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 janetinc@us.ibm.com
11 * added readv/writev support.
12 * 29Oct2002 Andrew Morton
13 * 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>
30#include <linux/task_io_accounting_ops.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 * Flags for dio_complete()
49 */
50#define DIO_COMPLETE_ASYNC 0x01 /* This is async IO */
51#define DIO_COMPLETE_INVALIDATE 0x02 /* Can invalidate pages */
52
53/*
54 * This code generally works in units of "dio_blocks". A dio_block is
55 * somewhere between the hard sector size and the filesystem block size. it
56 * is determined on a per-invocation basis. When talking to the filesystem
57 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
58 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
59 * to bio_block quantities by shifting left by blkfactor.
60 *
61 * If blkfactor is zero then the user's request was aligned to the filesystem's
62 * blocksize.
63 */
64
65/* dio_state only used in the submission path */
66
67struct dio_submit {
68 struct bio *bio; /* bio under assembly */
69 unsigned blkbits; /* doesn't change */
70 unsigned blkfactor; /* When we're using an alignment which
71 is finer than the filesystem's soft
72 blocksize, this specifies how much
73 finer. blkfactor=2 means 1/4-block
74 alignment. Does not change */
75 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
76 been performed at the start of a
77 write */
78 int pages_in_io; /* approximate total IO pages */
79 sector_t block_in_file; /* Current offset into the underlying
80 file in dio_block units. */
81 unsigned blocks_available; /* At block_in_file. changes */
82 int reap_counter; /* rate limit reaping */
83 sector_t final_block_in_request;/* doesn't change */
84 int boundary; /* prev block is at a boundary */
85 get_block_t *get_block; /* block mapping function */
86 dio_submit_t *submit_io; /* IO submition function */
87
88 loff_t logical_offset_in_bio; /* current first logical block in bio */
89 sector_t final_block_in_bio; /* current final block in bio + 1 */
90 sector_t next_block_for_io; /* next block to be put under IO,
91 in dio_blocks units */
92
93 /*
94 * Deferred addition of a page to the dio. These variables are
95 * private to dio_send_cur_page(), submit_page_section() and
96 * dio_bio_add_page().
97 */
98 struct page *cur_page; /* The page */
99 unsigned cur_page_offset; /* Offset into it, in bytes */
100 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
101 sector_t cur_page_block; /* Where it starts */
102 loff_t cur_page_fs_offset; /* Offset in file */
103
104 struct iov_iter *iter;
105 /*
106 * Page queue. These variables belong to dio_refill_pages() and
107 * dio_get_page().
108 */
109 unsigned head; /* next page to process */
110 unsigned tail; /* last valid page + 1 */
111 size_t from, to;
112};
113
114/* dio_state communicated between submission path and end_io */
115struct dio {
116 int flags; /* doesn't change */
117 int op;
118 int op_flags;
119 blk_qc_t bio_cookie;
120 struct gendisk *bio_disk;
121 struct inode *inode;
122 loff_t i_size; /* i_size when submitted */
123 dio_iodone_t *end_io; /* IO completion function */
124
125 void *private; /* copy from map_bh.b_private */
126
127 /* BIO completion state */
128 spinlock_t bio_lock; /* protects BIO fields below */
129 int page_errors; /* errno from get_user_pages() */
130 int is_async; /* is IO async ? */
131 bool defer_completion; /* defer AIO completion to workqueue? */
132 bool should_dirty; /* if pages should be dirtied */
133 int io_error; /* IO error in completion path */
134 unsigned long refcount; /* direct_io_worker() and bios */
135 struct bio *bio_list; /* singly linked via bi_private */
136 struct task_struct *waiter; /* waiting task (NULL if none) */
137
138 /* AIO related stuff */
139 struct kiocb *iocb; /* kiocb */
140 ssize_t result; /* IO result */
141
142 /*
143 * pages[] (and any fields placed after it) are not zeroed out at
144 * allocation time. Don't add new fields after pages[] unless you
145 * wish that they not be zeroed.
146 */
147 union {
148 struct page *pages[DIO_PAGES]; /* page buffer */
149 struct work_struct complete_work;/* deferred AIO completion */
150 };
151} ____cacheline_aligned_in_smp;
152
153static struct kmem_cache *dio_cache __read_mostly;
154
155/*
156 * How many pages are in the queue?
157 */
158static inline unsigned dio_pages_present(struct dio_submit *sdio)
159{
160 return sdio->tail - sdio->head;
161}
162
163/*
164 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
165 */
166static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
167{
168 ssize_t ret;
169
170 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
171 &sdio->from);
172
173 if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
174 struct page *page = ZERO_PAGE(0);
175 /*
176 * A memory fault, but the filesystem has some outstanding
177 * mapped blocks. We need to use those blocks up to avoid
178 * leaking stale data in the file.
179 */
180 if (dio->page_errors == 0)
181 dio->page_errors = ret;
182 get_page(page);
183 dio->pages[0] = page;
184 sdio->head = 0;
185 sdio->tail = 1;
186 sdio->from = 0;
187 sdio->to = PAGE_SIZE;
188 return 0;
189 }
190
191 if (ret >= 0) {
192 iov_iter_advance(sdio->iter, ret);
193 ret += sdio->from;
194 sdio->head = 0;
195 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
196 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
197 return 0;
198 }
199 return ret;
200}
201
202/*
203 * Get another userspace page. Returns an ERR_PTR on error. Pages are
204 * buffered inside the dio so that we can call get_user_pages() against a
205 * decent number of pages, less frequently. To provide nicer use of the
206 * L1 cache.
207 */
208static inline struct page *dio_get_page(struct dio *dio,
209 struct dio_submit *sdio)
210{
211 if (dio_pages_present(sdio) == 0) {
212 int ret;
213
214 ret = dio_refill_pages(dio, sdio);
215 if (ret)
216 return ERR_PTR(ret);
217 BUG_ON(dio_pages_present(sdio) == 0);
218 }
219 return dio->pages[sdio->head];
220}
221
222/*
223 * Warn about a page cache invalidation failure during a direct io write.
224 */
225void dio_warn_stale_pagecache(struct file *filp)
226{
227 static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST);
228 char pathname[128];
229 struct inode *inode = file_inode(filp);
230 char *path;
231
232 errseq_set(&inode->i_mapping->wb_err, -EIO);
233 if (__ratelimit(&_rs)) {
234 path = file_path(filp, pathname, sizeof(pathname));
235 if (IS_ERR(path))
236 path = "(unknown)";
237 pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n");
238 pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid,
239 current->comm);
240 }
241}
242
243/**
244 * dio_complete() - called when all DIO BIO I/O has been completed
245 * @offset: the byte offset in the file of the completed operation
246 *
247 * This drops i_dio_count, lets interested parties know that a DIO operation
248 * has completed, and calculates the resulting return code for the operation.
249 *
250 * It lets the filesystem know if it registered an interest earlier via
251 * get_block. Pass the private field of the map buffer_head so that
252 * filesystems can use it to hold additional state between get_block calls and
253 * dio_complete.
254 */
255static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
256{
257 loff_t offset = dio->iocb->ki_pos;
258 ssize_t transferred = 0;
259 int err;
260
261 /*
262 * AIO submission can race with bio completion to get here while
263 * expecting to have the last io completed by bio completion.
264 * In that case -EIOCBQUEUED is in fact not an error we want
265 * to preserve through this call.
266 */
267 if (ret == -EIOCBQUEUED)
268 ret = 0;
269
270 if (dio->result) {
271 transferred = dio->result;
272
273 /* Check for short read case */
274 if ((dio->op == REQ_OP_READ) &&
275 ((offset + transferred) > dio->i_size))
276 transferred = dio->i_size - offset;
277 /* ignore EFAULT if some IO has been done */
278 if (unlikely(ret == -EFAULT) && transferred)
279 ret = 0;
280 }
281
282 if (ret == 0)
283 ret = dio->page_errors;
284 if (ret == 0)
285 ret = dio->io_error;
286 if (ret == 0)
287 ret = transferred;
288
289 if (dio->end_io) {
290 // XXX: ki_pos??
291 err = dio->end_io(dio->iocb, offset, ret, dio->private);
292 if (err)
293 ret = err;
294 }
295
296 /*
297 * Try again to invalidate clean pages which might have been cached by
298 * non-direct readahead, or faulted in by get_user_pages() if the source
299 * of the write was an mmap'ed region of the file we're writing. Either
300 * one is a pretty crazy thing to do, so we don't support it 100%. If
301 * this invalidation fails, tough, the write still worked...
302 *
303 * And this page cache invalidation has to be after dio->end_io(), as
304 * some filesystems convert unwritten extents to real allocations in
305 * end_io() when necessary, otherwise a racing buffer read would cache
306 * zeros from unwritten extents.
307 */
308 if (flags & DIO_COMPLETE_INVALIDATE &&
309 ret > 0 && dio->op == REQ_OP_WRITE &&
310 dio->inode->i_mapping->nrpages) {
311 err = invalidate_inode_pages2_range(dio->inode->i_mapping,
312 offset >> PAGE_SHIFT,
313 (offset + ret - 1) >> PAGE_SHIFT);
314 if (err)
315 dio_warn_stale_pagecache(dio->iocb->ki_filp);
316 }
317
318 inode_dio_end(dio->inode);
319
320 if (flags & DIO_COMPLETE_ASYNC) {
321 /*
322 * generic_write_sync expects ki_pos to have been updated
323 * already, but the submission path only does this for
324 * synchronous I/O.
325 */
326 dio->iocb->ki_pos += transferred;
327
328 if (ret > 0 && dio->op == REQ_OP_WRITE)
329 ret = generic_write_sync(dio->iocb, ret);
330 dio->iocb->ki_complete(dio->iocb, ret, 0);
331 }
332
333 kmem_cache_free(dio_cache, dio);
334 return ret;
335}
336
337static void dio_aio_complete_work(struct work_struct *work)
338{
339 struct dio *dio = container_of(work, struct dio, complete_work);
340
341 dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
342}
343
344static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
345
346/*
347 * Asynchronous IO callback.
348 */
349static void dio_bio_end_aio(struct bio *bio)
350{
351 struct dio *dio = bio->bi_private;
352 unsigned long remaining;
353 unsigned long flags;
354 bool defer_completion = false;
355
356 /* cleanup the bio */
357 dio_bio_complete(dio, bio);
358
359 spin_lock_irqsave(&dio->bio_lock, flags);
360 remaining = --dio->refcount;
361 if (remaining == 1 && dio->waiter)
362 wake_up_process(dio->waiter);
363 spin_unlock_irqrestore(&dio->bio_lock, flags);
364
365 if (remaining == 0) {
366 /*
367 * Defer completion when defer_completion is set or
368 * when the inode has pages mapped and this is AIO write.
369 * We need to invalidate those pages because there is a
370 * chance they contain stale data in the case buffered IO
371 * went in between AIO submission and completion into the
372 * same region.
373 */
374 if (dio->result)
375 defer_completion = dio->defer_completion ||
376 (dio->op == REQ_OP_WRITE &&
377 dio->inode->i_mapping->nrpages);
378 if (defer_completion) {
379 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
380 queue_work(dio->inode->i_sb->s_dio_done_wq,
381 &dio->complete_work);
382 } else {
383 dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
384 }
385 }
386}
387
388/*
389 * The BIO completion handler simply queues the BIO up for the process-context
390 * handler.
391 *
392 * During I/O bi_private points at the dio. After I/O, bi_private is used to
393 * implement a singly-linked list of completed BIOs, at dio->bio_list.
394 */
395static void dio_bio_end_io(struct bio *bio)
396{
397 struct dio *dio = bio->bi_private;
398 unsigned long flags;
399
400 spin_lock_irqsave(&dio->bio_lock, flags);
401 bio->bi_private = dio->bio_list;
402 dio->bio_list = bio;
403 if (--dio->refcount == 1 && dio->waiter)
404 wake_up_process(dio->waiter);
405 spin_unlock_irqrestore(&dio->bio_lock, flags);
406}
407
408/**
409 * dio_end_io - handle the end io action for the given bio
410 * @bio: The direct io bio thats being completed
411 *
412 * This is meant to be called by any filesystem that uses their own dio_submit_t
413 * so that the DIO specific endio actions are dealt with after the filesystem
414 * has done it's completion work.
415 */
416void dio_end_io(struct bio *bio)
417{
418 struct dio *dio = bio->bi_private;
419
420 if (dio->is_async)
421 dio_bio_end_aio(bio);
422 else
423 dio_bio_end_io(bio);
424}
425EXPORT_SYMBOL_GPL(dio_end_io);
426
427static inline void
428dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
429 struct block_device *bdev,
430 sector_t first_sector, int nr_vecs)
431{
432 struct bio *bio;
433
434 /*
435 * bio_alloc() is guaranteed to return a bio when allowed to sleep and
436 * we request a valid number of vectors.
437 */
438 bio = bio_alloc(GFP_KERNEL, nr_vecs);
439
440 bio_set_dev(bio, bdev);
441 bio->bi_iter.bi_sector = first_sector;
442 bio_set_op_attrs(bio, dio->op, dio->op_flags);
443 if (dio->is_async)
444 bio->bi_end_io = dio_bio_end_aio;
445 else
446 bio->bi_end_io = dio_bio_end_io;
447
448 bio->bi_write_hint = dio->iocb->ki_hint;
449
450 sdio->bio = bio;
451 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
452}
453
454/*
455 * In the AIO read case we speculatively dirty the pages before starting IO.
456 * During IO completion, any of these pages which happen to have been written
457 * back will be redirtied by bio_check_pages_dirty().
458 *
459 * bios hold a dio reference between submit_bio and ->end_io.
460 */
461static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
462{
463 struct bio *bio = sdio->bio;
464 unsigned long flags;
465
466 bio->bi_private = dio;
467
468 spin_lock_irqsave(&dio->bio_lock, flags);
469 dio->refcount++;
470 spin_unlock_irqrestore(&dio->bio_lock, flags);
471
472 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
473 bio_set_pages_dirty(bio);
474
475 dio->bio_disk = bio->bi_disk;
476
477 if (sdio->submit_io) {
478 sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
479 dio->bio_cookie = BLK_QC_T_NONE;
480 } else
481 dio->bio_cookie = submit_bio(bio);
482
483 sdio->bio = NULL;
484 sdio->boundary = 0;
485 sdio->logical_offset_in_bio = 0;
486}
487
488/*
489 * Release any resources in case of a failure
490 */
491static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
492{
493 while (sdio->head < sdio->tail)
494 put_page(dio->pages[sdio->head++]);
495}
496
497/*
498 * Wait for the next BIO to complete. Remove it and return it. NULL is
499 * returned once all BIOs have been completed. This must only be called once
500 * all bios have been issued so that dio->refcount can only decrease. This
501 * requires that that the caller hold a reference on the dio.
502 */
503static struct bio *dio_await_one(struct dio *dio)
504{
505 unsigned long flags;
506 struct bio *bio = NULL;
507
508 spin_lock_irqsave(&dio->bio_lock, flags);
509
510 /*
511 * Wait as long as the list is empty and there are bios in flight. bio
512 * completion drops the count, maybe adds to the list, and wakes while
513 * holding the bio_lock so we don't need set_current_state()'s barrier
514 * and can call it after testing our condition.
515 */
516 while (dio->refcount > 1 && dio->bio_list == NULL) {
517 __set_current_state(TASK_UNINTERRUPTIBLE);
518 dio->waiter = current;
519 spin_unlock_irqrestore(&dio->bio_lock, flags);
520 if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
521 !blk_poll(dio->bio_disk->queue, dio->bio_cookie, true))
522 io_schedule();
523 /* wake up sets us TASK_RUNNING */
524 spin_lock_irqsave(&dio->bio_lock, flags);
525 dio->waiter = NULL;
526 }
527 if (dio->bio_list) {
528 bio = dio->bio_list;
529 dio->bio_list = bio->bi_private;
530 }
531 spin_unlock_irqrestore(&dio->bio_lock, flags);
532 return bio;
533}
534
535/*
536 * Process one completed BIO. No locks are held.
537 */
538static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
539{
540 struct bio_vec *bvec;
541 unsigned i;
542 blk_status_t err = bio->bi_status;
543
544 if (err) {
545 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
546 dio->io_error = -EAGAIN;
547 else
548 dio->io_error = -EIO;
549 }
550
551 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) {
552 bio_check_pages_dirty(bio); /* transfers ownership */
553 } else {
554 struct bvec_iter_all iter_all;
555
556 bio_for_each_segment_all(bvec, bio, i, iter_all) {
557 struct page *page = bvec->bv_page;
558
559 if (dio->op == REQ_OP_READ && !PageCompound(page) &&
560 dio->should_dirty)
561 set_page_dirty_lock(page);
562 put_page(page);
563 }
564 bio_put(bio);
565 }
566 return err;
567}
568
569/*
570 * Wait on and process all in-flight BIOs. This must only be called once
571 * all bios have been issued so that the refcount can only decrease.
572 * This just waits for all bios to make it through dio_bio_complete. IO
573 * errors are propagated through dio->io_error and should be propagated via
574 * dio_complete().
575 */
576static void dio_await_completion(struct dio *dio)
577{
578 struct bio *bio;
579 do {
580 bio = dio_await_one(dio);
581 if (bio)
582 dio_bio_complete(dio, bio);
583 } while (bio);
584}
585
586/*
587 * A really large O_DIRECT read or write can generate a lot of BIOs. So
588 * to keep the memory consumption sane we periodically reap any completed BIOs
589 * during the BIO generation phase.
590 *
591 * This also helps to limit the peak amount of pinned userspace memory.
592 */
593static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
594{
595 int ret = 0;
596
597 if (sdio->reap_counter++ >= 64) {
598 while (dio->bio_list) {
599 unsigned long flags;
600 struct bio *bio;
601 int ret2;
602
603 spin_lock_irqsave(&dio->bio_lock, flags);
604 bio = dio->bio_list;
605 dio->bio_list = bio->bi_private;
606 spin_unlock_irqrestore(&dio->bio_lock, flags);
607 ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
608 if (ret == 0)
609 ret = ret2;
610 }
611 sdio->reap_counter = 0;
612 }
613 return ret;
614}
615
616/*
617 * Create workqueue for deferred direct IO completions. We allocate the
618 * workqueue when it's first needed. This avoids creating workqueue for
619 * filesystems that don't need it and also allows us to create the workqueue
620 * late enough so the we can include s_id in the name of the workqueue.
621 */
622int sb_init_dio_done_wq(struct super_block *sb)
623{
624 struct workqueue_struct *old;
625 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
626 WQ_MEM_RECLAIM, 0,
627 sb->s_id);
628 if (!wq)
629 return -ENOMEM;
630 /*
631 * This has to be atomic as more DIOs can race to create the workqueue
632 */
633 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
634 /* Someone created workqueue before us? Free ours... */
635 if (old)
636 destroy_workqueue(wq);
637 return 0;
638}
639
640static int dio_set_defer_completion(struct dio *dio)
641{
642 struct super_block *sb = dio->inode->i_sb;
643
644 if (dio->defer_completion)
645 return 0;
646 dio->defer_completion = true;
647 if (!sb->s_dio_done_wq)
648 return sb_init_dio_done_wq(sb);
649 return 0;
650}
651
652/*
653 * Call into the fs to map some more disk blocks. We record the current number
654 * of available blocks at sdio->blocks_available. These are in units of the
655 * fs blocksize, i_blocksize(inode).
656 *
657 * The fs is allowed to map lots of blocks at once. If it wants to do that,
658 * it uses the passed inode-relative block number as the file offset, as usual.
659 *
660 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
661 * has remaining to do. The fs should not map more than this number of blocks.
662 *
663 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
664 * indicate how much contiguous disk space has been made available at
665 * bh->b_blocknr.
666 *
667 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
668 * This isn't very efficient...
669 *
670 * In the case of filesystem holes: the fs may return an arbitrarily-large
671 * hole by returning an appropriate value in b_size and by clearing
672 * buffer_mapped(). However the direct-io code will only process holes one
673 * block at a time - it will repeatedly call get_block() as it walks the hole.
674 */
675static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
676 struct buffer_head *map_bh)
677{
678 int ret;
679 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
680 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
681 unsigned long fs_count; /* Number of filesystem-sized blocks */
682 int create;
683 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
684 loff_t i_size;
685
686 /*
687 * If there was a memory error and we've overwritten all the
688 * mapped blocks then we can now return that memory error
689 */
690 ret = dio->page_errors;
691 if (ret == 0) {
692 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
693 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
694 fs_endblk = (sdio->final_block_in_request - 1) >>
695 sdio->blkfactor;
696 fs_count = fs_endblk - fs_startblk + 1;
697
698 map_bh->b_state = 0;
699 map_bh->b_size = fs_count << i_blkbits;
700
701 /*
702 * For writes that could fill holes inside i_size on a
703 * DIO_SKIP_HOLES filesystem we forbid block creations: only
704 * overwrites are permitted. We will return early to the caller
705 * once we see an unmapped buffer head returned, and the caller
706 * will fall back to buffered I/O.
707 *
708 * Otherwise the decision is left to the get_blocks method,
709 * which may decide to handle it or also return an unmapped
710 * buffer head.
711 */
712 create = dio->op == REQ_OP_WRITE;
713 if (dio->flags & DIO_SKIP_HOLES) {
714 i_size = i_size_read(dio->inode);
715 if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
716 create = 0;
717 }
718
719 ret = (*sdio->get_block)(dio->inode, fs_startblk,
720 map_bh, create);
721
722 /* Store for completion */
723 dio->private = map_bh->b_private;
724
725 if (ret == 0 && buffer_defer_completion(map_bh))
726 ret = dio_set_defer_completion(dio);
727 }
728 return ret;
729}
730
731/*
732 * There is no bio. Make one now.
733 */
734static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
735 sector_t start_sector, struct buffer_head *map_bh)
736{
737 sector_t sector;
738 int ret, nr_pages;
739
740 ret = dio_bio_reap(dio, sdio);
741 if (ret)
742 goto out;
743 sector = start_sector << (sdio->blkbits - 9);
744 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
745 BUG_ON(nr_pages <= 0);
746 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
747 sdio->boundary = 0;
748out:
749 return ret;
750}
751
752/*
753 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
754 * that was successful then update final_block_in_bio and take a ref against
755 * the just-added page.
756 *
757 * Return zero on success. Non-zero means the caller needs to start a new BIO.
758 */
759static inline int dio_bio_add_page(struct dio_submit *sdio)
760{
761 int ret;
762
763 ret = bio_add_page(sdio->bio, sdio->cur_page,
764 sdio->cur_page_len, sdio->cur_page_offset);
765 if (ret == sdio->cur_page_len) {
766 /*
767 * Decrement count only, if we are done with this page
768 */
769 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
770 sdio->pages_in_io--;
771 get_page(sdio->cur_page);
772 sdio->final_block_in_bio = sdio->cur_page_block +
773 (sdio->cur_page_len >> sdio->blkbits);
774 ret = 0;
775 } else {
776 ret = 1;
777 }
778 return ret;
779}
780
781/*
782 * Put cur_page under IO. The section of cur_page which is described by
783 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
784 * starts on-disk at cur_page_block.
785 *
786 * We take a ref against the page here (on behalf of its presence in the bio).
787 *
788 * The caller of this function is responsible for removing cur_page from the
789 * dio, and for dropping the refcount which came from that presence.
790 */
791static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
792 struct buffer_head *map_bh)
793{
794 int ret = 0;
795
796 if (sdio->bio) {
797 loff_t cur_offset = sdio->cur_page_fs_offset;
798 loff_t bio_next_offset = sdio->logical_offset_in_bio +
799 sdio->bio->bi_iter.bi_size;
800
801 /*
802 * See whether this new request is contiguous with the old.
803 *
804 * Btrfs cannot handle having logically non-contiguous requests
805 * submitted. For example if you have
806 *
807 * Logical: [0-4095][HOLE][8192-12287]
808 * Physical: [0-4095] [4096-8191]
809 *
810 * We cannot submit those pages together as one BIO. So if our
811 * current logical offset in the file does not equal what would
812 * be the next logical offset in the bio, submit the bio we
813 * have.
814 */
815 if (sdio->final_block_in_bio != sdio->cur_page_block ||
816 cur_offset != bio_next_offset)
817 dio_bio_submit(dio, sdio);
818 }
819
820 if (sdio->bio == NULL) {
821 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
822 if (ret)
823 goto out;
824 }
825
826 if (dio_bio_add_page(sdio) != 0) {
827 dio_bio_submit(dio, sdio);
828 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
829 if (ret == 0) {
830 ret = dio_bio_add_page(sdio);
831 BUG_ON(ret != 0);
832 }
833 }
834out:
835 return ret;
836}
837
838/*
839 * An autonomous function to put a chunk of a page under deferred IO.
840 *
841 * The caller doesn't actually know (or care) whether this piece of page is in
842 * a BIO, or is under IO or whatever. We just take care of all possible
843 * situations here. The separation between the logic of do_direct_IO() and
844 * that of submit_page_section() is important for clarity. Please don't break.
845 *
846 * The chunk of page starts on-disk at blocknr.
847 *
848 * We perform deferred IO, by recording the last-submitted page inside our
849 * private part of the dio structure. If possible, we just expand the IO
850 * across that page here.
851 *
852 * If that doesn't work out then we put the old page into the bio and add this
853 * page to the dio instead.
854 */
855static inline int
856submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
857 unsigned offset, unsigned len, sector_t blocknr,
858 struct buffer_head *map_bh)
859{
860 int ret = 0;
861
862 if (dio->op == REQ_OP_WRITE) {
863 /*
864 * Read accounting is performed in submit_bio()
865 */
866 task_io_account_write(len);
867 }
868
869 /*
870 * Can we just grow the current page's presence in the dio?
871 */
872 if (sdio->cur_page == page &&
873 sdio->cur_page_offset + sdio->cur_page_len == offset &&
874 sdio->cur_page_block +
875 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
876 sdio->cur_page_len += len;
877 goto out;
878 }
879
880 /*
881 * If there's a deferred page already there then send it.
882 */
883 if (sdio->cur_page) {
884 ret = dio_send_cur_page(dio, sdio, map_bh);
885 put_page(sdio->cur_page);
886 sdio->cur_page = NULL;
887 if (ret)
888 return ret;
889 }
890
891 get_page(page); /* It is in dio */
892 sdio->cur_page = page;
893 sdio->cur_page_offset = offset;
894 sdio->cur_page_len = len;
895 sdio->cur_page_block = blocknr;
896 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
897out:
898 /*
899 * If sdio->boundary then we want to schedule the IO now to
900 * avoid metadata seeks.
901 */
902 if (sdio->boundary) {
903 ret = dio_send_cur_page(dio, sdio, map_bh);
904 if (sdio->bio)
905 dio_bio_submit(dio, sdio);
906 put_page(sdio->cur_page);
907 sdio->cur_page = NULL;
908 }
909 return ret;
910}
911
912/*
913 * If we are not writing the entire block and get_block() allocated
914 * the block for us, we need to fill-in the unused portion of the
915 * block with zeros. This happens only if user-buffer, fileoffset or
916 * io length is not filesystem block-size multiple.
917 *
918 * `end' is zero if we're doing the start of the IO, 1 at the end of the
919 * IO.
920 */
921static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
922 int end, struct buffer_head *map_bh)
923{
924 unsigned dio_blocks_per_fs_block;
925 unsigned this_chunk_blocks; /* In dio_blocks */
926 unsigned this_chunk_bytes;
927 struct page *page;
928
929 sdio->start_zero_done = 1;
930 if (!sdio->blkfactor || !buffer_new(map_bh))
931 return;
932
933 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
934 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
935
936 if (!this_chunk_blocks)
937 return;
938
939 /*
940 * We need to zero out part of an fs block. It is either at the
941 * beginning or the end of the fs block.
942 */
943 if (end)
944 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
945
946 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
947
948 page = ZERO_PAGE(0);
949 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
950 sdio->next_block_for_io, map_bh))
951 return;
952
953 sdio->next_block_for_io += this_chunk_blocks;
954}
955
956/*
957 * Walk the user pages, and the file, mapping blocks to disk and generating
958 * a sequence of (page,offset,len,block) mappings. These mappings are injected
959 * into submit_page_section(), which takes care of the next stage of submission
960 *
961 * Direct IO against a blockdev is different from a file. Because we can
962 * happily perform page-sized but 512-byte aligned IOs. It is important that
963 * blockdev IO be able to have fine alignment and large sizes.
964 *
965 * So what we do is to permit the ->get_block function to populate bh.b_size
966 * with the size of IO which is permitted at this offset and this i_blkbits.
967 *
968 * For best results, the blockdev should be set up with 512-byte i_blkbits and
969 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
970 * fine alignment but still allows this function to work in PAGE_SIZE units.
971 */
972static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
973 struct buffer_head *map_bh)
974{
975 const unsigned blkbits = sdio->blkbits;
976 const unsigned i_blkbits = blkbits + sdio->blkfactor;
977 int ret = 0;
978
979 while (sdio->block_in_file < sdio->final_block_in_request) {
980 struct page *page;
981 size_t from, to;
982
983 page = dio_get_page(dio, sdio);
984 if (IS_ERR(page)) {
985 ret = PTR_ERR(page);
986 goto out;
987 }
988 from = sdio->head ? 0 : sdio->from;
989 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
990 sdio->head++;
991
992 while (from < to) {
993 unsigned this_chunk_bytes; /* # of bytes mapped */
994 unsigned this_chunk_blocks; /* # of blocks */
995 unsigned u;
996
997 if (sdio->blocks_available == 0) {
998 /*
999 * Need to go and map some more disk
1000 */
1001 unsigned long blkmask;
1002 unsigned long dio_remainder;
1003
1004 ret = get_more_blocks(dio, sdio, map_bh);
1005 if (ret) {
1006 put_page(page);
1007 goto out;
1008 }
1009 if (!buffer_mapped(map_bh))
1010 goto do_holes;
1011
1012 sdio->blocks_available =
1013 map_bh->b_size >> blkbits;
1014 sdio->next_block_for_io =
1015 map_bh->b_blocknr << sdio->blkfactor;
1016 if (buffer_new(map_bh)) {
1017 clean_bdev_aliases(
1018 map_bh->b_bdev,
1019 map_bh->b_blocknr,
1020 map_bh->b_size >> i_blkbits);
1021 }
1022
1023 if (!sdio->blkfactor)
1024 goto do_holes;
1025
1026 blkmask = (1 << sdio->blkfactor) - 1;
1027 dio_remainder = (sdio->block_in_file & blkmask);
1028
1029 /*
1030 * If we are at the start of IO and that IO
1031 * starts partway into a fs-block,
1032 * dio_remainder will be non-zero. If the IO
1033 * is a read then we can simply advance the IO
1034 * cursor to the first block which is to be
1035 * read. But if the IO is a write and the
1036 * block was newly allocated we cannot do that;
1037 * the start of the fs block must be zeroed out
1038 * on-disk
1039 */
1040 if (!buffer_new(map_bh))
1041 sdio->next_block_for_io += dio_remainder;
1042 sdio->blocks_available -= dio_remainder;
1043 }
1044do_holes:
1045 /* Handle holes */
1046 if (!buffer_mapped(map_bh)) {
1047 loff_t i_size_aligned;
1048
1049 /* AKPM: eargh, -ENOTBLK is a hack */
1050 if (dio->op == REQ_OP_WRITE) {
1051 put_page(page);
1052 return -ENOTBLK;
1053 }
1054
1055 /*
1056 * Be sure to account for a partial block as the
1057 * last block in the file
1058 */
1059 i_size_aligned = ALIGN(i_size_read(dio->inode),
1060 1 << blkbits);
1061 if (sdio->block_in_file >=
1062 i_size_aligned >> blkbits) {
1063 /* We hit eof */
1064 put_page(page);
1065 goto out;
1066 }
1067 zero_user(page, from, 1 << blkbits);
1068 sdio->block_in_file++;
1069 from += 1 << blkbits;
1070 dio->result += 1 << blkbits;
1071 goto next_block;
1072 }
1073
1074 /*
1075 * If we're performing IO which has an alignment which
1076 * is finer than the underlying fs, go check to see if
1077 * we must zero out the start of this block.
1078 */
1079 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1080 dio_zero_block(dio, sdio, 0, map_bh);
1081
1082 /*
1083 * Work out, in this_chunk_blocks, how much disk we
1084 * can add to this page
1085 */
1086 this_chunk_blocks = sdio->blocks_available;
1087 u = (to - from) >> blkbits;
1088 if (this_chunk_blocks > u)
1089 this_chunk_blocks = u;
1090 u = sdio->final_block_in_request - sdio->block_in_file;
1091 if (this_chunk_blocks > u)
1092 this_chunk_blocks = u;
1093 this_chunk_bytes = this_chunk_blocks << blkbits;
1094 BUG_ON(this_chunk_bytes == 0);
1095
1096 if (this_chunk_blocks == sdio->blocks_available)
1097 sdio->boundary = buffer_boundary(map_bh);
1098 ret = submit_page_section(dio, sdio, page,
1099 from,
1100 this_chunk_bytes,
1101 sdio->next_block_for_io,
1102 map_bh);
1103 if (ret) {
1104 put_page(page);
1105 goto out;
1106 }
1107 sdio->next_block_for_io += this_chunk_blocks;
1108
1109 sdio->block_in_file += this_chunk_blocks;
1110 from += this_chunk_bytes;
1111 dio->result += this_chunk_bytes;
1112 sdio->blocks_available -= this_chunk_blocks;
1113next_block:
1114 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1115 if (sdio->block_in_file == sdio->final_block_in_request)
1116 break;
1117 }
1118
1119 /* Drop the ref which was taken in get_user_pages() */
1120 put_page(page);
1121 }
1122out:
1123 return ret;
1124}
1125
1126static inline int drop_refcount(struct dio *dio)
1127{
1128 int ret2;
1129 unsigned long flags;
1130
1131 /*
1132 * Sync will always be dropping the final ref and completing the
1133 * operation. AIO can if it was a broken operation described above or
1134 * in fact if all the bios race to complete before we get here. In
1135 * that case dio_complete() translates the EIOCBQUEUED into the proper
1136 * return code that the caller will hand to ->complete().
1137 *
1138 * This is managed by the bio_lock instead of being an atomic_t so that
1139 * completion paths can drop their ref and use the remaining count to
1140 * decide to wake the submission path atomically.
1141 */
1142 spin_lock_irqsave(&dio->bio_lock, flags);
1143 ret2 = --dio->refcount;
1144 spin_unlock_irqrestore(&dio->bio_lock, flags);
1145 return ret2;
1146}
1147
1148/*
1149 * This is a library function for use by filesystem drivers.
1150 *
1151 * The locking rules are governed by the flags parameter:
1152 * - if the flags value contains DIO_LOCKING we use a fancy locking
1153 * scheme for dumb filesystems.
1154 * For writes this function is called under i_mutex and returns with
1155 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1156 * taken and dropped again before returning.
1157 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1158 * internal locking but rather rely on the filesystem to synchronize
1159 * direct I/O reads/writes versus each other and truncate.
1160 *
1161 * To help with locking against truncate we incremented the i_dio_count
1162 * counter before starting direct I/O, and decrement it once we are done.
1163 * Truncate can wait for it to reach zero to provide exclusion. It is
1164 * expected that filesystem provide exclusion between new direct I/O
1165 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1166 * but other filesystems need to take care of this on their own.
1167 *
1168 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1169 * is always inlined. Otherwise gcc is unable to split the structure into
1170 * individual fields and will generate much worse code. This is important
1171 * for the whole file.
1172 */
1173static inline ssize_t
1174do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1175 struct block_device *bdev, struct iov_iter *iter,
1176 get_block_t get_block, dio_iodone_t end_io,
1177 dio_submit_t submit_io, int flags)
1178{
1179 unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
1180 unsigned blkbits = i_blkbits;
1181 unsigned blocksize_mask = (1 << blkbits) - 1;
1182 ssize_t retval = -EINVAL;
1183 const size_t count = iov_iter_count(iter);
1184 loff_t offset = iocb->ki_pos;
1185 const loff_t end = offset + count;
1186 struct dio *dio;
1187 struct dio_submit sdio = { 0, };
1188 struct buffer_head map_bh = { 0, };
1189 struct blk_plug plug;
1190 unsigned long align = offset | iov_iter_alignment(iter);
1191
1192 /*
1193 * Avoid references to bdev if not absolutely needed to give
1194 * the early prefetch in the caller enough time.
1195 */
1196
1197 if (align & blocksize_mask) {
1198 if (bdev)
1199 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1200 blocksize_mask = (1 << blkbits) - 1;
1201 if (align & blocksize_mask)
1202 goto out;
1203 }
1204
1205 /* watch out for a 0 len io from a tricksy fs */
1206 if (iov_iter_rw(iter) == READ && !count)
1207 return 0;
1208
1209 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1210 retval = -ENOMEM;
1211 if (!dio)
1212 goto out;
1213 /*
1214 * Believe it or not, zeroing out the page array caused a .5%
1215 * performance regression in a database benchmark. So, we take
1216 * care to only zero out what's needed.
1217 */
1218 memset(dio, 0, offsetof(struct dio, pages));
1219
1220 dio->flags = flags;
1221 if (dio->flags & DIO_LOCKING) {
1222 if (iov_iter_rw(iter) == READ) {
1223 struct address_space *mapping =
1224 iocb->ki_filp->f_mapping;
1225
1226 /* will be released by direct_io_worker */
1227 inode_lock(inode);
1228
1229 retval = filemap_write_and_wait_range(mapping, offset,
1230 end - 1);
1231 if (retval) {
1232 inode_unlock(inode);
1233 kmem_cache_free(dio_cache, dio);
1234 goto out;
1235 }
1236 }
1237 }
1238
1239 /* Once we sampled i_size check for reads beyond EOF */
1240 dio->i_size = i_size_read(inode);
1241 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1242 if (dio->flags & DIO_LOCKING)
1243 inode_unlock(inode);
1244 kmem_cache_free(dio_cache, dio);
1245 retval = 0;
1246 goto out;
1247 }
1248
1249 /*
1250 * For file extending writes updating i_size before data writeouts
1251 * complete can expose uninitialized blocks in dumb filesystems.
1252 * In that case we need to wait for I/O completion even if asked
1253 * for an asynchronous write.
1254 */
1255 if (is_sync_kiocb(iocb))
1256 dio->is_async = false;
1257 else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1258 dio->is_async = false;
1259 else
1260 dio->is_async = true;
1261
1262 dio->inode = inode;
1263 if (iov_iter_rw(iter) == WRITE) {
1264 dio->op = REQ_OP_WRITE;
1265 dio->op_flags = REQ_SYNC | REQ_IDLE;
1266 if (iocb->ki_flags & IOCB_NOWAIT)
1267 dio->op_flags |= REQ_NOWAIT;
1268 } else {
1269 dio->op = REQ_OP_READ;
1270 }
1271 if (iocb->ki_flags & IOCB_HIPRI)
1272 dio->op_flags |= REQ_HIPRI;
1273
1274 /*
1275 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1276 * so that we can call ->fsync.
1277 */
1278 if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1279 retval = 0;
1280 if (iocb->ki_flags & IOCB_DSYNC)
1281 retval = dio_set_defer_completion(dio);
1282 else if (!dio->inode->i_sb->s_dio_done_wq) {
1283 /*
1284 * In case of AIO write racing with buffered read we
1285 * need to defer completion. We can't decide this now,
1286 * however the workqueue needs to be initialized here.
1287 */
1288 retval = sb_init_dio_done_wq(dio->inode->i_sb);
1289 }
1290 if (retval) {
1291 /*
1292 * We grab i_mutex only for reads so we don't have
1293 * to release it here
1294 */
1295 kmem_cache_free(dio_cache, dio);
1296 goto out;
1297 }
1298 }
1299
1300 /*
1301 * Will be decremented at I/O completion time.
1302 */
1303 inode_dio_begin(inode);
1304
1305 retval = 0;
1306 sdio.blkbits = blkbits;
1307 sdio.blkfactor = i_blkbits - blkbits;
1308 sdio.block_in_file = offset >> blkbits;
1309
1310 sdio.get_block = get_block;
1311 dio->end_io = end_io;
1312 sdio.submit_io = submit_io;
1313 sdio.final_block_in_bio = -1;
1314 sdio.next_block_for_io = -1;
1315
1316 dio->iocb = iocb;
1317
1318 spin_lock_init(&dio->bio_lock);
1319 dio->refcount = 1;
1320
1321 dio->should_dirty = iter_is_iovec(iter) && iov_iter_rw(iter) == READ;
1322 sdio.iter = iter;
1323 sdio.final_block_in_request = end >> blkbits;
1324
1325 /*
1326 * In case of non-aligned buffers, we may need 2 more
1327 * pages since we need to zero out first and last block.
1328 */
1329 if (unlikely(sdio.blkfactor))
1330 sdio.pages_in_io = 2;
1331
1332 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1333
1334 blk_start_plug(&plug);
1335
1336 retval = do_direct_IO(dio, &sdio, &map_bh);
1337 if (retval)
1338 dio_cleanup(dio, &sdio);
1339
1340 if (retval == -ENOTBLK) {
1341 /*
1342 * The remaining part of the request will be
1343 * be handled by buffered I/O when we return
1344 */
1345 retval = 0;
1346 }
1347 /*
1348 * There may be some unwritten disk at the end of a part-written
1349 * fs-block-sized block. Go zero that now.
1350 */
1351 dio_zero_block(dio, &sdio, 1, &map_bh);
1352
1353 if (sdio.cur_page) {
1354 ssize_t ret2;
1355
1356 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1357 if (retval == 0)
1358 retval = ret2;
1359 put_page(sdio.cur_page);
1360 sdio.cur_page = NULL;
1361 }
1362 if (sdio.bio)
1363 dio_bio_submit(dio, &sdio);
1364
1365 blk_finish_plug(&plug);
1366
1367 /*
1368 * It is possible that, we return short IO due to end of file.
1369 * In that case, we need to release all the pages we got hold on.
1370 */
1371 dio_cleanup(dio, &sdio);
1372
1373 /*
1374 * All block lookups have been performed. For READ requests
1375 * we can let i_mutex go now that its achieved its purpose
1376 * of protecting us from looking up uninitialized blocks.
1377 */
1378 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1379 inode_unlock(dio->inode);
1380
1381 /*
1382 * The only time we want to leave bios in flight is when a successful
1383 * partial aio read or full aio write have been setup. In that case
1384 * bio completion will call aio_complete. The only time it's safe to
1385 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1386 * This had *better* be the only place that raises -EIOCBQUEUED.
1387 */
1388 BUG_ON(retval == -EIOCBQUEUED);
1389 if (dio->is_async && retval == 0 && dio->result &&
1390 (iov_iter_rw(iter) == READ || dio->result == count))
1391 retval = -EIOCBQUEUED;
1392 else
1393 dio_await_completion(dio);
1394
1395 if (drop_refcount(dio) == 0) {
1396 retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
1397 } else
1398 BUG_ON(retval != -EIOCBQUEUED);
1399
1400out:
1401 return retval;
1402}
1403
1404ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1405 struct block_device *bdev, struct iov_iter *iter,
1406 get_block_t get_block,
1407 dio_iodone_t end_io, dio_submit_t submit_io,
1408 int flags)
1409{
1410 /*
1411 * The block device state is needed in the end to finally
1412 * submit everything. Since it's likely to be cache cold
1413 * prefetch it here as first thing to hide some of the
1414 * latency.
1415 *
1416 * Attempt to prefetch the pieces we likely need later.
1417 */
1418 prefetch(&bdev->bd_disk->part_tbl);
1419 prefetch(bdev->bd_queue);
1420 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1421
1422 return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
1423 end_io, submit_io, flags);
1424}
1425
1426EXPORT_SYMBOL(__blockdev_direct_IO);
1427
1428static __init int dio_init(void)
1429{
1430 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1431 return 0;
1432}
1433module_init(dio_init)
1434