[PATCH v3] fuse: O_DIRECT support for files

[PATCH v3] fuse: O_DIRECT support for files

[Anand Avati]

Implement ->direct_IO() method in aops. The ->direct_IO() method combines
the existing fuse_direct_read/fuse_direct_write methods to implement
O_DIRECT functionality.

Reaching ->direct_IO() in the read path via generic_file_aio_read ensures
proper synchronization with page cache with its existing framework.

Reaching ->direct_IO() in the write path via fuse_file_aio_write is made
to come via generic_file_direct_write() which makes it play nice with
the page cache w.r.t other mmap pages etc.

On files marked 'direct_io' by the filesystem server, IO always follows
the fuse_direct_read/write path. There is no effect of fcntl(O_DIRECT)
and it always succeeds.

On files not marked with 'direct_io' by the filesystem server, the IO
path depends on O_DIRECT flag by the application. This can be passed
at the time of open() as well as via fcntl().

Note that asynchronous O_DIRECT iocb jobs are completed synchronously
always (this has been the case with FUSE even before this patch)

Signed-off-by: Anand Avati <avati@redhat.com>
---

Tested with concurrent read and write DDs with oflag=direct and iflag=direct set
in a few writes and a few reads

 fs/fuse/dir.c  |    3 -
 fs/fuse/file.c |  124 ++++++++++++++++++++++++++++++++++++++++++++++++--------
 2 files changed, 107 insertions(+), 20 deletions(-)

diff --git a/fs/fuse/dir.c b/fs/fuse/dir.c
index 2066328..7e5dbd0 100644
--- a/fs/fuse/dir.c
+++ b/fs/fuse/dir.c
@@ -387,9 +387,6 @@ static int fuse_create_open(struct inode *dir, struct dentry *entry,
 	if (fc->no_create)
 		return -ENOSYS;
 
-	if (flags & O_DIRECT)
-		return -EINVAL;
-
 	forget = fuse_alloc_forget();
 	if (!forget)
 		return -ENOMEM;
diff --git a/fs/fuse/file.c b/fs/fuse/file.c
index 4a199fd..0f426b5 100644
--- a/fs/fuse/file.c
+++ b/fs/fuse/file.c
@@ -194,10 +194,6 @@ int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 	struct fuse_conn *fc = get_fuse_conn(inode);
 	int err;
 
-	/* VFS checks this, but only _after_ ->open() */
-	if (file->f_flags & O_DIRECT)
-		return -EINVAL;
-
 	err = generic_file_open(inode, file);
 	if (err)
 		return err;
@@ -932,17 +928,23 @@ static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 	struct file *file = iocb->ki_filp;
 	struct address_space *mapping = file->f_mapping;
 	size_t count = 0;
+	size_t ocount = 0;
 	ssize_t written = 0;
+	ssize_t written_buffered = 0;
 	struct inode *inode = mapping->host;
 	ssize_t err;
 	struct iov_iter i;
+	loff_t endbyte = 0;
 
 	WARN_ON(iocb->ki_pos != pos);
 
-	err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
+	ocount = 0;
+	err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
 	if (err)
 		return err;
 
+	count = ocount;
+
 	mutex_lock(&inode->i_mutex);
 	vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
 
@@ -962,11 +964,36 @@ static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 
 	file_update_time(file);
 
-	iov_iter_init(&i, iov, nr_segs, count, 0);
-	written = fuse_perform_write(file, mapping, &i, pos);
-	if (written >= 0)
-		iocb->ki_pos = pos + written;
+	if (file->f_flags & O_DIRECT) {
+		written = generic_file_direct_write(iocb, iov, &nr_segs,
+						    pos, &iocb->ki_pos,
+						    count, ocount);
+		if (written < 0 || written == count)
+			goto out;
+
+		pos += written;
+		count -= written;
+
+		iov_iter_init(&i, iov, nr_segs, count, 0);
+		written_buffered = fuse_perform_write(file, mapping, &i, pos);
+		if (written_buffered < 0) {
+			err = written_buffered;
+			goto out;
+		}
+		endbyte = pos + written_buffered - 1;
 
+		err = filemap_write_and_wait_range(file->f_mapping, pos,
+						   endbyte);
+		if (err)
+			goto out;
+		written += written_buffered;
+		iocb->ki_pos = pos + written_buffered;
+	} else {
+		iov_iter_init(&i, iov, nr_segs, count, 0);
+		written = fuse_perform_write(file, mapping, &i, pos);
+		if (written >= 0)
+			iocb->ki_pos = pos + written;
+	}
 out:
 	current->backing_dev_info = NULL;
 	mutex_unlock(&inode->i_mutex);
@@ -1101,30 +1128,41 @@ static ssize_t fuse_direct_read(struct file *file, char __user *buf,
 	return res;
 }
 
-static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
-				 size_t count, loff_t *ppos)
+static ssize_t __fuse_direct_write(struct file *file, const char __user *buf,
+				   size_t count, loff_t *ppos)
 {
 	struct inode *inode = file->f_path.dentry->d_inode;
 	ssize_t res;
 
-	if (is_bad_inode(inode))
-		return -EIO;
-
-	/* Don't allow parallel writes to the same file */
-	mutex_lock(&inode->i_mutex);
 	res = generic_write_checks(file, ppos, &count, 0);
 	if (!res) {
 		res = fuse_direct_io(file, buf, count, ppos, 1);
 		if (res > 0)
 			fuse_write_update_size(inode, *ppos);
 	}
-	mutex_unlock(&inode->i_mutex);
 
 	fuse_invalidate_attr(inode);
 
 	return res;
 }
 
+static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
+				 size_t count, loff_t *ppos)
+{
+	struct inode *inode = file->f_path.dentry->d_inode;
+	ssize_t res;
+
+	if (is_bad_inode(inode))
+		return -EIO;
+
+	/* Don't allow parallel writes to the same file */
+	mutex_lock(&inode->i_mutex);
+	res = __fuse_direct_write(file, buf, count, ppos);
+	mutex_unlock(&inode->i_mutex);
+
+	return res;
+}
+
 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
 {
 	__free_page(req->pages[0]);
@@ -2077,6 +2115,57 @@ int fuse_notify_poll_wakeup(struct fuse_conn *fc,
 	return 0;
 }
 
+static ssize_t fuse_loop_dio(struct file *filp, const struct iovec *iov,
+			     unsigned long nr_segs, loff_t *ppos, int rw)
+{
+	const struct iovec *vector = iov;
+	ssize_t ret = 0;
+
+	while (nr_segs > 0) {
+		void __user *base;
+		size_t len;
+		ssize_t nr;
+
+		base = vector->iov_base;
+		len = vector->iov_len;
+		vector++;
+		nr_segs--;
+
+		if (rw == WRITE)
+			nr = __fuse_direct_write(filp, base, len, ppos);
+		else
+			nr = fuse_direct_read(filp, base, len, ppos);
+
+		if (nr < 0) {
+			if (!ret)
+				ret = nr;
+			break;
+		}
+		ret += nr;
+		if (nr != len)
+			break;
+	}
+
+	return ret;
+}
+
+
+static ssize_t
+fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
+			loff_t offset, unsigned long nr_segs)
+{
+	ssize_t ret = 0;
+	struct file *file = NULL;
+	loff_t pos = 0;
+
+	file = iocb->ki_filp;
+	pos = offset;
+
+	ret = fuse_loop_dio(file, iov, nr_segs, &pos, rw);
+
+	return ret;
+}
+
 static const struct file_operations fuse_file_operations = {
 	.llseek		= fuse_file_llseek,
 	.read		= do_sync_read,
@@ -2120,6 +2209,7 @@ static const struct address_space_operations fuse_file_aops  = {
 	.readpages	= fuse_readpages,
 	.set_page_dirty	= __set_page_dirty_nobuffers,
 	.bmap		= fuse_bmap,
+	.direct_IO	= fuse_direct_IO,
 };
 
 void fuse_init_file_inode(struct inode *inode)
-- 
1.7.4.4

Re: [PATCH v3] fuse: O_DIRECT support for files

[Miklos Szeredi]

Anand Avati <avati@redhat.com> writes:

> Implement ->direct_IO() method in aops. The ->direct_IO() method combines
> the existing fuse_direct_read/fuse_direct_write methods to implement
> O_DIRECT functionality.
>
> Reaching ->direct_IO() in the read path via generic_file_aio_read ensures
> proper synchronization with page cache with its existing framework.
>
> Reaching ->direct_IO() in the write path via fuse_file_aio_write is made
> to come via generic_file_direct_write() which makes it play nice with
> the page cache w.r.t other mmap pages etc.
>
> On files marked 'direct_io' by the filesystem server, IO always follows
> the fuse_direct_read/write path. There is no effect of fcntl(O_DIRECT)
> and it always succeeds.
>
> On files not marked with 'direct_io' by the filesystem server, the IO
> path depends on O_DIRECT flag by the application. This can be passed
> at the time of open() as well as via fcntl().
>
> Note that asynchronous O_DIRECT iocb jobs are completed synchronously
> always (this has been the case with FUSE even before this patch)
>
> Signed-off-by: Anand Avati <avati@redhat.com>
> ---
>
> Tested with concurrent read and write DDs with oflag=direct and iflag=direct set
> in a few writes and a few reads
>
>  fs/fuse/dir.c  |    3 -
>  fs/fuse/file.c |  124 ++++++++++++++++++++++++++++++++++++++++++++++++--------
>  2 files changed, 107 insertions(+), 20 deletions(-)
>
> diff --git a/fs/fuse/dir.c b/fs/fuse/dir.c
> index 2066328..7e5dbd0 100644
> --- a/fs/fuse/dir.c
> +++ b/fs/fuse/dir.c
> @@ -387,9 +387,6 @@ static int fuse_create_open(struct inode *dir, struct dentry *entry,
>  	if (fc->no_create)
>  		return -ENOSYS;
>  
> -	if (flags & O_DIRECT)
> -		return -EINVAL;
> -
>  	forget = fuse_alloc_forget();
>  	if (!forget)
>  		return -ENOMEM;
> diff --git a/fs/fuse/file.c b/fs/fuse/file.c
> index 4a199fd..0f426b5 100644
> --- a/fs/fuse/file.c
> +++ b/fs/fuse/file.c
> @@ -194,10 +194,6 @@ int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
>  	struct fuse_conn *fc = get_fuse_conn(inode);
>  	int err;
>  
> -	/* VFS checks this, but only _after_ ->open() */
> -	if (file->f_flags & O_DIRECT)
> -		return -EINVAL;
> -
>  	err = generic_file_open(inode, file);
>  	if (err)
>  		return err;
> @@ -932,17 +928,23 @@ static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
>  	struct file *file = iocb->ki_filp;
>  	struct address_space *mapping = file->f_mapping;
>  	size_t count = 0;
> +	size_t ocount = 0;
>  	ssize_t written = 0;
> +	ssize_t written_buffered = 0;
>  	struct inode *inode = mapping->host;
>  	ssize_t err;
>  	struct iov_iter i;
> +	loff_t endbyte = 0;
>  
>  	WARN_ON(iocb->ki_pos != pos);
>  
> -	err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
> +	ocount = 0;
> +	err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
>  	if (err)
>  		return err;
>  
> +	count = ocount;
> +
>  	mutex_lock(&inode->i_mutex);
>  	vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
>  
> @@ -962,11 +964,36 @@ static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
>  
>  	file_update_time(file);
>  
> -	iov_iter_init(&i, iov, nr_segs, count, 0);
> -	written = fuse_perform_write(file, mapping, &i, pos);
> -	if (written >= 0)
> -		iocb->ki_pos = pos + written;
> +	if (file->f_flags & O_DIRECT) {
> +		written = generic_file_direct_write(iocb, iov, &nr_segs,
> +						    pos, &iocb->ki_pos,
> +						    count, ocount);
> +		if (written < 0 || written == count)
> +			goto out;
> +
> +		pos += written;
> +		count -= written;
> +
> +		iov_iter_init(&i, iov, nr_segs, count, 0);

This codepath is still untested (hint: that last argument in
iov_iter_init() is wrong).  Yeah, it's not easy to test, but you can do
something like passing a fraction of count to
generic_file_direct_write() so one half is written through the direct IO
and the other half through the buffered IO.

Thanks,
Miklos

> +		written_buffered = fuse_perform_write(file, mapping, &i, pos);
> +		if (written_buffered < 0) {
> +			err = written_buffered;
> +			goto out;
> +		}
> +		endbyte = pos + written_buffered - 1;
>  
> +		err = filemap_write_and_wait_range(file->f_mapping, pos,
> +						   endbyte);
> +		if (err)
> +			goto out;
> +		written += written_buffered;
> +		iocb->ki_pos = pos + written_buffered;
> +	} else {
> +		iov_iter_init(&i, iov, nr_segs, count, 0);
> +		written = fuse_perform_write(file, mapping, &i, pos);
> +		if (written >= 0)
> +			iocb->ki_pos = pos + written;
> +	}
>  out:
>  	current->backing_dev_info = NULL;
>  	mutex_unlock(&inode->i_mutex);
> @@ -1101,30 +1128,41 @@ static ssize_t fuse_direct_read(struct file *file, char __user *buf,
>  	return res;
>  }
>  
> -static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
> -				 size_t count, loff_t *ppos)
> +static ssize_t __fuse_direct_write(struct file *file, const char __user *buf,
> +				   size_t count, loff_t *ppos)
>  {
>  	struct inode *inode = file->f_path.dentry->d_inode;
>  	ssize_t res;
>  
> -	if (is_bad_inode(inode))
> -		return -EIO;
> -
> -	/* Don't allow parallel writes to the same file */
> -	mutex_lock(&inode->i_mutex);
>  	res = generic_write_checks(file, ppos, &count, 0);
>  	if (!res) {
>  		res = fuse_direct_io(file, buf, count, ppos, 1);
>  		if (res > 0)
>  			fuse_write_update_size(inode, *ppos);
>  	}
> -	mutex_unlock(&inode->i_mutex);
>  
>  	fuse_invalidate_attr(inode);
>  
>  	return res;
>  }
>  
> +static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
> +				 size_t count, loff_t *ppos)
> +{
> +	struct inode *inode = file->f_path.dentry->d_inode;
> +	ssize_t res;
> +
> +	if (is_bad_inode(inode))
> +		return -EIO;
> +
> +	/* Don't allow parallel writes to the same file */
> +	mutex_lock(&inode->i_mutex);
> +	res = __fuse_direct_write(file, buf, count, ppos);
> +	mutex_unlock(&inode->i_mutex);
> +
> +	return res;
> +}
> +
>  static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
>  {
>  	__free_page(req->pages[0]);
> @@ -2077,6 +2115,57 @@ int fuse_notify_poll_wakeup(struct fuse_conn *fc,
>  	return 0;
>  }
>  
> +static ssize_t fuse_loop_dio(struct file *filp, const struct iovec *iov,
> +			     unsigned long nr_segs, loff_t *ppos, int rw)
> +{
> +	const struct iovec *vector = iov;
> +	ssize_t ret = 0;
> +
> +	while (nr_segs > 0) {
> +		void __user *base;
> +		size_t len;
> +		ssize_t nr;
> +
> +		base = vector->iov_base;
> +		len = vector->iov_len;
> +		vector++;
> +		nr_segs--;
> +
> +		if (rw == WRITE)
> +			nr = __fuse_direct_write(filp, base, len, ppos);
> +		else
> +			nr = fuse_direct_read(filp, base, len, ppos);
> +
> +		if (nr < 0) {
> +			if (!ret)
> +				ret = nr;
> +			break;
> +		}
> +		ret += nr;
> +		if (nr != len)
> +			break;
> +	}
> +
> +	return ret;
> +}
> +
> +
> +static ssize_t
> +fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
> +			loff_t offset, unsigned long nr_segs)
> +{
> +	ssize_t ret = 0;
> +	struct file *file = NULL;
> +	loff_t pos = 0;
> +
> +	file = iocb->ki_filp;
> +	pos = offset;
> +
> +	ret = fuse_loop_dio(file, iov, nr_segs, &pos, rw);
> +
> +	return ret;
> +}
> +
>  static const struct file_operations fuse_file_operations = {
>  	.llseek		= fuse_file_llseek,
>  	.read		= do_sync_read,
> @@ -2120,6 +2209,7 @@ static const struct address_space_operations fuse_file_aops  = {
>  	.readpages	= fuse_readpages,
>  	.set_page_dirty	= __set_page_dirty_nobuffers,
>  	.bmap		= fuse_bmap,
> +	.direct_IO	= fuse_direct_IO,
>  };
>  
>  void fuse_init_file_inode(struct inode *inode)