Certificate automation has permission modules that are designed to
prevent inappropriate issuance of unbounded or wildcard certificates.
When an explicit cert manager is used, no additional permission should
be necessary. For example, this should be a valid caddyfile:
https:// {
tls {
get_certificate tailscale
}
respond OK
}
This is accomplished when provisioning an AutomationPolicy by tracking
whether there were explicit managers configured directly on the policy
(in the ManagersRaw field). Only when a number of potentially unsafe
conditions are present AND no explicit cert managers are configured is
an error returned.
The problem arises from the fact that ctx.LoadModule deletes the raw
bytes after loading in order to save memory. The first time an
AutomationPolicy is provisioned, the ManagersRaw field is populated, and
everything is fine.
An AutomationPolicy with no subjects is treated as a special "catch-all"
policy. App.createAutomationPolicies ensures that this catch-all policy
has an ACME issuer, and then calls its Provision method again because it
may have changed. This second time Provision is called, ManagesRaw is no
longer populated, and the permission check fails because it appears as
though the policy has no explicit managers.
Address this by storing a new boolean on AutomationPolicy recording
whether it had explicit cert managers configured on it.
Also fix an inverted boolean check on this value when setting
failClosed.
Updates #6060
Updates #6229
Updates #6327
Signed-off-by: Will Norris <will@tailscale.com>
Set the requested server name in a context value for CertGetter
implementations to use. Pass ctx to tscert.GetCertificateWithContext.
Signed-off-by: Will Norris <will@tailscale.com>
* caddytls: Evict internal certs from cache based on issuer
During a config reload, we would keep certs in the cache fi they were used by the next config. If one config uses InternalIssuer and the other uses a public CA, this behavior is problematic / unintuitive, because there is a big difference between private/public CAs.
This change should ensure that internal issuers are considered when deciding whether to keep or evict from the cache during a reload, by making them distinct from each other and certs from public CAs.
* Make sure new TLS app manages configured certs
* Actually make it work
* Add option to configure certificate lifetime
* Bump CertMagic dep to latest master commit
* Apply suggestions and ran go mod tidy
* Update modules/caddytls/acmeissuer.go
Co-authored-by: Matt Holt <mholt@users.noreply.github.com>
---------
Co-authored-by: Matt Holt <mholt@users.noreply.github.com>
* caddytls: Make on-demand 'ask' permission modular
This makes the 'ask' endpoint a module, which means that developers can
write custom plugins for granting permission for on-demand certificates.
Kicking myself that we didn't do it this way at the beginning, but who coulda known...
* Lint
* Error on conflicting config
* Fix bad merge
---------
Co-authored-by: Francis Lavoie <lavofr@gmail.com>
* tls: modularize client authentication trusted CA
* add `omitempty` to `CARaw`
* docs
* initial caddyfile support
* revert anything related to leaf cert validation
The certs are used differently than the CA pool flow
* complete caddyfile unmarshalling implementation
* Caddyfile syntax documentation
* enhance caddyfile parsing and documentation
Apply suggestions from code review
Co-authored-by: Francis Lavoie <lavofr@gmail.com>
* add client_auth caddyfile tests
* add caddyfile unmarshalling tests
* fix and add missed adapt tests
* fix rebase issue
---------
Co-authored-by: Francis Lavoie <lavofr@gmail.com>
* use gofmput to format code
* use gci to format imports
* reconfigure gci
* linter autofixes
* rearrange imports a little
* export GOOS=windows golangci-lint run ./... --fix
Ideally I'd just remove the parameter to caddy.Context.Logger(), but
this would break most Caddy plugins.
Instead, I'm making it variadic and marking it as partially deprecated.
In the future, I might completely remove the parameter once most
plugins have updated.
Errors returned from the DecisionFunc (whether to get a cert on-demand)
are used as a signal whether to allow a cert or not; *any* error
will forbid cert issuance.
We bubble up the error all the way to the caller, but that caller is the
Go standard library which might gobble it up.
Now we explicitly log connection errors so sysadmins can
ensure their ask endpoints are working.
Thanks to our sponsor AppCove for reporting this!
