# Time Namespace {{#include ../../../../banners/hacktricks-training.md}} ## Basic Information The time namespace in Linux allows for per-namespace offsets to the system monotonic and boot-time clocks. It is commonly used in Linux containers to change the date/time within a container and adjust clocks after restoring from a checkpoint or snapshot. ## Lab: ### Create different Namespaces #### CLI ```bash sudo unshare -T [--mount-proc] /bin/bash ``` By mounting a new instance of the `/proc` filesystem if you use the param `--mount-proc`, you ensure that the new mount namespace has an **accurate and isolated view of the process information specific to that namespace**.
Error: bash: fork: Cannot allocate memory When `unshare` is executed without the `-f` option, an error is encountered due to the way Linux handles new PID (Process ID) namespaces. The key details and the solution are outlined below: 1. **Problem Explanation**: - The Linux kernel allows a process to create new namespaces using the `unshare` system call. However, the process that initiates the creation of a new PID namespace (referred to as the "unshare" process) does not enter the new namespace; only its child processes do. - Running `%unshare -p /bin/bash%` starts `/bin/bash` in the same process as `unshare`. Consequently, `/bin/bash` and its child processes are in the original PID namespace. - The first child process of `/bin/bash` in the new namespace becomes PID 1. When this process exits, it triggers the cleanup of the namespace if there are no other processes, as PID 1 has the special role of adopting orphan processes. The Linux kernel will then disable PID allocation in that namespace. 2. **Consequence**: - The exit of PID 1 in a new namespace leads to the cleaning of the `PIDNS_HASH_ADDING` flag. This results in the `alloc_pid` function failing to allocate a new PID when creating a new process, producing the "Cannot allocate memory" error. 3. **Solution**: - The issue can be resolved by using the `-f` option with `unshare`. This option makes `unshare` fork a new process after creating the new PID namespace. - Executing `%unshare -fp /bin/bash%` ensures that the `unshare` command itself becomes PID 1 in the new namespace. `/bin/bash` and its child processes are then safely contained within this new namespace, preventing the premature exit of PID 1 and allowing normal PID allocation. By ensuring that `unshare` runs with the `-f` flag, the new PID namespace is correctly maintained, allowing `/bin/bash` and its sub-processes to operate without encountering the memory allocation error.
#### Docker ```bash docker run -ti --name ubuntu1 -v /usr:/ubuntu1 ubuntu bash ``` ### Check which namespace is your process in ```bash ls -l /proc/self/ns/time lrwxrwxrwx 1 root root 0 Apr 4 21:16 /proc/self/ns/time -> 'time:[4026531834]' ``` ### Find all Time namespaces ```bash sudo find /proc -maxdepth 3 -type l -name time -exec readlink {} \; 2>/dev/null | sort -u # Find the processes with an specific namespace sudo find /proc -maxdepth 3 -type l -name time -exec ls -l {} \; 2>/dev/null | grep ``` ### Enter inside a Time namespace ```bash nsenter -T TARGET_PID --pid /bin/bash ``` ## Manipulating Time Offsets Starting with Linux 5.6, two clocks can be virtualised per time namespace: * `CLOCK_MONOTONIC` * `CLOCK_BOOTTIME` Their per-namespace deltas are exposed (and can be modified) through the file `/proc//timens_offsets`: ``` $ sudo unshare -Tr --mount-proc bash # -T creates a new timens, -r drops capabilities $ cat /proc/$$/timens_offsets monotonic 0 boottime 0 ``` The file contains two lines – one per clock – with the offset in **nanoseconds**. Processes that hold **CAP_SYS_TIME** _in the time namespace_ can change the value: ``` # advance CLOCK_MONOTONIC by two days (172 800 s) echo "monotonic 172800000000000" > /proc/$$/timens_offsets # verify $ cat /proc/$$/uptime # first column uses CLOCK_MONOTONIC 172801.37 13.57 ``` If you need the wall clock (`CLOCK_REALTIME`) to change as well you still have to rely on classic mechanisms (`date`, `hwclock`, `chronyd`, …); it is **not** namespaced. ### `unshare(1)` helper flags (util-linux ≥ 2.38) ``` sudo unshare -T \ --monotonic="+24h" \ --boottime="+7d" \ --mount-proc \ bash ``` The long options automatically write the chosen deltas to `timens_offsets` right after the namespace is created, saving a manual `echo`. --- ## OCI & Runtime support * The **OCI Runtime Specification v1.1** (Nov 2023) added a dedicated `time` namespace type and the `linux.timeOffsets` field so that container engines can request time virtualisation in a portable way. * **runc >= 1.2.0** implements that part of the spec. A minimal `config.json` fragment looks like: ```json { "linux": { "namespaces": [ {"type": "time"} ], "timeOffsets": { "monotonic": 86400, "boottime": 600 } } } ``` Then run the container with `runc run `. > NOTE: runc **1.2.6** (Feb 2025) fixed an "exec into container with private timens" bug that could lead to a hang and potential DoS. Make sure you are on ≥ 1.2.6 in production. --- ## Security considerations 1. **Required capability** – A process needs **CAP_SYS_TIME** inside its user/time namespace to change the offsets. Dropping that capability in the container (default in Docker & Kubernetes) prevents tampering. 2. **No wall-clock changes** – Because `CLOCK_REALTIME` is shared with the host, attackers cannot spoof certificate lifetimes, JWT expiry, etc. via timens alone. 3. **Log / detection evasion** – Software that relies on `CLOCK_MONOTONIC` (e.g. rate-limiters based on uptime) can be confused if the namespace user adjusts the offset. Prefer `CLOCK_REALTIME` for security-relevant timestamps. 4. **Kernel attack surface** – Even with `CAP_SYS_TIME` removed, the kernel code remains accessible; keep the host patched. Linux 5.6 → 5.12 received multiple timens bug-fixes (NULL-deref, signedness issues). ### Hardening checklist * Drop `CAP_SYS_TIME` in your container runtime default profile. * Keep runtimes updated (runc ≥ 1.2.6, crun ≥ 1.12). * Pin util-linux ≥ 2.38 if you rely on the `--monotonic/--boottime` helpers. * Audit in-container software that reads **uptime** or **CLOCK_MONOTONIC** for security-critical logic. ## References * man7.org – Time namespaces manual page: * OCI blog – "OCI v1.1: new time and RDT namespaces" (Nov 15 2023): {{#include ../../../../banners/hacktricks-training.md}}