[4/8] btrfs: Switch to iomap_dio_rw() for dio
diff mbox series

Message ID 20191213195750.32184-5-rgoldwyn@suse.de
State New
Headers show
Series
  • btrfs direct-io using iomap
Related show

Commit Message

Goldwyn Rodrigues Dec. 13, 2019, 7:57 p.m. UTC
From: Goldwyn Rodrigues <rgoldwyn@suse.com>

Switch from __blockdev_direct_IO() to iomap_dio_rw().
Rename btrfs_get_blocks_direct() to btrfs_dio_iomap_begin() and use it
as iomap_begin() for iomap direct I/O functions. This function
allocates and locks all the blocks required for the I/O.
btrfs_submit_direct() is used as the submit_io() hook for direct I/O
ops.

Since we need direct I/O reads to go through iomap_dio_rw(), we change
file_operations.read_iter() to a btrfs_file_read_iter() which calls
btrfs_direct_IO() for direct reads and falls back to
generic_file_buffered_read() for incomplete reads and buffered reads.

We don't need address_space.direct_IO() anymore so set it to noop.
Similarly, we don't need flags used in __blockdev_direct_IO(). iomap is
capable of direct I/O reads from a hole, so we don't need to return
-ENOENT.

BTRFS direct I/O is now done under i_rwsem, shared in case of
reads and exclusive in case of writes. This guards against simultaneous
truncates.

Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
---
 fs/btrfs/ctree.h |   1 +
 fs/btrfs/file.c  |  21 +++++-
 fs/btrfs/inode.c | 192 ++++++++++++++++++++++++++-----------------------------
 3 files changed, 110 insertions(+), 104 deletions(-)

Comments

Christoph Hellwig Dec. 21, 2019, 2:42 p.m. UTC | #1
So Ilooked into the "unlocked" direct I/O case, and I think the current
code using dio_sem is really sketchy.  What btrfs really needs to do is
take i_rwsem shared by default for direct writes, and only upgrade to
the exclusive lock when needed, similar to xfs and the WIP ext4 code.

While looking for that I also noticed two other things:

 - check_direct_IO looks pretty bogus
 - btrfs_direct_IO really should be split and folded into the two
   callers

Untested patches attached.  The first should probably go into a prep
patch, and the second could be folded into this one.
From bc285e440a50140beb456f11e545a049bdf51ec1 Mon Sep 17 00:00:00 2001
From: Christoph Hellwig <hch@lst.de>
Date: Sat, 21 Dec 2019 15:17:26 +0100
Subject: btrfs: remove direct I/O aligment checks

The direct I/O code itself already checks for the proper sector
size alignment, so remove the duplicate checks.  The remainder of
check_direct_IO is not ony needed for reads and can be moved to
file.c and outside of i_rwsem.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 fs/btrfs/file.c  | 34 +++++++++++++++++++++++++++-------
 fs/btrfs/inode.c | 37 -------------------------------------
 2 files changed, 27 insertions(+), 44 deletions(-)

diff --git a/fs/btrfs/file.c b/fs/btrfs/file.c
index a6d41d7bf362..0522f6d45a98 100644
--- a/fs/btrfs/file.c
+++ b/fs/btrfs/file.c
@@ -3444,21 +3444,41 @@ static int btrfs_file_open(struct inode *inode, struct file *filp)
 	return generic_file_open(inode, filp);
 }
 
-static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
+/*
+ * If there are duplicate iov_base's in this iovec, fall back to buffered I/O
+ * to avoid checksum errors.
+ */
+static bool btrfs_direct_read_ok(struct kiocb *iocb, struct iov_iter *iter)
 {
-	ssize_t ret = 0;
+	int seg, i;
 
-	if (iocb->ki_flags & IOCB_DIRECT) {
+	if (!iter_is_iovec(iter))
+		return true;
+
+	for (seg = 0; seg < iter->nr_segs; seg++) {
+		for (i = seg + 1; i < iter->nr_segs; i++) {
+			if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
+				return false;
+		}
+	}
+
+	return true;
+}
+
+
+static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
+{
+	if ((iocb->ki_flags & IOCB_DIRECT) && btrfs_direct_read_ok(iocb, to)) {
 		struct inode *inode = file_inode(iocb->ki_filp);
+		ssize_t ret;
 
 		inode_lock_shared(inode);
 		ret = btrfs_direct_IO(iocb, to);
 		inode_unlock_shared(inode);
-		if (ret < 0)
-			return ret;
-	}
 
-	return generic_file_buffered_read(iocb, to, ret);
+		return ret;
+	}
+	return generic_file_buffered_read(iocb, to, 0);
 }
 
 const struct file_operations btrfs_file_operations = {
diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index 824f318cee5e..18d153a62655 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c
@@ -8581,39 +8581,6 @@ static blk_qc_t btrfs_submit_direct(struct bio *dio_bio, struct file *file,
 	return BLK_QC_T_NONE;
 }
 
-static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
-			       const struct iov_iter *iter, loff_t offset)
-{
-	int seg;
-	int i;
-	unsigned int blocksize_mask = fs_info->sectorsize - 1;
-	ssize_t retval = -EINVAL;
-
-	if (offset & blocksize_mask)
-		goto out;
-
-	if (iov_iter_alignment(iter) & blocksize_mask)
-		goto out;
-
-	/* If this is a write we don't need to check anymore */
-	if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
-		return 0;
-	/*
-	 * Check to make sure we don't have duplicate iov_base's in this
-	 * iovec, if so return EINVAL, otherwise we'll get csum errors
-	 * when reading back.
-	 */
-	for (seg = 0; seg < iter->nr_segs; seg++) {
-		for (i = seg + 1; i < iter->nr_segs; i++) {
-			if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
-				goto out;
-		}
-	}
-	retval = 0;
-out:
-	return retval;
-}
-
 static const struct iomap_ops btrfs_dio_iomap_ops = {
 	.iomap_begin            = btrfs_dio_iomap_begin,
 	.iomap_end		= btrfs_dio_iomap_end,
@@ -8635,7 +8602,6 @@ ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 {
 	struct file *file = iocb->ki_filp;
 	struct inode *inode = file->f_mapping->host;
-	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 	struct extent_changeset *data_reserved = NULL;
 	loff_t offset = iocb->ki_pos;
 	size_t count = 0;
@@ -8644,9 +8610,6 @@ ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 
 	lockdep_assert_held(&inode->i_rwsem);
 
-	if (check_direct_IO(fs_info, iter, offset))
-		return 0;
-
 	count = iov_iter_count(iter);
 	if (iov_iter_rw(iter) == WRITE) {
 		/*
Goldwyn Rodrigues Jan. 2, 2020, 6:01 p.m. UTC | #2
On  6:42 21/12, Christoph Hellwig wrote:
> So Ilooked into the "unlocked" direct I/O case, and I think the current
> code using dio_sem is really sketchy.  What btrfs really needs to do is
> take i_rwsem shared by default for direct writes, and only upgrade to
> the exclusive lock when needed, similar to xfs and the WIP ext4 code.

Sketchy in what sense? I am not trying to second-guess, but I want to
know where it could fail. I would want it to be simpler as well, but if
we can perform direct writes without locking, why should we introduce
locks.

> 
> While looking for that I also noticed two other things:
> 
>  - check_direct_IO looks pretty bogus
>  - btrfs_direct_IO really should be split and folded into the two
>    callers

Thanks for the cleanups. I will incorporate these.

> 
> Untested patches attached.  The first should probably go into a prep
> patch, and the second could be folded into this one.

> From bc285e440a50140beb456f11e545a049bdf51ec1 Mon Sep 17 00:00:00 2001
> From: Christoph Hellwig <hch@lst.de>
> Date: Sat, 21 Dec 2019 15:17:26 +0100
> Subject: btrfs: remove direct I/O aligment checks
> 
> The direct I/O code itself already checks for the proper sector
> size alignment, so remove the duplicate checks.  The remainder of
> check_direct_IO is not ony needed for reads and can be moved to
> file.c and outside of i_rwsem.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> ---
>  fs/btrfs/file.c  | 34 +++++++++++++++++++++++++++-------
>  fs/btrfs/inode.c | 37 -------------------------------------
>  2 files changed, 27 insertions(+), 44 deletions(-)
> 
> diff --git a/fs/btrfs/file.c b/fs/btrfs/file.c
> index a6d41d7bf362..0522f6d45a98 100644
> --- a/fs/btrfs/file.c
> +++ b/fs/btrfs/file.c
> @@ -3444,21 +3444,41 @@ static int btrfs_file_open(struct inode *inode, struct file *filp)
>  	return generic_file_open(inode, filp);
>  }
>  
> -static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
> +/*
> + * If there are duplicate iov_base's in this iovec, fall back to buffered I/O
> + * to avoid checksum errors.
> + */
> +static bool btrfs_direct_read_ok(struct kiocb *iocb, struct iov_iter *iter)
>  {
> -	ssize_t ret = 0;
> +	int seg, i;
>  
> -	if (iocb->ki_flags & IOCB_DIRECT) {
> +	if (!iter_is_iovec(iter))
> +		return true;
> +
> +	for (seg = 0; seg < iter->nr_segs; seg++) {
> +		for (i = seg + 1; i < iter->nr_segs; i++) {
> +			if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
> +				return false;
> +		}
> +	}
> +
> +	return true;
> +}
> +
> +
> +static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
> +{
> +	if ((iocb->ki_flags & IOCB_DIRECT) && btrfs_direct_read_ok(iocb, to)) {
>  		struct inode *inode = file_inode(iocb->ki_filp);
> +		ssize_t ret;
>  
>  		inode_lock_shared(inode);
>  		ret = btrfs_direct_IO(iocb, to);
>  		inode_unlock_shared(inode);
> -		if (ret < 0)
> -			return ret;
> -	}
>  
> -	return generic_file_buffered_read(iocb, to, ret);
> +		return ret;
> +	}
> +	return generic_file_buffered_read(iocb, to, 0);
>  }
>  
>  const struct file_operations btrfs_file_operations = {
> diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
> index 824f318cee5e..18d153a62655 100644
> --- a/fs/btrfs/inode.c
> +++ b/fs/btrfs/inode.c
> @@ -8581,39 +8581,6 @@ static blk_qc_t btrfs_submit_direct(struct bio *dio_bio, struct file *file,
>  	return BLK_QC_T_NONE;
>  }
>  
> -static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
> -			       const struct iov_iter *iter, loff_t offset)
> -{
> -	int seg;
> -	int i;
> -	unsigned int blocksize_mask = fs_info->sectorsize - 1;
> -	ssize_t retval = -EINVAL;
> -
> -	if (offset & blocksize_mask)
> -		goto out;
> -
> -	if (iov_iter_alignment(iter) & blocksize_mask)
> -		goto out;
> -
> -	/* If this is a write we don't need to check anymore */
> -	if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
> -		return 0;
> -	/*
> -	 * Check to make sure we don't have duplicate iov_base's in this
> -	 * iovec, if so return EINVAL, otherwise we'll get csum errors
> -	 * when reading back.
> -	 */
> -	for (seg = 0; seg < iter->nr_segs; seg++) {
> -		for (i = seg + 1; i < iter->nr_segs; i++) {
> -			if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
> -				goto out;
> -		}
> -	}
> -	retval = 0;
> -out:
> -	return retval;
> -}
> -
>  static const struct iomap_ops btrfs_dio_iomap_ops = {
>  	.iomap_begin            = btrfs_dio_iomap_begin,
>  	.iomap_end		= btrfs_dio_iomap_end,
> @@ -8635,7 +8602,6 @@ ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
>  {
>  	struct file *file = iocb->ki_filp;
>  	struct inode *inode = file->f_mapping->host;
> -	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
>  	struct extent_changeset *data_reserved = NULL;
>  	loff_t offset = iocb->ki_pos;
>  	size_t count = 0;
> @@ -8644,9 +8610,6 @@ ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
>  
>  	lockdep_assert_held(&inode->i_rwsem);
>  
> -	if (check_direct_IO(fs_info, iter, offset))
> -		return 0;
> -
>  	count = iov_iter_count(iter);
>  	if (iov_iter_rw(iter) == WRITE) {
>  		/*
> -- 
> 2.24.0
> 

> From 7194fa1986a48af46d2b01457865066cdbd14e35 Mon Sep 17 00:00:00 2001
> From: Christoph Hellwig <hch@lst.de>
> Date: Sat, 21 Dec 2019 15:23:41 +0100
> Subject: btrfs: split btrfs_direct_IO
> 
> The read and write versions don't have anything in common except
> for the call to iomap_dio_rw.  So split this function, and merge
> each half into its only caller.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> ---
>  fs/btrfs/ctree.h |  4 ++-
>  fs/btrfs/file.c  | 44 +++++++++++++++++++++++++----
>  fs/btrfs/inode.c | 72 ++++--------------------------------------------
>  3 files changed, 48 insertions(+), 72 deletions(-)
> 
> diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h
> index 8faa069b0a73..fccbbfebdf88 100644
> --- a/fs/btrfs/ctree.h
> +++ b/fs/btrfs/ctree.h
> @@ -28,6 +28,7 @@
>  #include <linux/dynamic_debug.h>
>  #include <linux/refcount.h>
>  #include <linux/crc32c.h>
> +#include <linux/iomap.h>
>  #include "extent-io-tree.h"
>  #include "extent_io.h"
>  #include "extent_map.h"
> @@ -2904,7 +2905,8 @@ int btrfs_writepage_cow_fixup(struct page *page, u64 start, u64 end);
>  void btrfs_writepage_endio_finish_ordered(struct page *page, u64 start,
>  					  u64 end, int uptodate);
>  extern const struct dentry_operations btrfs_dentry_operations;
> -ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
> +const struct iomap_ops btrfs_dio_iomap_ops;
> +const struct iomap_dio_ops btrfs_dio_ops;
>  
>  /* ioctl.c */
>  long btrfs_ioctl(struct file *file, unsigned int cmd, unsigned long arg);
> diff --git a/fs/btrfs/file.c b/fs/btrfs/file.c
> index 0522f6d45a98..ed0b2e015d8d 100644
> --- a/fs/btrfs/file.c
> +++ b/fs/btrfs/file.c
> @@ -1822,17 +1822,50 @@ static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
>  	return num_written ? num_written : ret;
>  }
>  
> -static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
> +static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
>  {
>  	struct file *file = iocb->ki_filp;
>  	struct inode *inode = file_inode(file);
> -	loff_t pos;
> +	size_t count = iov_iter_count(from);
> +	struct extent_changeset *data_reserved = NULL;
> +	loff_t pos = iocb->ki_pos;
>  	ssize_t written;
>  	ssize_t written_buffered;
>  	loff_t endbyte;
> +	bool relock = false;
>  	int err;
>  
> -	written = btrfs_direct_IO(iocb, from);
> +	/*
> +	 * If the write DIO is beyond the EOF, we need update the isize, but
> +	 * it is protected by i_mutex. So we can not unlock the i_mutex in
> +	 * this case.
> +	 */
> +	if (pos + count <= inode->i_size) {
> +		inode_unlock(inode);
> +		relock = true;
> +	} else {
> +		if (iocb->ki_flags & IOCB_NOWAIT)
> +			return -EAGAIN;
> +	}
> +
> +	err = btrfs_delalloc_reserve_space(inode, &data_reserved, pos, count);
> +	if (err) {
> +		if (relock)
> +			inode_lock(inode);
> +		return err;
> +	}
> +
> +	down_read(&BTRFS_I(inode)->dio_sem);
> +	written = iomap_dio_rw(iocb, from, &btrfs_dio_iomap_ops, &btrfs_dio_ops,
> +			is_sync_kiocb(iocb));
> +	up_read(&BTRFS_I(inode)->dio_sem);
> +	if (written >= 0 && (size_t)written < count)
> +		btrfs_delalloc_release_space(inode, data_reserved,
> +				pos, count - (size_t)written, true);
> +	btrfs_delalloc_release_extents(BTRFS_I(inode), count);
> +	if (relock)
> +		inode_lock(inode);
> +	extent_changeset_free(data_reserved);
>  
>  	if (written < 0 || !iov_iter_count(from))
>  		return written;
> @@ -1975,7 +2008,7 @@ static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
>  		atomic_inc(&BTRFS_I(inode)->sync_writers);
>  
>  	if (iocb->ki_flags & IOCB_DIRECT) {
> -		num_written = __btrfs_direct_write(iocb, from);
> +		num_written = btrfs_direct_write(iocb, from);
>  	} else {
>  		num_written = btrfs_buffered_write(iocb, from);
>  		if (num_written > 0)
> @@ -3473,7 +3506,8 @@ static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
>  		ssize_t ret;
>  
>  		inode_lock_shared(inode);
> -		ret = btrfs_direct_IO(iocb, to);
> +		ret = iomap_dio_rw(iocb, to, &btrfs_dio_iomap_ops,
> +				   &btrfs_dio_ops, is_sync_kiocb(iocb));
>  		inode_unlock_shared(inode);
>  
>  		return ret;
> diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
> index 18d153a62655..7b747270ec40 100644
> --- a/fs/btrfs/inode.c
> +++ b/fs/btrfs/inode.c
> @@ -29,7 +29,6 @@
>  #include <linux/iversion.h>
>  #include <linux/swap.h>
>  #include <linux/sched/mm.h>
> -#include <linux/iomap.h>
>  #include <asm/unaligned.h>
>  #include "misc.h"
>  #include "ctree.h"
> @@ -7856,6 +7855,11 @@ static int btrfs_dio_iomap_end(struct inode *inode, loff_t pos, loff_t length,
>  	return 0;
>  }
>  
> +const struct iomap_ops btrfs_dio_iomap_ops = {
> +	.iomap_begin            = btrfs_dio_iomap_begin,
> +	.iomap_end		= btrfs_dio_iomap_end,
> +};
> +
>  static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
>  						 struct bio *bio,
>  						 int mirror_num)
> @@ -8581,74 +8585,10 @@ static blk_qc_t btrfs_submit_direct(struct bio *dio_bio, struct file *file,
>  	return BLK_QC_T_NONE;
>  }
>  
> -static const struct iomap_ops btrfs_dio_iomap_ops = {
> -	.iomap_begin            = btrfs_dio_iomap_begin,
> -	.iomap_end		= btrfs_dio_iomap_end,
> -};
> -
> -static const struct iomap_dio_ops btrfs_dops = {
> +const struct iomap_dio_ops btrfs_dio_ops = {
>  	.submit_io		= btrfs_submit_direct,
>  };
>  
> -
> -/*
> - * btrfs_direct_IO - perform direct I/O
> - * inode->i_rwsem must be locked before calling this function, shared or exclusive.
> - * @iocb - kernel iocb
> - * @iter - iter to/from data is copied
> - */
> -
> -ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
> -{
> -	struct file *file = iocb->ki_filp;
> -	struct inode *inode = file->f_mapping->host;
> -	struct extent_changeset *data_reserved = NULL;
> -	loff_t offset = iocb->ki_pos;
> -	size_t count = 0;
> -	bool relock = false;
> -	ssize_t ret;
> -
> -	lockdep_assert_held(&inode->i_rwsem);
> -
> -	count = iov_iter_count(iter);
> -	if (iov_iter_rw(iter) == WRITE) {
> -		/*
> -		 * If the write DIO is beyond the EOF, we need update
> -		 * the isize, but it is protected by i_mutex. So we can
> -		 * not unlock the i_mutex at this case.
> -		 */
> -		if (offset + count <= inode->i_size) {
> -			inode_unlock(inode);
> -			relock = true;
> -		} else if (iocb->ki_flags & IOCB_NOWAIT) {
> -			ret = -EAGAIN;
> -			goto out;
> -		}
> -		ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
> -						   offset, count);
> -		if (ret)
> -			goto out;
> -
> -		down_read(&BTRFS_I(inode)->dio_sem);
> -	}
> -
> -	ret = iomap_dio_rw(iocb, iter, &btrfs_dio_iomap_ops, &btrfs_dops,
> -			is_sync_kiocb(iocb));
> -
> -	if (iov_iter_rw(iter) == WRITE) {
> -		up_read(&BTRFS_I(inode)->dio_sem);
> -		if (ret >= 0 && (size_t)ret < count)
> -			btrfs_delalloc_release_space(inode, data_reserved,
> -					offset, count - (size_t)ret, true);
> -		btrfs_delalloc_release_extents(BTRFS_I(inode), count);
> -	}
> -out:
> -	if (relock)
> -		inode_lock(inode);
> -	extent_changeset_free(data_reserved);
> -	return ret;
> -}
> -
>  #define BTRFS_FIEMAP_FLAGS	(FIEMAP_FLAG_SYNC)
>  
>  static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
> -- 
> 2.24.0
>
Goldwyn Rodrigues Jan. 7, 2020, 11:59 a.m. UTC | #3
On  6:42 21/12, Christoph Hellwig wrote:
> So Ilooked into the "unlocked" direct I/O case, and I think the current
> code using dio_sem is really sketchy.  What btrfs really needs to do is
> take i_rwsem shared by default for direct writes, and only upgrade to
> the exclusive lock when needed, similar to xfs and the WIP ext4 code.
> 
> While looking for that I also noticed two other things:
> 
>  - check_direct_IO looks pretty bogus
>  - btrfs_direct_IO really should be split and folded into the two
>    callers
> 
> Untested patches attached.  The first should probably go into a prep
> patch, and the second could be folded into this one.

Testing revealed that removing check_direct_IO will not work. We try and
reserve space as a whole for the entire direct write. These checks
safeguard from requests unaligned to fs_info->sectorsize.

I liked the patch to split and fold the direct_IO code. However to merge
it into this will make it difficult to understand the changes since we
are moving it to a different file rather than changing in-place. A
separate patch would better serve as a cleanup.
Christoph Hellwig Jan. 7, 2020, 5:21 p.m. UTC | #4
On Tue, Jan 07, 2020 at 05:59:09AM -0600, Goldwyn Rodrigues wrote:
> Testing revealed that removing check_direct_IO will not work. We try and
> reserve space as a whole for the entire direct write. These checks
> safeguard from requests unaligned to fs_info->sectorsize.

Ok.  The fact that a wrong sector size falls back to buffered I/O
instead of failing the I/O is still bogus, though.  Btrfs should align
with all other file systems there.

> 
> I liked the patch to split and fold the direct_IO code. However to merge
> it into this will make it difficult to understand the changes since we
> are moving it to a different file rather than changing in-place. A
> separate patch would better serve as a cleanup.

Sure, this can be added on top.
Christoph Hellwig Jan. 7, 2020, 5:23 p.m. UTC | #5
On Thu, Jan 02, 2020 at 12:01:27PM -0600, Goldwyn Rodrigues wrote:
> On  6:42 21/12, Christoph Hellwig wrote:
> > So Ilooked into the "unlocked" direct I/O case, and I think the current
> > code using dio_sem is really sketchy.  What btrfs really needs to do is
> > take i_rwsem shared by default for direct writes, and only upgrade to
> > the exclusive lock when needed, similar to xfs and the WIP ext4 code.
> 
> Sketchy in what sense? I am not trying to second-guess, but I want to
> know where it could fail. I would want it to be simpler as well, but if
> we can perform direct writes without locking, why should we introduce
> locks.

In that it needs yet another lock which doesn't really provide
exclusion guarantees in its own.  In many ways this lock plus the
historic i_mutex were abused to provide the shared/exclusiv lock
that now exists natively with i_rwsem.

Patch
diff mbox series

diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h
index b2e8fd8a8e59..113dcd1a11cd 100644
--- a/fs/btrfs/ctree.h
+++ b/fs/btrfs/ctree.h
@@ -2904,6 +2904,7 @@  int btrfs_writepage_cow_fixup(struct page *page, u64 start, u64 end);
 void btrfs_writepage_endio_finish_ordered(struct page *page, u64 start,
 					  u64 end, int uptodate);
 extern const struct dentry_operations btrfs_dentry_operations;
+ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
 
 /* ioctl.c */
 long btrfs_ioctl(struct file *file, unsigned int cmd, unsigned long arg);
diff --git a/fs/btrfs/file.c b/fs/btrfs/file.c
index 0cb43b682789..7010dd7beccc 100644
--- a/fs/btrfs/file.c
+++ b/fs/btrfs/file.c
@@ -1832,7 +1832,7 @@  static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
 	loff_t endbyte;
 	int err;
 
-	written = generic_file_direct_write(iocb, from);
+	written = btrfs_direct_IO(iocb, from);
 
 	if (written < 0 || !iov_iter_count(from))
 		return written;
@@ -3444,9 +3444,26 @@  static int btrfs_file_open(struct inode *inode, struct file *filp)
 	return generic_file_open(inode, filp);
 }
 
+static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
+{
+	ssize_t ret = 0;
+
+	if (iocb->ki_flags & IOCB_DIRECT) {
+		struct inode *inode = file_inode(iocb->ki_filp);
+
+		inode_lock_shared(inode);
+		ret = btrfs_direct_IO(iocb, to);
+		inode_unlock_shared(inode);
+		if (ret < 0)
+			return ret;
+	}
+
+	return generic_file_buffered_read(iocb, to, ret);
+}
+
 const struct file_operations btrfs_file_operations = {
 	.llseek		= btrfs_file_llseek,
-	.read_iter      = generic_file_read_iter,
+	.read_iter      = btrfs_file_read_iter,
 	.splice_read	= generic_file_splice_read,
 	.write_iter	= btrfs_file_write_iter,
 	.mmap		= btrfs_file_mmap,
diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index 56032c518b26..ff5ee99086f0 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c
@@ -29,6 +29,7 @@ 
 #include <linux/iversion.h>
 #include <linux/swap.h>
 #include <linux/sched/mm.h>
+#include <linux/iomap.h>
 #include <asm/unaligned.h>
 #include "misc.h"
 #include "ctree.h"
@@ -7510,7 +7511,7 @@  noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
 }
 
 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
-			      struct extent_state **cached_state, int writing)
+			      struct extent_state **cached_state, bool writing)
 {
 	struct btrfs_ordered_extent *ordered;
 	int ret = 0;
@@ -7648,30 +7649,7 @@  static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
 }
 
 
-static int btrfs_get_blocks_direct_read(struct extent_map *em,
-					struct buffer_head *bh_result,
-					struct inode *inode,
-					u64 start, u64 len)
-{
-	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
-
-	if (em->block_start == EXTENT_MAP_HOLE ||
-			test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
-		return -ENOENT;
-
-	len = min(len, em->len - (start - em->start));
-
-	bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
-		inode->i_blkbits;
-	bh_result->b_size = len;
-	bh_result->b_bdev = fs_info->fs_devices->latest_bdev;
-	set_buffer_mapped(bh_result);
-
-	return 0;
-}
-
 static int btrfs_get_blocks_direct_write(struct extent_map **map,
-					 struct buffer_head *bh_result,
 					 struct inode *inode,
 					 struct btrfs_dio_data *dio_data,
 					 u64 start, u64 len)
@@ -7733,7 +7711,6 @@  static int btrfs_get_blocks_direct_write(struct extent_map **map,
 	}
 
 	/* this will cow the extent */
-	len = bh_result->b_size;
 	free_extent_map(em);
 	*map = em = btrfs_new_extent_direct(inode, start, len);
 	if (IS_ERR(em)) {
@@ -7744,15 +7721,6 @@  static int btrfs_get_blocks_direct_write(struct extent_map **map,
 	len = min(len, em->len - (start - em->start));
 
 skip_cow:
-	bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
-		inode->i_blkbits;
-	bh_result->b_size = len;
-	bh_result->b_bdev = fs_info->fs_devices->latest_bdev;
-	set_buffer_mapped(bh_result);
-
-	if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
-		set_buffer_new(bh_result);
-
 	/*
 	 * Need to update the i_size under the extent lock so buffered
 	 * readers will get the updated i_size when we unlock.
@@ -7768,24 +7736,37 @@  static int btrfs_get_blocks_direct_write(struct extent_map **map,
 	return ret;
 }
 
-static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
-				   struct buffer_head *bh_result, int create)
+static int btrfs_dio_iomap_begin(struct inode *inode, loff_t start,
+		loff_t length, unsigned flags, struct iomap *iomap,
+		struct iomap *srcmap)
 {
 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 	struct extent_map *em;
 	struct extent_state *cached_state = NULL;
 	struct btrfs_dio_data *dio_data = NULL;
-	u64 start = iblock << inode->i_blkbits;
 	u64 lockstart, lockend;
-	u64 len = bh_result->b_size;
+	bool write = !!(flags & IOMAP_WRITE);
 	int ret = 0;
+	u64 len = length;
+	bool unlock_extents = false;
 
-	if (!create)
+	if (!write)
 		len = min_t(u64, len, fs_info->sectorsize);
 
 	lockstart = start;
 	lockend = start + len - 1;
 
+	/*
+	 * The generic stuff only does filemap_write_and_wait_range, which
+	 * isn't enough if we've written compressed pages to this area, so
+	 * we need to flush the dirty pages again to make absolutely sure
+	 * that any outstanding dirty pages are on disk.
+	 */
+	if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
+		     &BTRFS_I(inode)->runtime_flags))
+		ret = filemap_fdatawrite_range(inode->i_mapping, start,
+					 start + length - 1);
+
 	if (current->journal_info) {
 		/*
 		 * Need to pull our outstanding extents and set journal_info to NULL so
@@ -7801,7 +7782,7 @@  static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
 	 * this range and we need to fallback to buffered.
 	 */
 	if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
-			       create)) {
+			       write)) {
 		ret = -ENOTBLK;
 		goto err;
 	}
@@ -7833,35 +7814,52 @@  static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
 		goto unlock_err;
 	}
 
-	if (create) {
-		ret = btrfs_get_blocks_direct_write(&em, bh_result, inode,
+	len = min(len, em->len - (start - em->start));
+	if (write) {
+		ret = btrfs_get_blocks_direct_write(&em, inode,
 						    dio_data, start, len);
 		if (ret < 0)
 			goto unlock_err;
-
-		unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
-				     lockend, &cached_state);
+		unlock_extents = true;
+		/* Recalc len in case the new em is smaller than requested */
+		len = min(len, em->len - (start - em->start));
+	} else if (em->block_start == EXTENT_MAP_HOLE ||
+			test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
+		/* Unlock in case of direct reading from a hole */
+		unlock_extents = true;
 	} else {
-		ret = btrfs_get_blocks_direct_read(em, bh_result, inode,
-						   start, len);
-		/* Can be negative only if we read from a hole */
-		if (ret < 0) {
-			ret = 0;
-			free_extent_map(em);
-			goto unlock_err;
-		}
 		/*
 		 * We need to unlock only the end area that we aren't using.
 		 * The rest is going to be unlocked by the endio routine.
 		 */
-		lockstart = start + bh_result->b_size;
-		if (lockstart < lockend) {
-			unlock_extent_cached(&BTRFS_I(inode)->io_tree,
-					     lockstart, lockend, &cached_state);
-		} else {
-			free_extent_state(cached_state);
-		}
+		lockstart = start + len;
+		if (lockstart < lockend)
+			unlock_extents = true;
+	}
+
+	if (unlock_extents)
+		unlock_extent_cached(&BTRFS_I(inode)->io_tree,
+				lockstart, lockend, &cached_state);
+	else
+		free_extent_state(cached_state);
+
+	/*
+	 * Translate extent map information to iomap
+	 * We trim the extents (and move the addr) even though
+	 * iomap code does that, since we have locked only the parts
+	 * we are performing I/O in.
+	 */
+	if ((em->block_start == EXTENT_MAP_HOLE) ||
+	    (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) && !write)) {
+		iomap->addr = IOMAP_NULL_ADDR;
+		iomap->type = IOMAP_HOLE;
+	} else {
+		iomap->addr = em->block_start + (start - em->start);
+		iomap->type = IOMAP_MAPPED;
 	}
+	iomap->offset = start;
+	iomap->bdev = fs_info->fs_devices->latest_bdev;
+	iomap->length = len;
 
 	free_extent_map(em);
 
@@ -8229,10 +8227,9 @@  static void btrfs_endio_direct_read(struct bio *bio)
 
 	kfree(dip);
 
-	dio_bio->bi_status = err;
-	dio_end_io(dio_bio);
 	btrfs_io_bio_free_csum(io_bio);
-	bio_put(bio);
+	dio_bio->bi_status = err;
+	bio_endio(dio_bio);
 }
 
 static void __endio_write_update_ordered(struct inode *inode,
@@ -8289,8 +8286,7 @@  static void btrfs_endio_direct_write(struct bio *bio)
 	kfree(dip);
 
 	dio_bio->bi_status = bio->bi_status;
-	dio_end_io(dio_bio);
-	bio_put(bio);
+	bio_endio(dio_bio);
 }
 
 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
@@ -8522,9 +8518,10 @@  static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
 	return 0;
 }
 
-static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
+static blk_qc_t btrfs_submit_direct(struct bio *dio_bio, struct file *file,
 				loff_t file_offset)
 {
+	struct inode *inode = file_inode(file);
 	struct btrfs_dio_private *dip = NULL;
 	struct bio *bio = NULL;
 	struct btrfs_io_bio *io_bio;
@@ -8575,7 +8572,7 @@  static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
 
 	ret = btrfs_submit_direct_hook(dip);
 	if (!ret)
-		return;
+		return BLK_QC_T_NONE;
 
 	btrfs_io_bio_free_csum(io_bio);
 
@@ -8594,7 +8591,7 @@  static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
 		/*
 		 * The end io callbacks free our dip, do the final put on bio
 		 * and all the cleanup and final put for dio_bio (through
-		 * dio_end_io()).
+		 * end_io()).
 		 */
 		dip = NULL;
 		bio = NULL;
@@ -8609,15 +8606,12 @@  static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
 			      file_offset + dio_bio->bi_iter.bi_size - 1);
 
 		dio_bio->bi_status = BLK_STS_IOERR;
-		/*
-		 * Releases and cleans up our dio_bio, no need to bio_put()
-		 * nor bio_endio()/bio_io_error() against dio_bio.
-		 */
-		dio_end_io(dio_bio);
+		bio_endio(dio_bio);
 	}
 	if (bio)
 		bio_put(bio);
 	kfree(dip);
+	return BLK_QC_T_NONE;
 }
 
 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
@@ -8653,7 +8647,23 @@  static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
 	return retval;
 }
 
-static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
+static const struct iomap_ops btrfs_dio_iomap_ops = {
+	.iomap_begin            = btrfs_dio_iomap_begin,
+};
+
+static const struct iomap_dio_ops btrfs_dops = {
+	.submit_io		= btrfs_submit_direct,
+};
+
+
+/*
+ * btrfs_direct_IO - perform direct I/O
+ * inode->i_rwsem must be locked before calling this function, shared or exclusive.
+ * @iocb - kernel iocb
+ * @iter - iter to/from data is copied
+ */
+
+ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 {
 	struct file *file = iocb->ki_filp;
 	struct inode *inode = file->f_mapping->host;
@@ -8662,28 +8672,15 @@  static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 	struct extent_changeset *data_reserved = NULL;
 	loff_t offset = iocb->ki_pos;
 	size_t count = 0;
-	int flags = 0;
-	bool wakeup = true;
 	bool relock = false;
 	ssize_t ret;
 
+	lockdep_assert_held(&inode->i_rwsem);
+
 	if (check_direct_IO(fs_info, iter, offset))
 		return 0;
 
-	inode_dio_begin(inode);
-
-	/*
-	 * The generic stuff only does filemap_write_and_wait_range, which
-	 * isn't enough if we've written compressed pages to this area, so
-	 * we need to flush the dirty pages again to make absolutely sure
-	 * that any outstanding dirty pages are on disk.
-	 */
 	count = iov_iter_count(iter);
-	if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
-		     &BTRFS_I(inode)->runtime_flags))
-		filemap_fdatawrite_range(inode->i_mapping, offset,
-					 offset + count - 1);
-
 	if (iov_iter_rw(iter) == WRITE) {
 		/*
 		 * If the write DIO is beyond the EOF, we need update
@@ -8714,17 +8711,11 @@  static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 		dio_data.unsubmitted_oe_range_end = (u64)offset;
 		current->journal_info = &dio_data;
 		down_read(&BTRFS_I(inode)->dio_sem);
-	} else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
-				     &BTRFS_I(inode)->runtime_flags)) {
-		inode_dio_end(inode);
-		flags = DIO_LOCKING | DIO_SKIP_HOLES;
-		wakeup = false;
 	}
 
-	ret = __blockdev_direct_IO(iocb, inode,
-				   fs_info->fs_devices->latest_bdev,
-				   iter, btrfs_get_blocks_direct, NULL,
-				   btrfs_submit_direct, flags);
+	ret = iomap_dio_rw(iocb, iter, &btrfs_dio_iomap_ops, &btrfs_dops,
+			is_sync_kiocb(iocb));
+
 	if (iov_iter_rw(iter) == WRITE) {
 		up_read(&BTRFS_I(inode)->dio_sem);
 		current->journal_info = NULL;
@@ -8751,11 +8742,8 @@  static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 		btrfs_delalloc_release_extents(BTRFS_I(inode), count);
 	}
 out:
-	if (wakeup)
-		inode_dio_end(inode);
 	if (relock)
 		inode_lock(inode);
-
 	extent_changeset_free(data_reserved);
 	return ret;
 }
@@ -11045,7 +11033,7 @@  static const struct address_space_operations btrfs_aops = {
 	.writepage	= btrfs_writepage,
 	.writepages	= btrfs_writepages,
 	.readpages	= btrfs_readpages,
-	.direct_IO	= btrfs_direct_IO,
+	.direct_IO	= noop_direct_IO,
 	.invalidatepage = btrfs_invalidatepage,
 	.releasepage	= btrfs_releasepage,
 	.set_page_dirty	= btrfs_set_page_dirty,