blob: f4aadd15b61376ee97a4a7bababe2eb659850820 [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 Sharma60063492011-07-26 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
Christoph Hellwigbd5fe6c2011-06-24 14:29:43 -0400148static void __inode_dio_wait(struct inode *inode)
149{
150 wait_queue_head_t *wq = bit_waitqueue(&inode->i_state, __I_DIO_WAKEUP);
151 DEFINE_WAIT_BIT(q, &inode->i_state, __I_DIO_WAKEUP);
152
153 do {
154 prepare_to_wait(wq, &q.wait, TASK_UNINTERRUPTIBLE);
155 if (atomic_read(&inode->i_dio_count))
156 schedule();
157 } while (atomic_read(&inode->i_dio_count));
158 finish_wait(wq, &q.wait);
159}
160
161/**
162 * inode_dio_wait - wait for outstanding DIO requests to finish
163 * @inode: inode to wait for
164 *
165 * Waits for all pending direct I/O requests to finish so that we can
166 * proceed with a truncate or equivalent operation.
167 *
168 * Must be called under a lock that serializes taking new references
169 * to i_dio_count, usually by inode->i_mutex.
170 */
171void inode_dio_wait(struct inode *inode)
172{
173 if (atomic_read(&inode->i_dio_count))
174 __inode_dio_wait(inode);
175}
Anton Altaparmakov37fbf4b2012-02-23 23:40:05 +0000176EXPORT_SYMBOL(inode_dio_wait);
Christoph Hellwigbd5fe6c2011-06-24 14:29:43 -0400177
178/*
179 * inode_dio_done - signal finish of a direct I/O requests
180 * @inode: inode the direct I/O happens on
181 *
182 * This is called once we've finished processing a direct I/O request,
183 * and is used to wake up callers waiting for direct I/O to be quiesced.
184 */
185void inode_dio_done(struct inode *inode)
186{
187 if (atomic_dec_and_test(&inode->i_dio_count))
188 wake_up_bit(&inode->i_state, __I_DIO_WAKEUP);
189}
Anton Altaparmakov37fbf4b2012-02-23 23:40:05 +0000190EXPORT_SYMBOL(inode_dio_done);
Christoph Hellwigbd5fe6c2011-06-24 14:29:43 -0400191
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192/*
193 * How many pages are in the queue?
194 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700195static inline unsigned dio_pages_present(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196{
Andi Kleeneb28be22011-08-01 21:38:03 -0700197 return sdio->tail - sdio->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198}
199
200/*
201 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
202 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700203static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204{
205 int ret;
206 int nr_pages;
207
Andi Kleeneb28be22011-08-01 21:38:03 -0700208 nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
Nick Pigginf5dd33c2008-07-25 19:45:25 -0700209 ret = get_user_pages_fast(
Andi Kleeneb28be22011-08-01 21:38:03 -0700210 sdio->curr_user_address, /* Where from? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211 nr_pages, /* How many pages? */
212 dio->rw == READ, /* Write to memory? */
Nick Pigginf5dd33c2008-07-25 19:45:25 -0700213 &dio->pages[0]); /* Put results here */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214
Andi Kleeneb28be22011-08-01 21:38:03 -0700215 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
Nick Piggin557ed1f2007-10-16 01:24:40 -0700216 struct page *page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217 /*
218 * A memory fault, but the filesystem has some outstanding
219 * mapped blocks. We need to use those blocks up to avoid
220 * leaking stale data in the file.
221 */
222 if (dio->page_errors == 0)
223 dio->page_errors = ret;
Nick Pigginb5810032005-10-29 18:16:12 -0700224 page_cache_get(page);
225 dio->pages[0] = page;
Andi Kleeneb28be22011-08-01 21:38:03 -0700226 sdio->head = 0;
227 sdio->tail = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228 ret = 0;
229 goto out;
230 }
231
232 if (ret >= 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700233 sdio->curr_user_address += ret * PAGE_SIZE;
234 sdio->curr_page += ret;
235 sdio->head = 0;
236 sdio->tail = ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 ret = 0;
238 }
239out:
240 return ret;
241}
242
243/*
244 * Get another userspace page. Returns an ERR_PTR on error. Pages are
245 * buffered inside the dio so that we can call get_user_pages() against a
246 * decent number of pages, less frequently. To provide nicer use of the
247 * L1 cache.
248 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700249static inline struct page *dio_get_page(struct dio *dio,
250 struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251{
Andi Kleeneb28be22011-08-01 21:38:03 -0700252 if (dio_pages_present(sdio) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253 int ret;
254
Andi Kleeneb28be22011-08-01 21:38:03 -0700255 ret = dio_refill_pages(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256 if (ret)
257 return ERR_PTR(ret);
Andi Kleeneb28be22011-08-01 21:38:03 -0700258 BUG_ON(dio_pages_present(sdio) == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700260 return dio->pages[sdio->head++];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261}
262
Zach Brown6d544bb2006-12-10 02:20:54 -0800263/**
264 * dio_complete() - called when all DIO BIO I/O has been completed
265 * @offset: the byte offset in the file of the completed operation
266 *
267 * This releases locks as dictated by the locking type, lets interested parties
268 * know that a DIO operation has completed, and calculates the resulting return
269 * code for the operation.
270 *
271 * It lets the filesystem know if it registered an interest earlier via
272 * get_block. Pass the private field of the map buffer_head so that
273 * filesystems can use it to hold additional state between get_block calls and
274 * dio_complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275 */
Edward Shishkincd1c5842010-10-26 14:22:28 -0700276static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277{
Zach Brown6d544bb2006-12-10 02:20:54 -0800278 ssize_t transferred = 0;
279
Zach Brown8459d862006-12-10 02:21:05 -0800280 /*
281 * AIO submission can race with bio completion to get here while
282 * expecting to have the last io completed by bio completion.
283 * In that case -EIOCBQUEUED is in fact not an error we want
284 * to preserve through this call.
285 */
286 if (ret == -EIOCBQUEUED)
287 ret = 0;
288
Zach Brown6d544bb2006-12-10 02:20:54 -0800289 if (dio->result) {
290 transferred = dio->result;
291
292 /* Check for short read case */
293 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
294 transferred = dio->i_size - offset;
295 }
296
Zach Brown6d544bb2006-12-10 02:20:54 -0800297 if (ret == 0)
298 ret = dio->page_errors;
299 if (ret == 0)
300 ret = dio->io_error;
301 if (ret == 0)
302 ret = transferred;
303
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000304 if (dio->end_io && dio->result) {
305 dio->end_io(dio->iocb, offset, transferred,
Andi Kleen18772642011-08-01 21:38:07 -0700306 dio->private, ret, is_async);
Christoph Hellwig72c50522011-06-24 14:29:48 -0400307 } else {
308 if (is_async)
309 aio_complete(dio->iocb, ret, 0);
310 inode_dio_done(dio->inode);
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000311 }
312
Zach Brown6d544bb2006-12-10 02:20:54 -0800313 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314}
315
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316static int dio_bio_complete(struct dio *dio, struct bio *bio);
317/*
318 * Asynchronous IO callback.
319 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200320static void dio_bio_end_aio(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321{
322 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800323 unsigned long remaining;
324 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326 /* cleanup the bio */
327 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800328
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800329 spin_lock_irqsave(&dio->bio_lock, flags);
330 remaining = --dio->refcount;
331 if (remaining == 1 && dio->waiter)
Zach Brown20258b22006-12-10 02:21:01 -0800332 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800333 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b22006-12-10 02:21:01 -0800334
Zach Brown8459d862006-12-10 02:21:05 -0800335 if (remaining == 0) {
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000336 dio_complete(dio, dio->iocb->ki_pos, 0, true);
Andi Kleen6e8267f2011-08-01 21:38:06 -0700337 kmem_cache_free(dio_cache, dio);
Zach Brown8459d862006-12-10 02:21:05 -0800338 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339}
340
341/*
342 * The BIO completion handler simply queues the BIO up for the process-context
343 * handler.
344 *
345 * During I/O bi_private points at the dio. After I/O, bi_private is used to
346 * implement a singly-linked list of completed BIOs, at dio->bio_list.
347 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200348static void dio_bio_end_io(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349{
350 struct dio *dio = bio->bi_private;
351 unsigned long flags;
352
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353 spin_lock_irqsave(&dio->bio_lock, flags);
354 bio->bi_private = dio->bio_list;
355 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800356 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357 wake_up_process(dio->waiter);
358 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359}
360
Josef Bacikfacd07b2010-05-23 11:00:55 -0400361/**
362 * dio_end_io - handle the end io action for the given bio
363 * @bio: The direct io bio thats being completed
364 * @error: Error if there was one
365 *
366 * This is meant to be called by any filesystem that uses their own dio_submit_t
367 * so that the DIO specific endio actions are dealt with after the filesystem
368 * has done it's completion work.
369 */
370void dio_end_io(struct bio *bio, int error)
371{
372 struct dio *dio = bio->bi_private;
373
374 if (dio->is_async)
375 dio_bio_end_aio(bio, error);
376 else
377 dio_bio_end_io(bio, error);
378}
379EXPORT_SYMBOL_GPL(dio_end_io);
380
Andi Kleenba253fb2011-08-01 21:38:08 -0700381static inline void
Andi Kleeneb28be22011-08-01 21:38:03 -0700382dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
383 struct block_device *bdev,
384 sector_t first_sector, int nr_vecs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385{
386 struct bio *bio;
387
David Dillow20d96002011-01-20 14:44:22 -0800388 /*
389 * bio_alloc() is guaranteed to return a bio when called with
390 * __GFP_WAIT and we request a valid number of vectors.
391 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393
394 bio->bi_bdev = bdev;
395 bio->bi_sector = first_sector;
396 if (dio->is_async)
397 bio->bi_end_io = dio_bio_end_aio;
398 else
399 bio->bi_end_io = dio_bio_end_io;
400
Andi Kleeneb28be22011-08-01 21:38:03 -0700401 sdio->bio = bio;
402 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403}
404
405/*
406 * In the AIO read case we speculatively dirty the pages before starting IO.
407 * During IO completion, any of these pages which happen to have been written
408 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800409 *
410 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700412static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413{
Andi Kleeneb28be22011-08-01 21:38:03 -0700414 struct bio *bio = sdio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800415 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416
417 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800418
419 spin_lock_irqsave(&dio->bio_lock, flags);
420 dio->refcount++;
421 spin_unlock_irqrestore(&dio->bio_lock, flags);
422
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 if (dio->is_async && dio->rw == READ)
424 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800425
Andi Kleeneb28be22011-08-01 21:38:03 -0700426 if (sdio->submit_io)
427 sdio->submit_io(dio->rw, bio, dio->inode,
428 sdio->logical_offset_in_bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400429 else
430 submit_bio(dio->rw, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431
Andi Kleeneb28be22011-08-01 21:38:03 -0700432 sdio->bio = NULL;
433 sdio->boundary = 0;
434 sdio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435}
436
437/*
438 * Release any resources in case of a failure
439 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700440static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441{
Andi Kleeneb28be22011-08-01 21:38:03 -0700442 while (dio_pages_present(sdio))
443 page_cache_release(dio_get_page(dio, sdio));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444}
445
446/*
Zach Brown02732012006-12-10 02:20:59 -0800447 * Wait for the next BIO to complete. Remove it and return it. NULL is
448 * returned once all BIOs have been completed. This must only be called once
449 * all bios have been issued so that dio->refcount can only decrease. This
450 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451 */
452static struct bio *dio_await_one(struct dio *dio)
453{
454 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800455 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456
457 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800458
459 /*
460 * Wait as long as the list is empty and there are bios in flight. bio
461 * completion drops the count, maybe adds to the list, and wakes while
462 * holding the bio_lock so we don't need set_current_state()'s barrier
463 * and can call it after testing our condition.
464 */
465 while (dio->refcount > 1 && dio->bio_list == NULL) {
466 __set_current_state(TASK_UNINTERRUPTIBLE);
467 dio->waiter = current;
468 spin_unlock_irqrestore(&dio->bio_lock, flags);
469 io_schedule();
470 /* wake up sets us TASK_RUNNING */
471 spin_lock_irqsave(&dio->bio_lock, flags);
472 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 }
Zach Brown02732012006-12-10 02:20:59 -0800474 if (dio->bio_list) {
475 bio = dio->bio_list;
476 dio->bio_list = bio->bi_private;
477 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 spin_unlock_irqrestore(&dio->bio_lock, flags);
479 return bio;
480}
481
482/*
483 * Process one completed BIO. No locks are held.
484 */
485static int dio_bio_complete(struct dio *dio, struct bio *bio)
486{
487 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
488 struct bio_vec *bvec = bio->bi_io_vec;
489 int page_no;
490
491 if (!uptodate)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800492 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493
494 if (dio->is_async && dio->rw == READ) {
495 bio_check_pages_dirty(bio); /* transfers ownership */
496 } else {
497 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
498 struct page *page = bvec[page_no].bv_page;
499
500 if (dio->rw == READ && !PageCompound(page))
501 set_page_dirty_lock(page);
502 page_cache_release(page);
503 }
504 bio_put(bio);
505 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 return uptodate ? 0 : -EIO;
507}
508
509/*
Zach Brown02732012006-12-10 02:20:59 -0800510 * Wait on and process all in-flight BIOs. This must only be called once
511 * all bios have been issued so that the refcount can only decrease.
512 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200513 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800514 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800516static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517{
Zach Brown02732012006-12-10 02:20:59 -0800518 struct bio *bio;
519 do {
520 bio = dio_await_one(dio);
521 if (bio)
522 dio_bio_complete(dio, bio);
523 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524}
525
526/*
527 * A really large O_DIRECT read or write can generate a lot of BIOs. So
528 * to keep the memory consumption sane we periodically reap any completed BIOs
529 * during the BIO generation phase.
530 *
531 * This also helps to limit the peak amount of pinned userspace memory.
532 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700533static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534{
535 int ret = 0;
536
Andi Kleeneb28be22011-08-01 21:38:03 -0700537 if (sdio->reap_counter++ >= 64) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 while (dio->bio_list) {
539 unsigned long flags;
540 struct bio *bio;
541 int ret2;
542
543 spin_lock_irqsave(&dio->bio_lock, flags);
544 bio = dio->bio_list;
545 dio->bio_list = bio->bi_private;
546 spin_unlock_irqrestore(&dio->bio_lock, flags);
547 ret2 = dio_bio_complete(dio, bio);
548 if (ret == 0)
549 ret = ret2;
550 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700551 sdio->reap_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 }
553 return ret;
554}
555
556/*
557 * Call into the fs to map some more disk blocks. We record the current number
Andi Kleeneb28be22011-08-01 21:38:03 -0700558 * of available blocks at sdio->blocks_available. These are in units of the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 * fs blocksize, (1 << inode->i_blkbits).
560 *
561 * The fs is allowed to map lots of blocks at once. If it wants to do that,
562 * it uses the passed inode-relative block number as the file offset, as usual.
563 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800564 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565 * has remaining to do. The fs should not map more than this number of blocks.
566 *
567 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
568 * indicate how much contiguous disk space has been made available at
569 * bh->b_blocknr.
570 *
571 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
572 * This isn't very efficient...
573 *
574 * In the case of filesystem holes: the fs may return an arbitrarily-large
575 * hole by returning an appropriate value in b_size and by clearing
576 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800577 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578 */
Andi Kleen18772642011-08-01 21:38:07 -0700579static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
580 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581{
582 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
Tao Maae55e1a2012-01-12 17:20:33 -0800584 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 unsigned long fs_count; /* Number of filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586 int create;
587
588 /*
589 * If there was a memory error and we've overwritten all the
590 * mapped blocks then we can now return that memory error
591 */
592 ret = dio->page_errors;
593 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700594 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
595 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
Tao Maae55e1a2012-01-12 17:20:33 -0800596 fs_endblk = (sdio->final_block_in_request - 1) >>
597 sdio->blkfactor;
598 fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599
Nathan Scott3c674e72006-03-29 09:26:15 +1000600 map_bh->b_state = 0;
601 map_bh->b_size = fs_count << dio->inode->i_blkbits;
602
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800603 /*
604 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
605 * forbid block creations: only overwrites are permitted.
606 * We will return early to the caller once we see an
607 * unmapped buffer head returned, and the caller will fall
608 * back to buffered I/O.
609 *
610 * Otherwise the decision is left to the get_blocks method,
611 * which may decide to handle it or also return an unmapped
612 * buffer head.
613 */
Jens Axboeb31dc662006-06-13 08:26:10 +0200614 create = dio->rw & WRITE;
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800615 if (dio->flags & DIO_SKIP_HOLES) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700616 if (sdio->block_in_file < (i_size_read(dio->inode) >>
617 sdio->blkbits))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000620
Andi Kleeneb28be22011-08-01 21:38:03 -0700621 ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 map_bh, create);
Andi Kleen18772642011-08-01 21:38:07 -0700623
624 /* Store for completion */
625 dio->private = map_bh->b_private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626 }
627 return ret;
628}
629
630/*
631 * There is no bio. Make one now.
632 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700633static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
634 sector_t start_sector, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635{
636 sector_t sector;
637 int ret, nr_pages;
638
Andi Kleeneb28be22011-08-01 21:38:03 -0700639 ret = dio_bio_reap(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640 if (ret)
641 goto out;
Andi Kleeneb28be22011-08-01 21:38:03 -0700642 sector = start_sector << (sdio->blkbits - 9);
Andi Kleen18772642011-08-01 21:38:07 -0700643 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
David Dillow20d96002011-01-20 14:44:22 -0800644 nr_pages = min(nr_pages, BIO_MAX_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 BUG_ON(nr_pages <= 0);
Andi Kleen18772642011-08-01 21:38:07 -0700646 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
Andi Kleeneb28be22011-08-01 21:38:03 -0700647 sdio->boundary = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648out:
649 return ret;
650}
651
652/*
653 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
654 * that was successful then update final_block_in_bio and take a ref against
655 * the just-added page.
656 *
657 * Return zero on success. Non-zero means the caller needs to start a new BIO.
658 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700659static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660{
661 int ret;
662
Andi Kleeneb28be22011-08-01 21:38:03 -0700663 ret = bio_add_page(sdio->bio, sdio->cur_page,
664 sdio->cur_page_len, sdio->cur_page_offset);
665 if (ret == sdio->cur_page_len) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666 /*
667 * Decrement count only, if we are done with this page
668 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700669 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
670 sdio->pages_in_io--;
671 page_cache_get(sdio->cur_page);
672 sdio->final_block_in_bio = sdio->cur_page_block +
673 (sdio->cur_page_len >> sdio->blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674 ret = 0;
675 } else {
676 ret = 1;
677 }
678 return ret;
679}
680
681/*
682 * Put cur_page under IO. The section of cur_page which is described by
683 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
684 * starts on-disk at cur_page_block.
685 *
686 * We take a ref against the page here (on behalf of its presence in the bio).
687 *
688 * The caller of this function is responsible for removing cur_page from the
689 * dio, and for dropping the refcount which came from that presence.
690 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700691static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
692 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693{
694 int ret = 0;
695
Andi Kleeneb28be22011-08-01 21:38:03 -0700696 if (sdio->bio) {
697 loff_t cur_offset = sdio->cur_page_fs_offset;
698 loff_t bio_next_offset = sdio->logical_offset_in_bio +
699 sdio->bio->bi_size;
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400700
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400702 * See whether this new request is contiguous with the old.
703 *
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900704 * Btrfs cannot handle having logically non-contiguous requests
705 * submitted. For example if you have
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400706 *
707 * Logical: [0-4095][HOLE][8192-12287]
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900708 * Physical: [0-4095] [4096-8191]
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400709 *
710 * We cannot submit those pages together as one BIO. So if our
711 * current logical offset in the file does not equal what would
712 * be the next logical offset in the bio, submit the bio we
713 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700715 if (sdio->final_block_in_bio != sdio->cur_page_block ||
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400716 cur_offset != bio_next_offset)
Andi Kleeneb28be22011-08-01 21:38:03 -0700717 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718 /*
719 * Submit now if the underlying fs is about to perform a
720 * metadata read
721 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700722 else if (sdio->boundary)
723 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724 }
725
Andi Kleeneb28be22011-08-01 21:38:03 -0700726 if (sdio->bio == NULL) {
Andi Kleen18772642011-08-01 21:38:07 -0700727 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728 if (ret)
729 goto out;
730 }
731
Andi Kleeneb28be22011-08-01 21:38:03 -0700732 if (dio_bio_add_page(sdio) != 0) {
733 dio_bio_submit(dio, sdio);
Andi Kleen18772642011-08-01 21:38:07 -0700734 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700736 ret = dio_bio_add_page(sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 BUG_ON(ret != 0);
738 }
739 }
740out:
741 return ret;
742}
743
744/*
745 * An autonomous function to put a chunk of a page under deferred IO.
746 *
747 * The caller doesn't actually know (or care) whether this piece of page is in
748 * a BIO, or is under IO or whatever. We just take care of all possible
749 * situations here. The separation between the logic of do_direct_IO() and
750 * that of submit_page_section() is important for clarity. Please don't break.
751 *
752 * The chunk of page starts on-disk at blocknr.
753 *
754 * We perform deferred IO, by recording the last-submitted page inside our
755 * private part of the dio structure. If possible, we just expand the IO
756 * across that page here.
757 *
758 * If that doesn't work out then we put the old page into the bio and add this
759 * page to the dio instead.
760 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700761static inline int
Andi Kleeneb28be22011-08-01 21:38:03 -0700762submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
Andi Kleen18772642011-08-01 21:38:07 -0700763 unsigned offset, unsigned len, sector_t blocknr,
764 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765{
766 int ret = 0;
767
Andrew Morton98c4d572006-12-10 02:19:47 -0800768 if (dio->rw & WRITE) {
769 /*
770 * Read accounting is performed in submit_bio()
771 */
772 task_io_account_write(len);
773 }
774
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775 /*
776 * Can we just grow the current page's presence in the dio?
777 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700778 if (sdio->cur_page == page &&
779 sdio->cur_page_offset + sdio->cur_page_len == offset &&
780 sdio->cur_page_block +
781 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
782 sdio->cur_page_len += len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783
784 /*
Andi Kleeneb28be22011-08-01 21:38:03 -0700785 * If sdio->boundary then we want to schedule the IO now to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786 * avoid metadata seeks.
787 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700788 if (sdio->boundary) {
Andi Kleen18772642011-08-01 21:38:07 -0700789 ret = dio_send_cur_page(dio, sdio, map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -0700790 page_cache_release(sdio->cur_page);
791 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 }
793 goto out;
794 }
795
796 /*
797 * If there's a deferred page already there then send it.
798 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700799 if (sdio->cur_page) {
Andi Kleen18772642011-08-01 21:38:07 -0700800 ret = dio_send_cur_page(dio, sdio, map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -0700801 page_cache_release(sdio->cur_page);
802 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 if (ret)
804 goto out;
805 }
806
807 page_cache_get(page); /* It is in dio */
Andi Kleeneb28be22011-08-01 21:38:03 -0700808 sdio->cur_page = page;
809 sdio->cur_page_offset = offset;
810 sdio->cur_page_len = len;
811 sdio->cur_page_block = blocknr;
812 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813out:
814 return ret;
815}
816
817/*
818 * Clean any dirty buffers in the blockdev mapping which alias newly-created
819 * file blocks. Only called for S_ISREG files - blockdevs do not set
820 * buffer_new
821 */
Andi Kleen18772642011-08-01 21:38:07 -0700822static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823{
824 unsigned i;
825 unsigned nblocks;
826
Andi Kleen18772642011-08-01 21:38:07 -0700827 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828
829 for (i = 0; i < nblocks; i++) {
Andi Kleen18772642011-08-01 21:38:07 -0700830 unmap_underlying_metadata(map_bh->b_bdev,
831 map_bh->b_blocknr + i);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832 }
833}
834
835/*
836 * If we are not writing the entire block and get_block() allocated
837 * the block for us, we need to fill-in the unused portion of the
838 * block with zeros. This happens only if user-buffer, fileoffset or
839 * io length is not filesystem block-size multiple.
840 *
841 * `end' is zero if we're doing the start of the IO, 1 at the end of the
842 * IO.
843 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700844static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
845 int end, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846{
847 unsigned dio_blocks_per_fs_block;
848 unsigned this_chunk_blocks; /* In dio_blocks */
849 unsigned this_chunk_bytes;
850 struct page *page;
851
Andi Kleeneb28be22011-08-01 21:38:03 -0700852 sdio->start_zero_done = 1;
Andi Kleen18772642011-08-01 21:38:07 -0700853 if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 return;
855
Andi Kleeneb28be22011-08-01 21:38:03 -0700856 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
857 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858
859 if (!this_chunk_blocks)
860 return;
861
862 /*
863 * We need to zero out part of an fs block. It is either at the
864 * beginning or the end of the fs block.
865 */
866 if (end)
867 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
868
Andi Kleeneb28be22011-08-01 21:38:03 -0700869 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870
Nick Piggin557ed1f2007-10-16 01:24:40 -0700871 page = ZERO_PAGE(0);
Andi Kleeneb28be22011-08-01 21:38:03 -0700872 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700873 sdio->next_block_for_io, map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874 return;
875
Andi Kleeneb28be22011-08-01 21:38:03 -0700876 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877}
878
879/*
880 * Walk the user pages, and the file, mapping blocks to disk and generating
881 * a sequence of (page,offset,len,block) mappings. These mappings are injected
882 * into submit_page_section(), which takes care of the next stage of submission
883 *
884 * Direct IO against a blockdev is different from a file. Because we can
885 * happily perform page-sized but 512-byte aligned IOs. It is important that
886 * blockdev IO be able to have fine alignment and large sizes.
887 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800888 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 * with the size of IO which is permitted at this offset and this i_blkbits.
890 *
891 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800892 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 * fine alignment but still allows this function to work in PAGE_SIZE units.
894 */
Andi Kleen18772642011-08-01 21:38:07 -0700895static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
896 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897{
Andi Kleeneb28be22011-08-01 21:38:03 -0700898 const unsigned blkbits = sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
900 struct page *page;
901 unsigned block_in_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 int ret = 0;
903
904 /* The I/O can start at any block offset within the first page */
Andi Kleeneb28be22011-08-01 21:38:03 -0700905 block_in_page = sdio->first_block_in_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906
Andi Kleeneb28be22011-08-01 21:38:03 -0700907 while (sdio->block_in_file < sdio->final_block_in_request) {
908 page = dio_get_page(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 if (IS_ERR(page)) {
910 ret = PTR_ERR(page);
911 goto out;
912 }
913
914 while (block_in_page < blocks_per_page) {
915 unsigned offset_in_page = block_in_page << blkbits;
916 unsigned this_chunk_bytes; /* # of bytes mapped */
917 unsigned this_chunk_blocks; /* # of blocks */
918 unsigned u;
919
Andi Kleeneb28be22011-08-01 21:38:03 -0700920 if (sdio->blocks_available == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 /*
922 * Need to go and map some more disk
923 */
924 unsigned long blkmask;
925 unsigned long dio_remainder;
926
Andi Kleen18772642011-08-01 21:38:07 -0700927 ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 if (ret) {
929 page_cache_release(page);
930 goto out;
931 }
932 if (!buffer_mapped(map_bh))
933 goto do_holes;
934
Andi Kleeneb28be22011-08-01 21:38:03 -0700935 sdio->blocks_available =
936 map_bh->b_size >> sdio->blkbits;
937 sdio->next_block_for_io =
938 map_bh->b_blocknr << sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 if (buffer_new(map_bh))
Andi Kleen18772642011-08-01 21:38:07 -0700940 clean_blockdev_aliases(dio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941
Andi Kleeneb28be22011-08-01 21:38:03 -0700942 if (!sdio->blkfactor)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 goto do_holes;
944
Andi Kleeneb28be22011-08-01 21:38:03 -0700945 blkmask = (1 << sdio->blkfactor) - 1;
946 dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947
948 /*
949 * If we are at the start of IO and that IO
950 * starts partway into a fs-block,
951 * dio_remainder will be non-zero. If the IO
952 * is a read then we can simply advance the IO
953 * cursor to the first block which is to be
954 * read. But if the IO is a write and the
955 * block was newly allocated we cannot do that;
956 * the start of the fs block must be zeroed out
957 * on-disk
958 */
959 if (!buffer_new(map_bh))
Andi Kleeneb28be22011-08-01 21:38:03 -0700960 sdio->next_block_for_io += dio_remainder;
961 sdio->blocks_available -= dio_remainder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 }
963do_holes:
964 /* Handle holes */
965 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800966 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967
968 /* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboeb31dc662006-06-13 08:26:10 +0200969 if (dio->rw & WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970 page_cache_release(page);
971 return -ENOTBLK;
972 }
973
Jeff Moyer35dc8162006-02-03 03:04:27 -0800974 /*
975 * Be sure to account for a partial block as the
976 * last block in the file
977 */
978 i_size_aligned = ALIGN(i_size_read(dio->inode),
979 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -0700980 if (sdio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -0800981 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982 /* We hit eof */
983 page_cache_release(page);
984 goto out;
985 }
Christoph Lametereebd2aa2008-02-04 22:28:29 -0800986 zero_user(page, block_in_page << blkbits,
987 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -0700988 sdio->block_in_file++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989 block_in_page++;
990 goto next_block;
991 }
992
993 /*
994 * If we're performing IO which has an alignment which
995 * is finer than the underlying fs, go check to see if
996 * we must zero out the start of this block.
997 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700998 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
Andi Kleen18772642011-08-01 21:38:07 -0700999 dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000
1001 /*
1002 * Work out, in this_chunk_blocks, how much disk we
1003 * can add to this page
1004 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001005 this_chunk_blocks = sdio->blocks_available;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006 u = (PAGE_SIZE - offset_in_page) >> blkbits;
1007 if (this_chunk_blocks > u)
1008 this_chunk_blocks = u;
Andi Kleeneb28be22011-08-01 21:38:03 -07001009 u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010 if (this_chunk_blocks > u)
1011 this_chunk_blocks = u;
1012 this_chunk_bytes = this_chunk_blocks << blkbits;
1013 BUG_ON(this_chunk_bytes == 0);
1014
Andi Kleeneb28be22011-08-01 21:38:03 -07001015 sdio->boundary = buffer_boundary(map_bh);
1016 ret = submit_page_section(dio, sdio, page,
1017 offset_in_page,
1018 this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -07001019 sdio->next_block_for_io,
1020 map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021 if (ret) {
1022 page_cache_release(page);
1023 goto out;
1024 }
Andi Kleeneb28be22011-08-01 21:38:03 -07001025 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026
Andi Kleeneb28be22011-08-01 21:38:03 -07001027 sdio->block_in_file += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028 block_in_page += this_chunk_blocks;
Andi Kleeneb28be22011-08-01 21:38:03 -07001029 sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030next_block:
Andi Kleeneb28be22011-08-01 21:38:03 -07001031 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1032 if (sdio->block_in_file == sdio->final_block_in_request)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001033 break;
1034 }
1035
1036 /* Drop the ref which was taken in get_user_pages() */
1037 page_cache_release(page);
1038 block_in_page = 0;
1039 }
1040out:
1041 return ret;
1042}
1043
Andi Kleen847cc632011-08-01 21:38:09 -07001044static inline int drop_refcount(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045{
Andi Kleen847cc632011-08-01 21:38:09 -07001046 int ret2;
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001047 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048
Zach Brown8459d862006-12-10 02:21:05 -08001049 /*
1050 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001051 * operation. AIO can if it was a broken operation described above or
1052 * in fact if all the bios race to complete before we get here. In
1053 * that case dio_complete() translates the EIOCBQUEUED into the proper
1054 * return code that the caller will hand to aio_complete().
1055 *
1056 * This is managed by the bio_lock instead of being an atomic_t so that
1057 * completion paths can drop their ref and use the remaining count to
1058 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001059 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001060 spin_lock_irqsave(&dio->bio_lock, flags);
1061 ret2 = --dio->refcount;
1062 spin_unlock_irqrestore(&dio->bio_lock, flags);
Andi Kleen847cc632011-08-01 21:38:09 -07001063 return ret2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064}
1065
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001066/*
1067 * This is a library function for use by filesystem drivers.
1068 *
1069 * The locking rules are governed by the flags parameter:
1070 * - if the flags value contains DIO_LOCKING we use a fancy locking
1071 * scheme for dumb filesystems.
1072 * For writes this function is called under i_mutex and returns with
1073 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1074 * taken and dropped again before returning.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001075 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1076 * internal locking but rather rely on the filesystem to synchronize
1077 * direct I/O reads/writes versus each other and truncate.
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001078 *
1079 * To help with locking against truncate we incremented the i_dio_count
1080 * counter before starting direct I/O, and decrement it once we are done.
1081 * Truncate can wait for it to reach zero to provide exclusion. It is
1082 * expected that filesystem provide exclusion between new direct I/O
1083 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1084 * but other filesystems need to take care of this on their own.
Andi Kleenba253fb2011-08-01 21:38:08 -07001085 *
1086 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1087 * is always inlined. Otherwise gcc is unable to split the structure into
1088 * individual fields and will generate much worse code. This is important
1089 * for the whole file.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001090 */
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001091static inline ssize_t
1092do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 struct block_device *bdev, const struct iovec *iov, loff_t offset,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -08001094 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
Josef Bacikfacd07b2010-05-23 11:00:55 -04001095 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096{
1097 int seg;
1098 size_t size;
1099 unsigned long addr;
1100 unsigned blkbits = inode->i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 unsigned blocksize_mask = (1 << blkbits) - 1;
1102 ssize_t retval = -EINVAL;
1103 loff_t end = offset;
1104 struct dio *dio;
Andi Kleeneb28be22011-08-01 21:38:03 -07001105 struct dio_submit sdio = { 0, };
Andi Kleen847cc632011-08-01 21:38:09 -07001106 unsigned long user_addr;
1107 size_t bytes;
1108 struct buffer_head map_bh = { 0, };
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109
1110 if (rw & WRITE)
Jens Axboe721a9602011-03-09 11:56:30 +01001111 rw = WRITE_ODIRECT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001113 /*
1114 * Avoid references to bdev if not absolutely needed to give
1115 * the early prefetch in the caller enough time.
1116 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117
1118 if (offset & blocksize_mask) {
1119 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001120 blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 blocksize_mask = (1 << blkbits) - 1;
1122 if (offset & blocksize_mask)
1123 goto out;
1124 }
1125
1126 /* Check the memory alignment. Blocks cannot straddle pages */
1127 for (seg = 0; seg < nr_segs; seg++) {
1128 addr = (unsigned long)iov[seg].iov_base;
1129 size = iov[seg].iov_len;
1130 end += size;
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001131 if (unlikely((addr & blocksize_mask) ||
1132 (size & blocksize_mask))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001134 blkbits = blksize_bits(
1135 bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136 blocksize_mask = (1 << blkbits) - 1;
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001137 if ((addr & blocksize_mask) || (size & blocksize_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138 goto out;
1139 }
1140 }
1141
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001142 /* watch out for a 0 len io from a tricksy fs */
1143 if (rw == READ && end == offset)
1144 return 0;
1145
Andi Kleen6e8267f2011-08-01 21:38:06 -07001146 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 retval = -ENOMEM;
1148 if (!dio)
1149 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001150 /*
1151 * Believe it or not, zeroing out the page array caused a .5%
1152 * performance regression in a database benchmark. So, we take
1153 * care to only zero out what's needed.
1154 */
1155 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001157 dio->flags = flags;
1158 if (dio->flags & DIO_LOCKING) {
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001159 if (rw == READ) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001160 struct address_space *mapping =
1161 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001163 /* will be released by direct_io_worker */
1164 mutex_lock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165
1166 retval = filemap_write_and_wait_range(mapping, offset,
1167 end - 1);
1168 if (retval) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001169 mutex_unlock(&inode->i_mutex);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001170 kmem_cache_free(dio_cache, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171 goto out;
1172 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 }
1175
1176 /*
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001177 * Will be decremented at I/O completion time.
1178 */
1179 atomic_inc(&inode->i_dio_count);
1180
1181 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 * For file extending writes updating i_size before data
1183 * writeouts complete can expose uninitialized blocks. So
1184 * even for AIO, we need to wait for i/o to complete before
1185 * returning in this case.
1186 */
Jens Axboeb31dc662006-06-13 08:26:10 +02001187 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188 (end > i_size_read(inode)));
1189
Andi Kleen847cc632011-08-01 21:38:09 -07001190 retval = 0;
1191
1192 dio->inode = inode;
1193 dio->rw = rw;
1194 sdio.blkbits = blkbits;
1195 sdio.blkfactor = inode->i_blkbits - blkbits;
1196 sdio.block_in_file = offset >> blkbits;
1197
1198 sdio.get_block = get_block;
1199 dio->end_io = end_io;
1200 sdio.submit_io = submit_io;
1201 sdio.final_block_in_bio = -1;
1202 sdio.next_block_for_io = -1;
1203
1204 dio->iocb = iocb;
1205 dio->i_size = i_size_read(inode);
1206
1207 spin_lock_init(&dio->bio_lock);
1208 dio->refcount = 1;
1209
1210 /*
1211 * In case of non-aligned buffers, we may need 2 more
1212 * pages since we need to zero out first and last block.
1213 */
1214 if (unlikely(sdio.blkfactor))
1215 sdio.pages_in_io = 2;
1216
1217 for (seg = 0; seg < nr_segs; seg++) {
1218 user_addr = (unsigned long)iov[seg].iov_base;
1219 sdio.pages_in_io +=
1220 ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) /
1221 PAGE_SIZE - user_addr / PAGE_SIZE);
1222 }
1223
1224 for (seg = 0; seg < nr_segs; seg++) {
1225 user_addr = (unsigned long)iov[seg].iov_base;
1226 sdio.size += bytes = iov[seg].iov_len;
1227
1228 /* Index into the first page of the first block */
1229 sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1230 sdio.final_block_in_request = sdio.block_in_file +
1231 (bytes >> blkbits);
1232 /* Page fetching state */
1233 sdio.head = 0;
1234 sdio.tail = 0;
1235 sdio.curr_page = 0;
1236
1237 sdio.total_pages = 0;
1238 if (user_addr & (PAGE_SIZE-1)) {
1239 sdio.total_pages++;
1240 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1241 }
1242 sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1243 sdio.curr_user_address = user_addr;
1244
1245 retval = do_direct_IO(dio, &sdio, &map_bh);
1246
1247 dio->result += iov[seg].iov_len -
1248 ((sdio.final_block_in_request - sdio.block_in_file) <<
1249 blkbits);
1250
1251 if (retval) {
1252 dio_cleanup(dio, &sdio);
1253 break;
1254 }
1255 } /* end iovec loop */
1256
1257 if (retval == -ENOTBLK) {
1258 /*
1259 * The remaining part of the request will be
1260 * be handled by buffered I/O when we return
1261 */
1262 retval = 0;
1263 }
1264 /*
1265 * There may be some unwritten disk at the end of a part-written
1266 * fs-block-sized block. Go zero that now.
1267 */
1268 dio_zero_block(dio, &sdio, 1, &map_bh);
1269
1270 if (sdio.cur_page) {
1271 ssize_t ret2;
1272
1273 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1274 if (retval == 0)
1275 retval = ret2;
1276 page_cache_release(sdio.cur_page);
1277 sdio.cur_page = NULL;
1278 }
1279 if (sdio.bio)
1280 dio_bio_submit(dio, &sdio);
1281
1282 /*
1283 * It is possible that, we return short IO due to end of file.
1284 * In that case, we need to release all the pages we got hold on.
1285 */
1286 dio_cleanup(dio, &sdio);
1287
1288 /*
1289 * All block lookups have been performed. For READ requests
1290 * we can let i_mutex go now that its achieved its purpose
1291 * of protecting us from looking up uninitialized blocks.
1292 */
1293 if (rw == READ && (dio->flags & DIO_LOCKING))
1294 mutex_unlock(&dio->inode->i_mutex);
1295
1296 /*
1297 * The only time we want to leave bios in flight is when a successful
1298 * partial aio read or full aio write have been setup. In that case
1299 * bio completion will call aio_complete. The only time it's safe to
1300 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1301 * This had *better* be the only place that raises -EIOCBQUEUED.
1302 */
1303 BUG_ON(retval == -EIOCBQUEUED);
1304 if (dio->is_async && retval == 0 && dio->result &&
1305 ((rw & READ) || (dio->result == sdio.size)))
1306 retval = -EIOCBQUEUED;
1307
1308 if (retval != -EIOCBQUEUED)
1309 dio_await_completion(dio);
1310
1311 if (drop_refcount(dio) == 0) {
1312 retval = dio_complete(dio, offset, retval, false);
1313 kmem_cache_free(dio_cache, dio);
1314 } else
1315 BUG_ON(retval != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001317out:
1318 return retval;
1319}
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001320
1321ssize_t
1322__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1323 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1324 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1325 dio_submit_t submit_io, int flags)
1326{
1327 /*
1328 * The block device state is needed in the end to finally
1329 * submit everything. Since it's likely to be cache cold
1330 * prefetch it here as first thing to hide some of the
1331 * latency.
1332 *
1333 * Attempt to prefetch the pieces we likely need later.
1334 */
1335 prefetch(&bdev->bd_disk->part_tbl);
1336 prefetch(bdev->bd_queue);
1337 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1338
1339 return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset,
1340 nr_segs, get_block, end_io,
1341 submit_io, flags);
1342}
1343
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344EXPORT_SYMBOL(__blockdev_direct_IO);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001345
1346static __init int dio_init(void)
1347{
1348 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1349 return 0;
1350}
1351module_init(dio_init)