Skip to contentSkip to conversation

f2fs: fix multiple f2fs_add_link() calls having same name

PATCH In review Jaegeuk Kim 2 participants 6 messages Feb 14, 2017 – Feb 16, 2017

Subsystems (3)
lore
Share & tools
lore

Jaegeuk Kim

author maintainer
#0 lore raw

PATCH 1/1 f2fs: fix multiple f2fs_add_link() calls having same name

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.

---
 fs/f2fs/dir.c | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)
fs/f2fs/dir.c +12 -2
index 827c5daef4fc..bc7c48839bec 100644
@@ -643,14 +643,24 @@ 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;
+	struct f2fs_dir_entry *de;
 	int err;
 
 	err = fscrypt_setup_filename(dir, name, 0, &fname);
 	if (err)
 		return err;
 
-	err = __f2fs_do_add_link(dir, &fname, inode, ino, mode);
-
+	de = __f2fs_find_entry(dir, &fname, &page);
+	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;
 }
Signature · 1 line
2.11.0
#1 lore raw
Quoted 4 lines · Jaegeuk Kim

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?

Thanks,

Quoted 39 lines
---
 fs/f2fs/dir.c | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)
fs/f2fs/dir.c +12 -2
index 827c5daef4fc..bc7c48839bec 100644
@@ -643,14 +643,24 @@ 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;
+	struct f2fs_dir_entry *de;
 	int err;
 
 	err = fscrypt_setup_filename(dir, name, 0, &fname);
 	if (err)
 		return err;
 
-	err = __f2fs_do_add_link(dir, &fname, inode, ino, mode);
-
+	de = __f2fs_find_entry(dir, &fname, &page);
+	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;
 }
author maintainer
#2 lore raw
Quoted 8 lines · Chao Yu
Quoted 4 lines · Jaegeuk Kim

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.

Thanks,

Quoted 42 lines

Thanks,

Quoted 39 lines
---
 fs/f2fs/dir.c | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)
fs/f2fs/dir.c +12 -2
index 827c5daef4fc..bc7c48839bec 100644
@@ -643,14 +643,24 @@ 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;
+	struct f2fs_dir_entry *de;
 	int err;
 
 	err = fscrypt_setup_filename(dir, name, 0, &fname);
 	if (err)
 		return err;
 
-	err = __f2fs_do_add_link(dir, &fname, inode, ino, mode);
-
+	de = __f2fs_find_entry(dir, &fname, &page);
+	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;
 }
author maintainer
#3 lore raw
Quoted 7 lines · Jaegeuk Kim
Quoted 6 lines · Chao Yu
Quoted 4 lines · Jaegeuk Kim

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.

Quoted 5 lines
Quoted 2 lines

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,

Quoted 1 line

From 2f433dbfa491550e666a3d08f5a59ac5bed42b0f Mon Sep 17 00:00:00 2001

From: Jaegeuk Kim <[email protected]>
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.

---
 fs/f2fs/dir.c  | 26 ++++++++++++++++++++++++--
 fs/f2fs/f2fs.h |  1 +
 2 files changed, 25 insertions(+), 2 deletions(-)
fs/f2fs/dir.c +24 -2
index 827c5daef4fc..4f5b3bb514c6 100644
@@ -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;
 	}
 
 	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;
 }
fs/f2fs/f2fs.h +1 -0
index d263dade5e4c..cc22dc458896 100644
@@ -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 */
 
Signature · 1 line
2.11.0
#4 lore raw
Quoted 44 lines · Jaegeuk Kim
Quoted 7 lines · Jaegeuk Kim
Quoted 6 lines · Chao Yu
Quoted 4 lines · Jaegeuk Kim

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.

Quoted 5 lines
Quoted 2 lines

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,

Quoted 1 line

From 2f433dbfa491550e666a3d08f5a59ac5bed42b0f Mon Sep 17 00:00:00 2001

From: Jaegeuk Kim <[email protected]>
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.

---
 fs/f2fs/dir.c  | 26 ++++++++++++++++++++++++--
 fs/f2fs/f2fs.h |  1 +
 2 files changed, 25 insertions(+), 2 deletions(-)
fs/f2fs/dir.c +0 -0
index 827c5daef4fc..4f5b3bb514c6 100644

fs/f2fs/dir.c static struct f2fs_dir_entry *find_in_level(struct inode *dir, · line 215

Quoted 9 lines
fs/f2fs/dir.c +2 -0
@@ -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,

Quoted 52 lines
fs/f2fs/dir.c +22 -2
 	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;
 }
fs/f2fs/f2fs.h +1 -0
index d263dade5e4c..cc22dc458896 100644
@@ -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 */
 
author maintainer
#5 lore raw
Quoted 58 lines · Chao Yu
Quoted 44 lines · Jaegeuk Kim
Quoted 7 lines · Jaegeuk Kim
Quoted 6 lines · Chao Yu
Quoted 4 lines · Jaegeuk Kim

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.

Quoted 5 lines
Quoted 2 lines

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,

Quoted 1 line

From 2f433dbfa491550e666a3d08f5a59ac5bed42b0f Mon Sep 17 00:00:00 2001

From: Jaegeuk Kim <[email protected]>
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.

---
 fs/f2fs/dir.c  | 26 ++++++++++++++++++++++++--
 fs/f2fs/f2fs.h |  1 +
 2 files changed, 25 insertions(+), 2 deletions(-)
fs/f2fs/dir.c +0 -0
index 827c5daef4fc..4f5b3bb514c6 100644

fs/f2fs/dir.c static struct f2fs_dir_entry *find_in_level(struct inode *dir, · line 215

Quoted 9 lines
fs/f2fs/dir.c +2 -0
@@ -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:

Quoted 7 lines
if (!de && room) {
	.task = current;
	if (chash != namehash) {
		.chash = namehash;
		.clevel = level;
	}
}

Makes sense.
Will apply that. ;)

Thanks,

Quoted 55 lines

Thanks,

Quoted 52 lines
fs/f2fs/dir.c +22 -2
 	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;
 }
fs/f2fs/f2fs.h +1 -0
index d263dade5e4c..cc22dc458896 100644
@@ -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 */