debugfs: debugfs_create_* shouldn't be checking permissions

From: "J. Bruce Fields" <[email protected]>

Subject: [PATCH] debugfs: debugfs_create_* shouldn't be checking permissions

Debugfs files and and directories are created by kernel subsystems not directly by users, so we shouldn't be using lookup_one_len, which checks permissions.

This was causing krb5 mounts to fail to Fedora servers using gss-proxy if selinux was enabled, on kernels since 388f0c776781 "sunrpc: add a debugfs rpc_xprt directory with an info file in it", which creates a new debugfs directory for each new rpc client.

Reported-by Anthony Messina <amessina-jVqiHVlHlyFWk0Htik3J/[email protected]>

Reported-by Jason Tibbits <[email protected]>

Cc Jeff Layton <jlayton-7I+n7zu2hftEKMMhf/[email protected]>

Signed-off-by J. Bruce Fields <[email protected]>

---
 fs/debugfs/inode.c | 17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)

I swiped this code fragment from net/sunrpc/rpc_pipe.c, and it's gotten only minimal testing. (It does fix krb5 mounts, though.)

fs/debugfs/inode.c +1 -1

index 96400ab42d13..75e5daa6a63f 100644
@@ -251,6 +251,7 @@ static struct dentry *start_creating(const char *name, struct dentry *parent)
 {
 	struct dentry *dentry;
 	int error;
+	struct qstr q = QSTR_INIT(name, strlen(name));
 
 	pr_debug("debugfs: creating file '%s'\n",name);
 

@@ -268,11 +269,19 @@ static struct dentry *start_creating(const char *name, struct dentry *parent)
 		parent = debugfs_mount->mnt_root;
 
 	mutex_lock(&parent->d_inode->i_mutex);
-	dentry = lookup_one_len(name, parent, strlen(name));

Quoted 8 lines · J. Bruce Fields

From: "J. Bruce Fields" <[email protected]>

Subject: [PATCH] debugfs: debugfs_create_* shouldn't be checking permissions

Debugfs files and and directories are created by kernel subsystems not directly by users, so we shouldn't be using lookup_one_len, which checks permissions.

Quoted 4 lines

This was causing krb5 mounts to fail to Fedora servers using gss-proxy if selinux was enabled, on kernels since 388f0c776781 "sunrpc: add a debugfs rpc_xprt directory with an info file in it", which creates a new debugfs directory for each new rpc client.

No kernel code should care / fail if a debugfs function fails, so please fix up the sunrpc code first.

Quoted 28 lines

Reported-by Anthony Messina <[email protected]>

Reported-by Jason Tibbits <[email protected]>

Cc Jeff Layton <[email protected]>

Signed-off-by J. Bruce Fields <[email protected]>

---
 fs/debugfs/inode.c | 17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)

I swiped this code fragment from net/sunrpc/rpc_pipe.c, and it's gotten only minimal testing. (It does fix krb5 mounts, though.)

fs/debugfs/inode.c +1 -1

index 96400ab42d13..75e5daa6a63f 100644
@@ -251,6 +251,7 @@ static struct dentry *start_creating(const char *name, struct dentry *parent)
 {
 	struct dentry *dentry;
 	int error;
+	struct qstr q = QSTR_INIT(name, strlen(name));
 
 	pr_debug("debugfs: creating file '%s'\n",name);
 

@@ -268,11 +269,19 @@ static struct dentry *start_creating(const char *name, struct dentry *parent)
 		parent = debugfs_mount->mnt_root;
 
 	mutex_lock(&parent->d_inode->i_mutex);
-	dentry = lookup_one_len(name, parent, strlen(name));

fs/debugfs/inode.c static struct dentry *start_creating(const char *name, struct dentry *parent) · line 280

Quoted 10 lines

fs/debugfs/inode.c +9 -1

-	if (!IS_ERR(dentry) && dentry->d_inode) {
+	dentry = d_hash_and_lookup(parent, &q);
+	if (!dentry) {
+		dentry = d_alloc(parent, &q);
+		if (!dentry) {
+			dentry = ERR_PTR(-ENOMEM);
+			goto out;
+		}
+	}
+	if (dentry->d_inode) {

No, I'd rather not "open code" lookup_one_len() if at all possible please.

What exactly is the problem here that the sunrpc code is failing from?

Is it just interacting with selinux? How is the debugfs code to blame here?

Quoted 9 lines · Greg Kroah-Hartman

Quoted 8 lines · J. Bruce Fields

From: "J. Bruce Fields" <[email protected]>

Subject: [PATCH] debugfs: debugfs_create_* shouldn't be checking permissions

Debugfs files and and directories are created by kernel subsystems not directly by users, so we shouldn't be using lookup_one_len, which checks permissions.

Quoted 7 lines

Quoted 4 lines

This was causing krb5 mounts to fail to Fedora servers using gss-proxy if selinux was enabled, on kernels since 388f0c776781 "sunrpc: add a debugfs rpc_xprt directory with an info file in it", which creates a new debugfs directory for each new rpc client.

No kernel code should care / fail if a debugfs function fails, so please fix up the sunrpc code first.

The sunrpc code is using debugfs to keep a list of all rpc clients together with some basic information about each one.

If the permissions issue is fixed then I think the only possible failures are lack of debugfs support and ENOMEM?

If there are situations where ENOMEM's actually possible, I'd worry a bit about this debugging information becoming unreliable after you go through some extreme memory shortage (possibly because that's what was necessary to reproduce the bug you're chasing).

But, I don't know, maybe I'm excessively paranoid; Jeff?

Quoted 43 lines

Quoted 28 lines

Reported-by Anthony Messina <amessina-jVqiHVlHlyFWk0Htik3J/[email protected]>

Reported-by Jason Tibbits <[email protected]>

Cc Jeff Layton <jlayton-7I+n7zu2hftEKMMhf/[email protected]>

Signed-off-by J. Bruce Fields <[email protected]>

---
 fs/debugfs/inode.c | 17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)

I swiped this code fragment from net/sunrpc/rpc_pipe.c, and it's gotten only minimal testing. (It does fix krb5 mounts, though.)

fs/debugfs/inode.c +1 -1

index 96400ab42d13..75e5daa6a63f 100644
@@ -251,6 +251,7 @@ static struct dentry *start_creating(const char *name, struct dentry *parent)
 {
 	struct dentry *dentry;
 	int error;
+	struct qstr q = QSTR_INIT(name, strlen(name));
 
 	pr_debug("debugfs: creating file '%s'\n",name);
 

@@ -268,11 +269,19 @@ static struct dentry *start_creating(const char *name, struct dentry *parent)
 		parent = debugfs_mount->mnt_root;
 
 	mutex_lock(&parent->d_inode->i_mutex);
-	dentry = lookup_one_len(name, parent, strlen(name));

fs/debugfs/inode.c static struct dentry *start_creating(const char *name, struct dentry *parent) · line 280

Quoted 10 lines

fs/debugfs/inode.c +9 -1

-	if (!IS_ERR(dentry) && dentry->d_inode) {
+	dentry = d_hash_and_lookup(parent, &q);
+	if (!dentry) {
+		dentry = d_alloc(parent, &q);
+		if (!dentry) {
+			dentry = ERR_PTR(-ENOMEM);
+			goto out;
+		}
+	}
+	if (dentry->d_inode) {

No, I'd rather not "open code" lookup_one_len() if at all possible please.

Quoted 1 line

What exactly is the problem here that the sunrpc code is failing from?

Sunrpc is creating a new rpc client (to talk to an rpc server).

Debugfs will fail to do that whenever the user doesn't have execute permissions on a debugfs directory. The user in this case is whoever gss-proxy's running as, but I think it could just as well be someone doing an nfs mount, for example.

Quoted 2 lines

Is it just interacting with selinux? How is the debugfs code to blame here?

As long as the debugfs directories permit "x" to everone, I guess you're only going to hit this with selinux (or some other security module).

We don't want to require updated selinux policies just because some subsystem started creating debugfs entries as a side-effect of some system call (which is what happened here).

--b.

From: "J. Bruce Fields" <[email protected]>

Subject: [PATCH] debugfs: debugfs_create_* shouldn't be checking permissions

Quoted 3 lines · J. Bruce Fields

From: "J. Bruce Fields" <[email protected]>

Subject: [PATCH] debugfs: debugfs_create_* shouldn't be checking permissions

Side-note, why are you putting the Subject: again here in the body of the patch?