[CRIU] [PATCHv5 3/3] crtools: cr_service() meat and a few fixes to properly dump cr_service socket

[Pavel Emelyanov]

On 09/09/2013 11:46 PM, Ruslan Kuprieiev wrote:
> On 09/09/2013 10:51 PM, Pavel Emelyanov wrote:
>> On 09/09/2013 04:30 PM, Ruslan Kuprieiev wrote:
>>> +		pr_perror("Can't stat images direcrtory");
>>> +		goto err;
>>> +	}
>>> +
>>> +	if (ids.uid == st.st_uid || ids.gid == st.st_gid || ids.uid == 0) {
>> This check is pointless. Kernel would do it itself properly.
> 
> But criu will open dir as root. And here we are checking if client has 
> permissions to write to this directory.

Crap :( Well, this check is wrong anyway -- the permissions to access
a directory are not just about ids matching.

Then we have to do a trick -- an application opens the directory and sends
us just the file descriptor number in the protobuf message. Criu service then
goes and opens /proc/$pid/fd/$img_fd and sets one as destination directory.

Got it?

>> I think in RPC we should use raw pb messages w/o u32 size-s pb_foo 
>> engine adds. 
> 
> Why? Isn't sending sizes of msgs good?

For RPC -- no. This engine is planned to be used as-is by python/java/ruby
and other apps, and forcing them to mess with uint32-s and alike is not nice.
Just protobuf messages.