We used to have ptmx be owned by the inner uid and gid 0. Change
this: if the owner and group are both mapped but are not both 0,
then use the owner instead.
For container-style namespaces (LXC, etc), this should have no
effect -- UID 0 is will either be the owner or will be unmapped.
The important behavior change is for sandboxes: many sandboxes
intentionally do not create an inner uid 0. Without this patch,
mounting devpts in such a sandbox is awkward. With this patch, it
will just work and ptmx will be owned by the namespace owner.
Cc: Alexander Larsson <[email protected]>
Cc: [email protected]
Cc: "Eric W. Biederman" <[email protected]>
Cc: Linux Containers <[email protected]>
Signed-off-by: Andy Lutomirski <[email protected]>
---
fs/devpts/inode.c | 34 ++++++++++++++++++++++++++++++----
1 file changed, 30 insertions(+), 4 deletions(-)
diff --git a/fs/devpts/inode.c b/fs/devpts/inode.c
index 655f21f99160..d6fa2d1beee3 100644
--- a/fs/devpts/inode.c
+++ b/fs/devpts/inode.c
@@ -27,6 +27,7 @@
#include <linux/parser.h>
#include <linux/fsnotify.h>
#include <linux/seq_file.h>
+#include <linux/user_namespace.h>
#define DEVPTS_DEFAULT_MODE 0600
/*
@@ -250,10 +251,35 @@ static int mknod_ptmx(struct super_block *sb)
kuid_t root_uid;
kgid_t root_gid;
- root_uid = make_kuid(current_user_ns(), 0);
- root_gid = make_kgid(current_user_ns(), 0);
- if (!uid_valid(root_uid) || !gid_valid(root_gid))
- return -EINVAL;
+ /*
+ * For a new devpts instance, ptmx is owned by the creating user
+ * namespace's owner. Usually, that will be 0 as seen by the
+ * user namespace, but for unprivileged sandbox namespaces,
+ * there may not be a uid 0 or gid 0 at all.
+ */
+ root_uid = current_user_ns()->owner;
+ root_gid = current_user_ns()->group;
+
+ if (!uid_valid(root_uid) || !gid_valid(root_gid)) {
+ /*
+ * It's very unlikely for us to get here if the userns
+ * owner is not mapped, but it's possible -- we'd have
+ * to be running in the userns with capabilities granted
+ * by unshare or setns, since there is no inner
+ * privileged user. Nonetheless, this could happen, and
+ * we don't want ptmx to be owned by an unmapped user or
+ * group.
+ *
+ * If this happens fall back to historical behavior:
+ * try to have ptmx be owned by 0:0.
+ */
+ root_uid = make_kuid(current_user_ns(), 0);
+ root_gid = make_kgid(current_user_ns(), 0);
+
+ /* If this still doesn't work, give up. */
+ if (!uid_valid(root_uid) || !gid_valid(root_gid))
+ return -EINVAL;
+ }
inode_lock(d_inode(root));
--
2.5.0
Quoting Andy Lutomirski ([email protected]):
> We used to have ptmx be owned by the inner uid and gid 0. Change
> this: if the owner and group are both mapped but are not both 0,
> then use the owner instead.
>
> For container-style namespaces (LXC, etc), this should have no
> effect -- UID 0 is will either be the owner or will be unmapped.
This doesn't seem right - it's often the case that the owner is mapped
in as non-0 uid, safe or not. The actual namespace root uid should be
the owner (so long as it exists).
Why not reverse the cases? If 0 is not mapped, then check whether the
current_user_ns()->owner is mapped?
> The important behavior change is for sandboxes: many sandboxes
> intentionally do not create an inner uid 0. Without this patch,
> mounting devpts in such a sandbox is awkward. With this patch, it
> will just work and ptmx will be owned by the namespace owner.
>
> Cc: Alexander Larsson <[email protected]>
> Cc: [email protected]
> Cc: "Eric W. Biederman" <[email protected]>
> Cc: Linux Containers <[email protected]>
> Signed-off-by: Andy Lutomirski <[email protected]>
> ---
> fs/devpts/inode.c | 34 ++++++++++++++++++++++++++++++----
> 1 file changed, 30 insertions(+), 4 deletions(-)
>
> diff --git a/fs/devpts/inode.c b/fs/devpts/inode.c
> index 655f21f99160..d6fa2d1beee3 100644
> --- a/fs/devpts/inode.c
> +++ b/fs/devpts/inode.c
> @@ -27,6 +27,7 @@
> #include <linux/parser.h>
> #include <linux/fsnotify.h>
> #include <linux/seq_file.h>
> +#include <linux/user_namespace.h>
>
> #define DEVPTS_DEFAULT_MODE 0600
> /*
> @@ -250,10 +251,35 @@ static int mknod_ptmx(struct super_block *sb)
> kuid_t root_uid;
> kgid_t root_gid;
>
> - root_uid = make_kuid(current_user_ns(), 0);
> - root_gid = make_kgid(current_user_ns(), 0);
> - if (!uid_valid(root_uid) || !gid_valid(root_gid))
> - return -EINVAL;
> + /*
> + * For a new devpts instance, ptmx is owned by the creating user
> + * namespace's owner. Usually, that will be 0 as seen by the
> + * user namespace, but for unprivileged sandbox namespaces,
> + * there may not be a uid 0 or gid 0 at all.
> + */
> + root_uid = current_user_ns()->owner;
> + root_gid = current_user_ns()->group;
> +
> + if (!uid_valid(root_uid) || !gid_valid(root_gid)) {
> + /*
> + * It's very unlikely for us to get here if the userns
> + * owner is not mapped, but it's possible -- we'd have
> + * to be running in the userns with capabilities granted
> + * by unshare or setns, since there is no inner
> + * privileged user. Nonetheless, this could happen, and
> + * we don't want ptmx to be owned by an unmapped user or
> + * group.
> + *
> + * If this happens fall back to historical behavior:
> + * try to have ptmx be owned by 0:0.
> + */
> + root_uid = make_kuid(current_user_ns(), 0);
> + root_gid = make_kgid(current_user_ns(), 0);
> +
> + /* If this still doesn't work, give up. */
> + if (!uid_valid(root_uid) || !gid_valid(root_gid))
> + return -EINVAL;
> + }
>
> inode_lock(d_inode(root));
>
> --
> 2.5.0
>
> _______________________________________________
> Containers mailing list
> [email protected]
> https://lists.linuxfoundation.org/mailman/listinfo/containers
On sön, 2016-03-13 at 22:06 -0700, Andy Lutomirski wrote:
> We used to have ptmx be owned by the inner uid and gid 0. Change
> this: if the owner and group are both mapped but are not both 0,
> then use the owner instead.
>
> For container-style namespaces (LXC, etc), this should have no
> effect -- UID 0 is will either be the owner or will be unmapped.
>
> The important behavior change is for sandboxes: many sandboxes
> intentionally do not create an inner uid 0. Without this patch,
> mounting devpts in such a sandbox is awkward. With this patch, it
> will just work and ptmx will be owned by the namespace owner.
>
> Cc: Alexander Larsson <[email protected]>
> Cc: [email protected]
> Cc: "Eric W. Biederman" <[email protected]>
> Cc: Linux Containers <[email protected]>
> Signed-off-by: Andy Lutomirski <[email protected]>
Tested-by: Alexander Larsson <[email protected]>
Seems to work fine for me! Thanks!
--
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Alexander Larsson Red Hat, Inc
[email protected] [email protected]
He's an uncontrollable voodoo librarian with a robot buddy named Sparky.
She's a cynical winged journalist from the wrong side of the tracks. They
fight crime!
On Mar 14, 2016 1:35 AM, "Serge Hallyn" <[email protected]> wrote:
>
> Quoting Andy Lutomirski ([email protected]):
> > We used to have ptmx be owned by the inner uid and gid 0. Change
> > this: if the owner and group are both mapped but are not both 0,
> > then use the owner instead.
> >
> > For container-style namespaces (LXC, etc), this should have no
> > effect -- UID 0 is will either be the owner or will be unmapped.
>
> This doesn't seem right - it's often the case that the owner is mapped
> in as non-0 uid, safe or not. The actual namespace root uid should be
> the owner (so long as it exists).
>
> Why not reverse the cases? If 0 is not mapped, then check whether the
> current_user_ns()->owner is mapped?
Good point, and less chance of breakage that way as well.
--Andy