[PATCH 1/2] ext4: optimize ea_inode block expansion

[PATCH 1/2] ext4: optimize ea_inode block expansion

[Jun Nie]

Copy ea data from inode entry when expanding ea block if possible.
Then remove the ea entry if expansion success. Thus memcpy to a
temporary buffer may be avoided.

If the expansion fails, we do not need to recovery the removed ea
entry neither in this way.

Signed-off-by: Jun Nie <jun.nie@linaro.org>
---
 fs/ext4/xattr.c | 28 +++++++++++++++++-----------
 1 file changed, 17 insertions(+), 11 deletions(-)

diff --git a/fs/ext4/xattr.c b/fs/ext4/xattr.c
index 7decaaf27e82..235a517d9c17 100644
--- a/fs/ext4/xattr.c
+++ b/fs/ext4/xattr.c
@@ -2551,9 +2551,8 @@ static int ext4_xattr_move_to_block(handle_t *handle, struct inode *inode,
 
 	is = kzalloc(sizeof(struct ext4_xattr_ibody_find), GFP_NOFS);
 	bs = kzalloc(sizeof(struct ext4_xattr_block_find), GFP_NOFS);
-	buffer = kvmalloc(value_size, GFP_NOFS);
 	b_entry_name = kmalloc(entry->e_name_len + 1, GFP_NOFS);
-	if (!is || !bs || !buffer || !b_entry_name) {
+	if (!is || !bs || !b_entry_name) {
 		error = -ENOMEM;
 		goto out;
 	}
@@ -2565,12 +2564,18 @@ static int ext4_xattr_move_to_block(handle_t *handle, struct inode *inode,
 
 	/* Save the entry name and the entry value */
 	if (entry->e_value_inum) {
+		buffer = kvmalloc(value_size, GFP_NOFS);
+		if (!buffer) {
+			error = -ENOMEM;
+			goto out;
+		}
+
 		error = ext4_xattr_inode_get(inode, entry, buffer, value_size);
 		if (error)
 			goto out;
 	} else {
 		size_t value_offs = le16_to_cpu(entry->e_value_offs);
-		memcpy(buffer, (void *)IFIRST(header) + value_offs, value_size);
+		buffer = (void *)IFIRST(header) + value_offs;
 	}
 
 	memcpy(b_entry_name, entry->e_name, entry->e_name_len);
@@ -2585,25 +2590,26 @@ static int ext4_xattr_move_to_block(handle_t *handle, struct inode *inode,
 	if (error)
 		goto out;
 
-	/* Remove the chosen entry from the inode */
-	error = ext4_xattr_ibody_set(handle, inode, &i, is);
-	if (error)
-		goto out;
-
 	i.value = buffer;
 	i.value_len = value_size;
 	error = ext4_xattr_block_find(inode, &i, bs);
 	if (error)
 		goto out;
 
-	/* Add entry which was removed from the inode into the block */
+	/* Move ea entry from the inode into the block */
 	error = ext4_xattr_block_set(handle, inode, &i, bs);
 	if (error)
 		goto out;
-	error = 0;
+
+	/* Remove the chosen entry from the inode */
+	i.value = NULL;
+	i.value_len = 0;
+	error = ext4_xattr_ibody_set(handle, inode, &i, is);
+
 out:
 	kfree(b_entry_name);
-	kvfree(buffer);
+	if (entry->e_value_inum && buffer)
+		kvfree(buffer);
 	if (is)
 		brelse(is->iloc.bh);
 	if (bs)
-- 
2.34.1

[PATCH 2/2] ext4: refuse to create ea block when umounted

[Jun Nie]

The ea block expansion need to access s_root while it is
already set as NULL when umount is triggered. Refuse this
request to avoid panic.

Reported-by: syzbot+2dacb8f015bf1420155f@syzkaller.appspotmail.com
Link: https://syzkaller.appspot.com/bug?id=3613786cb88c93aa1c6a279b1df6a7b201347d08
Signed-off-by: Jun Nie <jun.nie@linaro.org>
---
 fs/ext4/xattr.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/fs/ext4/xattr.c b/fs/ext4/xattr.c
index 235a517d9c17..ac58494e49b6 100644
--- a/fs/ext4/xattr.c
+++ b/fs/ext4/xattr.c
@@ -1422,6 +1422,12 @@ static struct inode *ext4_xattr_inode_create(handle_t *handle,
 	uid_t owner[2] = { i_uid_read(inode), i_gid_read(inode) };
 	int err;
 
+	if (inode->i_sb->s_root == NULL) {
+		ext4_error(inode->i_sb,
+			   "refuse to create EA inode when umounting");
+		return ERR_PTR(-EINVAL);
+	}
+
 	/*
 	 * Let the next inode be the goal, so we try and allocate the EA inode
 	 * in the same group, or nearby one.
-- 
2.34.1

Re: [PATCH 2/2] ext4: refuse to create ea block when umounted

[Eric Biggers]

On Fri, Dec 30, 2022 at 07:00:16PM +0800, Jun Nie wrote:
> The ea block expansion need to access s_root while it is
> already set as NULL when umount is triggered. Refuse this
> request to avoid panic.
> 
> Reported-by: syzbot+2dacb8f015bf1420155f@syzkaller.appspotmail.com
> Link: https://syzkaller.appspot.com/bug?id=3613786cb88c93aa1c6a279b1df6a7b201347d08
> Signed-off-by: Jun Nie <jun.nie@linaro.org>
> ---
>  fs/ext4/xattr.c | 6 ++++++
>  1 file changed, 6 insertions(+)
> 
> diff --git a/fs/ext4/xattr.c b/fs/ext4/xattr.c
> index 235a517d9c17..ac58494e49b6 100644
> --- a/fs/ext4/xattr.c
> +++ b/fs/ext4/xattr.c
> @@ -1422,6 +1422,12 @@ static struct inode *ext4_xattr_inode_create(handle_t *handle,
>  	uid_t owner[2] = { i_uid_read(inode), i_gid_read(inode) };
>  	int err;
>  
> +	if (inode->i_sb->s_root == NULL) {
> +		ext4_error(inode->i_sb,
> +			   "refuse to create EA inode when umounting");
> +		return ERR_PTR(-EINVAL);
> +	}
> +

Why is an xattr being set during unmount?

- Eric

Re: [PATCH 2/2] ext4: refuse to create ea block when umounted

[Theodore Ts'o]

On Fri, Dec 30, 2022 at 12:13:25PM -0800, Eric Biggers wrote:
> > diff --git a/fs/ext4/xattr.c b/fs/ext4/xattr.c
> > index 235a517d9c17..ac58494e49b6 100644
> > --- a/fs/ext4/xattr.c
> > +++ b/fs/ext4/xattr.c
> > @@ -1422,6 +1422,12 @@ static struct inode *ext4_xattr_inode_create(handle_t *handle,
> >  	uid_t owner[2] = { i_uid_read(inode), i_gid_read(inode) };
> >  	int err;
> >  
> > +	if (inode->i_sb->s_root == NULL) {
> > +		ext4_error(inode->i_sb,
> > +			   "refuse to create EA inode when umounting");
> > +		return ERR_PTR(-EINVAL);
> > +	}

This should not be an ext4_error() since this is not a file system
corruption issue, but rather a kernel bug.  (With the one known
example being corrected by the first patch in this patch series.
Thanks Jun for working on a patch to things!)

This should be replaced by an ext4_warning() followed by a WARN_ON(1),
so we can get the stack trace.

> Why is an xattr being set during unmount?

The scenario was discovered by syzbot; the reproducer did the moral
equivalent of this (attached) shell script.  It required the
combination of lazytime and the debug_want_extra_isize mount options,
with a file system with (non-default) ea_inode feature enabled; so it
was highly unlikely to happen in real life.

For the detailed analysis, see [1]

[1] https://lore.kernel.org/all/Y5wGZG05uicAPscI@mit.edu

					- Ted

P.S.  Converting this into an xfstests test script to test for a
regression of this bug (or a failure to backport this into various
stable kernels) is also left as an exercise to the reader.  :-)

#!/bin/bash -vx
#
# This reproduces an ext4 bug caused by an unfortunate interaction
# between lazytime updates happening when a file system is being
# unmounted and expand_extra_isize
#
# Initially discovered via syzkaller:
# https://syzkaller.appspot.com/bug?id=3613786cb88c93aa1c6a279b1df6a7b201347d08
#

img=/tmp/foo.img
dir=/mnt
file=$dir/file0

rm -f $img
mke2fs -Fq -t ext4 -I 256 -O ea_inode -b 1024 $img 200k
mount $img $dir
v=$(dd if=/dev/zero bs=2000 count=1 2>/dev/null | tr '\0' =)
touch $file
attr -q -s test -V $v $file
umount $dir
mount -o debug_want_extra_isize=128,lazytime /tmp/foo.img $dir
cat $file
umount $dir