[PATCH 1/2] overlayfs: fix a deadlock in ovl_dir_read_merged

[Felix Fietkau]

Some filesystems (e.g. jffs2) lock the same resources for both readdir
and lookup, leading to a deadlock in ovl_dir_read_merged, which uses
a lookup in the filldir callback for the upper filesystem to identify
whiteout entries.
Fix this by moving all links to the head of the list and marking whiteouts
after readdir.

Signed-off-by: Felix Fietkau <***@openwrt.org>
---
fs/overlayfs/overlayfs.c | 61 +++++++++++++++++++++++++++-------------------
1 files changed, 36 insertions(+), 25 deletions(-)

diff --git a/fs/overlayfs/overlayfs.c b/fs/overlayfs/overlayfs.c
index dca953e..10ca9c0 100644
--- a/fs/overlayfs/overlayfs.c
+++ b/fs/overlayfs/overlayfs.c
@@ -253,8 +253,7 @@ static struct ovl_cache_entry *ovl_cache_entry_find(struct rb_root *root,
}

static struct ovl_cache_entry *ovl_cache_entry_new(const char *name, int len,
- u64 ino, unsigned int d_type,
- bool is_whiteout)
+ u64 ino, unsigned int d_type)
{
struct ovl_cache_entry *p;

@@ -267,7 +266,7 @@ static struct ovl_cache_entry *ovl_cache_entry_new(const char *name, int len,
p->len = len;
p->type = d_type;
p->ino = ino;
- p->is_whiteout = is_whiteout;
+ p->is_whiteout = false;
}

return p;
@@ -275,7 +274,7 @@ static struct ovl_cache_entry *ovl_cache_entry_new(const char *name, int len,

static int ovl_cache_entry_add_rb(struct ovl_readdir_data *rdd,
const char *name, int len, u64 ino,
- unsigned int d_type, bool is_whiteout)
+ unsigned int d_type)
{
struct rb_node **newp = &rdd->root->rb_node;
struct rb_node *parent = NULL;
@@ -296,11 +295,18 @@ static int ovl_cache_entry_add_rb(struct ovl_readdir_data *rdd,
return 0;
}

- p = ovl_cache_entry_new(name, len, ino, d_type, is_whiteout);
+ p = ovl_cache_entry_new(name, len, ino, d_type);
if (p == NULL)
return -ENOMEM;

- list_add_tail(&p->l_node, rdd->list);
+ /*
+ * Add links before other types to be able to quicky mark
+ * any whiteout entries
+ */
+ if (d_type == DT_LNK)
+ list_add(&p->l_node, rdd->list);
+ else
+ list_add_tail(&p->l_node, rdd->list);
rb_link_node(&p->node, parent, newp);
rb_insert_color(&p->node, rdd->root);

@@ -318,7 +324,7 @@ static int ovl_fill_lower(void *buf, const char *name, int namelen,
if (p) {
list_move_tail(&p->l_node, rdd->middle);
} else {
- p = ovl_cache_entry_new(name, namelen, ino, d_type, false);
+ p = ovl_cache_entry_new(name, namelen, ino, d_type);
if (p == NULL)
rdd->err = -ENOMEM;
else
@@ -343,26 +349,9 @@ static int ovl_fill_upper(void *buf, const char *name, int namelen,
loff_t offset, u64 ino, unsigned int d_type)
{
struct ovl_readdir_data *rdd = buf;
- bool is_whiteout = false;

rdd->count++;
- if (d_type == DT_LNK) {
- struct dentry *dentry;
-
- dentry = lookup_one_len(name, rdd->dir, namelen);
- if (IS_ERR(dentry)) {
- rdd->err = PTR_ERR(dentry);
- goto out;
- }
- is_whiteout = ovl_is_whiteout(dentry);
- dput(dentry);
- }
-
- rdd->err = ovl_cache_entry_add_rb(rdd, name, namelen, ino, d_type,
- is_whiteout);
-
-out:
- return rdd->err;
+ return ovl_cache_entry_add_rb(rdd, name, namelen, ino, d_type);
}

static int ovl_dir_read(struct path *realpath, struct ovl_readdir_data *rdd,
@@ -428,6 +417,26 @@ static void ovl_dir_reset(struct file *file)
}
}

+static void ovl_dir_mark_whiteouts(struct ovl_readdir_data *rdd)
+{
+ struct ovl_cache_entry *p;
+ struct dentry *dentry;
+
+ mutex_lock(&rdd->dir->d_inode->i_mutex);
+ list_for_each_entry(p, rdd->list, l_node) {
+ if (p->type != DT_LNK)
+ break;
+
+ dentry = lookup_one_len(p->name, rdd->dir, p->len);
+ if (IS_ERR(dentry))
+ continue;
+
+ p->is_whiteout = ovl_is_whiteout(dentry);
+ dput(dentry);
+ }
+ mutex_unlock(&rdd->dir->d_inode->i_mutex);
+}
+
static int ovl_dir_read_merged(struct path *upperpath, struct path *lowerpath,
struct ovl_readdir_data *rdd)
{
@@ -441,6 +450,8 @@ static int ovl_dir_read_merged(struct path *upperpath, struct path *lowerpath,
err = ovl_dir_read(upperpath, rdd, ovl_fill_upper);
if (err)
goto out;
+
+ ovl_dir_mark_whiteouts(rdd);
}
/*
* Insert lowerpath entries before upperpath ones, this allows

[Felix Fietkau]

Calling lookup or vfs_unlink from the filldir callback leads to a deadlock,
so when removing whiteouts for a directory, read them into a list first,
then remove all entries afterwards.

Signed-off-by: Felix Fietkau <***@openwrt.org>
---
fs/overlayfs/overlayfs.c | 53 +++++++++++++++++++++++++++++++--------------
1 files changed, 36 insertions(+), 17 deletions(-)

diff --git a/fs/overlayfs/overlayfs.c b/fs/overlayfs/overlayfs.c
index 10ca9c0..f795699 100644
--- a/fs/overlayfs/overlayfs.c
+++ b/fs/overlayfs/overlayfs.c
@@ -1700,37 +1700,56 @@ static int ovl_check_empty_dir(struct dentry *dentry)
return err;
}

-static int ovl_unlink_whiteout(void *buf, const char *name, int namelen,
- loff_t offset, u64 ino, unsigned int d_type)
+static int ovl_fill_links(void *buf, const char *name, int namelen,
+ loff_t offset, u64 ino, unsigned int d_type)
{
struct ovl_readdir_data *rdd = buf;
+ struct ovl_cache_entry *p;

- rdd->count++;
- /* check d_type to filter out "." and ".." */
- if (d_type == DT_LNK) {
- struct dentry *dentry;
+ if (d_type != DT_LNK)
+ return 0;

- dentry = lookup_one_len(name, rdd->dir, namelen);
- if (IS_ERR(dentry)) {
- rdd->err = PTR_ERR(dentry);
- } else {
- rdd->err = vfs_unlink(rdd->dir->d_inode, dentry);
- dput(dentry);
- }
- }
+ p = ovl_cache_entry_new(name, namelen, ino, d_type);
+ if (!p)
+ return -ENOMEM;

- return rdd->err;
+ list_add(&p->l_node, rdd->list);
+ return 0;
}

static int ovl_remove_whiteouts(struct dentry *dentry)
{
struct path upperpath;
- struct ovl_readdir_data rdd = { .list = NULL };
+ LIST_HEAD(list);
+ struct ovl_readdir_data rdd = { .list = &list };
+ struct ovl_cache_entry *p, *t;
+ int ret;

ovl_path_upper(dentry, &upperpath);
rdd.dir = upperpath.dentry;

- return ovl_dir_read(&upperpath, &rdd, ovl_unlink_whiteout);
+ ret = ovl_dir_read(&upperpath, &rdd, ovl_fill_links);
+
+ mutex_lock(&rdd.dir->d_inode->i_mutex);
+ list_for_each_entry_safe(p, t, &list, l_node) {
+ struct dentry *dentry;
+
+ if (!ret) {
+ dentry = lookup_one_len(p->name, rdd.dir, p->len);
+ if (IS_ERR(dentry)) {
+ ret = PTR_ERR(dentry);
+ } else {
+ ret = vfs_unlink(rdd.dir->d_inode, dentry);
+ dput(dentry);
+ }
+ }
+
+ list_del(&p->l_node);
+ kfree(p);
+ }
+ mutex_unlock(&rdd.dir->d_inode->i_mutex);
+
+ return ret;
}

static int ovl_rmdir(struct inode *dir, struct dentry *dentry)

[Miklos Szeredi]

Felix,

Post by Felix Fietkau
Some filesystems (e.g. jffs2) lock the same resources for both readdir
and lookup, leading to a deadlock in ovl_dir_read_merged, which uses
a lookup in the filldir callback for the upper filesystem to identify
whiteout entries.
Fix this by moving all links to the head of the list and marking whiteouts
after readdir.

Thanks for the patches.

I committed and pushed them out to the overlayfs.v6 branch at

git://git.kernel.org/pub/scm/linux/kernel/git/mszeredi/vfs.git overlayfs.v6

Also added one optimization and one fix. I'll post those two patches
here shortly. Please test.

Thanks,
Miklos
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html