This commit is contained in:
1 parent
7f26ca4f2f
commit
db23feada5
8 files changed
+561
-15
No files matched your search
@@ -0,0 +1,118 @@
|
||||
# Design
|
||||
|
||||
## Context
|
||||
|
||||
See `proposal.md` — Why.
|
||||
|
||||
The relevant facts about the running system, established by inspecting a live
|
||||
container of this image (CUPS 2.4.2 on Debian bookworm):
|
||||
|
||||
- CUPS advertises the queue on both `_ipp._tcp` and `_ipps._tcp`; macOS and iOS
|
||||
prefer the latter and check the certificate's subject alternative names
|
||||
against the name they connected to. CUPS' own clients (`lpadmin`, `ipptool`)
|
||||
do not perform that check.
|
||||
- The name clients use is the one Avahi advertises (`<host>.local`, queryable
|
||||
over D-Bus as `GetHostNameFqdn`), not the container host name. The two can
|
||||
differ: observed example, host name `host.containers.internal` served while
|
||||
Avahi advertised `podp.local`.
|
||||
- CUPS generates a certificate on demand for the *host name*, so that
|
||||
certificate cannot cover names the host name does not imply (in the example
|
||||
above, `podp.local` was absent, and so was a `.local` form, because CUPS only
|
||||
appends `.local` to dot-less host names).
|
||||
- In CUPS 2.4, `ServerCertificate`/`ServerKey` are accepted and preserved by
|
||||
`cupsctl` (which rewrites `cupsd.conf`), but cupsd still served the
|
||||
certificate kept at `<certificate dir>/<host name>.crt`. Writing our
|
||||
certificate to that path is what changed the served certificate.
|
||||
- `/run` survives a container restart (`podman restart`, systemd restart), and
|
||||
`/etc/cups` is a named volume that survives container replacement.
|
||||
|
||||
## Goals / Non-Goals
|
||||
|
||||
**Goals:**
|
||||
- The certificate CUPS presents for `_ipps` validates for every name a client
|
||||
may use, including the Avahi-advertised one.
|
||||
- Restarting the container is safe and does not reissue the certificate.
|
||||
- Everything stays self-contained: no external CA, no configuration file edits
|
||||
by the user, no new volumes.
|
||||
|
||||
**Non-Goals:**
|
||||
- Obtaining a publicly trusted certificate or running a private CA.
|
||||
- Obtaining the printer's own (non-AirPrint) capabilities over IPP.
|
||||
- Changing how jobs are converted or delivered to the printer.
|
||||
|
||||
## Decisions
|
||||
|
||||
**Generate the certificate in the entrypoint rather than shipping one.**
|
||||
A committed certificate would have a fixed key (shared by every deployment), a
|
||||
fixed expiry, and no knowledge of the local host name. Generating at first start
|
||||
keeps the key per deployment and lets the SANs reflect the actual names.
|
||||
|
||||
**Collect SANs from several sources instead of one "server name" setting.**
|
||||
Clients may reach the server by the Avahi FQDN, the container host name, a
|
||||
`.local` form, an IP address or a DNS name users add themselves. Adding all of
|
||||
them is cheap and only ever widens acceptance: `DNS:<host name>`,
|
||||
`DNS:<host name>.local` (when the host name has no dot), the Avahi FQDN,
|
||||
`localhost`, `localhost.localdomain`, `host.containers.internal`,
|
||||
`IP:127.0.0.1`, and the comma-separated `CERT_EXTRA_NAMES`. SANs are used
|
||||
because modern clients ignore the common name; the common name is set to the
|
||||
host name deliberately (see next decision).
|
||||
|
||||
**Keep the common name equal to the host name.**
|
||||
CUPS' certificate handling is name-based; keeping the common name equal to the
|
||||
host name keeps it consistent with CUPS' own notion of its identity, while the
|
||||
SANs carry the names clients actually use. `basicConstraints=CA:TRUE` is set so
|
||||
the file can also be imported as a trust anchor on macOS/iOS.
|
||||
|
||||
**Install the certificate in two places.** As `ServerCertificate`/`ServerKey`
|
||||
(the documented way, added to `cupsd.conf` when the directives are absent —
|
||||
which also repairs a `cupsd.conf` that `cupsctl` rewrote), *and* as
|
||||
`<name>.crt`/`.key` for both the host name and the Avahi FQDN, because that is
|
||||
where cupsd 2.4 actually loaded the certificate from in testing. Belt and
|
||||
braces: either path alone would be enough on some versions, together they hold
|
||||
on the version in the image. The copies are rewritten on every start, since
|
||||
cupsd may replace them.
|
||||
|
||||
**Reuse the certificate whenever it already covers the required names.**
|
||||
Reissuing forces every client that trusted the old certificate to trust a new
|
||||
one, so regeneration is limited to genuinely missing names, and the recorded
|
||||
SAN list is compared as a set (sorted, deduplicated) so the order in which
|
||||
names are discovered does not matter. A transient failure to query Avahi
|
||||
therefore does not churn the certificate. Consequence: names are never *removed*
|
||||
from an existing certificate; that requires deleting it.
|
||||
|
||||
**Treat TLS failures as non-fatal, but fix the restart defect properly.**
|
||||
If the certificate directory is not writable (for example a read-only bind
|
||||
mount), the entrypoint warns and lets CUPS fall back to its own certificate,
|
||||
because plain IPP printing still works. For startup, the stale-state problem is
|
||||
fixed at its root: D-Bus and Avahi are started based on whether their *process*
|
||||
is running, and their leftover socket/PID files are removed first. Avahi failing
|
||||
to start is reported as a warning rather than ending the entrypoint, so a
|
||||
discovery problem cannot take the whole print server down.
|
||||
|
||||
## Risks / Trade-offs
|
||||
|
||||
- [Self-signed certificate is rejected by validating clients] → Documented:
|
||||
trust the certificate on macOS/iOS, or keep using plain IPP. The entrypoint
|
||||
cannot influence a client's trust store.
|
||||
- [Only the common name is known on first start if Avahi is not up yet] → The
|
||||
name is re-checked on every start and the certificate is reissued when a
|
||||
required name is missing, so the next start heals it. Host networking keeps
|
||||
this rare.
|
||||
- [A name that later disappears stays in the certificate] → Accepted: extra
|
||||
SANs are harmless and churn would invalidate client trust.
|
||||
- [Destroying `server.san` loses the record of covered names] → Worst case the
|
||||
certificate is reissued once, which is the same as losing the certificate.
|
||||
- [Restarting rewrites `<name>.crt`, overwriting a certificate the user placed
|
||||
there by hand] → Accepted; this image owns the certificate directory.
|
||||
- [Host name changes are not picked up until a restart] → Accepted; the
|
||||
certificate is inspected at startup only.
|
||||
|
||||
## Migration Plan
|
||||
|
||||
Deploy by rebuilding the image; no data migration. An existing `/etc/cups`
|
||||
volume is repaired on first start: the `ServerCertificate`/`ServerKey`
|
||||
directives are appended to the existing `cupsd.conf` and the stale
|
||||
`<host name>.crt` is replaced. Clients that already added the queue as a
|
||||
*Generic PostScript printer* must remove and re-add it, because macOS caches the
|
||||
driver choice. Rollback is redeploying the previous image; a certificate written
|
||||
by this version is ignored by it.
|
||||
Reference in new issue
Block a user