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>
Linus Torvalds's avatar
Linus Torvalds committed
40 41 42

/*
 * How many user pages to map in one call to get_user_pages().  This determines
Andi Kleen's avatar
Andi Kleen committed
43
 * the size of a structure in the slab cache
Linus Torvalds's avatar
Linus Torvalds committed
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 {
Linus Torvalds's avatar
Linus Torvalds committed
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 */
Linus Torvalds's avatar
Linus Torvalds committed
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 */
Linus Torvalds's avatar
Linus Torvalds committed
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 */
Linus Torvalds's avatar
Linus Torvalds committed
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 */

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;
Linus Torvalds's avatar
Linus Torvalds committed
147 148 149 150

/*
 * How many pages are in the queue?
 */
151
static inline unsigned dio_pages_present(struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
152
{
153
	return sdio->tail - sdio->head;
Linus Torvalds's avatar
Linus Torvalds committed
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)
Linus Torvalds's avatar
Linus Torvalds committed
160 161 162 163
{
	int ret;
	int nr_pages;

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

171
	if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
Nick Piggin's avatar
Nick Piggin committed
172
		struct page *page = ZERO_PAGE(0);
Linus Torvalds's avatar
Linus Torvalds committed
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;
Nick Piggin's avatar
Nick Piggin committed
180 181
		page_cache_get(page);
		dio->pages[0] = page;
182 183
		sdio->head = 0;
		sdio->tail = 1;
Linus Torvalds's avatar
Linus Torvalds committed
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;
Linus Torvalds's avatar
Linus Torvalds committed
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)
Linus Torvalds's avatar
Linus Torvalds committed
207
{
208
	if (dio_pages_present(sdio) == 0) {
Linus Torvalds's avatar
Linus Torvalds committed
209 210
		int ret;

211
		ret = dio_refill_pages(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
212 213
		if (ret)
			return ERR_PTR(ret);
214
		BUG_ON(dio_pages_present(sdio) == 0);
Linus Torvalds's avatar
Linus Torvalds committed
215
	}
216
	return dio->pages[sdio->head++];
Linus Torvalds's avatar
Linus Torvalds committed
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.
Linus Torvalds's avatar
Linus Torvalds committed
231
 */
232
static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async)
Linus Torvalds's avatar
Linus Torvalds committed
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,
262
			    dio->private, ret, is_async);
263 264 265 266
	} else {
		if (is_async)
			aio_complete(dio->iocb, ret, 0);
		inode_dio_done(dio->inode);
267 268
	}

269
	return ret;
Linus Torvalds's avatar
Linus Torvalds committed
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)
Linus Torvalds's avatar
Linus Torvalds committed
277 278
{
	struct dio *dio = bio->bi_private;
279 280
	unsigned long remaining;
	unsigned long flags;
Linus Torvalds's avatar
Linus Torvalds committed
281 282 283

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

285 286 287
	spin_lock_irqsave(&dio->bio_lock, flags);
	remaining = --dio->refcount;
	if (remaining == 1 && dio->waiter)
288
		wake_up_process(dio->waiter);
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
	}
Linus Torvalds's avatar
Linus Torvalds committed
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)
Linus Torvalds's avatar
Linus Torvalds committed
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;
312
	if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds's avatar
Linus Torvalds committed
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)
Linus Torvalds's avatar
Linus Torvalds committed
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.
	 */
Linus Torvalds's avatar
Linus Torvalds committed
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;
Linus Torvalds's avatar
Linus Torvalds committed
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.
Linus Torvalds's avatar
Linus Torvalds committed
367
 */
368
static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
369
{
370
	struct bio *bio = sdio->bio;
371
	unsigned long flags;
Linus Torvalds's avatar
Linus Torvalds committed
372 373

	bio->bi_private = dio;
374 375 376 377 378

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

Linus Torvalds's avatar
Linus Torvalds committed
379 380
	if (dio->is_async && dio->rw == READ)
		bio_set_pages_dirty(bio);
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);
Linus Torvalds's avatar
Linus Torvalds committed
387

388 389 390
	sdio->bio = NULL;
	sdio->boundary = 0;
	sdio->logical_offset_in_bio = 0;
Linus Torvalds's avatar
Linus Torvalds committed
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)
Linus Torvalds's avatar
Linus Torvalds committed
397
{
398 399
	while (dio_pages_present(sdio))
		page_cache_release(dio_get_page(dio, sdio));
Linus Torvalds's avatar
Linus Torvalds committed
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.
Linus Torvalds's avatar
Linus Torvalds committed
407 408 409 410
 */
static struct bio *dio_await_one(struct dio *dio)
{
	unsigned long flags;
411
	struct bio *bio = NULL;
Linus Torvalds's avatar
Linus Torvalds committed
412 413

	spin_lock_irqsave(&dio->bio_lock, flags);
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;
Linus Torvalds's avatar
Linus Torvalds committed
429
	}
430 431 432 433
	if (dio->bio_list) {
		bio = dio->bio_list;
		dio->bio_list = bio->bi_private;
	}
Linus Torvalds's avatar
Linus Torvalds committed
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;
Linus Torvalds's avatar
Linus Torvalds committed
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().
Linus Torvalds's avatar
Linus Torvalds committed
471
 */
472
static void dio_await_completion(struct dio *dio)
Linus Torvalds's avatar
Linus Torvalds committed
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);
Linus Torvalds's avatar
Linus Torvalds committed
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)
Linus Torvalds's avatar
Linus Torvalds committed
490 491 492
{
	int ret = 0;

493
	if (sdio->reap_counter++ >= 64) {
Linus Torvalds's avatar
Linus Torvalds committed
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;
Linus Torvalds's avatar
Linus Torvalds committed
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
Linus Torvalds's avatar
Linus Torvalds committed
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
Linus Torvalds's avatar
Linus Torvalds committed
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.
Linus Torvalds's avatar
Linus Torvalds committed
534
 */
535 536
static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
			   struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
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 */
Linus Torvalds's avatar
Linus Torvalds committed
541 542 543 544 545 546 547 548 549
	unsigned long fs_count;	/* Number of filesystem-sized blocks */
	int create;

	/*
	 * 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) {
550 551
		BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
		fs_startblk = sdio->block_in_file >> sdio->blkfactor;
552 553 554
		fs_endblk = (sdio->final_block_in_request - 1) >>
					sdio->blkfactor;
		fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds's avatar
Linus Torvalds committed
555

556 557 558
		map_bh->b_state = 0;
		map_bh->b_size = fs_count << dio->inode->i_blkbits;

559 560 561 562 563 564 565 566 567 568 569
		/*
		 * 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
570
		create = dio->rw & WRITE;
571
		if (dio->flags & DIO_SKIP_HOLES) {
572 573
			if (sdio->block_in_file < (i_size_read(dio->inode) >>
							sdio->blkbits))
Linus Torvalds's avatar
Linus Torvalds committed
574 575
				create = 0;
		}
576

577
		ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds's avatar
Linus Torvalds committed
578
						map_bh, create);
579 580 581

		/* Store for completion */
		dio->private = map_bh->b_private;
Linus Torvalds's avatar
Linus Torvalds committed
582 583 584 585 586 587 588
	}
	return ret;
}

/*
 * There is no bio.  Make one now.
 */
589 590
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
591 592 593 594
{
	sector_t sector;
	int ret, nr_pages;

595
	ret = dio_bio_reap(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
596 597
	if (ret)
		goto out;
598
	sector = start_sector << (sdio->blkbits - 9);
599
	nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
600
	nr_pages = min(nr_pages, BIO_MAX_PAGES);
Linus Torvalds's avatar
Linus Torvalds committed
601
	BUG_ON(nr_pages <= 0);
602
	dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
603
	sdio->boundary = 0;
Linus Torvalds's avatar
Linus Torvalds committed
604 605 606 607 608 609 610 611 612 613 614
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.
 */
615
static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
616 617 618
{
	int ret;

619 620 621
	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
622 623 624
		/*
		 * Decrement count only, if we are done with this page
		 */
625 626 627 628 629
		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
630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646
		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.
 */
647 648
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
649 650 651
{
	int ret = 0;

652 653 654 655
	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;
656

Linus Torvalds's avatar
Linus Torvalds committed
657
		/*
658 659
		 * See whether this new request is contiguous with the old.
		 *
Namhyung Kim's avatar
Namhyung Kim committed
660 661
		 * Btrfs cannot handle having logically non-contiguous requests
		 * submitted.  For example if you have
662 663
		 *
		 * Logical:  [0-4095][HOLE][8192-12287]
Namhyung Kim's avatar
Namhyung Kim committed
664
		 * Physical: [0-4095]      [4096-8191]
665 666 667 668 669
		 *
		 * 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
670
		 */
671
		if (sdio->final_block_in_bio != sdio->cur_page_block ||
672
		    cur_offset != bio_next_offset)
673
			dio_bio_submit(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
674 675 676 677
		/*
		 * Submit now if the underlying fs is about to perform a
		 * metadata read
		 */
678 679
		else if (sdio->boundary)
			dio_bio_submit(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
680 681
	}

682
	if (sdio->bio == NULL) {
683
		ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
684 685 686 687
		if (ret)
			goto out;
	}

688 689
	if (dio_bio_add_page(sdio) != 0) {
		dio_bio_submit(dio, sdio);
690
		ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
691
		if (ret == 0) {
692
			ret = dio_bio_add_page(sdio);
Linus Torvalds's avatar
Linus Torvalds committed
693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716
			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.
 */
717
static inline int
718
submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
719 720
		    unsigned offset, unsigned len, sector_t blocknr,
		    struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
721 722 723
{
	int ret = 0;

724 725 726 727 728 729 730
	if (dio->rw & WRITE) {
		/*
		 * Read accounting is performed in submit_bio()
		 */
		task_io_account_write(len);
	}

Linus Torvalds's avatar
Linus Torvalds committed
731 732 733
	/*
	 * Can we just grow the current page's presence in the dio?
	 */
734 735 736 737 738
	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
739 740

		/*
741
		 * If sdio->boundary then we want to schedule the IO now to
Linus Torvalds's avatar
Linus Torvalds committed
742 743
		 * avoid metadata seeks.
		 */
744
		if (sdio->boundary) {
745
			ret = dio_send_cur_page(dio, sdio, map_bh);
746 747
			page_cache_release(sdio->cur_page);
			sdio->cur_page = NULL;
Linus Torvalds's avatar
Linus Torvalds committed
748 749 750 751 752 753 754
		}
		goto out;
	}

	/*
	 * If there's a deferred page already there then send it.
	 */
755
	if (sdio->cur_page) {
756
		ret = dio_send_cur_page(dio, sdio, map_bh);
757 758
		page_cache_release(sdio->cur_page);
		sdio->cur_page = NULL;
Linus Torvalds's avatar
Linus Torvalds committed
759 760 761 762 763
		if (ret)
			goto out;
	}

	page_cache_get(page);		/* It is in dio */
764 765 766 767 768
	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
769 770 771 772 773 774 775 776 777
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
 */
778
static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
779 780 781 782
{
	unsigned i;
	unsigned nblocks;

783
	nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds's avatar
Linus Torvalds committed
784 785

	for (i = 0; i < nblocks; i++) {
786 787
		unmap_underlying_metadata(map_bh->b_bdev,
					  map_bh->b_blocknr + i);
Linus Torvalds's avatar
Linus Torvalds committed
788 789 790 791 792 793 794 795 796 797 798 799
	}
}

/*
 * 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.
 */
800 801
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
802 803 804 805 806 807
{
	unsigned dio_blocks_per_fs_block;
	unsigned this_chunk_blocks;	/* In dio_blocks */
	unsigned this_chunk_bytes;
	struct page *page;

808
	sdio->start_zero_done = 1;
809
	if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds's avatar
Linus Torvalds committed
810 811
		return;

812 813
	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
814 815 816 817 818 819 820 821 822 823 824

	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;

825
	this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds's avatar
Linus Torvalds committed
826

Nick Piggin's avatar
Nick Piggin committed
827
	page = ZERO_PAGE(0);
828
	if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
829
				sdio->next_block_for_io, map_bh))
Linus Torvalds's avatar
Linus Torvalds committed
830 831
		return;

832
	sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds's avatar
Linus Torvalds committed
833 834 835 836 837 838 839 840 841 842 843
}

/*
 * 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.
 *
844
 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds's avatar
Linus Torvalds committed
845 846 847
 * 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
848
 * it should set b_size to PAGE_SIZE or more inside get_block().  This gives
Linus Torvalds's avatar
Linus Torvalds committed
849 850
 * fine alignment but still allows this function to work in PAGE_SIZE units.
 */
851 852
static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
			struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
853
{
854
	const unsigned blkbits = sdio->blkbits;
Linus Torvalds's avatar
Linus Torvalds committed
855 856 857 858 859 860
	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 */
861
	block_in_page = sdio->first_block_in_page;
Linus Torvalds's avatar
Linus Torvalds committed
862

863 864
	while (sdio->block_in_file < sdio->final_block_in_request) {
		page = dio_get_page(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
865 866 867 868 869 870 871 872 873 874 875
		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;

876
			if (sdio->blocks_available == 0) {
Linus Torvalds's avatar
Linus Torvalds committed
877 878 879 880 881 882
				/*
				 * Need to go and map some more disk
				 */
				unsigned long blkmask;
				unsigned long dio_remainder;

883
				ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
884 885 886 887 888 889 890
				if (ret) {
					page_cache_release(page);
					goto out;
				}
				if (!buffer_mapped(map_bh))
					goto do_holes;

891 892 893 894
				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
895
				if (buffer_new(map_bh))
896
					clean_blockdev_aliases(dio, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
897

898
				if (!sdio->blkfactor)
Linus Torvalds's avatar
Linus Torvalds committed
899 900
					goto do_holes;

901 902
				blkmask = (1 << sdio->blkfactor) - 1;
				dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds's avatar
Linus Torvalds committed
903 904 905 906 907 908 909 910 911 912 913 914 915

				/*
				 * 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))
916 917
					sdio->next_block_for_io += dio_remainder;
				sdio->blocks_available -= dio_remainder;
Linus Torvalds's avatar
Linus Torvalds committed
918 919 920 921
			}
do_holes:
			/* Handle holes */
			if (!buffer_mapped(map_bh)) {
922
				loff_t i_size_aligned;
Linus Torvalds's avatar
Linus Torvalds committed
923 924

				/* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboe's avatar
Jens Axboe committed
925
				if (dio->rw & WRITE) {
Linus Torvalds's avatar
Linus Torvalds committed
926 927 928 929
					page_cache_release(page);
					return -ENOTBLK;
				}

930 931 932 933 934 935
				/*
				 * 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);
936
				if (sdio->block_in_file >=
937
						i_size_aligned >> blkbits) {
Linus Torvalds's avatar
Linus Torvalds committed
938 939 940 941
					/* We hit eof */
					page_cache_release(page);
					goto out;
				}
942 943
				zero_user(page, block_in_page << blkbits,
						1 << blkbits);
944
				sdio->block_in_file++;
Linus Torvalds's avatar
Linus Torvalds committed
945 946 947 948 949 950 951 952 953
				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.
			 */
954
			if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
955
				dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
956 957 958 959 960

			/*
			 * Work out, in this_chunk_blocks, how much disk we
			 * can add to this page
			 */
961
			this_chunk_blocks = sdio->blocks_available;
Linus Torvalds's avatar
Linus Torvalds committed
962 963 964
			u = (PAGE_SIZE - offset_in_page) >> blkbits;
			if (this_chunk_blocks > u)
				this_chunk_blocks = u;
965
			u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds's avatar
Linus Torvalds committed
966 967 968 969 970
			if (this_chunk_blocks > u)
				this_chunk_blocks = u;
			this_chunk_bytes = this_chunk_blocks << blkbits;
			BUG_ON(this_chunk_bytes == 0);

971 972 973 974
			sdio->boundary = buffer_boundary(map_bh);
			ret = submit_page_section(dio, sdio, page,
						  offset_in_page,
						  this_chunk_bytes,
975 976
						  sdio->next_block_for_io,
						  map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
977 978 979 980
			if (ret) {
				page_cache_release(page);
				goto out;
			}
981
			sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds's avatar
Linus Torvalds committed
982

983
			sdio->block_in_file += this_chunk_blocks;
Linus Torvalds's avatar
Linus Torvalds committed
984
			block_in_page += this_chunk_blocks;
985
			sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds's avatar
Linus Torvalds committed
986
next_block:
987 988
			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
989 990 991 992 993 994 995 996 997 998 999
				break;
		}

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

1000
static inline int drop_refcount(struct dio *dio)
Linus Torvalds's avatar
Linus Torvalds committed
1001
{
1002
	int ret2;
1003
	unsigned long flags;
Linus Torvalds's avatar
Linus Torvalds committed
1004

1005 1006
	/*
	 * Sync will always be dropping the final ref and completing the
1007 1008 1009 1010 1011 1012 1013 1014
	 * 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.
1015
	 */
1016 1017 1018
	spin_lock_irqsave(&dio->bio_lock, flags);
	ret2 = --dio->refcount;
	spin_unlock_irqrestore(&dio->bio_lock, flags);
1019
	return ret2;
Linus Torvalds's avatar
Linus Torvalds committed
1020 1021
}

1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033
/*
 * 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.
1034 1035 1036 1037 1038 1039 1040
 *
 * 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.
1041 1042 1043 1044 1045
 *
 * 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.
1046
 */
1047 1048
static inline ssize_t
do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
Linus Torvalds's avatar
Linus Torvalds committed
1049
	struct block_device *bdev, const struct iovec *iov, loff_t offset, 
1050
	unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1051
	dio_submit_t submit_io,	int flags)
Linus Torvalds's avatar
Linus Torvalds committed
1052 1053 1054 1055 1056 1057 1058 1059 1060
{
	int seg;
	size_t size;
	unsigned long addr;
	unsigned blkbits = inode->i_blkbits;
	unsigned blocksize_mask = (1 << blkbits) - 1;
	ssize_t retval = -EINVAL;
	loff_t end = offset;
	struct dio *dio;
1061
	struct dio_submit sdio = { 0, };
1062 1063 1064
	unsigned long user_addr;
	size_t bytes;
	struct buffer_head map_bh = { 0, };
1065
	struct blk_plug plug;
Linus Torvalds's avatar
Linus Torvalds committed
1066 1067

	if (rw & WRITE)
Jens Axboe's avatar
Jens Axboe committed
1068
		rw = WRITE_ODIRECT;
Linus Torvalds's avatar
Linus Torvalds committed
1069

1070 1071 1072 1073
	/*
	 * 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
1074 1075 1076

	if (offset & blocksize_mask) {
		if (bdev)
1077
			blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds's avatar
Linus Torvalds committed
1078 1079 1080 1081 1082 1083 1084 1085 1086 1087
		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;
1088 1089
		if (unlikely((addr & blocksize_mask) ||
			     (size & blocksize_mask))) {
Linus Torvalds's avatar
Linus Torvalds committed
1090
			if (bdev)
1091 1092
				blkbits = blksize_bits(
					 bdev_logical_block_size(bdev));
Linus Torvalds's avatar
Linus Torvalds committed
1093
			blocksize_mask = (1 << blkbits) - 1;
1094
			if ((addr & blocksize_mask) || (size & blocksize_mask))
Linus Torvalds's avatar
Linus Torvalds committed
1095 1096 1097 1098
				goto out;
		}
	}

1099 1100 1101 1102
	/* watch out for a 0 len io from a tricksy fs */
	if (rw == READ && end == offset)
		return 0;

1103
	dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds's avatar
Linus Torvalds committed
1104 1105 1106
	retval = -ENOMEM;
	if (!dio)
		goto out;
1107 1108 1109 1110 1111 1112
	/*
	 * 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
1113

1114 1115
	dio->flags = flags;
	if (dio->flags & DIO_LOCKING) {
1116
		if (rw == READ) {
1117 1118
			struct address_space *mapping =
					iocb->ki_filp->f_mapping;
Linus Torvalds's avatar
Linus Torvalds committed
1119

1120 1121
			/* will be released by direct_io_worker */
			mutex_lock(&inode->i_mutex);
Linus Torvalds's avatar
Linus Torvalds committed
1122 1123 1124 1125

			retval = filemap_write_and_wait_range(mapping, offset,
							      end - 1);
			if (retval) {
1126
				mutex_unlock(&inode->i_mutex);
1127
				kmem_cache_free(dio_cache, dio);
Linus Torvalds's avatar
Linus Torvalds committed
1128 1129 1130 1131 1132
				goto out;
			}
		}
	}

1133 1134 1135 1136 1137
	/*
	 * Will be decremented at I/O completion time.
	 */
	atomic_inc(&inode->i_dio_count);

Linus Torvalds's avatar
Linus Torvalds committed
1138 1139 1140 1141 1142 1143
	/*
	 * 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
1144
	dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
Linus Torvalds's avatar
Linus Torvalds committed
1145 1146
		(end > i_size_read(inode)));

1147 1148 1149 1150 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 1179 1180
	retval = 0;

	dio->inode = inode;
	dio->rw = rw;
	sdio.blkbits = blkbits;
	sdio.blkfactor = inode->i_blkbits - blkbits;
	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);
	}

1181 1182
	blk_start_plug(&plug);

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 1239 1240
	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);

1241 1242
	blk_finish_plug(&plug);

1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265
	/*
	 * 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 &&
1266
	    ((rw == READ) || (dio->result == sdio.size)))
1267 1268 1269 1270 1271 1272 1273 1274 1275 1276
		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);
Linus Torvalds's avatar
Linus Torvalds committed
1277

1278 1279 1280
out:
	return retval;
}
1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304

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);