blob: a10cb91cadea04ac68ff5ce0db34d6d07bdd6371 [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>
38#include <asm/atomic.h>
39
40/*
41 * How many user pages to map in one call to get_user_pages(). This determines
42 * the size of a structure on the stack.
43 */
44#define DIO_PAGES 64
45
46/*
47 * This code generally works in units of "dio_blocks". A dio_block is
48 * somewhere between the hard sector size and the filesystem block size. it
49 * is determined on a per-invocation basis. When talking to the filesystem
50 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
51 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
52 * to bio_block quantities by shifting left by blkfactor.
53 *
54 * If blkfactor is zero then the user's request was aligned to the filesystem's
55 * blocksize.
Linus Torvalds1da177e2005-04-16 15:20:36 -070056 */
57
58struct dio {
59 /* BIO submission state */
60 struct bio *bio; /* bio under assembly */
61 struct inode *inode;
62 int rw;
Daniel McNeil29504ff2005-04-16 15:25:50 -070063 loff_t i_size; /* i_size when submitted */
Christoph Hellwig5fe878a2009-12-15 16:47:50 -080064 int flags; /* doesn't change */
Linus Torvalds1da177e2005-04-16 15:20:36 -070065 unsigned blkbits; /* doesn't change */
66 unsigned blkfactor; /* When we're using an alignment which
67 is finer than the filesystem's soft
68 blocksize, this specifies how much
69 finer. blkfactor=2 means 1/4-block
70 alignment. Does not change */
71 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
72 been performed at the start of a
73 write */
74 int pages_in_io; /* approximate total IO pages */
75 size_t size; /* total request size (doesn't change)*/
76 sector_t block_in_file; /* Current offset into the underlying
77 file in dio_block units. */
78 unsigned blocks_available; /* At block_in_file. changes */
79 sector_t final_block_in_request;/* doesn't change */
80 unsigned first_block_in_page; /* doesn't change, Used only once */
81 int boundary; /* prev block is at a boundary */
82 int reap_counter; /* rate limit reaping */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080083 get_block_t *get_block; /* block mapping function */
Linus Torvalds1da177e2005-04-16 15:20:36 -070084 dio_iodone_t *end_io; /* IO completion function */
Josef Bacikfacd07b2010-05-23 11:00:55 -040085 dio_submit_t *submit_io; /* IO submition function */
86 loff_t logical_offset_in_bio; /* current first logical block in bio */
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 sector_t final_block_in_bio; /* current final block in bio + 1 */
88 sector_t next_block_for_io; /* next block to be put under IO,
89 in dio_blocks units */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080090 struct buffer_head map_bh; /* last get_block() result */
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
92 /*
93 * Deferred addition of a page to the dio. These variables are
94 * private to dio_send_cur_page(), submit_page_section() and
95 * dio_bio_add_page().
96 */
97 struct page *cur_page; /* The page */
98 unsigned cur_page_offset; /* Offset into it, in bytes */
99 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
100 sector_t cur_page_block; /* Where it starts */
Josef Bacikfacd07b2010-05-23 11:00:55 -0400101 loff_t cur_page_fs_offset; /* Offset in file */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103 /* BIO completion state */
104 spinlock_t bio_lock; /* protects BIO fields below */
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800105 unsigned long refcount; /* direct_io_worker() and bios */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106 struct bio *bio_list; /* singly linked via bi_private */
107 struct task_struct *waiter; /* waiting task (NULL if none) */
108
109 /* AIO related stuff */
110 struct kiocb *iocb; /* kiocb */
111 int is_async; /* is IO async ? */
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800112 int io_error; /* IO error in completion path */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 ssize_t result; /* IO result */
Jeff Moyer23aee092009-12-15 16:47:49 -0800114
115 /*
116 * Page fetching state. These variables belong to dio_refill_pages().
117 */
118 int curr_page; /* changes */
119 int total_pages; /* doesn't change */
120 unsigned long curr_user_address;/* changes */
121
122 /*
123 * Page queue. These variables belong to dio_refill_pages() and
124 * dio_get_page().
125 */
126 unsigned head; /* next page to process */
127 unsigned tail; /* last valid page + 1 */
128 int page_errors; /* errno from get_user_pages() */
129
130 /*
131 * pages[] (and any fields placed after it) are not zeroed out at
132 * allocation time. Don't add new fields after pages[] unless you
133 * wish that they not be zeroed.
134 */
135 struct page *pages[DIO_PAGES]; /* page buffer */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136};
137
138/*
139 * How many pages are in the queue?
140 */
141static inline unsigned dio_pages_present(struct dio *dio)
142{
143 return dio->tail - dio->head;
144}
145
146/*
147 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
148 */
149static int dio_refill_pages(struct dio *dio)
150{
151 int ret;
152 int nr_pages;
153
154 nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES);
Nick Pigginf5dd33c2008-07-25 19:45:25 -0700155 ret = get_user_pages_fast(
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156 dio->curr_user_address, /* Where from? */
157 nr_pages, /* How many pages? */
158 dio->rw == READ, /* Write to memory? */
Nick Pigginf5dd33c2008-07-25 19:45:25 -0700159 &dio->pages[0]); /* Put results here */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160
Jens Axboeb31dc662006-06-13 08:26:10 +0200161 if (ret < 0 && dio->blocks_available && (dio->rw & WRITE)) {
Nick Piggin557ed1f2007-10-16 01:24:40 -0700162 struct page *page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163 /*
164 * A memory fault, but the filesystem has some outstanding
165 * mapped blocks. We need to use those blocks up to avoid
166 * leaking stale data in the file.
167 */
168 if (dio->page_errors == 0)
169 dio->page_errors = ret;
Nick Pigginb5810032005-10-29 18:16:12 -0700170 page_cache_get(page);
171 dio->pages[0] = page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 dio->head = 0;
173 dio->tail = 1;
174 ret = 0;
175 goto out;
176 }
177
178 if (ret >= 0) {
179 dio->curr_user_address += ret * PAGE_SIZE;
180 dio->curr_page += ret;
181 dio->head = 0;
182 dio->tail = ret;
183 ret = 0;
184 }
185out:
186 return ret;
187}
188
189/*
190 * Get another userspace page. Returns an ERR_PTR on error. Pages are
191 * buffered inside the dio so that we can call get_user_pages() against a
192 * decent number of pages, less frequently. To provide nicer use of the
193 * L1 cache.
194 */
195static struct page *dio_get_page(struct dio *dio)
196{
197 if (dio_pages_present(dio) == 0) {
198 int ret;
199
200 ret = dio_refill_pages(dio);
201 if (ret)
202 return ERR_PTR(ret);
203 BUG_ON(dio_pages_present(dio) == 0);
204 }
205 return dio->pages[dio->head++];
206}
207
Zach Brown6d544bb2006-12-10 02:20:54 -0800208/**
209 * dio_complete() - called when all DIO BIO I/O has been completed
210 * @offset: the byte offset in the file of the completed operation
211 *
212 * This releases locks as dictated by the locking type, lets interested parties
213 * know that a DIO operation has completed, and calculates the resulting return
214 * code for the operation.
215 *
216 * It lets the filesystem know if it registered an interest earlier via
217 * get_block. Pass the private field of the map buffer_head so that
218 * filesystems can use it to hold additional state between get_block calls and
219 * dio_complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220 */
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000221static int dio_complete(struct dio *dio, loff_t offset, int ret, bool is_async)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222{
Zach Brown6d544bb2006-12-10 02:20:54 -0800223 ssize_t transferred = 0;
224
Zach Brown8459d862006-12-10 02:21:05 -0800225 /*
226 * AIO submission can race with bio completion to get here while
227 * expecting to have the last io completed by bio completion.
228 * In that case -EIOCBQUEUED is in fact not an error we want
229 * to preserve through this call.
230 */
231 if (ret == -EIOCBQUEUED)
232 ret = 0;
233
Zach Brown6d544bb2006-12-10 02:20:54 -0800234 if (dio->result) {
235 transferred = dio->result;
236
237 /* Check for short read case */
238 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
239 transferred = dio->i_size - offset;
240 }
241
Zach Brown6d544bb2006-12-10 02:20:54 -0800242 if (ret == 0)
243 ret = dio->page_errors;
244 if (ret == 0)
245 ret = dio->io_error;
246 if (ret == 0)
247 ret = transferred;
248
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000249 if (dio->end_io && dio->result) {
250 dio->end_io(dio->iocb, offset, transferred,
251 dio->map_bh.b_private, ret, is_async);
252 } else if (is_async) {
253 aio_complete(dio->iocb, ret, 0);
254 }
255
256 if (dio->flags & DIO_LOCKING)
257 /* lockdep: non-owner release */
258 up_read_non_owner(&dio->inode->i_alloc_sem);
259
Zach Brown6d544bb2006-12-10 02:20:54 -0800260 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261}
262
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263static int dio_bio_complete(struct dio *dio, struct bio *bio);
264/*
265 * Asynchronous IO callback.
266 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200267static void dio_bio_end_aio(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268{
269 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800270 unsigned long remaining;
271 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273 /* cleanup the bio */
274 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800275
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800276 spin_lock_irqsave(&dio->bio_lock, flags);
277 remaining = --dio->refcount;
278 if (remaining == 1 && dio->waiter)
Zach Brown20258b22006-12-10 02:21:01 -0800279 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800280 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b22006-12-10 02:21:01 -0800281
Zach Brown8459d862006-12-10 02:21:05 -0800282 if (remaining == 0) {
Christoph Hellwig40e2e972010-07-18 21:17:09 +0000283 dio_complete(dio, dio->iocb->ki_pos, 0, true);
Zach Brown8459d862006-12-10 02:21:05 -0800284 kfree(dio);
285 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286}
287
288/*
289 * The BIO completion handler simply queues the BIO up for the process-context
290 * handler.
291 *
292 * During I/O bi_private points at the dio. After I/O, bi_private is used to
293 * implement a singly-linked list of completed BIOs, at dio->bio_list.
294 */
NeilBrown6712ecf2007-09-27 12:47:43 +0200295static void dio_bio_end_io(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296{
297 struct dio *dio = bio->bi_private;
298 unsigned long flags;
299
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300 spin_lock_irqsave(&dio->bio_lock, flags);
301 bio->bi_private = dio->bio_list;
302 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800303 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304 wake_up_process(dio->waiter);
305 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306}
307
Josef Bacikfacd07b2010-05-23 11:00:55 -0400308/**
309 * dio_end_io - handle the end io action for the given bio
310 * @bio: The direct io bio thats being completed
311 * @error: Error if there was one
312 *
313 * This is meant to be called by any filesystem that uses their own dio_submit_t
314 * so that the DIO specific endio actions are dealt with after the filesystem
315 * has done it's completion work.
316 */
317void dio_end_io(struct bio *bio, int error)
318{
319 struct dio *dio = bio->bi_private;
320
321 if (dio->is_async)
322 dio_bio_end_aio(bio, error);
323 else
324 dio_bio_end_io(bio, error);
325}
326EXPORT_SYMBOL_GPL(dio_end_io);
327
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328static int
329dio_bio_alloc(struct dio *dio, struct block_device *bdev,
330 sector_t first_sector, int nr_vecs)
331{
332 struct bio *bio;
333
334 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335
336 bio->bi_bdev = bdev;
337 bio->bi_sector = first_sector;
338 if (dio->is_async)
339 bio->bi_end_io = dio_bio_end_aio;
340 else
341 bio->bi_end_io = dio_bio_end_io;
342
343 dio->bio = bio;
Josef Bacikfacd07b2010-05-23 11:00:55 -0400344 dio->logical_offset_in_bio = dio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345 return 0;
346}
347
348/*
349 * In the AIO read case we speculatively dirty the pages before starting IO.
350 * During IO completion, any of these pages which happen to have been written
351 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800352 *
353 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354 */
355static void dio_bio_submit(struct dio *dio)
356{
357 struct bio *bio = dio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800358 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359
360 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800361
362 spin_lock_irqsave(&dio->bio_lock, flags);
363 dio->refcount++;
364 spin_unlock_irqrestore(&dio->bio_lock, flags);
365
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366 if (dio->is_async && dio->rw == READ)
367 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800368
Josef Bacikfacd07b2010-05-23 11:00:55 -0400369 if (dio->submit_io)
370 dio->submit_io(dio->rw, bio, dio->inode,
371 dio->logical_offset_in_bio);
372 else
373 submit_bio(dio->rw, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374
375 dio->bio = NULL;
376 dio->boundary = 0;
Josef Bacikfacd07b2010-05-23 11:00:55 -0400377 dio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378}
379
380/*
381 * Release any resources in case of a failure
382 */
383static void dio_cleanup(struct dio *dio)
384{
385 while (dio_pages_present(dio))
386 page_cache_release(dio_get_page(dio));
387}
388
389/*
Zach Brown02732012006-12-10 02:20:59 -0800390 * Wait for the next BIO to complete. Remove it and return it. NULL is
391 * returned once all BIOs have been completed. This must only be called once
392 * all bios have been issued so that dio->refcount can only decrease. This
393 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394 */
395static struct bio *dio_await_one(struct dio *dio)
396{
397 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800398 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399
400 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800401
402 /*
403 * Wait as long as the list is empty and there are bios in flight. bio
404 * completion drops the count, maybe adds to the list, and wakes while
405 * holding the bio_lock so we don't need set_current_state()'s barrier
406 * and can call it after testing our condition.
407 */
408 while (dio->refcount > 1 && dio->bio_list == NULL) {
409 __set_current_state(TASK_UNINTERRUPTIBLE);
410 dio->waiter = current;
411 spin_unlock_irqrestore(&dio->bio_lock, flags);
412 io_schedule();
413 /* wake up sets us TASK_RUNNING */
414 spin_lock_irqsave(&dio->bio_lock, flags);
415 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 }
Zach Brown02732012006-12-10 02:20:59 -0800417 if (dio->bio_list) {
418 bio = dio->bio_list;
419 dio->bio_list = bio->bi_private;
420 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 spin_unlock_irqrestore(&dio->bio_lock, flags);
422 return bio;
423}
424
425/*
426 * Process one completed BIO. No locks are held.
427 */
428static int dio_bio_complete(struct dio *dio, struct bio *bio)
429{
430 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
431 struct bio_vec *bvec = bio->bi_io_vec;
432 int page_no;
433
434 if (!uptodate)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800435 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436
437 if (dio->is_async && dio->rw == READ) {
438 bio_check_pages_dirty(bio); /* transfers ownership */
439 } else {
440 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
441 struct page *page = bvec[page_no].bv_page;
442
443 if (dio->rw == READ && !PageCompound(page))
444 set_page_dirty_lock(page);
445 page_cache_release(page);
446 }
447 bio_put(bio);
448 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449 return uptodate ? 0 : -EIO;
450}
451
452/*
Zach Brown02732012006-12-10 02:20:59 -0800453 * Wait on and process all in-flight BIOs. This must only be called once
454 * all bios have been issued so that the refcount can only decrease.
455 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200456 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800457 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800459static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460{
Zach Brown02732012006-12-10 02:20:59 -0800461 struct bio *bio;
462 do {
463 bio = dio_await_one(dio);
464 if (bio)
465 dio_bio_complete(dio, bio);
466 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467}
468
469/*
470 * A really large O_DIRECT read or write can generate a lot of BIOs. So
471 * to keep the memory consumption sane we periodically reap any completed BIOs
472 * during the BIO generation phase.
473 *
474 * This also helps to limit the peak amount of pinned userspace memory.
475 */
476static int dio_bio_reap(struct dio *dio)
477{
478 int ret = 0;
479
480 if (dio->reap_counter++ >= 64) {
481 while (dio->bio_list) {
482 unsigned long flags;
483 struct bio *bio;
484 int ret2;
485
486 spin_lock_irqsave(&dio->bio_lock, flags);
487 bio = dio->bio_list;
488 dio->bio_list = bio->bi_private;
489 spin_unlock_irqrestore(&dio->bio_lock, flags);
490 ret2 = dio_bio_complete(dio, bio);
491 if (ret == 0)
492 ret = ret2;
493 }
494 dio->reap_counter = 0;
495 }
496 return ret;
497}
498
499/*
500 * Call into the fs to map some more disk blocks. We record the current number
501 * of available blocks at dio->blocks_available. These are in units of the
502 * fs blocksize, (1 << inode->i_blkbits).
503 *
504 * The fs is allowed to map lots of blocks at once. If it wants to do that,
505 * it uses the passed inode-relative block number as the file offset, as usual.
506 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800507 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508 * has remaining to do. The fs should not map more than this number of blocks.
509 *
510 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
511 * indicate how much contiguous disk space has been made available at
512 * bh->b_blocknr.
513 *
514 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
515 * This isn't very efficient...
516 *
517 * In the case of filesystem holes: the fs may return an arbitrarily-large
518 * hole by returning an appropriate value in b_size and by clearing
519 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800520 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 */
522static int get_more_blocks(struct dio *dio)
523{
524 int ret;
525 struct buffer_head *map_bh = &dio->map_bh;
526 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
527 unsigned long fs_count; /* Number of filesystem-sized blocks */
528 unsigned long dio_count;/* Number of dio_block-sized blocks */
529 unsigned long blkmask;
530 int create;
531
532 /*
533 * If there was a memory error and we've overwritten all the
534 * mapped blocks then we can now return that memory error
535 */
536 ret = dio->page_errors;
537 if (ret == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 BUG_ON(dio->block_in_file >= dio->final_block_in_request);
539 fs_startblk = dio->block_in_file >> dio->blkfactor;
540 dio_count = dio->final_block_in_request - dio->block_in_file;
541 fs_count = dio_count >> dio->blkfactor;
542 blkmask = (1 << dio->blkfactor) - 1;
543 if (dio_count & blkmask)
544 fs_count++;
545
Nathan Scott3c674e72006-03-29 09:26:15 +1000546 map_bh->b_state = 0;
547 map_bh->b_size = fs_count << dio->inode->i_blkbits;
548
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800549 /*
550 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
551 * forbid block creations: only overwrites are permitted.
552 * We will return early to the caller once we see an
553 * unmapped buffer head returned, and the caller will fall
554 * back to buffered I/O.
555 *
556 * Otherwise the decision is left to the get_blocks method,
557 * which may decide to handle it or also return an unmapped
558 * buffer head.
559 */
Jens Axboeb31dc662006-06-13 08:26:10 +0200560 create = dio->rw & WRITE;
Christoph Hellwig5fe878a2009-12-15 16:47:50 -0800561 if (dio->flags & DIO_SKIP_HOLES) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 if (dio->block_in_file < (i_size_read(dio->inode) >>
563 dio->blkbits))
564 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000566
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800567 ret = (*dio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568 map_bh, create);
569 }
570 return ret;
571}
572
573/*
574 * There is no bio. Make one now.
575 */
576static int dio_new_bio(struct dio *dio, sector_t start_sector)
577{
578 sector_t sector;
579 int ret, nr_pages;
580
581 ret = dio_bio_reap(dio);
582 if (ret)
583 goto out;
584 sector = start_sector << (dio->blkbits - 9);
585 nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
586 BUG_ON(nr_pages <= 0);
587 ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
588 dio->boundary = 0;
589out:
590 return ret;
591}
592
593/*
594 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
595 * that was successful then update final_block_in_bio and take a ref against
596 * the just-added page.
597 *
598 * Return zero on success. Non-zero means the caller needs to start a new BIO.
599 */
600static int dio_bio_add_page(struct dio *dio)
601{
602 int ret;
603
604 ret = bio_add_page(dio->bio, dio->cur_page,
605 dio->cur_page_len, dio->cur_page_offset);
606 if (ret == dio->cur_page_len) {
607 /*
608 * Decrement count only, if we are done with this page
609 */
610 if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE)
611 dio->pages_in_io--;
612 page_cache_get(dio->cur_page);
613 dio->final_block_in_bio = dio->cur_page_block +
614 (dio->cur_page_len >> dio->blkbits);
615 ret = 0;
616 } else {
617 ret = 1;
618 }
619 return ret;
620}
621
622/*
623 * Put cur_page under IO. The section of cur_page which is described by
624 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
625 * starts on-disk at cur_page_block.
626 *
627 * We take a ref against the page here (on behalf of its presence in the bio).
628 *
629 * The caller of this function is responsible for removing cur_page from the
630 * dio, and for dropping the refcount which came from that presence.
631 */
632static int dio_send_cur_page(struct dio *dio)
633{
634 int ret = 0;
635
636 if (dio->bio) {
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400637 loff_t cur_offset = dio->block_in_file << dio->blkbits;
638 loff_t bio_next_offset = dio->logical_offset_in_bio +
639 dio->bio->bi_size;
640
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400642 * See whether this new request is contiguous with the old.
643 *
644 * Btrfs cannot handl having logically non-contiguous requests
645 * submitted. For exmple if you have
646 *
647 * Logical: [0-4095][HOLE][8192-12287]
648 * Phyiscal: [0-4095] [4096-8181]
649 *
650 * We cannot submit those pages together as one BIO. So if our
651 * current logical offset in the file does not equal what would
652 * be the next logical offset in the bio, submit the bio we
653 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 */
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400655 if (dio->final_block_in_bio != dio->cur_page_block ||
656 cur_offset != bio_next_offset)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 dio_bio_submit(dio);
658 /*
659 * Submit now if the underlying fs is about to perform a
660 * metadata read
661 */
662 if (dio->boundary)
663 dio_bio_submit(dio);
664 }
665
666 if (dio->bio == NULL) {
667 ret = dio_new_bio(dio, dio->cur_page_block);
668 if (ret)
669 goto out;
670 }
671
672 if (dio_bio_add_page(dio) != 0) {
673 dio_bio_submit(dio);
674 ret = dio_new_bio(dio, dio->cur_page_block);
675 if (ret == 0) {
676 ret = dio_bio_add_page(dio);
677 BUG_ON(ret != 0);
678 }
679 }
680out:
681 return ret;
682}
683
684/*
685 * An autonomous function to put a chunk of a page under deferred IO.
686 *
687 * The caller doesn't actually know (or care) whether this piece of page is in
688 * a BIO, or is under IO or whatever. We just take care of all possible
689 * situations here. The separation between the logic of do_direct_IO() and
690 * that of submit_page_section() is important for clarity. Please don't break.
691 *
692 * The chunk of page starts on-disk at blocknr.
693 *
694 * We perform deferred IO, by recording the last-submitted page inside our
695 * private part of the dio structure. If possible, we just expand the IO
696 * across that page here.
697 *
698 * If that doesn't work out then we put the old page into the bio and add this
699 * page to the dio instead.
700 */
701static int
702submit_page_section(struct dio *dio, struct page *page,
703 unsigned offset, unsigned len, sector_t blocknr)
704{
705 int ret = 0;
706
Andrew Morton98c4d572006-12-10 02:19:47 -0800707 if (dio->rw & WRITE) {
708 /*
709 * Read accounting is performed in submit_bio()
710 */
711 task_io_account_write(len);
712 }
713
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 /*
715 * Can we just grow the current page's presence in the dio?
716 */
717 if ( (dio->cur_page == page) &&
718 (dio->cur_page_offset + dio->cur_page_len == offset) &&
719 (dio->cur_page_block +
720 (dio->cur_page_len >> dio->blkbits) == blocknr)) {
721 dio->cur_page_len += len;
722
723 /*
724 * If dio->boundary then we want to schedule the IO now to
725 * avoid metadata seeks.
726 */
727 if (dio->boundary) {
728 ret = dio_send_cur_page(dio);
729 page_cache_release(dio->cur_page);
730 dio->cur_page = NULL;
731 }
732 goto out;
733 }
734
735 /*
736 * If there's a deferred page already there then send it.
737 */
738 if (dio->cur_page) {
739 ret = dio_send_cur_page(dio);
740 page_cache_release(dio->cur_page);
741 dio->cur_page = NULL;
742 if (ret)
743 goto out;
744 }
745
746 page_cache_get(page); /* It is in dio */
747 dio->cur_page = page;
748 dio->cur_page_offset = offset;
749 dio->cur_page_len = len;
750 dio->cur_page_block = blocknr;
Josef Bacikfacd07b2010-05-23 11:00:55 -0400751 dio->cur_page_fs_offset = dio->block_in_file << dio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752out:
753 return ret;
754}
755
756/*
757 * Clean any dirty buffers in the blockdev mapping which alias newly-created
758 * file blocks. Only called for S_ISREG files - blockdevs do not set
759 * buffer_new
760 */
761static void clean_blockdev_aliases(struct dio *dio)
762{
763 unsigned i;
764 unsigned nblocks;
765
766 nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
767
768 for (i = 0; i < nblocks; i++) {
769 unmap_underlying_metadata(dio->map_bh.b_bdev,
770 dio->map_bh.b_blocknr + i);
771 }
772}
773
774/*
775 * If we are not writing the entire block and get_block() allocated
776 * the block for us, we need to fill-in the unused portion of the
777 * block with zeros. This happens only if user-buffer, fileoffset or
778 * io length is not filesystem block-size multiple.
779 *
780 * `end' is zero if we're doing the start of the IO, 1 at the end of the
781 * IO.
782 */
783static void dio_zero_block(struct dio *dio, int end)
784{
785 unsigned dio_blocks_per_fs_block;
786 unsigned this_chunk_blocks; /* In dio_blocks */
787 unsigned this_chunk_bytes;
788 struct page *page;
789
790 dio->start_zero_done = 1;
791 if (!dio->blkfactor || !buffer_new(&dio->map_bh))
792 return;
793
794 dio_blocks_per_fs_block = 1 << dio->blkfactor;
795 this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
796
797 if (!this_chunk_blocks)
798 return;
799
800 /*
801 * We need to zero out part of an fs block. It is either at the
802 * beginning or the end of the fs block.
803 */
804 if (end)
805 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
806
807 this_chunk_bytes = this_chunk_blocks << dio->blkbits;
808
Nick Piggin557ed1f2007-10-16 01:24:40 -0700809 page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810 if (submit_page_section(dio, page, 0, this_chunk_bytes,
811 dio->next_block_for_io))
812 return;
813
814 dio->next_block_for_io += this_chunk_blocks;
815}
816
817/*
818 * Walk the user pages, and the file, mapping blocks to disk and generating
819 * a sequence of (page,offset,len,block) mappings. These mappings are injected
820 * into submit_page_section(), which takes care of the next stage of submission
821 *
822 * Direct IO against a blockdev is different from a file. Because we can
823 * happily perform page-sized but 512-byte aligned IOs. It is important that
824 * blockdev IO be able to have fine alignment and large sizes.
825 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800826 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 * with the size of IO which is permitted at this offset and this i_blkbits.
828 *
829 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800830 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 * fine alignment but still allows this function to work in PAGE_SIZE units.
832 */
833static int do_direct_IO(struct dio *dio)
834{
835 const unsigned blkbits = dio->blkbits;
836 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
837 struct page *page;
838 unsigned block_in_page;
839 struct buffer_head *map_bh = &dio->map_bh;
840 int ret = 0;
841
842 /* The I/O can start at any block offset within the first page */
843 block_in_page = dio->first_block_in_page;
844
845 while (dio->block_in_file < dio->final_block_in_request) {
846 page = dio_get_page(dio);
847 if (IS_ERR(page)) {
848 ret = PTR_ERR(page);
849 goto out;
850 }
851
852 while (block_in_page < blocks_per_page) {
853 unsigned offset_in_page = block_in_page << blkbits;
854 unsigned this_chunk_bytes; /* # of bytes mapped */
855 unsigned this_chunk_blocks; /* # of blocks */
856 unsigned u;
857
858 if (dio->blocks_available == 0) {
859 /*
860 * Need to go and map some more disk
861 */
862 unsigned long blkmask;
863 unsigned long dio_remainder;
864
865 ret = get_more_blocks(dio);
866 if (ret) {
867 page_cache_release(page);
868 goto out;
869 }
870 if (!buffer_mapped(map_bh))
871 goto do_holes;
872
873 dio->blocks_available =
874 map_bh->b_size >> dio->blkbits;
875 dio->next_block_for_io =
876 map_bh->b_blocknr << dio->blkfactor;
877 if (buffer_new(map_bh))
878 clean_blockdev_aliases(dio);
879
880 if (!dio->blkfactor)
881 goto do_holes;
882
883 blkmask = (1 << dio->blkfactor) - 1;
884 dio_remainder = (dio->block_in_file & blkmask);
885
886 /*
887 * If we are at the start of IO and that IO
888 * starts partway into a fs-block,
889 * dio_remainder will be non-zero. If the IO
890 * is a read then we can simply advance the IO
891 * cursor to the first block which is to be
892 * read. But if the IO is a write and the
893 * block was newly allocated we cannot do that;
894 * the start of the fs block must be zeroed out
895 * on-disk
896 */
897 if (!buffer_new(map_bh))
898 dio->next_block_for_io += dio_remainder;
899 dio->blocks_available -= dio_remainder;
900 }
901do_holes:
902 /* Handle holes */
903 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -0800904 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905
906 /* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboeb31dc662006-06-13 08:26:10 +0200907 if (dio->rw & WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908 page_cache_release(page);
909 return -ENOTBLK;
910 }
911
Jeff Moyer35dc8162006-02-03 03:04:27 -0800912 /*
913 * Be sure to account for a partial block as the
914 * last block in the file
915 */
916 i_size_aligned = ALIGN(i_size_read(dio->inode),
917 1 << blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 if (dio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -0800919 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 /* We hit eof */
921 page_cache_release(page);
922 goto out;
923 }
Christoph Lametereebd2aa2008-02-04 22:28:29 -0800924 zero_user(page, block_in_page << blkbits,
925 1 << blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926 dio->block_in_file++;
927 block_in_page++;
928 goto next_block;
929 }
930
931 /*
932 * If we're performing IO which has an alignment which
933 * is finer than the underlying fs, go check to see if
934 * we must zero out the start of this block.
935 */
936 if (unlikely(dio->blkfactor && !dio->start_zero_done))
937 dio_zero_block(dio, 0);
938
939 /*
940 * Work out, in this_chunk_blocks, how much disk we
941 * can add to this page
942 */
943 this_chunk_blocks = dio->blocks_available;
944 u = (PAGE_SIZE - offset_in_page) >> blkbits;
945 if (this_chunk_blocks > u)
946 this_chunk_blocks = u;
947 u = dio->final_block_in_request - dio->block_in_file;
948 if (this_chunk_blocks > u)
949 this_chunk_blocks = u;
950 this_chunk_bytes = this_chunk_blocks << blkbits;
951 BUG_ON(this_chunk_bytes == 0);
952
953 dio->boundary = buffer_boundary(map_bh);
954 ret = submit_page_section(dio, page, offset_in_page,
955 this_chunk_bytes, dio->next_block_for_io);
956 if (ret) {
957 page_cache_release(page);
958 goto out;
959 }
960 dio->next_block_for_io += this_chunk_blocks;
961
962 dio->block_in_file += this_chunk_blocks;
963 block_in_page += this_chunk_blocks;
964 dio->blocks_available -= this_chunk_blocks;
965next_block:
Eric Sesterhennd4569d22006-04-01 01:10:13 +0200966 BUG_ON(dio->block_in_file > dio->final_block_in_request);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967 if (dio->block_in_file == dio->final_block_in_request)
968 break;
969 }
970
971 /* Drop the ref which was taken in get_user_pages() */
972 page_cache_release(page);
973 block_in_page = 0;
974 }
975out:
976 return ret;
977}
978
979/*
Jes Sorensen1b1dcc12006-01-09 15:59:24 -0800980 * Releases both i_mutex and i_alloc_sem
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981 */
982static ssize_t
983direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
984 const struct iovec *iov, loff_t offset, unsigned long nr_segs,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800985 unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
Josef Bacikfacd07b2010-05-23 11:00:55 -0400986 dio_submit_t submit_io, struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987{
988 unsigned long user_addr;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800989 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990 int seg;
991 ssize_t ret = 0;
992 ssize_t ret2;
993 size_t bytes;
994
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995 dio->inode = inode;
996 dio->rw = rw;
997 dio->blkbits = blkbits;
998 dio->blkfactor = inode->i_blkbits - blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999 dio->block_in_file = offset >> blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -08001001 dio->get_block = get_block;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002 dio->end_io = end_io;
Josef Bacikfacd07b2010-05-23 11:00:55 -04001003 dio->submit_io = submit_io;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004 dio->final_block_in_bio = -1;
1005 dio->next_block_for_io = -1;
1006
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007 dio->iocb = iocb;
Daniel McNeil29504ff2005-04-16 15:25:50 -07001008 dio->i_size = i_size_read(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010 spin_lock_init(&dio->bio_lock);
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001011 dio->refcount = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012
1013 /*
1014 * In case of non-aligned buffers, we may need 2 more
1015 * pages since we need to zero out first and last block.
1016 */
1017 if (unlikely(dio->blkfactor))
1018 dio->pages_in_io = 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019
1020 for (seg = 0; seg < nr_segs; seg++) {
1021 user_addr = (unsigned long)iov[seg].iov_base;
1022 dio->pages_in_io +=
1023 ((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
1024 - user_addr/PAGE_SIZE);
1025 }
1026
1027 for (seg = 0; seg < nr_segs; seg++) {
1028 user_addr = (unsigned long)iov[seg].iov_base;
1029 dio->size += bytes = iov[seg].iov_len;
1030
1031 /* Index into the first page of the first block */
1032 dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1033 dio->final_block_in_request = dio->block_in_file +
1034 (bytes >> blkbits);
1035 /* Page fetching state */
1036 dio->head = 0;
1037 dio->tail = 0;
1038 dio->curr_page = 0;
1039
1040 dio->total_pages = 0;
1041 if (user_addr & (PAGE_SIZE-1)) {
1042 dio->total_pages++;
1043 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1044 }
1045 dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1046 dio->curr_user_address = user_addr;
1047
1048 ret = do_direct_IO(dio);
1049
1050 dio->result += iov[seg].iov_len -
1051 ((dio->final_block_in_request - dio->block_in_file) <<
1052 blkbits);
1053
1054 if (ret) {
1055 dio_cleanup(dio);
1056 break;
1057 }
1058 } /* end iovec loop */
1059
Josef Bacikfacd07b2010-05-23 11:00:55 -04001060 if (ret == -ENOTBLK) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061 /*
1062 * The remaining part of the request will be
1063 * be handled by buffered I/O when we return
1064 */
1065 ret = 0;
1066 }
1067 /*
1068 * There may be some unwritten disk at the end of a part-written
1069 * fs-block-sized block. Go zero that now.
1070 */
1071 dio_zero_block(dio, 1);
1072
1073 if (dio->cur_page) {
1074 ret2 = dio_send_cur_page(dio);
1075 if (ret == 0)
1076 ret = ret2;
1077 page_cache_release(dio->cur_page);
1078 dio->cur_page = NULL;
1079 }
1080 if (dio->bio)
1081 dio_bio_submit(dio);
1082
1083 /*
1084 * It is possible that, we return short IO due to end of file.
1085 * In that case, we need to release all the pages we got hold on.
1086 */
1087 dio_cleanup(dio);
1088
1089 /*
1090 * All block lookups have been performed. For READ requests
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001091 * we can let i_mutex go now that its achieved its purpose
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 * of protecting us from looking up uninitialized blocks.
1093 */
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001094 if (rw == READ && (dio->flags & DIO_LOCKING))
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001095 mutex_unlock(&dio->inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096
1097 /*
Zach Brown8459d862006-12-10 02:21:05 -08001098 * The only time we want to leave bios in flight is when a successful
1099 * partial aio read or full aio write have been setup. In that case
1100 * bio completion will call aio_complete. The only time it's safe to
1101 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1102 * This had *better* be the only place that raises -EIOCBQUEUED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103 */
Zach Brown8459d862006-12-10 02:21:05 -08001104 BUG_ON(ret == -EIOCBQUEUED);
1105 if (dio->is_async && ret == 0 && dio->result &&
1106 ((rw & READ) || (dio->result == dio->size)))
1107 ret = -EIOCBQUEUED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108
Jeff Moyercfb1e332009-10-02 18:57:36 -04001109 if (ret != -EIOCBQUEUED) {
1110 /* All IO is now issued, send it on its way */
1111 blk_run_address_space(inode->i_mapping);
Zach Brown6d544bb2006-12-10 02:20:54 -08001112 dio_await_completion(dio);
Jeff Moyercfb1e332009-10-02 18:57:36 -04001113 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114
Zach Brown8459d862006-12-10 02:21:05 -08001115 /*
1116 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001117 * operation. AIO can if it was a broken operation described above or
1118 * in fact if all the bios race to complete before we get here. In
1119 * that case dio_complete() translates the EIOCBQUEUED into the proper
1120 * return code that the caller will hand to aio_complete().
1121 *
1122 * This is managed by the bio_lock instead of being an atomic_t so that
1123 * completion paths can drop their ref and use the remaining count to
1124 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001125 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001126 spin_lock_irqsave(&dio->bio_lock, flags);
1127 ret2 = --dio->refcount;
1128 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brownfcb82f82007-07-03 15:28:55 -07001129
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001130 if (ret2 == 0) {
Christoph Hellwig40e2e972010-07-18 21:17:09 +00001131 ret = dio_complete(dio, offset, ret, false);
Zach Brown8459d862006-12-10 02:21:05 -08001132 kfree(dio);
1133 } else
1134 BUG_ON(ret != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136 return ret;
1137}
1138
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139ssize_t
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001140__blockdev_direct_IO_newtrunc(int rw, struct kiocb *iocb, struct inode *inode,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 struct block_device *bdev, const struct iovec *iov, loff_t offset,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -08001142 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
Josef Bacikfacd07b2010-05-23 11:00:55 -04001143 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144{
1145 int seg;
1146 size_t size;
1147 unsigned long addr;
1148 unsigned blkbits = inode->i_blkbits;
1149 unsigned bdev_blkbits = 0;
1150 unsigned blocksize_mask = (1 << blkbits) - 1;
1151 ssize_t retval = -EINVAL;
1152 loff_t end = offset;
1153 struct dio *dio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154
1155 if (rw & WRITE)
Vivek Goyald9449ce2009-11-26 09:45:40 +01001156 rw = WRITE_ODIRECT_PLUG;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157
1158 if (bdev)
Martin K. Petersene1defc42009-05-22 17:17:49 -04001159 bdev_blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160
1161 if (offset & blocksize_mask) {
1162 if (bdev)
1163 blkbits = bdev_blkbits;
1164 blocksize_mask = (1 << blkbits) - 1;
1165 if (offset & blocksize_mask)
1166 goto out;
1167 }
1168
1169 /* Check the memory alignment. Blocks cannot straddle pages */
1170 for (seg = 0; seg < nr_segs; seg++) {
1171 addr = (unsigned long)iov[seg].iov_base;
1172 size = iov[seg].iov_len;
1173 end += size;
1174 if ((addr & blocksize_mask) || (size & blocksize_mask)) {
1175 if (bdev)
1176 blkbits = bdev_blkbits;
1177 blocksize_mask = (1 << blkbits) - 1;
1178 if ((addr & blocksize_mask) || (size & blocksize_mask))
1179 goto out;
1180 }
1181 }
1182
Jeff Moyer23aee092009-12-15 16:47:49 -08001183 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184 retval = -ENOMEM;
1185 if (!dio)
1186 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001187 /*
1188 * Believe it or not, zeroing out the page array caused a .5%
1189 * performance regression in a database benchmark. So, we take
1190 * care to only zero out what's needed.
1191 */
1192 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001194 dio->flags = flags;
1195 if (dio->flags & DIO_LOCKING) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 /* watch out for a 0 len io from a tricksy fs */
1197 if (rw == READ && end > offset) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001198 struct address_space *mapping =
1199 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001201 /* will be released by direct_io_worker */
1202 mutex_lock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203
1204 retval = filemap_write_and_wait_range(mapping, offset,
1205 end - 1);
1206 if (retval) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001207 mutex_unlock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208 kfree(dio);
1209 goto out;
1210 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211 }
1212
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001213 /*
1214 * Will be released at I/O completion, possibly in a
1215 * different thread.
1216 */
1217 down_read_non_owner(&inode->i_alloc_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 }
1219
1220 /*
1221 * For file extending writes updating i_size before data
1222 * writeouts complete can expose uninitialized blocks. So
1223 * even for AIO, we need to wait for i/o to complete before
1224 * returning in this case.
1225 */
Jens Axboeb31dc662006-06-13 08:26:10 +02001226 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227 (end > i_size_read(inode)));
1228
1229 retval = direct_io_worker(rw, iocb, inode, iov, offset,
Josef Bacikfacd07b2010-05-23 11:00:55 -04001230 nr_segs, blkbits, get_block, end_io,
1231 submit_io, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001233out:
1234 return retval;
1235}
1236EXPORT_SYMBOL(__blockdev_direct_IO_newtrunc);
1237
1238/*
1239 * This is a library function for use by filesystem drivers.
1240 *
1241 * The locking rules are governed by the flags parameter:
1242 * - if the flags value contains DIO_LOCKING we use a fancy locking
1243 * scheme for dumb filesystems.
1244 * For writes this function is called under i_mutex and returns with
1245 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1246 * taken and dropped again before returning.
1247 * For reads and writes i_alloc_sem is taken in shared mode and released
1248 * on I/O completion (which may happen asynchronously after returning to
1249 * the caller).
1250 *
1251 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1252 * internal locking but rather rely on the filesystem to synchronize
1253 * direct I/O reads/writes versus each other and truncate.
1254 * For reads and writes both i_mutex and i_alloc_sem are not held on
1255 * entry and are never taken.
1256 */
1257ssize_t
1258__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1259 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1260 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1261 dio_submit_t submit_io, int flags)
1262{
1263 ssize_t retval;
1264
1265 retval = __blockdev_direct_IO_newtrunc(rw, iocb, inode, bdev, iov,
1266 offset, nr_segs, get_block, end_io, submit_io, flags);
Dmitri Monakhov0f644152009-01-06 14:40:04 -08001267 /*
1268 * In case of error extending write may have instantiated a few
1269 * blocks outside i_size. Trim these off again for DIO_LOCKING.
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001270 * NOTE: DIO_NO_LOCK/DIO_OWN_LOCK callers have to handle this in
1271 * their own manner. This is a further example of where the old
1272 * truncate sequence is inadequate.
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001273 *
1274 * NOTE: filesystems with their own locking have to handle this
1275 * on their own.
Dmitri Monakhov0f644152009-01-06 14:40:04 -08001276 */
Al Viro06777d32009-12-17 04:52:13 -05001277 if (flags & DIO_LOCKING) {
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001278 if (unlikely((rw & WRITE) && retval < 0)) {
1279 loff_t isize = i_size_read(inode);
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001280 loff_t end = offset + iov_length(iov, nr_segs);
1281
Christoph Hellwig5fe878a2009-12-15 16:47:50 -08001282 if (end > isize)
1283 vmtruncate(inode, isize);
1284 }
Dmitri Monakhov0f644152009-01-06 14:40:04 -08001285 }
1286
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 return retval;
1288}
1289EXPORT_SYMBOL(__blockdev_direct_IO);