Basic guest memory introspection support
cover Basic guest memory introspection support
Hello,
We had to modify Xen in order to be able to detect rootkits in HVM guests, in a way that allows an application that runs in dom0 (or a similarly privileged domain) to control what the guest is allowed to do once a threat is detected. This has been done over the mem_event mechanism.
To this end, we needed to:
1. Be able to execute the current instruction without allowing it to write to memory. Also, be able to skip the current instruction completely (increment the guest's instruction pointer by the length of the current instruction). This is done based on new mem_event response fields sent from the controlling application.
2. Have the guest as responsive as possible amid all the processing. So we had to cache some information, namely a set of registers, with each mem_event sent.
3. Not allow the hypervisor to disable sending information about interesting MSR events.
4. Add an additional mem_event type, namely a VMCALL event - in order to do its work, our application occasionally triggers VMCALLs in the guest.
5. Add an additional libxc function that allows triggering page faults in the guest.
6. Since we're emulating the current instruction so that the guest can continue to run without lifting memory access restrictions, we've encountered problems with some instructions that Xen couldn't emulate. Our solution was a combination of working the 'accessed' and 'dirty' bits, and adding a re-execution mechanism for execute faults.
7. To make sure that pages that we've marked as interesting don't have their permissions reset.
Please note that while these patches can be applied directly to Xen master, we've tested them against commit 4bd78937ec324bcef4e29ef951e0ff9815770de1 (which was roughly before the new xc_mem_access_enable() came into libxc).
Patch list:
xen: Emulate with no writes; compute current instruction length
xen: Optimize introspection access to guest state
xen: Force-enable relevant MSR events; optimize the number of sent MSR events
xenctrl: Make the headers C++ friendly
xen: Support for VMCALL mem_events
xen, libxc: Request page fault injection via libxc
xen: Handle resumed instruction based on previous mem_event reply
xen: Generic instruction re-execution mechanism for execute faults
mm: mark pages that have their permissions controlled by a domain
All the best,
Razvan Cojocaru
Quoted 3 lines
Quoted 1 line
Quoted 1 line · On 02.07.14 at 15:34, <[email protected]> wrote:
4. Add an additional
mem_eventtype, namely a VMCALL event - in order to do its work, our application occasionally triggers VMCALLs in the guest.
That seems like bending the mem_event interface quite a bit - VMCALL isn't really memory (or event) specific.
Jan
Quoted 6 lines · Jan Beulich
Quoted 3 lines
Quoted 1 line
Quoted 1 line · On 02.07.14 at 15:34, <[email protected]> wrote:
4. Add an additional
mem_eventtype, namely a VMCALL event - in order to do its work, our application occasionally triggers VMCALLs in the guest.That seems like bending the
mem_eventinterface quite a bit - VMCALL isn't really memory (or event) specific.
First, thank you for the reply.
I don't disagree that it's somewhat bending mem_event, however in the context of our application it became necessary to know exactly when a triggered VMCALL occured, so conceptually this was a noteworthy event for us.
As for using mem_event for it, we chose it because it seems to be the closest thing to a standard Xen way of handling events, it is quite stable, and we thought it might not be such a stretch since there's also MEM_EVENT_REASON_SINGLESTEP.
We thought that adding an extra mem_event type would solve the problem while adding as little unnecessary complications to the existing code, and making it as trouble-free as possible for Xen maintainers / testers.
Thanks again,
Razvan Cojocaru
Quoted 7 lines · On Wed, 02 Jul 2014 15:30:24 +0100 Jan Beulich wrote:
Quoted 4 lines
Quoted 1 line
Quoted 1 line · On 02.07.14 at 15:34, <[email protected]> wrote:
4. Add an additional
mem_eventtype, namely a VMCALL event - in order to do its work, our application occasionally triggers VMCALLs in the guest.That seems like bending the
mem_eventinterface quite a bit - VMCALL isn't really memory (or event) specific.
Indeed. We've used that existing communication channel to carry MSR change events as well. Maybe in the feature we could rework the whole thing (mem, MSRs, VMCALL) under a new name, like 'guest events'.
Signature · 6 lines
Mihai DONȚU _______________________________________________ Xen-devel mailing list [email protected] http://lists.xen.org/xen-devel
Quoted 6 lines · On Wednesday 02 July 2014 17:50:53 Mihai Donțu wrote:
Quoted 5 lines · On Wed, 02 Jul 2014 15:30:24 +0100 Jan Beulich wrote:
Quoted 4 lines
Quoted 1 line
Quoted 1 line · On 02.07.14 at 15:34, <[email protected]> wrote:
4. Add an additional
mem_eventtype, namely a VMCALL event - in order to do its work, our application occasionally triggers VMCALLs in the guest.
Quoted 6 lines
Quoted 2 lines
That seems like bending the
mem_eventinterface quite a bit - VMCALL isn't really memory (or event) specific.Indeed. We've used that existing communication channel to carry MSR change events as well. Maybe in the feature we could rework the whole thing (mem, MSRs, VMCALL) under a new name, like 'guest events'.
Corrected Răzvan's email address.
Signature · 1 line
Mihai DONȚU
_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xen.org/xen-devel