lkml.org 
[lkml]   [2008]   [Oct]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [RFC v7][PATCH 2/9] General infrastructure for checkpoint restart


    Serge E. Hallyn wrote:
    > Quoting Oren Laadan (orenl@cs.columbia.edu):
    >>
    >> Serge E. Hallyn wrote:
    >>> Quoting Oren Laadan (orenl@cs.columbia.edu):
    >>> Just thinking aloud...
    >>>
    >>> Is read mode appropriate? The user can edit the statefile and restart
    >>> it. Admittedly the restart code should then do all the appropriate
    >>> checks for recreating resources, but I'm having a hard time thinking
    >>> through this straight.
    >>>
    >>> Let's say hallyn is running passwd. ruid=500,euid=0. He quickly
    >>> checkpoints. Then he restarts. Will restart say "ok, the /bin/passwd
    >>> binary is setuid 0 so let hallyn take euid=0 for this?" I guess not.
    >>> But are there other resources for which this is harder to get right?
    >> I'd say that checkpoint and restart are separate.
    >>
    >> In checkpoint, you read the state and save it somewhere; you don't
    >> modify anything in the target task (container). This equivalent to
    >> ptrace read-mode. If you could do ptrace, you could save all that
    >> state. In fact, you could save it in a format that is suitable for
    >> a future restart ... (or just forge one !)
    >
    > Yeah, that's convincing.
    >
    >> In restart, we either don't trust the user and keep everything to
    >> be done with her credentials, of we trust the root user and allow
    >> all operations (like loading a kernel module).
    >>
    >> We can actually have both modes of operations. How to decide that
    >> we trust the user is a separate question: one option is to have
    >> both checkpoint and restart executables setuid - checkpoint will
    >> sign (in user space) the output image, and restart (in user space)
    >> will validate the signature, before passing it to the kenrel. Surely
    >> there are other ways...
    >
    > Makes sense.
    >
    > ...
    >
    >>> Hmm, so do you think we just always use the caller's credentials?
    >> Nope, since we will fail to restart in many cases. We will need a way
    >> to move from caller's credentials to saved credentials, and even from
    >> caller's credentials to privileged credentials (e.g. to reopen a file
    >> that was created by a setuid program prior to dropping privileges).
    >
    > Can we agree to worry about that much much later? :) Would you agree

    Definitely. Even more so - I believe that's a user-space issue :)

    > that for the majority of use-cases, restarting with caller's credentials
    > will work? Or am I wrong about that?

    That depends on your target audience. For HPC you're probably right.
    For server applications this may not be the case (e.g. apache needs
    a privileged port, and then it drops privileges).

    I agree that we may safely (...) defer this discussion until the
    implementation gets much beefier.

    >
    >> To do that, we will need to agree on a way to escalate/change the
    >> credentials. This however belongs to user-space (and then the binaries
    >> for checkpoint/restart will be setuid themselves).
    >
    > Ok those are less scary, and I have no problem with those.
    >
    >> There will also be the issue of mapping credentials: a user A may have
    >> one UID/GID on once system and another UID/GID on another system, and
    >> we may want to do the conversion. This, too, can be done in user space
    >> prior to restart by using an appropriate filter through the checkpoint
    >> stream.
    >
    > User namespaces may help here too. So user A can create a new user
    > namespace and restart as user B in that namespace. But right now that
    > sounds like overkill.

    Indeed, virtualization is probably the solution. Here, too, I think
    it's safe to defer the discussion.

    Oren.




    \
     
     \ /
      Last update: 2008-10-22 20:35    [W:0.030 / U:0.080 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site