xref: /openbmc/linux/fs/direct-io.c (revision 8ffdff6a)
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 	dio_submit_t *submit_io;	/* IO submition function */
90 
91 	loff_t logical_offset_in_bio;	/* current first logical block in bio */
92 	sector_t final_block_in_bio;	/* current final block in bio + 1 */
93 	sector_t next_block_for_io;	/* next block to be put under IO,
94 					   in dio_blocks units */
95 
96 	/*
97 	 * Deferred addition of a page to the dio.  These variables are
98 	 * private to dio_send_cur_page(), submit_page_section() and
99 	 * dio_bio_add_page().
100 	 */
101 	struct page *cur_page;		/* The page */
102 	unsigned cur_page_offset;	/* Offset into it, in bytes */
103 	unsigned cur_page_len;		/* Nr of bytes at cur_page_offset */
104 	sector_t cur_page_block;	/* Where it starts */
105 	loff_t cur_page_fs_offset;	/* Offset in file */
106 
107 	struct iov_iter *iter;
108 	/*
109 	 * Page queue.  These variables belong to dio_refill_pages() and
110 	 * dio_get_page().
111 	 */
112 	unsigned head;			/* next page to process */
113 	unsigned tail;			/* last valid page + 1 */
114 	size_t from, to;
115 };
116 
117 /* dio_state communicated between submission path and end_io */
118 struct dio {
119 	int flags;			/* doesn't change */
120 	int op;
121 	int op_flags;
122 	blk_qc_t bio_cookie;
123 	struct gendisk *bio_disk;
124 	struct inode *inode;
125 	loff_t i_size;			/* i_size when submitted */
126 	dio_iodone_t *end_io;		/* IO completion function */
127 
128 	void *private;			/* copy from map_bh.b_private */
129 
130 	/* BIO completion state */
131 	spinlock_t bio_lock;		/* protects BIO fields below */
132 	int page_errors;		/* errno from get_user_pages() */
133 	int is_async;			/* is IO async ? */
134 	bool defer_completion;		/* defer AIO completion to workqueue? */
135 	bool should_dirty;		/* if pages should be dirtied */
136 	int io_error;			/* IO error in completion path */
137 	unsigned long refcount;		/* direct_io_worker() and bios */
138 	struct bio *bio_list;		/* singly linked via bi_private */
139 	struct task_struct *waiter;	/* waiting task (NULL if none) */
140 
141 	/* AIO related stuff */
142 	struct kiocb *iocb;		/* kiocb */
143 	ssize_t result;                 /* IO result */
144 
145 	/*
146 	 * pages[] (and any fields placed after it) are not zeroed out at
147 	 * allocation time.  Don't add new fields after pages[] unless you
148 	 * wish that they not be zeroed.
149 	 */
150 	union {
151 		struct page *pages[DIO_PAGES];	/* page buffer */
152 		struct work_struct complete_work;/* deferred AIO completion */
153 	};
154 } ____cacheline_aligned_in_smp;
155 
156 static struct kmem_cache *dio_cache __read_mostly;
157 
158 /*
159  * How many pages are in the queue?
160  */
161 static inline unsigned dio_pages_present(struct dio_submit *sdio)
162 {
163 	return sdio->tail - sdio->head;
164 }
165 
166 /*
167  * Go grab and pin some userspace pages.   Typically we'll get 64 at a time.
168  */
169 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
170 {
171 	ssize_t ret;
172 
173 	ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
174 				&sdio->from);
175 
176 	if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
177 		struct page *page = ZERO_PAGE(0);
178 		/*
179 		 * A memory fault, but the filesystem has some outstanding
180 		 * mapped blocks.  We need to use those blocks up to avoid
181 		 * leaking stale data in the file.
182 		 */
183 		if (dio->page_errors == 0)
184 			dio->page_errors = ret;
185 		get_page(page);
186 		dio->pages[0] = page;
187 		sdio->head = 0;
188 		sdio->tail = 1;
189 		sdio->from = 0;
190 		sdio->to = PAGE_SIZE;
191 		return 0;
192 	}
193 
194 	if (ret >= 0) {
195 		iov_iter_advance(sdio->iter, ret);
196 		ret += sdio->from;
197 		sdio->head = 0;
198 		sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
199 		sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
200 		return 0;
201 	}
202 	return ret;
203 }
204 
205 /*
206  * Get another userspace page.  Returns an ERR_PTR on error.  Pages are
207  * buffered inside the dio so that we can call get_user_pages() against a
208  * decent number of pages, less frequently.  To provide nicer use of the
209  * L1 cache.
210  */
211 static inline struct page *dio_get_page(struct dio *dio,
212 					struct dio_submit *sdio)
213 {
214 	if (dio_pages_present(sdio) == 0) {
215 		int ret;
216 
217 		ret = dio_refill_pages(dio, sdio);
218 		if (ret)
219 			return ERR_PTR(ret);
220 		BUG_ON(dio_pages_present(sdio) == 0);
221 	}
222 	return dio->pages[sdio->head];
223 }
224 
225 /*
226  * dio_complete() - called when all DIO BIO I/O has been completed
227  *
228  * This drops i_dio_count, lets interested parties know that a DIO operation
229  * has completed, and calculates the resulting return code for the operation.
230  *
231  * It lets the filesystem know if it registered an interest earlier via
232  * get_block.  Pass the private field of the map buffer_head so that
233  * filesystems can use it to hold additional state between get_block calls and
234  * dio_complete.
235  */
236 static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
237 {
238 	loff_t offset = dio->iocb->ki_pos;
239 	ssize_t transferred = 0;
240 	int err;
241 
242 	/*
243 	 * AIO submission can race with bio completion to get here while
244 	 * expecting to have the last io completed by bio completion.
245 	 * In that case -EIOCBQUEUED is in fact not an error we want
246 	 * to preserve through this call.
247 	 */
248 	if (ret == -EIOCBQUEUED)
249 		ret = 0;
250 
251 	if (dio->result) {
252 		transferred = dio->result;
253 
254 		/* Check for short read case */
255 		if ((dio->op == REQ_OP_READ) &&
256 		    ((offset + transferred) > dio->i_size))
257 			transferred = dio->i_size - offset;
258 		/* ignore EFAULT if some IO has been done */
259 		if (unlikely(ret == -EFAULT) && transferred)
260 			ret = 0;
261 	}
262 
263 	if (ret == 0)
264 		ret = dio->page_errors;
265 	if (ret == 0)
266 		ret = dio->io_error;
267 	if (ret == 0)
268 		ret = transferred;
269 
270 	if (dio->end_io) {
271 		// XXX: ki_pos??
272 		err = dio->end_io(dio->iocb, offset, ret, dio->private);
273 		if (err)
274 			ret = err;
275 	}
276 
277 	/*
278 	 * Try again to invalidate clean pages which might have been cached by
279 	 * non-direct readahead, or faulted in by get_user_pages() if the source
280 	 * of the write was an mmap'ed region of the file we're writing.  Either
281 	 * one is a pretty crazy thing to do, so we don't support it 100%.  If
282 	 * this invalidation fails, tough, the write still worked...
283 	 *
284 	 * And this page cache invalidation has to be after dio->end_io(), as
285 	 * some filesystems convert unwritten extents to real allocations in
286 	 * end_io() when necessary, otherwise a racing buffer read would cache
287 	 * zeros from unwritten extents.
288 	 */
289 	if (flags & DIO_COMPLETE_INVALIDATE &&
290 	    ret > 0 && dio->op == REQ_OP_WRITE &&
291 	    dio->inode->i_mapping->nrpages) {
292 		err = invalidate_inode_pages2_range(dio->inode->i_mapping,
293 					offset >> PAGE_SHIFT,
294 					(offset + ret - 1) >> PAGE_SHIFT);
295 		if (err)
296 			dio_warn_stale_pagecache(dio->iocb->ki_filp);
297 	}
298 
299 	inode_dio_end(dio->inode);
300 
301 	if (flags & DIO_COMPLETE_ASYNC) {
302 		/*
303 		 * generic_write_sync expects ki_pos to have been updated
304 		 * already, but the submission path only does this for
305 		 * synchronous I/O.
306 		 */
307 		dio->iocb->ki_pos += transferred;
308 
309 		if (ret > 0 && dio->op == REQ_OP_WRITE)
310 			ret = generic_write_sync(dio->iocb, ret);
311 		dio->iocb->ki_complete(dio->iocb, ret, 0);
312 	}
313 
314 	kmem_cache_free(dio_cache, dio);
315 	return ret;
316 }
317 
318 static void dio_aio_complete_work(struct work_struct *work)
319 {
320 	struct dio *dio = container_of(work, struct dio, complete_work);
321 
322 	dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
323 }
324 
325 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
326 
327 /*
328  * Asynchronous IO callback.
329  */
330 static void dio_bio_end_aio(struct bio *bio)
331 {
332 	struct dio *dio = bio->bi_private;
333 	unsigned long remaining;
334 	unsigned long flags;
335 	bool defer_completion = false;
336 
337 	/* cleanup the bio */
338 	dio_bio_complete(dio, bio);
339 
340 	spin_lock_irqsave(&dio->bio_lock, flags);
341 	remaining = --dio->refcount;
342 	if (remaining == 1 && dio->waiter)
343 		wake_up_process(dio->waiter);
344 	spin_unlock_irqrestore(&dio->bio_lock, flags);
345 
346 	if (remaining == 0) {
347 		/*
348 		 * Defer completion when defer_completion is set or
349 		 * when the inode has pages mapped and this is AIO write.
350 		 * We need to invalidate those pages because there is a
351 		 * chance they contain stale data in the case buffered IO
352 		 * went in between AIO submission and completion into the
353 		 * same region.
354 		 */
355 		if (dio->result)
356 			defer_completion = dio->defer_completion ||
357 					   (dio->op == REQ_OP_WRITE &&
358 					    dio->inode->i_mapping->nrpages);
359 		if (defer_completion) {
360 			INIT_WORK(&dio->complete_work, dio_aio_complete_work);
361 			queue_work(dio->inode->i_sb->s_dio_done_wq,
362 				   &dio->complete_work);
363 		} else {
364 			dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
365 		}
366 	}
367 }
368 
369 /*
370  * The BIO completion handler simply queues the BIO up for the process-context
371  * handler.
372  *
373  * During I/O bi_private points at the dio.  After I/O, bi_private is used to
374  * implement a singly-linked list of completed BIOs, at dio->bio_list.
375  */
376 static void dio_bio_end_io(struct bio *bio)
377 {
378 	struct dio *dio = bio->bi_private;
379 	unsigned long flags;
380 
381 	spin_lock_irqsave(&dio->bio_lock, flags);
382 	bio->bi_private = dio->bio_list;
383 	dio->bio_list = bio;
384 	if (--dio->refcount == 1 && dio->waiter)
385 		wake_up_process(dio->waiter);
386 	spin_unlock_irqrestore(&dio->bio_lock, flags);
387 }
388 
389 static inline void
390 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
391 	      struct block_device *bdev,
392 	      sector_t first_sector, int nr_vecs)
393 {
394 	struct bio *bio;
395 
396 	/*
397 	 * bio_alloc() is guaranteed to return a bio when allowed to sleep and
398 	 * we request a valid number of vectors.
399 	 */
400 	bio = bio_alloc(GFP_KERNEL, nr_vecs);
401 
402 	bio_set_dev(bio, bdev);
403 	bio->bi_iter.bi_sector = first_sector;
404 	bio_set_op_attrs(bio, dio->op, dio->op_flags);
405 	if (dio->is_async)
406 		bio->bi_end_io = dio_bio_end_aio;
407 	else
408 		bio->bi_end_io = dio_bio_end_io;
409 
410 	bio->bi_write_hint = dio->iocb->ki_hint;
411 
412 	sdio->bio = bio;
413 	sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
414 }
415 
416 /*
417  * In the AIO read case we speculatively dirty the pages before starting IO.
418  * During IO completion, any of these pages which happen to have been written
419  * back will be redirtied by bio_check_pages_dirty().
420  *
421  * bios hold a dio reference between submit_bio and ->end_io.
422  */
423 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
424 {
425 	struct bio *bio = sdio->bio;
426 	unsigned long flags;
427 
428 	bio->bi_private = dio;
429 	/* don't account direct I/O as memory stall */
430 	bio_clear_flag(bio, BIO_WORKINGSET);
431 
432 	spin_lock_irqsave(&dio->bio_lock, flags);
433 	dio->refcount++;
434 	spin_unlock_irqrestore(&dio->bio_lock, flags);
435 
436 	if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
437 		bio_set_pages_dirty(bio);
438 
439 	dio->bio_disk = bio->bi_bdev->bd_disk;
440 
441 	if (sdio->submit_io) {
442 		sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
443 		dio->bio_cookie = BLK_QC_T_NONE;
444 	} else
445 		dio->bio_cookie = submit_bio(bio);
446 
447 	sdio->bio = NULL;
448 	sdio->boundary = 0;
449 	sdio->logical_offset_in_bio = 0;
450 }
451 
452 /*
453  * Release any resources in case of a failure
454  */
455 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
456 {
457 	while (sdio->head < sdio->tail)
458 		put_page(dio->pages[sdio->head++]);
459 }
460 
461 /*
462  * Wait for the next BIO to complete.  Remove it and return it.  NULL is
463  * returned once all BIOs have been completed.  This must only be called once
464  * all bios have been issued so that dio->refcount can only decrease.  This
465  * requires that the caller hold a reference on the dio.
466  */
467 static struct bio *dio_await_one(struct dio *dio)
468 {
469 	unsigned long flags;
470 	struct bio *bio = NULL;
471 
472 	spin_lock_irqsave(&dio->bio_lock, flags);
473 
474 	/*
475 	 * Wait as long as the list is empty and there are bios in flight.  bio
476 	 * completion drops the count, maybe adds to the list, and wakes while
477 	 * holding the bio_lock so we don't need set_current_state()'s barrier
478 	 * and can call it after testing our condition.
479 	 */
480 	while (dio->refcount > 1 && dio->bio_list == NULL) {
481 		__set_current_state(TASK_UNINTERRUPTIBLE);
482 		dio->waiter = current;
483 		spin_unlock_irqrestore(&dio->bio_lock, flags);
484 		if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
485 		    !blk_poll(dio->bio_disk->queue, dio->bio_cookie, true))
486 			blk_io_schedule();
487 		/* wake up sets us TASK_RUNNING */
488 		spin_lock_irqsave(&dio->bio_lock, flags);
489 		dio->waiter = NULL;
490 	}
491 	if (dio->bio_list) {
492 		bio = dio->bio_list;
493 		dio->bio_list = bio->bi_private;
494 	}
495 	spin_unlock_irqrestore(&dio->bio_lock, flags);
496 	return bio;
497 }
498 
499 /*
500  * Process one completed BIO.  No locks are held.
501  */
502 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
503 {
504 	blk_status_t err = bio->bi_status;
505 	bool should_dirty = dio->op == REQ_OP_READ && dio->should_dirty;
506 
507 	if (err) {
508 		if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
509 			dio->io_error = -EAGAIN;
510 		else
511 			dio->io_error = -EIO;
512 	}
513 
514 	if (dio->is_async && should_dirty) {
515 		bio_check_pages_dirty(bio);	/* transfers ownership */
516 	} else {
517 		bio_release_pages(bio, should_dirty);
518 		bio_put(bio);
519 	}
520 	return err;
521 }
522 
523 /*
524  * Wait on and process all in-flight BIOs.  This must only be called once
525  * all bios have been issued so that the refcount can only decrease.
526  * This just waits for all bios to make it through dio_bio_complete.  IO
527  * errors are propagated through dio->io_error and should be propagated via
528  * dio_complete().
529  */
530 static void dio_await_completion(struct dio *dio)
531 {
532 	struct bio *bio;
533 	do {
534 		bio = dio_await_one(dio);
535 		if (bio)
536 			dio_bio_complete(dio, bio);
537 	} while (bio);
538 }
539 
540 /*
541  * A really large O_DIRECT read or write can generate a lot of BIOs.  So
542  * to keep the memory consumption sane we periodically reap any completed BIOs
543  * during the BIO generation phase.
544  *
545  * This also helps to limit the peak amount of pinned userspace memory.
546  */
547 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
548 {
549 	int ret = 0;
550 
551 	if (sdio->reap_counter++ >= 64) {
552 		while (dio->bio_list) {
553 			unsigned long flags;
554 			struct bio *bio;
555 			int ret2;
556 
557 			spin_lock_irqsave(&dio->bio_lock, flags);
558 			bio = dio->bio_list;
559 			dio->bio_list = bio->bi_private;
560 			spin_unlock_irqrestore(&dio->bio_lock, flags);
561 			ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
562 			if (ret == 0)
563 				ret = ret2;
564 		}
565 		sdio->reap_counter = 0;
566 	}
567 	return ret;
568 }
569 
570 /*
571  * Create workqueue for deferred direct IO completions. We allocate the
572  * workqueue when it's first needed. This avoids creating workqueue for
573  * filesystems that don't need it and also allows us to create the workqueue
574  * late enough so the we can include s_id in the name of the workqueue.
575  */
576 int sb_init_dio_done_wq(struct super_block *sb)
577 {
578 	struct workqueue_struct *old;
579 	struct workqueue_struct *wq = alloc_workqueue("dio/%s",
580 						      WQ_MEM_RECLAIM, 0,
581 						      sb->s_id);
582 	if (!wq)
583 		return -ENOMEM;
584 	/*
585 	 * This has to be atomic as more DIOs can race to create the workqueue
586 	 */
587 	old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
588 	/* Someone created workqueue before us? Free ours... */
589 	if (old)
590 		destroy_workqueue(wq);
591 	return 0;
592 }
593 
594 static int dio_set_defer_completion(struct dio *dio)
595 {
596 	struct super_block *sb = dio->inode->i_sb;
597 
598 	if (dio->defer_completion)
599 		return 0;
600 	dio->defer_completion = true;
601 	if (!sb->s_dio_done_wq)
602 		return sb_init_dio_done_wq(sb);
603 	return 0;
604 }
605 
606 /*
607  * Call into the fs to map some more disk blocks.  We record the current number
608  * of available blocks at sdio->blocks_available.  These are in units of the
609  * fs blocksize, i_blocksize(inode).
610  *
611  * The fs is allowed to map lots of blocks at once.  If it wants to do that,
612  * it uses the passed inode-relative block number as the file offset, as usual.
613  *
614  * get_block() is passed the number of i_blkbits-sized blocks which direct_io
615  * has remaining to do.  The fs should not map more than this number of blocks.
616  *
617  * If the fs has mapped a lot of blocks, it should populate bh->b_size to
618  * indicate how much contiguous disk space has been made available at
619  * bh->b_blocknr.
620  *
621  * If *any* of the mapped blocks are new, then the fs must set buffer_new().
622  * This isn't very efficient...
623  *
624  * In the case of filesystem holes: the fs may return an arbitrarily-large
625  * hole by returning an appropriate value in b_size and by clearing
626  * buffer_mapped().  However the direct-io code will only process holes one
627  * block at a time - it will repeatedly call get_block() as it walks the hole.
628  */
629 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
630 			   struct buffer_head *map_bh)
631 {
632 	int ret;
633 	sector_t fs_startblk;	/* Into file, in filesystem-sized blocks */
634 	sector_t fs_endblk;	/* Into file, in filesystem-sized blocks */
635 	unsigned long fs_count;	/* Number of filesystem-sized blocks */
636 	int create;
637 	unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
638 	loff_t i_size;
639 
640 	/*
641 	 * If there was a memory error and we've overwritten all the
642 	 * mapped blocks then we can now return that memory error
643 	 */
644 	ret = dio->page_errors;
645 	if (ret == 0) {
646 		BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
647 		fs_startblk = sdio->block_in_file >> sdio->blkfactor;
648 		fs_endblk = (sdio->final_block_in_request - 1) >>
649 					sdio->blkfactor;
650 		fs_count = fs_endblk - fs_startblk + 1;
651 
652 		map_bh->b_state = 0;
653 		map_bh->b_size = fs_count << i_blkbits;
654 
655 		/*
656 		 * For writes that could fill holes inside i_size on a
657 		 * DIO_SKIP_HOLES filesystem we forbid block creations: only
658 		 * overwrites are permitted. We will return early to the caller
659 		 * once we see an unmapped buffer head returned, and the caller
660 		 * will fall back to buffered I/O.
661 		 *
662 		 * Otherwise the decision is left to the get_blocks method,
663 		 * which may decide to handle it or also return an unmapped
664 		 * buffer head.
665 		 */
666 		create = dio->op == REQ_OP_WRITE;
667 		if (dio->flags & DIO_SKIP_HOLES) {
668 			i_size = i_size_read(dio->inode);
669 			if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
670 				create = 0;
671 		}
672 
673 		ret = (*sdio->get_block)(dio->inode, fs_startblk,
674 						map_bh, create);
675 
676 		/* Store for completion */
677 		dio->private = map_bh->b_private;
678 
679 		if (ret == 0 && buffer_defer_completion(map_bh))
680 			ret = dio_set_defer_completion(dio);
681 	}
682 	return ret;
683 }
684 
685 /*
686  * There is no bio.  Make one now.
687  */
688 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
689 		sector_t start_sector, struct buffer_head *map_bh)
690 {
691 	sector_t sector;
692 	int ret, nr_pages;
693 
694 	ret = dio_bio_reap(dio, sdio);
695 	if (ret)
696 		goto out;
697 	sector = start_sector << (sdio->blkbits - 9);
698 	nr_pages = bio_max_segs(sdio->pages_in_io);
699 	BUG_ON(nr_pages <= 0);
700 	dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
701 	sdio->boundary = 0;
702 out:
703 	return ret;
704 }
705 
706 /*
707  * Attempt to put the current chunk of 'cur_page' into the current BIO.  If
708  * that was successful then update final_block_in_bio and take a ref against
709  * the just-added page.
710  *
711  * Return zero on success.  Non-zero means the caller needs to start a new BIO.
712  */
713 static inline int dio_bio_add_page(struct dio_submit *sdio)
714 {
715 	int ret;
716 
717 	ret = bio_add_page(sdio->bio, sdio->cur_page,
718 			sdio->cur_page_len, sdio->cur_page_offset);
719 	if (ret == sdio->cur_page_len) {
720 		/*
721 		 * Decrement count only, if we are done with this page
722 		 */
723 		if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
724 			sdio->pages_in_io--;
725 		get_page(sdio->cur_page);
726 		sdio->final_block_in_bio = sdio->cur_page_block +
727 			(sdio->cur_page_len >> sdio->blkbits);
728 		ret = 0;
729 	} else {
730 		ret = 1;
731 	}
732 	return ret;
733 }
734 
735 /*
736  * Put cur_page under IO.  The section of cur_page which is described by
737  * cur_page_offset,cur_page_len is put into a BIO.  The section of cur_page
738  * starts on-disk at cur_page_block.
739  *
740  * We take a ref against the page here (on behalf of its presence in the bio).
741  *
742  * The caller of this function is responsible for removing cur_page from the
743  * dio, and for dropping the refcount which came from that presence.
744  */
745 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
746 		struct buffer_head *map_bh)
747 {
748 	int ret = 0;
749 
750 	if (sdio->bio) {
751 		loff_t cur_offset = sdio->cur_page_fs_offset;
752 		loff_t bio_next_offset = sdio->logical_offset_in_bio +
753 			sdio->bio->bi_iter.bi_size;
754 
755 		/*
756 		 * See whether this new request is contiguous with the old.
757 		 *
758 		 * Btrfs cannot handle having logically non-contiguous requests
759 		 * submitted.  For example if you have
760 		 *
761 		 * Logical:  [0-4095][HOLE][8192-12287]
762 		 * Physical: [0-4095]      [4096-8191]
763 		 *
764 		 * We cannot submit those pages together as one BIO.  So if our
765 		 * current logical offset in the file does not equal what would
766 		 * be the next logical offset in the bio, submit the bio we
767 		 * have.
768 		 */
769 		if (sdio->final_block_in_bio != sdio->cur_page_block ||
770 		    cur_offset != bio_next_offset)
771 			dio_bio_submit(dio, sdio);
772 	}
773 
774 	if (sdio->bio == NULL) {
775 		ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
776 		if (ret)
777 			goto out;
778 	}
779 
780 	if (dio_bio_add_page(sdio) != 0) {
781 		dio_bio_submit(dio, sdio);
782 		ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
783 		if (ret == 0) {
784 			ret = dio_bio_add_page(sdio);
785 			BUG_ON(ret != 0);
786 		}
787 	}
788 out:
789 	return ret;
790 }
791 
792 /*
793  * An autonomous function to put a chunk of a page under deferred IO.
794  *
795  * The caller doesn't actually know (or care) whether this piece of page is in
796  * a BIO, or is under IO or whatever.  We just take care of all possible
797  * situations here.  The separation between the logic of do_direct_IO() and
798  * that of submit_page_section() is important for clarity.  Please don't break.
799  *
800  * The chunk of page starts on-disk at blocknr.
801  *
802  * We perform deferred IO, by recording the last-submitted page inside our
803  * private part of the dio structure.  If possible, we just expand the IO
804  * across that page here.
805  *
806  * If that doesn't work out then we put the old page into the bio and add this
807  * page to the dio instead.
808  */
809 static inline int
810 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
811 		    unsigned offset, unsigned len, sector_t blocknr,
812 		    struct buffer_head *map_bh)
813 {
814 	int ret = 0;
815 
816 	if (dio->op == REQ_OP_WRITE) {
817 		/*
818 		 * Read accounting is performed in submit_bio()
819 		 */
820 		task_io_account_write(len);
821 	}
822 
823 	/*
824 	 * Can we just grow the current page's presence in the dio?
825 	 */
826 	if (sdio->cur_page == page &&
827 	    sdio->cur_page_offset + sdio->cur_page_len == offset &&
828 	    sdio->cur_page_block +
829 	    (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
830 		sdio->cur_page_len += len;
831 		goto out;
832 	}
833 
834 	/*
835 	 * If there's a deferred page already there then send it.
836 	 */
837 	if (sdio->cur_page) {
838 		ret = dio_send_cur_page(dio, sdio, map_bh);
839 		put_page(sdio->cur_page);
840 		sdio->cur_page = NULL;
841 		if (ret)
842 			return ret;
843 	}
844 
845 	get_page(page);		/* It is in dio */
846 	sdio->cur_page = page;
847 	sdio->cur_page_offset = offset;
848 	sdio->cur_page_len = len;
849 	sdio->cur_page_block = blocknr;
850 	sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
851 out:
852 	/*
853 	 * If sdio->boundary then we want to schedule the IO now to
854 	 * avoid metadata seeks.
855 	 */
856 	if (sdio->boundary) {
857 		ret = dio_send_cur_page(dio, sdio, map_bh);
858 		if (sdio->bio)
859 			dio_bio_submit(dio, sdio);
860 		put_page(sdio->cur_page);
861 		sdio->cur_page = NULL;
862 	}
863 	return ret;
864 }
865 
866 /*
867  * If we are not writing the entire block and get_block() allocated
868  * the block for us, we need to fill-in the unused portion of the
869  * block with zeros. This happens only if user-buffer, fileoffset or
870  * io length is not filesystem block-size multiple.
871  *
872  * `end' is zero if we're doing the start of the IO, 1 at the end of the
873  * IO.
874  */
875 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
876 		int end, struct buffer_head *map_bh)
877 {
878 	unsigned dio_blocks_per_fs_block;
879 	unsigned this_chunk_blocks;	/* In dio_blocks */
880 	unsigned this_chunk_bytes;
881 	struct page *page;
882 
883 	sdio->start_zero_done = 1;
884 	if (!sdio->blkfactor || !buffer_new(map_bh))
885 		return;
886 
887 	dio_blocks_per_fs_block = 1 << sdio->blkfactor;
888 	this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
889 
890 	if (!this_chunk_blocks)
891 		return;
892 
893 	/*
894 	 * We need to zero out part of an fs block.  It is either at the
895 	 * beginning or the end of the fs block.
896 	 */
897 	if (end)
898 		this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
899 
900 	this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
901 
902 	page = ZERO_PAGE(0);
903 	if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
904 				sdio->next_block_for_io, map_bh))
905 		return;
906 
907 	sdio->next_block_for_io += this_chunk_blocks;
908 }
909 
910 /*
911  * Walk the user pages, and the file, mapping blocks to disk and generating
912  * a sequence of (page,offset,len,block) mappings.  These mappings are injected
913  * into submit_page_section(), which takes care of the next stage of submission
914  *
915  * Direct IO against a blockdev is different from a file.  Because we can
916  * happily perform page-sized but 512-byte aligned IOs.  It is important that
917  * blockdev IO be able to have fine alignment and large sizes.
918  *
919  * So what we do is to permit the ->get_block function to populate bh.b_size
920  * with the size of IO which is permitted at this offset and this i_blkbits.
921  *
922  * For best results, the blockdev should be set up with 512-byte i_blkbits and
923  * it should set b_size to PAGE_SIZE or more inside get_block().  This gives
924  * fine alignment but still allows this function to work in PAGE_SIZE units.
925  */
926 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
927 			struct buffer_head *map_bh)
928 {
929 	const unsigned blkbits = sdio->blkbits;
930 	const unsigned i_blkbits = blkbits + sdio->blkfactor;
931 	int ret = 0;
932 
933 	while (sdio->block_in_file < sdio->final_block_in_request) {
934 		struct page *page;
935 		size_t from, to;
936 
937 		page = dio_get_page(dio, sdio);
938 		if (IS_ERR(page)) {
939 			ret = PTR_ERR(page);
940 			goto out;
941 		}
942 		from = sdio->head ? 0 : sdio->from;
943 		to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
944 		sdio->head++;
945 
946 		while (from < to) {
947 			unsigned this_chunk_bytes;	/* # of bytes mapped */
948 			unsigned this_chunk_blocks;	/* # of blocks */
949 			unsigned u;
950 
951 			if (sdio->blocks_available == 0) {
952 				/*
953 				 * Need to go and map some more disk
954 				 */
955 				unsigned long blkmask;
956 				unsigned long dio_remainder;
957 
958 				ret = get_more_blocks(dio, sdio, map_bh);
959 				if (ret) {
960 					put_page(page);
961 					goto out;
962 				}
963 				if (!buffer_mapped(map_bh))
964 					goto do_holes;
965 
966 				sdio->blocks_available =
967 						map_bh->b_size >> blkbits;
968 				sdio->next_block_for_io =
969 					map_bh->b_blocknr << sdio->blkfactor;
970 				if (buffer_new(map_bh)) {
971 					clean_bdev_aliases(
972 						map_bh->b_bdev,
973 						map_bh->b_blocknr,
974 						map_bh->b_size >> i_blkbits);
975 				}
976 
977 				if (!sdio->blkfactor)
978 					goto do_holes;
979 
980 				blkmask = (1 << sdio->blkfactor) - 1;
981 				dio_remainder = (sdio->block_in_file & blkmask);
982 
983 				/*
984 				 * If we are at the start of IO and that IO
985 				 * starts partway into a fs-block,
986 				 * dio_remainder will be non-zero.  If the IO
987 				 * is a read then we can simply advance the IO
988 				 * cursor to the first block which is to be
989 				 * read.  But if the IO is a write and the
990 				 * block was newly allocated we cannot do that;
991 				 * the start of the fs block must be zeroed out
992 				 * on-disk
993 				 */
994 				if (!buffer_new(map_bh))
995 					sdio->next_block_for_io += dio_remainder;
996 				sdio->blocks_available -= dio_remainder;
997 			}
998 do_holes:
999 			/* Handle holes */
1000 			if (!buffer_mapped(map_bh)) {
1001 				loff_t i_size_aligned;
1002 
1003 				/* AKPM: eargh, -ENOTBLK is a hack */
1004 				if (dio->op == REQ_OP_WRITE) {
1005 					put_page(page);
1006 					return -ENOTBLK;
1007 				}
1008 
1009 				/*
1010 				 * Be sure to account for a partial block as the
1011 				 * last block in the file
1012 				 */
1013 				i_size_aligned = ALIGN(i_size_read(dio->inode),
1014 							1 << blkbits);
1015 				if (sdio->block_in_file >=
1016 						i_size_aligned >> blkbits) {
1017 					/* We hit eof */
1018 					put_page(page);
1019 					goto out;
1020 				}
1021 				zero_user(page, from, 1 << blkbits);
1022 				sdio->block_in_file++;
1023 				from += 1 << blkbits;
1024 				dio->result += 1 << blkbits;
1025 				goto next_block;
1026 			}
1027 
1028 			/*
1029 			 * If we're performing IO which has an alignment which
1030 			 * is finer than the underlying fs, go check to see if
1031 			 * we must zero out the start of this block.
1032 			 */
1033 			if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1034 				dio_zero_block(dio, sdio, 0, map_bh);
1035 
1036 			/*
1037 			 * Work out, in this_chunk_blocks, how much disk we
1038 			 * can add to this page
1039 			 */
1040 			this_chunk_blocks = sdio->blocks_available;
1041 			u = (to - from) >> blkbits;
1042 			if (this_chunk_blocks > u)
1043 				this_chunk_blocks = u;
1044 			u = sdio->final_block_in_request - sdio->block_in_file;
1045 			if (this_chunk_blocks > u)
1046 				this_chunk_blocks = u;
1047 			this_chunk_bytes = this_chunk_blocks << blkbits;
1048 			BUG_ON(this_chunk_bytes == 0);
1049 
1050 			if (this_chunk_blocks == sdio->blocks_available)
1051 				sdio->boundary = buffer_boundary(map_bh);
1052 			ret = submit_page_section(dio, sdio, page,
1053 						  from,
1054 						  this_chunk_bytes,
1055 						  sdio->next_block_for_io,
1056 						  map_bh);
1057 			if (ret) {
1058 				put_page(page);
1059 				goto out;
1060 			}
1061 			sdio->next_block_for_io += this_chunk_blocks;
1062 
1063 			sdio->block_in_file += this_chunk_blocks;
1064 			from += this_chunk_bytes;
1065 			dio->result += this_chunk_bytes;
1066 			sdio->blocks_available -= this_chunk_blocks;
1067 next_block:
1068 			BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1069 			if (sdio->block_in_file == sdio->final_block_in_request)
1070 				break;
1071 		}
1072 
1073 		/* Drop the ref which was taken in get_user_pages() */
1074 		put_page(page);
1075 	}
1076 out:
1077 	return ret;
1078 }
1079 
1080 static inline int drop_refcount(struct dio *dio)
1081 {
1082 	int ret2;
1083 	unsigned long flags;
1084 
1085 	/*
1086 	 * Sync will always be dropping the final ref and completing the
1087 	 * operation.  AIO can if it was a broken operation described above or
1088 	 * in fact if all the bios race to complete before we get here.  In
1089 	 * that case dio_complete() translates the EIOCBQUEUED into the proper
1090 	 * return code that the caller will hand to ->complete().
1091 	 *
1092 	 * This is managed by the bio_lock instead of being an atomic_t so that
1093 	 * completion paths can drop their ref and use the remaining count to
1094 	 * decide to wake the submission path atomically.
1095 	 */
1096 	spin_lock_irqsave(&dio->bio_lock, flags);
1097 	ret2 = --dio->refcount;
1098 	spin_unlock_irqrestore(&dio->bio_lock, flags);
1099 	return ret2;
1100 }
1101 
1102 /*
1103  * This is a library function for use by filesystem drivers.
1104  *
1105  * The locking rules are governed by the flags parameter:
1106  *  - if the flags value contains DIO_LOCKING we use a fancy locking
1107  *    scheme for dumb filesystems.
1108  *    For writes this function is called under i_mutex and returns with
1109  *    i_mutex held, for reads, i_mutex is not held on entry, but it is
1110  *    taken and dropped again before returning.
1111  *  - if the flags value does NOT contain DIO_LOCKING we don't use any
1112  *    internal locking but rather rely on the filesystem to synchronize
1113  *    direct I/O reads/writes versus each other and truncate.
1114  *
1115  * To help with locking against truncate we incremented the i_dio_count
1116  * counter before starting direct I/O, and decrement it once we are done.
1117  * Truncate can wait for it to reach zero to provide exclusion.  It is
1118  * expected that filesystem provide exclusion between new direct I/O
1119  * and truncates.  For DIO_LOCKING filesystems this is done by i_mutex,
1120  * but other filesystems need to take care of this on their own.
1121  *
1122  * NOTE: if you pass "sdio" to anything by pointer make sure that function
1123  * is always inlined. Otherwise gcc is unable to split the structure into
1124  * individual fields and will generate much worse code. This is important
1125  * for the whole file.
1126  */
1127 static inline ssize_t
1128 do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1129 		      struct block_device *bdev, struct iov_iter *iter,
1130 		      get_block_t get_block, dio_iodone_t end_io,
1131 		      dio_submit_t submit_io, int flags)
1132 {
1133 	unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
1134 	unsigned blkbits = i_blkbits;
1135 	unsigned blocksize_mask = (1 << blkbits) - 1;
1136 	ssize_t retval = -EINVAL;
1137 	const size_t count = iov_iter_count(iter);
1138 	loff_t offset = iocb->ki_pos;
1139 	const loff_t end = offset + count;
1140 	struct dio *dio;
1141 	struct dio_submit sdio = { 0, };
1142 	struct buffer_head map_bh = { 0, };
1143 	struct blk_plug plug;
1144 	unsigned long align = offset | iov_iter_alignment(iter);
1145 
1146 	/*
1147 	 * Avoid references to bdev if not absolutely needed to give
1148 	 * the early prefetch in the caller enough time.
1149 	 */
1150 
1151 	/* watch out for a 0 len io from a tricksy fs */
1152 	if (iov_iter_rw(iter) == READ && !count)
1153 		return 0;
1154 
1155 	dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1156 	if (!dio)
1157 		return -ENOMEM;
1158 	/*
1159 	 * Believe it or not, zeroing out the page array caused a .5%
1160 	 * performance regression in a database benchmark.  So, we take
1161 	 * care to only zero out what's needed.
1162 	 */
1163 	memset(dio, 0, offsetof(struct dio, pages));
1164 
1165 	dio->flags = flags;
1166 	if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1167 		/* will be released by direct_io_worker */
1168 		inode_lock(inode);
1169 	}
1170 
1171 	/* Once we sampled i_size check for reads beyond EOF */
1172 	dio->i_size = i_size_read(inode);
1173 	if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1174 		retval = 0;
1175 		goto fail_dio;
1176 	}
1177 
1178 	if (align & blocksize_mask) {
1179 		if (bdev)
1180 			blkbits = blksize_bits(bdev_logical_block_size(bdev));
1181 		blocksize_mask = (1 << blkbits) - 1;
1182 		if (align & blocksize_mask)
1183 			goto fail_dio;
1184 	}
1185 
1186 	if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1187 		struct address_space *mapping = iocb->ki_filp->f_mapping;
1188 
1189 		retval = filemap_write_and_wait_range(mapping, offset, end - 1);
1190 		if (retval)
1191 			goto fail_dio;
1192 	}
1193 
1194 	/*
1195 	 * For file extending writes updating i_size before data writeouts
1196 	 * complete can expose uninitialized blocks in dumb filesystems.
1197 	 * In that case we need to wait for I/O completion even if asked
1198 	 * for an asynchronous write.
1199 	 */
1200 	if (is_sync_kiocb(iocb))
1201 		dio->is_async = false;
1202 	else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1203 		dio->is_async = false;
1204 	else
1205 		dio->is_async = true;
1206 
1207 	dio->inode = inode;
1208 	if (iov_iter_rw(iter) == WRITE) {
1209 		dio->op = REQ_OP_WRITE;
1210 		dio->op_flags = REQ_SYNC | REQ_IDLE;
1211 		if (iocb->ki_flags & IOCB_NOWAIT)
1212 			dio->op_flags |= REQ_NOWAIT;
1213 	} else {
1214 		dio->op = REQ_OP_READ;
1215 	}
1216 	if (iocb->ki_flags & IOCB_HIPRI)
1217 		dio->op_flags |= REQ_HIPRI;
1218 
1219 	/*
1220 	 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1221 	 * so that we can call ->fsync.
1222 	 */
1223 	if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1224 		retval = 0;
1225 		if (iocb->ki_flags & IOCB_DSYNC)
1226 			retval = dio_set_defer_completion(dio);
1227 		else if (!dio->inode->i_sb->s_dio_done_wq) {
1228 			/*
1229 			 * In case of AIO write racing with buffered read we
1230 			 * need to defer completion. We can't decide this now,
1231 			 * however the workqueue needs to be initialized here.
1232 			 */
1233 			retval = sb_init_dio_done_wq(dio->inode->i_sb);
1234 		}
1235 		if (retval)
1236 			goto fail_dio;
1237 	}
1238 
1239 	/*
1240 	 * Will be decremented at I/O completion time.
1241 	 */
1242 	inode_dio_begin(inode);
1243 
1244 	retval = 0;
1245 	sdio.blkbits = blkbits;
1246 	sdio.blkfactor = i_blkbits - blkbits;
1247 	sdio.block_in_file = offset >> blkbits;
1248 
1249 	sdio.get_block = get_block;
1250 	dio->end_io = end_io;
1251 	sdio.submit_io = submit_io;
1252 	sdio.final_block_in_bio = -1;
1253 	sdio.next_block_for_io = -1;
1254 
1255 	dio->iocb = iocb;
1256 
1257 	spin_lock_init(&dio->bio_lock);
1258 	dio->refcount = 1;
1259 
1260 	dio->should_dirty = iter_is_iovec(iter) && iov_iter_rw(iter) == READ;
1261 	sdio.iter = iter;
1262 	sdio.final_block_in_request = end >> blkbits;
1263 
1264 	/*
1265 	 * In case of non-aligned buffers, we may need 2 more
1266 	 * pages since we need to zero out first and last block.
1267 	 */
1268 	if (unlikely(sdio.blkfactor))
1269 		sdio.pages_in_io = 2;
1270 
1271 	sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1272 
1273 	blk_start_plug(&plug);
1274 
1275 	retval = do_direct_IO(dio, &sdio, &map_bh);
1276 	if (retval)
1277 		dio_cleanup(dio, &sdio);
1278 
1279 	if (retval == -ENOTBLK) {
1280 		/*
1281 		 * The remaining part of the request will be
1282 		 * handled by buffered I/O when we return
1283 		 */
1284 		retval = 0;
1285 	}
1286 	/*
1287 	 * There may be some unwritten disk at the end of a part-written
1288 	 * fs-block-sized block.  Go zero that now.
1289 	 */
1290 	dio_zero_block(dio, &sdio, 1, &map_bh);
1291 
1292 	if (sdio.cur_page) {
1293 		ssize_t ret2;
1294 
1295 		ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1296 		if (retval == 0)
1297 			retval = ret2;
1298 		put_page(sdio.cur_page);
1299 		sdio.cur_page = NULL;
1300 	}
1301 	if (sdio.bio)
1302 		dio_bio_submit(dio, &sdio);
1303 
1304 	blk_finish_plug(&plug);
1305 
1306 	/*
1307 	 * It is possible that, we return short IO due to end of file.
1308 	 * In that case, we need to release all the pages we got hold on.
1309 	 */
1310 	dio_cleanup(dio, &sdio);
1311 
1312 	/*
1313 	 * All block lookups have been performed. For READ requests
1314 	 * we can let i_mutex go now that its achieved its purpose
1315 	 * of protecting us from looking up uninitialized blocks.
1316 	 */
1317 	if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1318 		inode_unlock(dio->inode);
1319 
1320 	/*
1321 	 * The only time we want to leave bios in flight is when a successful
1322 	 * partial aio read or full aio write have been setup.  In that case
1323 	 * bio completion will call aio_complete.  The only time it's safe to
1324 	 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1325 	 * This had *better* be the only place that raises -EIOCBQUEUED.
1326 	 */
1327 	BUG_ON(retval == -EIOCBQUEUED);
1328 	if (dio->is_async && retval == 0 && dio->result &&
1329 	    (iov_iter_rw(iter) == READ || dio->result == count))
1330 		retval = -EIOCBQUEUED;
1331 	else
1332 		dio_await_completion(dio);
1333 
1334 	if (drop_refcount(dio) == 0) {
1335 		retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
1336 	} else
1337 		BUG_ON(retval != -EIOCBQUEUED);
1338 
1339 	return retval;
1340 
1341 fail_dio:
1342 	if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ)
1343 		inode_unlock(inode);
1344 
1345 	kmem_cache_free(dio_cache, dio);
1346 	return retval;
1347 }
1348 
1349 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1350 			     struct block_device *bdev, struct iov_iter *iter,
1351 			     get_block_t get_block,
1352 			     dio_iodone_t end_io, dio_submit_t submit_io,
1353 			     int flags)
1354 {
1355 	/*
1356 	 * The block device state is needed in the end to finally
1357 	 * submit everything.  Since it's likely to be cache cold
1358 	 * prefetch it here as first thing to hide some of the
1359 	 * latency.
1360 	 *
1361 	 * Attempt to prefetch the pieces we likely need later.
1362 	 */
1363 	prefetch(&bdev->bd_disk->part_tbl);
1364 	prefetch(bdev->bd_disk->queue);
1365 	prefetch((char *)bdev->bd_disk->queue + SMP_CACHE_BYTES);
1366 
1367 	return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
1368 				     end_io, submit_io, flags);
1369 }
1370 
1371 EXPORT_SYMBOL(__blockdev_direct_IO);
1372 
1373 static __init int dio_init(void)
1374 {
1375 	dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1376 	return 0;
1377 }
1378 module_init(dio_init)
1379