Each collection of installed package instances is some kind of image. We envision three kinds:
- entire images
- An entire image contains (and can contain) all appropriate packages from a repository. It has a machine type and release.
- partial images
- A partial image contains (and can contain) all appropriate packages from a repository. An partial image is tied to an entire image and has an identified list of non-writeable directory paths. It inherits its machine type and release.
- user images
- A user image consists entirely of relocatable packages. It is tied to an entire image for dependency satisfaction of non-relocatable packages. It inherits its machine type and release.
Some aspects of configuration can be shared between all images. For instance, a user image may specify publishers beyond those encoded into the system defaults. So, a user image must have authoritative configuration, but be able to draw on the parent image's state for default settings.
Operations on partial images and non-live entire images may need to access image configuration data when smf(5) for that image is unavailable. So images of these kinds must have cached configuration state.
Roughly, these aspects are sufficient to constrain our configuration behaviour for user and entire images:
- [user image]
- look up local configuration if undefined, request properties from svc://application/pkg if unavailable, examine parent's configuration cache if undefined or unavailable, use hard-coded default
- [entire image]
- request properties from svc://application/pkg if unavailable, examine configuration cache if undefined or unavailable, use hard-coded default
Partial images could have differing behaviour depending on whether the operation is invoked from within the partial image's "packaging context" (as it would be for an operation issued from within the zone), or outside it (operations on the entire image and its associated partial images).
For the first case, the configuration strategy is the same as that for an entire image. For the second case, we could do
[partial image, external context]
- examine partial image configuration cache
- if unavailable, request properties from svc://application/pkg
- if undefined or unavailable, examine entire image configuration cache
- if undefined, use hard-coded default
For certain properties (or even certain packages), it may be inappropriate to let the partial image configurations drift from that of the entire image.
List of publishers. For each publisher, we have a prefix, an origin URL, a list of mirror URLs, and annotations.
- publisher_[escaped_name]/ Property group of type "com.sun.pkg,publisher"
/prefix pkg: publisher /origin http:, https:, or ftp: URL /mirrors list of URLs /disabled boolean indicating whether publisher should be used Image properties. The image has a collection of simple properties, like machine type and policies that control the behavior of the pkg(5) software within that image. Policies are properties with a boolean value.
- properties/
/pursue-latest (not currently used) /require-optional (not currently used) /display-copyrights (not currently used) /flush-content-cache-on-success should downloaded compressed files be removed after a successful install /preferred-publisher preferred publisher for unknown package lookups /title title of image for use in GUIs and other UIs /description longer description of the content of the image /[various]
Entire images have a property group for each tied partial image.
partial_[escaped_name]/ Property group of type "com.sun.pkg,partial"
/path Filesystem path /vdomainname Defined if this image is a virtual domain controlled on this system. /vdomaintype "xen", "zone", ... (XXX Should we instead assume that each of Zones and Xen will acquire service FMRIs per zone/domain?)
The intent is to utilize snapshot-capable filesystems to provide a rollback mechanism that includes both the pkg(1M)-driven modifications as well as subsequent modifications from configuration methods, etc. In most cases, we are considering a filesystem with capabilities similar to ZFS.
XXX Is fssnap_ufs(1M) sufficient to build something meaningful for UFS?
With appropriate policies set, the image plan, prior to execution must snapshot the involved filesystems.
There seem to be two options:
The image can build its list of filesystems and types (plus other attributes, like mutability.
A list of filesystems is given to each package plan, the package plan then evaluates its proposed actions against that list, and offers a method to return the affected subset of the list.
In this case, we must also determine whether or not we are restricted to clones (because one or more packages in the image plan require kernel-restart) or are potentially live.
XXX Either of these approaches is applicable in the image/substrate model, where the substrate owns the low-level properties specific to a particular "place to write data".
In the case that one or more filesystems in the image is or are not capable of snapshots, we have two choices:
- take no snapshot, as image-revert not possible in any safe or complete sense
- take a set of snapshots that lead to a revert point that requires manual forcing
We must warn about images of this kind, unless policy settings allow us otherwise. Since we want to allow and understand "zfs split"-style operations [1], we also need to determine if a snapshot set taken before a split can sensibly be restored after the split operation.
[1] (The zfs split RFE.)