procd/jail, branch master

jail: support linux.resources.hugepageLimits

2026-10-21T11:52:36Z

ujail refuses every OCI spec carrying hugepageLimits, although the cgroup v2 hugetlb controller provides the limits as hugetlb..max and the subtree control computation already enables the controller for hugetlb keys. Write each limit to that file. OCI and the kernel name page sizes alike (64KB, 2MB, 1GB); a size the kernel does not offer has no such file and fails the container like any other limit the host cannot apply. Signed-off-by: Daniel Golle

jail/seccomp: resolve syscalls by all their header names

2026-10-20T17:47:29Z

find_syscall() looks a name up in the table indexed by syscall number, which holds one name per number. musl renames the 32-bit time syscalls on 32-bit targets (__NR_clock_gettime32 for the kernel's clock_gettime), so a seccomp profile allowing clock_gettime or timer_settime resolves to nothing and the syscall stays denied: the Go runtime's clock fails in such a jail on i386 and MIPS and the scheduler aborts. Look names up in a separate table built from both the libc and the kernel headers. Fixes: dfcfcca7baf2 ("add initial version of ujail and utrace") Signed-off-by: Daniel Golle

jail/seccomp: honour the MIPS ISA bit of code addresses

2026-10-20T16:26:47Z

OpenWrt builds MIPS packages as MIPS16 by default, so main() of a jailed binary has bit 0 set. seccomp-inject then pokes a classic break into the middle of the compressed code at that odd address and the workload dies with SIGSEGV before main, so every jail with a seccomp profile fails on MIPS, and the trace-mode breakpoints break the same way. Place breakpoints at the even address with the 16-bit MIPS16e or microMIPS break that AT_HWCAP selects, keep the ISA bit in the resumed PC and run the injected classic syscall at the even address. Fixes: f15d0c407c09 ("jail: apply OCI seccomp filters via ptrace syscall injection") Signed-off-by: Daniel Golle

jail: keep signals pending until ujail can forward them

2026-10-18T12:03:43Z

procd resets the signal handlers of a new instance and unblocks its signals right after fork(). A jailed instance is ujail, which installs its forwarding handlers only late in main(). A signal that arrives in between, from the cgroup set-up to the end of the ujail option scan, uses the default action and kills ujail. odhcpd-update sends SIGHUP to all dnsmasq instances on each lease change, which happens while dnsmasq restarts. Each restart lost all three dnsmasq instances this way, and procd reported a crash for each. Keep the procd signal set blocked across exec for a jailed instance. ujail unblocks it after it installs its handlers, so a pending signal goes to jail_handle_signal(). Before the jail runs, the handler drops a reload signal and records a stop request. Signed-off-by: John Crispin

uxc-net: join a bridge through the device bridge options

2026-10-18T11:31:24Z

netifd adds a device to a bridge it does not own through the bridge and bridge_vlan options of the device itself, read from procd device data. uxc-net publishes the host end of a bridged container's veth as a separate bridge-port entry, which netifd does not read, so the veth never joins the bridge of the attach network. Set bridge and bridge_vlan on the published veth device and drop the bridge-port data. bridge_vlan follows netifd's hotplug vlan syntax: the attach network's own VLAN becomes ":u", the untagged PVID member, and each org.openwrt.network.vlans entry ":t" becomes "", a tagged member. An untagged entry is refused, as the syntax ties untagged to the PVID and a port has one PVID. Signed-off-by: Daniel Golle

uxc-net: take ingress entries in the podman port order

2026-10-18T11:31:24Z

An ingress entry names one port, which is both the external port and the port inside the container, so a container service cannot be published on a different external port. Accept :[:][/], the order podman -p and compose use, with tcp as the default protocol, and set the redirect's dest_port to . A port range is only accepted untranslated, as the DNAT keeps an external port only when it lies in the target range and does not shift a range. The :/ form sealed into existing images keeps its meaning. Signed-off-by: Daniel Golle

jail, uxc: keep the OCI creation time on the uptime clock

2026-10-18T11:25:50Z

ujail reports the created time of the OCI state as an RFC 3339 string of the wall-clock at create, which is wrong on a device that creates its containers before NTP has corrected the clock of a missing or battery-less RTC, and stays wrong afterwards. Report created in seconds of uptime, the value the instance events already carry, and let uxc turn it into an RFC 3339 UTC time with calendar_format() when it prints uxc state or uxc list, so the shown time follows the wall-clock in effect at output. Signed-off-by: Daniel Golle

jail: report state changes to procd on a status socket

2026-10-18T11:25:50Z

procd only learns that the process it started for a container exited, and ujail stays alive in the stopped state until the container is deleted, so procd can tell neither whether the container was created and its payload started nor that the payload died. When descriptor 3 is a socket at startup, take it as the status socket procd hands over, mark it close-on-exec so that hooks and the payload never inherit it, and keep it across the restart re-execution. Each instance.ready, instance.running and stop or create failure is sent as one struct jail_status record of type, flags (reaped, started, requested, escalated, restarting, registered), exit status and payload runtime in seconds, defined in jail-status.h for both sides. A send uses MSG_NOSIGNAL, so a peer that has gone away cannot kill ujail. Signed-off-by: Daniel Golle

jail: apply a resources override at create

2026-10-18T11:25:50Z

Limits that differ from the bundle's linux.resources can only be set with container..update once the container exists, so a container started immediately runs its payload under the bundle's limits until the update lands, and a container that re-executes itself after a netifd restart comes back without the update at all. Take an override file with -W, holding linux.resources members, and apply it right after the bundle is parsed, through the same ceiling check the update method uses, so the values are written when the cgroup is set up and before the payload runs. A missing file means no override, so the path can be passed before an override exists, and since the option is part of the saved argv, a re-execution reads the file again and follows later changes to it. An override the ceiling refuses leaves the bundle's limits in place, is logged, and is reported in instance.ready as resources_refused and resources_errno. Signed-off-by: Daniel Golle

jail: let the env file override the bundle environment

2026-10-18T11:25:50Z

append_envfile() appends every line of the env file after the bundle's process.env and keeps duplicates. getenv() returns the first match, so a variable the image already sets cannot be changed through the env file, and an operator or stack value for it is silently ignored. Let a line from the env file replace an entry of the same name. The env file then overrides the image environment, as an environment given to docker run or podman run does. Engines that pass the final environment in config.json use no env file and are not affected. Signed-off-by: Daniel Golle