Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
direct-io.c
Go to the documentation of this file.
1 /*
2  * fs/direct-io.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  *
6  * O_DIRECT
7  *
8  * 04Jul2002 Andrew Morton
9  * Initial version
10  * 11Sep2002 [email protected]
11  * added readv/writev support.
12  * 29Oct2002 Andrew Morton
13  * rewrote bio_add_page() support.
14  * 30Oct2002 [email protected]
15  * added support for non-aligned IO.
16  * 06Nov2002 [email protected]
17  * added asynchronous IO support.
18  * 21Jul2003 [email protected]
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>
31 #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 <linux/atomic.h>
39 #include <linux/prefetch.h>
40 
41 /*
42  * How many user pages to map in one call to get_user_pages(). This determines
43  * the size of a structure in the slab cache
44  */
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.
57  */
58 
59 /* dio_state only used in the submission path */
60 
61 struct dio_submit {
62  struct bio *bio; /* bio under assembly */
63  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 */
77  int reap_counter; /* rate limit reaping */
78  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 */
81  get_block_t *get_block; /* block mapping function */
82  dio_submit_t *submit_io; /* IO submition function */
83 
84  loff_t logical_offset_in_bio; /* current first logical block in bio */
85  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 */
88 
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 */
98  loff_t cur_page_fs_offset; /* Offset in file */
99 
100  /*
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 */
113 };
114 
115 /* dio_state communicated between submission path and end_io */
116 struct dio {
117  int flags; /* doesn't change */
118  int rw;
119  struct inode *inode;
120  loff_t i_size; /* i_size when submitted */
121  dio_iodone_t *end_io; /* IO completion function */
122 
123  void *private; /* copy from map_bh.b_private */
124 
125  /* BIO completion state */
126  spinlock_t bio_lock; /* protects BIO fields below */
127  int page_errors; /* errno from get_user_pages() */
128  int is_async; /* is IO async ? */
129  int io_error; /* IO error in completion path */
130  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 */
136  ssize_t result; /* IO result */
137 
138  /*
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 */
145 
146 static struct kmem_cache *dio_cache __read_mostly;
147 
148 /*
149  * How many pages are in the queue?
150  */
151 static inline unsigned dio_pages_present(struct dio_submit *sdio)
152 {
153  return sdio->tail - sdio->head;
154 }
155 
156 /*
157  * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
158  */
159 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
160 {
161  int ret;
162  int nr_pages;
163 
164  nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
165  ret = get_user_pages_fast(
166  sdio->curr_user_address, /* Where from? */
167  nr_pages, /* How many pages? */
168  dio->rw == READ, /* Write to memory? */
169  &dio->pages[0]); /* Put results here */
170 
171  if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
172  struct page *page = ZERO_PAGE(0);
173  /*
174  * A memory fault, but the filesystem has some outstanding
175  * mapped blocks. We need to use those blocks up to avoid
176  * leaking stale data in the file.
177  */
178  if (dio->page_errors == 0)
179  dio->page_errors = ret;
180  page_cache_get(page);
181  dio->pages[0] = page;
182  sdio->head = 0;
183  sdio->tail = 1;
184  ret = 0;
185  goto out;
186  }
187 
188  if (ret >= 0) {
189  sdio->curr_user_address += ret * PAGE_SIZE;
190  sdio->curr_page += ret;
191  sdio->head = 0;
192  sdio->tail = ret;
193  ret = 0;
194  }
195 out:
196  return ret;
197 }
198 
199 /*
200  * Get another userspace page. Returns an ERR_PTR on error. Pages are
201  * buffered inside the dio so that we can call get_user_pages() against a
202  * decent number of pages, less frequently. To provide nicer use of the
203  * L1 cache.
204  */
205 static inline struct page *dio_get_page(struct dio *dio,
206  struct dio_submit *sdio)
207 {
208  if (dio_pages_present(sdio) == 0) {
209  int ret;
210 
211  ret = dio_refill_pages(dio, sdio);
212  if (ret)
213  return ERR_PTR(ret);
214  BUG_ON(dio_pages_present(sdio) == 0);
215  }
216  return dio->pages[sdio->head++];
217 }
218 
232 static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async)
233 {
234  ssize_t transferred = 0;
235 
236  /*
237  * AIO submission can race with bio completion to get here while
238  * expecting to have the last io completed by bio completion.
239  * In that case -EIOCBQUEUED is in fact not an error we want
240  * to preserve through this call.
241  */
242  if (ret == -EIOCBQUEUED)
243  ret = 0;
244 
245  if (dio->result) {
246  transferred = dio->result;
247 
248  /* Check for short read case */
249  if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
250  transferred = dio->i_size - offset;
251  }
252 
253  if (ret == 0)
254  ret = dio->page_errors;
255  if (ret == 0)
256  ret = dio->io_error;
257  if (ret == 0)
258  ret = transferred;
259 
260  if (dio->end_io && dio->result) {
261  dio->end_io(dio->iocb, offset, transferred,
262  dio->private, ret, is_async);
263  } else {
264  if (is_async)
265  aio_complete(dio->iocb, ret, 0);
266  inode_dio_done(dio->inode);
267  }
268 
269  return ret;
270 }
271 
272 static int dio_bio_complete(struct dio *dio, struct bio *bio);
273 /*
274  * Asynchronous IO callback.
275  */
276 static void dio_bio_end_aio(struct bio *bio, int error)
277 {
278  struct dio *dio = bio->bi_private;
279  unsigned long remaining;
280  unsigned long flags;
281 
282  /* cleanup the bio */
283  dio_bio_complete(dio, bio);
284 
285  spin_lock_irqsave(&dio->bio_lock, flags);
286  remaining = --dio->refcount;
287  if (remaining == 1 && dio->waiter)
288  wake_up_process(dio->waiter);
289  spin_unlock_irqrestore(&dio->bio_lock, flags);
290 
291  if (remaining == 0) {
292  dio_complete(dio, dio->iocb->ki_pos, 0, true);
293  kmem_cache_free(dio_cache, dio);
294  }
295 }
296 
297 /*
298  * The BIO completion handler simply queues the BIO up for the process-context
299  * handler.
300  *
301  * During I/O bi_private points at the dio. After I/O, bi_private is used to
302  * implement a singly-linked list of completed BIOs, at dio->bio_list.
303  */
304 static void dio_bio_end_io(struct bio *bio, int error)
305 {
306  struct dio *dio = bio->bi_private;
307  unsigned long flags;
308 
309  spin_lock_irqsave(&dio->bio_lock, flags);
310  bio->bi_private = dio->bio_list;
311  dio->bio_list = bio;
312  if (--dio->refcount == 1 && dio->waiter)
313  wake_up_process(dio->waiter);
314  spin_unlock_irqrestore(&dio->bio_lock, flags);
315 }
316 
326 void dio_end_io(struct bio *bio, int error)
327 {
328  struct dio *dio = bio->bi_private;
329 
330  if (dio->is_async)
331  dio_bio_end_aio(bio, error);
332  else
333  dio_bio_end_io(bio, error);
334 }
336 
337 static inline void
338 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
339  struct block_device *bdev,
340  sector_t first_sector, int nr_vecs)
341 {
342  struct bio *bio;
343 
344  /*
345  * bio_alloc() is guaranteed to return a bio when called with
346  * __GFP_WAIT and we request a valid number of vectors.
347  */
348  bio = bio_alloc(GFP_KERNEL, nr_vecs);
349 
350  bio->bi_bdev = bdev;
351  bio->bi_sector = first_sector;
352  if (dio->is_async)
353  bio->bi_end_io = dio_bio_end_aio;
354  else
355  bio->bi_end_io = dio_bio_end_io;
356 
357  sdio->bio = bio;
359 }
360 
361 /*
362  * In the AIO read case we speculatively dirty the pages before starting IO.
363  * During IO completion, any of these pages which happen to have been written
364  * back will be redirtied by bio_check_pages_dirty().
365  *
366  * bios hold a dio reference between submit_bio and ->end_io.
367  */
368 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
369 {
370  struct bio *bio = sdio->bio;
371  unsigned long flags;
372 
373  bio->bi_private = dio;
374 
375  spin_lock_irqsave(&dio->bio_lock, flags);
376  dio->refcount++;
377  spin_unlock_irqrestore(&dio->bio_lock, flags);
378 
379  if (dio->is_async && dio->rw == READ)
380  bio_set_pages_dirty(bio);
381 
382  if (sdio->submit_io)
383  sdio->submit_io(dio->rw, bio, dio->inode,
384  sdio->logical_offset_in_bio);
385  else
386  submit_bio(dio->rw, bio);
387 
388  sdio->bio = NULL;
389  sdio->boundary = 0;
390  sdio->logical_offset_in_bio = 0;
391 }
392 
393 /*
394  * Release any resources in case of a failure
395  */
396 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
397 {
398  while (dio_pages_present(sdio))
399  page_cache_release(dio_get_page(dio, sdio));
400 }
401 
402 /*
403  * Wait for the next BIO to complete. Remove it and return it. NULL is
404  * returned once all BIOs have been completed. This must only be called once
405  * all bios have been issued so that dio->refcount can only decrease. This
406  * requires that that the caller hold a reference on the dio.
407  */
408 static struct bio *dio_await_one(struct dio *dio)
409 {
410  unsigned long flags;
411  struct bio *bio = NULL;
412 
413  spin_lock_irqsave(&dio->bio_lock, flags);
414 
415  /*
416  * Wait as long as the list is empty and there are bios in flight. bio
417  * completion drops the count, maybe adds to the list, and wakes while
418  * holding the bio_lock so we don't need set_current_state()'s barrier
419  * and can call it after testing our condition.
420  */
421  while (dio->refcount > 1 && dio->bio_list == NULL) {
423  dio->waiter = current;
424  spin_unlock_irqrestore(&dio->bio_lock, flags);
425  io_schedule();
426  /* wake up sets us TASK_RUNNING */
427  spin_lock_irqsave(&dio->bio_lock, flags);
428  dio->waiter = NULL;
429  }
430  if (dio->bio_list) {
431  bio = dio->bio_list;
432  dio->bio_list = bio->bi_private;
433  }
434  spin_unlock_irqrestore(&dio->bio_lock, flags);
435  return bio;
436 }
437 
438 /*
439  * Process one completed BIO. No locks are held.
440  */
441 static int dio_bio_complete(struct dio *dio, struct bio *bio)
442 {
443  const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
444  struct bio_vec *bvec = bio->bi_io_vec;
445  int page_no;
446 
447  if (!uptodate)
448  dio->io_error = -EIO;
449 
450  if (dio->is_async && dio->rw == READ) {
451  bio_check_pages_dirty(bio); /* transfers ownership */
452  } else {
453  for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
454  struct page *page = bvec[page_no].bv_page;
455 
456  if (dio->rw == READ && !PageCompound(page))
457  set_page_dirty_lock(page);
458  page_cache_release(page);
459  }
460  bio_put(bio);
461  }
462  return uptodate ? 0 : -EIO;
463 }
464 
465 /*
466  * Wait on and process all in-flight BIOs. This must only be called once
467  * all bios have been issued so that the refcount can only decrease.
468  * This just waits for all bios to make it through dio_bio_complete. IO
469  * errors are propagated through dio->io_error and should be propagated via
470  * dio_complete().
471  */
472 static void dio_await_completion(struct dio *dio)
473 {
474  struct bio *bio;
475  do {
476  bio = dio_await_one(dio);
477  if (bio)
478  dio_bio_complete(dio, bio);
479  } while (bio);
480 }
481 
482 /*
483  * A really large O_DIRECT read or write can generate a lot of BIOs. So
484  * to keep the memory consumption sane we periodically reap any completed BIOs
485  * during the BIO generation phase.
486  *
487  * This also helps to limit the peak amount of pinned userspace memory.
488  */
489 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
490 {
491  int ret = 0;
492 
493  if (sdio->reap_counter++ >= 64) {
494  while (dio->bio_list) {
495  unsigned long flags;
496  struct bio *bio;
497  int ret2;
498 
499  spin_lock_irqsave(&dio->bio_lock, flags);
500  bio = dio->bio_list;
501  dio->bio_list = bio->bi_private;
502  spin_unlock_irqrestore(&dio->bio_lock, flags);
503  ret2 = dio_bio_complete(dio, bio);
504  if (ret == 0)
505  ret = ret2;
506  }
507  sdio->reap_counter = 0;
508  }
509  return ret;
510 }
511 
512 /*
513  * Call into the fs to map some more disk blocks. We record the current number
514  * of available blocks at sdio->blocks_available. These are in units of the
515  * fs blocksize, (1 << inode->i_blkbits).
516  *
517  * The fs is allowed to map lots of blocks at once. If it wants to do that,
518  * it uses the passed inode-relative block number as the file offset, as usual.
519  *
520  * get_block() is passed the number of i_blkbits-sized blocks which direct_io
521  * has remaining to do. The fs should not map more than this number of blocks.
522  *
523  * If the fs has mapped a lot of blocks, it should populate bh->b_size to
524  * indicate how much contiguous disk space has been made available at
525  * bh->b_blocknr.
526  *
527  * If *any* of the mapped blocks are new, then the fs must set buffer_new().
528  * This isn't very efficient...
529  *
530  * In the case of filesystem holes: the fs may return an arbitrarily-large
531  * hole by returning an appropriate value in b_size and by clearing
532  * buffer_mapped(). However the direct-io code will only process holes one
533  * block at a time - it will repeatedly call get_block() as it walks the hole.
534  */
535 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
536  struct buffer_head *map_bh)
537 {
538  int ret;
539  sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
540  sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
541  unsigned long fs_count; /* Number of filesystem-sized blocks */
542  int create;
543  unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
544 
545  /*
546  * If there was a memory error and we've overwritten all the
547  * mapped blocks then we can now return that memory error
548  */
549  ret = dio->page_errors;
550  if (ret == 0) {
552  fs_startblk = sdio->block_in_file >> sdio->blkfactor;
553  fs_endblk = (sdio->final_block_in_request - 1) >>
554  sdio->blkfactor;
555  fs_count = fs_endblk - fs_startblk + 1;
556 
557  map_bh->b_state = 0;
558  map_bh->b_size = fs_count << i_blkbits;
559 
560  /*
561  * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
562  * forbid block creations: only overwrites are permitted.
563  * We will return early to the caller once we see an
564  * unmapped buffer head returned, and the caller will fall
565  * back to buffered I/O.
566  *
567  * Otherwise the decision is left to the get_blocks method,
568  * which may decide to handle it or also return an unmapped
569  * buffer head.
570  */
571  create = dio->rw & WRITE;
572  if (dio->flags & DIO_SKIP_HOLES) {
573  if (sdio->block_in_file < (i_size_read(dio->inode) >>
574  sdio->blkbits))
575  create = 0;
576  }
577 
578  ret = (*sdio->get_block)(dio->inode, fs_startblk,
579  map_bh, create);
580 
581  /* Store for completion */
582  dio->private = map_bh->b_private;
583  }
584  return ret;
585 }
586 
587 /*
588  * There is no bio. Make one now.
589  */
590 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
591  sector_t start_sector, struct buffer_head *map_bh)
592 {
594  int ret, nr_pages;
595 
596  ret = dio_bio_reap(dio, sdio);
597  if (ret)
598  goto out;
599  sector = start_sector << (sdio->blkbits - 9);
600  nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
601  nr_pages = min(nr_pages, BIO_MAX_PAGES);
602  BUG_ON(nr_pages <= 0);
603  dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
604  sdio->boundary = 0;
605 out:
606  return ret;
607 }
608 
609 /*
610  * Attempt to put the current chunk of 'cur_page' into the current BIO. If
611  * that was successful then update final_block_in_bio and take a ref against
612  * the just-added page.
613  *
614  * Return zero on success. Non-zero means the caller needs to start a new BIO.
615  */
616 static inline int dio_bio_add_page(struct dio_submit *sdio)
617 {
618  int ret;
619 
620  ret = bio_add_page(sdio->bio, sdio->cur_page,
621  sdio->cur_page_len, sdio->cur_page_offset);
622  if (ret == sdio->cur_page_len) {
623  /*
624  * Decrement count only, if we are done with this page
625  */
626  if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
627  sdio->pages_in_io--;
628  page_cache_get(sdio->cur_page);
629  sdio->final_block_in_bio = sdio->cur_page_block +
630  (sdio->cur_page_len >> sdio->blkbits);
631  ret = 0;
632  } else {
633  ret = 1;
634  }
635  return ret;
636 }
637 
638 /*
639  * Put cur_page under IO. The section of cur_page which is described by
640  * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
641  * starts on-disk at cur_page_block.
642  *
643  * We take a ref against the page here (on behalf of its presence in the bio).
644  *
645  * The caller of this function is responsible for removing cur_page from the
646  * dio, and for dropping the refcount which came from that presence.
647  */
648 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
649  struct buffer_head *map_bh)
650 {
651  int ret = 0;
652 
653  if (sdio->bio) {
654  loff_t cur_offset = sdio->cur_page_fs_offset;
655  loff_t bio_next_offset = sdio->logical_offset_in_bio +
656  sdio->bio->bi_size;
657 
658  /*
659  * See whether this new request is contiguous with the old.
660  *
661  * Btrfs cannot handle having logically non-contiguous requests
662  * submitted. For example if you have
663  *
664  * Logical: [0-4095][HOLE][8192-12287]
665  * Physical: [0-4095] [4096-8191]
666  *
667  * We cannot submit those pages together as one BIO. So if our
668  * current logical offset in the file does not equal what would
669  * be the next logical offset in the bio, submit the bio we
670  * have.
671  */
672  if (sdio->final_block_in_bio != sdio->cur_page_block ||
673  cur_offset != bio_next_offset)
674  dio_bio_submit(dio, sdio);
675  /*
676  * Submit now if the underlying fs is about to perform a
677  * metadata read
678  */
679  else if (sdio->boundary)
680  dio_bio_submit(dio, sdio);
681  }
682 
683  if (sdio->bio == NULL) {
684  ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
685  if (ret)
686  goto out;
687  }
688 
689  if (dio_bio_add_page(sdio) != 0) {
690  dio_bio_submit(dio, sdio);
691  ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
692  if (ret == 0) {
693  ret = dio_bio_add_page(sdio);
694  BUG_ON(ret != 0);
695  }
696  }
697 out:
698  return ret;
699 }
700 
701 /*
702  * An autonomous function to put a chunk of a page under deferred IO.
703  *
704  * The caller doesn't actually know (or care) whether this piece of page is in
705  * a BIO, or is under IO or whatever. We just take care of all possible
706  * situations here. The separation between the logic of do_direct_IO() and
707  * that of submit_page_section() is important for clarity. Please don't break.
708  *
709  * The chunk of page starts on-disk at blocknr.
710  *
711  * We perform deferred IO, by recording the last-submitted page inside our
712  * private part of the dio structure. If possible, we just expand the IO
713  * across that page here.
714  *
715  * If that doesn't work out then we put the old page into the bio and add this
716  * page to the dio instead.
717  */
718 static inline int
719 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
720  unsigned offset, unsigned len, sector_t blocknr,
721  struct buffer_head *map_bh)
722 {
723  int ret = 0;
724 
725  if (dio->rw & WRITE) {
726  /*
727  * Read accounting is performed in submit_bio()
728  */
729  task_io_account_write(len);
730  }
731 
732  /*
733  * Can we just grow the current page's presence in the dio?
734  */
735  if (sdio->cur_page == page &&
736  sdio->cur_page_offset + sdio->cur_page_len == offset &&
737  sdio->cur_page_block +
738  (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
739  sdio->cur_page_len += len;
740 
741  /*
742  * If sdio->boundary then we want to schedule the IO now to
743  * avoid metadata seeks.
744  */
745  if (sdio->boundary) {
746  ret = dio_send_cur_page(dio, sdio, map_bh);
748  sdio->cur_page = NULL;
749  }
750  goto out;
751  }
752 
753  /*
754  * If there's a deferred page already there then send it.
755  */
756  if (sdio->cur_page) {
757  ret = dio_send_cur_page(dio, sdio, map_bh);
759  sdio->cur_page = NULL;
760  if (ret)
761  goto out;
762  }
763 
764  page_cache_get(page); /* It is in dio */
765  sdio->cur_page = page;
766  sdio->cur_page_offset = offset;
767  sdio->cur_page_len = len;
768  sdio->cur_page_block = blocknr;
769  sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
770 out:
771  return ret;
772 }
773 
774 /*
775  * Clean any dirty buffers in the blockdev mapping which alias newly-created
776  * file blocks. Only called for S_ISREG files - blockdevs do not set
777  * buffer_new
778  */
779 static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
780 {
781  unsigned i;
782  unsigned nblocks;
783 
784  nblocks = map_bh->b_size >> dio->inode->i_blkbits;
785 
786  for (i = 0; i < nblocks; i++) {
787  unmap_underlying_metadata(map_bh->b_bdev,
788  map_bh->b_blocknr + i);
789  }
790 }
791 
792 /*
793  * If we are not writing the entire block and get_block() allocated
794  * the block for us, we need to fill-in the unused portion of the
795  * block with zeros. This happens only if user-buffer, fileoffset or
796  * io length is not filesystem block-size multiple.
797  *
798  * `end' is zero if we're doing the start of the IO, 1 at the end of the
799  * IO.
800  */
801 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
802  int end, struct buffer_head *map_bh)
803 {
804  unsigned dio_blocks_per_fs_block;
805  unsigned this_chunk_blocks; /* In dio_blocks */
806  unsigned this_chunk_bytes;
807  struct page *page;
808 
809  sdio->start_zero_done = 1;
810  if (!sdio->blkfactor || !buffer_new(map_bh))
811  return;
812 
813  dio_blocks_per_fs_block = 1 << sdio->blkfactor;
814  this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
815 
816  if (!this_chunk_blocks)
817  return;
818 
819  /*
820  * We need to zero out part of an fs block. It is either at the
821  * beginning or the end of the fs block.
822  */
823  if (end)
824  this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
825 
826  this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
827 
828  page = ZERO_PAGE(0);
829  if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
830  sdio->next_block_for_io, map_bh))
831  return;
832 
833  sdio->next_block_for_io += this_chunk_blocks;
834 }
835 
836 /*
837  * Walk the user pages, and the file, mapping blocks to disk and generating
838  * a sequence of (page,offset,len,block) mappings. These mappings are injected
839  * into submit_page_section(), which takes care of the next stage of submission
840  *
841  * Direct IO against a blockdev is different from a file. Because we can
842  * happily perform page-sized but 512-byte aligned IOs. It is important that
843  * blockdev IO be able to have fine alignment and large sizes.
844  *
845  * So what we do is to permit the ->get_block function to populate bh.b_size
846  * with the size of IO which is permitted at this offset and this i_blkbits.
847  *
848  * For best results, the blockdev should be set up with 512-byte i_blkbits and
849  * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
850  * fine alignment but still allows this function to work in PAGE_SIZE units.
851  */
852 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
853  struct buffer_head *map_bh)
854 {
855  const unsigned blkbits = sdio->blkbits;
856  const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
857  struct page *page;
858  unsigned block_in_page;
859  int ret = 0;
860 
861  /* The I/O can start at any block offset within the first page */
862  block_in_page = sdio->first_block_in_page;
863 
864  while (sdio->block_in_file < sdio->final_block_in_request) {
865  page = dio_get_page(dio, sdio);
866  if (IS_ERR(page)) {
867  ret = PTR_ERR(page);
868  goto out;
869  }
870 
871  while (block_in_page < blocks_per_page) {
872  unsigned offset_in_page = block_in_page << blkbits;
873  unsigned this_chunk_bytes; /* # of bytes mapped */
874  unsigned this_chunk_blocks; /* # of blocks */
875  unsigned u;
876 
877  if (sdio->blocks_available == 0) {
878  /*
879  * Need to go and map some more disk
880  */
881  unsigned long blkmask;
882  unsigned long dio_remainder;
883 
884  ret = get_more_blocks(dio, sdio, map_bh);
885  if (ret) {
886  page_cache_release(page);
887  goto out;
888  }
889  if (!buffer_mapped(map_bh))
890  goto do_holes;
891 
892  sdio->blocks_available =
893  map_bh->b_size >> sdio->blkbits;
894  sdio->next_block_for_io =
895  map_bh->b_blocknr << sdio->blkfactor;
896  if (buffer_new(map_bh))
897  clean_blockdev_aliases(dio, map_bh);
898 
899  if (!sdio->blkfactor)
900  goto do_holes;
901 
902  blkmask = (1 << sdio->blkfactor) - 1;
903  dio_remainder = (sdio->block_in_file & blkmask);
904 
905  /*
906  * If we are at the start of IO and that IO
907  * starts partway into a fs-block,
908  * dio_remainder will be non-zero. If the IO
909  * is a read then we can simply advance the IO
910  * cursor to the first block which is to be
911  * read. But if the IO is a write and the
912  * block was newly allocated we cannot do that;
913  * the start of the fs block must be zeroed out
914  * on-disk
915  */
916  if (!buffer_new(map_bh))
917  sdio->next_block_for_io += dio_remainder;
918  sdio->blocks_available -= dio_remainder;
919  }
920 do_holes:
921  /* Handle holes */
922  if (!buffer_mapped(map_bh)) {
923  loff_t i_size_aligned;
924 
925  /* AKPM: eargh, -ENOTBLK is a hack */
926  if (dio->rw & WRITE) {
927  page_cache_release(page);
928  return -ENOTBLK;
929  }
930 
931  /*
932  * Be sure to account for a partial block as the
933  * last block in the file
934  */
935  i_size_aligned = ALIGN(i_size_read(dio->inode),
936  1 << blkbits);
937  if (sdio->block_in_file >=
938  i_size_aligned >> blkbits) {
939  /* We hit eof */
940  page_cache_release(page);
941  goto out;
942  }
943  zero_user(page, block_in_page << blkbits,
944  1 << blkbits);
945  sdio->block_in_file++;
946  block_in_page++;
947  goto next_block;
948  }
949 
950  /*
951  * If we're performing IO which has an alignment which
952  * is finer than the underlying fs, go check to see if
953  * we must zero out the start of this block.
954  */
955  if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
956  dio_zero_block(dio, sdio, 0, map_bh);
957 
958  /*
959  * Work out, in this_chunk_blocks, how much disk we
960  * can add to this page
961  */
962  this_chunk_blocks = sdio->blocks_available;
963  u = (PAGE_SIZE - offset_in_page) >> blkbits;
964  if (this_chunk_blocks > u)
965  this_chunk_blocks = u;
966  u = sdio->final_block_in_request - sdio->block_in_file;
967  if (this_chunk_blocks > u)
968  this_chunk_blocks = u;
969  this_chunk_bytes = this_chunk_blocks << blkbits;
970  BUG_ON(this_chunk_bytes == 0);
971 
972  sdio->boundary = buffer_boundary(map_bh);
973  ret = submit_page_section(dio, sdio, page,
974  offset_in_page,
975  this_chunk_bytes,
976  sdio->next_block_for_io,
977  map_bh);
978  if (ret) {
979  page_cache_release(page);
980  goto out;
981  }
982  sdio->next_block_for_io += this_chunk_blocks;
983 
984  sdio->block_in_file += this_chunk_blocks;
985  block_in_page += this_chunk_blocks;
986  sdio->blocks_available -= this_chunk_blocks;
987 next_block:
989  if (sdio->block_in_file == sdio->final_block_in_request)
990  break;
991  }
992 
993  /* Drop the ref which was taken in get_user_pages() */
994  page_cache_release(page);
995  block_in_page = 0;
996  }
997 out:
998  return ret;
999 }
1000 
1001 static inline int drop_refcount(struct dio *dio)
1002 {
1003  int ret2;
1004  unsigned long flags;
1005 
1006  /*
1007  * Sync will always be dropping the final ref and completing the
1008  * operation. AIO can if it was a broken operation described above or
1009  * in fact if all the bios race to complete before we get here. In
1010  * that case dio_complete() translates the EIOCBQUEUED into the proper
1011  * return code that the caller will hand to aio_complete().
1012  *
1013  * This is managed by the bio_lock instead of being an atomic_t so that
1014  * completion paths can drop their ref and use the remaining count to
1015  * decide to wake the submission path atomically.
1016  */
1017  spin_lock_irqsave(&dio->bio_lock, flags);
1018  ret2 = --dio->refcount;
1019  spin_unlock_irqrestore(&dio->bio_lock, flags);
1020  return ret2;
1021 }
1022 
1023 /*
1024  * This is a library function for use by filesystem drivers.
1025  *
1026  * The locking rules are governed by the flags parameter:
1027  * - if the flags value contains DIO_LOCKING we use a fancy locking
1028  * scheme for dumb filesystems.
1029  * For writes this function is called under i_mutex and returns with
1030  * i_mutex held, for reads, i_mutex is not held on entry, but it is
1031  * taken and dropped again before returning.
1032  * - if the flags value does NOT contain DIO_LOCKING we don't use any
1033  * internal locking but rather rely on the filesystem to synchronize
1034  * direct I/O reads/writes versus each other and truncate.
1035  *
1036  * To help with locking against truncate we incremented the i_dio_count
1037  * counter before starting direct I/O, and decrement it once we are done.
1038  * Truncate can wait for it to reach zero to provide exclusion. It is
1039  * expected that filesystem provide exclusion between new direct I/O
1040  * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1041  * but other filesystems need to take care of this on their own.
1042  *
1043  * NOTE: if you pass "sdio" to anything by pointer make sure that function
1044  * is always inlined. Otherwise gcc is unable to split the structure into
1045  * individual fields and will generate much worse code. This is important
1046  * for the whole file.
1047  */
1048 static inline ssize_t
1049 do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1050  struct block_device *bdev, const struct iovec *iov, loff_t offset,
1051  unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1052  dio_submit_t submit_io, int flags)
1053 {
1054  int seg;
1055  size_t size;
1056  unsigned long addr;
1057  unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1058  unsigned blkbits = i_blkbits;
1059  unsigned blocksize_mask = (1 << blkbits) - 1;
1060  ssize_t retval = -EINVAL;
1061  loff_t end = offset;
1062  struct dio *dio;
1063  struct dio_submit sdio = { 0, };
1064  unsigned long user_addr;
1065  size_t bytes;
1066  struct buffer_head map_bh = { 0, };
1067  struct blk_plug plug;
1068 
1069  if (rw & WRITE)
1070  rw = WRITE_ODIRECT;
1071 
1072  /*
1073  * Avoid references to bdev if not absolutely needed to give
1074  * the early prefetch in the caller enough time.
1075  */
1076 
1077  if (offset & blocksize_mask) {
1078  if (bdev)
1079  blkbits = blksize_bits(bdev_logical_block_size(bdev));
1080  blocksize_mask = (1 << blkbits) - 1;
1081  if (offset & blocksize_mask)
1082  goto out;
1083  }
1084 
1085  /* Check the memory alignment. Blocks cannot straddle pages */
1086  for (seg = 0; seg < nr_segs; seg++) {
1087  addr = (unsigned long)iov[seg].iov_base;
1088  size = iov[seg].iov_len;
1089  end += size;
1090  if (unlikely((addr & blocksize_mask) ||
1091  (size & blocksize_mask))) {
1092  if (bdev)
1093  blkbits = blksize_bits(
1094  bdev_logical_block_size(bdev));
1095  blocksize_mask = (1 << blkbits) - 1;
1096  if ((addr & blocksize_mask) || (size & blocksize_mask))
1097  goto out;
1098  }
1099  }
1100 
1101  /* watch out for a 0 len io from a tricksy fs */
1102  if (rw == READ && end == offset)
1103  return 0;
1104 
1105  dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1106  retval = -ENOMEM;
1107  if (!dio)
1108  goto out;
1109  /*
1110  * Believe it or not, zeroing out the page array caused a .5%
1111  * performance regression in a database benchmark. So, we take
1112  * care to only zero out what's needed.
1113  */
1114  memset(dio, 0, offsetof(struct dio, pages));
1115 
1116  dio->flags = flags;
1117  if (dio->flags & DIO_LOCKING) {
1118  if (rw == READ) {
1119  struct address_space *mapping =
1120  iocb->ki_filp->f_mapping;
1121 
1122  /* will be released by direct_io_worker */
1123  mutex_lock(&inode->i_mutex);
1124 
1125  retval = filemap_write_and_wait_range(mapping, offset,
1126  end - 1);
1127  if (retval) {
1128  mutex_unlock(&inode->i_mutex);
1129  kmem_cache_free(dio_cache, dio);
1130  goto out;
1131  }
1132  }
1133  }
1134 
1135  /*
1136  * Will be decremented at I/O completion time.
1137  */
1138  atomic_inc(&inode->i_dio_count);
1139 
1140  /*
1141  * For file extending writes updating i_size before data
1142  * writeouts complete can expose uninitialized blocks. So
1143  * even for AIO, we need to wait for i/o to complete before
1144  * returning in this case.
1145  */
1146  dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
1147  (end > i_size_read(inode)));
1148 
1149  retval = 0;
1150 
1151  dio->inode = inode;
1152  dio->rw = rw;
1153  sdio.blkbits = blkbits;
1154  sdio.blkfactor = i_blkbits - blkbits;
1155  sdio.block_in_file = offset >> blkbits;
1156 
1157  sdio.get_block = get_block;
1158  dio->end_io = end_io;
1159  sdio.submit_io = submit_io;
1160  sdio.final_block_in_bio = -1;
1161  sdio.next_block_for_io = -1;
1162 
1163  dio->iocb = iocb;
1164  dio->i_size = i_size_read(inode);
1165 
1166  spin_lock_init(&dio->bio_lock);
1167  dio->refcount = 1;
1168 
1169  /*
1170  * In case of non-aligned buffers, we may need 2 more
1171  * pages since we need to zero out first and last block.
1172  */
1173  if (unlikely(sdio.blkfactor))
1174  sdio.pages_in_io = 2;
1175 
1176  for (seg = 0; seg < nr_segs; seg++) {
1177  user_addr = (unsigned long)iov[seg].iov_base;
1178  sdio.pages_in_io +=
1179  ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) /
1180  PAGE_SIZE - user_addr / PAGE_SIZE);
1181  }
1182 
1183  blk_start_plug(&plug);
1184 
1185  for (seg = 0; seg < nr_segs; seg++) {
1186  user_addr = (unsigned long)iov[seg].iov_base;
1187  sdio.size += bytes = iov[seg].iov_len;
1188 
1189  /* Index into the first page of the first block */
1190  sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1192  (bytes >> blkbits);
1193  /* Page fetching state */
1194  sdio.head = 0;
1195  sdio.tail = 0;
1196  sdio.curr_page = 0;
1197 
1198  sdio.total_pages = 0;
1199  if (user_addr & (PAGE_SIZE-1)) {
1200  sdio.total_pages++;
1201  bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1202  }
1203  sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1204  sdio.curr_user_address = user_addr;
1205 
1206  retval = do_direct_IO(dio, &sdio, &map_bh);
1207 
1208  dio->result += iov[seg].iov_len -
1209  ((sdio.final_block_in_request - sdio.block_in_file) <<
1210  blkbits);
1211 
1212  if (retval) {
1213  dio_cleanup(dio, &sdio);
1214  break;
1215  }
1216  } /* end iovec loop */
1217 
1218  if (retval == -ENOTBLK) {
1219  /*
1220  * The remaining part of the request will be
1221  * be handled by buffered I/O when we return
1222  */
1223  retval = 0;
1224  }
1225  /*
1226  * There may be some unwritten disk at the end of a part-written
1227  * fs-block-sized block. Go zero that now.
1228  */
1229  dio_zero_block(dio, &sdio, 1, &map_bh);
1230 
1231  if (sdio.cur_page) {
1232  ssize_t ret2;
1233 
1234  ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1235  if (retval == 0)
1236  retval = ret2;
1238  sdio.cur_page = NULL;
1239  }
1240  if (sdio.bio)
1241  dio_bio_submit(dio, &sdio);
1242 
1243  blk_finish_plug(&plug);
1244 
1245  /*
1246  * It is possible that, we return short IO due to end of file.
1247  * In that case, we need to release all the pages we got hold on.
1248  */
1249  dio_cleanup(dio, &sdio);
1250 
1251  /*
1252  * All block lookups have been performed. For READ requests
1253  * we can let i_mutex go now that its achieved its purpose
1254  * of protecting us from looking up uninitialized blocks.
1255  */
1256  if (rw == READ && (dio->flags & DIO_LOCKING))
1257  mutex_unlock(&dio->inode->i_mutex);
1258 
1259  /*
1260  * The only time we want to leave bios in flight is when a successful
1261  * partial aio read or full aio write have been setup. In that case
1262  * bio completion will call aio_complete. The only time it's safe to
1263  * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1264  * This had *better* be the only place that raises -EIOCBQUEUED.
1265  */
1266  BUG_ON(retval == -EIOCBQUEUED);
1267  if (dio->is_async && retval == 0 && dio->result &&
1268  ((rw == READ) || (dio->result == sdio.size)))
1269  retval = -EIOCBQUEUED;
1270 
1271  if (retval != -EIOCBQUEUED)
1272  dio_await_completion(dio);
1273 
1274  if (drop_refcount(dio) == 0) {
1275  retval = dio_complete(dio, offset, retval, false);
1276  kmem_cache_free(dio_cache, dio);
1277  } else
1278  BUG_ON(retval != -EIOCBQUEUED);
1279 
1280 out:
1281  return retval;
1282 }
1283 
1284 ssize_t
1285 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1286  struct block_device *bdev, const struct iovec *iov, loff_t offset,
1287  unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1288  dio_submit_t submit_io, int flags)
1289 {
1290  /*
1291  * The block device state is needed in the end to finally
1292  * submit everything. Since it's likely to be cache cold
1293  * prefetch it here as first thing to hide some of the
1294  * latency.
1295  *
1296  * Attempt to prefetch the pieces we likely need later.
1297  */
1298  prefetch(&bdev->bd_disk->part_tbl);
1299  prefetch(bdev->bd_queue);
1300  prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1301 
1302  return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset,
1303  nr_segs, get_block, end_io,
1304  submit_io, flags);
1305 }
1306 
1308 
1309 static __init int dio_init(void)
1310 {
1311  dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1312  return 0;
1313 }
1314 module_init(dio_init)