[libvirt] RFC: virInterface change transaction API

[Laine Stump]

I've been asked to implement what some people have termed as a
"transaction-oriented" API for host interface configuration (ie 
virInterface*()).
The basic intent is to allow rollback to a known-good config if anything 
goes
wrong when changing around the host network config with virInterface*()
functions.

The most straightforward way to achieve this is that prior to calling
virInterfaceDefine/virInterfaceUndefine, the current state of the
host's network configuration (ie the /etc/sysconfig/network-scripts/ifcfg-*
files in the case of Fedora and RHEL) would be saved off somewhere, and
kept around until we're sure the new config is good; once we know that,
we can just eliminate the backup. If, however, the user of virInterface*()
explicitly requests, we could copy the files back; alternately if the system
is rebooted without these known-good files being erased, we would assume
that something went wrong and restore the original config.

As with all other virInterface functions, the details of all this will
be handled by netcf (and below), but since libvirt is the main consumer
of netcf, I figure this is the appropriate place to discuss how it gets 
done,
so please let me know any opinions on any piece of this. I plan to start
the implementation "soon", as I want to be finished before the end of
May.

I see 3 layers to this:

1) libvirt

    At the libvirt layer, this feature just requires 3 new APIs, which
    are directly passed through to netcf:

        virInterfaceChangeStart(virConnectPtr conn, unsigned int flags);
        virInterfaceChangeCommit(virConnectPtr conn, unsigned int flags);
        virInterfaceChangeRollback(virConnectPtr conn, unsigned int flags);

    For the initial implementation, these will be simple passthroughs
    to similarly named netcf functions. (in the future, it would be
    useful for the server side of libvirt to determine if client<->server
    connectivity was lost due to the network changes, and automatically
    tell netcf to do a rollback).

2) netcf

    The netcf api will have these same three APIs, just named slightly
    differently:

         ncf_change_start(struct netcf *ncf, unsigned int flags);


            There are two possibilities for this. Either:

             A) call the initscript described below to save all config
                files that might possibly be changed (snapshot_config)

               or

             B) set a flag in *ncf indicating that all future calls
                to netcf that would end up modifying a particular
                config file should save off that file *if it hasn't
                already been saved*.

             (A) is simpler, but relies on the initscript having
             exact/complete matching knowledge of what files netcf may
             change. Should we worry about that and deal with the
             complexities of (B), or is (A) good enough for now?


         ncf_change_rollback(struct netcf *ncf, unsigned int flags);

            Again, two possbilities:

            A)
               a) save the config of all current interfaces (in memory)
               b) call the initscript below to restore the config to its
                  original state.
               c) compare the new config to the old, and:
                  * bring down any interfaces that no longer exist
                    (PROBLEM: once an interface has no config files, you can
                     no longer operate on it with "ifdown")
                  * bounce any interfaces that have changed
                  * bring up any interfaces that have been re-added
             or

            B)
                a) ifdown all interfaces
                b) call initscript to restore previous config 
(rollback_config)
                c) ifup all interfaces.

            (A) is much simpler, but may lead to unnecessary
            difficulties when we bounce interfaces that didn't really
            need it. So, the same question oas for ncf_change_start() -
            is the more exact operation worth the extra complexity?



         ncf_change_commit(struct netcf *ncf, unsigned int flags);

             The simplest function - this will just call the initscript
             to erase the backup (commit_config).


3) initscript

    This initscript will at first live in (be installed by) netcf
    (called /etc/init.d/networking-config?), but hopefully it will
    eventually be accepted by the initscripts package (which includes
    the networking-related initscripts), as it is of general use. (Dan
    Kenigsberg already already took a stab at this script last year,
    but received no reply from the initscripts maintainers, implying
    they may not be too keen on the idea right now - it might take some
    convincing ;-)

      
https://fedorahosted.org/pipermail/initscripts-devel/2010-February/000025.html

    It will have three commands, one of which will be called
    automatically by "start" (the command called automatically at boot
    time):

    snapshot_config

      This will save a copy of (what the script believes are - is this
      problematic?) all network-config related files. It may or may not
      be called by netcf (see the notes in ncf_start_change() above.

      If this function finds that a snapshot has already been taken,
      it should fail.


    rollback_config (automatically called from "start" at boottime)

      This will move back (from the saved copies) all files that were
      changed/removed since snapshot, *and delete any files that have
      been added*.

      Note that this command doesn't need to worry about ifup/ifdown,
      because it will be called prior to any other networking startup
      (part of the reason that netcf will need to deal with that).

      I notice that Dan K's version saves the modified files to a
      "rollback-${date}" directory. Does this seem like a good idea?
      It's nice to not lose anything, but there is no provision for
      eliminating old versions, so it could grow without bound.

    commit_config

      This will just remove all the files in the save directory.


So, the two problems I have right now:

1) Do we accept the inexact method of just saving all files that match
    a list of patterns during *start(), then in *rollback() erasing all
    files matching that pattern and copying the old file back? Or do we
    need to keep track of what files have been changed/removed and added,
    and copy back / delete only those files during rollback?

    (A version control system would keep track of this rather nicely,
    but that's too complex for something that's intended to be a
    failsafe (and that we would also like to eventually be in the base
    OS install). Dan B. at one point suggested using patchfiles if I
    wanted the save info to keep exact track of which files would need
    to be replaced/deleted on rollback, but on further thought this
    turns out to not be workable, since we would need to run diff (to
    create the patchfile) after all changes had been made, and any
    outside changes to any of the files would leave the patchfile
    un-appliable, thus causing our "failsafe" to fail :-( ). Therefore,
    we will need to rely on the list of globs to tell us what files
    need to be deleted, or keep our own list in a separate file.)

2) Is it going to be okay to ifdown all interfaces prior to the
    rollback, and ifup all interfaces afterwards? Or must we compare
    the new config to the original, and ifdown only those interfaces
    that had been previously added/changed, then ifup only those
    interfaces that had been previously removed/changed?

3) If anyone has ideas on making the initscript more palatable to the
    initscripts people, please speak up! :-) (one comment from an 
initscripts
    person was that 1) for the general case it would be difficult to 
draw the
    line on what parts of network connectivity should be included in this
    rollback functionality, and 2) at some point this becomes a general
    system config problem, and would really be better addressed by a
    general system wide config management system. These are both
    concerns that need well qualified answers. (I tend to think that this
    is intended as a failsafe to prevent unreachable systems, so it should
    be as simple as possible, and thus shouldn't be burdened with the
    complexity of a full system config management system (which could
    also co-exist at a higher level), but better answers are welcome.)