 u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967 if (this_chunk_blocks > u)
968 this_chunk_blocks = u;
969 this_chunk_bytes = this_chunk_blocks << blkbits;
970 BUG_ON(this_chunk_bytes == 0);
971
Andi Kleeneb28be22011-08-01 21:38:03 -0700972 sdio->boundary = buffer_boundary(map_bh);
973 ret = submit_page_section(dio, sdio, page,
974 offset_in_page,
975 this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700976 sdio->next_block_for_io,
977 map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978 if (ret) {
979 page_cache_release(page);
980 goto out;
981 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700982 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983
Andi Kleeneb28be22011-08-01 21:38:03 -0700984 sdio->block_in_file += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985 block_in_page += this_chunk_blocks;
Andi Kleeneb28be22011-08-01 21:38:03 -0700986 sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987next_block:
Andi Kleeneb28be22011-08-01 21:38:03 -0700988 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
989 if (sdio->block_in_file == sdio->final_block_in_request)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990 break;
991 }
992
993 /* Drop the ref which was taken in get_user_pages() */
994 page_cache_release(page);
995 block_in_page = 0;
996 }
997out:
998 return ret;
999}
1000
Andi Kleen847cc632011-08-01 21:38:09 -07001001static inline int drop_refcount(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002{
Andi Kleen847cc632011-08-01 21:38:09 -07001003 int ret2;
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001004 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005
Zach Brown8459d862006-12-10 02:21:05 -08001006 /*
1007 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001008 * operation. AIO can if it was a broken operation described above or
1009 * in fact if all the bios race to complete before we get here. In
1010 * that case dio_complete() translates the EIOCBQUEUED into the proper
1011 * return code that the caller will hand to aio_complete().
1012 *
1013 * This is managed by the bio_lock instead of being an atomic_t so that
1014 * completion paths can drop their ref and use the remaining count to
1015 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001016 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001017 spin_lock_irqsave(&dio->bio_lock, flags);
1018 ret2 = --dio->refcount;
1019 spin_unlock_irqrestore(&dio->bio_lock, flags);
Andi Kleen847cc632011-08-01 21:38:09 -07001020 return ret2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021}
1022
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001023/*
1024 * This is a library function for use by filesystem drivers.
1025 *
1026 * The locking rules are governed by the flags parameter:
1027 * - if the flags value contains DIO_LOCKING we use a fancy locking
1028 * scheme for dumb filesystems.
1029 * For writes this function is called under i_mutex and returns with
1030 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1031 * taken and dropped again before returning.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001032 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1033 * internal locking but rather rely on the filesystem to synchronize
1034 * direct I/O reads/writes versus each other and truncate.
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001035 *
1036 * To help with locking against truncate we incremented the i_dio_count
1037 * counter before starting direct I/O, and decrement it once we are done.
1038 * Truncate can wait for it to reach zero to provide exclusion. It is
1039 * expected that filesystem provide exclusion between new direct I/O
1040 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1041 * but other filesystems need to take care of this on their own.
Andi Kleenba253fb2011-08-01 21:38:08 -07001042 *
1043 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1044 * is always inlined. Otherwise gcc is unable to split the structure into
1045 * individual fields and will generate much worse code. This is important
1046 * for the whole file.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001047 */
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001048static inline ssize_t
1049do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050 struct block_device *bdev, const struct iovec *iov, loff_t offset,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -08001051 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
Josef Bacikfacd07b2010-05-23 11:00:55 -04001052 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053{
1054 int seg;
1055 size_t size;
1056 unsigned long addr;
Linus Torvaldsab738572012-11-29 12:27:00 -08001057 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1058 unsigned blkbits = i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059 unsigned blocksize_mask = (1 << blkbits) - 1;
1060 ssize_t retval = -EINVAL;
1061 loff_t end = offset;
1062 struct dio *dio;
Andi Kleeneb28be22011-08-01 21:38:03 -07001063 struct dio_submit sdio = { 0, };
Andi Kleen847cc632011-08-01 21:38:09 -07001064 unsigned long user_addr;
1065 size_t bytes;
1066 struct buffer_head map_bh = { 0, };
Fengguang Wu647d1e42012-08-09 15:23:09 +02001067 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068
1069 if (rw & WRITE)
Jens Axboe721a9602011-03-09 11:56:30 +01001070 rw = WRITE_ODIRECT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001072 /*
1073 * Avoid references to bdev if not absolutely needed to give
1074 * the early prefetch in the caller enough time.
1075 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076
1077 if (offset & blocksize_mask) {
1078 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001079 blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080 blocksize_mask = (1 << blkbits) - 1;
1081 if (offset & blocksize_mask)
1082 goto out;
1083 }
1084
1085 /* Check the memory alignment. Blocks cannot straddle pages */
1086 for (seg = 0; seg < nr_segs; seg++) {
1087 addr = (unsigned long)iov[seg].iov_base;
1088 size = iov[seg].iov_len;
1089 end += size;
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001090 if (unlikely((addr & blocksize_mask) ||
1091 (size & blocksize_mask))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001093 blkbits = blksize_bits(
1094 bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 blocksize_mask = (1 << blkbits) - 1;
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001096 if ((addr & blocksize_mask) || (size & blocksize_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 goto out;
1098 }
1099 }
1100
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001101 /* watch out for a 0 len io from a tricksy fs */
1102 if (rw == READ && end == offset)
1103 return 0;
1104
Andi Kleen6e8267f2011-08-01 21:38:06 -07001105 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106 retval = -ENOMEM;
1107 if (!dio)
1108 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001109 /*
1110 * Believe it or not, zeroing out the page array caused a .5%
1111 * performance regression in a database benchmark. So, we take
1112 * care to only zero out what's needed.
1113 */
1114 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001116 dio->flags = flags;
1117 if (dio->flags & DIO_LOCKING) {
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001118 if (rw == READ) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001119 struct address_space *mapping =
1120 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001122 /* will be released by direct_io_worker */
1123 mutex_lock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124
1125 retval = filemap_write_and_wait_range(mapping, offset,
1126 end - 1);
1127 if (retval) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001128 mutex_unlock(&inode->i_mutex);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001129 kmem_cache_free(dio_cache, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 goto out;
1131 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 }
1134
1135 /*
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001136 * Will be decremented at I/O completion time.
1137 */
1138 atomic_inc(&inode->i_dio_count);
1139
1140 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 * For file extending writes updating i_size before data
1142 * writeouts complete can expose uninitialized blocks. So
1143 * even for AIO, we need to wait for i/o to complete before
1144 * returning in this case.
1145 */
Jens Axboeb31dc662006-06-13 08:26:10 +02001146 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 (end > i_size_read(inode)));
1148
Andi Kleen847cc632011-08-01 21:38:09 -07001149 retval = 0;
1150
1151 dio->inode = inode;
1152 dio->rw = rw;
1153 sdio.blkbits = blkbits;
Linus Torvaldsab738572012-11-29 12:27:00 -08001154 sdio.blkfactor = i_blkbits - blkbits;
Andi Kleen847cc632011-08-01 21:38:09 -07001155 sdio.block_in_file = offset >> blkbits;
1156
1157 sdio.get_block = get_block;
1158 dio->end_io = end_io;
1159 sdio.submit_io = submit_io;
1160 sdio.final_block_in_bio = -1;
1161 sdio.next_block_for_io = -1;
1162
1163 dio->iocb = iocb;
1164 dio->i_size = i_size_read(inode);
1165
1166 spin_lock_init(&dio->bio_lock);
1167 dio->refcount = 1;
1168
1169 /*
1170 * In case of non-aligned buffers, we may need 2 more
1171 * pages since we need to zero out first and last block.
1172 */
1173 if (unlikely(sdio.blkfactor))
1174 sdio.pages_in_io = 2;
1175
1176 for (seg = 0; seg < nr_segs; seg++) {
1177 user_addr = (unsigned long)iov[seg].iov_base;
1178 sdio.pages_in_io +=
1179 ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) /
1180 PAGE_SIZE - user_addr / PAGE_SIZE);
1181 }
1182
Fengguang Wu647d1e42012-08-09 15:23:09 +02001183 blk_start_plug(&plug);
1184
Andi Kleen847cc632011-08-01 21:38:09 -07001185 for (seg = 0; seg < nr_segs; seg++) {
1186 user_addr = (unsigned long)iov[seg].iov_base;
1187 sdio.size += bytes = iov[seg].iov_len;
1188
1189 /* Index into the first page of the first block */
1190 sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1191 sdio.final_block_in_request = sdio.block_in_file +
1192 (bytes >> blkbits);
1193 /* Page fetching state */
1194 sdio.head = 0;
1195 sdio.tail = 0;
1196 sdio.curr_page = 0;
1197
1198 sdio.total_pages = 0;
1199 if (user_addr & (PAGE_SIZE-1)) {
1200 sdio.total_pages++;
1201 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1202 }
1203 sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1204 sdio.curr_user_address = user_addr;
1205
1206 retval = do_direct_IO(dio, &sdio, &map_bh);
1207
1208 dio->result += iov[seg].iov_len -
1209 ((sdio.final_block_in_request - sdio.block_in_file) <<
1210 blkbits);
1211
1212 if (retval) {
1213 dio_cleanup(dio, &sdio);
1214 break;
1215 }
1216 } /* end iovec loop */
1217
1218 if (retval == -ENOTBLK) {
1219 /*
1220 * The remaining part of the request will be
1221 * be handled by buffered I/O when we return
1222 */
1223 retval = 0;
1224 }
1225 /*
1226 * There may be some unwritten disk at the end of a part-written
1227 * fs-block-sized block. Go zero that now.
1228 */
1229 dio_zero_block(dio, &sdio, 1, &map_bh);
1230
1231 if (sdio.cur_page) {
1232 ssize_t ret2;
1233
1234 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1235 if (retval == 0)
1236 retval = ret2;
1237 page_cache_release(sdio.cur_page);
1238 sdio.cur_page = NULL;
1239 }
1240 if (sdio.bio)
1241 dio_bio_submit(dio, &sdio);
1242
Fengguang Wu647d1e42012-08-09 15:23:09 +02001243 blk_finish_plug(&plug);
1244
Andi Kleen847cc632011-08-01 21:38:09 -07001245 /*
1246 * It is possible that, we return short IO due to end of file.
1247 * In that case, we need to release all the pages we got hold on.
1248 */
1249 dio_cleanup(dio, &sdio);
1250
1251 /*
1252 * All block lookups have been performed. For READ requests
1253 * we can let i_mutex go now that its achieved its purpose
1254 * of protecting us from looking up uninitialized blocks.
1255 */
1256 if (rw == READ && (dio->flags & DIO_LOCKING))
1257 mutex_unlock(&dio->inode->i_mutex);
1258
1259 /*
1260 * The only time we want to leave bios in flight is when a successful
1261 * partial aio read or full aio write have been setup. In that case
1262 * bio completion will call aio_complete. The only time it's safe to
1263 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1264 * This had *better* be the only place that raises -EIOCBQUEUED.
1265 */
1266 BUG_ON(retval == -EIOCBQUEUED);
1267 if (dio->is_async && retval == 0 && dio->result &&
Julia Lawalld1876632012-06-07 15:45:00 -07001268 ((rw == READ) || (dio->result == sdio.size)))
Andi Kleen847cc632011-08-01 21:38:09 -07001269 retval = -EIOCBQUEUED;
1270
1271 if (retval != -EIOCBQUEUED)
1272 dio_await_completion(dio);
1273
1274 if (drop_refcount(dio) == 0) {
1275 retval = dio_complete(dio, offset, retval, false);
1276 kmem_cache_free(dio_cache, dio);
1277 } else
1278 BUG_ON(retval != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001280out:
1281 return retval;
1282}
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001283
1284ssize_t
1285__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1286 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1287 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1288 dio_submit_t submit_io, int flags)
1289{
1290 /*
1291 * The block device state is needed in the end to finally
1292 * submit everything. Since it's likely to be cache cold
1293 * prefetch it here as first thing to hide some of the
1294 * latency.
1295 *
1296 * Attempt to prefetch the pieces we likely need later.
1297 */
1298 prefetch(&bdev->bd_disk->part_tbl);
1299 prefetch(bdev->bd_queue);
1300 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1301
1302 return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset,
1303 nr_segs, get_block, end_io,
1304 submit_io, flags);
1305}
1306
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307EXPORT_SYMBOL(__blockdev_direct_IO);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001308
1309static __init int dio_init(void)
1310{
1311 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1312 return 0;
1313}
1314module_init(dio_init)