On 14/02/18 13:19, Andrew Cooper wrote:
> On 14/02/18 12:15, Juergen Gross wrote:
>> On 14/02/18 13:03, Juergen Gross wrote:
>>> On 14/02/18 12:48, Andrew Cooper wrote:
>>>> On 14/02/18 07:54, Juergen Gross wrote:
>>>>> On 13/02/18 20:45, Andrew Cooper wrote:
>>>>>> The current XPTI implementation isolates the directmap (and therefore a 
>>>>>> lot of
>>>>>> guest data), but a large quantity of CPU0's state (including its stack)
>>>>>> remains visible.
>>>>>> Furthermore, an attacker able to read .text is in a vastly superior 
>>>>>> position
>>>>>> to normal when it comes to fingerprinting Xen for known vulnerabilities, 
>>>>>> or
>>>>>> scanning for ROP/Spectre gadgets.
>>>>>> Collect together the entrypoints in .text.entry (currently 3x4k frames, 
>>>>>> but
>>>>>> can almost certainly be slimmed down), and create a common mapping which 
>>>>>> is
>>>>>> inserted into each per-cpu shadow.  The stubs are also inserted into this
>>>>>> mapping by pointing at the in-use L2.  This allows stubs allocated later 
>>>>>> (SMP
>>>>>> boot, or CPU hotplug) to work without further changes to the common 
>>>>>> mappings.
>>>>>> Signed-off-by: Andrew Cooper <andrew.coop...@citrix.com>
>>>>>> ---
>>>>>> CC: Jan Beulich <jbeul...@suse.com>
>>>>>> CC: Wei Liu <wei.l...@citrix.com>
>>>>>> CC: Juergen Gross <jgr...@suse.com>
>>>>>> RFC, because I don't think the stubs handling is particularly sensible.
>>>>>> We allocate 4k of virtual address space per CPU, but squash loads of CPUs
>>>>>> together onto a single MFN.  The stubs ought to be isolated as well (as 
>>>>>> they
>>>>>> leak the virtual addresses of each stack), which can be done by 
>>>>>> allocating an
>>>>>> MFN per CPU (and simplifies cpu_smpboot_alloc() somewhat).  At this 
>>>>>> point, we
>>>>>> can't use a common set of mappings, and will have to clone the single 
>>>>>> stub and
>>>>>> .entry.text into each PCPUs copy of the pagetables.
>>>>>> Also, my plan to cause .text.entry to straddle a 512TB boundary (and 
>>>>>> therefore
>>>>>> avoid any further pagetable allocations) has come a little unstuck 
>>>>>> because of
>>>>>> CONFIG_BIGMEM.  I'm still working out whether there is a sensible way to
>>>>>> rearrange the virtual layout for this plan to work.
>>>>>> ---
>>>>>>  xen/arch/x86/smpboot.c             | 37 
>>>>>> ++++++++++++++++++++++++++++++++-----
>>>>>>  xen/arch/x86/x86_64/compat/entry.S |  2 ++
>>>>>>  xen/arch/x86/x86_64/entry.S        |  4 +++-
>>>>>>  xen/arch/x86/xen.lds.S             |  7 +++++++
>>>>>>  4 files changed, 44 insertions(+), 6 deletions(-)
>>>>>> diff --git a/xen/arch/x86/smpboot.c b/xen/arch/x86/smpboot.c
>>>>>> index 2ebef03..2519141 100644
>>>>>> --- a/xen/arch/x86/smpboot.c
>>>>>> +++ b/xen/arch/x86/smpboot.c
>>>>>> @@ -622,6 +622,9 @@ unsigned long alloc_stub_page(unsigned int cpu, 
>>>>>> unsigned long *mfn)
>>>>>>          unmap_domain_page(memset(__map_domain_page(pg), 0xcc, 
>>>>>> PAGE_SIZE));
>>>>>>      }
>>>>>> +    /* Confirm that all stubs fit in a single L2 pagetable. */
>>>>> So we limit NR_CPUS to be max 512 now?
>>>> Not intentionally.  The PAGE_SIZE should be dropped.  (One full L2
>>>> pagetable allows us to map 512*512 pages).
>>> L2_PAGETABLE_SHIFT is 21. So I still don't get why dropping PAGE_SIZE
>>> will correct it. OTOH I'm quite sure the BUILD_BUG_ON() won't trigger
>>> any more with PAGE_SIZE being dropped. :-)
>>>>> Maybe you should use STUB_BUF_SIZE instead of PAGE_SIZE?
>>>> No - that would be incorrect because of the physical packing of stubs
>>>> which occurs.
>>>>> BTW: Is there any reason we don't use a common stub page mapped to each
>>>>> per-cpu stack area? The stack address can easily be obtained via %rip
>>>>> relative addressing then (see my patch for the per-vcpu stacks:
>>>>> https://lists.xen.org/archives/html/xen-devel/2018-02/msg00773.html )
>>>> I don't understand what you are asking here.  We cannot access the
>>>> per-cpu area with plain rip-retaliative addressing without using gs base
>>>> (and we really don't want to go down that route), or without per-cpu
>>>> pagetables (which would have to be a compile time choice).
>>> The stub-page of a cpu is currently mapped as the 3rd page of the
>>> stack area. So the distance to the primary stack would be the same
>>> for all cpus (a little bit less than 20kB).
>>>> As for why the per-cpu areas aren't mapped, that's because they aren't
>>>> needed at the moment.  Any decision to change this needs to weigh the
>>>> utility of mapping the areas vs the additional data leakage, which is
>>>> substantial.
>>> The stack area is mapped. And that's where the stub is living.
>> Oh, did I mix up something? I followed the comments in current.h. The
>> code suggests the syscall trampoline page isn't used at the moment for
>> the stubs...
> That will be stale from the work Jan did to make the stack fully NX. 
> The syscall stubs used to be on the stack, but are no longer.

Changing this to make the syscall stub RO and executable again isn't
impossible, I guess. Especially when this will make life easier at
other places we should at least consider that option.


Xen-devel mailing list

Reply via email to