On 2017/2/16 9:16, Jaegeuk Kim wrote:
> On 02/14, Jaegeuk Kim wrote:
>> On 02/15, Chao Yu wrote:
>>> On 2017/2/15 2:03, Jaegeuk Kim wrote:
>>>> VFS uses f2fs_lookup() to decide f2fs_add_link() call during file creation.
>>>> But, if there is a race condition, f2fs_add_link() can be called multiple
>>>> times, resulting in multiple dentries with a same name. This patches fixes
>>>> it by adding __f2fs_find_entry() in f2fs_add_link() path.
>>>
>>> As during ->lookup, i_mutex will be held all the time, so there is no race
>>> condition in between different file creators?
>>
>> Hehe, yup.
>> I dropped this patch.
>
> It turns out sdcardfs has a race condition between lookup and create, and it
> calls vfs_create() twice. Workaround by f2fs would be needed for whole AOSP as
> well. And I added to check current to avoid performance regression.
>
> Thanks,
>
>>From 2f433dbfa491550e666a3d08f5a59ac5bed42b0f Mon Sep 17 00:00:00 2001
> From: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> Date: Tue, 14 Feb 2017 09:54:37 -0800
> Subject: [PATCH] f2fs: fix multiple f2fs_add_link() calls having same name
>
> It turns out a stakable filesystem like sdcardfs in AOSP can trigger multiple
> vfs_create() to lower filesystem. In that case, f2fs will add multiple dentries
> having same name which breaks filesystem consistency.
>
> Until upper layer fixes, let's work around by f2fs, which shows actually not
> much performance regression.
>
> Cc: <stable@xxxxxxxxxxxxxxx>
> Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> ---
> fs/f2fs/dir.c | 26 ++++++++++++++++++++++++--
> fs/f2fs/f2fs.h | 1 +
> 2 files changed, 25 insertions(+), 2 deletions(-)
>
> diff --git a/fs/f2fs/dir.c b/fs/f2fs/dir.c
> index 827c5daef4fc..4f5b3bb514c6 100644
> --- a/fs/f2fs/dir.c
> +++ b/fs/f2fs/dir.c
> @@ -207,9 +207,11 @@ static struct f2fs_dir_entry *find_in_level(struct inode *dir,
> f2fs_put_page(dentry_page, 0);
> }
>
> + /* This is to increase the speed of f2fs_create */
> if (!de && room && F2FS_I(dir)->chash != namehash) {
> F2FS_I(dir)->chash = namehash;
> F2FS_I(dir)->clevel = level;
> + F2FS_I(dir)->task = current;
> }
Hash collision can happen on different file names, here, assignment of .task
should be more accurately to prevent racing between lookup and creat, so how about:
if (!de && room) {
.task = current;
if (chash != namehash) {
.chash = namehash;
.clevel = level;
}
}
Thanks,
>
> return de;
> @@ -643,14 +645,34 @@ int __f2fs_add_link(struct inode *dir, const struct qstr *name,
> struct inode *inode, nid_t ino, umode_t mode)
> {
> struct fscrypt_name fname;
> + struct page *page = NULL;
> + struct f2fs_dir_entry *de = NULL;
> int err;
>
> err = fscrypt_setup_filename(dir, name, 0, &fname);
> if (err)
> return err;
>
> - err = __f2fs_do_add_link(dir, &fname, inode, ino, mode);
> -
> + /*
> + * An immature stakable filesystem shows a race condition between lookup
> + * and create. If we have same task when doing lookup and create, it's
> + * definitely fine as expected by VFS normally. Otherwise, let's just
> + * verify on-disk dentry one more time, which guarantees filesystem
> + * consistency more.
> + */
> + if (current != F2FS_I(dir)->task) {
> + de = __f2fs_find_entry(dir, &fname, &page);
> + F2FS_I(dir)->task = NULL;
> + }
> + if (de) {
> + f2fs_dentry_kunmap(dir, page);
> + f2fs_put_page(page, 0);
> + err = -EEXIST;
> + } else if (!IS_ERR(page)) {
> + err = __f2fs_do_add_link(dir, &fname, inode, ino, mode);
> + } else {
> + err = PTR_ERR(page);
> + }
> fscrypt_free_filename(&fname);
> return err;
> }
> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> index d263dade5e4c..cc22dc458896 100644
> --- a/fs/f2fs/f2fs.h
> +++ b/fs/f2fs/f2fs.h
> @@ -454,6 +454,7 @@ struct f2fs_inode_info {
> atomic_t dirty_pages; /* # of dirty pages */
> f2fs_hash_t chash; /* hash value of given file name */
> unsigned int clevel; /* maximum level of given file name */
> + struct task_struct *task; /* lookup and create consistency */
> nid_t i_xattr_nid; /* node id that contains xattrs */
> loff_t last_disk_size; /* lastly written file size */
>
>