CSE Roadmap

Core tier · about 3 hours · 7 parts

Part 3 of 7 · 2 min read

The eight kinds

Think of a house with separate systems for post, telephone, water and so on, where each tenant has its own mailbox, extension number and tap. Each system is renamed on its own, so a tenant can have a private phone book but share the water. The analogy stops in one place: a house has no rule that links the systems, and the kernel has a flag for each kind, so a program picks the kinds it wants.

In plain words: Linux has eight kinds. ls /proc/self/ns on the test machine lists them: cgroup, ipc, mnt, net, pid, time, user and uts.

Kind Flag What the process sees privately
Mount CLONE_NEWNS The tree of mounts and so the file system view (Level 9)
PID CLONE_NEWPID Process numbers, with its own number 1
Network CLONE_NEWNET Network devices, addresses, routes, ports
UTS CLONE_NEWUTS Host name and domain name
IPC CLONE_NEWIPC Shared memory, semaphores and message queues (Level 7)
User CLONE_NEWUSER User and group IDs, and capabilities (Level 10)
Cgroup CLONE_NEWCGROUP The view of the control-group tree (page 5)
Time CLONE_NEWTIME Offsets of the monotonic and boot clocks

Three system calls handle them. clone creates a process in new namespaces, unshare moves the calling process into new ones, and setns joins an existing one (the way nsenter and docker exec enter a running container). lsns lists the namespaces that exist.

A worked example. The host's lsns on the test machine shows the first lines:

        NS TYPE NPROCS PID USER COMMAND
4026531832 mnt      66   2 root kthreadd
4026531833 net      67   2 root kthreadd
4026531836 pid      67   2 root kthreadd
4026531838 uts      67   2 root kthreadd

All 67 processes share one network, one pid and one UTS namespace, which are the initial ones, shown by their number 4026531836 and similar. One mount namespace (4026531860) holds a single process: a kernel thread with its own mount tree. A new container would add four or five more lines, with new numbers.

The key lines. A namespace is identified by an inode number, and two processes share a namespace exactly if the numbers in /proc/PID/ns/... are equal.

unshare, clone and setns

clone creates a new process, and with flags it puts the child into new namespaces. unshare moves the calling process into new namespaces. setns moves the calling process into an existing namespace named by a file descriptor.

Example: unshare --uts sh starts a shell with its own host name, and nsenter --target 517 --uts sh joins the namespace of process 517.

Check yourself: 1. Name five of the eight kinds. 2. Which call joins a namespace that already exists? 3. How can you tell that two processes share a network namespace?