xref: /openbmc/linux/fs/direct-io.c (revision c33c7948)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * fs/direct-io.c
4  *
5  * Copyright (C) 2002, Linus Torvalds.
6  *
7  * O_DIRECT
8  *
9  * 04Jul2002	Andrew Morton
10  *		Initial version
11  * 11Sep2002	janetinc@us.ibm.com
12  * 		added readv/writev support.
13  * 29Oct2002	Andrew Morton
14  *		rewrote bio_add_page() support.
15  * 30Oct2002	pbadari@us.ibm.com
16  *		added support for non-aligned IO.
17  * 06Nov2002	pbadari@us.ibm.com
18  *		added asynchronous IO support.
19  * 21Jul2003	nathans@sgi.com
20  *		added IO completion notifier.
21  */
22 
23 #include <linux/kernel.h>
24 #include <linux/module.h>
25 #include <linux/types.h>
26 #include <linux/fs.h>
27 #include <linux/mm.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/pagemap.h>
31 #include <linux/task_io_accounting_ops.h>
32 #include <linux/bio.h>
33 #include <linux/wait.h>
34 #include <linux/err.h>
35 #include <linux/blkdev.h>
36 #include <linux/buffer_head.h>
37 #include <linux/rwsem.h>
38 #include <linux/uio.h>
39 #include <linux/atomic.h>
40 #include <linux/prefetch.h>
41 
42 #include "internal.h"
43 
44 /*
45  * How many user pages to map in one call to get_user_pages().  This determines
46  * the size of a structure in the slab cache
47  */
48 #define DIO_PAGES	64
49 
50 /*
51  * Flags for dio_complete()
52  */
53 #define DIO_COMPLETE_ASYNC		0x01	/* This is async IO */
54 #define DIO_COMPLETE_INVALIDATE		0x02	/* Can invalidate pages */
55 
56 /*
57  * This code generally works in units of "dio_blocks".  A dio_block is
58  * somewhere between the hard sector size and the filesystem block size.  it
59  * is determined on a per-invocation basis.   When talking to the filesystem
60  * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
61  * down by dio->blkfactor.  Similarly, fs-blocksize quantities are converted
62  * to bio_block quantities by shifting left by blkfactor.
63  *
64  * If blkfactor is zero then the user's request was aligned to the filesystem's
65  * blocksize.
66  */
67 
68 /* dio_state only used in the submission path */
69 
70 struct dio_submit {
71 	struct bio *bio;		/* bio under assembly */
72 	unsigned blkbits;		/* doesn't change */
73 	unsigned blkfactor;		/* When we're using an alignment which
74 					   is finer than the filesystem's soft
75 					   blocksize, this specifies how much
76 					   finer.  blkfactor=2 means 1/4-block
77 					   alignment.  Does not change */
78 	unsigned start_zero_done;	/* flag: sub-blocksize zeroing has
79 					   been performed at the start of a
80 					   write */
81 	int pages_in_io;		/* approximate total IO pages */
82 	sector_t block_in_file;		/* Current offset into the underlying
83 					   file in dio_block units. */
84 	unsigned blocks_available;	/* At block_in_file.  changes */
85 	int reap_counter;		/* rate limit reaping */
86 	sector_t final_block_in_request;/* doesn't change */
87 	int boundary;			/* prev block is at a boundary */
88 	get_block_t *get_block;		/* block mapping function */
89 
90 	loff_t logical_offset_in_bio;	/* current first logical block in bio */
91 	sector_t final_block_in_bio;	/* current final block in bio + 1 */
92 	sector_t next_block_for_io;	/* next block to be put under IO,
93 					   in dio_blocks units */
94 
95 	/*
96 	 * Deferred addition of a page to the dio.  These variables are
97 	 * private to dio_send_cur_page(), submit_page_section() and
98 	 * dio_bio_add_page().
99 	 */
100 	struct page *cur_page;		/* The page */
101 	unsigned cur_page_offset;	/* Offset into it, in bytes */
102 	unsigned cur_page_len;		/* Nr of bytes at cur_page_offset */
103 	sector_t cur_page_block;	/* Where it starts */
104 	loff_t cur_page_fs_offset;	/* Offset in file */
105 
106 	struct iov_iter *iter;
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 	size_t from, to;
114 };
115 
116 /* dio_state communicated between submission path and end_io */
117 struct dio {
118 	int flags;			/* doesn't change */
119 	blk_opf_t opf;			/* request operation type and flags */
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 
153 static struct kmem_cache *dio_cache __read_mostly;
154 
155 /*
156  * How many pages are in the queue?
157  */
158 static 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  */
166 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
167 {
168 	const enum req_op dio_op = dio->opf & REQ_OP_MASK;
169 	ssize_t ret;
170 
171 	ret = iov_iter_get_pages2(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
172 				&sdio->from);
173 
174 	if (ret < 0 && sdio->blocks_available && dio_op == REQ_OP_WRITE) {
175 		struct page *page = ZERO_PAGE(0);
176 		/*
177 		 * A memory fault, but the filesystem has some outstanding
178 		 * mapped blocks.  We need to use those blocks up to avoid
179 		 * leaking stale data in the file.
180 		 */
181 		if (dio->page_errors == 0)
182 			dio->page_errors = ret;
183 		get_page(page);
184 		dio->pages[0] = page;
185 		sdio->head = 0;
186 		sdio->tail = 1;
187 		sdio->from = 0;
188 		sdio->to = PAGE_SIZE;
189 		return 0;
190 	}
191 
192 	if (ret >= 0) {
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  */
208 static 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  * dio_complete() - called when all DIO BIO I/O has been completed
224  *
225  * This drops i_dio_count, lets interested parties know that a DIO operation
226  * has completed, and calculates the resulting return code for the operation.
227  *
228  * It lets the filesystem know if it registered an interest earlier via
229  * get_block.  Pass the private field of the map buffer_head so that
230  * filesystems can use it to hold additional state between get_block calls and
231  * dio_complete.
232  */
233 static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
234 {
235 	const enum req_op dio_op = dio->opf & REQ_OP_MASK;
236 	loff_t offset = dio->iocb->ki_pos;
237 	ssize_t transferred = 0;
238 	int err;
239 
240 	/*
241 	 * AIO submission can race with bio completion to get here while
242 	 * expecting to have the last io completed by bio completion.
243 	 * In that case -EIOCBQUEUED is in fact not an error we want
244 	 * to preserve through this call.
245 	 */
246 	if (ret == -EIOCBQUEUED)
247 		ret = 0;
248 
249 	if (dio->result) {
250 		transferred = dio->result;
251 
252 		/* Check for short read case */
253 		if (dio_op == REQ_OP_READ &&
254 		    ((offset + transferred) > dio->i_size))
255 			transferred = dio->i_size - offset;
256 		/* ignore EFAULT if some IO has been done */
257 		if (unlikely(ret == -EFAULT) && transferred)
258 			ret = 0;
259 	}
260 
261 	if (ret == 0)
262 		ret = dio->page_errors;
263 	if (ret == 0)
264 		ret = dio->io_error;
265 	if (ret == 0)
266 		ret = transferred;
267 
268 	if (dio->end_io) {
269 		// XXX: ki_pos??
270 		err = dio->end_io(dio->iocb, offset, ret, dio->private);
271 		if (err)
272 			ret = err;
273 	}
274 
275 	/*
276 	 * Try again to invalidate clean pages which might have been cached by
277 	 * non-direct readahead, or faulted in by get_user_pages() if the source
278 	 * of the write was an mmap'ed region of the file we're writing.  Either
279 	 * one is a pretty crazy thing to do, so we don't support it 100%.  If
280 	 * this invalidation fails, tough, the write still worked...
281 	 *
282 	 * And this page cache invalidation has to be after dio->end_io(), as
283 	 * some filesystems convert unwritten extents to real allocations in
284 	 * end_io() when necessary, otherwise a racing buffer read would cache
285 	 * zeros from unwritten extents.
286 	 */
287 	if (flags & DIO_COMPLETE_INVALIDATE &&
288 	    ret > 0 && dio_op == REQ_OP_WRITE)
289 		kiocb_invalidate_post_direct_write(dio->iocb, ret);
290 
291 	inode_dio_end(dio->inode);
292 
293 	if (flags & DIO_COMPLETE_ASYNC) {
294 		/*
295 		 * generic_write_sync expects ki_pos to have been updated
296 		 * already, but the submission path only does this for
297 		 * synchronous I/O.
298 		 */
299 		dio->iocb->ki_pos += transferred;
300 
301 		if (ret > 0 && dio_op == REQ_OP_WRITE)
302 			ret = generic_write_sync(dio->iocb, ret);
303 		dio->iocb->ki_complete(dio->iocb, ret);
304 	}
305 
306 	kmem_cache_free(dio_cache, dio);
307 	return ret;
308 }
309 
310 static void dio_aio_complete_work(struct work_struct *work)
311 {
312 	struct dio *dio = container_of(work, struct dio, complete_work);
313 
314 	dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
315 }
316 
317 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
318 
319 /*
320  * Asynchronous IO callback.
321  */
322 static void dio_bio_end_aio(struct bio *bio)
323 {
324 	struct dio *dio = bio->bi_private;
325 	const enum req_op dio_op = dio->opf & REQ_OP_MASK;
326 	unsigned long remaining;
327 	unsigned long flags;
328 	bool defer_completion = false;
329 
330 	/* cleanup the bio */
331 	dio_bio_complete(dio, bio);
332 
333 	spin_lock_irqsave(&dio->bio_lock, flags);
334 	remaining = --dio->refcount;
335 	if (remaining == 1 && dio->waiter)
336 		wake_up_process(dio->waiter);
337 	spin_unlock_irqrestore(&dio->bio_lock, flags);
338 
339 	if (remaining == 0) {
340 		/*
341 		 * Defer completion when defer_completion is set or
342 		 * when the inode has pages mapped and this is AIO write.
343 		 * We need to invalidate those pages because there is a
344 		 * chance they contain stale data in the case buffered IO
345 		 * went in between AIO submission and completion into the
346 		 * same region.
347 		 */
348 		if (dio->result)
349 			defer_completion = dio->defer_completion ||
350 					   (dio_op == REQ_OP_WRITE &&
351 					    dio->inode->i_mapping->nrpages);
352 		if (defer_completion) {
353 			INIT_WORK(&dio->complete_work, dio_aio_complete_work);
354 			queue_work(dio->inode->i_sb->s_dio_done_wq,
355 				   &dio->complete_work);
356 		} else {
357 			dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
358 		}
359 	}
360 }
361 
362 /*
363  * The BIO completion handler simply queues the BIO up for the process-context
364  * handler.
365  *
366  * During I/O bi_private points at the dio.  After I/O, bi_private is used to
367  * implement a singly-linked list of completed BIOs, at dio->bio_list.
368  */
369 static void dio_bio_end_io(struct bio *bio)
370 {
371 	struct dio *dio = bio->bi_private;
372 	unsigned long flags;
373 
374 	spin_lock_irqsave(&dio->bio_lock, flags);
375 	bio->bi_private = dio->bio_list;
376 	dio->bio_list = bio;
377 	if (--dio->refcount == 1 && dio->waiter)
378 		wake_up_process(dio->waiter);
379 	spin_unlock_irqrestore(&dio->bio_lock, flags);
380 }
381 
382 static inline void
383 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
384 	      struct block_device *bdev,
385 	      sector_t first_sector, int nr_vecs)
386 {
387 	struct bio *bio;
388 
389 	/*
390 	 * bio_alloc() is guaranteed to return a bio when allowed to sleep and
391 	 * we request a valid number of vectors.
392 	 */
393 	bio = bio_alloc(bdev, nr_vecs, dio->opf, GFP_KERNEL);
394 	bio->bi_iter.bi_sector = first_sector;
395 	if (dio->is_async)
396 		bio->bi_end_io = dio_bio_end_aio;
397 	else
398 		bio->bi_end_io = dio_bio_end_io;
399 	sdio->bio = bio;
400 	sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
401 }
402 
403 /*
404  * In the AIO read case we speculatively dirty the pages before starting IO.
405  * During IO completion, any of these pages which happen to have been written
406  * back will be redirtied by bio_check_pages_dirty().
407  *
408  * bios hold a dio reference between submit_bio and ->end_io.
409  */
410 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
411 {
412 	const enum req_op dio_op = dio->opf & REQ_OP_MASK;
413 	struct bio *bio = sdio->bio;
414 	unsigned long flags;
415 
416 	bio->bi_private = dio;
417 
418 	spin_lock_irqsave(&dio->bio_lock, flags);
419 	dio->refcount++;
420 	spin_unlock_irqrestore(&dio->bio_lock, flags);
421 
422 	if (dio->is_async && dio_op == REQ_OP_READ && dio->should_dirty)
423 		bio_set_pages_dirty(bio);
424 
425 	dio->bio_disk = bio->bi_bdev->bd_disk;
426 
427 	submit_bio(bio);
428 
429 	sdio->bio = NULL;
430 	sdio->boundary = 0;
431 	sdio->logical_offset_in_bio = 0;
432 }
433 
434 /*
435  * Release any resources in case of a failure
436  */
437 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
438 {
439 	while (sdio->head < sdio->tail)
440 		put_page(dio->pages[sdio->head++]);
441 }
442 
443 /*
444  * Wait for the next BIO to complete.  Remove it and return it.  NULL is
445  * returned once all BIOs have been completed.  This must only be called once
446  * all bios have been issued so that dio->refcount can only decrease.  This
447  * requires that the caller hold a reference on the dio.
448  */
449 static struct bio *dio_await_one(struct dio *dio)
450 {
451 	unsigned long flags;
452 	struct bio *bio = NULL;
453 
454 	spin_lock_irqsave(&dio->bio_lock, flags);
455 
456 	/*
457 	 * Wait as long as the list is empty and there are bios in flight.  bio
458 	 * completion drops the count, maybe adds to the list, and wakes while
459 	 * holding the bio_lock so we don't need set_current_state()'s barrier
460 	 * and can call it after testing our condition.
461 	 */
462 	while (dio->refcount > 1 && dio->bio_list == NULL) {
463 		__set_current_state(TASK_UNINTERRUPTIBLE);
464 		dio->waiter = current;
465 		spin_unlock_irqrestore(&dio->bio_lock, flags);
466 		blk_io_schedule();
467 		/* wake up sets us TASK_RUNNING */
468 		spin_lock_irqsave(&dio->bio_lock, flags);
469 		dio->waiter = NULL;
470 	}
471 	if (dio->bio_list) {
472 		bio = dio->bio_list;
473 		dio->bio_list = bio->bi_private;
474 	}
475 	spin_unlock_irqrestore(&dio->bio_lock, flags);
476 	return bio;
477 }
478 
479 /*
480  * Process one completed BIO.  No locks are held.
481  */
482 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
483 {
484 	blk_status_t err = bio->bi_status;
485 	const enum req_op dio_op = dio->opf & REQ_OP_MASK;
486 	bool should_dirty = dio_op == REQ_OP_READ && dio->should_dirty;
487 
488 	if (err) {
489 		if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
490 			dio->io_error = -EAGAIN;
491 		else
492 			dio->io_error = -EIO;
493 	}
494 
495 	if (dio->is_async && should_dirty) {
496 		bio_check_pages_dirty(bio);	/* transfers ownership */
497 	} else {
498 		bio_release_pages(bio, should_dirty);
499 		bio_put(bio);
500 	}
501 	return err;
502 }
503 
504 /*
505  * Wait on and process all in-flight BIOs.  This must only be called once
506  * all bios have been issued so that the refcount can only decrease.
507  * This just waits for all bios to make it through dio_bio_complete.  IO
508  * errors are propagated through dio->io_error and should be propagated via
509  * dio_complete().
510  */
511 static void dio_await_completion(struct dio *dio)
512 {
513 	struct bio *bio;
514 	do {
515 		bio = dio_await_one(dio);
516 		if (bio)
517 			dio_bio_complete(dio, bio);
518 	} while (bio);
519 }
520 
521 /*
522  * A really large O_DIRECT read or write can generate a lot of BIOs.  So
523  * to keep the memory consumption sane we periodically reap any completed BIOs
524  * during the BIO generation phase.
525  *
526  * This also helps to limit the peak amount of pinned userspace memory.
527  */
528 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
529 {
530 	int ret = 0;
531 
532 	if (sdio->reap_counter++ >= 64) {
533 		while (dio->bio_list) {
534 			unsigned long flags;
535 			struct bio *bio;
536 			int ret2;
537 
538 			spin_lock_irqsave(&dio->bio_lock, flags);
539 			bio = dio->bio_list;
540 			dio->bio_list = bio->bi_private;
541 			spin_unlock_irqrestore(&dio->bio_lock, flags);
542 			ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
543 			if (ret == 0)
544 				ret = ret2;
545 		}
546 		sdio->reap_counter = 0;
547 	}
548 	return ret;
549 }
550 
551 static int dio_set_defer_completion(struct dio *dio)
552 {
553 	struct super_block *sb = dio->inode->i_sb;
554 
555 	if (dio->defer_completion)
556 		return 0;
557 	dio->defer_completion = true;
558 	if (!sb->s_dio_done_wq)
559 		return sb_init_dio_done_wq(sb);
560 	return 0;
561 }
562 
563 /*
564  * Call into the fs to map some more disk blocks.  We record the current number
565  * of available blocks at sdio->blocks_available.  These are in units of the
566  * fs blocksize, i_blocksize(inode).
567  *
568  * The fs is allowed to map lots of blocks at once.  If it wants to do that,
569  * it uses the passed inode-relative block number as the file offset, as usual.
570  *
571  * get_block() is passed the number of i_blkbits-sized blocks which direct_io
572  * has remaining to do.  The fs should not map more than this number of blocks.
573  *
574  * If the fs has mapped a lot of blocks, it should populate bh->b_size to
575  * indicate how much contiguous disk space has been made available at
576  * bh->b_blocknr.
577  *
578  * If *any* of the mapped blocks are new, then the fs must set buffer_new().
579  * This isn't very efficient...
580  *
581  * In the case of filesystem holes: the fs may return an arbitrarily-large
582  * hole by returning an appropriate value in b_size and by clearing
583  * buffer_mapped().  However the direct-io code will only process holes one
584  * block at a time - it will repeatedly call get_block() as it walks the hole.
585  */
586 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
587 			   struct buffer_head *map_bh)
588 {
589 	const enum req_op dio_op = dio->opf & REQ_OP_MASK;
590 	int ret;
591 	sector_t fs_startblk;	/* Into file, in filesystem-sized blocks */
592 	sector_t fs_endblk;	/* Into file, in filesystem-sized blocks */
593 	unsigned long fs_count;	/* Number of filesystem-sized blocks */
594 	int create;
595 	unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
596 	loff_t i_size;
597 
598 	/*
599 	 * If there was a memory error and we've overwritten all the
600 	 * mapped blocks then we can now return that memory error
601 	 */
602 	ret = dio->page_errors;
603 	if (ret == 0) {
604 		BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
605 		fs_startblk = sdio->block_in_file >> sdio->blkfactor;
606 		fs_endblk = (sdio->final_block_in_request - 1) >>
607 					sdio->blkfactor;
608 		fs_count = fs_endblk - fs_startblk + 1;
609 
610 		map_bh->b_state = 0;
611 		map_bh->b_size = fs_count << i_blkbits;
612 
613 		/*
614 		 * For writes that could fill holes inside i_size on a
615 		 * DIO_SKIP_HOLES filesystem we forbid block creations: only
616 		 * overwrites are permitted. We will return early to the caller
617 		 * once we see an unmapped buffer head returned, and the caller
618 		 * will fall back to buffered I/O.
619 		 *
620 		 * Otherwise the decision is left to the get_blocks method,
621 		 * which may decide to handle it or also return an unmapped
622 		 * buffer head.
623 		 */
624 		create = dio_op == REQ_OP_WRITE;
625 		if (dio->flags & DIO_SKIP_HOLES) {
626 			i_size = i_size_read(dio->inode);
627 			if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
628 				create = 0;
629 		}
630 
631 		ret = (*sdio->get_block)(dio->inode, fs_startblk,
632 						map_bh, create);
633 
634 		/* Store for completion */
635 		dio->private = map_bh->b_private;
636 
637 		if (ret == 0 && buffer_defer_completion(map_bh))
638 			ret = dio_set_defer_completion(dio);
639 	}
640 	return ret;
641 }
642 
643 /*
644  * There is no bio.  Make one now.
645  */
646 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
647 		sector_t start_sector, struct buffer_head *map_bh)
648 {
649 	sector_t sector;
650 	int ret, nr_pages;
651 
652 	ret = dio_bio_reap(dio, sdio);
653 	if (ret)
654 		goto out;
655 	sector = start_sector << (sdio->blkbits - 9);
656 	nr_pages = bio_max_segs(sdio->pages_in_io);
657 	BUG_ON(nr_pages <= 0);
658 	dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
659 	sdio->boundary = 0;
660 out:
661 	return ret;
662 }
663 
664 /*
665  * Attempt to put the current chunk of 'cur_page' into the current BIO.  If
666  * that was successful then update final_block_in_bio and take a ref against
667  * the just-added page.
668  *
669  * Return zero on success.  Non-zero means the caller needs to start a new BIO.
670  */
671 static inline int dio_bio_add_page(struct dio_submit *sdio)
672 {
673 	int ret;
674 
675 	ret = bio_add_page(sdio->bio, sdio->cur_page,
676 			sdio->cur_page_len, sdio->cur_page_offset);
677 	if (ret == sdio->cur_page_len) {
678 		/*
679 		 * Decrement count only, if we are done with this page
680 		 */
681 		if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
682 			sdio->pages_in_io--;
683 		get_page(sdio->cur_page);
684 		sdio->final_block_in_bio = sdio->cur_page_block +
685 			(sdio->cur_page_len >> sdio->blkbits);
686 		ret = 0;
687 	} else {
688 		ret = 1;
689 	}
690 	return ret;
691 }
692 
693 /*
694  * Put cur_page under IO.  The section of cur_page which is described by
695  * cur_page_offset,cur_page_len is put into a BIO.  The section of cur_page
696  * starts on-disk at cur_page_block.
697  *
698  * We take a ref against the page here (on behalf of its presence in the bio).
699  *
700  * The caller of this function is responsible for removing cur_page from the
701  * dio, and for dropping the refcount which came from that presence.
702  */
703 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
704 		struct buffer_head *map_bh)
705 {
706 	int ret = 0;
707 
708 	if (sdio->bio) {
709 		loff_t cur_offset = sdio->cur_page_fs_offset;
710 		loff_t bio_next_offset = sdio->logical_offset_in_bio +
711 			sdio->bio->bi_iter.bi_size;
712 
713 		/*
714 		 * See whether this new request is contiguous with the old.
715 		 *
716 		 * Btrfs cannot handle having logically non-contiguous requests
717 		 * submitted.  For example if you have
718 		 *
719 		 * Logical:  [0-4095][HOLE][8192-12287]
720 		 * Physical: [0-4095]      [4096-8191]
721 		 *
722 		 * We cannot submit those pages together as one BIO.  So if our
723 		 * current logical offset in the file does not equal what would
724 		 * be the next logical offset in the bio, submit the bio we
725 		 * have.
726 		 */
727 		if (sdio->final_block_in_bio != sdio->cur_page_block ||
728 		    cur_offset != bio_next_offset)
729 			dio_bio_submit(dio, sdio);
730 	}
731 
732 	if (sdio->bio == NULL) {
733 		ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
734 		if (ret)
735 			goto out;
736 	}
737 
738 	if (dio_bio_add_page(sdio) != 0) {
739 		dio_bio_submit(dio, sdio);
740 		ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
741 		if (ret == 0) {
742 			ret = dio_bio_add_page(sdio);
743 			BUG_ON(ret != 0);
744 		}
745 	}
746 out:
747 	return ret;
748 }
749 
750 /*
751  * An autonomous function to put a chunk of a page under deferred IO.
752  *
753  * The caller doesn't actually know (or care) whether this piece of page is in
754  * a BIO, or is under IO or whatever.  We just take care of all possible
755  * situations here.  The separation between the logic of do_direct_IO() and
756  * that of submit_page_section() is important for clarity.  Please don't break.
757  *
758  * The chunk of page starts on-disk at blocknr.
759  *
760  * We perform deferred IO, by recording the last-submitted page inside our
761  * private part of the dio structure.  If possible, we just expand the IO
762  * across that page here.
763  *
764  * If that doesn't work out then we put the old page into the bio and add this
765  * page to the dio instead.
766  */
767 static inline int
768 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
769 		    unsigned offset, unsigned len, sector_t blocknr,
770 		    struct buffer_head *map_bh)
771 {
772 	const enum req_op dio_op = dio->opf & REQ_OP_MASK;
773 	int ret = 0;
774 	int boundary = sdio->boundary;	/* dio_send_cur_page may clear it */
775 
776 	if (dio_op == REQ_OP_WRITE) {
777 		/*
778 		 * Read accounting is performed in submit_bio()
779 		 */
780 		task_io_account_write(len);
781 	}
782 
783 	/*
784 	 * Can we just grow the current page's presence in the dio?
785 	 */
786 	if (sdio->cur_page == page &&
787 	    sdio->cur_page_offset + sdio->cur_page_len == offset &&
788 	    sdio->cur_page_block +
789 	    (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
790 		sdio->cur_page_len += len;
791 		goto out;
792 	}
793 
794 	/*
795 	 * If there's a deferred page already there then send it.
796 	 */
797 	if (sdio->cur_page) {
798 		ret = dio_send_cur_page(dio, sdio, map_bh);
799 		put_page(sdio->cur_page);
800 		sdio->cur_page = NULL;
801 		if (ret)
802 			return ret;
803 	}
804 
805 	get_page(page);		/* It is in dio */
806 	sdio->cur_page = page;
807 	sdio->cur_page_offset = offset;
808 	sdio->cur_page_len = len;
809 	sdio->cur_page_block = blocknr;
810 	sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
811 out:
812 	/*
813 	 * If boundary then we want to schedule the IO now to
814 	 * avoid metadata seeks.
815 	 */
816 	if (boundary) {
817 		ret = dio_send_cur_page(dio, sdio, map_bh);
818 		if (sdio->bio)
819 			dio_bio_submit(dio, sdio);
820 		put_page(sdio->cur_page);
821 		sdio->cur_page = NULL;
822 	}
823 	return ret;
824 }
825 
826 /*
827  * If we are not writing the entire block and get_block() allocated
828  * the block for us, we need to fill-in the unused portion of the
829  * block with zeros. This happens only if user-buffer, fileoffset or
830  * io length is not filesystem block-size multiple.
831  *
832  * `end' is zero if we're doing the start of the IO, 1 at the end of the
833  * IO.
834  */
835 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
836 		int end, struct buffer_head *map_bh)
837 {
838 	unsigned dio_blocks_per_fs_block;
839 	unsigned this_chunk_blocks;	/* In dio_blocks */
840 	unsigned this_chunk_bytes;
841 	struct page *page;
842 
843 	sdio->start_zero_done = 1;
844 	if (!sdio->blkfactor || !buffer_new(map_bh))
845 		return;
846 
847 	dio_blocks_per_fs_block = 1 << sdio->blkfactor;
848 	this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
849 
850 	if (!this_chunk_blocks)
851 		return;
852 
853 	/*
854 	 * We need to zero out part of an fs block.  It is either at the
855 	 * beginning or the end of the fs block.
856 	 */
857 	if (end)
858 		this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
859 
860 	this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
861 
862 	page = ZERO_PAGE(0);
863 	if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
864 				sdio->next_block_for_io, map_bh))
865 		return;
866 
867 	sdio->next_block_for_io += this_chunk_blocks;
868 }
869 
870 /*
871  * Walk the user pages, and the file, mapping blocks to disk and generating
872  * a sequence of (page,offset,len,block) mappings.  These mappings are injected
873  * into submit_page_section(), which takes care of the next stage of submission
874  *
875  * Direct IO against a blockdev is different from a file.  Because we can
876  * happily perform page-sized but 512-byte aligned IOs.  It is important that
877  * blockdev IO be able to have fine alignment and large sizes.
878  *
879  * So what we do is to permit the ->get_block function to populate bh.b_size
880  * with the size of IO which is permitted at this offset and this i_blkbits.
881  *
882  * For best results, the blockdev should be set up with 512-byte i_blkbits and
883  * it should set b_size to PAGE_SIZE or more inside get_block().  This gives
884  * fine alignment but still allows this function to work in PAGE_SIZE units.
885  */
886 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
887 			struct buffer_head *map_bh)
888 {
889 	const enum req_op dio_op = dio->opf & REQ_OP_MASK;
890 	const unsigned blkbits = sdio->blkbits;
891 	const unsigned i_blkbits = blkbits + sdio->blkfactor;
892 	int ret = 0;
893 
894 	while (sdio->block_in_file < sdio->final_block_in_request) {
895 		struct page *page;
896 		size_t from, to;
897 
898 		page = dio_get_page(dio, sdio);
899 		if (IS_ERR(page)) {
900 			ret = PTR_ERR(page);
901 			goto out;
902 		}
903 		from = sdio->head ? 0 : sdio->from;
904 		to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
905 		sdio->head++;
906 
907 		while (from < to) {
908 			unsigned this_chunk_bytes;	/* # of bytes mapped */
909 			unsigned this_chunk_blocks;	/* # of blocks */
910 			unsigned u;
911 
912 			if (sdio->blocks_available == 0) {
913 				/*
914 				 * Need to go and map some more disk
915 				 */
916 				unsigned long blkmask;
917 				unsigned long dio_remainder;
918 
919 				ret = get_more_blocks(dio, sdio, map_bh);
920 				if (ret) {
921 					put_page(page);
922 					goto out;
923 				}
924 				if (!buffer_mapped(map_bh))
925 					goto do_holes;
926 
927 				sdio->blocks_available =
928 						map_bh->b_size >> blkbits;
929 				sdio->next_block_for_io =
930 					map_bh->b_blocknr << sdio->blkfactor;
931 				if (buffer_new(map_bh)) {
932 					clean_bdev_aliases(
933 						map_bh->b_bdev,
934 						map_bh->b_blocknr,
935 						map_bh->b_size >> i_blkbits);
936 				}
937 
938 				if (!sdio->blkfactor)
939 					goto do_holes;
940 
941 				blkmask = (1 << sdio->blkfactor) - 1;
942 				dio_remainder = (sdio->block_in_file & blkmask);
943 
944 				/*
945 				 * If we are at the start of IO and that IO
946 				 * starts partway into a fs-block,
947 				 * dio_remainder will be non-zero.  If the IO
948 				 * is a read then we can simply advance the IO
949 				 * cursor to the first block which is to be
950 				 * read.  But if the IO is a write and the
951 				 * block was newly allocated we cannot do that;
952 				 * the start of the fs block must be zeroed out
953 				 * on-disk
954 				 */
955 				if (!buffer_new(map_bh))
956 					sdio->next_block_for_io += dio_remainder;
957 				sdio->blocks_available -= dio_remainder;
958 			}
959 do_holes:
960 			/* Handle holes */
961 			if (!buffer_mapped(map_bh)) {
962 				loff_t i_size_aligned;
963 
964 				/* AKPM: eargh, -ENOTBLK is a hack */
965 				if (dio_op == REQ_OP_WRITE) {
966 					put_page(page);
967 					return -ENOTBLK;
968 				}
969 
970 				/*
971 				 * Be sure to account for a partial block as the
972 				 * last block in the file
973 				 */
974 				i_size_aligned = ALIGN(i_size_read(dio->inode),
975 							1 << blkbits);
976 				if (sdio->block_in_file >=
977 						i_size_aligned >> blkbits) {
978 					/* We hit eof */
979 					put_page(page);
980 					goto out;
981 				}
982 				zero_user(page, from, 1 << blkbits);
983 				sdio->block_in_file++;
984 				from += 1 << blkbits;
985 				dio->result += 1 << blkbits;
986 				goto next_block;
987 			}
988 
989 			/*
990 			 * If we're performing IO which has an alignment which
991 			 * is finer than the underlying fs, go check to see if
992 			 * we must zero out the start of this block.
993 			 */
994 			if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
995 				dio_zero_block(dio, sdio, 0, map_bh);
996 
997 			/*
998 			 * Work out, in this_chunk_blocks, how much disk we
999 			 * can add to this page
1000 			 */
1001 			this_chunk_blocks = sdio->blocks_available;
1002 			u = (to - from) >> blkbits;
1003 			if (this_chunk_blocks > u)
1004 				this_chunk_blocks = u;
1005 			u = sdio->final_block_in_request - sdio->block_in_file;
1006 			if (this_chunk_blocks > u)
1007 				this_chunk_blocks = u;
1008 			this_chunk_bytes = this_chunk_blocks << blkbits;
1009 			BUG_ON(this_chunk_bytes == 0);
1010 
1011 			if (this_chunk_blocks == sdio->blocks_available)
1012 				sdio->boundary = buffer_boundary(map_bh);
1013 			ret = submit_page_section(dio, sdio, page,
1014 						  from,
1015 						  this_chunk_bytes,
1016 						  sdio->next_block_for_io,
1017 						  map_bh);
1018 			if (ret) {
1019 				put_page(page);
1020 				goto out;
1021 			}
1022 			sdio->next_block_for_io += this_chunk_blocks;
1023 
1024 			sdio->block_in_file += this_chunk_blocks;
1025 			from += this_chunk_bytes;
1026 			dio->result += this_chunk_bytes;
1027 			sdio->blocks_available -= this_chunk_blocks;
1028 next_block:
1029 			BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1030 			if (sdio->block_in_file == sdio->final_block_in_request)
1031 				break;
1032 		}
1033 
1034 		/* Drop the ref which was taken in get_user_pages() */
1035 		put_page(page);
1036 	}
1037 out:
1038 	return ret;
1039 }
1040 
1041 static inline int drop_refcount(struct dio *dio)
1042 {
1043 	int ret2;
1044 	unsigned long flags;
1045 
1046 	/*
1047 	 * Sync will always be dropping the final ref and completing the
1048 	 * operation.  AIO can if it was a broken operation described above or
1049 	 * in fact if all the bios race to complete before we get here.  In
1050 	 * that case dio_complete() translates the EIOCBQUEUED into the proper
1051 	 * return code that the caller will hand to ->complete().
1052 	 *
1053 	 * This is managed by the bio_lock instead of being an atomic_t so that
1054 	 * completion paths can drop their ref and use the remaining count to
1055 	 * decide to wake the submission path atomically.
1056 	 */
1057 	spin_lock_irqsave(&dio->bio_lock, flags);
1058 	ret2 = --dio->refcount;
1059 	spin_unlock_irqrestore(&dio->bio_lock, flags);
1060 	return ret2;
1061 }
1062 
1063 /*
1064  * This is a library function for use by filesystem drivers.
1065  *
1066  * The locking rules are governed by the flags parameter:
1067  *  - if the flags value contains DIO_LOCKING we use a fancy locking
1068  *    scheme for dumb filesystems.
1069  *    For writes this function is called under i_mutex and returns with
1070  *    i_mutex held, for reads, i_mutex is not held on entry, but it is
1071  *    taken and dropped again before returning.
1072  *  - if the flags value does NOT contain DIO_LOCKING we don't use any
1073  *    internal locking but rather rely on the filesystem to synchronize
1074  *    direct I/O reads/writes versus each other and truncate.
1075  *
1076  * To help with locking against truncate we incremented the i_dio_count
1077  * counter before starting direct I/O, and decrement it once we are done.
1078  * Truncate can wait for it to reach zero to provide exclusion.  It is
1079  * expected that filesystem provide exclusion between new direct I/O
1080  * and truncates.  For DIO_LOCKING filesystems this is done by i_mutex,
1081  * but other filesystems need to take care of this on their own.
1082  *
1083  * NOTE: if you pass "sdio" to anything by pointer make sure that function
1084  * is always inlined. Otherwise gcc is unable to split the structure into
1085  * individual fields and will generate much worse code. This is important
1086  * for the whole file.
1087  */
1088 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1089 		struct block_device *bdev, struct iov_iter *iter,
1090 		get_block_t get_block, dio_iodone_t end_io,
1091 		int flags)
1092 {
1093 	unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
1094 	unsigned blkbits = i_blkbits;
1095 	unsigned blocksize_mask = (1 << blkbits) - 1;
1096 	ssize_t retval = -EINVAL;
1097 	const size_t count = iov_iter_count(iter);
1098 	loff_t offset = iocb->ki_pos;
1099 	const loff_t end = offset + count;
1100 	struct dio *dio;
1101 	struct dio_submit sdio = { 0, };
1102 	struct buffer_head map_bh = { 0, };
1103 	struct blk_plug plug;
1104 	unsigned long align = offset | iov_iter_alignment(iter);
1105 
1106 	/*
1107 	 * Avoid references to bdev if not absolutely needed to give
1108 	 * the early prefetch in the caller enough time.
1109 	 */
1110 
1111 	/* watch out for a 0 len io from a tricksy fs */
1112 	if (iov_iter_rw(iter) == READ && !count)
1113 		return 0;
1114 
1115 	dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1116 	if (!dio)
1117 		return -ENOMEM;
1118 	/*
1119 	 * Believe it or not, zeroing out the page array caused a .5%
1120 	 * performance regression in a database benchmark.  So, we take
1121 	 * care to only zero out what's needed.
1122 	 */
1123 	memset(dio, 0, offsetof(struct dio, pages));
1124 
1125 	dio->flags = flags;
1126 	if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1127 		/* will be released by direct_io_worker */
1128 		inode_lock(inode);
1129 	}
1130 
1131 	/* Once we sampled i_size check for reads beyond EOF */
1132 	dio->i_size = i_size_read(inode);
1133 	if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1134 		retval = 0;
1135 		goto fail_dio;
1136 	}
1137 
1138 	if (align & blocksize_mask) {
1139 		if (bdev)
1140 			blkbits = blksize_bits(bdev_logical_block_size(bdev));
1141 		blocksize_mask = (1 << blkbits) - 1;
1142 		if (align & blocksize_mask)
1143 			goto fail_dio;
1144 	}
1145 
1146 	if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1147 		struct address_space *mapping = iocb->ki_filp->f_mapping;
1148 
1149 		retval = filemap_write_and_wait_range(mapping, offset, end - 1);
1150 		if (retval)
1151 			goto fail_dio;
1152 	}
1153 
1154 	/*
1155 	 * For file extending writes updating i_size before data writeouts
1156 	 * complete can expose uninitialized blocks in dumb filesystems.
1157 	 * In that case we need to wait for I/O completion even if asked
1158 	 * for an asynchronous write.
1159 	 */
1160 	if (is_sync_kiocb(iocb))
1161 		dio->is_async = false;
1162 	else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1163 		dio->is_async = false;
1164 	else
1165 		dio->is_async = true;
1166 
1167 	dio->inode = inode;
1168 	if (iov_iter_rw(iter) == WRITE) {
1169 		dio->opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1170 		if (iocb->ki_flags & IOCB_NOWAIT)
1171 			dio->opf |= REQ_NOWAIT;
1172 	} else {
1173 		dio->opf = REQ_OP_READ;
1174 	}
1175 
1176 	/*
1177 	 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1178 	 * so that we can call ->fsync.
1179 	 */
1180 	if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1181 		retval = 0;
1182 		if (iocb_is_dsync(iocb))
1183 			retval = dio_set_defer_completion(dio);
1184 		else if (!dio->inode->i_sb->s_dio_done_wq) {
1185 			/*
1186 			 * In case of AIO write racing with buffered read we
1187 			 * need to defer completion. We can't decide this now,
1188 			 * however the workqueue needs to be initialized here.
1189 			 */
1190 			retval = sb_init_dio_done_wq(dio->inode->i_sb);
1191 		}
1192 		if (retval)
1193 			goto fail_dio;
1194 	}
1195 
1196 	/*
1197 	 * Will be decremented at I/O completion time.
1198 	 */
1199 	inode_dio_begin(inode);
1200 
1201 	retval = 0;
1202 	sdio.blkbits = blkbits;
1203 	sdio.blkfactor = i_blkbits - blkbits;
1204 	sdio.block_in_file = offset >> blkbits;
1205 
1206 	sdio.get_block = get_block;
1207 	dio->end_io = end_io;
1208 	sdio.final_block_in_bio = -1;
1209 	sdio.next_block_for_io = -1;
1210 
1211 	dio->iocb = iocb;
1212 
1213 	spin_lock_init(&dio->bio_lock);
1214 	dio->refcount = 1;
1215 
1216 	dio->should_dirty = user_backed_iter(iter) && iov_iter_rw(iter) == READ;
1217 	sdio.iter = iter;
1218 	sdio.final_block_in_request = end >> blkbits;
1219 
1220 	/*
1221 	 * In case of non-aligned buffers, we may need 2 more
1222 	 * pages since we need to zero out first and last block.
1223 	 */
1224 	if (unlikely(sdio.blkfactor))
1225 		sdio.pages_in_io = 2;
1226 
1227 	sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1228 
1229 	blk_start_plug(&plug);
1230 
1231 	retval = do_direct_IO(dio, &sdio, &map_bh);
1232 	if (retval)
1233 		dio_cleanup(dio, &sdio);
1234 
1235 	if (retval == -ENOTBLK) {
1236 		/*
1237 		 * The remaining part of the request will be
1238 		 * handled by buffered I/O when we return
1239 		 */
1240 		retval = 0;
1241 	}
1242 	/*
1243 	 * There may be some unwritten disk at the end of a part-written
1244 	 * fs-block-sized block.  Go zero that now.
1245 	 */
1246 	dio_zero_block(dio, &sdio, 1, &map_bh);
1247 
1248 	if (sdio.cur_page) {
1249 		ssize_t ret2;
1250 
1251 		ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1252 		if (retval == 0)
1253 			retval = ret2;
1254 		put_page(sdio.cur_page);
1255 		sdio.cur_page = NULL;
1256 	}
1257 	if (sdio.bio)
1258 		dio_bio_submit(dio, &sdio);
1259 
1260 	blk_finish_plug(&plug);
1261 
1262 	/*
1263 	 * It is possible that, we return short IO due to end of file.
1264 	 * In that case, we need to release all the pages we got hold on.
1265 	 */
1266 	dio_cleanup(dio, &sdio);
1267 
1268 	/*
1269 	 * All block lookups have been performed. For READ requests
1270 	 * we can let i_mutex go now that its achieved its purpose
1271 	 * of protecting us from looking up uninitialized blocks.
1272 	 */
1273 	if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1274 		inode_unlock(dio->inode);
1275 
1276 	/*
1277 	 * The only time we want to leave bios in flight is when a successful
1278 	 * partial aio read or full aio write have been setup.  In that case
1279 	 * bio completion will call aio_complete.  The only time it's safe to
1280 	 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1281 	 * This had *better* be the only place that raises -EIOCBQUEUED.
1282 	 */
1283 	BUG_ON(retval == -EIOCBQUEUED);
1284 	if (dio->is_async && retval == 0 && dio->result &&
1285 	    (iov_iter_rw(iter) == READ || dio->result == count))
1286 		retval = -EIOCBQUEUED;
1287 	else
1288 		dio_await_completion(dio);
1289 
1290 	if (drop_refcount(dio) == 0) {
1291 		retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
1292 	} else
1293 		BUG_ON(retval != -EIOCBQUEUED);
1294 
1295 	return retval;
1296 
1297 fail_dio:
1298 	if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ)
1299 		inode_unlock(inode);
1300 
1301 	kmem_cache_free(dio_cache, dio);
1302 	return retval;
1303 }
1304 EXPORT_SYMBOL(__blockdev_direct_IO);
1305 
1306 static __init int dio_init(void)
1307 {
1308 	dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1309 	return 0;
1310 }
1311 module_init(dio_init)
1312