xref: /openbmc/linux/fs/direct-io.c (revision e868d61272caa648214046a096e5a6bfc068dc8c)
1 /*
2  * fs/direct-io.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  *
6  * O_DIRECT
7  *
8  * 04Jul2002	akpm@zip.com.au
9  *		Initial version
10  * 11Sep2002	janetinc@us.ibm.com
11  * 		added readv/writev support.
12  * 29Oct2002	akpm@zip.com.au
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 <asm/atomic.h>
39 
40 /*
41  * How many user pages to map in one call to get_user_pages().  This determines
42  * the size of a structure on the stack.
43  */
44 #define DIO_PAGES	64
45 
46 /*
47  * This code generally works in units of "dio_blocks".  A dio_block is
48  * somewhere between the hard sector size and the filesystem block size.  it
49  * is determined on a per-invocation basis.   When talking to the filesystem
50  * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
51  * down by dio->blkfactor.  Similarly, fs-blocksize quantities are converted
52  * to bio_block quantities by shifting left by blkfactor.
53  *
54  * If blkfactor is zero then the user's request was aligned to the filesystem's
55  * blocksize.
56  *
57  * lock_type is DIO_LOCKING for regular files on direct-IO-naive filesystems.
58  * This determines whether we need to do the fancy locking which prevents
59  * direct-IO from being able to read uninitialised disk blocks.  If its zero
60  * (blockdev) this locking is not done, and if it is DIO_OWN_LOCKING i_mutex is
61  * not held for the entire direct write (taken briefly, initially, during a
62  * direct read though, but its never held for the duration of a direct-IO).
63  */
64 
65 struct dio {
66 	/* BIO submission state */
67 	struct bio *bio;		/* bio under assembly */
68 	struct inode *inode;
69 	int rw;
70 	loff_t i_size;			/* i_size when submitted */
71 	int lock_type;			/* doesn't change */
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 	size_t	size;			/* total request size (doesn't change)*/
83 	sector_t block_in_file;		/* Current offset into the underlying
84 					   file in dio_block units. */
85 	unsigned blocks_available;	/* At block_in_file.  changes */
86 	sector_t final_block_in_request;/* doesn't change */
87 	unsigned first_block_in_page;	/* doesn't change, Used only once */
88 	int boundary;			/* prev block is at a boundary */
89 	int reap_counter;		/* rate limit reaping */
90 	get_block_t *get_block;		/* block mapping function */
91 	dio_iodone_t *end_io;		/* IO completion function */
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 	struct buffer_head map_bh;	/* last get_block() result */
96 
97 	/*
98 	 * Deferred addition of a page to the dio.  These variables are
99 	 * private to dio_send_cur_page(), submit_page_section() and
100 	 * dio_bio_add_page().
101 	 */
102 	struct page *cur_page;		/* The page */
103 	unsigned cur_page_offset;	/* Offset into it, in bytes */
104 	unsigned cur_page_len;		/* Nr of bytes at cur_page_offset */
105 	sector_t cur_page_block;	/* Where it starts */
106 
107 	/*
108 	 * Page fetching state. These variables belong to dio_refill_pages().
109 	 */
110 	int curr_page;			/* changes */
111 	int total_pages;		/* doesn't change */
112 	unsigned long curr_user_address;/* changes */
113 
114 	/*
115 	 * Page queue.  These variables belong to dio_refill_pages() and
116 	 * dio_get_page().
117 	 */
118 	struct page *pages[DIO_PAGES];	/* page buffer */
119 	unsigned head;			/* next page to process */
120 	unsigned tail;			/* last valid page + 1 */
121 	int page_errors;		/* errno from get_user_pages() */
122 
123 	/* BIO completion state */
124 	spinlock_t bio_lock;		/* protects BIO fields below */
125 	unsigned long refcount;		/* direct_io_worker() and bios */
126 	struct bio *bio_list;		/* singly linked via bi_private */
127 	struct task_struct *waiter;	/* waiting task (NULL if none) */
128 
129 	/* AIO related stuff */
130 	struct kiocb *iocb;		/* kiocb */
131 	int is_async;			/* is IO async ? */
132 	int io_error;			/* IO error in completion path */
133 	ssize_t result;                 /* IO result */
134 };
135 
136 /*
137  * How many pages are in the queue?
138  */
139 static inline unsigned dio_pages_present(struct dio *dio)
140 {
141 	return dio->tail - dio->head;
142 }
143 
144 /*
145  * Go grab and pin some userspace pages.   Typically we'll get 64 at a time.
146  */
147 static int dio_refill_pages(struct dio *dio)
148 {
149 	int ret;
150 	int nr_pages;
151 
152 	nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES);
153 	down_read(&current->mm->mmap_sem);
154 	ret = get_user_pages(
155 		current,			/* Task for fault acounting */
156 		current->mm,			/* whose pages? */
157 		dio->curr_user_address,		/* Where from? */
158 		nr_pages,			/* How many pages? */
159 		dio->rw == READ,		/* Write to memory? */
160 		0,				/* force (?) */
161 		&dio->pages[0],
162 		NULL);				/* vmas */
163 	up_read(&current->mm->mmap_sem);
164 
165 	if (ret < 0 && dio->blocks_available && (dio->rw & WRITE)) {
166 		struct page *page = ZERO_PAGE(dio->curr_user_address);
167 		/*
168 		 * A memory fault, but the filesystem has some outstanding
169 		 * mapped blocks.  We need to use those blocks up to avoid
170 		 * leaking stale data in the file.
171 		 */
172 		if (dio->page_errors == 0)
173 			dio->page_errors = ret;
174 		page_cache_get(page);
175 		dio->pages[0] = page;
176 		dio->head = 0;
177 		dio->tail = 1;
178 		ret = 0;
179 		goto out;
180 	}
181 
182 	if (ret >= 0) {
183 		dio->curr_user_address += ret * PAGE_SIZE;
184 		dio->curr_page += ret;
185 		dio->head = 0;
186 		dio->tail = ret;
187 		ret = 0;
188 	}
189 out:
190 	return ret;
191 }
192 
193 /*
194  * Get another userspace page.  Returns an ERR_PTR on error.  Pages are
195  * buffered inside the dio so that we can call get_user_pages() against a
196  * decent number of pages, less frequently.  To provide nicer use of the
197  * L1 cache.
198  */
199 static struct page *dio_get_page(struct dio *dio)
200 {
201 	if (dio_pages_present(dio) == 0) {
202 		int ret;
203 
204 		ret = dio_refill_pages(dio);
205 		if (ret)
206 			return ERR_PTR(ret);
207 		BUG_ON(dio_pages_present(dio) == 0);
208 	}
209 	return dio->pages[dio->head++];
210 }
211 
212 /**
213  * dio_complete() - called when all DIO BIO I/O has been completed
214  * @offset: the byte offset in the file of the completed operation
215  *
216  * This releases locks as dictated by the locking type, lets interested parties
217  * know that a DIO operation has completed, and calculates the resulting return
218  * code for the operation.
219  *
220  * It lets the filesystem know if it registered an interest earlier via
221  * get_block.  Pass the private field of the map buffer_head so that
222  * filesystems can use it to hold additional state between get_block calls and
223  * dio_complete.
224  */
225 static int dio_complete(struct dio *dio, loff_t offset, int ret)
226 {
227 	ssize_t transferred = 0;
228 
229 	/*
230 	 * AIO submission can race with bio completion to get here while
231 	 * expecting to have the last io completed by bio completion.
232 	 * In that case -EIOCBQUEUED is in fact not an error we want
233 	 * to preserve through this call.
234 	 */
235 	if (ret == -EIOCBQUEUED)
236 		ret = 0;
237 
238 	if (dio->result) {
239 		transferred = dio->result;
240 
241 		/* Check for short read case */
242 		if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
243 			transferred = dio->i_size - offset;
244 	}
245 
246 	if (dio->end_io && dio->result)
247 		dio->end_io(dio->iocb, offset, transferred,
248 			    dio->map_bh.b_private);
249 	if (dio->lock_type == DIO_LOCKING)
250 		/* lockdep: non-owner release */
251 		up_read_non_owner(&dio->inode->i_alloc_sem);
252 
253 	if (ret == 0)
254 		ret = dio->page_errors;
255 	if (ret == 0)
256 		ret = dio->io_error;
257 	if (ret == 0)
258 		ret = transferred;
259 
260 	return ret;
261 }
262 
263 static int dio_bio_complete(struct dio *dio, struct bio *bio);
264 /*
265  * Asynchronous IO callback.
266  */
267 static int dio_bio_end_aio(struct bio *bio, unsigned int bytes_done, int error)
268 {
269 	struct dio *dio = bio->bi_private;
270 	unsigned long remaining;
271 	unsigned long flags;
272 
273 	if (bio->bi_size)
274 		return 1;
275 
276 	/* cleanup the bio */
277 	dio_bio_complete(dio, bio);
278 
279 	spin_lock_irqsave(&dio->bio_lock, flags);
280 	remaining = --dio->refcount;
281 	if (remaining == 1 && dio->waiter)
282 		wake_up_process(dio->waiter);
283 	spin_unlock_irqrestore(&dio->bio_lock, flags);
284 
285 	if (remaining == 0) {
286 		int ret = dio_complete(dio, dio->iocb->ki_pos, 0);
287 		aio_complete(dio->iocb, ret, 0);
288 		kfree(dio);
289 	}
290 
291 	return 0;
292 }
293 
294 /*
295  * The BIO completion handler simply queues the BIO up for the process-context
296  * handler.
297  *
298  * During I/O bi_private points at the dio.  After I/O, bi_private is used to
299  * implement a singly-linked list of completed BIOs, at dio->bio_list.
300  */
301 static int dio_bio_end_io(struct bio *bio, unsigned int bytes_done, int error)
302 {
303 	struct dio *dio = bio->bi_private;
304 	unsigned long flags;
305 
306 	if (bio->bi_size)
307 		return 1;
308 
309 	spin_lock_irqsave(&dio->bio_lock, flags);
310 	bio->bi_private = dio->bio_list;
311 	dio->bio_list = bio;
312 	if (--dio->refcount == 1 && dio->waiter)
313 		wake_up_process(dio->waiter);
314 	spin_unlock_irqrestore(&dio->bio_lock, flags);
315 	return 0;
316 }
317 
318 static int
319 dio_bio_alloc(struct dio *dio, struct block_device *bdev,
320 		sector_t first_sector, int nr_vecs)
321 {
322 	struct bio *bio;
323 
324 	bio = bio_alloc(GFP_KERNEL, nr_vecs);
325 	if (bio == NULL)
326 		return -ENOMEM;
327 
328 	bio->bi_bdev = bdev;
329 	bio->bi_sector = first_sector;
330 	if (dio->is_async)
331 		bio->bi_end_io = dio_bio_end_aio;
332 	else
333 		bio->bi_end_io = dio_bio_end_io;
334 
335 	dio->bio = bio;
336 	return 0;
337 }
338 
339 /*
340  * In the AIO read case we speculatively dirty the pages before starting IO.
341  * During IO completion, any of these pages which happen to have been written
342  * back will be redirtied by bio_check_pages_dirty().
343  *
344  * bios hold a dio reference between submit_bio and ->end_io.
345  */
346 static void dio_bio_submit(struct dio *dio)
347 {
348 	struct bio *bio = dio->bio;
349 	unsigned long flags;
350 
351 	bio->bi_private = dio;
352 
353 	spin_lock_irqsave(&dio->bio_lock, flags);
354 	dio->refcount++;
355 	spin_unlock_irqrestore(&dio->bio_lock, flags);
356 
357 	if (dio->is_async && dio->rw == READ)
358 		bio_set_pages_dirty(bio);
359 
360 	submit_bio(dio->rw, bio);
361 
362 	dio->bio = NULL;
363 	dio->boundary = 0;
364 }
365 
366 /*
367  * Release any resources in case of a failure
368  */
369 static void dio_cleanup(struct dio *dio)
370 {
371 	while (dio_pages_present(dio))
372 		page_cache_release(dio_get_page(dio));
373 }
374 
375 /*
376  * Wait for the next BIO to complete.  Remove it and return it.  NULL is
377  * returned once all BIOs have been completed.  This must only be called once
378  * all bios have been issued so that dio->refcount can only decrease.  This
379  * requires that that the caller hold a reference on the dio.
380  */
381 static struct bio *dio_await_one(struct dio *dio)
382 {
383 	unsigned long flags;
384 	struct bio *bio = NULL;
385 
386 	spin_lock_irqsave(&dio->bio_lock, flags);
387 
388 	/*
389 	 * Wait as long as the list is empty and there are bios in flight.  bio
390 	 * completion drops the count, maybe adds to the list, and wakes while
391 	 * holding the bio_lock so we don't need set_current_state()'s barrier
392 	 * and can call it after testing our condition.
393 	 */
394 	while (dio->refcount > 1 && dio->bio_list == NULL) {
395 		__set_current_state(TASK_UNINTERRUPTIBLE);
396 		dio->waiter = current;
397 		spin_unlock_irqrestore(&dio->bio_lock, flags);
398 		io_schedule();
399 		/* wake up sets us TASK_RUNNING */
400 		spin_lock_irqsave(&dio->bio_lock, flags);
401 		dio->waiter = NULL;
402 	}
403 	if (dio->bio_list) {
404 		bio = dio->bio_list;
405 		dio->bio_list = bio->bi_private;
406 	}
407 	spin_unlock_irqrestore(&dio->bio_lock, flags);
408 	return bio;
409 }
410 
411 /*
412  * Process one completed BIO.  No locks are held.
413  */
414 static int dio_bio_complete(struct dio *dio, struct bio *bio)
415 {
416 	const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
417 	struct bio_vec *bvec = bio->bi_io_vec;
418 	int page_no;
419 
420 	if (!uptodate)
421 		dio->io_error = -EIO;
422 
423 	if (dio->is_async && dio->rw == READ) {
424 		bio_check_pages_dirty(bio);	/* transfers ownership */
425 	} else {
426 		for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
427 			struct page *page = bvec[page_no].bv_page;
428 
429 			if (dio->rw == READ && !PageCompound(page))
430 				set_page_dirty_lock(page);
431 			page_cache_release(page);
432 		}
433 		bio_put(bio);
434 	}
435 	return uptodate ? 0 : -EIO;
436 }
437 
438 /*
439  * Wait on and process all in-flight BIOs.  This must only be called once
440  * all bios have been issued so that the refcount can only decrease.
441  * This just waits for all bios to make it through dio_bio_complete.  IO
442  * errors are propagated through dio->io_error and should be propagated via
443  * dio_complete().
444  */
445 static void dio_await_completion(struct dio *dio)
446 {
447 	struct bio *bio;
448 	do {
449 		bio = dio_await_one(dio);
450 		if (bio)
451 			dio_bio_complete(dio, bio);
452 	} while (bio);
453 }
454 
455 /*
456  * A really large O_DIRECT read or write can generate a lot of BIOs.  So
457  * to keep the memory consumption sane we periodically reap any completed BIOs
458  * during the BIO generation phase.
459  *
460  * This also helps to limit the peak amount of pinned userspace memory.
461  */
462 static int dio_bio_reap(struct dio *dio)
463 {
464 	int ret = 0;
465 
466 	if (dio->reap_counter++ >= 64) {
467 		while (dio->bio_list) {
468 			unsigned long flags;
469 			struct bio *bio;
470 			int ret2;
471 
472 			spin_lock_irqsave(&dio->bio_lock, flags);
473 			bio = dio->bio_list;
474 			dio->bio_list = bio->bi_private;
475 			spin_unlock_irqrestore(&dio->bio_lock, flags);
476 			ret2 = dio_bio_complete(dio, bio);
477 			if (ret == 0)
478 				ret = ret2;
479 		}
480 		dio->reap_counter = 0;
481 	}
482 	return ret;
483 }
484 
485 /*
486  * Call into the fs to map some more disk blocks.  We record the current number
487  * of available blocks at dio->blocks_available.  These are in units of the
488  * fs blocksize, (1 << inode->i_blkbits).
489  *
490  * The fs is allowed to map lots of blocks at once.  If it wants to do that,
491  * it uses the passed inode-relative block number as the file offset, as usual.
492  *
493  * get_block() is passed the number of i_blkbits-sized blocks which direct_io
494  * has remaining to do.  The fs should not map more than this number of blocks.
495  *
496  * If the fs has mapped a lot of blocks, it should populate bh->b_size to
497  * indicate how much contiguous disk space has been made available at
498  * bh->b_blocknr.
499  *
500  * If *any* of the mapped blocks are new, then the fs must set buffer_new().
501  * This isn't very efficient...
502  *
503  * In the case of filesystem holes: the fs may return an arbitrarily-large
504  * hole by returning an appropriate value in b_size and by clearing
505  * buffer_mapped().  However the direct-io code will only process holes one
506  * block at a time - it will repeatedly call get_block() as it walks the hole.
507  */
508 static int get_more_blocks(struct dio *dio)
509 {
510 	int ret;
511 	struct buffer_head *map_bh = &dio->map_bh;
512 	sector_t fs_startblk;	/* Into file, in filesystem-sized blocks */
513 	unsigned long fs_count;	/* Number of filesystem-sized blocks */
514 	unsigned long dio_count;/* Number of dio_block-sized blocks */
515 	unsigned long blkmask;
516 	int create;
517 
518 	/*
519 	 * If there was a memory error and we've overwritten all the
520 	 * mapped blocks then we can now return that memory error
521 	 */
522 	ret = dio->page_errors;
523 	if (ret == 0) {
524 		BUG_ON(dio->block_in_file >= dio->final_block_in_request);
525 		fs_startblk = dio->block_in_file >> dio->blkfactor;
526 		dio_count = dio->final_block_in_request - dio->block_in_file;
527 		fs_count = dio_count >> dio->blkfactor;
528 		blkmask = (1 << dio->blkfactor) - 1;
529 		if (dio_count & blkmask)
530 			fs_count++;
531 
532 		map_bh->b_state = 0;
533 		map_bh->b_size = fs_count << dio->inode->i_blkbits;
534 
535 		create = dio->rw & WRITE;
536 		if (dio->lock_type == DIO_LOCKING) {
537 			if (dio->block_in_file < (i_size_read(dio->inode) >>
538 							dio->blkbits))
539 				create = 0;
540 		} else if (dio->lock_type == DIO_NO_LOCKING) {
541 			create = 0;
542 		}
543 
544 		/*
545 		 * For writes inside i_size we forbid block creations: only
546 		 * overwrites are permitted.  We fall back to buffered writes
547 		 * at a higher level for inside-i_size block-instantiating
548 		 * writes.
549 		 */
550 		ret = (*dio->get_block)(dio->inode, fs_startblk,
551 						map_bh, create);
552 	}
553 	return ret;
554 }
555 
556 /*
557  * There is no bio.  Make one now.
558  */
559 static int dio_new_bio(struct dio *dio, sector_t start_sector)
560 {
561 	sector_t sector;
562 	int ret, nr_pages;
563 
564 	ret = dio_bio_reap(dio);
565 	if (ret)
566 		goto out;
567 	sector = start_sector << (dio->blkbits - 9);
568 	nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
569 	BUG_ON(nr_pages <= 0);
570 	ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
571 	dio->boundary = 0;
572 out:
573 	return ret;
574 }
575 
576 /*
577  * Attempt to put the current chunk of 'cur_page' into the current BIO.  If
578  * that was successful then update final_block_in_bio and take a ref against
579  * the just-added page.
580  *
581  * Return zero on success.  Non-zero means the caller needs to start a new BIO.
582  */
583 static int dio_bio_add_page(struct dio *dio)
584 {
585 	int ret;
586 
587 	ret = bio_add_page(dio->bio, dio->cur_page,
588 			dio->cur_page_len, dio->cur_page_offset);
589 	if (ret == dio->cur_page_len) {
590 		/*
591 		 * Decrement count only, if we are done with this page
592 		 */
593 		if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE)
594 			dio->pages_in_io--;
595 		page_cache_get(dio->cur_page);
596 		dio->final_block_in_bio = dio->cur_page_block +
597 			(dio->cur_page_len >> dio->blkbits);
598 		ret = 0;
599 	} else {
600 		ret = 1;
601 	}
602 	return ret;
603 }
604 
605 /*
606  * Put cur_page under IO.  The section of cur_page which is described by
607  * cur_page_offset,cur_page_len is put into a BIO.  The section of cur_page
608  * starts on-disk at cur_page_block.
609  *
610  * We take a ref against the page here (on behalf of its presence in the bio).
611  *
612  * The caller of this function is responsible for removing cur_page from the
613  * dio, and for dropping the refcount which came from that presence.
614  */
615 static int dio_send_cur_page(struct dio *dio)
616 {
617 	int ret = 0;
618 
619 	if (dio->bio) {
620 		/*
621 		 * See whether this new request is contiguous with the old
622 		 */
623 		if (dio->final_block_in_bio != dio->cur_page_block)
624 			dio_bio_submit(dio);
625 		/*
626 		 * Submit now if the underlying fs is about to perform a
627 		 * metadata read
628 		 */
629 		if (dio->boundary)
630 			dio_bio_submit(dio);
631 	}
632 
633 	if (dio->bio == NULL) {
634 		ret = dio_new_bio(dio, dio->cur_page_block);
635 		if (ret)
636 			goto out;
637 	}
638 
639 	if (dio_bio_add_page(dio) != 0) {
640 		dio_bio_submit(dio);
641 		ret = dio_new_bio(dio, dio->cur_page_block);
642 		if (ret == 0) {
643 			ret = dio_bio_add_page(dio);
644 			BUG_ON(ret != 0);
645 		}
646 	}
647 out:
648 	return ret;
649 }
650 
651 /*
652  * An autonomous function to put a chunk of a page under deferred IO.
653  *
654  * The caller doesn't actually know (or care) whether this piece of page is in
655  * a BIO, or is under IO or whatever.  We just take care of all possible
656  * situations here.  The separation between the logic of do_direct_IO() and
657  * that of submit_page_section() is important for clarity.  Please don't break.
658  *
659  * The chunk of page starts on-disk at blocknr.
660  *
661  * We perform deferred IO, by recording the last-submitted page inside our
662  * private part of the dio structure.  If possible, we just expand the IO
663  * across that page here.
664  *
665  * If that doesn't work out then we put the old page into the bio and add this
666  * page to the dio instead.
667  */
668 static int
669 submit_page_section(struct dio *dio, struct page *page,
670 		unsigned offset, unsigned len, sector_t blocknr)
671 {
672 	int ret = 0;
673 
674 	if (dio->rw & WRITE) {
675 		/*
676 		 * Read accounting is performed in submit_bio()
677 		 */
678 		task_io_account_write(len);
679 	}
680 
681 	/*
682 	 * Can we just grow the current page's presence in the dio?
683 	 */
684 	if (	(dio->cur_page == page) &&
685 		(dio->cur_page_offset + dio->cur_page_len == offset) &&
686 		(dio->cur_page_block +
687 			(dio->cur_page_len >> dio->blkbits) == blocknr)) {
688 		dio->cur_page_len += len;
689 
690 		/*
691 		 * If dio->boundary then we want to schedule the IO now to
692 		 * avoid metadata seeks.
693 		 */
694 		if (dio->boundary) {
695 			ret = dio_send_cur_page(dio);
696 			page_cache_release(dio->cur_page);
697 			dio->cur_page = NULL;
698 		}
699 		goto out;
700 	}
701 
702 	/*
703 	 * If there's a deferred page already there then send it.
704 	 */
705 	if (dio->cur_page) {
706 		ret = dio_send_cur_page(dio);
707 		page_cache_release(dio->cur_page);
708 		dio->cur_page = NULL;
709 		if (ret)
710 			goto out;
711 	}
712 
713 	page_cache_get(page);		/* It is in dio */
714 	dio->cur_page = page;
715 	dio->cur_page_offset = offset;
716 	dio->cur_page_len = len;
717 	dio->cur_page_block = blocknr;
718 out:
719 	return ret;
720 }
721 
722 /*
723  * Clean any dirty buffers in the blockdev mapping which alias newly-created
724  * file blocks.  Only called for S_ISREG files - blockdevs do not set
725  * buffer_new
726  */
727 static void clean_blockdev_aliases(struct dio *dio)
728 {
729 	unsigned i;
730 	unsigned nblocks;
731 
732 	nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
733 
734 	for (i = 0; i < nblocks; i++) {
735 		unmap_underlying_metadata(dio->map_bh.b_bdev,
736 					dio->map_bh.b_blocknr + i);
737 	}
738 }
739 
740 /*
741  * If we are not writing the entire block and get_block() allocated
742  * the block for us, we need to fill-in the unused portion of the
743  * block with zeros. This happens only if user-buffer, fileoffset or
744  * io length is not filesystem block-size multiple.
745  *
746  * `end' is zero if we're doing the start of the IO, 1 at the end of the
747  * IO.
748  */
749 static void dio_zero_block(struct dio *dio, int end)
750 {
751 	unsigned dio_blocks_per_fs_block;
752 	unsigned this_chunk_blocks;	/* In dio_blocks */
753 	unsigned this_chunk_bytes;
754 	struct page *page;
755 
756 	dio->start_zero_done = 1;
757 	if (!dio->blkfactor || !buffer_new(&dio->map_bh))
758 		return;
759 
760 	dio_blocks_per_fs_block = 1 << dio->blkfactor;
761 	this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
762 
763 	if (!this_chunk_blocks)
764 		return;
765 
766 	/*
767 	 * We need to zero out part of an fs block.  It is either at the
768 	 * beginning or the end of the fs block.
769 	 */
770 	if (end)
771 		this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
772 
773 	this_chunk_bytes = this_chunk_blocks << dio->blkbits;
774 
775 	page = ZERO_PAGE(dio->curr_user_address);
776 	if (submit_page_section(dio, page, 0, this_chunk_bytes,
777 				dio->next_block_for_io))
778 		return;
779 
780 	dio->next_block_for_io += this_chunk_blocks;
781 }
782 
783 /*
784  * Walk the user pages, and the file, mapping blocks to disk and generating
785  * a sequence of (page,offset,len,block) mappings.  These mappings are injected
786  * into submit_page_section(), which takes care of the next stage of submission
787  *
788  * Direct IO against a blockdev is different from a file.  Because we can
789  * happily perform page-sized but 512-byte aligned IOs.  It is important that
790  * blockdev IO be able to have fine alignment and large sizes.
791  *
792  * So what we do is to permit the ->get_block function to populate bh.b_size
793  * with the size of IO which is permitted at this offset and this i_blkbits.
794  *
795  * For best results, the blockdev should be set up with 512-byte i_blkbits and
796  * it should set b_size to PAGE_SIZE or more inside get_block().  This gives
797  * fine alignment but still allows this function to work in PAGE_SIZE units.
798  */
799 static int do_direct_IO(struct dio *dio)
800 {
801 	const unsigned blkbits = dio->blkbits;
802 	const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
803 	struct page *page;
804 	unsigned block_in_page;
805 	struct buffer_head *map_bh = &dio->map_bh;
806 	int ret = 0;
807 
808 	/* The I/O can start at any block offset within the first page */
809 	block_in_page = dio->first_block_in_page;
810 
811 	while (dio->block_in_file < dio->final_block_in_request) {
812 		page = dio_get_page(dio);
813 		if (IS_ERR(page)) {
814 			ret = PTR_ERR(page);
815 			goto out;
816 		}
817 
818 		while (block_in_page < blocks_per_page) {
819 			unsigned offset_in_page = block_in_page << blkbits;
820 			unsigned this_chunk_bytes;	/* # of bytes mapped */
821 			unsigned this_chunk_blocks;	/* # of blocks */
822 			unsigned u;
823 
824 			if (dio->blocks_available == 0) {
825 				/*
826 				 * Need to go and map some more disk
827 				 */
828 				unsigned long blkmask;
829 				unsigned long dio_remainder;
830 
831 				ret = get_more_blocks(dio);
832 				if (ret) {
833 					page_cache_release(page);
834 					goto out;
835 				}
836 				if (!buffer_mapped(map_bh))
837 					goto do_holes;
838 
839 				dio->blocks_available =
840 						map_bh->b_size >> dio->blkbits;
841 				dio->next_block_for_io =
842 					map_bh->b_blocknr << dio->blkfactor;
843 				if (buffer_new(map_bh))
844 					clean_blockdev_aliases(dio);
845 
846 				if (!dio->blkfactor)
847 					goto do_holes;
848 
849 				blkmask = (1 << dio->blkfactor) - 1;
850 				dio_remainder = (dio->block_in_file & blkmask);
851 
852 				/*
853 				 * If we are at the start of IO and that IO
854 				 * starts partway into a fs-block,
855 				 * dio_remainder will be non-zero.  If the IO
856 				 * is a read then we can simply advance the IO
857 				 * cursor to the first block which is to be
858 				 * read.  But if the IO is a write and the
859 				 * block was newly allocated we cannot do that;
860 				 * the start of the fs block must be zeroed out
861 				 * on-disk
862 				 */
863 				if (!buffer_new(map_bh))
864 					dio->next_block_for_io += dio_remainder;
865 				dio->blocks_available -= dio_remainder;
866 			}
867 do_holes:
868 			/* Handle holes */
869 			if (!buffer_mapped(map_bh)) {
870 				loff_t i_size_aligned;
871 
872 				/* AKPM: eargh, -ENOTBLK is a hack */
873 				if (dio->rw & WRITE) {
874 					page_cache_release(page);
875 					return -ENOTBLK;
876 				}
877 
878 				/*
879 				 * Be sure to account for a partial block as the
880 				 * last block in the file
881 				 */
882 				i_size_aligned = ALIGN(i_size_read(dio->inode),
883 							1 << blkbits);
884 				if (dio->block_in_file >=
885 						i_size_aligned >> blkbits) {
886 					/* We hit eof */
887 					page_cache_release(page);
888 					goto out;
889 				}
890 				zero_user_page(page, block_in_page << blkbits,
891 						1 << blkbits, KM_USER0);
892 				dio->block_in_file++;
893 				block_in_page++;
894 				goto next_block;
895 			}
896 
897 			/*
898 			 * If we're performing IO which has an alignment which
899 			 * is finer than the underlying fs, go check to see if
900 			 * we must zero out the start of this block.
901 			 */
902 			if (unlikely(dio->blkfactor && !dio->start_zero_done))
903 				dio_zero_block(dio, 0);
904 
905 			/*
906 			 * Work out, in this_chunk_blocks, how much disk we
907 			 * can add to this page
908 			 */
909 			this_chunk_blocks = dio->blocks_available;
910 			u = (PAGE_SIZE - offset_in_page) >> blkbits;
911 			if (this_chunk_blocks > u)
912 				this_chunk_blocks = u;
913 			u = dio->final_block_in_request - dio->block_in_file;
914 			if (this_chunk_blocks > u)
915 				this_chunk_blocks = u;
916 			this_chunk_bytes = this_chunk_blocks << blkbits;
917 			BUG_ON(this_chunk_bytes == 0);
918 
919 			dio->boundary = buffer_boundary(map_bh);
920 			ret = submit_page_section(dio, page, offset_in_page,
921 				this_chunk_bytes, dio->next_block_for_io);
922 			if (ret) {
923 				page_cache_release(page);
924 				goto out;
925 			}
926 			dio->next_block_for_io += this_chunk_blocks;
927 
928 			dio->block_in_file += this_chunk_blocks;
929 			block_in_page += this_chunk_blocks;
930 			dio->blocks_available -= this_chunk_blocks;
931 next_block:
932 			BUG_ON(dio->block_in_file > dio->final_block_in_request);
933 			if (dio->block_in_file == dio->final_block_in_request)
934 				break;
935 		}
936 
937 		/* Drop the ref which was taken in get_user_pages() */
938 		page_cache_release(page);
939 		block_in_page = 0;
940 	}
941 out:
942 	return ret;
943 }
944 
945 /*
946  * Releases both i_mutex and i_alloc_sem
947  */
948 static ssize_t
949 direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
950 	const struct iovec *iov, loff_t offset, unsigned long nr_segs,
951 	unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
952 	struct dio *dio)
953 {
954 	unsigned long user_addr;
955 	unsigned long flags;
956 	int seg;
957 	ssize_t ret = 0;
958 	ssize_t ret2;
959 	size_t bytes;
960 
961 	dio->bio = NULL;
962 	dio->inode = inode;
963 	dio->rw = rw;
964 	dio->blkbits = blkbits;
965 	dio->blkfactor = inode->i_blkbits - blkbits;
966 	dio->start_zero_done = 0;
967 	dio->size = 0;
968 	dio->block_in_file = offset >> blkbits;
969 	dio->blocks_available = 0;
970 	dio->cur_page = NULL;
971 
972 	dio->boundary = 0;
973 	dio->reap_counter = 0;
974 	dio->get_block = get_block;
975 	dio->end_io = end_io;
976 	dio->map_bh.b_private = NULL;
977 	dio->final_block_in_bio = -1;
978 	dio->next_block_for_io = -1;
979 
980 	dio->page_errors = 0;
981 	dio->io_error = 0;
982 	dio->result = 0;
983 	dio->iocb = iocb;
984 	dio->i_size = i_size_read(inode);
985 
986 	spin_lock_init(&dio->bio_lock);
987 	dio->refcount = 1;
988 	dio->bio_list = NULL;
989 	dio->waiter = NULL;
990 
991 	/*
992 	 * In case of non-aligned buffers, we may need 2 more
993 	 * pages since we need to zero out first and last block.
994 	 */
995 	if (unlikely(dio->blkfactor))
996 		dio->pages_in_io = 2;
997 	else
998 		dio->pages_in_io = 0;
999 
1000 	for (seg = 0; seg < nr_segs; seg++) {
1001 		user_addr = (unsigned long)iov[seg].iov_base;
1002 		dio->pages_in_io +=
1003 			((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
1004 				- user_addr/PAGE_SIZE);
1005 	}
1006 
1007 	for (seg = 0; seg < nr_segs; seg++) {
1008 		user_addr = (unsigned long)iov[seg].iov_base;
1009 		dio->size += bytes = iov[seg].iov_len;
1010 
1011 		/* Index into the first page of the first block */
1012 		dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1013 		dio->final_block_in_request = dio->block_in_file +
1014 						(bytes >> blkbits);
1015 		/* Page fetching state */
1016 		dio->head = 0;
1017 		dio->tail = 0;
1018 		dio->curr_page = 0;
1019 
1020 		dio->total_pages = 0;
1021 		if (user_addr & (PAGE_SIZE-1)) {
1022 			dio->total_pages++;
1023 			bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1024 		}
1025 		dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1026 		dio->curr_user_address = user_addr;
1027 
1028 		ret = do_direct_IO(dio);
1029 
1030 		dio->result += iov[seg].iov_len -
1031 			((dio->final_block_in_request - dio->block_in_file) <<
1032 					blkbits);
1033 
1034 		if (ret) {
1035 			dio_cleanup(dio);
1036 			break;
1037 		}
1038 	} /* end iovec loop */
1039 
1040 	if (ret == -ENOTBLK && (rw & WRITE)) {
1041 		/*
1042 		 * The remaining part of the request will be
1043 		 * be handled by buffered I/O when we return
1044 		 */
1045 		ret = 0;
1046 	}
1047 	/*
1048 	 * There may be some unwritten disk at the end of a part-written
1049 	 * fs-block-sized block.  Go zero that now.
1050 	 */
1051 	dio_zero_block(dio, 1);
1052 
1053 	if (dio->cur_page) {
1054 		ret2 = dio_send_cur_page(dio);
1055 		if (ret == 0)
1056 			ret = ret2;
1057 		page_cache_release(dio->cur_page);
1058 		dio->cur_page = NULL;
1059 	}
1060 	if (dio->bio)
1061 		dio_bio_submit(dio);
1062 
1063 	/* All IO is now issued, send it on its way */
1064 	blk_run_address_space(inode->i_mapping);
1065 
1066 	/*
1067 	 * It is possible that, we return short IO due to end of file.
1068 	 * In that case, we need to release all the pages we got hold on.
1069 	 */
1070 	dio_cleanup(dio);
1071 
1072 	/*
1073 	 * All block lookups have been performed. For READ requests
1074 	 * we can let i_mutex go now that its achieved its purpose
1075 	 * of protecting us from looking up uninitialized blocks.
1076 	 */
1077 	if ((rw == READ) && (dio->lock_type == DIO_LOCKING))
1078 		mutex_unlock(&dio->inode->i_mutex);
1079 
1080 	/*
1081 	 * The only time we want to leave bios in flight is when a successful
1082 	 * partial aio read or full aio write have been setup.  In that case
1083 	 * bio completion will call aio_complete.  The only time it's safe to
1084 	 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1085 	 * This had *better* be the only place that raises -EIOCBQUEUED.
1086 	 */
1087 	BUG_ON(ret == -EIOCBQUEUED);
1088 	if (dio->is_async && ret == 0 && dio->result &&
1089 	    ((rw & READ) || (dio->result == dio->size)))
1090 		ret = -EIOCBQUEUED;
1091 
1092 	if (ret != -EIOCBQUEUED)
1093 		dio_await_completion(dio);
1094 
1095 	/*
1096 	 * Sync will always be dropping the final ref and completing the
1097 	 * operation.  AIO can if it was a broken operation described above or
1098 	 * in fact if all the bios race to complete before we get here.  In
1099 	 * that case dio_complete() translates the EIOCBQUEUED into the proper
1100 	 * return code that the caller will hand to aio_complete().
1101 	 *
1102 	 * This is managed by the bio_lock instead of being an atomic_t so that
1103 	 * completion paths can drop their ref and use the remaining count to
1104 	 * decide to wake the submission path atomically.
1105 	 */
1106 	spin_lock_irqsave(&dio->bio_lock, flags);
1107 	ret2 = --dio->refcount;
1108 	spin_unlock_irqrestore(&dio->bio_lock, flags);
1109 	BUG_ON(!dio->is_async && ret2 != 0);
1110 	if (ret2 == 0) {
1111 		ret = dio_complete(dio, offset, ret);
1112 		kfree(dio);
1113 	} else
1114 		BUG_ON(ret != -EIOCBQUEUED);
1115 
1116 	return ret;
1117 }
1118 
1119 /*
1120  * This is a library function for use by filesystem drivers.
1121  * The locking rules are governed by the dio_lock_type parameter.
1122  *
1123  * DIO_NO_LOCKING (no locking, for raw block device access)
1124  * For writes, i_mutex is not held on entry; it is never taken.
1125  *
1126  * DIO_LOCKING (simple locking for regular files)
1127  * For writes we are called under i_mutex and return with i_mutex held, even
1128  * though it is internally dropped.
1129  * For reads, i_mutex is not held on entry, but it is taken and dropped before
1130  * returning.
1131  *
1132  * DIO_OWN_LOCKING (filesystem provides synchronisation and handling of
1133  *	uninitialised data, allowing parallel direct readers and writers)
1134  * For writes we are called without i_mutex, return without it, never touch it.
1135  * For reads we are called under i_mutex and return with i_mutex held, even
1136  * though it may be internally dropped.
1137  *
1138  * Additional i_alloc_sem locking requirements described inline below.
1139  */
1140 ssize_t
1141 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1142 	struct block_device *bdev, const struct iovec *iov, loff_t offset,
1143 	unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1144 	int dio_lock_type)
1145 {
1146 	int seg;
1147 	size_t size;
1148 	unsigned long addr;
1149 	unsigned blkbits = inode->i_blkbits;
1150 	unsigned bdev_blkbits = 0;
1151 	unsigned blocksize_mask = (1 << blkbits) - 1;
1152 	ssize_t retval = -EINVAL;
1153 	loff_t end = offset;
1154 	struct dio *dio;
1155 	int release_i_mutex = 0;
1156 	int acquire_i_mutex = 0;
1157 
1158 	if (rw & WRITE)
1159 		rw = WRITE_SYNC;
1160 
1161 	if (bdev)
1162 		bdev_blkbits = blksize_bits(bdev_hardsect_size(bdev));
1163 
1164 	if (offset & blocksize_mask) {
1165 		if (bdev)
1166 			 blkbits = bdev_blkbits;
1167 		blocksize_mask = (1 << blkbits) - 1;
1168 		if (offset & blocksize_mask)
1169 			goto out;
1170 	}
1171 
1172 	/* Check the memory alignment.  Blocks cannot straddle pages */
1173 	for (seg = 0; seg < nr_segs; seg++) {
1174 		addr = (unsigned long)iov[seg].iov_base;
1175 		size = iov[seg].iov_len;
1176 		end += size;
1177 		if ((addr & blocksize_mask) || (size & blocksize_mask))  {
1178 			if (bdev)
1179 				 blkbits = bdev_blkbits;
1180 			blocksize_mask = (1 << blkbits) - 1;
1181 			if ((addr & blocksize_mask) || (size & blocksize_mask))
1182 				goto out;
1183 		}
1184 	}
1185 
1186 	dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1187 	retval = -ENOMEM;
1188 	if (!dio)
1189 		goto out;
1190 
1191 	/*
1192 	 * For block device access DIO_NO_LOCKING is used,
1193 	 *	neither readers nor writers do any locking at all
1194 	 * For regular files using DIO_LOCKING,
1195 	 *	readers need to grab i_mutex and i_alloc_sem
1196 	 *	writers need to grab i_alloc_sem only (i_mutex is already held)
1197 	 * For regular files using DIO_OWN_LOCKING,
1198 	 *	neither readers nor writers take any locks here
1199 	 */
1200 	dio->lock_type = dio_lock_type;
1201 	if (dio_lock_type != DIO_NO_LOCKING) {
1202 		/* watch out for a 0 len io from a tricksy fs */
1203 		if (rw == READ && end > offset) {
1204 			struct address_space *mapping;
1205 
1206 			mapping = iocb->ki_filp->f_mapping;
1207 			if (dio_lock_type != DIO_OWN_LOCKING) {
1208 				mutex_lock(&inode->i_mutex);
1209 				release_i_mutex = 1;
1210 			}
1211 
1212 			retval = filemap_write_and_wait_range(mapping, offset,
1213 							      end - 1);
1214 			if (retval) {
1215 				kfree(dio);
1216 				goto out;
1217 			}
1218 
1219 			if (dio_lock_type == DIO_OWN_LOCKING) {
1220 				mutex_unlock(&inode->i_mutex);
1221 				acquire_i_mutex = 1;
1222 			}
1223 		}
1224 
1225 		if (dio_lock_type == DIO_LOCKING)
1226 			/* lockdep: not the owner will release it */
1227 			down_read_non_owner(&inode->i_alloc_sem);
1228 	}
1229 
1230 	/*
1231 	 * For file extending writes updating i_size before data
1232 	 * writeouts complete can expose uninitialized blocks. So
1233 	 * even for AIO, we need to wait for i/o to complete before
1234 	 * returning in this case.
1235 	 */
1236 	dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
1237 		(end > i_size_read(inode)));
1238 
1239 	retval = direct_io_worker(rw, iocb, inode, iov, offset,
1240 				nr_segs, blkbits, get_block, end_io, dio);
1241 
1242 	if (rw == READ && dio_lock_type == DIO_LOCKING)
1243 		release_i_mutex = 0;
1244 
1245 out:
1246 	if (release_i_mutex)
1247 		mutex_unlock(&inode->i_mutex);
1248 	else if (acquire_i_mutex)
1249 		mutex_lock(&inode->i_mutex);
1250 	return retval;
1251 }
1252 EXPORT_SYMBOL(__blockdev_direct_IO);
1253