On Tue, Feb 12, 2019 at 10:18:49AM +0800, Anand Jain wrote:
> btrfs_set_prop() is a redirect to __btrfs_set_prop() with the
> transaction handler equal to NULL. And __btrfs_set_prop() inturn diectly
> uses trans to do_setxattr() which when trans is NULL creates a transaction.
That's right and I think that some of the callsites could actually pass
the existing transaction to btrfs_set_prop instead of relying on the
fallback behaviour.
Also, the potential NULL as transaction handle is not a good thing from
the API point of view and I'd like to reduce that to minimum (there are
some justified cases).
> diff --git a/fs/btrfs/ioctl.c b/fs/btrfs/ioctl.c
> index 3f9d7be30bf4..5a4ed2f66e09 100644
> --- a/fs/btrfs/ioctl.c
> +++ b/fs/btrfs/ioctl.c
> @@ -284,7 +284,8 @@ static int btrfs_ioctl_setflags(struct file *file, void
> __user *arg)
> binode->flags &= ~BTRFS_INODE_COMPRESS;
> binode->flags |= BTRFS_INODE_NOCOMPRESS;
>
> - ret = btrfs_set_prop(inode, "btrfs.compression", NULL, 0, 0);
> + ret = btrfs_set_prop(NULL, inode, "btrfs.compression", NULL, 0,
> + 0);
> if (ret && ret != -ENODATA)
> goto out_drop;
> } else if (fsflags & FS_COMPR_FL) {
> @@ -302,13 +303,14 @@ static int btrfs_ioctl_setflags(struct file *file, void
> __user *arg)
> if (!comp || comp[0] == 0)
> comp = btrfs_compress_type2str(BTRFS_COMPRESS_ZLIB);
>
> - ret = btrfs_set_prop(inode, "btrfs.compression",
> - comp, strlen(comp), 0);
> + ret = btrfs_set_prop(NULL, inode, "btrfs.compression", comp,
> + strlen(comp), 0);
> if (ret)
> goto out_drop;
>
> } else {
> - ret = btrfs_set_prop(inode, "btrfs.compression", NULL, 0, 0);
> + ret = btrfs_set_prop(NULL, inode, "btrfs.compression", NULL, 0,
> + 0);
> if (ret && ret != -ENODATA)
> goto out_drop;
> binode->flags &= ~(BTRFS_INODE_COMPRESS |
> BTRFS_INODE_NOCOMPRESS);
When the if-else block ends, there's a new transaction started, this
seems unnecessary.
> diff --git a/fs/btrfs/props.c b/fs/btrfs/props.c
> index dc6140013ae8..4525a2a4d1cd 100644
> --- a/fs/btrfs/props.c
> +++ b/fs/btrfs/props.c
> @@ -85,12 +85,9 @@ static const struct hlist_head
> *find_prop_handlers_by_hash(const u64 hash)
> return NULL;
> }
>
> -static int __btrfs_set_prop(struct btrfs_trans_handle *trans,
> - struct inode *inode,
> - const char *name,
> - const char *value,
> - size_t value_len,
> - int flags)
> +int btrfs_set_prop(struct btrfs_trans_handle *trans, struct inode *inode,
> + const char *name, const char *value, size_t value_len,
> + int flags)
> {
> const struct prop_handler *handler;
> int ret;
> @@ -133,15 +130,6 @@ static int __btrfs_set_prop(struct btrfs_trans_handle
> *trans,
> return 0;
> }
>
> -int btrfs_set_prop(struct inode *inode,
> - const char *name,
> - const char *value,
> - size_t value_len,
> - int flags)
> -{
> - return __btrfs_set_prop(NULL, inode, name, value, value_len, flags);
I agree that one function would be better here, with defined semantics
of 'trans'.
There are more cleanups around the properties, also in the xattr
handling functions. This patchset is a step in the right direction and I
think the cleanups could be more extensive.
The idea for the xattr function:
- the VFS callbacks (like btrfs_xattr_handler_set_prop) will start the
transaction themselves and pass the handle to the prop function
- a xattr function that does not take a valid trans could be named like
btrfs_setxattr_notrans and will start the transaction, ie. lifting
that to the callers
- there's a maze of the xattr callbacks so assertions are needed
