[PATCH v2 1/3] capabilities: Introduce CAP_CHECKPOINT_RESTORE

[Cyrill Gorcunov]

On Tue, Jun 09, 2020 at 08:09:49PM +0000, Nicolas Viennot wrote:
> >>  proc_map_files_get_link(struct dentry *dentry,
> >>  			struct inode *inode,
> >>  		        struct delayed_call *done)
> >>  {
> >> -	if (!capable(CAP_SYS_ADMIN))
> >> +	if (!(capable(CAP_SYS_ADMIN) || capable(CAP_CHECKPOINT_RESTORE)))
> >>  		return ERR_PTR(-EPERM);
> 
> > First of all -- sorry for late reply. You know, looking into this code more I think
> this CAP_SYS_ADMIN is simply wrong: for example I can't even fetch links for /proc/self/map_files.
> Still /proc/$pid/maps (which as well points to the files opened) test for ptrace-read permission.
> I think we need ptrace-may-attach test here instead of these capabilities (if I can attach to
> a process I can read any data needed, including the content of the mapped files, if only
> I'm not missing something obvious).
> 

Nikolas, could you please split the text lines next time, I've had to add newlines into reply manually :)

> Currently /proc/pid/map_files/* have exactly the same permission checks as /proc/pid/fd/*, with the exception
> of the extra CAP_SYS_ADMIN check. The check originated from the following discussions where 3 security issues are discussed:
> http://lkml.iu.edu/hypermail/linux/kernel/1505.2/02524.html
> http://lkml.iu.edu/hypermail/linux/kernel/1505.2/04030.html
> 
> From what I understand, the extra CAP_SYS_ADMIN comes from the following issues:
> 1. Being able to open dma-buf / kdbus region (referred in the referenced email as problem #1).
> I don't fully understand what the dangers are, but perhaps we could do CAP_SYS_ADMIN check
> only for such dangerous files, as opposed to all files.

As far as I remember we only need to read the content of mmap'ed files and if I've ptrace-attach
permission we aready can inject own code into a process and read anything we wish. That said we probably
should fixup this interface like -- test for open mode and if it is read only then ptrace-attach
should be enough, if it is write mode -- then we require being node's admin instead of just adding
a new capability here. And thanks a huge for mail reference, I'll take a look once time permit.