[PATCH 11/11] direct-io: Rewrite based on immutable biovecs

[Kent Overstreet]

Near total rewrite of fs/direct-io.c. This makes use of our new bio
splitting functionality, and the fact that generic_make_request() will
take arbitrary size bios - we allocate a bio, pin pages to it directly,
then call the getblocks() function to map it wherever the filesystem
tells us - splitting as needed.

Appears to pass xfstests with CONFIG_XFS_DEBUG=y (would appreciate testing from
someone who uses xfstests more than I).

Doesn't quite work with btrfs yet - when running xfstests it eventually get
stuck in an infinite loop somewhere. First thing in the kernel log is a warning
at fs/btrfs/ordered-data.c:288 in btrfs_add_ordered_sum() - if any of the btrfs
people want to take a look that would be a great help.

It _definitely_ needs review/auditing from someone intimately familiar with the
expected behaviour w.r.t. everything related to filesystem semantics and the
handling of the getblocks() call - dio_send_bio() is what primarily implements
that behaviour.

Signed-off-by: Kent Overstreet <kmo@xxxxxxxxxxxxx>
---
 fs/direct-io.c | 1426 +++++++++++++++-----------------------------------------
 1 file changed, 384 insertions(+), 1042 deletions(-)

diff --git a/fs/direct-io.c b/fs/direct-io.c
index 160a548..00adc44 100644
--- a/fs/direct-io.c
+++ b/fs/direct-io.c
@@ -8,7 +8,7 @@
  * 04Jul2002	Andrew Morton
  *		Initial version
  * 11Sep2002	janetinc@xxxxxxxxxx
- * 		added readv/writev support.
+ *		added readv/writev support.
  * 29Oct2002	Andrew Morton
  *		rewrote bio_add_page() support.
  * 30Oct2002	pbadari@xxxxxxxxxx
@@ -39,187 +39,48 @@
 #include <linux/prefetch.h>
 #include <linux/aio.h>
 
-/*
- * How many user pages to map in one call to get_user_pages().  This determines
- * the size of a structure in the slab cache
- */
-#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.
- */
-
-/* dio_state only used in the submission path */
-
-struct dio_submit {
-	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 */
-	int reap_counter;		/* rate limit reaping */
-	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 */
-	get_block_t *get_block;		/* block mapping function */
-	dio_submit_t *submit_io;	/* IO submition function */
-
-	loff_t logical_offset_in_bio;	/* current first logical block in bio */
-	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 */
-	loff_t cur_page_fs_offset;	/* Offset in file */
-
-	/*
-	 * 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 */
-};
-
 /* dio_state communicated between submission path and end_io */
 struct dio {
-	int flags;			/* doesn't change */
-	int rw;
-	struct inode *inode;
-	loff_t i_size;			/* i_size when submitted */
-	dio_iodone_t *end_io;		/* IO completion function */
-
-	void *private;			/* copy from map_bh.b_private */
+	int		flags;		/* doesn't change */
+	int		rw;
+	struct inode	*inode;
+	loff_t		i_size;		/* i_size when submitted */
+	unsigned	i_blkbits;
 
 	/* BIO completion state */
-	spinlock_t bio_lock;		/* protects BIO fields below */
-	int page_errors;		/* errno from get_user_pages() */
-	int is_async;			/* is IO async ? */
+	int		page_error;	/* errno from get_user_pages() */
+	int		io_error;	/* IO error in completion path */
 	bool defer_completion;		/* defer AIO completion to workqueue? */
-	int io_error;			/* IO error in completion path */
-	unsigned long refcount;		/* direct_io_worker() and bios */
-	struct bio *bio_list;		/* singly linked via bi_private */
+	atomic_t	refcount;	/* direct_io_worker() and bios */
+
+	dio_iodone_t	*end_io;	/* IO completion function */
+	void		*private;	/* copy from map_bh.b_private */
 	struct task_struct *waiter;	/* waiting task (NULL if none) */
 
 	/* AIO related stuff */
-	struct kiocb *iocb;		/* kiocb */
-	ssize_t result;                 /* IO result */
-
-	/*
-	 * 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.
-	 */
-	union {
-		struct page *pages[DIO_PAGES];	/* page buffer */
-		struct work_struct complete_work;/* deferred AIO completion */
-	};
-} ____cacheline_aligned_in_smp;
+	struct kiocb	*iocb;		/* kiocb */
+	ssize_t		result;		/* IO result */
+	struct work_struct complete_work;/* deferred AIO completion */
 
-static struct kmem_cache *dio_cache __read_mostly;
+	struct bio	bio;
+};
 
-/*
- * How many pages are in the queue?
- */
-static inline unsigned dio_pages_present(struct dio_submit *sdio)
-{
-	return sdio->tail - sdio->head;
-}
+#define DIO_WAKEUP	(1U << 31)
 
-/*
- * Go grab and pin some userspace pages.   Typically we'll get 64 at a time.
- */
-static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
-{
-	int ret;
-	int nr_pages;
-
-	nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
-	ret = get_user_pages_fast(
-		sdio->curr_user_address,		/* Where from? */
-		nr_pages,			/* How many pages? */
-		dio->rw == READ,		/* Write to memory? */
-		&dio->pages[0]);		/* Put results here */
-
-	if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
-		struct page *page = ZERO_PAGE(0);
-		/*
-		 * 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;
-		page_cache_get(page);
-		dio->pages[0] = page;
-		sdio->head = 0;
-		sdio->tail = 1;
-		ret = 0;
-		goto out;
-	}
+struct dio_mapping {
+	/* Low bit for READ|WRITE */
+	enum {
+		MAP_MAPPED	= 0,
+		MAP_NEW		= 2,
+		MAP_UNMAPPED	= 4,
+	} state;
 
-	if (ret >= 0) {
-		sdio->curr_user_address += ret * PAGE_SIZE;
-		sdio->curr_page += ret;
-		sdio->head = 0;
-		sdio->tail = ret;
-		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.
- */
-static inline struct page *dio_get_page(struct dio *dio,
-		struct dio_submit *sdio)
-{
-	if (dio_pages_present(sdio) == 0) {
-		int ret;
+	struct block_device	*bdev;
+	loff_t			offset;
+	size_t			size;
+};
 
-		ret = dio_refill_pages(dio, sdio);
-		if (ret)
-			return ERR_PTR(ret);
-		BUG_ON(dio_pages_present(sdio) == 0);
-	}
-	return dio->pages[sdio->head++];
-}
+static struct bio_set *dio_pool __read_mostly;
 
 /**
  * dio_complete() - called when all DIO BIO I/O has been completed
@@ -233,37 +94,18 @@ static inline struct page *dio_get_page(struct dio *dio,
  * filesystems can use it to hold additional state between get_block calls and
  * dio_complete.
  */
-static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret,
-		bool is_async)
+static ssize_t dio_complete(struct dio *dio, loff_t offset,
+			    ssize_t ret, bool is_async)
 {
-	ssize_t transferred = 0;
-
-	/*
-	 * 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;
-
-	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;
+		ret = dio->page_error;
 	if (ret == 0)
 		ret = dio->io_error;
 	if (ret == 0)
-		ret = transferred;
+		ret = dio->result;
 
 	if (dio->end_io && dio->result)
-		dio->end_io(dio->iocb, offset, transferred, dio->private);
+		dio->end_io(dio->iocb, offset, dio->result, dio->private);
 
 	inode_dio_done(dio->inode);
 	if (is_async) {
@@ -271,7 +113,7 @@ static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret,
 			int err;
 
 			err = generic_write_sync(dio->iocb->ki_filp, offset,
-						 transferred);
+						 dio->result);
 			if (err < 0 && ret > 0)
 				ret = err;
 		}
@@ -279,7 +121,7 @@ static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret,
 		aio_complete(dio->iocb, ret, 0);
 	}
 
-	kmem_cache_free(dio_cache, dio);
+	bio_put(&dio->bio);
 	return ret;
 }
 
@@ -290,57 +132,6 @@ static void dio_aio_complete_work(struct work_struct *work)
 	dio_complete(dio, dio->iocb->ki_pos, 0, true);
 }
 
-static int dio_bio_complete(struct dio *dio, struct bio *bio);
-
-/*
- * Asynchronous IO callback. 
- */
-static void dio_bio_end_aio(struct bio *bio, int error)
-{
-	struct dio *dio = bio->bi_private;
-	unsigned long remaining;
-	unsigned long flags;
-
-	/* cleanup the bio */
-	dio_bio_complete(dio, bio);
-
-	spin_lock_irqsave(&dio->bio_lock, flags);
-	remaining = --dio->refcount;
-	if (remaining == 1 && dio->waiter)
-		wake_up_process(dio->waiter);
-	spin_unlock_irqrestore(&dio->bio_lock, flags);
-
-	if (remaining == 0) {
-		if (dio->result && dio->defer_completion) {
-			INIT_WORK(&dio->complete_work, dio_aio_complete_work);
-			queue_work(dio->inode->i_sb->s_dio_done_wq,
-				   &dio->complete_work);
-		} else {
-			dio_complete(dio, dio->iocb->ki_pos, 0, true);
-		}
-	}
-}
-
-/*
- * 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.
- */
-static void dio_bio_end_io(struct bio *bio, int error)
-{
-	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;
-	if (--dio->refcount == 1 && dio->waiter)
-		wake_up_process(dio->waiter);
-	spin_unlock_irqrestore(&dio->bio_lock, flags);
-}
-
 /**
  * dio_end_io - handle the end io action for the given bio
  * @bio: The direct io bio thats being completed
@@ -353,187 +144,67 @@ static void dio_bio_end_io(struct bio *bio, int error)
 void dio_end_io(struct bio *bio, int error)
 {
 	struct dio *dio = bio->bi_private;
+	unsigned remaining;
 
-	if (dio->is_async)
-		dio_bio_end_aio(bio, error);
-	else
-		dio_bio_end_io(bio, error);
-}
-EXPORT_SYMBOL_GPL(dio_end_io);
-
-static inline void
-dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
-	      struct block_device *bdev,
-	      sector_t first_sector, int nr_vecs)
-{
-	struct bio *bio;
-
-	/*
-	 * bio_alloc() is guaranteed to return a bio when called with
-	 * __GFP_WAIT and we request a valid number of vectors.
-	 */
-	bio = bio_alloc(GFP_KERNEL, nr_vecs);
-
-	bio->bi_bdev = bdev;
-	bio->bi_iter.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;
-
-	sdio->bio = bio;
-	sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
-}
-
-/*
- * 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().
- *
- * bios hold a dio reference between submit_bio and ->end_io.
- */
-static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
-{
-	struct bio *bio = sdio->bio;
-	unsigned long flags;
-
-	bio->bi_private = dio;
-
-	spin_lock_irqsave(&dio->bio_lock, flags);
-	dio->refcount++;
-	spin_unlock_irqrestore(&dio->bio_lock, flags);
+	if (error)
+		dio->io_error = -EIO;
 
-	if (dio->is_async && dio->rw == READ)
-		bio_set_pages_dirty(bio);
+	if (dio->rw == READ) {
+		bio_check_pages_dirty(bio);	/* transfers ownership */
+	} else {
+		struct bio_vec *bv;
+		int i;
 
-	if (sdio->submit_io)
-		sdio->submit_io(dio->rw, bio, dio->inode,
-			       sdio->logical_offset_in_bio);
-	else
-		submit_bio(dio->rw, bio);
+		bio_for_each_segment_all(bv, bio, i)
+			page_cache_release(bv->bv_page);
+		bio_put(bio);
+	}
 
-	sdio->bio = NULL;
-	sdio->boundary = 0;
-	sdio->logical_offset_in_bio = 0;
-}
+	remaining = atomic_dec_return(&dio->refcount);
 
-/*
- * Release any resources in case of a failure
- */
-static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
-{
-	while (dio_pages_present(sdio))
-		page_cache_release(dio_get_page(dio, sdio));
+	if (remaining == DIO_WAKEUP)
+		wake_up_process(dio->waiter);
+	else if (!remaining) {
+		if (dio->defer_completion) {
+			INIT_WORK(&dio->complete_work, dio_aio_complete_work);
+			queue_work(dio->inode->i_sb->s_dio_done_wq,
+				   &dio->complete_work);
+		} else {
+			dio_complete(dio, dio->iocb->ki_pos, 0, true);
+		}
+	}
 }
+EXPORT_SYMBOL_GPL(dio_end_io);
 
-/*
- * 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.
- */
-static struct bio *dio_await_one(struct dio *dio)
+static void dio_wait_completion(struct dio *dio)
 {
-	unsigned long flags;
-	struct bio *bio = NULL;
+	if (atomic_add_return(DIO_WAKEUP - 1, &dio->refcount) == DIO_WAKEUP)
+		return;
 
-	spin_lock_irqsave(&dio->bio_lock, flags);
+	while (1) {
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		if (atomic_read(&dio->refcount) == DIO_WAKEUP)
+			break;
 
-	/*
-	 * 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;
-	}
-	if (dio->bio_list) {
-		bio = dio->bio_list;
-		dio->bio_list = bio->bi_private;
 	}
-	spin_unlock_irqrestore(&dio->bio_lock, flags);
-	return bio;
+	__set_current_state(TASK_RUNNING);
 }
 
 /*
- * Process one completed BIO.  No locks are held.
+ * 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
  */
-static int dio_bio_complete(struct dio *dio, struct bio *bio)
+static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
 {
-	const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
-	struct bio_vec *bvec;
 	unsigned i;
+	unsigned nblocks;
 
-	if (!uptodate)
-		dio->io_error = -EIO;
-
-	if (dio->is_async && dio->rw == READ) {
-		bio_check_pages_dirty(bio);	/* transfers ownership */
-	} else {
-		bio_for_each_segment_all(bvec, bio, i) {
-			struct page *page = bvec->bv_page;
-
-			if (dio->rw == READ && !PageCompound(page))
-				set_page_dirty_lock(page);
-			page_cache_release(page);
-		}
-		bio_put(bio);
-	}
-	return uptodate ? 0 : -EIO;
-}
-
-/*
- * 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
- * errors are propagated through dio->io_error and should be propagated via
- * dio_complete().
- */
-static void dio_await_completion(struct dio *dio)
-{
-	struct bio *bio;
-	do {
-		bio = dio_await_one(dio);
-		if (bio)
-			dio_bio_complete(dio, bio);
-	} while (bio);
-}
+	nblocks = map_bh->b_size >> dio->i_blkbits;
 
-/*
- * 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.
- */
-static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
-{
-	int ret = 0;
-
-	if (sdio->reap_counter++ >= 64) {
-		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;
-		}
-		sdio->reap_counter = 0;
-	}
-	return ret;
+	for (i = 0; i < nblocks; i++)
+		unmap_underlying_metadata(map_bh->b_bdev,
+					  map_bh->b_blocknr + i);
 }
 
 /*
@@ -572,513 +243,309 @@ static int dio_set_defer_completion(struct dio *dio)
 	return 0;
 }
 
-/*
- * Call into the fs to map some more disk blocks.  We record the current number
- * of available blocks at sdio->blocks_available.  These are in units of the
- * 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.
- *
- * get_block() is passed the number of i_blkbits-sized blocks which direct_io
- * 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
- * block at a time - it will repeatedly call get_block() as it walks the hole.
- */
-static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
-			   struct buffer_head *map_bh)
+static int get_blocks(struct dio *dio, loff_t offset, size_t size,
+		      struct dio_mapping *map, get_block_t *get_block)
 {
-	int ret;
-	sector_t fs_startblk;	/* Into file, in filesystem-sized blocks */
-	sector_t fs_endblk;	/* Into file, in filesystem-sized blocks */
-	unsigned long fs_count;	/* Number of filesystem-sized blocks */
-	int create;
-	unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
+	struct buffer_head map_bh = { 0, };
+	int ret, create;
+	unsigned i_mask = (1 << dio->i_blkbits) - 1;
+	unsigned fs_offset = offset & i_mask;
+	sector_t fs_blocknr = offset >> dio->i_blkbits;
 
 	/*
-	 * If there was a memory error and we've overwritten all the
-	 * mapped blocks then we can now return that memory error
+	 * 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.
 	 */
-	ret = dio->page_errors;
-	if (ret == 0) {
-		BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
-		fs_startblk = sdio->block_in_file >> sdio->blkfactor;
-		fs_endblk = (sdio->final_block_in_request - 1) >>
-					sdio->blkfactor;
-		fs_count = fs_endblk - fs_startblk + 1;
+	create = dio->rw & WRITE;
+	if (dio->flags & DIO_SKIP_HOLES) {
+		if (fs_blocknr < dio->i_size >> dio->i_blkbits)
+			create = 0;
+	}
 
-		map_bh->b_state = 0;
-		map_bh->b_size = fs_count << i_blkbits;
+	/* fs expects units of fs_blocks */
+	map_bh.b_size = size + fs_offset;
+	map_bh.b_size = round_up(map_bh.b_size, 1 << dio->i_blkbits);
 
-		/*
-		 * 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.
-		 */
-		create = dio->rw & WRITE;
-		if (dio->flags & DIO_SKIP_HOLES) {
-			if (sdio->block_in_file < (i_size_read(dio->inode) >>
-							sdio->blkbits))
-				create = 0;
-		}
+	ret = get_block(dio->inode, fs_blocknr, &map_bh, create);
+	if (ret)
+		return ret;
 
-		ret = (*sdio->get_block)(dio->inode, fs_startblk,
-						map_bh, create);
+	/* Store for completion */
+	dio->private = map_bh.b_private;
 
-		/* Store for completion */
-		dio->private = map_bh->b_private;
+	if (ret == 0 && buffer_defer_completion(&map_bh))
+		ret = dio_set_defer_completion(dio);
 
-		if (ret == 0 && buffer_defer_completion(map_bh))
-			ret = dio_set_defer_completion(dio);
-	}
-	return ret;
-}
+	if (buffer_new(&map_bh))
+		clean_blockdev_aliases(dio, &map_bh);
 
-/*
- * There is no bio.  Make one now.
- */
-static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
-		sector_t start_sector, struct buffer_head *map_bh)
-{
-	sector_t sector;
-	int ret, nr_pages;
+	if (!buffer_mapped(&map_bh))
+		map->state = MAP_UNMAPPED;
+	else if (buffer_new(&map_bh))
+		map->state = MAP_NEW;
+	else
+		map->state = MAP_MAPPED;
+
+#if 0
+	/* Previous DIO code only handled holes one block at a time */
+	if (map->state == MAP_UNMAPPED)
+		map_bh.b_size = 1 << dio->i_blkbits;
+
+#endif
+	BUG_ON(map_bh.b_size <= fs_offset);
+
+	map->bdev = map_bh.b_bdev;
+	map->offset = (map_bh.b_blocknr << dio->i_blkbits) +
+		fs_offset;
+	map->size = min(map_bh.b_size - fs_offset, size);
 
-	ret = dio_bio_reap(dio, sdio);
-	if (ret)
-		goto out;
-	sector = start_sector << (sdio->blkbits - 9);
-	nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
-	nr_pages = min(nr_pages, BIO_MAX_PAGES);
-	BUG_ON(nr_pages <= 0);
-	dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
-	sdio->boundary = 0;
-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.
- */
-static inline int dio_bio_add_page(struct dio_submit *sdio)
+static void __dio_bio_submit(struct dio *dio, struct bio *bio,
+			     loff_t offset, dio_submit_t *submit_io)
 {
-	int ret;
+	/*
+	 * Read accounting is performed in submit_bio()
+	 */
+	if (dio->rw & WRITE)
+		task_io_account_write(bio->bi_iter.bi_size);
 
-	ret = bio_add_page(sdio->bio, sdio->cur_page,
-			sdio->cur_page_len, sdio->cur_page_offset);
-	if (ret == sdio->cur_page_len) {
-		/*
-		 * Decrement count only, if we are done with this page
-		 */
-		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);
-		ret = 0;
-	} else {
-		ret = 1;
-	}
-	return ret;
+	if (submit_io)
+		submit_io(dio->rw, bio, dio->inode,
+			  offset >> dio->i_blkbits);
+	else
+		submit_bio(dio->rw, bio);
 }
-		
+
 /*
- * 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).
+ * For reads 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().
  *
- * The caller of this function is responsible for removing cur_page from the
- * dio, and for dropping the refcount which came from that presence.
+ * bios hold a dio reference between submit_bio and ->end_io.
  */
-static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
-		struct buffer_head *map_bh)
+static int dio_bio_submit(struct dio *dio, struct bio *bio,
+			  struct dio_mapping *map,
+			  loff_t offset, dio_submit_t *submit_io)
 {
-	int ret = 0;
+	struct bio *split;
 
-	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_iter.bi_size;
+	split = bio_next_split(bio, map->size >> 9,
+			       GFP_KERNEL, fs_bio_set);
 
-		/*
-		 * See whether this new request is contiguous with the old.
-		 *
-		 * Btrfs cannot handle having logically non-contiguous requests
-		 * submitted.  For example if you have
-		 *
-		 * Logical:  [0-4095][HOLE][8192-12287]
-		 * Physical: [0-4095]      [4096-8191]
-		 *
-		 * 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.
-		 */
-		if (sdio->final_block_in_bio != sdio->cur_page_block ||
-		    cur_offset != bio_next_offset)
-			dio_bio_submit(dio, sdio);
-	}
+	if (split != bio)
+		bio_chain(split, bio);
+	else
+		atomic_inc(&bio->bi_remaining);
 
-	if (sdio->bio == NULL) {
-		ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
-		if (ret)
-			goto out;
-	}
+	split->bi_bdev = map->bdev;
+	split->bi_iter.bi_sector = map->offset >> 9;
 
-	if (dio_bio_add_page(sdio) != 0) {
-		dio_bio_submit(dio, sdio);
-		ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
-		if (ret == 0) {
-			ret = dio_bio_add_page(sdio);
-			BUG_ON(ret != 0);
-		}
-	}
-out:
-	return ret;
+	dio->result += map->size;
+
+	__dio_bio_submit(dio, bio, offset, submit_io);
+
+	return split == bio;
 }
 
-/*
- * 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.
- */
-static inline int
-submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
-		    unsigned offset, unsigned len, sector_t blocknr,
-		    struct buffer_head *map_bh)
+static void dio_write_zeroes(struct dio *dio, struct bio *parent,
+			     struct block_device *bdev,
+			     sector_t sector, size_t size,
+			     loff_t offset, dio_submit_t *submit_io)
 {
-	int ret = 0;
+	unsigned pages = DIV_ROUND_UP(size, PAGE_SIZE);
+	struct bio *bio = bio_alloc(GFP_KERNEL, pages);
 
-	if (dio->rw & WRITE) {
-		/*
-		 * Read accounting is performed in submit_bio()
-		 */
-		task_io_account_write(len);
+	while (pages--) {
+		bio->bi_io_vec[pages].bv_page = ZERO_PAGE(0);
+		bio->bi_io_vec[pages].bv_len = PAGE_SIZE;
+		bio->bi_io_vec[pages].bv_offset = 0;
 	}
 
-	/*
-	 * Can we just grow the current page's presence in the dio?
-	 */
-	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;
-		goto out;
-	}
-
-	/*
-	 * If there's a deferred page already there then send it.
-	 */
-	if (sdio->cur_page) {
-		ret = dio_send_cur_page(dio, sdio, map_bh);
-		page_cache_release(sdio->cur_page);
-		sdio->cur_page = NULL;
-		if (ret)
-			return ret;
-	}
+	bio->bi_bdev = bdev;
+	bio->bi_iter.bi_sector = sector;
+	bio->bi_iter.bi_size = size;
 
-	page_cache_get(page);		/* It is in dio */
-	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;
-out:
-	/*
-	 * If sdio->boundary then we want to schedule the IO now to
-	 * avoid metadata seeks.
-	 */
-	if (sdio->boundary) {
-		ret = dio_send_cur_page(dio, sdio, map_bh);
-		dio_bio_submit(dio, sdio);
-		page_cache_release(sdio->cur_page);
-		sdio->cur_page = NULL;
-	}
-	return ret;
+	bio_chain(bio, parent);
+	__dio_bio_submit(dio, bio, offset, submit_io);
 }
 
-/*
- * 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
- */
-static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
+static void dio_zero_partial_block_front(struct dio *dio, struct bio *bio,
+					 struct dio_mapping *map, loff_t offset,
+					 dio_submit_t *submit_io)
 {
-	unsigned i;
-	unsigned nblocks;
-
-	nblocks = map_bh->b_size >> dio->inode->i_blkbits;
-
-	for (i = 0; i < nblocks; i++) {
-		unmap_underlying_metadata(map_bh->b_bdev,
-					  map_bh->b_blocknr + i);
-	}
+	unsigned blksize = 1 << dio->i_blkbits;
+	unsigned blkmask = blksize - 1;
+	unsigned front = offset & blkmask;
+
+	if (front)
+		dio_write_zeroes(dio, bio, map->bdev,
+				 (map->offset - front) >> 9,
+				 front, offset, submit_io);
 }
 
-/*
- * 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.
- */
-static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
-		int end, struct buffer_head *map_bh)
+static void dio_zero_partial_block_back(struct dio *dio, struct bio *bio,
+					struct dio_mapping *map, loff_t offset,
+					dio_submit_t *submit_io)
 {
-	unsigned dio_blocks_per_fs_block;
-	unsigned this_chunk_blocks;	/* In dio_blocks */
-	unsigned this_chunk_bytes;
-	struct page *page;
+	unsigned blksize = 1 << dio->i_blkbits;
+	unsigned blkmask = blksize - 1;
+	unsigned back = (offset + map->size) & blkmask;
+
+	if (back)
+		dio_write_zeroes(dio, bio, map->bdev,
+				 (map->offset + map->size) >> 9,
+				 blksize - back, offset, submit_io);
+}
 
-	sdio->start_zero_done = 1;
-	if (!sdio->blkfactor || !buffer_new(map_bh))
-		return;
+static int dio_read_zeroes(struct dio *dio, struct bio *bio,
+			   struct dio_mapping *map)
+{
+	swap(bio->bi_iter.bi_size, map->size);
+	zero_fill_bio(bio);
+	swap(bio->bi_iter.bi_size, map->size);
 
-	dio_blocks_per_fs_block = 1 << sdio->blkfactor;
-	this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
+	dio->result += map->size;
+	bio_advance(bio, map->size);
 
-	if (!this_chunk_blocks)
-		return;
+	return !bio->bi_iter.bi_size;
+}
 
+static int dio_is_aligned(struct dio *dio, struct dio_mapping *map)
+{
 	/*
-	 * We need to zero out part of an fs block.  It is either at the
-	 * beginning or the end of the fs block.
+	 * XXX: have to make sure we're at least sector aligned, but maybe leave
+	 * the rest to generic_make_request()?
 	 */
-	if (end) 
-		this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
 
-	this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
+	unsigned blocksize_mask =
+		roundup_pow_of_two(bdev_logical_block_size(map->bdev)) - 1;
 
-	page = ZERO_PAGE(0);
-	if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
-				sdio->next_block_for_io, map_bh))
-		return;
-
-	sdio->next_block_for_io += this_chunk_blocks;
+	return !(map->offset & blocksize_mask) &&
+		!(map->size & blocksize_mask);
 }
 
-/*
- * 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.
- *
- * So what we do is to permit the ->get_block function to populate bh.b_size
- * 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
- * it should set b_size to PAGE_SIZE or more inside get_block().  This gives
- * fine alignment but still allows this function to work in PAGE_SIZE units.
- */
-static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
-			struct buffer_head *map_bh)
+static int dio_send_bio(struct dio *dio, struct bio *bio, loff_t offset,
+			get_block_t *get_block, dio_submit_t *submit_io)
 {
-	const unsigned blkbits = sdio->blkbits;
-	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 */
-	block_in_page = sdio->first_block_in_page;
-
-	while (sdio->block_in_file < sdio->final_block_in_request) {
-		page = dio_get_page(dio, sdio);
-		if (IS_ERR(page)) {
-			ret = PTR_ERR(page);
-			goto out;
-		}
+	struct dio_mapping map;
+	int ret = 0, rw = dio->rw & WRITE;
+	bool done;
 
-		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;
-
-			if (sdio->blocks_available == 0) {
-				/*
-				 * Need to go and map some more disk
-				 */
-				unsigned long blkmask;
-				unsigned long dio_remainder;
-
-				ret = get_more_blocks(dio, sdio, map_bh);
-				if (ret) {
-					page_cache_release(page);
-					goto out;
-				}
-				if (!buffer_mapped(map_bh))
-					goto do_holes;
-
-				sdio->blocks_available =
-						map_bh->b_size >> sdio->blkbits;
-				sdio->next_block_for_io =
-					map_bh->b_blocknr << sdio->blkfactor;
-				if (buffer_new(map_bh))
-					clean_blockdev_aliases(dio, map_bh);
-
-				if (!sdio->blkfactor)
-					goto do_holes;
-
-				blkmask = (1 << sdio->blkfactor) - 1;
-				dio_remainder = (sdio->block_in_file & blkmask);
-
-				/*
-				 * 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))
-					sdio->next_block_for_io += dio_remainder;
-				sdio->blocks_available -= dio_remainder;
-			}
-do_holes:
-			/* Handle holes */
-			if (!buffer_mapped(map_bh)) {
-				loff_t i_size_aligned;
-
-				/* AKPM: eargh, -ENOTBLK is a hack */
-				if (dio->rw & WRITE) {
-					page_cache_release(page);
-					return -ENOTBLK;
-				}
-
-				/*
-				 * 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);
-				if (sdio->block_in_file >=
-						i_size_aligned >> blkbits) {
-					/* We hit eof */
-					page_cache_release(page);
-					goto out;
-				}
-				zero_user(page, block_in_page << blkbits,
-						1 << blkbits);
-				sdio->block_in_file++;
-				block_in_page++;
-				goto next_block;
+	if (rw == READ)
+		bio_set_pages_dirty(bio);
+
+	while (1) {
+		if (rw == READ && offset >= dio->i_size)
+			break;
+
+		ret = get_blocks(dio, offset, bio->bi_iter.bi_size,
+				 &map, get_block);
+		if (ret)
+			break;
+
+		switch (map.state|rw) {
+		case MAP_MAPPED|READ:
+		case MAP_MAPPED|WRITE:
+			if (!dio_is_aligned(dio, &map)) {
+				ret = -EINVAL;
+				goto out;
 			}
 
-			/*
-			 * 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.
-			 */
-			if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
-				dio_zero_block(dio, sdio, 0, map_bh);
-
-			/*
-			 * Work out, in this_chunk_blocks, how much disk we
-			 * can add to this page
-			 */
-			this_chunk_blocks = sdio->blocks_available;
-			u = (PAGE_SIZE - offset_in_page) >> blkbits;
-			if (this_chunk_blocks > u)
-				this_chunk_blocks = u;
-			u = sdio->final_block_in_request - sdio->block_in_file;
-			if (this_chunk_blocks > u)
-				this_chunk_blocks = u;
-			this_chunk_bytes = this_chunk_blocks << blkbits;
-			BUG_ON(this_chunk_bytes == 0);
-
-			if (this_chunk_blocks == sdio->blocks_available)
-				sdio->boundary = buffer_boundary(map_bh);
-			ret = submit_page_section(dio, sdio, page,
-						  offset_in_page,
-						  this_chunk_bytes,
-						  sdio->next_block_for_io,
-						  map_bh);
-			if (ret) {
-				page_cache_release(page);
+			if (dio_bio_submit(dio, bio, &map, offset, submit_io))
+				goto out;
+			break;
+		case MAP_NEW|READ:
+		case MAP_UNMAPPED|READ:
+			if (dio_read_zeroes(dio, bio, &map))
+				goto out;
+
+			break;
+		case MAP_NEW|WRITE:
+			if (!dio_is_aligned(dio, &map)) {
+				ret = -EINVAL;
 				goto out;
 			}
-			sdio->next_block_for_io += this_chunk_blocks;
-
-			sdio->block_in_file += this_chunk_blocks;
-			block_in_page += this_chunk_blocks;
-			sdio->blocks_available -= this_chunk_blocks;
-next_block:
-			BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
-			if (sdio->block_in_file == sdio->final_block_in_request)
-				break;
+
+			dio_zero_partial_block_front(dio, bio, &map,
+						     offset, submit_io);
+
+			done = dio_bio_submit(dio, bio, &map, offset, submit_io);
+
+			dio_zero_partial_block_back(dio, bio, &map,
+						    offset, submit_io);
+
+			if (done)
+				goto out;
+
+			break;
+		case MAP_UNMAPPED|WRITE:
+			/* AKPM: eargh, -ENOTBLK is a hack */
+			ret = -ENOTBLK;
+			goto out;
 		}
 
-		/* Drop the ref which was taken in get_user_pages() */
-		page_cache_release(page);
-		block_in_page = 0;
+		offset += map.size;
 	}
 out:
+	if (rw == READ && offset + dio->result > dio->i_size) {
+		BUG_ON(offset > dio->i_size ||
+		       (offset == dio->i_size && dio->result));
+		dio->result = dio->i_size - offset;
+	}
+
+	bio_endio(bio, 0);
 	return ret;
 }
 
-static inline int drop_refcount(struct dio *dio)
+static int dio_alloc_bios(struct dio *dio, loff_t offset,
+			  struct iov_iter *iter, unsigned nr_pages,
+			  get_block_t *get_block, dio_submit_t *submit_io)
 {
-	int ret2;
-	unsigned long flags;
+	ssize_t ret;
+	struct bio *bio;
 
-	/*
-	 * Sync will always be dropping the final ref and completing the
-	 * 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.
-	 */
-	spin_lock_irqsave(&dio->bio_lock, flags);
-	ret2 = --dio->refcount;
-	spin_unlock_irqrestore(&dio->bio_lock, flags);
-	return ret2;
+	bio = &dio->bio;
+	bio_get(bio);
+	goto start;
+
+	while (iter->count) {
+		BUG_ON(!nr_pages);
+
+		bio = bio_alloc(GFP_KERNEL,
+				min_t(unsigned, BIO_MAX_PAGES, nr_pages));
+start:
+		bio->bi_private = dio;
+		bio->bi_end_io = dio_end_io;
+
+		ret = bio_get_user_pages(bio, iter, dio->rw == READ);
+		if (ret < 0) {
+			bio_put(bio);
+			dio->page_error = ret;
+			return 0;
+		}
+
+		nr_pages -= bio->bi_vcnt;
+
+		atomic_inc(&dio->refcount);
+		ret = dio_send_bio(dio, bio, offset + dio->result,
+				   get_block, submit_io);
+		if (ret)
+			return ret;
+
+		if (dio->rw == READ && offset + dio->result == dio->i_size)
+			break;
+	}
+
+	return 0;
 }
 
 /*
@@ -1100,82 +567,36 @@ static inline int drop_refcount(struct dio *dio)
  * 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.
- *
- * 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.
  */
-static inline ssize_t
+static ssize_t
 do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
-	struct block_device *bdev, const struct iovec *iov, loff_t offset, 
+	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)
 {
-	int seg;
-	size_t size;
-	unsigned long addr;
-	unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
-	unsigned blkbits = i_blkbits;
-	unsigned blocksize_mask = (1 << blkbits) - 1;
-	ssize_t retval = -EINVAL;
-	loff_t end = offset;
+	unsigned nr_pages = 0, i_blkbits;
+	size_t size = iocb->ki_nbytes;
+	ssize_t retval = 0;
 	struct dio *dio;
-	struct dio_submit sdio = { 0, };
-	unsigned long user_addr;
-	size_t bytes;
-	struct buffer_head map_bh = { 0, };
 	struct blk_plug plug;
+	struct iov_iter iter;
 
 	if (rw & WRITE)
 		rw = WRITE_ODIRECT;
 
-	/*
-	 * Avoid references to bdev if not absolutely needed to give
-	 * the early prefetch in the caller enough time.
-	 */
-
-	if (offset & blocksize_mask) {
-		if (bdev)
-			blkbits = blksize_bits(bdev_logical_block_size(bdev));
-		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;
-		if (unlikely((addr & blocksize_mask) ||
-			     (size & blocksize_mask))) {
-			if (bdev)
-				blkbits = blksize_bits(
-					 bdev_logical_block_size(bdev));
-			blocksize_mask = (1 << blkbits) - 1;
-			if ((addr & blocksize_mask) || (size & blocksize_mask))
-				goto out;
-		}
-	}
+	i_blkbits = ACCESS_ONCE(inode->i_blkbits);
 
 	/* watch out for a 0 len io from a tricksy fs */
-	if (rw == READ && end == offset)
+	if (rw == READ && !size)
 		return 0;
 
-	dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
-	retval = -ENOMEM;
-	if (!dio)
-		goto out;
-	/*
-	 * 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));
+	iov_iter_init(&iter, iov, nr_segs, size);
+
+	nr_pages = iov_count_pages(&iter, 511);
+	if (nr_pages < 0)
+		return nr_pages;
 
-	dio->flags = flags;
-	if (dio->flags & DIO_LOCKING) {
+	if (flags & DIO_LOCKING) {
 		if (rw == READ) {
 			struct address_space *mapping =
 					iocb->ki_filp->f_mapping;
@@ -1184,114 +605,42 @@ do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
 			mutex_lock(&inode->i_mutex);
 
 			retval = filemap_write_and_wait_range(mapping, offset,
-							      end - 1);
+							   offset + size - 1);
 			if (retval) {
 				mutex_unlock(&inode->i_mutex);
-				kmem_cache_free(dio_cache, dio);
-				goto out;
+				return retval;
 			}
 		}
 	}
 
 	/*
-	 * 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.
-	 */
-	dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
-		(end > i_size_read(inode)));
-	dio->inode = inode;
-	dio->rw = rw;
-
-	/*
-	 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
-	 * so that we can call ->fsync.
-	 */
-	if (dio->is_async && (rw & WRITE) &&
-	    ((iocb->ki_filp->f_flags & O_DSYNC) ||
-	     IS_SYNC(iocb->ki_filp->f_mapping->host))) {
-		retval = dio_set_defer_completion(dio);
-		if (retval) {
-			/*
-			 * We grab i_mutex only for reads so we don't have
-			 * to release it here
-			 */
-			kmem_cache_free(dio_cache, dio);
-			goto out;
-		}
-	}
-
-	/*
 	 * Will be decremented at I/O completion time.
 	 */
 	atomic_inc(&inode->i_dio_count);
 
-	retval = 0;
-	sdio.blkbits = blkbits;
-	sdio.blkfactor = 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);
-	}
+	dio = container_of(bio_alloc_bioset(GFP_KERNEL,
+				    min_t(unsigned, BIO_MAX_PAGES, nr_pages),
+				    dio_pool),
+			   struct dio, bio);
+
+	dio->flags	= flags;
+	dio->rw		= rw;
+	dio->inode	= inode;
+	dio->i_size	= i_size_read(inode);
+	dio->i_blkbits	= i_blkbits;
+	dio->end_io	= end_io;
+	dio->private	= NULL;
+	dio->page_error	= 0;
+	dio->io_error	= 0;
+	atomic_set(&dio->refcount, 1);
+	dio->waiter	= current;
+	dio->iocb	= iocb;
+	dio->result	= 0;
 
 	blk_start_plug(&plug);
 
-	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 */
+	retval = dio_alloc_bios(dio, offset, &iter, nr_pages,
+				get_block, submit_io);
 
 	if (retval == -ENOTBLK) {
 		/*
@@ -1300,33 +649,10 @@ do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
 		 */
 		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);
 
 	blk_finish_plug(&plug);
 
 	/*
-	 * 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.
@@ -1342,19 +668,28 @@ do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
 	 * This had *better* be the only place that raises -EIOCBQUEUED.
 	 */
 	BUG_ON(retval == -EIOCBQUEUED);
-	if (dio->is_async && retval == 0 && dio->result &&
-	    ((rw == READ) || (dio->result == sdio.size)))
-		retval = -EIOCBQUEUED;
-
-	if (retval != -EIOCBQUEUED)
-		dio_await_completion(dio);
 
-	if (drop_refcount(dio) == 0) {
+	/*
+	 * 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.
+	 */
+	if (!is_sync_kiocb(iocb) &&
+	    retval == 0 && dio->result &&
+	    ((rw == READ) ||
+	     (offset + size <= dio->i_size &&
+	      dio->result == size))) {
+		if (atomic_dec_and_test(&dio->refcount))
+			retval = dio_complete(dio, offset, retval, false);
+		else
+			retval = -EIOCBQUEUED;
+	} else {
+		dio_wait_completion(dio);
 		retval = dio_complete(dio, offset, retval, false);
-	} else
-		BUG_ON(retval != -EIOCBQUEUED);
+		BUG_ON(retval == -EIOCBQUEUED);
+	}
 
-out:
 	return retval;
 }
 
@@ -1376,16 +711,23 @@ __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
 	prefetch(bdev->bd_queue);
 	prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
 
-	return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset,
+	return do_blockdev_direct_IO(rw, iocb, inode, iov, offset,
 				     nr_segs, get_block, end_io,
 				     submit_io, flags);
 }
-
 EXPORT_SYMBOL(__blockdev_direct_IO);
 
 static __init int dio_init(void)
 {
-	dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
+	/*
+	 * First argument to bioset_create() is completely arbitrary - it just
+	 * has to be nonzero to always make forward progress.
+	 */
+
+	dio_pool = bioset_create(4, offsetof(struct dio, bio));
+	if (!dio_pool)
+		panic("dio: can't allocate bios\n");
+
 	return 0;
 }
-module_init(dio_init)
+subsys_initcall(dio_init);
-- 
1.8.4.4

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/