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 Andrew Morton (akpm@linux-foundation.org):
>>>> On Mon, 20 Oct 2008 01:40:30 -0400
>>>> Oren Laadan <orenl@cs.columbia.edu> wrote:
>>>>> asmlinkage long sys_checkpoint(pid_t pid, int fd, unsigned long flags)
>>>>> {
>>>>> - pr_debug("sys_checkpoint not implemented yet\n");
>>>>> - return -ENOSYS;
>>>>> + struct cr_ctx *ctx;
>>>>> + int ret;
>>>>> +
>>>>> + /* no flags for now */
>>>>> + if (flags)
>>>>> + return -EINVAL;
>>>>> +
>>>>> + ctx = cr_ctx_alloc(pid, fd, flags | CR_CTX_CKPT);
>>>>> + if (IS_ERR(ctx))
>>>>> + return PTR_ERR(ctx);
>>>>> +
>>>>> + ret = do_checkpoint(ctx);
>>>>> +
>>>>> + if (!ret)
>>>>> + ret = ctx->crid;
>>>>> +
>>>>> + cr_ctx_free(ctx);
>>>>> + return ret;
>>>>> }
>>>> Is it appropriate that this be an unprivileged operation?
>>> Early versions checked capable(CAP_SYS_ADMIN), and we reasoned that we
>>> would later attempt to remove the need for privilege so that all users
>>> could safely use it.
>>>
>>> Arnd Bergmann called us on that nonsense, pointing out that it'd make
>>> more sense to let unprivileged users use them now, so that we'll be
>>> more careful about the security as patches roll in.
>>>
>>> So, Oren's patchset right now only checkpoints current, despite pid
>>> being part of the API. So the task can access its own data. When
>>> the patch supports checkpointing another task (which Oren says he's
>>> doing right now), then our intent is to check for ptrace access to
>>> the target task. (Right, Oren?)
>> Correct. That's already in the additional patch in the git tree - first
>> I locate the task and if found, I check ptrace_may_access() (read mode).
>
> 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 !)

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...

>
> ...
>
>> This should be covered by ptrace_may_access() test.
>>
>> In the longer run, I suppose SElinux people would want a security hook
>> there to approve or disapprove the operation.
>
> I think we'll find the ptrace() checks to be so like what we're doing
> that no new check will be needed. But we should definately ask them.
>
> Now may be too early to ask, though. The answer will be clearer once
> more resources are supported.
>
>>>> What happens if I pass it a pid of a process which I _do_ own, but it
>>>> does not refer to a container's init process?
>>> I would assume that do_checkpoint() would return -EINVAL, but it's a
>>> great question: Oren, did you have another plan?
>> Since we intentional provide minimal functionality to keep the patchset
>> simple and allow easy review - we only checkpoint one task; it doesn't
>> really matter because we don't deal with the entire container.
>>
>> With the ability to checkpoint multiple process we will have to ensure
>> that we checkpoint an entire container. I planned to return -EINVAL if
>> the target task isn't a container init(1). Another option, if people
>> prefer, is to use any task in a container to "represent" the entire
>> container.
>
> Except we support nested containers, so unless we only support
> checkpoint of the deepest container, that doesn't work.
>
> ...

yeah.. I just though about it this mornig ;)

>
>>>> Again, this is scary stuff. We're allowing unprivileged userspace to
>>>> feed random numbers into kernel data structures.
>>> Yes, all of the file opens and mmaps must not skip the usual security
>>> checks. The task credentials are currently unsupported, meaning that
>>> euid, etc, come from the caller, not the checkpoint image. When the
>> Actually, the fact that task credentials are not restored makes it
>> more secure, because the user can't do anything beyond her current
>> capabilities.
>
> 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).

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).

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.

Oren.

>
> If we were to use some sort of tpm-signing of statefiles, then
> hallyn restarting a checkpointed /bin/passwd may become doable.
>
>> For the same reason, however, unless we agree on a secure way to
>> elevate credentials, there are various things that we cannot restore,
>> even though it may be something we would want to permit.
>>
>>> restoration of credentials becomes supported, then definately the
>>> caller (of sys_restore())'s ability to setresuid/setresgid to those
>>> values must be checked.
>>>
>>> So that's why we don't want CAP_SYS_ADMIN required up-front. That way
>>> we will be forced to more carefully review each of those features.
>>>
>>>> I'd like to see the security guys take a real close look at all of
>>>> this, and for them to do that effectively they should be provided with
>>>> a full description of the security design of this feature.
>>> Right, some of the above should be spelled out somewhere. Should it be
>>> in the patch description, in the Documentation/checkpoint.txt file,
>>> or someplace else? Oren, do you want to filter the above information
>>> into the right place, or do you want me to do it and send you a patch?
>> I'll add something to the Documentation/checkpoint.txt.
>
> Cool, thanks Oren.
>
> -serge


\
 
 \ /
  Last update: 2008-10-22 18:07    [W:0.096 / U:0.100 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site