[PATCH 2/4] cr: add generic LSM c/r support (v5)

[Oren Laadan]

Serge E. Hallyn wrote:
> Quoting Oren Laadan (orenl at librato.com):
> ...
>>> +	switch (sectype) {
>>> +	case CKPT_SECURITY_MSG_MSG:
>>> +		str = security_msg_msg_checkpoint(security);
>>> +		break;
>>> +	case CKPT_SECURITY_IPC:
>>> +		str = security_ipc_checkpoint(security);
>>> +		break;
>>> +	case CKPT_SECURITY_FILE:
>>> +		str = security_file_checkpoint(security);
>>> +		break;
>>> +	case CKPT_SECURITY_CRED:
>>> +		str = security_cred_checkpoint(security);
>>> +		break;
>>> +	default:
>>> +		str = ERR_PTR(-EINVAL);
>>> +		break;
>>> +	}
>> Let me suggest a different scheme (also last night's IRC); I think it's
>> less hackish and uses better the existing {checkpoint,restore}_obj().
>>
>> * Define one obj type CKPT_OBJ_SEC_{IPC, MSG_MSG, FILE, CRED}, with
>> matching c/r functions security_{c,r}_{ipc,msg_msg,file,cred}_obj()
>>
>> * Define one obj type for the string representation CKPT_OBJ_SEC_STR
>> with matchin c/r functions security_{c,r}_string_obj()
>>
>> * The helper will now:
>>
>> 	security_checkpoint_obj()
>> 	{
>> 		switch (type) {
>> 		case CKPT_OBJ_SEC_IPC:
>> 			ret = checkpoint_obj(ctx, sec, CKPT_OBJ_SEC_IPC);
>> 			break;
>> 		case CKPT_OBJ_SEC_CRED:
>> 			ret = checkpoint_obj(ctx, sec, CKPT_OBJ_SEC_CRED);
>> 		...
>> 	}
>>
>> 	security_checkpoint_ipc_obj()
>> 	{
>> 		...
>> 		ckpt_lsm_str = str_from_sec_ipc();   /* like you do now */
>> 		objref = checkpoint_obj(ctx, ckpt_lsm_str, CKPT_OBJ_SEC_STR);
>> 		...
>> 		h->objref = objref;
>> 		ckpt_write_obj();
>> 	}
>>
>> Perhaps a variation on this where the string is checkpoint_obj()'ed
>> first would also work.
>>
>> I haven't looked at all the details, but I hope something along these
>> lines would help untangle the current mess.
> 
> So as discussed on irc, that by itself won't work bc (a) smack
> will checkpoint the same void* as multiple objtypes, and the
> objhash will complain.
> 
> Since we've gone over several possibilities on irc, let me summarize
> some here:
> 
> 1. do the restore_security() in the code instead of using an objref
> to have it called automatically.  That stops me having to write an
> objref by hand before writing out the CKT_HDR_CRED.  That's fine
> with me, but then I won't be using checkpoint_obj() either, so I
> want to make sure I'm not going to change all the restore callers
> just to end up nixing that path.

This is similar to how pipes/fifo are handled: first check for a
common inode (CKPT_OBJ_INODE only used during restart).

I agree that it is not as pretty to use it for security void* that
are kind of an independent shared object. However, it is clean.

> 
> 2. alter the objhash to not complain if the same void* is checkpointed
> as a different type.  That may have safety implications for the rest
> of the objhash users, especially at restart where we can't really trust the
> input.

Agree with your concerns, this is not my favorite.

> 
> 3. have security_checkpoint_obj() 'reserve' a dummy objsec by
> stuffing the void* security, then assume that the objref for
> the string representation will be objref(void*security)+1.
> This might cause problems if we later parallelize checkpoint so
> that objref+1 is no longer valid.

I don't like this. Too hacky.

> 
> 4. Add a new field to the struct ckpt_obj which lets us store
> the objref for the string pointer in the ckpt_obj for the void*.

Can you elaborate on what this entails ?

E.g., do you want to be able to store an arbitrary data field for
an object, and add interface to set and get it ?  (If so, what is
the proposed api ?)

> 
> For completeness, the latest version which I actually sent out
> did:
> 
> 5. Define two objhash object types for the lsm obj, one to
> use at checkpoint, and one at restart.  At checkpoint, it
> stuffs the void* security into the objhash and manually writes
> out a checkpoint entry for the context string.  At restart, it
> places a struct containing the context string in the objhash.
> The type used at restart must have ->get/->drop defined so that
> the struct is freed at the end of restart, while at checkpoint
> we can't hvae ->get/->drop bc the void* is opaque (and persistand
> relative to the checkpoint operation).

:(

> 
> And what I was starting on until the latest irc conversation
> was (3).
> 
> At the moment (4) seems to me like the best path.

There is a 6th option: allow callers of checkpoint_obj() to pass
another arbitrary argument to the obj-type-specific checkpoint
function. So you'd pass the type of the security void* along to
it.

Oren.