Follow up on auditing cmdline

[Steve Grubb]

On Tuesday, November 05, 2013 08:43:03 PM William Roberts wrote:
> So this still seems to be lingering as unresolved in my mind.

Yes, it is.

> I need to find out what the remaining reservations are on this feature. I am
> going to try and summarize...
> 
> Steve Grub:
> 1. Anyway to use argv values as cmdline could be a page (too big)
> 2. Doesn't like disappearing audit entries

There's more to it than this. I think you are overlooking my main point. This 
is a generic problem for all arches. Why don't we just solve the problem once 
and for all?

pgname is limited to 16 bytes. Why? Because that was a reasonable compromise 
way back in time when program names were short. With the Android naming 
convention 16 bytes is insufficient for anything listing processes like lsof, 
ps, netstat, etc. Why can't a new define for prctl be created to set an 
extended name? PR_SET_PROCTITLE. From what I saw in past discussions on LKML, 
there was a lot of concern that a trojan could change the proc title to 
something innocent and hide.

i think that is a losing argument because there are quite a few ways to 
accomplish the same thing. That is basically arguing for the status quo and 
that means the problem never gets fixed.

I think with the rise of Android's popularity, there is a growing importance 
to fix the problem.


> Richard Briggs:
> 1. Can we make it dynamic on/off

I object to this because we do not have any precedent for configuring a rule to 
get something and then another configuration to get all of it. I have worked on 
the audit system since 2004. A lot of maintainers have come and gone in that 
time. I don't want to see this subsystem have this oddity when they (people 
currently contributing patches) lose interest and go away.


> Stephen Smalley:
> 1. Can we cache the data for performance reasons
> 
> So I addressed RGB's issues, which led to one of steve Grub's concerns.
> Which I can address both with if feature on then print cmdline=value else
> print cmdline=(null)
> 
> Unfortunately the data I want to audit, is the full proc/cmdline entry,
> which I think is the most
> generic way of getting at potential vm data through various fork mazes on
> Android, as well
> as gathering the data on other architectures as well.

You know, we have the same issue with sudo. We have to get the full command 
line. What was done is sudo was patched to collect it and it emits a 
AUDIT_USER_CMD event. Maybe this works for you?

-Steve

> This also prevents us from hitting the
> 16 char width issue on task->comm. Increasing that will result in more
> non-pageable kernel
> memory use, versus my transient use of a page. I also need to make sure I
> can get this
> data before the process terminates, which can happen if I try to acquire it
> in user-space.
> 
> Also, on error conditions, the last patch version will not print
> cmdline=(null) which is an error and can be trivially corrected.
> 
> But before I put more time into it, I want to make sure the underlying idea
> will be accepted, architectures, cacheing, print formats etc are all
> trivial.