direct-io.c 37.3 KB
Newer Older
Linus Torvalds's avatar
Linus Torvalds committed
1 2 3 4 5 6 7
/*
 * fs/direct-io.c
 *
 * Copyright (C) 2002, Linus Torvalds.
 *
 * O_DIRECT
 *
8
 * 04Jul2002	Andrew Morton
Linus Torvalds's avatar
Linus Torvalds committed
9 10 11
 *		Initial version
 * 11Sep2002	janetinc@us.ibm.com
 * 		added readv/writev support.
12
 * 29Oct2002	Andrew Morton
Linus Torvalds's avatar
Linus Torvalds committed
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>
30
#include <linux/task_io_accounting_ops.h>
Linus Torvalds's avatar
Linus Torvalds committed
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>
Arun Sharma's avatar
Arun Sharma committed
38
#include <linux/atomic.h>
39
#include <linux/prefetch.h>
40
#include <linux/aio.h>
Linus Torvalds's avatar
Linus Torvalds committed
41 42 43

/*
 * How many user pages to map in one call to get_user_pages().  This determines
Andi Kleen's avatar
Andi Kleen committed
44
 * the size of a structure in the slab cache
Linus Torvalds's avatar
Linus Torvalds committed
45 46 47 48 49 50 51 52 53 54 55 56 57 58 59
 */
#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.
 */

60 61 62
/* dio_state only used in the submission path */

struct dio_submit {
Linus Torvalds's avatar
Linus Torvalds committed
63 64 65 66 67 68 69 70 71 72 73 74 75 76 77
	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 */
78
	int reap_counter;		/* rate limit reaping */
Linus Torvalds's avatar
Linus Torvalds committed
79 80 81
	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 */
82
	get_block_t *get_block;		/* block mapping function */
83
	dio_submit_t *submit_io;	/* IO submition function */
84

85
	loff_t logical_offset_in_bio;	/* current first logical block in bio */
Linus Torvalds's avatar
Linus Torvalds committed
86 87 88 89 90 91 92 93 94 95 96 97 98
	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 */
99
	loff_t cur_page_fs_offset;	/* Offset in file */
Linus Torvalds's avatar
Linus Torvalds committed
100 101 102 103 104 105 106 107 108 109 110 111 112 113

	/*
	 * 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 */
114 115 116 117 118 119
};

/* dio_state communicated between submission path and end_io */
struct dio {
	int flags;			/* doesn't change */
	int rw;
120
	struct inode *inode;
121 122 123
	loff_t i_size;			/* i_size when submitted */
	dio_iodone_t *end_io;		/* IO completion function */

124
	void *private;			/* copy from map_bh.b_private */
125 126 127

	/* BIO completion state */
	spinlock_t bio_lock;		/* protects BIO fields below */
128 129 130
	int page_errors;		/* errno from get_user_pages() */
	int is_async;			/* is IO async ? */
	int io_error;			/* IO error in completion path */
131 132 133 134 135 136 137 138
	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 */

139 140 141 142 143 144
	/*
	 * 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 */
145 146 147
} ____cacheline_aligned_in_smp;

static struct kmem_cache *dio_cache __read_mostly;
Linus Torvalds's avatar
Linus Torvalds committed
148 149 150 151

/*
 * How many pages are in the queue?
 */
152
static inline unsigned dio_pages_present(struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
153
{
154
	return sdio->tail - sdio->head;
Linus Torvalds's avatar
Linus Torvalds committed
155 156 157 158 159
}

/*
 * Go grab and pin some userspace pages.   Typically we'll get 64 at a time.
 */
160
static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
161 162 163 164
{
	int ret;
	int nr_pages;

165
	nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
Nick Piggin's avatar
Nick Piggin committed
166
	ret = get_user_pages_fast(
167
		sdio->curr_user_address,		/* Where from? */
Linus Torvalds's avatar
Linus Torvalds committed
168 169
		nr_pages,			/* How many pages? */
		dio->rw == READ,		/* Write to memory? */
Nick Piggin's avatar
Nick Piggin committed
170
		&dio->pages[0]);		/* Put results here */
Linus Torvalds's avatar
Linus Torvalds committed
171

172
	if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
Nick Piggin's avatar
Nick Piggin committed
173
		struct page *page = ZERO_PAGE(0);
Linus Torvalds's avatar
Linus Torvalds committed
174 175 176 177 178 179 180
		/*
		 * 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;
Nick Piggin's avatar
Nick Piggin committed
181 182
		page_cache_get(page);
		dio->pages[0] = page;
183 184
		sdio->head = 0;
		sdio->tail = 1;
Linus Torvalds's avatar
Linus Torvalds committed
185 186 187 188 189
		ret = 0;
		goto out;
	}

	if (ret >= 0) {
190 191 192 193
		sdio->curr_user_address += ret * PAGE_SIZE;
		sdio->curr_page += ret;
		sdio->head = 0;
		sdio->tail = ret;
Linus Torvalds's avatar
Linus Torvalds committed
194 195 196 197 198 199 200 201 202 203 204 205
		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.
 */
206 207
static inline struct page *dio_get_page(struct dio *dio,
		struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
208
{
209
	if (dio_pages_present(sdio) == 0) {
Linus Torvalds's avatar
Linus Torvalds committed
210 211
		int ret;

212
		ret = dio_refill_pages(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
213 214
		if (ret)
			return ERR_PTR(ret);
215
		BUG_ON(dio_pages_present(sdio) == 0);
Linus Torvalds's avatar
Linus Torvalds committed
216
	}
217
	return dio->pages[sdio->head++];
Linus Torvalds's avatar
Linus Torvalds committed
218 219
}

220 221 222 223 224 225 226 227 228 229 230 231
/**
 * 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.
Linus Torvalds's avatar
Linus Torvalds committed
232
 */
233
static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async)
Linus Torvalds's avatar
Linus Torvalds committed
234
{
235 236
	ssize_t transferred = 0;

237 238 239 240 241 242 243 244 245
	/*
	 * 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;

246 247 248 249 250 251 252 253 254 255 256 257 258 259 260
	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;

261 262
	if (dio->end_io && dio->result) {
		dio->end_io(dio->iocb, offset, transferred,
263
			    dio->private, ret, is_async);
264
	} else {
265
		inode_dio_done(dio->inode);
266 267
		if (is_async)
			aio_complete(dio->iocb, ret, 0);
268 269
	}

270
	return ret;
Linus Torvalds's avatar
Linus Torvalds committed
271 272 273 274 275 276
}

static int dio_bio_complete(struct dio *dio, struct bio *bio);
/*
 * Asynchronous IO callback. 
 */
277
static void dio_bio_end_aio(struct bio *bio, int error)
Linus Torvalds's avatar
Linus Torvalds committed
278 279
{
	struct dio *dio = bio->bi_private;
280 281
	unsigned long remaining;
	unsigned long flags;
Linus Torvalds's avatar
Linus Torvalds committed
282 283 284

	/* cleanup the bio */
	dio_bio_complete(dio, bio);
285

286 287 288
	spin_lock_irqsave(&dio->bio_lock, flags);
	remaining = --dio->refcount;
	if (remaining == 1 && dio->waiter)
289
		wake_up_process(dio->waiter);
290
	spin_unlock_irqrestore(&dio->bio_lock, flags);
291

292
	if (remaining == 0) {
293
		dio_complete(dio, dio->iocb->ki_pos, 0, true);
294
		kmem_cache_free(dio_cache, dio);
295
	}
Linus Torvalds's avatar
Linus Torvalds committed
296 297 298 299 300 301 302 303 304
}

/*
 * 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.
 */
305
static void dio_bio_end_io(struct bio *bio, int error)
Linus Torvalds's avatar
Linus Torvalds committed
306 307 308 309 310 311 312
{
	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;
313
	if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds's avatar
Linus Torvalds committed
314 315 316 317
		wake_up_process(dio->waiter);
	spin_unlock_irqrestore(&dio->bio_lock, flags);
}

318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337
/**
 * 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);

338
static inline void
339 340 341
dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
	      struct block_device *bdev,
	      sector_t first_sector, int nr_vecs)
Linus Torvalds's avatar
Linus Torvalds committed
342 343 344
{
	struct bio *bio;

345 346 347 348
	/*
	 * bio_alloc() is guaranteed to return a bio when called with
	 * __GFP_WAIT and we request a valid number of vectors.
	 */
Linus Torvalds's avatar
Linus Torvalds committed
349 350 351 352 353 354 355 356 357
	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;

358 359
	sdio->bio = bio;
	sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds's avatar
Linus Torvalds committed
360 361 362 363 364 365
}

/*
 * 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().
366 367
 *
 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds's avatar
Linus Torvalds committed
368
 */
369
static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
370
{
371
	struct bio *bio = sdio->bio;
372
	unsigned long flags;
Linus Torvalds's avatar
Linus Torvalds committed
373 374

	bio->bi_private = dio;
375 376 377 378 379

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

Linus Torvalds's avatar
Linus Torvalds committed
380 381
	if (dio->is_async && dio->rw == READ)
		bio_set_pages_dirty(bio);
382

383 384 385
	if (sdio->submit_io)
		sdio->submit_io(dio->rw, bio, dio->inode,
			       sdio->logical_offset_in_bio);
386 387
	else
		submit_bio(dio->rw, bio);
Linus Torvalds's avatar
Linus Torvalds committed
388

389 390 391
	sdio->bio = NULL;
	sdio->boundary = 0;
	sdio->logical_offset_in_bio = 0;
Linus Torvalds's avatar
Linus Torvalds committed
392 393 394 395 396
}

/*
 * Release any resources in case of a failure
 */
397
static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
398
{
399 400
	while (dio_pages_present(sdio))
		page_cache_release(dio_get_page(dio, sdio));
Linus Torvalds's avatar
Linus Torvalds committed
401 402 403
}

/*
404 405 406 407
 * 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.
Linus Torvalds's avatar
Linus Torvalds committed
408 409 410 411
 */
static struct bio *dio_await_one(struct dio *dio)
{
	unsigned long flags;
412
	struct bio *bio = NULL;
Linus Torvalds's avatar
Linus Torvalds committed
413 414

	spin_lock_irqsave(&dio->bio_lock, flags);
415 416 417 418 419 420 421 422 423 424 425 426 427 428 429

	/*
	 * 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;
Linus Torvalds's avatar
Linus Torvalds committed
430
	}
431 432 433 434
	if (dio->bio_list) {
		bio = dio->bio_list;
		dio->bio_list = bio->bi_private;
	}
Linus Torvalds's avatar
Linus Torvalds committed
435 436 437 438 439 440 441 442 443 444
	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);
445 446
	struct bio_vec *bvec;
	unsigned i;
Linus Torvalds's avatar
Linus Torvalds committed
447 448

	if (!uptodate)
449
		dio->io_error = -EIO;
Linus Torvalds's avatar
Linus Torvalds committed
450 451 452 453

	if (dio->is_async && dio->rw == READ) {
		bio_check_pages_dirty(bio);	/* transfers ownership */
	} else {
454 455
		bio_for_each_segment_all(bvec, bio, i) {
			struct page *page = bvec->bv_page;
Linus Torvalds's avatar
Linus Torvalds committed
456 457 458 459 460 461 462 463 464 465 466

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

/*
467 468 469
 * 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
470
 * errors are propagated through dio->io_error and should be propagated via
471
 * dio_complete().
Linus Torvalds's avatar
Linus Torvalds committed
472
 */
473
static void dio_await_completion(struct dio *dio)
Linus Torvalds's avatar
Linus Torvalds committed
474
{
475 476 477 478 479 480
	struct bio *bio;
	do {
		bio = dio_await_one(dio);
		if (bio)
			dio_bio_complete(dio, bio);
	} while (bio);
Linus Torvalds's avatar
Linus Torvalds committed
481 482 483 484 485 486 487 488 489
}

/*
 * 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.
 */
490
static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
491 492 493
{
	int ret = 0;

494
	if (sdio->reap_counter++ >= 64) {
Linus Torvalds's avatar
Linus Torvalds committed
495 496 497 498 499 500 501 502 503 504 505 506 507
		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;
		}
508
		sdio->reap_counter = 0;
Linus Torvalds's avatar
Linus Torvalds committed
509 510 511 512 513 514
	}
	return ret;
}

/*
 * Call into the fs to map some more disk blocks.  We record the current number
515
 * of available blocks at sdio->blocks_available.  These are in units of the
Linus Torvalds's avatar
Linus Torvalds committed
516 517 518 519 520
 * 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.
 *
521
 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds's avatar
Linus Torvalds committed
522 523 524 525 526 527 528 529 530 531 532 533
 * 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
534
 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds's avatar
Linus Torvalds committed
535
 */
536 537
static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
			   struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
538 539 540
{
	int ret;
	sector_t fs_startblk;	/* Into file, in filesystem-sized blocks */
541
	sector_t fs_endblk;	/* Into file, in filesystem-sized blocks */
Linus Torvalds's avatar
Linus Torvalds committed
542 543
	unsigned long fs_count;	/* Number of filesystem-sized blocks */
	int create;
544
	unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
Linus Torvalds's avatar
Linus Torvalds committed
545 546 547 548 549 550 551

	/*
	 * 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) {
552 553
		BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
		fs_startblk = sdio->block_in_file >> sdio->blkfactor;
554 555 556
		fs_endblk = (sdio->final_block_in_request - 1) >>
					sdio->blkfactor;
		fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds's avatar
Linus Torvalds committed
557

558
		map_bh->b_state = 0;
559
		map_bh->b_size = fs_count << i_blkbits;
560

561 562 563 564 565 566 567 568 569 570 571
		/*
		 * 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.
		 */
Jens Axboe's avatar
Jens Axboe committed
572
		create = dio->rw & WRITE;
573
		if (dio->flags & DIO_SKIP_HOLES) {
574 575
			if (sdio->block_in_file < (i_size_read(dio->inode) >>
							sdio->blkbits))
Linus Torvalds's avatar
Linus Torvalds committed
576 577
				create = 0;
		}
578

579
		ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds's avatar
Linus Torvalds committed
580
						map_bh, create);
581 582 583

		/* Store for completion */
		dio->private = map_bh->b_private;
Linus Torvalds's avatar
Linus Torvalds committed
584 585 586 587 588 589 590
	}
	return ret;
}

/*
 * There is no bio.  Make one now.
 */
591 592
static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
		sector_t start_sector, struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
593 594 595 596
{
	sector_t sector;
	int ret, nr_pages;

597
	ret = dio_bio_reap(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
598 599
	if (ret)
		goto out;
600
	sector = start_sector << (sdio->blkbits - 9);
601
	nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
602
	nr_pages = min(nr_pages, BIO_MAX_PAGES);
Linus Torvalds's avatar
Linus Torvalds committed
603
	BUG_ON(nr_pages <= 0);
604
	dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
605
	sdio->boundary = 0;
Linus Torvalds's avatar
Linus Torvalds committed
606 607 608 609 610 611 612 613 614 615 616
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.
 */
617
static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
618 619 620
{
	int ret;

621 622 623
	ret = bio_add_page(sdio->bio, sdio->cur_page,
			sdio->cur_page_len, sdio->cur_page_offset);
	if (ret == sdio->cur_page_len) {
Linus Torvalds's avatar
Linus Torvalds committed
624 625 626
		/*
		 * Decrement count only, if we are done with this page
		 */
627 628 629 630 631
		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);
Linus Torvalds's avatar
Linus Torvalds committed
632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648
		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.
 */
649 650
static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
		struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
651 652 653
{
	int ret = 0;

654 655 656 657
	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;
658

Linus Torvalds's avatar
Linus Torvalds committed
659
		/*
660 661
		 * See whether this new request is contiguous with the old.
		 *
Namhyung Kim's avatar
Namhyung Kim committed
662 663
		 * Btrfs cannot handle having logically non-contiguous requests
		 * submitted.  For example if you have
664 665
		 *
		 * Logical:  [0-4095][HOLE][8192-12287]
Namhyung Kim's avatar
Namhyung Kim committed
666
		 * Physical: [0-4095]      [4096-8191]
667 668 669 670 671
		 *
		 * 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.
Linus Torvalds's avatar
Linus Torvalds committed
672
		 */
673
		if (sdio->final_block_in_bio != sdio->cur_page_block ||
674
		    cur_offset != bio_next_offset)
675
			dio_bio_submit(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
676 677
	}

678
	if (sdio->bio == NULL) {
679
		ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
680 681 682 683
		if (ret)
			goto out;
	}

684 685
	if (dio_bio_add_page(sdio) != 0) {
		dio_bio_submit(dio, sdio);
686
		ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
687
		if (ret == 0) {
688
			ret = dio_bio_add_page(sdio);
Linus Torvalds's avatar
Linus Torvalds committed
689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712
			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.
 */
713
static inline int
714
submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
715 716
		    unsigned offset, unsigned len, sector_t blocknr,
		    struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
717 718 719
{
	int ret = 0;

720 721 722 723 724 725 726
	if (dio->rw & WRITE) {
		/*
		 * Read accounting is performed in submit_bio()
		 */
		task_io_account_write(len);
	}

Linus Torvalds's avatar
Linus Torvalds committed
727 728 729
	/*
	 * Can we just grow the current page's presence in the dio?
	 */
730 731 732 733 734
	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;
Linus Torvalds's avatar
Linus Torvalds committed
735 736 737 738 739 740
		goto out;
	}

	/*
	 * If there's a deferred page already there then send it.
	 */
741
	if (sdio->cur_page) {
742
		ret = dio_send_cur_page(dio, sdio, map_bh);
743 744
		page_cache_release(sdio->cur_page);
		sdio->cur_page = NULL;
Linus Torvalds's avatar
Linus Torvalds committed
745
		if (ret)
746
			return ret;
Linus Torvalds's avatar
Linus Torvalds committed
747 748 749
	}

	page_cache_get(page);		/* It is in dio */
750 751 752 753 754
	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;
Linus Torvalds's avatar
Linus Torvalds committed
755
out:
756 757 758 759 760 761 762 763 764 765
	/*
	 * If sdio->boundary then we want to schedule the IO now to
	 * avoid metadata seeks.
	 */
	if (sdio->boundary) {
		ret = dio_send_cur_page(dio, sdio, map_bh);
		dio_bio_submit(dio, sdio);
		page_cache_release(sdio->cur_page);
		sdio->cur_page = NULL;
	}
Linus Torvalds's avatar
Linus Torvalds committed
766 767 768 769 770 771 772 773
	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
 */
774
static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
775 776 777 778
{
	unsigned i;
	unsigned nblocks;

779
	nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds's avatar
Linus Torvalds committed
780 781

	for (i = 0; i < nblocks; i++) {
782 783
		unmap_underlying_metadata(map_bh->b_bdev,
					  map_bh->b_blocknr + i);
Linus Torvalds's avatar
Linus Torvalds committed
784 785 786 787 788 789 790 791 792 793 794 795
	}
}

/*
 * 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.
 */
796 797
static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
		int end, struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
798 799 800 801 802 803
{
	unsigned dio_blocks_per_fs_block;
	unsigned this_chunk_blocks;	/* In dio_blocks */
	unsigned this_chunk_bytes;
	struct page *page;

804
	sdio->start_zero_done = 1;
805
	if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds's avatar
Linus Torvalds committed
806 807
		return;

808 809
	dio_blocks_per_fs_block = 1 << sdio->blkfactor;
	this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
Linus Torvalds's avatar
Linus Torvalds committed
810 811 812 813 814 815 816 817 818 819 820

	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;

821
	this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds's avatar
Linus Torvalds committed
822

Nick Piggin's avatar
Nick Piggin committed
823
	page = ZERO_PAGE(0);
824
	if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
825
				sdio->next_block_for_io, map_bh))
Linus Torvalds's avatar
Linus Torvalds committed
826 827
		return;

828
	sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds's avatar
Linus Torvalds committed
829 830 831 832 833 834 835 836 837 838 839
}

/*
 * 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.
 *
840
 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds's avatar
Linus Torvalds committed
841 842 843
 * 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
844
 * it should set b_size to PAGE_SIZE or more inside get_block().  This gives
Linus Torvalds's avatar
Linus Torvalds committed
845 846
 * fine alignment but still allows this function to work in PAGE_SIZE units.
 */
847 848
static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
			struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
849
{
850
	const unsigned blkbits = sdio->blkbits;
Linus Torvalds's avatar
Linus Torvalds committed
851 852 853 854 855 856
	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 */
857
	block_in_page = sdio->first_block_in_page;
Linus Torvalds's avatar
Linus Torvalds committed
858

859 860
	while (sdio->block_in_file < sdio->final_block_in_request) {
		page = dio_get_page(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
861 862 863 864 865 866 867 868 869 870 871
		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;

872
			if (sdio->blocks_available == 0) {
Linus Torvalds's avatar
Linus Torvalds committed
873 874 875 876 877 878
				/*
				 * Need to go and map some more disk
				 */
				unsigned long blkmask;
				unsigned long dio_remainder;

879
				ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
880 881 882 883 884 885 886
				if (ret) {
					page_cache_release(page);
					goto out;
				}
				if (!buffer_mapped(map_bh))
					goto do_holes;

887 888 889 890
				sdio->blocks_available =
						map_bh->b_size >> sdio->blkbits;
				sdio->next_block_for_io =
					map_bh->b_blocknr << sdio->blkfactor;
Linus Torvalds's avatar
Linus Torvalds committed
891
				if (buffer_new(map_bh))
892
					clean_blockdev_aliases(dio, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
893

894
				if (!sdio->blkfactor)
Linus Torvalds's avatar
Linus Torvalds committed
895 896
					goto do_holes;

897 898
				blkmask = (1 << sdio->blkfactor) - 1;
				dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds's avatar
Linus Torvalds committed
899 900 901 902 903 904 905 906 907 908 909 910 911

				/*
				 * 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))
912 913
					sdio->next_block_for_io += dio_remainder;
				sdio->blocks_available -= dio_remainder;
Linus Torvalds's avatar
Linus Torvalds committed
914 915 916 917
			}
do_holes:
			/* Handle holes */
			if (!buffer_mapped(map_bh)) {
918
				loff_t i_size_aligned;
Linus Torvalds's avatar
Linus Torvalds committed
919 920

				/* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboe's avatar
Jens Axboe committed
921
				if (dio->rw & WRITE) {
Linus Torvalds's avatar
Linus Torvalds committed
922 923 924 925
					page_cache_release(page);
					return -ENOTBLK;
				}

926 927 928 929 930 931
				/*
				 * 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);
932
				if (sdio->block_in_file >=
933
						i_size_aligned >> blkbits) {
Linus Torvalds's avatar
Linus Torvalds committed
934 935 936 937
					/* We hit eof */
					page_cache_release(page);
					goto out;
				}
938 939
				zero_user(page, block_in_page << blkbits,
						1 << blkbits);
940
				sdio->block_in_file++;
Linus Torvalds's avatar
Linus Torvalds committed
941 942 943 944 945 946 947 948 949
				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.
			 */
950
			if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
951
				dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
952 953 954 955 956

			/*
			 * Work out, in this_chunk_blocks, how much disk we
			 * can add to this page
			 */
957
			this_chunk_blocks = sdio->blocks_available;
Linus Torvalds's avatar
Linus Torvalds committed
958 959 960
			u = (PAGE_SIZE - offset_in_page) >> blkbits;
			if (this_chunk_blocks > u)
				this_chunk_blocks = u;
961
			u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds's avatar
Linus Torvalds committed
962 963 964 965 966
			if (this_chunk_blocks > u)
				this_chunk_blocks = u;
			this_chunk_bytes = this_chunk_blocks << blkbits;
			BUG_ON(this_chunk_bytes == 0);

967 968
			if (this_chunk_blocks == sdio->blocks_available)
				sdio->boundary = buffer_boundary(map_bh);
969 970 971
			ret = submit_page_section(dio, sdio, page,
						  offset_in_page,
						  this_chunk_bytes,
972 973
						  sdio->next_block_for_io,
						  map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
974 975 976 977
			if (ret) {
				page_cache_release(page);
				goto out;
			}
978
			sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds's avatar
Linus Torvalds committed
979

980
			sdio->block_in_file += this_chunk_blocks;
Linus Torvalds's avatar
Linus Torvalds committed
981
			block_in_page += this_chunk_blocks;
982
			sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds's avatar
Linus Torvalds committed
983
next_block:
984 985
			BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
			if (sdio->block_in_file == sdio->final_block_in_request)
Linus Torvalds's avatar
Linus Torvalds committed
986 987 988 989 990 991 992 993 994 995 996
				break;
		}

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

997
static inline int drop_refcount(struct dio *dio)
Linus Torvalds's avatar
Linus Torvalds committed
998
{
999
	int ret2;
1000
	unsigned long flags;
Linus Torvalds's avatar
Linus Torvalds committed
1001

1002 1003
	/*
	 * Sync will always be dropping the final ref and completing the
1004 1005 1006 1007 1008 1009 1010 1011
	 * 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.
1012
	 */
1013 1014 1015
	spin_lock_irqsave(&dio->bio_lock, flags);
	ret2 = --dio->refcount;
	spin_unlock_irqrestore(&dio->bio_lock, flags);
1016
	return ret2;
Linus Torvalds's avatar
Linus Torvalds committed
1017 1018
}

1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030
/*
 * 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.
1031 1032 1033 1034 1035 1036 1037
 *
 * 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.
1038 1039 1040 1041 1042
 *
 * 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.
1043
 */
1044 1045
static inline ssize_t
do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
Linus Torvalds's avatar
Linus Torvalds committed
1046
	struct block_device *bdev, const struct iovec *iov, loff_t offset, 
1047
	unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1048
	dio_submit_t submit_io,	int flags)
Linus Torvalds's avatar
Linus Torvalds committed
1049 1050 1051 1052
{
	int seg;
	size_t size;
	unsigned long addr;
1053 1054
	unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
	unsigned blkbits = i_blkbits;
Linus Torvalds's avatar
Linus Torvalds committed
1055 1056 1057 1058
	unsigned blocksize_mask = (1 << blkbits) - 1;
	ssize_t retval = -EINVAL;
	loff_t end = offset;
	struct dio *dio;
1059
	struct dio_submit sdio = { 0, };
1060 1061 1062
	unsigned long user_addr;
	size_t bytes;
	struct buffer_head map_bh = { 0, };
1063
	struct blk_plug plug;
Linus Torvalds's avatar
Linus Torvalds committed
1064 1065

	if (rw & WRITE)
Jens Axboe's avatar
Jens Axboe committed
1066
		rw = WRITE_ODIRECT;
Linus Torvalds's avatar
Linus Torvalds committed
1067

1068 1069 1070 1071
	/*
	 * Avoid references to bdev if not absolutely needed to give
	 * the early prefetch in the caller enough time.
	 */
Linus Torvalds's avatar
Linus Torvalds committed
1072 1073 1074

	if (offset & blocksize_mask) {
		if (bdev)
1075
			blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds's avatar
Linus Torvalds committed
1076 1077 1078 1079 1080 1081 1082 1083 1084 1085
		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;
1086 1087
		if (unlikely((addr & blocksize_mask) ||
			     (size & blocksize_mask))) {
Linus Torvalds's avatar
Linus Torvalds committed
1088
			if (bdev)
1089 1090
				blkbits = blksize_bits(
					 bdev_logical_block_size(bdev));
Linus Torvalds's avatar
Linus Torvalds committed
1091
			blocksize_mask = (1 << blkbits) - 1;
1092
			if ((addr & blocksize_mask) || (size & blocksize_mask))
Linus Torvalds's avatar
Linus Torvalds committed
1093 1094 1095 1096
				goto out;
		}
	}

1097 1098 1099 1100
	/* watch out for a 0 len io from a tricksy fs */
	if (rw == READ && end == offset)
		return 0;

1101
	dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds's avatar
Linus Torvalds committed
1102 1103 1104
	retval = -ENOMEM;
	if (!dio)
		goto out;
1105 1106 1107 1108 1109 1110
	/*
	 * 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));
Linus Torvalds's avatar
Linus Torvalds committed
1111

1112 1113
	dio->flags = flags;
	if (dio->flags & DIO_LOCKING) {
1114
		if (rw == READ) {
1115 1116
			struct address_space *mapping =
					iocb->ki_filp->f_mapping;
Linus Torvalds's avatar
Linus Torvalds committed
1117

1118 1119
			/* will be released by direct_io_worker */
			mutex_lock(&inode->i_mutex);
Linus Torvalds's avatar
Linus Torvalds committed
1120 1121 1122 1123

			retval = filemap_write_and_wait_range(mapping, offset,
							      end - 1);
			if (retval) {
1124
				mutex_unlock(&inode->i_mutex);
1125
				kmem_cache_free(dio_cache, dio);
Linus Torvalds's avatar
Linus Torvalds committed
1126 1127 1128 1129 1130
				goto out;
			}
		}
	}

1131 1132 1133 1134 1135
	/*
	 * Will be decremented at I/O completion time.
	 */
	atomic_inc(&inode->i_dio_count);

Linus Torvalds's avatar
Linus Torvalds committed
1136 1137 1138 1139 1140 1141
	/*
	 * 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.
	 */
Jens Axboe's avatar
Jens Axboe committed
1142
	dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
Linus Torvalds's avatar
Linus Torvalds committed
1143 1144
		(end > i_size_read(inode)));

1145 1146 1147 1148 1149
	retval = 0;

	dio->inode = inode;
	dio->rw = rw;
	sdio.blkbits = blkbits;
1150
	sdio.blkfactor = i_blkbits - blkbits;
1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178
	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);
	}

1179 1180
	blk_start_plug(&plug);

1181 1182 1183 1184 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
	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);

1239 1240
	blk_finish_plug(&plug);

1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263
	/*
	 * 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 &&
1264
	    ((rw == READ) || (dio->result == sdio.size)))
1265 1266 1267 1268 1269 1270 1271 1272 1273 1274
		retval = -EIOCBQUEUED;

	if (retval != -EIOCBQUEUED)
		dio_await_completion