direct-io.c 38 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>
Linus Torvalds's avatar
Linus Torvalds committed
39 40 41

/*
 * How many user pages to map in one call to get_user_pages().  This determines
Andi Kleen's avatar
Andi Kleen committed
42
 * the size of a structure in the slab cache
Linus Torvalds's avatar
Linus Torvalds committed
43 44 45 46 47 48 49 50 51 52 53 54 55 56 57
 */
#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.
 */

58 59 60
/* dio_state only used in the submission path */

struct dio_submit {
Linus Torvalds's avatar
Linus Torvalds committed
61 62 63 64 65 66 67 68 69 70 71 72 73 74 75
	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 */
76
	int reap_counter;		/* rate limit reaping */
Linus Torvalds's avatar
Linus Torvalds committed
77 78 79
	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 */
80
	get_block_t *get_block;		/* block mapping function */
81
	dio_submit_t *submit_io;	/* IO submition function */
82

83
	loff_t logical_offset_in_bio;	/* current first logical block in bio */
Linus Torvalds's avatar
Linus Torvalds committed
84 85 86 87 88 89 90 91 92 93 94 95 96
	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 */
97
	loff_t cur_page_fs_offset;	/* Offset in file */
Linus Torvalds's avatar
Linus Torvalds committed
98 99 100 101 102 103 104 105 106 107 108 109 110 111

	/*
	 * 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 */
112 113 114 115 116 117
};

/* dio_state communicated between submission path and end_io */
struct dio {
	int flags;			/* doesn't change */
	int rw;
118
	struct inode *inode;
119 120 121
	loff_t i_size;			/* i_size when submitted */
	dio_iodone_t *end_io;		/* IO completion function */

122
	void *private;			/* copy from map_bh.b_private */
123 124 125

	/* BIO completion state */
	spinlock_t bio_lock;		/* protects BIO fields below */
126 127 128
	int page_errors;		/* errno from get_user_pages() */
	int is_async;			/* is IO async ? */
	int io_error;			/* IO error in completion path */
129 130 131 132 133 134 135 136
	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 */

137 138 139 140 141 142
	/*
	 * 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 */
143 144 145
} ____cacheline_aligned_in_smp;

static struct kmem_cache *dio_cache __read_mostly;
Linus Torvalds's avatar
Linus Torvalds committed
146

Christoph Hellwig's avatar
Christoph Hellwig committed
147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190
static void __inode_dio_wait(struct inode *inode)
{
	wait_queue_head_t *wq = bit_waitqueue(&inode->i_state, __I_DIO_WAKEUP);
	DEFINE_WAIT_BIT(q, &inode->i_state, __I_DIO_WAKEUP);

	do {
		prepare_to_wait(wq, &q.wait, TASK_UNINTERRUPTIBLE);
		if (atomic_read(&inode->i_dio_count))
			schedule();
	} while (atomic_read(&inode->i_dio_count));
	finish_wait(wq, &q.wait);
}

/**
 * inode_dio_wait - wait for outstanding DIO requests to finish
 * @inode: inode to wait for
 *
 * Waits for all pending direct I/O requests to finish so that we can
 * proceed with a truncate or equivalent operation.
 *
 * Must be called under a lock that serializes taking new references
 * to i_dio_count, usually by inode->i_mutex.
 */
void inode_dio_wait(struct inode *inode)
{
	if (atomic_read(&inode->i_dio_count))
		__inode_dio_wait(inode);
}
EXPORT_SYMBOL_GPL(inode_dio_wait);

/*
 * inode_dio_done - signal finish of a direct I/O requests
 * @inode: inode the direct I/O happens on
 *
 * This is called once we've finished processing a direct I/O request,
 * and is used to wake up callers waiting for direct I/O to be quiesced.
 */
void inode_dio_done(struct inode *inode)
{
	if (atomic_dec_and_test(&inode->i_dio_count))
		wake_up_bit(&inode->i_state, __I_DIO_WAKEUP);
}
EXPORT_SYMBOL_GPL(inode_dio_done);

Linus Torvalds's avatar
Linus Torvalds committed
191 192 193
/*
 * How many pages are in the queue?
 */
194
static inline unsigned dio_pages_present(struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
195
{
196
	return sdio->tail - sdio->head;
Linus Torvalds's avatar
Linus Torvalds committed
197 198 199 200 201
}

/*
 * Go grab and pin some userspace pages.   Typically we'll get 64 at a time.
 */
202
static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
203 204 205 206
{
	int ret;
	int nr_pages;

207
	nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
Nick Piggin's avatar
Nick Piggin committed
208
	ret = get_user_pages_fast(
209
		sdio->curr_user_address,		/* Where from? */
Linus Torvalds's avatar
Linus Torvalds committed
210 211
		nr_pages,			/* How many pages? */
		dio->rw == READ,		/* Write to memory? */
Nick Piggin's avatar
Nick Piggin committed
212
		&dio->pages[0]);		/* Put results here */
Linus Torvalds's avatar
Linus Torvalds committed
213

214
	if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
Nick Piggin's avatar
Nick Piggin committed
215
		struct page *page = ZERO_PAGE(0);
Linus Torvalds's avatar
Linus Torvalds committed
216 217 218 219 220 221 222
		/*
		 * 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
223 224
		page_cache_get(page);
		dio->pages[0] = page;
225 226
		sdio->head = 0;
		sdio->tail = 1;
Linus Torvalds's avatar
Linus Torvalds committed
227 228 229 230 231
		ret = 0;
		goto out;
	}

	if (ret >= 0) {
232 233 234 235
		sdio->curr_user_address += ret * PAGE_SIZE;
		sdio->curr_page += ret;
		sdio->head = 0;
		sdio->tail = ret;
Linus Torvalds's avatar
Linus Torvalds committed
236 237 238 239 240 241 242 243 244 245 246 247
		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.
 */
248 249
static inline struct page *dio_get_page(struct dio *dio,
		struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
250
{
251
	if (dio_pages_present(sdio) == 0) {
Linus Torvalds's avatar
Linus Torvalds committed
252 253
		int ret;

254
		ret = dio_refill_pages(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
255 256
		if (ret)
			return ERR_PTR(ret);
257
		BUG_ON(dio_pages_present(sdio) == 0);
Linus Torvalds's avatar
Linus Torvalds committed
258
	}
259
	return dio->pages[sdio->head++];
Linus Torvalds's avatar
Linus Torvalds committed
260 261
}

262 263 264 265 266 267 268 269 270 271 272 273
/**
 * 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
274
 */
275
static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async)
Linus Torvalds's avatar
Linus Torvalds committed
276
{
277 278
	ssize_t transferred = 0;

279 280 281 282 283 284 285 286 287
	/*
	 * 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;

288 289 290 291 292 293 294 295 296 297 298 299 300 301 302
	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;

303 304
	if (dio->end_io && dio->result) {
		dio->end_io(dio->iocb, offset, transferred,
305
			    dio->private, ret, is_async);
306 307 308 309
	} else {
		if (is_async)
			aio_complete(dio->iocb, ret, 0);
		inode_dio_done(dio->inode);
310 311
	}

312
	return ret;
Linus Torvalds's avatar
Linus Torvalds committed
313 314 315 316 317 318
}

static int dio_bio_complete(struct dio *dio, struct bio *bio);
/*
 * Asynchronous IO callback. 
 */
319
static void dio_bio_end_aio(struct bio *bio, int error)
Linus Torvalds's avatar
Linus Torvalds committed
320 321
{
	struct dio *dio = bio->bi_private;
322 323
	unsigned long remaining;
	unsigned long flags;
Linus Torvalds's avatar
Linus Torvalds committed
324 325 326

	/* cleanup the bio */
	dio_bio_complete(dio, bio);
327

328 329 330
	spin_lock_irqsave(&dio->bio_lock, flags);
	remaining = --dio->refcount;
	if (remaining == 1 && dio->waiter)
331
		wake_up_process(dio->waiter);
332
	spin_unlock_irqrestore(&dio->bio_lock, flags);
333

334
	if (remaining == 0) {
335
		dio_complete(dio, dio->iocb->ki_pos, 0, true);
336
		kmem_cache_free(dio_cache, dio);
337
	}
Linus Torvalds's avatar
Linus Torvalds committed
338 339 340 341 342 343 344 345 346
}

/*
 * 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.
 */
347
static void dio_bio_end_io(struct bio *bio, int error)
Linus Torvalds's avatar
Linus Torvalds committed
348 349 350 351 352 353 354
{
	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;
355
	if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds's avatar
Linus Torvalds committed
356 357 358 359
		wake_up_process(dio->waiter);
	spin_unlock_irqrestore(&dio->bio_lock, flags);
}

360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379
/**
 * 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);

380
static inline void
381 382 383
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
384 385 386
{
	struct bio *bio;

387 388 389 390
	/*
	 * 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
391 392 393 394 395 396 397 398 399
	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;

400 401
	sdio->bio = bio;
	sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds's avatar
Linus Torvalds committed
402 403 404 405 406 407
}

/*
 * 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().
408 409
 *
 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds's avatar
Linus Torvalds committed
410
 */
411
static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
412
{
413
	struct bio *bio = sdio->bio;
414
	unsigned long flags;
Linus Torvalds's avatar
Linus Torvalds committed
415 416

	bio->bi_private = dio;
417 418 419 420 421

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

Linus Torvalds's avatar
Linus Torvalds committed
422 423
	if (dio->is_async && dio->rw == READ)
		bio_set_pages_dirty(bio);
424

425 426 427
	if (sdio->submit_io)
		sdio->submit_io(dio->rw, bio, dio->inode,
			       sdio->logical_offset_in_bio);
428 429
	else
		submit_bio(dio->rw, bio);
Linus Torvalds's avatar
Linus Torvalds committed
430

431 432 433
	sdio->bio = NULL;
	sdio->boundary = 0;
	sdio->logical_offset_in_bio = 0;
Linus Torvalds's avatar
Linus Torvalds committed
434 435 436 437 438
}

/*
 * Release any resources in case of a failure
 */
439
static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
440
{
441 442
	while (dio_pages_present(sdio))
		page_cache_release(dio_get_page(dio, sdio));
Linus Torvalds's avatar
Linus Torvalds committed
443 444 445
}

/*
446 447 448 449
 * 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
450 451 452 453
 */
static struct bio *dio_await_one(struct dio *dio)
{
	unsigned long flags;
454
	struct bio *bio = NULL;
Linus Torvalds's avatar
Linus Torvalds committed
455 456

	spin_lock_irqsave(&dio->bio_lock, flags);
457 458 459 460 461 462 463 464 465 466 467 468 469 470 471

	/*
	 * 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
472
	}
473 474 475 476
	if (dio->bio_list) {
		bio = dio->bio_list;
		dio->bio_list = bio->bi_private;
	}
Linus Torvalds's avatar
Linus Torvalds committed
477 478 479 480 481 482 483 484 485 486 487 488 489 490
	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)
491
		dio->io_error = -EIO;
Linus Torvalds's avatar
Linus Torvalds committed
492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508

	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;
}

/*
509 510 511
 * 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
512
 * errors are propagated through dio->io_error and should be propagated via
513
 * dio_complete().
Linus Torvalds's avatar
Linus Torvalds committed
514
 */
515
static void dio_await_completion(struct dio *dio)
Linus Torvalds's avatar
Linus Torvalds committed
516
{
517 518 519 520 521 522
	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
523 524 525 526 527 528 529 530 531
}

/*
 * 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.
 */
532
static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
533 534 535
{
	int ret = 0;

536
	if (sdio->reap_counter++ >= 64) {
Linus Torvalds's avatar
Linus Torvalds committed
537 538 539 540 541 542 543 544 545 546 547 548 549
		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;
		}
550
		sdio->reap_counter = 0;
Linus Torvalds's avatar
Linus Torvalds committed
551 552 553 554 555 556
	}
	return ret;
}

/*
 * Call into the fs to map some more disk blocks.  We record the current number
557
 * of available blocks at sdio->blocks_available.  These are in units of the
Linus Torvalds's avatar
Linus Torvalds committed
558 559 560 561 562
 * 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.
 *
563
 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds's avatar
Linus Torvalds committed
564 565 566 567 568 569 570 571 572 573 574 575
 * 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
576
 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds's avatar
Linus Torvalds committed
577
 */
578 579
static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
			   struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
580 581 582 583 584 585 586 587 588 589 590 591 592 593
{
	int ret;
	sector_t fs_startblk;	/* Into file, in filesystem-sized blocks */
	unsigned long fs_count;	/* Number of filesystem-sized blocks */
	unsigned long dio_count;/* Number of dio_block-sized blocks */
	unsigned long blkmask;
	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) {
594 595 596 597 598
		BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
		fs_startblk = sdio->block_in_file >> sdio->blkfactor;
		dio_count = sdio->final_block_in_request - sdio->block_in_file;
		fs_count = dio_count >> sdio->blkfactor;
		blkmask = (1 << sdio->blkfactor) - 1;
Linus Torvalds's avatar
Linus Torvalds committed
599 600 601
		if (dio_count & blkmask)	
			fs_count++;

602 603 604
		map_bh->b_state = 0;
		map_bh->b_size = fs_count << dio->inode->i_blkbits;

605 606 607 608 609 610 611 612 613 614 615
		/*
		 * 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
616
		create = dio->rw & WRITE;
617
		if (dio->flags & DIO_SKIP_HOLES) {
618 619
			if (sdio->block_in_file < (i_size_read(dio->inode) >>
							sdio->blkbits))
Linus Torvalds's avatar
Linus Torvalds committed
620 621
				create = 0;
		}
622

623
		ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds's avatar
Linus Torvalds committed
624
						map_bh, create);
625 626 627

		/* Store for completion */
		dio->private = map_bh->b_private;
Linus Torvalds's avatar
Linus Torvalds committed
628 629 630 631 632 633 634
	}
	return ret;
}

/*
 * There is no bio.  Make one now.
 */
635 636
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
637 638 639 640
{
	sector_t sector;
	int ret, nr_pages;

641
	ret = dio_bio_reap(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
642 643
	if (ret)
		goto out;
644
	sector = start_sector << (sdio->blkbits - 9);
645
	nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
646
	nr_pages = min(nr_pages, BIO_MAX_PAGES);
Linus Torvalds's avatar
Linus Torvalds committed
647
	BUG_ON(nr_pages <= 0);
648
	dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
649
	sdio->boundary = 0;
Linus Torvalds's avatar
Linus Torvalds committed
650 651 652 653 654 655 656 657 658 659 660
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.
 */
661
static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
662 663 664
{
	int ret;

665 666 667
	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
668 669 670
		/*
		 * Decrement count only, if we are done with this page
		 */
671 672 673 674 675
		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
676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692
		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.
 */
693 694
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
695 696 697
{
	int ret = 0;

698 699 700 701
	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;
702

Linus Torvalds's avatar
Linus Torvalds committed
703
		/*
704 705
		 * See whether this new request is contiguous with the old.
		 *
Namhyung Kim's avatar
Namhyung Kim committed
706 707
		 * Btrfs cannot handle having logically non-contiguous requests
		 * submitted.  For example if you have
708 709
		 *
		 * Logical:  [0-4095][HOLE][8192-12287]
Namhyung Kim's avatar
Namhyung Kim committed
710
		 * Physical: [0-4095]      [4096-8191]
711 712 713 714 715
		 *
		 * 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
716
		 */
717
		if (sdio->final_block_in_bio != sdio->cur_page_block ||
718
		    cur_offset != bio_next_offset)
719
			dio_bio_submit(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
720 721 722 723
		/*
		 * Submit now if the underlying fs is about to perform a
		 * metadata read
		 */
724 725
		else if (sdio->boundary)
			dio_bio_submit(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
726 727
	}

728
	if (sdio->bio == NULL) {
729
		ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
730 731 732 733
		if (ret)
			goto out;
	}

734 735
	if (dio_bio_add_page(sdio) != 0) {
		dio_bio_submit(dio, sdio);
736
		ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
737
		if (ret == 0) {
738
			ret = dio_bio_add_page(sdio);
Linus Torvalds's avatar
Linus Torvalds committed
739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762
			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.
 */
763
static inline int
764
submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
765 766
		    unsigned offset, unsigned len, sector_t blocknr,
		    struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
767 768 769
{
	int ret = 0;

770 771 772 773 774 775 776
	if (dio->rw & WRITE) {
		/*
		 * Read accounting is performed in submit_bio()
		 */
		task_io_account_write(len);
	}

Linus Torvalds's avatar
Linus Torvalds committed
777 778 779
	/*
	 * Can we just grow the current page's presence in the dio?
	 */
780 781 782 783 784
	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
785 786

		/*
787
		 * If sdio->boundary then we want to schedule the IO now to
Linus Torvalds's avatar
Linus Torvalds committed
788 789
		 * avoid metadata seeks.
		 */
790
		if (sdio->boundary) {
791
			ret = dio_send_cur_page(dio, sdio, map_bh);
792 793
			page_cache_release(sdio->cur_page);
			sdio->cur_page = NULL;
Linus Torvalds's avatar
Linus Torvalds committed
794 795 796 797 798 799 800
		}
		goto out;
	}

	/*
	 * If there's a deferred page already there then send it.
	 */
801
	if (sdio->cur_page) {
802
		ret = dio_send_cur_page(dio, sdio, map_bh);
803 804
		page_cache_release(sdio->cur_page);
		sdio->cur_page = NULL;
Linus Torvalds's avatar
Linus Torvalds committed
805 806 807 808 809
		if (ret)
			goto out;
	}

	page_cache_get(page);		/* It is in dio */
810 811 812 813 814
	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
815 816 817 818 819 820 821 822 823
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
 */
824
static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
825 826 827 828
{
	unsigned i;
	unsigned nblocks;

829
	nblocks = map_bh->b_size >> dio->inode->i_blkbits;
Linus Torvalds's avatar
Linus Torvalds committed
830 831

	for (i = 0; i < nblocks; i++) {
832 833
		unmap_underlying_metadata(map_bh->b_bdev,
					  map_bh->b_blocknr + i);
Linus Torvalds's avatar
Linus Torvalds committed
834 835 836 837 838 839 840 841 842 843 844 845
	}
}

/*
 * 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.
 */
846 847
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
848 849 850 851 852 853
{
	unsigned dio_blocks_per_fs_block;
	unsigned this_chunk_blocks;	/* In dio_blocks */
	unsigned this_chunk_bytes;
	struct page *page;

854
	sdio->start_zero_done = 1;
855
	if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds's avatar
Linus Torvalds committed
856 857
		return;

858 859
	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
860 861 862 863 864 865 866 867 868 869 870

	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;

871
	this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds's avatar
Linus Torvalds committed
872

Nick Piggin's avatar
Nick Piggin committed
873
	page = ZERO_PAGE(0);
874
	if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
875
				sdio->next_block_for_io, map_bh))
Linus Torvalds's avatar
Linus Torvalds committed
876 877
		return;

878
	sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds's avatar
Linus Torvalds committed
879 880 881 882 883 884 885 886 887 888 889
}

/*
 * 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.
 *
890
 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds's avatar
Linus Torvalds committed
891 892 893
 * 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
894
 * it should set b_size to PAGE_SIZE or more inside get_block().  This gives
Linus Torvalds's avatar
Linus Torvalds committed
895 896
 * fine alignment but still allows this function to work in PAGE_SIZE units.
 */
897 898
static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
			struct buffer_head *map_bh)
Linus Torvalds's avatar
Linus Torvalds committed
899
{
900
	const unsigned blkbits = sdio->blkbits;
Linus Torvalds's avatar
Linus Torvalds committed
901 902 903 904 905 906
	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 */
907
	block_in_page = sdio->first_block_in_page;
Linus Torvalds's avatar
Linus Torvalds committed
908

909 910
	while (sdio->block_in_file < sdio->final_block_in_request) {
		page = dio_get_page(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
911 912 913 914 915 916 917 918 919 920 921
		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;

922
			if (sdio->blocks_available == 0) {
Linus Torvalds's avatar
Linus Torvalds committed
923 924 925 926 927 928
				/*
				 * Need to go and map some more disk
				 */
				unsigned long blkmask;
				unsigned long dio_remainder;

929
				ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
930 931 932 933 934 935 936
				if (ret) {
					page_cache_release(page);
					goto out;
				}
				if (!buffer_mapped(map_bh))
					goto do_holes;

937 938 939 940
				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
941
				if (buffer_new(map_bh))
942
					clean_blockdev_aliases(dio, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
943

944
				if (!sdio->blkfactor)
Linus Torvalds's avatar
Linus Torvalds committed
945 946
					goto do_holes;

947 948
				blkmask = (1 << sdio->blkfactor) - 1;
				dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds's avatar
Linus Torvalds committed
949 950 951 952 953 954 955 956 957 958 959 960 961

				/*
				 * 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))
962 963
					sdio->next_block_for_io += dio_remainder;
				sdio->blocks_available -= dio_remainder;
Linus Torvalds's avatar
Linus Torvalds committed
964 965 966 967
			}
do_holes:
			/* Handle holes */
			if (!buffer_mapped(map_bh)) {
968
				loff_t i_size_aligned;
Linus Torvalds's avatar
Linus Torvalds committed
969 970

				/* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboe's avatar
Jens Axboe committed
971
				if (dio->rw & WRITE) {
Linus Torvalds's avatar
Linus Torvalds committed
972 973 974 975
					page_cache_release(page);
					return -ENOTBLK;
				}

976 977 978 979 980 981
				/*
				 * 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);
982
				if (sdio->block_in_file >=
983
						i_size_aligned >> blkbits) {
Linus Torvalds's avatar
Linus Torvalds committed
984 985 986 987
					/* We hit eof */
					page_cache_release(page);
					goto out;
				}
988 989
				zero_user(page, block_in_page << blkbits,
						1 << blkbits);
990
				sdio->block_in_file++;
Linus Torvalds's avatar
Linus Torvalds committed
991 992 993 994 995 996 997 998 999
				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.
			 */
1000
			if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1001
				dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
1002 1003 1004 1005 1006

			/*
			 * Work out, in this_chunk_blocks, how much disk we
			 * can add to this page
			 */
1007
			this_chunk_blocks = sdio->blocks_available;
Linus Torvalds's avatar
Linus Torvalds committed
1008 1009 1010
			u = (PAGE_SIZE - offset_in_page) >> blkbits;
			if (this_chunk_blocks > u)
				this_chunk_blocks = u;
1011
			u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds's avatar
Linus Torvalds committed
1012 1013 1014 1015 1016
			if (this_chunk_blocks > u)
				this_chunk_blocks = u;
			this_chunk_bytes = this_chunk_blocks << blkbits;
			BUG_ON(this_chunk_bytes == 0);

1017 1018 1019 1020
			sdio->boundary = buffer_boundary(map_bh);
			ret = submit_page_section(dio, sdio, page,
						  offset_in_page,
						  this_chunk_bytes,
1021 1022
						  sdio->next_block_for_io,
						  map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
1023 1024 1025 1026
			if (ret) {
				page_cache_release(page);
				goto out;
			}
1027
			sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds's avatar
Linus Torvalds committed
1028

1029
			sdio->block_in_file += this_chunk_blocks;
Linus Torvalds's avatar
Linus Torvalds committed
1030
			block_in_page += this_chunk_blocks;
1031
			sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds's avatar
Linus Torvalds committed
1032
next_block:
1033 1034
			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
1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045
				break;
		}

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

1046
static inline ssize_t
Linus Torvalds's avatar
Linus Torvalds committed
1047 1048
direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode, 
	const struct iovec *iov, loff_t offset, unsigned long nr_segs, 
1049
	unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
1050
	dio_submit_t submit_io, struct dio *dio, struct dio_submit *sdio)
Linus Torvalds's avatar
Linus Torvalds committed
1051 1052
{
	unsigned long user_addr; 
1053
	unsigned long flags;
Linus Torvalds's avatar
Linus Torvalds committed
1054 1055 1056 1057
	int seg;
	ssize_t ret = 0;
	ssize_t ret2;
	size_t bytes;
1058
	struct buffer_head map_bh = { 0, };
Linus Torvalds's avatar
Linus Torvalds committed
1059 1060 1061

	dio->inode = inode;
	dio->rw = rw;
1062 1063 1064
	sdio->blkbits = blkbits;
	sdio->blkfactor = inode->i_blkbits - blkbits;
	sdio->block_in_file = offset >> blkbits;
Linus Torvalds's avatar
Linus Torvalds committed
1065

1066
	sdio->get_block = get_block;
Linus Torvalds's avatar
Linus Torvalds committed
1067
	dio->end_io = end_io;
1068 1069 1070
	sdio->submit_io = submit_io;
	sdio->final_block_in_bio = -1;
	sdio->next_block_for_io = -1;
Linus Torvalds's avatar
Linus Torvalds committed
1071 1072

	dio->iocb = iocb;
1073
	dio->i_size = i_size_read(inode);
Linus Torvalds's avatar
Linus Torvalds committed
1074 1075

	spin_lock_init(&dio->bio_lock);
1076
	dio->refcount = 1;
Linus Torvalds's avatar
Linus Torvalds committed
1077 1078 1079 1080 1081

	/*
	 * In case of non-aligned buffers, we may need 2 more
	 * pages since we need to zero out first and last block.
	 */
1082 1083
	if (unlikely(sdio->blkfactor))
		sdio->pages_in_io = 2;
Linus Torvalds's avatar
Linus Torvalds committed
1084 1085 1086

	for (seg = 0; seg < nr_segs; seg++) {
		user_addr = (unsigned long)iov[seg].iov_base;
1087
		sdio->pages_in_io +=
Linus Torvalds's avatar
Linus Torvalds committed
1088 1089 1090 1091 1092 1093
			((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
				- user_addr/PAGE_SIZE);
	}

	for (seg = 0; seg < nr_segs; seg++) {
		user_addr = (unsigned long)iov[seg].iov_base;
1094
		sdio->size += bytes = iov[seg].iov_len;
Linus Torvalds's avatar
Linus Torvalds committed
1095 1096

		/* Index into the first page of the first block */
1097 1098
		sdio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
		sdio->final_block_in_request = sdio->block_in_file +
Linus Torvalds's avatar
Linus Torvalds committed
1099 1100
						(bytes >> blkbits);
		/* Page fetching state */
1101 1102 1103
		sdio->head = 0;
		sdio->tail = 0;
		sdio->curr_page = 0;
Linus Torvalds's avatar
Linus Torvalds committed
1104

1105
		sdio->total_pages = 0;
Linus Torvalds's avatar
Linus Torvalds committed
1106
		if (user_addr & (PAGE_SIZE-1)) {
1107
			sdio->total_pages++;
Linus Torvalds's avatar
Linus Torvalds committed
1108 1109
			bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
		}
1110 1111
		sdio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
		sdio->curr_user_address = user_addr;
Linus Torvalds's avatar
Linus Torvalds committed
1112
	
1113
		ret = do_direct_IO(dio, sdio, &map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
1114 1115

		dio->result += iov[seg].iov_len -
1116
			((sdio->final_block_in_request - sdio->block_in_file) <<
Linus Torvalds's avatar
Linus Torvalds committed
1117 1118 1119
					blkbits);

		if (ret) {
1120
			dio_cleanup(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
1121 1122 1123 1124
			break;
		}
	} /* end iovec loop */

1125
	if (ret == -ENOTBLK) {
Linus Torvalds's avatar
Linus Torvalds committed
1126 1127 1128 1129 1130 1131 1132 1133 1134 1135
		/*
		 * The remaining part of the request will be
		 * be handled by buffered I/O when we return
		 */
		ret = 0;
	}
	/*
	 * There may be some unwritten disk at the end of a part-written
	 * fs-block-sized block.  Go zero that now.
	 */
1136
	dio_zero_block(dio, sdio, 1, &map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
1137

1138
	if (sdio->cur_page) {
1139
		ret2 = dio_send_cur_page(dio, sdio, &map_bh);
Linus Torvalds's avatar
Linus Torvalds committed
1140 1141
		if (ret == 0)
			ret = ret2;
1142 1143
		page_cache_release(sdio->cur_page);
		sdio->cur_page = NULL;
Linus Torvalds's avatar
Linus Torvalds committed
1144
	}
1145 1146
	if (sdio->bio)
		dio_bio_submit(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
1147 1148 1149 1150 1151

	/*
	 * 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.
	 */
1152
	dio_cleanup(dio, sdio);
Linus Torvalds's avatar
Linus Torvalds committed
1153 1154 1155

	/*
	 * All block lookups have been performed. For READ requests
1156
	 * we can let i_mutex go now that its achieved its purpose
Linus Torvalds's avatar
Linus Torvalds committed
1157 1158
	 * of protecting us from looking up uninitialized blocks.
	 */
1159
	if (rw == READ && (dio->flags & DIO_LOCKING))
1160
		mutex_unlock(&dio->inode->i_mutex);
Linus Torvalds's avatar
Linus Torvalds committed
1161 1162

	/*
1163 1164 1165 1166 1167
	 * 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.
Linus Torvalds's avatar
Linus Torvalds committed
1168
	 */
1169 1170
	BUG_ON(ret == -EIOCBQUEUED);
	if (dio->is_async && ret == 0 && dio->result &&
1171
	    ((rw & READ) || (dio->result == sdio->size)))
1172
		ret = -EIOCBQUEUED;
1173

Jens Axboe's avatar
Jens Axboe committed
1174
	if (ret != -EIOCBQUEUED)
1175
		dio_await_completion(dio);
Linus Torvalds's avatar
Linus Torvalds committed
1176

1177 1178
	/*
	 * Sync will always be dropping the final ref and completing the
1179 1180 1181 1182 1183 1184 1185 1186
	 * 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.
1187
	 */
1188 1189 1190
	spin_lock_irqsave(&dio->bio_lock, flags);
	ret2 = --dio->refcount;
	spin_unlock_irqrestore(&dio->bio_lock, flags);
1191

1192
	if (ret2 == 0) {
1193
		ret = dio_complete(dio, offset, ret, false);
1194
		kmem_cache_free(dio_cache, dio);
1195 1196
	} else
		BUG_ON(ret != -EIOCBQUEUED);
Linus Torvalds's avatar
Linus Torvalds committed
1197 1198 1199 1200

	return ret;
}

1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212
/*
 * 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.
1213 1214 1215 1216 1217 1218 1219
 *
 * 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.
1220 1221 1222 1223 1224
 *
 * 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.
1225
 */
Linus Torvalds's avatar
Linus Torvalds committed
1226
ssize_t
1227
__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
Linus Torvalds's avatar
Linus Torvalds committed
1228
	struct block_device *bdev, const struct iovec *iov, loff_t offset, 
1229
	unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1230
	dio_submit_t submit_io,	int flags)
Linus Torvalds's avatar
Linus Torvalds committed
1231 1232 1233 1234 1235 1236 1237 1238