direct-io.c 37.4 KB
Newer Older
L
Linus Torvalds 已提交
1 2 3 4 5 6 7
/*
 * fs/direct-io.c
 *
 * Copyright (C) 2002, Linus Torvalds.
 *
 * O_DIRECT
 *
8
 * 04Jul2002	Andrew Morton
L
Linus Torvalds 已提交
9 10 11
 *		Initial version
 * 11Sep2002	janetinc@us.ibm.com
 * 		added readv/writev support.
12
 * 29Oct2002	Andrew Morton
L
Linus Torvalds 已提交
13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
 *		rewrote bio_add_page() support.
 * 30Oct2002	pbadari@us.ibm.com
 *		added support for non-aligned IO.
 * 06Nov2002	pbadari@us.ibm.com
 *		added asynchronous IO support.
 * 21Jul2003	nathans@sgi.com
 *		added IO completion notifier.
 */

#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/types.h>
#include <linux/fs.h>
#include <linux/mm.h>
#include <linux/slab.h>
#include <linux/highmem.h>
#include <linux/pagemap.h>
A
Andrew Morton 已提交
30
#include <linux/task_io_accounting_ops.h>
L
Linus Torvalds 已提交
31 32 33 34 35 36 37
#include <linux/bio.h>
#include <linux/wait.h>
#include <linux/err.h>
#include <linux/blkdev.h>
#include <linux/buffer_head.h>
#include <linux/rwsem.h>
#include <linux/uio.h>
A
Arun Sharma 已提交
38
#include <linux/atomic.h>
39
#include <linux/prefetch.h>
L
Linus Torvalds 已提交
40 41 42

/*
 * How many user pages to map in one call to get_user_pages().  This determines
A
Andi Kleen 已提交
43
 * the size of a structure in the slab cache
L
Linus Torvalds 已提交
44 45 46 47 48 49 50 51 52 53 54 55 56 57 58
 */
#define DIO_PAGES	64

/*
 * This code generally works in units of "dio_blocks".  A dio_block is
 * somewhere between the hard sector size and the filesystem block size.  it
 * is determined on a per-invocation basis.   When talking to the filesystem
 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
 * down by dio->blkfactor.  Similarly, fs-blocksize quantities are converted
 * to bio_block quantities by shifting left by blkfactor.
 *
 * If blkfactor is zero then the user's request was aligned to the filesystem's
 * blocksize.
 */

59 60 61
/* dio_state only used in the submission path */

struct dio_submit {
L
Linus Torvalds 已提交
62 63 64 65 66 67 68 69 70 71 72 73 74 75 76
	struct bio *bio;		/* bio under assembly */
	unsigned blkbits;		/* doesn't change */
	unsigned blkfactor;		/* When we're using an alignment which
					   is finer than the filesystem's soft
					   blocksize, this specifies how much
					   finer.  blkfactor=2 means 1/4-block
					   alignment.  Does not change */
	unsigned start_zero_done;	/* flag: sub-blocksize zeroing has
					   been performed at the start of a
					   write */
	int pages_in_io;		/* approximate total IO pages */
	size_t	size;			/* total request size (doesn't change)*/
	sector_t block_in_file;		/* Current offset into the underlying
					   file in dio_block units. */
	unsigned blocks_available;	/* At block_in_file.  changes */
77
	int reap_counter;		/* rate limit reaping */
L
Linus Torvalds 已提交
78 79 80
	sector_t final_block_in_request;/* doesn't change */
	unsigned first_block_in_page;	/* doesn't change, Used only once */
	int boundary;			/* prev block is at a boundary */
81
	get_block_t *get_block;		/* block mapping function */
82
	dio_submit_t *submit_io;	/* IO submition function */
83

84
	loff_t logical_offset_in_bio;	/* current first logical block in bio */
L
Linus Torvalds 已提交
85 86 87 88 89 90 91 92 93 94 95 96 97
	sector_t final_block_in_bio;	/* current final block in bio + 1 */
	sector_t next_block_for_io;	/* next block to be put under IO,
					   in dio_blocks units */

	/*
	 * Deferred addition of a page to the dio.  These variables are
	 * private to dio_send_cur_page(), submit_page_section() and
	 * dio_bio_add_page().
	 */
	struct page *cur_page;		/* The page */
	unsigned cur_page_offset;	/* Offset into it, in bytes */
	unsigned cur_page_len;		/* Nr of bytes at cur_page_offset */
	sector_t cur_page_block;	/* Where it starts */
98
	loff_t cur_page_fs_offset;	/* Offset in file */
L
Linus Torvalds 已提交
99 100 101 102 103 104 105 106 107 108 109 110 111 112

	/*
	 * Page fetching state. These variables belong to dio_refill_pages().
	 */
	int curr_page;			/* changes */
	int total_pages;		/* doesn't change */
	unsigned long curr_user_address;/* changes */

	/*
	 * Page queue.  These variables belong to dio_refill_pages() and
	 * dio_get_page().
	 */
	unsigned head;			/* next page to process */
	unsigned tail;			/* last valid page + 1 */
113 114 115 116 117 118
};

/* dio_state communicated between submission path and end_io */
struct dio {
	int flags;			/* doesn't change */
	int rw;
119
	struct inode *inode;
120 121 122
	loff_t i_size;			/* i_size when submitted */
	dio_iodone_t *end_io;		/* IO completion function */

A
Andi Kleen 已提交
123
	void *private;			/* copy from map_bh.b_private */
124 125 126

	/* BIO completion state */
	spinlock_t bio_lock;		/* protects BIO fields below */
127 128 129
	int page_errors;		/* errno from get_user_pages() */
	int is_async;			/* is IO async ? */
	int io_error;			/* IO error in completion path */
130 131 132 133 134 135 136 137
	unsigned long refcount;		/* direct_io_worker() and bios */
	struct bio *bio_list;		/* singly linked via bi_private */
	struct task_struct *waiter;	/* waiting task (NULL if none) */

	/* AIO related stuff */
	struct kiocb *iocb;		/* kiocb */
	ssize_t result;                 /* IO result */

138 139 140 141 142 143
	/*
	 * pages[] (and any fields placed after it) are not zeroed out at
	 * allocation time.  Don't add new fields after pages[] unless you
	 * wish that they not be zeroed.
	 */
	struct page *pages[DIO_PAGES];	/* page buffer */
144 145 146
} ____cacheline_aligned_in_smp;

static struct kmem_cache *dio_cache __read_mostly;
L
Linus Torvalds 已提交
147 148 149 150

/*
 * How many pages are in the queue?
 */
151
static inline unsigned dio_pages_present(struct dio_submit *sdio)
L
Linus Torvalds 已提交
152
{
153
	return sdio->tail - sdio->head;
L
Linus Torvalds 已提交
154 155 156 157 158
}

/*
 * Go grab and pin some userspace pages.   Typically we'll get 64 at a time.
 */
159
static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
L
Linus Torvalds 已提交
160 161 162 163
{
	int ret;
	int nr_pages;

164
	nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
N
Nick Piggin 已提交
165
	ret = get_user_pages_fast(
166
		sdio->curr_user_address,		/* Where from? */
L
Linus Torvalds 已提交
167 168
		nr_pages,			/* How many pages? */
		dio->rw == READ,		/* Write to memory? */
N
Nick Piggin 已提交
169
		&dio->pages[0]);		/* Put results here */
L
Linus Torvalds 已提交
170

171
	if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
N
Nick Piggin 已提交
172
		struct page *page = ZERO_PAGE(0);
L
Linus Torvalds 已提交
173 174 175 176 177 178 179
		/*
		 * A memory fault, but the filesystem has some outstanding
		 * mapped blocks.  We need to use those blocks up to avoid
		 * leaking stale data in the file.
		 */
		if (dio->page_errors == 0)
			dio->page_errors = ret;
N
Nick Piggin 已提交
180 181
		page_cache_get(page);
		dio->pages[0] = page;
182 183
		sdio->head = 0;
		sdio->tail = 1;
L
Linus Torvalds 已提交
184 185 186 187 188
		ret = 0;
		goto out;
	}

	if (ret >= 0) {
189 190 191 192
		sdio->curr_user_address += ret * PAGE_SIZE;
		sdio->curr_page += ret;
		sdio->head = 0;
		sdio->tail = ret;
L
Linus Torvalds 已提交
193 194 195 196 197 198 199 200 201 202 203 204
		ret = 0;
	}
out:
	return ret;	
}

/*
 * Get another userspace page.  Returns an ERR_PTR on error.  Pages are
 * buffered inside the dio so that we can call get_user_pages() against a
 * decent number of pages, less frequently.  To provide nicer use of the
 * L1 cache.
 */
205 206
static inline struct page *dio_get_page(struct dio *dio,
		struct dio_submit *sdio)
L
Linus Torvalds 已提交
207
{
208
	if (dio_pages_present(sdio) == 0) {
L
Linus Torvalds 已提交
209 210
		int ret;

211
		ret = dio_refill_pages(dio, sdio);
L
Linus Torvalds 已提交
212 213
		if (ret)
			return ERR_PTR(ret);
214
		BUG_ON(dio_pages_present(sdio) == 0);
L
Linus Torvalds 已提交
215
	}
216
	return dio->pages[sdio->head++];
L
Linus Torvalds 已提交
217 218
}

219 220 221 222 223 224 225 226 227 228 229 230
/**
 * dio_complete() - called when all DIO BIO I/O has been completed
 * @offset: the byte offset in the file of the completed operation
 *
 * This releases locks as dictated by the locking type, lets interested parties
 * know that a DIO operation has completed, and calculates the resulting return
 * code for the operation.
 *
 * It lets the filesystem know if it registered an interest earlier via
 * get_block.  Pass the private field of the map buffer_head so that
 * filesystems can use it to hold additional state between get_block calls and
 * dio_complete.
L
Linus Torvalds 已提交
231
 */
232
static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async)
L
Linus Torvalds 已提交
233
{
234 235
	ssize_t transferred = 0;

236 237 238 239 240 241 242 243 244
	/*
	 * AIO submission can race with bio completion to get here while
	 * expecting to have the last io completed by bio completion.
	 * In that case -EIOCBQUEUED is in fact not an error we want
	 * to preserve through this call.
	 */
	if (ret == -EIOCBQUEUED)
		ret = 0;

245 246 247 248 249 250 251 252 253 254 255 256 257 258 259
	if (dio->result) {
		transferred = dio->result;

		/* Check for short read case */
		if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
			transferred = dio->i_size - offset;
	}

	if (ret == 0)
		ret = dio->page_errors;
	if (ret == 0)
		ret = dio->io_error;
	if (ret == 0)
		ret = transferred;

260 261
	if (dio->end_io && dio->result) {
		dio->end_io(dio->iocb, offset, transferred,
A
Andi Kleen 已提交
262
			    dio->private, ret, is_async);
263
	} else {
264
		inode_dio_done(dio->inode);
265 266
		if (is_async)
			aio_complete(dio->iocb, ret, 0);
267 268
	}

269
	return ret;
L
Linus Torvalds 已提交
270 271 272 273 274 275
}

static int dio_bio_complete(struct dio *dio, struct bio *bio);
/*
 * Asynchronous IO callback. 
 */
276
static void dio_bio_end_aio(struct bio *bio, int error)
L
Linus Torvalds 已提交
277 278
{
	struct dio *dio = bio->bi_private;
Z
Zach Brown 已提交
279 280
	unsigned long remaining;
	unsigned long flags;
L
Linus Torvalds 已提交
281 282 283

	/* cleanup the bio */
	dio_bio_complete(dio, bio);
284

Z
Zach Brown 已提交
285 286 287
	spin_lock_irqsave(&dio->bio_lock, flags);
	remaining = --dio->refcount;
	if (remaining == 1 && dio->waiter)
288
		wake_up_process(dio->waiter);
Z
Zach Brown 已提交
289
	spin_unlock_irqrestore(&dio->bio_lock, flags);
290

291
	if (remaining == 0) {
292
		dio_complete(dio, dio->iocb->ki_pos, 0, true);
293
		kmem_cache_free(dio_cache, dio);
294
	}
L
Linus Torvalds 已提交
295 296 297 298 299 300 301 302 303
}

/*
 * The BIO completion handler simply queues the BIO up for the process-context
 * handler.
 *
 * During I/O bi_private points at the dio.  After I/O, bi_private is used to
 * implement a singly-linked list of completed BIOs, at dio->bio_list.
 */
304
static void dio_bio_end_io(struct bio *bio, int error)
L
Linus Torvalds 已提交
305 306 307 308 309 310 311
{
	struct dio *dio = bio->bi_private;
	unsigned long flags;

	spin_lock_irqsave(&dio->bio_lock, flags);
	bio->bi_private = dio->bio_list;
	dio->bio_list = bio;
Z
Zach Brown 已提交
312
	if (--dio->refcount == 1 && dio->waiter)
L
Linus Torvalds 已提交
313 314 315 316
		wake_up_process(dio->waiter);
	spin_unlock_irqrestore(&dio->bio_lock, flags);
}

317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336
/**
 * dio_end_io - handle the end io action for the given bio
 * @bio: The direct io bio thats being completed
 * @error: Error if there was one
 *
 * This is meant to be called by any filesystem that uses their own dio_submit_t
 * so that the DIO specific endio actions are dealt with after the filesystem
 * has done it's completion work.
 */
void dio_end_io(struct bio *bio, int error)
{
	struct dio *dio = bio->bi_private;

	if (dio->is_async)
		dio_bio_end_aio(bio, error);
	else
		dio_bio_end_io(bio, error);
}
EXPORT_SYMBOL_GPL(dio_end_io);

337
static inline void
338 339 340
dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
	      struct block_device *bdev,
	      sector_t first_sector, int nr_vecs)
L
Linus Torvalds 已提交
341 342 343
{
	struct bio *bio;

344 345 346 347
	/*
	 * bio_alloc() is guaranteed to return a bio when called with
	 * __GFP_WAIT and we request a valid number of vectors.
	 */
L
Linus Torvalds 已提交
348 349 350 351 352 353 354 355 356
	bio = bio_alloc(GFP_KERNEL, nr_vecs);

	bio->bi_bdev = bdev;
	bio->bi_sector = first_sector;
	if (dio->is_async)
		bio->bi_end_io = dio_bio_end_aio;
	else
		bio->bi_end_io = dio_bio_end_io;

357 358
	sdio->bio = bio;
	sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
L
Linus Torvalds 已提交
359 360 361 362 363 364
}

/*
 * In the AIO read case we speculatively dirty the pages before starting IO.
 * During IO completion, any of these pages which happen to have been written
 * back will be redirtied by bio_check_pages_dirty().
365 366
 *
 * bios hold a dio reference between submit_bio and ->end_io.
L
Linus Torvalds 已提交
367
 */
368
static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
L
Linus Torvalds 已提交
369
{
370
	struct bio *bio = sdio->bio;
Z
Zach Brown 已提交
371
	unsigned long flags;
L
Linus Torvalds 已提交
372 373

	bio->bi_private = dio;
Z
Zach Brown 已提交
374 375 376 377 378

	spin_lock_irqsave(&dio->bio_lock, flags);
	dio->refcount++;
	spin_unlock_irqrestore(&dio->bio_lock, flags);

L
Linus Torvalds 已提交
379 380
	if (dio->is_async && dio->rw == READ)
		bio_set_pages_dirty(bio);
Z
Zach Brown 已提交
381

382 383 384
	if (sdio->submit_io)
		sdio->submit_io(dio->rw, bio, dio->inode,
			       sdio->logical_offset_in_bio);
385 386
	else
		submit_bio(dio->rw, bio);
L
Linus Torvalds 已提交
387

388 389 390
	sdio->bio = NULL;
	sdio->boundary = 0;
	sdio->logical_offset_in_bio = 0;
L
Linus Torvalds 已提交
391 392 393 394 395
}

/*
 * Release any resources in case of a failure
 */
396
static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
L
Linus Torvalds 已提交
397
{
398 399
	while (dio_pages_present(sdio))
		page_cache_release(dio_get_page(dio, sdio));
L
Linus Torvalds 已提交
400 401 402
}

/*
403 404 405 406
 * Wait for the next BIO to complete.  Remove it and return it.  NULL is
 * returned once all BIOs have been completed.  This must only be called once
 * all bios have been issued so that dio->refcount can only decrease.  This
 * requires that that the caller hold a reference on the dio.
L
Linus Torvalds 已提交
407 408 409 410
 */
static struct bio *dio_await_one(struct dio *dio)
{
	unsigned long flags;
411
	struct bio *bio = NULL;
L
Linus Torvalds 已提交
412 413

	spin_lock_irqsave(&dio->bio_lock, flags);
Z
Zach Brown 已提交
414 415 416 417 418 419 420 421 422 423 424 425 426 427 428

	/*
	 * Wait as long as the list is empty and there are bios in flight.  bio
	 * completion drops the count, maybe adds to the list, and wakes while
	 * holding the bio_lock so we don't need set_current_state()'s barrier
	 * and can call it after testing our condition.
	 */
	while (dio->refcount > 1 && dio->bio_list == NULL) {
		__set_current_state(TASK_UNINTERRUPTIBLE);
		dio->waiter = current;
		spin_unlock_irqrestore(&dio->bio_lock, flags);
		io_schedule();
		/* wake up sets us TASK_RUNNING */
		spin_lock_irqsave(&dio->bio_lock, flags);
		dio->waiter = NULL;
L
Linus Torvalds 已提交
429
	}
430 431 432 433
	if (dio->bio_list) {
		bio = dio->bio_list;
		dio->bio_list = bio->bi_private;
	}
L
Linus Torvalds 已提交
434 435 436 437 438 439 440 441 442 443 444 445 446 447
	spin_unlock_irqrestore(&dio->bio_lock, flags);
	return bio;
}

/*
 * Process one completed BIO.  No locks are held.
 */
static int dio_bio_complete(struct dio *dio, struct bio *bio)
{
	const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
	struct bio_vec *bvec = bio->bi_io_vec;
	int page_no;

	if (!uptodate)
448
		dio->io_error = -EIO;
L
Linus Torvalds 已提交
449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465

	if (dio->is_async && dio->rw == READ) {
		bio_check_pages_dirty(bio);	/* transfers ownership */
	} else {
		for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
			struct page *page = bvec[page_no].bv_page;

			if (dio->rw == READ && !PageCompound(page))
				set_page_dirty_lock(page);
			page_cache_release(page);
		}
		bio_put(bio);
	}
	return uptodate ? 0 : -EIO;
}

/*
466 467 468
 * Wait on and process all in-flight BIOs.  This must only be called once
 * all bios have been issued so that the refcount can only decrease.
 * This just waits for all bios to make it through dio_bio_complete.  IO
469
 * errors are propagated through dio->io_error and should be propagated via
470
 * dio_complete().
L
Linus Torvalds 已提交
471
 */
472
static void dio_await_completion(struct dio *dio)
L
Linus Torvalds 已提交
473
{
474 475 476 477 478 479
	struct bio *bio;
	do {
		bio = dio_await_one(dio);
		if (bio)
			dio_bio_complete(dio, bio);
	} while (bio);
L
Linus Torvalds 已提交
480 481 482 483 484 485 486 487 488
}

/*
 * A really large O_DIRECT read or write can generate a lot of BIOs.  So
 * to keep the memory consumption sane we periodically reap any completed BIOs
 * during the BIO generation phase.
 *
 * This also helps to limit the peak amount of pinned userspace memory.
 */
489
static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
L
Linus Torvalds 已提交
490 491 492
{
	int ret = 0;

493
	if (sdio->reap_counter++ >= 64) {
L
Linus Torvalds 已提交
494 495 496 497 498 499 500 501 502 503 504 505 506
		while (dio->bio_list) {
			unsigned long flags;
			struct bio *bio;
			int ret2;

			spin_lock_irqsave(&dio->bio_lock, flags);
			bio = dio->bio_list;
			dio->bio_list = bio->bi_private;
			spin_unlock_irqrestore(&dio->bio_lock, flags);
			ret2 = dio_bio_complete(dio, bio);
			if (ret == 0)
				ret = ret2;
		}
507
		sdio->reap_counter = 0;
L
Linus Torvalds 已提交
508 509 510 511 512 513
	}
	return ret;
}

/*
 * Call into the fs to map some more disk blocks.  We record the current number
514
 * of available blocks at sdio->blocks_available.  These are in units of the
L
Linus Torvalds 已提交
515 516 517 518 519
 * fs blocksize, (1 << inode->i_blkbits).
 *
 * The fs is allowed to map lots of blocks at once.  If it wants to do that,
 * it uses the passed inode-relative block number as the file offset, as usual.
 *
520
 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
L
Linus Torvalds 已提交
521 522 523 524 525 526 527 528 529 530 531 532
 * has remaining to do.  The fs should not map more than this number of blocks.
 *
 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
 * indicate how much contiguous disk space has been made available at
 * bh->b_blocknr.
 *
 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
 * This isn't very efficient...
 *
 * In the case of filesystem holes: the fs may return an arbitrarily-large
 * hole by returning an appropriate value in b_size and by clearing
 * buffer_mapped().  However the direct-io code will only process holes one
533
 * block at a time - it will repeatedly call get_block() as it walks the hole.
L
Linus Torvalds 已提交
534
 */
A
Andi Kleen 已提交
535 536
static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
			   struct buffer_head *map_bh)
L
Linus Torvalds 已提交
537 538 539
{
	int ret;
	sector_t fs_startblk;	/* Into file, in filesystem-sized blocks */
540
	sector_t fs_endblk;	/* Into file, in filesystem-sized blocks */
L
Linus Torvalds 已提交
541 542
	unsigned long fs_count;	/* Number of filesystem-sized blocks */
	int create;
543
	unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
L
Linus Torvalds 已提交
544 545 546 547 548 549 550

	/*
	 * If there was a memory error and we've overwritten all the
	 * mapped blocks then we can now return that memory error
	 */
	ret = dio->page_errors;
	if (ret == 0) {
551 552
		BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
		fs_startblk = sdio->block_in_file >> sdio->blkfactor;
553 554 555
		fs_endblk = (sdio->final_block_in_request - 1) >>
					sdio->blkfactor;
		fs_count = fs_endblk - fs_startblk + 1;
L
Linus Torvalds 已提交
556

557
		map_bh->b_state = 0;
558
		map_bh->b_size = fs_count << i_blkbits;
559

560 561 562 563 564 565 566 567 568 569 570
		/*
		 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
		 * forbid block creations: only overwrites are permitted.
		 * We will return early to the caller once we see an
		 * unmapped buffer head returned, and the caller will fall
		 * back to buffered I/O.
		 *
		 * Otherwise the decision is left to the get_blocks method,
		 * which may decide to handle it or also return an unmapped
		 * buffer head.
		 */
J
Jens Axboe 已提交
571
		create = dio->rw & WRITE;
572
		if (dio->flags & DIO_SKIP_HOLES) {
573 574
			if (sdio->block_in_file < (i_size_read(dio->inode) >>
							sdio->blkbits))
L
Linus Torvalds 已提交
575 576
				create = 0;
		}
577

578
		ret = (*sdio->get_block)(dio->inode, fs_startblk,
L
Linus Torvalds 已提交
579
						map_bh, create);
A
Andi Kleen 已提交
580 581 582

		/* Store for completion */
		dio->private = map_bh->b_private;
L
Linus Torvalds 已提交
583 584 585 586 587 588 589
	}
	return ret;
}

/*
 * There is no bio.  Make one now.
 */
590 591
static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
		sector_t start_sector, struct buffer_head *map_bh)
L
Linus Torvalds 已提交
592 593 594 595
{
	sector_t sector;
	int ret, nr_pages;

596
	ret = dio_bio_reap(dio, sdio);
L
Linus Torvalds 已提交
597 598
	if (ret)
		goto out;
599
	sector = start_sector << (sdio->blkbits - 9);
A
Andi Kleen 已提交
600
	nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
601
	nr_pages = min(nr_pages, BIO_MAX_PAGES);
L
Linus Torvalds 已提交
602
	BUG_ON(nr_pages <= 0);
A
Andi Kleen 已提交
603
	dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
604
	sdio->boundary = 0;
L
Linus Torvalds 已提交
605 606 607 608 609 610 611 612 613 614 615
out:
	return ret;
}

/*
 * Attempt to put the current chunk of 'cur_page' into the current BIO.  If
 * that was successful then update final_block_in_bio and take a ref against
 * the just-added page.
 *
 * Return zero on success.  Non-zero means the caller needs to start a new BIO.
 */
616
static inline int dio_bio_add_page(struct dio_submit *sdio)
L
Linus Torvalds 已提交
617 618 619
{
	int ret;

620 621 622
	ret = bio_add_page(sdio->bio, sdio->cur_page,
			sdio->cur_page_len, sdio->cur_page_offset);
	if (ret == sdio->cur_page_len) {
L
Linus Torvalds 已提交
623 624 625
		/*
		 * Decrement count only, if we are done with this page
		 */
626 627 628 629 630
		if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
			sdio->pages_in_io--;
		page_cache_get(sdio->cur_page);
		sdio->final_block_in_bio = sdio->cur_page_block +
			(sdio->cur_page_len >> sdio->blkbits);
L
Linus Torvalds 已提交
631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647
		ret = 0;
	} else {
		ret = 1;
	}
	return ret;
}
		
/*
 * Put cur_page under IO.  The section of cur_page which is described by
 * cur_page_offset,cur_page_len is put into a BIO.  The section of cur_page
 * starts on-disk at cur_page_block.
 *
 * We take a ref against the page here (on behalf of its presence in the bio).
 *
 * The caller of this function is responsible for removing cur_page from the
 * dio, and for dropping the refcount which came from that presence.
 */
648 649
static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
		struct buffer_head *map_bh)
L
Linus Torvalds 已提交
650 651 652
{
	int ret = 0;

653 654 655 656
	if (sdio->bio) {
		loff_t cur_offset = sdio->cur_page_fs_offset;
		loff_t bio_next_offset = sdio->logical_offset_in_bio +
			sdio->bio->bi_size;
657

L
Linus Torvalds 已提交
658
		/*
659 660
		 * See whether this new request is contiguous with the old.
		 *
N
Namhyung Kim 已提交
661 662
		 * Btrfs cannot handle having logically non-contiguous requests
		 * submitted.  For example if you have
663 664
		 *
		 * Logical:  [0-4095][HOLE][8192-12287]
N
Namhyung Kim 已提交
665
		 * Physical: [0-4095]      [4096-8191]
666 667 668 669 670
		 *
		 * We cannot submit those pages together as one BIO.  So if our
		 * current logical offset in the file does not equal what would
		 * be the next logical offset in the bio, submit the bio we
		 * have.
L
Linus Torvalds 已提交
671
		 */
672
		if (sdio->final_block_in_bio != sdio->cur_page_block ||
673
		    cur_offset != bio_next_offset)
674
			dio_bio_submit(dio, sdio);
L
Linus Torvalds 已提交
675 676 677 678
		/*
		 * Submit now if the underlying fs is about to perform a
		 * metadata read
		 */
679 680
		else if (sdio->boundary)
			dio_bio_submit(dio, sdio);
L
Linus Torvalds 已提交
681 682
	}

683
	if (sdio->bio == NULL) {
A
Andi Kleen 已提交
684
		ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
L
Linus Torvalds 已提交
685 686 687 688
		if (ret)
			goto out;
	}

689 690
	if (dio_bio_add_page(sdio) != 0) {
		dio_bio_submit(dio, sdio);
A
Andi Kleen 已提交
691
		ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
L
Linus Torvalds 已提交
692
		if (ret == 0) {
693
			ret = dio_bio_add_page(sdio);
L
Linus Torvalds 已提交
694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717
			BUG_ON(ret != 0);
		}
	}
out:
	return ret;
}

/*
 * An autonomous function to put a chunk of a page under deferred IO.
 *
 * The caller doesn't actually know (or care) whether this piece of page is in
 * a BIO, or is under IO or whatever.  We just take care of all possible 
 * situations here.  The separation between the logic of do_direct_IO() and
 * that of submit_page_section() is important for clarity.  Please don't break.
 *
 * The chunk of page starts on-disk at blocknr.
 *
 * We perform deferred IO, by recording the last-submitted page inside our
 * private part of the dio structure.  If possible, we just expand the IO
 * across that page here.
 *
 * If that doesn't work out then we put the old page into the bio and add this
 * page to the dio instead.
 */
718
static inline int
719
submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
A
Andi Kleen 已提交
720 721
		    unsigned offset, unsigned len, sector_t blocknr,
		    struct buffer_head *map_bh)
L
Linus Torvalds 已提交
722 723 724
{
	int ret = 0;

A
Andrew Morton 已提交
725 726 727 728 729 730 731
	if (dio->rw & WRITE) {
		/*
		 * Read accounting is performed in submit_bio()
		 */
		task_io_account_write(len);
	}

L
Linus Torvalds 已提交
732 733 734
	/*
	 * Can we just grow the current page's presence in the dio?
	 */
735 736 737 738 739
	if (sdio->cur_page == page &&
	    sdio->cur_page_offset + sdio->cur_page_len == offset &&
	    sdio->cur_page_block +
	    (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
		sdio->cur_page_len += len;
L
Linus Torvalds 已提交
740 741

		/*
742
		 * If sdio->boundary then we want to schedule the IO now to
L
Linus Torvalds 已提交
743 744
		 * avoid metadata seeks.
		 */
745
		if (sdio->boundary) {
A
Andi Kleen 已提交
746
			ret = dio_send_cur_page(dio, sdio, map_bh);
747 748
			page_cache_release(sdio->cur_page);
			sdio->cur_page = NULL;
L
Linus Torvalds 已提交
749 750 751 752 753 754 755
		}
		goto out;
	}

	/*
	 * If there's a deferred page already there then send it.
	 */
756
	if (sdio->cur_page) {
A
Andi Kleen 已提交
757
		ret = dio_send_cur_page(dio, sdio, map_bh);
758 759
		page_cache_release(sdio->cur_page);
		sdio->cur_page = NULL;
L
Linus Torvalds 已提交
760 761 762 763 764
		if (ret)
			goto out;
	}

	page_cache_get(page);		/* It is in dio */
765 766 767 768 769
	sdio->cur_page = page;
	sdio->cur_page_offset = offset;
	sdio->cur_page_len = len;
	sdio->cur_page_block = blocknr;
	sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
L
Linus Torvalds 已提交
770 771 772 773 774 775 776 777 778
out:
	return ret;
}

/*
 * Clean any dirty buffers in the blockdev mapping which alias newly-created
 * file blocks.  Only called for S_ISREG files - blockdevs do not set
 * buffer_new
 */
A
Andi Kleen 已提交
779
static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
L
Linus Torvalds 已提交
780 781 782 783
{
	unsigned i;
	unsigned nblocks;

A
Andi Kleen 已提交
784
	nblocks = map_bh->b_size >> dio->inode->i_blkbits;
L
Linus Torvalds 已提交
785 786

	for (i = 0; i < nblocks; i++) {
A
Andi Kleen 已提交
787 788
		unmap_underlying_metadata(map_bh->b_bdev,
					  map_bh->b_blocknr + i);
L
Linus Torvalds 已提交
789 790 791 792 793 794 795 796 797 798 799 800
	}
}

/*
 * If we are not writing the entire block and get_block() allocated
 * the block for us, we need to fill-in the unused portion of the
 * block with zeros. This happens only if user-buffer, fileoffset or
 * io length is not filesystem block-size multiple.
 *
 * `end' is zero if we're doing the start of the IO, 1 at the end of the
 * IO.
 */
801 802
static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
		int end, struct buffer_head *map_bh)
L
Linus Torvalds 已提交
803 804 805 806 807 808
{
	unsigned dio_blocks_per_fs_block;
	unsigned this_chunk_blocks;	/* In dio_blocks */
	unsigned this_chunk_bytes;
	struct page *page;

809
	sdio->start_zero_done = 1;
A
Andi Kleen 已提交
810
	if (!sdio->blkfactor || !buffer_new(map_bh))
L
Linus Torvalds 已提交
811 812
		return;

813 814
	dio_blocks_per_fs_block = 1 << sdio->blkfactor;
	this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
L
Linus Torvalds 已提交
815 816 817 818 819 820 821 822 823 824 825

	if (!this_chunk_blocks)
		return;

	/*
	 * We need to zero out part of an fs block.  It is either at the
	 * beginning or the end of the fs block.
	 */
	if (end) 
		this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;

826
	this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
L
Linus Torvalds 已提交
827

N
Nick Piggin 已提交
828
	page = ZERO_PAGE(0);
829
	if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
A
Andi Kleen 已提交
830
				sdio->next_block_for_io, map_bh))
L
Linus Torvalds 已提交
831 832
		return;

833
	sdio->next_block_for_io += this_chunk_blocks;
L
Linus Torvalds 已提交
834 835 836 837 838 839 840 841 842 843 844
}

/*
 * Walk the user pages, and the file, mapping blocks to disk and generating
 * a sequence of (page,offset,len,block) mappings.  These mappings are injected
 * into submit_page_section(), which takes care of the next stage of submission
 *
 * Direct IO against a blockdev is different from a file.  Because we can
 * happily perform page-sized but 512-byte aligned IOs.  It is important that
 * blockdev IO be able to have fine alignment and large sizes.
 *
845
 * So what we do is to permit the ->get_block function to populate bh.b_size
L
Linus Torvalds 已提交
846 847 848
 * with the size of IO which is permitted at this offset and this i_blkbits.
 *
 * For best results, the blockdev should be set up with 512-byte i_blkbits and
849
 * it should set b_size to PAGE_SIZE or more inside get_block().  This gives
L
Linus Torvalds 已提交
850 851
 * fine alignment but still allows this function to work in PAGE_SIZE units.
 */
A
Andi Kleen 已提交
852 853
static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
			struct buffer_head *map_bh)
L
Linus Torvalds 已提交
854
{
855
	const unsigned blkbits = sdio->blkbits;
L
Linus Torvalds 已提交
856 857 858 859 860 861
	const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
	struct page *page;
	unsigned block_in_page;
	int ret = 0;

	/* The I/O can start at any block offset within the first page */
862
	block_in_page = sdio->first_block_in_page;
L
Linus Torvalds 已提交
863

864 865
	while (sdio->block_in_file < sdio->final_block_in_request) {
		page = dio_get_page(dio, sdio);
L
Linus Torvalds 已提交
866 867 868 869 870 871 872 873 874 875 876
		if (IS_ERR(page)) {
			ret = PTR_ERR(page);
			goto out;
		}

		while (block_in_page < blocks_per_page) {
			unsigned offset_in_page = block_in_page << blkbits;
			unsigned this_chunk_bytes;	/* # of bytes mapped */
			unsigned this_chunk_blocks;	/* # of blocks */
			unsigned u;

877
			if (sdio->blocks_available == 0) {
L
Linus Torvalds 已提交
878 879 880 881 882 883
				/*
				 * Need to go and map some more disk
				 */
				unsigned long blkmask;
				unsigned long dio_remainder;

A
Andi Kleen 已提交
884
				ret = get_more_blocks(dio, sdio, map_bh);
L
Linus Torvalds 已提交
885 886 887 888 889 890 891
				if (ret) {
					page_cache_release(page);
					goto out;
				}
				if (!buffer_mapped(map_bh))
					goto do_holes;

892 893 894 895
				sdio->blocks_available =
						map_bh->b_size >> sdio->blkbits;
				sdio->next_block_for_io =
					map_bh->b_blocknr << sdio->blkfactor;
L
Linus Torvalds 已提交
896
				if (buffer_new(map_bh))
A
Andi Kleen 已提交
897
					clean_blockdev_aliases(dio, map_bh);
L
Linus Torvalds 已提交
898

899
				if (!sdio->blkfactor)
L
Linus Torvalds 已提交
900 901
					goto do_holes;

902 903
				blkmask = (1 << sdio->blkfactor) - 1;
				dio_remainder = (sdio->block_in_file & blkmask);
L
Linus Torvalds 已提交
904 905 906 907 908 909 910 911 912 913 914 915 916

				/*
				 * If we are at the start of IO and that IO
				 * starts partway into a fs-block,
				 * dio_remainder will be non-zero.  If the IO
				 * is a read then we can simply advance the IO
				 * cursor to the first block which is to be
				 * read.  But if the IO is a write and the
				 * block was newly allocated we cannot do that;
				 * the start of the fs block must be zeroed out
				 * on-disk
				 */
				if (!buffer_new(map_bh))
917 918
					sdio->next_block_for_io += dio_remainder;
				sdio->blocks_available -= dio_remainder;
L
Linus Torvalds 已提交
919 920 921 922
			}
do_holes:
			/* Handle holes */
			if (!buffer_mapped(map_bh)) {
923
				loff_t i_size_aligned;
L
Linus Torvalds 已提交
924 925

				/* AKPM: eargh, -ENOTBLK is a hack */
J
Jens Axboe 已提交
926
				if (dio->rw & WRITE) {
L
Linus Torvalds 已提交
927 928 929 930
					page_cache_release(page);
					return -ENOTBLK;
				}

931 932 933 934 935 936
				/*
				 * Be sure to account for a partial block as the
				 * last block in the file
				 */
				i_size_aligned = ALIGN(i_size_read(dio->inode),
							1 << blkbits);
937
				if (sdio->block_in_file >=
938
						i_size_aligned >> blkbits) {
L
Linus Torvalds 已提交
939 940 941 942
					/* We hit eof */
					page_cache_release(page);
					goto out;
				}
943 944
				zero_user(page, block_in_page << blkbits,
						1 << blkbits);
945
				sdio->block_in_file++;
L
Linus Torvalds 已提交
946 947 948 949 950 951 952 953 954
				block_in_page++;
				goto next_block;
			}

			/*
			 * If we're performing IO which has an alignment which
			 * is finer than the underlying fs, go check to see if
			 * we must zero out the start of this block.
			 */
955
			if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
A
Andi Kleen 已提交
956
				dio_zero_block(dio, sdio, 0, map_bh);
L
Linus Torvalds 已提交
957 958 959 960 961

			/*
			 * Work out, in this_chunk_blocks, how much disk we
			 * can add to this page
			 */
962
			this_chunk_blocks = sdio->blocks_available;
L
Linus Torvalds 已提交
963 964 965
			u = (PAGE_SIZE - offset_in_page) >> blkbits;
			if (this_chunk_blocks > u)
				this_chunk_blocks = u;
966
			u = sdio->final_block_in_request - sdio->block_in_file;
L
Linus Torvalds 已提交
967 968 969 970 971
			if (this_chunk_blocks > u)
				this_chunk_blocks = u;
			this_chunk_bytes = this_chunk_blocks << blkbits;
			BUG_ON(this_chunk_bytes == 0);

972 973 974 975
			sdio->boundary = buffer_boundary(map_bh);
			ret = submit_page_section(dio, sdio, page,
						  offset_in_page,
						  this_chunk_bytes,
A
Andi Kleen 已提交
976 977
						  sdio->next_block_for_io,
						  map_bh);
L
Linus Torvalds 已提交
978 979 980 981
			if (ret) {
				page_cache_release(page);
				goto out;
			}
982
			sdio->next_block_for_io += this_chunk_blocks;
L
Linus Torvalds 已提交
983

984
			sdio->block_in_file += this_chunk_blocks;
L
Linus Torvalds 已提交
985
			block_in_page += this_chunk_blocks;
986
			sdio->blocks_available -= this_chunk_blocks;
L
Linus Torvalds 已提交
987
next_block:
988 989
			BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
			if (sdio->block_in_file == sdio->final_block_in_request)
L
Linus Torvalds 已提交
990 991 992 993 994 995 996 997 998 999 1000
				break;
		}

		/* Drop the ref which was taken in get_user_pages() */
		page_cache_release(page);
		block_in_page = 0;
	}
out:
	return ret;
}

1001
static inline int drop_refcount(struct dio *dio)
L
Linus Torvalds 已提交
1002
{
1003
	int ret2;
Z
Zach Brown 已提交
1004
	unsigned long flags;
L
Linus Torvalds 已提交
1005

1006 1007
	/*
	 * Sync will always be dropping the final ref and completing the
Z
Zach Brown 已提交
1008 1009 1010 1011 1012 1013 1014 1015
	 * operation.  AIO can if it was a broken operation described above or
	 * in fact if all the bios race to complete before we get here.  In
	 * that case dio_complete() translates the EIOCBQUEUED into the proper
	 * return code that the caller will hand to aio_complete().
	 *
	 * This is managed by the bio_lock instead of being an atomic_t so that
	 * completion paths can drop their ref and use the remaining count to
	 * decide to wake the submission path atomically.
1016
	 */
Z
Zach Brown 已提交
1017 1018 1019
	spin_lock_irqsave(&dio->bio_lock, flags);
	ret2 = --dio->refcount;
	spin_unlock_irqrestore(&dio->bio_lock, flags);
1020
	return ret2;
L
Linus Torvalds 已提交
1021 1022
}

1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034
/*
 * This is a library function for use by filesystem drivers.
 *
 * The locking rules are governed by the flags parameter:
 *  - if the flags value contains DIO_LOCKING we use a fancy locking
 *    scheme for dumb filesystems.
 *    For writes this function is called under i_mutex and returns with
 *    i_mutex held, for reads, i_mutex is not held on entry, but it is
 *    taken and dropped again before returning.
 *  - if the flags value does NOT contain DIO_LOCKING we don't use any
 *    internal locking but rather rely on the filesystem to synchronize
 *    direct I/O reads/writes versus each other and truncate.
1035 1036 1037 1038 1039 1040 1041
 *
 * To help with locking against truncate we incremented the i_dio_count
 * counter before starting direct I/O, and decrement it once we are done.
 * Truncate can wait for it to reach zero to provide exclusion.  It is
 * expected that filesystem provide exclusion between new direct I/O
 * and truncates.  For DIO_LOCKING filesystems this is done by i_mutex,
 * but other filesystems need to take care of this on their own.
1042 1043 1044 1045 1046
 *
 * NOTE: if you pass "sdio" to anything by pointer make sure that function
 * is always inlined. Otherwise gcc is unable to split the structure into
 * individual fields and will generate much worse code. This is important
 * for the whole file.
1047
 */
1048 1049
static inline ssize_t
do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
L
Linus Torvalds 已提交
1050
	struct block_device *bdev, const struct iovec *iov, loff_t offset, 
1051
	unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1052
	dio_submit_t submit_io,	int flags)
L
Linus Torvalds 已提交
1053 1054 1055 1056
{
	int seg;
	size_t size;
	unsigned long addr;
1057 1058
	unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
	unsigned blkbits = i_blkbits;
L
Linus Torvalds 已提交
1059 1060 1061 1062
	unsigned blocksize_mask = (1 << blkbits) - 1;
	ssize_t retval = -EINVAL;
	loff_t end = offset;
	struct dio *dio;
1063
	struct dio_submit sdio = { 0, };
1064 1065 1066
	unsigned long user_addr;
	size_t bytes;
	struct buffer_head map_bh = { 0, };
1067
	struct blk_plug plug;
L
Linus Torvalds 已提交
1068 1069

	if (rw & WRITE)
J
Jens Axboe 已提交
1070
		rw = WRITE_ODIRECT;
L
Linus Torvalds 已提交
1071

1072 1073 1074 1075
	/*
	 * Avoid references to bdev if not absolutely needed to give
	 * the early prefetch in the caller enough time.
	 */
L
Linus Torvalds 已提交
1076 1077 1078

	if (offset & blocksize_mask) {
		if (bdev)
1079
			blkbits = blksize_bits(bdev_logical_block_size(bdev));
L
Linus Torvalds 已提交
1080 1081 1082 1083 1084 1085 1086 1087 1088 1089
		blocksize_mask = (1 << blkbits) - 1;
		if (offset & blocksize_mask)
			goto out;
	}

	/* Check the memory alignment.  Blocks cannot straddle pages */
	for (seg = 0; seg < nr_segs; seg++) {
		addr = (unsigned long)iov[seg].iov_base;
		size = iov[seg].iov_len;
		end += size;
1090 1091
		if (unlikely((addr & blocksize_mask) ||
			     (size & blocksize_mask))) {
L
Linus Torvalds 已提交
1092
			if (bdev)
1093 1094
				blkbits = blksize_bits(
					 bdev_logical_block_size(bdev));
L
Linus Torvalds 已提交
1095
			blocksize_mask = (1 << blkbits) - 1;
1096
			if ((addr & blocksize_mask) || (size & blocksize_mask))
L
Linus Torvalds 已提交
1097 1098 1099 1100
				goto out;
		}
	}

1101 1102 1103 1104
	/* watch out for a 0 len io from a tricksy fs */
	if (rw == READ && end == offset)
		return 0;

1105
	dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
L
Linus Torvalds 已提交
1106 1107 1108
	retval = -ENOMEM;
	if (!dio)
		goto out;
1109 1110 1111 1112 1113 1114
	/*
	 * Believe it or not, zeroing out the page array caused a .5%
	 * performance regression in a database benchmark.  So, we take
	 * care to only zero out what's needed.
	 */
	memset(dio, 0, offsetof(struct dio, pages));
L
Linus Torvalds 已提交
1115

1116 1117
	dio->flags = flags;
	if (dio->flags & DIO_LOCKING) {
1118
		if (rw == READ) {
1119 1120
			struct address_space *mapping =
					iocb->ki_filp->f_mapping;
L
Linus Torvalds 已提交
1121

1122 1123
			/* will be released by direct_io_worker */
			mutex_lock(&inode->i_mutex);
L
Linus Torvalds 已提交
1124 1125 1126 1127

			retval = filemap_write_and_wait_range(mapping, offset,
							      end - 1);
			if (retval) {
1128
				mutex_unlock(&inode->i_mutex);
1129
				kmem_cache_free(dio_cache, dio);
L
Linus Torvalds 已提交
1130 1131 1132 1133 1134
				goto out;
			}
		}
	}

1135 1136 1137 1138 1139
	/*
	 * Will be decremented at I/O completion time.
	 */
	atomic_inc(&inode->i_dio_count);

L
Linus Torvalds 已提交
1140 1141 1142 1143 1144 1145
	/*
	 * For file extending writes updating i_size before data
	 * writeouts complete can expose uninitialized blocks. So
	 * even for AIO, we need to wait for i/o to complete before
	 * returning in this case.
	 */
J
Jens Axboe 已提交
1146
	dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
L
Linus Torvalds 已提交
1147 1148
		(end > i_size_read(inode)));

1149 1150 1151 1152 1153
	retval = 0;

	dio->inode = inode;
	dio->rw = rw;
	sdio.blkbits = blkbits;
1154
	sdio.blkfactor = i_blkbits - blkbits;
1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182
	sdio.block_in_file = offset >> blkbits;

	sdio.get_block = get_block;
	dio->end_io = end_io;
	sdio.submit_io = submit_io;
	sdio.final_block_in_bio = -1;
	sdio.next_block_for_io = -1;

	dio->iocb = iocb;
	dio->i_size = i_size_read(inode);

	spin_lock_init(&dio->bio_lock);
	dio->refcount = 1;

	/*
	 * In case of non-aligned buffers, we may need 2 more
	 * pages since we need to zero out first and last block.
	 */
	if (unlikely(sdio.blkfactor))
		sdio.pages_in_io = 2;

	for (seg = 0; seg < nr_segs; seg++) {
		user_addr = (unsigned long)iov[seg].iov_base;
		sdio.pages_in_io +=
			((user_addr + iov[seg].iov_len + PAGE_SIZE-1) /
				PAGE_SIZE - user_addr / PAGE_SIZE);
	}

1183 1184
	blk_start_plug(&plug);

1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242
	for (seg = 0; seg < nr_segs; seg++) {
		user_addr = (unsigned long)iov[seg].iov_base;
		sdio.size += bytes = iov[seg].iov_len;

		/* Index into the first page of the first block */
		sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
		sdio.final_block_in_request = sdio.block_in_file +
						(bytes >> blkbits);
		/* Page fetching state */
		sdio.head = 0;
		sdio.tail = 0;
		sdio.curr_page = 0;

		sdio.total_pages = 0;
		if (user_addr & (PAGE_SIZE-1)) {
			sdio.total_pages++;
			bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
		}
		sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
		sdio.curr_user_address = user_addr;

		retval = do_direct_IO(dio, &sdio, &map_bh);

		dio->result += iov[seg].iov_len -
			((sdio.final_block_in_request - sdio.block_in_file) <<
					blkbits);

		if (retval) {
			dio_cleanup(dio, &sdio);
			break;
		}
	} /* end iovec loop */

	if (retval == -ENOTBLK) {
		/*
		 * The remaining part of the request will be
		 * be handled by buffered I/O when we return
		 */
		retval = 0;
	}
	/*
	 * There may be some unwritten disk at the end of a part-written
	 * fs-block-sized block.  Go zero that now.
	 */
	dio_zero_block(dio, &sdio, 1, &map_bh);

	if (sdio.cur_page) {
		ssize_t ret2;

		ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
		if (retval == 0)
			retval = ret2;
		page_cache_release(sdio.cur_page);
		sdio.cur_page = NULL;
	}
	if (sdio.bio)
		dio_bio_submit(dio, &sdio);

1243 1244
	blk_finish_plug(&plug);

1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267
	/*
	 * It is possible that, we return short IO due to end of file.
	 * In that case, we need to release all the pages we got hold on.
	 */
	dio_cleanup(dio, &sdio);

	/*
	 * All block lookups have been performed. For READ requests
	 * we can let i_mutex go now that its achieved its purpose
	 * of protecting us from looking up uninitialized blocks.
	 */
	if (rw == READ && (dio->flags & DIO_LOCKING))
		mutex_unlock(&dio->inode->i_mutex);

	/*
	 * The only time we want to leave bios in flight is when a successful
	 * partial aio read or full aio write have been setup.  In that case
	 * bio completion will call aio_complete.  The only time it's safe to
	 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
	 * This had *better* be the only place that raises -EIOCBQUEUED.
	 */
	BUG_ON(retval == -EIOCBQUEUED);
	if (dio->is_async && retval == 0 && dio->result &&
1268
	    ((rw == READ) || (dio->result == sdio.size)))
1269 1270 1271 1272 1273 1274 1275 1276 1277 1278
		retval = -EIOCBQUEUED;

	if (retval != -EIOCBQUEUED)
		dio_await_completion(dio);

	if (drop_refcount(dio) == 0) {
		retval = dio_complete(dio, offset, retval, false);
		kmem_cache_free(dio_cache, dio);
	} else
		BUG_ON(retval != -EIOCBQUEUED);
L
Linus Torvalds 已提交
1279

1280 1281 1282
out:
	return retval;
}
1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306

ssize_t
__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
	struct block_device *bdev, const struct iovec *iov, loff_t offset,
	unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
	dio_submit_t submit_io,	int flags)
{
	/*
	 * The block device state is needed in the end to finally
	 * submit everything.  Since it's likely to be cache cold
	 * prefetch it here as first thing to hide some of the
	 * latency.
	 *
	 * Attempt to prefetch the pieces we likely need later.
	 */
	prefetch(&bdev->bd_disk->part_tbl);
	prefetch(bdev->bd_queue);
	prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);

	return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset,
				     nr_segs, get_block, end_io,
				     submit_io, flags);
}

L
Linus Torvalds 已提交
1307
EXPORT_SYMBOL(__blockdev_direct_IO);
1308 1309 1310 1311 1312 1313 1314

static __init int dio_init(void)
{
	dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
	return 0;
}
module_init(dio_init)