Time namespaces virtualize the values of two system clocks:
- CLOCK_MONOTONIC (and likewise CLOCK_MONOTONIC_COARSE and CLOCK_MONOTONIC_RAW), a nonsettable clock that represents monotonic time since—as described by POSIX—"some unspecified point in the past".
- CLOCK_BOOTTIME (and likewise CLOCK_BOOTTIME_ALARM), a nonsettable clock that is identical to CLOCK_MONOTONIC, except that it also includes any time that the system is suspended.
Thus, the processes in a time namespace share per-namespace values for these clocks. This affects various APIs that measure against these clocks, including: clock_gettime(2), clock_nanosleep(2), nanosleep(2), timer_settime(2), timerfd_settime(2), and /proc/uptime.
Currently, the only way to create a time namespace is by calling unshare(2) with the CLONE_NEWTIME flag. This call creates a new time namespace but does not place the calling process in the new namespace. Instead, the calling process's subsequently created children are placed in the new namespace. This allows clock offsets (see below) for the new namespace to be set before the first process is placed in the namespace. The /proc/pid/ns/time_for_children symbolic link shows the time namespace in which the children of a process will be created. (A process can use a file descriptor opened on this symbolic link in a call to setns(2) in order to move into the namespace.)
Associated with each time namespace are offsets, expressed with respect to the initial time namespace, that define the values of the monotonic and boot-time clocks in that namespace. These offsets are exposed via the file /proc/PID/timens_offsets. Within this file, the offsets are expressed as lines consisting of three space-delimited fields:
<clock-id> <offset-secs> <offset-nanosecs>
The clock-id is a string that identifies the clock whose offsets are being shown. This field is either monotonic, for CLOCK_MONOTONIC, or boottime, for CLOCK_BOOTTIME. The remaining fields express the offset (seconds plus nanoseconds) for the clock in this time namespace. These offsets are expressed relative to the clock values in the initial time namespace. The offset-secs value can be negative, subject to restrictions noted below; offset-nanosecs is an unsigned value.
In the initial time namespace, the contents of the timens_offsets file are as follows:
$ cat /proc/self/timens_offsets monotonic 0 0 boottime 0 0
In a new time namespace that has had no member processes, the clock offsets can be modified by writing newline-terminated records of the same form to the timens_offsets file. The file can be written to multiple times, but after the first process has been created in or has entered the namespace, write(2)s on this file fail with the error EACCES. In order to write to the timens_offsets file, a process must have the CAP_SYS_TIME capability in the user namespace that owns the time namespace.
Writes to the timens_offsets file can fail with the following errors:
An offset-nanosecs value is greater than 999,999,999.
A clock-id value is not valid.
The caller does not have the CAP_SYS_TIME capability.
An offset-secs value is out of range. In particular;
- offset-secs can't be set to a value which would make the current time on the corresponding clock inside the namespace a negative value; and
- offset-secs can't be set to a value such that the time on the corresponding clock inside the namespace would exceed half of the value of the kernel constant KTIME_SEC_MAX (this limits the clock value to a maximum of approximately 146 years).
In a new time namespace created by unshare(2), the contents of the timens_offsets file are inherited from the time namespace of the creating process.
Use of time namespaces requires a kernel that is configured with the CONFIG_TIME_NS option.
Note that time namespaces do not virtualize the CLOCK_REALTIME clock. Virtualization of this clock was avoided for reasons of complexity and overhead within the kernel.
For compatibility with the initial implementation, when writing a clock-id to the /proc/pid/timens_offsets file, the numerical values of the IDs can be written instead of the symbolic names show above; i.e., 1 instead of monotonic, and 7 instead of boottime. For readability, the use of the symbolic names over the numbers is preferred.
The motivation for adding time namespaces was to allow the monotonic and boot-time clocks to maintain consistent values during container migration and checkpoint/restore.
The following shell session demonstrates the operation of time namespaces. We begin by displaying the inode number of the time namespace of a shell in the initial time namespace:
$ readlink /proc/$$/ns/time time:
Continuing in the initial time namespace, we display the system uptime using uptime(1) and use the clock_times example program shown in clock_getres(2) to display the values of various clocks:
$ uptime --pretty up 21 hours, 17 minutes $ ./clock_times CLOCK_REALTIME : 1585989401.971 (18356 days + 8h 36m 41s) CLOCK_TAI : 1585989438.972 (18356 days + 8h 37m 18s) CLOCK_MONOTONIC: 56338.247 (15h 38m 58s) CLOCK_BOOTTIME : 76633.544 (21h 17m 13s)
We then use unshare(1) to create a time namespace and execute a bash(1) shell. From the new shell, we use the built-in echo command to write records to the timens_offsets file adjusting the offset for the CLOCK_MONOTONIC clock forward 2 days and the offset for the CLOCK_BOOTTIME clock forward 7 days:
$ PS1="ns2# " sudo unshare -T -- bash --norc ns2# echo "monotonic $((2*24*60*60)) 0" > /proc/$$/timens_offsets ns2# echo "boottime $((7*24*60*60)) 0" > /proc/$$/timens_offsets
Above, we started the bash(1) shell with the --norc options so that no start-up scripts were executed. This ensures that no child processes are created from the shell before we have a chance to update the timens_offsets file.
We then use cat(1) to display the contents of the timens_offsets file. The execution of cat(1) creates the first process in the new time namespace, after which further attempts to update the timens_offsets file produce an error.
ns2# cat /proc/$$/timens_offsets monotonic 172800 0 boottime 604800 0 ns2# echo "boottime $((9*24*60*60)) 0" > /proc/$$/timens_offsets bash: echo: write error: Permission denied
Continuing in the new namespace, we execute uptime(1) and the clock_times example program:
ns2# uptime --pretty up 1 week, 21 hours, 18 minutes ns2# ./clock_times CLOCK_REALTIME : 1585989457.056 (18356 days + 8h 37m 37s) CLOCK_TAI : 1585989494.057 (18356 days + 8h 38m 14s) CLOCK_MONOTONIC: 229193.332 (2 days + 15h 39m 53s) CLOCK_BOOTTIME : 681488.629 (7 days + 21h 18m 8s)
From the above output, we can see that the monotonic and boot-time clocks have different values in the new time namespace.
Examining the /proc/pid/ns/time and /proc/pid/ns/time_for_children symbolic links, we see that the shell is a member of the initial time namespace, but its children are created in the new namespace.
ns2# readlink /proc/$$/ns/time time: ns2# readlink /proc/$$/ns/time_for_children time: ns2# readlink /proc/self/ns/time # Creates a child process time:
Returning to the shell in the initial time namespace, we see that the monotonic and boot-time clocks are unaffected by the timens_offsets changes that were made in the other time namespace:
$ uptime --pretty up 21 hours, 19 minutes $ ./clock_times CLOCK_REALTIME : 1585989401.971 (18356 days + 8h 38m 51s) CLOCK_TAI : 1585989438.972 (18356 days + 8h 39m 28s) CLOCK_MONOTONIC: 56338.247 (15h 41m 8s) CLOCK_BOOTTIME : 76633.544 (21h 19m 23s)
nsenter(1), unshare(1), clock_settime(2), setns(2), unshare(2), namespaces(7), time(7)
clock_getres(2), lttng-ust(3), namespaces(7), nsenter(1), time(7), unshare(1), unshare(2).