Kubernetes' restricted pods drop every capability. The binary carried cap_net_bind_service as a file capability, and exec of such a binary fails outright when the capability isn't in the bounding set (exec /usr/local/bin/inbuxa: operation not permitted with rc.27 under --cap-drop ALL).
What changes
Dockerfile, Dockerfile.build (both stages), Dockerfile.fdb: no setcap, and no libcap package to run it.
INBUXA_LISTENER_PORTS moves the default listeners a new store gets, or leaves some out: smtp=2525,submissions=4465,imaps=9993,pop3s=off,https=8443. Names: smtp, submissions, imaps, pop3s, sieve, https, http. It's checked on every start, and a mistake (unknown name, bad port, a name twice, two listeners on one port) stops the server with a clear message. It's used only when the store has no listeners yet; after that the registry holds them, as before.
A bind refused on a port below 1024 now names what would allow it (CAP_NET_BIND_SERVICE, the sysctl, or a higher port).
Docker users: unchanged. Docker sets net.ipv4.ip_unprivileged_port_start=0 in a container's own network namespace, so the unprivileged server still binds 25/443/465/993. Not so under podman or --network host.
Tested
Unit tests for the parser (3).
rc.27 with its capability stripped, --cap-drop ALL --security-opt no-new-privileges, Docker bridge, fresh store: all seven default ports bound, CapEff 0.
This branch's binary, all caps dropped and ip_unprivileged_port_start=1024 (what a pod gets):
no variable: 25/443/465 refused with the new hint, 4190/8080 up;
the variable above: 2525/4465/9993/8443/4190/8080, no POP3;
imap=143: exits 1, there is no default listener 'imap'; they are smtp, ....
fmt, workspace --no-run, fork checks, changelog check: ok.
For the release notes (Before you upgrade)
The image no longer carries the capability to bind ports below 1024. Docker on its default or a user-defined network: nothing to do. Podman: add --sysctl net.ipv4.ip_unprivileged_port_start=0 (the installer's compose file sets it, inbuxa/inbuxa-installer PR linked below). --network host: add --cap-add NET_BIND_SERVICE. The binary releases and the systemd unit (AmbientCapabilities) are unaffected.
Companion PRs: installer (sysctl in compose), docs (install/server page; merge after the release).
Not touched: Dockerfile.relay; inbuxa-relay isn't part of the Kubernetes recipe.
Closes #424 (K-S1).
Kubernetes' restricted pods drop every capability. The binary carried `cap_net_bind_service` as a file capability, and exec of such a binary fails outright when the capability isn't in the bounding set (`exec /usr/local/bin/inbuxa: operation not permitted` with rc.27 under `--cap-drop ALL`).
**What changes**
- `Dockerfile`, `Dockerfile.build` (both stages), `Dockerfile.fdb`: no `setcap`, and no libcap package to run it.
- `INBUXA_LISTENER_PORTS` moves the default listeners a new store gets, or leaves some out: `smtp=2525,submissions=4465,imaps=9993,pop3s=off,https=8443`. Names: smtp, submissions, imaps, pop3s, sieve, https, http. It's checked on every start, and a mistake (unknown name, bad port, a name twice, two listeners on one port) stops the server with a clear message. It's *used* only when the store has no listeners yet; after that the registry holds them, as before.
- A bind refused on a port below 1024 now names what would allow it (CAP_NET_BIND_SERVICE, the sysctl, or a higher port).
**Docker users: unchanged.** Docker sets `net.ipv4.ip_unprivileged_port_start=0` in a container's own network namespace, so the unprivileged server still binds 25/443/465/993. Not so under podman or `--network host`.
**Tested**
- Unit tests for the parser (3).
- rc.27 with its capability stripped, `--cap-drop ALL --security-opt no-new-privileges`, Docker bridge, fresh store: all seven default ports bound, CapEff 0.
- This branch's binary, all caps dropped and `ip_unprivileged_port_start=1024` (what a pod gets):
- no variable: 25/443/465 refused with the new hint, 4190/8080 up;
- the variable above: 2525/4465/9993/8443/4190/8080, no POP3;
- `imap=143`: exits 1, `there is no default listener 'imap'; they are smtp, ...`.
- fmt, workspace `--no-run`, fork checks, changelog check: ok.
**For the release notes (Before you upgrade)**
The image no longer carries the capability to bind ports below 1024. Docker on its default or a user-defined network: nothing to do. Podman: add `--sysctl net.ipv4.ip_unprivileged_port_start=0` (the installer's compose file sets it, inbuxa/inbuxa-installer PR linked below). `--network host`: add `--cap-add NET_BIND_SERVICE`. The binary releases and the systemd unit (AmbientCapabilities) are unaffected.
Companion PRs: installer (sysctl in compose), docs (install/server page; merge after the release).
Not touched: `Dockerfile.relay`; inbuxa-relay isn't part of the Kubernetes recipe.
jcoffey-dev
added this to the 27.1 milestone 2026-11-18 14:08:54 +00:00
The binary carried cap_net_bind_service as a file capability, so exec
failed outright wherever the capability isn't in the bounding set, as in a
Kubernetes pod with every capability dropped. The images no longer set it
(or install libcap to set it).
Docker users are unaffected: Docker sets ip_unprivileged_port_start=0 in a
container's own network namespace, so the server still listens on 25, 443,
465 and 993 unprivileged. Under podman or --network host, grant the sysctl
or CAP_NET_BIND_SERVICE.
INBUXA_LISTENER_PORTS moves the default listeners a new store gets, or
leaves some out (smtp=2525,pop3s=off,...). It is checked on every start
and a mistake stops the server, but only used when the store has no
listeners yet; after that the registry holds them.
A bind refused on a port below 1024 now says what would allow it.
Closes#424
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Closes #424 (K-S1).
Kubernetes' restricted pods drop every capability. The binary carried
cap_net_bind_serviceas a file capability, and exec of such a binary fails outright when the capability isn't in the bounding set (exec /usr/local/bin/inbuxa: operation not permittedwith rc.27 under--cap-drop ALL).What changes
Dockerfile,Dockerfile.build(both stages),Dockerfile.fdb: nosetcap, and no libcap package to run it.INBUXA_LISTENER_PORTSmoves the default listeners a new store gets, or leaves some out:smtp=2525,submissions=4465,imaps=9993,pop3s=off,https=8443. Names: smtp, submissions, imaps, pop3s, sieve, https, http. It's checked on every start, and a mistake (unknown name, bad port, a name twice, two listeners on one port) stops the server with a clear message. It's used only when the store has no listeners yet; after that the registry holds them, as before.Docker users: unchanged. Docker sets
net.ipv4.ip_unprivileged_port_start=0in a container's own network namespace, so the unprivileged server still binds 25/443/465/993. Not so under podman or--network host.Tested
--cap-drop ALL --security-opt no-new-privileges, Docker bridge, fresh store: all seven default ports bound, CapEff 0.ip_unprivileged_port_start=1024(what a pod gets):imap=143: exits 1,there is no default listener 'imap'; they are smtp, ....--no-run, fork checks, changelog check: ok.For the release notes (Before you upgrade)
The image no longer carries the capability to bind ports below 1024. Docker on its default or a user-defined network: nothing to do. Podman: add
--sysctl net.ipv4.ip_unprivileged_port_start=0(the installer's compose file sets it, inbuxa/inbuxa-installer PR linked below).--network host: add--cap-add NET_BIND_SERVICE. The binary releases and the systemd unit (AmbientCapabilities) are unaffected.Companion PRs: installer (sysctl in compose), docs (install/server page; merge after the release).
Not touched:
Dockerfile.relay; inbuxa-relay isn't part of the Kubernetes recipe.jcoffey-dev referenced this pull request from inbuxa/inbuxa-charts2026-11-18 14:09:09 +00:00