Btrfs: add basic DIO read/write support

This provides basic DIO support for reading and writing.  It does not do the
work to recover from mismatching checksums, that will come later.  A few design
changes have been made from Jim's code (sorry Jim!)

1) Use the generic direct-io code.  Jim originally re-wrote all the generic DIO
code in order to account for all of BTRFS's oddities, but thanks to that work it
seems like the best bet is to just ignore compression and such and just opt to
fallback on buffered IO.

2) Fallback on buffered IO for compressed or inline extents.  Jim's code did
it's own buffering to make dio with compressed extents work.  Now we just
fallback onto normal buffered IO.

3) Use ordered extents for the writes so that all of the

lock_extent()
lookup_ordered()

type checks continue to work.

4) Do the lock_extent() lookup_ordered() loop in readpage so we don't race with
DIO writes.

I've tested this with fsx and everything works great.  This patch depends on my
dio and filemap.c patches to work.  Thanks,

Signed-off-by: Josef Bacik <josef@redhat.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>

6 files changed, 631 insertions, 36 deletions

diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h
index 5ed0223d1cb..e9bf86415e8 100644
--- a/fs/btrfs/ctree.h
+++ b/fs/btrfs/ctree.h
@@ -2317,6 +2317,8 @@ int btrfs_del_csums(struct btrfs_trans_handle *trans,
 		    struct btrfs_root *root, u64 bytenr, u64 len);
 int btrfs_lookup_bio_sums(struct btrfs_root *root, struct inode *inode,
 			  struct bio *bio, u32 *dst);
+int btrfs_lookup_bio_sums_dio(struct btrfs_root *root, struct inode *inode,
+			      struct bio *bio, u64 logical_offset, u32 *dst);
 int btrfs_insert_file_extent(struct btrfs_trans_handle *trans,
 			     struct btrfs_root *root,
 			     u64 objectid, u64 pos,
diff --git a/fs/btrfs/file-item.c b/fs/btrfs/file-item.c
index 21aead39a76..a562a250ae7 100644
--- a/fs/btrfs/file-item.c
+++ b/fs/btrfs/file-item.c
@@ -149,13 +149,14 @@ int btrfs_lookup_file_extent(struct btrfs_trans_handle *trans,
 }
 
 
-int btrfs_lookup_bio_sums(struct btrfs_root *root, struct inode *inode,
-			  struct bio *bio, u32 *dst)
+static int __btrfs_lookup_bio_sums(struct btrfs_root *root,
+				   struct inode *inode, struct bio *bio,
+				   u64 logical_offset, u32 *dst, int dio)
 {
 	u32 sum;
 	struct bio_vec *bvec = bio->bi_io_vec;
 	int bio_index = 0;
-	u64 offset;
+	u64 offset = 0;
 	u64 item_start_offset = 0;
 	u64 item_last_offset = 0;
 	u64 disk_bytenr;
@@ -174,8 +175,11 @@ int btrfs_lookup_bio_sums(struct btrfs_root *root, struct inode *inode,
 	WARN_ON(bio->bi_vcnt <= 0);
 
 	disk_bytenr = (u64)bio->bi_sector << 9;
+	if (dio)
+		offset = logical_offset;
 	while (bio_index < bio->bi_vcnt) {
-		offset = page_offset(bvec->bv_page) + bvec->bv_offset;
+		if (!dio)
+			offset = page_offset(bvec->bv_page) + bvec->bv_offset;
 		ret = btrfs_find_ordered_sum(inode, offset, disk_bytenr, &sum);
 		if (ret == 0)
 			goto found;
@@ -238,6 +242,7 @@ found:
 		else
 			set_state_private(io_tree, offset, sum);
 		disk_bytenr += bvec->bv_len;
+		offset += bvec->bv_len;
 		bio_index++;
 		bvec++;
 	}
@@ -245,6 +250,18 @@ found:
 	return 0;
 }
 
+int btrfs_lookup_bio_sums(struct btrfs_root *root, struct inode *inode,
+			  struct bio *bio, u32 *dst)
+{
+	return __btrfs_lookup_bio_sums(root, inode, bio, 0, dst, 0);
+}
+
+int btrfs_lookup_bio_sums_dio(struct btrfs_root *root, struct inode *inode,
+			      struct bio *bio, u64 offset, u32 *dst)
+{
+	return __btrfs_lookup_bio_sums(root, inode, bio, offset, dst, 1);
+}
+
 int btrfs_lookup_csums_range(struct btrfs_root *root, u64 start, u64 end,
 			     struct list_head *list)
 {
diff --git a/fs/btrfs/file.c b/fs/btrfs/file.c
index 6d8f817eadb..a28810abfb9 100644
--- a/fs/btrfs/file.c
+++ b/fs/btrfs/file.c
@@ -822,6 +822,47 @@ again:
 	return 0;
 }
 
+/* Copied from read-write.c */
+static void wait_on_retry_sync_kiocb(struct kiocb *iocb)
+{
+	set_current_state(TASK_UNINTERRUPTIBLE);
+	if (!kiocbIsKicked(iocb))
+		schedule();
+	else
+		kiocbClearKicked(iocb);
+	__set_current_state(TASK_RUNNING);
+}
+
+/*
+ * Just a copy of what do_sync_write does.
+ */
+static ssize_t __btrfs_direct_write(struct file *file, const char __user *buf,
+				    size_t count, loff_t pos, loff_t *ppos)
+{
+	struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
+	unsigned long nr_segs = 1;
+	struct kiocb kiocb;
+	ssize_t ret;
+
+	init_sync_kiocb(&kiocb, file);
+	kiocb.ki_pos = pos;
+	kiocb.ki_left = count;
+	kiocb.ki_nbytes = count;
+
+	while (1) {
+		ret = generic_file_direct_write(&kiocb, &iov, &nr_segs, pos,
+						ppos, count, count);
+		if (ret != -EIOCBRETRY)
+			break;
+		wait_on_retry_sync_kiocb(&kiocb);
+	}
+
+	if (ret == -EIOCBQUEUED)
+		ret = wait_on_sync_kiocb(&kiocb);
+	*ppos = kiocb.ki_pos;
+	return ret;
+}
+
 static ssize_t btrfs_file_write(struct file *file, const char __user *buf,
 				size_t count, loff_t *ppos)
 {
@@ -838,12 +879,11 @@ static ssize_t btrfs_file_write(struct file *file, const char __user *buf,
 	unsigned long first_index;
 	unsigned long last_index;
 	int will_write;
+	int buffered = 0;
 
 	will_write = ((file->f_flags & O_DSYNC) || IS_SYNC(inode) ||
 		      (file->f_flags & O_DIRECT));
 
-	nrptrs = min((count + PAGE_CACHE_SIZE - 1) / PAGE_CACHE_SIZE,
-		     PAGE_CACHE_SIZE / (sizeof(struct page *)));
 	pinned[0] = NULL;
 	pinned[1] = NULL;
 
@@ -867,13 +907,34 @@ static ssize_t btrfs_file_write(struct file *file, const char __user *buf,
 		goto out;
 
 	file_update_time(file);
+	BTRFS_I(inode)->sequence++;
+
+	if (unlikely(file->f_flags & O_DIRECT)) {
+		num_written = __btrfs_direct_write(file, buf, count, pos,
+						   ppos);
+		pos += num_written;
+		count -= num_written;
+
+		/* We've written everything we wanted to, exit */
+		if (num_written < 0 || !count)
+			goto out;
 
+		/*
+		 * We are going to do buffered for the rest of the range, so we
+		 * need to make sure to invalidate the buffered pages when we're
+		 * done.
+		 */
+		buffered = 1;
+		buf += num_written;
+	}
+
+	nrptrs = min((count + PAGE_CACHE_SIZE - 1) / PAGE_CACHE_SIZE,
+		     PAGE_CACHE_SIZE / (sizeof(struct page *)));
 	pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
 
 	/* generic_write_checks can change our pos */
 	start_pos = pos;
 
-	BTRFS_I(inode)->sequence++;
 	first_index = pos >> PAGE_CACHE_SHIFT;
 	last_index = (pos + count) >> PAGE_CACHE_SHIFT;
 
@@ -1007,7 +1068,7 @@ out:
 				btrfs_end_transaction(trans, root);
 			}
 		}
-		if (file->f_flags & O_DIRECT) {
+		if (file->f_flags & O_DIRECT && buffered) {
 			invalidate_mapping_pages(inode->i_mapping,
 			      start_pos >> PAGE_CACHE_SHIFT,
 			     (start_pos + num_written - 1) >> PAGE_CACHE_SHIFT);
diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index 460dd512eeb..1695440a59a 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c
@@ -698,6 +698,38 @@ retry:
 	return 0;
 }
 
+static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
+				      u64 num_bytes)
+{
+	struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
+	struct extent_map *em;
+	u64 alloc_hint = 0;
+
+	read_lock(&em_tree->lock);
+	em = search_extent_mapping(em_tree, start, num_bytes);
+	if (em) {
+		/*
+		 * if block start isn't an actual block number then find the
+		 * first block in this inode and use that as a hint.  If that
+		 * block is also bogus then just don't worry about it.
+		 */
+		if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
+			free_extent_map(em);
+			em = search_extent_mapping(em_tree, 0, 0);
+			if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
+				alloc_hint = em->block_start;
+			if (em)
+				free_extent_map(em);
+		} else {
+			alloc_hint = em->block_start;
+			free_extent_map(em);
+		}
+	}
+	read_unlock(&em_tree->lock);
+
+	return alloc_hint;
+}
+
 /*
  * when extent_io.c finds a delayed allocation range in the file,
  * the call backs end up in this code.  The basic idea is to
@@ -770,29 +802,7 @@ static noinline int cow_file_range(struct inode *inode,
 	BUG_ON(disk_num_bytes >
 	       btrfs_super_total_bytes(&root->fs_info->super_copy));
 
-
-	read_lock(&BTRFS_I(inode)->extent_tree.lock);
-	em = search_extent_mapping(&BTRFS_I(inode)->extent_tree,
-				   start, num_bytes);
-	if (em) {
-		/*
-		 * if block start isn't an actual block number then find the
-		 * first block in this inode and use that as a hint.  If that
-		 * block is also bogus then just don't worry about it.
-		 */
-		if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
-			free_extent_map(em);
-			em = search_extent_mapping(em_tree, 0, 0);
-			if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
-				alloc_hint = em->block_start;
-			if (em)
-				free_extent_map(em);
-		} else {
-			alloc_hint = em->block_start;
-			free_extent_map(em);
-		}
-	}
-	read_unlock(&BTRFS_I(inode)->extent_tree.lock);
+	alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 	btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 
 	while (disk_num_bytes > 0) {
@@ -5171,11 +5181,440 @@ out:
 	return em;
 }
 
+static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
+						  u64 start, u64 len)
+{
+	struct btrfs_root *root = BTRFS_I(inode)->root;
+	struct btrfs_trans_handle *trans;
+	struct extent_map *em;
+	struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
+	struct btrfs_key ins;
+	u64 alloc_hint;
+	int ret;
+
+	btrfs_drop_extent_cache(inode, start, start + len - 1, 0);
+
+	trans = btrfs_join_transaction(root, 0);
+	if (!trans)
+		return ERR_PTR(-ENOMEM);
+
+	trans->block_rsv = &root->fs_info->delalloc_block_rsv;
+
+	alloc_hint = get_extent_allocation_hint(inode, start, len);
+	ret = btrfs_reserve_extent(trans, root, len, root->sectorsize, 0,
+				   alloc_hint, (u64)-1, &ins, 1);
+	if (ret) {
+		em = ERR_PTR(ret);
+		goto out;
+	}
+
+	em = alloc_extent_map(GFP_NOFS);
+	if (!em) {
+		em = ERR_PTR(-ENOMEM);
+		goto out;
+	}
+
+	em->start = start;
+	em->orig_start = em->start;
+	em->len = ins.offset;
+
+	em->block_start = ins.objectid;
+	em->block_len = ins.offset;
+	em->bdev = root->fs_info->fs_devices->latest_bdev;
+	set_bit(EXTENT_FLAG_PINNED, &em->flags);
+
+	while (1) {
+		write_lock(&em_tree->lock);
+		ret = add_extent_mapping(em_tree, em);
+		write_unlock(&em_tree->lock);
+		if (ret != -EEXIST)
+			break;
+		btrfs_drop_extent_cache(inode, start, start + em->len - 1, 0);
+	}
+
+	ret = btrfs_add_ordered_extent_dio(inode, start, ins.objectid,
+					   ins.offset, ins.offset, 0);
+	if (ret) {
+		btrfs_free_reserved_extent(root, ins.objectid, ins.offset);
+		em = ERR_PTR(ret);
+	}
+out:
+	btrfs_end_transaction(trans, root);
+	return em;
+}
+
+static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
+				   struct buffer_head *bh_result, int create)
+{
+	struct extent_map *em;
+	struct btrfs_root *root = BTRFS_I(inode)->root;
+	u64 start = iblock << inode->i_blkbits;
+	u64 len = bh_result->b_size;
+
+	em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
+	if (IS_ERR(em))
+		return PTR_ERR(em);
+
+	/*
+	 * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
+	 * io.  INLINE is special, and we could probably kludge it in here, but
+	 * it's still buffered so for safety lets just fall back to the generic
+	 * buffered path.
+	 *
+	 * For COMPRESSED we _have_ to read the entire extent in so we can
+	 * decompress it, so there will be buffering required no matter what we
+	 * do, so go ahead and fallback to buffered.
+	 *
+	 * We return -ENOTBLK because thats what makes DIO go ahead and go back
+	 * to buffered IO.  Don't blame me, this is the price we pay for using
+	 * the generic code.
+	 */
+	if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
+	    em->block_start == EXTENT_MAP_INLINE) {
+		free_extent_map(em);
+		return -ENOTBLK;
+	}
+
+	/* Just a good old fashioned hole, return */
+	if (!create && (em->block_start == EXTENT_MAP_HOLE ||
+			test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
+		free_extent_map(em);
+		/* DIO will do one hole at a time, so just unlock a sector */
+		unlock_extent(&BTRFS_I(inode)->io_tree, start,
+			      start + root->sectorsize - 1, GFP_NOFS);
+		return 0;
+	}
+
+	/*
+	 * We don't allocate a new extent in the following cases
+	 *
+	 * 1) The inode is marked as NODATACOW.  In this case we'll just use the
+	 * existing extent.
+	 * 2) The extent is marked as PREALLOC.  We're good to go here and can
+	 * just use the extent.
+	 *
+	 */
+	if (!create)
+		goto map;
+
+	if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
+	    ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
+	     em->block_start != EXTENT_MAP_HOLE)) {
+		u64 block_start;
+		int type;
+		int ret;
+
+		if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
+			type = BTRFS_ORDERED_PREALLOC;
+		else
+			type = BTRFS_ORDERED_NOCOW;
+		len = min(len, em->block_len - (start - em->start));
+		block_start = em->block_start + (start - em->start);
+		ret = btrfs_add_ordered_extent_dio(inode, start,
+						   start, len, len, type);
+		if (ret) {
+			free_extent_map(em);
+			return ret;
+		}
+	} else {
+		free_extent_map(em);
+		em = btrfs_new_extent_direct(inode, start, len);
+		if (IS_ERR(em))
+			return PTR_ERR(em);
+		len = min(len, em->block_len);
+	}
+	unlock_extent(&BTRFS_I(inode)->io_tree, start, start + len - 1,
+		      GFP_NOFS);
+map:
+	bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
+		inode->i_blkbits;
+	bh_result->b_size = em->len - (start - em->start);
+	bh_result->b_bdev = em->bdev;
+	set_buffer_mapped(bh_result);
+	if (create && !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
+		set_buffer_new(bh_result);
+
+	free_extent_map(em);
+
+	return 0;
+}
+
+struct btrfs_dio_private {
+	struct inode *inode;
+	u64 logical_offset;
+	u64 disk_bytenr;
+	u64 bytes;
+	u32 *csums;
+	void *private;
+};
+
+static void btrfs_endio_direct_read(struct bio *bio, int err)
+{
+	struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
+	struct bio_vec *bvec = bio->bi_io_vec;
+	struct btrfs_dio_private *dip = bio->bi_private;
+	struct inode *inode = dip->inode;
+	struct btrfs_root *root = BTRFS_I(inode)->root;
+	u64 start;
+	u32 *private = dip->csums;
+
+	start = dip->logical_offset;
+	do {
+		if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
+			struct page *page = bvec->bv_page;
+			char *kaddr;
+			u32 csum = ~(u32)0;
+			unsigned long flags;
+
+			local_irq_save(flags);
+			kaddr = kmap_atomic(page, KM_IRQ0);
+			csum = btrfs_csum_data(root, kaddr + bvec->bv_offset,
+					       csum, bvec->bv_len);
+			btrfs_csum_final(csum, (char *)&csum);
+			kunmap_atomic(kaddr, KM_IRQ0);
+			local_irq_restore(flags);
+
+			flush_dcache_page(bvec->bv_page);
+			if (csum != *private) {
+				printk(KERN_ERR "btrfs csum failed ino %lu off"
+				      " %llu csum %u private %u\n",
+				      inode->i_ino, (unsigned long long)start,
+				      csum, *private);
+				err = -EIO;
+			}
+		}
+
+		start += bvec->bv_len;
+		private++;
+		bvec++;
+	} while (bvec <= bvec_end);
+
+	unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
+		      dip->logical_offset + dip->bytes - 1, GFP_NOFS);
+	bio->bi_private = dip->private;
+
+	kfree(dip->csums);
+	kfree(dip);
+	dio_end_io(bio, err);
+}
+
+static void btrfs_endio_direct_write(struct bio *bio, int err)
+{
+	struct btrfs_dio_private *dip = bio->bi_private;
+	struct inode *inode = dip->inode;
+	struct btrfs_root *root = BTRFS_I(inode)->root;
+	struct btrfs_trans_handle *trans;
+	struct btrfs_ordered_extent *ordered = NULL;
+	struct extent_state *cached_state = NULL;
+	int ret;
+
+	if (err)
+		goto out_done;
+
+	ret = btrfs_dec_test_ordered_pending(inode, &ordered,
+					     dip->logical_offset, dip->bytes);
+	if (!ret)
+		goto out_done;
+
+	BUG_ON(!ordered);
+
+	trans = btrfs_join_transaction(root, 1);
+	if (!trans) {
+		err = -ENOMEM;
+		goto out;
+	}
+	trans->block_rsv = &root->fs_info->delalloc_block_rsv;
+
+	if (test_bit(BTRFS_ORDERED_NOCOW, &ordered->flags)) {
+		ret = btrfs_ordered_update_i_size(inode, 0, ordered);
+		if (!ret)
+			ret = btrfs_update_inode(trans, root, inode);
+		err = ret;
+		goto out;
+	}
+
+	lock_extent_bits(&BTRFS_I(inode)->io_tree, ordered->file_offset,
+			 ordered->file_offset + ordered->len - 1, 0,
+			 &cached_state, GFP_NOFS);
+
+	if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered->flags)) {
+		ret = btrfs_mark_extent_written(trans, inode,
+						ordered->file_offset,
+						ordered->file_offset +
+						ordered->len);
+		if (ret) {
+			err = ret;
+			goto out_unlock;
+		}
+	} else {
+		ret = insert_reserved_file_extent(trans, inode,
+						  ordered->file_offset,
+						  ordered->start,
+						  ordered->disk_len,
+						  ordered->len,
+						  ordered->len,
+						  0, 0, 0,
+						  BTRFS_FILE_EXTENT_REG);
+		unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
+				   ordered->file_offset, ordered->len);
+		if (ret) {
+			err = ret;
+			WARN_ON(1);
+			goto out_unlock;
+		}
+	}
+
+	add_pending_csums(trans, inode, ordered->file_offset, &ordered->list);
+	btrfs_ordered_update_i_size(inode, 0, ordered);
+	btrfs_update_inode(trans, root, inode);
+out_unlock:
+	unlock_extent_cached(&BTRFS_I(inode)->io_tree, ordered->file_offset,
+			     ordered->file_offset + ordered->len - 1,
+			     &cached_state, GFP_NOFS);
+out:
+	btrfs_delalloc_release_metadata(inode, ordered->len);
+	btrfs_end_transaction(trans, root);
+	btrfs_put_ordered_extent(ordered);
+	btrfs_put_ordered_extent(ordered);
+out_done:
+	bio->bi_private = dip->private;
+
+	kfree(dip->csums);
+	kfree(dip);
+	dio_end_io(bio, err);
+}
+
+static void btrfs_submit_direct(int rw, struct bio *bio, struct inode *inode,
+				loff_t file_offset)
+{
+	struct btrfs_root *root = BTRFS_I(inode)->root;
+	struct btrfs_dio_private *dip;
+	struct bio_vec *bvec = bio->bi_io_vec;
+	u64 start;
+	int skip_sum;
+	int write = rw & (1 << BIO_RW);
+	int ret = 0;
+
+	skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
+
+	dip = kmalloc(sizeof(*dip), GFP_NOFS);
+	if (!dip) {
+		ret = -ENOMEM;
+		goto free_ordered;
+	}
+	dip->csums = NULL;
+
+	if (!skip_sum) {
+		dip->csums = kmalloc(sizeof(u32) * bio->bi_vcnt, GFP_NOFS);
+		if (!dip->csums) {
+			ret = -ENOMEM;
+			goto free_ordered;
+		}
+	}
+
+	dip->private = bio->bi_private;
+	dip->inode = inode;
+	dip->logical_offset = file_offset;
+
+	start = dip->logical_offset;
+	dip->bytes = 0;
+	do {
+		dip->bytes += bvec->bv_len;
+		bvec++;
+	} while (bvec <= (bio->bi_io_vec + bio->bi_vcnt - 1));
+
+	dip->disk_bytenr = bio->bi_sector << 9;
+	bio->bi_private = dip;
+
+	if (write)
+		bio->bi_end_io = btrfs_endio_direct_write;
+	else
+		bio->bi_end_io = btrfs_endio_direct_read;
+
+	ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
+	if (ret)
+		goto out_err;
+
+	if (write && !skip_sum)
+		btrfs_csum_one_bio(root, inode, bio, dip->logical_offset, 1);
+	else if (!skip_sum)
+		btrfs_lookup_bio_sums_dio(root, inode, bio,
+					  dip->logical_offset, dip->csums);
+
+	ret = btrfs_map_bio(root, rw, bio, 0, 0);
+	if (ret)
+		goto out_err;
+	return;
+out_err:
+	kfree(dip->csums);
+	kfree(dip);
+free_ordered:
+	/*
+	 * If this is a write, we need to clean up the reserved space and kill
+	 * the ordered extent.
+	 */
+	if (write) {
+		struct btrfs_ordered_extent *ordered;
+		ordered = btrfs_lookup_ordered_extent(inode,
+						      dip->logical_offset);
+		if (!test_bit(BTRFS_ORDERED_PREALLOC, &ordered->flags) &&
+		    !test_bit(BTRFS_ORDERED_NOCOW, &ordered->flags))
+			btrfs_free_reserved_extent(root, ordered->start,
+						   ordered->disk_len);
+		btrfs_put_ordered_extent(ordered);
+		btrfs_put_ordered_extent(ordered);
+	}
+	bio_endio(bio, ret);
+}
+
 static ssize_t btrfs_direct_IO(int rw, struct kiocb *iocb,
 			const struct iovec *iov, loff_t offset,
 			unsigned long nr_segs)
 {
-	return -EINVAL;
+	struct file *file = iocb->ki_filp;
+	struct inode *inode = file->f_mapping->host;
+	struct btrfs_ordered_extent *ordered;
+	u64 lockstart, lockend;
+	ssize_t ret;
+
+	lockstart = offset;
+	lockend = offset + iov_length(iov, nr_segs) - 1;
+	while (1) {
+		lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
+			    GFP_NOFS);
+		/*
+		 * We're concerned with the entire range that we're going to be
+		 * doing DIO to, so we need to make sure theres no ordered
+		 * extents in this range.
+		 */
+		ordered = btrfs_lookup_ordered_range(inode, lockstart,
+						     lockend - lockstart + 1);
+		if (!ordered)
+			break;
+		unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
+			      GFP_NOFS);
+		btrfs_start_ordered_extent(inode, ordered, 1);
+		btrfs_put_ordered_extent(ordered);
+		cond_resched();
+	}
+
+	ret = __blockdev_direct_IO(rw, iocb, inode, NULL, iov, offset, nr_segs,
+				   btrfs_get_blocks_direct, NULL,
+				   btrfs_submit_direct, 0);
+
+	if (ret < 0 && ret != -EIOCBQUEUED) {
+		unlock_extent(&BTRFS_I(inode)->io_tree, offset,
+			      offset + iov_length(iov, nr_segs) - 1, GFP_NOFS);
+	} else if (ret >= 0 && ret < iov_length(iov, nr_segs)) {
+		/*
+		 * We're falling back to buffered, unlock the section we didn't
+		 * do IO on.
+		 */
+		unlock_extent(&BTRFS_I(inode)->io_tree, offset + ret,
+			      offset + iov_length(iov, nr_segs) - 1, GFP_NOFS);
+	}
+
+	return ret;
 }
 
 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
diff --git a/fs/btrfs/ordered-data.c b/fs/btrfs/ordered-data.c
index c9f1020572f..e56c72bc5ad 100644
--- a/fs/btrfs/ordered-data.c
+++ b/fs/btrfs/ordered-data.c
@@ -124,6 +124,15 @@ static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
 	return 1;
 }
 
+static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
+			  u64 len)
+{
+	if (file_offset + len <= entry->file_offset ||
+	    entry->file_offset + entry->len <= file_offset)
+		return 0;
+	return 1;
+}
+
 /*
  * look find the first ordered struct that has this offset, otherwise
  * the first one less than this offset
@@ -161,8 +170,9 @@ static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
  * The tree is given a single reference on the ordered extent that was
  * inserted.
  */
-int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
-			     u64 start, u64 len, u64 disk_len, int type)
+static int __btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
+				      u64 start, u64 len, u64 disk_len,
+				      int type, int dio)
 {
 	struct btrfs_ordered_inode_tree *tree;
 	struct rb_node *node;
@@ -182,6 +192,9 @@ int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
 	if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
 		set_bit(type, &entry->flags);
 
+	if (dio)
+		set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
+
 	/* one ref for the tree */
 	atomic_set(&entry->refs, 1);
 	init_waitqueue_head(&entry->wait);
@@ -203,6 +216,20 @@ int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
 	return 0;
 }
 
+int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
+			     u64 start, u64 len, u64 disk_len, int type)
+{
+	return __btrfs_add_ordered_extent(inode, file_offset, start, len,
+					  disk_len, type, 0);
+}
+
+int btrfs_add_ordered_extent_dio(struct inode *inode, u64 file_offset,
+				 u64 start, u64 len, u64 disk_len, int type)
+{
+	return __btrfs_add_ordered_extent(inode, file_offset, start, len,
+					  disk_len, type, 1);
+}
+
 /*
  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
  * when an ordered extent is finished.  If the list covers more than one
@@ -484,7 +511,8 @@ void btrfs_start_ordered_extent(struct inode *inode,
 	 * start IO on any dirty ones so the wait doesn't stall waiting
 	 * for pdflush to find them
 	 */
-	filemap_fdatawrite_range(inode->i_mapping, start, end);
+	if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
+		filemap_fdatawrite_range(inode->i_mapping, start, end);
 	if (wait) {
 		wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
 						 &entry->flags));
@@ -581,6 +609,47 @@ out:
 	return entry;
 }
 
+/* Since the DIO code tries to lock a wide area we need to look for any ordered
+ * extents that exist in the range, rather than just the start of the range.
+ */
+struct btrfs_ordered_extent *btrfs_lookup_ordered_range(struct inode *inode,
+							u64 file_offset,
+							u64 len)
+{
+	struct btrfs_ordered_inode_tree *tree;
+	struct rb_node *node;
+	struct btrfs_ordered_extent *entry = NULL;
+
+	tree = &BTRFS_I(inode)->ordered_tree;
+	spin_lock(&tree->lock);
+	node = tree_search(tree, file_offset);
+	if (!node) {
+		node = tree_search(tree, file_offset + len);
+		if (!node)
+			goto out;
+	}
+
+	while (1) {
+		entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
+		if (range_overlaps(entry, file_offset, len))
+			break;
+
+		if (entry->file_offset >= file_offset + len) {
+			entry = NULL;
+			break;
+		}
+		entry = NULL;
+		node = rb_next(node);
+		if (!node)
+			break;
+	}
+out:
+	if (entry)
+		atomic_inc(&entry->refs);
+	spin_unlock(&tree->lock);
+	return entry;
+}
+
 /*
  * lookup and return any extent before 'file_offset'.  NULL is returned
  * if none is found
diff --git a/fs/btrfs/ordered-data.h b/fs/btrfs/ordered-data.h
index c82f76a9f04..8ac365492a3 100644
--- a/fs/btrfs/ordered-data.h
+++ b/fs/btrfs/ordered-data.h
@@ -72,6 +72,8 @@ struct btrfs_ordered_sum {
 
 #define BTRFS_ORDERED_PREALLOC 4 /* set when writing to prealloced extent */
 
+#define BTRFS_ORDERED_DIRECT 5 /* set when we're doing DIO with this extent */
+
 struct btrfs_ordered_extent {
 	/* logical offset in the file */
 	u64 file_offset;
@@ -140,7 +142,9 @@ int btrfs_dec_test_ordered_pending(struct inode *inode,
 				   struct btrfs_ordered_extent **cached,
 				   u64 file_offset, u64 io_size);
 int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
-			     u64 start, u64 len, u64 disk_len, int tyep);
+			     u64 start, u64 len, u64 disk_len, int type);
+int btrfs_add_ordered_extent_dio(struct inode *inode, u64 file_offset,
+				 u64 start, u64 len, u64 disk_len, int type);
 int btrfs_add_ordered_sum(struct inode *inode,
 			  struct btrfs_ordered_extent *entry,
 			  struct btrfs_ordered_sum *sum);
@@ -151,6 +155,9 @@ void btrfs_start_ordered_extent(struct inode *inode,
 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len);
 struct btrfs_ordered_extent *
 btrfs_lookup_first_ordered_extent(struct inode * inode, u64 file_offset);
+struct btrfs_ordered_extent *btrfs_lookup_ordered_range(struct inode *inode,
+							u64 file_offset,
+							u64 len);
 int btrfs_ordered_update_i_size(struct inode *inode, u64 offset,
 				struct btrfs_ordered_extent *ordered);
 int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr, u32 *sum);