Anand Avati <avati@xxxxxxxxxx> writes:
> Implement ->direct_IO() method in aops. The ->direct_IO() method combines
> the existing fuse_direct_read/fuse_direct_write with a generic
> fuse_loop_rwv (inspired by do_loop_readv/writev) to neatly re-use the existing
> direct_io framework.
>
> 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)
I'm generally fine with this approach, but see comments inline.
>
> Signed-off-by: Anand Avati <avati@xxxxxxxxxx>
> ---
> fs/fuse/dir.c | 3 --
> fs/fuse/file.c | 73 ++++++++++++++++++++++++++++++++++++++++++++++++++++----
> 2 files changed, 68 insertions(+), 8 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..3c6b8bc 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;
> @@ -932,6 +928,7 @@ 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;
> struct inode *inode = mapping->host;
> ssize_t err;
> @@ -939,10 +936,13 @@ static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
>
> 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);
>
> @@ -962,6 +962,13 @@ static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
>
> file_update_time(file);
>
> + if (file->f_flags & O_DIRECT) {
> + written = generic_file_direct_write(iocb, iov, &nr_segs,
> + pos, &iocb->ki_pos,
> + count, ocount);
It looks to me like this will deadlock on i_mutex (fuse_direct_write()
will try to lock it again, while it is already locked). Was this code
path tested at all?
Also generic_file_direct_write() may return a short count in which case
the write should be finished using buffered write.
Thanks,
Miklos
> + goto out;
> + }
> +
> iov_iter_init(&i, iov, nr_segs, count, 0);
> written = fuse_perform_write(file, mapping, &i, pos);
> if (written >= 0)
> @@ -2077,6 +2084,61 @@ int fuse_notify_poll_wakeup(struct fuse_conn *fc,
> return 0;
> }
>
> +
> +static ssize_t fuse_loop_rwv(struct file *filp, const struct iovec *iov,
> + unsigned long nr_segs, loff_t *ppos, io_fn_t fn)
> +{
> + 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--;
> +
> + nr = fn(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;
> + io_fn_t fn = NULL;
> +
> + file = iocb->ki_filp;
> + pos = offset;
> +
> + if (rw == WRITE)
> + fn = (io_fn_t)fuse_direct_write;
> + else
> + fn = fuse_direct_read;
> +
> + ret = fuse_loop_rwv(file, iov, nr_segs, &pos, fn);
> +
> + return ret;
> +}
> +
> static const struct file_operations fuse_file_operations = {
> .llseek = fuse_file_llseek,
> .read = do_sync_read,
> @@ -2120,6 +2182,7 @@ 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)
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
