reactive audit proposal

[Joe Nall]

> On May 12, 2020, at 7:36 PM, Steve Grubb <sgrubb at redhat.com> wrote:
> 
> On Tuesday, May 12, 2020 8:31:45 PM EDT Joe Nall wrote:
>>> On May 12, 2020, at 7:22 PM, Steve Grubb <sgrubb at redhat.com> wrote:
>>> 
>>> Hello,
>>> 
>>> I wanted to run this by the crowd to see what people's reaction might be.
>>> 
>>> The audit system sometimes needs to have rules applied when something
>>> happens. For example, if someone plugs in a USB flash drive, the system
>>> creates the device in /dev and then automatically mounts it under some
>>> circumstances.
>>> 
>>> I would propose 2 new additions to the audit rule syntax: on-mount and
>>> on-login.These rules would be in a separate file from the main audit
>>> rules. When a file system is mounted, /proc/mounts changes and the mount
>>> table can be scanned to see if something new is there. In this way we
>>> can reliably detect newly mounted filesystems. We can then match against
>>> a specifier to see if this is a file system in which we want to apply
>>> new rules. If so, we send the new rules to the kernel. When the device
>>> is unmounted, the kernel drops all watches on that file system. So, we
>>> only need to worry about when a device is mounted.
>>> 
>>> This works good for anything that gets mounted. But it is also possible
>>> for a USB flash drive to be accessed as a block device, such as the dd
>>> utility. If we had to detect device discovery, there is a netlink group,
>>> NETLINK_KOBJECT_UEVENT which we could monitor for events. The only thing
>>> is that we could only detect open/read/write/close/ioctl/lseek. And we
>>> probably do not want to monitor anything except block devices.
>>> 
>>> It may also be possible to poll /sys/block to watch for changes. This
>>> might be easier as the names are more friendly. This would take some
>>> research to see if its even possible.
>>> 
>>> The rule syntax could look something like:
>>> on=mount mount=/run/user/1000 : -a exit,always ...
>>> on=device device=/dev/sdd : -a exit,always ...
>>> 
>>> The on-login event would simply watch the audit trail for any AUDIT_LOGIN
>>> events. That event can be parsed to get the new auid. If the auid matches
>>> any rules, then it will load them into the kernel. To remove the rules,
>>> we
>>> could watch for the AUDIT_USER_END event. The only issue is that we would
>>> need to track how many sessions the user has open and remove the rules
>>> only when the last session closes out.
>>> 
>>> The rules for this might look something like this:
>>> on=login auid=1000 : -a exit,always ...
>>> 
>>> The question is whether or not this should be done as part of the audit
>>> daemon or as a plugin for the audit daemon. One advantage of doing this
>>> as
>>> a plugin is that it will keep the audit daemon focused on getting events
>>> and distributing them. Any programming mistake in the plugin will crash
>>> it
>>> and not the daemon. The tradeoff is that it will get the event slightly
>>> after auditd sees it. This only matters for the on-login functionality.
>>> The device and mount events come from an entirely different source. And
>>> I'm sure that in every case, the program will react faster than a user
>>> possibily can winning the race evry time.
>>> 
>>> Thoughts?
>> 
>> Would bind mounts trigger these rules? I'm sitting next to a box with 10k
>> polyinstantiated bind mounts right now.
> 
> If you do cat /proc/mounts do you see 10k entries?
No. This is over many users and applications launched with different contexts.

This morning, on a machine with one user, as a permissive root/sysadm_t:
cat /proc/*/mounts | grep inst | wc -l
2046
and from a user terminal session:
cat /proc/mounts | wc -l
218

> And do you want them or do you think they are harmful?

I thought they were all intentional when I started typing, but you have sent me down a rabbit hole looking into this :(

I like this idea, I just want to make sure that our atypical use case does not break it or vice versa.

cheers,
joe

> 
> -Steve